1 //===-- AMDGPUCodeGenPrepare.cpp ------------------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
11 /// This pass does misc. AMDGPU optimizations on IR before instruction
14 //===----------------------------------------------------------------------===//
17 #include "AMDGPUSubtarget.h"
18 #include "AMDGPUTargetMachine.h"
19 #include "llvm/ADT/StringRef.h"
20 #include "llvm/Analysis/DivergenceAnalysis.h"
21 #include "llvm/CodeGen/Passes.h"
22 #include "llvm/IR/Attributes.h"
23 #include "llvm/IR/BasicBlock.h"
24 #include "llvm/IR/Constants.h"
25 #include "llvm/IR/DerivedTypes.h"
26 #include "llvm/IR/Function.h"
27 #include "llvm/IR/InstrTypes.h"
28 #include "llvm/IR/Instruction.h"
29 #include "llvm/IR/Instructions.h"
30 #include "llvm/IR/InstVisitor.h"
31 #include "llvm/IR/IntrinsicInst.h"
32 #include "llvm/IR/Intrinsics.h"
33 #include "llvm/IR/IRBuilder.h"
34 #include "llvm/IR/LLVMContext.h"
35 #include "llvm/IR/Operator.h"
36 #include "llvm/IR/Type.h"
37 #include "llvm/IR/Value.h"
38 #include "llvm/Pass.h"
39 #include "llvm/Support/Casting.h"
43 #define DEBUG_TYPE "amdgpu-codegenprepare"
49 class AMDGPUCodeGenPrepare : public FunctionPass,
50 public InstVisitor<AMDGPUCodeGenPrepare, bool> {
51 const GCNTargetMachine *TM;
52 const SISubtarget *ST = nullptr;
53 DivergenceAnalysis *DA = nullptr;
54 Module *Mod = nullptr;
55 bool HasUnsafeFPMath = false;
57 /// \brief Copies exact/nsw/nuw flags (if any) from binary operation \p I to
58 /// binary operation \p V.
60 /// \returns Binary operation \p V.
61 /// \returns \p T's base element bit width.
62 unsigned getBaseElementBitWidth(const Type *T) const;
64 /// \returns Equivalent 32 bit integer type for given type \p T. For example,
65 /// if \p T is i7, then i32 is returned; if \p T is <3 x i12>, then <3 x i32>
67 Type *getI32Ty(IRBuilder<> &B, const Type *T) const;
69 /// \returns True if binary operation \p I is a signed binary operation, false
71 bool isSigned(const BinaryOperator &I) const;
73 /// \returns True if the condition of 'select' operation \p I comes from a
74 /// signed 'icmp' operation, false otherwise.
75 bool isSigned(const SelectInst &I) const;
77 /// \returns True if type \p T needs to be promoted to 32 bit integer type,
79 bool needsPromotionToI32(const Type *T) const;
81 /// \brief Promotes uniform binary operation \p I to equivalent 32 bit binary
84 /// \details \p I's base element bit width must be greater than 1 and less
85 /// than or equal 16. Promotion is done by sign or zero extending operands to
86 /// 32 bits, replacing \p I with equivalent 32 bit binary operation, and
87 /// truncating the result of 32 bit binary operation back to \p I's original
88 /// type. Division operation is not promoted.
90 /// \returns True if \p I is promoted to equivalent 32 bit binary operation,
92 bool promoteUniformOpToI32(BinaryOperator &I) const;
94 /// \brief Promotes uniform 'icmp' operation \p I to 32 bit 'icmp' operation.
96 /// \details \p I's base element bit width must be greater than 1 and less
97 /// than or equal 16. Promotion is done by sign or zero extending operands to
98 /// 32 bits, and replacing \p I with 32 bit 'icmp' operation.
101 bool promoteUniformOpToI32(ICmpInst &I) const;
103 /// \brief Promotes uniform 'select' operation \p I to 32 bit 'select'
106 /// \details \p I's base element bit width must be greater than 1 and less
107 /// than or equal 16. Promotion is done by sign or zero extending operands to
108 /// 32 bits, replacing \p I with 32 bit 'select' operation, and truncating the
109 /// result of 32 bit 'select' operation back to \p I's original type.
112 bool promoteUniformOpToI32(SelectInst &I) const;
114 /// \brief Promotes uniform 'bitreverse' intrinsic \p I to 32 bit 'bitreverse'
117 /// \details \p I's base element bit width must be greater than 1 and less
118 /// than or equal 16. Promotion is done by zero extending the operand to 32
119 /// bits, replacing \p I with 32 bit 'bitreverse' intrinsic, shifting the
120 /// result of 32 bit 'bitreverse' intrinsic to the right with zero fill (the
121 /// shift amount is 32 minus \p I's base element bit width), and truncating
122 /// the result of the shift operation back to \p I's original type.
