From 2e5f39ee0fb4b5d6303e4342b17146952b9bb79c Mon Sep 17 00:00:00 2001 From: Alexey Volkov Date: Wed, 23 Apr 2014 08:57:09 +0000 Subject: [PATCH] [X86] Silvermont new scheduler model This model is not final and work is still in progress. However there are substantial improvements on integer tests mainly because of better RAL with new scheduler. Differential Revision: http://reviews.llvm.org/D3451 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@206957 91177308-0d34-0410-b5e6-96231b3b80d8 --- lib/Target/X86/X86ScheduleSLM.td | 846 ++++++++----------------------- 1 file changed, 203 insertions(+), 643 deletions(-) diff --git a/lib/Target/X86/X86ScheduleSLM.td b/lib/Target/X86/X86ScheduleSLM.td index 6c2a3040240..f29c987c96e 100644 --- a/lib/Target/X86/X86ScheduleSLM.td +++ b/lib/Target/X86/X86ScheduleSLM.td @@ -1,4 +1,4 @@ -//===- X86ScheduleSLM.td - X86 Atom Scheduling Definitions -*- tablegen -*-==// +//=- X86ScheduleSLM.td - X86 Silvermont Scheduling -----------*- tablegen -*-=// // // The LLVM Compiler Infrastructure // @@ -7,662 +7,222 @@ // //===----------------------------------------------------------------------===// // -// This file defines the itinerary class data for the Intel Atom -// (Silvermont) processor. +// This file defines the machine model for Intal Silvermont to support +// instruction scheduling and other instruction cost heuristics. // //===----------------------------------------------------------------------===// -def IEC_RSV0 : FuncUnit; -def IEC_RSV1 : FuncUnit; -def FPC_RSV0 : FuncUnit; -def FPC_RSV1 : FuncUnit; -def MEC_RSV : FuncUnit; - - - - - - - - - - - - - - -def SLMItineraries : ProcessorItineraries< - [ IEC_RSV0, IEC_RSV1, FPC_RSV0, FPC_RSV1, MEC_RSV ], - [], [ - // [InstrStage] - // [InstrStage, InstrStage] - // [InstrStage] - // [InstrStage,InstrStage] - // - // Default is 1 cycle, IEC_RSV0 or IEC_RSV1 - //InstrItinData] >, - InstrItinData, - InstrStage<1, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - // mul - InstrItinData] >, - InstrItinData, - InstrStage<4, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData, - InstrStage<3, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData] >, - // imul by al, ax, eax, rax - InstrItinData] >, - InstrItinData, - InstrStage<6, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData, - InstrStage<6, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData] >, - // imul reg by reg|mem - InstrItinData, - InstrStage<4, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData, - InstrStage<3, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData, - InstrStage<4, [MEC_RSV]>] >, - InstrItinData] >, - // imul reg = reg/mem * imm - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData, - InstrStage<4, [MEC_RSV]>] >, - InstrItinData, - InstrStage<3, [MEC_RSV]>] >, - InstrItinData, - InstrStage<4, [MEC_RSV]>] >, - // idiv - min latency - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - // div - min latency - InstrItinData] >, - InstrItinData, - InstrStage<25, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - // neg/not/inc/dec - InstrItinData] >, - InstrItinData, - InstrStage<1, [MEC_RSV]>] >, - // add/sub/and/or/xor/adc/sbc/cmp/test - InstrItinData] >, - InstrItinData, - InstrStage<1, [MEC_RSV]>] >, - // adc/sbb - InstrItinData] >, - InstrItinData, - InstrStage<2, [MEC_RSV]>] >, - // shift/rotate - InstrItinData, - InstrStage<1, [MEC_RSV]>] >, - // shift double - InstrItinData] >, - InstrItinData] >, - InstrItinData, - InstrStage<2, [MEC_RSV]>] >, - InstrItinData, - InstrStage<4, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData, - InstrStage<2, [MEC_RSV]>] >, - InstrItinData, - InstrStage<4, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData, - InstrStage<2, [MEC_RSV]>] >, - InstrItinData, - InstrStage<4, [MEC_RSV]>] >, - // cmov - InstrItinData, - InstrStage<2, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData, - InstrStage<2, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData, - InstrStage<2, [MEC_RSV]>] >, - InstrItinData] >, - // set - InstrItinData, - InstrStage<1, [MEC_RSV]>] >, - InstrItinData] >, - // jcc - InstrItinData] >, - // jcxz/jecxz/jrcxz - InstrItinData] >, - // jmp rel - InstrItinData] >, - // jmp indirect - InstrItinData] >, - InstrItinData, - InstrStage<1, [MEC_RSV]>] >, - // jmp far - InstrItinData, - InstrStage<1, [MEC_RSV]>] >, - InstrItinData] >, - // loop/loope/loopne - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - // call - all but reg/imm - InstrItinData] >, - InstrItinData, - InstrStage<1, [MEC_RSV]>] >, - InstrItinData, - InstrStage<1, [MEC_RSV]>] >, - InstrItinData] >, - //ret - InstrItinData] >, - InstrItinData] >, - //sign extension movs - InstrItinData] >, - InstrItinData] >, - InstrItinData, - InstrStage<1, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData] >, - //zero extension movs - InstrItinData] >, - InstrItinData] >, - InstrItinData, - InstrStage<1, [MEC_RSV]>] >, - - InstrItinData] >, - InstrItinData] >, - - // SSE binary operations - // arithmetic fp scalar - InstrItinData] >, - InstrItinData, - InstrStage<3, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData, - InstrStage<3, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData, - InstrStage<1, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData, - InstrStage<2, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData, - InstrStage<13, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData, - InstrStage<13, [MEC_RSV]>] >, - - InstrItinData] >, - InstrItinData] >, - - InstrItinData] >, - InstrItinData, - InstrStage<6, [MEC_RSV]>] >, - - // arithmetic fp parallel - InstrItinData] >, - InstrItinData, - InstrStage<3, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData, - InstrStage<4, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData, - InstrStage<2, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData, - InstrStage<4, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData, - InstrStage<27, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData, - InstrStage<27, [MEC_RSV]>] >, - - // bitwise parallel - InstrItinData] >, - InstrItinData, - InstrStage<1, [MEC_RSV]>] >, - - // arithmetic int parallel - InstrItinData] >, - InstrItinData, - InstrStage<1, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData, - InstrStage<4, [MEC_RSV]>] >, - - // multiply int parallel - InstrItinData] >, - InstrItinData, - InstrStage<5, [MEC_RSV]>] >, - - // shift parallel - InstrItinData] >, - InstrItinData, - InstrStage<2, [MEC_RSV]>] >, - InstrItinData] >, - - InstrItinData] >, - - InstrItinData] >, - InstrItinData] >, - InstrItinData, - InstrStage<1, [MEC_RSV]>] >, - - InstrItinData] >, - - InstrItinData] >, - InstrItinData, - InstrStage<26, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData, - InstrStage<13, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData, - InstrStage<26, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData, - InstrStage<13, [MEC_RSV]>] >, - - InstrItinData] >, - InstrItinData, - InstrStage<9, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData, - InstrStage<4, [MEC_RSV]>] >, - - InstrItinData] >, - InstrItinData] >, - - InstrItinData] >, - InstrItinData] >, - - InstrItinData] >, - InstrItinData, - InstrStage<1, [MEC_RSV]>] >, - - InstrItinData] >, - InstrItinData, - InstrStage<1, [MEC_RSV]>] >, - InstrItinData, - InstrStage<1, [MEC_RSV]>] >, - - InstrItinData] >, - InstrItinData, - InstrStage<1, [MEC_RSV]>] >, - InstrItinData, - InstrStage<1, [MEC_RSV]>] >, - - InstrItinData] >, - InstrItinData, - InstrStage<1, [MEC_RSV]>] >, - InstrItinData, - InstrStage<1, [MEC_RSV]>] >, - - InstrItinData] >, - - InstrItinData] >, - - InstrItinData] >, - InstrItinData, - InstrStage<1, [FPC_RSV0, FPC_RSV1]>] >, - InstrItinData] >, - - InstrItinData] >, - - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - - InstrItinData] >, - InstrItinData, - InstrStage<6, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData, - InstrStage<9, [MEC_RSV]>] >, - InstrItinData, - InstrStage<9, [MEC_RSV]>] >, - InstrItinData, - InstrStage<9, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData, - InstrStage<5, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData, - InstrStage<1, [MEC_RSV]>] >, - - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData, - InstrStage<1, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData] >, - - // conversions - // to/from PD ... - InstrItinData] >, - InstrItinData, - InstrStage<5, [MEC_RSV]>] >, - // to/from PS except to/from PD and PS2PI - InstrItinData] >, - InstrItinData, - InstrStage<4, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData, - InstrStage<1, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData, - InstrStage<4, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData, - InstrStage<4, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData, - InstrStage<4, [MEC_RSV]>] >, - - // MMX