1 //===- CodeGenPrepare.cpp - Prepare a function for code generation --------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This pass munges the code in the input function to better prepare it for
11 // SelectionDAG-based code generation. This works around limitations in it's
12 // basic-block-at-a-time approach. It should eventually be removed.
13 //
14 //===----------------------------------------------------------------------===//
15 
16 #include "llvm/CodeGen/Passes.h"
17 #include "llvm/ADT/DenseMap.h"
18 #include "llvm/ADT/SmallSet.h"
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/Analysis/InstructionSimplify.h"
21 #include "llvm/Analysis/TargetLibraryInfo.h"
22 #include "llvm/Analysis/TargetTransformInfo.h"
23 #include "llvm/IR/CallSite.h"
24 #include "llvm/IR/Constants.h"
25 #include "llvm/IR/DataLayout.h"
26 #include "llvm/IR/DerivedTypes.h"
27 #include "llvm/IR/Dominators.h"
28 #include "llvm/IR/Function.h"
29 #include "llvm/IR/GetElementPtrTypeIterator.h"
30 #include "llvm/IR/IRBuilder.h"
31 #include "llvm/IR/InlineAsm.h"
32 #include "llvm/IR/Instructions.h"
33 #include "llvm/IR/IntrinsicInst.h"
34 #include "llvm/IR/MDBuilder.h"
35 #include "llvm/IR/PatternMatch.h"
36 #include "llvm/IR/Statepoint.h"
37 #include "llvm/IR/ValueHandle.h"
38 #include "llvm/IR/ValueMap.h"
39 #include "llvm/Pass.h"
40 #include "llvm/Support/CommandLine.h"
41 #include "llvm/Support/Debug.h"
42 #include "llvm/Support/raw_ostream.h"
43 #include "llvm/Target/TargetLowering.h"
44 #include "llvm/Target/TargetSubtargetInfo.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/BuildLibCalls.h"
47 #include "llvm/Transforms/Utils/BypassSlowDivision.h"
48 #include "llvm/Transforms/Utils/Local.h"
49 #include "llvm/Transforms/Utils/SimplifyLibCalls.h"
50 using namespace llvm;
51 using namespace llvm::PatternMatch;
52 
53 #define DEBUG_TYPE "codegenprepare"
54 
55 STATISTIC(NumBlocksElim, "Number of blocks eliminated");
56 STATISTIC(NumPHIsElim,   "Number of trivial PHIs eliminated");
57 STATISTIC(NumGEPsElim,   "Number of GEPs converted to casts");
58 STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "
59                       "sunken Cmps");
60 STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "
61                        "of sunken Casts");
62 STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "
63                           "computations were sunk");
64 STATISTIC(NumExtsMoved,  "Number of [s|z]ext instructions combined with loads");
65 STATISTIC(NumExtUses,    "Number of uses of [s|z]ext instructions optimized");
66 STATISTIC(NumRetsDup,    "Number of return instructions duplicated");
67 STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved");
68 STATISTIC(NumSelectsExpanded, "Number of selects turned into branches");
69 STATISTIC(NumAndCmpsMoved, "Number of and/cmp's pushed into branches");
70 STATISTIC(NumStoreExtractExposed, "Number of store(extractelement) exposed");
71 
72 static cl::opt<bool> DisableBranchOpts(
73   "disable-cgp-branch-opts", cl::Hidden, cl::init(false),
74   cl::desc("Disable branch optimizations in CodeGenPrepare"));
75 
76 static cl::opt<bool>
77     DisableGCOpts("disable-cgp-gc-opts", cl::Hidden, cl::init(false),
78                   cl::desc("Disable GC optimizations in CodeGenPrepare"));
79 
80 static cl::opt<bool> DisableSelectToBranch(
81   "disable-cgp-select2branch", cl::Hidden, cl::init(false),
82   cl::desc("Disable select to branch conversion."));
83 
84 static cl::opt<bool> AddrSinkUsingGEPs(
85   "addr-sink-using-gep", cl::Hidden, cl::init(false),
86   cl::desc("Address sinking in CGP using GEPs."));
87 
88 static cl::opt<bool> EnableAndCmpSinking(
89    "enable-andcmp-sinking", cl::Hidden, cl::init(true),
90    cl::desc("Enable sinkinig and/cmp into branches."));
91 
92 static cl::opt<bool> DisableStoreExtract(
93     "disable-cgp-store-extract", cl::Hidden, cl::init(false),
94     cl::desc("Disable store(extract) optimizations in CodeGenPrepare"));
95 
96 static cl::opt<bool> StressStoreExtract(
97     "stress-cgp-store-extract", cl::Hidden, cl::init(false),
98     cl::desc("Stress test store(extract) optimizations in CodeGenPrepare"));
99 
100 static cl::opt<bool> DisableExtLdPromotion(
101     "disable-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
102     cl::desc("Disable ext(promotable(ld)) -> promoted(ext(ld)) optimization in "
103              "CodeGenPrepare"));
104 
105 static cl::opt<bool> StressExtLdPromotion(
106     "stress-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
107     cl::desc("Stress test ext(promotable(ld)) -> promoted(ext(ld)) "
108              "optimization in CodeGenPrepare"));
109 
110 namespace {
111 typedef SmallPtrSet<Instruction *, 16> SetOfInstrs;
112 struct TypeIsSExt {
113   Type *Ty;
114   bool IsSExt;
TypeIsSExt__anonebd7bfba0111::TypeIsSExt115   TypeIsSExt(Type *Ty, bool IsSExt) : Ty(Ty), IsSExt(IsSExt) {}
116 };
117 typedef DenseMap<Instruction *, TypeIsSExt> InstrToOrigTy;
118 class TypePromotionTransaction;
119 
120   class CodeGenPrepare : public FunctionPass {
121     /// TLI - Keep a pointer of a TargetLowering to consult for determining
122     /// transformation profitability.
123     const TargetMachine *TM;
124     const TargetLowering *TLI;
125     const TargetTransformInfo *TTI;
126     const TargetLibraryInfo *TLInfo;
127 
128     /// CurInstIterator - As we scan instructions optimizing them, this is the
129     /// next instruction to optimize.  Xforms that can invalidate this should
130     /// update it.
131     BasicBlock::iterator CurInstIterator;
132 
133     /// Keeps track of non-local addresses that have been sunk into a block.
134     /// This allows us to avoid inserting duplicate code for blocks with
135     /// multiple load/stores of the same address.
136     ValueMap<Value*, Value*> SunkAddrs;
137 
138     /// Keeps track of all truncates inserted for the current function.
139     SetOfInstrs InsertedTruncsSet;
140     /// Keeps track of the type of the related instruction before their
141     /// promotion for the current function.
142     InstrToOrigTy PromotedInsts;
143 
144     /// ModifiedDT - If CFG is modified in anyway.
145     bool ModifiedDT;
146 
147     /// OptSize - True if optimizing for size.
148     bool OptSize;
149 
150   public:
151     static char ID; // Pass identification, replacement for typeid
CodeGenPrepare(const TargetMachine * TM=nullptr)152     explicit CodeGenPrepare(const TargetMachine *TM = nullptr)
153         : FunctionPass(ID), TM(TM), TLI(nullptr), TTI(nullptr) {
154         initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
155       }
156     bool runOnFunction(Function &F) override;
157 
getPassName() const158     const char *getPassName() const override { return "CodeGen Prepare"; }
159 
getAnalysisUsage(AnalysisUsage & AU) const160     void getAnalysisUsage(AnalysisUsage &AU) const override {
161       AU.addPreserved<DominatorTreeWrapperPass>();
162       AU.addRequired<TargetLibraryInfoWrapperPass>();
163       AU.addRequired<TargetTransformInfoWrapperPass>();
164     }
165 
166   private:
167     bool EliminateFallThrough(Function &F);
168     bool EliminateMostlyEmptyBlocks(Function &F);
169     bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
170     void EliminateMostlyEmptyBlock(BasicBlock *BB);
171     bool OptimizeBlock(BasicBlock &BB, bool& ModifiedDT);
172     bool OptimizeInst(Instruction *I, bool& ModifiedDT);
173     bool OptimizeMemoryInst(Instruction *I, Value *Addr, Type *AccessTy);
174     bool OptimizeInlineAsmInst(CallInst *CS);
175     bool OptimizeCallInst(CallInst *CI, bool& ModifiedDT);
176     bool MoveExtToFormExtLoad(Instruction *&I);
177     bool OptimizeExtUses(Instruction *I);
178     bool OptimizeSelectInst(SelectInst *SI);
179     bool OptimizeShuffleVectorInst(ShuffleVectorInst *SI);
180     bool OptimizeExtractElementInst(Instruction *Inst);
181     bool DupRetToEnableTailCallOpts(BasicBlock *BB);
182     bool PlaceDbgValues(Function &F);
183     bool sinkAndCmp(Function &F);
184     bool ExtLdPromotion(TypePromotionTransaction &TPT, LoadInst *&LI,
185                         Instruction *&Inst,
186                         const SmallVectorImpl<Instruction *> &Exts,
187                         unsigned CreatedInstCost);
188     bool splitBranchCondition(Function &F);
189     bool simplifyOffsetableRelocate(Instruction &I);
190   };
191 }
192 
193 char CodeGenPrepare::ID = 0;
194 INITIALIZE_TM_PASS(CodeGenPrepare, "codegenprepare",
195                    "Optimize for code generation", false, false)
196 
createCodeGenPreparePass(const TargetMachine * TM)197 FunctionPass *llvm::createCodeGenPreparePass(const TargetMachine *TM) {
198   return new CodeGenPrepare(TM);
199 }
200 
runOnFunction(Function & F)201 bool CodeGenPrepare::runOnFunction(Function &F) {
202   if (skipOptnoneFunction(F))
203     return false;
204 
205   bool EverMadeChange = false;
206   // Clear per function information.
207   InsertedTruncsSet.clear();
208   PromotedInsts.clear();
209 
210   ModifiedDT = false;
211   if (TM)
212     TLI = TM->getSubtargetImpl(F)->getTargetLowering();
213   TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
214   TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
215   OptSize = F.hasFnAttribute(Attribute::OptimizeForSize);
216 
217   /// This optimization identifies DIV instructions that can be
218   /// profitably bypassed and carried out with a shorter, faster divide.
219   if (!OptSize && TLI && TLI->isSlowDivBypassed()) {
220     const DenseMap<unsigned int, unsigned int> &BypassWidths =
221        TLI->getBypassSlowDivWidths();
222     for (Function::iterator I = F.begin(); I != F.end(); I++)
223       EverMadeChange |= bypassSlowDivision(F, I, BypassWidths);
224   }
225 
226   // Eliminate blocks that contain only PHI nodes and an
227   // unconditional branch.
228   EverMadeChange |= EliminateMostlyEmptyBlocks(F);
229 
230   // llvm.dbg.value is far away from the value then iSel may not be able
231   // handle it properly. iSel will drop llvm.dbg.value if it can not
232   // find a node corresponding to the value.
233   EverMadeChange |= PlaceDbgValues(F);
234 
235   // If there is a mask, compare against zero, and branch that can be combined
236   // into a single target instruction, push the mask and compare into branch
237   // users. Do this before OptimizeBlock -> OptimizeInst ->
238   // OptimizeCmpExpression, which perturbs the pattern being searched for.
239   if (!DisableBranchOpts) {
240     EverMadeChange |= sinkAndCmp(F);
241     EverMadeChange |= splitBranchCondition(F);
242   }
243 
244   bool MadeChange = true;
245   while (MadeChange) {
246     MadeChange = false;
247     for (Function::iterator I = F.begin(); I != F.end(); ) {
248       BasicBlock *BB = I++;
249       bool ModifiedDTOnIteration = false;
250       MadeChange |= OptimizeBlock(*BB, ModifiedDTOnIteration);
251 
252       // Restart BB iteration if the dominator tree of the Function was changed
253       if (ModifiedDTOnIteration)
254         break;
255     }
256     EverMadeChange |= MadeChange;
257   }
258 
259   SunkAddrs.clear();
260 
261   if (!DisableBranchOpts) {
262     MadeChange = false;
263     SmallPtrSet<BasicBlock*, 8> WorkList;
264     for (BasicBlock &BB : F) {
265       SmallVector<BasicBlock *, 2> Successors(succ_begin(&BB), succ_end(&BB));
266       MadeChange |= ConstantFoldTerminator(&BB, true);
267       if (!MadeChange) continue;
268 
269       for (SmallVectorImpl<BasicBlock*>::iterator
270              II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
271         if (pred_begin(*II) == pred_end(*II))
272           WorkList.insert(*II);
273     }
274 
275     // Delete the dead blocks and any of their dead successors.
276     MadeChange |= !WorkList.empty();
277     while (!WorkList.empty()) {
278       BasicBlock *BB = *WorkList.begin();
279       WorkList.erase(BB);
280       SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
281 
282       DeleteDeadBlock(BB);
283 
284       for (SmallVectorImpl<BasicBlock*>::iterator
285              II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
286         if (pred_begin(*II) == pred_end(*II))
287           WorkList.insert(*II);
288     }
289 
290     // Merge pairs of basic blocks with unconditional branches, connected by
291     // a single edge.
292     if (EverMadeChange || MadeChange)
293       MadeChange |= EliminateFallThrough(F);
294 
295     EverMadeChange |= MadeChange;
296   }
297 
298   if (!DisableGCOpts) {
299     SmallVector<Instruction *, 2> Statepoints;
300     for (BasicBlock &BB : F)
301       for (Instruction &I : BB)
302         if (isStatepoint(I))
303           Statepoints.push_back(&I);
304     for (auto &I : Statepoints)
305       EverMadeChange |= simplifyOffsetableRelocate(*I);
306   }
307 
308   return EverMadeChange;
309 }
310 
311 /// EliminateFallThrough - Merge basic blocks which are connected
312 /// by a single edge, where one of the basic blocks has a single successor
313 /// pointing to the other basic block, which has a single predecessor.
EliminateFallThrough(Function & F)314 bool CodeGenPrepare::EliminateFallThrough(Function &F) {
315   bool Changed = false;
316   // Scan all of the blocks in the function, except for the entry block.
317   for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
318     BasicBlock *BB = I++;
319     // If the destination block has a single pred, then this is a trivial
320     // edge, just collapse it.
321     BasicBlock *SinglePred = BB->getSinglePredecessor();
322 
323     // Don't merge if BB's address is taken.
324     if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
325 
326     BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
327     if (Term && !Term->isConditional()) {
328       Changed = true;
329       DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n");
330       // Remember if SinglePred was the entry block of the function.
331       // If so, we will need to move BB back to the entry position.
332       bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
333       MergeBasicBlockIntoOnlyPred(BB, nullptr);
334 
335       if (isEntry && BB != &BB->getParent()->getEntryBlock())
336         BB->moveBefore(&BB->getParent()->getEntryBlock());
337 
338       // We have erased a block. Update the iterator.
339       I = BB;
340     }
341   }
342   return Changed;
343 }
344 
345 /// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes,
346 /// debug info directives, and an unconditional branch.  Passes before isel
347 /// (e.g. LSR/loopsimplify) often split edges in ways that are non-optimal for
348 /// isel.  Start by eliminating these blocks so we can split them the way we
349 /// want them.
EliminateMostlyEmptyBlocks(Function & F)350 bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) {
351   bool MadeChange = false;
352   // Note that this intentionally skips the entry block.
353   for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
354     BasicBlock *BB = I++;
355 
356     // If this block doesn't end with an uncond branch, ignore it.
357     BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
358     if (!BI || !BI->isUnconditional())
359       continue;
360 
361     // If the instruction before the branch (skipping debug info) isn't a phi
362     // node, then other stuff is happening here.
363     BasicBlock::iterator BBI = BI;
364     if (BBI != BB->begin()) {
365       --BBI;
366       while (isa<DbgInfoIntrinsic>(BBI)) {
367         if (BBI == BB->begin())
368           break;
369         --BBI;
370       }
371       if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
372         continue;
373     }
374 
375     // Do not break infinite loops.
376     BasicBlock *DestBB = BI->getSuccessor(0);
377     if (DestBB == BB)
378       continue;
379 
380     if (!CanMergeBlocks(BB, DestBB))
381       continue;
382 
383     EliminateMostlyEmptyBlock(BB);
384     MadeChange = true;
385   }
386   return MadeChange;
387 }
388 
389 /// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a
390 /// single uncond branch between them, and BB contains no other non-phi
391 /// instructions.
CanMergeBlocks(const BasicBlock * BB,const BasicBlock * DestBB) const392 bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB,
393                                     const BasicBlock *DestBB) const {
394   // We only want to eliminate blocks whose phi nodes are used by phi nodes in
395   // the successor.  If there are more complex condition (e.g. preheaders),
396   // don't mess around with them.
397   BasicBlock::const_iterator BBI = BB->begin();
398   while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
399     for (const User *U : PN->users()) {
400       const Instruction *UI = cast<Instruction>(U);
401       if (UI->getParent() != DestBB || !isa<PHINode>(UI))
402         return false;
403       // If User is inside DestBB block and it is a PHINode then check
404       // incoming value. If incoming value is not from BB then this is
405       // a complex condition (e.g. preheaders) we want to avoid here.
406       if (UI->getParent() == DestBB) {
407         if (const PHINode *UPN = dyn_cast<PHINode>(UI))
408           for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
409             Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
410             if (Insn && Insn->getParent() == BB &&
411                 Insn->getParent() != UPN->getIncomingBlock(I))
412               return false;
413           }
414       }
415     }
416   }
417 
418   // If BB and DestBB contain any common predecessors, then the phi nodes in BB
419   // and DestBB may have conflicting incoming values for the block.  If so, we
420   // can't merge the block.
421   const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
422   if (!DestBBPN) return true;  // no conflict.
423 
424   // Collect the preds of BB.
425   SmallPtrSet<const BasicBlock*, 16> BBPreds;
426   if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
427     // It is faster to get preds from a PHI than with pred_iterator.
428     for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
429       BBPreds.insert(BBPN->getIncomingBlock(i));
430   } else {
431     BBPreds.insert(pred_begin(BB), pred_end(BB));
432   }
433 
434   // Walk the preds of DestBB.
435   for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
436     BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
437     if (BBPreds.count(Pred)) {   // Common predecessor?
438       BBI = DestBB->begin();
439       while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
440         const Value *V1 = PN->getIncomingValueForBlock(Pred);
441         const Value *V2 = PN->getIncomingValueForBlock(BB);
442 
443         // If V2 is a phi node in BB, look up what the mapped value will be.
444         if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
445           if (V2PN->getParent() == BB)
446             V2 = V2PN->getIncomingValueForBlock(Pred);
447 
448         // If there is a conflict, bail out.
449         if (V1 != V2) return false;
450       }
451     }
452   }
453 
454   return true;
455 }
456 
457 
458 /// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and
459 /// an unconditional branch in it.
EliminateMostlyEmptyBlock(BasicBlock * BB)460 void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) {
461   BranchInst *BI = cast<BranchInst>(BB->getTerminator());
462   BasicBlock *DestBB = BI->getSuccessor(0);
463 
464   DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
465 
466   // If the destination block has a single pred, then this is a trivial edge,
467   // just collapse it.
468   if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
469     if (SinglePred != DestBB) {
470       // Remember if SinglePred was the entry block of the function.  If so, we
471       // will need to move BB back to the entry position.
472       bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
473       MergeBasicBlockIntoOnlyPred(DestBB, nullptr);
474 
475       if (isEntry && BB != &BB->getParent()->getEntryBlock())
476         BB->moveBefore(&BB->getParent()->getEntryBlock());
477 
478       DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
479       return;
480     }
481   }
482 
483   // Otherwise, we have multiple predecessors of BB.  Update the PHIs in DestBB
484   // to handle the new incoming edges it is about to have.
485   PHINode *PN;
486   for (BasicBlock::iterator BBI = DestBB->begin();
487        (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
488     // Remove the incoming value for BB, and remember it.
489     Value *InVal = PN->removeIncomingValue(BB, false);
490 
491     // Two options: either the InVal is a phi node defined in BB or it is some
492     // value that dominates BB.
493     PHINode *InValPhi = dyn_cast<PHINode>(InVal);
494     if (InValPhi && InValPhi->getParent() == BB) {
495       // Add all of the input values of the input PHI as inputs of this phi.
496       for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
497         PN->addIncoming(InValPhi->getIncomingValue(i),
498                         InValPhi->getIncomingBlock(i));
499     } else {
500       // Otherwise, add one instance of the dominating value for each edge that
501       // we will be adding.
502       if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
503         for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
504           PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
505       } else {
506         for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
507           PN->addIncoming(InVal, *PI);
508       }
509     }
510   }
511 
512   // The PHIs are now updated, change everything that refers to BB to use
513   // DestBB and remove BB.
514   BB->replaceAllUsesWith(DestBB);
515   BB->eraseFromParent();
516   ++NumBlocksElim;
517 
518   DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
519 }
520 
521 // Computes a map of base pointer relocation instructions to corresponding
522 // derived pointer relocation instructions given a vector of all relocate calls
computeBaseDerivedRelocateMap(const SmallVectorImpl<User * > & AllRelocateCalls,DenseMap<IntrinsicInst *,SmallVector<IntrinsicInst *,2>> & RelocateInstMap)523 static void computeBaseDerivedRelocateMap(
524     const SmallVectorImpl<User *> &AllRelocateCalls,
525     DenseMap<IntrinsicInst *, SmallVector<IntrinsicInst *, 2>> &
526         RelocateInstMap) {
527   // Collect information in two maps: one primarily for locating the base object
528   // while filling the second map; the second map is the final structure holding
529   // a mapping between Base and corresponding Derived relocate calls
530   DenseMap<std::pair<unsigned, unsigned>, IntrinsicInst *> RelocateIdxMap;
531   for (auto &U : AllRelocateCalls) {
532     GCRelocateOperands ThisRelocate(U);
533     IntrinsicInst *I = cast<IntrinsicInst>(U);
534     auto K = std::make_pair(ThisRelocate.basePtrIndex(),
535                             ThisRelocate.derivedPtrIndex());
536     RelocateIdxMap.insert(std::make_pair(K, I));
537   }
538   for (auto &Item : RelocateIdxMap) {
539     std::pair<unsigned, unsigned> Key = Item.first;
540     if (Key.first == Key.second)
541       // Base relocation: nothing to insert
542       continue;
543 
544     IntrinsicInst *I = Item.second;
545     auto BaseKey = std::make_pair(Key.first, Key.first);
546 
547     // We're iterating over RelocateIdxMap so we cannot modify it.
548     auto MaybeBase = RelocateIdxMap.find(BaseKey);
549     if (MaybeBase == RelocateIdxMap.end())
550       // TODO: We might want to insert a new base object relocate and gep off
551       // that, if there are enough derived object relocates.
552       continue;
553 
554     RelocateInstMap[MaybeBase->second].push_back(I);
555   }
556 }
557 
558 // Accepts a GEP and extracts the operands into a vector provided they're all
559 // small integer constants
getGEPSmallConstantIntOffsetV(GetElementPtrInst * GEP,SmallVectorImpl<Value * > & OffsetV)560 static bool getGEPSmallConstantIntOffsetV(GetElementPtrInst *GEP,
561                                           SmallVectorImpl<Value *> &OffsetV) {
562   for (unsigned i = 1; i < GEP->getNumOperands(); i++) {
563     // Only accept small constant integer operands
564     auto Op = dyn_cast<ConstantInt>(GEP->getOperand(i));
565     if (!Op || Op->getZExtValue() > 20)
566       return false;
567   }
568 
569   for (unsigned i = 1; i < GEP->getNumOperands(); i++)
570     OffsetV.push_back(GEP->getOperand(i));
571   return true;
572 }
573 
574 // Takes a RelocatedBase (base pointer relocation instruction) and Targets to
575 // replace, computes a replacement, and affects it.
576 static bool
simplifyRelocatesOffABase(IntrinsicInst * RelocatedBase,const SmallVectorImpl<IntrinsicInst * > & Targets)577 simplifyRelocatesOffABase(IntrinsicInst *RelocatedBase,
578                           const SmallVectorImpl<IntrinsicInst *> &Targets) {
579   bool MadeChange = false;
580   for (auto &ToReplace : Targets) {
581     GCRelocateOperands MasterRelocate(RelocatedBase);
582     GCRelocateOperands ThisRelocate(ToReplace);
583 
584     assert(ThisRelocate.basePtrIndex() == MasterRelocate.basePtrIndex() &&
585            "Not relocating a derived object of the original base object");
586     if (ThisRelocate.basePtrIndex() == ThisRelocate.derivedPtrIndex()) {
587       // A duplicate relocate call. TODO: coalesce duplicates.