125 bool promoteUniformBitreverseToI32(IntrinsicInst &I) const;
130 AMDGPUCodeGenPrepare(const TargetMachine *TM = nullptr) :
131 FunctionPass(ID), TM(static_cast<const GCNTargetMachine *>(TM)) {}
133 bool visitFDiv(BinaryOperator &I);
135 bool visitInstruction(Instruction &I) { return false; }
136 bool visitBinaryOperator(BinaryOperator &I);
137 bool visitICmpInst(ICmpInst &I);
138 bool visitSelectInst(SelectInst &I);
140 bool visitIntrinsicInst(IntrinsicInst &I);
141 bool visitBitreverseIntrinsicInst(IntrinsicInst &I);
143 bool doInitialization(Module &M) override;
144 bool runOnFunction(Function &F) override;
146 StringRef getPassName() const override { return "AMDGPU IR optimizations"; }
148 void getAnalysisUsage(AnalysisUsage &AU) const override {
149 AU.addRequired<DivergenceAnalysis>();
150 AU.setPreservesAll();
154 } // end anonymous namespace
156 unsigned AMDGPUCodeGenPrepare::getBaseElementBitWidth(const Type *T) const {
157 assert(needsPromotionToI32(T) && "T does not need promotion to i32");
159 if (T->isIntegerTy())
160 return T->getIntegerBitWidth();
161 return cast<VectorType>(T)->getElementType()->getIntegerBitWidth();
164 Type *AMDGPUCodeGenPrepare::getI32Ty(IRBuilder<> &B, const Type *T) const {
165 assert(needsPromotionToI32(T) && "T does not need promotion to i32");
167 if (T->isIntegerTy())
168 return B.getInt32Ty();
169 return VectorType::get(B.getInt32Ty(), cast<VectorType>(T)->getNumElements());
172 bool AMDGPUCodeGenPrepare::isSigned(const BinaryOperator &I) const {
173 return I.getOpcode() == Instruction::AShr ||
174 I.getOpcode() == Instruction::SDiv || I.getOpcode() == Instruction::SRem;
177 bool AMDGPUCodeGenPrepare::isSigned(const SelectInst &I) const {
178 return isa<ICmpInst>(I.getOperand(0)) ?
179 cast<ICmpInst>(I.getOperand(0))->isSigned() : false;
182 bool AMDGPUCodeGenPrepare::needsPromotionToI32(const Type *T) const {
183 const IntegerType *IntTy = dyn_cast<IntegerType>(T);
184 if (IntTy && IntTy->getBitWidth() > 1 && IntTy->getBitWidth() <= 16)
187 if (const VectorType *VT = dyn_cast<VectorType>(T)) {
188 // TODO: The set of packed operations is more limited, so may want to
189 // promote some anyway.
190 if (ST->hasVOP3PInsts())
193 return needsPromotionToI32(VT->getElementType());
199 // Return true if the op promoted to i32 should have nsw set.
200 static bool promotedOpIsNSW(const Instruction &I) {
201 switch (I.getOpcode()) {
202 case Instruction::Shl:
203 case Instruction::Add:
204 case Instruction::Sub:
206 case Instruction::Mul:
207 return I.hasNoUnsignedWrap();
213 // Return true if the op promoted to i32 should have nuw set.
214 static bool promotedOpIsNUW(const Instruction &I) {
215 switch (I.getOpcode()) {
216 case Instruction::Shl:
217 case Instruction::Add:
218 case Instruction::Mul:
220 case Instruction::Sub:
221 return I.hasNoUnsignedWrap();
227 bool AMDGPUCodeGenPrepare::promoteUniformOpToI32(BinaryOperator &I) const {
228 assert(needsPromotionToI32(I.getType()) &&
229 "I does not need promotion to i32");
231 if (I.getOpcode() == Instruction::SDiv ||
232 I.getOpcode() == Instruction::UDiv)
235 IRBuilder<> Builder(&I);
236 Builder.SetCurrentDebugLocation(I.getDebugLoc());
238 Type *I32Ty = getI32Ty(Builder, I.getType());
239 Value *ExtOp0 = nullptr;
240 Value *ExtOp1 = nullptr;
241 Value *ExtRes = nullptr;
242 Value *TruncRes = nullptr;
245 ExtOp0 = Builder.CreateSExt(I.getOperand(0), I32Ty);
