1 //===- CodeGenPrepare.cpp - Prepare a function for code generation --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass munges the code in the input function to better prepare it for
11 // SelectionDAG-based code generation. This works around limitations in it's
12 // basic-block-at-a-time approach. It should eventually be removed.
14 //===----------------------------------------------------------------------===//
16 #include "llvm/CodeGen/Passes.h"
17 #include "llvm/ADT/DenseMap.h"
18 #include "llvm/ADT/SmallSet.h"
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/Analysis/InstructionSimplify.h"
21 #include "llvm/Analysis/TargetLibraryInfo.h"
22 #include "llvm/Analysis/TargetTransformInfo.h"
23 #include "llvm/IR/CallSite.h"
24 #include "llvm/IR/Constants.h"
25 #include "llvm/IR/DataLayout.h"
26 #include "llvm/IR/DerivedTypes.h"
27 #include "llvm/IR/Dominators.h"
28 #include "llvm/IR/Function.h"
29 #include "llvm/IR/GetElementPtrTypeIterator.h"
30 #include "llvm/IR/IRBuilder.h"
31 #include "llvm/IR/InlineAsm.h"
32 #include "llvm/IR/Instructions.h"
33 #include "llvm/IR/IntrinsicInst.h"
34 #include "llvm/IR/MDBuilder.h"
35 #include "llvm/IR/PatternMatch.h"
36 #include "llvm/IR/Statepoint.h"
37 #include "llvm/IR/ValueHandle.h"
38 #include "llvm/IR/ValueMap.h"
39 #include "llvm/Pass.h"
40 #include "llvm/Support/CommandLine.h"
41 #include "llvm/Support/Debug.h"
42 #include "llvm/Support/raw_ostream.h"
43 #include "llvm/Target/TargetLowering.h"
44 #include "llvm/Target/TargetSubtargetInfo.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/BuildLibCalls.h"
47 #include "llvm/Transforms/Utils/BypassSlowDivision.h"
48 #include "llvm/Transforms/Utils/Local.h"
49 #include "llvm/Transforms/Utils/SimplifyLibCalls.h"
51 using namespace llvm::PatternMatch;
53 #define DEBUG_TYPE "codegenprepare"
55 STATISTIC(NumBlocksElim, "Number of blocks eliminated");
56 STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated");
57 STATISTIC(NumGEPsElim, "Number of GEPs converted to casts");
58 STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "
60 STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "
62 STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "
63 "computations were sunk");
64 STATISTIC(NumExtsMoved, "Number of [s|z]ext instructions combined with loads");
65 STATISTIC(NumExtUses, "Number of uses of [s|z]ext instructions optimized");
66 STATISTIC(NumRetsDup, "Number of return instructions duplicated");
67 STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved");
68 STATISTIC(NumSelectsExpanded, "Number of selects turned into branches");
69 STATISTIC(NumAndCmpsMoved, "Number of and/cmp's pushed into branches");
70 STATISTIC(NumStoreExtractExposed, "Number of store(extractelement) exposed");
72 static cl::opt<bool> DisableBranchOpts(
73 "disable-cgp-branch-opts", cl::Hidden, cl::init(false),
74 cl::desc("Disable branch optimizations in CodeGenPrepare"));
77 DisableGCOpts("disable-cgp-gc-opts", cl::Hidden, cl::init(false),
78 cl::desc("Disable GC optimizations in CodeGenPrepare"));
80 static cl::opt<bool> DisableSelectToBranch(
81 "disable-cgp-select2branch", cl::Hidden, cl::init(false),
82 cl::desc("Disable select to branch conversion."));
84 static cl::opt<bool> AddrSinkUsingGEPs(
85 "addr-sink-using-gep", cl::Hidden, cl::init(false),
86 cl::desc("Address sinking in CGP using GEPs."));
88 static cl::opt<bool> EnableAndCmpSinking(
89 "enable-andcmp-sinking", cl::Hidden, cl::init(true),
90 cl::desc("Enable sinkinig and/cmp into branches."));
92 static cl::opt<bool> DisableStoreExtract(
93 "disable-cgp-store-extract", cl::Hidden, cl::init(false),
94 cl::desc("Disable store(extract) optimizations in CodeGenPrepare"));
96 static cl::opt<bool> StressStoreExtract(
97 "stress-cgp-store-extract", cl::Hidden, cl::init(false),
98 cl::desc("Stress test store(extract) optimizations in CodeGenPrepare"));
100 static cl::opt<bool> DisableExtLdPromotion(
101 "disable-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
102 cl::desc("Disable ext(promotable(ld)) -> promoted(ext(ld)) optimization in "
105 static cl::opt<bool> StressExtLdPromotion(
106 "stress-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
107 cl::desc("Stress test ext(promotable(ld)) -> promoted(ext(ld)) "
108 "optimization in CodeGenPrepare"));
111 typedef SmallPtrSet<Instruction *, 16> SetOfInstrs;
115 TypeIsSExt(Type *Ty, bool IsSExt) : Ty(Ty), IsSExt(IsSExt) {}
117 typedef DenseMap<Instruction *, TypeIsSExt> InstrToOrigTy;
118 class TypePromotionTransaction;
120 class CodeGenPrepare : public FunctionPass {
121 /// TLI - Keep a pointer of a TargetLowering to consult for determining
122 /// transformation profitability.
123 const TargetMachine *TM;
124 const TargetLowering *TLI;
125 const TargetTransformInfo *TTI;
126 const TargetLibraryInfo *TLInfo;
129 /// CurInstIterator - As we scan instructions optimizing them, this is the
130 /// next instruction to optimize. Xforms that can invalidate this should
132 BasicBlock::iterator CurInstIterator;
134 /// Keeps track of non-local addresses that have been sunk into a block.
135 /// This allows us to avoid inserting duplicate code for blocks with
136 /// multiple load/stores of the same address.
137 ValueMap<Value*, Value*> SunkAddrs;
139 /// Keeps track of all truncates inserted for the current function.
140 SetOfInstrs InsertedTruncsSet;
141 /// Keeps track of the type of the related instruction before their
142 /// promotion for the current function.
143 InstrToOrigTy PromotedInsts;
145 /// ModifiedDT - If CFG is modified in anyway, dominator tree may need to
149 /// OptSize - True if optimizing for size.
153 static char ID; // Pass identification, replacement for typeid
154 explicit CodeGenPrepare(const TargetMachine *TM = nullptr)
155 : FunctionPass(ID), TM(TM), TLI(nullptr), TTI(nullptr) {
156 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
158 bool runOnFunction(Function &F) override;
160 const char *getPassName() const override { return "CodeGen Prepare"; }
162 void getAnalysisUsage(AnalysisUsage &AU) const override {
163 AU.addPreserved<DominatorTreeWrapperPass>();
164 AU.addRequired<TargetLibraryInfoWrapperPass>();
165 AU.addRequired<TargetTransformInfoWrapperPass>();
169 bool EliminateFallThrough(Function &F);
170 bool EliminateMostlyEmptyBlocks(Function &F);
171 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
172 void EliminateMostlyEmptyBlock(BasicBlock *BB);
173 bool OptimizeBlock(BasicBlock &BB, bool& ModifiedDT);
174 bool OptimizeInst(Instruction *I, bool& ModifiedDT);
175 bool OptimizeMemoryInst(Instruction *I, Value *Addr, Type *AccessTy);
176 bool OptimizeInlineAsmInst(CallInst *CS);
177 bool OptimizeCallInst(CallInst *CI, bool& ModifiedDT);
178 bool MoveExtToFormExtLoad(Instruction *&I);
179 bool OptimizeExtUses(Instruction *I);
180 bool OptimizeSelectInst(SelectInst *SI);
181 bool OptimizeShuffleVectorInst(ShuffleVectorInst *SI);
182 bool OptimizeExtractElementInst(Instruction *Inst);
183 bool DupRetToEnableTailCallOpts(BasicBlock *BB);
184 bool PlaceDbgValues(Function &F);
185 bool sinkAndCmp(Function &F);
186 bool ExtLdPromotion(TypePromotionTransaction &TPT, LoadInst *&LI,
188 const SmallVectorImpl<Instruction *> &Exts,
189 unsigned CreatedInst);
190 bool splitBranchCondition(Function &F);
191 bool simplifyOffsetableRelocate(Instruction &I);
195 char CodeGenPrepare::ID = 0;
196 INITIALIZE_TM_PASS(CodeGenPrepare, "codegenprepare",
197 "Optimize for code generation", false, false)
199 FunctionPass *llvm::createCodeGenPreparePass(const TargetMachine *TM) {
200 return new CodeGenPrepare(TM);
203 bool CodeGenPrepare::runOnFunction(Function &F) {
204 if (skipOptnoneFunction(F))
207 bool EverMadeChange = false;
208 // Clear per function information.
209 InsertedTruncsSet.clear();
210 PromotedInsts.clear();
214 TLI = TM->getSubtargetImpl(F)->getTargetLowering();
215 TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
216 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
217 DominatorTreeWrapperPass *DTWP =
218 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
219 DT = DTWP ? &DTWP->getDomTree() : nullptr;
220 OptSize = F.hasFnAttribute(Attribute::OptimizeForSize);
222 /// This optimization identifies DIV instructions that can be
223 /// profitably bypassed and carried out with a shorter, faster divide.
224 if (!OptSize && TLI && TLI->isSlowDivBypassed()) {
225 const DenseMap<unsigned int, unsigned int> &BypassWidths =
226 TLI->getBypassSlowDivWidths();
227 for (Function::iterator I = F.begin(); I != F.end(); I++)
228 EverMadeChange |= bypassSlowDivision(F, I, BypassWidths);
231 // Eliminate blocks that contain only PHI nodes and an
232 // unconditional branch.
233 EverMadeChange |= EliminateMostlyEmptyBlocks(F);
235 // llvm.dbg.value is far away from the value then iSel may not be able
236 // handle it properly. iSel will drop llvm.dbg.value if it can not
237 // find a node corresponding to the value.
238 EverMadeChange |= PlaceDbgValues(F);
240 // If there is a mask, compare against zero, and branch that can be combined
241 // into a single target instruction, push the mask and compare into branch
242 // users. Do this before OptimizeBlock -> OptimizeInst ->
243 // OptimizeCmpExpression, which perturbs the pattern being searched for.
244 if (!DisableBranchOpts) {
245 EverMadeChange |= sinkAndCmp(F);
246 EverMadeChange |= splitBranchCondition(F);
249 bool MadeChange = true;
252 for (Function::iterator I = F.begin(); I != F.end(); ) {
253 BasicBlock *BB = I++;
254 bool ModifiedDTOnIteration = false;
255 MadeChange |= OptimizeBlock(*BB, ModifiedDTOnIteration);
257 // Restart BB iteration if the dominator tree of the Function was changed
258 ModifiedDT |= ModifiedDTOnIteration;
259 if (ModifiedDTOnIteration)
262 EverMadeChange |= MadeChange;
267 if (!DisableBranchOpts) {
269 SmallPtrSet<BasicBlock*, 8> WorkList;
270 for (BasicBlock &BB : F) {
271 SmallVector<BasicBlock *, 2> Successors(succ_begin(&BB), succ_end(&BB));
272 MadeChange |= ConstantFoldTerminator(&BB, true);
273 if (!MadeChange) continue;
275 for (SmallVectorImpl<BasicBlock*>::iterator
276 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
277 if (pred_begin(*II) == pred_end(*II))
278 WorkList.insert(*II);
281 // Delete the dead blocks and any of their dead successors.
282 MadeChange |= !WorkList.empty();
283 while (!WorkList.empty()) {
284 BasicBlock *BB = *WorkList.begin();
286 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
290 for (SmallVectorImpl<BasicBlock*>::iterator
291 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
292 if (pred_begin(*II) == pred_end(*II))
293 WorkList.insert(*II);
296 // Merge pairs of basic blocks with unconditional branches, connected by
298 if (EverMadeChange || MadeChange)
299 MadeChange |= EliminateFallThrough(F);
303 EverMadeChange |= MadeChange;
306 if (!DisableGCOpts) {
307 SmallVector<Instruction *, 2> Statepoints;
308 for (BasicBlock &BB : F)
309 for (Instruction &I : BB)
311 Statepoints.push_back(&I);
312 for (auto &I : Statepoints)
313 EverMadeChange |= simplifyOffsetableRelocate(*I);
316 if (ModifiedDT && DT)
319 return EverMadeChange;
322 /// EliminateFallThrough - Merge basic blocks which are connected
323 /// by a single edge, where one of the basic blocks has a single successor
324 /// pointing to the other basic block, which has a single predecessor.
325 bool CodeGenPrepare::EliminateFallThrough(Function &F) {
326 bool Changed = false;
327 // Scan all of the blocks in the function, except for the entry block.
328 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
329 BasicBlock *BB = I++;
330 // If the destination block has a single pred, then this is a trivial
331 // edge, just collapse it.
332 BasicBlock *SinglePred = BB->getSinglePredecessor();
334 // Don't merge if BB's address is taken.
335 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
337 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
338 if (Term && !Term->isConditional()) {
340 DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n");
341 // Remember if SinglePred was the entry block of the function.
342 // If so, we will need to move BB back to the entry position.
343 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
344 MergeBasicBlockIntoOnlyPred(BB, DT);
346 if (isEntry && BB != &BB->getParent()->getEntryBlock())
347 BB->moveBefore(&BB->getParent()->getEntryBlock());
349 // We have erased a block. Update the iterator.
356 /// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes,
357 /// debug info directives, and an unconditional branch. Passes before isel
358 /// (e.g. LSR/loopsimplify) often split edges in ways that are non-optimal for
359 /// isel. Start by eliminating these blocks so we can split them the way we
361 bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) {
362 bool MadeChange = false;
363 // Note that this intentionally skips the entry block.
364 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
365 BasicBlock *BB = I++;
367 // If this block doesn't end with an uncond branch, ignore it.
368 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
369 if (!BI || !BI->isUnconditional())
372 // If the instruction before the branch (skipping debug info) isn't a phi
373 // node, then other stuff is happening here.
374 BasicBlock::iterator BBI = BI;
375 if (BBI != BB->begin()) {
377 while (isa<DbgInfoIntrinsic>(BBI)) {
378 if (BBI == BB->begin())
382 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
386 // Do not break infinite loops.
387 BasicBlock *DestBB = BI->getSuccessor(0);
391 if (!CanMergeBlocks(BB, DestBB))
394 EliminateMostlyEmptyBlock(BB);
400 /// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a
401 /// single uncond branch between them, and BB contains no other non-phi
403 bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB,
404 const BasicBlock *DestBB) const {
405 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
406 // the successor. If there are more complex condition (e.g. preheaders),
407 // don't mess around with them.
408 BasicBlock::const_iterator BBI = BB->begin();
409 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
410 for (const User *U : PN->users()) {
411 const Instruction *UI = cast<Instruction>(U);
412 if (UI->getParent() != DestBB || !isa<PHINode>(UI))
414 // If User is inside DestBB block and it is a PHINode then check
415 // incoming value. If incoming value is not from BB then this is
416 // a complex condition (e.g. preheaders) we want to avoid here.
417 if (UI->getParent() == DestBB) {
418 if (const PHINode *UPN = dyn_cast<PHINode>(UI))
419 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
420 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
421 if (Insn && Insn->getParent() == BB &&
422 Insn->getParent() != UPN->getIncomingBlock(I))
429 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
430 // and DestBB may have conflicting incoming values for the block. If so, we
431 // can't merge the block.
432 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
433 if (!DestBBPN) return true; // no conflict.
435 // Collect the preds of BB.
436 SmallPtrSet<const BasicBlock*, 16> BBPreds;
437 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
438 // It is faster to get preds from a PHI than with pred_iterator.
439 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
440 BBPreds.insert(BBPN->getIncomingBlock(i));
442 BBPreds.insert(pred_begin(BB), pred_end(BB));
445 // Walk the preds of DestBB.
446 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
447 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
448 if (BBPreds.count(Pred)) { // Common predecessor?
449 BBI = DestBB->begin();
450 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
451 const Value *V1 = PN->getIncomingValueForBlock(Pred);
452 const Value *V2 = PN->getIncomingValueForBlock(BB);
454 // If V2 is a phi node in BB, look up what the mapped value will be.
455 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
456 if (V2PN->getParent() == BB)
457 V2 = V2PN->getIncomingValueForBlock(Pred);
459 // If there is a conflict, bail out.
460 if (V1 != V2) return false;
469 /// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and
470 /// an unconditional branch in it.
471 void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) {
472 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
473 BasicBlock *DestBB = BI->getSuccessor(0);
475 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
477 // If the destination block has a single pred, then this is a trivial edge,
479 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
480 if (SinglePred != DestBB) {
481 // Remember if SinglePred was the entry block of the function. If so, we
482 // will need to move BB back to the entry position.
483 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
484 MergeBasicBlockIntoOnlyPred(DestBB, DT);
486 if (isEntry && BB != &BB->getParent()->getEntryBlock())
487 BB->moveBefore(&BB->getParent()->getEntryBlock());
489 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
494 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
495 // to handle the new incoming edges it is about to have.
497 for (BasicBlock::iterator BBI = DestBB->begin();
498 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
499 // Remove the incoming value for BB, and remember it.
500 Value *InVal = PN->removeIncomingValue(BB, false);
502 // Two options: either the InVal is a phi node defined in BB or it is some
503 // value that dominates BB.
504 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
505 if (InValPhi && InValPhi->getParent() == BB) {
506 // Add all of the input values of the input PHI as inputs of this phi.
507 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
508 PN->addIncoming(InValPhi->getIncomingValue(i),
509 InValPhi->getIncomingBlock(i));
511 // Otherwise, add one instance of the dominating value for each edge that
512 // we will be adding.
513 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
514 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
515 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
517 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
518 PN->addIncoming(InVal, *PI);
523 // The PHIs are now updated, change everything that refers to BB to use
524 // DestBB and remove BB.
525 BB->replaceAllUsesWith(DestBB);
526 if (DT && !ModifiedDT) {
527 BasicBlock *BBIDom = DT->getNode(BB)->getIDom()->getBlock();
528 BasicBlock *DestBBIDom = DT->getNode(DestBB)->getIDom()->getBlock();
529 BasicBlock *NewIDom = DT->findNearestCommonDominator(BBIDom, DestBBIDom);
530 DT->changeImmediateDominator(DestBB, NewIDom);
533 BB->eraseFromParent();
536 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
539 // Computes a map of base pointer relocation instructions to corresponding
540 // derived pointer relocation instructions given a vector of all relocate calls
541 static void computeBaseDerivedRelocateMap(
542 const SmallVectorImpl<User *> &AllRelocateCalls,
543 DenseMap<IntrinsicInst *, SmallVector<IntrinsicInst *, 2>> &
545 // Collect information in two maps: one primarily for locating the base object
546 // while filling the second map; the second map is the final structure holding
547 // a mapping between Base and corresponding Derived relocate calls
548 DenseMap<std::pair<unsigned, unsigned>, IntrinsicInst *> RelocateIdxMap;
549 for (auto &U : AllRelocateCalls) {
550 GCRelocateOperands ThisRelocate(U);
551 IntrinsicInst *I = cast<IntrinsicInst>(U);
552 auto K = std::make_pair(ThisRelocate.basePtrIndex(),
553 ThisRelocate.derivedPtrIndex());
554 RelocateIdxMap.insert(std::make_pair(K, I));
556 for (auto &Item : RelocateIdxMap) {
557 std::pair<unsigned, unsigned> Key = Item.first;
558 if (Key.first == Key.second)
559 // Base relocation: nothing to insert
562 IntrinsicInst *I = Item.second;
563 auto BaseKey = std::make_pair(Key.first, Key.first);
565 // We're iterating over RelocateIdxMap so we cannot modify it.
