1649 lines · cpp
1//===- InstCombinePHI.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 visitPHINode function.10//11//===----------------------------------------------------------------------===//12 13#include "InstCombineInternal.h"14#include "llvm/ADT/STLExtras.h"15#include "llvm/ADT/SmallPtrSet.h"16#include "llvm/ADT/Statistic.h"17#include "llvm/Analysis/InstructionSimplify.h"18#include "llvm/Analysis/ValueTracking.h"19#include "llvm/IR/PatternMatch.h"20#include "llvm/Support/CommandLine.h"21#include "llvm/Transforms/InstCombine/InstCombiner.h"22#include "llvm/Transforms/Utils/Local.h"23#include <optional>24 25using namespace llvm;26using namespace llvm::PatternMatch;27 28#define DEBUG_TYPE "instcombine"29 30static cl::opt<unsigned>31MaxNumPhis("instcombine-max-num-phis", cl::init(512),32 cl::desc("Maximum number phis to handle in intptr/ptrint folding"));33 34STATISTIC(NumPHIsOfInsertValues,35 "Number of phi-of-insertvalue turned into insertvalue-of-phis");36STATISTIC(NumPHIsOfExtractValues,37 "Number of phi-of-extractvalue turned into extractvalue-of-phi");38STATISTIC(NumPHICSEs, "Number of PHI's that got CSE'd");39 40/// The PHI arguments will be folded into a single operation with a PHI node41/// as input. The debug location of the single operation will be the merged42/// locations of the original PHI node arguments.43void InstCombinerImpl::PHIArgMergedDebugLoc(Instruction *Inst, PHINode &PN) {44 auto *FirstInst = cast<Instruction>(PN.getIncomingValue(0));45 Inst->setDebugLoc(FirstInst->getDebugLoc());46 // We do not expect a CallInst here, otherwise, N-way merging of DebugLoc47 // will be inefficient.48 assert(!isa<CallInst>(Inst));49 50 for (Value *V : drop_begin(PN.incoming_values())) {51 auto *I = cast<Instruction>(V);52 Inst->applyMergedLocation(Inst->getDebugLoc(), I->getDebugLoc());53 }54}55 56/// If the phi is within a phi web, which is formed by the def-use chain57/// of phis and all the phis in the web are only used in the other phis.58/// In this case, these phis are dead and we will remove all of them.59bool InstCombinerImpl::foldDeadPhiWeb(PHINode &PN) {60 SmallVector<PHINode *, 16> Stack;61 SmallPtrSet<PHINode *, 16> Visited;62 Stack.push_back(&PN);63 Visited.insert(&PN);64 while (!Stack.empty()) {65 PHINode *Phi = Stack.pop_back_val();66 for (User *Use : Phi->users()) {67 if (PHINode *PhiUse = dyn_cast<PHINode>(Use)) {68 if (!Visited.insert(PhiUse).second)69 continue;70 // Early stop if the set of PHIs is large71 if (Visited.size() >= 16)72 return false;73 Stack.push_back(PhiUse);74 } else75 return false;76 }77 }78 for (PHINode *Phi : Visited)79 replaceInstUsesWith(*Phi, PoisonValue::get(Phi->getType()));80 for (PHINode *Phi : Visited)81 eraseInstFromFunction(*Phi);82 return true;83}84 85// Replace Integer typed PHI PN if the PHI's value is used as a pointer value.86// If there is an existing pointer typed PHI that produces the same value as PN,87// replace PN and the IntToPtr operation with it. Otherwise, synthesize a new88// PHI node:89//90// Case-1:91// bb1:92// int_init = PtrToInt(ptr_init)93// br label %bb294// bb2:95// int_val = PHI([int_init, %bb1], [int_val_inc, %bb2]96// ptr_val = PHI([ptr_init, %bb1], [ptr_val_inc, %bb2]97// ptr_val2 = IntToPtr(int_val)98// ...99// use(ptr_val2)100// ptr_val_inc = ...101// inc_val_inc = PtrToInt(ptr_val_inc)102//103// ==>104// bb1:105// br label %bb2106// bb2:107// ptr_val = PHI([ptr_init, %bb1], [ptr_val_inc, %bb2]108// ...109// use(ptr_val)110// ptr_val_inc = ...111//112// Case-2:113// bb1:114// int_ptr = BitCast(ptr_ptr)115// int_init = Load(int_ptr)116// br label %bb2117// bb2:118// int_val = PHI([int_init, %bb1], [int_val_inc, %bb2]119// ptr_val2 = IntToPtr(int_val)120// ...121// use(ptr_val2)122// ptr_val_inc = ...123// inc_val_inc = PtrToInt(ptr_val_inc)124// ==>125// bb1:126// ptr_init = Load(ptr_ptr)127// br label %bb2128// bb2:129// ptr_val = PHI([ptr_init, %bb1], [ptr_val_inc, %bb2]130// ...131// use(ptr_val)132// ptr_val_inc = ...133// ...134//135bool InstCombinerImpl::foldIntegerTypedPHI(PHINode &PN) {136 if (!PN.getType()->isIntegerTy())137 return false;138 if (!PN.hasOneUse())139 return false;140 141 auto *IntToPtr = dyn_cast<IntToPtrInst>(PN.user_back());142 if (!IntToPtr)143 return false;144 145 // Check if the pointer is actually used as pointer:146 auto HasPointerUse = [](Instruction *IIP) {147 for (User *U : IIP->users()) {148 Value *Ptr = nullptr;149 if (LoadInst *LoadI = dyn_cast<LoadInst>(U)) {150 Ptr = LoadI->getPointerOperand();151 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {152 Ptr = SI->getPointerOperand();153 } else if (GetElementPtrInst *GI = dyn_cast<GetElementPtrInst>(U)) {154 Ptr = GI->getPointerOperand();155 }156 157 if (Ptr && Ptr == IIP)158 return true;159 }160 return false;161 };162 163 if (!HasPointerUse(IntToPtr))164 return false;165 166 if (DL.getPointerSizeInBits(IntToPtr->getAddressSpace()) !=167 DL.getTypeSizeInBits(IntToPtr->getOperand(0)->getType()))168 return false;169 170 SmallVector<Value *, 4> AvailablePtrVals;171 for (auto Incoming : zip(PN.blocks(), PN.incoming_values())) {172 BasicBlock *BB = std::get<0>(Incoming);173 Value *Arg = std::get<1>(Incoming);174 175 // Arg could be a constant, constant expr, etc., which we don't cover here.176 if (!isa<Instruction>(Arg) && !isa<Argument>(Arg))177 return false;178 179 // First look backward:180 if (auto *PI = dyn_cast<PtrToIntInst>(Arg)) {181 AvailablePtrVals.emplace_back(PI->getOperand(0));182 continue;183 }184 185 // Next look forward:186 Value *ArgIntToPtr = nullptr;187 for (User *U : Arg->users()) {188 if (isa<IntToPtrInst>(U) && U->getType() == IntToPtr->getType() &&189 (DT.dominates(cast<Instruction>(U), BB) ||190 cast<Instruction>(U)->getParent() == BB)) {191 ArgIntToPtr = U;192 break;193 }194 }195 196 if (ArgIntToPtr) {197 AvailablePtrVals.emplace_back(ArgIntToPtr);198 continue;199 }200 201 // If Arg is defined by a PHI, allow it. This will also create202 // more opportunities iteratively.203 if (isa<PHINode>(Arg)) {204 AvailablePtrVals.emplace_back(Arg);205 continue;206 }207 208 // For a single use integer load:209 auto *LoadI = dyn_cast<LoadInst>(Arg);210 if (!LoadI)211 return false;212 213 if (!LoadI->hasOneUse())214 return false;215 216 // Push the integer typed Load instruction into the available217 // value set, and fix it up later when the pointer typed PHI218 // is synthesized.219 AvailablePtrVals.emplace_back(LoadI);220 }221 222 // Now search for a matching PHI223 auto *BB = PN.getParent();224 assert(AvailablePtrVals.size() == PN.getNumIncomingValues() &&225 "Not enough available ptr typed incoming values");226 PHINode *MatchingPtrPHI = nullptr;227 unsigned NumPhis = 0;228 for (PHINode &PtrPHI : BB->phis()) {229 // FIXME: consider handling this in AggressiveInstCombine230 if (NumPhis++ > MaxNumPhis)231 return false;232 if (&PtrPHI == &PN || PtrPHI.getType() != IntToPtr->getType())233 continue;234 if (any_of(zip(PN.blocks(), AvailablePtrVals),235 [&](const auto &BlockAndValue) {236 BasicBlock *BB = std::get<0>(BlockAndValue);237 Value *V = std::get<1>(BlockAndValue);238 return PtrPHI.getIncomingValueForBlock(BB) != V;239 }))240 continue;241 MatchingPtrPHI = &PtrPHI;242 break;243 }244 245 if (MatchingPtrPHI) {246 assert(MatchingPtrPHI->getType() == IntToPtr->getType() &&247 "Phi's Type does not match with IntToPtr");248 // Explicitly replace the inttoptr (rather than inserting a ptrtoint) here,249 // to make sure another transform can't undo it in the meantime.250 replaceInstUsesWith(*IntToPtr, MatchingPtrPHI);251 eraseInstFromFunction(*IntToPtr);252 eraseInstFromFunction(PN);253 