246 ExtOp1 = Builder.CreateSExt(I.getOperand(1), I32Ty);
248 ExtOp0 = Builder.CreateZExt(I.getOperand(0), I32Ty);
249 ExtOp1 = Builder.CreateZExt(I.getOperand(1), I32Ty);
252 ExtRes = Builder.CreateBinOp(I.getOpcode(), ExtOp0, ExtOp1);
253 if (Instruction *Inst = dyn_cast<Instruction>(ExtRes)) {
254 if (promotedOpIsNSW(cast<Instruction>(I)))
255 Inst->setHasNoSignedWrap();
257 if (promotedOpIsNUW(cast<Instruction>(I)))
258 Inst->setHasNoUnsignedWrap();
260 if (const auto *ExactOp = dyn_cast<PossiblyExactOperator>(&I))
261 Inst->setIsExact(ExactOp->isExact());
264 TruncRes = Builder.CreateTrunc(ExtRes, I.getType());
266 I.replaceAllUsesWith(TruncRes);
272 bool AMDGPUCodeGenPrepare::promoteUniformOpToI32(ICmpInst &I) const {
273 assert(needsPromotionToI32(I.getOperand(0)->getType()) &&
274 "I does not need promotion to i32");
276 IRBuilder<> Builder(&I);
277 Builder.SetCurrentDebugLocation(I.getDebugLoc());
279 Type *I32Ty = getI32Ty(Builder, I.getOperand(0)->getType());
280 Value *ExtOp0 = nullptr;
281 Value *ExtOp1 = nullptr;
282 Value *NewICmp = nullptr;
285 ExtOp0 = Builder.CreateSExt(I.getOperand(0), I32Ty);
286 ExtOp1 = Builder.CreateSExt(I.getOperand(1), I32Ty);
288 ExtOp0 = Builder.CreateZExt(I.getOperand(0), I32Ty);
289 ExtOp1 = Builder.CreateZExt(I.getOperand(1), I32Ty);
291 NewICmp = Builder.CreateICmp(I.getPredicate(), ExtOp0, ExtOp1);
293 I.replaceAllUsesWith(NewICmp);
299 bool AMDGPUCodeGenPrepare::promoteUniformOpToI32(SelectInst &I) const {
300 assert(needsPromotionToI32(I.getType()) &&
301 "I does not need promotion to i32");
303 IRBuilder<> Builder(&I);
304 Builder.SetCurrentDebugLocation(I.getDebugLoc());
306 Type *I32Ty = getI32Ty(Builder, I.getType());
307 Value *ExtOp1 = nullptr;
308 Value *ExtOp2 = nullptr;
309 Value *ExtRes = nullptr;
310 Value *TruncRes = nullptr;
313 ExtOp1 = Builder.CreateSExt(I.getOperand(1), I32Ty);
314 ExtOp2 = Builder.CreateSExt(I.getOperand(2), I32Ty);
316 ExtOp1 = Builder.CreateZExt(I.getOperand(1), I32Ty);
317 ExtOp2 = Builder.CreateZExt(I.getOperand(2), I32Ty);
319 ExtRes = Builder.CreateSelect(I.getOperand(0), ExtOp1, ExtOp2);
320 TruncRes = Builder.CreateTrunc(ExtRes, I.getType());
322 I.replaceAllUsesWith(TruncRes);
328 bool AMDGPUCodeGenPrepare::promoteUniformBitreverseToI32(
329 IntrinsicInst &I) const {
330 assert(I.getIntrinsicID() == Intrinsic::bitreverse &&
331 "I must be bitreverse intrinsic");
332 assert(needsPromotionToI32(I.getType()) &&
333 "I does not need promotion to i32");
335 IRBuilder<> Builder(&I);
336 Builder.SetCurrentDebugLocation(I.getDebugLoc());
338 Type *I32Ty = getI32Ty(Builder, I.getType());
340 Intrinsic::getDeclaration(Mod, Intrinsic::bitreverse, { I32Ty });
341 Value *ExtOp = Builder.CreateZExt(I.getOperand(0), I32Ty);
342 Value *ExtRes = Builder.CreateCall(I32, { ExtOp });
344 Builder.CreateLShr(ExtRes, 32 - getBaseElementBitWidth(I.getType()));
346 Builder.CreateTrunc(LShrOp, I.getType());
348 I.replaceAllUsesWith(TruncRes);
354 static bool shouldKeepFDivF32(Value *Num, bool UnsafeDiv) {
355 const ConstantFP *CNum = dyn_cast<ConstantFP>(Num);
359 // Reciprocal f32 is handled separately without denormals.
360 return UnsafeDiv || CNum->isExactlyValue(+1.0);
363 // Insert an intrinsic for fast fdiv for safe math situations where we can
364 // reduce precision. Leave fdiv for situations where the generic node is
365 // expected to be optimized.
