| 1 | //=== ReplaceWithVeclib.cpp - Replace vector intrinsics with veclib calls -===// |
| 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 | // Replaces calls to LLVM Intrinsics with matching calls to functions from a |
| 10 | // vector library (e.g libmvec, SVML) using TargetLibraryInfo interface. |
| 11 | // |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
| 14 | #include "llvm/CodeGen/ReplaceWithVeclib.h" |
| 15 | #include "llvm/ADT/STLExtras.h" |
| 16 | #include "llvm/ADT/Statistic.h" |
| 17 | #include "llvm/ADT/StringRef.h" |
| 18 | #include "llvm/Analysis/DemandedBits.h" |
| 19 | #include "llvm/Analysis/GlobalsModRef.h" |
| 20 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" |
| 21 | #include "llvm/Analysis/TargetLibraryInfo.h" |
| 22 | #include "llvm/Analysis/VectorUtils.h" |
| 23 | #include "llvm/CodeGen/Passes.h" |
| 24 | #include "llvm/IR/DerivedTypes.h" |
| 25 | #include "llvm/IR/IRBuilder.h" |
| 26 | #include "llvm/IR/InstIterator.h" |
| 27 | #include "llvm/IR/IntrinsicInst.h" |
| 28 | #include "llvm/IR/VFABIDemangler.h" |
| 29 | #include "llvm/InitializePasses.h" |
| 30 | #include "llvm/Support/TypeSize.h" |
| 31 | #include "llvm/Transforms/Utils/ModuleUtils.h" |
| 32 | |
| 33 | using namespace llvm; |
| 34 | |
| 35 | #define DEBUG_TYPE "replace-with-veclib" |
| 36 | |
| 37 | STATISTIC(NumCallsReplaced, |
| 38 | "Number of calls to intrinsics that have been replaced." ); |
| 39 | |
| 40 | STATISTIC(NumTLIFuncDeclAdded, |
| 41 | "Number of vector library function declarations added." ); |
| 42 | |
| 43 | STATISTIC(NumFuncUsedAdded, |
| 44 | "Number of functions added to `llvm.compiler.used`" ); |
| 45 | |
| 46 | /// Returns a vector Function that it adds to the Module \p M. When an \p |
| 47 | /// ScalarFunc is not null, it copies its attributes to the newly created |
| 48 | /// Function. |
| 49 | Function *getTLIFunction(Module *M, FunctionType *VectorFTy, |
| 50 | const StringRef TLIName, |
| 51 | std::optional<CallingConv::ID> CC, |
| 52 | Function *ScalarFunc = nullptr) { |
| 53 | Function *TLIFunc = M->getFunction(Name: TLIName); |
| 54 | if (!TLIFunc) { |
| 55 | TLIFunc = |
| 56 | Function::Create(Ty: VectorFTy, Linkage: Function::ExternalLinkage, N: TLIName, M&: *M); |
| 57 | if (ScalarFunc) |
| 58 | TLIFunc->copyAttributesFrom(Src: ScalarFunc); |
| 59 | if (CC) |
| 60 | TLIFunc->setCallingConv(*CC); |
| 61 | |
| 62 | LLVM_DEBUG(dbgs() << DEBUG_TYPE << ": Added vector library function `" |
| 63 | << TLIName << "` of type `" << *(TLIFunc->getType()) |
| 64 | << "` to module.\n" ); |
| 65 | |
| 66 | ++NumTLIFuncDeclAdded; |
| 67 | // Add the freshly created function to llvm.compiler.used, similar to as it |
| 68 | // is done in InjectTLIMappings. |
| 69 | appendToCompilerUsed(M&: *M, Values: {TLIFunc}); |
| 70 | LLVM_DEBUG(dbgs() << DEBUG_TYPE << ": Adding `" << TLIName |
| 71 | << "` to `@llvm.compiler.used`.\n" ); |
| 72 | ++NumFuncUsedAdded; |
| 73 | } |
| 74 | return TLIFunc; |
| 75 | } |
| 76 | |
| 77 | /// Replace the intrinsic call \p II to \p TLIVecFunc, which is the |