588       continue;
589     }
590 
591     Value *Base = ThisRelocate.basePtr();
592     auto Derived = dyn_cast<GetElementPtrInst>(ThisRelocate.derivedPtr());
593     if (!Derived || Derived->getPointerOperand() != Base)
594       continue;
595 
596     SmallVector<Value *, 2> OffsetV;
597     if (!getGEPSmallConstantIntOffsetV(Derived, OffsetV))
598       continue;
599 
600     // Create a Builder and replace the target callsite with a gep
601     IRBuilder<> Builder(ToReplace);
602     Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
603     Value *Replacement = Builder.CreateGEP(
604         Derived->getSourceElementType(), RelocatedBase, makeArrayRef(OffsetV));
605     Instruction *ReplacementInst = cast<Instruction>(Replacement);
606     ReplacementInst->removeFromParent();
607     ReplacementInst->insertAfter(RelocatedBase);
608     Replacement->takeName(ToReplace);
609     ToReplace->replaceAllUsesWith(Replacement);
610     ToReplace->eraseFromParent();
611 
612     MadeChange = true;
613   }
614   return MadeChange;
615 }
616 
617 // Turns this:
618 //
619 // %base = ...
620 // %ptr = gep %base + 15
621 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
622 // %base' = relocate(%tok, i32 4, i32 4)
623 // %ptr' = relocate(%tok, i32 4, i32 5)
624 // %val = load %ptr'
625 //
626 // into this:
627 //
628 // %base = ...
629 // %ptr = gep %base + 15
630 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
631 // %base' = gc.relocate(%tok, i32 4, i32 4)
632 // %ptr' = gep %base' + 15
633 // %val = load %ptr'
simplifyOffsetableRelocate(Instruction & I)634 bool CodeGenPrepare::simplifyOffsetableRelocate(Instruction &I) {
635   bool MadeChange = false;
636   SmallVector<User *, 2> AllRelocateCalls;
637 
638   for (auto *U : I.users())
639     if (isGCRelocate(dyn_cast<Instruction>(U)))
640       // Collect all the relocate calls associated with a statepoint
641       AllRelocateCalls.push_back(U);
642 
643   // We need atleast one base pointer relocation + one derived pointer
644   // relocation to mangle
645   if (AllRelocateCalls.size() < 2)
646     return false;
647 
648   // RelocateInstMap is a mapping from the base relocate instruction to the
649   // corresponding derived relocate instructions
650   DenseMap<IntrinsicInst *, SmallVector<IntrinsicInst *, 2>> RelocateInstMap;
651   computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap);
652   if (RelocateInstMap.empty())
653     return false;
654 
655   for (auto &Item : RelocateInstMap)
656     // Item.first is the RelocatedBase to offset against
657     // Item.second is the vector of Targets to replace
658     MadeChange = simplifyRelocatesOffABase(Item.first, Item.second);
659   return MadeChange;
660 }
661 
662 /// SinkCast - Sink the specified cast instruction into its user blocks
SinkCast(CastInst * CI)663 static bool SinkCast(CastInst *CI) {
664   BasicBlock *DefBB = CI->getParent();
665 
666   /// InsertedCasts - Only insert a cast in each block once.
667   DenseMap<BasicBlock*, CastInst*> InsertedCasts;
668 
669   bool MadeChange = false;
670   for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
671        UI != E; ) {
672     Use &TheUse = UI.getUse();
673     Instruction *User = cast<Instruction>(*UI);
674 
675     // Figure out which BB this cast is used in.  For PHI's this is the
676     // appropriate predecessor block.
677     BasicBlock *UserBB = User->getParent();
678     if (PHINode *PN = dyn_cast<PHINode>(User)) {
679       UserBB = PN->getIncomingBlock(TheUse);
680     }
681 
682     // Preincrement use iterator so we don't invalidate it.
683     ++UI;
684 
685     // If this user is in the same block as the cast, don't change the cast.
686     if (UserBB == DefBB) continue;
687 
688     // If we have already inserted a cast into this block, use it.
689     CastInst *&InsertedCast = InsertedCasts[UserBB];
690 
691     if (!InsertedCast) {
692       BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
693       InsertedCast =
694         CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "",
695                          InsertPt);
696     }
697 
698     // Replace a use of the cast with a use of the new cast.
699     TheUse = InsertedCast;
700     MadeChange = true;
701     ++NumCastUses;
702   }
703 
704   // If we removed all uses, nuke the cast.
705   if (CI->use_empty()) {
706     CI->eraseFromParent();
707     MadeChange = true;
708   }
709 
710   return MadeChange;
711 }
712 
713 /// OptimizeNoopCopyExpression - If the specified cast instruction is a noop
714 /// copy (e.g. it's casting from one pointer type to another, i32->i8 on PPC),
715 /// sink it into user blocks to reduce the number of virtual
716 /// registers that must be created and coalesced.
717 ///
718 /// Return true if any changes are made.
719 ///
OptimizeNoopCopyExpression(CastInst * CI,const TargetLowering & TLI)720 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){
721   // If this is a noop copy,
722   EVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType());
723   EVT DstVT = TLI.getValueType(CI->getType());
724 
725   // This is an fp<->int conversion?
726   if (SrcVT.isInteger() != DstVT.isInteger())
727     return false;
728 
729   // If this is an extension, it will be a zero or sign extension, which
730   // isn't a noop.
731   if (SrcVT.bitsLT(DstVT)) return false;
732 
733   // If these values will be promoted, find out what they will be promoted
734   // to.  This helps us consider truncates on PPC as noop copies when they
735   // are.
736   if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
737       TargetLowering::TypePromoteInteger)
738     SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
739   if (TLI.getTypeAction(CI->getContext(), DstVT) ==
740       TargetLowering::TypePromoteInteger)
741     DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
742 
743   // If, after promotion, these are the same types, this is a noop copy.
744   if (SrcVT != DstVT)
745     return false;
746 
747   return SinkCast(CI);
748 }
749 
750 /// CombineUAddWithOverflow - try to combine CI into a call to the
751 /// llvm.uadd.with.overflow intrinsic if possible.
752 ///
753 /// Return true if any changes were made.
CombineUAddWithOverflow(CmpInst * CI)754 static bool CombineUAddWithOverflow(CmpInst *CI) {
755   Value *A, *B;
756   Instruction *AddI;
757   if (!match(CI,
758              m_UAddWithOverflow(m_Value(A), m_Value(B), m_Instruction(AddI))))
759     return false;
760 
761   Type *Ty = AddI->getType();
762   if (!isa<IntegerType>(Ty))
763     return false;
764 
765   // We don't want to move around uses of condition values this late, so we we
766   // check if it is legal to create the call to the intrinsic in the basic
767   // block containing the icmp:
768 
769   if (AddI->getParent() != CI->getParent() && !AddI->hasOneUse())
770     return false;
771 
772 #ifndef NDEBUG
773   // Someday m_UAddWithOverflow may get smarter, but this is a safe assumption
774   // for now:
775   if (AddI->hasOneUse())
776     assert(*AddI->user_begin() == CI && "expected!");
777 #endif
778 
779   Module *M = CI->getParent()->getParent()->getParent();
780   Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
781 
782   auto *InsertPt = AddI->hasOneUse() ? CI : AddI;
783 
784   auto *UAddWithOverflow =
785       CallInst::Create(F, {A, B}, "uadd.overflow", InsertPt);
786   auto *UAdd = ExtractValueInst::Create(UAddWithOverflow, 0, "uadd", InsertPt);
787   auto *Overflow =
788       ExtractValueInst::Create(UAddWithOverflow, 1, "overflow", InsertPt);
789 
790   CI->replaceAllUsesWith(Overflow);
791   AddI->replaceAllUsesWith(UAdd);
792   CI->eraseFromParent();
793   AddI->eraseFromParent();
794   return true;
795 }
796 
797 /// SinkCmpExpression - Sink the given CmpInst into user blocks to reduce
798 /// the number of virtual registers that must be created and coalesced.  This is
799 /// a clear win except on targets with multiple condition code registers
800 ///  (PowerPC), where it might lose; some adjustment may be wanted there.
801 ///
802 /// Return true if any changes are made.
SinkCmpExpression(CmpInst * CI)803 static bool SinkCmpExpression(CmpInst *CI) {
804   BasicBlock *DefBB = CI->getParent();
805 
806   /// InsertedCmp - Only insert a cmp in each block once.
807   DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
808 
809   bool MadeChange = false;
810   for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
811        UI != E; ) {
812     Use &TheUse = UI.getUse();
813     Instruction *User = cast<Instruction>(*UI);
814 
815     // Preincrement use iterator so we don't invalidate it.
816     ++UI;
817 
818     // Don't bother for PHI nodes.
819     if (isa<PHINode>(User))
820       continue;
821 
822     // Figure out which BB this cmp is used in.
823     BasicBlock *UserBB = User->getParent();
824 
825     // If this user is in the same block as the cmp, don't change the cmp.
826     if (UserBB == DefBB) continue;
827 
828     // If we have already inserted a cmp into this block, use it.
829     CmpInst *&InsertedCmp = InsertedCmps[UserBB];
830 
831     if (!InsertedCmp) {
832       BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
833       InsertedCmp =
834         CmpInst::Create(CI->getOpcode(),
835                         CI->getPredicate(),  CI->getOperand(0),
836                         CI->getOperand(1), "", InsertPt);
837     }
838 
839     // Replace a use of the cmp with a use of the new cmp.
840     TheUse = InsertedCmp;
841     MadeChange = true;
842     ++NumCmpUses;
843   }
844 
845   // If we removed all uses, nuke the cmp.
846   if (CI->use_empty()) {
847     CI->eraseFromParent();
848     MadeChange = true;
849   }
850 
851   return MadeChange;
852 }
853 
OptimizeCmpExpression(CmpInst * CI)854 static bool OptimizeCmpExpression(CmpInst *CI) {
855   if (SinkCmpExpression(CI))
856     return true;
857 
858   if (CombineUAddWithOverflow(CI))
859     return true;
860 
861   return false;
862 }
863 
864 /// isExtractBitsCandidateUse - Check if the candidates could
865 /// be combined with shift instruction, which includes:
866 /// 1. Truncate instruction
867 /// 2. And instruction and the imm is a mask of the low bits:
868 /// imm & (imm+1) == 0
isExtractBitsCandidateUse(Instruction * User)869 static bool isExtractBitsCandidateUse(Instruction *User) {
870   if (!isa<TruncInst>(User)) {
871     if (User->getOpcode() != Instruction::And ||
872         !isa<ConstantInt>(User->getOperand(1)))
873       return false;
874 
875     const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
876 
877     if ((Cimm & (Cimm + 1)).getBoolValue())
878       return false;
879   }
880   return true;
881 }
882 
883 /// SinkShiftAndTruncate - sink both shift and truncate instruction
884 /// to the use of truncate's BB.
885 static bool
SinkShiftAndTruncate(BinaryOperator * ShiftI,Instruction * User,ConstantInt * CI,DenseMap<BasicBlock *,BinaryOperator * > & InsertedShifts,const TargetLowering & TLI)886 SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI,
887                      DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts,
888                      const TargetLowering &TLI) {
889   BasicBlock *UserBB = User->getParent();
890   DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
891   TruncInst *TruncI = dyn_cast<TruncInst>(User);
892   bool MadeChange = false;
893 
894   for (Value::user_iterator TruncUI = TruncI->user_begin(),
895                             TruncE = TruncI->user_end();
896        TruncUI != TruncE;) {
897 
898     Use &TruncTheUse = TruncUI.getUse();
899     Instruction *TruncUser = cast<Instruction>(*TruncUI);
900     // Preincrement use iterator so we don't invalidate it.
901 
902     ++TruncUI;
903 
904     int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
905     if (!ISDOpcode)
906       continue;
907 
908     // If the use is actually a legal node, there will not be an
909     // implicit truncate.
910     // FIXME: always querying the result type is just an
911     // approximation; some nodes' legality is determined by the
912     // operand or other means. There's no good way to find out though.
913     if (TLI.isOperationLegalOrCustom(
914             ISDOpcode, TLI.getValueType(TruncUser->getType(), true)))
915       continue;
916 
917     // Don't bother for PHI nodes.
918     if (isa<PHINode>(TruncUser))
919       continue;
920 
921     BasicBlock *TruncUserBB = TruncUser->getParent();
922 
923     if (UserBB == TruncUserBB)
924       continue;
925 
926     BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
927     CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
928 
929     if (!InsertedShift && !InsertedTrunc) {
930       BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
931       // Sink the shift
932       if (ShiftI->getOpcode() == Instruction::AShr)
933         InsertedShift =
934             BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI, "", InsertPt);
935       else
936         InsertedShift =
937             BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI, "", InsertPt);
938 
939       // Sink the trunc
940       BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
941       TruncInsertPt++;
942 
943       InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
944                                        TruncI->getType(), "", TruncInsertPt);
945 
946       MadeChange = true;
947 
948       TruncTheUse = InsertedTrunc;
949     }
950   }
951   return MadeChange;
952 }
953 
954 /// OptimizeExtractBits - sink the shift *right* instruction into user blocks if
955 /// the uses could potentially be combined with this shift instruction and
956 /// generate BitExtract instruction. It will only be applied if the architecture
957 /// supports BitExtract instruction. Here is an example:
958 /// BB1:
959 ///   %x.extract.shift = lshr i64 %arg1, 32
960 /// BB2:
961 ///   %x.extract.trunc = trunc i64 %x.extract.shift to i16
962 /// ==>
963 ///
964 /// BB2:
965 ///   %x.extract.shift.1 = lshr i64 %arg1, 32
966 ///   %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
967 ///
968 /// CodeGen will recoginze the pattern in BB2 and generate BitExtract
969 /// instruction.
970 /// Return true if any changes are made.
OptimizeExtractBits(BinaryOperator * ShiftI,ConstantInt * CI,const TargetLowering & TLI)971 static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI,
972                                 const TargetLowering &TLI) {
973   BasicBlock *DefBB = ShiftI->getParent();
974 
975   /// Only insert instructions in each block once.
976   DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts;
977 
978   bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(ShiftI->getType()));
979 
980   bool MadeChange = false;
981   for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
982        UI != E;) {
983     Use &TheUse = UI.getUse();
984     Instruction *User = cast<Instruction>(*UI);
985     // Preincrement use iterator so we don't invalidate it.
986     ++UI;
987 
988     // Don't bother for PHI nodes.
989     if (isa<PHINode>(User))
990       continue;
991 
992     if (!isExtractBitsCandidateUse(User))
993       continue;
994 
995     BasicBlock *UserBB = User->getParent();
996 
997     if (UserBB == DefBB) {
998       // If the shift and truncate instruction are in the same BB. The use of
999       // the truncate(TruncUse) may still introduce another truncate if not
1000       // legal. In this case, we would like to sink both shift and truncate
1001       // instruction to the BB of TruncUse.
1002       // for example:
1003       // BB1:
1004       // i64 shift.result = lshr i64 opnd, imm
1005       // trunc.result = trunc shift.result to i16
1006       //
1007       // BB2:
1008       //   ----> We will have an implicit truncate here if the architecture does
1009       //   not have i16 compare.
1010       // cmp i16 trunc.result, opnd2
1011       //
1012       if (isa<TruncInst>(User) && shiftIsLegal
1013           // If the type of the truncate is legal, no trucate will be
1014           // introduced in other basic blocks.
1015           && (!TLI.isTypeLegal(TLI.getValueType(User->getType()))))
1016         MadeChange =
1017             SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI);
1018 
1019       continue;
1020     }
1021     // If we have already inserted a shift into this block, use it.
1022     BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
1023 
1024     if (!InsertedShift) {
1025       BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1026 
1027       if (ShiftI->getOpcode() == Instruction::AShr)
1028         InsertedShift =
1029             BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI, "", InsertPt);
1030       else
1031         InsertedShift =
1032             BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI, "", InsertPt);
1033 
1034       MadeChange = true;
1035     }
1036 
1037     // Replace a use of the shift with a use of the new shift.
1038     TheUse = InsertedShift;
1039   }
1040 
1041   // If we removed all uses, nuke the shift.
1042   if (ShiftI->use_empty())
1043     ShiftI->eraseFromParent();
1044 
1045   return MadeChange;
1046 }
1047 
1048 //  ScalarizeMaskedLoad() translates masked load intrinsic, like
1049 // <16 x i32 > @llvm.masked.load( <16 x i32>* %addr, i32 align,
1050 //                               <16 x i1> %mask, <16 x i32> %passthru)
1051 // to a chain of basic blocks, whith loading element one-by-one if
1052 // the appropriate mask bit is set
1053 //
1054 //  %1 = bitcast i8* %addr to i32*
1055 //  %2 = extractelement <16 x i1> %mask, i32 0
1056 //  %3 = icmp eq i1 %2, true
1057 //  br i1 %3, label %cond.load, label %else
1058 //
1059 //cond.load:                                        ; preds = %0
1060 //  %4 = getelementptr i32* %1, i32 0
1061 //  %5 = load i32* %4
1062 //  %6 = insertelement <16 x i32> undef, i32 %5, i32 0
1063 //  br label %else
1064 //
1065 //else:                                             ; preds = %0, %cond.load
1066 //  %res.phi.else = phi <16 x i32> [ %6, %cond.load ], [ undef, %0 ]
1067 //  %7 = extractelement <16 x i1> %mask, i32 1
1068 //  %8 = icmp eq i1 %7, true
1069 //  br i1 %8, label %cond.load1, label %else2
1070 //
1071 //cond.load1:                                       ; preds = %else
1072 //  %9 = getelementptr i32* %1, i32 1
1073 //  %10 = load i32* %9
1074 //  %11 = insertelement <16 x i32> %res.phi.else, i32 %10, i32 1
1075 //  br label %else2
1076 //
1077 //else2:                                            ; preds = %else, %cond.load1
1078 //  %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
1079 //  %12 = extractelement <16 x i1> %mask, i32 2
1080 //  %13 = icmp eq i1 %12, true
1081 //  br i1 %13, label %cond.load4, label %else5
1082 //
ScalarizeMaskedLoad(CallInst * CI)1083 static void ScalarizeMaskedLoad(CallInst *CI) {
1084   Value *Ptr  = CI->getArgOperand(0);
1085   Value *Src0 = CI->getArgOperand(3);
1086   Value *Mask = CI->getArgOperand(2);
1087   VectorType *VecType = dyn_cast<VectorType>(CI->getType());
1088   Type *EltTy = VecType->getElementType();
1089 
1090   assert(VecType && "Unexpected return type of masked load intrinsic");
1091 
1092   IRBuilder<> Builder(CI->getContext());
1093   Instruction *InsertPt = CI;
1094   BasicBlock *IfBlock = CI->getParent();
1095   BasicBlock *CondBlock = nullptr;
1096   BasicBlock *PrevIfBlock = CI->getParent();
1097   Builder.SetInsertPoint(InsertPt);
1098 
1099   Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1100 
1101   // Bitcast %addr fron i8* to EltTy*
1102   Type *NewPtrType =
1103     EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
1104   Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
1105   Value *UndefVal = UndefValue::get(VecType);
1106 
1107   // The result vector
1108   Value *VResult = UndefVal;
1109 
1110   PHINode *Phi = nullptr;
1111   Value *PrevPhi = UndefVal;
1112 
1113   unsigned VectorWidth = VecType->getNumElements();
1114   for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1115 
1116     // Fill the "else" block, created in the previous iteration
1117     //
1118     //  %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
1119     //  %mask_1 = extractelement <16 x i1> %mask, i32 Idx
1120     //  %to_load = icmp eq i1 %mask_1, true
1121     //  br i1 %to_load, label %cond.load, label %else
1122     //
1123     if (Idx > 0) {
1124       Phi = Builder.CreatePHI(VecType, 2, "res.phi.else");
1125       Phi->addIncoming(VResult, CondBlock);
1126       Phi->addIncoming(PrevPhi, PrevIfBlock);
1127       PrevPhi = Phi;
1128       VResult = Phi;
1129     }
1130 
1131     Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
1132     Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1133                                     ConstantInt::get(Predicate->getType(), 1));
1134 
1135     // Create "cond" block
1136     //
1137     //  %EltAddr = getelementptr i32* %1, i32 0
1138     //  %Elt = load i32* %EltAddr
1139     //  VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx
1140     //
1141     CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.load");
1142     Builder.SetInsertPoint(InsertPt);
1143 
1144     Value *Gep =
1145         Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1146     LoadInst* Load = Builder.CreateLoad(Gep, false);
1147     VResult = Builder.CreateInsertElement(VResult, Load, Builder.getInt32(Idx));
1148 
1149     // Create "else" block, fill it in the next iteration
1150     BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1151     Builder.SetInsertPoint(InsertPt);
1152     Instruction *OldBr = IfBlock->getTerminator();
1153     BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1154     OldBr->eraseFromParent();
1155     PrevIfBlock = IfBlock;
1156     IfBlock = NewIfBlock;
1157   }
1158 
1159   Phi = Builder.CreatePHI(VecType, 2, "res.phi.select");
1160   Phi->addIncoming(VResult, CondBlock);
1161   Phi->addIncoming(PrevPhi, PrevIfBlock);
1162   Value *NewI = Builder.CreateSelect(Mask, Phi, Src0);
1163   CI->replaceAllUsesWith(NewI);
1164   CI->eraseFromParent();
1165 }
1166 
1167 //  ScalarizeMaskedStore() translates masked store intrinsic, like
1168 // void @llvm.masked.store(<16 x i32> %src, <16 x i32>* %addr, i32 align,
1169 //                               <16 x i1> %mask)
1170 // to a chain of basic blocks, that stores element one-by-one if
1171 // the appropriate mask bit is set
1172 //
1173 //   %1 = bitcast i8* %addr to i32*
1174 //   %2 = extractelement <16 x i1> %mask, i32 0
1175 //   %3 = icmp eq i1 %2, true
1176 //   br i1 %3, label %cond.store, label %else
1177 //
1178 // cond.store:                                       ; preds = %0
1179 //   %4 = extractelement <16 x i32> %val, i32 0
1180 //   %5 = getelementptr i32* %1, i32 0
1181 //   store i32 %4, i32* %5
1182 //   br label %else
1183 //
1184 // else:                                             ; preds = %0, %cond.store
1185 //   %6 = extractelement <16 x i1> %mask, i32 1
1186 //   %7 = icmp eq i1 %6, true
1187 //   br i1 %7, label %cond.store1, label %else2
1188 //
1189 // cond.store1:                                      ; preds = %else
1190 //   %8 = extractelement <16 x i32> %val, i32 1
1191 //   %9 = getelementptr i32* %1, i32 1
1192 //   store i32 %8, i32* %9
1193 //   br label %else2
1194 //   . . .
ScalarizeMaskedStore(CallInst * CI)1195 static void ScalarizeMaskedStore(CallInst *CI) {
1196   Value *Ptr  = CI->getArgOperand(1);
1197   Value *Src = CI->getArgOperand(0);
1198   Value *Mask = CI->getArgOperand(3);
1199 
1200   VectorType *VecType = dyn_cast<VectorType>(Src->getType());
1201   Type *EltTy = VecType->getElementType();
1202 
1203   assert(VecType && "Unexpected data type in masked store intrinsic");
1204 
1205   IRBuilder<> Builder(CI->getContext());
1206   Instruction *InsertPt = CI;
1207   BasicBlock *IfBlock = CI->getParent();
1208   Builder.SetInsertPoint(InsertPt);
1209   Builder.SetCurrentDebugLocation(CI->getDebugLoc());
1210 
1211   // Bitcast %addr fron i8* to EltTy*
1212   Type *NewPtrType =
1213     EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
1214   Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
1215 
1216   unsigned VectorWidth = VecType->getNumElements();
1217   for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
1218 
1219     // Fill the "else" block, created in the previous iteration
1220     //
1221     //  %mask_1 = extractelement <16 x i1> %mask, i32 Idx
1222     //  %to_store = icmp eq i1 %mask_1, true
1223     //  br i1 %to_load, label %cond.store, label %else
1224     //
1225     Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
1226     Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
1227                                     ConstantInt::get(Predicate->getType(), 1));
1228 
1229     // Create "cond" block
1230     //
1231     //  %OneElt = extractelement <16 x i32> %Src, i32 Idx
1232     //  %EltAddr = getelementptr i32* %1, i32 0
1233     //  %store i32 %OneElt, i32* %EltAddr
1234     //
1235     BasicBlock *CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1236     Builder.SetInsertPoint(InsertPt);
1237 
1238     Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx));
1239     Value *Gep =
1240         Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
1241     Builder.CreateStore(OneElt, Gep);
1242 
1243     // Create "else" block, fill it in the next iteration
1244     BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1245     Builder.SetInsertPoint(InsertPt);
1246     Instruction *OldBr = IfBlock->getTerminator();
1247     BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1248     OldBr->eraseFromParent();
1249     IfBlock = NewIfBlock;
1250   }
1251   CI->eraseFromParent();
1252 }
1253 
OptimizeCallInst(CallInst * CI,bool & ModifiedDT)1254 bool CodeGenPrepare::OptimizeCallInst(CallInst *CI, bool& ModifiedDT) {
1255   BasicBlock *BB = CI->getParent();
1256 
1257   // Lower inline assembly if we can.
