• 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/IndVarSimplify.h"
28 #include "llvm/Transforms/Scalar.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/LoopPassManager.h"
35 #include "llvm/Analysis/ScalarEvolutionExpander.h"
36 #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h"
37 #include "llvm/Analysis/TargetLibraryInfo.h"
38 #include "llvm/Analysis/TargetTransformInfo.h"
39 #include "llvm/IR/BasicBlock.h"
40 #include "llvm/IR/CFG.h"
41 #include "llvm/IR/Constants.h"
42 #include "llvm/IR/DataLayout.h"
43 #include "llvm/IR/Dominators.h"
44 #include "llvm/IR/Instructions.h"
45 #include "llvm/IR/IntrinsicInst.h"
46 #include "llvm/IR/LLVMContext.h"
47 #include "llvm/IR/PatternMatch.h"
48 #include "llvm/IR/Type.h"
49 #include "llvm/Support/CommandLine.h"
50 #include "llvm/Support/Debug.h"
51 #include "llvm/Support/raw_ostream.h"
52 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
53 #include "llvm/Transforms/Utils/Local.h"
54 #include "llvm/Transforms/Utils/LoopUtils.h"
55 #include "llvm/Transforms/Utils/SimplifyIndVar.h"
56 using namespace llvm;
57 
58 #define DEBUG_TYPE "indvars"
59 
60 STATISTIC(NumWidened     , "Number of indvars widened");
61 STATISTIC(NumReplaced    , "Number of exit values replaced");
62 STATISTIC(NumLFTR        , "Number of loop exit tests replaced");
63 STATISTIC(NumElimExt     , "Number of IV sign/zero extends eliminated");
64 STATISTIC(NumElimIV      , "Number of congruent IVs eliminated");
65 
66 // Trip count verification can be enabled by default under NDEBUG if we
67 // implement a strong expression equivalence checker in SCEV. Until then, we
68 // use the verify-indvars flag, which may assert in some cases.
69 static cl::opt<bool> VerifyIndvars(
70   "verify-indvars", cl::Hidden,
71   cl::desc("Verify the ScalarEvolution result after running indvars"));
72 
73 enum ReplaceExitVal { NeverRepl, OnlyCheapRepl, AlwaysRepl };
74 
75 static cl::opt<ReplaceExitVal> ReplaceExitValue(
76     "replexitval", cl::Hidden, cl::init(OnlyCheapRepl),
77     cl::desc("Choose the strategy to replace exit value in IndVarSimplify"),
78     cl::values(clEnumValN(NeverRepl, "never", "never replace exit value"),
79                clEnumValN(OnlyCheapRepl, "cheap",
80                           "only replace exit value when the cost is cheap"),
81                clEnumValN(AlwaysRepl, "always",
82                           "always replace exit value whenever possible"),
83                clEnumValEnd));
84 
85 namespace {
86 struct RewritePhi;
87 
88 class IndVarSimplify {
89   LoopInfo *LI;
90   ScalarEvolution *SE;
91   DominatorTree *DT;
92   const DataLayout &DL;
93   TargetLibraryInfo *TLI;
94   const TargetTransformInfo *TTI;
95 
96   SmallVector<WeakVH, 16> DeadInsts;
97   bool Changed = false;
98 
99   bool isValidRewrite(Value *FromVal, Value *ToVal);
100 
101   void handleFloatingPointIV(Loop *L, PHINode *PH);
102   void rewriteNonIntegerIVs(Loop *L);
103 
104   void simplifyAndExtend(Loop *L, SCEVExpander &Rewriter, LoopInfo *LI);
105 
106   bool canLoopBeDeleted(Loop *L, SmallVector<RewritePhi, 8> &RewritePhiSet);
107   void rewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter);
108   void rewriteFirstIterationLoopExitValues(Loop *L);
109 
110   Value *linearFunctionTestReplace(Loop *L, const SCEV *BackedgeTakenCount,
111                                    PHINode *IndVar, SCEVExpander &Rewriter);
112 
