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 optimizeSwitchInst(SwitchInst *CI);
179 bool optimizeExtractElementInst(Instruction *Inst);
180 bool dupRetToEnableTailCallOpts(BasicBlock *BB);
181 bool placeDbgValues(Function &F);
182 bool sinkAndCmp(Function &F);
183 bool extLdPromotion(TypePromotionTransaction &TPT, LoadInst *&LI,
185 const SmallVectorImpl<Instruction *> &Exts,
186 unsigned CreatedInstCost);
187 bool splitBranchCondition(Function &F);
188 bool simplifyOffsetableRelocate(Instruction &I);
189 void stripInvariantGroupMetadata(Instruction &I);
193 char CodeGenPrepare::ID = 0;
194 INITIALIZE_TM_PASS(CodeGenPrepare, "codegenprepare",
195 "Optimize for code generation", false, false)
197 FunctionPass *llvm::createCodeGenPreparePass(const TargetMachine *TM) {
198 return new CodeGenPrepare(TM);
201 bool CodeGenPrepare::runOnFunction(Function &F) {
202 if (skipOptnoneFunction(F))
205 DL = &F.getParent()->getDataLayout();
207 bool EverMadeChange = false;
208 // Clear per function information.
209 InsertedInsts.clear();
210 PromotedInsts.clear();
214 TLI = TM->getSubtargetImpl(F)->getTargetLowering();
215 TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
216 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
217 OptSize = F.optForSize();
219 /// This optimization identifies DIV instructions that can be
220 /// profitably bypassed and carried out with a shorter, faster divide.
221 if (!OptSize && TLI && TLI->isSlowDivBypassed()) {
222 const DenseMap<unsigned int, unsigned int> &BypassWidths =
223 TLI->getBypassSlowDivWidths();
224 for (Function::iterator I = F.begin(); I != F.end(); I++)
225 EverMadeChange |= bypassSlowDivision(F, I, BypassWidths);
228 // Eliminate blocks that contain only PHI nodes and an
229 // unconditional branch.
230 EverMadeChange |= eliminateMostlyEmptyBlocks(F);
232 // llvm.dbg.value is far away from the value then iSel may not be able
233 // handle it properly. iSel will drop llvm.dbg.value if it can not
234 // find a node corresponding to the value.
235 EverMadeChange |= placeDbgValues(F);
237 // If there is a mask, compare against zero, and branch that can be combined
238 // into a single target instruction, push the mask and compare into branch
239 // users. Do this before OptimizeBlock -> OptimizeInst ->
240 // OptimizeCmpExpression, which perturbs the pattern being searched for.
241 if (!DisableBranchOpts) {
242 EverMadeChange |= sinkAndCmp(F);
243 EverMadeChange |= splitBranchCondition(F);
246 bool MadeChange = true;
249 for (Function::iterator I = F.begin(); I != F.end(); ) {
250 BasicBlock *BB = &*I++;
251 bool ModifiedDTOnIteration = false;
252 MadeChange |= optimizeBlock(*BB, ModifiedDTOnIteration);
254 // Restart BB iteration if the dominator tree of the Function was changed
255 if (ModifiedDTOnIteration)
258 EverMadeChange |= MadeChange;
263 if (!DisableBranchOpts) {
265 SmallPtrSet<BasicBlock*, 8> WorkList;
266 for (BasicBlock &BB : F) {
267 SmallVector<BasicBlock *, 2> Successors(succ_begin(&BB), succ_end(&BB));
268 MadeChange |= ConstantFoldTerminator(&BB, true);
269 if (!MadeChange) continue;
271 for (SmallVectorImpl<BasicBlock*>::iterator
272 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
273 if (pred_begin(*II) == pred_end(*II))
274 WorkList.insert(*II);
277 // Delete the dead blocks and any of their dead successors.
278 MadeChange |= !WorkList.empty();
279 while (!WorkList.empty()) {
280 BasicBlock *BB = *WorkList.begin();
282 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
286 for (SmallVectorImpl<BasicBlock*>::iterator
287 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
288 if (pred_begin(*II) == pred_end(*II))
289 WorkList.insert(*II);
292 // Merge pairs of basic blocks with unconditional branches, connected by
294 if (EverMadeChange || MadeChange)
295 MadeChange |= eliminateFallThrough(F);
297 EverMadeChange |= MadeChange;
300 if (!DisableGCOpts) {
301 SmallVector<Instruction *, 2> Statepoints;
302 for (BasicBlock &BB : F)
303 for (Instruction &I : BB)
305 Statepoints.push_back(&I);
306 for (auto &I : Statepoints)
307 EverMadeChange |= simplifyOffsetableRelocate(*I);
310 return EverMadeChange;
313 /// Merge basic blocks which are connected by a single edge, where one of the
314 /// basic blocks has a single successor pointing to the other basic block,
315 /// which has a single predecessor.
316 bool CodeGenPrepare::eliminateFallThrough(Function &F) {
317 bool Changed = false;
318 // Scan all of the blocks in the function, except for the entry block.
319 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
320 BasicBlock *BB = &*I++;
321 // If the destination block has a single pred, then this is a trivial
322 // edge, just collapse it.
323 BasicBlock *SinglePred = BB->getSinglePredecessor();
325 // Don't merge if BB's address is taken.
326 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
328 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
329 if (Term && !Term->isConditional()) {
331 DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n");
332 // Remember if SinglePred was the entry block of the function.
333 // If so, we will need to move BB back to the entry position.
334 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
335 MergeBasicBlockIntoOnlyPred(BB, nullptr);
337 if (isEntry && BB != &BB->getParent()->getEntryBlock())
338 BB->moveBefore(&BB->getParent()->getEntryBlock());
340 // We have erased a block. Update the iterator.
341 I = BB->getIterator();
347 /// Eliminate blocks that contain only PHI nodes, debug info directives, and an
348 /// unconditional branch. Passes before isel (e.g. LSR/loopsimplify) often split
349 /// edges in ways that are non-optimal for isel. Start by eliminating these
350 /// blocks so we can split them the way we want them.
351 bool CodeGenPrepare::eliminateMostlyEmptyBlocks(Function &F) {
352 bool MadeChange = false;
353 // Note that this intentionally skips the entry block.
354 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
355 BasicBlock *BB = &*I++;
357 // If this block doesn't end with an uncond branch, ignore it.
358 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
359 if (!BI || !BI->isUnconditional())
362 // If the instruction before the branch (skipping debug info) isn't a phi
363 // node, then other stuff is happening here.
364 BasicBlock::iterator BBI = BI->getIterator();
365 if (BBI != BB->begin()) {
367 while (isa<DbgInfoIntrinsic>(BBI)) {
368 if (BBI == BB->begin())
372 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
376 // Do not break infinite loops.
377 BasicBlock *DestBB = BI->getSuccessor(0);
381 if (!canMergeBlocks(BB, DestBB))
384 eliminateMostlyEmptyBlock(BB);
390 /// Return true if we can merge BB into DestBB if there is a single
391 /// unconditional branch between them, and BB contains no other non-phi
393 bool CodeGenPrepare::canMergeBlocks(const BasicBlock *BB,
394 const BasicBlock *DestBB) const {
395 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
396 // the successor. If there are more complex condition (e.g. preheaders),
397 // don't mess around with them.
398 BasicBlock::const_iterator BBI = BB->begin();
399 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
400 for (const User *U : PN->users()) {
401 const Instruction *UI = cast<Instruction>(U);
402 if (UI->getParent() != DestBB || !isa<PHINode>(UI))
404 // If User is inside DestBB block and it is a PHINode then check
405 // incoming value. If incoming value is not from BB then this is
406 // a complex condition (e.g. preheaders) we want to avoid here.
407 if (UI->getParent() == DestBB) {
408 if (const PHINode *UPN = dyn_cast<PHINode>(UI))
409 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
410 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
411 if (Insn && Insn->getParent() == BB &&
412 Insn->getParent() != UPN->getIncomingBlock(I))
419 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
420 // and DestBB may have conflicting incoming values for the block. If so, we
421 // can't merge the block.
422 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
423 if (!DestBBPN) return true; // no conflict.
425 // Collect the preds of BB.
426 SmallPtrSet<const BasicBlock*, 16> BBPreds;
427 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
428 // It is faster to get preds from a PHI than with pred_iterator.
429 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
430 BBPreds.insert(BBPN->getIncomingBlock(i));
432 BBPreds.insert(pred_begin(BB), pred_end(BB));
435 // Walk the preds of DestBB.
436 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
437 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
438 if (BBPreds.count(Pred)) { // Common predecessor?
439 BBI = DestBB->begin();
440 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
441 const Value *V1 = PN->getIncomingValueForBlock(Pred);
442 const Value *V2 = PN->getIncomingValueForBlock(BB);
444 // If V2 is a phi node in BB, look up what the mapped value will be.
445 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
446 if (V2PN->getParent() == BB)
447 V2 = V2PN->getIncomingValueForBlock(Pred);
449 // If there is a conflict, bail out.
450 if (V1 != V2) return false;
459 /// Eliminate a basic block that has only phi's and an unconditional branch in
461 void CodeGenPrepare::eliminateMostlyEmptyBlock(BasicBlock *BB) {
462 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
463 BasicBlock *DestBB = BI->getSuccessor(0);
465 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
467 // If the destination block has a single pred, then this is a trivial edge,
469 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
470 if (SinglePred != DestBB) {
471 // Remember if SinglePred was the entry block of the function. If so, we
472 // will need to move BB back to the entry position.
473 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
474 MergeBasicBlockIntoOnlyPred(DestBB, nullptr);
476 if (isEntry && BB != &BB->getParent()->getEntryBlock())
477 BB->moveBefore(&BB->getParent()->getEntryBlock());
479 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
484 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
485 // to handle the new incoming edges it is about to have.
487 for (BasicBlock::iterator BBI = DestBB->begin();
488 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
489 // Remove the incoming value for BB, and remember it.
490 Value *InVal = PN->removeIncomingValue(BB, false);
492 // Two options: either the InVal is a phi node defined in BB or it is some
493 // value that dominates BB.
494 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
495 if (InValPhi && InValPhi->getParent() == BB) {
496 // Add all of the input values of the input PHI as inputs of this phi.
497 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
498 PN->addIncoming(InValPhi->getIncomingValue(i),
499 InValPhi->getIncomingBlock(i));
501 // Otherwise, add one instance of the dominating value for each edge that
502 // we will be adding.
503 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
504 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
505 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
507 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
508 PN->addIncoming(InVal, *PI);
513 // The PHIs are now updated, change everything that refers to BB to use
514 // DestBB and remove BB.
515 BB->replaceAllUsesWith(DestBB);
516 BB->eraseFromParent();
519 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
522 // Computes a map of base pointer relocation instructions to corresponding
523 // derived pointer relocation instructions given a vector of all relocate calls
524 static void computeBaseDerivedRelocateMap(
525 const SmallVectorImpl<User *> &AllRelocateCalls,
526 DenseMap<IntrinsicInst *, SmallVector<IntrinsicInst *, 2>> &
528 // Collect information in two maps: one primarily for locating the base object
529 // while filling the second map; the second map is the final structure holding
530 // a mapping between Base and corresponding Derived relocate calls
531 DenseMap<std::pair<unsigned, unsigned>, IntrinsicInst *> RelocateIdxMap;
532 for (auto &U : AllRelocateCalls) {
533 GCRelocateOperands ThisRelocate(U);
534 IntrinsicInst *I = cast<IntrinsicInst>(U);
535 auto K = std::make_pair(ThisRelocate.getBasePtrIndex(),
536 ThisRelocate.getDerivedPtrIndex());
537 RelocateIdxMap.insert(std::make_pair(K, I));
539 for (auto &Item : RelocateIdxMap) {
540 std::pair<unsigned, unsigned> Key = Item.first;
541 if (Key.first == Key.second)
542 // Base relocation: nothing to insert
545 IntrinsicInst *I = Item.second;
546 auto BaseKey = std::make_pair(Key.first, Key.first);
548 // We're iterating over RelocateIdxMap so we cannot modify it.
549 auto MaybeBase = RelocateIdxMap.find(BaseKey);
550 if (MaybeBase == RelocateIdxMap.end())
551 // TODO: We might want to insert a new base object relocate and gep off
552 // that, if there are enough derived object relocates.
555 RelocateInstMap[MaybeBase->second].push_back(I);
559 // Accepts a GEP and extracts the operands into a vector provided they're all
560 // small integer constants
561 static bool getGEPSmallConstantIntOffsetV(GetElementPtrInst *GEP,
562 SmallVectorImpl<Value *> &OffsetV) {
563 for (unsigned i = 1; i < GEP->getNumOperands(); i++) {
564 // Only accept small constant integer operands
565 auto Op = dyn_cast<ConstantInt>(GEP->getOperand(i));
566 if (!Op || Op->getZExtValue() > 20)
570 for (unsigned i = 1; i < GEP->getNumOperands(); i++)
571 OffsetV.push_back(GEP->getOperand(i));
575 // Takes a RelocatedBase (base pointer relocation instruction) and Targets to
576 // replace, computes a replacement, and affects it.
578 simplifyRelocatesOffABase(IntrinsicInst *RelocatedBase,
579 const SmallVectorImpl<IntrinsicInst *> &Targets) {
580 bool MadeChange = false;
581 for (auto &ToReplace : Targets) {
582 GCRelocateOperands MasterRelocate(RelocatedBase);
583 GCRelocateOperands ThisRelocate(ToReplace);
585 assert(ThisRelocate.getBasePtrIndex() == MasterRelocate.getBasePtrIndex() &&
586 "Not relocating a derived object of the original base object");
587 if (ThisRelocate.getBasePtrIndex() == ThisRelocate.getDerivedPtrIndex()) {
588 // A duplicate relocate call. TODO: coalesce duplicates.
592 Value *Base = ThisRelocate.getBasePtr();
593 auto Derived = dyn_cast<GetElementPtrInst>(ThisRelocate.getDerivedPtr());
594 if (!Derived || Derived->getPointerOperand() != Base)
597 SmallVector<Value *, 2> OffsetV;
598 if (!getGEPSmallConstantIntOffsetV(Derived, OffsetV))
601 // Create a Builder and replace the target callsite with a gep
602 assert(RelocatedBase->getNextNode() && "Should always have one since it's not a terminator");
604 // Insert after RelocatedBase
605 IRBuilder<> Builder(RelocatedBase->getNextNode());
606 Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
608 // If gc_relocate does not match the actual type, cast it to the right type.
609 // In theory, there must be a bitcast after gc_relocate if the type does not
610 // match, and we should reuse it to get the derived pointer. But it could be
614 // %g1 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
619 // %g2 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
623 // %p1 = phi i8 addrspace(1)* [ %g1, %bb1 ], [ %g2, %bb2 ]
624 // %cast = bitcast i8 addrspace(1)* %p1 in to i32 addrspace(1)*
626 // In this case, we can not find the bitcast any more. So we insert a new bitcast
627 // no matter there is already one or not. In this way, we can handle all cases, and
628 // the extra bitcast should be optimized away in later passes.
629 Instruction *ActualRelocatedBase = RelocatedBase;
630 if (RelocatedBase->getType() != Base->getType()) {
631 ActualRelocatedBase =
632 cast<Instruction>(Builder.CreateBitCast(RelocatedBase, Base->getType()));
634 Value *Replacement = Builder.CreateGEP(
635 Derived->getSourceElementType(), ActualRelocatedBase, makeArrayRef(OffsetV));
636 Instruction *ReplacementInst = cast<Instruction>(Replacement);
637 Replacement->takeName(ToReplace);
638 // If the newly generated derived pointer's type does not match the original derived
639 // pointer's type, cast the new derived pointer to match it. Same reasoning as above.
640 Instruction *ActualReplacement = ReplacementInst;
641 if (ReplacementInst->getType() != ToReplace->getType()) {
643 cast<Instruction>(Builder.CreateBitCast(ReplacementInst, ToReplace->getType()));
645 ToReplace->replaceAllUsesWith(ActualReplacement);
646 ToReplace->eraseFromParent();
656 // %ptr = gep %base + 15
657 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
658 // %base' = relocate(%tok, i32 4, i32 4)
659 // %ptr' = relocate(%tok, i32 4, i32 5)
665 // %ptr = gep %base + 15
666 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
667 // %base' = gc.relocate(%tok, i32 4, i32 4)
668 // %ptr' = gep %base' + 15
670 bool CodeGenPrepare::simplifyOffsetableRelocate(Instruction &I) {
671 bool MadeChange = false;
672 SmallVector<User *, 2> AllRelocateCalls;
674 for (auto *U : I.users())
675 if (isGCRelocate(dyn_cast<Instruction>(U)))
676 // Collect all the relocate calls associated with a statepoint
677 AllRelocateCalls.push_back(U);
679 // We need atleast one base pointer relocation + one derived pointer
680 // relocation to mangle
681 if (AllRelocateCalls.size() < 2)
684 // RelocateInstMap is a mapping from the base relocate instruction to the
685 // corresponding derived relocate instructions
686 DenseMap<IntrinsicInst *, SmallVector<IntrinsicInst *, 2>> RelocateInstMap;
687 computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap);
688 if (RelocateInstMap.empty())
691 for (auto &Item : RelocateInstMap)
692 // Item.first is the RelocatedBase to offset against
693 // Item.second is the vector of Targets to replace
694 MadeChange = simplifyRelocatesOffABase(Item.first, Item.second);
698 /// SinkCast - Sink the specified cast instruction into its user blocks
699 static bool SinkCast(CastInst *CI) {
700 BasicBlock *DefBB = CI->getParent();
702 /// InsertedCasts - Only insert a cast in each block once.
703 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
705 bool MadeChange = false;
706 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
708 Use &TheUse = UI.getUse();
709 Instruction *User = cast<Instruction>(*UI);
711 // Figure out which BB this cast is used in. For PHI's this is the
712 // appropriate predecessor block.
713 BasicBlock *UserBB = User->getParent();
714 if (PHINode *PN = dyn_cast<PHINode>(User)) {
715 UserBB = PN->getIncomingBlock(TheUse);
718 // Preincrement use iterator so we don't invalidate it.
721 // If this user is in the same block as the cast, don't change the cast.
722 if (UserBB == DefBB) continue;
724 // If we have already inserted a cast into this block, use it.
725 CastInst *&InsertedCast = InsertedCasts[UserBB];
728 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
729 assert(InsertPt != UserBB->end());
730 InsertedCast = CastInst::Create(CI->getOpcode(), CI->getOperand(0),
731 CI->getType(), "", &*InsertPt);
734 // Replace a use of the cast with a use of the new cast.
735 TheUse = InsertedCast;
740 // If we removed all uses, nuke the cast.
741 if (CI->use_empty()) {
742 CI->eraseFromParent();
749 /// If the specified cast instruction is a noop copy (e.g. it's casting from
750 /// one pointer type to another, i32->i8 on PPC), sink it into user blocks to
751 /// reduce the number of virtual registers that must be created and coalesced.
753 /// Return true if any changes are made.
755 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI,
756 const DataLayout &DL) {
757 // If this is a noop copy,
758 EVT SrcVT = TLI.getValueType(DL, CI->getOperand(0)->getType());
759 EVT DstVT = TLI.getValueType(DL, CI->getType());
761 // This is an fp<->int conversion?
762 if (SrcVT.isInteger() != DstVT.isInteger())
765 // If this is an extension, it will be a zero or sign extension, which
767 if (SrcVT.bitsLT(DstVT)) return false;
769 // If these values will be promoted, find out what they will be promoted
770 // to. This helps us consider truncates on PPC as noop copies when they
772 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
773 TargetLowering::TypePromoteInteger)
774 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
775 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
776 TargetLowering::TypePromoteInteger)
777 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
779 // If, after promotion, these are the same types, this is a noop copy.
786 /// Try to combine CI into a call to the llvm.uadd.with.overflow intrinsic if
789 /// Return true if any changes were made.
790 static bool CombineUAddWithOverflow(CmpInst *CI) {
794 m_UAddWithOverflow(m_Value(A), m_Value(B), m_Instruction(AddI))))
797 Type *Ty = AddI->getType();
798 if (!isa<IntegerType>(Ty))
801 // We don't want to move around uses of condition values this late, so we we
802 // check if it is legal to create the call to the intrinsic in the basic
803 // block containing the icmp:
805 if (AddI->getParent() != CI->getParent() && !AddI->hasOneUse())
809 // Someday m_UAddWithOverflow may get smarter, but this is a safe assumption
811 if (AddI->hasOneUse())
812 assert(*AddI->user_begin() == CI && "expected!");
815 Module *M = CI->getParent()->getParent()->getParent();
816 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
818 auto *InsertPt = AddI->hasOneUse() ? CI : AddI;
820 auto *UAddWithOverflow =
821 CallInst::Create(F, {A, B}, "uadd.overflow", InsertPt);
822 auto *UAdd = ExtractValueInst::Create(UAddWithOverflow, 0, "uadd", InsertPt);
824 ExtractValueInst::Create(UAddWithOverflow, 1, "overflow", InsertPt);
826 CI->replaceAllUsesWith(Overflow);
827 AddI->replaceAllUsesWith(UAdd);
828 CI->eraseFromParent();
829 AddI->eraseFromParent();
833 /// Sink the given CmpInst into user blocks to reduce the number of virtual
834 /// registers that must be created and coalesced. This is a clear win except on
835 /// targets with multiple condition code registers (PowerPC), where it might
836 /// lose; some adjustment may be wanted there.
