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"
35 #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(true),
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->getIndexSizeInBits(
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 DestAlign = getKnownAlignment(MI->getDest(), *DL);
1610  if (DestAlign > MI->getDestAlignment())
1611  MI->setDestAlignment(DestAlign);
1612  if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) {
1613  unsigned SrcAlign = getKnownAlignment(MTI->getSource(), *DL);
1614  if (SrcAlign > MTI->getSourceAlignment())
1615  MTI->setSourceAlignment(SrcAlign);
1616  }
1617  }
1618  }
1619 
1620  // If we have a cold call site, try to sink addressing computation into the
1621  // cold block. This interacts with our handling for loads and stores to
1622  // ensure that we can fold all uses of a potential addressing computation
1623  // into their uses. TODO: generalize this to work over profiling data
1624  if (!OptSize && CI->hasFnAttr(Attribute::Cold))
1625  for (auto &Arg : CI->arg_operands()) {
1626  if (!Arg->getType()->isPointerTy())
1627  continue;
1628  unsigned AS = Arg->getType()->getPointerAddressSpace();
1629  return optimizeMemoryInst(CI, Arg, Arg->getType(), AS);
1630  }
1631 
1633  if (II) {
1634  switch (II->getIntrinsicID()) {
1635  default: break;
1636  case Intrinsic::objectsize: {
1637  // Lower all uses of llvm.objectsize.*
1638  ConstantInt *RetVal =
1639  lowerObjectSizeCall(II, *DL, TLInfo, /*MustSucceed=*/true);
1640  // Substituting this can cause recursive simplifications, which can
1641  // invalidate our iterator. Use a WeakTrackingVH to hold onto it in case
1642  // this
1643  // happens.
1644  Value *CurValue = &*CurInstIterator;
1645  WeakTrackingVH IterHandle(CurValue);
1646 
1647  replaceAndRecursivelySimplify(CI, RetVal, TLInfo, nullptr);
1648 
1649  // If the iterator instruction was recursively deleted, start over at the
1650  // start of the block.
1651  if (IterHandle != CurValue) {
1652  CurInstIterator = BB->begin();
1653  SunkAddrs.clear();
1654  }
1655  return true;
1656  }
1657  case Intrinsic::aarch64_stlxr:
1658  case Intrinsic::aarch64_stxr: {
1659  ZExtInst *ExtVal = dyn_cast<ZExtInst>(CI->getArgOperand(0));
1660  if (!ExtVal || !ExtVal->hasOneUse() ||
1661  ExtVal->getParent() == CI->getParent())
1662  return false;
1663  // Sink a zext feeding stlxr/stxr before it, so it can be folded into it.
1664  ExtVal->moveBefore(CI);
1665  // Mark this instruction as "inserted by CGP", so that other
1666  // optimizations don't touch it.
1667  InsertedInsts.insert(ExtVal);
1668  return true;
1669  }
1670  case Intrinsic::invariant_group_barrier:
1671  II->replaceAllUsesWith(II->getArgOperand(0));
1672  II->eraseFromParent();
1673  return true;
1674 
1675  case Intrinsic::cttz:
1676  case Intrinsic::ctlz:
1677  // If counting zeros is expensive, try to avoid it.
1678  return despeculateCountZeros(II, TLI, DL, ModifiedDT);
1679  }
1680 
1681  if (TLI) {
1682  SmallVector<Value*, 2> PtrOps;
1683  Type *AccessTy;
1684  if (TLI->getAddrModeArguments(II, PtrOps, AccessTy))
1685  while (!PtrOps.empty()) {
1686  Value *PtrVal = PtrOps.pop_back_val();
1687  unsigned AS = PtrVal->getType()->getPointerAddressSpace();
1688  if (optimizeMemoryInst(II, PtrVal, AccessTy, AS))
1689  return true;
1690  }
1691  }
1692  }
1693 
1694  // From here on out we're working with named functions.
1695  if (!CI->getCalledFunction()) return false;
1696 
1697  // Lower all default uses of _chk calls. This is very similar
1698  // to what InstCombineCalls does, but here we are only lowering calls
1699  // to fortified library functions (e.g. __memcpy_chk) that have the default
1700  // "don't know" as the objectsize. Anything else should be left alone.
1701  FortifiedLibCallSimplifier Simplifier(TLInfo, true);
1702  if (Value *V = Simplifier.optimizeCall(CI)) {
1703  CI->replaceAllUsesWith(V);
1704  CI->eraseFromParent();
1705  return true;
1706  }
1707 
1708  return false;
1709 }
1710 
1711 /// Look for opportunities to duplicate return instructions to the predecessor
1712 /// to enable tail call optimizations. The case it is currently looking for is:
1713 /// @code
1714 /// bb0:
1715 /// %tmp0 = tail call i32 @f0()
1716 /// br label %return
1717 /// bb1:
1718 /// %tmp1 = tail call i32 @f1()
1719 /// br label %return
1720 /// bb2:
1721 /// %tmp2 = tail call i32 @f2()
1722 /// br label %return
1723 /// return:
1724 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
1725 /// ret i32 %retval
1726 /// @endcode
1727 ///
1728 /// =>
1729 ///
1730 /// @code
1731 /// bb0:
1732 /// %tmp0 = tail call i32 @f0()
1733 /// ret i32 %tmp0
1734 /// bb1:
1735 /// %tmp1 = tail call i32 @f1()
1736 /// ret i32 %tmp1
1737 /// bb2:
1738 /// %tmp2 = tail call i32 @f2()
1739 /// ret i32 %tmp2
1740 /// @endcode
1741 bool CodeGenPrepare::dupRetToEnableTailCallOpts(BasicBlock *BB) {
1742  if (!TLI)
1743  return false;
1744 
1745  ReturnInst *RetI = dyn_cast<ReturnInst>(BB->getTerminator());
1746  if (!RetI)
1747  return false;
1748 
1749  PHINode *PN = nullptr;
1750  BitCastInst *BCI = nullptr;
1751  Value *V = RetI->getReturnValue();
1752  if (V) {
1753  BCI = dyn_cast<BitCastInst>(V);
1754  if (BCI)
1755  V = BCI->getOperand(0);
1756 
1757  PN = dyn_cast<PHINode>(V);
1758  if (!PN)
1759  return false;
1760  }
1761 
1762  if (PN && PN->getParent() != BB)
1763  return false;
1764 
1765  // Make sure there are no instructions between the PHI and return, or that the
1766  // return is the first instruction in the block.
1767  if (PN) {
1768  BasicBlock::iterator BI = BB->begin();
1769  do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
1770  if (&*BI == BCI)
1771  // Also skip over the bitcast.
1772  ++BI;
1773  if (&*BI != RetI)
1774  return false;
1775  } else {
1776  BasicBlock::iterator BI = BB->begin();
1777  while (isa<DbgInfoIntrinsic>(BI)) ++BI;
1778  if (&*BI != RetI)
1779  return false;
1780  }
1781 
1782  /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
1783  /// call.
1784  const Function *F = BB->getParent();
1785  SmallVector<CallInst*, 4> TailCalls;
1786  if (PN) {
1787  for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
1788  CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
1789  // Make sure the phi value is indeed produced by the tail call.
1790  if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
1791  TLI->mayBeEmittedAsTailCall(CI) &&
1792  attributesPermitTailCall(F, CI, RetI, *TLI))
1793  TailCalls.push_back(CI);
1794  }
1795  } else {
1796  SmallPtrSet<BasicBlock*, 4> VisitedBBs;
1797  for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
1798  if (!VisitedBBs.insert(*PI).second)
1799  continue;
1800 
1801  BasicBlock::InstListType &InstList = (*PI)->getInstList();
1802  BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
1803  BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
1804  do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
1805  if (RI == RE)
1806  continue;
1807 
1808  CallInst *CI = dyn_cast<CallInst>(&*RI);
1809  if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI) &&
1810  attributesPermitTailCall(F, CI, RetI, *TLI))
1811  TailCalls.push_back(CI);
1812  }
1813  }
1814 
1815  bool Changed = false;
1816  for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
1817  CallInst *CI = TailCalls[i];
1818  CallSite CS(CI);
1819 
1820  // Conservatively require the attributes of the call to match those of the
1821  // return. Ignore noalias because it doesn't affect the call sequence.
1822  AttributeList CalleeAttrs = CS.getAttributes();
1823  if (AttrBuilder(CalleeAttrs, AttributeList::ReturnIndex)
1824  .removeAttribute(Attribute::NoAlias) !=
1826  .removeAttribute(Attribute::NoAlias))
1827  continue;
1828 
1829  // Make sure the call instruction is followed by an unconditional branch to
1830  // the return block.
1831  BasicBlock *CallBB = CI->getParent();
1832  BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
1833  if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
1834  continue;
1835 
1836  // Duplicate the return into CallBB.
1837  (void)FoldReturnIntoUncondBranch(RetI, BB, CallBB);
1838  ModifiedDT = Changed = true;
1839  ++NumRetsDup;
1840  }
1841 
1842  // If we eliminated all predecessors of the block, delete the block now.
1843  if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
1844  BB->eraseFromParent();
1845 
1846  return Changed;
1847 }
1848 
1849 //===----------------------------------------------------------------------===//
1850 // Memory Optimization
1851 //===----------------------------------------------------------------------===//
1852 
1853 namespace {
1854 
1855 /// This is an extended version of TargetLowering::AddrMode
1856 /// which holds actual Value*'s for register values.
1857 struct ExtAddrMode : public TargetLowering::AddrMode {
1858  Value *BaseReg = nullptr;
1859  Value *ScaledReg = nullptr;
1860  Value *OriginalValue = nullptr;
1861 
1862  enum FieldName {
1863  NoField = 0x00,
1864  BaseRegField = 0x01,
1865  BaseGVField = 0x02,
1866  BaseOffsField = 0x04,
1867  ScaledRegField = 0x08,
1868  ScaleField = 0x10,
1869  MultipleFields = 0xff
1870  };
1871 
1872  ExtAddrMode() = default;
1873 
1874  void print(raw_ostream &OS) const;
1875  void dump() const;
1876 
1877  FieldName compare(const ExtAddrMode &other) {
1878  // First check that the types are the same on each field, as differing types
1879  // is something we can't cope with later on.
1880  if (BaseReg && other.BaseReg &&
1881  BaseReg->getType() != other.BaseReg->getType())
1882  return MultipleFields;
1883  if (BaseGV && other.BaseGV &&
1884  BaseGV->getType() != other.BaseGV->getType())
1885  return MultipleFields;
1886  if (ScaledReg && other.ScaledReg &&
1887  ScaledReg->getType() != other.ScaledReg->getType())
1888  return MultipleFields;
1889 
1890  // Check each field to see if it differs.
1891  unsigned Result = NoField;
1892  if (BaseReg != other.BaseReg)
1893  Result |= BaseRegField;
1894  if (BaseGV != other.BaseGV)
1895  Result |= BaseGVField;
1896  if (BaseOffs != other.BaseOffs)
1897  Result |= BaseOffsField;
1898  if (ScaledReg != other.ScaledReg)
1899  Result |= ScaledRegField;
1900  // Don't count 0 as being a different scale, because that actually means
1901  // unscaled (which will already be counted by having no ScaledReg).
1902  if (Scale && other.Scale && Scale != other.Scale)
1903  Result |= ScaleField;
1904 
1905  if (countPopulation(Result) > 1)
1906  return MultipleFields;
1907  else
1908  return static_cast<FieldName>(Result);
1909  }
1910 
1911  // An AddrMode is trivial if it involves no calculation i.e. it is just a base
1912  // with no offset.
1913  bool isTrivial() {
1914  // An AddrMode is (BaseGV + BaseReg + BaseOffs + ScaleReg * Scale) so it is
1915  // trivial if at most one of these terms is nonzero, except that BaseGV and
1916  // BaseReg both being zero actually means a null pointer value, which we
1917  // consider to be 'non-zero' here.
1918  return !BaseOffs && !Scale && !(BaseGV && BaseReg);
1919  }
1920 
1921  Value *GetFieldAsValue(FieldName Field, Type *IntPtrTy) {
1922  switch (Field) {
1923  default:
1924  return nullptr;
1925  case BaseRegField:
1926  return BaseReg;
1927  case BaseGVField:
1928  return BaseGV;
1929  case ScaledRegField:
1930  return ScaledReg;
1931  case BaseOffsField:
1932  return ConstantInt::get(IntPtrTy, BaseOffs);
1933  }
1934  }
1935 
1936  void SetCombinedField(FieldName Field, Value *V,
1937  const SmallVectorImpl<ExtAddrMode> &AddrModes) {
1938  switch (Field) {
1939  default:
1940  llvm_unreachable("Unhandled fields are expected to be rejected earlier");
1941  break;
1942  case ExtAddrMode::BaseRegField:
1943  BaseReg = V;
1944  break;
1945  case ExtAddrMode::BaseGVField:
1946  // A combined BaseGV is an Instruction, not a GlobalValue, so it goes
1947  // in the BaseReg field.
1948  assert(BaseReg == nullptr);
1949  BaseReg = V;
1950  BaseGV = nullptr;
1951  break;
1952  case ExtAddrMode::ScaledRegField:
1953  ScaledReg = V;
1954  // If we have a mix of scaled and unscaled addrmodes then we want scale
1955  // to be the scale and not zero.
1956  if (!Scale)
1957  for (const ExtAddrMode &AM : AddrModes)
1958  if (AM.Scale) {
1959  Scale = AM.Scale;
1960  break;
1961  }
1962  break;
1963  case ExtAddrMode::BaseOffsField:
1964  // The offset is no longer a constant, so it goes in ScaledReg with a
1965  // scale of 1.
1966  assert(ScaledReg == nullptr);
1967  ScaledReg = V;
1968  Scale = 1;
1969  BaseOffs = 0;
1970  break;
1971  }
1972  }
1973 };
1974 
1975 } // end anonymous namespace
1976 
1977 #ifndef NDEBUG
1978 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
1979  AM.print(OS);
1980  return OS;
1981 }
1982 #endif
1983 
1984 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1985 void ExtAddrMode::print(raw_ostream &OS) const {
1986  bool NeedPlus = false;
1987  OS << "[";
1988  if (BaseGV) {
1989  OS << (NeedPlus ? " + " : "")
1990  << "GV:";
1991  BaseGV->printAsOperand(OS, /*PrintType=*/false);
1992  NeedPlus = true;
1993  }
1994 
1995  if (BaseOffs) {
1996  OS << (NeedPlus ? " + " : "")
1997  << BaseOffs;
1998  NeedPlus = true;
1999  }
2000 
2001  if (BaseReg) {
2002  OS << (NeedPlus ? " + " : "")
2003  << "Base:";
2004  BaseReg->printAsOperand(OS, /*PrintType=*/false);
2005  NeedPlus = true;
2006  }
2007  if (Scale) {
2008  OS << (NeedPlus ? " + " : "")
2009  << Scale << "*";
2010  ScaledReg->printAsOperand(OS, /*PrintType=*/false);
2011  }
2012 
2013  OS << ']';
2014 }
2015 
2016 LLVM_DUMP_METHOD void ExtAddrMode::dump() const {
2017  print(dbgs());
2018  dbgs() << '\n';
2019 }
2020 #endif
2021 
2022 namespace {
2023 
2024 /// \brief This class provides transaction based operation on the IR.
2025 /// Every change made through this class is recorded in the internal state and
2026 /// can be undone (rollback) until commit is called.
2027 class TypePromotionTransaction {
2028  /// \brief This represents the common interface of the individual transaction.
2029  /// Each class implements the logic for doing one specific modification on
2030  /// the IR via the TypePromotionTransaction.
2031  class TypePromotionAction {
2032  protected:
2033  /// The Instruction modified.
2034  Instruction *Inst;
2035 
2036  public:
2037  /// \brief Constructor of the action.
2038  /// The constructor performs the related action on the IR.
2039  TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
2040 
2041  virtual ~TypePromotionAction() = default;
2042 
2043  /// \brief Undo the modification done by this action.
2044  /// When this method is called, the IR must be in the same state as it was
2045  /// before this action was applied.
2046  /// \pre Undoing the action works if and only if the IR is in the exact same
2047  /// state as it was directly after this action was applied.
2048  virtual void undo() = 0;
2049 
2050  /// \brief Advocate every change made by this action.
2051  /// When the results on the IR of the action are to be kept, it is important
2052  /// to call this function, otherwise hidden information may be kept forever.
2053  virtual void commit() {
2054  // Nothing to be done, this action is not doing anything.
2055  }
2056  };
2057 
2058  /// \brief Utility to remember the position of an instruction.
2059  class InsertionHandler {
2060  /// Position of an instruction.
2061  /// Either an instruction:
2062  /// - Is the first in a basic block: BB is used.
2063  /// - Has a previous instructon: PrevInst is used.
2064  union {
2065  Instruction *PrevInst;
2066  BasicBlock *BB;
2067  } Point;
2068 
2069  /// Remember whether or not the instruction had a previous instruction.
2070  bool HasPrevInstruction;
2071 
2072  public:
2073  /// \brief Record the position of \p Inst.
2074  InsertionHandler(Instruction *Inst) {
2075  BasicBlock::iterator It = Inst->getIterator();
2076  HasPrevInstruction = (It != (Inst->getParent()->begin()));
2077  if (HasPrevInstruction)
2078  Point.PrevInst = &*--It;
2079  else
2080  Point.BB = Inst->getParent();
2081  }
2082 
2083  /// \brief Insert \p Inst at the recorded position.
2084  void insert(Instruction *Inst) {
2085  if (HasPrevInstruction) {
2086  if (Inst->getParent())
2087  Inst->removeFromParent();
2088  Inst->insertAfter(Point.PrevInst);
2089  } else {
2090  Instruction *Position = &*Point.BB->getFirstInsertionPt();
2091  if (Inst->getParent())
2092  Inst->moveBefore(Position);
2093  else
2094  Inst->insertBefore(Position);
2095  }
2096  }
2097  };
2098 
2099  /// \brief Move an instruction before another.
2100  class InstructionMoveBefore : public TypePromotionAction {
2101  /// Original position of the instruction.
2102  InsertionHandler Position;
2103 
2104  public:
2105  /// \brief Move \p Inst before \p Before.
2106  InstructionMoveBefore(Instruction *Inst, Instruction *Before)
2107  : TypePromotionAction(Inst), Position(Inst) {
2108  DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before << "\n");
2109  Inst->moveBefore(Before);
2110  }
2111 
2112  /// \brief Move the instruction back to its original position.
2113  void undo() override {
2114  DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
2115  Position.insert(Inst);
2116  }
2117  };
2118 
2119  /// \brief Set the operand of an instruction with a new value.
2120  class OperandSetter : public TypePromotionAction {
2121  /// Original operand of the instruction.
2122  Value *Origin;
2123 
2124  /// Index of the modified instruction.
2125  unsigned Idx;
2126 
2127  public:
2128  /// \brief Set \p Idx operand of \p Inst with \p NewVal.
2129  OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
2130  : TypePromotionAction(Inst), Idx(Idx) {
2131  DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
2132  << "for:" << *Inst << "\n"
2133  << "with:" << *NewVal << "\n");
2134  Origin = Inst->getOperand(Idx);
2135  Inst->setOperand(Idx, NewVal);
2136  }
2137 
2138  /// \brief Restore the original value of the instruction.
2139  void undo() override {
2140  DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
2141  << "for: " << *Inst << "\n"
2142  << "with: " << *Origin << "\n");
2143  Inst->setOperand(Idx, Origin);
2144  }
2145  };
2146 
2147  /// \brief Hide the operands of an instruction.
2148  /// Do as if this instruction was not using any of its operands.
2149  class OperandsHider : public TypePromotionAction {
2150  /// The list of original operands.
2151  SmallVector<Value *, 4> OriginalValues;
2152 
2153  public:
2154  /// \brief Remove \p Inst from the uses of the operands of \p Inst.
2155  OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
2156  DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
2157  unsigned NumOpnds = Inst->getNumOperands();
2158  OriginalValues.reserve(NumOpnds);
2159  for (unsigned It = 0; It < NumOpnds; ++It) {
2160  // Save the current operand.
2161  Value *Val = Inst->getOperand(It);
2162  OriginalValues.push_back(Val);
2163  // Set a dummy one.
2164  // We could use OperandSetter here, but that would imply an overhead
2165  // that we are not willing to pay.
2166  Inst->setOperand(It, UndefValue::get(Val->getType()));
2167  }
2168  }
2169 
2170  /// \brief Restore the original list of uses.
2171  void undo() override {
2172  DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
2173  for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
2174  Inst->setOperand(It, OriginalValues[It]);
2175  }
2176  };
2177 
2178  /// \brief Build a truncate instruction.