113   void sinkUnusedInvariants(Loop *L);
114 
115   Value *expandSCEVIfNeeded(SCEVExpander &Rewriter, const SCEV *S, Loop *L,
116                             Instruction *InsertPt, Type *Ty);
117 
118 public:
IndVarSimplify(LoopInfo * LI,ScalarEvolution * SE,DominatorTree * DT,const DataLayout & DL,TargetLibraryInfo * TLI,TargetTransformInfo * TTI)119   IndVarSimplify(LoopInfo *LI, ScalarEvolution *SE, DominatorTree *DT,
120                  const DataLayout &DL, TargetLibraryInfo *TLI,
121                  TargetTransformInfo *TTI)
122       : LI(LI), SE(SE), DT(DT), DL(DL), TLI(TLI), TTI(TTI) {}
123 
124   bool run(Loop *L);
125 };
126 }
127 
128 /// Return true if the SCEV expansion generated by the rewriter can replace the
129 /// original value. SCEV guarantees that it produces the same value, but the way
130 /// it is produced may be illegal IR.  Ideally, this function will only be
131 /// called for verification.
isValidRewrite(Value * FromVal,Value * ToVal)132 bool IndVarSimplify::isValidRewrite(Value *FromVal, Value *ToVal) {
133   // If an SCEV expression subsumed multiple pointers, its expansion could
134   // reassociate the GEP changing the base pointer. This is illegal because the
135   // final address produced by a GEP chain must be inbounds relative to its
136   // underlying object. Otherwise basic alias analysis, among other things,
137   // could fail in a dangerous way. Ultimately, SCEV will be improved to avoid
138   // producing an expression involving multiple pointers. Until then, we must
139   // bail out here.
140   //
141   // Retrieve the pointer operand of the GEP. Don't use GetUnderlyingObject
142   // because it understands lcssa phis while SCEV does not.
143   Value *FromPtr = FromVal;
144   Value *ToPtr = ToVal;
145   if (auto *GEP = dyn_cast<GEPOperator>(FromVal)) {
146     FromPtr = GEP->getPointerOperand();
147   }
148   if (auto *GEP = dyn_cast<GEPOperator>(ToVal)) {
149     ToPtr = GEP->getPointerOperand();
150   }
151   if (FromPtr != FromVal || ToPtr != ToVal) {
152     // Quickly check the common case
153     if (FromPtr == ToPtr)
154       return true;
155 
156     // SCEV may have rewritten an expression that produces the GEP's pointer
157     // operand. That's ok as long as the pointer operand has the same base
158     // pointer. Unlike GetUnderlyingObject(), getPointerBase() will find the
159     // base of a recurrence. This handles the case in which SCEV expansion
160     // converts a pointer type recurrence into a nonrecurrent pointer base
161     // indexed by an integer recurrence.
162 
163     // If the GEP base pointer is a vector of pointers, abort.
164     if (!FromPtr->getType()->isPointerTy() || !ToPtr->getType()->isPointerTy())
165       return false;
166 
167     const SCEV *FromBase = SE->getPointerBase(SE->getSCEV(FromPtr));
168     const SCEV *ToBase = SE->getPointerBase(SE->getSCEV(ToPtr));
169     if (FromBase == ToBase)
170       return true;
171 
172     DEBUG(dbgs() << "INDVARS: GEP rewrite bail out "
173           << *FromBase << " != " << *ToBase << "\n");
174 
175     return false;
176   }
177   return true;
178 }
179 
180 /// Determine the insertion point for this user. By default, insert immediately
181 /// before the user. SCEVExpander or LICM will hoist loop invariants out of the
182 /// loop. For PHI nodes, there may be multiple uses, so compute the nearest
183 /// common dominator for the incoming blocks.
