LLVM  9.0.0svn
CodeGenPrepare.cpp
Go to the documentation of this file.
1 //===- CodeGenPrepare.cpp - Prepare a function for code generation --------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This pass munges the code in the input function to better prepare it for
10 // SelectionDAG-based code generation. This works around limitations in it's
11 // basic-block-at-a-time approach. It should eventually be removed.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #include "llvm/ADT/APInt.h"
16 #include "llvm/ADT/ArrayRef.h"
17 #include "llvm/ADT/DenseMap.h"
18 #include "llvm/ADT/MapVector.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/Config/llvm-config.h"
43 #include "llvm/IR/Argument.h"
44 #include "llvm/IR/Attributes.h"
45 #include "llvm/IR/BasicBlock.h"
46 #include "llvm/IR/CallSite.h"
47 #include "llvm/IR/Constant.h"
48 #include "llvm/IR/Constants.h"
49 #include "llvm/IR/DataLayout.h"
50 #include "llvm/IR/DerivedTypes.h"
51 #include "llvm/IR/Dominators.h"
52 #include "llvm/IR/Function.h"
54 #include "llvm/IR/GlobalValue.h"
55 #include "llvm/IR/GlobalVariable.h"
56 #include "llvm/IR/IRBuilder.h"
57 #include "llvm/IR/InlineAsm.h"
58 #include "llvm/IR/InstrTypes.h"
59 #include "llvm/IR/Instruction.h"
60 #include "llvm/IR/Instructions.h"
61 #include "llvm/IR/IntrinsicInst.h"
62 #include "llvm/IR/Intrinsics.h"
63 #include "llvm/IR/LLVMContext.h"
64 #include "llvm/IR/MDBuilder.h"
65 #include "llvm/IR/Module.h"
66 #include "llvm/IR/Operator.h"
67 #include "llvm/IR/PatternMatch.h"
68 #include "llvm/IR/Statepoint.h"
69 #include "llvm/IR/Type.h"
70 #include "llvm/IR/Use.h"
71 #include "llvm/IR/User.h"
72 #include "llvm/IR/Value.h"
73 #include "llvm/IR/ValueHandle.h"
74 #include "llvm/IR/ValueMap.h"
75 #include "llvm/Pass.h"
78 #include "llvm/Support/Casting.h"
80 #include "llvm/Support/Compiler.h"
81 #include "llvm/Support/Debug.h"
91 #include <algorithm>
92 #include <cassert>
93 #include <cstdint>
94 #include <iterator>
95 #include <limits>
96 #include <memory>
97 #include <utility>
98 #include <vector>
99 
100 using namespace llvm;
101 using namespace llvm::PatternMatch;
102 
103 #define DEBUG_TYPE "codegenprepare"
104 
105 STATISTIC(NumBlocksElim, "Number of blocks eliminated");
106 STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated");
107 STATISTIC(NumGEPsElim, "Number of GEPs converted to casts");
108 STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "
109  "sunken Cmps");
110 STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "
111  "of sunken Casts");
112 STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "
113  "computations were sunk");
114 STATISTIC(NumMemoryInstsPhiCreated,
115  "Number of phis created when address "
116  "computations were sunk to memory instructions");
117 STATISTIC(NumMemoryInstsSelectCreated,
118  "Number of select created when address "
119  "computations were sunk to memory instructions");
120 STATISTIC(NumExtsMoved, "Number of [s|z]ext instructions combined with loads");
121 STATISTIC(NumExtUses, "Number of uses of [s|z]ext instructions optimized");
122 STATISTIC(NumAndsAdded,
123  "Number of and mask instructions added to form ext loads");
124 STATISTIC(NumAndUses, "Number of uses of and mask instructions optimized");
125 STATISTIC(NumRetsDup, "Number of return instructions duplicated");
126 STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved");
127 STATISTIC(NumSelectsExpanded, "Number of selects turned into branches");
128 STATISTIC(NumStoreExtractExposed, "Number of store(extractelement) exposed");
129 
131  "disable-cgp-branch-opts", cl::Hidden, cl::init(false),
132  cl::desc("Disable branch optimizations in CodeGenPrepare"));
133 
134 static cl::opt<bool>
135  DisableGCOpts("disable-cgp-gc-opts", cl::Hidden, cl::init(false),
136  cl::desc("Disable GC optimizations in CodeGenPrepare"));
137 
139  "disable-cgp-select2branch", cl::Hidden, cl::init(false),
140  cl::desc("Disable select to branch conversion."));
141 
143  "addr-sink-using-gep", cl::Hidden, cl::init(true),
144  cl::desc("Address sinking in CGP using GEPs."));
145 
147  "enable-andcmp-sinking", cl::Hidden, cl::init(true),
148  cl::desc("Enable sinkinig and/cmp into branches."));
149 
151  "disable-cgp-store-extract", cl::Hidden, cl::init(false),
152  cl::desc("Disable store(extract) optimizations in CodeGenPrepare"));
153 
155  "stress-cgp-store-extract", cl::Hidden, cl::init(false),
156  cl::desc("Stress test store(extract) optimizations in CodeGenPrepare"));
157 
159  "disable-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
160  cl::desc("Disable ext(promotable(ld)) -> promoted(ext(ld)) optimization in "
161  "CodeGenPrepare"));
162 
164  "stress-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
165  cl::desc("Stress test ext(promotable(ld)) -> promoted(ext(ld)) "
166  "optimization in CodeGenPrepare"));
167 
169  "disable-preheader-prot", cl::Hidden, cl::init(false),
170  cl::desc("Disable protection against removing loop preheaders"));
171 
173  "profile-guided-section-prefix", cl::Hidden, cl::init(true), cl::ZeroOrMore,
174  cl::desc("Use profile info to add section prefix for hot/cold functions"));
175 
177  "cgp-freq-ratio-to-skip-merge", cl::Hidden, cl::init(2),
178  cl::desc("Skip merging empty blocks if (frequency of empty block) / "
179  "(frequency of destination block) is greater than this ratio"));
180 
182  "force-split-store", cl::Hidden, cl::init(false),
183  cl::desc("Force store splitting no matter what the target query says."));
184 
185 static cl::opt<bool>
186 EnableTypePromotionMerge("cgp-type-promotion-merge", cl::Hidden,
187  cl::desc("Enable merging of redundant sexts when one is dominating"
188  " the other."), cl::init(true));
189 
191  "disable-complex-addr-modes", cl::Hidden, cl::init(false),
192  cl::desc("Disables combining addressing modes with different parts "
193  "in optimizeMemoryInst."));
194 
195 static cl::opt<bool>
196 AddrSinkNewPhis("addr-sink-new-phis", cl::Hidden, cl::init(false),
197  cl::desc("Allow creation of Phis in Address sinking."));
198 
199 static cl::opt<bool>
200 AddrSinkNewSelects("addr-sink-new-select", cl::Hidden, cl::init(true),
201  cl::desc("Allow creation of selects in Address sinking."));
202 
204  "addr-sink-combine-base-reg", cl::Hidden, cl::init(true),
205  cl::desc("Allow combining of BaseReg field in Address sinking."));
206 
208  "addr-sink-combine-base-gv", cl::Hidden, cl::init(true),
209  cl::desc("Allow combining of BaseGV field in Address sinking."));
210 
212  "addr-sink-combine-base-offs", cl::Hidden, cl::init(true),
213  cl::desc("Allow combining of BaseOffs field in Address sinking."));
214 
216  "addr-sink-combine-scaled-reg", cl::Hidden, cl::init(true),
217  cl::desc("Allow combining of ScaledReg field in Address sinking."));
218 
219 static cl::opt<bool>
220  EnableGEPOffsetSplit("cgp-split-large-offset-gep", cl::Hidden,
221  cl::init(true),
222  cl::desc("Enable splitting large offset of GEP."));
223 
224 namespace {
225 
226 enum ExtType {
227  ZeroExtension, // Zero extension has been seen.
228  SignExtension, // Sign extension has been seen.
229  BothExtension // This extension type is used if we saw sext after
230  // ZeroExtension had been set, or if we saw zext after
231  // SignExtension had been set. It makes the type
232  // information of a promoted instruction invalid.
233 };
234 
235 using SetOfInstrs = SmallPtrSet<Instruction *, 16>;
236 using TypeIsSExt = PointerIntPair<Type *, 2, ExtType>;
237 using InstrToOrigTy = DenseMap<Instruction *, TypeIsSExt>;
238 using SExts = SmallVector<Instruction *, 16>;
239 using ValueToSExts = DenseMap<Value *, SExts>;
240 
241 class TypePromotionTransaction;
242 
243  class CodeGenPrepare : public FunctionPass {
244  const TargetMachine *TM = nullptr;
245  const TargetSubtargetInfo *SubtargetInfo;
246  const TargetLowering *TLI = nullptr;
247  const TargetRegisterInfo *TRI;
248  const TargetTransformInfo *TTI = nullptr;
249  const TargetLibraryInfo *TLInfo;
250  const LoopInfo *LI;
251  std::unique_ptr<BlockFrequencyInfo> BFI;
252  std::unique_ptr<BranchProbabilityInfo> BPI;
253 
254  /// As we scan instructions optimizing them, this is the next instruction
255  /// to optimize. Transforms that can invalidate this should update it.
256  BasicBlock::iterator CurInstIterator;
257 
258  /// Keeps track of non-local addresses that have been sunk into a block.
259  /// This allows us to avoid inserting duplicate code for blocks with
260  /// multiple load/stores of the same address. The usage of WeakTrackingVH
261  /// enables SunkAddrs to be treated as a cache whose entries can be
262  /// invalidated if a sunken address computation has been erased.
264 
265  /// Keeps track of all instructions inserted for the current function.
266  SetOfInstrs InsertedInsts;
267 
268  /// Keeps track of the type of the related instruction before their
269  /// promotion for the current function.
270  InstrToOrigTy PromotedInsts;
271 
272  /// Keep track of instructions removed during promotion.
273  SetOfInstrs RemovedInsts;
274 
275  /// Keep track of sext chains based on their initial value.
276  DenseMap<Value *, Instruction *> SeenChainsForSExt;
277 
278  /// Keep track of GEPs accessing the same data structures such as structs or
279  /// arrays that are candidates to be split later because of their large
280  /// size.
281  MapVector<
284  LargeOffsetGEPMap;
285 
286  /// Keep track of new GEP base after splitting the GEPs having large offset.
287  SmallSet<AssertingVH<Value>, 2> NewGEPBases;
288 
289  /// Map serial numbers to Large offset GEPs.
290  DenseMap<AssertingVH<GetElementPtrInst>, int> LargeOffsetGEPID;
291 
292  /// Keep track of SExt promoted.
293  ValueToSExts ValToSExtendedUses;
294 
295  /// True if optimizing for size.
296  bool OptSize;
297 
298  /// DataLayout for the Function being processed.
299  const DataLayout *DL = nullptr;
300 
301  /// Building the dominator tree can be expensive, so we only build it
302  /// lazily and update it when required.
303  std::unique_ptr<DominatorTree> DT;
304 
305  public:
306  static char ID; // Pass identification, replacement for typeid
307 
308  CodeGenPrepare() : FunctionPass(ID) {
310  }
311 
312  bool runOnFunction(Function &F) override;
313 
314  StringRef getPassName() const override { return "CodeGen Prepare"; }
315 
316  void getAnalysisUsage(AnalysisUsage &AU) const override {
317  // FIXME: When we can selectively preserve passes, preserve the domtree.
322  }
323 
324  private:
325  template <typename F>
326  void resetIteratorIfInvalidatedWhileCalling(BasicBlock *BB, F f) {
327  // Substituting can cause recursive simplifications, which can invalidate
328  // our iterator. Use a WeakTrackingVH to hold onto it in case this
329  // happens.
330  Value *CurValue = &*CurInstIterator;
331  WeakTrackingVH IterHandle(CurValue);
332 
333  f();
334 
335  // If the iterator instruction was recursively deleted, start over at the
336  // start of the block.
337  if (IterHandle != CurValue) {
338  CurInstIterator = BB->begin();
339  SunkAddrs.clear();
340  }
341  }
342 
343  // Get the DominatorTree, building if necessary.
344  DominatorTree &getDT(Function &F) {
345  if (!DT)
346  DT = llvm::make_unique<DominatorTree>(F);
347  return *DT;
348  }
349 
350  bool eliminateFallThrough(Function &F);
351  bool eliminateMostlyEmptyBlocks(Function &F);
352  BasicBlock *findDestBlockOfMergeableEmptyBlock(BasicBlock *BB);
353  bool canMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
354  void eliminateMostlyEmptyBlock(BasicBlock *BB);
355  bool isMergingEmptyBlockProfitable(BasicBlock *BB, BasicBlock *DestBB,
356  bool isPreheader);
357  bool optimizeBlock(BasicBlock &BB, bool &ModifiedDT);
358  bool optimizeInst(Instruction *I, bool &ModifiedDT);
359  bool optimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
360  Type *AccessTy, unsigned AddrSpace);
361  bool optimizeInlineAsmInst(CallInst *CS);
362  bool optimizeCallInst(CallInst *CI, bool &ModifiedDT);
363  bool optimizeExt(Instruction *&I);
364  bool optimizeExtUses(Instruction *I);
365  bool optimizeLoadExt(LoadInst *Load);
366  bool optimizeSelectInst(SelectInst *SI);
367  bool optimizeShuffleVectorInst(ShuffleVectorInst *SVI);
368  bool optimizeSwitchInst(SwitchInst *SI);
369  bool optimizeExtractElementInst(Instruction *Inst);
370  bool dupRetToEnableTailCallOpts(BasicBlock *BB, bool &ModifiedDT);
371  bool placeDbgValues(Function &F);
372  bool canFormExtLd(const SmallVectorImpl<Instruction *> &MovedExts,
373  LoadInst *&LI, Instruction *&Inst, bool HasPromoted);
374  bool tryToPromoteExts(TypePromotionTransaction &TPT,
375  const SmallVectorImpl<Instruction *> &Exts,
376  SmallVectorImpl<Instruction *> &ProfitablyMovedExts,
377  unsigned CreatedInstsCost = 0);
378  bool mergeSExts(Function &F);
379  bool splitLargeGEPOffsets();
380  bool performAddressTypePromotion(
381  Instruction *&Inst,
382  bool AllowPromotionWithoutCommonHeader,
383  bool HasPromoted, TypePromotionTransaction &TPT,
384  SmallVectorImpl<Instruction *> &SpeculativelyMovedExts);
385  bool splitBranchCondition(Function &F, bool &ModifiedDT);
386  bool simplifyOffsetableRelocate(Instruction &I);
387 
388  bool tryToSinkFreeOperands(Instruction *I);
389  bool replaceMathCmpWithIntrinsic(BinaryOperator *BO, CmpInst *Cmp,
390  Intrinsic::ID IID);
391  bool optimizeCmp(CmpInst *Cmp, bool &ModifiedDT);
392  bool combineToUSubWithOverflow(CmpInst *Cmp, bool &ModifiedDT);
393  bool combineToUAddWithOverflow(CmpInst *Cmp, bool &ModifiedDT);
394  };
395 
396 } // end anonymous namespace
397 
398 char CodeGenPrepare::ID = 0;
399 
400 INITIALIZE_PASS_BEGIN(CodeGenPrepare, DEBUG_TYPE,
401  "Optimize for code generation", false, false)
404  "Optimize for code generation", false, false)
405 
406 FunctionPass *llvm::createCodeGenPreparePass() { return new CodeGenPrepare(); }
407 
409  if (skipFunction(F))
410  return false;
411 
412  DL = &F.getParent()->getDataLayout();
413 
414  bool EverMadeChange = false;
415  // Clear per function information.
416  InsertedInsts.clear();
417  PromotedInsts.clear();
418 
419  if (auto *TPC = getAnalysisIfAvailable<TargetPassConfig>()) {
420  TM = &TPC->getTM<TargetMachine>();
421  SubtargetInfo = TM->getSubtargetImpl(F);
422  TLI = SubtargetInfo->getTargetLowering();
423  TRI = SubtargetInfo->getRegisterInfo();
424  }
425  TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
426  TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
427  LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
428  BPI.reset(new BranchProbabilityInfo(F, *LI));
429  BFI.reset(new BlockFrequencyInfo(F, *BPI, *LI));
430  OptSize = F.hasOptSize();
431 
432  ProfileSummaryInfo *PSI =
433  &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
435  if (PSI->isFunctionHotInCallGraph(&F, *BFI))
436  F.setSectionPrefix(".hot");
437  else if (PSI->isFunctionColdInCallGraph(&F, *BFI))
438  F.setSectionPrefix(".unlikely");
439  }
440 
441  /// This optimization identifies DIV instructions that can be
442  /// profitably bypassed and carried out with a shorter, faster divide.
443  if (!OptSize && !PSI->hasHugeWorkingSetSize() && TLI &&
444  TLI->isSlowDivBypassed()) {
445  const DenseMap<unsigned int, unsigned int> &BypassWidths =
446  TLI->getBypassSlowDivWidths();
447  BasicBlock* BB = &*F.begin();
448  while (BB != nullptr) {
449  // bypassSlowDivision may create new BBs, but we don't want to reapply the
450  // optimization to those blocks.
451  BasicBlock* Next = BB->getNextNode();
452  EverMadeChange |= bypassSlowDivision(BB, BypassWidths);
453  BB = Next;
454  }
455  }
456 
457  // Eliminate blocks that contain only PHI nodes and an
458  // unconditional branch.
459  EverMadeChange |= eliminateMostlyEmptyBlocks(F);
460 
461  bool ModifiedDT = false;
462  if (!DisableBranchOpts)
463  EverMadeChange |= splitBranchCondition(F, ModifiedDT);
464 
465  // Split some critical edges where one of the sources is an indirect branch,
466  // to help generate sane code for PHIs involving such edges.
467  EverMadeChange |= SplitIndirectBrCriticalEdges(F);
468 
469  bool MadeChange = true;
470  while (MadeChange) {
471  MadeChange = false;
472  DT.reset();
473  for (Function::iterator I = F.begin(); I != F.end(); ) {
474  BasicBlock *BB = &*I++;
475  bool ModifiedDTOnIteration = false;
476  MadeChange |= optimizeBlock(*BB, ModifiedDTOnIteration);
477 
478  // Restart BB iteration if the dominator tree of the Function was changed
479  if (ModifiedDTOnIteration)
480  break;
481  }
482  if (EnableTypePromotionMerge && !ValToSExtendedUses.empty())
483  MadeChange |= mergeSExts(F);
484  if (!LargeOffsetGEPMap.empty())
485  MadeChange |= splitLargeGEPOffsets();
486 
487  // Really free removed instructions during promotion.
488  for (Instruction *I : RemovedInsts)
489  I->deleteValue();
490 
491  EverMadeChange |= MadeChange;
492  SeenChainsForSExt.clear();
493  ValToSExtendedUses.clear();
494  RemovedInsts.clear();
495  LargeOffsetGEPMap.clear();
496  LargeOffsetGEPID.clear();
497  }
498 
499  SunkAddrs.clear();
500 
501  if (!DisableBranchOpts) {
502  MadeChange = false;
503  // Use a set vector to get deterministic iteration order. The order the
504  // blocks are removed may affect whether or not PHI nodes in successors
505  // are removed.
507  for (BasicBlock &BB : F) {
508  SmallVector<BasicBlock *, 2> Successors(succ_begin(&BB), succ_end(&BB));
509  MadeChange |= ConstantFoldTerminator(&BB, true);
510  if (!MadeChange) continue;
511 
513  II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
514  if (pred_begin(*II) == pred_end(*II))
515  WorkList.insert(*II);
516  }
517 
518  // Delete the dead blocks and any of their dead successors.
519  MadeChange |= !WorkList.empty();
520  while (!WorkList.empty()) {
521  BasicBlock *BB = WorkList.pop_back_val();
522  SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
523 
524  DeleteDeadBlock(BB);
525 
527  II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
528  if (pred_begin(*II) == pred_end(*II))
529  WorkList.insert(*II);
530  }
531 
532  // Merge pairs of basic blocks with unconditional branches, connected by
533  // a single edge.
534  if (EverMadeChange || MadeChange)
535  MadeChange |= eliminateFallThrough(F);
536 
537  EverMadeChange |= MadeChange;
538  }
539 
540  if (!DisableGCOpts) {
541  SmallVector<Instruction *, 2> Statepoints;
542  for (BasicBlock &BB : F)
543  for (Instruction &I : BB)
544  if (isStatepoint(I))
545  Statepoints.push_back(&I);
546  for (auto &I : Statepoints)
547  EverMadeChange |= simplifyOffsetableRelocate(*I);
548  }
549 
550  // Do this last to clean up use-before-def scenarios introduced by other
551  // preparatory transforms.
552  EverMadeChange |= placeDbgValues(F);
553 
554  return EverMadeChange;
555 }
556 
557 /// Merge basic blocks which are connected by a single edge, where one of the
558 /// basic blocks has a single successor pointing to the other basic block,
559 /// which has a single predecessor.
560 bool CodeGenPrepare::eliminateFallThrough(Function &F) {
561  bool Changed = false;
562  // Scan all of the blocks in the function, except for the entry block.
563  // Use a temporary array to avoid iterator being invalidated when
564  // deleting blocks.
566  for (auto &Block : llvm::make_range(std::next(F.begin()), F.end()))
567  Blocks.push_back(&Block);
568 
569  for (auto &Block : Blocks) {
570  auto *BB = cast_or_null<BasicBlock>(Block);
571  if (!BB)
572  continue;
573  // If the destination block has a single pred, then this is a trivial
574  // edge, just collapse it.
575  BasicBlock *SinglePred = BB->getSinglePredecessor();
576 
577  // Don't merge if BB's address is taken.
578  if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
579 
580  BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
581  if (Term && !Term->isConditional()) {
582  Changed = true;
583  LLVM_DEBUG(dbgs() << "To merge:\n" << *BB << "\n\n\n");
584 
585  // Merge BB into SinglePred and delete it.
587  }
588  }
589  return Changed;
590 }
591 
592 /// Find a destination block from BB if BB is mergeable empty block.
593 BasicBlock *CodeGenPrepare::findDestBlockOfMergeableEmptyBlock(BasicBlock *BB) {
594  // If this block doesn't end with an uncond branch, ignore it.
596  if (!BI || !BI->isUnconditional())
597  return nullptr;
598 
599  // If the instruction before the branch (skipping debug info) isn't a phi
600  // node, then other stuff is happening here.
601  BasicBlock::iterator BBI = BI->getIterator();
602  if (BBI != BB->begin()) {
603  --BBI;
604  while (isa<DbgInfoIntrinsic>(BBI)) {
605  if (BBI == BB->begin())
606  break;
607  --BBI;
608  }
609  if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
610  return nullptr;
611  }
612 
613  // Do not break infinite loops.
614  BasicBlock *DestBB = BI->getSuccessor(0);
615  if (DestBB == BB)
616  return nullptr;
617 
618  if (!canMergeBlocks(BB, DestBB))
619  DestBB = nullptr;
620 
621  return DestBB;
622 }
623 
624 /// Eliminate blocks that contain only PHI nodes, debug info directives, and an
625 /// unconditional branch. Passes before isel (e.g. LSR/loopsimplify) often split
626 /// edges in ways that are non-optimal for isel. Start by eliminating these
627 /// blocks so we can split them the way we want them.
628 bool CodeGenPrepare::eliminateMostlyEmptyBlocks(Function &F) {
630  SmallVector<Loop *, 16> LoopList(LI->begin(), LI->end());
631  while (!LoopList.empty()) {
632  Loop *L = LoopList.pop_back_val();
633  LoopList.insert(LoopList.end(), L->begin(), L->end());
634  if (BasicBlock *Preheader = L->getLoopPreheader())
635  Preheaders.insert(Preheader);
636  }
637 
638  bool MadeChange = false;
639  // Copy blocks into a temporary array to avoid iterator invalidation issues
640  // as we remove them.
641  // Note that this intentionally skips the entry block.
643  for (auto &Block : llvm::make_range(std::next(F.begin()), F.end()))
644  Blocks.push_back(&Block);
645 
646  for (auto &Block : Blocks) {
647  BasicBlock *BB = cast_or_null<BasicBlock>(Block);
648  if (!BB)
649  continue;
650  BasicBlock *DestBB = findDestBlockOfMergeableEmptyBlock(BB);
651  if (!DestBB ||
652  !isMergingEmptyBlockProfitable(BB, DestBB, Preheaders.count(BB)))
653  continue;
654 
655  eliminateMostlyEmptyBlock(BB);
656  MadeChange = true;
657  }
658  return MadeChange;
659 }
660 
661 bool CodeGenPrepare::isMergingEmptyBlockProfitable(BasicBlock *BB,
662  BasicBlock *DestBB,
663  bool isPreheader) {
664  // Do not delete loop preheaders if doing so would create a critical edge.
665  // Loop preheaders can be good locations to spill registers. If the
666  // preheader is deleted and we create a critical edge, registers may be
667  // spilled in the loop body instead.
668  if (!DisablePreheaderProtect && isPreheader &&
669  !(BB->getSinglePredecessor() &&
671  return false;
672 
673  // Skip merging if the block's successor is also a successor to any callbr
674  // that leads to this block.
675  // FIXME: Is this really needed? Is this a correctness issue?
676  for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
677  if (auto *CBI = dyn_cast<CallBrInst>((*PI)->getTerminator()))
678  for (unsigned i = 0, e = CBI->getNumSuccessors(); i != e; ++i)
679  if (DestBB == CBI->getSuccessor(i))
680  return false;
681  }
682 
683  // Try to skip merging if the unique predecessor of BB is terminated by a
684  // switch or indirect branch instruction, and BB is used as an incoming block
685  // of PHIs in DestBB. In such case, merging BB and DestBB would cause ISel to
686  // add COPY instructions in the predecessor of BB instead of BB (if it is not
687  // merged). Note that the critical edge created by merging such blocks wont be
688  // split in MachineSink because the jump table is not analyzable. By keeping
689  // such empty block (BB), ISel will place COPY instructions in BB, not in the
690  // predecessor of BB.
691  BasicBlock *Pred = BB->getUniquePredecessor();
692  if (!Pred ||
693  !(isa<SwitchInst>(Pred->getTerminator()) ||
694  isa<IndirectBrInst>(Pred->getTerminator())))
695  return true;
696 
697  if (BB->getTerminator() != BB->getFirstNonPHIOrDbg())
698  return true;
699 
700  // We use a simple cost heuristic which determine skipping merging is
701  // profitable if the cost of skipping merging is less than the cost of
702  // merging : Cost(skipping merging) < Cost(merging BB), where the
703  // Cost(skipping merging) is Freq(BB) * (Cost(Copy) + Cost(Branch)), and
704  // the Cost(merging BB) is Freq(Pred) * Cost(Copy).
705  // Assuming Cost(Copy) == Cost(Branch), we could simplify it to :
706  // Freq(Pred) / Freq(BB) > 2.
707  // Note that if there are multiple empty blocks sharing the same incoming
708  // value for the PHIs in the DestBB, we consider them together. In such
709  // case, Cost(merging BB) will be the sum of their frequencies.
710 
711  if (!isa<PHINode>(DestBB->begin()))
712  return true;
713 
714  SmallPtrSet<BasicBlock *, 16> SameIncomingValueBBs;
715 
716  // Find all other incoming blocks from which incoming values of all PHIs in
717  // DestBB are the same as the ones from BB.
718  for (pred_iterator PI = pred_begin(DestBB), E = pred_end(DestBB); PI != E;
719  ++PI) {
720  BasicBlock *DestBBPred = *PI;
721  if (DestBBPred == BB)
722  continue;
723 
724  if (llvm::all_of(DestBB->phis(), [&](const PHINode &DestPN) {
725  return DestPN.getIncomingValueForBlock(BB) ==
726  DestPN.getIncomingValueForBlock(DestBBPred);
727  }))
728  SameIncomingValueBBs.insert(DestBBPred);
729  }
730 
731  // See if all BB's incoming values are same as the value from Pred. In this
732  // case, no reason to skip merging because COPYs are expected to be place in
733  // Pred already.
734  if (SameIncomingValueBBs.count(Pred))
735  return true;
736 
737  BlockFrequency PredFreq = BFI->getBlockFreq(Pred);
738  BlockFrequency BBFreq = BFI->getBlockFreq(BB);
739 
740  for (auto SameValueBB : SameIncomingValueBBs)
741  if (SameValueBB->getUniquePredecessor() == Pred &&
742  DestBB == findDestBlockOfMergeableEmptyBlock(SameValueBB))
743  BBFreq += BFI->getBlockFreq(SameValueBB);
744 
745  return PredFreq.getFrequency() <=
747 }
748 
749 /// Return true if we can merge BB into DestBB if there is a single
750 /// unconditional branch between them, and BB contains no other non-phi
751 /// instructions.
752 bool CodeGenPrepare::canMergeBlocks(const BasicBlock *BB,
753  const BasicBlock *DestBB) const {
754  // We only want to eliminate blocks whose phi nodes are used by phi nodes in
755  // the successor. If there are more complex condition (e.g. preheaders),
756  // don't mess around with them.
