• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 //===- IndVarSimplify.cpp - Induction Variable Elimination ----------------===//
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 transformation analyzes and transforms the induction variables (and
11 // computations derived from them) into simpler forms suitable for subsequent
12 // analysis and transformation.
13 //
14 // If the trip count of a loop is computable, this pass also makes the following
15 // changes:
16 //   1. The exit condition for the loop is canonicalized to compare the
17 //      induction value against the exit value.  This turns loops like:
18 //        'for (i = 7; i*i < 1000; ++i)' into 'for (i = 0; i != 25; ++i)'
19 //   2. Any use outside of the loop of an expression derived from the indvar
20 //      is changed to compute the derived value outside of the loop, eliminating
21 //      the dependence on the exit value of the induction variable.  If the only
22 //      purpose of the loop is to compute the exit value of some derived
23 //      expression, this transformation will make the loop dead.
24 //
25 //===----------------------------------------------------------------------===//
26 
27 #include "llvm/Transforms/Scalar.h"
28 #include "llvm/ADT/DenseMap.h"
29 #include "llvm/ADT/SmallVector.h"
30 #include "llvm/ADT/Statistic.h"
31 #include "llvm/Analysis/GlobalsModRef.h"
32 #include "llvm/Analysis/LoopInfo.h"
33 #include "llvm/Analysis/LoopPass.h"
34 #include "llvm/Analysis/ScalarEvolutionExpander.h"
35 #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h"
36 #include "llvm/Analysis/TargetLibraryInfo.h"
37 #include "llvm/Analysis/TargetTransformInfo.h"
38 #include "llvm/IR/BasicBlock.h"
39 #include "llvm/IR/CFG.h"
40 #include "llvm/IR/Constants.h"
41 #include "llvm/IR/DataLayout.h"
42 #include "llvm/IR/Dominators.h"
43 #include "llvm/IR/Instructions.h"
44 #include "llvm/IR/IntrinsicInst.h"
45 #include "llvm/IR/LLVMContext.h"
46 #include "llvm/IR/PatternMatch.h"
47 #include "llvm/IR/Type.h"
48 #include "llvm/Support/CommandLine.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/raw_ostream.h"
51 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
52 #include "llvm/Transforms/Utils/Local.h"
53 #include "llvm/Transforms/Utils/LoopUtils.h"
54 #include "llvm/Transforms/Utils/SimplifyIndVar.h"
55 using namespace llvm;
56 
57 #define DEBUG_TYPE "indvars"
58 
59 STATISTIC(NumWidened     , "Number of indvars widened");
60 STATISTIC(NumReplaced    , "Number of exit values replaced");
61 STATISTIC(NumLFTR        , "Number of loop exit tests replaced");
62 STATISTIC(NumElimExt     , "Number of IV sign/zero extends eliminated");
63 STATISTIC(NumElimIV      , "Number of congruent IVs eliminated");
64 
65 // Trip count verification can be enabled by default under NDEBUG if we
66 // implement a strong expression equivalence checker in SCEV. Until then, we
67 // use the verify-indvars flag, which may assert in some cases.
68 static cl::opt<bool> VerifyIndvars(
69   "verify-indvars", cl::Hidden,
70   cl::desc("Verify the ScalarEvolution result after running indvars"));
71 
72 static cl::opt<bool> ReduceLiveIVs("liv-reduce", cl::Hidden,
73   cl::desc("Reduce live induction variables."));
74 
75 enum ReplaceExitVal { NeverRepl, OnlyCheapRepl, AlwaysRepl };
76 
77 static cl::opt<ReplaceExitVal> ReplaceExitValue(
78     "replexitval", cl::Hidden, cl::init(OnlyCheapRepl),
79     cl::desc("Choose the strategy to replace exit value in IndVarSimplify"),
80     cl::values(clEnumValN(NeverRepl, "never", "never replace exit value"),
81                clEnumValN(OnlyCheapRepl, "cheap",
82                           "only replace exit value when the cost is cheap"),
83                clEnumValN(AlwaysRepl, "always",
84                           "always replace exit value whenever possible"),
85                clEnumValEnd));
86 
87 namespace {
88 struct RewritePhi;
89 
90 class IndVarSimplify : public LoopPass {
91   LoopInfo                  *LI;
92   ScalarEvolution           *SE;
93   DominatorTree             *DT;
94   TargetLibraryInfo         *TLI;
95   const TargetTransformInfo *TTI;
96 
97   SmallVector<WeakVH, 16> DeadInsts;
98   bool Changed;
99 public:
100 
101   static char ID; // Pass identification, replacement for typeid
IndVarSimplify()102   IndVarSimplify()
103     : LoopPass(ID), LI(nullptr), SE(nullptr), DT(nullptr), Changed(false) {
104     initializeIndVarSimplifyPass(*PassRegistry::getPassRegistry());
105   }
106 
107   bool runOnLoop(Loop *L, LPPassManager &LPM) override;
108 
getAnalysisUsage(AnalysisUsage & AU) const109   void getAnalysisUsage(AnalysisUsage &AU) const override {
110     AU.addRequired<DominatorTreeWrapperPass>();
111     AU.addRequired<LoopInfoWrapperPass>();
112     AU.addRequired<ScalarEvolutionWrapperPass>();
113     AU.addRequiredID(LoopSimplifyID);
114     AU.addRequiredID(LCSSAID);
115     AU.addPreserved<GlobalsAAWrapperPass>();
116     AU.addPreserved<ScalarEvolutionWrapperPass>();
117     AU.addPreservedID(LoopSimplifyID);
118     AU.addPreservedID(LCSSAID);
119     AU.setPreservesCFG();
120   }
121 
122 private:
releaseMemory()123   void releaseMemory() override {
124     DeadInsts.clear();
125   }
126 
127   bool isValidRewrite(Value *FromVal, Value *ToVal);
128 
129   void handleFloatingPointIV(Loop *L, PHINode *PH);
130   void rewriteNonIntegerIVs(Loop *L);
131 
132   void simplifyAndExtend(Loop *L, SCEVExpander &Rewriter, LoopInfo *LI);
133 
134   bool canLoopBeDeleted(Loop *L, SmallVector<RewritePhi, 8> &RewritePhiSet);
135   void rewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter);
136 
137   Value *linearFunctionTestReplace(Loop *L, const SCEV *BackedgeTakenCount,
138                                    PHINode *IndVar, SCEVExpander &Rewriter);
139 
140   void sinkUnusedInvariants(Loop *L);
141 
142   Value *expandSCEVIfNeeded(SCEVExpander &Rewriter, const SCEV *S, Loop *L,
143                             Instruction *InsertPt, Type *Ty);
144 };
145 }
146 
147 char IndVarSimplify::ID = 0;
148 INITIALIZE_PASS_BEGIN(IndVarSimplify, "indvars",
149                 "Induction Variable Simplification", false, false)
INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)150 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
151 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
152 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
153 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
154 INITIALIZE_PASS_DEPENDENCY(LCSSA)
155 INITIALIZE_PASS_END(IndVarSimplify, "indvars",
156                 "Induction Variable Simplification", false, false)
157 
158 Pass *llvm::createIndVarSimplifyPass() {
159   return new IndVarSimplify();
160 }
161 
162 /// Return true if the SCEV expansion generated by the rewriter can replace the
163 /// original value. SCEV guarantees that it produces the same value, but the way
164 /// it is produced may be illegal IR.  Ideally, this function will only be
165 /// called for verification.
isValidRewrite(Value * FromVal,Value * ToVal)166 bool IndVarSimplify::isValidRewrite(Value *FromVal, Value *ToVal) {
167   // If an SCEV expression subsumed multiple pointers, its expansion could
168   // reassociate the GEP changing the base pointer. This is illegal because the
169   // final address produced by a GEP chain must be inbounds relative to its
170   // underlying object. Otherwise basic alias analysis, among other things,
171   // could fail in a dangerous way. Ultimately, SCEV will be improved to avoid
172   // producing an expression involving multiple pointers. Until then, we must
173   // bail out here.
174   //
175   // Retrieve the pointer operand of the GEP. Don't use GetUnderlyingObject
176   // because it understands lcssa phis while SCEV does not.
177   Value *FromPtr = FromVal;
178   Value *ToPtr = ToVal;
179   if (auto *GEP = dyn_cast<GEPOperator>(FromVal)) {
180     FromPtr = GEP->getPointerOperand();
181   }
182   if (auto *GEP = dyn_cast<GEPOperator>(ToVal)) {
183     ToPtr = GEP->getPointerOperand();
184   }
185   if (FromPtr != FromVal || ToPtr != ToVal) {
186     // Quickly check the common case
187     if (FromPtr == ToPtr)
188       return true;
189 
190     // SCEV may have rewritten an expression that produces the GEP's pointer
191     // operand. That's ok as long as the pointer operand has the same base
192     // pointer. Unlike GetUnderlyingObject(), getPointerBase() will find the
193     // base of a recurrence. This handles the case in which SCEV expansion
194     // converts a pointer type recurrence into a nonrecurrent pointer base
195     // indexed by an integer recurrence.
196 
197     // If the GEP base pointer is a vector of pointers, abort.
198     if (!FromPtr->getType()->isPointerTy() || !ToPtr->getType()->isPointerTy())
199       return false;
200 
201     const SCEV *FromBase = SE->getPointerBase(SE->getSCEV(FromPtr));
202     const SCEV *ToBase = SE->getPointerBase(SE->getSCEV(ToPtr));
203     if (FromBase == ToBase)
204       return true;
205 
206     DEBUG(dbgs() << "INDVARS: GEP rewrite bail out "
207           << *FromBase << " != " << *ToBase << "\n");
208 
209     return false;
210   }
211   return true;
212 }
213 
214 /// Determine the insertion point for this user. By default, insert immediately
215 /// before the user. SCEVExpander or LICM will hoist loop invariants out of the
216 /// loop. For PHI nodes, there may be multiple uses, so compute the nearest
217 /// common dominator for the incoming blocks.
getInsertPointForUses(Instruction * User,Value * Def,DominatorTree * DT,LoopInfo * LI)218 static Instruction *getInsertPointForUses(Instruction *User, Value *Def,
219                                           DominatorTree *DT, LoopInfo *LI) {
220   PHINode *PHI = dyn_cast<PHINode>(User);
221   if (!PHI)
222     return User;
223 
224   Instruction *InsertPt = nullptr;
225   for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i) {
226     if (PHI->getIncomingValue(i) != Def)
227       continue;
228 
229     BasicBlock *InsertBB = PHI->getIncomingBlock(i);
230     if (!InsertPt) {
231       InsertPt = InsertBB->getTerminator();
232       continue;
233     }
234     InsertBB = DT->findNearestCommonDominator(InsertPt->getParent(), InsertBB);
235     InsertPt = InsertBB->getTerminator();
236   }
237   assert(InsertPt && "Missing phi operand");
238 
239   auto *DefI = dyn_cast<Instruction>(Def);
240   if (!DefI)
241     return InsertPt;
242 
243   assert(DT->dominates(DefI, InsertPt) && "def does not dominate all uses");
244 
245   auto *L = LI->getLoopFor(DefI->getParent());
246   assert(!L || L->contains(LI->getLoopFor(InsertPt->getParent())));
247 
248   for (auto *DTN = (*DT)[InsertPt->getParent()]; DTN; DTN = DTN->getIDom())
249     if (LI->getLoopFor(DTN->getBlock()) == L)
250       return DTN->getBlock()->getTerminator();
251 
252   llvm_unreachable("DefI dominates InsertPt!");
253 }
254 
255 //===----------------------------------------------------------------------===//
256 // rewriteNonIntegerIVs and helpers. Prefer integer IVs.
257 //===----------------------------------------------------------------------===//
258 
259 /// Convert APF to an integer, if possible.
ConvertToSInt(const APFloat & APF,int64_t & IntVal)260 static bool ConvertToSInt(const APFloat &APF, int64_t &IntVal) {
261   bool isExact = false;
262   // See if we can convert this to an int64_t
263   uint64_t UIntVal;
264   if (APF.convertToInteger(&UIntVal, 64, true, APFloat::rmTowardZero,
265                            &isExact) != APFloat::opOK || !isExact)
266     return false;
267   IntVal = UIntVal;
268   return true;
269 }
270 
271 /// If the loop has floating induction variable then insert corresponding
272 /// integer induction variable if possible.
273 /// For example,
274 /// for(double i = 0; i < 10000; ++i)
275 ///   bar(i)
276 /// is converted into
277 /// for(int i = 0; i < 10000; ++i)
278 ///   bar((double)i);
279 ///
handleFloatingPointIV(Loop * L,PHINode * PN)280 void IndVarSimplify::handleFloatingPointIV(Loop *L, PHINode *PN) {
281   unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
282   unsigned BackEdge     = IncomingEdge^1;
283 
284   // Check incoming value.
