1 //===- CodeGenPrepare.cpp - Prepare a function for code generation --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass munges the code in the input function to better prepare it for
11 // SelectionDAG-based code generation. This works around limitations in it's
12 // basic-block-at-a-time approach. It should eventually be removed.
14 //===----------------------------------------------------------------------===//
16 #include "llvm/CodeGen/Passes.h"
17 #include "llvm/ADT/DenseMap.h"
18 #include "llvm/ADT/SmallSet.h"
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/Analysis/InstructionSimplify.h"
21 #include "llvm/Analysis/TargetLibraryInfo.h"
22 #include "llvm/Analysis/TargetTransformInfo.h"
23 #include "llvm/Analysis/ValueTracking.h"
24 #include "llvm/IR/CallSite.h"
25 #include "llvm/IR/Constants.h"
26 #include "llvm/IR/DataLayout.h"
27 #include "llvm/IR/DerivedTypes.h"
28 #include "llvm/IR/Dominators.h"
29 #include "llvm/IR/Function.h"
30 #include "llvm/IR/GetElementPtrTypeIterator.h"
31 #include "llvm/IR/IRBuilder.h"
32 #include "llvm/IR/InlineAsm.h"
33 #include "llvm/IR/Instructions.h"
34 #include "llvm/IR/IntrinsicInst.h"
35 #include "llvm/IR/MDBuilder.h"
36 #include "llvm/IR/PatternMatch.h"
37 #include "llvm/IR/Statepoint.h"
38 #include "llvm/IR/ValueHandle.h"
39 #include "llvm/IR/ValueMap.h"
40 #include "llvm/Pass.h"
41 #include "llvm/Support/CommandLine.h"
42 #include "llvm/Support/Debug.h"
43 #include "llvm/Support/raw_ostream.h"
44 #include "llvm/Target/TargetLowering.h"
45 #include "llvm/Target/TargetSubtargetInfo.h"
46 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
47 #include "llvm/Transforms/Utils/BuildLibCalls.h"
48 #include "llvm/Transforms/Utils/BypassSlowDivision.h"
49 #include "llvm/Transforms/Utils/Local.h"
50 #include "llvm/Transforms/Utils/SimplifyLibCalls.h"
52 using namespace llvm::PatternMatch;
54 #define DEBUG_TYPE "codegenprepare"
56 STATISTIC(NumBlocksElim, "Number of blocks eliminated");
57 STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated");
58 STATISTIC(NumGEPsElim, "Number of GEPs converted to casts");
59 STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "
61 STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "
63 STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "
64 "computations were sunk");
65 STATISTIC(NumExtsMoved, "Number of [s|z]ext instructions combined with loads");
66 STATISTIC(NumExtUses, "Number of uses of [s|z]ext instructions optimized");
67 STATISTIC(NumRetsDup, "Number of return instructions duplicated");
68 STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved");
69 STATISTIC(NumSelectsExpanded, "Number of selects turned into branches");
70 STATISTIC(NumAndCmpsMoved, "Number of and/cmp's pushed into branches");
71 STATISTIC(NumStoreExtractExposed, "Number of store(extractelement) exposed");
73 static cl::opt<bool> DisableBranchOpts(
74 "disable-cgp-branch-opts", cl::Hidden, cl::init(false),
75 cl::desc("Disable branch optimizations in CodeGenPrepare"));
78 DisableGCOpts("disable-cgp-gc-opts", cl::Hidden, cl::init(false),
79 cl::desc("Disable GC optimizations in CodeGenPrepare"));
81 static cl::opt<bool> DisableSelectToBranch(
82 "disable-cgp-select2branch", cl::Hidden, cl::init(false),
83 cl::desc("Disable select to branch conversion."));
85 static cl::opt<bool> AddrSinkUsingGEPs(
86 "addr-sink-using-gep", cl::Hidden, cl::init(false),
87 cl::desc("Address sinking in CGP using GEPs."));
89 static cl::opt<bool> EnableAndCmpSinking(
90 "enable-andcmp-sinking", cl::Hidden, cl::init(true),
91 cl::desc("Enable sinkinig and/cmp into branches."));
93 static cl::opt<bool> DisableStoreExtract(
94 "disable-cgp-store-extract", cl::Hidden, cl::init(false),
95 cl::desc("Disable store(extract) optimizations in CodeGenPrepare"));
97 static cl::opt<bool> StressStoreExtract(
98 "stress-cgp-store-extract", cl::Hidden, cl::init(false),
99 cl::desc("Stress test store(extract) optimizations in CodeGenPrepare"));
101 static cl::opt<bool> DisableExtLdPromotion(
102 "disable-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
103 cl::desc("Disable ext(promotable(ld)) -> promoted(ext(ld)) optimization in "
106 static cl::opt<bool> StressExtLdPromotion(
107 "stress-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
108 cl::desc("Stress test ext(promotable(ld)) -> promoted(ext(ld)) "
109 "optimization in CodeGenPrepare"));
112 typedef SmallPtrSet<Instruction *, 16> SetOfInstrs;
113 typedef PointerIntPair<Type *, 1, bool> TypeIsSExt;
114 typedef DenseMap<Instruction *, TypeIsSExt> InstrToOrigTy;
115 class TypePromotionTransaction;
117 class CodeGenPrepare : public FunctionPass {
118 const TargetMachine *TM;
119 const TargetLowering *TLI;
120 const TargetTransformInfo *TTI;
121 const TargetLibraryInfo *TLInfo;
123 /// As we scan instructions optimizing them, this is the next instruction
124 /// to optimize. Transforms that can invalidate this should update it.
125 BasicBlock::iterator CurInstIterator;
127 /// Keeps track of non-local addresses that have been sunk into a block.
128 /// This allows us to avoid inserting duplicate code for blocks with
129 /// multiple load/stores of the same address.
130 ValueMap<Value*, Value*> SunkAddrs;
132 /// Keeps track of all instructions inserted for the current function.
133 SetOfInstrs InsertedInsts;
134 /// Keeps track of the type of the related instruction before their
135 /// promotion for the current function.
136 InstrToOrigTy PromotedInsts;
138 /// True if CFG is modified in any way.
141 /// True if optimizing for size.
144 /// DataLayout for the Function being processed.
145 const DataLayout *DL;
148 static char ID; // Pass identification, replacement for typeid
149 explicit CodeGenPrepare(const TargetMachine *TM = nullptr)
150 : FunctionPass(ID), TM(TM), TLI(nullptr), TTI(nullptr), DL(nullptr) {
151 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
153 bool runOnFunction(Function &F) override;
155 const char *getPassName() const override { return "CodeGen Prepare"; }
157 void getAnalysisUsage(AnalysisUsage &AU) const override {
158 AU.addPreserved<DominatorTreeWrapperPass>();
159 AU.addRequired<TargetLibraryInfoWrapperPass>();
160 AU.addRequired<TargetTransformInfoWrapperPass>();
164 bool eliminateFallThrough(Function &F);
165 bool eliminateMostlyEmptyBlocks(Function &F);
166 bool canMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
167 void eliminateMostlyEmptyBlock(BasicBlock *BB);
168 bool optimizeBlock(BasicBlock &BB, bool& ModifiedDT);
169 bool optimizeInst(Instruction *I, bool& ModifiedDT);
170 bool optimizeMemoryInst(Instruction *I, Value *Addr,
171 Type *AccessTy, unsigned AS);
172 bool optimizeInlineAsmInst(CallInst *CS);
173 bool optimizeCallInst(CallInst *CI, bool& ModifiedDT);
174 bool moveExtToFormExtLoad(Instruction *&I);
175 bool optimizeExtUses(Instruction *I);
176 bool optimizeSelectInst(SelectInst *SI);
177 bool optimizeShuffleVectorInst(ShuffleVectorInst *SI);
178 bool optimizeExtractElementInst(Instruction *Inst);
179 bool dupRetToEnableTailCallOpts(BasicBlock *BB);
180 bool placeDbgValues(Function &F);
181 bool sinkAndCmp(Function &F);
182 bool extLdPromotion(TypePromotionTransaction &TPT, LoadInst *&LI,
184 const SmallVectorImpl<Instruction *> &Exts,
185 unsigned CreatedInstCost);
186 bool splitBranchCondition(Function &F);
187 bool simplifyOffsetableRelocate(Instruction &I);
188 void stripInvariantGroupMetadata(Instruction &I);
192 char CodeGenPrepare::ID = 0;
193 INITIALIZE_TM_PASS(CodeGenPrepare, "codegenprepare",
194 "Optimize for code generation", false, false)
196 FunctionPass *llvm::createCodeGenPreparePass(const TargetMachine *TM) {
197 return new CodeGenPrepare(TM);
200 bool CodeGenPrepare::runOnFunction(Function &F) {
201 if (skipOptnoneFunction(F))
204 DL = &F.getParent()->getDataLayout();
206 bool EverMadeChange = false;
207 // Clear per function information.
208 InsertedInsts.clear();
209 PromotedInsts.clear();
213 TLI = TM->getSubtargetImpl(F)->getTargetLowering();
214 TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
215 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
216 OptSize = F.optForSize();
218 /// This optimization identifies DIV instructions that can be
219 /// profitably bypassed and carried out with a shorter, faster divide.
220 if (!OptSize && TLI && TLI->isSlowDivBypassed()) {
221 const DenseMap<unsigned int, unsigned int> &BypassWidths =
222 TLI->getBypassSlowDivWidths();
223 for (Function::iterator I = F.begin(); I != F.end(); I++)
224 EverMadeChange |= bypassSlowDivision(F, I, BypassWidths);
227 // Eliminate blocks that contain only PHI nodes and an
228 // unconditional branch.
229 EverMadeChange |= eliminateMostlyEmptyBlocks(F);
231 // llvm.dbg.value is far away from the value then iSel may not be able
232 // handle it properly. iSel will drop llvm.dbg.value if it can not
233 // find a node corresponding to the value.
234 EverMadeChange |= placeDbgValues(F);
236 // If there is a mask, compare against zero, and branch that can be combined
237 // into a single target instruction, push the mask and compare into branch
238 // users. Do this before OptimizeBlock -> OptimizeInst ->
239 // OptimizeCmpExpression, which perturbs the pattern being searched for.
240 if (!DisableBranchOpts) {
241 EverMadeChange |= sinkAndCmp(F);
242 EverMadeChange |= splitBranchCondition(F);
245 bool MadeChange = true;
248 for (Function::iterator I = F.begin(); I != F.end(); ) {
249 BasicBlock *BB = &*I++;
250 bool ModifiedDTOnIteration = false;
251 MadeChange |= optimizeBlock(*BB, ModifiedDTOnIteration);
253 // Restart BB iteration if the dominator tree of the Function was changed
254 if (ModifiedDTOnIteration)
257 EverMadeChange |= MadeChange;
262 if (!DisableBranchOpts) {
264 SmallPtrSet<BasicBlock*, 8> WorkList;
265 for (BasicBlock &BB : F) {
266 SmallVector<BasicBlock *, 2> Successors(succ_begin(&BB), succ_end(&BB));
267 MadeChange |= ConstantFoldTerminator(&BB, true);
268 if (!MadeChange) continue;
270 for (SmallVectorImpl<BasicBlock*>::iterator
271 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
272 if (pred_begin(*II) == pred_end(*II))
273 WorkList.insert(*II);
276 // Delete the dead blocks and any of their dead successors.
277 MadeChange |= !WorkList.empty();
278 while (!WorkList.empty()) {
279 BasicBlock *BB = *WorkList.begin();
281 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
285 for (SmallVectorImpl<BasicBlock*>::iterator
286 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
287 if (pred_begin(*II) == pred_end(*II))
288 WorkList.insert(*II);
291 // Merge pairs of basic blocks with unconditional branches, connected by
293 if (EverMadeChange || MadeChange)
294 MadeChange |= eliminateFallThrough(F);
296 EverMadeChange |= MadeChange;
299 if (!DisableGCOpts) {
300 SmallVector<Instruction *, 2> Statepoints;
301 for (BasicBlock &BB : F)
302 for (Instruction &I : BB)
304 Statepoints.push_back(&I);
305 for (auto &I : Statepoints)
306 EverMadeChange |= simplifyOffsetableRelocate(*I);
309 return EverMadeChange;
312 /// Merge basic blocks which are connected by a single edge, where one of the
313 /// basic blocks has a single successor pointing to the other basic block,
314 /// which has a single predecessor.
315 bool CodeGenPrepare::eliminateFallThrough(Function &F) {
316 bool Changed = false;
317 // Scan all of the blocks in the function, except for the entry block.
318 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
319 BasicBlock *BB = &*I++;
320 // If the destination block has a single pred, then this is a trivial
321 // edge, just collapse it.
322 BasicBlock *SinglePred = BB->getSinglePredecessor();
324 // Don't merge if BB's address is taken.
325 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
327 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
328 if (Term && !Term->isConditional()) {
330 DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n");
331 // Remember if SinglePred was the entry block of the function.
332 // If so, we will need to move BB back to the entry position.
333 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
334 MergeBasicBlockIntoOnlyPred(BB, nullptr);
336 if (isEntry && BB != &BB->getParent()->getEntryBlock())
337 BB->moveBefore(&BB->getParent()->getEntryBlock());
339 // We have erased a block. Update the iterator.
340 I = BB->getIterator();
346 /// Eliminate blocks that contain only PHI nodes, debug info directives, and an
347 /// unconditional branch. Passes before isel (e.g. LSR/loopsimplify) often split
348 /// edges in ways that are non-optimal for isel. Start by eliminating these
349 /// blocks so we can split them the way we want them.
350 bool CodeGenPrepare::eliminateMostlyEmptyBlocks(Function &F) {
351 bool MadeChange = false;
352 // Note that this intentionally skips the entry block.
353 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
354 BasicBlock *BB = &*I++;
356 // If this block doesn't end with an uncond branch, ignore it.
357 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
358 if (!BI || !BI->isUnconditional())
361 // If the instruction before the branch (skipping debug info) isn't a phi
362 // node, then other stuff is happening here.
363 BasicBlock::iterator BBI = BI->getIterator();
364 if (BBI != BB->begin()) {
366 while (isa<DbgInfoIntrinsic>(BBI)) {
367 if (BBI == BB->begin())
371 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
375 // Do not break infinite loops.
376 BasicBlock *DestBB = BI->getSuccessor(0);
380 if (!canMergeBlocks(BB, DestBB))
383 eliminateMostlyEmptyBlock(BB);
389 /// Return true if we can merge BB into DestBB if there is a single
390 /// unconditional branch between them, and BB contains no other non-phi
392 bool CodeGenPrepare::canMergeBlocks(const BasicBlock *BB,
393 const BasicBlock *DestBB) const {
394 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
395 // the successor. If there are more complex condition (e.g. preheaders),
396 // don't mess around with them.
397 BasicBlock::const_iterator BBI = BB->begin();
398 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
399 for (const User *U : PN->users()) {
400 const Instruction *UI = cast<Instruction>(U);
401 if (UI->getParent() != DestBB || !isa<PHINode>(UI))
403 // If User is inside DestBB block and it is a PHINode then check
404 // incoming value. If incoming value is not from BB then this is
405 // a complex condition (e.g. preheaders) we want to avoid here.
406 if (UI->getParent() == DestBB) {
407 if (const PHINode *UPN = dyn_cast<PHINode>(UI))
408 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
409 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
410 if (Insn && Insn->getParent() == BB &&
411 Insn->getParent() != UPN->getIncomingBlock(I))
418 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
419 // and DestBB may have conflicting incoming values for the block. If so, we
420 // can't merge the block.
421 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
422 if (!DestBBPN) return true; // no conflict.
424 // Collect the preds of BB.
425 SmallPtrSet<const BasicBlock*, 16> BBPreds;
426 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
427 // It is faster to get preds from a PHI than with pred_iterator.
428 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
429 BBPreds.insert(BBPN->getIncomingBlock(i));
431 BBPreds.insert(pred_begin(BB), pred_end(BB));
434 // Walk the preds of DestBB.
435 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
436 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
437 if (BBPreds.count(Pred)) { // Common predecessor?
438 BBI = DestBB->begin();
439 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
440 const Value *V1 = PN->getIncomingValueForBlock(Pred);
441 const Value *V2 = PN->getIncomingValueForBlock(BB);
443 // If V2 is a phi node in BB, look up what the mapped value will be.
444 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
445 if (V2PN->getParent() == BB)
446 V2 = V2PN->getIncomingValueForBlock(Pred);
448 // If there is a conflict, bail out.
449 if (V1 != V2) return false;
458 /// Eliminate a basic block that has only phi's and an unconditional branch in
460 void CodeGenPrepare::eliminateMostlyEmptyBlock(BasicBlock *BB) {
461 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
462 BasicBlock *DestBB = BI->getSuccessor(0);
464 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
466 // If the destination block has a single pred, then this is a trivial edge,
468 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
469 if (SinglePred != DestBB) {
470 // Remember if SinglePred was the entry block of the function. If so, we
471 // will need to move BB back to the entry position.
472 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
473 MergeBasicBlockIntoOnlyPred(DestBB, nullptr);
475 if (isEntry && BB != &BB->getParent()->getEntryBlock())
476 BB->moveBefore(&BB->getParent()->getEntryBlock());
478 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
483 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
484 // to handle the new incoming edges it is about to have.
486 for (BasicBlock::iterator BBI = DestBB->begin();
487 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
488 // Remove the incoming value for BB, and remember it.
489 Value *InVal = PN->removeIncomingValue(BB, false);
491 // Two options: either the InVal is a phi node defined in BB or it is some
492 // value that dominates BB.
493 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
494 if (InValPhi && InValPhi->getParent() == BB) {
495 // Add all of the input values of the input PHI as inputs of this phi.
496 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
497 PN->addIncoming(InValPhi->getIncomingValue(i),
498 InValPhi->getIncomingBlock(i));
500 // Otherwise, add one instance of the dominating value for each edge that
501 // we will be adding.
502 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
503 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
504 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
506 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
507 PN->addIncoming(InVal, *PI);
512 // The PHIs are now updated, change everything that refers to BB to use
513 // DestBB and remove BB.
514 BB->replaceAllUsesWith(DestBB);
515 BB->eraseFromParent();
518 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
521 // Computes a map of base pointer relocation instructions to corresponding
522 // derived pointer relocation instructions given a vector of all relocate calls
523 static void computeBaseDerivedRelocateMap(
524 const SmallVectorImpl<User *> &AllRelocateCalls,
525 DenseMap<IntrinsicInst *, SmallVector<IntrinsicInst *, 2>> &
527 // Collect information in two maps: one primarily for locating the base object
528 // while filling the second map; the second map is the final structure holding
529 // a mapping between Base and corresponding Derived relocate calls
530 DenseMap<std::pair<unsigned, unsigned>, IntrinsicInst *> RelocateIdxMap;
531 for (auto &U : AllRelocateCalls) {
532 GCRelocateOperands ThisRelocate(U);
533 IntrinsicInst *I = cast<IntrinsicInst>(U);
534 auto K = std::make_pair(ThisRelocate.getBasePtrIndex(),
535 ThisRelocate.getDerivedPtrIndex());
536 RelocateIdxMap.insert(std::make_pair(K, I));
538 for (auto &Item : RelocateIdxMap) {
539 std::pair<unsigned, unsigned> Key = Item.first;
540 if (Key.first == Key.second)
541 // Base relocation: nothing to insert
544 IntrinsicInst *I = Item.second;
545 auto BaseKey = std::make_pair(Key.first, Key.first);
547 // We're iterating over RelocateIdxMap so we cannot modify it.
548 auto MaybeBase = RelocateIdxMap.find(BaseKey);
549 if (MaybeBase == RelocateIdxMap.end())
550 // TODO: We might want to insert a new base object relocate and gep off
551 // that, if there are enough derived object relocates.
554 RelocateInstMap[MaybeBase->second].push_back(I);
558 // Accepts a GEP and extracts the operands into a vector provided they're all
559 // small integer constants
560 static bool getGEPSmallConstantIntOffsetV(GetElementPtrInst *GEP,
561 SmallVectorImpl<Value *> &OffsetV) {
562 for (unsigned i = 1; i < GEP->getNumOperands(); i++) {
563 // Only accept small constant integer operands
564 auto Op = dyn_cast<ConstantInt>(GEP->getOperand(i));
565 if (!Op || Op->getZExtValue() > 20)
569 for (unsigned i = 1; i < GEP->getNumOperands(); i++)
570 OffsetV.push_back(GEP->getOperand(i));
574 // Takes a RelocatedBase (base pointer relocation instruction) and Targets to
575 // replace, computes a replacement, and affects it.
577 simplifyRelocatesOffABase(IntrinsicInst *RelocatedBase,
578 const SmallVectorImpl<IntrinsicInst *> &Targets) {
579 bool MadeChange = false;
580 for (auto &ToReplace : Targets) {
581 GCRelocateOperands MasterRelocate(RelocatedBase);
582 GCRelocateOperands ThisRelocate(ToReplace);
584 assert(ThisRelocate.getBasePtrIndex() == MasterRelocate.getBasePtrIndex() &&
585 "Not relocating a derived object of the original base object");
586 if (ThisRelocate.getBasePtrIndex() == ThisRelocate.getDerivedPtrIndex()) {
587 // A duplicate relocate call. TODO: coalesce duplicates.
591 Value *Base = ThisRelocate.getBasePtr();
592 auto Derived = dyn_cast<GetElementPtrInst>(ThisRelocate.getDerivedPtr());
593 if (!Derived || Derived->getPointerOperand() != Base)
596 SmallVector<Value *, 2> OffsetV;
597 if (!getGEPSmallConstantIntOffsetV(Derived, OffsetV))
600 // Create a Builder and replace the target callsite with a gep
601 assert(RelocatedBase->getNextNode() && "Should always have one since it's not a terminator");
603 // Insert after RelocatedBase
604 IRBuilder<> Builder(RelocatedBase->getNextNode());
605 Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
607 // If gc_relocate does not match the actual type, cast it to the right type.