838 /// Return true if any changes are made.
839 static bool SinkCmpExpression(CmpInst *CI) {
840 BasicBlock *DefBB = CI->getParent();
842 /// Only insert a cmp in each block once.
843 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
845 bool MadeChange = false;
846 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
848 Use &TheUse = UI.getUse();
849 Instruction *User = cast<Instruction>(*UI);
851 // Preincrement use iterator so we don't invalidate it.
854 // Don't bother for PHI nodes.
855 if (isa<PHINode>(User))
858 // Figure out which BB this cmp is used in.
859 BasicBlock *UserBB = User->getParent();
861 // If this user is in the same block as the cmp, don't change the cmp.
862 if (UserBB == DefBB) continue;
864 // If we have already inserted a cmp into this block, use it.
865 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
868 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
869 assert(InsertPt != UserBB->end());
871 CmpInst::Create(CI->getOpcode(), CI->getPredicate(),
872 CI->getOperand(0), CI->getOperand(1), "", &*InsertPt);
875 // Replace a use of the cmp with a use of the new cmp.
876 TheUse = InsertedCmp;
881 // If we removed all uses, nuke the cmp.
882 if (CI->use_empty()) {
883 CI->eraseFromParent();
890 static bool OptimizeCmpExpression(CmpInst *CI) {
891 if (SinkCmpExpression(CI))
894 if (CombineUAddWithOverflow(CI))
900 /// Check if the candidates could be combined with a shift instruction, which
902 /// 1. Truncate instruction
903 /// 2. And instruction and the imm is a mask of the low bits:
904 /// imm & (imm+1) == 0
905 static bool isExtractBitsCandidateUse(Instruction *User) {
906 if (!isa<TruncInst>(User)) {
907 if (User->getOpcode() != Instruction::And ||
908 !isa<ConstantInt>(User->getOperand(1)))
911 const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
913 if ((Cimm & (Cimm + 1)).getBoolValue())
919 /// Sink both shift and truncate instruction to the use of truncate's BB.
921 SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI,
922 DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts,
923 const TargetLowering &TLI, const DataLayout &DL) {
924 BasicBlock *UserBB = User->getParent();
925 DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
926 TruncInst *TruncI = dyn_cast<TruncInst>(User);
927 bool MadeChange = false;
929 for (Value::user_iterator TruncUI = TruncI->user_begin(),
930 TruncE = TruncI->user_end();
931 TruncUI != TruncE;) {
933 Use &TruncTheUse = TruncUI.getUse();
934 Instruction *TruncUser = cast<Instruction>(*TruncUI);
935 // Preincrement use iterator so we don't invalidate it.
939 int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
943 // If the use is actually a legal node, there will not be an
944 // implicit truncate.
945 // FIXME: always querying the result type is just an
946 // approximation; some nodes' legality is determined by the
947 // operand or other means. There's no good way to find out though.
948 if (TLI.isOperationLegalOrCustom(
949 ISDOpcode, TLI.getValueType(DL, TruncUser->getType(), true)))
952 // Don't bother for PHI nodes.
953 if (isa<PHINode>(TruncUser))
956 BasicBlock *TruncUserBB = TruncUser->getParent();
958 if (UserBB == TruncUserBB)
961 BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
962 CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
964 if (!InsertedShift && !InsertedTrunc) {
965 BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
966 assert(InsertPt != TruncUserBB->end());
968 if (ShiftI->getOpcode() == Instruction::AShr)
969 InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
972 InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
976 BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
978 assert(TruncInsertPt != TruncUserBB->end());
980 InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
981 TruncI->getType(), "", &*TruncInsertPt);
985 TruncTheUse = InsertedTrunc;
991 /// Sink the shift *right* instruction into user blocks if the uses could
992 /// potentially be combined with this shift instruction and generate BitExtract
993 /// instruction. It will only be applied if the architecture supports BitExtract
994 /// instruction. Here is an example:
996 /// %x.extract.shift = lshr i64 %arg1, 32
998 /// %x.extract.trunc = trunc i64 %x.extract.shift to i16
1002 /// %x.extract.shift.1 = lshr i64 %arg1, 32
1003 /// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
1005 /// CodeGen will recoginze the pattern in BB2 and generate BitExtract
1007 /// Return true if any changes are made.
1008 static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI,
1009 const TargetLowering &TLI,
1010 const DataLayout &DL) {
1011 BasicBlock *DefBB = ShiftI->getParent();
1013 /// Only insert instructions in each block once.
1014 DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts;
1016 bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(DL, ShiftI->getType()));
1018 bool MadeChange = false;
1019 for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
1021 Use &TheUse = UI.getUse();
1022 Instruction *User = cast<Instruction>(*UI);
1023 // Preincrement use iterator so we don't invalidate it.
1026 // Don't bother for PHI nodes.
1027 if (isa<PHINode>(User))
1030 if (!isExtractBitsCandidateUse(User))
1033 BasicBlock *UserBB = User->getParent();
1035 if (UserBB == DefBB) {
1036 // If the shift and truncate instruction are in the same BB. The use of
1037 // the truncate(TruncUse) may still introduce another truncate if not
1038 // legal. In this case, we would like to sink both shift and truncate
1039 // instruction to the BB of TruncUse.
1042 // i64 shift.result = lshr i64 opnd, imm
1043 // trunc.result = trunc shift.result to i16
1046 // ----> We will have an implicit truncate here if the architecture does
1047 // not have i16 compare.
1048 // cmp i16 trunc.result, opnd2
1050 if (isa<TruncInst>(User) && shiftIsLegal
1051 // If the type of the truncate is legal, no trucate will be
1052 // introduced in other basic blocks.
1054 (!TLI.isTypeLegal(TLI.getValueType(DL, User->getType()))))
1056 SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI, DL);
1060 // If we have already inserted a shift into this block, use it.
1061 BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
1063 if (!InsertedShift) {
1064 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1065 assert(InsertPt != UserBB->end());
1067 if (ShiftI->getOpcode() == Instruction::AShr)
1068 InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
1071 InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
1077 // Replace a use of the shift with a use of the new shift.
1078 TheUse = InsertedShift;
1081 // If we removed all uses, nuke the shift.
1082 if (ShiftI->use_empty())
1083 ShiftI->eraseFromParent();
1088 // Translate a masked load intrinsic like
1089 // <16 x i32 > @llvm.masked.load( <16 x i32>* %addr, i32 align,
1090 // <16 x i1> %mask, <16 x i32> %passthru)
1091 // to a chain of basic blocks, with loading element one-by-one if
1092 // the appropriate mask bit is set
1094 // %1 = bitcast i8* %addr to i32*
1095 // %2 = extractelement <16 x i1> %mask, i32 0
1096 // %3 = icmp eq i1 %2, true
1097 // br i1 %3, label %cond.load, label %else
1099 //cond.load: ; preds = %0
1100 // %4 = getelementptr i32* %1, i32 0
1101 // %5 = load i32* %4
1102 // %6 = insertelement <16 x i32> undef, i32 %5, i32 0
1105 //else: ; preds = %0, %cond.load
1106 // %res.phi.else = phi <16 x i32> [ %6, %cond.load ], [ undef, %0 ]
1107 // %7 = extractelement <16 x i1> %mask, i32 1
1108 // %8 = icmp eq i1 %7, true
1109 // br i1 %8, label %cond.load1, label %else2
1111 //cond.load1: ; preds = %else
1112 // %9 = getelementptr i32* %1, i32 1
1113 // %10 = load i32* %9
1114 // %11 = insertelement <16 x i32> %res.phi.else, i32 %10, i32 1
1117 //else2: ; preds = %else, %cond.load1
1118 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
1119 // %12 = extractelement <16 x i1> %mask, i32 2
1120 // %13 = icmp eq i1 %12, true
1121 // br i1 %13, label %cond.load4, label %else5
1123 static void ScalarizeMaskedLoad(CallInst *CI) {
1124 Value *Ptr = CI->getArgOperand(0);
1125 Value *Alignment = CI->getArgOperand(1);
1126 Value *Mask = CI->getArgOperand(2);
1127 Value *Src0 = CI->getArgOperand(3);
1129 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
1130 VectorType *VecType = dyn_cast<VectorType>(CI->getType());
1131 assert(VecType && "Unexpected return type of masked load intrinsic");
1133 Type *EltTy = CI->getType()->getVectorElementType();
1135 IRBuilder<> Builder(CI->getContext());
1136 Instruction *InsertPt = CI;
1137 BasicBlock *IfBlock = CI->getParent();
1138 BasicBlock *CondBlock = nullptr;
1139 BasicBlock *PrevIfBlock = CI->getParent();
1141 Builder.SetInsertPoint(InsertPt);
1142 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1144 // Short-cut if the mask is all-true.
1145 bool IsAllOnesMask = isa<Constant>(Mask) &&
1146 cast<Constant>(Mask)->isAllOnesValue();
1148 if (IsAllOnesMask) {
1149 Value *NewI = Builder.CreateAlignedLoad(Ptr, AlignVal);
1150 CI->replaceAllUsesWith(NewI);
1151 CI->eraseFromParent();
1155 // Adjust alignment for the scalar instruction.
1156 AlignVal = std::min(AlignVal, VecType->getScalarSizeInBits()/8);
1157 // Bitcast %addr fron i8* to EltTy*
1159 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
1160 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
1161 unsigned VectorWidth = VecType->getNumElements();
1163 Value *UndefVal = UndefValue::get(VecType);
1165 // The result vector
1166 Value *VResult = UndefVal;
1168 if (isa<ConstantVector>(Mask)) {
1169 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1170 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
1173 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1174 LoadInst* Load = Builder.CreateAlignedLoad(Gep, AlignVal);
1175 VResult = Builder.CreateInsertElement(VResult, Load,
1176 Builder.getInt32(Idx));
1178 Value *NewI = Builder.CreateSelect(Mask, VResult, Src0);
1179 CI->replaceAllUsesWith(NewI);
1180 CI->eraseFromParent();
1184 PHINode *Phi = nullptr;
1185 Value *PrevPhi = UndefVal;
1187 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1189 // Fill the "else" block, created in the previous iteration
1191 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
1192 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
1193 // %to_load = icmp eq i1 %mask_1, true
1194 // br i1 %to_load, label %cond.load, label %else
1197 Phi = Builder.CreatePHI(VecType, 2, "res.phi.else");
1198 Phi->addIncoming(VResult, CondBlock);
1199 Phi->addIncoming(PrevPhi, PrevIfBlock);
1204 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
1205 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1206 ConstantInt::get(Predicate->getType(), 1));
1208 // Create "cond" block
1210 // %EltAddr = getelementptr i32* %1, i32 0
1211 // %Elt = load i32* %EltAddr
1212 // VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx
1214 CondBlock = IfBlock->splitBasicBlock(InsertPt->getIterator(), "cond.load");
1215 Builder.SetInsertPoint(InsertPt);
1218 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1219 LoadInst *Load = Builder.CreateAlignedLoad(Gep, AlignVal);
1220 VResult = Builder.CreateInsertElement(VResult, Load, Builder.getInt32(Idx));
1222 // Create "else" block, fill it in the next iteration
1223 BasicBlock *NewIfBlock =
1224 CondBlock->splitBasicBlock(InsertPt->getIterator(), "else");
1225 Builder.SetInsertPoint(InsertPt);
1226 Instruction *OldBr = IfBlock->getTerminator();
1227 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1228 OldBr->eraseFromParent();
1229 PrevIfBlock = IfBlock;
1230 IfBlock = NewIfBlock;
1233 Phi = Builder.CreatePHI(VecType, 2, "res.phi.select");
1234 Phi->addIncoming(VResult, CondBlock);
1235 Phi->addIncoming(PrevPhi, PrevIfBlock);
1236 Value *NewI = Builder.CreateSelect(Mask, Phi, Src0);
1237 CI->replaceAllUsesWith(NewI);
1238 CI->eraseFromParent();
1241 // Translate a masked store intrinsic, like
1242 // void @llvm.masked.store(<16 x i32> %src, <16 x i32>* %addr, i32 align,
1244 // to a chain of basic blocks, that stores element one-by-one if
1245 // the appropriate mask bit is set
1247 // %1 = bitcast i8* %addr to i32*
1248 // %2 = extractelement <16 x i1> %mask, i32 0
1249 // %3 = icmp eq i1 %2, true
1250 // br i1 %3, label %cond.store, label %else
1252 // cond.store: ; preds = %0
1253 // %4 = extractelement <16 x i32> %val, i32 0
1254 // %5 = getelementptr i32* %1, i32 0
1255 // store i32 %4, i32* %5
1258 // else: ; preds = %0, %cond.store
1259 // %6 = extractelement <16 x i1> %mask, i32 1
1260 // %7 = icmp eq i1 %6, true
1261 // br i1 %7, label %cond.store1, label %else2
1263 // cond.store1: ; preds = %else
1264 // %8 = extractelement <16 x i32> %val, i32 1
1265 // %9 = getelementptr i32* %1, i32 1
1266 // store i32 %8, i32* %9
1269 static void ScalarizeMaskedStore(CallInst *CI) {
1270 Value *Src = CI->getArgOperand(0);
1271 Value *Ptr = CI->getArgOperand(1);
1272 Value *Alignment = CI->getArgOperand(2);
1273 Value *Mask = CI->getArgOperand(3);
1275 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
1276 VectorType *VecType = dyn_cast<VectorType>(Src->getType());
1277 assert(VecType && "Unexpected data type in masked store intrinsic");
1279 Type *EltTy = VecType->getElementType();
1281 IRBuilder<> Builder(CI->getContext());
1282 Instruction *InsertPt = CI;
1283 BasicBlock *IfBlock = CI->getParent();
1284 Builder.SetInsertPoint(InsertPt);
1285 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1287 // Short-cut if the mask is all-true.
1288 bool IsAllOnesMask = isa<Constant>(Mask) &&
1289 cast<Constant>(Mask)->isAllOnesValue();
1291 if (IsAllOnesMask) {
1292 Builder.CreateAlignedStore(Src, Ptr, AlignVal);
1293 CI->eraseFromParent();
1297 // Adjust alignment for the scalar instruction.
1298 AlignVal = std::max(AlignVal, VecType->getScalarSizeInBits()/8);
1299 // Bitcast %addr fron i8* to EltTy*
1301 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
1302 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
1303 unsigned VectorWidth = VecType->getNumElements();
1305 if (isa<ConstantVector>(Mask)) {
1306 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1307 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
1309 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx));
1311 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1312 Builder.CreateAlignedStore(OneElt, Gep, AlignVal);
1314 CI->eraseFromParent();
1318 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1320 // Fill the "else" block, created in the previous iteration
1322 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
1323 // %to_store = icmp eq i1 %mask_1, true
1324 // br i1 %to_store, label %cond.store, label %else
1326 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
1327 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1328 ConstantInt::get(Predicate->getType(), 1));
1330 // Create "cond" block
1332 // %OneElt = extractelement <16 x i32> %Src, i32 Idx
1333 // %EltAddr = getelementptr i32* %1, i32 0
1334 // %store i32 %OneElt, i32* %EltAddr
1336 BasicBlock *CondBlock =
1337 IfBlock->splitBasicBlock(InsertPt->getIterator(), "cond.store");
1338 Builder.SetInsertPoint(InsertPt);
1340 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx));
1342 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1343 Builder.CreateAlignedStore(OneElt, Gep, AlignVal);
1345 // Create "else" block, fill it in the next iteration
1346 BasicBlock *NewIfBlock =
1347 CondBlock->splitBasicBlock(InsertPt->getIterator(), "else");
1348 Builder.SetInsertPoint(InsertPt);
1349 Instruction *OldBr = IfBlock->getTerminator();
1350 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1351 OldBr->eraseFromParent();
1352 IfBlock = NewIfBlock;
1354 CI->eraseFromParent();
1357 // Translate a masked gather intrinsic like
1358 // <16 x i32 > @llvm.masked.gather.v16i32( <16 x i32*> %Ptrs, i32 4,
1359 // <16 x i1> %Mask, <16 x i32> %Src)
1360 // to a chain of basic blocks, with loading element one-by-one if
1361 // the appropriate mask bit is set
1363 // % Ptrs = getelementptr i32, i32* %base, <16 x i64> %ind
1364 // % Mask0 = extractelement <16 x i1> %Mask, i32 0
1365 // % ToLoad0 = icmp eq i1 % Mask0, true
1366 // br i1 % ToLoad0, label %cond.load, label %else
1369 // % Ptr0 = extractelement <16 x i32*> %Ptrs, i32 0
1370 // % Load0 = load i32, i32* % Ptr0, align 4
1371 // % Res0 = insertelement <16 x i32> undef, i32 % Load0, i32 0
1375 // %res.phi.else = phi <16 x i32>[% Res0, %cond.load], [undef, % 0]
1376 // % Mask1 = extractelement <16 x i1> %Mask, i32 1
1377 // % ToLoad1 = icmp eq i1 % Mask1, true
1378 // br i1 % ToLoad1, label %cond.load1, label %else2
1381 // % Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1
1382 // % Load1 = load i32, i32* % Ptr1, align 4
1383 // % Res1 = insertelement <16 x i32> %res.phi.else, i32 % Load1, i32 1
1386 // % Result = select <16 x i1> %Mask, <16 x i32> %res.phi.select, <16 x i32> %Src
1387 // ret <16 x i32> %Result
1388 static void ScalarizeMaskedGather(CallInst *CI) {
1389 Value *Ptrs = CI->getArgOperand(0);
1390 Value *Alignment = CI->getArgOperand(1);
1391 Value *Mask = CI->getArgOperand(2);
1392 Value *Src0 = CI->getArgOperand(3);
1394 VectorType *VecType = dyn_cast<VectorType>(CI->getType());
1396 assert(VecType && "Unexpected return type of masked load intrinsic");
1398 IRBuilder<> Builder(CI->getContext());
1399 Instruction *InsertPt = CI;
1400 BasicBlock *IfBlock = CI->getParent();
1401 BasicBlock *CondBlock = nullptr;
1402 BasicBlock *PrevIfBlock = CI->getParent();
1403 Builder.SetInsertPoint(InsertPt);
1404 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
1406 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1408 Value *UndefVal = UndefValue::get(VecType);
1410 // The result vector
1411 Value *VResult = UndefVal;
1412 unsigned VectorWidth = VecType->getNumElements();
1414 // Shorten the way if the mask is a vector of constants.
1415 bool IsConstMask = isa<ConstantVector>(Mask);
1418 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1419 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
1421 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
1422 "Ptr" + Twine(Idx));
1423 LoadInst *Load = Builder.CreateAlignedLoad(Ptr, AlignVal,
1424 "Load" + Twine(Idx));
1425 VResult = Builder.CreateInsertElement(VResult, Load,
1426 Builder.getInt32(Idx),
1427 "Res" + Twine(Idx));
1429 Value *NewI = Builder.CreateSelect(Mask, VResult, Src0);
1430 CI->replaceAllUsesWith(NewI);
1431 CI->eraseFromParent();
1435 PHINode *Phi = nullptr;
1436 Value *PrevPhi = UndefVal;
1438 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1440 // Fill the "else" block, created in the previous iteration
1442 // %Mask1 = extractelement <16 x i1> %Mask, i32 1
1443 // %ToLoad1 = icmp eq i1 %Mask1, true
1444 // br i1 %ToLoad1, label %cond.load, label %else
1447 Phi = Builder.CreatePHI(VecType, 2, "res.phi.else");
1448 Phi->addIncoming(VResult, CondBlock);
1449 Phi->addIncoming(PrevPhi, PrevIfBlock);
1454 Value *Predicate = Builder.CreateExtractElement(Mask,
1455 Builder.getInt32(Idx),
1456 "Mask" + Twine(Idx));
1457 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1458 ConstantInt::get(Predicate->getType(), 1),
1459 "ToLoad" + Twine(Idx));
1461 // Create "cond" block
1463 // %EltAddr = getelementptr i32* %1, i32 0
1464 // %Elt = load i32* %EltAddr
1465 // VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx
1467 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.load");
1468 Builder.SetInsertPoint(InsertPt);
1470 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
1471 "Ptr" + Twine(Idx));
1472 LoadInst *Load = Builder.CreateAlignedLoad(Ptr, AlignVal,
1473 "Load" + Twine(Idx));
1474 VResult = Builder.CreateInsertElement(VResult, Load, Builder.getInt32(Idx),
1475 "Res" + Twine(Idx));
1477 // Create "else" block, fill it in the next iteration
1478 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1479 Builder.SetInsertPoint(InsertPt);
1480 Instruction *OldBr = IfBlock->getTerminator();
1481 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1482 OldBr->eraseFromParent();
1483 PrevIfBlock = IfBlock;
1484 IfBlock = NewIfBlock;
1487 Phi = Builder.CreatePHI(VecType, 2, "res.phi.select");
1488 Phi->addIncoming(VResult, CondBlock);
1489 Phi->addIncoming(PrevPhi, PrevIfBlock);
1490 Value *NewI = Builder.CreateSelect(Mask, Phi, Src0);
1491 CI->replaceAllUsesWith(NewI);
1492 CI->eraseFromParent();
1495 // Translate a masked scatter intrinsic, like
1496 // void @llvm.masked.scatter.v16i32(<16 x i32> %Src, <16 x i32*>* %Ptrs, i32 4,
1498 // to a chain of basic blocks, that stores element one-by-one if
1499 // the appropriate mask bit is set.
