1 //===- CodeGenPrepare.cpp - Prepare a function for code generation --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass munges the code in the input function to better prepare it for
11 // SelectionDAG-based code generation. This works around limitations in it's
12 // basic-block-at-a-time approach. It should eventually be removed.
14 //===----------------------------------------------------------------------===//
16 #include "llvm/CodeGen/Passes.h"
17 #include "llvm/ADT/DenseMap.h"
18 #include "llvm/ADT/SmallSet.h"
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/Analysis/InstructionSimplify.h"
21 #include "llvm/Analysis/TargetLibraryInfo.h"
22 #include "llvm/Analysis/TargetTransformInfo.h"
23 #include "llvm/IR/CallSite.h"
24 #include "llvm/IR/Constants.h"
25 #include "llvm/IR/DataLayout.h"
26 #include "llvm/IR/DerivedTypes.h"
27 #include "llvm/IR/Dominators.h"
28 #include "llvm/IR/Function.h"
29 #include "llvm/IR/GetElementPtrTypeIterator.h"
30 #include "llvm/IR/IRBuilder.h"
31 #include "llvm/IR/InlineAsm.h"
32 #include "llvm/IR/Instructions.h"
33 #include "llvm/IR/IntrinsicInst.h"
34 #include "llvm/IR/MDBuilder.h"
35 #include "llvm/IR/PatternMatch.h"
36 #include "llvm/IR/Statepoint.h"
37 #include "llvm/IR/ValueHandle.h"
38 #include "llvm/IR/ValueMap.h"
39 #include "llvm/Pass.h"
40 #include "llvm/Support/CommandLine.h"
41 #include "llvm/Support/Debug.h"
42 #include "llvm/Support/raw_ostream.h"
43 #include "llvm/Target/TargetLowering.h"
44 #include "llvm/Target/TargetSubtargetInfo.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/BuildLibCalls.h"
47 #include "llvm/Transforms/Utils/BypassSlowDivision.h"
48 #include "llvm/Transforms/Utils/Local.h"
49 #include "llvm/Transforms/Utils/SimplifyLibCalls.h"
51 using namespace llvm::PatternMatch;
53 #define DEBUG_TYPE "codegenprepare"
55 STATISTIC(NumBlocksElim, "Number of blocks eliminated");
56 STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated");
57 STATISTIC(NumGEPsElim, "Number of GEPs converted to casts");
58 STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "
60 STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "
62 STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "
63 "computations were sunk");
64 STATISTIC(NumExtsMoved, "Number of [s|z]ext instructions combined with loads");
65 STATISTIC(NumExtUses, "Number of uses of [s|z]ext instructions optimized");
66 STATISTIC(NumRetsDup, "Number of return instructions duplicated");
67 STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved");
68 STATISTIC(NumSelectsExpanded, "Number of selects turned into branches");
69 STATISTIC(NumAndCmpsMoved, "Number of and/cmp's pushed into branches");
70 STATISTIC(NumStoreExtractExposed, "Number of store(extractelement) exposed");
72 static cl::opt<bool> DisableBranchOpts(
73 "disable-cgp-branch-opts", cl::Hidden, cl::init(false),
74 cl::desc("Disable branch optimizations in CodeGenPrepare"));
77 DisableGCOpts("disable-cgp-gc-opts", cl::Hidden, cl::init(false),
78 cl::desc("Disable GC optimizations in CodeGenPrepare"));
80 static cl::opt<bool> DisableSelectToBranch(
81 "disable-cgp-select2branch", cl::Hidden, cl::init(false),
82 cl::desc("Disable select to branch conversion."));
84 static cl::opt<bool> AddrSinkUsingGEPs(
85 "addr-sink-using-gep", cl::Hidden, cl::init(false),
86 cl::desc("Address sinking in CGP using GEPs."));
88 static cl::opt<bool> EnableAndCmpSinking(
89 "enable-andcmp-sinking", cl::Hidden, cl::init(true),
90 cl::desc("Enable sinkinig and/cmp into branches."));
92 static cl::opt<bool> DisableStoreExtract(
93 "disable-cgp-store-extract", cl::Hidden, cl::init(false),
94 cl::desc("Disable store(extract) optimizations in CodeGenPrepare"));
96 static cl::opt<bool> StressStoreExtract(
97 "stress-cgp-store-extract", cl::Hidden, cl::init(false),
98 cl::desc("Stress test store(extract) optimizations in CodeGenPrepare"));
100 static cl::opt<bool> DisableExtLdPromotion(
101 "disable-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
102 cl::desc("Disable ext(promotable(ld)) -> promoted(ext(ld)) optimization in "
105 static cl::opt<bool> StressExtLdPromotion(
106 "stress-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
107 cl::desc("Stress test ext(promotable(ld)) -> promoted(ext(ld)) "
108 "optimization in CodeGenPrepare"));
111 typedef SmallPtrSet<Instruction *, 16> SetOfInstrs;
115 TypeIsSExt(Type *Ty, bool IsSExt) : Ty(Ty), IsSExt(IsSExt) {}
117 typedef DenseMap<Instruction *, TypeIsSExt> InstrToOrigTy;
118 class TypePromotionTransaction;
120 class CodeGenPrepare : public FunctionPass {
121 /// TLI - Keep a pointer of a TargetLowering to consult for determining
122 /// transformation profitability.
123 const TargetMachine *TM;
124 const TargetLowering *TLI;
125 const TargetTransformInfo *TTI;
126 const TargetLibraryInfo *TLInfo;
128 /// CurInstIterator - As we scan instructions optimizing them, this is the
129 /// next instruction to optimize. Xforms that can invalidate this should
131 BasicBlock::iterator CurInstIterator;
133 /// Keeps track of non-local addresses that have been sunk into a block.
134 /// This allows us to avoid inserting duplicate code for blocks with
135 /// multiple load/stores of the same address.
136 ValueMap<Value*, Value*> SunkAddrs;
138 /// Keeps track of all instructions inserted for the current function.
139 SetOfInstrs InsertedInsts;
140 /// Keeps track of the type of the related instruction before their
141 /// promotion for the current function.
142 InstrToOrigTy PromotedInsts;
144 /// ModifiedDT - If CFG is modified in anyway.
147 /// OptSize - True if optimizing for size.
150 /// DataLayout for the Function being processed.
151 const DataLayout *DL;
154 static char ID; // Pass identification, replacement for typeid
155 explicit CodeGenPrepare(const TargetMachine *TM = nullptr)
156 : FunctionPass(ID), TM(TM), TLI(nullptr), TTI(nullptr), DL(nullptr) {
157 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
159 bool runOnFunction(Function &F) override;
161 const char *getPassName() const override { return "CodeGen Prepare"; }
163 void getAnalysisUsage(AnalysisUsage &AU) const override {
164 AU.addPreserved<DominatorTreeWrapperPass>();
165 AU.addRequired<TargetLibraryInfoWrapperPass>();
166 AU.addRequired<TargetTransformInfoWrapperPass>();
170 bool EliminateFallThrough(Function &F);
171 bool EliminateMostlyEmptyBlocks(Function &F);
172 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
173 void EliminateMostlyEmptyBlock(BasicBlock *BB);
174 bool OptimizeBlock(BasicBlock &BB, bool& ModifiedDT);
175 bool OptimizeInst(Instruction *I, bool& ModifiedDT);
176 bool OptimizeMemoryInst(Instruction *I, Value *Addr,
177 Type *AccessTy, unsigned AS);
178 bool OptimizeInlineAsmInst(CallInst *CS);
179 bool OptimizeCallInst(CallInst *CI, bool& ModifiedDT);
180 bool MoveExtToFormExtLoad(Instruction *&I);
181 bool OptimizeExtUses(Instruction *I);
182 bool OptimizeSelectInst(SelectInst *SI);
183 bool OptimizeShuffleVectorInst(ShuffleVectorInst *SI);
184 bool OptimizeExtractElementInst(Instruction *Inst);
185 bool DupRetToEnableTailCallOpts(BasicBlock *BB);
186 bool PlaceDbgValues(Function &F);
187 bool sinkAndCmp(Function &F);
188 bool ExtLdPromotion(TypePromotionTransaction &TPT, LoadInst *&LI,
190 const SmallVectorImpl<Instruction *> &Exts,
191 unsigned CreatedInstCost);
192 bool splitBranchCondition(Function &F);
193 bool simplifyOffsetableRelocate(Instruction &I);
197 char CodeGenPrepare::ID = 0;
198 INITIALIZE_TM_PASS(CodeGenPrepare, "codegenprepare",
199 "Optimize for code generation", false, false)
201 FunctionPass *llvm::createCodeGenPreparePass(const TargetMachine *TM) {
202 return new CodeGenPrepare(TM);
205 bool CodeGenPrepare::runOnFunction(Function &F) {
206 if (skipOptnoneFunction(F))
209 DL = &F.getParent()->getDataLayout();
211 bool EverMadeChange = false;
212 // Clear per function information.
213 InsertedInsts.clear();
214 PromotedInsts.clear();
218 TLI = TM->getSubtargetImpl(F)->getTargetLowering();
219 TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
220 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
221 OptSize = F.hasFnAttribute(Attribute::OptimizeForSize);
223 /// This optimization identifies DIV instructions that can be
224 /// profitably bypassed and carried out with a shorter, faster divide.
225 if (!OptSize && TLI && TLI->isSlowDivBypassed()) {
226 const DenseMap<unsigned int, unsigned int> &BypassWidths =
227 TLI->getBypassSlowDivWidths();
228 for (Function::iterator I = F.begin(); I != F.end(); I++)
229 EverMadeChange |= bypassSlowDivision(F, I, BypassWidths);
232 // Eliminate blocks that contain only PHI nodes and an
233 // unconditional branch.
234 EverMadeChange |= EliminateMostlyEmptyBlocks(F);
236 // llvm.dbg.value is far away from the value then iSel may not be able
237 // handle it properly. iSel will drop llvm.dbg.value if it can not
238 // find a node corresponding to the value.
239 EverMadeChange |= PlaceDbgValues(F);
241 // If there is a mask, compare against zero, and branch that can be combined
242 // into a single target instruction, push the mask and compare into branch
243 // users. Do this before OptimizeBlock -> OptimizeInst ->
244 // OptimizeCmpExpression, which perturbs the pattern being searched for.
245 if (!DisableBranchOpts) {
246 EverMadeChange |= sinkAndCmp(F);
247 EverMadeChange |= splitBranchCondition(F);
250 bool MadeChange = true;
253 for (Function::iterator I = F.begin(); I != F.end(); ) {
254 BasicBlock *BB = I++;
255 bool ModifiedDTOnIteration = false;
256 MadeChange |= OptimizeBlock(*BB, ModifiedDTOnIteration);
258 // Restart BB iteration if the dominator tree of the Function was changed
259 if (ModifiedDTOnIteration)
262 EverMadeChange |= MadeChange;
267 if (!DisableBranchOpts) {
269 SmallPtrSet<BasicBlock*, 8> WorkList;
270 for (BasicBlock &BB : F) {
271 SmallVector<BasicBlock *, 2> Successors(succ_begin(&BB), succ_end(&BB));
272 MadeChange |= ConstantFoldTerminator(&BB, true);
273 if (!MadeChange) continue;
275 for (SmallVectorImpl<BasicBlock*>::iterator
276 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
277 if (pred_begin(*II) == pred_end(*II))
278 WorkList.insert(*II);
281 // Delete the dead blocks and any of their dead successors.
282 MadeChange |= !WorkList.empty();
283 while (!WorkList.empty()) {
284 BasicBlock *BB = *WorkList.begin();
286 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
290 for (SmallVectorImpl<BasicBlock*>::iterator
291 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
292 if (pred_begin(*II) == pred_end(*II))
293 WorkList.insert(*II);
296 // Merge pairs of basic blocks with unconditional branches, connected by
298 if (EverMadeChange || MadeChange)
299 MadeChange |= EliminateFallThrough(F);
301 EverMadeChange |= MadeChange;
304 if (!DisableGCOpts) {
305 SmallVector<Instruction *, 2> Statepoints;
306 for (BasicBlock &BB : F)
307 for (Instruction &I : BB)
309 Statepoints.push_back(&I);
310 for (auto &I : Statepoints)
311 EverMadeChange |= simplifyOffsetableRelocate(*I);
314 return EverMadeChange;
317 /// EliminateFallThrough - Merge basic blocks which are connected
318 /// by a single edge, where one of the basic blocks has a single successor
319 /// pointing to the other basic block, which has a single predecessor.
320 bool CodeGenPrepare::EliminateFallThrough(Function &F) {
321 bool Changed = false;
322 // Scan all of the blocks in the function, except for the entry block.
323 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
324 BasicBlock *BB = I++;
325 // If the destination block has a single pred, then this is a trivial
326 // edge, just collapse it.
327 BasicBlock *SinglePred = BB->getSinglePredecessor();
329 // Don't merge if BB's address is taken.
330 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
332 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
333 if (Term && !Term->isConditional()) {
335 DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n");
336 // Remember if SinglePred was the entry block of the function.
337 // If so, we will need to move BB back to the entry position.
338 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
339 MergeBasicBlockIntoOnlyPred(BB, nullptr);
341 if (isEntry && BB != &BB->getParent()->getEntryBlock())
342 BB->moveBefore(&BB->getParent()->getEntryBlock());
344 // We have erased a block. Update the iterator.
351 /// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes,
352 /// debug info directives, and an unconditional branch. Passes before isel
353 /// (e.g. LSR/loopsimplify) often split edges in ways that are non-optimal for
354 /// isel. Start by eliminating these blocks so we can split them the way we
356 bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) {
357 bool MadeChange = false;
358 // Note that this intentionally skips the entry block.
359 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
360 BasicBlock *BB = I++;
362 // If this block doesn't end with an uncond branch, ignore it.
363 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
364 if (!BI || !BI->isUnconditional())
367 // If the instruction before the branch (skipping debug info) isn't a phi
368 // node, then other stuff is happening here.
369 BasicBlock::iterator BBI = BI;
370 if (BBI != BB->begin()) {
372 while (isa<DbgInfoIntrinsic>(BBI)) {
373 if (BBI == BB->begin())
377 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
381 // Do not break infinite loops.
382 BasicBlock *DestBB = BI->getSuccessor(0);
386 if (!CanMergeBlocks(BB, DestBB))
389 EliminateMostlyEmptyBlock(BB);
395 /// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a
396 /// single uncond branch between them, and BB contains no other non-phi
398 bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB,
399 const BasicBlock *DestBB) const {
400 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
401 // the successor. If there are more complex condition (e.g. preheaders),
402 // don't mess around with them.
403 BasicBlock::const_iterator BBI = BB->begin();
404 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
405 for (const User *U : PN->users()) {
406 const Instruction *UI = cast<Instruction>(U);
407 if (UI->getParent() != DestBB || !isa<PHINode>(UI))
409 // If User is inside DestBB block and it is a PHINode then check
410 // incoming value. If incoming value is not from BB then this is
411 // a complex condition (e.g. preheaders) we want to avoid here.
412 if (UI->getParent() == DestBB) {
413 if (const PHINode *UPN = dyn_cast<PHINode>(UI))
414 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
415 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
416 if (Insn && Insn->getParent() == BB &&
417 Insn->getParent() != UPN->getIncomingBlock(I))
424 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
425 // and DestBB may have conflicting incoming values for the block. If so, we
426 // can't merge the block.
427 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
428 if (!DestBBPN) return true; // no conflict.
430 // Collect the preds of BB.
431 SmallPtrSet<const BasicBlock*, 16> BBPreds;
432 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
433 // It is faster to get preds from a PHI than with pred_iterator.
434 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
435 BBPreds.insert(BBPN->getIncomingBlock(i));
437 BBPreds.insert(pred_begin(BB), pred_end(BB));
440 // Walk the preds of DestBB.
441 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
442 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
443 if (BBPreds.count(Pred)) { // Common predecessor?
444 BBI = DestBB->begin();
445 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
446 const Value *V1 = PN->getIncomingValueForBlock(Pred);
447 const Value *V2 = PN->getIncomingValueForBlock(BB);
449 // If V2 is a phi node in BB, look up what the mapped value will be.
450 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
451 if (V2PN->getParent() == BB)
452 V2 = V2PN->getIncomingValueForBlock(Pred);
454 // If there is a conflict, bail out.
455 if (V1 != V2) return false;
464 /// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and
465 /// an unconditional branch in it.
466 void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) {
467 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
468 BasicBlock *DestBB = BI->getSuccessor(0);
470 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
472 // If the destination block has a single pred, then this is a trivial edge,
474 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
475 if (SinglePred != DestBB) {
476 // Remember if SinglePred was the entry block of the function. If so, we
477 // will need to move BB back to the entry position.
478 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
479 MergeBasicBlockIntoOnlyPred(DestBB, nullptr);
481 if (isEntry && BB != &BB->getParent()->getEntryBlock())
482 BB->moveBefore(&BB->getParent()->getEntryBlock());
484 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
489 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
490 // to handle the new incoming edges it is about to have.
492 for (BasicBlock::iterator BBI = DestBB->begin();
493 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
494 // Remove the incoming value for BB, and remember it.
495 Value *InVal = PN->removeIncomingValue(BB, false);
497 // Two options: either the InVal is a phi node defined in BB or it is some
498 // value that dominates BB.
499 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
500 if (InValPhi && InValPhi->getParent() == BB) {
501 // Add all of the input values of the input PHI as inputs of this phi.
502 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
503 PN->addIncoming(InValPhi->getIncomingValue(i),
504 InValPhi->getIncomingBlock(i));
506 // Otherwise, add one instance of the dominating value for each edge that
507 // we will be adding.
508 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
509 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
510 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
512 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
513 PN->addIncoming(InVal, *PI);
518 // The PHIs are now updated, change everything that refers to BB to use
519 // DestBB and remove BB.
520 BB->replaceAllUsesWith(DestBB);
521 BB->eraseFromParent();
524 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
527 // Computes a map of base pointer relocation instructions to corresponding
528 // derived pointer relocation instructions given a vector of all relocate calls
529 static void computeBaseDerivedRelocateMap(
530 const SmallVectorImpl<User *> &AllRelocateCalls,
531 DenseMap<IntrinsicInst *, SmallVector<IntrinsicInst *, 2>> &
533 // Collect information in two maps: one primarily for locating the base object
534 // while filling the second map; the second map is the final structure holding
535 // a mapping between Base and corresponding Derived relocate calls
536 DenseMap<std::pair<unsigned, unsigned>, IntrinsicInst *> RelocateIdxMap;
537 for (auto &U : AllRelocateCalls) {
538 GCRelocateOperands ThisRelocate(U);
539 IntrinsicInst *I = cast<IntrinsicInst>(U);
540 auto K = std::make_pair(ThisRelocate.getBasePtrIndex(),
541 ThisRelocate.getDerivedPtrIndex());
542 RelocateIdxMap.insert(std::make_pair(K, I));
544 for (auto &Item : RelocateIdxMap) {
545 std::pair<unsigned, unsigned> Key = Item.first;
546 if (Key.first == Key.second)
547 // Base relocation: nothing to insert
550 IntrinsicInst *I = Item.second;
551 auto BaseKey = std::make_pair(Key.first, Key.first);
553 // We're iterating over RelocateIdxMap so we cannot modify it.
554 auto MaybeBase = RelocateIdxMap.find(BaseKey);
555 if (MaybeBase == RelocateIdxMap.end())
556 // TODO: We might want to insert a new base object relocate and gep off
557 // that, if there are enough derived object relocates.