getInsertPointForUses(Instruction * User,Value * Def,DominatorTree * DT,LoopInfo * LI)184 static Instruction *getInsertPointForUses(Instruction *User, Value *Def,
185                                           DominatorTree *DT, LoopInfo *LI) {
186   PHINode *PHI = dyn_cast<PHINode>(User);
187   if (!PHI)
188     return User;
189 
190   Instruction *InsertPt = nullptr;
191   for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i) {
192     if (PHI->getIncomingValue(i) != Def)
193       continue;
194 
195     BasicBlock *InsertBB = PHI->getIncomingBlock(i);
196     if (!InsertPt) {
197       InsertPt = InsertBB->getTerminator();
198       continue;
199     }
200     InsertBB = DT->findNearestCommonDominator(InsertPt->getParent(), InsertBB);
201     InsertPt = InsertBB->getTerminator();
202   }
203   assert(InsertPt && "Missing phi operand");
204 
205   auto *DefI = dyn_cast<Instruction>(Def);
206   if (!DefI)
207     return InsertPt;
208 
209   assert(DT->dominates(DefI, InsertPt) && "def does not dominate all uses");
210 
211   auto *L = LI->getLoopFor(DefI->getParent());
212   assert(!L || L->contains(LI->getLoopFor(InsertPt->getParent())));
213 
214   for (auto *DTN = (*DT)[InsertPt->getParent()]; DTN; DTN = DTN->getIDom())
215     if (LI->getLoopFor(DTN->getBlock()) == L)
216       return DTN->getBlock()->getTerminator();
217 
218   llvm_unreachable("DefI dominates InsertPt!");
219 }
220 
221 //===----------------------------------------------------------------------===//
222 // rewriteNonIntegerIVs and helpers. Prefer integer IVs.
223 //===----------------------------------------------------------------------===//
224 
225 /// Convert APF to an integer, if possible.
ConvertToSInt(const APFloat & APF,int64_t & IntVal)226 static bool ConvertToSInt(const APFloat &APF, int64_t &IntVal) {
227   bool isExact = false;
228   // See if we can convert this to an int64_t
229   uint64_t UIntVal;
230   if (APF.convertToInteger(&UIntVal, 64, true, APFloat::rmTowardZero,
231                            &isExact) != APFloat::opOK || !isExact)
232     return false;
233   IntVal = UIntVal;
234   return true;
235 }
236 
237 /// If the loop has floating induction variable then insert corresponding
238 /// integer induction variable if possible.
239 /// For example,
240 /// for(double i = 0; i < 10000; ++i)
241 ///   bar(i)
242 /// is converted into
243 /// for(int i = 0; i < 10000; ++i)
244 ///   bar((double)i);
245 ///
handleFloatingPointIV(Loop * L,PHINode * PN)246 void IndVarSimplify::handleFloatingPointIV(Loop *L, PHINode *PN) {
247   unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
248   unsigned BackEdge     = IncomingEdge^1;
249 
250   // Check incoming value.
251   auto *InitValueVal = dyn_cast<ConstantFP>(PN->getIncomingValue(IncomingEdge));
252 
253   int64_t InitValue;
254   if (!InitValueVal || !ConvertToSInt(InitValueVal->getValueAPF(), InitValue))
255     return;
256 
257   // Check IV increment. Reject this PN if increment operation is not
258   // an add or increment value can not be represented by an integer.
259   auto *Incr = dyn_cast<BinaryOperator>(PN->getIncomingValue(BackEdge));
260   if (Incr == nullptr || Incr->getOpcode() != Instruction::FAdd) return;
261 
262   // If this is not an add of the PHI with a constantfp, or if the constant fp
263   // is not an integer, bail out.
264   ConstantFP *IncValueVal = dyn_cast<ConstantFP>(Incr->getOperand(1));
265   int64_t IncValue;
266   if (IncValueVal == nullptr || Incr->getOperand(0) != PN ||
267       !ConvertToSInt(IncValueVal->getValueAPF(), IncValue))
268     return;
269 
270   // Check Incr uses. One user is PN and the other user is an exit condition
271   // used by the conditional terminator.