| 78 | /// corresponding function from the vector library. |
| 79 | static void replaceWithTLIFunction(IntrinsicInst *II, VFInfo &Info, |
| 80 | Function *TLIVecFunc) { |
| 81 | IRBuilder<> IRBuilder(II); |
| 82 | SmallVector<Value *> Args(II->args()); |
| 83 | if (Info.isMasked()) { |
| 84 | auto *MaskTy = |
| 85 | VectorType::get(ElementType: Type::getInt1Ty(C&: II->getContext()), EC: Info.Shape.VF); |
| 86 | Args.push_back(Elt: Constant::getAllOnesValue(Ty: MaskTy)); |
| 87 | } |
| 88 | |
| 89 | // Preserve the operand bundles. |
| 90 | SmallVector<OperandBundleDef, 1> OpBundles; |
| 91 | II->getOperandBundlesAsDefs(Defs&: OpBundles); |
| 92 | |
| 93 | // Preserve fast math flags for FP math (getFastMathFlagsOrNone keeps this |
| 94 | // safe for non-FP intrinsics, whose flags are simply empty). |
| 95 | auto *Replacement = IRBuilder.CreateCall( |
| 96 | Callee: TLIVecFunc, Args, OpBundles, /*FMFSource=*/II->getFastMathFlagsOrNone()); |
| 97 | // Preserve fpmath for FP math |
| 98 | if (isa<FPMathOperator>(Val: Replacement)) |
| 99 | Replacement->copyMetadata(SrcInst: *II, WL: {LLVMContext::MD_fpmath}); |
| 100 | II->replaceAllUsesWith(V: Replacement); |
| 101 | Replacement->setCallingConv(TLIVecFunc->getCallingConv()); |
| 102 | } |
| 103 | |
| 104 | /// Returns true when successfully replaced \p II, which is a call to a |
| 105 | /// vectorized intrinsic, with a suitable function taking vector arguments, |
| 106 | /// based on available mappings in the \p TLI. |
| 107 | static bool replaceWithCallToVeclib(const TargetLibraryInfo &TLI, |
| 108 | IntrinsicInst *II) { |
| 109 | assert(II != nullptr && "Intrinsic cannot be null" ); |
| 110 | Intrinsic::ID IID = II->getIntrinsicID(); |
| 111 | Type *RetTy = II->getType(); |
| 112 | Type *ScalarRetTy = RetTy->getScalarType(); |
| 113 | // At the moment VFABI assumes the return type is always widened unless it is |
| 114 | // a void type. |
| 115 | auto *VTy = dyn_cast<VectorType>(Val: RetTy); |
| 116 | ElementCount EC(VTy ? VTy->getElementCount() : ElementCount::getFixed(MinVal: 0)); |
| 117 | |
| 118 | // OloadTys collects types used in scalar intrinsic overload name. |
| 119 | SmallVector<Type *, 3> OloadTys; |
| 120 | if (!RetTy->isVoidTy() && |
| 121 | isVectorIntrinsicWithOverloadTypeAtArg(ID: IID, OpdIdx: -1, /*TTI=*/nullptr)) |
| 122 | OloadTys.push_back(Elt: ScalarRetTy); |
| 123 | |
| 124 | // Compute the argument types of the corresponding scalar call and check that |
| 125 | // all vector operands match the previously found EC. |
| 126 | SmallVector<Type *, 8> ScalarArgTypes; |
| 127 | for (auto Arg : enumerate(First: II->args())) { |
| 128 | auto *ArgTy = Arg.value()->getType(); |
| 129 | bool IsOloadTy = isVectorIntrinsicWithOverloadTypeAtArg(ID: IID, OpdIdx: Arg.index(), |
| 130 | /*TTI=*/nullptr); |
| 131 | if (isVectorIntrinsicWithScalarOpAtArg(ID: IID, ScalarOpdIdx: Arg.index(), /*TTI=*/nullptr)) { |
| 132 | ScalarArgTypes.push_back(Elt: ArgTy); |
| 133 | if (IsOloadTy) |
| 134 | OloadTys.push_back(Elt: ArgTy); |
