Stanislav Mekhanoshin | 8e45acf | 2017-03-17 23:56:58 +0000 | [diff] [blame] | 1 | //===- AMDGPUAliasAnalysis ---------------------------------------*- C++ -*-==// |
| 2 | // |
| 3 | // The LLVM Compiler Infrastructure |
| 4 | // |
| 5 | // This file is distributed under the University of Illinois Open Source |
| 6 | // License. See LICENSE.TXT for details. |
| 7 | // |
| 8 | //===----------------------------------------------------------------------===// |
| 9 | /// \file |
| 10 | /// This is the AMGPU address space based alias analysis pass. |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #include "AMDGPU.h" |
| 14 | #include "AMDGPUAliasAnalysis.h" |
| 15 | #include "llvm/Analysis/AliasAnalysis.h" |
| 16 | #include "llvm/Analysis/ValueTracking.h" |
| 17 | #include "llvm/Analysis/Passes.h" |
| 18 | #include "llvm/Support/raw_ostream.h" |
| 19 | #include "llvm/IR/Function.h" |
| 20 | #include "llvm/IR/Module.h" |
| 21 | #include "llvm/Pass.h" |
| 22 | |
| 23 | using namespace llvm; |
| 24 | |
| 25 | #define DEBUG_TYPE "amdgpu-aa" |
| 26 | |
| 27 | // Register this pass... |
| 28 | char AMDGPUAAWrapperPass::ID = 0; |
| 29 | INITIALIZE_PASS(AMDGPUAAWrapperPass, "amdgpu-aa", |
| 30 | "AMDGPU Address space based Alias Analysis", false, true) |
| 31 | |
| 32 | ImmutablePass *llvm::createAMDGPUAAWrapperPass() { |
| 33 | return new AMDGPUAAWrapperPass(); |
| 34 | } |
| 35 | |
| 36 | void AMDGPUAAWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const { |
| 37 | AU.setPreservesAll(); |
| 38 | } |
| 39 | |
| 40 | AliasResult AMDGPUAAResult::alias(const MemoryLocation &LocA, |
| 41 | const MemoryLocation &LocB) { |
| 42 | // This array is indexed by the AMDGPUAS::AddressSpaces |
| 43 | // enum elements PRIVATE_ADDRESS ... to FLAT_ADDRESS |
| 44 | // see "llvm/Transforms/AMDSPIRUtils.h" |
| 45 | static const AliasResult ASAliasRules[5][5] = { |
| 46 | /* Private Global Constant Group Flat */ |
| 47 | /* Private */ {MayAlias, NoAlias , NoAlias , NoAlias , MayAlias}, |
| 48 | /* Global */ {NoAlias , MayAlias, NoAlias , NoAlias , MayAlias}, |
| 49 | /* Constant */ {NoAlias , NoAlias , MayAlias, NoAlias , MayAlias}, |
| 50 | /* Group */ {NoAlias , NoAlias , NoAlias , MayAlias, MayAlias}, |
| 51 | /* Flat */ {MayAlias, MayAlias, MayAlias, MayAlias, MayAlias} |
| 52 | }; |
| 53 | unsigned asA = LocA.Ptr->getType()->getPointerAddressSpace(); |
| 54 | unsigned asB = LocB.Ptr->getType()->getPointerAddressSpace(); |
| 55 | if (asA > AMDGPUAS::AddressSpaces::FLAT_ADDRESS || |
| 56 | asB > AMDGPUAS::AddressSpaces::FLAT_ADDRESS) |
| 57 | report_fatal_error("Pointer address space out of range"); |
| 58 | |
| 59 | AliasResult Result = ASAliasRules[asA][asB]; |
| 60 | if (Result == NoAlias) return Result; |
| 61 | |
| 62 | if (isa<Argument>(LocA.Ptr) && isa<Argument>(LocB.Ptr)) { |
| 63 | Type *T1 = cast<PointerType>(LocA.Ptr->getType())->getElementType(); |
| 64 | Type *T2 = cast<PointerType>(LocB.Ptr->getType())->getElementType(); |
| 65 | |
| 66 | if ((T1->isVectorTy() && !T2->isVectorTy()) || |
| 67 | (T2->isVectorTy() && !T1->isVectorTy())) |
| 68 | return NoAlias; |
| 69 | } |
| 70 | // Forward the query to the next alias analysis. |
| 71 | return AAResultBase::alias(LocA, LocB); |
| 72 | } |
| 73 | |
| 74 | bool AMDGPUAAResult::pointsToConstantMemory(const MemoryLocation &Loc, |
| 75 | bool OrLocal) { |
| 76 | const Value *Base = GetUnderlyingObject(Loc.Ptr, DL); |
| 77 | |
| 78 | if (Base->getType()->getPointerAddressSpace() == |
| 79 | AMDGPUAS::AddressSpaces::CONSTANT_ADDRESS) { |
| 80 | return true; |
| 81 | } |
| 82 | |
| 83 | if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(Base)) { |
| 84 | if (GV->isConstant()) |
| 85 | return true; |
| 86 | } else if (const Argument *Arg = dyn_cast<Argument>(Base)) { |
| 87 | const Function *F = Arg->getParent(); |
| 88 | |
| 89 | // Only assume constant memory for arguments on kernels. |
| 90 | switch (F->getCallingConv()) { |
| 91 | default: |
| 92 | return AAResultBase::pointsToConstantMemory(Loc, OrLocal); |
| 93 | case CallingConv::AMDGPU_VS: |
| 94 | case CallingConv::AMDGPU_GS: |
| 95 | case CallingConv::AMDGPU_PS: |
| 96 | case CallingConv::AMDGPU_CS: |
| 97 | case CallingConv::AMDGPU_KERNEL: |
| 98 | case CallingConv::SPIR_KERNEL: |
| 99 | break; |
| 100 | } |
| 101 | |
| 102 | unsigned ArgNo = Arg->getArgNo(); |
| 103 | /* On an argument, ReadOnly attribute indicates that the function does |
| 104 | not write through this pointer argument, even though it may write |
| 105 | to the memory that the pointer points to. |
| 106 | On an argument, ReadNone attribute indicates that the function does |
| 107 | not dereference that pointer argument, even though it may read or write |
| 108 | the memory that the pointer points to if accessed through other pointers. |
| 109 | */ |
| 110 | if (F->getAttributes().hasAttribute(ArgNo + 1, Attribute::NoAlias) && |
| 111 | (F->getAttributes().hasAttribute(ArgNo + 1, Attribute::ReadNone) || |
| 112 | F->getAttributes().hasAttribute(ArgNo + 1, Attribute::ReadOnly))) { |
| 113 | return true; |
| 114 | } |
| 115 | } |
| 116 | return AAResultBase::pointsToConstantMemory(Loc, OrLocal); |
| 117 | } |