1 //===- ScalarizeMaskedMemIntrin.cpp - Scalarize unsupported masked mem ----===// 2 // intrinsics 3 // 4 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 5 // See https://llvm.org/LICENSE.txt for license information. 6 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This pass replaces masked memory intrinsics - when unsupported by the target 11 // - with a chain of basic blocks, that deal with the elements one-by-one if the 12 // appropriate mask bit is set. 13 // 14 //===----------------------------------------------------------------------===// 15 16 #include "llvm/Transforms/Scalar/ScalarizeMaskedMemIntrin.h" 17 #include "llvm/ADT/Twine.h" 18 #include "llvm/Analysis/DomTreeUpdater.h" 19 #include "llvm/Analysis/TargetTransformInfo.h" 20 #include "llvm/IR/BasicBlock.h" 21 #include "llvm/IR/Constant.h" 22 #include "llvm/IR/Constants.h" 23 #include "llvm/IR/DerivedTypes.h" 24 #include "llvm/IR/Dominators.h" 25 #include "llvm/IR/Function.h" 26 #include "llvm/IR/IRBuilder.h" 27 #include "llvm/IR/Instruction.h" 28 #include "llvm/IR/Instructions.h" 29 #include "llvm/IR/IntrinsicInst.h" 30 #include "llvm/IR/Type.h" 31 #include "llvm/IR/Value.h" 32 #include "llvm/InitializePasses.h" 33 #include "llvm/Pass.h" 34 #include "llvm/Support/Casting.h" 35 #include "llvm/Transforms/Scalar.h" 36 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 37 #include <cassert> 38 39 using namespace llvm; 40 41 #define DEBUG_TYPE "scalarize-masked-mem-intrin" 42 43 namespace { 44 45 class ScalarizeMaskedMemIntrinLegacyPass : public FunctionPass { 46 public: 47 static char ID; // Pass identification, replacement for typeid 48 49 explicit ScalarizeMaskedMemIntrinLegacyPass() : FunctionPass(ID) { 50 initializeScalarizeMaskedMemIntrinLegacyPassPass( 51 *PassRegistry::getPassRegistry()); 52 } 53 54 bool runOnFunction(Function &F) override; 55 56 StringRef getPassName() const override { 57 return "Scalarize Masked Memory Intrinsics"; 58 } 59 60 void getAnalysisUsage(AnalysisUsage &AU) const override { 61 AU.addRequired<TargetTransformInfoWrapperPass>(); 62 AU.addPreserved<DominatorTreeWrapperPass>(); 63 } 64 }; 65 66 } // end anonymous namespace 67 68 static bool optimizeBlock(BasicBlock &BB, bool &ModifiedDT, 69 const TargetTransformInfo &TTI, const DataLayout &DL, 70 DomTreeUpdater *DTU); 71 static bool optimizeCallInst(CallInst *CI, bool &ModifiedDT, 72 const TargetTransformInfo &TTI, 73 const DataLayout &DL, DomTreeUpdater *DTU); 74 75 char ScalarizeMaskedMemIntrinLegacyPass::ID = 0; 76 77 INITIALIZE_PASS_BEGIN(ScalarizeMaskedMemIntrinLegacyPass, DEBUG_TYPE, 78 "Scalarize unsupported masked memory intrinsics", false, 79 false) 80 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) 81 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 82 INITIALIZE_PASS_END(ScalarizeMaskedMemIntrinLegacyPass, DEBUG_TYPE, 83 "Scalarize unsupported masked memory intrinsics", false, 84 false) 85 86 FunctionPass *llvm::createScalarizeMaskedMemIntrinLegacyPass() { 87 return new ScalarizeMaskedMemIntrinLegacyPass(); 88 } 89 90 static bool isConstantIntVector(Value *Mask) { 91 Constant *C = dyn_cast<Constant>(Mask); 92 if (!C) 93 return false; 94 95 unsigned NumElts = cast<FixedVectorType>(Mask->getType())->getNumElements(); 96 for (unsigned i = 0; i != NumElts; ++i) { 97 Constant *CElt = C->getAggregateElement(i); 98 if (!CElt || !isa<ConstantInt>(CElt)) 99 return false; 100 } 101 102 return true; 103 } 104 105 static unsigned adjustForEndian(const DataLayout &DL, unsigned VectorWidth, 106 unsigned Idx) { 107 return DL.isBigEndian() ? VectorWidth - 1 - Idx : Idx; 108 } 109 110 // Translate a masked load intrinsic like 111 // <16 x i32 > @llvm.masked.load( <16 x i32>* %addr, i32 align, 112 // <16 x i1> %mask, <16 x i32> %passthru) 113 // to a chain of basic blocks, with loading element one-by-one if 114 // the appropriate mask bit is set 115 // 116 // %1 = bitcast i8* %addr to i32* 117 // %2 = extractelement <16 x i1> %mask, i32 0 118 // br i1 %2, label %cond.load, label %else 119 // 120 // cond.load: ; preds = %0 121 // %3 = getelementptr i32* %1, i32 0 122 // %4 = load i32* %3 123 // %5 = insertelement <16 x i32> %passthru, i32 %4, i32 0 124 // br label %else 125 // 126 // else: ; preds = %0, %cond.load 127 // %res.phi.else = phi <16 x i32> [ %5, %cond.load ], [ undef, %0 ] 128 // %6 = extractelement <16 x i1> %mask, i32 1 129 // br i1 %6, label %cond.load1, label %else2 130 // 