| 135 | } else if (auto *VectorArgTy = dyn_cast<VectorType>(Val: ArgTy)) { |
| 136 | auto *ScalarArgTy = VectorArgTy->getElementType(); |
| 137 | ScalarArgTypes.push_back(Elt: ScalarArgTy); |
| 138 | if (IsOloadTy) |
| 139 | OloadTys.push_back(Elt: ScalarArgTy); |
| 140 | // When return type is void, set EC to the first vector argument, and |
| 141 | // disallow vector arguments with different ECs. |
| 142 | if (EC.isZero()) |
| 143 | EC = VectorArgTy->getElementCount(); |
| 144 | else if (EC != VectorArgTy->getElementCount()) |
| 145 | return false; |
| 146 | } else |
| 147 | // Exit when it is supposed to be a vector argument but it isn't. |
| 148 | return false; |
| 149 | } |
| 150 | |
| 151 | // Try to reconstruct the name for the scalar version of the instruction, |
| 152 | // using scalar argument types. |
| 153 | std::string ScalarName = |
| 154 | Intrinsic::isOverloaded(id: IID) |
| 155 | ? Intrinsic::getName(Id: IID, OverloadTys: OloadTys, M: II->getModule()) |
| 156 | : Intrinsic::getName(id: IID).str(); |
| 157 | |
| 158 | // Try to find the mapping for the scalar version of this intrinsic and the |
| 159 | // exact vector width of the call operands in the TargetLibraryInfo. First, |
| 160 | // check with a non-masked variant, and if that fails try with a masked one. |
| 161 | const VecDesc *VD = |
| 162 | TLI.getVectorMappingInfo(F: ScalarName, VF: EC, /*Masked*/ false); |
| 163 | if (!VD && !(VD = TLI.getVectorMappingInfo(F: ScalarName, VF: EC, /*Masked*/ true))) |
| 164 | return false; |
| 165 | |
| 166 | LLVM_DEBUG(dbgs() << DEBUG_TYPE << ": Found TLI mapping from: `" << ScalarName |
| 167 | << "` and vector width " << EC << " to: `" |
| 168 | << VD->getVectorFnName() << "`.\n" ); |
| 169 | |
| 170 | // Replace the call to the intrinsic with a call to the vector library |
| 171 | // function. |
| 172 | FunctionType *ScalarFTy = |
| 173 | FunctionType::get(Result: ScalarRetTy, Params: ScalarArgTypes, /*isVarArg*/ false); |
| 174 | const std::string MangledName = VD->getVectorFunctionABIVariantString(); |
| 175 | auto OptInfo = VFABI::tryDemangleForVFABI(MangledName, FTy: ScalarFTy); |
| 176 | if (!OptInfo) |
| 177 | return false; |
| 178 | |
| 179 | // There is no guarantee that the vectorized instructions followed the VFABI |
| 180 | // specification when being created, this is why we need to add extra check to |
| 181 | // make sure that the operands of the vector function obtained via VFABI match |
| 182 | // the operands of the original vector instruction. |
| 183 | for (auto &VFParam : OptInfo->Shape.Parameters) { |
| 184 | if (VFParam.ParamKind == VFParamKind::GlobalPredicate) |
| 185 | continue; |
| 186 | |
| 187 | // tryDemangleForVFABI must return valid ParamPos, otherwise it could be |
| 188 | // a bug in the VFABI parser. |
| 189 | assert(VFParam.ParamPos < II->arg_size() && "ParamPos has invalid range" ); |
| 190 | Type *OrigTy = II->getArgOperand(i: VFParam.ParamPos)->getType(); |
| 191 | if (OrigTy->isVectorTy() != (VFParam.ParamKind == VFParamKind::Vector)) { |
| 192 | LLVM_DEBUG(dbgs() << DEBUG_TYPE << ": Will not replace: " << ScalarName |