757  for (const PHINode &PN : BB->phis()) {
758  for (const User *U : PN.users()) {
759  const Instruction *UI = cast<Instruction>(U);
760  if (UI->getParent() != DestBB || !isa<PHINode>(UI))
761  return false;
762  // If User is inside DestBB block and it is a PHINode then check
763  // incoming value. If incoming value is not from BB then this is
764  // a complex condition (e.g. preheaders) we want to avoid here.
765  if (UI->getParent() == DestBB) {
766  if (const PHINode *UPN = dyn_cast<PHINode>(UI))
767  for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
768  Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
769  if (Insn && Insn->getParent() == BB &&
770  Insn->getParent() != UPN->getIncomingBlock(I))
771  return false;
772  }
773  }
774  }
775  }
776 
777  // If BB and DestBB contain any common predecessors, then the phi nodes in BB
778  // and DestBB may have conflicting incoming values for the block. If so, we
779  // can't merge the block.
780  const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
781  if (!DestBBPN) return true; // no conflict.
782 
783  // Collect the preds of BB.
785  if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
786  // It is faster to get preds from a PHI than with pred_iterator.
787  for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
788  BBPreds.insert(BBPN->getIncomingBlock(i));
789  } else {
790  BBPreds.insert(pred_begin(BB), pred_end(BB));
791  }
792 
793  // Walk the preds of DestBB.
794  for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
795  BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
796  if (BBPreds.count(Pred)) { // Common predecessor?
797  for (const PHINode &PN : DestBB->phis()) {
798  const Value *V1 = PN.getIncomingValueForBlock(Pred);
799  const Value *V2 = PN.getIncomingValueForBlock(BB);
800 
801  // If V2 is a phi node in BB, look up what the mapped value will be.
802  if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
803  if (V2PN->getParent() == BB)
804  V2 = V2PN->getIncomingValueForBlock(Pred);
805 
806  // If there is a conflict, bail out.
807  if (V1 != V2) return false;
808  }
809  }
810  }
811 
812  return true;
813 }
814 
815 /// Eliminate a basic block that has only phi's and an unconditional branch in
816 /// it.
817 void CodeGenPrepare::eliminateMostlyEmptyBlock(BasicBlock *BB) {
818  BranchInst *BI = cast<BranchInst>(BB->getTerminator());
819  BasicBlock *DestBB = BI->getSuccessor(0);
820 
821  LLVM_DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n"
822  << *BB << *DestBB);
823 
824  // If the destination block has a single pred, then this is a trivial edge,
825  // just collapse it.
826  if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
827  if (SinglePred != DestBB) {
828  assert(SinglePred == BB &&
829  "Single predecessor not the same as predecessor");
830  // Merge DestBB into SinglePred/BB and delete it.
832  // Note: BB(=SinglePred) will not be deleted on this path.
833  // DestBB(=its single successor) is the one that was deleted.
834  LLVM_DEBUG(dbgs() << "AFTER:\n" << *SinglePred << "\n\n\n");
835  return;
836  }
837  }
838 
839  // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
840  // to handle the new incoming edges it is about to have.
841  for (PHINode &PN : DestBB->phis()) {
842  // Remove the incoming value for BB, and remember it.
843  Value *InVal = PN.removeIncomingValue(BB, false);
844 
845  // Two options: either the InVal is a phi node defined in BB or it is some
846  // value that dominates BB.
847  PHINode *InValPhi = dyn_cast<PHINode>(InVal);
848  if (InValPhi && InValPhi->getParent() == BB) {
849  // Add all of the input values of the input PHI as inputs of this phi.
850  for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
851  PN.addIncoming(InValPhi->getIncomingValue(i),
852  InValPhi->getIncomingBlock(i));
853  } else {
854  // Otherwise, add one instance of the dominating value for each edge that
855  // we will be adding.
856  if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
857  for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
858  PN.addIncoming(InVal, BBPN->getIncomingBlock(i));
859  } else {
860  for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
861  PN.addIncoming(InVal, *PI);
862  }
863  }
864  }
865 
866  // The PHIs are now updated, change everything that refers to BB to use
867  // DestBB and remove BB.
868  BB->replaceAllUsesWith(DestBB);
869  BB->eraseFromParent();
870  ++NumBlocksElim;
871 
872  LLVM_DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
873 }
874 
875 // Computes a map of base pointer relocation instructions to corresponding
876 // derived pointer relocation instructions given a vector of all relocate calls
878  const SmallVectorImpl<GCRelocateInst *> &AllRelocateCalls,
880  &RelocateInstMap) {
881  // Collect information in two maps: one primarily for locating the base object
882  // while filling the second map; the second map is the final structure holding
883  // a mapping between Base and corresponding Derived relocate calls
885  for (auto *ThisRelocate : AllRelocateCalls) {
886  auto K = std::make_pair(ThisRelocate->getBasePtrIndex(),
887  ThisRelocate->getDerivedPtrIndex());
888  RelocateIdxMap.insert(std::make_pair(K, ThisRelocate));
889  }
890  for (auto &Item : RelocateIdxMap) {
891  std::pair<unsigned, unsigned> Key = Item.first;
892  if (Key.first == Key.second)
893  // Base relocation: nothing to insert
894  continue;
895 
896  GCRelocateInst *I = Item.second;
897  auto BaseKey = std::make_pair(Key.first, Key.first);
898 
899  // We're iterating over RelocateIdxMap so we cannot modify it.
900  auto MaybeBase = RelocateIdxMap.find(BaseKey);
901  if (MaybeBase == RelocateIdxMap.end())
902  // TODO: We might want to insert a new base object relocate and gep off
903  // that, if there are enough derived object relocates.
904  continue;
905 
906  RelocateInstMap[MaybeBase->second].push_back(I);
907  }
908 }
909 
910 // Accepts a GEP and extracts the operands into a vector provided they're all
911 // small integer constants
913  SmallVectorImpl<Value *> &OffsetV) {
914  for (unsigned i = 1; i < GEP->getNumOperands(); i++) {
915  // Only accept small constant integer operands
916  auto Op = dyn_cast<ConstantInt>(GEP->getOperand(i));
917  if (!Op || Op->getZExtValue() > 20)
918  return false;
919  }
920 
921  for (unsigned i = 1; i < GEP->getNumOperands(); i++)
922  OffsetV.push_back(GEP->getOperand(i));
923  return true;
924 }
925 
926 // Takes a RelocatedBase (base pointer relocation instruction) and Targets to
927 // replace, computes a replacement, and affects it.
928 static bool
930  const SmallVectorImpl<GCRelocateInst *> &Targets) {
931  bool MadeChange = false;
932  // We must ensure the relocation of derived pointer is defined after
933  // relocation of base pointer. If we find a relocation corresponding to base
934  // defined earlier than relocation of base then we move relocation of base
935  // right before found relocation. We consider only relocation in the same
936  // basic block as relocation of base. Relocations from other basic block will
937  // be skipped by optimization and we do not care about them.
938  for (auto R = RelocatedBase->getParent()->getFirstInsertionPt();
939  &*R != RelocatedBase; ++R)
940  if (auto RI = dyn_cast<GCRelocateInst>(R))
941  if (RI->getStatepoint() == RelocatedBase->getStatepoint())
942  if (RI->getBasePtrIndex() == RelocatedBase->getBasePtrIndex()) {
943  RelocatedBase->moveBefore(RI);
944  break;
945  }
946 
947  for (GCRelocateInst *ToReplace : Targets) {
948  assert(ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex() &&
949  "Not relocating a derived object of the original base object");
950  if (ToReplace->getBasePtrIndex() == ToReplace->getDerivedPtrIndex()) {
951  // A duplicate relocate call. TODO: coalesce duplicates.
952  continue;
953  }
954 
955  if (RelocatedBase->getParent() != ToReplace->getParent()) {
956  // Base and derived relocates are in different basic blocks.
957  // In this case transform is only valid when base dominates derived
958  // relocate. However it would be too expensive to check dominance
959  // for each such relocate, so we skip the whole transformation.
960  continue;
961  }
962 
963  Value *Base = ToReplace->getBasePtr();
964  auto Derived = dyn_cast<GetElementPtrInst>(ToReplace->getDerivedPtr());
965  if (!Derived || Derived->getPointerOperand() != Base)
966  continue;
967 
968  SmallVector<Value *, 2> OffsetV;
969  if (!getGEPSmallConstantIntOffsetV(Derived, OffsetV))
970  continue;
971 
972  // Create a Builder and replace the target callsite with a gep
973  assert(RelocatedBase->getNextNode() &&
974  "Should always have one since it's not a terminator");
975 
976  // Insert after RelocatedBase
977  IRBuilder<> Builder(RelocatedBase->getNextNode());
978  Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
979 
980  // If gc_relocate does not match the actual type, cast it to the right type.
981  // In theory, there must be a bitcast after gc_relocate if the type does not
982  // match, and we should reuse it to get the derived pointer. But it could be
983  // cases like this:
984  // bb1:
985  // ...
986  // %g1 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
987  // br label %merge
988  //
989  // bb2:
990  // ...
991  // %g2 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
992  // br label %merge
993  //
994  // merge:
995  // %p1 = phi i8 addrspace(1)* [ %g1, %bb1 ], [ %g2, %bb2 ]
996  // %cast = bitcast i8 addrspace(1)* %p1 in to i32 addrspace(1)*
997  //
998  // In this case, we can not find the bitcast any more. So we insert a new bitcast
999  // no matter there is already one or not. In this way, we can handle all cases, and
1000  // the extra bitcast should be optimized away in later passes.
1001  Value *ActualRelocatedBase = RelocatedBase;
1002  if (RelocatedBase->getType() != Base->getType()) {
1003  ActualRelocatedBase =
1004  Builder.CreateBitCast(RelocatedBase, Base->getType());
1005  }
1006  Value *Replacement = Builder.CreateGEP(
1007  Derived->getSourceElementType(), ActualRelocatedBase, makeArrayRef(OffsetV));
1008  Replacement->takeName(ToReplace);
1009  // If the newly generated derived pointer's type does not match the original derived
1010  // pointer's type, cast the new derived pointer to match it. Same reasoning as above.
1011  Value *ActualReplacement = Replacement;
1012  if (Replacement->getType() != ToReplace->getType()) {
1013  ActualReplacement =
1014  Builder.CreateBitCast(Replacement, ToReplace->getType());
1015  }
1016  ToReplace->replaceAllUsesWith(ActualReplacement);
1017  ToReplace->eraseFromParent();
1018 
1019  MadeChange = true;
1020  }
1021  return MadeChange;
1022 }
1023 
1024 // Turns this:
1025 //
1026 // %base = ...
1027 // %ptr = gep %base + 15
1028 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
1029 // %base' = relocate(%tok, i32 4, i32 4)
1030 // %ptr' = relocate(%tok, i32 4, i32 5)
1031 // %val = load %ptr'
1032 //
1033 // into this:
1034 //
1035 // %base = ...
1036 // %ptr = gep %base + 15
1037 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
1038 // %base' = gc.relocate(%tok, i32 4, i32 4)
1039 // %ptr' = gep %base' + 15
1040 // %val = load %ptr'
1041 bool CodeGenPrepare::simplifyOffsetableRelocate(Instruction &I) {
1042  bool MadeChange = false;
1043  SmallVector<GCRelocateInst *, 2> AllRelocateCalls;
1044 
1045  for (auto *U : I.users())
1046  if (GCRelocateInst *Relocate = dyn_cast<GCRelocateInst>(U))
1047  // Collect all the relocate calls associated with a statepoint
1048  AllRelocateCalls.push_back(Relocate);
1049 
1050  // We need atleast one base pointer relocation + one derived pointer
1051  // relocation to mangle
1052  if (AllRelocateCalls.size() < 2)
1053  return false;
1054 
1055  // RelocateInstMap is a mapping from the base relocate instruction to the
1056  // corresponding derived relocate instructions
1058  computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap);
1059  if (RelocateInstMap.empty())
1060  return false;
1061 
1062  for (auto &Item : RelocateInstMap)
1063  // Item.first is the RelocatedBase to offset against
1064  // Item.second is the vector of Targets to replace
1065  MadeChange = simplifyRelocatesOffABase(Item.first, Item.second);
1066  return MadeChange;
1067 }
1068 
1069 /// Sink the specified cast instruction into its user blocks.
1070 static bool SinkCast(CastInst *CI) {
1071  BasicBlock *DefBB = CI->getParent();
1072 
1073  /// InsertedCasts - Only insert a cast in each block once.
1074  DenseMap<BasicBlock*, CastInst*> InsertedCasts;
1075 
1076  bool MadeChange = false;
1077  for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
1078  UI != E; ) {
1079  Use &TheUse = UI.getUse();
1080  Instruction *User = cast<Instruction>(*UI);
1081 
1082  // Figure out which BB this cast is used in. For PHI's this is the
1083  // appropriate predecessor block.
1084  BasicBlock *UserBB = User->getParent();
1085  if (PHINode *PN = dyn_cast<PHINode>(User)) {
1086  UserBB = PN->getIncomingBlock(TheUse);
1087  }
1088 
1089  // Preincrement use iterator so we don't invalidate it.
1090  ++UI;
1091 
1092  // The first insertion point of a block containing an EH pad is after the
1093  // pad. If the pad is the user, we cannot sink the cast past the pad.
1094  if (User->isEHPad())
1095  continue;
1096 
1097  // If the block selected to receive the cast is an EH pad that does not
1098  // allow non-PHI instructions before the terminator, we can't sink the
1099  // cast.
1100  if (UserBB->getTerminator()->isEHPad())
1101  continue;
1102 
1103  // If this user is in the same block as the cast, don't change the cast.
1104  if (UserBB == DefBB) continue;
1105 
1106  // If we have already inserted a cast into this block, use it.
1107  CastInst *&InsertedCast = InsertedCasts[UserBB];
1108 
1109  if (!InsertedCast) {
1110  BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1111  assert(InsertPt != UserBB->end());
1112  InsertedCast = CastInst::Create(CI->getOpcode(), CI->getOperand(0),
1113  CI->getType(), "", &*InsertPt);
1114  InsertedCast->setDebugLoc(CI->getDebugLoc());
1115  }
1116 
1117  // Replace a use of the cast with a use of the new cast.
1118  TheUse = InsertedCast;
1119  MadeChange = true;
1120  ++NumCastUses;
1121  }
1122 
1123  // If we removed all uses, nuke the cast.
1124  if (CI->use_empty()) {
1125  salvageDebugInfo(*CI);
1126  CI->eraseFromParent();
1127  MadeChange = true;
1128  }
1129 
1130  return MadeChange;
1131 }
1132 
1133 /// If the specified cast instruction is a noop copy (e.g. it's casting from
1134 /// one pointer type to another, i32->i8 on PPC), sink it into user blocks to
1135 /// reduce the number of virtual registers that must be created and coalesced.
1136 ///
1137 /// Return true if any changes are made.
1139  const DataLayout &DL) {
1140  // Sink only "cheap" (or nop) address-space casts. This is a weaker condition
1141  // than sinking only nop casts, but is helpful on some platforms.
1142  if (auto *ASC = dyn_cast<AddrSpaceCastInst>(CI)) {
1143  if (!TLI.isCheapAddrSpaceCast(ASC->getSrcAddressSpace(),
1144  ASC->getDestAddressSpace()))
1145  return false;
1146  }
1147 
1148  // If this is a noop copy,
1149  EVT SrcVT = TLI.getValueType(DL, CI->getOperand(0)->getType());
1150  EVT DstVT = TLI.getValueType(DL, CI->getType());
1151 
1152  // This is an fp<->int conversion?
1153  if (SrcVT.isInteger() != DstVT.isInteger())
1154  return false;
1155 
1156  // If this is an extension, it will be a zero or sign extension, which
1157  // isn't a noop.
1158  if (SrcVT.bitsLT(DstVT)) return false;
1159 
1160  // If these values will be promoted, find out what they will be promoted
1161  // to. This helps us consider truncates on PPC as noop copies when they
1162  // are.
1163  if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
1165  SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
1166  if (TLI.getTypeAction(CI->getContext(), DstVT) ==
1168  DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
1169 
1170  // If, after promotion, these are the same types, this is a noop copy.
1171  if (SrcVT != DstVT)
1172  return false;
1173 
1174  return SinkCast(CI);
1175 }
1176 
1177 bool CodeGenPrepare::replaceMathCmpWithIntrinsic(BinaryOperator *BO,
1178  CmpInst *Cmp,
1179  Intrinsic::ID IID) {
1180  if (BO->getParent() != Cmp->getParent()) {
1181  // We used to use a dominator tree here to allow multi-block optimization.
1182  // But that was problematic because:
1183  // 1. It could cause a perf regression by hoisting the math op into the
1184  // critical path.
1185  // 2. It could cause a perf regression by creating a value that was live
1186  // across multiple blocks and increasing register pressure.
1187  // 3. Use of a dominator tree could cause large compile-time regression.
1188  // This is because we recompute the DT on every change in the main CGP
1189  // run-loop. The recomputing is probably unnecessary in many cases, so if
1190  // that was fixed, using a DT here would be ok.
1191  return false;
1192  }
1193 
1194  // We allow matching the canonical IR (add X, C) back to (usubo X, -C).
1195  Value *Arg0 = BO->getOperand(0);
1196  Value *Arg1 = BO->getOperand(1);
1197  if (BO->getOpcode() == Instruction::Add &&
1198  IID == Intrinsic::usub_with_overflow) {
1199  assert(isa<Constant>(Arg1) && "Unexpected input for usubo");
1200  Arg1 = ConstantExpr::getNeg(cast<Constant>(Arg1));
1201  }
1202 
1203  // Insert at the first instruction of the pair.
1204  Instruction *InsertPt = nullptr;
1205  for (Instruction &Iter : *Cmp->getParent()) {
1206  if (&Iter == BO || &Iter == Cmp) {
1207  InsertPt = &Iter;
1208  break;
1209  }
1210  }
1211  assert(InsertPt != nullptr && "Parent block did not contain cmp or binop");
1212 
1213  IRBuilder<> Builder(InsertPt);
1214  Value *MathOV = Builder.CreateBinaryIntrinsic(IID, Arg0, Arg1);
1215  Value *Math = Builder.CreateExtractValue(MathOV, 0, "math");
1216  Value *OV = Builder.CreateExtractValue(MathOV, 1, "ov");
1217  BO->replaceAllUsesWith(Math);
1218  Cmp->replaceAllUsesWith(OV);
1219  BO->eraseFromParent();
1220  Cmp->eraseFromParent();
1221  return true;
1222 }
1223 
1224 /// Match special-case patterns that check for unsigned add overflow.
1226  BinaryOperator *&Add) {
1227  // Add = add A, 1; Cmp = icmp eq A,-1 (overflow if A is max val)
1228  // Add = add A,-1; Cmp = icmp ne A, 0 (overflow if A is non-zero)
1229  Value *A = Cmp->getOperand(0), *B = Cmp->getOperand(1);
1230 
1231  // We are not expecting non-canonical/degenerate code. Just bail out.
1232  if (isa<Constant>(A))
1233  return false;
1234 
1235  ICmpInst::Predicate Pred = Cmp->getPredicate();
1236  if (Pred == ICmpInst::ICMP_EQ && match(B, m_AllOnes()))
1237  B = ConstantInt::get(B->getType(), 1);
1238  else if (Pred == ICmpInst::ICMP_NE && match(B, m_ZeroInt()))
1239  B = ConstantInt::get(B->getType(), -1);
1240  else
1241  return false;
1242 
1243  // Check the users of the variable operand of the compare looking for an add
1244  // with the adjusted constant.
1245  for (User *U : A->users()) {
1246  if (match(U, m_Add(m_Specific(A), m_Specific(B)))) {
1247  Add = cast<BinaryOperator>(U);
1248  return true;
1249  }
1250  }
1251  return false;
1252 }
1253 
1254 /// Try to combine the compare into a call to the llvm.uadd.with.overflow
1255 /// intrinsic. Return true if any changes were made.
1256 bool CodeGenPrepare::combineToUAddWithOverflow(CmpInst *Cmp,
1257  bool &ModifiedDT) {
1258  Value *A, *B;
1260  if (!match(Cmp, m_UAddWithOverflow(m_Value(A), m_Value(B), m_BinOp(Add))))
1262  return false;
1263 
1264  if (!TLI->shouldFormOverflowOp(ISD::UADDO,
1265  TLI->getValueType(*DL, Add->getType())))
1266  return false;
1267 
1268  // We don't want to move around uses of condition values this late, so we
1269  // check if it is legal to create the call to the intrinsic in the basic
1270  // block containing the icmp.
1271  if (Add->getParent() != Cmp->getParent() && !Add->hasOneUse())
1272  return false;
1273 
1274  if (!replaceMathCmpWithIntrinsic(Add, Cmp, Intrinsic::uadd_with_overflow))
1275  return false;
1276 
1277  // Reset callers - do not crash by iterating over a dead instruction.
1278  ModifiedDT = true;
1279  return true;
1280 }
1281 
1282 bool CodeGenPrepare::combineToUSubWithOverflow(CmpInst *Cmp,
1283  bool &ModifiedDT) {
1284  // We are not expecting non-canonical/degenerate code. Just bail out.
1285  Value *A = Cmp->getOperand(0), *B = Cmp->getOperand(1);
1286  if (isa<Constant>(A) && isa<Constant>(B))
1287  return false;
1288 
1289  // Convert (A u> B) to (A u< B) to simplify pattern matching.
1290  ICmpInst::Predicate Pred = Cmp->getPredicate();
1291  if (Pred == ICmpInst::ICMP_UGT) {
1292  std::swap(A, B);
1293  Pred = ICmpInst::ICMP_ULT;
1294  }
1295  // Convert special-case: (A == 0) is the same as (A u< 1).
1296  if (Pred == ICmpInst::ICMP_EQ && match(B, m_ZeroInt())) {
1297  B = ConstantInt::get(B->getType(), 1);
1298  Pred = ICmpInst::ICMP_ULT;
1299  }
1300  // Convert special-case: (A != 0) is the same as (0 u< A).
1301  if (Pred == ICmpInst::ICMP_NE && match(B, m_ZeroInt())) {
1302  std::swap(A, B);
1303  Pred = ICmpInst::ICMP_ULT;
1304  }
1305  if (Pred != ICmpInst::ICMP_ULT)
1306  return false;
1307 
1308  // Walk the users of a variable operand of a compare looking for a subtract or
1309  // add with that same operand. Also match the 2nd operand of the compare to
1310  // the add/sub, but that may be a negated constant operand of an add.
1311  Value *CmpVariableOperand = isa<Constant>(A) ? B : A;
1312  BinaryOperator *Sub = nullptr;
1313  for (User *U : CmpVariableOperand->users()) {
1314  // A - B, A u< B --> usubo(A, B)
1315  if (match(U, m_Sub(m_Specific(A), m_Specific(B)))) {
1316  Sub = cast<BinaryOperator>(U);
1317  break;
1318  }
1319 
1320  // A + (-C), A u< C (canonicalized form of (sub A, C))
1321  const APInt *CmpC, *AddC;
1322  if (match(U, m_Add(m_Specific(A), m_APInt(AddC))) &&
1323  match(B, m_APInt(CmpC)) && *AddC == -(*CmpC)) {
1324  Sub = cast<BinaryOperator>(U);
1325  break;
1326  }
1327  }
1328  if (!Sub)
1329  return false;
1330 
1331  if (!TLI->shouldFormOverflowOp(ISD::USUBO,
1332  TLI->getValueType(*DL, Sub->getType())))
1333  return false;
1334 
1335  if (!replaceMathCmpWithIntrinsic(Sub, Cmp, Intrinsic::usub_with_overflow))
1336  return false;
1337 
1338  // Reset callers - do not crash by iterating over a dead instruction.
1339  ModifiedDT = true;
1340  return true;
1341 }
1342 
1343 /// Sink the given CmpInst into user blocks to reduce the number of virtual
1344 /// registers that must be created and coalesced. This is a clear win except on
1345 /// targets with multiple condition code registers (PowerPC), where it might
1346 /// lose; some adjustment may be wanted there.
1347 ///
1348 /// Return true if any changes are made.
1349 static bool sinkCmpExpression(CmpInst *Cmp, const TargetLowering &TLI) {
1351  return false;
1352 
1353  // Avoid sinking soft-FP comparisons, since this can move them into a loop.
1354  if (TLI.useSoftFloat() && isa<FCmpInst>(Cmp))
1355  return false;
1356 
1357  // Only insert a cmp in each block once.
1358  DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
1359 
1360  bool MadeChange = false;
1361  for (Value::user_iterator UI = Cmp->user_begin(), E = Cmp->user_end();
1362  UI != E; ) {
1363  Use &TheUse = UI.getUse();
1364  Instruction *User = cast<Instruction>(*UI);
1365 
1366  // Preincrement use iterator so we don't invalidate it.
1367  ++UI;
1368 
1369  // Don't bother for PHI nodes.
1370  if (isa<PHINode>(User))
1371  continue;
1372 
1373  // Figure out which BB this cmp is used in.
1374  BasicBlock *UserBB = User->getParent();
1375  BasicBlock *DefBB = Cmp->getParent();
1376 
1377  // If this user is in the same block as the cmp, don't change the cmp.
1378  if (UserBB == DefBB) continue;
1379 
1380  // If we have already inserted a cmp into this block, use it.
1381  CmpInst *&InsertedCmp = InsertedCmps[UserBB];
1382 
1383  if (!InsertedCmp) {
1384  BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1385  assert(InsertPt != UserBB->end());
1386  InsertedCmp =
1387  CmpInst::Create(Cmp->getOpcode(), Cmp->getPredicate(),
1388  Cmp->getOperand(0), Cmp->getOperand(1), "",
1389  &*InsertPt);
1390  // Propagate the debug info.
1391  InsertedCmp->setDebugLoc(Cmp->getDebugLoc());
1392  }
1393 
1394  // Replace a use of the cmp with a use of the new cmp.
1395  TheUse = InsertedCmp;
1396  MadeChange = true;
1397  ++NumCmpUses;
1398  }
1399 
1400  // If we removed all uses, nuke the cmp.
1401  if (Cmp->use_empty()) {
1402  Cmp->eraseFromParent();
1403  MadeChange = true;
1404  }
1405 
1406  return MadeChange;
1407 }
1408 
1409 bool CodeGenPrepare::optimizeCmp(CmpInst *Cmp, bool &ModifiedDT) {
1410  if (sinkCmpExpression(Cmp, *TLI))
1411  return true;
1412 
1413  if (combineToUAddWithOverflow(Cmp, ModifiedDT))
1414  return true;
1415 
1416  if (combineToUSubWithOverflow(Cmp, ModifiedDT))
1417  return true;
1418 
1419  return false;
1420 }
1421 
1422 /// Duplicate and sink the given 'and' instruction into user blocks where it is
1423 /// used in a compare to allow isel to generate better code for targets where
1424 /// this operation can be combined.
1425 ///
1426 /// Return true if any changes are made.
1428  const TargetLowering &TLI,
1429  SetOfInstrs &InsertedInsts) {
1430  // Double-check that we're not trying to optimize an instruction that was
1431  // already optimized by some other part of this pass.
1432  assert(!InsertedInsts.count(AndI) &&
1433  "Attempting to optimize already optimized and instruction");
1434  (void) InsertedInsts;
1435 
1436  // Nothing to do for single use in same basic block.
1437  if (AndI->hasOneUse() &&
1438  AndI->getParent() == cast<Instruction>(*AndI->user_begin())->getParent())
1439  return false;
1440 
1441  // Try to avoid cases where sinking/duplicating is likely to increase register
1442  // pressure.
1443  if (!isa<ConstantInt>(AndI->getOperand(0)) &&
1444  !isa<ConstantInt>(AndI->getOperand(1)) &&
1445  AndI->getOperand(0)->hasOneUse() && AndI->getOperand(1)->hasOneUse())
1446  return false;
1447 
1448  for (auto *U : AndI->users()) {
1449  Instruction *User = cast<Instruction>(U);
1450 
1451  // Only sink 'and' feeding icmp with 0.
1452  if (!isa<ICmpInst>(User))
1453  return false;
1454 
1455  auto *CmpC = dyn_cast<ConstantInt>(User->getOperand(1));
1456  if (!CmpC || !CmpC->isZero())
1457  return false;
1458  }
1459 
1460  if (!TLI.isMaskAndCmp0FoldingBeneficial(*AndI))
1461  return false;
1462 
1463  LLVM_DEBUG(dbgs() << "found 'and' feeding only icmp 0;\n");
1464  LLVM_DEBUG(AndI->getParent()->dump());
1465 
1466  // Push the 'and' into the same block as the icmp 0. There should only be
1467  // one (icmp (and, 0)) in each block, since CSE/GVN should have removed any
1468  // others, so we don't need to keep track of which BBs we insert into.