285   auto *InitValueVal = dyn_cast<ConstantFP>(PN->getIncomingValue(IncomingEdge));
286 
287   int64_t InitValue;
288   if (!InitValueVal || !ConvertToSInt(InitValueVal->getValueAPF(), InitValue))
289     return;
290 
291   // Check IV increment. Reject this PN if increment operation is not
292   // an add or increment value can not be represented by an integer.
293   auto *Incr = dyn_cast<BinaryOperator>(PN->getIncomingValue(BackEdge));
294   if (Incr == nullptr || Incr->getOpcode() != Instruction::FAdd) return;
295 
296   // If this is not an add of the PHI with a constantfp, or if the constant fp
297   // is not an integer, bail out.
298   ConstantFP *IncValueVal = dyn_cast<ConstantFP>(Incr->getOperand(1));
299   int64_t IncValue;
300   if (IncValueVal == nullptr || Incr->getOperand(0) != PN ||
301       !ConvertToSInt(IncValueVal->getValueAPF(), IncValue))
302     return;
303 
304   // Check Incr uses. One user is PN and the other user is an exit condition
305   // used by the conditional terminator.
306   Value::user_iterator IncrUse = Incr->user_begin();
307   Instruction *U1 = cast<Instruction>(*IncrUse++);
308   if (IncrUse == Incr->user_end()) return;
309   Instruction *U2 = cast<Instruction>(*IncrUse++);
310   if (IncrUse != Incr->user_end()) return;
311 
312   // Find exit condition, which is an fcmp.  If it doesn't exist, or if it isn't
313   // only used by a branch, we can't transform it.
314   FCmpInst *Compare = dyn_cast<FCmpInst>(U1);
315   if (!Compare)
316     Compare = dyn_cast<FCmpInst>(U2);
317   if (!Compare || !Compare->hasOneUse() ||
318       !isa<BranchInst>(Compare->user_back()))
319     return;
320 
321   BranchInst *TheBr = cast<BranchInst>(Compare->user_back());
322 
323   // We need to verify that the branch actually controls the iteration count
324   // of the loop.  If not, the new IV can overflow and no one will notice.
325   // The branch block must be in the loop and one of the successors must be out
326   // of the loop.
327   assert(TheBr->isConditional() && "Can't use fcmp if not conditional");
328   if (!L->contains(TheBr->getParent()) ||
329       (L->contains(TheBr->getSuccessor(0)) &&
330        L->contains(TheBr->getSuccessor(1))))
331     return;
332 
333 
334   // If it isn't a comparison with an integer-as-fp (the exit value), we can't
335   // transform it.
336   ConstantFP *ExitValueVal = dyn_cast<ConstantFP>(Compare->getOperand(1));
337   int64_t ExitValue;
338   if (ExitValueVal == nullptr ||
339       !ConvertToSInt(ExitValueVal->getValueAPF(), ExitValue))
340     return;
341 
342   // Find new predicate for integer comparison.
343   CmpInst::Predicate NewPred = CmpInst::BAD_ICMP_PREDICATE;
344   switch (Compare->getPredicate()) {
345   default: return;  // Unknown comparison.
346   case CmpInst::FCMP_OEQ:
347   case CmpInst::FCMP_UEQ: NewPred = CmpInst::ICMP_EQ; break;
348   case CmpInst::FCMP_ONE:
349   case CmpInst::FCMP_UNE: NewPred = CmpInst::ICMP_NE; break;
350   case CmpInst::FCMP_OGT:
351   case CmpInst::FCMP_UGT: NewPred = CmpInst::ICMP_SGT; break;
352   case CmpInst::FCMP_OGE:
353   case CmpInst::FCMP_UGE: NewPred = CmpInst::ICMP_SGE; break;
354   case CmpInst::FCMP_OLT:
355   case CmpInst::FCMP_ULT: NewPred = CmpInst::ICMP_SLT; break;
356   case CmpInst::FCMP_OLE:
357   case CmpInst::FCMP_ULE: NewPred = CmpInst::ICMP_SLE; break;
358   }
359 
360   // We convert the floating point induction variable to a signed i32 value if
361   // we can.  This is only safe if the comparison will not overflow in a way
362   // that won't be trapped by the integer equivalent operations.  Check for this
363   // now.
364   // TODO: We could use i64 if it is native and the range requires it.
365 
366   // The start/stride/exit values must all fit in signed i32.
367   if (!isInt<32>(InitValue) || !isInt<32>(IncValue) || !isInt<32>(ExitValue))
368     return;
369 
370   // If not actually striding (add x, 0.0), avoid touching the code.
371   if (IncValue == 0)
372     return;
373 
374   // Positive and negative strides have different safety conditions.
375   if (IncValue > 0) {
376     // If we have a positive stride, we require the init to be less than the
377     // exit value.
378     if (InitValue >= ExitValue)
379       return;
380 
381     uint32_t Range = uint32_t(ExitValue-InitValue);
382     // Check for infinite loop, either:
383     // while (i <= Exit) or until (i > Exit)
384     if (NewPred == CmpInst::ICMP_SLE || NewPred == CmpInst::ICMP_SGT) {
385       if (++Range == 0) return;  // Range overflows.
386     }
387 
388     unsigned Leftover = Range % uint32_t(IncValue);
389 
390     // If this is an equality comparison, we require that the strided value
391     // exactly land on the exit value, otherwise the IV condition will wrap
392     // around and do things the fp IV wouldn't.
393     if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
394         Leftover != 0)
395       return;
396 
397     // If the stride would wrap around the i32 before exiting, we can't
398     // transform the IV.
399     if (Leftover != 0 && int32_t(ExitValue+IncValue) < ExitValue)
400       return;
401 
402   } else {
403     // If we have a negative stride, we require the init to be greater than the
404     // exit value.
405     if (InitValue <= ExitValue)
406       return;
407 
408     uint32_t Range = uint32_t(InitValue-ExitValue);
409     // Check for infinite loop, either:
410     // while (i >= Exit) or until (i < Exit)
411     if (NewPred == CmpInst::ICMP_SGE || NewPred == CmpInst::ICMP_SLT) {
412       if (++Range == 0) return;  // Range overflows.
413     }
414 
415     unsigned Leftover = Range % uint32_t(-IncValue);
416 
417     // If this is an equality comparison, we require that the strided value
418     // exactly land on the exit value, otherwise the IV condition will wrap
419     // around and do things the fp IV wouldn't.
420     if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
421         Leftover != 0)
422       return;
423 
424     // If the stride would wrap around the i32 before exiting, we can't
425     // transform the IV.
426     if (Leftover != 0 && int32_t(ExitValue+IncValue) > ExitValue)
427       return;
428   }
429 
430   IntegerType *Int32Ty = Type::getInt32Ty(PN->getContext());
431 
432   // Insert new integer induction variable.
433   PHINode *NewPHI = PHINode::Create(Int32Ty, 2, PN->getName()+".int", PN);
434   NewPHI->addIncoming(ConstantInt::get(Int32Ty, InitValue),
435                       PN->getIncomingBlock(IncomingEdge));
436 
437   Value *NewAdd =
438     BinaryOperator::CreateAdd(NewPHI, ConstantInt::get(Int32Ty, IncValue),
439                               Incr->getName()+".int", Incr);
440   NewPHI->addIncoming(NewAdd, PN->getIncomingBlock(BackEdge));
441 
442   ICmpInst *NewCompare = new ICmpInst(TheBr, NewPred, NewAdd,
443                                       ConstantInt::get(Int32Ty, ExitValue),
444                                       Compare->getName());
445 
446   // In the following deletions, PN may become dead and may be deleted.
447   // Use a WeakVH to observe whether this happens.
448   WeakVH WeakPH = PN;
449 
450   // Delete the old floating point exit comparison.  The branch starts using the
451   // new comparison.
452   NewCompare->takeName(Compare);
453   Compare->replaceAllUsesWith(NewCompare);
454   RecursivelyDeleteTriviallyDeadInstructions(Compare, TLI);
455 
456   // Delete the old floating point increment.
457   Incr->replaceAllUsesWith(UndefValue::get(Incr->getType()));
458   RecursivelyDeleteTriviallyDeadInstructions(Incr, TLI);
459 
460   // If the FP induction variable still has uses, this is because something else
461   // in the loop uses its value.  In order to canonicalize the induction
462   // variable, we chose to eliminate the IV and rewrite it in terms of an
463   // int->fp cast.
464   //
465   // We give preference to sitofp over uitofp because it is faster on most
466   // platforms.
467   if (WeakPH) {
468     Value *Conv = new SIToFPInst(NewPHI, PN->getType(), "indvar.conv",
469                                  &*PN->getParent()->getFirstInsertionPt());
470     PN->replaceAllUsesWith(Conv);
471     RecursivelyDeleteTriviallyDeadInstructions(PN, TLI);
472   }
473   Changed = true;
474 }
475 
rewriteNonIntegerIVs(Loop * L)476 void IndVarSimplify::rewriteNonIntegerIVs(Loop *L) {
477   // First step.  Check to see if there are any floating-point recurrences.
478   // If there are, change them into integer recurrences, permitting analysis by
479   // the SCEV routines.
480   //
481   BasicBlock *Header = L->getHeader();
482 
483   SmallVector<WeakVH, 8> PHIs;
484   for (BasicBlock::iterator I = Header->begin();
485        PHINode *PN = dyn_cast<PHINode>(I); ++I)
486     PHIs.push_back(PN);
487 
488   for (unsigned i = 0, e = PHIs.size(); i != e; ++i)
489     if (PHINode *PN = dyn_cast_or_null<PHINode>(&*PHIs[i]))
490       handleFloatingPointIV(L, PN);
491 
492   // If the loop previously had floating-point IV, ScalarEvolution
493   // may not have been able to compute a trip count. Now that we've done some
494   // re-writing, the trip count may be computable.
495   if (Changed)
496     SE->forgetLoop(L);
497 }
498 
499 namespace {
500 // Collect information about PHI nodes which can be transformed in
501 // rewriteLoopExitValues.
502 struct RewritePhi {
503   PHINode *PN;
504   unsigned Ith;  // Ith incoming value.
505   Value *Val;    // Exit value after expansion.
506   bool HighCost; // High Cost when expansion.
507   bool SafePhi;  // LCSSASafePhiForRAUW.
508 
RewritePhi__anon4bb1d1520211::RewritePhi509   RewritePhi(PHINode *P, unsigned I, Value *V, bool H, bool S)
510       : PN(P), Ith(I), Val(V), HighCost(H), SafePhi(S) {}
511 };
512 }
513 
expandSCEVIfNeeded(SCEVExpander & Rewriter,const SCEV * S,Loop * L,Instruction * InsertPt,Type * ResultTy)514 Value *IndVarSimplify::expandSCEVIfNeeded(SCEVExpander &Rewriter, const SCEV *S,
515                                           Loop *L, Instruction *InsertPt,
516                                           Type *ResultTy) {
517   // Before expanding S into an expensive LLVM expression, see if we can use an
518   // already existing value as the expansion for S.
519   if (Value *ExistingValue = Rewriter.findExistingExpansion(S, InsertPt, L))
520     if (ExistingValue->getType() == ResultTy)
521       return ExistingValue;
522 
523   // We didn't find anything, fall back to using SCEVExpander.
524   return Rewriter.expandCodeFor(S, ResultTy, InsertPt);
525 }
526 
527 //===----------------------------------------------------------------------===//
528 // rewriteLoopExitValues - Optimize IV users outside the loop.
529 // As a side effect, reduces the amount of IV processing within the loop.
530 //===----------------------------------------------------------------------===//
531 
532 /// Check to see if this loop has a computable loop-invariant execution count.
533 /// If so, this means that we can compute the final value of any expressions
534 /// that are recurrent in the loop, and substitute the exit values from the loop
535 /// into any instructions outside of the loop that use the final values of the
536 /// current expressions.
537 ///
538 /// This is mostly redundant with the regular IndVarSimplify activities that
539 /// happen later, except that it's more powerful in some cases, because it's
540 /// able to brute-force evaluate arbitrary instructions as long as they have
541 /// constant operands at the beginning of the loop.
rewriteLoopExitValues(Loop * L,SCEVExpander & Rewriter)542 void IndVarSimplify::rewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter) {
543   // Check a pre-condition.
544   assert(L->isRecursivelyLCSSAForm(*DT) && "Indvars did not preserve LCSSA!");
545 
546   SmallVector<BasicBlock*, 8> ExitBlocks;
547   L->getUniqueExitBlocks(ExitBlocks);
548 
549   SmallVector<RewritePhi, 8> RewritePhiSet;
550   // Find all values that are computed inside the loop, but used outside of it.
551   // Because of LCSSA, these values will only occur in LCSSA PHI Nodes.  Scan
552   // the exit blocks of the loop to find them.
553   for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
554     BasicBlock *ExitBB = ExitBlocks[i];
555 
556     // If there are no PHI nodes in this exit block, then no values defined
557     // inside the loop are used on this path, skip it.
558     PHINode *PN = dyn_cast<PHINode>(ExitBB->begin());
559     if (!PN) continue;
560 
561     unsigned NumPreds = PN->getNumIncomingValues();
562 
563     // We would like to be able to RAUW single-incoming value PHI nodes. We
564     // have to be certain this is safe even when this is an LCSSA PHI node.
