llvm/lib/Target/AMDGPU/AMDGPUPreLegalizerCombiner.cpp

//=== lib/CodeGen/GlobalISel/AMDGPUPreLegalizerCombiner.cpp ---------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// This pass does combining of machine instructions at the generic MI level,
// before the legalizer.
//
//===----------------------------------------------------------------------===//

#include "AMDGPU.h"
#include "AMDGPUCombinerHelper.h"
#include "AMDGPULegalizerInfo.h"
#include "GCNSubtarget.h"
#include "llvm/CodeGen/GlobalISel/CSEInfo.h"
#include "llvm/CodeGen/GlobalISel/Combiner.h"
#include "llvm/CodeGen/GlobalISel/CombinerHelper.h"
#include "llvm/CodeGen/GlobalISel/CombinerInfo.h"
#include "llvm/CodeGen/GlobalISel/GIMatchTableExecutorImpl.h"
#include "llvm/CodeGen/GlobalISel/GISelValueTracking.h"
#include "llvm/CodeGen/GlobalISel/MIPatternMatch.h"
#include "llvm/CodeGen/MachineDominators.h"
#include "llvm/CodeGen/MachineFunctionAnalysisManager.h"
#include "llvm/CodeGen/MachinePassManager.h"
#include "llvm/CodeGen/TargetPassConfig.h"
#include "llvm/Target/TargetMachine.h"

#define GET_GICOMBINER_DEPS
#include "AMDGPUGenPreLegalizeGICombiner.inc"
#undef GET_GICOMBINER_DEPS

#define DEBUG_TYPE "amdgpu-prelegalizer-combiner"

using namespace llvm;
using namespace MIPatternMatch;
namespace {

#define GET_GICOMBINER_TYPES
#include "AMDGPUGenPreLegalizeGICombiner.inc"
#undef GET_GICOMBINER_TYPES

class AMDGPUPreLegalizerCombinerImpl : public Combiner {
protected:
  const AMDGPUPreLegalizerCombinerImplRuleConfig &RuleConfig;
  const GCNSubtarget &STI;
  const AMDGPUCombinerHelper Helper;

public:
  AMDGPUPreLegalizerCombinerImpl(
      MachineFunction &MF, CombinerInfo &CInfo, GISelValueTracking &VT,
      GISelCSEInfo *CSEInfo,
      const AMDGPUPreLegalizerCombinerImplRuleConfig &RuleConfig,
      const GCNSubtarget &STI, MachineDominatorTree *MDT,
      const LegalizerInfo *LI);

  static const char *getName() { return "AMDGPUPreLegalizerCombinerImpl"; }

  bool tryCombineAllImpl(MachineInstr &MI) const;
  bool tryCombineAll(MachineInstr &I) const override;

  struct ClampI64ToI16MatchInfo {
    int64_t Cmp1 = 0;
    int64_t Cmp2 = 0;
    Register Origin;
  };

  bool matchClampI64ToI16(MachineInstr &MI, const MachineRegisterInfo &MRI,
                          const MachineFunction &MF,
                          ClampI64ToI16MatchInfo &MatchInfo) const;

  void applyClampI64ToI16(MachineInstr &MI,
                          const ClampI64ToI16MatchInfo &MatchInfo) const;

private:
#define GET_GICOMBINER_CLASS_MEMBERS
#define AMDGPUSubtarget GCNSubtarget
#include "AMDGPUGenPreLegalizeGICombiner.inc"
#undef GET_GICOMBINER_CLASS_MEMBERS
#undef AMDGPUSubtarget
};

#define GET_GICOMBINER_IMPL
#define AMDGPUSubtarget GCNSubtarget
#include "AMDGPUGenPreLegalizeGICombiner.inc"
#undef AMDGPUSubtarget
#undef GET_GICOMBINER_IMPL