1469  for (Value::user_iterator UI = AndI->user_begin(), E = AndI->user_end();
1470  UI != E; ) {
1471  Use &TheUse = UI.getUse();
1472  Instruction *User = cast<Instruction>(*UI);
1473 
1474  // Preincrement use iterator so we don't invalidate it.
1475  ++UI;
1476 
1477  LLVM_DEBUG(dbgs() << "sinking 'and' use: " << *User << "\n");
1478 
1479  // Keep the 'and' in the same place if the use is already in the same block.
1480  Instruction *InsertPt =
1481  User->getParent() == AndI->getParent() ? AndI : User;
1482  Instruction *InsertedAnd =
1483  BinaryOperator::Create(Instruction::And, AndI->getOperand(0),
1484  AndI->getOperand(1), "", InsertPt);
1485  // Propagate the debug info.
1486  InsertedAnd->setDebugLoc(AndI->getDebugLoc());
1487 
1488  // Replace a use of the 'and' with a use of the new 'and'.
1489  TheUse = InsertedAnd;
1490  ++NumAndUses;
1491  LLVM_DEBUG(User->getParent()->dump());
1492  }
1493 
1494  // We removed all uses, nuke the and.
1495  AndI->eraseFromParent();
1496  return true;
1497 }
1498 
1499 /// Check if the candidates could be combined with a shift instruction, which
1500 /// includes:
1501 /// 1. Truncate instruction
1502 /// 2. And instruction and the imm is a mask of the low bits:
1503 /// imm & (imm+1) == 0
1505  if (!isa<TruncInst>(User)) {
1506  if (User->getOpcode() != Instruction::And ||
1507  !isa<ConstantInt>(User->getOperand(1)))
1508  return false;
1509 
1510  const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
1511 
1512  if ((Cimm & (Cimm + 1)).getBoolValue())
1513  return false;
1514  }
1515  return true;
1516 }
1517 
1518 /// Sink both shift and truncate instruction to the use of truncate's BB.
1519 static bool
1522  const TargetLowering &TLI, const DataLayout &DL) {
1523  BasicBlock *UserBB = User->getParent();
1524  DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
1525  TruncInst *TruncI = dyn_cast<TruncInst>(User);
1526  bool MadeChange = false;
1527 
1528  for (Value::user_iterator TruncUI = TruncI->user_begin(),
1529  TruncE = TruncI->user_end();
1530  TruncUI != TruncE;) {
1531 
1532  Use &TruncTheUse = TruncUI.getUse();
1533  Instruction *TruncUser = cast<Instruction>(*TruncUI);
1534  // Preincrement use iterator so we don't invalidate it.
1535 
1536  ++TruncUI;
1537 
1538  int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
1539  if (!ISDOpcode)
1540  continue;
1541 
1542  // If the use is actually a legal node, there will not be an
1543  // implicit truncate.
1544  // FIXME: always querying the result type is just an
1545  // approximation; some nodes' legality is determined by the
1546  // operand or other means. There's no good way to find out though.
1547  if (TLI.isOperationLegalOrCustom(
1548  ISDOpcode, TLI.getValueType(DL, TruncUser->getType(), true)))
1549  continue;
1550 
1551  // Don't bother for PHI nodes.
1552  if (isa<PHINode>(TruncUser))
1553  continue;
1554 
1555  BasicBlock *TruncUserBB = TruncUser->getParent();
1556 
1557  if (UserBB == TruncUserBB)
1558  continue;
1559 
1560  BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
1561  CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
1562 
1563  if (!InsertedShift && !InsertedTrunc) {
1564  BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
1565  assert(InsertPt != TruncUserBB->end());
1566  // Sink the shift
1567  if (ShiftI->getOpcode() == Instruction::AShr)
1568  InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
1569  "", &*InsertPt);
1570  else
1571  InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
1572  "", &*InsertPt);
1573  InsertedShift->setDebugLoc(ShiftI->getDebugLoc());
1574 
1575  // Sink the trunc
1576  BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
1577  TruncInsertPt++;
1578  assert(TruncInsertPt != TruncUserBB->end());
1579 
1580  InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
1581  TruncI->getType(), "", &*TruncInsertPt);
1582  InsertedTrunc->setDebugLoc(TruncI->getDebugLoc());
1583 
1584  MadeChange = true;
1585 
1586  TruncTheUse = InsertedTrunc;
1587  }
1588  }
1589  return MadeChange;
1590 }
1591 
1592 /// Sink the shift *right* instruction into user blocks if the uses could
1593 /// potentially be combined with this shift instruction and generate BitExtract
1594 /// instruction. It will only be applied if the architecture supports BitExtract
1595 /// instruction. Here is an example:
1596 /// BB1:
1597 /// %x.extract.shift = lshr i64 %arg1, 32
1598 /// BB2:
1599 /// %x.extract.trunc = trunc i64 %x.extract.shift to i16
1600 /// ==>
1601 ///
1602 /// BB2:
1603 /// %x.extract.shift.1 = lshr i64 %arg1, 32
1604 /// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
1605 ///
1606 /// CodeGen will recognize the pattern in BB2 and generate BitExtract
1607 /// instruction.
1608 /// Return true if any changes are made.
1610  const TargetLowering &TLI,
1611  const DataLayout &DL) {
1612  BasicBlock *DefBB = ShiftI->getParent();
1613 
1614  /// Only insert instructions in each block once.
1616 
1617  bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(DL, ShiftI->getType()));
1618 
1619  bool MadeChange = false;
1620  for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
1621  UI != E;) {
1622  Use &TheUse = UI.getUse();
1623  Instruction *User = cast<Instruction>(*UI);
1624  // Preincrement use iterator so we don't invalidate it.
1625  ++UI;
1626 
1627  // Don't bother for PHI nodes.
1628  if (isa<PHINode>(User))
1629  continue;
1630 
1631  if (!isExtractBitsCandidateUse(User))
1632  continue;
1633 
1634  BasicBlock *UserBB = User->getParent();
1635 
1636  if (UserBB == DefBB) {
1637  // If the shift and truncate instruction are in the same BB. The use of
1638  // the truncate(TruncUse) may still introduce another truncate if not
1639  // legal. In this case, we would like to sink both shift and truncate
1640  // instruction to the BB of TruncUse.
1641  // for example:
1642  // BB1:
1643  // i64 shift.result = lshr i64 opnd, imm
1644  // trunc.result = trunc shift.result to i16
1645  //
1646  // BB2:
1647  // ----> We will have an implicit truncate here if the architecture does
1648  // not have i16 compare.
1649  // cmp i16 trunc.result, opnd2
1650  //
1651  if (isa<TruncInst>(User) && shiftIsLegal
1652  // If the type of the truncate is legal, no truncate will be
1653  // introduced in other basic blocks.
1654  &&
1655  (!TLI.isTypeLegal(TLI.getValueType(DL, User->getType()))))
1656  MadeChange =
1657  SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI, DL);
1658 
1659  continue;
1660  }
1661  // If we have already inserted a shift into this block, use it.
1662  BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
1663 
1664  if (!InsertedShift) {
1665  BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1666  assert(InsertPt != UserBB->end());
1667 
1668  if (ShiftI->getOpcode() == Instruction::AShr)
1669  InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
1670  "", &*InsertPt);
1671  else
1672  InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
1673  "", &*InsertPt);
1674  InsertedShift->setDebugLoc(ShiftI->getDebugLoc());
1675 
1676  MadeChange = true;
1677  }
1678 
1679  // Replace a use of the shift with a use of the new shift.
1680  TheUse = InsertedShift;
1681  }
1682 
1683  // If we removed all uses, nuke the shift.
1684  if (ShiftI->use_empty()) {
1685  salvageDebugInfo(*ShiftI);
1686  ShiftI->eraseFromParent();
1687  }
1688 
1689  return MadeChange;
1690 }
1691 
1692 /// If counting leading or trailing zeros is an expensive operation and a zero
1693 /// input is defined, add a check for zero to avoid calling the intrinsic.
1694 ///
1695 /// We want to transform:
1696 /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 false)
1697 ///
1698 /// into:
1699 /// entry:
1700 /// %cmpz = icmp eq i64 %A, 0
1701 /// br i1 %cmpz, label %cond.end, label %cond.false
1702 /// cond.false:
1703 /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 true)
1704 /// br label %cond.end
1705 /// cond.end:
1706 /// %ctz = phi i64 [ 64, %entry ], [ %z, %cond.false ]
1707 ///
1708 /// If the transform is performed, return true and set ModifiedDT to true.
1709 static bool despeculateCountZeros(IntrinsicInst *CountZeros,
1710  const TargetLowering *TLI,
1711  const DataLayout *DL,
1712  bool &ModifiedDT) {
1713  if (!TLI || !DL)
1714  return false;
1715 
1716  // If a zero input is undefined, it doesn't make sense to despeculate that.
1717  if (match(CountZeros->getOperand(1), m_One()))
1718  return false;
1719 
1720  // If it's cheap to speculate, there's nothing to do.
1721  auto IntrinsicID = CountZeros->getIntrinsicID();
1722  if ((IntrinsicID == Intrinsic::cttz && TLI->isCheapToSpeculateCttz()) ||
1723  (IntrinsicID == Intrinsic::ctlz && TLI->isCheapToSpeculateCtlz()))
1724  return false;
1725 
1726  // Only handle legal scalar cases. Anything else requires too much work.
1727  Type *Ty = CountZeros->getType();
1728  unsigned SizeInBits = Ty->getPrimitiveSizeInBits();
1729  if (Ty->isVectorTy() || SizeInBits > DL->getLargestLegalIntTypeSizeInBits())
1730  return false;
1731 
1732  // The intrinsic will be sunk behind a compare against zero and branch.
1733  BasicBlock *StartBlock = CountZeros->getParent();
1734  BasicBlock *CallBlock = StartBlock->splitBasicBlock(CountZeros, "cond.false");
1735 
1736  // Create another block after the count zero intrinsic. A PHI will be added
1737  // in this block to select the result of the intrinsic or the bit-width
1738  // constant if the input to the intrinsic is zero.
1739  BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(CountZeros));
1740  BasicBlock *EndBlock = CallBlock->splitBasicBlock(SplitPt, "cond.end");
1741 
1742  // Set up a builder to create a compare, conditional branch, and PHI.
1743  IRBuilder<> Builder(CountZeros->getContext());
1744  Builder.SetInsertPoint(StartBlock->getTerminator());
1745  Builder.SetCurrentDebugLocation(CountZeros->getDebugLoc());
1746 
1747  // Replace the unconditional branch that was created by the first split with
1748  // a compare against zero and a conditional branch.
1749  Value *Zero = Constant::getNullValue(Ty);
1750  Value *Cmp = Builder.CreateICmpEQ(CountZeros->getOperand(0), Zero, "cmpz");
1751  Builder.CreateCondBr(Cmp, EndBlock, CallBlock);
1752  StartBlock->getTerminator()->eraseFromParent();
1753 
1754  // Create a PHI in the end block to select either the output of the intrinsic
1755  // or the bit width of the operand.
1756  Builder.SetInsertPoint(&EndBlock->front());
1757  PHINode *PN = Builder.CreatePHI(Ty, 2, "ctz");
1758  CountZeros->replaceAllUsesWith(PN);
1759  Value *BitWidth = Builder.getInt(APInt(SizeInBits, SizeInBits));
1760  PN->addIncoming(BitWidth, StartBlock);
1761  PN->addIncoming(CountZeros, CallBlock);
1762 
1763  // We are explicitly handling the zero case, so we can set the intrinsic's
1764  // undefined zero argument to 'true'. This will also prevent reprocessing the
1765  // intrinsic; we only despeculate when a zero input is defined.
1766  CountZeros->setArgOperand(1, Builder.getTrue());
1767  ModifiedDT = true;
1768  return true;
1769 }
1770 
1771 bool CodeGenPrepare::optimizeCallInst(CallInst *CI, bool &ModifiedDT) {
1772  BasicBlock *BB = CI->getParent();
1773 
1774  // Lower inline assembly if we can.
1775  // If we found an inline asm expession, and if the target knows how to
1776  // lower it to normal LLVM code, do so now.
1777  if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
1778  if (TLI->ExpandInlineAsm(CI)) {
1779  // Avoid invalidating the iterator.
1780  CurInstIterator = BB->begin();
1781  // Avoid processing instructions out of order, which could cause
1782  // reuse before a value is defined.
1783  SunkAddrs.clear();
1784  return true;
1785  }
1786  // Sink address computing for memory operands into the block.
1787  if (optimizeInlineAsmInst(CI))
1788  return true;
1789  }
1790 
1791  // Align the pointer arguments to this call if the target thinks it's a good
1792  // idea
1793  unsigned MinSize, PrefAlign;
1794  if (TLI && TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) {
1795  for (auto &Arg : CI->arg_operands()) {
1796  // We want to align both objects whose address is used directly and
1797  // objects whose address is used in casts and GEPs, though it only makes
1798  // sense for GEPs if the offset is a multiple of the desired alignment and
1799  // if size - offset meets the size threshold.
1800  if (!Arg->getType()->isPointerTy())
1801  continue;
1802  APInt Offset(DL->getIndexSizeInBits(
1803  cast<PointerType>(Arg->getType())->getAddressSpace()),
1804  0);
1805  Value *Val = Arg->stripAndAccumulateInBoundsConstantOffsets(*DL, Offset);
1806  uint64_t Offset2 = Offset.getLimitedValue();
1807  if ((Offset2 & (PrefAlign-1)) != 0)
1808  continue;
1809  AllocaInst *AI;
1810  if ((AI = dyn_cast<AllocaInst>(Val)) && AI->getAlignment() < PrefAlign &&
1811  DL->getTypeAllocSize(AI->getAllocatedType()) >= MinSize + Offset2)
1812  AI->setAlignment(PrefAlign);
1813  // Global variables can only be aligned if they are defined in this
1814  // object (i.e. they are uniquely initialized in this object), and
1815  // over-aligning global variables that have an explicit section is
1816  // forbidden.
1817  GlobalVariable *GV;
1818  if ((GV = dyn_cast<GlobalVariable>(Val)) && GV->canIncreaseAlignment() &&
1819  GV->getPointerAlignment(*DL) < PrefAlign &&
1820  DL->getTypeAllocSize(GV->getValueType()) >=
1821  MinSize + Offset2)
1822  GV->setAlignment(PrefAlign);
1823  }
1824  // If this is a memcpy (or similar) then we may be able to improve the
1825  // alignment
1826  if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(CI)) {
1827  unsigned DestAlign = getKnownAlignment(MI->getDest(), *DL);
1828  if (DestAlign > MI->getDestAlignment())
1829  MI->setDestAlignment(DestAlign);
1830  if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) {
1831  unsigned SrcAlign = getKnownAlignment(MTI->getSource(), *DL);
1832  if (SrcAlign > MTI->getSourceAlignment())
1833  MTI->setSourceAlignment(SrcAlign);
1834  }
1835  }
1836  }
1837 
1838  // If we have a cold call site, try to sink addressing computation into the
1839  // cold block. This interacts with our handling for loads and stores to
1840  // ensure that we can fold all uses of a potential addressing computation
1841  // into their uses. TODO: generalize this to work over profiling data
1842  if (!OptSize && CI->hasFnAttr(Attribute::Cold))
1843  for (auto &Arg : CI->arg_operands()) {
1844  if (!Arg->getType()->isPointerTy())
1845  continue;
1846  unsigned AS = Arg->getType()->getPointerAddressSpace();
1847  return optimizeMemoryInst(CI, Arg, Arg->getType(), AS);
1848  }
1849 
1851  if (II) {
1852  switch (II->getIntrinsicID()) {
1853  default: break;
1854  case Intrinsic::experimental_widenable_condition: {
1855  // Give up on future widening oppurtunties so that we can fold away dead
1856  // paths and merge blocks before going into block-local instruction
1857  // selection.
1858  if (II->use_empty()) {
1859  II->eraseFromParent();
1860  return true;
1861  }
1862  Constant *RetVal = ConstantInt::getTrue(II->getContext());
1863  resetIteratorIfInvalidatedWhileCalling(BB, [&]() {
1864  replaceAndRecursivelySimplify(CI, RetVal, TLInfo, nullptr);
1865  });
1866  return true;
1867  }
1868  case Intrinsic::objectsize: {
1869  // Lower all uses of llvm.objectsize.*
1870  Value *RetVal =
1871  lowerObjectSizeCall(II, *DL, TLInfo, /*MustSucceed=*/true);
1872 
1873  resetIteratorIfInvalidatedWhileCalling(BB, [&]() {
1874  replaceAndRecursivelySimplify(CI, RetVal, TLInfo, nullptr);
1875  });
1876  return true;
1877  }
1878  case Intrinsic::is_constant: {
1879  // If is_constant hasn't folded away yet, lower it to false now.
1880  Constant *RetVal = ConstantInt::get(II->getType(), 0);
1881  resetIteratorIfInvalidatedWhileCalling(BB, [&]() {
1882  replaceAndRecursivelySimplify(CI, RetVal, TLInfo, nullptr);
1883  });
1884  return true;
1885  }
1886  case Intrinsic::aarch64_stlxr:
1887  case Intrinsic::aarch64_stxr: {
1888  ZExtInst *ExtVal = dyn_cast<ZExtInst>(CI->getArgOperand(0));
1889  if (!ExtVal || !ExtVal->hasOneUse() ||
1890  ExtVal->getParent() == CI->getParent())
1891  return false;
1892  // Sink a zext feeding stlxr/stxr before it, so it can be folded into it.
1893  ExtVal->moveBefore(CI);
1894  // Mark this instruction as "inserted by CGP", so that other
1895  // optimizations don't touch it.
1896  InsertedInsts.insert(ExtVal);
1897  return true;
1898  }
1899 
1900  case Intrinsic::launder_invariant_group:
1901  case Intrinsic::strip_invariant_group: {
1902  Value *ArgVal = II->getArgOperand(0);
1903  auto it = LargeOffsetGEPMap.find(II);
1904  if (it != LargeOffsetGEPMap.end()) {
1905  // Merge entries in LargeOffsetGEPMap to reflect the RAUW.
1906  // Make sure not to have to deal with iterator invalidation
1907  // after possibly adding ArgVal to LargeOffsetGEPMap.
1908  auto GEPs = std::move(it->second);
1909  LargeOffsetGEPMap[ArgVal].append(GEPs.begin(), GEPs.end());
1910  LargeOffsetGEPMap.erase(II);
1911  }
1912 
1913  II->replaceAllUsesWith(ArgVal);
1914  II->eraseFromParent();
1915  return true;
1916  }
1917  case Intrinsic::cttz:
1918  case Intrinsic::ctlz:
1919  // If counting zeros is expensive, try to avoid it.
1920  return despeculateCountZeros(II, TLI, DL, ModifiedDT);
1921  }
1922 
1923  if (TLI) {
1924  SmallVector<Value*, 2> PtrOps;
1925  Type *AccessTy;
1926  if (TLI->getAddrModeArguments(II, PtrOps, AccessTy))
1927  while (!PtrOps.empty()) {
1928  Value *PtrVal = PtrOps.pop_back_val();
1929  unsigned AS = PtrVal->getType()->getPointerAddressSpace();
1930  if (optimizeMemoryInst(II, PtrVal, AccessTy, AS))
1931  return true;
1932  }
1933  }
1934  }
1935 
1936  // From here on out we're working with named functions.
1937  if (!CI->getCalledFunction()) return false;
1938 
1939  // Lower all default uses of _chk calls. This is very similar
1940  // to what InstCombineCalls does, but here we are only lowering calls
1941  // to fortified library functions (e.g. __memcpy_chk) that have the default
1942  // "don't know" as the objectsize. Anything else should be left alone.
1943  FortifiedLibCallSimplifier Simplifier(TLInfo, true);
1944  if (Value *V = Simplifier.optimizeCall(CI)) {
1945  CI->replaceAllUsesWith(V);
1946  CI->eraseFromParent();
1947  return true;
1948  }
1949 
1950  return false;
1951 }
1952 
1953 /// Look for opportunities to duplicate return instructions to the predecessor
1954 /// to enable tail call optimizations. The case it is currently looking for is:
1955 /// @code
1956 /// bb0:
1957 /// %tmp0 = tail call i32 @f0()
1958 /// br label %return
1959 /// bb1:
1960 /// %tmp1 = tail call i32 @f1()
1961 /// br label %return
1962 /// bb2:
1963 /// %tmp2 = tail call i32 @f2()
1964 /// br label %return
1965 /// return:
1966 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
1967 /// ret i32 %retval
1968 /// @endcode
1969 ///
1970 /// =>
1971 ///
1972 /// @code
1973 /// bb0:
1974 /// %tmp0 = tail call i32 @f0()
1975 /// ret i32 %tmp0
1976 /// bb1:
1977 /// %tmp1 = tail call i32 @f1()
1978 /// ret i32 %tmp1
1979 /// bb2:
1980 /// %tmp2 = tail call i32 @f2()
1981 /// ret i32 %tmp2
1982 /// @endcode
1983 bool CodeGenPrepare::dupRetToEnableTailCallOpts(BasicBlock *BB, bool &ModifiedDT) {
1984  if (!TLI)
1985  return false;
1986 
1987  ReturnInst *RetI = dyn_cast<ReturnInst>(BB->getTerminator());
1988  if (!RetI)
1989  return false;
1990 
1991  PHINode *PN = nullptr;
1992  BitCastInst *BCI = nullptr;
1993  Value *V = RetI->getReturnValue();
1994  if (V) {
1995  BCI = dyn_cast<BitCastInst>(V);
1996  if (BCI)
1997  V = BCI->getOperand(0);
1998 
1999  PN = dyn_cast<PHINode>(V);
2000  if (!PN)
2001  return false;
2002  }
2003 
2004  if (PN && PN->getParent() != BB)
2005  return false;
2006 
2007  // Make sure there are no instructions between the PHI and return, or that the
2008  // return is the first instruction in the block.
2009  if (PN) {
2010  BasicBlock::iterator BI = BB->begin();
2011  // Skip over debug and the bitcast.
2012  do { ++BI; } while (isa<DbgInfoIntrinsic>(BI) || &*BI == BCI);
2013  if (&*BI != RetI)
2014  return false;
2015  } else {
2016  BasicBlock::iterator BI = BB->begin();
2017  while (isa<DbgInfoIntrinsic>(BI)) ++BI;
2018  if (&*BI != RetI)
2019  return false;
2020  }
2021 
2022  /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
2023  /// call.
2024  const Function *F = BB->getParent();
2025  SmallVector<CallInst*, 4> TailCalls;
2026  if (PN) {
2027  for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
2028  // Look through bitcasts.
2029  Value *IncomingVal = PN->getIncomingValue(I)->stripPointerCasts();
2030  CallInst *CI = dyn_cast<CallInst>(IncomingVal);
2031  // Make sure the phi value is indeed produced by the tail call.
2032  if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
2033  TLI->mayBeEmittedAsTailCall(CI) &&
2034  attributesPermitTailCall(F, CI, RetI, *TLI))
2035  TailCalls.push_back(CI);
2036  }
2037  } else {
2038  SmallPtrSet<BasicBlock*, 4> VisitedBBs;
2039  for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
2040  if (!VisitedBBs.insert(*PI).second)
2041  continue;
2042 
2043  BasicBlock::InstListType &InstList = (*PI)->getInstList();
2044  BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
2045  BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
2046  do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
2047  if (RI == RE)
2048  continue;
2049 
2050  CallInst *CI = dyn_cast<CallInst>(&*RI);
2051  if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI) &&
2052  attributesPermitTailCall(F, CI, RetI, *TLI))
2053  TailCalls.push_back(CI);
2054  }
2055  }
2056 
2057  bool Changed = false;
2058  for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
2059  CallInst *CI = TailCalls[i];
2060  CallSite CS(CI);
2061 
2062  // Make sure the call instruction is followed by an unconditional branch to
2063  // the return block.
2064  BasicBlock *CallBB = CI->getParent();
2065  BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
2066  if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
2067  continue;
2068 
2069  // Duplicate the return into CallBB.
2070  (void)FoldReturnIntoUncondBranch(RetI, BB, CallBB);
2071  ModifiedDT = Changed = true;
2072  ++NumRetsDup;
2073  }
2074 
2075  // If we eliminated all predecessors of the block, delete the block now.
2076  if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
2077  BB->eraseFromParent();
2078 
2079  return Changed;
2080 }
2081 
2082 //===----------------------------------------------------------------------===//
2083 // Memory Optimization
2084 //===----------------------------------------------------------------------===//
2085 
2086 namespace {
2087 
2088 /// This is an extended version of TargetLowering::AddrMode
2089 /// which holds actual Value*'s for register values.
2090 struct ExtAddrMode : public TargetLowering::AddrMode {
2091  Value *BaseReg = nullptr;
2092  Value *ScaledReg = nullptr;
2093  Value *OriginalValue = nullptr;
2094  bool InBounds = true;
2095 
2096  enum FieldName {
2097  NoField = 0x00,
2098  BaseRegField = 0x01,
2099  BaseGVField = 0x02,
2100  BaseOffsField = 0x04,
2101  ScaledRegField = 0x08,
2102  ScaleField = 0x10,
2103  MultipleFields = 0xff
2104  };
2105 
2106 
2107  ExtAddrMode() = default;
2108 
2109  void print(raw_ostream &OS) const;
2110  void dump() const;
2111 
2112  FieldName compare(const ExtAddrMode &other) {
2113  // First check that the types are the same on each field, as differing types
2114  // is something we can't cope with later on.
2115  if (BaseReg && other.BaseReg &&
2116  BaseReg->getType() != other.BaseReg->getType())
2117  return MultipleFields;
2118  if (BaseGV && other.BaseGV &&
2119  BaseGV->getType() != other.BaseGV->getType())
2120  return MultipleFields;
2121  if (ScaledReg && other.ScaledReg &&
2122  ScaledReg->getType() != other.ScaledReg->getType())
2123  return MultipleFields;
2124 
2125  // Conservatively reject 'inbounds' mismatches.
2126  if (InBounds != other.InBounds)
2127  return MultipleFields;
2128 
2129  // Check each field to see if it differs.
2130  unsigned Result = NoField;
2131  if (BaseReg != other.BaseReg)
2132  Result |= BaseRegField;
2133  if (BaseGV != other.BaseGV)
2134  Result |= BaseGVField;
2135  if (BaseOffs != other.BaseOffs)
2136  Result |= BaseOffsField;
2137  if (ScaledReg != other.ScaledReg)
2138  Result |= ScaledRegField;
2139  // Don't count 0 as being a different scale, because that actually means
2140  // unscaled (which will already be counted by having no ScaledReg).
2141  if (Scale && other.Scale && Scale != other.Scale)
2142  Result |= ScaleField;
2143 
2144  if (countPopulation(Result) > 1)
2145  return MultipleFields;
2146  else
2147  return static_cast<FieldName>(Result);
2148  }
2149 
2150  // An AddrMode is trivial if it involves no calculation i.e. it is just a base
2151  // with no offset.
2152  bool isTrivial() {
2153  // An AddrMode is (BaseGV + BaseReg + BaseOffs + ScaleReg * Scale) so it is
2154  // trivial if at most one of these terms is nonzero, except that BaseGV and
2155  // BaseReg both being zero actually means a null pointer value, which we
2156  // consider to be 'non-zero' here.
2157  return !BaseOffs && !Scale && !(BaseGV && BaseReg);
2158  }
2159 
2160  Value *GetFieldAsValue(FieldName Field, Type *IntPtrTy) {
2161  switch (Field) {
2162  default:
2163  return nullptr;
2164  case BaseRegField:
2165  return BaseReg;
2166  case BaseGVField:
2167  return BaseGV;
2168  case ScaledRegField:
2169  return ScaledReg;
2170  case BaseOffsField:
2171  return ConstantInt::get(IntPtrTy, BaseOffs);
2172  }
2173  }
2174 
2175  void SetCombinedField(FieldName Field, Value *V,
2176  const SmallVectorImpl<ExtAddrMode> &AddrModes) {
2177  switch (Field) {
2178  default:
2179  llvm_unreachable("Unhandled fields are expected to be rejected earlier");
2180  break;
2181  case ExtAddrMode::BaseRegField:
2182  BaseReg = V;
2183  break;
2184  case ExtAddrMode::BaseGVField:
2185  // A combined BaseGV is an Instruction, not a GlobalValue, so it goes
2186  // in the BaseReg field.
2187  assert(BaseReg == nullptr);
2188  BaseReg = V;
2189  BaseGV = nullptr;
2190  break;
2191  case ExtAddrMode::ScaledRegField:
2192  ScaledReg = V;
2193  // If we have a mix of scaled and unscaled addrmodes then we want scale
2194  // to be the scale and not zero.
2195  if (!Scale)
2196  for (const ExtAddrMode &AM : AddrModes)
2197  if (AM.Scale) {
2198  Scale = AM.Scale;
2199  break;
2200  }
2201  break;
2202  case ExtAddrMode::BaseOffsField:
2203  // The offset is no longer a constant, so it goes in ScaledReg with a
2204  // scale of 1.