1501 // % Ptrs = getelementptr i32, i32* %ptr, <16 x i64> %ind
1502 // % Mask0 = extractelement <16 x i1> % Mask, i32 0
1503 // % ToStore0 = icmp eq i1 % Mask0, true
1504 // br i1 %ToStore0, label %cond.store, label %else
1507 // % Elt0 = extractelement <16 x i32> %Src, i32 0
1508 // % Ptr0 = extractelement <16 x i32*> %Ptrs, i32 0
1509 // store i32 %Elt0, i32* % Ptr0, align 4
1513 // % Mask1 = extractelement <16 x i1> % Mask, i32 1
1514 // % ToStore1 = icmp eq i1 % Mask1, true
1515 // br i1 % ToStore1, label %cond.store1, label %else2
1518 // % Elt1 = extractelement <16 x i32> %Src, i32 1
1519 // % Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1
1520 // store i32 % Elt1, i32* % Ptr1, align 4
1523 static void ScalarizeMaskedScatter(CallInst *CI) {
1524 Value *Src = CI->getArgOperand(0);
1525 Value *Ptrs = CI->getArgOperand(1);
1526 Value *Alignment = CI->getArgOperand(2);
1527 Value *Mask = CI->getArgOperand(3);
1529 assert(isa<VectorType>(Src->getType()) &&
1530 "Unexpected data type in masked scatter intrinsic");
1531 assert(isa<VectorType>(Ptrs->getType()) &&
1532 isa<PointerType>(Ptrs->getType()->getVectorElementType()) &&
1533 "Vector of pointers is expected in masked scatter intrinsic");
1535 IRBuilder<> Builder(CI->getContext());
1536 Instruction *InsertPt = CI;
1537 BasicBlock *IfBlock = CI->getParent();
1538 Builder.SetInsertPoint(InsertPt);
1539 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1541 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
1542 unsigned VectorWidth = Src->getType()->getVectorNumElements();
1544 // Shorten the way if the mask is a vector of constants.
1545 bool IsConstMask = isa<ConstantVector>(Mask);
1548 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1549 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
1551 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx),
1552 "Elt" + Twine(Idx));
1553 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
1554 "Ptr" + Twine(Idx));
1555 Builder.CreateAlignedStore(OneElt, Ptr, AlignVal);
1557 CI->eraseFromParent();
1560 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1561 // Fill the "else" block, created in the previous iteration
1563 // % Mask1 = extractelement <16 x i1> % Mask, i32 Idx
1564 // % ToStore = icmp eq i1 % Mask1, true
1565 // br i1 % ToStore, label %cond.store, label %else
1567 Value *Predicate = Builder.CreateExtractElement(Mask,
1568 Builder.getInt32(Idx),
1569 "Mask" + Twine(Idx));
1571 Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1572 ConstantInt::get(Predicate->getType(), 1),
1573 "ToStore" + Twine(Idx));
1575 // Create "cond" block
1577 // % Elt1 = extractelement <16 x i32> %Src, i32 1
1578 // % Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1
1579 // %store i32 % Elt1, i32* % Ptr1
1581 BasicBlock *CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1582 Builder.SetInsertPoint(InsertPt);
1584 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx),
1585 "Elt" + Twine(Idx));
1586 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
1587 "Ptr" + Twine(Idx));
1588 Builder.CreateAlignedStore(OneElt, Ptr, AlignVal);
1590 // Create "else" block, fill it in the next iteration
1591 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1592 Builder.SetInsertPoint(InsertPt);
1593 Instruction *OldBr = IfBlock->getTerminator();
1594 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1595 OldBr->eraseFromParent();
1596 IfBlock = NewIfBlock;
1598 CI->eraseFromParent();
1601 bool CodeGenPrepare::optimizeCallInst(CallInst *CI, bool& ModifiedDT) {
1602 BasicBlock *BB = CI->getParent();
1604 // Lower inline assembly if we can.
1605 // If we found an inline asm expession, and if the target knows how to
1606 // lower it to normal LLVM code, do so now.
1607 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
1608 if (TLI->ExpandInlineAsm(CI)) {
1609 // Avoid invalidating the iterator.
1610 CurInstIterator = BB->begin();
1611 // Avoid processing instructions out of order, which could cause
1612 // reuse before a value is defined.
1616 // Sink address computing for memory operands into the block.
1617 if (optimizeInlineAsmInst(CI))
1621 // Align the pointer arguments to this call if the target thinks it's a good
1623 unsigned MinSize, PrefAlign;
1624 if (TLI && TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) {
1625 for (auto &Arg : CI->arg_operands()) {
1626 // We want to align both objects whose address is used directly and
1627 // objects whose address is used in casts and GEPs, though it only makes
1628 // sense for GEPs if the offset is a multiple of the desired alignment and
1629 // if size - offset meets the size threshold.
1630 if (!Arg->getType()->isPointerTy())
1632 APInt Offset(DL->getPointerSizeInBits(
1633 cast<PointerType>(Arg->getType())->getAddressSpace()),
1635 Value *Val = Arg->stripAndAccumulateInBoundsConstantOffsets(*DL, Offset);
1636 uint64_t Offset2 = Offset.getLimitedValue();
1637 if ((Offset2 & (PrefAlign-1)) != 0)
1640 if ((AI = dyn_cast<AllocaInst>(Val)) && AI->getAlignment() < PrefAlign &&
1641 DL->getTypeAllocSize(AI->getAllocatedType()) >= MinSize + Offset2)
1642 AI->setAlignment(PrefAlign);
1643 // Global variables can only be aligned if they are defined in this
1644 // object (i.e. they are uniquely initialized in this object), and
1645 // over-aligning global variables that have an explicit section is
1648 if ((GV = dyn_cast<GlobalVariable>(Val)) && GV->hasUniqueInitializer() &&
1649 !GV->hasSection() && GV->getAlignment() < PrefAlign &&
1650 DL->getTypeAllocSize(GV->getType()->getElementType()) >=
1652 GV->setAlignment(PrefAlign);
1654 // If this is a memcpy (or similar) then we may be able to improve the
1656 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(CI)) {
1657 unsigned Align = getKnownAlignment(MI->getDest(), *DL);
1658 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI))
1659 Align = std::min(Align, getKnownAlignment(MTI->getSource(), *DL));
1660 if (Align > MI->getAlignment())
1661 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), Align));
1665 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
1667 switch (II->getIntrinsicID()) {
1669 case Intrinsic::objectsize: {
1670 // Lower all uses of llvm.objectsize.*
1671 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
1672 Type *ReturnTy = CI->getType();
1673 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
1675 // Substituting this can cause recursive simplifications, which can
1676 // invalidate our iterator. Use a WeakVH to hold onto it in case this
1678 WeakVH IterHandle(&*CurInstIterator);
1680 replaceAndRecursivelySimplify(CI, RetVal,
1683 // If the iterator instruction was recursively deleted, start over at the
1684 // start of the block.
1685 if (IterHandle != CurInstIterator.getNodePtrUnchecked()) {
1686 CurInstIterator = BB->begin();
1691 case Intrinsic::masked_load: {
1692 // Scalarize unsupported vector masked load
1693 if (!TTI->isLegalMaskedLoad(CI->getType())) {
1694 ScalarizeMaskedLoad(CI);
1700 case Intrinsic::masked_store: {
1701 if (!TTI->isLegalMaskedStore(CI->getArgOperand(0)->getType())) {
1702 ScalarizeMaskedStore(CI);
1708 case Intrinsic::masked_gather: {
1709 if (!TTI->isLegalMaskedGather(CI->getType())) {
1710 ScalarizeMaskedGather(CI);
1716 case Intrinsic::masked_scatter: {
1717 if (!TTI->isLegalMaskedScatter(CI->getArgOperand(0)->getType())) {
1718 ScalarizeMaskedScatter(CI);
1724 case Intrinsic::aarch64_stlxr:
1725 case Intrinsic::aarch64_stxr: {
1726 ZExtInst *ExtVal = dyn_cast<ZExtInst>(CI->getArgOperand(0));
1727 if (!ExtVal || !ExtVal->hasOneUse() ||
1728 ExtVal->getParent() == CI->getParent())
1730 // Sink a zext feeding stlxr/stxr before it, so it can be folded into it.
1731 ExtVal->moveBefore(CI);
1732 // Mark this instruction as "inserted by CGP", so that other
1733 // optimizations don't touch it.
1734 InsertedInsts.insert(ExtVal);
1737 case Intrinsic::invariant_group_barrier:
1738 II->replaceAllUsesWith(II->getArgOperand(0));
1739 II->eraseFromParent();
1744 // Unknown address space.
1745 // TODO: Target hook to pick which address space the intrinsic cares
1747 unsigned AddrSpace = ~0u;
1748 SmallVector<Value*, 2> PtrOps;
1750 if (TLI->GetAddrModeArguments(II, PtrOps, AccessTy, AddrSpace))
1751 while (!PtrOps.empty())
1752 if (optimizeMemoryInst(II, PtrOps.pop_back_val(), AccessTy, AddrSpace))
1757 // From here on out we're working with named functions.
1758 if (!CI->getCalledFunction()) return false;
1760 // Lower all default uses of _chk calls. This is very similar
1761 // to what InstCombineCalls does, but here we are only lowering calls
1762 // to fortified library functions (e.g. __memcpy_chk) that have the default
1763 // "don't know" as the objectsize. Anything else should be left alone.
1764 FortifiedLibCallSimplifier Simplifier(TLInfo, true);
1765 if (Value *V = Simplifier.optimizeCall(CI)) {
1766 CI->replaceAllUsesWith(V);
1767 CI->eraseFromParent();
1773 /// Look for opportunities to duplicate return instructions to the predecessor
1774 /// to enable tail call optimizations. The case it is currently looking for is:
1777 /// %tmp0 = tail call i32 @f0()
1778 /// br label %return
1780 /// %tmp1 = tail call i32 @f1()
1781 /// br label %return
1783 /// %tmp2 = tail call i32 @f2()
1784 /// br label %return
1786 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
1794 /// %tmp0 = tail call i32 @f0()
1797 /// %tmp1 = tail call i32 @f1()
1800 /// %tmp2 = tail call i32 @f2()
1803 bool CodeGenPrepare::dupRetToEnableTailCallOpts(BasicBlock *BB) {
1807 ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator());
1811 PHINode *PN = nullptr;
1812 BitCastInst *BCI = nullptr;
1813 Value *V = RI->getReturnValue();
1815 BCI = dyn_cast<BitCastInst>(V);
1817 V = BCI->getOperand(0);
1819 PN = dyn_cast<PHINode>(V);
1824 if (PN && PN->getParent() != BB)
1827 // It's not safe to eliminate the sign / zero extension of the return value.
1828 // See llvm::isInTailCallPosition().
1829 const Function *F = BB->getParent();
1830 AttributeSet CallerAttrs = F->getAttributes();
1831 if (CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt) ||
1832 CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
1835 // Make sure there are no instructions between the PHI and return, or that the
1836 // return is the first instruction in the block.
1838 BasicBlock::iterator BI = BB->begin();
1839 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
1841 // Also skip over the bitcast.
1846 BasicBlock::iterator BI = BB->begin();
1847 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
1852 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
1854 SmallVector<CallInst*, 4> TailCalls;
1856 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
1857 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
1858 // Make sure the phi value is indeed produced by the tail call.
1859 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
1860 TLI->mayBeEmittedAsTailCall(CI))
1861 TailCalls.push_back(CI);
1864 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
1865 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
1866 if (!VisitedBBs.insert(*PI).second)
1869 BasicBlock::InstListType &InstList = (*PI)->getInstList();
1870 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
1871 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
1872 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
1876 CallInst *CI = dyn_cast<CallInst>(&*RI);
1877 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
1878 TailCalls.push_back(CI);
1882 bool Changed = false;
1883 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
1884 CallInst *CI = TailCalls[i];
1887 // Conservatively require the attributes of the call to match those of the
1888 // return. Ignore noalias because it doesn't affect the call sequence.
1889 AttributeSet CalleeAttrs = CS.getAttributes();
1890 if (AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1891 removeAttribute(Attribute::NoAlias) !=
1892 AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1893 removeAttribute(Attribute::NoAlias))
1896 // Make sure the call instruction is followed by an unconditional branch to
1897 // the return block.
1898 BasicBlock *CallBB = CI->getParent();
1899 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
1900 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
1903 // Duplicate the return into CallBB.
1904 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
1905 ModifiedDT = Changed = true;
1909 // If we eliminated all predecessors of the block, delete the block now.
1910 if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
1911 BB->eraseFromParent();
1916 //===----------------------------------------------------------------------===//
1917 // Memory Optimization
1918 //===----------------------------------------------------------------------===//
1922 /// This is an extended version of TargetLowering::AddrMode
1923 /// which holds actual Value*'s for register values.
1924 struct ExtAddrMode : public TargetLowering::AddrMode {
1927 ExtAddrMode() : BaseReg(nullptr), ScaledReg(nullptr) {}
1928 void print(raw_ostream &OS) const;
1931 bool operator==(const ExtAddrMode& O) const {
1932 return (BaseReg == O.BaseReg) && (ScaledReg == O.ScaledReg) &&
1933 (BaseGV == O.BaseGV) && (BaseOffs == O.BaseOffs) &&
1934 (HasBaseReg == O.HasBaseReg) && (Scale == O.Scale);
1939 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
1945 void ExtAddrMode::print(raw_ostream &OS) const {
1946 bool NeedPlus = false;
1949 OS << (NeedPlus ? " + " : "")
1951 BaseGV->printAsOperand(OS, /*PrintType=*/false);
1956 OS << (NeedPlus ? " + " : "")
1962 OS << (NeedPlus ? " + " : "")
1964 BaseReg->printAsOperand(OS, /*PrintType=*/false);
1968 OS << (NeedPlus ? " + " : "")
1970 ScaledReg->printAsOperand(OS, /*PrintType=*/false);
1976 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1977 void ExtAddrMode::dump() const {
1983 /// \brief This class provides transaction based operation on the IR.
1984 /// Every change made through this class is recorded in the internal state and
1985 /// can be undone (rollback) until commit is called.
1986 class TypePromotionTransaction {
1988 /// \brief This represents the common interface of the individual transaction.
1989 /// Each class implements the logic for doing one specific modification on
1990 /// the IR via the TypePromotionTransaction.
1991 class TypePromotionAction {
1993 /// The Instruction modified.
1997 /// \brief Constructor of the action.
1998 /// The constructor performs the related action on the IR.
1999 TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
2001 virtual ~TypePromotionAction() {}
2003 /// \brief Undo the modification done by this action.
2004 /// When this method is called, the IR must be in the same state as it was
2005 /// before this action was applied.
2006 /// \pre Undoing the action works if and only if the IR is in the exact same
2007 /// state as it was directly after this action was applied.
2008 virtual void undo() = 0;
2010 /// \brief Advocate every change made by this action.
2011 /// When the results on the IR of the action are to be kept, it is important
2012 /// to call this function, otherwise hidden information may be kept forever.
2013 virtual void commit() {
2014 // Nothing to be done, this action is not doing anything.
2018 /// \brief Utility to remember the position of an instruction.
2019 class InsertionHandler {
2020 /// Position of an instruction.
2021 /// Either an instruction:
2022 /// - Is the first in a basic block: BB is used.
2023 /// - Has a previous instructon: PrevInst is used.
2025 Instruction *PrevInst;
2028 /// Remember whether or not the instruction had a previous instruction.
2029 bool HasPrevInstruction;
2032 /// \brief Record the position of \p Inst.
2033 InsertionHandler(Instruction *Inst) {
2034 BasicBlock::iterator It = Inst->getIterator();
2035 HasPrevInstruction = (It != (Inst->getParent()->begin()));
2036 if (HasPrevInstruction)
2037 Point.PrevInst = &*--It;
2039 Point.BB = Inst->getParent();
2042 /// \brief Insert \p Inst at the recorded position.
2043 void insert(Instruction *Inst) {
2044 if (HasPrevInstruction) {
2045 if (Inst->getParent())
2046 Inst->removeFromParent();
2047 Inst->insertAfter(Point.PrevInst);
2049 Instruction *Position = &*Point.BB->getFirstInsertionPt();
2050 if (Inst->getParent())
2051 Inst->moveBefore(Position);
2053 Inst->insertBefore(Position);
2058 /// \brief Move an instruction before another.
2059 class InstructionMoveBefore : public TypePromotionAction {
2060 /// Original position of the instruction.
2061 InsertionHandler Position;
2064 /// \brief Move \p Inst before \p Before.
2065 InstructionMoveBefore(Instruction *Inst, Instruction *Before)
2066 : TypePromotionAction(Inst), Position(Inst) {
2067 DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before << "\n");
2068 Inst->moveBefore(Before);
2071 /// \brief Move the instruction back to its original position.
2072 void undo() override {
2073 DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
2074 Position.insert(Inst);
2078 /// \brief Set the operand of an instruction with a new value.
2079 class OperandSetter : public TypePromotionAction {
2080 /// Original operand of the instruction.
2082 /// Index of the modified instruction.
2086 /// \brief Set \p Idx operand of \p Inst with \p NewVal.
2087 OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
2088 : TypePromotionAction(Inst), Idx(Idx) {
2089 DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
2090 << "for:" << *Inst << "\n"
2091 << "with:" << *NewVal << "\n");
2092 Origin = Inst->getOperand(Idx);
2093 Inst->setOperand(Idx, NewVal);
2096 /// \brief Restore the original value of the instruction.
2097 void undo() override {
2098 DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
2099 << "for: " << *Inst << "\n"
2100 << "with: " << *Origin << "\n");
2101 Inst->setOperand(Idx, Origin);
2105 /// \brief Hide the operands of an instruction.
2106 /// Do as if this instruction was not using any of its operands.
2107 class OperandsHider : public TypePromotionAction {
2108 /// The list of original operands.
2109 SmallVector<Value *, 4> OriginalValues;
2112 /// \brief Remove \p Inst from the uses of the operands of \p Inst.
2113 OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
2114 DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
2115 unsigned NumOpnds = Inst->getNumOperands();
2116 OriginalValues.reserve(NumOpnds);
2117 for (unsigned It = 0; It < NumOpnds; ++It) {
2118 // Save the current operand.
2119 Value *Val = Inst->getOperand(It);
2120 OriginalValues.push_back(Val);
2122 // We could use OperandSetter here, but that would imply an overhead
2123 // that we are not willing to pay.
2124 Inst->setOperand(It, UndefValue::get(Val->getType()));
2128 /// \brief Restore the original list of uses.
2129 void undo() override {
2130 DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
2131 for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
2132 Inst->setOperand(It, OriginalValues[It]);
2136 /// \brief Build a truncate instruction.
2137 class TruncBuilder : public TypePromotionAction {
2140 /// \brief Build a truncate instruction of \p Opnd producing a \p Ty
2142 /// trunc Opnd to Ty.
2143 TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
2144 IRBuilder<> Builder(Opnd);
2145 Val = Builder.CreateTrunc(Opnd, Ty, "promoted");
2146 DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n");
2149 /// \brief Get the built value.
2150 Value *getBuiltValue() { return Val; }
2152 /// \brief Remove the built instruction.
2153 void undo() override {
2154 DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n");
2155 if (Instruction *IVal = dyn_cast<Instruction>(Val))
2156 IVal->eraseFromParent();
2160 /// \brief Build a sign extension instruction.
2161 class SExtBuilder : public TypePromotionAction {
2164 /// \brief Build a sign extension instruction of \p Opnd producing a \p Ty
2166 /// sext Opnd to Ty.
2167 SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
2168 : TypePromotionAction(InsertPt) {
2169 IRBuilder<> Builder(InsertPt);
2170 Val = Builder.CreateSExt(Opnd, Ty, "promoted");
2171 DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n");
2174 /// \brief Get the built value.
2175 Value *getBuiltValue() { return Val; }
2177 /// \brief Remove the built instruction.
2178 void undo() override {
2179 DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n");
2180 if (Instruction *IVal = dyn_cast<Instruction>(Val))
2181 IVal->eraseFromParent();
2185 /// \brief Build a zero extension instruction.
2186 class ZExtBuilder : public TypePromotionAction {
2189 /// \brief Build a zero extension instruction of \p Opnd producing a \p Ty
2191 /// zext Opnd to Ty.
2192 ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
2193 : TypePromotionAction(InsertPt) {
2194 IRBuilder<> Builder(InsertPt);
2195 Val = Builder.CreateZExt(Opnd, Ty, "promoted");
2196 DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n");
2199 /// \brief Get the built value.
