1 //===- CodeGenPrepare.cpp - Prepare a function for code generation --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass munges the code in the input function to better prepare it for
11 // SelectionDAG-based code generation. This works around limitations in it's
12 // basic-block-at-a-time approach. It should eventually be removed.
14 //===----------------------------------------------------------------------===//
16 #include "llvm/CodeGen/Passes.h"
17 #include "llvm/ADT/DenseMap.h"
18 #include "llvm/ADT/SmallSet.h"
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/Analysis/InstructionSimplify.h"
21 #include "llvm/Analysis/TargetLibraryInfo.h"
22 #include "llvm/Analysis/TargetTransformInfo.h"
23 #include "llvm/IR/CallSite.h"
24 #include "llvm/IR/Constants.h"
25 #include "llvm/IR/DataLayout.h"
26 #include "llvm/IR/DerivedTypes.h"
27 #include "llvm/IR/Dominators.h"
28 #include "llvm/IR/Function.h"
29 #include "llvm/IR/GetElementPtrTypeIterator.h"
30 #include "llvm/IR/IRBuilder.h"
31 #include "llvm/IR/InlineAsm.h"
32 #include "llvm/IR/Instructions.h"
33 #include "llvm/IR/IntrinsicInst.h"
34 #include "llvm/IR/MDBuilder.h"
35 #include "llvm/IR/PatternMatch.h"
36 #include "llvm/IR/Statepoint.h"
37 #include "llvm/IR/ValueHandle.h"
38 #include "llvm/IR/ValueMap.h"
39 #include "llvm/Pass.h"
40 #include "llvm/Support/CommandLine.h"
41 #include "llvm/Support/Debug.h"
42 #include "llvm/Support/raw_ostream.h"
43 #include "llvm/Target/TargetLowering.h"
44 #include "llvm/Target/TargetSubtargetInfo.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/BuildLibCalls.h"
47 #include "llvm/Transforms/Utils/BypassSlowDivision.h"
48 #include "llvm/Transforms/Utils/Local.h"
49 #include "llvm/Transforms/Utils/SimplifyLibCalls.h"
51 using namespace llvm::PatternMatch;
53 #define DEBUG_TYPE "codegenprepare"
55 STATISTIC(NumBlocksElim, "Number of blocks eliminated");
56 STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated");
57 STATISTIC(NumGEPsElim, "Number of GEPs converted to casts");
58 STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "
60 STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "
62 STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "
63 "computations were sunk");
64 STATISTIC(NumExtsMoved, "Number of [s|z]ext instructions combined with loads");
65 STATISTIC(NumExtUses, "Number of uses of [s|z]ext instructions optimized");
66 STATISTIC(NumRetsDup, "Number of return instructions duplicated");
67 STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved");
68 STATISTIC(NumSelectsExpanded, "Number of selects turned into branches");
69 STATISTIC(NumAndCmpsMoved, "Number of and/cmp's pushed into branches");
70 STATISTIC(NumStoreExtractExposed, "Number of store(extractelement) exposed");
72 static cl::opt<bool> DisableBranchOpts(
73 "disable-cgp-branch-opts", cl::Hidden, cl::init(false),
74 cl::desc("Disable branch optimizations in CodeGenPrepare"));
77 DisableGCOpts("disable-cgp-gc-opts", cl::Hidden, cl::init(false),
78 cl::desc("Disable GC optimizations in CodeGenPrepare"));
80 static cl::opt<bool> DisableSelectToBranch(
81 "disable-cgp-select2branch", cl::Hidden, cl::init(false),
82 cl::desc("Disable select to branch conversion."));
84 static cl::opt<bool> AddrSinkUsingGEPs(
85 "addr-sink-using-gep", cl::Hidden, cl::init(false),
86 cl::desc("Address sinking in CGP using GEPs."));
88 static cl::opt<bool> EnableAndCmpSinking(
89 "enable-andcmp-sinking", cl::Hidden, cl::init(true),
90 cl::desc("Enable sinkinig and/cmp into branches."));
92 static cl::opt<bool> DisableStoreExtract(
93 "disable-cgp-store-extract", cl::Hidden, cl::init(false),
94 cl::desc("Disable store(extract) optimizations in CodeGenPrepare"));
96 static cl::opt<bool> StressStoreExtract(
97 "stress-cgp-store-extract", cl::Hidden, cl::init(false),
98 cl::desc("Stress test store(extract) optimizations in CodeGenPrepare"));
100 static cl::opt<bool> DisableExtLdPromotion(
101 "disable-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
102 cl::desc("Disable ext(promotable(ld)) -> promoted(ext(ld)) optimization in "
105 static cl::opt<bool> StressExtLdPromotion(
106 "stress-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
107 cl::desc("Stress test ext(promotable(ld)) -> promoted(ext(ld)) "
108 "optimization in CodeGenPrepare"));
111 typedef SmallPtrSet<Instruction *, 16> SetOfInstrs;
115 TypeIsSExt(Type *Ty, bool IsSExt) : Ty(Ty), IsSExt(IsSExt) {}
117 typedef DenseMap<Instruction *, TypeIsSExt> InstrToOrigTy;
118 class TypePromotionTransaction;
120 class CodeGenPrepare : public FunctionPass {
121 /// TLI - Keep a pointer of a TargetLowering to consult for determining
122 /// transformation profitability.
123 const TargetMachine *TM;
124 const TargetLowering *TLI;
125 const TargetTransformInfo *TTI;
126 const TargetLibraryInfo *TLInfo;
129 /// CurInstIterator - As we scan instructions optimizing them, this is the
130 /// next instruction to optimize. Xforms that can invalidate this should
132 BasicBlock::iterator CurInstIterator;
134 /// Keeps track of non-local addresses that have been sunk into a block.
135 /// This allows us to avoid inserting duplicate code for blocks with
136 /// multiple load/stores of the same address.
137 ValueMap<Value*, Value*> SunkAddrs;
139 /// Keeps track of all truncates inserted for the current function.
140 SetOfInstrs InsertedTruncsSet;
141 /// Keeps track of the type of the related instruction before their
142 /// promotion for the current function.
143 InstrToOrigTy PromotedInsts;
145 /// ModifiedDT - If CFG is modified in anyway, dominator tree may need to
149 /// OptSize - True if optimizing for size.
153 static char ID; // Pass identification, replacement for typeid
154 explicit CodeGenPrepare(const TargetMachine *TM = nullptr)
155 : FunctionPass(ID), TM(TM), TLI(nullptr), TTI(nullptr) {
156 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
158 bool runOnFunction(Function &F) override;
160 const char *getPassName() const override { return "CodeGen Prepare"; }
162 void getAnalysisUsage(AnalysisUsage &AU) const override {
163 AU.addPreserved<DominatorTreeWrapperPass>();
164 AU.addRequired<TargetLibraryInfoWrapperPass>();
165 AU.addRequired<TargetTransformInfoWrapperPass>();
169 bool EliminateFallThrough(Function &F);
170 bool EliminateMostlyEmptyBlocks(Function &F);
171 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
172 void EliminateMostlyEmptyBlock(BasicBlock *BB);
173 bool OptimizeBlock(BasicBlock &BB, bool& ModifiedDT);
174 bool OptimizeInst(Instruction *I, bool& ModifiedDT);
175 bool OptimizeMemoryInst(Instruction *I, Value *Addr, Type *AccessTy);
176 bool OptimizeInlineAsmInst(CallInst *CS);
177 bool OptimizeCallInst(CallInst *CI, bool& ModifiedDT);
178 bool MoveExtToFormExtLoad(Instruction *&I);
179 bool OptimizeExtUses(Instruction *I);
180 bool OptimizeSelectInst(SelectInst *SI);
181 bool OptimizeShuffleVectorInst(ShuffleVectorInst *SI);
182 bool OptimizeExtractElementInst(Instruction *Inst);
183 bool DupRetToEnableTailCallOpts(BasicBlock *BB);
184 bool PlaceDbgValues(Function &F);
185 bool sinkAndCmp(Function &F);
186 bool ExtLdPromotion(TypePromotionTransaction &TPT, LoadInst *&LI,
188 const SmallVectorImpl<Instruction *> &Exts,
189 unsigned CreatedInst);
190 bool splitBranchCondition(Function &F);
191 bool simplifyOffsetableRelocate(Instruction &I);
195 char CodeGenPrepare::ID = 0;
196 INITIALIZE_TM_PASS(CodeGenPrepare, "codegenprepare",
197 "Optimize for code generation", false, false)
199 FunctionPass *llvm::createCodeGenPreparePass(const TargetMachine *TM) {
200 return new CodeGenPrepare(TM);
203 bool CodeGenPrepare::runOnFunction(Function &F) {
204 if (skipOptnoneFunction(F))
207 bool EverMadeChange = false;
208 // Clear per function information.
209 InsertedTruncsSet.clear();
210 PromotedInsts.clear();
214 TLI = TM->getSubtargetImpl(F)->getTargetLowering();
215 TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
216 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
217 DominatorTreeWrapperPass *DTWP =
218 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
219 DT = DTWP ? &DTWP->getDomTree() : nullptr;
220 OptSize = F.hasFnAttribute(Attribute::OptimizeForSize);
222 /// This optimization identifies DIV instructions that can be
223 /// profitably bypassed and carried out with a shorter, faster divide.
224 if (!OptSize && TLI && TLI->isSlowDivBypassed()) {
225 const DenseMap<unsigned int, unsigned int> &BypassWidths =
226 TLI->getBypassSlowDivWidths();
227 for (Function::iterator I = F.begin(); I != F.end(); I++)
228 EverMadeChange |= bypassSlowDivision(F, I, BypassWidths);
231 // Eliminate blocks that contain only PHI nodes and an
232 // unconditional branch.
233 EverMadeChange |= EliminateMostlyEmptyBlocks(F);
235 // llvm.dbg.value is far away from the value then iSel may not be able
236 // handle it properly. iSel will drop llvm.dbg.value if it can not
237 // find a node corresponding to the value.
238 EverMadeChange |= PlaceDbgValues(F);
240 // If there is a mask, compare against zero, and branch that can be combined
241 // into a single target instruction, push the mask and compare into branch
242 // users. Do this before OptimizeBlock -> OptimizeInst ->
243 // OptimizeCmpExpression, which perturbs the pattern being searched for.
244 if (!DisableBranchOpts) {
245 EverMadeChange |= sinkAndCmp(F);
246 EverMadeChange |= splitBranchCondition(F);
249 bool MadeChange = true;
252 for (Function::iterator I = F.begin(); I != F.end(); ) {
253 BasicBlock *BB = I++;
254 bool ModifiedDTOnIteration = false;
255 MadeChange |= OptimizeBlock(*BB, ModifiedDTOnIteration);
257 // Restart BB iteration if the dominator tree of the Function was changed
258 ModifiedDT |= ModifiedDTOnIteration;
259 if (ModifiedDTOnIteration)
262 EverMadeChange |= MadeChange;
267 if (!DisableBranchOpts) {
269 SmallPtrSet<BasicBlock*, 8> WorkList;
270 for (BasicBlock &BB : F) {
271 SmallVector<BasicBlock *, 2> Successors(succ_begin(&BB), succ_end(&BB));
272 MadeChange |= ConstantFoldTerminator(&BB, true);
273 if (!MadeChange) continue;
275 for (SmallVectorImpl<BasicBlock*>::iterator
276 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
277 if (pred_begin(*II) == pred_end(*II))
278 WorkList.insert(*II);
281 // Delete the dead blocks and any of their dead successors.
282 MadeChange |= !WorkList.empty();
283 while (!WorkList.empty()) {
284 BasicBlock *BB = *WorkList.begin();
286 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
290 for (SmallVectorImpl<BasicBlock*>::iterator
291 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
292 if (pred_begin(*II) == pred_end(*II))
293 WorkList.insert(*II);
296 // Merge pairs of basic blocks with unconditional branches, connected by
298 if (EverMadeChange || MadeChange)
299 MadeChange |= EliminateFallThrough(F);
303 EverMadeChange |= MadeChange;
306 if (!DisableGCOpts) {
307 SmallVector<Instruction *, 2> Statepoints;
308 for (BasicBlock &BB : F)
309 for (Instruction &I : BB)
311 Statepoints.push_back(&I);
312 for (auto &I : Statepoints)
313 EverMadeChange |= simplifyOffsetableRelocate(*I);
316 if (ModifiedDT && DT)
319 return EverMadeChange;
322 /// EliminateFallThrough - Merge basic blocks which are connected
323 /// by a single edge, where one of the basic blocks has a single successor
324 /// pointing to the other basic block, which has a single predecessor.
325 bool CodeGenPrepare::EliminateFallThrough(Function &F) {
326 bool Changed = false;
327 // Scan all of the blocks in the function, except for the entry block.
328 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
329 BasicBlock *BB = I++;
330 // If the destination block has a single pred, then this is a trivial
331 // edge, just collapse it.
332 BasicBlock *SinglePred = BB->getSinglePredecessor();
334 // Don't merge if BB's address is taken.
335 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
337 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
338 if (Term && !Term->isConditional()) {
340 DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n");
341 // Remember if SinglePred was the entry block of the function.
342 // If so, we will need to move BB back to the entry position.
343 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
344 MergeBasicBlockIntoOnlyPred(BB, DT);
346 if (isEntry && BB != &BB->getParent()->getEntryBlock())
347 BB->moveBefore(&BB->getParent()->getEntryBlock());
349 // We have erased a block. Update the iterator.
356 /// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes,
357 /// debug info directives, and an unconditional branch. Passes before isel
358 /// (e.g. LSR/loopsimplify) often split edges in ways that are non-optimal for
359 /// isel. Start by eliminating these blocks so we can split them the way we
361 bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) {
362 bool MadeChange = false;
363 // Note that this intentionally skips the entry block.
364 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
365 BasicBlock *BB = I++;
367 // If this block doesn't end with an uncond branch, ignore it.
368 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
369 if (!BI || !BI->isUnconditional())
372 // If the instruction before the branch (skipping debug info) isn't a phi
373 // node, then other stuff is happening here.
374 BasicBlock::iterator BBI = BI;
375 if (BBI != BB->begin()) {
377 while (isa<DbgInfoIntrinsic>(BBI)) {
378 if (BBI == BB->begin())
382 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
386 // Do not break infinite loops.
387 BasicBlock *DestBB = BI->getSuccessor(0);
391 if (!CanMergeBlocks(BB, DestBB))
394 EliminateMostlyEmptyBlock(BB);
400 /// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a
401 /// single uncond branch between them, and BB contains no other non-phi
403 bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB,
404 const BasicBlock *DestBB) const {
405 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
406 // the successor. If there are more complex condition (e.g. preheaders),
407 // don't mess around with them.
408 BasicBlock::const_iterator BBI = BB->begin();
409 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
410 for (const User *U : PN->users()) {
411 const Instruction *UI = cast<Instruction>(U);
412 if (UI->getParent() != DestBB || !isa<PHINode>(UI))
414 // If User is inside DestBB block and it is a PHINode then check
415 // incoming value. If incoming value is not from BB then this is
416 // a complex condition (e.g. preheaders) we want to avoid here.
417 if (UI->getParent() == DestBB) {
418 if (const PHINode *UPN = dyn_cast<PHINode>(UI))
419 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
420 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
421 if (Insn && Insn->getParent() == BB &&
422 Insn->getParent() != UPN->getIncomingBlock(I))
429 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
430 // and DestBB may have conflicting incoming values for the block. If so, we
431 // can't merge the block.
432 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
433 if (!DestBBPN) return true; // no conflict.
435 // Collect the preds of BB.
436 SmallPtrSet<const BasicBlock*, 16> BBPreds;
437 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
438 // It is faster to get preds from a PHI than with pred_iterator.
439 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
440 BBPreds.insert(BBPN->getIncomingBlock(i));
442 BBPreds.insert(pred_begin(BB), pred_end(BB));
445 // Walk the preds of DestBB.
446 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
447 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
448 if (BBPreds.count(Pred)) { // Common predecessor?
449 BBI = DestBB->begin();
450 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
451 const Value *V1 = PN->getIncomingValueForBlock(Pred);
452 const Value *V2 = PN->getIncomingValueForBlock(BB);
454 // If V2 is a phi node in BB, look up what the mapped value will be.
455 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
456 if (V2PN->getParent() == BB)
457 V2 = V2PN->getIncomingValueForBlock(Pred);
459 // If there is a conflict, bail out.
460 if (V1 != V2) return false;
469 /// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and
470 /// an unconditional branch in it.
471 void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) {
472 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
473 BasicBlock *DestBB = BI->getSuccessor(0);
475 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
477 // If the destination block has a single pred, then this is a trivial edge,
479 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
480 if (SinglePred != DestBB) {
481 // Remember if SinglePred was the entry block of the function. If so, we
482 // will need to move BB back to the entry position.
483 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
484 MergeBasicBlockIntoOnlyPred(DestBB, DT);
486 if (isEntry && BB != &BB->getParent()->getEntryBlock())
487 BB->moveBefore(&BB->getParent()->getEntryBlock());
489 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
494 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
495 // to handle the new incoming edges it is about to have.
497 for (BasicBlock::iterator BBI = DestBB->begin();
498 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
499 // Remove the incoming value for BB, and remember it.
500 Value *InVal = PN->removeIncomingValue(BB, false);
502 // Two options: either the InVal is a phi node defined in BB or it is some
503 // value that dominates BB.
504 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
505 if (InValPhi && InValPhi->getParent() == BB) {
506 // Add all of the input values of the input PHI as inputs of this phi.
507 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
508 PN->addIncoming(InValPhi->getIncomingValue(i),
509 InValPhi->getIncomingBlock(i));
511 // Otherwise, add one instance of the dominating value for each edge that
512 // we will be adding.
513 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
514 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
515 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
517 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
518 PN->addIncoming(InVal, *PI);
523 // The PHIs are now updated, change everything that refers to BB to use
524 // DestBB and remove BB.
525 BB->replaceAllUsesWith(DestBB);
526 if (DT && !ModifiedDT) {
527 BasicBlock *BBIDom = DT->getNode(BB)->getIDom()->getBlock();
528 BasicBlock *DestBBIDom = DT->getNode(DestBB)->getIDom()->getBlock();
529 BasicBlock *NewIDom = DT->findNearestCommonDominator(BBIDom, DestBBIDom);
530 DT->changeImmediateDominator(DestBB, NewIDom);
533 BB->eraseFromParent();
536 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
539 // Computes a map of base pointer relocation instructions to corresponding
540 // derived pointer relocation instructions given a vector of all relocate calls
541 static void computeBaseDerivedRelocateMap(
542 const SmallVectorImpl<User *> &AllRelocateCalls,
543 DenseMap<IntrinsicInst *, SmallVector<IntrinsicInst *, 2>> &
545 // Collect information in two maps: one primarily for locating the base object
546 // while filling the second map; the second map is the final structure holding
547 // a mapping between Base and corresponding Derived relocate calls
548 DenseMap<std::pair<unsigned, unsigned>, IntrinsicInst *> RelocateIdxMap;
549 for (auto &U : AllRelocateCalls) {
550 GCRelocateOperands ThisRelocate(U);
551 IntrinsicInst *I = cast<IntrinsicInst>(U);
552 auto K = std::make_pair(ThisRelocate.basePtrIndex(),
553 ThisRelocate.derivedPtrIndex());
554 RelocateIdxMap.insert(std::make_pair(K, I));
556 for (auto &Item : RelocateIdxMap) {
557 std::pair<unsigned, unsigned> Key = Item.first;
558 if (Key.first == Key.second)
559 // Base relocation: nothing to insert
562 IntrinsicInst *I = Item.second;
563 auto BaseKey = std::make_pair(Key.first, Key.first);
564 IntrinsicInst *Base = RelocateIdxMap[BaseKey];
566 // TODO: We might want to insert a new base object relocate and gep off
567 // that, if there are enough derived object relocates.
