LLVM  7.0.0svn
CodeGenPrepare.cpp
Go to the documentation of this file.
1 //===- CodeGenPrepare.cpp - Prepare a function for code generation --------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
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.
13 //
14 //===----------------------------------------------------------------------===//
15 
16 #include "llvm/ADT/APInt.h"
17 #include "llvm/ADT/ArrayRef.h"
18 #include "llvm/ADT/DenseMap.h"
20 #include "llvm/ADT/STLExtras.h"
21 #include "llvm/ADT/SmallPtrSet.h"
22 #include "llvm/ADT/SmallVector.h"
23 #include "llvm/ADT/Statistic.h"
28 #include "llvm/Analysis/LoopInfo.h"
34 #include "llvm/CodeGen/Analysis.h"
42 #include "llvm/IR/Argument.h"
43 #include "llvm/IR/Attributes.h"
44 #include "llvm/IR/BasicBlock.h"
45 #include "llvm/IR/CallSite.h"
46 #include "llvm/IR/Constant.h"
47 #include "llvm/IR/Constants.h"
48 #include "llvm/IR/DataLayout.h"
49 #include "llvm/IR/DerivedTypes.h"
50 #include "llvm/IR/Dominators.h"
51 #include "llvm/IR/Function.h"
53 #include "llvm/IR/GlobalValue.h"
54 #include "llvm/IR/GlobalVariable.h"
55 #include "llvm/IR/IRBuilder.h"
56 #include "llvm/IR/InlineAsm.h"
57 #include "llvm/IR/InstrTypes.h"
58 #include "llvm/IR/Instruction.h"
59 #include "llvm/IR/Instructions.h"
60 #include "llvm/IR/IntrinsicInst.h"
61 #include "llvm/IR/Intrinsics.h"
62 #include "llvm/IR/LLVMContext.h"
63 #include "llvm/IR/MDBuilder.h"
64 #include "llvm/IR/Module.h"
65 #include "llvm/IR/Operator.h"
66 #include "llvm/IR/PatternMatch.h"
67 #include "llvm/IR/Statepoint.h"
68 #include "llvm/IR/Type.h"
69 #include "llvm/IR/Use.h"
70 #include "llvm/IR/User.h"
71 #include "llvm/IR/Value.h"
72 #include "llvm/IR/ValueHandle.h"
73 #include "llvm/IR/ValueMap.h"
74 #include "llvm/Pass.h"
77 #include "llvm/Support/Casting.h"
79 #include "llvm/Support/Compiler.h"
80 #include "llvm/Support/Debug.h"
90 #include <algorithm>
91 #include <cassert>
92 #include <cstdint>
93 #include <iterator>
94 #include <limits>
95 #include <memory>
96 #include <utility>
97 #include <vector>
98 
99 using namespace llvm;
100 using namespace llvm::PatternMatch;
101 
102 #define DEBUG_TYPE "codegenprepare"
103 
104 STATISTIC(NumBlocksElim, "Number of blocks eliminated");
105 STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated");
106 STATISTIC(NumGEPsElim, "Number of GEPs converted to casts");
107 STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "
108  "sunken Cmps");
109 STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "
110  "of sunken Casts");
111 STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "
112  "computations were sunk");
113 STATISTIC(NumMemoryInstsPhiCreated,
114  "Number of phis created when address "
115  "computations were sunk to memory instructions");
116 STATISTIC(NumMemoryInstsSelectCreated,
117  "Number of select created when address "
118  "computations were sunk to memory instructions");
119 STATISTIC(NumExtsMoved, "Number of [s|z]ext instructions combined with loads");
120 STATISTIC(NumExtUses, "Number of uses of [s|z]ext instructions optimized");
121 STATISTIC(NumAndsAdded,
122  "Number of and mask instructions added to form ext loads");
123 STATISTIC(NumAndUses, "Number of uses of and mask instructions optimized");
124 STATISTIC(NumRetsDup, "Number of return instructions duplicated");
125 STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved");
126 STATISTIC(NumSelectsExpanded, "Number of selects turned into branches");
127 STATISTIC(NumStoreExtractExposed, "Number of store(extractelement) exposed");
128 
130  "disable-cgp-branch-opts", cl::Hidden, cl::init(false),
131  cl::desc("Disable branch optimizations in CodeGenPrepare"));
132 
133 static cl::opt<bool>
134  DisableGCOpts("disable-cgp-gc-opts", cl::Hidden, cl::init(false),
135  cl::desc("Disable GC optimizations in CodeGenPrepare"));
136 
138  "disable-cgp-select2branch", cl::Hidden, cl::init(false),
139  cl::desc("Disable select to branch conversion."));
140 
142  "addr-sink-using-gep", cl::Hidden, cl::init(true),
143  cl::desc("Address sinking in CGP using GEPs."));
144 
146  "enable-andcmp-sinking", cl::Hidden, cl::init(true),
147  cl::desc("Enable sinkinig and/cmp into branches."));
148 
150  "disable-cgp-store-extract", cl::Hidden, cl::init(false),
151  cl::desc("Disable store(extract) optimizations in CodeGenPrepare"));
152 
154  "stress-cgp-store-extract", cl::Hidden, cl::init(false),
155  cl::desc("Stress test store(extract) optimizations in CodeGenPrepare"));
156 
158  "disable-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
159  cl::desc("Disable ext(promotable(ld)) -> promoted(ext(ld)) optimization in "
160  "CodeGenPrepare"));
161 
163  "stress-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
164  cl::desc("Stress test ext(promotable(ld)) -> promoted(ext(ld)) "
165  "optimization in CodeGenPrepare"));
166 
168  "disable-preheader-prot", cl::Hidden, cl::init(false),
169  cl::desc("Disable protection against removing loop preheaders"));
170 
172  "profile-guided-section-prefix", cl::Hidden, cl::init(true), cl::ZeroOrMore,
173  cl::desc("Use profile info to add section prefix for hot/cold functions"));
174 
176  "cgp-freq-ratio-to-skip-merge", cl::Hidden, cl::init(2),
177  cl::desc("Skip merging empty blocks if (frequency of empty block) / "
178  "(frequency of destination block) is greater than this ratio"));
179 
181  "force-split-store", cl::Hidden, cl::init(false),
182  cl::desc("Force store splitting no matter what the target query says."));
183 
184 static cl::opt<bool>
185 EnableTypePromotionMerge("cgp-type-promotion-merge", cl::Hidden,
186  cl::desc("Enable merging of redundant sexts when one is dominating"
187  " the other."), cl::init(true));
188 
190  "disable-complex-addr-modes", cl::Hidden, cl::init(false),
191  cl::desc("Disables combining addressing modes with different parts "
192  "in optimizeMemoryInst."));
193 
194 static cl::opt<bool>
195 AddrSinkNewPhis("addr-sink-new-phis", cl::Hidden, cl::init(false),
196  cl::desc("Allow creation of Phis in Address sinking."));
197 
198 static cl::opt<bool>
199 AddrSinkNewSelects("addr-sink-new-select", cl::Hidden, cl::init(false),
200  cl::desc("Allow creation of selects in Address sinking."));
201 
203  "addr-sink-combine-base-reg", cl::Hidden, cl::init(true),
204  cl::desc("Allow combining of BaseReg field in Address sinking."));
205 
207  "addr-sink-combine-base-gv", cl::Hidden, cl::init(true),
208  cl::desc("Allow combining of BaseGV field in Address sinking."));
209 
211  "addr-sink-combine-base-offs", cl::Hidden, cl::init(true),
212  cl::desc("Allow combining of BaseOffs field in Address sinking."));
213 
215  "addr-sink-combine-scaled-reg", cl::Hidden, cl::init(true),
216  cl::desc("Allow combining of ScaledReg field in Address sinking."));
217 
218 namespace {
219 
220 using SetOfInstrs = SmallPtrSet<Instruction *, 16>;
221 using TypeIsSExt = PointerIntPair<Type *, 1, bool>;
222 using InstrToOrigTy = DenseMap<Instruction *, TypeIsSExt>;
223 using SExts = SmallVector<Instruction *, 16>;
224 using ValueToSExts = DenseMap<Value *, SExts>;
225 
226 class TypePromotionTransaction;
227 
228  class CodeGenPrepare : public FunctionPass {
229  const TargetMachine *TM = nullptr;
230  const TargetSubtargetInfo *SubtargetInfo;
231  const TargetLowering *TLI = nullptr;
232  const TargetRegisterInfo *TRI;
233  const TargetTransformInfo *TTI = nullptr;
234  const TargetLibraryInfo *TLInfo;
235  const LoopInfo *LI;
236  std::unique_ptr<BlockFrequencyInfo> BFI;
237  std::unique_ptr<BranchProbabilityInfo> BPI;
238 
239  /// As we scan instructions optimizing them, this is the next instruction
240  /// to optimize. Transforms that can invalidate this should update it.
241  BasicBlock::iterator CurInstIterator;
242 
243  /// Keeps track of non-local addresses that have been sunk into a block.
244  /// This allows us to avoid inserting duplicate code for blocks with
245  /// multiple load/stores of the same address. The usage of WeakTrackingVH
246  /// enables SunkAddrs to be treated as a cache whose entries can be
247  /// invalidated if a sunken address computation has been erased.
249 
250  /// Keeps track of all instructions inserted for the current function.
251  SetOfInstrs InsertedInsts;
252 
253  /// Keeps track of the type of the related instruction before their
254  /// promotion for the current function.
255  InstrToOrigTy PromotedInsts;
256 
257  /// Keep track of instructions removed during promotion.
258  SetOfInstrs RemovedInsts;
259 
260  /// Keep track of sext chains based on their initial value.
261  DenseMap<Value *, Instruction *> SeenChainsForSExt;
262 
263  /// Keep track of SExt promoted.
264  ValueToSExts ValToSExtendedUses;
265 
266  /// True if CFG is modified in any way.
267  bool ModifiedDT;
268 
269  /// True if optimizing for size.
270  bool OptSize;
271 
272  /// DataLayout for the Function being processed.
273  const DataLayout *DL = nullptr;
274 
275  public:
276  static char ID; // Pass identification, replacement for typeid
277 
278  CodeGenPrepare() : FunctionPass(ID) {
280  }
281 
282  bool runOnFunction(Function &F) override;
283 
284  StringRef getPassName() const override { return "CodeGen Prepare"; }
285 
286  void getAnalysisUsage(AnalysisUsage &AU) const override {
287  // FIXME: When we can selectively preserve passes, preserve the domtree.
292  }
293 
294  private:
295  bool eliminateFallThrough(Function &F);
296  bool eliminateMostlyEmptyBlocks(Function &F);
297  BasicBlock *findDestBlockOfMergeableEmptyBlock(BasicBlock *BB);
298  bool canMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
299  void eliminateMostlyEmptyBlock(BasicBlock *BB);
300  bool isMergingEmptyBlockProfitable(BasicBlock *BB, BasicBlock *DestBB,
301  bool isPreheader);
302  bool optimizeBlock(BasicBlock &BB, bool &ModifiedDT);
303  bool optimizeInst(Instruction *I, bool &ModifiedDT);
304  bool optimizeMemoryInst(Instruction *I, Value *Addr,
305  Type *AccessTy, unsigned AS);
306  bool optimizeInlineAsmInst(CallInst *CS);
307  bool optimizeCallInst(CallInst *CI, bool &ModifiedDT);
308  bool optimizeExt(Instruction *&I);
309  bool optimizeExtUses(Instruction *I);
310  bool optimizeLoadExt(LoadInst *I);
311  bool optimizeSelectInst(SelectInst *SI);
312  bool optimizeShuffleVectorInst(ShuffleVectorInst *SI);
313  bool optimizeSwitchInst(SwitchInst *CI);
314  bool optimizeExtractElementInst(Instruction *Inst);
315  bool dupRetToEnableTailCallOpts(BasicBlock *BB);
316  bool placeDbgValues(Function &F);
317  bool canFormExtLd(const SmallVectorImpl<Instruction *> &MovedExts,
318  LoadInst *&LI, Instruction *&Inst, bool HasPromoted);
319  bool tryToPromoteExts(TypePromotionTransaction &TPT,
320  const SmallVectorImpl<Instruction *> &Exts,
321  SmallVectorImpl<Instruction *> &ProfitablyMovedExts,
322  unsigned CreatedInstsCost = 0);
323  bool mergeSExts(Function &F);
324  bool performAddressTypePromotion(
325  Instruction *&Inst,
326  bool AllowPromotionWithoutCommonHeader,
327  bool HasPromoted, TypePromotionTransaction &TPT,
328  SmallVectorImpl<Instruction *> &SpeculativelyMovedExts);
329  bool splitBranchCondition(Function &F);
330  bool simplifyOffsetableRelocate(Instruction &I);
331  };
332 
333 } // end anonymous namespace
334 
335 char CodeGenPrepare::ID = 0;
336 
337 INITIALIZE_PASS_BEGIN(CodeGenPrepare, DEBUG_TYPE,
338  "Optimize for code generation", false, false)
341  "Optimize for code generation", false, false)
342 
343 FunctionPass *llvm::createCodeGenPreparePass() { return new CodeGenPrepare(); }
344 
346  if (skipFunction(F))
347  return false;
348 
349  DL = &F.getParent()->getDataLayout();
350 
351  bool EverMadeChange = false;
352  // Clear per function information.
353  InsertedInsts.clear();
354  PromotedInsts.clear();
355 
356  ModifiedDT = false;
357  if (auto *TPC = getAnalysisIfAvailable<TargetPassConfig>()) {
358  TM = &TPC->getTM<TargetMachine>();
359  SubtargetInfo = TM->getSubtargetImpl(F);
360  TLI = SubtargetInfo->getTargetLowering();
361  TRI = SubtargetInfo->getRegisterInfo();
362  }
363  TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
364  TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
365  LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
366  BPI.reset(new BranchProbabilityInfo(F, *LI));
367  BFI.reset(new BlockFrequencyInfo(F, *BPI, *LI));
368  OptSize = F.optForSize();
369 
370  ProfileSummaryInfo *PSI =
371  getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
373  if (PSI->isFunctionHotInCallGraph(&F, *BFI))
374  F.setSectionPrefix(".hot");
375  else if (PSI->isFunctionColdInCallGraph(&F, *BFI))
376  F.setSectionPrefix(".unlikely");
377  }
378 
379  /// This optimization identifies DIV instructions that can be
380  /// profitably bypassed and carried out with a shorter, faster divide.
381  if (!OptSize && !PSI->hasHugeWorkingSetSize() && TLI &&
382  TLI->isSlowDivBypassed()) {
383  const DenseMap<unsigned int, unsigned int> &BypassWidths =
384  TLI->getBypassSlowDivWidths();
385  BasicBlock* BB = &*F.begin();
386  while (BB != nullptr) {
387  // bypassSlowDivision may create new BBs, but we don't want to reapply the
388  // optimization to those blocks.
389  BasicBlock* Next = BB->getNextNode();
390  EverMadeChange |= bypassSlowDivision(BB, BypassWidths);
391  BB = Next;
392  }
393  }
394 
395  // Eliminate blocks that contain only PHI nodes and an
396  // unconditional branch.
397  EverMadeChange |= eliminateMostlyEmptyBlocks(F);
398 
399  // llvm.dbg.value is far away from the value then iSel may not be able
400  // handle it properly. iSel will drop llvm.dbg.value if it can not
401  // find a node corresponding to the value.
402  EverMadeChange |= placeDbgValues(F);
403 
404  if (!DisableBranchOpts)
405  EverMadeChange |= splitBranchCondition(F);
406 
407  // Split some critical edges where one of the sources is an indirect branch,
408  // to help generate sane code for PHIs involving such edges.
409  EverMadeChange |= SplitIndirectBrCriticalEdges(F);
410 
411  bool MadeChange = true;
412  while (MadeChange) {
413  MadeChange = false;
414  SeenChainsForSExt.clear();
415  ValToSExtendedUses.clear();
416  RemovedInsts.clear();
417  for (Function::iterator I = F.begin(); I != F.end(); ) {
418  BasicBlock *BB = &*I++;
419  bool ModifiedDTOnIteration = false;
420  MadeChange |= optimizeBlock(*BB, ModifiedDTOnIteration);
421 
422  // Restart BB iteration if the dominator tree of the Function was changed
423  if (ModifiedDTOnIteration)
424  break;
425  }
426  if (EnableTypePromotionMerge && !ValToSExtendedUses.empty())
427  MadeChange |= mergeSExts(F);
428 
429  // Really free removed instructions during promotion.
430  for (Instruction *I : RemovedInsts)
431  I->deleteValue();
432 
433  EverMadeChange |= MadeChange;
434  }
435 
436  SunkAddrs.clear();
437 
438  if (!DisableBranchOpts) {
439  MadeChange = false;
441  for (BasicBlock &BB : F) {
442  SmallVector<BasicBlock *, 2> Successors(succ_begin(&BB), succ_end(&BB));
443  MadeChange |= ConstantFoldTerminator(&BB, true);
444  if (!MadeChange) continue;
445 
447  II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
448  if (pred_begin(*II) == pred_end(*II))
449  WorkList.insert(*II);
450  }
451 
452  // Delete the dead blocks and any of their dead successors.
453  MadeChange |= !WorkList.empty();
454  while (!WorkList.empty()) {
455  BasicBlock *BB = *WorkList.begin();
456  WorkList.erase(BB);
457  SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
458 
459  DeleteDeadBlock(BB);
460 
462  II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
463  if (pred_begin(*II) == pred_end(*II))
464  WorkList.insert(*II);
465  }
466 
467  // Merge pairs of basic blocks with unconditional branches, connected by
468  // a single edge.
469  if (EverMadeChange || MadeChange)
470  MadeChange |= eliminateFallThrough(F);
471 
472  EverMadeChange |= MadeChange;
473  }
474 
475  if (!DisableGCOpts) {
476  SmallVector<Instruction *, 2> Statepoints;
477  for (BasicBlock &BB : F)
478  for (Instruction &I : BB)
479  if (isStatepoint(I))
480  Statepoints.push_back(&I);
481  for (auto &I : Statepoints)
482  EverMadeChange |= simplifyOffsetableRelocate(*I);
483  }
484 
485  return EverMadeChange;
486 }
487 
488 /// Merge basic blocks which are connected by a single edge, where one of the
489 /// basic blocks has a single successor pointing to the other basic block,
490 /// which has a single predecessor.
491 bool CodeGenPrepare::eliminateFallThrough(Function &F) {
492  bool Changed = false;
493  // Scan all of the blocks in the function, except for the entry block.
494  for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
495  BasicBlock *BB = &*I++;
496  // If the destination block has a single pred, then this is a trivial
497  // edge, just collapse it.
498  BasicBlock *SinglePred = BB->getSinglePredecessor();
499 
500  // Don't merge if BB's address is taken.
501  if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
502 
503  BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
504  if (Term && !Term->isConditional()) {
505  Changed = true;
506  DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n");
507  // Remember if SinglePred was the entry block of the function.
508  // If so, we will need to move BB back to the entry position.
509  bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
510  MergeBasicBlockIntoOnlyPred(BB, nullptr);
511 
512  if (isEntry && BB != &BB->getParent()->getEntryBlock())
513  BB->moveBefore(&BB->getParent()->getEntryBlock());
514 
515  // We have erased a block. Update the iterator.
516  I = BB->getIterator();
517  }
518  }
519  return Changed;
520 }
521 
522 /// Find a destination block from BB if BB is mergeable empty block.
523 BasicBlock *CodeGenPrepare::findDestBlockOfMergeableEmptyBlock(BasicBlock *BB) {
524  // If this block doesn't end with an uncond branch, ignore it.
526  if (!BI || !BI->isUnconditional())
527  return nullptr;
528 
529  // If the instruction before the branch (skipping debug info) isn't a phi
530  // node, then other stuff is happening here.
531  BasicBlock::iterator BBI = BI->getIterator();
532  if (BBI != BB->begin()) {
533  --BBI;
534  while (isa<DbgInfoIntrinsic>(BBI)) {
535  if (BBI == BB->begin())
536  break;
537  --BBI;
538  }
539  if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
540  return nullptr;
541  }
542 
543  // Do not break infinite loops.
544  BasicBlock *DestBB = BI->getSuccessor(0);
545  if (DestBB == BB)
546  return nullptr;
547 
548  if (!canMergeBlocks(BB, DestBB))
549  DestBB = nullptr;
550 
551  return DestBB;
552 }
553 
554 /// Eliminate blocks that contain only PHI nodes, debug info directives, and an
555 /// unconditional branch. Passes before isel (e.g. LSR/loopsimplify) often split
556 /// edges in ways that are non-optimal for isel. Start by eliminating these
557 /// blocks so we can split them the way we want them.
558 bool CodeGenPrepare::eliminateMostlyEmptyBlocks(Function &F) {
560  SmallVector<Loop *, 16> LoopList(LI->begin(), LI->end());
561  while (!LoopList.empty()) {
562  Loop *L = LoopList.pop_back_val();
563  LoopList.insert(LoopList.end(), L->begin(), L->end());
564  if (BasicBlock *Preheader = L->getLoopPreheader())
565  Preheaders.insert(Preheader);
566  }
567 
568  bool MadeChange = false;
569  // Note that this intentionally skips the entry block.
570  for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
571  BasicBlock *BB = &*I++;
572  BasicBlock *DestBB = findDestBlockOfMergeableEmptyBlock(BB);
573  if (!DestBB ||
574  !isMergingEmptyBlockProfitable(BB, DestBB, Preheaders.count(BB)))
575  continue;
576 
577  eliminateMostlyEmptyBlock(BB);
578  MadeChange = true;
579  }
580  return MadeChange;
581 }
582 
583 bool CodeGenPrepare::isMergingEmptyBlockProfitable(BasicBlock *BB,
584  BasicBlock *DestBB,
585  bool isPreheader) {
586  // Do not delete loop preheaders if doing so would create a critical edge.
587  // Loop preheaders can be good locations to spill registers. If the
588  // preheader is deleted and we create a critical edge, registers may be
589  // spilled in the loop body instead.
590  if (!DisablePreheaderProtect && isPreheader &&
591  !(BB->getSinglePredecessor() &&
593  return false;
594 
595  // Try to skip merging if the unique predecessor of BB is terminated by a
596  // switch or indirect branch instruction, and BB is used as an incoming block
597  // of PHIs in DestBB. In such case, merging BB and DestBB would cause ISel to
598  // add COPY instructions in the predecessor of BB instead of BB (if it is not
599  // merged). Note that the critical edge created by merging such blocks wont be
600  // split in MachineSink because the jump table is not analyzable. By keeping
601  // such empty block (BB), ISel will place COPY instructions in BB, not in the
602  // predecessor of BB.
603  BasicBlock *Pred = BB->getUniquePredecessor();
604  if (!Pred ||
605  !(isa<SwitchInst>(Pred->getTerminator()) ||
606  isa<IndirectBrInst>(Pred->getTerminator())))
607  return true;
608 
609  if (BB->getTerminator() != BB->getFirstNonPHI())
610  return true;
611 
612  // We use a simple cost heuristic which determine skipping merging is
613  // profitable if the cost of skipping merging is less than the cost of
614  // merging : Cost(skipping merging) < Cost(merging BB), where the
615  // Cost(skipping merging) is Freq(BB) * (Cost(Copy) + Cost(Branch)), and
616  // the Cost(merging BB) is Freq(Pred) * Cost(Copy).
617  // Assuming Cost(Copy) == Cost(Branch), we could simplify it to :
618  // Freq(Pred) / Freq(BB) > 2.
619  // Note that if there are multiple empty blocks sharing the same incoming
620  // value for the PHIs in the DestBB, we consider them together. In such
621  // case, Cost(merging BB) will be the sum of their frequencies.
622 
623  if (!isa<PHINode>(DestBB->begin()))
624  return true;
625 
626  SmallPtrSet<BasicBlock *, 16> SameIncomingValueBBs;
627 
628  // Find all other incoming blocks from which incoming values of all PHIs in
629  // DestBB are the same as the ones from BB.
630  for (pred_iterator PI = pred_begin(DestBB), E = pred_end(DestBB); PI != E;
631  ++PI) {
632  BasicBlock *DestBBPred = *PI;
633  if (DestBBPred == BB)
634  continue;
635 
636  if (llvm::all_of(DestBB->phis(), [&](const PHINode &DestPN) {
637  return DestPN.getIncomingValueForBlock(BB) ==
638  DestPN.getIncomingValueForBlock(DestBBPred);
639  }))
640  SameIncomingValueBBs.insert(DestBBPred);
641  }
642 
643  // See if all BB's incoming values are same as the value from Pred. In this
644  // case, no reason to skip merging because COPYs are expected to be place in
645  // Pred already.
646  if (SameIncomingValueBBs.count(Pred))
647  return true;
648 
649  BlockFrequency PredFreq = BFI->getBlockFreq(Pred);
650  BlockFrequency BBFreq = BFI->getBlockFreq(BB);
651 
652  for (auto SameValueBB : SameIncomingValueBBs)
653  if (SameValueBB->getUniquePredecessor() == Pred &&
654  DestBB == findDestBlockOfMergeableEmptyBlock(SameValueBB))
655  BBFreq += BFI->getBlockFreq(SameValueBB);
656 
657  return PredFreq.getFrequency() <=
659 }
660 
661 /// Return true if we can merge BB into DestBB if there is a single
662 /// unconditional branch between them, and BB contains no other non-phi
663 /// instructions.
664 bool CodeGenPrepare::canMergeBlocks(const BasicBlock *BB,
665  const BasicBlock *DestBB) const {
666  // We only want to eliminate blocks whose phi nodes are used by phi nodes in
667  // the successor. If there are more complex condition (e.g. preheaders),
668  // don't mess around with them.
669  for (const PHINode &PN : BB->phis()) {
670  for (const User *U : PN.users()) {
671  const Instruction *UI = cast<Instruction>(U);
672  if (UI->getParent() != DestBB || !isa<PHINode>(UI))
673  return false;
674  // If User is inside DestBB block and it is a PHINode then check
675  // incoming value. If incoming value is not from BB then this is
676  // a complex condition (e.g. preheaders) we want to avoid here.
677  if (UI->getParent() == DestBB) {
678  if (const PHINode *UPN = dyn_cast<PHINode>(UI))
679  for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
680  Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
681  if (Insn && Insn->getParent() == BB &&
682  Insn->getParent() != UPN->getIncomingBlock(I))
683  return false;
684  }
685  }
686  }
687  }
688 
689  // If BB and DestBB contain any common predecessors, then the phi nodes in BB
690  // and DestBB may have conflicting incoming values for the block. If so, we
691  // can't merge the block.
692  const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
693  if (!DestBBPN) return true; // no conflict.
694 
695  // Collect the preds of BB.
697  if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
698  // It is faster to get preds from a PHI than with pred_iterator.
699  for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
700  BBPreds.insert(BBPN->getIncomingBlock(i));
701  } else {
702  BBPreds.insert(pred_begin(BB), pred_end(BB));
703  }
704 
705  // Walk the preds of DestBB.
706  for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
707  BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
708  if (BBPreds.count(Pred)) { // Common predecessor?
709  for (const PHINode &PN : DestBB->phis()) {
710  const Value *V1 = PN.getIncomingValueForBlock(Pred);
711  const Value *V2 = PN.getIncomingValueForBlock(BB);
712 
713  // If V2 is a phi node in BB, look up what the mapped value will be.
714  if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
715  if (V2PN->getParent() == BB)
716  V2 = V2PN->getIncomingValueForBlock(Pred);
717 
718  // If there is a conflict, bail out.
719  if (V1 != V2) return false;
720  }
721  }
722  }
723 
724  return true;
725 }
726 
727 /// Eliminate a basic block that has only phi's and an unconditional branch in
728 /// it.
729 void CodeGenPrepare::eliminateMostlyEmptyBlock(BasicBlock *BB) {
730  BranchInst *BI = cast<BranchInst>(BB->getTerminator());
731  BasicBlock *DestBB = BI->getSuccessor(0);
732 
733  DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
734 
735  // If the destination block has a single pred, then this is a trivial edge,
736  // just collapse it.
737  if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
738  if (SinglePred != DestBB) {
739  // Remember if SinglePred was the entry block of the function. If so, we
740  // will need to move BB back to the entry position.