131 // cond.load1: ; preds = %else 132 // %7 = getelementptr i32* %1, i32 1 133 // %8 = load i32* %7 134 // %9 = insertelement <16 x i32> %res.phi.else, i32 %8, i32 1 135 // br label %else2 136 // 137 // else2: ; preds = %else, %cond.load1 138 // %res.phi.else3 = phi <16 x i32> [ %9, %cond.load1 ], [ %res.phi.else, %else ] 139 // %10 = extractelement <16 x i1> %mask, i32 2 140 // br i1 %10, label %cond.load4, label %else5 141 // 142 static void scalarizeMaskedLoad(const DataLayout &DL, CallInst *CI, 143 DomTreeUpdater *DTU, bool &ModifiedDT) { 144 Value *Ptr = CI->getArgOperand(0); 145 Value *Alignment = CI->getArgOperand(1); 146 Value *Mask = CI->getArgOperand(2); 147 Value *Src0 = CI->getArgOperand(3); 148 149 const Align AlignVal = cast<ConstantInt>(Alignment)->getAlignValue(); 150 VectorType *VecType = cast<FixedVectorType>(CI->getType()); 151 152 Type *EltTy = VecType->getElementType(); 153 154 IRBuilder<> Builder(CI->getContext()); 155 Instruction *InsertPt = CI; 156 BasicBlock *IfBlock = CI->getParent(); 157 158 Builder.SetInsertPoint(InsertPt); 159 Builder.SetCurrentDebugLocation(CI->getDebugLoc()); 160 161 // Short-cut if the mask is all-true. 162 if (isa<Constant>(Mask) && cast<Constant>(Mask)->isAllOnesValue()) { 163 Value *NewI = Builder.CreateAlignedLoad(VecType, Ptr, AlignVal); 164 CI->replaceAllUsesWith(NewI); 165 CI->eraseFromParent(); 166 return; 167 } 168 169 // Adjust alignment for the scalar instruction. 170 const Align AdjustedAlignVal = 171 commonAlignment(AlignVal, EltTy->getPrimitiveSizeInBits() / 8); 172 // Bitcast %addr from i8* to EltTy* 173 Type *NewPtrType = 174 EltTy->getPointerTo(Ptr->getType()->getPointerAddressSpace()); 175 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType); 176 unsigned VectorWidth = cast<FixedVectorType>(VecType)->getNumElements(); 177 178 // The result vector 179 Value *VResult = Src0; 180 181 if (isConstantIntVector(Mask)) { 182 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) { 183 if (cast<Constant>(Mask)->getAggregateElement(Idx)->isNullValue()) 184 continue; 185 Value *Gep = Builder.CreateConstInBoundsGEP1_32(EltTy, FirstEltPtr, Idx); 186 LoadInst *Load = Builder.CreateAlignedLoad(EltTy, Gep, AdjustedAlignVal); 187 VResult = Builder.CreateInsertElement(VResult, Load, Idx); 188 } 189 CI->replaceAllUsesWith(VResult); 190 CI->eraseFromParent(); 191 return; 192 } 193 194 // If the mask is not v1i1, use scalar bit test operations. This generates 195 // better results on X86 at least. 196 Value *SclrMask; 197 if (VectorWidth != 1) { 198 Type *SclrMaskTy = Builder.getIntNTy(VectorWidth); 199 SclrMask = Builder.CreateBitCast(Mask, SclrMaskTy, "scalar_mask"); 200 } 201 202 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) { 203 // Fill the "else" block, created in the previous iteration 204 // 205 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ] 206 // %mask_1 = and i16 %scalar_mask, i32 1 << Idx 207 // %cond = icmp ne i16 %mask_1, 0 208 // br i1 %mask_1, label %cond.load, label %else 209 // 210 Value *Predicate; 211 if (VectorWidth != 1) { 212 Value *Mask = Builder.getInt(APInt::getOneBitSet( 213 VectorWidth, adjustForEndian(DL, VectorWidth, Idx))); 214 Predicate = Builder.CreateICmpNE(Builder.CreateAnd(SclrMask, Mask), 215 Builder.getIntN(VectorWidth, 0)); 216 } else { 217 Predicate = Builder.CreateExtractElement(Mask, Idx); 218 } 219 220 // Create "cond" block 221 // 222 // %EltAddr = getelementptr i32* %1, i32 0 223 // %Elt = load i32* %EltAddr 224 // VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx 225 // 226 Instruction *ThenTerm = 227 SplitBlockAndInsertIfThen(Predicate, InsertPt, /*Unreachable=*/false, 228 /*BranchWeights=*/nullptr, DTU); 229 230 BasicBlock *CondBlock = ThenTerm->getParent(); 231 CondBlock->setName("cond.load"); 232 233 Builder.SetInsertPoint(CondBlock->getTerminator()); 234 Value *Gep = Builder.CreateConstInBoundsGEP1_32(EltTy, FirstEltPtr, Idx); 235 LoadInst *Load = Builder.CreateAlignedLoad(EltTy, Gep, AdjustedAlignVal); 236 Value *NewVResult = Builder.CreateInsertElement(VResult, Load, Idx); 237 238 // Create "else" block, fill it in the next iteration 239 BasicBlock *NewIfBlock = ThenTerm->getSuccessor(0); 240 NewIfBlock->setName("else"); 241 BasicBlock *PrevIfBlock = IfBlock; 242 IfBlock = NewIfBlock; 243 244 // Create the phi to join the new and previous value. 