272   Value::user_iterator IncrUse = Incr->user_begin();
273   Instruction *U1 = cast<Instruction>(*IncrUse++);
274   if (IncrUse == Incr->user_end()) return;
275   Instruction *U2 = cast<Instruction>(*IncrUse++);
276   if (IncrUse != Incr->user_end()) return;
277 
278   // Find exit condition, which is an fcmp.  If it doesn't exist, or if it isn't
279   // only used by a branch, we can't transform it.
280   FCmpInst *Compare = dyn_cast<FCmpInst>(U1);
281   if (!Compare)
282     Compare = dyn_cast<FCmpInst>(U2);
283   if (!Compare || !Compare->hasOneUse() ||
284       !isa<BranchInst>(Compare->user_back()))
285     return;
286 
287   BranchInst *TheBr = cast<BranchInst>(Compare->user_back());
288 
289   // We need to verify that the branch actually controls the iteration count
290   // of the loop.  If not, the new IV can overflow and no one will notice.
291   // The branch block must be in the loop and one of the successors must be out
292   // of the loop.
293   assert(TheBr->isConditional() && "Can't use fcmp if not conditional");
294   if (!L->contains(TheBr->getParent()) ||
295       (L->contains(TheBr->getSuccessor(0)) &&
296        L->contains(TheBr->getSuccessor(1))))
297     return;
298 
299 
300   // If it isn't a comparison with an integer-as-fp (the exit value), we can't
301   // transform it.
302   ConstantFP *ExitValueVal = dyn_cast<ConstantFP>(Compare->getOperand(1));
303   int64_t ExitValue;
304   if (ExitValueVal == nullptr ||
305       !ConvertToSInt(ExitValueVal->getValueAPF(), ExitValue))
306     return;
307 
308   // Find new predicate for integer comparison.
309   CmpInst::Predicate NewPred = CmpInst::BAD_ICMP_PREDICATE;
310   switch (Compare->getPredicate()) {
311   default: return;  // Unknown comparison.
312   case CmpInst::FCMP_OEQ:
313   case CmpInst::FCMP_UEQ: NewPred = CmpInst::ICMP_EQ; break;
314   case CmpInst::FCMP_ONE:
315   case CmpInst::FCMP_UNE: NewPred = CmpInst::ICMP_NE; break;
316   case CmpInst::FCMP_OGT:
317   case CmpInst::FCMP_UGT: NewPred = CmpInst::ICMP_SGT; break;
318   case CmpInst::FCMP_OGE:
319   case CmpInst::FCMP_UGE: NewPred = CmpInst::ICMP_SGE; break;
320   case CmpInst::FCMP_OLT:
321   case CmpInst::FCMP_ULT: NewPred = CmpInst::ICMP_SLT; break;
322   case CmpInst::FCMP_OLE:
323   case CmpInst::FCMP_ULE: NewPred = CmpInst::ICMP_SLE; break;
324   }
325 
326   // We convert the floating point induction variable to a signed i32 value if
327   // we can.  This is only safe if the comparison will not overflow in a way
328   // that won't be trapped by the integer equivalent operations.  Check for this
329   // now.
330   // TODO: We could use i64 if it is native and the range requires it.
331 
332   // The start/stride/exit values must all fit in signed i32.
333   if (!isInt<32>(InitValue) || !isInt<32>(IncValue) || !isInt<32>(ExitValue))
334     return;
335 
336   // If not actually striding (add x, 0.0), avoid touching the code.
337   if (IncValue == 0)
338     return;
339 
340   // Positive and negative strides have different safety conditions.
341   if (IncValue > 0) {
342     // If we have a positive stride, we require the init to be less than the
343     // exit value.
