RISCVCodeGenPrepare.cpp

Go to the documentation of this file.

//===----- RISCVCodeGenPrepare.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 is a RISC-V specific version of CodeGenPrepare.
// It munges the code in the input function to better prepare it for
// SelectionDAG-based code generation. This works around limitations in it's
// basic-block-at-a-time approach.
//
//===----------------------------------------------------------------------===//
 
#include "RISCV.h"
#include "RISCVTargetMachine.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/Analysis/ValueTracking.h"
#include "llvm/CodeGen/TargetPassConfig.h"
#include "llvm/IR/Dominators.h"
#include "llvm/IR/IRBuilder.h"
#include "llvm/IR/InstVisitor.h"
#include "llvm/IR/IntrinsicInst.h"
#include "llvm/IR/Intrinsics.h"
#include "llvm/IR/PatternMatch.h"
#include "llvm/InitializePasses.h"
#include "llvm/Pass.h"
#include "llvm/Transforms/Utils/Local.h"
 
using namespace llvm;
 
#define DEBUG_TYPE "riscv-codegenprepare"
#define PASS_NAME "RISC-V CodeGenPrepare"
 
namespace {
class RISCVCodeGenPrepareImpl
    : public InstVisitor<RISCVCodeGenPrepareImpl, bool> {
  Function &F;
  const DataLayout *DL;
  const DominatorTree *DT;
  const RISCVSubtarget *ST;
 
public:
  RISCVCodeGenPrepareImpl(Function &F, const DominatorTree *DT,
                          const RISCVSubtarget *ST)
      : F(F), DL(&F.getDataLayout()), DT(DT), ST(ST) {}
  bool run();
  bool visitInstruction(Instruction &I) { return false; }
  bool visitAnd(BinaryOperator &BO);
  bool visitMul(BinaryOperator &BO);
  bool visitIntrinsicInst(IntrinsicInst &I);
  bool expandVPStrideLoad(IntrinsicInst &I);
  bool expandMulReduction(IntrinsicInst &I);
  bool widenVPMerge(Instruction *I);
  bool visitFreezeInst(FreezeInst &BO);
};
} // namespace
 
namespace {
class RISCVCodeGenPrepareLegacy : public FunctionPass {
public:
  static char ID;
 
  RISCVCodeGenPrepareLegacy() : FunctionPass(ID) {}
 
  bool runOnFunction(Function &F) override;
  StringRef getPassName() const override { return PASS_NAME; }
 
  void getAnalysisUsage(AnalysisUsage &AU) const override {
    AU.setPreservesCFG();
    AU.addRequired<DominatorTreeWrapperPass>();
    AU.addRequired<TargetPassConfig>();
  }
};
} // namespace
 
// Try to optimize (i64 (and (zext/sext (i32 X), C1))) if C1 has bit 31 set,
// but bits 63:32 are zero. If we know that bit 31 of X is 0, we can fill
// the upper 32 bits with ones.
bool RISCVCodeGenPrepareImpl::visitAnd(BinaryOperator &BO) {
  if (!ST->is64Bit())
    return false;
 
  if (!BO.getType()->isIntegerTy(64))
    return false;
 
  using namespace PatternMatch;
 
  // Left hand side should be a zext nneg.
  Value *LHSSrc;
  if (!match(BO.getOperand(0), m_NNegZExt(m_Value(LHSSrc))))
    return false;
 
  if (!LHSSrc->getType()->isIntegerTy(32))
    return false;
 
  // Right hand side should be a constant.
  Value *RHS = BO.getOperand(1);
 
  auto *CI = dyn_cast<ConstantInt>(RHS);
  if (!CI)
    return false;
  uint64_t C = CI->getZExtValue();
 
  // Look for constants that fit in 32 bits but not simm12, and can be made
  // into simm12 by sign extending bit 31. This will allow use of ANDI.
  // TODO: Is worth making simm32?
  if (!isUInt<32>(C) || isInt<12>(C) || !isInt<12>(SignExtend64<32>(C)))
    return false;
 
  // Sign extend the constant and replace the And operand.
  C = SignExtend64<32>(C);
  BO.setOperand(1, ConstantInt::get(RHS->getType(), C));
 
  return true;
}
 
// Look for (mul (zext X), (splat Y)) where zext is doubling the
// element size and Y is known to be zero extended. Replace with
// (mul (zext X), (zext (splat (trunc Y)))) to encourage the use of
// widening multiply. Only do this if the splat isn't already in the same
// basic block as the mul. We put the zext with the mul and the new
// splat and trunc in the basic block with the original splat.
// We only need to do this when sinkSplatOperands() is false, otherwise
// CodeGenPrepare will just sink the splat and SelectionDAG can sort it out.
// TODO: Extend to sign extend or mismatched extends.
bool RISCVCodeGenPrepareImpl::visitMul(BinaryOperator &BO) {
  using namespace PatternMatch;
 
  if (!ST->hasVInstructions() || ST->sinkSplatOperands())
    return false;
 
