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/Analysis/ValueTracking.h"
24 #include "llvm/IR/CallSite.h"
25 #include "llvm/IR/Constants.h"
26 #include "llvm/IR/DataLayout.h"
27 #include "llvm/IR/DerivedTypes.h"
28 #include "llvm/IR/Dominators.h"
29 #include "llvm/IR/Function.h"
30 #include "llvm/IR/GetElementPtrTypeIterator.h"
31 #include "llvm/IR/IRBuilder.h"
32 #include "llvm/IR/InlineAsm.h"
33 #include "llvm/IR/Instructions.h"
34 #include "llvm/IR/IntrinsicInst.h"
35 #include "llvm/IR/MDBuilder.h"
36 #include "llvm/IR/PatternMatch.h"
37 #include "llvm/IR/Statepoint.h"
38 #include "llvm/IR/ValueHandle.h"
39 #include "llvm/IR/ValueMap.h"
40 #include "llvm/Pass.h"
41 #include "llvm/Support/CommandLine.h"
42 #include "llvm/Support/Debug.h"
43 #include "llvm/Support/raw_ostream.h"
44 #include "llvm/Target/TargetLowering.h"
45 #include "llvm/Target/TargetSubtargetInfo.h"
46 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
47 #include "llvm/Transforms/Utils/BuildLibCalls.h"
48 #include "llvm/Transforms/Utils/BypassSlowDivision.h"
49 #include "llvm/Transforms/Utils/Local.h"
50 #include "llvm/Transforms/Utils/SimplifyLibCalls.h"
52 using namespace llvm::PatternMatch;
54 #define DEBUG_TYPE "codegenprepare"
56 STATISTIC(NumBlocksElim, "Number of blocks eliminated");
57 STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated");
58 STATISTIC(NumGEPsElim, "Number of GEPs converted to casts");
59 STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "
61 STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "
63 STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "
64 "computations were sunk");
65 STATISTIC(NumExtsMoved, "Number of [s|z]ext instructions combined with loads");
66 STATISTIC(NumExtUses, "Number of uses of [s|z]ext instructions optimized");
67 STATISTIC(NumRetsDup, "Number of return instructions duplicated");
68 STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved");
69 STATISTIC(NumSelectsExpanded, "Number of selects turned into branches");
70 STATISTIC(NumAndCmpsMoved, "Number of and/cmp's pushed into branches");
71 STATISTIC(NumStoreExtractExposed, "Number of store(extractelement) exposed");
73 static cl::opt<bool> DisableBranchOpts(
74 "disable-cgp-branch-opts", cl::Hidden, cl::init(false),
75 cl::desc("Disable branch optimizations in CodeGenPrepare"));
78 DisableGCOpts("disable-cgp-gc-opts", cl::Hidden, cl::init(false),
79 cl::desc("Disable GC optimizations in CodeGenPrepare"));
81 static cl::opt<bool> DisableSelectToBranch(
82 "disable-cgp-select2branch", cl::Hidden, cl::init(false),
83 cl::desc("Disable select to branch conversion."));
85 static cl::opt<bool> AddrSinkUsingGEPs(
86 "addr-sink-using-gep", cl::Hidden, cl::init(false),
87 cl::desc("Address sinking in CGP using GEPs."));
89 static cl::opt<bool> EnableAndCmpSinking(
90 "enable-andcmp-sinking", cl::Hidden, cl::init(true),
91 cl::desc("Enable sinkinig and/cmp into branches."));
93 static cl::opt<bool> DisableStoreExtract(
94 "disable-cgp-store-extract", cl::Hidden, cl::init(false),
95 cl::desc("Disable store(extract) optimizations in CodeGenPrepare"));
97 static cl::opt<bool> StressStoreExtract(
98 "stress-cgp-store-extract", cl::Hidden, cl::init(false),
99 cl::desc("Stress test store(extract) optimizations in CodeGenPrepare"));
101 static cl::opt<bool> DisableExtLdPromotion(
102 "disable-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
103 cl::desc("Disable ext(promotable(ld)) -> promoted(ext(ld)) optimization in "
106 static cl::opt<bool> StressExtLdPromotion(
107 "stress-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
108 cl::desc("Stress test ext(promotable(ld)) -> promoted(ext(ld)) "
109 "optimization in CodeGenPrepare"));
112 typedef SmallPtrSet<Instruction *, 16> SetOfInstrs;
113 typedef PointerIntPair<Type *, 1, bool> TypeIsSExt;
114 typedef DenseMap<Instruction *, TypeIsSExt> InstrToOrigTy;
115 class TypePromotionTransaction;
117 class CodeGenPrepare : public FunctionPass {
118 const TargetMachine *TM;
119 const TargetLowering *TLI;
120 const TargetTransformInfo *TTI;
121 const TargetLibraryInfo *TLInfo;
123 /// As we scan instructions optimizing them, this is the next instruction
124 /// to optimize. Transforms that can invalidate this should update it.
125 BasicBlock::iterator CurInstIterator;
127 /// Keeps track of non-local addresses that have been sunk into a block.
128 /// This allows us to avoid inserting duplicate code for blocks with
129 /// multiple load/stores of the same address.
130 ValueMap<Value*, Value*> SunkAddrs;
132 /// Keeps track of all instructions inserted for the current function.
133 SetOfInstrs InsertedInsts;
134 /// Keeps track of the type of the related instruction before their
135 /// promotion for the current function.
136 InstrToOrigTy PromotedInsts;
138 /// True if CFG is modified in any way.
141 /// True if optimizing for size.
144 /// DataLayout for the Function being processed.
145 const DataLayout *DL;
148 static char ID; // Pass identification, replacement for typeid
149 explicit CodeGenPrepare(const TargetMachine *TM = nullptr)
150 : FunctionPass(ID), TM(TM), TLI(nullptr), TTI(nullptr), DL(nullptr) {
151 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
153 bool runOnFunction(Function &F) override;
155 const char *getPassName() const override { return "CodeGen Prepare"; }
157 void getAnalysisUsage(AnalysisUsage &AU) const override {
158 AU.addPreserved<DominatorTreeWrapperPass>();
159 AU.addRequired<TargetLibraryInfoWrapperPass>();
160 AU.addRequired<TargetTransformInfoWrapperPass>();
164 bool eliminateFallThrough(Function &F);
165 bool eliminateMostlyEmptyBlocks(Function &F);
166 bool canMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
167 void eliminateMostlyEmptyBlock(BasicBlock *BB);
168 bool optimizeBlock(BasicBlock &BB, bool& ModifiedDT);
169 bool optimizeInst(Instruction *I, bool& ModifiedDT);
170 bool optimizeMemoryInst(Instruction *I, Value *Addr,
171 Type *AccessTy, unsigned AS);
172 bool optimizeInlineAsmInst(CallInst *CS);
173 bool optimizeCallInst(CallInst *CI, bool& ModifiedDT);
174 bool moveExtToFormExtLoad(Instruction *&I);
175 bool optimizeExtUses(Instruction *I);
176 bool optimizeSelectInst(SelectInst *SI);
177 bool optimizeShuffleVectorInst(ShuffleVectorInst *SI);
178 bool optimizeExtractElementInst(Instruction *Inst);
179 bool dupRetToEnableTailCallOpts(BasicBlock *BB);
180 bool placeDbgValues(Function &F);
181 bool sinkAndCmp(Function &F);
182 bool extLdPromotion(TypePromotionTransaction &TPT, LoadInst *&LI,
184 const SmallVectorImpl<Instruction *> &Exts,
185 unsigned CreatedInstCost);
186 bool splitBranchCondition(Function &F);
187 bool simplifyOffsetableRelocate(Instruction &I);
188 void stripInvariantGroupMetadata(Instruction &I);
192 char CodeGenPrepare::ID = 0;
193 INITIALIZE_TM_PASS(CodeGenPrepare, "codegenprepare",
194 "Optimize for code generation", false, false)
196 FunctionPass *llvm::createCodeGenPreparePass(const TargetMachine *TM) {
197 return new CodeGenPrepare(TM);
200 bool CodeGenPrepare::runOnFunction(Function &F) {
201 if (skipOptnoneFunction(F))
204 DL = &F.getParent()->getDataLayout();
206 bool EverMadeChange = false;
207 // Clear per function information.
208 InsertedInsts.clear();
209 PromotedInsts.clear();
213 TLI = TM->getSubtargetImpl(F)->getTargetLowering();
214 TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
215 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
216 OptSize = F.optForSize();
218 /// This optimization identifies DIV instructions that can be
219 /// profitably bypassed and carried out with a shorter, faster divide.
220 if (!OptSize && TLI && TLI->isSlowDivBypassed()) {
221 const DenseMap<unsigned int, unsigned int> &BypassWidths =
222 TLI->getBypassSlowDivWidths();
223 for (Function::iterator I = F.begin(); I != F.end(); I++)
224 EverMadeChange |= bypassSlowDivision(F, I, BypassWidths);
227 // Eliminate blocks that contain only PHI nodes and an
228 // unconditional branch.
229 EverMadeChange |= eliminateMostlyEmptyBlocks(F);
231 // llvm.dbg.value is far away from the value then iSel may not be able
232 // handle it properly. iSel will drop llvm.dbg.value if it can not
233 // find a node corresponding to the value.
234 EverMadeChange |= placeDbgValues(F);
236 // If there is a mask, compare against zero, and branch that can be combined
237 // into a single target instruction, push the mask and compare into branch
238 // users. Do this before OptimizeBlock -> OptimizeInst ->
239 // OptimizeCmpExpression, which perturbs the pattern being searched for.
240 if (!DisableBranchOpts) {
241 EverMadeChange |= sinkAndCmp(F);
242 EverMadeChange |= splitBranchCondition(F);
245 bool MadeChange = true;
248 for (Function::iterator I = F.begin(); I != F.end(); ) {
249 BasicBlock *BB = &*I++;
250 bool ModifiedDTOnIteration = false;
251 MadeChange |= optimizeBlock(*BB, ModifiedDTOnIteration);
253 // Restart BB iteration if the dominator tree of the Function was changed
254 if (ModifiedDTOnIteration)
257 EverMadeChange |= MadeChange;
262 if (!DisableBranchOpts) {
264 SmallPtrSet<BasicBlock*, 8> WorkList;
265 for (BasicBlock &BB : F) {
266 SmallVector<BasicBlock *, 2> Successors(succ_begin(&BB), succ_end(&BB));
267 MadeChange |= ConstantFoldTerminator(&BB, true);
268 if (!MadeChange) continue;
270 for (SmallVectorImpl<BasicBlock*>::iterator
271 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
272 if (pred_begin(*II) == pred_end(*II))
273 WorkList.insert(*II);
276 // Delete the dead blocks and any of their dead successors.
277 MadeChange |= !WorkList.empty();
278 while (!WorkList.empty()) {
279 BasicBlock *BB = *WorkList.begin();
281 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
285 for (SmallVectorImpl<BasicBlock*>::iterator
286 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
287 if (pred_begin(*II) == pred_end(*II))
288 WorkList.insert(*II);
291 // Merge pairs of basic blocks with unconditional branches, connected by
293 if (EverMadeChange || MadeChange)
294 MadeChange |= eliminateFallThrough(F);
296 EverMadeChange |= MadeChange;
299 if (!DisableGCOpts) {
300 SmallVector<Instruction *, 2> Statepoints;
301 for (BasicBlock &BB : F)
302 for (Instruction &I : BB)
304 Statepoints.push_back(&I);
305 for (auto &I : Statepoints)
306 EverMadeChange |= simplifyOffsetableRelocate(*I);
309 return EverMadeChange;
312 /// Merge basic blocks which are connected by a single edge, where one of the
313 /// basic blocks has a single successor pointing to the other basic block,
314 /// which has a single predecessor.
315 bool CodeGenPrepare::eliminateFallThrough(Function &F) {
316 bool Changed = false;
317 // Scan all of the blocks in the function, except for the entry block.
318 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
319 BasicBlock *BB = &*I++;
320 // If the destination block has a single pred, then this is a trivial
321 // edge, just collapse it.
322 BasicBlock *SinglePred = BB->getSinglePredecessor();
324 // Don't merge if BB's address is taken.
325 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
327 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
328 if (Term && !Term->isConditional()) {
330 DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n");
331 // Remember if SinglePred was the entry block of the function.
332 // If so, we will need to move BB back to the entry position.
333 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
334 MergeBasicBlockIntoOnlyPred(BB, nullptr);
336 if (isEntry && BB != &BB->getParent()->getEntryBlock())
337 BB->moveBefore(&BB->getParent()->getEntryBlock());
339 // We have erased a block. Update the iterator.
340 I = BB->getIterator();
346 /// Eliminate blocks that contain only PHI nodes, debug info directives, and an
347 /// unconditional branch. Passes before isel (e.g. LSR/loopsimplify) often split
348 /// edges in ways that are non-optimal for isel. Start by eliminating these
349 /// blocks so we can split them the way we want them.
350 bool CodeGenPrepare::eliminateMostlyEmptyBlocks(Function &F) {
351 bool MadeChange = false;
352 // Note that this intentionally skips the entry block.
353 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
354 BasicBlock *BB = &*I++;
356 // If this block doesn't end with an uncond branch, ignore it.
357 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
358 if (!BI || !BI->isUnconditional())
361 // If the instruction before the branch (skipping debug info) isn't a phi
362 // node, then other stuff is happening here.
363 BasicBlock::iterator BBI = BI->getIterator();
364 if (BBI != BB->begin()) {
366 while (isa<DbgInfoIntrinsic>(BBI)) {
367 if (BBI == BB->begin())
371 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
375 // Do not break infinite loops.
376 BasicBlock *DestBB = BI->getSuccessor(0);
380 if (!canMergeBlocks(BB, DestBB))
383 eliminateMostlyEmptyBlock(BB);
389 /// Return true if we can merge BB into DestBB if there is a single
390 /// unconditional branch between them, and BB contains no other non-phi
392 bool CodeGenPrepare::canMergeBlocks(const BasicBlock *BB,
393 const BasicBlock *DestBB) const {
394 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
395 // the successor. If there are more complex condition (e.g. preheaders),
396 // don't mess around with them.
397 BasicBlock::const_iterator BBI = BB->begin();
398 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
399 for (const User *U : PN->users()) {
400 const Instruction *UI = cast<Instruction>(U);
401 if (UI->getParent() != DestBB || !isa<PHINode>(UI))
403 // If User is inside DestBB block and it is a PHINode then check
404 // incoming value. If incoming value is not from BB then this is
405 // a complex condition (e.g. preheaders) we want to avoid here.
406 if (UI->getParent() == DestBB) {
407 if (const PHINode *UPN = dyn_cast<PHINode>(UI))
408 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
409 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
410 if (Insn && Insn->getParent() == BB &&
411 Insn->getParent() != UPN->getIncomingBlock(I))
418 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
419 // and DestBB may have conflicting incoming values for the block. If so, we
420 // can't merge the block.
421 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
422 if (!DestBBPN) return true; // no conflict.
424 // Collect the preds of BB.
425 SmallPtrSet<const BasicBlock*, 16> BBPreds;
426 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
427 // It is faster to get preds from a PHI than with pred_iterator.
428 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
429 BBPreds.insert(BBPN->getIncomingBlock(i));
431 BBPreds.insert(pred_begin(BB), pred_end(BB));
434 // Walk the preds of DestBB.
435 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
436 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
437 if (BBPreds.count(Pred)) { // Common predecessor?
438 BBI = DestBB->begin();
439 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
440 const Value *V1 = PN->getIncomingValueForBlock(Pred);
441 const Value *V2 = PN->getIncomingValueForBlock(BB);
443 // If V2 is a phi node in BB, look up what the mapped value will be.
444 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
445 if (V2PN->getParent() == BB)
446 V2 = V2PN->getIncomingValueForBlock(Pred);
448 // If there is a conflict, bail out.
449 if (V1 != V2) return false;
458 /// Eliminate a basic block that has only phi's and an unconditional branch in
460 void CodeGenPrepare::eliminateMostlyEmptyBlock(BasicBlock *BB) {
461 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
462 BasicBlock *DestBB = BI->getSuccessor(0);
464 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
466 // If the destination block has a single pred, then this is a trivial edge,
468 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
469 if (SinglePred != DestBB) {
470 // Remember if SinglePred was the entry block of the function. If so, we
471 // will need to move BB back to the entry position.
472 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
473 MergeBasicBlockIntoOnlyPred(DestBB, nullptr);
475 if (isEntry && BB != &BB->getParent()->getEntryBlock())
476 BB->moveBefore(&BB->getParent()->getEntryBlock());
478 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
483 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
484 // to handle the new incoming edges it is about to have.
486 for (BasicBlock::iterator BBI = DestBB->begin();
487 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
488 // Remove the incoming value for BB, and remember it.
489 Value *InVal = PN->removeIncomingValue(BB, false);
491 // Two options: either the InVal is a phi node defined in BB or it is some
492 // value that dominates BB.
493 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
494 if (InValPhi && InValPhi->getParent() == BB) {
495 // Add all of the input values of the input PHI as inputs of this phi.
496 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
497 PN->addIncoming(InValPhi->getIncomingValue(i),
498 InValPhi->getIncomingBlock(i));
500 // Otherwise, add one instance of the dominating value for each edge that
501 // we will be adding.
502 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
503 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
504 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
506 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
507 PN->addIncoming(InVal, *PI);
512 // The PHIs are now updated, change everything that refers to BB to use
513 // DestBB and remove BB.
514 BB->replaceAllUsesWith(DestBB);
515 BB->eraseFromParent();
518 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
521 // Computes a map of base pointer relocation instructions to corresponding
522 // derived pointer relocation instructions given a vector of all relocate calls
523 static void computeBaseDerivedRelocateMap(
524 const SmallVectorImpl<User *> &AllRelocateCalls,
525 DenseMap<IntrinsicInst *, SmallVector<IntrinsicInst *, 2>> &
527 // Collect information in two maps: one primarily for locating the base object
528 // while filling the second map; the second map is the final structure holding
529 // a mapping between Base and corresponding Derived relocate calls
530 DenseMap<std::pair<unsigned, unsigned>, IntrinsicInst *> RelocateIdxMap;
531 for (auto &U : AllRelocateCalls) {
532 GCRelocateOperands ThisRelocate(U);
533 IntrinsicInst *I = cast<IntrinsicInst>(U);
534 auto K = std::make_pair(ThisRelocate.getBasePtrIndex(),
535 ThisRelocate.getDerivedPtrIndex());
536 RelocateIdxMap.insert(std::make_pair(K, I));
538 for (auto &Item : RelocateIdxMap) {
539 std::pair<unsigned, unsigned> Key = Item.first;
540 if (Key.first == Key.second)
541 // Base relocation: nothing to insert
544 IntrinsicInst *I = Item.second;
545 auto BaseKey = std::make_pair(Key.first, Key.first);
547 // We're iterating over RelocateIdxMap so we cannot modify it.
548 auto MaybeBase = RelocateIdxMap.find(BaseKey);
549 if (MaybeBase == RelocateIdxMap.end())
550 // TODO: We might want to insert a new base object relocate and gep off
551 // that, if there are enough derived object relocates.
554 RelocateInstMap[MaybeBase->second].push_back(I);
558 // Accepts a GEP and extracts the operands into a vector provided they're all
559 // small integer constants
560 static bool getGEPSmallConstantIntOffsetV(GetElementPtrInst *GEP,
561 SmallVectorImpl<Value *> &OffsetV) {
562 for (unsigned i = 1; i < GEP->getNumOperands(); i++) {
563 // Only accept small constant integer operands
564 auto Op = dyn_cast<ConstantInt>(GEP->getOperand(i));
565 if (!Op || Op->getZExtValue() > 20)
569 for (unsigned i = 1; i < GEP->getNumOperands(); i++)
570 OffsetV.push_back(GEP->getOperand(i));
574 // Takes a RelocatedBase (base pointer relocation instruction) and Targets to
575 // replace, computes a replacement, and affects it.
577 simplifyRelocatesOffABase(IntrinsicInst *RelocatedBase,
578 const SmallVectorImpl<IntrinsicInst *> &Targets) {
579 bool MadeChange = false;
580 for (auto &ToReplace : Targets) {
581 GCRelocateOperands MasterRelocate(RelocatedBase);
582 GCRelocateOperands ThisRelocate(ToReplace);
584 assert(ThisRelocate.getBasePtrIndex() == MasterRelocate.getBasePtrIndex() &&
585 "Not relocating a derived object of the original base object");
586 if (ThisRelocate.getBasePtrIndex() == ThisRelocate.getDerivedPtrIndex()) {
587 // A duplicate relocate call. TODO: coalesce duplicates.