2179  class TruncBuilder : public TypePromotionAction {
2180  Value *Val;
2181 
2182  public:
2183  /// \brief Build a truncate instruction of \p Opnd producing a \p Ty
2184  /// result.
2185  /// trunc Opnd to Ty.
2186  TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
2187  IRBuilder<> Builder(Opnd);
2188  Val = Builder.CreateTrunc(Opnd, Ty, "promoted");
2189  DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n");
2190  }
2191 
2192  /// \brief Get the built value.
2193  Value *getBuiltValue() { return Val; }
2194 
2195  /// \brief Remove the built instruction.
2196  void undo() override {
2197  DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n");
2198  if (Instruction *IVal = dyn_cast<Instruction>(Val))
2199  IVal->eraseFromParent();
2200  }
2201  };
2202 
2203  /// \brief Build a sign extension instruction.
2204  class SExtBuilder : public TypePromotionAction {
2205  Value *Val;
2206 
2207  public:
2208  /// \brief Build a sign extension instruction of \p Opnd producing a \p Ty
2209  /// result.
2210  /// sext Opnd to Ty.
2211  SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
2212  : TypePromotionAction(InsertPt) {
2213  IRBuilder<> Builder(InsertPt);
2214  Val = Builder.CreateSExt(Opnd, Ty, "promoted");
2215  DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n");
2216  }
2217 
2218  /// \brief Get the built value.
2219  Value *getBuiltValue() { return Val; }
2220 
2221  /// \brief Remove the built instruction.
2222  void undo() override {
2223  DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n");
2224  if (Instruction *IVal = dyn_cast<Instruction>(Val))
2225  IVal->eraseFromParent();
2226  }
2227  };
2228 
2229  /// \brief Build a zero extension instruction.
2230  class ZExtBuilder : public TypePromotionAction {
2231  Value *Val;
2232 
2233  public:
2234  /// \brief Build a zero extension instruction of \p Opnd producing a \p Ty
2235  /// result.
2236  /// zext Opnd to Ty.
2237  ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
2238  : TypePromotionAction(InsertPt) {
2239  IRBuilder<> Builder(InsertPt);
2240  Val = Builder.CreateZExt(Opnd, Ty, "promoted");
2241  DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n");
2242  }
2243 
2244  /// \brief Get the built value.
2245  Value *getBuiltValue() { return Val; }
2246 
2247  /// \brief Remove the built instruction.
2248  void undo() override {
2249  DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n");
2250  if (Instruction *IVal = dyn_cast<Instruction>(Val))
2251  IVal->eraseFromParent();
2252  }
2253  };
2254 
2255  /// \brief Mutate an instruction to another type.
2256  class TypeMutator : public TypePromotionAction {
2257  /// Record the original type.
2258  Type *OrigTy;
2259 
2260  public:
2261  /// \brief Mutate the type of \p Inst into \p NewTy.
2262  TypeMutator(Instruction *Inst, Type *NewTy)
2263  : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
2264  DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
2265  << "\n");
2266  Inst->mutateType(NewTy);
2267  }
2268 
2269  /// \brief Mutate the instruction back to its original type.
2270  void undo() override {
2271  DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
2272  << "\n");
2273  Inst->mutateType(OrigTy);
2274  }
2275  };
2276 
2277  /// \brief Replace the uses of an instruction by another instruction.
2278  class UsesReplacer : public TypePromotionAction {
2279  /// Helper structure to keep track of the replaced uses.
2280  struct InstructionAndIdx {
2281  /// The instruction using the instruction.
2282  Instruction *Inst;
2283 
2284  /// The index where this instruction is used for Inst.
2285  unsigned Idx;
2286 
2287  InstructionAndIdx(Instruction *Inst, unsigned Idx)
2288  : Inst(Inst), Idx(Idx) {}
2289  };
2290 
2291  /// Keep track of the original uses (pair Instruction, Index).
2292  SmallVector<InstructionAndIdx, 4> OriginalUses;
2293 
2294  using use_iterator = SmallVectorImpl<InstructionAndIdx>::iterator;
2295 
2296  public:
2297  /// \brief Replace all the use of \p Inst by \p New.
2298  UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
2299  DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
2300  << "\n");
2301  // Record the original uses.
2302  for (Use &U : Inst->uses()) {
2303  Instruction *UserI = cast<Instruction>(U.getUser());
2304  OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
2305  }
2306  // Now, we can replace the uses.
2307  Inst->replaceAllUsesWith(New);
2308  }
2309 
2310  /// \brief Reassign the original uses of Inst to Inst.
2311  void undo() override {
2312  DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
2313  for (use_iterator UseIt = OriginalUses.begin(),
2314  EndIt = OriginalUses.end();
2315  UseIt != EndIt; ++UseIt) {
2316  UseIt->Inst->setOperand(UseIt->Idx, Inst);
2317  }
2318  }
2319  };
2320 
2321  /// \brief Remove an instruction from the IR.
2322  class InstructionRemover : public TypePromotionAction {
2323  /// Original position of the instruction.
2324  InsertionHandler Inserter;
2325 
2326  /// Helper structure to hide all the link to the instruction. In other
2327  /// words, this helps to do as if the instruction was removed.
2328  OperandsHider Hider;
2329 
2330  /// Keep track of the uses replaced, if any.
2331  UsesReplacer *Replacer = nullptr;
2332 
2333  /// Keep track of instructions removed.
2334  SetOfInstrs &RemovedInsts;
2335 
2336  public:
2337  /// \brief Remove all reference of \p Inst and optinally replace all its
2338  /// uses with New.
2339  /// \p RemovedInsts Keep track of the instructions removed by this Action.
2340  /// \pre If !Inst->use_empty(), then New != nullptr
2341  InstructionRemover(Instruction *Inst, SetOfInstrs &RemovedInsts,
2342  Value *New = nullptr)
2343  : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
2344  RemovedInsts(RemovedInsts) {
2345  if (New)
2346  Replacer = new UsesReplacer(Inst, New);
2347  DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
2348  RemovedInsts.insert(Inst);
2349  /// The instructions removed here will be freed after completing
2350  /// optimizeBlock() for all blocks as we need to keep track of the
2351  /// removed instructions during promotion.
2352  Inst->removeFromParent();
2353  }
2354 
2355  ~InstructionRemover() override { delete Replacer; }
2356 
2357  /// \brief Resurrect the instruction and reassign it to the proper uses if
2358  /// new value was provided when build this action.
2359  void undo() override {
2360  DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
2361  Inserter.insert(Inst);
2362  if (Replacer)
2363  Replacer->undo();
2364  Hider.undo();
2365  RemovedInsts.erase(Inst);
2366  }
2367  };
2368 
2369 public:
2370  /// Restoration point.
2371  /// The restoration point is a pointer to an action instead of an iterator
2372  /// because the iterator may be invalidated but not the pointer.
2373  using ConstRestorationPt = const TypePromotionAction *;
2374 
2375  TypePromotionTransaction(SetOfInstrs &RemovedInsts)
2376  : RemovedInsts(RemovedInsts) {}
2377 
2378  /// Advocate every changes made in that transaction.
2379  void commit();
2380 
2381  /// Undo all the changes made after the given point.
2382  void rollback(ConstRestorationPt Point);
2383 
2384  /// Get the current restoration point.
2385  ConstRestorationPt getRestorationPoint() const;
2386 
2387  /// \name API for IR modification with state keeping to support rollback.
2388  /// @{
2389  /// Same as Instruction::setOperand.
2390  void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
2391 
2392  /// Same as Instruction::eraseFromParent.
2393  void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
2394 
2395  /// Same as Value::replaceAllUsesWith.
2396  void replaceAllUsesWith(Instruction *Inst, Value *New);
2397 
2398  /// Same as Value::mutateType.
2399  void mutateType(Instruction *Inst, Type *NewTy);
2400 
2401  /// Same as IRBuilder::createTrunc.
2402  Value *createTrunc(Instruction *Opnd, Type *Ty);
2403 
2404  /// Same as IRBuilder::createSExt.
2405  Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
2406 
2407  /// Same as IRBuilder::createZExt.
2408  Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
2409 
2410  /// Same as Instruction::moveBefore.
2411  void moveBefore(Instruction *Inst, Instruction *Before);
2412  /// @}
2413 
2414 private:
2415  /// The ordered list of actions made so far.
2417 
2418  using CommitPt = SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator;
2419 
2420  SetOfInstrs &RemovedInsts;
2421 };
2422 
2423 } // end anonymous namespace
2424 
2425 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
2426  Value *NewVal) {
2427  Actions.push_back(llvm::make_unique<TypePromotionTransaction::OperandSetter>(
2428  Inst, Idx, NewVal));
2429 }
2430 
2431 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
2432  Value *NewVal) {
2433  Actions.push_back(
2434  llvm::make_unique<TypePromotionTransaction::InstructionRemover>(
2435  Inst, RemovedInsts, NewVal));
2436 }
2437 
2438 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
2439  Value *New) {
2440  Actions.push_back(
2441  llvm::make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
2442 }
2443 
2444 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
2445  Actions.push_back(
2446  llvm::make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
2447 }
2448 
2449 Value *TypePromotionTransaction::createTrunc(Instruction *Opnd,
2450  Type *Ty) {
2451  std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
2452  Value *Val = Ptr->getBuiltValue();
2453  Actions.push_back(std::move(Ptr));
2454  return Val;
2455 }
2456 
2457 Value *TypePromotionTransaction::createSExt(Instruction *Inst,
2458  Value *Opnd, Type *Ty) {
2459  std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
2460  Value *Val = Ptr->getBuiltValue();
2461  Actions.push_back(std::move(Ptr));
2462  return Val;
2463 }
2464 
2465 Value *TypePromotionTransaction::createZExt(Instruction *Inst,
2466  Value *Opnd, Type *Ty) {
2467  std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
2468  Value *Val = Ptr->getBuiltValue();
2469  Actions.push_back(std::move(Ptr));
2470  return Val;
2471 }
2472 
2473 void TypePromotionTransaction::moveBefore(Instruction *Inst,
2474  Instruction *Before) {
2475  Actions.push_back(
2476  llvm::make_unique<TypePromotionTransaction::InstructionMoveBefore>(
2477  Inst, Before));
2478 }
2479 
2480 TypePromotionTransaction::ConstRestorationPt
2481 TypePromotionTransaction::getRestorationPoint() const {
2482  return !Actions.empty() ? Actions.back().get() : nullptr;
2483 }
2484 
2485 void TypePromotionTransaction::commit() {
2486  for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
2487  ++It)
2488  (*It)->commit();
2489  Actions.clear();
2490 }
2491 
2492 void TypePromotionTransaction::rollback(
2493  TypePromotionTransaction::ConstRestorationPt Point) {
2494  while (!Actions.empty() && Point != Actions.back().get()) {
2495  std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
2496  Curr->undo();
2497  }
2498 }
2499 
2500 namespace {
2501 
2502 /// \brief A helper class for matching addressing modes.
2503 ///
2504 /// This encapsulates the logic for matching the target-legal addressing modes.
2505 class AddressingModeMatcher {
2506  SmallVectorImpl<Instruction*> &AddrModeInsts;
2507  const TargetLowering &TLI;
2508  const TargetRegisterInfo &TRI;
2509  const DataLayout &DL;
2510 
2511  /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
2512  /// the memory instruction that we're computing this address for.
2513  Type *AccessTy;
2514  unsigned AddrSpace;
2515  Instruction *MemoryInst;
2516 
2517  /// This is the addressing mode that we're building up. This is
2518  /// part of the return value of this addressing mode matching stuff.
2519  ExtAddrMode &AddrMode;
2520 
2521  /// The instructions inserted by other CodeGenPrepare optimizations.
2522  const SetOfInstrs &InsertedInsts;
2523 
2524  /// A map from the instructions to their type before promotion.
2525  InstrToOrigTy &PromotedInsts;
2526 
2527  /// The ongoing transaction where every action should be registered.
2528  TypePromotionTransaction &TPT;
2529 
2530  /// This is set to true when we should not do profitability checks.
2531  /// When true, IsProfitableToFoldIntoAddressingMode always returns true.
2532  bool IgnoreProfitability;
2533 
2534  AddressingModeMatcher(SmallVectorImpl<Instruction *> &AMI,
2535  const TargetLowering &TLI,
2536  const TargetRegisterInfo &TRI,
2537  Type *AT, unsigned AS,
2538  Instruction *MI, ExtAddrMode &AM,
2539  const SetOfInstrs &InsertedInsts,
2540  InstrToOrigTy &PromotedInsts,
2541  TypePromotionTransaction &TPT)
2542  : AddrModeInsts(AMI), TLI(TLI), TRI(TRI),
2543  DL(MI->getModule()->getDataLayout()), AccessTy(AT), AddrSpace(AS),
2544  MemoryInst(MI), AddrMode(AM), InsertedInsts(InsertedInsts),
2545  PromotedInsts(PromotedInsts), TPT(TPT) {
2546  IgnoreProfitability = false;
2547  }
2548 
2549 public:
2550  /// Find the maximal addressing mode that a load/store of V can fold,
2551  /// give an access type of AccessTy. This returns a list of involved
2552  /// instructions in AddrModeInsts.
2553  /// \p InsertedInsts The instructions inserted by other CodeGenPrepare
2554  /// optimizations.
2555  /// \p PromotedInsts maps the instructions to their type before promotion.
2556  /// \p The ongoing transaction where every action should be registered.
2557  static ExtAddrMode Match(Value *V, Type *AccessTy, unsigned AS,
2558  Instruction *MemoryInst,
2559  SmallVectorImpl<Instruction*> &AddrModeInsts,
2560  const TargetLowering &TLI,
2561  const TargetRegisterInfo &TRI,
2562  const SetOfInstrs &InsertedInsts,
2563  InstrToOrigTy &PromotedInsts,
2564  TypePromotionTransaction &TPT) {
2565  ExtAddrMode Result;
2566 
2567  bool Success = AddressingModeMatcher(AddrModeInsts, TLI, TRI,
2568  AccessTy, AS,
2569  MemoryInst, Result, InsertedInsts,
2570  PromotedInsts, TPT).matchAddr(V, 0);
2571  (void)Success; assert(Success && "Couldn't select *anything*?");
2572  return Result;
2573  }
2574 
2575 private:
2576  bool matchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
2577  bool matchAddr(Value *V, unsigned Depth);
2578  bool matchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth,
2579  bool *MovedAway = nullptr);
2580  bool isProfitableToFoldIntoAddressingMode(Instruction *I,
2581  ExtAddrMode &AMBefore,
2582  ExtAddrMode &AMAfter);
2583  bool valueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
2584  bool isPromotionProfitable(unsigned NewCost, unsigned OldCost,
2585  Value *PromotedOperand) const;
2586 };
2587 
2588 /// \brief Keep track of simplification of Phi nodes.
2589 /// Accept the set of all phi nodes and erase phi node from this set
2590 /// if it is simplified.
2591 class SimplificationTracker {
2593  const SimplifyQuery &SQ;
2594  // Tracks newly created Phi nodes. We use a SetVector to get deterministic
2595  // order when iterating over the set in MatchPhiSet.
2596  SmallSetVector<PHINode *, 32> AllPhiNodes;
2597  // Tracks newly created Select nodes.
2598  SmallPtrSet<SelectInst *, 32> AllSelectNodes;
2599 
2600 public:
2601  SimplificationTracker(const SimplifyQuery &sq)
2602  : SQ(sq) {}
2603 
2604  Value *Get(Value *V) {
2605  do {
2606  auto SV = Storage.find(V);
2607  if (SV == Storage.end())
2608  return V;
2609  V = SV->second;
2610  } while (true);
2611  }
2612 
2613  Value *Simplify(Value *Val) {
2614  SmallVector<Value *, 32> WorkList;
2615  SmallPtrSet<Value *, 32> Visited;
2616  WorkList.push_back(Val);
2617  while (!WorkList.empty()) {
2618  auto P = WorkList.pop_back_val();
2619  if (!Visited.insert(P).second)
2620  continue;
2621  if (auto *PI = dyn_cast<Instruction>(P))
2622  if (Value *V = SimplifyInstruction(cast<Instruction>(PI), SQ)) {
2623  for (auto *U : PI->users())
2624  WorkList.push_back(cast<Value>(U));
2625  Put(PI, V);
2626  PI->replaceAllUsesWith(V);
2627  if (auto *PHI = dyn_cast<PHINode>(PI))
2628  AllPhiNodes.remove(PHI);
2629  if (auto *Select = dyn_cast<SelectInst>(PI))
2630  AllSelectNodes.erase(Select);
2631  PI->eraseFromParent();
2632  }
2633  }
2634  return Get(Val);
2635  }
2636 
2637  void Put(Value *From, Value *To) {
2638  Storage.insert({ From, To });
2639  }
2640 
2641  void ReplacePhi(PHINode *From, PHINode *To) {
2642  Value* OldReplacement = Get(From);
2643  while (OldReplacement != From) {
2644  From = To;
2645  To = dyn_cast<PHINode>(OldReplacement);
2646  OldReplacement = Get(From);
2647  }
2648  assert(Get(To) == To && "Replacement PHI node is already replaced.");
2649  Put(From, To);
2650  From->replaceAllUsesWith(To);
2651  AllPhiNodes.remove(From);
2652  From->eraseFromParent();
2653  }
2654 
2655  SmallSetVector<PHINode *, 32>& newPhiNodes() { return AllPhiNodes; }
2656 
2657  void insertNewPhi(PHINode *PN) { AllPhiNodes.insert(PN); }
2658 
2659  void insertNewSelect(SelectInst *SI) { AllSelectNodes.insert(SI); }
2660 
2661  unsigned countNewPhiNodes() const { return AllPhiNodes.size(); }
2662 
2663  unsigned countNewSelectNodes() const { return AllSelectNodes.size(); }
2664 
2665  void destroyNewNodes(Type *CommonType) {
2666  // For safe erasing, replace the uses with dummy value first.
2667  auto Dummy = UndefValue::get(CommonType);
2668  for (auto I : AllPhiNodes) {
2670  I->eraseFromParent();
2671  }
2672  AllPhiNodes.clear();
2673  for (auto I : AllSelectNodes) {
2675  I->eraseFromParent();
2676  }
2677  AllSelectNodes.clear();
2678  }
2679 };
2680 
2681 /// \brief A helper class for combining addressing modes.
2682 class AddressingModeCombiner {
2683  typedef std::pair<Value *, BasicBlock *> ValueInBB;
2684  typedef DenseMap<ValueInBB, Value *> FoldAddrToValueMapping;
2685  typedef std::pair<PHINode *, PHINode *> PHIPair;
2686 
2687 private:
2688  /// The addressing modes we've collected.
2689  SmallVector<ExtAddrMode, 16> AddrModes;
2690 
2691  /// The field in which the AddrModes differ, when we have more than one.
2692  ExtAddrMode::FieldName DifferentField = ExtAddrMode::NoField;
2693 
2694  /// Are the AddrModes that we have all just equal to their original values?
2695  bool AllAddrModesTrivial = true;
2696 
2697  /// Common Type for all different fields in addressing modes.
2698  Type *CommonType;
2699 
2700  /// SimplifyQuery for simplifyInstruction utility.
2701  const SimplifyQuery &SQ;
2702 
2703  /// Original Address.
2704  ValueInBB Original;
2705 
2706 public:
2707  AddressingModeCombiner(const SimplifyQuery &_SQ, ValueInBB OriginalValue)
2708  : CommonType(nullptr), SQ(_SQ), Original(OriginalValue) {}
2709 
2710  /// \brief Get the combined AddrMode
2711  const ExtAddrMode &getAddrMode() const {
2712  return AddrModes[0];
2713  }
2714 
2715  /// \brief Add a new AddrMode if it's compatible with the AddrModes we already
2716  /// have.
2717  /// \return True iff we succeeded in doing so.
2718  bool addNewAddrMode(ExtAddrMode &NewAddrMode) {
2719  // Take note of if we have any non-trivial AddrModes, as we need to detect
2720  // when all AddrModes are trivial as then we would introduce a phi or select
2721  // which just duplicates what's already there.
2722  AllAddrModesTrivial = AllAddrModesTrivial && NewAddrMode.isTrivial();
2723 
2724  // If this is the first addrmode then everything is fine.
2725  if (AddrModes.empty()) {
2726  AddrModes.emplace_back(NewAddrMode);
2727  return true;
2728  }
2729 
2730  // Figure out how different this is from the other address modes, which we
2731  // can do just by comparing against the first one given that we only care
2732  // about the cumulative difference.
2733  ExtAddrMode::FieldName ThisDifferentField =
2734  AddrModes[0].compare(NewAddrMode);
2735  if (DifferentField == ExtAddrMode::NoField)
2736  DifferentField = ThisDifferentField;
2737  else if (DifferentField != ThisDifferentField)
2738  DifferentField = ExtAddrMode::MultipleFields;
2739 
2740  // If NewAddrMode differs in more than one dimension we cannot handle it.