  Value *LHS = BO.getOperand(0);
  Value *RHS = BO.getOperand(1);
 
  // Canonicalize zext to the LHS.
  if (isa<ZExtInst>(RHS))
    std::swap(LHS, RHS);
 
  // LHS should be a zext.
  Value *ZExtSrc;
  if (!match(LHS, m_ZExt(m_Value(ZExtSrc))))
    return false;
 
  // RHS should be a splat shuffle.
  Value *SplatVal;
  Instruction *RHSI;
  if (!match(RHS, m_Instruction(RHSI, m_OneUse(m_Splat(m_Value(SplatVal))))))
    return false;
 
  // zext should be in the same basic block as the mul.
  if (cast<Instruction>(LHS)->getParent() != BO.getParent())
    return false;
 
  // The extend should be doubling.
  unsigned Size = BO.getType()->getScalarSizeInBits();
  unsigned SrcSize = ZExtSrc->getType()->getScalarSizeInBits();
  if (Size != SrcSize * 2)
    return false;
 
  // Types should be legal.
  if (Size != 64 && Size != 32 && Size != 16)
    return false;
  if (Size == 64 && !ST->hasVInstructionsI64())
    return false;
 
  // Splat should be in another basic block.
  if (RHSI->getParent() == BO.getParent())
    return false;
 
  // Make sure we can freely truncate the value.
  KnownBits Known = computeKnownBits(SplatVal, *DL);
  if (Known.countMaxActiveBits() > SrcSize)
    return false;
 
  VectorType *VecTy = cast<VectorType>(ZExtSrc->getType());
  Type *ScalarTy = VecTy->getElementType();
  IRBuilder<> Builder(RHSI);
  Value *Splat = Builder.CreateVectorSplat(
      VecTy->getElementCount(), Builder.CreateTrunc(SplatVal, ScalarTy));
 
  Builder.SetInsertPoint(&BO);
  Value *NewZExt = Builder.CreateZExt(Splat, BO.getType());
 
  RHSI->replaceAllUsesWith(NewZExt);
  RHSI->eraseFromParent();
 
  return true;
}
 
// With EVL tail folding, an AnyOf reduction will generate an i1 vp.merge like
// follows:
//
// loop:
//   %phi = phi <vscale x 4 x i1> [zeroinitializer, %entry], [%freeze, %loop]
//   %cmp = icmp ...
//   %rec = call <vscale x 4 x i1> @llvm.vp.merge(%cmp, i1 true, %phi, %evl)
//   %freeze = freeze <vscale x 4 x i1> %rec [optional]
//   ...
// middle:
//   %res = call i1 @llvm.vector.reduce.or(<vscale x 4 x i1> %freeze)
//
// However RVV doesn't have any tail undisturbed mask instructions and so we
// need a convoluted sequence of mask instructions to lower the i1 vp.merge: see
// llvm/test/CodeGen/RISCV/rvv/vpmerge-sdnode.ll.
//
// To avoid that this widens the i1 vp.merge to an i8 vp.merge, which will
// generate a single vmerge.vim:
//
// loop:
//   %phi = phi <vscale x 4 x i8> [zeroinitializer, %entry], [%freeze, %loop]
//   %cmp = icmp ...
//   %rec = call <vscale x 4 x i8> @llvm.vp.merge(%cmp, i8 true, %phi, %evl)
//   %freeze = freeze <vscale x 4 x i8> %rec
//   %trunc = trunc <vscale x 4 x i8> %freeze to <vscale x 4 x i1>
//   ...
// middle:
//   %res = call i1 @llvm.vector.reduce.or(<vscale x 4 x i1> %trunc)
//
// The trunc will normally be sunk outside of the loop, but even if there are
// users inside the loop it is still profitable.
bool RISCVCodeGenPrepareImpl::widenVPMerge(Instruction *Root) {
  if (!Root->getType()->getScalarType()->isIntegerTy(1))
    return false;
 
  Value *Mask, *True, *EVL;
  PHINode *Phi;
  using namespace PatternMatch;
  auto m_VPMerge = m_Intrinsic<Intrinsic::vp_merge>(
      m_Value(Mask), m_Value(True), m_Phi(Phi), m_Value(EVL));
  if (!match(Root, m_CombineOr(m_VPMerge, m_Freeze(m_VPMerge))))
    return false;
 
  if (!Phi->hasOneUse() || Phi->getNumIncomingValues() != 2 ||
      !match(Phi->getIncomingValue(0), m_Zero()) ||
      Phi->getIncomingValue(1) != Root)
    return false;
 
