1 //===-- SITypeRewriter.cpp - Remove unwanted types ------------------------===//
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 removes performs the following type substitution on all
12 /// non-compute shaders:
15 /// - v16i8 is used for constant memory resource descriptors. This type is
16 /// legal for some compute APIs, and we don't want to declare it as legal
17 /// in the backend, because we want the legalizer to expand all v16i8
20 /// - Having v1* types complicates the legalizer and we can easily replace
21 /// - them with the element type.
22 //===----------------------------------------------------------------------===//
25 #include "Utils/AMDGPUBaseInfo.h"
26 #include "llvm/IR/IRBuilder.h"
27 #include "llvm/IR/InstVisitor.h"
33 class SITypeRewriter : public FunctionPass,
34 public InstVisitor<SITypeRewriter> {
42 SITypeRewriter() : FunctionPass(ID) { }
43 bool doInitialization(Module &M) override;
44 bool runOnFunction(Function &F) override;
45 const char *getPassName() const override {
46 return "SI Type Rewriter";
48 void visitLoadInst(LoadInst &I);
49 void visitCallInst(CallInst &I);
50 void visitBitCast(BitCastInst &I);
53 } // End anonymous namespace
55 char SITypeRewriter::ID = 0;
57 bool SITypeRewriter::doInitialization(Module &M) {
59 v16i8 = VectorType::get(Type::getInt8Ty(M.getContext()), 16);
60 v4i32 = VectorType::get(Type::getInt32Ty(M.getContext()), 4);
64 bool SITypeRewriter::runOnFunction(Function &F) {
65 if (AMDGPU::getShaderType(F) == ShaderType::COMPUTE)
74 void SITypeRewriter::visitLoadInst(LoadInst &I) {
75 Value *Ptr = I.getPointerOperand();
76 Type *PtrTy = Ptr->getType();
77 Type *ElemTy = PtrTy->getPointerElementType();
78 IRBuilder<> Builder(&I);
79 if (ElemTy == v16i8) {
80 Value *BitCast = Builder.CreateBitCast(Ptr,
81 PointerType::get(v4i32,PtrTy->getPointerAddressSpace()));
82 LoadInst *Load = Builder.CreateLoad(BitCast);
83 SmallVector<std::pair<unsigned, MDNode *>, 8> MD;
84 I.getAllMetadataOtherThanDebugLoc(MD);
85 for (unsigned i = 0, e = MD.size(); i != e; ++i) {
86 Load->setMetadata(MD[i].first, MD[i].second);
88 Value *BitCastLoad = Builder.CreateBitCast(Load, I.getType());
89 I.replaceAllUsesWith(BitCastLoad);
94 void SITypeRewriter::visitCallInst(CallInst &I) {
95 IRBuilder<> Builder(&I);
97 SmallVector <Value*, 8> Args;
98 SmallVector <Type*, 8> Types;
99 bool NeedToReplace = false;
100 Function *F = I.getCalledFunction();
101 std::string Name = F->getName();
102 for (unsigned i = 0, e = I.getNumArgOperands(); i != e; ++i) {
103 Value *Arg = I.getArgOperand(i);
104 if (Arg->getType() == v16i8) {
105 Args.push_back(Builder.CreateBitCast(Arg, v4i32));
106 Types.push_back(v4i32);
107 NeedToReplace = true;
108 Name = Name + ".v4i32";
109 } else if (Arg->getType()->isVectorTy() &&
110 Arg->getType()->getVectorNumElements() == 1 &&
111 Arg->getType()->getVectorElementType() ==
112 Type::getInt32Ty(I.getContext())){
113 Type *ElementTy = Arg->getType()->getVectorElementType();
114 std::string TypeName = "i32";
115 InsertElementInst *Def = cast<InsertElementInst>(Arg);
116 Args.push_back(Def->getOperand(1));
117 Types.push_back(ElementTy);
118 std::string VecTypeName = "v1" + TypeName;
119 Name = Name.replace(Name.find(VecTypeName), VecTypeName.length(), TypeName);
120 NeedToReplace = true;
123 Types.push_back(Arg->getType());
127 if (!NeedToReplace) {
130 Function *NewF = Mod->getFunction(Name);
132 NewF = Function::Create(FunctionType::get(F->getReturnType(), Types, false), GlobalValue::ExternalLinkage, Name, Mod);
133 NewF->setAttributes(F->getAttributes());
135 I.replaceAllUsesWith(Builder.CreateCall(NewF, Args));
139 void SITypeRewriter::visitBitCast(BitCastInst &I) {
140 IRBuilder<> Builder(&I);
141 if (I.getDestTy() != v4i32) {
145 if (BitCastInst *Op = dyn_cast<BitCastInst>(I.getOperand(0))) {
146 if (Op->getSrcTy() == v4i32) {
147 I.replaceAllUsesWith(Op->getOperand(0));
153 FunctionPass *llvm::createSITypeRewriter() {
154 return new SITypeRewriter();