2200 Value *getBuiltValue() { return Val; }
2202 /// \brief Remove the built instruction.
2203 void undo() override {
2204 DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n");
2205 if (Instruction *IVal = dyn_cast<Instruction>(Val))
2206 IVal->eraseFromParent();
2210 /// \brief Mutate an instruction to another type.
2211 class TypeMutator : public TypePromotionAction {
2212 /// Record the original type.
2216 /// \brief Mutate the type of \p Inst into \p NewTy.
2217 TypeMutator(Instruction *Inst, Type *NewTy)
2218 : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
2219 DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
2221 Inst->mutateType(NewTy);
2224 /// \brief Mutate the instruction back to its original type.
2225 void undo() override {
2226 DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
2228 Inst->mutateType(OrigTy);
2232 /// \brief Replace the uses of an instruction by another instruction.
2233 class UsesReplacer : public TypePromotionAction {
2234 /// Helper structure to keep track of the replaced uses.
2235 struct InstructionAndIdx {
2236 /// The instruction using the instruction.
2238 /// The index where this instruction is used for Inst.
2240 InstructionAndIdx(Instruction *Inst, unsigned Idx)
2241 : Inst(Inst), Idx(Idx) {}
2244 /// Keep track of the original uses (pair Instruction, Index).
2245 SmallVector<InstructionAndIdx, 4> OriginalUses;
2246 typedef SmallVectorImpl<InstructionAndIdx>::iterator use_iterator;
2249 /// \brief Replace all the use of \p Inst by \p New.
2250 UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
2251 DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
2253 // Record the original uses.
2254 for (Use &U : Inst->uses()) {
2255 Instruction *UserI = cast<Instruction>(U.getUser());
2256 OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
2258 // Now, we can replace the uses.
2259 Inst->replaceAllUsesWith(New);
2262 /// \brief Reassign the original uses of Inst to Inst.
2263 void undo() override {
2264 DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
2265 for (use_iterator UseIt = OriginalUses.begin(),
2266 EndIt = OriginalUses.end();
2267 UseIt != EndIt; ++UseIt) {
2268 UseIt->Inst->setOperand(UseIt->Idx, Inst);
2273 /// \brief Remove an instruction from the IR.
2274 class InstructionRemover : public TypePromotionAction {
2275 /// Original position of the instruction.
2276 InsertionHandler Inserter;
2277 /// Helper structure to hide all the link to the instruction. In other
2278 /// words, this helps to do as if the instruction was removed.
2279 OperandsHider Hider;
2280 /// Keep track of the uses replaced, if any.
2281 UsesReplacer *Replacer;
2284 /// \brief Remove all reference of \p Inst and optinally replace all its
2286 /// \pre If !Inst->use_empty(), then New != nullptr
2287 InstructionRemover(Instruction *Inst, Value *New = nullptr)
2288 : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
2291 Replacer = new UsesReplacer(Inst, New);
2292 DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
2293 Inst->removeFromParent();
2296 ~InstructionRemover() override { delete Replacer; }
2298 /// \brief Really remove the instruction.
2299 void commit() override { delete Inst; }
2301 /// \brief Resurrect the instruction and reassign it to the proper uses if
2302 /// new value was provided when build this action.
2303 void undo() override {
2304 DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
2305 Inserter.insert(Inst);
2313 /// Restoration point.
2314 /// The restoration point is a pointer to an action instead of an iterator
2315 /// because the iterator may be invalidated but not the pointer.
2316 typedef const TypePromotionAction *ConstRestorationPt;
2317 /// Advocate every changes made in that transaction.
2319 /// Undo all the changes made after the given point.
2320 void rollback(ConstRestorationPt Point);
2321 /// Get the current restoration point.
2322 ConstRestorationPt getRestorationPoint() const;
2324 /// \name API for IR modification with state keeping to support rollback.
2326 /// Same as Instruction::setOperand.
2327 void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
2328 /// Same as Instruction::eraseFromParent.
2329 void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
2330 /// Same as Value::replaceAllUsesWith.
2331 void replaceAllUsesWith(Instruction *Inst, Value *New);
2332 /// Same as Value::mutateType.
2333 void mutateType(Instruction *Inst, Type *NewTy);
2334 /// Same as IRBuilder::createTrunc.
2335 Value *createTrunc(Instruction *Opnd, Type *Ty);
2336 /// Same as IRBuilder::createSExt.
2337 Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
2338 /// Same as IRBuilder::createZExt.
2339 Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
2340 /// Same as Instruction::moveBefore.
2341 void moveBefore(Instruction *Inst, Instruction *Before);
2345 /// The ordered list of actions made so far.
2346 SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions;
2347 typedef SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator CommitPt;
2350 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
2353 make_unique<TypePromotionTransaction::OperandSetter>(Inst, Idx, NewVal));
2356 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
2359 make_unique<TypePromotionTransaction::InstructionRemover>(Inst, NewVal));
2362 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
2364 Actions.push_back(make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
2367 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
2368 Actions.push_back(make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
2371 Value *TypePromotionTransaction::createTrunc(Instruction *Opnd,
2373 std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
2374 Value *Val = Ptr->getBuiltValue();
2375 Actions.push_back(std::move(Ptr));
2379 Value *TypePromotionTransaction::createSExt(Instruction *Inst,
2380 Value *Opnd, Type *Ty) {
2381 std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
2382 Value *Val = Ptr->getBuiltValue();
2383 Actions.push_back(std::move(Ptr));
2387 Value *TypePromotionTransaction::createZExt(Instruction *Inst,
2388 Value *Opnd, Type *Ty) {
2389 std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
2390 Value *Val = Ptr->getBuiltValue();
2391 Actions.push_back(std::move(Ptr));
2395 void TypePromotionTransaction::moveBefore(Instruction *Inst,
2396 Instruction *Before) {
2398 make_unique<TypePromotionTransaction::InstructionMoveBefore>(Inst, Before));
2401 TypePromotionTransaction::ConstRestorationPt
2402 TypePromotionTransaction::getRestorationPoint() const {
2403 return !Actions.empty() ? Actions.back().get() : nullptr;
2406 void TypePromotionTransaction::commit() {
2407 for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
2413 void TypePromotionTransaction::rollback(
2414 TypePromotionTransaction::ConstRestorationPt Point) {
2415 while (!Actions.empty() && Point != Actions.back().get()) {
2416 std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
2421 /// \brief A helper class for matching addressing modes.
2423 /// This encapsulates the logic for matching the target-legal addressing modes.
2424 class AddressingModeMatcher {
2425 SmallVectorImpl<Instruction*> &AddrModeInsts;
2426 const TargetMachine &TM;
2427 const TargetLowering &TLI;
2428 const DataLayout &DL;
2430 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
2431 /// the memory instruction that we're computing this address for.
2434 Instruction *MemoryInst;
2436 /// This is the addressing mode that we're building up. This is
2437 /// part of the return value of this addressing mode matching stuff.
2438 ExtAddrMode &AddrMode;
2440 /// The instructions inserted by other CodeGenPrepare optimizations.
2441 const SetOfInstrs &InsertedInsts;
2442 /// A map from the instructions to their type before promotion.
2443 InstrToOrigTy &PromotedInsts;
2444 /// The ongoing transaction where every action should be registered.
2445 TypePromotionTransaction &TPT;
2447 /// This is set to true when we should not do profitability checks.
2448 /// When true, IsProfitableToFoldIntoAddressingMode always returns true.
2449 bool IgnoreProfitability;
2451 AddressingModeMatcher(SmallVectorImpl<Instruction *> &AMI,
2452 const TargetMachine &TM, Type *AT, unsigned AS,
2453 Instruction *MI, ExtAddrMode &AM,
2454 const SetOfInstrs &InsertedInsts,
2455 InstrToOrigTy &PromotedInsts,
2456 TypePromotionTransaction &TPT)
2457 : AddrModeInsts(AMI), TM(TM),
2458 TLI(*TM.getSubtargetImpl(*MI->getParent()->getParent())
2459 ->getTargetLowering()),
2460 DL(MI->getModule()->getDataLayout()), AccessTy(AT), AddrSpace(AS),
2461 MemoryInst(MI), AddrMode(AM), InsertedInsts(InsertedInsts),
2462 PromotedInsts(PromotedInsts), TPT(TPT) {
2463 IgnoreProfitability = false;
2467 /// Find the maximal addressing mode that a load/store of V can fold,
2468 /// give an access type of AccessTy. This returns a list of involved
2469 /// instructions in AddrModeInsts.
2470 /// \p InsertedInsts The instructions inserted by other CodeGenPrepare
2472 /// \p PromotedInsts maps the instructions to their type before promotion.
2473 /// \p The ongoing transaction where every action should be registered.
2474 static ExtAddrMode Match(Value *V, Type *AccessTy, unsigned AS,
2475 Instruction *MemoryInst,
2476 SmallVectorImpl<Instruction*> &AddrModeInsts,
2477 const TargetMachine &TM,
2478 const SetOfInstrs &InsertedInsts,
2479 InstrToOrigTy &PromotedInsts,
2480 TypePromotionTransaction &TPT) {
2483 bool Success = AddressingModeMatcher(AddrModeInsts, TM, AccessTy, AS,
2484 MemoryInst, Result, InsertedInsts,
2485 PromotedInsts, TPT).matchAddr(V, 0);
2486 (void)Success; assert(Success && "Couldn't select *anything*?");
2490 bool matchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
2491 bool matchAddr(Value *V, unsigned Depth);
2492 bool matchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth,
2493 bool *MovedAway = nullptr);
2494 bool isProfitableToFoldIntoAddressingMode(Instruction *I,
2495 ExtAddrMode &AMBefore,
2496 ExtAddrMode &AMAfter);
2497 bool valueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
2498 bool isPromotionProfitable(unsigned NewCost, unsigned OldCost,
2499 Value *PromotedOperand) const;
2502 /// Try adding ScaleReg*Scale to the current addressing mode.
2503 /// Return true and update AddrMode if this addr mode is legal for the target,
2505 bool AddressingModeMatcher::matchScaledValue(Value *ScaleReg, int64_t Scale,
2507 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
2508 // mode. Just process that directly.
2510 return matchAddr(ScaleReg, Depth);
2512 // If the scale is 0, it takes nothing to add this.
2516 // If we already have a scale of this value, we can add to it, otherwise, we
2517 // need an available scale field.
2518 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
2521 ExtAddrMode TestAddrMode = AddrMode;
2523 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
2524 // [A+B + A*7] -> [B+A*8].
2525 TestAddrMode.Scale += Scale;
2526 TestAddrMode.ScaledReg = ScaleReg;
2528 // If the new address isn't legal, bail out.
2529 if (!TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace))
2532 // It was legal, so commit it.
2533 AddrMode = TestAddrMode;
2535 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
2536 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
2537 // X*Scale + C*Scale to addr mode.
2538 ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
2539 if (isa<Instruction>(ScaleReg) && // not a constant expr.
2540 match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
2541 TestAddrMode.ScaledReg = AddLHS;
2542 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
2544 // If this addressing mode is legal, commit it and remember that we folded
2545 // this instruction.
2546 if (TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace)) {
2547 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
2548 AddrMode = TestAddrMode;
2553 // Otherwise, not (x+c)*scale, just return what we have.
2557 /// This is a little filter, which returns true if an addressing computation
2558 /// involving I might be folded into a load/store accessing it.
2559 /// This doesn't need to be perfect, but needs to accept at least
2560 /// the set of instructions that MatchOperationAddr can.
2561 static bool MightBeFoldableInst(Instruction *I) {
2562 switch (I->getOpcode()) {
2563 case Instruction::BitCast:
2564 case Instruction::AddrSpaceCast:
2565 // Don't touch identity bitcasts.
2566 if (I->getType() == I->getOperand(0)->getType())
2568 return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
2569 case Instruction::PtrToInt:
2570 // PtrToInt is always a noop, as we know that the int type is pointer sized.
2572 case Instruction::IntToPtr:
2573 // We know the input is intptr_t, so this is foldable.
2575 case Instruction::Add:
2577 case Instruction::Mul:
2578 case Instruction::Shl:
2579 // Can only handle X*C and X << C.
2580 return isa<ConstantInt>(I->getOperand(1));
2581 case Instruction::GetElementPtr:
2588 /// \brief Check whether or not \p Val is a legal instruction for \p TLI.
2589 /// \note \p Val is assumed to be the product of some type promotion.
2590 /// Therefore if \p Val has an undefined state in \p TLI, this is assumed
2591 /// to be legal, as the non-promoted value would have had the same state.
2592 static bool isPromotedInstructionLegal(const TargetLowering &TLI,
2593 const DataLayout &DL, Value *Val) {
2594 Instruction *PromotedInst = dyn_cast<Instruction>(Val);
2597 int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
2598 // If the ISDOpcode is undefined, it was undefined before the promotion.
2601 // Otherwise, check if the promoted instruction is legal or not.
2602 return TLI.isOperationLegalOrCustom(
2603 ISDOpcode, TLI.getValueType(DL, PromotedInst->getType()));
2606 /// \brief Hepler class to perform type promotion.
2607 class TypePromotionHelper {
2608 /// \brief Utility function to check whether or not a sign or zero extension
2609 /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
2610 /// either using the operands of \p Inst or promoting \p Inst.
2611 /// The type of the extension is defined by \p IsSExt.
2612 /// In other words, check if:
2613 /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
2614 /// #1 Promotion applies:
2615 /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
2616 /// #2 Operand reuses:
2617 /// ext opnd1 to ConsideredExtType.
2618 /// \p PromotedInsts maps the instructions to their type before promotion.
2619 static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
2620 const InstrToOrigTy &PromotedInsts, bool IsSExt);
2622 /// \brief Utility function to determine if \p OpIdx should be promoted when
2623 /// promoting \p Inst.
2624 static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
2625 return !(isa<SelectInst>(Inst) && OpIdx == 0);
2628 /// \brief Utility function to promote the operand of \p Ext when this
2629 /// operand is a promotable trunc or sext or zext.
2630 /// \p PromotedInsts maps the instructions to their type before promotion.
2631 /// \p CreatedInstsCost[out] contains the cost of all instructions
2632 /// created to promote the operand of Ext.
2633 /// Newly added extensions are inserted in \p Exts.
2634 /// Newly added truncates are inserted in \p Truncs.
2635 /// Should never be called directly.
2636 /// \return The promoted value which is used instead of Ext.
2637 static Value *promoteOperandForTruncAndAnyExt(
2638 Instruction *Ext, TypePromotionTransaction &TPT,
2639 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2640 SmallVectorImpl<Instruction *> *Exts,
2641 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI);
2643 /// \brief Utility function to promote the operand of \p Ext when this
2644 /// operand is promotable and is not a supported trunc or sext.
2645 /// \p PromotedInsts maps the instructions to their type before promotion.
2646 /// \p CreatedInstsCost[out] contains the cost of all the instructions
2647 /// created to promote the operand of Ext.
2648 /// Newly added extensions are inserted in \p Exts.
2649 /// Newly added truncates are inserted in \p Truncs.
2650 /// Should never be called directly.
2651 /// \return The promoted value which is used instead of Ext.
2652 static Value *promoteOperandForOther(Instruction *Ext,
2653 TypePromotionTransaction &TPT,
2654 InstrToOrigTy &PromotedInsts,
2655 unsigned &CreatedInstsCost,
2656 SmallVectorImpl<Instruction *> *Exts,
2657 SmallVectorImpl<Instruction *> *Truncs,
2658 const TargetLowering &TLI, bool IsSExt);
2660 /// \see promoteOperandForOther.
2661 static Value *signExtendOperandForOther(
2662 Instruction *Ext, TypePromotionTransaction &TPT,
2663 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2664 SmallVectorImpl<Instruction *> *Exts,
2665 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2666 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
2667 Exts, Truncs, TLI, true);
2670 /// \see promoteOperandForOther.
2671 static Value *zeroExtendOperandForOther(
2672 Instruction *Ext, TypePromotionTransaction &TPT,
2673 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2674 SmallVectorImpl<Instruction *> *Exts,
2675 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2676 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
2677 Exts, Truncs, TLI, false);
2681 /// Type for the utility function that promotes the operand of Ext.
2682 typedef Value *(*Action)(Instruction *Ext, TypePromotionTransaction &TPT,
2683 InstrToOrigTy &PromotedInsts,
2684 unsigned &CreatedInstsCost,
2685 SmallVectorImpl<Instruction *> *Exts,
2686 SmallVectorImpl<Instruction *> *Truncs,
2687 const TargetLowering &TLI);
2688 /// \brief Given a sign/zero extend instruction \p Ext, return the approriate
2689 /// action to promote the operand of \p Ext instead of using Ext.
2690 /// \return NULL if no promotable action is possible with the current
2692 /// \p InsertedInsts keeps track of all the instructions inserted by the
2693 /// other CodeGenPrepare optimizations. This information is important
2694 /// because we do not want to promote these instructions as CodeGenPrepare
2695 /// will reinsert them later. Thus creating an infinite loop: create/remove.
2696 /// \p PromotedInsts maps the instructions to their type before promotion.
2697 static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedInsts,
2698 const TargetLowering &TLI,
2699 const InstrToOrigTy &PromotedInsts);
2702 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
2703 Type *ConsideredExtType,
2704 const InstrToOrigTy &PromotedInsts,
2706 // The promotion helper does not know how to deal with vector types yet.
2707 // To be able to fix that, we would need to fix the places where we
2708 // statically extend, e.g., constants and such.
2709 if (Inst->getType()->isVectorTy())
2712 // We can always get through zext.
2713 if (isa<ZExtInst>(Inst))
2716 // sext(sext) is ok too.
2717 if (IsSExt && isa<SExtInst>(Inst))
2720 // We can get through binary operator, if it is legal. In other words, the
2721 // binary operator must have a nuw or nsw flag.
2722 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
2723 if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
2724 ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
2725 (IsSExt && BinOp->hasNoSignedWrap())))
2728 // Check if we can do the following simplification.
2729 // ext(trunc(opnd)) --> ext(opnd)
2730 if (!isa<TruncInst>(Inst))
2733 Value *OpndVal = Inst->getOperand(0);
2734 // Check if we can use this operand in the extension.
2735 // If the type is larger than the result type of the extension, we cannot.
2736 if (!OpndVal->getType()->isIntegerTy() ||
2737 OpndVal->getType()->getIntegerBitWidth() >
2738 ConsideredExtType->getIntegerBitWidth())
2741 // If the operand of the truncate is not an instruction, we will not have
2742 // any information on the dropped bits.
2743 // (Actually we could for constant but it is not worth the extra logic).
2744 Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
2748 // Check if the source of the type is narrow enough.
2749 // I.e., check that trunc just drops extended bits of the same kind of
2751 // #1 get the type of the operand and check the kind of the extended bits.
2752 const Type *OpndType;
2753 InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
2754 if (It != PromotedInsts.end() && It->second.getInt() == IsSExt)
2755 OpndType = It->second.getPointer();
2756 else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd)))
2757 OpndType = Opnd->getOperand(0)->getType();
2761 // #2 check that the truncate just drops extended bits.
2762 return Inst->getType()->getIntegerBitWidth() >=
2763 OpndType->getIntegerBitWidth();
2766 TypePromotionHelper::Action TypePromotionHelper::getAction(
2767 Instruction *Ext, const SetOfInstrs &InsertedInsts,
2768 const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
2769 assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
2770 "Unexpected instruction type");
2771 Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0));
2772 Type *ExtTy = Ext->getType();
2773 bool IsSExt = isa<SExtInst>(Ext);
2774 // If the operand of the extension is not an instruction, we cannot
2776 // If it, check we can get through.
2777 if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt))
2780 // Do not promote if the operand has been added by codegenprepare.
2781 // Otherwise, it means we are undoing an optimization that is likely to be
2782 // redone, thus causing potential infinite loop.
2783 if (isa<TruncInst>(ExtOpnd) && InsertedInsts.count(ExtOpnd))
2786 // SExt or Trunc instructions.
2787 // Return the related handler.
2788 if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) ||
2789 isa<ZExtInst>(ExtOpnd))
2790 return promoteOperandForTruncAndAnyExt;
2792 // Regular instruction.
2793 // Abort early if we will have to insert non-free instructions.
2794 if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType()))
2796 return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
2799 Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
2800 llvm::Instruction *SExt, TypePromotionTransaction &TPT,
2801 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2802 SmallVectorImpl<Instruction *> *Exts,
2803 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2804 // By construction, the operand of SExt is an instruction. Otherwise we cannot
2805 // get through it and this method should not be called.
2806 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
2807 Value *ExtVal = SExt;
2808 bool HasMergedNonFreeExt = false;
2809 if (isa<ZExtInst>(SExtOpnd)) {
2810 // Replace s|zext(zext(opnd))
2812 HasMergedNonFreeExt = !TLI.isExtFree(SExtOpnd);
2814 TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
2815 TPT.replaceAllUsesWith(SExt, ZExt);
2816 TPT.eraseInstruction(SExt);
2819 // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
2821 TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
2823 CreatedInstsCost = 0;
2825 // Remove dead code.
2826 if (SExtOpnd->use_empty())
2827 TPT.eraseInstruction(SExtOpnd);
2829 // Check if the extension is still needed.