608 // In theory, there must be a bitcast after gc_relocate if the type does not
609 // match, and we should reuse it to get the derived pointer. But it could be
613 // %g1 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
618 // %g2 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
622 // %p1 = phi i8 addrspace(1)* [ %g1, %bb1 ], [ %g2, %bb2 ]
623 // %cast = bitcast i8 addrspace(1)* %p1 in to i32 addrspace(1)*
625 // In this case, we can not find the bitcast any more. So we insert a new bitcast
626 // no matter there is already one or not. In this way, we can handle all cases, and
627 // the extra bitcast should be optimized away in later passes.
628 Instruction *ActualRelocatedBase = RelocatedBase;
629 if (RelocatedBase->getType() != Base->getType()) {
630 ActualRelocatedBase =
631 cast<Instruction>(Builder.CreateBitCast(RelocatedBase, Base->getType()));
633 Value *Replacement = Builder.CreateGEP(
634 Derived->getSourceElementType(), ActualRelocatedBase, makeArrayRef(OffsetV));
635 Instruction *ReplacementInst = cast<Instruction>(Replacement);
636 Replacement->takeName(ToReplace);
637 // If the newly generated derived pointer's type does not match the original derived
638 // pointer's type, cast the new derived pointer to match it. Same reasoning as above.
639 Instruction *ActualReplacement = ReplacementInst;
640 if (ReplacementInst->getType() != ToReplace->getType()) {
642 cast<Instruction>(Builder.CreateBitCast(ReplacementInst, ToReplace->getType()));
644 ToReplace->replaceAllUsesWith(ActualReplacement);
645 ToReplace->eraseFromParent();
655 // %ptr = gep %base + 15
656 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
657 // %base' = relocate(%tok, i32 4, i32 4)
658 // %ptr' = relocate(%tok, i32 4, i32 5)
664 // %ptr = gep %base + 15
665 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
666 // %base' = gc.relocate(%tok, i32 4, i32 4)
667 // %ptr' = gep %base' + 15
669 bool CodeGenPrepare::simplifyOffsetableRelocate(Instruction &I) {
670 bool MadeChange = false;
671 SmallVector<User *, 2> AllRelocateCalls;
673 for (auto *U : I.users())
674 if (isGCRelocate(dyn_cast<Instruction>(U)))
675 // Collect all the relocate calls associated with a statepoint
676 AllRelocateCalls.push_back(U);
678 // We need atleast one base pointer relocation + one derived pointer
679 // relocation to mangle
680 if (AllRelocateCalls.size() < 2)
683 // RelocateInstMap is a mapping from the base relocate instruction to the
684 // corresponding derived relocate instructions
685 DenseMap<IntrinsicInst *, SmallVector<IntrinsicInst *, 2>> RelocateInstMap;
686 computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap);
687 if (RelocateInstMap.empty())
690 for (auto &Item : RelocateInstMap)
691 // Item.first is the RelocatedBase to offset against
692 // Item.second is the vector of Targets to replace
693 MadeChange = simplifyRelocatesOffABase(Item.first, Item.second);
697 /// SinkCast - Sink the specified cast instruction into its user blocks
698 static bool SinkCast(CastInst *CI) {
699 BasicBlock *DefBB = CI->getParent();
701 /// InsertedCasts - Only insert a cast in each block once.
702 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
704 bool MadeChange = false;
705 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
707 Use &TheUse = UI.getUse();
708 Instruction *User = cast<Instruction>(*UI);
710 // Figure out which BB this cast is used in. For PHI's this is the
711 // appropriate predecessor block.
712 BasicBlock *UserBB = User->getParent();
713 if (PHINode *PN = dyn_cast<PHINode>(User)) {
714 UserBB = PN->getIncomingBlock(TheUse);
717 // Preincrement use iterator so we don't invalidate it.
720 // If this user is in the same block as the cast, don't change the cast.
721 if (UserBB == DefBB) continue;
723 // If we have already inserted a cast into this block, use it.
724 CastInst *&InsertedCast = InsertedCasts[UserBB];
727 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
728 assert(InsertPt != UserBB->end());
729 InsertedCast = CastInst::Create(CI->getOpcode(), CI->getOperand(0),
730 CI->getType(), "", &*InsertPt);
733 // Replace a use of the cast with a use of the new cast.
734 TheUse = InsertedCast;
739 // If we removed all uses, nuke the cast.
740 if (CI->use_empty()) {
741 CI->eraseFromParent();
748 /// If the specified cast instruction is a noop copy (e.g. it's casting from
749 /// one pointer type to another, i32->i8 on PPC), sink it into user blocks to
750 /// reduce the number of virtual registers that must be created and coalesced.
752 /// Return true if any changes are made.
754 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI,
755 const DataLayout &DL) {
756 // If this is a noop copy,
757 EVT SrcVT = TLI.getValueType(DL, CI->getOperand(0)->getType());
758 EVT DstVT = TLI.getValueType(DL, CI->getType());
760 // This is an fp<->int conversion?
761 if (SrcVT.isInteger() != DstVT.isInteger())
764 // If this is an extension, it will be a zero or sign extension, which
766 if (SrcVT.bitsLT(DstVT)) return false;
768 // If these values will be promoted, find out what they will be promoted
769 // to. This helps us consider truncates on PPC as noop copies when they
771 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
772 TargetLowering::TypePromoteInteger)
773 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
774 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
775 TargetLowering::TypePromoteInteger)
776 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
778 // If, after promotion, these are the same types, this is a noop copy.
785 /// Try to combine CI into a call to the llvm.uadd.with.overflow intrinsic if
788 /// Return true if any changes were made.
789 static bool CombineUAddWithOverflow(CmpInst *CI) {
793 m_UAddWithOverflow(m_Value(A), m_Value(B), m_Instruction(AddI))))
796 Type *Ty = AddI->getType();
797 if (!isa<IntegerType>(Ty))
800 // We don't want to move around uses of condition values this late, so we we
801 // check if it is legal to create the call to the intrinsic in the basic
802 // block containing the icmp:
804 if (AddI->getParent() != CI->getParent() && !AddI->hasOneUse())
808 // Someday m_UAddWithOverflow may get smarter, but this is a safe assumption
810 if (AddI->hasOneUse())
811 assert(*AddI->user_begin() == CI && "expected!");
814 Module *M = CI->getParent()->getParent()->getParent();
815 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
817 auto *InsertPt = AddI->hasOneUse() ? CI : AddI;
819 auto *UAddWithOverflow =
820 CallInst::Create(F, {A, B}, "uadd.overflow", InsertPt);
821 auto *UAdd = ExtractValueInst::Create(UAddWithOverflow, 0, "uadd", InsertPt);
823 ExtractValueInst::Create(UAddWithOverflow, 1, "overflow", InsertPt);
825 CI->replaceAllUsesWith(Overflow);
826 AddI->replaceAllUsesWith(UAdd);
827 CI->eraseFromParent();
828 AddI->eraseFromParent();
832 /// Sink the given CmpInst into user blocks to reduce the number of virtual
833 /// registers that must be created and coalesced. This is a clear win except on
834 /// targets with multiple condition code registers (PowerPC), where it might
835 /// lose; some adjustment may be wanted there.
837 /// Return true if any changes are made.
838 static bool SinkCmpExpression(CmpInst *CI) {
839 BasicBlock *DefBB = CI->getParent();
841 /// Only insert a cmp in each block once.
842 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
844 bool MadeChange = false;
845 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
847 Use &TheUse = UI.getUse();
848 Instruction *User = cast<Instruction>(*UI);
850 // Preincrement use iterator so we don't invalidate it.
853 // Don't bother for PHI nodes.
854 if (isa<PHINode>(User))
857 // Figure out which BB this cmp is used in.
858 BasicBlock *UserBB = User->getParent();
860 // If this user is in the same block as the cmp, don't change the cmp.
861 if (UserBB == DefBB) continue;
863 // If we have already inserted a cmp into this block, use it.
864 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
867 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
868 assert(InsertPt != UserBB->end());
870 CmpInst::Create(CI->getOpcode(), CI->getPredicate(),
871 CI->getOperand(0), CI->getOperand(1), "", &*InsertPt);
874 // Replace a use of the cmp with a use of the new cmp.
875 TheUse = InsertedCmp;
880 // If we removed all uses, nuke the cmp.
881 if (CI->use_empty()) {
882 CI->eraseFromParent();
889 static bool OptimizeCmpExpression(CmpInst *CI) {
890 if (SinkCmpExpression(CI))
893 if (CombineUAddWithOverflow(CI))
899 /// Check if the candidates could be combined with a shift instruction, which
901 /// 1. Truncate instruction
902 /// 2. And instruction and the imm is a mask of the low bits:
903 /// imm & (imm+1) == 0
904 static bool isExtractBitsCandidateUse(Instruction *User) {
905 if (!isa<TruncInst>(User)) {
906 if (User->getOpcode() != Instruction::And ||
907 !isa<ConstantInt>(User->getOperand(1)))
910 const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
912 if ((Cimm & (Cimm + 1)).getBoolValue())
918 /// Sink both shift and truncate instruction to the use of truncate's BB.
920 SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI,
921 DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts,
922 const TargetLowering &TLI, const DataLayout &DL) {
923 BasicBlock *UserBB = User->getParent();
924 DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
925 TruncInst *TruncI = dyn_cast<TruncInst>(User);
926 bool MadeChange = false;
928 for (Value::user_iterator TruncUI = TruncI->user_begin(),
929 TruncE = TruncI->user_end();
930 TruncUI != TruncE;) {
932 Use &TruncTheUse = TruncUI.getUse();
933 Instruction *TruncUser = cast<Instruction>(*TruncUI);
934 // Preincrement use iterator so we don't invalidate it.
938 int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
942 // If the use is actually a legal node, there will not be an
943 // implicit truncate.
944 // FIXME: always querying the result type is just an
945 // approximation; some nodes' legality is determined by the
946 // operand or other means. There's no good way to find out though.
947 if (TLI.isOperationLegalOrCustom(
948 ISDOpcode, TLI.getValueType(DL, TruncUser->getType(), true)))
951 // Don't bother for PHI nodes.
952 if (isa<PHINode>(TruncUser))
955 BasicBlock *TruncUserBB = TruncUser->getParent();
957 if (UserBB == TruncUserBB)
960 BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
961 CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
963 if (!InsertedShift && !InsertedTrunc) {
964 BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
965 assert(InsertPt != TruncUserBB->end());
967 if (ShiftI->getOpcode() == Instruction::AShr)
968 InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
971 InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
975 BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
977 assert(TruncInsertPt != TruncUserBB->end());
979 InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
980 TruncI->getType(), "", &*TruncInsertPt);
984 TruncTheUse = InsertedTrunc;
990 /// Sink the shift *right* instruction into user blocks if the uses could
991 /// potentially be combined with this shift instruction and generate BitExtract
992 /// instruction. It will only be applied if the architecture supports BitExtract
993 /// instruction. Here is an example:
995 /// %x.extract.shift = lshr i64 %arg1, 32
997 /// %x.extract.trunc = trunc i64 %x.extract.shift to i16
1001 /// %x.extract.shift.1 = lshr i64 %arg1, 32
1002 /// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
1004 /// CodeGen will recoginze the pattern in BB2 and generate BitExtract
1006 /// Return true if any changes are made.
1007 static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI,
1008 const TargetLowering &TLI,
1009 const DataLayout &DL) {
1010 BasicBlock *DefBB = ShiftI->getParent();
1012 /// Only insert instructions in each block once.
1013 DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts;
1015 bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(DL, ShiftI->getType()));
1017 bool MadeChange = false;
1018 for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
1020 Use &TheUse = UI.getUse();
1021 Instruction *User = cast<Instruction>(*UI);
1022 // Preincrement use iterator so we don't invalidate it.
1025 // Don't bother for PHI nodes.
1026 if (isa<PHINode>(User))
1029 if (!isExtractBitsCandidateUse(User))
1032 BasicBlock *UserBB = User->getParent();
1034 if (UserBB == DefBB) {
1035 // If the shift and truncate instruction are in the same BB. The use of
1036 // the truncate(TruncUse) may still introduce another truncate if not
1037 // legal. In this case, we would like to sink both shift and truncate
1038 // instruction to the BB of TruncUse.
1041 // i64 shift.result = lshr i64 opnd, imm
1042 // trunc.result = trunc shift.result to i16
1045 // ----> We will have an implicit truncate here if the architecture does
1046 // not have i16 compare.
1047 // cmp i16 trunc.result, opnd2
1049 if (isa<TruncInst>(User) && shiftIsLegal
1050 // If the type of the truncate is legal, no trucate will be
1051 // introduced in other basic blocks.
1053 (!TLI.isTypeLegal(TLI.getValueType(DL, User->getType()))))
1055 SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI, DL);
1059 // If we have already inserted a shift into this block, use it.
1060 BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
1062 if (!InsertedShift) {
1063 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1064 assert(InsertPt != UserBB->end());
1066 if (ShiftI->getOpcode() == Instruction::AShr)
1067 InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
1070 InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
1076 // Replace a use of the shift with a use of the new shift.
1077 TheUse = InsertedShift;
1080 // If we removed all uses, nuke the shift.
1081 if (ShiftI->use_empty())
1082 ShiftI->eraseFromParent();
1087 // Translate a masked load intrinsic like
1088 // <16 x i32 > @llvm.masked.load( <16 x i32>* %addr, i32 align,
1089 // <16 x i1> %mask, <16 x i32> %passthru)
1090 // to a chain of basic blocks, with loading element one-by-one if
1091 // the appropriate mask bit is set
1093 // %1 = bitcast i8* %addr to i32*
1094 // %2 = extractelement <16 x i1> %mask, i32 0
1095 // %3 = icmp eq i1 %2, true
1096 // br i1 %3, label %cond.load, label %else
1098 //cond.load: ; preds = %0
1099 // %4 = getelementptr i32* %1, i32 0
1100 // %5 = load i32* %4
1101 // %6 = insertelement <16 x i32> undef, i32 %5, i32 0
1104 //else: ; preds = %0, %cond.load
1105 // %res.phi.else = phi <16 x i32> [ %6, %cond.load ], [ undef, %0 ]
1106 // %7 = extractelement <16 x i1> %mask, i32 1
1107 // %8 = icmp eq i1 %7, true
1108 // br i1 %8, label %cond.load1, label %else2
1110 //cond.load1: ; preds = %else
1111 // %9 = getelementptr i32* %1, i32 1
1112 // %10 = load i32* %9
1113 // %11 = insertelement <16 x i32> %res.phi.else, i32 %10, i32 1
1116 //else2: ; preds = %else, %cond.load1
1117 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
1118 // %12 = extractelement <16 x i1> %mask, i32 2
1119 // %13 = icmp eq i1 %12, true
1120 // br i1 %13, label %cond.load4, label %else5
1122 static void ScalarizeMaskedLoad(CallInst *CI) {
1123 Value *Ptr = CI->getArgOperand(0);
1124 Value *Alignment = CI->getArgOperand(1);
1125 Value *Mask = CI->getArgOperand(2);
1126 Value *Src0 = CI->getArgOperand(3);
1128 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
1129 VectorType *VecType = dyn_cast<VectorType>(CI->getType());
1130 assert(VecType && "Unexpected return type of masked load intrinsic");
1132 Type *EltTy = CI->getType()->getVectorElementType();
1134 IRBuilder<> Builder(CI->getContext());
1135 Instruction *InsertPt = CI;
1136 BasicBlock *IfBlock = CI->getParent();
1137 BasicBlock *CondBlock = nullptr;
1138 BasicBlock *PrevIfBlock = CI->getParent();
1140 Builder.SetInsertPoint(InsertPt);
1141 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1143 // Short-cut if the mask is all-true.
1144 bool IsAllOnesMask = isa<Constant>(Mask) &&
1145 cast<Constant>(Mask)->isAllOnesValue();
1147 if (IsAllOnesMask) {
1148 Value *NewI = Builder.CreateAlignedLoad(Ptr, AlignVal);
1149 CI->replaceAllUsesWith(NewI);
1150 CI->eraseFromParent();
1154 // Adjust alignment for the scalar instruction.
1155 AlignVal = std::min(AlignVal, VecType->getScalarSizeInBits()/8);
1156 // Bitcast %addr fron i8* to EltTy*
1158 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
1159 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
1160 unsigned VectorWidth = VecType->getNumElements();
1162 Value *UndefVal = UndefValue::get(VecType);
1164 // The result vector
1165 Value *VResult = UndefVal;
1167 if (isa<ConstantVector>(Mask)) {
1168 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1169 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
1172 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1173 LoadInst* Load = Builder.CreateAlignedLoad(Gep, AlignVal);
1174 VResult = Builder.CreateInsertElement(VResult, Load,
1175 Builder.getInt32(Idx));
1177 Value *NewI = Builder.CreateSelect(Mask, VResult, Src0);
1178 CI->replaceAllUsesWith(NewI);
1179 CI->eraseFromParent();
1183 PHINode *Phi = nullptr;
1184 Value *PrevPhi = UndefVal;
1186 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1188 // Fill the "else" block, created in the previous iteration
1190 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
1191 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
1192 // %to_load = icmp eq i1 %mask_1, true
1193 // br i1 %to_load, label %cond.load, label %else
1196 Phi = Builder.CreatePHI(VecType, 2, "res.phi.else");
1197 Phi->addIncoming(VResult, CondBlock);
1198 Phi->addIncoming(PrevPhi, PrevIfBlock);
1203 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
1204 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1205 ConstantInt::get(Predicate->getType(), 1));
1207 // Create "cond" block
1209 // %EltAddr = getelementptr i32* %1, i32 0
1210 // %Elt = load i32* %EltAddr
1211 // VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx
1213 CondBlock = IfBlock->splitBasicBlock(InsertPt->getIterator(), "cond.load");
1214 Builder.SetInsertPoint(InsertPt);
1217 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1218 LoadInst *Load = Builder.CreateAlignedLoad(Gep, AlignVal);
1219 VResult = Builder.CreateInsertElement(VResult, Load, Builder.getInt32(Idx));
1221 // Create "else" block, fill it in the next iteration
1222 BasicBlock *NewIfBlock =
1223 CondBlock->splitBasicBlock(InsertPt->getIterator(), "else");
1224 Builder.SetInsertPoint(InsertPt);
1225 Instruction *OldBr = IfBlock->getTerminator();
1226 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1227 OldBr->eraseFromParent();
1228 PrevIfBlock = IfBlock;
1229 IfBlock = NewIfBlock;
1232 Phi = Builder.CreatePHI(VecType, 2, "res.phi.select");
1233 Phi->addIncoming(VResult, CondBlock);
1234 Phi->addIncoming(PrevPhi, PrevIfBlock);
1235 Value *NewI = Builder.CreateSelect(Mask, Phi, Src0);
1236 CI->replaceAllUsesWith(NewI);
1237 CI->eraseFromParent();
1240 // Translate a masked store intrinsic, like
1241 // void @llvm.masked.store(<16 x i32> %src, <16 x i32>* %addr, i32 align,
1243 // to a chain of basic blocks, that stores element one-by-one if
1244 // the appropriate mask bit is set
1246 // %1 = bitcast i8* %addr to i32*
1247 // %2 = extractelement <16 x i1> %mask, i32 0
1248 // %3 = icmp eq i1 %2, true
1249 // br i1 %3, label %cond.store, label %else
1251 // cond.store: ; preds = %0
1252 // %4 = extractelement <16 x i32> %val, i32 0
1253 // %5 = getelementptr i32* %1, i32 0
1254 // store i32 %4, i32* %5
1257 // else: ; preds = %0, %cond.store
1258 // %6 = extractelement <16 x i1> %mask, i32 1
1259 // %7 = icmp eq i1 %6, true
1260 // br i1 %7, label %cond.store1, label %else2
1262 // cond.store1: ; preds = %else
1263 // %8 = extractelement <16 x i32> %val, i32 1
1264 // %9 = getelementptr i32* %1, i32 1
1265 // store i32 %8, i32* %9
1268 static void ScalarizeMaskedStore(CallInst *CI) {
1269 Value *Src = CI->getArgOperand(0);
1270 Value *Ptr = CI->getArgOperand(1);
1271 Value *Alignment = CI->getArgOperand(2);
1272 Value *Mask = CI->getArgOperand(3);
1274 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
1275 VectorType *VecType = dyn_cast<VectorType>(Src->getType());
1276 assert(VecType && "Unexpected data type in masked store intrinsic");
1278 Type *EltTy = VecType->getElementType();
1280 IRBuilder<> Builder(CI->getContext());
1281 Instruction *InsertPt = CI;
1282 BasicBlock *IfBlock = CI->getParent();
1283 Builder.SetInsertPoint(InsertPt);
1284 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1286 // Short-cut if the mask is all-true.
1287 bool IsAllOnesMask = isa<Constant>(Mask) &&
1288 cast<Constant>(Mask)->isAllOnesValue();
1290 if (IsAllOnesMask) {
1291 Builder.CreateAlignedStore(Src, Ptr, AlignVal);
1292 CI->eraseFromParent();
1296 // Adjust alignment for the scalar instruction.
