Инструкции для AMD Typewriter x86

Содержание

• Contents

  3
• Revision History

  15
• Introduction

  17
  • About this Document

    17
    • Chapter 1: Introduction

      18
    • Chapter 2: Top Optimizations

      18
    • Chapter 3: C Source Level Optimizations

      18
    • Chapter 4: Instruction Decoding Optimizations

      18
    • Chapter 5: Cache and Memory Optimizations

      18
    • Chapter 6: Branch Optimizations

      18
    • Chapter 7: Scheduling Optimizations

      18
    • Chapter 8: Integer Optimizations

      18
    • Chapter 9: Floating-Point Optimizations

      18
    • Chapter 10: 3DNow!™ and MMX™ Optimizations

      18
    • Chapter 11: General x86 Optimizations Guidelines

      18
    • Appendix A: AMDAthlon Processor Microarchitecture

      18
    • Appendix B: Pipeline and Execution Unit Resources Overview

      19
    • Appendix C: Implementation of Write Combining

      19
    • Appendix D: Performance Monitoring Counters

      19
    • Appendix E: Programming the MTRR and PAT

      19
    • Appendix F: Instruction Dispatch and Execution Resources

      19
    • Appendix G: DirectPath versus VectorPath Instructions

      19
  • AMDAthlon™ Processor Family

    19
  • AMDAthlon™ Processor Microarchitecture Summary

    20
• Top Optimizations

  23
  • Group I — Essential Optimizations

    23
  • Group II — Secondary Optimizations

    23
  • Optimization Star

    24
  • Group I Optimizations — Essential Optimizations

    24
    • Memory Size and Alignment Issues

      24
      • Avoid Memory Size Mismatches

        24
      • Align Data Where Possible

        24
    • Use the 3DNow!™ PREFETCH and PREFETCHW Instructions

      24
    • Select DirectPath Over VectorPath Instructions

      25
  • Group II Optimizations—Secondary Optimizations

    25
    • LoadExecute Instruction Usage

      25
      • Use LoadExecute Instructions

        25
      • Avoid LoadExecute FloatingPoint Instructions with Integer Operands

        26
    • Take Advantage of Write Combining

      26
    • Use 3DNow!™ Instructions

      26
    • Avoid Branches Dependent on Random Data

      26
    • Avoid Placing Code and Data in the Same 64Byte Cache Line

      27
• C Source Level Optimizations

  29
  • Ensure FloatingPoint Variables and Expressions are of Type Float

    29
  • Use 32Bit Data Types for Integer Code

    29
  • Consider the Sign of Integer Operands

    30
    • Example 1 (Avoid):

      30
    • Example (Preferred):

      30
    • Example (Avoid):

      31
    • Example (Preferred):

      31
    • Use unsigned types for:

      31
    • Use signed types for:

      31
  • Use Array Style Instead of Pointer Style Code

    31
    • Example 1 (Avoid):

      32
    • Example 2 (Preferred):

      33
  • Completely Unroll Small Loops

    34
    • Example 1 (Avoid):

      34
    • Example 2 (Preferred):

      34
  • Avoid Unnecessary Store-to-Load Dependencies

    34
    • Example 1 (Avoid):

      35
    • Example 2 (Preferred):

      35
  • Consider Expression Order in Compound Branch Conditions

    36
  • Switch Statement Usage

    37
    • Optimize Switch Statements

      37
      • Example 1 (Avoid):

        37
      • Example 2 (Preferred):

        37
  • Use Prototypes for All Functions

    37
  • Use Const Type Qualifier

    38
  • Generic Loop Hoisting

    38
    • Example 1:

      38
    • Generalization for Multiple Constant Control Code

      39
      • Example 2:

        39
  • Declare Local Functions as Static

    40
  • Dynamic Memory Allocation Consideration

    41
    • Example:

      41
  • Introduce Explicit Parallelism into Code

    41
    • Example 1 (Avoid):

      42
    • Example 2 (Preferred):

      42
  • Explicitly Extract Common Subexpressions

    42
    • Example 1

      43
      • Avoid:

        43
      • Preferred:

        43
    • Example 2

      43
      • Avoid:

        43
      • Preferred:

        43
  • C Language Structure Component Considerations

    43
    • Sort by Base Type Size

      43
    • Pad by Multiple of Largest Base Type Size

      44
      • Original ordering (Avoid):

        44
      • New ordering, with padding (Preferred):

        44
  • Sort Local Variables According to Base Type Size

    44
    • Original ordering (Avoid):

      45
    • Improved ordering (Preferred):

      45
  • Accelerating Floating-Point Divides and Square Roots

    45
    • Example:

      46
  • Avoid Unnecessary Integer Division

    47
    • Example 1 (Avoid):

      47
    • Example 2 (Preferred):

      47
  • Copy Frequently De-referenced Pointer Arguments to Local Variables

    47
    • Example 1 (Avoid):

      48
    • Example 2 (Preferred):

      48
• Instruction Decoding Optimizations

  49
  • Overview

    49
  • Select DirectPath Over VectorPath Instructions

    50
  • LoadExecute Instruction Usage

    50
    • Use LoadExecute Integer Instructions

      50
    • Use LoadExecute FloatingPoint Instructions with FloatingPoint Operands

      51
      • Example 1 (Avoid):

        51
      • Example 2 (Preferred):

        51
    • Avoid LoadExecute FloatingPoint Instructions with Integer Operands

      51
      • Example 1 (Avoid):

        52
      • Example 2 (Preferred):

        52
  • Align Branch Targets in Program Hot Spots

    52
  • Use Short Instruction Lengths

    52
    • Example 1 (Avoid):

      52
    • Example 2 (Preferred):

      53
  • Avoid Partial Register Reads and Writes

    53
  • Replace Certain SHLD Instructions with Alternative Code

    54
    • Example 1

      54
      • (Avoid):

        54
      • (Preferred):

        54
    • Example 2

      54
      • (Avoid):

        54
      • (Preferred):

        54
    • Example 3

      54
      • (Avoid):

        54
      • (Preferred):

        54
  • Use 8Bit SignExtended Immediates

    54
  • Use 8Bit SignExtended Displacements

    55
  • Code Padding Using Neutral Code Fillers

    55
    • Recommendations for the AMDAthlon™ Processor

      56
    • Recommendations for AMDK6® Family and AMDAthlon™ Processor Blended Code

      57
• Cache and Memory Optimizations

  61
  • Memory Size and Alignment Issues

    61
    • Avoid Memory Size Mismatches

      61
      • Example 1 (Avoid):

        61
      • Example 2 (Avoid):

        61
    • Align Data Where Possible

      62
  • Use the 3DNow!™ PREFETCH and PREFETCHW Instructions

    62
    • PREFETCH/W versus PREFETCHNTA/T0/T1 /T2

      63
    • PREFETCHW Usage

      63
    • Multiple Prefetches

      63
      • Example (Multiple Prefetches):

        63
    • Determining Prefetch Distance

      65
    • Prefetch at Least 64 Bytes Away from Surrounding Stores

      65
  • Take Advantage of Write Combining

    66
  • Avoid Placing Code and Data in the Same 64Byte Cache Line

    66
  • StoretoLoad Forwarding Restrictions

    67
    • StoretoLoad Forwarding Pitfalls—True Dependencies

      67
      • NarrowtoWide StoreBuffer Data Forwarding Restriction

        68
        • Example 1 (Avoid):

          68
        • Example 2 (Avoid):

          68
      • WidetoNarrow StoreBuffer Data Forwarding Restriction

        68
        • Example 3 (Avoid):

          68
        • Example 4 (Avoid):

          68
        • Example 5 (Preferred):

          69
      • Misaligned StoreBuffer Data Forwarding Restriction

        69
        • Example 6 (Avoid):

          69
      • HighByte StoreBuffer Data Forwarding Restriction

        69
        • Example 7 (Avoid):

          69
      • One Supported Store- to-Load Forwarding Case

        70
        • Example 8 (Allowed):

          70
    • Summary of StoretoLoad Forwarding Pitfalls to Avoid

      70
  • Stack Alignment Considerations

    70
    • Extend to 32 Bits Before Pushing onto Stack

      70
      • Example (Preferred):

        71
  • Align TBYTE Variables on Quadword Aligned Addresses

    71
  • C Language Structure Component Considerations

    71
    • Example:

      72
  • Sort Variables According to Base Type Size

    72
    • Example:

      72
• Branch Optimizations

  73
  • Avoid Branches Dependent on Random Data

    73
    • AMDAthlon™ Processor Specific Code

      74
      • Example 1 — Signed integer ABS function (X = labs(X)):

        74
      • Example 2 — Unsigned integer min function (z = x < y ? x : y):

        74
    • Blended AMDK6® and AMDAthlon™ Processor Code

      74
      • Example 3 — Signed integer ABS function (X = labs(X)):

        74
      • Example 4 — Unsigned integer min function (z = x < y ? x : y):

        74
      • Example 5 — Hexadecimal to ASCII conversion (y=x < 10 ? x + 0x30: x + 0x41):