560 RelocateInstMap[MaybeBase->second].push_back(I);
564 // Accepts a GEP and extracts the operands into a vector provided they're all
565 // small integer constants
566 static bool getGEPSmallConstantIntOffsetV(GetElementPtrInst *GEP,
567 SmallVectorImpl<Value *> &OffsetV) {
568 for (unsigned i = 1; i < GEP->getNumOperands(); i++) {
569 // Only accept small constant integer operands
570 auto Op = dyn_cast<ConstantInt>(GEP->getOperand(i));
571 if (!Op || Op->getZExtValue() > 20)
575 for (unsigned i = 1; i < GEP->getNumOperands(); i++)
576 OffsetV.push_back(GEP->getOperand(i));
580 // Takes a RelocatedBase (base pointer relocation instruction) and Targets to
581 // replace, computes a replacement, and affects it.
583 simplifyRelocatesOffABase(IntrinsicInst *RelocatedBase,
584 const SmallVectorImpl<IntrinsicInst *> &Targets) {
585 bool MadeChange = false;
586 for (auto &ToReplace : Targets) {
587 GCRelocateOperands MasterRelocate(RelocatedBase);
588 GCRelocateOperands ThisRelocate(ToReplace);
590 assert(ThisRelocate.getBasePtrIndex() == MasterRelocate.getBasePtrIndex() &&
591 "Not relocating a derived object of the original base object");
592 if (ThisRelocate.getBasePtrIndex() == ThisRelocate.getDerivedPtrIndex()) {
593 // A duplicate relocate call. TODO: coalesce duplicates.
597 Value *Base = ThisRelocate.getBasePtr();
598 auto Derived = dyn_cast<GetElementPtrInst>(ThisRelocate.getDerivedPtr());
599 if (!Derived || Derived->getPointerOperand() != Base)
602 SmallVector<Value *, 2> OffsetV;
603 if (!getGEPSmallConstantIntOffsetV(Derived, OffsetV))
606 // Create a Builder and replace the target callsite with a gep
607 assert(RelocatedBase->getNextNode() && "Should always have one since it's not a terminator");
609 // Insert after RelocatedBase
610 IRBuilder<> Builder(RelocatedBase->getNextNode());
611 Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
613 // If gc_relocate does not match the actual type, cast it to the right type.
614 // In theory, there must be a bitcast after gc_relocate if the type does not
615 // match, and we should reuse it to get the derived pointer. But it could be
619 // %g1 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
624 // %g2 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
628 // %p1 = phi i8 addrspace(1)* [ %g1, %bb1 ], [ %g2, %bb2 ]
629 // %cast = bitcast i8 addrspace(1)* %p1 in to i32 addrspace(1)*
631 // In this case, we can not find the bitcast any more. So we insert a new bitcast
632 // no matter there is already one or not. In this way, we can handle all cases, and
633 // the extra bitcast should be optimized away in later passes.
634 Instruction *ActualRelocatedBase = RelocatedBase;
635 if (RelocatedBase->getType() != Base->getType()) {
636 ActualRelocatedBase =
637 cast<Instruction>(Builder.CreateBitCast(RelocatedBase, Base->getType()));
639 Value *Replacement = Builder.CreateGEP(
640 Derived->getSourceElementType(), ActualRelocatedBase, makeArrayRef(OffsetV));
641 Instruction *ReplacementInst = cast<Instruction>(Replacement);
642 Replacement->takeName(ToReplace);
643 // If the newly generated derived pointer's type does not match the original derived
644 // pointer's type, cast the new derived pointer to match it. Same reasoning as above.
645 Instruction *ActualReplacement = ReplacementInst;
646 if (ReplacementInst->getType() != ToReplace->getType()) {
648 cast<Instruction>(Builder.CreateBitCast(ReplacementInst, ToReplace->getType()));
650 ToReplace->replaceAllUsesWith(ActualReplacement);
651 ToReplace->eraseFromParent();
661 // %ptr = gep %base + 15
662 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
663 // %base' = relocate(%tok, i32 4, i32 4)
664 // %ptr' = relocate(%tok, i32 4, i32 5)
670 // %ptr = gep %base + 15
671 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
672 // %base' = gc.relocate(%tok, i32 4, i32 4)
673 // %ptr' = gep %base' + 15
675 bool CodeGenPrepare::simplifyOffsetableRelocate(Instruction &I) {
676 bool MadeChange = false;
677 SmallVector<User *, 2> AllRelocateCalls;
679 for (auto *U : I.users())
680 if (isGCRelocate(dyn_cast<Instruction>(U)))
681 // Collect all the relocate calls associated with a statepoint
682 AllRelocateCalls.push_back(U);
684 // We need atleast one base pointer relocation + one derived pointer
685 // relocation to mangle
686 if (AllRelocateCalls.size() < 2)
689 // RelocateInstMap is a mapping from the base relocate instruction to the
690 // corresponding derived relocate instructions
691 DenseMap<IntrinsicInst *, SmallVector<IntrinsicInst *, 2>> RelocateInstMap;
692 computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap);
693 if (RelocateInstMap.empty())
696 for (auto &Item : RelocateInstMap)
697 // Item.first is the RelocatedBase to offset against
698 // Item.second is the vector of Targets to replace
699 MadeChange = simplifyRelocatesOffABase(Item.first, Item.second);
703 /// SinkCast - Sink the specified cast instruction into its user blocks
704 static bool SinkCast(CastInst *CI) {
705 BasicBlock *DefBB = CI->getParent();
707 /// InsertedCasts - Only insert a cast in each block once.
708 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
710 bool MadeChange = false;
711 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
713 Use &TheUse = UI.getUse();
714 Instruction *User = cast<Instruction>(*UI);
716 // Figure out which BB this cast is used in. For PHI's this is the
717 // appropriate predecessor block.
718 BasicBlock *UserBB = User->getParent();
719 if (PHINode *PN = dyn_cast<PHINode>(User)) {
720 UserBB = PN->getIncomingBlock(TheUse);
723 // Preincrement use iterator so we don't invalidate it.
726 // If this user is in the same block as the cast, don't change the cast.
727 if (UserBB == DefBB) continue;
729 // If we have already inserted a cast into this block, use it.
730 CastInst *&InsertedCast = InsertedCasts[UserBB];
733 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
735 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "",
739 // Replace a use of the cast with a use of the new cast.
740 TheUse = InsertedCast;
745 // If we removed all uses, nuke the cast.
746 if (CI->use_empty()) {
747 CI->eraseFromParent();
754 /// OptimizeNoopCopyExpression - If the specified cast instruction is a noop
755 /// copy (e.g. it's casting from one pointer type to another, i32->i8 on PPC),
756 /// sink it into user blocks to reduce the number of virtual
757 /// registers that must be created and coalesced.
759 /// Return true if any changes are made.
761 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){
762 // If this is a noop copy,
763 EVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType());
764 EVT DstVT = TLI.getValueType(CI->getType());
766 // This is an fp<->int conversion?
767 if (SrcVT.isInteger() != DstVT.isInteger())
770 // If this is an extension, it will be a zero or sign extension, which
772 if (SrcVT.bitsLT(DstVT)) return false;
774 // If these values will be promoted, find out what they will be promoted
775 // to. This helps us consider truncates on PPC as noop copies when they
777 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
778 TargetLowering::TypePromoteInteger)
779 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
780 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
781 TargetLowering::TypePromoteInteger)
782 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
784 // If, after promotion, these are the same types, this is a noop copy.
791 /// CombineUAddWithOverflow - try to combine CI into a call to the
792 /// llvm.uadd.with.overflow intrinsic if possible.
794 /// Return true if any changes were made.
795 static bool CombineUAddWithOverflow(CmpInst *CI) {
799 m_UAddWithOverflow(m_Value(A), m_Value(B), m_Instruction(AddI))))
802 Type *Ty = AddI->getType();
803 if (!isa<IntegerType>(Ty))
806 // We don't want to move around uses of condition values this late, so we we
807 // check if it is legal to create the call to the intrinsic in the basic
808 // block containing the icmp:
810 if (AddI->getParent() != CI->getParent() && !AddI->hasOneUse())
814 // Someday m_UAddWithOverflow may get smarter, but this is a safe assumption
816 if (AddI->hasOneUse())
817 assert(*AddI->user_begin() == CI && "expected!");
820 Module *M = CI->getParent()->getParent()->getParent();
821 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
823 auto *InsertPt = AddI->hasOneUse() ? CI : AddI;
825 auto *UAddWithOverflow =
826 CallInst::Create(F, {A, B}, "uadd.overflow", InsertPt);
827 auto *UAdd = ExtractValueInst::Create(UAddWithOverflow, 0, "uadd", InsertPt);
829 ExtractValueInst::Create(UAddWithOverflow, 1, "overflow", InsertPt);
831 CI->replaceAllUsesWith(Overflow);
832 AddI->replaceAllUsesWith(UAdd);
833 CI->eraseFromParent();
834 AddI->eraseFromParent();
838 /// SinkCmpExpression - Sink the given CmpInst into user blocks to reduce
839 /// the number of virtual registers that must be created and coalesced. This is
840 /// a clear win except on targets with multiple condition code registers
841 /// (PowerPC), where it might lose; some adjustment may be wanted there.
843 /// Return true if any changes are made.
844 static bool SinkCmpExpression(CmpInst *CI) {
845 BasicBlock *DefBB = CI->getParent();
847 /// InsertedCmp - Only insert a cmp in each block once.
848 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
850 bool MadeChange = false;
851 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
853 Use &TheUse = UI.getUse();
854 Instruction *User = cast<Instruction>(*UI);
856 // Preincrement use iterator so we don't invalidate it.
859 // Don't bother for PHI nodes.
860 if (isa<PHINode>(User))
863 // Figure out which BB this cmp is used in.
864 BasicBlock *UserBB = User->getParent();
866 // If this user is in the same block as the cmp, don't change the cmp.
867 if (UserBB == DefBB) continue;
869 // If we have already inserted a cmp into this block, use it.
870 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
873 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
875 CmpInst::Create(CI->getOpcode(),
876 CI->getPredicate(), CI->getOperand(0),
877 CI->getOperand(1), "", InsertPt);
880 // Replace a use of the cmp with a use of the new cmp.
881 TheUse = InsertedCmp;
886 // If we removed all uses, nuke the cmp.
887 if (CI->use_empty()) {
888 CI->eraseFromParent();
895 static bool OptimizeCmpExpression(CmpInst *CI) {
896 if (SinkCmpExpression(CI))
899 if (CombineUAddWithOverflow(CI))
905 /// isExtractBitsCandidateUse - Check if the candidates could
906 /// be combined with shift instruction, which includes:
907 /// 1. Truncate instruction
908 /// 2. And instruction and the imm is a mask of the low bits:
909 /// imm & (imm+1) == 0
910 static bool isExtractBitsCandidateUse(Instruction *User) {
911 if (!isa<TruncInst>(User)) {
912 if (User->getOpcode() != Instruction::And ||
913 !isa<ConstantInt>(User->getOperand(1)))
916 const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
918 if ((Cimm & (Cimm + 1)).getBoolValue())
924 /// SinkShiftAndTruncate - sink both shift and truncate instruction
925 /// to the use of truncate's BB.
927 SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI,
928 DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts,
929 const TargetLowering &TLI) {
930 BasicBlock *UserBB = User->getParent();
931 DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
932 TruncInst *TruncI = dyn_cast<TruncInst>(User);
933 bool MadeChange = false;
935 for (Value::user_iterator TruncUI = TruncI->user_begin(),
936 TruncE = TruncI->user_end();
937 TruncUI != TruncE;) {
939 Use &TruncTheUse = TruncUI.getUse();
940 Instruction *TruncUser = cast<Instruction>(*TruncUI);
941 // Preincrement use iterator so we don't invalidate it.
945 int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
949 // If the use is actually a legal node, there will not be an
950 // implicit truncate.
951 // FIXME: always querying the result type is just an
952 // approximation; some nodes' legality is determined by the
953 // operand or other means. There's no good way to find out though.
954 if (TLI.isOperationLegalOrCustom(
955 ISDOpcode, TLI.getValueType(TruncUser->getType(), true)))
958 // Don't bother for PHI nodes.
959 if (isa<PHINode>(TruncUser))
962 BasicBlock *TruncUserBB = TruncUser->getParent();
964 if (UserBB == TruncUserBB)
967 BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
968 CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
970 if (!InsertedShift && !InsertedTrunc) {
971 BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
973 if (ShiftI->getOpcode() == Instruction::AShr)
975 BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI, "", InsertPt);
978 BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI, "", InsertPt);
981 BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
984 InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
985 TruncI->getType(), "", TruncInsertPt);
989 TruncTheUse = InsertedTrunc;
995 /// OptimizeExtractBits - sink the shift *right* instruction into user blocks if
996 /// the uses could potentially be combined with this shift instruction and
997 /// generate BitExtract instruction. It will only be applied if the architecture
998 /// supports BitExtract instruction. Here is an example:
1000 /// %x.extract.shift = lshr i64 %arg1, 32
1002 /// %x.extract.trunc = trunc i64 %x.extract.shift to i16
1006 /// %x.extract.shift.1 = lshr i64 %arg1, 32
1007 /// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
1009 /// CodeGen will recoginze the pattern in BB2 and generate BitExtract
1011 /// Return true if any changes are made.
1012 static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI,
1013 const TargetLowering &TLI) {
1014 BasicBlock *DefBB = ShiftI->getParent();
1016 /// Only insert instructions in each block once.
1017 DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts;
1019 bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(ShiftI->getType()));
1021 bool MadeChange = false;
1022 for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
1024 Use &TheUse = UI.getUse();
1025 Instruction *User = cast<Instruction>(*UI);
1026 // Preincrement use iterator so we don't invalidate it.
1029 // Don't bother for PHI nodes.
1030 if (isa<PHINode>(User))
1033 if (!isExtractBitsCandidateUse(User))
1036 BasicBlock *UserBB = User->getParent();
1038 if (UserBB == DefBB) {
1039 // If the shift and truncate instruction are in the same BB. The use of
1040 // the truncate(TruncUse) may still introduce another truncate if not
1041 // legal. In this case, we would like to sink both shift and truncate
1042 // instruction to the BB of TruncUse.
1045 // i64 shift.result = lshr i64 opnd, imm
1046 // trunc.result = trunc shift.result to i16
1049 // ----> We will have an implicit truncate here if the architecture does
1050 // not have i16 compare.
1051 // cmp i16 trunc.result, opnd2
1053 if (isa<TruncInst>(User) && shiftIsLegal
1054 // If the type of the truncate is legal, no trucate will be
1055 // introduced in other basic blocks.
1056 && (!TLI.isTypeLegal(TLI.getValueType(User->getType()))))
1058 SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI);
1062 // If we have already inserted a shift into this block, use it.
1063 BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
1065 if (!InsertedShift) {
1066 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1068 if (ShiftI->getOpcode() == Instruction::AShr)
1070 BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI, "", InsertPt);
1073 BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI, "", InsertPt);
1078 // Replace a use of the shift with a use of the new shift.
1079 TheUse = InsertedShift;
1082 // If we removed all uses, nuke the shift.
1083 if (ShiftI->use_empty())
1084 ShiftI->eraseFromParent();
1089 // ScalarizeMaskedLoad() translates masked load intrinsic, like
1090 // <16 x i32 > @llvm.masked.load( <16 x i32>* %addr, i32 align,
1091 // <16 x i1> %mask, <16 x i32> %passthru)
1092 // to a chain of basic blocks, whith loading element one-by-one if
1093 // the appropriate mask bit is set
1095 // %1 = bitcast i8* %addr to i32*
1096 // %2 = extractelement <16 x i1> %mask, i32 0
1097 // %3 = icmp eq i1 %2, true
1098 // br i1 %3, label %cond.load, label %else
1100 //cond.load: ; preds = %0
1101 // %4 = getelementptr i32* %1, i32 0
1102 // %5 = load i32* %4
1103 // %6 = insertelement <16 x i32> undef, i32 %5, i32 0
1106 //else: ; preds = %0, %cond.load
1107 // %res.phi.else = phi <16 x i32> [ %6, %cond.load ], [ undef, %0 ]
1108 // %7 = extractelement <16 x i1> %mask, i32 1
1109 // %8 = icmp eq i1 %7, true
1110 // br i1 %8, label %cond.load1, label %else2
1112 //cond.load1: ; preds = %else
1113 // %9 = getelementptr i32* %1, i32 1
1114 // %10 = load i32* %9
1115 // %11 = insertelement <16 x i32> %res.phi.else, i32 %10, i32 1
1118 //else2: ; preds = %else, %cond.load1
1119 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
1120 // %12 = extractelement <16 x i1> %mask, i32 2
1121 // %13 = icmp eq i1 %12, true
1122 // br i1 %13, label %cond.load4, label %else5
1124 static void ScalarizeMaskedLoad(CallInst *CI) {
1125 Value *Ptr = CI->getArgOperand(0);
1126 Value *Src0 = CI->getArgOperand(3);
1127 Value *Mask = CI->getArgOperand(2);
1128 VectorType *VecType = dyn_cast<VectorType>(CI->getType());
1129 Type *EltTy = VecType->getElementType();
1131 assert(VecType && "Unexpected return type of masked load intrinsic");
1133 IRBuilder<> Builder(CI->getContext());
1134 Instruction *InsertPt = CI;
1135 BasicBlock *IfBlock = CI->getParent();
1136 BasicBlock *CondBlock = nullptr;
1137 BasicBlock *PrevIfBlock = CI->getParent();
1138 Builder.SetInsertPoint(InsertPt);
1140 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1142 // Bitcast %addr fron i8* to EltTy*
1144 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
1145 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
1146 Value *UndefVal = UndefValue::get(VecType);
1148 // The result vector
1149 Value *VResult = UndefVal;
1151 PHINode *Phi = nullptr;
1152 Value *PrevPhi = UndefVal;
1154 unsigned VectorWidth = VecType->getNumElements();
1155 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1157 // Fill the "else" block, created in the previous iteration
1159 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
1160 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
1161 // %to_load = icmp eq i1 %mask_1, true
1162 // br i1 %to_load, label %cond.load, label %else
1165 Phi = Builder.CreatePHI(VecType, 2, "res.phi.else");
1166 Phi->addIncoming(VResult, CondBlock);
1167 Phi->addIncoming(PrevPhi, PrevIfBlock);
1172 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
1173 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1174 ConstantInt::get(Predicate->getType(), 1));
1176 // Create "cond" block
1178 // %EltAddr = getelementptr i32* %1, i32 0
1179 // %Elt = load i32* %EltAddr
1180 // VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx
1182 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.load");
1183 Builder.SetInsertPoint(InsertPt);
1186 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1187 LoadInst* Load = Builder.CreateLoad(Gep, false);
1188 VResult = Builder.CreateInsertElement(VResult, Load, Builder.getInt32(Idx));
1190 // Create "else" block, fill it in the next iteration
1191 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1192 Builder.SetInsertPoint(InsertPt);
1193 Instruction *OldBr = IfBlock->getTerminator();
1194 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1195 OldBr->eraseFromParent();
1196 PrevIfBlock = IfBlock;
1197 IfBlock = NewIfBlock;
1200 Phi = Builder.CreatePHI(VecType, 2, "res.phi.select");
1201 Phi->addIncoming(VResult, CondBlock);
1202 Phi->addIncoming(PrevPhi, PrevIfBlock);
1203 Value *NewI = Builder.CreateSelect(Mask, Phi, Src0);
1204 CI->replaceAllUsesWith(NewI);
1205 CI->eraseFromParent();
1208 // ScalarizeMaskedStore() translates masked store intrinsic, like
1209 // void @llvm.masked.store(<16 x i32> %src, <16 x i32>* %addr, i32 align,
1211 // to a chain of basic blocks, that stores element one-by-one if
1212 // the appropriate mask bit is set
1214 // %1 = bitcast i8* %addr to i32*
1215 // %2 = extractelement <16 x i1> %mask, i32 0
1216 // %3 = icmp eq i1 %2, true
1217 // br i1 %3, label %cond.store, label %else
1219 // cond.store: ; preds = %0
1220 // %4 = extractelement <16 x i32> %val, i32 0
1221 // %5 = getelementptr i32* %1, i32 0
1222 // store i32 %4, i32* %5
1225 // else: ; preds = %0, %cond.store
1226 // %6 = extractelement <16 x i1> %mask, i32 1
1227 // %7 = icmp eq i1 %6, true
1228 // br i1 %7, label %cond.store1, label %else2
1230 // cond.store1: ; preds = %else
1231 // %8 = extractelement <16 x i32> %val, i32 1
1232 // %9 = getelementptr i32* %1, i32 1
1233 // store i32 %8, i32* %9
1236 static void ScalarizeMaskedStore(CallInst *CI) {
1237 Value *Ptr = CI->getArgOperand(1);
1238 Value *Src = CI->getArgOperand(0);
1239 Value *Mask = CI->getArgOperand(3);
1241 VectorType *VecType = dyn_cast<VectorType>(Src->getType());
1242 Type *EltTy = VecType->getElementType();
1244 assert(VecType && "Unexpected data type in masked store intrinsic");
1246 IRBuilder<> Builder(CI->getContext());
1247 Instruction *InsertPt = CI;
1248 BasicBlock *IfBlock = CI->getParent();
1249 Builder.SetInsertPoint(InsertPt);
1250 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1252 // Bitcast %addr fron i8* to EltTy*
1254 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
1255 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
1257 unsigned VectorWidth = VecType->getNumElements();
1258 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1260 // Fill the "else" block, created in the previous iteration
1262 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
1263 // %to_store = icmp eq i1 %mask_1, true
1264 // br i1 %to_load, label %cond.store, label %else
1266 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
1267 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1268 ConstantInt::get(Predicate->getType(), 1));
1270 // Create "cond" block
1272 // %OneElt = extractelement <16 x i32> %Src, i32 Idx
1273 // %EltAddr = getelementptr i32* %1, i32 0
1274 // %store i32 %OneElt, i32* %EltAddr
1276 BasicBlock *CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1277 Builder.SetInsertPoint(InsertPt);
1279 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx));
1281 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1282 Builder.CreateStore(OneElt, Gep);
1284 // Create "else" block, fill it in the next iteration
1285 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1286 Builder.SetInsertPoint(InsertPt);
1287 Instruction *OldBr = IfBlock->getTerminator();
1288 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1289 OldBr->eraseFromParent();
1290 IfBlock = NewIfBlock;
1292 CI->eraseFromParent();
1295 bool CodeGenPrepare::OptimizeCallInst(CallInst *CI, bool& ModifiedDT) {
1296 BasicBlock *BB = CI->getParent();
1298 // Lower inline assembly if we can.