2205  assert(ScaledReg == nullptr);
2206  ScaledReg = V;
2207  Scale = 1;
2208  BaseOffs = 0;
2209  break;
2210  }
2211  }
2212 };
2213 
2214 } // end anonymous namespace
2215 
2216 #ifndef NDEBUG
2217 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
2218  AM.print(OS);
2219  return OS;
2220 }
2221 #endif
2222 
2223 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
2224 void ExtAddrMode::print(raw_ostream &OS) const {
2225  bool NeedPlus = false;
2226  OS << "[";
2227  if (InBounds)
2228  OS << "inbounds ";
2229  if (BaseGV) {
2230  OS << (NeedPlus ? " + " : "")
2231  << "GV:";
2232  BaseGV->printAsOperand(OS, /*PrintType=*/false);
2233  NeedPlus = true;
2234  }
2235 
2236  if (BaseOffs) {
2237  OS << (NeedPlus ? " + " : "")
2238  << BaseOffs;
2239  NeedPlus = true;
2240  }
2241 
2242  if (BaseReg) {
2243  OS << (NeedPlus ? " + " : "")
2244  << "Base:";
2245  BaseReg->printAsOperand(OS, /*PrintType=*/false);
2246  NeedPlus = true;
2247  }
2248  if (Scale) {
2249  OS << (NeedPlus ? " + " : "")
2250  << Scale << "*";
2251  ScaledReg->printAsOperand(OS, /*PrintType=*/false);
2252  }
2253 
2254  OS << ']';
2255 }
2256 
2257 LLVM_DUMP_METHOD void ExtAddrMode::dump() const {
2258  print(dbgs());
2259  dbgs() << '\n';
2260 }
2261 #endif
2262 
2263 namespace {
2264 
2265 /// This class provides transaction based operation on the IR.
2266 /// Every change made through this class is recorded in the internal state and
2267 /// can be undone (rollback) until commit is called.
2268 class TypePromotionTransaction {
2269  /// This represents the common interface of the individual transaction.
2270  /// Each class implements the logic for doing one specific modification on
2271  /// the IR via the TypePromotionTransaction.
2272  class TypePromotionAction {
2273  protected:
2274  /// The Instruction modified.
2275  Instruction *Inst;
2276 
2277  public:
2278  /// Constructor of the action.
2279  /// The constructor performs the related action on the IR.
2280  TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
2281 
2282  virtual ~TypePromotionAction() = default;
2283 
2284  /// Undo the modification done by this action.
2285  /// When this method is called, the IR must be in the same state as it was
2286  /// before this action was applied.
2287  /// \pre Undoing the action works if and only if the IR is in the exact same
2288  /// state as it was directly after this action was applied.
2289  virtual void undo() = 0;
2290 
2291  /// Advocate every change made by this action.
2292  /// When the results on the IR of the action are to be kept, it is important
2293  /// to call this function, otherwise hidden information may be kept forever.
2294  virtual void commit() {
2295  // Nothing to be done, this action is not doing anything.
2296  }
2297  };
2298 
2299  /// Utility to remember the position of an instruction.
2300  class InsertionHandler {
2301  /// Position of an instruction.
2302  /// Either an instruction:
2303  /// - Is the first in a basic block: BB is used.
2304  /// - Has a previous instruction: PrevInst is used.
2305  union {
2306  Instruction *PrevInst;
2307  BasicBlock *BB;
2308  } Point;
2309 
2310  /// Remember whether or not the instruction had a previous instruction.
2311  bool HasPrevInstruction;
2312 
2313  public:
2314  /// Record the position of \p Inst.
2315  InsertionHandler(Instruction *Inst) {
2316  BasicBlock::iterator It = Inst->getIterator();
2317  HasPrevInstruction = (It != (Inst->getParent()->begin()));
2318  if (HasPrevInstruction)
2319  Point.PrevInst = &*--It;
2320  else
2321  Point.BB = Inst->getParent();
2322  }
2323 
2324  /// Insert \p Inst at the recorded position.
2325  void insert(Instruction *Inst) {
2326  if (HasPrevInstruction) {
2327  if (Inst->getParent())
2328  Inst->removeFromParent();
2329  Inst->insertAfter(Point.PrevInst);
2330  } else {
2331  Instruction *Position = &*Point.BB->getFirstInsertionPt();
2332  if (Inst->getParent())
2333  Inst->moveBefore(Position);
2334  else
2335  Inst->insertBefore(Position);
2336  }
2337  }
2338  };
2339 
2340  /// Move an instruction before another.
2341  class InstructionMoveBefore : public TypePromotionAction {
2342  /// Original position of the instruction.
2343  InsertionHandler Position;
2344 
2345  public:
2346  /// Move \p Inst before \p Before.
2347  InstructionMoveBefore(Instruction *Inst, Instruction *Before)
2348  : TypePromotionAction(Inst), Position(Inst) {
2349  LLVM_DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before
2350  << "\n");
2351  Inst->moveBefore(Before);
2352  }
2353 
2354  /// Move the instruction back to its original position.
2355  void undo() override {
2356  LLVM_DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
2357  Position.insert(Inst);
2358  }
2359  };
2360 
2361  /// Set the operand of an instruction with a new value.
2362  class OperandSetter : public TypePromotionAction {
2363  /// Original operand of the instruction.
2364  Value *Origin;
2365 
2366  /// Index of the modified instruction.
2367  unsigned Idx;
2368 
2369  public:
2370  /// Set \p Idx operand of \p Inst with \p NewVal.
2371  OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
2372  : TypePromotionAction(Inst), Idx(Idx) {
2373  LLVM_DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
2374  << "for:" << *Inst << "\n"
2375  << "with:" << *NewVal << "\n");
2376  Origin = Inst->getOperand(Idx);
2377  Inst->setOperand(Idx, NewVal);
2378  }
2379 
2380  /// Restore the original value of the instruction.
2381  void undo() override {
2382  LLVM_DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
2383  << "for: " << *Inst << "\n"
2384  << "with: " << *Origin << "\n");
2385  Inst->setOperand(Idx, Origin);
2386  }
2387  };
2388 
2389  /// Hide the operands of an instruction.
2390  /// Do as if this instruction was not using any of its operands.
2391  class OperandsHider : public TypePromotionAction {
2392  /// The list of original operands.
2393  SmallVector<Value *, 4> OriginalValues;
2394 
2395  public:
2396  /// Remove \p Inst from the uses of the operands of \p Inst.
2397  OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
2398  LLVM_DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
2399  unsigned NumOpnds = Inst->getNumOperands();
2400  OriginalValues.reserve(NumOpnds);
2401  for (unsigned It = 0; It < NumOpnds; ++It) {
2402  // Save the current operand.
2403  Value *Val = Inst->getOperand(It);
2404  OriginalValues.push_back(Val);
2405  // Set a dummy one.
2406  // We could use OperandSetter here, but that would imply an overhead
2407  // that we are not willing to pay.
2408  Inst->setOperand(It, UndefValue::get(Val->getType()));
2409  }
2410  }
2411 
2412  /// Restore the original list of uses.
2413  void undo() override {
2414  LLVM_DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
2415  for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
2416  Inst->setOperand(It, OriginalValues[It]);
2417  }
2418  };
2419 
2420  /// Build a truncate instruction.
2421  class TruncBuilder : public TypePromotionAction {
2422  Value *Val;
2423 
2424  public:
2425  /// Build a truncate instruction of \p Opnd producing a \p Ty
2426  /// result.
2427  /// trunc Opnd to Ty.
2428  TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
2429  IRBuilder<> Builder(Opnd);
2430  Val = Builder.CreateTrunc(Opnd, Ty, "promoted");
2431  LLVM_DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n");
2432  }
2433 
2434  /// Get the built value.
2435  Value *getBuiltValue() { return Val; }
2436 
2437  /// Remove the built instruction.
2438  void undo() override {
2439  LLVM_DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n");
2440  if (Instruction *IVal = dyn_cast<Instruction>(Val))
2441  IVal->eraseFromParent();
2442  }
2443  };
2444 
2445  /// Build a sign extension instruction.
2446  class SExtBuilder : public TypePromotionAction {
2447  Value *Val;
2448 
2449  public:
2450  /// Build a sign extension instruction of \p Opnd producing a \p Ty
2451  /// result.
2452  /// sext Opnd to Ty.
2453  SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
2454  : TypePromotionAction(InsertPt) {
2455  IRBuilder<> Builder(InsertPt);
2456  Val = Builder.CreateSExt(Opnd, Ty, "promoted");
2457  LLVM_DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n");
2458  }
2459 
2460  /// Get the built value.
2461  Value *getBuiltValue() { return Val; }
2462 
2463  /// Remove the built instruction.
2464  void undo() override {
2465  LLVM_DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n");
2466  if (Instruction *IVal = dyn_cast<Instruction>(Val))
2467  IVal->eraseFromParent();
2468  }
2469  };
2470 
2471  /// Build a zero extension instruction.
2472  class ZExtBuilder : public TypePromotionAction {
2473  Value *Val;
2474 
2475  public:
2476  /// Build a zero extension instruction of \p Opnd producing a \p Ty
2477  /// result.
2478  /// zext Opnd to Ty.
2479  ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
2480  : TypePromotionAction(InsertPt) {
2481  IRBuilder<> Builder(InsertPt);
2482  Val = Builder.CreateZExt(Opnd, Ty, "promoted");
2483  LLVM_DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n");
2484  }
2485 
2486  /// Get the built value.
2487  Value *getBuiltValue() { return Val; }
2488 
2489  /// Remove the built instruction.
2490  void undo() override {
2491  LLVM_DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n");
2492  if (Instruction *IVal = dyn_cast<Instruction>(Val))
2493  IVal->eraseFromParent();
2494  }
2495  };
2496 
2497  /// Mutate an instruction to another type.
2498  class TypeMutator : public TypePromotionAction {
2499  /// Record the original type.
2500  Type *OrigTy;
2501 
2502  public:
2503  /// Mutate the type of \p Inst into \p NewTy.
2504  TypeMutator(Instruction *Inst, Type *NewTy)
2505  : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
2506  LLVM_DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
2507  << "\n");
2508  Inst->mutateType(NewTy);
2509  }
2510 
2511  /// Mutate the instruction back to its original type.
2512  void undo() override {
2513  LLVM_DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
2514  << "\n");
2515  Inst->mutateType(OrigTy);
2516  }
2517  };
2518 
2519  /// Replace the uses of an instruction by another instruction.
2520  class UsesReplacer : public TypePromotionAction {
2521  /// Helper structure to keep track of the replaced uses.
2522  struct InstructionAndIdx {
2523  /// The instruction using the instruction.
2524  Instruction *Inst;
2525 
2526  /// The index where this instruction is used for Inst.
2527  unsigned Idx;
2528 
2529  InstructionAndIdx(Instruction *Inst, unsigned Idx)
2530  : Inst(Inst), Idx(Idx) {}
2531  };
2532 
2533  /// Keep track of the original uses (pair Instruction, Index).
2534  SmallVector<InstructionAndIdx, 4> OriginalUses;
2535  /// Keep track of the debug users.
2537 
2538  using use_iterator = SmallVectorImpl<InstructionAndIdx>::iterator;
2539 
2540  public:
2541  /// Replace all the use of \p Inst by \p New.
2542  UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
2543  LLVM_DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
2544  << "\n");
2545  // Record the original uses.
2546  for (Use &U : Inst->uses()) {
2547  Instruction *UserI = cast<Instruction>(U.getUser());
2548  OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
2549  }
2550  // Record the debug uses separately. They are not in the instruction's
2551  // use list, but they are replaced by RAUW.
2552  findDbgValues(DbgValues, Inst);
2553 
2554  // Now, we can replace the uses.
2555  Inst->replaceAllUsesWith(New);
2556  }
2557 
2558  /// Reassign the original uses of Inst to Inst.
2559  void undo() override {
2560  LLVM_DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
2561  for (use_iterator UseIt = OriginalUses.begin(),
2562  EndIt = OriginalUses.end();
2563  UseIt != EndIt; ++UseIt) {
2564  UseIt->Inst->setOperand(UseIt->Idx, Inst);
2565  }
2566  // RAUW has replaced all original uses with references to the new value,
2567  // including the debug uses. Since we are undoing the replacements,
2568  // the original debug uses must also be reinstated to maintain the
2569  // correctness and utility of debug value instructions.
2570  for (auto *DVI: DbgValues) {
2571  LLVMContext &Ctx = Inst->getType()->getContext();
2572  auto *MV = MetadataAsValue::get(Ctx, ValueAsMetadata::get(Inst));
2573  DVI->setOperand(0, MV);
2574  }
2575  }
2576  };
2577 
2578  /// Remove an instruction from the IR.
2579  class InstructionRemover : public TypePromotionAction {
2580  /// Original position of the instruction.
2581  InsertionHandler Inserter;
2582 
2583  /// Helper structure to hide all the link to the instruction. In other
2584  /// words, this helps to do as if the instruction was removed.
2585  OperandsHider Hider;
2586 
2587  /// Keep track of the uses replaced, if any.
2588  UsesReplacer *Replacer = nullptr;
2589 
2590  /// Keep track of instructions removed.
2591  SetOfInstrs &RemovedInsts;
2592 
2593  public:
2594  /// Remove all reference of \p Inst and optionally replace all its
2595  /// uses with New.
2596  /// \p RemovedInsts Keep track of the instructions removed by this Action.
2597  /// \pre If !Inst->use_empty(), then New != nullptr
2598  InstructionRemover(Instruction *Inst, SetOfInstrs &RemovedInsts,
2599  Value *New = nullptr)
2600  : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
2601  RemovedInsts(RemovedInsts) {
2602  if (New)
2603  Replacer = new UsesReplacer(Inst, New);
2604  LLVM_DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
2605  RemovedInsts.insert(Inst);
2606  /// The instructions removed here will be freed after completing
2607  /// optimizeBlock() for all blocks as we need to keep track of the
2608  /// removed instructions during promotion.
2609  Inst->removeFromParent();
2610  }
2611 
2612  ~InstructionRemover() override { delete Replacer; }
2613 
2614  /// Resurrect the instruction and reassign it to the proper uses if
2615  /// new value was provided when build this action.
2616  void undo() override {
2617  LLVM_DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
2618  Inserter.insert(Inst);
2619  if (Replacer)
2620  Replacer->undo();
2621  Hider.undo();
2622  RemovedInsts.erase(Inst);
2623  }
2624  };
2625 
2626 public:
2627  /// Restoration point.
2628  /// The restoration point is a pointer to an action instead of an iterator
2629  /// because the iterator may be invalidated but not the pointer.
2630  using ConstRestorationPt = const TypePromotionAction *;
2631 
2632  TypePromotionTransaction(SetOfInstrs &RemovedInsts)
2633  : RemovedInsts(RemovedInsts) {}
2634 
2635  /// Advocate every changes made in that transaction.
2636  void commit();
2637 
2638  /// Undo all the changes made after the given point.
2639  void rollback(ConstRestorationPt Point);
2640 
2641  /// Get the current restoration point.
2642  ConstRestorationPt getRestorationPoint() const;
2643 
2644  /// \name API for IR modification with state keeping to support rollback.
2645  /// @{
2646  /// Same as Instruction::setOperand.
2647  void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
2648 
2649  /// Same as Instruction::eraseFromParent.
2650  void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
2651 
2652  /// Same as Value::replaceAllUsesWith.
2653  void replaceAllUsesWith(Instruction *Inst, Value *New);
2654 
2655  /// Same as Value::mutateType.
2656  void mutateType(Instruction *Inst, Type *NewTy);
2657 
2658  /// Same as IRBuilder::createTrunc.
2659  Value *createTrunc(Instruction *Opnd, Type *Ty);
2660 
2661  /// Same as IRBuilder::createSExt.
2662  Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
2663 
2664  /// Same as IRBuilder::createZExt.
2665  Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
2666 
2667  /// Same as Instruction::moveBefore.
2668  void moveBefore(Instruction *Inst, Instruction *Before);
2669  /// @}
2670 
2671 private:
2672  /// The ordered list of actions made so far.
2674 
2675  using CommitPt = SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator;
2676 
2677  SetOfInstrs &RemovedInsts;
2678 };
2679 
2680 } // end anonymous namespace
2681 
2682 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
2683  Value *NewVal) {
2684  Actions.push_back(llvm::make_unique<TypePromotionTransaction::OperandSetter>(
2685  Inst, Idx, NewVal));
2686 }
2687 
2689  Value *NewVal) {
2690  Actions.push_back(
2691  llvm::make_unique<TypePromotionTransaction::InstructionRemover>(
2692  Inst, RemovedInsts, NewVal));
2693 }
2694 
2695 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
2696  Value *New) {
2697  Actions.push_back(
2698  llvm::make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
2699 }
2700 
2701 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
2702  Actions.push_back(
2703  llvm::make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
2704 }
2705 
2706 Value *TypePromotionTransaction::createTrunc(Instruction *Opnd,
2707  Type *Ty) {
2708  std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
2709  Value *Val = Ptr->getBuiltValue();
2710  Actions.push_back(std::move(Ptr));
2711  return Val;
2712 }
2713 
2714 Value *TypePromotionTransaction::createSExt(Instruction *Inst,
2715  Value *Opnd, Type *Ty) {
2716  std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
2717  Value *Val = Ptr->getBuiltValue();
2718  Actions.push_back(std::move(Ptr));
2719  return Val;
2720 }
2721 
2722 Value *TypePromotionTransaction::createZExt(Instruction *Inst,
2723  Value *Opnd, Type *Ty) {
2724  std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
2725  Value *Val = Ptr->getBuiltValue();
2726  Actions.push_back(std::move(Ptr));
2727  return Val;
2728 }
2729 
2730 void TypePromotionTransaction::moveBefore(Instruction *Inst,
2731  Instruction *Before) {
2732  Actions.push_back(
2733  llvm::make_unique<TypePromotionTransaction::InstructionMoveBefore>(
2734  Inst, Before));
2735 }
2736 
2737 TypePromotionTransaction::ConstRestorationPt
2738 TypePromotionTransaction::getRestorationPoint() const {
2739  return !Actions.empty() ? Actions.back().get() : nullptr;
2740 }
2741 
2742 void TypePromotionTransaction::commit() {
2743  for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
2744  ++It)
2745  (*It)->commit();
2746  Actions.clear();
2747 }
2748 
2749 void TypePromotionTransaction::rollback(
2750  TypePromotionTransaction::ConstRestorationPt Point) {
2751  while (!Actions.empty() && Point != Actions.back().get()) {
2752  std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
2753  Curr->undo();
2754  }
2755 }
2756 
2757 namespace {
2758 
2759 /// A helper class for matching addressing modes.
2760 ///
2761 /// This encapsulates the logic for matching the target-legal addressing modes.
2762 class AddressingModeMatcher {
2763  SmallVectorImpl<Instruction*> &AddrModeInsts;
2764  const TargetLowering &TLI;
2765  const TargetRegisterInfo &TRI;
2766  const DataLayout &DL;
2767 
2768  /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
2769  /// the memory instruction that we're computing this address for.
2770  Type *AccessTy;
2771  unsigned AddrSpace;
2772  Instruction *MemoryInst;
2773 
2774  /// This is the addressing mode that we're building up. This is
2775  /// part of the return value of this addressing mode matching stuff.
2776  ExtAddrMode &AddrMode;
2777 
2778  /// The instructions inserted by other CodeGenPrepare optimizations.
2779  const SetOfInstrs &InsertedInsts;
2780 
2781  /// A map from the instructions to their type before promotion.
2782  InstrToOrigTy &PromotedInsts;
2783 
2784  /// The ongoing transaction where every action should be registered.
2785  TypePromotionTransaction &TPT;
2786 
2787  // A GEP which has too large offset to be folded into the addressing mode.
2788  std::pair<AssertingVH<GetElementPtrInst>, int64_t> &LargeOffsetGEP;
2789 
2790  /// This is set to true when we should not do profitability checks.
2791  /// When true, IsProfitableToFoldIntoAddressingMode always returns true.
2792  bool IgnoreProfitability;
2793 
2794  AddressingModeMatcher(
2796  const TargetRegisterInfo &TRI, Type *AT, unsigned AS, Instruction *MI,
2797  ExtAddrMode &AM, const SetOfInstrs &InsertedInsts,
2798  InstrToOrigTy &PromotedInsts, TypePromotionTransaction &TPT,
2799  std::pair<AssertingVH<GetElementPtrInst>, int64_t> &LargeOffsetGEP)
2800  : AddrModeInsts(AMI), TLI(TLI), TRI(TRI),
2801  DL(MI->getModule()->getDataLayout()), AccessTy(AT), AddrSpace(AS),
2802  MemoryInst(MI), AddrMode(AM), InsertedInsts(InsertedInsts),
2803  PromotedInsts(PromotedInsts), TPT(TPT), LargeOffsetGEP(LargeOffsetGEP) {
2804  IgnoreProfitability = false;
2805  }
2806 
2807 public:
2808  /// Find the maximal addressing mode that a load/store of V can fold,
2809  /// give an access type of AccessTy. This returns a list of involved
2810  /// instructions in AddrModeInsts.
2811  /// \p InsertedInsts The instructions inserted by other CodeGenPrepare
2812  /// optimizations.
2813  /// \p PromotedInsts maps the instructions to their type before promotion.
2814  /// \p The ongoing transaction where every action should be registered.
2815  static ExtAddrMode
2816  Match(Value *V, Type *AccessTy, unsigned AS, Instruction *MemoryInst,
2817  SmallVectorImpl<Instruction *> &AddrModeInsts,
2818  const TargetLowering &TLI, const TargetRegisterInfo &TRI,
2819  const SetOfInstrs &InsertedInsts, InstrToOrigTy &PromotedInsts,
2820  TypePromotionTransaction &TPT,
2821  std::pair<AssertingVH<GetElementPtrInst>, int64_t> &LargeOffsetGEP) {
2822  ExtAddrMode Result;
2823 
2824  bool Success = AddressingModeMatcher(AddrModeInsts, TLI, TRI, AccessTy, AS,
2825  MemoryInst, Result, InsertedInsts,
2826  PromotedInsts, TPT, LargeOffsetGEP)
2827  .matchAddr(V, 0);
2828  (void)Success; assert(Success && "Couldn't select *anything*?");
2829  return Result;
2830  }
2831 
2832 private:
2833  bool matchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
2834  bool matchAddr(Value *Addr, unsigned Depth);
2835  bool matchOperationAddr(User *AddrInst, unsigned Opcode, unsigned Depth,
2836  bool *MovedAway = nullptr);
2837  bool isProfitableToFoldIntoAddressingMode(Instruction *I,
2838  ExtAddrMode &AMBefore,
2839  ExtAddrMode &AMAfter);
2840  bool valueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
2841  bool isPromotionProfitable(unsigned NewCost, unsigned OldCost,
2842  Value *PromotedOperand) const;
2843 };
2844 
2845 class PhiNodeSet;
2846 
2847 /// An iterator for PhiNodeSet.
2848 class PhiNodeSetIterator {
2849  PhiNodeSet * const Set;
2850  size_t CurrentIndex = 0;
2851 
2852 public:
2853  /// The constructor. Start should point to either a valid element, or be equal
2854  /// to the size of the underlying SmallVector of the PhiNodeSet.
2855  PhiNodeSetIterator(PhiNodeSet * const Set, size_t Start);
2856  PHINode * operator*() const;
2857  PhiNodeSetIterator& operator++();
2858  bool operator==(const PhiNodeSetIterator &RHS) const;
2859  bool operator!=(const PhiNodeSetIterator &RHS) const;
2860 };
2861 
2862 /// Keeps a set of PHINodes.
2863 ///
2864 /// This is a minimal set implementation for a specific use case:
2865 /// It is very fast when there are very few elements, but also provides good
2866 /// performance when there are many. It is similar to SmallPtrSet, but also
2867 /// provides iteration by insertion order, which is deterministic and stable
2868 /// across runs. It is also similar to SmallSetVector, but provides removing
2869 /// elements in O(1) time. This is achieved by not actually removing the element
2870 /// from the underlying vector, so comes at the cost of using more memory, but
2871 /// that is fine, since PhiNodeSets are used as short lived objects.
2872 class PhiNodeSet {
2873  friend class PhiNodeSetIterator;
2874 
2875  using MapType = SmallDenseMap<PHINode *, size_t, 32>;
2876  using iterator = PhiNodeSetIterator;
2877 
2878  /// Keeps the elements in the order of their insertion in the underlying
2879  /// vector. To achieve constant time removal, it never deletes any element.
2881 
2882  /// Keeps the elements in the underlying set implementation. This (and not the
2883  /// NodeList defined above) is the source of truth on whether an element
2884  /// is actually in the collection.
2885  MapType NodeMap;
2886 
2887  /// Points to the first valid (not deleted) element when the set is not empty
2888  /// and the value is not zero. Equals to the size of the underlying vector
2889  /// when the set is empty. When the value is 0, as in the beginning, the
2890  /// first element may or may not be valid.
2891  size_t FirstValidElement = 0;
2892 
2893 public:
2894  /// Inserts a new element to the collection.
2895  /// \returns true if the element is actually added, i.e. was not in the
2896  /// collection before the operation.
2897  bool insert(PHINode *Ptr) {
2898  if (NodeMap.insert(std::make_pair(Ptr, NodeList.size())).second) {
2899  NodeList.push_back(Ptr);
2900  return true;
2901  }
2902  return false;
2903  }
2904 
2905  /// Removes the element from the collection.
2906  /// \returns whether the element is actually removed, i.e. was in the
2907  /// collection before the operation.
2908  bool erase(PHINode *Ptr) {
2909  auto it = NodeMap.find(Ptr);
2910  if (it != NodeMap.end()) {
2911  NodeMap.erase(Ptr);
2912  SkipRemovedElements(FirstValidElement);
2913  return true;
2914  }
2915  return false;
2916  }
2917 
2918  /// Removes all elements and clears the collection.
2919  void clear() {
2920  NodeMap.clear();
2921  NodeList.clear();
2922  FirstValidElement = 0;
2923  }
2924 
2925  /// \returns an iterator that will iterate the elements in the order of
2926  /// insertion.
2927  iterator begin() {
2928  if (FirstValidElement == 0)
2929  SkipRemovedElements(FirstValidElement);
2930  return PhiNodeSetIterator(this, FirstValidElement);
2931  }
2932 
2933  /// \returns an iterator that points to the end of the collection.
2934  iterator end() { return PhiNodeSetIterator(this, NodeList.size()); }
2935 
2936  /// Returns the number of elements in the collection.
2937  size_t size() const {
2938  return NodeMap.size();
2939  }
2940 
2941  /// \returns 1 if the given element is in the collection, and 0 if otherwise.
2942  size_t count(PHINode *Ptr) const {
2943  return NodeMap.count(Ptr);
2944  }
2945 
2946 private:
2947  /// Updates the CurrentIndex so that it will point to a valid element.
2948  ///
2949  /// If the element of NodeList at CurrentIndex is valid, it does not
2950  /// change it. If there are no more valid elements, it updates CurrentIndex
2951  /// to point to the end of the NodeList.
2952  void SkipRemovedElements(size_t &CurrentIndex) {
2953  while (CurrentIndex < NodeList.size()) {
2954  auto it = NodeMap.find(NodeList[CurrentIndex]);
2955  // If the element has been deleted and added again later, NodeMap will
2956  // point to a different index, so CurrentIndex will still be invalid.
2957  if (it != NodeMap.end() && it->second == CurrentIndex)
2958  break;
2959  ++CurrentIndex;
2960  }
2961  }
2962 };
2963 
2964 PhiNodeSetIterator::PhiNodeSetIterator(PhiNodeSet *const Set, size_t Start)
2965  : Set(Set), CurrentIndex(Start) {}
2966 
2968  assert(CurrentIndex < Set->NodeList.size() &&
2969  "PhiNodeSet access out of range");
2970  return Set->NodeList[CurrentIndex];
2971 }
2972 
2973 PhiNodeSetIterator& PhiNodeSetIterator::operator++() {
2974  assert(CurrentIndex < Set->NodeList.size() &&
2975  "PhiNodeSet access out of range");
2976  ++CurrentIndex;
2977  Set->SkipRemovedElements(CurrentIndex);
2978  return *this;
2979 }
2980 
2981 bool PhiNodeSetIterator::operator==(const PhiNodeSetIterator &RHS) const {
2982  return CurrentIndex == RHS.CurrentIndex;
2983 }
2984 
2985 bool PhiNodeSetIterator::operator!=(const PhiNodeSetIterator &RHS) const {
2986  return !((*this) == RHS);
2987 }
2988 
2989 /// Keep track of simplification of Phi nodes.
2990 /// Accept the set of all phi nodes and erase phi node from this set
2991 /// if it is simplified.
2992 class SimplificationTracker {
2994  const SimplifyQuery &SQ;
2995  // Tracks newly created Phi nodes. The elements are iterated by insertion
2996  // order.
2997  PhiNodeSet AllPhiNodes;
2998  // Tracks newly created Select nodes.