569 RelocateInstMap[Base].push_back(I);
573 // Accepts a GEP and extracts the operands into a vector provided they're all
574 // small integer constants
575 static bool getGEPSmallConstantIntOffsetV(GetElementPtrInst *GEP,
576 SmallVectorImpl<Value *> &OffsetV) {
577 for (unsigned i = 1; i < GEP->getNumOperands(); i++) {
578 // Only accept small constant integer operands
579 auto Op = dyn_cast<ConstantInt>(GEP->getOperand(i));
580 if (!Op || Op->getZExtValue() > 20)
584 for (unsigned i = 1; i < GEP->getNumOperands(); i++)
585 OffsetV.push_back(GEP->getOperand(i));
589 // Takes a RelocatedBase (base pointer relocation instruction) and Targets to
590 // replace, computes a replacement, and affects it.
592 simplifyRelocatesOffABase(IntrinsicInst *RelocatedBase,
593 const SmallVectorImpl<IntrinsicInst *> &Targets) {
594 bool MadeChange = false;
595 for (auto &ToReplace : Targets) {
596 GCRelocateOperands MasterRelocate(RelocatedBase);
597 GCRelocateOperands ThisRelocate(ToReplace);
599 assert(ThisRelocate.basePtrIndex() == MasterRelocate.basePtrIndex() &&
600 "Not relocating a derived object of the original base object");
601 if (ThisRelocate.basePtrIndex() == ThisRelocate.derivedPtrIndex()) {
602 // A duplicate relocate call. TODO: coalesce duplicates.
606 Value *Base = ThisRelocate.basePtr();
607 auto Derived = dyn_cast<GetElementPtrInst>(ThisRelocate.derivedPtr());
608 if (!Derived || Derived->getPointerOperand() != Base)
611 SmallVector<Value *, 2> OffsetV;
612 if (!getGEPSmallConstantIntOffsetV(Derived, OffsetV))
615 // Create a Builder and replace the target callsite with a gep
616 IRBuilder<> Builder(ToReplace);
617 Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
619 Builder.CreateGEP(RelocatedBase, makeArrayRef(OffsetV));
620 Instruction *ReplacementInst = cast<Instruction>(Replacement);
621 ReplacementInst->removeFromParent();
622 ReplacementInst->insertAfter(RelocatedBase);
623 Replacement->takeName(ToReplace);
624 ToReplace->replaceAllUsesWith(Replacement);
625 ToReplace->eraseFromParent();
635 // %ptr = gep %base + 15
636 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
637 // %base' = relocate(%tok, i32 4, i32 4)
638 // %ptr' = relocate(%tok, i32 4, i32 5)
644 // %ptr = gep %base + 15
645 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
646 // %base' = gc.relocate(%tok, i32 4, i32 4)
647 // %ptr' = gep %base' + 15
649 bool CodeGenPrepare::simplifyOffsetableRelocate(Instruction &I) {
650 bool MadeChange = false;
651 SmallVector<User *, 2> AllRelocateCalls;
653 for (auto *U : I.users())
654 if (isGCRelocate(dyn_cast<Instruction>(U)))
655 // Collect all the relocate calls associated with a statepoint
656 AllRelocateCalls.push_back(U);
658 // We need atleast one base pointer relocation + one derived pointer
659 // relocation to mangle
660 if (AllRelocateCalls.size() < 2)
663 // RelocateInstMap is a mapping from the base relocate instruction to the
664 // corresponding derived relocate instructions
665 DenseMap<IntrinsicInst *, SmallVector<IntrinsicInst *, 2>> RelocateInstMap;
666 computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap);
667 if (RelocateInstMap.empty())
670 for (auto &Item : RelocateInstMap)
671 // Item.first is the RelocatedBase to offset against
672 // Item.second is the vector of Targets to replace
673 MadeChange = simplifyRelocatesOffABase(Item.first, Item.second);
677 /// SinkCast - Sink the specified cast instruction into its user blocks
678 static bool SinkCast(CastInst *CI) {
679 BasicBlock *DefBB = CI->getParent();
681 /// InsertedCasts - Only insert a cast in each block once.
682 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
684 bool MadeChange = false;
685 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
687 Use &TheUse = UI.getUse();
688 Instruction *User = cast<Instruction>(*UI);
690 // Figure out which BB this cast is used in. For PHI's this is the
691 // appropriate predecessor block.
692 BasicBlock *UserBB = User->getParent();
693 if (PHINode *PN = dyn_cast<PHINode>(User)) {
694 UserBB = PN->getIncomingBlock(TheUse);
697 // Preincrement use iterator so we don't invalidate it.
700 // If this user is in the same block as the cast, don't change the cast.
701 if (UserBB == DefBB) continue;
703 // If we have already inserted a cast into this block, use it.
704 CastInst *&InsertedCast = InsertedCasts[UserBB];
707 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
709 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "",
714 // Replace a use of the cast with a use of the new cast.
715 TheUse = InsertedCast;
719 // If we removed all uses, nuke the cast.
720 if (CI->use_empty()) {
721 CI->eraseFromParent();
728 /// OptimizeNoopCopyExpression - If the specified cast instruction is a noop
729 /// copy (e.g. it's casting from one pointer type to another, i32->i8 on PPC),
730 /// sink it into user blocks to reduce the number of virtual
731 /// registers that must be created and coalesced.
733 /// Return true if any changes are made.
735 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){
736 // If this is a noop copy,
737 EVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType());
738 EVT DstVT = TLI.getValueType(CI->getType());
740 // This is an fp<->int conversion?
741 if (SrcVT.isInteger() != DstVT.isInteger())
744 // If this is an extension, it will be a zero or sign extension, which
746 if (SrcVT.bitsLT(DstVT)) return false;
748 // If these values will be promoted, find out what they will be promoted
749 // to. This helps us consider truncates on PPC as noop copies when they
751 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
752 TargetLowering::TypePromoteInteger)
753 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
754 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
755 TargetLowering::TypePromoteInteger)
756 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
758 // If, after promotion, these are the same types, this is a noop copy.
765 /// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce
766 /// the number of virtual registers that must be created and coalesced. This is
767 /// a clear win except on targets with multiple condition code registers
768 /// (PowerPC), where it might lose; some adjustment may be wanted there.
770 /// Return true if any changes are made.
771 static bool OptimizeCmpExpression(CmpInst *CI) {
772 BasicBlock *DefBB = CI->getParent();
774 /// InsertedCmp - Only insert a cmp in each block once.
775 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
777 bool MadeChange = false;
778 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
780 Use &TheUse = UI.getUse();
781 Instruction *User = cast<Instruction>(*UI);
783 // Preincrement use iterator so we don't invalidate it.
786 // Don't bother for PHI nodes.
787 if (isa<PHINode>(User))
790 // Figure out which BB this cmp is used in.
791 BasicBlock *UserBB = User->getParent();
793 // If this user is in the same block as the cmp, don't change the cmp.
794 if (UserBB == DefBB) continue;
796 // If we have already inserted a cmp into this block, use it.
797 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
800 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
802 CmpInst::Create(CI->getOpcode(),
803 CI->getPredicate(), CI->getOperand(0),
804 CI->getOperand(1), "", InsertPt);
808 // Replace a use of the cmp with a use of the new cmp.
809 TheUse = InsertedCmp;
813 // If we removed all uses, nuke the cmp.
815 CI->eraseFromParent();
820 /// isExtractBitsCandidateUse - Check if the candidates could
821 /// be combined with shift instruction, which includes:
822 /// 1. Truncate instruction
823 /// 2. And instruction and the imm is a mask of the low bits:
824 /// imm & (imm+1) == 0
825 static bool isExtractBitsCandidateUse(Instruction *User) {
826 if (!isa<TruncInst>(User)) {
827 if (User->getOpcode() != Instruction::And ||
828 !isa<ConstantInt>(User->getOperand(1)))
831 const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
833 if ((Cimm & (Cimm + 1)).getBoolValue())
839 /// SinkShiftAndTruncate - sink both shift and truncate instruction
840 /// to the use of truncate's BB.
842 SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI,
843 DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts,
844 const TargetLowering &TLI) {
845 BasicBlock *UserBB = User->getParent();
846 DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
847 TruncInst *TruncI = dyn_cast<TruncInst>(User);
848 bool MadeChange = false;
850 for (Value::user_iterator TruncUI = TruncI->user_begin(),
851 TruncE = TruncI->user_end();
852 TruncUI != TruncE;) {
854 Use &TruncTheUse = TruncUI.getUse();
855 Instruction *TruncUser = cast<Instruction>(*TruncUI);
856 // Preincrement use iterator so we don't invalidate it.
860 int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
864 // If the use is actually a legal node, there will not be an
865 // implicit truncate.
866 // FIXME: always querying the result type is just an
867 // approximation; some nodes' legality is determined by the
868 // operand or other means. There's no good way to find out though.
869 if (TLI.isOperationLegalOrCustom(
870 ISDOpcode, TLI.getValueType(TruncUser->getType(), true)))
873 // Don't bother for PHI nodes.
874 if (isa<PHINode>(TruncUser))
877 BasicBlock *TruncUserBB = TruncUser->getParent();
879 if (UserBB == TruncUserBB)
882 BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
883 CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
885 if (!InsertedShift && !InsertedTrunc) {
886 BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
888 if (ShiftI->getOpcode() == Instruction::AShr)
890 BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI, "", InsertPt);
893 BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI, "", InsertPt);
896 BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
899 InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
900 TruncI->getType(), "", TruncInsertPt);
904 TruncTheUse = InsertedTrunc;
910 /// OptimizeExtractBits - sink the shift *right* instruction into user blocks if
911 /// the uses could potentially be combined with this shift instruction and
912 /// generate BitExtract instruction. It will only be applied if the architecture
913 /// supports BitExtract instruction. Here is an example:
915 /// %x.extract.shift = lshr i64 %arg1, 32
917 /// %x.extract.trunc = trunc i64 %x.extract.shift to i16
921 /// %x.extract.shift.1 = lshr i64 %arg1, 32
922 /// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
924 /// CodeGen will recoginze the pattern in BB2 and generate BitExtract
926 /// Return true if any changes are made.
927 static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI,
928 const TargetLowering &TLI) {
929 BasicBlock *DefBB = ShiftI->getParent();
931 /// Only insert instructions in each block once.
932 DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts;
934 bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(ShiftI->getType()));
936 bool MadeChange = false;
937 for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
939 Use &TheUse = UI.getUse();
940 Instruction *User = cast<Instruction>(*UI);
941 // Preincrement use iterator so we don't invalidate it.
944 // Don't bother for PHI nodes.
945 if (isa<PHINode>(User))
948 if (!isExtractBitsCandidateUse(User))
951 BasicBlock *UserBB = User->getParent();
953 if (UserBB == DefBB) {
954 // If the shift and truncate instruction are in the same BB. The use of
955 // the truncate(TruncUse) may still introduce another truncate if not
956 // legal. In this case, we would like to sink both shift and truncate
957 // instruction to the BB of TruncUse.
960 // i64 shift.result = lshr i64 opnd, imm
961 // trunc.result = trunc shift.result to i16
964 // ----> We will have an implicit truncate here if the architecture does
965 // not have i16 compare.
966 // cmp i16 trunc.result, opnd2
968 if (isa<TruncInst>(User) && shiftIsLegal
969 // If the type of the truncate is legal, no trucate will be
970 // introduced in other basic blocks.
971 && (!TLI.isTypeLegal(TLI.getValueType(User->getType()))))
973 SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI);
977 // If we have already inserted a shift into this block, use it.
978 BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
980 if (!InsertedShift) {
981 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
983 if (ShiftI->getOpcode() == Instruction::AShr)
985 BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI, "", InsertPt);
988 BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI, "", InsertPt);
993 // Replace a use of the shift with a use of the new shift.
994 TheUse = InsertedShift;
997 // If we removed all uses, nuke the shift.
998 if (ShiftI->use_empty())
999 ShiftI->eraseFromParent();
1004 // ScalarizeMaskedLoad() translates masked load intrinsic, like
1005 // <16 x i32 > @llvm.masked.load( <16 x i32>* %addr, i32 align,
1006 // <16 x i1> %mask, <16 x i32> %passthru)
1007 // to a chain of basic blocks, whith loading element one-by-one if
1008 // the appropriate mask bit is set
1010 // %1 = bitcast i8* %addr to i32*
1011 // %2 = extractelement <16 x i1> %mask, i32 0
1012 // %3 = icmp eq i1 %2, true
1013 // br i1 %3, label %cond.load, label %else
1015 //cond.load: ; preds = %0
1016 // %4 = getelementptr i32* %1, i32 0
1017 // %5 = load i32* %4
1018 // %6 = insertelement <16 x i32> undef, i32 %5, i32 0
1021 //else: ; preds = %0, %cond.load
1022 // %res.phi.else = phi <16 x i32> [ %6, %cond.load ], [ undef, %0 ]
1023 // %7 = extractelement <16 x i1> %mask, i32 1
1024 // %8 = icmp eq i1 %7, true
1025 // br i1 %8, label %cond.load1, label %else2
1027 //cond.load1: ; preds = %else
1028 // %9 = getelementptr i32* %1, i32 1
1029 // %10 = load i32* %9
1030 // %11 = insertelement <16 x i32> %res.phi.else, i32 %10, i32 1
1033 //else2: ; preds = %else, %cond.load1
1034 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
1035 // %12 = extractelement <16 x i1> %mask, i32 2
1036 // %13 = icmp eq i1 %12, true
1037 // br i1 %13, label %cond.load4, label %else5
1039 static void ScalarizeMaskedLoad(CallInst *CI) {
1040 Value *Ptr = CI->getArgOperand(0);
1041 Value *Src0 = CI->getArgOperand(3);
1042 Value *Mask = CI->getArgOperand(2);
1043 VectorType *VecType = dyn_cast<VectorType>(CI->getType());
1044 Type *EltTy = VecType->getElementType();
1046 assert(VecType && "Unexpected return type of masked load intrinsic");
1048 IRBuilder<> Builder(CI->getContext());
1049 Instruction *InsertPt = CI;
1050 BasicBlock *IfBlock = CI->getParent();
1051 BasicBlock *CondBlock = nullptr;
1052 BasicBlock *PrevIfBlock = CI->getParent();
1053 Builder.SetInsertPoint(InsertPt);
1055 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1057 // Bitcast %addr fron i8* to EltTy*
1059 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
1060 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
1061 Value *UndefVal = UndefValue::get(VecType);
1063 // The result vector
1064 Value *VResult = UndefVal;
1066 PHINode *Phi = nullptr;
1067 Value *PrevPhi = UndefVal;
1069 unsigned VectorWidth = VecType->getNumElements();
1070 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1072 // Fill the "else" block, created in the previous iteration
1074 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
1075 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
1076 // %to_load = icmp eq i1 %mask_1, true
1077 // br i1 %to_load, label %cond.load, label %else
1080 Phi = Builder.CreatePHI(VecType, 2, "res.phi.else");
1081 Phi->addIncoming(VResult, CondBlock);
1082 Phi->addIncoming(PrevPhi, PrevIfBlock);
1087 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
1088 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1089 ConstantInt::get(Predicate->getType(), 1));
1091 // Create "cond" block
1093 // %EltAddr = getelementptr i32* %1, i32 0
1094 // %Elt = load i32* %EltAddr
1095 // VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx
1097 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.load");
1098 Builder.SetInsertPoint(InsertPt);
1100 Value* Gep = Builder.CreateInBoundsGEP(FirstEltPtr, Builder.getInt32(Idx));
1101 LoadInst* Load = Builder.CreateLoad(Gep, false);
1102 VResult = Builder.CreateInsertElement(VResult, Load, Builder.getInt32(Idx));
1104 // Create "else" block, fill it in the next iteration
1105 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1106 Builder.SetInsertPoint(InsertPt);
1107 Instruction *OldBr = IfBlock->getTerminator();
1108 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1109 OldBr->eraseFromParent();
1110 PrevIfBlock = IfBlock;
1111 IfBlock = NewIfBlock;
1114 Phi = Builder.CreatePHI(VecType, 2, "res.phi.select");
1115 Phi->addIncoming(VResult, CondBlock);
1116 Phi->addIncoming(PrevPhi, PrevIfBlock);
1117 Value *NewI = Builder.CreateSelect(Mask, Phi, Src0);
1118 CI->replaceAllUsesWith(NewI);
1119 CI->eraseFromParent();
1122 // ScalarizeMaskedStore() translates masked store intrinsic, like
1123 // void @llvm.masked.store(<16 x i32> %src, <16 x i32>* %addr, i32 align,
1125 // to a chain of basic blocks, that stores element one-by-one if
1126 // the appropriate mask bit is set
1128 // %1 = bitcast i8* %addr to i32*
1129 // %2 = extractelement <16 x i1> %mask, i32 0
1130 // %3 = icmp eq i1 %2, true
1131 // br i1 %3, label %cond.store, label %else
1133 // cond.store: ; preds = %0
1134 // %4 = extractelement <16 x i32> %val, i32 0
1135 // %5 = getelementptr i32* %1, i32 0
1136 // store i32 %4, i32* %5
1139 // else: ; preds = %0, %cond.store
1140 // %6 = extractelement <16 x i1> %mask, i32 1
1141 // %7 = icmp eq i1 %6, true
1142 // br i1 %7, label %cond.store1, label %else2
1144 // cond.store1: ; preds = %else
1145 // %8 = extractelement <16 x i32> %val, i32 1
1146 // %9 = getelementptr i32* %1, i32 1
1147 // store i32 %8, i32* %9
1150 static void ScalarizeMaskedStore(CallInst *CI) {
1151 Value *Ptr = CI->getArgOperand(1);
1152 Value *Src = CI->getArgOperand(0);
1153 Value *Mask = CI->getArgOperand(3);
1155 VectorType *VecType = dyn_cast<VectorType>(Src->getType());
1156 Type *EltTy = VecType->getElementType();
1158 assert(VecType && "Unexpected data type in masked store intrinsic");
1160 IRBuilder<> Builder(CI->getContext());
1161 Instruction *InsertPt = CI;
1162 BasicBlock *IfBlock = CI->getParent();
1163 Builder.SetInsertPoint(InsertPt);
1164 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1166 // Bitcast %addr fron i8* to EltTy*
1168 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
1169 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
1171 unsigned VectorWidth = VecType->getNumElements();
1172 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1174 // Fill the "else" block, created in the previous iteration
1176 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
1177 // %to_store = icmp eq i1 %mask_1, true
1178 // br i1 %to_load, label %cond.store, label %else
1180 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
1181 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1182 ConstantInt::get(Predicate->getType(), 1));
1184 // Create "cond" block
1186 // %OneElt = extractelement <16 x i32> %Src, i32 Idx
1187 // %EltAddr = getelementptr i32* %1, i32 0
1188 // %store i32 %OneElt, i32* %EltAddr
1190 BasicBlock *CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1191 Builder.SetInsertPoint(InsertPt);
1193 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx));
1194 Value* Gep = Builder.CreateInBoundsGEP(FirstEltPtr, Builder.getInt32(Idx));
1195 Builder.CreateStore(OneElt, Gep);
1197 // Create "else" block, fill it in the next iteration
1198 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1199 Builder.SetInsertPoint(InsertPt);
1200 Instruction *OldBr = IfBlock->getTerminator();
1201 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1202 OldBr->eraseFromParent();
1203 IfBlock = NewIfBlock;
1205 CI->eraseFromParent();
1208 bool CodeGenPrepare::OptimizeCallInst(CallInst *CI, bool& ModifiedDT) {
1209 BasicBlock *BB = CI->getParent();
1211 // Lower inline assembly if we can.