AMDGPUPreLegalizerCombinerImpl::AMDGPUPreLegalizerCombinerImpl(
    MachineFunction &MF, CombinerInfo &CInfo, GISelValueTracking &VT,
    GISelCSEInfo *CSEInfo,
    const AMDGPUPreLegalizerCombinerImplRuleConfig &RuleConfig,
    const GCNSubtarget &STI, MachineDominatorTree *MDT, const LegalizerInfo *LI)
    : Combiner(MF, CInfo, &VT, CSEInfo), RuleConfig(RuleConfig), STI(STI),
      Helper(Observer, B, /*IsPreLegalize*/ true, &VT, MDT, LI, STI),
#define GET_GICOMBINER_CONSTRUCTOR_INITS
#include "AMDGPUGenPreLegalizeGICombiner.inc"
#undef GET_GICOMBINER_CONSTRUCTOR_INITS
{
}

bool AMDGPUPreLegalizerCombinerImpl::tryCombineAll(MachineInstr &MI) const {
  if (tryCombineAllImpl(MI))
    return true;
  return false;
}

bool AMDGPUPreLegalizerCombinerImpl::matchClampI64ToI16(
    MachineInstr &MI, const MachineRegisterInfo &MRI, const MachineFunction &MF,
    ClampI64ToI16MatchInfo &MatchInfo) const {
  assert(MI.getOpcode() == TargetOpcode::G_TRUNC && "Invalid instruction!");

  // Try to find a pattern where an i64 value should get clamped to short.
  const LLT SrcType = MRI.getType(MI.getOperand(1).getReg());
  if (SrcType != LLT::scalar(64))
    return false;

  const LLT DstType = MRI.getType(MI.getOperand(0).getReg());
  if (DstType != LLT::scalar(16))
    return false;

  Register Base;

  // Lo must not exceed Hi: with inverted bounds smin(smax(X, Lo), Hi) is
  // constant, but the med3 built below would still clamp X to [Hi, Lo].
  auto IsApplicableForCombine = [&MatchInfo](bool OuterIsMin) -> bool {
    const int64_t Lo = OuterIsMin ? MatchInfo.Cmp2 : MatchInfo.Cmp1;
    const int64_t Hi = OuterIsMin ? MatchInfo.Cmp1 : MatchInfo.Cmp2;

    // Range-check first so Hi - Lo below can't overflow.
    const int64_t Min = std::numeric_limits<int16_t>::min();
    const int64_t Max = std::numeric_limits<int16_t>::max();
    if (Lo < Min || Lo > Max || Hi < Min || Hi > Max)
      return false;

    // Reject inverted bounds, and bounds so close there is no need to clamp.
    return Hi - Lo > 1;
  };

  // Try to match a combination of min / max MIR opcodes.
  if (mi_match(MI.getOperand(1).getReg(), MRI,
               m_GSMin(m_Reg(Base), m_ICst(MatchInfo.Cmp1)))) {
    if (mi_match(Base, MRI,
                 m_GSMax(m_Reg(MatchInfo.Origin), m_ICst(MatchInfo.Cmp2)))) {
      return IsApplicableForCombine(/*OuterIsMin=*/true);
    }
  }

  if (mi_match(MI.getOperand(1).getReg(), MRI,
               m_GSMax(m_Reg(Base), m_ICst(MatchInfo.Cmp1)))) {
    if (mi_match(Base, MRI,
                 m_GSMin(m_Reg(MatchInfo.Origin), m_ICst(MatchInfo.Cmp2)))) {
      return IsApplicableForCombine(/*OuterIsMin=*/false);
    }
  }

  return false;
}

// We want to find a combination of instructions that
// gets generated when an i64 gets clamped to i16.
// The corresponding pattern is:
// G_MAX / G_MAX for i16 <= G_TRUNC i64.
// This can be efficiently written as following:
// v_cvt_pk_i16_i32 v0, v0, v1
// v_med3_i32 v0, Clamp_Min, v0, Clamp_Max
void AMDGPUPreLegalizerCombinerImpl::applyClampI64ToI16(
    MachineInstr &MI, const ClampI64ToI16MatchInfo &MatchInfo) const {

