1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248
|
/*========================== begin_copyright_notice ============================
Copyright (C) 2024 Intel Corporation
SPDX-License-Identifier: MIT
============================= end_copyright_notice ===========================*/
#include "MergeAllocas.h"
#include "Compiler/IGCPassSupport.h"
#include "common/igc_regkeys.hpp"
#include "Probe/Assertion.h"
#include "debug/DebugMacros.hpp"
#include "common/LLVMWarningsPush.hpp"
#include <llvm/ADT/SetOperations.h>
#include <llvm/ADT/SetVector.h>
#include <llvm/ADT/SmallSet.h>
#include <llvm/ADT/SmallVector.h>
#include <llvm/Analysis/LoopInfo.h>
#include <llvm/IR/Constants.h>
#include <llvm/IR/DataLayout.h>
#include <llvm/IR/DerivedTypes.h>
#include <llvm/IR/Dominators.h>
#include <llvm/IR/Function.h>
#include <llvm/IR/IRBuilder.h>
#include <llvm/IR/InstIterator.h>
#include <llvm/IR/Instructions.h>
#include "common/LLVMWarningsPop.hpp"
using namespace llvm;
using namespace IGC;
// Get size of bytes allocated for type including padding.
static size_t GetByteSize(Type *T, const DataLayout *DL) {
if (T->isSized())
return static_cast<size_t>(DL->getTypeAllocSize(T));
return 0;
}
static MergeAllocas::AllocaInfo GetAllocaInfo(AllocaInst *allocaI, AllocationLivenessAnalyzer::LivenessData *LD,
const DataLayout *DL) {
size_t allocationSize = GetByteSize(allocaI->getAllocatedType(), DL);
return {{},
allocaI,
LD,
allocaI->getAddressSpace(),
allocationSize,
allocationSize,
static_cast<size_t>(DL->getPrefTypeAlign(allocaI->getAllocatedType()).value()),
0,
allocaI->getMetadata("uniform") != nullptr};
}
static size_t GetStartingOffset(size_t startOffset, size_t alignment) {
size_t remainder = startOffset % alignment;
if (remainder == 0) {
return startOffset;
}
return startOffset + (alignment - remainder);
}
static bool AddNonOverlappingAlloca(MergeAllocas::AllocaInfo *MergableAlloca, MergeAllocas::AllocaInfo *NewAlloca) {
if (MergableAlloca->isUniform != NewAlloca->isUniform) {
return false;
}
if (MergableAlloca->addressSpace != NewAlloca->addressSpace) {
return false;
}
if (MergableAlloca->allocationSize < NewAlloca->allocationSize) {
return false;
}
if (MergableAlloca->livenessData->OverlapsWith(*NewAlloca->livenessData)) {
return false;
}
if (IGC_IS_FLAG_ENABLED(DisableMergingOfAllocasWithDifferentType)) {
auto *AllocatedType = MergableAlloca->allocaI->getAllocatedType();
auto *AllocatedTypeNew = NewAlloca->allocaI->getAllocatedType();
bool IsArray = AllocatedType->isArrayTy() ? true : false;
bool IsArrayNew = AllocatedTypeNew->isArrayTy() ? true : false;
bool AreBothArrays = IsArray && IsArrayNew;
if (AreBothArrays && AllocatedType->getArrayElementType() != AllocatedTypeNew->getArrayElementType()) {
return false;
}
if (!AreBothArrays && AllocatedType != AllocatedTypeNew) {
return false;
}
}
// Check if we can merge alloca to one of existing non-overlapping allocas.
for (auto *NonOverlappingAlloca : MergableAlloca->nonOverlapingAllocas) {
bool added = AddNonOverlappingAlloca(NonOverlappingAlloca, NewAlloca);
if (added) {
return true;
}
}
// Check if we have still space in existing alloca to add new alloca
if (MergableAlloca->remainingSize >= NewAlloca->allocationSize) {
size_t currentOffset = MergableAlloca->allocationSize - MergableAlloca->remainingSize;
size_t newStartingOffset = GetStartingOffset(currentOffset, NewAlloca->alignment);
size_t sizeWithPadding = NewAlloca->allocationSize + (newStartingOffset - currentOffset);
// When adding alignment in consideration we can't fit new alloca.