return true;254 }255 256 // If it requires a conversion for every PHI operand, do not do it.257 if (all_of(AvailablePtrVals, [&](Value *V) {258 return (V->getType() != IntToPtr->getType()) || isa<IntToPtrInst>(V);259 }))260 return false;261 262 // If any of the operand that requires casting is a terminator263 // instruction, do not do it. Similarly, do not do the transform if the value264 // is PHI in a block with no insertion point, for example, a catchswitch265 // block, since we will not be able to insert a cast after the PHI.266 if (any_of(AvailablePtrVals, [&](Value *V) {267 if (V->getType() == IntToPtr->getType())268 return false;269 auto *Inst = dyn_cast<Instruction>(V);270 if (!Inst)271 return false;272 if (Inst->isTerminator())273 return true;274 auto *BB = Inst->getParent();275 if (isa<PHINode>(Inst) && BB->getFirstInsertionPt() == BB->end())276 return true;277 return false;278 }))279 return false;280 281 PHINode *NewPtrPHI = PHINode::Create(282 IntToPtr->getType(), PN.getNumIncomingValues(), PN.getName() + ".ptr");283 284 InsertNewInstBefore(NewPtrPHI, PN.getIterator());285 SmallDenseMap<Value *, Instruction *> Casts;286 for (auto Incoming : zip(PN.blocks(), AvailablePtrVals)) {287 auto *IncomingBB = std::get<0>(Incoming);288 auto *IncomingVal = std::get<1>(Incoming);289 290 if (IncomingVal->getType() == IntToPtr->getType()) {291 NewPtrPHI->addIncoming(IncomingVal, IncomingBB);292 continue;293 }294 295#ifndef NDEBUG296 LoadInst *LoadI = dyn_cast<LoadInst>(IncomingVal);297 assert((isa<PHINode>(IncomingVal) ||298 IncomingVal->getType()->isPointerTy() ||299 (LoadI && LoadI->hasOneUse())) &&300 "Can not replace LoadInst with multiple uses");301#endif302 // Need to insert a BitCast.303 // For an integer Load instruction with a single use, the load + IntToPtr304 // cast will be simplified into a pointer load:305 // %v = load i64, i64* %a.ip, align 8306 // %v.cast = inttoptr i64 %v to float **307 // ==>308 // %v.ptrp = bitcast i64 * %a.ip to float **309 // %v.cast = load float *, float ** %v.ptrp, align 8310 Instruction *&CI = Casts[IncomingVal];311 if (!CI) {312 CI = CastInst::CreateBitOrPointerCast(IncomingVal, IntToPtr->getType(),313 IncomingVal->getName() + ".ptr");314 if (auto *IncomingI = dyn_cast<Instruction>(IncomingVal)) {315 BasicBlock::iterator InsertPos(IncomingI);316 InsertPos++;317 BasicBlock *BB = IncomingI->getParent();318 if (isa<PHINode>(IncomingI))319 InsertPos = BB->getFirstInsertionPt();320 assert(InsertPos != BB->end() && "should have checked above");321 InsertNewInstBefore(CI, InsertPos);322 } else {323 auto *InsertBB = &IncomingBB->getParent()->getEntryBlock();324 InsertNewInstBefore(CI, InsertBB->getFirstInsertionPt());325 }326 }327 NewPtrPHI->addIncoming(CI, IncomingBB);328 }329 330 // Explicitly replace the inttoptr (rather than inserting a ptrtoint) here,331 // to make sure another transform can't undo it in the meantime.332 replaceInstUsesWith(*IntToPtr, NewPtrPHI);333 eraseInstFromFunction(*IntToPtr);334 eraseInstFromFunction(PN);335 return true;336}337 338// Remove RoundTrip IntToPtr/PtrToInt Cast on PHI-Operand and339// fold Phi-operand to bitcast.340Instruction *InstCombinerImpl::foldPHIArgIntToPtrToPHI(PHINode &PN) {341 // convert ptr2int ( phi[ int2ptr(ptr2int(x))] ) --> ptr2int ( phi [ x ] )342 // Make sure all uses of phi are ptr2int.343 if (!all_of(PN.users(), IsaPred<PtrToIntInst>))344 return nullptr;345 346 // Iterating over all operands to check presence of target pointers for347 // optimization.348 bool OperandWithRoundTripCast = false;349 for (unsigned OpNum = 0; OpNum != PN.getNumIncomingValues(); ++OpNum) {350 if (auto *NewOp =351 simplifyIntToPtrRoundTripCast(PN.getIncomingValue(OpNum))) {352 replaceOperand(PN, OpNum, NewOp);353 OperandWithRoundTripCast = true;354 }355 }356 if (!OperandWithRoundTripCast)357 return nullptr;358 return &PN;359}360 361/// If we have something like phi [insertvalue(a,b,0), insertvalue(c,d,0)],362/// turn this into a phi[a,c] and phi[b,d] and a single insertvalue.363Instruction *364InstCombinerImpl::foldPHIArgInsertValueInstructionIntoPHI(PHINode &PN) {365 auto *FirstIVI = cast<InsertValueInst>(PN.getIncomingValue(0));366 367 // Scan to see if all operands are `insertvalue`'s with the same indices,368 // and all have a single use.369 for (Value *V : drop_begin(PN.incoming_values())) {370 auto *I = dyn_cast<InsertValueInst>(V);371 if (!I || !I->hasOneUser() || I->getIndices() != FirstIVI->getIndices())372 return nullptr;373 }374 375 // For each operand of an `insertvalue`376 std::array<PHINode *, 2> NewOperands;377 for (int OpIdx : {0, 1}) {378 auto *&NewOperand = NewOperands[OpIdx];379 // Create a new PHI node to receive the values the operand has in each380 // incoming basic block.381 NewOperand = PHINode::Create(382 FirstIVI->getOperand(OpIdx)->getType(), PN.getNumIncomingValues(),383 FirstIVI->getOperand(OpIdx)->getName() + ".pn");384 // And populate each operand's PHI with said values.385 for (auto Incoming : zip(PN.blocks(), PN.incoming_values()))386 NewOperand->addIncoming(387 cast<InsertValueInst>(std::get<1>(Incoming))->getOperand(OpIdx),388 std::get<0>(Incoming));389 InsertNewInstBefore(NewOperand, PN.getIterator());390 }391 392 // And finally, create `insertvalue` over the newly-formed PHI nodes.393 auto *NewIVI = InsertValueInst::Create(NewOperands[0], NewOperands[1],394 FirstIVI->getIndices(), PN.getName());395 396 PHIArgMergedDebugLoc(NewIVI, PN);397 ++NumPHIsOfInsertValues;398 return NewIVI;399}400 401/// If we have something like phi [extractvalue(a,0), extractvalue(b,0)],402/// turn this into a phi[a,b] and a single extractvalue.403Instruction *404InstCombinerImpl::foldPHIArgExtractValueInstructionIntoPHI(PHINode &PN) {405 auto *FirstEVI = cast<ExtractValueInst>(PN.getIncomingValue(0));406 407 // Scan to see if all operands are `extractvalue`'s with the same indices,408 // and all have a single use.409 for (Value *V : drop_begin(PN.incoming_values())) {410 auto *I = dyn_cast<ExtractValueInst>(V);411 if (!I || !I->hasOneUser() || I->getIndices() != FirstEVI->getIndices() ||412 I->getAggregateOperand()->getType() !=413 FirstEVI->getAggregateOperand()->getType())414 return nullptr;415 }416 417 // Create a new PHI node to receive the values the aggregate operand has418 // in each incoming basic block.419 auto *NewAggregateOperand = PHINode::Create(420 FirstEVI->getAggregateOperand()->getType(), PN.getNumIncomingValues(),421 FirstEVI->getAggregateOperand()->getName() + ".pn");422 // And populate the PHI with said values.423 for (auto Incoming : zip(PN.blocks(), PN.incoming_values()))424 NewAggregateOperand->addIncoming(425 cast<ExtractValueInst>(std::get<1>(Incoming))->getAggregateOperand(),426 std::get<0>(Incoming));427 InsertNewInstBefore(NewAggregateOperand, PN.getIterator());428 429 // And finally, create `extractvalue` over the newly-formed PHI nodes.430 auto *NewEVI = ExtractValueInst::Create(NewAggregateOperand,431 FirstEVI->getIndices(), PN.getName());432 433 PHIArgMergedDebugLoc(NewEVI, PN);434 ++NumPHIsOfExtractValues;435 return NewEVI;436}437 438/// If we have something like phi [add (a,b), add(a,c)] and if a/b/c and the439/// adds all have a single user, turn this into a phi and a single binop.440Instruction *InstCombinerImpl::foldPHIArgBinOpIntoPHI(PHINode &PN) {441 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));442 assert(isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst));443 unsigned Opc = FirstInst->getOpcode();444 Value *LHSVal = FirstInst->getOperand(0);445 Value *RHSVal = FirstInst->getOperand(1);446 447 Type *LHSType = LHSVal->getType();448 Type *RHSType = RHSVal->getType();449 450 // Scan to see if all operands are the same opcode, and all have one user.451 for (Value *V : drop_begin(PN.incoming_values())) {452 Instruction *I = dyn_cast<Instruction>(V);453 if (!I || I->getOpcode() != Opc || !I->hasOneUser() ||454 // Verify type of the LHS matches so we don't fold cmp's of different455 // types.456 I->getOperand(0)->getType() != LHSType ||457 I->getOperand(1)->getType() != RHSType)458 return nullptr;459 460 // If they are CmpInst instructions, check their predicates461 if (CmpInst *CI = dyn_cast<CmpInst>(I))462 if (CI->getPredicate() != cast<CmpInst>(FirstInst)->getPredicate())463 return nullptr;464 465 // Keep track of which operand needs a phi node.466 if (I->getOperand(0) != LHSVal) LHSVal = nullptr;467 if (I->getOperand(1) != RHSVal) RHSVal = nullptr;468 }469 470 // If both LHS and RHS would need a PHI, don't do this transformation,471 // because it would increase the number of PHIs entering the block,472 // which leads to higher register pressure. This is especially473 // bad when the PHIs are in the header of a loop.474 if (!LHSVal && !RHSVal)475 return nullptr;476 477 // Otherwise, this is safe to transform!478 479 Value *InLHS = FirstInst->getOperand(0);480 Value *InRHS = FirstInst->getOperand(1);481 PHINode *NewLHS = nullptr, *NewRHS = nullptr;482 if (!LHSVal) {483 NewLHS = PHINode::Create(LHSType, PN.getNumIncomingValues(),484 FirstInst->getOperand(0)->getName() + ".pn");485 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));486 InsertNewInstBefore(NewLHS, PN.getIterator());487 LHSVal = NewLHS;488 }489 490 if (!RHSVal) {491 NewRHS = PHINode::Create(RHSType, PN.getNumIncomingValues(),492 FirstInst->getOperand(1)->getName() + ".pn");493 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));494 InsertNewInstBefore(NewRHS, PN.getIterator());495 RHSVal = NewRHS;496 }497 498 // Add all operands to the new PHIs.499 if (NewLHS || NewRHS) {500 for (auto Incoming : drop_begin(zip(PN.blocks(), PN.incoming_values()))) {501 BasicBlock *InBB = std::get<0>(Incoming);502 Value *InVal = std::get<1>(Incoming);503 Instruction *InInst = cast<Instruction>(InVal);504 if (NewLHS) {505 Value *NewInLHS = InInst->getOperand(0);506 NewLHS->addIncoming(NewInLHS, InBB);507 }508 if (NewRHS) {509 Value *NewInRHS = InInst->getOperand(1);510 NewRHS->addIncoming(NewInRHS, InBB);511 }512 }513 }514 515 if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst)) {516 CmpInst *NewCI = CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),517 LHSVal, RHSVal);518 PHIArgMergedDebugLoc(NewCI, PN);519 return NewCI;520 }521 522 BinaryOperator *BinOp = cast<BinaryOperator>(FirstInst);523 BinaryOperator *NewBinOp =524 BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal);525 526 NewBinOp->copyIRFlags(PN.getIncomingValue(0));527 528 for (Value *V : drop_begin(PN.incoming_values()))529 NewBinOp->andIRFlags(V);530 531 PHIArgMergedDebugLoc(NewBinOp, PN);532 return NewBinOp;533}534 535Instruction *InstCombinerImpl::foldPHIArgGEPIntoPHI(PHINode &PN) {536 GetElementPtrInst *FirstInst =cast<GetElementPtrInst>(PN.getIncomingValue(0));537 538 SmallVector<Value*, 16> FixedOperands(FirstInst->op_begin(),539 FirstInst->op_end());540 // This is true if all GEP bases are allocas and if all indices into them are541 // constants.542 bool AllBasePointersAreAllocas = true;543 544 // We don't want to replace this phi if the replacement would require545 // more than one phi, which leads to higher register pressure. This is546 // especially bad when the PHIs are in the header of a loop.547 bool NeededPhi = false;548 549 // Remember flags of the first phi-operand getelementptr.550 GEPNoWrapFlags NW = FirstInst->getNoWrapFlags();551 552 // Scan to see if all operands are the same opcode, and all have one user.553 for (Value *V : drop_begin(PN.incoming_values())) {554 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V);555 if (!GEP || !GEP->hasOneUser() ||556 GEP->getSourceElementType() != FirstInst->getSourceElementType() ||557 GEP->getNumOperands() != FirstInst->getNumOperands())558 return nullptr;559 560 NW &= GEP->getNoWrapFlags();561 562 // Keep track of whether or not all GEPs are of alloca pointers.563 if (AllBasePointersAreAllocas &&564 (!isa<AllocaInst>(GEP->getOperand(0)) ||565 !GEP->hasAllConstantIndices()))566 AllBasePointersAreAllocas = false;567 568 // Compare the operand lists.569 for (unsigned Op = 0, E = FirstInst->getNumOperands(); Op != E; ++Op) {570 if (FirstInst->getOperand(Op) == GEP->getOperand(Op))571 continue;572 573 // Don't merge two GEPs when two operands differ (introducing phi nodes)574 // if one of the PHIs has a constant for the index. The index may be575 // substantially cheaper to compute for the constants, so making it a576 // variable index could pessimize the path. This also handles the case577 // for struct indices, which must always be constant.578 if (isa<Constant>(FirstInst->getOperand(Op)) ||579 isa<Constant>(GEP->getOperand(Op)))580 return nullptr;581 582 if (FirstInst->getOperand(Op)->getType() !=583 GEP->getOperand(Op)->getType())584 return nullptr;585 586 // If we already needed a PHI for an earlier operand, and another operand587 // also requires a PHI, we'd be introducing more PHIs than we're588 // eliminating, which increases register pressure on entry to the PHI's589 // block.590 if (NeededPhi)591 return nullptr;592 593 FixedOperands[Op] = nullptr; // Needs a PHI.594 NeededPhi = true;595 }596 }597 598 // If all of the base pointers of the PHI'd GEPs are from allocas, don't599 // bother doing this transformation. At best, this will just save a bit of600 // offset calculation, but all the predecessors will have to materialize the601 // stack address into a register anyway. We'd actually rather *clone* the602 // load up into the predecessors so that we have a load of a gep of an alloca,603 // which can usually all be folded into the load.604 if (AllBasePointersAreAllocas)605 return nullptr;606 607 // Otherwise, this is safe to transform. Insert PHI nodes for each operand608 // that is variable.609 SmallVector<PHINode*, 16> OperandPhis(FixedOperands.size());610 611 bool HasAnyPHIs = false;612 for (unsigned I = 0, E = FixedOperands.size(); I != E; ++I) {613 if (FixedOperands[I])614 continue; // operand doesn't need a phi.615 Value *FirstOp = FirstInst->getOperand(I);616 PHINode *NewPN =617 PHINode::Create(FirstOp->getType(), E, FirstOp->getName() + ".pn");618 InsertNewInstBefore(NewPN, PN.getIterator());619 620 NewPN->addIncoming(FirstOp, PN.getIncomingBlock(0));621 OperandPhis[I] = NewPN;622 FixedOperands[I] = NewPN;623 HasAnyPHIs = true;624 }625 626 // Add all operands to the new PHIs.627 if (HasAnyPHIs) {628 for (auto Incoming : drop_begin(zip(PN.blocks(), PN.incoming_values()))) {629 BasicBlock *InBB = std::get<0>(Incoming);630 Value *InVal = std::get<1>(Incoming);631 GetElementPtrInst *InGEP = cast<GetElementPtrInst>(InVal);632 633 for (unsigned Op = 0, E = OperandPhis.size(); Op != E; ++Op)634 if (PHINode *OpPhi = OperandPhis[Op])635 OpPhi->addIncoming(InGEP->getOperand(Op), InBB);636 }637 }638 639 Value *Base = FixedOperands[0];640 GetElementPtrInst *NewGEP =641 GetElementPtrInst::Create(FirstInst->getSourceElementType(), Base,642 ArrayRef(FixedOperands).slice(1), NW);643 PHIArgMergedDebugLoc(NewGEP, PN);644 return NewGEP;645}646 647/// Return true if we know that it is safe to sink the load out of the block648/// that defines it. This means that it must be obvious the value of the load is649/// not changed from the point of the load to the end of the block it