591 Value *Base = ThisRelocate.getBasePtr();
592 auto Derived = dyn_cast<GetElementPtrInst>(ThisRelocate.getDerivedPtr());
593 if (!Derived || Derived->getPointerOperand() != Base)
596 SmallVector<Value *, 2> OffsetV;
597 if (!getGEPSmallConstantIntOffsetV(Derived, OffsetV))
600 // Create a Builder and replace the target callsite with a gep
601 assert(RelocatedBase->getNextNode() && "Should always have one since it's not a terminator");
603 // Insert after RelocatedBase
604 IRBuilder<> Builder(RelocatedBase->getNextNode());
605 Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
607 // If gc_relocate does not match the actual type, cast it to the right type.
608 // In theory, there must be a bitcast after gc_relocate if the type does not
609 // match, and we should reuse it to get the derived pointer. But it could be
613 // %g1 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
618 // %g2 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
622 // %p1 = phi i8 addrspace(1)* [ %g1, %bb1 ], [ %g2, %bb2 ]
623 // %cast = bitcast i8 addrspace(1)* %p1 in to i32 addrspace(1)*
625 // In this case, we can not find the bitcast any more. So we insert a new bitcast
626 // no matter there is already one or not. In this way, we can handle all cases, and
627 // the extra bitcast should be optimized away in later passes.
628 Instruction *ActualRelocatedBase = RelocatedBase;
629 if (RelocatedBase->getType() != Base->getType()) {
630 ActualRelocatedBase =
631 cast<Instruction>(Builder.CreateBitCast(RelocatedBase, Base->getType()));
633 Value *Replacement = Builder.CreateGEP(
634 Derived->getSourceElementType(), ActualRelocatedBase, makeArrayRef(OffsetV));
635 Instruction *ReplacementInst = cast<Instruction>(Replacement);
636 Replacement->takeName(ToReplace);
637 // If the newly generated derived pointer's type does not match the original derived
638 // pointer's type, cast the new derived pointer to match it. Same reasoning as above.
639 Instruction *ActualReplacement = ReplacementInst;
640 if (ReplacementInst->getType() != ToReplace->getType()) {
642 cast<Instruction>(Builder.CreateBitCast(ReplacementInst, ToReplace->getType()));
644 ToReplace->replaceAllUsesWith(ActualReplacement);
645 ToReplace->eraseFromParent();
655 // %ptr = gep %base + 15
656 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
657 // %base' = relocate(%tok, i32 4, i32 4)
658 // %ptr' = relocate(%tok, i32 4, i32 5)
664 // %ptr = gep %base + 15
665 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
666 // %base' = gc.relocate(%tok, i32 4, i32 4)
667 // %ptr' = gep %base' + 15
669 bool CodeGenPrepare::simplifyOffsetableRelocate(Instruction &I) {
670 bool MadeChange = false;
671 SmallVector<User *, 2> AllRelocateCalls;
673 for (auto *U : I.users())
674 if (isGCRelocate(dyn_cast<Instruction>(U)))
675 // Collect all the relocate calls associated with a statepoint
676 AllRelocateCalls.push_back(U);
678 // We need atleast one base pointer relocation + one derived pointer
679 // relocation to mangle
680 if (AllRelocateCalls.size() < 2)
683 // RelocateInstMap is a mapping from the base relocate instruction to the
684 // corresponding derived relocate instructions
685 DenseMap<IntrinsicInst *, SmallVector<IntrinsicInst *, 2>> RelocateInstMap;
686 computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap);
687 if (RelocateInstMap.empty())
690 for (auto &Item : RelocateInstMap)
691 // Item.first is the RelocatedBase to offset against
692 // Item.second is the vector of Targets to replace
693 MadeChange = simplifyRelocatesOffABase(Item.first, Item.second);
697 /// SinkCast - Sink the specified cast instruction into its user blocks
698 static bool SinkCast(CastInst *CI) {
699 BasicBlock *DefBB = CI->getParent();
701 /// InsertedCasts - Only insert a cast in each block once.
702 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
704 bool MadeChange = false;
705 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
707 Use &TheUse = UI.getUse();
708 Instruction *User = cast<Instruction>(*UI);
710 // Figure out which BB this cast is used in. For PHI's this is the
711 // appropriate predecessor block.
712 BasicBlock *UserBB = User->getParent();
713 if (PHINode *PN = dyn_cast<PHINode>(User)) {
714 UserBB = PN->getIncomingBlock(TheUse);
717 // Preincrement use iterator so we don't invalidate it.
720 // If this user is in the same block as the cast, don't change the cast.
721 if (UserBB == DefBB) continue;
723 // If we have already inserted a cast into this block, use it.
724 CastInst *&InsertedCast = InsertedCasts[UserBB];
727 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
728 assert(InsertPt != UserBB->end());
729 InsertedCast = CastInst::Create(CI->getOpcode(), CI->getOperand(0),
730 CI->getType(), "", &*InsertPt);
733 // Replace a use of the cast with a use of the new cast.
734 TheUse = InsertedCast;
739 // If we removed all uses, nuke the cast.
740 if (CI->use_empty()) {
741 CI->eraseFromParent();
748 /// If the specified cast instruction is a noop copy (e.g. it's casting from
749 /// one pointer type to another, i32->i8 on PPC), sink it into user blocks to
750 /// reduce the number of virtual registers that must be created and coalesced.
752 /// Return true if any changes are made.
754 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI,
755 const DataLayout &DL) {
756 // If this is a noop copy,
757 EVT SrcVT = TLI.getValueType(DL, CI->getOperand(0)->getType());
758 EVT DstVT = TLI.getValueType(DL, CI->getType());
760 // This is an fp<->int conversion?
761 if (SrcVT.isInteger() != DstVT.isInteger())
764 // If this is an extension, it will be a zero or sign extension, which
766 if (SrcVT.bitsLT(DstVT)) return false;
768 // If these values will be promoted, find out what they will be promoted
769 // to. This helps us consider truncates on PPC as noop copies when they
771 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
772 TargetLowering::TypePromoteInteger)
773 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
774 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
775 TargetLowering::TypePromoteInteger)
776 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
778 // If, after promotion, these are the same types, this is a noop copy.
785 /// Try to combine CI into a call to the llvm.uadd.with.overflow intrinsic if
788 /// Return true if any changes were made.
789 static bool CombineUAddWithOverflow(CmpInst *CI) {
793 m_UAddWithOverflow(m_Value(A), m_Value(B), m_Instruction(AddI))))
796 Type *Ty = AddI->getType();
797 if (!isa<IntegerType>(Ty))
800 // We don't want to move around uses of condition values this late, so we we
801 // check if it is legal to create the call to the intrinsic in the basic
802 // block containing the icmp:
804 if (AddI->getParent() != CI->getParent() && !AddI->hasOneUse())
808 // Someday m_UAddWithOverflow may get smarter, but this is a safe assumption
810 if (AddI->hasOneUse())
811 assert(*AddI->user_begin() == CI && "expected!");
814 Module *M = CI->getParent()->getParent()->getParent();
815 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
817 auto *InsertPt = AddI->hasOneUse() ? CI : AddI;
819 auto *UAddWithOverflow =
820 CallInst::Create(F, {A, B}, "uadd.overflow", InsertPt);
821 auto *UAdd = ExtractValueInst::Create(UAddWithOverflow, 0, "uadd", InsertPt);
823 ExtractValueInst::Create(UAddWithOverflow, 1, "overflow", InsertPt);
825 CI->replaceAllUsesWith(Overflow);
826 AddI->replaceAllUsesWith(UAdd);
827 CI->eraseFromParent();
828 AddI->eraseFromParent();
832 /// Sink the given CmpInst into user blocks to reduce the number of virtual
833 /// registers that must be created and coalesced. This is a clear win except on
834 /// targets with multiple condition code registers (PowerPC), where it might
835 /// lose; some adjustment may be wanted there.
837 /// Return true if any changes are made.
838 static bool SinkCmpExpression(CmpInst *CI) {
839 BasicBlock *DefBB = CI->getParent();
841 /// Only insert a cmp in each block once.
842 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
844 bool MadeChange = false;
845 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
847 Use &TheUse = UI.getUse();
848 Instruction *User = cast<Instruction>(*UI);
850 // Preincrement use iterator so we don't invalidate it.
853 // Don't bother for PHI nodes.
854 if (isa<PHINode>(User))
857 // Figure out which BB this cmp is used in.
858 BasicBlock *UserBB = User->getParent();
860 // If this user is in the same block as the cmp, don't change the cmp.
861 if (UserBB == DefBB) continue;
863 // If we have already inserted a cmp into this block, use it.
864 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
867 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
868 assert(InsertPt != UserBB->end());
870 CmpInst::Create(CI->getOpcode(), CI->getPredicate(),
871 CI->getOperand(0), CI->getOperand(1), "", &*InsertPt);
874 // Replace a use of the cmp with a use of the new cmp.
875 TheUse = InsertedCmp;
880 // If we removed all uses, nuke the cmp.
881 if (CI->use_empty()) {
882 CI->eraseFromParent();
889 static bool OptimizeCmpExpression(CmpInst *CI) {
890 if (SinkCmpExpression(CI))
893 if (CombineUAddWithOverflow(CI))
899 /// Check if the candidates could be combined with a shift instruction, which
901 /// 1. Truncate instruction
902 /// 2. And instruction and the imm is a mask of the low bits:
903 /// imm & (imm+1) == 0
904 static bool isExtractBitsCandidateUse(Instruction *User) {
905 if (!isa<TruncInst>(User)) {
906 if (User->getOpcode() != Instruction::And ||
907 !isa<ConstantInt>(User->getOperand(1)))
910 const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
912 if ((Cimm & (Cimm + 1)).getBoolValue())
918 /// Sink both shift and truncate instruction to the use of truncate's BB.
920 SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI,
921 DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts,
922 const TargetLowering &TLI, const DataLayout &DL) {
923 BasicBlock *UserBB = User->getParent();
924 DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
925 TruncInst *TruncI = dyn_cast<TruncInst>(User);
926 bool MadeChange = false;
928 for (Value::user_iterator TruncUI = TruncI->user_begin(),
929 TruncE = TruncI->user_end();
930 TruncUI != TruncE;) {
932 Use &TruncTheUse = TruncUI.getUse();
933 Instruction *TruncUser = cast<Instruction>(*TruncUI);
934 // Preincrement use iterator so we don't invalidate it.
938 int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
942 // If the use is actually a legal node, there will not be an
943 // implicit truncate.
944 // FIXME: always querying the result type is just an
945 // approximation; some nodes' legality is determined by the
946 // operand or other means. There's no good way to find out though.
947 if (TLI.isOperationLegalOrCustom(
948 ISDOpcode, TLI.getValueType(DL, TruncUser->getType(), true)))
951 // Don't bother for PHI nodes.
952 if (isa<PHINode>(TruncUser))
955 BasicBlock *TruncUserBB = TruncUser->getParent();
957 if (UserBB == TruncUserBB)
960 BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
961 CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
963 if (!InsertedShift && !InsertedTrunc) {
964 BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
965 assert(InsertPt != TruncUserBB->end());
967 if (ShiftI->getOpcode() == Instruction::AShr)
968 InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
971 InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
975 BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
977 assert(TruncInsertPt != TruncUserBB->end());
979 InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
980 TruncI->getType(), "", &*TruncInsertPt);
984 TruncTheUse = InsertedTrunc;
990 /// Sink the shift *right* instruction into user blocks if the uses could
991 /// potentially be combined with this shift instruction and generate BitExtract
992 /// instruction. It will only be applied if the architecture supports BitExtract
993 /// instruction. Here is an example:
995 /// %x.extract.shift = lshr i64 %arg1, 32
997 /// %x.extract.trunc = trunc i64 %x.extract.shift to i16
1001 /// %x.extract.shift.1 = lshr i64 %arg1, 32
1002 /// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
1004 /// CodeGen will recoginze the pattern in BB2 and generate BitExtract
1006 /// Return true if any changes are made.
1007 static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI,
1008 const TargetLowering &TLI,
1009 const DataLayout &DL) {
1010 BasicBlock *DefBB = ShiftI->getParent();
1012 /// Only insert instructions in each block once.
1013 DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts;
1015 bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(DL, ShiftI->getType()));
1017 bool MadeChange = false;
1018 for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
1020 Use &TheUse = UI.getUse();
1021 Instruction *User = cast<Instruction>(*UI);
1022 // Preincrement use iterator so we don't invalidate it.
1025 // Don't bother for PHI nodes.
1026 if (isa<PHINode>(User))
1029 if (!isExtractBitsCandidateUse(User))
1032 BasicBlock *UserBB = User->getParent();
1034 if (UserBB == DefBB) {
1035 // If the shift and truncate instruction are in the same BB. The use of
1036 // the truncate(TruncUse) may still introduce another truncate if not
1037 // legal. In this case, we would like to sink both shift and truncate
1038 // instruction to the BB of TruncUse.
1041 // i64 shift.result = lshr i64 opnd, imm
1042 // trunc.result = trunc shift.result to i16
1045 // ----> We will have an implicit truncate here if the architecture does
1046 // not have i16 compare.
1047 // cmp i16 trunc.result, opnd2
1049 if (isa<TruncInst>(User) && shiftIsLegal
1050 // If the type of the truncate is legal, no trucate will be
1051 // introduced in other basic blocks.
1053 (!TLI.isTypeLegal(TLI.getValueType(DL, User->getType()))))
1055 SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI, DL);
1059 // If we have already inserted a shift into this block, use it.
1060 BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
1062 if (!InsertedShift) {
1063 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1064 assert(InsertPt != UserBB->end());
1066 if (ShiftI->getOpcode() == Instruction::AShr)
1067 InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
1070 InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
1076 // Replace a use of the shift with a use of the new shift.
1077 TheUse = InsertedShift;
1080 // If we removed all uses, nuke the shift.
1081 if (ShiftI->use_empty())
1082 ShiftI->eraseFromParent();
1087 // Translate a masked load intrinsic like
1088 // <16 x i32 > @llvm.masked.load( <16 x i32>* %addr, i32 align,
1089 // <16 x i1> %mask, <16 x i32> %passthru)
1090 // to a chain of basic blocks, with loading element one-by-one if
1091 // the appropriate mask bit is set
1093 // %1 = bitcast i8* %addr to i32*
1094 // %2 = extractelement <16 x i1> %mask, i32 0
1095 // %3 = icmp eq i1 %2, true
1096 // br i1 %3, label %cond.load, label %else
1098 //cond.load: ; preds = %0
1099 // %4 = getelementptr i32* %1, i32 0
1100 // %5 = load i32* %4
1101 // %6 = insertelement <16 x i32> undef, i32 %5, i32 0
1104 //else: ; preds = %0, %cond.load
1105 // %res.phi.else = phi <16 x i32> [ %6, %cond.load ], [ undef, %0 ]
1106 // %7 = extractelement <16 x i1> %mask, i32 1
1107 // %8 = icmp eq i1 %7, true
1108 // br i1 %8, label %cond.load1, label %else2
1110 //cond.load1: ; preds = %else
1111 // %9 = getelementptr i32* %1, i32 1
1112 // %10 = load i32* %9
1113 // %11 = insertelement <16 x i32> %res.phi.else, i32 %10, i32 1
1116 //else2: ; preds = %else, %cond.load1
1117 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
1118 // %12 = extractelement <16 x i1> %mask, i32 2
1119 // %13 = icmp eq i1 %12, true
1120 // br i1 %13, label %cond.load4, label %else5
1122 static void ScalarizeMaskedLoad(CallInst *CI) {
1123 Value *Ptr = CI->getArgOperand(0);
1124 Value *Alignment = CI->getArgOperand(1);
1125 Value *Mask = CI->getArgOperand(2);
1126 Value *Src0 = CI->getArgOperand(3);
1128 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
1129 VectorType *VecType = dyn_cast<VectorType>(CI->getType());
1130 assert(VecType && "Unexpected return type of masked load intrinsic");
1132 Type *EltTy = CI->getType()->getVectorElementType();
1134 IRBuilder<> Builder(CI->getContext());
1135 Instruction *InsertPt = CI;
1136 BasicBlock *IfBlock = CI->getParent();
1137 BasicBlock *CondBlock = nullptr;
1138 BasicBlock *PrevIfBlock = CI->getParent();
1140 Builder.SetInsertPoint(InsertPt);
1141 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1143 // Short-cut if the mask is all-true.
1144 bool IsAllOnesMask = isa<Constant>(Mask) &&
1145 cast<Constant>(Mask)->isAllOnesValue();
1147 if (IsAllOnesMask) {
1148 Value *NewI = Builder.CreateAlignedLoad(Ptr, AlignVal);
1149 CI->replaceAllUsesWith(NewI);
1150 CI->eraseFromParent();
1154 // Adjust alignment for the scalar instruction.
1155 AlignVal = std::min(AlignVal, VecType->getScalarSizeInBits()/8);
1156 // Bitcast %addr fron i8* to EltTy*
1158 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
1159 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
1160 unsigned VectorWidth = VecType->getNumElements();
1162 Value *UndefVal = UndefValue::get(VecType);
1164 // The result vector
1165 Value *VResult = UndefVal;
1167 if (isa<ConstantVector>(Mask)) {
1168 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1169 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
1172 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1173 LoadInst* Load = Builder.CreateAlignedLoad(Gep, AlignVal);
1174 VResult = Builder.CreateInsertElement(VResult, Load,
1175 Builder.getInt32(Idx));
1177 Value *NewI = Builder.CreateSelect(Mask, VResult, Src0);
1178 CI->replaceAllUsesWith(NewI);
1179 CI->eraseFromParent();
1183 PHINode *Phi = nullptr;
1184 Value *PrevPhi = UndefVal;
1186 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1188 // Fill the "else" block, created in the previous iteration
1190 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
1191 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
1192 // %to_load = icmp eq i1 %mask_1, true
1193 // br i1 %to_load, label %cond.load, label %else
1196 Phi = Builder.CreatePHI(VecType, 2, "res.phi.else");
1197 Phi->addIncoming(VResult, CondBlock);
1198 Phi->addIncoming(PrevPhi, PrevIfBlock);
1203 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
1204 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1205 ConstantInt::get(Predicate->getType(), 1));
1207 // Create "cond" block
1209 // %EltAddr = getelementptr i32* %1, i32 0
1210 // %Elt = load i32* %EltAddr
1211 // VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx
1213 CondBlock = IfBlock->splitBasicBlock(InsertPt->getIterator(), "cond.load");
1214 Builder.SetInsertPoint(InsertPt);
1217 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1218 LoadInst* Load = Builder.CreateAlignedLoad(Gep, AlignVal);
1219 VResult = Builder.CreateInsertElement(VResult, Load, Builder.getInt32(Idx));
1221 // Create "else" block, fill it in the next iteration
1222 BasicBlock *NewIfBlock =
1223 CondBlock->splitBasicBlock(InsertPt->getIterator(), "else");
1224 Builder.SetInsertPoint(InsertPt);
1225 Instruction *OldBr = IfBlock->getTerminator();
1226 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1227 OldBr->eraseFromParent();
1228 PrevIfBlock = IfBlock;
1229 IfBlock = NewIfBlock;
1232 Phi = Builder.CreatePHI(VecType, 2, "res.phi.select");
1233 Phi->addIncoming(VResult, CondBlock);
1234 Phi->addIncoming(PrevPhi, PrevIfBlock);
1235 Value *NewI = Builder.CreateSelect(Mask, Phi, Src0);
1236 CI->replaceAllUsesWith(NewI);
1237 CI->eraseFromParent();
1240 // Translate a masked store intrinsic, like
1241 // void @llvm.masked.store(<16 x i32> %src, <16 x i32>* %addr, i32 align,
1243 // to a chain of basic blocks, that stores element one-by-one if
1244 // the appropriate mask bit is set
1246 // %1 = bitcast i8* %addr to i32*
1247 // %2 = extractelement <16 x i1> %mask, i32 0
1248 // %3 = icmp eq i1 %2, true
1249 // br i1 %3, label %cond.store, label %else
1251 // cond.store: ; preds = %0
1252 // %4 = extractelement <16 x i32> %val, i32 0
1253 // %5 = getelementptr i32* %1, i32 0
1254 // store i32 %4, i32* %5
1257 // else: ; preds = %0, %cond.store
1258 // %6 = extractelement <16 x i1> %mask, i32 1
1259 // %7 = icmp eq i1 %6, true
1260 // br i1 %7, label %cond.store1, label %else2
1262 // cond.store1: ; preds = %else
1263 // %8 = extractelement <16 x i32> %val, i32 1
1264 // %9 = getelementptr i32* %1, i32 1
1265 // store i32 %8, i32* %9
1268 static void ScalarizeMaskedStore(CallInst *CI) {
1269 Value *Src = CI->getArgOperand(0);
1270 Value *Ptr = CI->getArgOperand(1);
1271 Value *Alignment = CI->getArgOperand(2);
1272 Value *Mask = CI->getArgOperand(3);
1274 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
1275 VectorType *VecType = dyn_cast<VectorType>(Src->getType());
1276 assert(VecType && "Unexpected data type in masked store intrinsic");
1278 Type *EltTy = VecType->getElementType();
1280 IRBuilder<> Builder(CI->getContext());
1281 Instruction *InsertPt = CI;
1282 BasicBlock *IfBlock = CI->getParent();
1283 Builder.SetInsertPoint(InsertPt);
1284 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1286 // Short-cut if the mask is all-true.
