Digital PDFs

EC-QAEQD-TE

July 1996

450 pages

Original

2.1MB

Document:	Alpha 21164 Microprocessor Hardware Reference Manual
Order Number:	EC-QAEQD-TE
Revision:	0
Pages:	450
Original Filename:

OCR Text

Alpha 21164 Microprocessor
Hardware Reference Manual
Order Number: EC–QAEQD–TE
Revision/Update Information:

Digital Equipment Corporation
Maynard, Massachusetts

This preliminary document supersedes
the Alpha 21164 Microprocessor Hardware
Reference Manual (EC–QAEQC–TE).

July 1996
Possession, use, or copying of the software described in this publication is authorized only
pursuant to a valid written license from Digital or an authorized sublicensor.
While Digital believes the information included in this publication is correct as of the date of
publication, it is subject to change without notice.
Digital Equipment Corporation makes no representations that the use of its products in the
manner described in this publication will not infringe on existing or future patent rights, nor do
the descriptions contained in this publication imply the granting of licenses to make, use, or sell
equipment or software in accordance with the description.
© Digital Equipment Corporation 1994, 1995, 1996.
All rights reserved.
Printed in U.S.A.
AlphaGeneration, DEC, DECchip, Digital, Digital Semiconductor, OpenVMS, VAX, VAX
DOCUMENT, the AlphaGeneration design mark, and the DIGITAL logo are trademarks of
Digital Equipment Corporation.
Digital Semiconductor is a Digital Equipment Corporation business.
GRAFOIL is a registered trademark of Union Carbide Corporation.
Hewlett-Packard is a registered trademark of Hewlett-Packard Company.
IEEE is a registered trademark of The Institute of Electrical and Electronics Engineers, Inc.
Prentice Hall is a registered trademark of Prentice-Hall, Inc. of Englewood Cliffs, NJ.
Windows NT is a trademark of Microsoft Corporation.
All other trademarks and registered trademarks are the property of their respective owners.

This document was prepared using VAX DOCUMENT Version 2.1.

Contents
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

xxi

1 Introduction
1.1
1.1.1
1.1.2
1.1.3
1.2

The Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Addressing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Integer Data Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Floating-Point Data Types . . . . . . . . . . . . . . . . . . . . . . . . . . .
Alpha 21164 Microprocessor Features . . . . . . . . . . . . . . . . . . . . .

1–1
1–2
1–2
1–3
1–3

2 Internal Architecture
2.1
Alpha 21164 Microarchitecture . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.1
Instruction Fetch/Decode Unit and Branch Unit . . . . . . . . . .
2.1.1.1
Instruction Decode and Issue . . . . . . . . . . . . . . . . . . . . . .
2.1.1.2
Instruction Prefetch . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.1.3
Branch Execution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.1.4
Instruction Translation Buffer . . . . . . . . . . . . . . . . . . . . .
2.1.1.5
Interrupts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.2
Integer Execution Unit . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.3
Floating-Point Execution Unit . . . . . . . . . . . . . . . . . . . . . . . .
2.1.4
Memory Address Translation Unit . . . . . . . . . . . . . . . . . . . . .
2.1.4.1
Data Translation Buffer . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.4.2
Load Instruction and the Miss Address File . . . . . . . . . .
2.1.4.3
Dcache Control and Store Instructions . . . . . . . . . . . . . . .
2.1.4.4
Write Buffer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.5
Cache Control and Bus Interface Unit . . . . . . . . . . . . . . . . . .
2.1.6
Cache Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.6.1
Data Cache . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.6.2
Instruction Cache . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.6.3
Second-Level Cache . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.6.4
External Cache . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.7
Serial Read-Only Memory Interface . . . . . . . . . . . . . . . . . . . .

2–2
2–4
2–4
2–5
2–6
2–7
2–8
2–9
2–10
2–10
2–11
2–11
2–12
2–12
2–12
2–13
2–13
2–13
2–13
2–14
2–14

iii

2.2
Pipeline Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2.1
Pipeline Stages and Instruction Issue . . . . . . . . . . . . . . . . . .
2.2.2
Aborts and Exceptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2.3
Nonissue Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3
Scheduling and Issuing Rules . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3.1
Instruction Class Definition and Instruction Slotting . . . . . .
2.3.2
Coding Guidelines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3.3
Instruction Latencies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3.3.1
Producer–Producer Latency . . . . . . . . . . . . . . . . . . . . . . .
2.3.4
Issue Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.4
Replay Traps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.5
Miss Address File and Load-Merging Rules . . . . . . . . . . . . . . . . .
2.5.1
Merging Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.5.2
Read Requests to the Cbox . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.5.3
Load Instructions to Noncacheable Space . . . . . . . . . . . . . . .
2.5.4
MAF Entries and MAF Full Conditions . . . . . . . . . . . . . . . . .
2.5.5
Fill Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.6
Mbox Store Instruction Execution . . . . . . . . . . . . . . . . . . . . . . . .
2.7
Write Buffer and the WMB Instruction . . . . . . . . . . . . . . . . . . . .
2.7.1
The Write Buffer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.7.2
The Write Memory Barrier (WMB) Instruction . . . . . . . . . . .
2.7.3
Entry-Pointer Queues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.7.4
Write Buffer Entry Processing . . . . . . . . . . . . . . . . . . . . . . . .
2.7.5
Ordering of Noncacheable Space Write Instructions . . . . . . .
2.8
Performance Measurement Support–Performance Counters . . . .
2.9
Floating-Point Control Register . . . . . . . . . . . . . . . . . . . . . . . . . .
2.10
Design Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2–14
2–18
2–18
2–20
2–20
2–20
2–23
2–24
2–27
2–28
2–29
2–30
2–30
2–31
2–31
2–32
2–32
2–33
2–35
2–35
2–35
2–36
2–36
2–37
2–38
2–38
2–40

3 Hardware Interface
3.1
3.2

Alpha 21164 Microprocessor Logic Symbol . . . . . . . . . . . . . . . . .
Alpha 21164 Signal Names and Functions . . . . . . . . . . . . . . . . .

3–1
3–3

4 Clocks, Cache, and External Interface Functional Description
4.1
Introduction to the External Interface . . . . . . . . . . . . . . . . . . . . .
4.1.1
System Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.1.1.1
Commands and Addresses . . . . . . . . . . . . . . . . . . . . . . . .
4.1.2
Bcache Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2
Clocks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2.1
CPU Clock . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2.2
System Clock . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2.3
Delayed System Clock . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

4–2
4–2
4–4
4–4
4–5
4–5
4–6
4–8

Reference Clock . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Reference Clock Examples . . . . . . . . . . . . . . . . . . . . . . . .
Case 1: ref_clk_in_h Initially Sampled Low by
DPLL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2.4.1.2
Case 2: ref_clk_in_h Initially Sampled High by
DPLL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.3
Physical Address Considerations . . . . . . . . . . . . . . . . . . . . . . . . .
4.3.1
Physical Address Regions . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.3.2
Data Wrapping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.3.3
Noncached Read Operations . . . . . . . . . . . . . . . . . . . . . . . . .
4.3.4
Noncached Write Operations . . . . . . . . . . . . . . . . . . . . . . . . .
4.4
Bcache Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.4.1
Duplicate Tag Store . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.4.1.1
Full Duplicate Tag Store . . . . . . . . . . . . . . . . . . . . . . . . .
4.4.1.2
Partial Scache Duplicate Tag Store . . . . . . . . . . . . . . . . .
4.4.2
Bcache Victim Buffers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.5
Systems Without a Bcache . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.6
Cache Coherency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.6.1
Cache Coherency Basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.6.2
Write Invalidate Cache Coherency Protocol Systems . . . . . . .
4.6.3
Write Invalidate Cache Coherency States . . . . . . . . . . . . . . .
4.6.3.1
Write Invalidate Protocol State Machines . . . . . . . . . . . .
4.6.4
Flush Cache Coherency Protocol Systems . . . . . . . . . . . . . . .
4.6.5
Flush-Based Protocol State Machines . . . . . . . . . . . . . . . . . .
4.6.6
Cache Coherency Transaction Conflicts . . . . . . . . . . . . . . . . .
4.6.6.1
Case 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.6.6.2
Case 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.7
Lock Mechanisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.8
Alpha 21164-to-Bcache Transactions . . . . . . . . . . . . . . . . . . . . . .
4.8.1
Bcache Timing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.8.2
Bcache Read Transaction (Private Read Operation) . . . . . . . .
4.8.3
Wave Pipeline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.8.4
Bcache Write Transaction (Private Write Operation) . . . . . . .
4.8.5
Selecting Bcache Options . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.9
Alpha 21164-Initiated System Transactions . . . . . . . . . . . . . . . .
4.9.1
READ MISS—No Bcache . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.9.2
READ MISS—Bcache . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.9.3
FILL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.9.4
READ MISS with Victim . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.9.4.1
READ MISS with Victim (Victim Buffer) . . . . . . . . . . . . .
4.9.4.2
READ MISS with Victim (Without Victim Buffer) . . . . . .
4.9.5
WRITE BLOCK and WRITE BLOCK LOCK . . . . . . . . . . . . .
4.9.6
SET DIRTY and LOCK . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2.4
4.2.4.1
4.2.4.1.1

4–8
4–9
4–10
4–11
4–12
4–12
4–13
4–14
4–14
4–15
4–15
4–16
4–18
4–18
4–19
4–19
4–19
4–22
4–23
4–24
4–25
4–27
4–28
4–28
4–29
4–30
4–31
4–31
4–32
4–33
4–34
4–35
4–36
4–40
4–41
4–43
4–43
4–44
4–46
4–48
4–50

4.9.7
Memory Barrier (MB) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.9.7.1
When to Use a MEMORY BARRIER Command . . . . . . . .
4.9.8
FETCH . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.9.9
FETCH_M . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.10
System-Initiated Transactions . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.10.1
Sending Commands to the 21164 . . . . . . . . . . . . . . . . . . . . . .
4.10.2
Write Invalidate Protocol Commands . . . . . . . . . . . . . . . . . . .
4.10.2.1
Alpha 21164 Responses to Write Invalidate Protocol
Commands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.10.2.2
READ DIRTY and READ DIRTY/INVALIDATE . . . . . . .
4.10.2.3
INVALIDATE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.10.2.4
SET SHARED . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.10.3
Flush-Based Cache Coherency Protocol Commands . . . . . . . .
4.10.3.1
Alpha 21164 Responses to Flush-Based Protocol
Commands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.10.3.2
FLUSH . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.10.3.3
READ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.11
Data Bus and Command/Address Bus Contention . . . . . . . . . . . .
4.11.1
Command/Address Bus . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.11.2
Read/Write Spacing—Data Bus Contention . . . . . . . . . . . . . .
4.11.3
Using idle_bc_h and fill_h . . . . . . . . . . . . . . . . . . . . . . . . . .
4.11.4
Using data_bus_req_h . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.11.5
Tristate Overlap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.11.5.1
READ or WRITE to FILL . . . . . . . . . . . . . . . . . . . . . . . .
4.11.5.2
BCACHE VICTIM to FILL . . . . . . . . . . . . . . . . . . . . . . .
4.11.5.3
System Bcache Command to FILL . . . . . . . . . . . . . . . . . .
4.11.5.4
FILL to Private Read or Write Operation . . . . . . . . . . . .
4.12
Alpha 21164 Interface Restrictions . . . . . . . . . . . . . . . . . . . . . . .
4.12.1
FILL Operations after Other Transactions . . . . . . . . . . . . . . .
4.12.2
Command Acknowledge for WRITE BLOCK Commands . . . .
4.12.3
Systems Without a Bcache . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.12.4
Fast Probes with No Bcache . . . . . . . . . . . . . . . . . . . . . . . . .
4.12.5
WRITE BLOCK LOCK . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.13
Alpha 21164/System Race Conditions . . . . . . . . . . . . . . . . . . . . .
4.13.1
Rules for 21164 and System Use of External Interface . . . . .
4.13.2
READ MISS with Victim Example . . . . . . . . . . . . . . . . . . . . .
4.13.3
idle_bc_h and cack_h Race Example . . . . . . . . . . . . . . . . . .
4.13.4
READ MISS with idle_bc_h Asserted Example . . . . . . . . . .
4.13.5
READ MISS with Victim Abort Example . . . . . . . . . . . . . . . .
4.13.6
Bcache Hit Under READ MISS Example . . . . . . . . . . . . . . . .
4.14
Data Integrity, Bcache Errors, and Command/Address Errors . . .
4.14.1
Data ECC and Parity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.14.2
Force Correction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

4–52
4–52
4–52
4–52
4–53
4–53
4–55
4–56
4–58
4–60
4–62
4–64
4–65
4–66
4–68
4–70
4–70
4–71
4–72
4–74
4–75
4–75
4–75
4–78
4–80
4–81
4–81
4–81
4–81
4–81
4–82
4–83
4–83
4–84
4–86
4–88
4–89
4–90
4–92
4–92
4–94

4.14.3
Bcache Tag Data Parity . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.14.4
Bcache Tag Control Parity . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.14.5
Address and Command Parity . . . . . . . . . . . . . . . . . . . . . . . .
4.14.6
Fill Error . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.14.7
Forcing 21164 Reset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.15
Interrupts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.15.1
Interrupt Signals During Initialization . . . . . . . . . . . . . . . . .
4.15.2
Interrupt Signals During Normal Operation . . . . . . . . . . . . .
4.15.3
Interrupt Priority Level . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

4–94
4–94
4–95
4–95
4–95
4–96
4–96
4–96
4–96

5 Internal Processor Registers
5.1
5.1.1
5.1.2
5.1.3
5.1.4
5.1.5
5.1.6
5.1.7
5.1.8
5.1.9
5.1.10
5.1.11
5.1.12
5.1.13
5.1.14
5.1.15
5.1.16
5.1.17
5.1.18
5.1.19
5.1.20
5.1.21
5.1.22
5.1.23
5.1.24

Instruction Fetch/Decode Unit and Branch Unit (Ibox) IPRs . . . .
Istream Translation Buffer Tag Register (ITB_TAG) . . . . . . .
Instruction Translation Buffer Page Table Entry (ITB_PTE)
Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Instruction Translation Buffer Address Space Number
(ITB_ASN) Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Instruction Translation Buffer Page Table Entry Temporary
(ITB_PTE_TEMP) Register . . . . . . . . . . . . . . . . . . . . . . . . . .
Instruction Translation Buffer Invalidate All Process
(ITB_IAP) Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Instruction Translation Buffer Invalidate All (ITB_IA)
Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Instruction Translation Buffer IS (ITB_IS) Register . . . . . . .
Formatted Faulting Virtual Address (IFAULT_VA_FORM)
Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Virtual Page Table Base Register (IVPTBR) . . . . . . . . . . . . .
Icache Parity Error Status (ICPERR_STAT) Register . . . . . .
Icache Flush Control (IC_FLUSH_CTL) Register . . . . . . . . .
Exception Address (EXC_ADDR) Register . . . . . . . . . . . . . .
Exception Summary (EXC_SUM) Register . . . . . . . . . . . . . .
Exception Mask (EXC_MASK) Register . . . . . . . . . . . . . . . .
PAL Base Address (PAL_BASE) Register . . . . . . . . . . . . . . .
Ibox Current Mode (ICM) Register . . . . . . . . . . . . . . . . . . . . .
Ibox Control and Status Register (ICSR) . . . . . . . . . . . . . . . .
Interrupt Priority Level Register (IPLR) . . . . . . . . . . . . . . . .
Interrupt ID (INTID) Register . . . . . . . . . . . . . . . . . . . . . . .
Asynchronous System Trap Request Register (ASTRR) . . . . .
Asynchronous System Trap Enable Register (ASTER) . . . . . .
Software Interrupt Request Register (SIRR) . . . . . . . . . . . . .
Hardware Interrupt Clear (HWINT_CLR) Register . . . . . . .
Interrupt Summary Register (ISR) . . . . . . . . . . . . . . . . . . . .

5–5
5–5
5–6
5–8
5–9
5–9
5–9
5–10
5–11
5–12
5–13
5–13
5–14
5–15
5–17
5–18
5–19
5–20
5–23
5–24
5–25
5–26
5–27
5–28
5–29

vii

5.1.25
Serial Line Transmit (SL_XMIT) Register . . . . . . . . . . . . . . .
5.1.26
Serial Line Receive (SL_RCV) Register . . . . . . . . . . . . . . . . .
5.1.27
Performance Counter (PMCTR) Register . . . . . . . . . . . . . . .
5.2
Memory Address Translation Unit (Mbox) IPRs . . . . . . . . . . . . . .
5.2.1
Dstream Translation Buffer Address Space Number
(DTB_ASN) Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2.2
Dstream Translation Buffer Current Mode (DTB_CM)
Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2.3
Dstream Translation Buffer Tag (DTB_TAG) Register . . . . .
5.2.4
Dstream Translation Buffer Page Table Entry (DTB_PTE)
Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2.5
Dstream Translation Buffer Page Table Entry Temporary
(DTB_PTE_TEMP) Register . . . . . . . . . . . . . . . . . . . . . . . . .
5.2.6
Dstream Memory Management Fault Status (MM_STAT)
Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2.7
Faulting Virtual Address (VA) Register . . . . . . . . . . . . . . . . .
5.2.8
Formatted Virtual Address (VA_FORM) Register . . . . . . . . .
5.2.9
Mbox Virtual Page Table Base Register (MVPTBR) . . . . . . . .
5.2.10
Dcache Parity Error Status (DC_PERR_STAT) Register . . . .
5.2.11
Dstream Translation Buffer Invalidate All Process
(DTB_IAP) Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2.12
Dstream Translation Buffer Invalidate All (DTB_IA) Register
................................................
5.2.13
Dstream Translation Buffer Invalidate Single (DTB_IS)
Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2.14
Mbox Control Register (MCSR) . . . . . . . . . . . . . . . . . . . . . . .
5.2.15
Dcache Mode (DC_MODE) Register . . . . . . . . . . . . . . . . . . .
5.2.16
Miss Address File Mode (MAF_MODE) Register . . . . . . . . .
5.2.17
Dcache Flush (DC_FLUSH) Register . . . . . . . . . . . . . . . . . .
5.2.18
Alternate Mode (ALT_MODE) Register . . . . . . . . . . . . . . . . .
5.2.19
Cycle Counter (CC) Register . . . . . . . . . . . . . . . . . . . . . . . . .
5.2.20
Cycle Counter Control (CC_CTL) Register . . . . . . . . . . . . . .
5.2.21
Dcache Test Tag Control (DC_TEST_CTL) Register . . . . . . .
5.2.22
Dcache Test Tag (DC_TEST_TAG) Register . . . . . . . . . . . . .
5.2.23
Dcache Test Tag Temporary (DC_TEST_TAG_TEMP)
Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.3
External Interface Control (Cbox) IPRs . . . . . . . . . . . . . . . . . . . .
5.3.1
Scache Control (SC_CTL) Register (FF FFF0 00A8) . . . . . . .
5.3.2
Scache Status (SC_STAT) Register (FF FFF0 00E8) . . . . . . .
5.3.3
Scache Address (SC_ADDR) Register (FF FFF0 0188) . . . . . .
5.3.4
Bcache Control (BC_CONTROL) Register
(FF FFF0 0128) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

viii

5–31
5–32
5–33
5–38
5–38
5–39
5–40
5–41
5–43
5–44
5–46
5–47
5–49
5–50
5–52
5–52
5–53
5–54
5–56
5–58
5–60
5–60
5–61
5–62
5–63
5–64
5–66
5–68
5–69
5–72
5–75
5–78

5.3.5
5.3.6
5.3.7
5.3.8
5.3.9
5.4
5.5
5.5.1
5.5.2

Bcache Configuration (BC_CONFIG) Register
(FF FFF0 01C8) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Bcache Tag Address (BC_TAG_ADDR) Register
(FF FFF0 0108) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
External Interface Status (EI_STAT) Register
(FF FFF0 0168) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
External Interface Address (EI_ADDR) Register
(FF FFF0 0148) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Fill Syndrome (FILL_SYN) Register (FF FFF0 0068) . . . . . .
PALcode Storage Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Restrictions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Cbox IPR PALcode Restrictions . . . . . . . . . . . . . . . . . . . . . . .
PALcode Restrictions—Instruction Definitions . . . . . . . . . . . .

5–84
5–89
5–91
5–94
5–95
5–99
5–100
5–100
5–101

6 Privileged Architecture Library Code
6.1
6.2
6.3
6.4
6.4.1
6.4.2
6.5
6.6
6.6.1
6.6.2
6.6.3
6.6.4

PALcode Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
PALmode Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Invoking PALcode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
PALcode Entry Points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
CALL_PAL Entry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
PALcode Trap Entry Points . . . . . . . . . . . . . . . . . . . . . . . . . .
Required PALcode Function Codes . . . . . . . . . . . . . . . . . . . . . . . .
Alpha 21164 Implementation of the Architecturally Reserved
Opcodes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
HW_LD Instruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
HW_ST Instruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
HW_REI Instruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
HW_MFPR and HW_MTPR Instructions . . . . . . . . . . . . . . . .

6–1
6–2
6–3
6–5
6–5
6–6
6–7
6–7
6–8
6–10
6–11
6–11

7 Initialization and Configuration
7.1
7.1.1
7.2
7.3
7.4
7.4.1
7.5
7.6
7.6.1
7.6.2
7.7

Input Signals sys_reset_l and dc_ok_h and Booting . . . . . . . . .
Pin State with dc_ok_h Not Asserted . . . . . . . . . . . . . . . . . .
Sysclk Ratio and Delay . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Built-In Self-Test (BiSt) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Serial Read-Only Memory Interface Port . . . . . . . . . . . . . . . . . . .
Serial Instruction Cache Load Operation . . . . . . . . . . . . . . . .
Serial Terminal Port . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Cache Initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Icache Initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Flushing Dirty Blocks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
External Interface Initialization . . . . . . . . . . . . . . . . . . . . . . . . . .

7–1
7–5
7–6
7–6
7–6
7–7
7–8
7–8
7–8
7–9
7–10

7.8
7.9
7.10

Internal Processor Register Reset State . . . . . . . . . . . . . . . . . . . .
Timeout Reset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
IEEE 1149.1 Test Port Reset . . . . . . . . . . . . . . . . . . . . . . . . . . . .

7–10
7–13
7–13

8 Error Detection and Error Handling
8.1
8.1.1
8.1.2
8.1.3
8.1.4
8.1.5
8.1.6
8.1.7
8.1.8
8.1.9
8.1.10
8.1.11
8.1.12
8.1.13
8.1.14
8.1.15
8.1.16
8.1.17
8.1.18
8.2
8.3
8.4
8.5

Error Flows . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Icache Data or Tag Parity Error . . . . . . . . . . . . . . . . . . . . . .
Scache Data Parity Error—Istream . . . . . . . . . . . . . . . . . . . .
Scache Tag Parity Error—Istream . . . . . . . . . . . . . . . . . . . . .
Scache Data Parity Error—Dstream Read/Write,
READ_DIRTY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Scache Tag Parity Error—Dstream or System Commands . . .
Dcache Data Parity Error . . . . . . . . . . . . . . . . . . . . . . . . . . .
Dcache Tag Parity Error . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Istream Uncorrectable ECC or Data Parity Errors (Bcache or
Memory) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Dstream Uncorrectable ECC or Data Parity Errors (Bcache
or Memory) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Bcache Tag Parity Errors—Istream . . . . . . . . . . . . . . . . . . . .
Bcache Tag Parity Errors—Dstream . . . . . . . . . . . . . . . . . . .
System Command/Address Parity Error . . . . . . . . . . . . . . . .
System Read Operations of the Bcache . . . . . . . . . . . . . . . . .
Istream or Dstream Correctable ECC Error (Bcache or
Memory) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Fill Timeout (FILL_ERROR_H) . . . . . . . . . . . . . . . . . . . . . . .
System Machine Check . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Ibox Timeout . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
cfail_h and Not cack_h . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
MCHK Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Processor-Correctable Error Interrupt Flow (IPL 31) . . . . . . . . . .
MCK_INTERRUPT Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
System-Correctable Error Interrupt Flow (IPL 20) . . . . . . . . . . .

8–1
8–1
8–2
8–3
8–3
8–4
8–4
8–5
8–5
8–6
8–7
8–7
8–8
8–8
8–9
8–9
8–10
8–10
8–10
8–11
8–13
8–14
8–14

9 Electrical Data
9.1
9.2
9.2.1
9.2.2
9.2.3
9.3
9.3.1

Electrical Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
dc Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Power Supply . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Input Signal Pins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Output Signal Pins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Clocking Scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Input Clocks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

9–1
9–2
9–2
9–2
9–2
9–4
9–4

9.3.2
Clock Termination and Impedance Levels . . . . . . . . . . . . . . .
9.3.3
ac Coupling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9.4
ac Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9.4.1
Test Configuration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9.4.2
Pin Timing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9.4.2.1
Backup Cache Loop Timing . . . . . . . . . . . . . . . . . . . . . . .
9.4.2.2
sys_clk-Based Systems . . . . . . . . . . . . . . . . . . . . . . . . . . .
9.4.2.3
Reference Clock-Based Systems . . . . . . . . . . . . . . . . . . . .
9.4.3
Digital Phase-Locked Loop . . . . . . . . . . . . . . . . . . . . . . . . . . .
9.4.4
Timing—Additional Signals . . . . . . . . . . . . . . . . . . . . . . . . . .
9.4.5
Timing of Test Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9.4.6
Icache BiSt Operation Timing . . . . . . . . . . . . . . . . . . . . . . . .
9.4.7
Automatic SROM Load Timing . . . . . . . . . . . . . . . . . . . . . . .
9.4.8
Clock Test Modes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9.4.8.1
Normal Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9.4.8.2
Chip Test Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9.4.8.3
Module Test Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9.4.8.4
Clock Test Reset Mode . . . . . . . . . . . . . . . . . . . . . . . . . . .
9.4.9
IEEE 1149.1 (JTAG) Performance . . . . . . . . . . . . . . . . . . . . .
9.5
Power Supply Considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9.5.1
Decoupling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9.5.2
Power Supply Sequencing . . . . . . . . . . . . . . . . . . . . . . . . . . .

9–6
9–6
9–9
9–9
9–10
9–10
9–12
9–15
9–16
9–18
9–21
9–21
9–23
9–25
9–25
9–25
9–25
9–25
9–26
9–26
9–26
9–27

10 Thermal Management
10.1
10.2
10.3

Operating Temperature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Heat Sink Specifications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Thermal Design Considerations . . . . . . . . . . . . . . . . . . . . . . . . . .

10–1
10–3
10–4

11 Mechanical Data and Packaging Information
11.1
Mechanical Specifications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11.2
Signal Descriptions and Pin Assignment . . . . . . . . . . . . . . . . . . .
11.2.1
Signal Pin Lists . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11.2.2
Pin Assignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

11–1
11–3
11–3
11–8

12 Testability and Diagnostics
12.1
Test Port Pins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
12.2
Test Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
12.2.1
IEEE 1149.1 Test Access Port . . . . . . . . . . . . . . . . . . . . . . . .
12.2.2
Test Status Pins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
12.3
Boundary Scan Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

12–1
12–2
12–2
12–6
12–7

A Alpha Instruction Set
A.1
A.1.1
A.1.2
A.2
A.3
A.4
A.5
A.6

Alpha Instruction Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Opcodes Reserved for Digital . . . . . . . . . . . . . . . . . . . . . . . . .
Opcodes Reserved for PALcode . . . . . . . . . . . . . . . . . . . . . . . .
IEEE Floating-Point Instructions . . . . . . . . . . . . . . . . . . . . . . . . .
VAX Floating-Point Instructions . . . . . . . . . . . . . . . . . . . . . . . . .
Opcode Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Required PALcode Function Codes . . . . . . . . . . . . . . . . . . . . . . . .
Alpha 21164 Microprocessor IEEE Floating-Point
Conformance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

A–1
A–6
A–7
A–7
A–9
A–10
A–12
A–12

Alpha 21164 Microprocessor Specifications

C Serial Icache Load Predecode Values
D Errata Sheet
E Technical Support and Ordering Information
E.1
E.2
E.3
E.4
E.5

Glossary

xii

Technical Support . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Ordering Digital Semiconductor Products . . . . . . . . . . . . . . . . . .
Ordering Digital Semiconductor Sample Kits . . . . . . . . . . . . . . .
Ordering Associated Literature . . . . . . . . . . . . . . . . . . . . . . . . . .
Ordering Associated Third-Party Literature . . . . . . . . . . . . . . . .

E–1
E–1
E–2
E–2
E–3

Index
Figures
2–1
2–2
2–3
2–4
2–5
2–6
3–1
4–1
4–2
4–3
4–4
4–5
4–6
4–7
4–8
4–9
4–10
4–11
4–12
4–13
4–14
4–15
4–16
4–17
4–18
4–19
4–20
4–21
4–22
4–23
4–24
4–25

Alpha 21164 Microprocessor Block/Pipe Flow Diagram . . . . .
Instruction Pipeline Stages . . . . . . . . . . . . . . . . . . . . . . . . . .
Floating-Point Control Register (FPCR) Format . . . . . . . . . .
Typical Uniprocessor Configuration . . . . . . . . . . . . . . . . . . . .
Typical Multiprocessor Configuration . . . . . . . . . . . . . . . . . . .
Cacheless Multiprocessor Configuration . . . . . . . . . . . . . . . . .
Alpha 21164 Microprocessor Logic Symbol . . . . . . . . . . . . . .
Alpha 21164 System/Bcache Interface . . . . . . . . . . . . . . . . . .
Clock Signals and Functions . . . . . . . . . . . . . . . . . . . . . . . . .
Alpha 21164 Uniprocessor Clock . . . . . . . . . . . . . . . . . . . . . .
Alpha 21164 Reference Clock for Multiprocessor Systems . .
ref_clk_in_h Initially Sampled Low . . . . . . . . . . . . . . . . . . .
ref_clk_in_h Initially Sampled High . . . . . . . . . . . . . . . . . . .
Full Scache Duplicate Tag Store . . . . . . . . . . . . . . . . . . . . . .
Duplicate Tag Store Algorithm . . . . . . . . . . . . . . . . . . . . . . . .
Partial Scache Duplicate Tag Store . . . . . . . . . . . . . . . . . . . .
Cache Subset Hierarchy . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Write Invalidate Protocol: 21164 State Transitions . . . . . . . .
Write Invalidate Protocol: System/Bus State Transitions . . .
Flush-Based Protocol 21164 States . . . . . . . . . . . . . . . . . . . .
Flush-Based Protocol System/Bus States . . . . . . . . . . . . . . . .
Bcache Read Transaction . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Wave Pipeline Timing Diagram . . . . . . . . . . . . . . . . . . . . . . .
Bcache Write Transaction . . . . . . . . . . . . . . . . . . . . . . . . . . .
READ MISS—No Bcache Timing Diagram . . . . . . . . . . . . . .
READ MISS MOD—Bcache Timing Diagram . . . . . . . . . . . .
READ MISS with Victim (Victim Buffer) Timing Diagram . .
READ MISS with Victim (without Victim Buffer) Timing
Diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
WRITE BLOCK Timing Diagram . . . . . . . . . . . . . . . . . . . . . .
SET DIRTY and LOCK Timing Diagram . . . . . . . . . . . . . . . .
Algorithm for System Sending Commands to the 21164 . . . .
READ DIRTY Timing Diagram (Scache Hit) . . . . . . . . . . . . .

2–3
2–15
2–39
2–41
2–42
2–43
3–2
4–3
4–6
4–7
4–9
4–10
4–11
4–16
4–17
4–18
4–20
4–24
4–25
4–28
4–28
4–32
4–33
4–34
4–40
4–42
4–45
4–47
4–49
4–51
4–54
4–59

xiii

4–26
4–27
4–28
4–29
4–30
4–31
4–32
4–33
4–34
4–35
4–36
4–37
4–38
4–39
4–40
4–41
4–42
4–43
4–44
5–1
5–2
5–3
5–4
5–5
5–6
5–7
5–8
5–9
5–10
5–11
5–12
5–13
5–14
xiv

INVALIDATE Timing Diagram—Bcache Hit . . . . . . . . . . . . .
SET SHARED Timing Diagram . . . . . . . . . . . . . . . . . . . . . . .
FLUSH Timing Diagram (Scache Hit) . . . . . . . . . . . . . . . . . .
READ Timing Diagram (Scache Hit) . . . . . . . . . . . . . . . . . . .
Driving the Command/Address Bus . . . . . . . . . . . . . . . . . . . .
Example of Using idle_bc_h and fill_h . . . . . . . . . . . . . . . .
Using data_bus_req_h . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
READ MISS Completed First—Victim Buffer . . . . . . . . . . . .
READ MISS Second—No Victim Buffer . . . . . . . . . . . . . . . . .
System Command to FILL Example 1 . . . . . . . . . . . . . . . . . .
System Command to FILL Example 2 . . . . . . . . . . . . . . . . . .
FILL to Private Read or Write Operation . . . . . . . . . . . . . . .
READ MISS with Victim Example . . . . . . . . . . . . . . . . . . . . .
idle_bc_h and cack_h Race Example . . . . . . . . . . . . . . . . . . .
READ MISS with idle_bc_h Asserted Example . . . . . . . . . .
READ MISS with Victim Abort Example . . . . . . . . . . . . . . . .
Bcache Hit Under READ MISS Example . . . . . . . . . . . . . . . .
ECC Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Alpha 21164 Interrupt Signals . . . . . . . . . . . . . . . . . . . . . . .
Istream Translation Buffer Tag Register (ITB_TAG) . . . . . . .
Instruction Translation Buffer Page Table Entry (ITB_PTE)
Register Write Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Instruction Translation Buffer Page Table Entry (ITB_PTE)
Register Read Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Instruction Translation Buffer Address Space Number
(ITB_ASN) Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Instruction Translation Buffer IS (ITB_IS) Register . . . . . . .
Formatted Faulting Virtual Address (IFAULT_VA_FORM)
Register (NT_Mode=0) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Formatted Faulting Virtual Address (IFAULT_VA_FORM)
Register (NT_Mode=1) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Virtual Page Table Base Register (IVPTBR) (NT_Mode=0) . . .
Virtual Page Table Base Register (IVPTBR) (NT_Mode=1) . . .
Icache Parity Error Status (ICPERR_STAT) Register . . . . . .
Exception Address (EXC_ADDR) Register . . . . . . . . . . . . . .
Exception Summary (EXC_SUM) Register . . . . . . . . . . . . . .
Exception Mask (EXC_MASK) Register . . . . . . . . . . . . . . . .
PAL Base Address (PAL_BASE) Register . . . . . . . . . . . . . . .

4–61
4–63
4–67
4–69
4–70
4–73
4–74
4–76
4–77
4–78
4–79
4–80
4–85
4–87
4–88
4–90
4–91
4–93
4–96
5–5
5–6
5–7
5–8
5–10
5–11
5–11
5–12
5–12
5–13
5–14
5–15
5–17
5–18

5–15
5–16
5–17
5–18
5–19
5–20
5–21
5–22
5–23
5–24
5–25
5–26
5–27
5–28
5–29
5–30
5–31
5–32
5–33
5–34
5–35
5–36
5–37
5–38
5–39
5–40
5–41
5–42
5–43
5–44

Ibox Current Mode (ICM) Register . . . . . . . . . . . . . . . . . . . .
Ibox Control and Status Register (ICSR) . . . . . . . . . . . . . . . .
Interrupt Priority Level Register (IPLR) . . . . . . . . . . . . . . . .
Interrupt ID (INTID) Register . . . . . . . . . . . . . . . . . . . . . . .
Asynchronous System Trap Request Register (ASTRR) . . . . .
Asynchronous System Trap Enable Register (ASTER) . . . . . .
Software Interrupt Request Register (SIRR) . . . . . . . . . . . . .
Hardware Interrupt Clear (HWINT_CLR) Register . . . . . . .
Interrupt Summary Register (ISR) . . . . . . . . . . . . . . . . . . . .
Serial Line Transmit (SL_XMIT) Register . . . . . . . . . . . . . .
Serial Line Receive (SL_RCV) Register . . . . . . . . . . . . . . . . .
Performance Counter (PMCTR) Register . . . . . . . . . . . . . . .
Dstream Translation Buffer Address Space Number
(DTB_ASN) Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Dstream Translation Buffer Current Mode (DTB_CM)
Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Dstream Translation Buffer Tag (DTB_TAG) Register . . . . .
Dstream Translation Buffer Page Table Entry (DTB_PTE)
Register—Write Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Dstream Translation Buffer Page Table Entry Temporary
(DTB_PTE_TEMP) Register . . . . . . . . . . . . . . . . . . . . . . . . .
Dstream Memory Management Fault Status (MM_STAT)
Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Faulting Virtual Address (VA) Register . . . . . . . . . . . . . . . . .
Formatted Virtual Address (VA_FORM) Register
(NT_Mode=1) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Formatted Virtual Address (VA_FORM) Register
(NT_Mode=0) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Mbox Virtual Page Table Base Register (MVPTBR) . . . . . . . .
Dcache Parity Error Status (DC_PERR_STAT) Register . . . .
Dstream Translation Buffer Invalidate Single (DTB_IS)
Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Mbox Control Register (MCSR) . . . . . . . . . . . . . . . . . . . . . . .
Dcache Mode (DC_MODE) Register . . . . . . . . . . . . . . . . . . .
Miss Address File Mode (MAF_MODE) Register . . . . . . . . .
Alternate Mode (ALT_MODE) Register . . . . . . . . . . . . . . . . .
Cycle Counter (CC) Register . . . . . . . . . . . . . . . . . . . . . . . . .
Cycle Counter Control (CC_CTL) Register . . . . . . . . . . . . . .

5–19
5–20
5–23
5–24
5–25
5–26
5–27
5–28
5–29
5–31
5–32
5–33
5–38
5–39
5–40
5–42
5–43
5–44
5–46
5–47
5–47
5–49
5–50
5–53
5–54
5–56
5–58
5–60
5–61
5–62

5–45
5–46
5–47
5–48
5–49
5–50
5–51
5–52
5–53
5–54
5–55
5–56
6–1
6–2
6–3
6–4
9–1
9–2
9–3
9–4
9–5
9–6
9–7
9–8
9–9
10–1
10–2
11–1
11–2
11–3
12–1
12–2

xvi

Dcache Test Tag Control (DC_TEST_CTL) Register . . . . . . .
Dcache Test Tag (DC_TEST_TAG) Register . . . . . . . . . . . . .
Dcache Test Tag Temporary (DC_TEST_TAG_TEMP)
Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Scache Control (SC_CTL) Register . . . . . . . . . . . . . . . . . . . .
Scache Status (SC_STAT) Register . . . . . . . . . . . . . . . . . . . .
Scache Address (SC_ADDR) Register . . . . . . . . . . . . . . . . . .
Bcache Control (BC_CONTROL) Register . . . . . . . . . . . . . . .
Bcache Configuration (BC_CONFIG) Register . . . . . . . . . . .
Bcache Tag Address (BC_TAG_ADDR) Register . . . . . . . . . .
External Interface Status (EI_STAT) Register . . . . . . . . . . .
External Interface Address (EI_ADDR) Register . . . . . . . . .
Fill Syndrome (FILL_SYN) Register . . . . . . . . . . . . . . . . . . .
HW_LD Instruction Format . . . . . . . . . . . . . . . . . . . . . . . . . .
HW_ST Instruction Format . . . . . . . . . . . . . . . . . . . . . . . . . .
HW_REI Instruction Format . . . . . . . . . . . . . . . . . . . . . . . . .
HW_MFPR and HW_MTPR Instruction Format . . . . . . . . . .
osc_clk_in_h,l Input Network and Terminations . . . . . . . . . .
Clock Input Differential Impedance . . . . . . . . . . . . . . . . . . . .
Input/Output Pin Timing . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Bcache Timing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
sys_clk System Timing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
ref_clk System Timing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
BiSt Timing Event–Time Line . . . . . . . . . . . . . . . . . . . . . . . .
SROM Load Timing Event–Time Line . . . . . . . . . . . . . . . . . .
Serial ROM Load Timing . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Type 1 Heat Sink . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Type 2 Heat Sink . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Package Dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Alpha 21164 Top View (Pin Down) . . . . . . . . . . . . . . . . . . . .
Alpha 21164 Bottom View (Pin Up) . . . . . . . . . . . . . . . . . . .
IEEE 1149.1 Test Access Port . . . . . . . . . . . . . . . . . . . . . . . .
TAP Controller State Machine . . . . . . . . . . . . . . . . . . . . . . . .

5–63
5–64
5–66
5–69
5–72
5–76
5–78
5–84
5–89
5–92
5–94
5–96
6–9
6–10
6–11
6–12
9–5
9–8
9–9
9–12
9–14
9–17
9–22
9–23
9–24
10–3
10–4
11–2
11–8
11–9
12–4
12–5

Tables
1
2
2–1
2–2
2–3
2–4
2–5
2–6
2–7
2–8
2–9
2–10
3–1
3–2
4–1
4–2
4–3
4–4
4–5
4–6
4–7
4–8
4–9
4–10
4–11
4–12
4–13
4–14
4–15

Register Field Type Notation . . . . . . . . . . . . . . . . . . . . . . . . .
Register Field Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Effect of Branching Instructions on the Branch-Prediction
Stack . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Pipeline Examples—All Cases . . . . . . . . . . . . . . . . . . . . . . . .
Pipeline Examples—Integer Add . . . . . . . . . . . . . . . . . . . . . .
Pipeline Examples—Floating Add . . . . . . . . . . . . . . . . . . . . .
Pipeline Examples—Load (Dcache Hit) . . . . . . . . . . . . . . . . .
Pipeline Examples—Load (Dcache Miss) . . . . . . . . . . . . . . . .
Pipeline Examples—Store (Dcache Hit) . . . . . . . . . . . . . . . . .
Instruction Classes and Slotting . . . . . . . . . . . . . . . . . . . . . .
Instruction Latencies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Floating-Point Control Register Bit Descriptions . . . . . . . . . .
Alpha 21164 Signal Descriptions . . . . . . . . . . . . . . . . . . . . . .
Alpha 21164 Signal Descriptions by Function . . . . . . . . . . . .
CPU Clock Generation Control . . . . . . . . . . . . . . . . . . . . . . .
System Clock Divisor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
System Clock Delay . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Physical Memory Regions . . . . . . . . . . . . . . . . . . . . . . . . . . .
Components for 21164 Write Invalidate Systems . . . . . . . . . .
Bcache States for Cache Coherency Protocols . . . . . . . . . . . .
Components for 21164 Flush Cache Protocol Systems . . . . .
Bcache Options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Alpha 21164-Initiated Interface Commands . . . . . . . . . . . . .
System-Initiated Interface Commands (Write Invalidate
Protocol) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Alpha 21164 Responses on addr_res_h<1:0> to Write
Invalidate Protocol Commands . . . . . . . . . . . . . . . . . . . . . . . .
Alpha 21164 Responses on addr_res_h<2> to 21164
Commands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Alpha 21164 Minimum Response Time to Write Invalidate
Protocol Commands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
System-Initiated Interface Commands (Flush Protocol) . . . . .
Alpha 21164 Responses to Flush-Based Protocol Commands
................................................

xxvi
xxvii
2–7
2–16
2–16
2–16
2–17
2–17
2–18
2–20
2–25
2–39
3–3
3–13
4–5
4–7
4–8
4–13
4–22
4–23
4–26
4–35
4–37
4–55
4–57
4–57
4–58
4–64
4–65

xvii

4–16
4–17
4–18
4–19
5–1
5–2
5–3
5–4
5–5
5–6
5–7
5–8
5–9
5–10
5–11
5–12
5–13
5–14
5–15
5–16
5–17
5–18
5–19
5–20
5–21
5–22
5–23
5–24
5–25
5–26
5–27
5–28
5–29
5–30

xviii

Alpha 21164 Responses on addr_res_h<2> to 21164
Commands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Minimum 21164 Response Time to Flush Protocol Commands
................................................
Data Check Bit Correspondence to CBn . . . . . . . . . . . . . . . . .
Interrupt Priority Level Effect . . . . . . . . . . . . . . . . . . . . . . . .
Ibox, Mbox, Dcache, and PALtemp IPR Encodings . . . . . . . . .
Granularity Hint Bits in ITB_PTE_TEMP Read Format . . .
Icache Parity Error Status Register Fields . . . . . . . . . . . . . .
Exception Summary Register Fields . . . . . . . . . . . . . . . . . . .
Ibox Control and Status Register Fields . . . . . . . . . . . . . . . .
Software Interrupt Request Register Fields . . . . . . . . . . . . . .
Hardware Interrupt Clear Register Fields . . . . . . . . . . . . . . .
Interrupt Summary Register Fields . . . . . . . . . . . . . . . . . . . .
Serial Line Transmit Register Fields . . . . . . . . . . . . . . . . . . .
Serial Line Receive Register Fields . . . . . . . . . . . . . . . . . . . .
Performance Counter Register Fields . . . . . . . . . . . . . . . . . . .
PMCTR Counter Select Options . . . . . . . . . . . . . . . . . . . . . . .
Measurement Mode Control . . . . . . . . . . . . . . . . . . . . . . . . . .
Dstream Memory Management Fault Status Register
Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Formatted Virtual Address Register Fields . . . . . . . . . . . . . .
Dcache Parity Error Status Register Fields . . . . . . . . . . . . . .
Mbox Control Register Fields . . . . . . . . . . . . . . . . . . . . . . . . .
Dcache Mode Register Fields . . . . . . . . . . . . . . . . . . . . . . . . .
Miss Address File Mode Register Fields . . . . . . . . . . . . . . . .
Alternate Mode Register Settings . . . . . . . . . . . . . . . . . . . . .
Cycle Counter Control Register Fields . . . . . . . . . . . . . . . . . .
Dcache Test Tag Control Register Fields . . . . . . . . . . . . . . . .
Dcache Test Tag Register Fields . . . . . . . . . . . . . . . . . . . . . . .
Dcache Test Tag Temporary Register Fields . . . . . . . . . . . . . .
Cbox Internal Processor Register Descriptions . . . . . . . . . . . .
Scache Control Register Fields . . . . . . . . . . . . . . . . . . . . . . . .
Scache Status Register Fields . . . . . . . . . . . . . . . . . . . . . . . .
SC_CMD Field Descriptions . . . . . . . . . . . . . . . . . . . . . . . . . .
Scache Address Register Fields . . . . . . . . . . . . . . . . . . . . . . .
Bcache Control Register Fields . . . . . . . . . . . . . . . . . . . . . . .

4–65
4–66
4–93
4–97
5–2
5–9
5–13
5–15
5–21
5–27
5–28
5–30
5–31
5–32
5–34
5–35
5–37
5–44
5–48
5–51
5–55
5–57
5–59
5–60
5–62
5–63
5–65
5–67
5–68
5–70
5–73
5–74
5–77
5–79

5–31
5–32
5–33
5–34
5–35
5–36
5–37
5–38
6–1
6–2
6–3
6–4
6–5
6–6
6–7
7–1
7–2
9–1
9–2
9–3
9–4
9–5
9–6
9–7
9–8
9–9
9–10
9–11
9–12
9–13
9–14

PM_MUX_SEL Register Fields . . . . . . . . . . . . . . . . . . . . . . .
Bcache Configuration Register Fields . . . . . . . . . . . . . . . . . . .
Bcache Tag Address Register Fields . . . . . . . . . . . . . . . . . . . .
Loading and Locking Rules for External Interface
Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
EI_STAT Register Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Syndromes for Single-Bit Errors . . . . . . . . . . . . . . . . . . . . . .
Cbox IPR PALcode Restrictions . . . . . . . . . . . . . . . . . . . . . . .
PALcode Restrictions Table . . . . . . . . . . . . . . . . . . . . . . . . . .
PALcode Trap Entry Points . . . . . . . . . . . . . . . . . . . . . . . . . .
Required PALcode Function Codes . . . . . . . . . . . . . . . . . . . . .
Opcodes Reserved for PALcode . . . . . . . . . . . . . . . . . . . . . . . .
HW_LD Format Description . . . . . . . . . . . . . . . . . . . . . . . . .
HW_ST Format Description . . . . . . . . . . . . . . . . . . . . . . . . . .
HW_REI Format Description . . . . . . . . . . . . . . . . . . . . . . . . .
HW_MTPR and HW_MFPR Format Description . . . . . . . . . .
Alpha 21164 Signal Pin Reset State . . . . . . . . . . . . . . . . . . .
Internal Processor Register Reset State . . . . . . . . . . . . . . . . .
Alpha 21164 Absolute Maximum Ratings . . . . . . . . . . . . . . .
CMOS dc Input/Output Characteristics . . . . . . . . . . . . . . . . .
Input Clock Specification . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Bcache Loop Timing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Output Driver Characteristics . . . . . . . . . . . . . . . . . . . . . . . .
Alpha 21164 System Clock Output Timing (sysclk=Tø ) . . . . .
Alpha 21164 Reference Clock Input Timing . . . . . . . . . . . . .
ref_clk System Timing Stages . . . . . . . . . . . . . . . . . . . . . . . .
Input Timing for sys_clk_out- or ref_clk_in-Based
Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Output Timing for sys_clk_out- or ref_clk_in-Based
Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Bcache Control Signal Timing . . . . . . . . . . . . . . . . . . . . . . . .
BiSt Timing for Some System Clock Ratios, Port
Mode=Normal (System Cycles) . . . . . . . . . . . . . . . . . . . . . . .
BiSt Timing for Some System Clock Ratios, Port
Mode=Normal (CPU Cycles) . . . . . . . . . . . . . . . . . . . . . . . . .
SROM Load Timing for Some System Clock Ratios (System
Cycles) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

5–83
5–85
5–90
5–92
5–93
5–96
5–100
5–101
6–6
6–7
6–8
6–9
6–10
6–11
6–12
7–3
7–10
9–1
9–3
9–7
9–11
9–11
9–13
9–15
9–17
9–18
9–19
9–21
9–22
9–23
9–24

xix

9–15
9–16
9–17
10–1
10–2
11–1
12–1
12–2
12–3
12–4
A–1
A–2
A–3
A–4
A–5
A–6
A–7
A–8
B–1
D–1

SROM Load Timing for Some System Clock Ratios (CPU
Cycles) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Test Modes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
IEEE 1149.1 Circuit Performance Specifications . . . . . . . . . .
c a at Various Airflows . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Maximum Ta at Various Airflows . . . . . . . . . . . . . . . . . . . . . .
Alphabetic Signal Pin List . . . . . . . . . . . . . . . . . . . . . . . . . . .
Alpha 21164 Test Port Pins . . . . . . . . . . . . . . . . . . . . . . . . . .
Compliance Enable Inputs . . . . . . . . . . . . . . . . . . . . . . . . . . .
Instruction Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Boundary Scan Register Organization . . . . . . . . . . . . . . . . . .
Instruction Format and Opcode Notation . . . . . . . . . . . . . . . .
Architecture Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Opcodes Reserved for Digital . . . . . . . . . . . . . . . . . . . . . . . . .
Opcodes Reserved for PALcode . . . . . . . . . . . . . . . . . . . . . . . .
IEEE Floating-Point Instruction Function Codes . . . . . . . . . .
VAX Floating-Point Instruction Function Codes . . . . . . . . . . .
Opcode Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Required PALcode Function Codes . . . . . . . . . . . . . . . . . . . . .
Alpha 21164 Microprocessor Specifications . . . . . . . . . . . . . .
Document Revision History . . . . . . . . . . . . . . . . . . . . . . . . . .

9–24
9–25
9–26
10–2
10–2
11–3
12–1
12–2
12–6
12–8
A–1
A–2
A–6
A–7
A–7
A–9
A–11
A–12
B–2
D–1

Preface
Audience
This reference manual is for system designers and programmers who use the
Alpha 21164 microprocessor.

Content
This reference manual contains the following chapters and appendixes:
•

Chapter 1 introduces the 21164 and provides an overview of the Alpha
architecture.

•

Chapter 2 describes the major hardware functions and the internal chip
architecture. It describes performance measurement facilities, coding rules,
and design examples.

•

Chapter 3 lists and describes the external hardware interface signals.

•

Chapter 4 describes the external bus functions and transactions, lists bus
commands, and describes the clock functions.

•

Chapter 5 lists and describes the 21164 internal processor register set.

•

Chapter 6 describes the privileged architecture library code (PALcode).

•

Chapter 7 describes the initialization and configuration sequence.

•

Chapter 8 describes error detection and error handling.

•

Chapter 9 provides electrical data and describes signal integrity issues.

•

Chapter 10 provides information about thermal management.

•

Chapter 11 provides mechanical data and packaging information, including
signal pin lists.

•

Chapter 12 describes chip and system testability features.

•

Appendix A summarizes the Alpha instruction set.

xxi

•

Appendix B summarizes the 21164 specifications.

•

Appendix C provides a C code example that calculates the predecode values
of a serial Icache load.

•

Appendix D lists changes and revisions to this manual.

•

Appendix E provides phone numbers for support and lists related Digital
and third-party publications with order information.

•

The Glossary lists and defines terms associated with the 21164.

The companion volume to this manual, the Alpha Architecture Reference
Manual, contains the Alpha architecture information.

Terminology and Conventions
The following sections describe the terminology and conventions used in this
manual.

Numbering
All numbers are decimal unless otherwise indicated. Where there is ambiguity,
numbers other than decimal are indicated with the name of the base following
the number in parentheses, for example FF (hex).

Security Holes
Security holes exist when unprivileged software (that is, software running
outside of kernel mode) can:
•

Affect the operation of another process without authorization from the
operating system.

•

Amplify its privilege without authorization from the operating system.

•

Communicate with another process, either overtly or covertly, without
authorization from the operating system.

UNPREDICTABLE and UNDEFINED
Throughout this manual, the terms UNPREDICTABLE and UNDEFINED are
used. Their meanings are quite different and must be carefully distinguished.
In particular, only privileged software (that is, software running in kernel
mode) can trigger UNDEFINED operations. Unprivileged software cannot
trigger UNDEFINED operations. However, either privileged or unprivileged
software can trigger UNPREDICTABLE results or occurrences.

xxii

UNPREDICTABLE results or occurrences do not disrupt the basic operation
of the processor. The processor continues to execute instructions in its normal
manner. In contrast, UNDEFINED operations can halt the processor or cause
it to lose information.
The terms UNPREDICTABLE and UNDEFINED can be further described as
follows:
UNPREDICTABLE
•

Results or occurrences specified as UNPREDICTABLE may vary from
moment to moment, implementation to implementation, and instruction to
instruction within implementations. Software can never depend on results
specified as UNPREDICTABLE.

•

An UNPREDICTABLE result may acquire an arbitrary value subject to a
few constraints. Such a result may be an arbitrary function of the input
operands or of any state information that is accessible to the process in its
current access mode. UNPREDICTABLE results may be unchanged from
their previous values.
Operations that produce UNPREDICTABLE results may also produce
exceptions.

•

An occurrence specified as UNPREDICTABLE may happen or not based
on an arbitrary choice function. The choice function is subject to the same
constraints as are UNPREDICTABLE results and, in particular, must not
constitute a security hole.
Specifically, UNPREDICTABLE results must not depend upon, or be
a function of the contents of memory locations or registers that are
inaccessible to the current process in the current access mode.
Also, operations that may produce UNPREDICTABLE results must not:
Write or modify the contents of memory locations or registers to which
the current process in the current access mode does not have access.
Halt or hang the system or any of its components.
For example, a security hole would exist if some UNPREDICTABLE result
depended on the value of a register in another process, on the contents
of processor temporary registers left behind by some previously running
process, or on a sequence of actions of different processes.

xxiii

UNDEFINED
•

Operations specified as UNDEFINED may vary from moment to moment,
implementation to implementation, and instruction to instruction within
implementations. The operation may vary in effect from nothing, to
stopping system operation.

•

UNDEFINED operations may halt the processor or cause it to lose
information. However, UNDEFINED operations must not cause the
processor to hang, that is, reach an unhalted state from which there is no
transition to a normal state in which the machine executes instructions.
Only privileged software (that is, software running in kernel mode) may
trigger UNDEFINED operations.

Data Field Size
The term INTnn, where nn is one of 2, 4, 8, 16, 32, or 64, refers to a data field
of nn contiguous NATURALLY ALIGNED bytes. For example, INT4 refers to a
NATURALLY ALIGNED longword.

Ranges and Extents
Ranges are specified by a pair of numbers separated by three periods ( . . . )
and are inclusive. For example, a range of integers 0 . . . 4 includes the
integers 0, 1, 2, 3, and 4.
Extents are specified by a pair of numbers in angle brackets separated by a
colon ( : ) and are inclusive. For example, bits <7:3> specify an extent of bits
including bits 7, 6, 5, 4, and 3.

ALIGNED and UNALIGNED
In this manual the terms ALIGNED and NATURALLY ALIGNED are used
interchangeably to refer to data objects that are powers of two in size. An
ALIGNED datum of size 2**N is stored in memory at a byte address that is a
multiple of 2**N, that is, one that has N low-order zeros. Thus, an ALIGNED
64-byte stack frame has a memory address that is a multiple of 64.
If a datum of size 2**N is stored at a byte address that is not a multiple of
2**N, it is called UNALIGNED.

xxiv

Register Format Notation
This manual contains illustrations that show the format of various registers.
Some registers are followed by a description of each field. The fields on the
register are labeled with either a name or a mnemonic. The description of each
field includes the name or mnemonic, the bit extent, and the type.
The ‘‘Type’’ column in the field description includes both the actual type of the
field, and an optional initialized value, separated from the type by a comma.
The type denotes the functional operation of the field, and may be one of
the values shown in Table 1. If present, the initialized value indicates that
the field is initialized by hardware to the specified value at power-up. If the
initialized value is not present, the field is not initialized at power-up.

xxv

Table 1 Register Field Type Notation

xxvi

Notation

Description

A read-to-clear field. The value is written by hardware and remains
unchanged until read. The value may be read by software at which
point, hardware may write a new value into the field.

A read-only bit or field. The value may be read by software. It is
written by hardware. Software write operations are ignored.

A read/write bit or field. The value may be read and written by
software.

W0C

A write-zero-to-clear bit. If read operations are allowed to the register,
then the value may be read by software. If it is a write-only register,
then a read operation by software returns an UNPREDICTABLE
result. Software write operations of a 0 cause the bit to be cleared by
hardware. Software write operations of a 1 do not modify the state of
the bit.

W1C

A write-one-to-clear bit. If read operations are allowed to the register,
then the value may be read by software. If it is a write-only register,
then a read operation by software returns an UNPREDICTABLE
result. Software write operations of a 1 cause the bit to be cleared by
hardware. Software write operations of a 0 do not modify the state of
the bit.

A write-anything-to-the-register-to-clear bit. If read operations are
allowed to the register, then the value may be read by software. If it
is a write-only register, then a read operation by software returns an
UNPREDICTABLE result. Software write operations of any value to
the register cause the bit to be cleared by hardware.

A write-only bit or field. The value may be written by software
and is used by hardware. Read operations by software return an
UNPREDICTABLE result.

A write bit or field. The value may be written by software and is used
by hardware. Read operations by software return a 0.

In addition to named fields in registers, other bits of the register may be
labeled with one of the five symbols listed in Table 2. These symbols denote
the type of the unnamed fields in the register.
Table 2 Register Field Notation
Notation

Description

IGN

Register bits specified as ignore (IGN) are ignored when written and
are UNPREDICTABLE when read if not otherwise specified.

MBZ

Register bits specified as MBZ (must be zero) must never be filled by
software with a non-zero value. If the processor encounters a non-zero
value in a field specified as MBZ, an UNDEFINED operation may
result.

RAO

RAZ

SBZ

Register bits specified as SBZ (should be zero) should be filled by
software with a zero value. Non-zero values in SBZ fields produce
UNDEFINED results and may produce extraneous instruction-issue
delays.

xxvii

1
Introduction
This chapter provides a brief introduction to the Alpha architecture, Digital’s
RISC (reduced instruction set computing) architecture designed for high
performance. The chapter then summarizes the specific features of the Alpha
21164 (hereafter called the 21164), a microprocessor that implements the
Alpha architecture. Appendix A provides a list of Alpha instructions.
For a complete definition of the Alpha architecture, refer to the companion
volume, the Alpha Architecture Reference Manual.

1.1 The Architecture
The Alpha architecture is a 64-bit load and store RISC architecture designed
with particular emphasis on speed, multiple instruction issue, multiple
processors, and software migration from many operating systems.
All registers are 64 bits in length and all operations are performed between
64-bit registers. All instructions are 32 bits in length. Memory operations
are either load or store operations. All data manipulation is done between
registers.
The Alpha architecture supports the following data types:
•

8-, 16-, 32-, and 64-bit integers

•

IEEE 32-bit and 64-bit floating-point formats

•

VAX architecture 32-bit and 64-bit floating-point formats

In the Alpha architecture, instructions interact with each other only by one
instruction writing to a register or memory location and another instruction
reading from that register or memory location. This use of resources makes
it easy to build implementations that issue multiple instructions every CPU
cycle.

Preliminary—Subject to Change—July 1996 1–1

1.1 The Architecture

The 21164 uses a set of subroutines, called privileged architecture library
code (PALcode), that is specific to a particular Alpha operating system
implementation and hardware platform. These subroutines provide operating
system primitives for context switching, interrupts, exceptions, and memory
management. These subroutines can be invoked by hardware or CALL_PAL
instructions. CALL_PAL instructions use the function field of the instruction
to vector to a specified subroutine. PALcode is written in standard machine
code with some implementation-specific extensions to provide direct access to
low-level hardware functions. PALcode supports optimizations for multiple
operating systems, flexible memory management implementations, and
multi-instruction atomic sequences.
The Alpha architecture performs byte shifting and masking with normal 64-bit,
register-to-register instructions; it does not include single-byte load and store
instructions.

1.1.1 Addressing
The basic addressable unit in the Alpha architecture is the 8-bit byte. The
21164 supports a 43-bit virtual address.
Virtual addresses as seen by the program are translated into physical memory
addresses by the memory management mechanism. The 21164 supports a
40-bit physical address.

1.1.2 Integer Data Types
Alpha architecture supports four integer data types:
Data Type

Description

Byte

A byte is 8 contiguous bits that start at an addressable byte boundary.
A byte is an 8-bit value. A byte is supported in Alpha architecture by
the EXTRACT, MASK, INSERT, and ZAP instructions.

Word

A word is 2 contiguous bytes that start at an arbitrary byte boundary.
A word is a 16-bit value. A word is supported in Alpha architecture by
the EXTRACT, MASK, and INSERT instructions.

Longword

A longword is 4 contiguous bytes that start at an arbitrary byte
boundary. A longword is a 32-bit value. A longword is supported in
the Alpha architecture by sign-extended load and store instructions and
by longword arithmetic instructions.

Quadword

A quadword is 8 contiguous bytes that start at an arbitrary byte
boundary. A quadword is supported in Alpha architecture by load and
store instructions and quadword integer operate instructions.

1–2 Preliminary—Subject to Change—July 1996

1.1 The Architecture

Note
Alpha implementations may impose a significant performance penalty
when accessing operands that are not NATURALLY ALIGNED. Refer
to the Alpha Architecture Reference Manual for details.

1.1.3 Floating-Point Data Types
The 21164 supports the following floating-point data types:
•

Longword integer format in floating-point unit

•

Quadword integer format in floating-point unit

•

IEEE floating-point formats

•

–

S_floating

–

T_floating

VAX floating-point formats
–

F_floating

–

G_floating

–

D_floating (limited support)

1.2 Alpha 21164 Microprocessor Features
The 21164 microprocessor is a superscalar pipelined processor manufactured
using 0.5-micron CMOS technology. It is packaged in a 499-pin IPGA
carrier and has removable application-specific heat sinks. A number of
configuration options allow its use in a range of system designs ranging from
extremely simple uniprocessor systems with minimum component count to
high-performance multiprocessor systems with very high cache and memory
bandwidth.
The 21164 can issue four Alpha instructions in a single cycle, thereby
minimizing the average cycles per instruction (CPI). A number of low-latency
and/or high-throughput features in the instruction issue unit and the onchip
components of the memory subsystem further reduce the average CPI.
The 21164 and associated PALcode implements IEEE single- and doubleprecision, VAX F_floating and G_floating data types, and supports longword
(32-bit) and quadword (64-bit) integers. Byte (8-bit) and word (16-bit) support
is provided by byte-manipulation instructions. Limited hardware support is

Preliminary—Subject to Change—July 1996 1–3

1.2 Alpha 21164 Microprocessor Features

provided for the VAX D_floating data type. Partial hardware implementation is
provided for the architecturally optional FETCH and FETCH_M instructions.
Other 21164 features include:
•

A peak instruction execution rate of four times the CPU clock frequency.

•

The ability to issue up to four instructions during each clock cycle.

•

An onchip, demand-paged memory-management unit with translation
buffer, which, when used with PALcode, can implement a variety of page
table structures and translation algorithms. The unit consists of a 64-entry
data translation buffer (DTB) and a 48-entry instruction translation buffer
(ITB), with each entry able to map a single 8K-byte page or a group of 8,
64, or 512 8K-byte pages. The size of each translation buffer entry’s group
is specified by hint bits stored in the entry. The DTB and ITB implement
7-bit address space numbers (ASN), (MAX_ASN=127).

•

Two onchip, high-throughput pipelined floating-point units, capable of
executing both Digital and IEEE floating-point data types.

•

An onchip, 8K-byte virtual instruction cache with 7-bit ASNs (MAX_
ASN=127).

•

An onchip, dual-read-ported, 8K-byte data cache.

•

An onchip write buffer with six 32-byte entries.

•

An onchip, 96K-byte, 3-way, set-associative, write-back, second-level mixed
instruction and data cache.

•

A 128-bit data bus with onchip parity and error correction code (ECC)
support.

•

Support for an optional external third-level cache. The size and access time
of the external third-level cache is programmable.

•

An internal clock generator providing a high-speed clock used by the 21164,
and a pair of programmable system clocks for use by the CPU module.

•

Onchip performance counters to measure and analyze CPU and system
performance.

•

Chip and module level test support, including an instruction cache test
interface to support chip and module level testing.

•

A 3.3-V power supply. (Direct connection to 5-V logic supported.)

Refer to Chapter 9 for 21164 dc and ac electrical characteristics. Refer to
the Alpha Architecture Reference Manual for a description of address space
numbers (ASNs).

1–4 Preliminary—Subject to Change—July 1996

2
Internal Architecture
This chapter provides both an overview of the 21164 microarchitecture
and a system designer’s view of the 21164 implementation of the Alpha
architecture. The combination of the 21164 microarchitecture and privileged
architecture library code (PALcode) defines the chip’s implementation of the
Alpha architecture. If a certain piece of hardware seems to be ‘‘architecturally
incomplete,’’ the missing functionality is implemented in PALcode. Chapter 6
provides more information on PALcode.
This chapter describes the major functional hardware units and is not intended
to be a detailed hardware description of the chip. It is organized as follows:
•

21164 microarchitecture

•

Pipeline organization

•

Scheduling and issuing rules

•

Replay traps

•

Miss address file (MAF) and load-merging rules

•

Mbox store execution

•

Write buffer and the WMB instruction

•

Performance measurement support

•

Floating-point control register

•

Design examples

Preliminary—Subject to Change—July 1996 2–1

2.1 Alpha 21164 Microarchitecture

2.1 Alpha 21164 Microarchitecture
The 21164 microprocessor is a high-performance implementation of Digital’s
Alpha architecture. Figure 2–1 is a block diagram of the 21164 showing
the major functional blocks relative to pipeline stage flow. Please see the
end of this book for an enlarged foldout version of this figure. The following
paragraphs provide an overview of the chip’s architecture and major functional
units.
The 21164 microprocessor consists of the following internal sections:
•

Clock generation logic (Section 4.2)

•

Instruction fetch/decode unit and branch unit (Ibox) (Section 2.1.1), which
includes:
Instruction prefetcher and instruction decoder
Instruction translation buffer
Branch prediction
Instruction slotting/issue
Interrupt support

•

Integer execution unit (Ebox) (Section 2.1.2)

•

Floating-point execution unit (Fbox) (Section 2.1.3)

•

Memory address translation unit (Mbox) (Section 2.1.4), which includes:
Data translation buffer (DTB)
Miss address file (MAF)
Write buffer
Dcache control

•

Cache control and bus interface unit (Cbox) with interface to external cache
(Section 2.1.5)

•

Data cache (Dcache) (Section 2.1.6.1)

•

Instruction cache (Icache) (Section 2.1.6.2)

•

Second-level cache (Scache) (Section 2.1.6.3)

•

Serial read-only memory (SROM) interface (Section 2.1.7)

2–2 Preliminary—Subject to Change—July 1996

48−Entry
Associative

Instruction
Translation
Buffer

8K Bytes
32−Byte Block
Direct−Mapped

Program
Counter
Logic

Instruction
Buffer

Instruction
Cache

Instruction Fetch/Decode Unit

Next
Index
Logic

Refill
Buffer

Store and
Fill Data

Issue
Scoreboard
Logic

Integer
Register
File

Load Data

Instruction Stream Miss (Physical Address)

Instruction
Slot
Logic

Dual−Read
Translation Buffer

6, 32−Byte
Entries

Write Buffer

6 Data Misses
4 Istream
Misses

Miss
Address
File

Memory Address Translation Unit

Store
Data

64−Entry
Associative
Dual−Ported

Data from
Pins

2 Entries

Backup Cache (Bcache)
1M Byte to
64M Bytes
Direct−Mapped
(Offchip)
Bus Address
File

Instruction
and Data
Fills

MK−1455−13

Cache Control and Bus Interface Unit

Address to Pins

96K Bytes
64−Byte Block
3−Way Set−Associative

Second−Level Cache (Scache)

To Floating−Point Unit

Integer Unit
Store Data

Integer Execution Unit

Floating−Point Multiply Pipe
Floating−Point
Store Data

ADD, LOG, LD, BR,
CMP, CMOV

8K Bytes
32−Byte Block
Direct−Mapped
Dual Read−Ported

Floating−Point Add Pipe and Divider

ADD, LOG, SHIFT, LD,
ST, IMUL, CMP,
CMOV, BYTE, WORD

Floating−Point
Divider

Floating−Point Execution Unit

Data Cache (Dcache)

Integer Pipe 1

Integer Pipe 0

Integer
Multiplier

Floating−
Point
Register
File

Figure 2–1

Istream
Fill

Pipe Stages
S−1

2.1 Alpha 21164 Microarchitecture

Alpha 21164 Microprocessor Block/Pipe Flow Diagram

Preliminary—Subject to Change—July 1996 2–3

2.1 Alpha 21164 Microarchitecture

2.1.1 Instruction Fetch/Decode Unit and Branch Unit
The primary function of the instruction fetch/decode unit and branch unit
(Ibox) is to manage and issue instructions to the Ebox, Mbox, and Fbox. It also
manages the instruction cache. The Ibox contains:
•

Prefetcher and instruction buffer

•

Instruction slot and issue logic

•

Program counter (PC) and branch prediction logic

•

48-entry instruction translation buffers (ITBs)

•

Abort logic

•

Interrupt and exception logic

2.1.1.1 Instruction Decode and Issue
The Ibox decodes up to four instructions in parallel and checks that the
required resources are available for each instruction. The Ibox issues only
the instructions for which all required resources are available. The Ibox does
not issue instructions out of order, even if the resources are available for a
later instruction and not for an earlier one.
In other words:
•

If resources are available, and multiple issue is possible, then all four
instructions are issued.

•

If resources are available only for a later instruction and not for an earlier
one, then only the instructions up to the latest one for which resources are
available are issued.

The Ibox handles only NATURALLY ALIGNED groups of four instructions
(INT16). The Ibox does not advance to a new group of four instructions until
all instructions in a group are issued. If a branch to the middle of an INT16
group occurs, then the Ibox attempts to issue the instructions from the branch
target to the end of the current INT16, then it proceeds to the next INT16 of
instructions after all the instructions in the target INT16 are issued. Thus,
achieving maximum issue rate and optimal performance requires that code be
be scheduled properly and that floating or integer NOP instructions be used to
fill empty slots in the scheduled instruction stream.
For more information on instruction scheduling and issuing, including detailed
rules governing multiple instruction issue, refer to Section 2.3.

2–4 Preliminary—Subject to Change—July 1996

2.1 Alpha 21164 Microarchitecture

2.1.1.2 Instruction Prefetch
The Ibox contains an instruction prefetcher and a 4-entry, 32-byte-per-entry,
prefetch buffer called the refill buffer. Each instruction cache (Icache) miss is
checked in the refill buffer. If the refill buffer contains the instruction data, it
fills the Icache and instruction buffer simultaneously. If the refill buffer does
not contain the necessary data, a fetch and a number of prefetches are sent
to the Mbox. One prefetch is sent per cycle until each of the four entries in
the refill buffer is filled or has a pending fill. If these requests are all Scache
hits, it is possible for instruction data to stream into the Ibox at the rate of
one INT16 (four instructions) per cycle. The Ibox can sustain up to quadinstruction issue from this Scache fill stream, filling the Icache simultaneously.
The refill buffer holds all returned fill data until the data is required by the
Ibox pipeline.
When there is a hit in the refill buffer, the 21164 waits until there is a ‘‘true’’
miss. A ‘‘true’’ miss is one that misses in the Icache and then in the refill
buffer. If an Icache miss results in a refill buffer hit, prefetching is not started
until all the data has been moved from the refill buffer entry into the pipeline.
Each fill of the Icache by the refill buffer occurs when the instruction buffer
stage in the Ibox pipeline requires a new INT16. The INT16 is written into the
Icache and the instruction buffer simultaneously. This can occur at a maximum
rate of one Icache fill per cycle. The actual rate depends on how frequently the
instruction buffer stage requires a new INT16, and on availability of data in
the refill buffer.
Once an Icache miss occurs, the Icache enters fill mode. When the Icache is
in fill mode, the refill buffer is checked each cycle to see if it contains the next
INT16 required by the instruction buffer.
When the required data is not available in the refill buffer (also a miss), the
Icache is checked for a hit while it awaits the arrival of the data from the
Scache or beyond. The Ibox sends a read request to the Cbox by means of the
Mbox. The Cbox checks the Scache and Bcache, and if the request misses in all
caches, the Cbox drives a main memory request.
If there is an Icache hit at this time, the Icache returns to access mode and the
prefetcher stops sending fetches to the Mbox. When a new program counter
(PC) is loaded (that is, taken branches), the Icache returns to access mode until
the first miss. The refill buffer receives and holds instruction data from fetches
initiated before the Icache returned to access mode.
The Icache has a 32-byte block size, whereas the refill buffer is able to load the
Icache with only one INT16 (16 bytes) per cycle. Therefore, each Icache block
has two valid bits, one for each 16-byte subblock.

Preliminary—Subject to Change—July 1996 2–5

2.1 Alpha 21164 Microarchitecture

2.1.1.3 Branch Execution
When a branch or jump instruction is fetched from the Icache by the prefetcher,
the Ibox needs one cycle to calculate the target PC before it is ready to fetch
the target instruction stream. In the second cycle after the fetch, the Icache
is accessed at the target address. Branch and PC prediction are necessary to
predict and begin fetching the target instruction stream before the branch or
jump instruction is issued.
The Icache records the outcome of branch instructions in a 2-bit history state
provided for each instruction location in the Icache. This information is used
as the prediction for the next execution of the branch instruction. The 2-bit
history state is a saturating counter that increments on taken branches and
decrements on not-taken branches. The branch is predicted taken on the top
two count values and is predicted not-taken on the bottom two count values.
The history status is not initialized on Icache fill, therefore it may ‘‘remember’’
a branch that was evicted from the Icache and subsequently reloaded.
The 21164 does not limit the number of branch predictions outstanding to one.
It predicts branches even while waiting to confirm the prediction of previously
predicted branches. There can be one branch prediction pending for each of
pipeline stages 3 and 4, plus up to four in pipeline stage 2. Refer to Section 2.2
for a description of pipeline stages.
When a predicted branch is issued, the Ebox or Fbox checks the prediction.
The branch history table is updated accordingly. On branch mispredict, a
mispredict trap occurs and the Ibox restarts execution from the correct PC.
The 21164 provides a 12-entry subroutine return stack that is controlled by
decoding the opcode (BSR, HW_REI and JMP/JSR/RET/JSR_COROUTINE),
and DISP<15:14> in JMP/JSR/RET/JSR_COROUTINE. The stack stores an
Icache index in each entry. The stack is implemented as a circular queue that
wraps around in the overflow and underflow cases. Table 2–1 lists the effect
each of these instructions has on the state of the branch-prediction stack.

2–6 Preliminary—Subject to Change—July 1996

2.1 Alpha 21164 Microarchitecture

Table 2–1 Effect of Branching Instructions on the Branch-Prediction Stack
Instruction

Stack Used for
Prediction?

