3236 lines · cpp
1//===- InstCombineCasts.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-exception6//7//===----------------------------------------------------------------------===//8//9// This file implements the visit functions for cast operations.10//11//===----------------------------------------------------------------------===//12 13#include "InstCombineInternal.h"14#include "llvm/ADT/APInt.h"15#include "llvm/ADT/DenseMap.h"16#include "llvm/ADT/STLExtras.h"17#include "llvm/ADT/STLFunctionalExtras.h"18#include "llvm/ADT/SetVector.h"19#include "llvm/ADT/SmallVector.h"20#include "llvm/Analysis/ConstantFolding.h"21#include "llvm/IR/DataLayout.h"22#include "llvm/IR/DebugInfo.h"23#include "llvm/IR/Instruction.h"24#include "llvm/IR/PatternMatch.h"25#include "llvm/IR/Type.h"26#include "llvm/IR/Value.h"27#include "llvm/Support/KnownBits.h"28#include "llvm/Transforms/InstCombine/InstCombiner.h"29#include <iterator>30#include <optional>31 32using namespace llvm;33using namespace PatternMatch;34 35#define DEBUG_TYPE "instcombine"36 37using EvaluatedMap = SmallDenseMap<Value *, Value *, 8>;38 39static Value *EvaluateInDifferentTypeImpl(Value *V, Type *Ty, bool isSigned,40 InstCombinerImpl &IC,41 EvaluatedMap &Processed) {42 // Since we cover transformation of instructions with multiple users, we might43 // come to the same node via multiple paths. We should not create a44 // replacement for every single one of them though.45 if (Value *Result = Processed.lookup(V))46 return Result;47 48 if (Constant *C = dyn_cast<Constant>(V))49 return ConstantFoldIntegerCast(C, Ty, isSigned, IC.getDataLayout());50 51 // Otherwise, it must be an instruction.52 Instruction *I = cast<Instruction>(V);53 Instruction *Res = nullptr;54 unsigned Opc = I->getOpcode();55 switch (Opc) {56 case Instruction::Add:57 case Instruction::Sub:58 case Instruction::Mul:59 case Instruction::And:60 case Instruction::Or:61 case Instruction::Xor:62 case Instruction::AShr:63 case Instruction::LShr:64 case Instruction::Shl:65 case Instruction::UDiv:66 case Instruction::URem: {67 Value *LHS = EvaluateInDifferentTypeImpl(I->getOperand(0), Ty, isSigned, IC,68 Processed);69 Value *RHS = EvaluateInDifferentTypeImpl(I->getOperand(1), Ty, isSigned, IC,70 Processed);71 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS);72 if (Opc == Instruction::LShr || Opc == Instruction::AShr)73 Res->setIsExact(I->isExact());74 break;75 }76 case Instruction::Trunc:77 case Instruction::ZExt:78 case Instruction::SExt:79 // If the source type of the cast is the type we're trying for then we can80 // just return the source. There's no need to insert it because it is not81 // new.82 if (I->getOperand(0)->getType() == Ty)83 return I->getOperand(0);84 85 // Otherwise, must be the same type of cast, so just reinsert a new one.86 // This also handles the case of zext(trunc(x)) -> zext(x).87 Res = CastInst::CreateIntegerCast(I->getOperand(0), Ty,88 Opc == Instruction::SExt);89 break;90 case Instruction::Select: {91 Value *True = EvaluateInDifferentTypeImpl(I->getOperand(1), Ty, isSigned,92 IC, Processed);93 Value *False = EvaluateInDifferentTypeImpl(I->getOperand(2), Ty, isSigned,94 IC, Processed);95 Res = SelectInst::Create(I->getOperand(0), True, False);96 break;97 }98 case Instruction::PHI: {99 PHINode *OPN = cast<PHINode>(I);100 PHINode *NPN = PHINode::Create(Ty, OPN->getNumIncomingValues());101 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {102 Value *V = EvaluateInDifferentTypeImpl(OPN->getIncomingValue(i), Ty,103 isSigned, IC, Processed);104 NPN->addIncoming(V, OPN->getIncomingBlock(i));105 }106 Res = NPN;107 break;108 }109 case Instruction::FPToUI:110 case Instruction::FPToSI:111 Res = CastInst::Create(static_cast<Instruction::CastOps>(Opc),112 I->getOperand(0), Ty);113 break;114 case Instruction::Call:115 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {116 switch (II->getIntrinsicID()) {117 default:118 llvm_unreachable("Unsupported call!");119 case Intrinsic::vscale: {120 Function *Fn = Intrinsic::getOrInsertDeclaration(121 I->getModule(), Intrinsic::vscale, {Ty});122 Res = CallInst::Create(Fn->getFunctionType(), Fn);123 break;124 }125 }126 }127 break;128 case Instruction::ShuffleVector: {129 auto *ScalarTy = cast<VectorType>(Ty)->getElementType();130 auto *VTy = cast<VectorType>(I->getOperand(0)->getType());131 auto *FixedTy = VectorType::get(ScalarTy, VTy->getElementCount());132 Value *Op0 = EvaluateInDifferentTypeImpl(I->getOperand(0), FixedTy,133 isSigned, IC, Processed);134 Value *Op1 = EvaluateInDifferentTypeImpl(I->getOperand(1), FixedTy,135 isSigned, IC, Processed);136 Res = new ShuffleVectorInst(Op0, Op1,137 cast<ShuffleVectorInst>(I)->getShuffleMask());138 break;139 }140 default:141 // TODO: Can handle more cases here.142 llvm_unreachable("Unreachable!");143 }144 145 Res->takeName(I);146 Value *Result = IC.InsertNewInstWith(Res, I->getIterator());147 // There is no need in keeping track of the old value/new value relationship148 // when we have only one user, we came have here from that user and no-one149 // else cares.150 if (!V->hasOneUse())151 Processed[V] = Result;152 153 return Result;154}155 156/// Given an expression that CanEvaluateTruncated or CanEvaluateSExtd returns157/// true for, actually insert the code to evaluate the expression.158Value *InstCombinerImpl::EvaluateInDifferentType(Value *V, Type *Ty,159 bool isSigned) {160 EvaluatedMap Processed;161 return EvaluateInDifferentTypeImpl(V, Ty, isSigned, *this, Processed);162}163 164Instruction::CastOps165InstCombinerImpl::isEliminableCastPair(const CastInst *CI1,166 const CastInst *CI2) {167 Type *SrcTy = CI1->getSrcTy();168 Type *MidTy = CI1->getDestTy();169 Type *DstTy = CI2->getDestTy();170 171 Instruction::CastOps firstOp = CI1->getOpcode();172 Instruction::CastOps secondOp = CI2->getOpcode();173 Type *SrcIntPtrTy =174 SrcTy->isPtrOrPtrVectorTy() ? DL.getIntPtrType(SrcTy) : nullptr;175 Type *DstIntPtrTy =176 DstTy->isPtrOrPtrVectorTy() ? DL.getIntPtrType(DstTy) : nullptr;177 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,178 DstTy, &DL);179 180 // We don't want to form an inttoptr or ptrtoint that converts to an integer181 // type that differs from the pointer size.182 if ((Res == Instruction::IntToPtr && SrcTy != DstIntPtrTy) ||183 (Res == Instruction::PtrToInt && DstTy != SrcIntPtrTy))184 Res = 0;185 186 return Instruction::CastOps(Res);187}188 189/// Implement the transforms common to all CastInst visitors.190Instruction *InstCombinerImpl::commonCastTransforms(CastInst &CI) {191 Value *Src = CI.getOperand(0);192 Type *Ty = CI.getType();193 194 if (Value *Res =195 simplifyCastInst(CI.getOpcode(), Src, Ty, SQ.getWithInstruction(&CI)))196 return replaceInstUsesWith(CI, Res);197 198 // Try to eliminate a cast of a cast.199 if (auto *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast200 if (Instruction::CastOps NewOpc = isEliminableCastPair(CSrc, &CI)) {201 // The first cast (CSrc) is eliminable so we need to fix up or replace202 // the second cast (CI). CSrc will then have a good chance of being dead.203 auto *Res = CastInst::Create(NewOpc, CSrc->getOperand(0), Ty);204 // Point debug users of the dying cast to the new one.205 if (CSrc->hasOneUse())206 replaceAllDbgUsesWith(*CSrc, *Res, CI, DT);207 return Res;208 }209 }210 211 if (auto *Sel = dyn_cast<SelectInst>(Src)) {212 // We are casting a select. Try to fold the cast into the select if the213 // select does not have a compare instruction with matching operand types214 // or the select is likely better done in a narrow type.215 // Creating a select with operands that are different sizes than its216 // condition may inhibit other folds and lead to worse codegen.217 auto *Cmp = dyn_cast<CmpInst>(Sel->getCondition());218 if (!Cmp || Cmp->getOperand(0)->getType() != Sel->getType() ||219 (CI.getOpcode() == Instruction::Trunc &&220 shouldChangeType(CI.getSrcTy(), CI.getType()))) {221 222 // If it's a bitcast involving vectors, make sure it has the same number223 // of elements on both sides.224 if (CI.getOpcode() != Instruction::BitCast ||225 match(&CI, m_ElementWiseBitCast(m_Value()))) {226 if (Instruction *NV = FoldOpIntoSelect(CI, Sel)) {227 replaceAllDbgUsesWith(*Sel, *NV, CI, DT);228 return NV;229 }230 }231 }232 }233 234 // If we are casting a PHI, then fold the cast into the PHI.235 if (auto *PN = dyn_cast<PHINode>(Src)) {236 // Don't do this if it would create a PHI node with an illegal type from a237 // legal type.238 if (!Src->getType()->isIntegerTy() || !CI.getType()->isIntegerTy() ||239 shouldChangeType(CI.getSrcTy(), CI.getType()))240 if (Instruction *NV = foldOpIntoPhi(CI, PN))241 return NV;242 }243 244 // Canonicalize a unary shuffle after the cast if neither operation changes245 // the size or element size of the input vector.246 // TODO: We could allow size-changing ops if that doesn't harm codegen.247 // cast (shuffle X, Mask) --> shuffle (cast X), Mask248 Value *X;249 ArrayRef<int> Mask;250 if (match(Src, m_OneUse(m_Shuffle(m_Value(X), m_Undef(), m_Mask(Mask))))) {251 // TODO: Allow scalable vectors?252 auto *SrcTy = dyn_cast<FixedVectorType>(X->getType());253 auto *DestTy = dyn_cast<FixedVectorType>(Ty);254 if (SrcTy && DestTy &&255 SrcTy->getNumElements() == DestTy->getNumElements() &&256 SrcTy->getPrimitiveSizeInBits() == DestTy->getPrimitiveSizeInBits()) {257 Value *CastX = Builder.CreateCast(CI.getOpcode(), X, DestTy);258 return new ShuffleVectorInst(CastX, Mask);259 }260 }261 262 return nullptr;263}264 265namespace {266 267/// Helper class for evaluating whether a value can be computed in a different268/// type without changing its value. Used by cast simplification transforms.269class TypeEvaluationHelper {270public:271 /// Return true if we can evaluate the specified expression tree as type Ty272 /// instead of its larger type, and arrive with the same value.273 /// This is used by code that tries to eliminate truncates.274 [[nodiscard]] static bool canEvaluateTruncated(Value *V, Type *Ty,275 InstCombinerImpl &IC,276 Instruction *CxtI);277 278 /// Determine if the specified value can be computed in the specified wider279 /// type and produce the same low bits. If not, return false.280 [[nodiscard]] static bool canEvaluateZExtd(Value *V, Type *Ty,281 unsigned &BitsToClear,282 InstCombinerImpl &IC,283 Instruction *CxtI);284 285 /// Return true if we can take the specified value and return it as type Ty286 /// without inserting any new casts and without changing the value of the287 /// common low bits.288 [[nodiscard]] static bool canEvaluateSExtd(Value *V, Type *Ty);289 290private:291 /// Constants and extensions/truncates from the destination type are always292 /// free to be evaluated in that type.293 [[nodiscard]] static bool canAlwaysEvaluateInType(Value *V, Type *Ty);294 295 /// Check if we traversed all the users of the multi-use values we've seen.296 [[nodiscard]] bool allPendingVisited() const {297 return llvm::all_of(Pending,298 [this](Value *V) { return Visited.contains(V); });299 }300 301 /// A generic wrapper for canEvaluate* recursions to inject visitation302 /// tracking and enforce correct multi-use value evaluations.303 [[nodiscard]] bool304 canEvaluate(Value *V, Type *Ty,305 llvm::function_ref<bool(Value *, Type *Type)> Pred) {306 if (canAlwaysEvaluateInType(V, Ty))307 return true;308 309 auto *I = dyn_cast<Instruction>(V);310 311 if (I == nullptr)312 return false;313 314 // We insert false by default to return false when we encounter user loops.315 const auto [It, Inserted] = Visited.insert({V, false});316 317 // There are three possible cases for us having information on this value318 // in the Visited map:319 // 1. We properly checked it and concluded that we can evaluate it (true)320 // 2. We properly checked it and concluded that we can't (false)321 // 3. We started to check it, but during the recursive traversal we came322 // back to it.323 //324 // For cases 1 and 2, we can safely return the stored result. For case 3, we325 // can potentially have a situation where we can evaluate recursive user326 // chains, but that can be quite tricky to do properly and isntead, we327 // return false.328 //329 // In any case, we should return whatever was there in the map to begin330 // with.331 if (!Inserted)332 return It->getSecond();333 334 // We can easily make a decision about single-user values whether they can335 // be evaluated in a different type or not, we came from that user. This is336 // not as simple for multi-user values.337 //338 // In general, we have the following case (inverted control-flow, users are339 // at the top):340 //341 // Cast %A342 // ____|343 // /344 // %A = Use %B, %C345 // ________| |346 // / |347 // %B = Use %D |348 // ________| |349 // / |350 // %D = Use %C |351 // ________|___|352 // /353 // %C = ...354 //355 // In this case, when we check %A, %B and %D, we are confident that we can356 // make the decision here and now, since we came from their only users.357 //358 // For %C, it is harder. We come there twice, and when we come the first359 // time, it's hard to tell if we will visit the second user (technically360 // it's not hard, but we might need a lot of repetitive checks with non-zero361 // cost).362 //363 // In the case above, we are allowed to evaluate %C in different type364 // because all of it users were part of the traversal.365 //366 // In the following case, however, we can't make this conclusion:367 //368 // Cast %A369 // ____|370 // /371 // %A = Use %B, %C372 // ________| |373 // / |374 // %B = Use %D |375 // ________| |376 // / |377 // %D = Use %C |378 // | |379 // foo(%C) | | <- never traversing foo(%C)380 // ________|___|381 // /382 // %C = ...383 //384 // In this case, we still can evaluate %C in a different type, but we'd need385 // to create a copy of the original %C to be used in foo(%C). Such386 // duplication might be not profitable.387 //388 // For this reason, we collect all users of the mult-user values and mark389 // them as "pending" and defer this decision to the very end. When we are390 // done and and ready to have a positive verdict, we should double-check all391 // of the pending users and ensure that we visited them. allPendingVisited392 // predicate checks exactly that.393 if (!I->hasOneUse())394 llvm::append_range(Pending, I->users());395 396 const bool Result = Pred(V, Ty);397 // We have to set result this way and not via It because Pred is recursive398 // and it is very likely that we grew Visited and invalidated It.399 Visited[V] = Result;400 return Result;401 }402 403 /// Filter out values that we can not evaluate in the destination type for404 /// free.405 [[nodiscard]] bool canNotEvaluateInType(Value *V, Type *Ty);406 407 [[nodiscard]] bool canEvaluateTruncatedImpl(Value *V, Type *Ty,408 InstCombinerImpl &IC,409 Instruction *CxtI);410 [[nodiscard]] bool canEvaluateTruncatedPred(Value *V, Type *Ty,411 InstCombinerImpl &IC,412 Instruction *CxtI);413 [[nodiscard]] bool canEvaluateZExtdImpl(Value *V, Type *Ty,414 unsigned &BitsToClear,415 InstCombinerImpl &IC,416 Instruction *CxtI);417 [[nodiscard]] bool canEvaluateSExtdImpl(Value *V, Type *Ty);418 [[nodiscard]] bool canEvaluateSExtdPred(Value *V, Type *Ty);419 420 /// A bookkeeping map to memorize an already made decision for a traversed421 /// value.422 SmallDenseMap<Value *, bool, 8> Visited;423 424 /// A list of pending values to check in the end.425 SmallVector<Value *, 8> Pending;426};427 428} // anonymous namespace429 430/// Constants and extensions/truncates from the destination type are always431/// free to be evaluated in that type. This is a helper for canEvaluate*.432bool TypeEvaluationHelper::canAlwaysEvaluateInType(Value *V, Type *Ty) {433 if (isa<Constant>(V))434 return match(V, m_ImmConstant());435 436 Value *X;437 if ((match(V, m_ZExtOrSExt(m_Value(X))) || match(V, m_Trunc(m_Value(X)))) &&438 X->getType() == Ty)439 return true;440 441 return false;442}443 444/// Filter out values that we can not evaluate in the destination type