| 1 | //===------ PollyIRBuilder.cpp --------------------------------------------===// |
| 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 | // |
| 9 | // The Polly IRBuilder file contains Polly specific extensions for the IRBuilder |
| 10 | // that are used e.g. to emit the llvm.loop.parallel metadata. |
| 11 | // |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
| 14 | #include "polly/CodeGen/IRBuilder.h" |
| 15 | #include "polly/ScopInfo.h" |
| 16 | #include "polly/Support/ScopHelper.h" |
| 17 | #include "llvm/ADT/SmallVector.h" |
| 18 | #include "llvm/IR/Metadata.h" |
| 19 | |
| 20 | using namespace llvm; |
| 21 | using namespace polly; |
| 22 | |
| 23 | static const int MaxArraysInAliasScops = 10; |
| 24 | |
| 25 | /// Get a self referencing id metadata node. |
| 26 | /// |
| 27 | /// The MDNode looks like this (if arg0/arg1 are not null): |
| 28 | /// |
| 29 | /// '!n = distinct !{!n, arg0, arg1}' |
| 30 | /// |
| 31 | /// @return The self referencing id metadata node. |
| 32 | static MDNode *getID(LLVMContext &Ctx, Metadata *arg0 = nullptr, |
| 33 | Metadata *arg1 = nullptr) { |
| 34 | MDNode *ID; |
| 35 | SmallVector<Metadata *, 3> Args; |
| 36 | // Reserve operand 0 for loop id self reference. |
| 37 | Args.push_back(Elt: nullptr); |
| 38 | |
| 39 | if (arg0) |
| 40 | Args.push_back(Elt: arg0); |
| 41 | if (arg1) |
| 42 | Args.push_back(Elt: arg1); |
| 43 | |
| 44 | ID = MDNode::getDistinct(Context&: Ctx, MDs: Args); |
| 45 | ID->replaceOperandWith(I: 0, New: ID); |
| 46 | return ID; |
| 47 | } |
| 48 | |
| 49 | ScopAnnotator::ScopAnnotator() : SE(nullptr), AliasScopeDomain(nullptr) { |
| 50 | // Push an empty staging BandAttr. |
| 51 | LoopAttrEnv.emplace_back(); |
| 52 | } |
| 53 | |
| 54 | ScopAnnotator::~ScopAnnotator() { |
| 55 | assert(LoopAttrEnv.size() == 1 && "Loop stack imbalance" ); |
| 56 | assert(!getStagingAttrEnv() && "Forgot to clear staging attr env" ); |
| 57 | } |
| 58 | |
| 59 | void ScopAnnotator::buildAliasScopes(Scop &S) { |
| 60 | SE = S.getSE(); |
| 61 | |
| 62 | LLVMContext &Ctx = SE->getContext(); |
| 63 | AliasScopeDomain = getID(Ctx, arg0: MDString::get(Context&: Ctx, Str: "polly.alias.scope.domain" )); |
| 64 | |
| 65 | AliasScopeMap.clear(); |
| 66 | OtherAliasScopeListMap.clear(); |
| 67 | |
| 68 | // We are only interested in arrays, but no scalar references. Scalars should |
| 69 | // be handled easily by basicaa. |
| 70 | SmallVector<ScopArrayInfo *, 10> Arrays; |
| 71 | for (ScopArrayInfo *Array : S.arrays()) |
| 72 | if (Array->isArrayKind()) |
| 73 | Arrays.push_back(Elt: Array); |
| 74 | |
| 75 | // The construction of alias scopes is quadratic in the number of arrays |
| 76 | // involved. In case of too many arrays, skip the construction of alias |
| 77 | // information to avoid quadratic increases in compile time and code size. |
| 78 | if (Arrays.size() > MaxArraysInAliasScops) |
| 79 | return; |
| 80 | |
| 81 | std::string AliasScopeStr = "polly.alias.scope." ; |
| 82 | for (const ScopArrayInfo *Array : Arrays) { |
| 83 | assert(Array->getBasePtr() && "Base pointer must be present" ); |
| 84 | AliasScopeMap[Array->getBasePtr()] = |
| 85 | getID(Ctx, arg0: AliasScopeDomain, |
| 86 | arg1: MDString::get(Context&: Ctx, Str: (AliasScopeStr + Array->getName()).c_str())); |
| 87 | } |
| 88 | |
| 89 | for (const ScopArrayInfo *Array : Arrays) { |
| 90 | MDNode *AliasScopeList = MDNode::get(Context&: Ctx, MDs: {}); |
| 91 | for (const auto &AliasScopePair : AliasScopeMap) { |
| 92 | if (Array->getBasePtr() == AliasScopePair.first) |
| 93 | continue; |
| 94 | |
| 95 | Metadata *Args = {AliasScopePair.second}; |
| 96 | AliasScopeList = |