1287 bool IsAllOnesMask = isa<Constant>(Mask) &&
1288 cast<Constant>(Mask)->isAllOnesValue();
1290 if (IsAllOnesMask) {
1291 Builder.CreateAlignedStore(Src, Ptr, AlignVal);
1292 CI->eraseFromParent();
1296 // Adjust alignment for the scalar instruction.
1297 AlignVal = std::max(AlignVal, VecType->getScalarSizeInBits()/8);
1298 // Bitcast %addr fron i8* to EltTy*
1300 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
1301 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
1302 unsigned VectorWidth = VecType->getNumElements();
1304 if (isa<ConstantVector>(Mask)) {
1305 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1306 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
1308 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx));
1310 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1311 Builder.CreateAlignedStore(OneElt, Gep, AlignVal);
1313 CI->eraseFromParent();
1317 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1319 // Fill the "else" block, created in the previous iteration
1321 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
1322 // %to_store = icmp eq i1 %mask_1, true
1323 // br i1 %to_store, label %cond.store, label %else
1325 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
1326 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1327 ConstantInt::get(Predicate->getType(), 1));
1329 // Create "cond" block
1331 // %OneElt = extractelement <16 x i32> %Src, i32 Idx
1332 // %EltAddr = getelementptr i32* %1, i32 0
1333 // %store i32 %OneElt, i32* %EltAddr
1335 BasicBlock *CondBlock =
1336 IfBlock->splitBasicBlock(InsertPt->getIterator(), "cond.store");
1337 Builder.SetInsertPoint(InsertPt);
1339 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx));
1341 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1342 Builder.CreateAlignedStore(OneElt, Gep, AlignVal);
1344 // Create "else" block, fill it in the next iteration
1345 BasicBlock *NewIfBlock =
1346 CondBlock->splitBasicBlock(InsertPt->getIterator(), "else");
1347 Builder.SetInsertPoint(InsertPt);
1348 Instruction *OldBr = IfBlock->getTerminator();
1349 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1350 OldBr->eraseFromParent();
1351 IfBlock = NewIfBlock;
1353 CI->eraseFromParent();
1356 bool CodeGenPrepare::optimizeCallInst(CallInst *CI, bool& ModifiedDT) {
1357 BasicBlock *BB = CI->getParent();
1359 // Lower inline assembly if we can.
1360 // If we found an inline asm expession, and if the target knows how to
1361 // lower it to normal LLVM code, do so now.
1362 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
1363 if (TLI->ExpandInlineAsm(CI)) {
1364 // Avoid invalidating the iterator.
1365 CurInstIterator = BB->begin();
1366 // Avoid processing instructions out of order, which could cause
1367 // reuse before a value is defined.
1371 // Sink address computing for memory operands into the block.
1372 if (optimizeInlineAsmInst(CI))
1376 // Align the pointer arguments to this call if the target thinks it's a good
1378 unsigned MinSize, PrefAlign;
1379 if (TLI && TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) {
1380 for (auto &Arg : CI->arg_operands()) {
1381 // We want to align both objects whose address is used directly and
1382 // objects whose address is used in casts and GEPs, though it only makes
1383 // sense for GEPs if the offset is a multiple of the desired alignment and
1384 // if size - offset meets the size threshold.
1385 if (!Arg->getType()->isPointerTy())
1387 APInt Offset(DL->getPointerSizeInBits(
1388 cast<PointerType>(Arg->getType())->getAddressSpace()),
1390 Value *Val = Arg->stripAndAccumulateInBoundsConstantOffsets(*DL, Offset);
1391 uint64_t Offset2 = Offset.getLimitedValue();
1392 if ((Offset2 & (PrefAlign-1)) != 0)
1395 if ((AI = dyn_cast<AllocaInst>(Val)) && AI->getAlignment() < PrefAlign &&
1396 DL->getTypeAllocSize(AI->getAllocatedType()) >= MinSize + Offset2)
1397 AI->setAlignment(PrefAlign);
1398 // Global variables can only be aligned if they are defined in this
1399 // object (i.e. they are uniquely initialized in this object), and
1400 // over-aligning global variables that have an explicit section is
1403 if ((GV = dyn_cast<GlobalVariable>(Val)) && GV->hasUniqueInitializer() &&
1404 !GV->hasSection() && GV->getAlignment() < PrefAlign &&
1405 DL->getTypeAllocSize(GV->getType()->getElementType()) >=
1407 GV->setAlignment(PrefAlign);
1409 // If this is a memcpy (or similar) then we may be able to improve the
1411 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(CI)) {
1412 unsigned Align = getKnownAlignment(MI->getDest(), *DL);
1413 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI))
1414 Align = std::min(Align, getKnownAlignment(MTI->getSource(), *DL));
1415 if (Align > MI->getAlignment())
1416 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), Align));
1420 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
1422 switch (II->getIntrinsicID()) {
1424 case Intrinsic::objectsize: {
1425 // Lower all uses of llvm.objectsize.*
1426 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
1427 Type *ReturnTy = CI->getType();
1428 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
1430 // Substituting this can cause recursive simplifications, which can
1431 // invalidate our iterator. Use a WeakVH to hold onto it in case this
1433 WeakVH IterHandle(&*CurInstIterator);
1435 replaceAndRecursivelySimplify(CI, RetVal,
1438 // If the iterator instruction was recursively deleted, start over at the
1439 // start of the block.
1440 if (IterHandle != CurInstIterator.getNodePtrUnchecked()) {
1441 CurInstIterator = BB->begin();
1446 case Intrinsic::masked_load: {
1447 // Scalarize unsupported vector masked load
1448 if (!TTI->isLegalMaskedLoad(CI->getType())) {
1449 ScalarizeMaskedLoad(CI);
1455 case Intrinsic::masked_store: {
1456 if (!TTI->isLegalMaskedStore(CI->getArgOperand(0)->getType())) {
1457 ScalarizeMaskedStore(CI);
1463 case Intrinsic::aarch64_stlxr:
1464 case Intrinsic::aarch64_stxr: {
1465 ZExtInst *ExtVal = dyn_cast<ZExtInst>(CI->getArgOperand(0));
1466 if (!ExtVal || !ExtVal->hasOneUse() ||
1467 ExtVal->getParent() == CI->getParent())
1469 // Sink a zext feeding stlxr/stxr before it, so it can be folded into it.
1470 ExtVal->moveBefore(CI);
1471 // Mark this instruction as "inserted by CGP", so that other
1472 // optimizations don't touch it.
1473 InsertedInsts.insert(ExtVal);
1476 case Intrinsic::invariant_group_barrier:
1477 II->replaceAllUsesWith(II->getArgOperand(0));
1478 II->eraseFromParent();
1483 // Unknown address space.
1484 // TODO: Target hook to pick which address space the intrinsic cares
1486 unsigned AddrSpace = ~0u;
1487 SmallVector<Value*, 2> PtrOps;
1489 if (TLI->GetAddrModeArguments(II, PtrOps, AccessTy, AddrSpace))
1490 while (!PtrOps.empty())
1491 if (optimizeMemoryInst(II, PtrOps.pop_back_val(), AccessTy, AddrSpace))
1496 // From here on out we're working with named functions.
1497 if (!CI->getCalledFunction()) return false;
1499 // Lower all default uses of _chk calls. This is very similar
1500 // to what InstCombineCalls does, but here we are only lowering calls
1501 // to fortified library functions (e.g. __memcpy_chk) that have the default
1502 // "don't know" as the objectsize. Anything else should be left alone.
1503 FortifiedLibCallSimplifier Simplifier(TLInfo, true);
1504 if (Value *V = Simplifier.optimizeCall(CI)) {
1505 CI->replaceAllUsesWith(V);
1506 CI->eraseFromParent();
1512 /// Look for opportunities to duplicate return instructions to the predecessor
1513 /// to enable tail call optimizations. The case it is currently looking for is:
1516 /// %tmp0 = tail call i32 @f0()
1517 /// br label %return
1519 /// %tmp1 = tail call i32 @f1()
1520 /// br label %return
1522 /// %tmp2 = tail call i32 @f2()
1523 /// br label %return
1525 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
1533 /// %tmp0 = tail call i32 @f0()
1536 /// %tmp1 = tail call i32 @f1()
1539 /// %tmp2 = tail call i32 @f2()
1542 bool CodeGenPrepare::dupRetToEnableTailCallOpts(BasicBlock *BB) {
1546 ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator());
1550 PHINode *PN = nullptr;
1551 BitCastInst *BCI = nullptr;
1552 Value *V = RI->getReturnValue();
1554 BCI = dyn_cast<BitCastInst>(V);
1556 V = BCI->getOperand(0);
1558 PN = dyn_cast<PHINode>(V);
1563 if (PN && PN->getParent() != BB)
1566 // It's not safe to eliminate the sign / zero extension of the return value.
1567 // See llvm::isInTailCallPosition().
1568 const Function *F = BB->getParent();
1569 AttributeSet CallerAttrs = F->getAttributes();
1570 if (CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt) ||
1571 CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
1574 // Make sure there are no instructions between the PHI and return, or that the
1575 // return is the first instruction in the block.
1577 BasicBlock::iterator BI = BB->begin();
1578 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
1580 // Also skip over the bitcast.
1585 BasicBlock::iterator BI = BB->begin();
1586 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
1591 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
1593 SmallVector<CallInst*, 4> TailCalls;
1595 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
1596 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
1597 // Make sure the phi value is indeed produced by the tail call.
1598 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
1599 TLI->mayBeEmittedAsTailCall(CI))
1600 TailCalls.push_back(CI);
1603 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
1604 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
1605 if (!VisitedBBs.insert(*PI).second)
1608 BasicBlock::InstListType &InstList = (*PI)->getInstList();
1609 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
1610 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
1611 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
1615 CallInst *CI = dyn_cast<CallInst>(&*RI);
1616 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
1617 TailCalls.push_back(CI);
1621 bool Changed = false;
1622 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
1623 CallInst *CI = TailCalls[i];
1626 // Conservatively require the attributes of the call to match those of the
1627 // return. Ignore noalias because it doesn't affect the call sequence.
1628 AttributeSet CalleeAttrs = CS.getAttributes();
1629 if (AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1630 removeAttribute(Attribute::NoAlias) !=
1631 AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1632 removeAttribute(Attribute::NoAlias))
1635 // Make sure the call instruction is followed by an unconditional branch to
1636 // the return block.
1637 BasicBlock *CallBB = CI->getParent();
1638 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
1639 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
1642 // Duplicate the return into CallBB.
1643 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
1644 ModifiedDT = Changed = true;
1648 // If we eliminated all predecessors of the block, delete the block now.
1649 if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
1650 BB->eraseFromParent();
1655 //===----------------------------------------------------------------------===//
1656 // Memory Optimization
1657 //===----------------------------------------------------------------------===//
1661 /// This is an extended version of TargetLowering::AddrMode
1662 /// which holds actual Value*'s for register values.
1663 struct ExtAddrMode : public TargetLowering::AddrMode {
1666 ExtAddrMode() : BaseReg(nullptr), ScaledReg(nullptr) {}
1667 void print(raw_ostream &OS) const;
1670 bool operator==(const ExtAddrMode& O) const {
1671 return (BaseReg == O.BaseReg) && (ScaledReg == O.ScaledReg) &&
1672 (BaseGV == O.BaseGV) && (BaseOffs == O.BaseOffs) &&
1673 (HasBaseReg == O.HasBaseReg) && (Scale == O.Scale);
1678 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
1684 void ExtAddrMode::print(raw_ostream &OS) const {
1685 bool NeedPlus = false;
1688 OS << (NeedPlus ? " + " : "")
1690 BaseGV->printAsOperand(OS, /*PrintType=*/false);
1695 OS << (NeedPlus ? " + " : "")
1701 OS << (NeedPlus ? " + " : "")
1703 BaseReg->printAsOperand(OS, /*PrintType=*/false);
1707 OS << (NeedPlus ? " + " : "")
1709 ScaledReg->printAsOperand(OS, /*PrintType=*/false);
1715 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1716 void ExtAddrMode::dump() const {
1722 /// \brief This class provides transaction based operation on the IR.
1723 /// Every change made through this class is recorded in the internal state and
1724 /// can be undone (rollback) until commit is called.
1725 class TypePromotionTransaction {
1727 /// \brief This represents the common interface of the individual transaction.
1728 /// Each class implements the logic for doing one specific modification on
1729 /// the IR via the TypePromotionTransaction.
1730 class TypePromotionAction {
1732 /// The Instruction modified.
1736 /// \brief Constructor of the action.
1737 /// The constructor performs the related action on the IR.
1738 TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
1740 virtual ~TypePromotionAction() {}
1742 /// \brief Undo the modification done by this action.
1743 /// When this method is called, the IR must be in the same state as it was
1744 /// before this action was applied.
1745 /// \pre Undoing the action works if and only if the IR is in the exact same
1746 /// state as it was directly after this action was applied.
1747 virtual void undo() = 0;
1749 /// \brief Advocate every change made by this action.
1750 /// When the results on the IR of the action are to be kept, it is important
1751 /// to call this function, otherwise hidden information may be kept forever.
1752 virtual void commit() {
1753 // Nothing to be done, this action is not doing anything.
1757 /// \brief Utility to remember the position of an instruction.
1758 class InsertionHandler {
1759 /// Position of an instruction.
1760 /// Either an instruction:
1761 /// - Is the first in a basic block: BB is used.
1762 /// - Has a previous instructon: PrevInst is used.
1764 Instruction *PrevInst;
1767 /// Remember whether or not the instruction had a previous instruction.
1768 bool HasPrevInstruction;
1771 /// \brief Record the position of \p Inst.
1772 InsertionHandler(Instruction *Inst) {
1773 BasicBlock::iterator It = Inst->getIterator();
1774 HasPrevInstruction = (It != (Inst->getParent()->begin()));
1775 if (HasPrevInstruction)
1776 Point.PrevInst = &*--It;
1778 Point.BB = Inst->getParent();
1781 /// \brief Insert \p Inst at the recorded position.
1782 void insert(Instruction *Inst) {
1783 if (HasPrevInstruction) {
1784 if (Inst->getParent())
1785 Inst->removeFromParent();
1786 Inst->insertAfter(Point.PrevInst);
1788 Instruction *Position = &*Point.BB->getFirstInsertionPt();
1789 if (Inst->getParent())
1790 Inst->moveBefore(Position);
1792 Inst->insertBefore(Position);
1797 /// \brief Move an instruction before another.
1798 class InstructionMoveBefore : public TypePromotionAction {
1799 /// Original position of the instruction.
1800 InsertionHandler Position;
1803 /// \brief Move \p Inst before \p Before.
1804 InstructionMoveBefore(Instruction *Inst, Instruction *Before)
1805 : TypePromotionAction(Inst), Position(Inst) {
1806 DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before << "\n");
1807 Inst->moveBefore(Before);
1810 /// \brief Move the instruction back to its original position.
1811 void undo() override {
1812 DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
1813 Position.insert(Inst);
1817 /// \brief Set the operand of an instruction with a new value.
1818 class OperandSetter : public TypePromotionAction {
1819 /// Original operand of the instruction.
1821 /// Index of the modified instruction.
1825 /// \brief Set \p Idx operand of \p Inst with \p NewVal.
1826 OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
1827 : TypePromotionAction(Inst), Idx(Idx) {
1828 DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
1829 << "for:" << *Inst << "\n"
1830 << "with:" << *NewVal << "\n");
1831 Origin = Inst->getOperand(Idx);
1832 Inst->setOperand(Idx, NewVal);
1835 /// \brief Restore the original value of the instruction.
1836 void undo() override {
1837 DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
1838 << "for: " << *Inst << "\n"
1839 << "with: " << *Origin << "\n");
1840 Inst->setOperand(Idx, Origin);
1844 /// \brief Hide the operands of an instruction.
1845 /// Do as if this instruction was not using any of its operands.
1846 class OperandsHider : public TypePromotionAction {
1847 /// The list of original operands.
1848 SmallVector<Value *, 4> OriginalValues;
1851 /// \brief Remove \p Inst from the uses of the operands of \p Inst.
1852 OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
1853 DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
1854 unsigned NumOpnds = Inst->getNumOperands();
1855 OriginalValues.reserve(NumOpnds);
1856 for (unsigned It = 0; It < NumOpnds; ++It) {
1857 // Save the current operand.
1858 Value *Val = Inst->getOperand(It);
1859 OriginalValues.push_back(Val);
1861 // We could use OperandSetter here, but that would imply an overhead
1862 // that we are not willing to pay.
1863 Inst->setOperand(It, UndefValue::get(Val->getType()));
1867 /// \brief Restore the original list of uses.
1868 void undo() override {
1869 DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
1870 for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
1871 Inst->setOperand(It, OriginalValues[It]);
1875 /// \brief Build a truncate instruction.
1876 class TruncBuilder : public TypePromotionAction {
1879 /// \brief Build a truncate instruction of \p Opnd producing a \p Ty
1881 /// trunc Opnd to Ty.
1882 TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
1883 IRBuilder<> Builder(Opnd);
1884 Val = Builder.CreateTrunc(Opnd, Ty, "promoted");
1885 DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n");
1888 /// \brief Get the built value.
1889 Value *getBuiltValue() { return Val; }
1891 /// \brief Remove the built instruction.
1892 void undo() override {
1893 DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n");
1894 if (Instruction *IVal = dyn_cast<Instruction>(Val))
1895 IVal->eraseFromParent();
1899 /// \brief Build a sign extension instruction.
1900 class SExtBuilder : public TypePromotionAction {
1903 /// \brief Build a sign extension instruction of \p Opnd producing a \p Ty
1905 /// sext Opnd to Ty.
1906 SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
1907 : TypePromotionAction(InsertPt) {
1908 IRBuilder<> Builder(InsertPt);
1909 Val = Builder.CreateSExt(Opnd, Ty, "promoted");
1910 DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n");
1913 /// \brief Get the built value.
1914 Value *getBuiltValue() { return Val; }
1916 /// \brief Remove the built instruction.
1917 void undo() override {
1918 DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n");
1919 if (Instruction *IVal = dyn_cast<Instruction>(Val))
1920 IVal->eraseFromParent();
1924 /// \brief Build a zero extension instruction.
1925 class ZExtBuilder : public TypePromotionAction {
1928 /// \brief Build a zero extension instruction of \p Opnd producing a \p Ty
1930 /// zext Opnd to Ty.
1931 ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
1932 : TypePromotionAction(InsertPt) {
1933 IRBuilder<> Builder(InsertPt);
1934 Val = Builder.CreateZExt(Opnd, Ty, "promoted");
1935 DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n");
1938 /// \brief Get the built value.
1939 Value *getBuiltValue() { return Val; }
1941 /// \brief Remove the built instruction.
1942 void undo() override {
1943 DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n");
1944 if (Instruction *IVal = dyn_cast<Instruction>(Val))
1945 IVal->eraseFromParent();
1949 /// \brief Mutate an instruction to another type.
1950 class TypeMutator : public TypePromotionAction {
1951 /// Record the original type.
1955 /// \brief Mutate the type of \p Inst into \p NewTy.
1956 TypeMutator(Instruction *Inst, Type *NewTy)
1957 : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
1958 DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
1960 Inst->mutateType(NewTy);
1963 /// \brief Mutate the instruction back to its original type.
1964 void undo() override {
1965 DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
1967 Inst->mutateType(OrigTy);
1971 /// \brief Replace the uses of an instruction by another instruction.
1972 class UsesReplacer : public TypePromotionAction {
1973 /// Helper structure to keep track of the replaced uses.