1299 // If we found an inline asm expession, and if the target knows how to
1300 // lower it to normal LLVM code, do so now.
1301 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
1302 if (TLI->ExpandInlineAsm(CI)) {
1303 // Avoid invalidating the iterator.
1304 CurInstIterator = BB->begin();
1305 // Avoid processing instructions out of order, which could cause
1306 // reuse before a value is defined.
1310 // Sink address computing for memory operands into the block.
1311 if (OptimizeInlineAsmInst(CI))
1315 // Align the pointer arguments to this call if the target thinks it's a good
1317 unsigned MinSize, PrefAlign;
1318 if (TLI && TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) {
1319 for (auto &Arg : CI->arg_operands()) {
1320 // We want to align both objects whose address is used directly and
1321 // objects whose address is used in casts and GEPs, though it only makes
1322 // sense for GEPs if the offset is a multiple of the desired alignment and
1323 // if size - offset meets the size threshold.
1324 if (!Arg->getType()->isPointerTy())
1326 APInt Offset(DL->getPointerSizeInBits(
1327 cast<PointerType>(Arg->getType())->getAddressSpace()),
1329 Value *Val = Arg->stripAndAccumulateInBoundsConstantOffsets(*DL, Offset);
1330 uint64_t Offset2 = Offset.getLimitedValue();
1331 if ((Offset2 & (PrefAlign-1)) != 0)
1334 if ((AI = dyn_cast<AllocaInst>(Val)) && AI->getAlignment() < PrefAlign &&
1335 DL->getTypeAllocSize(AI->getAllocatedType()) >= MinSize + Offset2)
1336 AI->setAlignment(PrefAlign);
1337 // Global variables can only be aligned if they are defined in this
1338 // object (i.e. they are uniquely initialized in this object), and
1339 // over-aligning global variables that have an explicit section is
1342 if ((GV = dyn_cast<GlobalVariable>(Val)) && GV->hasUniqueInitializer() &&
1343 !GV->hasSection() && GV->getAlignment() < PrefAlign &&
1344 DL->getTypeAllocSize(GV->getType()->getElementType()) >=
1346 GV->setAlignment(PrefAlign);
1348 // If this is a memcpy (or similar) then we may be able to improve the
1350 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(CI)) {
1351 unsigned Align = getKnownAlignment(MI->getDest(), *DL);
1352 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI))
1353 Align = std::min(Align, getKnownAlignment(MTI->getSource(), *DL));
1354 if (Align > MI->getAlignment())
1355 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), Align));
1359 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
1361 switch (II->getIntrinsicID()) {
1363 case Intrinsic::objectsize: {
1364 // Lower all uses of llvm.objectsize.*
1365 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
1366 Type *ReturnTy = CI->getType();
1367 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
1369 // Substituting this can cause recursive simplifications, which can
1370 // invalidate our iterator. Use a WeakVH to hold onto it in case this
1372 WeakVH IterHandle(CurInstIterator);
1374 replaceAndRecursivelySimplify(CI, RetVal,
1377 // If the iterator instruction was recursively deleted, start over at the
1378 // start of the block.
1379 if (IterHandle != CurInstIterator) {
1380 CurInstIterator = BB->begin();
1385 case Intrinsic::masked_load: {
1386 // Scalarize unsupported vector masked load
1387 if (!TTI->isLegalMaskedLoad(CI->getType(), 1)) {
1388 ScalarizeMaskedLoad(CI);
1394 case Intrinsic::masked_store: {
1395 if (!TTI->isLegalMaskedStore(CI->getArgOperand(0)->getType(), 1)) {
1396 ScalarizeMaskedStore(CI);
1402 case Intrinsic::aarch64_stlxr:
1403 case Intrinsic::aarch64_stxr: {
1404 ZExtInst *ExtVal = dyn_cast<ZExtInst>(CI->getArgOperand(0));
1405 if (!ExtVal || !ExtVal->hasOneUse() ||
1406 ExtVal->getParent() == CI->getParent())
1408 // Sink a zext feeding stlxr/stxr before it, so it can be folded into it.
1409 ExtVal->moveBefore(CI);
1410 // Mark this instruction as "inserted by CGP", so that other
1411 // optimizations don't touch it.
1412 InsertedInsts.insert(ExtVal);
1418 // Unknown address space.
1419 // TODO: Target hook to pick which address space the intrinsic cares
1421 unsigned AddrSpace = ~0u;
1422 SmallVector<Value*, 2> PtrOps;
1424 if (TLI->GetAddrModeArguments(II, PtrOps, AccessTy, AddrSpace))
1425 while (!PtrOps.empty())
1426 if (OptimizeMemoryInst(II, PtrOps.pop_back_val(), AccessTy, AddrSpace))
1431 // From here on out we're working with named functions.
1432 if (!CI->getCalledFunction()) return false;
1434 // Lower all default uses of _chk calls. This is very similar
1435 // to what InstCombineCalls does, but here we are only lowering calls
1436 // to fortified library functions (e.g. __memcpy_chk) that have the default
1437 // "don't know" as the objectsize. Anything else should be left alone.
1438 FortifiedLibCallSimplifier Simplifier(TLInfo, true);
1439 if (Value *V = Simplifier.optimizeCall(CI)) {
1440 CI->replaceAllUsesWith(V);
1441 CI->eraseFromParent();
1447 /// DupRetToEnableTailCallOpts - Look for opportunities to duplicate return
1448 /// instructions to the predecessor to enable tail call optimizations. The
1449 /// case it is currently looking for is:
1452 /// %tmp0 = tail call i32 @f0()
1453 /// br label %return
1455 /// %tmp1 = tail call i32 @f1()
1456 /// br label %return
1458 /// %tmp2 = tail call i32 @f2()
1459 /// br label %return
1461 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
1469 /// %tmp0 = tail call i32 @f0()
1472 /// %tmp1 = tail call i32 @f1()
1475 /// %tmp2 = tail call i32 @f2()
1478 bool CodeGenPrepare::DupRetToEnableTailCallOpts(BasicBlock *BB) {
1482 ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator());
1486 PHINode *PN = nullptr;
1487 BitCastInst *BCI = nullptr;
1488 Value *V = RI->getReturnValue();
1490 BCI = dyn_cast<BitCastInst>(V);
1492 V = BCI->getOperand(0);
1494 PN = dyn_cast<PHINode>(V);
1499 if (PN && PN->getParent() != BB)
1502 // It's not safe to eliminate the sign / zero extension of the return value.
1503 // See llvm::isInTailCallPosition().
1504 const Function *F = BB->getParent();
1505 AttributeSet CallerAttrs = F->getAttributes();
1506 if (CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt) ||
1507 CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
1510 // Make sure there are no instructions between the PHI and return, or that the
1511 // return is the first instruction in the block.
1513 BasicBlock::iterator BI = BB->begin();
1514 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
1516 // Also skip over the bitcast.
1521 BasicBlock::iterator BI = BB->begin();
1522 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
1527 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
1529 SmallVector<CallInst*, 4> TailCalls;
1531 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
1532 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
1533 // Make sure the phi value is indeed produced by the tail call.
1534 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
1535 TLI->mayBeEmittedAsTailCall(CI))
1536 TailCalls.push_back(CI);
1539 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
1540 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
1541 if (!VisitedBBs.insert(*PI).second)
1544 BasicBlock::InstListType &InstList = (*PI)->getInstList();
1545 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
1546 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
1547 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
1551 CallInst *CI = dyn_cast<CallInst>(&*RI);
1552 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
1553 TailCalls.push_back(CI);
1557 bool Changed = false;
1558 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
1559 CallInst *CI = TailCalls[i];
1562 // Conservatively require the attributes of the call to match those of the
1563 // return. Ignore noalias because it doesn't affect the call sequence.
1564 AttributeSet CalleeAttrs = CS.getAttributes();
1565 if (AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1566 removeAttribute(Attribute::NoAlias) !=
1567 AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1568 removeAttribute(Attribute::NoAlias))
1571 // Make sure the call instruction is followed by an unconditional branch to
1572 // the return block.
1573 BasicBlock *CallBB = CI->getParent();
1574 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
1575 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
1578 // Duplicate the return into CallBB.
1579 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
1580 ModifiedDT = Changed = true;
1584 // If we eliminated all predecessors of the block, delete the block now.
1585 if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
1586 BB->eraseFromParent();
1591 //===----------------------------------------------------------------------===//
1592 // Memory Optimization
1593 //===----------------------------------------------------------------------===//
1597 /// ExtAddrMode - This is an extended version of TargetLowering::AddrMode
1598 /// which holds actual Value*'s for register values.
1599 struct ExtAddrMode : public TargetLowering::AddrMode {
1602 ExtAddrMode() : BaseReg(nullptr), ScaledReg(nullptr) {}
1603 void print(raw_ostream &OS) const;
1606 bool operator==(const ExtAddrMode& O) const {
1607 return (BaseReg == O.BaseReg) && (ScaledReg == O.ScaledReg) &&
1608 (BaseGV == O.BaseGV) && (BaseOffs == O.BaseOffs) &&
1609 (HasBaseReg == O.HasBaseReg) && (Scale == O.Scale);
1614 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
1620 void ExtAddrMode::print(raw_ostream &OS) const {
1621 bool NeedPlus = false;
1624 OS << (NeedPlus ? " + " : "")
1626 BaseGV->printAsOperand(OS, /*PrintType=*/false);
1631 OS << (NeedPlus ? " + " : "")
1637 OS << (NeedPlus ? " + " : "")
1639 BaseReg->printAsOperand(OS, /*PrintType=*/false);
1643 OS << (NeedPlus ? " + " : "")
1645 ScaledReg->printAsOperand(OS, /*PrintType=*/false);
1651 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1652 void ExtAddrMode::dump() const {
1658 /// \brief This class provides transaction based operation on the IR.
1659 /// Every change made through this class is recorded in the internal state and
1660 /// can be undone (rollback) until commit is called.
1661 class TypePromotionTransaction {
1663 /// \brief This represents the common interface of the individual transaction.
1664 /// Each class implements the logic for doing one specific modification on
1665 /// the IR via the TypePromotionTransaction.
1666 class TypePromotionAction {
1668 /// The Instruction modified.
1672 /// \brief Constructor of the action.
1673 /// The constructor performs the related action on the IR.
1674 TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
1676 virtual ~TypePromotionAction() {}
1678 /// \brief Undo the modification done by this action.
1679 /// When this method is called, the IR must be in the same state as it was
1680 /// before this action was applied.
1681 /// \pre Undoing the action works if and only if the IR is in the exact same
1682 /// state as it was directly after this action was applied.
1683 virtual void undo() = 0;
1685 /// \brief Advocate every change made by this action.
1686 /// When the results on the IR of the action are to be kept, it is important
1687 /// to call this function, otherwise hidden information may be kept forever.
1688 virtual void commit() {
1689 // Nothing to be done, this action is not doing anything.
1693 /// \brief Utility to remember the position of an instruction.
1694 class InsertionHandler {
1695 /// Position of an instruction.
1696 /// Either an instruction:
1697 /// - Is the first in a basic block: BB is used.
1698 /// - Has a previous instructon: PrevInst is used.
1700 Instruction *PrevInst;
1703 /// Remember whether or not the instruction had a previous instruction.
1704 bool HasPrevInstruction;
1707 /// \brief Record the position of \p Inst.
1708 InsertionHandler(Instruction *Inst) {
1709 BasicBlock::iterator It = Inst;
1710 HasPrevInstruction = (It != (Inst->getParent()->begin()));
1711 if (HasPrevInstruction)
1712 Point.PrevInst = --It;
1714 Point.BB = Inst->getParent();
1717 /// \brief Insert \p Inst at the recorded position.
1718 void insert(Instruction *Inst) {
1719 if (HasPrevInstruction) {
1720 if (Inst->getParent())
1721 Inst->removeFromParent();
1722 Inst->insertAfter(Point.PrevInst);
1724 Instruction *Position = Point.BB->getFirstInsertionPt();
1725 if (Inst->getParent())
1726 Inst->moveBefore(Position);
1728 Inst->insertBefore(Position);
1733 /// \brief Move an instruction before another.
1734 class InstructionMoveBefore : public TypePromotionAction {
1735 /// Original position of the instruction.
1736 InsertionHandler Position;
1739 /// \brief Move \p Inst before \p Before.
1740 InstructionMoveBefore(Instruction *Inst, Instruction *Before)
1741 : TypePromotionAction(Inst), Position(Inst) {
1742 DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before << "\n");
1743 Inst->moveBefore(Before);
1746 /// \brief Move the instruction back to its original position.
1747 void undo() override {
1748 DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
1749 Position.insert(Inst);
1753 /// \brief Set the operand of an instruction with a new value.
1754 class OperandSetter : public TypePromotionAction {
1755 /// Original operand of the instruction.
1757 /// Index of the modified instruction.
1761 /// \brief Set \p Idx operand of \p Inst with \p NewVal.
1762 OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
1763 : TypePromotionAction(Inst), Idx(Idx) {
1764 DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
1765 << "for:" << *Inst << "\n"
1766 << "with:" << *NewVal << "\n");
1767 Origin = Inst->getOperand(Idx);
1768 Inst->setOperand(Idx, NewVal);
1771 /// \brief Restore the original value of the instruction.
1772 void undo() override {
1773 DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
1774 << "for: " << *Inst << "\n"
1775 << "with: " << *Origin << "\n");
1776 Inst->setOperand(Idx, Origin);
1780 /// \brief Hide the operands of an instruction.
1781 /// Do as if this instruction was not using any of its operands.
1782 class OperandsHider : public TypePromotionAction {
1783 /// The list of original operands.
1784 SmallVector<Value *, 4> OriginalValues;
1787 /// \brief Remove \p Inst from the uses of the operands of \p Inst.
1788 OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
1789 DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
1790 unsigned NumOpnds = Inst->getNumOperands();
1791 OriginalValues.reserve(NumOpnds);
1792 for (unsigned It = 0; It < NumOpnds; ++It) {
1793 // Save the current operand.
1794 Value *Val = Inst->getOperand(It);
1795 OriginalValues.push_back(Val);
1797 // We could use OperandSetter here, but that would implied an overhead
1798 // that we are not willing to pay.
1799 Inst->setOperand(It, UndefValue::get(Val->getType()));
1803 /// \brief Restore the original list of uses.
1804 void undo() override {
1805 DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
1806 for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
1807 Inst->setOperand(It, OriginalValues[It]);
1811 /// \brief Build a truncate instruction.
1812 class TruncBuilder : public TypePromotionAction {
1815 /// \brief Build a truncate instruction of \p Opnd producing a \p Ty
1817 /// trunc Opnd to Ty.
1818 TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
1819 IRBuilder<> Builder(Opnd);
1820 Val = Builder.CreateTrunc(Opnd, Ty, "promoted");
1821 DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n");
1824 /// \brief Get the built value.
1825 Value *getBuiltValue() { return Val; }
1827 /// \brief Remove the built instruction.
1828 void undo() override {
1829 DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n");
1830 if (Instruction *IVal = dyn_cast<Instruction>(Val))
1831 IVal->eraseFromParent();
1835 /// \brief Build a sign extension instruction.
1836 class SExtBuilder : public TypePromotionAction {
1839 /// \brief Build a sign extension instruction of \p Opnd producing a \p Ty
1841 /// sext Opnd to Ty.
1842 SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
1843 : TypePromotionAction(InsertPt) {
1844 IRBuilder<> Builder(InsertPt);
1845 Val = Builder.CreateSExt(Opnd, Ty, "promoted");
1846 DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n");
1849 /// \brief Get the built value.
1850 Value *getBuiltValue() { return Val; }
1852 /// \brief Remove the built instruction.
1853 void undo() override {
1854 DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n");
1855 if (Instruction *IVal = dyn_cast<Instruction>(Val))
1856 IVal->eraseFromParent();
1860 /// \brief Build a zero extension instruction.
1861 class ZExtBuilder : public TypePromotionAction {
1864 /// \brief Build a zero extension instruction of \p Opnd producing a \p Ty
1866 /// zext Opnd to Ty.
1867 ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
1868 : TypePromotionAction(InsertPt) {
1869 IRBuilder<> Builder(InsertPt);
1870 Val = Builder.CreateZExt(Opnd, Ty, "promoted");
1871 DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n");
1874 /// \brief Get the built value.
1875 Value *getBuiltValue() { return Val; }
1877 /// \brief Remove the built instruction.
1878 void undo() override {
1879 DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n");
1880 if (Instruction *IVal = dyn_cast<Instruction>(Val))
1881 IVal->eraseFromParent();
1885 /// \brief Mutate an instruction to another type.
1886 class TypeMutator : public TypePromotionAction {
1887 /// Record the original type.
1891 /// \brief Mutate the type of \p Inst into \p NewTy.
1892 TypeMutator(Instruction *Inst, Type *NewTy)
1893 : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
1894 DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
1896 Inst->mutateType(NewTy);
1899 /// \brief Mutate the instruction back to its original type.
1900 void undo() override {
1901 DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
1903 Inst->mutateType(OrigTy);
1907 /// \brief Replace the uses of an instruction by another instruction.
1908 class UsesReplacer : public TypePromotionAction {
1909 /// Helper structure to keep track of the replaced uses.
1910 struct InstructionAndIdx {
1911 /// The instruction using the instruction.
1913 /// The index where this instruction is used for Inst.