566 auto MaybeBase = RelocateIdxMap.find(BaseKey);
567 if (MaybeBase == RelocateIdxMap.end())
568 // TODO: We might want to insert a new base object relocate and gep off
569 // that, if there are enough derived object relocates.
572 RelocateInstMap[MaybeBase->second].push_back(I);
576 // Accepts a GEP and extracts the operands into a vector provided they're all
577 // small integer constants
578 static bool getGEPSmallConstantIntOffsetV(GetElementPtrInst *GEP,
579 SmallVectorImpl<Value *> &OffsetV) {
580 for (unsigned i = 1; i < GEP->getNumOperands(); i++) {
581 // Only accept small constant integer operands
582 auto Op = dyn_cast<ConstantInt>(GEP->getOperand(i));
583 if (!Op || Op->getZExtValue() > 20)
587 for (unsigned i = 1; i < GEP->getNumOperands(); i++)
588 OffsetV.push_back(GEP->getOperand(i));
592 // Takes a RelocatedBase (base pointer relocation instruction) and Targets to
593 // replace, computes a replacement, and affects it.
595 simplifyRelocatesOffABase(IntrinsicInst *RelocatedBase,
596 const SmallVectorImpl<IntrinsicInst *> &Targets) {
597 bool MadeChange = false;
598 for (auto &ToReplace : Targets) {
599 GCRelocateOperands MasterRelocate(RelocatedBase);
600 GCRelocateOperands ThisRelocate(ToReplace);
602 assert(ThisRelocate.basePtrIndex() == MasterRelocate.basePtrIndex() &&
603 "Not relocating a derived object of the original base object");
604 if (ThisRelocate.basePtrIndex() == ThisRelocate.derivedPtrIndex()) {
605 // A duplicate relocate call. TODO: coalesce duplicates.
609 Value *Base = ThisRelocate.basePtr();
610 auto Derived = dyn_cast<GetElementPtrInst>(ThisRelocate.derivedPtr());
611 if (!Derived || Derived->getPointerOperand() != Base)
614 SmallVector<Value *, 2> OffsetV;
615 if (!getGEPSmallConstantIntOffsetV(Derived, OffsetV))
618 // Create a Builder and replace the target callsite with a gep
619 IRBuilder<> Builder(ToReplace);
620 Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
622 Builder.CreateGEP(RelocatedBase, makeArrayRef(OffsetV));
623 Instruction *ReplacementInst = cast<Instruction>(Replacement);
624 ReplacementInst->removeFromParent();
625 ReplacementInst->insertAfter(RelocatedBase);
626 Replacement->takeName(ToReplace);
627 ToReplace->replaceAllUsesWith(Replacement);
628 ToReplace->eraseFromParent();
638 // %ptr = gep %base + 15
639 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
640 // %base' = relocate(%tok, i32 4, i32 4)
641 // %ptr' = relocate(%tok, i32 4, i32 5)
647 // %ptr = gep %base + 15
648 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
649 // %base' = gc.relocate(%tok, i32 4, i32 4)
650 // %ptr' = gep %base' + 15
652 bool CodeGenPrepare::simplifyOffsetableRelocate(Instruction &I) {
653 bool MadeChange = false;
654 SmallVector<User *, 2> AllRelocateCalls;
656 for (auto *U : I.users())
657 if (isGCRelocate(dyn_cast<Instruction>(U)))
658 // Collect all the relocate calls associated with a statepoint
659 AllRelocateCalls.push_back(U);
661 // We need atleast one base pointer relocation + one derived pointer
662 // relocation to mangle
663 if (AllRelocateCalls.size() < 2)
666 // RelocateInstMap is a mapping from the base relocate instruction to the
667 // corresponding derived relocate instructions
668 DenseMap<IntrinsicInst *, SmallVector<IntrinsicInst *, 2>> RelocateInstMap;
669 computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap);
670 if (RelocateInstMap.empty())
673 for (auto &Item : RelocateInstMap)
674 // Item.first is the RelocatedBase to offset against
675 // Item.second is the vector of Targets to replace
676 MadeChange = simplifyRelocatesOffABase(Item.first, Item.second);
680 /// SinkCast - Sink the specified cast instruction into its user blocks
681 static bool SinkCast(CastInst *CI) {
682 BasicBlock *DefBB = CI->getParent();
684 /// InsertedCasts - Only insert a cast in each block once.
685 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
687 bool MadeChange = false;
688 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
690 Use &TheUse = UI.getUse();
691 Instruction *User = cast<Instruction>(*UI);
693 // Figure out which BB this cast is used in. For PHI's this is the
694 // appropriate predecessor block.
695 BasicBlock *UserBB = User->getParent();
696 if (PHINode *PN = dyn_cast<PHINode>(User)) {
697 UserBB = PN->getIncomingBlock(TheUse);
700 // Preincrement use iterator so we don't invalidate it.
703 // If this user is in the same block as the cast, don't change the cast.
704 if (UserBB == DefBB) continue;
706 // If we have already inserted a cast into this block, use it.
707 CastInst *&InsertedCast = InsertedCasts[UserBB];
710 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
712 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "",
717 // Replace a use of the cast with a use of the new cast.
718 TheUse = InsertedCast;
722 // If we removed all uses, nuke the cast.
723 if (CI->use_empty()) {
724 CI->eraseFromParent();
731 /// OptimizeNoopCopyExpression - If the specified cast instruction is a noop
732 /// copy (e.g. it's casting from one pointer type to another, i32->i8 on PPC),
733 /// sink it into user blocks to reduce the number of virtual
734 /// registers that must be created and coalesced.
736 /// Return true if any changes are made.
738 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){
739 // If this is a noop copy,
740 EVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType());
741 EVT DstVT = TLI.getValueType(CI->getType());
743 // This is an fp<->int conversion?
744 if (SrcVT.isInteger() != DstVT.isInteger())
747 // If this is an extension, it will be a zero or sign extension, which
749 if (SrcVT.bitsLT(DstVT)) return false;
751 // If these values will be promoted, find out what they will be promoted
752 // to. This helps us consider truncates on PPC as noop copies when they
754 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
755 TargetLowering::TypePromoteInteger)
756 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
757 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
758 TargetLowering::TypePromoteInteger)
759 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
761 // If, after promotion, these are the same types, this is a noop copy.
768 /// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce
769 /// the number of virtual registers that must be created and coalesced. This is
770 /// a clear win except on targets with multiple condition code registers
771 /// (PowerPC), where it might lose; some adjustment may be wanted there.
773 /// Return true if any changes are made.
774 static bool OptimizeCmpExpression(CmpInst *CI) {
775 BasicBlock *DefBB = CI->getParent();
777 /// InsertedCmp - Only insert a cmp in each block once.
778 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
780 bool MadeChange = false;
781 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
783 Use &TheUse = UI.getUse();
784 Instruction *User = cast<Instruction>(*UI);
786 // Preincrement use iterator so we don't invalidate it.
789 // Don't bother for PHI nodes.
790 if (isa<PHINode>(User))
793 // Figure out which BB this cmp is used in.
794 BasicBlock *UserBB = User->getParent();
796 // If this user is in the same block as the cmp, don't change the cmp.
797 if (UserBB == DefBB) continue;
799 // If we have already inserted a cmp into this block, use it.
800 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
803 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
805 CmpInst::Create(CI->getOpcode(),
806 CI->getPredicate(), CI->getOperand(0),
807 CI->getOperand(1), "", InsertPt);
811 // Replace a use of the cmp with a use of the new cmp.
812 TheUse = InsertedCmp;
816 // If we removed all uses, nuke the cmp.
818 CI->eraseFromParent();
823 /// isExtractBitsCandidateUse - Check if the candidates could
824 /// be combined with shift instruction, which includes:
825 /// 1. Truncate instruction
826 /// 2. And instruction and the imm is a mask of the low bits:
827 /// imm & (imm+1) == 0
828 static bool isExtractBitsCandidateUse(Instruction *User) {
829 if (!isa<TruncInst>(User)) {
830 if (User->getOpcode() != Instruction::And ||
831 !isa<ConstantInt>(User->getOperand(1)))
834 const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
836 if ((Cimm & (Cimm + 1)).getBoolValue())
842 /// SinkShiftAndTruncate - sink both shift and truncate instruction
843 /// to the use of truncate's BB.
845 SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI,
846 DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts,
847 const TargetLowering &TLI) {
848 BasicBlock *UserBB = User->getParent();
849 DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
850 TruncInst *TruncI = dyn_cast<TruncInst>(User);
851 bool MadeChange = false;
853 for (Value::user_iterator TruncUI = TruncI->user_begin(),
854 TruncE = TruncI->user_end();
855 TruncUI != TruncE;) {
857 Use &TruncTheUse = TruncUI.getUse();
858 Instruction *TruncUser = cast<Instruction>(*TruncUI);
859 // Preincrement use iterator so we don't invalidate it.
863 int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
867 // If the use is actually a legal node, there will not be an
868 // implicit truncate.
869 // FIXME: always querying the result type is just an
870 // approximation; some nodes' legality is determined by the
871 // operand or other means. There's no good way to find out though.
872 if (TLI.isOperationLegalOrCustom(
873 ISDOpcode, TLI.getValueType(TruncUser->getType(), true)))
876 // Don't bother for PHI nodes.
877 if (isa<PHINode>(TruncUser))
880 BasicBlock *TruncUserBB = TruncUser->getParent();
882 if (UserBB == TruncUserBB)
885 BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
886 CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
888 if (!InsertedShift && !InsertedTrunc) {
889 BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
891 if (ShiftI->getOpcode() == Instruction::AShr)
893 BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI, "", InsertPt);
896 BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI, "", InsertPt);
899 BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
902 InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
903 TruncI->getType(), "", TruncInsertPt);
907 TruncTheUse = InsertedTrunc;
913 /// OptimizeExtractBits - sink the shift *right* instruction into user blocks if
914 /// the uses could potentially be combined with this shift instruction and
915 /// generate BitExtract instruction. It will only be applied if the architecture
916 /// supports BitExtract instruction. Here is an example:
918 /// %x.extract.shift = lshr i64 %arg1, 32
920 /// %x.extract.trunc = trunc i64 %x.extract.shift to i16
924 /// %x.extract.shift.1 = lshr i64 %arg1, 32
925 /// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
927 /// CodeGen will recoginze the pattern in BB2 and generate BitExtract
929 /// Return true if any changes are made.
930 static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI,
931 const TargetLowering &TLI) {
932 BasicBlock *DefBB = ShiftI->getParent();
934 /// Only insert instructions in each block once.
935 DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts;
937 bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(ShiftI->getType()));
939 bool MadeChange = false;
940 for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
942 Use &TheUse = UI.getUse();
943 Instruction *User = cast<Instruction>(*UI);
944 // Preincrement use iterator so we don't invalidate it.
947 // Don't bother for PHI nodes.
948 if (isa<PHINode>(User))
951 if (!isExtractBitsCandidateUse(User))
954 BasicBlock *UserBB = User->getParent();
956 if (UserBB == DefBB) {
957 // If the shift and truncate instruction are in the same BB. The use of
958 // the truncate(TruncUse) may still introduce another truncate if not
959 // legal. In this case, we would like to sink both shift and truncate
960 // instruction to the BB of TruncUse.
963 // i64 shift.result = lshr i64 opnd, imm
964 // trunc.result = trunc shift.result to i16
967 // ----> We will have an implicit truncate here if the architecture does
968 // not have i16 compare.
969 // cmp i16 trunc.result, opnd2
971 if (isa<TruncInst>(User) && shiftIsLegal
972 // If the type of the truncate is legal, no trucate will be
973 // introduced in other basic blocks.
974 && (!TLI.isTypeLegal(TLI.getValueType(User->getType()))))
976 SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI);
980 // If we have already inserted a shift into this block, use it.
981 BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
983 if (!InsertedShift) {
984 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
986 if (ShiftI->getOpcode() == Instruction::AShr)
988 BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI, "", InsertPt);
991 BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI, "", InsertPt);
996 // Replace a use of the shift with a use of the new shift.
997 TheUse = InsertedShift;
1000 // If we removed all uses, nuke the shift.
1001 if (ShiftI->use_empty())
1002 ShiftI->eraseFromParent();
1007 // ScalarizeMaskedLoad() translates masked load intrinsic, like
1008 // <16 x i32 > @llvm.masked.load( <16 x i32>* %addr, i32 align,
1009 // <16 x i1> %mask, <16 x i32> %passthru)
1010 // to a chain of basic blocks, whith loading element one-by-one if
1011 // the appropriate mask bit is set
1013 // %1 = bitcast i8* %addr to i32*
1014 // %2 = extractelement <16 x i1> %mask, i32 0
1015 // %3 = icmp eq i1 %2, true
1016 // br i1 %3, label %cond.load, label %else
1018 //cond.load: ; preds = %0
1019 // %4 = getelementptr i32* %1, i32 0
1020 // %5 = load i32* %4
1021 // %6 = insertelement <16 x i32> undef, i32 %5, i32 0
1024 //else: ; preds = %0, %cond.load
1025 // %res.phi.else = phi <16 x i32> [ %6, %cond.load ], [ undef, %0 ]
1026 // %7 = extractelement <16 x i1> %mask, i32 1
1027 // %8 = icmp eq i1 %7, true
1028 // br i1 %8, label %cond.load1, label %else2
1030 //cond.load1: ; preds = %else
1031 // %9 = getelementptr i32* %1, i32 1
1032 // %10 = load i32* %9
1033 // %11 = insertelement <16 x i32> %res.phi.else, i32 %10, i32 1
1036 //else2: ; preds = %else, %cond.load1
1037 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
1038 // %12 = extractelement <16 x i1> %mask, i32 2
1039 // %13 = icmp eq i1 %12, true
1040 // br i1 %13, label %cond.load4, label %else5
1042 static void ScalarizeMaskedLoad(CallInst *CI) {
1043 Value *Ptr = CI->getArgOperand(0);
1044 Value *Src0 = CI->getArgOperand(3);
1045 Value *Mask = CI->getArgOperand(2);
1046 VectorType *VecType = dyn_cast<VectorType>(CI->getType());
1047 Type *EltTy = VecType->getElementType();
1049 assert(VecType && "Unexpected return type of masked load intrinsic");
1051 IRBuilder<> Builder(CI->getContext());
1052 Instruction *InsertPt = CI;
1053 BasicBlock *IfBlock = CI->getParent();
1054 BasicBlock *CondBlock = nullptr;
1055 BasicBlock *PrevIfBlock = CI->getParent();
1056 Builder.SetInsertPoint(InsertPt);
1058 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1060 // Bitcast %addr fron i8* to EltTy*
1062 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
1063 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
1064 Value *UndefVal = UndefValue::get(VecType);
1066 // The result vector
1067 Value *VResult = UndefVal;
1069 PHINode *Phi = nullptr;
1070 Value *PrevPhi = UndefVal;
1072 unsigned VectorWidth = VecType->getNumElements();
1073 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1075 // Fill the "else" block, created in the previous iteration
1077 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
1078 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
1079 // %to_load = icmp eq i1 %mask_1, true
1080 // br i1 %to_load, label %cond.load, label %else
1083 Phi = Builder.CreatePHI(VecType, 2, "res.phi.else");
1084 Phi->addIncoming(VResult, CondBlock);
1085 Phi->addIncoming(PrevPhi, PrevIfBlock);
1090 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
1091 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1092 ConstantInt::get(Predicate->getType(), 1));
1094 // Create "cond" block
1096 // %EltAddr = getelementptr i32* %1, i32 0
1097 // %Elt = load i32* %EltAddr
1098 // VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx
1100 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.load");
1101 Builder.SetInsertPoint(InsertPt);
1103 Value* Gep = Builder.CreateInBoundsGEP(FirstEltPtr, Builder.getInt32(Idx));
1104 LoadInst* Load = Builder.CreateLoad(Gep, false);
1105 VResult = Builder.CreateInsertElement(VResult, Load, Builder.getInt32(Idx));
1107 // Create "else" block, fill it in the next iteration
1108 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1109 Builder.SetInsertPoint(InsertPt);
1110 Instruction *OldBr = IfBlock->getTerminator();
1111 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1112 OldBr->eraseFromParent();
1113 PrevIfBlock = IfBlock;
1114 IfBlock = NewIfBlock;
1117 Phi = Builder.CreatePHI(VecType, 2, "res.phi.select");
1118 Phi->addIncoming(VResult, CondBlock);
1119 Phi->addIncoming(PrevPhi, PrevIfBlock);
1120 Value *NewI = Builder.CreateSelect(Mask, Phi, Src0);
1121 CI->replaceAllUsesWith(NewI);
1122 CI->eraseFromParent();
1125 // ScalarizeMaskedStore() translates masked store intrinsic, like
1126 // void @llvm.masked.store(<16 x i32> %src, <16 x i32>* %addr, i32 align,
1128 // to a chain of basic blocks, that stores element one-by-one if
1129 // the appropriate mask bit is set
1131 // %1 = bitcast i8* %addr to i32*
1132 // %2 = extractelement <16 x i1> %mask, i32 0
1133 // %3 = icmp eq i1 %2, true
1134 // br i1 %3, label %cond.store, label %else
1136 // cond.store: ; preds = %0
1137 // %4 = extractelement <16 x i32> %val, i32 0
1138 // %5 = getelementptr i32* %1, i32 0
1139 // store i32 %4, i32* %5
1142 // else: ; preds = %0, %cond.store
1143 // %6 = extractelement <16 x i1> %mask, i32 1
1144 // %7 = icmp eq i1 %6, true
1145 // br i1 %7, label %cond.store1, label %else2
1147 // cond.store1: ; preds = %else
1148 // %8 = extractelement <16 x i32> %val, i32 1
1149 // %9 = getelementptr i32* %1, i32 1
1150 // store i32 %8, i32* %9
1153 static void ScalarizeMaskedStore(CallInst *CI) {
1154 Value *Ptr = CI->getArgOperand(1);
1155 Value *Src = CI->getArgOperand(0);
1156 Value *Mask = CI->getArgOperand(3);
1158 VectorType *VecType = dyn_cast<VectorType>(Src->getType());
1159 Type *EltTy = VecType->getElementType();
1161 assert(VecType && "Unexpected data type in masked store intrinsic");
1163 IRBuilder<> Builder(CI->getContext());
1164 Instruction *InsertPt = CI;
1165 BasicBlock *IfBlock = CI->getParent();
1166 Builder.SetInsertPoint(InsertPt);
1167 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1169 // Bitcast %addr fron i8* to EltTy*
1171 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
1172 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
1174 unsigned VectorWidth = VecType->getNumElements();
1175 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1177 // Fill the "else" block, created in the previous iteration
1179 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
1180 // %to_store = icmp eq i1 %mask_1, true
1181 // br i1 %to_load, label %cond.store, label %else
1183 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
1184 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1185 ConstantInt::get(Predicate->getType(), 1));
1187 // Create "cond" block
1189 // %OneElt = extractelement <16 x i32> %Src, i32 Idx
1190 // %EltAddr = getelementptr i32* %1, i32 0
1191 // %store i32 %OneElt, i32* %EltAddr
1193 BasicBlock *CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1194 Builder.SetInsertPoint(InsertPt);
1196 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx));
1197 Value* Gep = Builder.CreateInBoundsGEP(FirstEltPtr, Builder.getInt32(Idx));
1198 Builder.CreateStore(OneElt, Gep);
1200 // Create "else" block, fill it in the next iteration
1201 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1202 Builder.SetInsertPoint(InsertPt);
1203 Instruction *OldBr = IfBlock->getTerminator();
1204 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1205 OldBr->eraseFromParent();
1206 IfBlock = NewIfBlock;
1208 CI->eraseFromParent();
1211 bool CodeGenPrepare::OptimizeCallInst(CallInst *CI, bool& ModifiedDT) {
1212 BasicBlock *BB = CI->getParent();
1214 // Lower inline assembly if we can.