2741  bool CanHandle = DifferentField != ExtAddrMode::MultipleFields;
2742 
2743  // If Scale Field is different then we reject.
2744  CanHandle = CanHandle && DifferentField != ExtAddrMode::ScaleField;
2745 
2746  // We also must reject the case when base offset is different and
2747  // scale reg is not null, we cannot handle this case due to merge of
2748  // different offsets will be used as ScaleReg.
2749  CanHandle = CanHandle && (DifferentField != ExtAddrMode::BaseOffsField ||
2750  !NewAddrMode.ScaledReg);
2751 
2752  // We also must reject the case when GV is different and BaseReg installed
2753  // due to we want to use base reg as a merge of GV values.
2754  CanHandle = CanHandle && (DifferentField != ExtAddrMode::BaseGVField ||
2755  !NewAddrMode.HasBaseReg);
2756 
2757  // Even if NewAddMode is the same we still need to collect it due to
2758  // original value is different. And later we will need all original values
2759  // as anchors during finding the common Phi node.
2760  if (CanHandle)
2761  AddrModes.emplace_back(NewAddrMode);
2762  else
2763  AddrModes.clear();
2764 
2765  return CanHandle;
2766  }
2767 
2768  /// \brief Combine the addressing modes we've collected into a single
2769  /// addressing mode.
2770  /// \return True iff we successfully combined them or we only had one so
2771  /// didn't need to combine them anyway.
2772  bool combineAddrModes() {
2773  // If we have no AddrModes then they can't be combined.
2774  if (AddrModes.size() == 0)
2775  return false;
2776 
2777  // A single AddrMode can trivially be combined.
2778  if (AddrModes.size() == 1 || DifferentField == ExtAddrMode::NoField)
2779  return true;
2780 
2781  // If the AddrModes we collected are all just equal to the value they are
2782  // derived from then combining them wouldn't do anything useful.
2783  if (AllAddrModesTrivial)
2784  return false;
2785 
2786  if (!addrModeCombiningAllowed())
2787  return false;
2788 
2789  // Build a map between <original value, basic block where we saw it> to
2790  // value of base register.
2791  // Bail out if there is no common type.
2792  FoldAddrToValueMapping Map;
2793  if (!initializeMap(Map))
2794  return false;
2795 
2796  Value *CommonValue = findCommon(Map);
2797  if (CommonValue)
2798  AddrModes[0].SetCombinedField(DifferentField, CommonValue, AddrModes);
2799  return CommonValue != nullptr;
2800  }
2801 
2802 private:
2803  /// \brief Initialize Map with anchor values. For address seen in some BB
2804  /// we set the value of different field saw in this address.
2805  /// If address is not an instruction than basic block is set to null.
2806  /// At the same time we find a common type for different field we will
2807  /// use to create new Phi/Select nodes. Keep it in CommonType field.
2808  /// Return false if there is no common type found.
2809  bool initializeMap(FoldAddrToValueMapping &Map) {
2810  // Keep track of keys where the value is null. We will need to replace it
2811  // with constant null when we know the common type.
2812  SmallVector<ValueInBB, 2> NullValue;
2813  Type *IntPtrTy = SQ.DL.getIntPtrType(AddrModes[0].OriginalValue->getType());
2814  for (auto &AM : AddrModes) {
2815  BasicBlock *BB = nullptr;
2816  if (Instruction *I = dyn_cast<Instruction>(AM.OriginalValue))
2817  BB = I->getParent();
2818 
2819  Value *DV = AM.GetFieldAsValue(DifferentField, IntPtrTy);
2820  if (DV) {
2821  auto *Type = DV->getType();
2822  if (CommonType && CommonType != Type)
2823  return false;
2824  CommonType = Type;
2825  Map[{ AM.OriginalValue, BB }] = DV;
2826  } else {
2827  NullValue.push_back({ AM.OriginalValue, BB });
2828  }
2829  }
2830  assert(CommonType && "At least one non-null value must be!");
2831  for (auto VIBB : NullValue)
2832  Map[VIBB] = Constant::getNullValue(CommonType);
2833  return true;
2834  }
2835 
2836  /// \brief We have mapping between value A and basic block where value A
2837  /// seen to other value B where B was a field in addressing mode represented
2838  /// by A. Also we have an original value C representin an address in some
2839  /// basic block. Traversing from C through phi and selects we ended up with
2840  /// A's in a map. This utility function tries to find a value V which is a
2841  /// field in addressing mode C and traversing through phi nodes and selects
2842  /// we will end up in corresponded values B in a map.
2843  /// The utility will create a new Phi/Selects if needed.
2844  // The simple example looks as follows:
2845  // BB1:
2846  // p1 = b1 + 40
2847  // br cond BB2, BB3
2848  // BB2:
2849  // p2 = b2 + 40
2850  // br BB3
2851  // BB3:
2852  // p = phi [p1, BB1], [p2, BB2]
2853  // v = load p
2854  // Map is
2855  // <p1, BB1> -> b1
2856  // <p2, BB2> -> b2
2857  // Request is
2858  // <p, BB3> -> ?
2859  // The function tries to find or build phi [b1, BB1], [b2, BB2] in BB3
2860  Value *findCommon(FoldAddrToValueMapping &Map) {
2861  // Tracks the simplification of newly created phi nodes. The reason we use
2862  // this mapping is because we will add new created Phi nodes in AddrToBase.
2863  // Simplification of Phi nodes is recursive, so some Phi node may
2864  // be simplified after we added it to AddrToBase.
2865  // Using this mapping we can find the current value in AddrToBase.
2866  SimplificationTracker ST(SQ);
2867 
2868  // First step, DFS to create PHI nodes for all intermediate blocks.
2869  // Also fill traverse order for the second step.
2870  SmallVector<ValueInBB, 32> TraverseOrder;
2871  InsertPlaceholders(Map, TraverseOrder, ST);
2872 
2873  // Second Step, fill new nodes by merged values and simplify if possible.
2874  FillPlaceholders(Map, TraverseOrder, ST);
2875 
2876  if (!AddrSinkNewSelects && ST.countNewSelectNodes() > 0) {
2877  ST.destroyNewNodes(CommonType);
2878  return nullptr;
2879  }
2880 
2881  // Now we'd like to match New Phi nodes to existed ones.
2882  unsigned PhiNotMatchedCount = 0;
2883  if (!MatchPhiSet(ST, AddrSinkNewPhis, PhiNotMatchedCount)) {
2884  ST.destroyNewNodes(CommonType);
2885  return nullptr;
2886  }
2887 
2888  auto *Result = ST.Get(Map.find(Original)->second);
2889  if (Result) {
2890  NumMemoryInstsPhiCreated += ST.countNewPhiNodes() + PhiNotMatchedCount;
2891  NumMemoryInstsSelectCreated += ST.countNewSelectNodes();
2892  }
2893  return Result;
2894  }
2895 
2896  /// \brief Try to match PHI node to Candidate.
2897  /// Matcher tracks the matched Phi nodes.
2898  bool MatchPhiNode(PHINode *PHI, PHINode *Candidate,
2899  SmallSetVector<PHIPair, 8> &Matcher,
2900  SmallSetVector<PHINode *, 32> &PhiNodesToMatch) {
2901  SmallVector<PHIPair, 8> WorkList;
2902  Matcher.insert({ PHI, Candidate });
2903  WorkList.push_back({ PHI, Candidate });
2904  SmallSet<PHIPair, 8> Visited;
2905  while (!WorkList.empty()) {
2906  auto Item = WorkList.pop_back_val();
2907  if (!Visited.insert(Item).second)
2908  continue;
2909  // We iterate over all incoming values to Phi to compare them.
2910  // If values are different and both of them Phi and the first one is a
2911  // Phi we added (subject to match) and both of them is in the same basic
2912  // block then we can match our pair if values match. So we state that
2913  // these values match and add it to work list to verify that.
2914  for (auto B : Item.first->blocks()) {
2915  Value *FirstValue = Item.first->getIncomingValueForBlock(B);
2916  Value *SecondValue = Item.second->getIncomingValueForBlock(B);
2917  if (FirstValue == SecondValue)
2918  continue;
2919 
2920  PHINode *FirstPhi = dyn_cast<PHINode>(FirstValue);
2921  PHINode *SecondPhi = dyn_cast<PHINode>(SecondValue);
2922 
2923  // One of them is not Phi or
2924  // The first one is not Phi node from the set we'd like to match or
2925  // Phi nodes from different basic blocks then
2926  // we will not be able to match.
2927  if (!FirstPhi || !SecondPhi || !PhiNodesToMatch.count(FirstPhi) ||
2928  FirstPhi->getParent() != SecondPhi->getParent())
2929  return false;
2930 
2931  // If we already matched them then continue.
2932  if (Matcher.count({ FirstPhi, SecondPhi }))
2933  continue;
2934  // So the values are different and does not match. So we need them to
2935  // match.
2936  Matcher.insert({ FirstPhi, SecondPhi });
2937  // But me must check it.
2938  WorkList.push_back({ FirstPhi, SecondPhi });
2939  }
2940  }
2941  return true;
2942  }
2943 
2944  /// \brief For the given set of PHI nodes (in the SimplificationTracker) try
2945  /// to find their equivalents.
2946  /// Returns false if this matching fails and creation of new Phi is disabled.
2947  bool MatchPhiSet(SimplificationTracker &ST, bool AllowNewPhiNodes,
2948  unsigned &PhiNotMatchedCount) {
2949  // Use a SetVector for Matched to make sure we do replacements (ReplacePhi)
2950  // in a deterministic order below.
2952  SmallPtrSet<PHINode *, 8> WillNotMatch;
2953  SmallSetVector<PHINode *, 32> &PhiNodesToMatch = ST.newPhiNodes();
2954  while (PhiNodesToMatch.size()) {
2955  PHINode *PHI = *PhiNodesToMatch.begin();
2956 
2957  // Add us, if no Phi nodes in the basic block we do not match.
2958  WillNotMatch.clear();
2959  WillNotMatch.insert(PHI);
2960 
2961  // Traverse all Phis until we found equivalent or fail to do that.
2962  bool IsMatched = false;
2963  for (auto &P : PHI->getParent()->phis()) {
2964  if (&P == PHI)
2965  continue;
2966  if ((IsMatched = MatchPhiNode(PHI, &P, Matched, PhiNodesToMatch)))
2967  break;
2968  // If it does not match, collect all Phi nodes from matcher.
2969  // if we end up with no match, them all these Phi nodes will not match
2970  // later.
2971  for (auto M : Matched)
2972  WillNotMatch.insert(M.first);
2973  Matched.clear();
2974  }
2975  if (IsMatched) {
2976  // Replace all matched values and erase them.
2977  for (auto MV : Matched)
2978  ST.ReplacePhi(MV.first, MV.second);
2979  Matched.clear();
2980  continue;
2981  }
2982  // If we are not allowed to create new nodes then bail out.
2983  if (!AllowNewPhiNodes)
2984  return false;
2985  // Just remove all seen values in matcher. They will not match anything.
2986  PhiNotMatchedCount += WillNotMatch.size();
2987  for (auto *P : WillNotMatch)
2988  PhiNodesToMatch.remove(P);
2989  }
2990  return true;
2991  }
2992  /// \brief Fill the placeholder with values from predecessors and simplify it.
2993  void FillPlaceholders(FoldAddrToValueMapping &Map,
2994  SmallVectorImpl<ValueInBB> &TraverseOrder,
2995  SimplificationTracker &ST) {
2996  while (!TraverseOrder.empty()) {
2997  auto Current = TraverseOrder.pop_back_val();
2998  assert(Map.find(Current) != Map.end() && "No node to fill!!!");
2999  Value *CurrentValue = Current.first;
3000  BasicBlock *CurrentBlock = Current.second;
3001  Value *V = Map[Current];
3002 
3003  if (SelectInst *Select = dyn_cast<SelectInst>(V)) {
3004  // CurrentValue also must be Select.
3005  auto *CurrentSelect = cast<SelectInst>(CurrentValue);
3006  auto *TrueValue = CurrentSelect->getTrueValue();
3007  ValueInBB TrueItem = { TrueValue, isa<Instruction>(TrueValue)
3008  ? CurrentBlock
3009  : nullptr };
3010  assert(Map.find(TrueItem) != Map.end() && "No True Value!");
3011  Select->setTrueValue(ST.Get(Map[TrueItem]));
3012  auto *FalseValue = CurrentSelect->getFalseValue();
3013  ValueInBB FalseItem = { FalseValue, isa<Instruction>(FalseValue)
3014  ? CurrentBlock
3015  : nullptr };
3016  assert(Map.find(FalseItem) != Map.end() && "No False Value!");
3017  Select->setFalseValue(ST.Get(Map[FalseItem]));
3018  } else {
3019  // Must be a Phi node then.
3020  PHINode *PHI = cast<PHINode>(V);
3021  // Fill the Phi node with values from predecessors.
3022  bool IsDefinedInThisBB =
3023  cast<Instruction>(CurrentValue)->getParent() == CurrentBlock;
3024  auto *CurrentPhi = dyn_cast<PHINode>(CurrentValue);
3025  for (auto B : predecessors(CurrentBlock)) {
3026  Value *PV = IsDefinedInThisBB
3027  ? CurrentPhi->getIncomingValueForBlock(B)
3028  : CurrentValue;
3029  ValueInBB item = { PV, isa<Instruction>(PV) ? B : nullptr };
3030  assert(Map.find(item) != Map.end() && "No predecessor Value!");
3031  PHI->addIncoming(ST.Get(Map[item]), B);
3032  }
3033  }
3034  // Simplify if possible.
3035  Map[Current] = ST.Simplify(V);
3036  }
3037  }
3038 
3039  /// Starting from value recursively iterates over predecessors up to known
3040  /// ending values represented in a map. For each traversed block inserts
3041  /// a placeholder Phi or Select.
3042  /// Reports all new created Phi/Select nodes by adding them to set.
3043  /// Also reports and order in what basic blocks have been traversed.
3044  void InsertPlaceholders(FoldAddrToValueMapping &Map,
3045  SmallVectorImpl<ValueInBB> &TraverseOrder,
3046  SimplificationTracker &ST) {
3047  SmallVector<ValueInBB, 32> Worklist;
3048  assert((isa<PHINode>(Original.first) || isa<SelectInst>(Original.first)) &&
3049  "Address must be a Phi or Select node");
3050  auto *Dummy = UndefValue::get(CommonType);
3051  Worklist.push_back(Original);
3052  while (!Worklist.empty()) {
3053  auto Current = Worklist.pop_back_val();
3054  // If value is not an instruction it is something global, constant,
3055  // parameter and we can say that this value is observable in any block.
3056  // Set block to null to denote it.
3057  // Also please take into account that it is how we build anchors.
3058  if (!isa<Instruction>(Current.first))
3059  Current.second = nullptr;
3060  // if it is already visited or it is an ending value then skip it.
3061  if (Map.find(Current) != Map.end())
3062  continue;
3063  TraverseOrder.push_back(Current);
3064 
3065  Value *CurrentValue = Current.first;
3066  BasicBlock *CurrentBlock = Current.second;
3067  // CurrentValue must be a Phi node or select. All others must be covered
3068  // by anchors.
3069  Instruction *CurrentI = cast<Instruction>(CurrentValue);
3070  bool IsDefinedInThisBB = CurrentI->getParent() == CurrentBlock;
3071 
3072  unsigned PredCount =
3073  std::distance(pred_begin(CurrentBlock), pred_end(CurrentBlock));
3074  // if Current Value is not defined in this basic block we are interested
3075  // in values in predecessors.
3076  if (!IsDefinedInThisBB) {
3077  assert(PredCount && "Unreachable block?!");
3078  PHINode *PHI = PHINode::Create(CommonType, PredCount, "sunk_phi",
3079  &CurrentBlock->front());
3080  Map[Current] = PHI;
3081  ST.insertNewPhi(PHI);
3082  // Add all predecessors in work list.
3083  for (auto B : predecessors(CurrentBlock))
3084  Worklist.push_back({ CurrentValue, B });
3085  continue;
3086  }
3087  // Value is defined in this basic block.
3088  if (SelectInst *OrigSelect = dyn_cast<SelectInst>(CurrentI)) {
3089  // Is it OK to get metadata from OrigSelect?!
3090  // Create a Select placeholder with dummy value.
3091  SelectInst *Select =
3092  SelectInst::Create(OrigSelect->getCondition(), Dummy, Dummy,
3093  OrigSelect->getName(), OrigSelect, OrigSelect);
3094  Map[Current] = Select;
3095  ST.insertNewSelect(Select);
3096  // We are interested in True and False value in this basic block.
3097  Worklist.push_back({ OrigSelect->getTrueValue(), CurrentBlock });
3098  Worklist.push_back({ OrigSelect->getFalseValue(), CurrentBlock });
3099  } else {
3100  // It must be a Phi node then.
3101  auto *CurrentPhi = cast<PHINode>(CurrentI);
3102  // Create new Phi node for merge of bases.
3103  assert(PredCount && "Unreachable block?!");
3104  PHINode *PHI = PHINode::Create(CommonType, PredCount, "sunk_phi",
3105  &CurrentBlock->front());
3106  Map[Current] = PHI;
3107  ST.insertNewPhi(PHI);
3108 
3109  // Add all predecessors in work list.
3110  for (auto B : predecessors(CurrentBlock))
3111  Worklist.push_back({ CurrentPhi->getIncomingValueForBlock(B), B });
3112  }
3113  }
3114  }
3115 
3116  bool addrModeCombiningAllowed() {
3118  return false;
3119  switch (DifferentField) {
3120  default:
3121  return false;
3122  case ExtAddrMode::BaseRegField:
3123  return AddrSinkCombineBaseReg;
3124  case ExtAddrMode::BaseGVField:
3125  return AddrSinkCombineBaseGV;
3126  case ExtAddrMode::BaseOffsField:
3127  return AddrSinkCombineBaseOffs;
3128  case ExtAddrMode::ScaledRegField:
3129  return AddrSinkCombineScaledReg;
3130  }
3131  }
3132 };
3133 } // end anonymous namespace
3134 
3135 /// Try adding ScaleReg*Scale to the current addressing mode.
3136 /// Return true and update AddrMode if this addr mode is legal for the target,
3137 /// false if not.
3138 bool AddressingModeMatcher::matchScaledValue(Value *ScaleReg, int64_t Scale,
3139  unsigned Depth) {
3140  // If Scale is 1, then this is the same as adding ScaleReg to the addressing
3141  // mode. Just process that directly.
3142  if (Scale == 1)
3143  return matchAddr(ScaleReg, Depth);
3144 
3145  // If the scale is 0, it takes nothing to add this.
3146  if (Scale == 0)
3147  return true;
3148 
3149  // If we already have a scale of this value, we can add to it, otherwise, we
3150  // need an available scale field.
3151  if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
3152  return false;
3153 
3154  ExtAddrMode TestAddrMode = AddrMode;
3155 
3156  // Add scale to turn X*4+X*3 -> X*7. This could also do things like
3157  // [A+B + A*7] -> [B+A*8].
3158  TestAddrMode.Scale += Scale;
3159  TestAddrMode.ScaledReg = ScaleReg;
3160 
3161  // If the new address isn't legal, bail out.
3162  if (!TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace))
3163  return false;
3164 
3165  // It was legal, so commit it.
3166  AddrMode = TestAddrMode;
3167 
3168  // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
3169  // to see if ScaleReg is actually X+C. If so, we can turn this into adding
3170  // X*Scale + C*Scale to addr mode.
3171  ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
3172  if (isa<Instruction>(ScaleReg) && // not a constant expr.
3173  match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
3174  TestAddrMode.ScaledReg = AddLHS;
3175  TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
3176 
3177  // If this addressing mode is legal, commit it and remember that we folded
3178  // this instruction.
3179  if (TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace)) {
3180  AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
3181  AddrMode = TestAddrMode;
3182  return true;
3183  }
3184  }
3185 
3186  // Otherwise, not (x+c)*scale, just return what we have.
3187  return true;
3188 }
3189 
3190 /// This is a little filter, which returns true if an addressing computation
3191 /// involving I might be folded into a load/store accessing it.
3192 /// This doesn't need to be perfect, but needs to accept at least
3193 /// the set of instructions that MatchOperationAddr can.
3195  switch (I->getOpcode()) {
3196  case Instruction::BitCast:
3197  case Instruction::AddrSpaceCast:
3198  // Don't touch identity bitcasts.
3199  if (I->getType() == I->getOperand(0)->getType())
3200  return false;
3201  return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
3202  case Instruction::PtrToInt:
3203  // PtrToInt is always a noop, as we know that the int type is pointer sized.