1212 // If we found an inline asm expession, and if the target knows how to
1213 // lower it to normal LLVM code, do so now.
1214 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
1215 if (TLI->ExpandInlineAsm(CI)) {
1216 // Avoid invalidating the iterator.
1217 CurInstIterator = BB->begin();
1218 // Avoid processing instructions out of order, which could cause
1219 // reuse before a value is defined.
1223 // Sink address computing for memory operands into the block.
1224 if (OptimizeInlineAsmInst(CI))
1228 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
1230 switch (II->getIntrinsicID()) {
1232 case Intrinsic::objectsize: {
1233 // Lower all uses of llvm.objectsize.*
1234 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
1235 Type *ReturnTy = CI->getType();
1236 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
1238 // Substituting this can cause recursive simplifications, which can
1239 // invalidate our iterator. Use a WeakVH to hold onto it in case this
1241 WeakVH IterHandle(CurInstIterator);
1243 replaceAndRecursivelySimplify(CI, RetVal,
1244 TLI ? TLI->getDataLayout() : nullptr,
1245 TLInfo, ModifiedDT ? nullptr : DT);
1247 // If the iterator instruction was recursively deleted, start over at the
1248 // start of the block.
1249 if (IterHandle != CurInstIterator) {
1250 CurInstIterator = BB->begin();
1255 case Intrinsic::masked_load: {
1256 // Scalarize unsupported vector masked load
1257 if (!TTI->isLegalMaskedLoad(CI->getType(), 1)) {
1258 ScalarizeMaskedLoad(CI);
1264 case Intrinsic::masked_store: {
1265 if (!TTI->isLegalMaskedStore(CI->getArgOperand(0)->getType(), 1)) {
1266 ScalarizeMaskedStore(CI);
1275 SmallVector<Value*, 2> PtrOps;
1277 if (TLI->GetAddrModeArguments(II, PtrOps, AccessTy))
1278 while (!PtrOps.empty())
1279 if (OptimizeMemoryInst(II, PtrOps.pop_back_val(), AccessTy))
1284 // From here on out we're working with named functions.
1285 if (!CI->getCalledFunction()) return false;
1287 // We'll need DataLayout from here on out.
1288 const DataLayout *TD = TLI ? TLI->getDataLayout() : nullptr;
1289 if (!TD) return false;
1291 // Lower all default uses of _chk calls. This is very similar
1292 // to what InstCombineCalls does, but here we are only lowering calls
1293 // to fortified library functions (e.g. __memcpy_chk) that have the default
1294 // "don't know" as the objectsize. Anything else should be left alone.
1295 FortifiedLibCallSimplifier Simplifier(TD, TLInfo, true);
1296 if (Value *V = Simplifier.optimizeCall(CI)) {
1297 CI->replaceAllUsesWith(V);
1298 CI->eraseFromParent();
1304 /// DupRetToEnableTailCallOpts - Look for opportunities to duplicate return
1305 /// instructions to the predecessor to enable tail call optimizations. The
1306 /// case it is currently looking for is:
1309 /// %tmp0 = tail call i32 @f0()
1310 /// br label %return
1312 /// %tmp1 = tail call i32 @f1()
1313 /// br label %return
1315 /// %tmp2 = tail call i32 @f2()
1316 /// br label %return
1318 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
1326 /// %tmp0 = tail call i32 @f0()
1329 /// %tmp1 = tail call i32 @f1()
1332 /// %tmp2 = tail call i32 @f2()
1335 bool CodeGenPrepare::DupRetToEnableTailCallOpts(BasicBlock *BB) {
1339 ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator());
1343 PHINode *PN = nullptr;
1344 BitCastInst *BCI = nullptr;
1345 Value *V = RI->getReturnValue();
1347 BCI = dyn_cast<BitCastInst>(V);
1349 V = BCI->getOperand(0);
1351 PN = dyn_cast<PHINode>(V);
1356 if (PN && PN->getParent() != BB)
1359 // It's not safe to eliminate the sign / zero extension of the return value.
1360 // See llvm::isInTailCallPosition().
1361 const Function *F = BB->getParent();
1362 AttributeSet CallerAttrs = F->getAttributes();
1363 if (CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt) ||
1364 CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
1367 // Make sure there are no instructions between the PHI and return, or that the
1368 // return is the first instruction in the block.
1370 BasicBlock::iterator BI = BB->begin();
1371 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
1373 // Also skip over the bitcast.
1378 BasicBlock::iterator BI = BB->begin();
1379 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
1384 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
1386 SmallVector<CallInst*, 4> TailCalls;
1388 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
1389 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
1390 // Make sure the phi value is indeed produced by the tail call.
1391 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
1392 TLI->mayBeEmittedAsTailCall(CI))
1393 TailCalls.push_back(CI);
1396 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
1397 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
1398 if (!VisitedBBs.insert(*PI).second)
1401 BasicBlock::InstListType &InstList = (*PI)->getInstList();
1402 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
1403 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
1404 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
1408 CallInst *CI = dyn_cast<CallInst>(&*RI);
1409 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
1410 TailCalls.push_back(CI);
1414 bool Changed = false;
1415 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
1416 CallInst *CI = TailCalls[i];
1419 // Conservatively require the attributes of the call to match those of the
1420 // return. Ignore noalias because it doesn't affect the call sequence.
1421 AttributeSet CalleeAttrs = CS.getAttributes();
1422 if (AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1423 removeAttribute(Attribute::NoAlias) !=
1424 AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1425 removeAttribute(Attribute::NoAlias))
1428 // Make sure the call instruction is followed by an unconditional branch to
1429 // the return block.
1430 BasicBlock *CallBB = CI->getParent();
1431 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
1432 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
1435 // Duplicate the return into CallBB.
1436 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
1437 ModifiedDT = Changed = true;
1441 // If we eliminated all predecessors of the block, delete the block now.
1442 if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
1443 BB->eraseFromParent();
1448 //===----------------------------------------------------------------------===//
1449 // Memory Optimization
1450 //===----------------------------------------------------------------------===//
1454 /// ExtAddrMode - This is an extended version of TargetLowering::AddrMode
1455 /// which holds actual Value*'s for register values.
1456 struct ExtAddrMode : public TargetLowering::AddrMode {
1459 ExtAddrMode() : BaseReg(nullptr), ScaledReg(nullptr) {}
1460 void print(raw_ostream &OS) const;
1463 bool operator==(const ExtAddrMode& O) const {
1464 return (BaseReg == O.BaseReg) && (ScaledReg == O.ScaledReg) &&
1465 (BaseGV == O.BaseGV) && (BaseOffs == O.BaseOffs) &&
1466 (HasBaseReg == O.HasBaseReg) && (Scale == O.Scale);
1471 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
1477 void ExtAddrMode::print(raw_ostream &OS) const {
1478 bool NeedPlus = false;
1481 OS << (NeedPlus ? " + " : "")
1483 BaseGV->printAsOperand(OS, /*PrintType=*/false);
1488 OS << (NeedPlus ? " + " : "")
1494 OS << (NeedPlus ? " + " : "")
1496 BaseReg->printAsOperand(OS, /*PrintType=*/false);
1500 OS << (NeedPlus ? " + " : "")
1502 ScaledReg->printAsOperand(OS, /*PrintType=*/false);
1508 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1509 void ExtAddrMode::dump() const {
1515 /// \brief This class provides transaction based operation on the IR.
1516 /// Every change made through this class is recorded in the internal state and
1517 /// can be undone (rollback) until commit is called.
1518 class TypePromotionTransaction {
1520 /// \brief This represents the common interface of the individual transaction.
1521 /// Each class implements the logic for doing one specific modification on
1522 /// the IR via the TypePromotionTransaction.
1523 class TypePromotionAction {
1525 /// The Instruction modified.
1529 /// \brief Constructor of the action.
1530 /// The constructor performs the related action on the IR.
1531 TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
1533 virtual ~TypePromotionAction() {}
1535 /// \brief Undo the modification done by this action.
1536 /// When this method is called, the IR must be in the same state as it was
1537 /// before this action was applied.
1538 /// \pre Undoing the action works if and only if the IR is in the exact same
1539 /// state as it was directly after this action was applied.
1540 virtual void undo() = 0;
1542 /// \brief Advocate every change made by this action.
1543 /// When the results on the IR of the action are to be kept, it is important
1544 /// to call this function, otherwise hidden information may be kept forever.
1545 virtual void commit() {
1546 // Nothing to be done, this action is not doing anything.
1550 /// \brief Utility to remember the position of an instruction.
1551 class InsertionHandler {
1552 /// Position of an instruction.
1553 /// Either an instruction:
1554 /// - Is the first in a basic block: BB is used.
1555 /// - Has a previous instructon: PrevInst is used.
1557 Instruction *PrevInst;
1560 /// Remember whether or not the instruction had a previous instruction.
1561 bool HasPrevInstruction;
1564 /// \brief Record the position of \p Inst.
1565 InsertionHandler(Instruction *Inst) {
1566 BasicBlock::iterator It = Inst;
1567 HasPrevInstruction = (It != (Inst->getParent()->begin()));
1568 if (HasPrevInstruction)
1569 Point.PrevInst = --It;
1571 Point.BB = Inst->getParent();
1574 /// \brief Insert \p Inst at the recorded position.
1575 void insert(Instruction *Inst) {
1576 if (HasPrevInstruction) {
1577 if (Inst->getParent())
1578 Inst->removeFromParent();
1579 Inst->insertAfter(Point.PrevInst);
1581 Instruction *Position = Point.BB->getFirstInsertionPt();
1582 if (Inst->getParent())
1583 Inst->moveBefore(Position);
1585 Inst->insertBefore(Position);
1590 /// \brief Move an instruction before another.
1591 class InstructionMoveBefore : public TypePromotionAction {
1592 /// Original position of the instruction.
1593 InsertionHandler Position;
1596 /// \brief Move \p Inst before \p Before.
1597 InstructionMoveBefore(Instruction *Inst, Instruction *Before)
1598 : TypePromotionAction(Inst), Position(Inst) {
1599 DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before << "\n");
1600 Inst->moveBefore(Before);
1603 /// \brief Move the instruction back to its original position.
1604 void undo() override {
1605 DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
1606 Position.insert(Inst);
1610 /// \brief Set the operand of an instruction with a new value.
1611 class OperandSetter : public TypePromotionAction {
1612 /// Original operand of the instruction.
1614 /// Index of the modified instruction.
1618 /// \brief Set \p Idx operand of \p Inst with \p NewVal.
1619 OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
1620 : TypePromotionAction(Inst), Idx(Idx) {
1621 DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
1622 << "for:" << *Inst << "\n"
1623 << "with:" << *NewVal << "\n");
1624 Origin = Inst->getOperand(Idx);
1625 Inst->setOperand(Idx, NewVal);
1628 /// \brief Restore the original value of the instruction.
1629 void undo() override {
1630 DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
1631 << "for: " << *Inst << "\n"
1632 << "with: " << *Origin << "\n");
1633 Inst->setOperand(Idx, Origin);
1637 /// \brief Hide the operands of an instruction.
1638 /// Do as if this instruction was not using any of its operands.
1639 class OperandsHider : public TypePromotionAction {
1640 /// The list of original operands.
1641 SmallVector<Value *, 4> OriginalValues;
1644 /// \brief Remove \p Inst from the uses of the operands of \p Inst.
1645 OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
1646 DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
1647 unsigned NumOpnds = Inst->getNumOperands();
1648 OriginalValues.reserve(NumOpnds);
1649 for (unsigned It = 0; It < NumOpnds; ++It) {
1650 // Save the current operand.
1651 Value *Val = Inst->getOperand(It);
1652 OriginalValues.push_back(Val);
1654 // We could use OperandSetter here, but that would implied an overhead
1655 // that we are not willing to pay.
1656 Inst->setOperand(It, UndefValue::get(Val->getType()));
1660 /// \brief Restore the original list of uses.
1661 void undo() override {
1662 DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
1663 for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
1664 Inst->setOperand(It, OriginalValues[It]);
1668 /// \brief Build a truncate instruction.
1669 class TruncBuilder : public TypePromotionAction {
1672 /// \brief Build a truncate instruction of \p Opnd producing a \p Ty
1674 /// trunc Opnd to Ty.
1675 TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
1676 IRBuilder<> Builder(Opnd);
1677 Val = Builder.CreateTrunc(Opnd, Ty, "promoted");
1678 DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n");
1681 /// \brief Get the built value.
1682 Value *getBuiltValue() { return Val; }
1684 /// \brief Remove the built instruction.
1685 void undo() override {
1686 DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n");
1687 if (Instruction *IVal = dyn_cast<Instruction>(Val))
1688 IVal->eraseFromParent();
1692 /// \brief Build a sign extension instruction.
1693 class SExtBuilder : public TypePromotionAction {
1696 /// \brief Build a sign extension instruction of \p Opnd producing a \p Ty
1698 /// sext Opnd to Ty.
1699 SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
1700 : TypePromotionAction(InsertPt) {
1701 IRBuilder<> Builder(InsertPt);
1702 Val = Builder.CreateSExt(Opnd, Ty, "promoted");
1703 DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n");
1706 /// \brief Get the built value.
1707 Value *getBuiltValue() { return Val; }
1709 /// \brief Remove the built instruction.
1710 void undo() override {
1711 DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n");
1712 if (Instruction *IVal = dyn_cast<Instruction>(Val))
1713 IVal->eraseFromParent();
1717 /// \brief Build a zero extension instruction.
1718 class ZExtBuilder : public TypePromotionAction {
1721 /// \brief Build a zero extension instruction of \p Opnd producing a \p Ty
1723 /// zext Opnd to Ty.
1724 ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
1725 : TypePromotionAction(InsertPt) {
1726 IRBuilder<> Builder(InsertPt);
1727 Val = Builder.CreateZExt(Opnd, Ty, "promoted");
1728 DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n");
1731 /// \brief Get the built value.
1732 Value *getBuiltValue() { return Val; }
1734 /// \brief Remove the built instruction.
1735 void undo() override {
1736 DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n");
1737 if (Instruction *IVal = dyn_cast<Instruction>(Val))
1738 IVal->eraseFromParent();
1742 /// \brief Mutate an instruction to another type.
1743 class TypeMutator : public TypePromotionAction {
1744 /// Record the original type.
1748 /// \brief Mutate the type of \p Inst into \p NewTy.
1749 TypeMutator(Instruction *Inst, Type *NewTy)
1750 : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
1751 DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
1753 Inst->mutateType(NewTy);
1756 /// \brief Mutate the instruction back to its original type.
1757 void undo() override {
1758 DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
1760 Inst->mutateType(OrigTy);
1764 /// \brief Replace the uses of an instruction by another instruction.
1765 class UsesReplacer : public TypePromotionAction {
1766 /// Helper structure to keep track of the replaced uses.
1767 struct InstructionAndIdx {
1768 /// The instruction using the instruction.
1770 /// The index where this instruction is used for Inst.
1772 InstructionAndIdx(Instruction *Inst, unsigned Idx)
1773 : Inst(Inst), Idx(Idx) {}
1776 /// Keep track of the original uses (pair Instruction, Index).
1777 SmallVector<InstructionAndIdx, 4> OriginalUses;
1778 typedef SmallVectorImpl<InstructionAndIdx>::iterator use_iterator;
1781 /// \brief Replace all the use of \p Inst by \p New.
1782 UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
1783 DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
1785 // Record the original uses.
1786 for (Use &U : Inst->uses()) {
1787 Instruction *UserI = cast<Instruction>(U.getUser());
1788 OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
1790 // Now, we can replace the uses.
1791 Inst->replaceAllUsesWith(New);
1794 /// \brief Reassign the original uses of Inst to Inst.
1795 void undo() override {
1796 DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
1797 for (use_iterator UseIt = OriginalUses.begin(),
1798 EndIt = OriginalUses.end();
1799 UseIt != EndIt; ++UseIt) {
1800 UseIt->Inst->setOperand(UseIt->Idx, Inst);
1805 /// \brief Remove an instruction from the IR.
1806 class InstructionRemover : public TypePromotionAction {
1807 /// Original position of the instruction.
1808 InsertionHandler Inserter;
1809 /// Helper structure to hide all the link to the instruction. In other
1810 /// words, this helps to do as if the instruction was removed.
1811 OperandsHider Hider;
1812 /// Keep track of the uses replaced, if any.
1813 UsesReplacer *Replacer;
1816 /// \brief Remove all reference of \p Inst and optinally replace all its
1818 /// \pre If !Inst->use_empty(), then New != nullptr
1819 InstructionRemover(Instruction *Inst, Value *New = nullptr)
1820 : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
1823 Replacer = new UsesReplacer(Inst, New);
1824 DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
1825 Inst->removeFromParent();
1828 ~InstructionRemover() { delete Replacer; }
1830 /// \brief Really remove the instruction.
1831 void commit() override { delete Inst; }
1833 /// \brief Resurrect the instruction and reassign it to the proper uses if
1834 /// new value was provided when build this action.
1835 void undo() override {
1836 DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
1837 Inserter.insert(Inst);
1845 /// Restoration point.
1846 /// The restoration point is a pointer to an action instead of an iterator
1847 /// because the iterator may be invalidated but not the pointer.
1848 typedef const TypePromotionAction *ConstRestorationPt;
1849 /// Advocate every changes made in that transaction.
1851 /// Undo all the changes made after the given point.
1852 void rollback(ConstRestorationPt Point);
1853 /// Get the current restoration point.
1854 ConstRestorationPt getRestorationPoint() const;
1856 /// \name API for IR modification with state keeping to support rollback.