1215 // If we found an inline asm expession, and if the target knows how to
1216 // lower it to normal LLVM code, do so now.
1217 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
1218 if (TLI->ExpandInlineAsm(CI)) {
1219 // Avoid invalidating the iterator.
1220 CurInstIterator = BB->begin();
1221 // Avoid processing instructions out of order, which could cause
1222 // reuse before a value is defined.
1226 // Sink address computing for memory operands into the block.
1227 if (OptimizeInlineAsmInst(CI))
1231 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
1233 switch (II->getIntrinsicID()) {
1235 case Intrinsic::objectsize: {
1236 // Lower all uses of llvm.objectsize.*
1237 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
1238 Type *ReturnTy = CI->getType();
1239 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
1241 // Substituting this can cause recursive simplifications, which can
1242 // invalidate our iterator. Use a WeakVH to hold onto it in case this
1244 WeakVH IterHandle(CurInstIterator);
1246 replaceAndRecursivelySimplify(CI, RetVal,
1247 TLI ? TLI->getDataLayout() : nullptr,
1248 TLInfo, ModifiedDT ? nullptr : DT);
1250 // If the iterator instruction was recursively deleted, start over at the
1251 // start of the block.
1252 if (IterHandle != CurInstIterator) {
1253 CurInstIterator = BB->begin();
1258 case Intrinsic::masked_load: {
1259 // Scalarize unsupported vector masked load
1260 if (!TTI->isLegalMaskedLoad(CI->getType(), 1)) {
1261 ScalarizeMaskedLoad(CI);
1267 case Intrinsic::masked_store: {
1268 if (!TTI->isLegalMaskedStore(CI->getArgOperand(0)->getType(), 1)) {
1269 ScalarizeMaskedStore(CI);
1278 SmallVector<Value*, 2> PtrOps;
1280 if (TLI->GetAddrModeArguments(II, PtrOps, AccessTy))
1281 while (!PtrOps.empty())
1282 if (OptimizeMemoryInst(II, PtrOps.pop_back_val(), AccessTy))
1287 // From here on out we're working with named functions.
1288 if (!CI->getCalledFunction()) return false;
1290 // We'll need DataLayout from here on out.
1291 const DataLayout *TD = TLI ? TLI->getDataLayout() : nullptr;
1292 if (!TD) return false;
1294 // Lower all default uses of _chk calls. This is very similar
1295 // to what InstCombineCalls does, but here we are only lowering calls
1296 // to fortified library functions (e.g. __memcpy_chk) that have the default
1297 // "don't know" as the objectsize. Anything else should be left alone.
1298 FortifiedLibCallSimplifier Simplifier(TD, TLInfo, true);
1299 if (Value *V = Simplifier.optimizeCall(CI)) {
1300 CI->replaceAllUsesWith(V);
1301 CI->eraseFromParent();
1307 /// DupRetToEnableTailCallOpts - Look for opportunities to duplicate return
1308 /// instructions to the predecessor to enable tail call optimizations. The
1309 /// case it is currently looking for is:
1312 /// %tmp0 = tail call i32 @f0()
1313 /// br label %return
1315 /// %tmp1 = tail call i32 @f1()
1316 /// br label %return
1318 /// %tmp2 = tail call i32 @f2()
1319 /// br label %return
1321 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
1329 /// %tmp0 = tail call i32 @f0()
1332 /// %tmp1 = tail call i32 @f1()
1335 /// %tmp2 = tail call i32 @f2()
1338 bool CodeGenPrepare::DupRetToEnableTailCallOpts(BasicBlock *BB) {
1342 ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator());
1346 PHINode *PN = nullptr;
1347 BitCastInst *BCI = nullptr;
1348 Value *V = RI->getReturnValue();
1350 BCI = dyn_cast<BitCastInst>(V);
1352 V = BCI->getOperand(0);
1354 PN = dyn_cast<PHINode>(V);
1359 if (PN && PN->getParent() != BB)
1362 // It's not safe to eliminate the sign / zero extension of the return value.
1363 // See llvm::isInTailCallPosition().
1364 const Function *F = BB->getParent();
1365 AttributeSet CallerAttrs = F->getAttributes();
1366 if (CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt) ||
1367 CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
1370 // Make sure there are no instructions between the PHI and return, or that the
1371 // return is the first instruction in the block.
1373 BasicBlock::iterator BI = BB->begin();
1374 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
1376 // Also skip over the bitcast.
1381 BasicBlock::iterator BI = BB->begin();
1382 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
1387 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
1389 SmallVector<CallInst*, 4> TailCalls;
1391 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
1392 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
1393 // Make sure the phi value is indeed produced by the tail call.
1394 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
1395 TLI->mayBeEmittedAsTailCall(CI))
1396 TailCalls.push_back(CI);
1399 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
1400 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
1401 if (!VisitedBBs.insert(*PI).second)
1404 BasicBlock::InstListType &InstList = (*PI)->getInstList();
1405 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
1406 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
1407 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
1411 CallInst *CI = dyn_cast<CallInst>(&*RI);
1412 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
1413 TailCalls.push_back(CI);
1417 bool Changed = false;
1418 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
1419 CallInst *CI = TailCalls[i];
1422 // Conservatively require the attributes of the call to match those of the
1423 // return. Ignore noalias because it doesn't affect the call sequence.
1424 AttributeSet CalleeAttrs = CS.getAttributes();
1425 if (AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1426 removeAttribute(Attribute::NoAlias) !=
1427 AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1428 removeAttribute(Attribute::NoAlias))
1431 // Make sure the call instruction is followed by an unconditional branch to
1432 // the return block.
1433 BasicBlock *CallBB = CI->getParent();
1434 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
1435 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
1438 // Duplicate the return into CallBB.
1439 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
1440 ModifiedDT = Changed = true;
1444 // If we eliminated all predecessors of the block, delete the block now.
1445 if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
1446 BB->eraseFromParent();
1451 //===----------------------------------------------------------------------===//
1452 // Memory Optimization
1453 //===----------------------------------------------------------------------===//
1457 /// ExtAddrMode - This is an extended version of TargetLowering::AddrMode
1458 /// which holds actual Value*'s for register values.
1459 struct ExtAddrMode : public TargetLowering::AddrMode {
1462 ExtAddrMode() : BaseReg(nullptr), ScaledReg(nullptr) {}
1463 void print(raw_ostream &OS) const;
1466 bool operator==(const ExtAddrMode& O) const {
1467 return (BaseReg == O.BaseReg) && (ScaledReg == O.ScaledReg) &&
1468 (BaseGV == O.BaseGV) && (BaseOffs == O.BaseOffs) &&
1469 (HasBaseReg == O.HasBaseReg) && (Scale == O.Scale);
1474 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
1480 void ExtAddrMode::print(raw_ostream &OS) const {
1481 bool NeedPlus = false;
1484 OS << (NeedPlus ? " + " : "")
1486 BaseGV->printAsOperand(OS, /*PrintType=*/false);
1491 OS << (NeedPlus ? " + " : "")
1497 OS << (NeedPlus ? " + " : "")
1499 BaseReg->printAsOperand(OS, /*PrintType=*/false);
1503 OS << (NeedPlus ? " + " : "")
1505 ScaledReg->printAsOperand(OS, /*PrintType=*/false);
1511 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1512 void ExtAddrMode::dump() const {
1518 /// \brief This class provides transaction based operation on the IR.
1519 /// Every change made through this class is recorded in the internal state and
1520 /// can be undone (rollback) until commit is called.
1521 class TypePromotionTransaction {
1523 /// \brief This represents the common interface of the individual transaction.
1524 /// Each class implements the logic for doing one specific modification on
1525 /// the IR via the TypePromotionTransaction.
1526 class TypePromotionAction {
1528 /// The Instruction modified.
1532 /// \brief Constructor of the action.
1533 /// The constructor performs the related action on the IR.
1534 TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
1536 virtual ~TypePromotionAction() {}
1538 /// \brief Undo the modification done by this action.
1539 /// When this method is called, the IR must be in the same state as it was
1540 /// before this action was applied.
1541 /// \pre Undoing the action works if and only if the IR is in the exact same
1542 /// state as it was directly after this action was applied.
1543 virtual void undo() = 0;
1545 /// \brief Advocate every change made by this action.
1546 /// When the results on the IR of the action are to be kept, it is important
1547 /// to call this function, otherwise hidden information may be kept forever.
1548 virtual void commit() {
1549 // Nothing to be done, this action is not doing anything.
1553 /// \brief Utility to remember the position of an instruction.
1554 class InsertionHandler {
1555 /// Position of an instruction.
1556 /// Either an instruction:
1557 /// - Is the first in a basic block: BB is used.
1558 /// - Has a previous instructon: PrevInst is used.
1560 Instruction *PrevInst;
1563 /// Remember whether or not the instruction had a previous instruction.
1564 bool HasPrevInstruction;
1567 /// \brief Record the position of \p Inst.
1568 InsertionHandler(Instruction *Inst) {
1569 BasicBlock::iterator It = Inst;
1570 HasPrevInstruction = (It != (Inst->getParent()->begin()));
1571 if (HasPrevInstruction)
1572 Point.PrevInst = --It;
1574 Point.BB = Inst->getParent();
1577 /// \brief Insert \p Inst at the recorded position.
1578 void insert(Instruction *Inst) {
1579 if (HasPrevInstruction) {
1580 if (Inst->getParent())
1581 Inst->removeFromParent();
1582 Inst->insertAfter(Point.PrevInst);
1584 Instruction *Position = Point.BB->getFirstInsertionPt();
1585 if (Inst->getParent())
1586 Inst->moveBefore(Position);
1588 Inst->insertBefore(Position);
1593 /// \brief Move an instruction before another.
1594 class InstructionMoveBefore : public TypePromotionAction {
1595 /// Original position of the instruction.
1596 InsertionHandler Position;
1599 /// \brief Move \p Inst before \p Before.
1600 InstructionMoveBefore(Instruction *Inst, Instruction *Before)
1601 : TypePromotionAction(Inst), Position(Inst) {
1602 DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before << "\n");
1603 Inst->moveBefore(Before);
1606 /// \brief Move the instruction back to its original position.
1607 void undo() override {
1608 DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
1609 Position.insert(Inst);
1613 /// \brief Set the operand of an instruction with a new value.
1614 class OperandSetter : public TypePromotionAction {
1615 /// Original operand of the instruction.
1617 /// Index of the modified instruction.
1621 /// \brief Set \p Idx operand of \p Inst with \p NewVal.
1622 OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
1623 : TypePromotionAction(Inst), Idx(Idx) {
1624 DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
1625 << "for:" << *Inst << "\n"
1626 << "with:" << *NewVal << "\n");
1627 Origin = Inst->getOperand(Idx);
1628 Inst->setOperand(Idx, NewVal);
1631 /// \brief Restore the original value of the instruction.
1632 void undo() override {
1633 DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
1634 << "for: " << *Inst << "\n"
1635 << "with: " << *Origin << "\n");
1636 Inst->setOperand(Idx, Origin);
1640 /// \brief Hide the operands of an instruction.
1641 /// Do as if this instruction was not using any of its operands.
1642 class OperandsHider : public TypePromotionAction {
1643 /// The list of original operands.
1644 SmallVector<Value *, 4> OriginalValues;
1647 /// \brief Remove \p Inst from the uses of the operands of \p Inst.
1648 OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
1649 DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
1650 unsigned NumOpnds = Inst->getNumOperands();
1651 OriginalValues.reserve(NumOpnds);
1652 for (unsigned It = 0; It < NumOpnds; ++It) {
1653 // Save the current operand.
1654 Value *Val = Inst->getOperand(It);
1655 OriginalValues.push_back(Val);
1657 // We could use OperandSetter here, but that would implied an overhead
1658 // that we are not willing to pay.
1659 Inst->setOperand(It, UndefValue::get(Val->getType()));
1663 /// \brief Restore the original list of uses.
1664 void undo() override {
1665 DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
1666 for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
1667 Inst->setOperand(It, OriginalValues[It]);
1671 /// \brief Build a truncate instruction.
1672 class TruncBuilder : public TypePromotionAction {
1675 /// \brief Build a truncate instruction of \p Opnd producing a \p Ty
1677 /// trunc Opnd to Ty.
1678 TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
1679 IRBuilder<> Builder(Opnd);
1680 Val = Builder.CreateTrunc(Opnd, Ty, "promoted");
1681 DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n");
1684 /// \brief Get the built value.
1685 Value *getBuiltValue() { return Val; }
1687 /// \brief Remove the built instruction.
1688 void undo() override {
1689 DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n");
1690 if (Instruction *IVal = dyn_cast<Instruction>(Val))
1691 IVal->eraseFromParent();
1695 /// \brief Build a sign extension instruction.
1696 class SExtBuilder : public TypePromotionAction {
1699 /// \brief Build a sign extension instruction of \p Opnd producing a \p Ty
1701 /// sext Opnd to Ty.
1702 SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
1703 : TypePromotionAction(InsertPt) {
1704 IRBuilder<> Builder(InsertPt);
1705 Val = Builder.CreateSExt(Opnd, Ty, "promoted");
1706 DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n");
1709 /// \brief Get the built value.
1710 Value *getBuiltValue() { return Val; }
1712 /// \brief Remove the built instruction.
1713 void undo() override {
1714 DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n");
1715 if (Instruction *IVal = dyn_cast<Instruction>(Val))
1716 IVal->eraseFromParent();
1720 /// \brief Build a zero extension instruction.
1721 class ZExtBuilder : public TypePromotionAction {
1724 /// \brief Build a zero extension instruction of \p Opnd producing a \p Ty
1726 /// zext Opnd to Ty.
1727 ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
1728 : TypePromotionAction(InsertPt) {
1729 IRBuilder<> Builder(InsertPt);
1730 Val = Builder.CreateZExt(Opnd, Ty, "promoted");
1731 DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n");
1734 /// \brief Get the built value.
1735 Value *getBuiltValue() { return Val; }
1737 /// \brief Remove the built instruction.
1738 void undo() override {
1739 DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n");
1740 if (Instruction *IVal = dyn_cast<Instruction>(Val))
1741 IVal->eraseFromParent();
1745 /// \brief Mutate an instruction to another type.
1746 class TypeMutator : public TypePromotionAction {
1747 /// Record the original type.
1751 /// \brief Mutate the type of \p Inst into \p NewTy.
1752 TypeMutator(Instruction *Inst, Type *NewTy)
1753 : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
1754 DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
1756 Inst->mutateType(NewTy);
1759 /// \brief Mutate the instruction back to its original type.
1760 void undo() override {
1761 DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
1763 Inst->mutateType(OrigTy);
1767 /// \brief Replace the uses of an instruction by another instruction.
1768 class UsesReplacer : public TypePromotionAction {
1769 /// Helper structure to keep track of the replaced uses.
1770 struct InstructionAndIdx {
1771 /// The instruction using the instruction.
1773 /// The index where this instruction is used for Inst.
1775 InstructionAndIdx(Instruction *Inst, unsigned Idx)
1776 : Inst(Inst), Idx(Idx) {}
1779 /// Keep track of the original uses (pair Instruction, Index).
1780 SmallVector<InstructionAndIdx, 4> OriginalUses;
1781 typedef SmallVectorImpl<InstructionAndIdx>::iterator use_iterator;
1784 /// \brief Replace all the use of \p Inst by \p New.
1785 UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
1786 DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
1788 // Record the original uses.
1789 for (Use &U : Inst->uses()) {
1790 Instruction *UserI = cast<Instruction>(U.getUser());
1791 OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
1793 // Now, we can replace the uses.
1794 Inst->replaceAllUsesWith(New);
1797 /// \brief Reassign the original uses of Inst to Inst.
1798 void undo() override {
1799 DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
1800 for (use_iterator UseIt = OriginalUses.begin(),
1801 EndIt = OriginalUses.end();
1802 UseIt != EndIt; ++UseIt) {
1803 UseIt->Inst->setOperand(UseIt->Idx, Inst);
1808 /// \brief Remove an instruction from the IR.
1809 class InstructionRemover : public TypePromotionAction {
1810 /// Original position of the instruction.
1811 InsertionHandler Inserter;
1812 /// Helper structure to hide all the link to the instruction. In other
1813 /// words, this helps to do as if the instruction was removed.
1814 OperandsHider Hider;
1815 /// Keep track of the uses replaced, if any.
1816 UsesReplacer *Replacer;
1819 /// \brief Remove all reference of \p Inst and optinally replace all its
1821 /// \pre If !Inst->use_empty(), then New != nullptr
1822 InstructionRemover(Instruction *Inst, Value *New = nullptr)
1823 : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
1826 Replacer = new UsesReplacer(Inst, New);
1827 DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
1828 Inst->removeFromParent();
1831 ~InstructionRemover() { delete Replacer; }
1833 /// \brief Really remove the instruction.
1834 void commit() override { delete Inst; }
1836 /// \brief Resurrect the instruction and reassign it to the proper uses if
1837 /// new value was provided when build this action.
1838 void undo() override {
1839 DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
1840 Inserter.insert(Inst);
1848 /// Restoration point.
1849 /// The restoration point is a pointer to an action instead of an iterator
1850 /// because the iterator may be invalidated but not the pointer.
1851 typedef const TypePromotionAction *ConstRestorationPt;
1852 /// Advocate every changes made in that transaction.
1854 /// Undo all the changes made after the given point.
1855 void rollback(ConstRestorationPt Point);
1856 /// Get the current restoration point.
1857 ConstRestorationPt getRestorationPoint() const;
1859 /// \name API for IR modification with state keeping to support rollback.
1861 /// Same as Instruction::setOperand.
1862 void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
1863 /// Same as Instruction::eraseFromParent.