3204  return true;
3205  case Instruction::IntToPtr:
3206  // We know the input is intptr_t, so this is foldable.
3207  return true;
3208  case Instruction::Add:
3209  return true;
3210  case Instruction::Mul:
3211  case Instruction::Shl:
3212  // Can only handle X*C and X << C.
3213  return isa<ConstantInt>(I->getOperand(1));
3214  case Instruction::GetElementPtr:
3215  return true;
3216  default:
3217  return false;
3218  }
3219 }
3220 
3221 /// \brief Check whether or not \p Val is a legal instruction for \p TLI.
3222 /// \note \p Val is assumed to be the product of some type promotion.
3223 /// Therefore if \p Val has an undefined state in \p TLI, this is assumed
3224 /// to be legal, as the non-promoted value would have had the same state.
3226  const DataLayout &DL, Value *Val) {
3227  Instruction *PromotedInst = dyn_cast<Instruction>(Val);
3228  if (!PromotedInst)
3229  return false;
3230  int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
3231  // If the ISDOpcode is undefined, it was undefined before the promotion.
3232  if (!ISDOpcode)
3233  return true;
3234  // Otherwise, check if the promoted instruction is legal or not.
3235  return TLI.isOperationLegalOrCustom(
3236  ISDOpcode, TLI.getValueType(DL, PromotedInst->getType()));
3237 }
3238 
3239 namespace {
3240 
3241 /// \brief Hepler class to perform type promotion.
3242 class TypePromotionHelper {
3243  /// \brief Utility function to check whether or not a sign or zero extension
3244  /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
3245  /// either using the operands of \p Inst or promoting \p Inst.
3246  /// The type of the extension is defined by \p IsSExt.
3247  /// In other words, check if:
3248  /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
3249  /// #1 Promotion applies:
3250  /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
3251  /// #2 Operand reuses:
3252  /// ext opnd1 to ConsideredExtType.
3253  /// \p PromotedInsts maps the instructions to their type before promotion.
3254  static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
3255  const InstrToOrigTy &PromotedInsts, bool IsSExt);
3256 
3257  /// \brief Utility function to determine if \p OpIdx should be promoted when
3258  /// promoting \p Inst.
3259  static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
3260  return !(isa<SelectInst>(Inst) && OpIdx == 0);
3261  }
3262 
3263  /// \brief Utility function to promote the operand of \p Ext when this
3264  /// operand is a promotable trunc or sext or zext.
3265  /// \p PromotedInsts maps the instructions to their type before promotion.
3266  /// \p CreatedInstsCost[out] contains the cost of all instructions
3267  /// created to promote the operand of Ext.
3268  /// Newly added extensions are inserted in \p Exts.
3269  /// Newly added truncates are inserted in \p Truncs.
3270  /// Should never be called directly.
3271  /// \return The promoted value which is used instead of Ext.
3272  static Value *promoteOperandForTruncAndAnyExt(
3273  Instruction *Ext, TypePromotionTransaction &TPT,
3274  InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3276  SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI);
3277 
3278  /// \brief Utility function to promote the operand of \p Ext when this
3279  /// operand is promotable and is not a supported trunc or sext.
3280  /// \p PromotedInsts maps the instructions to their type before promotion.
3281  /// \p CreatedInstsCost[out] contains the cost of all the instructions
3282  /// created to promote the operand of Ext.
3283  /// Newly added extensions are inserted in \p Exts.
3284  /// Newly added truncates are inserted in \p Truncs.
3285  /// Should never be called directly.
3286  /// \return The promoted value which is used instead of Ext.
3287  static Value *promoteOperandForOther(Instruction *Ext,
3288  TypePromotionTransaction &TPT,
3289  InstrToOrigTy &PromotedInsts,
3290  unsigned &CreatedInstsCost,
3293  const TargetLowering &TLI, bool IsSExt);
3294 
3295  /// \see promoteOperandForOther.
3296  static Value *signExtendOperandForOther(
3297  Instruction *Ext, TypePromotionTransaction &TPT,
3298  InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3300  SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3301  return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
3302  Exts, Truncs, TLI, true);
3303  }
3304 
3305  /// \see promoteOperandForOther.
3306  static Value *zeroExtendOperandForOther(
3307  Instruction *Ext, TypePromotionTransaction &TPT,
3308  InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3310  SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3311  return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
3312  Exts, Truncs, TLI, false);
3313  }
3314 
3315 public:
3316  /// Type for the utility function that promotes the operand of Ext.
3317  using Action = Value *(*)(Instruction *Ext, TypePromotionTransaction &TPT,
3318  InstrToOrigTy &PromotedInsts,
3319  unsigned &CreatedInstsCost,
3322  const TargetLowering &TLI);
3323 
3324  /// \brief Given a sign/zero extend instruction \p Ext, return the approriate
3325  /// action to promote the operand of \p Ext instead of using Ext.
3326  /// \return NULL if no promotable action is possible with the current
3327  /// sign extension.
3328  /// \p InsertedInsts keeps track of all the instructions inserted by the
3329  /// other CodeGenPrepare optimizations. This information is important
3330  /// because we do not want to promote these instructions as CodeGenPrepare
3331  /// will reinsert them later. Thus creating an infinite loop: create/remove.
3332  /// \p PromotedInsts maps the instructions to their type before promotion.
3333  static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedInsts,
3334  const TargetLowering &TLI,
3335  const InstrToOrigTy &PromotedInsts);
3336 };
3337 
3338 } // end anonymous namespace
3339 
3340 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
3341  Type *ConsideredExtType,
3342  const InstrToOrigTy &PromotedInsts,
3343  bool IsSExt) {
3344  // The promotion helper does not know how to deal with vector types yet.
3345  // To be able to fix that, we would need to fix the places where we
3346  // statically extend, e.g., constants and such.
3347  if (Inst->getType()->isVectorTy())
3348  return false;
3349 
3350  // We can always get through zext.
3351  if (isa<ZExtInst>(Inst))
3352  return true;
3353 
3354  // sext(sext) is ok too.
3355  if (IsSExt && isa<SExtInst>(Inst))
3356  return true;
3357 
3358  // We can get through binary operator, if it is legal. In other words, the
3359  // binary operator must have a nuw or nsw flag.
3360  const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
3361  if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
3362  ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
3363  (IsSExt && BinOp->hasNoSignedWrap())))
3364  return true;
3365 
3366  // Check if we can do the following simplification.
3367  // ext(trunc(opnd)) --> ext(opnd)
3368  if (!isa<TruncInst>(Inst))
3369  return false;
3370 
3371  Value *OpndVal = Inst->getOperand(0);
3372  // Check if we can use this operand in the extension.
3373  // If the type is larger than the result type of the extension, we cannot.
3374  if (!OpndVal->getType()->isIntegerTy() ||
3375  OpndVal->getType()->getIntegerBitWidth() >
3376  ConsideredExtType->getIntegerBitWidth())
3377  return false;
3378 
3379  // If the operand of the truncate is not an instruction, we will not have
3380  // any information on the dropped bits.
3381  // (Actually we could for constant but it is not worth the extra logic).
3382  Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
3383  if (!Opnd)
3384  return false;
3385 
3386  // Check if the source of the type is narrow enough.
3387  // I.e., check that trunc just drops extended bits of the same kind of
3388  // the extension.
3389  // #1 get the type of the operand and check the kind of the extended bits.
3390  const Type *OpndType;
3391  InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
3392  if (It != PromotedInsts.end() && It->second.getInt() == IsSExt)
3393  OpndType = It->second.getPointer();
3394  else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd)))
3395  OpndType = Opnd->getOperand(0)->getType();
3396  else
3397  return false;
3398 
3399  // #2 check that the truncate just drops extended bits.
3400  return Inst->getType()->getIntegerBitWidth() >=
3401  OpndType->getIntegerBitWidth();
3402 }
3403 
3404 TypePromotionHelper::Action TypePromotionHelper::getAction(
3405  Instruction *Ext, const SetOfInstrs &InsertedInsts,
3406  const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
3407  assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
3408  "Unexpected instruction type");
3409  Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0));
3410  Type *ExtTy = Ext->getType();
3411  bool IsSExt = isa<SExtInst>(Ext);
3412  // If the operand of the extension is not an instruction, we cannot
3413  // get through.
3414  // If it, check we can get through.
3415  if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt))
3416  return nullptr;
3417 
3418  // Do not promote if the operand has been added by codegenprepare.
3419  // Otherwise, it means we are undoing an optimization that is likely to be
3420  // redone, thus causing potential infinite loop.
3421  if (isa<TruncInst>(ExtOpnd) && InsertedInsts.count(ExtOpnd))
3422  return nullptr;
3423 
3424  // SExt or Trunc instructions.
3425  // Return the related handler.
3426  if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) ||
3427  isa<ZExtInst>(ExtOpnd))
3428  return promoteOperandForTruncAndAnyExt;
3429 
3430  // Regular instruction.
3431  // Abort early if we will have to insert non-free instructions.
3432  if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType()))
3433  return nullptr;
3434  return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
3435 }
3436 
3437 Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
3438  Instruction *SExt, TypePromotionTransaction &TPT,
3439  InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3441  SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3442  // By construction, the operand of SExt is an instruction. Otherwise we cannot
3443  // get through it and this method should not be called.
3444  Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
3445  Value *ExtVal = SExt;
3446  bool HasMergedNonFreeExt = false;
3447  if (isa<ZExtInst>(SExtOpnd)) {
3448  // Replace s|zext(zext(opnd))
3449  // => zext(opnd).
3450  HasMergedNonFreeExt = !TLI.isExtFree(SExtOpnd);
3451  Value *ZExt =
3452  TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
3453  TPT.replaceAllUsesWith(SExt, ZExt);
3454  TPT.eraseInstruction(SExt);
3455  ExtVal = ZExt;
3456  } else {
3457  // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
3458  // => z|sext(opnd).
3459  TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
3460  }
3461  CreatedInstsCost = 0;
3462 
3463  // Remove dead code.
3464  if (SExtOpnd->use_empty())
3465  TPT.eraseInstruction(SExtOpnd);
3466 
3467  // Check if the extension is still needed.
3468  Instruction *ExtInst = dyn_cast<Instruction>(ExtVal);
3469  if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) {
3470  if (ExtInst) {
3471  if (Exts)
3472  Exts->push_back(ExtInst);
3473  CreatedInstsCost = !TLI.isExtFree(ExtInst) && !HasMergedNonFreeExt;
3474  }
3475  return ExtVal;
3476  }
3477 
3478  // At this point we have: ext ty opnd to ty.
3479  // Reassign the uses of ExtInst to the opnd and remove ExtInst.
3480  Value *NextVal = ExtInst->getOperand(0);
3481  TPT.eraseInstruction(ExtInst, NextVal);
3482  return NextVal;
3483 }
3484 
3485 Value *TypePromotionHelper::promoteOperandForOther(
3486  Instruction *Ext, TypePromotionTransaction &TPT,
3487  InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3489  SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI,
3490  bool IsSExt) {
3491  // By construction, the operand of Ext is an instruction. Otherwise we cannot
3492  // get through it and this method should not be called.
3493  Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0));
3494  CreatedInstsCost = 0;
3495  if (!ExtOpnd->hasOneUse()) {
3496  // ExtOpnd will be promoted.
3497  // All its uses, but Ext, will need to use a truncated value of the
3498  // promoted version.
3499  // Create the truncate now.
3500  Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType());
3501  if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) {
3502  // Insert it just after the definition.
3503  ITrunc->moveAfter(ExtOpnd);
3504  if (Truncs)
3505  Truncs->push_back(ITrunc);
3506  }
3507 
3508  TPT.replaceAllUsesWith(ExtOpnd, Trunc);
3509  // Restore the operand of Ext (which has been replaced by the previous call
3510  // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
3511  TPT.setOperand(Ext, 0, ExtOpnd);
3512  }
3513 
3514  // Get through the Instruction:
3515  // 1. Update its type.
3516  // 2. Replace the uses of Ext by Inst.
3517  // 3. Extend each operand that needs to be extended.
3518 
3519  // Remember the original type of the instruction before promotion.
3520  // This is useful to know that the high bits are sign extended bits.
3521  PromotedInsts.insert(std::pair<Instruction *, TypeIsSExt>(
3522  ExtOpnd, TypeIsSExt(ExtOpnd->getType(), IsSExt)));
3523  // Step #1.
3524  TPT.mutateType(ExtOpnd, Ext->getType());
3525  // Step #2.
3526  TPT.replaceAllUsesWith(Ext, ExtOpnd);
3527  // Step #3.
3528  Instruction *ExtForOpnd = Ext;
3529 
3530  DEBUG(dbgs() << "Propagate Ext to operands\n");
3531  for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
3532  ++OpIdx) {
3533  DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n');
3534  if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() ||
3535  !shouldExtOperand(ExtOpnd, OpIdx)) {
3536  DEBUG(dbgs() << "No need to propagate\n");
3537  continue;
3538  }
3539  // Check if we can statically extend the operand.
3540  Value *Opnd = ExtOpnd->getOperand(OpIdx);
3541  if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
3542  DEBUG(dbgs() << "Statically extend\n");
3543  unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
3544  APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth)
3545  : Cst->getValue().zext(BitWidth);
3546  TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal));
3547  continue;
3548  }
3549  // UndefValue are typed, so we have to statically sign extend them.
3550  if (isa<UndefValue>(Opnd)) {
3551  DEBUG(dbgs() << "Statically extend\n");
3552  TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType()));
3553  continue;
3554  }
3555 
3556  // Otherwise we have to explicity sign extend the operand.
3557  // Check if Ext was reused to extend an operand.
3558  if (!ExtForOpnd) {
3559  // If yes, create a new one.
3560  DEBUG(dbgs() << "More operands to ext\n");
3561  Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType())
3562  : TPT.createZExt(Ext, Opnd, Ext->getType());
3563  if (!isa<Instruction>(ValForExtOpnd)) {
3564  TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd);
3565  continue;
3566  }
3567  ExtForOpnd = cast<Instruction>(ValForExtOpnd);
3568  }
3569  if (Exts)
3570  Exts->push_back(ExtForOpnd);
3571  TPT.setOperand(ExtForOpnd, 0, Opnd);
3572 
3573  // Move the sign extension before the insertion point.
3574  TPT.moveBefore(ExtForOpnd, ExtOpnd);
3575  TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd);
3576  CreatedInstsCost += !TLI.isExtFree(ExtForOpnd);
3577  // If more sext are required, new instructions will have to be created.
3578  ExtForOpnd = nullptr;
3579  }
3580  if (ExtForOpnd == Ext) {
3581  DEBUG(dbgs() << "Extension is useless now\n");
3582  TPT.eraseInstruction(Ext);
3583  }
3584  return ExtOpnd;
3585 }
3586 
3587 /// Check whether or not promoting an instruction to a wider type is profitable.
3588 /// \p NewCost gives the cost of extension instructions created by the
3589 /// promotion.
3590 /// \p OldCost gives the cost of extension instructions before the promotion
3591 /// plus the number of instructions that have been
3592 /// matched in the addressing mode the promotion.
3593 /// \p PromotedOperand is the value that has been promoted.
3594 /// \return True if the promotion is profitable, false otherwise.
3595 bool AddressingModeMatcher::isPromotionProfitable(
3596  unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const {
3597  DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCost << '\n');
3598  // The cost of the new extensions is greater than the cost of the
3599  // old extension plus what we folded.
3600  // This is not profitable.
3601  if (NewCost > OldCost)
3602  return false;
3603  if (NewCost < OldCost)
3604  return true;
3605  // The promotion is neutral but it may help folding the sign extension in
3606  // loads for instance.
3607  // Check that we did not create an illegal instruction.
3608  return isPromotedInstructionLegal(TLI, DL, PromotedOperand);
3609 }
3610 
3611 /// Given an instruction or constant expr, see if we can fold the operation
3612 /// into the addressing mode. If so, update the addressing mode and return
3613 /// true, otherwise return false without modifying AddrMode.
3614 /// If \p MovedAway is not NULL, it contains the information of whether or
3615 /// not AddrInst has to be folded into the addressing mode on success.
3616 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
3617 /// because it has been moved away.
3618 /// Thus AddrInst must not be added in the matched instructions.
3619 /// This state can happen when AddrInst is a sext, since it may be moved away.
3620 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
3621 /// not be referenced anymore.
3622 bool AddressingModeMatcher::matchOperationAddr(User *AddrInst, unsigned Opcode,
3623  unsigned Depth,
3624  bool *MovedAway) {
3625  // Avoid exponential behavior on extremely deep expression trees.
3626  if (Depth >= 5) return false;
3627 
3628  // By default, all matched instructions stay in place.
3629  if (MovedAway)
3630  *MovedAway = false;
3631 
3632  switch (Opcode) {
3633  case Instruction::PtrToInt:
3634  // PtrToInt is always a noop, as we know that the int type is pointer sized.
3635  return matchAddr(AddrInst->getOperand(0), Depth);
3636  case Instruction::IntToPtr: {
3637  auto AS = AddrInst->getType()->getPointerAddressSpace();
3638  auto PtrTy = MVT::getIntegerVT(DL.getPointerSizeInBits(AS));
3639  // This inttoptr is a no-op if the integer type is pointer sized.
3640  if (TLI.getValueType(DL, AddrInst->getOperand(0)->getType()) == PtrTy)
3641  return matchAddr(AddrInst->getOperand(0), Depth);
3642  return false;
3643  }
3644  case Instruction::BitCast:
3645  // BitCast is always a noop, and we can handle it as long as it is
3646  // int->int or pointer->pointer (we don't want int<->fp or something).
3647  if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
3648  AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
3649  // Don't touch identity bitcasts. These were probably put here by LSR,
3650  // and we don't want to mess around with them. Assume it knows what it
3651  // is doing.
3652  AddrInst->getOperand(0)->getType() != AddrInst->getType())
3653  return matchAddr(AddrInst->getOperand(0), Depth);
3654  return false;
3655  case Instruction::AddrSpaceCast: {
3656  unsigned SrcAS
3657  = AddrInst->getOperand(0)->getType()->getPointerAddressSpace();
3658  unsigned DestAS = AddrInst->getType()->getPointerAddressSpace();
3659  if (TLI.isNoopAddrSpaceCast(SrcAS, DestAS))
3660  return matchAddr(AddrInst->getOperand(0), Depth);
3661  return false;
3662  }
3663  case Instruction::Add: {
3664  // Check to see if we can merge in the RHS then the LHS. If so, we win.
3665  ExtAddrMode BackupAddrMode = AddrMode;
3666  unsigned OldSize = AddrModeInsts.size();
3667  // Start a transaction at this point.
3668  // The LHS may match but not the RHS.
3669  // Therefore, we need a higher level restoration point to undo partially
3670  // matched operation.
3671  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3672  TPT.getRestorationPoint();
3673 
3674  if (matchAddr(AddrInst->getOperand(1), Depth+1) &&
3675  matchAddr(AddrInst->getOperand(0), Depth+1))
3676  return true;
3677 
3678  // Restore the old addr mode info.
3679  AddrMode = BackupAddrMode;
3680  AddrModeInsts.resize(OldSize);
3681  TPT.rollback(LastKnownGood);
3682 
3683  // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
3684  if (matchAddr(AddrInst->getOperand(0), Depth+1) &&
3685  matchAddr(AddrInst->getOperand(1), Depth+1))
3686  return true;
3687 
3688  // Otherwise we definitely can't merge the ADD in.
3689  AddrMode = BackupAddrMode;
3690  AddrModeInsts.resize(OldSize);
3691  TPT.rollback(LastKnownGood);
3692  break;
3693  }
3694  //case Instruction::Or:
3695  // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
3696  //break;
3697  case Instruction::Mul:
3698  case Instruction::Shl: {
3699  // Can only handle X*C and X << C.
3700  ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
3701  if (!RHS || RHS->getBitWidth() > 64)
3702  return false;
3703  int64_t Scale = RHS->getSExtValue();
3704  if (Opcode == Instruction::Shl)
3705  Scale = 1LL << Scale;
3706 
3707  return matchScaledValue(AddrInst->getOperand(0), Scale, Depth);
3708  }
3709  case Instruction::GetElementPtr: {
3710  // Scan the GEP. We check it if it contains constant offsets and at most
3711  // one variable offset.
3712  int VariableOperand = -1;
3713  unsigned VariableScale = 0;
3714 
3715  int64_t ConstantOffset = 0;
3716  gep_type_iterator GTI = gep_type_begin(AddrInst);
3717  for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
3718  if (StructType *STy = GTI.getStructTypeOrNull()) {
3719  const StructLayout *SL = DL.getStructLayout(STy);
3720  unsigned Idx =
3721  cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
3722  ConstantOffset += SL->getElementOffset(Idx);
3723  } else {
3724  uint64_t TypeSize = DL.getTypeAllocSize(GTI.getIndexedType());
3725  if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
3726  ConstantOffset += CI->getSExtValue() * TypeSize;
3727  } else if (TypeSize) { // Scales of zero don't do anything.