1258   // If we found an inline asm expession, and if the target knows how to
1259   // lower it to normal LLVM code, do so now.
1260   if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
1261     if (TLI->ExpandInlineAsm(CI)) {
1262       // Avoid invalidating the iterator.
1263       CurInstIterator = BB->begin();
1264       // Avoid processing instructions out of order, which could cause
1265       // reuse before a value is defined.
1266       SunkAddrs.clear();
1267       return true;
1268     }
1269     // Sink address computing for memory operands into the block.
1270     if (OptimizeInlineAsmInst(CI))
1271       return true;
1272   }
1273 
1274   const DataLayout *TD = TLI ? TLI->getDataLayout() : nullptr;
1275 
1276   // Align the pointer arguments to this call if the target thinks it's a good
1277   // idea
1278   unsigned MinSize, PrefAlign;
1279   if (TLI && TD && TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) {
1280     for (auto &Arg : CI->arg_operands()) {
1281       // We want to align both objects whose address is used directly and
1282       // objects whose address is used in casts and GEPs, though it only makes
1283       // sense for GEPs if the offset is a multiple of the desired alignment and
1284       // if size - offset meets the size threshold.
1285       if (!Arg->getType()->isPointerTy())
1286         continue;
1287       APInt Offset(TD->getPointerSizeInBits(
1288                      cast<PointerType>(Arg->getType())->getAddressSpace()), 0);
1289       Value *Val = Arg->stripAndAccumulateInBoundsConstantOffsets(*TD, Offset);
1290       uint64_t Offset2 = Offset.getLimitedValue();
1291       if ((Offset2 & (PrefAlign-1)) != 0)
1292         continue;
1293       AllocaInst *AI;
1294       if ((AI = dyn_cast<AllocaInst>(Val)) &&
1295           AI->getAlignment() < PrefAlign &&
1296           TD->getTypeAllocSize(AI->getAllocatedType()) >= MinSize + Offset2)
1297         AI->setAlignment(PrefAlign);
1298       // Global variables can only be aligned if they are defined in this
1299       // object (i.e. they are uniquely initialized in this object), and
1300       // over-aligning global variables that have an explicit section is
1301       // forbidden.
1302       GlobalVariable *GV;
1303       if ((GV = dyn_cast<GlobalVariable>(Val)) &&
1304           GV->hasUniqueInitializer() &&
1305           !GV->hasSection() &&
1306           GV->getAlignment() < PrefAlign &&
1307           TD->getTypeAllocSize(
1308             GV->getType()->getElementType()) >= MinSize + Offset2)
1309         GV->setAlignment(PrefAlign);
1310     }
1311     // If this is a memcpy (or similar) then we may be able to improve the
1312     // alignment
1313     if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(CI)) {
1314       unsigned Align = getKnownAlignment(MI->getDest(), *TD);
1315       if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI))
1316         Align = std::min(Align, getKnownAlignment(MTI->getSource(), *TD));
1317       if (Align > MI->getAlignment())
1318         MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), Align));
1319     }
1320   }
1321 
1322   IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
1323   if (II) {
1324     switch (II->getIntrinsicID()) {
1325     default: break;
1326     case Intrinsic::objectsize: {
1327       // Lower all uses of llvm.objectsize.*
1328       bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
1329       Type *ReturnTy = CI->getType();
1330       Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
1331 
1332       // Substituting this can cause recursive simplifications, which can
1333       // invalidate our iterator.  Use a WeakVH to hold onto it in case this
1334       // happens.
1335       WeakVH IterHandle(CurInstIterator);
1336 
1337       replaceAndRecursivelySimplify(CI, RetVal,
1338                                     TLInfo, nullptr);
1339 
1340       // If the iterator instruction was recursively deleted, start over at the
1341       // start of the block.
1342       if (IterHandle != CurInstIterator) {
1343         CurInstIterator = BB->begin();
1344         SunkAddrs.clear();
1345       }
1346       return true;
1347     }
1348     case Intrinsic::masked_load: {
1349       // Scalarize unsupported vector masked load
1350       if (!TTI->isLegalMaskedLoad(CI->getType(), 1)) {
1351         ScalarizeMaskedLoad(CI);
1352         ModifiedDT = true;
1353         return true;
1354       }
1355       return false;
1356     }
1357     case Intrinsic::masked_store: {
1358       if (!TTI->isLegalMaskedStore(CI->getArgOperand(0)->getType(), 1)) {
1359         ScalarizeMaskedStore(CI);
1360         ModifiedDT = true;
1361         return true;
1362       }
1363       return false;
1364     }
1365     }
1366 
1367     if (TLI) {
1368       SmallVector<Value*, 2> PtrOps;
1369       Type *AccessTy;
1370       if (TLI->GetAddrModeArguments(II, PtrOps, AccessTy))
1371         while (!PtrOps.empty())
1372           if (OptimizeMemoryInst(II, PtrOps.pop_back_val(), AccessTy))
1373             return true;
1374     }
1375   }
1376 
1377   // From here on out we're working with named functions.
1378   if (!CI->getCalledFunction()) return false;
1379 
1380   // Lower all default uses of _chk calls.  This is very similar
1381   // to what InstCombineCalls does, but here we are only lowering calls
1382   // to fortified library functions (e.g. __memcpy_chk) that have the default
1383   // "don't know" as the objectsize.  Anything else should be left alone.
1384   FortifiedLibCallSimplifier Simplifier(TLInfo, true);
1385   if (Value *V = Simplifier.optimizeCall(CI)) {
1386     CI->replaceAllUsesWith(V);
1387     CI->eraseFromParent();
1388     return true;
1389   }
1390   return false;
1391 }
1392 
1393 /// DupRetToEnableTailCallOpts - Look for opportunities to duplicate return
1394 /// instructions to the predecessor to enable tail call optimizations. The
1395 /// case it is currently looking for is:
1396 /// @code
1397 /// bb0:
1398 ///   %tmp0 = tail call i32 @f0()
1399 ///   br label %return
1400 /// bb1:
1401 ///   %tmp1 = tail call i32 @f1()
1402 ///   br label %return
1403 /// bb2:
1404 ///   %tmp2 = tail call i32 @f2()
1405 ///   br label %return
1406 /// return:
1407 ///   %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
1408 ///   ret i32 %retval
1409 /// @endcode
1410 ///
1411 /// =>
1412 ///
1413 /// @code
1414 /// bb0:
1415 ///   %tmp0 = tail call i32 @f0()
1416 ///   ret i32 %tmp0
1417 /// bb1:
1418 ///   %tmp1 = tail call i32 @f1()
1419 ///   ret i32 %tmp1
1420 /// bb2:
1421 ///   %tmp2 = tail call i32 @f2()
1422 ///   ret i32 %tmp2
1423 /// @endcode
DupRetToEnableTailCallOpts(BasicBlock * BB)1424 bool CodeGenPrepare::DupRetToEnableTailCallOpts(BasicBlock *BB) {
1425   if (!TLI)
1426     return false;
1427 
1428   ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator());
1429   if (!RI)
1430     return false;
1431 
1432   PHINode *PN = nullptr;
1433   BitCastInst *BCI = nullptr;
1434   Value *V = RI->getReturnValue();
1435   if (V) {
1436     BCI = dyn_cast<BitCastInst>(V);
1437     if (BCI)
1438       V = BCI->getOperand(0);
1439 
1440     PN = dyn_cast<PHINode>(V);
1441     if (!PN)
1442       return false;
1443   }
1444 
1445   if (PN && PN->getParent() != BB)
1446     return false;
1447 
1448   // It's not safe to eliminate the sign / zero extension of the return value.
1449   // See llvm::isInTailCallPosition().
1450   const Function *F = BB->getParent();
1451   AttributeSet CallerAttrs = F->getAttributes();
1452   if (CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt) ||
1453       CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
1454     return false;
1455 
1456   // Make sure there are no instructions between the PHI and return, or that the
1457   // return is the first instruction in the block.
1458   if (PN) {
1459     BasicBlock::iterator BI = BB->begin();
1460     do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
1461     if (&*BI == BCI)
1462       // Also skip over the bitcast.
1463       ++BI;
1464     if (&*BI != RI)
1465       return false;
1466   } else {
1467     BasicBlock::iterator BI = BB->begin();
1468     while (isa<DbgInfoIntrinsic>(BI)) ++BI;
1469     if (&*BI != RI)
1470       return false;
1471   }
1472 
1473   /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
1474   /// call.
1475   SmallVector<CallInst*, 4> TailCalls;
1476   if (PN) {
1477     for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
1478       CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
1479       // Make sure the phi value is indeed produced by the tail call.
1480       if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
1481           TLI->mayBeEmittedAsTailCall(CI))
1482         TailCalls.push_back(CI);
1483     }
1484   } else {
1485     SmallPtrSet<BasicBlock*, 4> VisitedBBs;
1486     for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
1487       if (!VisitedBBs.insert(*PI).second)
1488         continue;
1489 
1490       BasicBlock::InstListType &InstList = (*PI)->getInstList();
1491       BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
1492       BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
1493       do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
1494       if (RI == RE)
1495         continue;
1496 
1497       CallInst *CI = dyn_cast<CallInst>(&*RI);
1498       if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
1499         TailCalls.push_back(CI);
1500     }
1501   }
1502 
1503   bool Changed = false;
1504   for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
1505     CallInst *CI = TailCalls[i];
1506     CallSite CS(CI);
1507 
1508     // Conservatively require the attributes of the call to match those of the
1509     // return. Ignore noalias because it doesn't affect the call sequence.
1510     AttributeSet CalleeAttrs = CS.getAttributes();
1511     if (AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1512           removeAttribute(Attribute::NoAlias) !=
1513         AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
1514           removeAttribute(Attribute::NoAlias))
1515       continue;
1516 
1517     // Make sure the call instruction is followed by an unconditional branch to
1518     // the return block.
1519     BasicBlock *CallBB = CI->getParent();
1520     BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
1521     if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
1522       continue;
1523 
1524     // Duplicate the return into CallBB.
1525     (void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
1526     ModifiedDT = Changed = true;
1527     ++NumRetsDup;
1528   }
1529 
1530   // If we eliminated all predecessors of the block, delete the block now.
1531   if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
1532     BB->eraseFromParent();
1533 
1534   return Changed;
1535 }
1536 
1537 //===----------------------------------------------------------------------===//
1538 // Memory Optimization
1539 //===----------------------------------------------------------------------===//
1540 
1541 namespace {
1542 
1543 /// ExtAddrMode - This is an extended version of TargetLowering::AddrMode
1544 /// which holds actual Value*'s for register values.
1545 struct ExtAddrMode : public TargetLowering::AddrMode {
1546   Value *BaseReg;
1547   Value *ScaledReg;
ExtAddrMode__anonebd7bfba0211::ExtAddrMode1548   ExtAddrMode() : BaseReg(nullptr), ScaledReg(nullptr) {}
1549   void print(raw_ostream &OS) const;
1550   void dump() const;
1551 
operator ==__anonebd7bfba0211::ExtAddrMode1552   bool operator==(const ExtAddrMode& O) const {
1553     return (BaseReg == O.BaseReg) && (ScaledReg == O.ScaledReg) &&
1554            (BaseGV == O.BaseGV) && (BaseOffs == O.BaseOffs) &&
1555            (HasBaseReg == O.HasBaseReg) && (Scale == O.Scale);
1556   }
1557 };
1558 
1559 #ifndef NDEBUG
operator <<(raw_ostream & OS,const ExtAddrMode & AM)1560 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
1561   AM.print(OS);
1562   return OS;
1563 }
1564 #endif
1565 
print(raw_ostream & OS) const1566 void ExtAddrMode::print(raw_ostream &OS) const {
1567   bool NeedPlus = false;
1568   OS << "[";
1569   if (BaseGV) {
1570     OS << (NeedPlus ? " + " : "")
1571        << "GV:";
1572     BaseGV->printAsOperand(OS, /*PrintType=*/false);
1573     NeedPlus = true;
1574   }
1575 
1576   if (BaseOffs) {
1577     OS << (NeedPlus ? " + " : "")
1578        << BaseOffs;
1579     NeedPlus = true;
1580   }
1581 
1582   if (BaseReg) {
1583     OS << (NeedPlus ? " + " : "")
1584        << "Base:";
1585     BaseReg->printAsOperand(OS, /*PrintType=*/false);
1586     NeedPlus = true;
1587   }
1588   if (Scale) {
1589     OS << (NeedPlus ? " + " : "")
1590        << Scale << "*";
1591     ScaledReg->printAsOperand(OS, /*PrintType=*/false);
1592   }
1593 
1594   OS << ']';
1595 }
1596 
1597 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
dump() const1598 void ExtAddrMode::dump() const {
1599   print(dbgs());
1600   dbgs() << '\n';
1601 }
1602 #endif
1603 
1604 /// \brief This class provides transaction based operation on the IR.
1605 /// Every change made through this class is recorded in the internal state and
1606 /// can be undone (rollback) until commit is called.
1607 class TypePromotionTransaction {
1608 
1609   /// \brief This represents the common interface of the individual transaction.
1610   /// Each class implements the logic for doing one specific modification on
1611   /// the IR via the TypePromotionTransaction.
1612   class TypePromotionAction {
1613   protected:
1614     /// The Instruction modified.
1615     Instruction *Inst;
1616 
1617   public:
1618     /// \brief Constructor of the action.
1619     /// The constructor performs the related action on the IR.
TypePromotionAction(Instruction * Inst)1620     TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
1621 
~TypePromotionAction()1622     virtual ~TypePromotionAction() {}
1623 
1624     /// \brief Undo the modification done by this action.
1625     /// When this method is called, the IR must be in the same state as it was
1626     /// before this action was applied.
1627     /// \pre Undoing the action works if and only if the IR is in the exact same
1628     /// state as it was directly after this action was applied.
1629     virtual void undo() = 0;
1630 
1631     /// \brief Advocate every change made by this action.
1632     /// When the results on the IR of the action are to be kept, it is important
1633     /// to call this function, otherwise hidden information may be kept forever.
commit()1634     virtual void commit() {
1635       // Nothing to be done, this action is not doing anything.
1636     }
1637   };
1638 
1639   /// \brief Utility to remember the position of an instruction.
1640   class InsertionHandler {
1641     /// Position of an instruction.
1642     /// Either an instruction:
1643     /// - Is the first in a basic block: BB is used.
1644     /// - Has a previous instructon: PrevInst is used.
1645     union {
1646       Instruction *PrevInst;
1647       BasicBlock *BB;
1648     } Point;
1649     /// Remember whether or not the instruction had a previous instruction.
1650     bool HasPrevInstruction;
1651 
1652   public:
1653     /// \brief Record the position of \p Inst.
InsertionHandler(Instruction * Inst)1654     InsertionHandler(Instruction *Inst) {
1655       BasicBlock::iterator It = Inst;
1656       HasPrevInstruction = (It != (Inst->getParent()->begin()));
1657       if (HasPrevInstruction)
1658         Point.PrevInst = --It;
1659       else
1660         Point.BB = Inst->getParent();
1661     }
1662 
1663     /// \brief Insert \p Inst at the recorded position.
insert(Instruction * Inst)1664     void insert(Instruction *Inst) {
1665       if (HasPrevInstruction) {
1666         if (Inst->getParent())
1667           Inst->removeFromParent();
1668         Inst->insertAfter(Point.PrevInst);
1669       } else {
1670         Instruction *Position = Point.BB->getFirstInsertionPt();
1671         if (Inst->getParent())
1672           Inst->moveBefore(Position);
1673         else
1674           Inst->insertBefore(Position);
1675       }
1676     }
1677   };
1678 
1679   /// \brief Move an instruction before another.
1680   class InstructionMoveBefore : public TypePromotionAction {
1681     /// Original position of the instruction.
1682     InsertionHandler Position;
1683 
1684   public:
1685     /// \brief Move \p Inst before \p Before.
InstructionMoveBefore(Instruction * Inst,Instruction * Before)1686     InstructionMoveBefore(Instruction *Inst, Instruction *Before)
1687         : TypePromotionAction(Inst), Position(Inst) {
1688       DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before << "\n");
1689       Inst->moveBefore(Before);
1690     }
1691 
1692     /// \brief Move the instruction back to its original position.
undo()1693     void undo() override {
1694       DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
1695       Position.insert(Inst);
1696     }
1697   };
1698 
1699   /// \brief Set the operand of an instruction with a new value.
1700   class OperandSetter : public TypePromotionAction {
1701     /// Original operand of the instruction.
1702     Value *Origin;
1703     /// Index of the modified instruction.
1704     unsigned Idx;
1705 
1706   public:
1707     /// \brief Set \p Idx operand of \p Inst with \p NewVal.
OperandSetter(Instruction * Inst,unsigned Idx,Value * NewVal)1708     OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
1709         : TypePromotionAction(Inst), Idx(Idx) {
1710       DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
1711                    << "for:" << *Inst << "\n"
1712                    << "with:" << *NewVal << "\n");
1713       Origin = Inst->getOperand(Idx);
1714       Inst->setOperand(Idx, NewVal);
1715     }
1716 
1717     /// \brief Restore the original value of the instruction.
undo()1718     void undo() override {
1719       DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
1720                    << "for: " << *Inst << "\n"
1721                    << "with: " << *Origin << "\n");
1722       Inst->setOperand(Idx, Origin);
1723     }
1724   };
1725 
1726   /// \brief Hide the operands of an instruction.
1727   /// Do as if this instruction was not using any of its operands.
1728   class OperandsHider : public TypePromotionAction {
1729     /// The list of original operands.
1730     SmallVector<Value *, 4> OriginalValues;
1731 
1732   public:
1733     /// \brief Remove \p Inst from the uses of the operands of \p Inst.
OperandsHider(Instruction * Inst)1734     OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
1735       DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
1736       unsigned NumOpnds = Inst->getNumOperands();
1737       OriginalValues.reserve(NumOpnds);
1738       for (unsigned It = 0; It < NumOpnds; ++It) {
1739         // Save the current operand.
1740         Value *Val = Inst->getOperand(It);
1741         OriginalValues.push_back(Val);
1742         // Set a dummy one.
1743         // We could use OperandSetter here, but that would implied an overhead
1744         // that we are not willing to pay.
1745         Inst->setOperand(It, UndefValue::get(Val->getType()));
1746       }
1747     }
1748 
1749     /// \brief Restore the original list of uses.
undo()1750     void undo() override {
1751       DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
1752       for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
1753         Inst->setOperand(It, OriginalValues[It]);
1754     }
1755   };
1756 
1757   /// \brief Build a truncate instruction.
1758   class TruncBuilder : public TypePromotionAction {
1759     Value *Val;
1760   public:
1761     /// \brief Build a truncate instruction of \p Opnd producing a \p Ty
1762     /// result.
1763     /// trunc Opnd to Ty.
TruncBuilder(Instruction * Opnd,Type * Ty)1764     TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
1765       IRBuilder<> Builder(Opnd);
1766       Val = Builder.CreateTrunc(Opnd, Ty, "promoted");
1767       DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n");
1768     }
1769 
1770     /// \brief Get the built value.
getBuiltValue()1771     Value *getBuiltValue() { return Val; }
1772 
1773     /// \brief Remove the built instruction.
undo()1774     void undo() override {
1775       DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n");
1776       if (Instruction *IVal = dyn_cast<Instruction>(Val))
1777         IVal->eraseFromParent();
1778     }
1779   };
1780 
1781   /// \brief Build a sign extension instruction.
1782   class SExtBuilder : public TypePromotionAction {
1783     Value *Val;
1784   public:
1785     /// \brief Build a sign extension instruction of \p Opnd producing a \p Ty
1786     /// result.
1787     /// sext Opnd to Ty.
SExtBuilder(Instruction * InsertPt,Value * Opnd,Type * Ty)1788     SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
1789         : TypePromotionAction(InsertPt) {
1790       IRBuilder<> Builder(InsertPt);
1791       Val = Builder.CreateSExt(Opnd, Ty, "promoted");
1792       DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n");
1793     }
1794 
1795     /// \brief Get the built value.
getBuiltValue()1796     Value *getBuiltValue() { return Val; }
1797 
1798     /// \brief Remove the built instruction.
undo()1799     void undo() override {
1800       DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n");
1801       if (Instruction *IVal = dyn_cast<Instruction>(Val))
1802         IVal->eraseFromParent();
1803     }
1804   };
1805 
1806   /// \brief Build a zero extension instruction.
1807   class ZExtBuilder : public TypePromotionAction {
1808     Value *Val;
1809   public:
1810     /// \brief Build a zero extension instruction of \p Opnd producing a \p Ty
1811     /// result.
1812     /// zext Opnd to Ty.
ZExtBuilder(Instruction * InsertPt,Value * Opnd,Type * Ty)1813     ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
1814         : TypePromotionAction(InsertPt) {
1815       IRBuilder<> Builder(InsertPt);
1816       Val = Builder.CreateZExt(Opnd, Ty, "promoted");
1817       DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n");
1818     }
1819 
1820     /// \brief Get the built value.
getBuiltValue()1821     Value *getBuiltValue() { return Val; }
1822 
1823     /// \brief Remove the built instruction.
undo()1824     void undo() override {
1825       DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n");
1826       if (Instruction *IVal = dyn_cast<Instruction>(Val))
1827         IVal->eraseFromParent();
1828     }
1829   };
1830 
1831   /// \brief Mutate an instruction to another type.
1832   class TypeMutator : public TypePromotionAction {
1833     /// Record the original type.
1834     Type *OrigTy;
1835 
1836   public:
1837     /// \brief Mutate the type of \p Inst into \p NewTy.
TypeMutator(Instruction * Inst,Type * NewTy)1838     TypeMutator(Instruction *Inst, Type *NewTy)
1839         : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
1840       DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
1841                    << "\n");
1842       Inst->mutateType(NewTy);
1843     }
1844 
1845     /// \brief Mutate the instruction back to its original type.
undo()1846     void undo() override {
1847       DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
1848                    << "\n");
1849       Inst->mutateType(OrigTy);
1850     }
1851   };
1852 
1853   /// \brief Replace the uses of an instruction by another instruction.
1854   class UsesReplacer : public TypePromotionAction {
1855     /// Helper structure to keep track of the replaced uses.
1856     struct InstructionAndIdx {
1857       /// The instruction using the instruction.
1858       Instruction *Inst;
1859       /// The index where this instruction is used for Inst.
1860       unsigned Idx;
InstructionAndIdx__anonebd7bfba0211::TypePromotionTransaction::UsesReplacer::InstructionAndIdx1861       InstructionAndIdx(Instruction *Inst, unsigned Idx)
1862           : Inst(Inst), Idx(Idx) {}
1863     };
1864 
1865     /// Keep track of the original uses (pair Instruction, Index).
1866     SmallVector<InstructionAndIdx, 4> OriginalUses;
1867     typedef SmallVectorImpl<InstructionAndIdx>::iterator use_iterator;
1868 
1869   public:
1870     /// \brief Replace all the use of \p Inst by \p New.
UsesReplacer(Instruction * Inst,Value * New)1871     UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
1872       DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
1873                    << "\n");
1874       // Record the original uses.
1875       for (Use &U : Inst->uses()) {
1876         Instruction *UserI = cast<Instruction>(U.getUser());
1877         OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
1878       }
1879       // Now, we can replace the uses.