Effect on Stack

BSR, JSR

Push PC+4

RET

Yes

Pop

JMP, BR, BRxx

No effect

JSR_COROUTINE

Yes

Pop, then push PC+4

PAL entry

Push PC+4

HW_REI

Yes

Pop

The 21164 uses the Icache index hint in the JMP and JSR instructions to
predict the target PC. The Icache index hint in the instruction’s displacement
field is used to access the direct-mapped Icache. The upper bits of the PC
are formed from the data in the Icache tag store at that index. Later in the
pipeline, the PC prediction is checked against the actual PC generated by the
Ebox. A mismatch causes a PC mispredict trap and restart from the correct
PC. This is similar to branch prediction.
The RET, JSR_COROUTINE, and HW_REI instructions predict the next PC
by using the index from the subroutine return stack. The upper bits of the PC
are formed from the data in the Icache tag at that index. These predictions
are checked against the actual PC in exactly the same way that JMP and JSR
predictions are checked.
Changes from PALmode to native mode and vice versa are predicted on all
PC predictions that use the subroutine return stack. In all cases, if the PC
prediction is correct, the mode prediction will also be correct. Instruction
stream (Istream) prefetching is disabled when a PC prediction is outstanding.
2.1.1.4 Instruction Translation Buffer
The Ibox includes a 48-entry, fully associative instruction translation buffer
(ITB). The buffer stores recently used Istream address translations and
protection information for pages ranging from 8K bytes to 4M bytes and
uses a not-last-used replacement algorithm.
PALcode fills and maintains the ITB. Each entry supports all four granularity
hint bit combinations, so that any single ITB entry can provide translation for
up to 512 contiguously mapped 8K-byte pages. The operating system, using
PALcode, must ensure that virtual addresses can only be mapped through a
single ITB entry or superpage mapping at one time. Multiple simultaneous
mapping can cause UNDEFINED results.

Preliminary—Subject to Change—July 1996 2–7

2.1 Alpha 21164 Microarchitecture

While not executing in PALmode, the 43-bit virtual PC is routed to the ITB
each cycle. If the page table entry (PTE) associated with the PC is cached in
the ITB, the protection bits for the page that contains the PC are used by the
Ibox to do the necessary access checks. If there is an Icache miss and the PC
is cached in the ITB, the page frame number (PFN) and protection bits for the
page that contains the PC are used by the Ibox to do the address translation
and access checks.
The 21164’s ITB supports 128 address space numbers (ASNs) (MAX_ASN=127)
by means of a 7-bit ASN field in each ITB entry. PALcode uses the hardwarespecific HW_MTPR instruction to write to the architecturally defined ITB_IAP
register. This has the effect of invalidating ITB entries that do not have their
ASM bit set.
The 21164 provides two optional translation extensions called superpages.
Access to superpages is enabled using ICSR<SPE> and is allowed only while
executing in privileged mode.
•

One superpage maps virtual address bits <39:13> to physical address bits
<39:13>, on a one-to-one basis, when virtual address bits <42:41> equal
2. This maps the entire physical address space four times over to the
quadrant of the virtual address space.

•

The other superpage maps virtual address bits <29:13> to physical address
bits <29:13>, on a one-to-one basis, and forces physical address bits <39:30>
to 0 when virtual address bits <42:30> equal 1FFE16 . This effectively maps
a 30-bit region of physical address space to a single region of the virtual
address space defined by virtual address bits <42:30> = 1FFE16 .

Access to either superpage mapping is allowed only while executing in kernel
mode. Superpage mapping allows the operating system to map all physical
memory to a privileged virtual memory region.
2.1.1.5 Interrupts
The Ibox exception logic supports three sources of interrupts:
•

Hardware interrupts
There are seven level-sensitive hardware interrupt sources supplied by the
following signals:
irq_h<3:0>
mch_hlt_irq_h
pwr_fail_irq_h
sys_mch_chk_irq_h

•

Software interrupts

2–8 Preliminary—Subject to Change—July 1996

2.1 Alpha 21164 Microarchitecture

There are 15 prioritized software interrupts sourced by the software
interrupt request register (SIRR) (see Section 5.1.22).
•

Asynchronous system traps (ASTs)
There are four ASTs sourced by the asynchronous system trap request
(ASTRR) register.

The serial interrupt, the internally detected correctable error interrupt, the
performance counter interrupts, and irq_h<3:0> are all maskable by bits in
the ICSR (see Section 5.1.17). The four AST traps are maskable by bits in
the ASTER (see Section 5.1.21). In addition, the AST traps are qualified by
the current processor mode. All interrupts are disabled when the processor is
executing PALcode.
Each interrupt source, or group of sources, is assigned an interrupt priority
level (IPL), as shown in Table 4–19. The current IPL is set using the IPLR
register (see Section 5.1.18). Any interrupts that have an equal or lower
IPL are masked. When an interrupt occurs that has an IPL greater than
the value in the IPLR register, program control passes to the INTERRUPT
PALcode entry point. PALcode processes the interrupt by reading the ISR (see
Section 5.1.24) and the INTID register (see Section 5.1.19).

2.1.2 Integer Execution Unit
The integer execution unit (Ebox) contains two 64-bit integer execution
pipelines, E0 and E1, which include the following:
•

Two adders

•

Two logic boxes

•

A barrel shifter

•

Byte-manipulation logic

•

An integer multiplier

The Ebox also includes the 40-entry, 64-bit integer register file (IRF) that
contains the 32 integer registers defined by the Alpha architecture and 8 PAL
shadow registers. The register file has four read ports and two write ports that
provide operands to both integer execution pipelines and accept results from
both pipes. The register file also accepts load instruction results (memory data)
on the same two write ports.

Preliminary—Subject to Change—July 1996 2–9

2.1 Alpha 21164 Microarchitecture

2.1.3 Floating-Point Execution Unit
The onchip, pipelined floating-point unit (FPU) can execute both IEEE and
VAX floating-point instructions. The 21164 supports IEEE S_floating and
T_floating data types, and all rounding modes. It also supports VAX F_floating
and G_floating data types, and provides limited support for the D_floating
format. The FPU contains:
•

A 32-entry, 64-bit floating-point register file.

•

A user-accessible control register.

•

A floating-point multiply pipeline.

•

A floating-point add pipeline—The floating-point divide unit is associated
with the floating-point add pipeline but is not pipelined.

The FPU can accept two instructions every cycle, with the exception of floatingpoint divide instructions. The result latency for nondivide, floating-point
instructions is four cycles.
The floating-point register file (FRF) has five read ports and four write ports.
Four of the read ports are used by the two pipelines to source operands. The
remaining read port is used by floating-point stores. Two of the write ports
are used to write results from the two pipelines. The other two write ports are
used to write fills from floating-point loads.

2.1.4 Memory Address Translation Unit
The memory address translation unit (Mbox) contains three major sections:
•

Data translation buffer (dual ported)

•

Miss address file

•

Write buffer address file

There are a pair of write ports on the floating-point register file devoted to
loads and fills for previous loads that missed. The Mbox arbitrates between
floating-point loads that hit in the Dcache and floating-point fills from the
Cbox, making certain that only one register is written per fill port in each
cycle. Floating-point loads that conflict with Cbox fills for use of these write
ports are forced to miss in the Dcache so that the Cbox fill can execute.
The Mbox receives up to two virtual addresses every cycle from the Ebox. The
translation buffer generates the corresponding physical addresses and access
control information for each virtual address. The 21164 implements a 43-bit
virtual address and a 40-bit physical address.

2–10 Preliminary—Subject to Change—July 1996

2.1 Alpha 21164 Microarchitecture

2.1.4.1 Data Translation Buffer
The 64-entry, fully associative, dual-read-ported data translation buffer (DTB)
stores recently used data stream (Dstream) page table entries (PTEs). Each
entry supports all four granularity hint-bit combinations, so that a single DTB
entry can provide translation for up to 512 contiguously mapped, 8K-byte
pages. The translation buffer uses a not-last-used replacement algorithm.
For load and store instructions, and other Mbox instructions requiring address
translation, the effective 43-bit virtual address is presented to the DTB. If
the PTE of the supplied virtual address is cached in the DTB, the page frame
number (PFN) and protection bits for the page that contains the address are
used by the Mbox to complete the address translation and access checks.
The DTB also supports the optional superpage extensions that are enabled
using ICSR<SPE>. The DTB superpage maps provide virtual-to-physical
address translation for two regions of the virtual address space, as described in
Section 2.1.1.4.
PALcode fills and maintains the DTB. The operating system, using PALcode,
must ensure that virtual addresses be mapped either through a single DTB
entry or through superpage mapping. Multiple simultaneous mapping can
cause UNDEFINED results. The only exception to this rule is that any given
virtual page may be mapped twice with identical data in two different DTB
entries. This occurs in operating systems, such as OpenVMS, which utilize
virtually accessible page tables. If the level 1 page table is accessed virtually,
PALcode loads the translation information twice; once in the double-miss
handler, and once in the primary handler. The PTE mapping the level 1
page table must remain constant during accesses to this page to meet this
requirement.
2.1.4.2 Load Instruction and the Miss Address File
The Mbox begins the execution of each load instruction by translating the
virtual address and by accessing the data cache (Dcache). Translation and
Dcache tag read operations occur in parallel. If the addressed location is found
in the Dcache (a hit), then the data from the Dcache is formatted and written
to either the integer register file (IRF) or floating-point register file (FRF). The
formatting required depends on the particular load instruction executed. If
the data is not found in the Dcache (a miss), then the address, target register
number, and formatting information are entered in the miss address file (MAF).
The MAF performs a load-merging function. When a load miss occurs, each
MAF entry is checked to see if it contains a load miss that addresses the same
Dcache (32-byte) block. If it does, and certain merging rules are satisfied, then
the new load miss is merged with an existing MAF entry. This allows the Mbox
to service two or more load misses with one data fill from the Cbox.

Preliminary—Subject to Change—July 1996 2–11

2.1 Alpha 21164 Microarchitecture

There are six MAF entries for load misses and four more for Ibox instruction
fetches and prefetches. Load misses are usually the highest Mbox priority.
Refer to Section 2.5 for information on load-merging rules.
2.1.4.3 Dcache Control and Store Instructions
The Dcache follows a write-through protocol. During the execution of a store
instruction, the Mbox probes the Dcache to determine whether the location to
be overwritten is currently cached. If so (a Dcache hit), the Dcache is updated.
Regardless of the Dcache state, the Mbox forwards the data to the Cbox.
A load instruction that is issued one cycle after a store instruction in the
pipeline creates a conflict if both the load and store operations access the
same memory location. (The store instruction has not yet updated the location
when the load instruction reads it.) This conflict is handled by forcing the load
instruction to take a replay trap; that is, the Ibox flushes the pipeline and
restarts execution from the load instruction. By the time the load instruction
arrives at the Dcache the second time, the conflicting store instruction has
written the Dcache and the load instruction is executed normally.
Replay traps can be avoided by scheduling the load instruction to issue three
cycles after the store instruction. If the load instruction is scheduled to issue
two cycles after the store instruction, then it will be issue-stalled for one cycle.
2.1.4.4 Write Buffer
The Mbox contains a write buffer that has six 32-byte entries, each of which
holds the data from one or more store instructions that access the same
32-byte block in memory until the data is written into the Scache. The write
buffer provides a finite, high-bandwidth resource for receiving store data to
minimize the number of CPU stall cycles. The write buffer and associated
WMB instruction are described in Section 2.7.

2.1.5 Cache Control and Bus Interface Unit
The cache control and bus interface unit (Cbox) processes all accesses sent by
the Mbox and implements all memory-related external interface functions,
particularly the coherence protocol functions for write-back caching. It
controls the second-level cache (Scache) and the optional board-level backup
cache (Bcache). The Cbox handles all instruction and primary Dcache read
misses, performs the function of writing data from the write buffer into the
shared coherent memory subsystem, and has a major role in executing the
Alpha memory barrier (MB) instruction. The Cbox also controls the 128-bit
bidirectional data bus, address bus, and I/O control. Chapter 4 describes the
external interface.

2–12 Preliminary—Subject to Change—July 1996

2.1 Alpha 21164 Microarchitecture

2.1.6 Cache Organization
The 21164 has three onchip caches—a primary data cache (Dcache), a primary
instruction cache (Icache), and a second-level data and instruction cache
(Scache). All memory cells in the onchip caches are fully static, 6-transistor,
CMOS structures.
The 21164 also provides control for an optional board-level, external cache
(Bcache).
2.1.6.1 Data Cache
The data cache (Dcache) is a dual-read-ported, single-write-ported, 8K-byte
cache. It is a write-through, read-allocate, direct-mapped, physical cache with
32-byte blocks.
2.1.6.2 Instruction Cache
The instruction cache (Icache) is an 8K-byte, virtual, direct-mapped cache with
32-byte blocks. Each block tag contains:
•

A 7-bit address space number (ASN) field as defined by the Alpha
architecture

•

A 1-bit address space match (ASM) field as defined by the Alpha
architecture

•

A 1-bit PALcode (physically addressed) indicator

Software, rather than Icache hardware, maintains Icache coherence with
memory.
2.1.6.3 Second-Level Cache
The second-level cache (Scache) is a 96K-byte, 3-way, set-associative, physical,
write-back, write-allocate cache with 32- or 64-byte blocks. It is a mixed data
and instruction cache. The Scache is fully pipelined; it processes read and
write operations at the rate of one INT16 per CPU cycle and can alternate
between read and write accesses without bubble cycles.
When operating in 32-byte block mode, the Scache has 64-byte blocks with
32-byte subblocks, one tag per block. If configured to 32 bytes, the Scache is
organized as three sets of 512 blocks, with each block divided into two 32-byte
subblocks. If configured to 64 bytes, the Scache is three sets of 512 64-byte
blocks.

Preliminary—Subject to Change—July 1996 2–13

2.1 Alpha 21164 Microarchitecture

2.1.6.4 External Cache
The Cbox implements control for an optional, external, direct-mapped, physical,
write-back, write-allocate cache with 32- or 64-byte blocks. The 21164 supports
board-level cache sizes of 1, 2, 4, 8, 16, 32, and 64 megabytes.

2.1.7 Serial Read-Only Memory Interface
The serial read-only memory (SROM) interface provides the initialization data
load path from a system SROM to the Icache. Chapter 7 provides information
about the SROM interface.

2.2 Pipeline Organization
The 21164 has a 7-stage (or 7-cycle) pipeline for integer operate and memory
reference instructions, and a 9-stage pipeline for floating-point operate
instructions. The Ibox maintains state for all pipeline stages to track
outstanding register write operations.
Figure 2–2 shows the integer operate, memory reference, and floating-point
operate pipelines for the Ibox, FPU, Ebox, and Mbox. The first four stages are
executed in the Ibox. Remaining stages are executed by the Ebox, Fbox, Mbox,
and Cbox. There are bypass paths that allow the result of one instruction to be
used as a source operand of a following instruction before it is written to the
register file.
Tables 2–2, 2–3, 2–4, 2–5, 2–6, and 2–7 provide examples of events at various
stages of pipelining during instruction execution.

2–14 Preliminary—Subject to Change—July 1996

2.2 Pipeline Organization

Figure 2–2 Instruction Pipeline Stages
Instruction Cache Read
Instruction Buffer, Branch Decode,
Determine Next PC
Slot by Function Unit
Register File Access Checks,
Integer Register File Access
Integer
Operate
Pipeline

IC
0

IB
1

SL
2

AC
3

First Integer
Operate Stage
If Needed, Second Integer
Operate Stage
Write Integer Register File
FloatingPoint
Pipeline

IC
0

IB
1

SL
2

Arithmetic, logical, shift and
compare instructions complete in pipeline
stage 4 (1-cycle latency). CMOV completes
in stage 5 (2-cycle latency). IMULL has
an 8- or 9-cycle latency. CMOV or BR
can issue in parallel (0-cycle latency)
with a dependent CMP instruction.

Floating-Point Register
File Access
First Floating-Point
Operate Stage
Write Floating-Point
Register File, Last
Floating-Point
Operate Stage
Memory
Reference
Pipeline

IC
0

IB
1

SL
2

Dcache Read Begins
Dcache Read Ends
Use Dcache Data, Store Writes
Dcache, Scache, Tag Access
Scache Data Access Begins
Scache Data Access Ends
Fill Dcache
Use Scache Data
LJ-03560-TI0A

Preliminary—Subject to Change—July 1996 2–15

2.2 Pipeline Organization

Table 2–2 Pipeline Examples—All Cases
Pipeline Stage

Events

Access Icache tag and data.

Buffer four instructions, check for branches, calculate branch
displacements, and check for Icache hit.

Slot-swap instructions around so they are headed for pipelines
capable of executing them. Stall preceding stages if all instructions
in this stage cannot issue simultaneously because of function unit
conflicts.

Check the operands of each instruction to see that the source is
valid and available and that no write-write hazards exist. Read the
IRF. Stall preceding stages if any instruction cannot be issued. All
source operands must be available at the end of this stage for the
instruction to issue.

Table 2–3 Pipeline Examples—Integer Add
Pipeline Stage

Events

Perform the add operation.

Result is available for use by an operate function in this cycle.

Write the IRF. Result is available for use by an operate function in
this cycle.

Table 2–4 Pipeline Examples—Floating Add
Pipeline Stage

Events

Read the FRF.

First stage of Fbox add pipeline.

Second stage of Fbox add pipeline.

Third stage of Fbox add pipeline.

Fourth stage of Fbox add pipeline. Write the FRF.

Result is available for use by an operate function in this cycle. For
instance, pipeline stage 5 of the user instruction can coincide with
pipeline stage 9 of the producer (latency of 4).

2–16 Preliminary—Subject to Change—July 1996

2.2 Pipeline Organization

Table 2–5 Pipeline Examples—Load (Dcache Hit)
Pipeline Stage1

Events

Calculate the effective address. Begin the Dcache data and tag store
access.

Finish the Dcache data and tag store access. Detect Dcache hit.
Format the data as required. Scache arbitration defaults to pipe E0
in anticipation of a possible miss.

Write the IRF or FRF. Data is available for use by an operate
function in this cycle.

1 Pipe E0 has not been defined at this point.

Table 2–6 Pipeline Examples—Load (Dcache Miss)
Pipeline Stage1

Events

Calculate the effective address. Begin the Dcache data and tag store
access.

Finish the Dcache data and tag store access. Detect Dcache miss.
Scache arbitration defaults to pipe E0 in anticipation of a possible
miss. If there are load instructions in both E0 and E1, the load
instruction in E1 would be delayed at least one more cycle because
default arbitration speculatively assumes the load in E0 will miss.

Begin Scache tag read operation.

Finish Scache tag read operation. Begin detecting Scache hit.

Finish detecting Scache hit. Begin accessing the correct Scache data
bank. (Bcache index at interface—Bcache access begins.)

Finish the Scache data bank access. Begin sending fill data from the
Scache.

Finish sending fill data from the Scache. Begin Dcache fill. Format
the data as required.

Finish the Dcache fill. Write the integer or floating-point register
file.

Data is available for use by an operate function in this cycle.

1 Pipes E0 and E1 have not been defined at this point.

Preliminary—Subject to Change—July 1996 2–17

2.2 Pipeline Organization

Table 2–7 Pipeline Examples—Store (Dcache Hit)
Pipeline Stage

Events

Calculate the effective address. Begin the Dcache tag store access.

Finish the Dcache tag store access. Detect Dcache hit. Send store to
the write buffer simultaneously.

Write the Dcache data store if hit (write begins this cycle).

2.2.1 Pipeline Stages and Instruction Issue
The 21164 pipeline divides instruction processing into four static and a number
of dynamic stages of execution. The first four stages consist of the instruction
fetch, buffer and decode, slotting, and issue-check logic. These stages are static
in that instructions may remain valid in the same pipeline stage for multiple
cycles while waiting for a resource or stalling for other reasons. Dynamic
stages (Ebox and Fbox) always advance state and are unaffected by any stall in
the pipeline. A pipeline stall may occur while zero instructions issue, or while
some instructions of a set of four issue and the others are held at the issue
stage. A pipeline stall implies that a valid instruction is (or instructions are)
presented to be issued but cannot proceed.
Upon satisfying all issue requirements, instructions are issued into their
slotted pipeline. After issuing, instructions cannot stall in a subsequent
pipeline stage. The issue stage is responsible for ensuring that all resource
conflicts are resolved before an instruction is allowed to continue. The only
means of stopping instructions after the issue stage is an abort condition. (The
term abort as used here is different from its use in the Alpha Architecture
Reference Manual.)

2.2.2 Aborts and Exceptions
Aborts result from a number of causes. In general, they can be grouped
into two classes, exceptions (including interrupts) and nonexceptions. The
difference between the two is that exceptions require that the pipeline be
drained of all outstanding instructions before restarting the pipeline at
a redirected address. In either case, the pipeline must be flushed of all
instructions that were fetched subsequent to the instruction that caused the
abort condition (arithmetic exceptions are an exception to this rule). This
includes aborting some instructions of a multiple-issued set in the case of an
abort condition on the one instruction in the set.

2–18 Preliminary—Subject to Change—July 1996

2.2 Pipeline Organization

The nonexception case does not need to drain the pipeline of all outstanding
instructions ahead of the aborting instruction. The pipeline can be restarted
immediately at a redirected address. Examples of nonexception abort
conditions are branch mispredictions, subroutine call/return mispredictions,
and replay traps. Data cache misses can cause aborts or issue stalls depending
on the cycle-by-cycle timing.
In the event of an exception other than an arithmetic exception, the processor
aborts all instructions issued after the exceptional instruction, as described in
the preceding paragraphs. Due to the nature of some exception conditions, this
may occur as late as the integer register file (IRF) write cycle. In the case of
an arithmetic exception, the processor may execute instructions issued after
the exceptional instruction.
After aborting, the address of the exceptional instruction or the immediately
subsequent instruction is latched in the EXC_ADDR internal processor
register (IPR). In the case of an arithmetic exception, EXC_ADDR contains
the address of the instruction immediately after the last instruction executed.
(Every instruction prior to the last instruction executed was also executed.)
For machine check and interrupts, EXC_ADDR points to the instruction
immediately following the last instruction executed. For the remaining cases,
EXC_ADDR points to the exceptional instruction; where, in all cases, its
execution should naturally restart.
When the pipeline is fully drained, the processor begins instruction execution
at the address given by the PALcode dispatch. The pipeline is drained when all
outstanding write operations to both the IRF and FRF have completed and all
outstanding instructions have passed the point in the pipeline such that they
are guaranteed to complete without an exception in the absence of a machine
check.
Replay traps are aborts that occur when an instruction requires a resource
that is not available at some point in the pipeline. These are usually Mbox
resources whose availability could not be anticipated accurately at issue
time (refer to Section 2.4). If the necessary resource is not available when
the instruction requires it, the instruction is aborted and the Ibox begins
fetching at exactly that instruction, thereby replaying the instruction in the
pipeline. A slight variation on this is the load-miss-and-use replay trap in
which an operate instruction is issued just as a Dcache hit is being evaluated
to determine if one of the instruction’s operands is valid. If the result is a
Dcache miss, then the operate instruction is aborted and replayed.

Preliminary—Subject to Change—July 1996 2–19

2.2 Pipeline Organization

2.2.3 Nonissue Conditions
There are two reasons for nonissue conditions. The first is a pipeline stall
wherein a valid instruction or set of instructions are prepared to issue but
cannot due to a resource conflict (register conflict or function unit conflict).
These types of nonissue cycles can be minimized through code scheduling.
The second type of nonissue conditions consists of pipeline bubbles where there
is no valid instruction in the pipeline to issue. Pipeline bubbles result from
the abort conditions described in the previous section. In addition, a single
pipeline bubble is produced whenever a branch type instruction is predicted to
be taken, including subroutine calls and returns.
Pipeline bubbles are reduced directly by the instruction buffer hardware and
through bubble squashing, but can also be effectively minimized through
careful coding practices. Bubble squashing involves the ability of the first four
pipeline stages to advance whenever a bubble or buffer slot is detected in the
pipeline stage immediately ahead of it while the pipeline is otherwise stalled.

2.3 Scheduling and Issuing Rules
The following sections define the classes of instructions and provide rules for
instruction slotting, instruction issuing, and latency.

2.3.1 Instruction Class Definition and Instruction Slotting
The scheduling and multiple issue rules presented here are performance
related only; that is, there are no functional dependencies related to scheduling
or multiple issuing. The rules are defined in terms of instruction classes.
Table 2–8 specifies all of the instruction classes and the pipeline that executes
the particular class. With a few additional rules, the table provides the
information necessary to determine the functional resource conflicts that
determine which instructions can issue in a given cycle.
Table 2–8 Instruction Classes and Slotting
Class Name

Pipeline
1

E0 or E1

Instruction List
2

All loads except LDx_L
All stores except STx_C

1 Ebox pipeline 0.
2 Ebox pipeline 1.

(continued on next page)

2–20 Preliminary—Subject to Change—July 1996

2.3 Scheduling and Issuing Rules

Table 2–8 (Cont.) Instruction Classes and Slotting
Class Name

Pipeline

Instruction List

MBX

LDx_L, MB, WMB, STx_C, HW_LD-lock, HW_ST-cond,
FETCH

RS, RC

MXPR

E0 or E1
(depends on
the IPR)

HW_MFPR, HW_MTPR

IBR

Integer conditional branches

FBR

FA3

Floating-point conditional branches

JSR

Jump-to-subroutine instructions: JMP, JSR, RET, or
JSR_COROUTINE, BSR, BR, HW_REI, CALLPAL

IADD

E0 or E1

ADDL, ADDL/V, ADDQ, ADDQ/V, SUBL, SUBL/V,
SUBQ, SUBQ/V, S4ADDL, S4ADDQ, S8ADDL,
S8ADDQ, S4SUBL, S4SUBQ, S8SUBL, S8SUBQ,
LDA, LDAH

ILOG

E0 or E1

AND, BIS, XOR, BIC, ORNOT, EQV

SHIFT

SLL, SRL, SRA, EXTQL, EXTLL, EXTWL, EXTBL,
EXTQH, EXTLH, EXTWH, MSKQL, MSKLL,
MSKWL, MSKBL, MSKQH, MSKLH, MSKWH,
INSQL, INSLL, INSWL, INSBL, INSQH, INSLH,
INSWH, ZAP, ZAPNOT

CMOV

E0 or E1

CMOVEQ, CMOVNE, CMOVLT, CMOVLE, CMOVGT,
CMOVGE, CMOVLBS, CMOVLBC

ICMP

E0 or E1

CMPEQ, CMPLT, CMPLE, CMPULT, CMPULE,
CMPBGE

IMULL

MULL, MULL/V

IMULQ

MULQ, MULQ/V

IMULH

UMULH

FADD

Floating-point operates, including CPYSN and
CPYSE, except multiply, divide, and CPYS

FDIV

Floating-point divide
4

FMUL

Floating-point multiply

FCPYS

FM or FA

CPYS, not including CPYSN or CPYSE

3 Fbox add pipeline.
4 Fbox multiply pipeline.

(continued on next page)

Preliminary—Subject to Change—July 1996 2–21

2.3 Scheduling and Issuing Rules

Table 2–8 (Cont.) Instruction Classes and Slotting
Class Name

Pipeline

Instruction List

MISC

RPCC, TRAPB

UNOP

None

UNOP5

5 UNOP is LDQ_U R31,0(Rx).

Slotting
The slotting function in the Ibox determines which instructions will be sent
forward to attempt to issue. The slotting function detects and removes all
static functional resource conflicts. The set of instructions output by the
slotting function will issue if no register or other dynamic resource conflict is
detected in stage 3 of the pipeline. The slotting algorithm follows:
Starting from the first (lowest addressed) valid instruction in the INT16 in
stage 2 of the 21164 Ibox pipeline, attempt to assign that instruction to one
of the four pipelines (E0, E1, FA, FM). If it is an instruction that can issue
in either E0 or E1, assign it to E0. However, if one of the following is true,
assign it to E1:
•

E0 is not free and E1 is free.

•

The next integer instruction1 in this INT16 can issue only in E0.

If the current instruction is one that can issue in either FA or FM, assign
it to FA unless FA is not free. If it is an FA-only instruction, it must be
assigned to FA. If it is FM-only instruction, it must be assigned to FM.
Mark the pipeline selected by this process as taken and resume with the
next sequential instruction. Stop when an instruction cannot be allocated
in an execution pipeline because any pipeline it can use is already taken.
The slotting logic does not send instructions forward out of logical instruction
order because the 21164 always issues instructions in order. The slotting logic
also enforces the special rules in the following list, stopping the slotting process
when a rule would be violated by allocating the next instruction an execution
pipeline:
•

An instruction of class LD cannot be issued simultaneously with an
instruction of class ST.

In this context, an integer instruction is one that can issue in one or both of E0 or E1,
but not FA or FM.

2–22 Preliminary—Subject to Change—July 1996

2.3 Scheduling and Issuing Rules

•

All instructions are discarded at the slotting stage after a predicted-taken
IBR or FBR class instruction, or a JSR class instruction.

•

After a predicted not-taken IBR or FBR, no other IBR, FBR, or JSR class
can be slotted together.

•

The following cases are detected by the slotting logic:
From lowest address to highest within an INT16, with the following
arrangement:

I-instruction, F-instruction, I-instruction, I-instruction
I-instruction is any instruction that can issue in one or both of E0 or
E1. F-instruction is any instruction that can issue in one or both of FA
or FM.
From lowest address to highest within an INT16, with the following
arrangement:

F-instruction, I-instruction, I-instruction, I-instruction
When this type of case is detected, the first two instructions are forwarded
to the issue point in one cycle. The second two are sent only when the first
two have both issued, provided no other slotting rule would prevent the
second two from being slotted in the same cycle.

2.3.2 Coding Guidelines
Code should be scheduled according to latency and function unit availability.
This is good practice in most RISC architectures. Code alignment and the
effects of split-issue1 should be considered.
Instructions [a] (the LDL) and [b] (the first ADDL) in the following example are
slotted together. Instruction [b] stalls (split-issue), thus preventing instruction
[c] from advancing to the issue stage:
Code example showing
incorrect ordering

Code example showing
correct ordering

(1) [a] LDL R2,0(R1)
(3) [b] ADDL R2,R3,R4
(4) [c] ADDL R2,R5,R6

(1) [d] LDL R2,0(R1)
(1) [e] NOP
(3) [f] ADDL R2,R3,R4
(3) [g] ADDL R2,R5,R6

NOTES: The instruction examples are assumed to begin on an
INT16 alignment. (n) = Expected execute cycle.

Split-issue is the situation in which not all instructions sent from the slotting stage
to the issue stage issue. One or more stalls result.

Preliminary—Subject to Change—July 1996 2–23

2.3 Scheduling and Issuing Rules

Eventually [b] issues when the result of [a] is returned from a presumed
Dcache hit. Instruction [c] is delayed because it cannot advance to the issue
stage until [b] issues.
In the improved sequence, the LDL [d] is slotted with the NOP [e]. Then the
first ADDL [f] is slotted with the second ADDL [g] and those two instructions
dual-issue. This sequence takes one less cycle to complete than the first
sequence.

2.3.3 Instruction Latencies
After slotting, instruction issue is governed by the availability of registers for
read or write operations, and the availability of the floating divide unit and the
integer multiply unit. There are producer–consumer dependencies, producer–
producer dependencies (also known as write-after-write conflicts), and dynamic
function unit availability dependencies (integer multiply and floating divide).
The Ibox logic in stage 3 of the 21164 pipeline detects all these conflicts.
The latency to produce a valid result for most instructions is fixed. The
exceptions are loads that miss, floating-point divides, and integer multiplies.
Table 2–9 gives the latencies for each instruction class. A latency of 1 means
that the result may be used by an instruction issued one cycle after the
producing instruction. Most latencies are only a property of the producer. An
exception is integer multiply latencies. There are no variations in latency due
to which a particular unit produces a given result relative to the particular
unit that consumes it. In the case of integer multiply, the instruction is issued
at the time determined by the standard latency numbers. The multiply’s
latency is dependent on which previous instructions produced its operands and
when they executed.

2–24 Preliminary—Subject to Change—July 1996

2.3 Scheduling and Issuing Rules

Table 2–9 Instruction Latencies
Additional Time Before
Result Available to Integer
Multiply Unit1

Class

Latency

Dcache hits, latency=2.
Dcache miss/Scache hit, latency=8 or longer.2

1 cycle

Store operations produce no result.

—

MBX

LDx_L always Dcache misses, latency depends on
memory subsystem state.
STx_C, latency depends on memory subsystem
state.
MB, WMB, and FETCH produce no result.

—

RS, RC, latency=1.

2 cycles

MXPR

HW_MFPR, latency=1, 2, or longer, depending on
the IPR.
HW_MTPR, produces no result.

1 or 2 cycles

IBR

Produces no result. (Taken branch issue latency
minimum = 1 cycle, branch mispredict penalty =
5 cycles.)

—

FBR

Produces no result. (Taken branch issue latency
minimum = 1 cycle, branch mispredict penalty =
5 cycles.)

—

JSR

All but HW_REI, latency=1.
HW_REI produces no result.
(Issue latency—minimum 1 cycle.)

2 cycles

IADD

Latency=1.

2 cycles

ILOG

2 cycles

Latency=1.

1 The multiplier is unable to receive data from Ebox bypass paths. The instruction issues at the expected time,

but its latency is increased by the time it takes for the input data to become available to the multiplier. For
example, an IMULL instruction issued one cycle later than an ADDL instruction, which produced one of its
operands, has a latency of 10 (8 + 2). If the IMULL instruction is issued two cycles later than the ADDL
instruction, the latency is 9 (8 + 1).
2 When idle, Scache arbitration predicts a load miss in E0.

If a load actually does miss in E0, it is sent to
the Scache immediately. If it hits, and no other event in the Cbox affects the operation, the requested data is
available for use in eight cycles. Otherwise, the request takes longer (possibly much longer, depending on the
state of the Scache and Cbox). It should be possible to schedule some unrolled code loops for Scache by using
a data access pattern that takes advantage of the Mbox load-merging function, achieving high throughput
with large data sets.

4 A special bypass provides an effective latency of 0 (zero) cycles for an ICMP or ILOG instruction producing

the test operand of an IBR or CMOV instruction. This is true only when the IBR or CMOV instruction issues
in the same cycle as the ICMP or ILOG instruction that produced the test operand of the IBR or CMOV
instruction. In all other cases, the effective latency of ICMP and ILOG instruction is one cycle.

(continued on next page)

Preliminary—Subject to Change—July 1996 2–25

2.3 Scheduling and Issuing Rules

Table 2–9 (Cont.) Instruction Latencies

Class

Latency

Additional Time Before
Result Available to Integer
Multiply Unit1

SHIFT

Latency=1.

2 cycles

CMOV

Latency=2.

1 cycle

ICMP

Latency=1.

2 cycles

IMULL

Latency=8, plus up to 2 cycles of added latency,
depending on the source of the data.1
Latency until next IMULL, IMULQ, or IMULH
instruction can issue (if there are no data
dependencies) is 4 cycles plus the number of
cycles added to the latency.

1 cycle

IMULQ

Latency=12, plus up to 2 cycles of added latency,
depending on the source of the data.1
Latency until next IMULL, IMULQ, or IMULH
instruction can issue (if there are no data
dependencies) is 8 cycles plus the number of
cycles added to the latency.

1 cycle

IMULH

Latency=14, plus up to 2 cycles of added latency,
depending on the source of the data.1
Latency until next IMULL, IMULQ, or IMULH
instruction can issue (if there are no data
dependencies) is 8 cycles plus the number of
cycles added to the latency.

1 cycle

FADD

Latency=4.

—

FDIV

Data-dependent latency: 15 to 31 single
precision, 22 to 60 double precision. Next floating
divide can be issued in the same cycle. The result
of the previous divide is available, regardless of
data dependencies.

—

FMUL

Latency=4.

—

FCPYS

Latency=4.

—

1 The multiplier is unable to receive data from Ebox bypass paths. The instruction issues at the expected time,

but its latency is increased by the time it takes for the input data to become available to the multiplier. For
example, an IMULL instruction issued one cycle later than an ADDL instruction, which produced one of its
operands, has a latency of 10 (8 + 2). If the IMULL instruction is issued two cycles later than the ADDL
instruction, the latency is 9 (8 + 1).
4 A special bypass provides an effective latency of 0 (zero) cycles for an ICMP or ILOG instruction producing

(continued on next page)

2–26 Preliminary—Subject to Change—July 1996

2.3 Scheduling and Issuing Rules

Table 2–9 (Cont.) Instruction Latencies

Class

Latency

Additional Time Before
Result Available to Integer
Multiply Unit1

MISC

RPCC, latency=2. TRAPB produces no result.

1 cycle

UNOP

UNOP produces no result.

—

1 The multiplier is unable to receive data from Ebox bypass paths. The instruction issues at the expected time,

2.3.3.1 Producer–Producer Latency
Producer–producer latency, also known as write-after-write conflicts, cause
issue-stalls to preserve write order. If two instructions write the same
register, they are forced to do so in different cycles by the Ibox. This is
necessary to ensure that the correct result is left in the register file after both
instructions have executed. For most instructions, the order in which they
write the register file is dictated by issue order. However IMUL, FDIV, and
LD instructions may require more time than other instructions to complete.
Subsequent instructions that write the same destination register are issuestalled to preserve write ordering at the register file.
Conditions that involve an intervening producer–consumer conflict can occur
commonly in a multiple-issue situation when a register is reused. In these
cases, producer–consumer latencies are equal to or greater than the required
producer–producer latency as determined by write ordering and therefore
dictate the overall latency.
An example of this case is shown in the following code:
LDQ R2,0(R0)
ADDQ R2,R3,R4
LDQ R2,D(R1)

; R2 destination
; wr-rd conflict stalls execution waiting for R2
; wr-wr conflict may dual issue when ADDQ issues

Producer–producer latency is generally determined by applying the rule that
register file write operations must occur in the correct order (enforced by
Ibox hardware). Two IADD or ILOG class instructions that write the same
register issue at least one cycle apart. The same is true of a pair of CMOVclass instructions, even though their latency is 2. For IMUL, FDIV, and LD
instructions, producer–producer conflicts with any subsequent instruction
results in the second instruction being issue-stalled until the IMUL, FDIV, or
LD instruction is about to complete. The second instruction is issued as soon

Preliminary—Subject to Change—July 1996 2–27

2.3 Scheduling and Issuing Rules

as it is guaranteed to write the register file at least one cycle after the IMUL,
FDIV, or LD instruction.
If a load writes a register, and within two cycles a subsequent instruction
writes the same register, the subsequent instruction is issued speculatively,
assuming the load hits. If the load misses, a load-miss-and-use trap is
generated. This causes the second instruction to be replayed by the Ibox.
When the second instruction again reaches the issue point, it is issue-stalled
until the load fill occurs.

2.3.4 Issue Rules
The following is a list of conditions that prevent the 21164 from issuing an
instruction:
•

No instruction can be issued until all of its source and destination registers
are clean; that is, all outstanding write operations to the destination
register are guaranteed to complete in issue order and there are no
outstanding write operations to the source registers, or those write
operations can be bypassed.
Technically, load-miss-and-use replay traps are an exception to this rule.
The consumer of the load’s result issues, and is aborted, because a load
was predicted to hit and was discovered to miss just as the consumer
instruction issued. In practice, the only difference is that the latency of
the consumer may be longer than it would have been had the issue logic
‘‘known’’ the load would miss in time to prevent issue.

•

An instruction of class LD cannot be issued in the second cycle after an
instruction of class ST is issued.

•

No LD, ST, MXPR (to an Mbox register), or MBX class instructions can be
issued after an MB instruction has been issued until the MB instruction
has been acknowledged by the Cbox.

•

No LD, ST, MXPR (to an Mbox register), or MBX class instructions can be
issued after a STx_C (or HW_ST-cond) instruction has been issued until
the Mbox writes the success/failure result of the STx_C (HW_ST-cond) in
its destination register.

•

No IMUL instructions can be issued if the integer multiplier is busy.

•

No floating-point divide instructions can be issued if the floating-point
divider is busy.

•

No instruction can be issued to pipe E0 exactly two cycles before an integer
multiplication completes.

2–28 Preliminary—Subject to Change—July 1996

2.3 Scheduling and Issuing Rules

•

No instruction can be issued to pipe FA exactly five cycles before a floatingpoint divide completes.

•

No instruction can be issued to pipe E0 or E1 exactly two cycles before an
integer register fill is requested (speculatively) by the Cbox, except IMULL,
IMULQ, and IMULH instructions and instructions that do not produce any
result.

•

No LD, ST, or MBX class instructions can be issued to pipe E0 or E1
exactly one cycle before a integer register fill is requested (speculatively) by
the Cbox.

•

No instruction issues after a TRAPB instruction until all previously issued
instructions are guaranteed to finish without generating a trap other than
a machine check.

All instructions sent to the issue stage (stage 3) by the slotting logic (stage 2)
are issued subject to the previous rules. If issue is prevented for a given
instruction at the issue stage, all logically subsequent instructions at that stage
are prevented from issuing automatically. The 21164 only issues instructions
in order.

2.4 Replay Traps
There are no stalls after the instruction issue point in the pipeline. In some
situations, an Mbox instruction cannot be executed because of insufficient
resources (or some other reason). These instructions trap and the Ibox restarts
their execution from the beginning of the pipeline. This is called a replay trap.
Replay traps occur in the following cases:
•

The write buffer is full when a store instruction is executed and there are
already six write buffer entries allocated. The trap occurs even if the entry
would have merged in the write buffer.

•

A load instruction is issued in pipe E0 when all six MAF entries are valid
(not available), or a load instruction issued in pipe E1 when five of the six
MAF entries are valid. The trap occurs even if the load instruction would
have hit in the Dcache or merged with an MAF entry.

•

Alpha shared memory model order trap (Litmus test 1 trap): If a load
instruction issues that address matches with any miss in the MAF, the
load instruction is aborted through a replay trap regardless of whether the
newly issued load instruction hits or misses in the Dcache. The address
match is precise except that it includes the case in which a longword
access matches within a quadword access. This ensures that the two loads
execute in issue order.

Preliminary—Subject to Change—July 1996 2–29

2.4 Replay Traps

•

Load-after-store trap: A replay trap occurs if a load instruction is issued in
the cycle immediately following a store instruction that hits in the Dcache,
and both access the same location. The address match is exact for address
bits <12:2> (longword granularity), but ignores address bits <42:13>.

•

When a load instruction is followed, within one cycle, by any instruction
that uses the result of that load, and the load misses in the Dcache,
the consumer instruction traps and is restarted from the beginning of
the pipeline. This occurs because the consumer instruction is issued
speculatively while the Dcache hit is being evaluated. If the load misses
in the Dcache, the speculative issue of the consumer instruction was
incorrect. The replay trap generally brings the consumer instruction to the
issue point before or simultaneously with the availability of fill data.

2.5 Miss Address File and Load-Merging Rules
The following sections describe the miss address file (MAF) and its loadmerging function, and the load-merging rules that apply after a load miss.

2.5.1 Merging Rules
When a load miss occurs, each MAF entry is checked to see if it contains
a load miss that addresses the same 32-byte Dcache block. If it does, and
certain merging rules 2 are satisfied, then the new load miss is merged with an
existing MAF entry. This allows the Mbox to service two or more load misses
with one data fill from the Cbox. The merging rules for an individual MAF
entry are as follows:
•

Merging only occurs if the new load miss addresses a different INT8 from
all loads previously entered or merged to that MAF entry.

•

Merging only occurs if the new load miss is the same access size as the
load instructions previously entered in that MAF entry. That is, quadword
load instructions merge only with other quadword load instructions
and longword load instructions merge only with other longword load
instructions.

•

In the case of longword load instructions, both <02> address bits must
be the same. That is, longword load instructions with even addresses
merge only with other even longword load instructions, and longword load
instructions with odd addresses merge only with other odd longword load
instructions.

Merging rules result primarily from limitations of the implementation.

2–30 Preliminary—Subject to Change—July 1996

2.5 Miss Address File and Load-Merging Rules

•

The MAF does not merge floating-point and integer load misses in the
same entry.

•

Merging is prevented for the MAF entry a certain number of cycles after
the Scache access corresponding to the MAF entry begins. Merging is
prevented for that entry only if the Scache access hits. The minimum
number of cycles of merging is three; the cycle in which the first load
is issued, and the two subsequent cycles. This corresponds to the most
optimistic case of a load miss being forwarded to the Scache without delay
(accounting for the cycle saved by the bypass that sends new load misses
directly to the Scache when there is nothing else pending).

2.5.2 Read Requests to the Cbox
When merging does not occur, a new MAF entry is allocated for the new load
miss. Merging is done for two load instructions issued simultaneously, which
both miss in effect as if they were issued sequentially with the load from Ebox
pipe E0 first. The Mbox sends a read request to the Cbox for each MAF entry
allocated.
A bypass is provided so that if the load instruction issues in Ebox pipe E0, and
no MAF requests are pending, the load instruction’s read request is sent to the
Cbox immediately. Similarly, if a load instruction from Ebox pipe E1 misses,
and there was no load instruction in pipe E0 to begin with, the E1 load miss
is sent to the Cbox immediately. In either case, the bypassed read request is
aborted if the load hits in the Dcache or merges in the MAF.

2.5.3 Load Instructions to Noncacheable Space
Merging is normally allowed for load instructions to noncacheable space
(physical address bit <39> = 1). It is prevented when MAF_MODE<03>=1
(see Section 5.2.16). At the external interface, these read instructions
tell the system environment which INT32 is addressed and which of the
INT8s within the INT32 are actually accessed. Merging stops for a load
instruction to noncacheable space as soon as the Cbox accepts the reference.
This permits the system environment to access only those INT8s that are
actually requested by load instructions. For memory-mapped INT4 registers,
the system environment must return the result of reading each register within
the INT8. This occurs because the 21164 only indicates those INT8s that
are accessed, not the exact length and offset of the access within each INT8.
Systems implementing memory-mapped registers with side effects from read
instructions should place each such register in a separate INT8 in memory.

Preliminary—Subject to Change—July 1996 2–31

2.5 Miss Address File and Load-Merging Rules

2.5.4 MAF Entries and MAF Full Conditions
There are six MAF entries for load misses and four for Ibox instruction fetches
and prefetches. Load misses are usually the highest Mbox priority request.
If the MAF is full and a load instruction issues in pipe E0, or if five of the six
MAF entries are valid and a load instruction issues in pipe E1, an MAF full
trap occurs causing the Ibox to restart execution with the load instruction that
caused the MAF overflow. When the load instruction arrives at the MAF the
second time, an MAF entry may have become available. If not, the MAF full
trap occurs again.

2.5.5 Fill Operation
Eventually, the Cbox provides the data requested for a given MAF entry (a fill).
If the fill is integer data and not floating-point data, the Cbox requests that the
Ibox allocate two consecutive ‘‘bubble’’ cycles in the Ebox pipelines. The first
bubble prevents any instruction from issuing. The second bubble prevents only
Mbox instructions (particularly load and store instructions) from issuing. The
fill uses the first bubble cycle as it progresses down the Ebox/Mbox pipelines to
format the data and load the register file. It uses the second bubble cycle to fill
the Dcache.
An instruction typically writes the register file in pipeline stage 6 (see
Figure 2–2). Because there is only one register file write port per integer
pipeline, a no-instruction bubble cycle is required to reserve a register file
write port for the fill. A load or store instruction accesses the Dcache in the
second half of stage 4 and the first half of stage 5. The fill operation writes the
Dcache, making it unavailable for other accesses at that time. Relative to the
register file write operation, the Dcache (write) access for a fill occurs a cycle
later than the Dcache access for a load hit. Only load and store instructions
use the Dcache in the pipeline. Therefore, the second bubble reserved for a fill
is a no-Mbox-instruction bubble.
The second bubble is a subset of the first bubble. When two fills are in
consecutive cycles, as in an Scache hit, then three total bubbles are allocated:
two no-instruction bubbles, followed by one no-Mbox-instruction bubble. The
bubbles are requested speculatively before it is known whether the Scache or
the optional external Bcache will hit.
For fills from the Cbox to floating-point registers, no cycle is allocated. Load
instructions that conflict with the fill in the pipeline are forced to miss. Store
instructions that conflict in the pipeline force the fill to be aborted in order
to keep the Dcache available to the store operation. In all cases, the floatingpoint registers are filled as dictated by the associated MAF entry. The Fbox
has separate write ports for fill data as is necessary for this fill scheme.

2–32 Preliminary—Subject to Change—July 1996

2.5 Miss Address File and Load-Merging Rules

Up to two floating or integer registers may be written for each Cbox fill
cycle. Fills deliver 32 bytes in two cycles: two INT8s per cycle. The MAF
merging rules ensure that there is no more than one register to write for each
INT8, so that there is a register file write port available for each INT8. After
appropriate formatting, data from each INT8 is written into the IRF or FRF
provided there is a miss recorded for that INT8.
Load misses are all checked against the write buffer contents for conflicts
between new load instructions and previously issued store instructions. Refer
to Section 2.7 for more information on write operations.
LDL_L and LDQ_L instructions always allocate a new MAF entry. No load
instructions that follow an LDL_L or LDQ_L instruction are allowed to merge
with it. After an LDL_L or LDQ_L instruction is issued, the Ibox does not
issue any more Mbox instructions until the Mbox has successfully sent the
LDL_L or LDQ_L instruction to the Cbox. This guarantees correct ordering
between an LDL_L or LDQ_L instruction and a subsequent STL_C or STQ_C
instruction even if they access different addresses.

2.6 Mbox Store Instruction Execution
Store instructions execute in the Mbox by:
1. Reading the Dcache tag store instruction in the pipeline stage in which a
load instruction would read the Dcache
2. Checking for a hit in the next stage
3. Writing the Dcache data store instruction if there is a hit in the second
(following) pipeline stage.
Load instructions are not allowed to issue in the second cycle after a store
instruction (one bubble cycle). Other instructions can be issued in that
cycle. Store instructions can issue at the rate of one per cycle because store
instructions in the Dstream do not conflict in their use of resources. The
Dcache tag store and Dcache data store are the principal resources. However,
a load instruction uses the Dcache data store in the same early stage that
it uses the Dcache tag store. Therefore, a load instruction would conflict
with a store instruction if it were issued in the second cycle after any store
instruction. Refer to Section 2.2 for more information on store instruction
execution in the pipeline.
A load instruction that is issued one cycle after a store instruction in the
pipeline creates a conflict if both access exactly the same memory location.
This occurs because the store instruction has not yet updated the location
when the load instruction reads it. This conflict is handled by forcing the

Preliminary—Subject to Change—July 1996 2–33

2.6 Mbox Store Instruction Execution

load instruction to replay trap. The Ibox flushes the pipeline and restarts
execution from the load instruction. By the time the load instruction arrives
at the Dcache the second time, the conflicting store instruction has written the
Dcache and the load instruction is executed normally.
Software should not load data immediately after storing it. The replay trap
that is incurred ‘‘costs’’ seven cycles. The best solution is to schedule the load
instruction to issue three cycles after the store. No issue stalls or replay traps
will occur in that case. If the load instruction is scheduled to issue two cycles
after the store instruction, it will be issue-stalled for one cycle. This is not
an optimal solution, but is preferred over incurring a replay trap on the load
instruction.
For three cycles during store instruction execution, fills from the Cbox are not
placed in the Dcache. Register fills are unaffected. There are conflicts that
make it impossible to fill the Dcache in each of these cycles. Fills are prevented
in cycles in which a store instruction is in pipeline stage 4, 5, or 6. This always
applies to fills of floating-point data. Fills of integer data allocate bubble cycles,
such that an integer fill never conflicts with a store instruction in pipeline
stages 4 or 5. Instead, a store instruction that would have conflicted in stage 4
or 5 is issue-stalled but an integer fill will conflict with a store instruction in
pipeline stage 6.
If a store instruction is stalled at the issue point for any reason, it interferes
with fills just as if it had been issued. This applies only to fills of floating-point
data.
For each store instruction, a search of the MAF is done to detect load-beforestore hazards. If a store instruction is executed, and a load of the same address
is present in the MAF, two things happen:
1. Bits are set in each conflicting MAF entry to prevent its fill from being
placed in the Dcache when it arrives, and to prevent subsequent load
instructions from merging with that MAF entry.
2. Conflict bits are set with the store instruction in the write buffer to prevent
the store instruction from being issued until all conflicting load instructions
have been issued to the Cbox.
Conflict checking is done at the 32-byte block granularity. This ensures proper
results from the load instructions and prevents incorrect data from being
cached in the Dcache.
A check is performed for each new store against store instructions in the write
buffer that have already been sent to the Cbox but have not been completed.
Section 2.7 describes this process.

2–34 Preliminary—Subject to Change—July 1996

2.7 Write Buffer and the WMB Instruction

2.7 Write Buffer and the WMB Instruction
The following sections describe the write buffer and the WMB instruction.

2.7.1 The Write Buffer
The write buffer contains six fully associative 32-byte entries. The purpose of
the write buffer is to minimize the number of CPU stall cycles by providing
a finite, high-bandwidth resource for receiving store data. This is required
because the 21164 can generate store data at the peak rate of one INT8 every
CPU cycle. This is greater than the average rate at which the Scache can
accept the data if Scache misses occur.
In addition to HW_ST and other store instructions, the STQ_C, STL_C,
FETCH, and FETCH_M instructions are also written into the write buffer and
sent offchip. However, unlike store instructions, these write buffer-directed
instructions are never merged into a write buffer entry with other instructions.
A write buffer entry is invalid if it does not contain one of these instructions.

2.7.2 The Write Memory Barrier (WMB) Instruction
The memory barrier (MB) instruction is suitable for ordering memory
references of any kind. The WMB instruction forces ordering of write
operations only (store instructions). The WMB instruction has a special
effect on the write buffer. When it is executed, a bit is set in every write buffer
entry containing valid store data that will prevent future store instructions
from merging with any of the entries. Also, the next entry to be allocated is
marked with a WMB flag. At this point, the entry marked with the WMB flag
does not yet have valid data in it. When an entry marked with a WMB flag is
ready to issue to the Cbox, the entry is not issued until every previously issued
write instruction is complete. This ensures correct ordering between store
instructions issued before the WMB instruction and store instructions issued
after it.
Each write buffer entry contains a content-addressable memory (CAM) for
holding physical address bits <39:05>, 32 bytes of data, eight INT4 mask bits
(that indicate which of the eight INT4s in the entry contain valid data), and
miscellaneous control bits. Among the control bits are the WMB flag, and a
no-merge bit, which indicates that the entry is closed to further merging.

Preliminary—Subject to Change—July 1996 2–35

2.7 Write Buffer and the WMB Instruction

2.7.3 Entry-Pointer Queues
Two entry-pointer queues are associated with the write buffer: a free-entry
queue and a pending-request queue. The free-entry queue contains pointers
to available invalid write buffer entries. The pending-request queue contains
pointers to valid write buffer entries that have not yet been issued to the Cbox.
The pending-request queue is ordered in allocation order.
Each time the write buffer is presented with a store instruction, the physical
address generated by the instruction is compared to the address in each valid
write buffer entry that is open for merging. If the address is in the same
INT32 as an address in a valid write buffer entry (that also contains a store
instruction), and the entry is open for merging, then the new store data is
merged into that entry and the entry’s INT4 mask bits are updated. If no
matching address is found, or all entries are closed to merging, then the store
data is written into the entry at the top of the free-entry queue. This entry is
validated, and a pointer to the entry is moved from the free-entry queue to the
pending-request queue.

2.7.4 Write Buffer Entry Processing
When two or more entries are in the pending-request queue, the Mbox requests
that the Cbox process the write buffer entry at the head of the pending-request
queue. Then the Mbox removes the entry from the pending-request queue
without placing it in the free-entry queue. When the Cbox has completely
processed the write buffer entry, it notifies the Mbox, and the now invalid write
buffer entry is placed in the free-entry queue. The Mbox may request that
a second write buffer entry be processed while waiting for the Cbox to finish
the first. The write buffer entries are invalidated and placed in the free-entry
queue in the order that the requests complete. This order may be different
from the order in which the requests were made.
The Mbox sends write requests from the write buffer to the Cbox. The Cbox
processes these requests according to the cache coherence protocol. Typically,
this involves loading the target block into the Scache, making it writable, and
then writing it. Because the Scache is write-back, this completes the operation.
The Mbox requests that a write buffer entry be processed every 64 cycles, even
if there is only one valid entry. This ensures that write instructions do not wait
forever to be written to memory. (This is triggered by a free running timer.)
When an LDL_L or LDQ_L instruction is processed by the Mbox, the Mbox
requests processing of the next pending write buffer request. This increases
the chances of the write buffer being empty when an STL_C or STQ_C
instruction is issued.

2–36 Preliminary—Subject to Change—July 1996

2.7 Write Buffer and the WMB Instruction

The Mbox continues to request that write buffer entries be processed as long
as one of the following occurs:
•

One buffer contains an STQ_C, STL_C, FETCH, or FETCH_M instruction

•

One buffer is marked by a WMB flag

•

An MB instruction is being executed by the Mbox.

This ensures that these instructions complete as quickly as possible.
Every store instruction that does not merge in the write buffer is checked
against every valid entry. If any entry is an address match, then the WMB flag
is set on the newly allocated write buffer entry. This prevents the Mbox from
concurrently sending two write instructions to exactly the same block in the
Cbox.
Load misses are checked in the write buffer for conflicts. The granularity of
this check is an INT32. Any load instruction matching any write buffer entry’s
address is considered a hit even if it does not access an INT4 marked for
update in that write buffer entry. If a load hits in the write buffer, a conflict bit
is set in the load instruction’s MAF entry, which prevents the load instruction
from being issued to the Cbox before the conflicting write buffer entry has been
issued and completed. At the same time, the no-merge bit is set in every write
buffer entry with which the load hit. A write buffer flush flag is also set. The
Mbox continues to request that write buffer entries be processed until all the
entries that were ahead of, and including, the conflicting write instructions at
the time of the load hit have been processed.
Some write instructions cannot be processed in the Scache without external
environment involvement. To support this, the Mbox retransmits a write
instruction at the Cbox’s request. This situation arises when the Scache block
is not dirty when the write instruction is issued, or when the access misses in
the Scache.

2.7.5 Ordering of Noncacheable Space Write Instructions
Special logic ensures that write instructions to noncacheable space are sent
offchip in the order in which their corresponding buffers were allocated (placed
in the pending-request queue).

Preliminary—Subject to Change—July 1996 2–37

2.8 Performance Measurement Support–Performance Counters

2.8 Performance Measurement Support–Performance
Counters
The 21164 contains a performance recording feature. The implementation
of this feature provides a mechanism to count various hardware events and
causes an interrupt upon counter overflow. Interrupts are triggered six cycles
after the event, and therefore, the exception PC may not reflect the exact
instruction causing counter overflow. Three counters are provided to allow
accurate comparison of two variables under a potentially nonrepeatable
experimental condition. Counter inputs include:
•

Issues

•

Nonissues

•

Total cycles

•

Pipe dry

•

Pipe freeze

•

Mispredicts and cache misses

•

Counts for various instruction classifications

In addition, the 21164 provides one signal-pin input (perf_mon_h) to measure
external events at a maximum rate determined by the selected system clock
speed (see Table 5–12).
For information about counter control, refer to the following IPR descriptions:
•

Hardware interrupt clear (HWINT_CLR) register (see Section 5.1.23)

•

Interrupt summary register (ISR) (see Section 5.1.24)

•

Performance counter (PMCTR) register (see Section 5.1.27)

•

Bcache control (BC_CONTROL) register bits <24:19> (see Section 5.3.4 and
Table 5–31)

2.9 Floating-Point Control Register
Figure 2–3 shows the format of the floating-point control register (FPCR) and
Table 2–10 describes the fields.

2–38 Preliminary—Subject to Change—July 1996

2.9 Floating-Point Control Register

Figure 2–3 Floating-Point Control Register (FPCR) Format

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48

S I U U DYN I I U O D I O D I
U N N N _RM O N N V Z N V Z N
VE F F E V F E V
ME F D
DDD
DD Z

RAZ/IGN
MLO-011301

Table 2–10 Floating-Point Control Register Bit Descriptions
Bit

Description (Meaning When Set)

<63>

Summary bit (SUM). Records bitwise OR of FPCR exception bits. Equal to
FPCR<57 | 56 | 55 | 54 | 53 | 52>

<62>

Inexact disable (INED). Suppress INE trap and place correct IEEE
nontrapping result in the destination register if the 21164 is capable of
producing correct IEEE nontrapping result.

<61>

Underflow disable (UNFD). Subset support: Suppress UNF trap if UNDZ
is also set and the /S qualifier is set on the instruction.

<60>

Underflow to zero (UNDZ). When set together with UNFD, on underflow,
the hardware places a true zero (all 64 bits zero) in the destination register
rather than the denormal number specified by the IEEE standard.

<59,58>

Dynamic routing mode (DYN). Indicates the rounding mode to be used by
an IEEE floating-point operate instruction when the instruction’s function
field specifies dynamic mode (/D). The assignments are:

DYN

IEEE Rounding Mode Selected

Chopped rounding mode

Minus infinity

Normal rounding

Plus infinity

<57>

Integer overflow (IOV). An integer arithmetic operation or a conversion
from floating to integer overflowed the destination precision.

<56>

Inexact result (INE). A floating arithmetic or conversion operation gave a
result that differed from the mathematically exact result.
(continued on next page)

Preliminary—Subject to Change—July 1996 2–39

2.9 Floating-Point Control Register

Table 2–10 (Cont.) Floating-Point Control Register Bit Descriptions
Bit

Description (Meaning When Set)

<55>

Underflow (UNF). A floating arithmetic or conversion operation
underflowed the destination exponent.

<54>

Overflow (OVF). A floating arithmetic or conversion operation overflowed
the destination exponent.

<53>

Division by zero (DZE). An attempt was made to perform a floating divide
operation with a divisor of zero.

<52>

Invalid operation (INV). An attempt was made to perform a floating
arithmetic, conversion, or comparison operation, and one or more of the
operand values were illegal.

<51>

Overflow disable (OVFD). Not supported.

<50>

Division by zero disable (DZED). Not supported.

<49>

Invalid operation disable (INVD). Not supported.

<48:0>

Reserved. Read as zero; ignored when written.

2.10 Design Examples
The 21164 can be designed into many different uniprocessor and multiprocessor
system configurations. Figures 2–4, 2–5, and 2–6 illustrate three possible
configurations. These configurations employ additional system/memory
controller chipsets.
Figure 2–4 shows a typical uniprocessor system with a board-level cache. This
system configuration could be used in standalone or networked workstations.

2–40 Preliminary—Subject to Change—July 1996

2.10 Design Examples

Figure 2–4 Typical Uniprocessor Configuration

External
Cache
Tag

21164

Addr/cmd

External
Cache
Data
Data

Main Memory
I/O Bus

Memory
and I/O
Interface
DRAM
Banks

DRAM
Bank
LJ-04040.AI

Figure 2–5 shows a typical multiprocessor system, each processor with a
board-level cache. Each interface controller must employ a duplicate tag store
to maintain cache coherency. This system configuration could be used in a
networked database server application.

Preliminary—Subject to Change—July 1996 2–41

2.10 Design Examples

Figure 2–5 Typical Multiprocessor Configuration

21164

Addr/cmd

External
Cache
Tag

21164

Addr/cmd

External
Cache
Data
Data

Bus
Interface

External
Cache
Tag

External
Cache
Data
Data

Duplicate
Tag
Store

Bus
Interface

Duplicate
Tag
Store

System Bus

I/O
Bridge

Memory

I/O Bus

Memory

I/O
Bridge

I/O Bus

LJ-04041.AI

Figure 2–6 shows a cacheless multiprocessor system. This system configuration
could be used in high-bandwith dedicated server applications.

2–42 Preliminary—Subject to Change—July 1996

2.10 Design Examples

Figure 2–6 Cacheless Multiprocessor Configuration

21164

Bus
Interface

System Bus

I/O
Bridge

I/O Bus

Memory

I/O
Bridge

I/O Bus

LJ-04039.AI

Preliminary—Subject to Change—July 1996 2–43

3
Hardware Interface
This chapter contains the 21164 microprocessor logic symbol and provides a
list of signal names and their functions.

3.1 Alpha 21164 Microprocessor Logic Symbol
Figure 3–1 shows the logic symbol for the 21164 chip.

Preliminary—Subject to Change—July 1996 3–1

3.1 Alpha 21164 Microprocessor Logic Symbol

Figure 3–1

Alpha 21164 Microprocessor Logic Symbol

21164
addr_bus_req_h
cack_h
cfail_h
dack_h
data_bus_req_h
fill_h

addr_h<39:4>

System/Bcache
Interface

addr_cmd_par_h
addr_res_h<2:0>
cmd_h<3:0>
data_h<127:0>

fill_error_h

data_check_h<15:0>

fill_id_h
fill_nocheck_h
idle_bc_h
shared_h
system_lock_flag_h

data_ram_oe_h
data_ram_we_h
index_h<25:4>
int4_valid_h<3:0>
scache_set_h<1:0>
st_clk_h
tag_ctl_par_h
tag_data_h<38:20>
tag_data_par_h
tag_dirty_h
tag_ram_oe_h
tag_ram_we_h
tag_shared_h
tag_valid_h
victim_pending_h

irq_h<3:0>
mch_hlt_irq_h
pwr_fail_irq_h

Interrupts

sys_mch_chk_irq_h
clk_mode_h<1:0>
osc_clk_in_h

cpu_clk_out_h

Clocks

osc_clk_in_l
ref_clk_in_h
sys_reset_l
dc_ok_h
perf_mon_h
port_mode_h<1:0>
srom_data_h

sys_clk_out1_h
sys_clk_out1_l
sys_clk_out2_h
sys_clk_out2_l
srom_clk_h

Test Modes and

srom_oe_l

Miscellaneous

srom_present_l

tdi_h
temp_sense
tms_h
Vdd

tck_h
tdo_h
test_status_h<1:0>
trst_l

Vss
MK145506

3–2 Preliminary—Subject to Change—July 1996

3.2 Alpha 21164 Signal Names and Functions

3.2 Alpha 21164 Signal Names and Functions
The 21164 is contained in a 499-pin IPGA package. Of these pins, 292 are
used for functional signals. There are two spare (unused) signal pins. The
remaining pins are used for power (104) and ground (101).
The following table defines the 21164 signal types referred to in this section:
Signal Type

Definition

Bidirectional

Input only

Output only

The remaining two tables describe the function of each 21164 external signal.
Table 3–1 lists all signals in alphanumeric order. This table provides full signal
descriptions. Table 3–2 lists signals by function and provides an abbreviated
description.
Table 3–1

Alpha 21164 Signal Descriptions

Signal

Type

Count Description

addr_h<39:4>

Address bus. These bidirectional signals provide the
address of the requested data or operation between the
21164 and the system. If bit 39 is asserted, then the
reference is to noncached, I/O memory space.

addr_bus_req_h

Address bus request. The system interface uses
this signal to gain control of the addr_h<39:4>,
addr_cmd_par_h, and cmd_h<3:0> pins (see
Figure 4–30).

addr_cmd_par_h

Address command parity. This is the odd parity bit on
the current command and address buses. The 21164
takes a machine check if a parity error is detected. The
system should do the same if it detects an error.
(continued on next page)

Preliminary—Subject to Change—July 1996 3–3

3.2 Alpha 21164 Signal Names and Functions

Table 3–1 (Cont.)

Alpha 21164 Signal Descriptions

Signal

Type

Count Description

addr_res_h<1:0>

Address response bits <1> and <0>. For system
commands, the 21164 uses these pins to indicate the
state of the block in the Scache:
Bits

Command

Meaning

NOP

Nothing.

NOACK

Data not found or clean.

ACK/Scache

Data from Scache.

ACK/Bcache

Data from Bcache.

addr_res_h<2>

Address response bit <2>. For system commands, the
21164 uses this pin to indicate if the command hits in
the Scache or onchip load lock register.

cack_h

Command acknowledge. The system interface uses this
signal to acknowledge any one of the commands driven
by the 21164.

cfail_h

Command fail. This signal has two uses. It can be
asserted during a cack cycle of a WRITE BLOCK LOCK
command to indicate that the write operation is not
successful. In this case, both cack_h and cfail_h are
asserted together. It can also be asserted instead of
cack_h to force an instruction fetch/decode unit (Ibox)
timeout event. This causes the 21164 to do a partial
reset and trap to the machine check (MCHK) PALcode
entry point, which indicates a serious hardware error.

clk_mode_h<1:0>

Clock test mode. These signals specify a relationship
between osc_clk_in_h,l and the CPU cycle time. These
signals should be deasserted in normal operation mode.

cmd_h<3:0>

Command bus. These signals drive and receive the
commands from the command bus. The following
tables define the commands that can be driven on the
cmd_h<3:0> bus by the 21164 or the system. For
additional information, refer to Section 4.1.1.1.
(continued on next page)

3–4 Preliminary—Subject to Change—July 1996

3.2 Alpha 21164 Signal Names and Functions

Table 3–1 (Cont.)
Signal

Alpha 21164 Signal Descriptions
Type

Count Description
21164 Commands to System:
cmd_h
<3:0>

Command

Meaning

0000

NOP

Nothing.

0001

LOCK

Lock register address.

0010

FETCH

The 21164 passes a
FETCH instruction to
the system.

0011

FETCH_M

The 21164 passes a
FETCH_M instruction
to the system.

0100

MEMORY
BARRIER

MB instruction.

0101

SET DIRTY

Dirty bit set if shared
bit is clear.

0110

WRITE BLOCK

Request to write a
block.

0111

WRITE BLOCK
LOCK

Request to write a
block with lock.

1000

READ MISS0

Request for data.

1001

READ MISS1

Request for data.

1010

READ MISS MOD0

Request for data;
modify intent.

1011

READ MISS MOD1

Request for data;
modify intent.

1100

BCACHE VICTIM

Bcache victim should
be removed.

1101

—

Reserved.

1110

READ MISS MOD
STC0

Request for data,
STx_C data.

1111

READ MISS MOD
STC1

Request for data,
STx_C data.
(continued on next page)

Preliminary—Subject to Change—July 1996 3–5

3.2 Alpha 21164 Signal Names and Functions

Table 3–1 (Cont.)
Signal

Alpha 21164 Signal Descriptions
Type

Count Description
System Commands to 21164:
cmd_h
<3:0>

Command

Meaning

0000

NOP

Nothing.

0001

FLUSH

Remove block from
caches; return dirty
data.

0010

INVALIDATE

Invalidate the block
from caches.

0011

SET SHARED

Block goes to the
shared state.

0100

READ

Read a block.

0101

READ DIRTY

Read a block; set
shared.

0111

READ DIRTY/INV

Read a block;
invalidate.

cpu_clk_out_h

CPU clock output. This signal is used for test purposes.

dack_h

Data acknowledge. The system interface uses this
signal to control data transfer between the 21164 and
the system.

data_h<127:0>

128

Data bus. These signals are used to move data between
the 21164, the system, and the Bcache.

data_bus_req_h

Data bus request. If the 21164 samples this signal
asserted on the rising edge of sysclk n, then the 21164
does not drive the data bus on the rising edge of sysclk
n+1. Before asserting this signal, the system should
assert idle_bc_h for the correct number of cycles. If the
21164 samples this signal deasserted on the rising edge
of sysclk n, then the 21164 drives the data bus on the
rising edge of sysclk n+1. For timing details, refer to
Section 4.11.4.
(continued on next page)

3–6 Preliminary—Subject to Change—July 1996

3.2 Alpha 21164 Signal Names and Functions

Table 3–1 (Cont.)

Alpha 21164 Signal Descriptions

Signal

Type

Count Description

data_check_h<15:0>

Data check. These signals set even byte parity or INT8
ECC for the current data cycle. Refer to Section 4.14.1
for information on the purpose of each data_check_h
bit.

data_ram_oe_h

Data RAM output enable. This signal is asserted for
Bcache read operations.

data_ram_we_h

Data RAM write-enable. This signal is asserted for any
Bcache write operation. Refer to Section 5.3.5 for timing
details.

dc_ok_h

dc voltage OK. Must be deasserted until dc voltage
reaches proper operating level. After that, dc_ok_h is
asserted.

fill_h

Fill warning. If the 21164 samples this signal asserted
on the rising edge of sysclk n, then the 21164 provides
the address indicated by fill_id_h to the Bcache on the
rising edge of sysclk n+1. The Bcache begins to write in
that sysclk. At the end of sysclk n+1, the 21164 waits
for the next sysclk and then begins the write operation
again if dack_h is not asserted. Refer to Section 4.11.3
for timing details.

fill_error_h

Fill error. If this signal is asserted during a fill from
memory, it indicates to the 21164 that the system has
detected an invalid address or hard error. The system
still provides an apparently normal read sequence with
correct ECC/parity though the data is not valid. The
21164 traps to the machine check (MCHK) PALcode
entry point and indicates a serious hardware error.
fill_error_h should be asserted when the data is
returned. Each assertion produces a MCHK trap.

fill_id_h

Fill identification. Asserted with fill_h to indicate which
register is used. The 21164 supports two outstanding
load instructions. If this signal is asserted when the
21164 samples fill_h asserted, then the 21164 provides
the address from miss register 1. If it is deasserted,
then the address in miss register 0 is used for the read
operation.

fill_nocheck_h

Fill checking off. If this signal is asserted, then the
21164 does not check the parity or ECC for the current
data cycle on a fill.
(continued on next page)

Preliminary—Subject to Change—July 1996 3–7

3.2 Alpha 21164 Signal Names and Functions

Table 3–1 (Cont.)

Alpha 21164 Signal Descriptions

Signal

Type

Count Description

idle_bc_h

Idle Bcache. When asserted, the 21164 finishes the
current Bcache read or write operation but does not
start a new read or write operation until the signal
is deasserted. The system interface must assert this
signal in time to idle the Bcache before fill data arrives.

index_h<25:4>

Index. These signals index the Bcache.

int4_valid_h<3:0>

INT4 data valid. During write operations to noncached
space, these signals are used to indicate which INT4
bytes of data are valid. This is useful for noncached
write operations that have been merged in the write
buffer.
int4_valid_h<3:0>

Write Meaning

xxx1

data_h<31:0> valid

xx1x

data_h<63:32> valid

x1xx

data_h<95:64> valid

1xxx

data_h<127:96> valid

During read operations to noncached space, these
signals indicate which INT8 bytes of a 32-byte block
need to be read and returned to the processor. This is
useful for read operations to noncached memory.
int4_valid_h<3:0>

Read Meaning

xxx1

data_h<63:0> valid

xx1x

data_h<127:64> valid

x1xx

data_h<191:128> valid

1xxx

data_h<255:192> valid

Note: For both read and write operations, multiple
int4_valid_h<3:0> bits can be set simultaneously.
(continued on next page)

3–8 Preliminary—Subject to Change—July 1996

3.2 Alpha 21164 Signal Names and Functions

Table 3–1 (Cont.)

Alpha 21164 Signal Descriptions

Signal

Type

Count Description

irq_h<3:0>

System interrupt requests. These signals have multiple
modes of operation. During normal operation, these
level-sensitive signals are used to signal interrupt
requests. During initialization, these signals are used to
set up the CPU cycle time divisor for sys_clk_out1_h,l
as follows:
irq_h
<3>

<2>

<1>

<0>

Ratio

Low

High

Low

High

Low

High

Low

High

Low

High

Low

High

Low

High

Low

High

Low

High

Low

High

Low

High

Low

High

Low

High

Low

High

mch_hlt_irq_h

Machine halt interrupt request. This signal has
multiple modes of operation. During initialization,
this signal is used to set up sys_clk_out2_h,l delay (see
Table 4–3). During normal operation, it is used to signal
a halt request.

osc_clk_in_h
osc_clk_in_l

I
I

1
1

Oscillator clock inputs. These signals provide the
differential clock input that is the fundamental timing
of the 21164. These signals are driven at twice the
desired internal clock frequency. (Under normal
operating conditions the CPU cycle time is one-half
the frequency of osc_clk_in.)
(continued on next page)

Preliminary—Subject to Change—July 1996 3–9

3.2 Alpha 21164 Signal Names and Functions

Table 3–1 (Cont.)

Alpha 21164 Signal Descriptions

Signal

Type

Count Description

perf_mon_h

Performance monitor. This signal can be used as an
input to the 21164 internal performance monitoring
hardware from offchip events (such as bus activity).
Refer to Section 5.1.27 for information on the PMCTR
register.

port_mode_h<1:0>

Select test port interface modes (normal, manufacturing,
and debug). For normal operation, both signals must be
deasserted.

pwr_fail_irq_h

Power failure interrupt request. This signal has
multiple modes of operation. During initialization,
this signal is used to set up sys_clk_out2_h,l delay (see
Table 4–3). During normal operation, this signal is used
to signal a power failure.

ref_clk_in_h

Reference clock input. Optional. Used to synchronize
the timing of multiple microprocessors to a single
reference clock. If this signal is not used, it must be tied
to Vdd for proper operation.

scache_set_h<1:0>

Secondary cache set. During a read miss request, these
signals indicate the Scache set number that will be
filled when the data is returned. This information can
be used by the system to maintain a duplicate copy of
the Scache tag store.

shared_h

Keep block status shared. For systems without a
Bcache, when a WRITE BLOCK/NO VICTIM PENDING
or WRITE BLOCK LOCK command is acknowledged,
this pin can be used to keep the block status shared or
private in the Scache.

srom_clk_h

Serial ROM clock. Supplies the clock that causes the
SROM to advance to the next bit. The cycle time of this
clock is 128 times the cycle time of the CPU clock.

srom_data_h

Serial ROM data. Input for the SROM.

srom_oe_l

Serial ROM output enable. Supplies the output enable
to the SROM.

srom_present_l1

Serial ROM present. Indicates that SROM is present
and ready to load the Icache.

1 This signal is shown as bidirectional. However, for normal operation it is input only. The output function is

used during manufacturing test and verification only.

(continued on next page)

3–10 Preliminary—Subject to Change—July 1996

3.2 Alpha 21164 Signal Names and Functions

Table 3–1 (Cont.)