741  bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
742  MergeBasicBlockIntoOnlyPred(DestBB, nullptr);
743 
744  if (isEntry && BB != &BB->getParent()->getEntryBlock())
745  BB->moveBefore(&BB->getParent()->getEntryBlock());
746 
747  DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
748  return;
749  }
750  }
751 
752  // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
753  // to handle the new incoming edges it is about to have.
754  for (PHINode &PN : DestBB->phis()) {
755  // Remove the incoming value for BB, and remember it.
756  Value *InVal = PN.removeIncomingValue(BB, false);
757 
758  // Two options: either the InVal is a phi node defined in BB or it is some
759  // value that dominates BB.
760  PHINode *InValPhi = dyn_cast<PHINode>(InVal);
761  if (InValPhi && InValPhi->getParent() == BB) {
762  // Add all of the input values of the input PHI as inputs of this phi.
763  for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
764  PN.addIncoming(InValPhi->getIncomingValue(i),
765  InValPhi->getIncomingBlock(i));
766  } else {
767  // Otherwise, add one instance of the dominating value for each edge that
768  // we will be adding.
769  if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
770  for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
771  PN.addIncoming(InVal, BBPN->getIncomingBlock(i));
772  } else {
773  for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
774  PN.addIncoming(InVal, *PI);
775  }
776  }
777  }
778 
779  // The PHIs are now updated, change everything that refers to BB to use
780  // DestBB and remove BB.
781  BB->replaceAllUsesWith(DestBB);
782  BB->eraseFromParent();
783  ++NumBlocksElim;
784 
785  DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
786 }
787 
788 // Computes a map of base pointer relocation instructions to corresponding
789 // derived pointer relocation instructions given a vector of all relocate calls
791  const SmallVectorImpl<GCRelocateInst *> &AllRelocateCalls,
793  &RelocateInstMap) {
794  // Collect information in two maps: one primarily for locating the base object
795  // while filling the second map; the second map is the final structure holding
796  // a mapping between Base and corresponding Derived relocate calls
798  for (auto *ThisRelocate : AllRelocateCalls) {
799  auto K = std::make_pair(ThisRelocate->getBasePtrIndex(),
800  ThisRelocate->getDerivedPtrIndex());
801  RelocateIdxMap.insert(std::make_pair(K, ThisRelocate));
802  }
803  for (auto &Item : RelocateIdxMap) {
804  std::pair<unsigned, unsigned> Key = Item.first;
805  if (Key.first == Key.second)
806  // Base relocation: nothing to insert
807  continue;
808 
809  GCRelocateInst *I = Item.second;
810  auto BaseKey = std::make_pair(Key.first, Key.first);
811 
812  // We're iterating over RelocateIdxMap so we cannot modify it.
813  auto MaybeBase = RelocateIdxMap.find(BaseKey);
814  if (MaybeBase == RelocateIdxMap.end())
815  // TODO: We might want to insert a new base object relocate and gep off
816  // that, if there are enough derived object relocates.
817  continue;
818 
819  RelocateInstMap[MaybeBase->second].push_back(I);
820  }
821 }
822 
823 // Accepts a GEP and extracts the operands into a vector provided they're all
824 // small integer constants
826  SmallVectorImpl<Value *> &OffsetV) {
827  for (unsigned i = 1; i < GEP->getNumOperands(); i++) {
828  // Only accept small constant integer operands
829  auto Op = dyn_cast<ConstantInt>(GEP->getOperand(i));
830  if (!Op || Op->getZExtValue() > 20)
831  return false;
832  }
833 
834  for (unsigned i = 1; i < GEP->getNumOperands(); i++)
835  OffsetV.push_back(GEP->getOperand(i));
836  return true;
837 }
838 
839 // Takes a RelocatedBase (base pointer relocation instruction) and Targets to
840 // replace, computes a replacement, and affects it.
841 static bool
843  const SmallVectorImpl<GCRelocateInst *> &Targets) {
844  bool MadeChange = false;
845  // We must ensure the relocation of derived pointer is defined after
846  // relocation of base pointer. If we find a relocation corresponding to base
847  // defined earlier than relocation of base then we move relocation of base
848  // right before found relocation. We consider only relocation in the same
849  // basic block as relocation of base. Relocations from other basic block will
850  // be skipped by optimization and we do not care about them.
851  for (auto R = RelocatedBase->getParent()->getFirstInsertionPt();
852  &*R != RelocatedBase; ++R)
853  if (auto RI = dyn_cast<GCRelocateInst>(R))
854  if (RI->getStatepoint() == RelocatedBase->getStatepoint())
855  if (RI->getBasePtrIndex() == RelocatedBase->getBasePtrIndex()) {
856  RelocatedBase->moveBefore(RI);
857  break;
858  }
859 
860  for (GCRelocateInst *ToReplace : Targets) {
861  assert(ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex() &&
862  "Not relocating a derived object of the original base object");
863  if (ToReplace->getBasePtrIndex() == ToReplace->getDerivedPtrIndex()) {
864  // A duplicate relocate call. TODO: coalesce duplicates.
865  continue;
866  }
867 
868  if (RelocatedBase->getParent() != ToReplace->getParent()) {
869  // Base and derived relocates are in different basic blocks.
870  // In this case transform is only valid when base dominates derived
871  // relocate. However it would be too expensive to check dominance
872  // for each such relocate, so we skip the whole transformation.
873  continue;
874  }
875 
876  Value *Base = ToReplace->getBasePtr();
877  auto Derived = dyn_cast<GetElementPtrInst>(ToReplace->getDerivedPtr());
878  if (!Derived || Derived->getPointerOperand() != Base)
879  continue;
880 
881  SmallVector<Value *, 2> OffsetV;
882  if (!getGEPSmallConstantIntOffsetV(Derived, OffsetV))
883  continue;
884 
885  // Create a Builder and replace the target callsite with a gep
886  assert(RelocatedBase->getNextNode() &&
887  "Should always have one since it's not a terminator");
888 
889  // Insert after RelocatedBase
890  IRBuilder<> Builder(RelocatedBase->getNextNode());
891  Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
892 
893  // If gc_relocate does not match the actual type, cast it to the right type.
894  // In theory, there must be a bitcast after gc_relocate if the type does not
895  // match, and we should reuse it to get the derived pointer. But it could be
896  // cases like this:
897  // bb1:
898  // ...
899  // %g1 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
900  // br label %merge
901  //
902  // bb2:
903  // ...
904  // %g2 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
905  // br label %merge
906  //
907  // merge:
908  // %p1 = phi i8 addrspace(1)* [ %g1, %bb1 ], [ %g2, %bb2 ]
909  // %cast = bitcast i8 addrspace(1)* %p1 in to i32 addrspace(1)*
910  //
911  // In this case, we can not find the bitcast any more. So we insert a new bitcast
912  // no matter there is already one or not. In this way, we can handle all cases, and
913  // the extra bitcast should be optimized away in later passes.
914  Value *ActualRelocatedBase = RelocatedBase;
915  if (RelocatedBase->getType() != Base->getType()) {
916  ActualRelocatedBase =
917  Builder.CreateBitCast(RelocatedBase, Base->getType());
918  }
919  Value *Replacement = Builder.CreateGEP(
920  Derived->getSourceElementType(), ActualRelocatedBase, makeArrayRef(OffsetV));
921  Replacement->takeName(ToReplace);
922  // If the newly generated derived pointer's type does not match the original derived
923  // pointer's type, cast the new derived pointer to match it. Same reasoning as above.
924  Value *ActualReplacement = Replacement;
925  if (Replacement->getType() != ToReplace->getType()) {
926  ActualReplacement =
927  Builder.CreateBitCast(Replacement, ToReplace->getType());
928  }
929  ToReplace->replaceAllUsesWith(ActualReplacement);
930  ToReplace->eraseFromParent();
931 
932  MadeChange = true;
933  }
934  return MadeChange;
935 }
936 
937 // Turns this:
938 //
939 // %base = ...
940 // %ptr = gep %base + 15
941 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
942 // %base' = relocate(%tok, i32 4, i32 4)
943 // %ptr' = relocate(%tok, i32 4, i32 5)
944 // %val = load %ptr'
945 //
946 // into this:
947 //
948 // %base = ...
949 // %ptr = gep %base + 15
950 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
951 // %base' = gc.relocate(%tok, i32 4, i32 4)
952 // %ptr' = gep %base' + 15
953 // %val = load %ptr'
954 bool CodeGenPrepare::simplifyOffsetableRelocate(Instruction &I) {
955  bool MadeChange = false;
956  SmallVector<GCRelocateInst *, 2> AllRelocateCalls;
957 
958  for (auto *U : I.users())
959  if (GCRelocateInst *Relocate = dyn_cast<GCRelocateInst>(U))
960  // Collect all the relocate calls associated with a statepoint
961  AllRelocateCalls.push_back(Relocate);
962 
963  // We need atleast one base pointer relocation + one derived pointer
964  // relocation to mangle
965  if (AllRelocateCalls.size() < 2)
966  return false;
967 
968  // RelocateInstMap is a mapping from the base relocate instruction to the
969  // corresponding derived relocate instructions
971  computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap);
972  if (RelocateInstMap.empty())
973  return false;
974 
975  for (auto &Item : RelocateInstMap)
976  // Item.first is the RelocatedBase to offset against
977  // Item.second is the vector of Targets to replace
978  MadeChange = simplifyRelocatesOffABase(Item.first, Item.second);
979  return MadeChange;
980 }
981 
982 /// SinkCast - Sink the specified cast instruction into its user blocks
983 static bool SinkCast(CastInst *CI) {
984  BasicBlock *DefBB = CI->getParent();
985 
986  /// InsertedCasts - Only insert a cast in each block once.
987  DenseMap<BasicBlock*, CastInst*> InsertedCasts;
988 
989  bool MadeChange = false;
990  for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
991  UI != E; ) {
992  Use &TheUse = UI.getUse();
993  Instruction *User = cast<Instruction>(*UI);
994 
995  // Figure out which BB this cast is used in. For PHI's this is the
996  // appropriate predecessor block.
997  BasicBlock *UserBB = User->getParent();
998  if (PHINode *PN = dyn_cast<PHINode>(User)) {
999  UserBB = PN->getIncomingBlock(TheUse);
1000  }
1001 
1002  // Preincrement use iterator so we don't invalidate it.
1003  ++UI;
1004 
1005  // The first insertion point of a block containing an EH pad is after the
1006  // pad. If the pad is the user, we cannot sink the cast past the pad.
1007  if (User->isEHPad())
1008  continue;
1009 
1010  // If the block selected to receive the cast is an EH pad that does not
1011  // allow non-PHI instructions before the terminator, we can't sink the
1012  // cast.
1013  if (UserBB->getTerminator()->isEHPad())
1014  continue;
1015 
1016  // If this user is in the same block as the cast, don't change the cast.
1017  if (UserBB == DefBB) continue;
1018 
1019  // If we have already inserted a cast into this block, use it.
1020  CastInst *&InsertedCast = InsertedCasts[UserBB];
1021 
1022  if (!InsertedCast) {
1023  BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1024  assert(InsertPt != UserBB->end());
1025  InsertedCast = CastInst::Create(CI->getOpcode(), CI->getOperand(0),
1026  CI->getType(), "", &*InsertPt);
1027  }
1028 
1029  // Replace a use of the cast with a use of the new cast.
1030  TheUse = InsertedCast;
1031  MadeChange = true;
1032  ++NumCastUses;
1033  }
1034 
1035  // If we removed all uses, nuke the cast.
1036  if (CI->use_empty()) {
1037  salvageDebugInfo(*CI);
1038  CI->eraseFromParent();
1039  MadeChange = true;
1040  }
1041 
1042  return MadeChange;
1043 }
1044 
1045 /// If the specified cast instruction is a noop copy (e.g. it's casting from
1046 /// one pointer type to another, i32->i8 on PPC), sink it into user blocks to
1047 /// reduce the number of virtual registers that must be created and coalesced.
1048 ///
1049 /// Return true if any changes are made.
1051  const DataLayout &DL) {
1052  // Sink only "cheap" (or nop) address-space casts. This is a weaker condition
1053  // than sinking only nop casts, but is helpful on some platforms.
1054  if (auto *ASC = dyn_cast<AddrSpaceCastInst>(CI)) {
1055  if (!TLI.isCheapAddrSpaceCast(ASC->getSrcAddressSpace(),
1056  ASC->getDestAddressSpace()))
1057  return false;
1058  }
1059 
1060  // If this is a noop copy,
1061  EVT SrcVT = TLI.getValueType(DL, CI->getOperand(0)->getType());
1062  EVT DstVT = TLI.getValueType(DL, CI->getType());
1063 
1064  // This is an fp<->int conversion?
1065  if (SrcVT.isInteger() != DstVT.isInteger())
1066  return false;
1067 
1068  // If this is an extension, it will be a zero or sign extension, which
1069  // isn't a noop.
1070  if (SrcVT.bitsLT(DstVT)) return false;
1071 
1072  // If these values will be promoted, find out what they will be promoted
1073  // to. This helps us consider truncates on PPC as noop copies when they
1074  // are.
1075  if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
1077  SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
1078  if (TLI.getTypeAction(CI->getContext(), DstVT) ==
1080  DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
1081 
1082  // If, after promotion, these are the same types, this is a noop copy.
1083  if (SrcVT != DstVT)
1084  return false;
1085 
1086  return SinkCast(CI);
1087 }
1088 
1089 /// Try to combine CI into a call to the llvm.uadd.with.overflow intrinsic if
1090 /// possible.
1091 ///
1092 /// Return true if any changes were made.
1094  Value *A, *B;
1095  Instruction *AddI;
1096  if (!match(CI,
1098  return false;
1099 
1100  Type *Ty = AddI->getType();
1101  if (!isa<IntegerType>(Ty))
1102  return false;
1103 
1104  // We don't want to move around uses of condition values this late, so we we
1105  // check if it is legal to create the call to the intrinsic in the basic
1106  // block containing the icmp:
1107 
1108  if (AddI->getParent() != CI->getParent() && !AddI->hasOneUse())
1109  return false;
1110 
1111 #ifndef NDEBUG
1112  // Someday m_UAddWithOverflow may get smarter, but this is a safe assumption
1113  // for now:
1114  if (AddI->hasOneUse())
1115  assert(*AddI->user_begin() == CI && "expected!");
1116 #endif
1117 
1118  Module *M = CI->getModule();
1119  Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
1120 
1121  auto *InsertPt = AddI->hasOneUse() ? CI : AddI;
1122 
1123  auto *UAddWithOverflow =
1124  CallInst::Create(F, {A, B}, "uadd.overflow", InsertPt);
1125  auto *UAdd = ExtractValueInst::Create(UAddWithOverflow, 0, "uadd", InsertPt);
1126  auto *Overflow =
1127  ExtractValueInst::Create(UAddWithOverflow, 1, "overflow", InsertPt);
1128 
1129  CI->replaceAllUsesWith(Overflow);
1130  AddI->replaceAllUsesWith(UAdd);
1131  CI->eraseFromParent();
1132  AddI->eraseFromParent();
1133  return true;
1134 }
1135 
1136 /// Sink the given CmpInst into user blocks to reduce the number of virtual
1137 /// registers that must be created and coalesced. This is a clear win except on
1138 /// targets with multiple condition code registers (PowerPC), where it might
1139 /// lose; some adjustment may be wanted there.
1140 ///
1141 /// Return true if any changes are made.
1142 static bool SinkCmpExpression(CmpInst *CI, const TargetLowering *TLI) {
1143  BasicBlock *DefBB = CI->getParent();
1144 
1145  // Avoid sinking soft-FP comparisons, since this can move them into a loop.
1146  if (TLI && TLI->useSoftFloat() && isa<FCmpInst>(CI))
1147  return false;
1148 
1149  // Only insert a cmp in each block once.
1150  DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
1151 
1152  bool MadeChange = false;
1153  for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
1154  UI != E; ) {
1155  Use &TheUse = UI.getUse();
1156  Instruction *User = cast<Instruction>(*UI);
1157 
1158  // Preincrement use iterator so we don't invalidate it.
1159  ++UI;
1160 
1161  // Don't bother for PHI nodes.
1162  if (isa<PHINode>(User))
1163  continue;
1164 
1165  // Figure out which BB this cmp is used in.
1166  BasicBlock *UserBB = User->getParent();
1167 
1168  // If this user is in the same block as the cmp, don't change the cmp.
1169  if (UserBB == DefBB) continue;
1170 
1171  // If we have already inserted a cmp into this block, use it.
1172  CmpInst *&InsertedCmp = InsertedCmps[UserBB];
1173 
1174  if (!InsertedCmp) {
1175  BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1176  assert(InsertPt != UserBB->end());
1177  InsertedCmp =
1178  CmpInst::Create(CI->getOpcode(), CI->getPredicate(),
1179  CI->getOperand(0), CI->getOperand(1), "", &*InsertPt);
1180  // Propagate the debug info.
1181  InsertedCmp->setDebugLoc(CI->getDebugLoc());
1182  }
1183 
1184  // Replace a use of the cmp with a use of the new cmp.
1185  TheUse = InsertedCmp;
1186  MadeChange = true;
1187  ++NumCmpUses;
1188  }
1189 
1190  // If we removed all uses, nuke the cmp.
1191  if (CI->use_empty()) {
1192  CI->eraseFromParent();
1193  MadeChange = true;
1194  }
1195 
1196  return MadeChange;
1197 }
1198 
1199 static bool OptimizeCmpExpression(CmpInst *CI, const TargetLowering *TLI) {
1200  if (SinkCmpExpression(CI, TLI))
1201  return true;
1202 
1203  if (CombineUAddWithOverflow(CI))
1204  return true;
1205 
1206  return false;
1207 }
1208 
1209 /// Duplicate and sink the given 'and' instruction into user blocks where it is
1210 /// used in a compare to allow isel to generate better code for targets where
1211 /// this operation can be combined.
1212 ///
1213 /// Return true if any changes are made.
1215  const TargetLowering &TLI,
1216  SetOfInstrs &InsertedInsts) {
1217  // Double-check that we're not trying to optimize an instruction that was
1218  // already optimized by some other part of this pass.
1219  assert(!InsertedInsts.count(AndI) &&
1220  "Attempting to optimize already optimized and instruction");
1221  (void) InsertedInsts;
1222 
1223  // Nothing to do for single use in same basic block.
1224  if (AndI->hasOneUse() &&
1225  AndI->getParent() == cast<Instruction>(*AndI->user_begin())->getParent())
1226  return false;
1227 
1228  // Try to avoid cases where sinking/duplicating is likely to increase register
1229  // pressure.
1230  if (!isa<ConstantInt>(AndI->getOperand(0)) &&
1231  !isa<ConstantInt>(AndI->getOperand(1)) &&
1232  AndI->getOperand(0)->hasOneUse() && AndI->getOperand(1)->hasOneUse())
1233  return false;
1234 
1235  for (auto *U : AndI->users()) {
1236  Instruction *User = cast<Instruction>(U);
1237 
1238  // Only sink for and mask feeding icmp with 0.
1239  if (!isa<ICmpInst>(User))
1240  return false;
1241 
1242  auto *CmpC = dyn_cast<ConstantInt>(User->getOperand(1));
1243  if (!CmpC || !CmpC->isZero())
1244  return false;
1245  }
1246 
1247  if (!TLI.isMaskAndCmp0FoldingBeneficial(*AndI))
1248  return false;
1249 
1250  DEBUG(dbgs() << "found 'and' feeding only icmp 0;\n");
1251  DEBUG(AndI->getParent()->dump());
1252 
1253  // Push the 'and' into the same block as the icmp 0. There should only be
1254  // one (icmp (and, 0)) in each block, since CSE/GVN should have removed any
1255  // others, so we don't need to keep track of which BBs we insert into.
1256  for (Value::user_iterator UI = AndI->user_begin(), E = AndI->user_end();
1257  UI != E; ) {
1258  Use &TheUse = UI.getUse();
1259  Instruction *User = cast<Instruction>(*UI);
1260 
1261  // Preincrement use iterator so we don't invalidate it.
1262  ++UI;
1263 
1264  DEBUG(dbgs() << "sinking 'and' use: " << *User << "\n");
1265 
1266  // Keep the 'and' in the same place if the use is already in the same block.
1267  Instruction *InsertPt =
1268  User->getParent() == AndI->getParent() ? AndI : User;
1269  Instruction *InsertedAnd =
1270  BinaryOperator::Create(Instruction::And, AndI->getOperand(0),
1271  AndI->getOperand(1), "", InsertPt);
1272  // Propagate the debug info.
1273  InsertedAnd->setDebugLoc(AndI->getDebugLoc());
1274 
1275  // Replace a use of the 'and' with a use of the new 'and'.
1276  TheUse = InsertedAnd;
1277  ++NumAndUses;
1278  DEBUG(User->getParent()->dump());
1279  }
1280 
1281  // We removed all uses, nuke the and.
1282  AndI->eraseFromParent();
1283  return true;
1284 }
1285 
1286 /// Check if the candidates could be combined with a shift instruction, which
1287 /// includes:
1288 /// 1. Truncate instruction
1289 /// 2. And instruction and the imm is a mask of the low bits:
1290 /// imm & (imm+1) == 0
1292  if (!isa<TruncInst>(User)) {
1293  if (User->getOpcode() != Instruction::And ||
1294  !isa<ConstantInt>(User->getOperand(1)))
1295  return false;
1296 
1297  const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
1298 
1299  if ((Cimm & (Cimm + 1)).getBoolValue())
1300  return false;
1301  }
1302  return true;
1303 }
1304 
1305 /// Sink both shift and truncate instruction to the use of truncate's BB.
1306 static bool
1309  const TargetLowering &TLI, const DataLayout &DL) {
1310  BasicBlock *UserBB = User->getParent();
1311  DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
1312  TruncInst *TruncI = dyn_cast<TruncInst>(User);
1313  bool MadeChange = false;
1314 
1315  for (Value::user_iterator TruncUI = TruncI->user_begin(),
1316  TruncE = TruncI->user_end();
1317  TruncUI != TruncE;) {
1318 
1319  Use &TruncTheUse = TruncUI.getUse();
1320  Instruction *TruncUser = cast<Instruction>(*TruncUI);
1321  // Preincrement use iterator so we don't invalidate it.
1322 
1323  ++TruncUI;
1324 
1325  int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
1326  if (!ISDOpcode)
1327  continue;
1328 
1329  // If the use is actually a legal node, there will not be an
1330  // implicit truncate.
1331  // FIXME: always querying the result type is just an
1332  // approximation; some nodes' legality is determined by the
1333  // operand or other means. There's no good way to find out though.
1334  if (TLI.isOperationLegalOrCustom(
1335  ISDOpcode, TLI.getValueType(DL, TruncUser->getType(), true)))
1336  continue;
1337 
1338  // Don't bother for PHI nodes.
1339  if (isa<PHINode>(TruncUser))
1340  continue;
1341 
1342  BasicBlock *TruncUserBB = TruncUser->getParent();
1343 
1344  if (UserBB == TruncUserBB)
1345  continue;
1346 
1347  BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
1348  CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
1349 
1350  if (!InsertedShift && !InsertedTrunc) {
1351  BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
1352  assert(InsertPt != TruncUserBB->end());
1353  // Sink the shift
1354  if (ShiftI->getOpcode() == Instruction::AShr)
1355  InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
1356  "", &*InsertPt);
1357  else
1358  InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
1359  "", &*InsertPt);
1360 
1361  // Sink the trunc
1362  BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
1363  TruncInsertPt++;
1364  assert(TruncInsertPt != TruncUserBB->end());
1365 
1366  InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
1367  TruncI->getType(), "", &*TruncInsertPt);
1368 
1369  MadeChange = true;
1370 
1371  TruncTheUse = InsertedTrunc;
1372  }
1373  }
1374  return MadeChange;
1375 }
1376 
1377 /// Sink the shift *right* instruction into user blocks if the uses could
1378 /// potentially be combined with this shift instruction and generate BitExtract
1379 /// instruction. It will only be applied if the architecture supports BitExtract
1380 /// instruction. Here is an example:
1381 /// BB1:
1382 /// %x.extract.shift = lshr i64 %arg1, 32
1383 /// BB2:
1384 /// %x.extract.trunc = trunc i64 %x.extract.shift to i16
1385 /// ==>
1386 ///
1387 /// BB2:
1388 /// %x.extract.shift.1 = lshr i64 %arg1, 32
1389 /// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
1390 ///
1391 /// CodeGen will recoginze the pattern in BB2 and generate BitExtract
1392 /// instruction.
1393 /// Return true if any changes are made.
1395  const TargetLowering &TLI,
1396  const DataLayout &DL) {
1397  BasicBlock *DefBB = ShiftI->getParent();
1398 
1399  /// Only insert instructions in each block once.
1401 
1402  bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(DL, ShiftI->getType()));
1403 
1404  bool MadeChange = false;
1405  for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
1406  UI != E;) {
1407  Use &TheUse = UI.getUse();
1408  Instruction *User = cast<Instruction>(*UI);
1409  // Preincrement use iterator so we don't invalidate it.
1410  ++UI;
1411 
1412  // Don't bother for PHI nodes.
1413  if (isa<PHINode>(User))
1414  continue;
1415 
1416  if (!isExtractBitsCandidateUse(User))
1417  continue;
1418 
1419  BasicBlock *UserBB = User->getParent();
1420 
1421  if (UserBB == DefBB) {
1422  // If the shift and truncate instruction are in the same BB. The use of
1423  // the truncate(TruncUse) may still introduce another truncate if not
1424  // legal. In this case, we would like to sink both shift and truncate
1425  // instruction to the BB of TruncUse.
1426  // for example:
1427  // BB1:
1428  // i64 shift.result = lshr i64 opnd, imm
1429  // trunc.result = trunc shift.result to i16
1430  //
1431  // BB2:
1432  // ----> We will have an implicit truncate here if the architecture does
1433  // not have i16 compare.
1434  // cmp i16 trunc.result, opnd2
1435  //
1436  if (isa<TruncInst>(User) && shiftIsLegal
1437  // If the type of the truncate is legal, no trucate will be
1438  // introduced in other basic blocks.
1439  &&
1440  (!TLI.isTypeLegal(TLI.getValueType(DL, User->getType()))))
1441  MadeChange =
1442  SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI, DL);
1443 
1444  continue;
1445  }
1446  // If we have already inserted a shift into this block, use it.
1447  BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
1448 
1449  if (!InsertedShift) {
1450  BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1451  assert(InsertPt != UserBB->end());
1452 
1453  if (ShiftI->getOpcode() == Instruction::AShr)
1454  InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
1455  "", &*InsertPt);
1456  else
1457  InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
1458  "", &*InsertPt);
1459 
1460  MadeChange = true;
1461  }
1462 
1463  // Replace a use of the shift with a use of the new shift.
1464  TheUse = InsertedShift;
1465  }
1466 
1467  // If we removed all uses, nuke the shift.
1468  if (ShiftI->use_empty())
1469  ShiftI->eraseFromParent();
1470 
1471  return MadeChange;
1472 }
1473 
1474 /// If counting leading or trailing zeros is an expensive operation and a zero
1475 /// input is defined, add a check for zero to avoid calling the intrinsic.
1476 ///
1477 /// We want to transform:
1478 /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 false)
1479 ///
1480 /// into:
1481 /// entry:
1482 /// %cmpz = icmp eq i64 %A, 0
1483 /// br i1 %cmpz, label %cond.end, label %cond.false
1484 /// cond.false:
1485 /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 true)
1486 /// br label %cond.end
1487 /// cond.end:
1488 /// %ctz = phi i64 [ 64, %entry ], [ %z, %cond.false ]
1489 ///
1490 /// If the transform is performed, return true and set ModifiedDT to true.
1491 static bool despeculateCountZeros(IntrinsicInst *CountZeros,
1492  const TargetLowering *TLI,
1493  const DataLayout *DL,
1494  bool &ModifiedDT) {
1495  if (!TLI || !DL)
1496  return false;
1497 
1498  // If a zero input is undefined, it doesn't make sense to despeculate that.
1499  if (match(CountZeros->getOperand(1), m_One()))
1500  return false;
1501 
1502  // If it's cheap to speculate, there's nothing to do.
1503  auto IntrinsicID = CountZeros->getIntrinsicID();
1504  if ((IntrinsicID == Intrinsic::cttz && TLI->isCheapToSpeculateCttz()) ||
1505  (IntrinsicID == Intrinsic::ctlz && TLI->isCheapToSpeculateCtlz()))
1506  return false;
1507 
1508  // Only handle legal scalar cases. Anything else requires too much work.