1880       Inst->replaceAllUsesWith(New);
1881     }
1882 
1883     /// \brief Reassign the original uses of Inst to Inst.
undo()1884     void undo() override {
1885       DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
1886       for (use_iterator UseIt = OriginalUses.begin(),
1887                         EndIt = OriginalUses.end();
1888            UseIt != EndIt; ++UseIt) {
1889         UseIt->Inst->setOperand(UseIt->Idx, Inst);
1890       }
1891     }
1892   };
1893 
1894   /// \brief Remove an instruction from the IR.
1895   class InstructionRemover : public TypePromotionAction {
1896     /// Original position of the instruction.
1897     InsertionHandler Inserter;
1898     /// Helper structure to hide all the link to the instruction. In other
1899     /// words, this helps to do as if the instruction was removed.
1900     OperandsHider Hider;
1901     /// Keep track of the uses replaced, if any.
1902     UsesReplacer *Replacer;
1903 
1904   public:
1905     /// \brief Remove all reference of \p Inst and optinally replace all its
1906     /// uses with New.
1907     /// \pre If !Inst->use_empty(), then New != nullptr
InstructionRemover(Instruction * Inst,Value * New=nullptr)1908     InstructionRemover(Instruction *Inst, Value *New = nullptr)
1909         : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
1910           Replacer(nullptr) {
1911       if (New)
1912         Replacer = new UsesReplacer(Inst, New);
1913       DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
1914       Inst->removeFromParent();
1915     }
1916 
~InstructionRemover()1917     ~InstructionRemover() override { delete Replacer; }
1918 
1919     /// \brief Really remove the instruction.
commit()1920     void commit() override { delete Inst; }
1921 
1922     /// \brief Resurrect the instruction and reassign it to the proper uses if
1923     /// new value was provided when build this action.
undo()1924     void undo() override {
1925       DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
1926       Inserter.insert(Inst);
1927       if (Replacer)
1928         Replacer->undo();
1929       Hider.undo();
1930     }
1931   };
1932 
1933 public:
1934   /// Restoration point.
1935   /// The restoration point is a pointer to an action instead of an iterator
1936   /// because the iterator may be invalidated but not the pointer.
1937   typedef const TypePromotionAction *ConstRestorationPt;
1938   /// Advocate every changes made in that transaction.
1939   void commit();
1940   /// Undo all the changes made after the given point.
1941   void rollback(ConstRestorationPt Point);
1942   /// Get the current restoration point.
1943   ConstRestorationPt getRestorationPoint() const;
1944 
1945   /// \name API for IR modification with state keeping to support rollback.
1946   /// @{
1947   /// Same as Instruction::setOperand.
1948   void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
1949   /// Same as Instruction::eraseFromParent.
1950   void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
1951   /// Same as Value::replaceAllUsesWith.
1952   void replaceAllUsesWith(Instruction *Inst, Value *New);
1953   /// Same as Value::mutateType.
1954   void mutateType(Instruction *Inst, Type *NewTy);
1955   /// Same as IRBuilder::createTrunc.
1956   Value *createTrunc(Instruction *Opnd, Type *Ty);
1957   /// Same as IRBuilder::createSExt.
1958   Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
1959   /// Same as IRBuilder::createZExt.
1960   Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
1961   /// Same as Instruction::moveBefore.
1962   void moveBefore(Instruction *Inst, Instruction *Before);
1963   /// @}
1964 
1965 private:
1966   /// The ordered list of actions made so far.
1967   SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions;
1968   typedef SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator CommitPt;
1969 };
1970 
setOperand(Instruction * Inst,unsigned Idx,Value * NewVal)1971 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
1972                                           Value *NewVal) {
1973   Actions.push_back(
1974       make_unique<TypePromotionTransaction::OperandSetter>(Inst, Idx, NewVal));
1975 }
1976 
eraseInstruction(Instruction * Inst,Value * NewVal)1977 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
1978                                                 Value *NewVal) {
1979   Actions.push_back(
1980       make_unique<TypePromotionTransaction::InstructionRemover>(Inst, NewVal));
1981 }
1982 
replaceAllUsesWith(Instruction * Inst,Value * New)1983 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
1984                                                   Value *New) {
1985   Actions.push_back(make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
1986 }
1987 
mutateType(Instruction * Inst,Type * NewTy)1988 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
1989   Actions.push_back(make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
1990 }
1991 
createTrunc(Instruction * Opnd,Type * Ty)1992 Value *TypePromotionTransaction::createTrunc(Instruction *Opnd,
1993                                              Type *Ty) {
1994   std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
1995   Value *Val = Ptr->getBuiltValue();
1996   Actions.push_back(std::move(Ptr));
1997   return Val;
1998 }
1999 
createSExt(Instruction * Inst,Value * Opnd,Type * Ty)2000 Value *TypePromotionTransaction::createSExt(Instruction *Inst,
2001                                             Value *Opnd, Type *Ty) {
2002   std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
2003   Value *Val = Ptr->getBuiltValue();
2004   Actions.push_back(std::move(Ptr));
2005   return Val;
2006 }
2007 
createZExt(Instruction * Inst,Value * Opnd,Type * Ty)2008 Value *TypePromotionTransaction::createZExt(Instruction *Inst,
2009                                             Value *Opnd, Type *Ty) {
2010   std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
2011   Value *Val = Ptr->getBuiltValue();
2012   Actions.push_back(std::move(Ptr));
2013   return Val;
2014 }
2015 
moveBefore(Instruction * Inst,Instruction * Before)2016 void TypePromotionTransaction::moveBefore(Instruction *Inst,
2017                                           Instruction *Before) {
2018   Actions.push_back(
2019       make_unique<TypePromotionTransaction::InstructionMoveBefore>(Inst, Before));
2020 }
2021 
2022 TypePromotionTransaction::ConstRestorationPt
getRestorationPoint() const2023 TypePromotionTransaction::getRestorationPoint() const {
2024   return !Actions.empty() ? Actions.back().get() : nullptr;
2025 }
2026 
commit()2027 void TypePromotionTransaction::commit() {
2028   for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
2029        ++It)
2030     (*It)->commit();
2031   Actions.clear();
2032 }
2033 
rollback(TypePromotionTransaction::ConstRestorationPt Point)2034 void TypePromotionTransaction::rollback(
2035     TypePromotionTransaction::ConstRestorationPt Point) {
2036   while (!Actions.empty() && Point != Actions.back().get()) {
2037     std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
2038     Curr->undo();
2039   }
2040 }
2041 
2042 /// \brief A helper class for matching addressing modes.
2043 ///
2044 /// This encapsulates the logic for matching the target-legal addressing modes.
2045 class AddressingModeMatcher {
2046   SmallVectorImpl<Instruction*> &AddrModeInsts;
2047   const TargetMachine &TM;
2048   const TargetLowering &TLI;
2049 
2050   /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
2051   /// the memory instruction that we're computing this address for.
2052   Type *AccessTy;
2053   Instruction *MemoryInst;
2054 
2055   /// AddrMode - This is the addressing mode that we're building up.  This is
2056   /// part of the return value of this addressing mode matching stuff.
2057   ExtAddrMode &AddrMode;
2058 
2059   /// The truncate instruction inserted by other CodeGenPrepare optimizations.
2060   const SetOfInstrs &InsertedTruncs;
2061   /// A map from the instructions to their type before promotion.
2062   InstrToOrigTy &PromotedInsts;
2063   /// The ongoing transaction where every action should be registered.
2064   TypePromotionTransaction &TPT;
2065 
2066   /// IgnoreProfitability - This is set to true when we should not do
2067   /// profitability checks.  When true, IsProfitableToFoldIntoAddressingMode
2068   /// always returns true.
2069   bool IgnoreProfitability;
2070 
AddressingModeMatcher(SmallVectorImpl<Instruction * > & AMI,const TargetMachine & TM,Type * AT,Instruction * MI,ExtAddrMode & AM,const SetOfInstrs & InsertedTruncs,InstrToOrigTy & PromotedInsts,TypePromotionTransaction & TPT)2071   AddressingModeMatcher(SmallVectorImpl<Instruction *> &AMI,
2072                         const TargetMachine &TM, Type *AT, Instruction *MI,
2073                         ExtAddrMode &AM, const SetOfInstrs &InsertedTruncs,
2074                         InstrToOrigTy &PromotedInsts,
2075                         TypePromotionTransaction &TPT)
2076       : AddrModeInsts(AMI), TM(TM),
2077         TLI(*TM.getSubtargetImpl(*MI->getParent()->getParent())
2078                  ->getTargetLowering()),
2079         AccessTy(AT), MemoryInst(MI), AddrMode(AM),
2080         InsertedTruncs(InsertedTruncs), PromotedInsts(PromotedInsts), TPT(TPT) {
2081     IgnoreProfitability = false;
2082   }
2083 public:
2084 
2085   /// Match - Find the maximal addressing mode that a load/store of V can fold,
2086   /// give an access type of AccessTy.  This returns a list of involved
2087   /// instructions in AddrModeInsts.
2088   /// \p InsertedTruncs The truncate instruction inserted by other
2089   /// CodeGenPrepare
2090   /// optimizations.
2091   /// \p PromotedInsts maps the instructions to their type before promotion.
2092   /// \p The ongoing transaction where every action should be registered.
Match(Value * V,Type * AccessTy,Instruction * MemoryInst,SmallVectorImpl<Instruction * > & AddrModeInsts,const TargetMachine & TM,const SetOfInstrs & InsertedTruncs,InstrToOrigTy & PromotedInsts,TypePromotionTransaction & TPT)2093   static ExtAddrMode Match(Value *V, Type *AccessTy,
2094                            Instruction *MemoryInst,
2095                            SmallVectorImpl<Instruction*> &AddrModeInsts,
2096                            const TargetMachine &TM,
2097                            const SetOfInstrs &InsertedTruncs,
2098                            InstrToOrigTy &PromotedInsts,
2099                            TypePromotionTransaction &TPT) {
2100     ExtAddrMode Result;
2101 
2102     bool Success = AddressingModeMatcher(AddrModeInsts, TM, AccessTy,
2103                                          MemoryInst, Result, InsertedTruncs,
2104                                          PromotedInsts, TPT).MatchAddr(V, 0);
2105     (void)Success; assert(Success && "Couldn't select *anything*?");
2106     return Result;
2107   }
2108 private:
2109   bool MatchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
2110   bool MatchAddr(Value *V, unsigned Depth);
2111   bool MatchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth,
2112                           bool *MovedAway = nullptr);
2113   bool IsProfitableToFoldIntoAddressingMode(Instruction *I,
2114                                             ExtAddrMode &AMBefore,
2115                                             ExtAddrMode &AMAfter);
2116   bool ValueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
2117   bool IsPromotionProfitable(unsigned NewCost, unsigned OldCost,
2118                              Value *PromotedOperand) const;
2119 };
2120 
2121 /// MatchScaledValue - Try adding ScaleReg*Scale to the current addressing mode.
2122 /// Return true and update AddrMode if this addr mode is legal for the target,
2123 /// false if not.
MatchScaledValue(Value * ScaleReg,int64_t Scale,unsigned Depth)2124 bool AddressingModeMatcher::MatchScaledValue(Value *ScaleReg, int64_t Scale,
2125                                              unsigned Depth) {
2126   // If Scale is 1, then this is the same as adding ScaleReg to the addressing
2127   // mode.  Just process that directly.
2128   if (Scale == 1)
2129     return MatchAddr(ScaleReg, Depth);
2130 
2131   // If the scale is 0, it takes nothing to add this.
2132   if (Scale == 0)
2133     return true;
2134 
2135   // If we already have a scale of this value, we can add to it, otherwise, we
2136   // need an available scale field.
2137   if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
2138     return false;
2139 
2140   ExtAddrMode TestAddrMode = AddrMode;
2141 
2142   // Add scale to turn X*4+X*3 -> X*7.  This could also do things like
2143   // [A+B + A*7] -> [B+A*8].
2144   TestAddrMode.Scale += Scale;
2145   TestAddrMode.ScaledReg = ScaleReg;
2146 
2147   // If the new address isn't legal, bail out.
2148   if (!TLI.isLegalAddressingMode(TestAddrMode, AccessTy))
2149     return false;
2150 
2151   // It was legal, so commit it.
2152   AddrMode = TestAddrMode;
2153 
2154   // Okay, we decided that we can add ScaleReg+Scale to AddrMode.  Check now
2155   // to see if ScaleReg is actually X+C.  If so, we can turn this into adding
2156   // X*Scale + C*Scale to addr mode.
2157   ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
2158   if (isa<Instruction>(ScaleReg) &&  // not a constant expr.
2159       match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
2160     TestAddrMode.ScaledReg = AddLHS;
2161     TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
2162 
2163     // If this addressing mode is legal, commit it and remember that we folded
2164     // this instruction.
2165     if (TLI.isLegalAddressingMode(TestAddrMode, AccessTy)) {
2166       AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
2167       AddrMode = TestAddrMode;
2168       return true;
2169     }
2170   }
2171 
2172   // Otherwise, not (x+c)*scale, just return what we have.
2173   return true;
2174 }
2175 
2176 /// MightBeFoldableInst - This is a little filter, which returns true if an
2177 /// addressing computation involving I might be folded into a load/store
2178 /// accessing it.  This doesn't need to be perfect, but needs to accept at least
2179 /// the set of instructions that MatchOperationAddr can.
MightBeFoldableInst(Instruction * I)2180 static bool MightBeFoldableInst(Instruction *I) {
2181   switch (I->getOpcode()) {
2182   case Instruction::BitCast:
2183   case Instruction::AddrSpaceCast:
2184     // Don't touch identity bitcasts.
2185     if (I->getType() == I->getOperand(0)->getType())
2186       return false;
2187     return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
2188   case Instruction::PtrToInt:
2189     // PtrToInt is always a noop, as we know that the int type is pointer sized.
2190     return true;
2191   case Instruction::IntToPtr:
2192     // We know the input is intptr_t, so this is foldable.
2193     return true;
2194   case Instruction::Add:
2195     return true;
2196   case Instruction::Mul:
2197   case Instruction::Shl:
2198     // Can only handle X*C and X << C.
2199     return isa<ConstantInt>(I->getOperand(1));
2200   case Instruction::GetElementPtr:
2201     return true;
2202   default:
2203     return false;
2204   }
2205 }
2206 
2207 /// \brief Check whether or not \p Val is a legal instruction for \p TLI.
2208 /// \note \p Val is assumed to be the product of some type promotion.
2209 /// Therefore if \p Val has an undefined state in \p TLI, this is assumed
2210 /// to be legal, as the non-promoted value would have had the same state.
isPromotedInstructionLegal(const TargetLowering & TLI,Value * Val)2211 static bool isPromotedInstructionLegal(const TargetLowering &TLI, Value *Val) {
2212   Instruction *PromotedInst = dyn_cast<Instruction>(Val);
2213   if (!PromotedInst)
2214     return false;
2215   int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
2216   // If the ISDOpcode is undefined, it was undefined before the promotion.
2217   if (!ISDOpcode)
2218     return true;
2219   // Otherwise, check if the promoted instruction is legal or not.
2220   return TLI.isOperationLegalOrCustom(
2221       ISDOpcode, TLI.getValueType(PromotedInst->getType()));
2222 }
2223 
2224 /// \brief Hepler class to perform type promotion.
2225 class TypePromotionHelper {
2226   /// \brief Utility function to check whether or not a sign or zero extension
2227   /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
2228   /// either using the operands of \p Inst or promoting \p Inst.
2229   /// The type of the extension is defined by \p IsSExt.
2230   /// In other words, check if:
2231   /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
2232   /// #1 Promotion applies:
2233   /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
2234   /// #2 Operand reuses:
2235   /// ext opnd1 to ConsideredExtType.
2236   /// \p PromotedInsts maps the instructions to their type before promotion.
2237   static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
2238                             const InstrToOrigTy &PromotedInsts, bool IsSExt);
2239 
2240   /// \brief Utility function to determine if \p OpIdx should be promoted when
2241   /// promoting \p Inst.
shouldExtOperand(const Instruction * Inst,int OpIdx)2242   static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
2243     if (isa<SelectInst>(Inst) && OpIdx == 0)
2244       return false;
2245     return true;
2246   }
2247 
2248   /// \brief Utility function to promote the operand of \p Ext when this
2249   /// operand is a promotable trunc or sext or zext.
2250   /// \p PromotedInsts maps the instructions to their type before promotion.
2251   /// \p CreatedInstsCost[out] contains the cost of all instructions
2252   /// created to promote the operand of Ext.
2253   /// Newly added extensions are inserted in \p Exts.
2254   /// Newly added truncates are inserted in \p Truncs.
2255   /// Should never be called directly.
2256   /// \return The promoted value which is used instead of Ext.
2257   static Value *promoteOperandForTruncAndAnyExt(
2258       Instruction *Ext, TypePromotionTransaction &TPT,
2259       InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2260       SmallVectorImpl<Instruction *> *Exts,
2261       SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI);
2262 
2263   /// \brief Utility function to promote the operand of \p Ext when this
2264   /// operand is promotable and is not a supported trunc or sext.
2265   /// \p PromotedInsts maps the instructions to their type before promotion.
2266   /// \p CreatedInstsCost[out] contains the cost of all the instructions
2267   /// created to promote the operand of Ext.
2268   /// Newly added extensions are inserted in \p Exts.
2269   /// Newly added truncates are inserted in \p Truncs.
2270   /// Should never be called directly.
2271   /// \return The promoted value which is used instead of Ext.
2272   static Value *promoteOperandForOther(Instruction *Ext,
2273                                        TypePromotionTransaction &TPT,
2274                                        InstrToOrigTy &PromotedInsts,
2275                                        unsigned &CreatedInstsCost,
2276                                        SmallVectorImpl<Instruction *> *Exts,
2277                                        SmallVectorImpl<Instruction *> *Truncs,
2278                                        const TargetLowering &TLI, bool IsSExt);
2279 
2280   /// \see promoteOperandForOther.
signExtendOperandForOther(Instruction * Ext,TypePromotionTransaction & TPT,InstrToOrigTy & PromotedInsts,unsigned & CreatedInstsCost,SmallVectorImpl<Instruction * > * Exts,SmallVectorImpl<Instruction * > * Truncs,const TargetLowering & TLI)2281   static Value *signExtendOperandForOther(
2282       Instruction *Ext, TypePromotionTransaction &TPT,
2283       InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2284       SmallVectorImpl<Instruction *> *Exts,
2285       SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2286     return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
2287                                   Exts, Truncs, TLI, true);
2288   }
2289 
2290   /// \see promoteOperandForOther.
zeroExtendOperandForOther(Instruction * Ext,TypePromotionTransaction & TPT,InstrToOrigTy & PromotedInsts,unsigned & CreatedInstsCost,SmallVectorImpl<Instruction * > * Exts,SmallVectorImpl<Instruction * > * Truncs,const TargetLowering & TLI)2291   static Value *zeroExtendOperandForOther(
2292       Instruction *Ext, TypePromotionTransaction &TPT,
2293       InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2294       SmallVectorImpl<Instruction *> *Exts,
2295       SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2296     return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
2297                                   Exts, Truncs, TLI, false);
2298   }
2299 
2300 public:
2301   /// Type for the utility function that promotes the operand of Ext.
2302   typedef Value *(*Action)(Instruction *Ext, TypePromotionTransaction &TPT,
2303                            InstrToOrigTy &PromotedInsts,
2304                            unsigned &CreatedInstsCost,
2305                            SmallVectorImpl<Instruction *> *Exts,
2306                            SmallVectorImpl<Instruction *> *Truncs,
2307                            const TargetLowering &TLI);
2308   /// \brief Given a sign/zero extend instruction \p Ext, return the approriate
2309   /// action to promote the operand of \p Ext instead of using Ext.
2310   /// \return NULL if no promotable action is possible with the current
2311   /// sign extension.
2312   /// \p InsertedTruncs keeps track of all the truncate instructions inserted by
2313   /// the others CodeGenPrepare optimizations. This information is important
2314   /// because we do not want to promote these instructions as CodeGenPrepare
2315   /// will reinsert them later. Thus creating an infinite loop: create/remove.
2316   /// \p PromotedInsts maps the instructions to their type before promotion.
2317   static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedTruncs,
2318                           const TargetLowering &TLI,
2319                           const InstrToOrigTy &PromotedInsts);
2320 };
2321 
canGetThrough(const Instruction * Inst,Type * ConsideredExtType,const InstrToOrigTy & PromotedInsts,bool IsSExt)2322 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
2323                                         Type *ConsideredExtType,
2324                                         const InstrToOrigTy &PromotedInsts,
2325                                         bool IsSExt) {
2326   // The promotion helper does not know how to deal with vector types yet.
2327   // To be able to fix that, we would need to fix the places where we
2328   // statically extend, e.g., constants and such.
2329   if (Inst->getType()->isVectorTy())
2330     return false;
2331 
2332   // We can always get through zext.
2333   if (isa<ZExtInst>(Inst))
2334     return true;
2335 
2336   // sext(sext) is ok too.
2337   if (IsSExt && isa<SExtInst>(Inst))
2338     return true;
2339 
2340   // We can get through binary operator, if it is legal. In other words, the
2341   // binary operator must have a nuw or nsw flag.
2342   const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
2343   if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
2344       ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
2345        (IsSExt && BinOp->hasNoSignedWrap())))
2346     return true;
2347 
2348   // Check if we can do the following simplification.
2349   // ext(trunc(opnd)) --> ext(opnd)
2350   if (!isa<TruncInst>(Inst))
2351     return false;
2352 
2353   Value *OpndVal = Inst->getOperand(0);
2354   // Check if we can use this operand in the extension.
2355   // If the type is larger than the result type of the extension,
2356   // we cannot.
2357   if (!OpndVal->getType()->isIntegerTy() ||
2358       OpndVal->getType()->getIntegerBitWidth() >
2359           ConsideredExtType->getIntegerBitWidth())
2360     return false;
2361 
2362   // If the operand of the truncate is not an instruction, we will not have
2363   // any information on the dropped bits.
2364   // (Actually we could for constant but it is not worth the extra logic).
2365   Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
2366   if (!Opnd)
2367     return false;
2368 
2369   // Check if the source of the type is narrow enough.
2370   // I.e., check that trunc just drops extended bits of the same kind of
2371   // the extension.
2372   // #1 get the type of the operand and check the kind of the extended bits.
2373   const Type *OpndType;
2374   InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
2375   if (It != PromotedInsts.end() && It->second.IsSExt == IsSExt)
2376     OpndType = It->second.Ty;
2377   else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd)))
2378     OpndType = Opnd->getOperand(0)->getType();
2379   else
2380     return false;
2381 
2382   // #2 check that the truncate just drop extended bits.
2383   if (Inst->getType()->getIntegerBitWidth() >= OpndType->getIntegerBitWidth())
2384     return true;
2385 
2386   return false;
2387 }
2388 
getAction(Instruction * Ext,const SetOfInstrs & InsertedTruncs,const TargetLowering & TLI,const InstrToOrigTy & PromotedInsts)2389 TypePromotionHelper::Action TypePromotionHelper::getAction(
2390     Instruction *Ext, const SetOfInstrs &InsertedTruncs,
2391     const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
2392   assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
2393          "Unexpected instruction type");
2394   Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0));
2395   Type *ExtTy = Ext->getType();
2396   bool IsSExt = isa<SExtInst>(Ext);
2397   // If the operand of the extension is not an instruction, we cannot
2398   // get through.
2399   // If it, check we can get through.
2400   if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt))
2401     return nullptr;
2402 
2403   // Do not promote if the operand has been added by codegenprepare.
2404   // Otherwise, it means we are undoing an optimization that is likely to be
2405   // redone, thus causing potential infinite loop.
2406   if (isa<TruncInst>(ExtOpnd) && InsertedTruncs.count(ExtOpnd))
2407     return nullptr;
2408 
2409   // SExt or Trunc instructions.
2410   // Return the related handler.
2411   if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) ||
2412       isa<ZExtInst>(ExtOpnd))
2413     return promoteOperandForTruncAndAnyExt;
2414 
2415   // Regular instruction.
2416   // Abort early if we will have to insert non-free instructions.
2417   if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType()))
2418     return nullptr;
2419   return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
2420 }
2421 
promoteOperandForTruncAndAnyExt(llvm::Instruction * SExt,TypePromotionTransaction & TPT,InstrToOrigTy & PromotedInsts,unsigned & CreatedInstsCost,SmallVectorImpl<Instruction * > * Exts,SmallVectorImpl<Instruction * > * Truncs,const TargetLowering & TLI)2422 Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
2423     llvm::Instruction *SExt, TypePromotionTransaction &TPT,
2424     InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2425     SmallVectorImpl<Instruction *> *Exts,
2426     SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
2427   // By construction, the operand of SExt is an instruction. Otherwise we cannot
2428   // get through it and this method should not be called.
