1 //===- TruncInstCombine.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 // TruncInstCombine - looks for expression graphs post-dominated by TruncInst 10 // and for each eligible graph, it will create a reduced bit-width expression, 11 // replace the old expression with this new one and remove the old expression. 12 // Eligible expression graph is such that: 13 // 1. Contains only supported instructions. 14 // 2. Supported leaves: ZExtInst, SExtInst, TruncInst and Constant value. 15 // 3. Can be evaluated into type with reduced legal bit-width. 16 // 4. All instructions in the graph must not have users outside the graph. 17 // The only exception is for {ZExt, SExt}Inst with operand type equal to 18 // the new reduced type evaluated in (3). 19 // 20 // The motivation for this optimization is that evaluating and expression using 21 // smaller bit-width is preferable, especially for vectorization where we can 22 // fit more values in one vectorized instruction. In addition, this optimization 23 // may decrease the number of cast instructions, but will not increase it. 24 // 25 //===----------------------------------------------------------------------===// 26 27 #include "AggressiveInstCombineInternal.h" 28 #include "llvm/ADT/STLExtras.h" 29 #include "llvm/ADT/Statistic.h" 30 #include "llvm/Analysis/ConstantFolding.h" 31 #include "llvm/IR/DataLayout.h" 32 #include "llvm/IR/Dominators.h" 33 #include "llvm/IR/IRBuilder.h" 34 #include "llvm/IR/Instruction.h" 35 #include "llvm/Support/KnownBits.h" 36 37 using namespace llvm; 38 39 #define DEBUG_TYPE "aggressive-instcombine" 40 41 STATISTIC(NumExprsReduced, "Number of truncations eliminated by reducing bit " 42 "width of expression graph"); 43 STATISTIC(NumInstrsReduced, 44 "Number of instructions whose bit width was reduced"); 45 46 /// Given an instruction and a container, it fills all the relevant operands of 47 /// that instruction, with respect to the Trunc expression graph optimizaton. 48 static void getRelevantOperands(Instruction *I, SmallVectorImpl<Value *> &Ops) { 49 unsigned Opc = I->getOpcode(); 50 switch (Opc) { 51 case Instruction::Trunc: 52 case Instruction::ZExt: 53 case Instruction::SExt: 54 // These CastInst are considered leaves of the evaluated expression, thus, 55 // their operands are not relevent. 56 break; 57 case Instruction::Add: 58 case Instruction::Sub: 59 case Instruction::Mul: 60 case Instruction::And: 61 case Instruction::Or: 62 case Instruction::Xor: 63 case Instruction::Shl: 64 case Instruction::LShr: 65 case Instruction::AShr: 66 case Instruction::UDiv: 67 case Instruction::URem: 68 case Instruction::InsertElement: 69 Ops.push_back(I->getOperand(0)); 70 Ops.push_back(I->getOperand(1)); 71 break; 72 case Instruction::ExtractElement: 73 Ops.push_back(I->getOperand(0)); 74 break; 75 case Instruction::Select: 76 Ops.push_back(I->getOperand(1)); 77 Ops.push_back(I->getOperand(2)); 78 break; 79 case Instruction::PHI: 80 for (Value *V : cast<PHINode>(I)->incoming_values()) 81 Ops.push_back(V); 82 break; 83 default: 84 llvm_unreachable("Unreachable!"); 85 } 86 } 87 88 bool TruncInstCombine::buildTruncExpressionGraph() { 89 SmallVector<Value *, 8> Worklist; 90 SmallVector<Instruction *, 8> Stack; 91 // Clear old instructions info. 92 InstInfoMap.clear(); 93 94 Worklist.push_back(CurrentTruncInst->getOperand(0)); 