245 Builder.SetInsertPoint(NewIfBlock, NewIfBlock->begin()); 246 PHINode *Phi = Builder.CreatePHI(VecType, 2, "res.phi.else"); 247 Phi->addIncoming(NewVResult, CondBlock); 248 Phi->addIncoming(VResult, PrevIfBlock); 249 VResult = Phi; 250 } 251 252 CI->replaceAllUsesWith(VResult); 253 CI->eraseFromParent(); 254 255 ModifiedDT = true; 256 } 257 258 // Translate a masked store intrinsic, like 259 // void @llvm.masked.store(<16 x i32> %src, <16 x i32>* %addr, i32 align, 260 // <16 x i1> %mask) 261 // to a chain of basic blocks, that stores element one-by-one if 262 // the appropriate mask bit is set 263 // 264 // %1 = bitcast i8* %addr to i32* 265 // %2 = extractelement <16 x i1> %mask, i32 0 266 // br i1 %2, label %cond.store, label %else 267 // 268 // cond.store: ; preds = %0 269 // %3 = extractelement <16 x i32> %val, i32 0 270 // %4 = getelementptr i32* %1, i32 0 271 // store i32 %3, i32* %4 272 // br label %else 273 // 274 // else: ; preds = %0, %cond.store 275 // %5 = extractelement <16 x i1> %mask, i32 1 276 // br i1 %5, label %cond.store1, label %else2 277 // 278 // cond.store1: ; preds = %else 279 // %6 = extractelement <16 x i32> %val, i32 1 280 // %7 = getelementptr i32* %1, i32 1 281 // store i32 %6, i32* %7 282 // br label %else2 283 // . . . 284 static void scalarizeMaskedStore(const DataLayout &DL, CallInst *CI, 285 DomTreeUpdater *DTU, bool &ModifiedDT) { 286 Value *Src = CI->getArgOperand(0); 287 Value *Ptr = CI->getArgOperand(1); 288 Value *Alignment = CI->getArgOperand(2); 289 Value *Mask = CI->getArgOperand(3); 290 291 const Align AlignVal = cast<ConstantInt>(Alignment)->getAlignValue(); 292 auto *VecType = cast<VectorType>(Src->getType()); 293 294 Type *EltTy = VecType->getElementType(); 295 296 IRBuilder<> Builder(CI->getContext()); 297 Instruction *InsertPt = CI; 298 Builder.SetInsertPoint(InsertPt); 299 Builder.SetCurrentDebugLocation(CI->getDebugLoc()); 300 301 // Short-cut if the mask is all-true. 302 if (isa<Constant>(Mask) && cast<Constant>(Mask)->isAllOnesValue()) { 303 Builder.CreateAlignedStore(Src, Ptr, AlignVal); 304 CI->eraseFromParent(); 305 return; 306 } 307 308 // Adjust alignment for the scalar instruction. 309 const Align AdjustedAlignVal = 310 commonAlignment(AlignVal, EltTy->getPrimitiveSizeInBits() / 8); 311 // Bitcast %addr from i8* to EltTy* 312 Type *NewPtrType = 313 EltTy->getPointerTo(Ptr->getType()->getPointerAddressSpace()); 314 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType); 315 unsigned VectorWidth = cast<FixedVectorType>(VecType)->getNumElements(); 316 317 if (isConstantIntVector(Mask)) { 318 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) { 319 if (cast<Constant>(Mask)->getAggregateElement(Idx)->isNullValue()) 320 continue; 321 Value *OneElt = Builder.CreateExtractElement(Src, Idx); 322 Value *Gep = Builder.CreateConstInBoundsGEP1_32(EltTy, FirstEltPtr, Idx); 323 Builder.CreateAlignedStore(OneElt, Gep, AdjustedAlignVal); 324 } 325 CI->eraseFromParent(); 326 return; 327 } 328 329 // If the mask is not v1i1, use scalar bit test operations. This generates 330 // better results on X86 at least. 331 Value *SclrMask; 332 if (VectorWidth != 1) { 333 Type *SclrMaskTy = Builder.getIntNTy(VectorWidth); 334 SclrMask = Builder.CreateBitCast(Mask, SclrMaskTy, "scalar_mask"); 335 } 336 337 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) { 338 // Fill the "else" block, created in the previous iteration 339 // 340 // %mask_1 = and i16 %scalar_mask, i32 1 << Idx 341 // %cond = icmp ne i16 %mask_1, 0 342 // br i1 %mask_1, label %cond.store, label %else 343 // 344 Value *Predicate; 345 if (VectorWidth != 1) { 346 Value *Mask = Builder.getInt(APInt::getOneBitSet( 347 VectorWidth, adjustForEndian(DL, VectorWidth, Idx))); 348 Predicate = Builder.CreateICmpNE(Builder.CreateAnd(SclrMask, Mask), 349 Builder.getIntN(VectorWidth, 0)); 350 } else { 351 Predicate = Builder.CreateExtractElement(Mask, Idx); 352 } 353 354 // Create "cond" block 355 // 356 // %OneElt = extractelement <16 x i32> %Src, i32 Idx 357 // %EltAddr = getelementptr i32* %1, i32 0 358 // %store i32 %OneElt, i32* %EltAddr 359 // 360 Instruction *ThenTerm = 361 SplitBlockAndInsertIfThen(Predicate, InsertPt, /*Unreachable=*/false, 362 /*BranchWeights=*/nullptr, DTU); 363 364 BasicBlock *CondBlock = ThenTerm->getParent(); 365 CondBlock->setName("cond.store"); 366 367 Builder.SetInsertPoint(CondBlock->getTerminator()); 368 Value *OneElt = Builder.CreateExtractElement(Src, Idx); 369 Value *Gep = Builder.CreateConstInBoundsGEP1_32(EltTy, FirstEltPtr, Idx); 370 Builder.CreateAlignedStore(OneElt, Gep, AdjustedAlignVal); 371 372 // Create "else" block, fill it in the next iteration 373 BasicBlock *NewIfBlock = ThenTerm->getSuccessor(0); 374 NewIfBlock->setName("else"); 375 376 Builder.SetInsertPoint(NewIfBlock, NewIfBlock->begin()); 377 } 378 CI->eraseFromParent(); 379 380 ModifiedDT = true; 381 } 382 383 // Translate a masked gather intrinsic like 384 // <16 x i32 > @llvm.masked.gather.v16i32( <16 x i32*> %Ptrs, i32 4, 385 // <16 x i1> %Mask, <16 x i32> %Src) 386 // to a chain of basic blocks, with loading element one-by-one if 387 // the appropriate mask bit is set 388 // 389 // %Ptrs = getelementptr i32, i32* %base, <16 x i64> %ind 390 // %Mask0 = extractelement <16 x i1> %Mask, i32 0 391 // br i1 %Mask0, label %cond.load, label %else 392 // 393 // cond.load: 394 // %Ptr0 = extractelement <16 x i32*> %Ptrs, i32 0 395 // %Load0 = load i32, i32* %Ptr0, align 4 396 // %Res0 = insertelement <16 x i32> undef, i32 %Load0, i32 0 397 // br label %else 398 // 399 // else: 400 // %res.phi.else = phi <16 x i32>[%Res0, %cond.load], [undef, %0] 401 // %Mask1 = extractelement <16 x i1> %Mask, i32 1 402 // br i1 %Mask1, label %cond.load1, label %else2 403 // 404 // cond.load1: 405 // %Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1 406 // %Load1 = load i32, i32* %Ptr1, align 4 407 // %Res1 = insertelement <16 x i32> %res.phi.else, i32 %Load1, i32 1 408 // br label %else2 409 // . . . 410 // %Result = select <16 x i1> %Mask, <16 x i32> %res.phi.select, <16 x i32> %Src 411 // ret <16 x i32> %Result 412 static void scalarizeMaskedGather(const DataLayout &DL, CallInst *CI, 413 DomTreeUpdater *DTU, bool &ModifiedDT) { 414 Value *Ptrs = CI->getArgOperand(0); 415 Value *Alignment = CI->getArgOperand(1); 416 Value *Mask = CI->getArgOperand(2); 417 Value *Src0 = CI->getArgOperand(3); 418 419 auto *VecType = cast<FixedVectorType>(CI->getType()); 420 Type *EltTy = VecType->getElementType(); 421 422 IRBuilder<> Builder(CI->getContext()); 423 Instruction *InsertPt = CI; 424 BasicBlock *IfBlock = CI->getParent(); 425 Builder.SetInsertPoint(InsertPt); 426 MaybeAlign AlignVal = cast<ConstantInt>(Alignment)->getMaybeAlignValue(); 427 428 Builder.SetCurrentDebugLocation(CI->getDebugLoc()); 429 430 // The result vector 431 Value *VResult = Src0; 432 unsigned VectorWidth = VecType->getNumElements(); 433 434 // Shorten the way if the mask is a vector of constants. 435 if (isConstantIntVector(Mask)) { 436 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) { 437 if (cast<Constant>(Mask)->getAggregateElement(Idx)->isNullValue()) 438 continue; 439 Value *Ptr = Builder.CreateExtractElement(Ptrs, Idx, "Ptr" + Twine(Idx)); 440 LoadInst *Load = 441 Builder.CreateAlignedLoad(EltTy, Ptr, AlignVal, "Load" + Twine(Idx)); 442 VResult = 443 Builder.CreateInsertElement(VResult, Load, Idx, "Res" + Twine(Idx)); 444 } 445 CI->replaceAllUsesWith(VResult); 446 CI->eraseFromParent(); 447 return; 448 } 449 450 // If the mask is not v1i1, use scalar bit test operations. This generates 451 // better results on X86 at least. 452 Value *SclrMask; 453 if (VectorWidth != 1) { 454 Type *SclrMaskTy = Builder.getIntNTy(VectorWidth); 455 SclrMask = Builder.CreateBitCast(Mask, SclrMaskTy, "scalar_mask"); 456 } 457 458 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) { 459 // Fill the "else" block, created in the previous iteration 460 // 461 // %Mask1 = and i16 %scalar_mask, i32 1 << Idx 462 // %cond = icmp ne i16 %mask_1, 0 463 // br i1 %Mask1, label %cond.load, label %else 464 // 465 466 Value *Predicate; 467 if (VectorWidth != 1) { 468 Value *Mask = Builder.getInt(APInt::getOneBitSet( 469 VectorWidth, adjustForEndian(DL, VectorWidth, Idx))); 470 Predicate = Builder.CreateICmpNE(Builder.CreateAnd(SclrMask, Mask), 471 Builder.getIntN(VectorWidth, 0)); 472 } else { 473 Predicate = Builder.CreateExtractElement(Mask, Idx, "Mask" + Twine(Idx)); 474 } 475 476 // Create "cond" block 477 // 478 // %EltAddr = getelementptr i32* %1, i32 0 479 // %Elt = load i32* %EltAddr 480 // VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx 481 // 482 Instruction *ThenTerm = 483 SplitBlockAndInsertIfThen(Predicate, InsertPt, /*Unreachable=*/false, 484 /*BranchWeights=*/nullptr, DTU); 485 486 BasicBlock *CondBlock = ThenTerm->getParent(); 487 CondBlock->setName("cond.load"); 488 489 Builder.SetInsertPoint(CondBlock->getTerminator()); 490 Value *Ptr = Builder.CreateExtractElement(Ptrs, Idx, "Ptr" + Twine(Idx)); 491 LoadInst *Load = 492 Builder.CreateAlignedLoad(EltTy, Ptr, AlignVal, "Load" + Twine(Idx)); 493 Value *NewVResult = 494 Builder.CreateInsertElement(VResult, Load, Idx, "Res" + Twine(Idx)); 495 496 // Create "else" block, fill it in the next iteration 497 BasicBlock *NewIfBlock = ThenTerm->getSuccessor(0); 498 NewIfBlock->setName("else"); 499 BasicBlock *PrevIfBlock = IfBlock; 500 IfBlock = NewIfBlock; 501 502 // Create the phi to join the new and previous value. 