is in.650///651/// Finally, it is safe, but not profitable, to sink a load targeting a652/// non-address-taken alloca. Doing so will cause us to not promote the alloca653/// to a register.654static bool isSafeAndProfitableToSinkLoad(LoadInst *L) {655 BasicBlock::iterator BBI = L->getIterator(), E = L->getParent()->end();656 657 for (++BBI; BBI != E; ++BBI)658 if (BBI->mayWriteToMemory()) {659 // Calls that only access inaccessible memory do not block sinking the660 // load.661 if (auto *CB = dyn_cast<CallBase>(BBI))662 if (CB->onlyAccessesInaccessibleMemory())663 continue;664 return false;665 }666 667 // Check for non-address taken alloca. If not address-taken already, it isn't668 // profitable to do this xform.669 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {670 bool IsAddressTaken = false;671 for (User *U : AI->users()) {672 if (isa<LoadInst>(U)) continue;673 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {674 // If storing TO the alloca, then the address isn't taken.675 if (SI->getOperand(1) == AI) continue;676 }677 IsAddressTaken = true;678 break;679 }680 681 if (!IsAddressTaken && AI->isStaticAlloca())682 return false;683 }684 685 // If this load is a load from a GEP with a constant offset from an alloca,686 // then we don't want to sink it. In its present form, it will be687 // load [constant stack offset]. Sinking it will cause us to have to688 // materialize the stack addresses in each predecessor in a register only to689 // do a shared load from register in the successor.690 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(L->getOperand(0)))691 if (AllocaInst *AI = dyn_cast<AllocaInst>(GEP->getOperand(0)))692 if (AI->isStaticAlloca() && GEP->hasAllConstantIndices())693 return false;694 695 return true;696}697 698Instruction *InstCombinerImpl::foldPHIArgLoadIntoPHI(PHINode &PN) {699 LoadInst *FirstLI = cast<LoadInst>(PN.getIncomingValue(0));700 701 // Can't forward swifterror through a phi.702 if (FirstLI->getOperand(0)->isSwiftError())703 return nullptr;704 705 // FIXME: This is overconservative; this transform is allowed in some cases706 // for atomic operations.707 if (FirstLI->isAtomic())708 return nullptr;709 710 // When processing loads, we need to propagate two bits of information to the711 // sunk load: whether it is volatile, and what its alignment is.712 bool IsVolatile = FirstLI->isVolatile();713 Align LoadAlignment = FirstLI->getAlign();714 const unsigned LoadAddrSpace = FirstLI->getPointerAddressSpace();715 716 // We can't sink the load if the loaded value could be modified between the717 // load and the PHI.718 if (FirstLI->getParent() != PN.getIncomingBlock(0) ||719 !isSafeAndProfitableToSinkLoad(FirstLI))720 return nullptr;721 722 // If the PHI is of volatile loads and the load block has multiple723 // successors, sinking it would remove a load of the volatile value from724 // the path through the other successor.725 if (IsVolatile &&726 FirstLI->getParent()->getTerminator()->getNumSuccessors() != 1)727 return nullptr;728 729 for (auto Incoming : drop_begin(zip(PN.blocks(), PN.incoming_values()))) {730 BasicBlock *InBB = std::get<0>(Incoming);731 Value *InVal = std::get<1>(Incoming);732 LoadInst *LI = dyn_cast<LoadInst>(InVal);733 if (!LI || !LI->hasOneUser() || LI->isAtomic())734 return nullptr;735 736 // Make sure all arguments are the same type of operation.737 if (LI->isVolatile() != IsVolatile ||738 LI->getPointerAddressSpace() != LoadAddrSpace)739 return nullptr;740 741 // Can't forward swifterror through a phi.742 if (LI->getOperand(0)->isSwiftError())743 return nullptr;744 745 // We can't sink the load if the loaded value could be modified between746 // the load and the PHI.747 if (LI->getParent() != InBB || !isSafeAndProfitableToSinkLoad(LI))748 return nullptr;749 750 LoadAlignment = std::min(LoadAlignment, LI->getAlign());751 752 // If the PHI is of volatile loads and the load block has multiple753 // successors, sinking it would remove a load of the volatile value from754 // the path through the other successor.755 if (IsVolatile && LI->getParent()->getTerminator()->getNumSuccessors() != 1)756 return nullptr;757 }758 759 // Okay, they are all the same operation. Create a new PHI node of the760 // correct type, and PHI together all of the LHS's of the instructions.761 PHINode *NewPN = PHINode::Create(FirstLI->getOperand(0)->getType(),762 PN.getNumIncomingValues(),763 PN.getName()+".in");764 765 Value *InVal = FirstLI->getOperand(0);766 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));767 LoadInst *NewLI =768 new LoadInst(FirstLI->getType(), NewPN, "", IsVolatile, LoadAlignment);769 NewLI->copyMetadata(*FirstLI);770 771 // Add all operands to the new PHI and combine TBAA metadata.772 for (auto Incoming : drop_begin(zip(PN.blocks(), PN.incoming_values()))) {773 BasicBlock *BB = std::get<0>(Incoming);774 Value *V = std::get<1>(Incoming);775 LoadInst *LI = cast<LoadInst>(V);776 combineMetadataForCSE(NewLI, LI, true);777 Value *NewInVal = LI->getOperand(0);778 if (NewInVal != InVal)779 InVal = nullptr;780 NewPN->addIncoming(NewInVal, BB);781 }782 783 if (InVal) {784 // The new PHI unions all of the same values together. This is really785 // common, so we handle it intelligently here for compile-time speed.786 NewLI->setOperand(0, InVal);787 delete NewPN;788 } else {789 InsertNewInstBefore(NewPN, PN.getIterator());790 }791 792 // If this was a volatile load that we are merging, make sure to loop through793 // and mark all the input loads as non-volatile. If we don't do this, we will794 // insert a new volatile load and the old ones will not be deletable.795 if (IsVolatile)796 for (Value *IncValue : PN.incoming_values())797 cast<LoadInst>(IncValue)->setVolatile(false);798 799 PHIArgMergedDebugLoc(NewLI, PN);800 return NewLI;801}802 803/// TODO: This function could handle other cast types, but then it might804/// require special-casing a cast from the 'i1' type. See the comment in805/// FoldPHIArgOpIntoPHI() about pessimizing illegal integer types.806Instruction *InstCombinerImpl::foldPHIArgZextsIntoPHI(PHINode &Phi) {807 // We cannot create a new instruction after the PHI if the terminator is an808 // EHPad because there is no valid insertion point.809 if (Instruction *TI = Phi.getParent()->getTerminator())810 if (TI->isEHPad())811 return nullptr;812 813 // Early exit for the common case of a phi with two operands. These are814 // handled elsewhere. See the comment below where we check the count of zexts815 // and constants for more details.816 unsigned NumIncomingValues = Phi.getNumIncomingValues();817 if (NumIncomingValues < 3)818 return nullptr;819 820 // Find the narrower type specified by the first zext.821 Type *NarrowType = nullptr;822 for (Value *V : Phi.incoming_values()) {823 if (auto *Zext = dyn_cast<ZExtInst>(V)) {824 NarrowType = Zext->getSrcTy();825 break;826 }827 }828 if (!NarrowType)829 return nullptr;830 831 // Walk the phi operands checking that we only have zexts or constants that832 // we can shrink for free. Store the new operands for the new phi.833 SmallVector<Value *, 4> NewIncoming;834 unsigned NumZexts = 0;835 unsigned NumConsts = 0;836 for (Value *V : Phi.incoming_values()) {837 if (auto *Zext = dyn_cast<ZExtInst>(V)) {838 // All zexts must be identical and have one user.839 if (Zext->getSrcTy() != NarrowType || !Zext->hasOneUser())840 return nullptr;841 NewIncoming.push_back(Zext->getOperand(0));842 NumZexts++;843 } else if (auto *C = dyn_cast<Constant>(V)) {844 // Make sure that constants can fit in the new type.845 Constant *Trunc = getLosslessUnsignedTrunc(C, NarrowType, DL);846 