565     // While the computed exit value is no longer varying in *this* loop, the
566     // exit block may be an exit block for an outer containing loop as well,
567     // the exit value may be varying in the outer loop, and thus it may still
568     // require an LCSSA PHI node. The safe case is when this is
569     // single-predecessor PHI node (LCSSA) and the exit block containing it is
570     // part of the enclosing loop, or this is the outer most loop of the nest.
571     // In either case the exit value could (at most) be varying in the same
572     // loop body as the phi node itself. Thus if it is in turn used outside of
573     // an enclosing loop it will only be via a separate LCSSA node.
574     bool LCSSASafePhiForRAUW =
575         NumPreds == 1 &&
576         (!L->getParentLoop() || L->getParentLoop() == LI->getLoopFor(ExitBB));
577 
578     // Iterate over all of the PHI nodes.
579     BasicBlock::iterator BBI = ExitBB->begin();
580     while ((PN = dyn_cast<PHINode>(BBI++))) {
581       if (PN->use_empty())
582         continue; // dead use, don't replace it
583 
584       // SCEV only supports integer expressions for now.
585       if (!PN->getType()->isIntegerTy() && !PN->getType()->isPointerTy())
586         continue;
587 
588       // It's necessary to tell ScalarEvolution about this explicitly so that
589       // it can walk the def-use list and forget all SCEVs, as it may not be
590       // watching the PHI itself. Once the new exit value is in place, there
591       // may not be a def-use connection between the loop and every instruction
592       // which got a SCEVAddRecExpr for that loop.
593       SE->forgetValue(PN);
594 
595       // Iterate over all of the values in all the PHI nodes.
596       for (unsigned i = 0; i != NumPreds; ++i) {
597         // If the value being merged in is not integer or is not defined
598         // in the loop, skip it.
599         Value *InVal = PN->getIncomingValue(i);
600         if (!isa<Instruction>(InVal))
601           continue;
602 
603         // If this pred is for a subloop, not L itself, skip it.
604         if (LI->getLoopFor(PN->getIncomingBlock(i)) != L)
605           continue; // The Block is in a subloop, skip it.
606 
607         // Check that InVal is defined in the loop.
608         Instruction *Inst = cast<Instruction>(InVal);
609         if (!L->contains(Inst))
610           continue;
611 
612         // Okay, this instruction has a user outside of the current loop
613         // and varies predictably *inside* the loop.  Evaluate the value it
614         // contains when the loop exits, if possible.
615         const SCEV *ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop());
616         if (!SE->isLoopInvariant(ExitValue, L) ||
617             !isSafeToExpand(ExitValue, *SE))
618           continue;
619 
620         // Computing the value outside of the loop brings no benefit if :
621         //  - it is definitely used inside the loop in a way which can not be
622         //    optimized away.
623         //  - no use outside of the loop can take advantage of hoisting the
624         //    computation out of the loop
625         if (ExitValue->getSCEVType()>=scMulExpr) {
626           unsigned NumHardInternalUses = 0;
627           unsigned NumSoftExternalUses = 0;
628           unsigned NumUses = 0;
629           for (auto IB = Inst->user_begin(), IE = Inst->user_end();
630                IB != IE && NumUses <= 6; ++IB) {
631             Instruction *UseInstr = cast<Instruction>(*IB);
632             unsigned Opc = UseInstr->getOpcode();
633             NumUses++;
634             if (L->contains(UseInstr)) {
635               if (Opc == Instruction::Call || Opc == Instruction::Ret)
636                 NumHardInternalUses++;
637             } else {
638               if (Opc == Instruction::PHI) {
639                 // Do not count the Phi as a use. LCSSA may have inserted
640                 // plenty of trivial ones.
641                 NumUses--;
642                 for (auto PB = UseInstr->user_begin(),
643                           PE = UseInstr->user_end();
644                      PB != PE && NumUses <= 6; ++PB, ++NumUses) {
645                   unsigned PhiOpc = cast<Instruction>(*PB)->getOpcode();
646                   if (PhiOpc != Instruction::Call && PhiOpc != Instruction::Ret)
647                     NumSoftExternalUses++;
648                 }
649                 continue;
650               }
651               if (Opc != Instruction::Call && Opc != Instruction::Ret)
652                 NumSoftExternalUses++;
653             }
654           }
655           if (NumUses <= 6 && NumHardInternalUses && !NumSoftExternalUses)
656             continue;
657         }
658 
659         bool HighCost = Rewriter.isHighCostExpansion(ExitValue, L, Inst);
660         Value *ExitVal =
661             expandSCEVIfNeeded(Rewriter, ExitValue, L, Inst, PN->getType());
662 
663         DEBUG(dbgs() << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal << '\n'
664                      << "  LoopVal = " << *Inst << "\n");
665 
666         if (!isValidRewrite(Inst, ExitVal)) {
667           DeadInsts.push_back(ExitVal);
668           continue;
669         }
670 
671         // Collect all the candidate PHINodes to be rewritten.
672         RewritePhiSet.push_back(
673             RewritePhi(PN, i, ExitVal, HighCost, LCSSASafePhiForRAUW));
674       }
675     }
676   }
677 
678   bool LoopCanBeDel = canLoopBeDeleted(L, RewritePhiSet);
679 
680   // Transformation.
681   for (const RewritePhi &Phi : RewritePhiSet) {
682     PHINode *PN = Phi.PN;
683     Value *ExitVal = Phi.Val;
684 
685     // Only do the rewrite when the ExitValue can be expanded cheaply.
686     // If LoopCanBeDel is true, rewrite exit value aggressively.
687     if (ReplaceExitValue == OnlyCheapRepl && !LoopCanBeDel && Phi.HighCost) {
688       DeadInsts.push_back(ExitVal);
689       continue;
690     }
691 
692     Changed = true;
693     ++NumReplaced;
694     Instruction *Inst = cast<Instruction>(PN->getIncomingValue(Phi.Ith));
695     PN->setIncomingValue(Phi.Ith, ExitVal);
696 
697     // If this instruction is dead now, delete it. Don't do it now to avoid
698     // invalidating iterators.
699     if (isInstructionTriviallyDead(Inst, TLI))
700       DeadInsts.push_back(Inst);
701 
702     // If we determined that this PHI is safe to replace even if an LCSSA
703     // PHI, do so.
704     if (Phi.SafePhi) {
705       PN->replaceAllUsesWith(ExitVal);
706       PN->eraseFromParent();
707     }
708   }
709 
710   // The insertion point instruction may have been deleted; clear it out
711   // so that the rewriter doesn't trip over it later.
712   Rewriter.clearInsertPoint();
713 }
714 
715 /// Check whether it is possible to delete the loop after rewriting exit
716 /// value. If it is possible, ignore ReplaceExitValue and do rewriting
717 /// aggressively.
canLoopBeDeleted(Loop * L,SmallVector<RewritePhi,8> & RewritePhiSet)718 bool IndVarSimplify::canLoopBeDeleted(
719     Loop *L, SmallVector<RewritePhi, 8> &RewritePhiSet) {
720 
721   BasicBlock *Preheader = L->getLoopPreheader();
722   // If there is no preheader, the loop will not be deleted.
723   if (!Preheader)
724     return false;
725 
726   // In LoopDeletion pass Loop can be deleted when ExitingBlocks.size() > 1.
727   // We obviate multiple ExitingBlocks case for simplicity.
728   // TODO: If we see testcase with multiple ExitingBlocks can be deleted
729   // after exit value rewriting, we can enhance the logic here.
730   SmallVector<BasicBlock *, 4> ExitingBlocks;
731   L->getExitingBlocks(ExitingBlocks);
732   SmallVector<BasicBlock *, 8> ExitBlocks;
733   L->getUniqueExitBlocks(ExitBlocks);
734   if (ExitBlocks.size() > 1 || ExitingBlocks.size() > 1)
735     return false;
736 
737   BasicBlock *ExitBlock = ExitBlocks[0];
738   BasicBlock::iterator BI = ExitBlock->begin();
739   while (PHINode *P = dyn_cast<PHINode>(BI)) {
740     Value *Incoming = P->getIncomingValueForBlock(ExitingBlocks[0]);
741 
742     // If the Incoming value of P is found in RewritePhiSet, we know it
743     // could be rewritten to use a loop invariant value in transformation
744     // phase later. Skip it in the loop invariant check below.
745     bool found = false;
746     for (const RewritePhi &Phi : RewritePhiSet) {
747       unsigned i = Phi.Ith;
748       if (Phi.PN == P && (Phi.PN)->getIncomingValue(i) == Incoming) {
749         found = true;
750         break;
751       }
752     }
753 
754     Instruction *I;
755     if (!found && (I = dyn_cast<Instruction>(Incoming)))
756       if (!L->hasLoopInvariantOperands(I))
757         return false;
758 
759     ++BI;
760   }
761 
762   for (auto *BB : L->blocks())
763     if (any_of(*BB, [](Instruction &I) { return I.mayHaveSideEffects(); }))
764       return false;
765 
766   return true;
767 }
768 
769 //===----------------------------------------------------------------------===//
770 //  IV Widening - Extend the width of an IV to cover its widest uses.
771 //===----------------------------------------------------------------------===//
772 
773 namespace {
774 // Collect information about induction variables that are used by sign/zero
775 // extend operations. This information is recorded by CollectExtend and provides
776 // the input to WidenIV.
777 struct WideIVInfo {
778   PHINode *NarrowIV = nullptr;
779   Type *WidestNativeType = nullptr; // Widest integer type created [sz]ext
780   bool IsSigned = false;            // Was a sext user seen before a zext?
781 };
782 }
783 
784 /// Update information about the induction variable that is extended by this
785 /// sign or zero extend operation. This is used to determine the final width of
786 /// the IV before actually widening it.
visitIVCast(CastInst * Cast,WideIVInfo & WI,ScalarEvolution * SE,const TargetTransformInfo * TTI)787 static void visitIVCast(CastInst *Cast, WideIVInfo &WI, ScalarEvolution *SE,
788                         const TargetTransformInfo *TTI) {
789   bool IsSigned = Cast->getOpcode() == Instruction::SExt;
790   if (!IsSigned && Cast->getOpcode() != Instruction::ZExt)
791     return;
792 
793   Type *Ty = Cast->getType();
794   uint64_t Width = SE->getTypeSizeInBits(Ty);
795   if (!Cast->getModule()->getDataLayout().isLegalInteger(Width))
796     return;
797 
798   // Cast is either an sext or zext up to this point.
799   // We should not widen an indvar if arithmetics on the wider indvar are more
800   // expensive than those on the narrower indvar. We check only the cost of ADD
801   // because at least an ADD is required to increment the induction variable. We
802   // could compute more comprehensively the cost of all instructions on the
803   // induction variable when necessary.
804   if (TTI &&
805       TTI->getArithmeticInstrCost(Instruction::Add, Ty) >
806           TTI->getArithmeticInstrCost(Instruction::Add,
807                                       Cast->getOperand(0)->getType())) {
808     return;
809   }
810 
811   if (!WI.WidestNativeType) {
812     WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
813     WI.IsSigned = IsSigned;
814     return;
815   }
816 
817   // We extend the IV to satisfy the sign of its first user, arbitrarily.
818   if (WI.IsSigned != IsSigned)
819     return;
820 
821   if (Width > SE->getTypeSizeInBits(WI.WidestNativeType))
822     WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
823 }
824 
825 namespace {
826 
827 /// Record a link in the Narrow IV def-use chain along with the WideIV that
828 /// computes the same value as the Narrow IV def.  This avoids caching Use*
829 /// pointers.
830 struct NarrowIVDefUse {
831   Instruction *NarrowDef = nullptr;
832   Instruction *NarrowUse = nullptr;
833   Instruction *WideDef = nullptr;
834 
835   // True if the narrow def is never negative.  Tracking this information lets
836   // us use a sign extension instead of a zero extension or vice versa, when
837   // profitable and legal.
838   bool NeverNegative = false;
839 
NarrowIVDefUse__anon4bb1d1520511::NarrowIVDefUse840   NarrowIVDefUse(Instruction *ND, Instruction *NU, Instruction *WD,
841                  bool NeverNegative)
842       : NarrowDef(ND), NarrowUse(NU), WideDef(WD),
843         NeverNegative(NeverNegative) {}
844 };
845 
846 /// The goal of this transform is to remove sign and zero extends without
847 /// creating any new induction variables. To do this, it creates a new phi of
848 /// the wider type and redirects all users, either removing extends or inserting
849 /// truncs whenever we stop propagating the type.