503 Builder.SetInsertPoint(NewIfBlock, NewIfBlock->begin()); 504 PHINode *Phi = Builder.CreatePHI(VecType, 2, "res.phi.else"); 505 Phi->addIncoming(NewVResult, CondBlock); 506 Phi->addIncoming(VResult, PrevIfBlock); 507 VResult = Phi; 508 } 509 510 CI->replaceAllUsesWith(VResult); 511 CI->eraseFromParent(); 512 513 ModifiedDT = true; 514 } 515 516 // Translate a masked scatter intrinsic, like 517 // void @llvm.masked.scatter.v16i32(<16 x i32> %Src, <16 x i32*>* %Ptrs, i32 4, 518 // <16 x i1> %Mask) 519 // to a chain of basic blocks, that stores element one-by-one if 520 // the appropriate mask bit is set. 521 // 522 // %Ptrs = getelementptr i32, i32* %ptr, <16 x i64> %ind 523 // %Mask0 = extractelement <16 x i1> %Mask, i32 0 524 // br i1 %Mask0, label %cond.store, label %else 525 // 526 // cond.store: 527 // %Elt0 = extractelement <16 x i32> %Src, i32 0 528 // %Ptr0 = extractelement <16 x i32*> %Ptrs, i32 0 529 // store i32 %Elt0, i32* %Ptr0, align 4 530 // br label %else 531 // 532 // else: 533 // %Mask1 = extractelement <16 x i1> %Mask, i32 1 534 // br i1 %Mask1, label %cond.store1, label %else2 535 // 536 // cond.store1: 537 // %Elt1 = extractelement <16 x i32> %Src, i32 1 538 // %Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1 539 // store i32 %Elt1, i32* %Ptr1, align 4 540 // br label %else2 541 // . . . 542 static void scalarizeMaskedScatter(const DataLayout &DL, CallInst *CI, 543 DomTreeUpdater *DTU, bool &ModifiedDT) { 544 Value *Src = CI->getArgOperand(0); 545 Value *Ptrs = CI->getArgOperand(1); 546 Value *Alignment = CI->getArgOperand(2); 547 Value *Mask = CI->getArgOperand(3); 548 549 auto *SrcFVTy = cast<FixedVectorType>(Src->getType()); 550 551 assert( 552 isa<VectorType>(Ptrs->getType()) && 553 isa<PointerType>(cast<VectorType>(Ptrs->getType())->getElementType()) && 554 "Vector of pointers is expected in masked scatter intrinsic"); 555 556 IRBuilder<> Builder(CI->getContext()); 557 Instruction *InsertPt = CI; 558 Builder.SetInsertPoint(InsertPt); 559 Builder.SetCurrentDebugLocation(CI->getDebugLoc()); 560 561 MaybeAlign AlignVal = cast<ConstantInt>(Alignment)->getMaybeAlignValue(); 562 unsigned VectorWidth = SrcFVTy->getNumElements(); 563 564 // Shorten the way if the mask is a vector of constants. 565 if (isConstantIntVector(Mask)) { 566 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) { 567 if (cast<Constant>(Mask)->getAggregateElement(Idx)->isNullValue()) 568 continue; 569 Value *OneElt = 570 Builder.CreateExtractElement(Src, Idx, "Elt" + Twine(Idx)); 571 Value *Ptr = Builder.CreateExtractElement(Ptrs, Idx, "Ptr" + Twine(Idx)); 572 Builder.CreateAlignedStore(OneElt, Ptr, AlignVal); 573 } 574 CI->eraseFromParent(); 575 return; 576 } 577 578 // If the mask is not v1i1, use scalar bit test operations. This generates 579 // better results on X86 at least. 580 Value *SclrMask; 581 if (VectorWidth != 1) { 582 Type *SclrMaskTy = Builder.getIntNTy(VectorWidth); 583 SclrMask = Builder.CreateBitCast(Mask, SclrMaskTy, "scalar_mask"); 584 } 585 586 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) { 587 // Fill the "else" block, created in the previous iteration 588 // 589 // %Mask1 = and i16 %scalar_mask, i32 1 << Idx 590 // %cond = icmp ne i16 %mask_1, 0 591 // br i1 %Mask1, label %cond.store, label %else 592 // 593 Value *Predicate; 594 if (VectorWidth != 1) { 595 Value *Mask = Builder.getInt(APInt::getOneBitSet( 596 VectorWidth, adjustForEndian(DL, VectorWidth, Idx))); 597 Predicate = Builder.CreateICmpNE(Builder.CreateAnd(SclrMask, Mask), 598 Builder.getIntN(VectorWidth, 0)); 599 } else { 600 Predicate = Builder.CreateExtractElement(Mask, Idx, "Mask" + Twine(Idx)); 601 } 602 603 // Create "cond" block 604 // 605 // %Elt1 = extractelement <16 x i32> %Src, i32 1 606 // %Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1 607 // %store i32 %Elt1, i32* %Ptr1 608 // 609 Instruction *ThenTerm = 610 SplitBlockAndInsertIfThen(Predicate, InsertPt, /*Unreachable=*/false, 611 /*BranchWeights=*/nullptr, DTU); 612 613 BasicBlock *CondBlock = ThenTerm->getParent(); 614 CondBlock->setName("cond.store"); 615 616 Builder.SetInsertPoint(CondBlock->getTerminator()); 617 Value *OneElt = Builder.CreateExtractElement(Src, Idx, "Elt" + Twine(Idx)); 618 Value *Ptr = Builder.CreateExtractElement(Ptrs, Idx, "Ptr" + Twine(Idx)); 619 Builder.CreateAlignedStore(OneElt, Ptr, AlignVal); 620 621 // Create "else" block, fill it in the next iteration 622 BasicBlock *NewIfBlock = ThenTerm->getSuccessor(0); 623 NewIfBlock->setName("else"); 624 625 Builder.SetInsertPoint(NewIfBlock, NewIfBlock->begin()); 626 } 627 CI->eraseFromParent(); 628 629 ModifiedDT = true; 630 } 631 632 static void scalarizeMaskedExpandLoad(const DataLayout &DL, CallInst *CI, 633 DomTreeUpdater *DTU, bool &ModifiedDT) { 634 Value *Ptr = CI->getArgOperand(0); 635 Value *Mask = CI->getArgOperand(1); 636 Value *PassThru = CI->getArgOperand(2); 637 638 auto *VecType = cast<FixedVectorType>(CI->getType()); 639 640 Type *EltTy = VecType->getElementType(); 641 642 IRBuilder<> Builder(CI->getContext()); 643 Instruction *InsertPt = CI; 644 BasicBlock *IfBlock = CI->getParent(); 645 646 Builder.SetInsertPoint(InsertPt); 647 Builder.SetCurrentDebugLocation(CI->getDebugLoc()); 648 649 unsigned VectorWidth = VecType->getNumElements(); 650 651 // The result vector 652 Value *VResult = PassThru; 653 654 // Shorten the way if the mask is a vector of constants. 655 // Create a build_vector pattern, with loads/undefs as necessary and then 656 // shuffle blend with the pass through value. 657 if (isConstantIntVector(Mask)) { 658 unsigned MemIndex = 0; 659 VResult = UndefValue::get(VecType); 660 SmallVector<int, 16> ShuffleMask(VectorWidth, UndefMaskElem); 661 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) { 662 Value *InsertElt; 663 if (cast<Constant>(Mask)->getAggregateElement(Idx)->isNullValue()) { 664 InsertElt = UndefValue::get(EltTy); 665 ShuffleMask[Idx] = Idx + VectorWidth; 666 } else { 667 Value *NewPtr = 668 Builder.CreateConstInBoundsGEP1_32(EltTy, Ptr, MemIndex); 669 InsertElt = Builder.CreateAlignedLoad(EltTy, NewPtr, Align(1), 670 "Load" + Twine(Idx)); 671 ShuffleMask[Idx] = Idx; 672 ++MemIndex; 673 } 674 VResult = Builder.CreateInsertElement(VResult, InsertElt, Idx, 675 "Res" + Twine(Idx)); 676 } 677 VResult = Builder.CreateShuffleVector(VResult, PassThru, ShuffleMask); 678 CI->replaceAllUsesWith(VResult); 679 CI->eraseFromParent(); 680 return; 681 } 682 683 // If the mask is not v1i1, use scalar bit test operations. This generates 684 // better results on X86 at least. 685 Value *SclrMask; 686 if (VectorWidth != 1) { 687 Type *SclrMaskTy = Builder.getIntNTy(VectorWidth); 688 SclrMask = Builder.CreateBitCast(Mask, SclrMaskTy, "scalar_mask"); 689 } 690 691 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) { 692 // Fill the "else" block, created in the previous iteration 693 // 694 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ] 695 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx 696 // br i1 %mask_1, label %cond.load, label %else 697 // 698 699 Value *Predicate; 700 if (VectorWidth != 1) { 701 Value *Mask = Builder.getInt(APInt::getOneBitSet( 702 VectorWidth, adjustForEndian(DL, VectorWidth, Idx))); 703 Predicate = Builder.CreateICmpNE(Builder.CreateAnd(SclrMask, Mask), 704 Builder.getIntN(VectorWidth, 0)); 705 } else { 706 Predicate = Builder.CreateExtractElement(Mask, Idx, "Mask" + Twine(Idx)); 707 } 708 709 // Create "cond" block 710 // 711 // %EltAddr = getelementptr i32* %1, i32 0 712 // %Elt = load i32* %EltAddr 713 // VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx 714 // 715 Instruction *ThenTerm = 716 SplitBlockAndInsertIfThen(Predicate, InsertPt, /*Unreachable=*/false, 717 /*BranchWeights=*/nullptr, DTU); 718 719 BasicBlock *CondBlock = ThenTerm->getParent(); 720 CondBlock->setName("cond.load"); 721 722 Builder.SetInsertPoint(CondBlock->getTerminator()); 723 LoadInst *Load = Builder.CreateAlignedLoad(EltTy, Ptr, Align(1)); 724 Value *NewVResult = Builder.CreateInsertElement(VResult, Load, Idx); 725 726 // Move the pointer if there are more blocks to come. 727 Value *NewPtr; 728 if ((Idx + 1) != VectorWidth) 729 NewPtr = Builder.CreateConstInBoundsGEP1_32(EltTy, Ptr, 1); 730 731 // Create "else" block, fill it in the next iteration 732 BasicBlock *NewIfBlock = ThenTerm->getSuccessor(0); 733 NewIfBlock->setName("else"); 734 BasicBlock *PrevIfBlock = IfBlock; 735 IfBlock = NewIfBlock; 736 737 // Create the phi to join the new and previous value. 