2999  SmallPtrSet<SelectInst *, 32> AllSelectNodes;
3000 
3001 public:
3002  SimplificationTracker(const SimplifyQuery &sq)
3003  : SQ(sq) {}
3004 
3005  Value *Get(Value *V) {
3006  do {
3007  auto SV = Storage.find(V);
3008  if (SV == Storage.end())
3009  return V;
3010  V = SV->second;
3011  } while (true);
3012  }
3013 
3014  Value *Simplify(Value *Val) {
3015  SmallVector<Value *, 32> WorkList;
3016  SmallPtrSet<Value *, 32> Visited;
3017  WorkList.push_back(Val);
3018  while (!WorkList.empty()) {
3019  auto P = WorkList.pop_back_val();
3020  if (!Visited.insert(P).second)
3021  continue;
3022  if (auto *PI = dyn_cast<Instruction>(P))
3023  if (Value *V = SimplifyInstruction(cast<Instruction>(PI), SQ)) {
3024  for (auto *U : PI->users())
3025  WorkList.push_back(cast<Value>(U));
3026  Put(PI, V);
3027  PI->replaceAllUsesWith(V);
3028  if (auto *PHI = dyn_cast<PHINode>(PI))
3029  AllPhiNodes.erase(PHI);
3030  if (auto *Select = dyn_cast<SelectInst>(PI))
3031  AllSelectNodes.erase(Select);
3032  PI->eraseFromParent();
3033  }
3034  }
3035  return Get(Val);
3036  }
3037 
3038  void Put(Value *From, Value *To) {
3039  Storage.insert({ From, To });
3040  }
3041 
3042  void ReplacePhi(PHINode *From, PHINode *To) {
3043  Value* OldReplacement = Get(From);
3044  while (OldReplacement != From) {
3045  From = To;
3046  To = dyn_cast<PHINode>(OldReplacement);
3047  OldReplacement = Get(From);
3048  }
3049  assert(Get(To) == To && "Replacement PHI node is already replaced.");
3050  Put(From, To);
3051  From->replaceAllUsesWith(To);
3052  AllPhiNodes.erase(From);
3053  From->eraseFromParent();
3054  }
3055 
3056  PhiNodeSet& newPhiNodes() { return AllPhiNodes; }
3057 
3058  void insertNewPhi(PHINode *PN) { AllPhiNodes.insert(PN); }
3059 
3060  void insertNewSelect(SelectInst *SI) { AllSelectNodes.insert(SI); }
3061 
3062  unsigned countNewPhiNodes() const { return AllPhiNodes.size(); }
3063 
3064  unsigned countNewSelectNodes() const { return AllSelectNodes.size(); }
3065 
3066  void destroyNewNodes(Type *CommonType) {
3067  // For safe erasing, replace the uses with dummy value first.
3068  auto Dummy = UndefValue::get(CommonType);
3069  for (auto I : AllPhiNodes) {
3071  I->eraseFromParent();
3072  }
3073  AllPhiNodes.clear();
3074  for (auto I : AllSelectNodes) {
3076  I->eraseFromParent();
3077  }
3078  AllSelectNodes.clear();
3079  }
3080 };
3081 
3082 /// A helper class for combining addressing modes.
3083 class AddressingModeCombiner {
3084  typedef DenseMap<Value *, Value *> FoldAddrToValueMapping;
3085  typedef std::pair<PHINode *, PHINode *> PHIPair;
3086 
3087 private:
3088  /// The addressing modes we've collected.
3089  SmallVector<ExtAddrMode, 16> AddrModes;
3090 
3091  /// The field in which the AddrModes differ, when we have more than one.
3092  ExtAddrMode::FieldName DifferentField = ExtAddrMode::NoField;
3093 
3094  /// Are the AddrModes that we have all just equal to their original values?
3095  bool AllAddrModesTrivial = true;
3096 
3097  /// Common Type for all different fields in addressing modes.
3098  Type *CommonType;
3099 
3100  /// SimplifyQuery for simplifyInstruction utility.
3101  const SimplifyQuery &SQ;
3102 
3103  /// Original Address.
3104  Value *Original;
3105 
3106 public:
3107  AddressingModeCombiner(const SimplifyQuery &_SQ, Value *OriginalValue)
3108  : CommonType(nullptr), SQ(_SQ), Original(OriginalValue) {}
3109 
3110  /// Get the combined AddrMode
3111  const ExtAddrMode &getAddrMode() const {
3112  return AddrModes[0];
3113  }
3114 
3115  /// Add a new AddrMode if it's compatible with the AddrModes we already
3116  /// have.
3117  /// \return True iff we succeeded in doing so.
3118  bool addNewAddrMode(ExtAddrMode &NewAddrMode) {
3119  // Take note of if we have any non-trivial AddrModes, as we need to detect
3120  // when all AddrModes are trivial as then we would introduce a phi or select
3121  // which just duplicates what's already there.
3122  AllAddrModesTrivial = AllAddrModesTrivial && NewAddrMode.isTrivial();
3123 
3124  // If this is the first addrmode then everything is fine.
3125  if (AddrModes.empty()) {
3126  AddrModes.emplace_back(NewAddrMode);
3127  return true;
3128  }
3129 
3130  // Figure out how different this is from the other address modes, which we
3131  // can do just by comparing against the first one given that we only care
3132  // about the cumulative difference.
3133  ExtAddrMode::FieldName ThisDifferentField =
3134  AddrModes[0].compare(NewAddrMode);
3135  if (DifferentField == ExtAddrMode::NoField)
3136  DifferentField = ThisDifferentField;
3137  else if (DifferentField != ThisDifferentField)
3138  DifferentField = ExtAddrMode::MultipleFields;
3139 
3140  // If NewAddrMode differs in more than one dimension we cannot handle it.
3141  bool CanHandle = DifferentField != ExtAddrMode::MultipleFields;
3142 
3143  // If Scale Field is different then we reject.
3144  CanHandle = CanHandle && DifferentField != ExtAddrMode::ScaleField;
3145 
3146  // We also must reject the case when base offset is different and
3147  // scale reg is not null, we cannot handle this case due to merge of
3148  // different offsets will be used as ScaleReg.
3149  CanHandle = CanHandle && (DifferentField != ExtAddrMode::BaseOffsField ||
3150  !NewAddrMode.ScaledReg);
3151 
3152  // We also must reject the case when GV is different and BaseReg installed
3153  // due to we want to use base reg as a merge of GV values.
3154  CanHandle = CanHandle && (DifferentField != ExtAddrMode::BaseGVField ||
3155  !NewAddrMode.HasBaseReg);
3156 
3157  // Even if NewAddMode is the same we still need to collect it due to
3158  // original value is different. And later we will need all original values
3159  // as anchors during finding the common Phi node.
3160  if (CanHandle)
3161  AddrModes.emplace_back(NewAddrMode);
3162  else
3163  AddrModes.clear();
3164 
3165  return CanHandle;
3166  }
3167 
3168  /// Combine the addressing modes we've collected into a single
3169  /// addressing mode.
3170  /// \return True iff we successfully combined them or we only had one so
3171  /// didn't need to combine them anyway.
3172  bool combineAddrModes() {
3173  // If we have no AddrModes then they can't be combined.
3174  if (AddrModes.size() == 0)
3175  return false;
3176 
3177  // A single AddrMode can trivially be combined.
3178  if (AddrModes.size() == 1 || DifferentField == ExtAddrMode::NoField)
3179  return true;
3180 
3181  // If the AddrModes we collected are all just equal to the value they are
3182  // derived from then combining them wouldn't do anything useful.
3183  if (AllAddrModesTrivial)
3184  return false;
3185 
3186  if (!addrModeCombiningAllowed())
3187  return false;
3188 
3189  // Build a map between <original value, basic block where we saw it> to
3190  // value of base register.
3191  // Bail out if there is no common type.
3192  FoldAddrToValueMapping Map;
3193  if (!initializeMap(Map))
3194  return false;
3195 
3196  Value *CommonValue = findCommon(Map);
3197  if (CommonValue)
3198  AddrModes[0].SetCombinedField(DifferentField, CommonValue, AddrModes);
3199  return CommonValue != nullptr;
3200  }
3201 
3202 private:
3203  /// Initialize Map with anchor values. For address seen
3204  /// we set the value of different field saw in this address.
3205  /// At the same time we find a common type for different field we will
3206  /// use to create new Phi/Select nodes. Keep it in CommonType field.
3207  /// Return false if there is no common type found.
3208  bool initializeMap(FoldAddrToValueMapping &Map) {
3209  // Keep track of keys where the value is null. We will need to replace it
3210  // with constant null when we know the common type.
3211  SmallVector<Value *, 2> NullValue;
3212  Type *IntPtrTy = SQ.DL.getIntPtrType(AddrModes[0].OriginalValue->getType());
3213  for (auto &AM : AddrModes) {
3214  Value *DV = AM.GetFieldAsValue(DifferentField, IntPtrTy);
3215  if (DV) {
3216  auto *Type = DV->getType();
3217  if (CommonType && CommonType != Type)
3218  return false;
3219  CommonType = Type;
3220  Map[AM.OriginalValue] = DV;
3221  } else {
3222  NullValue.push_back(AM.OriginalValue);
3223  }
3224  }
3225  assert(CommonType && "At least one non-null value must be!");
3226  for (auto *V : NullValue)
3227  Map[V] = Constant::getNullValue(CommonType);
3228  return true;
3229  }
3230 
3231  /// We have mapping between value A and other value B where B was a field in
3232  /// addressing mode represented by A. Also we have an original value C
3233  /// representing an address we start with. Traversing from C through phi and
3234  /// selects we ended up with A's in a map. This utility function tries to find
3235  /// a value V which is a field in addressing mode C and traversing through phi
3236  /// nodes and selects we will end up in corresponded values B in a map.
3237  /// The utility will create a new Phi/Selects if needed.
3238  // The simple example looks as follows:
3239  // BB1:
3240  // p1 = b1 + 40
3241  // br cond BB2, BB3
3242  // BB2:
3243  // p2 = b2 + 40
3244  // br BB3
3245  // BB3:
3246  // p = phi [p1, BB1], [p2, BB2]
3247  // v = load p
3248  // Map is
3249  // p1 -> b1
3250  // p2 -> b2
3251  // Request is
3252  // p -> ?
3253  // The function tries to find or build phi [b1, BB1], [b2, BB2] in BB3.
3254  Value *findCommon(FoldAddrToValueMapping &Map) {
3255  // Tracks the simplification of newly created phi nodes. The reason we use
3256  // this mapping is because we will add new created Phi nodes in AddrToBase.
3257  // Simplification of Phi nodes is recursive, so some Phi node may
3258  // be simplified after we added it to AddrToBase. In reality this
3259  // simplification is possible only if original phi/selects were not
3260  // simplified yet.
3261  // Using this mapping we can find the current value in AddrToBase.
3262  SimplificationTracker ST(SQ);
3263 
3264  // First step, DFS to create PHI nodes for all intermediate blocks.
3265  // Also fill traverse order for the second step.
3266  SmallVector<Value *, 32> TraverseOrder;
3267  InsertPlaceholders(Map, TraverseOrder, ST);
3268 
3269  // Second Step, fill new nodes by merged values and simplify if possible.
3270  FillPlaceholders(Map, TraverseOrder, ST);
3271 
3272  if (!AddrSinkNewSelects && ST.countNewSelectNodes() > 0) {
3273  ST.destroyNewNodes(CommonType);
3274  return nullptr;
3275  }
3276 
3277  // Now we'd like to match New Phi nodes to existed ones.
3278  unsigned PhiNotMatchedCount = 0;
3279  if (!MatchPhiSet(ST, AddrSinkNewPhis, PhiNotMatchedCount)) {
3280  ST.destroyNewNodes(CommonType);
3281  return nullptr;
3282  }
3283 
3284  auto *Result = ST.Get(Map.find(Original)->second);
3285  if (Result) {
3286  NumMemoryInstsPhiCreated += ST.countNewPhiNodes() + PhiNotMatchedCount;
3287  NumMemoryInstsSelectCreated += ST.countNewSelectNodes();
3288  }
3289  return Result;
3290  }
3291 
3292  /// Try to match PHI node to Candidate.
3293  /// Matcher tracks the matched Phi nodes.
3294  bool MatchPhiNode(PHINode *PHI, PHINode *Candidate,
3295  SmallSetVector<PHIPair, 8> &Matcher,
3296  PhiNodeSet &PhiNodesToMatch) {
3297  SmallVector<PHIPair, 8> WorkList;
3298  Matcher.insert({ PHI, Candidate });
3299  SmallSet<PHINode *, 8> MatchedPHIs;
3300  MatchedPHIs.insert(PHI);
3301  WorkList.push_back({ PHI, Candidate });
3302  SmallSet<PHIPair, 8> Visited;
3303  while (!WorkList.empty()) {
3304  auto Item = WorkList.pop_back_val();
3305  if (!Visited.insert(Item).second)
3306  continue;
3307  // We iterate over all incoming values to Phi to compare them.
3308  // If values are different and both of them Phi and the first one is a
3309  // Phi we added (subject to match) and both of them is in the same basic
3310  // block then we can match our pair if values match. So we state that
3311  // these values match and add it to work list to verify that.
3312  for (auto B : Item.first->blocks()) {
3313  Value *FirstValue = Item.first->getIncomingValueForBlock(B);
3314  Value *SecondValue = Item.second->getIncomingValueForBlock(B);
3315  if (FirstValue == SecondValue)
3316  continue;
3317 
3318  PHINode *FirstPhi = dyn_cast<PHINode>(FirstValue);
3319  PHINode *SecondPhi = dyn_cast<PHINode>(SecondValue);
3320 
3321  // One of them is not Phi or
3322  // The first one is not Phi node from the set we'd like to match or
3323  // Phi nodes from different basic blocks then
3324  // we will not be able to match.
3325  if (!FirstPhi || !SecondPhi || !PhiNodesToMatch.count(FirstPhi) ||
3326  FirstPhi->getParent() != SecondPhi->getParent())
3327  return false;
3328 
3329  // If we already matched them then continue.
3330  if (Matcher.count({ FirstPhi, SecondPhi }))
3331  continue;
3332  // So the values are different and does not match. So we need them to
3333  // match. (But we register no more than one match per PHI node, so that
3334  // we won't later try to replace them twice.)
3335  if (!MatchedPHIs.insert(FirstPhi).second)
3336  Matcher.insert({ FirstPhi, SecondPhi });
3337  // But me must check it.
3338  WorkList.push_back({ FirstPhi, SecondPhi });
3339  }
3340  }
3341  return true;
3342  }
3343 
3344  /// For the given set of PHI nodes (in the SimplificationTracker) try
3345  /// to find their equivalents.
3346  /// Returns false if this matching fails and creation of new Phi is disabled.
3347  bool MatchPhiSet(SimplificationTracker &ST, bool AllowNewPhiNodes,
3348  unsigned &PhiNotMatchedCount) {
3349  // Matched and PhiNodesToMatch iterate their elements in a deterministic
3350  // order, so the replacements (ReplacePhi) are also done in a deterministic
3351  // order.
3353  SmallPtrSet<PHINode *, 8> WillNotMatch;
3354  PhiNodeSet &PhiNodesToMatch = ST.newPhiNodes();
3355  while (PhiNodesToMatch.size()) {
3356  PHINode *PHI = *PhiNodesToMatch.begin();
3357 
3358  // Add us, if no Phi nodes in the basic block we do not match.
3359  WillNotMatch.clear();
3360  WillNotMatch.insert(PHI);
3361 
3362  // Traverse all Phis until we found equivalent or fail to do that.
3363  bool IsMatched = false;
3364  for (auto &P : PHI->getParent()->phis()) {
3365  if (&P == PHI)
3366  continue;
3367  if ((IsMatched = MatchPhiNode(PHI, &P, Matched, PhiNodesToMatch)))
3368  break;
3369  // If it does not match, collect all Phi nodes from matcher.
3370  // if we end up with no match, them all these Phi nodes will not match
3371  // later.
3372  for (auto M : Matched)
3373  WillNotMatch.insert(M.first);
3374  Matched.clear();
3375  }
3376  if (IsMatched) {
3377  // Replace all matched values and erase them.
3378  for (auto MV : Matched)
3379  ST.ReplacePhi(MV.first, MV.second);
3380  Matched.clear();
3381  continue;
3382  }
3383  // If we are not allowed to create new nodes then bail out.
3384  if (!AllowNewPhiNodes)
3385  return false;
3386  // Just remove all seen values in matcher. They will not match anything.
3387  PhiNotMatchedCount += WillNotMatch.size();
3388  for (auto *P : WillNotMatch)
3389  PhiNodesToMatch.erase(P);
3390  }
3391  return true;
3392  }
3393  /// Fill the placeholders with values from predecessors and simplify them.
3394  void FillPlaceholders(FoldAddrToValueMapping &Map,
3395  SmallVectorImpl<Value *> &TraverseOrder,
3396  SimplificationTracker &ST) {
3397  while (!TraverseOrder.empty()) {
3398  Value *Current = TraverseOrder.pop_back_val();
3399  assert(Map.find(Current) != Map.end() && "No node to fill!!!");
3400  Value *V = Map[Current];
3401 
3402  if (SelectInst *Select = dyn_cast<SelectInst>(V)) {
3403  // CurrentValue also must be Select.
3404  auto *CurrentSelect = cast<SelectInst>(Current);
3405  auto *TrueValue = CurrentSelect->getTrueValue();
3406  assert(Map.find(TrueValue) != Map.end() && "No True Value!");
3407  Select->setTrueValue(ST.Get(Map[TrueValue]));
3408  auto *FalseValue = CurrentSelect->getFalseValue();
3409  assert(Map.find(FalseValue) != Map.end() && "No False Value!");
3410  Select->setFalseValue(ST.Get(Map[FalseValue]));
3411  } else {
3412  // Must be a Phi node then.
3413  PHINode *PHI = cast<PHINode>(V);
3414  auto *CurrentPhi = dyn_cast<PHINode>(Current);
3415  // Fill the Phi node with values from predecessors.
3416  for (auto B : predecessors(PHI->getParent())) {
3417  Value *PV = CurrentPhi->getIncomingValueForBlock(B);
3418  assert(Map.find(PV) != Map.end() && "No predecessor Value!");
3419  PHI->addIncoming(ST.Get(Map[PV]), B);
3420  }
3421  }
3422  Map[Current] = ST.Simplify(V);
3423  }
3424  }
3425 
3426  /// Starting from original value recursively iterates over def-use chain up to
3427  /// known ending values represented in a map. For each traversed phi/select
3428  /// inserts a placeholder Phi or Select.
3429  /// Reports all new created Phi/Select nodes by adding them to set.
3430  /// Also reports and order in what values have been traversed.
3431  void InsertPlaceholders(FoldAddrToValueMapping &Map,
3432  SmallVectorImpl<Value *> &TraverseOrder,
3433  SimplificationTracker &ST) {
3434  SmallVector<Value *, 32> Worklist;
3435  assert((isa<PHINode>(Original) || isa<SelectInst>(Original)) &&
3436  "Address must be a Phi or Select node");
3437  auto *Dummy = UndefValue::get(CommonType);
3438  Worklist.push_back(Original);
3439  while (!Worklist.empty()) {
3440  Value *Current = Worklist.pop_back_val();
3441  // if it is already visited or it is an ending value then skip it.
3442  if (Map.find(Current) != Map.end())
3443  continue;
3444  TraverseOrder.push_back(Current);
3445 
3446  // CurrentValue must be a Phi node or select. All others must be covered
3447  // by anchors.
3448  if (SelectInst *CurrentSelect = dyn_cast<SelectInst>(Current)) {
3449  // Is it OK to get metadata from OrigSelect?!
3450  // Create a Select placeholder with dummy value.
3452  CurrentSelect->getCondition(), Dummy, Dummy,
3453  CurrentSelect->getName(), CurrentSelect, CurrentSelect);
3454  Map[Current] = Select;
3455  ST.insertNewSelect(Select);
3456  // We are interested in True and False values.
3457  Worklist.push_back(CurrentSelect->getTrueValue());
3458  Worklist.push_back(CurrentSelect->getFalseValue());
3459  } else {
3460  // It must be a Phi node then.
3461  PHINode *CurrentPhi = cast<PHINode>(Current);
3462  unsigned PredCount = CurrentPhi->getNumIncomingValues();
3463  PHINode *PHI =
3464  PHINode::Create(CommonType, PredCount, "sunk_phi", CurrentPhi);
3465  Map[Current] = PHI;
3466  ST.insertNewPhi(PHI);
3467  for (Value *P : CurrentPhi->incoming_values())
3468  Worklist.push_back(P);
3469  }
3470  }
3471  }
3472 
3473  bool addrModeCombiningAllowed() {
3475  return false;
3476  switch (DifferentField) {
3477  default:
3478  return false;
3479  case ExtAddrMode::BaseRegField:
3480  return AddrSinkCombineBaseReg;
3481  case ExtAddrMode::BaseGVField:
3482  return AddrSinkCombineBaseGV;
3483  case ExtAddrMode::BaseOffsField:
3484  return AddrSinkCombineBaseOffs;
3485  case ExtAddrMode::ScaledRegField:
3486  return AddrSinkCombineScaledReg;
3487  }
3488  }
3489 };
3490 } // end anonymous namespace
3491 
3492 /// Try adding ScaleReg*Scale to the current addressing mode.
3493 /// Return true and update AddrMode if this addr mode is legal for the target,
3494 /// false if not.
3495 bool AddressingModeMatcher::matchScaledValue(Value *ScaleReg, int64_t Scale,
3496  unsigned Depth) {
3497  // If Scale is 1, then this is the same as adding ScaleReg to the addressing
3498  // mode. Just process that directly.
3499  if (Scale == 1)
3500  return matchAddr(ScaleReg, Depth);
3501 
3502  // If the scale is 0, it takes nothing to add this.
3503  if (Scale == 0)
3504  return true;
3505 
3506  // If we already have a scale of this value, we can add to it, otherwise, we
3507  // need an available scale field.
3508  if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
3509  return false;
3510 
3511  ExtAddrMode TestAddrMode = AddrMode;
3512 
3513  // Add scale to turn X*4+X*3 -> X*7. This could also do things like
3514  // [A+B + A*7] -> [B+A*8].
3515  TestAddrMode.Scale += Scale;
3516  TestAddrMode.ScaledReg = ScaleReg;
3517 
3518  // If the new address isn't legal, bail out.
3519  if (!TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace))
3520  return false;
3521 
3522  // It was legal, so commit it.
3523  AddrMode = TestAddrMode;
3524 
3525  // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
3526  // to see if ScaleReg is actually X+C. If so, we can turn this into adding
3527  // X*Scale + C*Scale to addr mode.
3528  ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
3529  if (isa<Instruction>(ScaleReg) && // not a constant expr.
3530  match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
3531  TestAddrMode.InBounds = false;
3532  TestAddrMode.ScaledReg = AddLHS;
3533  TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
3534 
3535  // If this addressing mode is legal, commit it and remember that we folded
3536  // this instruction.
3537  if (TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace)) {
3538  AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
3539  AddrMode = TestAddrMode;
3540  return true;
3541  }
3542  }
3543 
3544  // Otherwise, not (x+c)*scale, just return what we have.
3545  return true;
3546 }
3547 
3548 /// This is a little filter, which returns true if an addressing computation
3549 /// involving I might be folded into a load/store accessing it.
3550 /// This doesn't need to be perfect, but needs to accept at least
3551 /// the set of instructions that MatchOperationAddr can.
3553  switch (I->getOpcode()) {
3554  case Instruction::BitCast:
3555  case Instruction::AddrSpaceCast:
3556  // Don't touch identity bitcasts.
3557  if (I->getType() == I->getOperand(0)->getType())
3558  return false;
3559  return I->getType()->isIntOrPtrTy();
3560  case Instruction::PtrToInt:
3561  // PtrToInt is always a noop, as we know that the int type is pointer sized.
3562  return true;
3563  case Instruction::IntToPtr:
3564  // We know the input is intptr_t, so this is foldable.
3565  return true;
3566  case Instruction::Add:
3567  return true;
3568  case Instruction::Mul:
3569  case Instruction::Shl:
3570  // Can only handle X*C and X << C.
3571  return isa<ConstantInt>(I->getOperand(1));
3572  case Instruction::GetElementPtr:
3573  return true;
3574  default:
3575  return false;
3576  }
3577 }
3578 
3579 /// Check whether or not \p Val is a legal instruction for \p TLI.
3580 /// \note \p Val is assumed to be the product of some type promotion.
3581 /// Therefore if \p Val has an undefined state in \p TLI, this is assumed
3582 /// to be legal, as the non-promoted value would have had the same state.
3584  const DataLayout &DL, Value *Val) {
3585  Instruction *PromotedInst = dyn_cast<Instruction>(Val);
3586  if (!PromotedInst)
3587  return false;
3588  int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
3589  // If the ISDOpcode is undefined, it was undefined before the promotion.
3590  if (!ISDOpcode)
3591  return true;
3592  // Otherwise, check if the promoted instruction is legal or not.
3593  return TLI.isOperationLegalOrCustom(
3594  ISDOpcode, TLI.getValueType(DL, PromotedInst->getType()));
3595 }
3596 
3597 namespace {
3598 
3599 /// Hepler class to perform type promotion.
3600 class TypePromotionHelper {
3601  /// Utility function to add a promoted instruction \p ExtOpnd to
3602  /// \p PromotedInsts and record the type of extension we have seen.
3603  static void addPromotedInst(InstrToOrigTy &PromotedInsts,
3604  Instruction *ExtOpnd,
3605  bool IsSExt) {
3606  ExtType ExtTy = IsSExt ? SignExtension : ZeroExtension;
3607  InstrToOrigTy::iterator It = PromotedInsts.find(ExtOpnd);
3608  if (It != PromotedInsts.end()) {
3609  // If the new extension is same as original, the information in
3610  // PromotedInsts[ExtOpnd] is still correct.
3611  if (It->second.getInt() == ExtTy)
3612  return;
3613 
3614  // Now the new extension is different from old extension, we make
3615  // the type information invalid by setting extension type to
3616  // BothExtension.
3617  ExtTy = BothExtension;
3618  }
3619  PromotedInsts[ExtOpnd] = TypeIsSExt(ExtOpnd->getType(), ExtTy);
3620  }
3621 
3622  /// Utility function to query the original type of instruction \p Opnd
3623  /// with a matched extension type. If the extension doesn't match, we
3624  /// cannot use the information we had on the original type.
3625  /// BothExtension doesn't match any extension type.
3626  static const Type *getOrigType(const InstrToOrigTy &PromotedInsts,
3627  Instruction *Opnd,
3628  bool IsSExt) {
3629  ExtType ExtTy = IsSExt ? SignExtension : ZeroExtension;
3630  InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
3631  if (It != PromotedInsts.end() && It->second.getInt() == ExtTy)
3632  return It->second.getPointer();
3633  return nullptr;
3634  }
3635 
3636  /// Utility function to check whether or not a sign or zero extension
3637  /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
3638  /// either using the operands of \p Inst or promoting \p Inst.
3639  /// The type of the extension is defined by \p IsSExt.
3640  /// In other words, check if:
3641  /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
3642  /// #1 Promotion applies:
3643  /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
3644  /// #2 Operand reuses:
3645  /// ext opnd1 to ConsideredExtType.
3646  /// \p PromotedInsts maps the instructions to their type before promotion.
3647  static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
3648  const InstrToOrigTy &PromotedInsts, bool IsSExt);
3649 
3650  /// Utility function to determine if \p OpIdx should be promoted when
3651  /// promoting \p Inst.
3652  static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
3653  return !(isa<SelectInst>(Inst) && OpIdx == 0);
3654  }
3655 
3656  /// Utility function to promote the operand of \p Ext when this
3657  /// operand is a promotable trunc or sext or zext.
3658  /// \p PromotedInsts maps the instructions to their type before promotion.
3659  /// \p CreatedInstsCost[out] contains the cost of all instructions
3660  /// created to promote the operand of Ext.
3661  /// Newly added extensions are inserted in \p Exts.
3662  /// Newly added truncates are inserted in \p Truncs.
3663  /// Should never be called directly.
3664  /// \return The promoted value which is used instead of Ext.
3665  static Value *promoteOperandForTruncAndAnyExt(
3666  Instruction *Ext, TypePromotionTransaction &TPT,
3667  InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3669  SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI);
3670 
3671  /// Utility function to promote the operand of \p Ext when this
3672  /// operand is promotable and is not a supported trunc or sext.
3673  /// \p PromotedInsts maps the instructions to their type before promotion.
3674  /// \p CreatedInstsCost[out] contains the cost of all the instructions
3675  /// created to promote the operand of Ext.
3676  /// Newly added extensions are inserted in \p Exts.
3677  /// Newly added truncates are inserted in \p Truncs.
3678  /// Should never be called directly.
3679  /// \return The promoted value which is used instead of Ext.
3680  static Value *promoteOperandForOther(Instruction *Ext,
3681  TypePromotionTransaction &TPT,
3682  InstrToOrigTy &PromotedInsts,
3683  unsigned &CreatedInstsCost,
3686  const TargetLowering &TLI, bool IsSExt);
3687 
3688  /// \see promoteOperandForOther.