2830 Instruction *ExtInst = dyn_cast<Instruction>(ExtVal);
2831 if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) {
2834 Exts->push_back(ExtInst);
2835 CreatedInstsCost = !TLI.isExtFree(ExtInst) && !HasMergedNonFreeExt;
2840 // At this point we have: ext ty opnd to ty.
2841 // Reassign the uses of ExtInst to the opnd and remove ExtInst.
2842 Value *NextVal = ExtInst->getOperand(0);
2843 TPT.eraseInstruction(ExtInst, NextVal);
2847 Value *TypePromotionHelper::promoteOperandForOther(
2848 Instruction *Ext, TypePromotionTransaction &TPT,
2849 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2850 SmallVectorImpl<Instruction *> *Exts,
2851 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI,
2853 // By construction, the operand of Ext is an instruction. Otherwise we cannot
2854 // get through it and this method should not be called.
2855 Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0));
2856 CreatedInstsCost = 0;
2857 if (!ExtOpnd->hasOneUse()) {
2858 // ExtOpnd will be promoted.
2859 // All its uses, but Ext, will need to use a truncated value of the
2860 // promoted version.
2861 // Create the truncate now.
2862 Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType());
2863 if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) {
2864 ITrunc->removeFromParent();
2865 // Insert it just after the definition.
2866 ITrunc->insertAfter(ExtOpnd);
2868 Truncs->push_back(ITrunc);
2871 TPT.replaceAllUsesWith(ExtOpnd, Trunc);
2872 // Restore the operand of Ext (which has been replaced by the previous call
2873 // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
2874 TPT.setOperand(Ext, 0, ExtOpnd);
2877 // Get through the Instruction:
2878 // 1. Update its type.
2879 // 2. Replace the uses of Ext by Inst.
2880 // 3. Extend each operand that needs to be extended.
2882 // Remember the original type of the instruction before promotion.
2883 // This is useful to know that the high bits are sign extended bits.
2884 PromotedInsts.insert(std::pair<Instruction *, TypeIsSExt>(
2885 ExtOpnd, TypeIsSExt(ExtOpnd->getType(), IsSExt)));
2887 TPT.mutateType(ExtOpnd, Ext->getType());
2889 TPT.replaceAllUsesWith(Ext, ExtOpnd);
2891 Instruction *ExtForOpnd = Ext;
2893 DEBUG(dbgs() << "Propagate Ext to operands\n");
2894 for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
2896 DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n');
2897 if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() ||
2898 !shouldExtOperand(ExtOpnd, OpIdx)) {
2899 DEBUG(dbgs() << "No need to propagate\n");
2902 // Check if we can statically extend the operand.
2903 Value *Opnd = ExtOpnd->getOperand(OpIdx);
2904 if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
2905 DEBUG(dbgs() << "Statically extend\n");
2906 unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
2907 APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth)
2908 : Cst->getValue().zext(BitWidth);
2909 TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal));
2912 // UndefValue are typed, so we have to statically sign extend them.
2913 if (isa<UndefValue>(Opnd)) {
2914 DEBUG(dbgs() << "Statically extend\n");
2915 TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType()));
2919 // Otherwise we have to explicity sign extend the operand.
2920 // Check if Ext was reused to extend an operand.
2922 // If yes, create a new one.
2923 DEBUG(dbgs() << "More operands to ext\n");
2924 Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType())
2925 : TPT.createZExt(Ext, Opnd, Ext->getType());
2926 if (!isa<Instruction>(ValForExtOpnd)) {
2927 TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd);
2930 ExtForOpnd = cast<Instruction>(ValForExtOpnd);
2933 Exts->push_back(ExtForOpnd);
2934 TPT.setOperand(ExtForOpnd, 0, Opnd);
2936 // Move the sign extension before the insertion point.
2937 TPT.moveBefore(ExtForOpnd, ExtOpnd);
2938 TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd);
2939 CreatedInstsCost += !TLI.isExtFree(ExtForOpnd);
2940 // If more sext are required, new instructions will have to be created.
2941 ExtForOpnd = nullptr;
2943 if (ExtForOpnd == Ext) {
2944 DEBUG(dbgs() << "Extension is useless now\n");
2945 TPT.eraseInstruction(Ext);
2950 /// Check whether or not promoting an instruction to a wider type is profitable.
2951 /// \p NewCost gives the cost of extension instructions created by the
2953 /// \p OldCost gives the cost of extension instructions before the promotion
2954 /// plus the number of instructions that have been
2955 /// matched in the addressing mode the promotion.
2956 /// \p PromotedOperand is the value that has been promoted.
2957 /// \return True if the promotion is profitable, false otherwise.
2958 bool AddressingModeMatcher::isPromotionProfitable(
2959 unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const {
2960 DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCost << '\n');
2961 // The cost of the new extensions is greater than the cost of the
2962 // old extension plus what we folded.
2963 // This is not profitable.
2964 if (NewCost > OldCost)
2966 if (NewCost < OldCost)
2968 // The promotion is neutral but it may help folding the sign extension in
2969 // loads for instance.
2970 // Check that we did not create an illegal instruction.
2971 return isPromotedInstructionLegal(TLI, DL, PromotedOperand);
2974 /// Given an instruction or constant expr, see if we can fold the operation
2975 /// into the addressing mode. If so, update the addressing mode and return
2976 /// true, otherwise return false without modifying AddrMode.
2977 /// If \p MovedAway is not NULL, it contains the information of whether or
2978 /// not AddrInst has to be folded into the addressing mode on success.
2979 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
2980 /// because it has been moved away.
2981 /// Thus AddrInst must not be added in the matched instructions.
2982 /// This state can happen when AddrInst is a sext, since it may be moved away.
2983 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
2984 /// not be referenced anymore.
2985 bool AddressingModeMatcher::matchOperationAddr(User *AddrInst, unsigned Opcode,
2988 // Avoid exponential behavior on extremely deep expression trees.
2989 if (Depth >= 5) return false;
2991 // By default, all matched instructions stay in place.
2996 case Instruction::PtrToInt:
2997 // PtrToInt is always a noop, as we know that the int type is pointer sized.
2998 return matchAddr(AddrInst->getOperand(0), Depth);
2999 case Instruction::IntToPtr: {
3000 auto AS = AddrInst->getType()->getPointerAddressSpace();
3001 auto PtrTy = MVT::getIntegerVT(DL.getPointerSizeInBits(AS));
3002 // This inttoptr is a no-op if the integer type is pointer sized.
3003 if (TLI.getValueType(DL, AddrInst->getOperand(0)->getType()) == PtrTy)
3004 return matchAddr(AddrInst->getOperand(0), Depth);
3007 case Instruction::BitCast:
3008 // BitCast is always a noop, and we can handle it as long as it is
3009 // int->int or pointer->pointer (we don't want int<->fp or something).
3010 if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
3011 AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
3012 // Don't touch identity bitcasts. These were probably put here by LSR,
3013 // and we don't want to mess around with them. Assume it knows what it
3015 AddrInst->getOperand(0)->getType() != AddrInst->getType())
3016 return matchAddr(AddrInst->getOperand(0), Depth);
3018 case Instruction::AddrSpaceCast: {
3020 = AddrInst->getOperand(0)->getType()->getPointerAddressSpace();
3021 unsigned DestAS = AddrInst->getType()->getPointerAddressSpace();
3022 if (TLI.isNoopAddrSpaceCast(SrcAS, DestAS))
3023 return matchAddr(AddrInst->getOperand(0), Depth);
3026 case Instruction::Add: {
3027 // Check to see if we can merge in the RHS then the LHS. If so, we win.
3028 ExtAddrMode BackupAddrMode = AddrMode;
3029 unsigned OldSize = AddrModeInsts.size();
3030 // Start a transaction at this point.
3031 // The LHS may match but not the RHS.
3032 // Therefore, we need a higher level restoration point to undo partially
3033 // matched operation.
3034 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3035 TPT.getRestorationPoint();
3037 if (matchAddr(AddrInst->getOperand(1), Depth+1) &&
3038 matchAddr(AddrInst->getOperand(0), Depth+1))
3041 // Restore the old addr mode info.
3042 AddrMode = BackupAddrMode;
3043 AddrModeInsts.resize(OldSize);
3044 TPT.rollback(LastKnownGood);
3046 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
3047 if (matchAddr(AddrInst->getOperand(0), Depth+1) &&
3048 matchAddr(AddrInst->getOperand(1), Depth+1))
3051 // Otherwise we definitely can't merge the ADD in.
3052 AddrMode = BackupAddrMode;
3053 AddrModeInsts.resize(OldSize);
3054 TPT.rollback(LastKnownGood);
3057 //case Instruction::Or:
3058 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
3060 case Instruction::Mul:
3061 case Instruction::Shl: {
3062 // Can only handle X*C and X << C.
3063 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
3066 int64_t Scale = RHS->getSExtValue();
3067 if (Opcode == Instruction::Shl)
3068 Scale = 1LL << Scale;
3070 return matchScaledValue(AddrInst->getOperand(0), Scale, Depth);
3072 case Instruction::GetElementPtr: {
3073 // Scan the GEP. We check it if it contains constant offsets and at most
3074 // one variable offset.
3075 int VariableOperand = -1;
3076 unsigned VariableScale = 0;
3078 int64_t ConstantOffset = 0;
3079 gep_type_iterator GTI = gep_type_begin(AddrInst);
3080 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
3081 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
3082 const StructLayout *SL = DL.getStructLayout(STy);
3084 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
3085 ConstantOffset += SL->getElementOffset(Idx);
3087 uint64_t TypeSize = DL.getTypeAllocSize(GTI.getIndexedType());
3088 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
3089 ConstantOffset += CI->getSExtValue()*TypeSize;
3090 } else if (TypeSize) { // Scales of zero don't do anything.
3091 // We only allow one variable index at the moment.
3092 if (VariableOperand != -1)
3095 // Remember the variable index.
3096 VariableOperand = i;
3097 VariableScale = TypeSize;
3102 // A common case is for the GEP to only do a constant offset. In this case,
3103 // just add it to the disp field and check validity.
3104 if (VariableOperand == -1) {
3105 AddrMode.BaseOffs += ConstantOffset;
3106 if (ConstantOffset == 0 ||
3107 TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) {
3108 // Check to see if we can fold the base pointer in too.
3109 if (matchAddr(AddrInst->getOperand(0), Depth+1))
3112 AddrMode.BaseOffs -= ConstantOffset;
3116 // Save the valid addressing mode in case we can't match.
3117 ExtAddrMode BackupAddrMode = AddrMode;
3118 unsigned OldSize = AddrModeInsts.size();
3120 // See if the scale and offset amount is valid for this target.
3121 AddrMode.BaseOffs += ConstantOffset;
3123 // Match the base operand of the GEP.
3124 if (!matchAddr(AddrInst->getOperand(0), Depth+1)) {
3125 // If it couldn't be matched, just stuff the value in a register.
3126 if (AddrMode.HasBaseReg) {
3127 AddrMode = BackupAddrMode;
3128 AddrModeInsts.resize(OldSize);
3131 AddrMode.HasBaseReg = true;
3132 AddrMode.BaseReg = AddrInst->getOperand(0);
3135 // Match the remaining variable portion of the GEP.
3136 if (!matchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
3138 // If it couldn't be matched, try stuffing the base into a register
3139 // instead of matching it, and retrying the match of the scale.
3140 AddrMode = BackupAddrMode;
3141 AddrModeInsts.resize(OldSize);
3142 if (AddrMode.HasBaseReg)
3144 AddrMode.HasBaseReg = true;
3145 AddrMode.BaseReg = AddrInst->getOperand(0);
3146 AddrMode.BaseOffs += ConstantOffset;
3147 if (!matchScaledValue(AddrInst->getOperand(VariableOperand),
3148 VariableScale, Depth)) {
3149 // If even that didn't work, bail.
3150 AddrMode = BackupAddrMode;
3151 AddrModeInsts.resize(OldSize);
3158 case Instruction::SExt:
3159 case Instruction::ZExt: {
3160 Instruction *Ext = dyn_cast<Instruction>(AddrInst);
3164 // Try to move this ext out of the way of the addressing mode.
3165 // Ask for a method for doing so.
3166 TypePromotionHelper::Action TPH =
3167 TypePromotionHelper::getAction(Ext, InsertedInsts, TLI, PromotedInsts);
3171 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3172 TPT.getRestorationPoint();
3173 unsigned CreatedInstsCost = 0;
3174 unsigned ExtCost = !TLI.isExtFree(Ext);
3175 Value *PromotedOperand =
3176 TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI);
3177 // SExt has been moved away.
3178 // Thus either it will be rematched later in the recursive calls or it is
3179 // gone. Anyway, we must not fold it into the addressing mode at this point.
3183 // addr = gep base, idx
3185 // promotedOpnd = ext opnd <- no match here
3186 // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
3187 // addr = gep base, op <- match
3191 assert(PromotedOperand &&
3192 "TypePromotionHelper should have filtered out those cases");
3194 ExtAddrMode BackupAddrMode = AddrMode;
3195 unsigned OldSize = AddrModeInsts.size();
3197 if (!matchAddr(PromotedOperand, Depth) ||
3198 // The total of the new cost is equal to the cost of the created
3200 // The total of the old cost is equal to the cost of the extension plus
3201 // what we have saved in the addressing mode.
3202 !isPromotionProfitable(CreatedInstsCost,
3203 ExtCost + (AddrModeInsts.size() - OldSize),
3205 AddrMode = BackupAddrMode;
3206 AddrModeInsts.resize(OldSize);
3207 DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
3208 TPT.rollback(LastKnownGood);
3217 /// If we can, try to add the value of 'Addr' into the current addressing mode.
3218 /// If Addr can't be added to AddrMode this returns false and leaves AddrMode
3219 /// unmodified. This assumes that Addr is either a pointer type or intptr_t
3222 bool AddressingModeMatcher::matchAddr(Value *Addr, unsigned Depth) {
3223 // Start a transaction at this point that we will rollback if the matching
3225 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3226 TPT.getRestorationPoint();
3227 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
3228 // Fold in immediates if legal for the target.
3229 AddrMode.BaseOffs += CI->getSExtValue();
3230 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3232 AddrMode.BaseOffs -= CI->getSExtValue();
3233 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
3234 // If this is a global variable, try to fold it into the addressing mode.
3235 if (!AddrMode.BaseGV) {
3236 AddrMode.BaseGV = GV;
3237 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3239 AddrMode.BaseGV = nullptr;
3241 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
3242 ExtAddrMode BackupAddrMode = AddrMode;
3243 unsigned OldSize = AddrModeInsts.size();
3245 // Check to see if it is possible to fold this operation.
3246 bool MovedAway = false;
3247 if (matchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
3248 // This instruction may have been moved away. If so, there is nothing
3252 // Okay, it's possible to fold this. Check to see if it is actually
3253 // *profitable* to do so. We use a simple cost model to avoid increasing
3254 // register pressure too much.
3255 if (I->hasOneUse() ||
3256 isProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
3257 AddrModeInsts.push_back(I);
3261 // It isn't profitable to do this, roll back.
3262 //cerr << "NOT FOLDING: " << *I;
3263 AddrMode = BackupAddrMode;
3264 AddrModeInsts.resize(OldSize);
3265 TPT.rollback(LastKnownGood);
3267 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
3268 if (matchOperationAddr(CE, CE->getOpcode(), Depth))
3270 TPT.rollback(LastKnownGood);
3271 } else if (isa<ConstantPointerNull>(Addr)) {
3272 // Null pointer gets folded without affecting the addressing mode.
3276 // Worse case, the target should support [reg] addressing modes. :)
3277 if (!AddrMode.HasBaseReg) {
3278 AddrMode.HasBaseReg = true;
3279 AddrMode.BaseReg = Addr;
3280 // Still check for legality in case the target supports [imm] but not [i+r].
3281 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3283 AddrMode.HasBaseReg = false;
3284 AddrMode.BaseReg = nullptr;
3287 // If the base register is already taken, see if we can do [r+r].
3288 if (AddrMode.Scale == 0) {
3290 AddrMode.ScaledReg = Addr;
3291 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3294 AddrMode.ScaledReg = nullptr;
3297 TPT.rollback(LastKnownGood);
3301 /// Check to see if all uses of OpVal by the specified inline asm call are due
3302 /// to memory operands. If so, return true, otherwise return false.
3303 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
3304 const TargetMachine &TM) {
3305 const Function *F = CI->getParent()->getParent();
3306 const TargetLowering *TLI = TM.getSubtargetImpl(*F)->getTargetLowering();
3307 const TargetRegisterInfo *TRI = TM.getSubtargetImpl(*F)->getRegisterInfo();
3308 TargetLowering::AsmOperandInfoVector TargetConstraints =
3309 TLI->ParseConstraints(F->getParent()->getDataLayout(), TRI,
3310 ImmutableCallSite(CI));
3311 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
3312 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
3314 // Compute the constraint code and ConstraintType to use.
3315 TLI->ComputeConstraintToUse(OpInfo, SDValue());
3317 // If this asm operand is our Value*, and if it isn't an indirect memory
3318 // operand, we can't fold it!
3319 if (OpInfo.CallOperandVal == OpVal &&
3320 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
3321 !OpInfo.isIndirect))
3328 /// Recursively walk all the uses of I until we find a memory use.
3329 /// If we find an obviously non-foldable instruction, return true.
3330 /// Add the ultimately found memory instructions to MemoryUses.
3331 static bool FindAllMemoryUses(
3333 SmallVectorImpl<std::pair<Instruction *, unsigned>> &MemoryUses,
3334 SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetMachine &TM) {
3335 // If we already considered this instruction, we're done.
3336 if (!ConsideredInsts.insert(I).second)
3339 // If this is an obviously unfoldable instruction, bail out.
3340 if (!MightBeFoldableInst(I))
3343 // Loop over all the uses, recursively processing them.
3344 for (Use &U : I->uses()) {
3345 Instruction *UserI = cast<Instruction>(U.getUser());
3347 if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
3348 MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
3352 if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
3353 unsigned opNo = U.getOperandNo();
3354 if (opNo == 0) return true; // Storing addr, not into addr.
3355 MemoryUses.push_back(std::make_pair(SI, opNo));
3359 if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
3360 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
3361 if (!IA) return true;
3363 // If this is a memory operand, we're cool, otherwise bail out.
3364 if (!IsOperandAMemoryOperand(CI, IA, I, TM))
3369 if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TM))
3376 /// Return true if Val is already known to be live at the use site that we're
3377 /// folding it into. If so, there is no cost to include it in the addressing
3378 /// mode. KnownLive1 and KnownLive2 are two values that we know are live at the
3379 /// instruction already.
3380 bool AddressingModeMatcher::valueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
3381 Value *KnownLive2) {
3382 // If Val is either of the known-live values, we know it is live!
3383 if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
3386 // All values other than instructions and arguments (e.g. constants) are live.
3387 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
3389 // If Val is a constant sized alloca in the entry block, it is live, this is
3390 // true because it is just a reference to the stack/frame pointer, which is
3391 // live for the whole function.
3392 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
3393 if (AI->isStaticAlloca())
3396 // Check to see if this value is already used in the memory instruction's
3397 // block. If so, it's already live into the block at the very least, so we
3398 // can reasonably fold it.
3399 return Val->isUsedInBasicBlock(MemoryInst->getParent());
3402 /// It is possible for the addressing mode of the machine to fold the specified
3403 /// instruction into a load or store that ultimately uses it.
3404 /// However, the specified instruction has multiple uses.
3405 /// Given this, it may actually increase register pressure to fold it
3406 /// into the load. For example, consider this code:
3410 /// use(Y) -> nonload/store
3414 /// In this case, Y has multiple uses, and can be folded into the load of Z
3415 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
3416 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
3417 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
3418 /// number of computations either.
3420 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
3421 /// X was live across 'load Z' for other reasons, we actually *would* want to
3422 /// fold the addressing mode in the Z case. This would make Y die earlier.
3423 bool AddressingModeMatcher::
3424 isProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
3425 ExtAddrMode &AMAfter) {
3426 if (IgnoreProfitability) return true;
3428 // AMBefore is the addressing mode before this instruction was folded into it,
3429 // and AMAfter is the addressing mode after the instruction was folded. Get
3430 // the set of registers referenced by AMAfter and subtract out those
3431 // referenced by AMBefore: this is the set of values which folding in this
3432 // address extends the lifetime of.
3434 // Note that there are only two potential values being referenced here,
3435 // BaseReg and ScaleReg (global addresses are always available, as are any
3436 // folded immediates).
3437 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
3439 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
3440 // lifetime wasn't extended by adding this instruction.
3441 if (valueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
3443 if (valueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
3444 ScaledReg = nullptr;
3446 // If folding this instruction (and it's subexprs) didn't extend any live
3447 // ranges, we're ok with it.
3448 if (!BaseReg && !ScaledReg)
3451 // If all uses of this instruction are ultimately load/store/inlineasm's,
3452 // check to see if their addressing modes will include this instruction. If
3453 // so, we can fold it into all uses, so it doesn't matter if it has multiple
3455 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
3456 SmallPtrSet<Instruction*, 16> ConsideredInsts;
3457 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TM))
3458 return false; // Has a non-memory, non-foldable use!
