Add a scheduling model for AMD 16H Jaguar (btver2)

[spatel]

Summary

This is a first pass at a scheduling model for Jaguar. It's structured largely on the existing SandyBridge and SLM sched models.

Using this model, in addition to turning on the PostRA scheduler, results in some perf wins on internal and 3rd party benchmarks. There's not much difference in LLVM's test-suite benchmarking subset of tests.

[andreadb]

Hi Sanjay,

The patch LGTM, thanks!
As you said, the scheduling model is largely based on the SandyBridge model. I verified that the numbers (latencies and values in the BtVer2Model) make sense.

I have tested your model on our internal codebase. On average I have seen runtime performance improvements in the range of ~1-1.5%. In particular, your scheduling model improves the runtime of a large internal codebase by the 1% (~2.5% with LTO).
As you said, this is a first pass at a scheduling model for Jaguar, so overall, the patch looks good.

Thanks!
Andrea

[spatel]

Closed by commit rL217457 (authored by @spatel).

[spatel]

Thanks, Andrea - checked in at r217457.

llvm/trunk/lib/Target/X86/X86.td

 def : Proc<"amdfam10",        [FeatureSSE4A,
                                Feature3DNowA, FeatureCMPXCHG16B, FeatureLZCNT,
                                FeaturePOPCNT, FeatureSlowBTMem,
                                FeatureSlowSHLD]>;
 // Bobcat
 def : Proc<"btver1",          [FeatureSSSE3, FeatureSSE4A, FeatureCMPXCHG16B,
                                FeaturePRFCHW, FeatureLZCNT, FeaturePOPCNT,
                                FeatureSlowSHLD]>;
+
 // Jaguar
-def : Proc<"btver2",          [FeatureAVX, FeatureSSE4A, FeatureCMPXCHG16B,
-                               FeaturePRFCHW, FeatureAES, FeaturePCLMUL,
-                               FeatureBMI, FeatureF16C, FeatureMOVBE,
-                               FeatureLZCNT, FeaturePOPCNT, FeatureSlowSHLD]>;
+def : ProcessorModel<"btver2", BtVer2Model,
+                     [FeatureAVX, FeatureSSE4A, FeatureCMPXCHG16B,
+
 // Bulldozer
 def : Proc<"bdver1",          [FeatureXOP, FeatureFMA4, FeatureCMPXCHG16B,
                                FeatureAES, FeaturePRFCHW, FeaturePCLMUL,
                                FeatureLZCNT, FeaturePOPCNT, FeatureSlowSHLD]>;
 // Piledriver
 def : Proc<"bdver2",          [FeatureXOP, FeatureFMA4, FeatureCMPXCHG16B,
                                FeatureAES, FeaturePRFCHW, FeaturePCLMUL,
                                FeatureF16C, FeatureLZCNT,

llvm/trunk/lib/Target/X86/X86Schedule.td

def GenericModel : SchedMachineModel {
   let HighLatency = 10;
   let PostRAScheduler = 0;
 }
 
 include "X86ScheduleAtom.td"
 include "X86SchedSandyBridge.td"
 include "X86SchedHaswell.td"
 include "X86ScheduleSLM.td"
+include "X86ScheduleBtVer2.td"
+