1974 struct InstructionAndIdx {
1975 /// The instruction using the instruction.
1977 /// The index where this instruction is used for Inst.
1979 InstructionAndIdx(Instruction *Inst, unsigned Idx)
1980 : Inst(Inst), Idx(Idx) {}
1983 /// Keep track of the original uses (pair Instruction, Index).
1984 SmallVector<InstructionAndIdx, 4> OriginalUses;
1985 typedef SmallVectorImpl<InstructionAndIdx>::iterator use_iterator;
1988 /// \brief Replace all the use of \p Inst by \p New.
1989 UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
1990 DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
1992 // Record the original uses.
1993 for (Use &U : Inst->uses()) {
1994 Instruction *UserI = cast<Instruction>(U.getUser());
1995 OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
1997 // Now, we can replace the uses.
1998 Inst->replaceAllUsesWith(New);
2001 /// \brief Reassign the original uses of Inst to Inst.
2002 void undo() override {
2003 DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
2004 for (use_iterator UseIt = OriginalUses.begin(),
2005 EndIt = OriginalUses.end();
2006 UseIt != EndIt; ++UseIt) {
2007 UseIt->Inst->setOperand(UseIt->Idx, Inst);
2012 /// \brief Remove an instruction from the IR.
2013 class InstructionRemover : public TypePromotionAction {
2014 /// Original position of the instruction.
2015 InsertionHandler Inserter;
2016 /// Helper structure to hide all the link to the instruction. In other
2017 /// words, this helps to do as if the instruction was removed.
2018 OperandsHider Hider;
2019 /// Keep track of the uses replaced, if any.
2020 UsesReplacer *Replacer;
2023 /// \brief Remove all reference of \p Inst and optinally replace all its
2025 /// \pre If !Inst->use_empty(), then New != nullptr
2026 InstructionRemover(Instruction *Inst, Value *New = nullptr)
2027 : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
2030 Replacer = new UsesReplacer(Inst, New);
2031 DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
2032 Inst->removeFromParent();
2035 ~InstructionRemover() override { delete Replacer; }
2037 /// \brief Really remove the instruction.
2038 void commit() override { delete Inst; }
2040 /// \brief Resurrect the instruction and reassign it to the proper uses if
2041 /// new value was provided when build this action.
2042 void undo() override {
2043 DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
2044 Inserter.insert(Inst);
2052 /// Restoration point.
2053 /// The restoration point is a pointer to an action instead of an iterator
2054 /// because the iterator may be invalidated but not the pointer.
2055 typedef const TypePromotionAction *ConstRestorationPt;
2056 /// Advocate every changes made in that transaction.
2058 /// Undo all the changes made after the given point.
2059 void rollback(ConstRestorationPt Point);
2060 /// Get the current restoration point.
2061 ConstRestorationPt getRestorationPoint() const;
2063 /// \name API for IR modification with state keeping to support rollback.
2065 /// Same as Instruction::setOperand.
2066 void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
2067 /// Same as Instruction::eraseFromParent.
2068 void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
2069 /// Same as Value::replaceAllUsesWith.
2070 void replaceAllUsesWith(Instruction *Inst, Value *New);
2071 /// Same as Value::mutateType.
2072 void mutateType(Instruction *Inst, Type *NewTy);
2073 /// Same as IRBuilder::createTrunc.
2074 Value *createTrunc(Instruction *Opnd, Type *Ty);
2075 /// Same as IRBuilder::createSExt.
2076 Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
2077 /// Same as IRBuilder::createZExt.
2078 Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
2079 /// Same as Instruction::moveBefore.
2080 void moveBefore(Instruction *Inst, Instruction *Before);
2084 /// The ordered list of actions made so far.
2085 SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions;
2086 typedef SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator CommitPt;
2089 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
2092 make_unique<TypePromotionTransaction::OperandSetter>(Inst, Idx, NewVal));
2095 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
2098 make_unique<TypePromotionTransaction::InstructionRemover>(Inst, NewVal));
2101 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
2103 Actions.push_back(make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
2106 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
2107 Actions.push_back(make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
2110 Value *TypePromotionTransaction::createTrunc(Instruction *Opnd,
2112 std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
2113 Value *Val = Ptr->getBuiltValue();
2114 Actions.push_back(std::move(Ptr));
2118 Value *TypePromotionTransaction::createSExt(Instruction *Inst,
2119 Value *Opnd, Type *Ty) {
2120 std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
2121 Value *Val = Ptr->getBuiltValue();
2122 Actions.push_back(std::move(Ptr));
2126 Value *TypePromotionTransaction::createZExt(Instruction *Inst,
2127 Value *Opnd, Type *Ty) {
2128 std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
2129 Value *Val = Ptr->getBuiltValue();
2130 Actions.push_back(std::move(Ptr));
2134 void TypePromotionTransaction::moveBefore(Instruction *Inst,
2135 Instruction *Before) {
2137 make_unique<TypePromotionTransaction::InstructionMoveBefore>(Inst, Before));
2140 TypePromotionTransaction::ConstRestorationPt
2141 TypePromotionTransaction::getRestorationPoint() const {
2142 return !Actions.empty() ? Actions.back().get() : nullptr;
2145 void TypePromotionTransaction::commit() {
2146 for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
2152 void TypePromotionTransaction::rollback(
2153 TypePromotionTransaction::ConstRestorationPt Point) {
2154 while (!Actions.empty() && Point != Actions.back().get()) {
2155 std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
2160 /// \brief A helper class for matching addressing modes.
2162 /// This encapsulates the logic for matching the target-legal addressing modes.
2163 class AddressingModeMatcher {
2164 SmallVectorImpl<Instruction*> &AddrModeInsts;
2165 const TargetMachine &TM;
2166 const TargetLowering &TLI;
2167 const DataLayout &DL;
2169 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
2170 /// the memory instruction that we're computing this address for.
2173 Instruction *MemoryInst;
2175 /// This is the addressing mode that we're building up. This is
2176 /// part of the return value of this addressing mode matching stuff.
2177 ExtAddrMode &AddrMode;
2179 /// The instructions inserted by other CodeGenPrepare optimizations.
2180 const SetOfInstrs &InsertedInsts;
2181 /// A map from the instructions to their type before promotion.
2182 InstrToOrigTy &PromotedInsts;
2183 /// The ongoing transaction where every action should be registered.
2184 TypePromotionTransaction &TPT;
2186 /// This is set to true when we should not do profitability checks.
2187 /// When true, IsProfitableToFoldIntoAddressingMode always returns true.
2188 bool IgnoreProfitability;
2190 AddressingModeMatcher(SmallVectorImpl<Instruction *> &AMI,
2191 const TargetMachine &TM, Type *AT, unsigned AS,
2192 Instruction *MI, ExtAddrMode &AM,
2193 const SetOfInstrs &InsertedInsts,
2194 InstrToOrigTy &PromotedInsts,
2195 TypePromotionTransaction &TPT)
2196 : AddrModeInsts(AMI), TM(TM),
2197 TLI(*TM.getSubtargetImpl(*MI->getParent()->getParent())
2198 ->getTargetLowering()),
2199 DL(MI->getModule()->getDataLayout()), AccessTy(AT), AddrSpace(AS),
2200 MemoryInst(MI), AddrMode(AM), InsertedInsts(InsertedInsts),
2201 PromotedInsts(PromotedInsts), TPT(TPT) {
2202 IgnoreProfitability = false;
2206 /// Find the maximal addressing mode that a load/store of V can fold,
2207 /// give an access type of AccessTy. This returns a list of involved
2208 /// instructions in AddrModeInsts.
2209 /// \p InsertedInsts The instructions inserted by other CodeGenPrepare
2211 /// \p PromotedInsts maps the instructions to their type before promotion.
2212 /// \p The ongoing transaction where every action should be registered.
2213 static ExtAddrMode Match(Value *V, Type *AccessTy, unsigned AS,
2214 Instruction *MemoryInst,
2215 SmallVectorImpl<Instruction*> &AddrModeInsts,
2216 const TargetMachine &TM,
2217 const SetOfInstrs &InsertedInsts,
2218 InstrToOrigTy &PromotedInsts,
2219 TypePromotionTransaction &TPT) {
2222 bool Success = AddressingModeMatcher(AddrModeInsts, TM, AccessTy, AS,
2223 MemoryInst, Result, InsertedInsts,
2224 PromotedInsts, TPT).matchAddr(V, 0);
2225 (void)Success; assert(Success && "Couldn't select *anything*?");
2229 bool matchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
2230 bool matchAddr(Value *V, unsigned Depth);
2231 bool matchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth,
2232 bool *MovedAway = nullptr);
2233 bool isProfitableToFoldIntoAddressingMode(Instruction *I,
2234 ExtAddrMode &AMBefore,
2235 ExtAddrMode &AMAfter);
2236 bool valueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
2237 bool isPromotionProfitable(unsigned NewCost, unsigned OldCost,
2238 Value *PromotedOperand) const;
2241 /// Try adding ScaleReg*Scale to the current addressing mode.
2242 /// Return true and update AddrMode if this addr mode is legal for the target,
2244 bool AddressingModeMatcher::matchScaledValue(Value *ScaleReg, int64_t Scale,
2246 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
2247 // mode. Just process that directly.
2249 return matchAddr(ScaleReg, Depth);
2251 // If the scale is 0, it takes nothing to add this.
2255 // If we already have a scale of this value, we can add to it, otherwise, we
2256 // need an available scale field.
2257 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
2260 ExtAddrMode TestAddrMode = AddrMode;
2262 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
2263 // [A+B + A*7] -> [B+A*8].
2264 TestAddrMode.Scale += Scale;
2265 TestAddrMode.ScaledReg = ScaleReg;
2267 // If the new address isn't legal, bail out.
2268 if (!TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace))
2271 // It was legal, so commit it.
2272 AddrMode = TestAddrMode;
2274 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
2275 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
2276 // X*Scale + C*Scale to addr mode.
2277 ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
2278 if (isa<Instruction>(ScaleReg) && // not a constant expr.
2279 match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
2280 TestAddrMode.ScaledReg = AddLHS;
2281 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
2283 // If this addressing mode is legal, commit it and remember that we folded
2284 // this instruction.
2285 if (TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace)) {
2286 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
2287 AddrMode = TestAddrMode;
2292 // Otherwise, not (x+c)*scale, just return what we have.
2296 /// This is a little filter, which returns true if an addressing computation
2297 /// involving I might be folded into a load/store accessing it.
2298 /// This doesn't need to be perfect, but needs to accept at least
2299 /// the set of instructions that MatchOperationAddr can.
2300 static bool MightBeFoldableInst(Instruction *I) {
2301 switch (I->getOpcode()) {
2302 case Instruction::BitCast:
2303 case Instruction::AddrSpaceCast:
2304 // Don't touch identity bitcasts.
2305 if (I->getType() == I->getOperand(0)->getType())
2307 return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
2308 case Instruction::PtrToInt:
2309 // PtrToInt is always a noop, as we know that the int type is pointer sized.
2311 case Instruction::IntToPtr:
2312 // We know the input is intptr_t, so this is foldable.
2314 case Instruction::Add:
2316 case Instruction::Mul:
2317 case Instruction::Shl:
2318 // Can only handle X*C and X << C.
2319 return isa<ConstantInt>(I->getOperand(1));
2320 case Instruction::GetElementPtr:
2327 /// \brief Check whether or not \p Val is a legal instruction for \p TLI.
2328 /// \note \p Val is assumed to be the product of some type promotion.
2329 /// Therefore if \p Val has an undefined state in \p TLI, this is assumed
2330 /// to be legal, as the non-promoted value would have had the same state.
2331 static bool isPromotedInstructionLegal(const TargetLowering &TLI,
2332 const DataLayout &DL, Value *Val) {
2333 Instruction *PromotedInst = dyn_cast<Instruction>(Val);
2336 int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
2337 // If the ISDOpcode is undefined, it was undefined before the promotion.
2340 // Otherwise, check if the promoted instruction is legal or not.
2341 return TLI.isOperationLegalOrCustom(
2342 ISDOpcode, TLI.getValueType(DL, PromotedInst->getType()));
2345 /// \brief Hepler class to perform type promotion.
2346 class TypePromotionHelper {
2347 /// \brief Utility function to check whether or not a sign or zero extension
2348 /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
2349 /// either using the operands of \p Inst or promoting \p Inst.
2350 /// The type of the extension is defined by \p IsSExt.
2351 /// In other words, check if:
2352 /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
2353 /// #1 Promotion applies:
2354 /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
2355 /// #2 Operand reuses:
2356 /// ext opnd1 to ConsideredExtType.
2357 /// \p PromotedInsts maps the instructions to their type before promotion.
2358 static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
2359 const InstrToOrigTy &PromotedInsts, bool IsSExt);
2361 /// \brief Utility function to determine if \p OpIdx should be promoted when
2362 /// promoting \p Inst.
2363 static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
2364 return !(isa<SelectInst>(Inst) && OpIdx == 0);
2367 /// \brief Utility function to promote the operand of \p Ext when this
2368 /// operand is a promotable trunc or sext or zext.
2369 /// \p PromotedInsts maps the instructions to their type before promotion.
2370 /// \p CreatedInstsCost[out] contains the cost of all instructions
2371 /// created to promote the operand of Ext.
2372 /// Newly added extensions are inserted in \p Exts.
2373 /// Newly added truncates are inserted in \p Truncs.
2374 /// Should never be called directly.
2375 /// \return The promoted value which is used instead of Ext.
2376 static Value *promoteOperandForTruncAndAnyExt(
2377 Instruction *Ext, TypePromotionTransaction &TPT,
2378 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2379 SmallVectorImpl<Instruction *> *Exts,
2380 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI);
2382 /// \brief Utility function to promote the operand of \p Ext when this
2383 /// operand is promotable and is not a supported trunc or sext.
2384 /// \p PromotedInsts maps the instructions to their type before promotion.
2385 /// \p CreatedInstsCost[out] contains the cost of all the instructions
2386 /// created to promote the operand of Ext.
2387 /// Newly added extensions are inserted in \p Exts.
2388 /// Newly added truncates are inserted in \p Truncs.
2389 /// Should never be called directly.
2390 /// \return The promoted value which is used instead of Ext.
2391 static Value *promoteOperandForOther(Instruction *Ext,
2392 TypePromotionTransaction &TPT,
2393 InstrToOrigTy &PromotedInsts,
2394 unsigned &CreatedInstsCost,
2395 SmallVectorImpl<Instruction *> *Exts,
2396 SmallVectorImpl<Instruction *> *Truncs,
2397 const TargetLowering &TLI, bool IsSExt);
2399 /// \see promoteOperandForOther.
2400 static Value *signExtendOperandForOther(
2401 Instruction *Ext, TypePromotionTransaction &TPT,
2402 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2403 SmallVectorImpl<Instruction *> *Exts,
2404 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2405 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
2406 Exts, Truncs, TLI, true);
2409 /// \see promoteOperandForOther.
2410 static Value *zeroExtendOperandForOther(
2411 Instruction *Ext, TypePromotionTransaction &TPT,
2412 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2413 SmallVectorImpl<Instruction *> *Exts,
2414 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2415 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
2416 Exts, Truncs, TLI, false);
2420 /// Type for the utility function that promotes the operand of Ext.
2421 typedef Value *(*Action)(Instruction *Ext, TypePromotionTransaction &TPT,
2422 InstrToOrigTy &PromotedInsts,
2423 unsigned &CreatedInstsCost,
2424 SmallVectorImpl<Instruction *> *Exts,
2425 SmallVectorImpl<Instruction *> *Truncs,
2426 const TargetLowering &TLI);
2427 /// \brief Given a sign/zero extend instruction \p Ext, return the approriate
2428 /// action to promote the operand of \p Ext instead of using Ext.
2429 /// \return NULL if no promotable action is possible with the current
2431 /// \p InsertedInsts keeps track of all the instructions inserted by the
2432 /// other CodeGenPrepare optimizations. This information is important
2433 /// because we do not want to promote these instructions as CodeGenPrepare
2434 /// will reinsert them later. Thus creating an infinite loop: create/remove.
2435 /// \p PromotedInsts maps the instructions to their type before promotion.
2436 static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedInsts,
2437 const TargetLowering &TLI,
2438 const InstrToOrigTy &PromotedInsts);
2441 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
2442 Type *ConsideredExtType,
2443 const InstrToOrigTy &PromotedInsts,
2445 // The promotion helper does not know how to deal with vector types yet.
2446 // To be able to fix that, we would need to fix the places where we
2447 // statically extend, e.g., constants and such.
2448 if (Inst->getType()->isVectorTy())
2451 // We can always get through zext.
2452 if (isa<ZExtInst>(Inst))
2455 // sext(sext) is ok too.
2456 if (IsSExt && isa<SExtInst>(Inst))
2459 // We can get through binary operator, if it is legal. In other words, the
2460 // binary operator must have a nuw or nsw flag.
2461 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
2462 if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
2463 ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
2464 (IsSExt && BinOp->hasNoSignedWrap())))
2467 // Check if we can do the following simplification.
2468 // ext(trunc(opnd)) --> ext(opnd)
2469 if (!isa<TruncInst>(Inst))
2472 Value *OpndVal = Inst->getOperand(0);
2473 // Check if we can use this operand in the extension.
2474 // If the type is larger than the result type of the extension, we cannot.
2475 if (!OpndVal->getType()->isIntegerTy() ||
2476 OpndVal->getType()->getIntegerBitWidth() >
2477 ConsideredExtType->getIntegerBitWidth())
2480 // If the operand of the truncate is not an instruction, we will not have
2481 // any information on the dropped bits.
2482 // (Actually we could for constant but it is not worth the extra logic).
2483 Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
2487 // Check if the source of the type is narrow enough.
2488 // I.e., check that trunc just drops extended bits of the same kind of
2490 // #1 get the type of the operand and check the kind of the extended bits.
2491 const Type *OpndType;
2492 InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
2493 if (It != PromotedInsts.end() && It->second.getInt() == IsSExt)
2494 OpndType = It->second.getPointer();
2495 else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd)))
2496 OpndType = Opnd->getOperand(0)->getType();
2500 // #2 check that the truncate just drops extended bits.
2501 return Inst->getType()->getIntegerBitWidth() >=
2502 OpndType->getIntegerBitWidth();
2505 TypePromotionHelper::Action TypePromotionHelper::getAction(
2506 Instruction *Ext, const SetOfInstrs &InsertedInsts,
2507 const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
2508 assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
2509 "Unexpected instruction type");
2510 Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0));
2511 Type *ExtTy = Ext->getType();
2512 bool IsSExt = isa<SExtInst>(Ext);
2513 // If the operand of the extension is not an instruction, we cannot
2515 // If it, check we can get through.
2516 if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt))
2519 // Do not promote if the operand has been added by codegenprepare.
2520 // Otherwise, it means we are undoing an optimization that is likely to be
2521 // redone, thus causing potential infinite loop.
2522 if (isa<TruncInst>(ExtOpnd) && InsertedInsts.count(ExtOpnd))
2525 // SExt or Trunc instructions.
2526 // Return the related handler.
2527 if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) ||
2528 isa<ZExtInst>(ExtOpnd))
2529 return promoteOperandForTruncAndAnyExt;
2531 // Regular instruction.
2532 // Abort early if we will have to insert non-free instructions.
2533 if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType()))
2535 return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
2538 Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
2539 llvm::Instruction *SExt, TypePromotionTransaction &TPT,
2540 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2541 SmallVectorImpl<Instruction *> *Exts,
2542 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2543 // By construction, the operand of SExt is an instruction. Otherwise we cannot
2544 // get through it and this method should not be called.
2545 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
2546 Value *ExtVal = SExt;
2547 bool HasMergedNonFreeExt = false;
2548 if (isa<ZExtInst>(SExtOpnd)) {
2549 // Replace s|zext(zext(opnd))
2551 HasMergedNonFreeExt = !TLI.isExtFree(SExtOpnd);
2553 TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
2554 TPT.replaceAllUsesWith(SExt, ZExt);
2555 TPT.eraseInstruction(SExt);
2558 // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
2560 TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
2562 CreatedInstsCost = 0;
2564 // Remove dead code.
2565 if (SExtOpnd->use_empty())
2566 TPT.eraseInstruction(SExtOpnd);
2568 // Check if the extension is still needed.
2569 Instruction *ExtInst = dyn_cast<Instruction>(ExtVal);
2570 if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) {
2573 Exts->push_back(ExtInst);
2574 CreatedInstsCost = !TLI.isExtFree(ExtInst) && !HasMergedNonFreeExt;
2579 // At this point we have: ext ty opnd to ty.