1864 void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
1865 /// Same as Value::replaceAllUsesWith.
1866 void replaceAllUsesWith(Instruction *Inst, Value *New);
1867 /// Same as Value::mutateType.
1868 void mutateType(Instruction *Inst, Type *NewTy);
1869 /// Same as IRBuilder::createTrunc.
1870 Value *createTrunc(Instruction *Opnd, Type *Ty);
1871 /// Same as IRBuilder::createSExt.
1872 Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
1873 /// Same as IRBuilder::createZExt.
1874 Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
1875 /// Same as Instruction::moveBefore.
1876 void moveBefore(Instruction *Inst, Instruction *Before);
1880 /// The ordered list of actions made so far.
1881 SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions;
1882 typedef SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator CommitPt;
1885 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
1888 make_unique<TypePromotionTransaction::OperandSetter>(Inst, Idx, NewVal));
1891 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
1894 make_unique<TypePromotionTransaction::InstructionRemover>(Inst, NewVal));
1897 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
1899 Actions.push_back(make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
1902 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
1903 Actions.push_back(make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
1906 Value *TypePromotionTransaction::createTrunc(Instruction *Opnd,
1908 std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
1909 Value *Val = Ptr->getBuiltValue();
1910 Actions.push_back(std::move(Ptr));
1914 Value *TypePromotionTransaction::createSExt(Instruction *Inst,
1915 Value *Opnd, Type *Ty) {
1916 std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
1917 Value *Val = Ptr->getBuiltValue();
1918 Actions.push_back(std::move(Ptr));
1922 Value *TypePromotionTransaction::createZExt(Instruction *Inst,
1923 Value *Opnd, Type *Ty) {
1924 std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
1925 Value *Val = Ptr->getBuiltValue();
1926 Actions.push_back(std::move(Ptr));
1930 void TypePromotionTransaction::moveBefore(Instruction *Inst,
1931 Instruction *Before) {
1933 make_unique<TypePromotionTransaction::InstructionMoveBefore>(Inst, Before));
1936 TypePromotionTransaction::ConstRestorationPt
1937 TypePromotionTransaction::getRestorationPoint() const {
1938 return !Actions.empty() ? Actions.back().get() : nullptr;
1941 void TypePromotionTransaction::commit() {
1942 for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
1948 void TypePromotionTransaction::rollback(
1949 TypePromotionTransaction::ConstRestorationPt Point) {
1950 while (!Actions.empty() && Point != Actions.back().get()) {
1951 std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
1956 /// \brief A helper class for matching addressing modes.
1958 /// This encapsulates the logic for matching the target-legal addressing modes.
1959 class AddressingModeMatcher {
1960 SmallVectorImpl<Instruction*> &AddrModeInsts;
1961 const TargetMachine &TM;
1962 const TargetLowering &TLI;
1964 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
1965 /// the memory instruction that we're computing this address for.
1967 Instruction *MemoryInst;
1969 /// AddrMode - This is the addressing mode that we're building up. This is
1970 /// part of the return value of this addressing mode matching stuff.
1971 ExtAddrMode &AddrMode;
1973 /// The truncate instruction inserted by other CodeGenPrepare optimizations.
1974 const SetOfInstrs &InsertedTruncs;
1975 /// A map from the instructions to their type before promotion.
1976 InstrToOrigTy &PromotedInsts;
1977 /// The ongoing transaction where every action should be registered.
1978 TypePromotionTransaction &TPT;
1980 /// IgnoreProfitability - This is set to true when we should not do
1981 /// profitability checks. When true, IsProfitableToFoldIntoAddressingMode
1982 /// always returns true.
1983 bool IgnoreProfitability;
1985 AddressingModeMatcher(SmallVectorImpl<Instruction *> &AMI,
1986 const TargetMachine &TM, Type *AT, Instruction *MI,
1987 ExtAddrMode &AM, const SetOfInstrs &InsertedTruncs,
1988 InstrToOrigTy &PromotedInsts,
1989 TypePromotionTransaction &TPT)
1990 : AddrModeInsts(AMI), TM(TM),
1991 TLI(*TM.getSubtargetImpl(*MI->getParent()->getParent())
1992 ->getTargetLowering()),
1993 AccessTy(AT), MemoryInst(MI), AddrMode(AM),
1994 InsertedTruncs(InsertedTruncs), PromotedInsts(PromotedInsts), TPT(TPT) {
1995 IgnoreProfitability = false;
1999 /// Match - Find the maximal addressing mode that a load/store of V can fold,
2000 /// give an access type of AccessTy. This returns a list of involved
2001 /// instructions in AddrModeInsts.
2002 /// \p InsertedTruncs The truncate instruction inserted by other
2005 /// \p PromotedInsts maps the instructions to their type before promotion.
2006 /// \p The ongoing transaction where every action should be registered.
2007 static ExtAddrMode Match(Value *V, Type *AccessTy,
2008 Instruction *MemoryInst,
2009 SmallVectorImpl<Instruction*> &AddrModeInsts,
2010 const TargetMachine &TM,
2011 const SetOfInstrs &InsertedTruncs,
2012 InstrToOrigTy &PromotedInsts,
2013 TypePromotionTransaction &TPT) {
2016 bool Success = AddressingModeMatcher(AddrModeInsts, TM, AccessTy,
2017 MemoryInst, Result, InsertedTruncs,
2018 PromotedInsts, TPT).MatchAddr(V, 0);
2019 (void)Success; assert(Success && "Couldn't select *anything*?");
2023 bool MatchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
2024 bool MatchAddr(Value *V, unsigned Depth);
2025 bool MatchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth,
2026 bool *MovedAway = nullptr);
2027 bool IsProfitableToFoldIntoAddressingMode(Instruction *I,
2028 ExtAddrMode &AMBefore,
2029 ExtAddrMode &AMAfter);
2030 bool ValueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
2031 bool IsPromotionProfitable(unsigned MatchedSize, unsigned SizeWithPromotion,
2032 Value *PromotedOperand) const;
2035 /// MatchScaledValue - Try adding ScaleReg*Scale to the current addressing mode.
2036 /// Return true and update AddrMode if this addr mode is legal for the target,
2038 bool AddressingModeMatcher::MatchScaledValue(Value *ScaleReg, int64_t Scale,
2040 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
2041 // mode. Just process that directly.
2043 return MatchAddr(ScaleReg, Depth);
2045 // If the scale is 0, it takes nothing to add this.
2049 // If we already have a scale of this value, we can add to it, otherwise, we
2050 // need an available scale field.
2051 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
2054 ExtAddrMode TestAddrMode = AddrMode;
2056 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
2057 // [A+B + A*7] -> [B+A*8].
2058 TestAddrMode.Scale += Scale;
2059 TestAddrMode.ScaledReg = ScaleReg;
2061 // If the new address isn't legal, bail out.
2062 if (!TLI.isLegalAddressingMode(TestAddrMode, AccessTy))
2065 // It was legal, so commit it.
2066 AddrMode = TestAddrMode;
2068 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
2069 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
2070 // X*Scale + C*Scale to addr mode.
2071 ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
2072 if (isa<Instruction>(ScaleReg) && // not a constant expr.
2073 match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
2074 TestAddrMode.ScaledReg = AddLHS;
2075 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
2077 // If this addressing mode is legal, commit it and remember that we folded
2078 // this instruction.
2079 if (TLI.isLegalAddressingMode(TestAddrMode, AccessTy)) {
2080 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
2081 AddrMode = TestAddrMode;
2086 // Otherwise, not (x+c)*scale, just return what we have.
2090 /// MightBeFoldableInst - This is a little filter, which returns true if an
2091 /// addressing computation involving I might be folded into a load/store
2092 /// accessing it. This doesn't need to be perfect, but needs to accept at least
2093 /// the set of instructions that MatchOperationAddr can.
2094 static bool MightBeFoldableInst(Instruction *I) {
2095 switch (I->getOpcode()) {
2096 case Instruction::BitCast:
2097 case Instruction::AddrSpaceCast:
2098 // Don't touch identity bitcasts.
2099 if (I->getType() == I->getOperand(0)->getType())
2101 return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
2102 case Instruction::PtrToInt:
2103 // PtrToInt is always a noop, as we know that the int type is pointer sized.
2105 case Instruction::IntToPtr:
2106 // We know the input is intptr_t, so this is foldable.
2108 case Instruction::Add:
2110 case Instruction::Mul:
2111 case Instruction::Shl:
2112 // Can only handle X*C and X << C.
2113 return isa<ConstantInt>(I->getOperand(1));
2114 case Instruction::GetElementPtr:
2121 /// \brief Check whether or not \p Val is a legal instruction for \p TLI.
2122 /// \note \p Val is assumed to be the product of some type promotion.
2123 /// Therefore if \p Val has an undefined state in \p TLI, this is assumed
2124 /// to be legal, as the non-promoted value would have had the same state.
2125 static bool isPromotedInstructionLegal(const TargetLowering &TLI, Value *Val) {
2126 Instruction *PromotedInst = dyn_cast<Instruction>(Val);
2129 int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
2130 // If the ISDOpcode is undefined, it was undefined before the promotion.
2133 // Otherwise, check if the promoted instruction is legal or not.
2134 return TLI.isOperationLegalOrCustom(
2135 ISDOpcode, TLI.getValueType(PromotedInst->getType()));
2138 /// \brief Hepler class to perform type promotion.
2139 class TypePromotionHelper {
2140 /// \brief Utility function to check whether or not a sign or zero extension
2141 /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
2142 /// either using the operands of \p Inst or promoting \p Inst.
2143 /// The type of the extension is defined by \p IsSExt.
2144 /// In other words, check if:
2145 /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
2146 /// #1 Promotion applies:
2147 /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
2148 /// #2 Operand reuses:
2149 /// ext opnd1 to ConsideredExtType.
2150 /// \p PromotedInsts maps the instructions to their type before promotion.
2151 static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
2152 const InstrToOrigTy &PromotedInsts, bool IsSExt);
2154 /// \brief Utility function to determine if \p OpIdx should be promoted when
2155 /// promoting \p Inst.
2156 static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
2157 if (isa<SelectInst>(Inst) && OpIdx == 0)
2162 /// \brief Utility function to promote the operand of \p Ext when this
2163 /// operand is a promotable trunc or sext or zext.
2164 /// \p PromotedInsts maps the instructions to their type before promotion.
2165 /// \p CreatedInsts[out] contains how many non-free instructions have been
2166 /// created to promote the operand of Ext.
2167 /// Newly added extensions are inserted in \p Exts.
2168 /// Newly added truncates are inserted in \p Truncs.
2169 /// Should never be called directly.
2170 /// \return The promoted value which is used instead of Ext.
2171 static Value *promoteOperandForTruncAndAnyExt(
2172 Instruction *Ext, TypePromotionTransaction &TPT,
2173 InstrToOrigTy &PromotedInsts, unsigned &CreatedInsts,
2174 SmallVectorImpl<Instruction *> *Exts,
2175 SmallVectorImpl<Instruction *> *Truncs);
2177 /// \brief Utility function to promote the operand of \p Ext when this
2178 /// operand is promotable and is not a supported trunc or sext.
2179 /// \p PromotedInsts maps the instructions to their type before promotion.
2180 /// \p CreatedInsts[out] contains how many non-free instructions have been
2181 /// created to promote the operand of Ext.
2182 /// Newly added extensions are inserted in \p Exts.
2183 /// Newly added truncates are inserted in \p Truncs.
2184 /// Should never be called directly.
2185 /// \return The promoted value which is used instead of Ext.
2187 promoteOperandForOther(Instruction *Ext, TypePromotionTransaction &TPT,
2188 InstrToOrigTy &PromotedInsts, unsigned &CreatedInsts,
2189 SmallVectorImpl<Instruction *> *Exts,
2190 SmallVectorImpl<Instruction *> *Truncs, bool IsSExt);
2192 /// \see promoteOperandForOther.
2194 signExtendOperandForOther(Instruction *Ext, TypePromotionTransaction &TPT,
2195 InstrToOrigTy &PromotedInsts,
2196 unsigned &CreatedInsts,
2197 SmallVectorImpl<Instruction *> *Exts,
2198 SmallVectorImpl<Instruction *> *Truncs) {
2199 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInsts, Exts,
2203 /// \see promoteOperandForOther.
2205 zeroExtendOperandForOther(Instruction *Ext, TypePromotionTransaction &TPT,
2206 InstrToOrigTy &PromotedInsts,
2207 unsigned &CreatedInsts,
2208 SmallVectorImpl<Instruction *> *Exts,
2209 SmallVectorImpl<Instruction *> *Truncs) {
2210 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInsts, Exts,
2215 /// Type for the utility function that promotes the operand of Ext.
2216 typedef Value *(*Action)(Instruction *Ext, TypePromotionTransaction &TPT,
2217 InstrToOrigTy &PromotedInsts, unsigned &CreatedInsts,
2218 SmallVectorImpl<Instruction *> *Exts,
2219 SmallVectorImpl<Instruction *> *Truncs);
2220 /// \brief Given a sign/zero extend instruction \p Ext, return the approriate
2221 /// action to promote the operand of \p Ext instead of using Ext.
2222 /// \return NULL if no promotable action is possible with the current
2224 /// \p InsertedTruncs keeps track of all the truncate instructions inserted by
2225 /// the others CodeGenPrepare optimizations. This information is important
2226 /// because we do not want to promote these instructions as CodeGenPrepare
2227 /// will reinsert them later. Thus creating an infinite loop: create/remove.
2228 /// \p PromotedInsts maps the instructions to their type before promotion.
2229 static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedTruncs,
2230 const TargetLowering &TLI,
2231 const InstrToOrigTy &PromotedInsts);
2234 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
2235 Type *ConsideredExtType,
2236 const InstrToOrigTy &PromotedInsts,
2238 // The promotion helper does not know how to deal with vector types yet.
2239 // To be able to fix that, we would need to fix the places where we
2240 // statically extend, e.g., constants and such.
2241 if (Inst->getType()->isVectorTy())
2244 // We can always get through zext.
2245 if (isa<ZExtInst>(Inst))
2248 // sext(sext) is ok too.
2249 if (IsSExt && isa<SExtInst>(Inst))
2252 // We can get through binary operator, if it is legal. In other words, the
2253 // binary operator must have a nuw or nsw flag.
2254 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
2255 if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
2256 ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
2257 (IsSExt && BinOp->hasNoSignedWrap())))
2260 // Check if we can do the following simplification.
2261 // ext(trunc(opnd)) --> ext(opnd)
2262 if (!isa<TruncInst>(Inst))
2265 Value *OpndVal = Inst->getOperand(0);
2266 // Check if we can use this operand in the extension.
2267 // If the type is larger than the result type of the extension,
2269 if (!OpndVal->getType()->isIntegerTy() ||
2270 OpndVal->getType()->getIntegerBitWidth() >
2271 ConsideredExtType->getIntegerBitWidth())
2274 // If the operand of the truncate is not an instruction, we will not have
2275 // any information on the dropped bits.
2276 // (Actually we could for constant but it is not worth the extra logic).
2277 Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
2281 // Check if the source of the type is narrow enough.
2282 // I.e., check that trunc just drops extended bits of the same kind of
2284 // #1 get the type of the operand and check the kind of the extended bits.
2285 const Type *OpndType;
2286 InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
2287 if (It != PromotedInsts.end() && It->second.IsSExt == IsSExt)
2288 OpndType = It->second.Ty;
2289 else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd)))
2290 OpndType = Opnd->getOperand(0)->getType();
2294 // #2 check that the truncate just drop extended bits.
2295 if (Inst->getType()->getIntegerBitWidth() >= OpndType->getIntegerBitWidth())
2301 TypePromotionHelper::Action TypePromotionHelper::getAction(
2302 Instruction *Ext, const SetOfInstrs &InsertedTruncs,
2303 const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
2304 assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
2305 "Unexpected instruction type");
2306 Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0));
2307 Type *ExtTy = Ext->getType();
2308 bool IsSExt = isa<SExtInst>(Ext);
2309 // If the operand of the extension is not an instruction, we cannot
2311 // If it, check we can get through.
2312 if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt))
2315 // Do not promote if the operand has been added by codegenprepare.
2316 // Otherwise, it means we are undoing an optimization that is likely to be
2317 // redone, thus causing potential infinite loop.
2318 if (isa<TruncInst>(ExtOpnd) && InsertedTruncs.count(ExtOpnd))
2321 // SExt or Trunc instructions.
2322 // Return the related handler.
2323 if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) ||
2324 isa<ZExtInst>(ExtOpnd))
2325 return promoteOperandForTruncAndAnyExt;
2327 // Regular instruction.
2328 // Abort early if we will have to insert non-free instructions.
2329 if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType()))
2331 return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
2334 Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
2335 llvm::Instruction *SExt, TypePromotionTransaction &TPT,
2336 InstrToOrigTy &PromotedInsts, unsigned &CreatedInsts,
2337 SmallVectorImpl<Instruction *> *Exts,
2338 SmallVectorImpl<Instruction *> *Truncs) {
2339 // By construction, the operand of SExt is an instruction. Otherwise we cannot
2340 // get through it and this method should not be called.
2341 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
2342 Value *ExtVal = SExt;
2343 if (isa<ZExtInst>(SExtOpnd)) {
2344 // Replace s|zext(zext(opnd))
2347 TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
2348 TPT.replaceAllUsesWith(SExt, ZExt);
2349 TPT.eraseInstruction(SExt);
2352 // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
2354 TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
2358 // Remove dead code.
2359 if (SExtOpnd->use_empty())
2360 TPT.eraseInstruction(SExtOpnd);
2362 // Check if the extension is still needed.
2363 Instruction *ExtInst = dyn_cast<Instruction>(ExtVal);
2364 if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) {
2365 if (ExtInst && Exts)
2366 Exts->push_back(ExtInst);
2370 // At this point we have: ext ty opnd to ty.
2371 // Reassign the uses of ExtInst to the opnd and remove ExtInst.
2372 Value *NextVal = ExtInst->getOperand(0);
2373 TPT.eraseInstruction(ExtInst, NextVal);
2377 Value *TypePromotionHelper::promoteOperandForOther(
2378 Instruction *Ext, TypePromotionTransaction &TPT,
2379 InstrToOrigTy &PromotedInsts, unsigned &CreatedInsts,
2380 SmallVectorImpl<Instruction *> *Exts,
2381 SmallVectorImpl<Instruction *> *Truncs, bool IsSExt) {
2382 // By construction, the operand of Ext is an instruction. Otherwise we cannot
2383 // get through it and this method should not be called.
2384 Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0));
2386 if (!ExtOpnd->hasOneUse()) {
2387 // ExtOpnd will be promoted.
2388 // All its uses, but Ext, will need to use a truncated value of the
2389 // promoted version.
2390 // Create the truncate now.
2391 Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType());
2392 if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) {
2393 ITrunc->removeFromParent();
2394 // Insert it just after the definition.
2395 ITrunc->insertAfter(ExtOpnd);
2397 Truncs->push_back(ITrunc);
2400 TPT.replaceAllUsesWith(ExtOpnd, Trunc);
2401 // Restore the operand of Ext (which has been replace by the previous call
2402 // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
2403 TPT.setOperand(Ext, 0, ExtOpnd);
2406 // Get through the Instruction:
2407 // 1. Update its type.