1858 /// Same as Instruction::setOperand.
1859 void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
1860 /// Same as Instruction::eraseFromParent.
1861 void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
1862 /// Same as Value::replaceAllUsesWith.
1863 void replaceAllUsesWith(Instruction *Inst, Value *New);
1864 /// Same as Value::mutateType.
1865 void mutateType(Instruction *Inst, Type *NewTy);
1866 /// Same as IRBuilder::createTrunc.
1867 Value *createTrunc(Instruction *Opnd, Type *Ty);
1868 /// Same as IRBuilder::createSExt.
1869 Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
1870 /// Same as IRBuilder::createZExt.
1871 Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
1872 /// Same as Instruction::moveBefore.
1873 void moveBefore(Instruction *Inst, Instruction *Before);
1877 /// The ordered list of actions made so far.
1878 SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions;
1879 typedef SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator CommitPt;
1882 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
1885 make_unique<TypePromotionTransaction::OperandSetter>(Inst, Idx, NewVal));
1888 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
1891 make_unique<TypePromotionTransaction::InstructionRemover>(Inst, NewVal));
1894 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
1896 Actions.push_back(make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
1899 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
1900 Actions.push_back(make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
1903 Value *TypePromotionTransaction::createTrunc(Instruction *Opnd,
1905 std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
1906 Value *Val = Ptr->getBuiltValue();
1907 Actions.push_back(std::move(Ptr));
1911 Value *TypePromotionTransaction::createSExt(Instruction *Inst,
1912 Value *Opnd, Type *Ty) {
1913 std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
1914 Value *Val = Ptr->getBuiltValue();
1915 Actions.push_back(std::move(Ptr));
1919 Value *TypePromotionTransaction::createZExt(Instruction *Inst,
1920 Value *Opnd, Type *Ty) {
1921 std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
1922 Value *Val = Ptr->getBuiltValue();
1923 Actions.push_back(std::move(Ptr));
1927 void TypePromotionTransaction::moveBefore(Instruction *Inst,
1928 Instruction *Before) {
1930 make_unique<TypePromotionTransaction::InstructionMoveBefore>(Inst, Before));
1933 TypePromotionTransaction::ConstRestorationPt
1934 TypePromotionTransaction::getRestorationPoint() const {
1935 return !Actions.empty() ? Actions.back().get() : nullptr;
1938 void TypePromotionTransaction::commit() {
1939 for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
1945 void TypePromotionTransaction::rollback(
1946 TypePromotionTransaction::ConstRestorationPt Point) {
1947 while (!Actions.empty() && Point != Actions.back().get()) {
1948 std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
1953 /// \brief A helper class for matching addressing modes.
1955 /// This encapsulates the logic for matching the target-legal addressing modes.
1956 class AddressingModeMatcher {
1957 SmallVectorImpl<Instruction*> &AddrModeInsts;
1958 const TargetLowering &TLI;
1960 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
1961 /// the memory instruction that we're computing this address for.
1963 Instruction *MemoryInst;
1965 /// AddrMode - This is the addressing mode that we're building up. This is
1966 /// part of the return value of this addressing mode matching stuff.
1967 ExtAddrMode &AddrMode;
1969 /// The truncate instruction inserted by other CodeGenPrepare optimizations.
1970 const SetOfInstrs &InsertedTruncs;
1971 /// A map from the instructions to their type before promotion.
1972 InstrToOrigTy &PromotedInsts;
1973 /// The ongoing transaction where every action should be registered.
1974 TypePromotionTransaction &TPT;
1976 /// IgnoreProfitability - This is set to true when we should not do
1977 /// profitability checks. When true, IsProfitableToFoldIntoAddressingMode
1978 /// always returns true.
1979 bool IgnoreProfitability;
1981 AddressingModeMatcher(SmallVectorImpl<Instruction*> &AMI,
1982 const TargetLowering &T, Type *AT,
1983 Instruction *MI, ExtAddrMode &AM,
1984 const SetOfInstrs &InsertedTruncs,
1985 InstrToOrigTy &PromotedInsts,
1986 TypePromotionTransaction &TPT)
1987 : AddrModeInsts(AMI), TLI(T), AccessTy(AT), MemoryInst(MI), AddrMode(AM),
1988 InsertedTruncs(InsertedTruncs), PromotedInsts(PromotedInsts), TPT(TPT) {
1989 IgnoreProfitability = false;
1993 /// Match - Find the maximal addressing mode that a load/store of V can fold,
1994 /// give an access type of AccessTy. This returns a list of involved
1995 /// instructions in AddrModeInsts.
1996 /// \p InsertedTruncs The truncate instruction inserted by other
1999 /// \p PromotedInsts maps the instructions to their type before promotion.
2000 /// \p The ongoing transaction where every action should be registered.
2001 static ExtAddrMode Match(Value *V, Type *AccessTy,
2002 Instruction *MemoryInst,
2003 SmallVectorImpl<Instruction*> &AddrModeInsts,
2004 const TargetLowering &TLI,
2005 const SetOfInstrs &InsertedTruncs,
2006 InstrToOrigTy &PromotedInsts,
2007 TypePromotionTransaction &TPT) {
2010 bool Success = AddressingModeMatcher(AddrModeInsts, TLI, AccessTy,
2011 MemoryInst, Result, InsertedTruncs,
2012 PromotedInsts, TPT).MatchAddr(V, 0);
2013 (void)Success; assert(Success && "Couldn't select *anything*?");
2017 bool MatchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
2018 bool MatchAddr(Value *V, unsigned Depth);
2019 bool MatchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth,
2020 bool *MovedAway = nullptr);
2021 bool IsProfitableToFoldIntoAddressingMode(Instruction *I,
2022 ExtAddrMode &AMBefore,
2023 ExtAddrMode &AMAfter);
2024 bool ValueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
2025 bool IsPromotionProfitable(unsigned MatchedSize, unsigned SizeWithPromotion,
2026 Value *PromotedOperand) const;
2029 /// MatchScaledValue - Try adding ScaleReg*Scale to the current addressing mode.
2030 /// Return true and update AddrMode if this addr mode is legal for the target,
2032 bool AddressingModeMatcher::MatchScaledValue(Value *ScaleReg, int64_t Scale,
2034 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
2035 // mode. Just process that directly.
2037 return MatchAddr(ScaleReg, Depth);
2039 // If the scale is 0, it takes nothing to add this.
2043 // If we already have a scale of this value, we can add to it, otherwise, we
2044 // need an available scale field.
2045 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
2048 ExtAddrMode TestAddrMode = AddrMode;
2050 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
2051 // [A+B + A*7] -> [B+A*8].
2052 TestAddrMode.Scale += Scale;
2053 TestAddrMode.ScaledReg = ScaleReg;
2055 // If the new address isn't legal, bail out.
2056 if (!TLI.isLegalAddressingMode(TestAddrMode, AccessTy))
2059 // It was legal, so commit it.
2060 AddrMode = TestAddrMode;
2062 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
2063 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
2064 // X*Scale + C*Scale to addr mode.
2065 ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
2066 if (isa<Instruction>(ScaleReg) && // not a constant expr.
2067 match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
2068 TestAddrMode.ScaledReg = AddLHS;
2069 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
2071 // If this addressing mode is legal, commit it and remember that we folded
2072 // this instruction.
2073 if (TLI.isLegalAddressingMode(TestAddrMode, AccessTy)) {
2074 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
2075 AddrMode = TestAddrMode;
2080 // Otherwise, not (x+c)*scale, just return what we have.
2084 /// MightBeFoldableInst - This is a little filter, which returns true if an
2085 /// addressing computation involving I might be folded into a load/store
2086 /// accessing it. This doesn't need to be perfect, but needs to accept at least
2087 /// the set of instructions that MatchOperationAddr can.
2088 static bool MightBeFoldableInst(Instruction *I) {
2089 switch (I->getOpcode()) {
2090 case Instruction::BitCast:
2091 case Instruction::AddrSpaceCast:
2092 // Don't touch identity bitcasts.
2093 if (I->getType() == I->getOperand(0)->getType())
2095 return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
2096 case Instruction::PtrToInt:
2097 // PtrToInt is always a noop, as we know that the int type is pointer sized.
2099 case Instruction::IntToPtr:
2100 // We know the input is intptr_t, so this is foldable.
2102 case Instruction::Add:
2104 case Instruction::Mul:
2105 case Instruction::Shl:
2106 // Can only handle X*C and X << C.
2107 return isa<ConstantInt>(I->getOperand(1));
2108 case Instruction::GetElementPtr:
2115 /// \brief Check whether or not \p Val is a legal instruction for \p TLI.
2116 /// \note \p Val is assumed to be the product of some type promotion.
2117 /// Therefore if \p Val has an undefined state in \p TLI, this is assumed
2118 /// to be legal, as the non-promoted value would have had the same state.
2119 static bool isPromotedInstructionLegal(const TargetLowering &TLI, Value *Val) {
2120 Instruction *PromotedInst = dyn_cast<Instruction>(Val);
2123 int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
2124 // If the ISDOpcode is undefined, it was undefined before the promotion.
2127 // Otherwise, check if the promoted instruction is legal or not.
2128 return TLI.isOperationLegalOrCustom(
2129 ISDOpcode, TLI.getValueType(PromotedInst->getType()));
2132 /// \brief Hepler class to perform type promotion.
2133 class TypePromotionHelper {
2134 /// \brief Utility function to check whether or not a sign or zero extension
2135 /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
2136 /// either using the operands of \p Inst or promoting \p Inst.
2137 /// The type of the extension is defined by \p IsSExt.
2138 /// In other words, check if:
2139 /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
2140 /// #1 Promotion applies:
2141 /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
2142 /// #2 Operand reuses:
2143 /// ext opnd1 to ConsideredExtType.
2144 /// \p PromotedInsts maps the instructions to their type before promotion.
2145 static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
2146 const InstrToOrigTy &PromotedInsts, bool IsSExt);
2148 /// \brief Utility function to determine if \p OpIdx should be promoted when
2149 /// promoting \p Inst.
2150 static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
2151 if (isa<SelectInst>(Inst) && OpIdx == 0)
2156 /// \brief Utility function to promote the operand of \p Ext when this
2157 /// operand is a promotable trunc or sext or zext.
2158 /// \p PromotedInsts maps the instructions to their type before promotion.
2159 /// \p CreatedInsts[out] contains how many non-free instructions have been
2160 /// created to promote the operand of Ext.
2161 /// Newly added extensions are inserted in \p Exts.
2162 /// Newly added truncates are inserted in \p Truncs.
2163 /// Should never be called directly.
2164 /// \return The promoted value which is used instead of Ext.
2165 static Value *promoteOperandForTruncAndAnyExt(
2166 Instruction *Ext, TypePromotionTransaction &TPT,
2167 InstrToOrigTy &PromotedInsts, unsigned &CreatedInsts,
2168 SmallVectorImpl<Instruction *> *Exts,
2169 SmallVectorImpl<Instruction *> *Truncs);
2171 /// \brief Utility function to promote the operand of \p Ext when this
2172 /// operand is promotable and is not a supported trunc or sext.
2173 /// \p PromotedInsts maps the instructions to their type before promotion.
2174 /// \p CreatedInsts[out] contains how many non-free instructions have been
2175 /// created to promote the operand of Ext.
2176 /// Newly added extensions are inserted in \p Exts.
2177 /// Newly added truncates are inserted in \p Truncs.
2178 /// Should never be called directly.
2179 /// \return The promoted value which is used instead of Ext.
2181 promoteOperandForOther(Instruction *Ext, TypePromotionTransaction &TPT,
2182 InstrToOrigTy &PromotedInsts, unsigned &CreatedInsts,
2183 SmallVectorImpl<Instruction *> *Exts,
2184 SmallVectorImpl<Instruction *> *Truncs, bool IsSExt);
2186 /// \see promoteOperandForOther.
2188 signExtendOperandForOther(Instruction *Ext, TypePromotionTransaction &TPT,
2189 InstrToOrigTy &PromotedInsts,
2190 unsigned &CreatedInsts,
2191 SmallVectorImpl<Instruction *> *Exts,
2192 SmallVectorImpl<Instruction *> *Truncs) {
2193 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInsts, Exts,
2197 /// \see promoteOperandForOther.
2199 zeroExtendOperandForOther(Instruction *Ext, TypePromotionTransaction &TPT,
2200 InstrToOrigTy &PromotedInsts,
2201 unsigned &CreatedInsts,
2202 SmallVectorImpl<Instruction *> *Exts,
2203 SmallVectorImpl<Instruction *> *Truncs) {
2204 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInsts, Exts,
2209 /// Type for the utility function that promotes the operand of Ext.
2210 typedef Value *(*Action)(Instruction *Ext, TypePromotionTransaction &TPT,
2211 InstrToOrigTy &PromotedInsts, unsigned &CreatedInsts,
2212 SmallVectorImpl<Instruction *> *Exts,
2213 SmallVectorImpl<Instruction *> *Truncs);
2214 /// \brief Given a sign/zero extend instruction \p Ext, return the approriate
2215 /// action to promote the operand of \p Ext instead of using Ext.
2216 /// \return NULL if no promotable action is possible with the current
2218 /// \p InsertedTruncs keeps track of all the truncate instructions inserted by
2219 /// the others CodeGenPrepare optimizations. This information is important
2220 /// because we do not want to promote these instructions as CodeGenPrepare
2221 /// will reinsert them later. Thus creating an infinite loop: create/remove.
2222 /// \p PromotedInsts maps the instructions to their type before promotion.
2223 static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedTruncs,
2224 const TargetLowering &TLI,
2225 const InstrToOrigTy &PromotedInsts);
2228 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
2229 Type *ConsideredExtType,
2230 const InstrToOrigTy &PromotedInsts,
2232 // The promotion helper does not know how to deal with vector types yet.
2233 // To be able to fix that, we would need to fix the places where we
2234 // statically extend, e.g., constants and such.
2235 if (Inst->getType()->isVectorTy())
2238 // We can always get through zext.
2239 if (isa<ZExtInst>(Inst))
2242 // sext(sext) is ok too.
2243 if (IsSExt && isa<SExtInst>(Inst))
2246 // We can get through binary operator, if it is legal. In other words, the
2247 // binary operator must have a nuw or nsw flag.
2248 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
2249 if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
2250 ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
2251 (IsSExt && BinOp->hasNoSignedWrap())))
2254 // Check if we can do the following simplification.
2255 // ext(trunc(opnd)) --> ext(opnd)
2256 if (!isa<TruncInst>(Inst))
2259 Value *OpndVal = Inst->getOperand(0);
2260 // Check if we can use this operand in the extension.
2261 // If the type is larger than the result type of the extension,
2263 if (!OpndVal->getType()->isIntegerTy() ||
2264 OpndVal->getType()->getIntegerBitWidth() >
2265 ConsideredExtType->getIntegerBitWidth())
2268 // If the operand of the truncate is not an instruction, we will not have
2269 // any information on the dropped bits.
2270 // (Actually we could for constant but it is not worth the extra logic).
2271 Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
2275 // Check if the source of the type is narrow enough.
2276 // I.e., check that trunc just drops extended bits of the same kind of
2278 // #1 get the type of the operand and check the kind of the extended bits.
2279 const Type *OpndType;
2280 InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
2281 if (It != PromotedInsts.end() && It->second.IsSExt == IsSExt)
2282 OpndType = It->second.Ty;
2283 else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd)))
2284 OpndType = Opnd->getOperand(0)->getType();
2288 // #2 check that the truncate just drop extended bits.
2289 if (Inst->getType()->getIntegerBitWidth() >= OpndType->getIntegerBitWidth())
2295 TypePromotionHelper::Action TypePromotionHelper::getAction(
2296 Instruction *Ext, const SetOfInstrs &InsertedTruncs,
2297 const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
2298 assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
2299 "Unexpected instruction type");
2300 Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0));
2301 Type *ExtTy = Ext->getType();
2302 bool IsSExt = isa<SExtInst>(Ext);
2303 // If the operand of the extension is not an instruction, we cannot
2305 // If it, check we can get through.
2306 if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt))
2309 // Do not promote if the operand has been added by codegenprepare.
2310 // Otherwise, it means we are undoing an optimization that is likely to be
2311 // redone, thus causing potential infinite loop.
2312 if (isa<TruncInst>(ExtOpnd) && InsertedTruncs.count(ExtOpnd))
2315 // SExt or Trunc instructions.
2316 // Return the related handler.
2317 if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) ||
2318 isa<ZExtInst>(ExtOpnd))
2319 return promoteOperandForTruncAndAnyExt;
2321 // Regular instruction.
2322 // Abort early if we will have to insert non-free instructions.
2323 if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType()))
2325 return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
2328 Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
2329 llvm::Instruction *SExt, TypePromotionTransaction &TPT,
2330 InstrToOrigTy &PromotedInsts, unsigned &CreatedInsts,
2331 SmallVectorImpl<Instruction *> *Exts,
2332 SmallVectorImpl<Instruction *> *Truncs) {
2333 // By construction, the operand of SExt is an instruction. Otherwise we cannot
2334 // get through it and this method should not be called.
2335 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
2336 Value *ExtVal = SExt;
2337 if (isa<ZExtInst>(SExtOpnd)) {
2338 // Replace s|zext(zext(opnd))
2341 TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
2342 TPT.replaceAllUsesWith(SExt, ZExt);
2343 TPT.eraseInstruction(SExt);
2346 // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
2348 TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
2352 // Remove dead code.
2353 if (SExtOpnd->use_empty())
2354 TPT.eraseInstruction(SExtOpnd);
2356 // Check if the extension is still needed.
2357 Instruction *ExtInst = dyn_cast<Instruction>(ExtVal);
2358 if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) {
2359 if (ExtInst && Exts)
2360 Exts->push_back(ExtInst);
2364 // At this point we have: ext ty opnd to ty.
2365 // Reassign the uses of ExtInst to the opnd and remove ExtInst.
2366 Value *NextVal = ExtInst->getOperand(0);
2367 TPT.eraseInstruction(ExtInst, NextVal);
2371 Value *TypePromotionHelper::promoteOperandForOther(
2372 Instruction *Ext, TypePromotionTransaction &TPT,
2373 InstrToOrigTy &PromotedInsts, unsigned &CreatedInsts,
2374 SmallVectorImpl<Instruction *> *Exts,
2375 SmallVectorImpl<Instruction *> *Truncs, bool IsSExt) {
2376 // By construction, the operand of Ext is an instruction. Otherwise we cannot
2377 // get through it and this method should not be called.
2378 Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0));
2380 if (!ExtOpnd->hasOneUse()) {
2381 // ExtOpnd will be promoted.
2382 // All its uses, but Ext, will need to use a truncated value of the
2383 // promoted version.
2384 // Create the truncate now.
2385 Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType());
2386 if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) {
2387 ITrunc->removeFromParent();
2388 // Insert it just after the definition.
2389 ITrunc->insertAfter(ExtOpnd);
2391 Truncs->push_back(ITrunc);
2394 TPT.replaceAllUsesWith(ExtOpnd, Trunc);
2395 // Restore the operand of Ext (which has been replace by the previous call
2396 // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
2397 TPT.setOperand(Ext, 0, ExtOpnd);
2400 // Get through the Instruction:
2401 // 1. Update its type.