850 ///
851 class WidenIV {
852   // Parameters
853   PHINode *OrigPhi;
854   Type *WideType;
855   bool IsSigned;
856 
857   // Context
858   LoopInfo        *LI;
859   Loop            *L;
860   ScalarEvolution *SE;
861   DominatorTree   *DT;
862 
863   // Result
864   PHINode *WidePhi;
865   Instruction *WideInc;
866   const SCEV *WideIncExpr;
867   SmallVectorImpl<WeakVH> &DeadInsts;
868 
869   SmallPtrSet<Instruction*,16> Widened;
870   SmallVector<NarrowIVDefUse, 8> NarrowIVUsers;
871 
872 public:
WidenIV(const WideIVInfo & WI,LoopInfo * LInfo,ScalarEvolution * SEv,DominatorTree * DTree,SmallVectorImpl<WeakVH> & DI)873   WidenIV(const WideIVInfo &WI, LoopInfo *LInfo,
874           ScalarEvolution *SEv, DominatorTree *DTree,
875           SmallVectorImpl<WeakVH> &DI) :
876     OrigPhi(WI.NarrowIV),
877     WideType(WI.WidestNativeType),
878     IsSigned(WI.IsSigned),
879     LI(LInfo),
880     L(LI->getLoopFor(OrigPhi->getParent())),
881     SE(SEv),
882     DT(DTree),
883     WidePhi(nullptr),
884     WideInc(nullptr),
885     WideIncExpr(nullptr),
886     DeadInsts(DI) {
887     assert(L->getHeader() == OrigPhi->getParent() && "Phi must be an IV");
888   }
889 
890   PHINode *createWideIV(SCEVExpander &Rewriter);
891 
892 protected:
893   Value *createExtendInst(Value *NarrowOper, Type *WideType, bool IsSigned,
894                           Instruction *Use);
895 
896   Instruction *cloneIVUser(NarrowIVDefUse DU, const SCEVAddRecExpr *WideAR);
897   Instruction *cloneArithmeticIVUser(NarrowIVDefUse DU,
898                                      const SCEVAddRecExpr *WideAR);
899   Instruction *cloneBitwiseIVUser(NarrowIVDefUse DU);
900 
901   const SCEVAddRecExpr *getWideRecurrence(Instruction *NarrowUse);
902 
903   const SCEVAddRecExpr* getExtendedOperandRecurrence(NarrowIVDefUse DU);
904 
905   const SCEV *getSCEVByOpCode(const SCEV *LHS, const SCEV *RHS,
906                               unsigned OpCode) const;
907 
908   Instruction *widenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter);
909 
910   bool widenLoopCompare(NarrowIVDefUse DU);
911 
912   void pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef);
913 };
914 } // anonymous namespace
915 
916 /// Perform a quick domtree based check for loop invariance assuming that V is
917 /// used within the loop. LoopInfo::isLoopInvariant() seems gratuitous for this
918 /// purpose.
isLoopInvariant(Value * V,const Loop * L,const DominatorTree * DT)919 static bool isLoopInvariant(Value *V, const Loop *L, const DominatorTree *DT) {
920   Instruction *Inst = dyn_cast<Instruction>(V);
921   if (!Inst)
922     return true;
923 
924   return DT->properlyDominates(Inst->getParent(), L->getHeader());
925 }
926 
createExtendInst(Value * NarrowOper,Type * WideType,bool IsSigned,Instruction * Use)927 Value *WidenIV::createExtendInst(Value *NarrowOper, Type *WideType,
928                                  bool IsSigned, Instruction *Use) {
929   // Set the debug location and conservative insertion point.
930   IRBuilder<> Builder(Use);
931   // Hoist the insertion point into loop preheaders as far as possible.
932   for (const Loop *L = LI->getLoopFor(Use->getParent());
933        L && L->getLoopPreheader() && isLoopInvariant(NarrowOper, L, DT);
934        L = L->getParentLoop())
935     Builder.SetInsertPoint(L->getLoopPreheader()->getTerminator());
936 
937   return IsSigned ? Builder.CreateSExt(NarrowOper, WideType) :
938                     Builder.CreateZExt(NarrowOper, WideType);
939 }
940 
941 /// Instantiate a wide operation to replace a narrow operation. This only needs
942 /// to handle operations that can evaluation to SCEVAddRec. It can safely return
943 /// 0 for any operation we decide not to clone.
cloneIVUser(NarrowIVDefUse DU,const SCEVAddRecExpr * WideAR)944 Instruction *WidenIV::cloneIVUser(NarrowIVDefUse DU,
945                                   const SCEVAddRecExpr *WideAR) {
946   unsigned Opcode = DU.NarrowUse->getOpcode();
947   switch (Opcode) {
948   default:
949     return nullptr;
950   case Instruction::Add:
951   case Instruction::Mul:
952   case Instruction::UDiv:
953   case Instruction::Sub:
954     return cloneArithmeticIVUser(DU, WideAR);
955 
956   case Instruction::And:
957   case Instruction::Or:
958   case Instruction::Xor:
959   case Instruction::Shl:
960   case Instruction::LShr:
961   case Instruction::AShr:
962     return cloneBitwiseIVUser(DU);
963   }
964 }
965 
cloneBitwiseIVUser(NarrowIVDefUse DU)966 Instruction *WidenIV::cloneBitwiseIVUser(NarrowIVDefUse DU) {
967   Instruction *NarrowUse = DU.NarrowUse;
968   Instruction *NarrowDef = DU.NarrowDef;
969   Instruction *WideDef = DU.WideDef;
970 
971   DEBUG(dbgs() << "Cloning bitwise IVUser: " << *NarrowUse << "\n");
972 
973   // Replace NarrowDef operands with WideDef. Otherwise, we don't know anything
974   // about the narrow operand yet so must insert a [sz]ext. It is probably loop
975   // invariant and will be folded or hoisted. If it actually comes from a
976   // widened IV, it should be removed during a future call to widenIVUse.
977   Value *LHS = (NarrowUse->getOperand(0) == NarrowDef)
978                    ? WideDef
979                    : createExtendInst(NarrowUse->getOperand(0), WideType,
980                                       IsSigned, NarrowUse);
981   Value *RHS = (NarrowUse->getOperand(1) == NarrowDef)
982                    ? WideDef
983                    : createExtendInst(NarrowUse->getOperand(1), WideType,
984                                       IsSigned, NarrowUse);
985 
986   auto *NarrowBO = cast<BinaryOperator>(NarrowUse);
987   auto *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(), LHS, RHS,
988                                         NarrowBO->getName());
989   IRBuilder<> Builder(NarrowUse);
990   Builder.Insert(WideBO);
991   WideBO->copyIRFlags(NarrowBO);
992   return WideBO;
993 }
994 
cloneArithmeticIVUser(NarrowIVDefUse DU,const SCEVAddRecExpr * WideAR)995 Instruction *WidenIV::cloneArithmeticIVUser(NarrowIVDefUse DU,
996                                             const SCEVAddRecExpr *WideAR) {
997   Instruction *NarrowUse = DU.NarrowUse;
998   Instruction *NarrowDef = DU.NarrowDef;
999   Instruction *WideDef = DU.WideDef;
1000 
1001   DEBUG(dbgs() << "Cloning arithmetic IVUser: " << *NarrowUse << "\n");
1002 
1003   unsigned IVOpIdx = (NarrowUse->getOperand(0) == NarrowDef) ? 0 : 1;
1004 
1005   // We're trying to find X such that
1006   //
1007   //  Widen(NarrowDef `op` NonIVNarrowDef) == WideAR == WideDef `op.wide` X
1008   //
1009   // We guess two solutions to X, sext(NonIVNarrowDef) and zext(NonIVNarrowDef),
1010   // and check using SCEV if any of them are correct.
1011 
1012   // Returns true if extending NonIVNarrowDef according to `SignExt` is a
1013   // correct solution to X.
1014   auto GuessNonIVOperand = [&](bool SignExt) {
1015     const SCEV *WideLHS;
1016     const SCEV *WideRHS;
1017 
1018     auto GetExtend = [this, SignExt](const SCEV *S, Type *Ty) {
1019       if (SignExt)
1020         return SE->getSignExtendExpr(S, Ty);
1021       return SE->getZeroExtendExpr(S, Ty);
1022     };
1023 
1024     if (IVOpIdx == 0) {
1025       WideLHS = SE->getSCEV(WideDef);
1026       const SCEV *NarrowRHS = SE->getSCEV(NarrowUse->getOperand(1));
1027       WideRHS = GetExtend(NarrowRHS, WideType);
1028     } else {
1029       const SCEV *NarrowLHS = SE->getSCEV(NarrowUse->getOperand(0));
1030       WideLHS = GetExtend(NarrowLHS, WideType);
1031       WideRHS = SE->getSCEV(WideDef);
1032     }
1033 
1034     // WideUse is "WideDef `op.wide` X" as described in the comment.
1035     const SCEV *WideUse = nullptr;
1036 
1037     switch (NarrowUse->getOpcode()) {
1038     default:
1039       llvm_unreachable("No other possibility!");
1040 
1041     case Instruction::Add:
1042       WideUse = SE->getAddExpr(WideLHS, WideRHS);
1043       break;
1044 
1045     case Instruction::Mul:
1046       WideUse = SE->getMulExpr(WideLHS, WideRHS);
1047       break;
1048 
1049     case Instruction::UDiv:
1050       WideUse = SE->getUDivExpr(WideLHS, WideRHS);
1051       break;
1052 
1053     case Instruction::Sub:
1054       WideUse = SE->getMinusSCEV(WideLHS, WideRHS);
1055       break;
1056     }
1057 
1058     return WideUse == WideAR;
1059   };
1060 
1061   bool SignExtend = IsSigned;
1062   if (!GuessNonIVOperand(SignExtend)) {
1063     SignExtend = !SignExtend;
1064     if (!GuessNonIVOperand(SignExtend))
1065       return nullptr;
1066   }
1067 
1068   Value *LHS = (NarrowUse->getOperand(0) == NarrowDef)
1069                    ? WideDef
1070                    : createExtendInst(NarrowUse->getOperand(0), WideType,
1071                                       SignExtend, NarrowUse);
1072   Value *RHS = (NarrowUse->getOperand(1) == NarrowDef)
1073                    ? WideDef
1074                    : createExtendInst(NarrowUse->getOperand(1), WideType,
1075                                       SignExtend, NarrowUse);
1076 
1077   auto *NarrowBO = cast<BinaryOperator>(NarrowUse);
1078   auto *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(), LHS, RHS,
1079                                         NarrowBO->getName());
1080 
1081   IRBuilder<> Builder(NarrowUse);
1082   Builder.Insert(WideBO);
1083   WideBO->copyIRFlags(NarrowBO);
1084   return WideBO;
1085 }
1086 
getSCEVByOpCode(const SCEV * LHS,const SCEV * RHS,unsigned OpCode) const1087 const SCEV *WidenIV::getSCEVByOpCode(const SCEV *LHS, const SCEV *RHS,
1088                                      unsigned OpCode) const {
1089   if (OpCode == Instruction::Add)
1090     return SE->getAddExpr(LHS, RHS);
1091   if (OpCode == Instruction::Sub)
1092     return SE->getMinusSCEV(LHS, RHS);
1093   if (OpCode == Instruction::Mul)
1094     return SE->getMulExpr(LHS, RHS);
1095 
1096   llvm_unreachable("Unsupported opcode.");
1097 }
1098 
1099 /// No-wrap operations can transfer sign extension of their result to their
1100 /// operands. Generate the SCEV value for the widened operation without
1101 /// actually modifying the IR yet. If the expression after extending the
1102 /// operands is an AddRec for this loop, return it.
getExtendedOperandRecurrence(NarrowIVDefUse DU)1103 const SCEVAddRecExpr* WidenIV::getExtendedOperandRecurrence(NarrowIVDefUse DU) {
1104 
1105   // Handle the common case of add<nsw/nuw>
1106   const unsigned OpCode = DU.NarrowUse->getOpcode();
1107   // Only Add/Sub/Mul instructions supported yet.
1108   if (OpCode != Instruction::Add && OpCode != Instruction::Sub &&
1109       OpCode != Instruction::Mul)
1110     return nullptr;
1111 
1112   // One operand (NarrowDef) has already been extended to WideDef. Now determine
1113   // if extending the other will lead to a recurrence.
1114   const unsigned ExtendOperIdx =
1115       DU.NarrowUse->getOperand(0) == DU.NarrowDef ? 1 : 0;
1116   assert(DU.NarrowUse->getOperand(1-ExtendOperIdx) == DU.NarrowDef && "bad DU");
1117 
1118   const SCEV *ExtendOperExpr = nullptr;
1119   const OverflowingBinaryOperator *OBO =
1120     cast<OverflowingBinaryOperator>(DU.NarrowUse);
1121   if (IsSigned && OBO->hasNoSignedWrap())
1122     ExtendOperExpr = SE->getSignExtendExpr(
1123       SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType);
1124   else if(!IsSigned && OBO->hasNoUnsignedWrap())
1125     ExtendOperExpr = SE->getZeroExtendExpr(
1126       SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType);
1127   else
1128     return nullptr;
1129 
1130   // When creating this SCEV expr, don't apply the current operations NSW or NUW
1131   // flags. This instruction may be guarded by control flow that the no-wrap
1132   // behavior depends on. Non-control-equivalent instructions can be mapped to
1133   // the same SCEV expression, and it would be incorrect to transfer NSW/NUW
1134   // semantics to those operations.