2580 // Reassign the uses of ExtInst to the opnd and remove ExtInst.
2581 Value *NextVal = ExtInst->getOperand(0);
2582 TPT.eraseInstruction(ExtInst, NextVal);
2586 Value *TypePromotionHelper::promoteOperandForOther(
2587 Instruction *Ext, TypePromotionTransaction &TPT,
2588 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2589 SmallVectorImpl<Instruction *> *Exts,
2590 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI,
2592 // By construction, the operand of Ext is an instruction. Otherwise we cannot
2593 // get through it and this method should not be called.
2594 Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0));
2595 CreatedInstsCost = 0;
2596 if (!ExtOpnd->hasOneUse()) {
2597 // ExtOpnd will be promoted.
2598 // All its uses, but Ext, will need to use a truncated value of the
2599 // promoted version.
2600 // Create the truncate now.
2601 Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType());
2602 if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) {
2603 ITrunc->removeFromParent();
2604 // Insert it just after the definition.
2605 ITrunc->insertAfter(ExtOpnd);
2607 Truncs->push_back(ITrunc);
2610 TPT.replaceAllUsesWith(ExtOpnd, Trunc);
2611 // Restore the operand of Ext (which has been replaced by the previous call
2612 // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
2613 TPT.setOperand(Ext, 0, ExtOpnd);
2616 // Get through the Instruction:
2617 // 1. Update its type.
2618 // 2. Replace the uses of Ext by Inst.
2619 // 3. Extend each operand that needs to be extended.
2621 // Remember the original type of the instruction before promotion.
2622 // This is useful to know that the high bits are sign extended bits.
2623 PromotedInsts.insert(std::pair<Instruction *, TypeIsSExt>(
2624 ExtOpnd, TypeIsSExt(ExtOpnd->getType(), IsSExt)));
2626 TPT.mutateType(ExtOpnd, Ext->getType());
2628 TPT.replaceAllUsesWith(Ext, ExtOpnd);
2630 Instruction *ExtForOpnd = Ext;
2632 DEBUG(dbgs() << "Propagate Ext to operands\n");
2633 for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
2635 DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n');
2636 if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() ||
2637 !shouldExtOperand(ExtOpnd, OpIdx)) {
2638 DEBUG(dbgs() << "No need to propagate\n");
2641 // Check if we can statically extend the operand.
2642 Value *Opnd = ExtOpnd->getOperand(OpIdx);
2643 if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
2644 DEBUG(dbgs() << "Statically extend\n");
2645 unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
2646 APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth)
2647 : Cst->getValue().zext(BitWidth);
2648 TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal));
2651 // UndefValue are typed, so we have to statically sign extend them.
2652 if (isa<UndefValue>(Opnd)) {
2653 DEBUG(dbgs() << "Statically extend\n");
2654 TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType()));
2658 // Otherwise we have to explicity sign extend the operand.
2659 // Check if Ext was reused to extend an operand.
2661 // If yes, create a new one.
2662 DEBUG(dbgs() << "More operands to ext\n");
2663 Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType())
2664 : TPT.createZExt(Ext, Opnd, Ext->getType());
2665 if (!isa<Instruction>(ValForExtOpnd)) {
2666 TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd);
2669 ExtForOpnd = cast<Instruction>(ValForExtOpnd);
2672 Exts->push_back(ExtForOpnd);
2673 TPT.setOperand(ExtForOpnd, 0, Opnd);
2675 // Move the sign extension before the insertion point.
2676 TPT.moveBefore(ExtForOpnd, ExtOpnd);
2677 TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd);
2678 CreatedInstsCost += !TLI.isExtFree(ExtForOpnd);
2679 // If more sext are required, new instructions will have to be created.
2680 ExtForOpnd = nullptr;
2682 if (ExtForOpnd == Ext) {
2683 DEBUG(dbgs() << "Extension is useless now\n");
2684 TPT.eraseInstruction(Ext);
2689 /// Check whether or not promoting an instruction to a wider type is profitable.
2690 /// \p NewCost gives the cost of extension instructions created by the
2692 /// \p OldCost gives the cost of extension instructions before the promotion
2693 /// plus the number of instructions that have been
2694 /// matched in the addressing mode the promotion.
2695 /// \p PromotedOperand is the value that has been promoted.
2696 /// \return True if the promotion is profitable, false otherwise.
2697 bool AddressingModeMatcher::isPromotionProfitable(
2698 unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const {
2699 DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCost << '\n');
2700 // The cost of the new extensions is greater than the cost of the
2701 // old extension plus what we folded.
2702 // This is not profitable.
2703 if (NewCost > OldCost)
2705 if (NewCost < OldCost)
2707 // The promotion is neutral but it may help folding the sign extension in
2708 // loads for instance.
2709 // Check that we did not create an illegal instruction.
2710 return isPromotedInstructionLegal(TLI, DL, PromotedOperand);
2713 /// Given an instruction or constant expr, see if we can fold the operation
2714 /// into the addressing mode. If so, update the addressing mode and return
2715 /// true, otherwise return false without modifying AddrMode.
2716 /// If \p MovedAway is not NULL, it contains the information of whether or
2717 /// not AddrInst has to be folded into the addressing mode on success.
2718 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
2719 /// because it has been moved away.
2720 /// Thus AddrInst must not be added in the matched instructions.
2721 /// This state can happen when AddrInst is a sext, since it may be moved away.
2722 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
2723 /// not be referenced anymore.
2724 bool AddressingModeMatcher::matchOperationAddr(User *AddrInst, unsigned Opcode,
2727 // Avoid exponential behavior on extremely deep expression trees.
2728 if (Depth >= 5) return false;
2730 // By default, all matched instructions stay in place.
2735 case Instruction::PtrToInt:
2736 // PtrToInt is always a noop, as we know that the int type is pointer sized.
2737 return matchAddr(AddrInst->getOperand(0), Depth);
2738 case Instruction::IntToPtr: {
2739 auto AS = AddrInst->getType()->getPointerAddressSpace();
2740 auto PtrTy = MVT::getIntegerVT(DL.getPointerSizeInBits(AS));
2741 // This inttoptr is a no-op if the integer type is pointer sized.
2742 if (TLI.getValueType(DL, AddrInst->getOperand(0)->getType()) == PtrTy)
2743 return matchAddr(AddrInst->getOperand(0), Depth);
2746 case Instruction::BitCast:
2747 // BitCast is always a noop, and we can handle it as long as it is
2748 // int->int or pointer->pointer (we don't want int<->fp or something).
2749 if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
2750 AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
2751 // Don't touch identity bitcasts. These were probably put here by LSR,
2752 // and we don't want to mess around with them. Assume it knows what it
2754 AddrInst->getOperand(0)->getType() != AddrInst->getType())
2755 return matchAddr(AddrInst->getOperand(0), Depth);
2757 case Instruction::AddrSpaceCast: {
2759 = AddrInst->getOperand(0)->getType()->getPointerAddressSpace();
2760 unsigned DestAS = AddrInst->getType()->getPointerAddressSpace();
2761 if (TLI.isNoopAddrSpaceCast(SrcAS, DestAS))
2762 return matchAddr(AddrInst->getOperand(0), Depth);
2765 case Instruction::Add: {
2766 // Check to see if we can merge in the RHS then the LHS. If so, we win.
2767 ExtAddrMode BackupAddrMode = AddrMode;
2768 unsigned OldSize = AddrModeInsts.size();
2769 // Start a transaction at this point.
2770 // The LHS may match but not the RHS.
2771 // Therefore, we need a higher level restoration point to undo partially
2772 // matched operation.
2773 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2774 TPT.getRestorationPoint();
2776 if (matchAddr(AddrInst->getOperand(1), Depth+1) &&
2777 matchAddr(AddrInst->getOperand(0), Depth+1))
2780 // Restore the old addr mode info.
2781 AddrMode = BackupAddrMode;
2782 AddrModeInsts.resize(OldSize);
2783 TPT.rollback(LastKnownGood);
2785 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
2786 if (matchAddr(AddrInst->getOperand(0), Depth+1) &&
2787 matchAddr(AddrInst->getOperand(1), Depth+1))
2790 // Otherwise we definitely can't merge the ADD in.
2791 AddrMode = BackupAddrMode;
2792 AddrModeInsts.resize(OldSize);
2793 TPT.rollback(LastKnownGood);
2796 //case Instruction::Or:
2797 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
2799 case Instruction::Mul:
2800 case Instruction::Shl: {
2801 // Can only handle X*C and X << C.
2802 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
2805 int64_t Scale = RHS->getSExtValue();
2806 if (Opcode == Instruction::Shl)
2807 Scale = 1LL << Scale;
2809 return matchScaledValue(AddrInst->getOperand(0), Scale, Depth);
2811 case Instruction::GetElementPtr: {
2812 // Scan the GEP. We check it if it contains constant offsets and at most
2813 // one variable offset.
2814 int VariableOperand = -1;
2815 unsigned VariableScale = 0;
2817 int64_t ConstantOffset = 0;
2818 gep_type_iterator GTI = gep_type_begin(AddrInst);
2819 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
2820 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
2821 const StructLayout *SL = DL.getStructLayout(STy);
2823 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
2824 ConstantOffset += SL->getElementOffset(Idx);
2826 uint64_t TypeSize = DL.getTypeAllocSize(GTI.getIndexedType());
2827 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
2828 ConstantOffset += CI->getSExtValue()*TypeSize;
2829 } else if (TypeSize) { // Scales of zero don't do anything.
2830 // We only allow one variable index at the moment.
2831 if (VariableOperand != -1)
2834 // Remember the variable index.
2835 VariableOperand = i;
2836 VariableScale = TypeSize;
2841 // A common case is for the GEP to only do a constant offset. In this case,
2842 // just add it to the disp field and check validity.
2843 if (VariableOperand == -1) {
2844 AddrMode.BaseOffs += ConstantOffset;
2845 if (ConstantOffset == 0 ||
2846 TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) {
2847 // Check to see if we can fold the base pointer in too.
2848 if (matchAddr(AddrInst->getOperand(0), Depth+1))
2851 AddrMode.BaseOffs -= ConstantOffset;
2855 // Save the valid addressing mode in case we can't match.
2856 ExtAddrMode BackupAddrMode = AddrMode;
2857 unsigned OldSize = AddrModeInsts.size();
2859 // See if the scale and offset amount is valid for this target.
2860 AddrMode.BaseOffs += ConstantOffset;
2862 // Match the base operand of the GEP.
2863 if (!matchAddr(AddrInst->getOperand(0), Depth+1)) {
2864 // If it couldn't be matched, just stuff the value in a register.
2865 if (AddrMode.HasBaseReg) {
2866 AddrMode = BackupAddrMode;
2867 AddrModeInsts.resize(OldSize);
2870 AddrMode.HasBaseReg = true;
2871 AddrMode.BaseReg = AddrInst->getOperand(0);
2874 // Match the remaining variable portion of the GEP.
2875 if (!matchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
2877 // If it couldn't be matched, try stuffing the base into a register
2878 // instead of matching it, and retrying the match of the scale.
2879 AddrMode = BackupAddrMode;
2880 AddrModeInsts.resize(OldSize);
2881 if (AddrMode.HasBaseReg)
2883 AddrMode.HasBaseReg = true;
2884 AddrMode.BaseReg = AddrInst->getOperand(0);
2885 AddrMode.BaseOffs += ConstantOffset;
2886 if (!matchScaledValue(AddrInst->getOperand(VariableOperand),
2887 VariableScale, Depth)) {
2888 // If even that didn't work, bail.
2889 AddrMode = BackupAddrMode;
2890 AddrModeInsts.resize(OldSize);
2897 case Instruction::SExt:
2898 case Instruction::ZExt: {
2899 Instruction *Ext = dyn_cast<Instruction>(AddrInst);
2903 // Try to move this ext out of the way of the addressing mode.
2904 // Ask for a method for doing so.
2905 TypePromotionHelper::Action TPH =
2906 TypePromotionHelper::getAction(Ext, InsertedInsts, TLI, PromotedInsts);
2910 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2911 TPT.getRestorationPoint();
2912 unsigned CreatedInstsCost = 0;
2913 unsigned ExtCost = !TLI.isExtFree(Ext);
2914 Value *PromotedOperand =
2915 TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI);
2916 // SExt has been moved away.
2917 // Thus either it will be rematched later in the recursive calls or it is
2918 // gone. Anyway, we must not fold it into the addressing mode at this point.
2922 // addr = gep base, idx
2924 // promotedOpnd = ext opnd <- no match here
2925 // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
2926 // addr = gep base, op <- match
2930 assert(PromotedOperand &&
2931 "TypePromotionHelper should have filtered out those cases");
2933 ExtAddrMode BackupAddrMode = AddrMode;
2934 unsigned OldSize = AddrModeInsts.size();
2936 if (!matchAddr(PromotedOperand, Depth) ||
2937 // The total of the new cost is equal to the cost of the created
2939 // The total of the old cost is equal to the cost of the extension plus
2940 // what we have saved in the addressing mode.
2941 !isPromotionProfitable(CreatedInstsCost,
2942 ExtCost + (AddrModeInsts.size() - OldSize),
2944 AddrMode = BackupAddrMode;
2945 AddrModeInsts.resize(OldSize);
2946 DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
2947 TPT.rollback(LastKnownGood);
2956 /// If we can, try to add the value of 'Addr' into the current addressing mode.
2957 /// If Addr can't be added to AddrMode this returns false and leaves AddrMode
2958 /// unmodified. This assumes that Addr is either a pointer type or intptr_t
2961 bool AddressingModeMatcher::matchAddr(Value *Addr, unsigned Depth) {
2962 // Start a transaction at this point that we will rollback if the matching
2964 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2965 TPT.getRestorationPoint();
2966 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
2967 // Fold in immediates if legal for the target.
2968 AddrMode.BaseOffs += CI->getSExtValue();
2969 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
2971 AddrMode.BaseOffs -= CI->getSExtValue();
2972 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
2973 // If this is a global variable, try to fold it into the addressing mode.
2974 if (!AddrMode.BaseGV) {
2975 AddrMode.BaseGV = GV;
2976 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
2978 AddrMode.BaseGV = nullptr;
2980 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
2981 ExtAddrMode BackupAddrMode = AddrMode;
2982 unsigned OldSize = AddrModeInsts.size();
2984 // Check to see if it is possible to fold this operation.
2985 bool MovedAway = false;
2986 if (matchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
2987 // This instruction may have been moved away. If so, there is nothing
2991 // Okay, it's possible to fold this. Check to see if it is actually
2992 // *profitable* to do so. We use a simple cost model to avoid increasing
2993 // register pressure too much.
2994 if (I->hasOneUse() ||
2995 isProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
2996 AddrModeInsts.push_back(I);
3000 // It isn't profitable to do this, roll back.
3001 //cerr << "NOT FOLDING: " << *I;
3002 AddrMode = BackupAddrMode;
3003 AddrModeInsts.resize(OldSize);
3004 TPT.rollback(LastKnownGood);
3006 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
3007 if (matchOperationAddr(CE, CE->getOpcode(), Depth))
3009 TPT.rollback(LastKnownGood);
3010 } else if (isa<ConstantPointerNull>(Addr)) {
3011 // Null pointer gets folded without affecting the addressing mode.
3015 // Worse case, the target should support [reg] addressing modes. :)
3016 if (!AddrMode.HasBaseReg) {
3017 AddrMode.HasBaseReg = true;
3018 AddrMode.BaseReg = Addr;
3019 // Still check for legality in case the target supports [imm] but not [i+r].
3020 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3022 AddrMode.HasBaseReg = false;
3023 AddrMode.BaseReg = nullptr;
3026 // If the base register is already taken, see if we can do [r+r].
3027 if (AddrMode.Scale == 0) {
3029 AddrMode.ScaledReg = Addr;
3030 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3033 AddrMode.ScaledReg = nullptr;
3036 TPT.rollback(LastKnownGood);
3040 /// Check to see if all uses of OpVal by the specified inline asm call are due
3041 /// to memory operands. If so, return true, otherwise return false.
3042 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
3043 const TargetMachine &TM) {
3044 const Function *F = CI->getParent()->getParent();
3045 const TargetLowering *TLI = TM.getSubtargetImpl(*F)->getTargetLowering();
3046 const TargetRegisterInfo *TRI = TM.getSubtargetImpl(*F)->getRegisterInfo();
3047 TargetLowering::AsmOperandInfoVector TargetConstraints =
3048 TLI->ParseConstraints(F->getParent()->getDataLayout(), TRI,
3049 ImmutableCallSite(CI));
3050 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
3051 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
3053 // Compute the constraint code and ConstraintType to use.
3054 TLI->ComputeConstraintToUse(OpInfo, SDValue());
3056 // If this asm operand is our Value*, and if it isn't an indirect memory
3057 // operand, we can't fold it!
3058 if (OpInfo.CallOperandVal == OpVal &&
3059 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
3060 !OpInfo.isIndirect))
3067 /// Recursively walk all the uses of I until we find a memory use.
3068 /// If we find an obviously non-foldable instruction, return true.
3069 /// Add the ultimately found memory instructions to MemoryUses.
3070 static bool FindAllMemoryUses(
3072 SmallVectorImpl<std::pair<Instruction *, unsigned>> &MemoryUses,
3073 SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetMachine &TM) {
3074 // If we already considered this instruction, we're done.
3075 if (!ConsideredInsts.insert(I).second)
3078 // If this is an obviously unfoldable instruction, bail out.
3079 if (!MightBeFoldableInst(I))
3082 // Loop over all the uses, recursively processing them.
3083 for (Use &U : I->uses()) {
3084 Instruction *UserI = cast<Instruction>(U.getUser());
3086 if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
3087 MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
3091 if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
3092 unsigned opNo = U.getOperandNo();
3093 if (opNo == 0) return true; // Storing addr, not into addr.
3094 MemoryUses.push_back(std::make_pair(SI, opNo));
3098 if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
3099 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
3100 if (!IA) return true;
3102 // If this is a memory operand, we're cool, otherwise bail out.
3103 if (!IsOperandAMemoryOperand(CI, IA, I, TM))
3108 if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TM))
3115 /// Return true if Val is already known to be live at the use site that we're
3116 /// folding it into. If so, there is no cost to include it in the addressing
3117 /// mode. KnownLive1 and KnownLive2 are two values that we know are live at the
3118 /// instruction already.
3119 bool AddressingModeMatcher::valueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
3120 Value *KnownLive2) {
3121 // If Val is either of the known-live values, we know it is live!
3122 if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
3125 // All values other than instructions and arguments (e.g. constants) are live.
3126 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
3128 // If Val is a constant sized alloca in the entry block, it is live, this is
3129 // true because it is just a reference to the stack/frame pointer, which is
3130 // live for the whole function.
3131 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
3132 if (AI->isStaticAlloca())
3135 // Check to see if this value is already used in the memory instruction's
3136 // block. If so, it's already live into the block at the very least, so we
3137 // can reasonably fold it.
3138 return Val->isUsedInBasicBlock(MemoryInst->getParent());
3141 /// It is possible for the addressing mode of the machine to fold the specified
3142 /// instruction into a load or store that ultimately uses it.
3143 /// However, the specified instruction has multiple uses.
3144 /// Given this, it may actually increase register pressure to fold it
3145 /// into the load. For example, consider this code:
3149 /// use(Y) -> nonload/store
3153 /// In this case, Y has multiple uses, and can be folded into the load of Z
3154 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
3155 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
3156 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
3157 /// number of computations either.
3159 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
3160 /// X was live across 'load Z' for other reasons, we actually *would* want to
3161 /// fold the addressing mode in the Z case. This would make Y die earlier.
3162 bool AddressingModeMatcher::
3163 isProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
3164 ExtAddrMode &AMAfter) {
3165 if (IgnoreProfitability) return true;
3167 // AMBefore is the addressing mode before this instruction was folded into it,
3168 // and AMAfter is the addressing mode after the instruction was folded. Get
3169 // the set of registers referenced by AMAfter and subtract out those
3170 // referenced by AMBefore: this is the set of values which folding in this
3171 // address extends the lifetime of.
3173 // Note that there are only two potential values being referenced here,
3174 // BaseReg and ScaleReg (global addresses are always available, as are any
3175 // folded immediates).
3176 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
3178 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
3179 // lifetime wasn't extended by adding this instruction.
3180 if (valueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
3182 if (valueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
3183 ScaledReg = nullptr;
3185 // If folding this instruction (and it's subexprs) didn't extend any live
3186 // ranges, we're ok with it.
