1 //===- AMDGPUAliasAnalysis ------------------------------------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 /// \file 9 /// This is the AMGPU address space based alias analysis pass. 10 //===----------------------------------------------------------------------===// 11 12 #include "AMDGPUAliasAnalysis.h" 13 #include "AMDGPU.h" 14 #include "llvm/ADT/Triple.h" 15 #include "llvm/Analysis/AliasAnalysis.h" 16 #include "llvm/Analysis/MemoryLocation.h" 17 #include "llvm/Analysis/ValueTracking.h" 18 #include "llvm/IR/Argument.h" 19 #include "llvm/IR/Attributes.h" 20 #include "llvm/IR/CallingConv.h" 21 #include "llvm/IR/Function.h" 22 #include "llvm/IR/GlobalVariable.h" 23 #include "llvm/IR/Instructions.h" 24 #include "llvm/IR/Type.h" 25 #include "llvm/IR/Value.h" 26 #include "llvm/Pass.h" 27 #include "llvm/Support/Casting.h" 28 #include "llvm/Support/ErrorHandling.h" 29 #include <cassert> 30 31 using namespace llvm; 32 33 #define DEBUG_TYPE "amdgpu-aa" 34 35 // Register this pass... 36 char AMDGPUAAWrapperPass::ID = 0; 37 char AMDGPUExternalAAWrapper::ID = 0; 38 39 INITIALIZE_PASS(AMDGPUAAWrapperPass, "amdgpu-aa", 40 "AMDGPU Address space based Alias Analysis", false, true) 41 42 INITIALIZE_PASS(AMDGPUExternalAAWrapper, "amdgpu-aa-wrapper", 43 "AMDGPU Address space based Alias Analysis Wrapper", false, true) 44 45 ImmutablePass *llvm::createAMDGPUAAWrapperPass() { 46 return new AMDGPUAAWrapperPass(); 47 } 48 49 ImmutablePass *llvm::createAMDGPUExternalAAWrapperPass() { 50 return new AMDGPUExternalAAWrapper(); 51 } 52 53 void AMDGPUAAWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const { 54 AU.setPreservesAll(); 55 } 56 57 // These arrays are indexed by address space value enum elements 0 ... to 7 58 static const AliasResult ASAliasRules[8][8] = { 59 /* Flat Global Region Group Constant Private Constant 32-bit Buffer Fat Ptr */ 60 /* Flat */ {MayAlias, MayAlias, NoAlias, MayAlias, MayAlias, MayAlias, MayAlias, MayAlias}, 61 /* Global */ {MayAlias, MayAlias, NoAlias , NoAlias , MayAlias, NoAlias , MayAlias, MayAlias}, 62 /* Region */ {NoAlias, NoAlias , MayAlias, NoAlias , NoAlias, NoAlias , NoAlias, NoAlias}, 63 /* Group */ {MayAlias, NoAlias , NoAlias , MayAlias, NoAlias , NoAlias , NoAlias , NoAlias}, 64 /* Constant */ {MayAlias, MayAlias, NoAlias, NoAlias , NoAlias , NoAlias , MayAlias, MayAlias}, 65 /* Private */ {MayAlias, NoAlias , NoAlias , NoAlias , NoAlias , MayAlias, NoAlias , NoAlias}, 66 /* Constant 32-bit */ {MayAlias, MayAlias, NoAlias, NoAlias , MayAlias, NoAlias , NoAlias , MayAlias}, 67 /* Buffer Fat Ptr */ {MayAlias, MayAlias, NoAlias , NoAlias , MayAlias, NoAlias , MayAlias, MayAlias} 68 }; 69 70 static AliasResult getAliasResult(unsigned AS1, unsigned AS2) { 71 static_assert(AMDGPUAS::MAX_AMDGPU_ADDRESS <= 7, "Addr space out of range"); 72 73 if (AS1 > AMDGPUAS::MAX_AMDGPU_ADDRESS || AS2 > AMDGPUAS::MAX_AMDGPU_ADDRESS) 74 return MayAlias; 75 76 return ASAliasRules[AS1][AS2]; 77 } 78 79 AliasResult AMDGPUAAResult::alias(const MemoryLocation &LocA, 80 const MemoryLocation &LocB, 81 AAQueryInfo &AAQI) { 82 unsigned asA = LocA.Ptr->getType()->getPointerAddressSpace(); 83 unsigned asB = LocB.Ptr->getType()->getPointerAddressSpace(); 84 85 AliasResult Result = getAliasResult(asA, asB); 86 if (Result == NoAlias) 87 return Result; 88 89 // In general, FLAT (generic) pointers could be aliased to LOCAL or PRIVATE 90 // pointers. However, as LOCAL or PRIVATE pointers point to local objects, in 91 // certain cases, it's still viable to check whether a FLAT pointer won't 92 // alias to a LOCAL or PRIVATE pointer. 