95 96 while (!Worklist.empty()) { 97 Value *Curr = Worklist.back(); 98 99 if (isa<Constant>(Curr)) { 100 Worklist.pop_back(); 101 continue; 102 } 103 104 auto *I = dyn_cast<Instruction>(Curr); 105 if (!I) 106 return false; 107 108 if (!Stack.empty() && Stack.back() == I) { 109 // Already handled all instruction operands, can remove it from both the 110 // Worklist and the Stack, and add it to the instruction info map. 111 Worklist.pop_back(); 112 Stack.pop_back(); 113 // Insert I to the Info map. 114 InstInfoMap.insert(std::make_pair(I, Info())); 115 continue; 116 } 117 118 if (InstInfoMap.count(I)) { 119 Worklist.pop_back(); 120 continue; 121 } 122 123 // Add the instruction to the stack before start handling its operands. 124 Stack.push_back(I); 125 126 unsigned Opc = I->getOpcode(); 127 switch (Opc) { 128 case Instruction::Trunc: 129 case Instruction::ZExt: 130 case Instruction::SExt: 131 // trunc(trunc(x)) -> trunc(x) 132 // trunc(ext(x)) -> ext(x) if the source type is smaller than the new dest 133 // trunc(ext(x)) -> trunc(x) if the source type is larger than the new 134 // dest 135 break; 136 case Instruction::Add: 137 case Instruction::Sub: 138 case Instruction::Mul: 139 case Instruction::And: 140 case Instruction::Or: 141 case Instruction::Xor: 142 case Instruction::Shl: 143 case Instruction::LShr: 144 case Instruction::AShr: 145 case Instruction::UDiv: 146 case Instruction::URem: 147 case Instruction::InsertElement: 148 case Instruction::ExtractElement: 149 case Instruction::Select: { 150 SmallVector<Value *, 2> Operands; 151 getRelevantOperands(I, Operands); 152 append_range(Worklist, Operands); 153 break; 154 } 155 case Instruction::PHI: { 156 SmallVector<Value *, 2> Operands; 157 getRelevantOperands(I, Operands); 158 // Add only operands not in Stack to prevent cycle 159 for (auto *Op : Operands) 160 if (!llvm::is_contained(Stack, Op)) 161 Worklist.push_back(Op); 162 break; 163 } 164 default: 165 // TODO: Can handle more cases here: 166 // 1. shufflevector 167 // 2. sdiv, srem 168 // ... 169 return false; 170 } 171 } 172 return true; 173 } 174 175 unsigned TruncInstCombine::getMinBitWidth() { 176 SmallVector<Value *, 8> Worklist; 177 SmallVector<Instruction *, 8> Stack; 178 179 Value *Src = CurrentTruncInst->getOperand(0); 180 Type *DstTy = CurrentTruncInst->getType(); 181 unsigned TruncBitWidth = DstTy->getScalarSizeInBits(); 182 unsigned OrigBitWidth = 183 CurrentTruncInst->getOperand(0)->getType()->getScalarSizeInBits(); 184 185 if (isa<Constant>(Src)) 186 return TruncBitWidth; 187 188 Worklist.push_back(Src); 189 InstInfoMap[cast<Instruction>(Src)].ValidBitWidth = TruncBitWidth; 190 191 while (!Worklist.empty()) { 192 Value *Curr = Worklist.back(); 193 194 if (isa<Constant>(Curr)) { 195 Worklist.pop_back(); 196 continue; 197 } 198 199 // Otherwise, it must be an instruction. 200 auto *I = cast<Instruction>(Curr); 201 202 auto &Info = InstInfoMap[I]; 203 204 SmallVector<Value *, 2> Operands; 205 getRelevantOperands(I, Operands); 206 207 if (!Stack.empty() && Stack.back() == I) { 208 // Already handled all instruction operands, can remove it from both, the 209 // Worklist and the Stack, and update MinBitWidth. 