1509  Type *Ty = CountZeros->getType();
1510  unsigned SizeInBits = Ty->getPrimitiveSizeInBits();
1511  if (Ty->isVectorTy() || SizeInBits > DL->getLargestLegalIntTypeSizeInBits())
1512  return false;
1513 
1514  // The intrinsic will be sunk behind a compare against zero and branch.
1515  BasicBlock *StartBlock = CountZeros->getParent();
1516  BasicBlock *CallBlock = StartBlock->splitBasicBlock(CountZeros, "cond.false");
1517 
1518  // Create another block after the count zero intrinsic. A PHI will be added
1519  // in this block to select the result of the intrinsic or the bit-width
1520  // constant if the input to the intrinsic is zero.
1521  BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(CountZeros));
1522  BasicBlock *EndBlock = CallBlock->splitBasicBlock(SplitPt, "cond.end");
1523 
1524  // Set up a builder to create a compare, conditional branch, and PHI.
1525  IRBuilder<> Builder(CountZeros->getContext());
1526  Builder.SetInsertPoint(StartBlock->getTerminator());
1527  Builder.SetCurrentDebugLocation(CountZeros->getDebugLoc());
1528 
1529  // Replace the unconditional branch that was created by the first split with
1530  // a compare against zero and a conditional branch.
1531  Value *Zero = Constant::getNullValue(Ty);
1532  Value *Cmp = Builder.CreateICmpEQ(CountZeros->getOperand(0), Zero, "cmpz");
1533  Builder.CreateCondBr(Cmp, EndBlock, CallBlock);
1534  StartBlock->getTerminator()->eraseFromParent();
1535 
1536  // Create a PHI in the end block to select either the output of the intrinsic
1537  // or the bit width of the operand.
1538  Builder.SetInsertPoint(&EndBlock->front());
1539  PHINode *PN = Builder.CreatePHI(Ty, 2, "ctz");
1540  CountZeros->replaceAllUsesWith(PN);
1541  Value *BitWidth = Builder.getInt(APInt(SizeInBits, SizeInBits));
1542  PN->addIncoming(BitWidth, StartBlock);
1543  PN->addIncoming(CountZeros, CallBlock);
1544 
1545  // We are explicitly handling the zero case, so we can set the intrinsic's
1546  // undefined zero argument to 'true'. This will also prevent reprocessing the
1547  // intrinsic; we only despeculate when a zero input is defined.
1548  CountZeros->setArgOperand(1, Builder.getTrue());
1549  ModifiedDT = true;
1550  return true;
1551 }
1552 
1553 bool CodeGenPrepare::optimizeCallInst(CallInst *CI, bool &ModifiedDT) {
1554  BasicBlock *BB = CI->getParent();
1555 
1556  // Lower inline assembly if we can.
1557  // If we found an inline asm expession, and if the target knows how to
1558  // lower it to normal LLVM code, do so now.
1559  if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
1560  if (TLI->ExpandInlineAsm(CI)) {
1561  // Avoid invalidating the iterator.
1562  CurInstIterator = BB->begin();
1563  // Avoid processing instructions out of order, which could cause
1564  // reuse before a value is defined.
1565  SunkAddrs.clear();
1566  return true;
1567  }
1568  // Sink address computing for memory operands into the block.
1569  if (optimizeInlineAsmInst(CI))
1570  return true;
1571  }
1572 
1573  // Align the pointer arguments to this call if the target thinks it's a good
1574  // idea
1575  unsigned MinSize, PrefAlign;
1576  if (TLI && TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) {
1577  for (auto &Arg : CI->arg_operands()) {
1578  // We want to align both objects whose address is used directly and
1579  // objects whose address is used in casts and GEPs, though it only makes
1580  // sense for GEPs if the offset is a multiple of the desired alignment and
1581  // if size - offset meets the size threshold.
1582  if (!Arg->getType()->isPointerTy())
1583  continue;
1584  APInt Offset(DL->getPointerSizeInBits(
1585  cast<PointerType>(Arg->getType())->getAddressSpace()),
1586  0);
1588  uint64_t Offset2 = Offset.getLimitedValue();
1589  if ((Offset2 & (PrefAlign-1)) != 0)
1590  continue;
1591  AllocaInst *AI;
1592  if ((AI = dyn_cast<AllocaInst>(Val)) && AI->getAlignment() < PrefAlign &&
1593  DL->getTypeAllocSize(AI->getAllocatedType()) >= MinSize + Offset2)
1594  AI->setAlignment(PrefAlign);
1595  // Global variables can only be aligned if they are defined in this
1596  // object (i.e. they are uniquely initialized in this object), and
1597  // over-aligning global variables that have an explicit section is
1598  // forbidden.
1599  GlobalVariable *GV;
1600  if ((GV = dyn_cast<GlobalVariable>(Val)) && GV->canIncreaseAlignment() &&
1601  GV->getPointerAlignment(*DL) < PrefAlign &&
1602  DL->getTypeAllocSize(GV->getValueType()) >=
1603  MinSize + Offset2)
1604  GV->setAlignment(PrefAlign);
1605  }
1606  // If this is a memcpy (or similar) then we may be able to improve the
1607  // alignment
1608  if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(CI)) {
1609  unsigned Align = getKnownAlignment(MI->getDest(), *DL);
1610  if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI))
1611  Align = std::min(Align, getKnownAlignment(MTI->getSource(), *DL));
1612  if (Align > MI->getAlignment())
1613  MI->setAlignment(Align);
1614  }
1615  }
1616 
1617  // If we have a cold call site, try to sink addressing computation into the
1618  // cold block. This interacts with our handling for loads and stores to
1619  // ensure that we can fold all uses of a potential addressing computation
1620  // into their uses. TODO: generalize this to work over profiling data
1621  if (!OptSize && CI->hasFnAttr(Attribute::Cold))
1622  for (auto &Arg : CI->arg_operands()) {
1623  if (!Arg->getType()->isPointerTy())
1624  continue;
1625  unsigned AS = Arg->getType()->getPointerAddressSpace();
1626  return optimizeMemoryInst(CI, Arg, Arg->getType(), AS);
1627  }
1628 
1630  if (II) {
1631  switch (II->getIntrinsicID()) {
1632  default: break;
1633  case Intrinsic::objectsize: {
1634  // Lower all uses of llvm.objectsize.*
1635  ConstantInt *RetVal =
1636  lowerObjectSizeCall(II, *DL, TLInfo, /*MustSucceed=*/true);
1637  // Substituting this can cause recursive simplifications, which can
1638  // invalidate our iterator. Use a WeakTrackingVH to hold onto it in case
1639  // this
1640  // happens.
1641  Value *CurValue = &*CurInstIterator;
1642  WeakTrackingVH IterHandle(CurValue);
1643 
1644  replaceAndRecursivelySimplify(CI, RetVal, TLInfo, nullptr);
1645 
1646  // If the iterator instruction was recursively deleted, start over at the
1647  // start of the block.
1648  if (IterHandle != CurValue) {
1649  CurInstIterator = BB->begin();
1650  SunkAddrs.clear();
1651  }
1652  return true;
1653  }
1654  case Intrinsic::aarch64_stlxr:
1655  case Intrinsic::aarch64_stxr: {
1656  ZExtInst *ExtVal = dyn_cast<ZExtInst>(CI->getArgOperand(0));
1657  if (!ExtVal || !ExtVal->hasOneUse() ||
1658  ExtVal->getParent() == CI->getParent())
1659  return false;
1660  // Sink a zext feeding stlxr/stxr before it, so it can be folded into it.
1661  ExtVal->moveBefore(CI);
1662  // Mark this instruction as "inserted by CGP", so that other
1663  // optimizations don't touch it.
1664  InsertedInsts.insert(ExtVal);
1665  return true;
1666  }
1667  case Intrinsic::invariant_group_barrier:
1668  II->replaceAllUsesWith(II->getArgOperand(0));
1669  II->eraseFromParent();
1670  return true;
1671 
1672  case Intrinsic::cttz:
1673  case Intrinsic::ctlz:
1674  // If counting zeros is expensive, try to avoid it.
1675  return despeculateCountZeros(II, TLI, DL, ModifiedDT);
1676  }
1677 
1678  if (TLI) {
1679  SmallVector<Value*, 2> PtrOps;
1680  Type *AccessTy;
1681  if (TLI->getAddrModeArguments(II, PtrOps, AccessTy))
1682  while (!PtrOps.empty()) {
1683  Value *PtrVal = PtrOps.pop_back_val();
1684  unsigned AS = PtrVal->getType()->getPointerAddressSpace();
1685  if (optimizeMemoryInst(II, PtrVal, AccessTy, AS))
1686  return true;
1687  }
1688  }
1689  }
1690 
1691  // From here on out we're working with named functions.
1692  if (!CI->getCalledFunction()) return false;
1693 
1694  // Lower all default uses of _chk calls. This is very similar
1695  // to what InstCombineCalls does, but here we are only lowering calls
1696  // to fortified library functions (e.g. __memcpy_chk) that have the default
1697  // "don't know" as the objectsize. Anything else should be left alone.
1698  FortifiedLibCallSimplifier Simplifier(TLInfo, true);
1699  if (Value *V = Simplifier.optimizeCall(CI)) {
1700  CI->replaceAllUsesWith(V);
1701  CI->eraseFromParent();
1702  return true;
1703  }
1704 
1705  return false;
1706 }
1707 
1708 /// Look for opportunities to duplicate return instructions to the predecessor
1709 /// to enable tail call optimizations. The case it is currently looking for is:
1710 /// @code
1711 /// bb0:
1712 /// %tmp0 = tail call i32 @f0()
1713 /// br label %return
1714 /// bb1:
1715 /// %tmp1 = tail call i32 @f1()
1716 /// br label %return
1717 /// bb2:
1718 /// %tmp2 = tail call i32 @f2()
1719 /// br label %return
1720 /// return:
1721 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
1722 /// ret i32 %retval
1723 /// @endcode
1724 ///
1725 /// =>
1726 ///
1727 /// @code
1728 /// bb0:
1729 /// %tmp0 = tail call i32 @f0()
1730 /// ret i32 %tmp0
1731 /// bb1:
1732 /// %tmp1 = tail call i32 @f1()
1733 /// ret i32 %tmp1
1734 /// bb2:
1735 /// %tmp2 = tail call i32 @f2()
1736 /// ret i32 %tmp2
1737 /// @endcode
1738 bool CodeGenPrepare::dupRetToEnableTailCallOpts(BasicBlock *BB) {
1739  if (!TLI)
1740  return false;
1741 
1742  ReturnInst *RetI = dyn_cast<ReturnInst>(BB->getTerminator());
1743  if (!RetI)
1744  return false;
1745 
1746  PHINode *PN = nullptr;
1747  BitCastInst *BCI = nullptr;
1748  Value *V = RetI->getReturnValue();
1749  if (V) {
1750  BCI = dyn_cast<BitCastInst>(V);
1751  if (BCI)
1752  V = BCI->getOperand(0);
1753 
1754  PN = dyn_cast<PHINode>(V);
1755  if (!PN)
1756  return false;
1757  }
1758 
1759  if (PN && PN->getParent() != BB)
1760  return false;
1761 
1762  // Make sure there are no instructions between the PHI and return, or that the
1763  // return is the first instruction in the block.
1764  if (PN) {
1765  BasicBlock::iterator BI = BB->begin();
1766  do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
1767  if (&*BI == BCI)
1768  // Also skip over the bitcast.
1769  ++BI;
1770  if (&*BI != RetI)
1771  return false;
1772  } else {
1773  BasicBlock::iterator BI = BB->begin();
1774  while (isa<DbgInfoIntrinsic>(BI)) ++BI;
1775  if (&*BI != RetI)
1776  return false;
1777  }
1778 
1779  /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
1780  /// call.
1781  const Function *F = BB->getParent();
1782  SmallVector<CallInst*, 4> TailCalls;
1783  if (PN) {
1784  for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
1785  CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
1786  // Make sure the phi value is indeed produced by the tail call.
1787  if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
1788  TLI->mayBeEmittedAsTailCall(CI) &&
1789  attributesPermitTailCall(F, CI, RetI, *TLI))
1790  TailCalls.push_back(CI);
1791  }
1792  } else {
1793  SmallPtrSet<BasicBlock*, 4> VisitedBBs;
1794  for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
1795  if (!VisitedBBs.insert(*PI).second)
1796  continue;
1797 
1798  BasicBlock::InstListType &InstList = (*PI)->getInstList();
1799  BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
1800  BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
1801  do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
1802  if (RI == RE)
1803  continue;
1804 
1805  CallInst *CI = dyn_cast<CallInst>(&*RI);
1806  if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI) &&
1807  attributesPermitTailCall(F, CI, RetI, *TLI))
1808  TailCalls.push_back(CI);
1809  }
1810  }
1811 
1812  bool Changed = false;
1813  for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
1814  CallInst *CI = TailCalls[i];
1815  CallSite CS(CI);
1816 
1817  // Conservatively require the attributes of the call to match those of the
1818  // return. Ignore noalias because it doesn't affect the call sequence.
1819  AttributeList CalleeAttrs = CS.getAttributes();
1820  if (AttrBuilder(CalleeAttrs, AttributeList::ReturnIndex)
1821  .removeAttribute(Attribute::NoAlias) !=
1823  .removeAttribute(Attribute::NoAlias))
1824  continue;
1825 
1826  // Make sure the call instruction is followed by an unconditional branch to
1827  // the return block.
1828  BasicBlock *CallBB = CI->getParent();
1829  BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
1830  if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
1831  continue;
1832 
1833  // Duplicate the return into CallBB.
1834  (void)FoldReturnIntoUncondBranch(RetI, BB, CallBB);
1835  ModifiedDT = Changed = true;
1836  ++NumRetsDup;
1837  }
1838 
1839  // If we eliminated all predecessors of the block, delete the block now.
1840  if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
1841  BB->eraseFromParent();
1842 
1843  return Changed;
1844 }
1845 
1846 //===----------------------------------------------------------------------===//
1847 // Memory Optimization
1848 //===----------------------------------------------------------------------===//
1849 
1850 namespace {
1851 
1852 /// This is an extended version of TargetLowering::AddrMode
1853 /// which holds actual Value*'s for register values.
1854 struct ExtAddrMode : public TargetLowering::AddrMode {
1855  Value *BaseReg = nullptr;
1856  Value *ScaledReg = nullptr;
1857  Value *OriginalValue = nullptr;
1858 
1859  enum FieldName {
1860  NoField = 0x00,
1861  BaseRegField = 0x01,
1862  BaseGVField = 0x02,
1863  BaseOffsField = 0x04,
1864  ScaledRegField = 0x08,
1865  ScaleField = 0x10,
1866  MultipleFields = 0xff
1867  };
1868 
1869  ExtAddrMode() = default;
1870 
1871  void print(raw_ostream &OS) const;
1872  void dump() const;
1873 
1874  FieldName compare(const ExtAddrMode &other) {
1875  // First check that the types are the same on each field, as differing types
1876  // is something we can't cope with later on.
1877  if (BaseReg && other.BaseReg &&
1878  BaseReg->getType() != other.BaseReg->getType())
1879  return MultipleFields;
1880  if (BaseGV && other.BaseGV &&
1881  BaseGV->getType() != other.BaseGV->getType())
1882  return MultipleFields;
1883  if (ScaledReg && other.ScaledReg &&
1884  ScaledReg->getType() != other.ScaledReg->getType())
1885  return MultipleFields;
1886 
1887  // Check each field to see if it differs.
1888  unsigned Result = NoField;
1889  if (BaseReg != other.BaseReg)
1890  Result |= BaseRegField;
1891  if (BaseGV != other.BaseGV)
1892  Result |= BaseGVField;
1893  if (BaseOffs != other.BaseOffs)
1894  Result |= BaseOffsField;
1895  if (ScaledReg != other.ScaledReg)
1896  Result |= ScaledRegField;
1897  // Don't count 0 as being a different scale, because that actually means
1898  // unscaled (which will already be counted by having no ScaledReg).
1899  if (Scale && other.Scale && Scale != other.Scale)
1900  Result |= ScaleField;
1901 
1902  if (countPopulation(Result) > 1)
1903  return MultipleFields;
1904  else
1905  return static_cast<FieldName>(Result);
1906  }
1907 
1908  // An AddrMode is trivial if it involves no calculation i.e. it is just a base
1909  // with no offset.
1910  bool isTrivial() {
1911  // An AddrMode is (BaseGV + BaseReg + BaseOffs + ScaleReg * Scale) so it is
1912  // trivial if at most one of these terms is nonzero, except that BaseGV and
1913  // BaseReg both being zero actually means a null pointer value, which we
1914  // consider to be 'non-zero' here.
1915  return !BaseOffs && !Scale && !(BaseGV && BaseReg);
1916  }
1917 
1918  Value *GetFieldAsValue(FieldName Field, Type *IntPtrTy) {
1919  switch (Field) {
1920  default:
1921  return nullptr;
1922  case BaseRegField:
1923  return BaseReg;
1924  case BaseGVField:
1925  return BaseGV;
1926  case ScaledRegField:
1927  return ScaledReg;
1928  case BaseOffsField:
1929  return ConstantInt::get(IntPtrTy, BaseOffs);
1930  }
1931  }
1932 
1933  void SetCombinedField(FieldName Field, Value *V,
1934  const SmallVectorImpl<ExtAddrMode> &AddrModes) {
1935  switch (Field) {
1936  default:
1937  llvm_unreachable("Unhandled fields are expected to be rejected earlier");
1938  break;
1939  case ExtAddrMode::BaseRegField:
1940  BaseReg = V;
1941  break;
1942  case ExtAddrMode::BaseGVField:
1943  // A combined BaseGV is an Instruction, not a GlobalValue, so it goes
1944  // in the BaseReg field.
1945  assert(BaseReg == nullptr);
1946  BaseReg = V;
1947  BaseGV = nullptr;
1948  break;
1949  case ExtAddrMode::ScaledRegField:
1950  ScaledReg = V;
1951  // If we have a mix of scaled and unscaled addrmodes then we want scale
1952  // to be the scale and not zero.
1953  if (!Scale)
1954  for (const ExtAddrMode &AM : AddrModes)
1955  if (AM.Scale) {
1956  Scale = AM.Scale;
1957  break;
1958  }
1959  break;
1960  case ExtAddrMode::BaseOffsField:
1961  // The offset is no longer a constant, so it goes in ScaledReg with a
1962  // scale of 1.
1963  assert(ScaledReg == nullptr);
1964  ScaledReg = V;
1965  Scale = 1;
1966  BaseOffs = 0;
1967  break;
1968  }
1969  }
1970 };
1971 
1972 } // end anonymous namespace
1973 
1974 #ifndef NDEBUG
1975 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
1976  AM.print(OS);
1977  return OS;
1978 }
1979 #endif
1980 
1981 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1982 void ExtAddrMode::print(raw_ostream &OS) const {
1983  bool NeedPlus = false;
1984  OS << "[";
1985  if (BaseGV) {
1986  OS << (NeedPlus ? " + " : "")
1987  << "GV:";
1988  BaseGV->printAsOperand(OS, /*PrintType=*/false);
1989  NeedPlus = true;
1990  }
1991 
1992  if (BaseOffs) {
1993  OS << (NeedPlus ? " + " : "")
1994  << BaseOffs;
1995  NeedPlus = true;
1996  }
1997 
1998  if (BaseReg) {
1999  OS << (NeedPlus ? " + " : "")
2000  << "Base:";
2001  BaseReg->printAsOperand(OS, /*PrintType=*/false);
2002  NeedPlus = true;
2003  }
2004  if (Scale) {
2005  OS << (NeedPlus ? " + " : "")
2006  << Scale << "*";
2007  ScaledReg->printAsOperand(OS, /*PrintType=*/false);
2008  }
2009 
2010  OS << ']';
2011 }
2012 
2013 LLVM_DUMP_METHOD void ExtAddrMode::dump() const {
2014  print(dbgs());
2015  dbgs() << '\n';
2016 }
2017 #endif
2018 
2019 namespace {
2020 
2021 /// \brief This class provides transaction based operation on the IR.
2022 /// Every change made through this class is recorded in the internal state and
2023 /// can be undone (rollback) until commit is called.
2024 class TypePromotionTransaction {
2025  /// \brief This represents the common interface of the individual transaction.
2026  /// Each class implements the logic for doing one specific modification on
2027  /// the IR via the TypePromotionTransaction.
2028  class TypePromotionAction {
2029  protected:
2030  /// The Instruction modified.
2031  Instruction *Inst;
2032 
2033  public:
2034  /// \brief Constructor of the action.
2035  /// The constructor performs the related action on the IR.
2036  TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
2037 
2038  virtual ~TypePromotionAction() = default;
2039 
2040  /// \brief Undo the modification done by this action.
2041  /// When this method is called, the IR must be in the same state as it was
2042  /// before this action was applied.
2043  /// \pre Undoing the action works if and only if the IR is in the exact same
2044  /// state as it was directly after this action was applied.
2045  virtual void undo() = 0;
2046 
2047  /// \brief Advocate every change made by this action.
2048  /// When the results on the IR of the action are to be kept, it is important
2049  /// to call this function, otherwise hidden information may be kept forever.
2050  virtual void commit() {
2051  // Nothing to be done, this action is not doing anything.
2052  }
2053  };
2054 
2055  /// \brief Utility to remember the position of an instruction.
2056  class InsertionHandler {
2057  /// Position of an instruction.
2058  /// Either an instruction:
2059  /// - Is the first in a basic block: BB is used.
2060  /// - Has a previous instructon: PrevInst is used.
2061  union {
2062  Instruction *PrevInst;
2063  BasicBlock *BB;
2064  } Point;
2065 
2066  /// Remember whether or not the instruction had a previous instruction.
2067  bool HasPrevInstruction;
2068 
2069  public:
2070  /// \brief Record the position of \p Inst.
2071  InsertionHandler(Instruction *Inst) {
2072  BasicBlock::iterator It = Inst->getIterator();
2073  HasPrevInstruction = (It != (Inst->getParent()->begin()));
2074  if (HasPrevInstruction)
2075  Point.PrevInst = &*--It;
2076  else
2077  Point.BB = Inst->getParent();
2078  }
2079 
2080  /// \brief Insert \p Inst at the recorded position.
2081  void insert(Instruction *Inst) {
2082  if (HasPrevInstruction) {
2083  if (Inst->getParent())
2084  Inst->removeFromParent();
2085  Inst->insertAfter(Point.PrevInst);
2086  } else {
2087  Instruction *Position = &*Point.BB->getFirstInsertionPt();
2088  if (Inst->getParent())
2089  Inst->moveBefore(Position);
2090  else
2091  Inst->insertBefore(Position);
2092  }
2093  }
2094  };
2095 
2096  /// \brief Move an instruction before another.
2097  class InstructionMoveBefore : public TypePromotionAction {
2098  /// Original position of the instruction.
2099  InsertionHandler Position;
2100 
2101  public:
2102  /// \brief Move \p Inst before \p Before.
2103  InstructionMoveBefore(Instruction *Inst, Instruction *Before)
2104  : TypePromotionAction(Inst), Position(Inst) {
2105  DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before << "\n");
2106  Inst->moveBefore(Before);
2107  }
2108 
2109  /// \brief Move the instruction back to its original position.
2110  void undo() override {
2111  DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
2112  Position.insert(Inst);
2113  }
2114  };
2115 
2116  /// \brief Set the operand of an instruction with a new value.
2117  class OperandSetter : public TypePromotionAction {
2118  /// Original operand of the instruction.
2119  Value *Origin;
2120 
2121  /// Index of the modified instruction.
2122  unsigned Idx;
2123 
2124  public:
2125  /// \brief Set \p Idx operand of \p Inst with \p NewVal.
2126  OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
2127  : TypePromotionAction(Inst), Idx(Idx) {
2128  DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
2129  << "for:" << *Inst << "\n"
2130  << "with:" << *NewVal << "\n");
2131  Origin = Inst->getOperand(Idx);
2132  Inst->setOperand(Idx, NewVal);
2133  }
2134 
2135  /// \brief Restore the original value of the instruction.
2136  void undo() override {
2137  DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
2138  << "for: " << *Inst << "\n"
2139  << "with: " << *Origin << "\n");
2140  Inst->setOperand(Idx, Origin);
2141  }
2142  };
2143 
2144  /// \brief Hide the operands of an instruction.
2145  /// Do as if this instruction was not using any of its operands.
2146  class OperandsHider : public TypePromotionAction {
2147  /// The list of original operands.
2148  SmallVector<Value *, 4> OriginalValues;
2149 
2150  public:
2151  /// \brief Remove \p Inst from the uses of the operands of \p Inst.
2152  OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
2153  DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
2154  unsigned NumOpnds = Inst->getNumOperands();
2155  OriginalValues.reserve(NumOpnds);
2156  for (unsigned It = 0; It < NumOpnds; ++It) {
2157  // Save the current operand.
2158  Value *Val = Inst->getOperand(It);
2159  OriginalValues.push_back(Val);
2160  // Set a dummy one.
2161  // We could use OperandSetter here, but that would imply an overhead
2162  // that we are not willing to pay.
2163  Inst->setOperand(It, UndefValue::get(Val->getType()));
2164  }
2165  }
2166 
2167  /// \brief Restore the original list of uses.
2168  void undo() override {
2169  DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
2170  for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
2171  Inst->setOperand(It, OriginalValues[It]);
2172  }
2173  };
2174 
2175  /// \brief Build a truncate instruction.
2176  class TruncBuilder : public TypePromotionAction {
2177  Value *Val;
2178 
2179  public:
2180  /// \brief Build a truncate instruction of \p Opnd producing a \p Ty
2181  /// result.
2182  /// trunc Opnd to Ty.
2183  TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
2184  IRBuilder<> Builder(Opnd);
2185  Val = Builder.CreateTrunc(Opnd, Ty, "promoted");
2186  DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n");
2187  }
2188 
2189  /// \brief Get the built value.
2190  Value *getBuiltValue() { return Val; }
2191 
2192  /// \brief Remove the built instruction.
2193  void undo() override {
2194  DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n");
2195  if (Instruction *IVal = dyn_cast<Instruction>(Val))
2196  IVal->eraseFromParent();
2197  }
2198  };
2199 
2200  /// \brief Build a sign extension instruction.
2201  class SExtBuilder : public TypePromotionAction {
2202  Value *Val;
2203 
2204  public:
2205  /// \brief Build a sign extension instruction of \p Opnd producing a \p Ty
2206  /// result.
2207  /// sext Opnd to Ty.
2208  SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
2209  : TypePromotionAction(InsertPt) {
2210  IRBuilder<> Builder(InsertPt);
2211  Val = Builder.CreateSExt(Opnd, Ty, "promoted");
2212  DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n");
2213  }
2214 
2215  /// \brief Get the built value.
2216  Value *getBuiltValue() { return Val; }
2217 
2218  /// \brief Remove the built instruction.
2219  void undo() override {
2220  DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n");
2221  if (Instruction *IVal = dyn_cast<Instruction>(Val))
2222  IVal->eraseFromParent();
2223  }
2224  };
2225 
2226  /// \brief Build a zero extension instruction.
2227  class ZExtBuilder : public TypePromotionAction {
2228  Value *Val;
2229 
2230  public:
2231  /// \brief Build a zero extension instruction of \p Opnd producing a \p Ty
2232  /// result.
2233  /// zext Opnd to Ty.
2234  ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
2235  : TypePromotionAction(InsertPt) {
2236  IRBuilder<> Builder(InsertPt);
2237  Val = Builder.CreateZExt(Opnd, Ty, "promoted");
2238  DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n");
2239  }
2240 
2241  /// \brief Get the built value.
2242  Value *getBuiltValue() { return Val; }
2243 
2244  /// \brief Remove the built instruction.
2245  void undo() override {
2246  DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n");
2247  if (Instruction *IVal = dyn_cast<Instruction>(Val))
2248  IVal->eraseFromParent();
2249  }
2250  };
2251 
2252  /// \brief Mutate an instruction to another type.
2253  class TypeMutator : public TypePromotionAction {
2254  /// Record the original type.
2255  Type *OrigTy;
2256 
2257  public:
2258  /// \brief Mutate the type of \p Inst into \p NewTy.