if (sizeWithPadding > MergableAlloca->remainingSize) {
return false;
}
size_t newAllocaOffset = newStartingOffset + MergableAlloca->offset;
if (newAllocaOffset != 0 && IGC_IS_FLAG_ENABLED(DisableMergingOfMultipleAllocasWithOffset)) {
return false;
}
NewAlloca->offset = newAllocaOffset;
MergableAlloca->nonOverlapingAllocas.push_back(NewAlloca);
MergableAlloca->remainingSize -= sizeWithPadding;
return true;
}
return false;
}
static void ReplaceAllocas(const MergeAllocas::AllocaInfo &MergableAlloca, Function &F) {
Instruction *topAlloca = MergableAlloca.allocaI;
topAlloca->moveBefore(F.getEntryBlock().getFirstNonPHI());
topAlloca->setName(VALUE_NAME("MergedAlloca"));
IRBuilder<> Builder(topAlloca->getParent());
Instruction *topAllocaBitcast = nullptr;
SmallVector<MergeAllocas::AllocaInfo *> allocasToReplace;
allocasToReplace.insert(allocasToReplace.end(), MergableAlloca.nonOverlapingAllocas.begin(),
MergableAlloca.nonOverlapingAllocas.end());
while (!allocasToReplace.empty()) {
auto *subAlloca = allocasToReplace.pop_back_val();
auto *subInst = subAlloca->allocaI;
auto *ReplacementValue = topAlloca;
if (topAlloca->getType() != subInst->getType()) {
auto *InsertionPoint = (topAllocaBitcast != nullptr) ? topAllocaBitcast : topAlloca;
Builder.SetInsertPoint(InsertionPoint->getNextNode());
Value *ValueToCast = nullptr;
// If we have offset from original alloca we need to create GEP
if (subAlloca->offset != 0) {
// We can re-use same bitcast
if (topAllocaBitcast == nullptr) {
topAllocaBitcast = cast<Instruction>(
Builder.CreateBitCast(topAlloca, Builder.getInt8PtrTy(topAlloca->getType()->getPointerAddressSpace())));
}
auto *Offset = Builder.getInt32(subAlloca->offset);
auto *GEP = Builder.CreateGEP(Builder.getInt8Ty(), topAllocaBitcast, Offset);
ValueToCast = GEP;
} else {
// If no offset is needed we can directly cast to target type
ValueToCast = Builder.CreateBitCast(topAlloca, subInst->getType());
}
auto *CastedValue = llvm::cast<Instruction>(Builder.CreateBitCast(ValueToCast, subInst->getType()));
ReplacementValue = CastedValue;
}
subInst->replaceAllUsesWith(ReplacementValue);
subInst->eraseFromParent();
allocasToReplace.insert(allocasToReplace.end(), subAlloca->nonOverlapingAllocas.begin(),
subAlloca->nonOverlapingAllocas.end());
}
}
bool MergeAllocas::runOnFunction(Function &F) {
if (skipFunction(F)) {
return false;
}
auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
// in the past, the analysis pass used to be responsible for the liveness data objects
// the pass got deleted as a part of refactor (it was leaking memory anyway),
// so to avoid changing the logic, we create a backing storage for the liveness data objects
// we should revisit this at some point to clean it up, but for now
// we use std::list so it doesn't invalidate the references when inserting
std::list<LivenessData> storage;
llvm::SmallVector<std::pair<llvm::Instruction *, LivenessData *>> ABLA;
for (auto &I : make_filter_range(instructions(F), [](auto &I) { return isa<AllocaInst>(&I); })) {
storage.push_back(ProcessInstruction(&I, DT, LI));
ABLA.push_back(std::make_pair(&I, &storage.back()));
}
const auto *DataLayout = &F.getParent()->getDataLayout();
// We group non-overlapping allocas for replacements.
SmallVector<AllocaInfo *> MergableAllocas;
// First we sort analysis results based on allocation size, from larger to
// smaller.
llvm::sort(ABLA, [&](const auto &a, const auto &b) {
return GetByteSize(cast<AllocaInst>(a.first)->getAllocatedType(), DataLayout) >
GetByteSize(cast<AllocaInst>(b.first)->getAllocatedType(), DataLayout);
});
// Reserve space for all alloca infos so we can use pointers to them.
AllAllocasInfos.resize(ABLA.size());
size_t currentIndex = 0;
// We iterate over analysis results collecting non-overlapping allocas.
for (const auto &A : ABLA) {
const auto &[currI, currLD] = A;
if (skipInstruction(F, *currLD))
continue;
AllAllocasInfos[currentIndex] = GetAllocaInfo(cast<AllocaInst>(currI), currLD, DataLayout);
AllocaInfo &AllocaInfo = AllAllocasInfos[currentIndex];
currentIndex++;
// We check if the current alloca overlaps with any of the previously added.
bool added = false;
for (auto *MergableAlloca : MergableAllocas) {
if (AllocaInfo.livenessData->OverlapsWith(*MergableAlloca->livenessData)) {
continue;
}
added = AddNonOverlappingAlloca(MergableAlloca, &AllocaInfo);
if (added) {
break;
}
}
// Alloca overlaps with all of the current ones so it will be added as new
// element.
if (!added && AllocaInfo.allocationSize != 0) {
MergableAllocas.push_back(&AllocaInfo);
}
}
bool changed = false;
// Replace alloca usages
for (auto *MergableAlloca : MergableAllocas) {
if (MergableAlloca->nonOverlapingAllocas.empty()) {
continue;
}
changed = true;
ReplaceAllocas(*MergableAlloca, F);
}
return changed;
}
|