  Register Src = MatchInfo.Origin;
  assert(MI.getMF()->getRegInfo().getType(Src) == LLT::scalar(64));
  const LLT I32 = LLT::integer(32);

  auto Unmerge = B.buildUnmerge(I32, Src);

  assert(MI.getOpcode() != AMDGPU::G_AMDGPU_CVT_PK_I16_I32);

  const LLT V2S16 = LLT::fixed_vector(2, 16);
  auto CvtPk =
      B.buildInstr(AMDGPU::G_AMDGPU_CVT_PK_I16_I32, {V2S16},
                   {Unmerge.getReg(0), Unmerge.getReg(1)}, MI.getFlags());

  auto MinBoundary = std::min(MatchInfo.Cmp1, MatchInfo.Cmp2);
  auto MaxBoundary = std::max(MatchInfo.Cmp1, MatchInfo.Cmp2);
  auto MinBoundaryDst = B.buildConstant(I32, MinBoundary);
  auto MaxBoundaryDst = B.buildConstant(I32, MaxBoundary);

  auto Bitcast = B.buildBitcast({I32}, CvtPk);

  auto Med3 = B.buildInstr(
      AMDGPU::G_AMDGPU_SMED3, {I32},
      {MinBoundaryDst.getReg(0), Bitcast.getReg(0), MaxBoundaryDst.getReg(0)},
      MI.getFlags());

  B.buildTrunc(MI.getOperand(0).getReg(), Med3);

  MI.eraseFromParent();
}

static bool runCombiner(MachineFunction &MF,
                        function_ref<GISelCSEInfo *()> GetCSEInfo,
                        function_ref<GISelValueTracking *()> GetVT,
                        function_ref<MachineDominatorTree *()> GetMDT,
                        bool EnableOpt) {
  AMDGPUPreLegalizerCombinerImplRuleConfig RuleConfig;
  if (!RuleConfig.parseCommandLineOption())
    reportFatalUsageError("Invalid rule identifier");

  // If the ISel pipeline failed, do not bother running that pass.
  if (MF.getProperties().hasFailedISel())
    return false;

  const GCNSubtarget &STI = MF.getSubtarget<GCNSubtarget>();
  const Function &F = MF.getFunction();
  CombinerInfo CInfo(/*AllowIllegalOps=*/true, /*ShouldLegalizeIllegal=*/false,
                     nullptr, EnableOpt, F.hasOptSize(), F.hasMinSize());
  // Disable fixed-point iteration to reduce compile-time
  CInfo.MaxIterations = 1;
  CInfo.ObserverLvl = CombinerInfo::ObserverLevel::SinglePass;
  // This is the first Combiner, so the input IR might contain dead
  // instructions.
  CInfo.EnableFullDCE = true;

  GISelValueTracking *VT = GetVT();
  GISelCSEInfo *CSEInfo = GetCSEInfo();
  MachineDominatorTree *MDT = GetMDT();
  AMDGPUPreLegalizerCombinerImpl Impl(MF, CInfo, *VT, CSEInfo, RuleConfig, STI,
                                      MDT, STI.getLegalizerInfo());
  return Impl.combineMachineInstrs();
}

// Pass boilerplate
// ================

class AMDGPUPreLegalizerCombinerLegacy : public MachineFunctionPass {
public:
  static char ID;

  AMDGPUPreLegalizerCombinerLegacy(bool IsOptLevelNone = false)
      : MachineFunctionPass(ID), IsOptLevelNone(IsOptLevelNone) {}