3728  // We only allow one variable index at the moment.
3729  if (VariableOperand != -1)
3730  return false;
3731 
3732  // Remember the variable index.
3733  VariableOperand = i;
3734  VariableScale = TypeSize;
3735  }
3736  }
3737  }
3738 
3739  // A common case is for the GEP to only do a constant offset. In this case,
3740  // just add it to the disp field and check validity.
3741  if (VariableOperand == -1) {
3742  AddrMode.BaseOffs += ConstantOffset;
3743  if (ConstantOffset == 0 ||
3744  TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) {
3745  // Check to see if we can fold the base pointer in too.
3746  if (matchAddr(AddrInst->getOperand(0), Depth+1))
3747  return true;
3748  }
3749  AddrMode.BaseOffs -= ConstantOffset;
3750  return false;
3751  }
3752 
3753  // Save the valid addressing mode in case we can't match.
3754  ExtAddrMode BackupAddrMode = AddrMode;
3755  unsigned OldSize = AddrModeInsts.size();
3756 
3757  // See if the scale and offset amount is valid for this target.
3758  AddrMode.BaseOffs += ConstantOffset;
3759 
3760  // Match the base operand of the GEP.
3761  if (!matchAddr(AddrInst->getOperand(0), Depth+1)) {
3762  // If it couldn't be matched, just stuff the value in a register.
3763  if (AddrMode.HasBaseReg) {
3764  AddrMode = BackupAddrMode;
3765  AddrModeInsts.resize(OldSize);
3766  return false;
3767  }
3768  AddrMode.HasBaseReg = true;
3769  AddrMode.BaseReg = AddrInst->getOperand(0);
3770  }
3771 
3772  // Match the remaining variable portion of the GEP.
3773  if (!matchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
3774  Depth)) {
3775  // If it couldn't be matched, try stuffing the base into a register
3776  // instead of matching it, and retrying the match of the scale.
3777  AddrMode = BackupAddrMode;
3778  AddrModeInsts.resize(OldSize);
3779  if (AddrMode.HasBaseReg)
3780  return false;
3781  AddrMode.HasBaseReg = true;
3782  AddrMode.BaseReg = AddrInst->getOperand(0);
3783  AddrMode.BaseOffs += ConstantOffset;
3784  if (!matchScaledValue(AddrInst->getOperand(VariableOperand),
3785  VariableScale, Depth)) {
3786  // If even that didn't work, bail.
3787  AddrMode = BackupAddrMode;
3788  AddrModeInsts.resize(OldSize);
3789  return false;
3790  }
3791  }
3792 
3793  return true;
3794  }
3795  case Instruction::SExt:
3796  case Instruction::ZExt: {
3797  Instruction *Ext = dyn_cast<Instruction>(AddrInst);
3798  if (!Ext)
3799  return false;
3800 
3801  // Try to move this ext out of the way of the addressing mode.
3802  // Ask for a method for doing so.
3803  TypePromotionHelper::Action TPH =
3804  TypePromotionHelper::getAction(Ext, InsertedInsts, TLI, PromotedInsts);
3805  if (!TPH)
3806  return false;
3807 
3808  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3809  TPT.getRestorationPoint();
3810  unsigned CreatedInstsCost = 0;
3811  unsigned ExtCost = !TLI.isExtFree(Ext);
3812  Value *PromotedOperand =
3813  TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI);
3814  // SExt has been moved away.
3815  // Thus either it will be rematched later in the recursive calls or it is
3816  // gone. Anyway, we must not fold it into the addressing mode at this point.
3817  // E.g.,
3818  // op = add opnd, 1
3819  // idx = ext op
3820  // addr = gep base, idx
3821  // is now:
3822  // promotedOpnd = ext opnd <- no match here
3823  // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
3824  // addr = gep base, op <- match
3825  if (MovedAway)
3826  *MovedAway = true;
3827 
3828  assert(PromotedOperand &&
3829  "TypePromotionHelper should have filtered out those cases");
3830 
3831  ExtAddrMode BackupAddrMode = AddrMode;
3832  unsigned OldSize = AddrModeInsts.size();
3833 
3834  if (!matchAddr(PromotedOperand, Depth) ||
3835  // The total of the new cost is equal to the cost of the created
3836  // instructions.
3837  // The total of the old cost is equal to the cost of the extension plus
3838  // what we have saved in the addressing mode.
3839  !isPromotionProfitable(CreatedInstsCost,
3840  ExtCost + (AddrModeInsts.size() - OldSize),
3841  PromotedOperand)) {
3842  AddrMode = BackupAddrMode;
3843  AddrModeInsts.resize(OldSize);
3844  DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
3845  TPT.rollback(LastKnownGood);
3846  return false;
3847  }
3848  return true;
3849  }
3850  }
3851  return false;
3852 }
3853 
3854 /// If we can, try to add the value of 'Addr' into the current addressing mode.
3855 /// If Addr can't be added to AddrMode this returns false and leaves AddrMode
3856 /// unmodified. This assumes that Addr is either a pointer type or intptr_t
3857 /// for the target.
3858 ///
3859 bool AddressingModeMatcher::matchAddr(Value *Addr, unsigned Depth) {
3860  // Start a transaction at this point that we will rollback if the matching
3861  // fails.
3862  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3863  TPT.getRestorationPoint();
3864  if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
3865  // Fold in immediates if legal for the target.
3866  AddrMode.BaseOffs += CI->getSExtValue();
3867  if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3868  return true;
3869  AddrMode.BaseOffs -= CI->getSExtValue();
3870  } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
3871  // If this is a global variable, try to fold it into the addressing mode.
3872  if (!AddrMode.BaseGV) {
3873  AddrMode.BaseGV = GV;
3874  if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3875  return true;
3876  AddrMode.BaseGV = nullptr;
3877  }
3878  } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
3879  ExtAddrMode BackupAddrMode = AddrMode;
3880  unsigned OldSize = AddrModeInsts.size();
3881 
3882  // Check to see if it is possible to fold this operation.
3883  bool MovedAway = false;
3884  if (matchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
3885  // This instruction may have been moved away. If so, there is nothing
3886  // to check here.
3887  if (MovedAway)
3888  return true;
3889  // Okay, it's possible to fold this. Check to see if it is actually
3890  // *profitable* to do so. We use a simple cost model to avoid increasing
3891  // register pressure too much.
3892  if (I->hasOneUse() ||
3893  isProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
3894  AddrModeInsts.push_back(I);
3895  return true;
3896  }
3897 
3898  // It isn't profitable to do this, roll back.
3899  //cerr << "NOT FOLDING: " << *I;
3900  AddrMode = BackupAddrMode;
3901  AddrModeInsts.resize(OldSize);
3902  TPT.rollback(LastKnownGood);
3903  }
3904  } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
3905  if (matchOperationAddr(CE, CE->getOpcode(), Depth))
3906  return true;
3907  TPT.rollback(LastKnownGood);
3908  } else if (isa<ConstantPointerNull>(Addr)) {
3909  // Null pointer gets folded without affecting the addressing mode.
3910  return true;
3911  }
3912 
3913  // Worse case, the target should support [reg] addressing modes. :)
3914  if (!AddrMode.HasBaseReg) {
3915  AddrMode.HasBaseReg = true;
3916  AddrMode.BaseReg = Addr;
3917  // Still check for legality in case the target supports [imm] but not [i+r].
3918  if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3919  return true;
3920  AddrMode.HasBaseReg = false;
3921  AddrMode.BaseReg = nullptr;
3922  }
3923 
3924  // If the base register is already taken, see if we can do [r+r].
3925  if (AddrMode.Scale == 0) {
3926  AddrMode.Scale = 1;
3927  AddrMode.ScaledReg = Addr;
3928  if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
3929  return true;
3930  AddrMode.Scale = 0;
3931  AddrMode.ScaledReg = nullptr;
3932  }
3933  // Couldn't match.
3934  TPT.rollback(LastKnownGood);
3935  return false;
3936 }
3937 
3938 /// Check to see if all uses of OpVal by the specified inline asm call are due
3939 /// to memory operands. If so, return true, otherwise return false.
3940 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
3941  const TargetLowering &TLI,
3942  const TargetRegisterInfo &TRI) {
3943  const Function *F = CI->getFunction();
3944  TargetLowering::AsmOperandInfoVector TargetConstraints =
3945  TLI.ParseConstraints(F->getParent()->getDataLayout(), &TRI,
3946  ImmutableCallSite(CI));
3947 
3948  for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
3949  TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
3950 
3951  // Compute the constraint code and ConstraintType to use.
3952  TLI.ComputeConstraintToUse(OpInfo, SDValue());
3953 
3954  // If this asm operand is our Value*, and if it isn't an indirect memory
3955  // operand, we can't fold it!
3956  if (OpInfo.CallOperandVal == OpVal &&
3958  !OpInfo.isIndirect))
3959  return false;
3960  }
3961 
3962  return true;
3963 }
3964 
3965 // Max number of memory uses to look at before aborting the search to conserve
3966 // compile time.
3967 static constexpr int MaxMemoryUsesToScan = 20;
3968 
3969 /// Recursively walk all the uses of I until we find a memory use.
3970 /// If we find an obviously non-foldable instruction, return true.
3971 /// Add the ultimately found memory instructions to MemoryUses.
3972 static bool FindAllMemoryUses(
3973  Instruction *I,
3974  SmallVectorImpl<std::pair<Instruction *, unsigned>> &MemoryUses,
3975  SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetLowering &TLI,
3976  const TargetRegisterInfo &TRI, int SeenInsts = 0) {
3977  // If we already considered this instruction, we're done.
3978  if (!ConsideredInsts.insert(I).second)
3979  return false;
3980 
3981  // If this is an obviously unfoldable instruction, bail out.
3982  if (!MightBeFoldableInst(I))
3983  return true;
3984 
3985  const bool OptSize = I->getFunction()->optForSize();
3986 
3987  // Loop over all the uses, recursively processing them.
3988  for (Use &U : I->uses()) {
3989  // Conservatively return true if we're seeing a large number or a deep chain
3990  // of users. This avoids excessive compilation times in pathological cases.
3991  if (SeenInsts++ >= MaxMemoryUsesToScan)
3992  return true;
3993 
3994  Instruction *UserI = cast<Instruction>(U.getUser());
3995  if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
3996  MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
3997  continue;
3998  }
3999 
4000  if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
4001  unsigned opNo = U.getOperandNo();
4002  if (opNo != StoreInst::getPointerOperandIndex())
4003  return true; // Storing addr, not into addr.
4004  MemoryUses.push_back(std::make_pair(SI, opNo));
4005  continue;
4006  }
4007 
4008  if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(UserI)) {
4009  unsigned opNo = U.getOperandNo();
4011  return true; // Storing addr, not into addr.
4012  MemoryUses.push_back(std::make_pair(RMW, opNo));
4013  continue;
4014  }
4015 
4016  if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(UserI)) {
4017  unsigned opNo = U.getOperandNo();
4019  return true; // Storing addr, not into addr.
4020  MemoryUses.push_back(std::make_pair(CmpX, opNo));
4021  continue;
4022  }
4023 
4024  if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
4025  // If this is a cold call, we can sink the addressing calculation into
4026  // the cold path. See optimizeCallInst
4027  if (!OptSize && CI->hasFnAttr(Attribute::Cold))
4028  continue;
4029 
4031  if (!IA) return true;
4032 
4033  // If this is a memory operand, we're cool, otherwise bail out.
4034  if (!IsOperandAMemoryOperand(CI, IA, I, TLI, TRI))
4035  return true;
4036  continue;
4037  }
4038 
4039  if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TLI, TRI,
4040  SeenInsts))
4041  return true;
4042  }
4043 
4044  return false;
4045 }
4046 
4047 /// Return true if Val is already known to be live at the use site that we're
4048 /// folding it into. If so, there is no cost to include it in the addressing
4049 /// mode. KnownLive1 and KnownLive2 are two values that we know are live at the
4050 /// instruction already.
4051 bool AddressingModeMatcher::valueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
4052  Value *KnownLive2) {
4053  // If Val is either of the known-live values, we know it is live!
4054  if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
4055  return true;
4056 
4057  // All values other than instructions and arguments (e.g. constants) are live.
4058  if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
4059 
4060  // If Val is a constant sized alloca in the entry block, it is live, this is
4061  // true because it is just a reference to the stack/frame pointer, which is
4062  // live for the whole function.
4063  if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
4064  if (AI->isStaticAlloca())
4065  return true;
4066 
4067  // Check to see if this value is already used in the memory instruction's
4068  // block. If so, it's already live into the block at the very least, so we
4069  // can reasonably fold it.
4070  return Val->isUsedInBasicBlock(MemoryInst->getParent());
4071 }
4072 
4073 /// It is possible for the addressing mode of the machine to fold the specified
4074 /// instruction into a load or store that ultimately uses it.
4075 /// However, the specified instruction has multiple uses.
4076 /// Given this, it may actually increase register pressure to fold it
4077 /// into the load. For example, consider this code:
4078 ///
4079 /// X = ...
4080 /// Y = X+1
4081 /// use(Y) -> nonload/store
4082 /// Z = Y+1
4083 /// load Z
4084 ///
4085 /// In this case, Y has multiple uses, and can be folded into the load of Z
4086 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
4087 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
4088 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
4089 /// number of computations either.
4090 ///
4091 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
4092 /// X was live across 'load Z' for other reasons, we actually *would* want to
4093 /// fold the addressing mode in the Z case. This would make Y die earlier.
4094 bool AddressingModeMatcher::
4095 isProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
4096  ExtAddrMode &AMAfter) {
4097  if (IgnoreProfitability) return true;
4098 
4099  // AMBefore is the addressing mode before this instruction was folded into it,
4100  // and AMAfter is the addressing mode after the instruction was folded. Get
4101  // the set of registers referenced by AMAfter and subtract out those
4102  // referenced by AMBefore: this is the set of values which folding in this
4103  // address extends the lifetime of.
4104  //
4105  // Note that there are only two potential values being referenced here,
4106  // BaseReg and ScaleReg (global addresses are always available, as are any
4107  // folded immediates).
4108  Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
4109 
4110  // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
4111  // lifetime wasn't extended by adding this instruction.
4112  if (valueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
4113  BaseReg = nullptr;
4114  if (valueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
4115  ScaledReg = nullptr;
4116 
4117  // If folding this instruction (and it's subexprs) didn't extend any live
4118  // ranges, we're ok with it.
4119  if (!BaseReg && !ScaledReg)
4120  return true;
4121 
4122  // If all uses of this instruction can have the address mode sunk into them,
4123  // we can remove the addressing mode and effectively trade one live register
4124  // for another (at worst.) In this context, folding an addressing mode into
4125  // the use is just a particularly nice way of sinking it.
4127  SmallPtrSet<Instruction*, 16> ConsideredInsts;
4128  if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TLI, TRI))
4129  return false; // Has a non-memory, non-foldable use!
4130 
4131  // Now that we know that all uses of this instruction are part of a chain of
4132  // computation involving only operations that could theoretically be folded
4133  // into a memory use, loop over each of these memory operation uses and see
4134  // if they could *actually* fold the instruction. The assumption is that
4135  // addressing modes are cheap and that duplicating the computation involved
4136  // many times is worthwhile, even on a fastpath. For sinking candidates
4137  // (i.e. cold call sites), this serves as a way to prevent excessive code
4138  // growth since most architectures have some reasonable small and fast way to
4139  // compute an effective address. (i.e LEA on x86)
4140  SmallVector<Instruction*, 32> MatchedAddrModeInsts;
4141  for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
4142  Instruction *User = MemoryUses[i].first;
4143  unsigned OpNo = MemoryUses[i].second;
4144 
4145  // Get the access type of this use. If the use isn't a pointer, we don't
4146  // know what it accesses.
4147  Value *Address = User->getOperand(OpNo);
4148  PointerType *AddrTy = dyn_cast<PointerType>(Address->getType());
4149  if (!AddrTy)
4150  return false;
4151  Type *AddressAccessTy = AddrTy->getElementType();
4152  unsigned AS = AddrTy->getAddressSpace();
4153 
4154  // Do a match against the root of this address, ignoring profitability. This
4155  // will tell us if the addressing mode for the memory operation will
4156  // *actually* cover the shared instruction.
4157  ExtAddrMode Result;
4158  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4159  TPT.getRestorationPoint();
4160  AddressingModeMatcher Matcher(MatchedAddrModeInsts, TLI, TRI,
4161  AddressAccessTy, AS,
4162  MemoryInst, Result, InsertedInsts,
4163  PromotedInsts, TPT);
4164  Matcher.IgnoreProfitability = true;
4165  bool Success = Matcher.matchAddr(Address, 0);
4166  (void)Success; assert(Success && "Couldn't select *anything*?");
4167 
4168  // The match was to check the profitability, the changes made are not
4169  // part of the original matcher. Therefore, they should be dropped
4170  // otherwise the original matcher will not present the right state.
4171  TPT.rollback(LastKnownGood);
4172 
4173  // If the match didn't cover I, then it won't be shared by it.
4174  if (!is_contained(MatchedAddrModeInsts, I))
4175  return false;
4176 
4177  MatchedAddrModeInsts.clear();
4178  }
4179 
4180  return true;
4181 }
4182 
4183 /// Return true if the specified values are defined in a
4184 /// different basic block than BB.
4185 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
4186  if (Instruction *I = dyn_cast<Instruction>(V))
4187  return I->getParent() != BB;
4188  return false;
4189 }
4190 
4191 /// Sink addressing mode computation immediate before MemoryInst if doing so
4192 /// can be done without increasing register pressure. The need for the
4193 /// register pressure constraint means this can end up being an all or nothing
4194 /// decision for all uses of the same addressing computation.
4195 ///
4196 /// Load and Store Instructions often have addressing modes that can do
4197 /// significant amounts of computation. As such, instruction selection will try
4198 /// to get the load or store to do as much computation as possible for the
4199 /// program. The problem is that isel can only see within a single block. As
4200 /// such, we sink as much legal addressing mode work into the block as possible.
4201 ///
4202 /// This method is used to optimize both load/store and inline asms with memory
4203 /// operands. It's also used to sink addressing computations feeding into cold
4204 /// call sites into their (cold) basic block.
4205 ///
4206 /// The motivation for handling sinking into cold blocks is that doing so can
4207 /// both enable other address mode sinking (by satisfying the register pressure
4208 /// constraint above), and reduce register pressure globally (by removing the
4209 /// addressing mode computation from the fast path entirely.).
4210 bool CodeGenPrepare::optimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
4211  Type *AccessTy, unsigned AddrSpace) {
4212  Value *Repl = Addr;
4213 
4214  // Try to collapse single-value PHI nodes. This is necessary to undo
4215  // unprofitable PRE transformations.
4216  SmallVector<Value*, 8> worklist;
4217  SmallPtrSet<Value*, 16> Visited;
4218  worklist.push_back(Addr);
4219 
4220  // Use a worklist to iteratively look through PHI and select nodes, and
4221  // ensure that the addressing mode obtained from the non-PHI/select roots of
4222  // the graph are compatible.
4223  bool PhiOrSelectSeen = false;
4224  SmallVector<Instruction*, 16> AddrModeInsts;
4225  const SimplifyQuery SQ(*DL, TLInfo);
4226  AddressingModeCombiner AddrModes(SQ, { Addr, MemoryInst->getParent() });
4227  TypePromotionTransaction TPT(RemovedInsts);
4228  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4229  TPT.getRestorationPoint();
4230  while (!worklist.empty()) {
4231  Value *V = worklist.back();
4232  worklist.pop_back();
4233 
4234  // We allow traversing cyclic Phi nodes.
4235  // In case of success after this loop we ensure that traversing through
4236  // Phi nodes ends up with all cases to compute address of the form
4237  // BaseGV + Base + Scale * Index + Offset
4238  // where Scale and Offset are constans and BaseGV, Base and Index
4239  // are exactly the same Values in all cases.
4240  // It means that BaseGV, Scale and Offset dominate our memory instruction
4241  // and have the same value as they had in address computation represented
4242  // as Phi. So we can safely sink address computation to memory instruction.
4243  if (!Visited.insert(V).second)
4244  continue;
4245 
4246  // For a PHI node, push all of its incoming values.
4247  if (PHINode *P = dyn_cast<PHINode>(V)) {
4248  for (Value *IncValue : P->incoming_values())
4249  worklist.push_back(IncValue);
4250  PhiOrSelectSeen = true;
4251  continue;
4252  }
4253  // Similar for select.