Alpha 21164 Signal Descriptions

Signal

Type

Count Description

st_clk_h

STRAM clock. Clock for Bcache synchronously
timed RAMs (STRAMs). This signal is synchronous
with index_h<25:4> during private read and write
operations, and with sys_clk_out1_h,l during read and
fill operations.

sys_clk_out1_h
sys_clk_out1_l

O
O

1
1

System clock outputs. Programmable system clock
(cpu_clk_out_h divided by a value of 3 to 15) is used
for board-level cache and system logic.

sys_clk_out2_h
sys_clk_out2_l

O
O

1
1

System clock outputs. A version of sys_clk_out1_h,l
delayed by a programmable amount from 0 to 7 CPU
cycles.

sys_mch_chk_irq_h

System machine check interrupt request. This signal
has multiple modes of operation. During initialization,
it is used to set up sys_clk_out2_h,l delay (see
Table 4–3). During normal operation, it is used to
signal a machine interrupt check request.

sys_reset_l

System reset. This signal protects the 21164 from
damage during initial power-up. It must be asserted
until dc_ok_h is asserted. After that, it is deasserted
and the 21164 begins its reset sequence.

system_lock_flag_h

System lock flag. During fills, the 21164 logically ANDs
the value of the system copy with its own copy to
produce the true value of the lock flag.

tag_ctl_par_h

Tag control parity. This signal indicates odd parity for
tag_valid_h, tag_shared_h, and tag_dirty_h. During
fills, the system should drive the correct parity based on
the state of the valid, shared, and dirty bits.

tag_data_h<38:20>

Bcache tag data bits. This bit range supports 1M-byte
to 64M-byte Bcaches.

tag_data_par_h

Tag data parity bit. This signal indicates odd parity for
tag_data_h<38:20>.

tag_dirty_h

Tag dirty state bit. During fills, the system should
assert this signal if the 21164 request is a READ MISS
MOD, and the shared bit is not asserted. Refer to
Table 4–6 for information about Bcache protocol.

tag_ram_oe_h

Tag RAM output enable. This signal is asserted during
any Bcache read operation.
(continued on next page)

Preliminary—Subject to Change—July 1996 3–11

3.2 Alpha 21164 Signal Names and Functions

Table 3–1 (Cont.)

Alpha 21164 Signal Descriptions

Signal

Type

Count Description

tag_ram_we_h

Tag RAM write-enable. This signal is asserted
during any tag write operation. During the first
CPU cycle of a write operation, the write pulse is
deasserted. In the second and following CPU cycles
of a write operation, the write pulse is asserted if the
corresponding bit in the write pulse register is asserted.
Bits BC_WE_CTL<8:0> control the shape of the pulse
(see Section 5.3.5).

tag_shared_h

Tag shared bit. During fills, the system should drive
this signal with the correct value to mark the cache
block as shared. See Table 4–6 for information about
Bcache protocol.

tag_valid_h

Tag valid bit. During fills, this signal is asserted to
indicate that the block has valid data. See Table 4–6 for
information about Bcache protocol.

tck_h

JTAG boundary scan clock.

tdi_h

JTAG serial boundary scan data-in signal.

tdo_h

JTAG serial boundary scan data-out signal.

temp_sense

Temperature sense. This signal is used to measure the
die temperature and is for manufacturing use only. For
normal operation, this signal must be left disconnected.

test_status_h<1:0>

Icache test status. These signals are used for
manufacturing test purposes only to extract Icache test
status information from the chip. test_status_h<0>
is asserted if ICSR<39> is true, on Ibox timeout, or
remains asserted if the Icache built-in self-test (BiSt)
fails. Also, test_status_h<0> outputs the value written
by PALcode to test_status_h<1> through IPR access.
For additional information, refer to Section 12.2.2.

tms_h

JTAG test mode select signal.

trst_l

JTAG test access port (TAP) reset signal.

victim_pending_h

Victim pending. When asserted, this signal indicates
that the current read miss has generated a victim.

1 This signal is shown as bidirectional. However, for normal operation it is input only. The output function is

used during manufacturing test and verification only.

3–12 Preliminary—Subject to Change—July 1996

3.2 Alpha 21164 Signal Names and Functions

Table 3–2 lists signals by function and provides an abbreviated description.
Table 3–2

Alpha 21164 Signal Descriptions by Function

Signal

Type

Count Description

clk_mode_h<1:0>

Clock test mode.

cpu_clk_out_h

CPU clock output.

osc_clk_in_h,l

Oscillator clock inputs.

ref_clk_in_h

Reference clock input.

st_clk_h

Bcache STRAM clock output.

sys_clk_out1_h,l

System clock outputs.

sys_clk_out2_h,l

System clock outputs.

sys_reset_l

System reset.

data_h<127:0>

128

Data bus.

data_check_h<15:0>

Data check.

data_ram_oe_h

Data RAM output enable.

data_ram_we_h

Data RAM write-enable.

index_h<25:4>

Index.

tag_ctl_par_h

Tag control parity.

tag_data_h<38:20>

Bcache tag data bits.

tag_data_par_h

Tag data parity bit.

tag_dirty_h

Tag dirty state bit.

tag_ram_oe_h

Tag RAM output enable.

tag_ram_we_h

Tag RAM write-enable.

tag_shared_h

Tag shared bit.

tag_valid_h

Tag valid bit.

Clocks

Bcache

(continued on next page)

Preliminary—Subject to Change—July 1996 3–13

3.2 Alpha 21164 Signal Names and Functions

Table 3–2 (Cont.)
Signal

Alpha 21164 Signal Descriptions by Function
Type

Count Description

addr_h<39:4>

addr_bus_req_h

Address bus request.

addr_cmd_par_h

Address command parity.

addr_res_h<2:0>

Address response.

cack_h

Command acknowledge.

cfail_h

Command fail.

cmd_h<3:0>

Command bus.

dack_h

Data acknowledge.

data_bus_req_h

Data bus request.

fill_h

Fill warning.

fill_error_h

Fill error.

fill_id_h

Fill identification.

fill_nocheck_h

Fill checking off.

idle_bc_h

Idle Bcache.

int4_valid_h<3:0>

INT4 data valid.

scache_set_h<1:0>

Secondary cache set.

shared_h

Keep block status shared.

system_lock_flag_h

System lock flag.

victim_pending_h

Victim pending.

irq_h<3:0>

System interrupt requests.

mch_hlt_irq_h

Machine halt interrupt request.

pwr_fail_irq_h

Power failure interrupt request.

sys_mch_chk_irq_h

System machine check interrupt request.

System Interface
Address bus.

Interrupts

(continued on next page)

3–14 Preliminary—Subject to Change—July 1996

3.2 Alpha 21164 Signal Names and Functions

Table 3–2 (Cont.)
Signal

Alpha 21164 Signal Descriptions by Function
Type

Count Description

Test Modes and Miscellaneous
dc_ok_h

dc voltage OK.

perf_mon_h

Performance monitor.

port_mode_h<1:0>

Select test port interface modes (normal,
manufacturing, and debug).

srom_clk_h

Serial ROM clock.

srom_data_h

Serial ROM data.

Serial ROM output enable.

Serial ROM present.

tck_h

JTAG boundary scan clock.

tdi_h

JTAG serial boundary scan data in.

tdo_h

JTAG serial boundary scan data out.

temp_sense

Temperature sense.

test_status_h<1:0>

Icache test status.

tms_h

JTAG test mode select.

JTAG test access port (TAP) reset.

srom_oe_l
srom_present_l

trst_l

1 This signal is shown as bidirectional. However, for normal operation is is input only. The output

function is used during manufacturing test and verification only.

Preliminary—Subject to Change—July 1996 3–15

4
Clocks, Cache, and External Interface
Functional Description
This chapter describes the 21164 microprocessor external interface, which
includes the backup cache (Bcache) and system interfaces. It also describes
the clock circuitry, locks, interrupt signals, and ECC/parity generation. It is
organized as follows:
•

Introduction to the external interface

•

Clocks

•

Physical address considerations

•

Bcache structure and operation

•

Cache coherency

•

Locks mechanisms

•

21164-to-Bcache transactions

•

21164-initiated system transactions

•

System-initiated transactions

•

Data bus and command/address bus contention

•

21164 interface restrictions

•

21164/system race conditions

•

Data integrity, Bcache errors, and command/address errors

•

Interrupts

Chapter 3 lists and defines all 21164 hardware interface signal pins. Chapter 9
describes the 21164 hardware interface electrical requirements.

Preliminary—Subject to Change—July 1996 4–1

4.1 Introduction to the External Interface

4.1 Introduction to the External Interface
A 21164-based system can be divided into three major sections:
•

21164 microprocessor

•

Optional external Bcache

•

System interface logic
Optional duplicate tag store
Optional lock register
Optional victim buffers

The 21164 external interface is flexible and mandates few design rules,
allowing a wide range of prospective systems. The interface includes a 128-bit
bidirectional data bus, a 36-bit bidirectional address bus, and several control
signals.
Read and write speeds of the optional Bcache array can be programmed by
means of register bits. Read and write speeds are independent of each other
and the system interface clock frequency.
The cache system supports a selectable 32-byte or 64-byte block size.
Figure 4–1 shows a simplified view of the external interface. The function and
purpose of each signal is described in Chapter 3.

4.1.1 System Interface
This section describes the system or external bus interface. The system
interface is made up of bidirectional address and command buses, a data bus
that is shared with the Bcache interface, and several control signals.
The system interface is under the control of the bus interface unit (BIU) in the
Cbox. The system interface is a 128-bit bidirectional data bus.
The cycle time of the system interface is programmable to speeds of 3 to 15
times the CPU cycle time (sysclk ratio). All system interface signals are driven
or sampled by the 21164 on the rising edge of signal sys_clk_out1_h. In
this chapter, this edge is sometimes referred to as ‘‘sysclk.’’ Precisely when
interface signals rise and fall does not matter as long as they meet the setup
and hold times specified in Chapter 9.

4–2 Preliminary—Subject to Change—July 1996

4.1 Introduction to the External Interface

Figure 4–1

21164

Alpha 21164 System/Bcache Interface

System Memory
and I/O

addr_h<39:4>
addr_bus_req_h
addr_cmd_par_h
addr_res_h<2:0>
cack_h
cfail_h
cmd_h<3:0>
dack_h
data_bus_req_h
fill_h

System
Interface

fill_error_h
fill_id_h
fill_nocheck_h
idle_bc_h

Optional

int4_valid_h<3:0>
scache_set_h<1:0>

Lock
Register

shared_h
Scache
Miss

system_lock_flag_h
victim_pending_h

Victim
Buffers

data_h<127:0>

Duplicate
Tag
Store

index_h<25:4>
Optional

Victim
Buffer

Tag
SRAM

State
V,D,S,P

Bcache
Data
SRAM
Bcache
Interface

tag_data_h<38:20>,p
tag_valid_h
tag_dirty_h
tag_shared_h
tag_ctl_par_h
data_check_h<15:0>
irq_h<3:0>
mch_hlt_irq_h
pwr_fail_irq_h
sys_mch_chk_irq_h

Interrupts

MK−1455−04

Preliminary—Subject to Change—July 1996 4–3

4.1 Introduction to the External Interface

4.1.1.1 Commands and Addresses
The 21164 can take up to two commands from the system at a time. The
Scache or Bcache or both are probed to determine what must be done with the
command.
•

If nothing is to be done, the 21164 acknowledges receiving the command.

•

If a Bcache read, set shared, or invalidate operation is required, the
21164 performs the task as soon as the Bcache becomes free. The
21164 acknowledges receiving the command at the start of the Bcache
transaction.

There are two miss and two victim buffers in the BIU. They can hold one or
two miss addresses and one or two Scache victim addresses or up to two shared
write operations at a time.
•

A miss occurs when the 21164 searches its caches but does not find the
addressed block. The 21164 can queue two misses to the system.

•

An Scache victim occurs when the 21164 deallocates a dirty block from the
Scache.

4.1.2 Bcache Interface
The 21164 includes an interface and control for an optional backup cache
(Bcache). The Bcache interface is made up of the following:
•

A 128-bit data bus (which it shares with the system interface)

•

Index address bits (index_h<25:4>)

•

Tag and state bits for determining hit and coherence

•

SRAM output and write control signals

4–4 Preliminary—Subject to Change—July 1996

4.2 Clocks

4.2 Clocks
The 21164 develops three clock signals that are available at output pins:
Signal

Description

cpu_clk_out_h

A 21164 internal clock that may or may not drive the system clock.

sys_clk_out1_h,l

A clock of programmable speed supplied to the external interface.

sys_clk_out2_h,l

A delayed copy of sys_clk_out1_h,l. The delay is programmable
and is an integer number of cpu_clk_out_h periods.

The 21164 may use ref_clk_in_h as a reference clock when generating
sys_clk_out1_h,l and sys_clk_out2_h,l. The behavior of the programmable
clocks during the reset sequence is described in Section 7.1.

4.2.1 CPU Clock
The 21164 uses the differential input clock lines osc_clk_in_h,l as a
source to generate its CPU clock. The input signals clk_mode_h<1:0>
control generation of the CPU clock as listed in Table 4–1 and as shown in
Figure 4–2.
Table 4–1 CPU Clock Generation Control
Mode

clk_mode_h<1:0>

Divisor

Description

Normal

Usual operation—CPU clock frequency
is ½ input frequency.

Chip test

CPU clock frequency is the same as the
input clock frequency to accommodate
chip testers.

Module test

CPU clock frequency is ¼ input
frequency to accommodate module
testers.

Reset

—

Initializes CPU clock, allowing system
clock to be synchronized to a stable
reference clock.

1 Divide by 2 or 4 should be used to obtain the best internal clock.

Caution
A clock source should always be provided on osc_clk_in_h,l when
signal dc_ok_h is asserted.

Preliminary—Subject to Change—July 1996 4–5

4.2 Clocks

Figure 4–2 Clock Signals and Functions

21164
osc_clk_in_h, l
clk_mode_h<1:0>

CPU Clock
Divider
(/1, /2, or /4)

Digital
PLL

cpu_clk_out_h

ref_clk_in_h

irq_h<3:0>

System Clock
Divider
(/3 through /15)

mch_hlt_irq_h
pwr_fail_irq_h
sys_mch_chk_irq_h

System Clock
Delay
(0 through 7)

sys_clk_out1_h, l

sys_clk_out2_h, l

sys_reset_l
dc_okay_h
MK−1455−02

4.2.2 System Clock
The CPU clock is the source clock used to generate the system clock
sys_clk_out1_h,l. The system clock divider controls the frequency of
sys_clk_out1_h,l. The divisor, 3 to 15, is obtained from the four interrupt lines
irq_h<3:0> at power-up as listed in Table 4–2. The system clock frequency
is determined by dividing the ratio into the CPU clock frequency. Refer to
Section 7.2 for information on sysclk behavior during reset.

4–6 Preliminary—Subject to Change—July 1996

4.2 Clocks

Table 4–2 System Clock Divisor
irq_h<3>

irq_h<2>

irq_h<1>

irq_h<0>

Ratio

Low

High

Low

High

Low

High

Low

High

Low

High

Low

High

Low

High

Low

High

Low

High

Low

High

Low

High

Low

High

Low

High

Low

High

Figure 4–3 shows the 21164 driving the system clock on a uniprocessor
system.
Figure 4–3

Alpha 21164 Uniprocessor Clock

Memory
ASIC
sys_clk_out
21164
Bus
ASIC

LJ-03676-TI0

Preliminary—Subject to Change—July 1996 4–7

4.2 Clocks

4.2.3 Delayed System Clock
The system clock sys_clk_out1_h,l is the source clock for the delayed system
clock sys_clk_out2_h,l. These clock signals provide flexible timing for system
use. The delay unit, 0 to 7, is obtained from the three interrupt signals:
mch_hlt_irq_h, pwr_fail_irq_h, and sys_mch_chk_irq_h at power-up, as
listed in Table 4–3. The output of this programmable divider is symmetric
if the divisor is even. The output is asymmetric if the divisor is odd. When
the divisor is odd, the clock is high for an extra cycle. Refer to Section 7.2 for
information on sysclk behavior during reset.
Table 4–3 System Clock Delay
sys_mch_chk_irq_h

pwr_fail_irq_h

mch_hlt_irq_h

Delay Cycles

Low

High

Low

High

Low

High

Low

High

Low

High

Low

High

4.2.4 Reference Clock
The 21164 provides a reference clock input so that other CPUs and system
devices can be synchronized in multiprocessor systems. If a clock is asserted
on signal ref_clk_in_h, then the sys_clk_out1_h,l signals are synchronized to
that reference clock. The reference clock input should be connected to Vdd if
the input is not to be used.
The 21164 synchronizes the sys_clk_out1_h frequency with the ref_clk_in_h
signal by means of a digital phase-locked loop (DPLL). The DPLL does not
lock the two frequencies, but rather, creates a window. To accomplish this, the
frequency of signal sys_clk_out1 must be slightly higher, but no greater than
0.35% higher, than that of signal ref_clk_in_h. This causes the rising edge of
sys_clk_out1 to drift back toward the rising edge of ref_clk_in_h. The 21164
detects when the edges meet and stalls the internal clock generator for one
osc_clk_in cycle. This moves the rising edge of sys_clk_out1 back in front of
ref_clk_in_h.

4–8 Preliminary—Subject to Change—July 1996

4.2 Clocks

Figure 4–4 shows a multiprocessor 21164 system synchronized to a reference
clock.
Figure 4–4

Alpha 21164 Reference Clock for Multiprocessor Systems

Memory
ASIC
sys_clk_out

ref_clk_in
21164

Bus
ASIC
Reference
Clock
Memory
ASIC
sys_clk_out

ref_clk_in
21164

Bus
ASIC

LJ-03675-TI0

4.2.4.1 Reference Clock Examples
This section contains example calculations of setting time in systems that use
the DPLL for synchronization.
After sys_clk_out1_h,l has stabilized (20 cycles after irq_h<3:0> have
settled) there will be a delay before sys_clk_out1_h,l comes into lock with
ref_clk_in_h. The two cases for this event are described in Section 4.2.4.1.1
and Section 4.2.4.1.2.

Preliminary—Subject to Change—July 1996 4–9

4.2 Clocks

4.2.4.1.1 Case 1: ref_clk_in_h Initially Sampled Low by DPLL
When the
DPLL initially samples ref_clk_in_h in the low state, as shown in Figure 4–5,
it slips its internal cycle repeatedly until it samples ref_clk_in_h in the high
state. After it samples ref_clk_in_h in the high state, the DPLL stays in lock
mode.
Figure 4–5 ref_clk_in_h Initially Sampled Low

CPU Clock
(Internal)

sys_clk_out 1_h
ref_clk_in_h

LJ-04000.AI

Note
The timing diagram shows a sys_clk_out1_h,l ratio of 4.

The worst case (slowest) maximum rate at which the DPLL will slip its
internal cycle (the frequency of phase slips) is calculated from the lock range
specification of 0.35%. In effect, an average of 0.35% period is added to each
sys_clk_out1_h,l period until lock mode is reached.

SettlingTime =

Ref C lockLowRatio Ref C lockP eriod

0:0035

Note
The reference clock low ratio equals the portion of the reference clock
period that ref_clk_in_h is low.

Assuming the worst case ref_clk_in_h duty cycle is 60/40 to 40/60:

SettlingTime = 0 63
:

Ref C lockP eriod

0:0035

4–10 Preliminary—Subject to Change—July 1996

= 171 3 RefClockPeriod

4.2 Clocks

Depending upon the sys_clk_out1_h,l ratio, the DPLL may come into lock
much more quickly. The DPLL may insert phase slips more frequently at
smaller sys_clk_out1_h,l ratios.
4.2.4.1.2 Case 2: ref_clk_in_h Initially Sampled High by DPLL
When the
DPLL initially samples ref_clk_in_h in the high state, as shown in Figure 4–6,
it will not slip its internal cycle until it samples ref_clk_in_h in the low state.
After it samples ref_clk_in_h in the low state, the DPLL stays in lock mode.
Figure 4–6 ref_clk_in_h Initially Sampled High

CPU Clock
(Internal)

sys_clk_out1_h
ref_clk_in_h

LJ-04001.AI

The rate at which sys_clk_out1_h,l gains on ref_clk_in_h depends on the
difference in frequency of the two signals. Assuming that:
ref_clk_in_h is nominally selected to run 0.175% slower than
sys_clk_out1_h,l (in the center of the specified lock range),
and that worst case deviation of 200 PPM from the specified frequency for
ref_clk_in_h and osc_clk_in_h,l,
Then the worst case (smallest) frequency difference is calculated to be,
0:00175 0 200PPM 0 200PPM = 0:00135 = 0:135%

SettlingTime =

Ref C lockH ighRatio Ref C lockP eriod

0:00135

Note
The reference clock high ratio equals the portion of the ref_clk_in_h
period that ref_clk_in_h is high.

Preliminary—Subject to Change—July 1996 4–11

4.2 Clocks

Assuming the worst case ref_clk_in_h duty cycle is 60/40 to 40/60:

SettlingTime = 0 63 0 00135
:

Ref C lockP eriod
:

= 444 3 RefClockPeriod

4.3 Physical Address Considerations
This section lists and describes the physical address regions. Cache and data
wrapping characteristics of physical addresses are also described.

4.3.1 Physical Address Regions
Physical memory of the 21164 is divided into three regions:
1. The first region is the first half of the physical address space. It is treated
by the 21164 as memory-like.
2. The second region is the second half of the physical address space except
for a 1M-byte region reserved for Cbox IPRs. It is treated by the 21164 as
noncachable.
3. The third region is the 1M-byte region reserved for Cbox IPRs.
In the first region, write invalidate caching, write merging, and load merging
are all permitted. All 21164 accesses in this region are 32- or 64-byte
depending on the programmable block size.
The 21164 does not cache data accessed in the second and third region of
the physical address space; 21164 read accesses in these regions are always
INT32 requests. Load merging is permitted, but the request includes a mask
to tell the system environment which INT8s are accessed. Write merging is
permitted. Write accesses are INT32 requests with a mask indicating which
INT4s are actually modified. The 21164 never writes more than 32 bytes at a
time in noncached space.
The 21164 does not broadcast accesses to the Cbox IPR region if they map to a
Cbox IPR. Accesses in this region, that are not to a defined Cbox IPR, produce
UNDEFINED results. The system should not probe this region.
Table 4–4 shows the 21164 physical memory regions.

4–12 Preliminary—Subject to Change—July 1996

4.3 Physical Address Considerations

Table 4–4 Physical Memory Regions
Region

Address Range

Description

Memory-like

00 0000 0000–
7F FFFF FFFF16

Write invalidate cached, load, and store
merging allowed.

Noncacheable

80 0000 0000–
FF FFEF FFFF16

Not cached, load merging limited.

IPR region

FF FFF0 0000–
FF FFFF FFFF16

Accesses do not appear on the interface
unless an undefined location is accessed
(which produces UNDEFINED results).

4.3.2 Data Wrapping
The 21164 requires that wrapped read operations be performed on INT16
boundaries. READ, READ DIRTY, and FLUSH commands are all wrapped on
INT16 boundaries as described here. The valid wrap orders for 64-byte blocks
are selected by addr_h<5:4>. They are:
0, 1, 2, 3
1, 0, 3, 2
2, 3, 0, 1
3, 2, 1, 0
For 32-byte blocks, the valid wrap orders are selected by addr_h<4>. They
are:
0, 1
1, 0
Similarly, when the system interface supplies a command that returns data
from the 21164 caches, the values that the system drives on addr_h<5:4>
determine the order in which data is supplied by the 21164.
WRITE BLOCK and WRITE BLOCK LOCK commands from the 21164 are
not wrapped. They always write INT16 0, 1, 2, and 3. BCACHE VICTIM
commands provide the data with the same wrap order as the read miss that
produced them.

Preliminary—Subject to Change—July 1996 4–13

4.3 Physical Address Considerations

4.3.3 Noncached Read Operations
Read operations to physical addresses that have addr_h<39> asserted are
not cached in the Dcache, Scache, or Bcache. They are merged like other read
operations in the miss address file (MAF). To prevent several read operations
to noncached memory from being merged into a single 32-byte bus request,
software must insert memory barrier (MB) instructions or set MAF_MODE
IPR bit (IO_NMERGE). The MAF merges as many Dstream read operations
together as it can and sends the request to the BIU through the Scache.
Rather than merging two 32-byte requests into a single 64-byte request, the
BIU requests a READ MISS from the system. Signals int4_valid_h<3:0>
indicate which of the four quadwords are being requested by software. The
system should return the fill data to the 21164 as usual. The 21164 does not
write the Dcache, Scache, or Bcache with the fill data. The requested data is
written in the register file or Icache.
Note
A special case using int4_valid_h<3:0> occurs during an Icache fill. In
this case the entire returned block is valid although int4_valid_h<3:0>
indicates zero.

4.3.4 Noncached Write Operations
Write operations to physical addresses that have addr_h<39> asserted are
not written to any of the caches. These write operations are merged in the
write buffer before being sent to the system. If software does not want write
operations to merge, it must insert MB or WMB instructions between them.
When the write buffer decides to write data to noncached memory, the BIU
requests a WRITE BLOCK. During each data cycle, int4_valid_h<3:0>
indicates which INT4s within the INT16 are valid.

4–14 Preliminary—Subject to Change—July 1996

4.4 Bcache Structure

4.4 Bcache Structure
The 21164 supports a 1M-byte, 2M-byte, . . . , 32M-byte and 64M-byte
Bcache. The size is under program control and is specified by BC_CONF<2:0>
(BC_SIZE<2:0>).
The Bcache block size may consist of 32-byte or 64-byte blocks. The Scache
also supports either 32-byte or 64-byte blocks. The block size must be the same
for both and is selected using SC_CTL<SC_BLK_SIZE>.
Industry-standard static RAMs (SRAMs) may be connected to the 21164
without many extra components although fanout buffers may be required for
the index lines. The SRAMs are directly controlled by the 21164, and the
Bcache data lines are connected to the 21164 data bus.
The 21164 partitions physical address (addr_h<39:5>) into an index field and
a tag field. The 21164 presents index_h<25:4> and tag_data_h<38:20> to the
Bcache interface. The tag size required is Bcache_size/block_size.
The system designer uses the signal lines needed for a particular size Bcache.
For example the smallest Bcache (1 MB) needs index_h<19:4> to address the
cache block while the tag field would be tag_data_h<38:20>.
Only those bits that are actually needed for the amount of cached system main
memory need to be stored in the Bcache tag, although the 21164 uses all the
relevant tag address bits for that Bcache size on its tag compare. A larger
Bcache uses more index bits and fewer tag address bits.
The CPU data bus is 16 bytes wide (128 bits) and thus each Bcache transaction
requires two data cycles for a 32-byte block or four data cycles for a 64-byte
block.

4.4.1 Duplicate Tag Store
In systems that have a Bcache, it is possible to build a full copy of the Bcache
tag store. This data can then be used to filter requests coming off the system
bus to the 21164.
In systems without a Bcache it is possible to build a full or partial copy of the
Scache tag store and to model the contents of the Scache victim buffers.

Preliminary—Subject to Change—July 1996 4–15

4.4 Bcache Structure

4.4.1.1 Full Duplicate Tag Store
The complete Bcache duplicate tag store would contain an entry for each
Bcache block and each victim buffer. Each entry would contain state bits
for the VALID, SHARED, and DIRTY status bits along with part or all of
addr_h<38:20> for a Bcache block. The part of addr_h<38:20> stored in an
entry depends upon the size of the Bcache.
In a system without a Bcache a full Scache duplicate tag store may be
maintained. The full Scache duplicate tag store should contain three sets
of 512 entries—one for each of the three Scache sets. It should also have
two entries for the two Scache victim buffers. Signal victim_pending_h is
used to indicate that the current READ command displaced a dirty block from
the Scache, scache_set_h<1:0>, into the Scache victim buffer. The Scache
duplicate tag store should be updated accordingly.
Figure 4–7 is a simplified diagram showing the signal lines of interest.
Figure 4–7 Full Scache Duplicate Tag Store

scache_set_h<1:0>
Set 0

Set 2

Set 1

addr_h<14:6>
(Index)

tag_shared_h,
tag_dirty_h,
tag_valid_h
addr_h<39:15>
(Tag Data)
Victim
Buffer 0

Victim
Buffer 1

victim_pending_h
LJ-04002.AI

4–16 Preliminary—Subject to Change—July 1996

4.4 Bcache Structure

The system should use the algorithm shown in Figure 4–8 to maintain the
duplicate tag store.
Figure 4–8 Duplicate Tag Store Algorithm

Init

21164
issues
command
?

Yes

Read
?
Yes
Push new entry into
duplicate tag store.

Victim
?
Yes
Put BUF0 into BUF1.
Put victim in BUF0.

LJ-04003.AI

Preliminary—Subject to Change—July 1996 4–17

4.4 Bcache Structure

4.4.1.2 Partial Scache Duplicate Tag Store
System designers may also choose to build a partial Scache duplicate tag store
such as that shown in Figure 4–9. This store contains one or more bits of
tag data for each block in the Scache, and for the two victim buffers inside
21164. If a system bus transaction hits in the partial duplicate tag store,
then the block may be in the Scache. If a system bus transaction misses in
the partial duplicate tag store, then the block is not in the Scache. Signal
victim_pending_h is used to indicate that the current READ command
displaced a dirty block from the Scache, scache_set_h<1:0>, into the Scache
victim buffer. The Scache duplicate tag store should be updated acordingly.
Figure 4–9 Partial Scache Duplicate Tag Store

scache_set_h<1:0>
Set 0

Set 2

Set 1

addr_h<14:6>
(Index)

addr_h<m:n>
(Part of <39:15> Tag Data)

Victim
Buffer
0

Victim
Buffer
1

victim_pending_h
LJ-04004.AI

4.4.2 Bcache Victim Buffers
A Bcache victim is generated when the 21164 deallocates a dirty block from
the Bcache. Each time a Bcache victim is produced, the 21164 asserts
victim_pending_h and stops reading the Bcache until the system takes
the current victim. Then Bcache transactions resume.
External logic may help improve system performance by implementing any
number of victim buffers that act as temporary storage that can be written
faster and with lower latency than system memory. The victim buffers hold
Bcache victims and enable the Bcache location to be filled with data from the
desired address. Data in the victim buffers will be written to memory at a
later time. This action reduces the time that the 21164 is waiting for data.

4–18 Preliminary—Subject to Change—July 1996

4.5 Systems Without a Bcache

4.5 Systems Without a Bcache
Systems that do not employ a Bcache should leave the bidirectional
signals tag_data_par_h, tag_dirty_h, tag_valid_h, tag_shared_h, and
tag_data_h<38:20> disconnected. Pull-down structures within the 21164
prevent these signals from attaining undefined logic levels.
In systems with no Bcache, the Scache block size must be set to 64 bytes.
In systems with no Bcache, signal idle_bc_h is not required and should be
permanently deasserted.

4.6 Cache Coherency
Cache coherency is a concern for single and multiprocessor 21164-based
systems as there may be several caches on a processor module and several
more in multiprocessor systems.
The system hardware designer need not be concerned about Icache and Dcache
coherency. Coherency of the Icache is a software concern—it is flushed with
an IMB (PALcode) instruction. The 21164 maintains coherency between the
Dcache and the Scache.
If the system does not have a Bcache, the system designer must create
mechanisms in the system interface logic to support cache coherency between
the Scache, main memory, and other caches in the system.
If the system has a Bcache, the 21164 maintains cache coherency between the
Scache and the Bcache. The Scache is a subset of the Bcache. In this case
the designer must create mechanisms in the system interface logic to support
cache coherency between the Bcache, main memory, and other caches in the
system.

4.6.1 Cache Coherency Basics
The 21164 systems maintain the cache coherency and hierarchy shown in
Figure 4–10.

Preliminary—Subject to Change—July 1996 4–19

4.6 Cache Coherency

Figure 4–10 Cache Subset Hierarchy

System

Main Memory

Bcache
(Optional)

Scache

Icache

Dcache

MK−1455−01

The following tasks must be performed to maintain cache coherency:
•

The Cbox in the 21164 maintains coherency in the Dcache and keeps it as
a subset of the Scache.

•

If an optional Bcache is present, then the 21164 maintains the Scache as a
subset of the Bcache. The Scache is set-associative but is kept a subset of
the larger externally implemented direct-mapped Bcache.

•

System logic must help the 21164 to keep the Bcache coherent with main
memory and other caches in the system.

•

The Icache is not a subset of any cache and also is not kept coherent with
the memory system.

The 21164 requires the system to allow only one change to a block at a time.
This means that if the 21164 gains the bus to read or write a block, no other
node on the bus should be allowed to access that block until the data has been
moved.

4–20 Preliminary—Subject to Change—July 1996

4.6 Cache Coherency

The 21164 provides hardware mechanisms to support several cache coherency
protocols. The protocols can be separated into two classes: write invalidate
cache coherency protocol and flush cache coherency protocol.
Write Invalidate Cache Coherency Protocol
The write invalidate cache coherency protocol is best suited for shared memory
multiprocessors.
The write invalidate protocol allows for shared data in the cache. If a Bcache
(optional) is used, then a duplicate tag store is required. If a Bcache is not
used, the duplicate tag store is not required but the module designer may
include an Scache duplicate tag store.
Requiring the duplicate tag store if there is a Bcache allows the 21164 to
process system commands in the Bcache without probing to see if the block
is present (system logic knows the block is present). This results in higher
performance for these transactions.
If a Bcache is not used, the module designer may include an Scache duplicate
tag store to improve system performance.
Flush Cache Coherency Protocol
This protocol is best suited for low-cost single-processor systems. It is typically
used by an I/O subsystem to ensure that data coherence is maintained when
DMA transactions are performed. Flush protocol does not allow shared data in
the cache.
Flush protocol does not require a duplicate tag store. Because the duplicate tag
store is optional for this protocol, the Bcache is probed for each transaction to
determine if the block is present. If the block is present, the requested action
is taken; if the block is not present, the command is still acknowledged, but no
other action is taken.
Section 4.6.2 and Section 4.6.3 describe the write invalidate cache coherency
protocol in more detail while Section 4.6.4 and Section 4.6.5 provide a more
detailed description of flush cache coherency protocol. The system commands
that are used to maintain cache coherency are described in more detail in
Section 4.10.

Preliminary—Subject to Change—July 1996 4–21

4.6 Cache Coherency

4.6.2 Write Invalidate Cache Coherency Protocol Systems
All 21164-based systems that implement the write invalidate cache protocol
must have the combinations of components listed in Table 4–5. For example,
a system such as that listed in write invalidate (3), having an Scache and
Bcache, is required to have a Bcache duplicate tag store and a lock register.
Table 4–5 Components for 21164 Write Invalidate Systems

Cache Protocol

Scache

Scache
Duplicate
Tag

Write invalidate (1)

Yes

Write invalidate (2)

Yes

Yes
(full or partial)

Required

Write invalidate (3)

Yes

Required (full)

Required

Bcache

Bcache
Duplicate
Tag

Lock
Register

Write Invalidate 1
This system has no external cache, duplicate tag store, or lock register. The
21164 must be made aware of all memory data transactions that occur on
the system bus. System logic uses an INVALIDATE, READ DIRTY, or READ
DIRTY/INVALIDATE transaction to the 21164 to maintain cache coherency
and to support the lock mechanism.
Write Invalidate 2
This system has an external Scache duplicate tag store and lock register.
System logic uses the duplicate Scache tag store and lock register to partially
or completely filter out unneeded transactions to the 21164. System logic
maintains the lock mechanism status and initiates transactions that affect
Scache coherency.
Write Invalidate 3
This system has an external Bcache, Bcache duplicate tag store and lock
register. An Scache duplicate tag store is not needed because the Scache is a
subset of the Bcache. This system operates similarly to the write invalidate
2 system, except that the cache is larger. Write invalidate systems with a
Bcache require a full Bcache duplicate tag store because the 21164 assumes
that a duplicate tag store has been used to completely filter out unneeded
transactions. Therefore, the 21164 does not probe the Bcache when system
commands are received, but assumes that they will hit in the Bcache.

4–22 Preliminary—Subject to Change—July 1996

4.6 Cache Coherency

4.6.3 Write Invalidate Cache Coherency States
Each processor in the system must be able to read and write data as if all
transactions were going onto the system bus to memory or I/O modules.
Therefore, the system bus is the point at which cache coherency must be
maintained.
Table 4–6 describes the Bcache states that determine cache coherency protocol
for 21164 systems.
Table 4–6 Bcache States for Cache Coherency Protocols
Valid1

Shared1

Dirty1

State of Cache Line

Not valid.

Valid for read or write operations. This cache line
contains the only cached copy of the block and the
copy in memory is identical to this line.

Valid for read or write operations. This cache line
contains the only cached copy of the block. The
contents of the block have been modified more recently
than the copy in memory.

Valid for read or write operations. This block may be
in another CPU’s cache.

Valid for read or write operations. This block may
be in another CPU’s cache. The contents of the block
have been modified more recently than the copy in
memory.

1 The tag_valid_h, tag_shared_h, and tag_dirty_h signals are described in Table 3–1.

Note
Unlike some other systems, the 21164 will not take an update to a
shared block but instead will invalidate the block.

Preliminary—Subject to Change—July 1996 4–23

4.6 Cache Coherency

4.6.3.1 Write Invalidate Protocol State Machines
Figure 4–11 shows the 21164 cache state transitions that can occur as a result
of 21164 transactions to the system. Figure 4–12 shows the 21164 cache state
transitions maintained by the 21164 as a result of transactions by other nodes
on the system bus. These two figures both represent the same state machine.
They show transitions caused by the 21164, and by the system, separately for
clarity.
Note
The abbreviations ‘‘I,S,D’’ indicate the INVALID, SHARED, and DIRTY
states.

Figure 4–11 Write Invalidate Protocol: 21164 State Transitions

READ (S)
(CPU Read Operation)

)
o
(S ati
D per
A
O
RE ad
Re
PU

I
READ MISS MOD
(CPU read for
write intent.)

SET DIRTY*
(CPU Write Operation)

WR
PU ITE
Wr BLO
WRITE BLOCK (S)
ite
Op CK (
(CPU Write Operation)
era S)
tio
n)
WRITE BLOCK** (S)
(CPU Write Operation)
(C

Write Block** (S)
(CPU Write Operation)

* Optionally this transition can be configured to occur without
a SET DIRTY command being issued externally.
** Only allowed in no_Bcache systems.

4–24 Preliminary—Subject to Change—July 1996

LJ-04036.AI

4.6 Cache Coherency

Figure 4–12 Write Invalidate Protocol: System/Bus State Transitions

SET SHARED
(Bus Read Operation)

)
(Bu INV
s W AL
rite IDA
Op TE
era
tion

AD
od (Bu DIR
ify s R T
O ea Y/I
pe d NV
ra /
tio
n)

)
INV ion
TY erat
DIR Op
AD dify
RE d/Mo
ea
sR

(Bu

E
AT tio
ID pera
L
VA O
IN rite
W
us

READ DIRTY
(Bus Read Operation)

READ DIRTY
(Bus Read Operation)
SD

LJ-04042.AI

4.6.4 Flush Cache Coherency Protocol Systems
All 21164-based systems that implement the flush cache protocol must have
the combinations of components listed in Table 4–7. For example, a system
such as that listed in flush (3), having a Bcache and a Bcache duplicate tag
store, is required to have a lock register.

Preliminary—Subject to Change—July 1996 4–25

4.6 Cache Coherency

Table 4–7 Components for 21164 Flush Cache Protocol Systems

Cache Protocol

Scache

Scache
Duplicate
Tag

Flush protocol (1)

Yes

Flush protocol (1.5)

Yes

Yes
(full or partial)

Required

Flush protocol (2)

Yes

Flush protocol (3)

Yes

Yes (partial/full)

Required

Bcache

Bcache
Duplicate
Tag

Lock
Register

Flush-Based 1
This system has no external cache, duplicate tag store, or lock register. System
logic notifies the 21164 of all memory data read operations that occur on the
system bus by using the interface READ command. The 21164 returns data if
the block is dirty.
System logic notifies the 21164 of all memory data write operations that
occur on the system bus by using the interface FLUSH command. The 21164
invalidates the block in cache, provides the data to the system if the block was
dirty, and updates the lock mechanism status.
Flush-Based 1.5
This system has no external cache, but does contain a partial or full duplicate
tag store for the Scache and the onchip Scache victim buffers. The SET_DIRTY
and LOCK commands should be enabled. The LOCK register is required.
System logic notifies the 21164 of all memory data read operations that hit in
the duplicate tag store by using the READ command. The 21164 provides the
system with a copy of the dirty data.
System logic notifies the 21164 of all memory data write operations that hit in
the duplicate tag store by using the FLUSH command. The 21164 provides the
dirty data and then invalidates the block.
Flush-Based 2
This system has an external cache but no duplicate tag store or lock register.
System logic and 21164 operation is identical to operation for the flush-based 1
system.

4–26 Preliminary—Subject to Change—July 1996

4.6 Cache Coherency

Flush-Based 3
This system has an external cache, a Bcache duplicate tag store, and lock
register. System logic notifies the 21164 of all memory data read operations
that occur on the system bus to addresses that are valid in the Bcache
duplicate tag store. System logic uses the READ command and the 21164
returns data if the block is dirty.
System logic uses the FLUSH command to notify the 21164 of all memory data
write transactions that occur on the system bus to addresses that are valid
in the Bcache duplicate tag store. If the block is dirty, the 21164 provides the
block data and invalidates the block in cache in any case.
System logic updates its lock mechanism status.
Flush-based systems with a Bcache do not require a full Bcache duplicate tag
because the 21164 always probes the Bcache in response to system commands.

4.6.5 Flush-Based Protocol State Machines
Figure 4–13 shows the 21164 cache state transitions that can occur as a result
of transactions with the system. Figure 4–14 shows the 21164 cache state
transitions maintained by the 21164 as a result of transactions by other nodes
on the system bus. These two figures both represent the same state machine.
They show transitions caused by the 21164, and by the system, separately for
clarity.
Note
The abbreviations ‘‘I,S,D’’ indicate the INVALID, SHARED, and DIRTY
states.

Preliminary—Subject to Change—July 1996 4–27

4.6 Cache Coherency

Figure 4–13 Flush-Based Protocol 21164 States

(C
SD

t
AD era
RE d Op
a
Re

RE
AD
ite (CP M
int U ISS
en rea M
to d O
pe for D
ra
tio
n.
)

SET DIRTY*
(CPU Write Operation)

*Optionally this transition can be configured to
occur without a SET DIRTY command being issued externally.
Refer to BC_CONTROL<EI_CMD_GRP2>.
LJ-04038.AI

Figure 4–14 Flush-Based Protocol System/Bus States

)
on
H rati d
S
U pe e
FL te O turn
ri Re m
W ta
te
MA Da ys
(D No to S

M
A F
Da Wr LU
ta ite SH
to Re Op
Sy tu er
ste rne ati
m d on)

READ
(DMA Read Operation)

READ
(DMA Read Operation)
LJ-04037.AI

4.6.6 Cache Coherency Transaction Conflicts
Cache coherency conflicts that can occur during system operation are described
here. Systems should be designed to avoid these conflicts.
4.6.6.1 Case 1
If the 21164 requests a READ MISS MOD transaction, it expects the block
to be returned SHARED; DIRTY. However, if the system returns the data
SHARED; DIRTY , the 21164 follows with a WRITE BLOCK command. This
might cause a multiprocessor system to have live-lock problems, a condition
that can cause long delays in writing from the 21164 to memory.
4–28 Preliminary—Subject to Change—July 1996

4.6 Cache Coherency

4.6.6.2 Case 2
If the 21164 attempts to write a clean/private block of memory, it sends a SET
DIRTY command to the system. The system could be sending a SET SHARED
or INVALIDATE command to the 21164 at the same time for the same block.
The bus is the coherence point in the system; therefore, if the bus has already
changed the state of the block to shared, setting the dirty bit is incorrect. The
21164 will not resend the SET DIRTY command when the ownership of the
ADDRESS/CMD bus is returned. The write will be restarted and will use the
new tag state to generate a new system request.
Another possibility is for the system to send an INVALIDATE instruction at
the same time the 21164 is attempting to do a WRITE BLOCK transaction to
the same block. In this case, the 21164 aborts the WRITE BLOCK transaction,
services the INVALIDATE instruction, then restarts the write transaction,
which produces a READ MISS command.
In both of these cases, if the SET DIRTY or WRITE BLOCK transaction is
started by the 21164 and then interrupted by the system, the 21164 resumes
the same transaction unless the system request was to the same block as the
request the 21164 had started. In this case, the 21164 request is restarted
internally by the CPU and it is UNPREDICTABLE what transaction the 21164
presents next to the system.

Preliminary—Subject to Change—July 1996 4–29

4.7 Lock Mechanisms

4.7 Lock Mechanisms
The LDx_L instruction is forced to miss in the Dcache. When the Scache is
read, the BIU’s lock IPR is loaded with the physical address and the lock flag
set. The BIU sends a LOCK command to the system so that it can load its
own lock register. The system lock register is used only if the locked block is
displaced from the cache system.
The lock flag is cleared if any of the following events occur:
•

Any write operation from the bus addresses the locked block (FLUSH,
INVALIDATE, or READ DIRTY/INV).

•

An STx_C is executed by the processor.

•

The locked block is refilled from memory and system_lock_flag_h is
cleared.

The system copy of the lock register is required on systems that have a
duplicate tag store to filter write traffic. The direct-mapped Icache, Dcache,
and Bcache; along with the subsetting rules, branch prediction, and Istream
prefetching, can cause a lock to always fail because of constant Scache
thrashing of the locked block. Each time a block is loaded into the Scache,
the value of the lock register is logically ANDed with the value of signal
system_lock_flag_h. If the locked block is displaced from the cache system,
the 21164 does not ‘‘see’’ bus write operations to the locked block. In this case,
the system’s copy of the lock register corrects the processor copy of the lock flag
when the block is filled into the cache, using signal system_lock_flag_h.
Systems that do not have duplicate tag stores, and send all probe traffic to the
21164, are not required to implement a lock register or lock flag. Such systems
should permanently assert signal system_lock_flag_h.
When the STx_C instruction is issued, the Ibox stops issuing memory-type
instructions. The store updates the Dcache in the usual way, and places itself
in the write buffer. It is not merged with other pending write operations. The
write buffer is flushed.
When the write buffer arrives at an STx_C instruction in cached memory, it
probes the Scache to check the block state. When the STx_C passes through
the Scache, an INVALIDATE command is sent to the Dcache. If the lock flag is
clear, the STx_C fails. If the block is SHARED; DIRTY, the write buffer writes
the STx_C data into the Scache. Success is written to the register file and the
Ibox begins issuing memory instructions again. If the block is in the shared
state, the BIU requests a WRITE BLOCK transaction. If the system CACKs
the WRITE BLOCK transaction, the Scache is written and the Ibox starts as
previously stated.

4–30 Preliminary—Subject to Change—July 1996

4.7 Lock Mechanisms

When the write buffer arrives at an STx_C instruction in noncached memory,
it probes the Scache to check the block state. The Scache misses, the state
of the lock flag is ignored, and the BIU requests a WRITE BLOCK LOCK
transaction. If the system CACKs the WRITE BLOCK LOCK transaction, the
Ibox starts as stated previously. If cfail_h is asserted along with cack_h, then
the STx_C fails.

4.8 Alpha 21164-to-Bcache Transactions
When initiating an Istream or Dstream data transaction, the 21164 first tries
the Icache or Dcache, respectively. If that access is unsuccessful, then the
Scache will be tried next. If that fails, then the 21164 tries the Bcache.
The 21164 interface to the system and Bcache is in the Cbox. The Cbox
provides address and control signals for transactions to and from the Bcache
and the system interface logic. The Cbox also transfers data across the 128-bit
bidirectional data bus.
The 21164 controls all Bcache transactions and will be able to process read and
write hits to the Bcache without assistance from the system. When system
logic writes to or reads from the Bcache, it transfers data to and from the
Bcache but only under the direct control of the 21164.
Note
Timing diagrams do not explicitly show tristated buses. For examples
of tristate timing, refer to Section 4.11.

4.8.1 Bcache Timing
Bcache cycle time may be faster than, identical to, or slower than, that of the
sysclk. If the system is involved in a Bcache transaction, each read or write
operation starts on a sysclk edge. It is the responsibility of the system to
control the rate of Bcache transactions by using the dack_h signal. Read and
write operations that are private to the 21164 and Bcache may start on any
CPU clock. There is no relation between sysclk and private Bcache accesses.
Bcache timing is configured using the BC_CONFIG and BC_CONTROL IPRs.
Section 5.3.5 and Section 5.3.4 show the layout of these registers. These
registers are normally configured by 21164 initialization code.
Bcache read timing and write timing are programmable. Read speed is selected
using BC_CONFIG<7:4> (BC_RD_SPD<3:0>). Write speed is selected using
BC_CONFIG<11:8> (BC_WR_SPD<3:0>).

Preliminary—Subject to Change—July 1996 4–31

4.8 Alpha 21164-to-Bcache Transactions

4.8.2 Bcache Read Transaction (Private Read Operation)
Figure 4–15 shows an example of the timing for a private read operation to
Bcache by the 21164. BC_CONFIG<BC_RD_SPD> (read speed) is set to 4 CPU
cycles, the minimum read time (maximum read speed).
Figure 4–15 Bcache Read Transaction

Arrows indicate when 21164
clocks Bcache data into the
pad ring.
CPU Clock Cycles

index_h<25:4>

data_h<127:0>

tag_ram_oe_h

data_ram_oe_h
LJ-04005.AI5

The index increments through four 16-byte addresses, each being asserted for
four CPU cycles. The Bcache logic waits BC_CONFIG<BC_RD_SPD<3:0>>
cycles before recieving the data.
The 21164 always delays one cycle before asserting the tag_ram_oe_h and
data_ram_oe_h lines. The lines are deasserted when the fourth index address
is deasserted.

4–32 Preliminary—Subject to Change—July 1996

4.8 Alpha 21164-to-Bcache Transactions

4.8.3 Wave Pipeline
The wave pipeline is implemented to improve performance for systems that use
64-byte block size. It is not supported for systems with 32-byte block size.
The wave pipeline is controlled using BC_CONFIG<7:4> (BC_RD_SPD<3:0>)
and BC_CTL<18:17> (BC_WAVE<1:0>).
BC_CONFIG<7:4> (BC_RD_SPD<3:0>) is set to the latency of the Bcache read
transaction. BC_CTL<18:17> (BC_WAVE<1:0>) is set to the number of cycles
to subtract from BC_RD_SPD to get the Bcache repetition rate.
For example, if BC_RD_SPD is set to 6 and BC_WAVE<1:0> is set to 2, it
takes 6 cycles for valid data to arrive at the pins, but a new read starts every
4 cycles.
The read repetition rate must be greater than 3. For example it is not
permitted to set BC_RD_SPD to 5 and BC_WAVE<1:0> to 2.
The example shown in Figure 4–16 has BC_RD_SPD=6, BC_WAVE<1:0>=2.
Figure 4–16 Wave Pipeline Timing Diagram

Arrows indicate when 21164
clocks Bcache data into the
pad ring.
CPU Clock Cycles
index_h<25:4>

data_h<127:0>

tag_ram_oe_h

data_ram_oe_h
LJ-04034.AI5

Preliminary—Subject to Change—July 1996 4–33

4.8 Alpha 21164-to-Bcache Transactions

4.8.4 Bcache Write Transaction (Private Write Operation)
Figure 4–17 shows an example of the timing for a private write operation to
Bcache by the 21164. BC_CONFIG<BC_WR_SPD> (write speed) is set to 4
CPU cycles, the minimum time.
Figure 4–17 Bcache Write Transaction

CPU Clock Cycles

index_h<25.4>

data_h<127:0>

data_ram_we_h

20 21 22 23

tag_ram-we_h

20 21 22 23

LJ-04007.AI

The index increments through four 16-byte addresses, each being asserted for
four cycles. The 21164 always delays one cycle then drives the data associated
with each index.
Signals tag_ram_we_h and data_ram_we_h are asserted high for two
cycles because the BC_CONFIG<28:20> (BC_WE_CTL<8:0>) is set to 6.
BC_CONFIG<22:21> being set causes the write-enable lines to be asserted
during the second and third CPU cycles. BC_CONFIG<20,23> being clear
causes the write-enable lines to not be asserted during the first and fourth
CPU cycles.
The Bcache maximum read or write time is 15 cycles. The minimum
read or write time is 4 cycles; except that in 32-byte mode, the minimum
read time is 5 cycles. So the index and data can be asserted from 4 to
15 cycles. The write-enable signals can be asserted from 0 to 9 cycles. If
BC_CONFIG(BC_WE_CTL) is set to 0, the write-enable signals will not be
asserted. If the 9-bit field is set to 1FF16 , then the write-enable signals will be
asserted for 9 CPU cycles.

4–34 Preliminary—Subject to Change—July 1996

4.8 Alpha 21164-to-Bcache Transactions

4.8.5 Selecting Bcache Options
Table 4–8 lists the variables to consider when designing and implementing a
Bcache.
Table 4–8 Bcache Options
Parameter

Selection

Sysclk ratio (3–15)

CPU cycles

Cache protocol, write invalidate or flush
Cache block size 64/32

–byte block

ECC or byte parity
Bcache present?
Bcache size (1M byte to 64M bytes)

M byte

Bcache read speed (4–15)

CPU cycles

Bcache wave pipelining (0–3)

CPU cycles

Bcache victim buffer?
Bcache write speed (4–15)
Bcache read to write spacing (1–7)
Bcache fill write pulse offset (1–7)
Bcache write pulse (bit mask 9–0)
Enable LOCK and SET DIRTY
commands?
Enable memory barrier (MB) commands?

Preliminary—Subject to Change—July 1996 4–35

4.9 Alpha 21164-Initiated System Transactions

4.9 Alpha 21164-Initiated System Transactions
This section describes how commands are used to move data between the
21164 and its cache system.
Note
Timing diagrams do not explicitly show tristated buses. For examples
of tristate timing, refer to Section 4.11.

The 21164 starts an external transaction when:
•

It encounters a ‘‘miss.’’

•

A LOCK command is invoked.

•

A WRITE command is directed at a shared block.

•

A WRITE command is directed at a clean block in Scache.

•

The CPU addresses a noncached region of memory.

•

The 21164 executes a FETCH, FETCH_M, or MB instruction.

For example, the sequence for a 21164-initiated transaction caused by a Bcache
miss is:
•

At the start of a Bcache transaction, the 21164 checks the tag and tag
control status of the target block.

•

If there is a tag mismatch or the Valid bit is clear, a Bcache miss has
occurred and the 21164 starts an external READ MISS transaction that
tells the system logic to access and return data.

•

System logic acknowledges acceptance of the command from the 21164 by
asserting cack_h.

•

Because the transaction is a read operation, requiring a FILL transaction,
the transaction is broken (pended) while system logic obtains the FILL
data.

•

At a later time, the system asserts fill_h.

•

The 21164 will assert the tag and tag control bits, and will control the
write action during the FILL transaction.

•

The system logic provides the data. As each of the two (or four) data cycles
becomes valid, the system logic asserts dack_h to cause the 21164 to
sample to data and write it into the Bcache.

4–36 Preliminary—Subject to Change—July 1996

4.9 Alpha 21164-Initiated System Transactions

Interface commands from the 21164 to the system are driven on the
cmd_h<3:0> signals. Table 4–9 lists and describes the set of interface
commands.
Table 4–9

Alpha 21164-Initiated Interface Commands

Command

cmd_h
<3:0>

NOP

0000

The NOP command is driven by the owner of the
cmd_h bus when it has no tasks queued.

LOCK

0001

The LOCK command is used to load the system lock
register with a new lock register address. The state
of the system lock register flag is used on each fill
to update the 21164’s copy of the lock flag. Refer to
Section 4.7 for more information.

FETCH

0010

The 21164 passes a FETCH instruction to the
system when the FETCH instruction is executed.

FETCH_M

0011

The 21164 passes a FETCH_M instruction to the
system when the FETCH_M instruction is executed.

MEMORY
BARRIER

0100

The 21164 issues the MEMORY BARRIER command
when an MB instruction is executed. This command
should be used to synchronize read and write
accesses with other processors in the system. The
21164 stops issuing memory reference instructions
and waits for the command to be acknowledged
before continuing.

SET DIRTY

0101

Dirty bit set if shared bit is clear. The 21164 uses
the SET DIRTY command when it wants to write
a clean, private block in its Scache and it wants
the dirty bit set in the duplicate tag store. The
21164 does not proceed with the write until a CACK
response is received from the system. When the
CACK is received, the 21164 attempts to set the
dirty bit. If the shared bit is still clear, the dirty
bit is set and the write operation is completed. If
the shared bit is set, the dirty bit is not set and the
21164 requests a WRITE BLOCK transaction. The
copy of the dirty bit in the Bcache is not updated
until the block is removed from the Scache.

Description

(continued on next page)

Preliminary—Subject to Change—July 1996 4–37

4.9 Alpha 21164-Initiated System Transactions

Table 4–9 (Cont.)

Alpha 21164-Initiated Interface Commands

Command

cmd_h
<3:0>

WRITE BLOCK

0110

Request to write a block. When the 21164 wants
to write a block of data back to memory, it drives
the command, address, and first INT16 of data on
a sysclk edge. The 21164 outputs the next INT16
of data when dack_h is received. When the system
asserts cack_h, the 21164 removes the command
and address from the bus and begins the write of
the Scache. Signal cack_h can be asserted before
all the data is removed.

WRITE BLOCK
LOCK

0111

Request to write a block with lock. This command
is identical to a WRITE BLOCK command except
that the cfail_h signal may be asserted by the
system, indicating that the data cannot be written.
This command is only used for STx_C in noncached
space.

READ MISS0

1000

Request for data. This command indicates that the
21164 has probed its caches and that the addressed
block is not present.

READ MISS1

1001

Request for data. This command indicates that the
21164 has probed its caches and that the addressed
block is not present.

READ MISS
MOD0

1010

Request for data; modify intent. This command
indicates that the 21164 plans to write to the
returned cache block. Normally, the dirty bit should
be set when the tag status is returned to the 21164
on a Bcache fill.

READ MISS
MOD1

1011

Description

(continued on next page)

4–38 Preliminary—Subject to Change—July 1996

4.9 Alpha 21164-Initiated System Transactions

Table 4–9 (Cont.)

Alpha 21164-Initiated Interface Commands

Command

cmd_h
<3:0>

BCACHE
VICTIM

1100

Description
Bcache victim should be removed. If there is a
victim buffer in the system, this command is used
to pass the address of the victim to the system. The
READ MISS command that produced the victim
precedes the BCACHE VICTIM command. Signal
victim_pending_h is asserted during the READ
MISS command to indicate that a BCACHE VICTIM
command is waiting, and that the Bcache is starting
the read of the victim data.
If the system does not have a victim buffer, the
BCACHE VICTIM command precedes the READ
MISS commands. The BCACHE VICTIM command
is driven, along with the address of the victim. At
the same time, the Bcache is read to provide the
victim data.
If the system does have a victim buffer, and it
asserts signal dack_h any time before the BCACHE
VICTIM command is driven, then address bits
addr_h<5:4> of the address sent with the BCACHE
VICTIM command are UNPREDICTABLE.
The system must use the values of addr_h<5:4>
that were sent with the READ MISS command that
produced the victim.

—

1101

Spare.

READ MISS
MOD STC0

1110

Request for data, STx_C data.

READ MISS
MOD STC1

1111

Request for data, STx_C data.

Preliminary—Subject to Change—July 1996 4–39

4.9 Alpha 21164-Initiated System Transactions

4.9.1 READ MISS—No Bcache
A read operation to the Dcache misses causing a read operation to the Scache,
which also misses. After the Scache miss there is no Bcache probe—the 21164
sends a READ MISS command to the system. The system acknowledges
receipt of the READ MISS by asserting cack_h as shown in Figure 4–18.
Figure 4–18 READ MISS—No Bcache Timing Diagram

sys_clk_out
RM1

RM0

RM1

RM0

cmd_h<3.0>
addr_h<394>

FFF0

cack_h
fill_h
fill_id_h
data_h<127:0>

dack_h
LJ-04008.AI

4–40 Preliminary—Subject to Change—July 1996

4.9 Alpha 21164-Initiated System Transactions

4.9.2 READ MISS—Bcache
The 21164 starts a Bcache read operation on any CPU clock. The index is
asserted to the RAM for a programmable number of CPU cycles in the range
of 4 to 15. The tag is accessed at the same time. At the end of the first read
operation, the 21164 latches the data and tag information and begins the read
operation of the next 16 bytes of data. The tag is checked for a hit. If there is
a miss, a READ MISS or READ MISS MOD command, along with the address,
is queued to the cmd_h<3:0> bus. It appears on the interface at the next
sysclk edge.
Figure 4–19 shows the timing of a Bcache read and the resulting READ MISS
MOD request. The system immediately asserts cack_h to acknowledge the
command. This allows the 21164 to make additional READ MISS requests. It
is also possible for the system to defer assertion of cack_h until the fill data is
returned. This allows the system to use cmd_h<0> for the value of fill_id_h.
The assertion of cack_h should arrive no later than the last fill dack_h.
The only difference between a READ MISS and a READ MISS MOD sequence
on the bus is that tag_dirty_h should be asserted during the Bcache fill
associated with a READ MISS MOD.
Note
A READ MISS command with int4_valid_h<3:0> of zero is a request
for Istream data while int4_valid_h<3:0> of non-zero is a request for
Dstream data.

Preliminary—Subject to Change—July 1996 4–41

4.9 Alpha 21164-Initiated System Transactions

Figure 4–19 READ MISS MOD—Bcache Timing Diagram

sys_clk_out1_h
addr_bus_req_h
cmd_h<3:0>

RMM1

RMM0

victim_pending_h
addr_h<39:4>

9900

FFF0

5CE0

cack_h
addr_res_h<2:0>
fill_h
fill_id_h
idle_bc_h
index_h<25:4> 6000

9900

5CE0

9910

5CF0

5CE0

data_h<127:0>

9900

9910

9920

9930

5CE0

dack_h
data_ram_oe_h
data_ram_we_h
tag_ram_oe_h
tag_ram_we_h
tag_data_h<38:20>
tag_dirty_h
tag_shared_h
tag_valid_h
LJ-04009.AI5

4–42 Preliminary—Subject to Change—July 1996

4.9 Alpha 21164-Initiated System Transactions

4.9.3 FILL
Signals fill_h, fill_id_h, and fill_error_h are used to control the return of
fill data to the 21164 and the Bcache, if it is present. Signal idle_bc_h must
be used to stop CPU requests in the Bcache in such a way that the Bcache
will be idle when the fill data arrives (but not the FILL command). Signal
fill_h should be asserted at least two sysclk periods before the fill data arrives.
Signal fill_id_h should be asserted at the same time to indicate whether the
FILL is for a READ MISS0 or READ MISS1 operation. The 21164 uses this
information to select the correct fill address. Figure 4–19 shows the timing of
a FILL command. Refer also to Section 4.11.3 for more information on using
signals idle_bc_h and fill_h.
If signals fill_h and fill_id_h are asserted at the rising edge of sysclk N, then
at the rising edge of sysclk N+1, the 21164 tristates data_h<127:0>, asserts
the Bcache index, and begins a Bcache write operation. The system should
drive the data onto the data bus and assert dack_h before the end of the
sysclk cycle. At the end of the write time, the 21164 waits for the next sysclk
edge. If dack_h has not been asserted, the Bcache write operation starts again
at the same index. If dack_h is asserted, the index advances to the next part
of the fill and the write operation begins again. The system must provide
the data and dack_h signal at the correct sysclk edges to complete the fill
correctly. For example, if the Bcache requires 17 ns to write, and the sysclk is
12 ns, then two sysclk cycles are required for each write operation.
The 21164 calculates and asserts tag_valid_h and writes the Bcache tag
store with each INT16 of data. The system is required to drive signals
tag_shared_h, tag_dirty_h, and tag_ctl_par_h with the correct value for the
entire FILL transaction.
At the end of the FILL transaction, the 21164 will not assert data_ram_oe_h
or begin to drive the data bus until the fifth CPU cycle after the sysclk that
loads the last DACK. If systems require more time to turn off their drivers,
they must use idle_bc_h in combination with data_bus_req_h to stop 21164
requests, and not send any system requests.

4.9.4 READ MISS with Victim
The 21164 supports two models for removing displaced dirty blocks from the
Bcache. The first assumes that the system does not contain a victim buffer. In
this case, the victim must be read from the Bcache before the new block can be
requested. In the second case, where the system has a victim buffer, the 21164
requests the new block from memory while it starts to read the victim from the
Bcache. The VICTIM command and address follows the miss request.

Preliminary—Subject to Change—July 1996 4–43

4.9 Alpha 21164-Initiated System Transactions

In either case, the 21164 treats a miss/victim as a single transaction. If the
assertion of addr_bus_req_h or idle_bc_h causes the BIU sequencer to reset,
both the READ MISS and BCACHE VICTIM transactions are restarted from
the beginning. For example, if the 21164 is operating in victim first mode, and
it sends a BCACHE VICTIM command to the system, then the system sends
an INVALIDATE request to the 21164. The 21164 processes the INVALIDATE
request and then restarts the READ operation and resends the BCACHE
VICTIM command and data, and then processes the READ MISS.
Sections 4.9.4.1 and 4.9.4.2 describe each of these methods of victim processing.
4.9.4.1 READ MISS with Victim (Victim Buffer)
When the miss is detected, if the system has a victim buffer, the 21164 waits
for the next sysclk, then asserts a READ MISS command, the read miss
address, the victim_pending_h signal, and indexes the Bcache to begin the
read operation of the victim. When the system asserts cack_h, the 21164
sends out a NOP command along with the victim address. In the following
cycle the BCACHE_VICTIM command is driven. Each assertion of dack_h
causes the Bcache index to advance to the next part of the block. Figure 4–20
shows the timing of a READ MISS command with a victim.

4–44 Preliminary—Subject to Change—July 1996

4.9 Alpha 21164-Initiated System Transactions

Figure 4–20 READ MISS with Victim (Victim Buffer) Timing Diagram

sys_clk_out1_h
addr_bus_req_h
cmd_h<3:0>

RM0

Bcache Victim

5FA0

4005F80

victim_pending_h
addr_h<39:4>

FFF0

5FA0

cack_h
addr_res_h<2:0>
fill_h
fill_id_h
idle_bc_h
index_h<25:4>
data_h<127:0>

9B60

5FB0

5FA0
D0

5FB0
D1

5F80

5F90
D2

5FA0

5FB0

dack_h
data_ram_oe_h
data_ram_we_h
tag_ram_oe_h
tag_ram_we_h
tag_data_h<38:20>
tag_dirty_h
tag_shared_h
tag_valid_h
LJ-04010.AI5

Preliminary—Subject to Change—July 1996 4–45

4.9 Alpha 21164-Initiated System Transactions

4.9.4.2 READ MISS with Victim (Without Victim Buffer)
If the system does not contain a victim buffer, the 21164 stops reading the
Bcache as soon as the miss is detected. This occurs while the second INT16
data is on data_h<127:0>, as shown in Figure 4–21.
A BCACHE VICTIM command is asserted at the next sysclk along with the
victim address. A Bcache read operation of the victim is also started at the
sysclk edge.
When dack_h is received for the first INT16 of the victim, the 21164 begins
reading the next INT16 of the victim. cack_h can be sent any time before the
last dack_h is asserted or with the last dack_h assertion.
The 21164 sends the READ MISS command after the last dack_h is received.
Figure 4–21 shows the timing of a victim being removed.
Notice the data wrap sequence of this transaction—D2, D3, D0, and D1.

4–46 Preliminary—Subject to Change—July 1996

4.9 Alpha 21164-Initiated System Transactions

Figure 4–21 READ MISS with Victim (without Victim Buffer) Timing Diagram

sys_clk_out1_h
addr_bus_req_h
cmd_h<3:0>

Bcache Victim

RM0

victim_pending_h
addr_h<39:4>

5FB0

4005FA0

5F80

5FA0

9900

cack_h
addr_res_h<2:0>
fill_h
fill_id_h
idle_bc_h
index_h<25:4>
data_h<127:0>

B8C0

5FB0

5FA0

5FB0
D2

5F80
D3

5F90
D0

5FA0

dack_h
data_ram_oe_h
data_ram_we_h
tag_ram_oe_h
tag_ram_we_h
tag_data_h<38:20>
tag_dirty_h
tag_shared_h
tag_valid_h
LJ-04011.AI

Preliminary—Subject to Change—July 1996 4–47

4.9 Alpha 21164-Initiated System Transactions

4.9.5 WRITE BLOCK and WRITE BLOCK LOCK
The WRITE BLOCK command is used to complete write operations to shared
data, to remove Scache victims in systems without a Bcache, and to complete
write operations to noncached memory.
The WRITE BLOCK LOCK command follows the same protocol. The LOCK
qualifier allows the system to be more ‘‘conservative’’ on interlocked write
operations to noncached memory space. Refer to Section 4.7 for more
information on lock mechanisms.
The WRITE BLOCK command to cached memory regions that source data from
the Scache sends data to the system and also causes the data to be written in
the Bcache.
The 21164 asserts the WRITE BLOCK command, along with the address
and the first 16 bytes of data, at the start of a sysclk. If the system removes
ownership of the cmd_h<3:0> bus, the 21164 retains the WRITE command and
waits for bus ownership to be returned. If the block in question is invalidated,
the 21164 restarts the write operation. This results in the READ MISS MOD
request instead.
When the system takes the first part of the data, it asserts dack_h. This
causes the 21164 to drive the next 16 bytes of data on the same sysclk edge.
If the system asserts cack_h, the 21164 outputs the next command in the
next sysclk. Receipt of signal cack_h indicates to the 21164 that the write
operation will be taken, and that it is safe to update the Scache with the new
version of the block.
During each cycle, the int4_valid_h<3:0> signals indicate which INT4 parts
of the write operation are really being written by the processor. For write
operations to cached memory, all of the data is valid. For write operations to
noncached memory, only those INT4 with the int4_valid_h<n> signal asserted
are valid. See the definition for int4_valid_h<n> in Table 3–1.
Figure 4–22 shows the timing of a WRITE BLOCK command.

4–48 Preliminary—Subject to Change—July 1996

4.9 Alpha 21164-Initiated System Transactions

Figure 4–22 WRITE BLOCK Timing Diagram

sys_clk_out1_h
addr_bus_req_h
cmd_h<3:0>

WRITE BLOCK

WRITE BLOCK LOCK
7

victim_pending_h
FFF0000040

addr_h<39:4>

4700

0180

cack_h
addr_res_h<2:0>

fill_h
idle_bc_h
index_h<25:4>

5100

data_h<127:0>

0180

0190

01A0

01B0

0180

0040

0050

FFFF

dack_h
data_ram_oe_h
data_ram_we_h
tag_ram_oe_h
tag_ram_we_h
tag_data_h<38:20>

0000

0030

tag_dirty_h
tag_shared_h
tag_valid_h
LJ-04012.AI

Preliminary—Subject to Change—July 1996 4–49

4.9 Alpha 21164-Initiated System Transactions

4.9.6 SET DIRTY and LOCK
Figure 4–23 shows the timing of a SET DIRTY and a LOCK operation.
The 21164 uses the SET DIRTY transaction to inform a duplicate tag store
that a cached block is changing from the SHARED, DIRTY state to the
SHARED; DIRTY state. When cack_h is received from the system, the 21164
sets the dirty bit. If a SET SHARED or INVALIDATE command is received
for the same block, the 21164 responds with a WRITE BLOCK or READ MISS
MOD command.
The SET DIRTY and LOCK commands must be enabled in any system that
contains a duplicate tag store. The 21164 uses the SET DIRTY command to
update the dirty bit in the duplicate tag store.
The 21164 uses the LOCK command to pass the address of a LDx_L to the
system. A system lock register is required in any system that filters write
traffic with a duplicate tag store. If the locked block is displaced from the
21164 caches, the 21164 uses the value of the system lock register to determine
if the LDx_L/STx_C sequence should pass or fail.
The system may use BC_CONTROL<EI_CMD_GRP2> to modify operation for
these commands.
•

If BC_CONTROL<EI_CMD_GRP2> is set, the 21164 is allowed to issue
SET DIRTY and LOCK commands to the system interface. The system
logic acknowledges receipt of these commands.

•

If BC_CONTROL<EI_CMD_GRP2> is clear, the SET_DIRTY command
will never be driven by the 21164. It is UNPREDICTABLE if the LOCK
command is driven. However, the system should never assert cack_h for
the command when BC_CONTROL<EI_CMD_GRP2> is clear.

4–50 Preliminary—Subject to Change—July 1996

4.9 Alpha 21164-Initiated System Transactions

Figure 4–23 SET DIRTY and LOCK Timing Diagram

sys_clk_out1_h
addr_bus_req_h
cmd_h<3:0>

LOCK

FETCH
0

MB
0

SET_DIRTY
0

victim_pending_h
addr_h<39:4>

cack_h

V = Valid
LJ-04013.AI5

Preliminary—Subject to Change—July 1996 4–51

4.9 Alpha 21164-Initiated System Transactions

4.9.7 Memory Barrier (MB)
The 21164 may encounter a memory barrier (MB) instruction when executing
the instruction stream. The action taken by the 21164 depends upon the state
of BC_CONTROL<3> (EI_CMD_GRP3).
•

If BC_CONTROL<EI_CMD_GRP3> is set, the 21164 drains its pipeline and
buffers, then issues an MB command to the system interface. The system
logic must empty its buffers and complete all pending transactions before
acknowledging receipt for the MB command by asserting cack_h.

•

If BC_CONTROL<EI_CMD_GRP3> is clear, the 21164 never drives a MB
command to the interface command pins.
Note
The address presented on addr_h<39:4> during a MB transaction is
UNPREDICTABLE.

4.9.7.1 When to Use a MEMORY BARRIER Command
If the system interface buffers invalidate between the duplicate tag store and
the 21164, then the system interface must enable the MB command and drain
all invalidates before asserting cack_h in response to an MB command.

4.9.8 FETCH
The 21164 passes a FETCH command to the system when it executes a FETCH
instruction. The system responds to the command by asserting cack_h. This
command acts as a ‘‘hint’’ to the system. The system may respond with
optional behavior as a result of this hint (refer to the Alpha Architecture
Reference Manual).

4.9.9 FETCH_M
The 21164 passes a FETCH_M (fetch with modify intent) command to the
system when it executes a FETCH_M instruction.

4–52 Preliminary—Subject to Change—July 1996

4.10 System-Initiated Transactions

4.10 System-Initiated Transactions
System commands to the 21164, are driven on the cmd_h<3:0> signal lines.
Before driving these signals, the system must gain control of the command and
address buses by using addr_bus_req_h, as described in Section 4.11.1. The
algorithm used by the 21164 for accepting system commands to be processed in
parallel by the 21164 is presented in Section 4.10.1.
System-initiated commands may be separated into two protocol groups. The
group of commands used by write invalidate protocol systems is listed and
described in Section 4.10.2. The group of commands used by flush-based
protocol systems is listed and described in Section 4.10.3.
Note
Timing diagrams do not explicitly show tristated buses. For examples
of tristate timing, refer to Section 4.11.

4.10.1 Sending Commands to the 21164
The rules used by the Cbox BIU to process commands sent by the system to
the 21164 are listed in Section 4.13.1.
The 21164 can hold two outstanding commands from the system at any time.
The algorithm used by the system to send commands to the 21164 without
overflowing the two Cbox BIU command buffers is shown in Figure 4–24.

Preliminary—Subject to Change—July 1996 4–53

4.10 System-Initiated Transactions

Figure 4–24 Algorithm for System Sending Commands to the 21164

Start

Yes

Init
?

Set count to zero.

Is
(CMD Not NOP)
and
(Count=2)
?

Yes

Send command.
Increment count.

Is
CPU response
equal to
ACK/Scache
?
No

Yes

Receive data.

Yes

Decrement count.

CPU
response equals
(ACK/Bcache
or NACK)
?

Yes

CMD is
READ or FLUSH
or READ DIRTY INV
or READ DIRTY
?
No

Decrement count.

No
LJ-04014.AI

4–54 Preliminary—Subject to Change—July 1996

4.10 System-Initiated Transactions

4.10.2 Write Invalidate Protocol Commands
All 21164-based systems that use the write invalidate protocol are expected
to use the READ DIRTY, READ DIRTY/INVALIDATE, INVALIDATE, and
SET SHARED commands to keep the state of each block up to date. These
commands are defined in Table 4–10.
Table 4–10 System-Initiated Interface Commands (Write Invalidate Protocol)
Command

cmd_h
<3:0>

NOP

0000

The NOP command is driven by the owner of the
cmd_h<3:0> bus when it has no tasks queued.

INVALIDATE

0010

Remove the block. When the system issues the
INVALIDATE command, the 21164 probes its
Scache. If the block is found, the 21164 responds
with ACK/Scache and invalidates the block. If the
block is not found, and the system does not contain
a Bcache, the 21164 responds with a NOACK.

