1 files changed, 294 insertions, 7 deletions

diff --git a/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp b/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp
index b955e231699a..e6230547a9b3 100644
--- a/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp
+++ b/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp
@@ -39,6 +39,78 @@ class AMDGPUCodeGenPrepare : public FunctionPass,
   Module *Mod;
   bool HasUnsafeFPMath;
 
+  /// \brief Copies exact/nsw/nuw flags (if any) from binary operation \p I to
+  /// binary operation \p V.
+  ///
+  /// \returns Binary operation \p V.
+  Value *copyFlags(const BinaryOperator &I, Value *V) const;
+
+  /// \returns \p T's base element bit width.
+  unsigned getBaseElementBitWidth(const Type *T) const;
+
+  /// \returns Equivalent 32 bit integer type for given type \p T. For example,
+  /// if \p T is i7, then i32 is returned; if \p T is <3 x i12>, then <3 x i32>
+  /// is returned.
+  Type *getI32Ty(IRBuilder<> &B, const Type *T) const;
+
+  /// \returns True if binary operation \p I is a signed binary operation, false
+  /// otherwise.
+  bool isSigned(const BinaryOperator &I) const;
+
+  /// \returns True if the condition of 'select' operation \p I comes from a
+  /// signed 'icmp' operation, false otherwise.
+  bool isSigned(const SelectInst &I) const;
+
+  /// \returns True if type \p T needs to be promoted to 32 bit integer type,
+  /// false otherwise.
+  bool needsPromotionToI32(const Type *T) const;
+
+  /// \brief Promotes uniform binary operation \p I to equivalent 32 bit binary
+  /// operation.
+  ///
+  /// \details \p I's base element bit width must be greater than 1 and less
+  /// than or equal 16. Promotion is done by sign or zero extending operands to
+  /// 32 bits, replacing \p I with equivalent 32 bit binary operation, and
+  /// truncating the result of 32 bit binary operation back to \p I's original
+  /// type. Division operation is not promoted.
+  ///
+  /// \returns True if \p I is promoted to equivalent 32 bit binary operation,
+  /// false otherwise.
+  bool promoteUniformOpToI32(BinaryOperator &I) const;
+
+  /// \brief Promotes uniform 'icmp' operation \p I to 32 bit 'icmp' operation.
+  ///
+  /// \details \p I's base element bit width must be greater than 1 and less
+  /// than or equal 16. Promotion is done by sign or zero extending operands to
+  /// 32 bits, and replacing \p I with 32 bit 'icmp' operation.
+  ///
+  /// \returns True.
+  bool promoteUniformOpToI32(ICmpInst &I) const;
+
+  /// \brief Promotes uniform 'select' operation \p I to 32 bit 'select'
+  /// operation.
+  ///
+  /// \details \p I's base element bit width must be greater than 1 and less
+  /// than or equal 16. Promotion is done by sign or zero extending operands to
+  /// 32 bits, replacing \p I with 32 bit 'select' operation, and truncating the
+  /// result of 32 bit 'select' operation back to \p I's original type.
+  ///
+  /// \returns True.
+  bool promoteUniformOpToI32(SelectInst &I) const;
+
+  /// \brief Promotes uniform 'bitreverse' intrinsic \p I to 32 bit 'bitreverse'
+  /// intrinsic.
+  ///
+  /// \details \p I's base element bit width must be greater than 1 and less
+  /// than or equal 16. Promotion is done by zero extending the operand to 32
+  /// bits, replacing \p I with 32 bit 'bitreverse' intrinsic, shifting the
+  /// result of 32 bit 'bitreverse' intrinsic to the right with zero fill (the
+  /// shift amount is 32 minus \p I's base element bit width), and truncating
+  /// the result of the shift operation back to \p I's original type.
+  ///
+  /// \returns True.
+  bool promoteUniformBitreverseToI32(IntrinsicInst &I) const;
+
 public:
   static char ID;
   AMDGPUCodeGenPrepare(const TargetMachine *TM = nullptr) :
@@ -51,16 +123,18 @@ public:
 
   bool visitFDiv(BinaryOperator &I);
 
-  bool visitInstruction(Instruction &I) {
-    return false;
-  }
+  bool visitInstruction(Instruction &I) { return false; }
+  bool visitBinaryOperator(BinaryOperator &I);
+  bool visitICmpInst(ICmpInst &I);
+  bool visitSelectInst(SelectInst &I);
+
+  bool visitIntrinsicInst(IntrinsicInst &I);
+  bool visitBitreverseIntrinsicInst(IntrinsicInst &I);
 
   bool doInitialization(Module &M) override;
   bool runOnFunction(Function &F) override;
 
-  const char *getPassName() const override {
-    return "AMDGPU IR optimizations";
-  }
+  StringRef getPassName() const override { return "AMDGPU IR optimizations"; }
 
   void getAnalysisUsage(AnalysisUsage &AU) const override {
     AU.addRequired<DivergenceAnalysis>();
@@ -70,6 +144,171 @@ public:
 