2429   Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
2430   Value *ExtVal = SExt;
2431   bool HasMergedNonFreeExt = false;
2432   if (isa<ZExtInst>(SExtOpnd)) {
2433     // Replace s|zext(zext(opnd))
2434     // => zext(opnd).
2435     HasMergedNonFreeExt = !TLI.isExtFree(SExtOpnd);
2436     Value *ZExt =
2437         TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
2438     TPT.replaceAllUsesWith(SExt, ZExt);
2439     TPT.eraseInstruction(SExt);
2440     ExtVal = ZExt;
2441   } else {
2442     // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
2443     // => z|sext(opnd).
2444     TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
2445   }
2446   CreatedInstsCost = 0;
2447 
2448   // Remove dead code.
2449   if (SExtOpnd->use_empty())
2450     TPT.eraseInstruction(SExtOpnd);
2451 
2452   // Check if the extension is still needed.
2453   Instruction *ExtInst = dyn_cast<Instruction>(ExtVal);
2454   if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) {
2455     if (ExtInst) {
2456       if (Exts)
2457         Exts->push_back(ExtInst);
2458       CreatedInstsCost = !TLI.isExtFree(ExtInst) && !HasMergedNonFreeExt;
2459     }
2460     return ExtVal;
2461   }
2462 
2463   // At this point we have: ext ty opnd to ty.
2464   // Reassign the uses of ExtInst to the opnd and remove ExtInst.
2465   Value *NextVal = ExtInst->getOperand(0);
2466   TPT.eraseInstruction(ExtInst, NextVal);
2467   return NextVal;
2468 }
2469 
promoteOperandForOther(Instruction * Ext,TypePromotionTransaction & TPT,InstrToOrigTy & PromotedInsts,unsigned & CreatedInstsCost,SmallVectorImpl<Instruction * > * Exts,SmallVectorImpl<Instruction * > * Truncs,const TargetLowering & TLI,bool IsSExt)2470 Value *TypePromotionHelper::promoteOperandForOther(
2471     Instruction *Ext, TypePromotionTransaction &TPT,
2472     InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
2473     SmallVectorImpl<Instruction *> *Exts,
2474     SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI,
2475     bool IsSExt) {
2476   // By construction, the operand of Ext is an instruction. Otherwise we cannot
2477   // get through it and this method should not be called.
2478   Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0));
2479   CreatedInstsCost = 0;
2480   if (!ExtOpnd->hasOneUse()) {
2481     // ExtOpnd will be promoted.
2482     // All its uses, but Ext, will need to use a truncated value of the
2483     // promoted version.
2484     // Create the truncate now.
2485     Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType());
2486     if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) {
2487       ITrunc->removeFromParent();
2488       // Insert it just after the definition.
2489       ITrunc->insertAfter(ExtOpnd);
2490       if (Truncs)
2491         Truncs->push_back(ITrunc);
2492     }
2493 
2494     TPT.replaceAllUsesWith(ExtOpnd, Trunc);
2495     // Restore the operand of Ext (which has been replace by the previous call
2496     // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
2497     TPT.setOperand(Ext, 0, ExtOpnd);
2498   }
2499 
2500   // Get through the Instruction:
2501   // 1. Update its type.
2502   // 2. Replace the uses of Ext by Inst.
2503   // 3. Extend each operand that needs to be extended.
2504 
2505   // Remember the original type of the instruction before promotion.
2506   // This is useful to know that the high bits are sign extended bits.
2507   PromotedInsts.insert(std::pair<Instruction *, TypeIsSExt>(
2508       ExtOpnd, TypeIsSExt(ExtOpnd->getType(), IsSExt)));
2509   // Step #1.
2510   TPT.mutateType(ExtOpnd, Ext->getType());
2511   // Step #2.
2512   TPT.replaceAllUsesWith(Ext, ExtOpnd);
2513   // Step #3.
2514   Instruction *ExtForOpnd = Ext;
2515 
2516   DEBUG(dbgs() << "Propagate Ext to operands\n");
2517   for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
2518        ++OpIdx) {
2519     DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n');
2520     if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() ||
2521         !shouldExtOperand(ExtOpnd, OpIdx)) {
2522       DEBUG(dbgs() << "No need to propagate\n");
2523       continue;
2524     }
2525     // Check if we can statically extend the operand.
2526     Value *Opnd = ExtOpnd->getOperand(OpIdx);
2527     if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
2528       DEBUG(dbgs() << "Statically extend\n");
2529       unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
2530       APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth)
2531                             : Cst->getValue().zext(BitWidth);
2532       TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal));
2533       continue;
2534     }
2535     // UndefValue are typed, so we have to statically sign extend them.
2536     if (isa<UndefValue>(Opnd)) {
2537       DEBUG(dbgs() << "Statically extend\n");
2538       TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType()));
2539       continue;
2540     }
2541 
2542     // Otherwise we have to explicity sign extend the operand.
2543     // Check if Ext was reused to extend an operand.
2544     if (!ExtForOpnd) {
2545       // If yes, create a new one.
2546       DEBUG(dbgs() << "More operands to ext\n");
2547       Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType())
2548         : TPT.createZExt(Ext, Opnd, Ext->getType());
2549       if (!isa<Instruction>(ValForExtOpnd)) {
2550         TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd);
2551         continue;
2552       }
2553       ExtForOpnd = cast<Instruction>(ValForExtOpnd);
2554     }
2555     if (Exts)
2556       Exts->push_back(ExtForOpnd);
2557     TPT.setOperand(ExtForOpnd, 0, Opnd);
2558 
2559     // Move the sign extension before the insertion point.
2560     TPT.moveBefore(ExtForOpnd, ExtOpnd);
2561     TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd);
2562     CreatedInstsCost += !TLI.isExtFree(ExtForOpnd);
2563     // If more sext are required, new instructions will have to be created.
2564     ExtForOpnd = nullptr;
2565   }
2566   if (ExtForOpnd == Ext) {
2567     DEBUG(dbgs() << "Extension is useless now\n");
2568     TPT.eraseInstruction(Ext);
2569   }
2570   return ExtOpnd;
2571 }
2572 
2573 /// IsPromotionProfitable - Check whether or not promoting an instruction
2574 /// to a wider type was profitable.
2575 /// \p NewCost gives the cost of extension instructions created by the
2576 /// promotion.
2577 /// \p OldCost gives the cost of extension instructions before the promotion
2578 /// plus the number of instructions that have been
2579 /// matched in the addressing mode the promotion.
2580 /// \p PromotedOperand is the value that has been promoted.
2581 /// \return True if the promotion is profitable, false otherwise.
IsPromotionProfitable(unsigned NewCost,unsigned OldCost,Value * PromotedOperand) const2582 bool AddressingModeMatcher::IsPromotionProfitable(
2583     unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const {
2584   DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCost << '\n');
2585   // The cost of the new extensions is greater than the cost of the
2586   // old extension plus what we folded.
2587   // This is not profitable.
2588   if (NewCost > OldCost)
2589     return false;
2590   if (NewCost < OldCost)
2591     return true;
2592   // The promotion is neutral but it may help folding the sign extension in
2593   // loads for instance.
2594   // Check that we did not create an illegal instruction.
2595   return isPromotedInstructionLegal(TLI, PromotedOperand);
2596 }
2597 
2598 /// MatchOperationAddr - Given an instruction or constant expr, see if we can
2599 /// fold the operation into the addressing mode.  If so, update the addressing
2600 /// mode and return true, otherwise return false without modifying AddrMode.
2601 /// If \p MovedAway is not NULL, it contains the information of whether or
2602 /// not AddrInst has to be folded into the addressing mode on success.
2603 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
2604 /// because it has been moved away.
2605 /// Thus AddrInst must not be added in the matched instructions.
2606 /// This state can happen when AddrInst is a sext, since it may be moved away.
2607 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
2608 /// not be referenced anymore.
MatchOperationAddr(User * AddrInst,unsigned Opcode,unsigned Depth,bool * MovedAway)2609 bool AddressingModeMatcher::MatchOperationAddr(User *AddrInst, unsigned Opcode,
2610                                                unsigned Depth,
2611                                                bool *MovedAway) {
2612   // Avoid exponential behavior on extremely deep expression trees.
2613   if (Depth >= 5) return false;
2614 
2615   // By default, all matched instructions stay in place.
2616   if (MovedAway)
2617     *MovedAway = false;
2618 
2619   switch (Opcode) {
2620   case Instruction::PtrToInt:
2621     // PtrToInt is always a noop, as we know that the int type is pointer sized.
2622     return MatchAddr(AddrInst->getOperand(0), Depth);
2623   case Instruction::IntToPtr:
2624     // This inttoptr is a no-op if the integer type is pointer sized.
2625     if (TLI.getValueType(AddrInst->getOperand(0)->getType()) ==
2626         TLI.getPointerTy(AddrInst->getType()->getPointerAddressSpace()))
2627       return MatchAddr(AddrInst->getOperand(0), Depth);
2628     return false;
2629   case Instruction::BitCast:
2630   case Instruction::AddrSpaceCast:
2631     // BitCast is always a noop, and we can handle it as long as it is
2632     // int->int or pointer->pointer (we don't want int<->fp or something).
2633     if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
2634          AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
2635         // Don't touch identity bitcasts.  These were probably put here by LSR,
2636         // and we don't want to mess around with them.  Assume it knows what it
2637         // is doing.
2638         AddrInst->getOperand(0)->getType() != AddrInst->getType())
2639       return MatchAddr(AddrInst->getOperand(0), Depth);
2640     return false;
2641   case Instruction::Add: {
2642     // Check to see if we can merge in the RHS then the LHS.  If so, we win.
2643     ExtAddrMode BackupAddrMode = AddrMode;
2644     unsigned OldSize = AddrModeInsts.size();
2645     // Start a transaction at this point.
2646     // The LHS may match but not the RHS.
2647     // Therefore, we need a higher level restoration point to undo partially
2648     // matched operation.
2649     TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2650         TPT.getRestorationPoint();
2651 
2652     if (MatchAddr(AddrInst->getOperand(1), Depth+1) &&
2653         MatchAddr(AddrInst->getOperand(0), Depth+1))
2654       return true;
2655 
2656     // Restore the old addr mode info.
2657     AddrMode = BackupAddrMode;
2658     AddrModeInsts.resize(OldSize);
2659     TPT.rollback(LastKnownGood);
2660 
2661     // Otherwise this was over-aggressive.  Try merging in the LHS then the RHS.
2662     if (MatchAddr(AddrInst->getOperand(0), Depth+1) &&
2663         MatchAddr(AddrInst->getOperand(1), Depth+1))
2664       return true;
2665 
2666     // Otherwise we definitely can't merge the ADD in.
2667     AddrMode = BackupAddrMode;
2668     AddrModeInsts.resize(OldSize);
2669     TPT.rollback(LastKnownGood);
2670     break;
2671   }
2672   //case Instruction::Or:
2673   // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
2674   //break;
2675   case Instruction::Mul:
2676   case Instruction::Shl: {
2677     // Can only handle X*C and X << C.
2678     ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
2679     if (!RHS)
2680       return false;
2681     int64_t Scale = RHS->getSExtValue();
2682     if (Opcode == Instruction::Shl)
2683       Scale = 1LL << Scale;
2684 
2685     return MatchScaledValue(AddrInst->getOperand(0), Scale, Depth);
2686   }
2687   case Instruction::GetElementPtr: {
2688     // Scan the GEP.  We check it if it contains constant offsets and at most
2689     // one variable offset.
2690     int VariableOperand = -1;
2691     unsigned VariableScale = 0;
2692 
2693     int64_t ConstantOffset = 0;
2694     const DataLayout *TD = TLI.getDataLayout();
2695     gep_type_iterator GTI = gep_type_begin(AddrInst);
2696     for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
2697       if (StructType *STy = dyn_cast<StructType>(*GTI)) {
2698         const StructLayout *SL = TD->getStructLayout(STy);
2699         unsigned Idx =
2700           cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
2701         ConstantOffset += SL->getElementOffset(Idx);
2702       } else {
2703         uint64_t TypeSize = TD->getTypeAllocSize(GTI.getIndexedType());
2704         if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
2705           ConstantOffset += CI->getSExtValue()*TypeSize;
2706         } else if (TypeSize) {  // Scales of zero don't do anything.
2707           // We only allow one variable index at the moment.
2708           if (VariableOperand != -1)
2709             return false;
2710 
2711           // Remember the variable index.
2712           VariableOperand = i;
2713           VariableScale = TypeSize;
2714         }
2715       }
2716     }
2717 
2718     // A common case is for the GEP to only do a constant offset.  In this case,
2719     // just add it to the disp field and check validity.
2720     if (VariableOperand == -1) {
2721       AddrMode.BaseOffs += ConstantOffset;
2722       if (ConstantOffset == 0 || TLI.isLegalAddressingMode(AddrMode, AccessTy)){
2723         // Check to see if we can fold the base pointer in too.
2724         if (MatchAddr(AddrInst->getOperand(0), Depth+1))
2725           return true;
2726       }
2727       AddrMode.BaseOffs -= ConstantOffset;
2728       return false;
2729     }
2730 
2731     // Save the valid addressing mode in case we can't match.
2732     ExtAddrMode BackupAddrMode = AddrMode;
2733     unsigned OldSize = AddrModeInsts.size();
2734 
2735     // See if the scale and offset amount is valid for this target.
2736     AddrMode.BaseOffs += ConstantOffset;
2737 
2738     // Match the base operand of the GEP.
2739     if (!MatchAddr(AddrInst->getOperand(0), Depth+1)) {
2740       // If it couldn't be matched, just stuff the value in a register.
2741       if (AddrMode.HasBaseReg) {
2742         AddrMode = BackupAddrMode;
2743         AddrModeInsts.resize(OldSize);
2744         return false;
2745       }
2746       AddrMode.HasBaseReg = true;
2747       AddrMode.BaseReg = AddrInst->getOperand(0);
2748     }
2749 
2750     // Match the remaining variable portion of the GEP.
2751     if (!MatchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
2752                           Depth)) {
2753       // If it couldn't be matched, try stuffing the base into a register
2754       // instead of matching it, and retrying the match of the scale.
2755       AddrMode = BackupAddrMode;
2756       AddrModeInsts.resize(OldSize);
2757       if (AddrMode.HasBaseReg)
2758         return false;
2759       AddrMode.HasBaseReg = true;
2760       AddrMode.BaseReg = AddrInst->getOperand(0);
2761       AddrMode.BaseOffs += ConstantOffset;
2762       if (!MatchScaledValue(AddrInst->getOperand(VariableOperand),
2763                             VariableScale, Depth)) {
2764         // If even that didn't work, bail.
2765         AddrMode = BackupAddrMode;
2766         AddrModeInsts.resize(OldSize);
2767         return false;
2768       }
2769     }
2770 
2771     return true;
2772   }
2773   case Instruction::SExt:
2774   case Instruction::ZExt: {
2775     Instruction *Ext = dyn_cast<Instruction>(AddrInst);
2776     if (!Ext)
2777       return false;
2778 
2779     // Try to move this ext out of the way of the addressing mode.
2780     // Ask for a method for doing so.
2781     TypePromotionHelper::Action TPH =
2782         TypePromotionHelper::getAction(Ext, InsertedTruncs, TLI, PromotedInsts);
2783     if (!TPH)
2784       return false;
2785 
2786     TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2787         TPT.getRestorationPoint();
2788     unsigned CreatedInstsCost = 0;
2789     unsigned ExtCost = !TLI.isExtFree(Ext);
2790     Value *PromotedOperand =
2791         TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI);
2792     // SExt has been moved away.
2793     // Thus either it will be rematched later in the recursive calls or it is
2794     // gone. Anyway, we must not fold it into the addressing mode at this point.
2795     // E.g.,
2796     // op = add opnd, 1
2797     // idx = ext op
2798     // addr = gep base, idx
2799     // is now:
2800     // promotedOpnd = ext opnd            <- no match here
2801     // op = promoted_add promotedOpnd, 1  <- match (later in recursive calls)
2802     // addr = gep base, op                <- match
2803     if (MovedAway)
2804       *MovedAway = true;
2805 
2806     assert(PromotedOperand &&
2807            "TypePromotionHelper should have filtered out those cases");
2808 
2809     ExtAddrMode BackupAddrMode = AddrMode;
2810     unsigned OldSize = AddrModeInsts.size();
2811 
2812     if (!MatchAddr(PromotedOperand, Depth) ||
2813         // The total of the new cost is equals to the cost of the created
2814         // instructions.
2815         // The total of the old cost is equals to the cost of the extension plus
2816         // what we have saved in the addressing mode.
2817         !IsPromotionProfitable(CreatedInstsCost,
2818                                ExtCost + (AddrModeInsts.size() - OldSize),
2819                                PromotedOperand)) {
2820       AddrMode = BackupAddrMode;
2821       AddrModeInsts.resize(OldSize);
2822       DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
2823       TPT.rollback(LastKnownGood);
2824       return false;
2825     }
2826     return true;
2827   }
2828   }
2829   return false;
2830 }
2831 
2832 /// MatchAddr - If we can, try to add the value of 'Addr' into the current
2833 /// addressing mode.  If Addr can't be added to AddrMode this returns false and
2834 /// leaves AddrMode unmodified.  This assumes that Addr is either a pointer type
2835 /// or intptr_t for the target.
2836 ///
MatchAddr(Value * Addr,unsigned Depth)2837 bool AddressingModeMatcher::MatchAddr(Value *Addr, unsigned Depth) {
2838   // Start a transaction at this point that we will rollback if the matching
2839   // fails.
2840   TypePromotionTransaction::ConstRestorationPt LastKnownGood =
2841       TPT.getRestorationPoint();
2842   if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
2843     // Fold in immediates if legal for the target.
2844     AddrMode.BaseOffs += CI->getSExtValue();
2845     if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2846       return true;
2847     AddrMode.BaseOffs -= CI->getSExtValue();
2848   } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
2849     // If this is a global variable, try to fold it into the addressing mode.
2850     if (!AddrMode.BaseGV) {
2851       AddrMode.BaseGV = GV;
2852       if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2853         return true;
2854       AddrMode.BaseGV = nullptr;
2855     }
2856   } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
2857     ExtAddrMode BackupAddrMode = AddrMode;
2858     unsigned OldSize = AddrModeInsts.size();
2859 
2860     // Check to see if it is possible to fold this operation.
2861     bool MovedAway = false;
2862     if (MatchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
2863       // This instruction may have been move away. If so, there is nothing
2864       // to check here.
2865       if (MovedAway)
2866         return true;
2867       // Okay, it's possible to fold this.  Check to see if it is actually
2868       // *profitable* to do so.  We use a simple cost model to avoid increasing
2869       // register pressure too much.
2870       if (I->hasOneUse() ||
2871           IsProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
2872         AddrModeInsts.push_back(I);
2873         return true;
2874       }
2875 
2876       // It isn't profitable to do this, roll back.
2877       //cerr << "NOT FOLDING: " << *I;
2878       AddrMode = BackupAddrMode;
2879       AddrModeInsts.resize(OldSize);
2880       TPT.rollback(LastKnownGood);
2881     }
2882   } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
2883     if (MatchOperationAddr(CE, CE->getOpcode(), Depth))
2884       return true;
2885     TPT.rollback(LastKnownGood);
2886   } else if (isa<ConstantPointerNull>(Addr)) {
2887     // Null pointer gets folded without affecting the addressing mode.
2888     return true;
2889   }
2890 
2891   // Worse case, the target should support [reg] addressing modes. :)
2892   if (!AddrMode.HasBaseReg) {
2893     AddrMode.HasBaseReg = true;
2894     AddrMode.BaseReg = Addr;
2895     // Still check for legality in case the target supports [imm] but not [i+r].
2896     if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2897       return true;
2898     AddrMode.HasBaseReg = false;
2899     AddrMode.BaseReg = nullptr;
2900   }
2901 
2902   // If the base register is already taken, see if we can do [r+r].
2903   if (AddrMode.Scale == 0) {
2904     AddrMode.Scale = 1;
2905     AddrMode.ScaledReg = Addr;
2906     if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
2907       return true;
2908     AddrMode.Scale = 0;
2909     AddrMode.ScaledReg = nullptr;
2910   }
2911   // Couldn't match.
2912   TPT.rollback(LastKnownGood);
2913   return false;
2914 }
2915 
2916 /// IsOperandAMemoryOperand - Check to see if all uses of OpVal by the specified
2917 /// inline asm call are due to memory operands.  If so, return true, otherwise
2918 /// return false.
IsOperandAMemoryOperand(CallInst * CI,InlineAsm * IA,Value * OpVal,const TargetMachine & TM)2919 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
2920                                     const TargetMachine &TM) {
2921   const Function *F = CI->getParent()->getParent();
2922   const TargetLowering *TLI = TM.getSubtargetImpl(*F)->getTargetLowering();
2923   const TargetRegisterInfo *TRI = TM.getSubtargetImpl(*F)->getRegisterInfo();
2924   TargetLowering::AsmOperandInfoVector TargetConstraints =
2925       TLI->ParseConstraints(TRI, ImmutableCallSite(CI));
2926   for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
2927     TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
2928 
2929     // Compute the constraint code and ConstraintType to use.
2930     TLI->ComputeConstraintToUse(OpInfo, SDValue());
2931 
2932     // If this asm operand is our Value*, and if it isn't an indirect memory
2933     // operand, we can't fold it!
2934     if (OpInfo.CallOperandVal == OpVal &&
2935         (OpInfo.ConstraintType != TargetLowering::C_Memory ||
2936          !OpInfo.isIndirect))
2937       return false;
2938   }
2939 
2940   return true;
2941 }
2942 
2943 /// FindAllMemoryUses - Recursively walk all the uses of I until we find a
2944 /// memory use.  If we find an obviously non-foldable instruction, return true.
2945 /// Add the ultimately found memory instructions to MemoryUses.
FindAllMemoryUses(Instruction * I,SmallVectorImpl<std::pair<Instruction *,unsigned>> & MemoryUses,SmallPtrSetImpl<Instruction * > & ConsideredInsts,const TargetMachine & TM)2946 static bool FindAllMemoryUses(
2947     Instruction *I,
2948     SmallVectorImpl<std::pair<Instruction *, unsigned>> &MemoryUses,
2949     SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetMachine &TM) {
2950   // If we already considered this instruction, we're done.
2951   if (!ConsideredInsts.insert(I).second)
2952     return false;
2953 
2954   // If this is an obviously unfoldable instruction, bail out.
2955   if (!MightBeFoldableInst(I))
2956     return true;
2957 
2958   // Loop over all the uses, recursively processing them.
2959   for (Use &U : I->uses()) {
2960     Instruction *UserI = cast<Instruction>(U.getUser());
2961 
2962     if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
2963       MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
2964       continue;
2965     }
2966 
2967     if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
2968       unsigned opNo = U.getOperandNo();
2969       if (opNo == 0) return true; // Storing addr, not into addr.
2970       MemoryUses.push_back(std::make_pair(SI, opNo));
2971       continue;
2972     }
2973 
2974     if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
2975       InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
2976       if (!IA) return true;
2977 
2978       // If this is a memory operand, we're cool, otherwise bail out.
2979       if (!IsOperandAMemoryOperand(CI, IA, I, TM))
2980         return true;
2981       continue;
2982     }
2983 
2984     if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TM))
2985       return true;
2986   }
2987 
2988   return false;
2989 }
2990 
2991 /// ValueAlreadyLiveAtInst - Retrn true if Val is already known to be live at
2992 /// the use site that we're folding it into.  If so, there is no cost to
2993 /// include it in the addressing mode.  KnownLive1 and KnownLive2 are two values
2994 /// that we know are live at the instruction already.
ValueAlreadyLiveAtInst(Value * Val,Value * KnownLive1,Value * KnownLive2)2995 bool AddressingModeMatcher::ValueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
2996                                                    Value *KnownLive2) {
2997   // If Val is either of the known-live values, we know it is live!
2998   if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
2999     return true;
3000 
3001   // All values other than instructions and arguments (e.g. constants) are live.