344     if (InitValue >= ExitValue)
345       return;
346 
347     uint32_t Range = uint32_t(ExitValue-InitValue);
348     // Check for infinite loop, either:
349     // while (i <= Exit) or until (i > Exit)
350     if (NewPred == CmpInst::ICMP_SLE || NewPred == CmpInst::ICMP_SGT) {
351       if (++Range == 0) return;  // Range overflows.
352     }
353 
354     unsigned Leftover = Range % uint32_t(IncValue);
355 
356     // If this is an equality comparison, we require that the strided value
357     // exactly land on the exit value, otherwise the IV condition will wrap
358     // around and do things the fp IV wouldn't.
359     if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
360         Leftover != 0)
361       return;
362 
363     // If the stride would wrap around the i32 before exiting, we can't
364     // transform the IV.
365     if (Leftover != 0 && int32_t(ExitValue+IncValue) < ExitValue)
366       return;
367 
368   } else {
369     // If we have a negative stride, we require the init to be greater than the
370     // exit value.
371     if (InitValue <= ExitValue)
372       return;
373 
374     uint32_t Range = uint32_t(InitValue-ExitValue);
375     // Check for infinite loop, either:
376     // while (i >= Exit) or until (i < Exit)
377     if (NewPred == CmpInst::ICMP_SGE || NewPred == CmpInst::ICMP_SLT) {
378       if (++Range == 0) return;  // Range overflows.
379     }
380 
381     unsigned Leftover = Range % uint32_t(-IncValue);
382 
383     // If this is an equality comparison, we require that the strided value
384     // exactly land on the exit value, otherwise the IV condition will wrap
385     // around and do things the fp IV wouldn't.
386     if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
387         Leftover != 0)
388       return;
389 
390     // If the stride would wrap around the i32 before exiting, we can't
391     // transform the IV.
392     if (Leftover != 0 && int32_t(ExitValue+IncValue) > ExitValue)
393       return;
394   }
395 
396   IntegerType *Int32Ty = Type::getInt32Ty(PN->getContext());
397 
398   // Insert new integer induction variable.
399   PHINode *NewPHI = PHINode::Create(Int32Ty, 2, PN->getName()+".int", PN);
400   NewPHI->addIncoming(ConstantInt::get(Int32Ty, InitValue),
401                       PN->getIncomingBlock(IncomingEdge));
402 
403   Value *NewAdd =
404     BinaryOperator::CreateAdd(NewPHI, ConstantInt::get(Int32Ty, IncValue),
405                               Incr->getName()+".int", Incr);
406   NewPHI->addIncoming(NewAdd, PN->getIncomingBlock(BackEdge));
407 
408   ICmpInst *NewCompare = new ICmpInst(TheBr, NewPred, NewAdd,
409                                       ConstantInt::get(Int32Ty, ExitValue),
410                                       Compare->getName());
411 
412   // In the following deletions, PN may become dead and may be deleted.
413   // Use a WeakVH to observe whether this happens.
414   WeakVH WeakPH = PN;
415 
416   // Delete the old floating point exit comparison.  The branch starts using the
417   // new comparison.
418   NewCompare->takeName(Compare);
419   Compare->replaceAllUsesWith(NewCompare);
420   RecursivelyDeleteTriviallyDeadInstructions(Compare, TLI);
421 
422   // Delete the old floating point increment.
423   Incr->replaceAllUsesWith(UndefValue::get(Incr->getType()));
424   RecursivelyDeleteTriviallyDeadInstructions(Incr, TLI);
425 
426   // If the FP induction variable still has uses, this is because something else
427   // in the loop uses its value.  In order to canonicalize the induction
428   // variable, we chose to eliminate the IV and rewrite it in terms of an
429   // int->fp cast.
430   //
431   // We give preference to sitofp over uitofp because it is faster on most
432   // platforms.