210 Worklist.pop_back(); 211 Stack.pop_back(); 212 for (auto *Operand : Operands) 213 if (auto *IOp = dyn_cast<Instruction>(Operand)) 214 Info.MinBitWidth = 215 std::max(Info.MinBitWidth, InstInfoMap[IOp].MinBitWidth); 216 continue; 217 } 218 219 // Add the instruction to the stack before start handling its operands. 220 Stack.push_back(I); 221 unsigned ValidBitWidth = Info.ValidBitWidth; 222 223 // Update minimum bit-width before handling its operands. This is required 224 // when the instruction is part of a loop. 225 Info.MinBitWidth = std::max(Info.MinBitWidth, Info.ValidBitWidth); 226 227 for (auto *Operand : Operands) 228 if (auto *IOp = dyn_cast<Instruction>(Operand)) { 229 // If we already calculated the minimum bit-width for this valid 230 // bit-width, or for a smaller valid bit-width, then just keep the 231 // answer we already calculated. 232 unsigned IOpBitwidth = InstInfoMap.lookup(IOp).ValidBitWidth; 233 if (IOpBitwidth >= ValidBitWidth) 234 continue; 235 InstInfoMap[IOp].ValidBitWidth = ValidBitWidth; 236 Worklist.push_back(IOp); 237 } 238 } 239 unsigned MinBitWidth = InstInfoMap.lookup(cast<Instruction>(Src)).MinBitWidth; 240 assert(MinBitWidth >= TruncBitWidth); 241 242 if (MinBitWidth > TruncBitWidth) { 243 // In this case reducing expression with vector type might generate a new 244 // vector type, which is not preferable as it might result in generating 245 // sub-optimal code. 246 if (DstTy->isVectorTy()) 247 return OrigBitWidth; 248 // Use the smallest integer type in the range [MinBitWidth, OrigBitWidth). 249 Type *Ty = DL.getSmallestLegalIntType(DstTy->getContext(), MinBitWidth); 250 // Update minimum bit-width with the new destination type bit-width if 251 // succeeded to find such, otherwise, with original bit-width. 252 MinBitWidth = Ty ? Ty->getScalarSizeInBits() : OrigBitWidth; 253 } else { // MinBitWidth == TruncBitWidth 254 // In this case the expression can be evaluated with the trunc instruction 255 // destination type, and trunc instruction can be omitted. However, we 256 // should not perform the evaluation if the original type is a legal scalar 257 // type and the target type is illegal. 258 bool FromLegal = MinBitWidth == 1 || DL.isLegalInteger(OrigBitWidth); 259 bool ToLegal = MinBitWidth == 1 || DL.isLegalInteger(MinBitWidth); 260 if (!DstTy->isVectorTy() && FromLegal && !ToLegal) 261 return OrigBitWidth; 262 } 263 return MinBitWidth; 264 } 265 266 Type *TruncInstCombine::getBestTruncatedType() { 267 if (!buildTruncExpressionGraph()) 268 return nullptr; 269 270 // We don't want to duplicate instructions, which isn't profitable. Thus, we 271 // can't shrink something that has multiple users, unless all users are 272 // post-dominated by the trunc instruction, i.e., were visited during the 273 // expression evaluation. 274 unsigned DesiredBitWidth = 0; 275 for (auto Itr : InstInfoMap) { 276 Instruction *I = Itr.first; 277 if (I->hasOneUse()) 278 continue; 279 bool IsExtInst = (isa<ZExtInst>(I) || isa<SExtInst>(I)); 280 for (auto *U : I->users()) 281 if (auto *UI = dyn_cast<Instruction>(U)) 282 if (UI != CurrentTruncInst && !InstInfoMap.count(UI)) { 283 if (!IsExtInst) 284 return nullptr; 285 // If this is an extension from the dest type, we can eliminate it, 286 // even if it has multiple users. Thus, update the DesiredBitWidth and 287 // validate all extension instructions agrees on same DesiredBitWidth. 