| 97 | MDNode::concatenate(A: AliasScopeList, B: MDNode::get(Context&: Ctx, MDs: Args)); |
| 98 | } |
| 99 | |
| 100 | OtherAliasScopeListMap[Array->getBasePtr()] = AliasScopeList; |
| 101 | } |
| 102 | } |
| 103 | |
| 104 | void ScopAnnotator::pushLoop(Loop *L, bool IsParallel) { |
| 105 | ActiveLoops.push_back(Elt: L); |
| 106 | |
| 107 | if (IsParallel) { |
| 108 | LLVMContext &Ctx = SE->getContext(); |
| 109 | MDNode *AccessGroup = MDNode::getDistinct(Context&: Ctx, MDs: {}); |
| 110 | ParallelLoops.push_back(Elt: AccessGroup); |
| 111 | } |
| 112 | |
| 113 | // Open an empty BandAttr context for loops nested in this one. |
| 114 | LoopAttrEnv.emplace_back(); |
| 115 | } |
| 116 | |
| 117 | void ScopAnnotator::popLoop(bool IsParallel) { |
| 118 | ActiveLoops.pop_back(); |
| 119 | |
| 120 | if (IsParallel) { |
| 121 | assert(!ParallelLoops.empty() && "Expected a parallel loop to pop" ); |
| 122 | ParallelLoops.pop_back(); |
| 123 | } |
| 124 | |
| 125 | // Exit the subloop context. |
| 126 | assert(!getStagingAttrEnv() && "Forgot to clear staging attr env" ); |
| 127 | assert(LoopAttrEnv.size() >= 2 && "Popped too many" ); |
| 128 | LoopAttrEnv.pop_back(); |
| 129 | } |
| 130 | |
| 131 | static void addVectorizeMetadata(LLVMContext &Ctx, |
| 132 | SmallVector<Metadata *, 3> *Args, |
| 133 | bool EnableLoopVectorizer) { |
| 134 | MDString *PropName = MDString::get(Context&: Ctx, Str: "llvm.loop.vectorize.enable" ); |
| 135 | ConstantInt *Value = |
| 136 | ConstantInt::get(Ty: Type::getInt1Ty(C&: Ctx), V: EnableLoopVectorizer); |
| 137 | ValueAsMetadata *PropValue = ValueAsMetadata::get(V: Value); |
| 138 | Args->push_back(Elt: MDNode::get(Context&: Ctx, MDs: {PropName, PropValue})); |
| 139 | } |
| 140 | |
| 141 | void addParallelMetadata(LLVMContext &Ctx, SmallVector<Metadata *, 3> *Args, |
| 142 | llvm::SmallVector<llvm::MDNode *, 8> ParallelLoops) { |
| 143 | MDString *PropName = MDString::get(Context&: Ctx, Str: "llvm.loop.parallel_accesses" ); |
| 144 | MDNode *AccGroup = ParallelLoops.back(); |
| 145 | Args->push_back(Elt: MDNode::get(Context&: Ctx, MDs: {PropName, AccGroup})); |
| 146 | } |
| 147 | |
| 148 | void ScopAnnotator::annotateLoopLatch( |
| 149 | BranchInst *B, bool IsParallel, |
| 150 | std::optional<bool> EnableVectorizeMetadata) const { |
| 151 | LLVMContext &Ctx = SE->getContext(); |
| 152 | SmallVector<Metadata *, 3> Args; |
| 153 | |
| 154 | // For the LoopID self-reference. |
| 155 | Args.push_back(Elt: nullptr); |
| 156 | |
| 157 | // Add the user-defined loop properties to the annotation, if any. Any |
| 158 | // additional properties are appended. |
| 159 | // FIXME: What to do if these conflict? |
| 160 | MDNode *MData = nullptr; |
| 161 | if (BandAttr *AttrEnv = getActiveAttrEnv()) { |
| 162 | MData = AttrEnv->Metadata; |
| 163 | if (MData) |
| 164 | llvm::append_range(C&: Args, R: drop_begin(RangeOrContainer: MData->operands(), N: 1)); |
| 165 | } |
| 166 | if (IsParallel) |
| 167 | addParallelMetadata(Ctx, Args: &Args, ParallelLoops); |
| 168 | if (EnableVectorizeMetadata.has_value()) |
| 169 | addVectorizeMetadata(Ctx, Args: &Args, EnableLoopVectorizer: *EnableVectorizeMetadata); |
| 170 | |
| 171 | // No metadata to annotate. |
| 172 | if (!MData && Args.size() <= 1) |
| 173 | return; |
| 174 | |
| 175 | // Reuse the MData node if possible, this will avoid having to create another |
| 176 | // one that cannot be merged because LoopIDs are 'distinct'. However, we have |
| 177 | // to create a new one if we add properties. |
| 178 | if (!MData || Args.size() > MData->getNumOperands()) { |
| 179 | MData = MDNode::getDistinct(Context&: Ctx, MDs: Args); |