2408 // 2. Replace the uses of Ext by Inst.
2409 // 3. Extend each operand that needs to be extended.
2411 // Remember the original type of the instruction before promotion.
2412 // This is useful to know that the high bits are sign extended bits.
2413 PromotedInsts.insert(std::pair<Instruction *, TypeIsSExt>(
2414 ExtOpnd, TypeIsSExt(ExtOpnd->getType(), IsSExt)));
2416 TPT.mutateType(ExtOpnd, Ext->getType());
2418 TPT.replaceAllUsesWith(Ext, ExtOpnd);
2420 Instruction *ExtForOpnd = Ext;
2422 DEBUG(dbgs() << "Propagate Ext to operands\n");
2423 for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
2425 DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n');
2426 if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() ||
2427 !shouldExtOperand(ExtOpnd, OpIdx)) {
2428 DEBUG(dbgs() << "No need to propagate\n");
2431 // Check if we can statically extend the operand.
2432 Value *Opnd = ExtOpnd->getOperand(OpIdx);
2433 if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
2434 DEBUG(dbgs() << "Statically extend\n");
2435 unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
2436 APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth)
2437 : Cst->getValue().zext(BitWidth);
2438 TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal));
2441 // UndefValue are typed, so we have to statically sign extend them.
2442 if (isa<UndefValue>(Opnd)) {
2443 DEBUG(dbgs() << "Statically extend\n");
2444 TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType()));
2448 // Otherwise we have to explicity sign extend the operand.
2449 // Check if Ext was reused to extend an operand.
2451 // If yes, create a new one.
2452 DEBUG(dbgs() << "More operands to ext\n");
2453 Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType())
2454 : TPT.createZExt(Ext, Opnd, Ext->getType());
2455 if (!isa<Instruction>(ValForExtOpnd)) {
2456 TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd);
2459 ExtForOpnd = cast<Instruction>(ValForExtOpnd);
2463 Exts->push_back(ExtForOpnd);
2464 TPT.setOperand(ExtForOpnd, 0, Opnd);
2466 // Move the sign extension before the insertion point.
2467 TPT.moveBefore(ExtForOpnd, ExtOpnd);
2468 TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd);
2469 // If more sext are required, new instructions will have to be created.
2470 ExtForOpnd = nullptr;
2472 if (ExtForOpnd == Ext) {
2473 DEBUG(dbgs() << "Extension is useless now\n");
2474 TPT.eraseInstruction(Ext);
2479 /// IsPromotionProfitable - Check whether or not promoting an instruction
2480 /// to a wider type was profitable.
2481 /// \p MatchedSize gives the number of instructions that have been matched
2482 /// in the addressing mode after the promotion was applied.
2483 /// \p SizeWithPromotion gives the number of created instructions for
2484 /// the promotion plus the number of instructions that have been
2485 /// matched in the addressing mode before the promotion.
2486 /// \p PromotedOperand is the value that has been promoted.
2487 /// \return True if the promotion is profitable, false otherwise.
2489 AddressingModeMatcher::IsPromotionProfitable(unsigned MatchedSize,
2490 unsigned SizeWithPromotion,
2491 Value *PromotedOperand) const {
2492 // We folded less instructions than what we created to promote the operand.
2493 // This is not profitable.
2494 if (MatchedSize < SizeWithPromotion)
2496 if (MatchedSize > SizeWithPromotion)
2498 // The promotion is neutral but it may help folding the sign extension in
2499 // loads for instance.
2500 // Check that we did not create an illegal instruction.
2501 return isPromotedInstructionLegal(TLI, PromotedOperand);
2504 /// MatchOperationAddr - Given an instruction or constant expr, see if we can
2505 /// fold the operation into the addressing mode. If so, update the addressing
2506 /// mode and return true, otherwise return false without modifying AddrMode.
2507 /// If \p MovedAway is not NULL, it contains the information of whether or
2508 /// not AddrInst has to be folded into the addressing mode on success.
2509 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
2510 /// because it has been moved away.
2511 /// Thus AddrInst must not be added in the matched instructions.
2512 /// This state can happen when AddrInst is a sext, since it may be moved away.
2513 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
2514 /// not be referenced anymore.
2515 bool AddressingModeMatcher::MatchOperationAddr(User *AddrInst, unsigned Opcode,
2518 // Avoid exponential behavior on extremely deep expression trees.
2519 if (Depth >= 5) return false;
2521 // By default, all matched instructions stay in place.
2526 case Instruction::PtrToInt:
2527 // PtrToInt is always a noop, as we know that the int type is pointer sized.
2528 return MatchAddr(AddrInst->getOperand(0), Depth);
2529 case Instruction::IntToPtr:
2530 // This inttoptr is a no-op if the integer type is pointer sized.
2531 if (TLI.getValueType(AddrInst->getOperand(0)->getType()) ==
2532 TLI.getPointerTy(AddrInst->getType()->getPointerAddressSpace()))
2533 return MatchAddr(AddrInst->getOperand(0), Depth);
2535 case Instruction::BitCast:
2536 case Instruction::AddrSpaceCast:
2537 // BitCast is always a noop, and we can handle it as long as it is
2538 // int->int or pointer->pointer (we don't want int<->fp or something).
2539 if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
2540 AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
2541 // Don't touch identity bitcasts. These were probably put here by LSR,
2542 // and we don't want to mess around with them. Assume it knows what it
2544 AddrInst->getOperand(0)->getType() != AddrInst->getType())
2545 return MatchAddr(AddrInst->getOperand(0), Depth);
2547 case Instruction::Add: {
2548 // Check to see if we can merge in the RHS then the LHS. If so, we win.
2549 ExtAddrMode BackupAddrMode = AddrMode;
2550 unsigned OldSize = AddrModeInsts.size();
2551 // Start a transaction at this point.
2552 // The LHS may match but not the RHS.
2553 // Therefore, we need a higher level restoration point to undo partially
2554 // matched operation.
2555 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2556 TPT.getRestorationPoint();
2558 if (MatchAddr(AddrInst->getOperand(1), Depth+1) &&
2559 MatchAddr(AddrInst->getOperand(0), Depth+1))
2562 // Restore the old addr mode info.
2563 AddrMode = BackupAddrMode;
2564 AddrModeInsts.resize(OldSize);
2565 TPT.rollback(LastKnownGood);
2567 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
2568 if (MatchAddr(AddrInst->getOperand(0), Depth+1) &&
2569 MatchAddr(AddrInst->getOperand(1), Depth+1))
2572 // Otherwise we definitely can't merge the ADD in.
2573 AddrMode = BackupAddrMode;
2574 AddrModeInsts.resize(OldSize);
2575 TPT.rollback(LastKnownGood);
2578 //case Instruction::Or:
2579 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
2581 case Instruction::Mul:
2582 case Instruction::Shl: {
2583 // Can only handle X*C and X << C.
2584 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
2587 int64_t Scale = RHS->getSExtValue();
2588 if (Opcode == Instruction::Shl)
2589 Scale = 1LL << Scale;
2591 return MatchScaledValue(AddrInst->getOperand(0), Scale, Depth);
2593 case Instruction::GetElementPtr: {
2594 // Scan the GEP. We check it if it contains constant offsets and at most
2595 // one variable offset.
2596 int VariableOperand = -1;
2597 unsigned VariableScale = 0;
2599 int64_t ConstantOffset = 0;
2600 const DataLayout *TD = TLI.getDataLayout();
2601 gep_type_iterator GTI = gep_type_begin(AddrInst);
2602 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
2603 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
2604 const StructLayout *SL = TD->getStructLayout(STy);
2606 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
2607 ConstantOffset += SL->getElementOffset(Idx);
2609 uint64_t TypeSize = TD->getTypeAllocSize(GTI.getIndexedType());
2610 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
2611 ConstantOffset += CI->getSExtValue()*TypeSize;
2612 } else if (TypeSize) { // Scales of zero don't do anything.
2613 // We only allow one variable index at the moment.
2614 if (VariableOperand != -1)
2617 // Remember the variable index.
2618 VariableOperand = i;
2619 VariableScale = TypeSize;
2624 // A common case is for the GEP to only do a constant offset. In this case,
2625 // just add it to the disp field and check validity.
2626 if (VariableOperand == -1) {
2627 AddrMode.BaseOffs += ConstantOffset;
2628 if (ConstantOffset == 0 || TLI.isLegalAddressingMode(AddrMode, AccessTy)){
2629 // Check to see if we can fold the base pointer in too.
2630 if (MatchAddr(AddrInst->getOperand(0), Depth+1))
2633 AddrMode.BaseOffs -= ConstantOffset;
2637 // Save the valid addressing mode in case we can't match.
2638 ExtAddrMode BackupAddrMode = AddrMode;
2639 unsigned OldSize = AddrModeInsts.size();
2641 // See if the scale and offset amount is valid for this target.
2642 AddrMode.BaseOffs += ConstantOffset;
2644 // Match the base operand of the GEP.
2645 if (!MatchAddr(AddrInst->getOperand(0), Depth+1)) {
2646 // If it couldn't be matched, just stuff the value in a register.
2647 if (AddrMode.HasBaseReg) {
2648 AddrMode = BackupAddrMode;
2649 AddrModeInsts.resize(OldSize);
2652 AddrMode.HasBaseReg = true;
2653 AddrMode.BaseReg = AddrInst->getOperand(0);
2656 // Match the remaining variable portion of the GEP.
2657 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
2659 // If it couldn't be matched, try stuffing the base into a register
2660 // instead of matching it, and retrying the match of the scale.
2661 AddrMode = BackupAddrMode;
2662 AddrModeInsts.resize(OldSize);
2663 if (AddrMode.HasBaseReg)
2665 AddrMode.HasBaseReg = true;
2666 AddrMode.BaseReg = AddrInst->getOperand(0);
2667 AddrMode.BaseOffs += ConstantOffset;
2668 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand),
2669 VariableScale, Depth)) {
2670 // If even that didn't work, bail.
2671 AddrMode = BackupAddrMode;
2672 AddrModeInsts.resize(OldSize);
2679 case Instruction::SExt:
2680 case Instruction::ZExt: {
2681 Instruction *Ext = dyn_cast<Instruction>(AddrInst);
2685 // Try to move this ext out of the way of the addressing mode.
2686 // Ask for a method for doing so.
2687 TypePromotionHelper::Action TPH =
2688 TypePromotionHelper::getAction(Ext, InsertedTruncs, TLI, PromotedInsts);
2692 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2693 TPT.getRestorationPoint();
2694 unsigned CreatedInsts = 0;
2695 Value *PromotedOperand =
2696 TPH(Ext, TPT, PromotedInsts, CreatedInsts, nullptr, nullptr);
2697 // SExt has been moved away.
2698 // Thus either it will be rematched later in the recursive calls or it is
2699 // gone. Anyway, we must not fold it into the addressing mode at this point.
2703 // addr = gep base, idx
2705 // promotedOpnd = ext opnd <- no match here
2706 // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
2707 // addr = gep base, op <- match
2711 assert(PromotedOperand &&
2712 "TypePromotionHelper should have filtered out those cases");
2714 ExtAddrMode BackupAddrMode = AddrMode;
2715 unsigned OldSize = AddrModeInsts.size();
2717 if (!MatchAddr(PromotedOperand, Depth) ||
2718 !IsPromotionProfitable(AddrModeInsts.size(), OldSize + CreatedInsts,
2720 AddrMode = BackupAddrMode;
2721 AddrModeInsts.resize(OldSize);
2722 DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
2723 TPT.rollback(LastKnownGood);
2732 /// MatchAddr - If we can, try to add the value of 'Addr' into the current
2733 /// addressing mode. If Addr can't be added to AddrMode this returns false and
2734 /// leaves AddrMode unmodified. This assumes that Addr is either a pointer type
2735 /// or intptr_t for the target.
2737 bool AddressingModeMatcher::MatchAddr(Value *Addr, unsigned Depth) {
2738 // Start a transaction at this point that we will rollback if the matching
2740 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2741 TPT.getRestorationPoint();
2742 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
2743 // Fold in immediates if legal for the target.
2744 AddrMode.BaseOffs += CI->getSExtValue();
2745 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2747 AddrMode.BaseOffs -= CI->getSExtValue();
2748 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
2749 // If this is a global variable, try to fold it into the addressing mode.
2750 if (!AddrMode.BaseGV) {
2751 AddrMode.BaseGV = GV;
2752 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2754 AddrMode.BaseGV = nullptr;
2756 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
2757 ExtAddrMode BackupAddrMode = AddrMode;
2758 unsigned OldSize = AddrModeInsts.size();
2760 // Check to see if it is possible to fold this operation.
2761 bool MovedAway = false;
2762 if (MatchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
2763 // This instruction may have been move away. If so, there is nothing
2767 // Okay, it's possible to fold this. Check to see if it is actually
2768 // *profitable* to do so. We use a simple cost model to avoid increasing
2769 // register pressure too much.
2770 if (I->hasOneUse() ||
2771 IsProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
2772 AddrModeInsts.push_back(I);
2776 // It isn't profitable to do this, roll back.
2777 //cerr << "NOT FOLDING: " << *I;
2778 AddrMode = BackupAddrMode;
2779 AddrModeInsts.resize(OldSize);
2780 TPT.rollback(LastKnownGood);
2782 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
2783 if (MatchOperationAddr(CE, CE->getOpcode(), Depth))
2785 TPT.rollback(LastKnownGood);
2786 } else if (isa<ConstantPointerNull>(Addr)) {
2787 // Null pointer gets folded without affecting the addressing mode.
2791 // Worse case, the target should support [reg] addressing modes. :)
2792 if (!AddrMode.HasBaseReg) {
2793 AddrMode.HasBaseReg = true;
2794 AddrMode.BaseReg = Addr;
2795 // Still check for legality in case the target supports [imm] but not [i+r].
2796 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2798 AddrMode.HasBaseReg = false;
2799 AddrMode.BaseReg = nullptr;
2802 // If the base register is already taken, see if we can do [r+r].
2803 if (AddrMode.Scale == 0) {
2805 AddrMode.ScaledReg = Addr;
2806 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2809 AddrMode.ScaledReg = nullptr;
2812 TPT.rollback(LastKnownGood);
2816 /// IsOperandAMemoryOperand - Check to see if all uses of OpVal by the specified
2817 /// inline asm call are due to memory operands. If so, return true, otherwise
2819 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
2820 const TargetMachine &TM) {
2821 const Function *F = CI->getParent()->getParent();
2822 const TargetLowering *TLI = TM.getSubtargetImpl(*F)->getTargetLowering();
2823 const TargetRegisterInfo *TRI = TM.getSubtargetImpl(*F)->getRegisterInfo();
2824 TargetLowering::AsmOperandInfoVector TargetConstraints =
2825 TLI->ParseConstraints(TRI, ImmutableCallSite(CI));
2826 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
2827 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
2829 // Compute the constraint code and ConstraintType to use.
2830 TLI->ComputeConstraintToUse(OpInfo, SDValue());
2832 // If this asm operand is our Value*, and if it isn't an indirect memory
2833 // operand, we can't fold it!
2834 if (OpInfo.CallOperandVal == OpVal &&
2835 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
2836 !OpInfo.isIndirect))
2843 /// FindAllMemoryUses - Recursively walk all the uses of I until we find a
2844 /// memory use. If we find an obviously non-foldable instruction, return true.
2845 /// Add the ultimately found memory instructions to MemoryUses.
2846 static bool FindAllMemoryUses(
2848 SmallVectorImpl<std::pair<Instruction *, unsigned>> &MemoryUses,
2849 SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetMachine &TM) {
2850 // If we already considered this instruction, we're done.
2851 if (!ConsideredInsts.insert(I).second)
2854 // If this is an obviously unfoldable instruction, bail out.
2855 if (!MightBeFoldableInst(I))
2858 // Loop over all the uses, recursively processing them.
2859 for (Use &U : I->uses()) {
2860 Instruction *UserI = cast<Instruction>(U.getUser());
2862 if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
2863 MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
2867 if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
2868 unsigned opNo = U.getOperandNo();
2869 if (opNo == 0) return true; // Storing addr, not into addr.
2870 MemoryUses.push_back(std::make_pair(SI, opNo));
2874 if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
2875 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
2876 if (!IA) return true;
2878 // If this is a memory operand, we're cool, otherwise bail out.
2879 if (!IsOperandAMemoryOperand(CI, IA, I, TM))
2884 if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TM))
2891 /// ValueAlreadyLiveAtInst - Retrn true if Val is already known to be live at
2892 /// the use site that we're folding it into. If so, there is no cost to
2893 /// include it in the addressing mode. KnownLive1 and KnownLive2 are two values
2894 /// that we know are live at the instruction already.
2895 bool AddressingModeMatcher::ValueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
2896 Value *KnownLive2) {
2897 // If Val is either of the known-live values, we know it is live!
2898 if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
2901 // All values other than instructions and arguments (e.g. constants) are live.
2902 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
2904 // If Val is a constant sized alloca in the entry block, it is live, this is
2905 // true because it is just a reference to the stack/frame pointer, which is
2906 // live for the whole function.
2907 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
2908 if (AI->isStaticAlloca())
2911 // Check to see if this value is already used in the memory instruction's
2912 // block. If so, it's already live into the block at the very least, so we
2913 // can reasonably fold it.
2914 return Val->isUsedInBasicBlock(MemoryInst->getParent());
2917 /// IsProfitableToFoldIntoAddressingMode - It is possible for the addressing
2918 /// mode of the machine to fold the specified instruction into a load or store
2919 /// that ultimately uses it. However, the specified instruction has multiple
2920 /// uses. Given this, it may actually increase register pressure to fold it
2921 /// into the load. For example, consider this code:
2925 /// use(Y) -> nonload/store
2929 /// In this case, Y has multiple uses, and can be folded into the load of Z
2930 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
2931 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
2932 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
2933 /// number of computations either.
2935 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
2936 /// X was live across 'load Z' for other reasons, we actually *would* want to
2937 /// fold the addressing mode in the Z case. This would make Y die earlier.
2938 bool AddressingModeMatcher::
2939 IsProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
2940 ExtAddrMode &AMAfter) {
2941 if (IgnoreProfitability) return true;
2943 // AMBefore is the addressing mode before this instruction was folded into it,
2944 // and AMAfter is the addressing mode after the instruction was folded. Get
2945 // the set of registers referenced by AMAfter and subtract out those
2946 // referenced by AMBefore: this is the set of values which folding in this
2947 // address extends the lifetime of.
2949 // Note that there are only two potential values being referenced here,
2950 // BaseReg and ScaleReg (global addresses are always available, as are any
2951 // folded immediates).
2952 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
2954 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
2955 // lifetime wasn't extended by adding this instruction.
2956 if (ValueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
2958 if (ValueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
2959 ScaledReg = nullptr;
2961 // If folding this instruction (and it's subexprs) didn't extend any live
2962 // ranges, we're ok with it.