for free.445/// This is a helper for canEvaluate*.446bool TypeEvaluationHelper::canNotEvaluateInType(Value *V, Type *Ty) {447 if (!isa<Instruction>(V))448 return true;449 // We don't extend or shrink something that has multiple uses -- doing so450 // would require duplicating the instruction which isn't profitable.451 if (!V->hasOneUse())452 return true;453 454 return false;455}456 457/// Return true if we can evaluate the specified expression tree as type Ty458/// instead of its larger type, and arrive with the same value.459/// This is used by code that tries to eliminate truncates.460///461/// Ty will always be a type smaller than V. We should return true if trunc(V)462/// can be computed by computing V in the smaller type. If V is an instruction,463/// then trunc(inst(x,y)) can be computed as inst(trunc(x),trunc(y)), which only464/// makes sense if x and y can be efficiently truncated.465///466/// This function works on both vectors and scalars.467///468bool TypeEvaluationHelper::canEvaluateTruncated(Value *V, Type *Ty,469 InstCombinerImpl &IC,470 Instruction *CxtI) {471 TypeEvaluationHelper TYH;472 return TYH.canEvaluateTruncatedImpl(V, Ty, IC, CxtI) &&473 // We need to check whether we visited all users of multi-user values,474 // and we have to do it at the very end, outside of the recursion.475 TYH.allPendingVisited();476}477 478bool TypeEvaluationHelper::canEvaluateTruncatedImpl(Value *V, Type *Ty,479 InstCombinerImpl &IC,480 Instruction *CxtI) {481 return canEvaluate(V, Ty, [this, &IC, CxtI](Value *V, Type *Ty) {482 return canEvaluateTruncatedPred(V, Ty, IC, CxtI);483 });484}485 486bool TypeEvaluationHelper::canEvaluateTruncatedPred(Value *V, Type *Ty,487 InstCombinerImpl &IC,488 Instruction *CxtI) {489 auto *I = cast<Instruction>(V);490 Type *OrigTy = V->getType();491 switch (I->getOpcode()) {492 case Instruction::Add:493 case Instruction::Sub:494 case Instruction::Mul:495 case Instruction::And:496 case Instruction::Or:497 case Instruction::Xor:498 // These operators can all arbitrarily be extended or truncated.499 return canEvaluateTruncatedImpl(I->getOperand(0), Ty, IC, CxtI) &&500 canEvaluateTruncatedImpl(I->getOperand(1), Ty, IC, CxtI);501 502 case Instruction::UDiv:503 case Instruction::URem: {504 // UDiv and URem can be truncated if all the truncated bits are zero.505 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();506 uint32_t BitWidth = Ty->getScalarSizeInBits();507 assert(BitWidth < OrigBitWidth && "Unexpected bitwidths!");508 APInt Mask = APInt::getBitsSetFrom(OrigBitWidth, BitWidth);509 // Do not preserve the original context instruction. Simplifying div/rem510 // based on later context may introduce a trap.511 if (IC.MaskedValueIsZero(I->getOperand(0), Mask, I) &&512 IC.MaskedValueIsZero(I->getOperand(1), Mask, I)) {513 return canEvaluateTruncatedImpl(I->getOperand(0), Ty, IC, CxtI) &&514 canEvaluateTruncatedImpl(I->getOperand(1), Ty, IC, CxtI);515 }516 break;517 }518 case Instruction::Shl: {519 // If we are truncating the result of this SHL, and if it's a shift of an520 // inrange amount, we can always perform a SHL in a smaller type.521 uint32_t BitWidth = Ty->getScalarSizeInBits();522 KnownBits AmtKnownBits =523 llvm::computeKnownBits(I->getOperand(1), IC.getDataLayout());524 if (AmtKnownBits.getMaxValue().ult(BitWidth))525 return canEvaluateTruncatedImpl(I->getOperand(0), Ty, IC, CxtI) &&526 canEvaluateTruncatedImpl(I->getOperand(1), Ty, IC, CxtI);527 break;528 }529 case Instruction::LShr: {530 // If this is a truncate of a logical shr, we can truncate it to a smaller531 // lshr iff we know that the bits we would otherwise be shifting in are532 // already zeros.533 // TODO: It is enough to check that the bits we would be shifting in are534 // zero - use AmtKnownBits.getMaxValue().535 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();536 uint32_t BitWidth = Ty->getScalarSizeInBits();537 KnownBits AmtKnownBits = IC.computeKnownBits(I->getOperand(1), CxtI);538 APInt MaxShiftAmt = AmtKnownBits.getMaxValue();539 APInt ShiftedBits = APInt::getBitsSetFrom(OrigBitWidth, BitWidth);540 if (MaxShiftAmt.ult(BitWidth)) {541 // If the only user is a trunc then we can narrow the shift if any new542 // MSBs are not going to be used.543 if (auto *Trunc = dyn_cast<TruncInst>(V->user_back())) {544 auto DemandedBits = Trunc->getType()->getScalarSizeInBits();545 if ((MaxShiftAmt + DemandedBits).ule(BitWidth))546 return canEvaluateTruncatedImpl(I->getOperand(0), Ty, IC, CxtI) &&547 canEvaluateTruncatedImpl(I->getOperand(1), Ty, IC, CxtI);548 }549 if (IC.MaskedValueIsZero(I->getOperand(0), ShiftedBits, CxtI))550 return canEvaluateTruncatedImpl(I->getOperand(0), Ty, IC, CxtI) &&551 canEvaluateTruncatedImpl(I->getOperand(1), Ty, IC, CxtI);552 }553 break;554 }555 case Instruction::AShr: {556 // If this is a truncate of an arithmetic shr, we can truncate it to a557 // smaller ashr iff we know that all the bits from the sign bit of the558 // original type and the sign bit of the truncate type are similar.559 // TODO: It is enough to check that the bits we would be shifting in are560 // similar to sign bit of the truncate type.561 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();562 uint32_t BitWidth = Ty->getScalarSizeInBits();563 KnownBits AmtKnownBits =564 llvm::computeKnownBits(I->getOperand(1), IC.getDataLayout());565 unsigned ShiftedBits = OrigBitWidth - BitWidth;566 if (AmtKnownBits.getMaxValue().ult(BitWidth) &&567 ShiftedBits < IC.ComputeNumSignBits(I->getOperand(0), CxtI))568 return canEvaluateTruncatedImpl(I->getOperand(0), Ty, IC, CxtI) &&569 canEvaluateTruncatedImpl(I->getOperand(1), Ty, IC, CxtI);570 break;571 }572 case Instruction::Trunc:573 // trunc(trunc(x)) -> trunc(x)574 return true;575 case Instruction::ZExt:576 case Instruction::SExt:577 // trunc(ext(x)) -> ext(x) if the source type is smaller than the new dest578 // trunc(ext(x)) -> trunc(x) if the source type is larger than the new dest579 return true;580 case Instruction::Select: {581 SelectInst *SI = cast<SelectInst>(I);582 return canEvaluateTruncatedImpl(SI->getTrueValue(), Ty, IC, CxtI) &&583 canEvaluateTruncatedImpl(SI->getFalseValue(), Ty, IC, CxtI);584 }585 case Instruction::PHI: {586 // We can change a phi if we can change all operands. Note that we never587 // get into trouble with cyclic PHIs here because canEvaluate handles use588 // chain loops.589 PHINode *PN = cast<PHINode>(I);590 return llvm::all_of(591 PN->incoming_values(), [this, Ty, &IC, CxtI](Value *IncValue) {592 return canEvaluateTruncatedImpl(IncValue, Ty, IC, CxtI);593 });594 }595 case Instruction::FPToUI:596 case Instruction::FPToSI: {597 // If the integer type can hold the max FP value, it is safe to cast598 // directly to that type. Otherwise, we may create poison via overflow599 // that did not exist in the original code.600 Type *InputTy = I->getOperand(0)->getType()->getScalarType();601 const fltSemantics &Semantics = InputTy->getFltSemantics();602 uint32_t MinBitWidth = APFloatBase::semanticsIntSizeInBits(603 Semantics, I->getOpcode() == Instruction::FPToSI);604 return Ty->getScalarSizeInBits() >= MinBitWidth;605 }606 case Instruction::ShuffleVector:607 return canEvaluateTruncatedImpl(I->getOperand(0), Ty, IC, CxtI) &&608 canEvaluateTruncatedImpl(I->getOperand(1), Ty, IC, CxtI);609 610 default:611 // TODO: Can handle more cases here.612 break;613 }614 615 return false;616}617 618/// Given a vector that is bitcast to an integer, optionally logically619/// right-shifted, and truncated, convert it to an extractelement.620/// Example (big endian):621/// trunc (lshr (bitcast <4 x i32> %X to i128), 32) to i32622/// --->623/// extractelement <4 x i32> %X, 1624static Instruction *foldVecTruncToExtElt(TruncInst &Trunc,625 InstCombinerImpl &IC) {626 Value *TruncOp = Trunc.getOperand(0);627 Type *DestType = Trunc.getType();628 if (!TruncOp->hasOneUse() || !isa<IntegerType>(DestType))629 return nullptr;630 631 Value *VecInput = nullptr;632 ConstantInt *ShiftVal = nullptr;633 if (!match(TruncOp, m_CombineOr(m_BitCast(m_Value(VecInput)),634 m_LShr(m_BitCast(m_Value(VecInput)),635 m_ConstantInt(ShiftVal)))) ||636 !isa<VectorType>(VecInput->getType()))637 return nullptr;638 639 VectorType *VecType = cast<VectorType>(VecInput->getType());640 unsigned VecWidth = VecType->getPrimitiveSizeInBits();641 unsigned DestWidth = DestType->getPrimitiveSizeInBits();642 unsigned ShiftAmount = ShiftVal ? ShiftVal->getZExtValue() : 0;643 644 if ((VecWidth % DestWidth != 0) || (ShiftAmount % DestWidth != 0))645 return nullptr;646 647 // If the element type of the vector doesn't match the result type,648 // bitcast it to a vector type that we can extract from.649 unsigned NumVecElts = VecWidth / DestWidth;650 if (VecType->getElementType() != DestType) {651 VecType = FixedVectorType::get(DestType, NumVecElts);652 VecInput = IC.Builder.CreateBitCast(VecInput, VecType, "bc");653 }654 655 unsigned Elt = ShiftAmount / DestWidth;656 if (IC.getDataLayout().isBigEndian())657 Elt = NumVecElts - 1 - Elt;658 659 return ExtractElementInst::Create(VecInput, IC.Builder.getInt32(Elt));660}661 662/// Whenever an element is extracted from a vector, optionally shifted down, and663/// then truncated, canonicalize by converting it to a bitcast followed by an664/// extractelement.665///666/// Examples (little endian):667/// trunc (extractelement <4 x i64> %X, 0) to i32668/// --->669/// extractelement <8 x i32> (bitcast <4 x i64> %X to <8 x i32>), i32 0670///671/// trunc (lshr (extractelement <4 x i32> %X, 0), 8) to i8672/// --->673/// extractelement <16 x i8> (bitcast <4 x i32> %X to <16 x i8>), i32 1674static Instruction *foldVecExtTruncToExtElt(TruncInst &Trunc,675 InstCombinerImpl &IC) {676 Value *Src = Trunc.getOperand(0);677 Type *SrcType = Src->getType();678 Type *DstType = Trunc.getType();679 680 // Only attempt this if we have simple aliasing of the vector elements.681 // A badly fit destination size would result in an invalid cast.682 unsigned SrcBits = SrcType->getScalarSizeInBits();683 unsigned DstBits = DstType->getScalarSizeInBits();684 unsigned TruncRatio = SrcBits / DstBits;685 if ((SrcBits % DstBits) != 0)686 return nullptr;687 688 Value *VecOp;689 ConstantInt *Cst;690 const APInt *ShiftAmount = nullptr;691 if (!match(Src, m_OneUse(m_ExtractElt(m_Value(VecOp), m_ConstantInt(Cst)))) &&692 !match(Src,693 m_OneUse(m_LShr(m_ExtractElt(m_Value(VecOp), m_ConstantInt(Cst)),694 m_APInt(ShiftAmount)))))695 return nullptr;696 697 auto *VecOpTy = cast<VectorType>(VecOp->getType());698 auto VecElts = VecOpTy->getElementCount();699 700 uint64_t BitCastNumElts = VecElts.getKnownMinValue() * TruncRatio;701 uint64_t VecOpIdx = Cst->getZExtValue();702 uint64_t NewIdx = IC.getDataLayout().isBigEndian()703 ? (VecOpIdx + 1) * TruncRatio - 1704 : VecOpIdx * TruncRatio;705 706 // Adjust index by the whole number of truncated elements.707 if (ShiftAmount) {708 // Check shift amount is in range and shifts a whole number of truncated709 // elements.710 if (ShiftAmount->uge(SrcBits) || ShiftAmount->urem(DstBits) != 0)711 return nullptr;712 713 uint64_t IdxOfs = ShiftAmount->udiv(DstBits).getZExtValue();714 NewIdx = IC.getDataLayout().isBigEndian() ? (NewIdx - IdxOfs)715 : (NewIdx + IdxOfs);716 }717 718 assert(BitCastNumElts <= std::numeric_limits<uint32_t>::max() &&719 NewIdx <= std::numeric_limits<uint32_t>::max() && "overflow 32-bits");720 721 auto *BitCastTo =722 VectorType::get(DstType, BitCastNumElts, VecElts.isScalable());723 Value *BitCast = IC.Builder.CreateBitCast(VecOp, BitCastTo);724 return ExtractElementInst::Create(BitCast, IC.Builder.getInt32(NewIdx));725}726 727/// Funnel/Rotate left/right may occur in a wider type than necessary because of728/// type promotion rules. Try to narrow the inputs and convert to funnel shift.729Instruction *InstCombinerImpl::narrowFunnelShift(TruncInst &Trunc) {730 assert((isa<VectorType>(Trunc.getSrcTy()) ||731 shouldChangeType(Trunc.getSrcTy(), Trunc.getType())) &&732 "Don't narrow to an illegal scalar type");733 734 // Bail out on strange types. It is possible to handle some of these patterns735 // even with non-power-of-2 sizes, but it is not a likely scenario.736 Type *DestTy = Trunc.getType();737 unsigned NarrowWidth = DestTy->getScalarSizeInBits();738 unsigned WideWidth = Trunc.getSrcTy()->getScalarSizeInBits();739 if (!isPowerOf2_32(NarrowWidth))740 return nullptr;741 742 // First, find an or'd pair of opposite shifts:743 // trunc (or (lshr ShVal0, ShAmt0), (shl ShVal1, ShAmt1))744 BinaryOperator *Or0, *Or1;745 if (!match(Trunc.getOperand(0), m_OneUse(m_Or(m_BinOp(Or0), m_BinOp(Or1)))))746 return nullptr;747 748 Value *ShVal0, *ShVal1, *ShAmt0, *ShAmt1;749 if (!match(Or0, m_OneUse(m_LogicalShift(m_Value(ShVal0), m_Value(ShAmt0)))) ||750 !match(Or1, m_OneUse(m_LogicalShift(m_Value(ShVal1), m_Value(ShAmt1)))) ||751 Or0->getOpcode() == Or1->getOpcode())752 return nullptr;753 754 // Canonicalize to or(shl(ShVal0, ShAmt0), lshr(ShVal1, ShAmt1)).755 if (Or0->getOpcode() == BinaryOperator::LShr) {756 std::swap(Or0, Or1);757 std::swap(ShVal0, ShVal1);758 std::swap(ShAmt0, ShAmt1);759 }760 assert(Or0->getOpcode() == BinaryOperator::Shl &&761 Or1->getOpcode() == BinaryOperator::LShr &&762 "Illegal or(shift,shift) pair");763 764 // Match the shift amount operands for a funnel/rotate pattern. This always765 // matches a subtraction on the R operand.766 auto matchShiftAmount = [&](Value *L, Value *R, unsigned Width) -> Value * {767 // The shift amounts may add up to the narrow bit width:768 // (shl ShVal0, L) | (lshr ShVal1, Width - L)769 // If this is a funnel shift (different operands are shifted), then the770 // shift amount can not over-shift (create poison) in the narrow type.771 unsigned MaxShiftAmountWidth = Log2_32(NarrowWidth);772 APInt HiBitMask = ~APInt::getLowBitsSet(WideWidth, MaxShiftAmountWidth);773 if (ShVal0 == ShVal1 || MaskedValueIsZero(L, HiBitMask))774 if (match(R, m_OneUse(m_Sub(m_SpecificInt(Width), m_Specific(L)))))775 return L;776 777 // The following patterns currently only work for rotation patterns.778 // TODO: Add more general funnel-shift compatible patterns.779 if (ShVal0 != ShVal1)780 return nullptr;781 782 // The shift amount may be masked with negation:783 // (shl ShVal0, (X & (Width - 1))) | (lshr ShVal1, ((-X) & (Width - 1)))784 Value *X;785 unsigned Mask = Width - 1;786 if (match(L, m_And(m_Value(X), m_SpecificInt(Mask))) &&787 match(R, m_And(m_Neg(m_Specific(X)), m_SpecificInt(Mask))))788 return X;789 790 // Same as above, but the shift amount may be extended after masking:791 if (match(L, m_ZExt(m_And(m_Value(X), m_SpecificInt(Mask)))) &&792 match(R, m_ZExt(m_And(m_Neg(m_Specific(X)), m_SpecificInt(Mask)))))793 return X;794 795 return nullptr;796 };797 798 Value *ShAmt = matchShiftAmount(ShAmt0, ShAmt1, NarrowWidth);799 bool IsFshl = true; // Sub on LSHR.800 if (!ShAmt) {801 ShAmt = matchShiftAmount(ShAmt1, ShAmt0, NarrowWidth);802 IsFshl = false; // Sub on SHL.803 }804 if (!ShAmt)805 return nullptr;806 807 // The right-shifted value must have high zeros in the wide type (for example808 // from 'zext', 'and' or 'shift'). High bits of the left-shifted value are809 // truncated, so those do not matter.810 APInt HiBitMask = APInt::getHighBitsSet(WideWidth, WideWidth - NarrowWidth);811 if (!MaskedValueIsZero(ShVal1, HiBitMask, &Trunc))812 return nullptr;813 814 // Adjust the width of ShAmt for narrowed funnel shift operation:815 // - Zero-extend if ShAmt is narrower than the destination type.816 // - Truncate if ShAmt is wider, discarding non-significant high-order bits.817 // This prepares ShAmt for llvm.fshl.i8(trunc(ShVal), trunc(ShVal),818 // zext/trunc(ShAmt)).819 Value *NarrowShAmt = Builder.CreateZExtOrTrunc(ShAmt, DestTy);820 821 Value *X, *Y;822 X = Y = Builder.CreateTrunc(ShVal0, DestTy);823 if (ShVal0 != ShVal1)824 Y = Builder.CreateTrunc(ShVal1, DestTy);825 Intrinsic::ID IID = IsFshl ? Intrinsic::fshl : Intrinsic::fshr;826 Function *F =827 Intrinsic::getOrInsertDeclaration(Trunc.getModule(), IID, DestTy);828 