 } // End anonymous namespace
 
+Value *AMDGPUCodeGenPrepare::copyFlags(
+    const BinaryOperator &I, Value *V) const {
+  BinaryOperator *BinOp = dyn_cast<BinaryOperator>(V);
+  if (!BinOp) // Possibly constant expression.
+    return V;
+
+  if (isa<OverflowingBinaryOperator>(BinOp)) {
+    BinOp->setHasNoSignedWrap(I.hasNoSignedWrap());
+    BinOp->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
+  } else if (isa<PossiblyExactOperator>(BinOp))
+    BinOp->setIsExact(I.isExact());
+
+  return V;
+}
+
+unsigned AMDGPUCodeGenPrepare::getBaseElementBitWidth(const Type *T) const {
+  assert(needsPromotionToI32(T) && "T does not need promotion to i32");
+
+  if (T->isIntegerTy())
+    return T->getIntegerBitWidth();
+  return cast<VectorType>(T)->getElementType()->getIntegerBitWidth();
+}
+
+Type *AMDGPUCodeGenPrepare::getI32Ty(IRBuilder<> &B, const Type *T) const {
+  assert(needsPromotionToI32(T) && "T does not need promotion to i32");
+
+  if (T->isIntegerTy())
+    return B.getInt32Ty();
+  return VectorType::get(B.getInt32Ty(), cast<VectorType>(T)->getNumElements());
+}
+
+bool AMDGPUCodeGenPrepare::isSigned(const BinaryOperator &I) const {
+  return I.getOpcode() == Instruction::AShr ||
+      I.getOpcode() == Instruction::SDiv || I.getOpcode() == Instruction::SRem;
+}
+
+bool AMDGPUCodeGenPrepare::isSigned(const SelectInst &I) const {
+  return isa<ICmpInst>(I.getOperand(0)) ?
+      cast<ICmpInst>(I.getOperand(0))->isSigned() : false;
+}
+
+bool AMDGPUCodeGenPrepare::needsPromotionToI32(const Type *T) const {
+  if (T->isIntegerTy() && T->getIntegerBitWidth() > 1 &&
+      T->getIntegerBitWidth() <= 16)
+    return true;
+  if (!T->isVectorTy())
+    return false;
+  return needsPromotionToI32(cast<VectorType>(T)->getElementType());
+}
+
+bool AMDGPUCodeGenPrepare::promoteUniformOpToI32(BinaryOperator &I) const {
+  assert(needsPromotionToI32(I.getType()) &&
+         "I does not need promotion to i32");
+
+  if (I.getOpcode() == Instruction::SDiv ||
+      I.getOpcode() == Instruction::UDiv)
+    return false;
+
+  IRBuilder<> Builder(&I);
+  Builder.SetCurrentDebugLocation(I.getDebugLoc());
+
+  Type *I32Ty = getI32Ty(Builder, I.getType());
+  Value *ExtOp0 = nullptr;
+  Value *ExtOp1 = nullptr;
+  Value *ExtRes = nullptr;
+  Value *TruncRes = nullptr;
+
+  if (isSigned(I)) {
+    ExtOp0 = Builder.CreateSExt(I.getOperand(0), I32Ty);
+    ExtOp1 = Builder.CreateSExt(I.getOperand(1), I32Ty);
+  } else {
+    ExtOp0 = Builder.CreateZExt(I.getOperand(0), I32Ty);
+    ExtOp1 = Builder.CreateZExt(I.getOperand(1), I32Ty);
+  }
+  ExtRes = copyFlags(I, Builder.CreateBinOp(I.getOpcode(), ExtOp0, ExtOp1));
+  TruncRes = Builder.CreateTrunc(ExtRes, I.getType());
+
+  I.replaceAllUsesWith(TruncRes);
+  I.eraseFromParent();
+
+  return true;
+}
+
+bool AMDGPUCodeGenPrepare::promoteUniformOpToI32(ICmpInst &I) const {
+  assert(needsPromotionToI32(I.getOperand(0)->getType()) &&
+         "I does not need promotion to i32");
+
+  IRBuilder<> Builder(&I);
+  Builder.SetCurrentDebugLocation(I.getDebugLoc());
+
+  Type *I32Ty = getI32Ty(Builder, I.getOperand(0)->getType());
+  Value *ExtOp0 = nullptr;
+  Value *ExtOp1 = nullptr;
+  Value *NewICmp  = nullptr;
+
+  if (I.isSigned()) {
+    ExtOp0 = Builder.CreateSExt(I.getOperand(0), I32Ty);
+    ExtOp1 = Builder.CreateSExt(I.getOperand(1), I32Ty);
+  } else {
+    ExtOp0 = Builder.CreateZExt(I.getOperand(0), I32Ty);
+    ExtOp1 = Builder.CreateZExt(I.getOperand(1), I32Ty);
+  }
+  NewICmp = Builder.CreateICmp(I.getPredicate(), ExtOp0, ExtOp1);
+
+  I.replaceAllUsesWith(NewICmp);
+  I.eraseFromParent();
+
+  return true;
+}
+
+bool AMDGPUCodeGenPrepare::promoteUniformOpToI32(SelectInst &I) const {
+  assert(needsPromotionToI32(I.getType()) &&
+         "I does not need promotion to i32");
+
+  IRBuilder<> Builder(&I);
+  Builder.SetCurrentDebugLocation(I.getDebugLoc());
+
+  Type *I32Ty = getI32Ty(Builder, I.getType());
+  Value *ExtOp1 = nullptr;
+  Value *ExtOp2 = nullptr;
+  Value *ExtRes = nullptr;
+  Value *TruncRes = nullptr;
+
+  if (isSigned(I)) {
+    ExtOp1 = Builder.CreateSExt(I.getOperand(1), I32Ty);
+    ExtOp2 = Builder.CreateSExt(I.getOperand(2), I32Ty);
+  } else {
+    ExtOp1 = Builder.CreateZExt(I.getOperand(1), I32Ty);
+    ExtOp2 = Builder.CreateZExt(I.getOperand(2), I32Ty);
+  }
+  ExtRes = Builder.CreateSelect(I.getOperand(0), ExtOp1, ExtOp2);
+  TruncRes = Builder.CreateTrunc(ExtRes, I.getType());
+
+  I.replaceAllUsesWith(TruncRes);
+  I.eraseFromParent();
+
+  return true;
+}
+
+bool AMDGPUCodeGenPrepare::promoteUniformBitreverseToI32(
+    IntrinsicInst &I) const {
+  assert(I.getIntrinsicID() == Intrinsic::bitreverse &&
+         "I must be bitreverse intrinsic");
+  assert(needsPromotionToI32(I.getType()) &&
+         "I does not need promotion to i32");
+
+  IRBuilder<> Builder(&I);
+  Builder.SetCurrentDebugLocation(I.getDebugLoc());
+
+  Type *I32Ty = getI32Ty(Builder, I.getType());
+  Function *I32 =
+      Intrinsic::getDeclaration(Mod, Intrinsic::bitreverse, { I32Ty });
+  Value *ExtOp = Builder.CreateZExt(I.getOperand(0), I32Ty);
+  Value *ExtRes = Builder.CreateCall(I32, { ExtOp });
+  Value *LShrOp =
+      Builder.CreateLShr(ExtRes, 32 - getBaseElementBitWidth(I.getType()));
+  Value *TruncRes =
+      Builder.CreateTrunc(LShrOp, I.getType());
+
+  I.replaceAllUsesWith(TruncRes);
+  I.eraseFromParent();
+
+  return true;
+}
+
 static bool shouldKeepFDivF32(Value *Num, bool UnsafeDiv) {
   const ConstantFP *CNum = dyn_cast<ConstantFP>(Num);
   if (!CNum)
@@ -85,7 +324,6 @@ static bool shouldKeepFDivF32(Value *Num, bool UnsafeDiv) {
 bool AMDGPUCodeGenPrepare::visitFDiv(BinaryOperator &FDiv) {
   Type *Ty = FDiv.getType();
 