1915 InstructionAndIdx(Instruction *Inst, unsigned Idx)
1916 : Inst(Inst), Idx(Idx) {}
1919 /// Keep track of the original uses (pair Instruction, Index).
1920 SmallVector<InstructionAndIdx, 4> OriginalUses;
1921 typedef SmallVectorImpl<InstructionAndIdx>::iterator use_iterator;
1924 /// \brief Replace all the use of \p Inst by \p New.
1925 UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
1926 DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
1928 // Record the original uses.
1929 for (Use &U : Inst->uses()) {
1930 Instruction *UserI = cast<Instruction>(U.getUser());
1931 OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
1933 // Now, we can replace the uses.
1934 Inst->replaceAllUsesWith(New);
1937 /// \brief Reassign the original uses of Inst to Inst.
1938 void undo() override {
1939 DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
1940 for (use_iterator UseIt = OriginalUses.begin(),
1941 EndIt = OriginalUses.end();
1942 UseIt != EndIt; ++UseIt) {
1943 UseIt->Inst->setOperand(UseIt->Idx, Inst);
1948 /// \brief Remove an instruction from the IR.
1949 class InstructionRemover : public TypePromotionAction {
1950 /// Original position of the instruction.
1951 InsertionHandler Inserter;
1952 /// Helper structure to hide all the link to the instruction. In other
1953 /// words, this helps to do as if the instruction was removed.
1954 OperandsHider Hider;
1955 /// Keep track of the uses replaced, if any.
1956 UsesReplacer *Replacer;
1959 /// \brief Remove all reference of \p Inst and optinally replace all its
1961 /// \pre If !Inst->use_empty(), then New != nullptr
1962 InstructionRemover(Instruction *Inst, Value *New = nullptr)
1963 : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
1966 Replacer = new UsesReplacer(Inst, New);
1967 DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
1968 Inst->removeFromParent();
1971 ~InstructionRemover() override { delete Replacer; }
1973 /// \brief Really remove the instruction.
1974 void commit() override { delete Inst; }
1976 /// \brief Resurrect the instruction and reassign it to the proper uses if
1977 /// new value was provided when build this action.
1978 void undo() override {
1979 DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
1980 Inserter.insert(Inst);
1988 /// Restoration point.
1989 /// The restoration point is a pointer to an action instead of an iterator
1990 /// because the iterator may be invalidated but not the pointer.
1991 typedef const TypePromotionAction *ConstRestorationPt;
1992 /// Advocate every changes made in that transaction.
1994 /// Undo all the changes made after the given point.
1995 void rollback(ConstRestorationPt Point);
1996 /// Get the current restoration point.
1997 ConstRestorationPt getRestorationPoint() const;
1999 /// \name API for IR modification with state keeping to support rollback.
2001 /// Same as Instruction::setOperand.
2002 void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
2003 /// Same as Instruction::eraseFromParent.
2004 void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
2005 /// Same as Value::replaceAllUsesWith.
2006 void replaceAllUsesWith(Instruction *Inst, Value *New);
2007 /// Same as Value::mutateType.
2008 void mutateType(Instruction *Inst, Type *NewTy);
2009 /// Same as IRBuilder::createTrunc.
2010 Value *createTrunc(Instruction *Opnd, Type *Ty);
2011 /// Same as IRBuilder::createSExt.
2012 Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
2013 /// Same as IRBuilder::createZExt.
2014 Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
2015 /// Same as Instruction::moveBefore.
2016 void moveBefore(Instruction *Inst, Instruction *Before);
2020 /// The ordered list of actions made so far.
2021 SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions;
2022 typedef SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator CommitPt;
2025 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
2028 make_unique<TypePromotionTransaction::OperandSetter>(Inst, Idx, NewVal));
2031 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
2034 make_unique<TypePromotionTransaction::InstructionRemover>(Inst, NewVal));
2037 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
2039 Actions.push_back(make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
2042 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
2043 Actions.push_back(make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
2046 Value *TypePromotionTransaction::createTrunc(Instruction *Opnd,
2048 std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
2049 Value *Val = Ptr->getBuiltValue();
2050 Actions.push_back(std::move(Ptr));
2054 Value *TypePromotionTransaction::createSExt(Instruction *Inst,
2055 Value *Opnd, Type *Ty) {
2056 std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
2057 Value *Val = Ptr->getBuiltValue();
2058 Actions.push_back(std::move(Ptr));
2062 Value *TypePromotionTransaction::createZExt(Instruction *Inst,
2063 Value *Opnd, Type *Ty) {
2064 std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
2065 Value *Val = Ptr->getBuiltValue();
2066 Actions.push_back(std::move(Ptr));
2070 void TypePromotionTransaction::moveBefore(Instruction *Inst,
2071 Instruction *Before) {
2073 make_unique<TypePromotionTransaction::InstructionMoveBefore>(Inst, Before));
2076 TypePromotionTransaction::ConstRestorationPt
2077 TypePromotionTransaction::getRestorationPoint() const {
2078 return !Actions.empty() ? Actions.back().get() : nullptr;
2081 void TypePromotionTransaction::commit() {
2082 for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
2088 void TypePromotionTransaction::rollback(
2089 TypePromotionTransaction::ConstRestorationPt Point) {
2090 while (!Actions.empty() && Point != Actions.back().get()) {
2091 std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
2096 /// \brief A helper class for matching addressing modes.
2098 /// This encapsulates the logic for matching the target-legal addressing modes.
2099 class AddressingModeMatcher {
2100 SmallVectorImpl<Instruction*> &AddrModeInsts;
2101 const TargetMachine &TM;
2102 const TargetLowering &TLI;
2103 const DataLayout &DL;
2105 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
2106 /// the memory instruction that we're computing this address for.
2109 Instruction *MemoryInst;
2111 /// AddrMode - This is the addressing mode that we're building up. This is
2112 /// part of the return value of this addressing mode matching stuff.
2113 ExtAddrMode &AddrMode;
2115 /// The instructions inserted by other CodeGenPrepare optimizations.
2116 const SetOfInstrs &InsertedInsts;
2117 /// A map from the instructions to their type before promotion.
2118 InstrToOrigTy &PromotedInsts;
2119 /// The ongoing transaction where every action should be registered.
2120 TypePromotionTransaction &TPT;
2122 /// IgnoreProfitability - This is set to true when we should not do
2123 /// profitability checks. When true, IsProfitableToFoldIntoAddressingMode
2124 /// always returns true.
2125 bool IgnoreProfitability;
2127 AddressingModeMatcher(SmallVectorImpl<Instruction *> &AMI,
2128 const TargetMachine &TM, Type *AT, unsigned AS,
2129 Instruction *MI, ExtAddrMode &AM,
2130 const SetOfInstrs &InsertedInsts,
2131 InstrToOrigTy &PromotedInsts,
2132 TypePromotionTransaction &TPT)
2133 : AddrModeInsts(AMI), TM(TM),
2134 TLI(*TM.getSubtargetImpl(*MI->getParent()->getParent())
2135 ->getTargetLowering()),
2136 DL(MI->getModule()->getDataLayout()), AccessTy(AT), AddrSpace(AS),
2137 MemoryInst(MI), AddrMode(AM), InsertedInsts(InsertedInsts),
2138 PromotedInsts(PromotedInsts), TPT(TPT) {
2139 IgnoreProfitability = false;
2143 /// Match - Find the maximal addressing mode that a load/store of V can fold,
2144 /// give an access type of AccessTy. This returns a list of involved
2145 /// instructions in AddrModeInsts.
2146 /// \p InsertedInsts The instructions inserted by other CodeGenPrepare
2148 /// \p PromotedInsts maps the instructions to their type before promotion.
2149 /// \p The ongoing transaction where every action should be registered.
2150 static ExtAddrMode Match(Value *V, Type *AccessTy, unsigned AS,
2151 Instruction *MemoryInst,
2152 SmallVectorImpl<Instruction*> &AddrModeInsts,
2153 const TargetMachine &TM,
2154 const SetOfInstrs &InsertedInsts,
2155 InstrToOrigTy &PromotedInsts,
2156 TypePromotionTransaction &TPT) {
2159 bool Success = AddressingModeMatcher(AddrModeInsts, TM, AccessTy, AS,
2160 MemoryInst, Result, InsertedInsts,
2161 PromotedInsts, TPT).MatchAddr(V, 0);
2162 (void)Success; assert(Success && "Couldn't select *anything*?");
2166 bool MatchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
2167 bool MatchAddr(Value *V, unsigned Depth);
2168 bool MatchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth,
2169 bool *MovedAway = nullptr);
2170 bool IsProfitableToFoldIntoAddressingMode(Instruction *I,
2171 ExtAddrMode &AMBefore,
2172 ExtAddrMode &AMAfter);
2173 bool ValueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
2174 bool IsPromotionProfitable(unsigned NewCost, unsigned OldCost,
2175 Value *PromotedOperand) const;
2178 /// MatchScaledValue - Try adding ScaleReg*Scale to the current addressing mode.
2179 /// Return true and update AddrMode if this addr mode is legal for the target,
2181 bool AddressingModeMatcher::MatchScaledValue(Value *ScaleReg, int64_t Scale,
2183 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
2184 // mode. Just process that directly.
2186 return MatchAddr(ScaleReg, Depth);
2188 // If the scale is 0, it takes nothing to add this.
2192 // If we already have a scale of this value, we can add to it, otherwise, we
2193 // need an available scale field.
2194 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
2197 ExtAddrMode TestAddrMode = AddrMode;
2199 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
2200 // [A+B + A*7] -> [B+A*8].
2201 TestAddrMode.Scale += Scale;
2202 TestAddrMode.ScaledReg = ScaleReg;
2204 // If the new address isn't legal, bail out.
2205 if (!TLI.isLegalAddressingMode(TestAddrMode, AccessTy, AddrSpace))
2208 // It was legal, so commit it.
2209 AddrMode = TestAddrMode;
2211 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
2212 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
2213 // X*Scale + C*Scale to addr mode.
2214 ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
2215 if (isa<Instruction>(ScaleReg) && // not a constant expr.
2216 match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
2217 TestAddrMode.ScaledReg = AddLHS;
2218 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
2220 // If this addressing mode is legal, commit it and remember that we folded
2221 // this instruction.
2222 if (TLI.isLegalAddressingMode(TestAddrMode, AccessTy, AddrSpace)) {
2223 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
2224 AddrMode = TestAddrMode;
2229 // Otherwise, not (x+c)*scale, just return what we have.
2233 /// MightBeFoldableInst - This is a little filter, which returns true if an
2234 /// addressing computation involving I might be folded into a load/store
2235 /// accessing it. This doesn't need to be perfect, but needs to accept at least
2236 /// the set of instructions that MatchOperationAddr can.
2237 static bool MightBeFoldableInst(Instruction *I) {
2238 switch (I->getOpcode()) {
2239 case Instruction::BitCast:
2240 case Instruction::AddrSpaceCast:
2241 // Don't touch identity bitcasts.
2242 if (I->getType() == I->getOperand(0)->getType())
2244 return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
2245 case Instruction::PtrToInt:
2246 // PtrToInt is always a noop, as we know that the int type is pointer sized.
2248 case Instruction::IntToPtr:
2249 // We know the input is intptr_t, so this is foldable.
2251 case Instruction::Add:
2253 case Instruction::Mul:
2254 case Instruction::Shl:
2255 // Can only handle X*C and X << C.
2256 return isa<ConstantInt>(I->getOperand(1));
2257 case Instruction::GetElementPtr:
2264 /// \brief Check whether or not \p Val is a legal instruction for \p TLI.
2265 /// \note \p Val is assumed to be the product of some type promotion.
2266 /// Therefore if \p Val has an undefined state in \p TLI, this is assumed
2267 /// to be legal, as the non-promoted value would have had the same state.
2268 static bool isPromotedInstructionLegal(const TargetLowering &TLI, Value *Val) {
2269 Instruction *PromotedInst = dyn_cast<Instruction>(Val);
2272 int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
2273 // If the ISDOpcode is undefined, it was undefined before the promotion.
2276 // Otherwise, check if the promoted instruction is legal or not.
2277 return TLI.isOperationLegalOrCustom(
2278 ISDOpcode, TLI.getValueType(PromotedInst->getType()));
2281 /// \brief Hepler class to perform type promotion.
2282 class TypePromotionHelper {
2283 /// \brief Utility function to check whether or not a sign or zero extension
2284 /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
2285 /// either using the operands of \p Inst or promoting \p Inst.
2286 /// The type of the extension is defined by \p IsSExt.
2287 /// In other words, check if:
2288 /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
2289 /// #1 Promotion applies:
2290 /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
2291 /// #2 Operand reuses:
2292 /// ext opnd1 to ConsideredExtType.
2293 /// \p PromotedInsts maps the instructions to their type before promotion.
2294 static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
2295 const InstrToOrigTy &PromotedInsts, bool IsSExt);
2297 /// \brief Utility function to determine if \p OpIdx should be promoted when
2298 /// promoting \p Inst.
2299 static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
2300 if (isa<SelectInst>(Inst) && OpIdx == 0)
2305 /// \brief Utility function to promote the operand of \p Ext when this
2306 /// operand is a promotable trunc or sext or zext.
2307 /// \p PromotedInsts maps the instructions to their type before promotion.
2308 /// \p CreatedInstsCost[out] contains the cost of all instructions
2309 /// created to promote the operand of Ext.
2310 /// Newly added extensions are inserted in \p Exts.
2311 /// Newly added truncates are inserted in \p Truncs.
2312 /// Should never be called directly.
2313 /// \return The promoted value which is used instead of Ext.
2314 static Value *promoteOperandForTruncAndAnyExt(
2315 Instruction *Ext, TypePromotionTransaction &TPT,
2316 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2317 SmallVectorImpl<Instruction *> *Exts,
2318 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI);
2320 /// \brief Utility function to promote the operand of \p Ext when this
2321 /// operand is promotable and is not a supported trunc or sext.
2322 /// \p PromotedInsts maps the instructions to their type before promotion.
2323 /// \p CreatedInstsCost[out] contains the cost of all the instructions
2324 /// created to promote the operand of Ext.
2325 /// Newly added extensions are inserted in \p Exts.
2326 /// Newly added truncates are inserted in \p Truncs.
2327 /// Should never be called directly.
2328 /// \return The promoted value which is used instead of Ext.
2329 static Value *promoteOperandForOther(Instruction *Ext,
2330 TypePromotionTransaction &TPT,
2331 InstrToOrigTy &PromotedInsts,
2332 unsigned &CreatedInstsCost,
2333 SmallVectorImpl<Instruction *> *Exts,
2334 SmallVectorImpl<Instruction *> *Truncs,
2335 const TargetLowering &TLI, bool IsSExt);
2337 /// \see promoteOperandForOther.
2338 static Value *signExtendOperandForOther(
2339 Instruction *Ext, TypePromotionTransaction &TPT,
2340 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2341 SmallVectorImpl<Instruction *> *Exts,
2342 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2343 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
2344 Exts, Truncs, TLI, true);
2347 /// \see promoteOperandForOther.
2348 static Value *zeroExtendOperandForOther(
2349 Instruction *Ext, TypePromotionTransaction &TPT,
2350 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2351 SmallVectorImpl<Instruction *> *Exts,
2352 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2353 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
2354 Exts, Truncs, TLI, false);
2358 /// Type for the utility function that promotes the operand of Ext.
2359 typedef Value *(*Action)(Instruction *Ext, TypePromotionTransaction &TPT,
2360 InstrToOrigTy &PromotedInsts,
2361 unsigned &CreatedInstsCost,
2362 SmallVectorImpl<Instruction *> *Exts,
2363 SmallVectorImpl<Instruction *> *Truncs,
2364 const TargetLowering &TLI);
2365 /// \brief Given a sign/zero extend instruction \p Ext, return the approriate
2366 /// action to promote the operand of \p Ext instead of using Ext.
2367 /// \return NULL if no promotable action is possible with the current
2369 /// \p InsertedInsts keeps track of all the instructions inserted by the
2370 /// other CodeGenPrepare optimizations. This information is important
2371 /// because we do not want to promote these instructions as CodeGenPrepare
2372 /// will reinsert them later. Thus creating an infinite loop: create/remove.
2373 /// \p PromotedInsts maps the instructions to their type before promotion.
2374 static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedInsts,
2375 const TargetLowering &TLI,
2376 const InstrToOrigTy &PromotedInsts);
2379 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
2380 Type *ConsideredExtType,
2381 const InstrToOrigTy &PromotedInsts,
2383 // The promotion helper does not know how to deal with vector types yet.
2384 // To be able to fix that, we would need to fix the places where we
2385 // statically extend, e.g., constants and such.
2386 if (Inst->getType()->isVectorTy())
2389 // We can always get through zext.
2390 if (isa<ZExtInst>(Inst))
2393 // sext(sext) is ok too.
2394 if (IsSExt && isa<SExtInst>(Inst))
2397 // We can get through binary operator, if it is legal. In other words, the
2398 // binary operator must have a nuw or nsw flag.
2399 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
2400 if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
2401 ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
2402 (IsSExt && BinOp->hasNoSignedWrap())))
2405 // Check if we can do the following simplification.
2406 // ext(trunc(opnd)) --> ext(opnd)
2407 if (!isa<TruncInst>(Inst))
2410 Value *OpndVal = Inst->getOperand(0);
2411 // Check if we can use this operand in the extension.
2412 // If the type is larger than the result type of the extension,
2414 if (!OpndVal->getType()->isIntegerTy() ||
2415 OpndVal->getType()->getIntegerBitWidth() >
2416 ConsideredExtType->getIntegerBitWidth())
2419 // If the operand of the truncate is not an instruction, we will not have
2420 // any information on the dropped bits.
2421 // (Actually we could for constant but it is not worth the extra logic).
2422 Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
2426 // Check if the source of the type is narrow enough.
2427 // I.e., check that trunc just drops extended bits of the same kind of
2429 // #1 get the type of the operand and check the kind of the extended bits.
2430 const Type *OpndType;
2431 InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
2432 if (It != PromotedInsts.end() && It->second.IsSExt == IsSExt)
2433 OpndType = It->second.Ty;
2434 else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd)))
2435 OpndType = Opnd->getOperand(0)->getType();
2439 // #2 check that the truncate just drop extended bits.
2440 if (Inst->getType()->getIntegerBitWidth() >= OpndType->getIntegerBitWidth())
2446 TypePromotionHelper::Action TypePromotionHelper::getAction(
2447 Instruction *Ext, const SetOfInstrs &InsertedInsts,
2448 const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
2449 assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
2450 "Unexpected instruction type");
2451 Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0));
2452 Type *ExtTy = Ext->getType();
2453 bool IsSExt = isa<SExtInst>(Ext);
2454 // If the operand of the extension is not an instruction, we cannot
2456 // If it, check we can get through.
2457 if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt))
2460 // Do not promote if the operand has been added by codegenprepare.
2461 // Otherwise, it means we are undoing an optimization that is likely to be
2462 // redone, thus causing potential infinite loop.
2463 if (isa<TruncInst>(ExtOpnd) && InsertedInsts.count(ExtOpnd))
2466 // SExt or Trunc instructions.
2467 // Return the related handler.
2468 if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) ||
2469 isa<ZExtInst>(ExtOpnd))
2470 return promoteOperandForTruncAndAnyExt;
2472 // Regular instruction.
2473 // Abort early if we will have to insert non-free instructions.
2474 if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType()))
2476 return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
2479 Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
2480 llvm::Instruction *SExt, TypePromotionTransaction &TPT,
2481 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2482 SmallVectorImpl<Instruction *> *Exts,
2483 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2484 // By construction, the operand of SExt is an instruction. Otherwise we cannot
2485 // get through it and this method should not be called.