        74
      • Example 6 — Increment Ring Buffer Offset:

        75
      • Example 7 — Integer Signum Function:

        75
  • Always Pair CALL and RETURN

    75
  • Replace Branches with Computation in 3DNow!™ Code

    76
    • Muxing Constructs

      76
      • Example 1 (Avoid):

        76
      • Example 2 (Preferred):

        77
    • Sample Code Translated into 3DNow!™ Code

      77
      • Example 1:

        77
        • C code:

          77
        • 3DNow! code:

          77
      • Example 2:

        78
        • C code:

          78
        • 3DNow! code:

          78
      • Example 3:

        78
        • C code:

          78
        • 3DNow! code:

          78
      • Example 4:

        79
        • C code:

          79
        • 3DNow! code:

          79
      • Example 5:

        80
        • C code:

          80
        • 3DNow! code:

          80
  • Avoid the Loop Instruction

    81
    • Example 1 (Avoid):

      81
    • Example 2 (Preferred):

      81
  • Avoid Far Control Transfer Instructions

    81
  • Avoid Recursive Functions

    82
    • Example 1 (Avoid):

      82
    • Example 2 (Preferred):

      82
• Scheduling Optimizations

  83
  • Schedule Instructions According to their Latency

    83
  • Unrolling Loops

    83
    • Complete Loop Unrolling

      83
    • Partial Loop Unrolling

      84
      • Without Loop Unrolling:

        85
      • With Partial Loop Unrolling:

        85
      • Deriving Loop Control For Partially Unrolled Loops

        86
        • Example 1 (rolled loop):

          86
        • Example 2 (partially unrolled loop):

          86
  • Use Function Inlining

    87
    • Overview

      87
    • Always Inline Functions if Called from One Site

      88
    • Always Inline Functions with Fewer than 25 Machine Instructions

      88
  • Avoid Address Generation Interlocks

    88
    • Example 1 (Avoid):

      89
    • Example 2 (Preferred):

      89
  • Use MOVZX and MOVSX

    89
    • Example 1 (Avoid):

      89
    • Example 2 (Preferred):

      89
  • Minimize Pointer Arithmetic in Loops

    89
    • Example 1 (Avoid):

      90
    • Example 2 (Preferred):

      90
    • Example 3 (Preferred):

      91
  • Push Memory Data Carefully

    91
    • Example 1 (Avoid):

      91
    • Example 2 (Preferred):

      91
• Integer Optimizations

  93
  • Replace Divides with Multiplies

    93
    • Multiplication by Reciprocal (Division) Utility

      93
      • Signed Division Utility

        94
      • Unsigned Division Utility

        94
    • Unsigned Division by Multiplication of Constant

      94
      • Algorithm: Divisors 1 <= d < 231, Odd d

        94
      • Derivation of a, m, s

        94
      • Algorithm: Divisors 231 <= d < 232

        94
        • Example 1:

          95
      • Simpler Code for Restricted Dividend

        95
    • Signed Division by Multiplication of Constant

      95
      • Algorithm: Divisors 2 <= d < 231

        95
      • Derivation for a, m, s

        96
      • Signed Division By 2

        96
      • Signed Division By 2n

        96
      • Signed Division By –2

        96
      • Signed Division By –(2n)

        96
      • Remainder of Signed Integer 2 or –2

        96
      • Remainder of Signed Integer 2n or –(2n)

        97
  • Use Alternative Code When Multiplying by a Constant

    97
  • Use MMX™ Instructions for IntegerOnly Work

    99
  • Repeated String Instruction Usage

    100
    • Latency of Repeated String Instructions

      100
    • Guidelines for Repeated String Instructions

      100
      • Use the Largest Possible Operand Size

        100
      • Ensure DF=0 (UP)

        101
      • Align Source and Destination with Operand Size

        101
      • Inline REP String with Low Counts

        101
      • Use Loop for REP String with Low Variable Counts

        101
      • Using MOVQ and MOVNTQ for Block Copy/Fill

        101
  • Use XOR Instruction to Clear Integer Registers

    102
    • Example 1 (Acceptable):

      102
    • Example 2 (Preferred):

      102
  • Efficient 64Bit Integer Arithmetic

    102
    • Example 1 (Addition):

      102
    • Example 2 (Subtraction):

      102
    • Example 3 (Negation):

      102
    • Example 4 (Left shift):

      103
    • Example 5 (Right shift):

      103
    • Example 6 (Multiplication):

      103
    • Example 7 (Division):

      104
    • Example 8 (Remainder):

      105
  • Efficient Implementation of Population Count Function

    107
    • Step 1

      107
    • Step 2

      107
    • Step 3

      108
    • Step 4

      108
      • Example:

        108
  • Derivation of Multiplier Used for Integer Division by Constants

    109
    • Unsigned Derivation for Algorithm, Multiplier, and Shift Factor

      109
    • Signed Derivation for Algorithm, Multiplier, and Shift Factor

      111
• FloatingPoint Optimizations

  113
  • Ensure All FPU Data is Aligned

    113
  • Use Multiplies Rather than Divides

    113
  • Use FFREEP Macro to Pop One Register from the FPU Stack

    114
  • FloatingPoint Compare Instructions

    114
  • Use the FXCH Instruction Rather than FST/FLD Pairs

    115
  • Avoid Using ExtendedPrecision Data

    115
  • Minimize FloatingPointtoInteger Conversions

    116
    • Example 1 (Fast):

      116
    • Example 2 (Potentially faster)

      117
    • Example 3 (Potentially faster):

      118
    • Example 4 (Fastest):

      118
  • Floating-Point Subexpression Elimination

    119
    • Example 1 (Avoid):

      119
    • Example 2 (Preferred):

      119
  • Check Argument Range of Trigonometric Instructions Efficiently

    119
    • Example 1 (Avoid):

      120
    • Example 2 (Preferred):

      120
  • Take Advantage of the FSINCOS Instruction

    121
    • Example 1 (Avoid):

      121
    • Example 2 (Preferred):

      121
• 3DNow!™ and MMX™ Optimizations

  123
  • Use 3DNow!™ Instructions

    123
  • Use FEMMS Instruction

    123
  • Use 3DNow!™ Instructions for Fast Division

    124
    • Optimized 14Bit Precision Divide

      124
      • Example:

        124
    • Optimized Full 24Bit Precision Divide

      124
      • Example:

        124
    • Pipelined Pair of 24Bit Precision Divides

      125
      • Example:

        125
    • NewtonRaphson Reciprocal

      125
  • Use 3DNow!™ Instructions for Fast Square Root and Reciprocal Square Root

    126
    • Optimized 15Bit Precision Square Root

      126
      • Example:

        126
    • Optimized 24Bit Precision Square Root

      126
      • Example:

        126
    • NewtonRaphson Reciprocal Square Root

      127
  • Use MMX™ PMADDWD Instruction to Perform Two 32-Bit Multiplies in Parallel

    127
    • Example:

      128
  • 3DNow!™ and MMX™ IntraOperand Swapping

    128
    • AMDAthlon™ Specific Code

      128
    • Blended Code

      128
      • Example 1 (Preferred, faster):

        128
      • Example 2 (Preferred, fast):

        128
  • Fast Conversion of Signed Words to Floating-Point

    129
    • Example 1 (AMD Athlon specific code using 3DNow! DSP extension):

      129
    • Example 2 (AMDK6 Family and AMD Athlon processor blended code):

      129
  • Use MMX™ PXOR to Negate 3DNow!™ Data

    129
  • Use MMX™ PCMP Instead of 3DNow!™ PFCMP

    130
    • Both Numbers Positive

      130
    • One Negative, One Positive

      130
    • Both Numbers Negative

      130
  • Use MMX™ Instructions for Block Copies and Block Fills

    131
    • AMDK6® and AMDAthlon™ Processor Blended Code

      131
      • Example 1:

        131
    • AMDAthlon™ Processor Specific Code

      133
      • Example 2:

        133
  • Use MMX™ PXOR to Clear All Bits in an MMX™ Register

    134
  • Use MMX™ PCMPEQD to Set All Bits in an MMX™ Register

    135
  • Use MMX™ PAND to Find Absolute Value in 3DNow!™ Code

    135
  • Optimized Matrix Multiplication

    135
  • Efficient 3D-Clipping Code Computation Using 3DNow!™ Instructions

    138
  • Use 3DNow!™ PAVGUSB for MPEG2 Motion Compensation

    139
    • Example 1 (Avoid):

      140
    • Example 2 (Preferred):

      141
  • Stream of Packed Unsigned Bytes

    141
    • Example:

      141
  • Complex Number Arithmetic

    142
    • Example:

      142
    • Example:

      142
• General x86 Optimization Guidelines

  143
  • Short Forms

    143
    • Example 1 (Avoid):

      143
    • Example 2 (Preferred):