| 193 | << ". Wrong type at index " << VFParam.ParamPos << ": " |
| 194 | << *OrigTy << "\n" ); |
| 195 | return false; |
| 196 | } |
| 197 | } |
| 198 | |
| 199 | FunctionType *VectorFTy = VFABI::createFunctionType(Info: *OptInfo, ScalarFTy); |
| 200 | if (!VectorFTy) |
| 201 | return false; |
| 202 | |
| 203 | Function *TLIFunc = |
| 204 | getTLIFunction(M: II->getModule(), VectorFTy, TLIName: VD->getVectorFnName(), |
| 205 | CC: VD->getCallingConv(), ScalarFunc: II->getCalledFunction()); |
| 206 | replaceWithTLIFunction(II, Info&: *OptInfo, TLIVecFunc: TLIFunc); |
| 207 | LLVM_DEBUG(dbgs() << DEBUG_TYPE << ": Replaced call to `" << ScalarName |
| 208 | << "` with call to `" << TLIFunc->getName() << "`.\n" ); |
| 209 | ++NumCallsReplaced; |
| 210 | return true; |
| 211 | } |
| 212 | |
| 213 | /// Returns true when \p TLI has a vector mapping for the scalar function name |
| 214 | /// \p Name at \p EC (matching either masked or unmasked variants). |
| 215 | static bool hasVectorMapping(const TargetLibraryInfo &TLI, StringRef Name, |
| 216 | ElementCount EC) { |
| 217 | return TLI.getVectorMappingInfo(F: Name, VF: EC, /*Masked=*/false) || |
| 218 | TLI.getVectorMappingInfo(F: Name, VF: EC, /*Masked=*/true); |
| 219 | } |
| 220 | |
| 221 | /// Returns true when \p TLI has a vector mapping for \p IID at the given |
| 222 | /// element type and \p EC. |
| 223 | static bool hasIntrinsicVectorMapping(const TargetLibraryInfo &TLI, |
| 224 | Intrinsic::ID IID, Type *ScalarTy, |
| 225 | ElementCount EC, Module *M) { |
| 226 | std::string Name = Intrinsic::getName(Id: IID, OverloadTys: {ScalarTy}, M); |
| 227 | return hasVectorMapping(TLI, Name, EC); |
| 228 | } |
| 229 | |
| 230 | /// If \p II is a vector llvm.sincos with no direct vector library mapping but |
| 231 | /// the target does have vector mappings for both llvm.sin and llvm.cos at the |
| 232 | /// same element count, replace it with separate llvm.sin and llvm.cos calls |
| 233 | /// and run the standard veclib replacement on each. |
| 234 | static bool trySplitVectorSinCos(const TargetLibraryInfo &TLI, |
| 235 | IntrinsicInst *II, |
| 236 | SmallVectorImpl<Instruction *> &Replaced) { |
| 237 | if (II->getIntrinsicID() != Intrinsic::sincos) |
| 238 | return false; |
| 239 | Value *Arg = II->getArgOperand(i: 0); |
| 240 | auto *VTy = dyn_cast<VectorType>(Val: Arg->getType()); |
| 241 | if (!VTy) |
| 242 | return false; |
| 243 | |
| 244 | ElementCount EC = VTy->getElementCount(); |
| 245 | Type *ScalarTy = VTy->getElementType(); |
| 246 | Module *M = II->getModule(); |
| 247 | |
| 248 | // If a vector sincos mapping exists for the intrinsic name (e.g. |
| 249 | // "llvm.sincos.f32") or for the scalar libcall name ("sincos"/"sincosf"), |
| 250 | // leave the call alone -- SelectionDAG legalization will handle it via |
| 251 | // expandMultipleResultFPLibCall when the runtime libcall impl is enabled. |
| 252 | if (hasIntrinsicVectorMapping(TLI, IID: Intrinsic::sincos, ScalarTy, EC, M)) |
| 253 | return false; |