Description

If the system contains a Bcache, the system is
assumed to have filtered all requests by using the
duplicate tag store. Therefore, the block is assumed
to be present in the Bcache. The 21164 responds
with ACK/Bcache, and the block is changed to the
invalid state without probing.
SET SHARED

0011

Block goes to the shared state. The SET SHARED
command is used by the system to change the state
of a block in the cache system to shared. The shared
bit in the Scache is set if the block is present.
The Bcache tag is written to the shared not dirty
state. The 21164 assumes that this action is correct,
because the system would have sent a READ DIRTY
command if the dirty bit were set.
If the block is found in the Scache, the 21164
responds with ACK/Scache. Otherwise, if the system
contains a Bcache, the block is assumed to be in the
Bcache, and the 21164 responds with ACK/Bcache.
If the system does not contain a Bcache, and the
block is not found in the Scache, the 21164 responds
with NOACK.
(continued on next page)

Preliminary—Subject to Change—July 1996 4–55

4.10 System-Initiated Transactions

Table 4–10 (Cont.) System-Initiated Interface Commands (Write Invalidate
Protocol)

4.10.2.1

Command

cmd_h
<3:0>

READ DIRTY

0101

Read a block; set shared. The READ DIRTY
command probes the Scache to see if the requested
block is present and dirty. If the block is not found,
or if the block is clean, and the system does not
contain a Bcache, the 21164 responds with NOACK.
If the block is found and dirty in the Scache, the
21164 responds with ACK/Scache and drives the
data on the data_h<127:0> bus. If the block is
not found in the Scache, and the system contains a
Bcache, the block is assumed to be in the Bcache.
The 21164 responds with ACK/Bcache, indexes the
Bcache to read the block, and changes the block
status to the shared dirty state.

READ DIRTY/
INVALIDATE

0111

Read a block; invalidate. This command is identical
to the READ DIRTY command except that if the
block is present in the caches, it will be invalidated
from the caches.

Description

Alpha 21164 Responses to Write Invalidate Protocol Commands
The 21164 responses on addr_res_h<1:0> to write invalidate protocol
commands are listed in Table 4–11.

4–56 Preliminary—Subject to Change—July 1996

4.10 System-Initiated Transactions

Table 4–11

Alpha 21164 Responses on addr_res_h<1:0> to Write Invalidate
Protocol Commands

Bcache

Scache

addr_res_h<1:0>

INVALIDATE and SET SHARED Commands
No Bcache

Scache_Miss

NOACK

No Bcache

Scache_Hit

ACK/Scache

Bcache_Hit/Miss

Scache_Hit/Miss

ACK/Bcache

READ DIRTY and READ DIRTY/INVALIDATE Commands
No Bcache

Scache_Miss

NOACK

No Bcache

Scache_Hit,Not Dirty

NOACK

No Bcache

Scache_Hit,Dirty

ACK/Scache

Bcache

Scache_Hit,Dirty

ACK/Scache

Bcache

Scache_Miss

ACK/Bcache

The signal addr_res_h<2> allows a system without a duplicate tag store to
determine if a block is present in the Scache or lock register. The system logic
can use this information to correctly assert tag_shared_h in a multiprocessor
system.
The 21164 responds to the READ, FLUSH, READ DIRTY, SET SHARED
and READ DIRTY/INVALIDATE commands on addr_res_h<2>, as listed in
Table 4–12.
Table 4–12

Alpha 21164 Responses on addr_res_h<2> to 21164 Commands

Scache

Lock Register

addr_res_h<2>

Miss

Hit

Miss

Hit

Table 4–13 presents the 21164 best-case response time to system commands in
a write invalidate protocol system.

Preliminary—Subject to Change—July 1996 4–57

4.10 System-Initiated Transactions

Table 4–13

Alpha 21164 Minimum Response Time to Write Invalidate
Protocol Commands

Cache Status

Response

Number of sys_clk_out1_h,l Cycles

No Bcache

NOACK

8 CPU cycles rounded up to next
sys_clk_out1_h,l cycles

No Bcache

ACK/Scache

12 CPU cycles rounded up to next
sys_clk_out1_h,l cycles

Bcache

NOACK, ACK/Scache,
ACK/Bcache

10 CPU cycles rounded up to next
sys_clk_out1_h,l cycles

4.10.2.2 READ DIRTY and READ DIRTY/INVALIDATE
The READ DIRTY command is used to read modified data from the cache
system. The block status changes from DIRTY; SHARED to DIRTY; SHARED.
Figure 4–25 shows the timing of a READ DIRTY command that hits in the
Scache. The 21164 drives data starting at the rising edge of the sysclk that
drives addr_res_h<2:0>. The Bcache data and tag state are updated as each
INT16 is passed to the system. If the data had not been found in the Scache,
the Bcache would have been indexed on the rising edge of the syclk that
drove addr_res_h<2:0>. The index would advance to the next INT16 data as
dack_h pulses arrive. The Bcache tag would be written with the updated state
during the second INT16 data cycle.
The READ DIRTY/INVALIDATE command is identical to the READ DIRTY
command except that the block is changed to VALID rather than to SHARED.

4–58 Preliminary—Subject to Change—July 1996

4.10 System-Initiated Transactions

Figure 4–25 READ DIRTY Timing Diagram (Scache Hit)

sys_clk_out1_h
addr_bus_req_h
cmd_h<3:0>

READ DIRTY

victim_pending_h
addr_h<39:4>

0000

cack_h
ACK/Scache

addr_res_h<2:0>

NOP

idle_bc_h
index_h<25:4>
data_h<127:0>

000

001

002

003

000

dack_h
data_ram_oe_h
data_ram_we_h
tag_ram_oe_h
tag_ram_we_h
tag_data_h<38:20>
tag_dirty_h
tag_shared_h
tag_valid_h
LJ-04015.AI

Preliminary—Subject to Change—July 1996 4–59

4.10 System-Initiated Transactions

4.10.2.3 INVALIDATE
The INVALIDATE command can be used to remove a block from the cache
system. Unlike the FLUSH command, any modified data will not be read.
The Scache is probed and invalidated if the block is found. The Bcache is
invalidated without probing. Figure 4–26 shows the timing of an INVALIDATE
transaction.

4–60 Preliminary—Subject to Change—July 1996

4.10 System-Initiated Transactions

Figure 4–26 INVALIDATE Timing Diagram—Bcache Hit

sys_clk_out1_h
addr_bus_req_h
cmd_h<3:0>

INVALIDATE

victim_pending_h
addr_h<39:4>

00D0

cack_h
ACK/Bcache

addr_res_h<2:0>
idle_bc_h
index_h<25:4>

0000

data_h<127:0>
dack_h
data_ram_oe_h
data_ram_we_h
tag_ram_oe_h
tag_ram_we_h
tag_data_h<38:20>

FC00

tag_dirty_h
tag_shared_h
tag_valid_h
LJ-04016.AI

Preliminary—Subject to Change—July 1996 4–61

4.10 System-Initiated Transactions

4.10.2.4 SET SHARED
When the 21164 receives a SET SHARED command, it probes the Scache
and changes the state of the block to SHARED if it is found. The 21164
‘‘assumes’’ that the block is in the Bcache and writes the state of the tag to
SHARED; DIRTY . Figure 4–27 shows the timing of a SET SHARED command.

4–62 Preliminary—Subject to Change—July 1996

4.10 System-Initiated Transactions

Figure 4–27 SET SHARED Timing Diagram

sys_clk_out1_h
addr_bus_req_h
cmd_h<3:0>

SET SHARED

victim_pending_h
addr_h<39:4>

00D0

0000

cack_h
ACK/Bcache

addr_res_h<2:0>
idle_bc_h
index_h<25:4>

00C0

0000

data_h<127:0>
dack_h
data_ram_oe_h
data_ram_we_h
tag_ram_oe_h
tag_ram_we_h
tag_data_h<38:20>
tag_dirty_h
tag_shared_h
tag_valid_h
LJ-04017.AI

Preliminary—Subject to Change—July 1996 4–63

4.10 System-Initiated Transactions

4.10.3 Flush-Based Cache Coherency Protocol Commands
All 21164-based systems that use the flush protocol are expected to use
the READ and FLUSH commands defined in Table 4–14 to maintain cache
coherency.
Table 4–14 System-Initiated Interface Commands (Flush Protocol)
Command

cmd_h
<3:0>

NOP

0000

The NOP command is driven by the owner of the
cmd_h<3:0> bus when it has no tasks queued.

FLUSH

0001

Remove block from caches; return dirty data. The
FLUSH command causes a block to be removed
from the 21164 cache system. If the block is not
found, the 21164 responds with NOACK. If the
block is found and the block is clean, the 21164
responds with NOACK. The block is invalidated
in the Dcache, Scache, and Bcache. If the block
is found and is dirty, the 21164 responds with
ACK/Scache or ACK/Bcache. If the data is found
dirty in the Scache, it is driven at the interface in
the same sysclk as the ACK/Scache. If the data is
found dirty in the Bcache, the Bcache read starts on
the same sysclk as ACK. The block is invalidated in
the Dcache, Scache, and Bcache.

READ

0100

Read a block. The READ command probes the
Scache and Bcache to see if the requested block is
present. If the block is present, the 21164 responds
with ACK/Scache or ACK/Bcache. If the data is in
Scache, the data is driven on the data_h<127:0>
bus in the same sysclk as the ACK. If the data is in
the Bcache, a Bcache read operation begins in the
same sysclk as the ACK. If the block is not present
in either cache, the 21164 responds with a NOACK
on addr_res_h<1:0>.

Description

4–64 Preliminary—Subject to Change—July 1996

4.10 System-Initiated Transactions

4.10.3.1

Alpha 21164 Responses to Flush-Based Protocol Commands
The system responds to flush-based protocol commands on addr_res_h<1:0>,
as shown in Table 4–15.
Table 4–15

Alpha 21164 Responses to Flush-Based Protocol Commands
READ and FLUSH Commands

Bcache Status

Scache Status

21164 Response

No Bcache

Scache_Miss

NOACK

No Bcache

Scache_Hit,Not Dirty

NOACK

No Bcache

Scache_Hit,Dirty

ACK/Scache

Bcache_Miss

Scache_Miss

NOACK

Bcache_Hit

Scache_Hit,Dirty

ACK/Scache

Bcache_Hit,
Not Dirty

Scache_Miss/Hit, Not Dirty

NOACK

Bcache_Hit,Dirty

Scache_Miss

ACK/Bcache

The signal addr_res_h<2> allows a system without a duplicate tag store to
determine if a block is present in the Scache or lock register. The system logic
can use this information to correctly assert tag_shared_h in a multiprocessor
system.
The 21164 responds to the READ, FLUSH, READ DIRTY, SET SHARED,
and READ DIRTY/INVALIDATE commands on addr_res_h<2>, as listed in
Table 4–16.
Table 4–16

Alpha 21164 Responses on addr_res_h<2> to 21164 Commands

Scache

Lock Register

addr_res_h<2>

Miss

Hit

Miss

Hit

Table 4–17 presents the 21164 best-case response time to system commands in
a flush protocol system.

Preliminary—Subject to Change—July 1996 4–65

4.10 System-Initiated Transactions

Table 4–17 Minimum 21164 Response Time to Flush Protocol Commands
Cache Status

Response

Number of sys_clk_out1_h,l Cycles

No Bcache

NOACK

8 CPU cycles rounded up to next
sys_clk_out1_h,l cycles

No Bcache

ACK/Scache

12 CPU cycles rounded up to next
sys_clk_out1_h,l cycles

Bcache

NOACK,
ACK/Scache,
ACK/Bcache

10 CPU cycles plus <BC_RD_SPD> rounded up
to next sys_clk_out1_h,l cycles

4.10.3.2 FLUSH
The FLUSH command is used to remove blocks from the 21164 cache system.
Figure 4–28 shows the timing of a FLUSH transaction.
If the block is DIRTY, the 21164 will respond with an ACK and the system
must read data from the cache, using dack_h to control the rate at which data
is supplied, and write it to memory.
In the timing diagram shown in Figure 4–28, the cache block state changes
from DIRTY; SHARED; VALID to DIRTY ; SHARED; VALID. When the block
state changes to VALID, the state of SHARED and DIRTY does not matter.

4–66 Preliminary—Subject to Change—July 1996

4.10 System-Initiated Transactions

Figure 4–28 FLUSH Timing Diagram (Scache Hit)

sys_clk_out1_h
addr_bus_req_h
cmd_h<3:0>

FLUSH

0080

victim_pending_h
addr_h<39:4>
cack_h
ACK/Scache

addr_res_h<2:0>
idle_bc_h
index_h<25:4>

008

data_h<127:0>

009

00A

00B

dack_h
data_ram_oe_h
data_ram_we_h
tag_ram_oe_h
tag_ram_we_h
tag_data_h<38:20>

FC00

tag_dirty_h
tag_shared_h
tag_valid_h
LJ-04018.AI

Preliminary—Subject to Change—July 1996 4–67

4.10 System-Initiated Transactions

4.10.3.3 READ
The READ command is used by the system to read DIRTY data from the
21164. The tag control status does not change. Figure 4–29 shows the timing
and tag control status of a READ transaction.

4–68 Preliminary—Subject to Change—July 1996

4.10 System-Initiated Transactions

Figure 4–29 READ Timing Diagram (Scache Hit)

sys_clk_out1_h
addr_bus_req_h
cmd_h<3:0>

READ

victim_pending_h
addr_h<39:4>

0040

cack_h
ACK/Scache

addr_res_h<2:0>
idle_bc_h
index_h<25:4>

004

data_h<127:0>

005

006

007

004

dack_h
data_ram_oe_h
data_ram_we_h
tag_ram_oe_h
tag_ram_we_h
tag_data_h<38:20>

FC00

tag_dirty_h
tag_shared_h
tag_valid_h
LJ-04019.AI

Preliminary—Subject to Change—July 1996 4–69

4.11 Data Bus and Command/Address Bus Contention

4.11 Data Bus and Command/Address Bus Contention
The data bus is composed of data_h<127:0> and data_check_h<15:0>.
The command/address bus is composed of cmd_h<3:0>, addr_h<39:4>, and
addr_cmd_par_h.
The following sections describe situations that have contention for use of the
data bus or contention for use of the command/address bus.

4.11.1 Command/Address Bus
Figure 4–30 shows the 21164 and the system alternately driving the
command/address bus. If signal addr_bus_req_h is asserted at the rising
edge of sysclk N, the next cycle on the command/address bus belongs to the
system. The 21164 turns off its drivers at the rising edge of sysclk N. While
the system must turn on its drivers between sysclk N and sysclk N+1, it must
ensure that the drivers do not turn on before the 21164 drivers turn off. The
21164 samples the state of the command/address bus at the end of sysclk N+1.
If addr_bus_req_h remains asserted, the system should continue to drive the
command/address bus.
Figure 4–30 Driving the Command/Address Bus

N+1

N+2

sys_clk_out1_h

addr_bus_req_h

21164 Drive

System Drive
21164 Sample Point
MK−1455−03

To pass control of the command/address bus back to the 21164, the system
should turn off its drivers during a sysclk and deassert addr_bus_req_h. The
21164 does not sample the state of the bus if addr_bus_req_h is deasserted.
The 21164 drives the command/address bus at the rising edge of sysclk N+2.

4–70 Preliminary—Subject to Change—July 1996

4.11 Data Bus and Command/Address Bus Contention

On every 21164 sample point, the cmd_h<3:0>, addr_h<39:4>, and
addr_cmd_par_h signals must be valid, and the parity must be correct
unless BC_CONTROL<DIS_SYS_PAR> is set. If DIS_SYS_PAR is clear,
addr_cmd_par_h must be valid for the address and command, even when the
address is irrelevant, because the system is driving a NOP on cmd_h<3:0>.

4.11.2 Read/Write Spacing—Data Bus Contention
The data bus, data_h<127:0>, can be driven by the 21164, the Bcache array,
or the system.
In the case of private Bcache write operations followed by private Bcache read
operations, the 21164 stops driving the data bus well in advance of the Bcache
turning on.
For private Bcache read operations followed by private Bcache write
operations, the 21164 inserts a programmable number of CPU cycles between
the read and the write operation. This allows time for the Bcache drivers to
turn off before the 21164 data drivers are turned on.
Note
This rule also applies to WRITE BLOCK, WRITE BLOCK LOCK,
READ, READ DIRTY, READ DIRTY/INV, and FLUSH commands.

Preliminary—Subject to Change—July 1996 4–71

4.11 Data Bus and Command/Address Bus Contention

4.11.3 Using idle_bc_h and fill_h
The 21164 uses the idle_bc_h and fill_h signals to fill data into the Scache,
the Bcache, or both. The system must assert the idle_bc_h signal early
enough to ensure that the 21164 completes any Bcache transaction it might
have started while waiting for the fill data.
Signal fill_h is asserted a fixed number of sysclk cycles before the start of a fill
transaction.
At the end of the fill, the 21164 waits five CPU cycles before starting a read or
write operation. This time should allow the system to turn off its drivers. If,
in practice, this is not enough time, the system may assert data_bus_req_h to
gain additional cycles.
Calculating Time to Assert idle_bc_h
The equations for calculating length of time to assert idle_bc_h are:
read_hit_idle = 2 + (block_size/16) 3 BC_RD_SPD +
tristate_ram_turn_off - 3 3 wave_pipelining;
read_miss_idle = 6 + BC_RD_SPD + Sysclk_ratio + tristate_RAM_turn_off;
write_idle

= 4 + (block_size/16) 3 BC_WRT_SPD + tristate_21164_turn_off;

When using these equations, the turn-off times should be expressed as an
integer number of CPU clock periods. Take the largest of the three times and
then round up to the next sysclk boundary.
When determining the tristate turn-off times, if the system will not turn on its
drivers for some number of nanoseconds after the 21164 starts driving Bcache
index_h<25:4>; this time can be used to reduce the tristate_turn_off time.
For example if the sysclk ratio is 6 (the caches use a 64-byte block size),
Bcache read/write speed is 5, with no wave pipelining, 2 cycles for tristate
read, 0 cycles for tristate_write, then the equations would work out to:
read_hit_idle = 2 + (64/16) 3 5 + 2 - 3 3 0 = 24
read_miss_idle = 6 + 5 + 6 + 2 = 19
write_idle

= 4 + (64/16) 3 5 + 0 = 24

Maximum of (24/6), (19/6), (24/6) = 4
In this example wave_pipelining = 0 makes only the partial product zero, not
the entire equation.

4–72 Preliminary—Subject to Change—July 1996

4.11 Data Bus and Command/Address Bus Contention

If the 21164 samples idle_bc_h asserted at sysclk edge N, the earliest time
that the system can allow the 21164 to sample fill_h asserted is at sysclk edge
N+3. The 21164 drives index_h<25:4> to fill the Bcache on sysclk edge N+4.
Systems without a Bcache are not required to assert idle_bc_h to use the
data_bus_req_h signal.
Figure 4–31 Example of Using idle_bc_h and fill_h

N+1

N+2

N+3

N+4

sys_clk_out1_h,l

idle_bc_h

fill_h

dack_h

index_h<25:4>

data<127:0>

D3
LJ-04020.AI5

Minimum idle_bc_h time
If the system contains a Bcache, and the write ratio of the Bcache is greater
than or equal to twice the sysclk ratio, then the minimum idle_bc_h assertion
time is two sysclk cycles.
For example, if the Bcache write speed is 10, and the sysclk ratio is 4, then any
assertion of idle_bc_h must be for two or more sysclk cycles.

Preliminary—Subject to Change—July 1996 4–73

4.11 Data Bus and Command/Address Bus Contention

4.11.4 Using data_bus_req_h
The signal data_bus_req_h can be used along with the idle_bc_h signal to
prevent the 21164 and the Bcache from driving the data bus. In general, the
system should not need to use this feature but it may be useful if the system
places other devices on the data bus.
To gain control of the data bus, the system must ensure that the Bcache
is idle by asserting idle_bc_h for the required time. It can then assert
data_bus_req_h. If data_bus_req_h is received asserted at the rising edge of
sysclk N, the 21164 stops driving the bus on the rising edge of sysclk N+1.
To return the bus to the 21164, the system should deassert data_bus_req_h
and then deassert idle_bc_h on the next sysclk.
Figure 4–32 Using data_bus_req_h
N

N+1

sys_clk_out1_h,l

idle_bc_h

data_bus_req_h

21164 Drive
LJ-04021.AI5

4–74 Preliminary—Subject to Change—July 1996

4.11 Data Bus and Command/Address Bus Contention

4.11.5 Tristate Overlap
The addr_h<39:4>, cmd_h<3:0>, data_h<127:0>, and tag_data_h<38:20>
buses must be operated in such a way that no more than one driver may
drive the bus at a time. This section describes particular cases where tristate
overlap may be a problem that needs to be corrected using features described
in previous sections.
The ‘‘owner’’ of each bus must drive the bus to some value for each cycle.
Tristate drivers in the 21164 turn on and off very fast (in the 0.5-ns to 1.0-ns
range). At the other end of the range, SRAM memory devices turn on and
off slowly (in the 7.0-ns to 10.0-ns range). Generally, system drivers fall
somewhere in the middle.
4.11.5.1 READ or WRITE to FILL
The time required to tristate the 21164 drivers at the end of a WRITE
command, or the Bcache drivers at the end of a READ command is part of
the idle_bc_h equation.
4.11.5.2 BCACHE VICTIM to FILL
The time to turn off the Bcache drivers at the end of a BCACHE VICTIM is
fixed by the 21164 design. The system must allow for this time before starting
a FILL.
There are two READ MISS with victim cases to consider. In one case, the
READ MISS operation will be completed first because the system logic contains
a victim buffer. In the other case the READ MISS operation will be completed
second because the system logic does not have a victim buffer.

Preliminary—Subject to Change—July 1996 4–75

4.11 Data Bus and Command/Address Bus Contention

READ MISS Completed First—Victim Buffer
The final dack_h will be sampled by the 21164 on the rising edge of sysclk. If
the corresponding rising CPU clock edge is labeled N, then data_ram_oe_h
will deassert at the rising edge of CPU clock N+4.
Figure 4–33 READ MISS Completed First—Victim Buffer

N N+1 N+2 N+3 N+4

CPU Clock Cycles

sys_clk_out1_h

dack_h

index<25:4>

data_h<127:0>

data_ram_oe_h
LJ-04022.AI5

4–76 Preliminary—Subject to Change—July 1996

4.11 Data Bus and Command/Address Bus Contention

READ MISS Second—No Victim Buffer
The final dack_h will be sampled by 21164 on the rising edge of sysclk. If
the corresponding rising CPU clock edge is labeled N, then the READ MISS
command will arrive on the next sysclk edge, and the data_ram_oe_h will
deassert at the rising edge of CPU clock N+S+1, where S is the sysclk ratio.
If the sysclk ratio is 3, it will take an extra sysclk to send the READ MISS
command, so the data_ram_oe_h will deassert at N+2S+1.
Figure 4–34 READ MISS Second—No Victim Buffer

N+S N+S+1

CPU Clock Cycles

sys_clk_out1_h

cmd_h<3:0>

READ MISS

dack_h

index<25:4>

data_h<127:0>

data_ram_oe_h

LJ-04023.AI5

Preliminary—Subject to Change—July 1996 4–77

4.11 Data Bus and Command/Address Bus Contention

4.11.5.3 System Bcache Command to FILL
At the end of a system command that uses the Bcache, the system must
provide enough time for the Bcache drivers to turn off before returning any fill
data.
The final dack_h will be sampled by the 21164 on the rising edge of sysclk.
If the corresponding rising CPU clock edge is labeled N, data_ram_oe_h will
deassert at the rising edge of CPU clock N+5.
Figure 4–35 System Command to FILL Example 1
N

N+1 N+2 N+3 N+4 N+5

CPU Clock Cycles

sys_clk_out1_h

dack_h

index<25:4>

data_h<127:0>

data_ram_oe_h

fill_h
LJ-04024.AI5

A side effect of this is the earliest assertion of fill_h after a system command.
The system must allow time for data_ram_oe_h to turn off and the RAMs to
stop driving the bus before the system drives the fill data.

4–78 Preliminary—Subject to Change—July 1996

4.11 Data Bus and Command/Address Bus Contention

If the system command was a SET SHARED or an INVALIDATE command,
the system must allow time for the 21164 to complete the Bcache tag write
operation and then for the drivers to turn off before driving the tag_shared_h,
tag_dirty_h, and tag_ctl_par_h lines.
The 21164 begins the tag write operation one CPU cycle after the response
is sent to the system. The write transaction will take BC_WRT_SPD cycles
to complete. During the write transaction, data_ram_oe_h will be asserted
but not tag_ram_oe_h. At the end of the write transaction, tag_ram_oe_h
will pulse for one CPU cycle, then both will go off. Refer to Figure 4–36 if the
response is driven at the rising edge of CPU clock N, then data_ram_oe_h
will fall at N+2+BC_WRT_SPD, or N+6 for a 4-cycle write speed.
Figure 4–36 System Command to FILL Example 2

N+2+BC_WRT_SPD

CPU Clock Cycles

sys_clk_out1_h

addr_res_h<1:0>

index<25:4>

ACK/Bcache

TAG WRITE

data_h<127:0>

tag_ram_oe_h

data_ram_oe_h
LJ-04025.AI5

Preliminary—Subject to Change—July 1996 4–79

4.11 Data Bus and Command/Address Bus Contention

4.11.5.4 FILL to Private Read or Write Operation
At the end of the fill, the 21164 does not begin to drive the data bus until the
fifth CPU cycle after the sysclk that loads the last dack_h. The 21164 does not
assert data_ram_oe_h until the fifth cycle after the sysclk that loads the last
dack_h.
Systems requiring more time to turn off their drivers must not send any more
requests and must use idle_bc_h and data_bus_req_h at the end of the fill to
stop 21164 requests.
Figure 4–37 FILL to Private Read or Write Operation

N+1

N+2

N+3

N+4

N+5

CPU Clock Cycles

sys_clk_out1_h

dack_h

index<25:4>

data_h<127:0>

data_ram_oe_h
LJ-04026.AI5

4–80 Preliminary—Subject to Change—July 1996

4.12 Alpha 21164 Interface Restrictions

4.12 Alpha 21164 Interface Restrictions
This section lists restrictions on the use of 21164 interface features.

4.12.1 FILL Operations after Other Transactions
If the system has removed data from the 21164 with any of the system
commands, or completed a WRITE_BLOCK, or removed a Bcache victim from
the Bcache, and wants to follow any of these transactions with a FILL, then
the earliest point the system can assert the fill_h signal is at the sysclk after
the last assertion of dack_h. However, fill_h can be asserted at the sysclk
with the last dack_h if the sysclk ratio is greater than 3.
FILL operations followed by FILL operations are special cases. FILL
operations can be pipelined back-to-back so that 100% of the data bus
bandwidth can be used.

4.12.2 Command Acknowledge for WRITE BLOCK Commands
When the 21164 requests a WRITE BLOCK or WRITE BLOCK LOCK
operation, the system can acknowledge the data by asserting dack_h before
asserting cack_h. The system must assert cack_h no later than the last
assertion of dack_h.

4.12.3 Systems Without a Bcache
Systems without a Bcache must set a 64-byte block size.
If systems without a Bcache have an Scache duplicate tag store, they are
required to maintain tags for the two blocks in the 21164 Scache victim buffer.

4.12.4 Fast Probes with No Bcache
If BC_CONTROL<BC_ENABLED>=0, then the 21164 processes system
requests while other commands are being processed by the interface. The
21164 does not wait for the interface to become idle before processing system
requests. This creates race conditions for the state of a cache block.
For example, if a certain block is being filled private-clean, and the system
sends a SET SHARED command for the block, the SET SHARED command
must be delayed until the fill completes and records the correct end state for
the block, shared-clean. The system must avoid changing the state of a block
that is in transit.
The restrictions are as follows:
•

The system may not send a request to the 21164 for a block that has been
filled until one sysclk after the last dack_h if the sysclk ratio is greater
than 3.

Preliminary—Subject to Change—July 1996 4–81

4.12 Alpha 21164 Interface Restrictions

•

The system may not send a request to the 21164 for a block that has been
filled until two sysclks after the last dack_h if the sysclk ratio is 3.

•

The system may not send a request to the 21164 for a block that has
completed a WRITE BLOCK command until one sysclk after the last
dack_h.

•

The system may not send a request to the 21164 for a block that has
completed a SET DIRTY command until one sysclk after the cack_h for
the SET DIRTY command.

If BC_CONTROL<BC_ENABLED>=1, all system requests are delayed to avoid
race conditions.

4.12.5 WRITE BLOCK LOCK
A WRITE BLOCK LOCK transaction is caused by a store conditional
instruction to I/O space. Two octawords of data are provided by the 21164,
each requiring the system to assert dack_h. If the system asserts dack_h for
the first octaword, and asserts cack_h and cfail_h together, the 21164 hangs.
If dack_h, cack_h, and cfail_h are asserted for the second INT16 of data, the
write operation will be failed correctly.
If cack_h and cfail_h are asserted at any time without asserting dack_h, the
write operation will be failed correctly.

4–82 Preliminary—Subject to Change—July 1996

4.13 Alpha 21164/System Race Conditions

4.13 Alpha 21164/System Race Conditions
When certain sequences of transactions occur on the interface between the
21164, the Bcache and the system race conditions may occur. The rules for use
of the interface by the 21164 and the system are listed in Section 4.13.1.
Examples of race conditions to be avoided are described and illustrated in
Section 4.13.2 through Section 4.13.6.

4.13.1 Rules for 21164 and System Use of External Interface
This section goes over the rules for determining the order in which 21164 and
system requests are allowed by the Cbox BIU. In general, the order allowed is
determined by use of cmd_h<3:0>, idle_bc_h, and fill_h.
1. If idle_bc_h is not asserted and there are no valid requests in the BIU
command buffer, then the BIU is free to perform any 21164 request.
2. If a FILL transaction is pending, the BIU only produces another READ
MISS command, with a possible BCACHE VICTIM command. The BIU
will not attempt any other command.
3. The assertion of idle_bc_h, or the sending of a system command other
than NOP to the 21164, causes the BIU to idle. If the BIU has a command
loaded in the pad ring, it removes the command and replaces it with a
NOP command. The state of cmd_h<3:0> is unpredictable until the idle
condition ends.
4. The idle condition ends when the 21164 receives a deasserted idle_bc_h,
and the 21164 has responded to all the system commands that were sent.
5. The system must not assert cack_h during the idle condition.
6. There is one exception to rules 3, 4, and 5. If idle_bc_h or a system
command arrives while the 21164 is reading the Bcache, and that read
transaction turns into a READ MISS transaction, and it does not produce
a victim, then the 21164 loads the miss into the pad ring. The system may
assert cack_h for this READ MISS request at any time.
7. If cack_h is asserted at the same time as idle_bc_h or a valid system
request, cack_h wins and the command is taken by the system. Signal
cack_h should not be asserted if idle_bc_h has been asserted or a valid
system command is under way.
8. A READ MISS with a BCACHE VICTIM transaction is treated as an
atomic pair. The command order, READ MISS then BCACHE VICTIM or
BCACHE VICTIM then READ MISS, is programmable. Either way, if the
first command is acknowledged with cack_h, then both commands must be

Preliminary—Subject to Change—July 1996 4–83

4.13 Alpha 21164/System Race Conditions

acknowledged with cack_h and all the data acknowledged with dack_h,
before the 21164 responds to any other request.
9. The cack_h acknowledgment for a WRITE BLOCK or BCACHE VICTIM
transaction must be received by the 21164 with or before the last dack_h
acknowledgment of the data. For WRITE BLOCK and BCACHE VICTIM
transactions, it is possible to acknowledge all but the last data, and then
decide to do something else.
10. For a READ MISS transaction, cack_h must be received with or before the
last data acknowledgment (dack_h) for the requested FILL operation.
11. If a 21164 request is interrupted by an idle condition, the 21164 restarts
the same command unless:
a. A system request is received that changes the state of the block made
by the original 21164 request.
For example, if the 21164 is requesting a WRITE BLOCK and the
system sends an INVALIDATE command to the same block, then the
WRITE BLOCK command will not be restarted.
b. If the system does not have a Bcache, and a WRITE BLOCK command
to write an Scache victim back is interrupted, then the WRITE BLOCK
command will not be restarted if a higher priority request arrives in
the BIU.

4.13.2 READ MISS with Victim Example
In this example, the 21164 asserts a READ MISS command with a victim. The
system asserts dack_h for two data cycles received from the Bcache and then
asserts idle_bc_h. This causes the 21164 to remove the READ MISS command
with victim pending. The 21164 reasserts the READ MISS and BCACHE
VICTIM commands, if needed, at a later time.

4–84 Preliminary—Subject to Change—July 1996

4.13 Alpha 21164/System Race Conditions

Figure 4–38 READ MISS with Victim Example

sys_clk_out1_h Cycles

cmd_h<3:0>

NOP

READ MISS

addr_h<39:4>

victim_pending_h

addr_bus_req_h

idle_bc_h

cack_h

dack_h

index<25:4>

data_h<127:0>

data_ram_oe_h
LJ-04027.AI

Preliminary—Subject to Change—July 1996 4–85

4.13 Alpha 21164/System Race Conditions

4.13.3 idle_bc_h and cack_h Race Example
In this example, idle_bc_h and cack_h are asserted in the same sysclk. The
system takes the READ MISS and BCACHE VICTIM commands before doing
anything else. The last dack_h meets the requirement that the cack_h arrive
before or with the last dack_h.

4–86 Preliminary—Subject to Change—July 1996

4.13 Alpha 21164/System Race Conditions

Figure 4–39 idle_bc_h and cack_h Race Example

sys_clk_out1_h Cycles

cmd_h<3:0>

NOP

READ MISS

NOP

BCACHE VICTIM

addr_h<39:4>

victim_pending_h

addr_bus_req_h

idle_bc_h

cack_h

dack_h

index<25:4>

data_h<127:0>

data_ram_oe_h

LJ-04028.AI

Preliminary—Subject to Change—July 1996 4–87

4.13 Alpha 21164/System Race Conditions

4.13.4 READ MISS with idle_bc_h Asserted Example
In this example, the 21164 has started a Bcache read operation that misses.
The signal idle_bc_h is asserted, but no victim was created, so the READ
MISS request is loaded into the pad ring. The system then takes the request.

Figure 4–40 READ MISS with idle_bc_h Asserted Example

sys_clk_out1_h Cycles

cmd_h<3:0>

READ MISS

NOP

addr_h<39:4>

NOP

victim_pending_h

addr_bus_req_h

idle_bc_h

cack_h

dack_h

index<25:4>

data_h<127:0>

data_ram_oe_h
LJ-04029.AI5

4–88 Preliminary—Subject to Change—July 1996

4.13 Alpha 21164/System Race Conditions

4.13.5 READ MISS with Victim Abort Example
In this example, the 21164 produces a READ MISS command with a victim
and is waiting for the system to take it when the system takes the bus and
requests a READ DIRTY transaction. The 21164 drives the READ MISS
request for one more cycle after it gets command of the bus and then removes
the request. The 21164 then responds to the READ DIRTY command and
drives index_h<25:4> to read the Bcache. The 21164 restarting the Bcache
read operation, requesting the read miss with victim, is not shown in the
timing diagram. If the victim block was invalidated by the system request, the
21164 produces a clean READ MISS transaction.

Preliminary—Subject to Change—July 1996 4–89

4.13 Alpha 21164/System Race Conditions

Figure 4–41 READ MISS with Victim Abort Example

sysclk
cmd_h<3:0>

NOP

addr_h<39:4>

READ MISS

READ
DIRTY

NOP

victim_pending_h
addr_bus_req_h
addr_res_h<2:0>

NOP

ACK

NOP

idle_bc_h
cack_h
dack_h
index<25:4>

data_h<127:0>

data_ram_oe_h
LJ-04030.AI

4.13.6 Bcache Hit Under READ MISS Example
In this example, the 21164 produces a READ MISS transaction and requests
a fill from the system. A Bcache hit to index j take places while waiting for
the fill. The system then returns the requested data in two bursts, asserting
cack_h at the same time as the last assertion of dack_h.

4–90 Preliminary—Subject to Change—July 1996

4.13 Alpha 21164/System Race Conditions

Figure 4–42 Bcache Hit Under READ MISS Example

sysclk
cmd_h<3:0>

READ MISS

NOP

addr_h<39:4>

NOP

victim_pending_h
addr_bus_req_h
fill_h
idle_bc_h
cack_h
dack_h
index<25:4>

data_h<127:0>

data_ram_oe_h
LJ-04031.AI

Preliminary—Subject to Change—July 1996 4–91

4.14 Data Integrity, Bcache Errors, and Command/Address Errors

4.14 Data Integrity, Bcache Errors, and Command/Address
Errors
Mechanisms for ensuring that errors on data received by the 21164 from the
Bcache, the system, or both are described in this section. Tag data and tag
control errors are described. Command/address bus parity protection is also
described.

4.14.1 Data ECC and Parity
The 21164 supports INT8 error correction code (ECC) for the external Bcache
and memory system. ECC is generated by the CPU for each INT8 that is
written into the Bcache. FILL data from the Bcache to the system is not
checked for errors. The receiving node detects any ECC errors.
Uncorrected data from the Bcache or system is sent to the Dcache, and register
files. If a correctable error is detected (single bit error) the machine traps and
the fill is replayed with corrected data.
Double bit errors are detected. If the system indicates that the data should not
be checked, then no checking or correcting is performed.
Each data bus cycle delivers one INT16 worth of data. ECC is calculated as
ECC(data<063:000>) and ECC(data<127:064>). Figure 4–43 shows the code.
Two IDT49C460 or AMD29C660 chips can be cascaded to produce this ECC
code. A single IDT49C466 chip also supports this ECC code.
The code provides single bit correct, double bit detect, and all 1s and all 0s
detect.
If the 21164 is in parity mode, it generates byte parity and places it on
data_check_h<15:0> for write operations. Parity is checked for read
operations. Parity for data_h<7:0> is driven on signal data_check_h<0>
and so on.

4–92 Preliminary—Subject to Change—July 1996

4.14 Data Integrity, Bcache Errors, and Command/Address Errors

Figure 4–43 ECC Code

11 1111 1111 2222 2222 2233 3333 3333 4444 4444 4455 5555 5555 6666 cccc cccc
0123 4567 8901 2345 6789 0123 4567 8901 2345 6789 0123 4567 8901 2345 6789 0123 01.23 4567
CB0
CB1
CB2
CB3

. 111 . 1 . . 11 . 1 . . 1 . . 111 . 1 . . 11 . 1 . . 1 . 1 . . . 1 . 11 . . 1 . 11 . 1 1 . . . 1 . 11 . . 1 . 11 . 1 1 . . . . . . .
111 . 1 . 1 . 1 . 1 . 1 . . . 111 . 1 . 1 . 1 . 1 . 1 . . . 111 . 1 . 1 . 1 . 1 . 1 . . . 111 . 1 . 1 . 1 . 1 . 1 . . . . 1 . . . . . .
1 . . 1 1 . . 1 . 11 . . 1 . 1 1 . . 1 1 . . 1 . 11 . . 1 . 1 1 . . 1 1 . . 1 . 11 . . 1 . 1 1 . . 1 1 . . 1 . 11 . . 1 . 1 . . 1 . . . . .
11 . . . 111 . . . 1 11 . . 11 . . . 111 . . . 1 11 . . 11 . . . 111 . . . 1 11 . . 11 . . . 111 . . . 1 11 . . . . . 1 . . . .

CB4
CB5
CB6
CB7

. . 11 1111 . . . . . . 11 . . 11 1111 . . . . . . 11 . . 11 1111 . . . . . . 11 . . 11 1111 . . . . . . 11 . . . .
. . . . . . . . 1111 1111 . . . . . . . . 1111 1111 . . . . . . . . 1111 1111 . . . . . . . . 1111 1111 . . . .
1111 1111 . . . . . . . . . . . . . . . . 1111 1111 1111 1111 . . . . . . . . . . . . . . . . 1111 1111 . . . .
1111 1111 . . . . . . . . . . . . . . . . 1111 1111 . . . . . . . . 1111 1111 1111 1111 . . . . . . . . . . . .

1. . .
.1. .
. .1.
. . .1

CB2 and CB3 are calculated for CDD parity (an odd number of 1s counting
the CB).
CB0, CB1, CB4, CB5, CB6, and CB7 are calculated for EVEN parity (an even
number of 1s counting the CB).

LJ-03461-TI0

The correspondence of data check bits to CBn is shown in Table 4–18.
Table 4–18 Data Check Bit Correspondence to CBn
data_check_h
CBn

Upper 64 bits

Lower 64 bits

CB0

<8>

<0>

CB1

<9>

<1>

CB2

<10>

<2>

CB3

<11>

<3>

CB4

<12>

<4>

CB5

<13>

<5>

CB6

<14>

<6>

CB7

<15>

<7>

Preliminary—Subject to Change—July 1996 4–93

4.14 Data Integrity, Bcache Errors, and Command/Address Errors

For x4 RAMs, the following bit arrangement detects nibble errors:
CB0
CB2
CB3
CB7
D1
D9
D12
D15
D24
D26
D32
D33
D39
D42
D48
D49
D55
D58

CB1
D0
CB4
D2
D6
D14
D16
D19
D25
D28
D34
D36
D41
D44
D50
D52
D57
D60

CB5
D4
D7
D3
D10
D18
D17
D20
D27
D29
D35
D38
D43
D45
D51
D54
D59
D61

CB6
D5
D8
D11
D13
D21
D22
D23
D30
D31
D37
D40
D46
D47
D53
D56
D62
D63

4.14.2 Force Correction
Setting BC_CTL<4> (CORR_FILL_DAT), forces the 21164 to route fill data
from the Bcache or memory through error correction logic before being driven
to the Scache or Dcache. If the error is correctable, it is transparent to the
21164.

4.14.3 Bcache Tag Data Parity
The signal line tag_data_par_h is used to maintain parity over
tag_data_h<38:20>. A Bcache tag data parity error is usually not recoverable.
A Bcache hit is determined based on the tag alone, not the tag parity bit.
The Cbox records the Bcache probe address and the tag value read from the
Bcache. A tag data parity error causes a trap to privileged architecture library
code (PALcode), which handles the error condition.

4.14.4 Bcache Tag Control Parity
The signal tag_ctl_par_h is used to maintain parity over tag_shared_h,
tag_valid_h, and tag_dirty_h. A Bcache tag control parity error is usually
not recoverable.
A Bcache victim is processed according to the tag control status alone, not the
tag control parity bit. The Cbox records the Bcache probe address and the tag
control value read from the Bcache. A tag control parity error causes a trap to
PALcode, which handles the error condition.

4–94 Preliminary—Subject to Change—July 1996

4.14 Data Integrity, Bcache Errors, and Command/Address Errors

4.14.5 Address and Command Parity
The signal line addr_cmd_par_h is used to maintain odd parity over
addr_h<39:4> and cmd_h<3:0>. These signals are driven by the 21164
or by the system, using the protocol described in Section 4.11.1.

4.14.6 Fill Error
The signal fill_error_h is asserted by the system to notify the 21164 that a fill
error has occurred.
Systems in which a fill error timeout is not expected, such as a small system
with fixed access time, it is likely that the 21164 internal Ibox timeout logic
would detect a stall if the system fails to complete a fill transaction.
Systems in which a fill error timeout could occur should contain logic to detect
fill timeouts and cleanly terminate the transaction with the 21164.
To properly terminate a fill in an error case, the fill_error_h line is asserted
for one cycle and the normal fill sequence involving lines fill_h, fill_id_h, and
dack_h is generated by the system.
Asserting fill_error_h forces a trap to the PALcode at the MCHK entry point
but has no other effect.

4.14.7 Forcing 21164 Reset
Assertion of cfail_h in a sysclk cycle in which cack_h is deasserted causes
the 21164 to execute a partial internal reset and then trap to the MCHK entry
point in PALcode. The current command, if any, and all pending fills, and all
pending system commands are cleared. The 21164 will complete its partial
reset in 128 CPU cycles, then begin execution of the machine check PALcode
flow. The system should not send a request to the 21164 during this time.
This mechanism is used by the 21164 to restore itself and the system to a
consistent state after command or address parity error or a timeout error.
Refer also to Section 8.1.18.

Preliminary—Subject to Change—July 1996 4–95

4.15 Interrupts

4.15 Interrupts
The 21164 has seven interrupt signals that have different uses during
initialization and normal operation.
Figure 4–44 shows the 21164 interrupt signals.
Figure 4–44

Alpha 21164 Interrupt Signals

irq_h<3:0>
sys_mch_chk_irq_h

21164

pwr_fail_irq_h
mch_hlt_irq_h

LJ-03669-TI0

4.15.1 Interrupt Signals During Initialization
The 21164 interrupt signals work in tandem with the sys_reset_l signal to
set the values for clock ratios and clock delays. During initialization, the
21164 reads system clock configuration parameters from the interrupt pins.
Section 4.2.2 and Section 4.2.3 describe how the interrupt signals are used to
set system clock values when the system is initialized.

4.15.2 Interrupt Signals During Normal Operation
During normal operation, interrupt signals indicate interrupt requests from
external devices such as the realtime clock and I/O controllers.

4.15.3 Interrupt Priority Level
Table 4–19 shows which interrupts are enabled for a given interrupt priority
level (IPL). An interrupt is enabled if the current IPL is less than the target
IPL of the interrupt.

4–96 Preliminary—Subject to Change—July 1996

4.15 Interrupts

Table 4–19 Interrupt Priority Level Effect
Interrupt Source

Target IPL

Source

Software Interrupt Request 1

Internal

Software Interrupt Request 2

Internal

Software Interrupt Request 3

Internal

Software Interrupt Request 4

Internal

Software Interrupt Request 5

Internal

Software Interrupt Request 6

Internal

Software Interrupt Request 7

Internal

Software Interrupt Request 8

Internal

Software Interrupt Request 9

Internal

Software Interrupt Request 10

Internal

Software Interrupt Request 11

Internal

Software Interrupt Request 12

Internal

Software Interrupt Request 13

Internal

Software Interrupt Request 14

Internal

Software Interrupt Request 15

Internal

Asynchronous system trap ATR pending (for
current or more privileged mode)

Internal

Performance counter interrupt

Internal

Powerfail interrupt1

pwr_fail_irq_h

System machine check interrupt , internally
detected correctable error interrupt pending

sys_mch_chk_irq_h
and internal

External interrupt 201

202

irq_h<0>

External interrupt 21

irq_h<1>

External interrupt 22

irq_h<2>

External interrupt 23

irq_h<3>

1 These interrupts are from external sources. In some cases, the system environment provides the

logic-OR of multiple interrupt sources at the same IPL to a particular pin.
2 The external interrupts 20–23 are separately maskable by setting the appropriate bits in the

ICSR register.

(continued on next page)

Preliminary—Subject to Change—July 1996 4–97

4.15 Interrupts

Table 4–19 (Cont.) Interrupt Priority Level Effect
Interrupt Source

Target IPL

Source

Halt

Masked
only by
executing
in
PALmode.

mch_hlt_irq_h

Serial line interrupt

Masked
only by
executing
in
PALmode.

Internal

1 These interrupts are from external sources. In some cases, the system environment provides the

logic-OR of multiple interrupt sources at the same IPL to a particular pin.

When the processor receives an interrupt request and that request is enabled,
an interrupt is reported or delivered to the exception logic if the processor is
not currently executing PALcode. Before vectoring to the interrupt service
PAL dispatch address, the pipeline is completely drained to the point that
instructions issued before entering the PALcode cannot trap (implied TRAPB).
The restart address is saved in the exception address (EXC_ADDR) IPR and
the processor enters PALmode. The cause of the interrupt can be determined
by examining the state of the INTID and ISR registers.
Hardware interrupt requests are level sensitive and therefore may be removed
before an interrupt is serviced. PALcode must verify that the interrupt actually
indicated in INTID is to be serviced at an IPL higher that the current IPL. If
it is not, PALcode should ignore the spurious interrupt.

4–98 Preliminary—Subject to Change—July 1996

5
Internal Processor Registers
This chapter describes the 21164 microprocessor internal processor registers
(IPRs). It is organized as follows:
•

Instruction fetch/decode unit and branch unit (Ibox) IPRs

•

Memory address translation unit (Mbox) IPRs

•

Cache control and bus interface unit (Cbox) IPRs

•

PAL storage registers

•

Restrictions

Ibox, Mbox, data cache (Dcache), and PALtemp IPRs are accessible to PALcode
by means of the HW_MTPR and HW_MFPR instructions. Table 5–1 lists the
IPR numbers for these instructions.
Cbox, second-level cache (Scache), and backup cache (Bcache) IPRs are
accessible in the physical address region FF FFF0 0000 to FF FFFF FFFF.
Table 5–25 summarizes the Cbox, Scache, and Bcache IPRs. Table 5–38 lists
restrictions on the IPRs.
Note for Windows NT
For 21164–P1 and 21164–P2 users, the following bits must be set:
•

IBOX control and status register (ICSR<28>) SPE<0> must always
be set (Section 5.1.17). Clearing this bit will cause 21164–Pn
operation to be UNPREDICTABLE.

•

MBOX control register (MCSR<01>) SP<0> must always be set
(Section 5.2.14). Clearing this bit will cause 21164–Pn operation to
be UNPREDICTABLE.

Preliminary—Subject to Change—July 1996 5–1

Note
Unless explicitly stated, IPRs are not cleared or set by hardware on
chip or timeout reset.

Table 5–1 Ibox, Mbox, Dcache, and PALtemp IPR Encodings
IPR Mnemonic

Access

Index16

Ibox Slots to Pipe

ISR

100

ITB_TAG

101

ITB_PTE

R/W

102

ITB_ASN

R/W

103

Ibox IPRs

ITB_PTE_TEMP

104

ITB_IA

105

ITB_IAP

106

ITB_IS

107

SIRR

R/W

108

ASTRR

R/W

109

ASTER

R/W

10A

EXC_ADDR

R/W

10B

EXC_SUM

R/W0C

10C

EXC_MASK

10D

PAL_BASE

R/W

10E

ICM

R/W

10F

IPLR

R/W

110

INTID

111

IFAULT_VA_FORM

112

IVPTBR

R/W

113

HWINT_CLR

115

SL_XMIT

116

SL_RCV

117

E1
(continued on next page)

5–2 Preliminary—Subject to Change—July 1996

Table 5–1 (Cont.) Ibox, Mbox, Dcache, and PALtemp IPR Encodings
IPR Mnemonic

Access

Index16

Ibox Slots to Pipe

ICSR

R/W

118

IC_FLUSH_CTL

119

ICPERR_STAT

R/W1C

11A

PMCTR

R/W

11C

PALtemp0

R/W

140

PALtemp1

R/W

141

PALtemp2

R/W

142

PALtemp3

R/W

143

PALtemp4

R/W

144

PALtemp5

R/W

145

PALtemp6

R/W

146

PALtemp7

R/W

147

PALtemp8

R/W

148

PALtemp9

R/W

149

PALtemp10

R/W

14A

PALtemp11

R/W

14B

PALtemp12

R/W

14C

PALtemp13

R/W

14D

PALtemp14

R/W

14E

PALtemp15

R/W

14F

PALtemp16

R/W

150

PALtemp17

R/W

151

PALtemp18

R/W

152

PALtemp19

R/W

153

PALtemp20

R/W

154

PALtemp21

R/W

155

PALtemp22

R/W

156

PALtemp IPRs

(continued on next page)

Preliminary—Subject to Change—July 1996 5–3

Table 5–1 (Cont.) Ibox, Mbox, Dcache, and PALtemp IPR Encodings
IPR Mnemonic

Access

Index16

Ibox Slots to Pipe

PALtemp23

R/W

157

DTB_ASN

200

DTB_CM

201

Mbox IPRs

DTB_TAG

202

DTB_PTE

R/W

203

DTB_PTE_TEMP

204

MM_STAT

205

206

VA_FORM

207

MVPTBR

208

DTB_IAP

209

DTB_IA

20A

DTB_IS

20B

ALT_MODE

20C

20D

CC_CTL

20E

MCSR

R/W

20F

DC_FLUSH

210

DC_PERR_STAT

R/W1C

212

DC_TEST_CTL

R/W

213

DC_TEST_TAG

R/W

214

DC_TEST_TAG_TEMP

R/W

215

DC_MODE

R/W

216

MAF_MODE

R/W

217

5–4 Preliminary—Subject to Change—July 1996

5.1 Instruction Fetch/Decode Unit and Branch Unit (Ibox) IPRs

5.1 Instruction Fetch/Decode Unit and Branch Unit (Ibox)
IPRs
The Ibox internal processor registers (IPRs) are described in Section 5.1.1
through Section 5.1.27.

5.1.1 Istream Translation Buffer Tag Register (ITB_TAG)
ITB_TAG is a write-only register written by hardware on an
ITBMISS/IACCVIO, with the tag field of the faulting virtual address. To
ensure the integrity of the instruction translation buffer (ITB), the TAG and
page table entry (PTE) fields of an ITB entry are updated simultaneously by
a write operation to the ITB_PTE register. This write operation causes the
contents of the ITB_TAG register to be written into the tag field of the ITB
location, which is determined by a not-last-used replacement algorithm. The
PTE field is obtained from the HW_MTPR ITB_PTE instruction. Figure 5–1
shows the ITB_TAG register format.
Figure 5–1 Istream Translation Buffer Tag Register (ITB_TAG)
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
VA<42:13>

IGN

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
IGN

VA<42:13>

LJ-03473-TI0

Preliminary—Subject to Change—July 1996 5–5

5.1 Instruction Fetch/Decode Unit and Branch Unit (Ibox) IPRs

5.1.2 Instruction Translation Buffer Page Table Entry (ITB_PTE)
Register
ITB_PTE is a read/write register.
Write Format
A write operation to this register writes both the PTE and TAG fields of an
ITB location determined by a not-last-used replacement algorithm. The TAG
and PTE fields are updated simultaneously to ensure the integrity of the ITB.
A write operation to the ITB_PTE register increments the not-last-used (NLU)
pointer, which allows for writing the entire set of ITB PTE and TAG entries.
If the HW_MTPR ITB_PTE instruction falls in the shadow of a trapping
instruction, the NLU pointer may be incremented multiple times. The TAG
field of the ITB location is determined by the contents of the ITB_TAG register.
The PTE field is provided by the HW_MTPR ITB_PTE instruction. Write
operations to this register use the memory format bits, as described in the
Alpha Architecture Reference Manual. Figure 5–2 shows the ITB_PTE register
write format.
Figure 5–2 Instruction Translation Buffer Page Table Entry (ITB_PTE)
Register Write Format
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
IGN

IGN
ASM
GH
IGN
KRE
ERE
SRE
URE

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
IGN

PFN<39:13>

LJ-03474-TI0

Read Format
A read of the ITB_PTE requires two instructions. A read of the ITB_PTE
register returns the PTE pointed to by the NLU pointer to the ITB_PTE_
TEMP register and increments the NLU pointer. If the HW_MFPR ITB_PTE
instruction falls in the shadow of a trapping instruction, the NLU pointer may
be incremented multiple times. A zero value is returned to the integer register
file. A second read of the ITB_PTE_TEMP register returns the PTE to the

5–6 Preliminary—Subject to Change—July 1996

5.1 Instruction Fetch/Decode Unit and Branch Unit (Ibox) IPRs

general purpose integer register file (IRF). Figure 5–3 shows the ITB_PTE
register read format.
Figure 5–3 Instruction Translation Buffer Page Table Entry (ITB_PTE) Register Read
Format
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
RAZ

RAZ

RAZ
ASM
KRE
ERE
SRE
URE
GHD<2:0>

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
RAZ

PFN<39:13>

LJ-03475-TI0

Preliminary—Subject to Change—July 1996 5–7

5.1 Instruction Fetch/Decode Unit and Branch Unit (Ibox) IPRs

5.1.3 Instruction Translation Buffer Address Space Number (ITB_ASN)
Register
ITB_ASN is a read/write register that contains the address space number
(ASN) of the current process. Figure 5–4 shows the ITB_ASN register format.
Figure 5–4 Instruction Translation Buffer Address Space Number (ITB_ASN)
Register
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
RAZ/IGN

ASN<6:0>

RAZ/IGN

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
RAZ/IGN

LJ-03476-TI0

5–8 Preliminary—Subject to Change—July 1996

5.1 Instruction Fetch/Decode Unit and Branch Unit (Ibox) IPRs

5.1.4 Instruction Translation Buffer Page Table Entry Temporary
(ITB_PTE_TEMP) Register
ITB_PTE_TEMP is a read-only holding register for ITB_PTE read data. A read
of the ITB_PTE register returns data to this register. A second read of the
ITB_PTE_TEMP register returns data to the general purpose integer register
file (IRF). Figure 5–3 shows the ITB_PTE register format.
Table 5–2 shows the GHD settings for the ITB_PTE_TEMP register.
Table 5–2 Granularity Hint Bits in ITB_PTE_TEMP Read Format
Name

Extent

Type

Description

GHD

<29>

Set if granularity hint equals 01, 10, or 11.

GHD

<30>

Set if granularity hint equals 10 or 11.

GHD

<31>

Set if granularity hint equals 11.

5.1.5 Instruction Translation Buffer Invalidate All Process (ITB_IAP)
Register
ITB_IAP is a write-only register. Any write operation to this register
invalidates all ITB entries that have an address space match (ASM) bit
that equals zero.

5.1.6 Instruction Translation Buffer Invalidate All (ITB_IA) Register
ITB_IA is a write-only register. A write operation to this register invalidates
all ITB entries, and resets the ITB not-last-used (NLU) pointer to its initial
state. RESET PALcode must execute an HW_MTPR ITB_IA instruction in
order to initialize the NLU pointer.

Preliminary—Subject to Change—July 1996 5–9

5.1 Instruction Fetch/Decode Unit and Branch Unit (Ibox) IPRs

5.1.7 Instruction Translation Buffer IS (ITB_IS) Register
ITB_IS is a write-only register. Writing a virtual address to this register
invalidates the ITB entry that meets either of the following criteria:
•

An ITB entry whose virtual address (VA) field matches ITB_IS<42:13> and
whose ASN field matches ITB_ASN<10:04>.

•

An ITB entry whose VA field matches ITB_IS<42:13> and whose ASM bit
is set.

Figure 5–5 shows the ITB_IS register format.
Figure 5–5 Instruction Translation Buffer IS (ITB_IS) Register
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
VA<42:13>

IGN

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
IGN

VA<42:13>

LJ-03478-TI0

5–10 Preliminary—Subject to Change—July 1996

5.1 Instruction Fetch/Decode Unit and Branch Unit (Ibox) IPRs

5.1.8 Formatted Faulting Virtual Address (IFAULT_VA_FORM) Register
IFAULT_VA_FORM is a read-only register containing the formatted faulting
virtual address on an ITBMISS/IACCVIO (except on IACCVIOs generated
by sign-check errors). The formatted faulting address generated depends
on whether NT superpage mapping is enabled through ICSR bit SPE<0>.
Figure 5–6 shows the IFAULT_VA_FORM register format in non-NT mode.
Figure 5–6 Formatted Faulting Virtual Address (IFAULT_VA_FORM) Register
(NT_Mode=0)
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
VA<42:13>

RAZ

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
VPTB<63:33>
VA<42:13>
LJ-03479-TI0

Figure 5–7 shows the IFAULT_VA_FORM register format in NT mode.
Figure 5–7 Formatted Faulting Virtual Address (IFAULT_VA_FORM) Register
(NT_Mode=1)
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
RAZ

VA<31:13>

RAZ
VPTB<63:30>

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
VPTB<63:30>

LJ-03480-TI0

Preliminary—Subject to Change—July 1996 5–11

5.1 Instruction Fetch/Decode Unit and Branch Unit (Ibox) IPRs

5.1.9 Virtual Page Table Base Register (IVPTBR)
IVPTBR is a read/write register. Bits <32:30> are UNDEFINED on a read of
this register in non-NT mode. Figure 5–8 shows the IVPTBR format in non-NT
mode.
Figure 5–8 Virtual Page Table Base Register (IVPTBR) (NT_Mode=0)
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
IGN

RAZ/IGN

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
VPTB<63:33>

I
G
N
MA0602

Figure 5–9 shows the IVPTBR format in NT mode.
Figure 5–9 Virtual Page Table Base Register (IVPTBR) (NT_Mode=1)
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
RAZ/IGN
VPTB<63:30>
63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
VPTB<63:30>
LJ-03481-TI0

5–12 Preliminary—Subject to Change—July 1996

5.1 Instruction Fetch/Decode Unit and Branch Unit (Ibox) IPRs

5.1.10 Icache Parity Error Status (ICPERR_STAT) Register
ICPERR_STAT is a read/write register. The Icache parity error status bits may
be cleared by writing a 1 to the appropriate bits. Figure 5–10 and Table 5–3
describe the ICPERR_STAT register format.
Figure 5–10 Icache Parity Error Status (ICPERR_STAT) Register
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
RAZ/IGN

RAZ/IGN
DPE
TPE
TMR

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
RAZ/IGN

LJ-03482-TI0

Table 5–3 Icache Parity Error Status Register Fields
Name

Extent

Type

Description

DPE

<11>

W1C

Data parity error

TPE

<12>

W1C

Tag parity error

TMR

<13>

W1C

Timeout reset error or cfail_h/no cack_h error

5.1.11 Icache Flush Control (IC_FLUSH_CTL) Register
IC_FLUSH_CTL is a write-only register. Writing any value to this register
flushes the entire Icache.

Preliminary—Subject to Change—July 1996 5–13

5.1 Instruction Fetch/Decode Unit and Branch Unit (Ibox) IPRs

5.1.12 Exception Address (EXC_ADDR) Register
EXC_ADDR is a read/write register used to restart the system after exceptions
or interrupts. The HW_REI instruction causes a return to the instruction
pointed to by the EXC_ADDR register. This register can be written both
by hardware and software. Hardware write operations occur as a result of
exceptions/interrupts and CALL_PAL instructions. Hardware write operations
that occur as a result of exceptions/interrupts take precedence over all other
write operations.
In case of an exception/interrupt, hardware writes a program counter
(PC) to this register. In case of precise exceptions, this is the PC
value of the instruction that caused the exception. In case of imprecise
exceptions/interrupts, this is the PC value of the next instruction that would
have issued if the exception/interrupt was not reported.
In case of a CALL_PAL instruction, the PC value of the next instruction after
the CALL_PAL is written to EXC_ADDR.
Bit <00> of this register is used to indicate PALmode. On a HW_REI
instruction, the mode of the system is determined by bit <00> of EXC_ADDR.
Figure 5–11 shows the EXC_ADDR register format.
Figure 5–11 Exception Address (EXC_ADDR) Register

31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
PC<63:2>
PAL
RAZ/IGN

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
PC<63:2>

LJ-03483-TI0

5–14 Preliminary—Subject to Change—July 1996

5.1 Instruction Fetch/Decode Unit and Branch Unit (Ibox) IPRs

5.1.13 Exception Summary (EXC_SUM) Register
EXC_SUM is a read/write register that records the different arithmetic traps
that occur between EXC_SUM write operations. Any write operation to this
register clears bits <16:10>. Figure 5–12 and Table 5–4 describe the EXC_SUM
register format.
Figure 5–12 Exception Summary (EXC_SUM) Register
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
RAZ/IGN

RAZ/IGN
SWC
INV
DZE
FOV
UNF
INE
IOV

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
RAZ/IGN

LJ-03484-TI0

Table 5–4 Exception Summary Register Fields
Name

Extent

Type

Description

SWC

<10>

Indicates software completion possible. This bit
is set after a floating-point instruction containing
the /S modifier completes with an arithmetic trap
and if all previous floating-point instructions that
trapped since the last HW_MTPR EXC_SUM
instruction also contained the /S modifier.
The SWC bit is cleared whenever a floating-point
instruction without the /S modifier completes
with an arithmetic trap. The bit remains cleared
regardless of additional arithmetic traps until the
register is written by an HW_MTPR instruction.
The bit is always cleared upon any HW_MTPR
write operation to the EXC_SUM register.
(continued on next page)

Preliminary—Subject to Change—July 1996 5–15

5.1 Instruction Fetch/Decode Unit and Branch Unit (Ibox) IPRs

Table 5–4 (Cont.) Exception Summary Register Fields
Name

Extent

Type

Description

INV

<11>

Indicates invalid operation.

DZE

<12>

Indicates divide by zero.

FOV

<13>

Indicates floating-point overflow.

UNF

<14>

Indicates floating-point underflow.

INE

<15>

Indicates floating inexact error.

IOV

<16>

Indicates floating-point execution unit (Fbox)
convert to integer overflow or integer arithmetic
overflow.

5–16 Preliminary—Subject to Change—July 1996

5.1 Instruction Fetch/Decode Unit and Branch Unit (Ibox) IPRs

5.1.14 Exception Mask (EXC_MASK) Register
EXC_MASK is a read/write register that records the destinations of
instructions that have caused an arithmetic trap between EXC_MASK write
operations. The destination is recorded as a single bit mask in the 64-bit IPR
representing F0–F31 and I0–I31. A write operation to EXC_SUM clears the
EXC_MASK register. Figure 5–13 shows the EXC_MASK register format.
Figure 5–13 Exception Mask (EXC_MASK) Register
31
131130129 . . .

63
F31F30 F29 . . .

00
I1 I0

32
F1 F0

LJ-03485-TI0

Preliminary—Subject to Change—July 1996 5–17

5.1 Instruction Fetch/Decode Unit and Branch Unit (Ibox) IPRs

5.1.15 PAL Base Address (PAL_BASE) Register
PAL_BASE is a read/write register containing the base address for PALcode.
The register is cleared by hardware on reset. Figure 5–14 shows the
PAL_BASE register format.
Figure 5–14 PAL Base Address (PAL_BASE) Register

31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
PAL_BASE<39:14>

RAZ/IGN

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
RAZ/IGN

PAL_BASE<39:14>

LJ-03486-TI0

5–18 Preliminary—Subject to Change—July 1996

5.1 Instruction Fetch/Decode Unit and Branch Unit (Ibox) IPRs

5.1.16 Ibox Current Mode (ICM) Register
ICM is a read/write register containing the current mode bits of the
architecturally defined processor status, as described in the Alpha Architecture
Reference Manual. Figure 5–15 shows the ICM register format.
Figure 5–15 Ibox Current Mode (ICM) Register
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
RAZ/IGN
RAZ/IGN
CM0
CM1

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
RAZ/IGN

LJ-03487-TI0

Preliminary—Subject to Change—July 1996 5–19

5.1 Instruction Fetch/Decode Unit and Branch Unit (Ibox) IPRs

5.1.17 Ibox Control and Status Register (ICSR)
ICSR is a read/write register containing Ibox-related control and status
information. Figure 5–16 and Table 5–5 describe ICSR format.
Figure 5–16 Ibox Control and Status Register (ICSR)
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
RAZ/IGN

RAZ/IGN

PME<1:0>
IMSK<3:0>
TMM
TMD
FPE
HWE
SPE<1:0>
SDE
RAZ/IGN
63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
RAZ/IGN
CRDE
SLE
FMS
FBT
FBD
MBO
ISTA
TST
LJ-03488-TI0

5–20 Preliminary—Subject to Change—July 1996

5.1 Instruction Fetch/Decode Unit and Branch Unit (Ibox) IPRs

Table 5–5 Ibox Control and Status Register Fields
Name

Extent

Type

Description

PME<1:0>

<09:08>

RW,0

Performance counter master enable bits. If both
PME<1> and PME<0> are clear, all performance
counters in the PMCTR IPR are disabled. If
either PME<1> or PME<0> are set, the counter is
enabled according to the settings of the PMCTR
CTL fields.

IMSK<3:0> <23:20>

RW,0

If set, each IMSK<3:0> signal disables the
corresponding IRQ_H<3:0> interrupt.

TMM

<24>

RW,0

If set, the timeout counter counts 5 thousand
cycles before asserting timeout reset. If clear,
the timeout counter counts 1 billion cycles before
asserting timeout reset.

TMD

<25>

RW,0

If set, disables the Ibox timeout counter. Does
not affect cfail_h/no cack_h error.

FPE

<26>

RW,0

If set, floating-point instructions may be issued.
If clear, floating-point instructions cause FEN
exceptions.

HWE

<27>

RW,0

If set, allows PALRES instructions to be issued
in kernel mode.

SPE<1:0>

<29:28>

RW,0

21164–266, 21164–300, and 21164–333
If SPE<1> is set, it enables superpage mapping
of Istream virtual address VA<39:13> directly
to physical address PA<39:13> assuming
VA<42:41> = 10. Virtual address bit VA<40>
is ignored in this translation. Access is allowed
only in kernel mode.
If SPE<0> is set (NT mode), it enables
superpage mapping of Istream virtual addresses
VA<42:30> = 1FFE16 directly to physical address
PA<39:30> = 016 . VA<30:13> is mapped directly
to PA<30:13>. Access is allowed only in kernel
mode.
21164–P1 and 21164–P2
SPE<0> must always be set. Clearing this
bit will cause 21164–Pn operation to be
UNPREDICTABLE.
(continued on next page)

Preliminary—Subject to Change—July 1996 5–21

5.1 Instruction Fetch/Decode Unit and Branch Unit (Ibox) IPRs

Table 5–5 (Cont.) Ibox Control and Status Register Fields
Name

Extent

Type

Description

SDE

<30>

RW,0

If set, enables PAL shadow registers.

CRDE

<32>

RW,0

If set, enables correctable error interrupts.

SLE

<33>

RW,0

If set, enables serial line interrupts.

FMS

<34>

RW,0

If set, forces miss on Icache references. MBZ in
normal operation.

FBT

<35>

RW,0

If set, forces bad Icache tag parity. MBZ in
normal operation.

FBD

<36>

RW,0

If set, forces bad Icache data parity. MBZ in
normal operation.

Reserved

<37>

RW,1

Reserved to Digital. Must be one.

ISTA

<38>

Reading this bit indicates ICACHE BIST status.
If set, ICACHE BIST was successful.

TST

<39>

RW,0

Writing a 1 to this bit asserts the
test_status_h<1> signal.

5–22 Preliminary—Subject to Change—July 1996

5.1 Instruction Fetch/Decode Unit and Branch Unit (Ibox) IPRs

5.1.18 Interrupt Priority Level Register (IPLR)
IPLR is a read/write register that is accessed by PALcode to set the value of
the interrupt priority level (IPL). Whenever hardware detects an interrupt
whose target IPL is greater than the value in IPLR<04:00>, an interrupt is
taken. Figure 5–17 shows the IPLR register format. Refer to Table 4–19 for
information on which interrupts are enabled for a given IPL.
Figure 5–17 Interrupt Priority Level Register (IPLR)
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
RAZ/IGN

IPL<4:0>

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
RAZ/IGN

LJ-03489-TI0

Preliminary—Subject to Change—July 1996 5–23

5.1 Instruction Fetch/Decode Unit and Branch Unit (Ibox) IPRs

5.1.19 Interrupt ID (INTID) Register
INTID is a read-only register that is written by hardware with the target
IPL of the highest priority pending interrupt. The hardware recognizes an
interrupt if the IPL being read is greater than the IPL given by IPLR<04:00>.
Interrupt service routines may use the value of this register to determine the
cause of the interrupt. PALcode, for the interrupt service, must ensure that
the IPL in INTID is greater than the IPL specified by IPLR. This restriction
is required because a level-sensitive hardware interrupt may disappear before
the interrupt service routine is entered (passive release).
The contents of INTID are not correct on a HALT interrupt because this
particular interrupt does not have a target IPL at which it can be masked.
When a HALT interrupt occurs, INTID indicates the next highest priority
pending interrupt. PALcode for interrupt service must check the interrupt
summary register (ISR) to determine if a HALT interrupt has occurred.
Figure 5–18 shows the INTID register format.
Figure 5–18 Interrupt ID (INTID) Register
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
RAZ/IGN

INTID<4:0>

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
RAZ/IGN

LJ-03490-TI0

5–24 Preliminary—Subject to Change—July 1996

5.1 Instruction Fetch/Decode Unit and Branch Unit (Ibox) IPRs

5.1.20 Asynchronous System Trap Request Register (ASTRR)
ASTRR is a read/write register containing bits to request asynchronous system
trap (AST) interrupts in each of the four processor modes (U,S,E,K). In order to
generate an AST interrupt, the corresponding enable bit in the ASTER must be
set and the current processor mode given in the ICM<04:03> should be equal to
or higher than the mode associated with the AST request. Figure 5–19 shows
the ASTRR format.
Figure 5–19 Asynchronous System Trap Request Register (ASTRR)
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
RAZ/IGN
KAR
EAR
SAR
UAR

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
RAZ/IGN

LJ-03491-TI0

Preliminary—Subject to Change—July 1996 5–25

5.1 Instruction Fetch/Decode Unit and Branch Unit (Ibox) IPRs

5.1.21 Asynchronous System Trap Enable Register (ASTER)
ASTER is a read/write register containing bits to enable corresponding
asynchronous system trap (AST) interrupt requests. Figure 5–20 shows the
ASTER format.
Figure 5–20 Asynchronous System Trap Enable Register (ASTER)
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
RAZ/IGN
KAE
EAE
SAE
UAE

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
RAZ/IGN

LJ-03492-TI0

5–26 Preliminary—Subject to Change—July 1996

5.1 Instruction Fetch/Decode Unit and Branch Unit (Ibox) IPRs

5.1.22 Software Interrupt Request Register (SIRR)
SIRR is a read/write register used to control software interrupt requests.
A software request for a particular IPL may be requested by setting the
appropriate bit in SIRR<15:01>. Figure 5–21 and Table 5–6 describe the SIRR
format.
Figure 5–21 Software Interrupt Request Register (SIRR)

31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
RAZ/IGN

SIRR<15:1>

RAZ/IGN

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
RAZ/IGN

LJ-03493-TI0

Table 5–6 Software Interrupt Request Register Fields
Name

Extent

Type

Description

SIRR<15:1>

<18:04>

Request software interrupts.

Preliminary—Subject to Change—July 1996 5–27

5.1 Instruction Fetch/Decode Unit and Branch Unit (Ibox) IPRs

5.1.23 Hardware Interrupt Clear (HWINT_CLR) Register
HWINT_CLR is a write-only register used to clear edge-sensitive hardware
interrupt requests. Figure 5–22 and Table 5–7 describe the HWINT_CLR
register format.
Figure 5–22 Hardware Interrupt Clear (HWINT_CLR) Register
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
IGN

IGN
PC0C
PC1C
PC2C

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
IGN
CRDC
SLC
LJ-03495-TI0

Table 5–7 Hardware Interrupt Clear Register Fields
Name

Extent

Type

Description

PC0C

<27>

W1C

Clears performance counter 0 interrupt requests.

PC1C

<28>

W1C

Clears performance counter 1 interrupt requests.

PC2C

<29>

W1C

Clears performance counter 2 interrupt requests.

CRDC

<32>

W1C

Clears correctable read data interrupt requests.

SLC

<33>

W1C

Clears serial line interrupt requests.

5–28 Preliminary—Subject to Change—July 1996

5.1 Instruction Fetch/Decode Unit and Branch Unit (Ibox) IPRs

5.1.24 Interrupt Summary Register (ISR)
ISR is a read-only register containing information about all pending hardware,
software, and asynchronous system trap (AST) interrupt requests. Figure 5–23
and Table 5–8 describe the ISR format. Refer to Table 4–19 for a description of
which interrupts are enabled for a given interrupt priority level (IPL).
Figure 5–23 Interrupt Summary Register (ISR)
31 30 29 28 27 26 25 24 23 22 21 20 19 18

04 03

SISR<15:1>

RAZ

ASTRR<3:0>
and ASTER<3:0>
ATR
I20
I21
I22
I23
PC0
PC1
PC2
PFL
MCK
63

32
RAZ
CRD
SLI
HLT
LJ-03496-TI0A

Preliminary—Subject to Change—July 1996 5–29

5.1 Instruction Fetch/Decode Unit and Branch Unit (Ibox) IPRs

Table 5–8 Interrupt Summary Register Fields
Name

Extent

Type

Description

ASTRR<3:0> <03:00>
and
ASTER<3:0>

Boolean AND of ASTRR<USEK> with
ASTER<USEK> used to indicate enabled AST
requests.

SISR<15:1>

<18:04>

RO,0

Software interrupt requests 15 through 1
corresponding to IPL 15 through 1.

ATR

<19>

Set if any AST request and corresponding
enable bit is set and if the processor mode is
equal to or higher than the AST request mode.

I20

<20>

External hardware interrupt—irq_h<0>.

I21

<21>

External hardware interrupt—irq_h<1>.

I22

<22>

External hardware interrupt—irq_h<2>.

I23

<23>

External hardware interrupt—irq_h<3>.

PC0

<27>

External hardware interrupt—performance
counter 0 (IPL 29).

PC1

<28>

External hardware interrupt—performance
counter 1 (IPL 29).

PC2

<29>

External hardware interrupt—performance
counter 2 (IPL 29).

PFL

<30>

External hardware interrupt—power failure
(IPL 30).

MCK

<31>

External hardware interrupt—system machine
check (IPL 31).

CRD

<32>

Correctable ECC errors (IPL 31).

SLI

<33>

Serial line interrupt.

HLT

<34>

External hardware interrupt—halt.