2486 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
2487 Value *ExtVal = SExt;
2488 bool HasMergedNonFreeExt = false;
2489 if (isa<ZExtInst>(SExtOpnd)) {
2490 // Replace s|zext(zext(opnd))
2492 HasMergedNonFreeExt = !TLI.isExtFree(SExtOpnd);
2494 TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
2495 TPT.replaceAllUsesWith(SExt, ZExt);
2496 TPT.eraseInstruction(SExt);
2499 // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
2501 TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
2503 CreatedInstsCost = 0;
2505 // Remove dead code.
2506 if (SExtOpnd->use_empty())
2507 TPT.eraseInstruction(SExtOpnd);
2509 // Check if the extension is still needed.
2510 Instruction *ExtInst = dyn_cast<Instruction>(ExtVal);
2511 if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) {
2514 Exts->push_back(ExtInst);
2515 CreatedInstsCost = !TLI.isExtFree(ExtInst) && !HasMergedNonFreeExt;
2520 // At this point we have: ext ty opnd to ty.
2521 // Reassign the uses of ExtInst to the opnd and remove ExtInst.
2522 Value *NextVal = ExtInst->getOperand(0);
2523 TPT.eraseInstruction(ExtInst, NextVal);
2527 Value *TypePromotionHelper::promoteOperandForOther(
2528 Instruction *Ext, TypePromotionTransaction &TPT,
2529 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2530 SmallVectorImpl<Instruction *> *Exts,
2531 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI,
2533 // By construction, the operand of Ext is an instruction. Otherwise we cannot
2534 // get through it and this method should not be called.
2535 Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0));
2536 CreatedInstsCost = 0;
2537 if (!ExtOpnd->hasOneUse()) {
2538 // ExtOpnd will be promoted.
2539 // All its uses, but Ext, will need to use a truncated value of the
2540 // promoted version.
2541 // Create the truncate now.
2542 Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType());
2543 if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) {
2544 ITrunc->removeFromParent();
2545 // Insert it just after the definition.
2546 ITrunc->insertAfter(ExtOpnd);
2548 Truncs->push_back(ITrunc);
2551 TPT.replaceAllUsesWith(ExtOpnd, Trunc);
2552 // Restore the operand of Ext (which has been replace by the previous call
2553 // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
2554 TPT.setOperand(Ext, 0, ExtOpnd);
2557 // Get through the Instruction:
2558 // 1. Update its type.
2559 // 2. Replace the uses of Ext by Inst.
2560 // 3. Extend each operand that needs to be extended.
2562 // Remember the original type of the instruction before promotion.
2563 // This is useful to know that the high bits are sign extended bits.
2564 PromotedInsts.insert(std::pair<Instruction *, TypeIsSExt>(
2565 ExtOpnd, TypeIsSExt(ExtOpnd->getType(), IsSExt)));
2567 TPT.mutateType(ExtOpnd, Ext->getType());
2569 TPT.replaceAllUsesWith(Ext, ExtOpnd);
2571 Instruction *ExtForOpnd = Ext;
2573 DEBUG(dbgs() << "Propagate Ext to operands\n");
2574 for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
2576 DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n');
2577 if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() ||
2578 !shouldExtOperand(ExtOpnd, OpIdx)) {
2579 DEBUG(dbgs() << "No need to propagate\n");
2582 // Check if we can statically extend the operand.
2583 Value *Opnd = ExtOpnd->getOperand(OpIdx);
2584 if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
2585 DEBUG(dbgs() << "Statically extend\n");
2586 unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
2587 APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth)
2588 : Cst->getValue().zext(BitWidth);
2589 TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal));
2592 // UndefValue are typed, so we have to statically sign extend them.
2593 if (isa<UndefValue>(Opnd)) {
2594 DEBUG(dbgs() << "Statically extend\n");
2595 TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType()));
2599 // Otherwise we have to explicity sign extend the operand.
2600 // Check if Ext was reused to extend an operand.
2602 // If yes, create a new one.
2603 DEBUG(dbgs() << "More operands to ext\n");
2604 Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType())
2605 : TPT.createZExt(Ext, Opnd, Ext->getType());
2606 if (!isa<Instruction>(ValForExtOpnd)) {
2607 TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd);
2610 ExtForOpnd = cast<Instruction>(ValForExtOpnd);
2613 Exts->push_back(ExtForOpnd);
2614 TPT.setOperand(ExtForOpnd, 0, Opnd);
2616 // Move the sign extension before the insertion point.
2617 TPT.moveBefore(ExtForOpnd, ExtOpnd);
2618 TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd);
2619 CreatedInstsCost += !TLI.isExtFree(ExtForOpnd);
2620 // If more sext are required, new instructions will have to be created.
2621 ExtForOpnd = nullptr;
2623 if (ExtForOpnd == Ext) {
2624 DEBUG(dbgs() << "Extension is useless now\n");
2625 TPT.eraseInstruction(Ext);
2630 /// IsPromotionProfitable - Check whether or not promoting an instruction
2631 /// to a wider type was profitable.
2632 /// \p NewCost gives the cost of extension instructions created by the
2634 /// \p OldCost gives the cost of extension instructions before the promotion
2635 /// plus the number of instructions that have been
2636 /// matched in the addressing mode the promotion.
2637 /// \p PromotedOperand is the value that has been promoted.
2638 /// \return True if the promotion is profitable, false otherwise.
2639 bool AddressingModeMatcher::IsPromotionProfitable(
2640 unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const {
2641 DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCost << '\n');
2642 // The cost of the new extensions is greater than the cost of the
2643 // old extension plus what we folded.
2644 // This is not profitable.
2645 if (NewCost > OldCost)
2647 if (NewCost < OldCost)
2649 // The promotion is neutral but it may help folding the sign extension in
2650 // loads for instance.
2651 // Check that we did not create an illegal instruction.
2652 return isPromotedInstructionLegal(TLI, PromotedOperand);
2655 /// MatchOperationAddr - Given an instruction or constant expr, see if we can
2656 /// fold the operation into the addressing mode. If so, update the addressing
2657 /// mode and return true, otherwise return false without modifying AddrMode.
2658 /// If \p MovedAway is not NULL, it contains the information of whether or
2659 /// not AddrInst has to be folded into the addressing mode on success.
2660 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
2661 /// because it has been moved away.
2662 /// Thus AddrInst must not be added in the matched instructions.
2663 /// This state can happen when AddrInst is a sext, since it may be moved away.
2664 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
2665 /// not be referenced anymore.
2666 bool AddressingModeMatcher::MatchOperationAddr(User *AddrInst, unsigned Opcode,
2669 // Avoid exponential behavior on extremely deep expression trees.
2670 if (Depth >= 5) return false;
2672 // By default, all matched instructions stay in place.
2677 case Instruction::PtrToInt:
2678 // PtrToInt is always a noop, as we know that the int type is pointer sized.
2679 return MatchAddr(AddrInst->getOperand(0), Depth);
2680 case Instruction::IntToPtr:
2681 // This inttoptr is a no-op if the integer type is pointer sized.
2682 if (TLI.getValueType(AddrInst->getOperand(0)->getType()) ==
2683 TLI.getPointerTy(AddrInst->getType()->getPointerAddressSpace()))
2684 return MatchAddr(AddrInst->getOperand(0), Depth);
2686 case Instruction::BitCast:
2687 // BitCast is always a noop, and we can handle it as long as it is
2688 // int->int or pointer->pointer (we don't want int<->fp or something).
2689 if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
2690 AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
2691 // Don't touch identity bitcasts. These were probably put here by LSR,
2692 // and we don't want to mess around with them. Assume it knows what it
2694 AddrInst->getOperand(0)->getType() != AddrInst->getType())
2695 return MatchAddr(AddrInst->getOperand(0), Depth);
2697 case Instruction::AddrSpaceCast: {
2699 = AddrInst->getOperand(0)->getType()->getPointerAddressSpace();
2700 unsigned DestAS = AddrInst->getType()->getPointerAddressSpace();
2701 if (TLI.isNoopAddrSpaceCast(SrcAS, DestAS))
2702 return MatchAddr(AddrInst->getOperand(0), Depth);
2705 case Instruction::Add: {
2706 // Check to see if we can merge in the RHS then the LHS. If so, we win.
2707 ExtAddrMode BackupAddrMode = AddrMode;
2708 unsigned OldSize = AddrModeInsts.size();
2709 // Start a transaction at this point.
2710 // The LHS may match but not the RHS.
2711 // Therefore, we need a higher level restoration point to undo partially
2712 // matched operation.
2713 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2714 TPT.getRestorationPoint();
2716 if (MatchAddr(AddrInst->getOperand(1), Depth+1) &&
2717 MatchAddr(AddrInst->getOperand(0), Depth+1))
2720 // Restore the old addr mode info.
2721 AddrMode = BackupAddrMode;
2722 AddrModeInsts.resize(OldSize);
2723 TPT.rollback(LastKnownGood);
2725 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
2726 if (MatchAddr(AddrInst->getOperand(0), Depth+1) &&
2727 MatchAddr(AddrInst->getOperand(1), Depth+1))
2730 // Otherwise we definitely can't merge the ADD in.
2731 AddrMode = BackupAddrMode;
2732 AddrModeInsts.resize(OldSize);
2733 TPT.rollback(LastKnownGood);
2736 //case Instruction::Or:
2737 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
2739 case Instruction::Mul:
2740 case Instruction::Shl: {
2741 // Can only handle X*C and X << C.
2742 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
2745 int64_t Scale = RHS->getSExtValue();
2746 if (Opcode == Instruction::Shl)
2747 Scale = 1LL << Scale;
2749 return MatchScaledValue(AddrInst->getOperand(0), Scale, Depth);
2751 case Instruction::GetElementPtr: {
2752 // Scan the GEP. We check it if it contains constant offsets and at most
2753 // one variable offset.
2754 int VariableOperand = -1;
2755 unsigned VariableScale = 0;
2757 int64_t ConstantOffset = 0;
2758 gep_type_iterator GTI = gep_type_begin(AddrInst);
2759 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
2760 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
2761 const StructLayout *SL = DL.getStructLayout(STy);
2763 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
2764 ConstantOffset += SL->getElementOffset(Idx);
2766 uint64_t TypeSize = DL.getTypeAllocSize(GTI.getIndexedType());
2767 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
2768 ConstantOffset += CI->getSExtValue()*TypeSize;
2769 } else if (TypeSize) { // Scales of zero don't do anything.
2770 // We only allow one variable index at the moment.
2771 if (VariableOperand != -1)
2774 // Remember the variable index.
2775 VariableOperand = i;
2776 VariableScale = TypeSize;
2781 // A common case is for the GEP to only do a constant offset. In this case,
2782 // just add it to the disp field and check validity.
2783 if (VariableOperand == -1) {
2784 AddrMode.BaseOffs += ConstantOffset;
2785 if (ConstantOffset == 0 ||
2786 TLI.isLegalAddressingMode(AddrMode, AccessTy, AddrSpace)) {
2787 // Check to see if we can fold the base pointer in too.
2788 if (MatchAddr(AddrInst->getOperand(0), Depth+1))
2791 AddrMode.BaseOffs -= ConstantOffset;
2795 // Save the valid addressing mode in case we can't match.
2796 ExtAddrMode BackupAddrMode = AddrMode;
2797 unsigned OldSize = AddrModeInsts.size();
2799 // See if the scale and offset amount is valid for this target.
2800 AddrMode.BaseOffs += ConstantOffset;
2802 // Match the base operand of the GEP.
2803 if (!MatchAddr(AddrInst->getOperand(0), Depth+1)) {
2804 // If it couldn't be matched, just stuff the value in a register.
2805 if (AddrMode.HasBaseReg) {
2806 AddrMode = BackupAddrMode;
2807 AddrModeInsts.resize(OldSize);
2810 AddrMode.HasBaseReg = true;
2811 AddrMode.BaseReg = AddrInst->getOperand(0);
2814 // Match the remaining variable portion of the GEP.
2815 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
2817 // If it couldn't be matched, try stuffing the base into a register
2818 // instead of matching it, and retrying the match of the scale.
2819 AddrMode = BackupAddrMode;
2820 AddrModeInsts.resize(OldSize);
2821 if (AddrMode.HasBaseReg)
2823 AddrMode.HasBaseReg = true;
2824 AddrMode.BaseReg = AddrInst->getOperand(0);
2825 AddrMode.BaseOffs += ConstantOffset;
2826 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand),
2827 VariableScale, Depth)) {
2828 // If even that didn't work, bail.
2829 AddrMode = BackupAddrMode;
2830 AddrModeInsts.resize(OldSize);
2837 case Instruction::SExt:
2838 case Instruction::ZExt: {
2839 Instruction *Ext = dyn_cast<Instruction>(AddrInst);
2843 // Try to move this ext out of the way of the addressing mode.
2844 // Ask for a method for doing so.
2845 TypePromotionHelper::Action TPH =
2846 TypePromotionHelper::getAction(Ext, InsertedInsts, TLI, PromotedInsts);
2850 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2851 TPT.getRestorationPoint();
2852 unsigned CreatedInstsCost = 0;
2853 unsigned ExtCost = !TLI.isExtFree(Ext);
2854 Value *PromotedOperand =
2855 TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI);
2856 // SExt has been moved away.
2857 // Thus either it will be rematched later in the recursive calls or it is
2858 // gone. Anyway, we must not fold it into the addressing mode at this point.
2862 // addr = gep base, idx
2864 // promotedOpnd = ext opnd <- no match here
2865 // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
2866 // addr = gep base, op <- match
2870 assert(PromotedOperand &&
2871 "TypePromotionHelper should have filtered out those cases");
2873 ExtAddrMode BackupAddrMode = AddrMode;
2874 unsigned OldSize = AddrModeInsts.size();
2876 if (!MatchAddr(PromotedOperand, Depth) ||
2877 // The total of the new cost is equals to the cost of the created
2879 // The total of the old cost is equals to the cost of the extension plus
2880 // what we have saved in the addressing mode.
2881 !IsPromotionProfitable(CreatedInstsCost,
2882 ExtCost + (AddrModeInsts.size() - OldSize),
2884 AddrMode = BackupAddrMode;
2885 AddrModeInsts.resize(OldSize);
2886 DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
2887 TPT.rollback(LastKnownGood);
2896 /// MatchAddr - If we can, try to add the value of 'Addr' into the current
2897 /// addressing mode. If Addr can't be added to AddrMode this returns false and
2898 /// leaves AddrMode unmodified. This assumes that Addr is either a pointer type
2899 /// or intptr_t for the target.
2901 bool AddressingModeMatcher::MatchAddr(Value *Addr, unsigned Depth) {
2902 // Start a transaction at this point that we will rollback if the matching
2904 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2905 TPT.getRestorationPoint();
2906 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
2907 // Fold in immediates if legal for the target.
2908 AddrMode.BaseOffs += CI->getSExtValue();
2909 if (TLI.isLegalAddressingMode(AddrMode, AccessTy, AddrSpace))
2911 AddrMode.BaseOffs -= CI->getSExtValue();
2912 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
2913 // If this is a global variable, try to fold it into the addressing mode.
2914 if (!AddrMode.BaseGV) {
2915 AddrMode.BaseGV = GV;
2916 if (TLI.isLegalAddressingMode(AddrMode, AccessTy, AddrSpace))
2918 AddrMode.BaseGV = nullptr;
2920 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
2921 ExtAddrMode BackupAddrMode = AddrMode;
2922 unsigned OldSize = AddrModeInsts.size();
2924 // Check to see if it is possible to fold this operation.
2925 bool MovedAway = false;
2926 if (MatchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
2927 // This instruction may have been move away. If so, there is nothing
2931 // Okay, it's possible to fold this. Check to see if it is actually
2932 // *profitable* to do so. We use a simple cost model to avoid increasing
2933 // register pressure too much.
2934 if (I->hasOneUse() ||
2935 IsProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
2936 AddrModeInsts.push_back(I);
2940 // It isn't profitable to do this, roll back.
2941 //cerr << "NOT FOLDING: " << *I;
2942 AddrMode = BackupAddrMode;
2943 AddrModeInsts.resize(OldSize);
2944 TPT.rollback(LastKnownGood);
2946 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
2947 if (MatchOperationAddr(CE, CE->getOpcode(), Depth))
2949 TPT.rollback(LastKnownGood);
2950 } else if (isa<ConstantPointerNull>(Addr)) {
2951 // Null pointer gets folded without affecting the addressing mode.
2955 // Worse case, the target should support [reg] addressing modes. :)
2956 if (!AddrMode.HasBaseReg) {
2957 AddrMode.HasBaseReg = true;
2958 AddrMode.BaseReg = Addr;
2959 // Still check for legality in case the target supports [imm] but not [i+r].
2960 if (TLI.isLegalAddressingMode(AddrMode, AccessTy, AddrSpace))
2962 AddrMode.HasBaseReg = false;
2963 AddrMode.BaseReg = nullptr;
2966 // If the base register is already taken, see if we can do [r+r].
2967 if (AddrMode.Scale == 0) {
2969 AddrMode.ScaledReg = Addr;
2970 if (TLI.isLegalAddressingMode(AddrMode, AccessTy, AddrSpace))
2973 AddrMode.ScaledReg = nullptr;
2976 TPT.rollback(LastKnownGood);
2980 /// IsOperandAMemoryOperand - Check to see if all uses of OpVal by the specified
2981 /// inline asm call are due to memory operands. If so, return true, otherwise
2983 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
2984 const TargetMachine &TM) {
2985 const Function *F = CI->getParent()->getParent();
2986 const TargetLowering *TLI = TM.getSubtargetImpl(*F)->getTargetLowering();
2987 const TargetRegisterInfo *TRI = TM.getSubtargetImpl(*F)->getRegisterInfo();
2988 TargetLowering::AsmOperandInfoVector TargetConstraints =
2989 TLI->ParseConstraints(F->getParent()->getDataLayout(), TRI,
2990 ImmutableCallSite(CI));
2991 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
2992 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
2994 // Compute the constraint code and ConstraintType to use.
2995 TLI->ComputeConstraintToUse(OpInfo, SDValue());
2997 // If this asm operand is our Value*, and if it isn't an indirect memory
2998 // operand, we can't fold it!
2999 if (OpInfo.CallOperandVal == OpVal &&
3000 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
3001 !OpInfo.isIndirect))
3008 /// FindAllMemoryUses - Recursively walk all the uses of I until we find a
3009 /// memory use. If we find an obviously non-foldable instruction, return true.
3010 /// Add the ultimately found memory instructions to MemoryUses.
3011 static bool FindAllMemoryUses(
3013 SmallVectorImpl<std::pair<Instruction *, unsigned>> &MemoryUses,
3014 SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetMachine &TM) {
3015 // If we already considered this instruction, we're done.
3016 if (!ConsideredInsts.insert(I).second)
3019 // If this is an obviously unfoldable instruction, bail out.
3020 if (!MightBeFoldableInst(I))
3023 // Loop over all the uses, recursively processing them.
3024 for (Use &U : I->uses()) {
3025 Instruction *UserI = cast<Instruction>(U.getUser());
3027 if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
3028 MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
3032 if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
3033 unsigned opNo = U.getOperandNo();
3034 if (opNo == 0) return true; // Storing addr, not into addr.
3035 MemoryUses.push_back(std::make_pair(SI, opNo));
3039 if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
3040 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
3041 if (!IA) return true;
3043 // If this is a memory operand, we're cool, otherwise bail out.
3044 if (!IsOperandAMemoryOperand(CI, IA, I, TM))
3049 if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TM))
3056 /// ValueAlreadyLiveAtInst - Retrn true if Val is already known to be live at
3057 /// the use site that we're folding it into. If so, there is no cost to
3058 /// include it in the addressing mode. KnownLive1 and KnownLive2 are two values
3059 /// that we know are live at the instruction already.
3060 bool AddressingModeMatcher::ValueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
3061 Value *KnownLive2) {
3062 // If Val is either of the known-live values, we know it is live!
3063 if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
3066 // All values other than instructions and arguments (e.g. constants) are live.