93 MemoryLocation A = LocA; 94 MemoryLocation B = LocB; 95 // Canonicalize the location order to simplify the following alias check. 96 if (asA != AMDGPUAS::FLAT_ADDRESS) { 97 std::swap(asA, asB); 98 std::swap(A, B); 99 } 100 if (asA == AMDGPUAS::FLAT_ADDRESS && 101 (asB == AMDGPUAS::LOCAL_ADDRESS || asB == AMDGPUAS::PRIVATE_ADDRESS)) { 102 const auto *ObjA = 103 getUnderlyingObject(A.Ptr->stripPointerCastsAndInvariantGroups()); 104 if (const LoadInst *LI = dyn_cast<LoadInst>(ObjA)) { 105 // If a generic pointer is loaded from the constant address space, it 106 // could only be a GLOBAL or CONSTANT one as that address space is soley 107 // prepared on the host side, where only GLOBAL or CONSTANT variables are 108 // visible. Note that this even holds for regular functions. 109 if (LI->getPointerAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS) 110 return NoAlias; 111 } else if (const Argument *Arg = dyn_cast<Argument>(ObjA)) { 112 const Function *F = Arg->getParent(); 113 switch (F->getCallingConv()) { 114 case CallingConv::AMDGPU_KERNEL: 115 // In the kernel function, kernel arguments won't alias to (local) 116 // variables in shared or private address space. 117 return NoAlias; 118 default: 119 // TODO: In the regular function, if that local variable in the 120 // location B is not captured, that argument pointer won't alias to it 121 // as well. 122 break; 123 } 124 } 125 } 126 127 // Forward the query to the next alias analysis. 128 return AAResultBase::alias(LocA, LocB, AAQI); 129 } 130 131 bool AMDGPUAAResult::pointsToConstantMemory(const MemoryLocation &Loc, 132 AAQueryInfo &AAQI, bool OrLocal) { 133 unsigned AS = Loc.Ptr->getType()->getPointerAddressSpace(); 134 if (AS == AMDGPUAS::CONSTANT_ADDRESS || 135 AS == AMDGPUAS::CONSTANT_ADDRESS_32BIT) 136 return true; 137 138 const Value *Base = getUnderlyingObject(Loc.Ptr); 139 AS = Base->getType()->getPointerAddressSpace(); 140 if (AS == AMDGPUAS::CONSTANT_ADDRESS || 141 AS == AMDGPUAS::CONSTANT_ADDRESS_32BIT) 142 return true; 143 144 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(Base)) { 145 if (GV->isConstant()) 146 return true; 147 } else if (const Argument *Arg = dyn_cast<Argument>(Base)) { 148 const Function *F = Arg->getParent(); 149 150 // Only assume constant memory for arguments on kernels. 151 switch (F->getCallingConv()) { 152 default: 153 return AAResultBase::pointsToConstantMemory(Loc, AAQI, OrLocal); 154 case CallingConv::AMDGPU_LS: 155 case CallingConv::AMDGPU_HS: 156 case CallingConv::AMDGPU_ES: 157 case CallingConv::AMDGPU_GS: 158 case CallingConv::AMDGPU_VS: 159 case CallingConv::AMDGPU_PS: 160 case CallingConv::AMDGPU_CS: 161 case CallingConv::AMDGPU_KERNEL: 162 case CallingConv::SPIR_KERNEL: 163 break; 164 } 165 166 unsigned ArgNo = Arg->getArgNo(); 167 /* On an argument, ReadOnly attribute indicates that the function does 168 not write through this pointer argument, even though it may write 169 to the memory that the pointer points to. 170 On an argument, ReadNone attribute indicates that the function does 171 not dereference that pointer argument, even though it may read or write 172 the memory that the pointer points to if accessed through other pointers. 173 */ 174 if (F->hasParamAttribute(ArgNo, Attribute::NoAlias) && 175 (F->hasParamAttribute(ArgNo, Attribute::ReadNone) || 176 F->hasParamAttribute(ArgNo, Attribute::ReadOnly))) { 177 return true; 178 } 179 } 180 return AAResultBase::pointsToConstantMemory(Loc, AAQI, OrLocal); 181 } 182