3460 // Now that we know that all uses of this instruction are part of a chain of
3461 // computation involving only operations that could theoretically be folded
3462 // into a memory use, loop over each of these uses and see if they could
3463 // *actually* fold the instruction.
3464 SmallVector<Instruction*, 32> MatchedAddrModeInsts;
3465 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
3466 Instruction *User = MemoryUses[i].first;
3467 unsigned OpNo = MemoryUses[i].second;
3469 // Get the access type of this use. If the use isn't a pointer, we don't
3470 // know what it accesses.
3471 Value *Address = User->getOperand(OpNo);
3472 PointerType *AddrTy = dyn_cast<PointerType>(Address->getType());
3475 Type *AddressAccessTy = AddrTy->getElementType();
3476 unsigned AS = AddrTy->getAddressSpace();
3478 // Do a match against the root of this address, ignoring profitability. This
3479 // will tell us if the addressing mode for the memory operation will
3480 // *actually* cover the shared instruction.
3482 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3483 TPT.getRestorationPoint();
3484 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TM, AddressAccessTy, AS,
3485 MemoryInst, Result, InsertedInsts,
3486 PromotedInsts, TPT);
3487 Matcher.IgnoreProfitability = true;
3488 bool Success = Matcher.matchAddr(Address, 0);
3489 (void)Success; assert(Success && "Couldn't select *anything*?");
3491 // The match was to check the profitability, the changes made are not
3492 // part of the original matcher. Therefore, they should be dropped
3493 // otherwise the original matcher will not present the right state.
3494 TPT.rollback(LastKnownGood);
3496 // If the match didn't cover I, then it won't be shared by it.
3497 if (std::find(MatchedAddrModeInsts.begin(), MatchedAddrModeInsts.end(),
3498 I) == MatchedAddrModeInsts.end())
3501 MatchedAddrModeInsts.clear();
3507 } // end anonymous namespace
3509 /// Return true if the specified values are defined in a
3510 /// different basic block than BB.
3511 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
3512 if (Instruction *I = dyn_cast<Instruction>(V))
3513 return I->getParent() != BB;
3517 /// Load and Store Instructions often have addressing modes that can do
3518 /// significant amounts of computation. As such, instruction selection will try
3519 /// to get the load or store to do as much computation as possible for the
3520 /// program. The problem is that isel can only see within a single block. As
3521 /// such, we sink as much legal addressing mode work into the block as possible.
3523 /// This method is used to optimize both load/store and inline asms with memory
3525 bool CodeGenPrepare::optimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
3526 Type *AccessTy, unsigned AddrSpace) {
3529 // Try to collapse single-value PHI nodes. This is necessary to undo
3530 // unprofitable PRE transformations.
3531 SmallVector<Value*, 8> worklist;
3532 SmallPtrSet<Value*, 16> Visited;
3533 worklist.push_back(Addr);
3535 // Use a worklist to iteratively look through PHI nodes, and ensure that
3536 // the addressing mode obtained from the non-PHI roots of the graph
3538 Value *Consensus = nullptr;
3539 unsigned NumUsesConsensus = 0;
3540 bool IsNumUsesConsensusValid = false;
3541 SmallVector<Instruction*, 16> AddrModeInsts;
3542 ExtAddrMode AddrMode;
3543 TypePromotionTransaction TPT;
3544 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3545 TPT.getRestorationPoint();
3546 while (!worklist.empty()) {
3547 Value *V = worklist.back();
3548 worklist.pop_back();
3550 // Break use-def graph loops.
3551 if (!Visited.insert(V).second) {
3552 Consensus = nullptr;
3556 // For a PHI node, push all of its incoming values.
3557 if (PHINode *P = dyn_cast<PHINode>(V)) {
3558 for (Value *IncValue : P->incoming_values())
3559 worklist.push_back(IncValue);
3563 // For non-PHIs, determine the addressing mode being computed.
3564 SmallVector<Instruction*, 16> NewAddrModeInsts;
3565 ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
3566 V, AccessTy, AddrSpace, MemoryInst, NewAddrModeInsts, *TM,
3567 InsertedInsts, PromotedInsts, TPT);
3569 // This check is broken into two cases with very similar code to avoid using
3570 // getNumUses() as much as possible. Some values have a lot of uses, so
3571 // calling getNumUses() unconditionally caused a significant compile-time
3575 AddrMode = NewAddrMode;
3576 AddrModeInsts = NewAddrModeInsts;
3578 } else if (NewAddrMode == AddrMode) {
3579 if (!IsNumUsesConsensusValid) {
3580 NumUsesConsensus = Consensus->getNumUses();
3581 IsNumUsesConsensusValid = true;
3584 // Ensure that the obtained addressing mode is equivalent to that obtained
3585 // for all other roots of the PHI traversal. Also, when choosing one
3586 // such root as representative, select the one with the most uses in order
3587 // to keep the cost modeling heuristics in AddressingModeMatcher
3589 unsigned NumUses = V->getNumUses();
3590 if (NumUses > NumUsesConsensus) {
3592 NumUsesConsensus = NumUses;
3593 AddrModeInsts = NewAddrModeInsts;
3598 Consensus = nullptr;
3602 // If the addressing mode couldn't be determined, or if multiple different
3603 // ones were determined, bail out now.
3605 TPT.rollback(LastKnownGood);
3610 // Check to see if any of the instructions supersumed by this addr mode are
3611 // non-local to I's BB.
3612 bool AnyNonLocal = false;
3613 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
3614 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
3620 // If all the instructions matched are already in this BB, don't do anything.
3622 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
3626 // Insert this computation right after this user. Since our caller is
3627 // scanning from the top of the BB to the bottom, reuse of the expr are
3628 // guaranteed to happen later.
3629 IRBuilder<> Builder(MemoryInst);
3631 // Now that we determined the addressing expression we want to use and know
3632 // that we have to sink it into this block. Check to see if we have already
3633 // done this for some other load/store instr in this block. If so, reuse the
3635 Value *&SunkAddr = SunkAddrs[Addr];
3637 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
3638 << *MemoryInst << "\n");
3639 if (SunkAddr->getType() != Addr->getType())
3640 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
3641 } else if (AddrSinkUsingGEPs ||
3642 (!AddrSinkUsingGEPs.getNumOccurrences() && TM &&
3643 TM->getSubtargetImpl(*MemoryInst->getParent()->getParent())
3645 // By default, we use the GEP-based method when AA is used later. This
3646 // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
3647 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
3648 << *MemoryInst << "\n");
3649 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
3650 Value *ResultPtr = nullptr, *ResultIndex = nullptr;
3652 // First, find the pointer.
3653 if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
3654 ResultPtr = AddrMode.BaseReg;
3655 AddrMode.BaseReg = nullptr;
3658 if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
3659 // We can't add more than one pointer together, nor can we scale a
3660 // pointer (both of which seem meaningless).
3661 if (ResultPtr || AddrMode.Scale != 1)
3664 ResultPtr = AddrMode.ScaledReg;
3668 if (AddrMode.BaseGV) {
3672 ResultPtr = AddrMode.BaseGV;
3675 // If the real base value actually came from an inttoptr, then the matcher
3676 // will look through it and provide only the integer value. In that case,
3678 if (!ResultPtr && AddrMode.BaseReg) {
3680 Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(), "sunkaddr");
3681 AddrMode.BaseReg = nullptr;
3682 } else if (!ResultPtr && AddrMode.Scale == 1) {
3684 Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(), "sunkaddr");
3689 !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
3690 SunkAddr = Constant::getNullValue(Addr->getType());
3691 } else if (!ResultPtr) {
3695 Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
3696 Type *I8Ty = Builder.getInt8Ty();
3698 // Start with the base register. Do this first so that subsequent address
3699 // matching finds it last, which will prevent it from trying to match it
3700 // as the scaled value in case it happens to be a mul. That would be
3701 // problematic if we've sunk a different mul for the scale, because then
3702 // we'd end up sinking both muls.
3703 if (AddrMode.BaseReg) {
3704 Value *V = AddrMode.BaseReg;
3705 if (V->getType() != IntPtrTy)
3706 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
3711 // Add the scale value.
3712 if (AddrMode.Scale) {
3713 Value *V = AddrMode.ScaledReg;
3714 if (V->getType() == IntPtrTy) {
3716 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
3717 cast<IntegerType>(V->getType())->getBitWidth()) {
3718 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
3720 // It is only safe to sign extend the BaseReg if we know that the math
3721 // required to create it did not overflow before we extend it. Since
3722 // the original IR value was tossed in favor of a constant back when
3723 // the AddrMode was created we need to bail out gracefully if widths
3724 // do not match instead of extending it.
3725 Instruction *I = dyn_cast_or_null<Instruction>(ResultIndex);
3726 if (I && (ResultIndex != AddrMode.BaseReg))
3727 I->eraseFromParent();
3731 if (AddrMode.Scale != 1)
3732 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
3735 ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
3740 // Add in the Base Offset if present.
3741 if (AddrMode.BaseOffs) {
3742 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
3744 // We need to add this separately from the scale above to help with
3745 // SDAG consecutive load/store merging.
3746 if (ResultPtr->getType() != I8PtrTy)
3747 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
3748 ResultPtr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
3755 SunkAddr = ResultPtr;
3757 if (ResultPtr->getType() != I8PtrTy)
3758 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
3759 SunkAddr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
3762 if (SunkAddr->getType() != Addr->getType())
3763 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
3766 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
3767 << *MemoryInst << "\n");
3768 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
3769 Value *Result = nullptr;
3771 // Start with the base register. Do this first so that subsequent address
3772 // matching finds it last, which will prevent it from trying to match it
3773 // as the scaled value in case it happens to be a mul. That would be
3774 // problematic if we've sunk a different mul for the scale, because then
3775 // we'd end up sinking both muls.
3776 if (AddrMode.BaseReg) {
3777 Value *V = AddrMode.BaseReg;
3778 if (V->getType()->isPointerTy())
3779 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
3780 if (V->getType() != IntPtrTy)
3781 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
3785 // Add the scale value.
3786 if (AddrMode.Scale) {
3787 Value *V = AddrMode.ScaledReg;
3788 if (V->getType() == IntPtrTy) {
3790 } else if (V->getType()->isPointerTy()) {
3791 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
3792 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
3793 cast<IntegerType>(V->getType())->getBitWidth()) {
3794 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
3796 // It is only safe to sign extend the BaseReg if we know that the math
3797 // required to create it did not overflow before we extend it. Since
3798 // the original IR value was tossed in favor of a constant back when
3799 // the AddrMode was created we need to bail out gracefully if widths
3800 // do not match instead of extending it.
3801 Instruction *I = dyn_cast_or_null<Instruction>(Result);
3802 if (I && (Result != AddrMode.BaseReg))
3803 I->eraseFromParent();
3806 if (AddrMode.Scale != 1)
3807 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
3810 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3815 // Add in the BaseGV if present.
3816 if (AddrMode.BaseGV) {
3817 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
3819 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3824 // Add in the Base Offset if present.
3825 if (AddrMode.BaseOffs) {
3826 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
3828 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3834 SunkAddr = Constant::getNullValue(Addr->getType());
3836 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
3839 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
3841 // If we have no uses, recursively delete the value and all dead instructions
3843 if (Repl->use_empty()) {
3844 // This can cause recursive deletion, which can invalidate our iterator.
3845 // Use a WeakVH to hold onto it in case this happens.
3846 WeakVH IterHandle(&*CurInstIterator);
3847 BasicBlock *BB = CurInstIterator->getParent();
3849 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
3851 if (IterHandle != CurInstIterator.getNodePtrUnchecked()) {
3852 // If the iterator instruction was recursively deleted, start over at the
3853 // start of the block.
3854 CurInstIterator = BB->begin();
3862 /// If there are any memory operands, use OptimizeMemoryInst to sink their
3863 /// address computing into the block when possible / profitable.
3864 bool CodeGenPrepare::optimizeInlineAsmInst(CallInst *CS) {
3865 bool MadeChange = false;
3867 const TargetRegisterInfo *TRI =
3868 TM->getSubtargetImpl(*CS->getParent()->getParent())->getRegisterInfo();
3869 TargetLowering::AsmOperandInfoVector TargetConstraints =
3870 TLI->ParseConstraints(*DL, TRI, CS);
3872 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
3873 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
3875 // Compute the constraint code and ConstraintType to use.
3876 TLI->ComputeConstraintToUse(OpInfo, SDValue());
3878 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
3879 OpInfo.isIndirect) {
3880 Value *OpVal = CS->getArgOperand(ArgNo++);
3881 MadeChange |= optimizeMemoryInst(CS, OpVal, OpVal->getType(), ~0u);
3882 } else if (OpInfo.Type == InlineAsm::isInput)
3889 /// \brief Check if all the uses of \p Inst are equivalent (or free) zero or
3890 /// sign extensions.
3891 static bool hasSameExtUse(Instruction *Inst, const TargetLowering &TLI) {
3892 assert(!Inst->use_empty() && "Input must have at least one use");
3893 const Instruction *FirstUser = cast<Instruction>(*Inst->user_begin());
3894 bool IsSExt = isa<SExtInst>(FirstUser);
3895 Type *ExtTy = FirstUser->getType();
3896 for (const User *U : Inst->users()) {
3897 const Instruction *UI = cast<Instruction>(U);
3898 if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI)))
3900 Type *CurTy = UI->getType();
3901 // Same input and output types: Same instruction after CSE.
3905 // If IsSExt is true, we are in this situation:
3907 // b = sext ty1 a to ty2
3908 // c = sext ty1 a to ty3
3909 // Assuming ty2 is shorter than ty3, this could be turned into:
3911 // b = sext ty1 a to ty2
3912 // c = sext ty2 b to ty3
3913 // However, the last sext is not free.
3917 // This is a ZExt, maybe this is free to extend from one type to another.
3918 // In that case, we would not account for a different use.
3921 if (ExtTy->getScalarType()->getIntegerBitWidth() >
3922 CurTy->getScalarType()->getIntegerBitWidth()) {
3930 if (!TLI.isZExtFree(NarrowTy, LargeTy))
3933 // All uses are the same or can be derived from one another for free.
3937 /// \brief Try to form ExtLd by promoting \p Exts until they reach a
3938 /// load instruction.
3939 /// If an ext(load) can be formed, it is returned via \p LI for the load
3940 /// and \p Inst for the extension.
3941 /// Otherwise LI == nullptr and Inst == nullptr.
3942 /// When some promotion happened, \p TPT contains the proper state to
3945 /// \return true when promoting was necessary to expose the ext(load)
3946 /// opportunity, false otherwise.
3950 /// %ld = load i32* %addr
3951 /// %add = add nuw i32 %ld, 4
3952 /// %zext = zext i32 %add to i64
3956 /// %ld = load i32* %addr
3957 /// %zext = zext i32 %ld to i64
3958 /// %add = add nuw i64 %zext, 4
3960 /// Thanks to the promotion, we can match zext(load i32*) to i64.
3961 bool CodeGenPrepare::extLdPromotion(TypePromotionTransaction &TPT,
3962 LoadInst *&LI, Instruction *&Inst,
3963 const SmallVectorImpl<Instruction *> &Exts,
3964 unsigned CreatedInstsCost = 0) {
3965 // Iterate over all the extensions to see if one form an ext(load).
3966 for (auto I : Exts) {
3967 // Check if we directly have ext(load).
3968 if ((LI = dyn_cast<LoadInst>(I->getOperand(0)))) {
3970 // No promotion happened here.
3973 // Check whether or not we want to do any promotion.
3974 if (!TLI || !TLI->enableExtLdPromotion() || DisableExtLdPromotion)
3976 // Get the action to perform the promotion.
3977 TypePromotionHelper::Action TPH = TypePromotionHelper::getAction(
3978 I, InsertedInsts, *TLI, PromotedInsts);
3979 // Check if we can promote.
3982 // Save the current state.
3983 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3984 TPT.getRestorationPoint();
3985 SmallVector<Instruction *, 4> NewExts;
3986 unsigned NewCreatedInstsCost = 0;
3987 unsigned ExtCost = !TLI->isExtFree(I);
3989 Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost,
3990 &NewExts, nullptr, *TLI);
3991 assert(PromotedVal &&
3992 "TypePromotionHelper should have filtered out those cases");
3994 // We would be able to merge only one extension in a load.
3995 // Therefore, if we have more than 1 new extension we heuristically
3996 // cut this search path, because it means we degrade the code quality.
3997 // With exactly 2, the transformation is neutral, because we will merge
3998 // one extension but leave one. However, we optimistically keep going,
3999 // because the new extension may be removed too.
4000 long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost;
4001 TotalCreatedInstsCost -= ExtCost;
4002 if (!StressExtLdPromotion &&
4003 (TotalCreatedInstsCost > 1 ||
4004 !isPromotedInstructionLegal(*TLI, *DL, PromotedVal))) {
4005 // The promotion is not profitable, rollback to the previous state.
4006 TPT.rollback(LastKnownGood);
4009 // The promotion is profitable.
4010 // Check if it exposes an ext(load).
4011 (void)extLdPromotion(TPT, LI, Inst, NewExts, TotalCreatedInstsCost);
4012 if (LI && (StressExtLdPromotion || NewCreatedInstsCost <= ExtCost ||
4013 // If we have created a new extension, i.e., now we have two
4014 // extensions. We must make sure one of them is merged with
4015 // the load, otherwise we may degrade the code quality.
4016 (LI->hasOneUse() || hasSameExtUse(LI, *TLI))))
4017 // Promotion happened.
4019 // If this does not help to expose an ext(load) then, rollback.
4020 TPT.rollback(LastKnownGood);
4022 // None of the extension can form an ext(load).
4028 /// Move a zext or sext fed by a load into the same basic block as the load,
4029 /// unless conditions are unfavorable. This allows SelectionDAG to fold the
4030 /// extend into the load.
4031 /// \p I[in/out] the extension may be modified during the process if some
4032 /// promotions apply.
4034 bool CodeGenPrepare::moveExtToFormExtLoad(Instruction *&I) {
4035 // Try to promote a chain of computation if it allows to form
4036 // an extended load.
4037 TypePromotionTransaction TPT;
4038 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4039 TPT.getRestorationPoint();
4040 SmallVector<Instruction *, 1> Exts;
4042 // Look for a load being extended.
4043 LoadInst *LI = nullptr;
4044 Instruction *OldExt = I;
4045 bool HasPromoted = extLdPromotion(TPT, LI, I, Exts);
4047 assert(!HasPromoted && !LI && "If we did not match any load instruction "
4048 "the code must remain the same");
4053 // If they're already in the same block, there's nothing to do.
4054 // Make the cheap checks first if we did not promote.
4055 // If we promoted, we need to check if it is indeed profitable.
4056 if (!HasPromoted && LI->getParent() == I->getParent())
4059 EVT VT = TLI->getValueType(*DL, I->getType());
4060 EVT LoadVT = TLI->getValueType(*DL, LI->getType());
4062 // If the load has other users and the truncate is not free, this probably
4063 // isn't worthwhile.
4064 if (!LI->hasOneUse() && TLI &&
4065 (TLI->isTypeLegal(LoadVT) || !TLI->isTypeLegal(VT)) &&
4066 !TLI->isTruncateFree(I->getType(), LI->getType())) {
4068 TPT.rollback(LastKnownGood);
4072 // Check whether the target supports casts folded into loads.
4074 if (isa<ZExtInst>(I))
4075 LType = ISD::ZEXTLOAD;
4077 assert(isa<SExtInst>(I) && "Unexpected ext type!");
4078 LType = ISD::SEXTLOAD;
4080 if (TLI && !TLI->isLoadExtLegal(LType, VT, LoadVT)) {
4082 TPT.rollback(LastKnownGood);
4086 // Move the extend into the same block as the load, so that SelectionDAG
4089 I->removeFromParent();
4095 bool CodeGenPrepare::optimizeExtUses(Instruction *I) {
4096 BasicBlock *DefBB = I->getParent();
4098 // If the result of a {s|z}ext and its source are both live out, rewrite all
4099 // other uses of the source with result of extension.
4100 Value *Src = I->getOperand(0);
4101 if (Src->hasOneUse())
4104 // Only do this xform if truncating is free.
4105 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
4108 // Only safe to perform the optimization if the source is also defined in
4110 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
4113 bool DefIsLiveOut = false;
4114 for (User *U : I->users()) {
4115 Instruction *UI = cast<Instruction>(U);
4117 // Figure out which BB this ext is used in.
4118 BasicBlock *UserBB = UI->getParent();
4119 if (UserBB == DefBB) continue;
4120 DefIsLiveOut = true;
4126 // Make sure none of the uses are PHI nodes.
4127 for (User *U : Src->users()) {
4128 Instruction *UI = cast<Instruction>(U);
4129 BasicBlock *UserBB = UI->getParent();
4130 if (UserBB == DefBB) continue;
4131 // Be conservative. We don't want this xform to end up introducing
4132 // reloads just before load / store instructions.
4133 if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
4137 // InsertedTruncs - Only insert one trunc in each block once.