1135   const SCEV *lhs = SE->getSCEV(DU.WideDef);
1136   const SCEV *rhs = ExtendOperExpr;
1137 
1138   // Let's swap operands to the initial order for the case of non-commutative
1139   // operations, like SUB. See PR21014.
1140   if (ExtendOperIdx == 0)
1141     std::swap(lhs, rhs);
1142   const SCEVAddRecExpr *AddRec =
1143       dyn_cast<SCEVAddRecExpr>(getSCEVByOpCode(lhs, rhs, OpCode));
1144 
1145   if (!AddRec || AddRec->getLoop() != L)
1146     return nullptr;
1147   return AddRec;
1148 }
1149 
1150 /// Is this instruction potentially interesting for further simplification after
1151 /// widening it's type? In other words, can the extend be safely hoisted out of
1152 /// the loop with SCEV reducing the value to a recurrence on the same loop. If
1153 /// so, return the sign or zero extended recurrence. Otherwise return NULL.
getWideRecurrence(Instruction * NarrowUse)1154 const SCEVAddRecExpr *WidenIV::getWideRecurrence(Instruction *NarrowUse) {
1155   if (!SE->isSCEVable(NarrowUse->getType()))
1156     return nullptr;
1157 
1158   const SCEV *NarrowExpr = SE->getSCEV(NarrowUse);
1159   if (SE->getTypeSizeInBits(NarrowExpr->getType())
1160       >= SE->getTypeSizeInBits(WideType)) {
1161     // NarrowUse implicitly widens its operand. e.g. a gep with a narrow
1162     // index. So don't follow this use.
1163     return nullptr;
1164   }
1165 
1166   const SCEV *WideExpr = IsSigned ?
1167     SE->getSignExtendExpr(NarrowExpr, WideType) :
1168     SE->getZeroExtendExpr(NarrowExpr, WideType);
1169   const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(WideExpr);
1170   if (!AddRec || AddRec->getLoop() != L)
1171     return nullptr;
1172   return AddRec;
1173 }
1174 
1175 /// This IV user cannot be widen. Replace this use of the original narrow IV
1176 /// with a truncation of the new wide IV to isolate and eliminate the narrow IV.
truncateIVUse(NarrowIVDefUse DU,DominatorTree * DT,LoopInfo * LI)1177 static void truncateIVUse(NarrowIVDefUse DU, DominatorTree *DT, LoopInfo *LI) {
1178   DEBUG(dbgs() << "INDVARS: Truncate IV " << *DU.WideDef
1179         << " for user " << *DU.NarrowUse << "\n");
1180   IRBuilder<> Builder(
1181       getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT, LI));
1182   Value *Trunc = Builder.CreateTrunc(DU.WideDef, DU.NarrowDef->getType());
1183   DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, Trunc);
1184 }
1185 
1186 /// If the narrow use is a compare instruction, then widen the compare
1187 //  (and possibly the other operand).  The extend operation is hoisted into the
1188 // loop preheader as far as possible.
widenLoopCompare(NarrowIVDefUse DU)1189 bool WidenIV::widenLoopCompare(NarrowIVDefUse DU) {
1190   ICmpInst *Cmp = dyn_cast<ICmpInst>(DU.NarrowUse);
1191   if (!Cmp)
1192     return false;
1193 
1194   // We can legally widen the comparison in the following two cases:
1195   //
1196   //  - The signedness of the IV extension and comparison match
1197   //
1198   //  - The narrow IV is always positive (and thus its sign extension is equal
1199   //    to its zero extension).  For instance, let's say we're zero extending
1200   //    %narrow for the following use
1201   //
1202   //      icmp slt i32 %narrow, %val   ... (A)
1203   //
1204   //    and %narrow is always positive.  Then
1205   //
1206   //      (A) == icmp slt i32 sext(%narrow), sext(%val)
1207   //          == icmp slt i32 zext(%narrow), sext(%val)
1208 
1209   if (!(DU.NeverNegative || IsSigned == Cmp->isSigned()))
1210     return false;
1211 
1212   Value *Op = Cmp->getOperand(Cmp->getOperand(0) == DU.NarrowDef ? 1 : 0);
1213   unsigned CastWidth = SE->getTypeSizeInBits(Op->getType());
1214   unsigned IVWidth = SE->getTypeSizeInBits(WideType);
1215   assert (CastWidth <= IVWidth && "Unexpected width while widening compare.");
1216 
1217   // Widen the compare instruction.
1218   IRBuilder<> Builder(
1219       getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT, LI));
1220   DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef);
1221 
1222   // Widen the other operand of the compare, if necessary.
1223   if (CastWidth < IVWidth) {
1224     Value *ExtOp = createExtendInst(Op, WideType, Cmp->isSigned(), Cmp);
1225     DU.NarrowUse->replaceUsesOfWith(Op, ExtOp);
1226   }
1227   return true;
1228 }
1229 
1230 /// Determine whether an individual user of the narrow IV can be widened. If so,
1231 /// return the wide clone of the user.
widenIVUse(NarrowIVDefUse DU,SCEVExpander & Rewriter)1232 Instruction *WidenIV::widenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter) {
1233 
1234   // Stop traversing the def-use chain at inner-loop phis or post-loop phis.
1235   if (PHINode *UsePhi = dyn_cast<PHINode>(DU.NarrowUse)) {
1236     if (LI->getLoopFor(UsePhi->getParent()) != L) {
1237       // For LCSSA phis, sink the truncate outside the loop.
1238       // After SimplifyCFG most loop exit targets have a single predecessor.
1239       // Otherwise fall back to a truncate within the loop.
1240       if (UsePhi->getNumOperands() != 1)
1241         truncateIVUse(DU, DT, LI);
1242       else {
1243         PHINode *WidePhi =
1244           PHINode::Create(DU.WideDef->getType(), 1, UsePhi->getName() + ".wide",
1245                           UsePhi);
1246         WidePhi->addIncoming(DU.WideDef, UsePhi->getIncomingBlock(0));
1247         IRBuilder<> Builder(&*WidePhi->getParent()->getFirstInsertionPt());
1248         Value *Trunc = Builder.CreateTrunc(WidePhi, DU.NarrowDef->getType());
1249         UsePhi->replaceAllUsesWith(Trunc);
1250         DeadInsts.emplace_back(UsePhi);
1251         DEBUG(dbgs() << "INDVARS: Widen lcssa phi " << *UsePhi
1252               << " to " << *WidePhi << "\n");
1253       }
1254       return nullptr;
1255     }
1256   }
1257   // Our raison d'etre! Eliminate sign and zero extension.
1258   if (IsSigned ? isa<SExtInst>(DU.NarrowUse) : isa<ZExtInst>(DU.NarrowUse)) {
1259     Value *NewDef = DU.WideDef;
1260     if (DU.NarrowUse->getType() != WideType) {
1261       unsigned CastWidth = SE->getTypeSizeInBits(DU.NarrowUse->getType());
1262       unsigned IVWidth = SE->getTypeSizeInBits(WideType);
1263       if (CastWidth < IVWidth) {
1264         // The cast isn't as wide as the IV, so insert a Trunc.
1265         IRBuilder<> Builder(DU.NarrowUse);
1266         NewDef = Builder.CreateTrunc(DU.WideDef, DU.NarrowUse->getType());
1267       }
1268       else {
1269         // A wider extend was hidden behind a narrower one. This may induce
1270         // another round of IV widening in which the intermediate IV becomes
1271         // dead. It should be very rare.
1272         DEBUG(dbgs() << "INDVARS: New IV " << *WidePhi
1273               << " not wide enough to subsume " << *DU.NarrowUse << "\n");
1274         DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef);
1275         NewDef = DU.NarrowUse;
1276       }
1277     }
1278     if (NewDef != DU.NarrowUse) {
1279       DEBUG(dbgs() << "INDVARS: eliminating " << *DU.NarrowUse
1280             << " replaced by " << *DU.WideDef << "\n");
1281       ++NumElimExt;
1282       DU.NarrowUse->replaceAllUsesWith(NewDef);
1283       DeadInsts.emplace_back(DU.NarrowUse);
1284     }
1285     // Now that the extend is gone, we want to expose it's uses for potential
1286     // further simplification. We don't need to directly inform SimplifyIVUsers
1287     // of the new users, because their parent IV will be processed later as a
1288     // new loop phi. If we preserved IVUsers analysis, we would also want to
1289     // push the uses of WideDef here.
1290 
1291     // No further widening is needed. The deceased [sz]ext had done it for us.
1292     return nullptr;
1293   }
1294 
1295   // Does this user itself evaluate to a recurrence after widening?
1296   const SCEVAddRecExpr *WideAddRec = getWideRecurrence(DU.NarrowUse);
1297   if (!WideAddRec)
1298     WideAddRec = getExtendedOperandRecurrence(DU);
1299 
1300   if (!WideAddRec) {
1301     // If use is a loop condition, try to promote the condition instead of
1302     // truncating the IV first.
1303     if (widenLoopCompare(DU))
1304       return nullptr;
1305 
1306     // This user does not evaluate to a recurence after widening, so don't
1307     // follow it. Instead insert a Trunc to kill off the original use,
1308     // eventually isolating the original narrow IV so it can be removed.
1309     truncateIVUse(DU, DT, LI);
1310     return nullptr;
1311   }
1312   // Assume block terminators cannot evaluate to a recurrence. We can't to
1313   // insert a Trunc after a terminator if there happens to be a critical edge.
1314   assert(DU.NarrowUse != DU.NarrowUse->getParent()->getTerminator() &&
1315          "SCEV is not expected to evaluate a block terminator");
1316 
1317   // Reuse the IV increment that SCEVExpander created as long as it dominates
1318   // NarrowUse.
1319   Instruction *WideUse = nullptr;
1320   if (WideAddRec == WideIncExpr
1321       && Rewriter.hoistIVInc(WideInc, DU.NarrowUse))
1322     WideUse = WideInc;
1323   else {
1324     WideUse = cloneIVUser(DU, WideAddRec);
1325     if (!WideUse)
1326       return nullptr;
1327   }
1328   // Evaluation of WideAddRec ensured that the narrow expression could be
1329   // extended outside the loop without overflow. This suggests that the wide use
1330   // evaluates to the same expression as the extended narrow use, but doesn't
1331   // absolutely guarantee it. Hence the following failsafe check. In rare cases
1332   // where it fails, we simply throw away the newly created wide use.
1333   if (WideAddRec != SE->getSCEV(WideUse)) {
1334     DEBUG(dbgs() << "Wide use expression mismatch: " << *WideUse
1335           << ": " << *SE->getSCEV(WideUse) << " != " << *WideAddRec << "\n");
1336     DeadInsts.emplace_back(WideUse);
1337     return nullptr;
1338   }
1339 
1340   // Returning WideUse pushes it on the worklist.
1341   return WideUse;
1342 }
1343 
1344 /// Add eligible users of NarrowDef to NarrowIVUsers.
1345 ///
pushNarrowIVUsers(Instruction * NarrowDef,Instruction * WideDef)1346 void WidenIV::pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef) {
1347   const SCEV *NarrowSCEV = SE->getSCEV(NarrowDef);
1348   bool NeverNegative =
1349       SE->isKnownPredicate(ICmpInst::ICMP_SGE, NarrowSCEV,
1350                            SE->getConstant(NarrowSCEV->getType(), 0));
1351   for (User *U : NarrowDef->users()) {
1352     Instruction *NarrowUser = cast<Instruction>(U);
1353 
1354     // Handle data flow merges and bizarre phi cycles.
1355     if (!Widened.insert(NarrowUser).second)
1356       continue;
1357 
1358     NarrowIVUsers.push_back(
1359         NarrowIVDefUse(NarrowDef, NarrowUser, WideDef, NeverNegative));
1360   }
1361 }
1362 
1363 /// Process a single induction variable. First use the SCEVExpander to create a
1364 /// wide induction variable that evaluates to the same recurrence as the
1365 /// original narrow IV. Then use a worklist to forward traverse the narrow IV's
1366 /// def-use chain. After widenIVUse has processed all interesting IV users, the
1367 /// narrow IV will be isolated for removal by DeleteDeadPHIs.
1368 ///
1369 /// It would be simpler to delete uses as they are processed, but we must avoid
1370 /// invalidating SCEV expressions.
1371 ///
createWideIV(SCEVExpander & Rewriter)1372 PHINode *WidenIV::createWideIV(SCEVExpander &Rewriter) {
1373   // Is this phi an induction variable?
1374   const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(OrigPhi));
1375   if (!AddRec)
1376     return nullptr;
1377 
1378   // Widen the induction variable expression.
1379   const SCEV *WideIVExpr = IsSigned ?
1380     SE->getSignExtendExpr(AddRec, WideType) :
1381     SE->getZeroExtendExpr(AddRec, WideType);
1382 
1383   assert(SE->getEffectiveSCEVType(WideIVExpr->getType()) == WideType &&
1384          "Expect the new IV expression to preserve its type");
1385 
1386   // Can the IV be extended outside the loop without overflow?