1297 AlignVal = std::max(AlignVal, VecType->getScalarSizeInBits()/8);
1298 // Bitcast %addr fron i8* to EltTy*
1300 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
1301 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
1302 unsigned VectorWidth = VecType->getNumElements();
1304 if (isa<ConstantVector>(Mask)) {
1305 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1306 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
1308 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx));
1310 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1311 Builder.CreateAlignedStore(OneElt, Gep, AlignVal);
1313 CI->eraseFromParent();
1317 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1319 // Fill the "else" block, created in the previous iteration
1321 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
1322 // %to_store = icmp eq i1 %mask_1, true
1323 // br i1 %to_store, label %cond.store, label %else
1325 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
1326 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1327 ConstantInt::get(Predicate->getType(), 1));
1329 // Create "cond" block
1331 // %OneElt = extractelement <16 x i32> %Src, i32 Idx
1332 // %EltAddr = getelementptr i32* %1, i32 0
1333 // %store i32 %OneElt, i32* %EltAddr
1335 BasicBlock *CondBlock =
1336 IfBlock->splitBasicBlock(InsertPt->getIterator(), "cond.store");
1337 Builder.SetInsertPoint(InsertPt);
1339 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx));
1341 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1342 Builder.CreateAlignedStore(OneElt, Gep, AlignVal);
1344 // Create "else" block, fill it in the next iteration
1345 BasicBlock *NewIfBlock =
1346 CondBlock->splitBasicBlock(InsertPt->getIterator(), "else");
1347 Builder.SetInsertPoint(InsertPt);
1348 Instruction *OldBr = IfBlock->getTerminator();
1349 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1350 OldBr->eraseFromParent();
1351 IfBlock = NewIfBlock;
1353 CI->eraseFromParent();
1356 // Translate a masked gather intrinsic like
1357 // <16 x i32 > @llvm.masked.gather.v16i32( <16 x i32*> %Ptrs, i32 4,
1358 // <16 x i1> %Mask, <16 x i32> %Src)
1359 // to a chain of basic blocks, with loading element one-by-one if
1360 // the appropriate mask bit is set
1362 // % Ptrs = getelementptr i32, i32* %base, <16 x i64> %ind
1363 // % Mask0 = extractelement <16 x i1> %Mask, i32 0
1364 // % ToLoad0 = icmp eq i1 % Mask0, true
1365 // br i1 % ToLoad0, label %cond.load, label %else
1368 // % Ptr0 = extractelement <16 x i32*> %Ptrs, i32 0
1369 // % Load0 = load i32, i32* % Ptr0, align 4
1370 // % Res0 = insertelement <16 x i32> undef, i32 % Load0, i32 0
1374 // %res.phi.else = phi <16 x i32>[% Res0, %cond.load], [undef, % 0]
1375 // % Mask1 = extractelement <16 x i1> %Mask, i32 1
1376 // % ToLoad1 = icmp eq i1 % Mask1, true
1377 // br i1 % ToLoad1, label %cond.load1, label %else2
1380 // % Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1
1381 // % Load1 = load i32, i32* % Ptr1, align 4
1382 // % Res1 = insertelement <16 x i32> %res.phi.else, i32 % Load1, i32 1
1385 // % Result = select <16 x i1> %Mask, <16 x i32> %res.phi.select, <16 x i32> %Src
1386 // ret <16 x i32> %Result
1387 static void ScalarizeMaskedGather(CallInst *CI) {
1388 Value *Ptrs = CI->getArgOperand(0);
1389 Value *Alignment = CI->getArgOperand(1);
1390 Value *Mask = CI->getArgOperand(2);
1391 Value *Src0 = CI->getArgOperand(3);
1393 VectorType *VecType = dyn_cast<VectorType>(CI->getType());
1395 assert(VecType && "Unexpected return type of masked load intrinsic");
1397 IRBuilder<> Builder(CI->getContext());
1398 Instruction *InsertPt = CI;
1399 BasicBlock *IfBlock = CI->getParent();
1400 BasicBlock *CondBlock = nullptr;
1401 BasicBlock *PrevIfBlock = CI->getParent();
1402 Builder.SetInsertPoint(InsertPt);
1403 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
1405 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1407 Value *UndefVal = UndefValue::get(VecType);
1409 // The result vector
1410 Value *VResult = UndefVal;
1411 unsigned VectorWidth = VecType->getNumElements();
1413 // Shorten the way if the mask is a vector of constants.
1414 bool IsConstMask = isa<ConstantVector>(Mask);
1417 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1418 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
1420 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
1421 "Ptr" + Twine(Idx));
1422 LoadInst *Load = Builder.CreateAlignedLoad(Ptr, AlignVal,
1423 "Load" + Twine(Idx));
1424 VResult = Builder.CreateInsertElement(VResult, Load,
1425 Builder.getInt32(Idx),
1426 "Res" + Twine(Idx));
1428 Value *NewI = Builder.CreateSelect(Mask, VResult, Src0);
1429 CI->replaceAllUsesWith(NewI);
1430 CI->eraseFromParent();
1434 PHINode *Phi = nullptr;
1435 Value *PrevPhi = UndefVal;
1437 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1439 // Fill the "else" block, created in the previous iteration
1441 // %Mask1 = extractelement <16 x i1> %Mask, i32 1
1442 // %ToLoad1 = icmp eq i1 %Mask1, true
1443 // br i1 %ToLoad1, label %cond.load, label %else
1446 Phi = Builder.CreatePHI(VecType, 2, "res.phi.else");
1447 Phi->addIncoming(VResult, CondBlock);
1448 Phi->addIncoming(PrevPhi, PrevIfBlock);
1453 Value *Predicate = Builder.CreateExtractElement(Mask,
1454 Builder.getInt32(Idx),
1455 "Mask" + Twine(Idx));
1456 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1457 ConstantInt::get(Predicate->getType(), 1),
1458 "ToLoad" + Twine(Idx));
1460 // Create "cond" block
1462 // %EltAddr = getelementptr i32* %1, i32 0
1463 // %Elt = load i32* %EltAddr
1464 // VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx
1466 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.load");
1467 Builder.SetInsertPoint(InsertPt);
1469 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
1470 "Ptr" + Twine(Idx));
1471 LoadInst *Load = Builder.CreateAlignedLoad(Ptr, AlignVal,
1472 "Load" + Twine(Idx));
1473 VResult = Builder.CreateInsertElement(VResult, Load, Builder.getInt32(Idx),
1474 "Res" + Twine(Idx));
1476 // Create "else" block, fill it in the next iteration
1477 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1478 Builder.SetInsertPoint(InsertPt);
1479 Instruction *OldBr = IfBlock->getTerminator();
1480 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1481 OldBr->eraseFromParent();
1482 PrevIfBlock = IfBlock;
1483 IfBlock = NewIfBlock;
1486 Phi = Builder.CreatePHI(VecType, 2, "res.phi.select");
1487 Phi->addIncoming(VResult, CondBlock);
1488 Phi->addIncoming(PrevPhi, PrevIfBlock);
1489 Value *NewI = Builder.CreateSelect(Mask, Phi, Src0);
1490 CI->replaceAllUsesWith(NewI);
1491 CI->eraseFromParent();
1494 // Translate a masked scatter intrinsic, like
1495 // void @llvm.masked.scatter.v16i32(<16 x i32> %Src, <16 x i32*>* %Ptrs, i32 4,
1497 // to a chain of basic blocks, that stores element one-by-one if
1498 // the appropriate mask bit is set.
1500 // % Ptrs = getelementptr i32, i32* %ptr, <16 x i64> %ind
1501 // % Mask0 = extractelement <16 x i1> % Mask, i32 0
1502 // % ToStore0 = icmp eq i1 % Mask0, true
1503 // br i1 %ToStore0, label %cond.store, label %else
1506 // % Elt0 = extractelement <16 x i32> %Src, i32 0
1507 // % Ptr0 = extractelement <16 x i32*> %Ptrs, i32 0
1508 // store i32 %Elt0, i32* % Ptr0, align 4
1512 // % Mask1 = extractelement <16 x i1> % Mask, i32 1
1513 // % ToStore1 = icmp eq i1 % Mask1, true
1514 // br i1 % ToStore1, label %cond.store1, label %else2
1517 // % Elt1 = extractelement <16 x i32> %Src, i32 1
1518 // % Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1
1519 // store i32 % Elt1, i32* % Ptr1, align 4
1522 static void ScalarizeMaskedScatter(CallInst *CI) {
1523 Value *Src = CI->getArgOperand(0);
1524 Value *Ptrs = CI->getArgOperand(1);
1525 Value *Alignment = CI->getArgOperand(2);
1526 Value *Mask = CI->getArgOperand(3);
1528 assert(isa<VectorType>(Src->getType()) &&
1529 "Unexpected data type in masked scatter intrinsic");
1530 assert(isa<VectorType>(Ptrs->getType()) &&
1531 isa<PointerType>(Ptrs->getType()->getVectorElementType()) &&
1532 "Vector of pointers is expected in masked scatter intrinsic");
1534 IRBuilder<> Builder(CI->getContext());
1535 Instruction *InsertPt = CI;
1536 BasicBlock *IfBlock = CI->getParent();
1537 Builder.SetInsertPoint(InsertPt);
1538 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1540 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
1541 unsigned VectorWidth = Src->getType()->getVectorNumElements();
1543 // Shorten the way if the mask is a vector of constants.
1544 bool IsConstMask = isa<ConstantVector>(Mask);
1547 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1548 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
1550 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx),
1551 "Elt" + Twine(Idx));
1552 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
1553 "Ptr" + Twine(Idx));
1554 Builder.CreateAlignedStore(OneElt, Ptr, AlignVal);
1556 CI->eraseFromParent();
1559 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1560 // Fill the "else" block, created in the previous iteration
1562 // % Mask1 = extractelement <16 x i1> % Mask, i32 Idx
1563 // % ToStore = icmp eq i1 % Mask1, true
1564 // br i1 % ToStore, label %cond.store, label %else
1566 Value *Predicate = Builder.CreateExtractElement(Mask,
1567 Builder.getInt32(Idx),
1568 "Mask" + Twine(Idx));
1570 Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1571 ConstantInt::get(Predicate->getType(), 1),
1572 "ToStore" + Twine(Idx));
1574 // Create "cond" block
1576 // % Elt1 = extractelement <16 x i32> %Src, i32 1
1577 // % Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1
1578 // %store i32 % Elt1, i32* % Ptr1
1580 BasicBlock *CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1581 Builder.SetInsertPoint(InsertPt);
1583 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx),
1584 "Elt" + Twine(Idx));
1585 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
1586 "Ptr" + Twine(Idx));
1587 Builder.CreateAlignedStore(OneElt, Ptr, AlignVal);
1589 // Create "else" block, fill it in the next iteration
1590 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1591 Builder.SetInsertPoint(InsertPt);
1592 Instruction *OldBr = IfBlock->getTerminator();
1593 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1594 OldBr->eraseFromParent();
1595 IfBlock = NewIfBlock;
1597 CI->eraseFromParent();
1600 bool CodeGenPrepare::optimizeCallInst(CallInst *CI, bool& ModifiedDT) {
1601 BasicBlock *BB = CI->getParent();
1603 // Lower inline assembly if we can.
1604 // If we found an inline asm expession, and if the target knows how to
1605 // lower it to normal LLVM code, do so now.
1606 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
1607 if (TLI->ExpandInlineAsm(CI)) {
1608 // Avoid invalidating the iterator.
1609 CurInstIterator = BB->begin();
1610 // Avoid processing instructions out of order, which could cause
1611 // reuse before a value is defined.
1615 // Sink address computing for memory operands into the block.
1616 if (optimizeInlineAsmInst(CI))
1620 // Align the pointer arguments to this call if the target thinks it's a good
1622 unsigned MinSize, PrefAlign;
1623 if (TLI && TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) {
1624 for (auto &Arg : CI->arg_operands()) {
1625 // We want to align both objects whose address is used directly and
1626 // objects whose address is used in casts and GEPs, though it only makes
1627 // sense for GEPs if the offset is a multiple of the desired alignment and
1628 // if size - offset meets the size threshold.
1629 if (!Arg->getType()->isPointerTy())
1631 APInt Offset(DL->getPointerSizeInBits(
1632 cast<PointerType>(Arg->getType())->getAddressSpace()),
1634 Value *Val = Arg->stripAndAccumulateInBoundsConstantOffsets(*DL, Offset);
1635 uint64_t Offset2 = Offset.getLimitedValue();
1636 if ((Offset2 & (PrefAlign-1)) != 0)
1639 if ((AI = dyn_cast<AllocaInst>(Val)) && AI->getAlignment() < PrefAlign &&
1640 DL->getTypeAllocSize(AI->getAllocatedType()) >= MinSize + Offset2)
1641 AI->setAlignment(PrefAlign);
1642 // Global variables can only be aligned if they are defined in this
1643 // object (i.e. they are uniquely initialized in this object), and
1644 // over-aligning global variables that have an explicit section is
1647 if ((GV = dyn_cast<GlobalVariable>(Val)) && GV->hasUniqueInitializer() &&
1648 !GV->hasSection() && GV->getAlignment() < PrefAlign &&
1649 DL->getTypeAllocSize(GV->getType()->getElementType()) >=
1651 GV->setAlignment(PrefAlign);
1653 // If this is a memcpy (or similar) then we may be able to improve the
1655 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(CI)) {
1656 unsigned Align = getKnownAlignment(MI->getDest(), *DL);
1657 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI))
1658 Align = std::min(Align, getKnownAlignment(MTI->getSource(), *DL));
1659 if (Align > MI->getAlignment())
1660 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), Align));
1664 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
1666 switch (II->getIntrinsicID()) {
1668 case Intrinsic::objectsize: {
1669 // Lower all uses of llvm.objectsize.*
1670 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
1671 Type *ReturnTy = CI->getType();
1672 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
1674 // Substituting this can cause recursive simplifications, which can
1675 // invalidate our iterator. Use a WeakVH to hold onto it in case this
1677 WeakVH IterHandle(&*CurInstIterator);
1679 replaceAndRecursivelySimplify(CI, RetVal,
1682 // If the iterator instruction was recursively deleted, start over at the
1683 // start of the block.
1684 if (IterHandle != CurInstIterator.getNodePtrUnchecked()) {
1685 CurInstIterator = BB->begin();
1690 case Intrinsic::masked_load: {
1691 // Scalarize unsupported vector masked load
1692 if (!TTI->isLegalMaskedLoad(CI->getType())) {
1693 ScalarizeMaskedLoad(CI);
1699 case Intrinsic::masked_store: {
1700 if (!TTI->isLegalMaskedStore(CI->getArgOperand(0)->getType())) {
1701 ScalarizeMaskedStore(CI);
1707 case Intrinsic::masked_gather: {
1708 if (!TTI->isLegalMaskedGather(CI->getType())) {
1709 ScalarizeMaskedGather(CI);
1715 case Intrinsic::masked_scatter: {
1716 if (!TTI->isLegalMaskedScatter(CI->getArgOperand(0)->getType())) {
1717 ScalarizeMaskedScatter(CI);
1723 case Intrinsic::aarch64_stlxr:
1724 case Intrinsic::aarch64_stxr: {
1725 ZExtInst *ExtVal = dyn_cast<ZExtInst>(CI->getArgOperand(0));
1726 if (!ExtVal || !ExtVal->hasOneUse() ||
1727 ExtVal->getParent() == CI->getParent())
1729 // Sink a zext feeding stlxr/stxr before it, so it can be folded into it.
1730 ExtVal->moveBefore(CI);
1731 // Mark this instruction as "inserted by CGP", so that other
1732 // optimizations don't touch it.
1733 InsertedInsts.insert(ExtVal);
1736 case Intrinsic::invariant_group_barrier:
1737 II->replaceAllUsesWith(II->getArgOperand(0));
1738 II->eraseFromParent();
1743 // Unknown address space.
1744 // TODO: Target hook to pick which address space the intrinsic cares
1746 unsigned AddrSpace = ~0u;
1747 SmallVector<Value*, 2> PtrOps;
1749 if (TLI->GetAddrModeArguments(II, PtrOps, AccessTy, AddrSpace))
1750 while (!PtrOps.empty())
1751 if (optimizeMemoryInst(II, PtrOps.pop_back_val(), AccessTy, AddrSpace))
1756 // From here on out we're working with named functions.
1757 if (!CI->getCalledFunction()) return false;
1759 // Lower all default uses of _chk calls. This is very similar
1760 // to what InstCombineCalls does, but here we are only lowering calls
1761 // to fortified library functions (e.g. __memcpy_chk) that have the default
1762 // "don't know" as the objectsize. Anything else should be left alone.
1763 FortifiedLibCallSimplifier Simplifier(TLInfo, true);
1764 if (Value *V = Simplifier.optimizeCall(CI)) {
1765 CI->replaceAllUsesWith(V);
1766 CI->eraseFromParent();
1772 /// Look for opportunities to duplicate return instructions to the predecessor
1773 /// to enable tail call optimizations. The case it is currently looking for is:
1776 /// %tmp0 = tail call i32 @f0()
1777 /// br label %return
1779 /// %tmp1 = tail call i32 @f1()
1780 /// br label %return
1782 /// %tmp2 = tail call i32 @f2()
1783 /// br label %return
1785 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
1793 /// %tmp0 = tail call i32 @f0()
1796 /// %tmp1 = tail call i32 @f1()
1799 /// %tmp2 = tail call i32 @f2()
1802 bool CodeGenPrepare::dupRetToEnableTailCallOpts(BasicBlock *BB) {
1806 ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator());
1810 PHINode *PN = nullptr;
1811 BitCastInst *BCI = nullptr;
1812 Value *V = RI->getReturnValue();
1814 BCI = dyn_cast<BitCastInst>(V);
1816 V = BCI->getOperand(0);
1818 PN = dyn_cast<PHINode>(V);
1823 if (PN && PN->getParent() != BB)
1826 // It's not safe to eliminate the sign / zero extension of the return value.
1827 // See llvm::isInTailCallPosition().
1828 const Function *F = BB->getParent();
1829 AttributeSet CallerAttrs = F->getAttributes();
1830 if (CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt) ||
1831 CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
1834 // Make sure there are no instructions between the PHI and return, or that the
1835 // return is the first instruction in the block.
1837 BasicBlock::iterator BI = BB->begin();
1838 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
1840 // Also skip over the bitcast.
1845 BasicBlock::iterator BI = BB->begin();
1846 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
1851 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
1853 SmallVector<CallInst*, 4> TailCalls;
1855 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
1856 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
1857 // Make sure the phi value is indeed produced by the tail call.
1858 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
1859 TLI->mayBeEmittedAsTailCall(CI))
1860 TailCalls.push_back(CI);
1863 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
1864 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
1865 if (!VisitedBBs.insert(*PI).second)
1868 BasicBlock::InstListType &InstList = (*PI)->getInstList();
1869 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
1870 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
1871 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
1875 CallInst *CI = dyn_cast<CallInst>(&*RI);
1876 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
1877 TailCalls.push_back(CI);
1881 bool Changed = false;
1882 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
1883 CallInst *CI = TailCalls[i];
1886 // Conservatively require the attributes of the call to match those of the
1887 // return. Ignore noalias because it doesn't affect the call sequence.
1888 AttributeSet CalleeAttrs = CS.getAttributes();
1889 if (AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1890 removeAttribute(Attribute::NoAlias) !=
1891 AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1892 removeAttribute(Attribute::NoAlias))
1895 // Make sure the call instruction is followed by an unconditional branch to
1896 // the return block.
1897 BasicBlock *CallBB = CI->getParent();
1898 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
1899 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
1902 // Duplicate the return into CallBB.
1903 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
1904 ModifiedDT = Changed = true;
1908 // If we eliminated all predecessors of the block, delete the block now.
1909 if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
1910 BB->eraseFromParent();
1915 //===----------------------------------------------------------------------===//
1916 // Memory Optimization
1917 //===----------------------------------------------------------------------===//
1921 /// This is an extended version of TargetLowering::AddrMode
1922 /// which holds actual Value*'s for register values.
1923 struct ExtAddrMode : public TargetLowering::AddrMode {
1926 ExtAddrMode() : BaseReg(nullptr), ScaledReg(nullptr) {}
1927 void print(raw_ostream &OS) const;
1930 bool operator==(const ExtAddrMode& O) const {
1931 return (BaseReg == O.BaseReg) && (ScaledReg == O.ScaledReg) &&
1932 (BaseGV == O.BaseGV) && (BaseOffs == O.BaseOffs) &&
1933 (HasBaseReg == O.HasBaseReg) && (Scale == O.Scale);
1938 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
1944 void ExtAddrMode::print(raw_ostream &OS) const {
1945 bool NeedPlus = false;
1948 OS << (NeedPlus ? " + " : "")
1950 BaseGV->printAsOperand(OS, /*PrintType=*/false);
1955 OS << (NeedPlus ? " + " : "")
1961 OS << (NeedPlus ? " + " : "")
1963 BaseReg->printAsOperand(OS, /*PrintType=*/false);
1967 OS << (NeedPlus ? " + " : "")
1969 ScaledReg->printAsOperand(OS, /*PrintType=*/false);
1975 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1976 void ExtAddrMode::dump() const {
1982 /// \brief This class provides transaction based operation on the IR.
1983 /// Every change made through this class is recorded in the internal state and
1984 /// can be undone (rollback) until commit is called.
1985 class TypePromotionTransaction {
1987 /// \brief This represents the common interface of the individual transaction.
1988 /// Each class implements the logic for doing one specific modification on
1989 /// the IR via the TypePromotionTransaction.
1990 class TypePromotionAction {
1992 /// The Instruction modified.
1996 /// \brief Constructor of the action.
1997 /// The constructor performs the related action on the IR.
1998 TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
2000 virtual ~TypePromotionAction() {}
2002 /// \brief Undo the modification done by this action.
2003 /// When this method is called, the IR must be in the same state as it was
2004 /// before this action was applied.
2005 /// \pre Undoing the action works if and only if the IR is in the exact same
2006 /// state as it was directly after this action was applied.
2007 virtual void undo() = 0;
2009 /// \brief Advocate every change made by this action.
2010 /// When the results on the IR of the action are to be kept, it is important
2011 /// to call this function, otherwise hidden information may be kept forever.
2012 virtual void commit() {
2013 // Nothing to be done, this action is not doing anything.