| 180 | MData->replaceOperandWith(I: 0, New: MData); |
| 181 | } |
| 182 | B->setMetadata(KindID: LLVMContext::MD_loop, Node: MData); |
| 183 | } |
| 184 | |
| 185 | /// Get the pointer operand |
| 186 | /// |
| 187 | /// @param Inst The instruction to be analyzed. |
| 188 | /// @return the pointer operand in case @p Inst is a memory access |
| 189 | /// instruction and nullptr otherwise. |
| 190 | static llvm::Value *getMemAccInstPointerOperand(Instruction *Inst) { |
| 191 | auto MemInst = MemAccInst::dyn_cast(V: Inst); |
| 192 | if (!MemInst) |
| 193 | return nullptr; |
| 194 | |
| 195 | return MemInst.getPointerOperand(); |
| 196 | } |
| 197 | |
| 198 | /// Find the base pointer of an array access. |
| 199 | /// |
| 200 | /// This should be equivalent to ScalarEvolution::getPointerBase, which we |
| 201 | /// cannot use here the IR is still under construction which ScalarEvolution |
| 202 | /// assumes to not be modified. |
| 203 | static Value *findBasePtr(Value *Val) { |
| 204 | while (true) { |
| 205 | if (auto *Gep = dyn_cast<GEPOperator>(Val)) { |
| 206 | Val = Gep->getPointerOperand(); |
| 207 | continue; |
| 208 | } |
| 209 | if (auto *Cast = dyn_cast<BitCastOperator>(Val)) { |
| 210 | Val = Cast->getOperand(i_nocapture: 0); |
| 211 | continue; |
| 212 | } |
| 213 | |
| 214 | break; |
| 215 | } |
| 216 | |
| 217 | return Val; |
| 218 | } |
| 219 | |
| 220 | void ScopAnnotator::annotate(Instruction *Inst) { |
| 221 | if (!Inst->mayReadOrWriteMemory()) |
| 222 | return; |
| 223 | |
| 224 | switch (ParallelLoops.size()) { |
| 225 | case 0: |
| 226 | // Not parallel to anything: no access group needed. |
| 227 | break; |
| 228 | case 1: |
| 229 | // Single parallel loop: use directly. |
| 230 | Inst->setMetadata(KindID: LLVMContext::MD_access_group, |
| 231 | Node: cast<MDNode>(Val: ParallelLoops.front())); |
| 232 | break; |
| 233 | default: |
| 234 | // Parallel to multiple loops: refer to list of access groups. |
| 235 | Inst->setMetadata(KindID: LLVMContext::MD_access_group, |
| 236 | Node: MDNode::get(Context&: SE->getContext(), |
| 237 | MDs: ArrayRef<Metadata *>( |
| 238 | (Metadata *const *)ParallelLoops.data(), |
| 239 | ParallelLoops.size()))); |
| 240 | break; |
| 241 | } |
| 242 | |
| 243 | // TODO: Use the ScopArrayInfo once available here. |
| 244 | if (!AliasScopeDomain) |
| 245 | return; |
| 246 | |
| 247 | // Do not apply annotations on memory operations that take more than one |
| 248 | // pointer. It would be ambiguous to which pointer the annotation applies. |
| 249 | // FIXME: How can we specify annotations for all pointer arguments? |
| 250 | if (isa<CallInst>(Val: Inst) && !isa<MemSetInst>(Val: Inst)) |
| 251 | return; |
| 252 | |
| 253 | auto *Ptr = getMemAccInstPointerOperand(Inst); |
| 254 | if (!Ptr) |
| 255 | return; |
| 256 | |
| 257 | Value *BasePtr = findBasePtr(Val: Ptr); |
| 258 | if (!BasePtr) |
| 259 | return; |
| 260 | |
| 261 | auto AliasScope = AliasScopeMap.lookup(Key: BasePtr); |
| 262 | |
| 263 | if (!AliasScope) { |
| 264 | BasePtr = AlternativeAliasBases.lookup(Val: BasePtr); |
| 265 | if (!BasePtr) |
| 266 | return; |
| 267 | |
| 268 | AliasScope = AliasScopeMap.lookup(Key: BasePtr); |
| 269 | if (!AliasScope) |
| 270 | return; |
| 271 | } |
| 272 | |
| 273 | assert(OtherAliasScopeListMap.count(BasePtr) && |
| 274 | "BasePtr either expected in AliasScopeMap and OtherAlias...Map" ); |
| 275 | auto *OtherAliasScopeList = OtherAliasScopeListMap[BasePtr]; |
| 276 | |
| 277 | Inst->setMetadata(Kind: "alias.scope" , Node: MDNode::get(Context&: SE->getContext(), MDs: AliasScope)); |
| 278 | Inst->setMetadata(Kind: "noalias" , Node: OtherAliasScopeList); |
| 279 | } |
| 280 | |