3002   if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
3003 
3004   // If Val is a constant sized alloca in the entry block, it is live, this is
3005   // true because it is just a reference to the stack/frame pointer, which is
3006   // live for the whole function.
3007   if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
3008     if (AI->isStaticAlloca())
3009       return true;
3010 
3011   // Check to see if this value is already used in the memory instruction's
3012   // block.  If so, it's already live into the block at the very least, so we
3013   // can reasonably fold it.
3014   return Val->isUsedInBasicBlock(MemoryInst->getParent());
3015 }
3016 
3017 /// IsProfitableToFoldIntoAddressingMode - It is possible for the addressing
3018 /// mode of the machine to fold the specified instruction into a load or store
3019 /// that ultimately uses it.  However, the specified instruction has multiple
3020 /// uses.  Given this, it may actually increase register pressure to fold it
3021 /// into the load.  For example, consider this code:
3022 ///
3023 ///     X = ...
3024 ///     Y = X+1
3025 ///     use(Y)   -> nonload/store
3026 ///     Z = Y+1
3027 ///     load Z
3028 ///
3029 /// In this case, Y has multiple uses, and can be folded into the load of Z
3030 /// (yielding load [X+2]).  However, doing this will cause both "X" and "X+1" to
3031 /// be live at the use(Y) line.  If we don't fold Y into load Z, we use one
3032 /// fewer register.  Since Y can't be folded into "use(Y)" we don't increase the
3033 /// number of computations either.
3034 ///
3035 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic.  If
3036 /// X was live across 'load Z' for other reasons, we actually *would* want to
3037 /// fold the addressing mode in the Z case.  This would make Y die earlier.
3038 bool AddressingModeMatcher::
IsProfitableToFoldIntoAddressingMode(Instruction * I,ExtAddrMode & AMBefore,ExtAddrMode & AMAfter)3039 IsProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
3040                                      ExtAddrMode &AMAfter) {
3041   if (IgnoreProfitability) return true;
3042 
3043   // AMBefore is the addressing mode before this instruction was folded into it,
3044   // and AMAfter is the addressing mode after the instruction was folded.  Get
3045   // the set of registers referenced by AMAfter and subtract out those
3046   // referenced by AMBefore: this is the set of values which folding in this
3047   // address extends the lifetime of.
3048   //
3049   // Note that there are only two potential values being referenced here,
3050   // BaseReg and ScaleReg (global addresses are always available, as are any
3051   // folded immediates).
3052   Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
3053 
3054   // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
3055   // lifetime wasn't extended by adding this instruction.
3056   if (ValueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
3057     BaseReg = nullptr;
3058   if (ValueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
3059     ScaledReg = nullptr;
3060 
3061   // If folding this instruction (and it's subexprs) didn't extend any live
3062   // ranges, we're ok with it.
3063   if (!BaseReg && !ScaledReg)
3064     return true;
3065 
3066   // If all uses of this instruction are ultimately load/store/inlineasm's,
3067   // check to see if their addressing modes will include this instruction.  If
3068   // so, we can fold it into all uses, so it doesn't matter if it has multiple
3069   // uses.
3070   SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
3071   SmallPtrSet<Instruction*, 16> ConsideredInsts;
3072   if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TM))
3073     return false;  // Has a non-memory, non-foldable use!
3074 
3075   // Now that we know that all uses of this instruction are part of a chain of
3076   // computation involving only operations that could theoretically be folded
3077   // into a memory use, loop over each of these uses and see if they could
3078   // *actually* fold the instruction.
3079   SmallVector<Instruction*, 32> MatchedAddrModeInsts;
3080   for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
3081     Instruction *User = MemoryUses[i].first;
3082     unsigned OpNo = MemoryUses[i].second;
3083 
3084     // Get the access type of this use.  If the use isn't a pointer, we don't
3085     // know what it accesses.
3086     Value *Address = User->getOperand(OpNo);
3087     if (!Address->getType()->isPointerTy())
3088       return false;
3089     Type *AddressAccessTy = Address->getType()->getPointerElementType();
3090 
3091     // Do a match against the root of this address, ignoring profitability. This
3092     // will tell us if the addressing mode for the memory operation will
3093     // *actually* cover the shared instruction.
3094     ExtAddrMode Result;
3095     TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3096         TPT.getRestorationPoint();
3097     AddressingModeMatcher Matcher(MatchedAddrModeInsts, TM, AddressAccessTy,
3098                                   MemoryInst, Result, InsertedTruncs,
3099                                   PromotedInsts, TPT);
3100     Matcher.IgnoreProfitability = true;
3101     bool Success = Matcher.MatchAddr(Address, 0);
3102     (void)Success; assert(Success && "Couldn't select *anything*?");
3103 
3104     // The match was to check the profitability, the changes made are not
3105     // part of the original matcher. Therefore, they should be dropped
3106     // otherwise the original matcher will not present the right state.
3107     TPT.rollback(LastKnownGood);
3108 
3109     // If the match didn't cover I, then it won't be shared by it.
3110     if (std::find(MatchedAddrModeInsts.begin(), MatchedAddrModeInsts.end(),
3111                   I) == MatchedAddrModeInsts.end())
3112       return false;
3113 
3114     MatchedAddrModeInsts.clear();
3115   }
3116 
3117   return true;
3118 }
3119 
3120 } // end anonymous namespace
3121 
3122 /// IsNonLocalValue - Return true if the specified values are defined in a
3123 /// different basic block than BB.
IsNonLocalValue(Value * V,BasicBlock * BB)3124 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
3125   if (Instruction *I = dyn_cast<Instruction>(V))
3126     return I->getParent() != BB;
3127   return false;
3128 }
3129 
3130 /// OptimizeMemoryInst - Load and Store Instructions often have
3131 /// addressing modes that can do significant amounts of computation.  As such,
3132 /// instruction selection will try to get the load or store to do as much
3133 /// computation as possible for the program.  The problem is that isel can only
3134 /// see within a single block.  As such, we sink as much legal addressing mode
3135 /// stuff into the block as possible.
3136 ///
3137 /// This method is used to optimize both load/store and inline asms with memory
3138 /// operands.
OptimizeMemoryInst(Instruction * MemoryInst,Value * Addr,Type * AccessTy)3139 bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
3140                                         Type *AccessTy) {
3141   Value *Repl = Addr;
3142 
3143   // Try to collapse single-value PHI nodes.  This is necessary to undo
3144   // unprofitable PRE transformations.
3145   SmallVector<Value*, 8> worklist;
3146   SmallPtrSet<Value*, 16> Visited;
3147   worklist.push_back(Addr);
3148 
3149   // Use a worklist to iteratively look through PHI nodes, and ensure that
3150   // the addressing mode obtained from the non-PHI roots of the graph
3151   // are equivalent.
3152   Value *Consensus = nullptr;
3153   unsigned NumUsesConsensus = 0;
3154   bool IsNumUsesConsensusValid = false;
3155   SmallVector<Instruction*, 16> AddrModeInsts;
3156   ExtAddrMode AddrMode;
3157   TypePromotionTransaction TPT;
3158   TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3159       TPT.getRestorationPoint();
3160   while (!worklist.empty()) {
3161     Value *V = worklist.back();
3162     worklist.pop_back();
3163 
3164     // Break use-def graph loops.
3165     if (!Visited.insert(V).second) {
3166       Consensus = nullptr;
3167       break;
3168     }
3169 
3170     // For a PHI node, push all of its incoming values.
3171     if (PHINode *P = dyn_cast<PHINode>(V)) {
3172       for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i)
3173         worklist.push_back(P->getIncomingValue(i));
3174       continue;
3175     }
3176 
3177     // For non-PHIs, determine the addressing mode being computed.
3178     SmallVector<Instruction*, 16> NewAddrModeInsts;
3179     ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
3180         V, AccessTy, MemoryInst, NewAddrModeInsts, *TM, InsertedTruncsSet,
3181         PromotedInsts, TPT);
3182 
3183     // This check is broken into two cases with very similar code to avoid using
3184     // getNumUses() as much as possible. Some values have a lot of uses, so
3185     // calling getNumUses() unconditionally caused a significant compile-time
3186     // regression.
3187     if (!Consensus) {
3188       Consensus = V;
3189       AddrMode = NewAddrMode;
3190       AddrModeInsts = NewAddrModeInsts;
3191       continue;
3192     } else if (NewAddrMode == AddrMode) {
3193       if (!IsNumUsesConsensusValid) {
3194         NumUsesConsensus = Consensus->getNumUses();
3195         IsNumUsesConsensusValid = true;
3196       }
3197 
3198       // Ensure that the obtained addressing mode is equivalent to that obtained
3199       // for all other roots of the PHI traversal.  Also, when choosing one
3200       // such root as representative, select the one with the most uses in order
3201       // to keep the cost modeling heuristics in AddressingModeMatcher
3202       // applicable.
3203       unsigned NumUses = V->getNumUses();
3204       if (NumUses > NumUsesConsensus) {
3205         Consensus = V;
3206         NumUsesConsensus = NumUses;
3207         AddrModeInsts = NewAddrModeInsts;
3208       }
3209       continue;
3210     }
3211 
3212     Consensus = nullptr;
3213     break;
3214   }
3215 
3216   // If the addressing mode couldn't be determined, or if multiple different
3217   // ones were determined, bail out now.
3218   if (!Consensus) {
3219     TPT.rollback(LastKnownGood);
3220     return false;
3221   }
3222   TPT.commit();
3223 
3224   // Check to see if any of the instructions supersumed by this addr mode are
3225   // non-local to I's BB.
3226   bool AnyNonLocal = false;
3227   for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
3228     if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
3229       AnyNonLocal = true;
3230       break;
3231     }
3232   }
3233 
3234   // If all the instructions matched are already in this BB, don't do anything.
3235   if (!AnyNonLocal) {
3236     DEBUG(dbgs() << "CGP: Found      local addrmode: " << AddrMode << "\n");
3237     return false;
3238   }
3239 
3240   // Insert this computation right after this user.  Since our caller is
3241   // scanning from the top of the BB to the bottom, reuse of the expr are
3242   // guaranteed to happen later.
3243   IRBuilder<> Builder(MemoryInst);
3244 
3245   // Now that we determined the addressing expression we want to use and know
3246   // that we have to sink it into this block.  Check to see if we have already
3247   // done this for some other load/store instr in this block.  If so, reuse the
3248   // computation.
3249   Value *&SunkAddr = SunkAddrs[Addr];
3250   if (SunkAddr) {
3251     DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
3252                  << *MemoryInst << "\n");
3253     if (SunkAddr->getType() != Addr->getType())
3254       SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
3255   } else if (AddrSinkUsingGEPs ||
3256              (!AddrSinkUsingGEPs.getNumOccurrences() && TM &&
3257               TM->getSubtargetImpl(*MemoryInst->getParent()->getParent())
3258                   ->useAA())) {
3259     // By default, we use the GEP-based method when AA is used later. This
3260     // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
3261     DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
3262                  << *MemoryInst << "\n");
3263     Type *IntPtrTy = TLI->getDataLayout()->getIntPtrType(Addr->getType());
3264     Value *ResultPtr = nullptr, *ResultIndex = nullptr;
3265 
3266     // First, find the pointer.
3267     if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
3268       ResultPtr = AddrMode.BaseReg;
3269       AddrMode.BaseReg = nullptr;
3270     }
3271 
3272     if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
3273       // We can't add more than one pointer together, nor can we scale a
3274       // pointer (both of which seem meaningless).
3275       if (ResultPtr || AddrMode.Scale != 1)
3276         return false;
3277 
3278       ResultPtr = AddrMode.ScaledReg;
3279       AddrMode.Scale = 0;
3280     }
3281 
3282     if (AddrMode.BaseGV) {
3283       if (ResultPtr)
3284         return false;
3285 
3286       ResultPtr = AddrMode.BaseGV;
3287     }
3288 
3289     // If the real base value actually came from an inttoptr, then the matcher
3290     // will look through it and provide only the integer value. In that case,
3291     // use it here.
3292     if (!ResultPtr && AddrMode.BaseReg) {
3293       ResultPtr =
3294         Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(), "sunkaddr");
3295       AddrMode.BaseReg = nullptr;
3296     } else if (!ResultPtr && AddrMode.Scale == 1) {
3297       ResultPtr =
3298         Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(), "sunkaddr");
3299       AddrMode.Scale = 0;
3300     }
3301 
3302     if (!ResultPtr &&
3303         !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
3304       SunkAddr = Constant::getNullValue(Addr->getType());
3305     } else if (!ResultPtr) {
3306       return false;
3307     } else {
3308       Type *I8PtrTy =
3309           Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
3310       Type *I8Ty = Builder.getInt8Ty();
3311 
3312       // Start with the base register. Do this first so that subsequent address
3313       // matching finds it last, which will prevent it from trying to match it
3314       // as the scaled value in case it happens to be a mul. That would be
3315       // problematic if we've sunk a different mul for the scale, because then
3316       // we'd end up sinking both muls.
3317       if (AddrMode.BaseReg) {
3318         Value *V = AddrMode.BaseReg;
3319         if (V->getType() != IntPtrTy)
3320           V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
3321 
3322         ResultIndex = V;
3323       }
3324 
3325       // Add the scale value.
3326       if (AddrMode.Scale) {
3327         Value *V = AddrMode.ScaledReg;
3328         if (V->getType() == IntPtrTy) {
3329           // done.
3330         } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
3331                    cast<IntegerType>(V->getType())->getBitWidth()) {
3332           V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
3333         } else {
3334           // It is only safe to sign extend the BaseReg if we know that the math
3335           // required to create it did not overflow before we extend it. Since
3336           // the original IR value was tossed in favor of a constant back when
3337           // the AddrMode was created we need to bail out gracefully if widths
3338           // do not match instead of extending it.
3339           Instruction *I = dyn_cast_or_null<Instruction>(ResultIndex);
3340           if (I && (ResultIndex != AddrMode.BaseReg))
3341             I->eraseFromParent();
3342           return false;
3343         }
3344 
3345         if (AddrMode.Scale != 1)
3346           V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
3347                                 "sunkaddr");
3348         if (ResultIndex)
3349           ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
3350         else
3351           ResultIndex = V;
3352       }
3353 
3354       // Add in the Base Offset if present.
3355       if (AddrMode.BaseOffs) {
3356         Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
3357         if (ResultIndex) {
3358           // We need to add this separately from the scale above to help with
3359           // SDAG consecutive load/store merging.
3360           if (ResultPtr->getType() != I8PtrTy)
3361             ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
3362           ResultPtr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
3363         }
3364 
3365         ResultIndex = V;
3366       }
3367 
3368       if (!ResultIndex) {
3369         SunkAddr = ResultPtr;
3370       } else {
3371         if (ResultPtr->getType() != I8PtrTy)
3372           ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
3373         SunkAddr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
3374       }
3375 
3376       if (SunkAddr->getType() != Addr->getType())
3377         SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
3378     }
3379   } else {
3380     DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
3381                  << *MemoryInst << "\n");
3382     Type *IntPtrTy = TLI->getDataLayout()->getIntPtrType(Addr->getType());
3383     Value *Result = nullptr;
3384 
3385     // Start with the base register. Do this first so that subsequent address
3386     // matching finds it last, which will prevent it from trying to match it
3387     // as the scaled value in case it happens to be a mul. That would be
3388     // problematic if we've sunk a different mul for the scale, because then
3389     // we'd end up sinking both muls.
3390     if (AddrMode.BaseReg) {
3391       Value *V = AddrMode.BaseReg;
3392       if (V->getType()->isPointerTy())
3393         V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
3394       if (V->getType() != IntPtrTy)
3395         V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
3396       Result = V;
3397     }
3398 
3399     // Add the scale value.
3400     if (AddrMode.Scale) {
3401       Value *V = AddrMode.ScaledReg;
3402       if (V->getType() == IntPtrTy) {
3403         // done.
3404       } else if (V->getType()->isPointerTy()) {
3405         V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
3406       } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
3407                  cast<IntegerType>(V->getType())->getBitWidth()) {
3408         V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
3409       } else {
3410         // It is only safe to sign extend the BaseReg if we know that the math
3411         // required to create it did not overflow before we extend it. Since
3412         // the original IR value was tossed in favor of a constant back when
3413         // the AddrMode was created we need to bail out gracefully if widths
3414         // do not match instead of extending it.
3415         Instruction *I = dyn_cast_or_null<Instruction>(Result);
3416         if (I && (Result != AddrMode.BaseReg))
3417           I->eraseFromParent();
3418         return false;
3419       }
3420       if (AddrMode.Scale != 1)
3421         V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
3422                               "sunkaddr");
3423       if (Result)
3424         Result = Builder.CreateAdd(Result, V, "sunkaddr");
3425       else
3426         Result = V;
3427     }
3428 
3429     // Add in the BaseGV if present.
3430     if (AddrMode.BaseGV) {
3431       Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
3432       if (Result)
3433         Result = Builder.CreateAdd(Result, V, "sunkaddr");
3434       else
3435         Result = V;
3436     }
3437 
3438     // Add in the Base Offset if present.
3439     if (AddrMode.BaseOffs) {
3440       Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
3441       if (Result)
3442         Result = Builder.CreateAdd(Result, V, "sunkaddr");
3443       else
3444         Result = V;
3445     }
3446 
3447     if (!Result)
3448       SunkAddr = Constant::getNullValue(Addr->getType());
3449     else
3450       SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
3451   }
3452 
3453   MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
3454 
3455   // If we have no uses, recursively delete the value and all dead instructions
3456   // using it.
3457   if (Repl->use_empty()) {
3458     // This can cause recursive deletion, which can invalidate our iterator.
3459     // Use a WeakVH to hold onto it in case this happens.
3460     WeakVH IterHandle(CurInstIterator);
3461     BasicBlock *BB = CurInstIterator->getParent();
3462 
3463     RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
3464 
3465     if (IterHandle != CurInstIterator) {
3466       // If the iterator instruction was recursively deleted, start over at the
3467       // start of the block.
3468       CurInstIterator = BB->begin();
3469       SunkAddrs.clear();
3470     }
3471   }
3472   ++NumMemoryInsts;
3473   return true;
3474 }
3475 
3476 /// OptimizeInlineAsmInst - If there are any memory operands, use
3477 /// OptimizeMemoryInst to sink their address computing into the block when
3478 /// possible / profitable.
OptimizeInlineAsmInst(CallInst * CS)3479 bool CodeGenPrepare::OptimizeInlineAsmInst(CallInst *CS) {
3480   bool MadeChange = false;
3481 
3482   const TargetRegisterInfo *TRI =
3483       TM->getSubtargetImpl(*CS->getParent()->getParent())->getRegisterInfo();
3484   TargetLowering::AsmOperandInfoVector
3485     TargetConstraints = TLI->ParseConstraints(TRI, CS);
3486   unsigned ArgNo = 0;
3487   for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
3488     TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
3489 
3490     // Compute the constraint code and ConstraintType to use.
3491     TLI->ComputeConstraintToUse(OpInfo, SDValue());
3492 
3493     if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
3494         OpInfo.isIndirect) {
3495       Value *OpVal = CS->getArgOperand(ArgNo++);
3496       MadeChange |= OptimizeMemoryInst(CS, OpVal, OpVal->getType());
3497     } else if (OpInfo.Type == InlineAsm::isInput)
3498       ArgNo++;
3499   }
3500 
3501   return MadeChange;
3502 }
3503 
3504 /// \brief Check if all the uses of \p Inst are equivalent (or free) zero or
3505 /// sign extensions.
hasSameExtUse(Instruction * Inst,const TargetLowering & TLI)3506 static bool hasSameExtUse(Instruction *Inst, const TargetLowering &TLI) {
3507   assert(!Inst->use_empty() && "Input must have at least one use");
3508   const Instruction *FirstUser = cast<Instruction>(*Inst->user_begin());
3509   bool IsSExt = isa<SExtInst>(FirstUser);
3510   Type *ExtTy = FirstUser->getType();
3511   for (const User *U : Inst->users()) {
3512     const Instruction *UI = cast<Instruction>(U);
3513     if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI)))
3514       return false;
3515     Type *CurTy = UI->getType();
3516     // Same input and output types: Same instruction after CSE.
3517     if (CurTy == ExtTy)
3518       continue;
3519 
3520     // If IsSExt is true, we are in this situation:
3521     // a = Inst
3522     // b = sext ty1 a to ty2
3523     // c = sext ty1 a to ty3
3524     // Assuming ty2 is shorter than ty3, this could be turned into:
3525     // a = Inst
3526     // b = sext ty1 a to ty2
3527     // c = sext ty2 b to ty3
3528     // However, the last sext is not free.
3529     if (IsSExt)
3530       return false;
3531 
3532     // This is a ZExt, maybe this is free to extend from one type to another.
3533     // In that case, we would not account for a different use.
3534     Type *NarrowTy;
3535     Type *LargeTy;
3536     if (ExtTy->getScalarType()->getIntegerBitWidth() >
3537         CurTy->getScalarType()->getIntegerBitWidth()) {
3538       NarrowTy = CurTy;
3539       LargeTy = ExtTy;
3540     } else {
3541       NarrowTy = ExtTy;
3542       LargeTy = CurTy;
3543     }
3544 
3545     if (!TLI.isZExtFree(NarrowTy, LargeTy))
3546       return false;
3547   }
3548   // All uses are the same or can be derived from one another for free.
3549   return true;
3550 }
3551 
3552 /// \brief Try to form ExtLd by promoting \p Exts until they reach a
3553 /// load instruction.
3554 /// If an ext(load) can be formed, it is returned via \p LI for the load
3555 /// and \p Inst for the extension.
3556 /// Otherwise LI == nullptr and Inst == nullptr.
3557 /// When some promotion happened, \p TPT contains the proper state to
3558 /// revert them.
3559 ///
3560 /// \return true when promoting was necessary to expose the ext(load)
3561 /// opportunity, false otherwise.
3562 ///
3563 /// Example:
3564 /// \code
3565 /// %ld = load i32* %addr
3566 /// %add = add nuw i32 %ld, 4
3567 /// %zext = zext i32 %add to i64
3568 /// \endcode
3569 /// =>
3570 /// \code
3571 /// %ld = load i32* %addr
3572 /// %zext = zext i32 %ld to i64
3573 /// %add = add nuw i64 %zext, 4
3574 /// \encode
3575 /// Thanks to the promotion, we can match zext(load i32*) to i64.
ExtLdPromotion(TypePromotionTransaction & TPT,LoadInst * & LI,Instruction * & Inst,const SmallVectorImpl<Instruction * > & Exts,unsigned CreatedInstsCost=0)3576 bool CodeGenPrepare::ExtLdPromotion(TypePromotionTransaction &TPT,
3577                                     LoadInst *&LI, Instruction *&Inst,
3578                                     const SmallVectorImpl<Instruction *> &Exts,
3579                                     unsigned CreatedInstsCost = 0) {
3580   // Iterate over all the extensions to see if one form an ext(load).
3581   for (auto I : Exts) {
3582     // Check if we directly have ext(load).
3583     if ((LI = dyn_cast<LoadInst>(I->getOperand(0)))) {
3584       Inst = I;
3585       // No promotion happened here.
3586       return false;
3587     }
3588     // Check whether or not we want to do any promotion.
3589     if (!TLI || !TLI->enableExtLdPromotion() || DisableExtLdPromotion)
3590       continue;
3591     // Get the action to perform the promotion.
3592     TypePromotionHelper::Action TPH = TypePromotionHelper::getAction(
3593         I, InsertedTruncsSet, *TLI, PromotedInsts);
3594     // Check if we can promote.
3595     if (!TPH)
3596       continue;
3597     // Save the current state.
3598     TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3599         TPT.getRestorationPoint();
3600     SmallVector<Instruction *, 4> NewExts;
3601     unsigned NewCreatedInstsCost = 0;
3602     unsigned ExtCost = !TLI->isExtFree(I);
3603     // Promote.
3604     Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost,
3605                              &NewExts, nullptr, *TLI);
3606     assert(PromotedVal &&
3607            "TypePromotionHelper should have filtered out those cases");
3608 
3609     // We would be able to merge only one extension in a load.
3610     // Therefore, if we have more than 1 new extension we heuristically
3611     // cut this search path, because it means we degrade the code quality.
3612     // With exactly 2, the transformation is neutral, because we will merge
3613     // one extension but leave one. However, we optimistically keep going,
3614     // because the new extension may be removed too.