2017 /// \brief Utility to remember the position of an instruction.
2018 class InsertionHandler {
2019 /// Position of an instruction.
2020 /// Either an instruction:
2021 /// - Is the first in a basic block: BB is used.
2022 /// - Has a previous instructon: PrevInst is used.
2024 Instruction *PrevInst;
2027 /// Remember whether or not the instruction had a previous instruction.
2028 bool HasPrevInstruction;
2031 /// \brief Record the position of \p Inst.
2032 InsertionHandler(Instruction *Inst) {
2033 BasicBlock::iterator It = Inst->getIterator();
2034 HasPrevInstruction = (It != (Inst->getParent()->begin()));
2035 if (HasPrevInstruction)
2036 Point.PrevInst = &*--It;
2038 Point.BB = Inst->getParent();
2041 /// \brief Insert \p Inst at the recorded position.
2042 void insert(Instruction *Inst) {
2043 if (HasPrevInstruction) {
2044 if (Inst->getParent())
2045 Inst->removeFromParent();
2046 Inst->insertAfter(Point.PrevInst);
2048 Instruction *Position = &*Point.BB->getFirstInsertionPt();
2049 if (Inst->getParent())
2050 Inst->moveBefore(Position);
2052 Inst->insertBefore(Position);
2057 /// \brief Move an instruction before another.
2058 class InstructionMoveBefore : public TypePromotionAction {
2059 /// Original position of the instruction.
2060 InsertionHandler Position;
2063 /// \brief Move \p Inst before \p Before.
2064 InstructionMoveBefore(Instruction *Inst, Instruction *Before)
2065 : TypePromotionAction(Inst), Position(Inst) {
2066 DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before << "\n");
2067 Inst->moveBefore(Before);
2070 /// \brief Move the instruction back to its original position.
2071 void undo() override {
2072 DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
2073 Position.insert(Inst);
2077 /// \brief Set the operand of an instruction with a new value.
2078 class OperandSetter : public TypePromotionAction {
2079 /// Original operand of the instruction.
2081 /// Index of the modified instruction.
2085 /// \brief Set \p Idx operand of \p Inst with \p NewVal.
2086 OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
2087 : TypePromotionAction(Inst), Idx(Idx) {
2088 DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
2089 << "for:" << *Inst << "\n"
2090 << "with:" << *NewVal << "\n");
2091 Origin = Inst->getOperand(Idx);
2092 Inst->setOperand(Idx, NewVal);
2095 /// \brief Restore the original value of the instruction.
2096 void undo() override {
2097 DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
2098 << "for: " << *Inst << "\n"
2099 << "with: " << *Origin << "\n");
2100 Inst->setOperand(Idx, Origin);
2104 /// \brief Hide the operands of an instruction.
2105 /// Do as if this instruction was not using any of its operands.
2106 class OperandsHider : public TypePromotionAction {
2107 /// The list of original operands.
2108 SmallVector<Value *, 4> OriginalValues;
2111 /// \brief Remove \p Inst from the uses of the operands of \p Inst.
2112 OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
2113 DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
2114 unsigned NumOpnds = Inst->getNumOperands();
2115 OriginalValues.reserve(NumOpnds);
2116 for (unsigned It = 0; It < NumOpnds; ++It) {
2117 // Save the current operand.
2118 Value *Val = Inst->getOperand(It);
2119 OriginalValues.push_back(Val);
2121 // We could use OperandSetter here, but that would imply an overhead
2122 // that we are not willing to pay.
2123 Inst->setOperand(It, UndefValue::get(Val->getType()));
2127 /// \brief Restore the original list of uses.
2128 void undo() override {
2129 DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
2130 for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
2131 Inst->setOperand(It, OriginalValues[It]);
2135 /// \brief Build a truncate instruction.
2136 class TruncBuilder : public TypePromotionAction {
2139 /// \brief Build a truncate instruction of \p Opnd producing a \p Ty
2141 /// trunc Opnd to Ty.
2142 TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
2143 IRBuilder<> Builder(Opnd);
2144 Val = Builder.CreateTrunc(Opnd, Ty, "promoted");
2145 DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n");
2148 /// \brief Get the built value.
2149 Value *getBuiltValue() { return Val; }
2151 /// \brief Remove the built instruction.
2152 void undo() override {
2153 DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n");
2154 if (Instruction *IVal = dyn_cast<Instruction>(Val))
2155 IVal->eraseFromParent();
2159 /// \brief Build a sign extension instruction.
2160 class SExtBuilder : public TypePromotionAction {
2163 /// \brief Build a sign extension instruction of \p Opnd producing a \p Ty
2165 /// sext Opnd to Ty.
2166 SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
2167 : TypePromotionAction(InsertPt) {
2168 IRBuilder<> Builder(InsertPt);
2169 Val = Builder.CreateSExt(Opnd, Ty, "promoted");
2170 DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n");
2173 /// \brief Get the built value.
2174 Value *getBuiltValue() { return Val; }
2176 /// \brief Remove the built instruction.
2177 void undo() override {
2178 DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n");
2179 if (Instruction *IVal = dyn_cast<Instruction>(Val))
2180 IVal->eraseFromParent();
2184 /// \brief Build a zero extension instruction.
2185 class ZExtBuilder : public TypePromotionAction {
2188 /// \brief Build a zero extension instruction of \p Opnd producing a \p Ty
2190 /// zext Opnd to Ty.
2191 ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
2192 : TypePromotionAction(InsertPt) {
2193 IRBuilder<> Builder(InsertPt);
2194 Val = Builder.CreateZExt(Opnd, Ty, "promoted");
2195 DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n");
2198 /// \brief Get the built value.
2199 Value *getBuiltValue() { return Val; }
2201 /// \brief Remove the built instruction.
2202 void undo() override {
2203 DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n");
2204 if (Instruction *IVal = dyn_cast<Instruction>(Val))
2205 IVal->eraseFromParent();
2209 /// \brief Mutate an instruction to another type.
2210 class TypeMutator : public TypePromotionAction {
2211 /// Record the original type.
2215 /// \brief Mutate the type of \p Inst into \p NewTy.
2216 TypeMutator(Instruction *Inst, Type *NewTy)
2217 : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
2218 DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
2220 Inst->mutateType(NewTy);
2223 /// \brief Mutate the instruction back to its original type.
2224 void undo() override {
2225 DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
2227 Inst->mutateType(OrigTy);
2231 /// \brief Replace the uses of an instruction by another instruction.
2232 class UsesReplacer : public TypePromotionAction {
2233 /// Helper structure to keep track of the replaced uses.
2234 struct InstructionAndIdx {
2235 /// The instruction using the instruction.
2237 /// The index where this instruction is used for Inst.
2239 InstructionAndIdx(Instruction *Inst, unsigned Idx)
2240 : Inst(Inst), Idx(Idx) {}
2243 /// Keep track of the original uses (pair Instruction, Index).
2244 SmallVector<InstructionAndIdx, 4> OriginalUses;
2245 typedef SmallVectorImpl<InstructionAndIdx>::iterator use_iterator;
2248 /// \brief Replace all the use of \p Inst by \p New.
2249 UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
2250 DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
2252 // Record the original uses.
2253 for (Use &U : Inst->uses()) {
2254 Instruction *UserI = cast<Instruction>(U.getUser());
2255 OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
2257 // Now, we can replace the uses.
2258 Inst->replaceAllUsesWith(New);
2261 /// \brief Reassign the original uses of Inst to Inst.
2262 void undo() override {
2263 DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
2264 for (use_iterator UseIt = OriginalUses.begin(),
2265 EndIt = OriginalUses.end();
2266 UseIt != EndIt; ++UseIt) {
2267 UseIt->Inst->setOperand(UseIt->Idx, Inst);
2272 /// \brief Remove an instruction from the IR.
2273 class InstructionRemover : public TypePromotionAction {
2274 /// Original position of the instruction.
2275 InsertionHandler Inserter;
2276 /// Helper structure to hide all the link to the instruction. In other
2277 /// words, this helps to do as if the instruction was removed.
2278 OperandsHider Hider;
2279 /// Keep track of the uses replaced, if any.
2280 UsesReplacer *Replacer;
2283 /// \brief Remove all reference of \p Inst and optinally replace all its
2285 /// \pre If !Inst->use_empty(), then New != nullptr
2286 InstructionRemover(Instruction *Inst, Value *New = nullptr)
2287 : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
2290 Replacer = new UsesReplacer(Inst, New);
2291 DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
2292 Inst->removeFromParent();
2295 ~InstructionRemover() override { delete Replacer; }
2297 /// \brief Really remove the instruction.
2298 void commit() override { delete Inst; }
2300 /// \brief Resurrect the instruction and reassign it to the proper uses if
2301 /// new value was provided when build this action.
2302 void undo() override {
2303 DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
2304 Inserter.insert(Inst);
2312 /// Restoration point.
2313 /// The restoration point is a pointer to an action instead of an iterator
2314 /// because the iterator may be invalidated but not the pointer.
2315 typedef const TypePromotionAction *ConstRestorationPt;
2316 /// Advocate every changes made in that transaction.
2318 /// Undo all the changes made after the given point.
2319 void rollback(ConstRestorationPt Point);
2320 /// Get the current restoration point.
2321 ConstRestorationPt getRestorationPoint() const;
2323 /// \name API for IR modification with state keeping to support rollback.
2325 /// Same as Instruction::setOperand.
2326 void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
2327 /// Same as Instruction::eraseFromParent.
2328 void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
2329 /// Same as Value::replaceAllUsesWith.
2330 void replaceAllUsesWith(Instruction *Inst, Value *New);
2331 /// Same as Value::mutateType.
2332 void mutateType(Instruction *Inst, Type *NewTy);
2333 /// Same as IRBuilder::createTrunc.
2334 Value *createTrunc(Instruction *Opnd, Type *Ty);
2335 /// Same as IRBuilder::createSExt.
2336 Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
2337 /// Same as IRBuilder::createZExt.
2338 Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
2339 /// Same as Instruction::moveBefore.
2340 void moveBefore(Instruction *Inst, Instruction *Before);
2344 /// The ordered list of actions made so far.
2345 SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions;
2346 typedef SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator CommitPt;
2349 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
2352 make_unique<TypePromotionTransaction::OperandSetter>(Inst, Idx, NewVal));
2355 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
2358 make_unique<TypePromotionTransaction::InstructionRemover>(Inst, NewVal));
2361 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
2363 Actions.push_back(make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
2366 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
2367 Actions.push_back(make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
2370 Value *TypePromotionTransaction::createTrunc(Instruction *Opnd,
2372 std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
2373 Value *Val = Ptr->getBuiltValue();
2374 Actions.push_back(std::move(Ptr));
2378 Value *TypePromotionTransaction::createSExt(Instruction *Inst,
2379 Value *Opnd, Type *Ty) {
2380 std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
2381 Value *Val = Ptr->getBuiltValue();
2382 Actions.push_back(std::move(Ptr));
2386 Value *TypePromotionTransaction::createZExt(Instruction *Inst,
2387 Value *Opnd, Type *Ty) {
2388 std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
2389 Value *Val = Ptr->getBuiltValue();
2390 Actions.push_back(std::move(Ptr));
2394 void TypePromotionTransaction::moveBefore(Instruction *Inst,
2395 Instruction *Before) {
2397 make_unique<TypePromotionTransaction::InstructionMoveBefore>(Inst, Before));
2400 TypePromotionTransaction::ConstRestorationPt
2401 TypePromotionTransaction::getRestorationPoint() const {
2402 return !Actions.empty() ? Actions.back().get() : nullptr;
2405 void TypePromotionTransaction::commit() {
2406 for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
2412 void TypePromotionTransaction::rollback(
2413 TypePromotionTransaction::ConstRestorationPt Point) {
2414 while (!Actions.empty() && Point != Actions.back().get()) {
2415 std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
2420 /// \brief A helper class for matching addressing modes.
2422 /// This encapsulates the logic for matching the target-legal addressing modes.
2423 class AddressingModeMatcher {
2424 SmallVectorImpl<Instruction*> &AddrModeInsts;
2425 const TargetMachine &TM;
2426 const TargetLowering &TLI;
2427 const DataLayout &DL;
2429 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
2430 /// the memory instruction that we're computing this address for.
2433 Instruction *MemoryInst;
2435 /// This is the addressing mode that we're building up. This is
2436 /// part of the return value of this addressing mode matching stuff.
2437 ExtAddrMode &AddrMode;
2439 /// The instructions inserted by other CodeGenPrepare optimizations.
2440 const SetOfInstrs &InsertedInsts;
2441 /// A map from the instructions to their type before promotion.
2442 InstrToOrigTy &PromotedInsts;
2443 /// The ongoing transaction where every action should be registered.
2444 TypePromotionTransaction &TPT;
2446 /// This is set to true when we should not do profitability checks.
2447 /// When true, IsProfitableToFoldIntoAddressingMode always returns true.
2448 bool IgnoreProfitability;
2450 AddressingModeMatcher(SmallVectorImpl<Instruction *> &AMI,
2451 const TargetMachine &TM, Type *AT, unsigned AS,
2452 Instruction *MI, ExtAddrMode &AM,
2453 const SetOfInstrs &InsertedInsts,
2454 InstrToOrigTy &PromotedInsts,
2455 TypePromotionTransaction &TPT)
2456 : AddrModeInsts(AMI), TM(TM),
2457 TLI(*TM.getSubtargetImpl(*MI->getParent()->getParent())
2458 ->getTargetLowering()),
2459 DL(MI->getModule()->getDataLayout()), AccessTy(AT), AddrSpace(AS),
2460 MemoryInst(MI), AddrMode(AM), InsertedInsts(InsertedInsts),
2461 PromotedInsts(PromotedInsts), TPT(TPT) {
2462 IgnoreProfitability = false;
2466 /// Find the maximal addressing mode that a load/store of V can fold,
2467 /// give an access type of AccessTy. This returns a list of involved
2468 /// instructions in AddrModeInsts.
2469 /// \p InsertedInsts The instructions inserted by other CodeGenPrepare
2471 /// \p PromotedInsts maps the instructions to their type before promotion.
2472 /// \p The ongoing transaction where every action should be registered.
2473 static ExtAddrMode Match(Value *V, Type *AccessTy, unsigned AS,
2474 Instruction *MemoryInst,
2475 SmallVectorImpl<Instruction*> &AddrModeInsts,
2476 const TargetMachine &TM,
2477 const SetOfInstrs &InsertedInsts,
2478 InstrToOrigTy &PromotedInsts,
2479 TypePromotionTransaction &TPT) {
2482 bool Success = AddressingModeMatcher(AddrModeInsts, TM, AccessTy, AS,
2483 MemoryInst, Result, InsertedInsts,
2484 PromotedInsts, TPT).matchAddr(V, 0);
2485 (void)Success; assert(Success && "Couldn't select *anything*?");
2489 bool matchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
2490 bool matchAddr(Value *V, unsigned Depth);
2491 bool matchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth,
2492 bool *MovedAway = nullptr);
2493 bool isProfitableToFoldIntoAddressingMode(Instruction *I,
2494 ExtAddrMode &AMBefore,
2495 ExtAddrMode &AMAfter);
2496 bool valueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
2497 bool isPromotionProfitable(unsigned NewCost, unsigned OldCost,
2498 Value *PromotedOperand) const;
2501 /// Try adding ScaleReg*Scale to the current addressing mode.
2502 /// Return true and update AddrMode if this addr mode is legal for the target,
2504 bool AddressingModeMatcher::matchScaledValue(Value *ScaleReg, int64_t Scale,
2506 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
2507 // mode. Just process that directly.
2509 return matchAddr(ScaleReg, Depth);
2511 // If the scale is 0, it takes nothing to add this.
2515 // If we already have a scale of this value, we can add to it, otherwise, we
2516 // need an available scale field.
2517 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
2520 ExtAddrMode TestAddrMode = AddrMode;
2522 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
2523 // [A+B + A*7] -> [B+A*8].
2524 TestAddrMode.Scale += Scale;
2525 TestAddrMode.ScaledReg = ScaleReg;
2527 // If the new address isn't legal, bail out.
2528 if (!TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace))
2531 // It was legal, so commit it.
2532 AddrMode = TestAddrMode;
2534 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
2535 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
2536 // X*Scale + C*Scale to addr mode.
2537 ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
2538 if (isa<Instruction>(ScaleReg) && // not a constant expr.
2539 match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
2540 TestAddrMode.ScaledReg = AddLHS;
2541 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
2543 // If this addressing mode is legal, commit it and remember that we folded
2544 // this instruction.
2545 if (TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace)) {
2546 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
2547 AddrMode = TestAddrMode;
2552 // Otherwise, not (x+c)*scale, just return what we have.
2556 /// This is a little filter, which returns true if an addressing computation
2557 /// involving I might be folded into a load/store accessing it.
2558 /// This doesn't need to be perfect, but needs to accept at least
2559 /// the set of instructions that MatchOperationAddr can.
2560 static bool MightBeFoldableInst(Instruction *I) {
2561 switch (I->getOpcode()) {
2562 case Instruction::BitCast:
2563 case Instruction::AddrSpaceCast:
2564 // Don't touch identity bitcasts.
2565 if (I->getType() == I->getOperand(0)->getType())
2567 return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
2568 case Instruction::PtrToInt:
2569 // PtrToInt is always a noop, as we know that the int type is pointer sized.
2571 case Instruction::IntToPtr:
2572 // We know the input is intptr_t, so this is foldable.
2574 case Instruction::Add:
2576 case Instruction::Mul:
2577 case Instruction::Shl:
2578 // Can only handle X*C and X << C.
2579 return isa<ConstantInt>(I->getOperand(1));
2580 case Instruction::GetElementPtr:
2587 /// \brief Check whether or not \p Val is a legal instruction for \p TLI.
2588 /// \note \p Val is assumed to be the product of some type promotion.
2589 /// Therefore if \p Val has an undefined state in \p TLI, this is assumed
2590 /// to be legal, as the non-promoted value would have had the same state.
2591 static bool isPromotedInstructionLegal(const TargetLowering &TLI,
2592 const DataLayout &DL, Value *Val) {
2593 Instruction *PromotedInst = dyn_cast<Instruction>(Val);
2596 int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
2597 // If the ISDOpcode is undefined, it was undefined before the promotion.
2600 // Otherwise, check if the promoted instruction is legal or not.
2601 return TLI.isOperationLegalOrCustom(
2602 ISDOpcode, TLI.getValueType(DL, PromotedInst->getType()));
2605 /// \brief Hepler class to perform type promotion.
2606 class TypePromotionHelper {
2607 /// \brief Utility function to check whether or not a sign or zero extension
2608 /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
2609 /// either using the operands of \p Inst or promoting \p Inst.
2610 /// The type of the extension is defined by \p IsSExt.
2611 /// In other words, check if:
2612 /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
2613 /// #1 Promotion applies:
2614 /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
2615 /// #2 Operand reuses:
2616 /// ext opnd1 to ConsideredExtType.
2617 /// \p PromotedInsts maps the instructions to their type before promotion.
2618 static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
2619 const InstrToOrigTy &PromotedInsts, bool IsSExt);
2621 /// \brief Utility function to determine if \p OpIdx should be promoted when
2622 /// promoting \p Inst.
2623 static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
2624 return !(isa<SelectInst>(Inst) && OpIdx == 0);
2627 /// \brief Utility function to promote the operand of \p Ext when this
2628 /// operand is a promotable trunc or sext or zext.
2629 /// \p PromotedInsts maps the instructions to their type before promotion.
2630 /// \p CreatedInstsCost[out] contains the cost of all instructions
2631 /// created to promote the operand of Ext.
2632 /// Newly added extensions are inserted in \p Exts.
2633 /// Newly added truncates are inserted in \p Truncs.
2634 /// Should never be called directly.
2635 /// \return The promoted value which is used instead of Ext.
2636 static Value *promoteOperandForTruncAndAnyExt(
2637 Instruction *Ext, TypePromotionTransaction &TPT,
2638 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2639 SmallVectorImpl<Instruction *> *Exts,
2640 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI);
2642 /// \brief Utility function to promote the operand of \p Ext when this
2643 /// operand is promotable and is not a supported trunc or sext.
2644 /// \p PromotedInsts maps the instructions to their type before promotion.
2645 /// \p CreatedInstsCost[out] contains the cost of all the instructions
2646 /// created to promote the operand of Ext.
2647 /// Newly added extensions are inserted in \p Exts.
2648 /// Newly added truncates are inserted in \p Truncs.
2649 /// Should never be called directly.
2650 /// \return The promoted value which is used instead of Ext.
2651 static Value *promoteOperandForOther(Instruction *Ext,
2652 TypePromotionTransaction &TPT,
2653 InstrToOrigTy &PromotedInsts,
2654 unsigned &CreatedInstsCost,
2655 SmallVectorImpl<Instruction *> *Exts,
2656 SmallVectorImpl<Instruction *> *Truncs,
2657 const TargetLowering &TLI, bool IsSExt);
2659 /// \see promoteOperandForOther.
2660 static Value *signExtendOperandForOther(
2661 Instruction *Ext, TypePromotionTransaction &TPT,
2662 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2663 SmallVectorImpl<Instruction *> *Exts,
2664 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2665 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
2666 Exts, Truncs, TLI, true);
2669 /// \see promoteOperandForOther.
2670 static Value *zeroExtendOperandForOther(
2671 Instruction *Ext, TypePromotionTransaction &TPT,
2672 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2673 SmallVectorImpl<Instruction *> *Exts,
2674 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2675 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
2676 Exts, Truncs, TLI, false);
2680 /// Type for the utility function that promotes the operand of Ext.