3187 if (!BaseReg && !ScaledReg)
3190 // If all uses of this instruction are ultimately load/store/inlineasm's,
3191 // check to see if their addressing modes will include this instruction. If
3192 // so, we can fold it into all uses, so it doesn't matter if it has multiple
3194 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
3195 SmallPtrSet<Instruction*, 16> ConsideredInsts;
3196 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TM))
3197 return false; // Has a non-memory, non-foldable use!
3199 // Now that we know that all uses of this instruction are part of a chain of
3200 // computation involving only operations that could theoretically be folded
3201 // into a memory use, loop over each of these uses and see if they could
3202 // *actually* fold the instruction.
3203 SmallVector<Instruction*, 32> MatchedAddrModeInsts;
3204 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
3205 Instruction *User = MemoryUses[i].first;
3206 unsigned OpNo = MemoryUses[i].second;
3208 // Get the access type of this use. If the use isn't a pointer, we don't
3209 // know what it accesses.
3210 Value *Address = User->getOperand(OpNo);
3211 PointerType *AddrTy = dyn_cast<PointerType>(Address->getType());
3214 Type *AddressAccessTy = AddrTy->getElementType();
3215 unsigned AS = AddrTy->getAddressSpace();
3217 // Do a match against the root of this address, ignoring profitability. This
3218 // will tell us if the addressing mode for the memory operation will
3219 // *actually* cover the shared instruction.
3221 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3222 TPT.getRestorationPoint();
3223 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TM, AddressAccessTy, AS,
3224 MemoryInst, Result, InsertedInsts,
3225 PromotedInsts, TPT);
3226 Matcher.IgnoreProfitability = true;
3227 bool Success = Matcher.matchAddr(Address, 0);
3228 (void)Success; assert(Success && "Couldn't select *anything*?");
3230 // The match was to check the profitability, the changes made are not
3231 // part of the original matcher. Therefore, they should be dropped
3232 // otherwise the original matcher will not present the right state.
3233 TPT.rollback(LastKnownGood);
3235 // If the match didn't cover I, then it won't be shared by it.
3236 if (std::find(MatchedAddrModeInsts.begin(), MatchedAddrModeInsts.end(),
3237 I) == MatchedAddrModeInsts.end())
3240 MatchedAddrModeInsts.clear();
3246 } // end anonymous namespace
3248 /// Return true if the specified values are defined in a
3249 /// different basic block than BB.
3250 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
3251 if (Instruction *I = dyn_cast<Instruction>(V))
3252 return I->getParent() != BB;
3256 /// Load and Store Instructions often have addressing modes that can do
3257 /// significant amounts of computation. As such, instruction selection will try
3258 /// to get the load or store to do as much computation as possible for the
3259 /// program. The problem is that isel can only see within a single block. As
3260 /// such, we sink as much legal addressing mode work into the block as possible.
3262 /// This method is used to optimize both load/store and inline asms with memory
3264 bool CodeGenPrepare::optimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
3265 Type *AccessTy, unsigned AddrSpace) {
3268 // Try to collapse single-value PHI nodes. This is necessary to undo
3269 // unprofitable PRE transformations.
3270 SmallVector<Value*, 8> worklist;
3271 SmallPtrSet<Value*, 16> Visited;
3272 worklist.push_back(Addr);
3274 // Use a worklist to iteratively look through PHI nodes, and ensure that
3275 // the addressing mode obtained from the non-PHI roots of the graph
3277 Value *Consensus = nullptr;
3278 unsigned NumUsesConsensus = 0;
3279 bool IsNumUsesConsensusValid = false;
3280 SmallVector<Instruction*, 16> AddrModeInsts;
3281 ExtAddrMode AddrMode;
3282 TypePromotionTransaction TPT;
3283 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3284 TPT.getRestorationPoint();
3285 while (!worklist.empty()) {
3286 Value *V = worklist.back();
3287 worklist.pop_back();
3289 // Break use-def graph loops.
3290 if (!Visited.insert(V).second) {
3291 Consensus = nullptr;
3295 // For a PHI node, push all of its incoming values.
3296 if (PHINode *P = dyn_cast<PHINode>(V)) {
3297 for (Value *IncValue : P->incoming_values())
3298 worklist.push_back(IncValue);
3302 // For non-PHIs, determine the addressing mode being computed.
3303 SmallVector<Instruction*, 16> NewAddrModeInsts;
3304 ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
3305 V, AccessTy, AddrSpace, MemoryInst, NewAddrModeInsts, *TM,
3306 InsertedInsts, PromotedInsts, TPT);
3308 // This check is broken into two cases with very similar code to avoid using
3309 // getNumUses() as much as possible. Some values have a lot of uses, so
3310 // calling getNumUses() unconditionally caused a significant compile-time
3314 AddrMode = NewAddrMode;
3315 AddrModeInsts = NewAddrModeInsts;
3317 } else if (NewAddrMode == AddrMode) {
3318 if (!IsNumUsesConsensusValid) {
3319 NumUsesConsensus = Consensus->getNumUses();
3320 IsNumUsesConsensusValid = true;
3323 // Ensure that the obtained addressing mode is equivalent to that obtained
3324 // for all other roots of the PHI traversal. Also, when choosing one
3325 // such root as representative, select the one with the most uses in order
3326 // to keep the cost modeling heuristics in AddressingModeMatcher
3328 unsigned NumUses = V->getNumUses();
3329 if (NumUses > NumUsesConsensus) {
3331 NumUsesConsensus = NumUses;
3332 AddrModeInsts = NewAddrModeInsts;
3337 Consensus = nullptr;
3341 // If the addressing mode couldn't be determined, or if multiple different
3342 // ones were determined, bail out now.
3344 TPT.rollback(LastKnownGood);
3349 // Check to see if any of the instructions supersumed by this addr mode are
3350 // non-local to I's BB.
3351 bool AnyNonLocal = false;
3352 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
3353 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
3359 // If all the instructions matched are already in this BB, don't do anything.
3361 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
3365 // Insert this computation right after this user. Since our caller is
3366 // scanning from the top of the BB to the bottom, reuse of the expr are
3367 // guaranteed to happen later.
3368 IRBuilder<> Builder(MemoryInst);
3370 // Now that we determined the addressing expression we want to use and know
3371 // that we have to sink it into this block. Check to see if we have already
3372 // done this for some other load/store instr in this block. If so, reuse the
3374 Value *&SunkAddr = SunkAddrs[Addr];
3376 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
3377 << *MemoryInst << "\n");
3378 if (SunkAddr->getType() != Addr->getType())
3379 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
3380 } else if (AddrSinkUsingGEPs ||
3381 (!AddrSinkUsingGEPs.getNumOccurrences() && TM &&
3382 TM->getSubtargetImpl(*MemoryInst->getParent()->getParent())
3384 // By default, we use the GEP-based method when AA is used later. This
3385 // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
3386 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
3387 << *MemoryInst << "\n");
3388 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
3389 Value *ResultPtr = nullptr, *ResultIndex = nullptr;
3391 // First, find the pointer.
3392 if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
3393 ResultPtr = AddrMode.BaseReg;
3394 AddrMode.BaseReg = nullptr;
3397 if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
3398 // We can't add more than one pointer together, nor can we scale a
3399 // pointer (both of which seem meaningless).
3400 if (ResultPtr || AddrMode.Scale != 1)
3403 ResultPtr = AddrMode.ScaledReg;
3407 if (AddrMode.BaseGV) {
3411 ResultPtr = AddrMode.BaseGV;
3414 // If the real base value actually came from an inttoptr, then the matcher
3415 // will look through it and provide only the integer value. In that case,
3417 if (!ResultPtr && AddrMode.BaseReg) {
3419 Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(), "sunkaddr");
3420 AddrMode.BaseReg = nullptr;
3421 } else if (!ResultPtr && AddrMode.Scale == 1) {
3423 Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(), "sunkaddr");
3428 !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
3429 SunkAddr = Constant::getNullValue(Addr->getType());
3430 } else if (!ResultPtr) {
3434 Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
3435 Type *I8Ty = Builder.getInt8Ty();
3437 // Start with the base register. Do this first so that subsequent address
3438 // matching finds it last, which will prevent it from trying to match it
3439 // as the scaled value in case it happens to be a mul. That would be
3440 // problematic if we've sunk a different mul for the scale, because then
3441 // we'd end up sinking both muls.
3442 if (AddrMode.BaseReg) {
3443 Value *V = AddrMode.BaseReg;
3444 if (V->getType() != IntPtrTy)
3445 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
3450 // Add the scale value.
3451 if (AddrMode.Scale) {
3452 Value *V = AddrMode.ScaledReg;
3453 if (V->getType() == IntPtrTy) {
3455 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
3456 cast<IntegerType>(V->getType())->getBitWidth()) {
3457 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
3459 // It is only safe to sign extend the BaseReg if we know that the math
3460 // required to create it did not overflow before we extend it. Since
3461 // the original IR value was tossed in favor of a constant back when
3462 // the AddrMode was created we need to bail out gracefully if widths
3463 // do not match instead of extending it.
3464 Instruction *I = dyn_cast_or_null<Instruction>(ResultIndex);
3465 if (I && (ResultIndex != AddrMode.BaseReg))
3466 I->eraseFromParent();
3470 if (AddrMode.Scale != 1)
3471 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
3474 ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
3479 // Add in the Base Offset if present.
3480 if (AddrMode.BaseOffs) {
3481 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
3483 // We need to add this separately from the scale above to help with
3484 // SDAG consecutive load/store merging.
3485 if (ResultPtr->getType() != I8PtrTy)
3486 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
3487 ResultPtr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
3494 SunkAddr = ResultPtr;
3496 if (ResultPtr->getType() != I8PtrTy)
3497 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
3498 SunkAddr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
3501 if (SunkAddr->getType() != Addr->getType())
3502 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
3505 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
3506 << *MemoryInst << "\n");
3507 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
3508 Value *Result = nullptr;
3510 // Start with the base register. Do this first so that subsequent address
3511 // matching finds it last, which will prevent it from trying to match it
3512 // as the scaled value in case it happens to be a mul. That would be
3513 // problematic if we've sunk a different mul for the scale, because then
3514 // we'd end up sinking both muls.
3515 if (AddrMode.BaseReg) {
3516 Value *V = AddrMode.BaseReg;
3517 if (V->getType()->isPointerTy())
3518 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
3519 if (V->getType() != IntPtrTy)
3520 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
3524 // Add the scale value.
3525 if (AddrMode.Scale) {
3526 Value *V = AddrMode.ScaledReg;
3527 if (V->getType() == IntPtrTy) {
3529 } else if (V->getType()->isPointerTy()) {
3530 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
3531 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
3532 cast<IntegerType>(V->getType())->getBitWidth()) {
3533 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
3535 // It is only safe to sign extend the BaseReg if we know that the math
3536 // required to create it did not overflow before we extend it. Since
3537 // the original IR value was tossed in favor of a constant back when
3538 // the AddrMode was created we need to bail out gracefully if widths
3539 // do not match instead of extending it.
3540 Instruction *I = dyn_cast_or_null<Instruction>(Result);
3541 if (I && (Result != AddrMode.BaseReg))
3542 I->eraseFromParent();
3545 if (AddrMode.Scale != 1)
3546 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
3549 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3554 // Add in the BaseGV if present.
3555 if (AddrMode.BaseGV) {
3556 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
3558 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3563 // Add in the Base Offset if present.
3564 if (AddrMode.BaseOffs) {
3565 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
3567 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3573 SunkAddr = Constant::getNullValue(Addr->getType());
3575 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
3578 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
3580 // If we have no uses, recursively delete the value and all dead instructions
3582 if (Repl->use_empty()) {
3583 // This can cause recursive deletion, which can invalidate our iterator.
3584 // Use a WeakVH to hold onto it in case this happens.
3585 WeakVH IterHandle(&*CurInstIterator);
3586 BasicBlock *BB = CurInstIterator->getParent();
3588 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
3590 if (IterHandle != CurInstIterator.getNodePtrUnchecked()) {
3591 // If the iterator instruction was recursively deleted, start over at the
3592 // start of the block.
3593 CurInstIterator = BB->begin();
3601 /// If there are any memory operands, use OptimizeMemoryInst to sink their
3602 /// address computing into the block when possible / profitable.
3603 bool CodeGenPrepare::optimizeInlineAsmInst(CallInst *CS) {
3604 bool MadeChange = false;
3606 const TargetRegisterInfo *TRI =
3607 TM->getSubtargetImpl(*CS->getParent()->getParent())->getRegisterInfo();
3608 TargetLowering::AsmOperandInfoVector TargetConstraints =
3609 TLI->ParseConstraints(*DL, TRI, CS);
3611 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
3612 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
3614 // Compute the constraint code and ConstraintType to use.
3615 TLI->ComputeConstraintToUse(OpInfo, SDValue());
3617 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
3618 OpInfo.isIndirect) {
3619 Value *OpVal = CS->getArgOperand(ArgNo++);
3620 MadeChange |= optimizeMemoryInst(CS, OpVal, OpVal->getType(), ~0u);
3621 } else if (OpInfo.Type == InlineAsm::isInput)
3628 /// \brief Check if all the uses of \p Inst are equivalent (or free) zero or
3629 /// sign extensions.
3630 static bool hasSameExtUse(Instruction *Inst, const TargetLowering &TLI) {
3631 assert(!Inst->use_empty() && "Input must have at least one use");
3632 const Instruction *FirstUser = cast<Instruction>(*Inst->user_begin());
3633 bool IsSExt = isa<SExtInst>(FirstUser);
3634 Type *ExtTy = FirstUser->getType();
3635 for (const User *U : Inst->users()) {
3636 const Instruction *UI = cast<Instruction>(U);
3637 if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI)))
3639 Type *CurTy = UI->getType();
3640 // Same input and output types: Same instruction after CSE.
3644 // If IsSExt is true, we are in this situation:
3646 // b = sext ty1 a to ty2
3647 // c = sext ty1 a to ty3
3648 // Assuming ty2 is shorter than ty3, this could be turned into:
3650 // b = sext ty1 a to ty2
3651 // c = sext ty2 b to ty3
3652 // However, the last sext is not free.
3656 // This is a ZExt, maybe this is free to extend from one type to another.
3657 // In that case, we would not account for a different use.
3660 if (ExtTy->getScalarType()->getIntegerBitWidth() >
3661 CurTy->getScalarType()->getIntegerBitWidth()) {
3669 if (!TLI.isZExtFree(NarrowTy, LargeTy))
3672 // All uses are the same or can be derived from one another for free.
3676 /// \brief Try to form ExtLd by promoting \p Exts until they reach a
3677 /// load instruction.
3678 /// If an ext(load) can be formed, it is returned via \p LI for the load
3679 /// and \p Inst for the extension.
3680 /// Otherwise LI == nullptr and Inst == nullptr.
3681 /// When some promotion happened, \p TPT contains the proper state to
3684 /// \return true when promoting was necessary to expose the ext(load)
3685 /// opportunity, false otherwise.
3689 /// %ld = load i32* %addr
3690 /// %add = add nuw i32 %ld, 4
3691 /// %zext = zext i32 %add to i64
3695 /// %ld = load i32* %addr
3696 /// %zext = zext i32 %ld to i64
3697 /// %add = add nuw i64 %zext, 4
3699 /// Thanks to the promotion, we can match zext(load i32*) to i64.
3700 bool CodeGenPrepare::extLdPromotion(TypePromotionTransaction &TPT,
3701 LoadInst *&LI, Instruction *&Inst,
3702 const SmallVectorImpl<Instruction *> &Exts,
3703 unsigned CreatedInstsCost = 0) {
3704 // Iterate over all the extensions to see if one form an ext(load).
3705 for (auto I : Exts) {
3706 // Check if we directly have ext(load).
3707 if ((LI = dyn_cast<LoadInst>(I->getOperand(0)))) {
3709 // No promotion happened here.
3712 // Check whether or not we want to do any promotion.
3713 if (!TLI || !TLI->enableExtLdPromotion() || DisableExtLdPromotion)
3715 // Get the action to perform the promotion.
3716 TypePromotionHelper::Action TPH = TypePromotionHelper::getAction(
3717 I, InsertedInsts, *TLI, PromotedInsts);
3718 // Check if we can promote.
3721 // Save the current state.
3722 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3723 TPT.getRestorationPoint();
3724 SmallVector<Instruction *, 4> NewExts;
3725 unsigned NewCreatedInstsCost = 0;
3726 unsigned ExtCost = !TLI->isExtFree(I);
3728 Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost,
3729 &NewExts, nullptr, *TLI);
3730 assert(PromotedVal &&
3731 "TypePromotionHelper should have filtered out those cases");
3733 // We would be able to merge only one extension in a load.
3734 // Therefore, if we have more than 1 new extension we heuristically
3735 // cut this search path, because it means we degrade the code quality.
3736 // With exactly 2, the transformation is neutral, because we will merge
3737 // one extension but leave one. However, we optimistically keep going,
3738 // because the new extension may be removed too.
3739 long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost;
3740 TotalCreatedInstsCost -= ExtCost;
3741 if (!StressExtLdPromotion &&
3742 (TotalCreatedInstsCost > 1 ||
3743 !isPromotedInstructionLegal(*TLI, *DL, PromotedVal))) {
3744 // The promotion is not profitable, rollback to the previous state.
3745 TPT.rollback(LastKnownGood);
3748 // The promotion is profitable.
3749 // Check if it exposes an ext(load).
3750 (void)extLdPromotion(TPT, LI, Inst, NewExts, TotalCreatedInstsCost);
3751 if (LI && (StressExtLdPromotion || NewCreatedInstsCost <= ExtCost ||
3752 // If we have created a new extension, i.e., now we have two
3753 // extensions. We must make sure one of them is merged with
3754 // the load, otherwise we may degrade the code quality.
3755 (LI->hasOneUse() || hasSameExtUse(LI, *TLI))))
3756 // Promotion happened.
3758 // If this does not help to expose an ext(load) then, rollback.
3759 TPT.rollback(LastKnownGood);
3761 // None of the extension can form an ext(load).
3767 /// Move a zext or sext fed by a load into the same basic block as the load,
3768 /// unless conditions are unfavorable. This allows SelectionDAG to fold the
3769 /// extend into the load.
3770 /// \p I[in/out] the extension may be modified during the process if some
3771 /// promotions apply.
3773 bool CodeGenPrepare::moveExtToFormExtLoad(Instruction *&I) {
3774 // Try to promote a chain of computation if it allows to form
3775 // an extended load.
3776 TypePromotionTransaction TPT;
3777 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3778 TPT.getRestorationPoint();
3779 SmallVector<Instruction *, 1> Exts;
3781 // Look for a load being extended.
3782 LoadInst *LI = nullptr;
3783 Instruction *OldExt = I;
3784 bool HasPromoted = extLdPromotion(TPT, LI, I, Exts);
3786 assert(!HasPromoted && !LI && "If we did not match any load instruction "
3787 "the code must remain the same");
3792 // If they're already in the same block, there's nothing to do.
3793 // Make the cheap checks first if we did not promote.
3794 // If we promoted, we need to check if it is indeed profitable.
3795 if (!HasPromoted && LI->getParent() == I->getParent())
3798 EVT VT = TLI->getValueType(*DL, I->getType());
3799 EVT LoadVT = TLI->getValueType(*DL, LI->getType());
3801 // If the load has other users and the truncate is not free, this probably
3802 // isn't worthwhile.
3803 if (!LI->hasOneUse() && TLI &&
3804 (TLI->isTypeLegal(LoadVT) || !TLI->isTypeLegal(VT)) &&
3805 !TLI->isTruncateFree(I->getType(), LI->getType())) {
3807 TPT.rollback(LastKnownGood);
3811 // Check whether the target supports casts folded into loads.
3813 if (isa<ZExtInst>(I))
3814 LType = ISD::ZEXTLOAD;
3816 assert(isa<SExtInst>(I) && "Unexpected ext type!");
3817 LType = ISD::SEXTLOAD;
3819 if (TLI && !TLI->isLoadExtLegal(LType, VT, LoadVT)) {
3821 TPT.rollback(LastKnownGood);
3825 // Move the extend into the same block as the load, so that SelectionDAG
3828 I->removeFromParent();
3834 bool CodeGenPrepare::optimizeExtUses(Instruction *I) {
3835 BasicBlock *DefBB = I->getParent();
3837 // If the result of a {s|z}ext and its source are both live out, rewrite all
3838 // other uses of the source with result of extension.
3839 Value *Src = I->getOperand(0);
3840 if (Src->hasOneUse())
3843 // Only do this xform if truncating is free.