  StringRef getPassName() const override {
    return "AMDGPUPreLegalizerCombiner";
  }

  bool runOnMachineFunction(MachineFunction &MF) override;

  void getAnalysisUsage(AnalysisUsage &AU) const override;

private:
  bool IsOptLevelNone;
};
} // end anonymous namespace

void AMDGPUPreLegalizerCombinerLegacy::getAnalysisUsage(
    AnalysisUsage &AU) const {
  AU.addRequired<TargetPassConfig>();
  AU.setPreservesCFG();
  getSelectionDAGFallbackAnalysisUsage(AU);
  AU.addRequired<GISelValueTrackingAnalysisLegacy>();
  AU.addPreserved<GISelValueTrackingAnalysisLegacy>();
  if (!IsOptLevelNone) {
    AU.addRequired<MachineDominatorTreeWrapperPass>();
  }

  AU.addRequired<GISelCSEAnalysisWrapperPass>();
  AU.addPreserved<GISelCSEAnalysisWrapperPass>();
  MachineFunctionPass::getAnalysisUsage(AU);
}

bool AMDGPUPreLegalizerCombinerLegacy::runOnMachineFunction(
    MachineFunction &MF) {
  const Function &F = MF.getFunction();
  bool EnableOpt =
      MF.getTarget().getOptLevel() != CodeGenOptLevel::None && !skipFunction(F);
  return runCombiner(
      MF,
      [&]() {
        // Enable CSE.
        GISelCSEAnalysisWrapper &Wrapper =
            getAnalysis<GISelCSEAnalysisWrapperPass>().getCSEWrapper();
        return &Wrapper.get(getAnalysis<TargetPassConfig>().getCSEConfig());
      },
      [&]() {
        return &getAnalysis<GISelValueTrackingAnalysisLegacy>().get(MF);
      },
      [&]() -> MachineDominatorTree * {
        return IsOptLevelNone ? nullptr
                              : &getAnalysis<MachineDominatorTreeWrapperPass>()
                                     .getDomTree();
      },
      EnableOpt);
}

char AMDGPUPreLegalizerCombinerLegacy::ID = 0;
INITIALIZE_PASS_BEGIN(AMDGPUPreLegalizerCombinerLegacy, DEBUG_TYPE,
                      "Combine AMDGPU machine instrs before legalization",
                      false, false)
INITIALIZE_PASS_DEPENDENCY(TargetPassConfig)
INITIALIZE_PASS_DEPENDENCY(GISelValueTrackingAnalysisLegacy)
INITIALIZE_PASS_END(AMDGPUPreLegalizerCombinerLegacy, DEBUG_TYPE,
                    "Combine AMDGPU machine instrs before legalization", false,
                    false)

FunctionPass *
llvm::createAMDGPUPreLegalizeCombinerLegacyPass(bool IsOptLevelNone) {
  return new AMDGPUPreLegalizerCombinerLegacy(IsOptLevelNone);
}

PreservedAnalyses
AMDGPUPreLegalizerCombinerPass::run(MachineFunction &MF,
                                    MachineFunctionAnalysisManager &MFAM) {
  bool IsOptLevelNone = MF.getTarget().getOptLevel() == CodeGenOptLevel::None ||
                        shouldSkipOptimizationForOptBisect(MF.getFunction());

  if (!runCombiner(
          MF, [&]() { return MFAM.getResult<GISelCSEAnalysis>(MF).get(); },
          [&]() { return &MFAM.getResult<GISelValueTrackingAnalysis>(MF); },
          [&]() -> MachineDominatorTree * {
            return IsOptLevelNone
                       ? nullptr
                       : &MFAM.getResult<MachineDominatorTreeAnalysis>(MF);
          },
          /*EnableOpt=*/!IsOptLevelNone))
    return PreservedAnalyses::all();

  PreservedAnalyses PA = getMachineFunctionPassPreservedAnalyses();
  PA.preserveSet<CFGAnalyses>();
  PA.preserve<GISelValueTrackingAnalysis>();
  PA.preserve<GISelCSEAnalysis>();
  return PA;
}