2259  TypeMutator(Instruction *Inst, Type *NewTy)
2260  : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
2261  DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
2262  << "\n");
2263  Inst->mutateType(NewTy);
2264  }
2265 
2266  /// \brief Mutate the instruction back to its original type.
2267  void undo() override {
2268  DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
2269  << "\n");
2270  Inst->mutateType(OrigTy);
2271  }
2272  };
2273 
2274  /// \brief Replace the uses of an instruction by another instruction.
2275  class UsesReplacer : public TypePromotionAction {
2276  /// Helper structure to keep track of the replaced uses.
2277  struct InstructionAndIdx {
2278  /// The instruction using the instruction.
2279  Instruction *Inst;
2280 
2281  /// The index where this instruction is used for Inst.
2282  unsigned Idx;
2283 
2284  InstructionAndIdx(Instruction *Inst, unsigned Idx)
2285  : Inst(Inst), Idx(Idx) {}
2286  };
2287 
2288  /// Keep track of the original uses (pair Instruction, Index).
2289  SmallVector<InstructionAndIdx, 4> OriginalUses;
2290 
2291  using use_iterator = SmallVectorImpl<InstructionAndIdx>::iterator;
2292 
2293  public:
2294  /// \brief Replace all the use of \p Inst by \p New.
2295  UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
2296  DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
2297  << "\n");
2298  // Record the original uses.
2299  for (Use &U : Inst->uses()) {
2300  Instruction *UserI = cast<Instruction>(U.getUser());
2301  OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
2302  }
2303  // Now, we can replace the uses.
2304  Inst->replaceAllUsesWith(New);
2305  }
2306 
2307  /// \brief Reassign the original uses of Inst to Inst.
2308  void undo() override {
2309  DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
2310  for (use_iterator UseIt = OriginalUses.begin(),
2311  EndIt = OriginalUses.end();
2312  UseIt != EndIt; ++UseIt) {
2313  UseIt->Inst->setOperand(UseIt->Idx, Inst);
2314  }
2315  }
2316  };
2317 
2318  /// \brief Remove an instruction from the IR.
2319  class InstructionRemover : public TypePromotionAction {
2320  /// Original position of the instruction.
2321  InsertionHandler Inserter;
2322 
2323  /// Helper structure to hide all the link to the instruction. In other
2324  /// words, this helps to do as if the instruction was removed.
2325  OperandsHider Hider;
2326 
2327  /// Keep track of the uses replaced, if any.
2328  UsesReplacer *Replacer = nullptr;
2329 
2330  /// Keep track of instructions removed.
2331  SetOfInstrs &RemovedInsts;
2332 
2333  public:
2334  /// \brief Remove all reference of \p Inst and optinally replace all its
2335  /// uses with New.
2336  /// \p RemovedInsts Keep track of the instructions removed by this Action.
2337  /// \pre If !Inst->use_empty(), then New != nullptr
2338  InstructionRemover(Instruction *Inst, SetOfInstrs &RemovedInsts,
2339  Value *New = nullptr)
2340  : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
2341  RemovedInsts(RemovedInsts) {
2342  if (New)
2343  Replacer = new UsesReplacer(Inst, New);
2344  DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
2345  RemovedInsts.insert(Inst);
2346  /// The instructions removed here will be freed after completing
2347  /// optimizeBlock() for all blocks as we need to keep track of the
2348  /// removed instructions during promotion.
2349  Inst->removeFromParent();
2350  }
2351 
2352  ~InstructionRemover() override { delete Replacer; }
2353 
2354  /// \brief Resurrect the instruction and reassign it to the proper uses if
2355  /// new value was provided when build this action.
2356  void undo() override {
2357  DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
2358  Inserter.insert(Inst);
2359  if (Replacer)
2360  Replacer->undo();
2361  Hider.undo();
2362  RemovedInsts.erase(Inst);
2363  }
2364  };
2365 
2366 public:
2367  /// Restoration point.
2368  /// The restoration point is a pointer to an action instead of an iterator
2369  /// because the iterator may be invalidated but not the pointer.
2370  using ConstRestorationPt = const TypePromotionAction *;
2371 
2372  TypePromotionTransaction(SetOfInstrs &RemovedInsts)
2373  : RemovedInsts(RemovedInsts) {}
2374 
2375  /// Advocate every changes made in that transaction.
2376  void commit();
2377 
2378  /// Undo all the changes made after the given point.
2379  void rollback(ConstRestorationPt Point);
2380 
2381  /// Get the current restoration point.
2382  ConstRestorationPt getRestorationPoint() const;
2383 
2384  /// \name API for IR modification with state keeping to support rollback.
2385  /// @{
2386  /// Same as Instruction::setOperand.
2387  void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
2388 
2389  /// Same as Instruction::eraseFromParent.
2390  void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
2391 
2392  /// Same as Value::replaceAllUsesWith.
2393  void replaceAllUsesWith(Instruction *Inst, Value *New);
2394 
2395  /// Same as Value::mutateType.
2396  void mutateType(Instruction *Inst, Type *NewTy);
2397 
2398  /// Same as IRBuilder::createTrunc.
2399  Value *createTrunc(Instruction *Opnd, Type *Ty);
2400 
2401  /// Same as IRBuilder::createSExt.
2402  Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
2403 
2404  /// Same as IRBuilder::createZExt.
2405  Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
2406 
2407  /// Same as Instruction::moveBefore.
2408  void moveBefore(Instruction *Inst, Instruction *Before);
2409  /// @}
2410 
2411 private:
2412  /// The ordered list of actions made so far.
2414 
2415  using CommitPt = SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator;
2416 
2417  SetOfInstrs &RemovedInsts;
2418 };
2419 
2420 } // end anonymous namespace
2421 
2422 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
2423  Value *NewVal) {
2424  Actions.push_back(llvm::make_unique<TypePromotionTransaction::OperandSetter>(
2425  Inst, Idx, NewVal));
2426 }
2427 
2428 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
2429  Value *NewVal) {
2430  Actions.push_back(
2431  llvm::make_unique<TypePromotionTransaction::InstructionRemover>(
2432  Inst, RemovedInsts, NewVal));
2433 }
2434 
2435 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
2436  Value *New) {
2437  Actions.push_back(
2438  llvm::make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
2439 }
2440 
2441 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
2442  Actions.push_back(
2443  llvm::make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
2444 }
2445 
2446 Value *TypePromotionTransaction::createTrunc(Instruction *Opnd,
2447  Type *Ty) {
2448  std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
2449  Value *Val = Ptr->getBuiltValue();
2450  Actions.push_back(std::move(Ptr));
2451  return Val;
2452 }
2453 
2454 Value *TypePromotionTransaction::createSExt(Instruction *Inst,
2455  Value *Opnd, Type *Ty) {
2456  std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
2457  Value *Val = Ptr->getBuiltValue();
2458  Actions.push_back(std::move(Ptr));
2459  return Val;
2460 }
2461 
2462 Value *TypePromotionTransaction::createZExt(Instruction *Inst,
2463  Value *Opnd, Type *Ty) {
2464  std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
2465  Value *Val = Ptr->getBuiltValue();
2466  Actions.push_back(std::move(Ptr));
2467  return Val;
2468 }
2469 
2470 void TypePromotionTransaction::moveBefore(Instruction *Inst,
2471  Instruction *Before) {
2472  Actions.push_back(
2473  llvm::make_unique<TypePromotionTransaction::InstructionMoveBefore>(
2474  Inst, Before));
2475 }
2476 
2477 TypePromotionTransaction::ConstRestorationPt
2478 TypePromotionTransaction::getRestorationPoint() const {
2479  return !Actions.empty() ? Actions.back().get() : nullptr;
2480 }
2481 
2482 void TypePromotionTransaction::commit() {
2483  for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
2484  ++It)
2485  (*It)->commit();
2486  Actions.clear();
2487 }
2488 
2489 void TypePromotionTransaction::rollback(
2490  TypePromotionTransaction::ConstRestorationPt Point) {
2491  while (!Actions.empty() && Point != Actions.back().get()) {
2492  std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
2493  Curr->undo();
2494  }
2495 }
2496 
2497 namespace {
2498 
2499 /// \brief A helper class for matching addressing modes.
2500 ///
2501 /// This encapsulates the logic for matching the target-legal addressing modes.
2502 class AddressingModeMatcher {
2503  SmallVectorImpl<Instruction*> &AddrModeInsts;
2504  const TargetLowering &TLI;
2505  const TargetRegisterInfo &TRI;
2506  const DataLayout &DL;
2507 
2508  /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
2509  /// the memory instruction that we're computing this address for.
2510  Type *AccessTy;
2511  unsigned AddrSpace;
2512  Instruction *MemoryInst;
2513 
2514  /// This is the addressing mode that we're building up. This is
2515  /// part of the return value of this addressing mode matching stuff.
2516  ExtAddrMode &AddrMode;
2517 
2518  /// The instructions inserted by other CodeGenPrepare optimizations.
2519  const SetOfInstrs &InsertedInsts;
2520 
2521  /// A map from the instructions to their type before promotion.
2522  InstrToOrigTy &PromotedInsts;
2523 
2524  /// The ongoing transaction where every action should be registered.
2525  TypePromotionTransaction &TPT;
2526 
2527  /// This is set to true when we should not do profitability checks.
2528  /// When true, IsProfitableToFoldIntoAddressingMode always returns true.
2529  bool IgnoreProfitability;
2530 
2531  AddressingModeMatcher(SmallVectorImpl<Instruction *> &AMI,
2532  const TargetLowering &TLI,
2533  const TargetRegisterInfo &TRI,
2534  Type *AT, unsigned AS,
2535  Instruction *MI, ExtAddrMode &AM,
2536  const SetOfInstrs &InsertedInsts,
2537  InstrToOrigTy &PromotedInsts,
2538  TypePromotionTransaction &TPT)
2539  : AddrModeInsts(AMI), TLI(TLI), TRI(TRI),
2540  DL(MI->getModule()->getDataLayout()), AccessTy(AT), AddrSpace(AS),
2541  MemoryInst(MI), AddrMode(AM), InsertedInsts(InsertedInsts),
2542  PromotedInsts(PromotedInsts), TPT(TPT) {
2543  IgnoreProfitability = false;
2544  }
2545 
2546 public:
2547  /// Find the maximal addressing mode that a load/store of V can fold,
2548  /// give an access type of AccessTy. This returns a list of involved
2549  /// instructions in AddrModeInsts.
2550  /// \p InsertedInsts The instructions inserted by other CodeGenPrepare
2551  /// optimizations.
2552  /// \p PromotedInsts maps the instructions to their type before promotion.
2553  /// \p The ongoing transaction where every action should be registered.
2554  static ExtAddrMode Match(Value *V, Type *AccessTy, unsigned AS,
2555  Instruction *MemoryInst,
2556  SmallVectorImpl<Instruction*> &AddrModeInsts,
2557  const TargetLowering &TLI,
2558  const TargetRegisterInfo &TRI,
2559  const SetOfInstrs &InsertedInsts,
2560  InstrToOrigTy &PromotedInsts,
2561  TypePromotionTransaction &TPT) {
2562  ExtAddrMode Result;
2563 
2564  bool Success = AddressingModeMatcher(AddrModeInsts, TLI, TRI,
2565  AccessTy, AS,
2566  MemoryInst, Result, InsertedInsts,
2567  PromotedInsts, TPT).matchAddr(V, 0);
2568  (void)Success; assert(Success && "Couldn't select *anything*?");
2569  return Result;
2570  }
2571 
2572 private:
2573  bool matchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
2574  bool matchAddr(Value *V, unsigned Depth);
2575  bool matchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth,
2576  bool *MovedAway = nullptr);
2577  bool isProfitableToFoldIntoAddressingMode(Instruction *I,
2578  ExtAddrMode &AMBefore,
2579  ExtAddrMode &AMAfter);
2580  bool valueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
2581  bool isPromotionProfitable(unsigned NewCost, unsigned OldCost,
2582  Value *PromotedOperand) const;
2583 };
2584 
2585 /// \brief Keep track of simplification of Phi nodes.
2586 /// Accept the set of all phi nodes and erase phi node from this set
2587 /// if it is simplified.
2588 class SimplificationTracker {
2590  const SimplifyQuery &SQ;
2591  SmallPtrSetImpl<PHINode *> &AllPhiNodes;
2592  SmallPtrSetImpl<SelectInst *> &AllSelectNodes;
2593 
2594 public:
2595  SimplificationTracker(const SimplifyQuery &sq,
2598  : SQ(sq), AllPhiNodes(APN), AllSelectNodes(ASN) {}
2599 
2600  Value *Get(Value *V) {
2601  do {
2602  auto SV = Storage.find(V);
2603  if (SV == Storage.end())
2604  return V;
2605  V = SV->second;
2606  } while (true);
2607  }
2608 
2609  Value *Simplify(Value *Val) {
2610  SmallVector<Value *, 32> WorkList;
2611  SmallPtrSet<Value *, 32> Visited;
2612  WorkList.push_back(Val);
2613  while (!WorkList.empty()) {
2614  auto P = WorkList.pop_back_val();
2615  if (!Visited.insert(P).second)
2616  continue;
2617  if (auto *PI = dyn_cast<Instruction>(P))
2618  if (Value *V = SimplifyInstruction(cast<Instruction>(PI), SQ)) {
2619  for (auto *U : PI->users())
2620  WorkList.push_back(cast<Value>(U));
2621  Put(PI, V);
2622  PI->replaceAllUsesWith(V);
2623  if (auto *PHI = dyn_cast<PHINode>(PI))
2624  AllPhiNodes.erase(PHI);
2625  if (auto *Select = dyn_cast<SelectInst>(PI))
2626  AllSelectNodes.erase(Select);
2627  PI->eraseFromParent();
2628  }
2629  }
2630  return Get(Val);
2631  }
2632 
2633  void Put(Value *From, Value *To) {
2634  Storage.insert({ From, To });
2635  }
2636 };
2637 
2638 /// \brief A helper class for combining addressing modes.
2639 class AddressingModeCombiner {
2640  typedef std::pair<Value *, BasicBlock *> ValueInBB;
2641  typedef DenseMap<ValueInBB, Value *> FoldAddrToValueMapping;
2642  typedef std::pair<PHINode *, PHINode *> PHIPair;
2643 
2644 private:
2645  /// The addressing modes we've collected.
2646  SmallVector<ExtAddrMode, 16> AddrModes;
2647 
2648  /// The field in which the AddrModes differ, when we have more than one.
2649  ExtAddrMode::FieldName DifferentField = ExtAddrMode::NoField;
2650 
2651  /// Are the AddrModes that we have all just equal to their original values?
2652  bool AllAddrModesTrivial = true;
2653 
2654  /// Common Type for all different fields in addressing modes.
2655  Type *CommonType;
2656 
2657  /// SimplifyQuery for simplifyInstruction utility.
2658  const SimplifyQuery &SQ;
2659 
2660  /// Original Address.
2661  ValueInBB Original;
2662 
2663 public:
2664  AddressingModeCombiner(const SimplifyQuery &_SQ, ValueInBB OriginalValue)
2665  : CommonType(nullptr), SQ(_SQ), Original(OriginalValue) {}
2666 
2667  /// \brief Get the combined AddrMode
2668  const ExtAddrMode &getAddrMode() const {
2669  return AddrModes[0];
2670  }
2671 
2672  /// \brief Add a new AddrMode if it's compatible with the AddrModes we already
2673  /// have.
2674  /// \return True iff we succeeded in doing so.
2675  bool addNewAddrMode(ExtAddrMode &NewAddrMode) {
2676  // Take note of if we have any non-trivial AddrModes, as we need to detect
2677  // when all AddrModes are trivial as then we would introduce a phi or select
2678  // which just duplicates what's already there.
2679  AllAddrModesTrivial = AllAddrModesTrivial && NewAddrMode.isTrivial();
2680 
2681  // If this is the first addrmode then everything is fine.
2682  if (AddrModes.empty()) {
2683  AddrModes.emplace_back(NewAddrMode);
2684  return true;
2685  }
2686 
2687  // Figure out how different this is from the other address modes, which we
2688  // can do just by comparing against the first one given that we only care
2689  // about the cumulative difference.
2690  ExtAddrMode::FieldName ThisDifferentField =
2691  AddrModes[0].compare(NewAddrMode);
2692  if (DifferentField == ExtAddrMode::NoField)
2693  DifferentField = ThisDifferentField;
2694  else if (DifferentField != ThisDifferentField)
2695  DifferentField = ExtAddrMode::MultipleFields;
2696 
2697  // If NewAddrMode differs in only one dimension, and that dimension isn't
2698  // the amount that ScaledReg is scaled by, then we can handle it by
2699  // inserting a phi/select later on. Even if NewAddMode is the same
2700  // we still need to collect it due to original value is different.
2701  // And later we will need all original values as anchors during
2702  // finding the common Phi node.
2703  // We also must reject the case when base offset is different and
2704  // scale reg is not null, we cannot handle this case due to merge of
2705  // different offsets will be used as ScaleReg.
2706  if (DifferentField != ExtAddrMode::MultipleFields &&
2707  DifferentField != ExtAddrMode::ScaleField &&
2708  (DifferentField != ExtAddrMode::BaseOffsField ||
2709  !NewAddrMode.ScaledReg)) {
2710  AddrModes.emplace_back(NewAddrMode);
2711  return true;
2712  }
2713 
2714  // We couldn't combine NewAddrMode with the rest, so return failure.
2715  AddrModes.clear();
2716  return false;
2717  }
2718 
2719  /// \brief Combine the addressing modes we've collected into a single
2720  /// addressing mode.
2721  /// \return True iff we successfully combined them or we only had one so
2722  /// didn't need to combine them anyway.
2723  bool combineAddrModes() {
2724  // If we have no AddrModes then they can't be combined.
2725  if (AddrModes.size() == 0)
2726  return false;
2727 
2728  // A single AddrMode can trivially be combined.
2729  if (AddrModes.size() == 1 || DifferentField == ExtAddrMode::NoField)
2730  return true;
2731 
2732  // If the AddrModes we collected are all just equal to the value they are
2733  // derived from then combining them wouldn't do anything useful.
2734  if (AllAddrModesTrivial)
2735  return false;
2736 
2737  if (!addrModeCombiningAllowed())
2738  return false;
2739 
2740  // Build a map between <original value, basic block where we saw it> to
2741  // value of base register.
2742  // Bail out if there is no common type.
2743  FoldAddrToValueMapping Map;
2744  if (!initializeMap(Map))
2745  return false;
2746 
2747  Value *CommonValue = findCommon(Map);
2748  if (CommonValue)
2749  AddrModes[0].SetCombinedField(DifferentField, CommonValue, AddrModes);
2750  return CommonValue != nullptr;
2751  }
2752 
2753 private:
2754  /// \brief Initialize Map with anchor values. For address seen in some BB
2755  /// we set the value of different field saw in this address.
2756  /// If address is not an instruction than basic block is set to null.
2757  /// At the same time we find a common type for different field we will
2758  /// use to create new Phi/Select nodes. Keep it in CommonType field.
2759  /// Return false if there is no common type found.
2760  bool initializeMap(FoldAddrToValueMapping &Map) {
2761  // Keep track of keys where the value is null. We will need to replace it
2762  // with constant null when we know the common type.
2763  SmallVector<ValueInBB, 2> NullValue;
2764  Type *IntPtrTy = SQ.DL.getIntPtrType(AddrModes[0].OriginalValue->getType());
2765  for (auto &AM : AddrModes) {
2766  BasicBlock *BB = nullptr;
2767  if (Instruction *I = dyn_cast<Instruction>(AM.OriginalValue))
2768  BB = I->getParent();
2769 
2770  Value *DV = AM.GetFieldAsValue(DifferentField, IntPtrTy);
2771  if (DV) {
2772  auto *Type = DV->getType();
2773  if (CommonType && CommonType != Type)
2774  return false;
2775  CommonType = Type;
2776  Map[{ AM.OriginalValue, BB }] = DV;
2777  } else {
2778  NullValue.push_back({ AM.OriginalValue, BB });
2779  }
2780  }
2781  assert(CommonType && "At least one non-null value must be!");
2782  for (auto VIBB : NullValue)
2783  Map[VIBB] = Constant::getNullValue(CommonType);
2784  return true;
2785  }
2786 
2787  /// \brief We have mapping between value A and basic block where value A
2788  /// seen to other value B where B was a field in addressing mode represented
2789  /// by A. Also we have an original value C representin an address in some
2790  /// basic block. Traversing from C through phi and selects we ended up with
2791  /// A's in a map. This utility function tries to find a value V which is a
2792  /// field in addressing mode C and traversing through phi nodes and selects
2793  /// we will end up in corresponded values B in a map.
2794  /// The utility will create a new Phi/Selects if needed.
2795  // The simple example looks as follows:
2796  // BB1:
2797  // p1 = b1 + 40
2798  // br cond BB2, BB3
2799  // BB2:
2800  // p2 = b2 + 40
2801  // br BB3
2802  // BB3:
2803  // p = phi [p1, BB1], [p2, BB2]
2804  // v = load p
2805  // Map is
2806  // <p1, BB1> -> b1
2807  // <p2, BB2> -> b2
2808  // Request is
2809  // <p, BB3> -> ?
2810  // The function tries to find or build phi [b1, BB1], [b2, BB2] in BB3
2811  Value *findCommon(FoldAddrToValueMapping &Map) {
2812  // Tracks newly created Phi nodes.
2813  SmallPtrSet<PHINode *, 32> NewPhiNodes;
2814  // Tracks newly created Select nodes.
2815  SmallPtrSet<SelectInst *, 32> NewSelectNodes;
2816  // Tracks the simplification of newly created phi nodes. The reason we use
2817  // this mapping is because we will add new created Phi nodes in AddrToBase.
2818  // Simplification of Phi nodes is recursive, so some Phi node may
2819  // be simplified after we added it to AddrToBase.
2820  // Using this mapping we can find the current value in AddrToBase.
2821  SimplificationTracker ST(SQ, NewPhiNodes, NewSelectNodes);
2822 
2823  // First step, DFS to create PHI nodes for all intermediate blocks.
2824  // Also fill traverse order for the second step.
2825  SmallVector<ValueInBB, 32> TraverseOrder;
2826  InsertPlaceholders(Map, TraverseOrder, NewPhiNodes, NewSelectNodes);
2827 
2828  // Second Step, fill new nodes by merged values and simplify if possible.
2829  FillPlaceholders(Map, TraverseOrder, ST);
2830 
2831  if (!AddrSinkNewSelects && NewSelectNodes.size() > 0) {
2832  DestroyNodes(NewPhiNodes);
2833  DestroyNodes(NewSelectNodes);
2834  return nullptr;
2835  }
2836 
2837  // Now we'd like to match New Phi nodes to existed ones.
2838  unsigned PhiNotMatchedCount = 0;
2839  if (!MatchPhiSet(NewPhiNodes, ST, AddrSinkNewPhis, PhiNotMatchedCount)) {
2840  DestroyNodes(NewPhiNodes);
2841  DestroyNodes(NewSelectNodes);
2842  return nullptr;
2843  }
2844 
2845  auto *Result = ST.Get(Map.find(Original)->second);
2846  if (Result) {
2847  NumMemoryInstsPhiCreated += NewPhiNodes.size() + PhiNotMatchedCount;
2848  NumMemoryInstsSelectCreated += NewSelectNodes.size();
2849  }
2850  return Result;
2851  }
2852 
2853  /// \brief Destroy nodes from a set.
2854  template <typename T> void DestroyNodes(SmallPtrSetImpl<T *> &Instructions) {
2855  // For safe erasing, replace the Phi with dummy value first.
2856  auto Dummy = UndefValue::get(CommonType);
2857  for (auto I : Instructions) {
2859  I->eraseFromParent();
2860  }
2861  }
2862 
2863  /// \brief Try to match PHI node to Candidate.
2864  /// Matcher tracks the matched Phi nodes.
2865  bool MatchPhiNode(PHINode *PHI, PHINode *Candidate,
2866  DenseSet<PHIPair> &Matcher,
2867  SmallPtrSetImpl<PHINode *> &PhiNodesToMatch) {
2868  SmallVector<PHIPair, 8> WorkList;
2869  Matcher.insert({ PHI, Candidate });
2870  WorkList.push_back({ PHI, Candidate });
2871  SmallSet<PHIPair, 8> Visited;
2872  while (!WorkList.empty()) {
2873  auto Item = WorkList.pop_back_val();
2874  if (!Visited.insert(Item).second)
2875  continue;
2876  // We iterate over all incoming values to Phi to compare them.
2877  // If values are different and both of them Phi and the first one is a
2878  // Phi we added (subject to match) and both of them is in the same basic
2879  // block then we can match our pair if values match. So we state that
2880  // these values match and add it to work list to verify that.
2881  for (auto B : Item.first->blocks()) {
2882  Value *FirstValue = Item.first->getIncomingValueForBlock(B);
2883  Value *SecondValue = Item.second->getIncomingValueForBlock(B);
2884  if (FirstValue == SecondValue)
2885  continue;
2886 
2887  PHINode *FirstPhi = dyn_cast<PHINode>(FirstValue);
2888  PHINode *SecondPhi = dyn_cast<PHINode>(SecondValue);
2889 
2890  // One of them is not Phi or
2891  // The first one is not Phi node from the set we'd like to match or
2892  // Phi nodes from different basic blocks then
2893  // we will not be able to match.
2894  if (!FirstPhi || !SecondPhi || !PhiNodesToMatch.count(FirstPhi) ||
2895  FirstPhi->getParent() != SecondPhi->getParent())
2896  return false;
2897 
2898  // If we already matched them then continue.
2899  if (Matcher.count({ FirstPhi, SecondPhi }))
2900  continue;
2901  // So the values are different and does not match. So we need them to
2902  // match.
2903  Matcher.insert({ FirstPhi, SecondPhi });
2904  // But me must check it.
2905  WorkList.push_back({ FirstPhi, SecondPhi });
2906  }
2907  }
2908  return true;
2909  }
2910 
2911  /// \brief For the given set of PHI nodes try to find their equivalents.
2912  /// Returns false if this matching fails and creation of new Phi is disabled.
2913  bool MatchPhiSet(SmallPtrSetImpl<PHINode *> &PhiNodesToMatch,
2914  SimplificationTracker &ST, bool AllowNewPhiNodes,
2915  unsigned &PhiNotMatchedCount) {
2916  DenseSet<PHIPair> Matched;
2917  SmallPtrSet<PHINode *, 8> WillNotMatch;
2918  while (PhiNodesToMatch.size()) {
2919  PHINode *PHI = *PhiNodesToMatch.begin();
2920 
2921  // Add us, if no Phi nodes in the basic block we do not match.
2922  WillNotMatch.clear();
2923  WillNotMatch.insert(PHI);
2924 
2925  // Traverse all Phis until we found equivalent or fail to do that.
2926  bool IsMatched = false;
2927  for (auto &P : PHI->getParent()->phis()) {
2928  if (&P == PHI)
2929  continue;
2930  if ((IsMatched = MatchPhiNode(PHI, &P, Matched, PhiNodesToMatch)))
2931  break;
2932  // If it does not match, collect all Phi nodes from matcher.
2933  // if we end up with no match, them all these Phi nodes will not match
2934  // later.
2935  for (auto M : Matched)
2936  WillNotMatch.insert(M.first);
2937  Matched.clear();
2938  }
2939  if (IsMatched) {
2940  // Replace all matched values and erase them.
2941  for (auto MV : Matched) {
2942  MV.first->replaceAllUsesWith(MV.second);
2943  PhiNodesToMatch.erase(MV.first);
2944  ST.Put(MV.first, MV.second);
2945  MV.first->eraseFromParent();
2946  }
2947  Matched.clear();
2948  continue;
2949  }
2950  // If we are not allowed to create new nodes then bail out.
2951  if (!AllowNewPhiNodes)
2952  return false;
2953  // Just remove all seen values in matcher. They will not match anything.
2954  PhiNotMatchedCount += WillNotMatch.size();
2955  for (auto *P : WillNotMatch)
2956  PhiNodesToMatch.erase(P);
2957  }
2958  return true;
2959  }
2960  /// \brief Fill the placeholder with values from predecessors and simplify it.