2963 if (!BaseReg && !ScaledReg)
2966 // If all uses of this instruction are ultimately load/store/inlineasm's,
2967 // check to see if their addressing modes will include this instruction. If
2968 // so, we can fold it into all uses, so it doesn't matter if it has multiple
2970 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
2971 SmallPtrSet<Instruction*, 16> ConsideredInsts;
2972 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TM))
2973 return false; // Has a non-memory, non-foldable use!
2975 // Now that we know that all uses of this instruction are part of a chain of
2976 // computation involving only operations that could theoretically be folded
2977 // into a memory use, loop over each of these uses and see if they could
2978 // *actually* fold the instruction.
2979 SmallVector<Instruction*, 32> MatchedAddrModeInsts;
2980 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
2981 Instruction *User = MemoryUses[i].first;
2982 unsigned OpNo = MemoryUses[i].second;
2984 // Get the access type of this use. If the use isn't a pointer, we don't
2985 // know what it accesses.
2986 Value *Address = User->getOperand(OpNo);
2987 if (!Address->getType()->isPointerTy())
2989 Type *AddressAccessTy = Address->getType()->getPointerElementType();
2991 // Do a match against the root of this address, ignoring profitability. This
2992 // will tell us if the addressing mode for the memory operation will
2993 // *actually* cover the shared instruction.
2995 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2996 TPT.getRestorationPoint();
2997 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TM, AddressAccessTy,
2998 MemoryInst, Result, InsertedTruncs,
2999 PromotedInsts, TPT);
3000 Matcher.IgnoreProfitability = true;
3001 bool Success = Matcher.MatchAddr(Address, 0);
3002 (void)Success; assert(Success && "Couldn't select *anything*?");
3004 // The match was to check the profitability, the changes made are not
3005 // part of the original matcher. Therefore, they should be dropped
3006 // otherwise the original matcher will not present the right state.
3007 TPT.rollback(LastKnownGood);
3009 // If the match didn't cover I, then it won't be shared by it.
3010 if (std::find(MatchedAddrModeInsts.begin(), MatchedAddrModeInsts.end(),
3011 I) == MatchedAddrModeInsts.end())
3014 MatchedAddrModeInsts.clear();
3020 } // end anonymous namespace
3022 /// IsNonLocalValue - Return true if the specified values are defined in a
3023 /// different basic block than BB.
3024 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
3025 if (Instruction *I = dyn_cast<Instruction>(V))
3026 return I->getParent() != BB;
3030 /// OptimizeMemoryInst - Load and Store Instructions often have
3031 /// addressing modes that can do significant amounts of computation. As such,
3032 /// instruction selection will try to get the load or store to do as much
3033 /// computation as possible for the program. The problem is that isel can only
3034 /// see within a single block. As such, we sink as much legal addressing mode
3035 /// stuff into the block as possible.
3037 /// This method is used to optimize both load/store and inline asms with memory
3039 bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
3043 // Try to collapse single-value PHI nodes. This is necessary to undo
3044 // unprofitable PRE transformations.
3045 SmallVector<Value*, 8> worklist;
3046 SmallPtrSet<Value*, 16> Visited;
3047 worklist.push_back(Addr);
3049 // Use a worklist to iteratively look through PHI nodes, and ensure that
3050 // the addressing mode obtained from the non-PHI roots of the graph
3052 Value *Consensus = nullptr;
3053 unsigned NumUsesConsensus = 0;
3054 bool IsNumUsesConsensusValid = false;
3055 SmallVector<Instruction*, 16> AddrModeInsts;
3056 ExtAddrMode AddrMode;
3057 TypePromotionTransaction TPT;
3058 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3059 TPT.getRestorationPoint();
3060 while (!worklist.empty()) {
3061 Value *V = worklist.back();
3062 worklist.pop_back();
3064 // Break use-def graph loops.
3065 if (!Visited.insert(V).second) {
3066 Consensus = nullptr;
3070 // For a PHI node, push all of its incoming values.
3071 if (PHINode *P = dyn_cast<PHINode>(V)) {
3072 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i)
3073 worklist.push_back(P->getIncomingValue(i));
3077 // For non-PHIs, determine the addressing mode being computed.
3078 SmallVector<Instruction*, 16> NewAddrModeInsts;
3079 ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
3080 V, AccessTy, MemoryInst, NewAddrModeInsts, *TM, InsertedTruncsSet,
3081 PromotedInsts, TPT);
3083 // This check is broken into two cases with very similar code to avoid using
3084 // getNumUses() as much as possible. Some values have a lot of uses, so
3085 // calling getNumUses() unconditionally caused a significant compile-time
3089 AddrMode = NewAddrMode;
3090 AddrModeInsts = NewAddrModeInsts;
3092 } else if (NewAddrMode == AddrMode) {
3093 if (!IsNumUsesConsensusValid) {
3094 NumUsesConsensus = Consensus->getNumUses();
3095 IsNumUsesConsensusValid = true;
3098 // Ensure that the obtained addressing mode is equivalent to that obtained
3099 // for all other roots of the PHI traversal. Also, when choosing one
3100 // such root as representative, select the one with the most uses in order
3101 // to keep the cost modeling heuristics in AddressingModeMatcher
3103 unsigned NumUses = V->getNumUses();
3104 if (NumUses > NumUsesConsensus) {
3106 NumUsesConsensus = NumUses;
3107 AddrModeInsts = NewAddrModeInsts;
3112 Consensus = nullptr;
3116 // If the addressing mode couldn't be determined, or if multiple different
3117 // ones were determined, bail out now.
3119 TPT.rollback(LastKnownGood);
3124 // Check to see if any of the instructions supersumed by this addr mode are
3125 // non-local to I's BB.
3126 bool AnyNonLocal = false;
3127 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
3128 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
3134 // If all the instructions matched are already in this BB, don't do anything.
3136 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
3140 // Insert this computation right after this user. Since our caller is
3141 // scanning from the top of the BB to the bottom, reuse of the expr are
3142 // guaranteed to happen later.
3143 IRBuilder<> Builder(MemoryInst);
3145 // Now that we determined the addressing expression we want to use and know
3146 // that we have to sink it into this block. Check to see if we have already
3147 // done this for some other load/store instr in this block. If so, reuse the
3149 Value *&SunkAddr = SunkAddrs[Addr];
3151 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
3152 << *MemoryInst << "\n");
3153 if (SunkAddr->getType() != Addr->getType())
3154 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
3155 } else if (AddrSinkUsingGEPs ||
3156 (!AddrSinkUsingGEPs.getNumOccurrences() && TM &&
3157 TM->getSubtargetImpl(*MemoryInst->getParent()->getParent())
3159 // By default, we use the GEP-based method when AA is used later. This
3160 // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
3161 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
3162 << *MemoryInst << "\n");
3163 Type *IntPtrTy = TLI->getDataLayout()->getIntPtrType(Addr->getType());
3164 Value *ResultPtr = nullptr, *ResultIndex = nullptr;
3166 // First, find the pointer.
3167 if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
3168 ResultPtr = AddrMode.BaseReg;
3169 AddrMode.BaseReg = nullptr;
3172 if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
3173 // We can't add more than one pointer together, nor can we scale a
3174 // pointer (both of which seem meaningless).
3175 if (ResultPtr || AddrMode.Scale != 1)
3178 ResultPtr = AddrMode.ScaledReg;
3182 if (AddrMode.BaseGV) {
3186 ResultPtr = AddrMode.BaseGV;
3189 // If the real base value actually came from an inttoptr, then the matcher
3190 // will look through it and provide only the integer value. In that case,
3192 if (!ResultPtr && AddrMode.BaseReg) {
3194 Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(), "sunkaddr");
3195 AddrMode.BaseReg = nullptr;
3196 } else if (!ResultPtr && AddrMode.Scale == 1) {
3198 Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(), "sunkaddr");
3203 !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
3204 SunkAddr = Constant::getNullValue(Addr->getType());
3205 } else if (!ResultPtr) {
3209 Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
3211 // Start with the base register. Do this first so that subsequent address
3212 // matching finds it last, which will prevent it from trying to match it
3213 // as the scaled value in case it happens to be a mul. That would be
3214 // problematic if we've sunk a different mul for the scale, because then
3215 // we'd end up sinking both muls.
3216 if (AddrMode.BaseReg) {
3217 Value *V = AddrMode.BaseReg;
3218 if (V->getType() != IntPtrTy)
3219 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
3224 // Add the scale value.
3225 if (AddrMode.Scale) {
3226 Value *V = AddrMode.ScaledReg;
3227 if (V->getType() == IntPtrTy) {
3229 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
3230 cast<IntegerType>(V->getType())->getBitWidth()) {
3231 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
3233 // It is only safe to sign extend the BaseReg if we know that the math
3234 // required to create it did not overflow before we extend it. Since
3235 // the original IR value was tossed in favor of a constant back when
3236 // the AddrMode was created we need to bail out gracefully if widths
3237 // do not match instead of extending it.
3238 Instruction *I = dyn_cast_or_null<Instruction>(ResultIndex);
3239 if (I && (ResultIndex != AddrMode.BaseReg))
3240 I->eraseFromParent();
3244 if (AddrMode.Scale != 1)
3245 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
3248 ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
3253 // Add in the Base Offset if present.
3254 if (AddrMode.BaseOffs) {
3255 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
3257 // We need to add this separately from the scale above to help with
3258 // SDAG consecutive load/store merging.
3259 if (ResultPtr->getType() != I8PtrTy)
3260 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
3261 ResultPtr = Builder.CreateGEP(ResultPtr, ResultIndex, "sunkaddr");
3268 SunkAddr = ResultPtr;
3270 if (ResultPtr->getType() != I8PtrTy)
3271 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
3272 SunkAddr = Builder.CreateGEP(ResultPtr, ResultIndex, "sunkaddr");
3275 if (SunkAddr->getType() != Addr->getType())
3276 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
3279 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
3280 << *MemoryInst << "\n");
3281 Type *IntPtrTy = TLI->getDataLayout()->getIntPtrType(Addr->getType());
3282 Value *Result = nullptr;
3284 // Start with the base register. Do this first so that subsequent address
3285 // matching finds it last, which will prevent it from trying to match it
3286 // as the scaled value in case it happens to be a mul. That would be
3287 // problematic if we've sunk a different mul for the scale, because then
3288 // we'd end up sinking both muls.
3289 if (AddrMode.BaseReg) {
3290 Value *V = AddrMode.BaseReg;
3291 if (V->getType()->isPointerTy())
3292 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
3293 if (V->getType() != IntPtrTy)
3294 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
3298 // Add the scale value.
3299 if (AddrMode.Scale) {
3300 Value *V = AddrMode.ScaledReg;
3301 if (V->getType() == IntPtrTy) {
3303 } else if (V->getType()->isPointerTy()) {
3304 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
3305 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
3306 cast<IntegerType>(V->getType())->getBitWidth()) {
3307 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
3309 // It is only safe to sign extend the BaseReg if we know that the math
3310 // required to create it did not overflow before we extend it. Since
3311 // the original IR value was tossed in favor of a constant back when
3312 // the AddrMode was created we need to bail out gracefully if widths
3313 // do not match instead of extending it.
3314 Instruction *I = dyn_cast_or_null<Instruction>(Result);
3315 if (I && (Result != AddrMode.BaseReg))
3316 I->eraseFromParent();
3319 if (AddrMode.Scale != 1)
3320 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
3323 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3328 // Add in the BaseGV if present.
3329 if (AddrMode.BaseGV) {
3330 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
3332 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3337 // Add in the Base Offset if present.
3338 if (AddrMode.BaseOffs) {
3339 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
3341 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3347 SunkAddr = Constant::getNullValue(Addr->getType());
3349 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
3352 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
3354 // If we have no uses, recursively delete the value and all dead instructions
3356 if (Repl->use_empty()) {
3357 // This can cause recursive deletion, which can invalidate our iterator.
3358 // Use a WeakVH to hold onto it in case this happens.
3359 WeakVH IterHandle(CurInstIterator);
3360 BasicBlock *BB = CurInstIterator->getParent();
3362 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
3364 if (IterHandle != CurInstIterator) {
3365 // If the iterator instruction was recursively deleted, start over at the
3366 // start of the block.
3367 CurInstIterator = BB->begin();
3375 /// OptimizeInlineAsmInst - If there are any memory operands, use
3376 /// OptimizeMemoryInst to sink their address computing into the block when
3377 /// possible / profitable.
3378 bool CodeGenPrepare::OptimizeInlineAsmInst(CallInst *CS) {
3379 bool MadeChange = false;
3381 const TargetRegisterInfo *TRI =
3382 TM->getSubtargetImpl(*CS->getParent()->getParent())->getRegisterInfo();
3383 TargetLowering::AsmOperandInfoVector
3384 TargetConstraints = TLI->ParseConstraints(TRI, CS);
3386 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
3387 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
3389 // Compute the constraint code and ConstraintType to use.
3390 TLI->ComputeConstraintToUse(OpInfo, SDValue());
3392 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
3393 OpInfo.isIndirect) {
3394 Value *OpVal = CS->getArgOperand(ArgNo++);
3395 MadeChange |= OptimizeMemoryInst(CS, OpVal, OpVal->getType());
3396 } else if (OpInfo.Type == InlineAsm::isInput)
3403 /// \brief Check if all the uses of \p Inst are equivalent (or free) zero or
3404 /// sign extensions.
3405 static bool hasSameExtUse(Instruction *Inst, const TargetLowering &TLI) {
3406 assert(!Inst->use_empty() && "Input must have at least one use");
3407 const Instruction *FirstUser = cast<Instruction>(*Inst->user_begin());
3408 bool IsSExt = isa<SExtInst>(FirstUser);
3409 Type *ExtTy = FirstUser->getType();
3410 for (const User *U : Inst->users()) {
3411 const Instruction *UI = cast<Instruction>(U);
3412 if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI)))
3414 Type *CurTy = UI->getType();
3415 // Same input and output types: Same instruction after CSE.
3419 // If IsSExt is true, we are in this situation:
3421 // b = sext ty1 a to ty2
3422 // c = sext ty1 a to ty3
3423 // Assuming ty2 is shorter than ty3, this could be turned into:
3425 // b = sext ty1 a to ty2
3426 // c = sext ty2 b to ty3
3427 // However, the last sext is not free.
3431 // This is a ZExt, maybe this is free to extend from one type to another.
3432 // In that case, we would not account for a different use.
3435 if (ExtTy->getScalarType()->getIntegerBitWidth() >
3436 CurTy->getScalarType()->getIntegerBitWidth()) {
3444 if (!TLI.isZExtFree(NarrowTy, LargeTy))
3447 // All uses are the same or can be derived from one another for free.
3451 /// \brief Try to form ExtLd by promoting \p Exts until they reach a
3452 /// load instruction.
3453 /// If an ext(load) can be formed, it is returned via \p LI for the load
3454 /// and \p Inst for the extension.
3455 /// Otherwise LI == nullptr and Inst == nullptr.
3456 /// When some promotion happened, \p TPT contains the proper state to
3459 /// \return true when promoting was necessary to expose the ext(load)
3460 /// opportunity, false otherwise.
3464 /// %ld = load i32* %addr
3465 /// %add = add nuw i32 %ld, 4
3466 /// %zext = zext i32 %add to i64
3470 /// %ld = load i32* %addr
3471 /// %zext = zext i32 %ld to i64
3472 /// %add = add nuw i64 %zext, 4
3474 /// Thanks to the promotion, we can match zext(load i32*) to i64.
3475 bool CodeGenPrepare::ExtLdPromotion(TypePromotionTransaction &TPT,
3476 LoadInst *&LI, Instruction *&Inst,
3477 const SmallVectorImpl<Instruction *> &Exts,
3478 unsigned CreatedInsts = 0) {
3479 // Iterate over all the extensions to see if one form an ext(load).
3480 for (auto I : Exts) {
3481 // Check if we directly have ext(load).
3482 if ((LI = dyn_cast<LoadInst>(I->getOperand(0)))) {
3484 // No promotion happened here.
3487 // Check whether or not we want to do any promotion.
3488 if (!TLI || !TLI->enableExtLdPromotion() || DisableExtLdPromotion)
3490 // Get the action to perform the promotion.
3491 TypePromotionHelper::Action TPH = TypePromotionHelper::getAction(
3492 I, InsertedTruncsSet, *TLI, PromotedInsts);
3493 // Check if we can promote.
3496 // Save the current state.
3497 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3498 TPT.getRestorationPoint();
3499 SmallVector<Instruction *, 4> NewExts;
3500 unsigned NewCreatedInsts = 0;
3502 Value *PromotedVal =
3503 TPH(I, TPT, PromotedInsts, NewCreatedInsts, &NewExts, nullptr);
3504 assert(PromotedVal &&
3505 "TypePromotionHelper should have filtered out those cases");
3507 // We would be able to merge only one extension in a load.
3508 // Therefore, if we have more than 1 new extension we heuristically
3509 // cut this search path, because it means we degrade the code quality.
3510 // With exactly 2, the transformation is neutral, because we will merge
3511 // one extension but leave one. However, we optimistically keep going,
3512 // because the new extension may be removed too.
3513 unsigned TotalCreatedInsts = CreatedInsts + NewCreatedInsts;
3514 if (!StressExtLdPromotion &&
3515 (TotalCreatedInsts > 1 ||
3516 !isPromotedInstructionLegal(*TLI, PromotedVal))) {
3517 // The promotion is not profitable, rollback to the previous state.
3518 TPT.rollback(LastKnownGood);
3521 // The promotion is profitable.
3522 // Check if it exposes an ext(load).
3523 (void)ExtLdPromotion(TPT, LI, Inst, NewExts, TotalCreatedInsts);
3524 if (LI && (StressExtLdPromotion || NewCreatedInsts == 0 ||
3525 // If we have created a new extension, i.e., now we have two
3526 // extensions. We must make sure one of them is merged with
3527 // the load, otherwise we may degrade the code quality.
3528 (LI->hasOneUse() || hasSameExtUse(LI, *TLI))))
3529 // Promotion happened.
3531 // If this does not help to expose an ext(load) then, rollback.
3532 TPT.rollback(LastKnownGood);
3534 // None of the extension can form an ext(load).
3540 /// MoveExtToFormExtLoad - Move a zext or sext fed by a load into the same
3541 /// basic block as the load, unless conditions are unfavorable. This allows
3542 /// SelectionDAG to fold the extend into the load.
3543 /// \p I[in/out] the extension may be modified during the process if some
3544 /// promotions apply.
3546 bool CodeGenPrepare::MoveExtToFormExtLoad(Instruction *&I) {
3547 // Try to promote a chain of computation if it allows to form
3548 // an extended load.
3549 TypePromotionTransaction TPT;
3550 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3551 TPT.getRestorationPoint();
3552 SmallVector<Instruction *, 1> Exts;
3554 // Look for a load being extended.
3555 LoadInst *LI = nullptr;
3556 Instruction *OldExt = I;
3557 bool HasPromoted = ExtLdPromotion(TPT, LI, I, Exts);
3559 assert(!HasPromoted && !LI && "If we did not match any load instruction "
3560 "the code must remain the same");
3565 // If they're already in the same block, there's nothing to do.
3566 // Make the cheap checks first if we did not promote.