2402 // 2. Replace the uses of Ext by Inst.
2403 // 3. Extend each operand that needs to be extended.
2405 // Remember the original type of the instruction before promotion.
2406 // This is useful to know that the high bits are sign extended bits.
2407 PromotedInsts.insert(std::pair<Instruction *, TypeIsSExt>(
2408 ExtOpnd, TypeIsSExt(ExtOpnd->getType(), IsSExt)));
2410 TPT.mutateType(ExtOpnd, Ext->getType());
2412 TPT.replaceAllUsesWith(Ext, ExtOpnd);
2414 Instruction *ExtForOpnd = Ext;
2416 DEBUG(dbgs() << "Propagate Ext to operands\n");
2417 for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
2419 DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n');
2420 if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() ||
2421 !shouldExtOperand(ExtOpnd, OpIdx)) {
2422 DEBUG(dbgs() << "No need to propagate\n");
2425 // Check if we can statically extend the operand.
2426 Value *Opnd = ExtOpnd->getOperand(OpIdx);
2427 if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
2428 DEBUG(dbgs() << "Statically extend\n");
2429 unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
2430 APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth)
2431 : Cst->getValue().zext(BitWidth);
2432 TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal));
2435 // UndefValue are typed, so we have to statically sign extend them.
2436 if (isa<UndefValue>(Opnd)) {
2437 DEBUG(dbgs() << "Statically extend\n");
2438 TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType()));
2442 // Otherwise we have to explicity sign extend the operand.
2443 // Check if Ext was reused to extend an operand.
2445 // If yes, create a new one.
2446 DEBUG(dbgs() << "More operands to ext\n");
2447 Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType())
2448 : TPT.createZExt(Ext, Opnd, Ext->getType());
2449 if (!isa<Instruction>(ValForExtOpnd)) {
2450 TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd);
2453 ExtForOpnd = cast<Instruction>(ValForExtOpnd);
2457 Exts->push_back(ExtForOpnd);
2458 TPT.setOperand(ExtForOpnd, 0, Opnd);
2460 // Move the sign extension before the insertion point.
2461 TPT.moveBefore(ExtForOpnd, ExtOpnd);
2462 TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd);
2463 // If more sext are required, new instructions will have to be created.
2464 ExtForOpnd = nullptr;
2466 if (ExtForOpnd == Ext) {
2467 DEBUG(dbgs() << "Extension is useless now\n");
2468 TPT.eraseInstruction(Ext);
2473 /// IsPromotionProfitable - Check whether or not promoting an instruction
2474 /// to a wider type was profitable.
2475 /// \p MatchedSize gives the number of instructions that have been matched
2476 /// in the addressing mode after the promotion was applied.
2477 /// \p SizeWithPromotion gives the number of created instructions for
2478 /// the promotion plus the number of instructions that have been
2479 /// matched in the addressing mode before the promotion.
2480 /// \p PromotedOperand is the value that has been promoted.
2481 /// \return True if the promotion is profitable, false otherwise.
2483 AddressingModeMatcher::IsPromotionProfitable(unsigned MatchedSize,
2484 unsigned SizeWithPromotion,
2485 Value *PromotedOperand) const {
2486 // We folded less instructions than what we created to promote the operand.
2487 // This is not profitable.
2488 if (MatchedSize < SizeWithPromotion)
2490 if (MatchedSize > SizeWithPromotion)
2492 // The promotion is neutral but it may help folding the sign extension in
2493 // loads for instance.
2494 // Check that we did not create an illegal instruction.
2495 return isPromotedInstructionLegal(TLI, PromotedOperand);
2498 /// MatchOperationAddr - Given an instruction or constant expr, see if we can
2499 /// fold the operation into the addressing mode. If so, update the addressing
2500 /// mode and return true, otherwise return false without modifying AddrMode.
2501 /// If \p MovedAway is not NULL, it contains the information of whether or
2502 /// not AddrInst has to be folded into the addressing mode on success.
2503 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
2504 /// because it has been moved away.
2505 /// Thus AddrInst must not be added in the matched instructions.
2506 /// This state can happen when AddrInst is a sext, since it may be moved away.
2507 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
2508 /// not be referenced anymore.
2509 bool AddressingModeMatcher::MatchOperationAddr(User *AddrInst, unsigned Opcode,
2512 // Avoid exponential behavior on extremely deep expression trees.
2513 if (Depth >= 5) return false;
2515 // By default, all matched instructions stay in place.
2520 case Instruction::PtrToInt:
2521 // PtrToInt is always a noop, as we know that the int type is pointer sized.
2522 return MatchAddr(AddrInst->getOperand(0), Depth);
2523 case Instruction::IntToPtr:
2524 // This inttoptr is a no-op if the integer type is pointer sized.
2525 if (TLI.getValueType(AddrInst->getOperand(0)->getType()) ==
2526 TLI.getPointerTy(AddrInst->getType()->getPointerAddressSpace()))
2527 return MatchAddr(AddrInst->getOperand(0), Depth);
2529 case Instruction::BitCast:
2530 case Instruction::AddrSpaceCast:
2531 // BitCast is always a noop, and we can handle it as long as it is
2532 // int->int or pointer->pointer (we don't want int<->fp or something).
2533 if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
2534 AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
2535 // Don't touch identity bitcasts. These were probably put here by LSR,
2536 // and we don't want to mess around with them. Assume it knows what it
2538 AddrInst->getOperand(0)->getType() != AddrInst->getType())
2539 return MatchAddr(AddrInst->getOperand(0), Depth);
2541 case Instruction::Add: {
2542 // Check to see if we can merge in the RHS then the LHS. If so, we win.
2543 ExtAddrMode BackupAddrMode = AddrMode;
2544 unsigned OldSize = AddrModeInsts.size();
2545 // Start a transaction at this point.
2546 // The LHS may match but not the RHS.
2547 // Therefore, we need a higher level restoration point to undo partially
2548 // matched operation.
2549 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2550 TPT.getRestorationPoint();
2552 if (MatchAddr(AddrInst->getOperand(1), Depth+1) &&
2553 MatchAddr(AddrInst->getOperand(0), Depth+1))
2556 // Restore the old addr mode info.
2557 AddrMode = BackupAddrMode;
2558 AddrModeInsts.resize(OldSize);
2559 TPT.rollback(LastKnownGood);
2561 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
2562 if (MatchAddr(AddrInst->getOperand(0), Depth+1) &&
2563 MatchAddr(AddrInst->getOperand(1), Depth+1))
2566 // Otherwise we definitely can't merge the ADD in.
2567 AddrMode = BackupAddrMode;
2568 AddrModeInsts.resize(OldSize);
2569 TPT.rollback(LastKnownGood);
2572 //case Instruction::Or:
2573 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
2575 case Instruction::Mul:
2576 case Instruction::Shl: {
2577 // Can only handle X*C and X << C.
2578 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
2581 int64_t Scale = RHS->getSExtValue();
2582 if (Opcode == Instruction::Shl)
2583 Scale = 1LL << Scale;
2585 return MatchScaledValue(AddrInst->getOperand(0), Scale, Depth);
2587 case Instruction::GetElementPtr: {
2588 // Scan the GEP. We check it if it contains constant offsets and at most
2589 // one variable offset.
2590 int VariableOperand = -1;
2591 unsigned VariableScale = 0;
2593 int64_t ConstantOffset = 0;
2594 const DataLayout *TD = TLI.getDataLayout();
2595 gep_type_iterator GTI = gep_type_begin(AddrInst);
2596 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
2597 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
2598 const StructLayout *SL = TD->getStructLayout(STy);
2600 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
2601 ConstantOffset += SL->getElementOffset(Idx);
2603 uint64_t TypeSize = TD->getTypeAllocSize(GTI.getIndexedType());
2604 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
2605 ConstantOffset += CI->getSExtValue()*TypeSize;
2606 } else if (TypeSize) { // Scales of zero don't do anything.
2607 // We only allow one variable index at the moment.
2608 if (VariableOperand != -1)
2611 // Remember the variable index.
2612 VariableOperand = i;
2613 VariableScale = TypeSize;
2618 // A common case is for the GEP to only do a constant offset. In this case,
2619 // just add it to the disp field and check validity.
2620 if (VariableOperand == -1) {
2621 AddrMode.BaseOffs += ConstantOffset;
2622 if (ConstantOffset == 0 || TLI.isLegalAddressingMode(AddrMode, AccessTy)){
2623 // Check to see if we can fold the base pointer in too.
2624 if (MatchAddr(AddrInst->getOperand(0), Depth+1))
2627 AddrMode.BaseOffs -= ConstantOffset;
2631 // Save the valid addressing mode in case we can't match.
2632 ExtAddrMode BackupAddrMode = AddrMode;
2633 unsigned OldSize = AddrModeInsts.size();
2635 // See if the scale and offset amount is valid for this target.
2636 AddrMode.BaseOffs += ConstantOffset;
2638 // Match the base operand of the GEP.
2639 if (!MatchAddr(AddrInst->getOperand(0), Depth+1)) {
2640 // If it couldn't be matched, just stuff the value in a register.
2641 if (AddrMode.HasBaseReg) {
2642 AddrMode = BackupAddrMode;
2643 AddrModeInsts.resize(OldSize);
2646 AddrMode.HasBaseReg = true;
2647 AddrMode.BaseReg = AddrInst->getOperand(0);
2650 // Match the remaining variable portion of the GEP.
2651 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
2653 // If it couldn't be matched, try stuffing the base into a register
2654 // instead of matching it, and retrying the match of the scale.
2655 AddrMode = BackupAddrMode;
2656 AddrModeInsts.resize(OldSize);
2657 if (AddrMode.HasBaseReg)
2659 AddrMode.HasBaseReg = true;
2660 AddrMode.BaseReg = AddrInst->getOperand(0);
2661 AddrMode.BaseOffs += ConstantOffset;
2662 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand),
2663 VariableScale, Depth)) {
2664 // If even that didn't work, bail.
2665 AddrMode = BackupAddrMode;
2666 AddrModeInsts.resize(OldSize);
2673 case Instruction::SExt:
2674 case Instruction::ZExt: {
2675 Instruction *Ext = dyn_cast<Instruction>(AddrInst);
2679 // Try to move this ext out of the way of the addressing mode.
2680 // Ask for a method for doing so.
2681 TypePromotionHelper::Action TPH =
2682 TypePromotionHelper::getAction(Ext, InsertedTruncs, TLI, PromotedInsts);
2686 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2687 TPT.getRestorationPoint();
2688 unsigned CreatedInsts = 0;
2689 Value *PromotedOperand =
2690 TPH(Ext, TPT, PromotedInsts, CreatedInsts, nullptr, nullptr);
2691 // SExt has been moved away.
2692 // Thus either it will be rematched later in the recursive calls or it is
2693 // gone. Anyway, we must not fold it into the addressing mode at this point.
2697 // addr = gep base, idx
2699 // promotedOpnd = ext opnd <- no match here
2700 // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
2701 // addr = gep base, op <- match
2705 assert(PromotedOperand &&
2706 "TypePromotionHelper should have filtered out those cases");
2708 ExtAddrMode BackupAddrMode = AddrMode;
2709 unsigned OldSize = AddrModeInsts.size();
2711 if (!MatchAddr(PromotedOperand, Depth) ||
2712 !IsPromotionProfitable(AddrModeInsts.size(), OldSize + CreatedInsts,
2714 AddrMode = BackupAddrMode;
2715 AddrModeInsts.resize(OldSize);
2716 DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
2717 TPT.rollback(LastKnownGood);
2726 /// MatchAddr - If we can, try to add the value of 'Addr' into the current
2727 /// addressing mode. If Addr can't be added to AddrMode this returns false and
2728 /// leaves AddrMode unmodified. This assumes that Addr is either a pointer type
2729 /// or intptr_t for the target.
2731 bool AddressingModeMatcher::MatchAddr(Value *Addr, unsigned Depth) {
2732 // Start a transaction at this point that we will rollback if the matching
2734 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2735 TPT.getRestorationPoint();
2736 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
2737 // Fold in immediates if legal for the target.
2738 AddrMode.BaseOffs += CI->getSExtValue();
2739 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2741 AddrMode.BaseOffs -= CI->getSExtValue();
2742 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
2743 // If this is a global variable, try to fold it into the addressing mode.
2744 if (!AddrMode.BaseGV) {
2745 AddrMode.BaseGV = GV;
2746 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2748 AddrMode.BaseGV = nullptr;
2750 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
2751 ExtAddrMode BackupAddrMode = AddrMode;
2752 unsigned OldSize = AddrModeInsts.size();
2754 // Check to see if it is possible to fold this operation.
2755 bool MovedAway = false;
2756 if (MatchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
2757 // This instruction may have been move away. If so, there is nothing
2761 // Okay, it's possible to fold this. Check to see if it is actually
2762 // *profitable* to do so. We use a simple cost model to avoid increasing
2763 // register pressure too much.
2764 if (I->hasOneUse() ||
2765 IsProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
2766 AddrModeInsts.push_back(I);
2770 // It isn't profitable to do this, roll back.
2771 //cerr << "NOT FOLDING: " << *I;
2772 AddrMode = BackupAddrMode;
2773 AddrModeInsts.resize(OldSize);
2774 TPT.rollback(LastKnownGood);
2776 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
2777 if (MatchOperationAddr(CE, CE->getOpcode(), Depth))
2779 TPT.rollback(LastKnownGood);
2780 } else if (isa<ConstantPointerNull>(Addr)) {
2781 // Null pointer gets folded without affecting the addressing mode.
2785 // Worse case, the target should support [reg] addressing modes. :)
2786 if (!AddrMode.HasBaseReg) {
2787 AddrMode.HasBaseReg = true;
2788 AddrMode.BaseReg = Addr;
2789 // Still check for legality in case the target supports [imm] but not [i+r].
2790 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2792 AddrMode.HasBaseReg = false;
2793 AddrMode.BaseReg = nullptr;
2796 // If the base register is already taken, see if we can do [r+r].
2797 if (AddrMode.Scale == 0) {
2799 AddrMode.ScaledReg = Addr;
2800 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2803 AddrMode.ScaledReg = nullptr;
2806 TPT.rollback(LastKnownGood);
2810 /// IsOperandAMemoryOperand - Check to see if all uses of OpVal by the specified
2811 /// inline asm call are due to memory operands. If so, return true, otherwise
2813 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
2814 const TargetLowering &TLI) {
2815 TargetLowering::AsmOperandInfoVector TargetConstraints = TLI.ParseConstraints(ImmutableCallSite(CI));
2816 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
2817 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
2819 // Compute the constraint code and ConstraintType to use.
2820 TLI.ComputeConstraintToUse(OpInfo, SDValue());
2822 // If this asm operand is our Value*, and if it isn't an indirect memory
2823 // operand, we can't fold it!
2824 if (OpInfo.CallOperandVal == OpVal &&
2825 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
2826 !OpInfo.isIndirect))
2833 /// FindAllMemoryUses - Recursively walk all the uses of I until we find a
2834 /// memory use. If we find an obviously non-foldable instruction, return true.
2835 /// Add the ultimately found memory instructions to MemoryUses.
2836 static bool FindAllMemoryUses(Instruction *I,
2837 SmallVectorImpl<std::pair<Instruction*,unsigned> > &MemoryUses,
2838 SmallPtrSetImpl<Instruction*> &ConsideredInsts,
2839 const TargetLowering &TLI) {
2840 // If we already considered this instruction, we're done.
2841 if (!ConsideredInsts.insert(I).second)
2844 // If this is an obviously unfoldable instruction, bail out.
2845 if (!MightBeFoldableInst(I))
2848 // Loop over all the uses, recursively processing them.
2849 for (Use &U : I->uses()) {
2850 Instruction *UserI = cast<Instruction>(U.getUser());
2852 if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
2853 MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
2857 if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
2858 unsigned opNo = U.getOperandNo();
2859 if (opNo == 0) return true; // Storing addr, not into addr.
2860 MemoryUses.push_back(std::make_pair(SI, opNo));
2864 if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
2865 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
2866 if (!IA) return true;
2868 // If this is a memory operand, we're cool, otherwise bail out.
2869 if (!IsOperandAMemoryOperand(CI, IA, I, TLI))
2874 if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TLI))
2881 /// ValueAlreadyLiveAtInst - Retrn true if Val is already known to be live at
2882 /// the use site that we're folding it into. If so, there is no cost to
2883 /// include it in the addressing mode. KnownLive1 and KnownLive2 are two values
2884 /// that we know are live at the instruction already.
2885 bool AddressingModeMatcher::ValueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
2886 Value *KnownLive2) {
2887 // If Val is either of the known-live values, we know it is live!
2888 if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
2891 // All values other than instructions and arguments (e.g. constants) are live.
2892 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
2894 // If Val is a constant sized alloca in the entry block, it is live, this is
2895 // true because it is just a reference to the stack/frame pointer, which is
2896 // live for the whole function.
2897 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
2898 if (AI->isStaticAlloca())
2901 // Check to see if this value is already used in the memory instruction's
2902 // block. If so, it's already live into the block at the very least, so we
2903 // can reasonably fold it.
2904 return Val->isUsedInBasicBlock(MemoryInst->getParent());
2907 /// IsProfitableToFoldIntoAddressingMode - It is possible for the addressing
2908 /// mode of the machine to fold the specified instruction into a load or store
2909 /// that ultimately uses it. However, the specified instruction has multiple
2910 /// uses. Given this, it may actually increase register pressure to fold it
2911 /// into the load. For example, consider this code:
2915 /// use(Y) -> nonload/store
2919 /// In this case, Y has multiple uses, and can be folded into the load of Z
2920 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
2921 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
2922 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
2923 /// number of computations either.
2925 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
2926 /// X was live across 'load Z' for other reasons, we actually *would* want to
2927 /// fold the addressing mode in the Z case. This would make Y die earlier.
2928 bool AddressingModeMatcher::
2929 IsProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
2930 ExtAddrMode &AMAfter) {
2931 if (IgnoreProfitability) return true;
2933 // AMBefore is the addressing mode before this instruction was folded into it,
2934 // and AMAfter is the addressing mode after the instruction was folded. Get
2935 // the set of registers referenced by AMAfter and subtract out those
2936 // referenced by AMBefore: this is the set of values which folding in this
2937 // address extends the lifetime of.
2939 // Note that there are only two potential values being referenced here,
2940 // BaseReg and ScaleReg (global addresses are always available, as are any
2941 // folded immediates).
2942 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
2944 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
2945 // lifetime wasn't extended by adding this instruction.
2946 if (ValueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
2948 if (ValueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
2949 ScaledReg = nullptr;
2951 // If folding this instruction (and it's subexprs) didn't extend any live
2952 // ranges, we're ok with it.