  Type *WideTy =
      VectorType::get(IntegerType::getInt8Ty(Root->getContext()),
                      cast<VectorType>(Root->getType())->getElementCount());
 
  IRBuilder<> Builder(Phi);
  PHINode *WidePhi = Builder.CreatePHI(WideTy, 2);
  WidePhi->addIncoming(ConstantAggregateZero::get(WideTy),
                       Phi->getIncomingBlock(0));
  Builder.SetInsertPoint(Root);
  Value *WideTrue = Builder.CreateZExt(True, WideTy);
  Value *WideMerge = Builder.CreateIntrinsic(Intrinsic::vp_merge, {WideTy},
                                             {Mask, WideTrue, WidePhi, EVL});
  if (isa<FreezeInst>(Root))
    WideMerge = Builder.CreateFreeze(WideMerge);
  WidePhi->addIncoming(WideMerge, Phi->getIncomingBlock(1));
  Value *Trunc = Builder.CreateTrunc(WideMerge, Root->getType());
 
  Root->replaceAllUsesWith(Trunc);
 
  // Break the cycle and delete the old chain.
  Phi->setIncomingValue(1, Phi->getIncomingValue(0));
  llvm::RecursivelyDeleteTriviallyDeadInstructions(Root);
 
  return true;
}
 
bool RISCVCodeGenPrepareImpl::visitFreezeInst(FreezeInst &I) {
  if (auto *II = dyn_cast<IntrinsicInst>(I.getOperand(0)))
    if (II->getIntrinsicID() == Intrinsic::vp_merge)
      return widenVPMerge(&I);
  return false;
}
 
// LLVM vector reduction intrinsics return a scalar result, but on RISC-V vector
// reduction instructions write the result in the first element of a vector
// register. So when a reduction in a loop uses a scalar phi, we end up with
// unnecessary scalar moves:
//
// loop:
// vfmv.s.f v10, fa0
// vfredosum.vs v8, v8, v10
// vfmv.f.s fa0, v8
//
// This mainly affects ordered fadd reductions and VP reductions that have a
// scalar start value, since other types of reduction typically use element-wise
// vectorisation in the loop body. This tries to vectorize any scalar phis that
// feed into these reductions:
//
// loop:
// %phi = phi <float> [ ..., %entry ], [ %acc, %loop ]
// %acc = call float @llvm.vector.reduce.fadd.nxv2f32(float %phi,
//                                                    <vscale x 2 x float> %vec)
//
// ->
//
// loop:
// %phi = phi <vscale x 2 x float> [ ..., %entry ], [ %acc.vec, %loop ]
// %phi.scalar = extractelement <vscale x 2 x float> %phi, i64 0
// %acc = call float @llvm.vector.reduce.fadd.nxv2f32(float %x,
//                                                    <vscale x 2 x float> %vec)
// %acc.vec = insertelement <vscale x 2 x float> poison, float %acc.next, i64 0
//
// Which eliminates the scalar -> vector -> scalar crossing during instruction
// selection.
bool RISCVCodeGenPrepareImpl::visitIntrinsicInst(IntrinsicInst &I) {
  if (expandVPStrideLoad(I))
    return true;
 
  if (expandMulReduction(I))
    return true;
 
  if (widenVPMerge(&I))
    return true;
 
  if (I.getIntrinsicID() != Intrinsic::vector_reduce_fadd &&
      !isa<VPReductionIntrinsic>(&I))
    return false;
 
  auto *PHI = dyn_cast<PHINode>(I.getOperand(0));
  if (!PHI || !PHI->hasOneUse() ||
      !llvm::is_contained(PHI->incoming_values(), &I))
    return false;
 
  Type *VecTy = I.getOperand(1)->getType();
  IRBuilder<> Builder(PHI);
  auto *VecPHI = Builder.CreatePHI(VecTy, PHI->getNumIncomingValues());
 
  for (auto *BB : PHI->blocks()) {
    Builder.SetInsertPoint(BB->getTerminator());
    Value *InsertElt = Builder.CreateInsertElement(
        VecTy, PHI->getIncomingValueForBlock(BB), (uint64_t)0);
    VecPHI->addIncoming(InsertElt, BB);
  }
 
  Builder.SetInsertPoint(&I);
  I.setOperand(0, Builder.CreateExtractElement(VecPHI, (uint64_t)0));
 