5–30 Preliminary—Subject to Change—July 1996

5.1 Instruction Fetch/Decode Unit and Branch Unit (Ibox) IPRs

5.1.25 Serial Line Transmit (SL_XMIT) Register
SL_XMIT is a write-only register used to transmit bit-serial data out of
the microprocessor chip under the control of a software timing loop. The
value of the TMT bit is transmitted offchip on the srom_clk_h signal. In
normal operation mode (not in debugging mode), the srom_clk_h signal
serves both the serial line transmission and the Icache serial ROM interface
(see Section 7.5). Figure 5–24 and Table 5–9 describe the SL_XMIT register
format.
Figure 5–24 Serial Line Transmit (SL_XMIT) Register
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
IGN

IGN
TMT

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
IGN

LJ-03497-TI0

Table 5–9 Serial Line Transmit Register Fields
Name

Extent

Type

Description

TMT

<07>

WO,1

Serial line transmit data

Preliminary—Subject to Change—July 1996 5–31

5.1 Instruction Fetch/Decode Unit and Branch Unit (Ibox) IPRs

5.1.26 Serial Line Receive (SL_RCV) Register
SL_RCV is a read-only register used to receive bit-serial data under the control
of a software timing loop. The RCV bit in the SL_RCV register is functionally
connected to the srom_data_h signal. A serial line interrupt is requested
whenever a transition is detected on the srom_data_h signal and the SLE
bit in the ICSR is set. During normal operations (not in test mode), the
srom_data_h signal serves both the serial line reception and the Icache serial
ROM (SROM) interface (see Section 7.5). Figure 5–25 and Table 5–10 describe
the SL_RCV register format.
Figure 5–25 Serial Line Receive (SL_RCV) Register
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
RAZ

RAZ
RCV

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
RAZ

LJ-03498-TI0

Table 5–10 Serial Line Receive Register Fields
Name

Extent

Type

Description

RCV

<06>

Serial line receive data

5–32 Preliminary—Subject to Change—July 1996

5.1 Instruction Fetch/Decode Unit and Branch Unit (Ibox) IPRs

5.1.27 Performance Counter (PMCTR) Register
PMCTR is a read/write register that controls the three onchip performance
counters. Figure 5–26 and Table 5–11 describe the PMCTR format.
Performance counter interrupt requests are summarized in Section 5.1.24.
Cbox inputs to the counter select options are described in Table 5–31.
Section 2.8 describes the performance measurement support features.
Note
The arrangement of the select option tables is not meant to imply any
restrictions on permitted combinations of selections. The only cases
in which the selection for one counter influences another’s count is
SEL1=8 (SEL 2=2, 3, other).

Figure 5–26 Performance Counter (PMCTR) Register

31 30 29
K
u

16 15 14 13 12 11 10 09 08 07
CTR2<13:0>

CTL0 CTL1 CTL2

04 03

K K
SEL1<3:0> SEL2<3:0>
p k

SEL0

48 47
CTR0<15:0>

32
CTR1<15:0>

MA-0601A

Preliminary—Subject to Change—July 1996 5–33

5.1 Instruction Fetch/Decode Unit and Branch Unit (Ibox) IPRs

Table 5–11 Performance Counter Register Fields
Name

Extent

Type

Description

CTR0<15:0>

<63:48>

A 16-bit counter of events selected by SEL0 and
enabled by CTL0<1:0>.

CTR1<15:0>

<47:32>

A 16-bit counter.

SEL0

<31>

Counter0 Select—refer to Table 5–12.

<30>

Kill user mode—disables all counters in user
mode (refer to Table 5–13).

CTR2<13:0>

<29:16>

14-bit counter

CTL0<1:0>

<15:14>

RW,0

CTR0 counter control:
00 counter disable, interrupt disable
01 counter enable, interrupt disable
10 counter enable, interrupt at count 65536
(Refer to Section 5.1.23 and Section 5.1.24.)
11 counter enable, interrupt at count 256

CTL1<1:0>

<13:12>

RW,0

CTR1 counter control:
00 counter disable,interrupt disable
01 counter enable, interrupt disable
10 counter enable, interrupt at count 65536
11 counter enable, interrupt at count 256

CTL2<1:0>

<11:10>

RW,0

CTR2 counter control:
00 counter disable,interrupt disable
01 counter enable, interrupt disable
10 counter enable, interrupt at count 16384
11 counter enable, interrupt at count 256

<09>

Kill PALmode—disables all counters in
PALmode (refer to Table 5–13).

<08>

Kill kernel, executive, supervisor mode—
disables all counters in kernel, executive,
and supervisor modes (refer to Table 5–13).
Ku=1, Kp=1, and Kk=1 enables counters in
executive and supervisor modes only.

SEL1<3:0>

<07:04>

Counter1 Select—refer to Table 5–12.

SEL2<3:0>

<03:00>

Counter2 Select—refer to Table 5–12.

5–34 Preliminary—Subject to Change—July 1996

5.1 Instruction Fetch/Decode Unit and Branch Unit (Ibox) IPRs

Table 5–12 shows the PMCTR counter select options.
Table 5–12 PMCTR Counter Select Options
Counter0
SEL0<0>

Counter1
SEL1<3:0>

Counter2
SEL2<3:0>

0:Cycles

0x0: nonissue cycles
Valid instruction in S3 but none
issued.

0x0: long(>15 cycle) stalls

0x1: split-issue cycles
Some, but not all, instructions at
S3 issued.

0x1: reserved

0x2: pipe-dry cycles
No valid instruction at S3.
0x3: replay trap
A replay trap occurred.
0x4: single-issue cycles
Exactly one instruction issued.
0x5: dual-issue cycles
Exactly two instructions issued.
0x6: triple-issue cycles
Exactly three instructions
issued.
0x7: quad-issue cycles
Exactly four instructions issued.
1:Instructions

0x8: jsr-ret if sel2=PC-M
Instruction issued if sel2 is
PC-M.

0x2: PC-mispredicts

0x8: cond-branch if sel2=BR-M
Instruction issued if sel2 is
BR-M

0x3: BR-mispredicts

0x8: all flow-change instructions
if sel2=! (PC-M or BR-M)
0x9: IntOps issued

0x4: Icache/RFB misses

0xA: FPOps issued

0x5: ITB misses

0xB: loads issued

0x6: Dcache LD misses

0xC: stores issued

0x7: DTB misses

0xD: Icache issued

0x8: LDs merged in MAF
(continued on next page)

Preliminary—Subject to Change—July 1996 5–35

5.1 Instruction Fetch/Decode Unit and Branch Unit (Ibox) IPRs

Table 5–12 (Cont.) PMCTR Counter Select Options
Counter0
SEL0<0>

Counter1
SEL1<3:0>
0xE: Dcache accesses

Counter2
SEL2<3:0>
0x9: LDU replay traps
0xA:WB/MAF full replay traps
0xB: external perf_mon_h
input. This counts in CPU
cycles, but input is sampled
in sysclk cycles. The external
status perf_mon_h is sampled
once per system clock and held
through the system clock period.
This means that ‘‘sysclock ratio’’
counts occur for each system
clock cycle in which the status is
true.
0xC: CPU cycles
0xD: MB stall cycles
0xE: LDxL instructions issued

0xF: pick CBOX input 1

5–36 Preliminary—Subject to Change—July 1996

0xF: pick CBOX input 2

5.1 Instruction Fetch/Decode Unit and Branch Unit (Ibox) IPRs

Table 5–13 Measurement Mode Control
Kill Bit Settings
Measurement Mode Desired

Program

PAL only

OS only (kernel, executive,
supervisor)

User only

All except PAL

OS + PAL (not user)

User + PAL (not kernel,
executive, and supervisor)

Executive and supervisor only1

1 In this instance, Kk means kill kernel only.

The combination Ku=1, Kp=1, and Kk=1 is used to
gather events for the executive and supervisor modes only.

Note
Both the user and the operating system can make PAL subroutine
calls that put the machine in PALmode. The ‘‘OS only,’’ ‘‘user only,’’
and ‘‘executive and supervisor only’’ modes do not measure the events
during the PAL subroutine calls made by the OS or user. The ‘‘OS +
PAL’’ and ‘‘user + PAL’’ modes should be used carefully. ‘‘OS + PAL’’
mode measures the events during the PAL calls made by the user,
whereas ‘‘user + PAL’’ mode measures the events during the PAL calls
made by the OS.

Preliminary—Subject to Change—July 1996 5–37

5.2 Memory Address Translation Unit (Mbox) IPRs

5.2 Memory Address Translation Unit (Mbox) IPRs
The Mbox internal processor registers (IPRs) are described in Section 5.2.1
through Section 5.2.23.

5.2.1 Dstream Translation Buffer Address Space Number (DTB_ASN)
Register
DTB_ASN is a write-only register that must be written with an exact duplicate
of the ITB_ASN register ASN field. Figure 5–27 shows the DTB_ASN register
format.
Figure 5–27 Dstream Translation Buffer Address Space Number (DTB_ASN)
Register
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
IGN

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
ASN<6:0>

IGN

LJ-03499-TI0

5–38 Preliminary—Subject to Change—July 1996

5.2 Memory Address Translation Unit (Mbox) IPRs

5.2.2 Dstream Translation Buffer Current Mode (DTB_CM) Register
DTB_CM is a write-only register that must be written with an exact duplicate
of the Ibox current mode (ICM) register CM field. These bits indicate the
current mode of the machine, as described in the Alpha Architecture Reference
Manual. Figure 5–28 shows the DTB_CM register format.
Figure 5–28 Dstream Translation Buffer Current Mode (DTB_CM) Register
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
IGN

IGN
CM0
CM1

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
IGN

LJ-03500-TI0

Preliminary—Subject to Change—July 1996 5–39

5.2 Memory Address Translation Unit (Mbox) IPRs

5.2.3 Dstream Translation Buffer Tag (DTB_TAG) Register
DTB_TAG is a write-only register that writes the DTB tag and the contents
of the DTB_PTE register to the DTB. To ensure the integrity of the DTBs,
the DTB’s PTE array is updated simultaneously from the internal DTB_PTE
register when the DTB_TAG register is written.
The entry to be written is chosen at the time of the DTB_TAG write operation
by a not-last-used replacement algorithm implemented in hardware. A write
operation to the DTB_TAG register increments the translation buffer (TB)
entry pointer of the DTB, which allows writing the entire set of DTB PTE and
TAG entries. The TB entry pointer is initialized to entry zero and the TB valid
bits are cleared on chip reset but not on timeout reset. Figure 5–29 shows the
DTB_TAG register format.
Figure 5–29 Dstream Translation Buffer Tag (DTB_TAG) Register

31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
VA<42:13>

IGN

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
IGN

VA<42:13>

LJ-03501-TI0

5–40 Preliminary—Subject to Change—July 1996

5.2 Memory Address Translation Unit (Mbox) IPRs

5.2.4 Dstream Translation Buffer Page Table Entry (DTB_PTE) Register
DTB_PTE is a read/write register representing the 64-entry DTB page
table entries (PTEs). The entry to be written is chosen by a not-last-used
replacement algorithm implemented in hardware. Write operations to
DTB_PTE use the memory format bit positions, as described in the Alpha
Architecture Reference Manual, with the exception that some fields are ignored.
In particular, the page frame number (PFN) valid bit is not stored in the DTB.
To ensure the integrity of the DTB, the PTE is actually written to a temporary
register and is not transferred to the DTB until the DTB_TAG register is
written. As a result, writing the DTB_PTE and then reading without an
intervening DTB_TAG write operation does not return the data previously
written to the DTB_PTE register.
Read operations of the DTB_PTE require two instructions. First, a read
from the DTB_PTE sends the PTE data to the DTB_PTE_TEMP register. A
zero value is returned to the integer register file (IRF) on a DTB_PTE read
operation. A second instruction reading from the DTB_PTE_TEMP register
returns the PTE entry to the register file. Reading the DTB_PTE register
increments the TB entry pointer of the DTB, which allows reading the entire
set of DTB PTE entries. Figure 5–30 shows the DTB_PTE register format.
Note
The Alpha Architecture Reference Manual provides descriptions of the
fields of the PTE.

Preliminary—Subject to Change—July 1996 5–41

5.2 Memory Address Translation Unit (Mbox) IPRs

Figure 5–30 Dstream Translation Buffer Page Table Entry (DTB_PTE) Register—Write
Format
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
IGN
IGN
FOR
FOW
IGN
ASM
GH<1:0>
IGN
KRE
ERE
SRE
URE
KWE
EWE
SWE
UWE

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
IGN

PFN<39:13>

LJ-03502-TI0

5–42 Preliminary—Subject to Change—July 1996

5.2 Memory Address Translation Unit (Mbox) IPRs

5.2.5 Dstream Translation Buffer Page Table Entry Temporary
(DTB_PTE_TEMP) Register
DTB_PTE_TEMP is a read-only holding register used for DTB_PTE data. Read
operations of the DTB_PTE require two instructions to return the PTE data to
the register file. The first reads the DTB_PTE register to the DTB_PTE_TEMP
register and returns zero to the register file. The second returns the DTB_
PTE_TEMP register to the integer register file (IRF). Figure 5–31 shows the
DTB_PTE_TEMP register format.
Figure 5–31 Dstream Translation Buffer Page Table Entry Temporary (DTB_PTE_TEMP)
Register
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
PFN<39:13>

RAZ
FOR
FOW
KRE
ERE
SRE
URE
KWE
EWE
SWE
UWE
PFN<39:13>

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
RAZ

PFN<39:13>

LJ-03503-TI0

Preliminary—Subject to Change—July 1996 5–43

5.2 Memory Address Translation Unit (Mbox) IPRs

5.2.6 Dstream Memory Management Fault Status (MM_STAT) Register
MM_STAT is a read-only register that stores information on Dstream faults
and Dcache parity errors. The VA, VA_FORM, and MM_STAT registers are
locked against further updates until software reads the VA register. The
MM_STAT bits are only modified by hardware when the register is not locked
and a memory management error, DTB miss, or Dcache parity error occurs.
The MM_STAT register is not unlocked or cleared on reset. Figure 5–32 and
Table 5–14 describe the MM_STAT register format.
Figure 5–32 Dstream Memory Management Fault Status (MM_STAT) Register
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
OPCODE

RAZ

RA
WR
ACV
FOR
FOW
DTB_MISS
BAD_VA

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
RAZ

LJ-03504-TI0

Table 5–14 Dstream Memory Management Fault Status Register Fields
Name

Extent

Type

Description

<00>

Set if reference that caused error was a write
operation.

ACV

<01>

Set if reference caused an access violation.
Includes bad virtual address.

FOR

<02>

Set if reference was a read operation and the
PTE FOR bit was set.

FOW

<03>

Set if reference was a write operation and the
PTE FOW bit was set.

DTB_MISS

<04>

Set if reference resulted in a DTB miss.

BAD_VA

<05>

Set if reference had a bad virtual address.
(continued on next page)

5–44 Preliminary—Subject to Change—July 1996

5.2 Memory Address Translation Unit (Mbox) IPRs

Table 5–14 (Cont.) Dstream Memory Management Fault Status Register
Fields
Name

Extent

Type

Description

<10:06>

RA field of the faulting instruction.

OPCODE

<16:11>

Opcode field of the faulting instruction.

Preliminary—Subject to Change—July 1996 5–45

5.2 Memory Address Translation Unit (Mbox) IPRs

5.2.7 Faulting Virtual Address (VA) Register
VA is a read-only register. When Dstream faults, DTB misses, or Dcache
parity errors occur, the effective virtual address associated with the fault, miss,
or error is latched in the VA register. The VA, VA_FORM, and MM_STAT
registers are locked against further updates until software reads the VA
register. The VA register is not unlocked on reset. Figure 5–33 shows the VA
register format.
Figure 5–33 Faulting Virtual Address (VA) Register
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
Virtual Address

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
Virtual Address

LJ-03505-TI0

5–46 Preliminary—Subject to Change—July 1996

5.2 Memory Address Translation Unit (Mbox) IPRs

5.2.8 Formatted Virtual Address (VA_FORM) Register
VA_FORM is a read-only register containing the virtual page table entry (PTE)
address calculated as a function of the faulting virtual address and the virtual
page table base (VA and MVPTBR registers). This is done as a performance
enhancement to the Dstream TBmiss PAL flow.
The virtual address is formatted as a 32-bit PTE when the NT_Mode bit
(MCSR<01>) is set (see Figure 5–34). VA_FORM is locked on any Dstream
fault, DTB miss, or Dcache parity error. The VA, VA_FORM, and MM_STAT
registers are locked against further updates until software reads the VA
register. The VA_FORM register is not unlocked on reset. Figure 5–35 shows
the VA_FORM register format when MCSR<01> is clear.
Figure 5–34 Formatted Virtual Address (VA_FORM) Register (NT_Mode=1)
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
VA<31:13>

RAZ

RAZ
VPTB<63:30>

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
VPTB<63:30>

LJ-03507-TI0

Figure 5–35 Formatted Virtual Address (VA_FORM) Register (NT_Mode=0)
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
VA<42:13>

RAZ

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
VPTB<63:33>
VA<42:13>
LJ-03506-TI0

Preliminary—Subject to Change—July 1996 5–47

5.2 Memory Address Translation Unit (Mbox) IPRs

Table 5–15 describes the VA_FORM register fields.
Table 5–15 Formatted Virtual Address Register Fields
Name

Extent

Type

Description

VPTB

<63:33>

Virtual page table base address as stored in
MVPTBR

VA<42:13>

<32:03>

Subset of the original faulting virtual address

VPTB

<63:30>

Virtual page table base address as stored in
MVPTBR

VA<31:13>

<21:03>

Subset of the original faulting virtual address

NT_Mode=0

NT_Mode=1

5–48 Preliminary—Subject to Change—July 1996

5.2 Memory Address Translation Unit (Mbox) IPRs

5.2.9 Mbox Virtual Page Table Base Register (MVPTBR)
MVPTBR is a write-only register containing the virtual address of the base of
the page table structure. It is stored in the Mbox to be used in calculating the
VA_FORM value for the Dstream TBmiss PAL flow. Unlike the VA register, the
MVPTBR is not locked against further updates when a Dstream fault, DTB
Miss, or Dcache parity error occurs. Figure 5–36 shows the MVPTBR format.
Figure 5–36 Mbox Virtual Page Table Base Register (MVPTBR)
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
IGN
VPTB<63:30>

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
VPTB<63:30>

LJ-03508-TI0

Preliminary—Subject to Change—July 1996 5–49

5.2 Memory Address Translation Unit (Mbox) IPRs

5.2.10 Dcache Parity Error Status (DC_PERR_STAT) Register
DC_PERR_STAT is a read/write register that locks and stores Dcache parity
error status. The VA, VA_FORM, and MM_STAT registers are locked against
further updates until software reads the VA register. If a Dcache parity error
is detected while the Dcache parity error status register is unlocked, the
error status is loaded into DC_PERR_STAT<05:02>. The LOCK bit is set and
the register is locked against further updates (except for the SEO bit) until
software writes a 1 to clear the LOCK bit.
The SEO bit is set when a Dcache parity error occurs while the Dcache parity
error status register is locked. Once the SEO bit is set, it is locked against
further updates until the software writes a 1 to DC_PERR_STAT<00> to
unlock and clear the bit. The SEO bit is not set when Dcache parity errors are
detected on both pipes within the same cycle. In this particular situation, the
pipe0/pipe1 Dcache parity error status bits indicate the existence of a second
parity error. The DC_PERR_STAT register is not unlocked or cleared on reset.
Figure 5–37 and Table 5–16 describe the DC_PERR_STAT register format.
Figure 5–37 Dcache Parity Error Status (DC_PERR_STAT) Register

31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
RAZ
SEO
LOCK
DP0
DP1
TP0
TP1

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
RAZ

LJ-03509-TI0

5–50 Preliminary—Subject to Change—July 1996

5.2 Memory Address Translation Unit (Mbox) IPRs

Table 5–16 Dcache Parity Error Status Register Fields
Name

Extent

Type

Description

SEO

<00>

W1C

Set if second Dcache parity error occurred in a
cycle after the register was locked. The SEO bit
is not set as a result of a second parity error that
occurs within the same cycle as the first.

LOCK

<01>

W1C

Set if parity error detected in Dcache. Bits
<05:02> are locked against further updates when
this bit is set. Bits <05:02> are cleared when the
LOCK bit is cleared.

DP0

<02>

Set on data parity error in Dcache bank 0.

DP1

<03>

Set on data parity error in Dcache bank 1.

TP0

<04>

Set on tag parity error in Dcache bank 0.

TP1

<05>

Set on tag parity error in Dcache bank 1.

Preliminary—Subject to Change—July 1996 5–51

5.2 Memory Address Translation Unit (Mbox) IPRs

5.2.11 Dstream Translation Buffer Invalidate All Process (DTB_IAP)
Register
DTB_IAP is a write-only register. Any write operation to this register
invalidates all data translation buffer (DTB) entries in which the address
space match (ASM) bit is equal to zero.

5.2.12 Dstream Translation Buffer Invalidate All (DTB_IA) Register
DTB_IA is a write-only register. Any write operation to this register
invalidates all 64 DTB entries, and resets the DTB not-last-used (NLU)
pointer to its initial state.

5–52 Preliminary—Subject to Change—July 1996

5.2 Memory Address Translation Unit (Mbox) IPRs

5.2.13 Dstream Translation Buffer Invalidate Single (DTB_IS) Register
DTB_IS is a write-only register. Writing a virtual address to this register
invalidates the DTB entry that meets either of the following criteria:
•

A DTB entry whose VA field matches DTB_IS<42:13> and whose ASN field
matches DTB_ASN<63:57>.

•

A DTB entry whose VA field matches DTB_IS<42:13> and whose ASM bit
is set.

Figure 5–38 shows the DTB_IS register format.
Figure 5–38 Dstream Translation Buffer Invalidate Single (DTB_IS) Register
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
VA<42:13>

IGN

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
IGN

VA<42:13>

LJ-03510-TI0

Note
The DTB_IS register is written before the normal Ibox trap point. The
DTB invalidate single operation is aborted by the Ibox only for the
following trap conditions:
•

ITB miss

•

PC mispredict

•

When the HW_MTPR DTB_IS is executed in user mode

Preliminary—Subject to Change—July 1996 5–53

5.2 Memory Address Translation Unit (Mbox) IPRs

5.2.14 Mbox Control Register (MCSR)
MCSR is a read/write register that controls features and records status in
the Mbox. This register is cleared on chip reset but not on timeout reset.
Figure 5–39 and Table 5–17 describe the MCSR format.
Figure 5–39 Mbox Control Register (MCSR)
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
RAZ/IGN
M_BIG_ENDIAN
SP<1:0>
MBZ
E_BIG_ENDIAN
MBZ

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
RAZ/IGN
LJ-03511-TI0

5–54 Preliminary—Subject to Change—July 1996

5.2 Memory Address Translation Unit (Mbox) IPRs

Table 5–17 Mbox Control Register Fields
Name

Extent

Type

Description

M_BIG_
ENDIAN

<00>

RW,0

Mbox Big Endian mode enable. When set, bit 2 of the
physical address is inverted for all longword Dstream
references.

SP<1:0>

<02:01>

RW,0

21164–266, 21164–300, and 21164–333
Superpage mode enables.
Note: Superpage access is only allowed in kernel mode.
SP<1> enables superpage mapping when VA<42:41> = 2.
In this mode, virtual addresses VA<39:13> are mapped
directly to physical addresses PA<39:13>. Virtual
address bit VA<40> is ignored in this translation.
SP<0> enables one-to-one superpage mapping of
Dstream virtual addresses with VA<42:30> = 1FFE16 .
In this mode, virtual addresses VA<29:13> are mapped
directly to physical addresses PA<29:13>, with bits
<39:30> of physical address set to 0. SP<0> is the
NT_Mode bit that is used to control virtual address
formatting on a read operation from the VA_FORM
register.
21164–P1 and 21164–P2
SP<0> must always be set. Clearing this bit will cause
21164–Pn operation to be UNPREDICTABLE.

Reserved

<03>

RW,0

Reserved to Digital. Must be zero (MBZ).

E_BIG_
ENDIAN

<04>

RW,0

Ebox Big Endian mode enable. This bit is sent to the
Ebox to enable Big Endian support for the EXTxx,
MSKxx and INSxx byte instructions. This bit causes the
shift amount to be inverted (one’s-complemented) prior
to the shifter operation.

Reserved

<05>

RW,0

Reserved to Digital. Must be zero (MBZ).

Preliminary—Subject to Change—July 1996 5–55

5.2 Memory Address Translation Unit (Mbox) IPRs

5.2.15 Dcache Mode (DC_MODE) Register
DC_MODE is a read/write register that controls diagnostic and test modes in
the Dcache. This register is cleared on chip reset but not on timeout reset.
Figure 5–40 and Table 5–18 describe the DC_MODE register format.
Note
The following bit settings are required for normal operation:
DC_ENA = 1
DC_FHIT = 0
DC_BAD_PARITY = 0
DC_PERR_DISABLE = 0

Figure 5–40 Dcache Mode (DC_MODE) Register

31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
RAZ/IGN
DC_ENA
DC_FHIT
DC_BAD_PARITY
DC_PERR_DISABLE

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
RAZ/IGN

LJ-03512-TI0

5–56 Preliminary—Subject to Change—July 1996

5.2 Memory Address Translation Unit (Mbox) IPRs

Table 5–18 Dcache Mode Register Fields
Name

Extent

Type

Description

DC_ENA

<00>

RW,0

Software Dcache enable. The DC_ENA bit
enables the Dcache unless the Dcache has
been disabled in hardware (DC_DOA is set).
(The Dcache is enabled if DC_ENA=1 and
DC_DOA=0). When clear, the Dcache command
is not updated by ST or FILL operations,
and all LD operations are forced to miss in
the Dcache. Must be one (MBO) in normal
operation.

DC_FHIT

<01>

RW,0

Dcache force hit. When set, the DC_FHIT
bit forces all Dstream references to hit in the
Dcache. Must be zero in normal operation.

DC_BAD_
PARITY

<02>

RW,0

When set, the DC_BAD_PARITY bit inverts
the data parity inputs to the Dcache on integer
stores. This has the effect of putting bad data
parity into the Dcache on integer stores that
hit in the Dcache. This bit has no effect on
the tag parity written to the Dcache during
FILL operations, or the data parity written
to the Cbox write data buffer on integer store
instructions.
Floating-point store instructions should not be
issued when this bit is set because it may result
in bad parity being written to the Cbox write
data buffer. Must be zero (MBZ) in normal
operation.

DC_PERR_
DISABLE

<03>

RW,0

When set, the DC_PERR_DISABLE bit disables
Dcache parity error reporting. When clear,
this bit enables all Dcache tag and data parity
errors. Parity error reporting is enabled during
all other Dcache test modes unless this bit is
explicitly set. Must be zero (MBZ) in normal
operation.

Preliminary—Subject to Change—July 1996 5–57

5.2 Memory Address Translation Unit (Mbox) IPRs

5.2.16 Miss Address File Mode (MAF_MODE) Register
MAF_MODE is a read/write register that controls diagnostic and test modes
in the Mbox miss address file (MAF). This register is cleared on chip reset.
MAF_MODE<05> is also cleared on timeout reset. Figure 5–41 and Table 5–19
describe the MAF_MODE register format.
Note
The following bit settings are required for normal operation:
DREAD_NOMERGE = 0
WB_FLUSH_ALWAYS = 0
WB_NOMERGE = 0
MAF_ARB_DISABLE = 0
WB_CNT_DISABLE = 0

Figure 5–41 Miss Address File Mode (MAF_MODE) Register

08 07 06 05 04 03 02 01 00
RAZ/IGN

DREAD_NOMERGE
WB_FLUSH_ALWAYS
WB_NOMERGE
IO_NMERGE
WB_CNT_DISABLE
MAF_ARB_DISABLE
DREAD_PENDING (Read-Only)
WB_PENDING (Read-Only)
63

32
RAZ/IGN
LJ-03513-TI0A

5–58 Preliminary—Subject to Change—July 1996

5.2 Memory Address Translation Unit (Mbox) IPRs

Table 5–19 Miss Address File Mode Register Fields
Name

Extent

Type

Description

DREAD_
NOMERGE

<00>

RW,0

Miss address file (MAF) DREAD Merge Disable. When
set, this bit disables all merging in the DREAD portion
of the MAF. Any load instruction that is issued when
DREAD_NOMERGE is set is forced to allocate a new
entry. Subsequent merging to that entry is not allowed
(even if DREAD_NOMERGE is cleared). Must be zero
(MBZ) in normal operation.

WB_FLUSH_
ALWAYS

<01>

RW,0

When set, this bit forces the write buffer to flush
whenever there is a valid WB entry. Must be zero
(MBZ) in normal operation.

WB_
NOMERGE

<02>

RW,0

When set, this bit disables all merging in the write
buffer. Any store instruction that is issued when WB_
NOMERGE is set is forced to allocate a new entry.
Subsequent merging to that entry is not allowed (even
if WB_NOMERGE is cleared). Must be zero (MBZ) in
normal operation.

IO_NMERGE

<03>

RW,0

When set, this bit prevents loads from I/O space
(address bit <39>=1) from merging in the MAF. Should
be zero (SBZ) in typical operation.

WB_CNT_
DISABLE

<04>

RW,0

When set, this bit disables the 64-cycle WB counter in
the MAF arbiter. The top entry of the WB arbitrates
at low priority only when a LDx_L instruction is issued
or a second WB entry is made. Must be zero (MBZ) in
normal operation.

MAF_ARB_
DISABLE

<05>

RW,0

When set, this bit disables all DREAD and WB requests
in the MAF arbiter. WB_Reissue, Replay, Iref and MB
requests are not blocked from arbitrating for the Scache.
This bit is cleared on both timeout and chip reset. Must
be zero (MBZ) in normal operation.

DREAD_
PENDING

<06>

R,0

Indicates the status of the MAF DREAD file. When set,
there are one or more outstanding DREAD requests
in the MAF file. When clear, there are no outstanding
DREAD requests.

WB_
PENDING

<07>

R,0

This bit indicates the status of the MAF WB file. When
set, there are one or more outstanding WB requests in
the MAF file. When clear, there are no outstanding WB
requests.

Preliminary—Subject to Change—July 1996 5–59

5.2 Memory Address Translation Unit (Mbox) IPRs

5.2.17 Dcache Flush (DC_FLUSH) Register
DC_FLUSH is a write-only register. A write operation to this register clears
all the valid bits in both banks of the Dcache.

5.2.18 Alternate Mode (ALT_MODE) Register
ALT_MODE is a write-only register that specifies the alternate processor mode
used by some HW_LD and HW_ST instructions. Figure 5–42 and Table 5–20
describe the ALT_MODE register format.
Figure 5–42 Alternate Mode (ALT_MODE) Register
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
IGN

IGN

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
IGN

LJ-03514-TI0

Table 5–20 Alternate Mode Register Settings
ALT_MODE<04:03>

Mode

Kernel

Executive

Supervisor

User

5–60 Preliminary—Subject to Change—July 1996

5.2 Memory Address Translation Unit (Mbox) IPRs

5.2.19 Cycle Counter (CC) Register
CC is a read/write register. The 21164 supports it as described in the Alpha
Architecture Reference Manual. The low half of the counter, when enabled,
increments once each CPU cycle. The upper half of the CC register is the
counter offset. An HW_MTPR instruction writes CC<63:32>. Bits <31:00> are
unchanged. CC_CTL<32> is used to enable or disable the cycle counter. The
CC<31:00> is written to CC_CTL by an HW_MTPR instruction.
The CC register is read by the RPCC instruction as defined in the Alpha
Architecture Reference Manual. The RPCC instruction returns a 64-bit
value. The cycle counter is enabled to increment only three cycles after the
MTPR CC_CTL (with CC_CTL<32> set) instruction is issued. This means
that an RPCC instruction issued four cycles after an HW_MTPR CC_CTL
instruction that enables the counter reads a value that is one greater than the
initial count.
The CC register is disabled on chip reset. Figure 5–43 shows the CC register
format.
Figure 5–43 Cycle Counter (CC) Register

31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
IGN

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
CC, OFFSET

LJ-03515-TI0

Preliminary—Subject to Change—July 1996 5–61

5.2 Memory Address Translation Unit (Mbox) IPRs

5.2.20 Cycle Counter Control (CC_CTL) Register
CC_CTL is a write-only register that writes the low 32 bits of the cycle counter
to enable or disable the counter. Bits CC<31:04> are written with the value
in CC_CTL<31:04> on a HW_MTPR instruction to the CC_CTL register.
Bits CC<03:00> are written with zero. Bits CC<63:32> are not changed. If
CC_CTL<32> is set, then the counter is enabled; otherwise, the counter is
disabled. Figure 5–44 and Table 5–21 describe the CC_CTL register format.
Figure 5–44 Cycle Counter Control (CC_CTL) Register
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
COUNT<31:04>

IGN

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
IGN
CC_ENA
LJ-03516-TI0

Table 5–21 Cycle Counter Control Register Fields
Name

Extent

Type

Description

COUNT<31:04> <31:04>

Cycle count. This value is loaded into
CC<31:04>.

CC_ENA

Cycle Counter enable. When set, this bit
enables the CC register to begin incrementing
3 cycles later. An RPCC issued 4 cycles after
CC_CTL<32> is written ‘‘sees’’ the initial count
incremented by 1.

<32>

5–62 Preliminary—Subject to Change—July 1996

5.2 Memory Address Translation Unit (Mbox) IPRs

5.2.21 Dcache Test Tag Control (DC_TEST_CTL) Register
DC_TEST_CTL is a read/write register used exclusively for testing and
diagnostics. An address written to this register is used to index into the Dcache
array when reading or writing to the DC_TEST_TAG register. Figure 5–45
and Table 5–22 describe the DC_TEST_CTL register format. Section 5.2.22
describes how this register is used.
Figure 5–45 Dcache Test Tag Control (DC_TEST_CTL) Register
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
RAZ/IGN

INDEX<12:3>
BANK0
BANK1
IGN/RAZ

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
RAZ/IGN

LJ-03517-TI0

Table 5–22 Dcache Test Tag Control Register Fields
Name

Extent

Type

Description

BANK0

<00>

Dcache Bank0 enable. When set, reads from
DC_TEST_TAG return the tag from Dcache
bank0, writes to DC_TEST_TAG write to Dcache
bank0. When clear, reads from DC_TEST_TAG
return the tag from Dcache bank1.

BANK1

<01>

Dcache Bank1 enable. When set, writes to
DC_TEST_TAG write to Dcache bank1. This
bit has no effect on reads.

Dcache tag index. This field is used on reads
from and writes to the DC_TEST_TAG register to
index into the Dcache tag array.

INDEX<12:3> <12:03>

Preliminary—Subject to Change—July 1996 5–63

5.2 Memory Address Translation Unit (Mbox) IPRs

5.2.22 Dcache Test Tag (DC_TEST_TAG) Register
DC_TEST_TAG is a read/write register used exclusively for testing and
diagnostics. When DC_TEST_TAG is read, the value in the DC_TEST_CTL
register is used to index into the Dcache. The value in the tag, tag parity, valid
and data parity bits for that index are read out of the Dcache and loaded into
the DC_TEST_TAG_TEMP register. A zero value is returned to the integer
register file (IRF). If BANK0 is set, the read operation is from Dcache bank0.
Otherwise, the read operation is from Dcache bank1.
When DC_TEST_TAG is written, the value written to DC_TEST_TAG is
written to the Dcache index referenced by the value in the DC_TEST_CTL
register. The tag, tag parity, and valid bits are affected by this write operation.
Data parity bits are not affected by this write operation (use DC_MODE<02>
and force hit modes). If BANK0 is set, the write operation is to Dcache bank0.
If BANK1 is set, the write operation is to Dcache bank1. If both are set, both
banks are written.
Figure 5–46 and Table 5–23 describe the DC_TEST_TAG register format.
Figure 5–46 Dcache Test Tag (DC_TEST_TAG) Register
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
TAG<38:13>

IGN

IGN
TAG_PARITY
OW0_VALID
OW1_VALID

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
IGN

TAG<38:13>

LJ-03518-TI0

5–64 Preliminary—Subject to Change—July 1996

5.2 Memory Address Translation Unit (Mbox) IPRs

Table 5–23 Dcache Test Tag Register Fields
Name

Extent

Type

Description

TAG_PARITY

<02>

Tag parity. This bit refers to the Dcache tag
parity bit that covers tag bits 38 through 13
(valid bits not covered).

OW0_VALID

<11>

Octaword valid bit 0. This bit refers to the
Dcache valid bit for the low-order octaword
within a Dcache 32-byte block.

OW1_VALID

<12>

Octaword valid bit 1. This bit refers to the
Dcache valid bit for the high-order octaword
within a Dcache 32-byte block.

TAG<38:13>

<38:13>

TAG<38:13>. These bits refer to the tag field in
the Dcache array.
Note: Bit 39 is not stored in the array.

Preliminary—Subject to Change—July 1996 5–65

5.2 Memory Address Translation Unit (Mbox) IPRs

5.2.23 Dcache Test Tag Temporary (DC_TEST_TAG_TEMP) Register
DC_TEST_TAG_TEMP is a read-only register used exclusively for testing and
diagnostics.
Reading the Dcache tag array requires a two-step read process:
1. The first read operation from DC_TEST_TAG reads the tag array and data
parity bits and loads them into the DC_TEST_TAG_TEMP register. An
UNDEFINED value is returned to the integer register file (IRF).
2. The second read operation of the DC_TEST_TAG_TEMP register returns
the Dcache test data to the integer register file (IRF).
Figure 5–47 and Table 5–24 describe the DC_TEST_TAG_TEMP register
format.
Figure 5–47 Dcache Test Tag Temporary (DC_TEST_TAG_TEMP) Register
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
TAG<38:13>

RAZ

RAZ
TAG_PARITY
DATA_PAR0<0>
DATA_PAR0<1>
DATA_PAR1<0>
DATA_PAR1<1>
OW0_VALID
OW1_VALID

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
RAZ

TAG<38:13>

LJ-03519-TI0

5–66 Preliminary—Subject to Change—July 1996

5.2 Memory Address Translation Unit (Mbox) IPRs

Table 5–24 Dcache Test Tag Temporary Register Fields
Name

Extent

Type

Description

TAG_PARITY

<02>

Tag parity. This bit refers to the Dcache tag parity
bit that covers tag bits 38 through 13 (valid bits not
covered).

DATA_PAR0<0>

<03>

Data parity. This bit refers to the Bank0 Dcache
data parity bit that covers the lower longword of data
indexed by DC_TEST_CTL<12:03>.

DATA_PAR0<1>

<04>

Data parity. This bit refers to the Bank0 Dcache data
parity bit that covers the upper longword of data
indexed by DC_TEST_CTL<12:03>.

DATA_PAR1<0>

<05>

Data parity. This bit refers to the Bank1 Dcache
data parity bit that covers the lower longword of data
indexed by DC_TEST_CTL<12:03>.

DATA_PAR1<1>

<06>

Data parity. This bit refers to the Bank1 Dcache data
parity bit that covers the upper longword of data
indexed by DC_TEST_CTL<12:03>.

OW0_VALID

<11>

Octaword valid bit 0. This bit refers to the Dcache valid
bit for the low-order octaword within a Dcache 32-byte
block.

OW1_VALID

<12>

Octaword valid bit 1. This bit refers to the Dcache valid
bit for the high-order octaword within a Dcache 32-byte
block.

TAG<38:13>

<38:13>

TAG<38:13>. These bits refer to the tag field in the
Dcache array.
Note: Bit 39 is not stored in the array.

Preliminary—Subject to Change—July 1996 5–67

5.3 External Interface Control (Cbox) IPRs

5.3 External Interface Control (Cbox) IPRs
Table 5–25 lists specific IPRs for controlling Scache, Bcache, system
configuration, and logging error information. These IPRs cannot be read
or written from the system. They are placed in the 1 MB region of 21164specific I/O address space ranging from FF FFF0 0000 to FF FFFF FFFF. Any
read or write operation to an undefined IPR in this address space produces
UNDEFINED behavior. The operating system should not map any address in
this region as writable in any mode.
The Cbox internal processor registers are described in Section 5.3.1 through
Section 5.3.9.
Table 5–25 Cbox Internal Processor Register Descriptions
Register

Address

Type1

Description

SC_CTL

FF FFF0 00A8

Controls Scache behavior.

SC_STAT

FF FFF0 00E8

Logs Scache-related errors.

SC_ADDR

FF FFF0 0188

Contains the address for Scacherelated errors.

BC_CONTROL

FF FFF0 0128

Controls Bcache/system interface
and Bcache testing.

BC_CONFIG

FF FFF0 01C8

Contains Bcache configuration
parameters.

BC_TAG_ADDR

FF FFF0 0108

Contains tag and control bits for
FILLs from Bcache.

EI_STAT

FF FFF0 0168

Logs Bcache/system-related errors.

EI_ADDR

FF FFF0 0148

Contains the address for
Bcache/system-related errors.

FILL_SYN

FF FFF0 0068

Contains fill syndrome or parity
bits for FILLs from Bcache or main
memory.

1 BC_CONTROL<01> must be 0 when reading any IPR in this table.

5–68 Preliminary—Subject to Change—July 1996

5.3 External Interface Control (Cbox) IPRs

5.3.1 Scache Control (SC_CTL) Register (FF FFF0 00A8)
SC_CTL is a read/write register that controls Scache activity. Figure 5–48 and
Table 5–26 describe the SC_CTL register format. The bits in this register are
initialized to the value indicated in Table 5–26 on reset, but not on timeout
reset.
Figure 5–48 Scache Control (SC_CTL) Register
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
RAZ/IGN

MBZ

S2 S1 S0

L3 L2 L1 L0
SC_FHIT
SC_FLUSH
SC_TAG_STAT<5:0>
SC_FB_DP<3:0>
SC_BLK_SIZE
SC_SET_EN<2:0>
Reserved

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
RAZ/IGN

LJ-03520-TI0

Preliminary—Subject to Change—July 1996 5–69

5.3 External Interface Control (Cbox) IPRs

Table 5–26 Scache Control Register Fields
Field

Extent

Type

Description

SC_FHIT

<00>

RW,0

When set, this bit forces cacheable load
and store instructions to hit in the
Scache, irrespective of the tag status bits.
Noncacheable references are not forced to
hit in the Scache and will be driven offchip.
In this mode, only one Scache set may be
enabled. The Scache tag and data parity
checking are disabled.
For store instructions, the value of the
tag status and parity bits are specified by
the SC_TAG_STAT<5:0> field. The tag is
written with the address provided to the
Scache with the store instruction.

SC_FLUSH

<01>

RW,0

All the Scache tag valid bits are cleared
every time this bit field is written to 1.

SC_TAG_
STAT<5:0>

<07:02>

RW,0

This field is used only in the SC_FHIT mode
to write any combination of tag status and
parity bits in the Scache. The parity bit can
be used to write bad tag parity. The correct
value of tag parity is even.
The following bits must be zero for normal
operation:
Scache Tag
Status<5:0>

Description

SC_TAG_
STAT<5:2>

Tag parity, valid,
shared, dirty;
bits 7, 6, 5, and 4
respectively

SC_TAG_
STAT<1:0>

Octaword modified
bits
(continued on next page)

5–70 Preliminary—Subject to Change—July 1996

5.3 External Interface Control (Cbox) IPRs

Table 5–26 (Cont.) Scache Control Register Fields
Field

Extent

Type

Description

SC_FB_DP<3:0>

<11:08>

RW,0

Force bad parity—This field is used to write
bad data parity for the selected longwords
within the octaword when writing the
Scache. If any one of these bits is set to one,
then the corresponding longword’s computed
parity value is inverted when writing the
Scache.
For Scache write transactions, the
Cbox allocates two consecutive cycles to
write up to two octawords based on the
longword valid bits received from the Mbox.
Therefore, the same longword parity control
bits are used for writing both octawords.
For example, SC_FB_DP<0> corresponds to
LW0 and LW4. This bit field must be zero
during normal operation.

SC_BLK_SIZE

<12>

RW,1

This bit selects the Scache and Bcache block
size to be either 64 bytes or 32 bytes. The
Scache and Bcache always have identical
block sizes. All the Bcache and main
memory FILLs or write transactions are
of the selected block size. At power-up time,
this bit is set and the default block size
is 64 bytes. When clear, the block size is
32 bytes. This bit must be set to the desired
value to reflect the correct Scache/Bcache
block size before the 21164 does the first
cacheable read or write transaction from
Bcache or system.

SC_SET_EN<2:0>

<15:13>

RW,7

This field is used to enable the Scache sets.
Only one or all three sets may be enabled
at a time. Enabling any combination of two
sets at a time results in UNPREDICTABLE
behavior. One of the Scache sets must
always be enabled irrespective of the
Bcache.

Reserved

<18:16>

RW,0

Reserved to Digital. Must be zero (MBZ).

Preliminary—Subject to Change—July 1996 5–71

5.3 External Interface Control (Cbox) IPRs

5.3.2 Scache Status (SC_STAT) Register (FF FFF0 00E8)
SC_STAT is a read-only register. It is not cleared or unlocked by reset. Any
PALcode read of this register unlocks SC_ADDR and SC_STAT and clears
SC_STAT.
If an Scache tag or data parity error is detected during an Scache lookup,
the SC_STAT register is locked against further updates from subsequent
transactions. Figure 5–49 and Table 5–27 describe the SC_STAT register
format.
Figure 5–49 Scache Status (SC_STAT) Register

31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
L7 L6 L5 L4 L3 L2 L1 L0 S2 S1 S0

RAZ

SC_TPERR<2:0>
SC_DPERR<7:0>
SC_CMD<4:0>
SC_SCND_ERR
63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
RAZ

LJ-03521-TI0

5–72 Preliminary—Subject to Change—July 1996

5.3 External Interface Control (Cbox) IPRs

Table 5–27 Scache Status Register Fields
Field

Extent

Type

Description

SC_TPERR<2:0>

<02:00>

When set, these bits indicate that an
Scache tag lookup resulted in a tag parity
error and identify the set that had the tag
parity error.

SC_DPERR<7:0>

<10:03>

When set, these bits indicate that an
Scache read transaction resulted in a data
parity error and indicate which longword
within the two octawords had the data
parity error. These bits are loaded if any
longword within two octawords read from
the Scache during lookup had a data parity
error. If SC_FHIT (SC_CTL<00>) is set,
this field is used for loading the longword
parity bits read out from the Scache.

SC_CMD<4:0>

<15:11>

This field indicates the Scache transaction
that resulted in a Scache tag or data
parity error. This field is written at the
time the actual Scache error bit is written.
The Scache transaction may be DREAD,
IREAD, or WRITE command from the
Mbox, Scache victim command, or the
system command being serviced. Refer to
Table 5–28 for field encoding.

SC_SCND_ERR

<16>

When set, this bit indicates that an Scache
transaction resulted in a parity error
while the SC_TPERR or SC_DPERR
bit was already set from the earlier
transaction. This bit is not set for two
errors in different octawords of the same
transaction.

Preliminary—Subject to Change—July 1996 5–73

5.3 External Interface Control (Cbox) IPRs

Table 5–28 SC_CMD Field Descriptions
SC_CMD<4:3> Source

SC_CMD<2:0> Encoding

Description

110

Set shared from
system

101

Read dirty from
system

100

Invalidate from
system

001

Scache victim

001

Scache IREAD

001

Scache DREAD

011

Scache DWRITE

5–74 Preliminary—Subject to Change—July 1996

5.3 External Interface Control (Cbox) IPRs

5.3.3 Scache Address (SC_ADDR) Register (FF FFF0 0188)
SC_ADDR is a read-only register. It is not cleared or unlocked by reset. The
address is loaded into this register every time the Scache is accessed if one
of the error bits in the SC_STAT register is not set. If an Scache tag or data
parity error is detected, then this register is locked preventing further updates.
This register is unlocked whenever SC_STAT is read.
For Scache read transactions, address bits <39:04> are valid to identify the
address being driven to the Scache. Address bit <04> identifies which octaword
was accessed first. For each Scache lookup, there is one tag access and two
data access cycles. If there is a hit, two octawords are read out in consecutive
CPU cycles. Tag parity error is detected only while reading the first octaword.
However, data parity error can be detected on either of the two octawords.
SC_ADDR<39> is always zero.
If SC_CTL<00> is set (force hit mode), SC_ADDR is used for storing the
Scache tag and status bits. For each tag in the Scache, there are unique valid,
shared, and dirty bits for a 32-byte subblock, and modify bits for each octaword
(16 bytes). There is a single tag and a parity bit for two consecutive 32-byte
subblocks. In force hit mode, only reads and probes load tag and status into
the SC_ADDR register. In this mode, tag and data parity checking are disabled
and the SC_ADDR and SC_STAT registers are not locked on an error.
In force hit mode, to write the Scache and read back the same block and
corresponding tag status bits, a minimum of 5-cycle spacing is required
between the Scache write and read of the SC_ADDR or SC_STAT.
Figure 5–50 and Table 5–29 describe the SC_ADDR register format.

Preliminary—Subject to Change—July 1996 5–75

5.3 External Interface Control (Cbox) IPRs

Figure 5–50 Scache Address (SC_ADDR) Register
Normal Mode
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
SC_ADDR<38:04>

RAO

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
0

RAO

SC_ADDR<38:04>
RAZ

Force Hit Mode
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
TAG<38:15>

M0 D1 S1 V1 D0 S0 V0 TP

RAO

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
RAO

TAG<38:15>
RAZ
LJ-03522-TI0

5–76 Preliminary—Subject to Change—July 1996

5.3 External Interface Control (Cbox) IPRs

Table 5–29 Scache Address Register Fields
Name

Extent

Type

Description

<38:04>

Scache address.

<04>

Scache tag parity bit.

<05>

Subblock0 tag valid bit.

<06>

Subblock0 tag shared bit.

<07>

Subblock0 tag dirty bit.

<08>

Subblock1 tag valid bit.

<09>

Subblock1 tag shared bit.

<10>

Subblock1 tag dirty bit.

<12,11>

Octawords modified for subblock0.

<14,13>

Octawords modified for subblock1.

TAG<38:15>

<38:15>

Scache tag.

Normal Mode
SC_ADDR<38:04>
Force Hit Mode

Preliminary—Subject to Change—July 1996 5–77

5.3 External Interface Control (Cbox) IPRs

5.3.4 Bcache Control (BC_CONTROL) Register (FF FFF0 0128)
BC_CONTROL is a write-only register. It is used to enable and control the
external Bcache. Figure 5–51 and Table 5–30 describe the BC_CONTROL
register format. The bits in this register are initialized to the value indicated
in Table 7–2 on reset, but not on timeout reset.
Figure 5–51 Bcache Control (BC_CONTROL) Register

29 28 27 26 25 24

19 18 17 16 15 14 13 12

08 07 06 05 04 03 02 01 00

T
TP C TV TS TD
P
BC_ENABLED
ALLOC_CYC
EI_CMD_GRP2
EI_CMD_GRP3
CORR_FILL_DAT
VTM_FIRST
EI_ECC_OR_PARITY
BC_FHIT
BC_TAG_STAT<4:0>
BC_BAD_DAT
EI_DIS_ERR
PIPE_LATCH
BC_WAVE<1:0>
PM_MUX_SEL<5:0>
MBZ
FLUSH_SC_VTM
MBZ
DIS_SYS_PAR
63

32
RAZ/IGN
LJ-03523-TI0

5–78 Preliminary—Subject to Change—July 1996

5.3 External Interface Control (Cbox) IPRs

Table 5–30 Bcache Control Register Fields
Field

Extent

Type

Description

BC_ENABLED

<00>

WO,0

When set, the external Bcache is enabled. When
clear, the Bcache is disabled. When the Bcache is
disabled, the BIU does not perform external cache
read or write transactions.

ALLOC_CYC

<01>

WO,0

When set, the issue unit does not allocate a
cycle for noncacheable fill data. When clear, the
instruction issue unit allocates a cycle for returning
noncacheable fill data to be written to the Dcache. In
either case, a cycle is always allocated for cacheable
integer fill data. If this bit is clear, the latency for
all noncacheable read operations increases by 1 CPU
cycle.

Note: This bit must be clear before reading any
Cbox IPR. It can be set when reading all other IPRs
and noncacheable LDs.
EI_CMD_GRP2

<02>

WO,0

When set, the optional commands, LOCK and
SET DIRTY are driven to the 21164 external
interface command pins to be acknowledged by
the system interface. When clear, the SET DIRTY
command is not driven to the command pins. It is
UNPREDICTABLE if the LOCK command is driven
to the pins. However, the system should never
CACK the LOCK command if this bit is clear.

EI_CMD_GRP3

<03>

WO,0

When set, the MB command is driven to the 21164
external interface command pins to be acknowledged
by the system interface. When clear, the MB
command is not driven to the command pins.

CORR_FILL_DAT

<04>

WO,1

Correct fill data from Bcache or main memory, in
ECC mode. When set, fill data from Bcache or main
memory first goes through error correction logic
before being driven to the Scache or Dcache. If the
error is correctable, it is transparent to the system.
When clear, fill data from Bcache or main memory is
driven directly to the Dcache before an ECC error is
detected. If the error is correctable, corrected data is
returned again, Dcache is invalidated, and an error
trap is taken.
This bit should be clear during normal operation.

1 When clear, the read speed (BC_RD_SPD<3:0>) and the write speed (BC_WR_SPD<3:0>) must be equal to

the sysclk to CPU clock ratio.

(continued on next page)

Preliminary—Subject to Change—July 1996 5–79

5.3 External Interface Control (Cbox) IPRs

Table 5–30 (Cont.) Bcache Control Register Fields
Field

Extent

Type

Description

VTM_FIRST

<05>

WO,1

This bit is set for systems without a victim buffer.
On a Bcache miss, the 21164 first drives out the
victimized block’s address on the system address bus,
followed by the read miss address and command.
This bit is cleared for systems with a victim buffer.
On a Bcache miss with victim, the 21164 first drives
out the read miss followed by the victim address and
command.

EI_ECC_OR_
PARITY

<06>

WO,1

When set, the 21164 generates or expects quadword
ECC on the data check pins. When clear, the 21164
generates or expects even-byte parity on the data
check pins.

BC_FHIT

<07>

WO,0

Bcache force hit. When set, and the Bcache is
enabled, all references in cached space are forced
to hit in the Bcache. A FILL to the Scache is
forced to be private. Software should turn off
BC_CONTROL<02> to allow clean to private
transitions without going to the system.
For write transactions, the values of tag status and
parity bits are specified by the BC_TAG_STAT field.
Bcache tag and index are the address received by
the BIU. The Bcache tag RAMs are written with the
address minus the Bcache index. This bit must be
zero during normal operation.

BC_TAG_
STAT<4:0>

<12:08>

This bit field is used only in BC_FHIT=1 mode to
write any combination of tag status and parity bits
in the Bcache. The parity bit can be used to write
bad tag parity. These bits are UNDEFINED on
reset. This bit field must be zero during normal
operation. The field encoding is as follows:
(continued on next page)

5–80 Preliminary—Subject to Change—July 1996

5.3 External Interface Control (Cbox) IPRs

Table 5–30 (Cont.) Bcache Control Register Fields
Field

Extent

Type

Description

Bcache Tag Status
Bit

Description

BC_TAG_STAT<4>

Parity for Bcache tag

BC_TAG_STAT<3>

Parity for Bcache tag status
bits

BC_TAG_STAT<2>

Bcache tag valid bit

BC_TAG_STAT<1>

Bcache tag shared bit

BC_TAG_STAT<0>

Bcache tag dirty bit

BC_BAD_DAT

<14:13>

WO,0

When set, bits in this field can be used to write
bad data with correctable or uncorrectable errors
in ECC mode. When bit <13> is set, data bit <0>
and <64> are inverted. When bit <14> is set, data
bit <1> and <65> are inverted. When the same
octaword is read from the Bcache, the 21164 detects
a correctable/uncorrectable ECC error on both the
quadwords based on the value of bits <14:13> used
when writing. This bit field must be zero during
normal operation.

EI_DIS_ERR

<15>

WO,1

When set, this bit causes the 21164 to ignore
any ECC (parity) error on fill data received from
the Bcache or main memory; or Bcache tag or
control parity error. It also ignores a system
command/address parity error. No machine check is
taken when this bit is set.

PIPE_LATCH

<16>

WO,0

When set, this bit causes the 21164 to pipe the
system control pins (addr_bus_req_h, cack_h, and
dack_h) for one system clock. Refer to Chapter 9 for
timing details.
(continued on next page)

Preliminary—Subject to Change—July 1996 5–81

5.3 External Interface Control (Cbox) IPRs

Table 5–30 (Cont.) Bcache Control Register Fields
Field

Extent

Type

Description

BC_WAVE<1:0>

<18:17>

WO,0

The bits in this field determine the number of
cycles of wave pipelining that should be used during
private read transactions of the Bcache. Wave
pipelining cannot be used in 32-byte block systems.
To enable wave pipelining, BC_CONFIG<07:04>
should be set to the latency of the Bcache read.
BC_CONTROL<18:17> should be set to the number
of cycles to subtract from BC_CONFIG<07:04> to
obtain the Bcache repetition rate. For example, if
BC_CONFIG<07:04>=7 and BC_CONTROL<18:17>=2,
it takes seven cycles for valid data to arrive at the
interface pins, but a new read will start every five
cycles.
The read repetition rate must be greater than
3. For example, it is not permitted to set
BC_CONFIG<07:04>=5 and BC_CONTROL<18:17>=2.
The value of BC_CONTROL<18:17> should be
added to the normal value of BC_CONFIG<14:12>
to increase the time between read and write
transactions. This prevents a write transaction from
starting before the last data of a read transaction is
received.

PM_MUX_
SEL<5:0>

<24:19>

WO,0

The bits in this field are used for selecting the BIU
parameters to be driven to the two performance
monitoring counters in the Ibox. Refer to Table 5–31
for the field encoding.

Reserved

<25>

WO,0

Reserved—MBZ.

FLUSH_SC_VTM

<26>

WO,0

Flush Scache victim buffer. For systems without
a Bcache, when this bit is clear, the 21164 flushes
the onchip victim buffer if it has to write-back any
entry from the victim buffer. When this bit is set,
the 21164 writes only one entry back from the victim
buffer as needed. This tends to cause read and write
operations to be batched rather than interleaved.
For systems with a Bcache, this bit must always be
clear. At power-up, this bit is initialized to a value of
0.

Reserved

<27>

WO,0

Reserved—MBZ.
(continued on next page)

5–82 Preliminary—Subject to Change—July 1996

5.3 External Interface Control (Cbox) IPRs

Table 5–30 (Cont.) Bcache Control Register Fields
Field

Extent

Type

Description

DIS_SYS_PAR

<28>

WO,0

When set, the 21164 does not check parity on the
system command/address bus. However, correct
parity will still be generated.

Table 5–31 describes the PM_MUX_SEL fields.
Table 5–31 PM_MUX_SEL Register Fields
PM_MUX_SEL<21:19>

Counter 1

0x0

Scache accesses

0x1

Scache read operations

0x2

Scache write operations

0x3

Scache victims

0x4

Undefined

0x5

Bcache accesses

0x6

Bcache victims

0x7

System command requests

PM_MUX_SEL<24:22>

Counter 2

0x0

Scache misses

0x1

Scache read misses

0x2

Scache write misses

0x3

Scache shared write operations

0x4

Scache write operations

0x5

Bcache misses

0x6

System invalidate operations

0x7

System read requests

Preliminary—Subject to Change—July 1996 5–83

5.3 External Interface Control (Cbox) IPRs

5.3.5 Bcache Configuration (BC_CONFIG) Register (FF FFF0 01C8)
BC_CONFIG is a write-only register used to configure the size and speed of
the external Bcache array. The bits in this register are initialized to the values
indicated in Table 5–32 on reset, but not on timeout reset. Figure 5–52 and
Table 5–32 describe the BC_CONFIG register format.
Figure 5–52 Bcache Configuration (BC_CONFIG) Register
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
IGN

BC_WE_CTL<8:0>
BC_SIZE<2:0>
MBZ
BC_RD_SPD<3:0>
BC_WR_SPD<3:0>
BC_RD_WR_SPC<2:0>
MBZ
FILL_WE_OFFSET<2:0>
MBZ

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
IGN

MLO-012926

5–84 Preliminary—Subject to Change—July 1996

5.3 External Interface Control (Cbox) IPRs

Table 5–32 Bcache Configuration Register Fields
Field

Extent

Type

Description

BC_SIZE<2:0>

<02:00>

WO,1

The bits in this field are used to indicate
the size of the Bcache. At power-on, this
field is initialized to a value representing
a 1M-byte Bcache. The field encoding is as
follows:

Reserved

<03>

WO,0

BC_
SIZE<2:0>1

Size

000

Invalid Bcache size

001

1 MB

010

2 MB

011

4 MB

100

8 MB

101

16 MB

110

32 MB

111

64 MB

Must be zero (MBZ).

(continued on next page)

Preliminary—Subject to Change—July 1996 5–85

5.3 External Interface Control (Cbox) IPRs

Table 5–32 (Cont.) Bcache Configuration Register Fields
Field

Extent

Type

Description

BC_RD_SPD<3:0>

<07:04>

WO,4

The bits in this field are used to indicate
to the BIU the read access time of the
Bcache, measured in CPU cycles, from the
start of a read transaction until data is
valid at the input pins. The Bcache read
speed must be within 4 to 10 CPU cycles.
At power-up, this field is initialized to a
value of 4 CPU cycles.
The Bcache read and write speeds
must be within three cycles of each
other (absolute value = (BC RD SPD 0
BC WR SPD) < 4).
For systems without a Bcache, the read
speed must be equal to the sysclk to
CPU clock ratio. In this configuration,
BC_RD_SPD can be set to a value ranging
from 3 to 15.

BC_WR_SPD<3:0>

<11:08>

WO,4

The bits in this field are used to indicate
to the BIU the write time of the Bcache,
measured in CPU cycles. The Bcache write
speed must be within 4 to 10 CPU cycles.
At power-up, this field is initialized to a
value of four CPU cycles.
For systems without a Bcache, the write
speed must be equal to sysclk to CPU clock
ratio.
(continued on next page)

5–86 Preliminary—Subject to Change—July 1996

5.3 External Interface Control (Cbox) IPRs

Table 5–32 (Cont.) Bcache Configuration Register Fields
Field

Extent

Type

Description

BC_RD_WR_
SPC<2:0>

<14:12>

WO,7

The bits in this field are used to indicate
to the BIU the number of CPU cycles to
wait when switching from a private read
to a private write Bcache transaction. For
other data movement commands, such as
READ DIRTY or FILL from main memory,
it is up to the system to direct systemwide
data movement in a way that is safe. A
value of 1 must be the minimum value for
this field.
The BIU always inserts three CPU
cycles between private Bcache read and
private Bcache write transactions, in
addition to the number of CPU cycles
specified by this field. The maximum value
(BC_RD_WR_SPC+3) should not be greater
than the Bcache READ speed when Bcache
is enabled.
At power-up, this field is initialized to a
read/write spacing of seven CPU cycles.

Reserved

<15>

WO,0

Must be zero (MBZ).

FILL_WE_
OFFSET<2:0>

<18:16>

WO,1

Bcache write-enable pulse offset, from
the sys_clk_outn_x edge, for FILL
transactions from the system. This field
does not affect private write transactions
to Bcache. It is used during FILLs from
the system when writing the Bcache to
determine the number of CPU cycles to
wait before shifting out the contents of the
write pulse field.
This field is programmed with a value in
the range of one to seven CPU cycles. It
must never exceed the sysclk ratio. For
example, if the sysclk ratio is 3, this field
must not be larger than 3. At power-up,
this field is initialized to a write offset
value of one CPU cycle.

Reserved

<19>

WO,0

Must be zero (MBZ).
(continued on next page)

Preliminary—Subject to Change—July 1996 5–87

5.3 External Interface Control (Cbox) IPRs

Table 5–32 (Cont.) Bcache Configuration Register Fields
Field

Extent

Type

Description

BC_WE_CTL<8:0>

<28:20>

WO,0

Bcache write-enable control. This field is
used to control the timing of the writeenable during a write or FILL transaction.
If the bit is set, the write pulse is asserted.
If the bit is clear, the write pulse is not
asserted. Each bit corresponds to a CPU
cycle. The least-significant bit corresponds
to the CPU cycle in which the 21164 starts
to drive the index for the write operation.
For private Bcache write and sharedwrite transactions, this field is used to
assert the write pulse without any writeenable pulse offset as indicated by the
FILL_WE_OFFSET<2:0> field.
For FILLs to the Bcache, the
FILL_WE_OFFSET<2:0> field determines
the number of CPU cycles to wait before
asserting the write pulse as programmed
in this field.
At power-up, all bits in this field are
cleared.

Reserved

<63:29>

5–88 Preliminary—Subject to Change—July 1996

Ignored.

5.3 External Interface Control (Cbox) IPRs

5.3.6 Bcache Tag Address (BC_TAG_ADDR) Register (FF FFF0 0108)
BC_TAG_ADDR is a read-only register. Unless locked, the BC_TAG_ADDR
register is loaded with the results of every Bcache tag read. When a tag or
tag control parity error occurs, this register is locked against further updates.
Software may read this register by using the 21164-specific I/O space address
instruction. This register is unlocked whenever the EI_STAT register is read,
or the user enters BC_FHIT mode. It is not unlocked by reset.
Note
The correct address is not loaded into BC_TAG_ADDR if a tag parity
error is detected when servicing a system command from the Bcache.

Unused tag bits in the TAG field of this register are always zero, based on the
size of the Bcache as determined by the BC_SIZE field of the BC_CONTROL
register. Figure 5–53 and Table 5–33 describe the BC_TAG_ADDR register
format.
Figure 5–53 Bcache Tag Address (BC_TAG_ADDR) Register

20 19 18 17 16 15 14 13 12 11
BC_TAG<38:20>

RAO

00
RAO
HIT
TAGCTL_P
TAGCTL_D
TAGCTL_S
TAGCTL_V
TAG_P
BC_TAG<38:20>

39 38
RAO

BC_TAG<38:20>
BC_TAG<38:20>
LJ-03526-TI0A

Preliminary—Subject to Change—July 1996 5–89

5.3 External Interface Control (Cbox) IPRs

Table 5–33 Bcache Tag Address Register Fields
Field

Extent

Type

Description

HIT

<12>

If set, Bcache access resulted in a hit in
the Bcache.

TAGCTL_P

<13>

Value of the parity bit for the Bcache tag
status bits.

TAGCTL_D

<14>

Value of the Bcache TAG dirty bit.

TAGCTL_S

<15>

Value of the Bcache TAG shared bit.

TAGCTL_V

<16>

Value of the Bcache TAG valid bit.

TAG_P

<17>

Value of the tag parity bit.

BC_TAG<38:20>

<38:20>

Bcache tag bits as read from the Bcache.
Unused bits are read as zero.

5–90 Preliminary—Subject to Change—July 1996

5.3 External Interface Control (Cbox) IPRs

5.3.7 External Interface Status (EI_STAT) Register (FF FFF0 0168)
EI_STAT is a read-only register. Any PALcode read access of this register
unlocks and clears it. A read access of EI_STAT also unlocks the EI_ADDR,
BC_TAG, and FILL_SYN registers subject to some restrictions. The EI_STAT
register is not unlocked or cleared by reset.
Fill data from Bcache or main memory could have correctable (c) or
uncorrectable (u) errors in ECC mode. In parity mode, fill data parity errors
are treated as uncorrectable hard errors. System address/command parity
errors are always treated as uncorrectable hard errors irrespective of the
mode. The sequence for reading, unlocking, and clearing EI_ADDR, BC_TAG,
FILL_SYN, and EI_STAT is as follows:
1. Read EI_ADDR, BC_TAG, and FILL_SYN in any order. Does not unlock or
clear any register.
2. Read EI_STAT register. Reading this register unlocks EI_ADDR, BC_TAG,
and FILL_SYN registers. EI_STAT is also unlocked and cleared when read,
subject to conditions described in Table 5–34.
Loading and locking rules for external interface registers are defined in
Table 5–34.
Note
If the first error is correctable, the registers are loaded but not locked.
On the second correctable error, registers are neither loaded nor locked.
Registers are locked on the first uncorrectable error except the
second hard error bit. The second hard error bit is set only for an
uncorrectable error followed by an uncorrectable error. If a correctable
error follows an uncorrectable error, it is not logged as a second error.
Bcache tag parity errors are uncorrectable in this context.

Preliminary—Subject to Change—July 1996 5–91

5.3 External Interface Control (Cbox) IPRs

Table 5–34 Loading and Locking Rules for External Interface Registers
Correctable
Error

Uncorrectable Second Hard
Error
Error

Load
Register

Lock
Register

Action when EI_STAT is read

Not possible

Clears and unlocks everything.

Not possible

Yes

Clears and unlocks everything.

Yes

Clears and unlocks everything.

Yes

Clear (c) bit does not unlock.
Transition to (0,1,0) state.

Already
locked

Clears and unlocks everything.

Already
locked

Clear (c) bit does not unlock.
Transition to (0,1,1) state.

1 These are special cases. It is possible that when EI_ADDR is read, only the correctable error bit is set and

the registers are not locked. By the time EI_STAT is read, an uncorrectable error is detected and the registers
are loaded again and locked. The value of EI_ADDR read earlier is no longer valid. Therefore, for the (1,1,x)
case, when EI_STAT is read correctable, the error bit is cleared and the registers are not unlocked or cleared.
Software must reexecute the IPR read sequence. On the second read operation, error bits are in (0,1,x) state,
all the related IPRs are unlocked, and EI_STAT is cleared.

The EI_STAT register is a read-only register used to control external interface
registers. Figure 5–54 and Table 5–35 describe the EI_STAT register format.
Figure 5–54 External Interface Status (EI_STAT) Register

31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
RAO
CHIP_ID<3:0>
BC_TPERR
BC_TC_PERR
EI_ES
COR_ECC_ERR
63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
RAO
UNC_ECC_ERR
EI_PAR_ERR
FIL_IRD
SEO_HRD_ERR
LJ-03524-TI0

5–92 Preliminary—Subject to Change—July 1996

5.3 External Interface Control (Cbox) IPRs

Table 5–35 EI_STAT Register Fields
Field

Extent

Type

Description

CHIP_ID<3:0>

<27:24> RO

Read as ‘‘4.’’ Future update revisions to the chip will
return new unique values.

BC_TPERR

<28>

Indicates that a Bcache read transaction encountered
bad parity in the tag address RAM.

BC_TC_PERR

<29>

Indicates that a Bcache read transaction encountered
bad parity in the tag control RAM.

EI_ES

<30>

When set, this bit indicates that the error source is fill
data from main memory or a system address/command
parity error.
When clear, the error source is fill data from the Bcache.
This bit is only meaningful when COR_ECC_ERR,
UNC_ECC_ERR, or EI_PAR_ERR is set.
This bit is not defined for a Bcache tag error
(BC_TPERR) or a Bcache tag control parity error
(BC_TC_ERR).

COR_ECC_ERR

<31>

Correctable ECC error. This bit indicates that a fill data
received from outside the CPU contained a correctable
ECC error.

UNC_ECC_ERR

<32>

Uncorrectable ECC error. This bit indicates that
fill data received from outside the CPU contained
an uncorrectable ECC error. In the parity mode, it
indicates data parity error.

EI_PAR_ERR

<33>

External interface command/address parity error. This
bit indicates that an address and command received by
the CPU has a parity error.

FIL_IRD

<34>

This bit has meaning only when one of the ECC or
parity error bit is set. It is set to indicate that the error
occurred during an I-ref FILL and clear to indicate that
the error occurred during a D-ref FILL.
This bit is not defined for a Bcache tag error
(BC_TPERR) or a Bcache tag control parity error
(BC_TC_ERR).

SEO_HRD_ERR

<35>

Second external interface hard error. This bit indicates
that a FILL from Bcache or main memory, or a system
address/command received by the CPU has a hard error
while one of the hard error bits in the EI_STAT register
is already set.

Preliminary—Subject to Change—July 1996 5–93

5.3 External Interface Control (Cbox) IPRs

5.3.8 External Interface Address (EI_ADDR) Register (FF FFF0 0148)
EI_ADDR is a read-only register that contains the physical address associated
with errors reported by the EI_STAT register. Its content is meaningful only
when one of the error bits is set. A read of EI_STAT unlocks the EI_ADDR
register. Figure 5–55 shows the EI_ADDR register format.
Figure 5–55 External Interface Address (EI_ADDR) Register

31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
EI_ADDR<39:4>

RAO

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
RAO

EI_ADDR<39:4>
LJ-03525-TI0

5–94 Preliminary—Subject to Change—July 1996

5.3 External Interface Control (Cbox) IPRs

5.3.9 Fill Syndrome (FILL_SYN) Register (FF FFF0 0068)
FILL_SYN is a 16-bit read-only register. It is loaded but not locked on a
correctable ECC error, so that another correctable error does not reload it. It is
loaded and locked if an uncorrectable ECC error or parity error is recognized
during a FILL from Bcache or main memory, as shown in Table 5–34. The
FILL_SYN register is unlocked when the EI_STAT register is read. This
register is not unlocked by reset.
If the 21164 is in ECC mode and an ECC error is recognized during a cache fill
transaction, the syndrome bits associated with the bad quadword are loaded in
the FILL_SYN register. FILL_SYN<07:00> contains the syndrome associated
with the lower quadword of the octaword. FILL_SYN<15:08> contains the
syndrome associated with the higher quadword of the octaword. A syndrome
value of 0 means that no errors where found in the associated quadword.
If the 21164 is in parity mode and a parity error is recognized during a cache
fill transaction, the FILL_SYN register indicates which of the bytes in the
octaword has bad parity. FILL_SYNDROME<07:00> is set appropriately to
indicate the bytes within the lower quadword that were corrupted. Likewise,
FILL_SYN<15:08> is set to indicate the corrupted bytes within the upper
quadword. Figure 5–56 shows the FILL_SYN register format.

Preliminary—Subject to Change—July 1996 5–95

5.3 External Interface Control (Cbox) IPRs

Figure 5–56 Fill Syndrome (FILL_SYN) Register
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
RAZ

HI<7:0>

LO<7:0>

63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
RAZ
LJ-03527-TI0

Table 5–36 lists the syndromes associated with correctable single-bit errors.
Table 5–36 Syndromes for Single-Bit Errors
Data Bit

Syndrome16

Check Bit

Syndrome16

0B
(continued on next page)

5–96 Preliminary—Subject to Change—July 1996

5.3 External Interface Control (Cbox) IPRs

Table 5–36 (Cont.) Syndromes for Single-Bit Errors
Data Bit

Syndrome16

Check Bit

Syndrome16

(continued on next page)

Preliminary—Subject to Change—July 1996 5–97

5.3 External Interface Control (Cbox) IPRs

Table 5–36 (Cont.) Syndromes for Single-Bit Errors
Data Bit

Syndrome16

5–98 Preliminary—Subject to Change—July 1996

Check Bit

Syndrome16

5.4 PALcode Storage Registers

5.4 PALcode Storage Registers
The 21164 Ebox register file has eight extra registers that are called the
PALshadow registers. The PALshadow registers overlay R8 through R14 and
R25 when the CPU is in PALmode and ICSR<SDE> is set. Thus, PALcode can
consider R8 through R14 and R25 as local scratch. PALshadow registers can
not be written in the last two cycles of a PALcode flow. The normal state of the
CPU is ICSR<SDE> = ON. PALcode disables SDE for the unaligned trap and
for error flows.
The Ibox holds a bank of 24 PALtemp registers. The PALtemp registers are
accessed with the HW_MTPR and HW_MFPR instructions. The latency from a
PALtemp read operation to availability is one cycle.

Preliminary—Subject to Change—July 1996 5–99

5.5 Restrictions

5.5 Restrictions
The following sections list all known register access restrictions. A software
tool called the PALcode violation checker (PVC) is available. This tool can be
used to verify adherence to many of the PALcode restrictions.

5.5.1 Cbox IPR PALcode Restrictions
Table 5–37 describes the Cbox IPR PALcode restrictions.
Table 5–37 Cbox IPR PALcode Restrictions
Condition

Restriction

Store to SC_CTL, BC_CONTROL, BC_
CONFIG except if no bit is changed other
than BC_CONTROL<ALLOC_CYC>,
BC_CONTROL<PM_MUX_SEL>, or BC_
CONTROL<DBG_MUX_SEL>.

Must be preceded by MB, must be followed
by MB, must have no concurrent cacheable
Istream references or concurrent system
commands.

Store to BC_CONTROL that only
changes bits BC_CONTROL<ALLOC_
CYC>, BC_CONTROL<PM_MUX_SEL>,
or BC_CONTROL<DBG_MUX_SEL>.

Must be preceded by MB and must be
followed by MB.

Load from SC_STAT.

Unlocks SC_ADDR and SC_STAT.

Load from EI_STAT.

Unlocks EI_ADDR, EI_STAT, FILL_SYN,
and BC_TAG_ADDR.

Any Cbox IPR address.

No LDx_L or STx_C.

Any undefined Cbox IPR address.

No store instructions.

Scache or Bcache in force hit mode.

No STx_C to cacheable space.

Clearing of SC_FHIT in SC_CTL.

Must be followed by MB, read operation
of SC_STAT, then MB prior to subsequent
store.

Clearing of BC_FHIT in BC_CONTROL.

Must be followed by MB, read operation
of EI_STAT, then MB prior to subsequent
store.

Load from any Cbox IPR.

BC_CONTROL<01> (ALLOC_CYCLE)
must be clear.