738 Builder.SetInsertPoint(NewIfBlock, NewIfBlock->begin()); 739 PHINode *ResultPhi = Builder.CreatePHI(VecType, 2, "res.phi.else"); 740 ResultPhi->addIncoming(NewVResult, CondBlock); 741 ResultPhi->addIncoming(VResult, PrevIfBlock); 742 VResult = ResultPhi; 743 744 // Add a PHI for the pointer if this isn't the last iteration. 745 if ((Idx + 1) != VectorWidth) { 746 PHINode *PtrPhi = Builder.CreatePHI(Ptr->getType(), 2, "ptr.phi.else"); 747 PtrPhi->addIncoming(NewPtr, CondBlock); 748 PtrPhi->addIncoming(Ptr, PrevIfBlock); 749 Ptr = PtrPhi; 750 } 751 } 752 753 CI->replaceAllUsesWith(VResult); 754 CI->eraseFromParent(); 755 756 ModifiedDT = true; 757 } 758 759 static void scalarizeMaskedCompressStore(const DataLayout &DL, CallInst *CI, 760 DomTreeUpdater *DTU, 761 bool &ModifiedDT) { 762 Value *Src = CI->getArgOperand(0); 763 Value *Ptr = CI->getArgOperand(1); 764 Value *Mask = CI->getArgOperand(2); 765 766 auto *VecType = cast<FixedVectorType>(Src->getType()); 767 768 IRBuilder<> Builder(CI->getContext()); 769 Instruction *InsertPt = CI; 770 BasicBlock *IfBlock = CI->getParent(); 771 772 Builder.SetInsertPoint(InsertPt); 773 Builder.SetCurrentDebugLocation(CI->getDebugLoc()); 774 775 Type *EltTy = VecType->getElementType(); 776 777 unsigned VectorWidth = VecType->getNumElements(); 778 779 // Shorten the way if the mask is a vector of constants. 780 if (isConstantIntVector(Mask)) { 781 unsigned MemIndex = 0; 782 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) { 783 if (cast<Constant>(Mask)->getAggregateElement(Idx)->isNullValue()) 784 continue; 785 Value *OneElt = 786 Builder.CreateExtractElement(Src, Idx, "Elt" + Twine(Idx)); 787 Value *NewPtr = Builder.CreateConstInBoundsGEP1_32(EltTy, Ptr, MemIndex); 788 Builder.CreateAlignedStore(OneElt, NewPtr, Align(1)); 789 ++MemIndex; 790 } 791 CI->eraseFromParent(); 792 return; 793 } 794 795 // If the mask is not v1i1, use scalar bit test operations. This generates 796 // better results on X86 at least. 797 Value *SclrMask; 798 if (VectorWidth != 1) { 799 Type *SclrMaskTy = Builder.getIntNTy(VectorWidth); 800 SclrMask = Builder.CreateBitCast(Mask, SclrMaskTy, "scalar_mask"); 801 } 802 803 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) { 804 // Fill the "else" block, created in the previous iteration 805 // 806 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx 807 // br i1 %mask_1, label %cond.store, label %else 808 // 809 Value *Predicate; 810 if (VectorWidth != 1) { 811 Value *Mask = Builder.getInt(APInt::getOneBitSet( 812 VectorWidth, adjustForEndian(DL, VectorWidth, Idx))); 813 Predicate = Builder.CreateICmpNE(Builder.CreateAnd(SclrMask, Mask), 814 Builder.getIntN(VectorWidth, 0)); 815 } else { 816 Predicate = Builder.CreateExtractElement(Mask, Idx, "Mask" + Twine(Idx)); 817 } 818 819 // Create "cond" block 820 // 821 // %OneElt = extractelement <16 x i32> %Src, i32 Idx 822 // %EltAddr = getelementptr i32* %1, i32 0 823 // %store i32 %OneElt, i32* %EltAddr 824 // 825 Instruction *ThenTerm = 826 SplitBlockAndInsertIfThen(Predicate, InsertPt, /*Unreachable=*/false, 827 /*BranchWeights=*/nullptr, DTU); 828 829 BasicBlock *CondBlock = ThenTerm->getParent(); 830 CondBlock->setName("cond.store"); 831 832 Builder.SetInsertPoint(CondBlock->getTerminator()); 833 Value *OneElt = Builder.CreateExtractElement(Src, Idx); 834 Builder.CreateAlignedStore(OneElt, Ptr, Align(1)); 835 836 // Move the pointer if there are more blocks to come. 837 Value *NewPtr; 838 if ((Idx + 1) != VectorWidth) 839 NewPtr = Builder.CreateConstInBoundsGEP1_32(EltTy, Ptr, 1); 840 841 // Create "else" block, fill it in the next iteration 842 BasicBlock *NewIfBlock = ThenTerm->getSuccessor(0); 843 NewIfBlock->setName("else"); 844 BasicBlock *PrevIfBlock = IfBlock; 845 IfBlock = NewIfBlock; 846 847 Builder.SetInsertPoint(NewIfBlock, NewIfBlock->begin()); 848 849 // Add a PHI for the pointer if this isn't the last iteration. 