3067 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
3069 // If Val is a constant sized alloca in the entry block, it is live, this is
3070 // true because it is just a reference to the stack/frame pointer, which is
3071 // live for the whole function.
3072 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
3073 if (AI->isStaticAlloca())
3076 // Check to see if this value is already used in the memory instruction's
3077 // block. If so, it's already live into the block at the very least, so we
3078 // can reasonably fold it.
3079 return Val->isUsedInBasicBlock(MemoryInst->getParent());
3082 /// IsProfitableToFoldIntoAddressingMode - It is possible for the addressing
3083 /// mode of the machine to fold the specified instruction into a load or store
3084 /// that ultimately uses it. However, the specified instruction has multiple
3085 /// uses. Given this, it may actually increase register pressure to fold it
3086 /// into the load. For example, consider this code:
3090 /// use(Y) -> nonload/store
3094 /// In this case, Y has multiple uses, and can be folded into the load of Z
3095 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
3096 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
3097 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
3098 /// number of computations either.
3100 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
3101 /// X was live across 'load Z' for other reasons, we actually *would* want to
3102 /// fold the addressing mode in the Z case. This would make Y die earlier.
3103 bool AddressingModeMatcher::
3104 IsProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
3105 ExtAddrMode &AMAfter) {
3106 if (IgnoreProfitability) return true;
3108 // AMBefore is the addressing mode before this instruction was folded into it,
3109 // and AMAfter is the addressing mode after the instruction was folded. Get
3110 // the set of registers referenced by AMAfter and subtract out those
3111 // referenced by AMBefore: this is the set of values which folding in this
3112 // address extends the lifetime of.
3114 // Note that there are only two potential values being referenced here,
3115 // BaseReg and ScaleReg (global addresses are always available, as are any
3116 // folded immediates).
3117 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
3119 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
3120 // lifetime wasn't extended by adding this instruction.
3121 if (ValueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
3123 if (ValueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
3124 ScaledReg = nullptr;
3126 // If folding this instruction (and it's subexprs) didn't extend any live
3127 // ranges, we're ok with it.
3128 if (!BaseReg && !ScaledReg)
3131 // If all uses of this instruction are ultimately load/store/inlineasm's,
3132 // check to see if their addressing modes will include this instruction. If
3133 // so, we can fold it into all uses, so it doesn't matter if it has multiple
3135 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
3136 SmallPtrSet<Instruction*, 16> ConsideredInsts;
3137 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TM))
3138 return false; // Has a non-memory, non-foldable use!
3140 // Now that we know that all uses of this instruction are part of a chain of
3141 // computation involving only operations that could theoretically be folded
3142 // into a memory use, loop over each of these uses and see if they could
3143 // *actually* fold the instruction.
3144 SmallVector<Instruction*, 32> MatchedAddrModeInsts;
3145 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
3146 Instruction *User = MemoryUses[i].first;
3147 unsigned OpNo = MemoryUses[i].second;
3149 // Get the access type of this use. If the use isn't a pointer, we don't
3150 // know what it accesses.
3151 Value *Address = User->getOperand(OpNo);
3152 PointerType *AddrTy = dyn_cast<PointerType>(Address->getType());
3155 Type *AddressAccessTy = AddrTy->getElementType();
3156 unsigned AS = AddrTy->getAddressSpace();
3158 // Do a match against the root of this address, ignoring profitability. This
3159 // will tell us if the addressing mode for the memory operation will
3160 // *actually* cover the shared instruction.
3162 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3163 TPT.getRestorationPoint();
3164 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TM, AddressAccessTy, AS,
3165 MemoryInst, Result, InsertedInsts,
3166 PromotedInsts, TPT);
3167 Matcher.IgnoreProfitability = true;
3168 bool Success = Matcher.MatchAddr(Address, 0);
3169 (void)Success; assert(Success && "Couldn't select *anything*?");
3171 // The match was to check the profitability, the changes made are not
3172 // part of the original matcher. Therefore, they should be dropped
3173 // otherwise the original matcher will not present the right state.
3174 TPT.rollback(LastKnownGood);
3176 // If the match didn't cover I, then it won't be shared by it.
3177 if (std::find(MatchedAddrModeInsts.begin(), MatchedAddrModeInsts.end(),
3178 I) == MatchedAddrModeInsts.end())
3181 MatchedAddrModeInsts.clear();
3187 } // end anonymous namespace
3189 /// IsNonLocalValue - Return true if the specified values are defined in a
3190 /// different basic block than BB.
3191 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
3192 if (Instruction *I = dyn_cast<Instruction>(V))
3193 return I->getParent() != BB;
3197 /// OptimizeMemoryInst - Load and Store Instructions often have
3198 /// addressing modes that can do significant amounts of computation. As such,
3199 /// instruction selection will try to get the load or store to do as much
3200 /// computation as possible for the program. The problem is that isel can only
3201 /// see within a single block. As such, we sink as much legal addressing mode
3202 /// stuff into the block as possible.
3204 /// This method is used to optimize both load/store and inline asms with memory
3206 bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
3207 Type *AccessTy, unsigned AddrSpace) {
3210 // Try to collapse single-value PHI nodes. This is necessary to undo
3211 // unprofitable PRE transformations.
3212 SmallVector<Value*, 8> worklist;
3213 SmallPtrSet<Value*, 16> Visited;
3214 worklist.push_back(Addr);
3216 // Use a worklist to iteratively look through PHI nodes, and ensure that
3217 // the addressing mode obtained from the non-PHI roots of the graph
3219 Value *Consensus = nullptr;
3220 unsigned NumUsesConsensus = 0;
3221 bool IsNumUsesConsensusValid = false;
3222 SmallVector<Instruction*, 16> AddrModeInsts;
3223 ExtAddrMode AddrMode;
3224 TypePromotionTransaction TPT;
3225 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3226 TPT.getRestorationPoint();
3227 while (!worklist.empty()) {
3228 Value *V = worklist.back();
3229 worklist.pop_back();
3231 // Break use-def graph loops.
3232 if (!Visited.insert(V).second) {
3233 Consensus = nullptr;
3237 // For a PHI node, push all of its incoming values.
3238 if (PHINode *P = dyn_cast<PHINode>(V)) {
3239 for (Value *IncValue : P->incoming_values())
3240 worklist.push_back(IncValue);
3244 // For non-PHIs, determine the addressing mode being computed.
3245 SmallVector<Instruction*, 16> NewAddrModeInsts;
3246 ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
3247 V, AccessTy, AddrSpace, MemoryInst, NewAddrModeInsts, *TM,
3248 InsertedInsts, PromotedInsts, TPT);
3250 // This check is broken into two cases with very similar code to avoid using
3251 // getNumUses() as much as possible. Some values have a lot of uses, so
3252 // calling getNumUses() unconditionally caused a significant compile-time
3256 AddrMode = NewAddrMode;
3257 AddrModeInsts = NewAddrModeInsts;
3259 } else if (NewAddrMode == AddrMode) {
3260 if (!IsNumUsesConsensusValid) {
3261 NumUsesConsensus = Consensus->getNumUses();
3262 IsNumUsesConsensusValid = true;
3265 // Ensure that the obtained addressing mode is equivalent to that obtained
3266 // for all other roots of the PHI traversal. Also, when choosing one
3267 // such root as representative, select the one with the most uses in order
3268 // to keep the cost modeling heuristics in AddressingModeMatcher
3270 unsigned NumUses = V->getNumUses();
3271 if (NumUses > NumUsesConsensus) {
3273 NumUsesConsensus = NumUses;
3274 AddrModeInsts = NewAddrModeInsts;
3279 Consensus = nullptr;
3283 // If the addressing mode couldn't be determined, or if multiple different
3284 // ones were determined, bail out now.
3286 TPT.rollback(LastKnownGood);
3291 // Check to see if any of the instructions supersumed by this addr mode are
3292 // non-local to I's BB.
3293 bool AnyNonLocal = false;
3294 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
3295 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
3301 // If all the instructions matched are already in this BB, don't do anything.
3303 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
3307 // Insert this computation right after this user. Since our caller is
3308 // scanning from the top of the BB to the bottom, reuse of the expr are
3309 // guaranteed to happen later.
3310 IRBuilder<> Builder(MemoryInst);
3312 // Now that we determined the addressing expression we want to use and know
3313 // that we have to sink it into this block. Check to see if we have already
3314 // done this for some other load/store instr in this block. If so, reuse the
3316 Value *&SunkAddr = SunkAddrs[Addr];
3318 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
3319 << *MemoryInst << "\n");
3320 if (SunkAddr->getType() != Addr->getType())
3321 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
3322 } else if (AddrSinkUsingGEPs ||
3323 (!AddrSinkUsingGEPs.getNumOccurrences() && TM &&
3324 TM->getSubtargetImpl(*MemoryInst->getParent()->getParent())
3326 // By default, we use the GEP-based method when AA is used later. This
3327 // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
3328 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
3329 << *MemoryInst << "\n");
3330 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
3331 Value *ResultPtr = nullptr, *ResultIndex = nullptr;
3333 // First, find the pointer.
3334 if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
3335 ResultPtr = AddrMode.BaseReg;
3336 AddrMode.BaseReg = nullptr;
3339 if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
3340 // We can't add more than one pointer together, nor can we scale a
3341 // pointer (both of which seem meaningless).
3342 if (ResultPtr || AddrMode.Scale != 1)
3345 ResultPtr = AddrMode.ScaledReg;
3349 if (AddrMode.BaseGV) {
3353 ResultPtr = AddrMode.BaseGV;
3356 // If the real base value actually came from an inttoptr, then the matcher
3357 // will look through it and provide only the integer value. In that case,
3359 if (!ResultPtr && AddrMode.BaseReg) {
3361 Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(), "sunkaddr");
3362 AddrMode.BaseReg = nullptr;
3363 } else if (!ResultPtr && AddrMode.Scale == 1) {
3365 Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(), "sunkaddr");
3370 !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
3371 SunkAddr = Constant::getNullValue(Addr->getType());
3372 } else if (!ResultPtr) {
3376 Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
3377 Type *I8Ty = Builder.getInt8Ty();
3379 // Start with the base register. Do this first so that subsequent address
3380 // matching finds it last, which will prevent it from trying to match it
3381 // as the scaled value in case it happens to be a mul. That would be
3382 // problematic if we've sunk a different mul for the scale, because then
3383 // we'd end up sinking both muls.
3384 if (AddrMode.BaseReg) {
3385 Value *V = AddrMode.BaseReg;
3386 if (V->getType() != IntPtrTy)
3387 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
3392 // Add the scale value.
3393 if (AddrMode.Scale) {
3394 Value *V = AddrMode.ScaledReg;
3395 if (V->getType() == IntPtrTy) {
3397 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
3398 cast<IntegerType>(V->getType())->getBitWidth()) {
3399 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
3401 // It is only safe to sign extend the BaseReg if we know that the math
3402 // required to create it did not overflow before we extend it. Since
3403 // the original IR value was tossed in favor of a constant back when
3404 // the AddrMode was created we need to bail out gracefully if widths
3405 // do not match instead of extending it.
3406 Instruction *I = dyn_cast_or_null<Instruction>(ResultIndex);
3407 if (I && (ResultIndex != AddrMode.BaseReg))
3408 I->eraseFromParent();
3412 if (AddrMode.Scale != 1)
3413 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
3416 ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
3421 // Add in the Base Offset if present.
3422 if (AddrMode.BaseOffs) {
3423 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
3425 // We need to add this separately from the scale above to help with
3426 // SDAG consecutive load/store merging.
3427 if (ResultPtr->getType() != I8PtrTy)
3428 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
3429 ResultPtr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
3436 SunkAddr = ResultPtr;
3438 if (ResultPtr->getType() != I8PtrTy)
3439 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
3440 SunkAddr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
3443 if (SunkAddr->getType() != Addr->getType())
3444 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
3447 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
3448 << *MemoryInst << "\n");
3449 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
3450 Value *Result = nullptr;
3452 // Start with the base register. Do this first so that subsequent address
3453 // matching finds it last, which will prevent it from trying to match it
3454 // as the scaled value in case it happens to be a mul. That would be
3455 // problematic if we've sunk a different mul for the scale, because then
3456 // we'd end up sinking both muls.
3457 if (AddrMode.BaseReg) {
3458 Value *V = AddrMode.BaseReg;
3459 if (V->getType()->isPointerTy())
3460 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
3461 if (V->getType() != IntPtrTy)
3462 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
3466 // Add the scale value.
3467 if (AddrMode.Scale) {
3468 Value *V = AddrMode.ScaledReg;
3469 if (V->getType() == IntPtrTy) {
3471 } else if (V->getType()->isPointerTy()) {
3472 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
3473 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
3474 cast<IntegerType>(V->getType())->getBitWidth()) {
3475 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
3477 // It is only safe to sign extend the BaseReg if we know that the math
3478 // required to create it did not overflow before we extend it. Since
3479 // the original IR value was tossed in favor of a constant back when
3480 // the AddrMode was created we need to bail out gracefully if widths
3481 // do not match instead of extending it.
3482 Instruction *I = dyn_cast_or_null<Instruction>(Result);
3483 if (I && (Result != AddrMode.BaseReg))
3484 I->eraseFromParent();
3487 if (AddrMode.Scale != 1)
3488 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
3491 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3496 // Add in the BaseGV if present.
3497 if (AddrMode.BaseGV) {
3498 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
3500 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3505 // Add in the Base Offset if present.
3506 if (AddrMode.BaseOffs) {
3507 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
3509 Result = Builder.CreateAdd(Result, V, "sunkaddr");
3515 SunkAddr = Constant::getNullValue(Addr->getType());
3517 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
3520 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
3522 // If we have no uses, recursively delete the value and all dead instructions
3524 if (Repl->use_empty()) {
3525 // This can cause recursive deletion, which can invalidate our iterator.
3526 // Use a WeakVH to hold onto it in case this happens.
3527 WeakVH IterHandle(CurInstIterator);
3528 BasicBlock *BB = CurInstIterator->getParent();
3530 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
3532 if (IterHandle != CurInstIterator) {
3533 // If the iterator instruction was recursively deleted, start over at the
3534 // start of the block.
3535 CurInstIterator = BB->begin();
3543 /// OptimizeInlineAsmInst - If there are any memory operands, use
3544 /// OptimizeMemoryInst to sink their address computing into the block when
3545 /// possible / profitable.
3546 bool CodeGenPrepare::OptimizeInlineAsmInst(CallInst *CS) {
3547 bool MadeChange = false;
3549 const TargetRegisterInfo *TRI =
3550 TM->getSubtargetImpl(*CS->getParent()->getParent())->getRegisterInfo();
3551 TargetLowering::AsmOperandInfoVector TargetConstraints =
3552 TLI->ParseConstraints(*DL, TRI, CS);
3554 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
3555 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
3557 // Compute the constraint code and ConstraintType to use.
3558 TLI->ComputeConstraintToUse(OpInfo, SDValue());
3560 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
3561 OpInfo.isIndirect) {
3562 Value *OpVal = CS->getArgOperand(ArgNo++);
3563 MadeChange |= OptimizeMemoryInst(CS, OpVal, OpVal->getType(), ~0u);
3564 } else if (OpInfo.Type == InlineAsm::isInput)
3571 /// \brief Check if all the uses of \p Inst are equivalent (or free) zero or
3572 /// sign extensions.
3573 static bool hasSameExtUse(Instruction *Inst, const TargetLowering &TLI) {
3574 assert(!Inst->use_empty() && "Input must have at least one use");
3575 const Instruction *FirstUser = cast<Instruction>(*Inst->user_begin());
3576 bool IsSExt = isa<SExtInst>(FirstUser);
3577 Type *ExtTy = FirstUser->getType();
3578 for (const User *U : Inst->users()) {
3579 const Instruction *UI = cast<Instruction>(U);
3580 if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI)))
3582 Type *CurTy = UI->getType();
3583 // Same input and output types: Same instruction after CSE.
3587 // If IsSExt is true, we are in this situation:
3589 // b = sext ty1 a to ty2
3590 // c = sext ty1 a to ty3
3591 // Assuming ty2 is shorter than ty3, this could be turned into:
3593 // b = sext ty1 a to ty2
3594 // c = sext ty2 b to ty3
3595 // However, the last sext is not free.
3599 // This is a ZExt, maybe this is free to extend from one type to another.
3600 // In that case, we would not account for a different use.
3603 if (ExtTy->getScalarType()->getIntegerBitWidth() >
3604 CurTy->getScalarType()->getIntegerBitWidth()) {
3612 if (!TLI.isZExtFree(NarrowTy, LargeTy))
3615 // All uses are the same or can be derived from one another for free.
3619 /// \brief Try to form ExtLd by promoting \p Exts until they reach a
3620 /// load instruction.
3621 /// If an ext(load) can be formed, it is returned via \p LI for the load
3622 /// and \p Inst for the extension.
3623 /// Otherwise LI == nullptr and Inst == nullptr.
3624 /// When some promotion happened, \p TPT contains the proper state to
3627 /// \return true when promoting was necessary to expose the ext(load)
3628 /// opportunity, false otherwise.
3632 /// %ld = load i32* %addr
3633 /// %add = add nuw i32 %ld, 4
3634 /// %zext = zext i32 %add to i64
3638 /// %ld = load i32* %addr
3639 /// %zext = zext i32 %ld to i64
3640 /// %add = add nuw i64 %zext, 4
3642 /// Thanks to the promotion, we can match zext(load i32*) to i64.
3643 bool CodeGenPrepare::ExtLdPromotion(TypePromotionTransaction &TPT,
3644 LoadInst *&LI, Instruction *&Inst,
3645 const SmallVectorImpl<Instruction *> &Exts,
3646 unsigned CreatedInstsCost = 0) {
3647 // Iterate over all the extensions to see if one form an ext(load).
3648 for (auto I : Exts) {
3649 // Check if we directly have ext(load).
3650 if ((LI = dyn_cast<LoadInst>(I->getOperand(0)))) {
3652 // No promotion happened here.
3655 // Check whether or not we want to do any promotion.
3656 if (!TLI || !TLI->enableExtLdPromotion() || DisableExtLdPromotion)
3658 // Get the action to perform the promotion.
3659 TypePromotionHelper::Action TPH = TypePromotionHelper::getAction(
3660 I, InsertedInsts, *TLI, PromotedInsts);
3661 // Check if we can promote.
3664 // Save the current state.
3665 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3666 TPT.getRestorationPoint();
3667 SmallVector<Instruction *, 4> NewExts;
3668 unsigned NewCreatedInstsCost = 0;
3669 unsigned ExtCost = !TLI->isExtFree(I);
3671 Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost,
3672 &NewExts, nullptr, *TLI);
3673 assert(PromotedVal &&
3674 "TypePromotionHelper should have filtered out those cases");
3676 // We would be able to merge only one extension in a load.
3677 // Therefore, if we have more than 1 new extension we heuristically
3678 // cut this search path, because it means we degrade the code quality.
3679 // With exactly 2, the transformation is neutral, because we will merge
3680 // one extension but leave one. However, we optimistically keep going,
3681 // because the new extension may be removed too.
3682 long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost;
3683 TotalCreatedInstsCost -= ExtCost;
3684 if (!StressExtLdPromotion &&
3685 (TotalCreatedInstsCost > 1 ||
3686 !isPromotedInstructionLegal(*TLI, PromotedVal))) {
3687 // The promotion is not profitable, rollback to the previous state.
3688 TPT.rollback(LastKnownGood);
3691 // The promotion is profitable.
3692 // Check if it exposes an ext(load).
3693 (void)ExtLdPromotion(TPT, LI, Inst, NewExts, TotalCreatedInstsCost);
3694 if (LI && (StressExtLdPromotion || NewCreatedInstsCost <= ExtCost ||
3695 // If we have created a new extension, i.e., now we have two
3696 // extensions. We must make sure one of them is merged with
3697 // the load, otherwise we may degrade the code quality.