return CallInst::Create(F, {X, Y, NarrowShAmt});829}830 831/// Try to narrow the width of math or bitwise logic instructions by pulling a832/// truncate ahead of binary operators.833Instruction *InstCombinerImpl::narrowBinOp(TruncInst &Trunc) {834 Type *SrcTy = Trunc.getSrcTy();835 Type *DestTy = Trunc.getType();836 unsigned SrcWidth = SrcTy->getScalarSizeInBits();837 unsigned DestWidth = DestTy->getScalarSizeInBits();838 839 if (!isa<VectorType>(SrcTy) && !shouldChangeType(SrcTy, DestTy))840 return nullptr;841 842 BinaryOperator *BinOp;843 if (!match(Trunc.getOperand(0), m_OneUse(m_BinOp(BinOp))))844 return nullptr;845 846 Value *BinOp0 = BinOp->getOperand(0);847 Value *BinOp1 = BinOp->getOperand(1);848 switch (BinOp->getOpcode()) {849 case Instruction::And:850 case Instruction::Or:851 case Instruction::Xor:852 case Instruction::Add:853 case Instruction::Sub:854 case Instruction::Mul: {855 Constant *C;856 if (match(BinOp0, m_Constant(C))) {857 // trunc (binop C, X) --> binop (trunc C', X)858 Constant *NarrowC = ConstantExpr::getTrunc(C, DestTy);859 Value *TruncX = Builder.CreateTrunc(BinOp1, DestTy);860 return BinaryOperator::Create(BinOp->getOpcode(), NarrowC, TruncX);861 }862 if (match(BinOp1, m_Constant(C))) {863 // trunc (binop X, C) --> binop (trunc X, C')864 Constant *NarrowC = ConstantExpr::getTrunc(C, DestTy);865 Value *TruncX = Builder.CreateTrunc(BinOp0, DestTy);866 return BinaryOperator::Create(BinOp->getOpcode(), TruncX, NarrowC);867 }868 Value *X;869 if (match(BinOp0, m_ZExtOrSExt(m_Value(X))) && X->getType() == DestTy) {870 // trunc (binop (ext X), Y) --> binop X, (trunc Y)871 Value *NarrowOp1 = Builder.CreateTrunc(BinOp1, DestTy);872 return BinaryOperator::Create(BinOp->getOpcode(), X, NarrowOp1);873 }874 if (match(BinOp1, m_ZExtOrSExt(m_Value(X))) && X->getType() == DestTy) {875 // trunc (binop Y, (ext X)) --> binop (trunc Y), X876 Value *NarrowOp0 = Builder.CreateTrunc(BinOp0, DestTy);877 return BinaryOperator::Create(BinOp->getOpcode(), NarrowOp0, X);878 }879 break;880 }881 case Instruction::LShr:882 case Instruction::AShr: {883 // trunc (*shr (trunc A), C) --> trunc(*shr A, C)884 Value *A;885 Constant *C;886 if (match(BinOp0, m_Trunc(m_Value(A))) && match(BinOp1, m_Constant(C))) {887 unsigned MaxShiftAmt = SrcWidth - DestWidth;888 // If the shift is small enough, all zero/sign bits created by the shift889 // are removed by the trunc.890 if (match(C, m_SpecificInt_ICMP(ICmpInst::ICMP_ULE,891 APInt(SrcWidth, MaxShiftAmt)))) {892 auto *OldShift = cast<Instruction>(Trunc.getOperand(0));893 bool IsExact = OldShift->isExact();894 if (Constant *ShAmt = ConstantFoldIntegerCast(C, A->getType(),895 /*IsSigned*/ true, DL)) {896 ShAmt = Constant::mergeUndefsWith(ShAmt, C);897 Value *Shift =898 OldShift->getOpcode() == Instruction::AShr899 ? Builder.CreateAShr(A, ShAmt, OldShift->getName(), IsExact)900 : Builder.CreateLShr(A, ShAmt, OldShift->getName(), IsExact);901 return CastInst::CreateTruncOrBitCast(Shift, DestTy);902 }903 }904 }905 break;906 }907 default: break;908 }909 910 if (Instruction *NarrowOr = narrowFunnelShift(Trunc))911 return NarrowOr;912 913 return nullptr;914}915 916/// Try to narrow the width of a splat shuffle. This could be generalized to any917/// shuffle with a constant operand, but we limit the transform to avoid918/// creating a shuffle type that targets may not be able to lower effectively.919static Instruction *shrinkSplatShuffle(TruncInst &Trunc,920 InstCombiner::BuilderTy &Builder) {921 auto *Shuf = dyn_cast<ShuffleVectorInst>(Trunc.getOperand(0));922 if (Shuf && Shuf->hasOneUse() && match(Shuf->getOperand(1), m_Undef()) &&923 all_equal(Shuf->getShuffleMask()) &&924 ElementCount::isKnownGE(Shuf->getType()->getElementCount(),925 cast<VectorType>(Shuf->getOperand(0)->getType())926 ->getElementCount())) {927 // trunc (shuf X, Undef, SplatMask) --> shuf (trunc X), Poison, SplatMask928 // trunc (shuf X, Poison, SplatMask) --> shuf (trunc X), Poison, SplatMask929 Type *NewTruncTy = Shuf->getOperand(0)->getType()->getWithNewType(930 Trunc.getType()->getScalarType());931 Value *NarrowOp = Builder.CreateTrunc(Shuf->getOperand(0), NewTruncTy);932 return new ShuffleVectorInst(NarrowOp, Shuf->getShuffleMask());933 }934 935 return nullptr;936}937 938/// Try to narrow the width of an insert element. This could be generalized for939/// any vector constant, but we limit the transform to insertion into undef to940/// avoid potential backend problems from unsupported insertion widths. This941/// could also be extended to handle the case of inserting a scalar constant942/// into a vector variable.943static Instruction *shrinkInsertElt(CastInst &Trunc,944 InstCombiner::BuilderTy &Builder) {945 Instruction::CastOps Opcode = Trunc.getOpcode();946 assert((Opcode == Instruction::Trunc || Opcode == Instruction::FPTrunc) &&947 "Unexpected instruction for shrinking");948 949 auto *InsElt = dyn_cast<InsertElementInst>(Trunc.getOperand(0));950 if (!InsElt || !InsElt->hasOneUse())951 return nullptr;952 953 Type *DestTy = Trunc.getType();954 Type *DestScalarTy = DestTy->getScalarType();955 Value *VecOp = InsElt->getOperand(0);956 Value *ScalarOp = InsElt->getOperand(1);957 Value *Index = InsElt->getOperand(2);958 959 if (match(VecOp, m_Undef())) {960 // trunc (inselt undef, X, Index) --> inselt undef, (trunc X), Index961 // fptrunc (inselt undef, X, Index) --> inselt undef, (fptrunc X), Index962 UndefValue *NarrowUndef = UndefValue::get(DestTy);963 Value *NarrowOp = Builder.CreateCast(Opcode, ScalarOp, DestScalarTy);964 return InsertElementInst::Create(NarrowUndef, NarrowOp, Index);965 }966 967 return nullptr;968}969 970Instruction *InstCombinerImpl::visitTrunc(TruncInst &Trunc) {971 if (Instruction *Result = commonCastTransforms(Trunc))972 return Result;973 974 Value *Src = Trunc.getOperand(0);975 Type *DestTy = Trunc.getType(), *SrcTy = Src->getType();976 unsigned DestWidth = DestTy->getScalarSizeInBits();977 unsigned SrcWidth = SrcTy->getScalarSizeInBits();978 979 // Attempt to truncate the entire input expression tree to the destination980 // type. Only do this if the dest type is a simple type, don't convert the981 // expression tree to something weird like i93 unless the source is also982 // strange.983 if ((DestTy->isVectorTy() || shouldChangeType(SrcTy, DestTy)) &&984 TypeEvaluationHelper::canEvaluateTruncated(Src, DestTy, *this, &Trunc)) {985 986 // If this cast is a truncate, evaluting in a different type always987 // eliminates the cast, so it is always a win.988 LLVM_DEBUG(989 dbgs() << "ICE: EvaluateInDifferentType converting expression type"990 " to avoid cast: "991 << Trunc << '\n');992 Value *Res = EvaluateInDifferentType(Src, DestTy, false);993 assert(Res->getType() == DestTy);994 return replaceInstUsesWith(Trunc, Res);995 }996 997 // For integer types, check if we can shorten the entire input expression to998 // DestWidth * 2, which won't allow removing the truncate, but reducing the999 // width may enable further optimizations, e.g. allowing for larger1000 // vectorization factors.1001 if (auto *DestITy = dyn_cast<IntegerType>(DestTy)) {1002 if (DestWidth * 2 < SrcWidth) {1003 auto *NewDestTy = DestITy->getExtendedType();1004 if (shouldChangeType(SrcTy, NewDestTy) &&1005 TypeEvaluationHelper::canEvaluateTruncated(Src, NewDestTy, *this,1006 &Trunc)) {1007 LLVM_DEBUG(1008 dbgs() << "ICE: EvaluateInDifferentType converting expression type"1009 " to reduce the width of operand of"1010 << Trunc << '\n');1011 Value *Res = EvaluateInDifferentType(Src, NewDestTy, false);1012 return new TruncInst(Res, DestTy);1013 }1014 }1015 }1016 1017 // See if we can simplify any instructions used by the input whose sole1018 // purpose is to compute bits we don't care about.1019 if (SimplifyDemandedInstructionBits(Trunc))1020 return &Trunc;1021 1022 if (DestWidth == 1) {1023 Value *Zero = Constant::getNullValue(SrcTy);1024 1025 Value *X;1026 const APInt *C1;1027 Constant *C2;1028 if (match(Src, m_OneUse(m_Shr(m_Shl(m_Power2(C1), m_Value(X)),1029 m_ImmConstant(C2))))) {1030 // trunc ((C1 << X) >> C2) to i1 --> X == (C2-cttz(C1)), where C1 is pow21031 Constant *Log2C1 = ConstantInt::get(SrcTy, C1->exactLogBase2());1032 Constant *CmpC = ConstantExpr::getSub(C2, Log2C1);1033 return new ICmpInst(ICmpInst::ICMP_EQ, X, CmpC);1034 }1035 1036 if (match(Src, m_Shr(m_Value(X), m_SpecificInt(SrcWidth - 1)))) {1037 // trunc (ashr X, BW-1) to i1 --> icmp slt X, 01038 // trunc (lshr X, BW-1) to i1 --> icmp slt X, 01039 return new ICmpInst(ICmpInst::ICMP_SLT, X, Zero);1040 }1041 1042 Constant *C;1043 if (match(Src, m_OneUse(m_LShr(m_Value(X), m_ImmConstant(C))))) {1044 // trunc (lshr X, C) to i1 --> icmp ne (and X, C'), 01045 Constant *One = ConstantInt::get(SrcTy, APInt(SrcWidth, 1));1046 Value *MaskC = Builder.CreateShl(One, C);1047 Value *And = Builder.CreateAnd(X, MaskC);1048 return new ICmpInst(ICmpInst::ICMP_NE, And, Zero);1049 }1050 if (match(Src, m_OneUse(m_c_Or(m_LShr(m_Value(X), m_ImmConstant(C)),1051 m_Deferred(X))))) {1052 // trunc (or (lshr X, C), X) to i1 --> icmp ne (and X, C'), 01053 Constant *One = ConstantInt::get(SrcTy, APInt(SrcWidth, 1));1054 Value *MaskC = Builder.CreateShl(One, C);1055 Value *And = Builder.CreateAnd(X, Builder.CreateOr(MaskC, One));1056 return new ICmpInst(ICmpInst::ICMP_NE, And, Zero);1057 }1058 1059 {1060 const APInt *C;1061 if (match(Src, m_Shl(m_APInt(C), m_Value(X))) && (*C)[0] == 1) {1062 // trunc (C << X) to i1 --> X == 0, where C is odd1063 return new ICmpInst(ICmpInst::Predicate::ICMP_EQ, X, Zero);1064 }1065 }1066 1067 if (Trunc.hasNoUnsignedWrap() || Trunc.hasNoSignedWrap()) {1068 Value *X, *Y;1069 if (match(Src, m_Xor(m_Value(X), m_Value(Y))))1070 return new ICmpInst(ICmpInst::ICMP_NE, X, Y);1071 }1072 }1073 1074 Value *A, *B;1075 Constant *C;1076 if (match(Src, m_LShr(m_SExt(m_Value(A)), m_Constant(C)))) {1077 unsigned AWidth = A->getType()->getScalarSizeInBits();1078 unsigned MaxShiftAmt = SrcWidth - std::max(DestWidth, AWidth);1079 auto *OldSh = cast<Instruction>(Src);1080 bool IsExact = OldSh->isExact();1081 1082 // If the shift is small enough, all zero bits created by the shift are1083 // removed by the trunc.1084 if (match(C, m_SpecificInt_ICMP(ICmpInst::ICMP_ULE,1085 APInt(SrcWidth, MaxShiftAmt)))) {1086 auto GetNewShAmt = [&](unsigned Width) {1087 Constant *MaxAmt = ConstantInt::get(SrcTy, Width - 1, false);1088 Constant *Cmp =1089 ConstantFoldCompareInstOperands(ICmpInst::ICMP_ULT, C, MaxAmt, DL);1090 Constant *ShAmt = ConstantFoldSelectInstruction(Cmp, C, MaxAmt);1091 return ConstantFoldCastOperand(Instruction::Trunc, ShAmt, A->getType(),1092 DL);1093 };1094 1095 // trunc (lshr (sext A), C) --> ashr A, C1096 if (A->getType() == DestTy) {1097 Constant *ShAmt = GetNewShAmt(DestWidth);1098 ShAmt = Constant::mergeUndefsWith(ShAmt, C);1099 return IsExact ? BinaryOperator::CreateExactAShr(A, ShAmt)1100 : BinaryOperator::CreateAShr(A, ShAmt);1101 }1102 // The types are mismatched, so create a cast after shifting:1103 // trunc (lshr (sext A), C) --> sext/trunc (ashr A, C)1104 if (Src->hasOneUse()) {1105 Constant *ShAmt = GetNewShAmt(AWidth);1106 Value *Shift = Builder.CreateAShr(A, ShAmt, "", IsExact);1107 return CastInst::CreateIntegerCast(Shift, DestTy, true);1108 }1109 }1110 // TODO: Mask high bits with 'and'.1111 }1112 1113 if (Instruction *I = narrowBinOp(Trunc))1114 return I;1115 1116 if (Instruction *I = shrinkSplatShuffle(Trunc, Builder))1117 return I;1118 1119 if (Instruction *I = shrinkInsertElt(Trunc, Builder))1120 return I;1121 1122 if (Src->hasOneUse() &&1123 (isa<VectorType>(SrcTy) || shouldChangeType(SrcTy, DestTy))) {1124 // Transform "trunc (shl X, cst)" -> "shl (trunc X), cst" so long as the1125 // dest type is native and cst < dest size.1126 if (match(Src, m_Shl(m_Value(A), m_Constant(C))) &&1127 !match(A, m_Shr(m_Value(), m_Constant()))) {1128 // Skip shifts of shift by constants. It undoes a combine in1129 // FoldShiftByConstant and is the extend in reg pattern.1130 APInt Threshold = APInt(C->getType()->getScalarSizeInBits(), DestWidth);1131 if (match(C, m_SpecificInt_ICMP(ICmpInst::ICMP_ULT, Threshold))) {1132 Value *NewTrunc = Builder.CreateTrunc(A, DestTy, A->getName() + ".tr");1133 return BinaryOperator::Create(Instruction::Shl, NewTrunc,1134 ConstantExpr::getTrunc(C, DestTy));1135 }1136 }1137 }1138 1139 if (Instruction *I = foldVecTruncToExtElt(Trunc, *this))1140 return I;1141 1142 if (Instruction *I = foldVecExtTruncToExtElt(Trunc, *this))1143 return I;1144 1145 // trunc (ctlz_i32(zext(A), B) --> add(ctlz_i16(A, B), C)1146 if (match(Src, m_OneUse(m_Intrinsic<Intrinsic::ctlz>(m_ZExt(m_Value(A)),1147 m_Value(B))))) {1148 unsigned AWidth = A->getType()->getScalarSizeInBits();1149 if (AWidth == DestWidth && AWidth > Log2_32(SrcWidth)) {1150 Value *WidthDiff = ConstantInt::get(A->getType(), SrcWidth - AWidth);1151 Value *NarrowCtlz =1152 Builder.CreateIntrinsic(Intrinsic::ctlz, {Trunc.getType()}, {A, B});1153 return BinaryOperator::CreateAdd(NarrowCtlz, WidthDiff);1154 }1155 }1156 1157 if (match(Src, m_VScale())) {1158 if (Trunc.getFunction() &&1159 Trunc.getFunction()->hasFnAttribute(Attribute::VScaleRange)) {1160 Attribute Attr =1161 Trunc.getFunction()->getFnAttribute(Attribute::VScaleRange);1162 if (std::optional<unsigned> MaxVScale = Attr.getVScaleRangeMax())1163 if (Log2_32(*MaxVScale) < DestWidth)1164 return replaceInstUsesWith(Trunc, Builder.CreateVScale(DestTy));1165 }1166 }1167 1168 if (DestWidth == 1 &&1169 (Trunc.hasNoUnsignedWrap() || Trunc.hasNoSignedWrap()) &&1170 isKnownNonZero(Src, SQ.getWithInstruction(&Trunc)))1171 return replaceInstUsesWith(Trunc, ConstantInt::getTrue(DestTy));1172 1173 bool Changed = false;1174 if (!Trunc.hasNoSignedWrap() &&1175 ComputeMaxSignificantBits(Src, &Trunc) <= DestWidth) {1176 Trunc.setHasNoSignedWrap(true);1177 Changed = true;1178 }1179 if (!Trunc.hasNoUnsignedWrap() &&1180 MaskedValueIsZero(Src, APInt::getBitsSetFrom(SrcWidth, DestWidth),1181 &Trunc)) {1182 Trunc.setHasNoUnsignedWrap(true);1183 Changed = true;1184 }1185 1186 const APInt *C1;1187 Value *V1;1188 // OP = { lshr, ashr }1189 // trunc ( OP i8 C1, V1) to i1 -> icmp eq V1, log_2(C1) iff C1 is power of 21190 if (DestWidth == 1 && match(Src, m_Shr(m_Power2(C1), m_Value(V1)))) {1191 Value *Right = ConstantInt::get(V1->getType(), C1->countr_zero());1192 return new ICmpInst(ICmpInst::ICMP_EQ, V1, Right);1193 }1194 1195 // OP = { lshr, ashr }1196 // trunc ( OP i8 C1, V1) to i1 -> icmp ult V1, log_2(C1 + 1) iff (C1 + 1) is1197 // power of 21198 if (DestWidth == 1 && match(Src, m_Shr(m_LowBitMask(C1), m_Value(V1)))) {1199 Value *Right = ConstantInt::get(V1->getType(), C1->countr_one());1200 return new ICmpInst(ICmpInst::ICMP_ULT, V1, Right);1201 }1202 1203 // OP = { lshr, ashr }1204 // trunc ( OP i8 C1, V1) to i1 -> icmp ugt V1, cttz(C1) - 1 iff (C1) is1205 // negative power of 21206 if (DestWidth == 1 && match(Src, m_Shr(m_NegatedPower2(C1), m_Value(V1)))) {1207 Value *Right = ConstantInt::get(V1->getType(), C1->countr_zero());1208 return new ICmpInst(ICmpInst::ICMP_UGE, V1, Right);1209 }1210 1211 return Changed ? &Trunc : nullptr;1212}1213 1214Instruction *InstCombinerImpl::transformZExtICmp(ICmpInst *Cmp,1215 ZExtInst &Zext) {1216 // If we are just checking for a icmp eq of a single bit and zext'ing it1217 // to an integer, then shift the bit to the appropriate place and then1218 // cast to integer to avoid the comparison.1219 1220 // FIXME: This set of transforms does not check for extra uses and/or creates1221 // an extra instruction (an optional final cast is not included1222 // in the transform comments). We may also want to favor icmp over1223 // shifts in cases of equal instructions because icmp has better1224 // analysis in general (invert the transform).1225 1226 const APInt *Op1CV;1227 if (match(Cmp->getOperand(1), m_APInt(Op1CV))) {1228 1229 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.1230 if (Cmp->getPredicate() == ICmpInst::ICMP_SLT && Op1CV->isZero()) {1231 Value *In = Cmp->getOperand(0);1232 Value *Sh = ConstantInt::get(In->getType(),1233 In->getType()->getScalarSizeInBits() - 1);1234 In = Builder.CreateLShr(In, Sh, In->getName() + ".lobit");1235 if (In->getType() != Zext.getType())1236 In = Builder.CreateIntCast(In, Zext.getType(), false /*ZExt*/);1237 1238 return replaceInstUsesWith(Zext, In);1239 }1240 1241 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.1242 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.1243 // zext (X != 0) to i32 --> X iff X has only the low bit set.1244 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.1245 1246 if (Op1CV->isZero() && Cmp->isEquality()) {1247 // Exactly 1 possible 1? But not the high-bit because that is1248 // canonicalized to this form.1249 KnownBits Known = computeKnownBits(Cmp->getOperand(0), &Zext);1250 APInt KnownZeroMask(~Known.Zero);1251 uint32_t ShAmt = KnownZeroMask.logBase2();1252 bool IsExpectShAmt = KnownZeroMask.isPowerOf2() &&1253 (Zext.getType()->getScalarSizeInBits() != ShAmt + 1);1254 if (IsExpectShAmt &&1255 (Cmp->getOperand(0)->getType() == Zext.getType() ||1256 Cmp->getPredicate() == ICmpInst::ICMP_NE || ShAmt == 0)) {1257 Value *In = Cmp->getOperand(0);1258 if (ShAmt) {1259 // Perform a logical shr by shiftamt.1260 // Insert the shift to put the result in the low bit.1261 In = Builder.CreateLShr(In, ConstantInt::get(In->getType(), ShAmt),1262 In->getName() + ".lobit");1263 }1264 1265 // Toggle the low bit for "X == 0".1266 if (Cmp->getPredicate() == ICmpInst::ICMP_EQ)1267 In = Builder.CreateXor(In, ConstantInt::get(In->getType(), 1));1268 1269 if (Zext.getType() == In->getType())1270 return replaceInstUsesWith(Zext, In);1271 1272 Value *IntCast = Builder.CreateIntCast(In, Zext.getType(), false);1273 return replaceInstUsesWith(Zext, IntCast);1274 }1275 }1276 }1277 1278 if (Cmp->isEquality()) {1279 // Test if a bit is clear/set using a shifted-one mask:1280 // zext (icmp eq (and X, (1 << ShAmt)), 0) --> and (lshr (not X), ShAmt), 11281 // zext (icmp ne (and X, (1 << ShAmt)), 0) --> and (lshr X, ShAmt), 11282 Value *X, *ShAmt;1283 if (Cmp->hasOneUse() && match(Cmp->getOperand(1), m_ZeroInt()) &&1284 match(Cmp->getOperand(0),1285 m_OneUse(m_c_And(m_Shl(m_One(), m_Value(ShAmt)), m_Value(X))))) {1286 auto *And = cast<BinaryOperator>(Cmp->getOperand(0));1287 Value *Shift = And->getOperand(X == And->getOperand(0) ? 