850 if ((Idx + 1) != VectorWidth) { 851 PHINode *PtrPhi = Builder.CreatePHI(Ptr->getType(), 2, "ptr.phi.else"); 852 PtrPhi->addIncoming(NewPtr, CondBlock); 853 PtrPhi->addIncoming(Ptr, PrevIfBlock); 854 Ptr = PtrPhi; 855 } 856 } 857 CI->eraseFromParent(); 858 859 ModifiedDT = true; 860 } 861 862 static bool runImpl(Function &F, const TargetTransformInfo &TTI, 863 DominatorTree *DT) { 864 Optional<DomTreeUpdater> DTU; 865 if (DT) 866 DTU.emplace(DT, DomTreeUpdater::UpdateStrategy::Lazy); 867 868 bool EverMadeChange = false; 869 bool MadeChange = true; 870 auto &DL = F.getParent()->getDataLayout(); 871 while (MadeChange) { 872 MadeChange = false; 873 for (BasicBlock &BB : llvm::make_early_inc_range(F)) { 874 bool ModifiedDTOnIteration = false; 875 MadeChange |= optimizeBlock(BB, ModifiedDTOnIteration, TTI, DL, 876 DTU ? DTU.getPointer() : nullptr); 877 878 // Restart BB iteration if the dominator tree of the Function was changed 879 if (ModifiedDTOnIteration) 880 break; 881 } 882 883 EverMadeChange |= MadeChange; 884 } 885 return EverMadeChange; 886 } 887 888 bool ScalarizeMaskedMemIntrinLegacyPass::runOnFunction(Function &F) { 889 auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); 890 DominatorTree *DT = nullptr; 891 if (auto *DTWP = getAnalysisIfAvailable<DominatorTreeWrapperPass>()) 892 DT = &DTWP->getDomTree(); 893 return runImpl(F, TTI, DT); 894 } 895 896 PreservedAnalyses 897 ScalarizeMaskedMemIntrinPass::run(Function &F, FunctionAnalysisManager &AM) { 898 auto &TTI = AM.getResult<TargetIRAnalysis>(F); 899 auto *DT = AM.getCachedResult<DominatorTreeAnalysis>(F); 900 if (!runImpl(F, TTI, DT)) 901 return PreservedAnalyses::all(); 902 PreservedAnalyses PA; 903 PA.preserve<TargetIRAnalysis>(); 904 PA.preserve<DominatorTreeAnalysis>(); 905 return PA; 906 } 907 908 static bool optimizeBlock(BasicBlock &BB, bool &ModifiedDT, 909 const TargetTransformInfo &TTI, const DataLayout &DL, 910 DomTreeUpdater *DTU) { 911 bool MadeChange = false; 912 913 BasicBlock::iterator CurInstIterator = BB.begin(); 914 while (CurInstIterator != BB.end()) { 915 if (CallInst *CI = dyn_cast<CallInst>(&*CurInstIterator++)) 916 MadeChange |= optimizeCallInst(CI, ModifiedDT, TTI, DL, DTU); 917 if (ModifiedDT) 918 return true; 919 } 920 921 return MadeChange; 922 } 923 924 static bool optimizeCallInst(CallInst *CI, bool &ModifiedDT, 925 const TargetTransformInfo &TTI, 926 const DataLayout &DL, DomTreeUpdater *DTU) { 927 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI); 928 if (II) { 929 // The scalarization code below does not work for scalable vectors. 930 if (isa<ScalableVectorType>(II->getType()) || 931 any_of(II->args(), 932 [](Value *V) { return isa<ScalableVectorType>(V->getType()); })) 933 return false; 934 935 switch (II->getIntrinsicID()) { 936 default: 937 break; 938 case Intrinsic::masked_load: 939 // Scalarize unsupported vector masked load 940 if (TTI.isLegalMaskedLoad( 941 CI->getType(), 942 cast<ConstantInt>(CI->getArgOperand(1))->getAlignValue())) 943 return false; 944 scalarizeMaskedLoad(DL, CI, DTU, ModifiedDT); 945 return true; 946 case Intrinsic::masked_store: 947 if (TTI.isLegalMaskedStore( 948 CI->getArgOperand(0)->getType(), 949 cast<ConstantInt>(CI->getArgOperand(2))->getAlignValue())) 950 return false; 951 scalarizeMaskedStore(DL, CI, DTU, ModifiedDT); 952 return true; 953 case Intrinsic::masked_gather: { 954 MaybeAlign MA = 955 cast<ConstantInt>(CI->getArgOperand(1))->getMaybeAlignValue(); 956 Type *LoadTy = CI->getType(); 957 Align Alignment = DL.getValueOrABITypeAlignment(MA, 958 LoadTy->getScalarType()); 959 if (TTI.isLegalMaskedGather(LoadTy, Alignment) && 960 !TTI.forceScalarizeMaskedGather(cast<VectorType>(LoadTy), Alignment)) 961 return false; 962 scalarizeMaskedGather(DL, CI, DTU, ModifiedDT); 963 return true; 964 } 965 case Intrinsic::masked_scatter: { 966 MaybeAlign MA = 967 cast<ConstantInt>(CI->getArgOperand(2))->getMaybeAlignValue(); 968 Type *StoreTy = CI->getArgOperand(0)->getType(); 969 Align Alignment = DL.getValueOrABITypeAlignment(MA, 970 StoreTy->getScalarType()); 971 if (TTI.isLegalMaskedScatter(StoreTy, Alignment) && 972 !TTI.forceScalarizeMaskedScatter(cast<VectorType>(StoreTy), 973 Alignment)) 974 return false; 975 scalarizeMaskedScatter(DL, CI, DTU, ModifiedDT); 976 return true; 977 } 978 case Intrinsic::masked_expandload: 979 if (TTI.isLegalMaskedExpandLoad(CI->getType())) 980 return false; 981 scalarizeMaskedExpandLoad(DL, CI, DTU, ModifiedDT); 982 return true; 983 case Intrinsic::masked_compressstore: 984 if (TTI.isLegalMaskedCompressStore(CI->getArgOperand(0)->getType())) 985 return false; 986 scalarizeMaskedCompressStore(DL, CI, DTU, ModifiedDT); 987 return true; 988 } 989 } 990 991 return false; 992 } 993