2681 typedef Value *(*Action)(Instruction *Ext, TypePromotionTransaction &TPT,
2682 InstrToOrigTy &PromotedInsts,
2683 unsigned &CreatedInstsCost,
2684 SmallVectorImpl<Instruction *> *Exts,
2685 SmallVectorImpl<Instruction *> *Truncs,
2686 const TargetLowering &TLI);
2687 /// \brief Given a sign/zero extend instruction \p Ext, return the approriate
2688 /// action to promote the operand of \p Ext instead of using Ext.
2689 /// \return NULL if no promotable action is possible with the current
2691 /// \p InsertedInsts keeps track of all the instructions inserted by the
2692 /// other CodeGenPrepare optimizations. This information is important
2693 /// because we do not want to promote these instructions as CodeGenPrepare
2694 /// will reinsert them later. Thus creating an infinite loop: create/remove.
2695 /// \p PromotedInsts maps the instructions to their type before promotion.
2696 static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedInsts,
2697 const TargetLowering &TLI,
2698 const InstrToOrigTy &PromotedInsts);
2701 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
2702 Type *ConsideredExtType,
2703 const InstrToOrigTy &PromotedInsts,
2705 // The promotion helper does not know how to deal with vector types yet.
2706 // To be able to fix that, we would need to fix the places where we
2707 // statically extend, e.g., constants and such.
2708 if (Inst->getType()->isVectorTy())
2711 // We can always get through zext.
2712 if (isa<ZExtInst>(Inst))
2715 // sext(sext) is ok too.
2716 if (IsSExt && isa<SExtInst>(Inst))
2719 // We can get through binary operator, if it is legal. In other words, the
2720 // binary operator must have a nuw or nsw flag.
2721 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
2722 if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
2723 ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
2724 (IsSExt && BinOp->hasNoSignedWrap())))
2727 // Check if we can do the following simplification.
2728 // ext(trunc(opnd)) --> ext(opnd)
2729 if (!isa<TruncInst>(Inst))
2732 Value *OpndVal = Inst->getOperand(0);
2733 // Check if we can use this operand in the extension.
2734 // If the type is larger than the result type of the extension, we cannot.
2735 if (!OpndVal->getType()->isIntegerTy() ||
2736 OpndVal->getType()->getIntegerBitWidth() >
2737 ConsideredExtType->getIntegerBitWidth())
2740 // If the operand of the truncate is not an instruction, we will not have
2741 // any information on the dropped bits.
2742 // (Actually we could for constant but it is not worth the extra logic).
2743 Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
2747 // Check if the source of the type is narrow enough.
2748 // I.e., check that trunc just drops extended bits of the same kind of
2750 // #1 get the type of the operand and check the kind of the extended bits.
2751 const Type *OpndType;
2752 InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
2753 if (It != PromotedInsts.end() && It->second.getInt() == IsSExt)
2754 OpndType = It->second.getPointer();
2755 else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd)))
2756 OpndType = Opnd->getOperand(0)->getType();
2760 // #2 check that the truncate just drops extended bits.
2761 return Inst->getType()->getIntegerBitWidth() >=
2762 OpndType->getIntegerBitWidth();
2765 TypePromotionHelper::Action TypePromotionHelper::getAction(
2766 Instruction *Ext, const SetOfInstrs &InsertedInsts,
2767 const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
2768 assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
2769 "Unexpected instruction type");
2770 Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0));
2771 Type *ExtTy = Ext->getType();
2772 bool IsSExt = isa<SExtInst>(Ext);
2773 // If the operand of the extension is not an instruction, we cannot
2775 // If it, check we can get through.
2776 if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt))
2779 // Do not promote if the operand has been added by codegenprepare.
2780 // Otherwise, it means we are undoing an optimization that is likely to be
2781 // redone, thus causing potential infinite loop.
2782 if (isa<TruncInst>(ExtOpnd) && InsertedInsts.count(ExtOpnd))
2785 // SExt or Trunc instructions.
2786 // Return the related handler.
2787 if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) ||
2788 isa<ZExtInst>(ExtOpnd))
2789 return promoteOperandForTruncAndAnyExt;
2791 // Regular instruction.
2792 // Abort early if we will have to insert non-free instructions.
2793 if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType()))
2795 return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
2798 Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
2799 llvm::Instruction *SExt, TypePromotionTransaction &TPT,
2800 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2801 SmallVectorImpl<Instruction *> *Exts,
2802 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2803 // By construction, the operand of SExt is an instruction. Otherwise we cannot
2804 // get through it and this method should not be called.
2805 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
2806 Value *ExtVal = SExt;
2807 bool HasMergedNonFreeExt = false;
2808 if (isa<ZExtInst>(SExtOpnd)) {
2809 // Replace s|zext(zext(opnd))
2811 HasMergedNonFreeExt = !TLI.isExtFree(SExtOpnd);
2813 TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
2814 TPT.replaceAllUsesWith(SExt, ZExt);
2815 TPT.eraseInstruction(SExt);
2818 // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
2820 TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
2822 CreatedInstsCost = 0;
2824 // Remove dead code.
2825 if (SExtOpnd->use_empty())
2826 TPT.eraseInstruction(SExtOpnd);
2828 // Check if the extension is still needed.
2829 Instruction *ExtInst = dyn_cast<Instruction>(ExtVal);
2830 if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) {
2833 Exts->push_back(ExtInst);
2834 CreatedInstsCost = !TLI.isExtFree(ExtInst) && !HasMergedNonFreeExt;
2839 // At this point we have: ext ty opnd to ty.
2840 // Reassign the uses of ExtInst to the opnd and remove ExtInst.
2841 Value *NextVal = ExtInst->getOperand(0);
2842 TPT.eraseInstruction(ExtInst, NextVal);
2846 Value *TypePromotionHelper::promoteOperandForOther(
2847 Instruction *Ext, TypePromotionTransaction &TPT,
2848 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2849 SmallVectorImpl<Instruction *> *Exts,
2850 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI,
2852 // By construction, the operand of Ext is an instruction. Otherwise we cannot
2853 // get through it and this method should not be called.
2854 Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0));
2855 CreatedInstsCost = 0;
2856 if (!ExtOpnd->hasOneUse()) {
2857 // ExtOpnd will be promoted.
2858 // All its uses, but Ext, will need to use a truncated value of the
2859 // promoted version.
2860 // Create the truncate now.
2861 Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType());
2862 if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) {
2863 ITrunc->removeFromParent();
2864 // Insert it just after the definition.
2865 ITrunc->insertAfter(ExtOpnd);
2867 Truncs->push_back(ITrunc);
2870 TPT.replaceAllUsesWith(ExtOpnd, Trunc);
2871 // Restore the operand of Ext (which has been replaced by the previous call
2872 // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
2873 TPT.setOperand(Ext, 0, ExtOpnd);
2876 // Get through the Instruction:
2877 // 1. Update its type.
2878 // 2. Replace the uses of Ext by Inst.
2879 // 3. Extend each operand that needs to be extended.
2881 // Remember the original type of the instruction before promotion.
2882 // This is useful to know that the high bits are sign extended bits.
2883 PromotedInsts.insert(std::pair<Instruction *, TypeIsSExt>(
2884 ExtOpnd, TypeIsSExt(ExtOpnd->getType(), IsSExt)));
2886 TPT.mutateType(ExtOpnd, Ext->getType());
2888 TPT.replaceAllUsesWith(Ext, ExtOpnd);
2890 Instruction *ExtForOpnd = Ext;
2892 DEBUG(dbgs() << "Propagate Ext to operands\n");
2893 for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
2895 DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n');
2896 if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() ||
2897 !shouldExtOperand(ExtOpnd, OpIdx)) {
2898 DEBUG(dbgs() << "No need to propagate\n");
2901 // Check if we can statically extend the operand.
2902 Value *Opnd = ExtOpnd->getOperand(OpIdx);
2903 if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
2904 DEBUG(dbgs() << "Statically extend\n");
2905 unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
2906 APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth)
2907 : Cst->getValue().zext(BitWidth);
2908 TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal));
2911 // UndefValue are typed, so we have to statically sign extend them.
2912 if (isa<UndefValue>(Opnd)) {
2913 DEBUG(dbgs() << "Statically extend\n");
2914 TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType()));
2918 // Otherwise we have to explicity sign extend the operand.
2919 // Check if Ext was reused to extend an operand.
2921 // If yes, create a new one.
2922 DEBUG(dbgs() << "More operands to ext\n");
2923 Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType())
2924 : TPT.createZExt(Ext, Opnd, Ext->getType());
2925 if (!isa<Instruction>(ValForExtOpnd)) {
2926 TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd);
2929 ExtForOpnd = cast<Instruction>(ValForExtOpnd);
2932 Exts->push_back(ExtForOpnd);
2933 TPT.setOperand(ExtForOpnd, 0, Opnd);
2935 // Move the sign extension before the insertion point.
2936 TPT.moveBefore(ExtForOpnd, ExtOpnd);
2937 TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd);
2938 CreatedInstsCost += !TLI.isExtFree(ExtForOpnd);
2939 // If more sext are required, new instructions will have to be created.
2940 ExtForOpnd = nullptr;
2942 if (ExtForOpnd == Ext) {
2943 DEBUG(dbgs() << "Extension is useless now\n");
2944 TPT.eraseInstruction(Ext);
2949 /// Check whether or not promoting an instruction to a wider type is profitable.
2950 /// \p NewCost gives the cost of extension instructions created by the
2952 /// \p OldCost gives the cost of extension instructions before the promotion
2953 /// plus the number of instructions that have been
2954 /// matched in the addressing mode the promotion.
2955 /// \p PromotedOperand is the value that has been promoted.
2956 /// \return True if the promotion is profitable, false otherwise.
2957 bool AddressingModeMatcher::isPromotionProfitable(
2958 unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const {
2959 DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCost << '\n');
2960 // The cost of the new extensions is greater than the cost of the
2961 // old extension plus what we folded.
2962 // This is not profitable.
2963 if (NewCost > OldCost)
2965 if (NewCost < OldCost)
2967 // The promotion is neutral but it may help folding the sign extension in
2968 // loads for instance.
2969 // Check that we did not create an illegal instruction.
2970 return isPromotedInstructionLegal(TLI, DL, PromotedOperand);
2973 /// Given an instruction or constant expr, see if we can fold the operation
2974 /// into the addressing mode. If so, update the addressing mode and return
2975 /// true, otherwise return false without modifying AddrMode.
2976 /// If \p MovedAway is not NULL, it contains the information of whether or
2977 /// not AddrInst has to be folded into the addressing mode on success.
2978 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
2979 /// because it has been moved away.
2980 /// Thus AddrInst must not be added in the matched instructions.
2981 /// This state can happen when AddrInst is a sext, since it may be moved away.
2982 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
2983 /// not be referenced anymore.
2984 bool AddressingModeMatcher::matchOperationAddr(User *AddrInst, unsigned Opcode,
2987 // Avoid exponential behavior on extremely deep expression trees.
2988 if (Depth >= 5) return false;
2990 // By default, all matched instructions stay in place.
2995 case Instruction::PtrToInt:
2996 // PtrToInt is always a noop, as we know that the int type is pointer sized.
2997 return matchAddr(AddrInst->getOperand(0), Depth);
2998 case Instruction::IntToPtr: {
2999 auto AS = AddrInst->getType()->getPointerAddressSpace();
3000 auto PtrTy = MVT::getIntegerVT(DL.getPointerSizeInBits(AS));
3001 // This inttoptr is a no-op if the integer type is pointer sized.
3002 if (TLI.getValueType(DL, AddrInst->getOperand(0)->getType()) == PtrTy)
3003 return matchAddr(AddrInst->getOperand(0), Depth);
3006 case Instruction::BitCast:
3007 // BitCast is always a noop, and we can handle it as long as it is
3008 // int->int or pointer->pointer (we don't want int<->fp or something).
3009 if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
3010 AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
3011 // Don't touch identity bitcasts. These were probably put here by LSR,
3012 // and we don't want to mess around with them. Assume it knows what it
3014 AddrInst->getOperand(0)->getType() != AddrInst->getType())
3015 return matchAddr(AddrInst->getOperand(0), Depth);
3017 case Instruction::AddrSpaceCast: {
3019 = AddrInst->getOperand(0)->getType()->getPointerAddressSpace();
3020 unsigned DestAS = AddrInst->getType()->getPointerAddressSpace();
3021 if (TLI.isNoopAddrSpaceCast(SrcAS, DestAS))
3022 return matchAddr(AddrInst->getOperand(0), Depth);
3025 case Instruction::Add: {
3026 // Check to see if we can merge in the RHS then the LHS. If so, we win.
3027 ExtAddrMode BackupAddrMode = AddrMode;
3028 unsigned OldSize = AddrModeInsts.size();
3029 // Start a transaction at this point.
3030 // The LHS may match but not the RHS.
3031 // Therefore, we need a higher level restoration point to undo partially
3032 // matched operation.
3033 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3034 TPT.getRestorationPoint();
3036 if (matchAddr(AddrInst->getOperand(1), Depth+1) &&
3037 matchAddr(AddrInst->getOperand(0), Depth+1))
3040 // Restore the old addr mode info.
3041 AddrMode = BackupAddrMode;
3042 AddrModeInsts.resize(OldSize);
3043 TPT.rollback(LastKnownGood);
3045 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
3046 if (matchAddr(AddrInst->getOperand(0), Depth+1) &&
3047 matchAddr(AddrInst->getOperand(1), Depth+1))
3050 // Otherwise we definitely can't merge the ADD in.
3051 AddrMode = BackupAddrMode;
3052 AddrModeInsts.resize(OldSize);
3053 TPT.rollback(LastKnownGood);
3056 //case Instruction::Or:
3057 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
3059 case Instruction::Mul:
3060 case Instruction::Shl: {
3061 // Can only handle X*C and X << C.
3062 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
3065 int64_t Scale = RHS->getSExtValue();
3066 if (Opcode == Instruction::Shl)
3067 Scale = 1LL << Scale;
3069 return matchScaledValue(AddrInst->getOperand(0), Scale, Depth);
3071 case Instruction::GetElementPtr: {
3072 // Scan the GEP. We check it if it contains constant offsets and at most
3073 // one variable offset.
3074 int VariableOperand = -1;
3075 unsigned VariableScale = 0;
3077 int64_t ConstantOffset = 0;
3078 gep_type_iterator GTI = gep_type_begin(AddrInst);
3079 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
3080 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
3081 const StructLayout *SL = DL.getStructLayout(STy);
3083 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
3084 ConstantOffset += SL->getElementOffset(Idx);
3086 uint64_t TypeSize = DL.getTypeAllocSize(GTI.getIndexedType());
3087 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
3088 ConstantOffset += CI->getSExtValue()*TypeSize;
3089 } else if (TypeSize) { // Scales of zero don't do anything.
3090 // We only allow one variable index at the moment.
3091 if (VariableOperand != -1)
3094 // Remember the variable index.
3095 VariableOperand = i;
3096 VariableScale = TypeSize;
3101 // A common case is for the GEP to only do a constant offset. In this case,
3102 // just add it to the disp field and check validity.
3103 if (VariableOperand == -1) {
3104 AddrMode.BaseOffs += ConstantOffset;
3105 if (ConstantOffset == 0 ||
3106 TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) {
3107 // Check to see if we can fold the base pointer in too.
3108 if (matchAddr(AddrInst->getOperand(0), Depth+1))
3111 AddrMode.BaseOffs -= ConstantOffset;
3115 // Save the valid addressing mode in case we can't match.
3116 ExtAddrMode BackupAddrMode = AddrMode;
3117 unsigned OldSize = AddrModeInsts.size();
3119 // See if the scale and offset amount is valid for this target.
3120 AddrMode.BaseOffs += ConstantOffset;
3122 // Match the base operand of the GEP.
3123 if (!matchAddr(AddrInst->getOperand(0), Depth+1)) {
3124 // If it couldn't be matched, just stuff the value in a register.
3125 if (AddrMode.HasBaseReg) {
3126 AddrMode = BackupAddrMode;
3127 AddrModeInsts.resize(OldSize);
3130 AddrMode.HasBaseReg = true;
3131 AddrMode.BaseReg = AddrInst->getOperand(0);
3134 // Match the remaining variable portion of the GEP.
3135 if (!matchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
3137 // If it couldn't be matched, try stuffing the base into a register
3138 // instead of matching it, and retrying the match of the scale.
3139 AddrMode = BackupAddrMode;
3140 AddrModeInsts.resize(OldSize);
3141 if (AddrMode.HasBaseReg)
3143 AddrMode.HasBaseReg = true;
3144 AddrMode.BaseReg = AddrInst->getOperand(0);
3145 AddrMode.BaseOffs += ConstantOffset;
3146 if (!matchScaledValue(AddrInst->getOperand(VariableOperand),
3147 VariableScale, Depth)) {
3148 // If even that didn't work, bail.
3149 AddrMode = BackupAddrMode;
3150 AddrModeInsts.resize(OldSize);
3157 case Instruction::SExt:
3158 case Instruction::ZExt: {
3159 Instruction *Ext = dyn_cast<Instruction>(AddrInst);
3163 // Try to move this ext out of the way of the addressing mode.
3164 // Ask for a method for doing so.
3165 TypePromotionHelper::Action TPH =
3166 TypePromotionHelper::getAction(Ext, InsertedInsts, TLI, PromotedInsts);
3170 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3171 TPT.getRestorationPoint();
3172 unsigned CreatedInstsCost = 0;
3173 unsigned ExtCost = !TLI.isExtFree(Ext);
3174 Value *PromotedOperand =
3175 TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI);
3176 // SExt has been moved away.
3177 // Thus either it will be rematched later in the recursive calls or it is
3178 // gone. Anyway, we must not fold it into the addressing mode at this point.
3182 // addr = gep base, idx
3184 // promotedOpnd = ext opnd <- no match here
3185 // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
3186 // addr = gep base, op <- match
3190 assert(PromotedOperand &&
3191 "TypePromotionHelper should have filtered out those cases");
3193 ExtAddrMode BackupAddrMode = AddrMode;
3194 unsigned OldSize = AddrModeInsts.size();
3196 if (!matchAddr(PromotedOperand, Depth) ||
3197 // The total of the new cost is equal to the cost of the created
3199 // The total of the old cost is equal to the cost of the extension plus
3200 // what we have saved in the addressing mode.
3201 !isPromotionProfitable(CreatedInstsCost,
3202 ExtCost + (AddrModeInsts.size() - OldSize),
3204 AddrMode = BackupAddrMode;
3205 AddrModeInsts.resize(OldSize);
3206 DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
3207 TPT.rollback(LastKnownGood);
3216 /// If we can, try to add the value of 'Addr' into the current addressing mode.
3217 /// If Addr can't be added to AddrMode this returns false and leaves AddrMode
3218 /// unmodified. This assumes that Addr is either a pointer type or intptr_t
3221 bool AddressingModeMatcher::matchAddr(Value *Addr, unsigned Depth) {
3222 // Start a transaction at this point that we will rollback if the matching
3224 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3225 TPT.getRestorationPoint();
3226 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
3227 // Fold in immediates if legal for the target.
3228 AddrMode.BaseOffs += CI->getSExtValue();
3229 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3231 AddrMode.BaseOffs -= CI->getSExtValue();
3232 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
3233 // If this is a global variable, try to fold it into the addressing mode.
3234 if (!AddrMode.BaseGV) {
3235 AddrMode.BaseGV = GV;
3236 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3238 AddrMode.BaseGV = nullptr;
3240 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
3241 ExtAddrMode BackupAddrMode = AddrMode;
3242 unsigned OldSize = AddrModeInsts.size();
3244 // Check to see if it is possible to fold this operation.
3245 bool MovedAway = false;
3246 if (matchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
3247 // This instruction may have been moved away. If so, there is nothing
3251 // Okay, it's possible to fold this. Check to see if it is actually
3252 // *profitable* to do so. We use a simple cost model to avoid increasing
3253 // register pressure too much.
3254 if (I->hasOneUse() ||
3255 isProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
3256 AddrModeInsts.push_back(I);
3260 // It isn't profitable to do this, roll back.
3261 //cerr << "NOT FOLDING: " << *I;
3262 AddrMode = BackupAddrMode;
3263 AddrModeInsts.resize(OldSize);
3264 TPT.rollback(LastKnownGood);
3266 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
3267 if (matchOperationAddr(CE, CE->getOpcode(), Depth))
3269 TPT.rollback(LastKnownGood);
3270 } else if (isa<ConstantPointerNull>(Addr)) {
3271 // Null pointer gets folded without affecting the addressing mode.
3275 // Worse case, the target should support [reg] addressing modes. :)
3276 if (!AddrMode.HasBaseReg) {
3277 AddrMode.HasBaseReg = true;
3278 AddrMode.BaseReg = Addr;
3279 // Still check for legality in case the target supports [imm] but not [i+r].
3280 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3282 AddrMode.HasBaseReg = false;
3283 AddrMode.BaseReg = nullptr;
3286 // If the base register is already taken, see if we can do [r+r].
3287 if (AddrMode.Scale == 0) {
3289 AddrMode.ScaledReg = Addr;
3290 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3293 AddrMode.ScaledReg = nullptr;
3296 TPT.rollback(LastKnownGood);
3300 /// Check to see if all uses of OpVal by the specified inline asm call are due
3301 /// to memory operands. If so, return true, otherwise return false.
3302 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
3303 const TargetMachine &TM) {
3304 const Function *F = CI->getParent()->getParent();
3305 const TargetLowering *TLI = TM.getSubtargetImpl(*F)->getTargetLowering();
3306 const TargetRegisterInfo *TRI = TM.getSubtargetImpl(*F)->getRegisterInfo();
3307 TargetLowering::AsmOperandInfoVector TargetConstraints =
3308 TLI->ParseConstraints(F->getParent()->getDataLayout(), TRI,
3309 ImmutableCallSite(CI));
3310 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
3311 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
3313 // Compute the constraint code and ConstraintType to use.
3314 TLI->ComputeConstraintToUse(OpInfo, SDValue());
3316 // If this asm operand is our Value*, and if it isn't an indirect memory
3317 // operand, we can't fold it!