3689  static Value *signExtendOperandForOther(
3690  Instruction *Ext, TypePromotionTransaction &TPT,
3691  InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3693  SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3694  return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
3695  Exts, Truncs, TLI, true);
3696  }
3697 
3698  /// \see promoteOperandForOther.
3699  static Value *zeroExtendOperandForOther(
3700  Instruction *Ext, TypePromotionTransaction &TPT,
3701  InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3703  SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3704  return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
3705  Exts, Truncs, TLI, false);
3706  }
3707 
3708 public:
3709  /// Type for the utility function that promotes the operand of Ext.
3710  using Action = Value *(*)(Instruction *Ext, TypePromotionTransaction &TPT,
3711  InstrToOrigTy &PromotedInsts,
3712  unsigned &CreatedInstsCost,
3715  const TargetLowering &TLI);
3716 
3717  /// Given a sign/zero extend instruction \p Ext, return the appropriate
3718  /// action to promote the operand of \p Ext instead of using Ext.
3719  /// \return NULL if no promotable action is possible with the current
3720  /// sign extension.
3721  /// \p InsertedInsts keeps track of all the instructions inserted by the
3722  /// other CodeGenPrepare optimizations. This information is important
3723  /// because we do not want to promote these instructions as CodeGenPrepare
3724  /// will reinsert them later. Thus creating an infinite loop: create/remove.
3725  /// \p PromotedInsts maps the instructions to their type before promotion.
3726  static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedInsts,
3727  const TargetLowering &TLI,
3728  const InstrToOrigTy &PromotedInsts);
3729 };
3730 
3731 } // end anonymous namespace
3732 
3733 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
3734  Type *ConsideredExtType,
3735  const InstrToOrigTy &PromotedInsts,
3736  bool IsSExt) {
3737  // The promotion helper does not know how to deal with vector types yet.
3738  // To be able to fix that, we would need to fix the places where we
3739  // statically extend, e.g., constants and such.
3740  if (Inst->getType()->isVectorTy())
3741  return false;
3742 
3743  // We can always get through zext.
3744  if (isa<ZExtInst>(Inst))
3745  return true;
3746 
3747  // sext(sext) is ok too.
3748  if (IsSExt && isa<SExtInst>(Inst))
3749  return true;
3750 
3751  // We can get through binary operator, if it is legal. In other words, the
3752  // binary operator must have a nuw or nsw flag.
3753  const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
3754  if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
3755  ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
3756  (IsSExt && BinOp->hasNoSignedWrap())))
3757  return true;
3758 
3759  // ext(and(opnd, cst)) --> and(ext(opnd), ext(cst))
3760  if ((Inst->getOpcode() == Instruction::And ||
3761  Inst->getOpcode() == Instruction::Or))
3762  return true;
3763 
3764  // ext(xor(opnd, cst)) --> xor(ext(opnd), ext(cst))
3765  if (Inst->getOpcode() == Instruction::Xor) {
3766  const ConstantInt *Cst = dyn_cast<ConstantInt>(Inst->getOperand(1));
3767  // Make sure it is not a NOT.
3768  if (Cst && !Cst->getValue().isAllOnesValue())
3769  return true;
3770  }
3771 
3772  // zext(shrl(opnd, cst)) --> shrl(zext(opnd), zext(cst))
3773  // It may change a poisoned value into a regular value, like
3774  // zext i32 (shrl i8 %val, 12) --> shrl i32 (zext i8 %val), 12
3775  // poisoned value regular value
3776  // It should be OK since undef covers valid value.
3777  if (Inst->getOpcode() == Instruction::LShr && !IsSExt)
3778  return true;
3779 
3780  // and(ext(shl(opnd, cst)), cst) --> and(shl(ext(opnd), ext(cst)), cst)
3781  // It may change a poisoned value into a regular value, like
3782  // zext i32 (shl i8 %val, 12) --> shl i32 (zext i8 %val), 12
3783  // poisoned value regular value
3784  // It should be OK since undef covers valid value.
3785  if (Inst->getOpcode() == Instruction::Shl && Inst->hasOneUse()) {
3786  const Instruction *ExtInst =
3787  dyn_cast<const Instruction>(*Inst->user_begin());
3788  if (ExtInst->hasOneUse()) {
3789  const Instruction *AndInst =
3790  dyn_cast<const Instruction>(*ExtInst->user_begin());
3791  if (AndInst && AndInst->getOpcode() == Instruction::And) {
3792  const ConstantInt *Cst = dyn_cast<ConstantInt>(AndInst->getOperand(1));
3793  if (Cst &&
3794  Cst->getValue().isIntN(Inst->getType()->getIntegerBitWidth()))
3795  return true;
3796  }
3797  }
3798  }
3799 
3800  // Check if we can do the following simplification.
3801  // ext(trunc(opnd)) --> ext(opnd)
3802  if (!isa<TruncInst>(Inst))
3803  return false;
3804 
3805  Value *OpndVal = Inst->getOperand(0);
3806  // Check if we can use this operand in the extension.
3807  // If the type is larger than the result type of the extension, we cannot.
3808  if (!OpndVal->getType()->isIntegerTy() ||
3809  OpndVal->getType()->getIntegerBitWidth() >
3810  ConsideredExtType->getIntegerBitWidth())
3811  return false;
3812 
3813  // If the operand of the truncate is not an instruction, we will not have
3814  // any information on the dropped bits.
3815  // (Actually we could for constant but it is not worth the extra logic).
3816  Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
3817  if (!Opnd)
3818  return false;
3819 
3820  // Check if the source of the type is narrow enough.
3821  // I.e., check that trunc just drops extended bits of the same kind of
3822  // the extension.
3823  // #1 get the type of the operand and check the kind of the extended bits.
3824  const Type *OpndType = getOrigType(PromotedInsts, Opnd, IsSExt);
3825  if (OpndType)
3826  ;
3827  else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd)))
3828  OpndType = Opnd->getOperand(0)->getType();
3829  else
3830  return false;
3831 
3832  // #2 check that the truncate just drops extended bits.
3833  return Inst->getType()->getIntegerBitWidth() >=
3834  OpndType->getIntegerBitWidth();
3835 }
3836 
3837 TypePromotionHelper::Action TypePromotionHelper::getAction(
3838  Instruction *Ext, const SetOfInstrs &InsertedInsts,
3839  const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
3840  assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
3841  "Unexpected instruction type");
3842  Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0));
3843  Type *ExtTy = Ext->getType();
3844  bool IsSExt = isa<SExtInst>(Ext);
3845  // If the operand of the extension is not an instruction, we cannot
3846  // get through.
3847  // If it, check we can get through.
3848  if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt))
3849  return nullptr;
3850 
3851  // Do not promote if the operand has been added by codegenprepare.
3852  // Otherwise, it means we are undoing an optimization that is likely to be
3853  // redone, thus causing potential infinite loop.
3854  if (isa<TruncInst>(ExtOpnd) && InsertedInsts.count(ExtOpnd))
3855  return nullptr;
3856 
3857  // SExt or Trunc instructions.
3858  // Return the related handler.
3859  if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) ||
3860  isa<ZExtInst>(ExtOpnd))
3861  return promoteOperandForTruncAndAnyExt;
3862 
3863  // Regular instruction.
3864  // Abort early if we will have to insert non-free instructions.
3865  if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType()))
3866  return nullptr;
3867  return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
3868 }
3869 
3870 Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
3871  Instruction *SExt, TypePromotionTransaction &TPT,
3872  InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3874  SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3875  // By construction, the operand of SExt is an instruction. Otherwise we cannot
3876  // get through it and this method should not be called.
3877  Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
3878  Value *ExtVal = SExt;
3879  bool HasMergedNonFreeExt = false;
3880  if (isa<ZExtInst>(SExtOpnd)) {
3881  // Replace s|zext(zext(opnd))
3882  // => zext(opnd).
3883  HasMergedNonFreeExt = !TLI.isExtFree(SExtOpnd);
3884  Value *ZExt =
3885  TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
3886  TPT.replaceAllUsesWith(SExt, ZExt);
3887  TPT.eraseInstruction(SExt);
3888  ExtVal = ZExt;
3889  } else {
3890  // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
3891  // => z|sext(opnd).
3892  TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
3893  }
3894  CreatedInstsCost = 0;
3895 
3896  // Remove dead code.
3897  if (SExtOpnd->use_empty())
3898  TPT.eraseInstruction(SExtOpnd);
3899 
3900  // Check if the extension is still needed.
3901  Instruction *ExtInst = dyn_cast<Instruction>(ExtVal);
3902  if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) {
3903  if (ExtInst) {
3904  if (Exts)
3905  Exts->push_back(ExtInst);
3906  CreatedInstsCost = !TLI.isExtFree(ExtInst) && !HasMergedNonFreeExt;
3907  }
3908  return ExtVal;
3909  }
3910 
3911  // At this point we have: ext ty opnd to ty.
3912  // Reassign the uses of ExtInst to the opnd and remove ExtInst.
3913  Value *NextVal = ExtInst->getOperand(0);
3914  TPT.eraseInstruction(ExtInst, NextVal);
3915  return NextVal;
3916 }
3917 
3918 Value *TypePromotionHelper::promoteOperandForOther(
3919  Instruction *Ext, TypePromotionTransaction &TPT,
3920  InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3922  SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI,
3923  bool IsSExt) {
3924  // By construction, the operand of Ext is an instruction. Otherwise we cannot
3925  // get through it and this method should not be called.
3926  Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0));
3927  CreatedInstsCost = 0;
3928  if (!ExtOpnd->hasOneUse()) {
3929  // ExtOpnd will be promoted.
3930  // All its uses, but Ext, will need to use a truncated value of the
3931  // promoted version.
3932  // Create the truncate now.
3933  Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType());
3934  if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) {
3935  // Insert it just after the definition.
3936  ITrunc->moveAfter(ExtOpnd);
3937  if (Truncs)
3938  Truncs->push_back(ITrunc);
3939  }
3940 
3941  TPT.replaceAllUsesWith(ExtOpnd, Trunc);
3942  // Restore the operand of Ext (which has been replaced by the previous call
3943  // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
3944  TPT.setOperand(Ext, 0, ExtOpnd);
3945  }
3946 
3947  // Get through the Instruction:
3948  // 1. Update its type.
3949  // 2. Replace the uses of Ext by Inst.
3950  // 3. Extend each operand that needs to be extended.
3951 
3952  // Remember the original type of the instruction before promotion.
3953  // This is useful to know that the high bits are sign extended bits.
3954  addPromotedInst(PromotedInsts, ExtOpnd, IsSExt);
3955  // Step #1.
3956  TPT.mutateType(ExtOpnd, Ext->getType());
3957  // Step #2.
3958  TPT.replaceAllUsesWith(Ext, ExtOpnd);
3959  // Step #3.
3960  Instruction *ExtForOpnd = Ext;
3961 
3962  LLVM_DEBUG(dbgs() << "Propagate Ext to operands\n");
3963  for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
3964  ++OpIdx) {
3965  LLVM_DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n');
3966  if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() ||
3967  !shouldExtOperand(ExtOpnd, OpIdx)) {
3968  LLVM_DEBUG(dbgs() << "No need to propagate\n");
3969  continue;
3970  }
3971  // Check if we can statically extend the operand.
3972  Value *Opnd = ExtOpnd->getOperand(OpIdx);
3973  if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
3974  LLVM_DEBUG(dbgs() << "Statically extend\n");
3975  unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
3976  APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth)
3977  : Cst->getValue().zext(BitWidth);
3978  TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal));
3979  continue;
3980  }
3981  // UndefValue are typed, so we have to statically sign extend them.
3982  if (isa<UndefValue>(Opnd)) {
3983  LLVM_DEBUG(dbgs() << "Statically extend\n");
3984  TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType()));
3985  continue;
3986  }
3987 
3988  // Otherwise we have to explicitly sign extend the operand.
3989  // Check if Ext was reused to extend an operand.
3990  if (!ExtForOpnd) {
3991  // If yes, create a new one.
3992  LLVM_DEBUG(dbgs() << "More operands to ext\n");
3993  Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType())
3994  : TPT.createZExt(Ext, Opnd, Ext->getType());
3995  if (!isa<Instruction>(ValForExtOpnd)) {
3996  TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd);
3997  continue;
3998  }
3999  ExtForOpnd = cast<Instruction>(ValForExtOpnd);
4000  }
4001  if (Exts)
4002  Exts->push_back(ExtForOpnd);
4003  TPT.setOperand(ExtForOpnd, 0, Opnd);
4004 
4005  // Move the sign extension before the insertion point.
4006  TPT.moveBefore(ExtForOpnd, ExtOpnd);
4007  TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd);
4008  CreatedInstsCost += !TLI.isExtFree(ExtForOpnd);
4009  // If more sext are required, new instructions will have to be created.
4010  ExtForOpnd = nullptr;
4011  }
4012  if (ExtForOpnd == Ext) {
4013  LLVM_DEBUG(dbgs() << "Extension is useless now\n");
4014  TPT.eraseInstruction(Ext);
4015  }
4016  return ExtOpnd;
4017 }
4018 
4019 /// Check whether or not promoting an instruction to a wider type is profitable.
4020 /// \p NewCost gives the cost of extension instructions created by the
4021 /// promotion.
4022 /// \p OldCost gives the cost of extension instructions before the promotion
4023 /// plus the number of instructions that have been
4024 /// matched in the addressing mode the promotion.
4025 /// \p PromotedOperand is the value that has been promoted.
4026 /// \return True if the promotion is profitable, false otherwise.
4027 bool AddressingModeMatcher::isPromotionProfitable(
4028  unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const {
4029  LLVM_DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCost
4030  << '\n');
4031  // The cost of the new extensions is greater than the cost of the
4032  // old extension plus what we folded.
4033  // This is not profitable.
4034  if (NewCost > OldCost)
4035  return false;
4036  if (NewCost < OldCost)
4037  return true;
4038  // The promotion is neutral but it may help folding the sign extension in
4039  // loads for instance.
4040  // Check that we did not create an illegal instruction.
4041  return isPromotedInstructionLegal(TLI, DL, PromotedOperand);
4042 }
4043 
4044 /// Given an instruction or constant expr, see if we can fold the operation
4045 /// into the addressing mode. If so, update the addressing mode and return
4046 /// true, otherwise return false without modifying AddrMode.
4047 /// If \p MovedAway is not NULL, it contains the information of whether or
4048 /// not AddrInst has to be folded into the addressing mode on success.
4049 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
4050 /// because it has been moved away.
4051 /// Thus AddrInst must not be added in the matched instructions.
4052 /// This state can happen when AddrInst is a sext, since it may be moved away.
4053 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
4054 /// not be referenced anymore.
4055 bool AddressingModeMatcher::matchOperationAddr(User *AddrInst, unsigned Opcode,
4056  unsigned Depth,
4057  bool *MovedAway) {
4058  // Avoid exponential behavior on extremely deep expression trees.
4059  if (Depth >= 5) return false;
4060 
4061  // By default, all matched instructions stay in place.
4062  if (MovedAway)
4063  *MovedAway = false;
4064 
4065  switch (Opcode) {
4066  case Instruction::PtrToInt:
4067  // PtrToInt is always a noop, as we know that the int type is pointer sized.
4068  return matchAddr(AddrInst->getOperand(0), Depth);
4069  case Instruction::IntToPtr: {
4070  auto AS = AddrInst->getType()->getPointerAddressSpace();
4071  auto PtrTy = MVT::getIntegerVT(DL.getPointerSizeInBits(AS));
4072  // This inttoptr is a no-op if the integer type is pointer sized.
4073  if (TLI.getValueType(DL, AddrInst->getOperand(0)->getType()) == PtrTy)
4074  return matchAddr(AddrInst->getOperand(0), Depth);
4075  return false;
4076  }
4077  case Instruction::BitCast:
4078  // BitCast is always a noop, and we can handle it as long as it is
4079  // int->int or pointer->pointer (we don't want int<->fp or something).
4080  if (AddrInst->getOperand(0)->getType()->isIntOrPtrTy() &&
4081  // Don't touch identity bitcasts. These were probably put here by LSR,
4082  // and we don't want to mess around with them. Assume it knows what it
4083  // is doing.
4084  AddrInst->getOperand(0)->getType() != AddrInst->getType())
4085  return matchAddr(AddrInst->getOperand(0), Depth);
4086  return false;
4087  case Instruction::AddrSpaceCast: {
4088  unsigned SrcAS
4089  = AddrInst->getOperand(0)->getType()->getPointerAddressSpace();
4090  unsigned DestAS = AddrInst->getType()->getPointerAddressSpace();
4091  if (TLI.isNoopAddrSpaceCast(SrcAS, DestAS))
4092  return matchAddr(AddrInst->getOperand(0), Depth);
4093  return false;
4094  }
4095  case Instruction::Add: {
4096  // Check to see if we can merge in the RHS then the LHS. If so, we win.
4097  ExtAddrMode BackupAddrMode = AddrMode;
4098  unsigned OldSize = AddrModeInsts.size();
4099  // Start a transaction at this point.
4100  // The LHS may match but not the RHS.
4101  // Therefore, we need a higher level restoration point to undo partially
4102  // matched operation.
4103  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4104  TPT.getRestorationPoint();
4105 
4106  AddrMode.InBounds = false;
4107  if (matchAddr(AddrInst->getOperand(1), Depth+1) &&
4108  matchAddr(AddrInst->getOperand(0), Depth+1))
4109  return true;
4110 
4111  // Restore the old addr mode info.
4112  AddrMode = BackupAddrMode;
4113  AddrModeInsts.resize(OldSize);
4114  TPT.rollback(LastKnownGood);
4115 
4116  // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
4117  if (matchAddr(AddrInst->getOperand(0), Depth+1) &&
4118  matchAddr(AddrInst->getOperand(1), Depth+1))
4119  return true;
4120 
4121  // Otherwise we definitely can't merge the ADD in.
4122  AddrMode = BackupAddrMode;
4123  AddrModeInsts.resize(OldSize);
4124  TPT.rollback(LastKnownGood);
4125  break;
4126  }
4127  //case Instruction::Or:
4128  // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
4129  //break;
4130  case Instruction::Mul:
4131  case Instruction::Shl: {
4132  // Can only handle X*C and X << C.
4133  AddrMode.InBounds = false;
4134  ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
4135  if (!RHS || RHS->getBitWidth() > 64)
4136  return false;
4137  int64_t Scale = RHS->getSExtValue();
4138  if (Opcode == Instruction::Shl)
4139  Scale = 1LL << Scale;
4140 
4141  return matchScaledValue(AddrInst->getOperand(0), Scale, Depth);
4142  }
4143  case Instruction::GetElementPtr: {
4144  // Scan the GEP. We check it if it contains constant offsets and at most
4145  // one variable offset.
4146  int VariableOperand = -1;
4147  unsigned VariableScale = 0;
4148 
4149  int64_t ConstantOffset = 0;
4150  gep_type_iterator GTI = gep_type_begin(AddrInst);
4151  for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
4152  if (StructType *STy = GTI.getStructTypeOrNull()) {
4153  const StructLayout *SL = DL.getStructLayout(STy);
4154  unsigned Idx =
4155  cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
4156  ConstantOffset += SL->getElementOffset(Idx);
4157  } else {
4158  uint64_t TypeSize = DL.getTypeAllocSize(GTI.getIndexedType());
4159  if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
4160  const APInt &CVal = CI->getValue();
4161  if (CVal.getMinSignedBits() <= 64) {
4162  ConstantOffset += CVal.getSExtValue() * TypeSize;
4163  continue;
4164  }
4165  }
4166  if (TypeSize) { // Scales of zero don't do anything.
4167  // We only allow one variable index at the moment.
4168  if (VariableOperand != -1)
4169  return false;
4170 
4171  // Remember the variable index.
4172  VariableOperand = i;
4173  VariableScale = TypeSize;
4174  }
4175  }
4176  }
4177 
4178  // A common case is for the GEP to only do a constant offset. In this case,
4179  // just add it to the disp field and check validity.
4180  if (VariableOperand == -1) {
4181  AddrMode.BaseOffs += ConstantOffset;
4182  if (ConstantOffset == 0 ||
4183  TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) {
4184  // Check to see if we can fold the base pointer in too.
4185  if (matchAddr(AddrInst->getOperand(0), Depth+1)) {
4186  if (!cast<GEPOperator>(AddrInst)->isInBounds())
4187  AddrMode.InBounds = false;
4188  return true;
4189  }
4190  } else if (EnableGEPOffsetSplit && isa<GetElementPtrInst>(AddrInst) &&
4191  TLI.shouldConsiderGEPOffsetSplit() && Depth == 0 &&
4192  ConstantOffset > 0) {
4193  // Record GEPs with non-zero offsets as candidates for splitting in the
4194  // event that the offset cannot fit into the r+i addressing mode.
4195  // Simple and common case that only one GEP is used in calculating the
4196  // address for the memory access.
4197  Value *Base = AddrInst->getOperand(0);
4198  auto *BaseI = dyn_cast<Instruction>(Base);
4199  auto *GEP = cast<GetElementPtrInst>(AddrInst);
4200  if (isa<Argument>(Base) || isa<GlobalValue>(Base) ||
4201  (BaseI && !isa<CastInst>(BaseI) &&
4202  !isa<GetElementPtrInst>(BaseI))) {
4203  // If the base is an instruction, make sure the GEP is not in the same
4204  // basic block as the base. If the base is an argument or global
4205  // value, make sure the GEP is not in the entry block. Otherwise,
4206  // instruction selection can undo the split. Also make sure the
4207  // parent block allows inserting non-PHI instructions before the
4208  // terminator.
4209  BasicBlock *Parent =
4210  BaseI ? BaseI->getParent() : &GEP->getFunction()->getEntryBlock();
4211  if (GEP->getParent() != Parent && !Parent->getTerminator()->isEHPad())
4212  LargeOffsetGEP = std::make_pair(GEP, ConstantOffset);
4213  }
4214  }
4215  AddrMode.BaseOffs -= ConstantOffset;
4216  return false;
4217  }
4218 
4219  // Save the valid addressing mode in case we can't match.
4220  ExtAddrMode BackupAddrMode = AddrMode;
4221  unsigned OldSize = AddrModeInsts.size();
4222 
4223  // See if the scale and offset amount is valid for this target.
4224  AddrMode.BaseOffs += ConstantOffset;
4225  if (!cast<GEPOperator>(AddrInst)->isInBounds())
4226  AddrMode.InBounds = false;
4227 
4228  // Match the base operand of the GEP.
4229  if (!matchAddr(AddrInst->getOperand(0), Depth+1)) {
4230  // If it couldn't be matched, just stuff the value in a register.
4231  if (AddrMode.HasBaseReg) {
4232  AddrMode = BackupAddrMode;
4233  AddrModeInsts.resize(OldSize);
4234  return false;
4235  }
4236  AddrMode.HasBaseReg = true;
4237  AddrMode.BaseReg = AddrInst->getOperand(0);
4238  }
4239 
4240  // Match the remaining variable portion of the GEP.
4241  if (!matchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
4242  Depth)) {
4243  // If it couldn't be matched, try stuffing the base into a register
4244  // instead of matching it, and retrying the match of the scale.
4245  AddrMode = BackupAddrMode;
4246  AddrModeInsts.resize(OldSize);
4247  if (AddrMode.HasBaseReg)
4248  return false;
4249  AddrMode.HasBaseReg = true;
4250  AddrMode.BaseReg = AddrInst->getOperand(0);
4251  AddrMode.BaseOffs += ConstantOffset;
4252  if (!matchScaledValue(AddrInst->getOperand(VariableOperand),
4253  VariableScale, Depth)) {
4254  // If even that didn't work, bail.
4255  AddrMode = BackupAddrMode;
4256  AddrModeInsts.resize(OldSize);
4257  return false;
4258  }
4259  }
4260 
4261  return true;
4262  }
4263  case Instruction::SExt:
4264  case Instruction::ZExt: {
4265  Instruction *Ext = dyn_cast<Instruction>(AddrInst);
4266  if (!Ext)
4267  return false;
4268 
4269  // Try to move this ext out of the way of the addressing mode.
4270  // Ask for a method for doing so.
4271  TypePromotionHelper::Action TPH =
4272  TypePromotionHelper::getAction(Ext, InsertedInsts, TLI, PromotedInsts);
4273  if (!TPH)
4274  return false;
4275 
4276  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4277  TPT.getRestorationPoint();
4278  unsigned CreatedInstsCost = 0;
4279  unsigned ExtCost = !TLI.isExtFree(Ext);
4280  Value *PromotedOperand =
4281  TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI);
4282  // SExt has been moved away.
4283  // Thus either it will be rematched later in the recursive calls or it is
4284  // gone. Anyway, we must not fold it into the addressing mode at this point.
4285  // E.g.,
4286  // op = add opnd, 1
4287  // idx = ext op
4288  // addr = gep base, idx
4289  // is now:
4290  // promotedOpnd = ext opnd <- no match here
4291  // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
4292  // addr = gep base, op <- match
4293  if (MovedAway)
4294  *MovedAway = true;
4295 
4296  assert(PromotedOperand &&
4297  "TypePromotionHelper should have filtered out those cases");
4298 
4299  ExtAddrMode BackupAddrMode = AddrMode;
4300  unsigned OldSize = AddrModeInsts.size();
4301 
4302  if (!matchAddr(PromotedOperand, Depth) ||
4303  // The total of the new cost is equal to the cost of the created
4304  // instructions.
4305  // The total of the old cost is equal to the cost of the extension plus
4306  // what we have saved in the addressing mode.
4307  !isPromotionProfitable(CreatedInstsCost,
4308  ExtCost + (AddrModeInsts.size() - OldSize),
4309  PromotedOperand)) {
4310  AddrMode = BackupAddrMode;
4311  AddrModeInsts.resize(OldSize);
4312  LLVM_DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
4313  TPT.rollback(LastKnownGood);
4314  return false;
4315  }
4316  return true;
4317  }
4318  }
4319  return false;
4320 }
4321 
4322 /// If we can, try to add the value of 'Addr' into the current addressing mode.
4323 /// If Addr can't be added to AddrMode this returns false and leaves AddrMode
4324 /// unmodified. This assumes that Addr is either a pointer type or intptr_t
4325 /// for the target.
4326 ///
4327 bool AddressingModeMatcher::matchAddr(Value *Addr, unsigned Depth) {
4328  // Start a transaction at this point that we will rollback if the matching
4329  // fails.
4330  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4331  TPT.getRestorationPoint();
4332  if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
4333  // Fold in immediates if legal for the target.
4334  AddrMode.BaseOffs += CI->getSExtValue();
4335  if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4336  return true;
4337  AddrMode.BaseOffs -= CI->getSExtValue();
4338  } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
4339  // If this is a global variable, try to fold it into the addressing mode.
4340  if (!AddrMode.BaseGV) {
4341  AddrMode.BaseGV = GV;
4342  if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4343  return true;
4344  AddrMode.BaseGV = nullptr;
4345  }
4346  } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
4347  ExtAddrMode BackupAddrMode = AddrMode;
4348  unsigned OldSize = AddrModeInsts.size();
4349 
4350  // Check to see if it is possible to fold this operation.
4351  bool MovedAway = false;
4352  if (matchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
4353  // This instruction may have been moved away. If so, there is nothing
4354  // to check here.
4355  if (MovedAway)
4356  return true;
4357  // Okay, it's possible to fold this. Check to see if it is actually
4358  // *profitable* to do so. We use a simple cost model to avoid increasing
4359  // register pressure too much.
4360  if (I->hasOneUse() ||
4361  isProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
4362  AddrModeInsts.push_back(I);
4363  return true;
4364  }
4365 
4366  // It isn't profitable to do this, roll back.
4367  //cerr << "NOT FOLDING: " << *I;
4368  AddrMode = BackupAddrMode;
4369  AddrModeInsts.resize(OldSize);
4370  TPT.rollback(LastKnownGood);
4371  }
4372  } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
4373  if (matchOperationAddr(CE, CE->getOpcode(), Depth))
4374  return true;
4375  TPT.rollback(LastKnownGood);
4376  } else if (isa<ConstantPointerNull>(Addr)) {
4377  // Null pointer gets folded without affecting the addressing mode.
4378  return true;
4379  }
4380 
4381  // Worse case, the target should support [reg] addressing modes. :)
4382  if (!AddrMode.HasBaseReg) {
4383  AddrMode.HasBaseReg = true;
4384  AddrMode.BaseReg = Addr;
4385  // Still check for legality in case the target supports [imm] but not [i+r].
4386  if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4387  return true;
4388  AddrMode.HasBaseReg = false;
4389  AddrMode.BaseReg = nullptr;
4390  }
4391 
4392  // If the base register is already taken, see if we can do [r+r].
4393  if (AddrMode.Scale == 0) {
4394  AddrMode.Scale = 1;
4395  AddrMode.ScaledReg = Addr;
4396  if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4397  return true;
4398  AddrMode.Scale = 0;
4399  AddrMode.ScaledReg = nullptr;
4400  }
4401  // Couldn't match.