2953 if (!BaseReg && !ScaledReg)
2956 // If all uses of this instruction are ultimately load/store/inlineasm's,
2957 // check to see if their addressing modes will include this instruction. If
2958 // so, we can fold it into all uses, so it doesn't matter if it has multiple
2960 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
2961 SmallPtrSet<Instruction*, 16> ConsideredInsts;
2962 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TLI))
2963 return false; // Has a non-memory, non-foldable use!
2965 // Now that we know that all uses of this instruction are part of a chain of
2966 // computation involving only operations that could theoretically be folded
2967 // into a memory use, loop over each of these uses and see if they could
2968 // *actually* fold the instruction.
2969 SmallVector<Instruction*, 32> MatchedAddrModeInsts;
2970 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
2971 Instruction *User = MemoryUses[i].first;
2972 unsigned OpNo = MemoryUses[i].second;
2974 // Get the access type of this use. If the use isn't a pointer, we don't
2975 // know what it accesses.
2976 Value *Address = User->getOperand(OpNo);
2977 if (!Address->getType()->isPointerTy())
2979 Type *AddressAccessTy = Address->getType()->getPointerElementType();
2981 // Do a match against the root of this address, ignoring profitability. This
2982 // will tell us if the addressing mode for the memory operation will
2983 // *actually* cover the shared instruction.
2985 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2986 TPT.getRestorationPoint();
2987 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TLI, AddressAccessTy,
2988 MemoryInst, Result, InsertedTruncs,
2989 PromotedInsts, TPT);
2990 Matcher.IgnoreProfitability = true;
2991 bool Success = Matcher.MatchAddr(Address, 0);
2992 (void)Success; assert(Success && "Couldn't select *anything*?");
2994 // The match was to check the profitability, the changes made are not
2995 // part of the original matcher. Therefore, they should be dropped
2996 // otherwise the original matcher will not present the right state.
2997 TPT.rollback(LastKnownGood);
2999 // If the match didn't cover I, then it won't be shared by it.
3000 if (std::find(MatchedAddrModeInsts.begin(), MatchedAddrModeInsts.end(),
3001 I) == MatchedAddrModeInsts.end())
3004 MatchedAddrModeInsts.clear();
3010 } // end anonymous namespace
3012 /// IsNonLocalValue - Return true if the specified values are defined in a
3013 /// different basic block than BB.
3014 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
3015 if (Instruction *I = dyn_cast<Instruction>(V))
3016 return I->getParent() != BB;
3020 /// OptimizeMemoryInst - Load and Store Instructions often have
3021 /// addressing modes that can do significant amounts of computation. As such,
3022 /// instruction selection will try to get the load or store to do as much
3023 /// computation as possible for the program. The problem is that isel can only
3024 /// see within a single block. As such, we sink as much legal addressing mode
3025 /// stuff into the block as possible.
3027 /// This method is used to optimize both load/store and inline asms with memory
3029 bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
3033 // Try to collapse single-value PHI nodes. This is necessary to undo
3034 // unprofitable PRE transformations.
3035 SmallVector<Value*, 8> worklist;
3036 SmallPtrSet<Value*, 16> Visited;
3037 worklist.push_back(Addr);
3039 // Use a worklist to iteratively look through PHI nodes, and ensure that
3040 // the addressing mode obtained from the non-PHI roots of the graph
3042 Value *Consensus = nullptr;
3043 unsigned NumUsesConsensus = 0;
3044 bool IsNumUsesConsensusValid = false;
3045 SmallVector<Instruction*, 16> AddrModeInsts;
3046 ExtAddrMode AddrMode;
3047 TypePromotionTransaction TPT;
3048 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3049 TPT.getRestorationPoint();
3050 while (!worklist.empty()) {
3051 Value *V = worklist.back();
3052 worklist.pop_back();
3054 // Break use-def graph loops.
3055 if (!Visited.insert(V).second) {
3056 Consensus = nullptr;
3060 // For a PHI node, push all of its incoming values.
3061 if (PHINode *P = dyn_cast<PHINode>(V)) {
3062 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i)
3063 worklist.push_back(P->getIncomingValue(i));
3067 // For non-PHIs, determine the addressing mode being computed.
3068 SmallVector<Instruction*, 16> NewAddrModeInsts;
3069 ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
3070 V, AccessTy, MemoryInst, NewAddrModeInsts, *TLI, InsertedTruncsSet,
3071 PromotedInsts, TPT);
3073 // This check is broken into two cases with very similar code to avoid using
3074 // getNumUses() as much as possible. Some values have a lot of uses, so
3075 // calling getNumUses() unconditionally caused a significant compile-time
3079 AddrMode = NewAddrMode;
3080 AddrModeInsts = NewAddrModeInsts;
3082 } else if (NewAddrMode == AddrMode) {
3083 if (!IsNumUsesConsensusValid) {
3084 NumUsesConsensus = Consensus->getNumUses();
3085 IsNumUsesConsensusValid = true;
3088 // Ensure that the obtained addressing mode is equivalent to that obtained
3089 // for all other roots of the PHI traversal. Also, when choosing one
3090 // such root as representative, select the one with the most uses in order
3091 // to keep the cost modeling heuristics in AddressingModeMatcher
3093 unsigned NumUses = V->getNumUses();
3094 if (NumUses > NumUsesConsensus) {
3096 NumUsesConsensus = NumUses;
3097 AddrModeInsts = NewAddrModeInsts;
3102 Consensus = nullptr;
3106 // If the addressing mode couldn't be determined, or if multiple different
3107 // ones were determined, bail out now.
3109 TPT.rollback(LastKnownGood);
3114 // Check to see if any of the instructions supersumed by this addr mode are
3115 // non-local to I's BB.
3116 bool AnyNonLocal = false;
3117 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
3118 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
3124 // If all the instructions matched are already in this BB, don't do anything.
3126 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
3130 // Insert this computation right after this user. Since our caller is
3131 // scanning from the top of the BB to the bottom, reuse of the expr are
3132 // guaranteed to happen later.
3133 IRBuilder<> Builder(MemoryInst);
3135 // Now that we determined the addressing expression we want to use and know
3136 // that we have to sink it into this block. Check to see if we have already
3137 // done this for some other load/store instr in this block. If so, reuse the
3139 Value *&SunkAddr = SunkAddrs[Addr];
3141 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
3142 << *MemoryInst << "\n");
3143 if (SunkAddr->getType() != Addr->getType())
3144 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
3145 } else if (AddrSinkUsingGEPs ||
3146 (!AddrSinkUsingGEPs.getNumOccurrences() && TM &&
3147 TM->getSubtargetImpl(*MemoryInst->getParent()->getParent())
3149 // By default, we use the GEP-based method when AA is used later. This
3150 // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
3151 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
3152 << *MemoryInst << "\n");
3153 Type *IntPtrTy = TLI->getDataLayout()->getIntPtrType(Addr->getType());
3154 Value *ResultPtr = nullptr, *ResultIndex = nullptr;
3156 // First, find the pointer.
3157 if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
3158 ResultPtr = AddrMode.BaseReg;
3159 AddrMode.BaseReg = nullptr;
3162 if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
3163 // We can't add more than one pointer together, nor can we scale a
3164 // pointer (both of which seem meaningless).
3165 if (ResultPtr || AddrMode.Scale != 1)
3168 ResultPtr = AddrMode.ScaledReg;
3172 if (AddrMode.BaseGV) {
3176 ResultPtr = AddrMode.BaseGV;
3179 // If the real base value actually came from an inttoptr, then the matcher
3180 // will look through it and provide only the integer value. In that case,
3182 if (!ResultPtr && AddrMode.BaseReg) {
3184 Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(), "sunkaddr");
3185 AddrMode.BaseReg = nullptr;
3186 } else if (!ResultPtr && AddrMode.Scale == 1) {
3188 Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(), "sunkaddr");
3193 !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
3194 SunkAddr = Constant::getNullValue(Addr->getType());
3195 } else if (!ResultPtr) {
3199 Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
3201 // Start with the base register. Do this first so that subsequent address
3202 // matching finds it last, which will prevent it from trying to match it
3203 // as the scaled value in case it happens to be a mul. That would be
3204 // problematic if we've sunk a different mul for the scale, because then
3205 // we'd end up sinking both muls.
3206 if (AddrMode.BaseReg) {
3207 Value *V = AddrMode.BaseReg;
3208 if (V->getType() != IntPtrTy)
3209 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
3214 // Add the scale value.
3215 if (AddrMode.Scale) {
3216 Value *V = AddrMode.ScaledReg;
3217 if (V->getType() == IntPtrTy) {
3219 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
3220 cast<IntegerType>(V->getType())->getBitWidth()) {
3221 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
3223 // It is only safe to sign extend the BaseReg if we know that the math
3224 // required to create it did not overflow before we extend it. Since
3225 // the original IR value was tossed in favor of a constant back when
3226 // the AddrMode was created we need to bail out gracefully if widths
3227 // do not match instead of extending it.
3228 Instruction *I = dyn_cast_or_null<Instruction>(ResultIndex);
3229 if (I && (ResultIndex != AddrMode.BaseReg))
3230 I->eraseFromParent();
3234 if (AddrMode.Scale != 1)
3235 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
3238 ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
3243 // Add in the Base Offset if present.
3244 if (AddrMode.BaseOffs) {
3245 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
3247 // We need to add this separately from the scale above to help with
3248 // SDAG consecutive load/store merging.
3249 if (ResultPtr->getType() != I8PtrTy)
3250 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
3251 ResultPtr = Builder.CreateGEP(ResultPtr, ResultIndex, "sunkaddr");
3258 SunkAddr = ResultPtr;
3260 if (ResultPtr->getType() != I8PtrTy)
3261 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
3262 SunkAddr = Builder.CreateGEP(ResultPtr, ResultIndex, "sunkaddr");
3265 if (SunkAddr->getType() != Addr->getType())
3266 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
3269 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
3270 << *MemoryInst << "\n");
3271 Type *IntPtrTy = TLI->getDataLayout()->getIntPtrType(Addr->getType());
3272 Value *Result = nullptr;
3274 // Start with the base register. Do this first so that subsequent address
3275 // matching finds it last, which will prevent it from trying to match it
3276 // as the scaled value in case it happens to be a mul. That would be
3277 // problematic if we've sunk a different mul for the scale, because then
3278 // we'd end up sinking both muls.
3279 if (AddrMode.BaseReg) {
3280 Value *V = AddrMode.BaseReg;
3281 if (V->getType()->isPointerTy())
3282 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
3283 if (V->getType() != IntPtrTy)
3284 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
3288 // Add the scale value.
3289 if (AddrMode.Scale) {
3290 Value *V = AddrMode.ScaledReg;
3291 if (V->getType() == IntPtrTy) {
3293 } else if (V->getType()->isPointerTy()) {
3294 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
3295 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
3296 cast<IntegerType>(V->getType())->getBitWidth()) {
3297 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
3299 // It is only safe to sign extend the BaseReg if we know that the math
3300 // required to create it did not overflow before we extend it. Since
3301 // the original IR value was tossed in favor of a constant back when
3302 // the AddrMode was created we need to bail out gracefully if widths
3303 // do not match instead of extending it.
3304 Instruction *I = dyn_cast_or_null<Instruction>(Result);
3305 if (I && (Result != AddrMode.BaseReg))
3306 I->eraseFromParent();
3309 if (AddrMode.Scale != 1)
3310 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
3313 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3318 // Add in the BaseGV if present.
3319 if (AddrMode.BaseGV) {
3320 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
3322 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3327 // Add in the Base Offset if present.
3328 if (AddrMode.BaseOffs) {
3329 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
3331 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3337 SunkAddr = Constant::getNullValue(Addr->getType());
3339 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
3342 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
3344 // If we have no uses, recursively delete the value and all dead instructions
3346 if (Repl->use_empty()) {
3347 // This can cause recursive deletion, which can invalidate our iterator.
3348 // Use a WeakVH to hold onto it in case this happens.
3349 WeakVH IterHandle(CurInstIterator);
3350 BasicBlock *BB = CurInstIterator->getParent();
3352 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
3354 if (IterHandle != CurInstIterator) {
3355 // If the iterator instruction was recursively deleted, start over at the
3356 // start of the block.
3357 CurInstIterator = BB->begin();
3365 /// OptimizeInlineAsmInst - If there are any memory operands, use
3366 /// OptimizeMemoryInst to sink their address computing into the block when
3367 /// possible / profitable.
3368 bool CodeGenPrepare::OptimizeInlineAsmInst(CallInst *CS) {
3369 bool MadeChange = false;
3371 TargetLowering::AsmOperandInfoVector
3372 TargetConstraints = TLI->ParseConstraints(CS);
3374 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
3375 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
3377 // Compute the constraint code and ConstraintType to use.
3378 TLI->ComputeConstraintToUse(OpInfo, SDValue());
3380 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
3381 OpInfo.isIndirect) {
3382 Value *OpVal = CS->getArgOperand(ArgNo++);
3383 MadeChange |= OptimizeMemoryInst(CS, OpVal, OpVal->getType());
3384 } else if (OpInfo.Type == InlineAsm::isInput)
3391 /// \brief Check if all the uses of \p Inst are equivalent (or free) zero or
3392 /// sign extensions.
3393 static bool hasSameExtUse(Instruction *Inst, const TargetLowering &TLI) {
3394 assert(!Inst->use_empty() && "Input must have at least one use");
3395 const Instruction *FirstUser = cast<Instruction>(*Inst->user_begin());
3396 bool IsSExt = isa<SExtInst>(FirstUser);
3397 Type *ExtTy = FirstUser->getType();
3398 for (const User *U : Inst->users()) {
3399 const Instruction *UI = cast<Instruction>(U);
3400 if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI)))
3402 Type *CurTy = UI->getType();
3403 // Same input and output types: Same instruction after CSE.
3407 // If IsSExt is true, we are in this situation:
3409 // b = sext ty1 a to ty2
3410 // c = sext ty1 a to ty3
3411 // Assuming ty2 is shorter than ty3, this could be turned into:
3413 // b = sext ty1 a to ty2
3414 // c = sext ty2 b to ty3
3415 // However, the last sext is not free.
3419 // This is a ZExt, maybe this is free to extend from one type to another.
3420 // In that case, we would not account for a different use.
3423 if (ExtTy->getScalarType()->getIntegerBitWidth() >
3424 CurTy->getScalarType()->getIntegerBitWidth()) {
3432 if (!TLI.isZExtFree(NarrowTy, LargeTy))
3435 // All uses are the same or can be derived from one another for free.
3439 /// \brief Try to form ExtLd by promoting \p Exts until they reach a
3440 /// load instruction.
3441 /// If an ext(load) can be formed, it is returned via \p LI for the load
3442 /// and \p Inst for the extension.
3443 /// Otherwise LI == nullptr and Inst == nullptr.
3444 /// When some promotion happened, \p TPT contains the proper state to
3447 /// \return true when promoting was necessary to expose the ext(load)
3448 /// opportunity, false otherwise.
3452 /// %ld = load i32* %addr
3453 /// %add = add nuw i32 %ld, 4
3454 /// %zext = zext i32 %add to i64
3458 /// %ld = load i32* %addr
3459 /// %zext = zext i32 %ld to i64
3460 /// %add = add nuw i64 %zext, 4
3462 /// Thanks to the promotion, we can match zext(load i32*) to i64.
3463 bool CodeGenPrepare::ExtLdPromotion(TypePromotionTransaction &TPT,
3464 LoadInst *&LI, Instruction *&Inst,
3465 const SmallVectorImpl<Instruction *> &Exts,
3466 unsigned CreatedInsts = 0) {
3467 // Iterate over all the extensions to see if one form an ext(load).
3468 for (auto I : Exts) {
3469 // Check if we directly have ext(load).
3470 if ((LI = dyn_cast<LoadInst>(I->getOperand(0)))) {
3472 // No promotion happened here.
3475 // Check whether or not we want to do any promotion.
3476 if (!TLI || !TLI->enableExtLdPromotion() || DisableExtLdPromotion)
3478 // Get the action to perform the promotion.
3479 TypePromotionHelper::Action TPH = TypePromotionHelper::getAction(
3480 I, InsertedTruncsSet, *TLI, PromotedInsts);
3481 // Check if we can promote.
3484 // Save the current state.
3485 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3486 TPT.getRestorationPoint();
3487 SmallVector<Instruction *, 4> NewExts;
3488 unsigned NewCreatedInsts = 0;
3490 Value *PromotedVal =
3491 TPH(I, TPT, PromotedInsts, NewCreatedInsts, &NewExts, nullptr);
3492 assert(PromotedVal &&
3493 "TypePromotionHelper should have filtered out those cases");
3495 // We would be able to merge only one extension in a load.
3496 // Therefore, if we have more than 1 new extension we heuristically
3497 // cut this search path, because it means we degrade the code quality.
3498 // With exactly 2, the transformation is neutral, because we will merge
3499 // one extension but leave one. However, we optimistically keep going,
3500 // because the new extension may be removed too.
3501 unsigned TotalCreatedInsts = CreatedInsts + NewCreatedInsts;
3502 if (!StressExtLdPromotion &&
3503 (TotalCreatedInsts > 1 ||
3504 !isPromotedInstructionLegal(*TLI, PromotedVal))) {
3505 // The promotion is not profitable, rollback to the previous state.
3506 TPT.rollback(LastKnownGood);
3509 // The promotion is profitable.
3510 // Check if it exposes an ext(load).
3511 (void)ExtLdPromotion(TPT, LI, Inst, NewExts, TotalCreatedInsts);
3512 if (LI && (StressExtLdPromotion || NewCreatedInsts == 0 ||
3513 // If we have created a new extension, i.e., now we have two
3514 // extensions. We must make sure one of them is merged with
3515 // the load, otherwise we may degrade the code quality.
3516 (LI->hasOneUse() || hasSameExtUse(LI, *TLI))))
3517 // Promotion happened.
3519 // If this does not help to expose an ext(load) then, rollback.
3520 TPT.rollback(LastKnownGood);
3522 // None of the extension can form an ext(load).
3528 /// MoveExtToFormExtLoad - Move a zext or sext fed by a load into the same
3529 /// basic block as the load, unless conditions are unfavorable. This allows
3530 /// SelectionDAG to fold the extend into the load.
3531 /// \p I[in/out] the extension may be modified during the process if some
3532 /// promotions apply.
3534 bool CodeGenPrepare::MoveExtToFormExtLoad(Instruction *&I) {
3535 // Try to promote a chain of computation if it allows to form
3536 // an extended load.
3537 TypePromotionTransaction TPT;
3538 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3539 TPT.getRestorationPoint();
3540 SmallVector<Instruction *, 1> Exts;
3542 // Look for a load being extended.
3543 LoadInst *LI = nullptr;
3544 Instruction *OldExt = I;
3545 bool HasPromoted = ExtLdPromotion(TPT, LI, I, Exts);
3547 assert(!HasPromoted && !LI && "If we did not match any load instruction "
3548 "the code must remain the same");
3553 // If they're already in the same block, there's nothing to do.
3554 // Make the cheap checks first if we did not promote.
3555 // If we promoted, we need to check if it is indeed profitable.