3318 if (OpInfo.CallOperandVal == OpVal &&
3319 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
3320 !OpInfo.isIndirect))
3327 /// Recursively walk all the uses of I until we find a memory use.
3328 /// If we find an obviously non-foldable instruction, return true.
3329 /// Add the ultimately found memory instructions to MemoryUses.
3330 static bool FindAllMemoryUses(
3332 SmallVectorImpl<std::pair<Instruction *, unsigned>> &MemoryUses,
3333 SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetMachine &TM) {
3334 // If we already considered this instruction, we're done.
3335 if (!ConsideredInsts.insert(I).second)
3338 // If this is an obviously unfoldable instruction, bail out.
3339 if (!MightBeFoldableInst(I))
3342 // Loop over all the uses, recursively processing them.
3343 for (Use &U : I->uses()) {
3344 Instruction *UserI = cast<Instruction>(U.getUser());
3346 if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
3347 MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
3351 if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
3352 unsigned opNo = U.getOperandNo();
3353 if (opNo == 0) return true; // Storing addr, not into addr.
3354 MemoryUses.push_back(std::make_pair(SI, opNo));
3358 if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
3359 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
3360 if (!IA) return true;
3362 // If this is a memory operand, we're cool, otherwise bail out.
3363 if (!IsOperandAMemoryOperand(CI, IA, I, TM))
3368 if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TM))
3375 /// Return true if Val is already known to be live at the use site that we're
3376 /// folding it into. If so, there is no cost to include it in the addressing
3377 /// mode. KnownLive1 and KnownLive2 are two values that we know are live at the
3378 /// instruction already.
3379 bool AddressingModeMatcher::valueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
3380 Value *KnownLive2) {
3381 // If Val is either of the known-live values, we know it is live!
3382 if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
3385 // All values other than instructions and arguments (e.g. constants) are live.
3386 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
3388 // If Val is a constant sized alloca in the entry block, it is live, this is
3389 // true because it is just a reference to the stack/frame pointer, which is
3390 // live for the whole function.
3391 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
3392 if (AI->isStaticAlloca())
3395 // Check to see if this value is already used in the memory instruction's
3396 // block. If so, it's already live into the block at the very least, so we
3397 // can reasonably fold it.
3398 return Val->isUsedInBasicBlock(MemoryInst->getParent());
3401 /// It is possible for the addressing mode of the machine to fold the specified
3402 /// instruction into a load or store that ultimately uses it.
3403 /// However, the specified instruction has multiple uses.
3404 /// Given this, it may actually increase register pressure to fold it
3405 /// into the load. For example, consider this code:
3409 /// use(Y) -> nonload/store
3413 /// In this case, Y has multiple uses, and can be folded into the load of Z
3414 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
3415 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
3416 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
3417 /// number of computations either.
3419 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
3420 /// X was live across 'load Z' for other reasons, we actually *would* want to
3421 /// fold the addressing mode in the Z case. This would make Y die earlier.
3422 bool AddressingModeMatcher::
3423 isProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
3424 ExtAddrMode &AMAfter) {
3425 if (IgnoreProfitability) return true;
3427 // AMBefore is the addressing mode before this instruction was folded into it,
3428 // and AMAfter is the addressing mode after the instruction was folded. Get
3429 // the set of registers referenced by AMAfter and subtract out those
3430 // referenced by AMBefore: this is the set of values which folding in this
3431 // address extends the lifetime of.
3433 // Note that there are only two potential values being referenced here,
3434 // BaseReg and ScaleReg (global addresses are always available, as are any
3435 // folded immediates).
3436 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
3438 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
3439 // lifetime wasn't extended by adding this instruction.
3440 if (valueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
3442 if (valueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
3443 ScaledReg = nullptr;
3445 // If folding this instruction (and it's subexprs) didn't extend any live
3446 // ranges, we're ok with it.
3447 if (!BaseReg && !ScaledReg)
3450 // If all uses of this instruction are ultimately load/store/inlineasm's,
3451 // check to see if their addressing modes will include this instruction. If
3452 // so, we can fold it into all uses, so it doesn't matter if it has multiple
3454 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
3455 SmallPtrSet<Instruction*, 16> ConsideredInsts;
3456 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TM))
3457 return false; // Has a non-memory, non-foldable use!
3459 // Now that we know that all uses of this instruction are part of a chain of
3460 // computation involving only operations that could theoretically be folded
3461 // into a memory use, loop over each of these uses and see if they could
3462 // *actually* fold the instruction.
3463 SmallVector<Instruction*, 32> MatchedAddrModeInsts;
3464 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
3465 Instruction *User = MemoryUses[i].first;
3466 unsigned OpNo = MemoryUses[i].second;
3468 // Get the access type of this use. If the use isn't a pointer, we don't
3469 // know what it accesses.
3470 Value *Address = User->getOperand(OpNo);
3471 PointerType *AddrTy = dyn_cast<PointerType>(Address->getType());
3474 Type *AddressAccessTy = AddrTy->getElementType();
3475 unsigned AS = AddrTy->getAddressSpace();
3477 // Do a match against the root of this address, ignoring profitability. This
3478 // will tell us if the addressing mode for the memory operation will
3479 // *actually* cover the shared instruction.
3481 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3482 TPT.getRestorationPoint();
3483 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TM, AddressAccessTy, AS,
3484 MemoryInst, Result, InsertedInsts,
3485 PromotedInsts, TPT);
3486 Matcher.IgnoreProfitability = true;
3487 bool Success = Matcher.matchAddr(Address, 0);
3488 (void)Success; assert(Success && "Couldn't select *anything*?");
3490 // The match was to check the profitability, the changes made are not
3491 // part of the original matcher. Therefore, they should be dropped
3492 // otherwise the original matcher will not present the right state.
3493 TPT.rollback(LastKnownGood);
3495 // If the match didn't cover I, then it won't be shared by it.
3496 if (std::find(MatchedAddrModeInsts.begin(), MatchedAddrModeInsts.end(),
3497 I) == MatchedAddrModeInsts.end())
3500 MatchedAddrModeInsts.clear();
3506 } // end anonymous namespace
3508 /// Return true if the specified values are defined in a
3509 /// different basic block than BB.
3510 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
3511 if (Instruction *I = dyn_cast<Instruction>(V))
3512 return I->getParent() != BB;
3516 /// Load and Store Instructions often have addressing modes that can do
3517 /// significant amounts of computation. As such, instruction selection will try
3518 /// to get the load or store to do as much computation as possible for the
3519 /// program. The problem is that isel can only see within a single block. As
3520 /// such, we sink as much legal addressing mode work into the block as possible.
3522 /// This method is used to optimize both load/store and inline asms with memory
3524 bool CodeGenPrepare::optimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
3525 Type *AccessTy, unsigned AddrSpace) {
3528 // Try to collapse single-value PHI nodes. This is necessary to undo
3529 // unprofitable PRE transformations.
3530 SmallVector<Value*, 8> worklist;
3531 SmallPtrSet<Value*, 16> Visited;
3532 worklist.push_back(Addr);
3534 // Use a worklist to iteratively look through PHI nodes, and ensure that
3535 // the addressing mode obtained from the non-PHI roots of the graph
3537 Value *Consensus = nullptr;
3538 unsigned NumUsesConsensus = 0;
3539 bool IsNumUsesConsensusValid = false;
3540 SmallVector<Instruction*, 16> AddrModeInsts;
3541 ExtAddrMode AddrMode;
3542 TypePromotionTransaction TPT;
3543 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3544 TPT.getRestorationPoint();
3545 while (!worklist.empty()) {
3546 Value *V = worklist.back();
3547 worklist.pop_back();
3549 // Break use-def graph loops.
3550 if (!Visited.insert(V).second) {
3551 Consensus = nullptr;
3555 // For a PHI node, push all of its incoming values.
3556 if (PHINode *P = dyn_cast<PHINode>(V)) {
3557 for (Value *IncValue : P->incoming_values())
3558 worklist.push_back(IncValue);
3562 // For non-PHIs, determine the addressing mode being computed.
3563 SmallVector<Instruction*, 16> NewAddrModeInsts;
3564 ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
3565 V, AccessTy, AddrSpace, MemoryInst, NewAddrModeInsts, *TM,
3566 InsertedInsts, PromotedInsts, TPT);
3568 // This check is broken into two cases with very similar code to avoid using
3569 // getNumUses() as much as possible. Some values have a lot of uses, so
3570 // calling getNumUses() unconditionally caused a significant compile-time
3574 AddrMode = NewAddrMode;
3575 AddrModeInsts = NewAddrModeInsts;
3577 } else if (NewAddrMode == AddrMode) {
3578 if (!IsNumUsesConsensusValid) {
3579 NumUsesConsensus = Consensus->getNumUses();
3580 IsNumUsesConsensusValid = true;
3583 // Ensure that the obtained addressing mode is equivalent to that obtained
3584 // for all other roots of the PHI traversal. Also, when choosing one
3585 // such root as representative, select the one with the most uses in order
3586 // to keep the cost modeling heuristics in AddressingModeMatcher
3588 unsigned NumUses = V->getNumUses();
3589 if (NumUses > NumUsesConsensus) {
3591 NumUsesConsensus = NumUses;
3592 AddrModeInsts = NewAddrModeInsts;
3597 Consensus = nullptr;
3601 // If the addressing mode couldn't be determined, or if multiple different
3602 // ones were determined, bail out now.
3604 TPT.rollback(LastKnownGood);
3609 // Check to see if any of the instructions supersumed by this addr mode are
3610 // non-local to I's BB.
3611 bool AnyNonLocal = false;
3612 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
3613 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
3619 // If all the instructions matched are already in this BB, don't do anything.
3621 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
3625 // Insert this computation right after this user. Since our caller is
3626 // scanning from the top of the BB to the bottom, reuse of the expr are
3627 // guaranteed to happen later.
3628 IRBuilder<> Builder(MemoryInst);
3630 // Now that we determined the addressing expression we want to use and know
3631 // that we have to sink it into this block. Check to see if we have already
3632 // done this for some other load/store instr in this block. If so, reuse the
3634 Value *&SunkAddr = SunkAddrs[Addr];
3636 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
3637 << *MemoryInst << "\n");
3638 if (SunkAddr->getType() != Addr->getType())
3639 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
3640 } else if (AddrSinkUsingGEPs ||
3641 (!AddrSinkUsingGEPs.getNumOccurrences() && TM &&
3642 TM->getSubtargetImpl(*MemoryInst->getParent()->getParent())
3644 // By default, we use the GEP-based method when AA is used later. This
3645 // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
3646 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
3647 << *MemoryInst << "\n");
3648 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
3649 Value *ResultPtr = nullptr, *ResultIndex = nullptr;
3651 // First, find the pointer.
3652 if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
3653 ResultPtr = AddrMode.BaseReg;
3654 AddrMode.BaseReg = nullptr;
3657 if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
3658 // We can't add more than one pointer together, nor can we scale a
3659 // pointer (both of which seem meaningless).
3660 if (ResultPtr || AddrMode.Scale != 1)
3663 ResultPtr = AddrMode.ScaledReg;
3667 if (AddrMode.BaseGV) {
3671 ResultPtr = AddrMode.BaseGV;
3674 // If the real base value actually came from an inttoptr, then the matcher
3675 // will look through it and provide only the integer value. In that case,
3677 if (!ResultPtr && AddrMode.BaseReg) {
3679 Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(), "sunkaddr");
3680 AddrMode.BaseReg = nullptr;
3681 } else if (!ResultPtr && AddrMode.Scale == 1) {
3683 Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(), "sunkaddr");
3688 !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
3689 SunkAddr = Constant::getNullValue(Addr->getType());
3690 } else if (!ResultPtr) {
3694 Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
3695 Type *I8Ty = Builder.getInt8Ty();
3697 // Start with the base register. Do this first so that subsequent address
3698 // matching finds it last, which will prevent it from trying to match it
3699 // as the scaled value in case it happens to be a mul. That would be
3700 // problematic if we've sunk a different mul for the scale, because then
3701 // we'd end up sinking both muls.
3702 if (AddrMode.BaseReg) {
3703 Value *V = AddrMode.BaseReg;
3704 if (V->getType() != IntPtrTy)
3705 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
3710 // Add the scale value.
3711 if (AddrMode.Scale) {
3712 Value *V = AddrMode.ScaledReg;
3713 if (V->getType() == IntPtrTy) {
3715 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
3716 cast<IntegerType>(V->getType())->getBitWidth()) {
3717 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
3719 // It is only safe to sign extend the BaseReg if we know that the math
3720 // required to create it did not overflow before we extend it. Since
3721 // the original IR value was tossed in favor of a constant back when
3722 // the AddrMode was created we need to bail out gracefully if widths
3723 // do not match instead of extending it.
3724 Instruction *I = dyn_cast_or_null<Instruction>(ResultIndex);
3725 if (I && (ResultIndex != AddrMode.BaseReg))
3726 I->eraseFromParent();
3730 if (AddrMode.Scale != 1)
3731 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
3734 ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
3739 // Add in the Base Offset if present.
3740 if (AddrMode.BaseOffs) {
3741 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
3743 // We need to add this separately from the scale above to help with
3744 // SDAG consecutive load/store merging.
3745 if (ResultPtr->getType() != I8PtrTy)
3746 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
3747 ResultPtr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
3754 SunkAddr = ResultPtr;
3756 if (ResultPtr->getType() != I8PtrTy)
3757 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
3758 SunkAddr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
3761 if (SunkAddr->getType() != Addr->getType())
3762 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
3765 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
3766 << *MemoryInst << "\n");
3767 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
3768 Value *Result = nullptr;
3770 // Start with the base register. Do this first so that subsequent address
3771 // matching finds it last, which will prevent it from trying to match it
3772 // as the scaled value in case it happens to be a mul. That would be
3773 // problematic if we've sunk a different mul for the scale, because then
3774 // we'd end up sinking both muls.
3775 if (AddrMode.BaseReg) {
3776 Value *V = AddrMode.BaseReg;
3777 if (V->getType()->isPointerTy())
3778 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
3779 if (V->getType() != IntPtrTy)
3780 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
3784 // Add the scale value.
3785 if (AddrMode.Scale) {
3786 Value *V = AddrMode.ScaledReg;
3787 if (V->getType() == IntPtrTy) {
3789 } else if (V->getType()->isPointerTy()) {
3790 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
3791 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
3792 cast<IntegerType>(V->getType())->getBitWidth()) {
3793 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
3795 // It is only safe to sign extend the BaseReg if we know that the math
3796 // required to create it did not overflow before we extend it. Since
3797 // the original IR value was tossed in favor of a constant back when
3798 // the AddrMode was created we need to bail out gracefully if widths
3799 // do not match instead of extending it.
3800 Instruction *I = dyn_cast_or_null<Instruction>(Result);
3801 if (I && (Result != AddrMode.BaseReg))
3802 I->eraseFromParent();
3805 if (AddrMode.Scale != 1)
3806 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
3809 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3814 // Add in the BaseGV if present.
3815 if (AddrMode.BaseGV) {
3816 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
3818 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3823 // Add in the Base Offset if present.
3824 if (AddrMode.BaseOffs) {
3825 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
3827 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3833 SunkAddr = Constant::getNullValue(Addr->getType());
3835 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
3838 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
3840 // If we have no uses, recursively delete the value and all dead instructions
3842 if (Repl->use_empty()) {
3843 // This can cause recursive deletion, which can invalidate our iterator.
3844 // Use a WeakVH to hold onto it in case this happens.
3845 WeakVH IterHandle(&*CurInstIterator);
3846 BasicBlock *BB = CurInstIterator->getParent();
3848 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
3850 if (IterHandle != CurInstIterator.getNodePtrUnchecked()) {
3851 // If the iterator instruction was recursively deleted, start over at the
3852 // start of the block.
3853 CurInstIterator = BB->begin();
3861 /// If there are any memory operands, use OptimizeMemoryInst to sink their
3862 /// address computing into the block when possible / profitable.
3863 bool CodeGenPrepare::optimizeInlineAsmInst(CallInst *CS) {
3864 bool MadeChange = false;
3866 const TargetRegisterInfo *TRI =
3867 TM->getSubtargetImpl(*CS->getParent()->getParent())->getRegisterInfo();
3868 TargetLowering::AsmOperandInfoVector TargetConstraints =
3869 TLI->ParseConstraints(*DL, TRI, CS);
3871 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
3872 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
3874 // Compute the constraint code and ConstraintType to use.
3875 TLI->ComputeConstraintToUse(OpInfo, SDValue());
3877 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
3878 OpInfo.isIndirect) {
3879 Value *OpVal = CS->getArgOperand(ArgNo++);
3880 MadeChange |= optimizeMemoryInst(CS, OpVal, OpVal->getType(), ~0u);
3881 } else if (OpInfo.Type == InlineAsm::isInput)
3888 /// \brief Check if all the uses of \p Inst are equivalent (or free) zero or
3889 /// sign extensions.
3890 static bool hasSameExtUse(Instruction *Inst, const TargetLowering &TLI) {
3891 assert(!Inst->use_empty() && "Input must have at least one use");
3892 const Instruction *FirstUser = cast<Instruction>(*Inst->user_begin());
3893 bool IsSExt = isa<SExtInst>(FirstUser);
3894 Type *ExtTy = FirstUser->getType();
3895 for (const User *U : Inst->users()) {
3896 const Instruction *UI = cast<Instruction>(U);
3897 if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI)))
3899 Type *CurTy = UI->getType();
3900 // Same input and output types: Same instruction after CSE.
3904 // If IsSExt is true, we are in this situation:
3906 // b = sext ty1 a to ty2
3907 // c = sext ty1 a to ty3
3908 // Assuming ty2 is shorter than ty3, this could be turned into:
3910 // b = sext ty1 a to ty2
3911 // c = sext ty2 b to ty3
3912 // However, the last sext is not free.
3916 // This is a ZExt, maybe this is free to extend from one type to another.
3917 // In that case, we would not account for a different use.
3920 if (ExtTy->getScalarType()->getIntegerBitWidth() >
3921 CurTy->getScalarType()->getIntegerBitWidth()) {
3929 if (!TLI.isZExtFree(NarrowTy, LargeTy))
3932 // All uses are the same or can be derived from one another for free.
3936 /// \brief Try to form ExtLd by promoting \p Exts until they reach a
3937 /// load instruction.
3938 /// If an ext(load) can be formed, it is returned via \p LI for the load
3939 /// and \p Inst for the extension.
3940 /// Otherwise LI == nullptr and Inst == nullptr.
3941 /// When some promotion happened, \p TPT contains the proper state to
3944 /// \return true when promoting was necessary to expose the ext(load)
3945 /// opportunity, false otherwise.
3949 /// %ld = load i32* %addr
3950 /// %add = add nuw i32 %ld, 4
3951 /// %zext = zext i32 %add to i64
3955 /// %ld = load i32* %addr
3956 /// %zext = zext i32 %ld to i64
3957 /// %add = add nuw i64 %zext, 4
3959 /// Thanks to the promotion, we can match zext(load i32*) to i64.
3960 bool CodeGenPrepare::extLdPromotion(TypePromotionTransaction &TPT,
3961 LoadInst *&LI, Instruction *&Inst,
3962 const SmallVectorImpl<Instruction *> &Exts,
3963 unsigned CreatedInstsCost = 0) {
3964 // Iterate over all the extensions to see if one form an ext(load).
3965 for (auto I : Exts) {
3966 // Check if we directly have ext(load).
3967 if ((LI = dyn_cast<LoadInst>(I->getOperand(0)))) {
3969 // No promotion happened here.
3972 // Check whether or not we want to do any promotion.
3973 if (!TLI || !TLI->enableExtLdPromotion() || DisableExtLdPromotion)
3975 // Get the action to perform the promotion.
3976 TypePromotionHelper::Action TPH = TypePromotionHelper::getAction(
3977 I, InsertedInsts, *TLI, PromotedInsts);
3978 // Check if we can promote.
3981 // Save the current state.
3982 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3983 TPT.getRestorationPoint();
3984 SmallVector<Instruction *, 4> NewExts;
3985 unsigned NewCreatedInstsCost = 0;
3986 unsigned ExtCost = !TLI->isExtFree(I);
3988 Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost,
3989 &NewExts, nullptr, *TLI);
3990 assert(PromotedVal &&
3991 "TypePromotionHelper should have filtered out those cases");
3993 // We would be able to merge only one extension in a load.
3994 // Therefore, if we have more than 1 new extension we heuristically
3995 // cut this search path, because it means we degrade the code quality.
3996 // With exactly 2, the transformation is neutral, because we will merge
3997 // one extension but leave one. However, we optimistically keep going,
3998 // because the new extension may be removed too.
3999 long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost;
4000 TotalCreatedInstsCost -= ExtCost;
4001 if (!StressExtLdPromotion &&
4002 (TotalCreatedInstsCost > 1 ||
4003 !isPromotedInstructionLegal(*TLI, *DL, PromotedVal))) {
4004 // The promotion is not profitable, rollback to the previous state.
4005 TPT.rollback(LastKnownGood);
4008 // The promotion is profitable.
4009 // Check if it exposes an ext(load).
4010 (void)extLdPromotion(TPT, LI, Inst, NewExts, TotalCreatedInstsCost);
4011 if (LI && (StressExtLdPromotion || NewCreatedInstsCost <= ExtCost ||
4012 // If we have created a new extension, i.e., now we have two
4013 // extensions. We must make sure one of them is merged with
4014 // the load, otherwise we may degrade the code quality.
4015 (LI->hasOneUse() || hasSameExtUse(LI, *TLI))))
4016 // Promotion happened.
4018 // If this does not help to expose an ext(load) then, rollback.
4019 TPT.rollback(LastKnownGood);
4021 // None of the extension can form an ext(load).