4138 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
4140 bool MadeChange = false;
4141 for (Use &U : Src->uses()) {
4142 Instruction *User = cast<Instruction>(U.getUser());
4144 // Figure out which BB this ext is used in.
4145 BasicBlock *UserBB = User->getParent();
4146 if (UserBB == DefBB) continue;
4148 // Both src and def are live in this block. Rewrite the use.
4149 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
4151 if (!InsertedTrunc) {
4152 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
4153 assert(InsertPt != UserBB->end());
4154 InsertedTrunc = new TruncInst(I, Src->getType(), "", &*InsertPt);
4155 InsertedInsts.insert(InsertedTrunc);
4158 // Replace a use of the {s|z}ext source with a use of the result.
4167 /// Check if V (an operand of a select instruction) is an expensive instruction
4168 /// that is only used once.
4169 static bool sinkSelectOperand(const TargetTransformInfo *TTI, Value *V) {
4170 auto *I = dyn_cast<Instruction>(V);
4171 // If it's safe to speculatively execute, then it should not have side
4172 // effects; therefore, it's safe to sink and possibly *not* execute.
4173 return I && I->hasOneUse() && isSafeToSpeculativelyExecute(I) &&
4174 TTI->getUserCost(I) >= TargetTransformInfo::TCC_Expensive;
4177 /// Returns true if a SelectInst should be turned into an explicit branch.
4178 static bool isFormingBranchFromSelectProfitable(const TargetTransformInfo *TTI,
4180 // FIXME: This should use the same heuristics as IfConversion to determine
4181 // whether a select is better represented as a branch. This requires that
4182 // branch probability metadata is preserved for the select, which is not the
4185 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
4187 // If a branch is predictable, an out-of-order CPU can avoid blocking on its
4188 // comparison condition. If the compare has more than one use, there's
4189 // probably another cmov or setcc around, so it's not worth emitting a branch.
4190 if (!Cmp || !Cmp->hasOneUse())
4193 Value *CmpOp0 = Cmp->getOperand(0);
4194 Value *CmpOp1 = Cmp->getOperand(1);
4196 // Emit "cmov on compare with a memory operand" as a branch to avoid stalls
4197 // on a load from memory. But if the load is used more than once, do not
4198 // change the select to a branch because the load is probably needed
4199 // regardless of whether the branch is taken or not.
4200 if ((isa<LoadInst>(CmpOp0) && CmpOp0->hasOneUse()) ||
4201 (isa<LoadInst>(CmpOp1) && CmpOp1->hasOneUse()))
4204 // If either operand of the select is expensive and only needed on one side
4205 // of the select, we should form a branch.
4206 if (sinkSelectOperand(TTI, SI->getTrueValue()) ||
4207 sinkSelectOperand(TTI, SI->getFalseValue()))
4214 /// If we have a SelectInst that will likely profit from branch prediction,
4215 /// turn it into a branch.
4216 bool CodeGenPrepare::optimizeSelectInst(SelectInst *SI) {
4217 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
4219 // Can we convert the 'select' to CF ?
4220 if (DisableSelectToBranch || OptSize || !TLI || VectorCond)
4223 TargetLowering::SelectSupportKind SelectKind;
4225 SelectKind = TargetLowering::VectorMaskSelect;
4226 else if (SI->getType()->isVectorTy())
4227 SelectKind = TargetLowering::ScalarCondVectorVal;
4229 SelectKind = TargetLowering::ScalarValSelect;
4231 // Do we have efficient codegen support for this kind of 'selects' ?
4232 if (TLI->isSelectSupported(SelectKind)) {
4233 // We have efficient codegen support for the select instruction.
4234 // Check if it is profitable to keep this 'select'.
4235 if (!TLI->isPredictableSelectExpensive() ||
4236 !isFormingBranchFromSelectProfitable(TTI, SI))
4242 // Transform a sequence like this:
4244 // %cmp = cmp uge i32 %a, %b
4245 // %sel = select i1 %cmp, i32 %c, i32 %d
4249 // %cmp = cmp uge i32 %a, %b
4250 // br i1 %cmp, label %select.true, label %select.false
4252 // br label %select.end
4254 // br label %select.end
4256 // %sel = phi i32 [ %c, %select.true ], [ %d, %select.false ]
4258 // In addition, we may sink instructions that produce %c or %d from
4259 // the entry block into the destination(s) of the new branch.
4260 // If the true or false blocks do not contain a sunken instruction, that
4261 // block and its branch may be optimized away. In that case, one side of the
4262 // first branch will point directly to select.end, and the corresponding PHI
4263 // predecessor block will be the start block.
4265 // First, we split the block containing the select into 2 blocks.
4266 BasicBlock *StartBlock = SI->getParent();
4267 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(SI));
4268 BasicBlock *EndBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
4270 // Delete the unconditional branch that was just created by the split.
4271 StartBlock->getTerminator()->eraseFromParent();
4273 // These are the new basic blocks for the conditional branch.
4274 // At least one will become an actual new basic block.
4275 BasicBlock *TrueBlock = nullptr;
4276 BasicBlock *FalseBlock = nullptr;
4278 // Sink expensive instructions into the conditional blocks to avoid executing
4279 // them speculatively.
4280 if (sinkSelectOperand(TTI, SI->getTrueValue())) {
4281 TrueBlock = BasicBlock::Create(SI->getContext(), "select.true.sink",
4282 EndBlock->getParent(), EndBlock);
4283 auto *TrueBranch = BranchInst::Create(EndBlock, TrueBlock);
4284 auto *TrueInst = cast<Instruction>(SI->getTrueValue());
4285 TrueInst->moveBefore(TrueBranch);
4287 if (sinkSelectOperand(TTI, SI->getFalseValue())) {
4288 FalseBlock = BasicBlock::Create(SI->getContext(), "select.false.sink",
4289 EndBlock->getParent(), EndBlock);
4290 auto *FalseBranch = BranchInst::Create(EndBlock, FalseBlock);
4291 auto *FalseInst = cast<Instruction>(SI->getFalseValue());
4292 FalseInst->moveBefore(FalseBranch);
4295 // If there was nothing to sink, then arbitrarily choose the 'false' side
4296 // for a new input value to the PHI.
4297 if (TrueBlock == FalseBlock) {
4298 assert(TrueBlock == nullptr &&
4299 "Unexpected basic block transform while optimizing select");
4301 FalseBlock = BasicBlock::Create(SI->getContext(), "select.false",
4302 EndBlock->getParent(), EndBlock);
4303 BranchInst::Create(EndBlock, FalseBlock);
4306 // Insert the real conditional branch based on the original condition.
4307 // If we did not create a new block for one of the 'true' or 'false' paths
4308 // of the condition, it means that side of the branch goes to the end block
4309 // directly and the path originates from the start block from the point of
4310 // view of the new PHI.
4311 if (TrueBlock == nullptr) {
4312 BranchInst::Create(EndBlock, FalseBlock, SI->getCondition(), SI);
4313 TrueBlock = StartBlock;
4314 } else if (FalseBlock == nullptr) {
4315 BranchInst::Create(TrueBlock, EndBlock, SI->getCondition(), SI);
4316 FalseBlock = StartBlock;
4318 BranchInst::Create(TrueBlock, FalseBlock, SI->getCondition(), SI);
4321 // The select itself is replaced with a PHI Node.
4322 PHINode *PN = PHINode::Create(SI->getType(), 2, "", &EndBlock->front());
4324 PN->addIncoming(SI->getTrueValue(), TrueBlock);
4325 PN->addIncoming(SI->getFalseValue(), FalseBlock);
4327 SI->replaceAllUsesWith(PN);
4328 SI->eraseFromParent();
4330 // Instruct OptimizeBlock to skip to the next block.
4331 CurInstIterator = StartBlock->end();
4332 ++NumSelectsExpanded;
4336 static bool isBroadcastShuffle(ShuffleVectorInst *SVI) {
4337 SmallVector<int, 16> Mask(SVI->getShuffleMask());
4339 for (unsigned i = 0; i < Mask.size(); ++i) {
4340 if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
4342 SplatElem = Mask[i];
4348 /// Some targets have expensive vector shifts if the lanes aren't all the same
4349 /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
4350 /// it's often worth sinking a shufflevector splat down to its use so that
4351 /// codegen can spot all lanes are identical.
4352 bool CodeGenPrepare::optimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
4353 BasicBlock *DefBB = SVI->getParent();
4355 // Only do this xform if variable vector shifts are particularly expensive.
4356 if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
4359 // We only expect better codegen by sinking a shuffle if we can recognise a
4361 if (!isBroadcastShuffle(SVI))
4364 // InsertedShuffles - Only insert a shuffle in each block once.
4365 DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
4367 bool MadeChange = false;
4368 for (User *U : SVI->users()) {
4369 Instruction *UI = cast<Instruction>(U);
4371 // Figure out which BB this ext is used in.
4372 BasicBlock *UserBB = UI->getParent();
4373 if (UserBB == DefBB) continue;
4375 // For now only apply this when the splat is used by a shift instruction.
4376 if (!UI->isShift()) continue;
4378 // Everything checks out, sink the shuffle if the user's block doesn't
4379 // already have a copy.
4380 Instruction *&InsertedShuffle = InsertedShuffles[UserBB];
4382 if (!InsertedShuffle) {
4383 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
4384 assert(InsertPt != UserBB->end());
4386 new ShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1),
4387 SVI->getOperand(2), "", &*InsertPt);
4390 UI->replaceUsesOfWith(SVI, InsertedShuffle);
4394 // If we removed all uses, nuke the shuffle.
4395 if (SVI->use_empty()) {
4396 SVI->eraseFromParent();
4403 bool CodeGenPrepare::optimizeSwitchInst(SwitchInst *SI) {
4407 Value *Cond = SI->getCondition();
4408 Type *OldType = Cond->getType();
4409 LLVMContext &Context = Cond->getContext();
4410 MVT RegType = TLI->getRegisterType(Context, TLI->getValueType(*DL, OldType));
4411 unsigned RegWidth = RegType.getSizeInBits();
4413 if (RegWidth <= cast<IntegerType>(OldType)->getBitWidth())
4416 // If the register width is greater than the type width, expand the condition
4417 // of the switch instruction and each case constant to the width of the
4418 // register. By widening the type of the switch condition, subsequent
4419 // comparisons (for case comparisons) will not need to be extended to the
4420 // preferred register width, so we will potentially eliminate N-1 extends,
4421 // where N is the number of cases in the switch.
4422 auto *NewType = Type::getIntNTy(Context, RegWidth);
4424 // Zero-extend the switch condition and case constants unless the switch
4425 // condition is a function argument that is already being sign-extended.
4426 // In that case, we can avoid an unnecessary mask/extension by sign-extending
4427 // everything instead.
4428 Instruction::CastOps ExtType = Instruction::ZExt;
4429 if (auto *Arg = dyn_cast<Argument>(Cond))
4430 if (Arg->hasSExtAttr())
4431 ExtType = Instruction::SExt;
4433 auto *ExtInst = CastInst::Create(ExtType, Cond, NewType);
4434 ExtInst->insertBefore(SI);
4435 SI->setCondition(ExtInst);
4436 for (SwitchInst::CaseIt Case : SI->cases()) {
4437 APInt NarrowConst = Case.getCaseValue()->getValue();
4438 APInt WideConst = (ExtType == Instruction::ZExt) ?
4439 NarrowConst.zext(RegWidth) : NarrowConst.sext(RegWidth);
4440 Case.setValue(ConstantInt::get(Context, WideConst));
4447 /// \brief Helper class to promote a scalar operation to a vector one.
4448 /// This class is used to move downward extractelement transition.
4450 /// a = vector_op <2 x i32>
4451 /// b = extractelement <2 x i32> a, i32 0
4456 /// a = vector_op <2 x i32>
4457 /// c = vector_op a (equivalent to scalar_op on the related lane)
4458 /// * d = extractelement <2 x i32> c, i32 0
4460 /// Assuming both extractelement and store can be combine, we get rid of the
4462 class VectorPromoteHelper {
4463 /// DataLayout associated with the current module.
4464 const DataLayout &DL;
4466 /// Used to perform some checks on the legality of vector operations.
4467 const TargetLowering &TLI;
4469 /// Used to estimated the cost of the promoted chain.
4470 const TargetTransformInfo &TTI;
4472 /// The transition being moved downwards.
4473 Instruction *Transition;
4474 /// The sequence of instructions to be promoted.
4475 SmallVector<Instruction *, 4> InstsToBePromoted;
4476 /// Cost of combining a store and an extract.
4477 unsigned StoreExtractCombineCost;
4478 /// Instruction that will be combined with the transition.
4479 Instruction *CombineInst;
4481 /// \brief The instruction that represents the current end of the transition.
4482 /// Since we are faking the promotion until we reach the end of the chain
4483 /// of computation, we need a way to get the current end of the transition.
4484 Instruction *getEndOfTransition() const {
4485 if (InstsToBePromoted.empty())
4487 return InstsToBePromoted.back();
4490 /// \brief Return the index of the original value in the transition.
4491 /// E.g., for "extractelement <2 x i32> c, i32 1" the original value,
4492 /// c, is at index 0.
4493 unsigned getTransitionOriginalValueIdx() const {
4494 assert(isa<ExtractElementInst>(Transition) &&
4495 "Other kind of transitions are not supported yet");
4499 /// \brief Return the index of the index in the transition.
4500 /// E.g., for "extractelement <2 x i32> c, i32 0" the index
4502 unsigned getTransitionIdx() const {
4503 assert(isa<ExtractElementInst>(Transition) &&
4504 "Other kind of transitions are not supported yet");
4508 /// \brief Get the type of the transition.
4509 /// This is the type of the original value.
4510 /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the
4511 /// transition is <2 x i32>.
4512 Type *getTransitionType() const {
4513 return Transition->getOperand(getTransitionOriginalValueIdx())->getType();
4516 /// \brief Promote \p ToBePromoted by moving \p Def downward through.
4517 /// I.e., we have the following sequence:
4518 /// Def = Transition <ty1> a to <ty2>
4519 /// b = ToBePromoted <ty2> Def, ...
4521 /// b = ToBePromoted <ty1> a, ...
4522 /// Def = Transition <ty1> ToBePromoted to <ty2>
4523 void promoteImpl(Instruction *ToBePromoted);
4525 /// \brief Check whether or not it is profitable to promote all the
4526 /// instructions enqueued to be promoted.
4527 bool isProfitableToPromote() {
4528 Value *ValIdx = Transition->getOperand(getTransitionOriginalValueIdx());
4529 unsigned Index = isa<ConstantInt>(ValIdx)
4530 ? cast<ConstantInt>(ValIdx)->getZExtValue()
4532 Type *PromotedType = getTransitionType();
4534 StoreInst *ST = cast<StoreInst>(CombineInst);
4535 unsigned AS = ST->getPointerAddressSpace();
4536 unsigned Align = ST->getAlignment();
4537 // Check if this store is supported.
4538 if (!TLI.allowsMisalignedMemoryAccesses(
4539 TLI.getValueType(DL, ST->getValueOperand()->getType()), AS,
4541 // If this is not supported, there is no way we can combine
4542 // the extract with the store.
4546 // The scalar chain of computation has to pay for the transition
4547 // scalar to vector.
4548 // The vector chain has to account for the combining cost.
4549 uint64_t ScalarCost =
4550 TTI.getVectorInstrCost(Transition->getOpcode(), PromotedType, Index);
4551 uint64_t VectorCost = StoreExtractCombineCost;
4552 for (const auto &Inst : InstsToBePromoted) {
4553 // Compute the cost.
4554 // By construction, all instructions being promoted are arithmetic ones.
4555 // Moreover, one argument is a constant that can be viewed as a splat
4557 Value *Arg0 = Inst->getOperand(0);
4558 bool IsArg0Constant = isa<UndefValue>(Arg0) || isa<ConstantInt>(Arg0) ||
4559 isa<ConstantFP>(Arg0);
4560 TargetTransformInfo::OperandValueKind Arg0OVK =
4561 IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
4562 : TargetTransformInfo::OK_AnyValue;
4563 TargetTransformInfo::OperandValueKind Arg1OVK =
4564 !IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
4565 : TargetTransformInfo::OK_AnyValue;
4566 ScalarCost += TTI.getArithmeticInstrCost(
4567 Inst->getOpcode(), Inst->getType(), Arg0OVK, Arg1OVK);
4568 VectorCost += TTI.getArithmeticInstrCost(Inst->getOpcode(), PromotedType,
4571 DEBUG(dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
4572 << ScalarCost << "\nVector: " << VectorCost << '\n');
4573 return ScalarCost > VectorCost;
4576 /// \brief Generate a constant vector with \p Val with the same
4577 /// number of elements as the transition.
4578 /// \p UseSplat defines whether or not \p Val should be replicated
4579 /// across the whole vector.
4580 /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>,
4581 /// otherwise we generate a vector with as many undef as possible:
4582 /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only
4583 /// used at the index of the extract.
4584 Value *getConstantVector(Constant *Val, bool UseSplat) const {
4585 unsigned ExtractIdx = UINT_MAX;
4587 // If we cannot determine where the constant must be, we have to
4588 // use a splat constant.
4589 Value *ValExtractIdx = Transition->getOperand(getTransitionIdx());
4590 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(ValExtractIdx))
4591 ExtractIdx = CstVal->getSExtValue();
4596 unsigned End = getTransitionType()->getVectorNumElements();
4598 return ConstantVector::getSplat(End, Val);
4600 SmallVector<Constant *, 4> ConstVec;
4601 UndefValue *UndefVal = UndefValue::get(Val->getType());
4602 for (unsigned Idx = 0; Idx != End; ++Idx) {
4603 if (Idx == ExtractIdx)
4604 ConstVec.push_back(Val);
4606 ConstVec.push_back(UndefVal);
4608 return ConstantVector::get(ConstVec);
4611 /// \brief Check if promoting to a vector type an operand at \p OperandIdx
4612 /// in \p Use can trigger undefined behavior.
4613 static bool canCauseUndefinedBehavior(const Instruction *Use,
4614 unsigned OperandIdx) {
4615 // This is not safe to introduce undef when the operand is on
4616 // the right hand side of a division-like instruction.
4617 if (OperandIdx != 1)
4619 switch (Use->getOpcode()) {
4622 case Instruction::SDiv:
4623 case Instruction::UDiv:
4624 case Instruction::SRem:
4625 case Instruction::URem:
4627 case Instruction::FDiv:
4628 case Instruction::FRem:
4629 return !Use->hasNoNaNs();
4631 llvm_unreachable(nullptr);
4635 VectorPromoteHelper(const DataLayout &DL, const TargetLowering &TLI,
4636 const TargetTransformInfo &TTI, Instruction *Transition,
4637 unsigned CombineCost)
4638 : DL(DL), TLI(TLI), TTI(TTI), Transition(Transition),
4639 StoreExtractCombineCost(CombineCost), CombineInst(nullptr) {
4640 assert(Transition && "Do not know how to promote null");
4643 /// \brief Check if we can promote \p ToBePromoted to \p Type.
4644 bool canPromote(const Instruction *ToBePromoted) const {
4645 // We could support CastInst too.
4646 return isa<BinaryOperator>(ToBePromoted);
4649 /// \brief Check if it is profitable to promote \p ToBePromoted
4650 /// by moving downward the transition through.
4651 bool shouldPromote(const Instruction *ToBePromoted) const {
4652 // Promote only if all the operands can be statically expanded.
4653 // Indeed, we do not want to introduce any new kind of transitions.
4654 for (const Use &U : ToBePromoted->operands()) {
4655 const Value *Val = U.get();
4656 if (Val == getEndOfTransition()) {
4657 // If the use is a division and the transition is on the rhs,
4658 // we cannot promote the operation, otherwise we may create a
4659 // division by zero.
4660 if (canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()))
4664 if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) &&
4665 !isa<ConstantFP>(Val))
4668 // Check that the resulting operation is legal.
4669 int ISDOpcode = TLI.InstructionOpcodeToISD(ToBePromoted->getOpcode());
4672 return StressStoreExtract ||
4673 TLI.isOperationLegalOrCustom(
4674 ISDOpcode, TLI.getValueType(DL, getTransitionType(), true));
4677 /// \brief Check whether or not \p Use can be combined
4678 /// with the transition.
4679 /// I.e., is it possible to do Use(Transition) => AnotherUse?
4680 bool canCombine(const Instruction *Use) { return isa<StoreInst>(Use); }
4682 /// \brief Record \p ToBePromoted as part of the chain to be promoted.
4683 void enqueueForPromotion(Instruction *ToBePromoted) {
4684 InstsToBePromoted.push_back(ToBePromoted);
4687 /// \brief Set the instruction that will be combined with the transition.
4688 void recordCombineInstruction(Instruction *ToBeCombined) {
4689 assert(canCombine(ToBeCombined) && "Unsupported instruction to combine");
4690 CombineInst = ToBeCombined;
4693 /// \brief Promote all the instructions enqueued for promotion if it is
4695 /// \return True if the promotion happened, false otherwise.
4697 // Check if there is something to promote.
4698 // Right now, if we do not have anything to combine with,
4699 // we assume the promotion is not profitable.