3556 if (!HasPromoted && LI->getParent() == I->getParent())
3559 EVT VT = TLI->getValueType(I->getType());
3560 EVT LoadVT = TLI->getValueType(LI->getType());
3562 // If the load has other users and the truncate is not free, this probably
3563 // isn't worthwhile.
3564 if (!LI->hasOneUse() && TLI &&
3565 (TLI->isTypeLegal(LoadVT) || !TLI->isTypeLegal(VT)) &&
3566 !TLI->isTruncateFree(I->getType(), LI->getType())) {
3568 TPT.rollback(LastKnownGood);
3572 // Check whether the target supports casts folded into loads.
3574 if (isa<ZExtInst>(I))
3575 LType = ISD::ZEXTLOAD;
3577 assert(isa<SExtInst>(I) && "Unexpected ext type!");
3578 LType = ISD::SEXTLOAD;
3580 if (TLI && !TLI->isLoadExtLegal(LType, VT, LoadVT)) {
3582 TPT.rollback(LastKnownGood);
3586 // Move the extend into the same block as the load, so that SelectionDAG
3589 I->removeFromParent();
3595 bool CodeGenPrepare::OptimizeExtUses(Instruction *I) {
3596 BasicBlock *DefBB = I->getParent();
3598 // If the result of a {s|z}ext and its source are both live out, rewrite all
3599 // other uses of the source with result of extension.
3600 Value *Src = I->getOperand(0);
3601 if (Src->hasOneUse())
3604 // Only do this xform if truncating is free.
3605 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
3608 // Only safe to perform the optimization if the source is also defined in
3610 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
3613 bool DefIsLiveOut = false;
3614 for (User *U : I->users()) {
3615 Instruction *UI = cast<Instruction>(U);
3617 // Figure out which BB this ext is used in.
3618 BasicBlock *UserBB = UI->getParent();
3619 if (UserBB == DefBB) continue;
3620 DefIsLiveOut = true;
3626 // Make sure none of the uses are PHI nodes.
3627 for (User *U : Src->users()) {
3628 Instruction *UI = cast<Instruction>(U);
3629 BasicBlock *UserBB = UI->getParent();
3630 if (UserBB == DefBB) continue;
3631 // Be conservative. We don't want this xform to end up introducing
3632 // reloads just before load / store instructions.
3633 if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
3637 // InsertedTruncs - Only insert one trunc in each block once.
3638 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
3640 bool MadeChange = false;
3641 for (Use &U : Src->uses()) {
3642 Instruction *User = cast<Instruction>(U.getUser());
3644 // Figure out which BB this ext is used in.
3645 BasicBlock *UserBB = User->getParent();
3646 if (UserBB == DefBB) continue;
3648 // Both src and def are live in this block. Rewrite the use.
3649 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
3651 if (!InsertedTrunc) {
3652 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
3653 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt);
3654 InsertedTruncsSet.insert(InsertedTrunc);
3657 // Replace a use of the {s|z}ext source with a use of the result.
3666 /// isFormingBranchFromSelectProfitable - Returns true if a SelectInst should be
3667 /// turned into an explicit branch.
3668 static bool isFormingBranchFromSelectProfitable(SelectInst *SI) {
3669 // FIXME: This should use the same heuristics as IfConversion to determine
3670 // whether a select is better represented as a branch. This requires that
3671 // branch probability metadata is preserved for the select, which is not the
3674 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
3676 // If the branch is predicted right, an out of order CPU can avoid blocking on
3677 // the compare. Emit cmovs on compares with a memory operand as branches to
3678 // avoid stalls on the load from memory. If the compare has more than one use
3679 // there's probably another cmov or setcc around so it's not worth emitting a
3684 Value *CmpOp0 = Cmp->getOperand(0);
3685 Value *CmpOp1 = Cmp->getOperand(1);
3687 // We check that the memory operand has one use to avoid uses of the loaded
3688 // value directly after the compare, making branches unprofitable.
3689 return Cmp->hasOneUse() &&
3690 ((isa<LoadInst>(CmpOp0) && CmpOp0->hasOneUse()) ||
3691 (isa<LoadInst>(CmpOp1) && CmpOp1->hasOneUse()));
3695 /// If we have a SelectInst that will likely profit from branch prediction,
3696 /// turn it into a branch.
3697 bool CodeGenPrepare::OptimizeSelectInst(SelectInst *SI) {
3698 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
3700 // Can we convert the 'select' to CF ?
3701 if (DisableSelectToBranch || OptSize || !TLI || VectorCond)
3704 TargetLowering::SelectSupportKind SelectKind;
3706 SelectKind = TargetLowering::VectorMaskSelect;
3707 else if (SI->getType()->isVectorTy())
3708 SelectKind = TargetLowering::ScalarCondVectorVal;
3710 SelectKind = TargetLowering::ScalarValSelect;
3712 // Do we have efficient codegen support for this kind of 'selects' ?
3713 if (TLI->isSelectSupported(SelectKind)) {
3714 // We have efficient codegen support for the select instruction.
3715 // Check if it is profitable to keep this 'select'.
3716 if (!TLI->isPredictableSelectExpensive() ||
3717 !isFormingBranchFromSelectProfitable(SI))
3723 // First, we split the block containing the select into 2 blocks.
3724 BasicBlock *StartBlock = SI->getParent();
3725 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(SI));
3726 BasicBlock *NextBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
3728 // Create a new block serving as the landing pad for the branch.
3729 BasicBlock *SmallBlock = BasicBlock::Create(SI->getContext(), "select.mid",
3730 NextBlock->getParent(), NextBlock);
3732 // Move the unconditional branch from the block with the select in it into our
3733 // landing pad block.
3734 StartBlock->getTerminator()->eraseFromParent();
3735 BranchInst::Create(NextBlock, SmallBlock);
3737 // Insert the real conditional branch based on the original condition.
3738 BranchInst::Create(NextBlock, SmallBlock, SI->getCondition(), SI);
3740 // The select itself is replaced with a PHI Node.
3741 PHINode *PN = PHINode::Create(SI->getType(), 2, "", NextBlock->begin());
3743 PN->addIncoming(SI->getTrueValue(), StartBlock);
3744 PN->addIncoming(SI->getFalseValue(), SmallBlock);
3745 SI->replaceAllUsesWith(PN);
3746 SI->eraseFromParent();
3748 // Instruct OptimizeBlock to skip to the next block.
3749 CurInstIterator = StartBlock->end();
3750 ++NumSelectsExpanded;
3754 static bool isBroadcastShuffle(ShuffleVectorInst *SVI) {
3755 SmallVector<int, 16> Mask(SVI->getShuffleMask());
3757 for (unsigned i = 0; i < Mask.size(); ++i) {
3758 if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
3760 SplatElem = Mask[i];
3766 /// Some targets have expensive vector shifts if the lanes aren't all the same
3767 /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
3768 /// it's often worth sinking a shufflevector splat down to its use so that
3769 /// codegen can spot all lanes are identical.
3770 bool CodeGenPrepare::OptimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
3771 BasicBlock *DefBB = SVI->getParent();
3773 // Only do this xform if variable vector shifts are particularly expensive.
3774 if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
3777 // We only expect better codegen by sinking a shuffle if we can recognise a
3779 if (!isBroadcastShuffle(SVI))
3782 // InsertedShuffles - Only insert a shuffle in each block once.
3783 DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
3785 bool MadeChange = false;
3786 for (User *U : SVI->users()) {
3787 Instruction *UI = cast<Instruction>(U);
3789 // Figure out which BB this ext is used in.
3790 BasicBlock *UserBB = UI->getParent();
3791 if (UserBB == DefBB) continue;
3793 // For now only apply this when the splat is used by a shift instruction.
3794 if (!UI->isShift()) continue;
3796 // Everything checks out, sink the shuffle if the user's block doesn't
3797 // already have a copy.
3798 Instruction *&InsertedShuffle = InsertedShuffles[UserBB];
3800 if (!InsertedShuffle) {
3801 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
3802 InsertedShuffle = new ShuffleVectorInst(SVI->getOperand(0),
3804 SVI->getOperand(2), "", InsertPt);
3807 UI->replaceUsesOfWith(SVI, InsertedShuffle);
3811 // If we removed all uses, nuke the shuffle.
3812 if (SVI->use_empty()) {
3813 SVI->eraseFromParent();
3821 /// \brief Helper class to promote a scalar operation to a vector one.
3822 /// This class is used to move downward extractelement transition.
3824 /// a = vector_op <2 x i32>
3825 /// b = extractelement <2 x i32> a, i32 0
3830 /// a = vector_op <2 x i32>
3831 /// c = vector_op a (equivalent to scalar_op on the related lane)
3832 /// * d = extractelement <2 x i32> c, i32 0
3834 /// Assuming both extractelement and store can be combine, we get rid of the
3836 class VectorPromoteHelper {
3837 /// Used to perform some checks on the legality of vector operations.
3838 const TargetLowering &TLI;
3840 /// Used to estimated the cost of the promoted chain.
3841 const TargetTransformInfo &TTI;
3843 /// The transition being moved downwards.
3844 Instruction *Transition;
3845 /// The sequence of instructions to be promoted.
3846 SmallVector<Instruction *, 4> InstsToBePromoted;
3847 /// Cost of combining a store and an extract.
3848 unsigned StoreExtractCombineCost;
3849 /// Instruction that will be combined with the transition.
3850 Instruction *CombineInst;
3852 /// \brief The instruction that represents the current end of the transition.
3853 /// Since we are faking the promotion until we reach the end of the chain
3854 /// of computation, we need a way to get the current end of the transition.
3855 Instruction *getEndOfTransition() const {
3856 if (InstsToBePromoted.empty())
3858 return InstsToBePromoted.back();
3861 /// \brief Return the index of the original value in the transition.
3862 /// E.g., for "extractelement <2 x i32> c, i32 1" the original value,
3863 /// c, is at index 0.
3864 unsigned getTransitionOriginalValueIdx() const {
3865 assert(isa<ExtractElementInst>(Transition) &&
3866 "Other kind of transitions are not supported yet");
3870 /// \brief Return the index of the index in the transition.
3871 /// E.g., for "extractelement <2 x i32> c, i32 0" the index
3873 unsigned getTransitionIdx() const {
3874 assert(isa<ExtractElementInst>(Transition) &&
3875 "Other kind of transitions are not supported yet");
3879 /// \brief Get the type of the transition.
3880 /// This is the type of the original value.
3881 /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the
3882 /// transition is <2 x i32>.
3883 Type *getTransitionType() const {
3884 return Transition->getOperand(getTransitionOriginalValueIdx())->getType();
3887 /// \brief Promote \p ToBePromoted by moving \p Def downward through.
3888 /// I.e., we have the following sequence:
3889 /// Def = Transition <ty1> a to <ty2>
3890 /// b = ToBePromoted <ty2> Def, ...
3892 /// b = ToBePromoted <ty1> a, ...
3893 /// Def = Transition <ty1> ToBePromoted to <ty2>
3894 void promoteImpl(Instruction *ToBePromoted);
3896 /// \brief Check whether or not it is profitable to promote all the
3897 /// instructions enqueued to be promoted.
3898 bool isProfitableToPromote() {
3899 Value *ValIdx = Transition->getOperand(getTransitionOriginalValueIdx());
3900 unsigned Index = isa<ConstantInt>(ValIdx)
3901 ? cast<ConstantInt>(ValIdx)->getZExtValue()
3903 Type *PromotedType = getTransitionType();
3905 StoreInst *ST = cast<StoreInst>(CombineInst);
3906 unsigned AS = ST->getPointerAddressSpace();
3907 unsigned Align = ST->getAlignment();
3908 // Check if this store is supported.
3909 if (!TLI.allowsMisalignedMemoryAccesses(
3910 TLI.getValueType(ST->getValueOperand()->getType()), AS, Align)) {
3911 // If this is not supported, there is no way we can combine
3912 // the extract with the store.
3916 // The scalar chain of computation has to pay for the transition
3917 // scalar to vector.
3918 // The vector chain has to account for the combining cost.
3919 uint64_t ScalarCost =
3920 TTI.getVectorInstrCost(Transition->getOpcode(), PromotedType, Index);
3921 uint64_t VectorCost = StoreExtractCombineCost;
3922 for (const auto &Inst : InstsToBePromoted) {
3923 // Compute the cost.
3924 // By construction, all instructions being promoted are arithmetic ones.
3925 // Moreover, one argument is a constant that can be viewed as a splat
3927 Value *Arg0 = Inst->getOperand(0);
3928 bool IsArg0Constant = isa<UndefValue>(Arg0) || isa<ConstantInt>(Arg0) ||
3929 isa<ConstantFP>(Arg0);
3930 TargetTransformInfo::OperandValueKind Arg0OVK =
3931 IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
3932 : TargetTransformInfo::OK_AnyValue;
3933 TargetTransformInfo::OperandValueKind Arg1OVK =
3934 !IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
3935 : TargetTransformInfo::OK_AnyValue;
3936 ScalarCost += TTI.getArithmeticInstrCost(
3937 Inst->getOpcode(), Inst->getType(), Arg0OVK, Arg1OVK);
3938 VectorCost += TTI.getArithmeticInstrCost(Inst->getOpcode(), PromotedType,
3941 DEBUG(dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
3942 << ScalarCost << "\nVector: " << VectorCost << '\n');
3943 return ScalarCost > VectorCost;
3946 /// \brief Generate a constant vector with \p Val with the same
3947 /// number of elements as the transition.
3948 /// \p UseSplat defines whether or not \p Val should be replicated
3949 /// accross the whole vector.
3950 /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>,
3951 /// otherwise we generate a vector with as many undef as possible:
3952 /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only
3953 /// used at the index of the extract.
3954 Value *getConstantVector(Constant *Val, bool UseSplat) const {
3955 unsigned ExtractIdx = UINT_MAX;
3957 // If we cannot determine where the constant must be, we have to
3958 // use a splat constant.
3959 Value *ValExtractIdx = Transition->getOperand(getTransitionIdx());
3960 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(ValExtractIdx))
3961 ExtractIdx = CstVal->getSExtValue();
3966 unsigned End = getTransitionType()->getVectorNumElements();
3968 return ConstantVector::getSplat(End, Val);
3970 SmallVector<Constant *, 4> ConstVec;
3971 UndefValue *UndefVal = UndefValue::get(Val->getType());
3972 for (unsigned Idx = 0; Idx != End; ++Idx) {
3973 if (Idx == ExtractIdx)
3974 ConstVec.push_back(Val);
3976 ConstVec.push_back(UndefVal);
3978 return ConstantVector::get(ConstVec);
3981 /// \brief Check if promoting to a vector type an operand at \p OperandIdx
3982 /// in \p Use can trigger undefined behavior.
3983 static bool canCauseUndefinedBehavior(const Instruction *Use,
3984 unsigned OperandIdx) {
3985 // This is not safe to introduce undef when the operand is on
3986 // the right hand side of a division-like instruction.
3987 if (OperandIdx != 1)
3989 switch (Use->getOpcode()) {
3992 case Instruction::SDiv:
3993 case Instruction::UDiv:
3994 case Instruction::SRem:
3995 case Instruction::URem:
3997 case Instruction::FDiv:
3998 case Instruction::FRem:
3999 return !Use->hasNoNaNs();
4001 llvm_unreachable(nullptr);
4005 VectorPromoteHelper(const TargetLowering &TLI, const TargetTransformInfo &TTI,
4006 Instruction *Transition, unsigned CombineCost)
4007 : TLI(TLI), TTI(TTI), Transition(Transition),
4008 StoreExtractCombineCost(CombineCost), CombineInst(nullptr) {
4009 assert(Transition && "Do not know how to promote null");
4012 /// \brief Check if we can promote \p ToBePromoted to \p Type.
4013 bool canPromote(const Instruction *ToBePromoted) const {
4014 // We could support CastInst too.
4015 return isa<BinaryOperator>(ToBePromoted);
4018 /// \brief Check if it is profitable to promote \p ToBePromoted
4019 /// by moving downward the transition through.
4020 bool shouldPromote(const Instruction *ToBePromoted) const {
4021 // Promote only if all the operands can be statically expanded.
4022 // Indeed, we do not want to introduce any new kind of transitions.
4023 for (const Use &U : ToBePromoted->operands()) {
4024 const Value *Val = U.get();
4025 if (Val == getEndOfTransition()) {
4026 // If the use is a division and the transition is on the rhs,
4027 // we cannot promote the operation, otherwise we may create a
4028 // division by zero.
4029 if (canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()))
4033 if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) &&
4034 !isa<ConstantFP>(Val))
4037 // Check that the resulting operation is legal.
4038 int ISDOpcode = TLI.InstructionOpcodeToISD(ToBePromoted->getOpcode());
4041 return StressStoreExtract ||
4042 TLI.isOperationLegalOrCustom(
4043 ISDOpcode, TLI.getValueType(getTransitionType(), true));
4046 /// \brief Check whether or not \p Use can be combined
4047 /// with the transition.
4048 /// I.e., is it possible to do Use(Transition) => AnotherUse?
4049 bool canCombine(const Instruction *Use) { return isa<StoreInst>(Use); }
4051 /// \brief Record \p ToBePromoted as part of the chain to be promoted.
4052 void enqueueForPromotion(Instruction *ToBePromoted) {
4053 InstsToBePromoted.push_back(ToBePromoted);
4056 /// \brief Set the instruction that will be combined with the transition.
4057 void recordCombineInstruction(Instruction *ToBeCombined) {
4058 assert(canCombine(ToBeCombined) && "Unsupported instruction to combine");
4059 CombineInst = ToBeCombined;
4062 /// \brief Promote all the instructions enqueued for promotion if it is
4064 /// \return True if the promotion happened, false otherwise.
4066 // Check if there is something to promote.
4067 // Right now, if we do not have anything to combine with,
4068 // we assume the promotion is not profitable.
4069 if (InstsToBePromoted.empty() || !CombineInst)
4073 if (!StressStoreExtract && !isProfitableToPromote())
4077 for (auto &ToBePromoted : InstsToBePromoted)
4078 promoteImpl(ToBePromoted);
4079 InstsToBePromoted.clear();
4083 } // End of anonymous namespace.
4085 void VectorPromoteHelper::promoteImpl(Instruction *ToBePromoted) {
4086 // At this point, we know that all the operands of ToBePromoted but Def
4087 // can be statically promoted.
4088 // For Def, we need to use its parameter in ToBePromoted:
4089 // b = ToBePromoted ty1 a
4090 // Def = Transition ty1 b to ty2
4091 // Move the transition down.
4092 // 1. Replace all uses of the promoted operation by the transition.
4093 // = ... b => = ... Def.
4094 assert(ToBePromoted->getType() == Transition->getType() &&
4095 "The type of the result of the transition does not match "
4097 ToBePromoted->replaceAllUsesWith(Transition);
4098 // 2. Update the type of the uses.
4099 // b = ToBePromoted ty2 Def => b = ToBePromoted ty1 Def.
4100 Type *TransitionTy = getTransitionType();
4101 ToBePromoted->mutateType(TransitionTy);
4102 // 3. Update all the operands of the promoted operation with promoted
4104 // b = ToBePromoted ty1 Def => b = ToBePromoted ty1 a.