4254  if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
4255  worklist.push_back(SI->getFalseValue());
4256  worklist.push_back(SI->getTrueValue());
4257  PhiOrSelectSeen = true;
4258  continue;
4259  }
4260 
4261  // For non-PHIs, determine the addressing mode being computed. Note that
4262  // the result may differ depending on what other uses our candidate
4263  // addressing instructions might have.
4264  AddrModeInsts.clear();
4265  ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
4266  V, AccessTy, AddrSpace, MemoryInst, AddrModeInsts, *TLI, *TRI,
4267  InsertedInsts, PromotedInsts, TPT);
4268  NewAddrMode.OriginalValue = V;
4269 
4270  if (!AddrModes.addNewAddrMode(NewAddrMode))
4271  break;
4272  }
4273 
4274  // Try to combine the AddrModes we've collected. If we couldn't collect any,
4275  // or we have multiple but either couldn't combine them or combining them
4276  // wouldn't do anything useful, bail out now.
4277  if (!AddrModes.combineAddrModes()) {
4278  TPT.rollback(LastKnownGood);
4279  return false;
4280  }
4281  TPT.commit();
4282 
4283  // Get the combined AddrMode (or the only AddrMode, if we only had one).
4284  ExtAddrMode AddrMode = AddrModes.getAddrMode();
4285 
4286  // If all the instructions matched are already in this BB, don't do anything.
4287  // If we saw a Phi node then it is not local definitely, and if we saw a select
4288  // then we want to push the address calculation past it even if it's already
4289  // in this BB.
4290  if (!PhiOrSelectSeen && none_of(AddrModeInsts, [&](Value *V) {
4291  return IsNonLocalValue(V, MemoryInst->getParent());
4292  })) {
4293  DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
4294  return false;
4295  }
4296 
4297  // Insert this computation right after this user. Since our caller is
4298  // scanning from the top of the BB to the bottom, reuse of the expr are
4299  // guaranteed to happen later.
4300  IRBuilder<> Builder(MemoryInst);
4301 
4302  // Now that we determined the addressing expression we want to use and know
4303  // that we have to sink it into this block. Check to see if we have already
4304  // done this for some other load/store instr in this block. If so, reuse
4305  // the computation. Before attempting reuse, check if the address is valid
4306  // as it may have been erased.
4307 
4308  WeakTrackingVH SunkAddrVH = SunkAddrs[Addr];
4309 
4310  Value * SunkAddr = SunkAddrVH.pointsToAliveValue() ? SunkAddrVH : nullptr;
4311  if (SunkAddr) {
4312  DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
4313  << *MemoryInst << "\n");
4314  if (SunkAddr->getType() != Addr->getType())
4315  SunkAddr = Builder.CreatePointerCast(SunkAddr, Addr->getType());
4316  } else if (AddrSinkUsingGEPs ||
4317  (!AddrSinkUsingGEPs.getNumOccurrences() && TM && TTI->useAA())) {
4318  // By default, we use the GEP-based method when AA is used later. This
4319  // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
4320  DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
4321  << *MemoryInst << "\n");
4322  Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
4323  Value *ResultPtr = nullptr, *ResultIndex = nullptr;
4324 
4325  // First, find the pointer.
4326  if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
4327  ResultPtr = AddrMode.BaseReg;
4328  AddrMode.BaseReg = nullptr;
4329  }
4330 
4331  if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
4332  // We can't add more than one pointer together, nor can we scale a
4333  // pointer (both of which seem meaningless).
4334  if (ResultPtr || AddrMode.Scale != 1)
4335  return false;
4336 
4337  ResultPtr = AddrMode.ScaledReg;
4338  AddrMode.Scale = 0;
4339  }
4340 
4341  // It is only safe to sign extend the BaseReg if we know that the math
4342  // required to create it did not overflow before we extend it. Since
4343  // the original IR value was tossed in favor of a constant back when
4344  // the AddrMode was created we need to bail out gracefully if widths
4345  // do not match instead of extending it.
4346  //
4347  // (See below for code to add the scale.)
4348  if (AddrMode.Scale) {
4349  Type *ScaledRegTy = AddrMode.ScaledReg->getType();
4350  if (cast<IntegerType>(IntPtrTy)->getBitWidth() >
4351  cast<IntegerType>(ScaledRegTy)->getBitWidth())
4352  return false;
4353  }
4354 
4355  if (AddrMode.BaseGV) {
4356  if (ResultPtr)
4357  return false;
4358 
4359  ResultPtr = AddrMode.BaseGV;
4360  }
4361 
4362  // If the real base value actually came from an inttoptr, then the matcher
4363  // will look through it and provide only the integer value. In that case,
4364  // use it here.
4365  if (!DL->isNonIntegralPointerType(Addr->getType())) {
4366  if (!ResultPtr && AddrMode.BaseReg) {
4367  ResultPtr = Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(),
4368  "sunkaddr");
4369  AddrMode.BaseReg = nullptr;
4370  } else if (!ResultPtr && AddrMode.Scale == 1) {
4371  ResultPtr = Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(),
4372  "sunkaddr");
4373  AddrMode.Scale = 0;
4374  }
4375  }
4376 
4377  if (!ResultPtr &&
4378  !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
4379  SunkAddr = Constant::getNullValue(Addr->getType());
4380  } else if (!ResultPtr) {
4381  return false;
4382  } else {
4383  Type *I8PtrTy =
4384  Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
4385  Type *I8Ty = Builder.getInt8Ty();
4386 
4387  // Start with the base register. Do this first so that subsequent address
4388  // matching finds it last, which will prevent it from trying to match it
4389  // as the scaled value in case it happens to be a mul. That would be
4390  // problematic if we've sunk a different mul for the scale, because then
4391  // we'd end up sinking both muls.
4392  if (AddrMode.BaseReg) {
4393  Value *V = AddrMode.BaseReg;
4394  if (V->getType() != IntPtrTy)
4395  V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
4396 
4397  ResultIndex = V;
4398  }
4399 
4400  // Add the scale value.
4401  if (AddrMode.Scale) {
4402  Value *V = AddrMode.ScaledReg;
4403  if (V->getType() == IntPtrTy) {
4404  // done.
4405  } else {
4406  assert(cast<IntegerType>(IntPtrTy)->getBitWidth() <
4407  cast<IntegerType>(V->getType())->getBitWidth() &&
4408  "We can't transform if ScaledReg is too narrow");
4409  V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
4410  }
4411 
4412  if (AddrMode.Scale != 1)
4413  V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
4414  "sunkaddr");
4415  if (ResultIndex)
4416  ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
4417  else
4418  ResultIndex = V;
4419  }
4420 
4421  // Add in the Base Offset if present.
4422  if (AddrMode.BaseOffs) {
4423  Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
4424  if (ResultIndex) {
4425  // We need to add this separately from the scale above to help with
4426  // SDAG consecutive load/store merging.
4427  if (ResultPtr->getType() != I8PtrTy)
4428  ResultPtr = Builder.CreatePointerCast(ResultPtr, I8PtrTy);
4429  ResultPtr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
4430  }
4431 
4432  ResultIndex = V;
4433  }
4434 
4435  if (!ResultIndex) {
4436  SunkAddr = ResultPtr;
4437  } else {
4438  if (ResultPtr->getType() != I8PtrTy)
4439  ResultPtr = Builder.CreatePointerCast(ResultPtr, I8PtrTy);
4440  SunkAddr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
4441  }
4442 
4443  if (SunkAddr->getType() != Addr->getType())
4444  SunkAddr = Builder.CreatePointerCast(SunkAddr, Addr->getType());
4445  }
4446  } else {
4447  // We'd require a ptrtoint/inttoptr down the line, which we can't do for
4448  // non-integral pointers, so in that case bail out now.
4449  Type *BaseTy = AddrMode.BaseReg ? AddrMode.BaseReg->getType() : nullptr;
4450  Type *ScaleTy = AddrMode.Scale ? AddrMode.ScaledReg->getType() : nullptr;
4451  PointerType *BasePtrTy = dyn_cast_or_null<PointerType>(BaseTy);
4452  PointerType *ScalePtrTy = dyn_cast_or_null<PointerType>(ScaleTy);
4453  if (DL->isNonIntegralPointerType(Addr->getType()) ||
4454  (BasePtrTy && DL->isNonIntegralPointerType(BasePtrTy)) ||
4455  (ScalePtrTy && DL->isNonIntegralPointerType(ScalePtrTy)) ||
4456  (AddrMode.BaseGV &&
4457  DL->isNonIntegralPointerType(AddrMode.BaseGV->getType())))
4458  return false;
4459 
4460  DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
4461  << *MemoryInst << "\n");
4462  Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
4463  Value *Result = nullptr;
4464 
4465  // Start with the base register. Do this first so that subsequent address
4466  // matching finds it last, which will prevent it from trying to match it
4467  // as the scaled value in case it happens to be a mul. That would be
4468  // problematic if we've sunk a different mul for the scale, because then
4469  // we'd end up sinking both muls.
4470  if (AddrMode.BaseReg) {
4471  Value *V = AddrMode.BaseReg;
4472  if (V->getType()->isPointerTy())
4473  V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
4474  if (V->getType() != IntPtrTy)
4475  V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
4476  Result = V;
4477  }
4478 
4479  // Add the scale value.
4480  if (AddrMode.Scale) {
4481  Value *V = AddrMode.ScaledReg;
4482  if (V->getType() == IntPtrTy) {
4483  // done.
4484  } else if (V->getType()->isPointerTy()) {
4485  V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
4486  } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
4487  cast<IntegerType>(V->getType())->getBitWidth()) {
4488  V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
4489  } else {
4490  // It is only safe to sign extend the BaseReg if we know that the math
4491  // required to create it did not overflow before we extend it. Since
4492  // the original IR value was tossed in favor of a constant back when
4493  // the AddrMode was created we need to bail out gracefully if widths
4494  // do not match instead of extending it.
4495  Instruction *I = dyn_cast_or_null<Instruction>(Result);
4496  if (I && (Result != AddrMode.BaseReg))
4497  I->eraseFromParent();
4498  return false;
4499  }
4500  if (AddrMode.Scale != 1)
4501  V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
4502  "sunkaddr");
4503  if (Result)
4504  Result = Builder.CreateAdd(Result, V, "sunkaddr");
4505  else
4506  Result = V;
4507  }
4508 
4509  // Add in the BaseGV if present.
4510  if (AddrMode.BaseGV) {
4511  Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
4512  if (Result)
4513  Result = Builder.CreateAdd(Result, V, "sunkaddr");
4514  else
4515  Result = V;
4516  }
4517 
4518  // Add in the Base Offset if present.
4519  if (AddrMode.BaseOffs) {
4520  Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
4521  if (Result)
4522  Result = Builder.CreateAdd(Result, V, "sunkaddr");
4523  else
4524  Result = V;
4525  }
4526 
4527  if (!Result)
4528  SunkAddr = Constant::getNullValue(Addr->getType());
4529  else
4530  SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
4531  }
4532 
4533  MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
4534  // Store the newly computed address into the cache. In the case we reused a
4535  // value, this should be idempotent.
4536  SunkAddrs[Addr] = WeakTrackingVH(SunkAddr);
4537 
4538  // If we have no uses, recursively delete the value and all dead instructions
4539  // using it.
4540  if (Repl->use_empty()) {
4541  // This can cause recursive deletion, which can invalidate our iterator.
4542  // Use a WeakTrackingVH to hold onto it in case this happens.
4543  Value *CurValue = &*CurInstIterator;
4544  WeakTrackingVH IterHandle(CurValue);
4545  BasicBlock *BB = CurInstIterator->getParent();
4546 
4548 
4549  if (IterHandle != CurValue) {
4550  // If the iterator instruction was recursively deleted, start over at the
4551  // start of the block.
4552  CurInstIterator = BB->begin();
4553  SunkAddrs.clear();
4554  }
4555  }
4556  ++NumMemoryInsts;
4557  return true;
4558 }
4559 
4560 /// If there are any memory operands, use OptimizeMemoryInst to sink their
4561 /// address computing into the block when possible / profitable.
4562 bool CodeGenPrepare::optimizeInlineAsmInst(CallInst *CS) {
4563  bool MadeChange = false;
4564 
4565  const TargetRegisterInfo *TRI =
4566  TM->getSubtargetImpl(*CS->getFunction())->getRegisterInfo();
4567  TargetLowering::AsmOperandInfoVector TargetConstraints =
4568  TLI->ParseConstraints(*DL, TRI, CS);
4569  unsigned ArgNo = 0;
4570  for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
4571  TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
4572 
4573  // Compute the constraint code and ConstraintType to use.
4574  TLI->ComputeConstraintToUse(OpInfo, SDValue());
4575 
4576  if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
4577  OpInfo.isIndirect) {
4578  Value *OpVal = CS->getArgOperand(ArgNo++);
4579  MadeChange |= optimizeMemoryInst(CS, OpVal, OpVal->getType(), ~0u);
4580  } else if (OpInfo.Type == InlineAsm::isInput)
4581  ArgNo++;
4582  }
4583 
4584  return MadeChange;
4585 }
4586 
4587 /// \brief Check if all the uses of \p Val are equivalent (or free) zero or
4588 /// sign extensions.
4589 static bool hasSameExtUse(Value *Val, const TargetLowering &TLI) {
4590  assert(!Val->use_empty() && "Input must have at least one use");
4591  const Instruction *FirstUser = cast<Instruction>(*Val->user_begin());
4592  bool IsSExt = isa<SExtInst>(FirstUser);
4593  Type *ExtTy = FirstUser->getType();
4594  for (const User *U : Val->users()) {
4595  const Instruction *UI = cast<Instruction>(U);
4596  if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI)))
4597  return false;
4598  Type *CurTy = UI->getType();
4599  // Same input and output types: Same instruction after CSE.
4600  if (CurTy == ExtTy)
4601  continue;
4602 
4603  // If IsSExt is true, we are in this situation:
4604  // a = Val
4605  // b = sext ty1 a to ty2
4606  // c = sext ty1 a to ty3
4607  // Assuming ty2 is shorter than ty3, this could be turned into:
4608  // a = Val
4609  // b = sext ty1 a to ty2
4610  // c = sext ty2 b to ty3
4611  // However, the last sext is not free.
4612  if (IsSExt)
4613  return false;
4614 
4615  // This is a ZExt, maybe this is free to extend from one type to another.
4616  // In that case, we would not account for a different use.
4617  Type *NarrowTy;
4618  Type *LargeTy;
4619  if (ExtTy->getScalarType()->getIntegerBitWidth() >
4620  CurTy->getScalarType()->getIntegerBitWidth()) {
4621  NarrowTy = CurTy;
4622  LargeTy = ExtTy;
4623  } else {
4624  NarrowTy = ExtTy;
4625  LargeTy = CurTy;
4626  }
4627 
4628  if (!TLI.isZExtFree(NarrowTy, LargeTy))
4629  return false;
4630  }
4631  // All uses are the same or can be derived from one another for free.
4632  return true;
4633 }
4634 
4635 /// \brief Try to speculatively promote extensions in \p Exts and continue
4636 /// promoting through newly promoted operands recursively as far as doing so is
4637 /// profitable. Save extensions profitably moved up, in \p ProfitablyMovedExts.
4638 /// When some promotion happened, \p TPT contains the proper state to revert
4639 /// them.
4640 ///
4641 /// \return true if some promotion happened, false otherwise.
4642 bool CodeGenPrepare::tryToPromoteExts(
4643  TypePromotionTransaction &TPT, const SmallVectorImpl<Instruction *> &Exts,
4644  SmallVectorImpl<Instruction *> &ProfitablyMovedExts,
4645  unsigned CreatedInstsCost) {
4646  bool Promoted = false;
4647 
4648  // Iterate over all the extensions to try to promote them.
4649  for (auto I : Exts) {
4650  // Early check if we directly have ext(load).
4651  if (isa<LoadInst>(I->getOperand(0))) {
4652  ProfitablyMovedExts.push_back(I);
4653  continue;
4654  }
4655 
4656  // Check whether or not we want to do any promotion. The reason we have
4657  // this check inside the for loop is to catch the case where an extension
4658  // is directly fed by a load because in such case the extension can be moved
4659  // up without any promotion on its operands.
4660  if (!TLI || !TLI->enableExtLdPromotion() || DisableExtLdPromotion)
4661  return false;
4662 
4663  // Get the action to perform the promotion.
4664  TypePromotionHelper::Action TPH =
4665  TypePromotionHelper::getAction(I, InsertedInsts, *TLI, PromotedInsts);
4666  // Check if we can promote.
4667  if (!TPH) {
4668  // Save the current extension as we cannot move up through its operand.
4669  ProfitablyMovedExts.push_back(I);
4670  continue;
4671  }
4672 
4673  // Save the current state.
4674  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4675  TPT.getRestorationPoint();
4677  unsigned NewCreatedInstsCost = 0;
4678  unsigned ExtCost = !TLI->isExtFree(I);
4679  // Promote.
4680  Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost,
4681  &NewExts, nullptr, *TLI);
4682  assert(PromotedVal &&
4683  "TypePromotionHelper should have filtered out those cases");
4684 
4685  // We would be able to merge only one extension in a load.
4686  // Therefore, if we have more than 1 new extension we heuristically
4687  // cut this search path, because it means we degrade the code quality.
4688  // With exactly 2, the transformation is neutral, because we will merge
4689  // one extension but leave one. However, we optimistically keep going,
4690  // because the new extension may be removed too.
4691  long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost;
4692  // FIXME: It would be possible to propagate a negative value instead of
4693  // conservatively ceiling it to 0.
4694  TotalCreatedInstsCost =
4695  std::max((long long)0, (TotalCreatedInstsCost - ExtCost));
4696  if (!StressExtLdPromotion &&
4697  (TotalCreatedInstsCost > 1 ||
4698  !isPromotedInstructionLegal(*TLI, *DL, PromotedVal))) {
4699  // This promotion is not profitable, rollback to the previous state, and
4700  // save the current extension in ProfitablyMovedExts as the latest
4701  // speculative promotion turned out to be unprofitable.
4702  TPT.rollback(LastKnownGood);
4703  ProfitablyMovedExts.push_back(I);
4704  continue;
4705  }
4706  // Continue promoting NewExts as far as doing so is profitable.
4707  SmallVector<Instruction *, 2> NewlyMovedExts;
4708  (void)tryToPromoteExts(TPT, NewExts, NewlyMovedExts, TotalCreatedInstsCost);
4709  bool NewPromoted = false;
4710  for (auto ExtInst : NewlyMovedExts) {
4711  Instruction *MovedExt = cast<Instruction>(ExtInst);
4712  Value *ExtOperand = MovedExt->getOperand(0);
4713  // If we have reached to a load, we need this extra profitability check
4714  // as it could potentially be merged into an ext(load).
4715  if (isa<LoadInst>(ExtOperand) &&
4716  !(StressExtLdPromotion || NewCreatedInstsCost <= ExtCost ||
4717  (ExtOperand->hasOneUse() || hasSameExtUse(ExtOperand, *TLI))))
4718  continue;
4719 
4720  ProfitablyMovedExts.push_back(MovedExt);
4721  NewPromoted = true;
4722  }
4723 
4724  // If none of speculative promotions for NewExts is profitable, rollback
4725  // and save the current extension (I) as the last profitable extension.
4726  if (!NewPromoted) {
4727  TPT.rollback(LastKnownGood);
4728  ProfitablyMovedExts.push_back(I);
4729  continue;
4730  }
4731  // The promotion is profitable.
4732  Promoted = true;
4733  }
4734  return Promoted;
4735 }
4736 
4737 /// Merging redundant sexts when one is dominating the other.
4738 bool CodeGenPrepare::mergeSExts(Function &F) {
4739  DominatorTree DT(F);
4740  bool Changed = false;
4741  for (auto &Entry : ValToSExtendedUses) {
4742  SExts &Insts = Entry.second;
4743  SExts CurPts;
4744  for (Instruction *Inst : Insts) {
4745  if (RemovedInsts.count(Inst) || !isa<SExtInst>(Inst) ||
4746  Inst->getOperand(0) != Entry.first)
4747  continue;
4748  bool inserted = false;
4749  for (auto &Pt : CurPts) {
4750  if (DT.dominates(Inst, Pt)) {
4751  Pt->replaceAllUsesWith(Inst);
4752  RemovedInsts.insert(Pt);
4753  Pt->removeFromParent();
4754  Pt = Inst;
4755  inserted = true;
4756  Changed = true;
4757  break;
4758  }
4759  if (!DT.dominates(Pt, Inst))
4760  // Give up if we need to merge in a common dominator as the
4761  // expermients show it is not profitable.