2961  void FillPlaceholders(FoldAddrToValueMapping &Map,
2962  SmallVectorImpl<ValueInBB> &TraverseOrder,
2963  SimplificationTracker &ST) {
2964  while (!TraverseOrder.empty()) {
2965  auto Current = TraverseOrder.pop_back_val();
2966  assert(Map.find(Current) != Map.end() && "No node to fill!!!");
2967  Value *CurrentValue = Current.first;
2968  BasicBlock *CurrentBlock = Current.second;
2969  Value *V = Map[Current];
2970 
2971  if (SelectInst *Select = dyn_cast<SelectInst>(V)) {
2972  // CurrentValue also must be Select.
2973  auto *CurrentSelect = cast<SelectInst>(CurrentValue);
2974  auto *TrueValue = CurrentSelect->getTrueValue();
2975  ValueInBB TrueItem = { TrueValue, isa<Instruction>(TrueValue)
2976  ? CurrentBlock
2977  : nullptr };
2978  assert(Map.find(TrueItem) != Map.end() && "No True Value!");
2979  Select->setTrueValue(ST.Get(Map[TrueItem]));
2980  auto *FalseValue = CurrentSelect->getFalseValue();
2981  ValueInBB FalseItem = { FalseValue, isa<Instruction>(FalseValue)
2982  ? CurrentBlock
2983  : nullptr };
2984  assert(Map.find(FalseItem) != Map.end() && "No False Value!");
2985  Select->setFalseValue(ST.Get(Map[FalseItem]));
2986  } else {
2987  // Must be a Phi node then.
2988  PHINode *PHI = cast<PHINode>(V);
2989  // Fill the Phi node with values from predecessors.
2990  bool IsDefinedInThisBB =
2991  cast<Instruction>(CurrentValue)->getParent() == CurrentBlock;
2992  auto *CurrentPhi = dyn_cast<PHINode>(CurrentValue);
2993  for (auto B : predecessors(CurrentBlock)) {
2994  Value *PV = IsDefinedInThisBB
2995  ? CurrentPhi->getIncomingValueForBlock(B)
2996  : CurrentValue;
2997  ValueInBB item = { PV, isa<Instruction>(PV) ? B : nullptr };
2998  assert(Map.find(item) != Map.end() && "No predecessor Value!");
2999  PHI->addIncoming(ST.Get(Map[item]), B);
3000  }
3001  }
3002  // Simplify if possible.
3003  Map[Current] = ST.Simplify(V);
3004  }
3005  }
3006 
3007  /// Starting from value recursively iterates over predecessors up to known
3008  /// ending values represented in a map. For each traversed block inserts
3009  /// a placeholder Phi or Select.
3010  /// Reports all new created Phi/Select nodes by adding them to set.
3011  /// Also reports and order in what basic blocks have been traversed.
3012  void InsertPlaceholders(FoldAddrToValueMapping &Map,
3013  SmallVectorImpl<ValueInBB> &TraverseOrder,
3014  SmallPtrSetImpl<PHINode *> &NewPhiNodes,
3015  SmallPtrSetImpl<SelectInst *> &NewSelectNodes) {
3016  SmallVector<ValueInBB, 32> Worklist;
3017  assert((isa<PHINode>(Original.first) || isa<SelectInst>(Original.first)) &&
3018  "Address must be a Phi or Select node");
3019  auto *Dummy = UndefValue::get(CommonType);
3020  Worklist.push_back(Original);
3021  while (!Worklist.empty()) {
3022  auto Current = Worklist.pop_back_val();
3023  // If value is not an instruction it is something global, constant,
3024  // parameter and we can say that this value is observable in any block.
3025  // Set block to null to denote it.
3026  // Also please take into account that it is how we build anchors.
3027  if (!isa<Instruction>(Current.first))
3028  Current.second = nullptr;
3029  // if it is already visited or it is an ending value then skip it.
3030  if (Map.find(Current) != Map.end())
3031  continue;
3032  TraverseOrder.push_back(Current);
3033 
3034  Value *CurrentValue = Current.first;
3035  BasicBlock *CurrentBlock = Current.second;
3036  // CurrentValue must be a Phi node or select. All others must be covered
3037  // by anchors.
3038  Instruction *CurrentI = cast<Instruction>(CurrentValue);
3039  bool IsDefinedInThisBB = CurrentI->getParent() == CurrentBlock;
3040 
3041  unsigned PredCount =
3042  std::distance(pred_begin(CurrentBlock), pred_end(CurrentBlock));
3043  // if Current Value is not defined in this basic block we are interested
3044  // in values in predecessors.
3045  if (!IsDefinedInThisBB) {
3046  assert(PredCount && "Unreachable block?!");
3047  PHINode *PHI = PHINode::Create(CommonType, PredCount, "sunk_phi",
3048  &CurrentBlock->front());
3049  Map[Current] = PHI;
3050  NewPhiNodes.insert(PHI);
3051  // Add all predecessors in work list.
3052  for (auto B : predecessors(CurrentBlock))
3053  Worklist.push_back({ CurrentValue, B });
3054  continue;
3055  }
3056  // Value is defined in this basic block.
3057  if (SelectInst *OrigSelect = dyn_cast<SelectInst>(CurrentI)) {
3058  // Is it OK to get metadata from OrigSelect?!
3059  // Create a Select placeholder with dummy value.
3060  SelectInst *Select =
3061  SelectInst::Create(OrigSelect->getCondition(), Dummy, Dummy,
3062  OrigSelect->getName(), OrigSelect, OrigSelect);
3063  Map[Current] = Select;
3064  NewSelectNodes.insert(Select);
3065  // We are interested in True and False value in this basic block.
3066  Worklist.push_back({ OrigSelect->getTrueValue(), CurrentBlock });
3067  Worklist.push_back({ OrigSelect->getFalseValue(), CurrentBlock });
3068  } else {
3069  // It must be a Phi node then.
3070  auto *CurrentPhi = cast<PHINode>(CurrentI);
3071  // Create new Phi node for merge of bases.
3072  assert(PredCount && "Unreachable block?!");
3073  PHINode *PHI = PHINode::Create(CommonType, PredCount, "sunk_phi",
3074  &CurrentBlock->front());
3075  Map[Current] = PHI;
3076  NewPhiNodes.insert(PHI);
3077 
3078  // Add all predecessors in work list.
3079  for (auto B : predecessors(CurrentBlock))
3080  Worklist.push_back({ CurrentPhi->getIncomingValueForBlock(B), B });
3081  }
3082  }
3083  }
3084 
3085  bool addrModeCombiningAllowed() {
3087  return false;
3088  switch (DifferentField) {
3089  default:
3090  return false;
3091  case ExtAddrMode::BaseRegField:
3092  return AddrSinkCombineBaseReg;
3093  case ExtAddrMode::BaseGVField:
3094  return AddrSinkCombineBaseGV;
3095  case ExtAddrMode::BaseOffsField:
3096  return AddrSinkCombineBaseOffs;
3097  case ExtAddrMode::ScaledRegField:
3098  return AddrSinkCombineScaledReg;
3099  }
3100  }
3101 };
3102 } // end anonymous namespace
3103 
3104 /// Try adding ScaleReg*Scale to the current addressing mode.
3105 /// Return true and update AddrMode if this addr mode is legal for the target,
3106 /// false if not.
3107 bool AddressingModeMatcher::matchScaledValue(Value *ScaleReg, int64_t Scale,
3108  unsigned Depth) {
3109  // If Scale is 1, then this is the same as adding ScaleReg to the addressing
3110  // mode. Just process that directly.
3111  if (Scale == 1)
3112  return matchAddr(ScaleReg, Depth);
3113 
3114  // If the scale is 0, it takes nothing to add this.
3115  if (Scale == 0)
3116  return true;
3117 
3118  // If we already have a scale of this value, we can add to it, otherwise, we
3119  // need an available scale field.
3120  if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
3121  return false;
3122 
3123  ExtAddrMode TestAddrMode = AddrMode;
3124 
3125  // Add scale to turn X*4+X*3 -> X*7. This could also do things like
3126  // [A+B + A*7] -> [B+A*8].
3127  TestAddrMode.Scale += Scale;
3128  TestAddrMode.ScaledReg = ScaleReg;
3129 
3130  // If the new address isn't legal, bail out.
3131  if (!TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace))
3132  return false;
3133 
3134  // It was legal, so commit it.
3135  AddrMode = TestAddrMode;
3136 
3137  // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
3138  // to see if ScaleReg is actually X+C. If so, we can turn this into adding
3139  // X*Scale + C*Scale to addr mode.
3140  ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
3141  if (isa<Instruction>(ScaleReg) && // not a constant expr.
3142  match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
3143  TestAddrMode.ScaledReg = AddLHS;
3144  TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
3145 
3146  // If this addressing mode is legal, commit it and remember that we folded
3147  // this instruction.
3148  if (TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace)) {
3149  AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
3150  AddrMode = TestAddrMode;
3151  return true;
3152  }
3153  }
3154 
3155  // Otherwise, not (x+c)*scale, just return what we have.
3156  return true;
3157 }
3158 
3159 /// This is a little filter, which returns true if an addressing computation
3160 /// involving I might be folded into a load/store accessing it.
3161 /// This doesn't need to be perfect, but needs to accept at least
3162 /// the set of instructions that MatchOperationAddr can.
3164  switch (I->getOpcode()) {
3165  case Instruction::BitCast:
3166  case Instruction::AddrSpaceCast:
3167  // Don't touch identity bitcasts.
3168  if (I->getType() == I->getOperand(0)->getType())
3169  return false;
3170  return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
3171  case Instruction::PtrToInt:
3172  // PtrToInt is always a noop, as we know that the int type is pointer sized.
3173  return true;
3174  case Instruction::IntToPtr:
3175  // We know the input is intptr_t, so this is foldable.
3176  return true;
3177  case Instruction::Add:
3178  return true;
3179  case Instruction::Mul:
3180  case Instruction::Shl:
3181  // Can only handle X*C and X << C.
3182  return isa<ConstantInt>(I->getOperand(1));
3183  case Instruction::GetElementPtr:
3184  return true;
3185  default:
3186  return false;
3187  }
3188 }
3189 
3190 /// \brief Check whether or not \p Val is a legal instruction for \p TLI.
3191 /// \note \p Val is assumed to be the product of some type promotion.
3192 /// Therefore if \p Val has an undefined state in \p TLI, this is assumed
3193 /// to be legal, as the non-promoted value would have had the same state.
3195  const DataLayout &DL, Value *Val) {
3196  Instruction *PromotedInst = dyn_cast<Instruction>(Val);
3197  if (!PromotedInst)
3198  return false;
3199  int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
3200  // If the ISDOpcode is undefined, it was undefined before the promotion.
3201  if (!ISDOpcode)
3202  return true;
3203  // Otherwise, check if the promoted instruction is legal or not.
3204  return TLI.isOperationLegalOrCustom(
3205  ISDOpcode, TLI.getValueType(DL, PromotedInst->getType()));
3206 }
3207 
3208 namespace {
3209 
3210 /// \brief Hepler class to perform type promotion.
3211 class TypePromotionHelper {
3212  /// \brief Utility function to check whether or not a sign or zero extension
3213  /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
3214  /// either using the operands of \p Inst or promoting \p Inst.
3215  /// The type of the extension is defined by \p IsSExt.
3216  /// In other words, check if:
3217  /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
3218  /// #1 Promotion applies:
3219  /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
3220  /// #2 Operand reuses:
3221  /// ext opnd1 to ConsideredExtType.
3222  /// \p PromotedInsts maps the instructions to their type before promotion.
3223  static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
3224  const InstrToOrigTy &PromotedInsts, bool IsSExt);
3225 
3226  /// \brief Utility function to determine if \p OpIdx should be promoted when
3227  /// promoting \p Inst.
3228  static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
3229  return !(isa<SelectInst>(Inst) && OpIdx == 0);
3230  }
3231 
3232  /// \brief Utility function to promote the operand of \p Ext when this
3233  /// operand is a promotable trunc or sext or zext.
3234  /// \p PromotedInsts maps the instructions to their type before promotion.
3235  /// \p CreatedInstsCost[out] contains the cost of all instructions
3236  /// created to promote the operand of Ext.
3237  /// Newly added extensions are inserted in \p Exts.
3238  /// Newly added truncates are inserted in \p Truncs.
3239  /// Should never be called directly.
3240  /// \return The promoted value which is used instead of Ext.
3241  static Value *promoteOperandForTruncAndAnyExt(
3242  Instruction *Ext, TypePromotionTransaction &TPT,
3243  InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3245  SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI);
3246 
3247  /// \brief Utility function to promote the operand of \p Ext when this
3248  /// operand is promotable and is not a supported trunc or sext.
3249  /// \p PromotedInsts maps the instructions to their type before promotion.
3250  /// \p CreatedInstsCost[out] contains the cost of all the instructions
3251  /// created to promote the operand of Ext.
3252  /// Newly added extensions are inserted in \p Exts.
3253  /// Newly added truncates are inserted in \p Truncs.
3254  /// Should never be called directly.
3255  /// \return The promoted value which is used instead of Ext.
3256  static Value *promoteOperandForOther(Instruction *Ext,
3257  TypePromotionTransaction &TPT,
3258  InstrToOrigTy &PromotedInsts,
3259  unsigned &CreatedInstsCost,
3262  const TargetLowering &TLI, bool IsSExt);
3263 
3264  /// \see promoteOperandForOther.
3265  static Value *signExtendOperandForOther(
3266  Instruction *Ext, TypePromotionTransaction &TPT,
3267  InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3269  SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3270  return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
3271  Exts, Truncs, TLI, true);
3272  }
3273 
3274  /// \see promoteOperandForOther.
3275  static Value *zeroExtendOperandForOther(
3276  Instruction *Ext, TypePromotionTransaction &TPT,
3277  InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3279  SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3280  return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
3281  Exts, Truncs, TLI, false);
3282  }
3283 
3284 public:
3285  /// Type for the utility function that promotes the operand of Ext.
3286  using Action = Value *(*)(Instruction *Ext, TypePromotionTransaction &TPT,
3287  InstrToOrigTy &PromotedInsts,
3288  unsigned &CreatedInstsCost,
3291  const TargetLowering &TLI);
3292 
3293  /// \brief Given a sign/zero extend instruction \p Ext, return the approriate
3294  /// action to promote the operand of \p Ext instead of using Ext.
3295  /// \return NULL if no promotable action is possible with the current
3296  /// sign extension.
3297  /// \p InsertedInsts keeps track of all the instructions inserted by the
3298  /// other CodeGenPrepare optimizations. This information is important
3299  /// because we do not want to promote these instructions as CodeGenPrepare
3300  /// will reinsert them later. Thus creating an infinite loop: create/remove.
3301  /// \p PromotedInsts maps the instructions to their type before promotion.
3302  static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedInsts,
3303  const TargetLowering &TLI,
3304  const InstrToOrigTy &PromotedInsts);
3305 };
3306 
3307 } // end anonymous namespace
3308 
3309 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
3310  Type *ConsideredExtType,
3311  const InstrToOrigTy &PromotedInsts,
3312  bool IsSExt) {
3313  // The promotion helper does not know how to deal with vector types yet.
3314  // To be able to fix that, we would need to fix the places where we
3315  // statically extend, e.g., constants and such.
3316  if (Inst->getType()->isVectorTy())
3317  return false;
3318 
3319  // We can always get through zext.
3320  if (isa<ZExtInst>(Inst))
3321  return true;
3322 
3323  // sext(sext) is ok too.
3324  if (IsSExt && isa<SExtInst>(Inst))
3325  return true;
3326 
3327  // We can get through binary operator, if it is legal. In other words, the
3328  // binary operator must have a nuw or nsw flag.
3329  const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
3330  if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
3331  ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
3332  (IsSExt && BinOp->hasNoSignedWrap())))
3333  return true;
3334 
3335  // Check if we can do the following simplification.
3336  // ext(trunc(opnd)) --> ext(opnd)
3337  if (!isa<TruncInst>(Inst))
3338  return false;
3339 
3340  Value *OpndVal = Inst->getOperand(0);
3341  // Check if we can use this operand in the extension.
3342  // If the type is larger than the result type of the extension, we cannot.
3343  if (!OpndVal->getType()->isIntegerTy() ||
3344  OpndVal->getType()->getIntegerBitWidth() >
3345  ConsideredExtType->getIntegerBitWidth())
3346  return false;
3347 
3348  // If the operand of the truncate is not an instruction, we will not have
3349  // any information on the dropped bits.
3350  // (Actually we could for constant but it is not worth the extra logic).
3351  Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
3352  if (!Opnd)
3353  return false;
3354 
3355  // Check if the source of the type is narrow enough.
3356  // I.e., check that trunc just drops extended bits of the same kind of
3357  // the extension.
3358  // #1 get the type of the operand and check the kind of the extended bits.
3359  const Type *OpndType;
3360  InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
3361  if (It != PromotedInsts.end() && It->second.getInt() == IsSExt)
3362  OpndType = It->second.getPointer();
3363  else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd)))
3364  OpndType = Opnd->getOperand(0)->getType();
3365  else
3366  return false;
3367 
3368  // #2 check that the truncate just drops extended bits.
3369  return Inst->getType()->getIntegerBitWidth() >=
3370  OpndType->getIntegerBitWidth();
3371 }
3372 
3373 TypePromotionHelper::Action TypePromotionHelper::getAction(
3374  Instruction *Ext, const SetOfInstrs &InsertedInsts,
3375  const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
3376  assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
3377  "Unexpected instruction type");
3378  Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0));
3379  Type *ExtTy = Ext->getType();
3380  bool IsSExt = isa<SExtInst>(Ext);
3381  // If the operand of the extension is not an instruction, we cannot
3382  // get through.
3383  // If it, check we can get through.
3384  if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt))
3385  return nullptr;
3386 
3387  // Do not promote if the operand has been added by codegenprepare.
3388  // Otherwise, it means we are undoing an optimization that is likely to be
3389  // redone, thus causing potential infinite loop.
3390  if (isa<TruncInst>(ExtOpnd) && InsertedInsts.count(ExtOpnd))
3391  return nullptr;
3392 
3393  // SExt or Trunc instructions.
3394  // Return the related handler.
3395  if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) ||
3396  isa<ZExtInst>(ExtOpnd))
3397  return promoteOperandForTruncAndAnyExt;
3398 
3399  // Regular instruction.
3400  // Abort early if we will have to insert non-free instructions.
3401  if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType()))
3402  return nullptr;
3403  return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
3404 }
3405 
3406 Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
3407  Instruction *SExt, TypePromotionTransaction &TPT,
3408  InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3410  SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3411  // By construction, the operand of SExt is an instruction. Otherwise we cannot
3412  // get through it and this method should not be called.
3413  Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
3414  Value *ExtVal = SExt;
3415  bool HasMergedNonFreeExt = false;
3416  if (isa<ZExtInst>(SExtOpnd)) {
3417  // Replace s|zext(zext(opnd))
3418  // => zext(opnd).
3419  HasMergedNonFreeExt = !TLI.isExtFree(SExtOpnd);
3420  Value *ZExt =
3421  TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
3422  TPT.replaceAllUsesWith(SExt, ZExt);
3423  TPT.eraseInstruction(SExt);
3424  ExtVal = ZExt;
3425  } else {
3426  // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
3427  // => z|sext(opnd).
3428  TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
3429  }
3430  CreatedInstsCost = 0;
3431 
3432  // Remove dead code.
3433  if (SExtOpnd->use_empty())
3434  TPT.eraseInstruction(SExtOpnd);
3435 
3436  // Check if the extension is still needed.
3437  Instruction *ExtInst = dyn_cast<Instruction>(ExtVal);
3438  if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) {
3439  if (ExtInst) {
3440  if (Exts)
3441  Exts->push_back(ExtInst);
3442  CreatedInstsCost = !TLI.isExtFree(ExtInst) && !HasMergedNonFreeExt;
3443  }
3444  return ExtVal;
3445  }
3446 
3447  // At this point we have: ext ty opnd to ty.
3448  // Reassign the uses of ExtInst to the opnd and remove ExtInst.
3449  Value *NextVal = ExtInst->getOperand(0);
3450  TPT.eraseInstruction(ExtInst, NextVal);
3451  return NextVal;
3452 }
3453 
3454 Value *TypePromotionHelper::promoteOperandForOther(
3455  Instruction *Ext, TypePromotionTransaction &TPT,
3456  InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3458  SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI,
3459  bool IsSExt) {
3460  // By construction, the operand of Ext is an instruction. Otherwise we cannot
3461  // get through it and this method should not be called.
3462  Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0));
3463  CreatedInstsCost = 0;
3464  if (!ExtOpnd->hasOneUse()) {
3465  // ExtOpnd will be promoted.
3466  // All its uses, but Ext, will need to use a truncated value of the
3467  // promoted version.
3468  // Create the truncate now.
3469  Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType());
3470  if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) {
3471  // Insert it just after the definition.
3472  ITrunc->moveAfter(ExtOpnd);
3473  if (Truncs)
3474  Truncs->push_back(ITrunc);
3475  }
3476 
3477  TPT.replaceAllUsesWith(ExtOpnd, Trunc);
3478  // Restore the operand of Ext (which has been replaced by the previous call
3479  // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
3480  TPT.setOperand(Ext, 0, ExtOpnd);
3481  }
3482 
3483  // Get through the Instruction:
3484  // 1. Update its type.
3485  // 2. Replace the uses of Ext by Inst.
3486  // 3. Extend each operand that needs to be extended.
3487 
3488  // Remember the original type of the instruction before promotion.
3489  // This is useful to know that the high bits are sign extended bits.
3490  PromotedInsts.insert(std::pair<Instruction *, TypeIsSExt>(
3491  ExtOpnd, TypeIsSExt(ExtOpnd->getType(), IsSExt)));
3492  // Step #1.
3493  TPT.mutateType(ExtOpnd, Ext->getType());
3494  // Step #2.
3495  TPT.replaceAllUsesWith(Ext, ExtOpnd);
3496  // Step #3.
3497  Instruction *ExtForOpnd = Ext;
3498 
3499  DEBUG(dbgs() << "Propagate Ext to operands\n");
3500  for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
3501  ++OpIdx) {
3502  DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n');
3503  if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() ||
3504  !shouldExtOperand(ExtOpnd, OpIdx)) {
3505  DEBUG(dbgs() << "No need to propagate\n");
3506  continue;
3507  }
3508  // Check if we can statically extend the operand.
3509  Value *Opnd = ExtOpnd->getOperand(OpIdx);
3510  if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
3511  DEBUG(dbgs() << "Statically extend\n");
3512  unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
3513  APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth)
3514  : Cst->getValue().zext(BitWidth);
3515  TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal));
3516  continue;
3517  }
3518  // UndefValue are typed, so we have to statically sign extend them.
3519  if (isa<UndefValue>(Opnd)) {
3520  DEBUG(dbgs() << "Statically extend\n");
3521  TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType()));
3522  continue;
3523  }
3524 
3525  // Otherwise we have to explicity sign extend the operand.
3526  // Check if Ext was reused to extend an operand.
3527  if (!ExtForOpnd) {
3528  // If yes, create a new one.
3529  DEBUG(dbgs() << "More operands to ext\n");
3530  Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType())
3531  : TPT.createZExt(Ext, Opnd, Ext->getType());
3532  if (!isa<Instruction>(ValForExtOpnd)) {
3533  TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd);
3534  continue;
3535  }
3536  ExtForOpnd = cast<Instruction>(ValForExtOpnd);
3537  }
3538  if (Exts)
3539  Exts->push_back(ExtForOpnd);
3540  TPT.setOperand(ExtForOpnd, 0, Opnd);
3541 
3542  // Move the sign extension before the insertion point.
3543  TPT.moveBefore(ExtForOpnd, ExtOpnd);
3544  TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd);
3545  CreatedInstsCost += !TLI.isExtFree(ExtForOpnd);
3546  // If more sext are required, new instructions will have to be created.
3547  ExtForOpnd = nullptr;
3548  }
3549  if (ExtForOpnd == Ext) {
3550  DEBUG(dbgs() << "Extension is useless now\n");
3551  TPT.eraseInstruction(Ext);
3552  }
3553  return ExtOpnd;
3554 }
3555 
3556 /// Check whether or not promoting an instruction to a wider type is profitable.
3557 /// \p NewCost gives the cost of extension instructions created by the
3558 /// promotion.
3559 /// \p OldCost gives the cost of extension instructions before the promotion
3560 /// plus the number of instructions that have been
3561 /// matched in the addressing mode the promotion.
3562 /// \p PromotedOperand is the value that has been promoted.
3563 /// \return True if the promotion is profitable, false otherwise.
3564 bool AddressingModeMatcher::isPromotionProfitable(
3565  unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const {
3566  DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCost << '\n');
3567  // The cost of the new extensions is greater than the cost of the
3568  // old extension plus what we folded.
3569  // This is not profitable.
3570  if (NewCost > OldCost)
3571  return false;
3572  if (NewCost < OldCost)
3573  return true;
3574  // The promotion is neutral but it may help folding the sign extension in
3575  // loads for instance.
3576  // Check that we did not create an illegal instruction.
3577  return isPromotedInstructionLegal(TLI, DL, PromotedOperand);
3578 }
3579 
3580 /// Given an instruction or constant expr, see if we can fold the operation
3581 /// into the addressing mode. If so, update the addressing mode and return
3582 /// true, otherwise return false without modifying AddrMode.
3583 /// If \p MovedAway is not NULL, it contains the information of whether or
3584 /// not AddrInst has to be folded into the addressing mode on success.
3585 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
3586 /// because it has been moved away.
3587 /// Thus AddrInst must not be added in the matched instructions.
3588 /// This state can happen when AddrInst is a sext, since it may be moved away.
3589 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
3590 /// not be referenced anymore.
3591 bool AddressingModeMatcher::matchOperationAddr(User *AddrInst, unsigned Opcode,
3592  unsigned Depth,
3593  bool *MovedAway) {
3594  // Avoid exponential behavior on extremely deep expression trees.
3595  if (Depth >= 5) return false;
3596 
3597  // By default, all matched instructions stay in place.
3598  if (MovedAway)
3599  *MovedAway = false;
3600 
3601  switch (Opcode) {
3602  case Instruction::PtrToInt:
3603  // PtrToInt is always a noop, as we know that the int type is pointer sized.
3604  return matchAddr(AddrInst->getOperand(0), Depth);
3605  case Instruction::IntToPtr: {
3606  auto AS = AddrInst->getType()->getPointerAddressSpace();
3607  auto PtrTy = MVT::getIntegerVT(DL.getPointerSizeInBits(AS));
3608  // This inttoptr is a no-op if the integer type is pointer sized.
3609  if (TLI.getValueType(DL, AddrInst->getOperand(0)->getType()) == PtrTy)
3610  return matchAddr(AddrInst->getOperand(0), Depth);
3611  return false;
3612  }
3613  case Instruction::BitCast:
3614  // BitCast is always a noop, and we can handle it as long as it is
3615  // int->int or pointer->pointer (we don't want int<->fp or something).
3616  if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
3617  AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
3618  // Don't touch identity bitcasts. These were probably put here by LSR,
3619  // and we don't want to mess around with them. Assume it knows what it
3620  // is doing.
3621  AddrInst->getOperand(0)->getType() != AddrInst->getType())
3622  return matchAddr(AddrInst->getOperand(0), Depth);
3623  return false;
3624  case Instruction::AddrSpaceCast: {
3625  unsigned SrcAS
3626  = AddrInst->getOperand(0)->getType()->getPointerAddressSpace();
3627  unsigned DestAS = AddrInst->getType()->getPointerAddressSpace();
3628  if (TLI.isNoopAddrSpaceCast(SrcAS, DestAS))
3629  return matchAddr(AddrInst->getOperand(0), Depth);
3630  return false;
3631  }
3632  case Instruction::Add: {
3633  // Check to see if we can merge in the RHS then the LHS. If so, we win.
3634  ExtAddrMode BackupAddrMode = AddrMode;
3635  unsigned OldSize = AddrModeInsts.size();
3636  // Start a transaction at this point.
3637  // The LHS may match but not the RHS.
3638  // Therefore, we need a higher level restoration point to undo partially
3639  // matched operation.
3640  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3641  TPT.getRestorationPoint();
3642 
3643  if (matchAddr(AddrInst->getOperand(1), Depth+1) &&
3644  matchAddr(AddrInst->getOperand(0), Depth+1))
3645  return true;
3646 
3647  // Restore the old addr mode info.