      143
  • Dependencies

    144
  • Register Operands

    144
  • Stack Allocation

    144
• Appendix A

  145
• AMDAthlon™Processor Microarchitecture

  145
  • Introduction

    145
  • AMDAthlon™ Processor Microarchitecture

    146
    • Superscalar Processor

      146
    • Instruction Cache

      147
    • Predecode

      148
    • Branch Prediction

      148
    • Early Decoding

      149
      • DirectPath Decoder

        149
      • VectorPath Decoder

        149
    • Instruction Control Unit

      150
    • Data Cache

      150
    • Integer Scheduler

      151
    • Integer Execution Unit

      151
    • FloatingPoint Scheduler

      152
    • FloatingPoint Execution Unit

      153
    • LoadStore Unit (LSU)

      154
    • L2 Cache Controller

      155
    • Write Combining

      155
    • AMDAthlon™ System Bus

      155
• Appendix B

  157
• Pipeline and Execution Unit Resources Overview

  157
  • Fetch and Decode Pipeline Stages

    157
    • Cycle 1–FETCH

      159
    • Cycle 2–SCAN

      159
    • Cycle 3 (DirectPath)– ALIGN1

      159
    • Cycle 3 (VectorPath)– MECTL

      159
    • Cycle 4 (DirectPath)– ALIGN2

      159
    • Cycle 4 (VectorPath)– MEROM

      159
    • Cycle 5 (DirectPath)– EDEC

      159
    • Cycle 5 (VectorPath)– MEDEC/MESEQ

      159
    • Cycle 6– IDEC/Rename

      159
  • Integer Pipeline Stages

    160
    • Cycle 7–SCHED

      161
    • Cycle 8–EXEC

      161
    • Cycle 9–ADDGEN

      161
    • Cycle 10–DCACC

      161
    • Cycle 11–RESP

      161
  • FloatingPoint Pipeline Stages

    162
    • Cycle 7–STKREN

      163
    • Cycle 8–REGREN

      163
    • Cycle 9–SCHEDW

      163
    • Cycle 10–SCHED

      163
    • Cycle 11–FREG

      163
    • Cycle 12–15– FloatingPoint Execution (FEXEC1–4)

      163
  • Execution Unit Resources

    164
    • Terminology

      164
      • Operands

        164
      • Results

        164
      • Examples

        164
    • Integer Pipeline Operations

      165
    • FloatingPoint Pipeline Operations

      166
    • Load/Store Pipeline Operations

      167
    • Code Sample Analysis

      168
• Appendix C

  171
• Implementation of Write Combining

  171
  • Introduction

    171
  • WriteCombining Definitions and Abbreviations

    172
  • What is Write Combining?

    172
  • Programming Details

    172
  • Write-Combining Operations

    173
    • Sending WriteBuffer Data to the System

      175
• Appendix D

  177
• Performance-Monitoring Counters

  177
  • Overview

    177
  • Performance Counter Usage

    177
    • PerfEvtSel[3:0] MSRs (MSR Addresses C001_0000h–C001_0003h)

      178
      • Event Select Field (Bits 0—7)

        178
      • Unit Mask Field (Bits 8—15)

        179
      • USR (User Mode) Flag (Bit 16)

        179
      • OS (Operating System Mode) Flag (Bit 17)

        179
      • E (Edge Detect) Flag (Bit 18)

        179
      • PC (Pin Control) Flag (Bit 19)

        179
      • INT (APIC Interrupt Enable) Flag (Bit 20)

        179
      • EN (Enable Counter) Flag (Bit 22)

        179
      • INV (Invert) Flag (Bit 23)

        179
      • Counter Mask Field (Bits 31–24)

        179
    • PerfCtr[3:0] MSRs (MSR Addresses C001_0004h–C001_0007h)

      183
    • Starting and Stopping the PerformanceMonitoring Counters

      184
  • Event and TimeStamp Monitoring Software

    184
  • Monitoring Counter Overflow

    185
• Appendix E

  187
• Programming the MTRR and PAT

  187
  • Introduction

    187
  • Memory Type Range Register (MTRR) Mechanism

    187
    • Memory Types

      190
    • MTRR Capability Register Format

      190
      • MTRR Default Type Register Format

        191
    • MTRR Overlapping

      192
  • Page Attribute Table (PAT)

    193
    • MSR Access

      193
    • Accessing the PAT

      194
    • MTRRs and PAT

      194
    • MTRR Fixed-Range Register Format

      198
    • Variable-Range MTRRs

      199
    • Variable-Range MTRR Register Format

      199
    • MTRR MSR Format

      201
• Appendix F

  203
• Instruction Dispatch and Execution Resources

  203
• Appendix G

  235
• DirectPath versus VectorPath Instructions

  235
  • Select DirectPath Over VectorPath Instructions

    235
  • DirectPath Instructions

    235
  • VectorPath Instructions

    247
  • Index

    253