-  // TODO: Handle half
   if (!Ty->getScalarType()->isFloatTy())
     return false;
 
@@ -154,6 +392,55 @@ static bool hasUnsafeFPMath(const Function &F) {
   return Attr.getValueAsString() == "true";
 }
 
+bool AMDGPUCodeGenPrepare::visitBinaryOperator(BinaryOperator &I) {
+  bool Changed = false;
+
+  if (ST->has16BitInsts() && needsPromotionToI32(I.getType()) &&
+      DA->isUniform(&I))
+    Changed |= promoteUniformOpToI32(I);
+
+  return Changed;
+}
+
+bool AMDGPUCodeGenPrepare::visitICmpInst(ICmpInst &I) {
+  bool Changed = false;
+
+  if (ST->has16BitInsts() && needsPromotionToI32(I.getOperand(0)->getType()) &&
+      DA->isUniform(&I))
+    Changed |= promoteUniformOpToI32(I);
+
+  return Changed;
+}
+
+bool AMDGPUCodeGenPrepare::visitSelectInst(SelectInst &I) {
+  bool Changed = false;
+
+  if (ST->has16BitInsts() && needsPromotionToI32(I.getType()) &&
+      DA->isUniform(&I))
+    Changed |= promoteUniformOpToI32(I);
+
+  return Changed;
+}
+
+bool AMDGPUCodeGenPrepare::visitIntrinsicInst(IntrinsicInst &I) {
+  switch (I.getIntrinsicID()) {
+  case Intrinsic::bitreverse:
+    return visitBitreverseIntrinsicInst(I);
+  default:
+    return false;
+  }
+}
+
+bool AMDGPUCodeGenPrepare::visitBitreverseIntrinsicInst(IntrinsicInst &I) {
+  bool Changed = false;
+
+  if (ST->has16BitInsts() && needsPromotionToI32(I.getType()) &&
+      DA->isUniform(&I))
+    Changed |= promoteUniformBitreverseToI32(I);
+
+  return Changed;
+}
+
 bool AMDGPUCodeGenPrepare::doInitialization(Module &M) {
   Mod = &M;
   return false;