3615     long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost;
3616     TotalCreatedInstsCost -= ExtCost;
3617     if (!StressExtLdPromotion &&
3618         (TotalCreatedInstsCost > 1 ||
3619          !isPromotedInstructionLegal(*TLI, PromotedVal))) {
3620       // The promotion is not profitable, rollback to the previous state.
3621       TPT.rollback(LastKnownGood);
3622       continue;
3623     }
3624     // The promotion is profitable.
3625     // Check if it exposes an ext(load).
3626     (void)ExtLdPromotion(TPT, LI, Inst, NewExts, TotalCreatedInstsCost);
3627     if (LI && (StressExtLdPromotion || NewCreatedInstsCost <= ExtCost ||
3628                // If we have created a new extension, i.e., now we have two
3629                // extensions. We must make sure one of them is merged with
3630                // the load, otherwise we may degrade the code quality.
3631                (LI->hasOneUse() || hasSameExtUse(LI, *TLI))))
3632       // Promotion happened.
3633       return true;
3634     // If this does not help to expose an ext(load) then, rollback.
3635     TPT.rollback(LastKnownGood);
3636   }
3637   // None of the extension can form an ext(load).
3638   LI = nullptr;
3639   Inst = nullptr;
3640   return false;
3641 }
3642 
3643 /// MoveExtToFormExtLoad - Move a zext or sext fed by a load into the same
3644 /// basic block as the load, unless conditions are unfavorable. This allows
3645 /// SelectionDAG to fold the extend into the load.
3646 /// \p I[in/out] the extension may be modified during the process if some
3647 /// promotions apply.
3648 ///
MoveExtToFormExtLoad(Instruction * & I)3649 bool CodeGenPrepare::MoveExtToFormExtLoad(Instruction *&I) {
3650   // Try to promote a chain of computation if it allows to form
3651   // an extended load.
3652   TypePromotionTransaction TPT;
3653   TypePromotionTransaction::ConstRestorationPt LastKnownGood =
3654     TPT.getRestorationPoint();
3655   SmallVector<Instruction *, 1> Exts;
3656   Exts.push_back(I);
3657   // Look for a load being extended.
3658   LoadInst *LI = nullptr;
3659   Instruction *OldExt = I;
3660   bool HasPromoted = ExtLdPromotion(TPT, LI, I, Exts);
3661   if (!LI || !I) {
3662     assert(!HasPromoted && !LI && "If we did not match any load instruction "
3663                                   "the code must remain the same");
3664     I = OldExt;
3665     return false;
3666   }
3667 
3668   // If they're already in the same block, there's nothing to do.
3669   // Make the cheap checks first if we did not promote.
3670   // If we promoted, we need to check if it is indeed profitable.
3671   if (!HasPromoted && LI->getParent() == I->getParent())
3672     return false;
3673 
3674   EVT VT = TLI->getValueType(I->getType());
3675   EVT LoadVT = TLI->getValueType(LI->getType());
3676 
3677   // If the load has other users and the truncate is not free, this probably
3678   // isn't worthwhile.
3679   if (!LI->hasOneUse() && TLI &&
3680       (TLI->isTypeLegal(LoadVT) || !TLI->isTypeLegal(VT)) &&
3681       !TLI->isTruncateFree(I->getType(), LI->getType())) {
3682     I = OldExt;
3683     TPT.rollback(LastKnownGood);
3684     return false;
3685   }
3686 
3687   // Check whether the target supports casts folded into loads.
3688   unsigned LType;
3689   if (isa<ZExtInst>(I))
3690     LType = ISD::ZEXTLOAD;
3691   else {
3692     assert(isa<SExtInst>(I) && "Unexpected ext type!");
3693     LType = ISD::SEXTLOAD;
3694   }
3695   if (TLI && !TLI->isLoadExtLegal(LType, VT, LoadVT)) {
3696     I = OldExt;
3697     TPT.rollback(LastKnownGood);
3698     return false;
3699   }
3700 
3701   // Move the extend into the same block as the load, so that SelectionDAG
3702   // can fold it.
3703   TPT.commit();
3704   I->removeFromParent();
3705   I->insertAfter(LI);
3706   ++NumExtsMoved;
3707   return true;
3708 }
3709 
OptimizeExtUses(Instruction * I)3710 bool CodeGenPrepare::OptimizeExtUses(Instruction *I) {
3711   BasicBlock *DefBB = I->getParent();
3712 
3713   // If the result of a {s|z}ext and its source are both live out, rewrite all
3714   // other uses of the source with result of extension.
3715   Value *Src = I->getOperand(0);
3716   if (Src->hasOneUse())
3717     return false;
3718 
3719   // Only do this xform if truncating is free.
3720   if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
3721     return false;
3722 
3723   // Only safe to perform the optimization if the source is also defined in
3724   // this block.
3725   if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
3726     return false;
3727 
3728   bool DefIsLiveOut = false;
3729   for (User *U : I->users()) {
3730     Instruction *UI = cast<Instruction>(U);
3731 
3732     // Figure out which BB this ext is used in.
3733     BasicBlock *UserBB = UI->getParent();
3734     if (UserBB == DefBB) continue;
3735     DefIsLiveOut = true;
3736     break;
3737   }
3738   if (!DefIsLiveOut)
3739     return false;
3740 
3741   // Make sure none of the uses are PHI nodes.
3742   for (User *U : Src->users()) {
3743     Instruction *UI = cast<Instruction>(U);
3744     BasicBlock *UserBB = UI->getParent();
3745     if (UserBB == DefBB) continue;
3746     // Be conservative. We don't want this xform to end up introducing
3747     // reloads just before load / store instructions.
3748     if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
3749       return false;
3750   }
3751 
3752   // InsertedTruncs - Only insert one trunc in each block once.
3753   DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
3754 
3755   bool MadeChange = false;
3756   for (Use &U : Src->uses()) {
3757     Instruction *User = cast<Instruction>(U.getUser());
3758 
3759     // Figure out which BB this ext is used in.
3760     BasicBlock *UserBB = User->getParent();
3761     if (UserBB == DefBB) continue;
3762 
3763     // Both src and def are live in this block. Rewrite the use.
3764     Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
3765 
3766     if (!InsertedTrunc) {
3767       BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
3768       InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt);
3769       InsertedTruncsSet.insert(InsertedTrunc);
3770     }
3771 
3772     // Replace a use of the {s|z}ext source with a use of the result.
3773     U = InsertedTrunc;
3774     ++NumExtUses;
3775     MadeChange = true;
3776   }
3777 
3778   return MadeChange;
3779 }
3780 
3781 /// isFormingBranchFromSelectProfitable - Returns true if a SelectInst should be
3782 /// turned into an explicit branch.
isFormingBranchFromSelectProfitable(SelectInst * SI)3783 static bool isFormingBranchFromSelectProfitable(SelectInst *SI) {
3784   // FIXME: This should use the same heuristics as IfConversion to determine
3785   // whether a select is better represented as a branch.  This requires that
3786   // branch probability metadata is preserved for the select, which is not the
3787   // case currently.
3788 
3789   CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
3790 
3791   // If the branch is predicted right, an out of order CPU can avoid blocking on
3792   // the compare.  Emit cmovs on compares with a memory operand as branches to
3793   // avoid stalls on the load from memory.  If the compare has more than one use
3794   // there's probably another cmov or setcc around so it's not worth emitting a
3795   // branch.
3796   if (!Cmp)
3797     return false;
3798 
3799   Value *CmpOp0 = Cmp->getOperand(0);
3800   Value *CmpOp1 = Cmp->getOperand(1);
3801 
3802   // We check that the memory operand has one use to avoid uses of the loaded
3803   // value directly after the compare, making branches unprofitable.
3804   return Cmp->hasOneUse() &&
3805          ((isa<LoadInst>(CmpOp0) && CmpOp0->hasOneUse()) ||
3806           (isa<LoadInst>(CmpOp1) && CmpOp1->hasOneUse()));
3807 }
3808 
3809 
3810 /// If we have a SelectInst that will likely profit from branch prediction,
3811 /// turn it into a branch.
OptimizeSelectInst(SelectInst * SI)3812 bool CodeGenPrepare::OptimizeSelectInst(SelectInst *SI) {
3813   bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
3814 
3815   // Can we convert the 'select' to CF ?
3816   if (DisableSelectToBranch || OptSize || !TLI || VectorCond)
3817     return false;
3818 
3819   TargetLowering::SelectSupportKind SelectKind;
3820   if (VectorCond)
3821     SelectKind = TargetLowering::VectorMaskSelect;
3822   else if (SI->getType()->isVectorTy())
3823     SelectKind = TargetLowering::ScalarCondVectorVal;
3824   else
3825     SelectKind = TargetLowering::ScalarValSelect;
3826 
3827   // Do we have efficient codegen support for this kind of 'selects' ?
3828   if (TLI->isSelectSupported(SelectKind)) {
3829     // We have efficient codegen support for the select instruction.
3830     // Check if it is profitable to keep this 'select'.
3831     if (!TLI->isPredictableSelectExpensive() ||
3832         !isFormingBranchFromSelectProfitable(SI))
3833       return false;
3834   }
3835 
3836   ModifiedDT = true;
3837 
3838   // First, we split the block containing the select into 2 blocks.
3839   BasicBlock *StartBlock = SI->getParent();
3840   BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(SI));
3841   BasicBlock *NextBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
3842 
3843   // Create a new block serving as the landing pad for the branch.
3844   BasicBlock *SmallBlock = BasicBlock::Create(SI->getContext(), "select.mid",
3845                                              NextBlock->getParent(), NextBlock);
3846 
3847   // Move the unconditional branch from the block with the select in it into our
3848   // landing pad block.
3849   StartBlock->getTerminator()->eraseFromParent();
3850   BranchInst::Create(NextBlock, SmallBlock);
3851 
3852   // Insert the real conditional branch based on the original condition.
3853   BranchInst::Create(NextBlock, SmallBlock, SI->getCondition(), SI);
3854 
3855   // The select itself is replaced with a PHI Node.
3856   PHINode *PN = PHINode::Create(SI->getType(), 2, "", NextBlock->begin());
3857   PN->takeName(SI);
3858   PN->addIncoming(SI->getTrueValue(), StartBlock);
3859   PN->addIncoming(SI->getFalseValue(), SmallBlock);
3860   SI->replaceAllUsesWith(PN);
3861   SI->eraseFromParent();
3862 
3863   // Instruct OptimizeBlock to skip to the next block.
3864   CurInstIterator = StartBlock->end();
3865   ++NumSelectsExpanded;
3866   return true;
3867 }
3868 
isBroadcastShuffle(ShuffleVectorInst * SVI)3869 static bool isBroadcastShuffle(ShuffleVectorInst *SVI) {
3870   SmallVector<int, 16> Mask(SVI->getShuffleMask());
3871   int SplatElem = -1;
3872   for (unsigned i = 0; i < Mask.size(); ++i) {
3873     if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
3874       return false;
3875     SplatElem = Mask[i];
3876   }
3877 
3878   return true;
3879 }
3880 
3881 /// Some targets have expensive vector shifts if the lanes aren't all the same
3882 /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
3883 /// it's often worth sinking a shufflevector splat down to its use so that
3884 /// codegen can spot all lanes are identical.
OptimizeShuffleVectorInst(ShuffleVectorInst * SVI)3885 bool CodeGenPrepare::OptimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
3886   BasicBlock *DefBB = SVI->getParent();
3887 
3888   // Only do this xform if variable vector shifts are particularly expensive.
3889   if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
3890     return false;
3891 
3892   // We only expect better codegen by sinking a shuffle if we can recognise a
3893   // constant splat.
3894   if (!isBroadcastShuffle(SVI))
3895     return false;
3896 
3897   // InsertedShuffles - Only insert a shuffle in each block once.
3898   DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
3899 
3900   bool MadeChange = false;
3901   for (User *U : SVI->users()) {
3902     Instruction *UI = cast<Instruction>(U);
3903 
3904     // Figure out which BB this ext is used in.
3905     BasicBlock *UserBB = UI->getParent();
3906     if (UserBB == DefBB) continue;
3907 
3908     // For now only apply this when the splat is used by a shift instruction.
3909     if (!UI->isShift()) continue;
3910 
3911     // Everything checks out, sink the shuffle if the user's block doesn't
3912     // already have a copy.
3913     Instruction *&InsertedShuffle = InsertedShuffles[UserBB];
3914 
3915     if (!InsertedShuffle) {
3916       BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
3917       InsertedShuffle = new ShuffleVectorInst(SVI->getOperand(0),
3918                                               SVI->getOperand(1),
3919                                               SVI->getOperand(2), "", InsertPt);
3920     }
3921 
3922     UI->replaceUsesOfWith(SVI, InsertedShuffle);
3923     MadeChange = true;
3924   }
3925 
3926   // If we removed all uses, nuke the shuffle.
3927   if (SVI->use_empty()) {
3928     SVI->eraseFromParent();
3929     MadeChange = true;
3930   }
3931 
3932   return MadeChange;
3933 }
3934 
3935 namespace {
3936 /// \brief Helper class to promote a scalar operation to a vector one.
3937 /// This class is used to move downward extractelement transition.
3938 /// E.g.,
3939 /// a = vector_op <2 x i32>
3940 /// b = extractelement <2 x i32> a, i32 0
3941 /// c = scalar_op b
3942 /// store c
3943 ///
3944 /// =>
3945 /// a = vector_op <2 x i32>
3946 /// c = vector_op a (equivalent to scalar_op on the related lane)
3947 /// * d = extractelement <2 x i32> c, i32 0
3948 /// * store d
3949 /// Assuming both extractelement and store can be combine, we get rid of the
3950 /// transition.
3951 class VectorPromoteHelper {
3952   /// Used to perform some checks on the legality of vector operations.
3953   const TargetLowering &TLI;
3954 
3955   /// Used to estimated the cost of the promoted chain.
3956   const TargetTransformInfo &TTI;
3957 
3958   /// The transition being moved downwards.
3959   Instruction *Transition;
3960   /// The sequence of instructions to be promoted.
3961   SmallVector<Instruction *, 4> InstsToBePromoted;
3962   /// Cost of combining a store and an extract.
3963   unsigned StoreExtractCombineCost;
3964   /// Instruction that will be combined with the transition.
3965   Instruction *CombineInst;
3966 
3967   /// \brief The instruction that represents the current end of the transition.
3968   /// Since we are faking the promotion until we reach the end of the chain
3969   /// of computation, we need a way to get the current end of the transition.
getEndOfTransition() const3970   Instruction *getEndOfTransition() const {
3971     if (InstsToBePromoted.empty())
3972       return Transition;
3973     return InstsToBePromoted.back();
3974   }
3975 
3976   /// \brief Return the index of the original value in the transition.
3977   /// E.g., for "extractelement <2 x i32> c, i32 1" the original value,
3978   /// c, is at index 0.
getTransitionOriginalValueIdx() const3979   unsigned getTransitionOriginalValueIdx() const {
3980     assert(isa<ExtractElementInst>(Transition) &&
3981            "Other kind of transitions are not supported yet");
3982     return 0;
3983   }
3984 
3985   /// \brief Return the index of the index in the transition.
3986   /// E.g., for "extractelement <2 x i32> c, i32 0" the index
3987   /// is at index 1.
getTransitionIdx() const3988   unsigned getTransitionIdx() const {
3989     assert(isa<ExtractElementInst>(Transition) &&
3990            "Other kind of transitions are not supported yet");
3991     return 1;
3992   }
3993 
3994   /// \brief Get the type of the transition.
3995   /// This is the type of the original value.
3996   /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the
3997   /// transition is <2 x i32>.
getTransitionType() const3998   Type *getTransitionType() const {
3999     return Transition->getOperand(getTransitionOriginalValueIdx())->getType();
4000   }
4001 
4002   /// \brief Promote \p ToBePromoted by moving \p Def downward through.
4003   /// I.e., we have the following sequence:
4004   /// Def = Transition <ty1> a to <ty2>
4005   /// b = ToBePromoted <ty2> Def, ...
4006   /// =>
4007   /// b = ToBePromoted <ty1> a, ...
4008   /// Def = Transition <ty1> ToBePromoted to <ty2>
4009   void promoteImpl(Instruction *ToBePromoted);
4010 
4011   /// \brief Check whether or not it is profitable to promote all the
4012   /// instructions enqueued to be promoted.
isProfitableToPromote()4013   bool isProfitableToPromote() {
4014     Value *ValIdx = Transition->getOperand(getTransitionOriginalValueIdx());
4015     unsigned Index = isa<ConstantInt>(ValIdx)
4016                          ? cast<ConstantInt>(ValIdx)->getZExtValue()
4017                          : -1;
4018     Type *PromotedType = getTransitionType();
4019 
4020     StoreInst *ST = cast<StoreInst>(CombineInst);
4021     unsigned AS = ST->getPointerAddressSpace();
4022     unsigned Align = ST->getAlignment();
4023     // Check if this store is supported.
4024     if (!TLI.allowsMisalignedMemoryAccesses(
4025             TLI.getValueType(ST->getValueOperand()->getType()), AS, Align)) {
4026       // If this is not supported, there is no way we can combine
4027       // the extract with the store.
4028       return false;
4029     }
4030 
4031     // The scalar chain of computation has to pay for the transition
4032     // scalar to vector.
4033     // The vector chain has to account for the combining cost.
4034     uint64_t ScalarCost =
4035         TTI.getVectorInstrCost(Transition->getOpcode(), PromotedType, Index);
4036     uint64_t VectorCost = StoreExtractCombineCost;
4037     for (const auto &Inst : InstsToBePromoted) {
4038       // Compute the cost.
4039       // By construction, all instructions being promoted are arithmetic ones.
4040       // Moreover, one argument is a constant that can be viewed as a splat
4041       // constant.
4042       Value *Arg0 = Inst->getOperand(0);
4043       bool IsArg0Constant = isa<UndefValue>(Arg0) || isa<ConstantInt>(Arg0) ||
4044                             isa<ConstantFP>(Arg0);
4045       TargetTransformInfo::OperandValueKind Arg0OVK =
4046           IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
4047                          : TargetTransformInfo::OK_AnyValue;
4048       TargetTransformInfo::OperandValueKind Arg1OVK =
4049           !IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
4050                           : TargetTransformInfo::OK_AnyValue;
4051       ScalarCost += TTI.getArithmeticInstrCost(
4052           Inst->getOpcode(), Inst->getType(), Arg0OVK, Arg1OVK);
4053       VectorCost += TTI.getArithmeticInstrCost(Inst->getOpcode(), PromotedType,
4054                                                Arg0OVK, Arg1OVK);
4055     }
4056     DEBUG(dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
4057                  << ScalarCost << "\nVector: " << VectorCost << '\n');
4058     return ScalarCost > VectorCost;
4059   }
4060 
4061   /// \brief Generate a constant vector with \p Val with the same
4062   /// number of elements as the transition.
4063   /// \p UseSplat defines whether or not \p Val should be replicated
4064   /// accross the whole vector.
4065   /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>,
4066   /// otherwise we generate a vector with as many undef as possible:
4067   /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only
4068   /// used at the index of the extract.
getConstantVector(Constant * Val,bool UseSplat) const4069   Value *getConstantVector(Constant *Val, bool UseSplat) const {
4070     unsigned ExtractIdx = UINT_MAX;
4071     if (!UseSplat) {
4072       // If we cannot determine where the constant must be, we have to
4073       // use a splat constant.
4074       Value *ValExtractIdx = Transition->getOperand(getTransitionIdx());
4075       if (ConstantInt *CstVal = dyn_cast<ConstantInt>(ValExtractIdx))
4076         ExtractIdx = CstVal->getSExtValue();
4077       else
4078         UseSplat = true;
4079     }
4080 
4081     unsigned End = getTransitionType()->getVectorNumElements();
4082     if (UseSplat)
4083       return ConstantVector::getSplat(End, Val);
4084 
4085     SmallVector<Constant *, 4> ConstVec;
4086     UndefValue *UndefVal = UndefValue::get(Val->getType());
4087     for (unsigned Idx = 0; Idx != End; ++Idx) {
4088       if (Idx == ExtractIdx)
4089         ConstVec.push_back(Val);
4090       else
4091         ConstVec.push_back(UndefVal);
4092     }
4093     return ConstantVector::get(ConstVec);
4094   }
4095 
4096   /// \brief Check if promoting to a vector type an operand at \p OperandIdx
4097   /// in \p Use can trigger undefined behavior.
canCauseUndefinedBehavior(const Instruction * Use,unsigned OperandIdx)4098   static bool canCauseUndefinedBehavior(const Instruction *Use,
4099                                         unsigned OperandIdx) {
4100     // This is not safe to introduce undef when the operand is on
4101     // the right hand side of a division-like instruction.
4102     if (OperandIdx != 1)
4103       return false;
4104     switch (Use->getOpcode()) {
4105     default:
4106       return false;
4107     case Instruction::SDiv:
4108     case Instruction::UDiv:
4109     case Instruction::SRem:
4110     case Instruction::URem:
4111       return true;
4112     case Instruction::FDiv:
4113     case Instruction::FRem:
4114       return !Use->hasNoNaNs();
4115     }
4116     llvm_unreachable(nullptr);
4117   }
4118 
4119 public:
VectorPromoteHelper(const TargetLowering & TLI,const TargetTransformInfo & TTI,Instruction * Transition,unsigned CombineCost)4120   VectorPromoteHelper(const TargetLowering &TLI, const TargetTransformInfo &TTI,
4121                       Instruction *Transition, unsigned CombineCost)
4122       : TLI(TLI), TTI(TTI), Transition(Transition),
4123         StoreExtractCombineCost(CombineCost), CombineInst(nullptr) {
4124     assert(Transition && "Do not know how to promote null");
4125   }
4126 
4127   /// \brief Check if we can promote \p ToBePromoted to \p Type.
canPromote(const Instruction * ToBePromoted) const4128   bool canPromote(const Instruction *ToBePromoted) const {
4129     // We could support CastInst too.
4130     return isa<BinaryOperator>(ToBePromoted);
4131   }
4132 
4133   /// \brief Check if it is profitable to promote \p ToBePromoted
4134   /// by moving downward the transition through.
shouldPromote(const Instruction * ToBePromoted) const4135   bool shouldPromote(const Instruction *ToBePromoted) const {
4136     // Promote only if all the operands can be statically expanded.
4137     // Indeed, we do not want to introduce any new kind of transitions.
4138     for (const Use &U : ToBePromoted->operands()) {
4139       const Value *Val = U.get();
4140       if (Val == getEndOfTransition()) {
4141         // If the use is a division and the transition is on the rhs,
4142         // we cannot promote the operation, otherwise we may create a
4143         // division by zero.
4144         if (canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()))
4145           return false;
4146         continue;
4147       }
4148       if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) &&
4149           !isa<ConstantFP>(Val))
4150         return false;
4151     }
4152     // Check that the resulting operation is legal.
4153     int ISDOpcode = TLI.InstructionOpcodeToISD(ToBePromoted->getOpcode());
4154     if (!ISDOpcode)
4155       return false;
4156     return StressStoreExtract ||
4157            TLI.isOperationLegalOrCustom(
4158                ISDOpcode, TLI.getValueType(getTransitionType(), true));
4159   }
4160 
4161   /// \brief Check whether or not \p Use can be combined
4162   /// with the transition.
4163   /// I.e., is it possible to do Use(Transition) => AnotherUse?
canCombine(const Instruction * Use)4164   bool canCombine(const Instruction *Use) { return isa<StoreInst>(Use); }
4165 
4166   /// \brief Record \p ToBePromoted as part of the chain to be promoted.
enqueueForPromotion(Instruction * ToBePromoted)4167   void enqueueForPromotion(Instruction *ToBePromoted) {
4168     InstsToBePromoted.push_back(ToBePromoted);
4169   }
4170 
4171   /// \brief Set the instruction that will be combined with the transition.
recordCombineInstruction(Instruction * ToBeCombined)4172   void recordCombineInstruction(Instruction *ToBeCombined) {
4173     assert(canCombine(ToBeCombined) && "Unsupported instruction to combine");
4174     CombineInst = ToBeCombined;
4175   }
4176 
4177   /// \brief Promote all the instructions enqueued for promotion if it is
4178   /// is profitable.
4179   /// \return True if the promotion happened, false otherwise.
promote()4180   bool promote() {
4181     // Check if there is something to promote.
4182     // Right now, if we do not have anything to combine with,
4183     // we assume the promotion is not profitable.
4184     if (InstsToBePromoted.empty() || !CombineInst)
4185       return false;
4186 
4187     // Check cost.