if (!Trunc)847 return nullptr;848 NewIncoming.push_back(Trunc);849 NumConsts++;850 } else {851 // If it's not a cast or a constant, bail out.852 return nullptr;853 }854 }855 856 // The more common cases of a phi with no constant operands or just one857 // variable operand are handled by FoldPHIArgOpIntoPHI() and foldOpIntoPhi()858 // respectively. foldOpIntoPhi() wants to do the opposite transform that is859 // performed here. It tries to replicate a cast in the phi operand's basic860 // block to expose other folding opportunities. Thus, InstCombine will861 // infinite loop without this check.862 if (NumConsts == 0 || NumZexts < 2)863 return nullptr;864 865 // All incoming values are zexts or constants that are safe to truncate.866 // Create a new phi node of the narrow type, phi together all of the new867 // operands, and zext the result back to the original type.868 PHINode *NewPhi = PHINode::Create(NarrowType, NumIncomingValues,869 Phi.getName() + ".shrunk");870 for (unsigned I = 0; I != NumIncomingValues; ++I)871 NewPhi->addIncoming(NewIncoming[I], Phi.getIncomingBlock(I));872 873 InsertNewInstBefore(NewPhi, Phi.getIterator());874 auto *CI = CastInst::CreateZExtOrBitCast(NewPhi, Phi.getType());875 876 // We use a dropped location here because the new ZExt is necessarily a merge877 // of ZExtInsts and at least one constant from incoming branches; the presence878 // of the constant means we have no viable DebugLoc from that branch, and879 // therefore we must use a dropped location.880 CI->setDebugLoc(DebugLoc::getDropped());881 return CI;882}883 884/// If all operands to a PHI node are the same "unary" operator and they all are885/// only used by the PHI, PHI together their inputs, and do the operation once,886/// to the result of the PHI.887Instruction *InstCombinerImpl::foldPHIArgOpIntoPHI(PHINode &PN) {888 // We cannot create a new instruction after the PHI if the terminator is an889 // EHPad because there is no valid insertion point.890 if (Instruction *TI = PN.getParent()->getTerminator())891 if (TI->isEHPad())892 return nullptr;893 894 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));895 896 if (isa<GetElementPtrInst>(FirstInst))897 return foldPHIArgGEPIntoPHI(PN);898 if (isa<LoadInst>(FirstInst))899 return foldPHIArgLoadIntoPHI(PN);900 if (isa<InsertValueInst>(FirstInst))901 return foldPHIArgInsertValueInstructionIntoPHI(PN);902 if (isa<ExtractValueInst>(FirstInst))903 return foldPHIArgExtractValueInstructionIntoPHI(PN);904 905 // Scan the instruction, looking for input operations that can be folded away.906 // If all input operands to the phi are the same instruction (e.g. a cast from907 // the same type or "+42") we can pull the operation through the PHI, reducing908 // code size and simplifying code.909 Constant *ConstantOp = nullptr;910 Type *CastSrcTy = nullptr;911 912 if (isa<CastInst>(FirstInst)) {913 CastSrcTy = FirstInst->getOperand(0)->getType();914 915 // Be careful about transforming integer PHIs. We don't want to pessimize916 // the code by turning an i32 into an i1293.917 if (PN.getType()->isIntegerTy() && CastSrcTy->isIntegerTy()) {918 if (!shouldChangeType(PN.getType(), CastSrcTy))919 return nullptr;920 }921 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {922 // Can fold binop, compare or shift here if the RHS is a constant,923 // otherwise call FoldPHIArgBinOpIntoPHI.924 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));925 if (!ConstantOp)926 return foldPHIArgBinOpIntoPHI(PN);927 } else {928 return nullptr; // Cannot fold this operation.929 }930 931 // Check to see if all arguments are the same operation.932 for (Value *V : drop_begin(PN.incoming_values())) {933 Instruction *I = dyn_cast<Instruction>(V);934 if (!I || !I->hasOneUser() || !I->isSameOperationAs(FirstInst))935 return nullptr;936 if (CastSrcTy) {937 if (I->getOperand(0)->getType() != CastSrcTy)938 return nullptr; // Cast operation must match.939 } else if (I->getOperand(1) != ConstantOp) {940 return nullptr;941 }942 }943 944 // Okay, they are all the same operation. Create a new PHI node of the945 // correct type, and PHI together all of the LHS's of the instructions.946 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),947 PN.getNumIncomingValues(),948 PN.getName()+".in");949 950 Value *InVal = FirstInst->getOperand(0);951 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));952 953 // Add all operands to the new PHI.954 for (auto Incoming : drop_begin(zip(PN.blocks(), PN.incoming_values()))) {955 BasicBlock *BB = std::get<0>(Incoming);956 Value *V = std::get<1>(Incoming);957 Value *NewInVal = cast<Instruction>(V)->getOperand(0);958 if (NewInVal != InVal)959 InVal = nullptr;960 NewPN->addIncoming(NewInVal, BB);961 }962 963 Value *PhiVal;964 if (InVal) {965 // The new PHI unions all of the same values together. This is really966 // common, so we handle it intelligently here for compile-time speed.967 PhiVal = InVal;968 delete NewPN;969 } else {970 InsertNewInstBefore(NewPN, PN.getIterator());971 PhiVal = NewPN;972 }973 974 // Insert and return the new operation.975 if (CastInst *FirstCI = dyn_cast<CastInst>(FirstInst)) {976 CastInst *NewCI = CastInst::Create(FirstCI->getOpcode(), PhiVal,977 PN.getType());978 PHIArgMergedDebugLoc(NewCI, PN);979 return NewCI;980 }981 982 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst)) {983 BinOp = BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);984 BinOp->copyIRFlags(PN.getIncomingValue(0));985 986 for (Value *V : drop_begin(PN.incoming_values()))987 BinOp->andIRFlags(V);988 989 PHIArgMergedDebugLoc(BinOp, PN);990 return BinOp;991 }992 993 CmpInst *CIOp = cast<CmpInst>(FirstInst);994 CmpInst *NewCI = CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),995 PhiVal, ConstantOp);996 PHIArgMergedDebugLoc(NewCI, PN);997 return NewCI;998}999 1000/// Return true if this phi node is always equal to NonPhiInVal.1001/// This happens with mutually cyclic phi nodes like:1002/// z = some value; x = phi (y, z); y = phi (x, z)1003static bool PHIsEqualValue(PHINode *PN, Value *&NonPhiInVal,1004 SmallPtrSetImpl<PHINode *> &ValueEqualPHIs) {1005 // See if we already saw this PHI node.1006 if (!ValueEqualPHIs.insert(PN).second)1007 return true;1008 1009 // Don't scan crazily complex things.1010 if (ValueEqualPHIs.size() == 16)1011 return false;1012 1013 // Scan the operands to see if they are either phi nodes or are equal to1014 // the value.1015 for (Value *Op : PN->incoming_values()) {1016 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {1017 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs)) {1018 if (NonPhiInVal)1019 return false;1020 NonPhiInVal = OpPN;1021 }1022 } else if (Op != NonPhiInVal)1023 return false;1024 }1025 1026 return true;1027}1028 1029/// Return an existing non-zero constant if this phi node has one, otherwise1030/// return constant 1.1031static ConstantInt *getAnyNonZeroConstInt(PHINode &PN) {1032 assert(isa<IntegerType>(PN.getType()) && "Expect only integer type phi");1033 for (Value *V : PN.operands())1034 if (auto *ConstVA = dyn_cast<ConstantInt>(V))1035 if (!ConstVA->isZero())1036 return ConstVA;1037 return ConstantInt::get(cast<IntegerType>(PN.getType()), 1);1038}1039 1040namespace {1041struct PHIUsageRecord {1042 unsigned PHIId; // The ID # of the PHI (something determinstic to sort on)1043 unsigned Shift; // The amount shifted.1044 Instruction *Inst; // The trunc instruction.1045 1046 PHIUsageRecord(unsigned Pn, unsigned Sh, Instruction *User)1047 : PHIId(Pn), Shift(Sh), Inst(User) {}1048 1049 bool operator<(const PHIUsageRecord &RHS) const {1050 if (PHIId < RHS.PHIId) return true;1051 if (PHIId > RHS.PHIId) return false;1052 if (Shift < RHS.Shift) return true;1053 if (Shift > RHS.Shift) return false;1054 return Inst->getType()->getPrimitiveSizeInBits() <1055 RHS.Inst->getType()->getPrimitiveSizeInBits();1056 }1057};1058 1059struct LoweredPHIRecord {1060 PHINode *PN; // The PHI that was lowered.1061 unsigned Shift; // The amount shifted.1062 unsigned Width; // The width extracted.1063 1064 LoweredPHIRecord(PHINode *Phi, unsigned Sh, Type *Ty)1065 : PN(Phi), Shift(Sh), Width(Ty->getPrimitiveSizeInBits()) {}1066 1067 // Ctor form used by DenseMap.1068 LoweredPHIRecord(PHINode *Phi, unsigned Sh) : PN(Phi), Shift(Sh), Width(0) {}1069};1070} // namespace1071 1072template <> struct llvm::DenseMapInfo<LoweredPHIRecord> {1073 static inline LoweredPHIRecord getEmptyKey() {1074 return LoweredPHIRecord(nullptr, 0);1075 }1076 static inline LoweredPHIRecord getTombstoneKey() {1077 return LoweredPHIRecord(nullptr, 1);1078 }1079 static unsigned getHashValue(const LoweredPHIRecord &Val) {1080 return DenseMapInfo<PHINode *>::getHashValue(Val.PN) ^ (Val.Shift >> 3) ^1081 (Val.Width >> 3);1082 }1083 static bool isEqual(const LoweredPHIRecord &LHS,1084 const LoweredPHIRecord &RHS) {1085 return LHS.PN == RHS.PN && LHS.Shift == RHS.Shift && LHS.Width == RHS.Width;1086 }1087};1088 1089/// This is an integer PHI and we know that it has an illegal type: see if it is1090/// only used by trunc or trunc(lshr) operations. If so, we split the PHI into1091/// the various pieces being extracted. This sort of thing is introduced when1092/// SROA promotes an aggregate to large integer values.1093///1094/// TODO: The user of the trunc may be an bitcast to float/double/vector or an1095/// inttoptr. We should produce new PHIs in the right type.1096///1097Instruction *InstCombinerImpl::SliceUpIllegalIntegerPHI(PHINode &FirstPhi) {1098 // PHIUsers - Keep track of all of the truncated values extracted from a set1099 // of PHIs, along with their offset. These are the things we want to rewrite.1100 SmallVector<PHIUsageRecord, 16> PHIUsers;1101 1102 // PHIs are often mutually cyclic, so we keep track of a whole set of PHI1103 // nodes which are extracted from. PHIsToSlice is a set we use to avoid1104 // revisiting PHIs, PHIsInspected is a ordered list of PHIs that we need to1105 // check the uses of (to ensure they are all extracts).1106 SmallVector<PHINode*, 8> PHIsToSlice;1107 SmallPtrSet<PHINode*, 8> PHIsInspected;1108 1109 PHIsToSlice.push_back(&FirstPhi);1110 PHIsInspected.insert(&FirstPhi);1111 1112 for (unsigned PHIId = 0; PHIId != PHIsToSlice.size(); ++PHIId) {1113 PHINode *PN = PHIsToSlice[PHIId];1114 1115 // Scan the input list of the PHI. If any input is an invoke, and if the1116 // input is defined in the predecessor, then we won't be split the critical1117 // edge which is required to insert a truncate. Because of this, we have to1118 // bail out.1119 for (auto Incoming : zip(PN->blocks(), PN->incoming_values())) {1120 BasicBlock *BB = std::get<0>(Incoming);1121 Value *V = std::get<1>(Incoming);1122 InvokeInst *II = dyn_cast<InvokeInst>(V);1123 if (!II)1124 continue;1125 if (II->getParent() != BB)1126 continue;1127 1128 // If we have a phi, and if it's directly in the predecessor, then we have1129 // a critical edge where we need to put the truncate. Since we can't1130 // split the edge in instcombine, we have to bail out.1131 return nullptr;1132 }1133 1134 // If the incoming value is a PHI node before a catchswitch, we cannot1135 // extract the value within that BB because we cannot insert any non-PHI1136 // instructions in the BB.1137 for (auto *Pred : PN->blocks())1138 if (Pred->getFirstInsertionPt() == Pred->end())1139 return nullptr;1140 1141 for (User *U : PN->users()) {1142 Instruction *UserI = cast<Instruction>(U);1143 1144 // If the user is a PHI, inspect its uses recursively.1145 if (PHINode *UserPN = dyn_cast<PHINode>(UserI)) {1146 if (PHIsInspected.insert(UserPN).second)1147 PHIsToSlice.push_back(UserPN);1148 continue;1149 }1150 1151 // Truncates are always ok.1152 if (isa<TruncInst>(UserI)) {1153 PHIUsers.push_back(PHIUsageRecord(PHIId, 0, UserI));1154 continue;1155 }1156 1157 // Otherwise it must be a lshr which can only be used by one trunc.1158 if (UserI->getOpcode() != Instruction::LShr ||1159 !UserI->hasOneUse() || !isa<TruncInst>(UserI->user_back()) ||1160 !isa<ConstantInt>(UserI->getOperand(1)))1161 return nullptr;1162 1163 // Bail on out of range shifts.1164 unsigned SizeInBits = UserI->getType()->getScalarSizeInBits();1165 if (cast<ConstantInt>(UserI->getOperand(1))->getValue().uge(SizeInBits))1166 return nullptr;1167 1168 unsigned Shift = cast<ConstantInt>(UserI->getOperand(1))->getZExtValue();1169 PHIUsers.push_back(PHIUsageRecord(PHIId, Shift, UserI->user_back()));1170 }1171 }1172 1173 // If we have no users, they must be all self uses, just nuke the PHI.1174 if (PHIUsers.empty())1175 return replaceInstUsesWith(FirstPhi, PoisonValue::get(FirstPhi.getType()));1176 1177 // If this phi node is transformable, create new PHIs for all the pieces1178 // extracted out of it. First, sort the users by their offset and size.1179 array_pod_sort(PHIUsers.begin(), PHIUsers.end());1180 1181 LLVM_DEBUG(dbgs() << "SLICING UP PHI: " << FirstPhi << '\n';1182 for (unsigned I = 1; I != PHIsToSlice.size(); ++I) dbgs()1183 << "AND USER PHI #" << I << ": " << *PHIsToSlice[I] << '\n');1184 1185 // PredValues - This is a temporary used when rewriting PHI nodes. It is1186 // hoisted out here to avoid construction/destruction thrashing.1187 DenseMap<BasicBlock*, Value*> PredValues;1188 1189 // ExtractedVals - Each new PHI we introduce is saved here so we don't1190 // introduce redundant PHIs.1191 DenseMap<LoweredPHIRecord, PHINode*> ExtractedVals;1192 1193 for (unsigned UserI = 0, UserE = PHIUsers.size(); UserI != UserE; ++UserI) {1194 unsigned PHIId = PHIUsers[UserI].PHIId;1195 PHINode *PN = PHIsToSlice[PHIId];1196 unsigned Offset = PHIUsers[UserI].Shift;1197 Type *Ty = PHIUsers[UserI].Inst->getType();1198 1199 PHINode *EltPHI;1200 1201 // If we've already lowered a user like this, reuse the previously lowered1202 // value.1203 if ((EltPHI = ExtractedVals[LoweredPHIRecord(PN, Offset, Ty)]) == nullptr) {1204 1205 // Otherwise, Create the new PHI node for this user.1206 EltPHI = PHINode::Create(Ty, PN->getNumIncomingValues(),1207 PN->getName() + ".off" + Twine(Offset),1208 PN->getIterator());1209 assert(EltPHI->getType() != PN->getType() &&1210 "Truncate didn't shrink phi?");1211 1212 for (auto Incoming : zip(PN->blocks(), PN->incoming_values())) {1213 BasicBlock *Pred = std::get<0>(Incoming);1214 Value *InVal = std::get<1>(Incoming);1215 Value *&PredVal = PredValues[Pred];1216 1217 // If we already have a value for this predecessor, reuse it.1218 if (PredVal) {1219 EltPHI->addIncoming(PredVal, Pred);1220 continue;1221 }1222 1223 // Handle the PHI self-reuse case.1224 if (InVal == PN) {1225 PredVal = EltPHI;1226 EltPHI->addIncoming(PredVal, Pred);1227 continue;1228 }1229 1230 if (PHINode *InPHI = dyn_cast<PHINode>(PN)) {1231 // If the incoming value was a PHI, and if it was one of the PHIs we1232 // already rewrote it, just use the lowered value.1233 if (Value *Res = ExtractedVals[LoweredPHIRecord(InPHI, Offset, Ty)]) {1234 PredVal = Res;1235 EltPHI->addIncoming(PredVal, Pred);1236 continue;1237 }1238 }1239 1240 // Otherwise, do an extract in the predecessor.1241 Builder.SetInsertPoint(Pred->getTerminator());1242 Value *Res = InVal;1243 if (Offset)1244 Res = Builder.CreateLShr(1245 Res, ConstantInt::get(InVal->getType(), Offset), "extract");1246 