4402  TPT.rollback(LastKnownGood);
4403  return false;
4404 }
4405 
4406 /// Check to see if all uses of OpVal by the specified inline asm call are due
4407 /// to memory operands. If so, return true, otherwise return false.
4408 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
4409  const TargetLowering &TLI,
4410  const TargetRegisterInfo &TRI) {
4411  const Function *F = CI->getFunction();
4412  TargetLowering::AsmOperandInfoVector TargetConstraints =
4414  ImmutableCallSite(CI));
4415 
4416  for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
4417  TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
4418 
4419  // Compute the constraint code and ConstraintType to use.
4420  TLI.ComputeConstraintToUse(OpInfo, SDValue());
4421 
4422  // If this asm operand is our Value*, and if it isn't an indirect memory
4423  // operand, we can't fold it!
4424  if (OpInfo.CallOperandVal == OpVal &&
4426  !OpInfo.isIndirect))
4427  return false;
4428  }
4429 
4430  return true;
4431 }
4432 
4433 // Max number of memory uses to look at before aborting the search to conserve
4434 // compile time.
4435 static constexpr int MaxMemoryUsesToScan = 20;
4436 
4437 /// Recursively walk all the uses of I until we find a memory use.
4438 /// If we find an obviously non-foldable instruction, return true.
4439 /// Add the ultimately found memory instructions to MemoryUses.
4440 static bool FindAllMemoryUses(
4441  Instruction *I,
4442  SmallVectorImpl<std::pair<Instruction *, unsigned>> &MemoryUses,
4443  SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetLowering &TLI,
4444  const TargetRegisterInfo &TRI, int SeenInsts = 0) {
4445  // If we already considered this instruction, we're done.
4446  if (!ConsideredInsts.insert(I).second)
4447  return false;
4448 
4449  // If this is an obviously unfoldable instruction, bail out.
4450  if (!MightBeFoldableInst(I))
4451  return true;
4452 
4453  const bool OptSize = I->getFunction()->hasOptSize();
4454 
4455  // Loop over all the uses, recursively processing them.
4456  for (Use &U : I->uses()) {
4457  // Conservatively return true if we're seeing a large number or a deep chain
4458  // of users. This avoids excessive compilation times in pathological cases.
4459  if (SeenInsts++ >= MaxMemoryUsesToScan)
4460  return true;
4461 
4462  Instruction *UserI = cast<Instruction>(U.getUser());
4463  if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
4464  MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
4465  continue;
4466  }
4467 
4468  if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
4469  unsigned opNo = U.getOperandNo();
4470  if (opNo != StoreInst::getPointerOperandIndex())
4471  return true; // Storing addr, not into addr.
4472  MemoryUses.push_back(std::make_pair(SI, opNo));
4473  continue;
4474  }
4475 
4476  if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(UserI)) {
4477  unsigned opNo = U.getOperandNo();
4479  return true; // Storing addr, not into addr.
4480  MemoryUses.push_back(std::make_pair(RMW, opNo));
4481  continue;
4482  }
4483 
4484  if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(UserI)) {
4485  unsigned opNo = U.getOperandNo();
4487  return true; // Storing addr, not into addr.
4488  MemoryUses.push_back(std::make_pair(CmpX, opNo));
4489  continue;
4490  }
4491 
4492  if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
4493  // If this is a cold call, we can sink the addressing calculation into
4494  // the cold path. See optimizeCallInst
4495  if (!OptSize && CI->hasFnAttr(Attribute::Cold))
4496  continue;
4497 
4499  if (!IA) return true;
4500 
4501  // If this is a memory operand, we're cool, otherwise bail out.
4502  if (!IsOperandAMemoryOperand(CI, IA, I, TLI, TRI))
4503  return true;
4504  continue;
4505  }
4506 
4507  if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TLI, TRI,
4508  SeenInsts))
4509  return true;
4510  }
4511 
4512  return false;
4513 }
4514 
4515 /// Return true if Val is already known to be live at the use site that we're
4516 /// folding it into. If so, there is no cost to include it in the addressing
4517 /// mode. KnownLive1 and KnownLive2 are two values that we know are live at the
4518 /// instruction already.
4519 bool AddressingModeMatcher::valueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
4520  Value *KnownLive2) {
4521  // If Val is either of the known-live values, we know it is live!
4522  if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
4523  return true;
4524 
4525  // All values other than instructions and arguments (e.g. constants) are live.
4526  if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
4527 
4528  // If Val is a constant sized alloca in the entry block, it is live, this is
4529  // true because it is just a reference to the stack/frame pointer, which is
4530  // live for the whole function.
4531  if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
4532  if (AI->isStaticAlloca())
4533  return true;
4534 
4535  // Check to see if this value is already used in the memory instruction's
4536  // block. If so, it's already live into the block at the very least, so we
4537  // can reasonably fold it.
4538  return Val->isUsedInBasicBlock(MemoryInst->getParent());
4539 }
4540 
4541 /// It is possible for the addressing mode of the machine to fold the specified
4542 /// instruction into a load or store that ultimately uses it.
4543 /// However, the specified instruction has multiple uses.
4544 /// Given this, it may actually increase register pressure to fold it
4545 /// into the load. For example, consider this code:
4546 ///
4547 /// X = ...
4548 /// Y = X+1
4549 /// use(Y) -> nonload/store
4550 /// Z = Y+1
4551 /// load Z
4552 ///
4553 /// In this case, Y has multiple uses, and can be folded into the load of Z
4554 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
4555 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
4556 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
4557 /// number of computations either.
4558 ///
4559 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
4560 /// X was live across 'load Z' for other reasons, we actually *would* want to
4561 /// fold the addressing mode in the Z case. This would make Y die earlier.
4562 bool AddressingModeMatcher::
4563 isProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
4564  ExtAddrMode &AMAfter) {
4565  if (IgnoreProfitability) return true;
4566 
4567  // AMBefore is the addressing mode before this instruction was folded into it,
4568  // and AMAfter is the addressing mode after the instruction was folded. Get
4569  // the set of registers referenced by AMAfter and subtract out those
4570  // referenced by AMBefore: this is the set of values which folding in this
4571  // address extends the lifetime of.
4572  //
4573  // Note that there are only two potential values being referenced here,
4574  // BaseReg and ScaleReg (global addresses are always available, as are any
4575  // folded immediates).
4576  Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
4577 
4578  // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
4579  // lifetime wasn't extended by adding this instruction.
4580  if (valueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
4581  BaseReg = nullptr;
4582  if (valueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
4583  ScaledReg = nullptr;
4584 
4585  // If folding this instruction (and it's subexprs) didn't extend any live
4586  // ranges, we're ok with it.
4587  if (!BaseReg && !ScaledReg)
4588  return true;
4589 
4590  // If all uses of this instruction can have the address mode sunk into them,
4591  // we can remove the addressing mode and effectively trade one live register
4592  // for another (at worst.) In this context, folding an addressing mode into
4593  // the use is just a particularly nice way of sinking it.
4595  SmallPtrSet<Instruction*, 16> ConsideredInsts;
4596  if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TLI, TRI))
4597  return false; // Has a non-memory, non-foldable use!
4598 
4599  // Now that we know that all uses of this instruction are part of a chain of
4600  // computation involving only operations that could theoretically be folded
4601  // into a memory use, loop over each of these memory operation uses and see
4602  // if they could *actually* fold the instruction. The assumption is that
4603  // addressing modes are cheap and that duplicating the computation involved
4604  // many times is worthwhile, even on a fastpath. For sinking candidates
4605  // (i.e. cold call sites), this serves as a way to prevent excessive code
4606  // growth since most architectures have some reasonable small and fast way to
4607  // compute an effective address. (i.e LEA on x86)
4608  SmallVector<Instruction*, 32> MatchedAddrModeInsts;
4609  for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
4610  Instruction *User = MemoryUses[i].first;
4611  unsigned OpNo = MemoryUses[i].second;
4612 
4613  // Get the access type of this use. If the use isn't a pointer, we don't
4614  // know what it accesses.
4615  Value *Address = User->getOperand(OpNo);
4616  PointerType *AddrTy = dyn_cast<PointerType>(Address->getType());
4617  if (!AddrTy)
4618  return false;
4619  Type *AddressAccessTy = AddrTy->getElementType();
4620  unsigned AS = AddrTy->getAddressSpace();
4621 
4622  // Do a match against the root of this address, ignoring profitability. This
4623  // will tell us if the addressing mode for the memory operation will
4624  // *actually* cover the shared instruction.
4625  ExtAddrMode Result;
4626  std::pair<AssertingVH<GetElementPtrInst>, int64_t> LargeOffsetGEP(nullptr,
4627  0);
4628  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4629  TPT.getRestorationPoint();
4630  AddressingModeMatcher Matcher(
4631  MatchedAddrModeInsts, TLI, TRI, AddressAccessTy, AS, MemoryInst, Result,
4632  InsertedInsts, PromotedInsts, TPT, LargeOffsetGEP);
4633  Matcher.IgnoreProfitability = true;
4634  bool Success = Matcher.matchAddr(Address, 0);
4635  (void)Success; assert(Success && "Couldn't select *anything*?");
4636 
4637  // The match was to check the profitability, the changes made are not
4638  // part of the original matcher. Therefore, they should be dropped
4639  // otherwise the original matcher will not present the right state.
4640  TPT.rollback(LastKnownGood);
4641 
4642  // If the match didn't cover I, then it won't be shared by it.
4643  if (!is_contained(MatchedAddrModeInsts, I))
4644  return false;
4645 
4646  MatchedAddrModeInsts.clear();
4647  }
4648 
4649  return true;
4650 }
4651 
4652 /// Return true if the specified values are defined in a
4653 /// different basic block than BB.
4654 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
4655  if (Instruction *I = dyn_cast<Instruction>(V))
4656  return I->getParent() != BB;
4657  return false;
4658 }
4659 
4660 /// Sink addressing mode computation immediate before MemoryInst if doing so
4661 /// can be done without increasing register pressure. The need for the
4662 /// register pressure constraint means this can end up being an all or nothing
4663 /// decision for all uses of the same addressing computation.
4664 ///
4665 /// Load and Store Instructions often have addressing modes that can do
4666 /// significant amounts of computation. As such, instruction selection will try
4667 /// to get the load or store to do as much computation as possible for the
4668 /// program. The problem is that isel can only see within a single block. As
4669 /// such, we sink as much legal addressing mode work into the block as possible.
4670 ///
4671 /// This method is used to optimize both load/store and inline asms with memory
4672 /// operands. It's also used to sink addressing computations feeding into cold
4673 /// call sites into their (cold) basic block.
4674 ///
4675 /// The motivation for handling sinking into cold blocks is that doing so can
4676 /// both enable other address mode sinking (by satisfying the register pressure
4677 /// constraint above), and reduce register pressure globally (by removing the
4678 /// addressing mode computation from the fast path entirely.).
4679 bool CodeGenPrepare::optimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
4680  Type *AccessTy, unsigned AddrSpace) {
4681  Value *Repl = Addr;
4682 
4683  // Try to collapse single-value PHI nodes. This is necessary to undo
4684  // unprofitable PRE transformations.
4685  SmallVector<Value*, 8> worklist;
4686  SmallPtrSet<Value*, 16> Visited;
4687  worklist.push_back(Addr);
4688 
4689  // Use a worklist to iteratively look through PHI and select nodes, and
4690  // ensure that the addressing mode obtained from the non-PHI/select roots of
4691  // the graph are compatible.
4692  bool PhiOrSelectSeen = false;
4693  SmallVector<Instruction*, 16> AddrModeInsts;
4694  const SimplifyQuery SQ(*DL, TLInfo);
4695  AddressingModeCombiner AddrModes(SQ, Addr);
4696  TypePromotionTransaction TPT(RemovedInsts);
4697  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4698  TPT.getRestorationPoint();
4699  while (!worklist.empty()) {
4700  Value *V = worklist.back();
4701  worklist.pop_back();
4702 
4703  // We allow traversing cyclic Phi nodes.
4704  // In case of success after this loop we ensure that traversing through
4705  // Phi nodes ends up with all cases to compute address of the form
4706  // BaseGV + Base + Scale * Index + Offset
4707  // where Scale and Offset are constans and BaseGV, Base and Index
4708  // are exactly the same Values in all cases.
4709  // It means that BaseGV, Scale and Offset dominate our memory instruction
4710  // and have the same value as they had in address computation represented
4711  // as Phi. So we can safely sink address computation to memory instruction.
4712  if (!Visited.insert(V).second)
4713  continue;
4714 
4715  // For a PHI node, push all of its incoming values.
4716  if (PHINode *P = dyn_cast<PHINode>(V)) {
4717  for (Value *IncValue : P->incoming_values())
4718  worklist.push_back(IncValue);
4719  PhiOrSelectSeen = true;
4720  continue;
4721  }
4722  // Similar for select.
4723  if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
4724  worklist.push_back(SI->getFalseValue());
4725  worklist.push_back(SI->getTrueValue());
4726  PhiOrSelectSeen = true;
4727  continue;
4728  }
4729 
4730  // For non-PHIs, determine the addressing mode being computed. Note that
4731  // the result may differ depending on what other uses our candidate
4732  // addressing instructions might have.
4733  AddrModeInsts.clear();
4734  std::pair<AssertingVH<GetElementPtrInst>, int64_t> LargeOffsetGEP(nullptr,
4735  0);
4736  ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
4737  V, AccessTy, AddrSpace, MemoryInst, AddrModeInsts, *TLI, *TRI,
4738  InsertedInsts, PromotedInsts, TPT, LargeOffsetGEP);
4739 
4740  GetElementPtrInst *GEP = LargeOffsetGEP.first;
4741  if (GEP && GEP->getParent() != MemoryInst->getParent() &&
4742  !NewGEPBases.count(GEP)) {
4743  // If splitting the underlying data structure can reduce the offset of a
4744  // GEP, collect the GEP. Skip the GEPs that are the new bases of
4745  // previously split data structures.
4746  LargeOffsetGEPMap[GEP->getPointerOperand()].push_back(LargeOffsetGEP);
4747  if (LargeOffsetGEPID.find(GEP) == LargeOffsetGEPID.end())
4748  LargeOffsetGEPID[GEP] = LargeOffsetGEPID.size();
4749  }
4750 
4751  NewAddrMode.OriginalValue = V;
4752  if (!AddrModes.addNewAddrMode(NewAddrMode))
4753  break;
4754  }
4755 
4756  // Try to combine the AddrModes we've collected. If we couldn't collect any,
4757  // or we have multiple but either couldn't combine them or combining them
4758  // wouldn't do anything useful, bail out now.
4759  if (!AddrModes.combineAddrModes()) {
4760  TPT.rollback(LastKnownGood);
4761  return false;
4762  }
4763  TPT.commit();
4764 
4765  // Get the combined AddrMode (or the only AddrMode, if we only had one).
4766  ExtAddrMode AddrMode = AddrModes.getAddrMode();
4767 
4768  // If all the instructions matched are already in this BB, don't do anything.
4769  // If we saw a Phi node then it is not local definitely, and if we saw a select
4770  // then we want to push the address calculation past it even if it's already
4771  // in this BB.
4772  if (!PhiOrSelectSeen && none_of(AddrModeInsts, [&](Value *V) {
4773  return IsNonLocalValue(V, MemoryInst->getParent());
4774  })) {
4775  LLVM_DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode
4776  << "\n");
4777  return false;
4778  }
4779 
4780  // Insert this computation right after this user. Since our caller is
4781  // scanning from the top of the BB to the bottom, reuse of the expr are
4782  // guaranteed to happen later.
4783  IRBuilder<> Builder(MemoryInst);
4784 
4785  // Now that we determined the addressing expression we want to use and know
4786  // that we have to sink it into this block. Check to see if we have already
4787  // done this for some other load/store instr in this block. If so, reuse
4788  // the computation. Before attempting reuse, check if the address is valid
4789  // as it may have been erased.
4790 
4791  WeakTrackingVH SunkAddrVH = SunkAddrs[Addr];
4792 
4793  Value * SunkAddr = SunkAddrVH.pointsToAliveValue() ? SunkAddrVH : nullptr;
4794  if (SunkAddr) {
4795  LLVM_DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode
4796  << " for " << *MemoryInst << "\n");
4797  if (SunkAddr->getType() != Addr->getType())
4798  SunkAddr = Builder.CreatePointerCast(SunkAddr, Addr->getType());
4799  } else if (AddrSinkUsingGEPs ||
4800  (!AddrSinkUsingGEPs.getNumOccurrences() && TM && TTI->useAA())) {
4801  // By default, we use the GEP-based method when AA is used later. This
4802  // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
4803  LLVM_DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode
4804  << " for " << *MemoryInst << "\n");
4805  Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
4806  Value *ResultPtr = nullptr, *ResultIndex = nullptr;
4807 
4808  // First, find the pointer.
4809  if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
4810  ResultPtr = AddrMode.BaseReg;
4811  AddrMode.BaseReg = nullptr;
4812  }
4813 
4814  if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
4815  // We can't add more than one pointer together, nor can we scale a
4816  // pointer (both of which seem meaningless).
4817  if (ResultPtr || AddrMode.Scale != 1)
4818  return false;
4819 
4820  ResultPtr = AddrMode.ScaledReg;
4821  AddrMode.Scale = 0;
4822  }
4823 
4824  // It is only safe to sign extend the BaseReg if we know that the math
4825  // required to create it did not overflow before we extend it. Since
4826  // the original IR value was tossed in favor of a constant back when
4827  // the AddrMode was created we need to bail out gracefully if widths
4828  // do not match instead of extending it.
4829  //
4830  // (See below for code to add the scale.)
4831  if (AddrMode.Scale) {
4832  Type *ScaledRegTy = AddrMode.ScaledReg->getType();
4833  if (cast<IntegerType>(IntPtrTy)->getBitWidth() >
4834  cast<IntegerType>(ScaledRegTy)->getBitWidth())
4835  return false;
4836  }
4837 
4838  if (AddrMode.BaseGV) {
4839  if (ResultPtr)
4840  return false;
4841 
4842  ResultPtr = AddrMode.BaseGV;
4843  }
4844 
4845  // If the real base value actually came from an inttoptr, then the matcher
4846  // will look through it and provide only the integer value. In that case,
4847  // use it here.
4848  if (!DL->isNonIntegralPointerType(Addr->getType())) {
4849  if (!ResultPtr && AddrMode.BaseReg) {
4850  ResultPtr = Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(),
4851  "sunkaddr");
4852  AddrMode.BaseReg = nullptr;
4853  } else if (!ResultPtr && AddrMode.Scale == 1) {
4854  ResultPtr = Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(),
4855  "sunkaddr");
4856  AddrMode.Scale = 0;
4857  }
4858  }
4859 
4860  if (!ResultPtr &&
4861  !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
4862  SunkAddr = Constant::getNullValue(Addr->getType());
4863  } else if (!ResultPtr) {
4864  return false;
4865  } else {
4866  Type *I8PtrTy =
4867  Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
4868  Type *I8Ty = Builder.getInt8Ty();
4869 
4870  // Start with the base register. Do this first so that subsequent address
4871  // matching finds it last, which will prevent it from trying to match it
4872  // as the scaled value in case it happens to be a mul. That would be
4873  // problematic if we've sunk a different mul for the scale, because then
4874  // we'd end up sinking both muls.
4875  if (AddrMode.BaseReg) {
4876  Value *V = AddrMode.BaseReg;
4877  if (V->getType() != IntPtrTy)
4878  V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
4879 
4880  ResultIndex = V;
4881  }
4882 
4883  // Add the scale value.
4884  if (AddrMode.Scale) {
4885  Value *V = AddrMode.ScaledReg;
4886  if (V->getType() == IntPtrTy) {
4887  // done.
4888  } else {
4889  assert(cast<IntegerType>(IntPtrTy)->getBitWidth() <
4890  cast<IntegerType>(V->getType())->getBitWidth() &&
4891  "We can't transform if ScaledReg is too narrow");
4892  V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
4893  }
4894 
4895  if (AddrMode.Scale != 1)
4896  V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
4897  "sunkaddr");
4898  if (ResultIndex)
4899  ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
4900  else
4901  ResultIndex = V;
4902  }
4903 
4904  // Add in the Base Offset if present.
4905  if (AddrMode.BaseOffs) {
4906  Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
4907  if (ResultIndex) {
4908  // We need to add this separately from the scale above to help with
4909  // SDAG consecutive load/store merging.
4910  if (ResultPtr->getType() != I8PtrTy)
4911  ResultPtr = Builder.CreatePointerCast(ResultPtr, I8PtrTy);
4912  ResultPtr =
4913  AddrMode.InBounds
4914  ? Builder.CreateInBoundsGEP(I8Ty, ResultPtr, ResultIndex,
4915  "sunkaddr")
4916  : Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
4917  }
4918 
4919  ResultIndex = V;
4920  }
4921 
4922  if (!ResultIndex) {
4923  SunkAddr = ResultPtr;
4924  } else {
4925  if (ResultPtr->getType() != I8PtrTy)
4926  ResultPtr = Builder.CreatePointerCast(ResultPtr, I8PtrTy);
4927  SunkAddr =
4928  AddrMode.InBounds
4929  ? Builder.CreateInBoundsGEP(I8Ty, ResultPtr, ResultIndex,
4930  "sunkaddr")
4931  : Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
4932  }
4933 
4934  if (SunkAddr->getType() != Addr->getType())
4935  SunkAddr = Builder.CreatePointerCast(SunkAddr, Addr->getType());
4936  }
4937  } else {
4938  // We'd require a ptrtoint/inttoptr down the line, which we can't do for
4939  // non-integral pointers, so in that case bail out now.
4940  Type *BaseTy = AddrMode.BaseReg ? AddrMode.BaseReg->getType() : nullptr;
4941  Type *ScaleTy = AddrMode.Scale ? AddrMode.ScaledReg->getType() : nullptr;
4942  PointerType *BasePtrTy = dyn_cast_or_null<PointerType>(BaseTy);
4943  PointerType *ScalePtrTy = dyn_cast_or_null<PointerType>(ScaleTy);
4944  if (DL->isNonIntegralPointerType(Addr->getType()) ||
4945  (BasePtrTy && DL->isNonIntegralPointerType(BasePtrTy)) ||
4946  (ScalePtrTy && DL->isNonIntegralPointerType(ScalePtrTy)) ||
4947  (AddrMode.BaseGV &&
4948  DL->isNonIntegralPointerType(AddrMode.BaseGV->getType())))
4949  return false;
4950 
4951  LLVM_DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode
4952  << " for " << *MemoryInst << "\n");
4953  Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
4954  Value *Result = nullptr;
4955 
4956  // Start with the base register. Do this first so that subsequent address
4957  // matching finds it last, which will prevent it from trying to match it
4958  // as the scaled value in case it happens to be a mul. That would be
4959  // problematic if we've sunk a different mul for the scale, because then
4960  // we'd end up sinking both muls.
4961  if (AddrMode.BaseReg) {
4962  Value *V = AddrMode.BaseReg;
4963  if (V->getType()->isPointerTy())
4964  V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
4965  if (V->getType() != IntPtrTy)
4966  V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
4967  Result = V;
4968  }
4969 
4970  // Add the scale value.
4971  if (AddrMode.Scale) {
4972  Value *V = AddrMode.ScaledReg;
4973  if (V->getType() == IntPtrTy) {
4974  // done.
4975  } else if (V->getType()->isPointerTy()) {
4976  V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
4977  } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
4978  cast<IntegerType>(V->getType())->getBitWidth()) {
4979  V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
4980  } else {
4981  // It is only safe to sign extend the BaseReg if we know that the math
4982  // required to create it did not overflow before we extend it. Since
4983  // the original IR value was tossed in favor of a constant back when
4984  // the AddrMode was created we need to bail out gracefully if widths
4985  // do not match instead of extending it.
4986  Instruction *I = dyn_cast_or_null<Instruction>(Result);
4987  if (I && (Result != AddrMode.BaseReg))
4988  I->eraseFromParent();
4989  return false;
4990  }
4991  if (AddrMode.Scale != 1)
4992  V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
4993  "sunkaddr");
4994  if (Result)
4995  Result = Builder.CreateAdd(Result, V, "sunkaddr");
4996  else
4997  Result = V;
4998  }
4999 
5000  // Add in the BaseGV if present.
5001  if (AddrMode.BaseGV) {
5002  Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
5003  if (Result)
5004  Result = Builder.CreateAdd(Result, V, "sunkaddr");
5005  else
5006  Result = V;
5007  }
5008 
5009  // Add in the Base Offset if present.
5010  if (AddrMode.BaseOffs) {
5011  Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
5012  if (Result)
5013  Result = Builder.CreateAdd(Result, V, "sunkaddr");
5014  else
5015  Result = V;
5016  }
5017 
5018  if (!Result)
5019  SunkAddr = Constant::getNullValue(Addr->getType());
5020  else
5021  SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
5022  }
5023 
5024  MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
5025  // Store the newly computed address into the cache. In the case we reused a
5026  // value, this should be idempotent.
5027  SunkAddrs[Addr] = WeakTrackingVH(SunkAddr);
5028 
5029  // If we have no uses, recursively delete the value and all dead instructions
5030  // using it.
5031  if (Repl->use_empty()) {
5032  // This can cause recursive deletion, which can invalidate our iterator.
5033  // Use a WeakTrackingVH to hold onto it in case this happens.
5034  Value *CurValue = &*CurInstIterator;
5035  WeakTrackingVH IterHandle(CurValue);
5036  BasicBlock *BB = CurInstIterator->getParent();
5037 
5039 
5040  if (IterHandle != CurValue) {
5041  // If the iterator instruction was recursively deleted, start over at the
5042  // start of the block.
5043  CurInstIterator = BB->begin();
5044  SunkAddrs.clear();
5045  }
5046  }
5047  ++NumMemoryInsts;
5048  return true;
5049 }
5050 
5051 /// If there are any memory operands, use OptimizeMemoryInst to sink their
5052 /// address computing into the block when possible / profitable.
5053 bool CodeGenPrepare::optimizeInlineAsmInst(CallInst *CS) {
5054  bool MadeChange = false;
5055 
5056  const TargetRegisterInfo *TRI =
5057  TM->getSubtargetImpl(*CS->getFunction())->getRegisterInfo();
5058  TargetLowering::AsmOperandInfoVector TargetConstraints =
5059  TLI->ParseConstraints(*DL, TRI, CS);
5060  unsigned ArgNo = 0;
5061  for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
5062  TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
5063 
5064  // Compute the constraint code and ConstraintType to use.
5065  TLI->ComputeConstraintToUse(OpInfo, SDValue());
5066 
5067  if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
5068  OpInfo.isIndirect) {
5069  Value *OpVal = CS->getArgOperand(ArgNo++);
5070  MadeChange |= optimizeMemoryInst(CS, OpVal, OpVal->getType(), ~0u);
5071  } else if (OpInfo.Type == InlineAsm::isInput)
5072  ArgNo++;
5073  }
5074 
5075  return MadeChange;
5076 }
5077 
5078 /// Check if all the uses of \p Val are equivalent (or free) zero or
5079 /// sign extensions.
5080 static bool hasSameExtUse(Value *Val, const TargetLowering &TLI) {
5081  assert(!Val->use_empty() && "Input must have at least one use");
5082  const Instruction *FirstUser = cast<Instruction>(*Val->user_begin());
5083  bool IsSExt = isa<SExtInst>(FirstUser);
5084  Type *ExtTy = FirstUser->getType();
5085  for (const User *U : Val->users()) {
5086  const Instruction *UI = cast<Instruction>(U);
5087  if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI)))
5088  return false;
5089  Type *CurTy = UI->getType();
5090  // Same input and output types: Same instruction after CSE.
5091  if (CurTy == ExtTy)
5092  continue;
5093 
5094  // If IsSExt is true, we are in this situation:
5095  // a = Val
5096  // b = sext ty1 a to ty2
5097  // c = sext ty1 a to ty3
5098  // Assuming ty2 is shorter than ty3, this could be turned into:
5099  // a = Val
5100  // b = sext ty1 a to ty2
5101  // c = sext ty2 b to ty3
5102  // However, the last sext is not free.
5103  if (IsSExt)
5104  return false;
5105 
5106  // This is a ZExt, maybe this is free to extend from one type to another.
5107  // In that case, we would not account for a different use.
5108  Type *NarrowTy;
5109  Type *LargeTy;
5110  if (ExtTy->getScalarType()->getIntegerBitWidth() >
5111  CurTy->getScalarType()->getIntegerBitWidth()) {
5112  NarrowTy = CurTy;
5113  LargeTy = ExtTy;
5114  } else {
5115  NarrowTy = ExtTy;
5116  LargeTy = CurTy;
5117  }
5118 
5119  if (!TLI.isZExtFree(NarrowTy, LargeTy))
5120  return false;
5121  }
5122  // All uses are the same or can be derived from one another for free.
5123  return true;
5124 }
5125 
5126 /// Try to speculatively promote extensions in \p Exts and continue
5127 /// promoting through newly promoted operands recursively as far as doing so is
5128 /// profitable. Save extensions profitably moved up, in \p ProfitablyMovedExts.