MOVs - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - // other MMX - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - // conversions - // from/to PD - InstrItinData] >, - InstrItinData] >, - // from/to PI - InstrItinData] >, - InstrItinData] >, - - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - - InstrItinData] >, - InstrItinData] >, - - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, +def SLMModel : SchedMachineModel { + // All x86 instructions are modeled as a single micro-op, and SLM can decode 2 + // instructions per cycle. + let IssueWidth = 2; + let MicroOpBufferSize = 32; // Based on the reorder buffer. + let LoadLatency = 3; + let MispredictPenalty = 10; + + // FIXME: SSE4 is unimplemented. This flag is set to allow + // the scheduler to assign a default model to unrecognized opcodes. + let CompleteModel = 0; +} - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, +let SchedModel = SLMModel in { + +// Silveromnt has 5 reservation stations for micro-ops + +def IEC_RSV0 : ProcResource<1>; +def IEC_RSV1 : ProcResource<1>; +def FPC_RSV0 : ProcResource<1> { let BufferSize = 1; } +def FPC_RSV1 : ProcResource<1> { let BufferSize = 1; } +def MEC_RSV : ProcResource<1>; + +// Many micro-ops are capable of issuing on multiple ports. +def IEC_RSV01 : ProcResGroup<[IEC_RSV0, IEC_RSV1]>; +def FPC_RSV01 : ProcResGroup<[FPC_RSV0, FPC_RSV1]>; + +def SMDivider : ProcResource<1>; +def SMFPMultiplier : ProcResource<1>; +def SMFPDivider : ProcResource<1>; + +// Loads are 3 cycles, so ReadAfterLd registers needn't be available until 3 +// cycles after the memory operand. +def : ReadAdvance; + +// 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 queued in the reservation station. +// This multiclass defines the resource usage for variants with and without +// folded loads. +multiclass SMWriteResPair { + // Register variant is using a single cycle on ExePort. + def : WriteRes { let Latency = Lat; } + + // Memory variant also uses a cycle on MEC_RSV and adds 3 cycles to the + // latency. + def : WriteRes { + let Latency = !add(Lat, 3); + } +} - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, +// A folded store needs a cycle on MEC_RSV for the store data, but it does not +// need an extra port cycle to recompute the address. +def : WriteRes; + +def : WriteRes; +def : WriteRes { let Latency = 3; } +def : WriteRes; +def : WriteRes; + +defm : SMWriteResPair; +defm : SMWriteResPair; +defm : SMWriteResPair; +defm : SMWriteResPair; + +// This is for simple LEAs with one or two input operands. +// The complex ones can only execute on port 1, and they require two cycles on +// the port to read all inputs. We don't model that. +def : WriteRes; + +// This is quite rough, latency depends on the dividend. +def : WriteRes { + let Latency = 25; + let ResourceCycles = [1, 25]; +} +def : WriteRes { + let Latency = 29; + let ResourceCycles = [1, 1, 25]; +} - // System instructions - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, +// Scalar and vector floating point. +defm : SMWriteResPair; +defm : SMWriteResPair; +defm : SMWriteResPair; +defm : SMWriteResPair; +defm : SMWriteResPair; +defm : SMWriteResPair; +defm : SMWriteResPair; +defm : SMWriteResPair; + +// This is quite rough, latency depends on precision +def : WriteRes { + let Latency = 5; + let ResourceCycles = [1, 2]; +} +def : WriteRes { + let Latency = 8; + let ResourceCycles = [1, 1, 2]; +} - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, +def : WriteRes { + let Latency = 34; + let ResourceCycles = [1, 34]; +} +def : WriteRes { + let Latency = 37; + let ResourceCycles = [1, 1, 34]; +} - InstrItinData] >, - InstrItinData] >, - // worst case for mov REG_CRx - InstrItinData] >, - InstrItinData] >, +// Vector integer operations. +defm : SMWriteResPair; +defm : SMWriteResPair; +defm : SMWriteResPair; +defm : SMWriteResPair; +defm : SMWriteResPair; +defm : SMWriteResPair; +defm : SMWriteResPair; + +// String instructions. +// Packed Compare Implicit Length Strings, Return Mask +def : WriteRes { + let Latency = 13; + let ResourceCycles = [13]; +} +def : WriteRes { + let Latency = 13; + let ResourceCycles = [13, 1]; +} - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - // LAR - InstrItinData] >, - InstrItinData] >, - // LSL - InstrItinData] >, - InstrItinData] >, +// Packed Compare Explicit Length Strings, Return Mask +def : WriteRes { + let Latency = 17; + let ResourceCycles = [17]; +} +def : WriteRes { + let Latency = 17; + let ResourceCycles = [17, 1]; +} - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - // push control register, segment registers - InstrItinData] >, - InstrItinData] >, - // pop control register, segment registers - InstrItinData] >, - InstrItinData] >, - // VERR, VERW - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - // WRMSR, RDMSR - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - // SMSW, LMSW - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, +// Packed Compare Implicit Length Strings, Return Index +def : WriteRes { + let Latency = 17; + let ResourceCycles = [17]; +} +def : WriteRes { + let Latency = 17; + let ResourceCycles = [17, 1]; +} - InstrItinData] >, - InstrItinData] >, +// Packed Compare Explicit Length Strings, Return Index +def : WriteRes { + let Latency = 21; + let ResourceCycles = [21]; +} +def : WriteRes { + let Latency = 21; + let ResourceCycles = [21, 1]; +} - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, +// AES Instructions. +def : WriteRes { + let Latency = 8; + let ResourceCycles = [5]; +} +def : WriteRes { + let Latency = 8; + let ResourceCycles = [5, 1]; +} - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, +def : WriteRes { + let Latency = 8; + let ResourceCycles = [5]; +} +def : WriteRes { + let Latency = 8; + let ResourceCycles = [5, 1]; +} - InstrItinData] >, - InstrItinData, - InstrStage<10, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData, - InstrStage<1, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData, - InstrStage<1, [MEC_RSV]>] >, - InstrItinData, - InstrStage<1, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData, - InstrStage<1, [MEC_RSV]>] >, - InstrItinData, - InstrStage<1, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData, - InstrStage<5, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData, - InstrStage<5, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData, - InstrStage<6, [MEC_RSV]>] >, - InstrItinData, - InstrStage<6, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData, - InstrStage<1, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData, - InstrStage<4, [MEC_RSV]>] >, - InstrItinData, - InstrStage<10, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData, - InstrStage<3, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData, - InstrStage<12, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData, - InstrStage<15, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData] >, - InstrItinData, - InstrStage<4, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData, - InstrStage<1, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData, - InstrStage<1, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData, - InstrStage<11, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData, - InstrStage<5, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData, - InstrStage<4, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData, - InstrStage<4, [MEC_RSV]>] >, - InstrItinData] >, - InstrItinData, - InstrStage<10, [MEC_RSV]>] >, +def : WriteRes { + let Latency = 8; + let ResourceCycles = [5]; +} +def : WriteRes { + let Latency = 8; + let ResourceCycles = [5, 1]; +} - InstrItinData] > - ]>; +// Carry-less multiplication instructions. +def : WriteRes { + let Latency = 10; + let ResourceCycles = [10]; +} +def : WriteRes { + let Latency = 10; + let ResourceCycles = [10, 1]; +} -// Silvermont machine model. -def SLMModel : SchedMachineModel { - let IssueWidth = 2; // Allows 2 instructions per scheduling group. - let MinLatency = 1; // InstrStage cycles overrides MinLatency. - // OperandCycles may be used for expected latency. - let LoadLatency = 3; // Expected cycles, may be overriden by OperandCycles. - let HighLatency = 30;// Expected, may be overriden by OperandCycles. - let Itineraries = SLMItineraries; -} +def : WriteRes { let Latency = 100; } +def : WriteRes { let Latency = 100; } +def : WriteRes; +def : WriteRes; + +// AVX is not supported on that architecture, but we should define the basic +// scheduling resources anyway. +def : WriteRes; +defm : SMWriteResPair; +defm : SMWriteResPair; +defm : SMWriteResPair; +defm : SMWriteResPair; +defm : SMWriteResPair; +} // SchedModel -- 2.34.1