4027 /// Move a zext or sext fed by a load into the same basic block as the load,
4028 /// unless conditions are unfavorable. This allows SelectionDAG to fold the
4029 /// extend into the load.
4030 /// \p I[in/out] the extension may be modified during the process if some
4031 /// promotions apply.
4033 bool CodeGenPrepare::moveExtToFormExtLoad(Instruction *&I) {
4034 // Try to promote a chain of computation if it allows to form
4035 // an extended load.
4036 TypePromotionTransaction TPT;
4037 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4038 TPT.getRestorationPoint();
4039 SmallVector<Instruction *, 1> Exts;
4041 // Look for a load being extended.
4042 LoadInst *LI = nullptr;
4043 Instruction *OldExt = I;
4044 bool HasPromoted = extLdPromotion(TPT, LI, I, Exts);
4046 assert(!HasPromoted && !LI && "If we did not match any load instruction "
4047 "the code must remain the same");
4052 // If they're already in the same block, there's nothing to do.
4053 // Make the cheap checks first if we did not promote.
4054 // If we promoted, we need to check if it is indeed profitable.
4055 if (!HasPromoted && LI->getParent() == I->getParent())
4058 EVT VT = TLI->getValueType(*DL, I->getType());
4059 EVT LoadVT = TLI->getValueType(*DL, LI->getType());
4061 // If the load has other users and the truncate is not free, this probably
4062 // isn't worthwhile.
4063 if (!LI->hasOneUse() && TLI &&
4064 (TLI->isTypeLegal(LoadVT) || !TLI->isTypeLegal(VT)) &&
4065 !TLI->isTruncateFree(I->getType(), LI->getType())) {
4067 TPT.rollback(LastKnownGood);
4071 // Check whether the target supports casts folded into loads.
4073 if (isa<ZExtInst>(I))
4074 LType = ISD::ZEXTLOAD;
4076 assert(isa<SExtInst>(I) && "Unexpected ext type!");
4077 LType = ISD::SEXTLOAD;
4079 if (TLI && !TLI->isLoadExtLegal(LType, VT, LoadVT)) {
4081 TPT.rollback(LastKnownGood);
4085 // Move the extend into the same block as the load, so that SelectionDAG
4088 I->removeFromParent();
4094 bool CodeGenPrepare::optimizeExtUses(Instruction *I) {
4095 BasicBlock *DefBB = I->getParent();
4097 // If the result of a {s|z}ext and its source are both live out, rewrite all
4098 // other uses of the source with result of extension.
4099 Value *Src = I->getOperand(0);
4100 if (Src->hasOneUse())
4103 // Only do this xform if truncating is free.
4104 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
4107 // Only safe to perform the optimization if the source is also defined in
4109 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
4112 bool DefIsLiveOut = false;
4113 for (User *U : I->users()) {
4114 Instruction *UI = cast<Instruction>(U);
4116 // Figure out which BB this ext is used in.
4117 BasicBlock *UserBB = UI->getParent();
4118 if (UserBB == DefBB) continue;
4119 DefIsLiveOut = true;
4125 // Make sure none of the uses are PHI nodes.
4126 for (User *U : Src->users()) {
4127 Instruction *UI = cast<Instruction>(U);
4128 BasicBlock *UserBB = UI->getParent();
4129 if (UserBB == DefBB) continue;
4130 // Be conservative. We don't want this xform to end up introducing
4131 // reloads just before load / store instructions.
4132 if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
4136 // InsertedTruncs - Only insert one trunc in each block once.
4137 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
4139 bool MadeChange = false;
4140 for (Use &U : Src->uses()) {
4141 Instruction *User = cast<Instruction>(U.getUser());
4143 // Figure out which BB this ext is used in.
4144 BasicBlock *UserBB = User->getParent();
4145 if (UserBB == DefBB) continue;
4147 // Both src and def are live in this block. Rewrite the use.
4148 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
4150 if (!InsertedTrunc) {
4151 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
4152 assert(InsertPt != UserBB->end());
4153 InsertedTrunc = new TruncInst(I, Src->getType(), "", &*InsertPt);
4154 InsertedInsts.insert(InsertedTrunc);
4157 // Replace a use of the {s|z}ext source with a use of the result.
4166 /// Check if V (an operand of a select instruction) is an expensive instruction
4167 /// that is only used once.
4168 static bool sinkSelectOperand(const TargetTransformInfo *TTI, Value *V) {
4169 auto *I = dyn_cast<Instruction>(V);
4170 // If it's safe to speculatively execute, then it should not have side
4171 // effects; therefore, it's safe to sink and possibly *not* execute.
4172 return I && I->hasOneUse() && isSafeToSpeculativelyExecute(I) &&
4173 TTI->getUserCost(I) >= TargetTransformInfo::TCC_Expensive;
4176 /// Returns true if a SelectInst should be turned into an explicit branch.
4177 static bool isFormingBranchFromSelectProfitable(const TargetTransformInfo *TTI,
4179 // FIXME: This should use the same heuristics as IfConversion to determine
4180 // whether a select is better represented as a branch. This requires that
4181 // branch probability metadata is preserved for the select, which is not the
4184 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
4186 // If a branch is predictable, an out-of-order CPU can avoid blocking on its
4187 // comparison condition. If the compare has more than one use, there's
4188 // probably another cmov or setcc around, so it's not worth emitting a branch.
4189 if (!Cmp || !Cmp->hasOneUse())
4192 Value *CmpOp0 = Cmp->getOperand(0);
4193 Value *CmpOp1 = Cmp->getOperand(1);
4195 // Emit "cmov on compare with a memory operand" as a branch to avoid stalls
4196 // on a load from memory. But if the load is used more than once, do not
4197 // change the select to a branch because the load is probably needed
4198 // regardless of whether the branch is taken or not.
4199 if ((isa<LoadInst>(CmpOp0) && CmpOp0->hasOneUse()) ||
4200 (isa<LoadInst>(CmpOp1) && CmpOp1->hasOneUse()))
4203 // If either operand of the select is expensive and only needed on one side
4204 // of the select, we should form a branch.
4205 if (sinkSelectOperand(TTI, SI->getTrueValue()) ||
4206 sinkSelectOperand(TTI, SI->getFalseValue()))
4213 /// If we have a SelectInst that will likely profit from branch prediction,
4214 /// turn it into a branch.
4215 bool CodeGenPrepare::optimizeSelectInst(SelectInst *SI) {
4216 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
4218 // Can we convert the 'select' to CF ?
4219 if (DisableSelectToBranch || OptSize || !TLI || VectorCond)
4222 TargetLowering::SelectSupportKind SelectKind;
4224 SelectKind = TargetLowering::VectorMaskSelect;
4225 else if (SI->getType()->isVectorTy())
4226 SelectKind = TargetLowering::ScalarCondVectorVal;
4228 SelectKind = TargetLowering::ScalarValSelect;
4230 // Do we have efficient codegen support for this kind of 'selects' ?
4231 if (TLI->isSelectSupported(SelectKind)) {
4232 // We have efficient codegen support for the select instruction.
4233 // Check if it is profitable to keep this 'select'.
4234 if (!TLI->isPredictableSelectExpensive() ||
4235 !isFormingBranchFromSelectProfitable(TTI, SI))
4241 // Transform a sequence like this:
4243 // %cmp = cmp uge i32 %a, %b
4244 // %sel = select i1 %cmp, i32 %c, i32 %d
4248 // %cmp = cmp uge i32 %a, %b
4249 // br i1 %cmp, label %select.true, label %select.false
4251 // br label %select.end
4253 // br label %select.end
4255 // %sel = phi i32 [ %c, %select.true ], [ %d, %select.false ]
4257 // In addition, we may sink instructions that produce %c or %d from
4258 // the entry block into the destination(s) of the new branch.
4259 // If the true or false blocks do not contain a sunken instruction, that
4260 // block and its branch may be optimized away. In that case, one side of the
4261 // first branch will point directly to select.end, and the corresponding PHI
4262 // predecessor block will be the start block.
4264 // First, we split the block containing the select into 2 blocks.
4265 BasicBlock *StartBlock = SI->getParent();
4266 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(SI));
4267 BasicBlock *EndBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
4269 // Delete the unconditional branch that was just created by the split.
4270 StartBlock->getTerminator()->eraseFromParent();
4272 // These are the new basic blocks for the conditional branch.
4273 // At least one will become an actual new basic block.
4274 BasicBlock *TrueBlock = nullptr;
4275 BasicBlock *FalseBlock = nullptr;
4277 // Sink expensive instructions into the conditional blocks to avoid executing
4278 // them speculatively.
4279 if (sinkSelectOperand(TTI, SI->getTrueValue())) {
4280 TrueBlock = BasicBlock::Create(SI->getContext(), "select.true.sink",
4281 EndBlock->getParent(), EndBlock);
4282 auto *TrueBranch = BranchInst::Create(EndBlock, TrueBlock);
4283 auto *TrueInst = cast<Instruction>(SI->getTrueValue());
4284 TrueInst->moveBefore(TrueBranch);
4286 if (sinkSelectOperand(TTI, SI->getFalseValue())) {
4287 FalseBlock = BasicBlock::Create(SI->getContext(), "select.false.sink",
4288 EndBlock->getParent(), EndBlock);
4289 auto *FalseBranch = BranchInst::Create(EndBlock, FalseBlock);
4290 auto *FalseInst = cast<Instruction>(SI->getFalseValue());
4291 FalseInst->moveBefore(FalseBranch);
4294 // If there was nothing to sink, then arbitrarily choose the 'false' side
4295 // for a new input value to the PHI.
4296 if (TrueBlock == FalseBlock) {
4297 assert(TrueBlock == nullptr &&
4298 "Unexpected basic block transform while optimizing select");
4300 FalseBlock = BasicBlock::Create(SI->getContext(), "select.false",
4301 EndBlock->getParent(), EndBlock);
4302 BranchInst::Create(EndBlock, FalseBlock);
4305 // Insert the real conditional branch based on the original condition.
4306 // If we did not create a new block for one of the 'true' or 'false' paths
4307 // of the condition, it means that side of the branch goes to the end block
4308 // directly and the path originates from the start block from the point of
4309 // view of the new PHI.
4310 if (TrueBlock == nullptr) {
4311 BranchInst::Create(EndBlock, FalseBlock, SI->getCondition(), SI);
4312 TrueBlock = StartBlock;
4313 } else if (FalseBlock == nullptr) {
4314 BranchInst::Create(TrueBlock, EndBlock, SI->getCondition(), SI);
4315 FalseBlock = StartBlock;
4317 BranchInst::Create(TrueBlock, FalseBlock, SI->getCondition(), SI);
4320 // The select itself is replaced with a PHI Node.
4321 PHINode *PN = PHINode::Create(SI->getType(), 2, "", &EndBlock->front());
4323 PN->addIncoming(SI->getTrueValue(), TrueBlock);
4324 PN->addIncoming(SI->getFalseValue(), FalseBlock);
4326 SI->replaceAllUsesWith(PN);
4327 SI->eraseFromParent();
4329 // Instruct OptimizeBlock to skip to the next block.
4330 CurInstIterator = StartBlock->end();
4331 ++NumSelectsExpanded;
4335 static bool isBroadcastShuffle(ShuffleVectorInst *SVI) {
4336 SmallVector<int, 16> Mask(SVI->getShuffleMask());
4338 for (unsigned i = 0; i < Mask.size(); ++i) {
4339 if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
4341 SplatElem = Mask[i];
4347 /// Some targets have expensive vector shifts if the lanes aren't all the same
4348 /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
4349 /// it's often worth sinking a shufflevector splat down to its use so that
4350 /// codegen can spot all lanes are identical.
4351 bool CodeGenPrepare::optimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
4352 BasicBlock *DefBB = SVI->getParent();
4354 // Only do this xform if variable vector shifts are particularly expensive.
4355 if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
4358 // We only expect better codegen by sinking a shuffle if we can recognise a
4360 if (!isBroadcastShuffle(SVI))
4363 // InsertedShuffles - Only insert a shuffle in each block once.
4364 DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
4366 bool MadeChange = false;
4367 for (User *U : SVI->users()) {
4368 Instruction *UI = cast<Instruction>(U);
4370 // Figure out which BB this ext is used in.
4371 BasicBlock *UserBB = UI->getParent();
4372 if (UserBB == DefBB) continue;
4374 // For now only apply this when the splat is used by a shift instruction.
4375 if (!UI->isShift()) continue;
4377 // Everything checks out, sink the shuffle if the user's block doesn't
4378 // already have a copy.
4379 Instruction *&InsertedShuffle = InsertedShuffles[UserBB];
4381 if (!InsertedShuffle) {
4382 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
4383 assert(InsertPt != UserBB->end());
4385 new ShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1),
4386 SVI->getOperand(2), "", &*InsertPt);
4389 UI->replaceUsesOfWith(SVI, InsertedShuffle);
4393 // If we removed all uses, nuke the shuffle.
4394 if (SVI->use_empty()) {
4395 SVI->eraseFromParent();
4403 /// \brief Helper class to promote a scalar operation to a vector one.
4404 /// This class is used to move downward extractelement transition.
4406 /// a = vector_op <2 x i32>
4407 /// b = extractelement <2 x i32> a, i32 0
4412 /// a = vector_op <2 x i32>
4413 /// c = vector_op a (equivalent to scalar_op on the related lane)
4414 /// * d = extractelement <2 x i32> c, i32 0
4416 /// Assuming both extractelement and store can be combine, we get rid of the
4418 class VectorPromoteHelper {
4419 /// DataLayout associated with the current module.
4420 const DataLayout &DL;
4422 /// Used to perform some checks on the legality of vector operations.
4423 const TargetLowering &TLI;
4425 /// Used to estimated the cost of the promoted chain.
4426 const TargetTransformInfo &TTI;
4428 /// The transition being moved downwards.
4429 Instruction *Transition;
4430 /// The sequence of instructions to be promoted.
4431 SmallVector<Instruction *, 4> InstsToBePromoted;
4432 /// Cost of combining a store and an extract.
4433 unsigned StoreExtractCombineCost;
4434 /// Instruction that will be combined with the transition.
4435 Instruction *CombineInst;
4437 /// \brief The instruction that represents the current end of the transition.
4438 /// Since we are faking the promotion until we reach the end of the chain
4439 /// of computation, we need a way to get the current end of the transition.
4440 Instruction *getEndOfTransition() const {
4441 if (InstsToBePromoted.empty())
4443 return InstsToBePromoted.back();
4446 /// \brief Return the index of the original value in the transition.
4447 /// E.g., for "extractelement <2 x i32> c, i32 1" the original value,
4448 /// c, is at index 0.
4449 unsigned getTransitionOriginalValueIdx() const {
4450 assert(isa<ExtractElementInst>(Transition) &&
4451 "Other kind of transitions are not supported yet");
4455 /// \brief Return the index of the index in the transition.
4456 /// E.g., for "extractelement <2 x i32> c, i32 0" the index
4458 unsigned getTransitionIdx() const {
4459 assert(isa<ExtractElementInst>(Transition) &&
4460 "Other kind of transitions are not supported yet");
4464 /// \brief Get the type of the transition.
4465 /// This is the type of the original value.
4466 /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the
4467 /// transition is <2 x i32>.
4468 Type *getTransitionType() const {
4469 return Transition->getOperand(getTransitionOriginalValueIdx())->getType();
4472 /// \brief Promote \p ToBePromoted by moving \p Def downward through.
4473 /// I.e., we have the following sequence:
4474 /// Def = Transition <ty1> a to <ty2>
4475 /// b = ToBePromoted <ty2> Def, ...
4477 /// b = ToBePromoted <ty1> a, ...
4478 /// Def = Transition <ty1> ToBePromoted to <ty2>
4479 void promoteImpl(Instruction *ToBePromoted);
4481 /// \brief Check whether or not it is profitable to promote all the
4482 /// instructions enqueued to be promoted.
4483 bool isProfitableToPromote() {
4484 Value *ValIdx = Transition->getOperand(getTransitionOriginalValueIdx());
4485 unsigned Index = isa<ConstantInt>(ValIdx)
4486 ? cast<ConstantInt>(ValIdx)->getZExtValue()
4488 Type *PromotedType = getTransitionType();
4490 StoreInst *ST = cast<StoreInst>(CombineInst);
4491 unsigned AS = ST->getPointerAddressSpace();
4492 unsigned Align = ST->getAlignment();
4493 // Check if this store is supported.
4494 if (!TLI.allowsMisalignedMemoryAccesses(
4495 TLI.getValueType(DL, ST->getValueOperand()->getType()), AS,
4497 // If this is not supported, there is no way we can combine
4498 // the extract with the store.
4502 // The scalar chain of computation has to pay for the transition
4503 // scalar to vector.
4504 // The vector chain has to account for the combining cost.
4505 uint64_t ScalarCost =
4506 TTI.getVectorInstrCost(Transition->getOpcode(), PromotedType, Index);
4507 uint64_t VectorCost = StoreExtractCombineCost;
4508 for (const auto &Inst : InstsToBePromoted) {
4509 // Compute the cost.
4510 // By construction, all instructions being promoted are arithmetic ones.
4511 // Moreover, one argument is a constant that can be viewed as a splat
4513 Value *Arg0 = Inst->getOperand(0);
4514 bool IsArg0Constant = isa<UndefValue>(Arg0) || isa<ConstantInt>(Arg0) ||
4515 isa<ConstantFP>(Arg0);
4516 TargetTransformInfo::OperandValueKind Arg0OVK =
4517 IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
4518 : TargetTransformInfo::OK_AnyValue;
4519 TargetTransformInfo::OperandValueKind Arg1OVK =
4520 !IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
4521 : TargetTransformInfo::OK_AnyValue;
4522 ScalarCost += TTI.getArithmeticInstrCost(
4523 Inst->getOpcode(), Inst->getType(), Arg0OVK, Arg1OVK);
4524 VectorCost += TTI.getArithmeticInstrCost(Inst->getOpcode(), PromotedType,
4527 DEBUG(dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
4528 << ScalarCost << "\nVector: " << VectorCost << '\n');
4529 return ScalarCost > VectorCost;
4532 /// \brief Generate a constant vector with \p Val with the same
4533 /// number of elements as the transition.
4534 /// \p UseSplat defines whether or not \p Val should be replicated
4535 /// across the whole vector.
4536 /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>,
4537 /// otherwise we generate a vector with as many undef as possible:
4538 /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only
4539 /// used at the index of the extract.
4540 Value *getConstantVector(Constant *Val, bool UseSplat) const {
4541 unsigned ExtractIdx = UINT_MAX;
4543 // If we cannot determine where the constant must be, we have to
4544 // use a splat constant.
4545 Value *ValExtractIdx = Transition->getOperand(getTransitionIdx());
4546 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(ValExtractIdx))
4547 ExtractIdx = CstVal->getSExtValue();
4552 unsigned End = getTransitionType()->getVectorNumElements();
4554 return ConstantVector::getSplat(End, Val);
4556 SmallVector<Constant *, 4> ConstVec;
4557 UndefValue *UndefVal = UndefValue::get(Val->getType());
4558 for (unsigned Idx = 0; Idx != End; ++Idx) {
4559 if (Idx == ExtractIdx)
4560 ConstVec.push_back(Val);
4562 ConstVec.push_back(UndefVal);
4564 return ConstantVector::get(ConstVec);
4567 /// \brief Check if promoting to a vector type an operand at \p OperandIdx
4568 /// in \p Use can trigger undefined behavior.
4569 static bool canCauseUndefinedBehavior(const Instruction *Use,
4570 unsigned OperandIdx) {
4571 // This is not safe to introduce undef when the operand is on
4572 // the right hand side of a division-like instruction.
4573 if (OperandIdx != 1)
4575 switch (Use->getOpcode()) {
4578 case Instruction::SDiv:
4579 case Instruction::UDiv:
4580 case Instruction::SRem:
4581 case Instruction::URem:
4583 case Instruction::FDiv:
4584 case Instruction::FRem:
4585 return !Use->hasNoNaNs();
4587 llvm_unreachable(nullptr);
4591 VectorPromoteHelper(const DataLayout &DL, const TargetLowering &TLI,
4592 const TargetTransformInfo &TTI, Instruction *Transition,
4593 unsigned CombineCost)
4594 : DL(DL), TLI(TLI), TTI(TTI), Transition(Transition),
4595 StoreExtractCombineCost(CombineCost), CombineInst(nullptr) {
4596 assert(Transition && "Do not know how to promote null");
4599 /// \brief Check if we can promote \p ToBePromoted to \p Type.
4600 bool canPromote(const Instruction *ToBePromoted) const {
4601 // We could support CastInst too.
4602 return isa<BinaryOperator>(ToBePromoted);
4605 /// \brief Check if it is profitable to promote \p ToBePromoted
4606 /// by moving downward the transition through.
4607 bool shouldPromote(const Instruction *ToBePromoted) const {
4608 // Promote only if all the operands can be statically expanded.
4609 // Indeed, we do not want to introduce any new kind of transitions.
4610 for (const Use &U : ToBePromoted->operands()) {
4611 const Value *Val = U.get();
4612 if (Val == getEndOfTransition()) {
4613 // If the use is a division and the transition is on the rhs,
4614 // we cannot promote the operation, otherwise we may create a
4615 // division by zero.
4616 if (canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()))
4620 if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) &&
4621 !isa<ConstantFP>(Val))
4624 // Check that the resulting operation is legal.
4625 int ISDOpcode = TLI.InstructionOpcodeToISD(ToBePromoted->getOpcode());
4628 return StressStoreExtract ||
4629 TLI.isOperationLegalOrCustom(
4630 ISDOpcode, TLI.getValueType(DL, getTransitionType(), true));
4633 /// \brief Check whether or not \p Use can be combined
4634 /// with the transition.
4635 /// I.e., is it possible to do Use(Transition) => AnotherUse?