3844 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
3847 // Only safe to perform the optimization if the source is also defined in
3849 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
3852 bool DefIsLiveOut = false;
3853 for (User *U : I->users()) {
3854 Instruction *UI = cast<Instruction>(U);
3856 // Figure out which BB this ext is used in.
3857 BasicBlock *UserBB = UI->getParent();
3858 if (UserBB == DefBB) continue;
3859 DefIsLiveOut = true;
3865 // Make sure none of the uses are PHI nodes.
3866 for (User *U : Src->users()) {
3867 Instruction *UI = cast<Instruction>(U);
3868 BasicBlock *UserBB = UI->getParent();
3869 if (UserBB == DefBB) continue;
3870 // Be conservative. We don't want this xform to end up introducing
3871 // reloads just before load / store instructions.
3872 if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
3876 // InsertedTruncs - Only insert one trunc in each block once.
3877 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
3879 bool MadeChange = false;
3880 for (Use &U : Src->uses()) {
3881 Instruction *User = cast<Instruction>(U.getUser());
3883 // Figure out which BB this ext is used in.
3884 BasicBlock *UserBB = User->getParent();
3885 if (UserBB == DefBB) continue;
3887 // Both src and def are live in this block. Rewrite the use.
3888 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
3890 if (!InsertedTrunc) {
3891 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
3892 assert(InsertPt != UserBB->end());
3893 InsertedTrunc = new TruncInst(I, Src->getType(), "", &*InsertPt);
3894 InsertedInsts.insert(InsertedTrunc);
3897 // Replace a use of the {s|z}ext source with a use of the result.
3906 /// Check if V (an operand of a select instruction) is an expensive instruction
3907 /// that is only used once.
3908 static bool sinkSelectOperand(const TargetTransformInfo *TTI, Value *V) {
3909 auto *I = dyn_cast<Instruction>(V);
3910 // If it's safe to speculatively execute, then it should not have side
3911 // effects; therefore, it's safe to sink and possibly *not* execute.
3912 return I && I->hasOneUse() && isSafeToSpeculativelyExecute(I) &&
3913 TTI->getUserCost(I) >= TargetTransformInfo::TCC_Expensive;
3916 /// Returns true if a SelectInst should be turned into an explicit branch.
3917 static bool isFormingBranchFromSelectProfitable(const TargetTransformInfo *TTI,
3919 // FIXME: This should use the same heuristics as IfConversion to determine
3920 // whether a select is better represented as a branch. This requires that
3921 // branch probability metadata is preserved for the select, which is not the
3924 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
3926 // If a branch is predictable, an out-of-order CPU can avoid blocking on its
3927 // comparison condition. If the compare has more than one use, there's
3928 // probably another cmov or setcc around, so it's not worth emitting a branch.
3929 if (!Cmp || !Cmp->hasOneUse())
3932 Value *CmpOp0 = Cmp->getOperand(0);
3933 Value *CmpOp1 = Cmp->getOperand(1);
3935 // Emit "cmov on compare with a memory operand" as a branch to avoid stalls
3936 // on a load from memory. But if the load is used more than once, do not
3937 // change the select to a branch because the load is probably needed
3938 // regardless of whether the branch is taken or not.
3939 if ((isa<LoadInst>(CmpOp0) && CmpOp0->hasOneUse()) ||
3940 (isa<LoadInst>(CmpOp1) && CmpOp1->hasOneUse()))
3943 // If either operand of the select is expensive and only needed on one side
3944 // of the select, we should form a branch.
3945 if (sinkSelectOperand(TTI, SI->getTrueValue()) ||
3946 sinkSelectOperand(TTI, SI->getFalseValue()))
3953 /// If we have a SelectInst that will likely profit from branch prediction,
3954 /// turn it into a branch.
3955 bool CodeGenPrepare::optimizeSelectInst(SelectInst *SI) {
3956 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
3958 // Can we convert the 'select' to CF ?
3959 if (DisableSelectToBranch || OptSize || !TLI || VectorCond)
3962 TargetLowering::SelectSupportKind SelectKind;
3964 SelectKind = TargetLowering::VectorMaskSelect;
3965 else if (SI->getType()->isVectorTy())
3966 SelectKind = TargetLowering::ScalarCondVectorVal;
3968 SelectKind = TargetLowering::ScalarValSelect;
3970 // Do we have efficient codegen support for this kind of 'selects' ?
3971 if (TLI->isSelectSupported(SelectKind)) {
3972 // We have efficient codegen support for the select instruction.
3973 // Check if it is profitable to keep this 'select'.
3974 if (!TLI->isPredictableSelectExpensive() ||
3975 !isFormingBranchFromSelectProfitable(TTI, SI))
3981 // Transform a sequence like this:
3983 // %cmp = cmp uge i32 %a, %b
3984 // %sel = select i1 %cmp, i32 %c, i32 %d
3988 // %cmp = cmp uge i32 %a, %b
3989 // br i1 %cmp, label %select.true, label %select.false
3991 // br label %select.end
3993 // br label %select.end
3995 // %sel = phi i32 [ %c, %select.true ], [ %d, %select.false ]
3997 // In addition, we may sink instructions that produce %c or %d from
3998 // the entry block into the destination(s) of the new branch.
3999 // If the true or false blocks do not contain a sunken instruction, that
4000 // block and its branch may be optimized away. In that case, one side of the
4001 // first branch will point directly to select.end, and the corresponding PHI
4002 // predecessor block will be the start block.
4004 // First, we split the block containing the select into 2 blocks.
4005 BasicBlock *StartBlock = SI->getParent();
4006 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(SI));
4007 BasicBlock *EndBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
4009 // Delete the unconditional branch that was just created by the split.
4010 StartBlock->getTerminator()->eraseFromParent();
4012 // These are the new basic blocks for the conditional branch.
4013 // At least one will become an actual new basic block.
4014 BasicBlock *TrueBlock = nullptr;
4015 BasicBlock *FalseBlock = nullptr;
4017 // Sink expensive instructions into the conditional blocks to avoid executing
4018 // them speculatively.
4019 if (sinkSelectOperand(TTI, SI->getTrueValue())) {
4020 TrueBlock = BasicBlock::Create(SI->getContext(), "select.true.sink",
4021 EndBlock->getParent(), EndBlock);
4022 auto *TrueBranch = BranchInst::Create(EndBlock, TrueBlock);
4023 auto *TrueInst = cast<Instruction>(SI->getTrueValue());
4024 TrueInst->moveBefore(TrueBranch);
4026 if (sinkSelectOperand(TTI, SI->getFalseValue())) {
4027 FalseBlock = BasicBlock::Create(SI->getContext(), "select.false.sink",
4028 EndBlock->getParent(), EndBlock);
4029 auto *FalseBranch = BranchInst::Create(EndBlock, FalseBlock);
4030 auto *FalseInst = cast<Instruction>(SI->getFalseValue());
4031 FalseInst->moveBefore(FalseBranch);
4034 // If there was nothing to sink, then arbitrarily choose the 'false' side
4035 // for a new input value to the PHI.
4036 if (TrueBlock == FalseBlock) {
4037 assert(TrueBlock == nullptr &&
4038 "Unexpected basic block transform while optimizing select");
4040 FalseBlock = BasicBlock::Create(SI->getContext(), "select.false",
4041 EndBlock->getParent(), EndBlock);
4042 BranchInst::Create(EndBlock, FalseBlock);
4045 // Insert the real conditional branch based on the original condition.
4046 // If we did not create a new block for one of the 'true' or 'false' paths
4047 // of the condition, it means that side of the branch goes to the end block
4048 // directly and the path originates from the start block from the point of
4049 // view of the new PHI.
4050 if (TrueBlock == nullptr) {
4051 BranchInst::Create(EndBlock, FalseBlock, SI->getCondition(), SI);
4052 TrueBlock = StartBlock;
4053 } else if (FalseBlock == nullptr) {
4054 BranchInst::Create(TrueBlock, EndBlock, SI->getCondition(), SI);
4055 FalseBlock = StartBlock;
4057 BranchInst::Create(TrueBlock, FalseBlock, SI->getCondition(), SI);
4060 // The select itself is replaced with a PHI Node.
4061 PHINode *PN = PHINode::Create(SI->getType(), 2, "", &EndBlock->front());
4063 PN->addIncoming(SI->getTrueValue(), TrueBlock);
4064 PN->addIncoming(SI->getFalseValue(), FalseBlock);
4066 SI->replaceAllUsesWith(PN);
4067 SI->eraseFromParent();
4069 // Instruct OptimizeBlock to skip to the next block.
4070 CurInstIterator = StartBlock->end();
4071 ++NumSelectsExpanded;
4075 static bool isBroadcastShuffle(ShuffleVectorInst *SVI) {
4076 SmallVector<int, 16> Mask(SVI->getShuffleMask());
4078 for (unsigned i = 0; i < Mask.size(); ++i) {
4079 if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
4081 SplatElem = Mask[i];
4087 /// Some targets have expensive vector shifts if the lanes aren't all the same
4088 /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
4089 /// it's often worth sinking a shufflevector splat down to its use so that
4090 /// codegen can spot all lanes are identical.
4091 bool CodeGenPrepare::optimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
4092 BasicBlock *DefBB = SVI->getParent();
4094 // Only do this xform if variable vector shifts are particularly expensive.
4095 if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
4098 // We only expect better codegen by sinking a shuffle if we can recognise a
4100 if (!isBroadcastShuffle(SVI))
4103 // InsertedShuffles - Only insert a shuffle in each block once.
4104 DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
4106 bool MadeChange = false;
4107 for (User *U : SVI->users()) {
4108 Instruction *UI = cast<Instruction>(U);
4110 // Figure out which BB this ext is used in.
4111 BasicBlock *UserBB = UI->getParent();
4112 if (UserBB == DefBB) continue;
4114 // For now only apply this when the splat is used by a shift instruction.
4115 if (!UI->isShift()) continue;
4117 // Everything checks out, sink the shuffle if the user's block doesn't
4118 // already have a copy.
4119 Instruction *&InsertedShuffle = InsertedShuffles[UserBB];
4121 if (!InsertedShuffle) {
4122 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
4123 assert(InsertPt != UserBB->end());
4125 new ShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1),
4126 SVI->getOperand(2), "", &*InsertPt);
4129 UI->replaceUsesOfWith(SVI, InsertedShuffle);
4133 // If we removed all uses, nuke the shuffle.
4134 if (SVI->use_empty()) {
4135 SVI->eraseFromParent();
4143 /// \brief Helper class to promote a scalar operation to a vector one.
4144 /// This class is used to move downward extractelement transition.
4146 /// a = vector_op <2 x i32>
4147 /// b = extractelement <2 x i32> a, i32 0
4152 /// a = vector_op <2 x i32>
4153 /// c = vector_op a (equivalent to scalar_op on the related lane)
4154 /// * d = extractelement <2 x i32> c, i32 0
4156 /// Assuming both extractelement and store can be combine, we get rid of the
4158 class VectorPromoteHelper {
4159 /// DataLayout associated with the current module.
4160 const DataLayout &DL;
4162 /// Used to perform some checks on the legality of vector operations.
4163 const TargetLowering &TLI;
4165 /// Used to estimated the cost of the promoted chain.
4166 const TargetTransformInfo &TTI;
4168 /// The transition being moved downwards.
4169 Instruction *Transition;
4170 /// The sequence of instructions to be promoted.
4171 SmallVector<Instruction *, 4> InstsToBePromoted;
4172 /// Cost of combining a store and an extract.
4173 unsigned StoreExtractCombineCost;
4174 /// Instruction that will be combined with the transition.
4175 Instruction *CombineInst;
4177 /// \brief The instruction that represents the current end of the transition.
4178 /// Since we are faking the promotion until we reach the end of the chain
4179 /// of computation, we need a way to get the current end of the transition.
4180 Instruction *getEndOfTransition() const {
4181 if (InstsToBePromoted.empty())
4183 return InstsToBePromoted.back();
4186 /// \brief Return the index of the original value in the transition.
4187 /// E.g., for "extractelement <2 x i32> c, i32 1" the original value,
4188 /// c, is at index 0.
4189 unsigned getTransitionOriginalValueIdx() const {
4190 assert(isa<ExtractElementInst>(Transition) &&
4191 "Other kind of transitions are not supported yet");
4195 /// \brief Return the index of the index in the transition.
4196 /// E.g., for "extractelement <2 x i32> c, i32 0" the index
4198 unsigned getTransitionIdx() const {
4199 assert(isa<ExtractElementInst>(Transition) &&
4200 "Other kind of transitions are not supported yet");
4204 /// \brief Get the type of the transition.
4205 /// This is the type of the original value.
4206 /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the
4207 /// transition is <2 x i32>.
4208 Type *getTransitionType() const {
4209 return Transition->getOperand(getTransitionOriginalValueIdx())->getType();
4212 /// \brief Promote \p ToBePromoted by moving \p Def downward through.
4213 /// I.e., we have the following sequence:
4214 /// Def = Transition <ty1> a to <ty2>
4215 /// b = ToBePromoted <ty2> Def, ...
4217 /// b = ToBePromoted <ty1> a, ...
4218 /// Def = Transition <ty1> ToBePromoted to <ty2>
4219 void promoteImpl(Instruction *ToBePromoted);
4221 /// \brief Check whether or not it is profitable to promote all the
4222 /// instructions enqueued to be promoted.
4223 bool isProfitableToPromote() {
4224 Value *ValIdx = Transition->getOperand(getTransitionOriginalValueIdx());
4225 unsigned Index = isa<ConstantInt>(ValIdx)
4226 ? cast<ConstantInt>(ValIdx)->getZExtValue()
4228 Type *PromotedType = getTransitionType();
4230 StoreInst *ST = cast<StoreInst>(CombineInst);
4231 unsigned AS = ST->getPointerAddressSpace();
4232 unsigned Align = ST->getAlignment();
4233 // Check if this store is supported.
4234 if (!TLI.allowsMisalignedMemoryAccesses(
4235 TLI.getValueType(DL, ST->getValueOperand()->getType()), AS,
4237 // If this is not supported, there is no way we can combine
4238 // the extract with the store.
4242 // The scalar chain of computation has to pay for the transition
4243 // scalar to vector.
4244 // The vector chain has to account for the combining cost.
4245 uint64_t ScalarCost =
4246 TTI.getVectorInstrCost(Transition->getOpcode(), PromotedType, Index);
4247 uint64_t VectorCost = StoreExtractCombineCost;
4248 for (const auto &Inst : InstsToBePromoted) {
4249 // Compute the cost.
4250 // By construction, all instructions being promoted are arithmetic ones.
4251 // Moreover, one argument is a constant that can be viewed as a splat
4253 Value *Arg0 = Inst->getOperand(0);
4254 bool IsArg0Constant = isa<UndefValue>(Arg0) || isa<ConstantInt>(Arg0) ||
4255 isa<ConstantFP>(Arg0);
4256 TargetTransformInfo::OperandValueKind Arg0OVK =
4257 IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
4258 : TargetTransformInfo::OK_AnyValue;
4259 TargetTransformInfo::OperandValueKind Arg1OVK =
4260 !IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
4261 : TargetTransformInfo::OK_AnyValue;
4262 ScalarCost += TTI.getArithmeticInstrCost(
4263 Inst->getOpcode(), Inst->getType(), Arg0OVK, Arg1OVK);
4264 VectorCost += TTI.getArithmeticInstrCost(Inst->getOpcode(), PromotedType,
4267 DEBUG(dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
4268 << ScalarCost << "\nVector: " << VectorCost << '\n');
4269 return ScalarCost > VectorCost;
4272 /// \brief Generate a constant vector with \p Val with the same
4273 /// number of elements as the transition.
4274 /// \p UseSplat defines whether or not \p Val should be replicated
4275 /// across the whole vector.
4276 /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>,
4277 /// otherwise we generate a vector with as many undef as possible:
4278 /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only
4279 /// used at the index of the extract.
4280 Value *getConstantVector(Constant *Val, bool UseSplat) const {
4281 unsigned ExtractIdx = UINT_MAX;
4283 // If we cannot determine where the constant must be, we have to
4284 // use a splat constant.
4285 Value *ValExtractIdx = Transition->getOperand(getTransitionIdx());
4286 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(ValExtractIdx))
4287 ExtractIdx = CstVal->getSExtValue();
4292 unsigned End = getTransitionType()->getVectorNumElements();
4294 return ConstantVector::getSplat(End, Val);
4296 SmallVector<Constant *, 4> ConstVec;
4297 UndefValue *UndefVal = UndefValue::get(Val->getType());
4298 for (unsigned Idx = 0; Idx != End; ++Idx) {
4299 if (Idx == ExtractIdx)
4300 ConstVec.push_back(Val);
4302 ConstVec.push_back(UndefVal);
4304 return ConstantVector::get(ConstVec);
4307 /// \brief Check if promoting to a vector type an operand at \p OperandIdx
4308 /// in \p Use can trigger undefined behavior.
4309 static bool canCauseUndefinedBehavior(const Instruction *Use,
4310 unsigned OperandIdx) {
4311 // This is not safe to introduce undef when the operand is on
4312 // the right hand side of a division-like instruction.
4313 if (OperandIdx != 1)
4315 switch (Use->getOpcode()) {
4318 case Instruction::SDiv:
4319 case Instruction::UDiv:
4320 case Instruction::SRem:
4321 case Instruction::URem:
4323 case Instruction::FDiv:
4324 case Instruction::FRem:
4325 return !Use->hasNoNaNs();
4327 llvm_unreachable(nullptr);
4331 VectorPromoteHelper(const DataLayout &DL, const TargetLowering &TLI,
4332 const TargetTransformInfo &TTI, Instruction *Transition,
4333 unsigned CombineCost)
4334 : DL(DL), TLI(TLI), TTI(TTI), Transition(Transition),
4335 StoreExtractCombineCost(CombineCost), CombineInst(nullptr) {
4336 assert(Transition && "Do not know how to promote null");
4339 /// \brief Check if we can promote \p ToBePromoted to \p Type.
4340 bool canPromote(const Instruction *ToBePromoted) const {
4341 // We could support CastInst too.
4342 return isa<BinaryOperator>(ToBePromoted);
4345 /// \brief Check if it is profitable to promote \p ToBePromoted
4346 /// by moving downward the transition through.
4347 bool shouldPromote(const Instruction *ToBePromoted) const {
4348 // Promote only if all the operands can be statically expanded.
4349 // Indeed, we do not want to introduce any new kind of transitions.
4350 for (const Use &U : ToBePromoted->operands()) {
4351 const Value *Val = U.get();
4352 if (Val == getEndOfTransition()) {
4353 // If the use is a division and the transition is on the rhs,
4354 // we cannot promote the operation, otherwise we may create a
4355 // division by zero.
4356 if (canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()))
4360 if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) &&
4361 !isa<ConstantFP>(Val))
4364 // Check that the resulting operation is legal.
4365 int ISDOpcode = TLI.InstructionOpcodeToISD(ToBePromoted->getOpcode());
4368 return StressStoreExtract ||
4369 TLI.isOperationLegalOrCustom(
4370 ISDOpcode, TLI.getValueType(DL, getTransitionType(), true));
4373 /// \brief Check whether or not \p Use can be combined
4374 /// with the transition.
4375 /// I.e., is it possible to do Use(Transition) => AnotherUse?
4376 bool canCombine(const Instruction *Use) { return isa<StoreInst>(Use); }
4378 /// \brief Record \p ToBePromoted as part of the chain to be promoted.
4379 void enqueueForPromotion(Instruction *ToBePromoted) {
4380 InstsToBePromoted.push_back(ToBePromoted);
4383 /// \brief Set the instruction that will be combined with the transition.
4384 void recordCombineInstruction(Instruction *ToBeCombined) {
4385 assert(canCombine(ToBeCombined) && "Unsupported instruction to combine");
4386 CombineInst = ToBeCombined;
4389 /// \brief Promote all the instructions enqueued for promotion if it is
4391 /// \return True if the promotion happened, false otherwise.
4393 // Check if there is something to promote.
4394 // Right now, if we do not have anything to combine with,
4395 // we assume the promotion is not profitable.
4396 if (InstsToBePromoted.empty() || !CombineInst)
4400 if (!StressStoreExtract && !isProfitableToPromote())
4404 for (auto &ToBePromoted : InstsToBePromoted)
4405 promoteImpl(ToBePromoted);
4406 InstsToBePromoted.clear();
4410 } // End of anonymous namespace.
4412 void VectorPromoteHelper::promoteImpl(Instruction *ToBePromoted) {
4413 // At this point, we know that all the operands of ToBePromoted but Def
4414 // can be statically promoted.