| 254 | LibFunc LF = NotLibFunc; |
| 255 | if (ScalarTy->isFloatTy()) |
| 256 | LF = LibFunc_sincosf; |
| 257 | else if (ScalarTy->isDoubleTy()) |
| 258 | LF = LibFunc_sincos; |
| 259 | if (LF != NotLibFunc && hasVectorMapping(TLI, Name: TLI.getName(F: LF), EC)) |
| 260 | return false; |
| 261 | |
| 262 | // Splitting is only worthwhile when both sin and cos have vector mappings. |
| 263 | if (!hasIntrinsicVectorMapping(TLI, IID: Intrinsic::sin, ScalarTy, EC, M) || |
| 264 | !hasIntrinsicVectorMapping(TLI, IID: Intrinsic::cos, ScalarTy, EC, M)) |
| 265 | return false; |
| 266 | |
| 267 | // All users must be extractvalue. |
| 268 | for (User *U : II->users()) { |
| 269 | if (!isa<ExtractValueInst>(Val: U)) |
| 270 | return false; |
| 271 | } |
| 272 | |
| 273 | IRBuilder<> B(II); |
| 274 | Function *SinFn = |
| 275 | Intrinsic::getOrInsertDeclaration(M, id: Intrinsic::sin, OverloadTys: Arg->getType()); |
| 276 | Function *CosFn = |
| 277 | Intrinsic::getOrInsertDeclaration(M, id: Intrinsic::cos, OverloadTys: Arg->getType()); |
| 278 | CallInst *SinCall = B.CreateCall(Callee: SinFn, Args: {Arg}, /*FMFSource=*/II, Name: "sin" ); |
| 279 | CallInst *CosCall = B.CreateCall(Callee: CosFn, Args: {Arg}, /*FMFSource=*/II, Name: "cos" ); |
| 280 | SinCall->copyMetadata(SrcInst: *II, WL: {LLVMContext::MD_fpmath}); |
| 281 | CosCall->copyMetadata(SrcInst: *II, WL: {LLVMContext::MD_fpmath}); |
| 282 | |
| 283 | // Forward extractvalue uses to the new calls. |
| 284 | for (User *U : make_early_inc_range(Range: II->users())) { |
| 285 | auto *EV = cast<ExtractValueInst>(Val: U); |
| 286 | EV->replaceAllUsesWith(V: EV->getIndices()[0] == 0 ? SinCall : CosCall); |
| 287 | EV->eraseFromParent(); |
| 288 | } |
| 289 | |
| 290 | // Replace each new call with the vector library function. |
| 291 | if (replaceWithCallToVeclib(TLI, II: cast<IntrinsicInst>(Val: SinCall))) |
| 292 | Replaced.push_back(Elt: SinCall); |
| 293 | if (replaceWithCallToVeclib(TLI, II: cast<IntrinsicInst>(Val: CosCall))) |
| 294 | Replaced.push_back(Elt: CosCall); |
| 295 | |
| 296 | return true; |
| 297 | } |
| 298 | |
| 299 | static bool runImpl(const TargetLibraryInfo &TLI, Function &F) { |
| 300 | SmallVector<Instruction *> ReplacedCalls; |
| 301 | for (auto &I : instructions(F)) { |
| 302 | auto *II = dyn_cast<IntrinsicInst>(Val: &I); |
| 303 | if (!II) |
| 304 | continue; |
| 305 | |
| 306 | // Vector llvm.sincos returns a struct so it does not fit the generic |
| 307 | // path below; try to split it into separate sin and cos calls when the |
| 308 | // target has vector mappings for them. |
| 309 | if (trySplitVectorSinCos(TLI, II, Replaced&: ReplacedCalls)) { |
| 310 | ReplacedCalls.push_back(Elt: &I); |
| 311 | continue; |
| 312 | } |
| 313 | |
| 314 | // Process only intrinsic calls that return void or a vector. |
| 315 | if (!II->getType()->isVectorTy() && !II->getType()->isVoidTy()) |
| 316 | continue; |
| 317 | |
| 318 | if (replaceWithCallToVeclib(TLI, II)) |
| 319 | ReplacedCalls.push_back(Elt: &I); |
| 320 | } |