3567 // If we promoted, we need to check if it is indeed profitable.
3568 if (!HasPromoted && LI->getParent() == I->getParent())
3571 EVT VT = TLI->getValueType(I->getType());
3572 EVT LoadVT = TLI->getValueType(LI->getType());
3574 // If the load has other users and the truncate is not free, this probably
3575 // isn't worthwhile.
3576 if (!LI->hasOneUse() && TLI &&
3577 (TLI->isTypeLegal(LoadVT) || !TLI->isTypeLegal(VT)) &&
3578 !TLI->isTruncateFree(I->getType(), LI->getType())) {
3580 TPT.rollback(LastKnownGood);
3584 // Check whether the target supports casts folded into loads.
3586 if (isa<ZExtInst>(I))
3587 LType = ISD::ZEXTLOAD;
3589 assert(isa<SExtInst>(I) && "Unexpected ext type!");
3590 LType = ISD::SEXTLOAD;
3592 if (TLI && !TLI->isLoadExtLegal(LType, VT, LoadVT)) {
3594 TPT.rollback(LastKnownGood);
3598 // Move the extend into the same block as the load, so that SelectionDAG
3601 I->removeFromParent();
3607 bool CodeGenPrepare::OptimizeExtUses(Instruction *I) {
3608 BasicBlock *DefBB = I->getParent();
3610 // If the result of a {s|z}ext and its source are both live out, rewrite all
3611 // other uses of the source with result of extension.
3612 Value *Src = I->getOperand(0);
3613 if (Src->hasOneUse())
3616 // Only do this xform if truncating is free.
3617 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
3620 // Only safe to perform the optimization if the source is also defined in
3622 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
3625 bool DefIsLiveOut = false;
3626 for (User *U : I->users()) {
3627 Instruction *UI = cast<Instruction>(U);
3629 // Figure out which BB this ext is used in.
3630 BasicBlock *UserBB = UI->getParent();
3631 if (UserBB == DefBB) continue;
3632 DefIsLiveOut = true;
3638 // Make sure none of the uses are PHI nodes.
3639 for (User *U : Src->users()) {
3640 Instruction *UI = cast<Instruction>(U);
3641 BasicBlock *UserBB = UI->getParent();
3642 if (UserBB == DefBB) continue;
3643 // Be conservative. We don't want this xform to end up introducing
3644 // reloads just before load / store instructions.
3645 if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
3649 // InsertedTruncs - Only insert one trunc in each block once.
3650 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
3652 bool MadeChange = false;
3653 for (Use &U : Src->uses()) {
3654 Instruction *User = cast<Instruction>(U.getUser());
3656 // Figure out which BB this ext is used in.
3657 BasicBlock *UserBB = User->getParent();
3658 if (UserBB == DefBB) continue;
3660 // Both src and def are live in this block. Rewrite the use.
3661 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
3663 if (!InsertedTrunc) {
3664 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
3665 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt);
3666 InsertedTruncsSet.insert(InsertedTrunc);
3669 // Replace a use of the {s|z}ext source with a use of the result.
3678 /// isFormingBranchFromSelectProfitable - Returns true if a SelectInst should be
3679 /// turned into an explicit branch.
3680 static bool isFormingBranchFromSelectProfitable(SelectInst *SI) {
3681 // FIXME: This should use the same heuristics as IfConversion to determine
3682 // whether a select is better represented as a branch. This requires that
3683 // branch probability metadata is preserved for the select, which is not the
3686 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
3688 // If the branch is predicted right, an out of order CPU can avoid blocking on
3689 // the compare. Emit cmovs on compares with a memory operand as branches to
3690 // avoid stalls on the load from memory. If the compare has more than one use
3691 // there's probably another cmov or setcc around so it's not worth emitting a
3696 Value *CmpOp0 = Cmp->getOperand(0);
3697 Value *CmpOp1 = Cmp->getOperand(1);
3699 // We check that the memory operand has one use to avoid uses of the loaded
3700 // value directly after the compare, making branches unprofitable.
3701 return Cmp->hasOneUse() &&
3702 ((isa<LoadInst>(CmpOp0) && CmpOp0->hasOneUse()) ||
3703 (isa<LoadInst>(CmpOp1) && CmpOp1->hasOneUse()));
3707 /// If we have a SelectInst that will likely profit from branch prediction,
3708 /// turn it into a branch.
3709 bool CodeGenPrepare::OptimizeSelectInst(SelectInst *SI) {
3710 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
3712 // Can we convert the 'select' to CF ?
3713 if (DisableSelectToBranch || OptSize || !TLI || VectorCond)
3716 TargetLowering::SelectSupportKind SelectKind;
3718 SelectKind = TargetLowering::VectorMaskSelect;
3719 else if (SI->getType()->isVectorTy())
3720 SelectKind = TargetLowering::ScalarCondVectorVal;
3722 SelectKind = TargetLowering::ScalarValSelect;
3724 // Do we have efficient codegen support for this kind of 'selects' ?
3725 if (TLI->isSelectSupported(SelectKind)) {
3726 // We have efficient codegen support for the select instruction.
3727 // Check if it is profitable to keep this 'select'.
3728 if (!TLI->isPredictableSelectExpensive() ||
3729 !isFormingBranchFromSelectProfitable(SI))
3735 // First, we split the block containing the select into 2 blocks.
3736 BasicBlock *StartBlock = SI->getParent();
3737 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(SI));
3738 BasicBlock *NextBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
3740 // Create a new block serving as the landing pad for the branch.
3741 BasicBlock *SmallBlock = BasicBlock::Create(SI->getContext(), "select.mid",
3742 NextBlock->getParent(), NextBlock);
3744 // Move the unconditional branch from the block with the select in it into our
3745 // landing pad block.
3746 StartBlock->getTerminator()->eraseFromParent();
3747 BranchInst::Create(NextBlock, SmallBlock);
3749 // Insert the real conditional branch based on the original condition.
3750 BranchInst::Create(NextBlock, SmallBlock, SI->getCondition(), SI);
3752 // The select itself is replaced with a PHI Node.
3753 PHINode *PN = PHINode::Create(SI->getType(), 2, "", NextBlock->begin());
3755 PN->addIncoming(SI->getTrueValue(), StartBlock);
3756 PN->addIncoming(SI->getFalseValue(), SmallBlock);
3757 SI->replaceAllUsesWith(PN);
3758 SI->eraseFromParent();
3760 // Instruct OptimizeBlock to skip to the next block.
3761 CurInstIterator = StartBlock->end();
3762 ++NumSelectsExpanded;
3766 static bool isBroadcastShuffle(ShuffleVectorInst *SVI) {
3767 SmallVector<int, 16> Mask(SVI->getShuffleMask());
3769 for (unsigned i = 0; i < Mask.size(); ++i) {
3770 if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
3772 SplatElem = Mask[i];
3778 /// Some targets have expensive vector shifts if the lanes aren't all the same
3779 /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
3780 /// it's often worth sinking a shufflevector splat down to its use so that
3781 /// codegen can spot all lanes are identical.
3782 bool CodeGenPrepare::OptimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
3783 BasicBlock *DefBB = SVI->getParent();
3785 // Only do this xform if variable vector shifts are particularly expensive.
3786 if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
3789 // We only expect better codegen by sinking a shuffle if we can recognise a
3791 if (!isBroadcastShuffle(SVI))
3794 // InsertedShuffles - Only insert a shuffle in each block once.
3795 DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
3797 bool MadeChange = false;
3798 for (User *U : SVI->users()) {
3799 Instruction *UI = cast<Instruction>(U);
3801 // Figure out which BB this ext is used in.
3802 BasicBlock *UserBB = UI->getParent();
3803 if (UserBB == DefBB) continue;
3805 // For now only apply this when the splat is used by a shift instruction.
3806 if (!UI->isShift()) continue;
3808 // Everything checks out, sink the shuffle if the user's block doesn't
3809 // already have a copy.
3810 Instruction *&InsertedShuffle = InsertedShuffles[UserBB];
3812 if (!InsertedShuffle) {
3813 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
3814 InsertedShuffle = new ShuffleVectorInst(SVI->getOperand(0),
3816 SVI->getOperand(2), "", InsertPt);
3819 UI->replaceUsesOfWith(SVI, InsertedShuffle);
3823 // If we removed all uses, nuke the shuffle.
3824 if (SVI->use_empty()) {
3825 SVI->eraseFromParent();
3833 /// \brief Helper class to promote a scalar operation to a vector one.
3834 /// This class is used to move downward extractelement transition.
3836 /// a = vector_op <2 x i32>
3837 /// b = extractelement <2 x i32> a, i32 0
3842 /// a = vector_op <2 x i32>
3843 /// c = vector_op a (equivalent to scalar_op on the related lane)
3844 /// * d = extractelement <2 x i32> c, i32 0
3846 /// Assuming both extractelement and store can be combine, we get rid of the
3848 class VectorPromoteHelper {
3849 /// Used to perform some checks on the legality of vector operations.
3850 const TargetLowering &TLI;
3852 /// Used to estimated the cost of the promoted chain.
3853 const TargetTransformInfo &TTI;
3855 /// The transition being moved downwards.
3856 Instruction *Transition;
3857 /// The sequence of instructions to be promoted.
3858 SmallVector<Instruction *, 4> InstsToBePromoted;
3859 /// Cost of combining a store and an extract.
3860 unsigned StoreExtractCombineCost;
3861 /// Instruction that will be combined with the transition.
3862 Instruction *CombineInst;
3864 /// \brief The instruction that represents the current end of the transition.
3865 /// Since we are faking the promotion until we reach the end of the chain
3866 /// of computation, we need a way to get the current end of the transition.
3867 Instruction *getEndOfTransition() const {
3868 if (InstsToBePromoted.empty())
3870 return InstsToBePromoted.back();
3873 /// \brief Return the index of the original value in the transition.
3874 /// E.g., for "extractelement <2 x i32> c, i32 1" the original value,
3875 /// c, is at index 0.
3876 unsigned getTransitionOriginalValueIdx() const {
3877 assert(isa<ExtractElementInst>(Transition) &&
3878 "Other kind of transitions are not supported yet");
3882 /// \brief Return the index of the index in the transition.
3883 /// E.g., for "extractelement <2 x i32> c, i32 0" the index
3885 unsigned getTransitionIdx() const {
3886 assert(isa<ExtractElementInst>(Transition) &&
3887 "Other kind of transitions are not supported yet");
3891 /// \brief Get the type of the transition.
3892 /// This is the type of the original value.
3893 /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the
3894 /// transition is <2 x i32>.
3895 Type *getTransitionType() const {
3896 return Transition->getOperand(getTransitionOriginalValueIdx())->getType();
3899 /// \brief Promote \p ToBePromoted by moving \p Def downward through.
3900 /// I.e., we have the following sequence:
3901 /// Def = Transition <ty1> a to <ty2>
3902 /// b = ToBePromoted <ty2> Def, ...
3904 /// b = ToBePromoted <ty1> a, ...
3905 /// Def = Transition <ty1> ToBePromoted to <ty2>
3906 void promoteImpl(Instruction *ToBePromoted);
3908 /// \brief Check whether or not it is profitable to promote all the
3909 /// instructions enqueued to be promoted.
3910 bool isProfitableToPromote() {
3911 Value *ValIdx = Transition->getOperand(getTransitionOriginalValueIdx());
3912 unsigned Index = isa<ConstantInt>(ValIdx)
3913 ? cast<ConstantInt>(ValIdx)->getZExtValue()
3915 Type *PromotedType = getTransitionType();
3917 StoreInst *ST = cast<StoreInst>(CombineInst);
3918 unsigned AS = ST->getPointerAddressSpace();
3919 unsigned Align = ST->getAlignment();
3920 // Check if this store is supported.
3921 if (!TLI.allowsMisalignedMemoryAccesses(
3922 TLI.getValueType(ST->getValueOperand()->getType()), AS, Align)) {
3923 // If this is not supported, there is no way we can combine
3924 // the extract with the store.
3928 // The scalar chain of computation has to pay for the transition
3929 // scalar to vector.
3930 // The vector chain has to account for the combining cost.
3931 uint64_t ScalarCost =
3932 TTI.getVectorInstrCost(Transition->getOpcode(), PromotedType, Index);
3933 uint64_t VectorCost = StoreExtractCombineCost;
3934 for (const auto &Inst : InstsToBePromoted) {
3935 // Compute the cost.
3936 // By construction, all instructions being promoted are arithmetic ones.
3937 // Moreover, one argument is a constant that can be viewed as a splat
3939 Value *Arg0 = Inst->getOperand(0);
3940 bool IsArg0Constant = isa<UndefValue>(Arg0) || isa<ConstantInt>(Arg0) ||
3941 isa<ConstantFP>(Arg0);
3942 TargetTransformInfo::OperandValueKind Arg0OVK =
3943 IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
3944 : TargetTransformInfo::OK_AnyValue;
3945 TargetTransformInfo::OperandValueKind Arg1OVK =
3946 !IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
3947 : TargetTransformInfo::OK_AnyValue;
3948 ScalarCost += TTI.getArithmeticInstrCost(
3949 Inst->getOpcode(), Inst->getType(), Arg0OVK, Arg1OVK);
3950 VectorCost += TTI.getArithmeticInstrCost(Inst->getOpcode(), PromotedType,
3953 DEBUG(dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
3954 << ScalarCost << "\nVector: " << VectorCost << '\n');
3955 return ScalarCost > VectorCost;
3958 /// \brief Generate a constant vector with \p Val with the same
3959 /// number of elements as the transition.
3960 /// \p UseSplat defines whether or not \p Val should be replicated
3961 /// accross the whole vector.
3962 /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>,
3963 /// otherwise we generate a vector with as many undef as possible:
3964 /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only
3965 /// used at the index of the extract.
3966 Value *getConstantVector(Constant *Val, bool UseSplat) const {
3967 unsigned ExtractIdx = UINT_MAX;
3969 // If we cannot determine where the constant must be, we have to
3970 // use a splat constant.
3971 Value *ValExtractIdx = Transition->getOperand(getTransitionIdx());
3972 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(ValExtractIdx))
3973 ExtractIdx = CstVal->getSExtValue();
3978 unsigned End = getTransitionType()->getVectorNumElements();
3980 return ConstantVector::getSplat(End, Val);
3982 SmallVector<Constant *, 4> ConstVec;
3983 UndefValue *UndefVal = UndefValue::get(Val->getType());
3984 for (unsigned Idx = 0; Idx != End; ++Idx) {
3985 if (Idx == ExtractIdx)
3986 ConstVec.push_back(Val);
3988 ConstVec.push_back(UndefVal);
3990 return ConstantVector::get(ConstVec);
3993 /// \brief Check if promoting to a vector type an operand at \p OperandIdx
3994 /// in \p Use can trigger undefined behavior.
3995 static bool canCauseUndefinedBehavior(const Instruction *Use,
3996 unsigned OperandIdx) {
3997 // This is not safe to introduce undef when the operand is on
3998 // the right hand side of a division-like instruction.
3999 if (OperandIdx != 1)
4001 switch (Use->getOpcode()) {
4004 case Instruction::SDiv:
4005 case Instruction::UDiv:
4006 case Instruction::SRem:
4007 case Instruction::URem:
4009 case Instruction::FDiv:
4010 case Instruction::FRem:
4011 return !Use->hasNoNaNs();
4013 llvm_unreachable(nullptr);
4017 VectorPromoteHelper(const TargetLowering &TLI, const TargetTransformInfo &TTI,
4018 Instruction *Transition, unsigned CombineCost)
4019 : TLI(TLI), TTI(TTI), Transition(Transition),
4020 StoreExtractCombineCost(CombineCost), CombineInst(nullptr) {
4021 assert(Transition && "Do not know how to promote null");
4024 /// \brief Check if we can promote \p ToBePromoted to \p Type.
4025 bool canPromote(const Instruction *ToBePromoted) const {
4026 // We could support CastInst too.
4027 return isa<BinaryOperator>(ToBePromoted);
4030 /// \brief Check if it is profitable to promote \p ToBePromoted
4031 /// by moving downward the transition through.
4032 bool shouldPromote(const Instruction *ToBePromoted) const {
4033 // Promote only if all the operands can be statically expanded.
4034 // Indeed, we do not want to introduce any new kind of transitions.
4035 for (const Use &U : ToBePromoted->operands()) {
4036 const Value *Val = U.get();
4037 if (Val == getEndOfTransition()) {
4038 // If the use is a division and the transition is on the rhs,
4039 // we cannot promote the operation, otherwise we may create a
4040 // division by zero.
4041 if (canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()))
4045 if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) &&
4046 !isa<ConstantFP>(Val))
4049 // Check that the resulting operation is legal.
4050 int ISDOpcode = TLI.InstructionOpcodeToISD(ToBePromoted->getOpcode());
4053 return StressStoreExtract ||
4054 TLI.isOperationLegalOrCustom(
4055 ISDOpcode, TLI.getValueType(getTransitionType(), true));
4058 /// \brief Check whether or not \p Use can be combined
4059 /// with the transition.
4060 /// I.e., is it possible to do Use(Transition) => AnotherUse?
4061 bool canCombine(const Instruction *Use) { return isa<StoreInst>(Use); }
4063 /// \brief Record \p ToBePromoted as part of the chain to be promoted.
4064 void enqueueForPromotion(Instruction *ToBePromoted) {
4065 InstsToBePromoted.push_back(ToBePromoted);
4068 /// \brief Set the instruction that will be combined with the transition.
4069 void recordCombineInstruction(Instruction *ToBeCombined) {
4070 assert(canCombine(ToBeCombined) && "Unsupported instruction to combine");
4071 CombineInst = ToBeCombined;
4074 /// \brief Promote all the instructions enqueued for promotion if it is
4076 /// \return True if the promotion happened, false otherwise.
4078 // Check if there is something to promote.
4079 // Right now, if we do not have anything to combine with,
4080 // we assume the promotion is not profitable.
4081 if (InstsToBePromoted.empty() || !CombineInst)
4085 if (!StressStoreExtract && !isProfitableToPromote())
4089 for (auto &ToBePromoted : InstsToBePromoted)
4090 promoteImpl(ToBePromoted);
4091 InstsToBePromoted.clear();
4095 } // End of anonymous namespace.
4097 void VectorPromoteHelper::promoteImpl(Instruction *ToBePromoted) {
4098 // At this point, we know that all the operands of ToBePromoted but Def
4099 // can be statically promoted.
4100 // For Def, we need to use its parameter in ToBePromoted:
4101 // b = ToBePromoted ty1 a
4102 // Def = Transition ty1 b to ty2
4103 // Move the transition down.
4104 // 1. Replace all uses of the promoted operation by the transition.
4105 // = ... b => = ... Def.
4106 assert(ToBePromoted->getType() == Transition->getType() &&
4107 "The type of the result of the transition does not match "
4109 ToBePromoted->replaceAllUsesWith(Transition);
4110 // 2. Update the type of the uses.
4111 // b = ToBePromoted ty2 Def => b = ToBePromoted ty1 Def.