llvm/trunk/lib/Target/X86/X86ScheduleBtVer2.td

+//=- X86ScheduleBtVer2.td - X86 BtVer2 (Jaguar) Scheduling ---*- tablegen -*-=//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file defines the machine model for AMD btver2 (Jaguar) to support
+// instruction scheduling and other instruction cost heuristics. Based off AMD Software
+// Optimization Guide for AMD Family 16h Processors & Instruction Latency appendix.
+//
+//===----------------------------------------------------------------------===//
+
+def BtVer2Model : SchedMachineModel {
+  // All x86 instructions are modeled as a single micro-op, and btver2 can
+  // decode 2 instructions per cycle.
+  let IssueWidth = 2;
+  let MicroOpBufferSize = 64; // Retire Control Unit
+  let LoadLatency = 5; // FPU latency (worse case cf Integer 3 cycle latency)
+  let HighLatency = 25;
+  let MispredictPenalty = 14; // Minimum branch misdirection penalty
+  let PostRAScheduler = 1;
+
+  // FIXME: SSE4/AVX is unimplemented. This flag is set to allow
+  // the scheduler to assign a default model to unrecognized opcodes.
+  let CompleteModel = 0;
+}
+
+let SchedModel = BtVer2Model in {
+
+// Jaguar can issue up to 6 micro-ops in one cycle
+def JALU0 : ProcResource<1>; // Integer Pipe0: integer ALU0 (also handle FP->INT jam)
+def JALU1 : ProcResource<1>; // Integer Pipe1: integer ALU1/MUL/DIV
+def JLAGU : ProcResource<1>; // Integer Pipe2: LAGU
+def JSAGU : ProcResource<1>; // Integer Pipe3: SAGU (also handles 3-operand LEA)
+def JFPU0 : ProcResource<1>; // Vector/FPU Pipe0: VALU0/VIMUL/FPA
+def JFPU1 : ProcResource<1>; // Vector/FPU Pipe1: VALU1/STC/FPM
+
+// Any pipe - FIXME we need this until we can discriminate between int/fpu load/store/moves properly
+def JAny : ProcResGroup<[JALU0, JALU1, JLAGU, JSAGU, JFPU0, JFPU1]>;
+
+// Integer Pipe Scheduler
+def JALU01 : ProcResGroup<[JALU0, JALU1]> {
+  let BufferSize=20;
+}
+
+// AGU Pipe Scheduler
+def JLSAGU : ProcResGroup<[JLAGU, JSAGU]> {
+  let BufferSize=12;
+}
+
+// Fpu Pipe Scheduler
+def JFPU01 : ProcResGroup<[JFPU0, JFPU1]> {
+  let BufferSize=18;
+}
+
+def JDiv    : ProcResource<1>; // integer division
+def JMul    : ProcResource<1>; // integer multiplication
+def JVALU0  : ProcResource<1>; // vector integer
+def JVALU1  : ProcResource<1>; // vector integer
+def JVIMUL  : ProcResource<1>; // vector integer multiplication
+def JSTC    : ProcResource<1>; // vector store/convert
+def JFPM    : ProcResource<1>; // FP multiplication
+def JFPA    : ProcResource<1>; // FP addition
+
+// Integer loads are 3 cycles, so ReadAfterLd registers needn't be available until 3
+// cycles after the memory operand.
+def : ReadAdvance<ReadAfterLd, 3>;
+
+// Many SchedWrites are defined in pairs with and without a folded load.
+// Instructions with folded loads are usually micro-fused, so they only appear
+// as two micro-ops when dispatched by the schedulers.
+// This multiclass defines the resource usage for variants with and without
+// folded loads.
+multiclass JWriteResIntPair<X86FoldableSchedWrite SchedRW,
+                          ProcResourceKind ExePort,
+                          int Lat> {
+  // Register variant is using a single cycle on ExePort.
+  def : WriteRes<SchedRW, [ExePort]> { let Latency = Lat; }
+
+  // Memory variant also uses a cycle on JLAGU and adds 3 cycles to the
+  // latency.
+  def : WriteRes<SchedRW.Folded, [JLAGU, ExePort]> {
+     let Latency = !add(Lat, 3);
+  }
+}
+
+multiclass JWriteResFpuPair<X86FoldableSchedWrite SchedRW,
+                          ProcResourceKind ExePort,
+                          int Lat> {
+  // Register variant is using a single cycle on ExePort.
+  def : WriteRes<SchedRW, [ExePort]> { let Latency = Lat; }
+
+  // Memory variant also uses a cycle on JLAGU and adds 5 cycles to the
+  // latency.
+  def : WriteRes<SchedRW.Folded, [JLAGU, ExePort]> {
+     let Latency = !add(Lat, 5);