| 321 | // Erase any intrinsic calls that were replaced with vector library calls. |
| 322 | for (auto *I : ReplacedCalls) |
| 323 | I->eraseFromParent(); |
| 324 | return !ReplacedCalls.empty(); |
| 325 | } |
| 326 | |
| 327 | //////////////////////////////////////////////////////////////////////////////// |
| 328 | // New pass manager implementation. |
| 329 | //////////////////////////////////////////////////////////////////////////////// |
| 330 | PreservedAnalyses ReplaceWithVeclib::run(Function &F, |
| 331 | FunctionAnalysisManager &AM) { |
| 332 | const TargetLibraryInfo &TLI = AM.getResult<TargetLibraryAnalysis>(IR&: F); |
| 333 | auto Changed = runImpl(TLI, F); |
| 334 | if (Changed) { |
| 335 | LLVM_DEBUG(dbgs() << "Intrinsic calls replaced with vector libraries: " |
| 336 | << NumCallsReplaced << "\n" ); |
| 337 | |
| 338 | PreservedAnalyses PA; |
| 339 | PA.preserveSet<CFGAnalyses>(); |
| 340 | PA.preserve<TargetLibraryAnalysis>(); |
| 341 | PA.preserve<ScalarEvolutionAnalysis>(); |
| 342 | PA.preserve<LoopAccessAnalysis>(); |
| 343 | PA.preserve<DemandedBitsAnalysis>(); |
| 344 | PA.preserve<OptimizationRemarkEmitterAnalysis>(); |
| 345 | return PA; |
| 346 | } |
| 347 | |
| 348 | // The pass did not replace any calls, hence it preserves all analyses. |
| 349 | return PreservedAnalyses::all(); |
| 350 | } |
| 351 | |
| 352 | //////////////////////////////////////////////////////////////////////////////// |
| 353 | // Legacy PM Implementation. |
| 354 | //////////////////////////////////////////////////////////////////////////////// |
| 355 | bool ReplaceWithVeclibLegacy::runOnFunction(Function &F) { |
| 356 | const TargetLibraryInfo &TLI = |
| 357 | getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F); |
| 358 | return runImpl(TLI, F); |
| 359 | } |
| 360 | |
| 361 | void ReplaceWithVeclibLegacy::getAnalysisUsage(AnalysisUsage &AU) const { |
| 362 | AU.setPreservesCFG(); |
| 363 | AU.addRequired<TargetLibraryInfoWrapperPass>(); |
| 364 | AU.addPreserved<TargetLibraryInfoWrapperPass>(); |
| 365 | AU.addPreserved<ScalarEvolutionWrapperPass>(); |
| 366 | AU.addPreserved<AAResultsWrapperPass>(); |
| 367 | AU.addPreserved<OptimizationRemarkEmitterWrapperPass>(); |
| 368 | AU.addPreserved<GlobalsAAWrapperPass>(); |
| 369 | } |
| 370 | |
| 371 | //////////////////////////////////////////////////////////////////////////////// |
| 372 | // Legacy Pass manager initialization |
| 373 | //////////////////////////////////////////////////////////////////////////////// |
| 374 | char ReplaceWithVeclibLegacy::ID = 0; |
| 375 | |
| 376 | INITIALIZE_PASS_BEGIN(ReplaceWithVeclibLegacy, DEBUG_TYPE, |
| 377 | "Replace intrinsics with calls to vector library" , false, |
| 378 | false) |
| 379 | INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) |
| 380 | INITIALIZE_PASS_END(ReplaceWithVeclibLegacy, DEBUG_TYPE, |
| 381 | "Replace intrinsics with calls to vector library" , false, |
| 382 | false) |
| 383 | |
| 384 | FunctionPass *llvm::createReplaceWithVeclibLegacyPass() { |
| 385 | return new ReplaceWithVeclibLegacy(); |
| 386 | } |
| 387 | |