5129 /// When some promotion happened, \p TPT contains the proper state to revert
5130 /// them.
5131 ///
5132 /// \return true if some promotion happened, false otherwise.
5133 bool CodeGenPrepare::tryToPromoteExts(
5134  TypePromotionTransaction &TPT, const SmallVectorImpl<Instruction *> &Exts,
5135  SmallVectorImpl<Instruction *> &ProfitablyMovedExts,
5136  unsigned CreatedInstsCost) {
5137  bool Promoted = false;
5138 
5139  // Iterate over all the extensions to try to promote them.
5140  for (auto I : Exts) {
5141  // Early check if we directly have ext(load).
5142  if (isa<LoadInst>(I->getOperand(0))) {
5143  ProfitablyMovedExts.push_back(I);
5144  continue;
5145  }
5146 
5147  // Check whether or not we want to do any promotion. The reason we have
5148  // this check inside the for loop is to catch the case where an extension
5149  // is directly fed by a load because in such case the extension can be moved
5150  // up without any promotion on its operands.
5151  if (!TLI || !TLI->enableExtLdPromotion() || DisableExtLdPromotion)
5152  return false;
5153 
5154  // Get the action to perform the promotion.
5155  TypePromotionHelper::Action TPH =
5156  TypePromotionHelper::getAction(I, InsertedInsts, *TLI, PromotedInsts);
5157  // Check if we can promote.
5158  if (!TPH) {
5159  // Save the current extension as we cannot move up through its operand.
5160  ProfitablyMovedExts.push_back(I);
5161  continue;
5162  }
5163 
5164  // Save the current state.
5165  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
5166  TPT.getRestorationPoint();
5168  unsigned NewCreatedInstsCost = 0;
5169  unsigned ExtCost = !TLI->isExtFree(I);
5170  // Promote.
5171  Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost,
5172  &NewExts, nullptr, *TLI);
5173  assert(PromotedVal &&
5174  "TypePromotionHelper should have filtered out those cases");
5175 
5176  // We would be able to merge only one extension in a load.
5177  // Therefore, if we have more than 1 new extension we heuristically
5178  // cut this search path, because it means we degrade the code quality.
5179  // With exactly 2, the transformation is neutral, because we will merge
5180  // one extension but leave one. However, we optimistically keep going,
5181  // because the new extension may be removed too.
5182  long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost;
5183  // FIXME: It would be possible to propagate a negative value instead of
5184  // conservatively ceiling it to 0.
5185  TotalCreatedInstsCost =
5186  std::max((long long)0, (TotalCreatedInstsCost - ExtCost));
5187  if (!StressExtLdPromotion &&
5188  (TotalCreatedInstsCost > 1 ||
5189  !isPromotedInstructionLegal(*TLI, *DL, PromotedVal))) {
5190  // This promotion is not profitable, rollback to the previous state, and
5191  // save the current extension in ProfitablyMovedExts as the latest
5192  // speculative promotion turned out to be unprofitable.
5193  TPT.rollback(LastKnownGood);
5194  ProfitablyMovedExts.push_back(I);
5195  continue;
5196  }
5197  // Continue promoting NewExts as far as doing so is profitable.
5198  SmallVector<Instruction *, 2> NewlyMovedExts;
5199  (void)tryToPromoteExts(TPT, NewExts, NewlyMovedExts, TotalCreatedInstsCost);
5200  bool NewPromoted = false;
5201  for (auto ExtInst : NewlyMovedExts) {
5202  Instruction *MovedExt = cast<Instruction>(ExtInst);
5203  Value *ExtOperand = MovedExt->getOperand(0);
5204  // If we have reached to a load, we need this extra profitability check
5205  // as it could potentially be merged into an ext(load).
5206  if (isa<LoadInst>(ExtOperand) &&
5207  !(StressExtLdPromotion || NewCreatedInstsCost <= ExtCost ||
5208  (ExtOperand->hasOneUse() || hasSameExtUse(ExtOperand, *TLI))))
5209  continue;
5210 
5211  ProfitablyMovedExts.push_back(MovedExt);
5212  NewPromoted = true;
5213  }
5214 
5215  // If none of speculative promotions for NewExts is profitable, rollback
5216  // and save the current extension (I) as the last profitable extension.
5217  if (!NewPromoted) {
5218  TPT.rollback(LastKnownGood);
5219  ProfitablyMovedExts.push_back(I);
5220  continue;
5221  }
5222  // The promotion is profitable.
5223  Promoted = true;
5224  }
5225  return Promoted;
5226 }
5227 
5228 /// Merging redundant sexts when one is dominating the other.
5229 bool CodeGenPrepare::mergeSExts(Function &F) {
5230  bool Changed = false;
5231  for (auto &Entry : ValToSExtendedUses) {
5232  SExts &Insts = Entry.second;
5233  SExts CurPts;
5234  for (Instruction *Inst : Insts) {
5235  if (RemovedInsts.count(Inst) || !isa<SExtInst>(Inst) ||
5236  Inst->getOperand(0) != Entry.first)
5237  continue;
5238  bool inserted = false;
5239  for (auto &Pt : CurPts) {
5240  if (getDT(F).dominates(Inst, Pt)) {
5241  Pt->replaceAllUsesWith(Inst);
5242  RemovedInsts.insert(Pt);
5243  Pt->removeFromParent();
5244  Pt = Inst;
5245  inserted = true;
5246  Changed = true;
5247  break;
5248  }
5249  if (!getDT(F).dominates(Pt, Inst))
5250  // Give up if we need to merge in a common dominator as the
5251  // experiments show it is not profitable.
5252  continue;
5253  Inst->replaceAllUsesWith(Pt);
5254  RemovedInsts.insert(Inst);
5255  Inst->removeFromParent();
5256  inserted = true;
5257  Changed = true;
5258  break;
5259  }
5260  if (!inserted)
5261  CurPts.push_back(Inst);
5262  }
5263  }
5264  return Changed;
5265 }
5266 
5267 // Spliting large data structures so that the GEPs accessing them can have
5268 // smaller offsets so that they can be sunk to the same blocks as their users.
5269 // For example, a large struct starting from %base is splitted into two parts
5270 // where the second part starts from %new_base.
5271 //
5272 // Before:
5273 // BB0:
5274 // %base =
5275 //
5276 // BB1:
5277 // %gep0 = gep %base, off0
5278 // %gep1 = gep %base, off1
5279 // %gep2 = gep %base, off2
5280 //
5281 // BB2:
5282 // %load1 = load %gep0
5283 // %load2 = load %gep1
5284 // %load3 = load %gep2
5285 //
5286 // After:
5287 // BB0:
5288 // %base =
5289 // %new_base = gep %base, off0
5290 //
5291 // BB1:
5292 // %new_gep0 = %new_base
5293 // %new_gep1 = gep %new_base, off1 - off0
5294 // %new_gep2 = gep %new_base, off2 - off0
5295 //
5296 // BB2:
5297 // %load1 = load i32, i32* %new_gep0
5298 // %load2 = load i32, i32* %new_gep1
5299 // %load3 = load i32, i32* %new_gep2
5300 //
5301 // %new_gep1 and %new_gep2 can be sunk to BB2 now after the splitting because
5302 // their offsets are smaller enough to fit into the addressing mode.
5303 bool CodeGenPrepare::splitLargeGEPOffsets() {
5304  bool Changed = false;
5305  for (auto &Entry : LargeOffsetGEPMap) {
5306  Value *OldBase = Entry.first;
5308  &LargeOffsetGEPs = Entry.second;
5309  auto compareGEPOffset =
5310  [&](const std::pair<GetElementPtrInst *, int64_t> &LHS,
5311  const std::pair<GetElementPtrInst *, int64_t> &RHS) {
5312  if (LHS.first == RHS.first)
5313  return false;
5314  if (LHS.second != RHS.second)
5315  return LHS.second < RHS.second;
5316  return LargeOffsetGEPID[LHS.first] < LargeOffsetGEPID[RHS.first];
5317  };
5318  // Sorting all the GEPs of the same data structures based on the offsets.
5319  llvm::sort(LargeOffsetGEPs, compareGEPOffset);
5320  LargeOffsetGEPs.erase(
5321  std::unique(LargeOffsetGEPs.begin(), LargeOffsetGEPs.end()),
5322  LargeOffsetGEPs.end());
5323  // Skip if all the GEPs have the same offsets.
5324  if (LargeOffsetGEPs.front().second == LargeOffsetGEPs.back().second)
5325  continue;
5326  GetElementPtrInst *BaseGEP = LargeOffsetGEPs.begin()->first;
5327  int64_t BaseOffset = LargeOffsetGEPs.begin()->second;
5328  Value *NewBaseGEP = nullptr;
5329 
5330  auto LargeOffsetGEP = LargeOffsetGEPs.begin();
5331  while (LargeOffsetGEP != LargeOffsetGEPs.end()) {
5332  GetElementPtrInst *GEP = LargeOffsetGEP->first;
5333  int64_t Offset = LargeOffsetGEP->second;
5334  if (Offset != BaseOffset) {
5336  AddrMode.BaseOffs = Offset - BaseOffset;
5337  // The result type of the GEP might not be the type of the memory
5338  // access.
5339  if (!TLI->isLegalAddressingMode(*DL, AddrMode,
5340  GEP->getResultElementType(),
5341  GEP->getAddressSpace())) {
5342  // We need to create a new base if the offset to the current base is
5343  // too large to fit into the addressing mode. So, a very large struct
5344  // may be splitted into several parts.
5345  BaseGEP = GEP;
5346  BaseOffset = Offset;
5347  NewBaseGEP = nullptr;
5348  }
5349  }
5350 
5351  // Generate a new GEP to replace the current one.
5352  LLVMContext &Ctx = GEP->getContext();
5353  Type *IntPtrTy = DL->getIntPtrType(GEP->getType());
5354  Type *I8PtrTy =
5356  Type *I8Ty = Type::getInt8Ty(Ctx);
5357 
5358  if (!NewBaseGEP) {
5359  // Create a new base if we don't have one yet. Find the insertion
5360  // pointer for the new base first.
5361  BasicBlock::iterator NewBaseInsertPt;
5362  BasicBlock *NewBaseInsertBB;
5363  if (auto *BaseI = dyn_cast<Instruction>(OldBase)) {
5364  // If the base of the struct is an instruction, the new base will be
5365  // inserted close to it.
5366  NewBaseInsertBB = BaseI->getParent();
5367  if (isa<PHINode>(BaseI))
5368  NewBaseInsertPt = NewBaseInsertBB->getFirstInsertionPt();
5369  else if (InvokeInst *Invoke = dyn_cast<InvokeInst>(BaseI)) {
5370  NewBaseInsertBB =
5371  SplitEdge(NewBaseInsertBB, Invoke->getNormalDest());
5372  NewBaseInsertPt = NewBaseInsertBB->getFirstInsertionPt();
5373  } else
5374  NewBaseInsertPt = std::next(BaseI->getIterator());
5375  } else {
5376  // If the current base is an argument or global value, the new base
5377  // will be inserted to the entry block.
5378  NewBaseInsertBB = &BaseGEP->getFunction()->getEntryBlock();
5379  NewBaseInsertPt = NewBaseInsertBB->getFirstInsertionPt();
5380  }
5381  IRBuilder<> NewBaseBuilder(NewBaseInsertBB, NewBaseInsertPt);
5382  // Create a new base.
5383  Value *BaseIndex = ConstantInt::get(IntPtrTy, BaseOffset);
5384  NewBaseGEP = OldBase;
5385  if (NewBaseGEP->getType() != I8PtrTy)
5386  NewBaseGEP = NewBaseBuilder.CreatePointerCast(NewBaseGEP, I8PtrTy);
5387  NewBaseGEP =
5388  NewBaseBuilder.CreateGEP(I8Ty, NewBaseGEP, BaseIndex, "splitgep");
5389  NewGEPBases.insert(NewBaseGEP);
5390  }
5391 
5392  IRBuilder<> Builder(GEP);
5393  Value *NewGEP = NewBaseGEP;
5394  if (Offset == BaseOffset) {
5395  if (GEP->getType() != I8PtrTy)
5396  NewGEP = Builder.CreatePointerCast(NewGEP, GEP->getType());
5397  } else {
5398  // Calculate the new offset for the new GEP.
5399  Value *Index = ConstantInt::get(IntPtrTy, Offset - BaseOffset);
5400  NewGEP = Builder.CreateGEP(I8Ty, NewBaseGEP, Index);
5401 
5402  if (GEP->getType() != I8PtrTy)
5403  NewGEP = Builder.CreatePointerCast(NewGEP, GEP->getType());
5404  }
5405  GEP->replaceAllUsesWith(NewGEP);
5406  LargeOffsetGEPID.erase(GEP);
5407  LargeOffsetGEP = LargeOffsetGEPs.erase(LargeOffsetGEP);
5408  GEP->eraseFromParent();
5409  Changed = true;
5410  }
5411  }
5412  return Changed;
5413 }
5414 
5415 /// Return true, if an ext(load) can be formed from an extension in
5416 /// \p MovedExts.
5417 bool CodeGenPrepare::canFormExtLd(
5418  const SmallVectorImpl<Instruction *> &MovedExts, LoadInst *&LI,
5419  Instruction *&Inst, bool HasPromoted) {
5420  for (auto *MovedExtInst : MovedExts) {
5421  if (isa<LoadInst>(MovedExtInst->getOperand(0))) {
5422  LI = cast<LoadInst>(MovedExtInst->getOperand(0));
5423  Inst = MovedExtInst;
5424  break;
5425  }
5426  }
5427  if (!LI)
5428  return false;
5429 
5430  // If they're already in the same block, there's nothing to do.
5431  // Make the cheap checks first if we did not promote.
5432  // If we promoted, we need to check if it is indeed profitable.
5433  if (!HasPromoted && LI->getParent() == Inst->getParent())
5434  return false;
5435 
5436  return TLI->isExtLoad(LI, Inst, *DL);
5437 }
5438 
5439 /// Move a zext or sext fed by a load into the same basic block as the load,
5440 /// unless conditions are unfavorable. This allows SelectionDAG to fold the
5441 /// extend into the load.
5442 ///
5443 /// E.g.,
5444 /// \code
5445 /// %ld = load i32* %addr
5446 /// %add = add nuw i32 %ld, 4
5447 /// %zext = zext i32 %add to i64
5448 // \endcode
5449 /// =>
5450 /// \code
5451 /// %ld = load i32* %addr
5452 /// %zext = zext i32 %ld to i64
5453 /// %add = add nuw i64 %zext, 4
5454 /// \encode
5455 /// Note that the promotion in %add to i64 is done in tryToPromoteExts(), which
5456 /// allow us to match zext(load i32*) to i64.
5457 ///
5458 /// Also, try to promote the computations used to obtain a sign extended
5459 /// value used into memory accesses.
5460 /// E.g.,
5461 /// \code
5462 /// a = add nsw i32 b, 3
5463 /// d = sext i32 a to i64
5464 /// e = getelementptr ..., i64 d
5465 /// \endcode
5466 /// =>
5467 /// \code
5468 /// f = sext i32 b to i64
5469 /// a = add nsw i64 f, 3
5470 /// e = getelementptr ..., i64 a
5471 /// \endcode
5472 ///
5473 /// \p Inst[in/out] the extension may be modified during the process if some
5474 /// promotions apply.
5475 bool CodeGenPrepare::optimizeExt(Instruction *&Inst) {
5476  // ExtLoad formation and address type promotion infrastructure requires TLI to
5477  // be effective.
5478  if (!TLI)
5479  return false;
5480 
5481  bool AllowPromotionWithoutCommonHeader = false;
5482  /// See if it is an interesting sext operations for the address type
5483  /// promotion before trying to promote it, e.g., the ones with the right
5484  /// type and used in memory accesses.
5485  bool ATPConsiderable = TTI->shouldConsiderAddressTypePromotion(
5486  *Inst, AllowPromotionWithoutCommonHeader);
5487  TypePromotionTransaction TPT(RemovedInsts);
5488  TypePromotionTransaction::ConstRestorationPt LastKnownGood =
5489  TPT.getRestorationPoint();
5491  SmallVector<Instruction *, 2> SpeculativelyMovedExts;
5492  Exts.push_back(Inst);
5493 
5494  bool HasPromoted = tryToPromoteExts(TPT, Exts, SpeculativelyMovedExts);
5495 
5496  // Look for a load being extended.
5497  LoadInst *LI = nullptr;
5498  Instruction *ExtFedByLoad;
5499 
5500  // Try to promote a chain of computation if it allows to form an extended
5501  // load.
5502  if (canFormExtLd(SpeculativelyMovedExts, LI, ExtFedByLoad, HasPromoted)) {
5503  assert(LI && ExtFedByLoad && "Expect a valid load and extension");
5504  TPT.commit();
5505  // Move the extend into the same block as the load
5506  ExtFedByLoad->moveAfter(LI);
5507  // CGP does not check if the zext would be speculatively executed when moved
5508  // to the same basic block as the load. Preserving its original location
5509  // would pessimize the debugging experience, as well as negatively impact
5510  // the quality of sample pgo. We don't want to use "line 0" as that has a
5511  // size cost in the line-table section and logically the zext can be seen as
5512  // part of the load. Therefore we conservatively reuse the same debug
5513  // location for the load and the zext.
5514  ExtFedByLoad->setDebugLoc(LI->getDebugLoc());
5515  ++NumExtsMoved;
5516  Inst = ExtFedByLoad;
5517  return true;
5518  }
5519 
5520  // Continue promoting SExts if known as considerable depending on targets.
5521  if (ATPConsiderable &&
5522  performAddressTypePromotion(Inst, AllowPromotionWithoutCommonHeader,
5523  HasPromoted, TPT, SpeculativelyMovedExts))
5524  return true;
5525 
5526  TPT.rollback(LastKnownGood);
5527  return false;
5528 }
5529 
5530 // Perform address type promotion if doing so is profitable.
5531 // If AllowPromotionWithoutCommonHeader == false, we should find other sext
5532 // instructions that sign extended the same initial value. However, if
5533 // AllowPromotionWithoutCommonHeader == true, we expect promoting the
5534 // extension is just profitable.
5535 bool CodeGenPrepare::performAddressTypePromotion(
5536  Instruction *&Inst, bool AllowPromotionWithoutCommonHeader,
5537  bool HasPromoted, TypePromotionTransaction &TPT,
5538  SmallVectorImpl<Instruction *> &SpeculativelyMovedExts) {
5539  bool Promoted = false;
5540  SmallPtrSet<Instruction *, 1> UnhandledExts;
5541  bool AllSeenFirst = true;
5542  for (auto I : SpeculativelyMovedExts) {
5543  Value *HeadOfChain = I->getOperand(0);
5545  SeenChainsForSExt.find(HeadOfChain);
5546  // If there is an unhandled SExt which has the same header, try to promote
5547  // it as well.
5548  if (AlreadySeen != SeenChainsForSExt.end()) {
5549  if (AlreadySeen->second != nullptr)
5550  UnhandledExts.insert(AlreadySeen->second);
5551  AllSeenFirst = false;
5552  }
5553  }
5554 
5555  if (!AllSeenFirst || (AllowPromotionWithoutCommonHeader &&
5556  SpeculativelyMovedExts.size() == 1)) {
5557  TPT.commit();
5558  if (HasPromoted)
5559  Promoted = true;
5560  for (auto I : SpeculativelyMovedExts) {
5561  Value *HeadOfChain = I->getOperand(0);
5562  SeenChainsForSExt[HeadOfChain] = nullptr;
5563  ValToSExtendedUses[HeadOfChain].push_back(I);
5564  }
5565  // Update Inst as promotion happen.
5566  Inst = SpeculativelyMovedExts.pop_back_val();
5567  } else {
5568  // This is the first chain visited from the header, keep the current chain
5569  // as unhandled. Defer to promote this until we encounter another SExt
5570  // chain derived from the same header.
5571  for (auto I : SpeculativelyMovedExts) {
5572  Value *HeadOfChain = I->getOperand(0);
5573  SeenChainsForSExt[HeadOfChain] = Inst;
5574  }
5575  return false;
5576  }
5577 
5578  if (!AllSeenFirst && !UnhandledExts.empty())
5579  for (auto VisitedSExt : UnhandledExts) {
5580  if (RemovedInsts.count(VisitedSExt))
5581  continue;
5582  TypePromotionTransaction TPT(RemovedInsts);
5585  Exts.push_back(VisitedSExt);
5586  bool HasPromoted = tryToPromoteExts(TPT, Exts, Chains);
5587  TPT.commit();
5588  if (HasPromoted)
5589  Promoted = true;
5590  for (auto I : Chains) {
5591  Value *HeadOfChain = I->getOperand(0);
5592  // Mark this as handled.
5593  SeenChainsForSExt[HeadOfChain] = nullptr;
5594  ValToSExtendedUses[HeadOfChain].push_back(I);
5595  }
5596  }
5597  return Promoted;
5598 }
5599 
5600 bool CodeGenPrepare::optimizeExtUses(Instruction *I) {
5601  BasicBlock *DefBB = I->getParent();
5602 
5603  // If the result of a {s|z}ext and its source are both live out, rewrite all
5604  // other uses of the source with result of extension.
5605  Value *Src = I->getOperand(0);
5606  if (Src->hasOneUse())
5607  return false;
5608 
5609  // Only do this xform if truncating is free.
5610  if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
5611  return false;
5612 
5613  // Only safe to perform the optimization if the source is also defined in
5614  // this block.
5615  if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
5616  return false;
5617 
5618  bool DefIsLiveOut = false;
5619  for (User *U : I->users()) {
5620  Instruction *UI = cast<Instruction>(U);
5621 
5622  // Figure out which BB this ext is used in.
5623  BasicBlock *UserBB = UI->getParent();
5624  if (UserBB == DefBB) continue;
5625  DefIsLiveOut = true;
5626  break;
5627  }
5628  if (!DefIsLiveOut)
5629  return false;
5630 
5631  // Make sure none of the uses are PHI nodes.
5632  for (User *U : Src->users()) {
5633  Instruction *UI = cast<Instruction>(U);
5634  BasicBlock *UserBB = UI->getParent();
5635  if (UserBB == DefBB) continue;
5636  // Be conservative. We don't want this xform to end up introducing
5637  // reloads just before load / store instructions.
5638  if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
5639  return false;
5640  }
5641 
5642  // InsertedTruncs - Only insert one trunc in each block once.
5643  DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
5644 
5645  bool MadeChange = false;
5646  for (Use &U : Src->uses()) {
5647  Instruction *User = cast<Instruction>(U.getUser());
5648 
5649  // Figure out which BB this ext is used in.
5650  BasicBlock *UserBB = User->getParent();
5651  if (UserBB == DefBB) continue;
5652 
5653  // Both src and def are live in this block. Rewrite the use.
5654  Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
5655 
5656  if (!InsertedTrunc) {
5657  BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
5658  assert(InsertPt != UserBB->end());
5659  InsertedTrunc = new TruncInst(I, Src->getType(), "", &*InsertPt);
5660  InsertedInsts.insert(InsertedTrunc);
5661  }
5662 
5663  // Replace a use of the {s|z}ext source with a use of the result.
5664  U = InsertedTrunc;
5665  ++NumExtUses;
5666  MadeChange = true;
5667  }
5668 
5669  return MadeChange;
5670 }
5671 
5672 // Find loads whose uses only use some of the loaded value's bits. Add an "and"
5673 // just after the load if the target can fold this into one extload instruction,
5674 // with the hope of eliminating some of the other later "and" instructions using
5675 // the loaded value. "and"s that are made trivially redundant by the insertion
5676 // of the new "and" are removed by this function, while others (e.g. those whose
5677 // path from the load goes through a phi) are left for isel to potentially
5678 // remove.
5679 //
5680 // For example:
5681 //
5682 // b0:
5683 // x = load i32
5684 // ...
5685 // b1:
5686 // y = and x, 0xff
5687 // z = use y
5688 //
5689 // becomes:
5690 //
5691 // b0:
5692 // x = load i32
5693 // x' = and x, 0xff
5694 // ...
5695 // b1:
5696 // z = use x'
5697 //
5698 // whereas:
5699 //
5700 // b0:
5701 // x1 = load i32
5702 // ...
5703 // b1:
5704 // x2 = load i32
5705 // ...
5706 // b2:
5707 // x = phi x1, x2
5708 // y = and x, 0xff
5709 //
5710 // becomes (after a call to optimizeLoadExt for each load):
5711 //
5712 // b0:
5713 // x1 = load i32
5714 // x1' = and x1, 0xff
5715 // ...
5716 // b1:
5717 // x2 = load i32
5718 // x2' = and x2, 0xff
5719 // ...
5720 // b2:
5721 // x = phi x1', x2'
5722 // y = and x, 0xff
5723 bool CodeGenPrepare::optimizeLoadExt(LoadInst *Load) {
5724  if (!Load->isSimple() || !Load->getType()->isIntOrPtrTy())
5725  return false;
5726 
5727  // Skip loads we've already transformed.
5728  if (Load->hasOneUse() &&
5729  InsertedInsts.count(cast<Instruction>(*Load->user_begin())))
5730  return false;
5731 
5732  // Look at all uses of Load, looking through phis, to determine how many bits
5733  // of the loaded value are needed.
5736  SmallVector<Instruction *, 8> AndsToMaybeRemove;
5737  for (auto *U : Load->users())
5738  WorkList.push_back(cast<Instruction>(U));
5739 
5740  EVT LoadResultVT = TLI->getValueType(*DL, Load->getType());
5741  unsigned BitWidth = LoadResultVT.getSizeInBits();
5742  APInt DemandBits(BitWidth, 0);
5743  APInt WidestAndBits(BitWidth, 0);
5744 
5745  while (!WorkList.empty()) {
5746  Instruction *I = WorkList.back();
5747  WorkList.pop_back();
5748 
5749  // Break use-def graph loops.
5750  if (!Visited.insert(I).second)
5751  continue;
5752 
5753  // For a PHI node, push all of its users.
5754  if (auto *Phi = dyn_cast<PHINode>(I)) {
5755  for (auto *U : Phi->users())
5756  WorkList.push_back(cast<Instruction>(U));
5757  continue;
5758  }
5759 
5760  switch (I->getOpcode()) {
5761  case Instruction::And: {
5762  auto *AndC = dyn_cast<ConstantInt>(I->getOperand(1));
5763  if (!AndC)
5764  return false;
5765  APInt AndBits = AndC->getValue();
5766  DemandBits |= AndBits;
5767  // Keep track of the widest and mask we see.
5768  if (AndBits.ugt(WidestAndBits))
5769  WidestAndBits = AndBits;
5770  if (AndBits == WidestAndBits && I->getOperand(0) == Load)
5771  AndsToMaybeRemove.push_back(I);
5772  break;
5773  }
5774 
5775  case Instruction::Shl: {
5776  auto *ShlC = dyn_cast<ConstantInt>(I->getOperand(1));
5777  if (!ShlC)
5778  return false;
5779  uint64_t ShiftAmt = ShlC->getLimitedValue(BitWidth - 1);
5780  DemandBits.setLowBits(BitWidth - ShiftAmt);
5781  break;
5782  }
5783 
5784  case Instruction::Trunc: {
5785  EVT TruncVT = TLI->getValueType(*DL, I->getType());
5786  unsigned TruncBitWidth = TruncVT.getSizeInBits();
5787  DemandBits.setLowBits(TruncBitWidth);
5788  break;
5789  }
5790 
5791  default:
5792  return false;
5793  }
5794  }
5795 
5796  uint32_t ActiveBits = DemandBits.getActiveBits();
5797  // Avoid hoisting (and (load x) 1) since it is unlikely to be folded by the
5798  // target even if isLoadExtLegal says an i1 EXTLOAD is valid. For example,
5799  // for the AArch64 target isLoadExtLegal(ZEXTLOAD, i32, i1) returns true, but
5800  // (and (load x) 1) is not matched as a single instruction, rather as a LDR
5801  // followed by an AND.
5802  // TODO: Look into removing this restriction by fixing backends to either
5803  // return false for isLoadExtLegal for i1 or have them select this pattern to
5804  // a single instruction.
5805  //
5806  // Also avoid hoisting if we didn't see any ands with the exact DemandBits
5807  // mask, since these are the only ands that will be removed by isel.
5808  if (ActiveBits <= 1 || !DemandBits.isMask(ActiveBits) ||
5809  WidestAndBits != DemandBits)
5810  return false;
5811 
5812  LLVMContext &Ctx = Load->getType()->getContext();
5813  Type *TruncTy = Type::getIntNTy(Ctx, ActiveBits);
5814  EVT TruncVT = TLI->getValueType(*DL, TruncTy);
5815 
5816  // Reject cases that won't be matched as extloads.
5817  if (!LoadResultVT.bitsGT(TruncVT) || !TruncVT.isRound() ||
5818  !TLI->isLoadExtLegal(ISD::ZEXTLOAD, LoadResultVT, TruncVT))
5819  return false;
5820 
5821  IRBuilder<> Builder(Load->