+  }
+}
+
+// A folded store needs a cycle on the SAGU for the store data.
+def : WriteRes<WriteRMW, [JSAGU]>;
+
+////////////////////////////////////////////////////////////////////////////////
+// Arithmetic.
+////////////////////////////////////////////////////////////////////////////////
+
+defm : JWriteResIntPair<WriteALU,   JALU01, 1>;
+defm : JWriteResIntPair<WriteIMul,  JALU1,  3>;
+
+def  : WriteRes<WriteIMulH, [JALU1]> {
+  let Latency = 6;
+  let ResourceCycles = [4];
+}
+
+// FIXME 8/16 bit divisions
+def : WriteRes<WriteIDiv, [JALU1, JDiv]> {
+  let Latency = 25;
+  let ResourceCycles = [1, 25];
+}
+def : WriteRes<WriteIDivLd, [JALU1, JLAGU, JDiv]> {
+  let Latency = 41;
+  let ResourceCycles = [1, 1, 25];
+}
+
+// This is for simple LEAs with one or two input operands.
+// FIXME: SAGU 3-operand LEA
+def : WriteRes<WriteLEA, [JALU01]>;
+
+////////////////////////////////////////////////////////////////////////////////
+// Integer shifts and rotates.
+////////////////////////////////////////////////////////////////////////////////
+
+defm : JWriteResIntPair<WriteShift, JALU01, 1>;
+
+////////////////////////////////////////////////////////////////////////////////
+// Loads, stores, and moves, not folded with other operations.
+// FIXME: Split x86 and SSE load/store/moves
+////////////////////////////////////////////////////////////////////////////////
+
+def : WriteRes<WriteLoad,  [JLAGU]> { let Latency = 5; }
+def : WriteRes<WriteStore, [JSAGU]>;
+def : WriteRes<WriteMove,  [JAny]>;
+
+////////////////////////////////////////////////////////////////////////////////
+// Idioms that clear a register, like xorps %xmm0, %xmm0.
+// These can often bypass execution ports completely.
+////////////////////////////////////////////////////////////////////////////////
+
+def : WriteRes<WriteZero,  []>;
+
+////////////////////////////////////////////////////////////////////////////////
+// Branches don't produce values, so they have no latency, but they still
+// consume resources. Indirect branches can fold loads.
+////////////////////////////////////////////////////////////////////////////////
+
+defm : JWriteResIntPair<WriteJump,  JALU01, 1>;
+
+////////////////////////////////////////////////////////////////////////////////
+// Floating point. This covers both scalar and vector operations.
+// FIXME: should we bother splitting JFPU pipe + unit stages for fast instructions?
+// FIXME: Double precision latencies
+// FIXME: SS vs PS latencies
+// FIXME: RSQRT latencies
+// FIXME: ymm latencies
+////////////////////////////////////////////////////////////////////////////////
+
+defm : JWriteResFpuPair<WriteFAdd,      JFPU0,  3>;
+defm : JWriteResFpuPair<WriteFMul,      JFPU1,  2>;
+defm : JWriteResFpuPair<WriteFRcp,      JFPU1,  2>;
+defm : JWriteResFpuPair<WriteFShuffle, JFPU01,  1>;
+defm : JWriteResFpuPair<WriteFBlend,   JFPU01,  1>;
+defm : JWriteResFpuPair<WriteFShuffle256, JFPU01, 1>;
+
+def : WriteRes<WriteFSqrt, [JFPU1, JLAGU, JFPM]> {
+  let Latency = 21;
+  let ResourceCycles = [1, 1, 21];
+}
+def : WriteRes<WriteFSqrtLd, [JFPU1, JLAGU, JFPM]> {
+  let Latency = 26;
+  let ResourceCycles = [1, 1, 21];
+}
+
+def : WriteRes<WriteFDiv, [JFPU1, JLAGU, JFPM]> {
+  let Latency = 19;
+  let ResourceCycles = [1, 1, 19];
+}
+def : WriteRes<WriteFDivLd, [JFPU1, JLAGU, JFPM]> {
+  let Latency = 24;
+  let ResourceCycles = [1, 1, 19];
+}
+
+// FIXME: integer pipes
+defm : JWriteResFpuPair<WriteCvtF2I,    JFPU1,  3>; // Float -> Integer.
+defm : JWriteResFpuPair<WriteCvtI2F,    JFPU1,  3>; // Integer -> Float.
+defm : JWriteResFpuPair<WriteCvtF2F,    JFPU1,  3>; // Float -> Float size conversion.
+
+def : WriteRes<WriteFVarBlend, [JFPU01]> {
+  let Latency = 2;
+  let ResourceCycles = [2];
+}
+def : WriteRes<WriteFVarBlendLd, [JLAGU, JFPU01]> {
+  let Latency = 7;
+  let ResourceCycles = [1, 2];
+}