4700 if (InstsToBePromoted.empty() || !CombineInst)
4704 if (!StressStoreExtract && !isProfitableToPromote())
4708 for (auto &ToBePromoted : InstsToBePromoted)
4709 promoteImpl(ToBePromoted);
4710 InstsToBePromoted.clear();
4714 } // End of anonymous namespace.
4716 void VectorPromoteHelper::promoteImpl(Instruction *ToBePromoted) {
4717 // At this point, we know that all the operands of ToBePromoted but Def
4718 // can be statically promoted.
4719 // For Def, we need to use its parameter in ToBePromoted:
4720 // b = ToBePromoted ty1 a
4721 // Def = Transition ty1 b to ty2
4722 // Move the transition down.
4723 // 1. Replace all uses of the promoted operation by the transition.
4724 // = ... b => = ... Def.
4725 assert(ToBePromoted->getType() == Transition->getType() &&
4726 "The type of the result of the transition does not match "
4728 ToBePromoted->replaceAllUsesWith(Transition);
4729 // 2. Update the type of the uses.
4730 // b = ToBePromoted ty2 Def => b = ToBePromoted ty1 Def.
4731 Type *TransitionTy = getTransitionType();
4732 ToBePromoted->mutateType(TransitionTy);
4733 // 3. Update all the operands of the promoted operation with promoted
4735 // b = ToBePromoted ty1 Def => b = ToBePromoted ty1 a.
4736 for (Use &U : ToBePromoted->operands()) {
4737 Value *Val = U.get();
4738 Value *NewVal = nullptr;
4739 if (Val == Transition)
4740 NewVal = Transition->getOperand(getTransitionOriginalValueIdx());
4741 else if (isa<UndefValue>(Val) || isa<ConstantInt>(Val) ||
4742 isa<ConstantFP>(Val)) {
4743 // Use a splat constant if it is not safe to use undef.
4744 NewVal = getConstantVector(
4745 cast<Constant>(Val),
4746 isa<UndefValue>(Val) ||
4747 canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()));
4749 llvm_unreachable("Did you modified shouldPromote and forgot to update "
4751 ToBePromoted->setOperand(U.getOperandNo(), NewVal);
4753 Transition->removeFromParent();
4754 Transition->insertAfter(ToBePromoted);
4755 Transition->setOperand(getTransitionOriginalValueIdx(), ToBePromoted);
4758 /// Some targets can do store(extractelement) with one instruction.
4759 /// Try to push the extractelement towards the stores when the target
4760 /// has this feature and this is profitable.
4761 bool CodeGenPrepare::optimizeExtractElementInst(Instruction *Inst) {
4762 unsigned CombineCost = UINT_MAX;
4763 if (DisableStoreExtract || !TLI ||
4764 (!StressStoreExtract &&
4765 !TLI->canCombineStoreAndExtract(Inst->getOperand(0)->getType(),
4766 Inst->getOperand(1), CombineCost)))
4769 // At this point we know that Inst is a vector to scalar transition.
4770 // Try to move it down the def-use chain, until:
4771 // - We can combine the transition with its single use
4772 // => we got rid of the transition.
4773 // - We escape the current basic block
4774 // => we would need to check that we are moving it at a cheaper place and
4775 // we do not do that for now.
4776 BasicBlock *Parent = Inst->getParent();
4777 DEBUG(dbgs() << "Found an interesting transition: " << *Inst << '\n');
4778 VectorPromoteHelper VPH(*DL, *TLI, *TTI, Inst, CombineCost);
4779 // If the transition has more than one use, assume this is not going to be
4781 while (Inst->hasOneUse()) {
4782 Instruction *ToBePromoted = cast<Instruction>(*Inst->user_begin());
4783 DEBUG(dbgs() << "Use: " << *ToBePromoted << '\n');
4785 if (ToBePromoted->getParent() != Parent) {
4786 DEBUG(dbgs() << "Instruction to promote is in a different block ("
4787 << ToBePromoted->getParent()->getName()
4788 << ") than the transition (" << Parent->getName() << ").\n");
4792 if (VPH.canCombine(ToBePromoted)) {
4793 DEBUG(dbgs() << "Assume " << *Inst << '\n'
4794 << "will be combined with: " << *ToBePromoted << '\n');
4795 VPH.recordCombineInstruction(ToBePromoted);
4796 bool Changed = VPH.promote();
4797 NumStoreExtractExposed += Changed;
4801 DEBUG(dbgs() << "Try promoting.\n");
4802 if (!VPH.canPromote(ToBePromoted) || !VPH.shouldPromote(ToBePromoted))
4805 DEBUG(dbgs() << "Promoting is possible... Enqueue for promotion!\n");
4807 VPH.enqueueForPromotion(ToBePromoted);
4808 Inst = ToBePromoted;
4813 bool CodeGenPrepare::optimizeInst(Instruction *I, bool& ModifiedDT) {
4814 // Bail out if we inserted the instruction to prevent optimizations from
4815 // stepping on each other's toes.
4816 if (InsertedInsts.count(I))
4819 if (PHINode *P = dyn_cast<PHINode>(I)) {
4820 // It is possible for very late stage optimizations (such as SimplifyCFG)
4821 // to introduce PHI nodes too late to be cleaned up. If we detect such a
4822 // trivial PHI, go ahead and zap it here.
4823 if (Value *V = SimplifyInstruction(P, *DL, TLInfo, nullptr)) {
4824 P->replaceAllUsesWith(V);
4825 P->eraseFromParent();
4832 if (CastInst *CI = dyn_cast<CastInst>(I)) {
4833 // If the source of the cast is a constant, then this should have
4834 // already been constant folded. The only reason NOT to constant fold
4835 // it is if something (e.g. LSR) was careful to place the constant
4836 // evaluation in a block other than then one that uses it (e.g. to hoist
4837 // the address of globals out of a loop). If this is the case, we don't
4838 // want to forward-subst the cast.
4839 if (isa<Constant>(CI->getOperand(0)))
4842 if (TLI && OptimizeNoopCopyExpression(CI, *TLI, *DL))
4845 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
4846 /// Sink a zext or sext into its user blocks if the target type doesn't
4847 /// fit in one register
4849 TLI->getTypeAction(CI->getContext(),
4850 TLI->getValueType(*DL, CI->getType())) ==
4851 TargetLowering::TypeExpandInteger) {
4852 return SinkCast(CI);
4854 bool MadeChange = moveExtToFormExtLoad(I);
4855 return MadeChange | optimizeExtUses(I);
4861 if (CmpInst *CI = dyn_cast<CmpInst>(I))
4862 if (!TLI || !TLI->hasMultipleConditionRegisters())
4863 return OptimizeCmpExpression(CI);
4865 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
4866 stripInvariantGroupMetadata(*LI);
4868 unsigned AS = LI->getPointerAddressSpace();
4869 return optimizeMemoryInst(I, I->getOperand(0), LI->getType(), AS);
4874 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
4875 stripInvariantGroupMetadata(*SI);
4877 unsigned AS = SI->getPointerAddressSpace();
4878 return optimizeMemoryInst(I, SI->getOperand(1),
4879 SI->getOperand(0)->getType(), AS);
4884 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(I);
4886 if (BinOp && (BinOp->getOpcode() == Instruction::AShr ||
4887 BinOp->getOpcode() == Instruction::LShr)) {
4888 ConstantInt *CI = dyn_cast<ConstantInt>(BinOp->getOperand(1));
4889 if (TLI && CI && TLI->hasExtractBitsInsn())
4890 return OptimizeExtractBits(BinOp, CI, *TLI, *DL);
4895 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
4896 if (GEPI->hasAllZeroIndices()) {
4897 /// The GEP operand must be a pointer, so must its result -> BitCast
4898 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
4899 GEPI->getName(), GEPI);
4900 GEPI->replaceAllUsesWith(NC);
4901 GEPI->eraseFromParent();
4903 optimizeInst(NC, ModifiedDT);
4909 if (CallInst *CI = dyn_cast<CallInst>(I))
4910 return optimizeCallInst(CI, ModifiedDT);
4912 if (SelectInst *SI = dyn_cast<SelectInst>(I))
4913 return optimizeSelectInst(SI);
4915 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I))
4916 return optimizeShuffleVectorInst(SVI);
4918 if (auto *Switch = dyn_cast<SwitchInst>(I))
4919 return optimizeSwitchInst(Switch);
4921 if (isa<ExtractElementInst>(I))
4922 return optimizeExtractElementInst(I);
4927 // In this pass we look for GEP and cast instructions that are used
4928 // across basic blocks and rewrite them to improve basic-block-at-a-time
4930 bool CodeGenPrepare::optimizeBlock(BasicBlock &BB, bool& ModifiedDT) {
4932 bool MadeChange = false;
4934 CurInstIterator = BB.begin();
4935 while (CurInstIterator != BB.end()) {
4936 MadeChange |= optimizeInst(&*CurInstIterator++, ModifiedDT);
4940 MadeChange |= dupRetToEnableTailCallOpts(&BB);
4945 // llvm.dbg.value is far away from the value then iSel may not be able
4946 // handle it properly. iSel will drop llvm.dbg.value if it can not
4947 // find a node corresponding to the value.
4948 bool CodeGenPrepare::placeDbgValues(Function &F) {
4949 bool MadeChange = false;
4950 for (BasicBlock &BB : F) {
4951 Instruction *PrevNonDbgInst = nullptr;
4952 for (BasicBlock::iterator BI = BB.begin(), BE = BB.end(); BI != BE;) {
4953 Instruction *Insn = &*BI++;
4954 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
4955 // Leave dbg.values that refer to an alloca alone. These
4956 // instrinsics describe the address of a variable (= the alloca)
4957 // being taken. They should not be moved next to the alloca
4958 // (and to the beginning of the scope), but rather stay close to
4959 // where said address is used.
4960 if (!DVI || (DVI->getValue() && isa<AllocaInst>(DVI->getValue()))) {
4961 PrevNonDbgInst = Insn;
4965 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
4966 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
4967 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
4968 DVI->removeFromParent();
4969 if (isa<PHINode>(VI))
4970 DVI->insertBefore(&*VI->getParent()->getFirstInsertionPt());
4972 DVI->insertAfter(VI);
4981 // If there is a sequence that branches based on comparing a single bit
4982 // against zero that can be combined into a single instruction, and the
4983 // target supports folding these into a single instruction, sink the
4984 // mask and compare into the branch uses. Do this before OptimizeBlock ->
4985 // OptimizeInst -> OptimizeCmpExpression, which perturbs the pattern being
4987 bool CodeGenPrepare::sinkAndCmp(Function &F) {
4988 if (!EnableAndCmpSinking)
4990 if (!TLI || !TLI->isMaskAndBranchFoldingLegal())
4992 bool MadeChange = false;
4993 for (Function::iterator I = F.begin(), E = F.end(); I != E; ) {
4994 BasicBlock *BB = &*I++;
4996 // Does this BB end with the following?
4997 // %andVal = and %val, #single-bit-set
4998 // %icmpVal = icmp %andResult, 0
4999 // br i1 %cmpVal label %dest1, label %dest2"
5000 BranchInst *Brcc = dyn_cast<BranchInst>(BB->getTerminator());
5001 if (!Brcc || !Brcc->isConditional())
5003 ICmpInst *Cmp = dyn_cast<ICmpInst>(Brcc->getOperand(0));
5004 if (!Cmp || Cmp->getParent() != BB)
5006 ConstantInt *Zero = dyn_cast<ConstantInt>(Cmp->getOperand(1));
5007 if (!Zero || !Zero->isZero())
5009 Instruction *And = dyn_cast<Instruction>(Cmp->getOperand(0));
5010 if (!And || And->getOpcode() != Instruction::And || And->getParent() != BB)
5012 ConstantInt* Mask = dyn_cast<ConstantInt>(And->getOperand(1));
5013 if (!Mask || !Mask->getUniqueInteger().isPowerOf2())
5015 DEBUG(dbgs() << "found and; icmp ?,0; brcc\n"); DEBUG(BB->dump());
5017 // Push the "and; icmp" for any users that are conditional branches.
5018 // Since there can only be one branch use per BB, we don't need to keep
5019 // track of which BBs we insert into.
5020 for (Value::use_iterator UI = Cmp->use_begin(), E = Cmp->use_end();
5024 BranchInst *BrccUser = dyn_cast<BranchInst>(*UI);
5026 if (!BrccUser || !BrccUser->isConditional())
5028 BasicBlock *UserBB = BrccUser->getParent();
5029 if (UserBB == BB) continue;
5030 DEBUG(dbgs() << "found Brcc use\n");
5032 // Sink the "and; icmp" to use.
5034 BinaryOperator *NewAnd =
5035 BinaryOperator::CreateAnd(And->getOperand(0), And->getOperand(1), "",
5038 CmpInst::Create(Cmp->getOpcode(), Cmp->getPredicate(), NewAnd, Zero,
5042 DEBUG(BrccUser->getParent()->dump());
5048 /// \brief Retrieve the probabilities of a conditional branch. Returns true on
5049 /// success, or returns false if no or invalid metadata was found.
5050 static bool extractBranchMetadata(BranchInst *BI,
5051 uint64_t &ProbTrue, uint64_t &ProbFalse) {
5052 assert(BI->isConditional() &&
5053 "Looking for probabilities on unconditional branch?");
5054 auto *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
5055 if (!ProfileData || ProfileData->getNumOperands() != 3)
5058 const auto *CITrue =
5059 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(1));
5060 const auto *CIFalse =
5061 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(2));
5062 if (!CITrue || !CIFalse)
5065 ProbTrue = CITrue->getValue().getZExtValue();
5066 ProbFalse = CIFalse->getValue().getZExtValue();
5071 /// \brief Scale down both weights to fit into uint32_t.
5072 static void scaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) {
5073 uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse;
5074 uint32_t Scale = (NewMax / UINT32_MAX) + 1;
5075 NewTrue = NewTrue / Scale;
5076 NewFalse = NewFalse / Scale;
5079 /// \brief Some targets prefer to split a conditional branch like:
5081 /// %0 = icmp ne i32 %a, 0
5082 /// %1 = icmp ne i32 %b, 0
5083 /// %or.cond = or i1 %0, %1
5084 /// br i1 %or.cond, label %TrueBB, label %FalseBB
5086 /// into multiple branch instructions like:
5089 /// %0 = icmp ne i32 %a, 0
5090 /// br i1 %0, label %TrueBB, label %bb2
5092 /// %1 = icmp ne i32 %b, 0
5093 /// br i1 %1, label %TrueBB, label %FalseBB
5095 /// This usually allows instruction selection to do even further optimizations
5096 /// and combine the compare with the branch instruction. Currently this is
5097 /// applied for targets which have "cheap" jump instructions.
5099 /// FIXME: Remove the (equivalent?) implementation in SelectionDAG.
5101 bool CodeGenPrepare::splitBranchCondition(Function &F) {
5102 if (!TM || !TM->Options.EnableFastISel || !TLI || TLI->isJumpExpensive())
5105 bool MadeChange = false;
5106 for (auto &BB : F) {
5107 // Does this BB end with the following?
5108 // %cond1 = icmp|fcmp|binary instruction ...
5109 // %cond2 = icmp|fcmp|binary instruction ...
5110 // %cond.or = or|and i1 %cond1, cond2
5111 // br i1 %cond.or label %dest1, label %dest2"
5112 BinaryOperator *LogicOp;
5113 BasicBlock *TBB, *FBB;
5114 if (!match(BB.getTerminator(), m_Br(m_OneUse(m_BinOp(LogicOp)), TBB, FBB)))
5117 auto *Br1 = cast<BranchInst>(BB.getTerminator());
5118 if (Br1->getMetadata(LLVMContext::MD_unpredictable))
5122 Value *Cond1, *Cond2;
5123 if (match(LogicOp, m_And(m_OneUse(m_Value(Cond1)),
5124 m_OneUse(m_Value(Cond2)))))
5125 Opc = Instruction::And;
5126 else if (match(LogicOp, m_Or(m_OneUse(m_Value(Cond1)),
5127 m_OneUse(m_Value(Cond2)))))
5128 Opc = Instruction::Or;
5132 if (!match(Cond1, m_CombineOr(m_Cmp(), m_BinOp())) ||
5133 !match(Cond2, m_CombineOr(m_Cmp(), m_BinOp())) )
5136 DEBUG(dbgs() << "Before branch condition splitting\n"; BB.dump());
5139 auto *InsertBefore = std::next(Function::iterator(BB))
5140 .getNodePtrUnchecked();
5141 auto TmpBB = BasicBlock::Create(BB.getContext(),
5142 BB.getName() + ".cond.split",
5143 BB.getParent(), InsertBefore);
5145 // Update original basic block by using the first condition directly by the
5146 // branch instruction and removing the no longer needed and/or instruction.
5147 Br1->setCondition(Cond1);
5148 LogicOp->eraseFromParent();
5150 // Depending on the conditon we have to either replace the true or the false
5151 // successor of the original branch instruction.
5152 if (Opc == Instruction::And)
5153 Br1->setSuccessor(0, TmpBB);
5155 Br1->setSuccessor(1, TmpBB);
5157 // Fill in the new basic block.
5158 auto *Br2 = IRBuilder<>(TmpBB).CreateCondBr(Cond2, TBB, FBB);
5159 if (auto *I = dyn_cast<Instruction>(Cond2)) {
5160 I->removeFromParent();
5161 I->insertBefore(Br2);
5164 // Update PHI nodes in both successors. The original BB needs to be
5165 // replaced in one succesor's PHI nodes, because the branch comes now from
5166 // the newly generated BB (NewBB). In the other successor we need to add one
5167 // incoming edge to the PHI nodes, because both branch instructions target
5168 // now the same successor. Depending on the original branch condition
5169 // (and/or) we have to swap the successors (TrueDest, FalseDest), so that
5170 // we perfrom the correct update for the PHI nodes.
5171 // This doesn't change the successor order of the just created branch
5172 // instruction (or any other instruction).
5173 if (Opc == Instruction::Or)
5174 std::swap(TBB, FBB);
5176 // Replace the old BB with the new BB.
5177 for (auto &I : *TBB) {
5178 PHINode *PN = dyn_cast<PHINode>(&I);
5182 while ((i = PN->getBasicBlockIndex(&BB)) >= 0)
5183 PN->setIncomingBlock(i, TmpBB);
5186 // Add another incoming edge form the new BB.
5187 for (auto &I : *FBB) {
5188 PHINode *PN = dyn_cast<PHINode>(&I);
5191 auto *Val = PN->getIncomingValueForBlock(&BB);
5192 PN->addIncoming(Val, TmpBB);
5195 // Update the branch weights (from SelectionDAGBuilder::
5196 // FindMergedConditions).
5197 if (Opc == Instruction::Or) {
5198 // Codegen X | Y as:
5207 // We have flexibility in setting Prob for BB1 and Prob for NewBB.
5208 // The requirement is that
5209 // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB)
5210 // = TrueProb for orignal BB.
5211 // Assuming the orignal weights are A and B, one choice is to set BB1's
5212 // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice
5214 // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB.
5215 // Another choice is to assume TrueProb for BB1 equals to TrueProb for
5216 // TmpBB, but the math is more complicated.
5217 uint64_t TrueWeight, FalseWeight;
5218 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
5219 uint64_t NewTrueWeight = TrueWeight;
5220 uint64_t NewFalseWeight = TrueWeight + 2 * FalseWeight;
5221 scaleWeights(NewTrueWeight, NewFalseWeight);
5222 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
5223 .createBranchWeights(TrueWeight, FalseWeight));
5225 NewTrueWeight = TrueWeight;
5226 NewFalseWeight = 2 * FalseWeight;
5227 scaleWeights(NewTrueWeight, NewFalseWeight);
5228 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
5229 .createBranchWeights(TrueWeight, FalseWeight));
5232 // Codegen X & Y as:
5240 // This requires creation of TmpBB after CurBB.
5242 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
5243 // The requirement is that
5244 // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB)
5245 // = FalseProb for orignal BB.
5246 // Assuming the orignal weights are A and B, one choice is to set BB1's
5247 // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice
5249 // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB.
5250 uint64_t TrueWeight, FalseWeight;
5251 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
5252 uint64_t NewTrueWeight = 2 * TrueWeight + FalseWeight;
5253 uint64_t NewFalseWeight = FalseWeight;
5254 scaleWeights(NewTrueWeight, NewFalseWeight);
5255 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
5256 .createBranchWeights(TrueWeight, FalseWeight));
5258 NewTrueWeight = 2 * TrueWeight;
5259 NewFalseWeight = FalseWeight;
5260 scaleWeights(NewTrueWeight, NewFalseWeight);
5261 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
5262 .createBranchWeights(TrueWeight, FalseWeight));
5266 // Note: No point in getting fancy here, since the DT info is never
5267 // available to CodeGenPrepare.
5272 DEBUG(dbgs() << "After branch condition splitting\n"; BB.dump();
5278 void CodeGenPrepare::stripInvariantGroupMetadata(Instruction &I) {
5279 if (auto *InvariantMD = I.getMetadata(LLVMContext::MD_invariant_group))
5280 I.dropUnknownNonDebugMetadata(InvariantMD->getMetadataID());