5–100 Preliminary—Subject to Change—July 1996

5.5 Restrictions

5.5.2 PALcode Restrictions—Instruction Definitions
Mbox instructions are: LDx, LDQ_U, LDx_L, HW_LD, STx, STQ_U, STx_C,
HW_ST, and FETCHx.
Virtual Mbox instructions are: LDx, LDQ_U, LDx_L, HW_LD (virtual), STx,
STQ_U, STx_C, HW_ST (virtual), and FETCHx.
Load instructions are: LDx, LDQ_U, LDx_L, and HW_LD.
Store instructions are: STx, STQ_U, STx_C, and HW_ST.
Table 5–38 lists PALcode restrictions.
Table 5–38 PALcode Restrictions Table
The following in cycle 0:

Restrictions (Note: Numbers refer to cycle number):

Y if checked
by PVC1

CALL_PAL entry

No HW_REI or HW_REI_STALL in cycle 0.
No HW_MFPR EXC_ADDR in cycle 0,1.

Y
Y

PALshadow write instruction

No HW_REI or HW_REI_STALL in 0, 1.

HW_LD, lock bit set

PAL must slot to E0.
No other Mbox instruction in 0.

HW_LD, VPTE bit set

No other virtual reference in 0.

Any load instruction

No Mbox HW_MTPR or HW_MFPR in 0.
No HW_MFPR MAF_MODE in 1,2 (DREAD_PENDING
may not be updated).
No HW_MFPR DC_PERR_STAT in 1,2.
No HW_MFPR DC_TEST_TAG slotted in 0.

Y
Y

Any store instruction

No HW_MFPR DC_PERR_STAT in 1,2.
No HW_MFPR MAF_MODE in 1,2 (WB_PENDING may
not be updated).

Y
Y

Any virtual Mbox instruction

No HW_MTPR DTB_IS in 1.

Any Mbox instruction or WMB,
if it traps

HW_MTPR any Ibox IPR not aborted in 0,1 (except that
EXC_ADDR is updated with correct faulting PC).
HW_MTPR DTB_IS not aborted in 0,1.

Any Ibox trap except PCmispredict, ITBMISS, or
OPCDEC due to user mode

HW_MTPR DTB_IS not aborted in 0,1.

HW_REI_STALL

Only one HW_REI_STALL in an aligned block of four
instructions.

1 PALcode violation checker

(continued on next page)

Preliminary—Subject to Change—July 1996 5–101

5.5 Restrictions

Table 5–38 (Cont.) PALcode Restrictions Table
Y if checked
by PVC1

The following in cycle 0:

Restrictions (Note: Numbers refer to cycle number):

HW_MTPR any undefined IPR
number

Illegal in any cycle.

ARITH trap entry

No HW_MFPR EXC_SUM or EXC_MASK in cycle 0,1.

Machine check trap entry

No register file read or write access in 0,1,2,3,4,5,6,7.
No HW_MFPR EXC_SUM or EXC_MASK in cycle 0,1.

HW_MTPR any Ibox IPR
(including PALtemp registers)

No HW_MFPR same IPR in cycle 1,2.
No floating-point conditional branch in 0.
No FEN or OPCDEC instruction in 0.

HW_MTPR ASTRR, ASTER

No HW_MFPR INTID in 0,1,2,3,4,5.
No HW_REI in 0,1.

Y
Y

HW_MTPR SIRR

No HW_MFPR INTID in 0,1,2,3,4.

HW_MTPR EXC_ADDR

No HW_REI in cycle 0,1.

HW_MTPR IC_FLUSH_CTL

Must be followed by 44 inline PALcode instructions.

HW_MTPR ICSR: HWE

No HW_REI in 0,1,2,3.

HW_MTPR ICSR: FPE

No floating-point instructions in 0, 1, 2, 3.
No HW_REI in 0,1,2.

HW_MTPR ICSR: SPE, FMS

If HW_REI_STALL, then no HW_REI_STALL in 0,1.
If HW_REI, then no HW_REI in 0,1,2,3,4.

HW_MTPR ICSR: SPE

Must flush Icache.

HW_MTPR ICSR: SDE

No PALshadow read/write access in 0,1,2,3.
No HW_REI in 0,1,2.

HW_MTPR ITB_ASN

Must be followed by HW_REI_STALL.
No HW_REI_STALL in cycle 0,1,2,3,4.
No HW_MTPR ITB_IS in 0,1,2,3.

Y
Y

HW_MTPR ITB_PTE

Must be followed by HW_REI_STALL.

HW_MTPR ITB_IAP, ITB_IS,
ITB_IA

Must be followed by HW_REI_STALL.

HW_MTPR ITB_IS

HW_REI_STALL must be in the same Istream octaword.

HW_MTPR IVPTBR

No HW_MFPR IFAULT_VA_FORM in 0,1,2.

HW_MTPR PAL_BASE

No CALL_PAL in 0,1,2,3,4,5,6,7.
No HW_REI in 0,1,2,3,4,5,6.

Y
Y

HW_MTPR ICM

No HW_REI in 0,1,2.
No private CALL_PAL in 0,1,2,3.

1 PALcode violation checker

(continued on next page)

5–102 Preliminary—Subject to Change—July 1996

5.5 Restrictions

Table 5–38 (Cont.) PALcode Restrictions Table
The following in cycle 0:

Restrictions (Note: Numbers refer to cycle number):

Y if checked
by PVC1

HW_MTPR CC, CC_CTL

No RPCC in 0,1,2.
No HW_REI in 0,1.

Y
Y

HW_MTPR DC_FLUSH

No Mbox instructions in 1,2.
No outstanding fills in 0.
No HW_REI in 0,1.

No Mbox instructions in 1,2,3,4.
No HW_MFPR DC_MODE in 1,2.
No outstanding fills in 0.
No HW_REI in 0,1,2,3.
No HW_REI_STALL in 0,1.

Y
Y

HW_MTPR DC_PERR_STAT

No load or store instructions in 1.
No HW_MFPR DC_PERR_STAT in 1,2.

Y
Y

HW_MTPR DC_TEST_CTL

No HW_MFPR DC_TEST_TAG in 1,2,3.
No HW_MFPR DC_TEST_CTL issued or slotted in 1,2.

HW_MTPR DC_TEST_TAG

No outstanding DC fills in 0.
No HW_MFPR DC_TEST_TAG in 1,2,3.

HW_MTPR DTB_ASN

No virtual Mbox instructions in 1,2,3.
No HW_REI in 0,1,2.

Y
Y

HW_MTPR DTB_CM, ALT_
MODE

No virtual Mbox instructions in 1,2.
No HW_REI in 0,1.

Y
Y

HW_MTPR DTB_PTE

No virtual Mbox instructions in 2.
No HW_MTPR DTB_ASN, DTB_CM, ALT_MODE, MCSR,
MAF_MODE, DC_MODE, DC_PERR_STAT,
DC_TEST_CTL, DC_TEST_TAG in 2.

Y
Y

HW_MTPR DTB_TAG

No virtual Mbox instructions in 1,2,3.
No HW_MTPR DTB_TAG in 1.
No HW_MFPR DTB_PTE in 1,2.
No HW_MTPR DTB_IS in 1,2.
No HW_REI in 0,1,2.

Y
Y
Y
Y
Y

HW_MTPR DTB_IAP, DTB_IA

No virtual Mbox instructions in 1,2,3.
No HW_MTPR DTB_IS in 0,1,2.
No HW_REI in 0,1,2.

Y
Y
Y

HW_MTPR DTB_IA

No HW_MFPR DTB_PTE in 1.

HW_MTPR MAF_MODE

No Mbox instructions in 1,2,3.
No WMB in 1,2,3.
No HW_MFPR MAF_MODE in 1,2.
No HW_REI in 0,1,2.

Y
Y
Y
Y

HW_MTPR DC_MODE

Y
Y

1 PALcode violation checker

(continued on next page)

Preliminary—Subject to Change—July 1996 5–103

5.5 Restrictions

Table 5–38 (Cont.) PALcode Restrictions Table
The following in cycle 0:

Restrictions (Note: Numbers refer to cycle number):

Y if checked
by PVC1

HW_MTPR MCSR

No virtual Mbox instructions in 0,1,2,3,4.
No HW_MFPR MCSR in 1,2.
No HW_MFPR VA_FORM in 1,2,3.
No HW_REI in 0,1,2,3.
No HW_REI_STALL in 0,1.

Y
Y
Y
Y
Y

HW_MTPR MVPTBR

No HW_MFPR VA_FORM in 1,2.

HW_MFPR ITB_PTE

No HW_MFPR ITB_PTE_TEMP in 1,2,3.

HW_MFPR DC_TEST_TAG

No outstanding DC fills in 0.
No HW_MFPR DC_TEST_TAG_TEMP issued or slotted
in 1.
No LDx instructions slotted in 0.
No HW_MTPR DC_TEST_CTL between HW_MFPR
DC_TEST_TAG and HW_MFPR DC_TEST_TAG_TEMP.

HW_MFPR DTB_PTE

No Mbox instructions in 0,1.
No HW_MTPR DC_TEST_CTL, DC_TEST_TAG in 0,1.
No HW_MFPR DTB_PTE_TEMP issued or slotted in
1,2,3.
No HW_MFPR DTB_PTE in 1.
No virtual Mbox instructions in 0,1,2.

HW_MFPR VA

Must be done in ARITH, MACHINE CHECK,
DTBMISS_SINGLE, UNALIGN, DFAULT traps and
ITBMISS flow after the VPTE load.

1 PALcode violation checker

5–104 Preliminary—Subject to Change—July 1996

Y
Y
Y
Y

6
Privileged Architecture Library Code
This chapter describes the 21164 privileged architecture library code
(PALcode). The chapter is organized as follows:
•

PALcode description

•

PALmode environment

•

Invoking PALcode

•

PALcode entry points

•

Required PALcode function codes

•

21164 implementation of the architecturally reserved opcodes

6.1 PALcode Description
Privileged architecture library code (PALcode) is macrocode that provides an
architecturally defined operating-system-specific programming interface that
is common across all Alpha microprocessors. The actual implementation of
PALcode differs for each operating system.
PALcode runs with privileges enabled, instruction stream mapping disabled,
and interrupts disabled. PALcode has privilege to use five special opcodes that
allow functions such as physical data stream references and internal processor
register (IPR) manipulation.
PALcode can be invoked by the following events:
•

Reset

•

System hardware exceptions (MCHK, ARITH)

•

Memory-management exceptions

•

Interrupts

•

CALL_PAL instructions

Preliminary—Subject to Change—July 1996 6–1

6.1 PALcode Description

PALcode has characteristics that make it appear to be a combination of
microcode, ROM BIOS, and system service routines, though the analogy to any
of these other items is not exact. PALcode exists for several major reasons:
•

There are some necessary support functions that are too complex to
implement directly in a processor chip’s hardware, but that cannot be
handled by a normal operating system software routine. Routines to
fill the translation buffer (TB), acknowledge interrupts, and dispatch
exceptions are some examples. In some architectures, these functions are
handled by microcode, but the Alpha architecture is careful not to mandate
the use of microcode so as to allow reasonable chip implementations.

•

There are functions that must run atomically, yet involve long sequences of
instructions that may need complete access to all the underlying computer
hardware. An example of this is the sequence that returns from an
exception or interrupt.

•

There are some instructions that are necessary for backward compatibility
or ease of programming; however, these are not used often enough to
dedicate them to hardware, or are so complex that they would jeopardize
the overall performance of the computer. For example, an instruction that
does a VAX style interlocked memory access might be familiar to someone
used to programming on a CISC machine, but is not included in the Alpha
architecture. Another example is the emulation of an instruction that has
no direct hardware support in a particular chip implementation.

In each of these cases, PALcode routines are used to provide the function. The
routines are nothing more than programs invoked at specified times, and read
in as Istream code in the same way that all other Alpha code is read. Once
invoked, however, PALcode runs in a special mode called PALmode.

6.2 PALmode Environment
PALcode runs in a special environment called PALmode, defined as follows:
•

Istream memory mapping is disabled. Because the PALcode is used to
implement translation buffer fill routines, Istream mapping clearly cannot
be enabled. Dstream mapping is still enabled.

•

The program has privileged access to all the computer hardware. Most of
the functions handled by PALcode are privileged and need control of the
lowest levels of the system.

•

Interrupts are disabled. If a long sequence of instructions need to be
executed atomically, interrupts cannot be allowed.

6–2 Preliminary—Subject to Change—July 1996

6.2 PALmode Environment

An important aspect of PALcode is that it uses normal Alpha instructions for
most of its operations; that is, the same instruction set that nonprivileged
Alpha programmers use. There are a few extra instructions that are only
available in PALmode, and will cause a dispatch to the OPCDEC PALcode
entry point if attempted while not in PALmode. The Alpha architecture allows
some flexibility in what these special PALmode instructions do. In the 21164
the special PALmode-only instructions perform the following functions:
•

Read or write internal processor registers (HW_MFPR, HW_MTPR).

•

Perform memory load or store operations without invoking the normal
memory-management routines (HW_LD, HW_ST).

•

Return from an exception or interrupt (HW_REI) .

When executing in PALmode, there are certain restrictions for using the
privileged instructions because PALmode gives the programmer complete
access to many of the internal details of the 21164. Refer to Section 6.6 for
information on these special PALmode instructions.
Caution
It is possible to cause unintended side effects by writing what appears
to be perfectly acceptable PALcode. As such, PALcode is not something
that many users will want to change.

6.3 Invoking PALcode
PALcode is invoked at specific entry points, under certain well-defined
conditions. These entry points provide access to a series of callable routines,
with each routine indexed as an offset from a base address. The base address
of the PALcode is programmable (stored in the PAL_BASE IPR), and is
normally set by the system reset code. Refer to Section 6.4 for additional
information on PALcode entry points.
PC<00> is used as the PALmode flag both to the hardware and to PALcode
itself. When the CPU enters a PALflow, the Ibox sets PC<00>. This bit
remains set as instructions are executed in the PAL Istream. The Ibox
hardware ignores this and behaves as if the PC were still longword aligned
for the purposes of Istream fetch and execute. On HW_REI, the new state of
PALmode is copied from EXC_ADDR<00>.

Preliminary—Subject to Change—July 1996 6–3

6.3 Invoking PALcode

When an event occurs that needs to invoke PALcode, the 21164 first drains
the pipeline. The current PC is loaded into the EXC_ADDR IPR, and the
appropriate PALcode routine is dispatched. These operations occur under
direct control of the chip hardware, and the machine is now in PALmode.
When the HW_REI instruction is executed at the end of the PALcode routine,
the hardware executes a jump to the address contained in the EXC_ADDR IPR.
The LSB is used to indicate PALmode to the hardware. Generally, the LSB is
clear upon return from a PALcode routine, in which case, the hardware loads
the new PC, enables interrupts, enables memory mapping, and dispatches back
to the user.
The most basic use of PALcode is to handle complex hardware events, and it is
called automatically when the particular hardware event is sensed. This use of
PALcode is similar to other architectures’ use of microcode.
There are several major categories of hardware-initiated invocations of
PALcode:
•

When the 21164 is reset, it enters PALmode and executes the RESET
PALcode. The system will remain in PALmode until a HW_REI instruction
is executed and EXC_ADDR<00> is cleared. It then continues execution
in non-PALmode (native mode), as just described. It is during this
initial RESET PALcode execution that the rest of the low-level system
initialization is performed, including any modification to the PALcode base
register.

•

When a system hardware error is detected by the 21164, it invokes one of
several PALcode routines, depending upon the type of error. Errors such as
machine checks, arithmetic exceptions, reserved or privileged instruction
decode, and data fetch errors are handled in this manner.

•

When the 21164 senses an interrupt, it dispatches the acknowledgment
of the interrupt to a PALcode routine that does the necessary information
gathering, then handles the situation appropriately for the given interrupt.

•

When a Dstream or Istream translation buffer miss occurs, one of several
PALcode routines is called to perform the TB fill.

The 21164 Ebox register file has eight extra registers that are called the
PALshadow registers. The PALshadow registers overlay R8, R9, R10, R11,
R12, R13, R14, and R25 when the CPU is in PALmode and ICSR<SDE>
is asserted. For additional PAL scratch, the Ibox has a register bank of 24
PALtemp registers, which are accessible via HW_MTPR and HW_MFPR
instructions.

6–4 Preliminary—Subject to Change—July 1996

6.4 PALcode Entry Points

6.4 PALcode Entry Points
PALcode is invoked at specific entry points. The 21164 has two types of
PALcode entry points: CALL_PAL and traps.

6.4.1 CALL_PAL Entry
CALL_PAL entry points are used whenever the Ibox encounters a CALL_PAL
instruction in the instruction stream (Istream). CALL_PAL instructions start
at the following offsets:
•

Privileged CALL_PAL instructions start at offset 200016 .

•

Nonnprivileged CALL_PAL instructions start at offset 300016 .

The CALL_PAL itself is issued into pipe E1 and the Ibox stalls for the
minimum number of cycles necessary to perform an implicit TRAPB. The
PC of the instruction immediately following the CALL_PAL is loaded into
EXC_ADDR and is pushed onto the return prediction stack.
The Ibox contains special hardware to minimize the number of cycles in
the TRAPB at the start of a CALL_PAL. Software can benefit from this by
scheduling CALL_PALs such that they do not fall in the shadow of:
•

IMUL

•

Any floating-point operate, especially FDIV

Each CALL_PAL instruction includes a function field that will be used in
the calculation of the next PC. The PAL OPCDEC flow will be started if the
CALL_PAL function field is:
•

In the range 4016 to 7F16 inclusive.

•

Greater than BF16 .

•

Between 0016 and 3F16 inclusive, and ICM<04:03> is not equal to kernel.

If no OPCDEC is detected on the CALL_PAL function, then the PC of the
instruction to execute after the CALL_PAL is calculated as follows:
•

PC<63:14> = PAL_BASE IPR<63:14>

•

PC<13> = 1

•

PC<12> = CALL_PAL function field<7>

•

PC<11:06> = CALL_PAL function field<5:0>

•

PC<05:01> = 0

•

PC<00> = 1 (PALmode)

Preliminary—Subject to Change—July 1996 6–5

6.4 PALcode Entry Points

The minimum number of cycles for a CALL_PAL execution is 4:
Number of
Cycles

Description

Minimum TRAPB for empty pipe. Typically this will be four cycles.

Issue the CALL_PAL instruction.

The minimum length of a PAL flow. However, in most cases there will
be more than two cycles of work for the CALL_PAL.

6.4.2 PALcode Trap Entry Points
Chip-specific trap entry points start PALcode. (No PALcode assist is required
for replay and mispredict type traps.) EXC_ADDR is loaded with the return
PC and the Ibox performs a TRAPB in the shadow of the trap. The return
prediction stack is pushed with the PC of the trapping instruction for precise
traps, and with some later PC for imprecise traps.
Table 6–1 shows the PALcode trap entry points and their offset from the
PAL_BASE IPR. Entry points are listed from highest to lowest priority.
(Prioritization among the Dstream traps works because DTBMISS is
suppressed when there is a sign check error. The priority of ITBMISS and
interrupt is reversed if there is an Icache miss.)
Table 6–1 PALcode Trap Entry Points
Entry Name

Offset16

Description

RESET

0000

Reset

IACCVIO

0080

Istream access violation or sign check error
on PC

INTERRUPT

0100

Interrupt: hardware, software, and AST

ITBMISS

0180

Istream TBMISS

DTBMISS_SINGLE

0200

Dstream TBMISS

DTBMISS_DOUBLE

0280

Dstream TBMISS during virtual page table
entry (PTE) fetch

UNALIGN

0300

Dstream unaligned reference

DFAULT

0380

Dstream fault or sign check error on virtual
address
(continued on next page)

6–6 Preliminary—Subject to Change—July 1996

6.4 PALcode Entry Points

Table 6–1 (Cont.) PALcode Trap Entry Points
Entry Name

Offset16

Description

MCHK

0400

Uncorrected hardware error

OPCDEC

0480

Illegal opcode

ARITH

0500

Arithmetic exception

FEN

0580

Floating-point operation attempted with:
•

Floating-point instructions (LD, ST, and
operates) disabled through FPE bit in the
ICSR IPR

•

Floating-point IEEE operation with data
type other than S, T, or Q

6.5 Required PALcode Function Codes
Table 6–2 lists opcodes required for all Alpha implementations. The notation
used is oo.ffff, where oo is the hexadecimal 6-bit opcode and ffff is the
hexadecimal 26-bit function code.
Table 6–2 Required PALcode Function Codes
Mnemonic

Type

Function Code

DRAINA

Privileged

00.0002

HALT

Privileged

00.0000

IMB

Unprivileged

00.0086

6.6 Alpha 21164 Implementation of the Architecturally
Reserved Opcodes
PALcode uses the Alpha instruction set for most of its operations. Table 6–3
lists the opcodes reserved by the Alpha architecture for implementation-specific
use. These opcodes are privileged and are only available in PALmode.
Note
These architecturally reserved opcodes contain different options to the
21064 opcodes of the same names.

Preliminary—Subject to Change—July 1996 6–7

6.6 Alpha 21164 Implementation of the Architecturally Reserved Opcodes

Table 6–3 Opcodes Reserved for PALcode
21164
Mnemonic

Opcode

Architecture
Mnemonic

HW_LD

PAL1B

Performs Dstream load
instructions.

HW_ST

PAL1F

Performs Dstream store
instructions.

HW_REI

PAL1E

Returns instruction flow to the
program counter (PC) pointed to
by EXC_ADDR IPR.

HW_MFPR

PAL19

Accesses the Ibox, Mbox, and
Dcache internal processor
registers (IPRs).

HW_MTPR

PAL1D

Accesses the Ibox, Mbox, and
Dcache IPRs.

Function

These instructions produce an OPCDEC exception if executed while not in
the PALmode environment. If ICSR<HWE> is set, these instructions can be
executed in kernel mode. Any software executing with ICSR<HWE> set must
use extreme care to obey all restrictions listed in this chapter and Chapter 5.
Register checking and bypassing logic is provided for PALcode instructions
as it is for non-PALcode instructions, when using general purpose registers
(GPRs).
Note
Explicit software timing is required for accessing the hardware-specific
IPRs and the PAL_TEMP registers. These constraints are described in
Table 5–38.

6.6.1 HW_LD Instruction
PALcode uses the HW_LD instruction to access memory outside of the realm
of normal Alpha memory management and to do special forms of Dstream
loads. Figure 6–1 and Table 6–4 describe the format and fields of the HW_LD
instruction. Data alignment traps are inhibited for HW_LD instructions.

6–8 Preliminary—Subject to Change—July 1996

6.6 Alpha 21164 Implementation of the Architecturally Reserved Opcodes

Figure 6–1 HW_LD Instruction Format
XXXXXX

Address

31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
OPCODE

DISP
LOCK
VPTE
QUAD
WRTCK
ALT
PHYS
LJ-03469-TI0

Table 6–4 HW_LD Format Description
Field

Value

Description

OPCODE

1B16

The OPCODE field contains 1B16 .

Destination register number.

Base register for memory address.

PHYS

0
1

The effective address for the HW_LD is virtual.
The effective address for the HW_LD is physical. Translation
and memory-management access checks are inhibited.

ALT

Memory-management checks use Mbox IPR DTB_CM for access
checks.
Memory-management checks use Mbox IPR ALT_MODE for
access checks.

1
WRTCK

0
1

Memory-management checks fault on read (FOR) and read
access violations.
Memory-management checks FOR, fault on write (FOW), read,
and write access violations.

QUAD

0
1

Length is longword.
Length is quadword.

VPTE

Flags a virtual PTE fetch. Used by trap logic to distinguish
single TBMISS from double TBMISS. Access checks are
performed in kernel mode.

LOCK

Load lock version of HW_LD. PAL must slot to E0 pipe.

DISP

Holds a 10-bit signed byte displacement.

Preliminary—Subject to Change—July 1996 6–9

6.6 Alpha 21164 Implementation of the Architecturally Reserved Opcodes

6.6.2 HW_ST Instruction
PALcode uses the HW_ST instruction to access memory outside of the realm
of normal Alpha memory management and to do special forms of Dstream
store instructions. Figure 6–2 and Table 6–5 describe the format and fields
of the HW_ST instruction. Data alignment traps are inhibited for HW_ST
instructions. The Ibox logic will always slot HW_ST to pipe E0.
Figure 6–2 HW_ST Instruction Format
XXXXXX

Address

31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
OPCODE

DISP
COND
MBZ
QUAD
MBZ
ALT
PHYS
LJ-03470-TI0

Table 6–5 HW_ST Format Description
Field

Value

Description

OPCODE

1F16

The OPCODE field contains 1F16 .

Write data register number.

Base register for memory address.

PHYS

0
1

The effective address for the HW_ST is virtual.
The effective address for the HW_ST is physical. Translation
and memory-management access checks are inhibited.

ALT

Memory-management checks use Mbox IPR DTB_CM for access
checks.
Memory-management checks use Mbox IPR ALT_MODE for
access checks.

1
QUAD

0
1

Length is longword.
Length is quadword.

COND

Store_conditional version of HW_ST. In this case, RA is written
with the value of LOCK_FLAG.

DISP

Holds a 10-bit signed byte displacement.

MBZ

HW_ST<13,11> must be zero.

6–10 Preliminary—Subject to Change—July 1996

6.6 Alpha 21164 Implementation of the Architecturally Reserved Opcodes

6.6.3 HW_REI Instruction
The HW_REI instruction is used to return instruction flow to the PC pointed to
by the EXC_ADDR IPR. The value in EXC_ADDR<0> will be used as the new
value of PALmode after the HW_REI instruction.
The Ibox uses the return prediction stack to speed the execution of HW_REI.
There are two different types of HW_REI:
•

Prefetch: In this case, the Ibox begins fetching the new Istream as soon as
possible. This is the version of HW_REI that is normally used.

•

Stall prefetch: This encoding of HW_REI inhibits Istream fetch until the
HW_REI itself is issued. Thus, this is the method used to synchronize Ibox
changes (such as ITB write instructions) with the HW_REI. There is a rule
that PALcode can have only one such HW_REI in an aligned block of four
instructions.

Figure 6–3 and Table 6–6 describe the format and fields of the HW_REI
instruction. The Ibox logic will slot HW_REI to pipe E1.
Figure 6–3 HW_REI Instruction Format
XXXXXX

Address

31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
OPCODE

TYP

MBZ
LJ-03471-TI0

Table 6–6 HW_REI Format Description
Field

Value

Description

OPCODE

1E16

The OPCODE field contains 1E16 .

TYP

10
11

Normal version.
Stall version.

MBZ

HW_REI<13:00> must be zero.

RA/RB

6.6.4 HW_MFPR and HW_MTPR Instructions
The HW_MFPR and HW_MTPR instructions are used to access internal state
from the Ibox, Mbox, and Dcache. The HW_MFPR from Ibox IPRs has a
latency of one cycle (HW_MFPR in cycle n results in data available to the
using instruction in cycle n+1). HW_MFPR from Mbox and Dcache IPRs has

Preliminary—Subject to Change—July 1996 6–11

6.6 Alpha 21164 Implementation of the Architecturally Reserved Opcodes

a latency of two cycles. Ibox hardware slots each type of MXPR to the correct
Ebox pipe (refer to Table 5–1).
Figure 6–4 and Table 6–7 describe the format and fields of the HW_MFPR and
HW_MTPR instructions.
Figure 6–4 HW_MFPR and HW_MTPR Instruction Format
XXXXXX

Address

31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15
OPCODE

12 11 10 09 08 07 06 05 04 03 02 01 00
Index
LJ-03472-TI0

Table 6–7 HW_MTPR and HW_MFPR Format Description
Field

Value

Description

OPCODE

1916
1D16

The OPCODE field contains 1916 for HW_MFPR.
The OPCODE field contains 1D16 for HW_MTPR.

RA/RB

Must be the same, source register for HW_MTPR and
destination register for HW_MFPR.

Index

Specifies the IPR. Refer to Table 5–1 for field encoding. Refer to
Chapter 5 for more details about specific IPRs.

6–12 Preliminary—Subject to Change—July 1996

7
Initialization and Configuration
This chapter provides information on 21164-specific microprocessor/system
initialization and configuration. It is organized as follows:
•

Input signals sys_reset_l and dc_ok_h and booting

•

Sysclk ratio and delay

•

Built-in self-test (BiSt)

•

Serial read-only memory (SROM) interface port

•

Serial terminal port

•

Cache initialization

•

External interface initialization

•

Internal processor register (IPR) reset state

•

Timeout reset

•

IEEE 1149.1 test port reset

7.1 Input Signals sys_reset_l and dc_ok_h and Booting
The 21164 reset sequence uses two input signals: sys_reset_l and dc_ok_h.
When transitioning from a powered-down state to a powered-up state, signal
dc_ok_h must be deasserted, and signal sys_reset_l must be asserted until
power has reached the proper operating point and the input clock to the
21164 is stable. If the input clock is derived from a PLL it may take many
milliseconds for the input oscillator to start and the PLL output to stabilize.
After power has reached the proper operating point, signal dc_ok_h must
be asserted. Then, signal sys_reset_l must be deasserted. At this point, the
21164 recognizes a powered on state. If signal dc_ok_h is not asserted, signal
sys_reset_l is forced asserted internally. After sys_reset_l is deasserted, the
21164 begins the following sequence of operations:
1. Icache built-in self-test (BiSt)

Preliminary—Subject to Change—July 1996 7–1

7.1 Input Signals sys_reset_l and dc_ok_h and Booting

2. An optional automatic Icache initialization, using an external serial ROM
(SROM) interface
3. Dispatch to the reset PALcode trap entry point (physical location 0)
a. If step 2 initialized the Icache by using the SROM interface, the cache
should contain code that appears to be at location 0, that is, the cache
should be initialized such that it hits on the dispatch. Typically the
code in the Icache should configure the 21164’s IPRs as necessary
before causing any offchip read or write commands. This allows the
21164 to be configured to match the external system implementation.
b. If step 2 did not initialize the Icache, the Icache has been flushed by
reset. The reset PALcode trap dispatch misses in the Icache and Scache
(also flushed by reset) and produces an offchip read command. The
external system implementation must be compatible with the 21164’s
default configuration after reset (refer to Section 7.8). The code that
is executed at this point should complete the 21164 configuration as
necessary.
4. After configuring the 21164, control can be transferred to code anywhere in
memory, including the noncacheable regions. If the SROM interface was
used to initialize the Icache, the Icache can be flushed by a write operation
to IC_FLUSH_CTL after control is transferred. This transfer of control
should be to addresses not loaded in the Icache by the SROM interface or
the Icache may provide unexpected instructions.
5. Typically, PALbase and any state required by PALcode are initialized
and the console is started (switching out of PALmode and into native
mode). The console code initializes and configures the system and boots an
operating system from an I/O device such as a disk or the network.
Signal sys_reset_l forces the CPU into a known state. Signal sys_reset_l
must remain asserted while signal dc_ok_h is deasserted, and for some period
of time after dc_ok_h assertion. It should remain asserted for at least 400
internal CPU cycles in length. Then, signal sys_reset_l may be deasserted.
Signal sys_reset_l deassertion need not be synchronous with respect to sysclk.
Section 7.8 lists the reset state of each IPR. Table 7–1 provides the reset state
of each external signal pin.

7–2 Preliminary—Subject to Change—July 1996

7.1 Input Signals sys_reset_l and dc_ok_h and Booting

Table 7–1

Alpha 21164 Signal Pin Reset State

Signal

Reset State

Clocks
clk_mode_h<1:0>

NA (input).

cpu_clk_out_h

Clock output.

dc_ok_h

NA (input).

osc_clk_in_h,l

Must be clocking.

ref_clk_in_h

NA (input).

sys_clk_out1_h,l

Clock output.

sys_clk_out2_h,l

Clock output.

sys_reset_l

NA (input).

Bcache
data_h<127:0>

Tristated.

data_check_h<15:0>

Tristated.

data_ram_oe_h

Deasserted.

data_ram_we_h

Deasserted.

index_h<25:4>

Unspecified.

tag_ctl_par_h

Tristated.

tag_data_h<38:20>

Tristated.

tag_data_par_h

Tristated.

tag_dirty_h

Tristated.

tag_ram_oe_h

Deasserted.

tag_ram_we_h

Deasserted.

tag_shared_h

Tristated.

tag_valid_h

Tristated.
(continued on next page)

Preliminary—Subject to Change—July 1996 7–3

7.1 Input Signals sys_reset_l and dc_ok_h and Booting

Table 7–1 (Cont.)
Signal

Alpha 21164 Signal Pin Reset State
Reset State

System Interface
addr_h<39:4>

Driven or tristated depending upon addr_bus_req_h at
most recent sysclk edge. If driven, the value is unspecified.

addr_bus_req_h

NA (input).

addr_cmd_par_h

Driven or tristated depending upon addr_bus_req_h at
most recent sysclk edge. If driven, the command is NOP.

addr_res_h<2:0>

NOP.

cack_h

Must be deasserted.

cfail_h

Must be deasserted.

cmd_h<3:0>

Driven or tristated depending upon addr_bus_req_h at
most recent sysclk edge. If driven, the command is NOP.

dack_h

Must be deasserted.

data_bus_req_h

NA (input).

fill_h

Must be deasserted.

fill_error_h

Must be deasserted.

fill_id_h

Must be deasserted.

fill_nocheck_h

Must be deasserted.

idle_bc_h

Must be deasserted.

int4_valid_h<3:0>

Unspecified.

scache_set_h<1:0>

Unspecified.

shared_h

NA (input).

system_lock_flag_h

Must be deasserted.

victim_pending_h

Unspecified.

Interrupts
irq_h<3:0>

Sysclk divisor ratio input.

mch_hlt_irq_h

Sysclk delay input.

pwr_fail_irq_h

Sysclk delay input.

sys_mch_chk_irq_h

Sysclk delay input.
(continued on next page)

7–4 Preliminary—Subject to Change—July 1996

7.1 Input Signals sys_reset_l and dc_ok_h and Booting

Table 7–1 (Cont.)
Signal

Alpha 21164 Signal Pin Reset State
Reset State

Test Modes
port_mode_h<1:0>

NA (input).

srom_clk_h

Deasserted.

srom_data_h

NA (input).

srom_oe_l

Deasserted.

srom_present_l

NA (input).

tck_h

NA (input).

tdi_h

NA (input).

tdo_h

NA (input).

temp_sense

NA (input).

test_status_h<1:0>

Deasserted.

tms_h

NA (input).

trst_l

Must be asserted (input).

Miscellaneous
perf_mon_h

NA (input).

spare_io

NA.

While signal dc_ok_h is deasserted, the 21164 provides its own internal clock
source from an onchip ring oscillator. When dc_ok_h is asserted, the 21164
clock source is the differential clock input pins osc_clk_in_h,l.
When the 21164 is free-running from the internal ring oscillator, the internal
clock frequency is in the range of 10 MHz to 100 MHz (varies from chip to
chip). The sysclk divisor and sys_clk_out2_x delay are determined by input
pins while signal sys_reset_l remains asserted. Refer to Section 4.2.2 and
Section 4.2.3 for ratio and delay values.

7.1.1 Pin State with dc_ok_h Not Asserted
While dc_ok_h is deasserted, and sys_reset_l is asserted, every output and
bidirectional 21164 pin is tristated and pulled weakly to ground by a small
pull-down transistor.

Preliminary—Subject to Change—July 1996 7–5

7.2 Sysclk Ratio and Delay

7.2 Sysclk Ratio and Delay
While in reset, the 21164 reads sysclk configuration parameters from the
interrupt signal pins. These inputs should be driven with the correct
configuration values whenever signal sys_reset_l is asserted. Refer to
Section 4.2.2 and Section 4.2.3 for relevant input signals and ratio/delay
values.
If the signal inputs reflecting configuration parameters change while
sys_reset_l is asserted, allow 20 internal CPU cycles before the new sysclk
behavior is correct.

7.3 Built-In Self-Test (BiSt)
Upon deassertion of signal sys_reset_l, the 21164 automatically executes the
Icache built-in self-test (BiSt). The Icache is automatically tested and the
result is made available in the ICSR IPR and on signal test_status_h<0>.
Internally, the CPU reset continues to be asserted throughout the BiSt process.
For additional information, refer to Section 9.4.6.

7.4 Serial Read-Only Memory Interface Port
The serial read-only memory (SROM) interface provides the initialization data
load path from a system SROM to the instruction cache (Icache). Following
initialization, this interface can function as a diagnostic port using privileged
architecture library code (PALcode).
The following signals make up the SROM interface:
srom_present_l
srom_data_h
srom_oe_l
srom_clk_h
During system reset, the 21164 samples the srom_present_l signal for
the presence of SROM. If srom_present_l is deasserted, the SROM load is
disabled and the reset sequence clears the Icache valid bits. This causes the
first instruction fetch to miss the Icache and read instructions from offchip
memory.
If srom_present_l is asserted during setup, then the system performs an
SROM load as follows:
1. The srom_oe_l signal supplies the output enable to the SROM.

7–6 Preliminary—Subject to Change—July 1996

7.4 Serial Read-Only Memory Interface Port

2. The srom_clk_h signal supplies the clock to the ROM that causes it to
advance to the next bit. The cycle time of this clock is 1266 times the CPU
clock period.
3. The srom_data_h signal inputs the SROM data.
Every data and tag bit in the Icache is loaded by this sequence.

7.4.1 Serial Instruction Cache Load Operation
All Icache bits, including each block’s tag, address space number (ASN),
address space match (ASM), valid and branch history bits can be loaded
serially from offchip serial ROMs. Once the serial load has been invoked by
the chip reset sequence, the entire cache is loaded automatically from the
lowest to the highest addresses.
The automatic serial Icache fill invoked by the chip reset sequence operates
internally at a frequency of 126*CPU clock period. However, due to the
synchronization with the system clocks, consecutive access cycles to SROM
may shrink or stretch by a system cycle. For example, for a system with a
system clock ratio of 15, the time between the two consecutive SROM accesses
may be anywhere in the range 111 to 141 CPU cycles. The SROM used
in the system must be able to support access times in this range. Refer to
Section 9.4.5 for additional SROM timing information.
The serial bits are received in a 200-bit-long fill scan path, from which they are
written in parallel into the Icache address. The fill scan path is organized as
shown in the text following this paragraph. The farthest bit (<42>) is shifted
in first and the nearest bit (BHT<0>) is shifted in last. The data and predecode
bits in the data array are interleaved.

srom_data_h
BHT Array
Data
Predecodes
Data parity
Predecodes
Data
Tag Parity
Tag Valids
TAG Phy.Address
TAG ASN
TAG ASM
TAGs

serial input ->
0 -> 1 -> ... ->
127 -> 95 -> 126 ->
19 -> 14 -> 18 ->
1 -> 0 ->
9 -> 4 ->
8 ->
63 -> 31 -> 62 ->
b ->
0 -> 1 ->
b ->
0 -> 1 -> ... ->
b ->
13 -> 14 -> ... ->

7 ->
94 -> ... -> 96 -> 64 ->
13 -> ... -> 15 -> 10 ->
3 -> ... -> 5 ->
30 -> ... -> 32 ->

0 ->
0 ->

6 ->
42

b = Single bit signal

Preliminary—Subject to Change—July 1996 7–7

7.4 Serial Read-Only Memory Interface Port

Refer to Appendix C for example C code that calculates the predecode values of
a serial Icache load.

7.5 Serial Terminal Port
After the SROM data is loaded into the Icache, the three SROM interface
signals can be used as a software ‘‘UART’’ and the pins become parallel I/O
pins that can drive a diagnostic terminal by using an interface such as RS232
or RS423.

7.6 Cache Initialization
Regardless of whether the Icache BiSt is executed, the Icache is flushed during
the reset sequence prior to the SROM load. If the SROM load is bypassed, the
Icache will be in the flushed state initially.
The second-level cache (Scache) is flushed and enabled by internal reset. This
is required if the SROM load is bypassed. The initial Istream reference after
reset is location 0. Because that is a cacheable-space reference, the Scache will
be probed.
The data cache (Dcache) is disabled by reset. It is not initialized or flushed by
reset. It should be initialized by PALcode before being enabled.
The external board-level Bcache is disabled by reset. It should be initialized by
PALcode before being enabled.

7.6.1 Icache Initialization
The Icache is not kept coherent with memory. When it is necessary to make it
coherent with memory, the following procedure is used. The CALL_PAL IMB
function performs this function by using this procedure.
1. Execute an MB instruction. This forces all write data in the write buffer
into memory.
Stall until write buffer is drained.
Carry load or issue a HW_MFPR from any Mbox IPR.
2. Write to IC_FLUSH_CTL with an HW_MTPR to flush the Icache.
3. Execute a total of 44 NOP instructions (BIS r31,r31,r31) to clear the
prefetch buffers and Ibox pipeline. The 44 NOP instructions must start on
an INT16 boundary. Pad with additional NOP instructions if necessary.

7–8 Preliminary—Subject to Change—July 1996

7.6 Cache Initialization

7.6.2 Flushing Dirty Blocks
During a power failure recovery, dirty blocks must be flushed out of the Scache
and backup cache (Bcache), if present.
Systems Without a Bcache
To flush out dirty blocks from the Scache on power failure, the following
sequence must be used to guarantee that all the dirty blocks have been
written back to main memory. The BC_CONFIG<BC_SIZE> field is used
for this function in systems without a Bcache. When powering up, this field
is initialized to a value representing a 1M-byte Bcache. During system
configuration flow, this field must be changed to a value of 0 for normal
operation.
To flush out the dirty blocks from all three sets in the Scache, perform the
following tasks:
1. Set BC_CONFIG<BC_SIZE><2:0> = 0x1; do loads at a stride of 64 bytes
through 128K bytes of continuous memory; guarantees all dirty blocks from
set0 are flushed out.
2. Set BC_CONFIG<BC_SIZE><2:0> = 0x2; do loads at a stride of 64 bytes
through 96K bytes of continuous memory; guarantees all dirty blocks from
set1 are flushed out.
3. Set BC_CONFIG<BC_SIZE><2:0> = 0x4; do loads at a stride of 64 bytes
through 64K bytes of continuous memory; guarantees all dirty blocks from
set2 are flushed out.
All other values of BC_CONFIG<BC_SIZE><2:0> are undefined in this mode.
Systems with a Bcache
To flush out dirty blocks from the Scache and Bcache on power failure, the
following sequence must be used to guarantee that all the dirty blocks have
been written back to main memory:
perform loads at a stride of Bcache block size = 22 size of the Bcache

Preliminary—Subject to Change—July 1996 7–9

7.7 External Interface Initialization

7.7 External Interface Initialization
After reset, the cache control and bus interface unit (Cbox) is in the default
configuration dictated by the reset state of the IPR bits that select the
configuration options. The Cbox response to system commands and internally
generated memory accesses is determined by this default configuration.
System environments that are not compatible with the default configuration
must use the SROM Icache load feature to initially load and execute a PALcode
program. This program configures the external interface control (Cbox) IPRs
as needed.

7.8 Internal Processor Register Reset State
Many IPR bits are not initialized by reset. They are located in error-reporting
registers and other IPR states. They must be initialized by initialization
PALcode. Table 7–2 lists the state of all internal processor registers (IPRs)
immediately following reset. The table also specifies which registers need to be
initialized by power-up PALcode.
Table 7–2 Internal Processor Register Reset State
IPR

Reset State

Comments

Ibox Registers
ITB_TAG

UNDEFINED

ITB_PTE

UNDEFINED

ITB_ASN

UNDEFINED

ITB_PTE_TEMP

UNDEFINED

ITB_IAP

UNDEFINED

ITB_IA

UNDEFINED

ITB_IS

UNDEFINED

IFAULT_VA_FORM

UNDEFINED

IVPTBR

UNDEFINED

PALcode must initialize.

ICPERR_STAT

UNDEFINED

PALcode must initialize.

IC_FLUSH_CTL

UNDEFINED

EXC_ADDR

UNDEFINED

PALcode must initialize.

(continued on next page)

7–10 Preliminary—Subject to Change—July 1996

7.8 Internal Processor Register Reset State

Table 7–2 (Cont.) Internal Processor Register Reset State
IPR

Reset State

Comments

EXC_SUM

UNDEFINED

PALcode must clear exception
summary and exception register
write mask by writing EXC_SUM.

EXC_MASK

UNDEFINED

PAL_BASE

Cleared

Cleared on reset.

ICM

UNDEFINED

PALcode must set current mode.

ICSR

See Comments

All bits are cleared on reset except
ICSR<37>, which is set, and
ICSR<38>, which is UNDEFINED.

IPLR

UNDEFINED

PALcode must initialize.

INTID

UNDEFINED

ASTRR

UNDEFINED

PALcode must initialize.

ASTER

UNDEFINED

PALcode must initialize.

SIRR

UNDEFINED

PALcode must initialize.

HWINT_CLR

UNDEFINED

PALcode must initialize.

ISR

UNDEFINED

SL_XMIT

Cleared

SL_RCV

UNDEFINED

PMCTR

See Comments

PMCTR<15:10> are cleared
on reset. All other bits are
UNDEFINED.

DTB_ASN

UNDEFINED

PALcode must initialize.

DTB_CM

UNDEFINED

PALcode must initialize.

DTB_TAG

Cleared

Valid bits are cleared on chip reset
but not on timeout reset.

DTB_PTE

UNDEFINED

DTB_PTE_TEMP

UNDEFINED

MM_STAT

UNDEFINED

Appears on external pin.

Mbox Registers

Must be unlocked by PALcode by
reading VA register.
(continued on next page)

Preliminary—Subject to Change—July 1996 7–11

7.8 Internal Processor Register Reset State

Table 7–2 (Cont.) Internal Processor Register Reset State
IPR

Reset State

Comments

UNDEFINED

Must be unlocked by PALcode by
reading VA register.

VA_FORM

UNDEFINED

Must be unlocked by PALcode by
reading VA register.

MVPTBR

UNDEFINED

PALcode must initialize.

DC_PERR_STAT

UNDEFINED

PALcode must initialize.

DTB_IAP

UNDEFINED

DTB_IA

UNDEFINED

DTB_IS

UNDEFINED

MCSR

Cleared

Cleared on chip reset but not on
timeout reset.

DC_MODE

Cleared

Cleared on chip reset but not on
timeout reset.

MAF_MODE

Cleared

Cleared on chip reset.
MAF_MODE<05> cleared on
timeout reset.

DC_FLUSH

UNDEFINED

PALcode must write this register to
clear Dcache valid bits.

ALT_MODE

UNDEFINED

CC_CTL

UNDEFINED

DC_TEST_CTL

UNDEFINED

DC_TEST_TAG

UNDEFINED

DC_TEST_TAG_TEMP

UNDEFINED

CC is disabled on chip reset.

Cbox Registers
SC_CTL

See Comments

SC_CTL<11:00> cleared on reset.
SC_CTL<12> is set at power-up.

SC_STAT

UNDEFINED

PALcode must read to unlock.

SC_ADDR

UNDEFINED
(continued on next page)

7–12 Preliminary—Subject to Change—July 1996

7.8 Internal Processor Register Reset State

Table 7–2 (Cont.) Internal Processor Register Reset State
IPR

Reset State

Comments

BC_CONTROL

See Comments

BC_CONTROL<01:00>, <07>,
<14:13>, <16>, and <27:19> cleared.
BC_CONTROL<06:04> and <15>
set on reset but not timeout reset.
All other bits are UNDEFINED and
must be initialized by PALcode.

BC_CONFIG

See Comments

At power-up, BC_CONFIG
is initialized to a value of
0000 0000 0001 744116 .

BC_TAG_ADDR

UNDEFINED

EI_STAT

UNDEFINED

EI_ADDR

UNDEFINED

FILL_SYN

UNDEFINED

PALcode must read twice to unlock.

Note
The Bcache parameters BC_SIZE (size), BC_RD_SPD (read speed),
BC_WR_SPD (write speed), and BC_WE_CTL (write-enable control) are
all configured to default values on reset and must be initialized in the
BC_CONFIG register before enabling the Bcache.

7.9 Timeout Reset
The instruction fetch/decode unit and branch unit (Ibox) contains a timer
that times out when a very long period of time passes with no instruction
completing. When this timeout occurs, an internal reset event occurs. This
clears sufficient internal state to allow the CPU to begin executing again.
Registers, IPRs (except as noted in Table 7–2), and caches are not affected.
Dispatch to the PALcode MCHK trap entry point occurs immediately.

7.10 IEEE 1149.1 Test Port Reset
Signal trst_l must be asserted when sys_reset_l is asserted or when dc_ok_h
is deasserted. Continuous trst_l assertion during normal operation is used to
guarantee that the IEEE 1149.1 test port does not affect 21164 operation.

Preliminary—Subject to Change—July 1996 7–13

8
Error Detection and Error Handling
This chapter provides an overview of the 21164’s error handling strategy.
Each internal cache (instruction cache [Icache], data cache [Dcache], and
second-level cache [Scache]) implements parity protection for tag and data.
Error correction code (ECC) protection is implemented for memory and backup
cache (Bcache) data. (The implementation provides detection of all double-bit
errors and correction of all single-bit errors.) Correctable instruction stream
(Istream) and data stream (Dstream) ECC errors are corrected in hardware
without privileged architecture library code (PALcode) intervention. Bcache
tags are parity protected. The instruction fetch/decode unit and branch unit
(Ibox) implements logic that detects when no progress has been made for a
very long time and forces a machine check trap.
PALcode handles all error traps (machine checks and correctable error
interrupts). Where possible, the address of affected data is latched in an IPR.
Most of the Istream errors can be retried by the operating system because the
machine check occurs before any part of the instruction causing the error is
executed. In some other cases, the system may be able to recover from an error
by terminating all processes that had access to the affected memory location.

8.1 Error Flows
The following flows describe the events that take place during an error, the
recommended responses necessary to determine the source of the error, and the
suggested actions to resolve them.

8.1.1 Icache Data or Tag Parity Error
•

Machine check occurs before the instruction causing the parity error is
executed.

•

EXC_ADDR contains either the PC of the instruction that caused the
parity error or that of an earlier trapping instruction.

•

ICPERR_STAT<TPE> or <DPE> is set.

Preliminary—Subject to Change—July 1996 8–1

8.1 Error Flows

•

Can be retried.
Note
The Icache is not flushed by hardware in this event. If an Icache parity
error occurs early in the PALcode routine at the machine check entry
point, an infinite loop may result.

•

Recommendation: Flush the Icache early in the MCHK routine.

8.1.2 Scache Data Parity Error—Istream
•

Machine check occurs before the instruction causing the parity error is
executed.

•

Bad data may be written to the Icache or Icache refill buffer and validated.

•

Can be retried if there are no multiple errors.

•

Recommendation: Flush the Icache to remove bad data. The Icache refill
buffer may be flushed by executing enough instructions to fill the refill
buffer with new data (32 instructions). Then flush the Icache again.

•

SC_STAT: SC_DPERR<7:0> is set; <SC_SCND_ERR> is set if there are
multiple errors.

•

SC_STAT: CBOX_CMD is IRD.

•

SC_ADDR: Contains the address of the 32-byte block containing the error.
(Bit 4 indicates which octaword was accessed first, but the error may be in
either octaword.)
Note
If the Istream parity error occurs early in the PALcode routine at the
machine check entry point, an infinite loop may result.

•

Recommendation: On data parity errors, it may be feasible for the
operating system to ‘‘flush’’ the block of data out of the Scache by
requesting a block of data with the same Bcache index, but a different
tag. This may not be feasible on tag parity errors, because the tag address
is suspect. If the requested block is loaded with no problems, then the
‘‘bad data’’ has been replaced. If the ‘‘bad data’’ is marked dirty, then when
the new data tries to replace the old data, another parity error may result

8–2 Preliminary—Subject to Change—July 1996

8.1 Error Flows

during the write-back (this is a reason not to attempt this in PALcode,
because a MCHK from PALcode is always fatal).

8.1.3 Scache Tag Parity Error—Istream
•

Machine check occurs before the instruction causing the parity error is
executed.

•

Bad data may be written to the Icache or Icache refill buffer and validated.

•

Cannot be retried. Probably will not be able to recover by deleting a single
process because the exact address is unknown.

•

SC_STAT: SC_TPERR<2:0> is set; <SC_SCND_ERR> is set if there are
multiple errors.

•

SC_STAT: CBOX_CMD is IRD.

•

8.1.4 Scache Data Parity Error—Dstream Read/Write, READ_DIRTY
•

Machine check occurs. Machine state may have changed.

•

Cannot be retried, but may only need to delete the process if data is
confined to a single process and no second error occurred.

•

SC_STAT: SC_DPERR<7:0> is set; SC_SCND_ERR is set if there are
multiple errors.

•

SC_STAT: CBOX_CMD is DRD, DWRITE, or READ_DIRTY.

•

SC_ADDR: Contains the address of the 32-byte block containing the error.
(Bit 4 indicates which octaword was accessed first, but the error may be in
either octaword.)

Preliminary—Subject to Change—July 1996 8–3

8.1 Error Flows

8.1.5 Scache Tag Parity Error—Dstream or System Commands
•

Machine check occurs. Machine state may have changed.

•

Cannot be retried. Probably will not be able to recover by deleting a single
process because the exact address is unknown.

•

SC_STAT: SC_TPERR<7:0> is set; <SC_SCND_ERR> is set if there are
multiple errors.

•

SC_STAT: CBOX_CMD is DRD, DWRITE, READ_DIRTY, SET_SHARED,
or INVAL.

•

SC_ADDR: records physical address bits <39:04> of location with error.

8.1.6 Dcache Data Parity Error
•

Machine check occurs. Machine state may have changed.

•

Cannot be retried, but may only need to delete the process if data is
confined to a single process and no second error occurred.

•

DCPERR_STAT: <DP0> or <DP1> is set. <LOCK> is set. <SEO> is set if
there are multiple errors.
Note
For multiple parity errors in the same cycle, the <SEO> bit is not set,
but more than one error bit will be set.

•

VA: Contains the virtual address of the quadword with the error.

•

MM_STAT locked. Contents contain information about instruction causing
parity error.
Note
Fault information on another instruction in same cycle may be lost.

8–4 Preliminary—Subject to Change—July 1996

8.1 Error Flows

8.1.7 Dcache Tag Parity Error
•

Machine check occurs. Machine state may have changed.

•

DCPERR_STAT: <TP0> or <TP1> is set. <LOCK> is set. <SEO> is set if
there are multiple errors.
Note
For multiple parity errors in the same cycle, the <SEO> bit is not set,
but more than one error bit will be set.

•

VA: Contains the virtual address of the Dcache block (hexword) with the
error.

•

MM_STAT locked. Contents contain information about instruction causing
parity error. <WR> bit is set if error occurred on a store instruction.
Note
Fault information on another instruction in the same cycle may be lost.

•

Probably will not be able to recover by deleting a single process, because
exact address is unknown, and a load may have falsely hit.

8.1.8 Istream Uncorrectable ECC or Data Parity Errors (Bcache or
Memory)
•

Machine check occurs before the instruction causing the error is executed.

•

Bad data may be written to the Icache or Icache refill buffer and validated.

•

Can be retried if there are no multiple errors.

•

Must flush Icache to remove bad data. The Icache refill buffer may be
flushed by executing enough instructions to fill the refill buffer with new
data (32 instructions). Then flush the Icache again.

•

EI_STAT: <UNC_ECC_ERR> is set; <SEO_HRD_ERR> is set if there are
multiple errors.

•

EI_STAT: <EI_ES> is set if source of fill data is memory/system, clear if
Bcache.

•

EI_STAT: <FIL_IRD> is set.

Preliminary—Subject to Change—July 1996 8–5

8.1 Error Flows

•

EI_ADDR: Contains the physical address bits <39:04> of the octaword
associated with the error.

•

FILL_SYN: Contains syndrome bits associated with the failing octaword.
This register contains byte parity error status if in parity mode.

•

BC_TAG_ADDR: Holds results of external cache tag probe if external cache
was enabled for this transaction.
Note
If the Istream ECC or parity error occurs early in the PALcode routine
at the machine check entry point, an infinite loop may result.

•

Recommendation: On data ECC/parity errors, it may be feasible for
the operating system to ‘‘flush’’ the block of data out of the Bcache by
requesting a block of data with the same Bcache index, but a different tag.
If the requested block is loaded with no problems, then the ‘‘bad data’’ has
been replaced. If the ‘‘bad data’’ is marked dirty, then when the new data
tries to replace the old data, another ECC/parity error may result during
the write-back (this is a reason not to attempt this in PALcode, because a
MCHK from PALcode is always fatal).

8.1.9 Dstream Uncorrectable ECC or Data Parity Errors (Bcache or
Memory)
•

Machine check occurs. Machine state may have changed.

•

Cannot be retried, but may only need to delete the process if data is
confined to a single process and no second error occurred.

•

EI_STAT: <UNC_ECC_ERR> is set; <SEO_HRD_ERR> is set if there are
multiple errors.

•

EI_STAT: <EI_ES> is set if source of fill data is memory/system, is clear if
Bcache.

•

EI_STAT: <FIL_IRD> is clear.

•

EI_ADDR: Contains the physical address bits <39:04> of the octaword
associated with the error.

•

FILL_SYN: Contains syndrome bits associated with the failing octaword.
This register contains byte parity error status if in parity mode.

8–6 Preliminary—Subject to Change—July 1996

8.1 Error Flows

•

BC_TAG_ADDR: Holds results of external cache tag probe if external cache
was enabled for this transaction.

8.1.10 Bcache Tag Parity Errors—Istream
•

Machine check occurs before the instruction causing the error is executed.

•

Bad data may be written to the Icache or Icache refill buffer and validated.

•

Can be retried if there are no multiple errors.

•

Must flush Icache to remove bad data. The Icache refill buffer may be
flushed by executing enough instructions to fill the refill buffer with new
data (32 instructions). Then flush the Icache again.

•

EI_STAT: <BC_TPERR> or <BC_TC_PERR> is set; <SEO_HRD_ERR> is
set if there are multiple errors.

•

EI_STAT: <EI_ES> is clear.

•

EI_STAT: <FIL_IRD> is set.

•

EI_ADDR: Contains the physical address bits <39:04> of the octaword
associated with the error.

•

BC_TAG_ADDR: Holds results of external cache tag probe.
Note
The Bcache hit is determined based on the tag alone, not the parity bit.
The victim is processed according to the status bits in the tag, ignoring
the control field parity. PALcode can distinguish fatal from nonfatal
occurrences by checking for the case in which a potentially dirty block
is replaced without the victim being properly written back and the case
of false hit when the tag parity is incorrect.

8.1.11 Bcache Tag Parity Errors—Dstream
•

Machine check occurs. Machine state may have changed.

•

Cannot be retried, but may only need to delete the process if data is
confined to a single process and no second error occurred. Bcache hit
is determined based on the tag alone, not the parity bit. The victim is
processed according to the status bits in the tag, ignoring the control
field parity. PALcode can distinguish fatal from nonfatal occurrences by
checking for the case in which a potentially dirty block is replaced without

Preliminary—Subject to Change—July 1996 8–7

8.1 Error Flows

the victim being properly written back and the case of false hit when the
tag parity is incorrect.
•

EI_STAT: <BC_TPERR> or <BC_TC_PERR> is set; <SEO_HRD_ERR> is
set if there are multiple errors.

•

EI_STAT: <EI_ES> is clear.

•

EI_STAT: <FIL_IRD> is clear.

•

EI_ADDR: Contains the physical address bits <39:04> of the octaword
associated with the error.

•

BC_TAG_ADDR: Holds results of external cache tag probe.

8.1.12 System Command/Address Parity Error
•

Machine check occurs. Machine state may have changed.

•

EI_STAT: <EI_PAR_ERR> is set; <SEO_HRD_ERR> is set if there are
multiple errors.

•

EI_STAT: <EI_ES> is set.

•

EI_ADDR: Contains the physical address bits <39:04> of the octaword
associated with the error.

•

BC_TAG_ADDR: Holds results of external cache tag probe if external cache
was enabled for this transaction.

•

When the 21164 detects a command or address parity error, the command
is unconditionally NOACKed.
Note
For a sysclk-to-CPU clock ratio of 3, if the 21164 detects a system
command/address parity error on a NOP, and immediately receives a
valid command from the system, then the 21164 may not acknowledge
the command. The 21164 does take the machine check.

8.1.13 System Read Operations of the Bcache
The 21164 does not check the ECC on outgoing Bcache data. If it is bad, the
receiving processor will detect it.

8–8 Preliminary—Subject to Change—July 1996

8.1 Error Flows

8.1.14 Istream or Dstream Correctable ECC Error (Bcache or Memory)
•

The 21164 hardware corrects the data before filling the Scache and Icache.
The Dcache is completely invalidated. The data in the Bcache contains the
ECC error, but is scrubbed by PALcode in the correctable error interrupt
routine. (Using LDxL or STxC, if the STxC fails, the location can be
assumed to be scrubbed.)

•

A separately maskable correctable error interrupt occurs at IPL 31 (same
as machine check). (Masked by clearing ICSR<CRDE>.)

•

ISR: <CRD> is set.

•

EI_STAT: <COR_ECC_ERR> is set.

•

EI_STAT: <FIL_IRD> is set if Istream; is clear if Dstream.

•

EI_STAT: <EI_ES> is clear if source of error is Bcache, is set otherwise.

•

EI_ADDR: Contains the physical address bits <39:04> of the octaword
associated with the error.

•

FILL_SYN: Contains syndrome bits associated with the octaword
containing the ECC error.

•

BC_TAG_ADDR: Unpredictable (not loaded on correctable errors).
Note
There will be performance degradation in systems when extremely
high rates of correctable ECC errors are present due to the internal
handling of this error (the implementation utilizes a replay trap and
automatic Dcache flush to prevent use of the incorrect data).

8.1.15 Fill Timeout (FILL_ERROR_H)
•

For systems in which fill timeout can occur, the system environment should
detect fill timeout and cleanly terminate the reference to 21164. If the
system environment expects fill timeout to occur, it should detect them.
If it does not expect them (as might be true in small systems with fixed
memory access timing), it is likely that the internal Ibox timeout will
eventually detect a stall if a fill fails to occur. To properly terminate a
fill in an error case, the fill_error_h pin is asserted for one cycle and the
normal fill sequence involving the fill_h, fill_id_h, and dack_h pins is
generated by the system environment.

Preliminary—Subject to Change—July 1996 8–9

8.1 Error Flows

•

A fill_error_h assertion forces a PALcode trap to the MCHK entry point,
but has no other effect.
Note
No internal status is saved to show that this happened. If necessary,
systems must save this status, and include read operations of the
appropriate status registers in the MCHK PALcode.

8.1.16 System Machine Check
•

The 21164 has a maskable machine check interrupt input pin. It is used by
system environments to signal fatal errors that are not directly connected
to a read access from the 21164. It is masked at IPL 31 and anytime the
21164 is in PALmode.

•

ISR: <MCK> is set.

8.1.17 Ibox Timeout
•

When the Ibox detects a timeout, it causes a PALcode trap to the MCHK
entry point.

•

Simultaneously, a partial internal reset occurs: most states except IPR
state is reset. This should not be depended on by systems in which fill
timeouts occur in typical use (such as, operating system or console code
probing locations to determine if certain hardware is present). The purpose
of this error detection mechanism is to attempt to prevent system hang in
order to write a machine check stack frame.

•

ICPERR_STAT: <TMR> is set.

8.1.18 cfail_h and Not cack_h
•

Assertion of cfail_h in a sysclk cycle in which cack_h is not asserted
causes the 21164 to immediately execute a partial internal reset.

•

PALcode trap to the MCHK entry point.

•

Simultaneously, a partial internal reset occurs: most states except IPR
state is reset.

•

ICPERR_STAT: <TMR> is set.

8–10 Preliminary—Subject to Change—July 1996

8.1 Error Flows

•

This can be used to restore 21164 and the external environment to a
consistent state after the external environment detects a command or
address parity error.
Note
There is no internal status saved to differentiate the cfail_h/no cack_
h case from the Ibox timeout reset case. If necessary, systems must
save this status, and include read operations of the appropriate status
registers in the MCHK PALcode.

8.2 MCHK Flow
The following flow is the recommended IPR access order to determine the
source of a machine check.
•

Must flush Icache to remove bad data on Istream errors. The Icache refill
buffer may be flushed by executing enough instructions to fill the refill
buffer with new data (32 instructions). Then flush the Icache again.

•

Read EXC_ADDR.

•

If EXC_ADDR=PAL, then halt.

•

Issue MB to clear out Mbox/Cbox before reading Cbox registers or issuing
DC_FLUSH.

•

Flush Dcache to remove bad data on Dstream errors.

•

Read ICSR.

•

Read ICPERR_STAT.

•

Read DCPERR_STAT.

•

Read SC_ADDR.

•

Use register dependencies or MB to ensure read operation of SC_ADDR
finishes before subsequent read operation of SC_STAT.

•

Read SC_STAT (unlocks SC_ADDR).

•

Read EI_ADDR, BC_TAG_ADDR, and FILL_SYN.

•

Use register dependencies or MB to ensure read operations of EI_ADDR,
BC_TAG_ADDR, and FILL_SYN finish before subsequent read operation of
EI_STAT.

•

Read EI_STAT and save (unlocks EI_ADDR, BC_TAG_ADDR, FILL_SYN).
Preliminary—Subject to Change—July 1996 8–11

8.2 MCHK Flow

•

Read EI_STAT again to be sure it is unlocked, discard result.

•

Check for cases that cannot be retried. If any one of the following are true,
then skip retry:
EI_STAT<TPERR>
EI_STAT<TC_PERR>
EI_STAT<EI_PAR_ERR>
EI_STAT<SEO_HRD_ERR>
EI_STAT<UNC_ECC_ERR> and not EI_STAT<FIL_IRD>
DCPERR_STAT<LOCK>
SC_STAT<SC_SCND_ERR>
SC_STAT<SC_TPERR>
Not (SC_STAT<CMD> = IRD) and SC_STAT<SC_DPERR>
ICPERR_STAT<TMR>
ISR<MCK>

•

If none of the previous conditions are true, then there is either an IRD that
can be retried or the source of the MCHK is a fill_error_h. Add code for
query of system status.

•

The case can be retried if any one or several of the following are true (and
none of the previous conditions were true):
EI_STAT<UNC_ECC_ERR> and EI_STAT<FIL_IRD>
SC_STAT<SC_DPERR> and (SC_STAT<CMD> = IRD)
ICPERR_STAT<TPE>
ICPERR_STAT<DPE>

•

Unlock the following IPRs:
ICPERR_STAT (write 0x1800)
DCPERR_STAT (write 0x03)
VA, SC_STAT, and EI_STAT are already unlocked.

•

Check for arithmetic exceptions:
Read EXC_SUM.

8–12 Preliminary—Subject to Change—July 1996

8.2 MCHK Flow

Check for arithmetic errors and handle according to operating-systemspecific requirements.
Clear EXC_SUM (unlocks EXC_MASK).
•

Report the processor-uncorrectable MCHK according to operating-systemspecific requirements.

8.3 Processor-Correctable Error Interrupt Flow (IPL 31)
The following flow is the recommended way to report correctable errors:
•

Arrived here through interrupt routine because ISR<CRD> bit set.

•

Read EI_ADDR and FILL_SYN.

•

Use register dependencies or MB to ensure read operations of EI_ADDR
and FILL_SYN finish before subsequent read operation of EI_STAT.

•

Read EI_STAT. (Unlocks EI_STAT, EI_ADDR, and FILL_SYN.)

•

Scrub the memory location by using LDQ_L/STQ_C to one of the quadwords
in each octaword of the Bcache block whose address is reported in
EI_ADDR. No need to scrub I/O space addresses as these are noncacheable.

•

ACK the CRD Interrupt by writing a ‘‘0’’ to HWINT_CLR<CRDC>.

•

No need to unlock any registers because conditions that would cause a lock
would also cause a MCHK. VA will not be locked because DTB_MISS and
FAULT PALcode routines will not ever be interrupted.

•

Report the processor-correctable MCHK according to operating-systemspecific requirements.
Note
Only read EI_STAT once in the CRD flow, and then only if ISR<CRD>
is set. If an uncorrectable error were to occur just after a second read
operation from EI_STAT was issued, then there could be a race between
the unlocking of the register and the loading of the new error status,
potentially resulting in the loss of the error status.

Preliminary—Subject to Change—July 1996 8–13

8.4 MCK_INTERRUPT Flow

8.4 MCK_INTERRUPT Flow
•

Arrived here through interrupt routine because ISR<MCK> bit set.

•

Report the system-uncorrectable MCHK according to operating-systemspecific requirements.

8.5 System-Correctable Error Interrupt Flow (IPL 20)
The system-correctable error interrupt is system specific.

8–14 Preliminary—Subject to Change—July 1996

9
Electrical Data
This chapter describes the electrical characteristics of the 21164 component
and its interface pins. It is organized as follows:
•

Electrical characteristics

•

dc characteristics

•

Clocking scheme

•

ac characteristics

•

Power supply considerations

9.1 Electrical Characteristics
Table 9–1 lists the maximum ratings for the 21164.
Table 9–1

Alpha 21164 Absolute Maximum Ratings

Characteristics

Ratings

Storage temperature

–55°C to 125°C (–67°F to 257°F)

Junction temperature

15°C to 90°C (59°F to 194°F)

Supply voltage
Input or output applied

Vss –0.5 V, Vdd 3.6 V
1

Typical worst case power
@Vdd = 3.3 V
Frequency = 266 MHz
Frequency = 300 MHz
Frequency = 333 MHz

–0.5 V to 6.3 V

46 W
51 W
56 W

1 Refer to Section 9.5.2.

Preliminary—Subject to Change—July 1996 9–1

9.1 Electrical Characteristics

Caution
Stress beyond the absolute maximum rating can cause permanent
damage to the 21164. Exposure to absolute maximum rating conditions
for extended periods of time can affect the 21164 reliability.

9.2 dc Characteristics
The 21164 is designed to run in a CMOS/TTL environment. The 21164 is
tested and characterized in a CMOS environment.

9.2.1 Power Supply
The Vss pins are connected to 0.0 V, and the Vdd pins are connected to 3.3 V,
65%.

9.2.2 Input Signal Pins
Nearly all input signals are ordinary CMOS inputs with standard TTL
levels (see Table 9–2). (See Section 9.3.1 for a description of an exception—
osc_clk_in_h,l.)
After power has been applied, input and bidirectional pins can be driven to a
maximum dc voltage of 6.3 V (6.8 V for 1 ns) without harming the 21164. (It is
not necessary to use static RAMs with 3.3-V outputs.)

9.2.3 Output Signal Pins
Output pins are ordinary 3.3-V CMOS outputs. Although output signals are
rail-to-rail, timing is specified to V 2dd .
Bidirectional pins are either input or output pins, depending on control timing.
When functioning as output pins, they are ordinary 3.3-V CMOS outputs.
Table 9–2 shows the CMOS dc input and output pins.

9–2 Preliminary—Subject to Change—July 1996

9.2 dc Characteristics

Table 9–2 CMOS dc Input/Output Characteristics
Parameter

Requirements

Symbol

Description

Min.

Max.

Units

Test Conditions

Vih

High-level input voltage

2.0

—

Vil

Low-level input voltage

—

0.8

—

Voh

High-level output voltage

2.4

—

Ioh = –6.0 mA

Vol

Low-level output voltage

—

Iol = 6.0 mA

Input with pull-down
leakage current

—

650

Iil_pd

Vin = 0 V

Iih_pd

Input with pull-down
current

—

200

Vin = 2.4 V

Iil_pu

Input with pull-up current

—

–800

Iih_pu

Input with pull-up leakage
current

—

650

A
A

Iozl_pd

Output with pull-down
leakage current (tristate)

—

6100

Vin = 0 V

Iozh_pd

Output with pull-down
current (tristate)

—

300

Vin = 2.4 V

Iozl_pu

Output with pull-up current
(tristate)

—

–800

Vin = 0.4 V

Iozh_pu

Output with pull-up
leakage current (tristate)

—

6100

Vin = Vdd V

Idd

Peak power supply current

—

Vdd = 3.465 V
Frequency = 266 MHz

Idd

Peak power supply current

—

Vdd = 3.465 V
Frequency = 300 MHz

Idd

Peak power supply current

—

Vdd = 3.465 V
Frequency = 333 MHz

0.4

Vin = 0.4 V
Vin = Vdd V

Most pins have low current pull-down devices to Vss. However, two pins have
a pull-up device to Vdd. The pull-downs (or pull-ups) are always enabled. This
means that some current will flow from the 21164 (if the pin has a pull-up
device) or into the 21164 (if the pin has a pull-down device) even when the pin
is in the high-impedance state. All pins have pull-down devices, except for the
pins in the following table:

Preliminary—Subject to Change—July 1996 9–3

9.2 dc Characteristics

Signal Name

Notes

tms_h

Has a pull-up device

tdi_h
osc_clk_in_h

Has a pull-up device
50 to Vterm ( V dd ) (See Figure 9–1)

osc_clk_in_l

to Vterm ( V 2dd ) (See Figure 9–1)

temp_sense

150

to Vss

9.3 Clocking Scheme
The differential input clock signals osc_clk_in_h,l run at two times the
internal frequency of the time base for the 21164. Input clocks are divided by
two onchip to generate a 50% duty cycle clock for internal distribution. The
output signal cpu_clk_out_h toggles with an unspecified propagation delay
relative to the transitions on osc_clk_in_h,l.
System designers have a choice of two system clocking schemes to run the
21164 synchronous to the system:
1. The 21164 generates and drives out a system clock, sys_clk_out1_h,l. It
runs synchronous to the internal clock at a selected ratio of the internal
clock frequency. There is a small clock skew between the internal clock and
sys_clk_out1_h,l.
2. The 21164 synchronizes to a system clock, ref_clk_in_h, supplied by
the system. The ref_clk_in_h clock runs at a selected ratio of the 21164
internal clock frequency. The internal clock is synchronized to the reference
clock by an onchip digital phase-locked loop (DPLL).
Refer to Section 4.2 for more information on clock functions.

9.3.1 Input Clocks
The differential input clocks osc_clk_in_h,l provide the time base for the
chip when dc_ok_h is asserted. These pins are self-biasing, and must be
capacitively coupled to the clock source on the module, or they can be directly
driven. The terminations on these signals are designed to be compatible with
system oscillators of arbitrary dc bias. The oscillator must have a duty cycle
of 60%/40% or tighter. Figure 9–1 shows the input network and the schematic
equivalent of osc_clk_in_h,l terminations.

9–4 Preliminary—Subject to Change—July 1996

9.3 Clocking Scheme

Figure 9–1 osc_clk_in_h,l Input Network and Terminations

Module Circuitry

Onchip Circuitry
6 nH

osc_clk_in_h

*
3.3 pF

3.3 pF

Vss

Vdd
2
130 to
600

50
Oscillator

To
Differential
Amplifier

47 pF
3.3 pF

3.3 pF

osc_clk_in_l
6 nH

Note:
Coupling Capacitors 47pF to 220 pF

LJ-04035.AI

Ring Oscillator
When signal dc_ok_h is deasserted, the clock outputs follow the internal ring
oscillator. The 21164 runs off the ring oscillator, just as it would when an
external clock is applied. The frequency of the ring oscillator varies from chip
to chip within a range of 10 MHz to 100 MHz. This corresponds to an internal
CPU clock frequency range of 5 MHz to 50 MHz. The system clock divisor is
forced to 8, and the sys_clk_out2 delay is forced to 3.
Clock Sniffer
A special onchip circuit monitors the osc_clk_in pins and detects when input
clocks are not present. When activated, this circuit switches the 21164 clock
generator from the osc_clk_in pins to the internal ring oscillator. This
happens independently of the state of the dc_ok_h pin. The dc_ok_h pin
functions normally if clocks are present on the osc_clk_in pins.

Preliminary—Subject to Change—July 1996 9–5

9.3 Clocking Scheme

9.3.2 Clock Termination and Impedance Levels
In Figure 9–1, the clock is designed to approximate a 50- termination for the
purpose of impedance matching for those systems that drive input clocks across
long traces. The clock input pins appear as a 50- series termination resistor
connected to a high impedance voltage source. The voltage source produces a
nominal voltage value of V 2dd . The source has an impedance of between 130
and 600 . This voltage is called the self-bias voltage and sources current when
the applied voltage at the clock input pins is less than the self-bias voltage.
It sinks current when the applied voltage exceeds the self-bias voltage. This
high impedance bias driver allows a clock source of arbitrary dc bias to be ac
coupled to the 21164. The peak-to-peak amplitude of the clock source must
be between 0.6 V and 3.0 V. Either a square-wave or a sinusoidal source may
be used. Full-rail clocks may be driven by testers. In any case, the oscillator
should be ac coupled to the osc_clk_in_h,l inputs by 47 pF through 220 pF
capacitors.

9.3.3 ac Coupling
Using series coupling (blocking) capacitors renders the 21164 clock input pins
insensitive to the oscillator’s dc level. When connected this way, oscillators
with any dc offset relative to Vss can be used provided they can drive a signal
into the osc_clk_in_h,l pins with a peak-to-peak level of at least 600 mV, but
no greater than 3.0 V peak to peak.
The value of the coupling capacitor is not overly critical. However, it should be
sufficiently low impedance at the clock frequency so that the oscillator’s output
signal (when measured at the osc_clk_in_h,l pins) is not attenuated below
the 600 mV peak-to-peak lower limit. For sine waves or oscillators producing
nearly sinusoidal (pseudo square wave) outputs, 220 pF is recommended at
533.3 MHz (266.6 MHz 2 2). A high quality dielectric such as NPO is required
to avoid dielectric losses.
Table 9–3 shows the input clock specification.