1387   AddRec = dyn_cast<SCEVAddRecExpr>(WideIVExpr);
1388   if (!AddRec || AddRec->getLoop() != L)
1389     return nullptr;
1390 
1391   // An AddRec must have loop-invariant operands. Since this AddRec is
1392   // materialized by a loop header phi, the expression cannot have any post-loop
1393   // operands, so they must dominate the loop header.
1394   assert(SE->properlyDominates(AddRec->getStart(), L->getHeader()) &&
1395          SE->properlyDominates(AddRec->getStepRecurrence(*SE), L->getHeader())
1396          && "Loop header phi recurrence inputs do not dominate the loop");
1397 
1398   // The rewriter provides a value for the desired IV expression. This may
1399   // either find an existing phi or materialize a new one. Either way, we
1400   // expect a well-formed cyclic phi-with-increments. i.e. any operand not part
1401   // of the phi-SCC dominates the loop entry.
1402   Instruction *InsertPt = &L->getHeader()->front();
1403   WidePhi = cast<PHINode>(Rewriter.expandCodeFor(AddRec, WideType, InsertPt));
1404 
1405   // Remembering the WideIV increment generated by SCEVExpander allows
1406   // widenIVUse to reuse it when widening the narrow IV's increment. We don't
1407   // employ a general reuse mechanism because the call above is the only call to
1408   // SCEVExpander. Henceforth, we produce 1-to-1 narrow to wide uses.
1409   if (BasicBlock *LatchBlock = L->getLoopLatch()) {
1410     WideInc =
1411       cast<Instruction>(WidePhi->getIncomingValueForBlock(LatchBlock));
1412     WideIncExpr = SE->getSCEV(WideInc);
1413   }
1414 
1415   DEBUG(dbgs() << "Wide IV: " << *WidePhi << "\n");
1416   ++NumWidened;
1417 
1418   // Traverse the def-use chain using a worklist starting at the original IV.
1419   assert(Widened.empty() && NarrowIVUsers.empty() && "expect initial state" );
1420 
1421   Widened.insert(OrigPhi);
1422   pushNarrowIVUsers(OrigPhi, WidePhi);
1423 
1424   while (!NarrowIVUsers.empty()) {
1425     NarrowIVDefUse DU = NarrowIVUsers.pop_back_val();
1426 
1427     // Process a def-use edge. This may replace the use, so don't hold a
1428     // use_iterator across it.
1429     Instruction *WideUse = widenIVUse(DU, Rewriter);
1430 
1431     // Follow all def-use edges from the previous narrow use.
1432     if (WideUse)
1433       pushNarrowIVUsers(DU.NarrowUse, WideUse);
1434 
1435     // widenIVUse may have removed the def-use edge.
1436     if (DU.NarrowDef->use_empty())
1437       DeadInsts.emplace_back(DU.NarrowDef);
1438   }
1439   return WidePhi;
1440 }
1441 
1442 //===----------------------------------------------------------------------===//
1443 //  Live IV Reduction - Minimize IVs live across the loop.
1444 //===----------------------------------------------------------------------===//
1445 
1446 
1447 //===----------------------------------------------------------------------===//
1448 //  Simplification of IV users based on SCEV evaluation.
1449 //===----------------------------------------------------------------------===//
1450 
1451 namespace {
1452 class IndVarSimplifyVisitor : public IVVisitor {
1453   ScalarEvolution *SE;
1454   const TargetTransformInfo *TTI;
1455   PHINode *IVPhi;
1456 
1457 public:
1458   WideIVInfo WI;
1459 
IndVarSimplifyVisitor(PHINode * IV,ScalarEvolution * SCEV,const TargetTransformInfo * TTI,const DominatorTree * DTree)1460   IndVarSimplifyVisitor(PHINode *IV, ScalarEvolution *SCEV,
1461                         const TargetTransformInfo *TTI,
1462                         const DominatorTree *DTree)
1463     : SE(SCEV), TTI(TTI), IVPhi(IV) {
1464     DT = DTree;
1465     WI.NarrowIV = IVPhi;
1466     if (ReduceLiveIVs)
1467       setSplitOverflowIntrinsics();
1468   }
1469 
1470   // Implement the interface used by simplifyUsersOfIV.
visitCast(CastInst * Cast)1471   void visitCast(CastInst *Cast) override { visitIVCast(Cast, WI, SE, TTI); }
1472 };
1473 }
1474 
1475 /// Iteratively perform simplification on a worklist of IV users. Each
1476 /// successive simplification may push more users which may themselves be
1477 /// candidates for simplification.
1478 ///
1479 /// Sign/Zero extend elimination is interleaved with IV simplification.
1480 ///
simplifyAndExtend(Loop * L,SCEVExpander & Rewriter,LoopInfo * LI)1481 void IndVarSimplify::simplifyAndExtend(Loop *L,
1482                                        SCEVExpander &Rewriter,
1483                                        LoopInfo *LI) {
1484   SmallVector<WideIVInfo, 8> WideIVs;
1485 
1486   SmallVector<PHINode*, 8> LoopPhis;
1487   for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
1488     LoopPhis.push_back(cast<PHINode>(I));
1489   }
1490   // Each round of simplification iterates through the SimplifyIVUsers worklist
1491   // for all current phis, then determines whether any IVs can be
1492   // widened. Widening adds new phis to LoopPhis, inducing another round of
1493   // simplification on the wide IVs.
1494   while (!LoopPhis.empty()) {
1495     // Evaluate as many IV expressions as possible before widening any IVs. This
1496     // forces SCEV to set no-wrap flags before evaluating sign/zero
1497     // extension. The first time SCEV attempts to normalize sign/zero extension,
1498     // the result becomes final. So for the most predictable results, we delay
1499     // evaluation of sign/zero extend evaluation until needed, and avoid running
1500     // other SCEV based analysis prior to simplifyAndExtend.
1501     do {
1502       PHINode *CurrIV = LoopPhis.pop_back_val();
1503 
1504       // Information about sign/zero extensions of CurrIV.
1505       IndVarSimplifyVisitor Visitor(CurrIV, SE, TTI, DT);
1506 
1507       Changed |= simplifyUsersOfIV(CurrIV, SE, DT, LI, DeadInsts, &Visitor);
1508 
1509       if (Visitor.WI.WidestNativeType) {
1510         WideIVs.push_back(Visitor.WI);
1511       }
1512     } while(!LoopPhis.empty());
1513 
1514     for (; !WideIVs.empty(); WideIVs.pop_back()) {
1515       WidenIV Widener(WideIVs.back(), LI, SE, DT, DeadInsts);
1516       if (PHINode *WidePhi = Widener.createWideIV(Rewriter)) {
1517         Changed = true;
1518         LoopPhis.push_back(WidePhi);
1519       }
1520     }
1521   }
1522 }
1523 
1524 //===----------------------------------------------------------------------===//
1525 //  linearFunctionTestReplace and its kin. Rewrite the loop exit condition.
1526 //===----------------------------------------------------------------------===//
1527 
1528 /// Return true if this loop's backedge taken count expression can be safely and
1529 /// cheaply expanded into an instruction sequence that can be used by
1530 /// linearFunctionTestReplace.
1531 ///
1532 /// TODO: This fails for pointer-type loop counters with greater than one byte
1533 /// strides, consequently preventing LFTR from running. For the purpose of LFTR
1534 /// we could skip this check in the case that the LFTR loop counter (chosen by
1535 /// FindLoopCounter) is also pointer type. Instead, we could directly convert
1536 /// the loop test to an inequality test by checking the target data's alignment
1537 /// of element types (given that the initial pointer value originates from or is
1538 /// used by ABI constrained operation, as opposed to inttoptr/ptrtoint).
1539 /// However, we don't yet have a strong motivation for converting loop tests
1540 /// into inequality tests.
canExpandBackedgeTakenCount(Loop * L,ScalarEvolution * SE,SCEVExpander & Rewriter)1541 static bool canExpandBackedgeTakenCount(Loop *L, ScalarEvolution *SE,
1542                                         SCEVExpander &Rewriter) {
1543   const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
1544   if (isa<SCEVCouldNotCompute>(BackedgeTakenCount) ||
1545       BackedgeTakenCount->isZero())
1546     return false;
1547 
1548   if (!L->getExitingBlock())
1549     return false;
1550 
1551   // Can't rewrite non-branch yet.
1552   if (!isa<BranchInst>(L->getExitingBlock()->getTerminator()))
1553     return false;
1554 
1555   if (Rewriter.isHighCostExpansion(BackedgeTakenCount, L))
1556     return false;
1557 
1558   return true;
1559 }
1560 
1561 /// Return the loop header phi IFF IncV adds a loop invariant value to the phi.
getLoopPhiForCounter(Value * IncV,Loop * L,DominatorTree * DT)1562 static PHINode *getLoopPhiForCounter(Value *IncV, Loop *L, DominatorTree *DT) {
1563   Instruction *IncI = dyn_cast<Instruction>(IncV);
1564   if (!IncI)
1565     return nullptr;
1566 
1567   switch (IncI->getOpcode()) {
1568   case Instruction::Add:
1569   case Instruction::Sub:
1570     break;
1571   case Instruction::GetElementPtr:
1572     // An IV counter must preserve its type.
1573     if (IncI->getNumOperands() == 2)
1574       break;
1575   default:
1576     return nullptr;
1577   }
1578 
1579   PHINode *Phi = dyn_cast<PHINode>(IncI->getOperand(0));
1580   if (Phi && Phi->getParent() == L->getHeader()) {
1581     if (isLoopInvariant(IncI->getOperand(1), L, DT))
1582       return Phi;
1583     return nullptr;
1584   }
1585   if (IncI->getOpcode() == Instruction::GetElementPtr)
1586     return nullptr;
1587 
1588   // Allow add/sub to be commuted.
1589   Phi = dyn_cast<PHINode>(IncI->getOperand(1));
1590   if (Phi && Phi->getParent() == L->getHeader()) {
1591     if (isLoopInvariant(IncI->getOperand(0), L, DT))
1592       return Phi;
1593   }
1594   return nullptr;
1595 }
1596 
1597 /// Return the compare guarding the loop latch, or NULL for unrecognized tests.
getLoopTest(Loop * L)1598 static ICmpInst *getLoopTest(Loop *L) {
1599   assert(L->getExitingBlock() && "expected loop exit");
1600 
1601   BasicBlock *LatchBlock = L->getLoopLatch();
1602   // Don't bother with LFTR if the loop is not properly simplified.
1603   if (!LatchBlock)
1604     return nullptr;
1605 
1606   BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
1607   assert(BI && "expected exit branch");
1608 
1609   return dyn_cast<ICmpInst>(BI->getCondition());
1610 }
1611 
1612 /// linearFunctionTestReplace policy. Return true unless we can show that the
1613 /// current exit test is already sufficiently canonical.
needsLFTR(Loop * L,DominatorTree * DT)1614 static bool needsLFTR(Loop *L, DominatorTree *DT) {
1615   // Do LFTR to simplify the exit condition to an ICMP.
1616   ICmpInst *Cond = getLoopTest(L);
1617   if (!Cond)
1618     return true;
1619 
1620   // Do LFTR to simplify the exit ICMP to EQ/NE
1621   ICmpInst::Predicate Pred = Cond->getPredicate();
1622   if (Pred != ICmpInst::ICMP_NE && Pred != ICmpInst::ICMP_EQ)
1623     return true;
1624 
1625   // Look for a loop invariant RHS
1626   Value *LHS = Cond->getOperand(0);
1627   Value *RHS = Cond->getOperand(1);
1628   if (!isLoopInvariant(RHS, L, DT)) {
1629     if (!isLoopInvariant(LHS, L, DT))
1630       return true;
1631     std::swap(LHS, RHS);
1632   }
1633   // Look for a simple IV counter LHS
1634   PHINode *Phi = dyn_cast<PHINode>(LHS);
1635   if (!Phi)
1636     Phi = getLoopPhiForCounter(LHS, L, DT);
1637 
1638   if (!Phi)
1639     return true;
1640 
1641   // Do LFTR if PHI node is defined in the loop, but is *not* a counter.
1642   int Idx = Phi->getBasicBlockIndex(L->getLoopLatch());
1643   if (Idx < 0)
1644     return true;
1645 
1646   // Do LFTR if the exit condition's IV is *not* a simple counter.
1647   Value *IncV = Phi->getIncomingValue(Idx);
1648   return Phi != getLoopPhiForCounter(IncV, L, DT);
1649 }
1650 
1651 /// Recursive helper for hasConcreteDef(). Unfortunately, this currently boils
1652 /// down to checking that all operands are constant and listing instructions
1653 /// that may hide undef.
hasConcreteDefImpl(Value * V,SmallPtrSetImpl<Value * > & Visited,unsigned Depth)1654 static bool hasConcreteDefImpl(Value *V, SmallPtrSetImpl<Value*> &Visited,
1655                                unsigned Depth) {
1656   if (isa<Constant>(V))
1657     return !isa<UndefValue>(V);
1658 
1659   if (Depth >= 6)
1660     return false;
1661 
1662   // Conservatively handle non-constant non-instructions. For example, Arguments
1663   // may be undef.