366 bool AMDGPUCodeGenPrepare::visitFDiv(BinaryOperator &FDiv) {
367 Type *Ty = FDiv.getType();
369 if (!Ty->getScalarType()->isFloatTy())
372 MDNode *FPMath = FDiv.getMetadata(LLVMContext::MD_fpmath);
376 const FPMathOperator *FPOp = cast<const FPMathOperator>(&FDiv);
377 float ULP = FPOp->getFPAccuracy();
381 FastMathFlags FMF = FPOp->getFastMathFlags();
382 bool UnsafeDiv = HasUnsafeFPMath || FMF.unsafeAlgebra() ||
383 FMF.allowReciprocal();
384 if (ST->hasFP32Denormals() && !UnsafeDiv)
387 IRBuilder<> Builder(FDiv.getParent(), std::next(FDiv.getIterator()), FPMath);
388 Builder.setFastMathFlags(FMF);
389 Builder.SetCurrentDebugLocation(FDiv.getDebugLoc());
391 Function *Decl = Intrinsic::getDeclaration(Mod, Intrinsic::amdgcn_fdiv_fast);
393 Value *Num = FDiv.getOperand(0);
394 Value *Den = FDiv.getOperand(1);
396 Value *NewFDiv = nullptr;
398 if (VectorType *VT = dyn_cast<VectorType>(Ty)) {
399 NewFDiv = UndefValue::get(VT);
401 // FIXME: Doesn't do the right thing for cases where the vector is partially
402 // constant. This works when the scalarizer pass is run first.
403 for (unsigned I = 0, E = VT->getNumElements(); I != E; ++I) {
404 Value *NumEltI = Builder.CreateExtractElement(Num, I);
405 Value *DenEltI = Builder.CreateExtractElement(Den, I);
408 if (shouldKeepFDivF32(NumEltI, UnsafeDiv)) {
409 NewElt = Builder.CreateFDiv(NumEltI, DenEltI);
411 NewElt = Builder.CreateCall(Decl, { NumEltI, DenEltI });
414 NewFDiv = Builder.CreateInsertElement(NewFDiv, NewElt, I);
417 if (!shouldKeepFDivF32(Num, UnsafeDiv))
418 NewFDiv = Builder.CreateCall(Decl, { Num, Den });
422 FDiv.replaceAllUsesWith(NewFDiv);
423 NewFDiv->takeName(&FDiv);
424 FDiv.eraseFromParent();
430 static bool hasUnsafeFPMath(const Function &F) {
431 Attribute Attr = F.getFnAttribute("unsafe-fp-math");
432 return Attr.getValueAsString() == "true";
435 bool AMDGPUCodeGenPrepare::visitBinaryOperator(BinaryOperator &I) {
436 bool Changed = false;
438 if (ST->has16BitInsts() && needsPromotionToI32(I.getType()) &&
440 Changed |= promoteUniformOpToI32(I);
445 bool AMDGPUCodeGenPrepare::visitICmpInst(ICmpInst &I) {
446 bool Changed = false;
448 if (ST->has16BitInsts() && needsPromotionToI32(I.getOperand(0)->getType()) &&
450 Changed |= promoteUniformOpToI32(I);
455 bool AMDGPUCodeGenPrepare::visitSelectInst(SelectInst &I) {
456 bool Changed = false;
458 if (ST->has16BitInsts() && needsPromotionToI32(I.getType()) &&
460 Changed |= promoteUniformOpToI32(I);
465 bool AMDGPUCodeGenPrepare::visitIntrinsicInst(IntrinsicInst &I) {
466 switch (I.getIntrinsicID()) {
467 case Intrinsic::bitreverse:
468 return visitBitreverseIntrinsicInst(I);
474 bool AMDGPUCodeGenPrepare::visitBitreverseIntrinsicInst(IntrinsicInst &I) {
475 bool Changed = false;
477 if (ST->has16BitInsts() && needsPromotionToI32(I.getType()) &&
479 Changed |= promoteUniformBitreverseToI32(I);
484 bool AMDGPUCodeGenPrepare::doInitialization(Module &M) {
489 bool AMDGPUCodeGenPrepare::runOnFunction(Function &F) {
490 if (!TM || skipFunction(F))
493 ST = &TM->getSubtarget<SISubtarget>(F);
494 DA = &getAnalysis<DivergenceAnalysis>();
495 HasUnsafeFPMath = hasUnsafeFPMath(F);
497 bool MadeChange = false;
499 for (BasicBlock &BB : F) {
500 BasicBlock::iterator Next;
501 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; I = Next) {
503 MadeChange |= visit(*I);
510 INITIALIZE_TM_PASS_BEGIN(AMDGPUCodeGenPrepare, DEBUG_TYPE,
511 "AMDGPU IR optimizations", false, false)
512 INITIALIZE_PASS_DEPENDENCY(DivergenceAnalysis)
513 INITIALIZE_TM_PASS_END(AMDGPUCodeGenPrepare, DEBUG_TYPE,
514 "AMDGPU IR optimizations", false, false)
516 char AMDGPUCodeGenPrepare::ID = 0;
518 FunctionPass *llvm::createAMDGPUCodeGenPreparePass(const GCNTargetMachine *TM) {
519 return new AMDGPUCodeGenPrepare(TM);