9–6 Preliminary—Subject to Change—July 1996

9.3 Clocking Scheme

Table 9–3 Input Clock Specification
Signal Parameter

Minimum

Maximum

Unit

osc_clk_in_h,l symmetry

osc_clk_in_h,l voltage

0.6

3.0

V (peak-to-peak)

osc_clk_in_h,l Z input

Refer to Figure 9–2, Clock Input Differential
Impedance.

Tfreq (CPU clock frequency)

100

3331

MHz

Tcycle ( T freq )

1 Maximum CPU clock frequency is either 333, 300, or 266 MHz, depending upon part variation.

Preliminary—Subject to Change—July 1996 9–7

9.3 Clocking Scheme

Figure 9–2 Clock Input Differential Impedance

140

120

Impedance in Ohms

100

0
10

100
Frequency in MHz

1000

Differential Impedance ocs_clk_in_h to osc_clk_in_l
LJ-04724.AI5

9–8 Preliminary—Subject to Change—July 1996

9.4 ac Characteristics

9.4 ac Characteristics
This section describes the ac timing specifications for the 21164.

9.4.1 Test Configuration
All input timing is specified relative to the crossing of standard TTL input
levels of 0.8 V and 2.0 V. Output timing is to the nominal CMOS switch point
of V 2dd (see Figure 9–3).
Figure 9–3 Input/Output Pin Timing
Tcycle
Internal
CPU Clock

50%

Tdsu

Tdh

Vdd

2.0 V

Input
Signals

0.8 V

Vss

Input Timing

Internal
CPU Clock

50%

Tdd

Vdd
Output
Signals

Vdd
2

Vss

Output Timing
MK−1455−12

Preliminary—Subject to Change—July 1996 9–9

9.4 ac Characteristics

Because the speed and complexity of microprocessors has increased
substantially over the years, it is necessary to change the way they are
tested. Traditional assumptions that all loads can be lumped into some
accumulation of capacitance cannot be employed any more. Rather, the model
of a transmission line with discrete loads is a much more realistic approach for
current test technology.
Typically, printed circuit board (PCB) etch has a characteristic impedance
of approximately 75 . This may vary from 60 to 90 with tolerances.
If the line is driven in the electrical center, the load could be as low as
30 . Therefore, a characteristic impedance range of 30 to 90 could be
experienced.
The 21164 output drivers are designed with typical printed circuit board
applications in mind rather than trying to accommodate a 40-pF test load
specification. As such, it ‘‘launches’’ a voltage step into a characteristic
impedance, ranging from 30 to 90 .
To prevent signal quality problems due to overshoot or ringing, ‘‘near end’’
terminated transmission line design rules are used. By combining the source
impedance of the driver transistors with an additional 20- onchip resistor, a
source impedance of approximately 40 is achieved. Additionally, a load value
of 10 pF, when added to the PCB etch delays, provides a realistic estimate of
actual system timing. When employing this test configuration, the signal at
the end of the line will transition cleanly through the TTL input specification
range of 0.8 V to 2.0 V without plateaus, or reversal into the range.

9.4.2 Pin Timing
The following sections describe Bcache loop timing, sys_clk-based system
timing, and reference clock-based system timing.
9.4.2.1 Backup Cache Loop Timing
The 21164 can be configured to support an optional offchip backup cache
(Bcache). Private Bcache read or write (Scache victims) transactions initiated
by the 21164 are independent of the system clocking scheme. Bcache loop
timing must be an integer multiple of the 21164 cycle time.
Table 9–4 lists the Bcache loop timing.

9–10 Preliminary—Subject to Change—July 1996

9.4 ac Characteristics

Table 9–4 Bcache Loop Timing
Signal

Specification

Value

Name

data_h<127:0>

Input setup

1.1 ns

Tdsu

data_h<127:0>

Input hold

0.0 ns

Tdh
1

index_h<25:4>

Output delay

Tdd + 0.4 ns

index_h<25:4>

Output hold time

Tmdd

Tioh

data_h<127:0>

Output delay

Tdd + Tcycle + 0.4 ns1

Tdod

data_h<127:0>

Output hold

Tmdd + Tcycle

Tdoh

Tiod

1 The value 0.4 ns accounts for onchip driver and clock skew.

Outgoing Bcache index and data signals are driven off the internal clock edge
and the incoming Bcache tag and data signals are latched on the same internal
clock edge. Table 9–5 shows the output driver characteristics.
Table 9–5 Output Driver Characteristics
Specification

40-pF Load

10-pF Load

Name

Maximum driver delay

2.6 ns

1.6 ns

Tdd

Minimum driver delay

1.0 ns

Tmdd

Output pin timing is specified for lumped 40-pF and 10-pF loads. In some
cases, the circuit may have loads higher than 40 pF. The 21164 can safely drive
higher loads provided the average charging or discharging current from each
pin is 10 mA or less. The following equation can be used to determine the
maximum capacitance that can be safely driven by each pin:
Cmax (in pF) = 3t, where t is the waveform period (measured from rising
to rising or falling to falling edge), in nanoseconds.
For example, if the waveform appearing on a given I/O pin has a 20.4-ns
period, it can safely drive up to and including 61 pF.
Figure 9–4 shows the Bcache read and write timing.

Preliminary—Subject to Change—July 1996 9–11

9.4 ac Characteristics

Figure 9–4 Bcache Timing

Bcache Loop (Read)
Tiod

Tdsu

Tioh

CPU Clock
Index Out
Data In
Bcache Cycle

Bcache Loop (Write)
Tdod
Tiod

Tdh

Tdoh
Tioh

CPU Clock
Index Out
Data Out
Bcache Cycle
LJ-03409-TI0

9.4.2.2 sys_clk-Based Systems
All timing is specified relative to the rising edge of the internal CPU clock.
Table 9–6 shows 21164 system clock sys_clk_out1_h,l output timing. Setup
and hold times are specified independent of the relative capacitive loading of
sys_clk_out1_h,l, addr_h<39:4>, data_h<127:0>, and cmd_h<3:0> signals.
The ref_clk_in_h signal must be tied to Vdd for proper operation.

9–12 Preliminary—Subject to Change—July 1996

9.4 ac Characteristics

Table 9–6

Alpha 21164 System Clock Output Timing (sysclk=Tø )

Signal

Specification

Value

Name

sys_clk_out1_h,l

Output delay

Tdd

Tsysd

sys_clk_out1_h,l

Minimum output delay

Tmdd

Tsysdm

data_bus_req_h,
data_h<127:0>,
addr_h<39:4>

Input setup

1.1 ns

Tdsu

data_bus_req_h,
data_h<127:0>,
addr_h<39:4>

Input hold

0 ns

Tdh

addr_h<39:4>

Output delay

Tdd + 0.4 ns1

Taod

addr_h<39:4>

Output hold time

Tmdd

Taoh
1

data_h<127:0>

Output delay

Tdd + Tcycle + 0.4 ns

Tdod2

data_h<127:0>

Output hold time

Tmdd + Tcycle1

Tdoh2

Non-Pipe_Latch Mode
addr_bus_req_h

Input setup

3.8 ns

Tabrsu

addr_bus_req_h

Input hold

–1.0 ns

Tabrh

dack_h

Input setup

3.4 ns

Tntacksu

cack_h

Input setup

3.7 ns

Tntcacksu

cack, dack

Input hold

–1.0 ns

Tntackh

Pipe_Latch Mode3
addr_bus_req_h,
cack_h, dack_h

Input setup

1.1 ns

Ttacksu

addr_bus_req_h,
cack_h, dack_h

Input hold

0 ns

Ttackh

1 The value 0.4 ns accounts for onchip driver and clock skew.
2 For all write transactions initiated by the 21164, data is driven one CPU cycle after the

sys_clk_out1 or index_h<25:4> pins.
3 In pipe_latch mode, control signals are piped onchip for one sys_clk_out1_h,l before usage.

Figure 9–5 shows sys_clk system timing.

Preliminary—Subject to Change—July 1996 9–13

9.4 ac Characteristics

Figure 9–5 sys_clk System Timing
Relationship of CPU Clock and sys_clk_out1
Tsysd
CPU Clock

sys_clk_out1

Memory Read (Turbo Mode)
Tsysd

Tsysd

sys_clk_out1
Taod

Ttacksu

Tdsu

Taoh

CPU Clock
Address/
Command Out
dack

Data In

Memory Read (Non-Turbo Mode)
Tsysd

Tsysd

sys_clk_out1

Tsysd
Tntacksu

Taod

Tdsu

Taoh

CPU Clock
Address/
Command Out

Tntcacksu

cack
dack

Data In
Tntackh
LJ-03410-TI0

9–14 Preliminary—Subject to Change—July 1996

9.4 ac Characteristics

9.4.2.3 Reference Clock-Based Systems
Systems that generate their own system clock expect the 21164 to synchronize
its sys_clk_out1_h,l outputs to their system clock. The 21164 uses a digital
phase-locked loop (DPLL) to synchronize its sys_clk_out1 signals to the
system clock that is applied to the ref_clk_in_h signal. For additional
information on reference clock timing, refer to Section 4.2.4.
Table 9–7 shows all timing relative to the rising edge of ref_clk_in_h.
Table 9–7

Alpha 21164 Reference Clock Input Timing

Signal

Specification

Value

Name

data_bus_req_h,
data_h<127:0>,
addr_h<39:4>

Input setup

1.1 ns

Tdsu

data_bus_req_h,
data_h<127:0>,
addr_h<39:4>

Input hold

0.5 x Tcycle

Troh

addr_h<39:4>

Output delay

Tdd + 0.5 x Tcycle + 0.9 ns1

Traod

addr_h<39:4>

Output hold
time

Tmdd

Traoh

data_h<127:0>

Output delay

Tdd + 1.5 x Tcycle + 0.9 ns1

Trdod2

data_h<127:0>

Output hold
time

Tmdd + Tcycle

Trdoh2

Non-Pipe_Latch Mode
addr_bus_req_h

Input setup

3.8 ns

Tntrabrsu

addr_bus_req_h

Input hold

0.5 x Tcycle

Tntrabrh

dack_h

Input setup

3.3 ns

Tntracksu

cack_h

Input setup

3.7 ns

Tntrcacksu

cack_h, dack_h

Input hold

(0.5 x Tcycle)

Tntrackh

1 The value 0.9 ns accounts for onchip skews that include 0.4 ns for driver and clock skew, phase

detector skews due to circuit delay (0.2 ns), and delay in ref_clk_in_h due to the package (0.3 ns).
2 For all write transactions initiated by the 21164, data is driven one CPU cycle later.

(continued on next page)

Preliminary—Subject to Change—July 1996 9–15

9.4 ac Characteristics

Table 9–7 (Cont.)

Alpha 21164 Reference Clock Input Timing

Signal

Specification

Value

Name

Pipe_Latch Mode3
addr_bus_req_h,
cack_h, dack_h

Input setup

1.1 ns

Ttracksu

addr_bus_req_h,
cack_h, dack_h

Input hold

0.5 x Tcycle

Ttrackh

3 In pipe_latch mode, control signals are piped onchip for one sys_clk_out1_h,l before usage.

9.4.3 Digital Phase-Locked Loop
Figure 9–6 and Table 9–8 describe the digital phase-locked loop (DPLL) stages
of operation.

9–16 Preliminary—Subject to Change—July 1996

9.4 ac Characteristics

Figure 9–6 ref_clk System Timing
Relationship of CPU Clock and ref_clk_in
1

CPU Clock
ref_clk_

Relationship of CPU Clock, ref_clk_in, and sys_clk_out1

CPU Clock
ref_clk_in
sys_clk_out1

Tsysd

Tsysd
LJ-03411-TI0

Table 9–8 ref_clk System Timing Stages
Stage

!
"
#
$

Description
The internal CPU clock rising edge coincides with the rising edge of
ref_clk_in_h.
The DPLL causes the internal CPU clock to stretch for one phase (1 cycle
of osc_clk_in_h,l).
The stretch causes ref_clk_in_h to lead the internal CPU clock by one
phase.
The CPU clock is always slightly faster than the external ref_clk_in_h
and gains on ref_clk_in_h over time. Eventually the gain equals one
phase and a new stretch phase follows.

Although systems that supply a ref_clk_in_h do not use sys_clk_out1_h,l,
a relationship between the two signals exists, just as in the sys_clk-based
systems, because the 21164 uses sys_clk_out1_h,l internally to determine
timing during system transactions.

Preliminary—Subject to Change—July 1996 9–17

9.4 ac Characteristics

9.4.4 Timing—Additional Signals
This section lists timing for all other signals.
Asynchronous Input Signals
The following is a list of the asynchronous input signals:
clk_mode_h
sys_reset_l1
perf_mon_h2
irq_h<3:0>2

dc_ok_h

ref_clk_in_h

mch_hlt_irq_h2

pwr_fail_irq_h2

sys_mch_chk_irq_h2

1 Signal sys_reset_l may be deasserted synchronously.
2 These signals can also be used synchronously.

Miscellaneous Signals
Table 9–9 and Table 9–10 list the timing for miscellaneous input-only and
output-only signals. All timing is expressed in nanoseconds.
Table 9–9 Input Timing for sys_clk_out- or ref_clk_in-Based Systems
Value

Name

Signal

Specification

sys_clk_out ref_clk_in

cfail_h, fill_h, fill_error_h, fill_id_h,
fill_nocheck_h, idle_bc_h, shared_h,
system_lock_flag_h

Input setup

1.1 ns

Tdsu

Input hold

0 ns

0.5 Tcycle Tdh

Troh

irq_h<3:0>, mch_hlt_irq_h, pwr_
fail_irq_h, sys_mch_chk_irq_h
Testability pins:
port_mode_h, srom_data_h,
srom_present_l

cfail_h, fill_h, fill_error_h, fill_id_h,
fill_nocheck_h, idle_bc_h, shared_h,
system_lock_flag_h
irq_h<3:0>, mch_hlt_irq_h, pwr_
fail_irq_h, sys_mch_chk_irq_h
sys_reset_l
Testability pins:
port_mode_h, srom_data_h,
srom_present_l

9–18 Preliminary—Subject to Change—July 1996

9.4 ac Characteristics

Table 9–10 Output Timing for sys_clk_out- or ref_clk_in-Based Systems
Clocking System Value
Signal

Specification sys_clk_out

ref_clk_in

Clocking System Name
sys_clk_out

ref_clk_in

Unidirectional Signals

addr_res_h,
Output
int4_valid_h,1
delay
scache_set_h,
srom_clk_h,
srom_oe_l,
victim_pending_h

Tdd+0.4 ns

Tdd+0.5 Tcycle+0.9 ns

Taod

Traod

addr_res_h,
Output
int4_valid_h,1
hold
scache_set_h,
srom_clk_h,
srom_oe_l,
victim_pending_h

Tmdd

Taoh

Traoh

int4_valid_h2

Output
delay

Tdd+Tcycle+0.4 ns Tdd+1.5 Tcycle+0.9 ns

Tdod

Trdod

int4_valid_h2

Output
hold

Tmdd+Tcycle

Tdoh

Trdoh

Input setup

1.1 ns

Tdsu

Input hold

0 ns

0.5 Tcycle

Tdh

Tsdadh

Bidirectional Signals
Input mode:
addr_cmd_par_h,
cmd_h,
data_check_h,1
tag_ctl_par_h,3
tag_dirty_h,3
tag_shared_h3
addr_cmd_par_h,
cmd_h,
data_check_h,1
tag_ctl_par_h,3
tag_dirty_h,3
tag_shared_h3
1 Read transaction
2 Write transaction
3 Fills from memory

(continued on next page)

Preliminary—Subject to Change—July 1996 9–19

9.4 ac Characteristics

Table 9–10 (Cont.) Output Timing for sys_clk_out- or ref_clk_in-Based Systems
Clocking System Value
Signal

Specification sys_clk_out

ref_clk_in

Output
delay

Tdd+0.4 ns

Tdd+0.5 Tcycle+0.9 ns

data_check_h2

Output
delay

addr_cmd_par_h,

Clocking System Name
sys_clk_out

ref_clk_in

Taod

Traod

Tdd+Tcycle+0.4 ns Tdd+1.5 Tcycle+0.9 ns

Tdod

Trdod

Output
hold

Tmdd

Taoh

Traoh

Output
hold

Tmdd+Tcycle

Tdoh

Trdoh

Bidirectional Signals
Output mode:
addr_cmd_par_h,
cmd_h,
tag_ctl_par_h,4
tag_dirty_h,4
tag_shared_h,4
tag_valid_h4

cmd_h,
tag_ctl_par_h,4
tag_dirty_h,4
tag_shared_h,4
tag_valid_h4
data_check_h2
2 Write transaction
4 Only for write broadcasts and system transactions

9–20 Preliminary—Subject to Change—July 1996

9.4 ac Characteristics

Signals in Table 9–11 are used to control Bcache data transfers. These signals
are driven off the CPU clock. The choice of sys_clk_out or ref_clk_in has no
impact on the timing of these signals.
Table 9–11 Bcache Control Signal Timing
Signal

Specification

Value

Name

tag_data_h, tag_data_par_h,
tag_valid_h

Input setup

1.1 ns

Tdsu

tag_data_h, tag_data_par_h,
tag_valid_h

Input hold

0 ns

Tdh

data_ram_oe_h, data_ram_we_h,1
tag_ram_oe_h, tag_ram_we_h1

Output delay

Tdd+0.4 ns

Taod

tag_data_h, tag_data_par_h,
tag_valid_h

Output delay

Tdd+0.4 ns

Taod

data_ram_oe_h, data_ram_we_h,1
tag_ram_oe_h, tag_ram_we_h1

Output hold

Tmdd

Taoh

tag_data_h, tag_data_par_h,
tag_valid_h

Output hold

Tmdd

Taoh

Input mode:

Output mode:

1 Pulse width for this signal is controlled through the BC_CONFIG IPR.

9.4.5 Timing of Test Features
Timing of 21164 testability features depends on the system clock rate and the
test port’s operating mode. This section provides timing information that may
be needed for most common operations.

9.4.6 Icache BiSt Operation Timing
The Icache BiSt is invoked by deasserting the external reset signal
sys_reset_l. Figure 9–7 shows the timing between various events relevant
to BiSt operations.

Preliminary—Subject to Change—July 1996 9–21

9.4 ac Characteristics

Figure 9–7 BiSt Timing Event–Time Line

Deassert
sys_reset_l

Deassert *
Internal Reset
BiSt Done
(T%Z_RESET_B_L) (test_status_h<1:0>=00)

BiSt Start
(test_status_h<1:0>=01)
t2
t1

MK−1455−09

The timing for deassertion of internal reset (time t2 , see asterisk) is valid
only if an SROM is not present (indicated by keeping signal srom_present_l
deasserted). If an SROM is present, the SROM load is performed once the
BiSt completes. The internal reset signal T%Z_RESET_B_L is extended until
the end of the SROM load (Section 9.4.7). In this case, the end of the time
line shown in Figure 9–7 connects to the beginning of the time line shown in
Figure 9–8.
Table 9–12 and Table 9–13 list timing shown in Figure 9–7 for some of the
system clock ratios. Time t1 is measured starting from the rising edge of sysclk
following the deassertion of the sys_reset_l signal.
Table 9–12 BiSt Timing for Some System Clock Ratios, Port Mode=Normal
(System Cycles)
Sysclk

System Cycles

Ratio

22644+2½

22645

19721+2½

19722

13291+14½ 13292

9–22 Preliminary—Subject to Change—July 1996

9.4 ac Characteristics

Table 9–13 BiSt Timing for Some System Clock Ratios, Port Mode=Normal
(CPU Cycles)
Sysclk

CPU Cycles

Ratio

67934½

67935

78886½

78888

105

199379½

199380

9.4.7 Automatic SROM Load Timing
The SROM load is triggered by the conclusion of BiSt if srom_present_l is
asserted. The SROM load occurs at the internal cycle time of approximately
126 CPU cycles for srom_clk_h, but the behavior at the pins may shift
slightly. Refer to Chapter 7 for more information on input signals, booting, and
the SROM interface port.
Timing events are shown in Figure 9–8 and are listed in Table 9–14 and
Table 9–15.
Figure 9–8 SROM Load Timing Event–Time Line

BiSt Done
(test_status_h
<1:0>=00)

Assert
srom_oe_l

Deassert
Internal Reset
Last Rise
srom_clk_h (T%Z_RESET_B_L)

First Rise
srom_clk_h

Deassert
srom_oe_l

t2
t3
t1

MK−1455−10

Preliminary—Subject to Change—July 1996 9–23

9.4 ac Characteristics

Table 9–14 SROM Load Timing for Some System Clock Ratios (System
Cycles)
Sysclk

System Cycles1

Ratio

4408090

4408216+½

4408217

3306099

3306193+2½ 3306194

881627

881651+9½

881652

1 Measured in sysclk cycles, where +n refers to an additional n CPU cycles.

Table 9–15 SROM Load Timing for Some System Clock Ratios (CPU Cycles)
Sysclk

CPU Cycles

Ratio

13224270

13224648½

13224651

192

13224396

13224774½

13224776

195

13224405

13224774½

13224780

Figure 9–9 is a timing diagram of an SROM load sequence.
Figure 9–9 Serial ROM Load Timing

sys_reset_l

srom_oe_l

srom_clk_h
t su

t ho

srom_data_h
t su = 4 x sysclk period + 1.1 ns
t ho = 0 ns

102,400 Bits Total
MK−1455−07

The minimum srom_clk_h cycle = (126 0 sysclk ratio) 3 (CPU cycle time).
The maximum srom_clk_h to srom_data_h delay allowable (in order to meet
the required setup time) = [126 0 (5 3 sysclk ratio)] 3 (CPU cycle time).

9–24 Preliminary—Subject to Change—July 1996

9.4 ac Characteristics

9.4.8 Clock Test Modes
This section describes the 21164 clock test modes.
9.4.8.1 Normal Mode
When the clk_mode_h<1:0> signals are not asserted, the osc_clk_in_h,l
frequency is divided by 2. This is the normal operational mode of the clock
circuitry.
9.4.8.2 Chip Test Mode
To lower the maximum frequency that the chip manufacturing tester is
required to supply, a divide-by-1 mode has been designed into the clock
generator circuitry. When the clk_mode_h<0> signal is asserted and
clk_mode_h<1> is not asserted, the clock frequency that is applied to the
input clock signals osc_clk_in_h,l bypasses the clock divider and is sent to
the chip clock driver. This allows the chip internal circuitry to be tested at full
speed with a one-half frequency osc_clk_in_h,l.
9.4.8.3 Module Test Mode
When the clk_mode_h<0> signal is not asserted and clk_mode_h<1>
is asserted, the clock frequency that is applied to the input clock signals
osc_clk_in_h,l is divided by 4 and is sent to the chip clock driver. The digital
phase-locked loop (DPLL) continues to keep the onchip sys_clk_out1_h,l
locked to ref_clk_in_h within the normal limits if a ref_clk_in_h signal is
applied (0 ns to 1 osc_clk_in_h,l cycle after ref_clk_in_h).
9.4.8.4 Clock Test Reset Mode
When both the clk_mode_h<0> and the clk_mode_h<1> signals are asserted,
the sys_clk_out generator circuit is forced to reset to a known state. This
allows the chip manufacturing tester to synchronize the chip to the tester
cycle. Table 9–16 lists the test modes.
Table 9–16 Test Modes
Mode

clk_mode_h<0>

clk_mode_h<1>

Normal

Chip test

Module test

Clock reset

Preliminary—Subject to Change—July 1996 9–25

9.4 ac Characteristics

9.4.9 IEEE 1149.1 (JTAG) Performance
Table 9–17 lists the standard mandated performance specifications for the
IEEE 1149.1 circuits.
Table 9–17 IEEE 1149.1 Circuit Performance Specifications
Item

Specification

trst_l is asynchronous. Minimum pulse width.

4 ns

trst_l setup time for deassertion before a transition on
tck_h.

4 ns

Maximum acceptable tck_h clock frequency.

16.6 MHz

tdi_h/tms_h setup time (referenced to tck_h rising edge).

4 ns

tdi_h/tms_h hold time (referenced to tck_h rising edge).

4 ns

Maximum propagation delay at pin tdo_h (referenced to
tck_h falling edge).

14 ns

Maximum propagation delay at system output pins
(referenced to tck_h falling edge).

20 ns

9.5 Power Supply Considerations
For correct operation of the 21164, all of the Vss pins must be connected to
ground and all of the Vdd pins must be connected to a 3.3 V ±5% power source.
This source voltage should be guaranteed (even under transient conditions) at
the 21164 pins, and not just at the PCB edge.
Plus 5 V is not used in the 21164. The voltage difference between the Vdd
pins and Vss pins must never be greater than 3.6 V. If the differential exceeds
this limit, the 21164 chip will be damaged.

9.5.1 Decoupling
The effectiveness of decoupling capacitors depends on the amount of inductance
placed in series with them. The inductance depends both on the capacitor style
(construction) and on the module design. In general, the use of small, high
frequency capacitors placed close to the chip package’s power and ground pins
with very short module etch will give best results. Depending on the user’s
power supply and power supply distribution system, bulk decoupling may also
be required on the module.

9–26 Preliminary—Subject to Change—July 1996

9.5 Power Supply Considerations

Each individual case must be separately analyzed, but generally designers
should plan to use at least 6 F of capacitance. Typically, 40 to 60 small,
high frequency 0.1-F capacitors are placed near the chip’s Vdd and Vss pins.
Actually placing the capacitors in the pin field is the best approach. Several
tens of F of bulk decoupling (comprised of tantalum and ceramic capacitors)
should be positioned near the 21164 chip.
Use capacitors that are as physically small as possible. Connect the capacitors
directly to the 21164 Vdd and Vss pins by short (0.64 cm [0.25 in] or
less) surface etch. The small capacitors generally have better electrical
characteristics than the larger units, and will more readily fit close to the
IPGA pin field.

9.5.2 Power Supply Sequencing
Although the 21164 uses a 3.3-V (nominal) power source, most of the other
logic on the PCB probably requires a 5-V power supply. These 5-V devices can
damage the 21164’s I/O circuits if the 5-V power source powering the PCB logic
and the 3.3-V (Vdd) supply feeding the 21164 are not sequenced correctly.
Caution
To avoid damaging the 21164’s I/O circuits, the I/O pin voltages must
not exceed 4 V until the Vdd supply is at least 3 V or greater.

This rule can be satisfied if the Vdd and the 5-V supplies come up together,
or if the Vdd supply comes up before the 5-V supply is asserted. Bringing
the lower voltage up before the higher voltage is the opposite of the way that
CMOS systems with multiple power supplies of different voltages are usually
sequenced, but it is required for the 21164.
A three-terminal voltage regulator can be used to make 3.3-V Vdd from the
5-V supply, provided the output of the regulator (Vdd) tracks the 5-V supply
with only a small offset. The requirement is that when the 5-V supply reaches
4.0 V, Vdd must be 3.0 V or higher. While the 5-V supply is below 4.0 V, Vdd
can be less than 3.0 V.
All 5-V sources on the 21164’s I/O pins should be disabled if the power supply
sequencing is such that the 5-V supply will exceed 4.0 V before Vdd is at least
3.0 V. The 5-V sources should remain disabled until the Vdd power supply is
equal to or greater than 3.0 V.

Preliminary—Subject to Change—July 1996 9–27

9.5 Power Supply Considerations

Disabling all 5-V sources can be very difficult because there are so many
possible sneak paths. Inputs, for example, on bipolar TTL logic can be a source
of current, and will put a voltage across a 21164 I/O pin high enough to violate
the (no higher than 4.0 V until there is 3.0 V) rule. TTL outputs are specified
to drive a logic one to at least 2.4 V, but usually drive voltages much higher.
CMOS logic and CMOS SRAMs usually drive ‘‘full rail’’ signals that match the
value of the 5-V power supply.
Another concern is parallel (dc) terminations or pull-ups connected between
the 21164 and the 5-V supply. The 3.3 V (Vdd) supply should be used to power
parallel terminations.
Disabling the non-21164 5-V outputs of PCB logic is generally possible, but
raises the PCB complexity and can reduce system performance by increasing
critical path timing. If the 5-V logic device has an enable pin, circuits (such
as power supply supervisor chips) on the PCB can monitor the Vdd and 5-V
supplies. When the supervision circuit detects that 5.0 V is increasing from
zero while the Vdd supply is below 3.0 V, the power supply supervisor circuit
produces a disable signal to force all PCB logic with 5-V outputs into the high
impedance state. This technique will not prevent bipolar TTL inputs from
acting as a 5-V source, but it can be used to disable sources such as cache RAM
outputs.

9–28 Preliminary—Subject to Change—July 1996

10
Thermal Management
This chapter describes the 21164 thermal management and thermal design
considerations.

10.1 Operating Temperature
The 21164 is specified to operate when the temperature at the center of the
heat sink (Tc ) is no higher than 72°C (266 MHz), 70°C (300 MHz), or 68°C
(333 MHz). Temperature (Tc ) should be measured at the center of the heat sink
(between the two package studs). The GRAFOIL pad is the interface material
between the package and the heat sink.
Table 10–1 lists the values for the center of heat-sink-to-ambient (c a) for the
499-pin grid array. Table 10–2 shows the allowable Ta (without exceeding Tc )
at various airflows.
Note
Digital recommends using the heat sink because it greatly improves
the ambient temperature requirement.

Preliminary—Subject to Change—July 1996 10–1

10.1 Operating Temperature

Table 10–1

c a at Various Airflows

Airflow (linear ft/min)

100

200

400

600

800

1000

2.30

1.30

0.70

0.53

0.45

0.41

1.25

0.75

0.48

0.40

0.35

0.32

400

600

800

1000

Frequency: 266, 300, and 333 MHz

c a with heat sink 1 (°C/W)
c a with heat sink 2 (°C/W)

Table 10–2 Maximum Ta at Various Airflows
Airflow (linear ft/min)

100

200

Frequency: 266 MHz, Power: 46 W @Vdd = 3.3 V
Ta with heat sink 1 (°C)

—

39.8

47.6

51.3

53.2

Ta with heat sink 2 (°C)

14.5

37.5

49.9

53.6

55.9

57.3

Frequency: 300 MHz, Power: 51 W @Vdd = 3.3 V
Ta with heat sink 1 (°C)

—

34.3

43.0

47.1

49.1

Ta with heat sink 2 (°C)

—

31.8

45.5

49.6

52.2

53.7

Frequency: 333 MHz, Power: 56 W @Vdd = 3.3 V
Ta with heat sink 1 (°C)

—

28.8

38.3

42.8

45.0

Ta with heat sink 2 (°C)

—

26.0

41.1

45.6

48.4

46.2

10–2 Preliminary—Subject to Change—July 1996

10.2 Heat Sink Specifications

10.2 Heat Sink Specifications
Two heat sinks are specified. Heat sink type 1 mounting holes are in line with
the cooling fins. Heat sink type 2 mounting holes are rotated 90° from the
cooling fins. The heat sink composition is aluminum alloy 6063. Type 1 heat
sink is shown in Figure 10–1, and type 2 heat sink is shown in Figure 10–2,
along with their approximate dimensions.
Figure 10–1 Type 1 Heat Sink
6.57 cm
(2.585 in)

2.54 cm
(1.0 in)

6.57 cm
(2.585 in)

3.25 cm
(1.280 in)

3.81 cm
(1.5 in)
sq.
LJ-04032.AI

Preliminary—Subject to Change—July 1996 10–3

10.3 Thermal Design Considerations

Figure 10–2 Type 2 Heat Sink
7.59 cm
(2.990 in)

3.80 cm
(1.495 in)

2.54 cm
(1.0 in)

4.45 cm
(1.75 in)

3.81 cm
(1.5 in)

LJ-04033.AI

10.3 Thermal Design Considerations
Follow these guidelines for printed circuit board (PCB) component placement:
•

Orient the 21164 on the PCB with the heat sink fins aligned with the
airflow direction.

•

Avoid preheating ambient air. Place the 21164 on the PCB so that inlet air
is not preheated by any other PCB components.

•

Do not place other high power devices in the vicinity of the 21164.

•

Do not restrict the airflow across the 21164 heat sink. Placement of other
devices must allow for maximum system airflow in order to maximize the
performance of the heat sink.

10–4 Preliminary—Subject to Change—July 1996

11
Mechanical Data and Packaging
Information
This chapter describes the 21164 mechanical packaging including chip package
physical specifications and a signal/pin list. For heat sink dimensions, refer to
Chapter 10.

11.1 Mechanical Specifications
Figure 11–1 shows the package physical dimensions without a heat sink.

Preliminary—Subject to Change—July 1996 11–1

11.1 Mechanical Specifications

Figure 11–1 Package Dimensions
1.27 mm (0.050 in) Typ
4.32 mm (0.170 in) Typ
2.54 mm (0.100 in) Typ

Standoff (4x)

BC
BB
BA
AY
AW
AV
AU
AT
AR
AP
AN
AM
AL
AK
AJ
AH
AG
AF
AE
AD
AC
AB
AA
Y
W
V
U
T
R
P
N
M
L
K
J
H
G
F
E
D
C
B
A

1.27 mm
(0.050 in) Typ

499x 1.40 mm (0.055 in) Typ
1.27 mm (0.050 in) Typ

26.67 mm
(1.050 in)

Lid

1/4-20
Stud (2x)

0.46 mm
(0.018 in) Typ

7.62 mm
(0.300 in) Typ

0.13 mm
(0.005 in) R

02 04 06 08 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 42
01 03 05 07 09 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43

2.69 mm (0.106 in) Typ

26.67 mm
(1.050 in)
57.40 mm (2.260 in) Typ
28.70 mm
(1.130 in) Typ

28.70 mm
(1.130 in) Typ

Capacitors (12x)

25.40 mm
(1.000 in) Typ
38.10 mm
(1.500 in) Typ

11–2 Preliminary—Subject to Change—July 1996

LJ-03457-TI0

11.2 Signal Descriptions and Pin Assignment

11.2 Signal Descriptions and Pin Assignment
This section provides detailed information about the 21164 pinout. The 21164
has 499 pins aligned in an interstitial pin grid array (IPGA) design.

11.2.1 Signal Pin Lists
Table 11–1 lists the 21164 signal pins and their corresponding pin grid array
(PGA) locations in alphabetic order. There are 292 functional signal pins, 2
spare (unused) signal pins, 104 power (Vdd) pins, and 101 ground (Vss) pins,
for a total of 499 pins in the array.
Table 11–1 Alphabetic Signal Pin List
Signal

PGA
Location

Signal

PGA
Location

Signal

PGA
Location

addr_bus_req_h

E23

addr_cmd_par_h

B20

addr_h<4>

BB14

addr_h<5>

BC13

addr_h<6>

BA13

addr_h<7>

AV14

addr_h<8>

AW13

addr_h<9>

BC11

addr_h<10>

BA11

addr_h<11>

AV12

addr_h<12>

AW11

addr_h<13>

BC09

addr_h<14>

BA09

addr_h<15>

AV10

addr_h<16>

AW09

addr_h<17>

BC07

addr_h<18>

BA07

addr_h<19>

AV08

addr_h<20>

AW07

addr_h<21>

BC05

addr_h<22>

BC39

addr_h<23>

AW37

addr_h<24>

AV36

addr_h<25>

BA37

addr_h<26>

BC37

addr_h<27>

AW35

addr_h<28>

AV34

addr_h<29>

BA35

addr_h<30>

BC35

addr_h<31>

AW33

addr_h<32>

AV32

addr_h<33>

BA33

addr_h<34>

BC33

addr_h<35>

AW31

addr_h<36>

AV30

addr_h<37>

BA31

addr_h<38>

BC31

addr_h<39>

BB30

addr_res_h<0>

C27

addr_res_h<1>

F26

addr_res_h<2>

E27

cack_h

G21

cfail_h

C25

clk_mode_h<0>

AU21

clk_mode_h<1>

BA23

cmd_h<0>

F20

cmd_h<1>

A19

cmd_h<2>

C19

cmd_h<3>

E19

cpu_clk_out_h

BA25

dack_h

B24

data_bus_req_h

E25

data_check_h<0>

J41

data_check_h<1>

K38

data_check_h<2>

J39

data_check_h<3>

G43

data_check_h<4>

G41

(continued on next page)

Preliminary—Subject to Change—July 1996 11–3

11.2 Signal Descriptions and Pin Assignment

Table 11–1 (Cont.) Alphabetic Signal Pin List
Signal

PGA
Location

Signal

PGA
Location

Signal

PGA
Location

data_check_h<5>

H38

data_check_h<6>

G39

data_check_h<7>

E43

data_check_h<8>

J03

data_check_h<9>

K06

data_check_h<10>

J05

data_check_h<11>

G01

data_check_h<12>

G03

data_check_h<13>

H06

data_check_h<14>

G05

data_check_h<15>

E01

data_h<0>

J43

data_h<1>

L39

data_h<2>

M38

data_h<3>

L41

data_h<4>

L43

data_h<5>

N39

data_h<6>

P38

data_h<7>

N41

data_h<8>

N43

data_h<9>

P42

data_h<10>

R39

data_h<11>

T38

data_h<12>

R41

data_h<13>

R43

data_h<14>

U39

data_h<15>

V38

data_h<16>

U41

data_h<17>

U43

data_h<18>

W39

data_h<19>

W41

data_h<20>

W43

data_h<21>

Y38

data_h<22>

Y42

data_h<23>

AA39

data_h<24>

AA41

data_h<25>

AA43

data_h<26>

AB38

data_h<27>

AC43

data_h<28>

AC41

data_h<29>

AC39

data_h<30>

AD42

data_h<31>

AD38

data_h<32>

AE43

data_h<33>

AE41

data_h<34>

AE39

data_h<35>

AG43

data_h<36>

AG41

data_h<37>

AF38

data_h<38>

AG39

data_h<39>

AJ43

data_h<40>

AJ41

data_h<41>

AH38

data_h<42>

AJ39

data_h<43>

AK42

data_h<44>

AL43

data_h<45>

AL41

data_h<46>

AK38

data_h<47>

AL39

data_h<48>

AN43

data_h<49>

AN41

data_h<50>

AM38

data_h<51>

AN39

data_h<52>

AR43

data_h<53>

AR41

data_h<54>

AP38

data_h<55>

AR39

data_h<56>

AU43

data_h<57>

AU41

data_h<58>

AT38

data_h<59>

AU39

data_h<60>

AW43

data_h<61>

AW41

data_h<62>

AV38

data_h<63>

AW39

data_h<64>

J01

data_h<65>

L05

data_h<66>

M06

data_h<67>

L03

data_h<68>

L01

data_h<69>

N05

data_h<70>

P06

data_h<71>

N03

data_h<72>

N01

(continued on next page)

11–4 Preliminary—Subject to Change—July 1996

11.2 Signal Descriptions and Pin Assignment

Table 11–1 (Cont.) Alphabetic Signal Pin List
Signal

PGA
Location

Signal

PGA
Location

Signal

PGA
Location

data_h<73>

P02

data_h<74>

R05

data_h<75>

T06

data_h<76>

R03

data_h<77>

R01

data_h<78>

U05

data_h<79>

V06

data_h<80>

U03

data_h<81>

U01

data_h<82>

W05

data_h<83>

W03

data_h<84>

W01

data_h<85>

Y06

data_h<86>

Y02

data_h<87>

AA05

data_h<88>

AA03

data_h<89>

AA01

data_h<90>

AB06

data_h<91>

AC01

data_h<92>

AC03

data_h<93>

AC05

data_h<94>

AD02

data_h<95>

AD06

data_h<96>

AE01

data_h<97>

AE03

data_h<98>

AE05

data_h<99>

AG01

data_h<100>

AG03

data_h<101>

AF06

data_h<102>

AG05

data_h<103>

AJ01

data_h<104>

AJ03

data_h<105>

AH06

data_h<106>

AJ05

data_h<107>

AK02

data_h<108>

AL01

data_h<109>

AL03

data_h<110>

AK06

data_h<111>

AL05

data_h<112>

AN01

data_h<113>

AN03

data_h<114>

AM06

data_h<115>

AN05

data_h<116>

AR01

data_h<117>

AR03

data_h<118>

AP06

data_h<119>

AR05

data_h<120>

AU01

data_h<121>

AU03

data_h<122>

AT06

data_h<123>

AU05

data_h<124>

AW01

data_h<125>

AW03

data_h<126>

AV06

data_h<127>

AW05

data_ram_oe_h

F22

data_ram_we_h

A23

dc_ok_h

AU23

fill_error_h

A25

fill_h

G23

fill_id_h

F24

fill_nocheck_h

G25

idle_bc_h

A27

index_h<4>

A29

index_h<5>

C29

index_h<6>

F28

index_h<7>

E29

index_h<8>

B30

index_h<9>

A31

index_h<10>

C31

index_h<11>

F30

index_h<12>

E31

index_h<13>

A33

index_h<14>

C33

index_h<15>

F32

index_h<16>

E33

index_h<17>

A35

index_h<18>

C35

index_h<19>

F34

index_h<20>

E35

index_h<21>

A37

index_h<22>

C37

index_h<23>

F36

index_h<24>

E37

(continued on next page)

Preliminary—Subject to Change—July 1996 11–5

11.2 Signal Descriptions and Pin Assignment

Table 11–1 (Cont.) Alphabetic Signal Pin List
Signal

PGA
Location

Signal

PGA
Location

Signal

PGA
Location

index_h<25>

A39

int4_valid_h<0>

F38

int4_valid_h<1>

E41

int4_valid_h<2>

F06

int4_valid_h<3>

E03

irq_h<0>

BA29

irq_h<1>

AU27

irq_h<2>

BC29

irq_h<3>

AW27

mch_hlt_irq_h

AU25

osc_clk_in_h

BC21

osc_clk_in_l

BB22

perf_mon_h

AW29

port_mode_h<0>

AY20

port_mode_h<1>

BB20

pwr_fail_irq_h

AV26

ref_clk_in_h

AW25

scache_set_h<0>

C17

scache_set_h<1>

A17

shared_h

C23

srom_clk_h

BA19

srom_data_h

BC19

srom_oe_l

AW19

srom_present_l

AV20

st_clk_h

E05

system_lock_flag_h

G27

sys_clk_out1_h

AW23

sys_clk_out1_l

BB24

sys_clk_out2_h

AV24

sys_clk_out2_l

BC25

sys_mch_chk_irq_h

BA27

sys_reset_l

BC27

tag_ctl_par_h

F18

tag_data_h<20>

A05

tag_data_h<21>

E07

tag_data_h<22>

F08

tag_data_h<23>

C07

tag_data_h<24>

A07

tag_data_h<25>

E09

tag_data_h<26>

F10

tag_data_h<27>

C09

tag_data_h<28>

A09

tag_data_h<29>

E11

tag_data_h<30>

F12

tag_data_h<31>

C11

tag_data_h<32>

A11

tag_data_h<33>

E13

tag_data_h<34>

F14

tag_data_h<35>

C13

tag_data_h<36>

A13

tag_data_h<37>

B14

tag_data_h<38>

E15

tag_data_par_h

C15

tag_dirty_h

E17

tag_ram_oe_h

C21

tag_ram_we_h

A21

tag_shared_h

A15

tag_valid_h

F16

tck_h

AW17

tdi_h

BC17

tdo_h

BA17

temp_sense

AW15

test_status_h<0>

BA15

test_status_h<1>

AV16

tms_h

AV18

trst_l

BC15

victim_pending_h

E21

spare_in<438>

E39

spare_io<250>

AV28

(continued on next page)

11–6 Preliminary—Subject to Change—July 1996

11.2 Signal Descriptions and Pin Assignment

Table 11–1 (Cont.) Alphabetic Signal Pin List
Signal

PGA Location

Vss
Metal planes 21 and
52

A03, A41, AA07, AA37, AC07, AC37, AD04, AD40, AF02, AF42, AG07, AG37, AH04,
AH40, AL07, AL37, AM04, AM40, AP02, AP42, AR07, AR37, AT04, AT40, AU09,
AU13, AU17, AU31, AU35, AV02, AV22, AV42, AW21, AY04, AY08, AY12, AY16,
AY22, AY24, AY28, AY32, AY36, AY40, B02, B06, B10, B18, B26, B34, B38, B42,
BA01, BA21, BA43, BB02, BB06, BB10, BB18, BB26, BB34, BB38, BB42, BC03,
BC41, C01, C43, D04, D08, D12, D16, D20, D24, D28, D32, D36, D40, F02, F42, G09,
G13, G17, G31, G35, H04, H40, J07, J37, K02, K42, M04, M40, N07, N37, T04, T40,
U07, U37, V02, V42, Y04, Y40

Vdd
Metal planes 4 and 6

AB02, AB04, AB40, AB42, AE07, AE37, AF04, AF40, AH02, AH42, AJ07, AJ37,
AK04, AK40, AM02, AM42, AN07, AN37, AP04, AP40, AT02, AT42, AU07, AU11,
AU15, AU19, AU29, AU33, AU37, AV04, AV40, AY02, AY06, AY10, AY14, AY18,
AY26, AY30, AY34, AY38, AY42, B04, B08, B12, B16, B22, B28, B32, B36, B40, BA03,
BA05, BA39, BA41, BB04, BB08, BB12, BB16, BB28, BB32, BB36, BB40, BC23, C03,
C05, C39, C41, D02, D06, D10, D14, D18, D22, D26, D30, D34, D38, D42, F04, F40,
G11, G15, G19, G29, G33, G37, H02, H42, K04, K40, L07, L37, M02, M42, P04, P40,
R07, R37, T02, T42, V04, V40, W07, W37

1 Metal plane 2—Seal ring connection tied to Vss
2 Metal plane 5—Heat slug braze pad connections tied to Vss

Preliminary—Subject to Change—July 1996 11–7

11.2 Signal Descriptions and Pin Assignment

11.2.2 Pin Assignment
Figure 11–2 shows the 21164 pinout from the top view with pins facing down.
Figure 11–2
BC
BA
AW
AU
AR
AN
AL
AJ
AG
AE
AC
AA
W
U
R
N
L
J
G
E
C
A

Alpha 21164 Top View (Pin Down)

BB
AY
AV
AT
AP
AM
AK
AH
AF
AD
AB
Y

21164
Top View
(Pin Down)

V
T
P
M
K
H
F
D
B

42 40 38 36 34 32 30 28 26 24 22 20 18 16 14 12 10 08 06 04 02
43 41 39 37 35 33 31 29 27 25 23 21 19 17 15 13 11 09 07 05 03 01
LJ-03453-TI0A

11–8 Preliminary—Subject to Change—July 1996

11.2 Signal Descriptions and Pin Assignment

Figure 11–3 shows the 21164 pinout from the bottom view with pins facing
up.
Figure 11–3
BC
BA
AW
AU
AR
AN
AL
AJ
AG
AE
AC
AA
W
U
R
N
L
J
G
E
C
A

Alpha 21164 Bottom View (Pin Up)

BB
AY
AV
AT
AP
AM
AK
AH
AF
AD
AB
Y

21164
Bottom View
(Pin Up)

V
T
P
M
K
H
F
D
B

02 04 06 08 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 42
01 03 05 07 09 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43
LJ-03413-TI0B

Preliminary—Subject to Change—July 1996 11–9

12
Testability and Diagnostics
This chapter describes the 21164 user-oriented testability features. The 21164
also has several internal testability features that are implemented for factory
use only. These features are beyond the scope of this document.

12.1 Test Port Pins
Table 12–1 summarizes the test port pins and their function.
Table 12–1

Alpha 21164 Test Port Pins

Pin Name

Type

Function

port_mode_h<1>

Must be false.

port_mode_h<0>

Must be false.

srom_present_l

Tied low if serial ROMs (SROMs) are present in
system.

srom_data_h/Rx

Receives SROM or serial terminal data.

srom_clk_h/Tx

Supplies clock to SROMs or transmits serial
terminal data.

srom_oe_l

SROM enable.

tdi_h

IEEE 1149.1 TDI port.

tdo_h

IEEE 1149.1 TDO port.

tms_h

IEEE 1149.1 TMS port.

tck_h

IEEE 1149.1 TCK port.

trst_l

IEEE 1149.1 optional TRST port.

test_status_h<0>

Indicates Icache BiSt status.

test_status_h<1>

Outputs an IPR-written value and timeout reset.

Preliminary—Subject to Change—July 1996 12–1

12.2 Test Interface

12.2 Test Interface
The 21164 test interface supports a serial ROM interface, a serial diagnostic
terminal interface, and an IEEE 1149.1 test access port. These ports are
available and set to normal test interface mode when port_mode_h<1:0>=00.
Driving these pins to a value of anything other than 00 redefines all other
test interface pins and invokes special factory test modes not covered in this
document.
The SROM port is described in Section 7.4 and the serial terminal port is
described in Section 7.5.

12.2.1 IEEE 1149.1 Test Access Port
Pins tdi_h, tdo_h, tck_h, tms_h, and trst_l constitute the IEEE 1149.1 test
access port. This port accesses the 21164 chip’s boundary scan register and
chip tristate functions for board level manufacturing test. The port also allows
access to factory manufacturing features not described in this document. The
port is compliant with most requirements of IEEE 1149.1 test access port.
Compliance Enable Inputs
Table 12–2 shows the compliance enable inputs and the pattern that must be
driven to those inputs in order to activate the 21164 IEEE 1149.1 circuits.
Table 12–2 Compliance Enable Inputs
Input

Compliance Enable Pattern

port_mode_h<1:0>

dc_ok_h

Exceptions to Compliance
The 21164 is compliant with IEEE Standard 1149.1–1993, with two exceptions.
Both exceptions provide enhanced value to the user.
1. trst_l pin
The optional trst_l pin has an internal pull-down, instead of a pull-up as
required by IEEE 1149.1 (non-complied spec 3.6.1(b) in IEEE 1149.1–1993).
The trst_l pull-down allows the chip to automatically force reset to
the IEEE 1149.1 circuits in a system in which the IEEE 1149.1 port is
unconnected. This may be considered a feature for most system designs
that use IEEE 1149.1 circuits solely during module manufacturing.

12–2 Preliminary—Subject to Change—July 1996

12.2 Test Interface

Note
Digital recommends that the trst_l pin be driven low (asserted) when
the JTAG (IEEE 1149.1) logic is not in use.

2. Coverage of oscillator differential input pins
The two differential clock input pins, osc_clk_in_h and osc_clk_in_l,
do not have any boundary scan cells associated with them (non-complied
spec 10.4.1(b) in IEEE 1149.1–1993). Instead, there is an extra input BSR
cell in the boundary scan register in bit position 255 (at pin dc_ok_h).
This cell captures the output of a ‘‘clock sniffer’’ circuit. It captures a ‘‘1’’
when the oscillator is connected, and captures a ‘‘0’’ if the chip’s oscillator
connections are broken.
This exception to the standard is made to permit a meaningful test of the
oscillator input pins.
Refer to IEEE Standard 1149.1-1993 A Test Access Port and Boundary Scan
Architecture for a full description of the specification.
Figure 12–1 shows the user-visible features from this port.

Preliminary—Subject to Change—July 1996 12–3

12.2 Test Interface

Figure 12–1 IEEE 1149.1 Test Access Port

TRST_L

TMS_H

TAP Controller
State Machine &
Control Dispatch
Logic

CONTROL

TCK_H
TDO_H

TDI_H

Instruction Register (IR)
Bypass Register (BPR)
Die-ID Register (IDR)
Boundary Scan Register (BSR)

LJ-03463-TI0

TAP Controller
The TAP controller contains a state machine. It interprets IEEE 1149.1
protocols received on signal tms_h and generates appropriate clocks and
control signals for the testability features under its jurisdiction. The state
machine is shown in Figure 12–2

12–4 Preliminary—Subject to Change—July 1996

12.2 Test Interface

Figure 12–2 TAP Controller State Machine

Test Logic
Reset
0
Run−Test/Idle

Select−DR−Scan

Select−IR−Scan 1

0
1

Capture−DR

Capture−IR
0

Shift−IR

Shift−DR

0
1

1
1

Exit1−DR

Values
shown
are for
TMS.

Pause−DR

Pause−IR
0

1
0

Exit2−DR

Exit2−IR

Update−DR
1

Exit1−IR

Update−IR

Scan Sequence

Scan Sequence
MK−1455−08

Instruction Register
The 5-bit-wide instruction register (IR) supports IEEE 1149.1 mandated
public instructions (EXTEST, SAMPLE, BYPASS, HIGHZ) and a number of optional
instructions for public and private factory use. Table 12–3 summarizes the
public instructions and their functions.
During the capture operation, the shift register stage of IR is loaded with the
value 00001. This automatic load feature is useful for testing the integrity of
the IEEE 1149.1 scan chain on the module.

Preliminary—Subject to Change—July 1996 12–5

12.2 Test Interface

Table 12–3 Instruction Register
IR<4:0>

Name

Scan Register
Selected

00000

EXTEST

BSR

BSR drives pins. Interconnect test
mode.

00010

SAMPLE/
PRELOAD

BSR

Preloads BSR.

00010

Private

BSR

Private.

00011

Private

BSR

Private.

00100

CLAMP

BPR

BSR drives pins.

00101

HIGHZ

BPR

Tristate all output and I/O pins.

00110

Private

IDR

Private.

00111

Private

IDR

Private.

01000
through
11110

Private

BPR

Private.

11111

BYPASS

BPR

Default.

Operation

Bypass Register
The bypass register is a 1-bit shift register. It provides a short single-bit scan
path through the port (chip).
Boundary Scan Register
The 289-bit boundary scan register is accessed during SAMPLE, EXTEST, and
CLAMP instructions. Refer to Section 12.3 for the organization of this register.

12.2.2 Test Status Pins
Two test status signal test_status_h<1:0> pins are used for extracting test
status information from the chip. System reset drives both test status pins low.
The default operation for test_status_h<0> is to output the BiSt results. The
default operation for test_status_h<1> is to output the IPR-written value.
•

During Icache BiSt Operation
test_status_h<0> is forced high at the start of the Icache BiSt. If the
Icache BiSt passes, the pin is deasserted at the end of the BiSt operation,
otherwise it remains high.

•

IPR read and write operations to test status pins

12–6 Preliminary—Subject to Change—July 1996

12.2 Test Interface

PALcode can write to the test_status_h<1> signal pin and can read the
test_status_h<0> signal pin through hardware IPR access. Refer to
Chapter 6.
•

Timeout Reset
The 21164 generates a timeout reset signal under two conditions:
1. If an instruction is not retired within 1 billion cycles.
2. If the system asserts cfail_h when cack_h is deasserted.
In either of these conditions, the CPU signals the timeout reset event by
outputting a 256 CPU cycle wide pulse on the test_status_h<1> pin. The
pulse on test_status_h<1> pin is clocked by sysclk and therefore appears
as an approximately 256 CPU cycle pulse that rises and falls on system
clock rising edges.

12.3 Boundary Scan Register
The 21164 boundary scan register (BSR) is 289 bits long. Table 12–4 provides
the boundary scan register organization. The BSR is connected between the
tdi_h and tdo_h pins whenever an instruction selects it (Table 12–3). The
scan register runs clockwise beginning at the upper left corner of the chip.
There are seven groups of bidirectional pins, each group controlled from a
group control cell. Loading a value of ‘‘1’’ in the control cell tristates the output
drivers and all bidirectional pins in the group are configured as input pins.
The bidirectional pin groups are identified as groups gr_1 through gr_7 in the
Control Group column in Table 12–4.
Information on Boundary Scan Description Language (BSDL) as it applies
to the 21164 boundary scan register is available through your local Digital
distributor (see Appendix E).
Notes
The following notes apply to Table 12–4:
•

The direction of shift is from top to bottom, and from left to right.

•

The bottom most signals appear first at the tdo_h pin when
shifting.

•

Given an arrayed signal of the form signal<a:b>, signal<b> appears
at the tdo_h pin prior to signal<a>.

Preliminary—Subject to Change—July 1996 12–7

12.3 Boundary Scan Register

Table 12–4 Boundary Scan Register Organization
Signal Name

Pin
Type

BSR
Count

BSR
Cell Type

Control
Group

Remarks

TR_ADL

Control

288

io_bcell

gr_1

Upper left corner.

addr_h<21:4>

287:270

io_bcell

gr_1

—

temp_sense

—

None

—

Analog pin.

test_status_h<1:0>

269:268

io_bcell

—

trst_l

—

None

—

tck_h

—

None

—

tms_h

—

None

—

tdo_h

—

None

—

tdi_h

—

None

—

srom_oe_l

267

io_bcell

—

srom_clk_h

266

io_bcell

—

srom_data_h

265

in_bcell

—

srom_present_l

264

in_bcell

—

port_mode_h<0:1>

—

None

clk_mode_h<0>

263

in_bcell

—

osc_clk_in_h,l

—

None

—

Analog pins.

clk_mode_h<1>

262

in_bcell

—

sys_clk_out1_h,l

261:260

io_bcell

—

sys_clk_out2_h,l

259:258

io_bcell

—

cpu_clk_out_h

—

none

—

For chip test.

ref_clk_in_h

257

in_bcell

—

sys_reset_l

256

in_bcell

—

dc_ok_h

—

None

—

Compliance enable pin.

OSC_SNIFFER_H

Internal

255

in_bcell

—

Captures 1 if osc is
connected, otherwise
captures 0.

sys_mch_chk_irq_h

254

in_bcell

—

pwr_fail_irq_h

253

in_bcell

—

mch_hlt_irq_h

252

in_bcell

—

Compliance enable pins.

(continued on next page)

12–8 Preliminary—Subject to Change—July 1996

12.3 Boundary Scan Register

Table 12–4 (Cont.) Boundary Scan Register Organization
Signal Name

Pin
Type

BSR
Count

BSR
Cell Type

Control
Group

Remarks

irq_h<3:0>

251:248

in_bcell

—

SPARE_IO<250>

247

io_bcell

—

Tied off as input.

perf_mon_h

246

in_bcell

—

TR_ADR

Control

245

io_bcell

gr_2

—

addr_h<39:22>

244:227

io_bcell

gr_2

Upper right corner.

TR_DDR

Control

226

io_bcell

gr_3

—

data_h<63:0>

225:162

io_bcell

gr_3

—

data_check_h<0:7>

161:154

io_bcell

gr_3

—

int4_valid_h<1:0>

153:152

io_bcell

—

SPARE_IO<438>

—

None

—

Lower right corner,
unpopulated.

index_h<25:4>

151:130

io_bcell

—

addr_res_h<2:0>

129:127

io_bcell

—

idle_bc_h

126

in_bcell

—

system_lock_flag_h

125

in_bcell

—

data_bus_req_h

124

in_bcell

—

cfail_h

123

in_bcell

—

fill_nocheck_h

122

in_bcell

—

fill_error_h

121

in_bcell

—

fill_id_h

120

in_bcell

—

fill_h

119

in_bcell

—

dack_h

118

in_bcell

—

addr_bus_req_h

117

in_bcell

—

cack_h

116

in_bcell

—

shared_h

115

in_bcell

—

data_ram_we_h

114

io_bcell

—

data_ram_oe_h

113

io_bcell

—

tag_ram_we_h

112

io_bcell

—

(continued on next page)

Preliminary—Subject to Change—July 1996 12–9

12.3 Boundary Scan Register

Table 12–4 (Cont.) Boundary Scan Register Organization
Signal Name

Pin
Type

BSR
Count

BSR
Cell Type

Control
Group

Remarks

tag_ram_oe_h

111

io_bcell

—

victim_pending_h

110

io_bcell

—

TMIS1

Control

109

io_bcell

gr_4

—

addr_cmd_par_h

108

io_bcell

gr_4

—

cmd_h<0:3>

107:104

io_bcell

gr_4

—

scache_set_h<1:0>

103:102

io_bcell

—

TTAG1

Control

101

io_bcell

gr_5

—

tag_ctl_par_h

100

io_bcell

gr_5

—

tag_dirty_h

io_bcell

gr_5

—

tag_shared_h

io_bcell

gr_5

—

TTAG2

Control

io_bcell

gr_6

—

tag_data_par_h

io_bcell

gr_6

—

tag_valid_h

io_bcell

gr_6

—

tag_data_h<38:20>

94:76

io_bcell

gr_6

—

st_clk_h

io_bcell

—

Lower left corner.

int4_valid_h<2:3>

74:73

io_bcell

—

TR_DDL

Control

io_bcell

gr_7

—

data_check_h<15:8>

71:64

io_bcell

gr_7

—

data_h<64:127>

63:0

io_bcell

gr_7

—

12–10 Preliminary—Subject to Change—July 1996

A
Alpha Instruction Set
A.1 Alpha Instruction Summary
This appendix contains a summary of all Alpha architecture instructions.
All values are in hexadecimal radix. Table A–1 describes the contents of the
Format and Opcode columns that are in Table A–2.
Table A–1 Instruction Format and Opcode Notation
Instruction
Format

Format
Symbol

Opcode
Notation

Meaning

Branch

Bra

oo is the 6-bit opcode field.

Floatingpoint

F-P

oo.fff

oo is the 6-bit opcode field.
fff is the 11-bit function code field.

Memory

Mem

oo is the 6-bit opcode field.

Memory/
function code

Mfc

oo.ffff

oo is the 6-bit opcode field.
ffff is the 16-bit function code in the
displacement field.

Memory/
branch

Mbr

oo.h

oo is the 6-bit opcode field.
h is the high-order 2 bits of the
displacement field.

Operate

Opr

oo.ff

oo is the 6-bit opcode field.
ff is the 7-bit function code field.

PALcode

Pcd

oo is the 6-bit opcode field; the
particular PALcode instruction is
specified in the 26-bit function code
field.

Preliminary—Subject to Change—July 1996 A–1

A.1 Alpha Instruction Summary

Qualifiers for operate instructions are shown in Table A–2. Qualifiers for
IEEE and VAX floating-point instructions are shown in Tables A–5 and A–6,
respectively.
Table A–2 Architecture Instructions
Mnemonic

Format

Opcode

Description

ADDF
ADDG
ADDL
ADDL/V
ADDQ
ADDQ/V
ADDS
ADDT
AND
BEQ
BGE
BGT
BIC
BIS
BLBC
BLBS
BLE
BLT
BNE
BR
BSR
CALL_PAL
CMOVEQ
CMOVGE
CMOVGT
CMOVLBC
CMOVLBS
CMOVLE
CMOVLT
CMOVNE
CMPBGE
CMPEQ
CMPGEQ
CMPGLE

F-P
F-P
Opr
Opr
Opr
Opr
F-P
F-P
Opr
Bra
Bra
Bra
Opr
Opr
Bra
Bra
Bra
Bra
Bra
Bra
Mbr
Pcd
Opr
Opr
Opr
Opr
Opr
Opr
Opr
Opr
Opr
Opr
F-P
F-P

15.080
15.0A0
10.00
10.40
10.20
10.60
16.080
16.0A0
11.00
39
3E
3F
11.0
11.20
38
3C
3B
3A
3D
30
34
00
11.24
11.46
11.66
11.16
11.14
11.64
11.44
11.26
10.0F
10.2D
15.0A5
15.0A7

Add F_floating
Add G_floating
Add longword
Add longword
Add quadword
Add quadword
Add S_floating
Add T_floating
Logical product
Branch if = zero
Branch if  zero
Branch if > zero
Bit clear
Logical sum
Branch if low bit clear
Branch if low bit set
Branch if  zero
Branch if < zero
Branch if 6= zero
Unconditional branch
Branch to subroutine
Trap to PALcode
CMOVE if = zero
CMOVE if  zero
CMOVE if > zero
CMOVE if low bit clear
CMOVE if low bit set
CMOVE if  zero
CMOVE if < zero
CMOVE if 6= zero
Compare byte
Compare signed quadword equal
Compare G_floating equal
Compare G_floating less than or
equal
(continued on next page)

A–2 Preliminary—Subject to Change—July 1996

A.1 Alpha Instruction Summary

Table A–2 (Cont.) Architecture Instructions
Mnemonic

Format

Opcode

Description

CMPGLT
CMPLE

F-P
Opr

15.0A6
10.6D

CMPLT

Opr

10.4D

CMPTEQ
CMPTLE

F-P
F-P

16.0A5
16.0A7

CMPTLT
CMPTUN
CMPULE

F-P
F-P
Opr

16.0A6
16.0A4
10.3D

CMPULT

Opr

10.1D

CPYS
CPYSE
CPYSN
CVTDG
CVTGD
CVTGF
CVTGQ
CVTLQ
CVTQF
CVTQG
CVTQL
CVTQL/SV
CVTQL/V
CVTQS
CVTQT
CVTST
CVTTQ
CVTTS
DIVF
DIVG
DIVS
DIVT
EQV
EXCB
EXTBL
EXTLH

F-P
F-P
F-P
F-P
F-P
F-P
F-P
F-P
F-P
F-P
F-P
F-P
F-P
F-P
F-P
F-P
F-P
F-P
F-P
F-P
F-P
F-P
Opr
Mfc
Opr
Opr

17.020
17.022
17.021
15.09E
15.0AD
15.0AC
15.0AF
17.010
15.0BC
15.0BE
17.030
17.530
17.130
16.0BC
16.0BE
16.2AC
16.0AF
16.0AC
15.083
15.0A3
16.083
16.0A3
11.48
18.0400
12.06
12.6A

Compare G_floating less than
Compare signed quadword less
than or equal
Compare signed quadword less
than
Compare T_floating equal
Compare T_floating less than or
equal
Compare T_floating less than
Compare T_floating unordered
Compare unsigned quadword
less than or equal
Compare unsigned quadword
less than
Copy sign
Copy sign and exponent
Copy sign negate
Convert D_floating to G_floating
Convert G_floating to D_floating
Convert G_floating to F_floating
Convert G_floating to quadword
Convert longword to quadword
Convert quadword to F_floating
Convert quadword to G_floating
Convert quadword to longword
Convert quadword to longword
Convert quadword to longword
Convert quadword to S_floating
Convert quadword to T_floating
Convert S_floating to T_floating
Convert T_floating to quadword
Convert T_floating to S_floating
Divide F_floating
Divide G_floating
Divide S_floating
Divide T_floating
Logical equivalence
Exception barrier
Extract byte low
Extract longword high
(continued on next page)

Preliminary—Subject to Change—July 1996 A–3

A.1 Alpha Instruction Summary

Table A–2 (Cont.) Architecture Instructions
Mnemonic

Format

Opcode

Description

EXTLL
EXTQH
EXTQL
EXTWH
EXTWL
FBEQ
FBGE
FBGT
FBLE
FBLT
FBNE
FCMOVEQ
FCMOVGE
FCMOVGT
FCMOVLE
FCMOVLT
FCMOVNE
FETCH
FETCH_M
INSBL
INSLH
INSLL
INSQH
INSQL
INSWH
INSWL
JMP
JSR
JSR_COROUTINE
LDA
LDAH
LDF
LDG
LDL
LDL_L

Opr
Opr
Opr
Opr
Opr
Bra
Bra
Bra
Bra
Bra
Bra
F-P
F-P
F-P
F-P
F-P
F-P
Mfc
Mfc
Opr
Opr
Opr
Opr
Opr
Opr
Opr
Mbr
Mbr
Mbr
Mem
Mem
Mem
Mem
Mem
Mem

12.26
12.7A
12.36
12.5A
12.16
31
36
37
33
32
35
17.02A
17.02D
17.02F
17.02E
17.02C
17.02B
18.80
18.A0
12.0B
12.67
12.2B
12.77
12.3B
12.57
12.1B
1A.0
1A.1
1A.3
08
09
20
21
28
2A

LDQ
LDQ_L
LDQ_U

Mem
Mem
Mem

29
2B
0B

Extract longword low
Extract quadword high
Extract quadword low
Extract word high
Extract word low
Floating branch if = zero
Floating branch if  zero
Floating branch if > zero
Floating branch if  zero
Floating branch if < zero
Floating branch if 6= zero
FCMOVE if = zero
FCMOVE if  zero
FCMOVE if > zero
FCMOVE if  zero
FCMOVE if < zero
FCMOVE if 6= zero
Prefetch data
Prefetch data, modify intent
Insert byte low
Insert longword high
Insert longword low
Insert quadword high
Insert quadword low
Insert word high
Insert word low
Jump
Jump to subroutine
Jump to subroutine return
Load address
Load address high
Load F_floating
Load G_floating
Load sign-extended longword
Load sign-extended longword
locked
Load quadword
Load quadword locked
Load unaligned quadword
(continued on next page)

A–4 Preliminary—Subject to Change—July 1996

A.1 Alpha Instruction Summary

Table A–2 (Cont.) Architecture Instructions
Mnemonic

Format

Opcode

Description

LDS
LDT
MB
MF_FPCR

Mem
Mem
Mfc
F-P

22
23
18.4000
17.025

MSKBL
MSKLH
MSKLL
MSKQH
MSKQL
MSKWH
MSKWL
MT_FPCR

Opr
Opr
Opr
Opr
Opr
Opr
Opr
F-P

12.02
12.62
12.22
12.72
12.32
12.52
12.12
17.024

MULF
MULG
MULL
MULL/V
MULQ
MULQ/V
MULS
MULT
ORNOT
RC
RET
RPCC
RS
S4ADDL
S4ADDQ
S4SUBL
S4SUBQ
S8ADDL
S8ADDQ
S8SUBL
S8SUBQ
SLL
SRA
SRL
STF

F-P
F-P
Opr
Opr
Opr
Opr
F-P
F-P
Opr
Mfc
Mbr
Mfc
Mfc
Opr
Opr
Opr
Opr
Opr
Opr
Opr
Opr
Opr
Opr
Opr
Mem

15.082
15.0A2
13.00
13.40
13.20
13.60
16.082
16.0A2
11.28
18.E0
1A.2
18.C0
18.F000
10.02
10.22
10.0B
10.2B
10.12
10.32
10.1B
10.3B
12.39
12.3C
12.34
24

Load S_floating
Load T_floating
Memory barrier
Move from floating-point control
register
Mask byte low
Mask longword high
Mask longword low
Mask quadword high
Mask quadword low
Mask word high
Mask word low
Move to floating-point control
register
Multiply F_floating
Multiply G_floating
Multiply longword
Multiply longword
Multiply quadword
Multiply quadword
Multiply S_floating
Multiply T_floating
Logical sum with complement
Read and clear
Return from subroutine
Read process cycle counter
Read and set
Scaled add longword by 4
Scaled add quadword by 4
Scaled subtract longword by 4
Scaled subtract quadword by 4
Scaled add longword by 8
Scaled add quadword by 8
Scaled subtract longword by 8
Scaled subtract quadword by 8
Shift left logical
Shift right arithmetic
Shift right logical
Store F_floating
(continued on next page)

Preliminary—Subject to Change—July 1996 A–5

A.1 Alpha Instruction Summary

Table A–2 (Cont.) Architecture Instructions
Mnemonic

Format

Opcode

Description

STG
STS
STL
STL_C
STQ
STQ_C
STQ_U
STT
SUBF
SUBG
SUBL
SUBL/V
SUBQ
SUBQ/V
SUBS
SUBT
TRAPB
UMULH

Mem
Mem
Mem
Mem
Mem
Mem
Mem
Mem
F-P
F-P
Opr

Store G_floating
Store S_floating
Store longword
Store longword conditional
Store quadword
Store quadword conditional
Store unaligned quadword
Store T_floating
Subtract F_floating
Subtract G_floating
Subtract longword

F-P
F-P
Mfc
Opr

25
26
2C
2E
2D
2F
0F
27
15.081
15.0A1
10.09
10.49
10.29
10.69
16.081
16.0A1
18.00
13.30

WMB
XOR
ZAP
ZAPNOT

Mfc
Opr
Opr
Opr

18.44
11.40
12.30
12.31

Opr

Subtract quadword
Subtract S_floating
Subtract T_floating
Trap barrier
Unsigned multiply quadword
high
Write memory barrier
Logical difference
Zero bytes
Zero bytes not

A.1.1 Opcodes Reserved for Digital
Table A–3 lists opcodes reserved for Digital.
Table A–3 Opcodes Reserved for Digital
Mnemonic

Opcode

Mnemonic

Opcode

Mnemonic

Opcode

OPC01

OPC05

OPC0B

OPC02

OPC06

OPC0C

OPC03

OPC07

OPC0D

OPC04

OPC0A

OPC14

A–6 Preliminary—Subject to Change—July 1996

A.1 Alpha Instruction Summary

A.1.2 Opcodes Reserved for PALcode
Table A–4 lists the 21164-specific instructions. For more information, refer to
Section 6.6.
Table A–4 Opcodes Reserved for PALcode
21164
Mnemonic

Opcode

Architecture
Mnemonic

HW_LD

PAL1B

Performs Dstream load
instructions.

HW_ST

PAL1F

Performs Dstream store
instructions.

HW_REI

PAL1E

Returns instruction flow to the
program counter (PC) pointed
to by EXC_ADDR internal
processor register (IPR).

HW_MFPR

PAL19

Accesses the Ibox, Mbox, and
Dcache IPRs.

HW_MTPR

PAL1D

Accesses the Ibox, Mbox, and
Dcache IPRs.

Function

A.2 IEEE Floating-Point Instructions
Table A–5 lists the hexadecimal value of the 11-bit function code field for the
IEEE floating-point instructions, with and without qualifiers. The opcode for
these instructions is 1616 .
Table A–5 IEEE Floating-Point Instruction Function Codes
Mnemonic

None

/UC

/UM

/UD

ADDS
ADDT
CMPTEQ
CMPTLT
CMPTLE
CMPTUN
CVTQS
CVTQT

080
0A0
0A5
0A6
0A7
0A4
0BC
0BE

000
020

040
060

0C0
0E0

180
1A0

100
120

140
160

1C0
1E0

03C
03E

07C
07E

0FC
0FE
(continued on next page)

Preliminary—Subject to Change—July 1996 A–7

A.2 IEEE Floating-Point Instructions

Table A–5 (Cont.) IEEE Floating-Point Instruction Function Codes
Mnemonic

None

/UC

/UM

/UD

CVTTS
DIVS
DIVT
MULS
MULT
SUBS
SUBT

0AC
083
0A3
082
0A2
081
0A1

02C
003
023
002
022
001
021

06C
043
063
042
062
041
061

0EC
0C3
0E3
0C2
0E2
0C1
0E1

1AC
183
1A3
182
1A2
181
1A1

12C
103
123
102
122
101
121

16C
143
163
142
162
141
161

1EC
1C3
1E3
1C2
1E2
1C1
1E1

Mnemonic

/SU

/SUC

/SUM

/SUD

/SUI

/SUIC

/SUIM

/SUID

ADDS
ADDT
CMPTEQ
CMPTLT
CMPTLE
CMPTUN
CVTQS
CVTQT
CVTTS
DIVS
DIVT
MULS
MULT
SUBS
SUBT

580
5A0
5A5
5A6
5A7
5A4

500
520

540
560

5C0
5E0

780
7A0

700
720

740
760

7C0
7E0

5AC
583
5A3
582
5A2
581
5A1

52C
503
523
502
522
501
521

56C
543
563
542
562
541
561

5EC
5C3
5E3
5C2
5E2
5C1
5E1

7BC
7BE
7AC
783
7A3
782
7A2
781
7A1

73C
73E
72C
703
723
702
722
701
721

77C
77E
76C
743
763
742
762
741
761

7FC
7FE
7EC
7C3
7E3
7C2
7E2
7C1
7E1

Mnemonic

None

CVTST

2AC

6AC

Mnemonic

None

/VC

/SV

/SVC

/SVI

/SVIC

CVTTQ

0AF

02F

1AF

12F

5AF

52F

7AF

72F

Mnemonic

/VD

/SVD

/SVID

/VM

/SVM

/SVIM

CVTTQ

0EF

1EF

5EF

7EF

06F

16F

56F

76F

A–8 Preliminary—Subject to Change—July 1996

A.2 IEEE Floating-Point Instructions

Programming Note
Because underflow cannot occur for CMPTxx, there is no difference
in function or performance between CMPTxx/S and CMPTxx/SU. It is
intended that software generate CMPTxx/SU in place of CMPTxx/S.
In the same manner, CVTQS and CVTQT can take an inexact
result trap, but not an underflow. Because there is no encoding for a
CVTQx/SI instruction, it is intended that software generate CVTQx/SUI
in place of CVTQx/SI.

A.3 VAX Floating-Point Instructions
Table A–6 lists the hexadecimal value of the 11-bit function code field for the
VAX floating-point instructions. The opcode for these instructions is 1516 .
Table A–6 VAX Floating-Point Instruction Function Codes
Mnemonic

None

/UC

/SC

/SU

/SUC

ADDF
CVTDG
ADDG
CMPGEQ
CMPGLT
CMPGLE
CVTGF
CVTGD
CVTQF
CVTQG
DIVF
DIVG
MULF
MULG
SUBF
SUBG

080
09E
0A0
0A5
0A6
0A7
0AC
0AD
0BC
0BE
083
0A3
082
0A2
081
0A1

000
01E
020

180
19E
1A0

100
11E
120

400
41E
420

580
59E
5A0

500
51E
520

02C
02D
03C
03E
003
023
002
022
001
021

1AC
1AD

12C
12D

480
49E
4A0
4A5
4A6
4A7
4AC
4AD

42C
42D

5AC
5AD

52C
52D

183
1A3
182
1A2
181
1A1

103
123
102
122
101
121

483
4A3
482
4A2
481
4A1

403
423
402
422
401
421

583
5A3
582
5A2
581
5A1

503
523
502
522
501
521

Mnemonic

None

/VC

/SC

/SV

/SVC

CVTGQ

0AF

02F

1AF

12F

4AF

42F

5AF

52F

Preliminary—Subject to Change—July 1996 A–9

A.4 Opcode Summary

A.4 Opcode Summary
Table A–7 lists all Alpha opcodes from 00 (CALL_PAL) through 3F (BGT).
In the table, the column headings that appear over the instructions have a
granularity of 816 . The rows beneath the Offset column supply the individual
hexadecimal number to resolve that granularity.
If an instruction column has a 0 in the right (low) hexadecimal digit, replace
that 0 with the number to the left of the backslash in the Offset column on
the instruction’s row. If an instruction column has an 8 in the right (low)
hexadecimal digit, replace that 8 with the number to the right of the backslash
in the Offset column.
For example, the third row (2/A) under the 1016 column contains the symbol
INTS*, representing the all-integer shift instructions. The opcode for those
instructions would then be 1216 because the 0 in 10 is replaced by the 2 in
the Offset column. Likewise, the third row under the 1816 column contains
the symbol JSR*, representing all jump instructions. The opcode for those
instructions is 1A because the 8 in the heading is replaced by the number to
the right of the backslash in the Offset column.
The instruction format is listed under the instruction symbol.