Res = Builder.CreateTrunc(Res, Ty, "extract.t");1247 PredVal = Res;1248 EltPHI->addIncoming(Res, Pred);1249 1250 // If the incoming value was a PHI, and if it was one of the PHIs we are1251 // rewriting, we will ultimately delete the code we inserted. This1252 // means we need to revisit that PHI to make sure we extract out the1253 // needed piece.1254 if (PHINode *OldInVal = dyn_cast<PHINode>(InVal))1255 if (PHIsInspected.count(OldInVal)) {1256 unsigned RefPHIId =1257 find(PHIsToSlice, OldInVal) - PHIsToSlice.begin();1258 PHIUsers.push_back(1259 PHIUsageRecord(RefPHIId, Offset, cast<Instruction>(Res)));1260 ++UserE;1261 }1262 }1263 PredValues.clear();1264 1265 LLVM_DEBUG(dbgs() << " Made element PHI for offset " << Offset << ": "1266 << *EltPHI << '\n');1267 ExtractedVals[LoweredPHIRecord(PN, Offset, Ty)] = EltPHI;1268 }1269 1270 // Replace the use of this piece with the PHI node.1271 replaceInstUsesWith(*PHIUsers[UserI].Inst, EltPHI);1272 }1273 1274 // Replace all the remaining uses of the PHI nodes (self uses and the lshrs)1275 // with poison.1276 Value *Poison = PoisonValue::get(FirstPhi.getType());1277 for (PHINode *PHI : drop_begin(PHIsToSlice))1278 replaceInstUsesWith(*PHI, Poison);1279 return replaceInstUsesWith(FirstPhi, Poison);1280}1281 1282static Value *simplifyUsingControlFlow(InstCombiner &Self, PHINode &PN,1283 const DominatorTree &DT) {1284 // Simplify the following patterns:1285 // if (cond)1286 // / \1287 // ... ...1288 // \ /1289 // phi [true] [false]1290 // and1291 // switch (cond)1292 // case v1: / \ case v2:1293 // ... ...1294 // \ /1295 // phi [v1] [v2]1296 // Make sure all inputs are constants.1297 if (!all_of(PN.operands(), IsaPred<ConstantInt>))1298 return nullptr;1299 1300 BasicBlock *BB = PN.getParent();1301 // Do not bother with unreachable instructions.1302 if (!DT.isReachableFromEntry(BB))1303 return nullptr;1304 1305 // Determine which value the condition of the idom has for which successor.1306 LLVMContext &Context = PN.getContext();1307 auto *IDom = DT.getNode(BB)->getIDom()->getBlock();1308 Value *Cond;1309 SmallDenseMap<ConstantInt *, BasicBlock *, 8> SuccForValue;1310 SmallDenseMap<BasicBlock *, unsigned, 8> SuccCount;1311 auto AddSucc = [&](ConstantInt *C, BasicBlock *Succ) {1312 SuccForValue[C] = Succ;1313 ++SuccCount[Succ];1314 };1315 if (auto *BI = dyn_cast<BranchInst>(IDom->getTerminator())) {1316 if (BI->isUnconditional())1317 return nullptr;1318 1319 Cond = BI->getCondition();1320 AddSucc(ConstantInt::getTrue(Context), BI->getSuccessor(0));1321 AddSucc(ConstantInt::getFalse(Context), BI->getSuccessor(1));1322 } else if (auto *SI = dyn_cast<SwitchInst>(IDom->getTerminator())) {1323 Cond = SI->getCondition();1324 ++SuccCount[SI->getDefaultDest()];1325 for (auto Case : SI->cases())1326 AddSucc(Case.getCaseValue(), Case.getCaseSuccessor());1327 } else {1328 return nullptr;1329 }1330 1331 if (Cond->getType() != PN.getType())1332 return nullptr;1333 1334 // Check that edges outgoing from the idom's terminators dominate respective1335 // inputs of the Phi.1336 std::optional<bool> Invert;1337 for (auto Pair : zip(PN.incoming_values(), PN.blocks())) {1338 auto *Input = cast<ConstantInt>(std::get<0>(Pair));1339 BasicBlock *Pred = std::get<1>(Pair);1340 auto IsCorrectInput = [&](ConstantInt *Input) {1341 // The input needs to be dominated by the corresponding edge of the idom.1342 // This edge cannot be a multi-edge, as that would imply that multiple1343 // different condition values follow the same edge.1344 auto It = SuccForValue.find(Input);1345 return It != SuccForValue.end() && SuccCount[It->second] == 1 &&1346 DT.dominates(BasicBlockEdge(IDom, It->second),1347 BasicBlockEdge(Pred, BB));1348 };1349 1350 // Depending on the constant, the condition may need to be inverted.1351 bool NeedsInvert;1352 if (IsCorrectInput(Input))1353 NeedsInvert = false;1354 else if (IsCorrectInput(cast<ConstantInt>(ConstantExpr::getNot(Input))))1355 NeedsInvert = true;1356 else1357 return nullptr;1358 1359 // Make sure the inversion requirement is always the same.1360 if (Invert && *Invert != NeedsInvert)1361 return nullptr;1362 1363 Invert = NeedsInvert;1364 }1365 1366 if (!*Invert)1367 return Cond;1368 1369 // This Phi is actually opposite to branching condition of IDom. We invert1370 // the condition that will potentially open up some opportunities for1371 // sinking.1372 auto InsertPt = BB->getFirstInsertionPt();1373 if (InsertPt != BB->end()) {1374 Self.Builder.SetInsertPoint(&*BB, InsertPt);1375 return Self.Builder.CreateNot(Cond);1376 }1377 1378 return nullptr;1379}1380 1381// Fold iv = phi(start, iv.next = iv2.next op start)1382// where iv2 = phi(iv2.start, iv2.next = iv2 + iv2.step)1383// and iv2.start op start = start1384// to iv = iv2 op start1385static Value *foldDependentIVs(PHINode &PN, IRBuilderBase &Builder) {1386 BasicBlock *BB = PN.getParent();1387 if (PN.getNumIncomingValues() != 2)1388 return nullptr;1389 1390 Value *Start;1391 Instruction *IvNext;1392 BinaryOperator *Iv2Next;1393 auto MatchOuterIV = [&](Value *V1, Value *V2) {1394 if (match(V2, m_c_BinOp(m_Specific(V1), m_BinOp(Iv2Next))) ||1395 match(V2, m_GEP(m_Specific(V1), m_BinOp(Iv2Next)))) {1396 Start = V1;1397 IvNext = cast<Instruction>(V2);1398 return true;1399 }1400 return false;1401 };1402 1403 if (!MatchOuterIV(PN.getIncomingValue(0), PN.getIncomingValue(1)) &&1404 !MatchOuterIV(PN.getIncomingValue(1), PN.getIncomingValue(0)))1405 return nullptr;1406 1407 PHINode *Iv2;1408 Value *Iv2Start, *Iv2Step;1409 if (!matchSimpleRecurrence(Iv2Next, Iv2, Iv2Start, Iv2Step) ||1410 Iv2->getParent() != BB)1411 return nullptr;1412 1413 auto *BO = dyn_cast<BinaryOperator>(IvNext);1414 Constant *Identity =1415 BO ? ConstantExpr::getBinOpIdentity(BO->getOpcode(), Iv2Start->getType())1416 : Constant::getNullValue(Iv2Start->getType());1417 if (Iv2Start != Identity)1418 return nullptr;1419 1420 Builder.SetInsertPoint(&*BB, BB->getFirstInsertionPt());1421 if (!BO) {1422 auto *GEP = cast<GEPOperator>(IvNext);1423 return Builder.CreateGEP(GEP->getSourceElementType(), Start, Iv2, "",1424 cast<GEPOperator>(IvNext)->getNoWrapFlags());1425 }1426 1427 assert(BO->isCommutative() && "Must be commutative");1428 Value *Res = Builder.CreateBinOp(BO->getOpcode(), Iv2, Start);1429 cast<Instruction>(Res)->copyIRFlags(BO);1430 return Res;1431}1432 1433// PHINode simplification1434//1435Instruction *InstCombinerImpl::visitPHINode(PHINode &PN) {1436 if (Value *V = simplifyInstruction(&PN, SQ.getWithInstruction(&PN)))1437 return replaceInstUsesWith(PN, V);1438 1439 if (Instruction *Result = foldPHIArgZextsIntoPHI(PN))1440 return Result;1441 1442 if (Instruction *Result = foldPHIArgIntToPtrToPHI(PN))1443 return Result;1444 1445 // If all PHI operands are the same operation, pull them through the PHI,1446 // reducing code size.1447 auto *Inst0 = dyn_cast<Instruction>(PN.getIncomingValue(0));1448 auto *Inst1 = dyn_cast<Instruction>(PN.getIncomingValue(1));1449 if (Inst0 && Inst1 && Inst0->getOpcode() == Inst1->getOpcode() &&1450 Inst0->hasOneUser())1451 if (Instruction *Result = foldPHIArgOpIntoPHI(PN))1452 return Result;1453 1454 // If the incoming values are pointer casts of the same original value,1455 // replace the phi with a single cast iff we can insert a non-PHI instruction.1456 if (PN.getType()->isPointerTy() &&1457 PN.getParent()->getFirstInsertionPt() != PN.getParent()->end()) {1458 Value *IV0 = PN.getIncomingValue(0);1459 Value *IV0Stripped = IV0->stripPointerCasts();1460 // Set to keep track of values known to be equal to IV0Stripped after1461 // stripping pointer casts.1462 SmallPtrSet<Value *, 4> CheckedIVs;1463 CheckedIVs.insert(IV0);1464 if (IV0 != IV0Stripped &&1465 all_of(PN.incoming_values(), [&CheckedIVs, IV0Stripped](Value *IV) {1466 return !CheckedIVs.insert(IV).second ||1467 IV0Stripped == IV->stripPointerCasts();1468 })) {1469 return CastInst::CreatePointerCast(IV0Stripped, PN.getType());1470 }1471 }1472 1473 if (foldDeadPhiWeb(PN))1474 return nullptr;1475 1476 // Optimization when the phi only has one use1477 if (PN.hasOneUse()) {1478 if (foldIntegerTypedPHI(PN))1479 return nullptr;1480 1481 // If this phi has a single use, and if that use just computes a value for1482 // the next iteration of a loop, delete the phi. This occurs with unused1483 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this1484 // common case here is good because the only other things that catch this1485 // are induction variable analysis (sometimes) and ADCE, which is only run1486 // late.1487 Instruction *PHIUser = cast<Instruction>(PN.user_back());1488 if (PHIUser->hasOneUse() &&1489 (isa<BinaryOperator>(PHIUser) || isa<UnaryOperator>(PHIUser) ||1490 isa<GetElementPtrInst>(PHIUser)) &&1491 PHIUser->user_back() == &PN) {1492 return replaceInstUsesWith(PN, PoisonValue::get(PN.getType()));1493 }1494 }1495 1496 // When a PHI is used only to be compared with zero, it is safe to replace1497 // an incoming value proved as known nonzero with any non-zero constant.1498 // For example, in the code below, the incoming value %v can be replaced1499 // with any non-zero constant based on the fact that the PHI is only used to1500 // be compared with zero and %v is a known non-zero value:1501 // %v = select %cond, 1, 21502 // %p = phi [%v, BB] ...1503 // icmp eq, %p, 01504 // FIXME: To be simple, handle only integer type for now.1505 // This handles a small number of uses to keep the complexity down, and an1506 // icmp(or(phi)) can equally be replaced with any non-zero constant as the1507 // "or" will only add bits.1508 if (!PN.hasNUsesOrMore(3)) {1509 SmallVector<Instruction *> DropPoisonFlags;1510 bool AllUsesOfPhiEndsInCmp = all_of(PN.users(), [&](User *U) {1511 auto *CmpInst = dyn_cast<ICmpInst>(U);1512 if (!CmpInst) {1513 // This is always correct as OR only add bits and we are checking1514 // against 0.1515 if (U->hasOneUse() && match(U, m_c_Or(m_Specific(&PN), m_Value()))) {1516 DropPoisonFlags.push_back(cast<Instruction>(U));1517 CmpInst = dyn_cast<ICmpInst>(U->user_back());1518 }1519 }1520 if (!CmpInst || !isa<IntegerType>(PN.getType()) ||1521 !CmpInst->isEquality() || !match(CmpInst->getOperand(1), m_Zero())) {1522 return false;1523 }1524 return true;1525 });1526 // All uses of PHI results in a compare with zero.1527 if (AllUsesOfPhiEndsInCmp) {1528 ConstantInt *NonZeroConst = nullptr;1529 bool MadeChange = false;1530 for (unsigned I = 0, E = PN.getNumIncomingValues(); I != E; ++I) {1531 Instruction *CtxI = PN.getIncomingBlock(I)->getTerminator();1532 Value *VA = PN.getIncomingValue(I);1533 if (isKnownNonZero(VA, getSimplifyQuery().getWithInstruction(CtxI))) {1534 if (!NonZeroConst)1535 NonZeroConst = getAnyNonZeroConstInt(PN);1536 if (NonZeroConst != VA) {1537 replaceOperand(PN, I, NonZeroConst);1538 // The "disjoint" flag may no longer hold after the transform.1539 for (Instruction *I : DropPoisonFlags)1540 I->dropPoisonGeneratingFlags();1541 MadeChange = true;1542 }1543 }1544 }1545 if (MadeChange)1546 return &PN;1547 }1548 }1549 1550 // We sometimes end up with phi cycles that non-obviously end up being the1551 // same value, for example:1552 // z = some value; x = phi (y, z); y = phi (x, z)1553 // where the phi nodes don't necessarily need to be in the same block. Do a1554 // quick check to see if the PHI node only contains a single non-phi value, if1555 // so, scan to see if the phi cycle is actually equal to that value. If the1556 // phi has no non-phi values then allow the "NonPhiInVal" to be set later if1557 // one of the phis itself does not have a single input.1558 {1559 unsigned InValNo = 0, NumIncomingVals = PN.getNumIncomingValues();1560 // Scan for the first non-phi operand.1561 while (InValNo != NumIncomingVals &&1562 isa<PHINode>(PN.getIncomingValue(InValNo)))1563 ++InValNo;1564 1565 Value *NonPhiInVal =1566 InValNo != NumIncomingVals ? PN.getIncomingValue(InValNo) : nullptr;1567 1568 // Scan the rest of the operands to see if there are any conflicts, if so1569 // there is no need to recursively scan other phis.1570 if (NonPhiInVal)1571 for (++InValNo; InValNo != NumIncomingVals; ++InValNo) {1572 Value *OpVal = PN.getIncomingValue(InValNo);1573 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))1574 break;1575 }1576 1577 // If we scanned over all operands, then we have one unique value plus1578 // phi values. Scan PHI nodes to see if they all merge in each other or1579 // the value.1580 if (InValNo == NumIncomingVals) {1581 SmallPtrSet<PHINode *, 16> ValueEqualPHIs;1582 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))1583 return replaceInstUsesWith(PN, NonPhiInVal);1584 }1585 }1586 1587 // If there are multiple PHIs, sort their operands so that they all list1588 // the blocks in the same order. This will help identical PHIs be eliminated1589 // by other passes. Other passes shouldn't depend on this for correctness1590 // however.1591 auto Res = PredOrder.try_emplace(PN.getParent());1592 if (!Res.second) {1593 const auto &Preds = Res.first->second;1594 for (unsigned I = 0, E = PN.getNumIncomingValues(); I != E; ++I) {1595 BasicBlock *BBA = PN.getIncomingBlock(I);1596 BasicBlock *BBB = Preds[I];1597 if (BBA != BBB) {1598 Value *VA = PN.getIncomingValue(I);1599 unsigned J = PN.getBasicBlockIndex(BBB);1600 Value *VB = PN.getIncomingValue(J);1601 PN.setIncomingBlock(I, BBB);1602 PN.setIncomingValue(I, VB);1603 PN.setIncomingBlock(J, BBA);1604 PN.setIncomingValue(J, VA);1605 // NOTE: Instcombine normally would want us to "return &PN" if we1606 // modified any of the operands of an instruction. However, since we1607 // aren't adding or removing uses (just rearranging them) we don't do1608 // this in this case.1609 }1610 }1611 } else {1612 // Remember the block order of the first encountered phi node.1613 append_range(Res.first->second, PN.blocks());1614 }1615 1616 // Is there an identical PHI node in this basic block?1617 for (PHINode &IdenticalPN : PN.getParent()->phis()) {1618 // Ignore the PHI node itself.1619 if (&IdenticalPN == &PN)1620 continue;1621 // Note that even though we've just canonicalized this PHI, due to the1622 // worklist visitation order, there are no guarantess that *every* PHI1623 // has been canonicalized, so we can't just compare operands ranges.1624 if (!PN.isIdenticalToWhenDefined(&IdenticalPN))1625 continue;1626 // Just use that PHI instead then.1627 ++NumPHICSEs;1628 return replaceInstUsesWith(PN, &IdenticalPN);1629 }1630 1631 // If this is an integer PHI and we know that it has an illegal type, see if1632 // it is only used by trunc or trunc(lshr) operations. If so, we split the1633 // PHI into the various pieces being extracted. This sort of thing is1634 // introduced when SROA promotes an aggregate to a single large integer type.1635 if (PN.getType()->isIntegerTy() &&1636 !DL.isLegalInteger(PN.getType()->getPrimitiveSizeInBits()))1637 if (Instruction *Res = SliceUpIllegalIntegerPHI(PN))1638 return Res;1639 1640 // Ultimately, try to replace this Phi with a dominating condition.1641 if (auto *V = simplifyUsingControlFlow(*this, PN, DT))1642 return replaceInstUsesWith(PN, V);1643 1644 if (Value *Res = foldDependentIVs(PN, Builder))1645 return replaceInstUsesWith(PN, Res);1646 1647 return nullptr;1648}1649