1 : 0);1288 if (Zext.getType() == And->getType() ||1289 Cmp->getPredicate() != ICmpInst::ICMP_EQ || Shift->hasOneUse()) {1290 if (Cmp->getPredicate() == ICmpInst::ICMP_EQ)1291 X = Builder.CreateNot(X);1292 Value *Lshr = Builder.CreateLShr(X, ShAmt);1293 Value *And1 =1294 Builder.CreateAnd(Lshr, ConstantInt::get(X->getType(), 1));1295 return replaceInstUsesWith(1296 Zext, Builder.CreateZExtOrTrunc(And1, Zext.getType()));1297 }1298 }1299 }1300 1301 return nullptr;1302}1303 1304/// Determine if the specified value can be computed in the specified wider type1305/// and produce the same low bits. If not, return false.1306///1307/// If this function returns true, it can also return a non-zero number of bits1308/// (in BitsToClear) which indicates that the value it computes is correct for1309/// the zero extend, but that the additional BitsToClear bits need to be zero'd1310/// out. For example, to promote something like:1311///1312/// %B = trunc i64 %A to i321313/// %C = lshr i32 %B, 81314/// %E = zext i32 %C to i641315///1316/// CanEvaluateZExtd for the 'lshr' will return true, and BitsToClear will be1317/// set to 8 to indicate that the promoted value needs to have bits 24-311318/// cleared in addition to bits 32-63. Since an 'and' will be generated to1319/// clear the top bits anyway, doing this has no extra cost.1320///1321/// This function works on both vectors and scalars.1322bool TypeEvaluationHelper::canEvaluateZExtd(Value *V, Type *Ty,1323 unsigned &BitsToClear,1324 InstCombinerImpl &IC,1325 Instruction *CxtI) {1326 TypeEvaluationHelper TYH;1327 return TYH.canEvaluateZExtdImpl(V, Ty, BitsToClear, IC, CxtI);1328}1329bool TypeEvaluationHelper::canEvaluateZExtdImpl(Value *V, Type *Ty,1330 unsigned &BitsToClear,1331 InstCombinerImpl &IC,1332 Instruction *CxtI) {1333 BitsToClear = 0;1334 if (canAlwaysEvaluateInType(V, Ty))1335 return true;1336 // We stick to the one-user limit for the ZExt transform due to the fact1337 // that this predicate returns two values: predicate result and BitsToClear.1338 if (canNotEvaluateInType(V, Ty))1339 return false;1340 1341 auto *I = cast<Instruction>(V);1342 unsigned Tmp;1343 switch (I->getOpcode()) {1344 case Instruction::ZExt: // zext(zext(x)) -> zext(x).1345 case Instruction::SExt: // zext(sext(x)) -> sext(x).1346 case Instruction::Trunc: // zext(trunc(x)) -> trunc(x) or zext(x)1347 return true;1348 case Instruction::And:1349 case Instruction::Or:1350 case Instruction::Xor:1351 case Instruction::Add:1352 case Instruction::Sub:1353 case Instruction::Mul:1354 if (!canEvaluateZExtdImpl(I->getOperand(0), Ty, BitsToClear, IC, CxtI) ||1355 !canEvaluateZExtdImpl(I->getOperand(1), Ty, Tmp, IC, CxtI))1356 return false;1357 // These can all be promoted if neither operand has 'bits to clear'.1358 if (BitsToClear == 0 && Tmp == 0)1359 return true;1360 1361 // If the operation is an AND/OR/XOR and the bits to clear are zero in the1362 // other side, BitsToClear is ok.1363 if (Tmp == 0 && I->isBitwiseLogicOp()) {1364 // We use MaskedValueIsZero here for generality, but the case we care1365 // about the most is constant RHS.1366 unsigned VSize = V->getType()->getScalarSizeInBits();1367 if (IC.MaskedValueIsZero(I->getOperand(1),1368 APInt::getHighBitsSet(VSize, BitsToClear),1369 CxtI)) {1370 // If this is an And instruction and all of the BitsToClear are1371 // known to be zero we can reset BitsToClear.1372 if (I->getOpcode() == Instruction::And)1373 BitsToClear = 0;1374 return true;1375 }1376 }1377 1378 // Otherwise, we don't know how to analyze this BitsToClear case yet.1379 return false;1380 1381 case Instruction::Shl: {1382 // We can promote shl(x, cst) if we can promote x. Since shl overwrites the1383 // upper bits we can reduce BitsToClear by the shift amount.1384 uint64_t ShiftAmt;1385 if (match(I->getOperand(1), m_ConstantInt(ShiftAmt))) {1386 if (!canEvaluateZExtdImpl(I->getOperand(0), Ty, BitsToClear, IC, CxtI))1387 return false;1388 BitsToClear = ShiftAmt < BitsToClear ? BitsToClear - ShiftAmt : 0;1389 return true;1390 }1391 return false;1392 }1393 case Instruction::LShr: {1394 // We can promote lshr(x, cst) if we can promote x. This requires the1395 // ultimate 'and' to clear out the high zero bits we're clearing out though.1396 uint64_t ShiftAmt;1397 if (match(I->getOperand(1), m_ConstantInt(ShiftAmt))) {1398 if (!canEvaluateZExtdImpl(I->getOperand(0), Ty, BitsToClear, IC, CxtI))1399 return false;1400 BitsToClear += ShiftAmt;1401 if (BitsToClear > V->getType()->getScalarSizeInBits())1402 BitsToClear = V->getType()->getScalarSizeInBits();1403 return true;1404 }1405 // Cannot promote variable LSHR.1406 return false;1407 }1408 case Instruction::Select:1409 if (!canEvaluateZExtdImpl(I->getOperand(1), Ty, Tmp, IC, CxtI) ||1410 !canEvaluateZExtdImpl(I->getOperand(2), Ty, BitsToClear, IC, CxtI) ||1411 // TODO: If important, we could handle the case when the BitsToClear are1412 // known zero in the disagreeing side.1413 Tmp != BitsToClear)1414 return false;1415 return true;1416 1417 case Instruction::PHI: {1418 // We can change a phi if we can change all operands. Note that we never1419 // get into trouble with cyclic PHIs here because we only consider1420 // instructions with a single use.1421 PHINode *PN = cast<PHINode>(I);1422 if (!canEvaluateZExtdImpl(PN->getIncomingValue(0), Ty, BitsToClear, IC,1423 CxtI))1424 return false;1425 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i)1426 if (!canEvaluateZExtdImpl(PN->getIncomingValue(i), Ty, Tmp, IC, CxtI) ||1427 // TODO: If important, we could handle the case when the BitsToClear1428 // are known zero in the disagreeing input.1429 Tmp != BitsToClear)1430 return false;1431 return true;1432 }1433 case Instruction::Call:1434 // llvm.vscale() can always be executed in larger type, because the1435 // value is automatically zero-extended.1436 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))1437 if (II->getIntrinsicID() == Intrinsic::vscale)1438 return true;1439 return false;1440 default:1441 // TODO: Can handle more cases here.1442 return false;1443 }1444}1445 1446Instruction *InstCombinerImpl::visitZExt(ZExtInst &Zext) {1447 // If this zero extend is only used by a truncate, let the truncate be1448 // eliminated before we try to optimize this zext.1449 if (Zext.hasOneUse() && isa<TruncInst>(Zext.user_back()) &&1450 !isa<Constant>(Zext.getOperand(0)))1451 return nullptr;1452 1453 // If one of the common conversion will work, do it.1454 if (Instruction *Result = commonCastTransforms(Zext))1455 return Result;1456 1457 Value *Src = Zext.getOperand(0);1458 Type *SrcTy = Src->getType(), *DestTy = Zext.getType();1459 1460 // zext nneg bool x -> 01461 if (SrcTy->isIntOrIntVectorTy(1) && Zext.hasNonNeg())1462 return replaceInstUsesWith(Zext, Constant::getNullValue(Zext.getType()));1463 1464 // Try to extend the entire expression tree to the wide destination type.1465 unsigned BitsToClear;1466 if (shouldChangeType(SrcTy, DestTy) &&1467 TypeEvaluationHelper::canEvaluateZExtd(Src, DestTy, BitsToClear, *this,1468 &Zext)) {1469 assert(BitsToClear <= SrcTy->getScalarSizeInBits() &&1470 "Can't clear more bits than in SrcTy");1471 1472 // Okay, we can transform this! Insert the new expression now.1473 LLVM_DEBUG(1474 dbgs() << "ICE: EvaluateInDifferentType converting expression type"1475 " to avoid zero extend: "1476 << Zext << '\n');1477 Value *Res = EvaluateInDifferentType(Src, DestTy, false);1478 assert(Res->getType() == DestTy);1479 1480 // Preserve debug values referring to Src if the zext is its last use.1481 if (auto *SrcOp = dyn_cast<Instruction>(Src))1482 if (SrcOp->hasOneUse())1483 replaceAllDbgUsesWith(*SrcOp, *Res, Zext, DT);1484 1485 uint32_t SrcBitsKept = SrcTy->getScalarSizeInBits() - BitsToClear;1486 uint32_t DestBitSize = DestTy->getScalarSizeInBits();1487 1488 // If the high bits are already filled with zeros, just replace this1489 // cast with the result.1490 if (MaskedValueIsZero(1491 Res, APInt::getHighBitsSet(DestBitSize, DestBitSize - SrcBitsKept),1492 &Zext))1493 return replaceInstUsesWith(Zext, Res);1494 1495 // We need to emit an AND to clear the high bits.1496 Constant *C = ConstantInt::get(Res->getType(),1497 APInt::getLowBitsSet(DestBitSize, SrcBitsKept));1498 return BinaryOperator::CreateAnd(Res, C);1499 }1500 1501 // If this is a TRUNC followed by a ZEXT then we are dealing with integral1502 // types and if the sizes are just right we can convert this into a logical1503 // 'and' which will be much cheaper than the pair of casts.1504 if (auto *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast1505 // TODO: Subsume this into EvaluateInDifferentType.1506 1507 // Get the sizes of the types involved. We know that the intermediate type1508 // will be smaller than A or C, but don't know the relation between A and C.1509 Value *A = CSrc->getOperand(0);1510 unsigned SrcSize = A->getType()->getScalarSizeInBits();1511 unsigned MidSize = CSrc->getType()->getScalarSizeInBits();1512 unsigned DstSize = DestTy->getScalarSizeInBits();1513 // If we're actually extending zero bits, then if1514 // SrcSize < DstSize: zext(a & mask)1515 // SrcSize == DstSize: a & mask1516 // SrcSize > DstSize: trunc(a) & mask1517 if (SrcSize < DstSize) {1518 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));1519 Constant *AndConst = ConstantInt::get(A->getType(), AndValue);1520 Value *And = Builder.CreateAnd(A, AndConst, CSrc->getName() + ".mask");1521 return new ZExtInst(And, DestTy);1522 }1523 1524 if (SrcSize == DstSize) {1525 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));1526 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(),1527 AndValue));1528 }1529 if (SrcSize > DstSize) {1530 Value *Trunc = Builder.CreateTrunc(A, DestTy);1531 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize));1532 return BinaryOperator::CreateAnd(Trunc,1533 ConstantInt::get(Trunc->getType(),1534 AndValue));1535 }1536 }1537 1538 if (auto *Cmp = dyn_cast<ICmpInst>(Src))1539 return transformZExtICmp(Cmp, Zext);1540 1541 // zext(trunc(X) & C) -> (X & zext(C)).1542 Constant *C;1543 Value *X;1544 if (match(Src, m_OneUse(m_And(m_Trunc(m_Value(X)), m_Constant(C)))) &&1545 X->getType() == DestTy)1546 return BinaryOperator::CreateAnd(X, Builder.CreateZExt(C, DestTy));1547 1548 // zext((trunc(X) & C) ^ C) -> ((X & zext(C)) ^ zext(C)).1549 Value *And;1550 if (match(Src, m_OneUse(m_Xor(m_Value(And), m_Constant(C)))) &&1551 match(And, m_OneUse(m_And(m_Trunc(m_Value(X)), m_Specific(C)))) &&1552 X->getType() == DestTy) {1553 Value *ZC = Builder.CreateZExt(C, DestTy);1554 return BinaryOperator::CreateXor(Builder.CreateAnd(X, ZC), ZC);1555 }1556 1557 // If we are truncating, masking, and then zexting back to the original type,1558 // that's just a mask. This is not handled by canEvaluateZextd if the1559 // intermediate values have extra uses. This could be generalized further for1560 // a non-constant mask operand.1561 // zext (and (trunc X), C) --> and X, (zext C)1562 if (match(Src, m_And(m_Trunc(m_Value(X)), m_Constant(C))) &&1563 X->getType() == DestTy) {1564 Value *ZextC = Builder.CreateZExt(C, DestTy);1565 return BinaryOperator::CreateAnd(X, ZextC);1566 }1567 1568 if (match(Src, m_VScale())) {1569 if (Zext.getFunction() &&1570 Zext.getFunction()->hasFnAttribute(Attribute::VScaleRange)) {1571 Attribute Attr =1572 Zext.getFunction()->getFnAttribute(Attribute::VScaleRange);1573 if (std::optional<unsigned> MaxVScale = Attr.getVScaleRangeMax()) {1574 unsigned TypeWidth = Src->getType()->getScalarSizeInBits();1575 if (Log2_32(*MaxVScale) < TypeWidth)1576 return replaceInstUsesWith(Zext, Builder.CreateVScale(DestTy));1577 }1578 }1579 }1580 1581 if (!Zext.hasNonNeg()) {1582 // If this zero extend is only used by a shift, add nneg flag.1583 if (Zext.hasOneUse() &&1584 SrcTy->getScalarSizeInBits() >1585 Log2_64_Ceil(DestTy->getScalarSizeInBits()) &&1586 match(Zext.user_back(), m_Shift(m_Value(), m_Specific(&Zext)))) {1587 Zext.setNonNeg();1588 return &Zext;1589 }1590 1591 if (isKnownNonNegative(Src, SQ.getWithInstruction(&Zext))) {1592 Zext.setNonNeg();1593 return &Zext;1594 }1595 }1596 1597 return nullptr;1598}1599 1600/// Transform (sext icmp) to bitwise / integer operations to eliminate the icmp.1601Instruction *InstCombinerImpl::transformSExtICmp(ICmpInst *Cmp,1602 SExtInst &Sext) {1603 Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1);1604 ICmpInst::Predicate Pred = Cmp->getPredicate();1605 1606 // Don't bother if Op1 isn't of vector or integer type.1607 if (!Op1->getType()->isIntOrIntVectorTy())1608 return nullptr;1609 1610 if (Pred == ICmpInst::ICMP_SLT && match(Op1, m_ZeroInt())) {1611 // sext (x <s 0) --> ashr x, 31 (all ones if negative)1612 Value *Sh = ConstantInt::get(Op0->getType(),1613 Op0->getType()->getScalarSizeInBits() - 1);1614 Value *In = Builder.CreateAShr(Op0, Sh, Op0->getName() + ".lobit");1615 if (In->getType() != Sext.getType())1616 In = Builder.CreateIntCast(In, Sext.getType(), true /*SExt*/);1617 1618 return replaceInstUsesWith(Sext, In);1619 }1620 1621 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {1622 // If we know that only one bit of the LHS of the icmp can be set and we1623 // have an equality comparison with zero or a power of 2, we can transform1624 // the icmp and sext into bitwise/integer operations.1625 if (Cmp->hasOneUse() &&1626 Cmp->isEquality() && (Op1C->isZero() || Op1C->getValue().isPowerOf2())){1627 KnownBits Known = computeKnownBits(Op0, &Sext);1628 1629 APInt KnownZeroMask(~Known.Zero);1630 if (KnownZeroMask.isPowerOf2()) {1631 Value *In = Cmp->getOperand(0);1632 1633 // If the icmp tests for a known zero bit we can constant fold it.1634 if (!Op1C->isZero() && Op1C->getValue() != KnownZeroMask) {1635 