4105 for (Use &U : ToBePromoted->operands()) {
4106 Value *Val = U.get();
4107 Value *NewVal = nullptr;
4108 if (Val == Transition)
4109 NewVal = Transition->getOperand(getTransitionOriginalValueIdx());
4110 else if (isa<UndefValue>(Val) || isa<ConstantInt>(Val) ||
4111 isa<ConstantFP>(Val)) {
4112 // Use a splat constant if it is not safe to use undef.
4113 NewVal = getConstantVector(
4114 cast<Constant>(Val),
4115 isa<UndefValue>(Val) ||
4116 canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()));
4118 llvm_unreachable("Did you modified shouldPromote and forgot to update "
4120 ToBePromoted->setOperand(U.getOperandNo(), NewVal);
4122 Transition->removeFromParent();
4123 Transition->insertAfter(ToBePromoted);
4124 Transition->setOperand(getTransitionOriginalValueIdx(), ToBePromoted);
4127 /// Some targets can do store(extractelement) with one instruction.
4128 /// Try to push the extractelement towards the stores when the target
4129 /// has this feature and this is profitable.
4130 bool CodeGenPrepare::OptimizeExtractElementInst(Instruction *Inst) {
4131 unsigned CombineCost = UINT_MAX;
4132 if (DisableStoreExtract || !TLI ||
4133 (!StressStoreExtract &&
4134 !TLI->canCombineStoreAndExtract(Inst->getOperand(0)->getType(),
4135 Inst->getOperand(1), CombineCost)))
4138 // At this point we know that Inst is a vector to scalar transition.
4139 // Try to move it down the def-use chain, until:
4140 // - We can combine the transition with its single use
4141 // => we got rid of the transition.
4142 // - We escape the current basic block
4143 // => we would need to check that we are moving it at a cheaper place and
4144 // we do not do that for now.
4145 BasicBlock *Parent = Inst->getParent();
4146 DEBUG(dbgs() << "Found an interesting transition: " << *Inst << '\n');
4147 VectorPromoteHelper VPH(*TLI, *TTI, Inst, CombineCost);
4148 // If the transition has more than one use, assume this is not going to be
4150 while (Inst->hasOneUse()) {
4151 Instruction *ToBePromoted = cast<Instruction>(*Inst->user_begin());
4152 DEBUG(dbgs() << "Use: " << *ToBePromoted << '\n');
4154 if (ToBePromoted->getParent() != Parent) {
4155 DEBUG(dbgs() << "Instruction to promote is in a different block ("
4156 << ToBePromoted->getParent()->getName()
4157 << ") than the transition (" << Parent->getName() << ").\n");
4161 if (VPH.canCombine(ToBePromoted)) {
4162 DEBUG(dbgs() << "Assume " << *Inst << '\n'
4163 << "will be combined with: " << *ToBePromoted << '\n');
4164 VPH.recordCombineInstruction(ToBePromoted);
4165 bool Changed = VPH.promote();
4166 NumStoreExtractExposed += Changed;
4170 DEBUG(dbgs() << "Try promoting.\n");
4171 if (!VPH.canPromote(ToBePromoted) || !VPH.shouldPromote(ToBePromoted))
4174 DEBUG(dbgs() << "Promoting is possible... Enqueue for promotion!\n");
4176 VPH.enqueueForPromotion(ToBePromoted);
4177 Inst = ToBePromoted;
4182 bool CodeGenPrepare::OptimizeInst(Instruction *I, bool& ModifiedDT) {
4183 if (PHINode *P = dyn_cast<PHINode>(I)) {
4184 // It is possible for very late stage optimizations (such as SimplifyCFG)
4185 // to introduce PHI nodes too late to be cleaned up. If we detect such a
4186 // trivial PHI, go ahead and zap it here.
4187 if (Value *V = SimplifyInstruction(P, TLI ? TLI->getDataLayout() : nullptr,
4189 P->replaceAllUsesWith(V);
4190 P->eraseFromParent();
4197 if (CastInst *CI = dyn_cast<CastInst>(I)) {
4198 // If the source of the cast is a constant, then this should have
4199 // already been constant folded. The only reason NOT to constant fold
4200 // it is if something (e.g. LSR) was careful to place the constant
4201 // evaluation in a block other than then one that uses it (e.g. to hoist
4202 // the address of globals out of a loop). If this is the case, we don't
4203 // want to forward-subst the cast.
4204 if (isa<Constant>(CI->getOperand(0)))
4207 if (TLI && OptimizeNoopCopyExpression(CI, *TLI))
4210 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
4211 /// Sink a zext or sext into its user blocks if the target type doesn't
4212 /// fit in one register
4213 if (TLI && TLI->getTypeAction(CI->getContext(),
4214 TLI->getValueType(CI->getType())) ==
4215 TargetLowering::TypeExpandInteger) {
4216 return SinkCast(CI);
4218 bool MadeChange = MoveExtToFormExtLoad(I);
4219 return MadeChange | OptimizeExtUses(I);
4225 if (CmpInst *CI = dyn_cast<CmpInst>(I))
4226 if (!TLI || !TLI->hasMultipleConditionRegisters())
4227 return OptimizeCmpExpression(CI);
4229 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
4231 return OptimizeMemoryInst(I, I->getOperand(0), LI->getType());
4235 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
4237 return OptimizeMemoryInst(I, SI->getOperand(1),
4238 SI->getOperand(0)->getType());
4242 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(I);
4244 if (BinOp && (BinOp->getOpcode() == Instruction::AShr ||
4245 BinOp->getOpcode() == Instruction::LShr)) {
4246 ConstantInt *CI = dyn_cast<ConstantInt>(BinOp->getOperand(1));
4247 if (TLI && CI && TLI->hasExtractBitsInsn())
4248 return OptimizeExtractBits(BinOp, CI, *TLI);
4253 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
4254 if (GEPI->hasAllZeroIndices()) {
4255 /// The GEP operand must be a pointer, so must its result -> BitCast
4256 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
4257 GEPI->getName(), GEPI);
4258 GEPI->replaceAllUsesWith(NC);
4259 GEPI->eraseFromParent();
4261 OptimizeInst(NC, ModifiedDT);
4267 if (CallInst *CI = dyn_cast<CallInst>(I))
4268 return OptimizeCallInst(CI, ModifiedDT);
4270 if (SelectInst *SI = dyn_cast<SelectInst>(I))
4271 return OptimizeSelectInst(SI);
4273 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I))
4274 return OptimizeShuffleVectorInst(SVI);
4276 if (isa<ExtractElementInst>(I))
4277 return OptimizeExtractElementInst(I);
4282 // In this pass we look for GEP and cast instructions that are used
4283 // across basic blocks and rewrite them to improve basic-block-at-a-time
4285 bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB, bool& ModifiedDT) {
4287 bool MadeChange = false;
4289 CurInstIterator = BB.begin();
4290 while (CurInstIterator != BB.end()) {
4291 MadeChange |= OptimizeInst(CurInstIterator++, ModifiedDT);
4295 MadeChange |= DupRetToEnableTailCallOpts(&BB);
4300 // llvm.dbg.value is far away from the value then iSel may not be able
4301 // handle it properly. iSel will drop llvm.dbg.value if it can not
4302 // find a node corresponding to the value.
4303 bool CodeGenPrepare::PlaceDbgValues(Function &F) {
4304 bool MadeChange = false;
4305 for (BasicBlock &BB : F) {
4306 Instruction *PrevNonDbgInst = nullptr;
4307 for (BasicBlock::iterator BI = BB.begin(), BE = BB.end(); BI != BE;) {
4308 Instruction *Insn = BI++;
4309 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
4310 // Leave dbg.values that refer to an alloca alone. These
4311 // instrinsics describe the address of a variable (= the alloca)
4312 // being taken. They should not be moved next to the alloca
4313 // (and to the beginning of the scope), but rather stay close to
4314 // where said address is used.
4315 if (!DVI || (DVI->getValue() && isa<AllocaInst>(DVI->getValue()))) {
4316 PrevNonDbgInst = Insn;
4320 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
4321 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
4322 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
4323 DVI->removeFromParent();
4324 if (isa<PHINode>(VI))
4325 DVI->insertBefore(VI->getParent()->getFirstInsertionPt());
4327 DVI->insertAfter(VI);
4336 // If there is a sequence that branches based on comparing a single bit
4337 // against zero that can be combined into a single instruction, and the
4338 // target supports folding these into a single instruction, sink the
4339 // mask and compare into the branch uses. Do this before OptimizeBlock ->
4340 // OptimizeInst -> OptimizeCmpExpression, which perturbs the pattern being
4342 bool CodeGenPrepare::sinkAndCmp(Function &F) {
4343 if (!EnableAndCmpSinking)
4345 if (!TLI || !TLI->isMaskAndBranchFoldingLegal())
4347 bool MadeChange = false;
4348 for (Function::iterator I = F.begin(), E = F.end(); I != E; ) {
4349 BasicBlock *BB = I++;
4351 // Does this BB end with the following?
4352 // %andVal = and %val, #single-bit-set
4353 // %icmpVal = icmp %andResult, 0
4354 // br i1 %cmpVal label %dest1, label %dest2"
4355 BranchInst *Brcc = dyn_cast<BranchInst>(BB->getTerminator());
4356 if (!Brcc || !Brcc->isConditional())
4358 ICmpInst *Cmp = dyn_cast<ICmpInst>(Brcc->getOperand(0));
4359 if (!Cmp || Cmp->getParent() != BB)
4361 ConstantInt *Zero = dyn_cast<ConstantInt>(Cmp->getOperand(1));
4362 if (!Zero || !Zero->isZero())
4364 Instruction *And = dyn_cast<Instruction>(Cmp->getOperand(0));
4365 if (!And || And->getOpcode() != Instruction::And || And->getParent() != BB)
4367 ConstantInt* Mask = dyn_cast<ConstantInt>(And->getOperand(1));
4368 if (!Mask || !Mask->getUniqueInteger().isPowerOf2())
4370 DEBUG(dbgs() << "found and; icmp ?,0; brcc\n"); DEBUG(BB->dump());
4372 // Push the "and; icmp" for any users that are conditional branches.
4373 // Since there can only be one branch use per BB, we don't need to keep
4374 // track of which BBs we insert into.
4375 for (Value::use_iterator UI = Cmp->use_begin(), E = Cmp->use_end();
4379 BranchInst *BrccUser = dyn_cast<BranchInst>(*UI);
4381 if (!BrccUser || !BrccUser->isConditional())
4383 BasicBlock *UserBB = BrccUser->getParent();
4384 if (UserBB == BB) continue;
4385 DEBUG(dbgs() << "found Brcc use\n");
4387 // Sink the "and; icmp" to use.
4389 BinaryOperator *NewAnd =
4390 BinaryOperator::CreateAnd(And->getOperand(0), And->getOperand(1), "",
4393 CmpInst::Create(Cmp->getOpcode(), Cmp->getPredicate(), NewAnd, Zero,
4397 DEBUG(BrccUser->getParent()->dump());
4403 /// \brief Retrieve the probabilities of a conditional branch. Returns true on
4404 /// success, or returns false if no or invalid metadata was found.
4405 static bool extractBranchMetadata(BranchInst *BI,
4406 uint64_t &ProbTrue, uint64_t &ProbFalse) {
4407 assert(BI->isConditional() &&
4408 "Looking for probabilities on unconditional branch?");
4409 auto *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
4410 if (!ProfileData || ProfileData->getNumOperands() != 3)
4413 const auto *CITrue =
4414 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(1));
4415 const auto *CIFalse =
4416 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(2));
4417 if (!CITrue || !CIFalse)
4420 ProbTrue = CITrue->getValue().getZExtValue();
4421 ProbFalse = CIFalse->getValue().getZExtValue();
4426 /// \brief Scale down both weights to fit into uint32_t.
4427 static void scaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) {
4428 uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse;
4429 uint32_t Scale = (NewMax / UINT32_MAX) + 1;
4430 NewTrue = NewTrue / Scale;
4431 NewFalse = NewFalse / Scale;
4434 /// \brief Some targets prefer to split a conditional branch like:
4436 /// %0 = icmp ne i32 %a, 0
4437 /// %1 = icmp ne i32 %b, 0
4438 /// %or.cond = or i1 %0, %1
4439 /// br i1 %or.cond, label %TrueBB, label %FalseBB
4441 /// into multiple branch instructions like:
4444 /// %0 = icmp ne i32 %a, 0
4445 /// br i1 %0, label %TrueBB, label %bb2
4447 /// %1 = icmp ne i32 %b, 0
4448 /// br i1 %1, label %TrueBB, label %FalseBB
4450 /// This usually allows instruction selection to do even further optimizations
4451 /// and combine the compare with the branch instruction. Currently this is
4452 /// applied for targets which have "cheap" jump instructions.
4454 /// FIXME: Remove the (equivalent?) implementation in SelectionDAG.
4456 bool CodeGenPrepare::splitBranchCondition(Function &F) {
4457 if (!TM || TM->Options.EnableFastISel != true ||
4458 !TLI || TLI->isJumpExpensive())
4461 bool MadeChange = false;
4462 for (auto &BB : F) {
4463 // Does this BB end with the following?
4464 // %cond1 = icmp|fcmp|binary instruction ...
4465 // %cond2 = icmp|fcmp|binary instruction ...
4466 // %cond.or = or|and i1 %cond1, cond2
4467 // br i1 %cond.or label %dest1, label %dest2"
4468 BinaryOperator *LogicOp;
4469 BasicBlock *TBB, *FBB;
4470 if (!match(BB.getTerminator(), m_Br(m_OneUse(m_BinOp(LogicOp)), TBB, FBB)))
4474 Value *Cond1, *Cond2;
4475 if (match(LogicOp, m_And(m_OneUse(m_Value(Cond1)),
4476 m_OneUse(m_Value(Cond2)))))
4477 Opc = Instruction::And;
4478 else if (match(LogicOp, m_Or(m_OneUse(m_Value(Cond1)),
4479 m_OneUse(m_Value(Cond2)))))
4480 Opc = Instruction::Or;
4484 if (!match(Cond1, m_CombineOr(m_Cmp(), m_BinOp())) ||
4485 !match(Cond2, m_CombineOr(m_Cmp(), m_BinOp())) )
4488 DEBUG(dbgs() << "Before branch condition splitting\n"; BB.dump());
4491 auto *InsertBefore = std::next(Function::iterator(BB))
4492 .getNodePtrUnchecked();
4493 auto TmpBB = BasicBlock::Create(BB.getContext(),
4494 BB.getName() + ".cond.split",
4495 BB.getParent(), InsertBefore);
4497 // Update original basic block by using the first condition directly by the
4498 // branch instruction and removing the no longer needed and/or instruction.
4499 auto *Br1 = cast<BranchInst>(BB.getTerminator());
4500 Br1->setCondition(Cond1);
4501 LogicOp->eraseFromParent();
4503 // Depending on the conditon we have to either replace the true or the false
4504 // successor of the original branch instruction.
4505 if (Opc == Instruction::And)
4506 Br1->setSuccessor(0, TmpBB);
4508 Br1->setSuccessor(1, TmpBB);
4510 // Fill in the new basic block.
4511 auto *Br2 = IRBuilder<>(TmpBB).CreateCondBr(Cond2, TBB, FBB);
4512 if (auto *I = dyn_cast<Instruction>(Cond2)) {
4513 I->removeFromParent();
4514 I->insertBefore(Br2);
4517 // Update PHI nodes in both successors. The original BB needs to be
4518 // replaced in one succesor's PHI nodes, because the branch comes now from
4519 // the newly generated BB (NewBB). In the other successor we need to add one
4520 // incoming edge to the PHI nodes, because both branch instructions target
4521 // now the same successor. Depending on the original branch condition
4522 // (and/or) we have to swap the successors (TrueDest, FalseDest), so that
4523 // we perfrom the correct update for the PHI nodes.
4524 // This doesn't change the successor order of the just created branch
4525 // instruction (or any other instruction).
4526 if (Opc == Instruction::Or)
4527 std::swap(TBB, FBB);
4529 // Replace the old BB with the new BB.
4530 for (auto &I : *TBB) {
4531 PHINode *PN = dyn_cast<PHINode>(&I);
4535 while ((i = PN->getBasicBlockIndex(&BB)) >= 0)
4536 PN->setIncomingBlock(i, TmpBB);
4539 // Add another incoming edge form the new BB.
4540 for (auto &I : *FBB) {
4541 PHINode *PN = dyn_cast<PHINode>(&I);
4544 auto *Val = PN->getIncomingValueForBlock(&BB);
4545 PN->addIncoming(Val, TmpBB);
4548 // Update the branch weights (from SelectionDAGBuilder::
4549 // FindMergedConditions).
4550 if (Opc == Instruction::Or) {
4551 // Codegen X | Y as:
4560 // We have flexibility in setting Prob for BB1 and Prob for NewBB.
4561 // The requirement is that
4562 // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB)
4563 // = TrueProb for orignal BB.
4564 // Assuming the orignal weights are A and B, one choice is to set BB1's
4565 // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice
4567 // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB.
4568 // Another choice is to assume TrueProb for BB1 equals to TrueProb for
4569 // TmpBB, but the math is more complicated.
4570 uint64_t TrueWeight, FalseWeight;
4571 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
4572 uint64_t NewTrueWeight = TrueWeight;
4573 uint64_t NewFalseWeight = TrueWeight + 2 * FalseWeight;
4574 scaleWeights(NewTrueWeight, NewFalseWeight);
4575 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
4576 .createBranchWeights(TrueWeight, FalseWeight));
4578 NewTrueWeight = TrueWeight;
4579 NewFalseWeight = 2 * FalseWeight;
4580 scaleWeights(NewTrueWeight, NewFalseWeight);
4581 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
4582 .createBranchWeights(TrueWeight, FalseWeight));
4585 // Codegen X & Y as:
4593 // This requires creation of TmpBB after CurBB.
4595 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
4596 // The requirement is that
4597 // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB)
4598 // = FalseProb for orignal BB.
4599 // Assuming the orignal weights are A and B, one choice is to set BB1's
4600 // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice
4602 // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB.
4603 uint64_t TrueWeight, FalseWeight;
4604 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
4605 uint64_t NewTrueWeight = 2 * TrueWeight + FalseWeight;
4606 uint64_t NewFalseWeight = FalseWeight;
4607 scaleWeights(NewTrueWeight, NewFalseWeight);
4608 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
4609 .createBranchWeights(TrueWeight, FalseWeight));
4611 NewTrueWeight = 2 * TrueWeight;
4612 NewFalseWeight = FalseWeight;
4613 scaleWeights(NewTrueWeight, NewFalseWeight);
4614 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
4615 .createBranchWeights(TrueWeight, FalseWeight));
4619 // Request DOM Tree update.
4620 // Note: No point in getting fancy here, since the DT info is never
4621 // available to CodeGenPrepare and the existing update code is broken
4627 DEBUG(dbgs() << "After branch condition splitting\n"; BB.dump();