3698 (LI->hasOneUse() || hasSameExtUse(LI, *TLI))))
3699 // Promotion happened.
3701 // If this does not help to expose an ext(load) then, rollback.
3702 TPT.rollback(LastKnownGood);
3704 // None of the extension can form an ext(load).
3710 /// MoveExtToFormExtLoad - Move a zext or sext fed by a load into the same
3711 /// basic block as the load, unless conditions are unfavorable. This allows
3712 /// SelectionDAG to fold the extend into the load.
3713 /// \p I[in/out] the extension may be modified during the process if some
3714 /// promotions apply.
3716 bool CodeGenPrepare::MoveExtToFormExtLoad(Instruction *&I) {
3717 // Try to promote a chain of computation if it allows to form
3718 // an extended load.
3719 TypePromotionTransaction TPT;
3720 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3721 TPT.getRestorationPoint();
3722 SmallVector<Instruction *, 1> Exts;
3724 // Look for a load being extended.
3725 LoadInst *LI = nullptr;
3726 Instruction *OldExt = I;
3727 bool HasPromoted = ExtLdPromotion(TPT, LI, I, Exts);
3729 assert(!HasPromoted && !LI && "If we did not match any load instruction "
3730 "the code must remain the same");
3735 // If they're already in the same block, there's nothing to do.
3736 // Make the cheap checks first if we did not promote.
3737 // If we promoted, we need to check if it is indeed profitable.
3738 if (!HasPromoted && LI->getParent() == I->getParent())
3741 EVT VT = TLI->getValueType(I->getType());
3742 EVT LoadVT = TLI->getValueType(LI->getType());
3744 // If the load has other users and the truncate is not free, this probably
3745 // isn't worthwhile.
3746 if (!LI->hasOneUse() && TLI &&
3747 (TLI->isTypeLegal(LoadVT) || !TLI->isTypeLegal(VT)) &&
3748 !TLI->isTruncateFree(I->getType(), LI->getType())) {
3750 TPT.rollback(LastKnownGood);
3754 // Check whether the target supports casts folded into loads.
3756 if (isa<ZExtInst>(I))
3757 LType = ISD::ZEXTLOAD;
3759 assert(isa<SExtInst>(I) && "Unexpected ext type!");
3760 LType = ISD::SEXTLOAD;
3762 if (TLI && !TLI->isLoadExtLegal(LType, VT, LoadVT)) {
3764 TPT.rollback(LastKnownGood);
3768 // Move the extend into the same block as the load, so that SelectionDAG
3771 I->removeFromParent();
3777 bool CodeGenPrepare::OptimizeExtUses(Instruction *I) {
3778 BasicBlock *DefBB = I->getParent();
3780 // If the result of a {s|z}ext and its source are both live out, rewrite all
3781 // other uses of the source with result of extension.
3782 Value *Src = I->getOperand(0);
3783 if (Src->hasOneUse())
3786 // Only do this xform if truncating is free.
3787 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
3790 // Only safe to perform the optimization if the source is also defined in
3792 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
3795 bool DefIsLiveOut = false;
3796 for (User *U : I->users()) {
3797 Instruction *UI = cast<Instruction>(U);
3799 // Figure out which BB this ext is used in.
3800 BasicBlock *UserBB = UI->getParent();
3801 if (UserBB == DefBB) continue;
3802 DefIsLiveOut = true;
3808 // Make sure none of the uses are PHI nodes.
3809 for (User *U : Src->users()) {
3810 Instruction *UI = cast<Instruction>(U);
3811 BasicBlock *UserBB = UI->getParent();
3812 if (UserBB == DefBB) continue;
3813 // Be conservative. We don't want this xform to end up introducing
3814 // reloads just before load / store instructions.
3815 if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
3819 // InsertedTruncs - Only insert one trunc in each block once.
3820 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
3822 bool MadeChange = false;
3823 for (Use &U : Src->uses()) {
3824 Instruction *User = cast<Instruction>(U.getUser());
3826 // Figure out which BB this ext is used in.
3827 BasicBlock *UserBB = User->getParent();
3828 if (UserBB == DefBB) continue;
3830 // Both src and def are live in this block. Rewrite the use.
3831 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
3833 if (!InsertedTrunc) {
3834 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
3835 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt);
3836 InsertedInsts.insert(InsertedTrunc);
3839 // Replace a use of the {s|z}ext source with a use of the result.
3848 /// isFormingBranchFromSelectProfitable - Returns true if a SelectInst should be
3849 /// turned into an explicit branch.
3850 static bool isFormingBranchFromSelectProfitable(SelectInst *SI) {
3851 // FIXME: This should use the same heuristics as IfConversion to determine
3852 // whether a select is better represented as a branch. This requires that
3853 // branch probability metadata is preserved for the select, which is not the
3856 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
3858 // If the branch is predicted right, an out of order CPU can avoid blocking on
3859 // the compare. Emit cmovs on compares with a memory operand as branches to
3860 // avoid stalls on the load from memory. If the compare has more than one use
3861 // there's probably another cmov or setcc around so it's not worth emitting a
3866 Value *CmpOp0 = Cmp->getOperand(0);
3867 Value *CmpOp1 = Cmp->getOperand(1);
3869 // We check that the memory operand has one use to avoid uses of the loaded
3870 // value directly after the compare, making branches unprofitable.
3871 return Cmp->hasOneUse() &&
3872 ((isa<LoadInst>(CmpOp0) && CmpOp0->hasOneUse()) ||
3873 (isa<LoadInst>(CmpOp1) && CmpOp1->hasOneUse()));
3877 /// If we have a SelectInst that will likely profit from branch prediction,
3878 /// turn it into a branch.
3879 bool CodeGenPrepare::OptimizeSelectInst(SelectInst *SI) {
3880 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
3882 // Can we convert the 'select' to CF ?
3883 if (DisableSelectToBranch || OptSize || !TLI || VectorCond)
3886 TargetLowering::SelectSupportKind SelectKind;
3888 SelectKind = TargetLowering::VectorMaskSelect;
3889 else if (SI->getType()->isVectorTy())
3890 SelectKind = TargetLowering::ScalarCondVectorVal;
3892 SelectKind = TargetLowering::ScalarValSelect;
3894 // Do we have efficient codegen support for this kind of 'selects' ?
3895 if (TLI->isSelectSupported(SelectKind)) {
3896 // We have efficient codegen support for the select instruction.
3897 // Check if it is profitable to keep this 'select'.
3898 if (!TLI->isPredictableSelectExpensive() ||
3899 !isFormingBranchFromSelectProfitable(SI))
3905 // First, we split the block containing the select into 2 blocks.
3906 BasicBlock *StartBlock = SI->getParent();
3907 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(SI));
3908 BasicBlock *NextBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
3910 // Create a new block serving as the landing pad for the branch.
3911 BasicBlock *SmallBlock = BasicBlock::Create(SI->getContext(), "select.mid",
3912 NextBlock->getParent(), NextBlock);
3914 // Move the unconditional branch from the block with the select in it into our
3915 // landing pad block.
3916 StartBlock->getTerminator()->eraseFromParent();
3917 BranchInst::Create(NextBlock, SmallBlock);
3919 // Insert the real conditional branch based on the original condition.
3920 BranchInst::Create(NextBlock, SmallBlock, SI->getCondition(), SI);
3922 // The select itself is replaced with a PHI Node.
3923 PHINode *PN = PHINode::Create(SI->getType(), 2, "", NextBlock->begin());
3925 PN->addIncoming(SI->getTrueValue(), StartBlock);
3926 PN->addIncoming(SI->getFalseValue(), SmallBlock);
3927 SI->replaceAllUsesWith(PN);
3928 SI->eraseFromParent();
3930 // Instruct OptimizeBlock to skip to the next block.
3931 CurInstIterator = StartBlock->end();
3932 ++NumSelectsExpanded;
3936 static bool isBroadcastShuffle(ShuffleVectorInst *SVI) {
3937 SmallVector<int, 16> Mask(SVI->getShuffleMask());
3939 for (unsigned i = 0; i < Mask.size(); ++i) {
3940 if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
3942 SplatElem = Mask[i];
3948 /// Some targets have expensive vector shifts if the lanes aren't all the same
3949 /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
3950 /// it's often worth sinking a shufflevector splat down to its use so that
3951 /// codegen can spot all lanes are identical.
3952 bool CodeGenPrepare::OptimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
3953 BasicBlock *DefBB = SVI->getParent();
3955 // Only do this xform if variable vector shifts are particularly expensive.
3956 if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
3959 // We only expect better codegen by sinking a shuffle if we can recognise a
3961 if (!isBroadcastShuffle(SVI))
3964 // InsertedShuffles - Only insert a shuffle in each block once.
3965 DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
3967 bool MadeChange = false;
3968 for (User *U : SVI->users()) {
3969 Instruction *UI = cast<Instruction>(U);
3971 // Figure out which BB this ext is used in.
3972 BasicBlock *UserBB = UI->getParent();
3973 if (UserBB == DefBB) continue;
3975 // For now only apply this when the splat is used by a shift instruction.
3976 if (!UI->isShift()) continue;
3978 // Everything checks out, sink the shuffle if the user's block doesn't
3979 // already have a copy.
3980 Instruction *&InsertedShuffle = InsertedShuffles[UserBB];
3982 if (!InsertedShuffle) {
3983 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
3984 InsertedShuffle = new ShuffleVectorInst(SVI->getOperand(0),
3986 SVI->getOperand(2), "", InsertPt);
3989 UI->replaceUsesOfWith(SVI, InsertedShuffle);
3993 // If we removed all uses, nuke the shuffle.
3994 if (SVI->use_empty()) {
3995 SVI->eraseFromParent();
4003 /// \brief Helper class to promote a scalar operation to a vector one.
4004 /// This class is used to move downward extractelement transition.
4006 /// a = vector_op <2 x i32>
4007 /// b = extractelement <2 x i32> a, i32 0
4012 /// a = vector_op <2 x i32>
4013 /// c = vector_op a (equivalent to scalar_op on the related lane)
4014 /// * d = extractelement <2 x i32> c, i32 0
4016 /// Assuming both extractelement and store can be combine, we get rid of the
4018 class VectorPromoteHelper {
4019 /// Used to perform some checks on the legality of vector operations.
4020 const TargetLowering &TLI;
4022 /// Used to estimated the cost of the promoted chain.
4023 const TargetTransformInfo &TTI;
4025 /// The transition being moved downwards.
4026 Instruction *Transition;
4027 /// The sequence of instructions to be promoted.
4028 SmallVector<Instruction *, 4> InstsToBePromoted;
4029 /// Cost of combining a store and an extract.
4030 unsigned StoreExtractCombineCost;
4031 /// Instruction that will be combined with the transition.
4032 Instruction *CombineInst;
4034 /// \brief The instruction that represents the current end of the transition.
4035 /// Since we are faking the promotion until we reach the end of the chain
4036 /// of computation, we need a way to get the current end of the transition.
4037 Instruction *getEndOfTransition() const {
4038 if (InstsToBePromoted.empty())
4040 return InstsToBePromoted.back();
4043 /// \brief Return the index of the original value in the transition.
4044 /// E.g., for "extractelement <2 x i32> c, i32 1" the original value,
4045 /// c, is at index 0.
4046 unsigned getTransitionOriginalValueIdx() const {
4047 assert(isa<ExtractElementInst>(Transition) &&
4048 "Other kind of transitions are not supported yet");
4052 /// \brief Return the index of the index in the transition.
4053 /// E.g., for "extractelement <2 x i32> c, i32 0" the index
4055 unsigned getTransitionIdx() const {
4056 assert(isa<ExtractElementInst>(Transition) &&
4057 "Other kind of transitions are not supported yet");
4061 /// \brief Get the type of the transition.
4062 /// This is the type of the original value.
4063 /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the
4064 /// transition is <2 x i32>.
4065 Type *getTransitionType() const {
4066 return Transition->getOperand(getTransitionOriginalValueIdx())->getType();
4069 /// \brief Promote \p ToBePromoted by moving \p Def downward through.
4070 /// I.e., we have the following sequence:
4071 /// Def = Transition <ty1> a to <ty2>
4072 /// b = ToBePromoted <ty2> Def, ...
4074 /// b = ToBePromoted <ty1> a, ...
4075 /// Def = Transition <ty1> ToBePromoted to <ty2>
4076 void promoteImpl(Instruction *ToBePromoted);
4078 /// \brief Check whether or not it is profitable to promote all the
4079 /// instructions enqueued to be promoted.
4080 bool isProfitableToPromote() {
4081 Value *ValIdx = Transition->getOperand(getTransitionOriginalValueIdx());
4082 unsigned Index = isa<ConstantInt>(ValIdx)
4083 ? cast<ConstantInt>(ValIdx)->getZExtValue()
4085 Type *PromotedType = getTransitionType();
4087 StoreInst *ST = cast<StoreInst>(CombineInst);
4088 unsigned AS = ST->getPointerAddressSpace();
4089 unsigned Align = ST->getAlignment();
4090 // Check if this store is supported.
4091 if (!TLI.allowsMisalignedMemoryAccesses(
4092 TLI.getValueType(ST->getValueOperand()->getType()), AS, Align)) {
4093 // If this is not supported, there is no way we can combine
4094 // the extract with the store.
4098 // The scalar chain of computation has to pay for the transition
4099 // scalar to vector.
4100 // The vector chain has to account for the combining cost.
4101 uint64_t ScalarCost =
4102 TTI.getVectorInstrCost(Transition->getOpcode(), PromotedType, Index);
4103 uint64_t VectorCost = StoreExtractCombineCost;
4104 for (const auto &Inst : InstsToBePromoted) {
4105 // Compute the cost.
4106 // By construction, all instructions being promoted are arithmetic ones.
4107 // Moreover, one argument is a constant that can be viewed as a splat
4109 Value *Arg0 = Inst->getOperand(0);
4110 bool IsArg0Constant = isa<UndefValue>(Arg0) || isa<ConstantInt>(Arg0) ||
4111 isa<ConstantFP>(Arg0);
4112 TargetTransformInfo::OperandValueKind Arg0OVK =
4113 IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
4114 : TargetTransformInfo::OK_AnyValue;
4115 TargetTransformInfo::OperandValueKind Arg1OVK =
4116 !IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
4117 : TargetTransformInfo::OK_AnyValue;
4118 ScalarCost += TTI.getArithmeticInstrCost(
4119 Inst->getOpcode(), Inst->getType(), Arg0OVK, Arg1OVK);
4120 VectorCost += TTI.getArithmeticInstrCost(Inst->getOpcode(), PromotedType,
4123 DEBUG(dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
4124 << ScalarCost << "\nVector: " << VectorCost << '\n');
4125 return ScalarCost > VectorCost;
4128 /// \brief Generate a constant vector with \p Val with the same
4129 /// number of elements as the transition.
4130 /// \p UseSplat defines whether or not \p Val should be replicated
4131 /// accross the whole vector.
4132 /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>,
4133 /// otherwise we generate a vector with as many undef as possible:
4134 /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only
4135 /// used at the index of the extract.
4136 Value *getConstantVector(Constant *Val, bool UseSplat) const {
4137 unsigned ExtractIdx = UINT_MAX;
4139 // If we cannot determine where the constant must be, we have to
4140 // use a splat constant.
4141 Value *ValExtractIdx = Transition->getOperand(getTransitionIdx());
4142 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(ValExtractIdx))
4143 ExtractIdx = CstVal->getSExtValue();
4148 unsigned End = getTransitionType()->getVectorNumElements();
4150 return ConstantVector::getSplat(End, Val);
4152 SmallVector<Constant *, 4> ConstVec;
4153 UndefValue *UndefVal = UndefValue::get(Val->getType());
4154 for (unsigned Idx = 0; Idx != End; ++Idx) {
4155 if (Idx == ExtractIdx)
4156 ConstVec.push_back(Val);
4158 ConstVec.push_back(UndefVal);
4160 return ConstantVector::get(ConstVec);
4163 /// \brief Check if promoting to a vector type an operand at \p OperandIdx
4164 /// in \p Use can trigger undefined behavior.
4165 static bool canCauseUndefinedBehavior(const Instruction *Use,
4166 unsigned OperandIdx) {
4167 // This is not safe to introduce undef when the operand is on
4168 // the right hand side of a division-like instruction.
4169 if (OperandIdx != 1)
4171 switch (Use->getOpcode()) {
4174 case Instruction::SDiv:
4175 case Instruction::UDiv:
4176 case Instruction::SRem:
4177 case Instruction::URem:
4179 case Instruction::FDiv:
4180 case Instruction::FRem:
4181 return !Use->hasNoNaNs();
4183 llvm_unreachable(nullptr);
4187 VectorPromoteHelper(const TargetLowering &TLI, const TargetTransformInfo &TTI,
4188 Instruction *Transition, unsigned CombineCost)
4189 : TLI(TLI), TTI(TTI), Transition(Transition),
4190 StoreExtractCombineCost(CombineCost), CombineInst(nullptr) {
4191 assert(Transition && "Do not know how to promote null");
4194 /// \brief Check if we can promote \p ToBePromoted to \p Type.
4195 bool canPromote(const Instruction *ToBePromoted) const {
4196 // We could support CastInst too.
4197 return isa<BinaryOperator>(ToBePromoted);
4200 /// \brief Check if it is profitable to promote \p ToBePromoted
4201 /// by moving downward the transition through.
4202 bool shouldPromote(const Instruction *ToBePromoted) const {
4203 // Promote only if all the operands can be statically expanded.
4204 // Indeed, we do not want to introduce any new kind of transitions.
4205 for (const Use &U : ToBePromoted->operands()) {
4206 const Value *Val = U.get();
4207 if (Val == getEndOfTransition()) {
4208 // If the use is a division and the transition is on the rhs,
4209 // we cannot promote the operation, otherwise we may create a
4210 // division by zero.
4211 if (canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()))
4215 if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) &&
4216 !isa<ConstantFP>(Val))
4219 // Check that the resulting operation is legal.
4220 int ISDOpcode = TLI.InstructionOpcodeToISD(ToBePromoted->getOpcode());
4223 return StressStoreExtract ||
4224 TLI.isOperationLegalOrCustom(
4225 ISDOpcode, TLI.getValueType(getTransitionType(), true));
4228 /// \brief Check whether or not \p Use can be combined
4229 /// with the transition.
4230 /// I.e., is it possible to do Use(Transition) => AnotherUse?
4231 bool canCombine(const Instruction *Use) { return isa<StoreInst>(Use); }
4233 /// \brief Record \p ToBePromoted as part of the chain to be promoted.
4234 void enqueueForPromotion(Instruction *ToBePromoted) {
4235 InstsToBePromoted.push_back(ToBePromoted);
4238 /// \brief Set the instruction that will be combined with the transition.
4239 void recordCombineInstruction(Instruction *ToBeCombined) {
4240 assert(canCombine(ToBeCombined) && "Unsupported instruction to combine");
4241 CombineInst = ToBeCombined;
4244 /// \brief Promote all the instructions enqueued for promotion if it is
4246 /// \return True if the promotion happened, false otherwise.
4248 // Check if there is something to promote.
4249 // Right now, if we do not have anything to combine with,
4250 // we assume the promotion is not profitable.
4251 if (InstsToBePromoted.empty() || !CombineInst)
4255 if (!StressStoreExtract && !isProfitableToPromote())
4259 for (auto &ToBePromoted : InstsToBePromoted)
4260 promoteImpl(ToBePromoted);
4261 InstsToBePromoted.clear();
4265 } // End of anonymous namespace.
4267 void VectorPromoteHelper::promoteImpl(Instruction *ToBePromoted) {
4268 // At this point, we know that all the operands of ToBePromoted but Def
4269 // can be statically promoted.
4270 // For Def, we need to use its parameter in ToBePromoted:
4271 // b = ToBePromoted ty1 a
4272 // Def = Transition ty1 b to ty2
4273 // Move the transition down.