Value *V = Pred == ICmpInst::ICMP_NE ?1636 ConstantInt::getAllOnesValue(Sext.getType()) :1637 ConstantInt::getNullValue(Sext.getType());1638 return replaceInstUsesWith(Sext, V);1639 }1640 1641 if (!Op1C->isZero() == (Pred == ICmpInst::ICMP_NE)) {1642 // sext ((x & 2^n) == 0) -> (x >> n) - 11643 // sext ((x & 2^n) != 2^n) -> (x >> n) - 11644 unsigned ShiftAmt = KnownZeroMask.countr_zero();1645 // Perform a right shift to place the desired bit in the LSB.1646 if (ShiftAmt)1647 In = Builder.CreateLShr(In,1648 ConstantInt::get(In->getType(), ShiftAmt));1649 1650 // At this point "In" is either 1 or 0. Subtract 1 to turn1651 // {1, 0} -> {0, -1}.1652 In = Builder.CreateAdd(In,1653 ConstantInt::getAllOnesValue(In->getType()),1654 "sext");1655 } else {1656 // sext ((x & 2^n) != 0) -> (x << bitwidth-n) a>> bitwidth-11657 // sext ((x & 2^n) == 2^n) -> (x << bitwidth-n) a>> bitwidth-11658 unsigned ShiftAmt = KnownZeroMask.countl_zero();1659 // Perform a left shift to place the desired bit in the MSB.1660 if (ShiftAmt)1661 In = Builder.CreateShl(In,1662 ConstantInt::get(In->getType(), ShiftAmt));1663 1664 // Distribute the bit over the whole bit width.1665 In = Builder.CreateAShr(In, ConstantInt::get(In->getType(),1666 KnownZeroMask.getBitWidth() - 1), "sext");1667 }1668 1669 if (Sext.getType() == In->getType())1670 return replaceInstUsesWith(Sext, In);1671 return CastInst::CreateIntegerCast(In, Sext.getType(), true/*SExt*/);1672 }1673 }1674 }1675 1676 return nullptr;1677}1678 1679/// Return true if we can take the specified value and return it as type Ty1680/// without inserting any new casts and without changing the value of the common1681/// low bits. This is used by code that tries to promote integer operations to1682/// a wider types will allow us to eliminate the extension.1683///1684/// This function works on both vectors and scalars.1685///1686bool TypeEvaluationHelper::canEvaluateSExtd(Value *V, Type *Ty) {1687 TypeEvaluationHelper TYH;1688 return TYH.canEvaluateSExtdImpl(V, Ty) && TYH.allPendingVisited();1689}1690 1691bool TypeEvaluationHelper::canEvaluateSExtdImpl(Value *V, Type *Ty) {1692 return canEvaluate(V, Ty, [this](Value *V, Type *Ty) {1693 return canEvaluateSExtdPred(V, Ty);1694 });1695}1696 1697bool TypeEvaluationHelper::canEvaluateSExtdPred(Value *V, Type *Ty) {1698 assert(V->getType()->getScalarSizeInBits() < Ty->getScalarSizeInBits() &&1699 "Can't sign extend type to a smaller type");1700 1701 auto *I = cast<Instruction>(V);1702 switch (I->getOpcode()) {1703 case Instruction::SExt: // sext(sext(x)) -> sext(x)1704 case Instruction::ZExt: // sext(zext(x)) -> zext(x)1705 case Instruction::Trunc: // sext(trunc(x)) -> trunc(x) or sext(x)1706 return true;1707 case Instruction::And:1708 case Instruction::Or:1709 case Instruction::Xor:1710 case Instruction::Add:1711 case Instruction::Sub:1712 case Instruction::Mul:1713 // These operators can all arbitrarily be extended if their inputs can.1714 return canEvaluateSExtdImpl(I->getOperand(0), Ty) &&1715 canEvaluateSExtdImpl(I->getOperand(1), Ty);1716 1717 // case Instruction::Shl: TODO1718 // case Instruction::LShr: TODO1719 1720 case Instruction::Select:1721 return canEvaluateSExtdImpl(I->getOperand(1), Ty) &&1722 canEvaluateSExtdImpl(I->getOperand(2), Ty);1723 1724 case Instruction::PHI: {1725 // We can change a phi if we can change all operands. Note that we never1726 // get into trouble with cyclic PHIs here because canEvaluate handles use1727 // chain loops.1728 PHINode *PN = cast<PHINode>(I);1729 for (Value *IncValue : PN->incoming_values())1730 if (!canEvaluateSExtdImpl(IncValue, Ty))1731 return false;1732 return true;1733 }1734 default:1735 // TODO: Can handle more cases here.1736 break;1737 }1738 1739 return false;1740}1741 1742Instruction *InstCombinerImpl::visitSExt(SExtInst &Sext) {1743 // If this sign extend is only used by a truncate, let the truncate be1744 // eliminated before we try to optimize this sext.1745 if (Sext.hasOneUse() && isa<TruncInst>(Sext.user_back()))1746 return nullptr;1747 1748 if (Instruction *I = commonCastTransforms(Sext))1749 return I;1750 1751 Value *Src = Sext.getOperand(0);1752 Type *SrcTy = Src->getType(), *DestTy = Sext.getType();1753 unsigned SrcBitSize = SrcTy->getScalarSizeInBits();1754 unsigned DestBitSize = DestTy->getScalarSizeInBits();1755 1756 // If the value being extended is zero or positive, use a zext instead.1757 if (isKnownNonNegative(Src, SQ.getWithInstruction(&Sext))) {1758 auto CI = CastInst::Create(Instruction::ZExt, Src, DestTy);1759 CI->setNonNeg(true);1760 return CI;1761 }1762 1763 // Try to extend the entire expression tree to the wide destination type.1764 bool ShouldExtendExpression = true;1765 Value *TruncSrc = nullptr;1766 // It is not desirable to extend expression in the trunc + sext pattern when1767 // destination type is narrower than original (pre-trunc) type.1768 if (match(Src, m_Trunc(m_Value(TruncSrc))))1769 if (TruncSrc->getType()->getScalarSizeInBits() > DestBitSize)1770 ShouldExtendExpression = false;1771 if (ShouldExtendExpression && shouldChangeType(SrcTy, DestTy) &&1772 TypeEvaluationHelper::canEvaluateSExtd(Src, DestTy)) {1773 // Okay, we can transform this! Insert the new expression now.1774 LLVM_DEBUG(1775 dbgs() << "ICE: EvaluateInDifferentType converting expression type"1776 " to avoid sign extend: "1777 << Sext << '\n');1778 Value *Res = EvaluateInDifferentType(Src, DestTy, true);1779 assert(Res->getType() == DestTy);1780 1781 // If the high bits are already filled with sign bit, just replace this1782 // cast with the result.1783 if (ComputeNumSignBits(Res, &Sext) > DestBitSize - SrcBitSize)1784 return replaceInstUsesWith(Sext, Res);1785 1786 // We need to emit a shl + ashr to do the sign extend.1787 Value *ShAmt = ConstantInt::get(DestTy, DestBitSize - SrcBitSize);1788 return BinaryOperator::CreateAShr(Builder.CreateShl(Res, ShAmt, "sext"),1789 ShAmt);1790 }1791 1792 Value *X = TruncSrc;1793 if (X) {1794 // If the input has more sign bits than bits truncated, then convert1795 // directly to final type.1796 unsigned XBitSize = X->getType()->getScalarSizeInBits();1797 bool HasNSW = cast<TruncInst>(Src)->hasNoSignedWrap();1798 if (HasNSW || (ComputeNumSignBits(X, &Sext) > XBitSize - SrcBitSize)) {1799 auto *Res = CastInst::CreateIntegerCast(X, DestTy, /* isSigned */ true);1800 if (auto *ResTrunc = dyn_cast<TruncInst>(Res); ResTrunc && HasNSW)1801 ResTrunc->setHasNoSignedWrap(true);1802 return Res;1803 }1804 1805 // If input is a trunc from the destination type, then convert into shifts.1806 if (Src->hasOneUse() && X->getType() == DestTy) {1807 // sext (trunc X) --> ashr (shl X, C), C1808 Constant *ShAmt = ConstantInt::get(DestTy, DestBitSize - SrcBitSize);1809 return BinaryOperator::CreateAShr(Builder.CreateShl(X, ShAmt), ShAmt);1810 }1811 1812 // If we are replacing shifted-in high zero bits with sign bits, convert1813 // the logic shift to arithmetic shift and eliminate the cast to1814 // intermediate type:1815 // sext (trunc (lshr Y, C)) --> sext/trunc (ashr Y, C)1816 Value *Y;1817 if (Src->hasOneUse() &&1818 match(X, m_LShr(m_Value(Y),1819 m_SpecificIntAllowPoison(XBitSize - SrcBitSize)))) {1820 Value *Ashr = Builder.CreateAShr(Y, XBitSize - SrcBitSize);1821 return CastInst::CreateIntegerCast(Ashr, DestTy, /* isSigned */ true);1822 }1823 }1824 1825 if (auto *Cmp = dyn_cast<ICmpInst>(Src))1826 return transformSExtICmp(Cmp, Sext);1827 1828 // If the input is a shl/ashr pair of a same constant, then this is a sign1829 // extension from a smaller value. If we could trust arbitrary bitwidth1830 // integers, we could turn this into a truncate to the smaller bit and then1831 // use a sext for the whole extension. Since we don't, look deeper and check1832 // for a truncate. If the source and dest are the same type, eliminate the1833 // trunc and extend and just do shifts. For example, turn:1834 // %a = trunc i32 %i to i81835 // %b = shl i8 %a, C1836 // %c = ashr i8 %b, C1837 // %d = sext i8 %c to i321838 // into:1839 // %a = shl i32 %i, 32-(8-C)1840 // %d = ashr i32 %a, 32-(8-C)1841 Value *A = nullptr;1842 // TODO: Eventually this could be subsumed by EvaluateInDifferentType.1843 Constant *BA = nullptr, *CA = nullptr;1844 if (match(Src, m_AShr(m_Shl(m_Trunc(m_Value(A)), m_Constant(BA)),1845 m_ImmConstant(CA))) &&1846 BA->isElementWiseEqual(CA) && A->getType() == DestTy) {1847 Constant *WideCurrShAmt =1848 ConstantFoldCastOperand(Instruction::SExt, CA, DestTy, DL);1849 assert(WideCurrShAmt && "Constant folding of ImmConstant cannot fail");1850 Constant *NumLowbitsLeft = ConstantExpr::getSub(1851 ConstantInt::get(DestTy, SrcTy->getScalarSizeInBits()), WideCurrShAmt);1852 Constant *NewShAmt = ConstantExpr::getSub(1853 ConstantInt::get(DestTy, DestTy->getScalarSizeInBits()),1854 NumLowbitsLeft);1855 NewShAmt =1856 Constant::mergeUndefsWith(Constant::mergeUndefsWith(NewShAmt, BA), CA);1857 A = Builder.CreateShl(A, NewShAmt, Sext.getName());1858 return BinaryOperator::CreateAShr(A, NewShAmt);1859 }1860 1861 // Splatting a bit of constant-index across a value:1862 // sext (ashr (trunc iN X to iM), M-1) to iN --> ashr (shl X, N-M), N-11863 // If the dest type is different, use a cast (adjust use check).1864 if (match(Src, m_OneUse(m_AShr(m_Trunc(m_Value(X)),1865 m_SpecificInt(SrcBitSize - 1))))) {1866 Type *XTy = X->getType();1867 unsigned XBitSize = XTy->getScalarSizeInBits();1868 Constant *ShlAmtC = ConstantInt::get(XTy, XBitSize - SrcBitSize);1869 Constant *AshrAmtC = ConstantInt::get(XTy, XBitSize - 1);1870 if (XTy == DestTy)1871 return BinaryOperator::CreateAShr(Builder.CreateShl(X, ShlAmtC),1872 AshrAmtC);1873 if (cast<BinaryOperator>(Src)->getOperand(0)->hasOneUse()) {1874 Value *Ashr = Builder.CreateAShr(Builder.CreateShl(X, ShlAmtC), AshrAmtC);1875 return CastInst::CreateIntegerCast(Ashr, DestTy, /* isSigned */ true);1876 }1877 }1878 1879 if (match(Src, m_VScale())) {1880 if (Sext.getFunction() &&1881 Sext.getFunction()->hasFnAttribute(Attribute::VScaleRange)) {1882 Attribute Attr =1883 Sext.getFunction()->getFnAttribute(Attribute::VScaleRange);1884 if (std::optional<unsigned> MaxVScale = Attr.getVScaleRangeMax())1885 if (Log2_32(*MaxVScale) < (SrcBitSize - 1))1886 return replaceInstUsesWith(Sext, Builder.CreateVScale(DestTy));1887 }1888 }1889 1890 return nullptr;1891}1892 1893/// Return a Constant* for the specified floating-point constant if it fits1894/// in the specified FP type without changing its value.1895static bool fitsInFPType(APFloat F, const fltSemantics &Sem) {1896 bool losesInfo;1897 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);1898 return !losesInfo;1899}1900 1901static Type *shrinkFPConstant(LLVMContext &Ctx, const APFloat &F,1902 bool PreferBFloat) {1903 // See if the value can be truncated to bfloat and then reextended.1904 if (PreferBFloat && fitsInFPType(F, APFloat::BFloat()))1905 return Type::getBFloatTy(Ctx);1906 // See if the value can be truncated to half and then reextended.1907 if (!PreferBFloat && fitsInFPType(F, APFloat::IEEEhalf()))1908 return Type::getHalfTy(Ctx);1909 // See if the value can be truncated to float and then reextended.1910 if (fitsInFPType(F, APFloat::IEEEsingle()))1911 return Type::getFloatTy(Ctx);1912 if (&F.getSemantics() == &APFloat::IEEEdouble())1913 return nullptr; // Won't shrink.1914 // See if the value can be truncated to double and then reextended.1915 if (fitsInFPType(F, APFloat::IEEEdouble()))1916 return Type::getDoubleTy(Ctx);1917 // Don't try to shrink to various long double types.1918 return nullptr;1919}1920 1921static Type *shrinkFPConstant(ConstantFP *CFP, bool PreferBFloat) {1922 Type *Ty = CFP->getType();1923 if (Ty->getScalarType()->isPPC_FP128Ty())1924 return nullptr; // No constant folding of this.1925 1926 Type *ShrinkTy =1927 shrinkFPConstant(CFP->getContext(), CFP->getValueAPF(), PreferBFloat);1928 if (ShrinkTy)1929 if (auto *VecTy = dyn_cast<VectorType>(Ty))1930 ShrinkTy = VectorType::get(ShrinkTy, VecTy);1931 1932 return ShrinkTy;1933}1934 1935// Determine if this is a vector of ConstantFPs and if so, return the minimal1936// type we can safely truncate all elements to.1937static Type *shrinkFPConstantVector(Value *V, bool PreferBFloat) {1938 auto *CV = dyn_cast<Constant>(V);1939 auto *CVVTy = dyn_cast<FixedVectorType>(V->getType());1940 if (!CV || !CVVTy)1941 return nullptr;1942 1943 Type *MinType = nullptr;1944 1945 unsigned NumElts = CVVTy->getNumElements();1946 1947 // For fixed-width vectors we find the minimal type by looking1948 // through the constant values of the vector.1949 for (unsigned i = 0; i != NumElts; ++i) {1950 if (isa<UndefValue>(CV->getAggregateElement(i)))1951 continue;1952 1953 auto *CFP = dyn_cast_or_null<ConstantFP>(CV->getAggregateElement(i));1954 if (!CFP)1955 return nullptr;1956 1957 Type *T = shrinkFPConstant(CFP, PreferBFloat);1958 if (!T)1959 return nullptr;1960 1961 // If we haven't found a type yet or this type has a larger mantissa than1962 // our previous type, this is our new minimal type.1963 if (!MinType || T->getFPMantissaWidth() > MinType->getFPMantissaWidth())1964 MinType = T;1965 }1966 1967 // Make a vector type from the minimal type.1968 return MinType ? FixedVectorType::get(MinType, NumElts) : nullptr;1969}1970 1971/// Find the minimum FP type we can safely truncate to.1972static Type *getMinimumFPType(Value *V, bool PreferBFloat) {1973 if (auto *FPExt = dyn_cast<FPExtInst>(V))1974 return FPExt->getOperand(0)->getType();1975 1976 // If this value is a constant, return the constant in the smallest FP type1977 // that can accurately represent it. This allows us to turn1978 // (float)((double)X+2.0) into x+2.0f.1979 if (auto *CFP = dyn_cast<ConstantFP>(V))1980 if (Type *T = shrinkFPConstant(CFP, PreferBFloat))1981 return T;1982 1983 // Try to shrink scalable and fixed splat vectors.1984 if (auto *FPC = dyn_cast<Constant>(V))1985 if (auto *VTy = dyn_cast<VectorType>(V->getType()))1986 if (auto *Splat = dyn_cast_or_null<ConstantFP>(FPC->getSplatValue()))1987 if (Type *T = shrinkFPConstant(Splat, PreferBFloat))1988 return VectorType::get(T, VTy);1989 1990 // Try to shrink a vector of FP constants. This returns nullptr on scalable1991 // vectors1992 if (Type *T = shrinkFPConstantVector(V, PreferBFloat))1993 return T;1994 1995 return V->getType();1996}1997 1998/// Return true if the cast from integer to FP can be proven to be exact for all1999/// possible inputs (the conversion does not lose any precision).2000static bool isKnownExactCastIntToFP(CastInst &I, InstCombinerImpl &IC) {2001 CastInst::CastOps Opcode = I.getOpcode();2002 assert((Opcode == CastInst::SIToFP || Opcode == CastInst::UIToFP) &&2003 "Unexpected cast");2004 Value *Src = I.getOperand(0);2005 Type *SrcTy = Src->getType();2006 Type *FPTy = I.getType();2007 bool IsSigned = Opcode == Instruction::SIToFP;2008 int SrcSize = (int)SrcTy->getScalarSizeInBits() - IsSigned;2009 2010 // Easy case - if the source integer type has less bits than the FP mantissa,2011 // then the cast must be exact.2012 int DestNumSigBits = FPTy->getFPMantissaWidth();2013 if (SrcSize <= DestNumSigBits)2014 return true;2015 2016 // Cast from FP to integer and back to FP is independent of the intermediate2017 // integer width because of poison on overflow.2018 Value *F;2019 if (match(Src, m_FPToSI(m_Value(F))) || match(Src, m_FPToUI(m_Value(F)))) {2020 // If this is uitofp (fptosi F), the source needs an extra bit to avoid2021 // potential rounding of negative FP input values.2022 int SrcNumSigBits = F->getType()->getFPMantissaWidth();2023 if (!IsSigned && match(Src, m_FPToSI(m_Value())))2024 SrcNumSigBits++;2025 2026 // [su]itofp (fpto[su]i F) --> exact if the source type has less or equal2027 // significant bits than the destination (and make sure neither type is2028 // weird -- ppc_fp128).2029 if (SrcNumSigBits > 0 && DestNumSigBits > 0 &&2030 SrcNumSigBits <= DestNumSigBits)2031 return true;2032 }2033 2034 // TODO:2035 // Try harder to find if the source integer type has less significant