4415 // For Def, we need to use its parameter in ToBePromoted:
4416 // b = ToBePromoted ty1 a
4417 // Def = Transition ty1 b to ty2
4418 // Move the transition down.
4419 // 1. Replace all uses of the promoted operation by the transition.
4420 // = ... b => = ... Def.
4421 assert(ToBePromoted->getType() == Transition->getType() &&
4422 "The type of the result of the transition does not match "
4424 ToBePromoted->replaceAllUsesWith(Transition);
4425 // 2. Update the type of the uses.
4426 // b = ToBePromoted ty2 Def => b = ToBePromoted ty1 Def.
4427 Type *TransitionTy = getTransitionType();
4428 ToBePromoted->mutateType(TransitionTy);
4429 // 3. Update all the operands of the promoted operation with promoted
4431 // b = ToBePromoted ty1 Def => b = ToBePromoted ty1 a.
4432 for (Use &U : ToBePromoted->operands()) {
4433 Value *Val = U.get();
4434 Value *NewVal = nullptr;
4435 if (Val == Transition)
4436 NewVal = Transition->getOperand(getTransitionOriginalValueIdx());
4437 else if (isa<UndefValue>(Val) || isa<ConstantInt>(Val) ||
4438 isa<ConstantFP>(Val)) {
4439 // Use a splat constant if it is not safe to use undef.
4440 NewVal = getConstantVector(
4441 cast<Constant>(Val),
4442 isa<UndefValue>(Val) ||
4443 canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()));
4445 llvm_unreachable("Did you modified shouldPromote and forgot to update "
4447 ToBePromoted->setOperand(U.getOperandNo(), NewVal);
4449 Transition->removeFromParent();
4450 Transition->insertAfter(ToBePromoted);
4451 Transition->setOperand(getTransitionOriginalValueIdx(), ToBePromoted);
4454 /// Some targets can do store(extractelement) with one instruction.
4455 /// Try to push the extractelement towards the stores when the target
4456 /// has this feature and this is profitable.
4457 bool CodeGenPrepare::optimizeExtractElementInst(Instruction *Inst) {
4458 unsigned CombineCost = UINT_MAX;
4459 if (DisableStoreExtract || !TLI ||
4460 (!StressStoreExtract &&
4461 !TLI->canCombineStoreAndExtract(Inst->getOperand(0)->getType(),
4462 Inst->getOperand(1), CombineCost)))
4465 // At this point we know that Inst is a vector to scalar transition.
4466 // Try to move it down the def-use chain, until:
4467 // - We can combine the transition with its single use
4468 // => we got rid of the transition.
4469 // - We escape the current basic block
4470 // => we would need to check that we are moving it at a cheaper place and
4471 // we do not do that for now.
4472 BasicBlock *Parent = Inst->getParent();
4473 DEBUG(dbgs() << "Found an interesting transition: " << *Inst << '\n');
4474 VectorPromoteHelper VPH(*DL, *TLI, *TTI, Inst, CombineCost);
4475 // If the transition has more than one use, assume this is not going to be
4477 while (Inst->hasOneUse()) {
4478 Instruction *ToBePromoted = cast<Instruction>(*Inst->user_begin());
4479 DEBUG(dbgs() << "Use: " << *ToBePromoted << '\n');
4481 if (ToBePromoted->getParent() != Parent) {
4482 DEBUG(dbgs() << "Instruction to promote is in a different block ("
4483 << ToBePromoted->getParent()->getName()
4484 << ") than the transition (" << Parent->getName() << ").\n");
4488 if (VPH.canCombine(ToBePromoted)) {
4489 DEBUG(dbgs() << "Assume " << *Inst << '\n'
4490 << "will be combined with: " << *ToBePromoted << '\n');
4491 VPH.recordCombineInstruction(ToBePromoted);
4492 bool Changed = VPH.promote();
4493 NumStoreExtractExposed += Changed;
4497 DEBUG(dbgs() << "Try promoting.\n");
4498 if (!VPH.canPromote(ToBePromoted) || !VPH.shouldPromote(ToBePromoted))
4501 DEBUG(dbgs() << "Promoting is possible... Enqueue for promotion!\n");
4503 VPH.enqueueForPromotion(ToBePromoted);
4504 Inst = ToBePromoted;
4509 bool CodeGenPrepare::optimizeInst(Instruction *I, bool& ModifiedDT) {
4510 // Bail out if we inserted the instruction to prevent optimizations from
4511 // stepping on each other's toes.
4512 if (InsertedInsts.count(I))
4515 if (PHINode *P = dyn_cast<PHINode>(I)) {
4516 // It is possible for very late stage optimizations (such as SimplifyCFG)
4517 // to introduce PHI nodes too late to be cleaned up. If we detect such a
4518 // trivial PHI, go ahead and zap it here.
4519 if (Value *V = SimplifyInstruction(P, *DL, TLInfo, nullptr)) {
4520 P->replaceAllUsesWith(V);
4521 P->eraseFromParent();
4528 if (CastInst *CI = dyn_cast<CastInst>(I)) {
4529 // If the source of the cast is a constant, then this should have
4530 // already been constant folded. The only reason NOT to constant fold
4531 // it is if something (e.g. LSR) was careful to place the constant
4532 // evaluation in a block other than then one that uses it (e.g. to hoist
4533 // the address of globals out of a loop). If this is the case, we don't
4534 // want to forward-subst the cast.
4535 if (isa<Constant>(CI->getOperand(0)))
4538 if (TLI && OptimizeNoopCopyExpression(CI, *TLI, *DL))
4541 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
4542 /// Sink a zext or sext into its user blocks if the target type doesn't
4543 /// fit in one register
4545 TLI->getTypeAction(CI->getContext(),
4546 TLI->getValueType(*DL, CI->getType())) ==
4547 TargetLowering::TypeExpandInteger) {
4548 return SinkCast(CI);
4550 bool MadeChange = moveExtToFormExtLoad(I);
4551 return MadeChange | optimizeExtUses(I);
4557 if (CmpInst *CI = dyn_cast<CmpInst>(I))
4558 if (!TLI || !TLI->hasMultipleConditionRegisters())
4559 return OptimizeCmpExpression(CI);
4561 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
4562 stripInvariantGroupMetadata(*LI);
4564 unsigned AS = LI->getPointerAddressSpace();
4565 return optimizeMemoryInst(I, I->getOperand(0), LI->getType(), AS);
4570 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
4571 stripInvariantGroupMetadata(*SI);
4573 unsigned AS = SI->getPointerAddressSpace();
4574 return optimizeMemoryInst(I, SI->getOperand(1),
4575 SI->getOperand(0)->getType(), AS);
4580 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(I);
4582 if (BinOp && (BinOp->getOpcode() == Instruction::AShr ||
4583 BinOp->getOpcode() == Instruction::LShr)) {
4584 ConstantInt *CI = dyn_cast<ConstantInt>(BinOp->getOperand(1));
4585 if (TLI && CI && TLI->hasExtractBitsInsn())
4586 return OptimizeExtractBits(BinOp, CI, *TLI, *DL);
4591 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
4592 if (GEPI->hasAllZeroIndices()) {
4593 /// The GEP operand must be a pointer, so must its result -> BitCast
4594 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
4595 GEPI->getName(), GEPI);
4596 GEPI->replaceAllUsesWith(NC);
4597 GEPI->eraseFromParent();
4599 optimizeInst(NC, ModifiedDT);
4605 if (CallInst *CI = dyn_cast<CallInst>(I))
4606 return optimizeCallInst(CI, ModifiedDT);
4608 if (SelectInst *SI = dyn_cast<SelectInst>(I))
4609 return optimizeSelectInst(SI);
4611 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I))
4612 return optimizeShuffleVectorInst(SVI);
4614 if (isa<ExtractElementInst>(I))
4615 return optimizeExtractElementInst(I);
4620 // In this pass we look for GEP and cast instructions that are used
4621 // across basic blocks and rewrite them to improve basic-block-at-a-time
4623 bool CodeGenPrepare::optimizeBlock(BasicBlock &BB, bool& ModifiedDT) {
4625 bool MadeChange = false;
4627 CurInstIterator = BB.begin();
4628 while (CurInstIterator != BB.end()) {
4629 MadeChange |= optimizeInst(&*CurInstIterator++, ModifiedDT);
4633 MadeChange |= dupRetToEnableTailCallOpts(&BB);
4638 // llvm.dbg.value is far away from the value then iSel may not be able
4639 // handle it properly. iSel will drop llvm.dbg.value if it can not
4640 // find a node corresponding to the value.
4641 bool CodeGenPrepare::placeDbgValues(Function &F) {
4642 bool MadeChange = false;
4643 for (BasicBlock &BB : F) {
4644 Instruction *PrevNonDbgInst = nullptr;
4645 for (BasicBlock::iterator BI = BB.begin(), BE = BB.end(); BI != BE;) {
4646 Instruction *Insn = &*BI++;
4647 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
4648 // Leave dbg.values that refer to an alloca alone. These
4649 // instrinsics describe the address of a variable (= the alloca)
4650 // being taken. They should not be moved next to the alloca
4651 // (and to the beginning of the scope), but rather stay close to
4652 // where said address is used.
4653 if (!DVI || (DVI->getValue() && isa<AllocaInst>(DVI->getValue()))) {
4654 PrevNonDbgInst = Insn;
4658 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
4659 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
4660 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
4661 DVI->removeFromParent();
4662 if (isa<PHINode>(VI))
4663 DVI->insertBefore(&*VI->getParent()->getFirstInsertionPt());
4665 DVI->insertAfter(VI);
4674 // If there is a sequence that branches based on comparing a single bit
4675 // against zero that can be combined into a single instruction, and the
4676 // target supports folding these into a single instruction, sink the
4677 // mask and compare into the branch uses. Do this before OptimizeBlock ->
4678 // OptimizeInst -> OptimizeCmpExpression, which perturbs the pattern being
4680 bool CodeGenPrepare::sinkAndCmp(Function &F) {
4681 if (!EnableAndCmpSinking)
4683 if (!TLI || !TLI->isMaskAndBranchFoldingLegal())
4685 bool MadeChange = false;
4686 for (Function::iterator I = F.begin(), E = F.end(); I != E; ) {
4687 BasicBlock *BB = &*I++;
4689 // Does this BB end with the following?
4690 // %andVal = and %val, #single-bit-set
4691 // %icmpVal = icmp %andResult, 0
4692 // br i1 %cmpVal label %dest1, label %dest2"
4693 BranchInst *Brcc = dyn_cast<BranchInst>(BB->getTerminator());
4694 if (!Brcc || !Brcc->isConditional())
4696 ICmpInst *Cmp = dyn_cast<ICmpInst>(Brcc->getOperand(0));
4697 if (!Cmp || Cmp->getParent() != BB)
4699 ConstantInt *Zero = dyn_cast<ConstantInt>(Cmp->getOperand(1));
4700 if (!Zero || !Zero->isZero())
4702 Instruction *And = dyn_cast<Instruction>(Cmp->getOperand(0));
4703 if (!And || And->getOpcode() != Instruction::And || And->getParent() != BB)
4705 ConstantInt* Mask = dyn_cast<ConstantInt>(And->getOperand(1));
4706 if (!Mask || !Mask->getUniqueInteger().isPowerOf2())
4708 DEBUG(dbgs() << "found and; icmp ?,0; brcc\n"); DEBUG(BB->dump());
4710 // Push the "and; icmp" for any users that are conditional branches.
4711 // Since there can only be one branch use per BB, we don't need to keep
4712 // track of which BBs we insert into.
4713 for (Value::use_iterator UI = Cmp->use_begin(), E = Cmp->use_end();
4717 BranchInst *BrccUser = dyn_cast<BranchInst>(*UI);
4719 if (!BrccUser || !BrccUser->isConditional())
4721 BasicBlock *UserBB = BrccUser->getParent();
4722 if (UserBB == BB) continue;
4723 DEBUG(dbgs() << "found Brcc use\n");
4725 // Sink the "and; icmp" to use.
4727 BinaryOperator *NewAnd =
4728 BinaryOperator::CreateAnd(And->getOperand(0), And->getOperand(1), "",
4731 CmpInst::Create(Cmp->getOpcode(), Cmp->getPredicate(), NewAnd, Zero,
4735 DEBUG(BrccUser->getParent()->dump());
4741 /// \brief Retrieve the probabilities of a conditional branch. Returns true on
4742 /// success, or returns false if no or invalid metadata was found.
4743 static bool extractBranchMetadata(BranchInst *BI,
4744 uint64_t &ProbTrue, uint64_t &ProbFalse) {
4745 assert(BI->isConditional() &&
4746 "Looking for probabilities on unconditional branch?");
4747 auto *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
4748 if (!ProfileData || ProfileData->getNumOperands() != 3)
4751 const auto *CITrue =
4752 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(1));
4753 const auto *CIFalse =
4754 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(2));
4755 if (!CITrue || !CIFalse)
4758 ProbTrue = CITrue->getValue().getZExtValue();
4759 ProbFalse = CIFalse->getValue().getZExtValue();
4764 /// \brief Scale down both weights to fit into uint32_t.
4765 static void scaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) {
4766 uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse;
4767 uint32_t Scale = (NewMax / UINT32_MAX) + 1;
4768 NewTrue = NewTrue / Scale;
4769 NewFalse = NewFalse / Scale;
4772 /// \brief Some targets prefer to split a conditional branch like:
4774 /// %0 = icmp ne i32 %a, 0
4775 /// %1 = icmp ne i32 %b, 0
4776 /// %or.cond = or i1 %0, %1
4777 /// br i1 %or.cond, label %TrueBB, label %FalseBB
4779 /// into multiple branch instructions like:
4782 /// %0 = icmp ne i32 %a, 0
4783 /// br i1 %0, label %TrueBB, label %bb2
4785 /// %1 = icmp ne i32 %b, 0
4786 /// br i1 %1, label %TrueBB, label %FalseBB
4788 /// This usually allows instruction selection to do even further optimizations
4789 /// and combine the compare with the branch instruction. Currently this is
4790 /// applied for targets which have "cheap" jump instructions.
4792 /// FIXME: Remove the (equivalent?) implementation in SelectionDAG.
4794 bool CodeGenPrepare::splitBranchCondition(Function &F) {
4795 if (!TM || !TM->Options.EnableFastISel || !TLI || TLI->isJumpExpensive())
4798 bool MadeChange = false;
4799 for (auto &BB : F) {
4800 // Does this BB end with the following?
4801 // %cond1 = icmp|fcmp|binary instruction ...
4802 // %cond2 = icmp|fcmp|binary instruction ...
4803 // %cond.or = or|and i1 %cond1, cond2
4804 // br i1 %cond.or label %dest1, label %dest2"
4805 BinaryOperator *LogicOp;
4806 BasicBlock *TBB, *FBB;
4807 if (!match(BB.getTerminator(), m_Br(m_OneUse(m_BinOp(LogicOp)), TBB, FBB)))
4810 auto *Br1 = cast<BranchInst>(BB.getTerminator());
4811 if (Br1->getMetadata(LLVMContext::MD_unpredictable))
4815 Value *Cond1, *Cond2;
4816 if (match(LogicOp, m_And(m_OneUse(m_Value(Cond1)),
4817 m_OneUse(m_Value(Cond2)))))
4818 Opc = Instruction::And;
4819 else if (match(LogicOp, m_Or(m_OneUse(m_Value(Cond1)),
4820 m_OneUse(m_Value(Cond2)))))
4821 Opc = Instruction::Or;
4825 if (!match(Cond1, m_CombineOr(m_Cmp(), m_BinOp())) ||
4826 !match(Cond2, m_CombineOr(m_Cmp(), m_BinOp())) )
4829 DEBUG(dbgs() << "Before branch condition splitting\n"; BB.dump());
4832 auto *InsertBefore = std::next(Function::iterator(BB))
4833 .getNodePtrUnchecked();
4834 auto TmpBB = BasicBlock::Create(BB.getContext(),
4835 BB.getName() + ".cond.split",
4836 BB.getParent(), InsertBefore);
4838 // Update original basic block by using the first condition directly by the
4839 // branch instruction and removing the no longer needed and/or instruction.
4840 Br1->setCondition(Cond1);
4841 LogicOp->eraseFromParent();
4843 // Depending on the conditon we have to either replace the true or the false
4844 // successor of the original branch instruction.
4845 if (Opc == Instruction::And)
4846 Br1->setSuccessor(0, TmpBB);
4848 Br1->setSuccessor(1, TmpBB);
4850 // Fill in the new basic block.
4851 auto *Br2 = IRBuilder<>(TmpBB).CreateCondBr(Cond2, TBB, FBB);
4852 if (auto *I = dyn_cast<Instruction>(Cond2)) {
4853 I->removeFromParent();
4854 I->insertBefore(Br2);
4857 // Update PHI nodes in both successors. The original BB needs to be
4858 // replaced in one succesor's PHI nodes, because the branch comes now from
4859 // the newly generated BB (NewBB). In the other successor we need to add one
4860 // incoming edge to the PHI nodes, because both branch instructions target
4861 // now the same successor. Depending on the original branch condition
4862 // (and/or) we have to swap the successors (TrueDest, FalseDest), so that
4863 // we perfrom the correct update for the PHI nodes.
4864 // This doesn't change the successor order of the just created branch
4865 // instruction (or any other instruction).
4866 if (Opc == Instruction::Or)
4867 std::swap(TBB, FBB);
4869 // Replace the old BB with the new BB.
4870 for (auto &I : *TBB) {
4871 PHINode *PN = dyn_cast<PHINode>(&I);
4875 while ((i = PN->getBasicBlockIndex(&BB)) >= 0)
4876 PN->setIncomingBlock(i, TmpBB);
4879 // Add another incoming edge form the new BB.
4880 for (auto &I : *FBB) {
4881 PHINode *PN = dyn_cast<PHINode>(&I);
4884 auto *Val = PN->getIncomingValueForBlock(&BB);
4885 PN->addIncoming(Val, TmpBB);
4888 // Update the branch weights (from SelectionDAGBuilder::
4889 // FindMergedConditions).
4890 if (Opc == Instruction::Or) {
4891 // Codegen X | Y as:
4900 // We have flexibility in setting Prob for BB1 and Prob for NewBB.
4901 // The requirement is that
4902 // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB)
4903 // = TrueProb for orignal BB.
4904 // Assuming the orignal weights are A and B, one choice is to set BB1's
4905 // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice
4907 // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB.
4908 // Another choice is to assume TrueProb for BB1 equals to TrueProb for
4909 // TmpBB, but the math is more complicated.
4910 uint64_t TrueWeight, FalseWeight;
4911 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
4912 uint64_t NewTrueWeight = TrueWeight;
4913 uint64_t NewFalseWeight = TrueWeight + 2 * FalseWeight;
4914 scaleWeights(NewTrueWeight, NewFalseWeight);
4915 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
4916 .createBranchWeights(TrueWeight, FalseWeight));
4918 NewTrueWeight = TrueWeight;
4919 NewFalseWeight = 2 * FalseWeight;
4920 scaleWeights(NewTrueWeight, NewFalseWeight);
4921 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
4922 .createBranchWeights(TrueWeight, FalseWeight));
4925 // Codegen X & Y as:
4933 // This requires creation of TmpBB after CurBB.
4935 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
4936 // The requirement is that
4937 // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB)
4938 // = FalseProb for orignal BB.
4939 // Assuming the orignal weights are A and B, one choice is to set BB1's
4940 // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice
4942 // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB.
4943 uint64_t TrueWeight, FalseWeight;
4944 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
4945 uint64_t NewTrueWeight = 2 * TrueWeight + FalseWeight;
4946 uint64_t NewFalseWeight = FalseWeight;
4947 scaleWeights(NewTrueWeight, NewFalseWeight);
4948 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
4949 .createBranchWeights(TrueWeight, FalseWeight));
4951 NewTrueWeight = 2 * TrueWeight;
4952 NewFalseWeight = FalseWeight;
4953 scaleWeights(NewTrueWeight, NewFalseWeight);
4954 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
4955 .createBranchWeights(TrueWeight, FalseWeight));
4959 // Note: No point in getting fancy here, since the DT info is never
4960 // available to CodeGenPrepare.
4965 DEBUG(dbgs() << "After branch condition splitting\n"; BB.dump();
4971 void CodeGenPrepare::stripInvariantGroupMetadata(Instruction &I) {
4972 if (auto *InvariantMD = I.getMetadata(LLVMContext::MD_invariant_group))
4973 I.dropUnknownNonDebugMetadata(InvariantMD->getMetadataID());