4112 Type *TransitionTy = getTransitionType();
4113 ToBePromoted->mutateType(TransitionTy);
4114 // 3. Update all the operands of the promoted operation with promoted
4116 // b = ToBePromoted ty1 Def => b = ToBePromoted ty1 a.
4117 for (Use &U : ToBePromoted->operands()) {
4118 Value *Val = U.get();
4119 Value *NewVal = nullptr;
4120 if (Val == Transition)
4121 NewVal = Transition->getOperand(getTransitionOriginalValueIdx());
4122 else if (isa<UndefValue>(Val) || isa<ConstantInt>(Val) ||
4123 isa<ConstantFP>(Val)) {
4124 // Use a splat constant if it is not safe to use undef.
4125 NewVal = getConstantVector(
4126 cast<Constant>(Val),
4127 isa<UndefValue>(Val) ||
4128 canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()));
4130 llvm_unreachable("Did you modified shouldPromote and forgot to update "
4132 ToBePromoted->setOperand(U.getOperandNo(), NewVal);
4134 Transition->removeFromParent();
4135 Transition->insertAfter(ToBePromoted);
4136 Transition->setOperand(getTransitionOriginalValueIdx(), ToBePromoted);
4139 /// Some targets can do store(extractelement) with one instruction.
4140 /// Try to push the extractelement towards the stores when the target
4141 /// has this feature and this is profitable.
4142 bool CodeGenPrepare::OptimizeExtractElementInst(Instruction *Inst) {
4143 unsigned CombineCost = UINT_MAX;
4144 if (DisableStoreExtract || !TLI ||
4145 (!StressStoreExtract &&
4146 !TLI->canCombineStoreAndExtract(Inst->getOperand(0)->getType(),
4147 Inst->getOperand(1), CombineCost)))
4150 // At this point we know that Inst is a vector to scalar transition.
4151 // Try to move it down the def-use chain, until:
4152 // - We can combine the transition with its single use
4153 // => we got rid of the transition.
4154 // - We escape the current basic block
4155 // => we would need to check that we are moving it at a cheaper place and
4156 // we do not do that for now.
4157 BasicBlock *Parent = Inst->getParent();
4158 DEBUG(dbgs() << "Found an interesting transition: " << *Inst << '\n');
4159 VectorPromoteHelper VPH(*TLI, *TTI, Inst, CombineCost);
4160 // If the transition has more than one use, assume this is not going to be
4162 while (Inst->hasOneUse()) {
4163 Instruction *ToBePromoted = cast<Instruction>(*Inst->user_begin());
4164 DEBUG(dbgs() << "Use: " << *ToBePromoted << '\n');
4166 if (ToBePromoted->getParent() != Parent) {
4167 DEBUG(dbgs() << "Instruction to promote is in a different block ("
4168 << ToBePromoted->getParent()->getName()
4169 << ") than the transition (" << Parent->getName() << ").\n");
4173 if (VPH.canCombine(ToBePromoted)) {
4174 DEBUG(dbgs() << "Assume " << *Inst << '\n'
4175 << "will be combined with: " << *ToBePromoted << '\n');
4176 VPH.recordCombineInstruction(ToBePromoted);
4177 bool Changed = VPH.promote();
4178 NumStoreExtractExposed += Changed;
4182 DEBUG(dbgs() << "Try promoting.\n");
4183 if (!VPH.canPromote(ToBePromoted) || !VPH.shouldPromote(ToBePromoted))
4186 DEBUG(dbgs() << "Promoting is possible... Enqueue for promotion!\n");
4188 VPH.enqueueForPromotion(ToBePromoted);
4189 Inst = ToBePromoted;
4194 bool CodeGenPrepare::OptimizeInst(Instruction *I, bool& ModifiedDT) {
4195 if (PHINode *P = dyn_cast<PHINode>(I)) {
4196 // It is possible for very late stage optimizations (such as SimplifyCFG)
4197 // to introduce PHI nodes too late to be cleaned up. If we detect such a
4198 // trivial PHI, go ahead and zap it here.
4199 if (Value *V = SimplifyInstruction(P, TLI ? TLI->getDataLayout() : nullptr,
4201 P->replaceAllUsesWith(V);
4202 P->eraseFromParent();
4209 if (CastInst *CI = dyn_cast<CastInst>(I)) {
4210 // If the source of the cast is a constant, then this should have
4211 // already been constant folded. The only reason NOT to constant fold
4212 // it is if something (e.g. LSR) was careful to place the constant
4213 // evaluation in a block other than then one that uses it (e.g. to hoist
4214 // the address of globals out of a loop). If this is the case, we don't
4215 // want to forward-subst the cast.
4216 if (isa<Constant>(CI->getOperand(0)))
4219 if (TLI && OptimizeNoopCopyExpression(CI, *TLI))
4222 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
4223 /// Sink a zext or sext into its user blocks if the target type doesn't
4224 /// fit in one register
4225 if (TLI && TLI->getTypeAction(CI->getContext(),
4226 TLI->getValueType(CI->getType())) ==
4227 TargetLowering::TypeExpandInteger) {
4228 return SinkCast(CI);
4230 bool MadeChange = MoveExtToFormExtLoad(I);
4231 return MadeChange | OptimizeExtUses(I);
4237 if (CmpInst *CI = dyn_cast<CmpInst>(I))
4238 if (!TLI || !TLI->hasMultipleConditionRegisters())
4239 return OptimizeCmpExpression(CI);
4241 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
4243 return OptimizeMemoryInst(I, I->getOperand(0), LI->getType());
4247 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
4249 return OptimizeMemoryInst(I, SI->getOperand(1),
4250 SI->getOperand(0)->getType());
4254 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(I);
4256 if (BinOp && (BinOp->getOpcode() == Instruction::AShr ||
4257 BinOp->getOpcode() == Instruction::LShr)) {
4258 ConstantInt *CI = dyn_cast<ConstantInt>(BinOp->getOperand(1));
4259 if (TLI && CI && TLI->hasExtractBitsInsn())
4260 return OptimizeExtractBits(BinOp, CI, *TLI);
4265 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
4266 if (GEPI->hasAllZeroIndices()) {
4267 /// The GEP operand must be a pointer, so must its result -> BitCast
4268 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
4269 GEPI->getName(), GEPI);
4270 GEPI->replaceAllUsesWith(NC);
4271 GEPI->eraseFromParent();
4273 OptimizeInst(NC, ModifiedDT);
4279 if (CallInst *CI = dyn_cast<CallInst>(I))
4280 return OptimizeCallInst(CI, ModifiedDT);
4282 if (SelectInst *SI = dyn_cast<SelectInst>(I))
4283 return OptimizeSelectInst(SI);
4285 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I))
4286 return OptimizeShuffleVectorInst(SVI);
4288 if (isa<ExtractElementInst>(I))
4289 return OptimizeExtractElementInst(I);
4294 // In this pass we look for GEP and cast instructions that are used
4295 // across basic blocks and rewrite them to improve basic-block-at-a-time
4297 bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB, bool& ModifiedDT) {
4299 bool MadeChange = false;
4301 CurInstIterator = BB.begin();
4302 while (CurInstIterator != BB.end()) {
4303 MadeChange |= OptimizeInst(CurInstIterator++, ModifiedDT);
4307 MadeChange |= DupRetToEnableTailCallOpts(&BB);
4312 // llvm.dbg.value is far away from the value then iSel may not be able
4313 // handle it properly. iSel will drop llvm.dbg.value if it can not
4314 // find a node corresponding to the value.
4315 bool CodeGenPrepare::PlaceDbgValues(Function &F) {
4316 bool MadeChange = false;
4317 for (BasicBlock &BB : F) {
4318 Instruction *PrevNonDbgInst = nullptr;
4319 for (BasicBlock::iterator BI = BB.begin(), BE = BB.end(); BI != BE;) {
4320 Instruction *Insn = BI++;
4321 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
4322 // Leave dbg.values that refer to an alloca alone. These
4323 // instrinsics describe the address of a variable (= the alloca)
4324 // being taken. They should not be moved next to the alloca
4325 // (and to the beginning of the scope), but rather stay close to
4326 // where said address is used.
4327 if (!DVI || (DVI->getValue() && isa<AllocaInst>(DVI->getValue()))) {
4328 PrevNonDbgInst = Insn;
4332 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
4333 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
4334 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
4335 DVI->removeFromParent();
4336 if (isa<PHINode>(VI))
4337 DVI->insertBefore(VI->getParent()->getFirstInsertionPt());
4339 DVI->insertAfter(VI);
4348 // If there is a sequence that branches based on comparing a single bit
4349 // against zero that can be combined into a single instruction, and the
4350 // target supports folding these into a single instruction, sink the
4351 // mask and compare into the branch uses. Do this before OptimizeBlock ->
4352 // OptimizeInst -> OptimizeCmpExpression, which perturbs the pattern being
4354 bool CodeGenPrepare::sinkAndCmp(Function &F) {
4355 if (!EnableAndCmpSinking)
4357 if (!TLI || !TLI->isMaskAndBranchFoldingLegal())
4359 bool MadeChange = false;
4360 for (Function::iterator I = F.begin(), E = F.end(); I != E; ) {
4361 BasicBlock *BB = I++;
4363 // Does this BB end with the following?
4364 // %andVal = and %val, #single-bit-set
4365 // %icmpVal = icmp %andResult, 0
4366 // br i1 %cmpVal label %dest1, label %dest2"
4367 BranchInst *Brcc = dyn_cast<BranchInst>(BB->getTerminator());
4368 if (!Brcc || !Brcc->isConditional())
4370 ICmpInst *Cmp = dyn_cast<ICmpInst>(Brcc->getOperand(0));
4371 if (!Cmp || Cmp->getParent() != BB)
4373 ConstantInt *Zero = dyn_cast<ConstantInt>(Cmp->getOperand(1));
4374 if (!Zero || !Zero->isZero())
4376 Instruction *And = dyn_cast<Instruction>(Cmp->getOperand(0));
4377 if (!And || And->getOpcode() != Instruction::And || And->getParent() != BB)
4379 ConstantInt* Mask = dyn_cast<ConstantInt>(And->getOperand(1));
4380 if (!Mask || !Mask->getUniqueInteger().isPowerOf2())
4382 DEBUG(dbgs() << "found and; icmp ?,0; brcc\n"); DEBUG(BB->dump());
4384 // Push the "and; icmp" for any users that are conditional branches.
4385 // Since there can only be one branch use per BB, we don't need to keep
4386 // track of which BBs we insert into.
4387 for (Value::use_iterator UI = Cmp->use_begin(), E = Cmp->use_end();
4391 BranchInst *BrccUser = dyn_cast<BranchInst>(*UI);
4393 if (!BrccUser || !BrccUser->isConditional())
4395 BasicBlock *UserBB = BrccUser->getParent();
4396 if (UserBB == BB) continue;
4397 DEBUG(dbgs() << "found Brcc use\n");
4399 // Sink the "and; icmp" to use.
4401 BinaryOperator *NewAnd =
4402 BinaryOperator::CreateAnd(And->getOperand(0), And->getOperand(1), "",
4405 CmpInst::Create(Cmp->getOpcode(), Cmp->getPredicate(), NewAnd, Zero,
4409 DEBUG(BrccUser->getParent()->dump());
4415 /// \brief Retrieve the probabilities of a conditional branch. Returns true on
4416 /// success, or returns false if no or invalid metadata was found.
4417 static bool extractBranchMetadata(BranchInst *BI,
4418 uint64_t &ProbTrue, uint64_t &ProbFalse) {
4419 assert(BI->isConditional() &&
4420 "Looking for probabilities on unconditional branch?");
4421 auto *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
4422 if (!ProfileData || ProfileData->getNumOperands() != 3)
4425 const auto *CITrue =
4426 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(1));
4427 const auto *CIFalse =
4428 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(2));
4429 if (!CITrue || !CIFalse)
4432 ProbTrue = CITrue->getValue().getZExtValue();
4433 ProbFalse = CIFalse->getValue().getZExtValue();
4438 /// \brief Scale down both weights to fit into uint32_t.
4439 static void scaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) {
4440 uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse;
4441 uint32_t Scale = (NewMax / UINT32_MAX) + 1;
4442 NewTrue = NewTrue / Scale;
4443 NewFalse = NewFalse / Scale;
4446 /// \brief Some targets prefer to split a conditional branch like:
4448 /// %0 = icmp ne i32 %a, 0
4449 /// %1 = icmp ne i32 %b, 0
4450 /// %or.cond = or i1 %0, %1
4451 /// br i1 %or.cond, label %TrueBB, label %FalseBB
4453 /// into multiple branch instructions like:
4456 /// %0 = icmp ne i32 %a, 0
4457 /// br i1 %0, label %TrueBB, label %bb2
4459 /// %1 = icmp ne i32 %b, 0
4460 /// br i1 %1, label %TrueBB, label %FalseBB
4462 /// This usually allows instruction selection to do even further optimizations
4463 /// and combine the compare with the branch instruction. Currently this is
4464 /// applied for targets which have "cheap" jump instructions.
4466 /// FIXME: Remove the (equivalent?) implementation in SelectionDAG.
4468 bool CodeGenPrepare::splitBranchCondition(Function &F) {
4469 if (!TM || TM->Options.EnableFastISel != true ||
4470 !TLI || TLI->isJumpExpensive())
4473 bool MadeChange = false;
4474 for (auto &BB : F) {
4475 // Does this BB end with the following?
4476 // %cond1 = icmp|fcmp|binary instruction ...
4477 // %cond2 = icmp|fcmp|binary instruction ...
4478 // %cond.or = or|and i1 %cond1, cond2
4479 // br i1 %cond.or label %dest1, label %dest2"
4480 BinaryOperator *LogicOp;
4481 BasicBlock *TBB, *FBB;
4482 if (!match(BB.getTerminator(), m_Br(m_OneUse(m_BinOp(LogicOp)), TBB, FBB)))
4486 Value *Cond1, *Cond2;
4487 if (match(LogicOp, m_And(m_OneUse(m_Value(Cond1)),
4488 m_OneUse(m_Value(Cond2)))))
4489 Opc = Instruction::And;
4490 else if (match(LogicOp, m_Or(m_OneUse(m_Value(Cond1)),
4491 m_OneUse(m_Value(Cond2)))))
4492 Opc = Instruction::Or;
4496 if (!match(Cond1, m_CombineOr(m_Cmp(), m_BinOp())) ||
4497 !match(Cond2, m_CombineOr(m_Cmp(), m_BinOp())) )
4500 DEBUG(dbgs() << "Before branch condition splitting\n"; BB.dump());
4503 auto *InsertBefore = std::next(Function::iterator(BB))
4504 .getNodePtrUnchecked();
4505 auto TmpBB = BasicBlock::Create(BB.getContext(),
4506 BB.getName() + ".cond.split",
4507 BB.getParent(), InsertBefore);
4509 // Update original basic block by using the first condition directly by the
4510 // branch instruction and removing the no longer needed and/or instruction.
4511 auto *Br1 = cast<BranchInst>(BB.getTerminator());
4512 Br1->setCondition(Cond1);
4513 LogicOp->eraseFromParent();
4515 // Depending on the conditon we have to either replace the true or the false
4516 // successor of the original branch instruction.
4517 if (Opc == Instruction::And)
4518 Br1->setSuccessor(0, TmpBB);
4520 Br1->setSuccessor(1, TmpBB);
4522 // Fill in the new basic block.
4523 auto *Br2 = IRBuilder<>(TmpBB).CreateCondBr(Cond2, TBB, FBB);
4524 if (auto *I = dyn_cast<Instruction>(Cond2)) {
4525 I->removeFromParent();
4526 I->insertBefore(Br2);
4529 // Update PHI nodes in both successors. The original BB needs to be
4530 // replaced in one succesor's PHI nodes, because the branch comes now from
4531 // the newly generated BB (NewBB). In the other successor we need to add one
4532 // incoming edge to the PHI nodes, because both branch instructions target
4533 // now the same successor. Depending on the original branch condition
4534 // (and/or) we have to swap the successors (TrueDest, FalseDest), so that
4535 // we perfrom the correct update for the PHI nodes.
4536 // This doesn't change the successor order of the just created branch
4537 // instruction (or any other instruction).
4538 if (Opc == Instruction::Or)
4539 std::swap(TBB, FBB);
4541 // Replace the old BB with the new BB.
4542 for (auto &I : *TBB) {
4543 PHINode *PN = dyn_cast<PHINode>(&I);
4547 while ((i = PN->getBasicBlockIndex(&BB)) >= 0)
4548 PN->setIncomingBlock(i, TmpBB);
4551 // Add another incoming edge form the new BB.
4552 for (auto &I : *FBB) {
4553 PHINode *PN = dyn_cast<PHINode>(&I);
4556 auto *Val = PN->getIncomingValueForBlock(&BB);
4557 PN->addIncoming(Val, TmpBB);
4560 // Update the branch weights (from SelectionDAGBuilder::
4561 // FindMergedConditions).
4562 if (Opc == Instruction::Or) {
4563 // Codegen X | Y as:
4572 // We have flexibility in setting Prob for BB1 and Prob for NewBB.
4573 // The requirement is that
4574 // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB)
4575 // = TrueProb for orignal BB.
4576 // Assuming the orignal weights are A and B, one choice is to set BB1's
4577 // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice
4579 // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB.
4580 // Another choice is to assume TrueProb for BB1 equals to TrueProb for
4581 // TmpBB, but the math is more complicated.
4582 uint64_t TrueWeight, FalseWeight;
4583 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
4584 uint64_t NewTrueWeight = TrueWeight;
4585 uint64_t NewFalseWeight = TrueWeight + 2 * FalseWeight;
4586 scaleWeights(NewTrueWeight, NewFalseWeight);
4587 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
4588 .createBranchWeights(TrueWeight, FalseWeight));
4590 NewTrueWeight = TrueWeight;
4591 NewFalseWeight = 2 * FalseWeight;
4592 scaleWeights(NewTrueWeight, NewFalseWeight);
4593 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
4594 .createBranchWeights(TrueWeight, FalseWeight));
4597 // Codegen X & Y as:
4605 // This requires creation of TmpBB after CurBB.
4607 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
4608 // The requirement is that
4609 // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB)
4610 // = FalseProb for orignal BB.
4611 // Assuming the orignal weights are A and B, one choice is to set BB1's
4612 // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice
4614 // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB.
4615 uint64_t TrueWeight, FalseWeight;
4616 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
4617 uint64_t NewTrueWeight = 2 * TrueWeight + FalseWeight;
4618 uint64_t NewFalseWeight = FalseWeight;
4619 scaleWeights(NewTrueWeight, NewFalseWeight);
4620 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
4621 .createBranchWeights(TrueWeight, FalseWeight));
4623 NewTrueWeight = 2 * TrueWeight;
4624 NewFalseWeight = FalseWeight;
4625 scaleWeights(NewTrueWeight, NewFalseWeight);
4626 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
4627 .createBranchWeights(TrueWeight, FalseWeight));
4631 // Request DOM Tree update.
4632 // Note: No point in getting fancy here, since the DT info is never
4633 // available to CodeGenPrepare and the existing update code is broken
4639 DEBUG(dbgs() << "After branch condition splitting\n"; BB.dump();