A–10 Preliminary—Subject to Change—July 1996

A.4 Opcode Summary

Table A–7 Opcode Summary
Offset

0/8

PAL*
(pal)

LDA
(mem)

INTA*
(op)

MISC*
(mem)

LDF
(mem)

LDL
(mem)

BR
(br)

BLBC
(br)

1/9

Res

LDAH
(mem)

INTL*
(op)

\ PAL\

LDG
(mem)

LDQ
(mem)

FBEQ
(br)

BEQ
(br)

2/A

Res

INTS*
(op)

JSR*
(mem)

LDS
(mem)

LDL_L
(mem)

FBLT
(br)

BLT
(br)

3/B

Res

LDQ_U
(mem)

INTM*
(op)

\ PAL\

LDT
(mem)

LDQ_L
(mem)

FBLE
(br)

BLE
(br)

4/C

Res

STF
(mem)

STL
(mem)

BSR
(br)

BLBS
(br)

5/D

Res

FLTV*
(op)

\ PAL\

STG
(mem)

STQ
(mem)

FBNE
(br)

BNE
(br)

6/E

Res

FLTI*
(op)

\ PAL\

STS
(mem)

STL_C
(mem)

FBGE
(br)

BGE
(br)

7/F

Res

STQ_U
(mem)

FLTL*
(op)

\ PAL\

STT
(mem)

STQ_C
(mem)

FBGT
(br)

BGT
(br)

Symbol
FLTI*
FLTL*
FLTV*
INTA*
INTL*
INTM*
INTS*
JSR*
MISC*
PAL*
\ PAL\
Res

Meaning
IEEE floating-point instruction opcodes
Floating-point operate instruction opcodes
VAX floating-point instruction opcodes
Integer arithmetic instruction opcodes
Integer logical instruction opcodes
Integer multiply instruction opcodes
Integer shift instruction opcodes
Jump instruction opcodes
Miscellaneous instruction opcodes
PALcode instruction (CALL_PAL) opcodes
Reserved for PALcode
Reserved for Digital

Preliminary—Subject to Change—July 1996 A–11

A.5 Required PALcode Function Codes

A.5 Required PALcode Function Codes
The opcodes listed in Table A–8 are required for all Alpha implementations.
The notation used is oo.ffff, where oo is the hexadecimal 6-bit opcode and ffff is
the hexadecimal 26-bit function code.
Table A–8 Required PALcode Function Codes
Mnemonic

Type

Function Code

DRAINA

Privileged

00.0002

HALT

Privileged

00.0000

IMB

Unprivileged

00.0086

A.6 Alpha 21164 Microprocessor IEEE Floating-Point
Conformance
The 21164 supports the IEEE floating-point operations as defined by the Alpha
architecture. Support for a complete implementation of the IEEE Standard
for Binary Floating-Point Arithmetic (ANSI/IEEE Standard 754 1985) is
provided by a combination of hardware and software as described in the Alpha
Architecture Reference Manual.
Additional information about writing code to support precise exception
handling (necessary for complete conformance to the standard) is in the Alpha
Architecture Reference Manual.
The following information is specific to the 21164:
•

Invalid operation (INV)
The invalid operation trap is always enabled. If the trap occurs, then the
destination register is UNPREDICTABLE. This exception is signaled if any
VAX architecture operand is non-finite (reserved operand or dirty zero) and
the operation can take an exception (that is, certain instructions, such as
CPYS, never take an exception). This exception is signaled if any IEEE
operand is non-finite (NAN, INF, denorm) and the operation can take an
exception. This trap is also signaled for an IEEE format divide of +/– 0
divided by +/– 0. If the exception occurs, then FPCR<INV> is set and the
trap is signaled to the Ibox.

A–12 Preliminary—Subject to Change—July 1996

A.6 Alpha 21164 Microprocessor IEEE Floating-Point Conformance

•

Divide-by-zero (DZE)
The divide-by-zero trap is always enabled. If the trap occurs, then the
destination register is UNPREDICTABLE. For VAX architecture format,
this exception is signaled whenever the numerator is valid and the
denominator is zero. For IEEE format, this exception is signaled whenever
the numerator is valid and non-zero, with a denominator of +/– 0. If the
exception occurs, then FPCR<DZE> is set and the trap is signaled to the
Ibox.
For IEEE format divides, 0/0 signals INV, not DZE.

•

Floating overflow (OVF)
The floating overflow trap is always enabled. If the trap occurs, then the
destination register is UNPREDICTABLE. The exception is signaled if the
rounded result exceeds in magnitude the largest finite number, which can
be represented by the destination format. This applies only to operations
whose destination is a floating-point data type. If the exception occurs,
then FPCR<OVF> is set and the trap is signaled to the Ibox.

•

Underflow (UNF)
The underflow trap can be disabled. If underflow occurs, then the
destination register is forced to a true zero, consisting of a full 64 bits of
zero. This is done even if the proper IEEE result would have been –0. The
exception is signaled if the rounded result is smaller in magnitude than the
smallest finite number that can be represented by the destination format.
If the exception occurs, then FPCR<UNF> is set. If the trap is enabled,
then the trap is signaled to the Ibox. The 21164 never produces a denormal
number; underflow occurs instead.

Preliminary—Subject to Change—July 1996 A–13

A.6 Alpha 21164 Microprocessor IEEE Floating-Point Conformance

•

Inexact (INE)
The inexact trap can be disabled. The destination register always contains
the properly rounded result, whether the trap is enabled. The exception
is signaled if the rounded result is different from what would have been
produced if infinite precision (infinitely wide data) were available. For
floating-point results, this requires both an infinite precision exponent and
fraction. For integer results, this requires an infinite precision integer and
an integral result. If the exception occurs, then FPCR<INE> is set. If the
trap is enabled, then the trap is signaled to the Ibox.
The IEEE-754 specification allows INE to occur concurrently with either
OVF or UNF. Whenever OVF is signaled (if the inexact trap is enabled),
INE is also signaled. Whenever UNF is signaled (if the inexact trap is
enabled), INE is also signaled. The inexact trap also occurs concurrently
with integer overflow. All valid opcodes that enable INE also enable both
overflow and underflow.
If a CVTQL results in an integer overflow (IOV), then FPCR<INE> is
automatically set. (The INE trap is never signaled to the Ibox because
there is no CVTQL opcode that enables the inexact trap.)

•

Integer overflow (IOV)
The integer overflow trap can be disabled. The destination register always
contains the low-order bits (<64> or <32>) of the true result (not the
truncated bits). Integer overflow can occur with CVTTQ, CVTGQ, or
CVTQL. In conversions from floating to quadword integer or longword
integer, an integer overflow occurs if the rounded result is outside the
range 0263 ..26301 . In conversions from quadword integer to longword
integer, an integer overflow occurs if the result is outside the range
0231 ..23101 . If the exception occurs, then the appropriate bit in the FPCR
is set. If the trap is enabled, then the trap is signaled to the Ibox.

•

Software completion (SWC)
The software completion signal is not recorded in the FPCR. The state
of this signal is always sent to the Ibox. If the Ibox detects the assertion
of any of the listed exceptions concurrent with the assertion of the SWC
signal, then it sets EXC_SUM<SWC>.

Input exceptions always take priority over output exceptions. If both exception
types occur, then only the input exception is recorded in the FPCR and only
the input exception is signaled to the Ibox.

A–14 Preliminary—Subject to Change—July 1996

B
Alpha 21164 Microprocessor
Specifications
Table B–1 lists specifications for the 21164.

Preliminary—Subject to Change—July 1996 B–1

Table B–1

Alpha 21164 Microprocessor Specifications

Feature

Description

Cycle time range

4.4 ns (227 MHz) to 3.0 ns (333 MHz).

Process technology

0.5-micron CMOS.

Transistor count

9.3 million.

Die size

664 2 732 mils.

Package

499-pin IPGA (interstitial pin grid array).

Number of signal pins

292.

Typical worst case power
@Vdd = 3.3 V

46 W @ 3.75 ns cycle time (266 MHz)1 ,
51 W @ 3.33 ns cycle time (300 MHz)1 ,
56 W @ 3.0 ns cycle time (333 MHz)1 .

Power supply

3.3 V dc.

Clocking input

Two times the internal clock speed (for example, 571.4 MHz at a
3.5-ns cycle time).

Virtual address size

43 bits.

Physical address size

40 bits.

Page size

8K bytes.

Issue rate

2 integer instructions and 2 floating-point instructions per cycle.

Integer instruction
pipeline

7 stage.

Floating instruction
pipeline

9 stage.

Onchip L1 Dcache

8K-byte, physical, direct-mapped, write-through, 32-byte block,
32-byte fill.

Onchip L1 Icache

8K-byte, virtual, direct-mapped, 32-byte block, 32-byte fill, 128
address space numbers (ASNs) (MAX_ASN=127).

Onchip L2 Scache

96K-byte, physical, 3-way set-associative, write-back, 32- or 64-byte
block, 32- or 64-byte fill.

Onchip data translation
buffer

64-entry, fully associative, not-last-used replacement, 8K pages, 128
ASNs (MAX_ASN=127), full granularity hint support.

Onchip instruction
translation buffer

48-entry, fully associative, not-last-used replacement, 128 ASNs
(MAX_ASN=127), full granularity hint support.

Floating-point unit

Onchip FPU supports both IEEE and Digital floating point.

Bus

Separate data and address bus, 128-bit/64-bit data bus.

Serial ROM interface

Allows microprocessor to access a serial ROM.

1 Power consumption scales linearly with frequency over the frequency range 225 MHz to 333 MHz.

B–2 Preliminary—Subject to Change—July 1996

C
Serial Icache Load Predecode Values
The following C code calculates the predecode values of a serial Icache load. A
software tool called the SROM Packer converts a binary image into a format
suitable for Icache serial loading. This tool is available from Digital.

#include <stdio.h>
/* fillmap [0 - 127] maps data 127:0, etc. */
/* fillmap[n] is bit position in output vector. */
/* bit 0 of this vector is first-in; bit 199 is last */
const int dfillmap [128] = {
42,44,46,48,50,52,54,56,
58,60,62,64,66,68,70,72,
74,76,78,80,82,84,86,88,
90,92,94,96,98,100,102,104,
43,45,47,49,51,53,55,57,
59,61,63,65,67,69,71,73,
75,77,79,81,83,85,87,89,
91,93,95,97,99,101,103,105,
128,130,132,134,136,138,140,142,
144,146,148,150,152,154,156,158,
160,162,164,166,168,170,172,174,
176,178,180,182,184,186,188,190,
129,131,133,135,137,139,141,143,
145,147,149,151,153,155,157,159,
161,163,165,167,169,171,173,175,
177,179,181,183,185,187,189,191
};

/* data 0:127 -- fillmap[0:127]*/
/* 0:7 */
/* 8:15 */
/* 16:23 */
/* 24:31 */
/* 32:39 */
/* 40:47 */
/* 48:55 */
/* 56:63 */
/* 64:71 */
/* 72:79 */
/* 80:87 */
/* 88:95 */
/* 96:103 */
/* 104:111 */
/* 112:119 */
/* 120:127 */

const int BHTfillmap[8] = {
199,198,197,196,195,194,193,192
};

/* BHT vector 0:7 -- BHTfillmap[0:7] */
/* 0:7 */

const int predfillmap[20] = {
106,108,110,112,114,
107,109,111,113,115,
118,120,122,124,126,
119,121,123,125,127
};

/* predecodes 0:19 -- predfillmap[0:19] */
/* 0:4 */
/* 5:9 */
/* 10:14 */
/* 15:19 */

const int octawpfillmap =
117;

/* octaword parity */

Preliminary—Subject to Change—July 1996 C–1

const int predpfillmap =
116;

/* predecode parity */

const int tagfillmap[30] = {
29,28,27,26,25,24,23,22,21,20,
19,18,17,16,15,14,13,12,11,10,
09,08,07,06,05,04,03,02,01,00
};

/* tag bits 13:42 -- tagfillmap[0:29] */
/* 13:22 */
/* 23:32 */
/* 33:42 */

const int asnfillmap[7] = {
37,36,35,34,33,32,31
};

/* asn 0:6 -- asnfillmap[0:6] */
/* 0:6 */

const int asmfillmap =
30;

/* asm -- asmfillmap */

const int tagphysfillmap =
38;

/* tagphysical address -- tagphysfillmap */

const int tagvalfillmap[2] = {
40,39
};

/* tag valid bits 0:1 -- tagvalfillmap */
/* 0:1 */

const int tagparfillmap =
41;

/* tag parity -- tagparfillmap */

main(argc, argv)
int argc;
char *argv[];
{
int i,j,k,t;
int status,instatus, instr_count;
char filename[256],ofilename[256],hfilename[256];
char *charptr;
int instr[4], outvector[7];
FILE* infile, outfile, hexfile;
int base, asm, asn, tag, predecodes,owparity,pdparity,tparity,
tvalids,tphysical,bhtvector, offset, chksum;
strcpy (filename ,"loadfile.exe");
strcpy(ofilename,"loadfile.srom");
base = 0;
tag = 0;
asn = 0;
asm = 1;
tphysical= 1;
bhtvector = 0;
offset = 0;
if (argc > 1)
strcpy(filename, argv[1]);
if (argc > 2)
strcpy(ofilename, argv[2]);
if (argc > 3)
{
base = strtol(argv[3],NULL,16) & (0xffffffff << 13);
tag = base >> 13;
}

C–2 Preliminary—Subject to Change—July 1996

if (argc > 4)
asn = strtol(argv[4],NULL,16) & 0x7f;
if (argc > 5)
asm = strtol(argv[5],NULL,16) & 1;
if (argc > 6)
tphysical = strtol(argv[6],NULL,16) & 1;
if (argc > 7)
bhtvector = strtol(argv[7],NULL,16) & 0xff;
if (NULL == (infile = fopen(filename,"rb")))
{
printf("input file open error: %s\n", filename);
exit(0);
}
if (NULL == (outfile = fopen(ofilename, "wb")))
{
printf("binary output file open error: %s\n", ofilename);
exit(0);
}
strcpy(hfilename,ofilename);
charptr = strpbrk(hfilename,".;");
if (charptr != NULL) *charptr = 0;
strcat(hfilename,".hex");
if (NULL == (hexfile = fopen(hfilename, "w")))
{
printf("hex output file open error: %s\n", hfilename);
exit(0);
}
fprintf(hexfile,":020000020000FC\n");
tparity = eparity(tag) ^ eparity(tphysical) ^ eparity(asn);
tvalids = 3;
instatus = 0;
instr_count = 0;
for (i=0; i<512; i++)
{
for (j=0;j<4;j++) instr[j] = 0;
for (j=0;j<7;j++) outvector[j]=0;
if (instatus == 0)
{
if (16 > (status = fread(&instr[0],1,16,infile)))
instatus = 1;
instr_count += status/4;
}

Preliminary—Subject to Change—July 1996 C–3

predecodes=0;
owparity = 0;
for (j=0;j<4;j++)
{
predecodes |= (4 ^ instrpredecode(instr[j])) << (j*5);
/* invert bit 2 to match fill scan chain attribute */
owparity ^= eparity(instr[j]);
}
pdparity = eparity(predecodes);
/* bhtvector */
for (j=0;j<8;j++)
{
t = BHTfillmap[j];
outvector[t>>5] |= ((bhtvector >> j) & 1) << (t&0x1f);
}
/* instructions */
for (k=0;k<4;k++)
{
for (j=0;j<32;j++)
{
t = dfillmap[j+k*32];
outvector[t>>5] |= ((instr[k] >> j) & 1) << (t&0x1f);
}
}
/* predecodes */
for (j=0;j<20;j++)
{
t = predfillmap[j];
outvector[t>>5] |= ((predecodes >> j) & 1) << (t&0x1f);
}
/* owparity */
outvector[octawpfillmap>>5] |= owparity << (octawpfillmap&0x1f);
/* pdparity */
outvector[predpfillmap>>5] |= pdparity << (predpfillmap&0x1f);
/* tparity */
outvector[tagparfillmap>>5] |= tparity << (tagparfillmap&0x1f);
/* tvalids */
for (j=0;j<2;j++)
{
t =tagvalfillmap[j];
outvector[t>>5] |= ((tvalids >> j) & 1) << (t&0x1f);
}
/* tphysical */
outvector[tagphysfillmap>>5] |= tphysical << (tagphysfillmap&0x1f);
/* asn */
for (j=0;j<7;j++)
{
t = asnfillmap[j];
outvector[t>>5] |= ((asn >> j) & 1) << (t&0x1f);
}

C–4 Preliminary—Subject to Change—July 1996

/* asm */
outvector[asmfillmap>>5] |= asm << (asmfillmap&0x1f);
/* tag */
for (j=0;j<30;j++)
{
t = tagfillmap[j];
outvector[t>>5] |= ((tag >> j) & 1) << (t&0x1f);
}
fwrite(&outvector[0],1,25,outfile);
fprintf(hexfile,":19%04X00",offset);
chksum = (offset & 0xff) + (offset >> 8) + 0x19;
for (j=0; j<25; j++)
{
charptr = ((char*) &outvector[0]) + j;
fprintf(hexfile,"%02X", (0xff& *charptr));
chksum += *charptr;
}
offset += 25;
fprintf(hexfile,"%02X\n", (-chksum) & 0xff);
}
fprintf(hexfile,":00000001FF\n");
if (instatus == 0)
if (fread(&instr[0],1,16,infile))
{
printf("There are more instructions in the input file than can");
printf("be fit in the output file: \n");
printf(" truncated the input file after 8K of instructions!!!\n");
}
printf("\n");
printf("Total intructions processed = %d\n", instr_count);
fclose(infile);
fclose(outfile);
fclose(hexfile);
exit(0);
}
int eparity(int x)
{
x = x ^ (x >> 16);
x = x ^ (x >> 8);
x = x ^ (x >> 4);
x = x ^ (x >> 2);
x = x ^ (x >> 1);
return (x&1);
}
#define EXT(data, bit)\
(((data) & ((unsigned) 1 << (bit))) != 0)
#define EXTV(data, hbit, lbit)\
(((data) >> (lbit)) & \
((((hbit) - (lbit) + 1) == 32) ? ((unsigned)0xffffffff) : \
(~((unsigned)0xffffffff << ((hbit) - (lbit) + 1)))))
#define INS(name, bit, data)\
(name) = (((name) & ~((unsigned) 1 << (bit))) | \
(((unsigned) (data) << (bit)) & ((unsigned) 1 << (bit))))

Preliminary—Subject to Change—July 1996 C–5

int instrpredecode(int inst)
{
int result;
int opcode;
int func;
int jsr_type;
int ra;
int out0;
int out1;
int out2;
int out3;
int out4;
int e0_only;
int e1_only;
int ee;
int lnoop;
int fadd;
int fmul;
int fe;
int br_type;
int ld;
int store;
int br;
int call_pal;
int bsr;
int ret_rei;
int jmp;
int jsr_cor;
int jsr;
int cond_br;
opcode = EXTV(inst, 31, 26 );
func = EXTV(inst, 12, 5);
jsr_type = EXTV(inst, 15,14);
ra = EXTV(inst,25,21);

C–6 Preliminary—Subject to Change—July 1996

e0_only = (opcode == 0x24) ||
/* STF */
(opcode == 0x25) ||
/* STG */
(opcode == 0x26) ||
/* STS */
(opcode == 0x27) ||
/* STT */
(opcode == 0x0F) ||
/* STQ_U */
(opcode == 0x2A) ||
/* LDL_L */
(opcode == 0x2B) ||
/* LDQ_L */
(opcode == 0x2C) ||
/* STL */
(opcode == 0x2D) ||
/* STQ */
(opcode == 0x2E) ||
/* STL_C */
(opcode == 0x2F) ||
/* STQ_C */
(opcode == 0x1F) ||
/* HW_ST*/
(opcode == 0x18) ||
/* MISC mem format: FETCH/_M, RS, RC, RPCC, TRAPB, MB) */
(opcode == 0x12) ||
/* EXT,MSK,INS,SRX,SLX,ZAP*/
(opcode == 0x13) ||
/* MULX */
((opcode == 0x1D) && (EXT(inst,8) == 0)) || /* MBOX HW_MTPR */
((opcode == 0x19) && (EXT(inst,8) == 0)) || /* MBOX HW_MFPR */
(opcode == 0x01) ||
/* RESDEC’s */
(opcode == 0x02) ||
/* RESDEC’s */
(opcode == 0x03) ||
/* RESDEC’s */
(opcode == 0x04) ||
/* RESDEC’s */
(opcode == 0x05) ||
/* RESDEC’s */
(opcode == 0x06) ||
/* RESDEC’s */
(opcode == 0x07) ||
/* RESDEC’s */
(opcode == 0x0a) ||
/* RESDEC’s */
(opcode == 0x0c) ||
/* RESDEC’s */
(opcode == 0x0d) ||
/* RESDEC’s */
(opcode == 0x0e) ||
/* RESDEC’s */
(opcode == 0x14) ||
/* RESDEC’s */
(opcode == 0x1c);
/* RESDEC’s */

Preliminary—Subject to Change—July 1996 C–7

e1_only = (opcode == 0x30) ||
/* BR */
(opcode == 0x34) ||
/* BSR */
(opcode == 0x38) ||
/* BLBC */
(opcode == 0x39) ||
/* BEQ */
(opcode == 0x3A) ||
/* BLT */
(opcode == 0x3B) ||
/* BLE */
(opcode == 0x3C) ||
/* BLBS */
(opcode == 0x3D) ||
/* BNE */
(opcode == 0x3E) ||
/* BGE */
(opcode == 0x3F) ||
/* BGT */
(opcode == 0x1A) ||
/* JMP,JSR,RET,JSR_COROT */
(opcode == 0x1E) ||
/* HW_REI */
(opcode == 0x00) ||
/* CALL_PAL */
((opcode == 0x1D) && (EXT(inst,8) == 1)) || /* IBOX HW_MTPR */
((opcode == 0x19) && (EXT(inst,8) == 1)); /* IBOX HW_MTPR */
ee =
(opcode == 0x10) ||
/* ADD, SUB, CMP */
(opcode == 0x11) ||
/* AND, BIC etc. logicals */
(opcode == 0x28) ||
/* LDL */
(opcode == 0x29) ||
/* LDQ */
(opcode == 0x0B)&(ra != 0x1F) || /* LDQ_U */
(opcode == 0x08) ||
/* LDA */
(opcode == 0x09) ||
/* LDAH */
(opcode == 0x20) ||
/* LDF */
(opcode == 0x21) ||
/* LDG */
(opcode == 0x22) ||
/* LDS */
(opcode == 0x23) ||
/* LDT */
(opcode == 0x1B);
/* HW_LD */
lnoop =

(opcode == 0x0B)&(ra == 0x1F);

/* LDQ_U R31, x(y) - NOOP*/

fadd =

((opcode == 0x17) && (func != 0x20)) ||
/* Flt, datatype indep excl CPYS */
((opcode == 0x15) && ((func & 0xf) != 0x2)) ||
/* VAX excl MUL’s */
((opcode == 0x16) && ((func & 0xf) != 0x2)) ||
/* IEEE excl MUL’s */
(opcode == 0x31) ||
/* FBEQ */
(opcode == 0x32) ||
/* FBLT */
(opcode == 0x33) ||
/* FBLE */
(opcode == 0x35) ||
/* FBNE */
(opcode == 0x36) ||
/* FBGE */
(opcode == 0x37);
/* FBGT */

fmul =
((opcode == 0x15) && ((func & 0xf) == 0x2)) || /* VAX MUL’s */
((opcode == 0x16) && ((func & 0xf) == 0x2));
/* IEEE MUL’s */
fe =

((opcode == 0x17) && (func == 0x20));

/* CPYS */

br_type = ((opcode & 0x30) == 0x30) || /* all branches */
(opcode == 0x1A) ||
/* JMP’s */
(opcode == 0x00) ||
/* CALL PAL */
(opcode == 0x1E);
/* HW_REI */

C–8 Preliminary—Subject to Change—July 1996

ld =
(opcode == 0x28) ||
/* LDL */
(opcode == 0x29) ||
/* LDQ */
/*
(opcode == 0x2A) ||
LDL_L */
/*
(opcode == 0x2B) ||
LDQ_L */
(opcode == 0x0B) ||
/* LDQ_U */
(opcode == 0x20) ||
/* LDF */
(opcode == 0x21) ||
/* LDG */
(opcode == 0x22) ||
/* LDS */
(opcode == 0x23) ||
/* LDT */
(opcode == 0x1B);
/* HW_LD */
store =
(opcode == 0x24) ||
/* STF */
(opcode == 0x25) ||
/* STG */
(opcode == 0x26) ||
/* STS */
(opcode == 0x27) ||
/* STT */
(opcode == 0x0F) ||
/* STQ_U */
(opcode == 0x2C) ||
/* STL */
(opcode == 0x2D) ||
/* STQ */
(opcode == 0x2E) ||
/* STL_C */
(opcode == 0x2F) ||
/* STQ_C */
(opcode == 0x18) ||
/* Misc: TRAPB, MB, RS, RC, RPCC etc. */
(opcode == 0x1F) ||
/* HW_ST */
(opcode == 0x2A) ||
/* LDL_L */
(opcode == 0x2B);
/* LDQ_L */
br =

(opcode == 0x30);

/* all branches */

call_pal = (opcode == 0x00);
bsr =

/* call PAL */

(opcode == 0x34);

ret_rei = ((opcode == 0x1A) && (jsr_type == 0x2)) ||
((opcode == 0x1E) && (jsr_type != 0x3));
jmp = ((opcode == 0x1A) && (jsr_type == 0x0));
jsr_cor = ((opcode == 0x1A) && (jsr_type == 0x3));
jsr = ((opcode == 0x1A) && (jsr_type == 0x1));
cond_br = (opcode == 0x31) ||
(opcode == 0x32) ||
(opcode == 0x33) ||
(opcode == 0x35) ||
(opcode == 0x36) ||
(opcode == 0x37) ||
(opcode == 0x38) ||
(opcode == 0x39) ||
(opcode == 0x3A) ||
(opcode == 0x3B) ||
(opcode == 0x3C) ||
(opcode == 0x3D) ||
(opcode == 0x3E) ||
(opcode == 0x3F);

Preliminary—Subject to Change—July 1996 C–9

out0 = br || bsr || jmp || jsr || (ee && !ld) || (e0_only && !store);
out1 = ret_rei ||(e1_only && !br_type)|| jmp ||jsr_cor|| jsr || lnoop ||
(fadd && !br_type) || fe;;
out2 = call_pal || bsr || jsr_cor || e0_only ||jsr ||fmul || fe;
out3 = (e1_only && cond_br) || (e1_only && !br_type) || fadd || fmul || fe;
out4 = ee || lnoop || e0_only || fadd || fmul || fe;
result = 0;
INS( result, 0, out0 );
INS( result, 1, out1 );
INS( result, 2, out2 );
INS( result, 3, out3 );
INS( result, 4, out4 );
return (result);
}

C–10 Preliminary—Subject to Change—July 1996

D
Errata Sheet
Table D–1 lists the revision history for this document.
Table D–1 Document Revision History
Date

Revision

March 28, 1994

First draft.

May 16, 1994

Second draft.

July 20, 1994

First Preliminary version.

September 12, 1994

Second Preliminary version. (First printing.)

December 22, 1994

Final draft.

April 3, 1995

Third Preliminary version. (Second printing.)

December 1995

Fourth Preliminary version. (Third printing.)

Preliminary—Subject to Change—July 1996 D–1

E
Technical Support and Ordering
Information
E.1 Technical Support
If you need technical support or help deciding which literature best meets your
needs, call the Semiconductor Information Line:
United States and Canada
Outside North America

1–800–332–2717
+1–508–628–4760

E.2 Ordering Digital Semiconductor Products
To order Alpha 21164 microprocessor evaluation boards and motherboards,
contact your local distributor.
You can order the following semiconductor products from Digital:
Product

Order Number

Alpha 21164 333-MHz Microprocessor

21164–333

Alpha 21164 300-MHz Microprocessor

21164–300

Alpha 21164 300-MHz Microprocessor for Windows NT

21164–P2

Alpha 21164 266-MHz Microprocessor

21164–266

Alpha 21164 266-MHz Microprocessor for Windows NT

21164–P1

Alpha 21164 Microprocessor Evaluation Board 266-MHz
Kit (Supports Digital UNIX, OpenVMS, and Windows NT
operating systems.)

21A04–01

Alpha 21164 Microprocessor Motherboard 266-MHz Kit
(Supports the Windows NT operating system.)

21A04–A0

Preliminary—Subject to Change—July 1996 E–1

E.3 Ordering Digital Semiconductor Sample Kits

E.3 Ordering Digital Semiconductor Sample Kits
To order an Alpha 21164 Microprocessor Sample Kit, which contains one
Alpha 21164 microprocessor, one heat sink, and supporting documentation,
call 1–800–DIGITAL. You will need a purchase order number or credit card to
order the following products:
Product

Order Number

Alpha 21164–266 Sample Kit

21164–SA

E.4 Ordering Associated Literature
The following table lists some of the available Digital Semiconductor literature.
For a complete list, contact the Digital Semiconductor Information Line.
Title
Alpha Architecture Reference Manual

Order Number
1

EY–L520E–DP–YCH

Alpha AXP Architecture Handbook

EC–QD2KA–TE

Alpha 21164 Microprocessor Data Sheet

EC–QAEPC–TE

Alpha 21164 Microprocessor Product Brief

EC–QAENB–TE

Alpha 21164 Evaluation Board Read Me First

EC–QD2VB–TE

Alpha 21164 Evaluation Board Product Brief

EC–QCZZD–TE

Alpha 21164 Evaluation Board User’s Guide

EC–QD2UC–TE

Alpha 21164 Microprocessor Motherboard Product Brief

EC–QSAGA–TE

Alpha 21164 Microprocessor Motherboard User’s Manual

EC–QLJLB–TE

DECchip 21171 Core Logic Chipset Product Brief

EC–QC3EB–TE

DECchip 21171 Core Logic Chipset Technical Reference
Manual

EC–QE18B–TE

Answers to Common Questions about PALcode for Alpha
AXP Systems

EC–N0647–72

PALcode for Alpha Microprocessors System Design Guide

EC–QFGLB–TE

Alpha Microprocessors Evaluation Board Windows NT
3.51 Installation Guide

EC–QLUAD–TE

1 To order and purchase the Alpha Architecture Reference Manual, call 1–800–DIGITAL from
the U.S. or Canada, or contact your local Digital office, or technical or reference bookstore where
Digital Press books are distributed by Prentice Hall.

E–2 Preliminary—Subject to Change—July 1996

E.4 Ordering Associated Literature

Title

Order Number

SPICE Models for Alpha Microprocessors and Peripheral
Chips: An Application Note

EC–QA4XC–TE

Alpha Microprocessors SROM Mini-Debugger User’s
Guide

EC–QHUXA–TE

Alpha Microprocessors Evaluation Board Debug Monitor
User’s Guide

EC–QHUVB–TE

Alpha Microprocessors Evaluation Board Software Design
Tools User’s Guide

EC–QHUWA–TE

E.5 Ordering Associated Third-Party Literature
You can order the following third-party literature directly from the vendor:
Title

Vendor

PCI System Design Guide

PCI Special Interest Group
1–800–433–5177 (U.S.)
1–503–797–4207 (International)
1–503–234–6762 (FAX)

PCI Local Bus Specification
Revision 2.1

See previous entry.

IEEE Standard 754, Standard for
Binary Floating-Point Arithmetic

IEEE Service Center
445 Hoes Lane
P.O. Box 1331
Piscataway, NJ 08855–1331
1–800–678–IEEE (U.S. and Canada)
908–562–3805 (Outside U.S. and Canada)

IEEE Standard 1149.1, A Test
Access Port and Boundary Scan
Architecture

See previous entry.

Preliminary—Subject to Change—July 1996 E–3

Glossary
The glossary provides definitions for specific terms and acronyms associated
with the Alpha 21164 microprocessor and chips in general.
abort
The unit stops the operation it is performing, without saving status, to perform
some other operation.
ABT
Advanced bipolar/CMOS technology.
address space number (ASN)
An optionally implemented register used to reduce the need for invalidation
of cached address translations for process-specific addresses when a context
switch occurs. ASNs are processor specific; the hardware makes no attempt to
maintain coherency across multiple processors.
address translation
The process of mapping addresses from one address space to another.
ALIGNED
A datum of size 2**N is stored in memory at a byte address that is a multiple
of 2**N (that is, one that has N low-order zeros).
ALU
Arithmetic logic unit.
ANSI
American National Standards Institute. An organization that develops and
publishes standards for the computer industry.
ASIC
Application-specific integrated circuit.
Glossary–1

ASN
See address space number.
assert
To cause a signal to change to its logical true state.
AST
See asynchronous system trap.
asynchronous system trap (AST)
A software-simulated interrupt to a user-defined routine. ASTs enable a user
process to be notified asynchronously, with respect to that process, of the
occurrence of a specific event. If a user process has defined an AST routine
for an event, the system interrupts the process and executes the AST routine
when that event occurs. When the AST routine exits, the system resumes
execution of the process at the point where it was interrupted.
backmap
A memory unit that is used to note addresses of valid entries within a cache.
bandwidth
Bandwidth is often used to express ‘‘high rate of data transfer’’ in a bus or an
I/O channel. This usage assumes that a wide bandwidth may contain a high
frequency, which can accommodate a high rate of data transfer.
Bcache
See external cache.
barrier transaction
A transaction on the external interface as a result of an MB (memory barrier)
instruction.
BCT
Bipolar/CMOS technology.
BiCMOS
Bipolar/CMOS. The combination of bipolar and MOSFET transistors in a
common integrated circuit.

Glossary–2

bidirectional
Flowing in two directions. The buses are bidirectional; they carry both input
and output signals.
BiSr
Built-in self-repair.
BiSt
Built-in self-test.
bit
Binary digit. The smallest unit of data in a binary notation system, designated
as 0 or 1.
BIU
Bus interface unit. See Cbox.
block exchange
Memory feature that improves bus bandwidth by paralleling a cache victim
write-back with a cache miss fill.
board-level cache
See external cache.
boot
Short for bootstrap. Loading an operating system into memory is called
booting.
BSR
Boundary scan register.
buffer
An internal memory area used for temporary storage of data records during
input or output operations.
bugcheck
A software condition, usually the response to software’s detection of an
‘‘internal inconsistency,’’ which results in the execution of the system bugcheck
code.

Glossary–3

bus
A group of signals that consists of many transmission lines or wires. It
interconnects computer system components to provide communications paths
for addresses, data, and control information.
byte
Eight contiguous bits starting on an addressable byte boundary. The bits are
numbered right to left, 0 through 7.
byte granularity
Memory systems are said to have byte granularity if adjacent bytes can be
written concurrently and independently by different processes or processors.
cache
See cache memory.
cache block
The smallest unit of storage that can be allocated or manipulated in a cache.
Also known as a cache line.
cache coherence
Maintaining cache coherence requires that when a processor accesses data
cached in another processor, it must not receive incorrect data and when
cached data is modified, all other processors that access that data receive
modified data. Schemes for maintaining consistency can be implemented in
hardware or software. Also called cache consistency.
cache fill
An operation that loads an entire cache block by using multiple read cycles
from main memory.
cache flush
An operation that marks all cache blocks as invalid.
cache hit
The status returned when a logic unit probes a cache memory and finds a valid
cache entry at the probed address.

Glossary–4

cache interference
The result of an operation that adversely affects the mechanisms and
procedures used to keep frequently used items in a cache. Such interference
may cause frequently used items to be removed from a cache or incur
significant overhead operations to ensure correct results. Either action
hampers performance.
cache line
See cache block.
cache line buffer
A buffer used to store a block of cache memory.
cache memory
A small, high-speed memory placed between slower main memory and the
processor. A cache increases effective memory transfer rates and processor
speed. It contains copies of data recently used by the processor and fetches
several bytes of data from memory in anticipation that the processor will
access the next sequential series of bytes. The Alpha 21164 microprocessor
contains three onchip internal caches. See also write-through cache and
write-back cache.
cache miss
The status returned when cache memory is probed with no valid cache entry at
the probed address.
CALL_PAL Instructions
Special instructions used to invoke PALcode.
Cbox
The external interface control logic unit. Provides the 21164 microprocessor
with an interface to the external data bus, board-level Bcache, and the onchip
Scache.
central processing unit (CPU)
The unit of the computer that is responsible for interpreting and executing
instructions.

Glossary–5

CISC
Complex instruction set computing. An instruction set consisting of a large
number of complex instructions that are managed by microcode. Contrast with
RISC.
clean
In the cache of a system bus node, refers to a cache line that is valid but has
not been written.
clock
A signal used to synchronize the circuits in a computer
CMOS
Complementary metal-oxide semiconductor. A silicon device formed by a
process that combines PMOS and NMOS semiconductor material.
conditonal branch instructions
Instructions that test a register for positive/negative or for zero/non-zero. They
can also test integer registers for even/odd.
control and status register (CSR)
A device or controller register that resides in the processor’s I/O space. The
CSR initiates device activity and records its status.
CPLD
Complex programmable logic device.
CPU
See central processing unit.
CSR
See control and status register.
cycle
One clock interval.
data bus
The bus used to carry data between the 21164 and external devices. Also
called the pin bus.

Glossary–6

Dcache
Data cache. A cache reserved for storage of data. The Dcache does not contain
instructions.
DIP
Dual inline package.
direct-mapping cache
A cache organization in which only one address comparison is needed to locate
any data in the cache, because any block of main memory data can be placed
in only one possible position in the cache.
direct memory access (DMA)
Access to memory by an I/O device that does not require processor intervention.
dirty
One status item for a cache block. The cache block is valid and has been
written so that it may differ from the copy in system main memory.
dirty victim
Used in reference to a cache block in the cache of a system bus node. The
cache block is valid but is about to be replaced due to a cache block resource
conflict. The data must therefore be written to memory.
DRAM
Dynamic random-access memory. Read/write memory that must be refreshed
(read from or written to) periodically to maintain the storage of information.
DTL
Diode-transistor logic.
dual issue
Two instructions are issued, in parallel, during the same microprocessor cycle.
The instructions use different resources and so do not conflict.
EB164
An evaluation board. A hardware/software applications development
platform for the Alpha program and a debug platform for the Alpha 21164
microprocessor.

Glossary–7

Ebox
The Ebox contains the 64-bit integer execution data path.
ECC
Error correction code. Code and algorithms used by logic to facilitate error
detection and correction. See also ECC error.
ECC error
An error detected by ECC logic, to indicate that data (or the protected ‘‘entity’’
has been corrupted. The error may be correctable (soft error) or uncorrectable
(hard error).
ECL
Emitter-coupled logic.
EEPROM
Electrically erasable programmable read-only memory. A memory device that
can be byte-erased, written to, and read from. Contrast with FEPROM.
EPLD
Erasable programmable logic device.
external cache
A cache memory provided outside of the microprocessor chip, usually located
on the same module. Also called board-level or module-level cache.
Fbox
The unit within the 21164 microprocessor that performs floating-point
calculations.
FEPROM
Flash-erasable programmable read-only memory. FEPROMs can be bank- or
bulk-erased. Contrast with EEPROM.
FET
Field-effect transistor.
firmware
Machine instructions stored in hardware.

Glossary–8

floating point
A number system in which the position of the radix point is indicated by the
exponent part and another part represents the significant digits or fractional
part.
flush
See cache flush.
FPGA
Field-programmable gate array.
FPLA
Field-programmable logic array.
granularity
A characteristic of storage systems that defines the amount of data that
can be read and/or written with a single instruction, or read and/or written
independently. VAX systems have byte or multibyte granularities, whereas
disk systems typically have 512-byte or greater granularities. For a given
storage device, a higher granularity generally yields a greater throughput.
hardware interrupt request (HIR)
An interrupt generated by a peripheral device.
high-impedance state
An electrical state of high resistance to current flow, which makes the device
appear not physically connected to the circuit.
hit
See cache hit.
Ibox
A logic unit within the 21164 microprocessor that fetches, decodes, and issues
instructions. It also controls the microprocessor pipeline.

Glossary–9

Icache
Instruction cache. A cache reserved for storage of instructions. One of the
three areas of primary cache (located on the 21164) used to store instructions.
The Icache contains 8K bytes of memory space. It is a direct-mapped cache.
Icache blocks, or lines, contain 32 bytes of instruction stream data with
associated tag as well as a 6-bit ASM field and an 8-bit branch history field per
block. Icache does not contain hardware for maintaining cache coherency with
memory and is unaffected by the invalidate bus.
IEEE Standard 754
A set of formats and operations that apply to floating-point numbers. The
formats cover 32-, 64-, and 80-bit operand sizes.
IEEE Standard 1149.1
A standard for the Test Access Port and Boundary Scan Architecture used in
board-level manufacturing test procedures. Commonly referred to as the Joint
Test Action Group (JTAG) standard.
INTnn
The term INTnn, where nn is one of 2, 4, 8, 16, 32, or 64, refers to a data field
size of nn contiguous NATURALLY ALIGNED bytes. For example, INT4 refers
to a NATURALLY ALIGNED longword.
internal processor register (IPR)
One of many registers internal to the Alpha 21164 microprocessor.
IPGA
Interstitial pin grid array.
JFET
Junction field-effect transistor.
latency
The amount of time it takes the system to respond to an event.
LCC
Leadless chip carrier.
LFSR
Linear feedback shift register.

Glossary–10

load/store architecture
A characteristic of a machine architecture where data items are first loaded
into a processor register, operated on, and then stored back to memory.
No operations on memory other than load and store are provided by the
instruction set.
longword
Four contiguous bytes starting on an arbitrary byte boundary. The bits are
numbered from right to left, 0 through 31.
LSB
Least significant bit.
machine check
An operating system action triggered by certain system hardware-detected
errors that can be fatal to system operation. Once triggered, machine check
handler software analyzes the error.
MAF
Miss address file.
main memory
The large memory, external to the microprocessor, used for holding most
instruction code and data. Usually built from cost-effective DRAM memory
chips. May be used in connection with the microprocessor’s internal caches
and an optional external cache.
masked write
A write cycle that only updates a subset of a nominal data block.
MBO
See must be one.
Mbox
This section of the processor unit performs address translation, interfaces to
the Dcache, and performs several other functions.
MBZ
See must be zero.

Glossary–11

MESI protocol
A cache consistency protocol with full support for multiprocessing. The MESI
protocol consists of four states that define whether a block is modified (M),
exclusive (E), shared (S), or invalid (I).
MIPS
Millions of instructions per second.
miss
See cache miss.
module
A board on which logic devices (such as transistors, resistors, and memory
chips) are mounted and connected to perform a specific system function.
module-level cache
See external cache.
MOS
Metal-oxide semiconductor.
MOSFET
Metal-oxide semiconductor field-effect transistor.
MSI
Medium-scale integration.
multiprocessing
A processing method that replicates the sequential computer and interconnects
the collection so that each processor can execute the same or a different
program at the same time.
Must be one (MBO)
A field that must be supplied as one.
Must be zero (MBZ)
A field that is reserved and must be supplied as zero. If examined, it must be
assumed to be UNDEFINED.
NATURALLY ALIGNED
See ALIGNED.
Glossary–12

NATURALLY ALIGNED data
Data stored in memory such that the address of the data is evenly divisible by
the size of the data in bytes. For example, an ALIGNED longword is stored
such that the address of the longword is evenly divisible by 4.
NMOS
N-type metal-oxide semiconductor.
NVRAM
Nonvolatile random-access memory.
OBL
Observability linear feedback shift register.
octaword
Sixteen contiguous bytes starting on an arbitrary byte boundary. The bits are
numbered from right to left, 0 through 127.
OpenVMS Alpha operating system
Digital’s open version of the VMS operating system, which runs on Alpha
platforms.
operand
The data or register upon which an operation is performed.
PAL
Privileged architecture library. See also PALcode. Also Programmable array
logic (hardware). A device that can be programmed by a process that blows
individual fuses to create a circuit.
PALcode
Alpha privileged architecture library code, written to support Alpha
microprocessors. PALcode implements architecturally defined behavior.
PALmode
A special environment for running PALcode routines.
parameter
A variable that is given a specific value that is passed to a program before
execution.

Glossary–13

parity
A method for checking the accuracy of data by calculating the sum of the
number of ones in a piece of binary data. Even parity requires the correct
sum to be an even number, odd parity requires the correct sum to be an odd
number.
PGA
Pin grid array.
pipeline
A CPU design technique whereby multiple instructions are simultaneously
overlapped in execution.
PLA
Programmable logic array.
PLCC
Plastic leadless chip carrier or plastic-leaded chip carrier.
PLD
Programmable logic device.
PLL
Phase-locked loop.
PMOS
P-type metal-oxide semiconductor.
PQFP
Plastic quad flat pack.
primary cache
The cache that is the fastest and closest to the processor. The first-level caches,
located on the CPU chip, composed of the Dcache, Icache, and Scache.
program counter
That portion of the CPU that contains the virtual address of the next
instruction to be executed. Most current CPUs implement the program counter
(PC) as a register. This register may be visible to the programmer through the
instruction set.

Glossary–14

PROM
Programmable read-only memory.
pull-down resistor
A resistor placed between a signal line and a negative voltage.
pull-up resistor
A resistor placed between a signal line to a positive voltage.
quad issue
Four instructions are issued, in parallel, during the same microprocessor cycle.
The instructions use different resources and so do not conflict.
quadword
Eight contiguous bytes starting on an arbitrary byte boundary. The bits are
numbered from right to left, 0 through 63.
RAM
Random-access memory.
READ_BLOCK
A transaction where the 21164 requests that an external logic unit fetch read
data.
read data wrapping
System feature that reduces apparent memory latency by allowing read data
cycles to differ the usual low-to-high sequence. Requires cooperation between
the 21164 and external hardware.
read stream buffers
Arrangement whereby each memory module independently prefetches DRAM
data prior to an actual read request for that data. Reduces average memory
latency while improving total memory bandwidth.
register
A temporary storage or control location in hardware logic.
reliability
The probability a device or system will not fail to perform its intended
functions during a specified time interval when operated under stated
conditions.

Glossary–15

reset
An action that causes a logic unit to interrupt the task it is performing and go
to its’ initialized state.
RISC
Reduced instruction set computing. A computer with an instruction set that
is paired down and reduced in complexity so that most can be performed in
a single processor cycle. High-level compilers synthesize the more complex,
least frequently used instructions by breaking them down into simpler
instructions. This approach allows the RISC architecture to implement
a small, hardware-assisted instruction set, thus eliminating the need for
microcode.
ROM
Read-only memory.
RTL
Register-transfer logic.
SAM
Serial access memory.
SBO
Should be one.
SBZ
Should be zero.
Scache
Secondary cache. A 3-way set-associative, second-level cache located on the
Alpha 21164 microprocessor.
scheduling
The process of ordering instruction execution to obtain optimum performance.

Glossary–16

set-associative
A form of cache organization in which the location of a data block in main
memory constrains, but does not completely determine, its location in the
cache. Set-associative organization is a compromise between direct-mapped
organization, in which data from a given address in main memory has only
one possible cache location, and fully associative organization, in which data
from anywhere in main memory can be put anywhere in the cache. An ‘‘n-way
set-associative’’ cache allows data from a given address in main memory to be
cached in any of n locations. The Scache in the 21164 microprocessor has a
3-way set-associative organization.
SIMM
Single inline memory module.
SIP
Single inline package.
SIPP
Single inline pin package.
SMD
Surface mount device.
SRAM
Static random-access memory.
SROM
Serial read-only memory.
SSI
Small-scale integration.
SSRAM
Synchronous static random-access memory.
stack
An area of memory set aside for temporary data storage or for procedure and
interrupt service linkages. A stack uses the last-in/first-out concept. As items
are added to (pushed on) the stack, the stack pointer decrements. As items are
retrieved from (popped off) the stack, the stack pointer increments.

Glossary–17

STRAM
Self-timed random-access memory.
superpipelined
Describes a pipelined machine that has a larger number of pipe stages and
more complex scheduling and control. See also pipeline.
superscalar
Describes a machine architecture that allows multiple independent instructions
to be issued in parallel during a given clock cycle.
tag
The part of a cache block that holds the address information used to determine
if a memory operation is a hit or a miss on that cache block.
TB
Translation buffer.
tristate
Refers to a bused line that has three states: high, low, and high-impedance.
TTL
Transistor–transistor logic.
UART
Universal asynchronous receiver-transmitter.
UNALIGNED
A datum of size 2**N stored at a byte address that is not a multiple of 2**N.
unconditional branch instructions
Instructions that write a return address into a register.
UNDEFINED
An operation that may halt the processor or cause it to lose information. Only
privileged software (that is, software running in kernel mode) can trigger an
UNDEFINED operation.

Glossary–18

UNPREDICTABLE
Results or occurrences that do not disrupt the basic operation of the processor;
the processor continues to execute instructions in its normal manner.
Privileged or unprivileged software can trigger UNPREDICTABLE results or
occurrences.
UVPROM
Ultraviolet (erasable) programmable read-only memory.
valid
Allocated. Valid cache blocks have been loaded with data and may return
cache hits when accessed.
victim
Used in reference to a cache block in the cache of a system bus node. The
cache block is valid but is about to be replaced due to a cache block resource
conflict.
virtual cache
A cache that is addressed with virtual addresses. The tag of the cache is a
virtual address. This process allows direct addressing of the cache without
having to go through the translation buffer making cache hit times faster.
VHSIC
Very-high-speed integrated circuit.
VLSI
Very-large-scale integration.
VRAM
Video random-access memory.
word
Two contiguous bytes (16 bits) starting on an arbitrary byte boundary. The bits
are numbered from right to left, 0 through 15.
write data wrapping
System feature that reduces apparent memory latency by allowing write data
cycles to differ the usual low-to-high sequence. Requires cooperation between
the 21164 and external hardware.

Glossary–19

write-back
A cache management technique in which write operation data is written into
cache but is not written into main memory in the same operation. This may
result in temporary differences between cache data and main memory data.
Some logic unit must maintain coherency between cache and main memory.
write-back cache
Copies are kept of any data in the region; read and write operations may use
the copies, and write operations use additional state to determine whether
there are other copies to invalidate or update.
write-through
A cache management technique in which a write operation to cache also causes
the same data to be written in main memory during the same operation.
write-through cache
Copies are kept of any data in the region; read operations may use the copies,
but write operations update the actual data location and either update or
invalidate all copies.
WRITE_BLOCK
A transaction where the 21164 requests that an external logic unit process
write data.

Glossary–20

Index
A
Aborts, 2–18
Absolute Maximum Rating, 9–1
ac coupling, 9–6
Addressing, 1–2
Address regions, physical, 4–12
Address translation, 2–10
addr_bus_req_h
description, 3–3
operation, 4–44, 4–53, 4–70, 5–81, 7–4,
9–13
addr_cmd_par_h
description, 3–3
operation, 3–3, 4–70, 4–71, 4–95, 7–4,
9–19, 9–20
addr_h<39:4>
description, 3–3
operation, 3–3, 4–13, 4–14, 4–15, 4–16,
4–39, 4–52, 4–70, 4–71, 4–75, 4–95,
7–3, 9–12, 9–13, 9–15
addr_res_h<2:0>
description, 3–4
operation, 4–56, 4–57, 4–58, 4–64, 4–65,
7–4, 9–19
Alpha documentation, E–2
ALT_MODE register, 5–60
Architecture, 1–1 to 1–3
Associated literature, E–2
AST, 2–9
ASTER register, 5–26
ASTRR register, 5–25

Asynchronous system trap
See AST

B
Bcache, 2–14
block size, 4–15
errors, 4–92
hit under READ MISS example, 4–90
interface, 4–4
introduction, 4–2 to 4–4
selecting options, 4–35
structure, 4–15
systems without, 4–19, 4–81
timing, 4–31
victim buffers, 4–18
Bcache read transaction
private read operation, 4–32
BCACHE VICTIM command, 4–38
Bcache write transaction
private write operation, 4–34
BC_CONFIG register, 5–84
BC_CONTROL register, 5–78
BC_TAG_ADDR register, 5–89
BIU, 4–2, 4–14, 4–30, 4–31, 4–44, 4–53,
4–83
See also Cbox
buffer, 4–4
Block diagram, 21164, 2–2
Boundaries
data wrap order, 4–13
Boundary scan register, 12–7

Index–1

Branch prediction, 2–5, 2–20
Bubble cycle, 2–32
Bubble squashing, 2–20
Bus contention
command/address bus, 4–70 to 4–80
data bus, 4–70 to 4–80
Bus interface unit
See BIU

C
Cache coherency, 4–19 to 4–29
basics, 4–19
flush protocol, 4–21
state machines, 4–27
systems, 4–25
transaction conflicts, 4–28
write invalidate protocol, 4–21
state machines, 4–24
states, 4–23
systems, 4–22
Cache control and bus interface unit
See Cbox
Cache organization, 2–13
cack_h
description, 3–4
operation, 3–4, 4–31, 4–36, 4–37, 4–40,
4–41, 4–44, 4–46, 4–48, 4–50, 4–52,
4–81, 4–82, 4–83, 4–84, 4–86, 4–90,
4–95, 5–13, 5–21, 5–81, 7–4, 8–10,
8–11, 9–13, 9–15, 9–16, 12–7
Cbox, 2–12
IPR PALcode restrictions, 5–100
IPRs, 5–68 to 5–98
read requests, 2–31
write buffer data store, 2–35
write ordering, 2–37
CC register, 5–61
CC_CTL register, 5–62
cfail_h
description, 3–4
operation, 4–31, 4–38, 4–82, 4–95, 5–13,
5–21, 7–4, 8–10, 8–11, 9–18, 12–7

Index–2

clk_mode_h<1:0>
description, 3–4
operation, 4–5, 7–3, 9–18, 9–25
Clocks, 4–5 to 4–12
CPU, 4–5
reference, 4–8, 4–9
system, 4–6
cmd_h<3:0>
description, 3–4
operation, 3–3, 4–37, 4–41, 4–48, 4–53,
4–55, 4–64, 4–71, 4–75, 4–83, 4–95,
7–4, 9–12, 9–19, 9–20
Coherency, caches, 4–19
Command/address
driving bus, 4–70
errors, 4–92
Commands
21164 initiated, 4–37
BCACHE VICTIM, 4–38
FETCH, 4–37
FETCH_M, 4–37
FLUSH, 4–64
INVALIDATE, 4–55
LOCK, 4–37
MEMORY BARRIER, 4–37
NOP, 4–37, 4–55, 4–64
READ, 4–64
READ DIRTY, 4–55
READ DIRTY/INVALIDATE, 4–56
READ MISS0, 4–38
READ MISS1, 4–38
READ MISS MOD0, 4–38
READ MISS MOD1, 4–38
READ MISS MOD STC0, 4–39
READ MISS MOD STC1, 4–39
SET DIRTY, 4–37
SET SHARED, 4–55
WRITE BLOCK, 4–37
WRITE BLOCK LOCK, 4–38
Commands, sending to 21164, 4–53
Conventions, xxii to xxvii
CPU
clock, 4–5
microarchitecture, 2–2

cpu_clk_out_h
description, 3–6
operation, 3–11, 4–5, 7–3, 9–4

D
dack_h
description, 3–6
operation, 3–7, 4–31, 4–36, 4–37, 4–39,
4–41, 4–43, 4–44, 4–46, 4–48, 4–58,
4–66, 4–76, 4–77, 4–78, 4–80, 4–81,
4–82, 4–83, 4–84, 4–86, 4–90, 4–95,
5–81, 7–4, 8–9, 9–13, 9–15, 9–16
Data cache
See Dcache
Data integrity, 4–92
address and command parity, 4–95
Bcache tag control parity, 4–94
Bcache tag data parity, 4–94
ECC and parity, 4–92
force correction, 4–94
Data translation buffer
See DTB
Data types, 1–1
floating-point, 1–3, 2–10
integer, 1–2
Data wrap order, 4–13
data_bus_req_h
description, 3–6
operation, 4–43, 4–72, 4–74, 4–80, 7–4,
9–13, 9–15, 9–16
data_check_h<15:0>
description, 3–7
operation, 4–70, 4–92, 7–3, 9–19, 9–20
data_h<127:0>
description, 3–6
operation, 3–8, 4–43, 4–46, 4–55, 4–64,
4–70, 4–71, 4–75, 4–92, 4–93, 7–3,
9–11, 9–12, 9–13, 9–15
data_ram_oe_h
description, 3–7
operation, 4–32, 4–43, 4–76, 4–77, 4–78,
4–79, 4–80, 7–3, 9–21

data_ram_we_h
description, 3–7
operation, 4–34, 7–3, 9–21
Dcache, 2–13
control, 2–12
DC_FLUSH register, 5–60
DC_MODE register, 5–56
dc_ok_h
description, 3–7
operation, 3–11, 4–5, 7–1, 7–2, 7–3, 7–5,
7–13, 9–4, 9–5, 9–18, 12–2, 12–3
DC_PERR_STAT register, 5–50
DC_TEST_CTL register, 5–63
DC_TEST_TAG register, 5–64
DC_TEST_TAG_TEMP register, 5–66
Decoupling, 9–26
Delayed system clock, 4–8
Design examples, 2–40
Documentation, E–2
DTB, 2–11
DTB_ASN register, 5–38
DTB_CM register, 5–39
DTB_IAP register, 5–52
DTB_IA register, 5–52
DTB_IS register, 5–53
DTB_PTE register, 5–41
DTB_PTE_TEMP register, 5–43
DTB_TAG register, 5–40
Duplicate tag store, 4–15
algorithm, 4–17
full, 4–16
partial Scache, 4–18

E
Ebox, 2–9
registers, 2–9, 5–99
ECC, 4–92 to 4–94
EI_ADDR register, 5–94
EI_STAT register, 5–91
Entry-pointer queues, 2–36
Environment instructions
PALcode, 6–7

Index–3

Error correction code
See ECC
Exceptions, 2–18
EXC_ADDR register, 5–14
EXC_MASK register, 5–17
EXC_SUM register, 5–15
External cache
See Bcache
External interface
rules for use, 4–83
External interface introduction, 4–2 to 4–4

F
Features, 1–3 to 1–4
FETCH command, 4–37, 4–52
FETCH_M command, 4–37, 4–52
Fill, 2–32
FILL, after other transactions, 4–81
FILL error, 4–95
FILL transaction, 4–43
fill_error_h
description, 3–7
operation, 4–43, 4–95, 7–4, 8–9, 8–12,
9–18
fill_h
description, 3–7
operation, 3–7, 4–36, 4–43, 4–72, 4–73,
4–78, 4–81, 4–83, 4–95, 7–4, 8–9,
9–18
fill_id_h
description, 3–7
operation, 3–7, 4–41, 4–43, 4–95, 7–4,
8–9, 9–18
fill_nocheck_h
description, 3–7
operation, 7–4, 9–18
FILL_SYN register, 5–95
Floating data types, 2–10
Floating-point unit
See FPU
FLUSH command, 4–64
Flush protocol, 4–21, 4–25, 4–26, 4–27
commands, 4–64
state machines, 4–27

Index–4

FLUSH timing diagram, 4–66
FLUSH transaction, 4–66
FPU, 2–10
Free-entry queue, 2–36

H
Heat sink, 10–3
Hint bits, 2–11
HWINT_CLR register, 5–28
HW_LD instruction, 6–3
HW_MFPR instruction, 6–3
HW_MTPR instruction, 6–3
HW_REI instruction, 6–3
HW_ST instruction, 6–3

I
Ibox, 2–2, 2–4
branch prediction, 2–5
instruction
decode, 2–4
issue, 2–4
instruction translation buffer, 2–7
interrupts, 2–8
IPRs, 5–5 to 5–37
encoding, 5–2
slotting, 2–22
Icache, 2–13
ICM register, 5–19
ICPERR_STAT register, 5–13
ICSR register, 5–20
IC_FLUSH_CTL register, 5–13
idle_bc_h
description, 3–8
operation, 3–6, 4–19, 4–43, 4–44, 4–72,
4–73, 4–74, 4–75, 4–80, 4–83, 4–84,
4–86, 4–88, 7–4, 9–18
IEEE floating-point conformance, A–12
IFAULT_VA_FORM register, 5–11
index_h<25:4>
description, 3–8
operation, 3–11, 4–4, 4–15, 4–72, 4–73,
4–89, 7–3, 9–11

Initialization
role of interrupt signals, 4–96
Input clock
ac coupling, 9–6
impedance levels, 9–6
termination, 9–6
Input clocks, 9–4
Instruction
decode, 2–4
issue, 2–4
Instruction cache
See Icache
Instruction fetch/decode unit and branch unit
See Ibox
Instruction issue, 1–3, 2–18
Instructions
classes, 2–20
issue rules, 2–28
latencies, 2–24
MB, 2–12
slotting, 2–20, 2–22
WMB, 2–12, 2–35
Instruction translation buffer, 2–7
See ITB
int4_valid_h<3:0>
description, 3–8
operation, 4–14, 4–41, 4–48, 7–4, 9–19
Integer execution unit
See Ebox
Integer register file
See IRF
Interface restrictions, 4–81
Interface transactions
21164 initiated, 4–36 to 4–52
system initiated, 4–53 to 4–69
Internal processor registers
See IPRs
Interrupts, 4–96 to 4–98
ASTs, 2–9
disabling, 2–9
hardware, 2–8
initialization, 4–96
normal operation, 4–96
priority level, 4–96

Interrupts (cont’d)
software, 2–8
Interrupt signals, 4–96
INTID register, 5–24
INTnn, xxiv
INVALIDATE command, 4–55
INVALIDATE timing diagram, 4–60
INVALIDATE transaction, 4–60
IPLR register, 5–23
IPRs
accessibility, 5–1
ALT_MODE, 5–60
ASTER, 5–26
ASTRR, 5–25
BC_CONFIG, 5–84
BC_CONTROL, 5–78
BC_TAG_ADDR, 5–89
CC, 5–61
CC_CTL, 5–62
DC_FLUSH, 5–60
DC_MODE, 5–56
DC_PERR_STAT, 5–50
DC_TEST_CTL, 5–63
DC_TEST_TAG, 5–64
DC_TEST_TAG_TEMP, 5–66
DTB_ASN, 5–38
DTB_CM, 5–39
DTB_IA, 5–52
DTB_IAP, 5–52
DTB_IS, 5–53
DTB_PTE, 5–41
DTB_PTE_TEMP, 5–43
DTB_TAG, 5–40
EI_ADDR, 5–94
EI_STAT, 5–91
EXC_ADDR, 2–19, 5–14
EXC_MASK, 5–17
EXC_SUM, 5–15
FILL_SYN, 5–95
HWINT_CLR, 5–28
ICM, 5–19
ICPERR_STAT, 5–13
ICSR, 2–9, 5–20
IC_FLUSH_CTL, 5–13
IFAULT_VA_FORM, 5–11

Index–5

IPRs (cont’d)
INTID, 5–24
IPLR, 2–9, 5–23
ISR, 5–29
ITB_ASN, 5–8
ITB_IA, 5–9
ITB_IAP, 5–9
ITB_IS, 5–10
ITB_PTE, 5–6
ITB_PTE_TEMP, 5–9
ITB_TAG, 5–5
IVPTBR, 5–12
MAF_MODE, 5–58
MCSR, 5–54
MM_STAT, 5–44
MVPTBR, 5–49
PAL_BASE, 5–18, 6–3
PMCTR, 5–33
reset state, 7–10
SC_ADDR, 5–75
SC_CTL, 5–69
SC_STAT, 5–72
SIRR, 5–27
SL_RCV, 5–32
SL_XMIT, 5–31
VA, 5–46
VA_FORM, 5–47
IRF, 2–9
irq_h<3:0>
description, 3–9
operation, 2–8, 2–9, 4–6, 4–9, 4–97, 5–30,
7–4, 9–18
ISR register, 5–29
Issue rules, 2–28
Issuing rules, 2–20 to 2–29
ITB, 2–7
ITB_ASN register, 5–8
ITB_IAP register, 5–9
ITB_IA register, 5–9
ITB_IS register, 5–10
ITB_PTE register, 5–6
ITB_PTE_TEMP register, 5–9
ITB_TAG register, 5–5

Index–6

IVPTBR register, 5–12

L
Latencies, 2–24
Literature, E–2
Live lock
cache conflict, 4–28
Load-after-store trap, 2–29
Load instructions
noncacheable space, 2–31
Load miss, 2–30
LOCK command, 4–37
Lock mechanisms, 4–30
LOCK timing diagram, 4–50
LOCK transaction, 4–50
Logic symbol, 3–1

M
MAF, 2–11, 2–30 to 2–33, 4–14
entries, 2–32
entry, 2–33
rules, 2–30
MAF_MODE register, 5–58
MB instruction, 2–12, 4–52
Mbox, 2–2, 2–10
address translation, 2–10
data translation buffer, 2–11
IPRs, 5–38 to 5–67
encoding, 5–4
load instruction, 2–11
miss address file, 2–11
store execution, 2–33 to 2–34
store instructions, 2–12
write buffer, 2–12
write buffer address file, 2–35
mch_hlt_irq_h
description, 3–9
operation, 2–8, 4–8, 4–97, 7–4, 9–18
MCSR register, 5–54
Memory address translation unit
See Mbox

MEMORY BARRIER command, 4–37
when to use, 4–52
Memory regions, physical, 4–12
Merge
write buffer, 4–14
Merging
loads to noncacheable space, 2–31
rules, 2–30
Microarchitecture, 2–2 to 2–14
Miss address file
See MAF
MM_STAT register, 5–44
Multiple instruction issue, 2–4
MVPTBR register, 5–49

N
Noncached read operations, 4–14
Noncached write operations, 4–14
Nonissue conditions, 2–20
NOP command, 4–37, 4–55, 4–64

O
Operating temperature, 10–1
Ordering products, E–1
osc_clk_in_h,l
description, 3–9
operation, 3–4, 4–5, 4–11, 7–3, 7–5, 9–2,
9–4, 9–5, 9–6, 9–17, 9–25, 12–3

P
Page table entry
See PTE
PAL
restrictions, 5–101
PALcode, 1–2
environment instructions, 6–7
invoke, 6–3
PALmode, 6–2
environment, 6–2
PALshadow registers, 5–99

PALtemp IPRs, 5–99
encoding, 5–3
PAL_BASE register, 5–18, 6–3
Parity, 4–92
Parts
ordering, E–1
Pending-request queue, 2–36
Performance counters, 2–38
perf_mon_h
description, 3–10
operation, 2–38, 5–36, 7–5, 9–18
Physical address considerations, 4–12
Physical address regions, 4–12
Physical memory regions, 4–12
Pipeline, wave, 4–33
Pipeline organization, 2–14 to 2–20
Pipelines, 2–9
bubbles, 2–20
examples, 2–14
floating add, 2–16
integer add, 2–16
load (Dcache hit), 2–16
load (Dcache miss), 2–17
store (Dcache hit), 2–17
instruction issue, 2–18
stages, 2–14, 2–18
stall, 2–18, 2–20
PMCTR register, 5–33
port_mode_h<1:0>
description, 3–10
operation, 7–5, 9–18, 12–1, 12–2
Power supply
considerations, 9–26
decoupling, 9–26
sequencing, 9–27
Private Bcache transactions
21164 to Bcache, 4–31 to 4–35
Privileged architecture library code
See PALcode
Producer–consumer dependencies, 2–24
Producer–producer dependencies, 2–24
Producer–producer latency, 2–27
PTE, 2–8, 2–11

Index–7

pwr_fail_irq_h
description, 3–10
operation, 2–8, 4–8, 4–97, 7–4, 9–18

Q
Queues
entry-pointer, 2–36

R
Race condition
21164 and system, 4–83
Race example
idle_bc_h and cack_h, 4–86
READ command, 4–64
READ DIRTY command, 4–55
READ DIRTY/INVALIDATE command, 4–56
READ DIRTY/INVALIDATE transaction,
4–58
READ DIRTY timing diagram, 4–58
READ DIRTY transaction, 4–58
READ MISS0 command, 4–38
READ MISS1 command, 4–38
READ MISS MOD0 command, 4–38
READ MISS MOD1 command, 4–38
READ MISS MOD STC0 command, 4–39
READ MISS MOD STC1 command, 4–39
READ MISS no Bcache timing diagram,
4–40
READ MISS timing diagram, 4–41
READ MISS transaction, 4–41
READ MISS transaction (no Bcache), 4–40
READ MISS with idle_bc_h asserted
example, 4–88
READ MISS with victim abort example,
4–89
READ MISS with victim example, 4–84
READ MISS with victim timing diagram,
4–45, 4–46
READ MISS with victim transaction, 4–43
READ timing diagram, 4–68
READ transaction, 4–68

Index–8

Read/write spacing
data bus contention, 4–71
Reference clock, 4–8, 4–9
example 1, 4–10
example 2, 4–11
examples, 4–9
ref_clk_in_h
description, 3–10
operation, 4–5, 4–8, 4–9, 4–10, 4–11,
4–12, 7–3, 9–4, 9–12, 9–15, 9–17,
9–18, 9–25
Registers
See also IPRs
accessibility, 5–1
integer, 2–9
PALshadow, 2–9, 5–99
PALtemp, 5–99
Related documentation, E–2
Replay traps, 2–29 to 2–30
as aborts, 2–19
load instruction, 2–12, 2–33
load-miss-and-use, 2–19
Reset
forcing, 4–95
Resource conflict, 2–20
Restrictions
interface, 4–81

S
Scache, 2–13
block size, 4–15
scache_set_h<1:0>
description, 3–10
operation, 4–16, 4–18, 7–4, 9–19
Scheduling rules, 2–20 to 2–29
SC_ADDR register, 5–75
SC_CTL register, 5–69
SC_STAT register, 5–72
Second-level cache
See Scache
Semiconductor Information Line, E–1
Serial read-only memory
See SROM

SET DIRTY command, 4–37
SET DIRTY timing diagram, 4–50
SET DIRTY transaction, 4–50
SET SHARED command, 4–55
SET SHARED timing diagram, 4–62
SET SHARED transaction, 4–62
shared_h
description, 3–10
operation, 7–4, 9–18
Signal descriptions, 3–3 to 3–15
SIRR register, 5–27
Slotting, 2–22
SL_RCV register, 5–32
SL_XMIT register, 5–31
Specifications
mechanical, 11–1
SROM, 2–14
srom_clk_h
description, 3–10
operation, 5–31, 7–5, 7–6, 9–19, 9–23,
9–24, 12–1
srom_data_h
description, 3–10
operation, 5–32, 7–5, 7–6, 7–7, 9–18,
9–24, 12–1
srom_oe_l
description, 3–10
operation, 7–5, 7–6, 9–19, 12–1
srom_present_l
description, 3–10
operation, 7–5, 7–6, 9–18, 9–22, 9–23,
12–1
Store instruction, 2–12
execution, 2–33
st_clk_h
description, 3–11
Superpages, 2–8
System clock, 4–6
delayed, 4–8
System clock delay, 4–8
System interface, 4–2
addresses, 4–4
commands, 4–4

System interface introduction, 4–2 to 4–4
system_lock_flag_h
description, 3–11
operation, 4–30, 7–4, 9–18
sys_clk_out1_h,l
description, 3–11
operation, 3–9, 3–11, 4–2, 4–5, 4–6, 4–8,
4–9, 4–11, 4–57, 4–65, 5–87, 7–3, 9–4,
9–12, 9–13, 9–15, 9–17, 9–25
sys_clk_out2_h,l
description, 3–11
operation, 3–9, 3–10, 3–11, 4–5, 5–87,
7–3, 7–5, 9–5
sys_mch_chk_irq_h
description, 3–11
operation, 2–8, 4–8, 4–97, 7–4, 9–18
sys_reset_l
description, 3–11
operation, 4–96, 7–1, 7–2, 7–3, 7–5, 7–6,
7–13, 9–18, 9–21, 9–22

T
Tag store, duplicate, 4–15
tag_ctl_par_h
description, 3–11
operation, 4–79, 4–94, 7–3, 9–19, 9–20
tag_data_h<38:20>
description, 3–11
operation, 4–15, 4–19, 4–75, 4–94, 7–3,
9–21
tag_data_par_h
description, 3–11
operation, 4–19, 4–43, 4–94, 7–3, 9–21
tag_dirty_h
description, 3–11
operation, 3–11, 4–19, 4–41, 4–43, 4–79,
4–94, 7–3, 9–19, 9–20
tag_ram_oe_h
description, 3–11
operation, 4–32, 4–79, 7–3, 9–21
tag_ram_we_h
description, 3–12
operation, 4–34, 7–3, 9–21

Index–9

tag_shared_h
description, 3–12
operation, 3–11, 4–19, 4–43, 4–57, 4–65,
4–79, 4–94, 7–3, 9–19, 9–20
tag_valid_h
description, 3–12
operation, 3–11, 4–19, 4–43, 4–94, 7–3,
9–20, 9–21
tck_h
description, 3–12
operation, 7–5, 9–26, 12–1, 12–2
tdi_h
description, 3–12
operation, 7–5, 9–4, 9–26, 12–1, 12–2,
12–7
tdo_h
description, 3–12
operation, 7–5, 9–26, 12–1, 12–2, 12–7
Technical support, E–1
Temperature, 10–1
temp_sense
description, 3–12
operation, 7–5, 9–4
Terminology, xxii to xxvii
test_status_h<1:0>
description, 3–12
operation, 5–22, 7–5, 7–6, 9–21, 12–1,
12–6, 12–7
Thermal design considerations, 10–4
Thermal heat sink, 10–3
Thermal management, 10–1
Thermal operating temperature, 10–1
Timing diagrams
Bcache hit under READ MISS, 4–90
Bcache read, 4–32
Bcache write, 4–34
bus contention, 4–70
FILL, 4–78, 4–79
FILL to private read or write, 4–80
FLUSH, 4–66
idle_bc_h and cack_h race, 4–86
INVALIDATE, 4–60
LOCK, 4–50
READ, 4–68
READ DIRTY, 4–58

Index–10

Timing diagrams (cont’d)
READ MISS, 4–41
READ MISS completed first—victim
buffer, 4–76
READ MISS—no Bcache, 4–40
READ MISS second—no victim buffer,
4–77
READ MISS with idle_bc_h asserted,
4–88
READ MISS with victim, 4–45, 4–46,
4–84
READ MISS with victim abort, 4–89
SET DIRTY, 4–50
SET SHARED, 4–62
using data_bus_req_h, 4–74
using idle_bc_h and fill_h, 4–73
wave pipeline, 4–33
WRITE BLOCK, 4–49
tms_h
description, 3–12
operation, 7–5, 9–4, 9–26, 12–1, 12–2,
12–4
Transactions
FILL, 4–43
FLUSH, 4–66
INVALIDATE, 4–60
LOCK, 4–50
READ, 4–68
READ DIRTY, 4–58
READ DIRTY/INVALIDATE, 4–58
READ MISS, 4–41
READ MISS (no Bcache), 4–40
READ MISS with victim, 4–43
SET DIRTY, 4–50
SET SHARED, 4–62
system initiated, 4–53
WRITE BLOCK, 4–48
WRITE BLOCK LOCK, 4–48
Traps
load-after-store, 2–29
load-miss-and-use, 2–28
replay, 2–19, 2–29, 2–33
Tristate
BCACHE VICTIM to fill, 4–75
FILL to private Bcache read or write,
4–80

Tristate (cont’d)
overlap, 4–70, 4–75
READ or WRITE to fill, 4–75
system Bcache command to fill, 4–78
trst_l
description, 3–12
operation, 7–5, 7–13, 9–26, 12–1, 12–2,
12–3

V
VA register, 5–46
VA_FORM register, 5–47
Victim buffers, 4–18, 4–44
victim_pending_h
description, 3–12
operation, 4–16, 4–18, 4–38, 4–44, 7–4,
9–19

W
Wave pipeline, 4–33
WMB instruction, 2–12, 2–35
Write-after-write conflicts
See Producer–producer dependencies
See Producer–producer latency
WRITE BLOCK command, 4–37
WRITE BLOCK command acknowledge,
4–81
WRITE BLOCK LOCK command, 4–38
WRITE BLOCK LOCK restriction, 4–82
WRITE BLOCK LOCK transaction, 4–48
WRITE BLOCK timing diagram, 4–49
WRITE BLOCK transaction, 4–48
Write buffer, 2–12, 2–35 to 2–37
entry processing, 2–36
Write invalidate protocol, 4–21, 4–22
commands, 4–55
states, 4–23
systems, 4–22
Write ordering, 2–37

Index–11