3648  AddrMode = BackupAddrMode;
3649  AddrModeInsts.resize(OldSize);
3650  TPT.rollback(LastKnownGood);
3651 
3652  // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
3653  if (matchAddr(AddrInst->getOperand(0), Depth+1) &&
3654  matchAddr(AddrInst->getOperand(1), Depth+1))
3655  return true;
3656 
3657  // Otherwise we definitely can't merge the ADD in.
3658  AddrMode = BackupAddrMode;
3659  AddrModeInsts.resize(OldSize);
3660  TPT.rollback(LastKnownGood);
3661  break;
3662  }
3663  //case Instruction::Or:
3664  // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
3665  //break;
3666  case Instruction::Mul:
3667  case Instruction::Shl: {
3668  // Can only handle X*C and X << C.
3669  ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
3670  if (!RHS || RHS->getBitWidth() > 64)
3671  return false;
3672  int64_t Scale = RHS->getSExtValue();
3673  if (Opcode == Instruction::Shl)
3674  Scale = 1LL << Scale;
3675 
3676  return matchScaledValue(AddrInst->getOperand(0), Scale, Depth);
3677  }
3678  case Instruction::GetElementPtr: {
3679  // Scan the GEP. We check it if it contains constant offsets and at most
3680  // one variable offset.
3681  int VariableOperand = -1;
3682  unsigned VariableScale = 0;
3683 
3684  int64_t ConstantOffset = 0;
3685  gep_type_iterator GTI = gep_type_begin(AddrInst);
3686  for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
3687  if (StructType *STy = GTI.getStructTypeOrNull()) {
3688  const StructLayout *SL = DL.getStructLayout(STy);
3689  unsigned Idx =
3690  cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
3691  ConstantOffset += SL->getElementOffset(Idx);
3692  } else {
3693  uint64_t TypeSize = DL.getTypeAllocSize(GTI.getIndexedType());
3694  if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
3695  ConstantOffset += CI->getSExtValue() * TypeSize;
3696  } else if (TypeSize) { // Scales of zero don't do anything.
3697  // We only allow one variable index at the moment.
3698  if (VariableOperand != -1)
3699  return false;
3700 
3701  // Remember the variable index.
3702  VariableOperand = i;
3703  VariableScale = TypeSize;
3704  }
3705  }
3706  }
3707 
3708  // A common case is for the GEP to only do a constant offset. In this case,
3709  // just add it to the disp field and check validity.
3710  if (VariableOperand == -1) {
3711  AddrMode.BaseOffs += ConstantOffset;
3712  if (ConstantOffset == 0 ||
3713  TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) {
3714  // Check to see if we can fold the base pointer in too.
3715  if (matchAddr(AddrInst->getOperand(0), Depth+1))
3716  return true;
3717  }
3718  AddrMode.BaseOffs -= ConstantOffset;
3719  return false;
3720  }
3721 
3722  // Save the valid addressing mode in case we can't match.
3723  ExtAddrMode BackupAddrMode = AddrMode;
3724  unsigned OldSize = AddrModeInsts.size();
3725 
3726  // See if the scale and offset amount is valid for this target.
3727  AddrMode.BaseOffs += ConstantOffset;
3728 
3729  // Match the base operand of the GEP.
3730  if (!matchAddr(AddrInst->getOperand(0), Depth+1)) {
3731  // If it couldn't be matched, just stuff the value in a register.
3732  if (AddrMode.HasBaseReg) {
3733  AddrMode = BackupAddrMode;
3734  AddrModeInsts.resize(OldSize);
3735  return false;
3736  }
3737  AddrMode.HasBaseReg = true;
3738  AddrMode.BaseReg = AddrInst->getOperand(0);
3739  }
3740 
3741  // Match the remaining variable portion of the GEP.
3742  if (!matchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
3743  Depth)) {
3744  // If it couldn't be matched, try stuffing the base into a register
3745  // instead of matching it, and retrying the match of the scale.
3746  AddrMode = BackupAddrMode;
3747  AddrModeInsts.resize(OldSize);
3748  if (AddrMode.HasBaseReg)
3749  return false;
3750  AddrMode.HasBaseReg = true;
3751  AddrMode.BaseReg = AddrInst->getOperand(0);
3752  AddrMode.BaseOffs += ConstantOffset;
3753  if (!matchScaledValue(AddrInst->getOperand(VariableOperand),
3754  VariableScale, Depth)) {
3755  // If even that didn't work, bail.
3756  AddrMode = BackupAddrMode;
3757  AddrModeInsts.resize(OldSize);
3758  return false;
3759  }
3760  }
3761 
3762  return true;
3763  }
3764  case Instruction::SExt:
3765  case Instruction::ZExt: {
3766  Instruction *Ext = dyn_cast<Instruction>(AddrInst);
3767  if (!Ext)
3768  return false;
3769 
3770  // Try to move this ext out of the way of the addressing mode.
3771  // Ask for a method for doing so.
3772  TypePromotionHelper::Action TPH =
3773  TypePromotionHelper::getAction(Ext, InsertedInsts, TLI, PromotedInsts);
3774  if (!TPH)
3775  return false;
3776 
3777  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3778  TPT.getRestorationPoint();
3779  unsigned CreatedInstsCost = 0;
3780  unsigned ExtCost = !TLI.isExtFree(Ext);
3781  Value *PromotedOperand =
3782  TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI);
3783  // SExt has been moved away.
3784  // Thus either it will be rematched later in the recursive calls or it is
3785  // gone. Anyway, we must not fold it into the addressing mode at this point.
3786  // E.g.,
3787  // op = add opnd, 1
3788  // idx = ext op
3789  // addr = gep base, idx
3790  // is now:
3791  // promotedOpnd = ext opnd <- no match here
3792  // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
3793  // addr = gep base, op <- match
3794  if (MovedAway)
3795  *MovedAway = true;
3796 
3797  assert(PromotedOperand &&
3798  "TypePromotionHelper should have filtered out those cases");
3799 
3800  ExtAddrMode BackupAddrMode = AddrMode;
3801  unsigned OldSize = AddrModeInsts.size();
3802 
3803  if (!matchAddr(PromotedOperand, Depth) ||
3804  // The total of the new cost is equal to the cost of the created
3805  // instructions.
3806  // The total of the old cost is equal to the cost of the extension plus
3807  // what we have saved in the addressing mode.
3808  !isPromotionProfitable(CreatedInstsCost,
3809  ExtCost + (AddrModeInsts.size() - OldSize),
3810  PromotedOperand)) {
3811  AddrMode = BackupAddrMode;
3812  AddrModeInsts.resize(OldSize);
3813  DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
3814  TPT.rollback(LastKnownGood);
3815  return false;
3816  }
3817  return true;
3818  }
3819  }
3820  return false;
3821 }
3822 
3823 /// If we can, try to add the value of 'Addr' into the current addressing mode.
3824 /// If Addr can't be added to AddrMode this returns false and leaves AddrMode
3825 /// unmodified. This assumes that Addr is either a pointer type or intptr_t
3826 /// for the target.
3827 ///
3828 bool AddressingModeMatcher::matchAddr(Value *Addr, unsigned Depth) {
3829  // Start a transaction at this point that we will rollback if the matching
3830  // fails.
3831  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3832  TPT.getRestorationPoint();
3833  if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
3834  // Fold in immediates if legal for the target.
3835  AddrMode.BaseOffs += CI->getSExtValue();
3836  if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3837  return true;
3838  AddrMode.BaseOffs -= CI->getSExtValue();
3839  } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
3840  // If this is a global variable, try to fold it into the addressing mode.
3841  if (!AddrMode.BaseGV) {
3842  AddrMode.BaseGV = GV;
3843  if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3844  return true;
3845  AddrMode.BaseGV = nullptr;
3846  }
3847  } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
3848  ExtAddrMode BackupAddrMode = AddrMode;
3849  unsigned OldSize = AddrModeInsts.size();
3850 
3851  // Check to see if it is possible to fold this operation.
3852  bool MovedAway = false;
3853  if (matchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
3854  // This instruction may have been moved away. If so, there is nothing
3855  // to check here.
3856  if (MovedAway)
3857  return true;
3858  // Okay, it's possible to fold this. Check to see if it is actually
3859  // *profitable* to do so. We use a simple cost model to avoid increasing
3860  // register pressure too much.
3861  if (I->hasOneUse() ||
3862  isProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
3863  AddrModeInsts.push_back(I);
3864  return true;
3865  }
3866 
3867  // It isn't profitable to do this, roll back.
3868  //cerr << "NOT FOLDING: " << *I;
3869  AddrMode = BackupAddrMode;
3870  AddrModeInsts.resize(OldSize);
3871  TPT.rollback(LastKnownGood);
3872  }
3873  } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
3874  if (matchOperationAddr(CE, CE->getOpcode(), Depth))
3875  return true;
3876  TPT.rollback(LastKnownGood);
3877  } else if (isa<ConstantPointerNull>(Addr)) {
3878  // Null pointer gets folded without affecting the addressing mode.
3879  return true;
3880  }
3881 
3882  // Worse case, the target should support [reg] addressing modes. :)
3883  if (!AddrMode.HasBaseReg) {
3884  AddrMode.HasBaseReg = true;
3885  AddrMode.BaseReg = Addr;
3886  // Still check for legality in case the target supports [imm] but not [i+r].
3887  if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3888  return true;
3889  AddrMode.HasBaseReg = false;
3890  AddrMode.BaseReg = nullptr;
3891  }
3892 
3893  // If the base register is already taken, see if we can do [r+r].
3894  if (AddrMode.Scale == 0) {
3895  AddrMode.Scale = 1;
3896  AddrMode.ScaledReg = Addr;
3897  if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3898  return true;
3899  AddrMode.Scale = 0;
3900  AddrMode.ScaledReg = nullptr;
3901  }
3902  // Couldn't match.
3903  TPT.rollback(LastKnownGood);
3904  return false;
3905 }
3906 
3907 /// Check to see if all uses of OpVal by the specified inline asm call are due
3908 /// to memory operands. If so, return true, otherwise return false.
3909 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
3910  const TargetLowering &TLI,
3911  const TargetRegisterInfo &TRI) {
3912  const Function *F = CI->getFunction();
3913  TargetLowering::AsmOperandInfoVector TargetConstraints =
3914  TLI.ParseConstraints(F->getParent()->getDataLayout(), &TRI,
3915  ImmutableCallSite(CI));
3916 
3917  for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
3918  TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
3919 
3920  // Compute the constraint code and ConstraintType to use.
3921  TLI.ComputeConstraintToUse(OpInfo, SDValue());
3922 
3923  // If this asm operand is our Value*, and if it isn't an indirect memory
3924  // operand, we can't fold it!
3925  if (OpInfo.CallOperandVal == OpVal &&
3927  !OpInfo.isIndirect))
3928  return false;
3929  }
3930 
3931  return true;
3932 }
3933 
3934 // Max number of memory uses to look at before aborting the search to conserve
3935 // compile time.
3936 static constexpr int MaxMemoryUsesToScan = 20;
3937 
3938 /// Recursively walk all the uses of I until we find a memory use.
3939 /// If we find an obviously non-foldable instruction, return true.
3940 /// Add the ultimately found memory instructions to MemoryUses.
3941 static bool FindAllMemoryUses(
3942  Instruction *I,
3943  SmallVectorImpl<std::pair<Instruction *, unsigned>> &MemoryUses,
3944  SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetLowering &TLI,
3945  const TargetRegisterInfo &TRI, int SeenInsts = 0) {
3946  // If we already considered this instruction, we're done.
3947  if (!ConsideredInsts.insert(I).second)
3948  return false;
3949 
3950  // If this is an obviously unfoldable instruction, bail out.
3951  if (!MightBeFoldableInst(I))
3952  return true;
3953 
3954  const bool OptSize = I->getFunction()->optForSize();
3955 
3956  // Loop over all the uses, recursively processing them.
3957  for (Use &U : I->uses()) {
3958  // Conservatively return true if we're seeing a large number or a deep chain
3959  // of users. This avoids excessive compilation times in pathological cases.
3960  if (SeenInsts++ >= MaxMemoryUsesToScan)
3961  return true;
3962 
3963  Instruction *UserI = cast<Instruction>(U.getUser());
3964  if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
3965  MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
3966  continue;
3967  }
3968 
3969  if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
3970  unsigned opNo = U.getOperandNo();
3971  if (opNo != StoreInst::getPointerOperandIndex())
3972  return true; // Storing addr, not into addr.
3973  MemoryUses.push_back(std::make_pair(SI, opNo));
3974  continue;
3975  }
3976 
3977  if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(UserI)) {
3978  unsigned opNo = U.getOperandNo();
3980  return true; // Storing addr, not into addr.
3981  MemoryUses.push_back(std::make_pair(RMW, opNo));
3982  continue;
3983  }
3984 
3985  if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(UserI)) {
3986  unsigned opNo = U.getOperandNo();
3988  return true; // Storing addr, not into addr.
3989  MemoryUses.push_back(std::make_pair(CmpX, opNo));
3990  continue;
3991  }
3992 
3993  if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
3994  // If this is a cold call, we can sink the addressing calculation into
3995  // the cold path. See optimizeCallInst
3996  if (!OptSize && CI->hasFnAttr(Attribute::Cold))
3997  continue;
3998 
4000  if (!IA) return true;
4001 
4002  // If this is a memory operand, we're cool, otherwise bail out.
4003  if (!IsOperandAMemoryOperand(CI, IA, I, TLI, TRI))
4004  return true;
4005  continue;
4006  }
4007 
4008  if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TLI, TRI,
4009  SeenInsts))
4010  return true;
4011  }
4012 
4013  return false;
4014 }
4015 
4016 /// Return true if Val is already known to be live at the use site that we're
4017 /// folding it into. If so, there is no cost to include it in the addressing
4018 /// mode. KnownLive1 and KnownLive2 are two values that we know are live at the
4019 /// instruction already.
4020 bool AddressingModeMatcher::valueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
4021  Value *KnownLive2) {
4022  // If Val is either of the known-live values, we know it is live!
4023  if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
4024  return true;
4025 
4026  // All values other than instructions and arguments (e.g. constants) are live.
4027  if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
4028 
4029  // If Val is a constant sized alloca in the entry block, it is live, this is
4030  // true because it is just a reference to the stack/frame pointer, which is
4031  // live for the whole function.
4032  if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
4033  if (AI->isStaticAlloca())
4034  return true;
4035 
4036  // Check to see if this value is already used in the memory instruction's
4037  // block. If so, it's already live into the block at the very least, so we
4038  // can reasonably fold it.
4039  return Val->isUsedInBasicBlock(MemoryInst->getParent());
4040 }
4041 
4042 /// It is possible for the addressing mode of the machine to fold the specified
4043 /// instruction into a load or store that ultimately uses it.
4044 /// However, the specified instruction has multiple uses.
4045 /// Given this, it may actually increase register pressure to fold it
4046 /// into the load. For example, consider this code:
4047 ///
4048 /// X = ...
4049 /// Y = X+1
4050 /// use(Y) -> nonload/store
4051 /// Z = Y+1
4052 /// load Z
4053 ///
4054 /// In this case, Y has multiple uses, and can be folded into the load of Z
4055 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
4056 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
4057 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
4058 /// number of computations either.
4059 ///
4060 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
4061 /// X was live across 'load Z' for other reasons, we actually *would* want to
4062 /// fold the addressing mode in the Z case. This would make Y die earlier.
4063 bool AddressingModeMatcher::
4064 isProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
4065  ExtAddrMode &AMAfter) {
4066  if (IgnoreProfitability) return true;
4067 
4068  // AMBefore is the addressing mode before this instruction was folded into it,
4069  // and AMAfter is the addressing mode after the instruction was folded. Get
4070  // the set of registers referenced by AMAfter and subtract out those
4071  // referenced by AMBefore: this is the set of values which folding in this
4072  // address extends the lifetime of.
4073  //
4074  // Note that there are only two potential values being referenced here,
4075  // BaseReg and ScaleReg (global addresses are always available, as are any
4076  // folded immediates).
4077  Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
4078 
4079  // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
4080  // lifetime wasn't extended by adding this instruction.
4081  if (valueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
4082  BaseReg = nullptr;
4083  if (valueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
4084  ScaledReg = nullptr;
4085 
4086  // If folding this instruction (and it's subexprs) didn't extend any live
4087  // ranges, we're ok with it.
4088  if (!BaseReg && !ScaledReg)
4089  return true;
4090 
4091  // If all uses of this instruction can have the address mode sunk into them,
4092  // we can remove the addressing mode and effectively trade one live register
4093  // for another (at worst.) In this context, folding an addressing mode into
4094  // the use is just a particularly nice way of sinking it.
4096  SmallPtrSet<Instruction*, 16> ConsideredInsts;
4097  if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TLI, TRI))
4098  return false; // Has a non-memory, non-foldable use!
4099 
4100  // Now that we know that all uses of this instruction are part of a chain of
4101  // computation involving only operations that could theoretically be folded
4102  // into a memory use, loop over each of these memory operation uses and see
4103  // if they could *actually* fold the instruction. The assumption is that
4104  // addressing modes are cheap and that duplicating the computation involved
4105  // many times is worthwhile, even on a fastpath. For sinking candidates
4106  // (i.e. cold call sites), this serves as a way to prevent excessive code
4107  // growth since most architectures have some reasonable small and fast way to
4108  // compute an effective address. (i.e LEA on x86)
4109  SmallVector<Instruction*, 32> MatchedAddrModeInsts;
4110  for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
4111  Instruction *User = MemoryUses[i].first;
4112  unsigned OpNo = MemoryUses[i].second;
4113 
4114  // Get the access type of this use. If the use isn't a pointer, we don't
4115  // know what it accesses.
4116  Value *Address = User->getOperand(OpNo);
4117  PointerType *AddrTy = dyn_cast<PointerType>(Address->getType());
4118  if (!AddrTy)
4119  return false;
4120  Type *AddressAccessTy = AddrTy->getElementType();
4121  unsigned AS = AddrTy->getAddressSpace();
4122 
4123  // Do a match against the root of this address, ignoring profitability. This
4124  // will tell us if the addressing mode for the memory operation will
4125  // *actually* cover the shared instruction.
4126  ExtAddrMode Result;
4127  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4128  TPT.getRestorationPoint();
4129  AddressingModeMatcher Matcher(MatchedAddrModeInsts, TLI, TRI,
4130  AddressAccessTy, AS,
4131  MemoryInst, Result, InsertedInsts,
4132  PromotedInsts, TPT);
4133  Matcher.IgnoreProfitability = true;
4134  bool Success = Matcher.matchAddr(Address, 0);
4135  (void)Success; assert(Success && "Couldn't select *anything*?");
4136 
4137  // The match was to check the profitability, the changes made are not
4138  // part of the original matcher. Therefore, they should be dropped
4139  // otherwise the original matcher will not present the right state.
4140  TPT.rollback(LastKnownGood);
4141 
4142  // If the match didn't cover I, then it won't be shared by it.
4143  if (!is_contained(MatchedAddrModeInsts, I))
4144  return false;
4145 
4146  MatchedAddrModeInsts.clear();
4147  }
4148 
4149  return true;
4150 }
4151 
4152 /// Return true if the specified values are defined in a
4153 /// different basic block than BB.
4154 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
4155  if (Instruction *I = dyn_cast<Instruction>(V))
4156  return I->getParent() != BB;
4157  return false;
4158 }
4159 
4160 /// Sink addressing mode computation immediate before MemoryInst if doing so
4161 /// can be done without increasing register pressure. The need for the
4162 /// register pressure constraint means this can end up being an all or nothing
4163 /// decision for all uses of the same addressing computation.
4164 ///
4165 /// Load and Store Instructions often have addressing modes that can do
4166 /// significant amounts of computation. As such, instruction selection will try
4167 /// to get the load or store to do as much computation as possible for the
4168 /// program. The problem is that isel can only see within a single block. As
4169 /// such, we sink as much legal addressing mode work into the block as possible.
4170 ///
4171 /// This method is used to optimize both load/store and inline asms with memory
4172 /// operands. It's also used to sink addressing computations feeding into cold
4173 /// call sites into their (cold) basic block.
4174 ///
4175 /// The motivation for handling sinking into cold blocks is that doing so can
4176 /// both enable other address mode sinking (by satisfying the register pressure
4177 /// constraint above), and reduce register pressure globally (by removing the
4178 /// addressing mode computation from the fast path entirely.).
4179 bool CodeGenPrepare::optimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
4180  Type *AccessTy, unsigned AddrSpace) {
4181  Value *Repl = Addr;
4182 
4183  // Try to collapse single-value PHI nodes. This is necessary to undo
4184  // unprofitable PRE transformations.
4185  SmallVector<Value*, 8> worklist;
4186  SmallPtrSet<Value*, 16> Visited;
4187  worklist.push_back(Addr);
4188 
4189  // Use a worklist to iteratively look through PHI and select nodes, and
4190  // ensure that the addressing mode obtained from the non-PHI/select roots of
4191  // the graph are compatible.
4192  bool PhiOrSelectSeen = false;
4193  SmallVector<Instruction*, 16> AddrModeInsts;
4194  const SimplifyQuery SQ(*DL, TLInfo);
4195  AddressingModeCombiner AddrModes(SQ, { Addr, MemoryInst->getParent() });
4196  TypePromotionTransaction TPT(RemovedInsts);
4197  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4198  TPT.getRestorationPoint();
4199  while (!worklist.empty()) {
4200  Value *V = worklist.back();
4201  worklist.pop_back();
4202 
4203  // We allow traversing cyclic Phi nodes.
4204  // In case of success after this loop we ensure that traversing through
4205  // Phi nodes ends up with all cases to compute address of the form
4206  // BaseGV + Base + Scale * Index + Offset
4207  // where Scale and Offset are constans and BaseGV, Base and Index
4208  // are exactly the same Values in all cases.
4209  // It means that BaseGV, Scale and Offset dominate our memory instruction
4210  // and have the same value as they had in address computation represented
4211  // as Phi. So we can safely sink address computation to memory instruction.
4212  if (!Visited.insert(V).second)
4213  continue;
4214 
4215  // For a PHI node, push all of its incoming values.
4216  if (PHINode *P = dyn_cast<PHINode>(V)) {
4217  for (Value *IncValue : P->incoming_values())
4218  worklist.push_back(IncValue);
4219  PhiOrSelectSeen = true;
4220  continue;
4221  }
4222  // Similar for select.
4223  if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
4224  worklist.push_back(SI->getFalseValue());
4225  worklist.push_back(SI->getTrueValue());
4226  PhiOrSelectSeen = true;
4227  continue;
4228  }
4229 
4230  // For non-PHIs, determine the addressing mode being computed. Note that
4231  // the result may differ depending on what other uses our candidate
4232  // addressing instructions might have.
4233  AddrModeInsts.clear();
4234  ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
4235  V, AccessTy, AddrSpace, MemoryInst, AddrModeInsts, *TLI, *TRI,
4236  InsertedInsts, PromotedInsts, TPT);
4237  NewAddrMode.OriginalValue = V;
4238 
4239  if (!AddrModes.addNewAddrMode(NewAddrMode))
4240  break;
4241  }
4242 
4243  // Try to combine the AddrModes we've collected. If we couldn't collect any,
4244  // or we have multiple but either couldn't combine them or combining them
4245  // wouldn't do anything useful, bail out now.
4246  if (!AddrModes.combineAddrModes()) {
4247  TPT.rollback(LastKnownGood);
4248  return false;
4249  }
4250  TPT.commit();
4251 
4252  // Get the combined AddrMode (or the only AddrMode, if we only had one).
4253  ExtAddrMode AddrMode = AddrModes.getAddrMode();
4254 
4255  // If all the instructions matched are already in this BB, don't do anything.
4256  // If we saw a Phi node then it is not local definitely, and if we saw a select
4257  // then we want to push the address calculation past it even if it's already
4258  // in this BB.
4259  if (!PhiOrSelectSeen && none_of(AddrModeInsts, [&](Value *V) {
4260  return IsNonLocalValue(V, MemoryInst->getParent());
4261  })) {
4262  DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
4263  return false;
4264  }
4265 
4266  // Insert this computation right after this user. Since our caller is
4267  // scanning from the top of the BB to the bottom, reuse of the expr are
4268  // guaranteed to happen later.
4269  IRBuilder<> Builder(MemoryInst);
4270 
4271  // Now that we determined the addressing expression we want to use and know
4272  // that we have to sink it into this block. Check to see if we have already
4273  // done this for some other load/store instr in this block. If so, reuse
4274  // the computation. Before attempting reuse, check if the address is valid
4275  // as it may have been erased.
4276 
4277  WeakTrackingVH SunkAddrVH = SunkAddrs[Addr];
4278 
4279  Value * SunkAddr = SunkAddrVH.pointsToAliveValue() ? SunkAddrVH : nullptr;
4280  if (SunkAddr) {
4281  DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
4282  << *MemoryInst << "\n");
4283  if (SunkAddr->getType() != Addr->getType())
4284  SunkAddr = Builder.CreatePointerCast(SunkAddr, Addr->getType());
4285  } else if (AddrSinkUsingGEPs ||
4286  (!AddrSinkUsingGEPs.getNumOccurrences() && TM &&
4287  SubtargetInfo->useAA())) {
4288  // By default, we use the GEP-based method when AA is used later. This
4289  // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
4290  DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
4291  << *MemoryInst << "\n");
4292  Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
4293  Value *ResultPtr = nullptr, *ResultIndex = nullptr;
4294 
4295  // First, find the pointer.
4296  if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
4297  ResultPtr = AddrMode.BaseReg;
4298  AddrMode.BaseReg = nullptr;
4299  }
4300 
4301  if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
4302  // We can't add more than one pointer together, nor can we scale a
4303  // pointer (both of which seem meaningless).
4304  if (ResultPtr || AddrMode.Scale != 1)
4305  return false;
4306 
4307  ResultPtr = AddrMode.ScaledReg;
4308  AddrMode.Scale = 0;
4309  }
4310 
4311  // It is only safe to sign extend the BaseReg if we know that the math
4312  // required to create it did not overflow before we extend it. Since
4313  // the original IR value was tossed in favor of a constant back when
4314  // the AddrMode was created we need to bail out gracefully if widths
4315  // do not match instead of extending it.
4316  //
4317  // (See below for code to add the scale.)
4318  if (AddrMode.Scale) {
4319  Type *ScaledRegTy = AddrMode.ScaledReg->getType();
4320  if (cast<IntegerType>(IntPtrTy)->getBitWidth() >
4321  cast<IntegerType>(ScaledRegTy)->getBitWidth())
4322  return false;
4323  }
4324 
4325  if (AddrMode.BaseGV) {
4326  if (ResultPtr)
4327  return false;
4328 
4329  ResultPtr = AddrMode.BaseGV;
4330  }
4331 
4332  // If the real base value actually came from an inttoptr, then the matcher
4333  // will look through it and provide only the integer value. In that case,
4334  // use it here.
4335  if (!DL->isNonIntegralPointerType(Addr->getType())) {
4336  if (!ResultPtr && AddrMode.BaseReg) {
4337  ResultPtr = Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(),
4338  "sunkaddr");
4339  AddrMode.BaseReg = nullptr;
4340  } else if (!ResultPtr && AddrMode.Scale == 1) {
4341  ResultPtr = Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(),
4342  "sunkaddr");
4343  AddrMode.Scale = 0;
4344  }
4345  }
4346 
4347  if (!ResultPtr &&
4348  !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
4349  SunkAddr = Constant::getNullValue(Addr->getType());
4350  } else if (!ResultPtr) {
4351  return false;
4352  } else {
4353  Type *I8PtrTy =
4354  Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
4355  Type *I8Ty = Builder.getInt8Ty();
4356 
4357  // Start with the base register. Do this first so that subsequent address
4358  // matching finds it last, which will prevent it from trying to match it
4359  // as the scaled value in case it happens to be a mul. That would be
4360  // problematic if we've sunk a different mul for the scale, because then
4361  // we'd end up sinking both muls.
4362  if (AddrMode.BaseReg) {
4363  Value *V = AddrMode.BaseReg;
4364  if (V->getType() != IntPtrTy)
4365  V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
4366 
4367  ResultIndex = V;
4368  }
4369 
4370  // Add the scale value.
4371  if (AddrMode.Scale) {
4372  Value *V = AddrMode.ScaledReg;
4373  if (V->getType() == IntPtrTy) {
4374  // done.