288 unsigned ExtInstBitWidth = 289 I->getOperand(0)->getType()->getScalarSizeInBits(); 290 if (DesiredBitWidth && DesiredBitWidth != ExtInstBitWidth) 291 return nullptr; 292 DesiredBitWidth = ExtInstBitWidth; 293 } 294 } 295 296 unsigned OrigBitWidth = 297 CurrentTruncInst->getOperand(0)->getType()->getScalarSizeInBits(); 298 299 // Initialize MinBitWidth for shift instructions with the minimum number 300 // that is greater than shift amount (i.e. shift amount + 1). 301 // For `lshr` adjust MinBitWidth so that all potentially truncated 302 // bits of the value-to-be-shifted are zeros. 303 // For `ashr` adjust MinBitWidth so that all potentially truncated 304 // bits of the value-to-be-shifted are sign bits (all zeros or ones) 305 // and even one (first) untruncated bit is sign bit. 306 // Exit early if MinBitWidth is not less than original bitwidth. 307 for (auto &Itr : InstInfoMap) { 308 Instruction *I = Itr.first; 309 if (I->isShift()) { 310 KnownBits KnownRHS = computeKnownBits(I->getOperand(1)); 311 unsigned MinBitWidth = KnownRHS.getMaxValue() 312 .uadd_sat(APInt(OrigBitWidth, 1)) 313 .getLimitedValue(OrigBitWidth); 314 if (MinBitWidth == OrigBitWidth) 315 return nullptr; 316 if (I->getOpcode() == Instruction::LShr) { 317 KnownBits KnownLHS = computeKnownBits(I->getOperand(0)); 318 MinBitWidth = 319 std::max(MinBitWidth, KnownLHS.getMaxValue().getActiveBits()); 320 } 321 if (I->getOpcode() == Instruction::AShr) { 322 unsigned NumSignBits = ComputeNumSignBits(I->getOperand(0)); 323 MinBitWidth = std::max(MinBitWidth, OrigBitWidth - NumSignBits + 1); 324 } 325 if (MinBitWidth >= OrigBitWidth) 326 return nullptr; 327 Itr.second.MinBitWidth = MinBitWidth; 328 } 329 if (I->getOpcode() == Instruction::UDiv || 330 I->getOpcode() == Instruction::URem) { 331 unsigned MinBitWidth = 0; 332 for (const auto &Op : I->operands()) { 333 KnownBits Known = computeKnownBits(Op); 334 MinBitWidth = 335 std::max(Known.getMaxValue().getActiveBits(), MinBitWidth); 336 if (MinBitWidth >= OrigBitWidth) 337 return nullptr; 338 } 339 Itr.second.MinBitWidth = MinBitWidth; 340 } 341 } 342 343 // Calculate minimum allowed bit-width allowed for shrinking the currently 344 // visited truncate's operand. 345 unsigned MinBitWidth = getMinBitWidth(); 346 347 // Check that we can shrink to smaller bit-width than original one and that 348 // it is similar to the DesiredBitWidth is such exists. 349 if (MinBitWidth >= OrigBitWidth || 350 (DesiredBitWidth && DesiredBitWidth != MinBitWidth)) 351 return nullptr; 352 353 return IntegerType::get(CurrentTruncInst->getContext(), MinBitWidth); 354 } 355 356 /// Given a reduced scalar type \p Ty and a \p V value, return a reduced type 357 /// for \p V, according to its type, if it vector type, return the vector 358 /// version of \p Ty, otherwise return \p Ty. 359 static Type *getReducedType(Value *V, Type *Ty) { 360 assert(Ty && !Ty->isVectorTy() && "Expect Scalar Type"); 361 if (auto *VTy = dyn_cast<VectorType>(V->getType())) 362 return VectorType::get(Ty, VTy->getElementCount()); 363 return Ty; 364 } 365 366 Value *TruncInstCombine::getReducedOperand(Value *V, Type *SclTy) { 367 Type *Ty = getReducedType(V, SclTy); 368 if (auto *C = dyn_cast<Constant>(V)) { 369 C = ConstantExpr::getIntegerCast(C, Ty, false); 370 // If we got a constantexpr back, try to simplify it with DL info. 371 return ConstantFoldConstant(C, DL, &TLI); 