4188     if (!StressStoreExtract && !isProfitableToPromote())
4189       return false;
4190 
4191     // Promote.
4192     for (auto &ToBePromoted : InstsToBePromoted)
4193       promoteImpl(ToBePromoted);
4194     InstsToBePromoted.clear();
4195     return true;
4196   }
4197 };
4198 } // End of anonymous namespace.
4199 
promoteImpl(Instruction * ToBePromoted)4200 void VectorPromoteHelper::promoteImpl(Instruction *ToBePromoted) {
4201   // At this point, we know that all the operands of ToBePromoted but Def
4202   // can be statically promoted.
4203   // For Def, we need to use its parameter in ToBePromoted:
4204   // b = ToBePromoted ty1 a
4205   // Def = Transition ty1 b to ty2
4206   // Move the transition down.
4207   // 1. Replace all uses of the promoted operation by the transition.
4208   // = ... b => = ... Def.
4209   assert(ToBePromoted->getType() == Transition->getType() &&
4210          "The type of the result of the transition does not match "
4211          "the final type");
4212   ToBePromoted->replaceAllUsesWith(Transition);
4213   // 2. Update the type of the uses.
4214   // b = ToBePromoted ty2 Def => b = ToBePromoted ty1 Def.
4215   Type *TransitionTy = getTransitionType();
4216   ToBePromoted->mutateType(TransitionTy);
4217   // 3. Update all the operands of the promoted operation with promoted
4218   // operands.
4219   // b = ToBePromoted ty1 Def => b = ToBePromoted ty1 a.
4220   for (Use &U : ToBePromoted->operands()) {
4221     Value *Val = U.get();
4222     Value *NewVal = nullptr;
4223     if (Val == Transition)
4224       NewVal = Transition->getOperand(getTransitionOriginalValueIdx());
4225     else if (isa<UndefValue>(Val) || isa<ConstantInt>(Val) ||
4226              isa<ConstantFP>(Val)) {
4227       // Use a splat constant if it is not safe to use undef.
4228       NewVal = getConstantVector(
4229           cast<Constant>(Val),
4230           isa<UndefValue>(Val) ||
4231               canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()));
4232     } else
4233       llvm_unreachable("Did you modified shouldPromote and forgot to update "
4234                        "this?");
4235     ToBePromoted->setOperand(U.getOperandNo(), NewVal);
4236   }
4237   Transition->removeFromParent();
4238   Transition->insertAfter(ToBePromoted);
4239   Transition->setOperand(getTransitionOriginalValueIdx(), ToBePromoted);
4240 }
4241 
4242 /// Some targets can do store(extractelement) with one instruction.
4243 /// Try to push the extractelement towards the stores when the target
4244 /// has this feature and this is profitable.
OptimizeExtractElementInst(Instruction * Inst)4245 bool CodeGenPrepare::OptimizeExtractElementInst(Instruction *Inst) {
4246   unsigned CombineCost = UINT_MAX;
4247   if (DisableStoreExtract || !TLI ||
4248       (!StressStoreExtract &&
4249        !TLI->canCombineStoreAndExtract(Inst->getOperand(0)->getType(),
4250                                        Inst->getOperand(1), CombineCost)))
4251     return false;
4252 
4253   // At this point we know that Inst is a vector to scalar transition.
4254   // Try to move it down the def-use chain, until:
4255   // - We can combine the transition with its single use
4256   //   => we got rid of the transition.
4257   // - We escape the current basic block
4258   //   => we would need to check that we are moving it at a cheaper place and
4259   //      we do not do that for now.
4260   BasicBlock *Parent = Inst->getParent();
4261   DEBUG(dbgs() << "Found an interesting transition: " << *Inst << '\n');
4262   VectorPromoteHelper VPH(*TLI, *TTI, Inst, CombineCost);
4263   // If the transition has more than one use, assume this is not going to be
4264   // beneficial.
4265   while (Inst->hasOneUse()) {
4266     Instruction *ToBePromoted = cast<Instruction>(*Inst->user_begin());
4267     DEBUG(dbgs() << "Use: " << *ToBePromoted << '\n');
4268 
4269     if (ToBePromoted->getParent() != Parent) {
4270       DEBUG(dbgs() << "Instruction to promote is in a different block ("
4271                    << ToBePromoted->getParent()->getName()
4272                    << ") than the transition (" << Parent->getName() << ").\n");
4273       return false;
4274     }
4275 
4276     if (VPH.canCombine(ToBePromoted)) {
4277       DEBUG(dbgs() << "Assume " << *Inst << '\n'
4278                    << "will be combined with: " << *ToBePromoted << '\n');
4279       VPH.recordCombineInstruction(ToBePromoted);
4280       bool Changed = VPH.promote();
4281       NumStoreExtractExposed += Changed;
4282       return Changed;
4283     }
4284 
4285     DEBUG(dbgs() << "Try promoting.\n");
4286     if (!VPH.canPromote(ToBePromoted) || !VPH.shouldPromote(ToBePromoted))
4287       return false;
4288 
4289     DEBUG(dbgs() << "Promoting is possible... Enqueue for promotion!\n");
4290 
4291     VPH.enqueueForPromotion(ToBePromoted);
4292     Inst = ToBePromoted;
4293   }
4294   return false;
4295 }
4296 
OptimizeInst(Instruction * I,bool & ModifiedDT)4297 bool CodeGenPrepare::OptimizeInst(Instruction *I, bool& ModifiedDT) {
4298   if (PHINode *P = dyn_cast<PHINode>(I)) {
4299     // It is possible for very late stage optimizations (such as SimplifyCFG)
4300     // to introduce PHI nodes too late to be cleaned up.  If we detect such a
4301     // trivial PHI, go ahead and zap it here.
4302     const DataLayout &DL = I->getModule()->getDataLayout();
4303     if (Value *V = SimplifyInstruction(P, DL, TLInfo, nullptr)) {
4304       P->replaceAllUsesWith(V);
4305       P->eraseFromParent();
4306       ++NumPHIsElim;
4307       return true;
4308     }
4309     return false;
4310   }
4311 
4312   if (CastInst *CI = dyn_cast<CastInst>(I)) {
4313     // If the source of the cast is a constant, then this should have
4314     // already been constant folded.  The only reason NOT to constant fold
4315     // it is if something (e.g. LSR) was careful to place the constant
4316     // evaluation in a block other than then one that uses it (e.g. to hoist
4317     // the address of globals out of a loop).  If this is the case, we don't
4318     // want to forward-subst the cast.
4319     if (isa<Constant>(CI->getOperand(0)))
4320       return false;
4321 
4322     if (TLI && OptimizeNoopCopyExpression(CI, *TLI))
4323       return true;
4324 
4325     if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
4326       /// Sink a zext or sext into its user blocks if the target type doesn't
4327       /// fit in one register
4328       if (TLI && TLI->getTypeAction(CI->getContext(),
4329                                     TLI->getValueType(CI->getType())) ==
4330                      TargetLowering::TypeExpandInteger) {
4331         return SinkCast(CI);
4332       } else {
4333         bool MadeChange = MoveExtToFormExtLoad(I);
4334         return MadeChange | OptimizeExtUses(I);
4335       }
4336     }
4337     return false;
4338   }
4339 
4340   if (CmpInst *CI = dyn_cast<CmpInst>(I))
4341     if (!TLI || !TLI->hasMultipleConditionRegisters())
4342       return OptimizeCmpExpression(CI);
4343 
4344   if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
4345     if (TLI)
4346       return OptimizeMemoryInst(I, I->getOperand(0), LI->getType());
4347     return false;
4348   }
4349 
4350   if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
4351     if (TLI)
4352       return OptimizeMemoryInst(I, SI->getOperand(1),
4353                                 SI->getOperand(0)->getType());
4354     return false;
4355   }
4356 
4357   BinaryOperator *BinOp = dyn_cast<BinaryOperator>(I);
4358 
4359   if (BinOp && (BinOp->getOpcode() == Instruction::AShr ||
4360                 BinOp->getOpcode() == Instruction::LShr)) {
4361     ConstantInt *CI = dyn_cast<ConstantInt>(BinOp->getOperand(1));
4362     if (TLI && CI && TLI->hasExtractBitsInsn())
4363       return OptimizeExtractBits(BinOp, CI, *TLI);
4364 
4365     return false;
4366   }
4367 
4368   if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
4369     if (GEPI->hasAllZeroIndices()) {
4370       /// The GEP operand must be a pointer, so must its result -> BitCast
4371       Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
4372                                         GEPI->getName(), GEPI);
4373       GEPI->replaceAllUsesWith(NC);
4374       GEPI->eraseFromParent();
4375       ++NumGEPsElim;
4376       OptimizeInst(NC, ModifiedDT);
4377       return true;
4378     }
4379     return false;
4380   }
4381 
4382   if (CallInst *CI = dyn_cast<CallInst>(I))
4383     return OptimizeCallInst(CI, ModifiedDT);
4384 
4385   if (SelectInst *SI = dyn_cast<SelectInst>(I))
4386     return OptimizeSelectInst(SI);
4387 
4388   if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I))
4389     return OptimizeShuffleVectorInst(SVI);
4390 
4391   if (isa<ExtractElementInst>(I))
4392     return OptimizeExtractElementInst(I);
4393 
4394   return false;
4395 }
4396 
4397 // In this pass we look for GEP and cast instructions that are used
4398 // across basic blocks and rewrite them to improve basic-block-at-a-time
4399 // selection.
OptimizeBlock(BasicBlock & BB,bool & ModifiedDT)4400 bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB, bool& ModifiedDT) {
4401   SunkAddrs.clear();
4402   bool MadeChange = false;
4403 
4404   CurInstIterator = BB.begin();
4405   while (CurInstIterator != BB.end()) {
4406     MadeChange |= OptimizeInst(CurInstIterator++, ModifiedDT);
4407     if (ModifiedDT)
4408       return true;
4409   }
4410   MadeChange |= DupRetToEnableTailCallOpts(&BB);
4411 
4412   return MadeChange;
4413 }
4414 
4415 // llvm.dbg.value is far away from the value then iSel may not be able
4416 // handle it properly. iSel will drop llvm.dbg.value if it can not
4417 // find a node corresponding to the value.
PlaceDbgValues(Function & F)4418 bool CodeGenPrepare::PlaceDbgValues(Function &F) {
4419   bool MadeChange = false;
4420   for (BasicBlock &BB : F) {
4421     Instruction *PrevNonDbgInst = nullptr;
4422     for (BasicBlock::iterator BI = BB.begin(), BE = BB.end(); BI != BE;) {
4423       Instruction *Insn = BI++;
4424       DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
4425       // Leave dbg.values that refer to an alloca alone. These
4426       // instrinsics describe the address of a variable (= the alloca)
4427       // being taken.  They should not be moved next to the alloca
4428       // (and to the beginning of the scope), but rather stay close to
4429       // where said address is used.
4430       if (!DVI || (DVI->getValue() && isa<AllocaInst>(DVI->getValue()))) {
4431         PrevNonDbgInst = Insn;
4432         continue;
4433       }
4434 
4435       Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
4436       if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
4437         DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
4438         DVI->removeFromParent();
4439         if (isa<PHINode>(VI))
4440           DVI->insertBefore(VI->getParent()->getFirstInsertionPt());
4441         else
4442           DVI->insertAfter(VI);
4443         MadeChange = true;
4444         ++NumDbgValueMoved;
4445       }
4446     }
4447   }
4448   return MadeChange;
4449 }
4450 
4451 // If there is a sequence that branches based on comparing a single bit
4452 // against zero that can be combined into a single instruction, and the
4453 // target supports folding these into a single instruction, sink the
4454 // mask and compare into the branch uses. Do this before OptimizeBlock ->
4455 // OptimizeInst -> OptimizeCmpExpression, which perturbs the pattern being
4456 // searched for.
sinkAndCmp(Function & F)4457 bool CodeGenPrepare::sinkAndCmp(Function &F) {
4458   if (!EnableAndCmpSinking)
4459     return false;
4460   if (!TLI || !TLI->isMaskAndBranchFoldingLegal())
4461     return false;
4462   bool MadeChange = false;
4463   for (Function::iterator I = F.begin(), E = F.end(); I != E; ) {
4464     BasicBlock *BB = I++;
4465 
4466     // Does this BB end with the following?
4467     //   %andVal = and %val, #single-bit-set
4468     //   %icmpVal = icmp %andResult, 0
4469     //   br i1 %cmpVal label %dest1, label %dest2"
4470     BranchInst *Brcc = dyn_cast<BranchInst>(BB->getTerminator());
4471     if (!Brcc || !Brcc->isConditional())
4472       continue;
4473     ICmpInst *Cmp = dyn_cast<ICmpInst>(Brcc->getOperand(0));
4474     if (!Cmp || Cmp->getParent() != BB)
4475       continue;
4476     ConstantInt *Zero = dyn_cast<ConstantInt>(Cmp->getOperand(1));
4477     if (!Zero || !Zero->isZero())
4478       continue;
4479     Instruction *And = dyn_cast<Instruction>(Cmp->getOperand(0));
4480     if (!And || And->getOpcode() != Instruction::And || And->getParent() != BB)
4481       continue;
4482     ConstantInt* Mask = dyn_cast<ConstantInt>(And->getOperand(1));
4483     if (!Mask || !Mask->getUniqueInteger().isPowerOf2())
4484       continue;
4485     DEBUG(dbgs() << "found and; icmp ?,0; brcc\n"); DEBUG(BB->dump());
4486 
4487     // Push the "and; icmp" for any users that are conditional branches.
4488     // Since there can only be one branch use per BB, we don't need to keep
4489     // track of which BBs we insert into.
4490     for (Value::use_iterator UI = Cmp->use_begin(), E = Cmp->use_end();
4491          UI != E; ) {
4492       Use &TheUse = *UI;
4493       // Find brcc use.
4494       BranchInst *BrccUser = dyn_cast<BranchInst>(*UI);
4495       ++UI;
4496       if (!BrccUser || !BrccUser->isConditional())
4497         continue;
4498       BasicBlock *UserBB = BrccUser->getParent();
4499       if (UserBB == BB) continue;
4500       DEBUG(dbgs() << "found Brcc use\n");
4501 
4502       // Sink the "and; icmp" to use.
4503       MadeChange = true;
4504       BinaryOperator *NewAnd =
4505         BinaryOperator::CreateAnd(And->getOperand(0), And->getOperand(1), "",
4506                                   BrccUser);
4507       CmpInst *NewCmp =
4508         CmpInst::Create(Cmp->getOpcode(), Cmp->getPredicate(), NewAnd, Zero,
4509                         "", BrccUser);
4510       TheUse = NewCmp;
4511       ++NumAndCmpsMoved;
4512       DEBUG(BrccUser->getParent()->dump());
4513     }
4514   }
4515   return MadeChange;
4516 }
4517 
4518 /// \brief Retrieve the probabilities of a conditional branch. Returns true on
4519 /// success, or returns false if no or invalid metadata was found.
extractBranchMetadata(BranchInst * BI,uint64_t & ProbTrue,uint64_t & ProbFalse)4520 static bool extractBranchMetadata(BranchInst *BI,
4521                                   uint64_t &ProbTrue, uint64_t &ProbFalse) {
4522   assert(BI->isConditional() &&
4523          "Looking for probabilities on unconditional branch?");
4524   auto *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
4525   if (!ProfileData || ProfileData->getNumOperands() != 3)
4526     return false;
4527 
4528   const auto *CITrue =
4529       mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(1));
4530   const auto *CIFalse =
4531       mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(2));
4532   if (!CITrue || !CIFalse)
4533     return false;
4534 
4535   ProbTrue = CITrue->getValue().getZExtValue();
4536   ProbFalse = CIFalse->getValue().getZExtValue();
4537 
4538   return true;
4539 }
4540 
4541 /// \brief Scale down both weights to fit into uint32_t.
scaleWeights(uint64_t & NewTrue,uint64_t & NewFalse)4542 static void scaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) {
4543   uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse;
4544   uint32_t Scale = (NewMax / UINT32_MAX) + 1;
4545   NewTrue = NewTrue / Scale;
4546   NewFalse = NewFalse / Scale;
4547 }
4548 
4549 /// \brief Some targets prefer to split a conditional branch like:
4550 /// \code
4551 ///   %0 = icmp ne i32 %a, 0
4552 ///   %1 = icmp ne i32 %b, 0
4553 ///   %or.cond = or i1 %0, %1
4554 ///   br i1 %or.cond, label %TrueBB, label %FalseBB
4555 /// \endcode
4556 /// into multiple branch instructions like:
4557 /// \code
4558 ///   bb1:
4559 ///     %0 = icmp ne i32 %a, 0
4560 ///     br i1 %0, label %TrueBB, label %bb2
4561 ///   bb2:
4562 ///     %1 = icmp ne i32 %b, 0
4563 ///     br i1 %1, label %TrueBB, label %FalseBB
4564 /// \endcode
4565 /// This usually allows instruction selection to do even further optimizations
4566 /// and combine the compare with the branch instruction. Currently this is
4567 /// applied for targets which have "cheap" jump instructions.
4568 ///
4569 /// FIXME: Remove the (equivalent?) implementation in SelectionDAG.
4570 ///
splitBranchCondition(Function & F)4571 bool CodeGenPrepare::splitBranchCondition(Function &F) {
4572   if (!TM || !TM->Options.EnableFastISel || !TLI || TLI->isJumpExpensive())
4573     return false;
4574 
4575   bool MadeChange = false;
4576   for (auto &BB : F) {
4577     // Does this BB end with the following?
4578     //   %cond1 = icmp|fcmp|binary instruction ...
4579     //   %cond2 = icmp|fcmp|binary instruction ...
4580     //   %cond.or = or|and i1 %cond1, cond2
4581     //   br i1 %cond.or label %dest1, label %dest2"
4582     BinaryOperator *LogicOp;
4583     BasicBlock *TBB, *FBB;
4584     if (!match(BB.getTerminator(), m_Br(m_OneUse(m_BinOp(LogicOp)), TBB, FBB)))
4585       continue;
4586 
4587     unsigned Opc;
4588     Value *Cond1, *Cond2;
4589     if (match(LogicOp, m_And(m_OneUse(m_Value(Cond1)),
4590                              m_OneUse(m_Value(Cond2)))))
4591       Opc = Instruction::And;
4592     else if (match(LogicOp, m_Or(m_OneUse(m_Value(Cond1)),
4593                                  m_OneUse(m_Value(Cond2)))))
4594       Opc = Instruction::Or;
4595     else
4596       continue;
4597 
4598     if (!match(Cond1, m_CombineOr(m_Cmp(), m_BinOp())) ||
4599         !match(Cond2, m_CombineOr(m_Cmp(), m_BinOp()))   )
4600       continue;
4601 
4602     DEBUG(dbgs() << "Before branch condition splitting\n"; BB.dump());
4603 
4604     // Create a new BB.
4605     auto *InsertBefore = std::next(Function::iterator(BB))
4606         .getNodePtrUnchecked();
4607     auto TmpBB = BasicBlock::Create(BB.getContext(),
4608                                     BB.getName() + ".cond.split",
4609                                     BB.getParent(), InsertBefore);
4610 
4611     // Update original basic block by using the first condition directly by the
4612     // branch instruction and removing the no longer needed and/or instruction.
4613     auto *Br1 = cast<BranchInst>(BB.getTerminator());
4614     Br1->setCondition(Cond1);
4615     LogicOp->eraseFromParent();
4616 
4617     // Depending on the conditon we have to either replace the true or the false
4618     // successor of the original branch instruction.
4619     if (Opc == Instruction::And)
4620       Br1->setSuccessor(0, TmpBB);
4621     else
4622       Br1->setSuccessor(1, TmpBB);
4623 
4624     // Fill in the new basic block.
4625     auto *Br2 = IRBuilder<>(TmpBB).CreateCondBr(Cond2, TBB, FBB);
4626     if (auto *I = dyn_cast<Instruction>(Cond2)) {
4627       I->removeFromParent();
4628       I->insertBefore(Br2);
4629     }
4630 
4631     // Update PHI nodes in both successors. The original BB needs to be
4632     // replaced in one succesor's PHI nodes, because the branch comes now from
4633     // the newly generated BB (NewBB). In the other successor we need to add one
4634     // incoming edge to the PHI nodes, because both branch instructions target
4635     // now the same successor. Depending on the original branch condition
4636     // (and/or) we have to swap the successors (TrueDest, FalseDest), so that
4637     // we perfrom the correct update for the PHI nodes.
4638     // This doesn't change the successor order of the just created branch
4639     // instruction (or any other instruction).
4640     if (Opc == Instruction::Or)
4641       std::swap(TBB, FBB);
4642 
4643     // Replace the old BB with the new BB.
4644     for (auto &I : *TBB) {
4645       PHINode *PN = dyn_cast<PHINode>(&I);
4646       if (!PN)
4647         break;
4648       int i;
4649       while ((i = PN->getBasicBlockIndex(&BB)) >= 0)
4650         PN->setIncomingBlock(i, TmpBB);
4651     }
4652 
4653     // Add another incoming edge form the new BB.
4654     for (auto &I : *FBB) {
4655       PHINode *PN = dyn_cast<PHINode>(&I);
4656       if (!PN)
4657         break;
4658       auto *Val = PN->getIncomingValueForBlock(&BB);
4659       PN->addIncoming(Val, TmpBB);
4660     }
4661 
4662     // Update the branch weights (from SelectionDAGBuilder::
4663     // FindMergedConditions).
4664     if (Opc == Instruction::Or) {
4665       // Codegen X | Y as:
4666       // BB1:
4667       //   jmp_if_X TBB
4668       //   jmp TmpBB
4669       // TmpBB:
4670       //   jmp_if_Y TBB
4671       //   jmp FBB
4672       //
4673 
4674       // We have flexibility in setting Prob for BB1 and Prob for NewBB.
4675       // The requirement is that
4676       //   TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB)
4677       //     = TrueProb for orignal BB.
4678       // Assuming the orignal weights are A and B, one choice is to set BB1's
4679       // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice
4680       // assumes that
4681       //   TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB.
4682       // Another choice is to assume TrueProb for BB1 equals to TrueProb for
4683       // TmpBB, but the math is more complicated.
4684       uint64_t TrueWeight, FalseWeight;
4685       if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
4686         uint64_t NewTrueWeight = TrueWeight;
4687         uint64_t NewFalseWeight = TrueWeight + 2 * FalseWeight;
4688         scaleWeights(NewTrueWeight, NewFalseWeight);
4689         Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
4690                          .createBranchWeights(TrueWeight, FalseWeight));
4691 
4692         NewTrueWeight = TrueWeight;
4693         NewFalseWeight = 2 * FalseWeight;
4694         scaleWeights(NewTrueWeight, NewFalseWeight);
4695         Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
4696                          .createBranchWeights(TrueWeight, FalseWeight));
4697       }
4698     } else {
4699       // Codegen X & Y as:
4700       // BB1:
4701       //   jmp_if_X TmpBB
4702       //   jmp FBB
4703       // TmpBB:
4704       //   jmp_if_Y TBB
4705       //   jmp FBB
4706       //
4707       //  This requires creation of TmpBB after CurBB.
4708 
4709       // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
4710       // The requirement is that
4711       //   FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB)
4712       //     = FalseProb for orignal BB.
4713       // Assuming the orignal weights are A and B, one choice is to set BB1's
4714       // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice
4715       // assumes that
4716       //   FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB.
4717       uint64_t TrueWeight, FalseWeight;
4718       if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
4719         uint64_t NewTrueWeight = 2 * TrueWeight + FalseWeight;
4720         uint64_t NewFalseWeight = FalseWeight;
4721         scaleWeights(NewTrueWeight, NewFalseWeight);
4722         Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
4723                          .createBranchWeights(TrueWeight, FalseWeight));
4724 
4725         NewTrueWeight = 2 * TrueWeight;
4726         NewFalseWeight = FalseWeight;
4727         scaleWeights(NewTrueWeight, NewFalseWeight);
4728         Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
4729                          .createBranchWeights(TrueWeight, FalseWeight));
4730       }
4731     }
4732 
4733     // Note: No point in getting fancy here, since the DT info is never
4734     // available to CodeGenPrepare.
4735     ModifiedDT = true;
4736 
4737     MadeChange = true;
4738 
4739     DEBUG(dbgs() << "After branch condition splitting\n"; BB.dump();
4740           TmpBB->dump());
4741   }
4742   return MadeChange;
4743 }
4744