  PHI->eraseFromParent();
 
  return true;
}
 
// Extract pieces of size PieceEC from Vec, then build a binary tree of
// element-wise multiplies reducing to a single piece.
static Value *buildMulTree(IRBuilder<> &Builder, ElementCount PieceEC,
                           Value *Vec) {
  auto *VecTy = cast<VectorType>(Vec->getType());
  auto *PieceTy = VectorType::get(VecTy->getElementType(), PieceEC);
  unsigned PieceElts = PieceEC.getKnownMinValue();
  unsigned NumPieces = VecTy->getElementCount().getKnownMinValue() / PieceElts;
  assert(isPowerOf2_32(NumPieces));
 
  SmallVector<Value *, 8> Pieces(NumPieces);
  for (unsigned i = 0; i < NumPieces; i++)
    Pieces[i] = Builder.CreateExtractVector(PieceTy, Vec, i * PieceElts);
 
  while (Pieces.size() > 1) {
    for (unsigned i = 0; i < Pieces.size() / 2; i++)
      Pieces[i] =
          Builder.CreateMul(Pieces[i * 2], Pieces[i * 2 + 1], "bin.rdx");
    Pieces.truncate(Pieces.size() / 2);
  }
  return Pieces[0];
}
 
// Partially expand a vector_reduce_mul wider than M1 to reduce
// register pressure and the number of vsetvlis required.
bool RISCVCodeGenPrepareImpl::expandMulReduction(IntrinsicInst &II) {
  if (II.getIntrinsicID() != Intrinsic::vector_reduce_mul)
    return false;
 
  if (!ST->hasVInstructions())
    return false;
 
  Value *TmpVec = II.getArgOperand(0);
  auto *VecTy = cast<VectorType>(TmpVec->getType());
  unsigned EltSize = VecTy->getScalarSizeInBits();
 
  if (auto *ScalTy = dyn_cast<ScalableVectorType>(VecTy)) {
    unsigned MinElts = ScalTy->getMinNumElements();
 
    if (auto VLen = ST->getRealVLen()) {
      // If VLEN is exactly known, convert to a fixed vector reduction and
      // recurse to let the fixed path handle it (shuffle reduction instead
      // of a scalar loop).
      unsigned VScale = *VLen / RISCV::RVVBitsPerBlock;
      auto *FixedTy =
          FixedVectorType::get(VecTy->getElementType(), MinElts * VScale);
      IRBuilder<> Builder(&II);
      Value *Fixed = Builder.CreateExtractVector(FixedTy, TmpVec, (uint64_t)0);
      auto *FixedRdx = cast<IntrinsicInst>(Builder.CreateIntrinsic(
          Intrinsic::vector_reduce_mul, {FixedTy}, {Fixed}));
      II.replaceAllUsesWith(FixedRdx);
      II.eraseFromParent();
      expandMulReduction(*FixedRdx);
      return true;
    }
 
    unsigned M1MinElts = RISCV::RVVBitsPerBlock / EltSize;
    if (MinElts <= M1MinElts || !isPowerOf2_32(MinElts / M1MinElts))
      return false;
 
    IRBuilder<> Builder(&II);
    auto M1EC = ElementCount::getScalable(M1MinElts);
    Value *Reduced = buildMulTree(Builder, M1EC, TmpVec);
    Value *Rdx = Builder.CreateIntrinsic(Intrinsic::vector_reduce_mul,
                                         {Reduced->getType()}, {Reduced});
    II.replaceAllUsesWith(Rdx);
    II.eraseFromParent();
    return true;
  }
 
  unsigned VF = cast<FixedVectorType>(VecTy)->getNumElements();
  unsigned MinVLen = ST->getRealMinVLen();
  unsigned M1VF = MinVLen / EltSize;
 
  if (!isPowerOf2_32(VF) || VF <= M1VF)
    return false;
 
  IRBuilder<> Builder(&II);
  auto M1EC = ElementCount::getFixed(M1VF);
  auto *M1Ty = VectorType::get(VecTy->getElementType(), M1EC);
 