bits.2036 // For example, compute number of sign bits.2037 KnownBits SrcKnown = IC.computeKnownBits(Src, &I);2038 int SigBits = (int)SrcTy->getScalarSizeInBits() -2039 SrcKnown.countMinLeadingZeros() -2040 SrcKnown.countMinTrailingZeros();2041 if (SigBits <= DestNumSigBits)2042 return true;2043 2044 return false;2045}2046 2047Instruction *InstCombinerImpl::visitFPTrunc(FPTruncInst &FPT) {2048 if (Instruction *I = commonCastTransforms(FPT))2049 return I;2050 2051 // If we have fptrunc(OpI (fpextend x), (fpextend y)), we would like to2052 // simplify this expression to avoid one or more of the trunc/extend2053 // operations if we can do so without changing the numerical results.2054 //2055 // The exact manner in which the widths of the operands interact to limit2056 // what we can and cannot do safely varies from operation to operation, and2057 // is explained below in the various case statements.2058 Type *Ty = FPT.getType();2059 auto *BO = dyn_cast<BinaryOperator>(FPT.getOperand(0));2060 if (BO && BO->hasOneUse()) {2061 bool PreferBFloat = Ty->getScalarType()->isBFloatTy();2062 Type *LHSMinType = getMinimumFPType(BO->getOperand(0), PreferBFloat);2063 Type *RHSMinType = getMinimumFPType(BO->getOperand(1), PreferBFloat);2064 unsigned OpWidth = BO->getType()->getFPMantissaWidth();2065 unsigned LHSWidth = LHSMinType->getFPMantissaWidth();2066 unsigned RHSWidth = RHSMinType->getFPMantissaWidth();2067 unsigned SrcWidth = std::max(LHSWidth, RHSWidth);2068 unsigned DstWidth = Ty->getFPMantissaWidth();2069 switch (BO->getOpcode()) {2070 default: break;2071 case Instruction::FAdd:2072 case Instruction::FSub:2073 // For addition and subtraction, the infinitely precise result can2074 // essentially be arbitrarily wide; proving that double rounding2075 // will not occur because the result of OpI is exact (as we will for2076 // FMul, for example) is hopeless. However, we *can* nonetheless2077 // frequently know that double rounding cannot occur (or that it is2078 // innocuous) by taking advantage of the specific structure of2079 // infinitely-precise results that admit double rounding.2080 //2081 // Specifically, if OpWidth >= 2*DstWdith+1 and DstWidth is sufficient2082 // to represent both sources, we can guarantee that the double2083 // rounding is innocuous (See p50 of Figueroa's 2000 PhD thesis,2084 // "A Rigorous Framework for Fully Supporting the IEEE Standard ..."2085 // for proof of this fact).2086 //2087 // Note: Figueroa does not consider the case where DstFormat !=2088 // SrcFormat. It's possible (likely even!) that this analysis2089 // could be tightened for those cases, but they are rare (the main2090 // case of interest here is (float)((double)float + float)).2091 if (OpWidth >= 2*DstWidth+1 && DstWidth >= SrcWidth) {2092 Value *LHS = Builder.CreateFPTrunc(BO->getOperand(0), Ty);2093 Value *RHS = Builder.CreateFPTrunc(BO->getOperand(1), Ty);2094 Instruction *RI = BinaryOperator::Create(BO->getOpcode(), LHS, RHS);2095 RI->copyFastMathFlags(BO);2096 return RI;2097 }2098 break;2099 case Instruction::FMul:2100 // For multiplication, the infinitely precise result has at most2101 // LHSWidth + RHSWidth significant bits; if OpWidth is sufficient2102 // that such a value can be exactly represented, then no double2103 // rounding can possibly occur; we can safely perform the operation2104 // in the destination format if it can represent both sources.2105 if (OpWidth >= LHSWidth + RHSWidth && DstWidth >= SrcWidth) {2106 Value *LHS = Builder.CreateFPTrunc(BO->getOperand(0), Ty);2107 Value *RHS = Builder.CreateFPTrunc(BO->getOperand(1), Ty);2108 return BinaryOperator::CreateFMulFMF(LHS, RHS, BO);2109 }2110 break;2111 case Instruction::FDiv:2112 // For division, we use again use the bound from Figueroa's2113 // dissertation. I am entirely certain that this bound can be2114 // tightened in the unbalanced operand case by an analysis based on2115 // the diophantine rational approximation bound, but the well-known2116 // condition used here is a good conservative first pass.2117 // TODO: Tighten bound via rigorous analysis of the unbalanced case.2118 if (OpWidth >= 2*DstWidth && DstWidth >= SrcWidth) {2119 Value *LHS = Builder.CreateFPTrunc(BO->getOperand(0), Ty);2120 Value *RHS = Builder.CreateFPTrunc(BO->getOperand(1), Ty);2121 return BinaryOperator::CreateFDivFMF(LHS, RHS, BO);2122 }2123 break;2124 case Instruction::FRem: {2125 // Remainder is straightforward. Remainder is always exact, so the2126 // type of OpI doesn't enter into things at all. We simply evaluate2127 // in whichever source type is larger, then convert to the2128 // destination type.2129 if (SrcWidth == OpWidth)2130 break;2131 Value *LHS, *RHS;2132 if (LHSWidth == SrcWidth) {2133 LHS = Builder.CreateFPTrunc(BO->getOperand(0), LHSMinType);2134 RHS = Builder.CreateFPTrunc(BO->getOperand(1), LHSMinType);2135 } else {2136 LHS = Builder.CreateFPTrunc(BO->getOperand(0), RHSMinType);2137 RHS = Builder.CreateFPTrunc(BO->getOperand(1), RHSMinType);2138 }2139 2140 Value *ExactResult = Builder.CreateFRemFMF(LHS, RHS, BO);2141 return CastInst::CreateFPCast(ExactResult, Ty);2142 }2143 }2144 }2145 2146 // (fptrunc (fneg x)) -> (fneg (fptrunc x))2147 Value *X;2148 Instruction *Op = dyn_cast<Instruction>(FPT.getOperand(0));2149 if (Op && Op->hasOneUse()) {2150 FastMathFlags FMF = FPT.getFastMathFlags();2151 if (auto *FPMO = dyn_cast<FPMathOperator>(Op))2152 FMF &= FPMO->getFastMathFlags();2153 2154 if (match(Op, m_FNeg(m_Value(X)))) {2155 Value *InnerTrunc = Builder.CreateFPTruncFMF(X, Ty, FMF);2156 Value *Neg = Builder.CreateFNegFMF(InnerTrunc, FMF);2157 return replaceInstUsesWith(FPT, Neg);2158 }2159 2160 // If we are truncating a select that has an extended operand, we can2161 // narrow the other operand and do the select as a narrow op.2162 Value *Cond, *X, *Y;2163 if (match(Op, m_Select(m_Value(Cond), m_FPExt(m_Value(X)), m_Value(Y))) &&2164 X->getType() == Ty) {2165 // fptrunc (select Cond, (fpext X), Y --> select Cond, X, (fptrunc Y)2166 Value *NarrowY = Builder.CreateFPTruncFMF(Y, Ty, FMF);2167 Value *Sel =2168 Builder.CreateSelectFMF(Cond, X, NarrowY, FMF, "narrow.sel", Op);2169 return replaceInstUsesWith(FPT, Sel);2170 }2171 if (match(Op, m_Select(m_Value(Cond), m_Value(Y), m_FPExt(m_Value(X)))) &&2172 X->getType() == Ty) {2173 // fptrunc (select Cond, Y, (fpext X) --> select Cond, (fptrunc Y), X2174 Value *NarrowY = Builder.CreateFPTruncFMF(Y, Ty, FMF);2175 Value *Sel =2176 Builder.CreateSelectFMF(Cond, NarrowY, X, FMF, "narrow.sel", Op);2177 return replaceInstUsesWith(FPT, Sel);2178 }2179 }2180 2181 if (auto *II = dyn_cast<IntrinsicInst>(FPT.getOperand(0))) {2182 switch (II->getIntrinsicID()) {2183 default: break;2184 case Intrinsic::ceil:2185 case Intrinsic::fabs:2186 case Intrinsic::floor:2187 case Intrinsic::nearbyint:2188 case Intrinsic::rint:2189 case Intrinsic::round:2190 case Intrinsic::roundeven:2191 case Intrinsic::trunc: {2192 Value *Src = II->getArgOperand(0);2193 if (!Src->hasOneUse())2194 break;2195 2196 // Except for fabs, this transformation requires the input of the unary FP2197 // operation to be itself an fpext from the type to which we're2198 // truncating.2199 if (II->getIntrinsicID() != Intrinsic::fabs) {2200 FPExtInst *FPExtSrc = dyn_cast<FPExtInst>(Src);2201 if (!FPExtSrc || FPExtSrc->getSrcTy() != Ty)2202 break;2203 }2204 2205 // Do unary FP operation on smaller type.2206 // (fptrunc (fabs x)) -> (fabs (fptrunc x))2207 Value *InnerTrunc = Builder.CreateFPTrunc(Src, Ty);2208 Function *Overload = Intrinsic::getOrInsertDeclaration(2209 FPT.getModule(), II->getIntrinsicID(), Ty);2210 SmallVector<OperandBundleDef, 1> OpBundles;2211 II->getOperandBundlesAsDefs(OpBundles);2212 CallInst *NewCI =2213 CallInst::Create(Overload, {InnerTrunc}, OpBundles, II->getName());2214 // A normal value may be converted to an infinity. It means that we cannot2215 // propagate ninf from the intrinsic. So we propagate FMF from fptrunc.2216 NewCI->copyFastMathFlags(&FPT);2217 return NewCI;2218 }2219 }2220 }2221 2222 if (Instruction *I = shrinkInsertElt(FPT, Builder))2223 return I;2224 2225 Value *Src = FPT.getOperand(0);2226 if (isa<SIToFPInst>(Src) || isa<UIToFPInst>(Src)) {2227 auto *FPCast = cast<CastInst>(Src);2228 if (isKnownExactCastIntToFP(*FPCast, *this))2229 return CastInst::Create(FPCast->getOpcode(), FPCast->getOperand(0), Ty);2230 }2231 2232 return nullptr;2233}2234 2235Instruction *InstCombinerImpl::visitFPExt(CastInst &FPExt) {2236 // If the source operand is a cast from integer to FP and known exact, then2237 // cast the integer operand directly to the destination type.2238 Type *Ty = FPExt.getType();2239 Value *Src = FPExt.getOperand(0);2240 if (isa<SIToFPInst>(Src) || isa<UIToFPInst>(Src)) {2241 auto *FPCast = cast<CastInst>(Src);2242 if (isKnownExactCastIntToFP(*FPCast, *this))2243 return CastInst::Create(FPCast->getOpcode(), FPCast->getOperand(0), Ty);2244 }2245 2246 return commonCastTransforms(FPExt);2247}2248 2249/// fpto{s/u}i({u/s}itofp(X)) --> X or zext(X) or sext(X) or trunc(X)2250/// This is safe if the intermediate type has enough bits in its mantissa to2251/// accurately represent all values of X. For example, this won't work with2252/// i64 -> float -> i64.2253Instruction *InstCombinerImpl::foldItoFPtoI(CastInst &FI) {2254 if (!isa<UIToFPInst>(FI.getOperand(0)) && !isa<SIToFPInst>(FI.getOperand(0)))2255 return nullptr;2256 2257 auto *OpI = cast<CastInst>(FI.getOperand(0));2258 Value *X = OpI->getOperand(0);2259 Type *XType = X->getType();2260 Type *DestType = FI.getType();2261 bool IsOutputSigned = isa<FPToSIInst>(FI);2262 2263 // Since we can assume the conversion won't overflow, our decision as to2264 // whether the input will fit in the float should depend on the minimum2265 // of the input range and output range.2266 2267 // This means this is also safe for a signed input and unsigned output, since2268 // a negative input would lead to undefined behavior.2269 if (!isKnownExactCastIntToFP(*OpI, *this)) {2270 // The first cast may not round exactly based on the source integer width2271 // and FP width, but the overflow UB rules can still allow this to fold.2272 // If the destination type is narrow, that means the intermediate FP value2273 // must be large enough to hold the source value exactly.2274 // For example, (uint8_t)((float)(uint32_t 16777217) is undefined behavior.2275 int OutputSize = (int)DestType->getScalarSizeInBits();2276 if (OutputSize > OpI->getType()->getFPMantissaWidth())2277 return nullptr;2278 }2279 2280 if (DestType->getScalarSizeInBits() > XType->getScalarSizeInBits()) {2281 bool IsInputSigned = isa<SIToFPInst>(OpI);2282 if (IsInputSigned && IsOutputSigned)2283 return new SExtInst(X, DestType);2284 return new ZExtInst(X, DestType);2285 }2286 if (DestType->getScalarSizeInBits() < XType->getScalarSizeInBits())2287 return new TruncInst(X, DestType);2288 2289 assert(XType == DestType && "Unexpected types for int to FP to int casts");2290 return replaceInstUsesWith(FI, X);2291}2292 2293static Instruction *foldFPtoI(Instruction &FI, InstCombiner &IC) {2294 // fpto{u/s}i non-norm --> 02295 FPClassTest Mask =2296 FI.getOpcode() == Instruction::FPToUI ? fcPosNormal : fcNormal;2297 KnownFPClass FPClass = computeKnownFPClass(2298 FI.getOperand(0), Mask, IC.getSimplifyQuery().getWithInstruction(&FI));2299 if (FPClass.isKnownNever(Mask))2300 return IC.replaceInstUsesWith(FI, ConstantInt::getNullValue(FI.getType()));2301 2302 return nullptr;2303}2304 2305Instruction *InstCombinerImpl::visitFPToUI(FPToUIInst &FI) {2306 if (Instruction *I = foldItoFPtoI(FI))2307 return I;2308 2309 if (Instruction *I = foldFPtoI(FI, *this))2310 return I;2311 2312 return commonCastTransforms(FI);2313}2314 2315Instruction *InstCombinerImpl::visitFPToSI(FPToSIInst &FI) {2316 if (Instruction *I = foldItoFPtoI(FI))2317 return I;2318 2319 if (Instruction *I = foldFPtoI(FI, *this))2320 return I;2321 2322 return commonCastTransforms(FI);2323}2324 2325Instruction *InstCombinerImpl::visitUIToFP(CastInst &CI) {2326 if (Instruction *R = commonCastTransforms(CI))2327 return R;2328 if (!CI.hasNonNeg() && isKnownNonNegative(CI.getOperand(0), SQ)) {2329 CI.setNonNeg();2330 return &CI;2331 }2332 return nullptr;2333}2334 2335Instruction *InstCombinerImpl::visitSIToFP(CastInst &CI) {2336 if (Instruction *R = commonCastTransforms(CI))2337 return R;2338 if (isKnownNonNegative(CI.getOperand(0), SQ)) {2339 auto *UI =2340 CastInst::Create(Instruction::UIToFP, CI.getOperand(0), CI.getType());2341 UI->setNonNeg(true);2342 return UI;2343 }2344 return nullptr;2345}2346 2347Instruction *InstCombinerImpl::visitIntToPtr(IntToPtrInst &CI) {2348 // If the source integer type is not the intptr_t type for this target, do a2349 // trunc or zext to the intptr_t type, then inttoptr of it. This allows the2350 // cast to be exposed to other transforms.2351 unsigned AS = CI.getAddressSpace();2352 if (CI.getOperand(0)->getType()->getScalarSizeInBits() !=2353 DL.getPointerSizeInBits(AS)) {2354 Type *Ty = CI.getOperand(0)->getType()->getWithNewType(2355 DL.getIntPtrType(CI.getContext(), AS));2356 Value *P = Builder.CreateZExtOrTrunc(CI.getOperand(0), Ty);2357 return new IntToPtrInst(P, CI.getType());2358 }2359 2360 // Replace (inttoptr (add (ptrtoint %Base), %Offset)) with2361 // (getelementptr i8, %Base, %Offset) if the pointer is only used as integer2362 // value.2363 Value *Base;2364 Value *Offset;2365 auto UsesPointerAsInt = [](User *U) {2366 if (isa<ICmpInst, PtrToIntInst>(U))2367 return true;2368 if (auto *P = dyn_cast<PHINode>(U))2369 return P->hasOneUse() && isa<ICmpInst, PtrToIntInst>(*P->user_begin());2370 return false;2371 };2372 if (match(CI.getOperand(0),2373 m_OneUse(m_c_Add(m_PtrToIntSameSize(DL, m_Value(Base)),2374 m_Value(Offset)))) &&2375 CI.getType()->getPointerAddressSpace() ==2376 Base->getType()->getPointerAddressSpace() &&2377 all_of(CI.users(), UsesPointerAsInt)) {2378 return GetElementPtrInst::Create(Builder.getInt8Ty(), Base, Offset);2379 }2380 2381 if (Instruction *I = commonCastTransforms(CI))2382 return I;2383 2384 return nullptr;2385}2386 2387Value *InstCombinerImpl::foldPtrToIntOrAddrOfGEP(Type *IntTy, Value *Ptr) {2388 // Look through chain of one-use GEPs.2389 Type *PtrTy = Ptr->getType();2390 SmallVector<GEPOperator *> GEPs;2391 while (true) {2392 auto *GEP = dyn_cast<GEPOperator>(Ptr);2393 if (!GEP || !GEP->hasOneUse())2394 break;2395 GEPs.push_back(GEP);2396 Ptr = GEP->getPointerOperand();2397 }2398 2399 // Don't handle case where GEP converts from pointer to vector.2400 if (GEPs.empty() || PtrTy != Ptr->getType())2401 return nullptr;2402 2403 // Check whether we know the integer value of the base pointer.2404 Value *Res;2405 Type *IdxTy = DL.getIndexType(PtrTy);2406 if (match(Ptr, m_OneUse(m_IntToPtr(m_Value(Res)))) &&2407 Res->getType() == IntTy && IntTy == IdxTy) {2408 // pass2409 } else if (isa<ConstantPointerNull>(Ptr)) {2410 Res = Constant::getNullValue(IdxTy);2411 } else {2412 return nullptr;2413 }2414 2415 // Perform the entire operation on integers instead.2416 for (GEPOperator *GEP : reverse(GEPs)) {2417 Value *Offset = EmitGEPOffset(GEP);2418 Res = Builder.CreateAdd(Res, Offset, "", GEP->hasNoUnsignedWrap());2419 }2420 return Builder.CreateZExtOrTrunc(Res, IntTy);2421}2422 2423Instruction *InstCombinerImpl::visitPtrToInt(PtrToIntInst &CI) {2424 // If the destination integer type is not the intptr_t type for this target,2425 // do a ptrtoint to intptr_t then do a trunc or zext. This allows the cast2426 // to be exposed to other transforms.2427 Value *SrcOp = CI.getPointerOperand();2428 Type *SrcTy = SrcOp->getType();2429 Type *Ty = CI.getType();2430 unsigned AS = CI.getPointerAddressSpace();2431 unsigned TySize = Ty->getScalarSizeInBits();2432 unsigned PtrSize = DL.getPointerSizeInBits(AS);2433 if (TySize != PtrSize) {2434 Type *IntPtrTy =2435 SrcTy->getWithNewType(DL.getIntPtrType(CI.getContext(), AS));2436 Value *P = Builder.CreatePtrToInt(SrcOp, IntPtrTy);2437 return CastInst::CreateIntegerCast(P, Ty, /*isSigned=*/false);2438 }2439 2440 // (ptrtoint (ptrmask P, M))2441 // -> (and (ptrtoint P), M)2442 // This is generally beneficial as `and` is better supported than `ptrmask`.2443 Value *Ptr, *Mask;2444 if (match(SrcOp, m_OneUse(m_Intrinsic<Intrinsic::ptrmask>(m_Value(Ptr),2445 m_Value(Mask)))) &&2446 Mask->getType() == Ty)2447 return BinaryOperator::CreateAnd(Builder.CreatePtrToInt(Ptr, Ty), Mask);2448 2449 if (Value *V = foldPtrToIntOrAddrOfGEP(Ty, SrcOp))2450 return replaceInstUsesWith(CI, V);2451 2452 Value *Vec, *Scalar, *Index;2453 if (match(SrcOp, m_OneUse(m_InsertElt(m_IntToPtr(m_Value(Vec)),2454 m_Value(Scalar), m_Value(Index)))) &&2455 Vec->getType() == Ty) {2456 assert(Vec->getType()->getScalarSizeInBits() == PtrSize && "Wrong type");2457 // Convert the scalar to int followed by insert to eliminate one cast:2458 // p2i (ins (i2p Vec), Scalar, Index --> ins Vec, (p2i Scalar), Index2459 Value *NewCast = Builder.CreatePtrToInt(Scalar, Ty->getScalarType());2460 return InsertElementInst::Create(Vec, NewCast, Index);2461 }2462 2463 return commonCastTransforms(CI);2464}2465 2466Instruction *InstCombinerImpl::visitPtrToAddr(PtrToAddrInst &CI) {2467 Value *SrcOp = CI.getPointerOperand();2468 Type *Ty = CI.getType();2469 2470 // (ptrtoaddr (ptrmask P, M))2471 // -> (and (ptrtoaddr P), M)2472 // This is generally beneficial as `and` is better supported than `ptrmask`.2473 Value *Ptr, *Mask;2474 if (match(SrcOp, m_OneUse(m_Intrinsic<Intrinsic::ptrmask>(m_Value(Ptr),2475 m_Value(Mask)))) &&2476 Mask->getType() == Ty)2477 return BinaryOperator::CreateAnd(Builder.CreatePtrToAddr(Ptr), Mask);2478 2479 if (Value *V = foldPtrToIntOrAddrOfGEP(Ty, SrcOp))2480 return replaceInstUsesWith(CI, V);2481 2482 // FIXME: Implement variants of ptrtoint folds.2483 return commonCastTransforms(CI);2484}2485 2486/// This input value (which is known to have vector type) is being zero extended2487/// or truncated to the specified vector type. Since the zext/trunc is done2488/// using an integer type, we have a (bitcast(cast(bitcast))) pattern,2489/// endianness will impact which end of the vector that is extended or2490/// truncated.2491///2492/// A vector is always stored with index 0 at the lowest address, which2493/// corresponds to the most significant bits for a big endian stored integer and2494/// the least significant bits for little endian. A trunc/zext of an integer2495/// impacts the big end of the integer. Thus, we need to add/remove elements at2496/// the front of the vector for big endian targets, and the back of the vector2497/// for little endian targets.2498///2499/// Try to replace it with a shuffle (and vector/vector bitcast) if possible.2500///2501/// The source and destination vector types may have different element types.2502static Instruction *2503optimizeVectorResizeWithIntegerBitCasts(Value *InVal, VectorType *DestTy,2504 InstCombinerImpl &IC) {2505 // We can only do this optimization if the output is a multiple of the input2506 // element size, or the input is a multiple of the output element size.2507 // Convert the input type to have the same element type as the output.2508 VectorType *SrcTy = cast<VectorType>(InVal->getType());2509 2510 if (SrcTy->getElementType() != DestTy->getElementType()) {2511 // The input types don't need to be identical, but for now they must be the2512 // same size. There is no specific reason we couldn't handle things like2513 // <4 x i16> -> <4 x i32> by bitcasting to <2 x i32> but haven't gotten2514 // there yet.2515 if (SrcTy->getElementType()->getPrimitiveSizeInBits() !=2516 DestTy->getElementType()->getPrimitiveSizeInBits())2517 return nullptr;2518 2519 SrcTy =2520 FixedVectorType::get(DestTy->getElementType(),2521 cast<FixedVectorType>(SrcTy)->getNumElements());2522 InVal = IC.Builder.CreateBitCast(InVal, SrcTy);2523 }2524 2525 bool IsBigEndian = IC.getDataLayout().isBigEndian();2526 unsigned SrcElts = cast<FixedVectorType>(SrcTy)->getNumElements();2527 unsigned DestElts = cast<FixedVectorType>(DestTy)->getNumElements();2528 2529 assert(SrcElts != DestElts && "Element counts should be different.");2530 2531 // Now that the element types match, get the shuffle mask and RHS of the2532 // shuffle to use, which depends on whether we're increasing or decreasing the2533 // size of the input.2534 auto ShuffleMaskStorage = llvm::to_vector<16>(llvm::seq<int>(0, SrcElts));2535 ArrayRef<int> ShuffleMask;2536 Value *V2;2537 2538 if (SrcElts > DestElts) {2539 // If we're shrinking the number of elements (rewriting an integer2540 // truncate), just shuffle in the elements corresponding to the least2541 // significant bits from the input and use poison as the second shuffle2542 // input.2543 V2 = PoisonValue::get(SrcTy);2544 // Make sure the shuffle mask selects the "least significant bits" by2545 // keeping elements from back of the src vector for big endian, and from the2546 // front for little endian.2547 ShuffleMask = ShuffleMaskStorage;2548 if (IsBigEndian)2549 ShuffleMask = ShuffleMask.take_back(DestElts);2550 else2551 ShuffleMask = ShuffleMask.take_front(DestElts);2552 } else {2553 // If we're increasing the number of elements (rewriting an integer zext),2554 // shuffle in all of the elements from InVal. Fill the rest of the result2555 // elements with zeros from a constant zero.2556 V2 = Constant::getNullValue(SrcTy);2557 // Use first elt from V2 when indicating zero in the shuffle mask.2558 uint32_t NullElt = SrcElts;2559 // Extend with null values in the "most significant bits" by adding elements2560 // in front of the src vector for big endian, and at the back for little2561 // endian.2562 unsigned DeltaElts = DestElts - SrcElts;2563 if (IsBigEndian)2564 ShuffleMaskStorage.insert(ShuffleMaskStorage.begin(), DeltaElts, NullElt);2565 else2566 ShuffleMaskStorage.append(DeltaElts, NullElt);2567 ShuffleMask = ShuffleMaskStorage;2568 }2569 2570 return new ShuffleVectorInst(InVal, V2, ShuffleMask);2571}2572 2573static bool isMultipleOfTypeSize(unsigned Value, Type *Ty) {2574 return Value % Ty->getPrimitiveSizeInBits() == 0;2575}2576 2577static unsigned getTypeSizeIndex(unsigned Value, Type *Ty) {2578 return Value / Ty->getPrimitiveSizeInBits();2579}2580 2581/// V is a value which is inserted into a vector of VecEltTy.2582/// Look through the value to see if we can decompose it into2583/// insertions into the vector. See the example in the comment for2584/// OptimizeIntegerToVectorInsertions for the pattern this handles.2585/// The type of V is always a non-zero multiple of VecEltTy's size.2586/// Shift is the number of bits between the lsb of V and the lsb of2587/// the vector.2588///2589/// This returns false if the pattern can't be matched or true if it can,2590/// filling in Elements with the elements found here.2591static bool collectInsertionElements(Value *V, unsigned Shift,2592 SmallVectorImpl<Value *> &Elements,2593 Type *VecEltTy, bool isBigEndian) {2594 assert(isMultipleOfTypeSize(Shift, VecEltTy) &&2595 "Shift should be a multiple of the element type size");2596 2597 // Undef values never contribute useful bits to the result.2598 if (isa<UndefValue>(V)) return true;2599 2600 // If we got down to a value of the right type, we win, try inserting into the2601 // right element.2602 if (V->getType() == VecEltTy) {2603 // Inserting null doesn't actually insert any elements.2604 if (Constant *C = dyn_cast<Constant>(V))2605 if (C->isNullValue())2606 return true;2607 2608 unsigned ElementIndex = getTypeSizeIndex(Shift, VecEltTy);2609 if (isBigEndian)2610 ElementIndex = Elements.size() - ElementIndex - 1;2611 2612 // Fail if multiple elements are inserted into this slot.2613 if (Elements[ElementIndex])2614 return false;2615 2616 Elements[ElementIndex] = V;2617 return true;2618 }2619 2620 if (Constant *C = dyn_cast<Constant>(V)) {2621 // Figure out the # elements this provides, and bitcast it or slice it up2622 // as required.2623 unsigned NumElts = getTypeSizeIndex(C->getType()->getPrimitiveSizeInBits(),2624 VecEltTy);2625 // If the constant is the size of a vector element, we just need to bitcast2626 // it to the right type so it gets properly inserted.2627 if (NumElts == 1)2628 return collectInsertionElements(ConstantExpr::getBitCast(C, VecEltTy),2629 Shift, Elements, VecEltTy, isBigEndian);2630 2631 // Okay, this is a constant that covers multiple elements. Slice it up into2632 // pieces and insert each element-sized piece into the vector.2633 if (!isa<IntegerType>(C->getType()))2634 C = ConstantExpr::getBitCast(C, IntegerType::get(V->getContext(),2635 C->getType()->getPrimitiveSizeInBits()));2636 unsigned ElementSize = VecEltTy->getPrimitiveSizeInBits();2637 Type *ElementIntTy = IntegerType::get(C->getContext(), ElementSize);2638 2639 for (unsigned i = 0; i != NumElts; ++i) {2640 unsigned ShiftI = i * ElementSize;2641 Constant *Piece = ConstantFoldBinaryInstruction(2642 Instruction::LShr, C, ConstantInt::get(C->getType(), ShiftI));2643 if (!Piece)2644 return false;2645 2646 Piece = ConstantExpr::getTrunc(Piece, ElementIntTy);2647 if (!collectInsertionElements(Piece, ShiftI + Shift, Elements, VecEltTy,2648 isBigEndian))2649 return false;2650 }2651 return true;2652 }2653 2654 if (!V->hasOneUse()) return false;2655 2656 Instruction *I = dyn_cast<Instruction>(V);2657 if (!I) return false;2658 switch (I->getOpcode()) {2659 default: return false; // Unhandled case.2660 case Instruction::BitCast:2661 if (I->getOperand(0)->getType()->isVectorTy())2662 return false;2663 return collectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy,2664 isBigEndian);2665 case Instruction::ZExt:2666 if (!isMultipleOfTypeSize(2667 I->getOperand(0)->getType()->getPrimitiveSizeInBits(),2668 VecEltTy))2669 return false;2670 return collectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy,2671 isBigEndian);2672 case Instruction::Or:2673 return collectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy,2674 isBigEndian) &&2675 collectInsertionElements(I->getOperand(1), Shift, Elements, VecEltTy,2676 isBigEndian);2677 case Instruction::Shl: {2678 // Must be shifting by a constant that is a multiple of the element size.2679 ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1));2680 if (!CI) return false;2681 Shift += CI->getZExtValue();2682 if (!isMultipleOfTypeSize(Shift, VecEltTy)) return false;2683 return collectInsertionElements(I->getOperand(0), Shift, Elements, VecEltTy,2684 isBigEndian);2685 }2686 2687 }2688}2689 2690 2691/// If the input is an 'or' instruction, we may be doing shifts and ors to2692/// assemble the elements of the vector manually.2693/// Try to rip the code out and replace it with insertelements. This is to2694/// optimize code like this:2695///2696/// %tmp37 = bitcast float %inc to i322697/// %tmp38 = zext i32 %tmp37 to i642698/// %tmp31 = bitcast float %inc5 to i322699/// %tmp32 = zext i32 %tmp31 to i642700/// %tmp33 = shl i64 %tmp32, 322701/// %ins35 = or i64 %tmp33, %tmp382702/// %tmp43 = bitcast i64 %ins35 to <2 x float>2703///2704/// Into two insertelements that do "buildvector{%inc, %inc5}".2705static Value *optimizeIntegerToVectorInsertions(BitCastInst &CI,2706 InstCombinerImpl &IC) {2707 auto *DestVecTy = cast<FixedVectorType>(CI.getType());2708 Value *IntInput = CI.getOperand(0);2709 2710 // if the int input is just an undef value do not try to optimize to vector2711 // insertions as it will prevent undef propagation2712 if (isa<UndefValue>(IntInput))2713 return nullptr;2714 2715 SmallVector<Value*, 8> Elements(DestVecTy->getNumElements());2716 if (!collectInsertionElements(IntInput, 0, Elements,2717 DestVecTy->getElementType(),2718 IC.getDataLayout().isBigEndian()))2719 return nullptr;2720 2721 // If we succeeded, we know that all of the element are specified by Elements2722 // or are zero if Elements has a null entry. Recast this as a set of2723 // insertions.2724 Value *Result = Constant::getNullValue(CI.getType());2725 for (unsigned i = 0, e = Elements.size(); i != e; ++i) {2726 if (!Elements[i]) continue; // Unset element.2727 2728 Result = IC.Builder.CreateInsertElement(Result, Elements[i],2729 IC.Builder.getInt32(i));2730 }2731 2732 return Result;2733}2734 2735/// Canonicalize scalar bitcasts of extracted elements into a bitcast of the2736/// vector followed by extract element. The backend tends to handle bitcasts of2737/// vectors better than bitcasts of scalars because vector registers are2738/// usually not type-specific like scalar integer or scalar floating-point.2739static Instruction *canonicalizeBitCastExtElt(BitCastInst &BitCast,2740 InstCombinerImpl &IC) {2741 Value *VecOp, *Index;2742 if (!match(BitCast.getOperand(0),2743 m_OneUse(m_ExtractElt(m_Value(VecOp), m_Value(Index)))))2744 return nullptr;2745 2746 // The bitcast must be to a vectorizable type, otherwise we can't make a new2747 // type to extract from.2748 Type *DestType = BitCast.getType();2749 VectorType *VecType = cast<VectorType>(VecOp->getType());2750 if (VectorType::isValidElementType(DestType)) {2751 auto *NewVecType = VectorType::get(DestType, VecType);2752 auto *NewBC = IC.Builder.CreateBitCast(VecOp, NewVecType, "bc");2753 return ExtractElementInst::Create(NewBC, Index);2754 }2755 2756 // Only solve DestType is vector to avoid inverse transform in visitBitCast.2757 // bitcast (extractelement <1 x elt>, dest) -> bitcast(<1 x elt>, dest)2758 auto *FixedVType = dyn_cast<FixedVectorType>(VecType);2759 if (DestType->isVectorTy() && FixedVType && FixedVType->getNumElements() == 1)2760 return CastInst::Create(Instruction::BitCast, VecOp, DestType);2761 2762 return nullptr;2763}2764 2765/// Change the type of a bitwise logic operation if we can eliminate a bitcast.2766static Instruction *foldBitCastBitwiseLogic(BitCastInst &BitCast,2767 InstCombiner::BuilderTy &Builder) {2768 Type *DestTy = BitCast.getType();2769 BinaryOperator *BO;2770 2771 if (!match(BitCast.getOperand(0), m_OneUse(m_BinOp(BO))) ||2772 !BO->isBitwiseLogicOp())2773 return nullptr;2774 2775 // FIXME: This transform is restricted to vector types to avoid backend2776 // problems caused by creating potentially illegal operations. If a fix-up is2777 // added to handle that situation, we can remove this check.2778 if (!DestTy->isVectorTy() || !BO->getType()->isVectorTy())2779 return nullptr;2780 2781 if (DestTy->isFPOrFPVectorTy()) {2782 Value *X, *Y;2783 // bitcast(logic(bitcast(X), bitcast(Y))) -> bitcast'(logic(bitcast'(X), Y))2784 if (match(BO->getOperand(0), m_OneUse(m_BitCast(m_Value(X)))) &&2785 match(BO->getOperand(1), m_OneUse(m_BitCast(m_Value(Y))))) {2786 if (X->getType()->isFPOrFPVectorTy() &&2787 Y->getType()->isIntOrIntVectorTy()) {2788 Value *CastedOp =2789 Builder.CreateBitCast(BO->getOperand(0), Y->getType());2790 Value *NewBO = Builder.CreateBinOp(BO->getOpcode(), CastedOp, Y);2791 return CastInst::CreateBitOrPointerCast(NewBO, DestTy);2792 }2793 if (X->getType()->isIntOrIntVectorTy() &&2794 Y->getType()->isFPOrFPVectorTy()) {2795 Value *CastedOp =2796 Builder.CreateBitCast(BO->getOperand(1), X->getType());2797 Value *NewBO = Builder.CreateBinOp(BO->getOpcode(), CastedOp, X);2798 return CastInst::CreateBitOrPointerCast(NewBO, DestTy);2799 }2800 }2801 return nullptr;2802 }2803 2804 if (!DestTy->isIntOrIntVectorTy())2805 return nullptr;2806 2807 Value *X;2808 if (match(BO->getOperand(0), m_OneUse(m_BitCast(m_Value(X)))) &&2809 X->getType() == DestTy && !isa<Constant>(X)) {2810 // bitcast(logic(bitcast(X), Y)) --> logic'(X, bitcast(Y))2811 Value *CastedOp1 = Builder.CreateBitCast(BO->getOperand(1), DestTy);2812 return BinaryOperator::Create(BO->getOpcode(), X, CastedOp1);2813 }2814 2815 if (match(BO->getOperand(1), m_OneUse(m_BitCast(m_Value(X)))) &&2816 X->getType() == DestTy && !isa<Constant>(X)) {2817 // bitcast(logic(Y, bitcast(X))) --> logic'(bitcast(Y), X)2818 Value *CastedOp0 = Builder.CreateBitCast(BO->getOperand(0), DestTy);2819 return BinaryOperator::Create(BO->getOpcode(), CastedOp0, X);2820 }2821 2822 // Canonicalize vector bitcasts to come before vector bitwise logic with a2823 // constant. This eases recognition of special constants for later ops.2824 // Example:2825 // icmp u/s (a ^ signmask), (b ^ signmask) --> icmp s/u a, b2826 Constant *C;2827 if (match(BO->getOperand(1), m_Constant(C))) {2828 // bitcast (logic X, C) --> logic (bitcast X, C')2829 Value *CastedOp0 = Builder.CreateBitCast(BO->getOperand(0), DestTy);2830 Value *CastedC = Builder.CreateBitCast(C, DestTy);2831 return BinaryOperator::Create(BO->getOpcode(), CastedOp0, CastedC);2832 }2833 2834 return nullptr;2835}2836 2837/// Change the type of a select if we can eliminate a bitcast.2838static Instruction *foldBitCastSelect(BitCastInst &BitCast,2839 InstCombiner::BuilderTy &Builder) {2840 Value *Cond, *TVal, *FVal;2841 if (!match(BitCast.getOperand(0),2842 m_OneUse(m_Select(m_Value(Cond), m_Value(TVal), m_Value(FVal)))))2843 return nullptr;2844 2845 // A vector select must maintain the same number of elements in its operands.2846 Type *CondTy = Cond->getType();2847 Type *DestTy = BitCast.getType();2848 if (auto *CondVTy = dyn_cast<VectorType>(CondTy))2849 if (!DestTy->isVectorTy() ||2850 CondVTy->getElementCount() !