372 } 373 374 auto *I = cast<Instruction>(V); 375 Info Entry = InstInfoMap.lookup(I); 376 assert(Entry.NewValue); 377 return Entry.NewValue; 378 } 379 380 void TruncInstCombine::ReduceExpressionGraph(Type *SclTy) { 381 NumInstrsReduced += InstInfoMap.size(); 382 // Pairs of old and new phi-nodes 383 SmallVector<std::pair<PHINode *, PHINode *>, 2> OldNewPHINodes; 384 for (auto &Itr : InstInfoMap) { // Forward 385 Instruction *I = Itr.first; 386 TruncInstCombine::Info &NodeInfo = Itr.second; 387 388 assert(!NodeInfo.NewValue && "Instruction has been evaluated"); 389 390 IRBuilder<> Builder(I); 391 Value *Res = nullptr; 392 unsigned Opc = I->getOpcode(); 393 switch (Opc) { 394 case Instruction::Trunc: 395 case Instruction::ZExt: 396 case Instruction::SExt: { 397 Type *Ty = getReducedType(I, SclTy); 398 // If the source type of the cast is the type we're trying for then we can 399 // just return the source. There's no need to insert it because it is not 400 // new. 401 if (I->getOperand(0)->getType() == Ty) { 402 assert(!isa<TruncInst>(I) && "Cannot reach here with TruncInst"); 403 NodeInfo.NewValue = I->getOperand(0); 404 continue; 405 } 406 // Otherwise, must be the same type of cast, so just reinsert a new one. 407 // This also handles the case of zext(trunc(x)) -> zext(x). 408 Res = Builder.CreateIntCast(I->getOperand(0), Ty, 409 Opc == Instruction::SExt); 410 411 // Update Worklist entries with new value if needed. 412 // There are three possible changes to the Worklist: 413 // 1. Update Old-TruncInst -> New-TruncInst. 414 // 2. Remove Old-TruncInst (if New node is not TruncInst). 415 // 3. Add New-TruncInst (if Old node was not TruncInst). 416 auto *Entry = find(Worklist, I); 417 if (Entry != Worklist.end()) { 418 if (auto *NewCI = dyn_cast<TruncInst>(Res)) 419 *Entry = NewCI; 420 else 421 Worklist.erase(Entry); 422 } else if (auto *NewCI = dyn_cast<TruncInst>(Res)) 423 Worklist.push_back(NewCI); 424 break; 425 } 426 case Instruction::Add: 427 case Instruction::Sub: 428 case Instruction::Mul: 429 case Instruction::And: 430 case Instruction::Or: 431 case Instruction::Xor: 432 case Instruction::Shl: 433 case Instruction::LShr: 434 case Instruction::AShr: 435 case Instruction::UDiv: 436 case Instruction::URem: { 437 Value *LHS = getReducedOperand(I->getOperand(0), SclTy); 438 Value *RHS = getReducedOperand(I->getOperand(1), SclTy); 439 Res = Builder.CreateBinOp((Instruction::BinaryOps)Opc, LHS, RHS); 440 // Preserve `exact` flag since truncation doesn't change exactness 441 if (auto *PEO = dyn_cast<PossiblyExactOperator>(I)) 442 if (auto *ResI = dyn_cast<Instruction>(Res)) 443 ResI->setIsExact(PEO->isExact()); 444 break; 445 } 446 case Instruction::ExtractElement: { 447 Value *Vec = getReducedOperand(I->getOperand(0), SclTy); 448 Value *Idx = I->getOperand(1); 449 Res = Builder.CreateExtractElement(Vec, Idx); 450 break; 451 } 452 case Instruction::InsertElement: { 453 Value *Vec = getReducedOperand(I->getOperand(0), SclTy); 454 Value *NewElt = getReducedOperand(I->getOperand(1), SclTy); 455 Value *Idx = I->getOperand(2); 456 Res = Builder.CreateInsertElement(Vec, NewElt, Idx); 457 break; 458 } 459 case Instruction::Select: { 460 Value *Op0 = I->getOperand(0); 461 Value *LHS = getReducedOperand(I->getOperand(1), SclTy); 462 Value *RHS = getReducedOperand(I->getOperand(2), SclTy); 