  // When VLEN is exactly known, extract m1 pieces and build a mul tree.
  // This greatly reduces register pressure during the reduction, and
  // avoids all but one vsetvli (the one from original LMUL to m1).
  // TODO: Generalize to handle the splitting case.
  if (MinVLen == ST->getRealMaxVLen() && VF <= 8 * M1VF) {
    TmpVec = buildMulTree(Builder, M1EC, TmpVec);
  } else {
    // For non-exact VLEN, shuffle-reduce at the original vector width down to
    // m1, then extract.  This prioritizes reducing the number of vsetvli
    // over maximal reduction of LMUL for the intermediate states.
    SmallVector<int, 32> ShuffleMask(VF);
    for (unsigned LiveElts = VF; LiveElts > M1VF; LiveElts /= 2) {
      unsigned Half = LiveElts / 2;
      std::iota(ShuffleMask.begin(), ShuffleMask.begin() + Half, Half);
      std::fill(ShuffleMask.begin() + Half, ShuffleMask.end(), -1);
      Value *Shuf =
          Builder.CreateShuffleVector(TmpVec, ShuffleMask, "rdx.shuf");
      TmpVec = Builder.CreateMul(TmpVec, Shuf, "bin.rdx");
    }
    // Extract the M1-sized subvector and emit the final reduction intrinsic.
    // This is the reason we're here - to force a vsetvli toggle once at m1.
    TmpVec = Builder.CreateExtractVector(M1Ty, TmpVec, (uint64_t)0, "rdx.sub");
  }
 
  Value *Rdx =
      Builder.CreateIntrinsic(Intrinsic::vector_reduce_mul, {M1Ty}, {TmpVec});
  II.replaceAllUsesWith(Rdx);
  II.eraseFromParent();
  return true;
}
 
// Always expand zero strided loads so we match more .vx splat patterns, even if
// we have +optimized-zero-stride-loads. RISCVDAGToDAGISel::Select will convert
// it back to a strided load if it's optimized.
bool RISCVCodeGenPrepareImpl::expandVPStrideLoad(IntrinsicInst &II) {
  Value *BasePtr, *VL;
 
  using namespace PatternMatch;
  if (!match(&II, m_Intrinsic<Intrinsic::experimental_vp_strided_load>(
                      m_Value(BasePtr), m_Zero(), m_AllOnes(), m_Value(VL))))
    return false;
 
  // If SEW>XLEN then a splat will get lowered as a zero strided load anyway, so
  // avoid expanding here.
  if (II.getType()->getScalarSizeInBits() > ST->getXLen())
    return false;
 
  if (!isKnownNonZero(VL, {*DL, DT, nullptr, &II}))
    return false;
 
  auto *VTy = cast<VectorType>(II.getType());
 
  IRBuilder<> Builder(&II);
  Type *STy = VTy->getElementType();
  Value *Val = Builder.CreateLoad(STy, BasePtr);
  Value *Res = Builder.CreateIntrinsic(
      Intrinsic::vp_merge, VTy,
      {II.getOperand(2), Builder.CreateVectorSplat(VTy->getElementCount(), Val),
       PoisonValue::get(VTy), VL});
 
  II.replaceAllUsesWith(Res);
  II.eraseFromParent();
  return true;
}
 
bool RISCVCodeGenPrepareImpl::run() {
  bool MadeChange = false;
  for (auto &BB : F)
    for (Instruction &I : llvm::make_early_inc_range(BB))
      MadeChange |= visit(I);
 
  return MadeChange;
}
 
bool RISCVCodeGenPrepareLegacy::runOnFunction(Function &F) {
  if (skipFunction(F))
    return false;
 
  auto &TPC = getAnalysis<TargetPassConfig>();
  auto &TM = TPC.getTM<RISCVTargetMachine>();
  auto ST = &TM.getSubtarget<RISCVSubtarget>(F);
  auto DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
 
  RISCVCodeGenPrepareImpl RVCGP(F, DT, ST);
  return RVCGP.run();
}
 
INITIALIZE_PASS_BEGIN(RISCVCodeGenPrepareLegacy, DEBUG_TYPE, PASS_NAME, false,
                      false)
INITIALIZE_PASS_DEPENDENCY(TargetPassConfig)
INITIALIZE_PASS_END(RISCVCodeGenPrepareLegacy, DEBUG_TYPE, PASS_NAME, false,
                    false)
 
char RISCVCodeGenPrepareLegacy::ID = 0;
 
FunctionPass *llvm::createRISCVCodeGenPrepareLegacyPass() {
  return new RISCVCodeGenPrepareLegacy();
}
 
PreservedAnalyses RISCVCodeGenPreparePass::run(Function &F,
                                               FunctionAnalysisManager &FAM) {
  DominatorTree *DT = &FAM.getResult<DominatorTreeAnalysis>(F);
  auto ST = &TM->getSubtarget<RISCVSubtarget>(F);
  bool Changed = RISCVCodeGenPrepareImpl(F, DT, ST).run();
  if (!Changed)
    return PreservedAnalyses::all();
 
  PreservedAnalyses PA = PreservedAnalyses::none();
  PA.preserveSet<CFGAnalyses>();
  return PA;
}