+
+// Vector integer operations.
+defm : JWriteResFpuPair<WriteVecALU,   JFPU01,  1>;
+defm : JWriteResFpuPair<WriteVecShift, JFPU01,  1>;
+defm : JWriteResFpuPair<WriteVecIMul,  JFPU0,   2>;
+defm : JWriteResFpuPair<WriteShuffle,  JFPU01,  1>;
+defm : JWriteResFpuPair<WriteBlend,    JFPU01,  1>;
+defm : JWriteResFpuPair<WriteVecLogic, JFPU01,  1>;
+defm : JWriteResFpuPair<WriteShuffle256, JFPU01, 1>;
+
+def : WriteRes<WriteVarBlend, [JFPU01]> {
+  let Latency = 2;
+  let ResourceCycles = [2];
+}
+def : WriteRes<WriteVarBlendLd, [JLAGU, JFPU01]> {
+  let Latency = 7;
+  let ResourceCycles = [1, 2];
+}
+
+// FIXME: why do we need to define AVX2 resource on CPU that doesn't have AVX2?
+def : WriteRes<WriteVarVecShift, [JFPU01]> {
+  let Latency = 1;
+  let ResourceCycles = [1];
+}
+def : WriteRes<WriteVarVecShiftLd, [JLAGU, JFPU01]> {
+  let Latency = 6;
+  let ResourceCycles = [1, 1];
+}
+
+def : WriteRes<WriteMPSAD, [JFPU0]> {
+  let Latency = 3;
+  let ResourceCycles = [2];
+}
+def : WriteRes<WriteMPSADLd, [JLAGU, JFPU0]> {
+  let Latency = 8;
+  let ResourceCycles = [1, 2];
+}
+
+////////////////////////////////////////////////////////////////////////////////
+// String instructions.
+// Packed Compare Implicit Length Strings, Return Mask
+// FIXME: approximate latencies + pipe dependencies
+////////////////////////////////////////////////////////////////////////////////
+
+def : WriteRes<WritePCmpIStrM, [JFPU01]> {
+  let Latency = 7;
+  let ResourceCycles = [2];
+}
+def : WriteRes<WritePCmpIStrMLd, [JLAGU, JFPU01]> {
+  let Latency = 12;
+  let ResourceCycles = [1, 2];
+}
+
+// Packed Compare Explicit Length Strings, Return Mask
+def : WriteRes<WritePCmpEStrM, [JFPU01]> {
+  let Latency = 13;
+  let ResourceCycles = [5];
+}
+def : WriteRes<WritePCmpEStrMLd, [JLAGU, JFPU01]> {
+  let Latency = 18;
+  let ResourceCycles = [1, 5];
+}
+
+// Packed Compare Implicit Length Strings, Return Index
+def : WriteRes<WritePCmpIStrI, [JFPU01]> {
+  let Latency = 6;
+  let ResourceCycles = [2];
+}
+def : WriteRes<WritePCmpIStrILd, [JLAGU, JFPU01]> {
+  let Latency = 11;
+  let ResourceCycles = [1, 2];
+}
+
+// Packed Compare Explicit Length Strings, Return Index
+def : WriteRes<WritePCmpEStrI, [JFPU01]> {
+  let Latency = 13;
+  let ResourceCycles = [5];
+}
+def : WriteRes<WritePCmpEStrILd, [JLAGU, JFPU01]> {
+  let Latency = 18;
+  let ResourceCycles = [1, 5];
+}
+
+////////////////////////////////////////////////////////////////////////////////
+// AES Instructions.
+////////////////////////////////////////////////////////////////////////////////
+
+def : WriteRes<WriteAESDecEnc, [JFPU01, JVIMUL]> {
+  let Latency = 3;
+  let ResourceCycles = [1, 1];
+}
+def : WriteRes<WriteAESDecEncLd, [JFPU01, JLAGU, JVIMUL]> {
+  let Latency = 8;
+  let ResourceCycles = [1, 1, 1];
+}
+
+def : WriteRes<WriteAESIMC, [JVIMUL]> {
+  let Latency = 2;
+  let ResourceCycles = [1];
+}
+def : WriteRes<WriteAESIMCLd, [JLAGU, JVIMUL]> {
+  let Latency = 7;
+  let ResourceCycles = [1, 1];
+}
+
+def : WriteRes<WriteAESKeyGen, [JVIMUL]> {
+  let Latency = 2;
+  let ResourceCycles = [1];
+}
+def : WriteRes<WriteAESKeyGenLd, [JLAGU, JVIMUL]> {
+  let Latency = 7;
+  let ResourceCycles = [1, 1];
+}
+
+////////////////////////////////////////////////////////////////////////////////
+// Carry-less multiplication instructions.
+////////////////////////////////////////////////////////////////////////////////
+
+def : WriteRes<WriteCLMul, [JVIMUL]> {
+  let Latency = 2;
+  let ResourceCycles = [1];
+}
+def : WriteRes<WriteCLMulLd, [JLAGU, JVIMUL]> {
+  let Latency = 7;
+  let ResourceCycles = [1, 1];
+}
+
+// FIXME: pipe for system/microcode?
+def : WriteRes<WriteSystem,     [JAny]> { let Latency = 100; }
+def : WriteRes<WriteMicrocoded, [JAny]> { let Latency = 100; }
+def : WriteRes<WriteFence,  [JSAGU]>;
+def : WriteRes<WriteNop, []>;
+} // SchedModel
+