=2851 cast<VectorType>(DestTy)->getElementCount())2852 return nullptr;2853 2854 // FIXME: This transform is restricted from changing the select between2855 // scalars and vectors to avoid backend problems caused by creating2856 // potentially illegal operations. If a fix-up is added to handle that2857 // situation, we can remove this check.2858 if (DestTy->isVectorTy() != TVal->getType()->isVectorTy())2859 return nullptr;2860 2861 auto *Sel = cast<Instruction>(BitCast.getOperand(0));2862 Value *X;2863 if (match(TVal, m_OneUse(m_BitCast(m_Value(X)))) && X->getType() == DestTy &&2864 !isa<Constant>(X)) {2865 // bitcast(select(Cond, bitcast(X), Y)) --> select'(Cond, X, bitcast(Y))2866 Value *CastedVal = Builder.CreateBitCast(FVal, DestTy);2867 return SelectInst::Create(Cond, X, CastedVal, "", nullptr, Sel);2868 }2869 2870 if (match(FVal, m_OneUse(m_BitCast(m_Value(X)))) && X->getType() == DestTy &&2871 !isa<Constant>(X)) {2872 // bitcast(select(Cond, Y, bitcast(X))) --> select'(Cond, bitcast(Y), X)2873 Value *CastedVal = Builder.CreateBitCast(TVal, DestTy);2874 return SelectInst::Create(Cond, CastedVal, X, "", nullptr, Sel);2875 }2876 2877 return nullptr;2878}2879 2880/// Check if all users of CI are StoreInsts.2881static bool hasStoreUsersOnly(CastInst &CI) {2882 for (User *U : CI.users()) {2883 if (!isa<StoreInst>(U))2884 return false;2885 }2886 return true;2887}2888 2889/// This function handles following case2890///2891/// A -> B cast2892/// PHI2893/// B -> A cast2894///2895/// All the related PHI nodes can be replaced by new PHI nodes with type A.2896/// The uses of \p CI can be changed to the new PHI node corresponding to \p PN.2897Instruction *InstCombinerImpl::optimizeBitCastFromPhi(CastInst &CI,2898 PHINode *PN) {2899 // BitCast used by Store can be handled in InstCombineLoadStoreAlloca.cpp.2900 if (hasStoreUsersOnly(CI))2901 return nullptr;2902 2903 Value *Src = CI.getOperand(0);2904 Type *SrcTy = Src->getType(); // Type B2905 Type *DestTy = CI.getType(); // Type A2906 2907 SmallVector<PHINode *, 4> PhiWorklist;2908 SmallSetVector<PHINode *, 4> OldPhiNodes;2909 2910 // Find all of the A->B casts and PHI nodes.2911 // We need to inspect all related PHI nodes, but PHIs can be cyclic, so2912 // OldPhiNodes is used to track all known PHI nodes, before adding a new2913 // PHI to PhiWorklist, it is checked against and added to OldPhiNodes first.2914 PhiWorklist.push_back(PN);2915 OldPhiNodes.insert(PN);2916 while (!PhiWorklist.empty()) {2917 auto *OldPN = PhiWorklist.pop_back_val();2918 for (Value *IncValue : OldPN->incoming_values()) {2919 if (isa<Constant>(IncValue))2920 continue;2921 2922 if (auto *LI = dyn_cast<LoadInst>(IncValue)) {2923 // If there is a sequence of one or more load instructions, each loaded2924 // value is used as address of later load instruction, bitcast is2925 // necessary to change the value type, don't optimize it. For2926 // simplicity we give up if the load address comes from another load.2927 Value *Addr = LI->getOperand(0);2928 if (Addr == &CI || isa<LoadInst>(Addr))2929 return nullptr;2930 // Don't tranform "load <256 x i32>, <256 x i32>*" to2931 // "load x86_amx, x86_amx*", because x86_amx* is invalid.2932 // TODO: Remove this check when bitcast between vector and x86_amx2933 // is replaced with a specific intrinsic.2934 if (DestTy->isX86_AMXTy())2935 return nullptr;2936 if (LI->hasOneUse() && LI->isSimple())2937 continue;2938 // If a LoadInst has more than one use, changing the type of loaded2939 // value may create another bitcast.2940 return nullptr;2941 }2942 2943 if (auto *PNode = dyn_cast<PHINode>(IncValue)) {2944 if (OldPhiNodes.insert(PNode))2945 PhiWorklist.push_back(PNode);2946 continue;2947 }2948 2949 auto *BCI = dyn_cast<BitCastInst>(IncValue);2950 // We can't handle other instructions.2951 if (!BCI)2952 return nullptr;2953 2954 // Verify it's a A->B cast.2955 Type *TyA = BCI->getOperand(0)->getType();2956 Type *TyB = BCI->getType();2957 if (TyA != DestTy || TyB != SrcTy)2958 return nullptr;2959 }2960 }2961 2962 // Check that each user of each old PHI node is something that we can2963 // rewrite, so that all of the old PHI nodes can be cleaned up afterwards.2964 for (auto *OldPN : OldPhiNodes) {2965 for (User *V : OldPN->users()) {2966 if (auto *SI = dyn_cast<StoreInst>(V)) {2967 if (!SI->isSimple() || SI->getOperand(0) != OldPN)2968 return nullptr;2969 } else if (auto *BCI = dyn_cast<BitCastInst>(V)) {2970 // Verify it's a B->A cast.2971 Type *TyB = BCI->getOperand(0)->getType();2972 Type *TyA = BCI->getType();2973 if (TyA != DestTy || TyB != SrcTy)2974 return nullptr;2975 } else if (auto *PHI = dyn_cast<PHINode>(V)) {2976 // As long as the user is another old PHI node, then even if we don't2977 // rewrite it, the PHI web we're considering won't have any users2978 // outside itself, so it'll be dead.2979 if (!OldPhiNodes.contains(PHI))2980 return nullptr;2981 } else {2982 return nullptr;2983 }2984 }2985 }2986 2987 // For each old PHI node, create a corresponding new PHI node with a type A.2988 SmallDenseMap<PHINode *, PHINode *> NewPNodes;2989 for (auto *OldPN : OldPhiNodes) {2990 Builder.SetInsertPoint(OldPN);2991 PHINode *NewPN = Builder.CreatePHI(DestTy, OldPN->getNumOperands());2992 NewPNodes[OldPN] = NewPN;2993 }2994 2995 // Fill in the operands of new PHI nodes.2996 for (auto *OldPN : OldPhiNodes) {2997 PHINode *NewPN = NewPNodes[OldPN];2998 for (unsigned j = 0, e = OldPN->getNumOperands(); j != e; ++j) {2999 Value *V = OldPN->getOperand(j);3000 Value *NewV = nullptr;3001 if (auto *C = dyn_cast<Constant>(V)) {3002 NewV = ConstantExpr::getBitCast(C, DestTy);3003 } else if (auto *LI = dyn_cast<LoadInst>(V)) {3004 // Explicitly perform load combine to make sure no opposing transform3005 // can remove the bitcast in the meantime and trigger an infinite loop.3006 Builder.SetInsertPoint(LI);3007 NewV = combineLoadToNewType(*LI, DestTy);3008 // Remove the old load and its use in the old phi, which itself becomes3009 // dead once the whole transform finishes.3010 replaceInstUsesWith(*LI, PoisonValue::get(LI->getType()));3011 eraseInstFromFunction(*LI);3012 } else if (auto *BCI = dyn_cast<BitCastInst>(V)) {3013 NewV = BCI->getOperand(0);3014 } else if (auto *PrevPN = dyn_cast<PHINode>(V)) {3015 NewV = NewPNodes[PrevPN];3016 }3017 assert(NewV);3018 NewPN->addIncoming(NewV, OldPN->getIncomingBlock(j));3019 }3020 }3021 3022 // Traverse all accumulated PHI nodes and process its users,3023 // which are Stores and BitcCasts. Without this processing3024 // NewPHI nodes could be replicated and could lead to extra3025 // moves generated after DeSSA.3026 // If there is a store with type B, change it to type A.3027 3028 3029 // Replace users of BitCast B->A with NewPHI. These will help3030 // later to get rid off a closure formed by OldPHI nodes.3031 Instruction *RetVal = nullptr;3032 for (auto *OldPN : OldPhiNodes) {3033 PHINode *NewPN = NewPNodes[OldPN];3034 for (User *V : make_early_inc_range(OldPN->users())) {3035 if (auto *SI = dyn_cast<StoreInst>(V)) {3036 assert(SI->isSimple() && SI->getOperand(0) == OldPN);3037 Builder.SetInsertPoint(SI);3038 auto *NewBC =3039 cast<BitCastInst>(Builder.CreateBitCast(NewPN, SrcTy));3040 SI->setOperand(0, NewBC);3041 Worklist.push(SI);3042 assert(hasStoreUsersOnly(*NewBC));3043 }3044 else if (auto *BCI = dyn_cast<BitCastInst>(V)) {3045 Type *TyB = BCI->getOperand(0)->getType();3046 Type *TyA = BCI->getType();3047 assert(TyA == DestTy && TyB == SrcTy);3048 (void) TyA;3049 (void) TyB;3050 Instruction *I = replaceInstUsesWith(*BCI, NewPN);3051 if (BCI == &CI)3052 RetVal = I;3053 } else if (auto *PHI = dyn_cast<PHINode>(V)) {3054 assert(OldPhiNodes.contains(PHI));3055 (void) PHI;3056 } else {3057 llvm_unreachable("all uses should be handled");3058 }3059 }3060 }3061 3062 return RetVal;3063}3064 3065/// Fold (bitcast (or (and (bitcast X to int), signmask), nneg Y) to fp) to3066/// copysign((bitcast Y to fp), X)3067static Value *foldCopySignIdioms(BitCastInst &CI,3068 InstCombiner::BuilderTy &Builder,3069 const SimplifyQuery &SQ) {3070 Value *X, *Y;3071 Type *FTy = CI.getType();3072 if (!FTy->isFPOrFPVectorTy())3073 return nullptr;3074 if (!match(&CI, m_ElementWiseBitCast(m_c_Or(3075 m_And(m_ElementWiseBitCast(m_Value(X)), m_SignMask()),3076 m_Value(Y)))))3077 return nullptr;3078 if (X->getType() != FTy)3079 return nullptr;3080 if (!isKnownNonNegative(Y, SQ))3081 return nullptr;3082 3083 return Builder.CreateCopySign(Builder.CreateBitCast(Y, FTy), X);3084}3085 3086Instruction *InstCombinerImpl::visitBitCast(BitCastInst &CI) {3087 // If the operands are integer typed then apply the integer transforms,3088 // otherwise just apply the common ones.3089 Value *Src = CI.getOperand(0);3090 Type *SrcTy = Src->getType();3091 Type *DestTy = CI.getType();3092 3093 // Get rid of casts from one type to the same type. These are useless and can3094 // be replaced by the operand.3095 if (DestTy == Src->getType())3096 return replaceInstUsesWith(CI, Src);3097 3098 if (isa<FixedVectorType>(DestTy)) {3099 if (isa<IntegerType>(SrcTy)) {3100 // If this is a cast from an integer to vector, check to see if the input3101 // is a trunc or zext of a bitcast from vector. If so, we can replace all3102 // the casts with a shuffle and (potentially) a bitcast.3103 if (isa<TruncInst>(Src) || isa<ZExtInst>(Src)) {3104 CastInst *SrcCast = cast<CastInst>(Src);3105 if (BitCastInst *BCIn = dyn_cast<BitCastInst>(SrcCast->getOperand(0)))3106 if (isa<VectorType>(BCIn->getOperand(0)->getType()))3107 if (Instruction *I = optimizeVectorResizeWithIntegerBitCasts(3108 BCIn->getOperand(0), cast<VectorType>(DestTy), *this))3109 return I;3110 }3111 3112 // If the input is an 'or' instruction, we may be doing shifts and ors to3113 // assemble the elements of the vector manually. Try to rip the code out3114 // and replace it with insertelements.3115 if (Value *V = optimizeIntegerToVectorInsertions(CI, *this))3116 return replaceInstUsesWith(CI, V);3117 }3118 }3119 3120 if (FixedVectorType *SrcVTy = dyn_cast<FixedVectorType>(SrcTy)) {3121 if (SrcVTy->getNumElements() == 1) {3122 // If our destination is not a vector, then make this a straight3123 // scalar-scalar cast.3124 if (!DestTy->isVectorTy()) {3125 Value *Elem =3126 Builder.CreateExtractElement(Src,3127 Constant::getNullValue(Type::getInt32Ty(CI.getContext())));3128 return CastInst::Create(Instruction::BitCast, Elem, DestTy);3129 }3130 3131 // Otherwise, see if our source is an insert. If so, then use the scalar3132 // component directly:3133 // bitcast (inselt <1 x elt> V, X, 0) to <n x m> --> bitcast X to <n x m>3134 if (auto *InsElt = dyn_cast<InsertElementInst>(Src))3135 return new BitCastInst(InsElt->getOperand(1), DestTy);3136 }3137 3138 // Convert an artificial vector insert into more analyzable bitwise logic.3139 unsigned BitWidth = DestTy->getScalarSizeInBits();3140 Value *X, *Y;3141 uint64_t IndexC;3142 if (match(Src, m_OneUse(m_InsertElt(m_OneUse(m_BitCast(m_Value(X))),3143 m_Value(Y), m_ConstantInt(IndexC)))) &&3144 DestTy->isIntegerTy() && X->getType() == DestTy &&3145 Y->getType()->isIntegerTy() && isDesirableIntType(BitWidth)) {3146 // Adjust for big endian - the LSBs are at the high index.3147 if (DL.isBigEndian())3148 IndexC = SrcVTy->getNumElements() - 1 - IndexC;3149 3150 // We only handle (endian-normalized) insert to index 0. Any other insert3151 // would require a left-shift, so that is an extra instruction.3152 if (IndexC == 0) {3153 // bitcast (inselt (bitcast X), Y, 0) --> or (and X, MaskC), (zext Y)3154 unsigned EltWidth = Y->getType()->getScalarSizeInBits();3155 APInt MaskC = APInt::getHighBitsSet(BitWidth, BitWidth - EltWidth);3156 Value *AndX = Builder.CreateAnd(X, MaskC);3157 Value *ZextY = Builder.CreateZExt(Y, DestTy);3158 return BinaryOperator::CreateOr(AndX, ZextY);3159 }3160 }3161 }3162 3163 if (auto *Shuf = dyn_cast<ShuffleVectorInst>(Src)) {3164 // Okay, we have (bitcast (shuffle ..)). Check to see if this is3165 // a bitcast to a vector with the same # elts.3166 Value *ShufOp0 = Shuf->getOperand(0);3167 Value *ShufOp1 = Shuf->getOperand(1);3168 auto ShufElts = cast<VectorType>(Shuf->getType())->getElementCount();3169 auto SrcVecElts = cast<VectorType>(ShufOp0->getType())->getElementCount();3170 if (Shuf->hasOneUse() && DestTy->isVectorTy() &&3171 cast<VectorType>(DestTy)->getElementCount() == ShufElts &&3172 ShufElts == SrcVecElts) {3173 BitCastInst *Tmp;3174 // If either of the operands is a cast from CI.getType(), then3175 // evaluating the shuffle in the casted destination's type will allow3176 // us to eliminate at least one cast.3177 if (((Tmp = dyn_cast<BitCastInst>(ShufOp0)) &&3178 Tmp->getOperand(0)->getType() == DestTy) ||3179 ((Tmp = dyn_cast<BitCastInst>(ShufOp1)) &&3180 Tmp->getOperand(0)->getType() == DestTy)) {3181 Value *LHS = Builder.CreateBitCast(ShufOp0, DestTy);3182 Value *RHS = Builder.CreateBitCast(ShufOp1, DestTy);3183 // Return a new shuffle vector. Use the same element ID's, as we3184 // know the vector types match #elts.3185 return new ShuffleVectorInst(LHS, RHS, Shuf->getShuffleMask());3186 }3187 }3188 3189 // A bitcasted-to-scalar and byte/bit reversing shuffle is better recognized3190 // as a byte/bit swap:3191 // bitcast <N x i8> (shuf X, undef, <N, N-1,...0>) -> bswap (bitcast X)3192 // bitcast <N x i1> (shuf X, undef, <N, N-1,...0>) -> bitreverse (bitcast X)3193 if (DestTy->isIntegerTy() && ShufElts.getKnownMinValue() % 2 == 0 &&3194 Shuf->hasOneUse() && Shuf->isReverse()) {3195 unsigned IntrinsicNum = 0;3196 if (DL.isLegalInteger(DestTy->getScalarSizeInBits()) &&3197 SrcTy->getScalarSizeInBits() == 8) {3198 IntrinsicNum = Intrinsic::bswap;3199 } else if (SrcTy->getScalarSizeInBits() == 1) {3200 IntrinsicNum = Intrinsic::bitreverse;3201 }3202 if (IntrinsicNum != 0) {3203 assert(ShufOp0->getType() == SrcTy && "Unexpected shuffle mask");3204 assert(match(ShufOp1, m_Undef()) && "Unexpected shuffle op");3205 Function *BswapOrBitreverse = Intrinsic::getOrInsertDeclaration(3206 CI.getModule(), IntrinsicNum, DestTy);3207 Value *ScalarX = Builder.CreateBitCast(ShufOp0, DestTy);3208 return CallInst::Create(BswapOrBitreverse, {ScalarX});3209 }3210 }3211 }3212 3213 // Handle the A->B->A cast, and there is an intervening PHI node.3214 if (PHINode *PN = dyn_cast<PHINode>(Src))3215 if (Instruction *I = optimizeBitCastFromPhi(CI, PN))3216 return I;3217 3218 if (Instruction *I = canonicalizeBitCastExtElt(CI, *this))3219 return I;3220 3221 if (Instruction *I = foldBitCastBitwiseLogic(CI, Builder))3222 return I;3223 3224 if (Instruction *I = foldBitCastSelect(CI, Builder))3225 return I;3226 3227 if (Value *V = foldCopySignIdioms(CI, Builder, SQ.getWithInstruction(&CI)))3228 return replaceInstUsesWith(CI, V);3229 3230 return commonCastTransforms(CI);3231}3232 3233Instruction *InstCombinerImpl::visitAddrSpaceCast(AddrSpaceCastInst &CI) {3234 return commonCastTransforms(CI);3235}3236