4274 // 1. Replace all uses of the promoted operation by the transition.
4275 // = ... b => = ... Def.
4276 assert(ToBePromoted->getType() == Transition->getType() &&
4277 "The type of the result of the transition does not match "
4279 ToBePromoted->replaceAllUsesWith(Transition);
4280 // 2. Update the type of the uses.
4281 // b = ToBePromoted ty2 Def => b = ToBePromoted ty1 Def.
4282 Type *TransitionTy = getTransitionType();
4283 ToBePromoted->mutateType(TransitionTy);
4284 // 3. Update all the operands of the promoted operation with promoted
4286 // b = ToBePromoted ty1 Def => b = ToBePromoted ty1 a.
4287 for (Use &U : ToBePromoted->operands()) {
4288 Value *Val = U.get();
4289 Value *NewVal = nullptr;
4290 if (Val == Transition)
4291 NewVal = Transition->getOperand(getTransitionOriginalValueIdx());
4292 else if (isa<UndefValue>(Val) || isa<ConstantInt>(Val) ||
4293 isa<ConstantFP>(Val)) {
4294 // Use a splat constant if it is not safe to use undef.
4295 NewVal = getConstantVector(
4296 cast<Constant>(Val),
4297 isa<UndefValue>(Val) ||
4298 canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()));
4300 llvm_unreachable("Did you modified shouldPromote and forgot to update "
4302 ToBePromoted->setOperand(U.getOperandNo(), NewVal);
4304 Transition->removeFromParent();
4305 Transition->insertAfter(ToBePromoted);
4306 Transition->setOperand(getTransitionOriginalValueIdx(), ToBePromoted);
4309 /// Some targets can do store(extractelement) with one instruction.
4310 /// Try to push the extractelement towards the stores when the target
4311 /// has this feature and this is profitable.
4312 bool CodeGenPrepare::OptimizeExtractElementInst(Instruction *Inst) {
4313 unsigned CombineCost = UINT_MAX;
4314 if (DisableStoreExtract || !TLI ||
4315 (!StressStoreExtract &&
4316 !TLI->canCombineStoreAndExtract(Inst->getOperand(0)->getType(),
4317 Inst->getOperand(1), CombineCost)))
4320 // At this point we know that Inst is a vector to scalar transition.
4321 // Try to move it down the def-use chain, until:
4322 // - We can combine the transition with its single use
4323 // => we got rid of the transition.
4324 // - We escape the current basic block
4325 // => we would need to check that we are moving it at a cheaper place and
4326 // we do not do that for now.
4327 BasicBlock *Parent = Inst->getParent();
4328 DEBUG(dbgs() << "Found an interesting transition: " << *Inst << '\n');
4329 VectorPromoteHelper VPH(*TLI, *TTI, Inst, CombineCost);
4330 // If the transition has more than one use, assume this is not going to be
4332 while (Inst->hasOneUse()) {
4333 Instruction *ToBePromoted = cast<Instruction>(*Inst->user_begin());
4334 DEBUG(dbgs() << "Use: " << *ToBePromoted << '\n');
4336 if (ToBePromoted->getParent() != Parent) {
4337 DEBUG(dbgs() << "Instruction to promote is in a different block ("
4338 << ToBePromoted->getParent()->getName()
4339 << ") than the transition (" << Parent->getName() << ").\n");
4343 if (VPH.canCombine(ToBePromoted)) {
4344 DEBUG(dbgs() << "Assume " << *Inst << '\n'
4345 << "will be combined with: " << *ToBePromoted << '\n');
4346 VPH.recordCombineInstruction(ToBePromoted);
4347 bool Changed = VPH.promote();
4348 NumStoreExtractExposed += Changed;
4352 DEBUG(dbgs() << "Try promoting.\n");
4353 if (!VPH.canPromote(ToBePromoted) || !VPH.shouldPromote(ToBePromoted))
4356 DEBUG(dbgs() << "Promoting is possible... Enqueue for promotion!\n");
4358 VPH.enqueueForPromotion(ToBePromoted);
4359 Inst = ToBePromoted;
4364 bool CodeGenPrepare::OptimizeInst(Instruction *I, bool& ModifiedDT) {
4365 // Bail out if we inserted the instruction to prevent optimizations from
4366 // stepping on each other's toes.
4367 if (InsertedInsts.count(I))
4370 if (PHINode *P = dyn_cast<PHINode>(I)) {
4371 // It is possible for very late stage optimizations (such as SimplifyCFG)
4372 // to introduce PHI nodes too late to be cleaned up. If we detect such a
4373 // trivial PHI, go ahead and zap it here.
4374 if (Value *V = SimplifyInstruction(P, *DL, TLInfo, nullptr)) {
4375 P->replaceAllUsesWith(V);
4376 P->eraseFromParent();
4383 if (CastInst *CI = dyn_cast<CastInst>(I)) {
4384 // If the source of the cast is a constant, then this should have
4385 // already been constant folded. The only reason NOT to constant fold
4386 // it is if something (e.g. LSR) was careful to place the constant
4387 // evaluation in a block other than then one that uses it (e.g. to hoist
4388 // the address of globals out of a loop). If this is the case, we don't
4389 // want to forward-subst the cast.
4390 if (isa<Constant>(CI->getOperand(0)))
4393 if (TLI && OptimizeNoopCopyExpression(CI, *TLI))
4396 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
4397 /// Sink a zext or sext into its user blocks if the target type doesn't
4398 /// fit in one register
4399 if (TLI && TLI->getTypeAction(CI->getContext(),
4400 TLI->getValueType(CI->getType())) ==
4401 TargetLowering::TypeExpandInteger) {
4402 return SinkCast(CI);
4404 bool MadeChange = MoveExtToFormExtLoad(I);
4405 return MadeChange | OptimizeExtUses(I);
4411 if (CmpInst *CI = dyn_cast<CmpInst>(I))
4412 if (!TLI || !TLI->hasMultipleConditionRegisters())
4413 return OptimizeCmpExpression(CI);
4415 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
4417 unsigned AS = LI->getPointerAddressSpace();
4418 return OptimizeMemoryInst(I, I->getOperand(0), LI->getType(), AS);
4423 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
4425 unsigned AS = SI->getPointerAddressSpace();
4426 return OptimizeMemoryInst(I, SI->getOperand(1),
4427 SI->getOperand(0)->getType(), AS);
4432 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(I);
4434 if (BinOp && (BinOp->getOpcode() == Instruction::AShr ||
4435 BinOp->getOpcode() == Instruction::LShr)) {
4436 ConstantInt *CI = dyn_cast<ConstantInt>(BinOp->getOperand(1));
4437 if (TLI && CI && TLI->hasExtractBitsInsn())
4438 return OptimizeExtractBits(BinOp, CI, *TLI);
4443 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
4444 if (GEPI->hasAllZeroIndices()) {
4445 /// The GEP operand must be a pointer, so must its result -> BitCast
4446 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
4447 GEPI->getName(), GEPI);
4448 GEPI->replaceAllUsesWith(NC);
4449 GEPI->eraseFromParent();
4451 OptimizeInst(NC, ModifiedDT);
4457 if (CallInst *CI = dyn_cast<CallInst>(I))
4458 return OptimizeCallInst(CI, ModifiedDT);
4460 if (SelectInst *SI = dyn_cast<SelectInst>(I))
4461 return OptimizeSelectInst(SI);
4463 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I))
4464 return OptimizeShuffleVectorInst(SVI);
4466 if (isa<ExtractElementInst>(I))
4467 return OptimizeExtractElementInst(I);
4472 // In this pass we look for GEP and cast instructions that are used
4473 // across basic blocks and rewrite them to improve basic-block-at-a-time
4475 bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB, bool& ModifiedDT) {
4477 bool MadeChange = false;
4479 CurInstIterator = BB.begin();
4480 while (CurInstIterator != BB.end()) {
4481 MadeChange |= OptimizeInst(CurInstIterator++, ModifiedDT);
4485 MadeChange |= DupRetToEnableTailCallOpts(&BB);
4490 // llvm.dbg.value is far away from the value then iSel may not be able
4491 // handle it properly. iSel will drop llvm.dbg.value if it can not
4492 // find a node corresponding to the value.
4493 bool CodeGenPrepare::PlaceDbgValues(Function &F) {
4494 bool MadeChange = false;
4495 for (BasicBlock &BB : F) {
4496 Instruction *PrevNonDbgInst = nullptr;
4497 for (BasicBlock::iterator BI = BB.begin(), BE = BB.end(); BI != BE;) {
4498 Instruction *Insn = BI++;
4499 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
4500 // Leave dbg.values that refer to an alloca alone. These
4501 // instrinsics describe the address of a variable (= the alloca)
4502 // being taken. They should not be moved next to the alloca
4503 // (and to the beginning of the scope), but rather stay close to
4504 // where said address is used.
4505 if (!DVI || (DVI->getValue() && isa<AllocaInst>(DVI->getValue()))) {
4506 PrevNonDbgInst = Insn;
4510 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
4511 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
4512 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
4513 DVI->removeFromParent();
4514 if (isa<PHINode>(VI))
4515 DVI->insertBefore(VI->getParent()->getFirstInsertionPt());
4517 DVI->insertAfter(VI);
4526 // If there is a sequence that branches based on comparing a single bit
4527 // against zero that can be combined into a single instruction, and the
4528 // target supports folding these into a single instruction, sink the
4529 // mask and compare into the branch uses. Do this before OptimizeBlock ->
4530 // OptimizeInst -> OptimizeCmpExpression, which perturbs the pattern being
4532 bool CodeGenPrepare::sinkAndCmp(Function &F) {
4533 if (!EnableAndCmpSinking)
4535 if (!TLI || !TLI->isMaskAndBranchFoldingLegal())
4537 bool MadeChange = false;
4538 for (Function::iterator I = F.begin(), E = F.end(); I != E; ) {
4539 BasicBlock *BB = I++;
4541 // Does this BB end with the following?
4542 // %andVal = and %val, #single-bit-set
4543 // %icmpVal = icmp %andResult, 0
4544 // br i1 %cmpVal label %dest1, label %dest2"
4545 BranchInst *Brcc = dyn_cast<BranchInst>(BB->getTerminator());
4546 if (!Brcc || !Brcc->isConditional())
4548 ICmpInst *Cmp = dyn_cast<ICmpInst>(Brcc->getOperand(0));
4549 if (!Cmp || Cmp->getParent() != BB)
4551 ConstantInt *Zero = dyn_cast<ConstantInt>(Cmp->getOperand(1));
4552 if (!Zero || !Zero->isZero())
4554 Instruction *And = dyn_cast<Instruction>(Cmp->getOperand(0));
4555 if (!And || And->getOpcode() != Instruction::And || And->getParent() != BB)
4557 ConstantInt* Mask = dyn_cast<ConstantInt>(And->getOperand(1));
4558 if (!Mask || !Mask->getUniqueInteger().isPowerOf2())
4560 DEBUG(dbgs() << "found and; icmp ?,0; brcc\n"); DEBUG(BB->dump());
4562 // Push the "and; icmp" for any users that are conditional branches.
4563 // Since there can only be one branch use per BB, we don't need to keep
4564 // track of which BBs we insert into.
4565 for (Value::use_iterator UI = Cmp->use_begin(), E = Cmp->use_end();
4569 BranchInst *BrccUser = dyn_cast<BranchInst>(*UI);
4571 if (!BrccUser || !BrccUser->isConditional())
4573 BasicBlock *UserBB = BrccUser->getParent();
4574 if (UserBB == BB) continue;
4575 DEBUG(dbgs() << "found Brcc use\n");
4577 // Sink the "and; icmp" to use.
4579 BinaryOperator *NewAnd =
4580 BinaryOperator::CreateAnd(And->getOperand(0), And->getOperand(1), "",
4583 CmpInst::Create(Cmp->getOpcode(), Cmp->getPredicate(), NewAnd, Zero,
4587 DEBUG(BrccUser->getParent()->dump());
4593 /// \brief Retrieve the probabilities of a conditional branch. Returns true on
4594 /// success, or returns false if no or invalid metadata was found.
4595 static bool extractBranchMetadata(BranchInst *BI,
4596 uint64_t &ProbTrue, uint64_t &ProbFalse) {
4597 assert(BI->isConditional() &&
4598 "Looking for probabilities on unconditional branch?");
4599 auto *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
4600 if (!ProfileData || ProfileData->getNumOperands() != 3)
4603 const auto *CITrue =
4604 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(1));
4605 const auto *CIFalse =
4606 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(2));
4607 if (!CITrue || !CIFalse)
4610 ProbTrue = CITrue->getValue().getZExtValue();
4611 ProbFalse = CIFalse->getValue().getZExtValue();
4616 /// \brief Scale down both weights to fit into uint32_t.
4617 static void scaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) {
4618 uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse;
4619 uint32_t Scale = (NewMax / UINT32_MAX) + 1;
4620 NewTrue = NewTrue / Scale;
4621 NewFalse = NewFalse / Scale;
4624 /// \brief Some targets prefer to split a conditional branch like:
4626 /// %0 = icmp ne i32 %a, 0
4627 /// %1 = icmp ne i32 %b, 0
4628 /// %or.cond = or i1 %0, %1
4629 /// br i1 %or.cond, label %TrueBB, label %FalseBB
4631 /// into multiple branch instructions like:
4634 /// %0 = icmp ne i32 %a, 0
4635 /// br i1 %0, label %TrueBB, label %bb2
4637 /// %1 = icmp ne i32 %b, 0
4638 /// br i1 %1, label %TrueBB, label %FalseBB
4640 /// This usually allows instruction selection to do even further optimizations
4641 /// and combine the compare with the branch instruction. Currently this is
4642 /// applied for targets which have "cheap" jump instructions.
4644 /// FIXME: Remove the (equivalent?) implementation in SelectionDAG.
4646 bool CodeGenPrepare::splitBranchCondition(Function &F) {
4647 if (!TM || !TM->Options.EnableFastISel || !TLI || TLI->isJumpExpensive())
4650 bool MadeChange = false;
4651 for (auto &BB : F) {
4652 // Does this BB end with the following?
4653 // %cond1 = icmp|fcmp|binary instruction ...
4654 // %cond2 = icmp|fcmp|binary instruction ...
4655 // %cond.or = or|and i1 %cond1, cond2
4656 // br i1 %cond.or label %dest1, label %dest2"
4657 BinaryOperator *LogicOp;
4658 BasicBlock *TBB, *FBB;
4659 if (!match(BB.getTerminator(), m_Br(m_OneUse(m_BinOp(LogicOp)), TBB, FBB)))
4663 Value *Cond1, *Cond2;
4664 if (match(LogicOp, m_And(m_OneUse(m_Value(Cond1)),
4665 m_OneUse(m_Value(Cond2)))))
4666 Opc = Instruction::And;
4667 else if (match(LogicOp, m_Or(m_OneUse(m_Value(Cond1)),
4668 m_OneUse(m_Value(Cond2)))))
4669 Opc = Instruction::Or;
4673 if (!match(Cond1, m_CombineOr(m_Cmp(), m_BinOp())) ||
4674 !match(Cond2, m_CombineOr(m_Cmp(), m_BinOp())) )
4677 DEBUG(dbgs() << "Before branch condition splitting\n"; BB.dump());
4680 auto *InsertBefore = std::next(Function::iterator(BB))
4681 .getNodePtrUnchecked();
4682 auto TmpBB = BasicBlock::Create(BB.getContext(),
4683 BB.getName() + ".cond.split",
4684 BB.getParent(), InsertBefore);
4686 // Update original basic block by using the first condition directly by the
4687 // branch instruction and removing the no longer needed and/or instruction.
4688 auto *Br1 = cast<BranchInst>(BB.getTerminator());
4689 Br1->setCondition(Cond1);
4690 LogicOp->eraseFromParent();
4692 // Depending on the conditon we have to either replace the true or the false
4693 // successor of the original branch instruction.
4694 if (Opc == Instruction::And)
4695 Br1->setSuccessor(0, TmpBB);
4697 Br1->setSuccessor(1, TmpBB);
4699 // Fill in the new basic block.
4700 auto *Br2 = IRBuilder<>(TmpBB).CreateCondBr(Cond2, TBB, FBB);
4701 if (auto *I = dyn_cast<Instruction>(Cond2)) {
4702 I->removeFromParent();
4703 I->insertBefore(Br2);
4706 // Update PHI nodes in both successors. The original BB needs to be
4707 // replaced in one succesor's PHI nodes, because the branch comes now from
4708 // the newly generated BB (NewBB). In the other successor we need to add one
4709 // incoming edge to the PHI nodes, because both branch instructions target
4710 // now the same successor. Depending on the original branch condition
4711 // (and/or) we have to swap the successors (TrueDest, FalseDest), so that
4712 // we perfrom the correct update for the PHI nodes.
4713 // This doesn't change the successor order of the just created branch
4714 // instruction (or any other instruction).
4715 if (Opc == Instruction::Or)
4716 std::swap(TBB, FBB);
4718 // Replace the old BB with the new BB.
4719 for (auto &I : *TBB) {
4720 PHINode *PN = dyn_cast<PHINode>(&I);
4724 while ((i = PN->getBasicBlockIndex(&BB)) >= 0)
4725 PN->setIncomingBlock(i, TmpBB);
4728 // Add another incoming edge form the new BB.
4729 for (auto &I : *FBB) {
4730 PHINode *PN = dyn_cast<PHINode>(&I);
4733 auto *Val = PN->getIncomingValueForBlock(&BB);
4734 PN->addIncoming(Val, TmpBB);
4737 // Update the branch weights (from SelectionDAGBuilder::
4738 // FindMergedConditions).
4739 if (Opc == Instruction::Or) {
4740 // Codegen X | Y as:
4749 // We have flexibility in setting Prob for BB1 and Prob for NewBB.
4750 // The requirement is that
4751 // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB)
4752 // = TrueProb for orignal BB.
4753 // Assuming the orignal weights are A and B, one choice is to set BB1's
4754 // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice
4756 // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB.
4757 // Another choice is to assume TrueProb for BB1 equals to TrueProb for
4758 // TmpBB, but the math is more complicated.
4759 uint64_t TrueWeight, FalseWeight;
4760 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
4761 uint64_t NewTrueWeight = TrueWeight;
4762 uint64_t NewFalseWeight = TrueWeight + 2 * FalseWeight;
4763 scaleWeights(NewTrueWeight, NewFalseWeight);
4764 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
4765 .createBranchWeights(TrueWeight, FalseWeight));
4767 NewTrueWeight = TrueWeight;
4768 NewFalseWeight = 2 * FalseWeight;
4769 scaleWeights(NewTrueWeight, NewFalseWeight);
4770 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
4771 .createBranchWeights(TrueWeight, FalseWeight));
4774 // Codegen X & Y as:
4782 // This requires creation of TmpBB after CurBB.
4784 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
4785 // The requirement is that
4786 // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB)
4787 // = FalseProb for orignal BB.
4788 // Assuming the orignal weights are A and B, one choice is to set BB1's
4789 // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice
4791 // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB.
4792 uint64_t TrueWeight, FalseWeight;
4793 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
4794 uint64_t NewTrueWeight = 2 * TrueWeight + FalseWeight;
4795 uint64_t NewFalseWeight = FalseWeight;
4796 scaleWeights(NewTrueWeight, NewFalseWeight);
4797 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
4798 .createBranchWeights(TrueWeight, FalseWeight));
4800 NewTrueWeight = 2 * TrueWeight;
4801 NewFalseWeight = FalseWeight;
4802 scaleWeights(NewTrueWeight, NewFalseWeight);
4803 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
4804 .createBranchWeights(TrueWeight, FalseWeight));
4808 // Note: No point in getting fancy here, since the DT info is never
4809 // available to CodeGenPrepare.
4814 DEBUG(dbgs() << "After branch condition splitting\n"; BB.dump();