4762  continue;
4763  Inst->replaceAllUsesWith(Pt);
4764  RemovedInsts.insert(Inst);
4765  Inst->removeFromParent();
4766  inserted = true;
4767  Changed = true;
4768  break;
4769  }
4770  if (!inserted)
4771  CurPts.push_back(Inst);
4772  }
4773  }
4774  return Changed;
4775 }
4776 
4777 /// Return true, if an ext(load) can be formed from an extension in
4778 /// \p MovedExts.
4779 bool CodeGenPrepare::canFormExtLd(
4780  const SmallVectorImpl<Instruction *> &MovedExts, LoadInst *&LI,
4781  Instruction *&Inst, bool HasPromoted) {
4782  for (auto *MovedExtInst : MovedExts) {
4783  if (isa<LoadInst>(MovedExtInst->getOperand(0))) {
4784  LI = cast<LoadInst>(MovedExtInst->getOperand(0));
4785  Inst = MovedExtInst;
4786  break;
4787  }
4788  }
4789  if (!LI)
4790  return false;
4791 
4792  // If they're already in the same block, there's nothing to do.
4793  // Make the cheap checks first if we did not promote.
4794  // If we promoted, we need to check if it is indeed profitable.
4795  if (!HasPromoted && LI->getParent() == Inst->getParent())
4796  return false;
4797 
4798  return TLI->isExtLoad(LI, Inst, *DL);
4799 }
4800 
4801 /// Move a zext or sext fed by a load into the same basic block as the load,
4802 /// unless conditions are unfavorable. This allows SelectionDAG to fold the
4803 /// extend into the load.
4804 ///
4805 /// E.g.,
4806 /// \code
4807 /// %ld = load i32* %addr
4808 /// %add = add nuw i32 %ld, 4
4809 /// %zext = zext i32 %add to i64
4810 // \endcode
4811 /// =>
4812 /// \code
4813 /// %ld = load i32* %addr
4814 /// %zext = zext i32 %ld to i64
4815 /// %add = add nuw i64 %zext, 4
4816 /// \encode
4817 /// Note that the promotion in %add to i64 is done in tryToPromoteExts(), which
4818 /// allow us to match zext(load i32*) to i64.
4819 ///
4820 /// Also, try to promote the computations used to obtain a sign extended
4821 /// value used into memory accesses.
4822 /// E.g.,
4823 /// \code
4824 /// a = add nsw i32 b, 3
4825 /// d = sext i32 a to i64
4826 /// e = getelementptr ..., i64 d
4827 /// \endcode
4828 /// =>
4829 /// \code
4830 /// f = sext i32 b to i64
4831 /// a = add nsw i64 f, 3
4832 /// e = getelementptr ..., i64 a
4833 /// \endcode
4834 ///
4835 /// \p Inst[in/out] the extension may be modified during the process if some
4836 /// promotions apply.
4837 bool CodeGenPrepare::optimizeExt(Instruction *&Inst) {
4838  // ExtLoad formation and address type promotion infrastructure requires TLI to
4839  // be effective.
4840  if (!TLI)
4841  return false;
4842 
4843  bool AllowPromotionWithoutCommonHeader = false;
4844  /// See if it is an interesting sext operations for the address type
4845  /// promotion before trying to promote it, e.g., the ones with the right
4846  /// type and used in memory accesses.
4847  bool ATPConsiderable = TTI->shouldConsiderAddressTypePromotion(
4848  *Inst, AllowPromotionWithoutCommonHeader);
4849  TypePromotionTransaction TPT(RemovedInsts);
4850  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4851  TPT.getRestorationPoint();
4853  SmallVector<Instruction *, 2> SpeculativelyMovedExts;
4854  Exts.push_back(Inst);
4855 
4856  bool HasPromoted = tryToPromoteExts(TPT, Exts, SpeculativelyMovedExts);
4857 
4858  // Look for a load being extended.
4859  LoadInst *LI = nullptr;
4860  Instruction *ExtFedByLoad;
4861 
4862  // Try to promote a chain of computation if it allows to form an extended
4863  // load.
4864  if (canFormExtLd(SpeculativelyMovedExts, LI, ExtFedByLoad, HasPromoted)) {
4865  assert(LI && ExtFedByLoad && "Expect a valid load and extension");
4866  TPT.commit();
4867  // Move the extend into the same block as the load
4868  ExtFedByLoad->moveAfter(LI);
4869  // CGP does not check if the zext would be speculatively executed when moved
4870  // to the same basic block as the load. Preserving its original location
4871  // would pessimize the debugging experience, as well as negatively impact
4872  // the quality of sample pgo. We don't want to use "line 0" as that has a
4873  // size cost in the line-table section and logically the zext can be seen as
4874  // part of the load. Therefore we conservatively reuse the same debug
4875  // location for the load and the zext.
4876  ExtFedByLoad->setDebugLoc(LI->getDebugLoc());
4877  ++NumExtsMoved;
4878  Inst = ExtFedByLoad;
4879  return true;
4880  }
4881 
4882  // Continue promoting SExts if known as considerable depending on targets.
4883  if (ATPConsiderable &&
4884  performAddressTypePromotion(Inst, AllowPromotionWithoutCommonHeader,
4885  HasPromoted, TPT, SpeculativelyMovedExts))
4886  return true;
4887 
4888  TPT.rollback(LastKnownGood);
4889  return false;
4890 }
4891 
4892 // Perform address type promotion if doing so is profitable.
4893 // If AllowPromotionWithoutCommonHeader == false, we should find other sext
4894 // instructions that sign extended the same initial value. However, if
4895 // AllowPromotionWithoutCommonHeader == true, we expect promoting the
4896 // extension is just profitable.
4897 bool CodeGenPrepare::performAddressTypePromotion(
4898  Instruction *&Inst, bool AllowPromotionWithoutCommonHeader,
4899  bool HasPromoted, TypePromotionTransaction &TPT,
4900  SmallVectorImpl<Instruction *> &SpeculativelyMovedExts) {
4901  bool Promoted = false;
4902  SmallPtrSet<Instruction *, 1> UnhandledExts;
4903  bool AllSeenFirst = true;
4904  for (auto I : SpeculativelyMovedExts) {
4905  Value *HeadOfChain = I->getOperand(0);
4907  SeenChainsForSExt.find(HeadOfChain);
4908  // If there is an unhandled SExt which has the same header, try to promote
4909  // it as well.
4910  if (AlreadySeen != SeenChainsForSExt.end()) {
4911  if (AlreadySeen->second != nullptr)
4912  UnhandledExts.insert(AlreadySeen->second);
4913  AllSeenFirst = false;
4914  }
4915  }
4916 
4917  if (!AllSeenFirst || (AllowPromotionWithoutCommonHeader &&
4918  SpeculativelyMovedExts.size() == 1)) {
4919  TPT.commit();
4920  if (HasPromoted)
4921  Promoted = true;
4922  for (auto I : SpeculativelyMovedExts) {
4923  Value *HeadOfChain = I->getOperand(0);
4924  SeenChainsForSExt[HeadOfChain] = nullptr;
4925  ValToSExtendedUses[HeadOfChain].push_back(I);
4926  }
4927  // Update Inst as promotion happen.
4928  Inst = SpeculativelyMovedExts.pop_back_val();
4929  } else {
4930  // This is the first chain visited from the header, keep the current chain
4931  // as unhandled. Defer to promote this until we encounter another SExt
4932  // chain derived from the same header.
4933  for (auto I : SpeculativelyMovedExts) {
4934  Value *HeadOfChain = I->getOperand(0);
4935  SeenChainsForSExt[HeadOfChain] = Inst;
4936  }
4937  return false;
4938  }
4939 
4940  if (!AllSeenFirst && !UnhandledExts.empty())
4941  for (auto VisitedSExt : UnhandledExts) {
4942  if (RemovedInsts.count(VisitedSExt))
4943  continue;
4944  TypePromotionTransaction TPT(RemovedInsts);
4947  Exts.push_back(VisitedSExt);
4948  bool HasPromoted = tryToPromoteExts(TPT, Exts, Chains);
4949  TPT.commit();
4950  if (HasPromoted)
4951  Promoted = true;
4952  for (auto I : Chains) {
4953  Value *HeadOfChain = I->getOperand(0);
4954  // Mark this as handled.
4955  SeenChainsForSExt[HeadOfChain] = nullptr;
4956  ValToSExtendedUses[HeadOfChain].push_back(I);
4957  }
4958  }
4959  return Promoted;
4960 }
4961 
4962 bool CodeGenPrepare::optimizeExtUses(Instruction *I) {
4963  BasicBlock *DefBB = I->getParent();
4964 
4965  // If the result of a {s|z}ext and its source are both live out, rewrite all
4966  // other uses of the source with result of extension.
4967  Value *Src = I->getOperand(0);
4968  if (Src->hasOneUse())
4969  return false;
4970 
4971  // Only do this xform if truncating is free.
4972  if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
4973  return false;
4974 
4975  // Only safe to perform the optimization if the source is also defined in
4976  // this block.
4977  if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
4978  return false;
4979 
4980  bool DefIsLiveOut = false;
4981  for (User *U : I->users()) {
4982  Instruction *UI = cast<Instruction>(U);
4983 
4984  // Figure out which BB this ext is used in.
4985  BasicBlock *UserBB = UI->getParent();
4986  if (UserBB == DefBB) continue;
4987  DefIsLiveOut = true;
4988  break;
4989  }
4990  if (!DefIsLiveOut)
4991  return false;
4992 
4993  // Make sure none of the uses are PHI nodes.
4994  for (User *U : Src->users()) {
4995  Instruction *UI = cast<Instruction>(U);
4996  BasicBlock *UserBB = UI->getParent();
4997  if (UserBB == DefBB) continue;
4998  // Be conservative. We don't want this xform to end up introducing
4999  // reloads just before load / store instructions.
5000  if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
5001  return false;
5002  }
5003 
5004  // InsertedTruncs - Only insert one trunc in each block once.
5005  DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
5006 
5007  bool MadeChange = false;
5008  for (Use &U : Src->uses()) {
5009  Instruction *User = cast<Instruction>(U.getUser());
5010 
5011  // Figure out which BB this ext is used in.
5012  BasicBlock *UserBB = User->getParent();
5013  if (UserBB == DefBB) continue;
5014 
5015  // Both src and def are live in this block. Rewrite the use.
5016  Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
5017 
5018  if (!InsertedTrunc) {
5019  BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
5020  assert(InsertPt != UserBB->end());
5021  InsertedTrunc = new TruncInst(I, Src->getType(), "", &*InsertPt);
5022  InsertedInsts.insert(InsertedTrunc);
5023  }
5024 
5025  // Replace a use of the {s|z}ext source with a use of the result.
5026  U = InsertedTrunc;
5027  ++NumExtUses;
5028  MadeChange = true;
5029  }
5030 
5031  return MadeChange;
5032 }
5033 
5034 // Find loads whose uses only use some of the loaded value's bits. Add an "and"
5035 // just after the load if the target can fold this into one extload instruction,
5036 // with the hope of eliminating some of the other later "and" instructions using
5037 // the loaded value. "and"s that are made trivially redundant by the insertion
5038 // of the new "and" are removed by this function, while others (e.g. those whose
5039 // path from the load goes through a phi) are left for isel to potentially
5040 // remove.
5041 //
5042 // For example:
5043 //
5044 // b0:
5045 // x = load i32
5046 // ...
5047 // b1:
5048 // y = and x, 0xff
5049 // z = use y
5050 //
5051 // becomes:
5052 //
5053 // b0:
5054 // x = load i32
5055 // x' = and x, 0xff
5056 // ...
5057 // b1:
5058 // z = use x'
5059 //
5060 // whereas:
5061 //
5062 // b0:
5063 // x1 = load i32
5064 // ...
5065 // b1:
5066 // x2 = load i32
5067 // ...
5068 // b2:
5069 // x = phi x1, x2
5070 // y = and x, 0xff
5071 //
5072 // becomes (after a call to optimizeLoadExt for each load):
5073 //
5074 // b0:
5075 // x1 = load i32
5076 // x1' = and x1, 0xff
5077 // ...
5078 // b1:
5079 // x2 = load i32
5080 // x2' = and x2, 0xff
5081 // ...
5082 // b2:
5083 // x = phi x1', x2'
5084 // y = and x, 0xff
5085 bool CodeGenPrepare::optimizeLoadExt(LoadInst *Load) {
5086  if (!Load->isSimple() ||
5087  !(Load->getType()->isIntegerTy() || Load->getType()->isPointerTy()))
5088  return false;
5089 
5090  // Skip loads we've already transformed.
5091  if (Load->hasOneUse() &&
5092  InsertedInsts.count(cast<Instruction>(*Load->user_begin())))
5093  return false;
5094 
5095  // Look at all uses of Load, looking through phis, to determine how many bits
5096  // of the loaded value are needed.
5099  SmallVector<Instruction *, 8> AndsToMaybeRemove;
5100  for (auto *U : Load->users())
5101  WorkList.push_back(cast<Instruction>(U));
5102 
5103  EVT LoadResultVT = TLI->getValueType(*DL, Load->getType());
5104  unsigned BitWidth = LoadResultVT.getSizeInBits();
5105  APInt DemandBits(BitWidth, 0);
5106  APInt WidestAndBits(BitWidth, 0);
5107 
5108  while (!WorkList.empty()) {
5109  Instruction *I = WorkList.back();
5110  WorkList.pop_back();
5111 
5112  // Break use-def graph loops.
5113  if (!Visited.insert(I).second)
5114  continue;
5115 
5116  // For a PHI node, push all of its users.
5117  if (auto *Phi = dyn_cast<PHINode>(I)) {
5118  for (auto *U : Phi->users())
5119  WorkList.push_back(cast<Instruction>(U));
5120  continue;
5121  }
5122 
5123  switch (I->getOpcode()) {
5124  case Instruction::And: {
5125  auto *AndC = dyn_cast<ConstantInt>(I->getOperand(1));
5126  if (!AndC)
5127  return false;
5128  APInt AndBits = AndC->getValue();
5129  DemandBits |= AndBits;
5130  // Keep track of the widest and mask we see.
5131  if (AndBits.ugt(WidestAndBits))
5132  WidestAndBits = AndBits;
5133  if (AndBits == WidestAndBits && I->getOperand(0) == Load)
5134  AndsToMaybeRemove.push_back(I);
5135  break;
5136  }
5137 
5138  case Instruction::Shl: {
5139  auto *ShlC = dyn_cast<ConstantInt>(I->getOperand(1));
5140  if (!ShlC)
5141  return false;
5142  uint64_t ShiftAmt = ShlC->getLimitedValue(BitWidth - 1);
5143  DemandBits.setLowBits(BitWidth - ShiftAmt);
5144  break;
5145  }
5146 
5147  case Instruction::Trunc: {
5148  EVT TruncVT = TLI->getValueType(*DL, I->getType());
5149  unsigned TruncBitWidth = TruncVT.getSizeInBits();
5150  DemandBits.setLowBits(TruncBitWidth);
5151  break;
5152  }
5153 
5154  default:
5155  return false;
5156  }
5157  }
5158 
5159  uint32_t ActiveBits = DemandBits.getActiveBits();
5160  // Avoid hoisting (and (load x) 1) since it is unlikely to be folded by the
5161  // target even if isLoadExtLegal says an i1 EXTLOAD is valid. For example,
5162  // for the AArch64 target isLoadExtLegal(ZEXTLOAD, i32, i1) returns true, but
5163  // (and (load x) 1) is not matched as a single instruction, rather as a LDR
5164  // followed by an AND.
5165  // TODO: Look into removing this restriction by fixing backends to either
5166  // return false for isLoadExtLegal for i1 or have them select this pattern to
5167  // a single instruction.
5168  //
5169  // Also avoid hoisting if we didn't see any ands with the exact DemandBits
5170  // mask, since these are the only ands that will be removed by isel.
5171  if (ActiveBits <= 1 || !DemandBits.isMask(ActiveBits) ||
5172  WidestAndBits != DemandBits)
5173  return false;
5174 
5175  LLVMContext &Ctx = Load->getType()->getContext();
5176  Type *TruncTy = Type::getIntNTy(Ctx, ActiveBits);
5177  EVT TruncVT = TLI->getValueType(*DL, TruncTy);
5178 
5179  // Reject cases that won't be matched as extloads.
5180  if (!LoadResultVT.bitsGT(TruncVT) || !TruncVT.isRound() ||
5181  !TLI->isLoadExtLegal(ISD::ZEXTLOAD, LoadResultVT, TruncVT))
5182  return false;
5183 
5184  IRBuilder<> Builder(Load->getNextNode());
5185  auto *NewAnd = dyn_cast<Instruction>(
5186  Builder.CreateAnd(Load, ConstantInt::get(Ctx, DemandBits)));
5187  // Mark this instruction as "inserted by CGP", so that other
5188  // optimizations don't touch it.
5189  InsertedInsts.insert(NewAnd);
5190 
5191  // Replace all uses of load with new and (except for the use of load in the
5192  // new and itself).
5193  Load->replaceAllUsesWith(NewAnd);
5194  NewAnd->setOperand(0, Load);
5195 
5196  // Remove any and instructions that are now redundant.
5197  for (auto *And : AndsToMaybeRemove)
5198  // Check that the and mask is the same as the one we decided to put on the
5199  // new and.
5200  if (cast<ConstantInt>(And->getOperand(1))->getValue() == DemandBits) {
5201  And->replaceAllUsesWith(NewAnd);
5202  if (&*CurInstIterator == And)
5203  CurInstIterator = std::next(And->getIterator());
5204  And->eraseFromParent();
5205  ++NumAndUses;
5206  }
5207 
5208  ++NumAndsAdded;
5209  return true;
5210 }
5211 
5212 /// Check if V (an operand of a select instruction) is an expensive instruction
5213 /// that is only used once.
5214 static bool sinkSelectOperand(const TargetTransformInfo *TTI, Value *V) {
5215  auto *I = dyn_cast<Instruction>(V);
5216  // If it's safe to speculatively execute, then it should not have side
5217  // effects; therefore, it's safe to sink and possibly *not* execute.
5218  return I && I->hasOneUse() && isSafeToSpeculativelyExecute(I) &&
5220 }
5221 
5222 /// Returns true if a SelectInst should be turned into an explicit branch.
5224  const TargetLowering *TLI,
5225  SelectInst *SI) {
5226  // If even a predictable select is cheap, then a branch can't be cheaper.
5227  if (!TLI->isPredictableSelectExpensive())
5228  return false;
5229 
5230  // FIXME: This should use the same heuristics as IfConversion to determine
5231  // whether a select is better represented as a branch.
5232 
5233  // If metadata tells us that the select condition is obviously predictable,
5234  // then we want to replace the select with a branch.
5235  uint64_t TrueWeight, FalseWeight;
5236  if (SI->extractProfMetadata(TrueWeight, FalseWeight)) {
5237  uint64_t Max = std::max(TrueWeight, FalseWeight);
5238  uint64_t Sum = TrueWeight + FalseWeight;
5239  if (Sum != 0) {
5240  auto Probability = BranchProbability::getBranchProbability(Max, Sum);
5241  if (Probability > TLI->getPredictableBranchThreshold())
5242  return true;
5243  }
5244  }
5245 
5246  CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
5247 
5248  // If a branch is predictable, an out-of-order CPU can avoid blocking on its
5249  // comparison condition. If the compare has more than one use, there's
5250  // probably another cmov or setcc around, so it's not worth emitting a branch.
5251  if (!Cmp || !Cmp->hasOneUse())
5252  return false;
5253 
5254  // If either operand of the select is expensive and only needed on one side
5255  // of the select, we should form a branch.
5256  if (sinkSelectOperand(TTI, SI->getTrueValue()) ||
5257  sinkSelectOperand(TTI, SI->getFalseValue()))
5258  return true;
5259 
5260  return false;
5261 }
5262 
5263 /// If \p isTrue is true, return the true value of \p SI, otherwise return
5264 /// false value of \p SI. If the true/false value of \p SI is defined by any
5265 /// select instructions in \p Selects, look through the defining select
5266 /// instruction until the true/false value is not defined in \p Selects.
5268  SelectInst *SI, bool isTrue,
5269  const SmallPtrSet<const Instruction *, 2> &Selects) {
5270  Value *V;
5271 
5272  for (SelectInst *DefSI = SI; DefSI != nullptr && Selects.count(DefSI);
5273  DefSI = dyn_cast<SelectInst>(V)) {
5274  assert(DefSI->getCondition() == SI->getCondition() &&
5275  "The condition of DefSI does not match with SI");
5276  V = (isTrue ? DefSI->getTrueValue() : DefSI->getFalseValue());
5277  }
5278  return V;
5279 }
5280 
5281 /// If we have a SelectInst that will likely profit from branch prediction,
5282 /// turn it into a branch.
