1 //===-- AMDGPUTargetTransformInfo.cpp - AMDGPU specific TTI pass ---------===//
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 file implements a TargetTransformInfo analysis pass specific to the
12 // AMDGPU target machine. It uses the target's detailed information to provide
13 // more precise answers to certain TTI queries, while letting the target
14 // independent and default TTI implementations handle the rest.
16 //===----------------------------------------------------------------------===//
19 #include "AMDGPUTargetMachine.h"
20 #include "llvm/Analysis/LoopInfo.h"
21 #include "llvm/Analysis/TargetTransformInfo.h"
22 #include "llvm/Analysis/ValueTracking.h"
23 #include "llvm/Support/Debug.h"
24 #include "llvm/Target/CostTable.h"
25 #include "llvm/Target/TargetLowering.h"
28 #define DEBUG_TYPE "AMDGPUtti"
30 // Declare the pass initialization routine locally as target-specific passes
31 // don't have a target-wide initialization entry point, and so we rely on the
32 // pass constructor initialization.
34 void initializeAMDGPUTTIPass(PassRegistry &);
39 class AMDGPUTTI final : public ImmutablePass, public TargetTransformInfo {
40 const AMDGPUTargetMachine *TM;
41 const AMDGPUSubtarget *ST;
42 const AMDGPUTargetLowering *TLI;
44 /// Estimate the overhead of scalarizing an instruction. Insert and Extract
45 /// are set if the result needs to be inserted and/or extracted from vectors.
46 unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract) const;
49 AMDGPUTTI() : ImmutablePass(ID), TM(nullptr), ST(nullptr), TLI(nullptr) {
50 llvm_unreachable("This pass cannot be directly constructed");
53 AMDGPUTTI(const AMDGPUTargetMachine *TM)
54 : ImmutablePass(ID), TM(TM), ST(TM->getSubtargetImpl()),
55 TLI(TM->getSubtargetImpl()->getTargetLowering()) {
56 initializeAMDGPUTTIPass(*PassRegistry::getPassRegistry());
59 void initializePass() override { pushTTIStack(this); }
61 void getAnalysisUsage(AnalysisUsage &AU) const override {
62 TargetTransformInfo::getAnalysisUsage(AU);
65 /// Pass identification.
68 /// Provide necessary pointer adjustments for the two base classes.
69 void *getAdjustedAnalysisPointer(const void *ID) override {
70 if (ID == &TargetTransformInfo::ID)
71 return (TargetTransformInfo *)this;
75 bool hasBranchDivergence() const override;
77 void getUnrollingPreferences(const Function *F, Loop *L,
78 UnrollingPreferences &UP) const override;
80 PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) const override;
82 unsigned getNumberOfRegisters(bool Vector) const override;
83 unsigned getRegisterBitWidth(bool Vector) const override;
84 unsigned getMaxInterleaveFactor() const override;
89 } // end anonymous namespace
91 INITIALIZE_AG_PASS(AMDGPUTTI, TargetTransformInfo, "AMDGPUtti",
92 "AMDGPU Target Transform Info", true, true, false)
93 char AMDGPUTTI::ID = 0;
96 llvm::createAMDGPUTargetTransformInfoPass(const AMDGPUTargetMachine *TM) {
97 return new AMDGPUTTI(TM);
100 bool AMDGPUTTI::hasBranchDivergence() const { return true; }
102 void AMDGPUTTI::getUnrollingPreferences(const Function *, Loop *L,
103 UnrollingPreferences &UP) const {
104 UP.Threshold = 300; // Twice the default.
108 // TODO: Do we want runtime unrolling?
110 for (const BasicBlock *BB : L->getBlocks()) {
111 for (const Instruction &I : *BB) {
112 const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(&I);
113 if (!GEP || GEP->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS)
116 const Value *Ptr = GEP->getPointerOperand();
117 const AllocaInst *Alloca = dyn_cast<AllocaInst>(GetUnderlyingObject(Ptr));
119 // We want to do whatever we can to limit the number of alloca
120 // instructions that make it through to the code generator. allocas
121 // require us to use indirect addressing, which is slow and prone to
122 // compiler bugs. If this loop does an address calculation on an
123 // alloca ptr, then we want to use a higher than normal loop unroll
124 // threshold. This will give SROA a better chance to eliminate these
127 // Don't use the maximum allowed value here as it will make some
128 // programs way too big.
135 AMDGPUTTI::PopcntSupportKind
136 AMDGPUTTI::getPopcntSupport(unsigned TyWidth) const {
137 assert(isPowerOf2_32(TyWidth) && "Ty width must be power of 2");
138 return ST->hasBCNT(TyWidth) ? PSK_FastHardware : PSK_Software;
141 unsigned AMDGPUTTI::getNumberOfRegisters(bool Vec) const {
145 // Number of VGPRs on SI.
146 if (ST->getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS)
149 return 4 * 128; // XXX - 4 channels. Should these count as vector instead?
152 unsigned AMDGPUTTI::getRegisterBitWidth(bool) const {
156 unsigned AMDGPUTTI::getMaxInterleaveFactor() const {
157 // Semi-arbitrary large amount.