4375  } else {
4376  assert(cast<IntegerType>(IntPtrTy)->getBitWidth() <
4377  cast<IntegerType>(V->getType())->getBitWidth() &&
4378  "We can't transform if ScaledReg is too narrow");
4379  V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
4380  }
4381 
4382  if (AddrMode.Scale != 1)
4383  V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
4384  "sunkaddr");
4385  if (ResultIndex)
4386  ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
4387  else
4388  ResultIndex = V;
4389  }
4390 
4391  // Add in the Base Offset if present.
4392  if (AddrMode.BaseOffs) {
4393  Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
4394  if (ResultIndex) {
4395  // We need to add this separately from the scale above to help with
4396  // SDAG consecutive load/store merging.
4397  if (ResultPtr->getType() != I8PtrTy)
4398  ResultPtr = Builder.CreatePointerCast(ResultPtr, I8PtrTy);
4399  ResultPtr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
4400  }
4401 
4402  ResultIndex = V;
4403  }
4404 
4405  if (!ResultIndex) {
4406  SunkAddr = ResultPtr;
4407  } else {
4408  if (ResultPtr->getType() != I8PtrTy)
4409  ResultPtr = Builder.CreatePointerCast(ResultPtr, I8PtrTy);
4410  SunkAddr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
4411  }
4412 
4413  if (SunkAddr->getType() != Addr->getType())
4414  SunkAddr = Builder.CreatePointerCast(SunkAddr, Addr->getType());
4415  }
4416  } else {
4417  // We'd require a ptrtoint/inttoptr down the line, which we can't do for
4418  // non-integral pointers, so in that case bail out now.
4419  Type *BaseTy = AddrMode.BaseReg ? AddrMode.BaseReg->getType() : nullptr;
4420  Type *ScaleTy = AddrMode.Scale ? AddrMode.ScaledReg->getType() : nullptr;
4421  PointerType *BasePtrTy = dyn_cast_or_null<PointerType>(BaseTy);
4422  PointerType *ScalePtrTy = dyn_cast_or_null<PointerType>(ScaleTy);
4423  if (DL->isNonIntegralPointerType(Addr->getType()) ||
4424  (BasePtrTy && DL->isNonIntegralPointerType(BasePtrTy)) ||
4425  (ScalePtrTy && DL->isNonIntegralPointerType(ScalePtrTy)) ||
4426  (AddrMode.BaseGV &&
4427  DL->isNonIntegralPointerType(AddrMode.BaseGV->getType())))
4428  return false;
4429 
4430  DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
4431  << *MemoryInst << "\n");
4432  Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
4433  Value *Result = nullptr;
4434 
4435  // Start with the base register. Do this first so that subsequent address
4436  // matching finds it last, which will prevent it from trying to match it
4437  // as the scaled value in case it happens to be a mul. That would be
4438  // problematic if we've sunk a different mul for the scale, because then
4439  // we'd end up sinking both muls.
4440  if (AddrMode.BaseReg) {
4441  Value *V = AddrMode.BaseReg;
4442  if (V->getType()->isPointerTy())
4443  V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
4444  if (V->getType() != IntPtrTy)
4445  V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
4446  Result = V;
4447  }
4448 
4449  // Add the scale value.
4450  if (AddrMode.Scale) {
4451  Value *V = AddrMode.ScaledReg;
4452  if (V->getType() == IntPtrTy) {
4453  // done.
4454  } else if (V->getType()->isPointerTy()) {
4455  V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
4456  } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
4457  cast<IntegerType>(V->getType())->getBitWidth()) {
4458  V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
4459  } else {
4460  // It is only safe to sign extend the BaseReg if we know that the math
4461  // required to create it did not overflow before we extend it. Since
4462  // the original IR value was tossed in favor of a constant back when
4463  // the AddrMode was created we need to bail out gracefully if widths
4464  // do not match instead of extending it.
4465  Instruction *I = dyn_cast_or_null<Instruction>(Result);
4466  if (I && (Result != AddrMode.BaseReg))
4467  I->eraseFromParent();
4468  return false;
4469  }
4470  if (AddrMode.Scale != 1)
4471  V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
4472  "sunkaddr");
4473  if (Result)
4474  Result = Builder.CreateAdd(Result, V, "sunkaddr");
4475  else
4476  Result = V;
4477  }
4478 
4479  // Add in the BaseGV if present.
4480  if (AddrMode.BaseGV) {
4481  Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
4482  if (Result)
4483  Result = Builder.CreateAdd(Result, V, "sunkaddr");
4484  else
4485  Result = V;
4486  }
4487 
4488  // Add in the Base Offset if present.
4489  if (AddrMode.BaseOffs) {
4490  Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
4491  if (Result)
4492  Result = Builder.CreateAdd(Result, V, "sunkaddr");
4493  else
4494  Result = V;
4495  }
4496 
4497  if (!Result)
4498  SunkAddr = Constant::getNullValue(Addr->getType());
4499  else
4500  SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
4501  }
4502 
4503  MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
4504  // Store the newly computed address into the cache. In the case we reused a
4505  // value, this should be idempotent.
4506  SunkAddrs[Addr] = WeakTrackingVH(SunkAddr);
4507 
4508  // If we have no uses, recursively delete the value and all dead instructions
4509  // using it.
4510  if (Repl->use_empty()) {
4511  // This can cause recursive deletion, which can invalidate our iterator.
4512  // Use a WeakTrackingVH to hold onto it in case this happens.
4513  Value *CurValue = &*CurInstIterator;
4514  WeakTrackingVH IterHandle(CurValue);
4515  BasicBlock *BB = CurInstIterator->getParent();
4516 
4518 
4519  if (IterHandle != CurValue) {
4520  // If the iterator instruction was recursively deleted, start over at the
4521  // start of the block.
4522  CurInstIterator = BB->begin();
4523  SunkAddrs.clear();
4524  }
4525  }
4526  ++NumMemoryInsts;
4527  return true;
4528 }
4529 
4530 /// If there are any memory operands, use OptimizeMemoryInst to sink their
4531 /// address computing into the block when possible / profitable.
4532 bool CodeGenPrepare::optimizeInlineAsmInst(CallInst *CS) {
4533  bool MadeChange = false;
4534 
4535  const TargetRegisterInfo *TRI =
4536  TM->getSubtargetImpl(*CS->getFunction())->getRegisterInfo();
4537  TargetLowering::AsmOperandInfoVector TargetConstraints =
4538  TLI->ParseConstraints(*DL, TRI, CS);
4539  unsigned ArgNo = 0;
4540  for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
4541  TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
4542 
4543  // Compute the constraint code and ConstraintType to use.
4544  TLI->ComputeConstraintToUse(OpInfo, SDValue());
4545 
4546  if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
4547  OpInfo.isIndirect) {
4548  Value *OpVal = CS->getArgOperand(ArgNo++);
4549  MadeChange |= optimizeMemoryInst(CS, OpVal, OpVal->getType(), ~0u);
4550  } else if (OpInfo.Type == InlineAsm::isInput)
4551  ArgNo++;
4552  }
4553 
4554  return MadeChange;
4555 }
4556 
4557 /// \brief Check if all the uses of \p Val are equivalent (or free) zero or
4558 /// sign extensions.
4559 static bool hasSameExtUse(Value *Val, const TargetLowering &TLI) {
4560  assert(!Val->use_empty() && "Input must have at least one use");
4561  const Instruction *FirstUser = cast<Instruction>(*Val->user_begin());
4562  bool IsSExt = isa<SExtInst>(FirstUser);
4563  Type *ExtTy = FirstUser->getType();
4564  for (const User *U : Val->users()) {
4565  const Instruction *UI = cast<Instruction>(U);
4566  if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI)))
4567  return false;
4568  Type *CurTy = UI->getType();
4569  // Same input and output types: Same instruction after CSE.
4570  if (CurTy == ExtTy)
4571  continue;
4572 
4573  // If IsSExt is true, we are in this situation:
4574  // a = Val
4575  // b = sext ty1 a to ty2
4576  // c = sext ty1 a to ty3
4577  // Assuming ty2 is shorter than ty3, this could be turned into:
4578  // a = Val
4579  // b = sext ty1 a to ty2
4580  // c = sext ty2 b to ty3
4581  // However, the last sext is not free.
4582  if (IsSExt)
4583  return false;
4584 
4585  // This is a ZExt, maybe this is free to extend from one type to another.
4586  // In that case, we would not account for a different use.
4587  Type *NarrowTy;
4588  Type *LargeTy;
4589  if (ExtTy->getScalarType()->getIntegerBitWidth() >
4590  CurTy->getScalarType()->getIntegerBitWidth()) {
4591  NarrowTy = CurTy;
4592  LargeTy = ExtTy;
4593  } else {
4594  NarrowTy = ExtTy;
4595  LargeTy = CurTy;
4596  }
4597 
4598  if (!TLI.isZExtFree(NarrowTy, LargeTy))
4599  return false;
4600  }
4601  // All uses are the same or can be derived from one another for free.
4602  return true;
4603 }
4604 
4605 /// \brief Try to speculatively promote extensions in \p Exts and continue
4606 /// promoting through newly promoted operands recursively as far as doing so is
4607 /// profitable. Save extensions profitably moved up, in \p ProfitablyMovedExts.
4608 /// When some promotion happened, \p TPT contains the proper state to revert
4609 /// them.
4610 ///
4611 /// \return true if some promotion happened, false otherwise.
4612 bool CodeGenPrepare::tryToPromoteExts(
4613  TypePromotionTransaction &TPT, const SmallVectorImpl<Instruction *> &Exts,
4614  SmallVectorImpl<Instruction *> &ProfitablyMovedExts,
4615  unsigned CreatedInstsCost) {
4616  bool Promoted = false;
4617 
4618  // Iterate over all the extensions to try to promote them.
4619  for (auto I : Exts) {
4620  // Early check if we directly have ext(load).
4621  if (isa<LoadInst>(I->getOperand(0))) {
4622  ProfitablyMovedExts.push_back(I);
4623  continue;
4624  }
4625 
4626  // Check whether or not we want to do any promotion. The reason we have
4627  // this check inside the for loop is to catch the case where an extension
4628  // is directly fed by a load because in such case the extension can be moved
4629  // up without any promotion on its operands.
4630  if (!TLI || !TLI->enableExtLdPromotion() || DisableExtLdPromotion)
4631  return false;
4632 
4633  // Get the action to perform the promotion.
4634  TypePromotionHelper::Action TPH =
4635  TypePromotionHelper::getAction(I, InsertedInsts, *TLI, PromotedInsts);
4636  // Check if we can promote.
4637  if (!TPH) {
4638  // Save the current extension as we cannot move up through its operand.
4639  ProfitablyMovedExts.push_back(I);
4640  continue;
4641  }
4642 
4643  // Save the current state.
4644  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4645  TPT.getRestorationPoint();
4647  unsigned NewCreatedInstsCost = 0;
4648  unsigned ExtCost = !TLI->isExtFree(I);
4649  // Promote.
4650  Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost,
4651  &NewExts, nullptr, *TLI);
4652  assert(PromotedVal &&
4653  "TypePromotionHelper should have filtered out those cases");
4654 
4655  // We would be able to merge only one extension in a load.
4656  // Therefore, if we have more than 1 new extension we heuristically
4657  // cut this search path, because it means we degrade the code quality.
4658  // With exactly 2, the transformation is neutral, because we will merge
4659  // one extension but leave one. However, we optimistically keep going,
4660  // because the new extension may be removed too.
4661  long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost;
4662  // FIXME: It would be possible to propagate a negative value instead of
4663  // conservatively ceiling it to 0.
4664  TotalCreatedInstsCost =
4665  std::max((long long)0, (TotalCreatedInstsCost - ExtCost));
4666  if (!StressExtLdPromotion &&
4667  (TotalCreatedInstsCost > 1 ||
4668  !isPromotedInstructionLegal(*TLI, *DL, PromotedVal))) {
4669  // This promotion is not profitable, rollback to the previous state, and
4670  // save the current extension in ProfitablyMovedExts as the latest
4671  // speculative promotion turned out to be unprofitable.
4672  TPT.rollback(LastKnownGood);
4673  ProfitablyMovedExts.push_back(I);
4674  continue;
4675  }
4676  // Continue promoting NewExts as far as doing so is profitable.
4677  SmallVector<Instruction *, 2> NewlyMovedExts;
4678  (void)tryToPromoteExts(TPT, NewExts, NewlyMovedExts, TotalCreatedInstsCost);
4679  bool NewPromoted = false;
4680  for (auto ExtInst : NewlyMovedExts) {
4681  Instruction *MovedExt = cast<Instruction>(ExtInst);
4682  Value *ExtOperand = MovedExt->getOperand(0);
4683  // If we have reached to a load, we need this extra profitability check
4684  // as it could potentially be merged into an ext(load).
4685  if (isa<LoadInst>(ExtOperand) &&
4686  !(StressExtLdPromotion || NewCreatedInstsCost <= ExtCost ||
4687  (ExtOperand->hasOneUse() || hasSameExtUse(ExtOperand, *TLI))))
4688  continue;
4689 
4690  ProfitablyMovedExts.push_back(MovedExt);
4691  NewPromoted = true;
4692  }
4693 
4694  // If none of speculative promotions for NewExts is profitable, rollback
4695  // and save the current extension (I) as the last profitable extension.
4696  if (!NewPromoted) {
4697  TPT.rollback(LastKnownGood);
4698  ProfitablyMovedExts.push_back(I);
4699  continue;
4700  }
4701  // The promotion is profitable.
4702  Promoted = true;
4703  }
4704  return Promoted;
4705 }
4706 
4707 /// Merging redundant sexts when one is dominating the other.
4708 bool CodeGenPrepare::mergeSExts(Function &F) {
4709  DominatorTree DT(F);
4710  bool Changed = false;
4711  for (auto &Entry : ValToSExtendedUses) {
4712  SExts &Insts = Entry.second;
4713  SExts CurPts;
4714  for (Instruction *Inst : Insts) {
4715  if (RemovedInsts.count(Inst) || !isa<SExtInst>(Inst) ||
4716  Inst->getOperand(0) != Entry.first)
4717  continue;
4718  bool inserted = false;
4719  for (auto &Pt : CurPts) {
4720  if (DT.dominates(Inst, Pt)) {
4721  Pt->replaceAllUsesWith(Inst);
4722  RemovedInsts.insert(Pt);
4723  Pt->removeFromParent();
4724  Pt = Inst;
4725  inserted = true;
4726  Changed = true;
4727  break;
4728  }
4729  if (!DT.dominates(Pt, Inst))
4730  // Give up if we need to merge in a common dominator as the
4731  // expermients show it is not profitable.
4732  continue;
4733  Inst->replaceAllUsesWith(Pt);
4734  RemovedInsts.insert(Inst);
4735  Inst->removeFromParent();
4736  inserted = true;
4737  Changed = true;
4738  break;
4739  }
4740  if (!inserted)
4741  CurPts.push_back(Inst);
4742  }
4743  }
4744  return Changed;
4745 }
4746 
4747 /// Return true, if an ext(load) can be formed from an extension in
4748 /// \p MovedExts.
4749 bool CodeGenPrepare::canFormExtLd(
4750  const SmallVectorImpl<Instruction *> &MovedExts, LoadInst *&LI,
4751  Instruction *&Inst, bool HasPromoted) {
4752  for (auto *MovedExtInst : MovedExts) {
4753  if (isa<LoadInst>(MovedExtInst->getOperand(0))) {
4754  LI = cast<LoadInst>(MovedExtInst->getOperand(0));
4755  Inst = MovedExtInst;
4756  break;
4757  }
4758  }
4759  if (!LI)
4760  return false;
4761 
4762  // If they're already in the same block, there's nothing to do.
4763  // Make the cheap checks first if we did not promote.
4764  // If we promoted, we need to check if it is indeed profitable.
4765  if (!HasPromoted && LI->getParent() == Inst->getParent())
4766  return false;
4767 
4768  return TLI->isExtLoad(LI, Inst, *DL);
4769 }
4770 
4771 /// Move a zext or sext fed by a load into the same basic block as the load,
4772 /// unless conditions are unfavorable. This allows SelectionDAG to fold the
4773 /// extend into the load.
4774 ///
4775 /// E.g.,
4776 /// \code
4777 /// %ld = load i32* %addr
4778 /// %add = add nuw i32 %ld, 4
4779 /// %zext = zext i32 %add to i64
4780 // \endcode
4781 /// =>
4782 /// \code
4783 /// %ld = load i32* %addr
4784 /// %zext = zext i32 %ld to i64
4785 /// %add = add nuw i64 %zext, 4
4786 /// \encode
4787 /// Note that the promotion in %add to i64 is done in tryToPromoteExts(), which
4788 /// allow us to match zext(load i32*) to i64.
4789 ///
4790 /// Also, try to promote the computations used to obtain a sign extended
4791 /// value used into memory accesses.
4792 /// E.g.,
4793 /// \code
4794 /// a = add nsw i32 b, 3
4795 /// d = sext i32 a to i64
4796 /// e = getelementptr ..., i64 d
4797 /// \endcode
4798 /// =>
4799 /// \code
4800 /// f = sext i32 b to i64
4801 /// a = add nsw i64 f, 3
4802 /// e = getelementptr ..., i64 a
4803 /// \endcode
4804 ///
4805 /// \p Inst[in/out] the extension may be modified during the process if some
4806 /// promotions apply.
4807 bool CodeGenPrepare::optimizeExt(Instruction *&Inst) {
4808  // ExtLoad formation and address type promotion infrastructure requires TLI to
4809  // be effective.
4810  if (!TLI)
4811  return false;
4812 
4813  bool AllowPromotionWithoutCommonHeader = false;
4814  /// See if it is an interesting sext operations for the address type
4815  /// promotion before trying to promote it, e.g., the ones with the right
4816  /// type and used in memory accesses.
4817  bool ATPConsiderable = TTI->shouldConsiderAddressTypePromotion(
4818  *Inst, AllowPromotionWithoutCommonHeader);
4819  TypePromotionTransaction TPT(RemovedInsts);
4820  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4821  TPT.getRestorationPoint();
4823  SmallVector<Instruction *, 2> SpeculativelyMovedExts;
4824  Exts.push_back(Inst);
4825 
4826  bool HasPromoted = tryToPromoteExts(TPT, Exts, SpeculativelyMovedExts);
4827 
4828  // Look for a load being extended.
4829  LoadInst *LI = nullptr;
4830  Instruction *ExtFedByLoad;
4831 
4832  // Try to promote a chain of computation if it allows to form an extended
4833  // load.
4834  if (canFormExtLd(SpeculativelyMovedExts, LI, ExtFedByLoad, HasPromoted)) {
4835  assert(LI && ExtFedByLoad && "Expect a valid load and extension");
4836  TPT.commit();
4837  // Move the extend into the same block as the load
4838  ExtFedByLoad->moveAfter(LI);
4839  // CGP does not check if the zext would be speculatively executed when moved
4840  // to the same basic block as the load. Preserving its original location
4841  // would pessimize the debugging experience, as well as negatively impact
4842  // the quality of sample pgo. We don't want to use "line 0" as that has a
4843  // size cost in the line-table section and logically the zext can be seen as
4844  // part of the load. Therefore we conservatively reuse the same debug
4845  // location for the load and the zext.
4846  ExtFedByLoad->setDebugLoc(LI->getDebugLoc());
4847  ++NumExtsMoved;
4848  Inst = ExtFedByLoad;
4849  return true;
4850  }
4851 
4852  // Continue promoting SExts if known as considerable depending on targets.
4853  if (ATPConsiderable &&
4854  performAddressTypePromotion(Inst, AllowPromotionWithoutCommonHeader,
4855  HasPromoted, TPT, SpeculativelyMovedExts))
4856  return true;
4857 
4858  TPT.rollback(LastKnownGood);
4859  return false;
4860 }
4861 
4862 // Perform address type promotion if doing so is profitable.
4863 // If AllowPromotionWithoutCommonHeader == false, we should find other sext
4864 // instructions that sign extended the same initial value. However, if
4865 // AllowPromotionWithoutCommonHeader == true, we expect promoting the
4866 // extension is just profitable.
4867 bool CodeGenPrepare::performAddressTypePromotion(
4868  Instruction *&Inst, bool AllowPromotionWithoutCommonHeader,
4869  bool HasPromoted, TypePromotionTransaction &TPT,
4870  SmallVectorImpl<Instruction *> &SpeculativelyMovedExts) {
4871  bool Promoted = false;
4872  SmallPtrSet<Instruction *, 1> UnhandledExts;
4873  bool AllSeenFirst = true;
4874  for (auto I : SpeculativelyMovedExts) {
4875  Value *HeadOfChain = I->getOperand(0);
4877  SeenChainsForSExt.find(HeadOfChain);
4878  // If there is an unhandled SExt which has the same header, try to promote
4879  // it as well.
4880  if (AlreadySeen != SeenChainsForSExt.end()) {
4881  if (AlreadySeen->second != nullptr)
4882  UnhandledExts.insert(AlreadySeen->second);
4883  AllSeenFirst = false;
4884  }
4885  }
4886 
4887  if (!AllSeenFirst || (AllowPromotionWithoutCommonHeader &&
4888  SpeculativelyMovedExts.size() == 1)) {
4889  TPT.commit();
4890  if (HasPromoted)
4891  Promoted = true;
4892  for (auto I : SpeculativelyMovedExts) {
4893  Value *HeadOfChain = I->getOperand(0);
4894  SeenChainsForSExt[HeadOfChain] = nullptr;
4895  ValToSExtendedUses[HeadOfChain].push_back(I);
4896  }
4897  // Update Inst as promotion happen.
4898  Inst = SpeculativelyMovedExts.pop_back_val();
4899  } else {
4900  // This is the first chain visited from the header, keep the current chain
4901  // as unhandled. Defer to promote this until we encounter another SExt
4902  // chain derived from the same header.
4903  for (auto I : SpeculativelyMovedExts) {
4904  Value *HeadOfChain = I->getOperand(0);
4905  SeenChainsForSExt[HeadOfChain] = Inst;
4906  }
4907  return false;
4908  }
4909 
4910  if (!AllSeenFirst && !UnhandledExts.empty())
4911  for (auto VisitedSExt : UnhandledExts) {
4912  if (RemovedInsts.count(VisitedSExt))
4913  continue;
4914  TypePromotionTransaction TPT(RemovedInsts);
4917  Exts.push_back(VisitedSExt);
4918  bool HasPromoted = tryToPromoteExts(TPT, Exts, Chains);
4919  TPT.commit();
4920  if (HasPromoted)
4921  Promoted = true;
4922  for (auto I : Chains) {
4923  Value *HeadOfChain = I->getOperand(0);
4924  // Mark this as handled.
4925  SeenChainsForSExt[HeadOfChain] = nullptr;
4926  ValToSExtendedUses[HeadOfChain].push_back(I);
4927  }
4928  }
4929  return Promoted;
4930 }
4931 
4932 bool CodeGenPrepare::optimizeExtUses(Instruction *I) {
4933  BasicBlock *DefBB = I->getParent();
4934 
4935  // If the result of a {s|z}ext and its source are both live out, rewrite all
4936  // other uses of the source with result of extension.
4937  Value *Src = I->getOperand(0);
4938  if (Src->hasOneUse())
4939  return false;
4940 
4941  // Only do this xform if truncating is free.
4942  if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
4943  return false;
4944 
4945  // Only safe to perform the optimization if the source is also defined in
4946  // this block.
4947  if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
4948  return false;
4949 
4950  bool DefIsLiveOut = false;
4951  for (User *U : I->users()) {
4952  Instruction *UI = cast<Instruction>(U);
4953 
4954  // Figure out which BB this ext is used in.
4955  BasicBlock *UserBB = UI->getParent();
4956  if (UserBB == DefBB) continue;
4957  DefIsLiveOut = true;
4958  break;
4959  }
4960  if (!DefIsLiveOut)
4961  return false;
4962 
4963  // Make sure none of the uses are PHI nodes.
4964  for (User *U : Src->users()) {
4965  Instruction *UI = cast<Instruction>(U);
4966  BasicBlock *UserBB = UI->getParent();
4967  if (UserBB == DefBB) continue;
4968  // Be conservative. We don't want this xform to end up introducing
4969  // reloads just before load / store instructions.
4970  if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
4971  return false;
4972  }
4973 
4974  // InsertedTruncs - Only insert one trunc in each block once.
4975  DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
4976 
4977  bool MadeChange = false;
4978  for (Use &U : Src->uses()) {
4979  Instruction *User = cast<Instruction>(U.getUser());
4980 
4981  // Figure out which BB this ext is used in.
4982  BasicBlock *UserBB = User->getParent();
4983  if (UserBB == DefBB) continue;
4984 
4985  // Both src and def are live in this block. Rewrite the use.
4986  Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
4987 
4988  if (!InsertedTrunc) {
4989  BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
4990  assert(InsertPt != UserBB->end());
4991  InsertedTrunc = new TruncInst(I, Src->getType(), "", &*InsertPt);
4992  InsertedInsts.insert(InsertedTrunc);
4993  }
4994 
4995  // Replace a use of the {s|z}ext source with a use of the result.
4996  U = InsertedTrunc;
4997  ++NumExtUses;
4998  MadeChange = true;
4999  }
5000 
5001  return MadeChange;
5002 }
5003 
5004 // Find loads whose uses only use some of the loaded value's bits. Add an "and"
5005 // just after the load if the target can fold this into one extload instruction,
5006 // with the hope of eliminating some of the other later "and" instructions using
5007 // the loaded value. "and"s that are made trivially redundant by the insertion
5008 // of the new "and" are removed by this function, while others (e.g. those whose
5009 // path from the load goes through a phi) are left for isel to potentially
5010 // remove.
5011 //
5012 // For example:
5013 //
5014 // b0:
5015 // x = load i32
5016 // ...
5017 // b1:
5018 // y = and x, 0xff
5019 // z = use y
5020 //
5021 // becomes:
5022 //
5023 // b0:
5024 // x = load i32
5025 // x' = and x, 0xff
5026 // ...
5027 // b1:
5028 // z = use x'
5029 //
5030 // whereas:
5031 //
5032 // b0:
5033 // x1 = load i32
5034 // ...
5035 // b1:
5036 // x2 = load i32
5037 // ...
5038 // b2:
5039 // x = phi x1, x2
5040 // y = and x, 0xff
5041 //
5042 // becomes (after a call to optimizeLoadExt for each load):
5043 //
5044 // b0:
5045 // x1 = load i32
5046 // x1' = and x1, 0xff
5047 // ...
5048 // b1:
5049 // x2 = load i32
5050 // x2' = and x2, 0xff
5051 // ...
5052 // b2:
5053 // x = phi x1', x2'
5054 // y = and x, 0xff
5055 bool CodeGenPrepare::optimizeLoadExt(LoadInst *Load) {
5056  if (!Load->isSimple() ||
5057  !(Load->getType()->isIntegerTy() || Load->getType()->isPointerTy()))
5058  return false;
5059 
5060  // Skip loads we've already transformed.
5061  if (Load->hasOneUse() &&
5062  InsertedInsts.count(cast<Instruction>(*Load->user_begin())))
5063  return false;
5064 
5065  // Look at all uses of Load, looking through phis, to determine how many bits
5066  // of the loaded value are needed.
5069  SmallVector<Instruction *, 8> AndsToMaybeRemove;
5070  for (auto *U : Load->users())
5071  WorkList.push_back(cast<Instruction>(U));
5072 
5073  EVT LoadResultVT = TLI->getValueType(*DL, Load->getType());
5074  unsigned BitWidth = LoadResultVT.getSizeInBits();
5075  APInt DemandBits(BitWidth, 0);
5076  APInt WidestAndBits(BitWidth, 0);
5077 
5078  while (!WorkList.empty()) {
5079  Instruction *I = WorkList.back();
5080  WorkList.pop_back();
5081 
5082  // Break use-def graph loops.
5083  if (!Visited.insert(I).second)
5084  continue;
5085 
5086  // For a PHI node, push all of its users.
5087  if (auto *Phi = dyn_cast<PHINode>(I)) {
5088  for (auto *U : Phi->users())
5089  WorkList.push_back(cast<Instruction>(U));
5090  continue;
5091  }
5092 
5093  switch (I->getOpcode()) {
5094  case Instruction::And: {
5095  auto *AndC = dyn_cast<ConstantInt>(I->getOperand(1));
5096  if (!AndC)
5097  return false;
5098  APInt AndBits = AndC->getValue();
5099  DemandBits |= AndBits;
5100  // Keep track of the widest and mask we see.