1664   Instruction *I = dyn_cast<Instruction>(V);
1665   if (!I)
1666     return false;
1667 
1668   // Load and return values may be undef.
1669   if(I->mayReadFromMemory() || isa<CallInst>(I) || isa<InvokeInst>(I))
1670     return false;
1671 
1672   // Optimistically handle other instructions.
1673   for (Value *Op : I->operands()) {
1674     if (!Visited.insert(Op).second)
1675       continue;
1676     if (!hasConcreteDefImpl(Op, Visited, Depth+1))
1677       return false;
1678   }
1679   return true;
1680 }
1681 
1682 /// Return true if the given value is concrete. We must prove that undef can
1683 /// never reach it.
1684 ///
1685 /// TODO: If we decide that this is a good approach to checking for undef, we
1686 /// may factor it into a common location.
hasConcreteDef(Value * V)1687 static bool hasConcreteDef(Value *V) {
1688   SmallPtrSet<Value*, 8> Visited;
1689   Visited.insert(V);
1690   return hasConcreteDefImpl(V, Visited, 0);
1691 }
1692 
1693 /// Return true if this IV has any uses other than the (soon to be rewritten)
1694 /// loop exit test.
AlmostDeadIV(PHINode * Phi,BasicBlock * LatchBlock,Value * Cond)1695 static bool AlmostDeadIV(PHINode *Phi, BasicBlock *LatchBlock, Value *Cond) {
1696   int LatchIdx = Phi->getBasicBlockIndex(LatchBlock);
1697   Value *IncV = Phi->getIncomingValue(LatchIdx);
1698 
1699   for (User *U : Phi->users())
1700     if (U != Cond && U != IncV) return false;
1701 
1702   for (User *U : IncV->users())
1703     if (U != Cond && U != Phi) return false;
1704   return true;
1705 }
1706 
1707 /// Find an affine IV in canonical form.
1708 ///
1709 /// BECount may be an i8* pointer type. The pointer difference is already
1710 /// valid count without scaling the address stride, so it remains a pointer
1711 /// expression as far as SCEV is concerned.
1712 ///
1713 /// Currently only valid for LFTR. See the comments on hasConcreteDef below.
1714 ///
1715 /// FIXME: Accept -1 stride and set IVLimit = IVInit - BECount
1716 ///
1717 /// FIXME: Accept non-unit stride as long as SCEV can reduce BECount * Stride.
1718 /// This is difficult in general for SCEV because of potential overflow. But we
1719 /// could at least handle constant BECounts.
FindLoopCounter(Loop * L,const SCEV * BECount,ScalarEvolution * SE,DominatorTree * DT)1720 static PHINode *FindLoopCounter(Loop *L, const SCEV *BECount,
1721                                 ScalarEvolution *SE, DominatorTree *DT) {
1722   uint64_t BCWidth = SE->getTypeSizeInBits(BECount->getType());
1723 
1724   Value *Cond =
1725     cast<BranchInst>(L->getExitingBlock()->getTerminator())->getCondition();
1726 
1727   // Loop over all of the PHI nodes, looking for a simple counter.
1728   PHINode *BestPhi = nullptr;
1729   const SCEV *BestInit = nullptr;
1730   BasicBlock *LatchBlock = L->getLoopLatch();
1731   assert(LatchBlock && "needsLFTR should guarantee a loop latch");
1732 
1733   for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
1734     PHINode *Phi = cast<PHINode>(I);
1735     if (!SE->isSCEVable(Phi->getType()))
1736       continue;
1737 
1738     // Avoid comparing an integer IV against a pointer Limit.
1739     if (BECount->getType()->isPointerTy() && !Phi->getType()->isPointerTy())
1740       continue;
1741 
1742     const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Phi));
1743     if (!AR || AR->getLoop() != L || !AR->isAffine())
1744       continue;
1745 
1746     // AR may be a pointer type, while BECount is an integer type.
1747     // AR may be wider than BECount. With eq/ne tests overflow is immaterial.
1748     // AR may not be a narrower type, or we may never exit.
1749     uint64_t PhiWidth = SE->getTypeSizeInBits(AR->getType());
1750     if (PhiWidth < BCWidth ||
1751         !L->getHeader()->getModule()->getDataLayout().isLegalInteger(PhiWidth))
1752       continue;
1753 
1754     const SCEV *Step = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE));
1755     if (!Step || !Step->isOne())
1756       continue;
1757 
1758     int LatchIdx = Phi->getBasicBlockIndex(LatchBlock);
1759     Value *IncV = Phi->getIncomingValue(LatchIdx);
1760     if (getLoopPhiForCounter(IncV, L, DT) != Phi)
1761       continue;
1762 
1763     // Avoid reusing a potentially undef value to compute other values that may
1764     // have originally had a concrete definition.
1765     if (!hasConcreteDef(Phi)) {
1766       // We explicitly allow unknown phis as long as they are already used by
1767       // the loop test. In this case we assume that performing LFTR could not
1768       // increase the number of undef users.
1769       if (ICmpInst *Cond = getLoopTest(L)) {
1770         if (Phi != getLoopPhiForCounter(Cond->getOperand(0), L, DT)
1771             && Phi != getLoopPhiForCounter(Cond->getOperand(1), L, DT)) {
1772           continue;
1773         }
1774       }
1775     }
1776     const SCEV *Init = AR->getStart();
1777 
1778     if (BestPhi && !AlmostDeadIV(BestPhi, LatchBlock, Cond)) {
1779       // Don't force a live loop counter if another IV can be used.
1780       if (AlmostDeadIV(Phi, LatchBlock, Cond))
1781         continue;
1782 
1783       // Prefer to count-from-zero. This is a more "canonical" counter form. It
1784       // also prefers integer to pointer IVs.
1785       if (BestInit->isZero() != Init->isZero()) {
1786         if (BestInit->isZero())
1787           continue;
1788       }
1789       // If two IVs both count from zero or both count from nonzero then the
1790       // narrower is likely a dead phi that has been widened. Use the wider phi
1791       // to allow the other to be eliminated.
1792       else if (PhiWidth <= SE->getTypeSizeInBits(BestPhi->getType()))
1793         continue;
1794     }
1795     BestPhi = Phi;
1796     BestInit = Init;
1797   }
1798   return BestPhi;
1799 }
1800 
1801 /// Help linearFunctionTestReplace by generating a value that holds the RHS of
1802 /// the new loop test.
genLoopLimit(PHINode * IndVar,const SCEV * IVCount,Loop * L,SCEVExpander & Rewriter,ScalarEvolution * SE)1803 static Value *genLoopLimit(PHINode *IndVar, const SCEV *IVCount, Loop *L,
1804                            SCEVExpander &Rewriter, ScalarEvolution *SE) {
1805   const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(IndVar));
1806   assert(AR && AR->getLoop() == L && AR->isAffine() && "bad loop counter");
1807   const SCEV *IVInit = AR->getStart();
1808 
1809   // IVInit may be a pointer while IVCount is an integer when FindLoopCounter
1810   // finds a valid pointer IV. Sign extend BECount in order to materialize a
1811   // GEP. Avoid running SCEVExpander on a new pointer value, instead reusing
1812   // the existing GEPs whenever possible.
1813   if (IndVar->getType()->isPointerTy()
1814       && !IVCount->getType()->isPointerTy()) {
1815 
1816     // IVOffset will be the new GEP offset that is interpreted by GEP as a
1817     // signed value. IVCount on the other hand represents the loop trip count,
1818     // which is an unsigned value. FindLoopCounter only allows induction
1819     // variables that have a positive unit stride of one. This means we don't
1820     // have to handle the case of negative offsets (yet) and just need to zero
1821     // extend IVCount.
1822     Type *OfsTy = SE->getEffectiveSCEVType(IVInit->getType());
1823     const SCEV *IVOffset = SE->getTruncateOrZeroExtend(IVCount, OfsTy);
1824 
1825     // Expand the code for the iteration count.
1826     assert(SE->isLoopInvariant(IVOffset, L) &&
1827            "Computed iteration count is not loop invariant!");
1828     BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1829     Value *GEPOffset = Rewriter.expandCodeFor(IVOffset, OfsTy, BI);
1830 
1831     Value *GEPBase = IndVar->getIncomingValueForBlock(L->getLoopPreheader());
1832     assert(AR->getStart() == SE->getSCEV(GEPBase) && "bad loop counter");
1833     // We could handle pointer IVs other than i8*, but we need to compensate for
1834     // gep index scaling. See canExpandBackedgeTakenCount comments.
1835     assert(SE->getSizeOfExpr(IntegerType::getInt64Ty(IndVar->getContext()),
1836              cast<PointerType>(GEPBase->getType())->getElementType())->isOne()
1837            && "unit stride pointer IV must be i8*");
1838 
1839     IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
1840     return Builder.CreateGEP(nullptr, GEPBase, GEPOffset, "lftr.limit");
1841   }
1842   else {
1843     // In any other case, convert both IVInit and IVCount to integers before
1844     // comparing. This may result in SCEV expension of pointers, but in practice
1845     // SCEV will fold the pointer arithmetic away as such:
1846     // BECount = (IVEnd - IVInit - 1) => IVLimit = IVInit (postinc).
1847     //
1848     // Valid Cases: (1) both integers is most common; (2) both may be pointers
1849     // for simple memset-style loops.
1850     //
1851     // IVInit integer and IVCount pointer would only occur if a canonical IV
1852     // were generated on top of case #2, which is not expected.
1853 
1854     const SCEV *IVLimit = nullptr;
1855     // For unit stride, IVCount = Start + BECount with 2's complement overflow.
1856     // For non-zero Start, compute IVCount here.
1857     if (AR->getStart()->isZero())
1858       IVLimit = IVCount;
1859     else {
1860       assert(AR->getStepRecurrence(*SE)->isOne() && "only handles unit stride");
1861       const SCEV *IVInit = AR->getStart();
1862 
1863       // For integer IVs, truncate the IV before computing IVInit + BECount.
1864       if (SE->getTypeSizeInBits(IVInit->getType())
1865           > SE->getTypeSizeInBits(IVCount->getType()))
1866         IVInit = SE->getTruncateExpr(IVInit, IVCount->getType());
1867 
1868       IVLimit = SE->getAddExpr(IVInit, IVCount);
1869     }
1870     // Expand the code for the iteration count.
1871     BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1872     IRBuilder<> Builder(BI);
1873     assert(SE->isLoopInvariant(IVLimit, L) &&
1874            "Computed iteration count is not loop invariant!");
1875     // Ensure that we generate the same type as IndVar, or a smaller integer
1876     // type. In the presence of null pointer values, we have an integer type
1877     // SCEV expression (IVInit) for a pointer type IV value (IndVar).
1878     Type *LimitTy = IVCount->getType()->isPointerTy() ?
1879       IndVar->getType() : IVCount->getType();
1880     return Rewriter.expandCodeFor(IVLimit, LimitTy, BI);
1881   }
1882 }
1883 
1884 /// This method rewrites the exit condition of the loop to be a canonical !=
1885 /// comparison against the incremented loop induction variable.  This pass is
1886 /// able to rewrite the exit tests of any loop where the SCEV analysis can
1887 /// determine a loop-invariant trip count of the loop, which is actually a much
1888 /// broader range than just linear tests.
1889 Value *IndVarSimplify::
linearFunctionTestReplace(Loop * L,const SCEV * BackedgeTakenCount,PHINode * IndVar,SCEVExpander & Rewriter)1890 linearFunctionTestReplace(Loop *L,
1891                           const SCEV *BackedgeTakenCount,
1892                           PHINode *IndVar,
1893                           SCEVExpander &Rewriter) {
1894   assert(canExpandBackedgeTakenCount(L, SE, Rewriter) && "precondition");
1895 
1896   // Initialize CmpIndVar and IVCount to their preincremented values.
1897   Value *CmpIndVar = IndVar;
1898   const SCEV *IVCount = BackedgeTakenCount;
1899 
1900   // If the exiting block is the same as the backedge block, we prefer to
1901   // compare against the post-incremented value, otherwise we must compare
1902   // against the preincremented value.
1903   if (L->getExitingBlock() == L->getLoopLatch()) {
1904     // Add one to the "backedge-taken" count to get the trip count.
1905     // This addition may overflow, which is valid as long as the comparison is
1906     // truncated to BackedgeTakenCount->getType().
1907     IVCount = SE->getAddExpr(BackedgeTakenCount,
1908                              SE->getOne(BackedgeTakenCount->getType()));
1909     // The BackedgeTaken expression contains the number of times that the
1910     // backedge branches to the loop header.  This is one less than the
1911     // number of times the loop executes, so use the incremented indvar.