5283 bool CodeGenPrepare::optimizeSelectInst(SelectInst *SI) {
5284  // Find all consecutive select instructions that share the same condition.
5286  ASI.push_back(SI);
5288  It != SI->getParent()->end(); ++It) {
5289  SelectInst *I = dyn_cast<SelectInst>(&*It);
5290  if (I && SI->getCondition() == I->getCondition()) {
5291  ASI.push_back(I);
5292  } else {
5293  break;
5294  }
5295  }
5296 
5297  SelectInst *LastSI = ASI.back();
5298  // Increment the current iterator to skip all the rest of select instructions
5299  // because they will be either "not lowered" or "all lowered" to branch.
5300  CurInstIterator = std::next(LastSI->getIterator());
5301 
5302  bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
5303 
5304  // Can we convert the 'select' to CF ?
5305  if (DisableSelectToBranch || OptSize || !TLI || VectorCond ||
5307  return false;
5308 
5310  if (VectorCond)
5311  SelectKind = TargetLowering::VectorMaskSelect;
5312  else if (SI->getType()->isVectorTy())
5314  else
5315  SelectKind = TargetLowering::ScalarValSelect;
5316 
5317  if (TLI->isSelectSupported(SelectKind) &&
5318  !isFormingBranchFromSelectProfitable(TTI, TLI, SI))
5319  return false;
5320 
5321  ModifiedDT = true;
5322 
5323  // Transform a sequence like this:
5324  // start:
5325  // %cmp = cmp uge i32 %a, %b
5326  // %sel = select i1 %cmp, i32 %c, i32 %d
5327  //
5328  // Into:
5329  // start:
5330  // %cmp = cmp uge i32 %a, %b
5331  // br i1 %cmp, label %select.true, label %select.false
5332  // select.true:
5333  // br label %select.end
5334  // select.false:
5335  // br label %select.end
5336  // select.end:
5337  // %sel = phi i32 [ %c, %select.true ], [ %d, %select.false ]
5338  //
5339  // In addition, we may sink instructions that produce %c or %d from
5340  // the entry block into the destination(s) of the new branch.
5341  // If the true or false blocks do not contain a sunken instruction, that
5342  // block and its branch may be optimized away. In that case, one side of the
5343  // first branch will point directly to select.end, and the corresponding PHI
5344  // predecessor block will be the start block.
5345 
5346  // First, we split the block containing the select into 2 blocks.
5347  BasicBlock *StartBlock = SI->getParent();
5348  BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(LastSI));
5349  BasicBlock *EndBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
5350 
5351  // Delete the unconditional branch that was just created by the split.
5352  StartBlock->getTerminator()->eraseFromParent();
5353 
5354  // These are the new basic blocks for the conditional branch.
5355  // At least one will become an actual new basic block.
5356  BasicBlock *TrueBlock = nullptr;
5357  BasicBlock *FalseBlock = nullptr;
5358  BranchInst *TrueBranch = nullptr;
5359  BranchInst *FalseBranch = nullptr;
5360 
5361  // Sink expensive instructions into the conditional blocks to avoid executing
5362  // them speculatively.
5363  for (SelectInst *SI : ASI) {
5364  if (sinkSelectOperand(TTI, SI->getTrueValue())) {
5365  if (TrueBlock == nullptr) {
5366  TrueBlock = BasicBlock::Create(SI->getContext(), "select.true.sink",
5367  EndBlock->getParent(), EndBlock);
5368  TrueBranch = BranchInst::Create(EndBlock, TrueBlock);
5369  }
5370  auto *TrueInst = cast<Instruction>(SI->getTrueValue());
5371  TrueInst->moveBefore(TrueBranch);
5372  }
5373  if (sinkSelectOperand(TTI, SI->getFalseValue())) {
5374  if (FalseBlock == nullptr) {
5375  FalseBlock = BasicBlock::Create(SI->getContext(), "select.false.sink",
5376  EndBlock->getParent(), EndBlock);
5377  FalseBranch = BranchInst::Create(EndBlock, FalseBlock);
5378  }
5379  auto *FalseInst = cast<Instruction>(SI->getFalseValue());
5380  FalseInst->moveBefore(FalseBranch);
5381  }
5382  }
5383 
5384  // If there was nothing to sink, then arbitrarily choose the 'false' side
5385  // for a new input value to the PHI.
5386  if (TrueBlock == FalseBlock) {
5387  assert(TrueBlock == nullptr &&
5388  "Unexpected basic block transform while optimizing select");
5389 
5390  FalseBlock = BasicBlock::Create(SI->getContext(), "select.false",
5391  EndBlock->getParent(), EndBlock);
5392  BranchInst::Create(EndBlock, FalseBlock);
5393  }
5394 
5395  // Insert the real conditional branch based on the original condition.
5396  // If we did not create a new block for one of the 'true' or 'false' paths
5397  // of the condition, it means that side of the branch goes to the end block
5398  // directly and the path originates from the start block from the point of
5399  // view of the new PHI.
5400  BasicBlock *TT, *FT;
5401  if (TrueBlock == nullptr) {
5402  TT = EndBlock;
5403  FT = FalseBlock;
5404  TrueBlock = StartBlock;
5405  } else if (FalseBlock == nullptr) {
5406  TT = TrueBlock;
5407  FT = EndBlock;
5408  FalseBlock = StartBlock;
5409  } else {
5410  TT = TrueBlock;
5411  FT = FalseBlock;
5412  }
5413  IRBuilder<>(SI).CreateCondBr(SI->getCondition(), TT, FT, SI);
5414 
5416  INS.insert(ASI.begin(), ASI.end());
5417  // Use reverse iterator because later select may use the value of the
5418  // earlier select, and we need to propagate value through earlier select
5419  // to get the PHI operand.
5420  for (auto It = ASI.rbegin(); It != ASI.rend(); ++It) {
5421  SelectInst *SI = *It;
5422  // The select itself is replaced with a PHI Node.
5423  PHINode *PN = PHINode::Create(SI->getType(), 2, "", &EndBlock->front());
5424  PN->takeName(SI);
5425  PN->addIncoming(getTrueOrFalseValue(SI, true, INS), TrueBlock);
5426  PN->addIncoming(getTrueOrFalseValue(SI, false, INS), FalseBlock);
5427 
5428  SI->replaceAllUsesWith(PN);
5429  SI->eraseFromParent();
5430  INS.erase(SI);
5431  ++NumSelectsExpanded;
5432  }
5433 
5434  // Instruct OptimizeBlock to skip to the next block.
5435  CurInstIterator = StartBlock->end();
5436  return true;
5437 }
5438 
5441  int SplatElem = -1;
5442  for (unsigned i = 0; i < Mask.size(); ++i) {
5443  if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
5444  return false;
5445  SplatElem = Mask[i];
5446  }
5447 
5448  return true;
5449 }
5450 
5451 /// Some targets have expensive vector shifts if the lanes aren't all the same
5452 /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
5453 /// it's often worth sinking a shufflevector splat down to its use so that
5454 /// codegen can spot all lanes are identical.
5455 bool CodeGenPrepare::optimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
5456  BasicBlock *DefBB = SVI->getParent();
5457 
5458  // Only do this xform if variable vector shifts are particularly expensive.
5459  if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
5460  return false;
5461 
5462  // We only expect better codegen by sinking a shuffle if we can recognise a
5463  // constant splat.
5464  if (!isBroadcastShuffle(SVI))
5465  return false;
5466 
5467  // InsertedShuffles - Only insert a shuffle in each block once.
5468  DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
5469 
5470  bool MadeChange = false;
5471  for (User *U : SVI->users()) {
5472  Instruction *UI = cast<Instruction>(U);
5473 
5474  // Figure out which BB this ext is used in.
5475  BasicBlock *UserBB = UI->getParent();
5476  if (UserBB == DefBB) continue;
5477 
5478  // For now only apply this when the splat is used by a shift instruction.
5479  if (!UI->isShift()) continue;
5480 
5481  // Everything checks out, sink the shuffle if the user's block doesn't
5482  // already have a copy.
5483  Instruction *&InsertedShuffle = InsertedShuffles[UserBB];
5484 
5485  if (!InsertedShuffle) {
5486  BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
5487  assert(InsertPt != UserBB->end());
5488  InsertedShuffle =
5489  new ShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1),
5490  SVI->getOperand(2), "", &*InsertPt);
5491  }
5492 
5493  UI->replaceUsesOfWith(SVI, InsertedShuffle);
5494  MadeChange = true;
5495  }
5496 
5497  // If we removed all uses, nuke the shuffle.
5498  if (SVI->use_empty()) {
5499  SVI->eraseFromParent();
5500  MadeChange = true;
5501  }
5502 
5503  return MadeChange;
5504 }
5505 
5506 bool CodeGenPrepare::optimizeSwitchInst(SwitchInst *SI) {
5507  if (!TLI || !DL)
5508  return false;
5509 
5510  Value *Cond = SI->getCondition();
5511  Type *OldType = Cond->getType();
5512  LLVMContext &Context = Cond->getContext();
5513  MVT RegType = TLI->getRegisterType(Context, TLI->getValueType(*DL, OldType));
5514  unsigned RegWidth = RegType.getSizeInBits();
5515 
5516  if (RegWidth <= cast<IntegerType>(OldType)->getBitWidth())
5517  return false;
5518 
5519  // If the register width is greater than the type width, expand the condition
5520  // of the switch instruction and each case constant to the width of the
5521  // register. By widening the type of the switch condition, subsequent
5522  // comparisons (for case comparisons) will not need to be extended to the
5523  // preferred register width, so we will potentially eliminate N-1 extends,
5524  // where N is the number of cases in the switch.
5525  auto *NewType = Type::getIntNTy(Context, RegWidth);
5526 
5527  // Zero-extend the switch condition and case constants unless the switch
5528  // condition is a function argument that is already being sign-extended.
5529  // In that case, we can avoid an unnecessary mask/extension by sign-extending
5530  // everything instead.
5531  Instruction::CastOps ExtType = Instruction::ZExt;
5532  if (auto *Arg = dyn_cast<Argument>(Cond))
5533  if (Arg->hasSExtAttr())
5534  ExtType = Instruction::SExt;
5535 
5536  auto *ExtInst = CastInst::Create(ExtType, Cond, NewType);
5537  ExtInst->insertBefore(SI);
5538  SI->setCondition(ExtInst);
5539  for (auto Case : SI->cases()) {
5540  APInt NarrowConst = Case.getCaseValue()->getValue();
5541  APInt WideConst = (ExtType == Instruction::ZExt) ?
5542  NarrowConst.zext(RegWidth) : NarrowConst.sext(RegWidth);
5543  Case.setValue(ConstantInt::get(Context, WideConst));
5544  }
5545 
5546  return true;
5547 }
5548 
5549 
5550 namespace {
5551 
5552 /// \brief Helper class to promote a scalar operation to a vector one.
5553 /// This class is used to move downward extractelement transition.
5554 /// E.g.,
5555 /// a = vector_op <2 x i32>
5556 /// b = extractelement <2 x i32> a, i32 0
5557 /// c = scalar_op b
5558 /// store c
5559 ///
5560 /// =>
5561 /// a = vector_op <2 x i32>
5562 /// c = vector_op a (equivalent to scalar_op on the related lane)
5563 /// * d = extractelement <2 x i32> c, i32 0
5564 /// * store d
5565 /// Assuming both extractelement and store can be combine, we get rid of the
5566 /// transition.
5567 class VectorPromoteHelper {
5568  /// DataLayout associated with the current module.
5569  const DataLayout &DL;
5570 
5571  /// Used to perform some checks on the legality of vector operations.
5572  const TargetLowering &TLI;
5573 
5574  /// Used to estimated the cost of the promoted chain.
5575  const TargetTransformInfo &TTI;
5576 
5577  /// The transition being moved downwards.
5578  Instruction *Transition;
5579 
5580  /// The sequence of instructions to be promoted.
5581  SmallVector<Instruction *, 4> InstsToBePromoted;
5582 
5583  /// Cost of combining a store and an extract.
5584  unsigned StoreExtractCombineCost;
5585 
5586  /// Instruction that will be combined with the transition.
5587  Instruction *CombineInst = nullptr;
5588 
5589  /// \brief The instruction that represents the current end of the transition.
5590  /// Since we are faking the promotion until we reach the end of the chain
5591  /// of computation, we need a way to get the current end of the transition.
5592  Instruction *getEndOfTransition() const {
5593  if (InstsToBePromoted.empty())
5594  return Transition;
5595  return InstsToBePromoted.back();
5596  }
5597 
5598  /// \brief Return the index of the original value in the transition.
5599  /// E.g., for "extractelement <2 x i32> c, i32 1" the original value,
5600  /// c, is at index 0.
5601  unsigned getTransitionOriginalValueIdx() const {
5602  assert(isa<ExtractElementInst>(Transition) &&
5603  "Other kind of transitions are not supported yet");
5604  return 0;
5605  }
5606 
5607  /// \brief Return the index of the index in the transition.
5608  /// E.g., for "extractelement <2 x i32> c, i32 0" the index
5609  /// is at index 1.
5610  unsigned getTransitionIdx() const {
5611  assert(isa<ExtractElementInst>(Transition) &&
5612  "Other kind of transitions are not supported yet");
5613  return 1;
5614  }
5615 
5616  /// \brief Get the type of the transition.
5617  /// This is the type of the original value.
5618  /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the
5619  /// transition is <2 x i32>.
5620  Type *getTransitionType() const {
5621  return Transition->getOperand(getTransitionOriginalValueIdx())->getType();
5622  }
5623 
5624  /// \brief Promote \p ToBePromoted by moving \p Def downward through.
5625  /// I.e., we have the following sequence:
5626  /// Def = Transition <ty1> a to <ty2>
5627  /// b = ToBePromoted <ty2> Def, ...
5628  /// =>
5629  /// b = ToBePromoted <ty1> a, ...
5630  /// Def = Transition <ty1> ToBePromoted to <ty2>
5631  void promoteImpl(Instruction *ToBePromoted);
5632 
5633  /// \brief Check whether or not it is profitable to promote all the
5634  /// instructions enqueued to be promoted.
5635  bool isProfitableToPromote() {
5636  Value *ValIdx = Transition->getOperand(getTransitionOriginalValueIdx());
5637  unsigned Index = isa<ConstantInt>(ValIdx)
5638  ? cast<ConstantInt>(ValIdx)->getZExtValue()
5639  : -1;
5640  Type *PromotedType = getTransitionType();
5641 
5642  StoreInst *ST = cast<StoreInst>(CombineInst);
5643  unsigned AS = ST->getPointerAddressSpace();
5644  unsigned Align = ST->getAlignment();
5645  // Check if this store is supported.
5647  TLI.getValueType(DL, ST->getValueOperand()->getType()), AS,
5648  Align)) {
5649  // If this is not supported, there is no way we can combine
5650  // the extract with the store.
5651  return false;
5652  }
5653 
5654  // The scalar chain of computation has to pay for the transition
5655  // scalar to vector.
5656  // The vector chain has to account for the combining cost.
5657  uint64_t ScalarCost =
5658  TTI.getVectorInstrCost(Transition->getOpcode(), PromotedType, Index);
5659  uint64_t VectorCost = StoreExtractCombineCost;
5660  for (const auto &Inst : InstsToBePromoted) {
5661  // Compute the cost.
5662  // By construction, all instructions being promoted are arithmetic ones.
5663  // Moreover, one argument is a constant that can be viewed as a splat
5664  // constant.
5665  Value *Arg0 = Inst->getOperand(0);
5666  bool IsArg0Constant = isa<UndefValue>(Arg0) || isa<ConstantInt>(Arg0) ||
5667  isa<ConstantFP>(Arg0);
5674  ScalarCost += TTI.getArithmeticInstrCost(
5675  Inst->getOpcode(), Inst->getType(), Arg0OVK, Arg1OVK);
5676  VectorCost += TTI.getArithmeticInstrCost(Inst->getOpcode(), PromotedType,
5677  Arg0OVK, Arg1OVK);
5678  }
5679  DEBUG(dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
5680  << ScalarCost << "\nVector: " << VectorCost << '\n');
5681  return ScalarCost > VectorCost;
5682  }
5683 
5684  /// \brief Generate a constant vector with \p Val with the same
5685  /// number of elements as the transition.
5686  /// \p UseSplat defines whether or not \p Val should be replicated
5687  /// across the whole vector.
5688  /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>,
5689  /// otherwise we generate a vector with as many undef as possible:
5690  /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only
5691  /// used at the index of the extract.
5692  Value *getConstantVector(Constant *Val, bool UseSplat) const {
5693  unsigned ExtractIdx = std::numeric_limits<unsigned>::max();
5694  if (!UseSplat) {
5695  // If we cannot determine where the constant must be, we have to
5696  // use a splat constant.
5697  Value *ValExtractIdx = Transition->getOperand(getTransitionIdx());
5698  if (ConstantInt *CstVal = dyn_cast<ConstantInt>(ValExtractIdx))
5699  ExtractIdx = CstVal->getSExtValue();
5700  else
5701  UseSplat = true;
5702  }
5703 
5704  unsigned End = getTransitionType()->getVectorNumElements();
5705  if (UseSplat)
5706  return ConstantVector::getSplat(End, Val);
5707 
5708  SmallVector<Constant *, 4> ConstVec;
5709  UndefValue *UndefVal = UndefValue::get(Val->getType());
5710  for (unsigned Idx = 0; Idx != End; ++Idx) {
5711  if (Idx == ExtractIdx)
5712  ConstVec.push_back(Val);
5713  else
5714  ConstVec.push_back(UndefVal);
5715  }
5716  return ConstantVector::get(ConstVec);
5717  }
5718 
5719  /// \brief Check if promoting to a vector type an operand at \p OperandIdx
5720  /// in \p Use can trigger undefined behavior.
5721  static bool canCauseUndefinedBehavior(const Instruction *Use,
5722  unsigned OperandIdx) {
5723  // This is not safe to introduce undef when the operand is on
5724  // the right hand side of a division-like instruction.
5725  if (OperandIdx != 1)
5726  return false;
5727  switch (Use->getOpcode()) {
5728  default:
5729  return false;
5730  case Instruction::SDiv:
5731  case Instruction::UDiv:
5732  case Instruction::SRem:
5733  case Instruction::URem:
5734  return true;
5735  case Instruction::FDiv:
5736  case Instruction::FRem:
5737  return !Use->hasNoNaNs();
5738  }
5739  llvm_unreachable(nullptr);
5740  }
5741 
5742 public:
5743  VectorPromoteHelper(const DataLayout &DL, const TargetLowering &TLI,
5744  const TargetTransformInfo &TTI, Instruction *Transition,
5745  unsigned CombineCost)
5746  : DL(DL), TLI(TLI), TTI(TTI), Transition(Transition),
5747  StoreExtractCombineCost(CombineCost) {
5748  assert(Transition && "Do not know how to promote null");
5749  }
5750 
5751  /// \brief Check if we can promote \p ToBePromoted to \p Type.
5752  bool canPromote(const Instruction *ToBePromoted) const {
5753  // We could support CastInst too.
5754  return isa<BinaryOperator>(ToBePromoted);
5755  }
5756 
5757  /// \brief Check if it is profitable to promote \p ToBePromoted
5758  /// by moving downward the transition through.
5759  bool shouldPromote(const Instruction *ToBePromoted) const {
5760  // Promote only if all the operands can be statically expanded.
5761  // Indeed, we do not want to introduce any new kind of transitions.
5762  for (const Use &U : ToBePromoted->operands()) {
5763  const Value *Val = U.get();
5764  if (Val == getEndOfTransition()) {
5765  // If the use is a division and the transition is on the rhs,
5766  // we cannot promote the operation, otherwise we may create a
5767  // division by zero.
5768  if (canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()))
5769  return false;
5770  continue;
5771  }
5772  if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) &&
5773  !isa<ConstantFP>(Val))
5774  return false;
5775  }
5776  // Check that the resulting operation is legal.
5777  int ISDOpcode = TLI.InstructionOpcodeToISD(ToBePromoted->getOpcode());
5778  if (!ISDOpcode)
5779  return false;
5780  return StressStoreExtract ||
5782  ISDOpcode, TLI.getValueType(DL, getTransitionType(), true));
5783  }
5784 
5785  /// \brief Check whether or not \p Use can be combined
5786  /// with the transition.
5787  /// I.e., is it possible to do Use(Transition) => AnotherUse?
5788  bool canCombine(const Instruction *Use) { return isa<StoreInst>(Use); }
5789 
5790  /// \brief Record \p ToBePromoted as part of the chain to be promoted.
5791  void enqueueForPromotion(Instruction *ToBePromoted) {
5792  InstsToBePromoted.push_back(ToBePromoted);
5793  }
5794