5101  if (AndBits.ugt(WidestAndBits))
5102  WidestAndBits = AndBits;
5103  if (AndBits == WidestAndBits && I->getOperand(0) == Load)
5104  AndsToMaybeRemove.push_back(I);
5105  break;
5106  }
5107 
5108  case Instruction::Shl: {
5109  auto *ShlC = dyn_cast<ConstantInt>(I->getOperand(1));
5110  if (!ShlC)
5111  return false;
5112  uint64_t ShiftAmt = ShlC->getLimitedValue(BitWidth - 1);
5113  DemandBits.setLowBits(BitWidth - ShiftAmt);
5114  break;
5115  }
5116 
5117  case Instruction::Trunc: {
5118  EVT TruncVT = TLI->getValueType(*DL, I->getType());
5119  unsigned TruncBitWidth = TruncVT.getSizeInBits();
5120  DemandBits.setLowBits(TruncBitWidth);
5121  break;
5122  }
5123 
5124  default:
5125  return false;
5126  }
5127  }
5128 
5129  uint32_t ActiveBits = DemandBits.getActiveBits();
5130  // Avoid hoisting (and (load x) 1) since it is unlikely to be folded by the
5131  // target even if isLoadExtLegal says an i1 EXTLOAD is valid. For example,
5132  // for the AArch64 target isLoadExtLegal(ZEXTLOAD, i32, i1) returns true, but
5133  // (and (load x) 1) is not matched as a single instruction, rather as a LDR
5134  // followed by an AND.
5135  // TODO: Look into removing this restriction by fixing backends to either
5136  // return false for isLoadExtLegal for i1 or have them select this pattern to
5137  // a single instruction.
5138  //
5139  // Also avoid hoisting if we didn't see any ands with the exact DemandBits
5140  // mask, since these are the only ands that will be removed by isel.
5141  if (ActiveBits <= 1 || !DemandBits.isMask(ActiveBits) ||
5142  WidestAndBits != DemandBits)
5143  return false;
5144 
5145  LLVMContext &Ctx = Load->getType()->getContext();
5146  Type *TruncTy = Type::getIntNTy(Ctx, ActiveBits);
5147  EVT TruncVT = TLI->getValueType(*DL, TruncTy);
5148 
5149  // Reject cases that won't be matched as extloads.
5150  if (!LoadResultVT.bitsGT(TruncVT) || !TruncVT.isRound() ||
5151  !TLI->isLoadExtLegal(ISD::ZEXTLOAD, LoadResultVT, TruncVT))
5152  return false;
5153 
5154  IRBuilder<> Builder(Load->getNextNode());
5155  auto *NewAnd = dyn_cast<Instruction>(
5156  Builder.CreateAnd(Load, ConstantInt::get(Ctx, DemandBits)));
5157  // Mark this instruction as "inserted by CGP", so that other
5158  // optimizations don't touch it.
5159  InsertedInsts.insert(NewAnd);
5160 
5161  // Replace all uses of load with new and (except for the use of load in the
5162  // new and itself).
5163  Load->replaceAllUsesWith(NewAnd);
5164  NewAnd->setOperand(0, Load);
5165 
5166  // Remove any and instructions that are now redundant.
5167  for (auto *And : AndsToMaybeRemove)
5168  // Check that the and mask is the same as the one we decided to put on the
5169  // new and.
5170  if (cast<ConstantInt>(And->getOperand(1))->getValue() == DemandBits) {
5171  And->replaceAllUsesWith(NewAnd);
5172  if (&*CurInstIterator == And)
5173  CurInstIterator = std::next(And->getIterator());
5174  And->eraseFromParent();
5175  ++NumAndUses;
5176  }
5177 
5178  ++NumAndsAdded;
5179  return true;
5180 }
5181 
5182 /// Check if V (an operand of a select instruction) is an expensive instruction
5183 /// that is only used once.
5184 static bool sinkSelectOperand(const TargetTransformInfo *TTI, Value *V) {
5185  auto *I = dyn_cast<Instruction>(V);
5186  // If it's safe to speculatively execute, then it should not have side
5187  // effects; therefore, it's safe to sink and possibly *not* execute.
5188  return I && I->hasOneUse() && isSafeToSpeculativelyExecute(I) &&
5190 }
5191 
5192 /// Returns true if a SelectInst should be turned into an explicit branch.
5194  const TargetLowering *TLI,
5195  SelectInst *SI) {
5196  // If even a predictable select is cheap, then a branch can't be cheaper.
5197  if (!TLI->isPredictableSelectExpensive())
5198  return false;
5199 
5200  // FIXME: This should use the same heuristics as IfConversion to determine
5201  // whether a select is better represented as a branch.
5202 
5203  // If metadata tells us that the select condition is obviously predictable,
5204  // then we want to replace the select with a branch.
5205  uint64_t TrueWeight, FalseWeight;
5206  if (SI->extractProfMetadata(TrueWeight, FalseWeight)) {
5207  uint64_t Max = std::max(TrueWeight, FalseWeight);
5208  uint64_t Sum = TrueWeight + FalseWeight;
5209  if (Sum != 0) {
5210  auto Probability = BranchProbability::getBranchProbability(Max, Sum);
5211  if (Probability > TLI->getPredictableBranchThreshold())
5212  return true;
5213  }
5214  }
5215 
5216  CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
5217 
5218  // If a branch is predictable, an out-of-order CPU can avoid blocking on its
5219  // comparison condition. If the compare has more than one use, there's
5220  // probably another cmov or setcc around, so it's not worth emitting a branch.
5221  if (!Cmp || !Cmp->hasOneUse())
5222  return false;
5223 
5224  // If either operand of the select is expensive and only needed on one side
5225  // of the select, we should form a branch.
5226  if (sinkSelectOperand(TTI, SI->getTrueValue()) ||
5227  sinkSelectOperand(TTI, SI->getFalseValue()))
5228  return true;
5229 
5230  return false;
5231 }
5232 
5233 /// If \p isTrue is true, return the true value of \p SI, otherwise return
5234 /// false value of \p SI. If the true/false value of \p SI is defined by any
5235 /// select instructions in \p Selects, look through the defining select
5236 /// instruction until the true/false value is not defined in \p Selects.
5238  SelectInst *SI, bool isTrue,
5239  const SmallPtrSet<const Instruction *, 2> &Selects) {
5240  Value *V;
5241 
5242  for (SelectInst *DefSI = SI; DefSI != nullptr && Selects.count(DefSI);
5243  DefSI = dyn_cast<SelectInst>(V)) {
5244  assert(DefSI->getCondition() == SI->getCondition() &&
5245  "The condition of DefSI does not match with SI");
5246  V = (isTrue ? DefSI->getTrueValue() : DefSI->getFalseValue());
5247  }
5248  return V;
5249 }
5250 
5251 /// If we have a SelectInst that will likely profit from branch prediction,
5252 /// turn it into a branch.
5253 bool CodeGenPrepare::optimizeSelectInst(SelectInst *SI) {
5254  // Find all consecutive select instructions that share the same condition.
5256  ASI.push_back(SI);
5258  It != SI->getParent()->end(); ++It) {
5259  SelectInst *I = dyn_cast<SelectInst>(&*It);
5260  if (I && SI->getCondition() == I->getCondition()) {
5261  ASI.push_back(I);
5262  } else {
5263  break;
5264  }
5265  }
5266 
5267  SelectInst *LastSI = ASI.back();
5268  // Increment the current iterator to skip all the rest of select instructions
5269  // because they will be either "not lowered" or "all lowered" to branch.
5270  CurInstIterator = std::next(LastSI->getIterator());
5271 
5272  bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
5273 
5274  // Can we convert the 'select' to CF ?
5275  if (DisableSelectToBranch || OptSize || !TLI || VectorCond ||
5277  return false;
5278 
5280  if (VectorCond)
5281  SelectKind = TargetLowering::VectorMaskSelect;
5282  else if (SI->getType()->isVectorTy())
5284  else
5285  SelectKind = TargetLowering::ScalarValSelect;
5286 
5287  if (TLI->isSelectSupported(SelectKind) &&
5288  !isFormingBranchFromSelectProfitable(TTI, TLI, SI))
5289  return false;
5290 
5291  ModifiedDT = true;
5292 
5293  // Transform a sequence like this:
5294  // start:
5295  // %cmp = cmp uge i32 %a, %b
5296  // %sel = select i1 %cmp, i32 %c, i32 %d
5297  //
5298  // Into:
5299  // start:
5300  // %cmp = cmp uge i32 %a, %b
5301  // br i1 %cmp, label %select.true, label %select.false
5302  // select.true:
5303  // br label %select.end
5304  // select.false:
5305  // br label %select.end
5306  // select.end:
5307  // %sel = phi i32 [ %c, %select.true ], [ %d, %select.false ]
5308  //
5309  // In addition, we may sink instructions that produce %c or %d from
5310  // the entry block into the destination(s) of the new branch.
5311  // If the true or false blocks do not contain a sunken instruction, that
5312  // block and its branch may be optimized away. In that case, one side of the
5313  // first branch will point directly to select.end, and the corresponding PHI
5314  // predecessor block will be the start block.
5315 
5316  // First, we split the block containing the select into 2 blocks.
5317  BasicBlock *StartBlock = SI->getParent();
5318  BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(LastSI));
5319  BasicBlock *EndBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
5320 
5321  // Delete the unconditional branch that was just created by the split.
5322  StartBlock->getTerminator()->eraseFromParent();
5323 
5324  // These are the new basic blocks for the conditional branch.
5325  // At least one will become an actual new basic block.
5326  BasicBlock *TrueBlock = nullptr;
5327  BasicBlock *FalseBlock = nullptr;
5328  BranchInst *TrueBranch = nullptr;
5329  BranchInst *FalseBranch = nullptr;
5330 
5331  // Sink expensive instructions into the conditional blocks to avoid executing
5332  // them speculatively.
5333  for (SelectInst *SI : ASI) {
5334  if (sinkSelectOperand(TTI, SI->getTrueValue())) {
5335  if (TrueBlock == nullptr) {
5336  TrueBlock = BasicBlock::Create(SI->getContext(), "select.true.sink",
5337  EndBlock->getParent(), EndBlock);
5338  TrueBranch = BranchInst::Create(EndBlock, TrueBlock);
5339  }
5340  auto *TrueInst = cast<Instruction>(SI->getTrueValue());
5341  TrueInst->moveBefore(TrueBranch);
5342  }
5343  if (sinkSelectOperand(TTI, SI->getFalseValue())) {
5344  if (FalseBlock == nullptr) {
5345  FalseBlock = BasicBlock::Create(SI->getContext(), "select.false.sink",
5346  EndBlock->getParent(), EndBlock);
5347  FalseBranch = BranchInst::Create(EndBlock, FalseBlock);
5348  }
5349  auto *FalseInst = cast<Instruction>(SI->getFalseValue());
5350  FalseInst->moveBefore(FalseBranch);
5351  }
5352  }
5353 
5354  // If there was nothing to sink, then arbitrarily choose the 'false' side
5355  // for a new input value to the PHI.
5356  if (TrueBlock == FalseBlock) {
5357  assert(TrueBlock == nullptr &&
5358  "Unexpected basic block transform while optimizing select");
5359 
5360  FalseBlock = BasicBlock::Create(SI->getContext(), "select.false",
5361  EndBlock->getParent(), EndBlock);
5362  BranchInst::Create(EndBlock, FalseBlock);
5363  }
5364 
5365  // Insert the real conditional branch based on the original condition.
5366  // If we did not create a new block for one of the 'true' or 'false' paths
5367  // of the condition, it means that side of the branch goes to the end block
5368  // directly and the path originates from the start block from the point of
5369  // view of the new PHI.
5370  BasicBlock *TT, *FT;
5371  if (TrueBlock == nullptr) {
5372  TT = EndBlock;
5373  FT = FalseBlock;
5374  TrueBlock = StartBlock;
5375  } else if (FalseBlock == nullptr) {
5376  TT = TrueBlock;
5377  FT = EndBlock;
5378  FalseBlock = StartBlock;
5379  } else {
5380  TT = TrueBlock;
5381  FT = FalseBlock;
5382  }
5383  IRBuilder<>(SI).CreateCondBr(SI->getCondition(), TT, FT, SI);
5384 
5386  INS.insert(ASI.begin(), ASI.end());
5387  // Use reverse iterator because later select may use the value of the
5388  // earlier select, and we need to propagate value through earlier select
5389  // to get the PHI operand.
5390  for (auto It = ASI.rbegin(); It != ASI.rend(); ++It) {
5391  SelectInst *SI = *It;
5392  // The select itself is replaced with a PHI Node.
5393  PHINode *PN = PHINode::Create(SI->getType(), 2, "", &EndBlock->front());
5394  PN->takeName(SI);
5395  PN->addIncoming(getTrueOrFalseValue(SI, true, INS), TrueBlock);
5396  PN->addIncoming(getTrueOrFalseValue(SI, false, INS), FalseBlock);
5397 
5398  SI->replaceAllUsesWith(PN);
5399  SI->eraseFromParent();
5400  INS.erase(SI);
5401  ++NumSelectsExpanded;
5402  }
5403 
5404  // Instruct OptimizeBlock to skip to the next block.
5405  CurInstIterator = StartBlock->end();
5406  return true;
5407 }
5408 
5411  int SplatElem = -1;
5412  for (unsigned i = 0; i < Mask.size(); ++i) {
5413  if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
5414  return false;
5415  SplatElem = Mask[i];
5416  }
5417 
5418  return true;
5419 }
5420 
5421 /// Some targets have expensive vector shifts if the lanes aren't all the same
5422 /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
5423 /// it's often worth sinking a shufflevector splat down to its use so that
5424 /// codegen can spot all lanes are identical.
5425 bool CodeGenPrepare::optimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
5426  BasicBlock *DefBB = SVI->getParent();
5427 
5428  // Only do this xform if variable vector shifts are particularly expensive.
5429  if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
5430  return false;
5431 
5432  // We only expect better codegen by sinking a shuffle if we can recognise a
5433  // constant splat.
5434  if (!isBroadcastShuffle(SVI))
5435  return false;
5436 
5437  // InsertedShuffles - Only insert a shuffle in each block once.
5438  DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
5439 
5440  bool MadeChange = false;
5441  for (User *U : SVI->users()) {
5442  Instruction *UI = cast<Instruction>(U);
5443 
5444  // Figure out which BB this ext is used in.
5445  BasicBlock *UserBB = UI->getParent();
5446  if (UserBB == DefBB) continue;
5447 
5448  // For now only apply this when the splat is used by a shift instruction.
5449  if (!UI->isShift()) continue;
5450 
5451  // Everything checks out, sink the shuffle if the user's block doesn't
5452  // already have a copy.
5453  Instruction *&InsertedShuffle = InsertedShuffles[UserBB];
5454 
5455  if (!InsertedShuffle) {
5456  BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
5457  assert(InsertPt != UserBB->end());
5458  InsertedShuffle =
5459  new ShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1),
5460  SVI->getOperand(2), "", &*InsertPt);
5461  }
5462 
5463  UI->replaceUsesOfWith(SVI, InsertedShuffle);
5464  MadeChange = true;
5465  }
5466 
5467  // If we removed all uses, nuke the shuffle.
5468  if (SVI->use_empty()) {
5469  SVI->eraseFromParent();
5470  MadeChange = true;
5471  }
5472 
5473  return MadeChange;
5474 }
5475 
5476 bool CodeGenPrepare::optimizeSwitchInst(SwitchInst *SI) {
5477  if (!TLI || !DL)
5478  return false;
5479 
5480  Value *Cond = SI->getCondition();
5481  Type *OldType = Cond->getType();
5482  LLVMContext &Context = Cond->getContext();
5483  MVT RegType = TLI->getRegisterType(Context, TLI->getValueType(*DL, OldType));
5484  unsigned RegWidth = RegType.getSizeInBits();
5485 
5486  if (RegWidth <= cast<IntegerType>(OldType)->getBitWidth())
5487  return false;
5488 
5489  // If the register width is greater than the type width, expand the condition
5490  // of the switch instruction and each case constant to the width of the
5491  // register. By widening the type of the switch condition, subsequent
5492  // comparisons (for case comparisons) will not need to be extended to the
5493  // preferred register width, so we will potentially eliminate N-1 extends,
5494  // where N is the number of cases in the switch.
5495  auto *NewType = Type::getIntNTy(Context, RegWidth);
5496 
5497  // Zero-extend the switch condition and case constants unless the switch
5498  // condition is a function argument that is already being sign-extended.
5499  // In that case, we can avoid an unnecessary mask/extension by sign-extending
5500  // everything instead.
5501  Instruction::CastOps ExtType = Instruction::ZExt;
5502  if (auto *Arg = dyn_cast<Argument>(Cond))
5503  if (Arg->hasSExtAttr())
5504  ExtType = Instruction::SExt;
5505 
5506  auto *ExtInst = CastInst::Create(ExtType, Cond, NewType);
5507  ExtInst->insertBefore(SI);
5508  SI->setCondition(ExtInst);
5509  for (auto Case : SI->cases()) {
5510  APInt NarrowConst = Case.getCaseValue()->getValue();
5511  APInt WideConst = (ExtType == Instruction::ZExt) ?
5512  NarrowConst.zext(RegWidth) : NarrowConst.sext(RegWidth);
5513  Case.setValue(ConstantInt::get(Context, WideConst));
5514  }
5515 
5516  return true;
5517 }
5518 
5519 
5520 namespace {
5521 
5522 /// \brief Helper class to promote a scalar operation to a vector one.
5523 /// This class is used to move downward extractelement transition.
5524 /// E.g.,
5525 /// a = vector_op <2 x i32>
5526 /// b = extractelement <2 x i32> a, i32 0
5527 /// c = scalar_op b
5528 /// store c
5529 ///
5530 /// =>
5531 /// a = vector_op <2 x i32>
5532 /// c = vector_op a (equivalent to scalar_op on the related lane)
5533 /// * d = extractelement <2 x i32> c, i32 0
5534 /// * store d
5535 /// Assuming both extractelement and store can be combine, we get rid of the
5536 /// transition.
5537 class VectorPromoteHelper {
5538  /// DataLayout associated with the current module.
5539  const DataLayout &DL;
5540 
5541  /// Used to perform some checks on the legality of vector operations.
5542  const TargetLowering &TLI;
5543 
5544  /// Used to estimated the cost of the promoted chain.
5545  const TargetTransformInfo &TTI;
5546 
5547  /// The transition being moved downwards.
5548  Instruction *Transition;
5549 
5550  /// The sequence of instructions to be promoted.
5551  SmallVector<Instruction *, 4> InstsToBePromoted;
5552 
5553  /// Cost of combining a store and an extract.
5554  unsigned StoreExtractCombineCost;
5555 
5556  /// Instruction that will be combined with the transition.
5557  Instruction *CombineInst = nullptr;
5558 
5559  /// \brief The instruction that represents the current end of the transition.
5560  /// Since we are faking the promotion until we reach the end of the chain
5561  /// of computation, we need a way to get the current end of the transition.
5562  Instruction *getEndOfTransition() const {
5563  if (InstsToBePromoted.empty())
5564  return Transition;
5565  return InstsToBePromoted.back();
5566  }
5567 
5568  /// \brief Return the index of the original value in the transition.
5569  /// E.g., for "extractelement <2 x i32> c, i32 1" the original value,
5570  /// c, is at index 0.
5571  unsigned getTransitionOriginalValueIdx() const {
5572  assert(isa<ExtractElementInst>(Transition) &&
5573  "Other kind of transitions are not supported yet");
5574  return 0;
5575  }
5576 
5577  /// \brief Return the index of the index in the transition.
5578  /// E.g., for "extractelement <2 x i32> c, i32 0" the index
5579  /// is at index 1.
5580  unsigned getTransitionIdx() const {
5581  assert(isa<ExtractElementInst>(Transition) &&
5582  "Other kind of transitions are not supported yet");
5583  return 1;
5584  }
5585 
5586  /// \brief Get the type of the transition.
5587  /// This is the type of the original value.
5588  /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the
5589  /// transition is <2 x i32>.
5590  Type *getTransitionType() const {
5591  return Transition->getOperand(getTransitionOriginalValueIdx())->getType();
5592  }
5593 
5594  /// \brief Promote \p ToBePromoted by moving \p Def downward through.
5595  /// I.e., we have the following sequence:
5596  /// Def = Transition <ty1> a to <ty2>
5597  /// b = ToBePromoted <ty2> Def, ...
5598  /// =>
5599  /// b = ToBePromoted <ty1> a, ...
5600  /// Def = Transition <ty1> ToBePromoted to <ty2>
5601  void promoteImpl(Instruction *ToBePromoted);
5602 
5603  /// \brief Check whether or not it is profitable to promote all the
5604  /// instructions enqueued to be promoted.
5605  bool isProfitableToPromote() {
5606  Value *ValIdx = Transition->getOperand(getTransitionOriginalValueIdx());
5607  unsigned Index = isa<ConstantInt>(ValIdx)
5608  ? cast<ConstantInt>(ValIdx)->getZExtValue()
5609  : -1;
5610  Type *PromotedType = getTransitionType();
5611 
5612  StoreInst *ST = cast<StoreInst>(CombineInst);
5613  unsigned AS = ST->getPointerAddressSpace();
5614  unsigned Align = ST->getAlignment();
5615  // Check if this store is supported.
5617  TLI.getValueType(DL, ST->getValueOperand()->getType()), AS,
5618  Align)) {
5619  // If this is not supported, there is no way we can combine
5620  // the extract with the store.
5621  return false;
5622  }
5623 
5624  // The scalar chain of computation has to pay for the transition
5625  // scalar to vector.
5626  // The vector chain has to account for the combining cost.
5627  uint64_t ScalarCost =
5628  TTI.getVectorInstrCost(Transition->getOpcode(), PromotedType, Index);
5629  uint64_t VectorCost = StoreExtractCombineCost;
5630  for (const auto &Inst : InstsToBePromoted) {
5631  // Compute the cost.
5632  // By construction, all instructions being promoted are arithmetic ones.
5633  // Moreover, one argument is a constant that can be viewed as a splat
5634  // constant.
5635  Value *Arg0 = Inst->getOperand(0);
5636  bool IsArg0Constant = isa<UndefValue>(Arg0) || isa<ConstantInt>(Arg0) ||
5637  isa<ConstantFP>(Arg0);
5644  ScalarCost += TTI.getArithmeticInstrCost(
5645  Inst->getOpcode(), Inst->getType(), Arg0OVK, Arg1OVK);
5646  VectorCost += TTI.getArithmeticInstrCost(Inst->getOpcode(), PromotedType,
5647  Arg0OVK, Arg1OVK);
5648  }
5649  DEBUG(dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
5650  << ScalarCost << "\nVector: " << VectorCost << '\n');
5651  return ScalarCost > VectorCost;
5652  }
5653 
5654  /// \brief Generate a constant vector with \p Val with the same
5655  /// number of elements as the transition.
5656  /// \p UseSplat defines whether or not \p Val should be replicated
5657  /// across the whole vector.
5658  /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>,
5659  /// otherwise we generate a vector with as many undef as possible:
5660  /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only
5661  /// used at the index of the extract.
5662  Value *getConstantVector(Constant *Val, bool UseSplat) const {
5663  unsigned ExtractIdx = std::numeric_limits<unsigned>::max();
5664  if (!UseSplat) {
5665  // If we cannot determine where the constant must be, we have to
5666  // use a splat constant.
5667  Value *ValExtractIdx = Transition->getOperand(getTransitionIdx());
5668  if (ConstantInt *CstVal = dyn_cast<ConstantInt>(ValExtractIdx))
5669  ExtractIdx = CstVal->getSExtValue();
5670  else
5671  UseSplat = true;
5672  }
5673 
5674  unsigned End = getTransitionType()->getVectorNumElements();
5675  if (UseSplat)
5676  return ConstantVector::getSplat(End, Val);
5677 
5678  SmallVector<Constant *, 4> ConstVec;
5679  UndefValue *UndefVal = UndefValue::get(Val->getType());
5680  for (unsigned Idx = 0; Idx != End; ++Idx) {
5681  if (Idx == ExtractIdx)
5682  ConstVec.push_back(Val);
5683  else
5684  ConstVec.push_back(UndefVal);
5685  }
5686  return ConstantVector::get(ConstVec);
5687  }
5688 
5689  /// \brief Check if promoting to a vector type an operand at \p OperandIdx
5690  /// in \p Use can trigger undefined behavior.
5691  static bool canCauseUndefinedBehavior(const Instruction *Use,
5692  unsigned OperandIdx) {
5693  // This is not safe to introduce undef when the operand is on
5694  // the right hand side of a division-like instruction.
5695  if (OperandIdx != 1)
5696  return false;
5697  switch (Use->getOpcode()) {
5698  default:
5699  return false;
5700  case Instruction::SDiv:
5701  case Instruction::UDiv:
5702  case Instruction::SRem:
5703  case Instruction::URem:
5704  return true;
5705  case Instruction::FDiv:
5706  case Instruction::FRem:
5707  return !Use->hasNoNaNs();
5708  }
5709  llvm_unreachable(nullptr);
5710  }
5711 
5712 public:
5713  VectorPromoteHelper(const DataLayout &DL, const TargetLowering &TLI,
5714  const TargetTransformInfo &TTI, Instruction *Transition,
5715  unsigned CombineCost)
5716  : DL(DL), TLI(TLI), TTI(TTI), Transition(Transition),
5717  StoreExtractCombineCost(CombineCost) {
5718  assert(Transition && "Do not know how to promote null");
5719  }
5720 
5721  /// \brief Check if we can promote \p ToBePromoted to \p Type.
5722  bool canPromote(const Instruction *ToBePromoted) const {
5723  // We could support CastInst too.
5724  return isa<BinaryOperator>(ToBePromoted);
5725  }
5726 
5727  /// \brief Check if it is profitable to promote \p ToBePromoted
5728  /// by moving downward the transition through.
5729  bool shouldPromote(const Instruction *ToBePromoted) const {
5730  // Promote only if all the operands can be statically expanded.
5731  // Indeed, we do not want to introduce any new kind of transitions.
5732  for (const Use &U : ToBePromoted->operands()) {
5733  const Value *Val = U.get();
5734  if (Val == getEndOfTransition()) {
5735  // If the use is a division and the transition is on the rhs,
5736  // we cannot promote the operation, otherwise we may create a
5737  // division by zero.
5738  if (canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()))
5739  return false;
5740  continue;
5741  }
5742  if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) &&
5743  !isa<ConstantFP>(Val))
5744  return false;
5745  }
5746  // Check that the resulting operation is legal.
5747  int ISDOpcode = TLI.InstructionOpcodeToISD(ToBePromoted->getOpcode());
5748  if (!ISDOpcode)
5749  return false;
5750  return StressStoreExtract ||
5752  ISDOpcode, TLI.getValueType(DL, getTransitionType(), true));
5753  }
5754 
5755  /// \brief Check whether or not \p Use can be combined
5756  /// with the transition.
5757  /// I.e., is it possible to do Use(Transition) => AnotherUse?
5758  bool canCombine(const Instruction *Use) { return isa<StoreInst>(Use); }
5759 
5760  /// \brief Record \p ToBePromoted as part of the chain to be promoted.
5761  void enqueueForPromotion(Instruction *ToBePromoted) {
5762  InstsToBePromoted.push_back(ToBePromoted);
5763  }
5764 
5765  /// \brief Set the instruction that will be combined with the transition.
5766  void recordCombineInstruction(Instruction *ToBeCombined) {
5767  assert(canCombine(ToBeCombined) && "Unsupported instruction to combine");
5768  CombineInst = ToBeCombined;
5769  }
5770 
5771  /// \brief Promote all the instructions enqueued for promotion if it is
5772  /// is profitable.
5773  /// \return True if the promotion happened, false otherwise.
5774  bool promote() {
5775  // Check if there is something to promote.
5776  // Right now, if we do not have anything to combine with,
5777  // we assume the promotion is not profitable.
5778  if (InstsToBePromoted.empty() || !CombineInst)
5779  return false;
5780 
5781  // Check cost.
5782  if (!StressStoreExtract && !isProfitableToPromote())
5783  return false;
5784 
5785  // Promote.
5786  for (auto &ToBePromoted : InstsToBePromoted)
5787  promoteImpl(ToBePromoted);
5788  InstsToBePromoted.clear();
5789  return true;
5790  }
5791 };
5792