463 Res = Builder.CreateSelect(Op0, LHS, RHS); 464 break; 465 } 466 case Instruction::PHI: { 467 Res = Builder.CreatePHI(getReducedType(I, SclTy), I->getNumOperands()); 468 OldNewPHINodes.push_back( 469 std::make_pair(cast<PHINode>(I), cast<PHINode>(Res))); 470 break; 471 } 472 default: 473 llvm_unreachable("Unhandled instruction"); 474 } 475 476 NodeInfo.NewValue = Res; 477 if (auto *ResI = dyn_cast<Instruction>(Res)) 478 ResI->takeName(I); 479 } 480 481 for (auto &Node : OldNewPHINodes) { 482 PHINode *OldPN = Node.first; 483 PHINode *NewPN = Node.second; 484 for (auto Incoming : zip(OldPN->incoming_values(), OldPN->blocks())) 485 NewPN->addIncoming(getReducedOperand(std::get<0>(Incoming), SclTy), 486 std::get<1>(Incoming)); 487 } 488 489 Value *Res = getReducedOperand(CurrentTruncInst->getOperand(0), SclTy); 490 Type *DstTy = CurrentTruncInst->getType(); 491 if (Res->getType() != DstTy) { 492 IRBuilder<> Builder(CurrentTruncInst); 493 Res = Builder.CreateIntCast(Res, DstTy, false); 494 if (auto *ResI = dyn_cast<Instruction>(Res)) 495 ResI->takeName(CurrentTruncInst); 496 } 497 CurrentTruncInst->replaceAllUsesWith(Res); 498 499 // Erase old expression graph, which was replaced by the reduced expression 500 // graph. 501 CurrentTruncInst->eraseFromParent(); 502 // First, erase old phi-nodes and its uses 503 for (auto &Node : OldNewPHINodes) { 504 PHINode *OldPN = Node.first; 505 OldPN->replaceAllUsesWith(PoisonValue::get(OldPN->getType())); 506 InstInfoMap.erase(OldPN); 507 OldPN->eraseFromParent(); 508 } 509 // Now we have expression graph turned into dag. 510 // We iterate backward, which means we visit the instruction before we 511 // visit any of its operands, this way, when we get to the operand, we already 512 // removed the instructions (from the expression dag) that uses it. 513 for (auto &I : llvm::reverse(InstInfoMap)) { 514 // We still need to check that the instruction has no users before we erase 515 // it, because {SExt, ZExt}Inst Instruction might have other users that was 516 // not reduced, in such case, we need to keep that instruction. 517 if (I.first->use_empty()) 518 I.first->eraseFromParent(); 519 else 520 assert((isa<SExtInst>(I.first) || isa<ZExtInst>(I.first)) && 521 "Only {SExt, ZExt}Inst might have unreduced users"); 522 } 523 } 524 525 bool TruncInstCombine::run(Function &F) { 526 bool MadeIRChange = false; 527 528 // Collect all TruncInst in the function into the Worklist for evaluating. 529 for (auto &BB : F) { 530 // Ignore unreachable basic block. 531 if (!DT.isReachableFromEntry(&BB)) 532 continue; 533 for (auto &I : BB) 534 if (auto *CI = dyn_cast<TruncInst>(&I)) 535 Worklist.push_back(CI); 536 } 537 538 // Process all TruncInst in the Worklist, for each instruction: 539 // 1. Check if it dominates an eligible expression graph to be reduced. 540 // 2. Create a reduced expression graph and replace the old one with it. 541 while (!Worklist.empty()) { 542 CurrentTruncInst = Worklist.pop_back_val(); 543 544 if (Type *NewDstSclTy = getBestTruncatedType()) { 545 LLVM_DEBUG( 546 dbgs() << "ICE: TruncInstCombine reducing type of expression graph " 547 "dominated by: " 548 << CurrentTruncInst << '\n'); 549 ReduceExpressionGraph(NewDstSclTy); 550 ++NumExprsReduced; 551 MadeIRChange = true; 552 } 553 } 554 555 return MadeIRChange; 556 } 557