1912     CmpIndVar = IndVar->getIncomingValueForBlock(L->getExitingBlock());
1913   }
1914 
1915   Value *ExitCnt = genLoopLimit(IndVar, IVCount, L, Rewriter, SE);
1916   assert(ExitCnt->getType()->isPointerTy() == IndVar->getType()->isPointerTy()
1917          && "genLoopLimit missed a cast");
1918 
1919   // Insert a new icmp_ne or icmp_eq instruction before the branch.
1920   BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1921   ICmpInst::Predicate P;
1922   if (L->contains(BI->getSuccessor(0)))
1923     P = ICmpInst::ICMP_NE;
1924   else
1925     P = ICmpInst::ICMP_EQ;
1926 
1927   DEBUG(dbgs() << "INDVARS: Rewriting loop exit condition to:\n"
1928                << "      LHS:" << *CmpIndVar << '\n'
1929                << "       op:\t"
1930                << (P == ICmpInst::ICMP_NE ? "!=" : "==") << "\n"
1931                << "      RHS:\t" << *ExitCnt << "\n"
1932                << "  IVCount:\t" << *IVCount << "\n");
1933 
1934   IRBuilder<> Builder(BI);
1935 
1936   // LFTR can ignore IV overflow and truncate to the width of
1937   // BECount. This avoids materializing the add(zext(add)) expression.
1938   unsigned CmpIndVarSize = SE->getTypeSizeInBits(CmpIndVar->getType());
1939   unsigned ExitCntSize = SE->getTypeSizeInBits(ExitCnt->getType());
1940   if (CmpIndVarSize > ExitCntSize) {
1941     const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(SE->getSCEV(IndVar));
1942     const SCEV *ARStart = AR->getStart();
1943     const SCEV *ARStep = AR->getStepRecurrence(*SE);
1944     // For constant IVCount, avoid truncation.
1945     if (isa<SCEVConstant>(ARStart) && isa<SCEVConstant>(IVCount)) {
1946       const APInt &Start = cast<SCEVConstant>(ARStart)->getAPInt();
1947       APInt Count = cast<SCEVConstant>(IVCount)->getAPInt();
1948       // Note that the post-inc value of BackedgeTakenCount may have overflowed
1949       // above such that IVCount is now zero.
1950       if (IVCount != BackedgeTakenCount && Count == 0) {
1951         Count = APInt::getMaxValue(Count.getBitWidth()).zext(CmpIndVarSize);
1952         ++Count;
1953       }
1954       else
1955         Count = Count.zext(CmpIndVarSize);
1956       APInt NewLimit;
1957       if (cast<SCEVConstant>(ARStep)->getValue()->isNegative())
1958         NewLimit = Start - Count;
1959       else
1960         NewLimit = Start + Count;
1961       ExitCnt = ConstantInt::get(CmpIndVar->getType(), NewLimit);
1962 
1963       DEBUG(dbgs() << "  Widen RHS:\t" << *ExitCnt << "\n");
1964     } else {
1965       CmpIndVar = Builder.CreateTrunc(CmpIndVar, ExitCnt->getType(),
1966                                       "lftr.wideiv");
1967     }
1968   }
1969   Value *Cond = Builder.CreateICmp(P, CmpIndVar, ExitCnt, "exitcond");
1970   Value *OrigCond = BI->getCondition();
1971   // It's tempting to use replaceAllUsesWith here to fully replace the old
1972   // comparison, but that's not immediately safe, since users of the old
1973   // comparison may not be dominated by the new comparison. Instead, just
1974   // update the branch to use the new comparison; in the common case this
1975   // will make old comparison dead.
1976   BI->setCondition(Cond);
1977   DeadInsts.push_back(OrigCond);
1978 
1979   ++NumLFTR;
1980   Changed = true;
1981   return Cond;
1982 }
1983 
1984 //===----------------------------------------------------------------------===//
1985 //  sinkUnusedInvariants. A late subpass to cleanup loop preheaders.
1986 //===----------------------------------------------------------------------===//
1987 
1988 /// If there's a single exit block, sink any loop-invariant values that
1989 /// were defined in the preheader but not used inside the loop into the
1990 /// exit block to reduce register pressure in the loop.
sinkUnusedInvariants(Loop * L)1991 void IndVarSimplify::sinkUnusedInvariants(Loop *L) {
1992   BasicBlock *ExitBlock = L->getExitBlock();
1993   if (!ExitBlock) return;
1994 
1995   BasicBlock *Preheader = L->getLoopPreheader();
1996   if (!Preheader) return;
1997 
1998   Instruction *InsertPt = &*ExitBlock->getFirstInsertionPt();
1999   BasicBlock::iterator I(Preheader->getTerminator());
2000   while (I != Preheader->begin()) {
2001     --I;
2002     // New instructions were inserted at the end of the preheader.
2003     if (isa<PHINode>(I))
2004       break;
2005 
2006     // Don't move instructions which might have side effects, since the side
2007     // effects need to complete before instructions inside the loop.  Also don't
2008     // move instructions which might read memory, since the loop may modify
2009     // memory. Note that it's okay if the instruction might have undefined
2010     // behavior: LoopSimplify guarantees that the preheader dominates the exit
2011     // block.
2012     if (I->mayHaveSideEffects() || I->mayReadFromMemory())
2013       continue;
2014 
2015     // Skip debug info intrinsics.
2016     if (isa<DbgInfoIntrinsic>(I))
2017       continue;
2018 
2019     // Skip eh pad instructions.
2020     if (I->isEHPad())
2021       continue;
2022 
2023     // Don't sink alloca: we never want to sink static alloca's out of the
2024     // entry block, and correctly sinking dynamic alloca's requires
2025     // checks for stacksave/stackrestore intrinsics.
2026     // FIXME: Refactor this check somehow?
2027     if (isa<AllocaInst>(I))
2028       continue;
2029 
2030     // Determine if there is a use in or before the loop (direct or
2031     // otherwise).
2032     bool UsedInLoop = false;
2033     for (Use &U : I->uses()) {
2034       Instruction *User = cast<Instruction>(U.getUser());
2035       BasicBlock *UseBB = User->getParent();
2036       if (PHINode *P = dyn_cast<PHINode>(User)) {
2037         unsigned i =
2038           PHINode::getIncomingValueNumForOperand(U.getOperandNo());
2039         UseBB = P->getIncomingBlock(i);
2040       }
2041       if (UseBB == Preheader || L->contains(UseBB)) {
2042         UsedInLoop = true;
2043         break;
2044       }
2045     }
2046 
2047     // If there is, the def must remain in the preheader.
2048     if (UsedInLoop)
2049       continue;
2050 
2051     // Otherwise, sink it to the exit block.
2052     Instruction *ToMove = &*I;
2053     bool Done = false;
2054 
2055     if (I != Preheader->begin()) {
2056       // Skip debug info intrinsics.
2057       do {
2058         --I;
2059       } while (isa<DbgInfoIntrinsic>(I) && I != Preheader->begin());
2060 
2061       if (isa<DbgInfoIntrinsic>(I) && I == Preheader->begin())
2062         Done = true;
2063     } else {
2064       Done = true;
2065     }
2066 
2067     ToMove->moveBefore(InsertPt);
2068     if (Done) break;
2069     InsertPt = ToMove;
2070   }
2071 }
2072 
2073 //===----------------------------------------------------------------------===//
2074 //  IndVarSimplify driver. Manage several subpasses of IV simplification.
2075 //===----------------------------------------------------------------------===//
2076 
runOnLoop(Loop * L,LPPassManager & LPM)2077 bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) {
2078   if (skipOptnoneFunction(L))
2079     return false;
2080 
2081   // If LoopSimplify form is not available, stay out of trouble. Some notes:
2082   //  - LSR currently only supports LoopSimplify-form loops. Indvars'
2083   //    canonicalization can be a pessimization without LSR to "clean up"
2084   //    afterwards.
2085   //  - We depend on having a preheader; in particular,
2086   //    Loop::getCanonicalInductionVariable only supports loops with preheaders,
2087   //    and we're in trouble if we can't find the induction variable even when
2088   //    we've manually inserted one.
2089   if (!L->isLoopSimplifyForm())
2090     return false;
2091 
2092   LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
2093   SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
2094   DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2095   auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
2096   TLI = TLIP ? &TLIP->getTLI() : nullptr;
2097   auto *TTIP = getAnalysisIfAvailable<TargetTransformInfoWrapperPass>();
2098   TTI = TTIP ? &TTIP->getTTI(*L->getHeader()->getParent()) : nullptr;
2099   const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
2100 
2101   DeadInsts.clear();
2102   Changed = false;
2103 
2104   // If there are any floating-point recurrences, attempt to
2105   // transform them to use integer recurrences.
2106   rewriteNonIntegerIVs(L);
2107 
2108   const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
2109 
2110   // Create a rewriter object which we'll use to transform the code with.
2111   SCEVExpander Rewriter(*SE, DL, "indvars");
2112 #ifndef NDEBUG
2113   Rewriter.setDebugType(DEBUG_TYPE);
2114 #endif
2115 
2116   // Eliminate redundant IV users.
2117   //
2118   // Simplification works best when run before other consumers of SCEV. We
2119   // attempt to avoid evaluating SCEVs for sign/zero extend operations until
2120   // other expressions involving loop IVs have been evaluated. This helps SCEV
2121   // set no-wrap flags before normalizing sign/zero extension.
2122   Rewriter.disableCanonicalMode();
2123   simplifyAndExtend(L, Rewriter, LI);
2124 
2125   // Check to see if this loop has a computable loop-invariant execution count.
2126   // If so, this means that we can compute the final value of any expressions
2127   // that are recurrent in the loop, and substitute the exit values from the
2128   // loop into any instructions outside of the loop that use the final values of
2129   // the current expressions.
2130   //
2131   if (ReplaceExitValue != NeverRepl &&
2132       !isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2133     rewriteLoopExitValues(L, Rewriter);
2134 
2135   // Eliminate redundant IV cycles.
2136   NumElimIV += Rewriter.replaceCongruentIVs(L, DT, DeadInsts);
2137 
2138   // If we have a trip count expression, rewrite the loop's exit condition
2139   // using it.  We can currently only handle loops with a single exit.
2140   if (canExpandBackedgeTakenCount(L, SE, Rewriter) && needsLFTR(L, DT)) {
2141     PHINode *IndVar = FindLoopCounter(L, BackedgeTakenCount, SE, DT);
2142     if (IndVar) {
2143       // Check preconditions for proper SCEVExpander operation. SCEV does not
2144       // express SCEVExpander's dependencies, such as LoopSimplify. Instead any
2145       // pass that uses the SCEVExpander must do it. This does not work well for
2146       // loop passes because SCEVExpander makes assumptions about all loops,
2147       // while LoopPassManager only forces the current loop to be simplified.
2148       //
2149       // FIXME: SCEV expansion has no way to bail out, so the caller must
2150       // explicitly check any assumptions made by SCEV. Brittle.
2151       const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(BackedgeTakenCount);
2152       if (!AR || AR->getLoop()->getLoopPreheader())
2153         (void)linearFunctionTestReplace(L, BackedgeTakenCount, IndVar,
2154                                         Rewriter);
2155     }
2156   }
2157   // Clear the rewriter cache, because values that are in the rewriter's cache
2158   // can be deleted in the loop below, causing the AssertingVH in the cache to
2159   // trigger.
2160   Rewriter.clear();
2161 
2162   // Now that we're done iterating through lists, clean up any instructions
2163   // which are now dead.
2164   while (!DeadInsts.empty())
2165     if (Instruction *Inst =
2166             dyn_cast_or_null<Instruction>(DeadInsts.pop_back_val()))
2167       RecursivelyDeleteTriviallyDeadInstructions(Inst, TLI);
2168 
2169   // The Rewriter may not be used from this point on.
2170 
2171   // Loop-invariant instructions in the preheader that aren't used in the
2172   // loop may be sunk below the loop to reduce register pressure.
2173   sinkUnusedInvariants(L);
2174 
2175   // Clean up dead instructions.
2176   Changed |= DeleteDeadPHIs(L->getHeader(), TLI);
2177 
2178   // Check a post-condition.
2179   assert(L->isRecursivelyLCSSAForm(*DT) && "Indvars did not preserve LCSSA!");
2180 
2181   // Verify that LFTR, and any other change have not interfered with SCEV's
2182   // ability to compute trip count.
2183 #ifndef NDEBUG
2184   if (VerifyIndvars && !isa<SCEVCouldNotCompute>(BackedgeTakenCount)) {
2185     SE->forgetLoop(L);
2186     const SCEV *NewBECount = SE->getBackedgeTakenCount(L);
2187     if (SE->getTypeSizeInBits(BackedgeTakenCount->getType()) <
2188         SE->getTypeSizeInBits(NewBECount->getType()))
2189       NewBECount = SE->getTruncateOrNoop(NewBECount,
2190                                          BackedgeTakenCount->getType());
2191     else
2192       BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount,
2193                                                  NewBECount->getType());
2194     assert(BackedgeTakenCount == NewBECount && "indvars must preserve SCEV");
2195   }
2196 #endif
2197 
2198   return Changed;
2199 }
2200