4636 bool canCombine(const Instruction *Use) { return isa<StoreInst>(Use); }
4638 /// \brief Record \p ToBePromoted as part of the chain to be promoted.
4639 void enqueueForPromotion(Instruction *ToBePromoted) {
4640 InstsToBePromoted.push_back(ToBePromoted);
4643 /// \brief Set the instruction that will be combined with the transition.
4644 void recordCombineInstruction(Instruction *ToBeCombined) {
4645 assert(canCombine(ToBeCombined) && "Unsupported instruction to combine");
4646 CombineInst = ToBeCombined;
4649 /// \brief Promote all the instructions enqueued for promotion if it is
4651 /// \return True if the promotion happened, false otherwise.
4653 // Check if there is something to promote.
4654 // Right now, if we do not have anything to combine with,
4655 // we assume the promotion is not profitable.
4656 if (InstsToBePromoted.empty() || !CombineInst)
4660 if (!StressStoreExtract && !isProfitableToPromote())
4664 for (auto &ToBePromoted : InstsToBePromoted)
4665 promoteImpl(ToBePromoted);
4666 InstsToBePromoted.clear();
4670 } // End of anonymous namespace.
4672 void VectorPromoteHelper::promoteImpl(Instruction *ToBePromoted) {
4673 // At this point, we know that all the operands of ToBePromoted but Def
4674 // can be statically promoted.
4675 // For Def, we need to use its parameter in ToBePromoted:
4676 // b = ToBePromoted ty1 a
4677 // Def = Transition ty1 b to ty2
4678 // Move the transition down.
4679 // 1. Replace all uses of the promoted operation by the transition.
4680 // = ... b => = ... Def.
4681 assert(ToBePromoted->getType() == Transition->getType() &&
4682 "The type of the result of the transition does not match "
4684 ToBePromoted->replaceAllUsesWith(Transition);
4685 // 2. Update the type of the uses.
4686 // b = ToBePromoted ty2 Def => b = ToBePromoted ty1 Def.
4687 Type *TransitionTy = getTransitionType();
4688 ToBePromoted->mutateType(TransitionTy);
4689 // 3. Update all the operands of the promoted operation with promoted
4691 // b = ToBePromoted ty1 Def => b = ToBePromoted ty1 a.
4692 for (Use &U : ToBePromoted->operands()) {
4693 Value *Val = U.get();
4694 Value *NewVal = nullptr;
4695 if (Val == Transition)
4696 NewVal = Transition->getOperand(getTransitionOriginalValueIdx());
4697 else if (isa<UndefValue>(Val) || isa<ConstantInt>(Val) ||
4698 isa<ConstantFP>(Val)) {
4699 // Use a splat constant if it is not safe to use undef.
4700 NewVal = getConstantVector(
4701 cast<Constant>(Val),
4702 isa<UndefValue>(Val) ||
4703 canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()));
4705 llvm_unreachable("Did you modified shouldPromote and forgot to update "
4707 ToBePromoted->setOperand(U.getOperandNo(), NewVal);
4709 Transition->removeFromParent();
4710 Transition->insertAfter(ToBePromoted);
4711 Transition->setOperand(getTransitionOriginalValueIdx(), ToBePromoted);
4714 /// Some targets can do store(extractelement) with one instruction.
4715 /// Try to push the extractelement towards the stores when the target
4716 /// has this feature and this is profitable.
4717 bool CodeGenPrepare::optimizeExtractElementInst(Instruction *Inst) {
4718 unsigned CombineCost = UINT_MAX;
4719 if (DisableStoreExtract || !TLI ||
4720 (!StressStoreExtract &&
4721 !TLI->canCombineStoreAndExtract(Inst->getOperand(0)->getType(),
4722 Inst->getOperand(1), CombineCost)))
4725 // At this point we know that Inst is a vector to scalar transition.
4726 // Try to move it down the def-use chain, until:
4727 // - We can combine the transition with its single use
4728 // => we got rid of the transition.
4729 // - We escape the current basic block
4730 // => we would need to check that we are moving it at a cheaper place and
4731 // we do not do that for now.
4732 BasicBlock *Parent = Inst->getParent();
4733 DEBUG(dbgs() << "Found an interesting transition: " << *Inst << '\n');
4734 VectorPromoteHelper VPH(*DL, *TLI, *TTI, Inst, CombineCost);
4735 // If the transition has more than one use, assume this is not going to be
4737 while (Inst->hasOneUse()) {
4738 Instruction *ToBePromoted = cast<Instruction>(*Inst->user_begin());
4739 DEBUG(dbgs() << "Use: " << *ToBePromoted << '\n');
4741 if (ToBePromoted->getParent() != Parent) {
4742 DEBUG(dbgs() << "Instruction to promote is in a different block ("
4743 << ToBePromoted->getParent()->getName()
4744 << ") than the transition (" << Parent->getName() << ").\n");
4748 if (VPH.canCombine(ToBePromoted)) {
4749 DEBUG(dbgs() << "Assume " << *Inst << '\n'
4750 << "will be combined with: " << *ToBePromoted << '\n');
4751 VPH.recordCombineInstruction(ToBePromoted);
4752 bool Changed = VPH.promote();
4753 NumStoreExtractExposed += Changed;
4757 DEBUG(dbgs() << "Try promoting.\n");
4758 if (!VPH.canPromote(ToBePromoted) || !VPH.shouldPromote(ToBePromoted))
4761 DEBUG(dbgs() << "Promoting is possible... Enqueue for promotion!\n");
4763 VPH.enqueueForPromotion(ToBePromoted);
4764 Inst = ToBePromoted;
4769 bool CodeGenPrepare::optimizeInst(Instruction *I, bool& ModifiedDT) {
4770 // Bail out if we inserted the instruction to prevent optimizations from
4771 // stepping on each other's toes.
4772 if (InsertedInsts.count(I))
4775 if (PHINode *P = dyn_cast<PHINode>(I)) {
4776 // It is possible for very late stage optimizations (such as SimplifyCFG)
4777 // to introduce PHI nodes too late to be cleaned up. If we detect such a
4778 // trivial PHI, go ahead and zap it here.
4779 if (Value *V = SimplifyInstruction(P, *DL, TLInfo, nullptr)) {
4780 P->replaceAllUsesWith(V);
4781 P->eraseFromParent();
4788 if (CastInst *CI = dyn_cast<CastInst>(I)) {
4789 // If the source of the cast is a constant, then this should have
4790 // already been constant folded. The only reason NOT to constant fold
4791 // it is if something (e.g. LSR) was careful to place the constant
4792 // evaluation in a block other than then one that uses it (e.g. to hoist
4793 // the address of globals out of a loop). If this is the case, we don't
4794 // want to forward-subst the cast.
4795 if (isa<Constant>(CI->getOperand(0)))
4798 if (TLI && OptimizeNoopCopyExpression(CI, *TLI, *DL))
4801 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
4802 /// Sink a zext or sext into its user blocks if the target type doesn't
4803 /// fit in one register
4805 TLI->getTypeAction(CI->getContext(),
4806 TLI->getValueType(*DL, CI->getType())) ==
4807 TargetLowering::TypeExpandInteger) {
4808 return SinkCast(CI);
4810 bool MadeChange = moveExtToFormExtLoad(I);
4811 return MadeChange | optimizeExtUses(I);
4817 if (CmpInst *CI = dyn_cast<CmpInst>(I))
4818 if (!TLI || !TLI->hasMultipleConditionRegisters())
4819 return OptimizeCmpExpression(CI);
4821 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
4822 stripInvariantGroupMetadata(*LI);
4824 unsigned AS = LI->getPointerAddressSpace();
4825 return optimizeMemoryInst(I, I->getOperand(0), LI->getType(), AS);
4830 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
4831 stripInvariantGroupMetadata(*SI);
4833 unsigned AS = SI->getPointerAddressSpace();
4834 return optimizeMemoryInst(I, SI->getOperand(1),
4835 SI->getOperand(0)->getType(), AS);
4840 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(I);
4842 if (BinOp && (BinOp->getOpcode() == Instruction::AShr ||
4843 BinOp->getOpcode() == Instruction::LShr)) {
4844 ConstantInt *CI = dyn_cast<ConstantInt>(BinOp->getOperand(1));
4845 if (TLI && CI && TLI->hasExtractBitsInsn())
4846 return OptimizeExtractBits(BinOp, CI, *TLI, *DL);
4851 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
4852 if (GEPI->hasAllZeroIndices()) {
4853 /// The GEP operand must be a pointer, so must its result -> BitCast
4854 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
4855 GEPI->getName(), GEPI);
4856 GEPI->replaceAllUsesWith(NC);
4857 GEPI->eraseFromParent();
4859 optimizeInst(NC, ModifiedDT);
4865 if (CallInst *CI = dyn_cast<CallInst>(I))
4866 return optimizeCallInst(CI, ModifiedDT);
4868 if (SelectInst *SI = dyn_cast<SelectInst>(I))
4869 return optimizeSelectInst(SI);
4871 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I))
4872 return optimizeShuffleVectorInst(SVI);
4874 if (isa<ExtractElementInst>(I))
4875 return optimizeExtractElementInst(I);
4880 // In this pass we look for GEP and cast instructions that are used
4881 // across basic blocks and rewrite them to improve basic-block-at-a-time
4883 bool CodeGenPrepare::optimizeBlock(BasicBlock &BB, bool& ModifiedDT) {
4885 bool MadeChange = false;
4887 CurInstIterator = BB.begin();
4888 while (CurInstIterator != BB.end()) {
4889 MadeChange |= optimizeInst(&*CurInstIterator++, ModifiedDT);
4893 MadeChange |= dupRetToEnableTailCallOpts(&BB);
4898 // llvm.dbg.value is far away from the value then iSel may not be able
4899 // handle it properly. iSel will drop llvm.dbg.value if it can not
4900 // find a node corresponding to the value.
4901 bool CodeGenPrepare::placeDbgValues(Function &F) {
4902 bool MadeChange = false;
4903 for (BasicBlock &BB : F) {
4904 Instruction *PrevNonDbgInst = nullptr;
4905 for (BasicBlock::iterator BI = BB.begin(), BE = BB.end(); BI != BE;) {
4906 Instruction *Insn = &*BI++;
4907 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
4908 // Leave dbg.values that refer to an alloca alone. These
4909 // instrinsics describe the address of a variable (= the alloca)
4910 // being taken. They should not be moved next to the alloca
4911 // (and to the beginning of the scope), but rather stay close to
4912 // where said address is used.
4913 if (!DVI || (DVI->getValue() && isa<AllocaInst>(DVI->getValue()))) {
4914 PrevNonDbgInst = Insn;
4918 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
4919 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
4920 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
4921 DVI->removeFromParent();
4922 if (isa<PHINode>(VI))
4923 DVI->insertBefore(&*VI->getParent()->getFirstInsertionPt());
4925 DVI->insertAfter(VI);
4934 // If there is a sequence that branches based on comparing a single bit
4935 // against zero that can be combined into a single instruction, and the
4936 // target supports folding these into a single instruction, sink the
4937 // mask and compare into the branch uses. Do this before OptimizeBlock ->
4938 // OptimizeInst -> OptimizeCmpExpression, which perturbs the pattern being
4940 bool CodeGenPrepare::sinkAndCmp(Function &F) {
4941 if (!EnableAndCmpSinking)
4943 if (!TLI || !TLI->isMaskAndBranchFoldingLegal())
4945 bool MadeChange = false;
4946 for (Function::iterator I = F.begin(), E = F.end(); I != E; ) {
4947 BasicBlock *BB = &*I++;
4949 // Does this BB end with the following?
4950 // %andVal = and %val, #single-bit-set
4951 // %icmpVal = icmp %andResult, 0
4952 // br i1 %cmpVal label %dest1, label %dest2"
4953 BranchInst *Brcc = dyn_cast<BranchInst>(BB->getTerminator());
4954 if (!Brcc || !Brcc->isConditional())
4956 ICmpInst *Cmp = dyn_cast<ICmpInst>(Brcc->getOperand(0));
4957 if (!Cmp || Cmp->getParent() != BB)
4959 ConstantInt *Zero = dyn_cast<ConstantInt>(Cmp->getOperand(1));
4960 if (!Zero || !Zero->isZero())
4962 Instruction *And = dyn_cast<Instruction>(Cmp->getOperand(0));
4963 if (!And || And->getOpcode() != Instruction::And || And->getParent() != BB)
4965 ConstantInt* Mask = dyn_cast<ConstantInt>(And->getOperand(1));
4966 if (!Mask || !Mask->getUniqueInteger().isPowerOf2())
4968 DEBUG(dbgs() << "found and; icmp ?,0; brcc\n"); DEBUG(BB->dump());
4970 // Push the "and; icmp" for any users that are conditional branches.
4971 // Since there can only be one branch use per BB, we don't need to keep
4972 // track of which BBs we insert into.
4973 for (Value::use_iterator UI = Cmp->use_begin(), E = Cmp->use_end();
4977 BranchInst *BrccUser = dyn_cast<BranchInst>(*UI);
4979 if (!BrccUser || !BrccUser->isConditional())
4981 BasicBlock *UserBB = BrccUser->getParent();
4982 if (UserBB == BB) continue;
4983 DEBUG(dbgs() << "found Brcc use\n");
4985 // Sink the "and; icmp" to use.
4987 BinaryOperator *NewAnd =
4988 BinaryOperator::CreateAnd(And->getOperand(0), And->getOperand(1), "",
4991 CmpInst::Create(Cmp->getOpcode(), Cmp->getPredicate(), NewAnd, Zero,
4995 DEBUG(BrccUser->getParent()->dump());
5001 /// \brief Retrieve the probabilities of a conditional branch. Returns true on
5002 /// success, or returns false if no or invalid metadata was found.
5003 static bool extractBranchMetadata(BranchInst *BI,
5004 uint64_t &ProbTrue, uint64_t &ProbFalse) {
5005 assert(BI->isConditional() &&
5006 "Looking for probabilities on unconditional branch?");
5007 auto *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
5008 if (!ProfileData || ProfileData->getNumOperands() != 3)
5011 const auto *CITrue =
5012 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(1));
5013 const auto *CIFalse =
5014 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(2));
5015 if (!CITrue || !CIFalse)
5018 ProbTrue = CITrue->getValue().getZExtValue();
5019 ProbFalse = CIFalse->getValue().getZExtValue();
5024 /// \brief Scale down both weights to fit into uint32_t.
5025 static void scaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) {
5026 uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse;
5027 uint32_t Scale = (NewMax / UINT32_MAX) + 1;
5028 NewTrue = NewTrue / Scale;
5029 NewFalse = NewFalse / Scale;
5032 /// \brief Some targets prefer to split a conditional branch like:
5034 /// %0 = icmp ne i32 %a, 0
5035 /// %1 = icmp ne i32 %b, 0
5036 /// %or.cond = or i1 %0, %1
5037 /// br i1 %or.cond, label %TrueBB, label %FalseBB
5039 /// into multiple branch instructions like:
5042 /// %0 = icmp ne i32 %a, 0
5043 /// br i1 %0, label %TrueBB, label %bb2
5045 /// %1 = icmp ne i32 %b, 0
5046 /// br i1 %1, label %TrueBB, label %FalseBB
5048 /// This usually allows instruction selection to do even further optimizations
5049 /// and combine the compare with the branch instruction. Currently this is
5050 /// applied for targets which have "cheap" jump instructions.
5052 /// FIXME: Remove the (equivalent?) implementation in SelectionDAG.
5054 bool CodeGenPrepare::splitBranchCondition(Function &F) {
5055 if (!TM || !TM->Options.EnableFastISel || !TLI || TLI->isJumpExpensive())
5058 bool MadeChange = false;
5059 for (auto &BB : F) {
5060 // Does this BB end with the following?
5061 // %cond1 = icmp|fcmp|binary instruction ...
5062 // %cond2 = icmp|fcmp|binary instruction ...
5063 // %cond.or = or|and i1 %cond1, cond2
5064 // br i1 %cond.or label %dest1, label %dest2"
5065 BinaryOperator *LogicOp;
5066 BasicBlock *TBB, *FBB;
5067 if (!match(BB.getTerminator(), m_Br(m_OneUse(m_BinOp(LogicOp)), TBB, FBB)))
5070 auto *Br1 = cast<BranchInst>(BB.getTerminator());
5071 if (Br1->getMetadata(LLVMContext::MD_unpredictable))
5075 Value *Cond1, *Cond2;
5076 if (match(LogicOp, m_And(m_OneUse(m_Value(Cond1)),
5077 m_OneUse(m_Value(Cond2)))))
5078 Opc = Instruction::And;
5079 else if (match(LogicOp, m_Or(m_OneUse(m_Value(Cond1)),
5080 m_OneUse(m_Value(Cond2)))))
5081 Opc = Instruction::Or;
5085 if (!match(Cond1, m_CombineOr(m_Cmp(), m_BinOp())) ||
5086 !match(Cond2, m_CombineOr(m_Cmp(), m_BinOp())) )
5089 DEBUG(dbgs() << "Before branch condition splitting\n"; BB.dump());
5092 auto *InsertBefore = std::next(Function::iterator(BB))
5093 .getNodePtrUnchecked();
5094 auto TmpBB = BasicBlock::Create(BB.getContext(),
5095 BB.getName() + ".cond.split",
5096 BB.getParent(), InsertBefore);
5098 // Update original basic block by using the first condition directly by the
5099 // branch instruction and removing the no longer needed and/or instruction.
5100 Br1->setCondition(Cond1);
5101 LogicOp->eraseFromParent();
5103 // Depending on the conditon we have to either replace the true or the false
5104 // successor of the original branch instruction.
5105 if (Opc == Instruction::And)
5106 Br1->setSuccessor(0, TmpBB);
5108 Br1->setSuccessor(1, TmpBB);
5110 // Fill in the new basic block.
5111 auto *Br2 = IRBuilder<>(TmpBB).CreateCondBr(Cond2, TBB, FBB);
5112 if (auto *I = dyn_cast<Instruction>(Cond2)) {
5113 I->removeFromParent();
5114 I->insertBefore(Br2);
5117 // Update PHI nodes in both successors. The original BB needs to be
5118 // replaced in one succesor's PHI nodes, because the branch comes now from
5119 // the newly generated BB (NewBB). In the other successor we need to add one
5120 // incoming edge to the PHI nodes, because both branch instructions target
5121 // now the same successor. Depending on the original branch condition
5122 // (and/or) we have to swap the successors (TrueDest, FalseDest), so that
5123 // we perfrom the correct update for the PHI nodes.
5124 // This doesn't change the successor order of the just created branch
5125 // instruction (or any other instruction).
5126 if (Opc == Instruction::Or)
5127 std::swap(TBB, FBB);
5129 // Replace the old BB with the new BB.
5130 for (auto &I : *TBB) {
5131 PHINode *PN = dyn_cast<PHINode>(&I);
5135 while ((i = PN->getBasicBlockIndex(&BB)) >= 0)
5136 PN->setIncomingBlock(i, TmpBB);
5139 // Add another incoming edge form the new BB.
5140 for (auto &I : *FBB) {
5141 PHINode *PN = dyn_cast<PHINode>(&I);
5144 auto *Val = PN->getIncomingValueForBlock(&BB);
5145 PN->addIncoming(Val, TmpBB);
5148 // Update the branch weights (from SelectionDAGBuilder::
5149 // FindMergedConditions).
5150 if (Opc == Instruction::Or) {
5151 // Codegen X | Y as:
5160 // We have flexibility in setting Prob for BB1 and Prob for NewBB.
5161 // The requirement is that
5162 // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB)
5163 // = TrueProb for orignal BB.
5164 // Assuming the orignal weights are A and B, one choice is to set BB1's
5165 // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice
5167 // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB.
5168 // Another choice is to assume TrueProb for BB1 equals to TrueProb for
5169 // TmpBB, but the math is more complicated.
5170 uint64_t TrueWeight, FalseWeight;
5171 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
5172 uint64_t NewTrueWeight = TrueWeight;
5173 uint64_t NewFalseWeight = TrueWeight + 2 * FalseWeight;
5174 scaleWeights(NewTrueWeight, NewFalseWeight);
5175 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
5176 .createBranchWeights(TrueWeight, FalseWeight));
5178 NewTrueWeight = TrueWeight;
5179 NewFalseWeight = 2 * FalseWeight;
5180 scaleWeights(NewTrueWeight, NewFalseWeight);
5181 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
5182 .createBranchWeights(TrueWeight, FalseWeight));
5185 // Codegen X & Y as:
5193 // This requires creation of TmpBB after CurBB.
5195 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
5196 // The requirement is that
5197 // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB)
5198 // = FalseProb for orignal BB.
5199 // Assuming the orignal weights are A and B, one choice is to set BB1's
5200 // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice
5202 // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB.
5203 uint64_t TrueWeight, FalseWeight;
5204 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
5205 uint64_t NewTrueWeight = 2 * TrueWeight + FalseWeight;
5206 uint64_t NewFalseWeight = FalseWeight;
5207 scaleWeights(NewTrueWeight, NewFalseWeight);
5208 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
5209 .createBranchWeights(TrueWeight, FalseWeight));
5211 NewTrueWeight = 2 * TrueWeight;
5212 NewFalseWeight = FalseWeight;
5213 scaleWeights(NewTrueWeight, NewFalseWeight);
5214 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
5215 .createBranchWeights(TrueWeight, FalseWeight));
5219 // Note: No point in getting fancy here, since the DT info is never
5220 // available to CodeGenPrepare.
5225 DEBUG(dbgs() << "After branch condition splitting\n"; BB.dump();
5231 void CodeGenPrepare::stripInvariantGroupMetadata(Instruction &I) {
5232 if (auto *InvariantMD = I.getMetadata(LLVMContext::MD_invariant_group))
5233 I.dropUnknownNonDebugMetadata(InvariantMD->getMetadataID());