433   if (WeakPH) {
434     Value *Conv = new SIToFPInst(NewPHI, PN->getType(), "indvar.conv",
435                                  &*PN->getParent()->getFirstInsertionPt());
436     PN->replaceAllUsesWith(Conv);
437     RecursivelyDeleteTriviallyDeadInstructions(PN, TLI);
438   }
439   Changed = true;
440 }
441 
rewriteNonIntegerIVs(Loop * L)442 void IndVarSimplify::rewriteNonIntegerIVs(Loop *L) {
443   // First step.  Check to see if there are any floating-point recurrences.
444   // If there are, change them into integer recurrences, permitting analysis by
445   // the SCEV routines.
446   //
447   BasicBlock *Header = L->getHeader();
448 
449   SmallVector<WeakVH, 8> PHIs;
450   for (BasicBlock::iterator I = Header->begin();
451        PHINode *PN = dyn_cast<PHINode>(I); ++I)
452     PHIs.push_back(PN);
453 
454   for (unsigned i = 0, e = PHIs.size(); i != e; ++i)
455     if (PHINode *PN = dyn_cast_or_null<PHINode>(&*PHIs[i]))
456       handleFloatingPointIV(L, PN);
457 
458   // If the loop previously had floating-point IV, ScalarEvolution
459   // may not have been able to compute a trip count. Now that we've done some
460   // re-writing, the trip count may be computable.
461   if (Changed)
462     SE->forgetLoop(L);
463 }
464 
465 namespace {
466 // Collect information about PHI nodes which can be transformed in
467 // rewriteLoopExitValues.
468 struct RewritePhi {
469   PHINode *PN;
470   unsigned Ith;  // Ith incoming value.
471   Value *Val;    // Exit value after expansion.
472   bool HighCost; // High Cost when expansion.
473 
RewritePhi__anon49122e510211::RewritePhi474   RewritePhi(PHINode *P, unsigned I, Value *V, bool H)
475       : PN(P), Ith(I), Val(V), HighCost(H) {}
476 };
477 }
478 
expandSCEVIfNeeded(SCEVExpander & Rewriter,const SCEV * S,Loop * L,Instruction * InsertPt,Type * ResultTy)479 Value *IndVarSimplify::expandSCEVIfNeeded(SCEVExpander &Rewriter, const SCEV *S,
480                                           Loop *L, Instruction *InsertPt,
481                                           Type *ResultTy) {
482   // Before expanding S into an expensive LLVM expression, see if we can use an
483   // already existing value as the expansion for S.
484   if (Value *ExistingValue = Rewriter.findExistingExpansion(S, InsertPt, L))
485     if (ExistingValue->getType() == ResultTy)
486       return ExistingValue;
487 
488   // We didn't find anything, fall back to using SCEVExpander.
489   return Rewriter.expandCodeFor(S, ResultTy, InsertPt);
490 }
491 
492 //===----------------------------------------------------------------------===//
493 // rewriteLoopExitValues - Optimize IV users outside the loop.
494 // As a side effect, reduces the amount of IV processing within the loop.
495 //===----------------------------------------------------------------------===//
496 
497 /// Check to see if this loop has a computable loop-invariant execution count.
498 /// If so, this means that we can compute the final value of any expressions
499 /// that are recurrent in the loop, and substitute the exit values from the loop
500 /// into any instructions outside of the loop that use the final values of the
501 /// current expressions.
502 ///
503 /// This is mostly redundant with the regular IndVarSimplify activities that
504 /// happen later, except that it's more powerful in some cases, because it's
505 /// able to brute-force evaluate arbitrary instructions as long as they have
506 /// constant operands at the beginning of the loop.
rewriteLoopExitValues(Loop * L,SCEVExpander & Rewriter)507 void IndVarSimplify::rewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter) {
508   // Check a pre-condition.
509   assert(L->isRecursivelyLCSSAForm(*DT) && "Indvars did not preserve LCSSA!");
510 
511   SmallVector<BasicBlock*, 8> ExitBlocks;
512   L->getUniqueExitBlocks(ExitBlocks);
513 
514   SmallVector<RewritePhi, 8> RewritePhiSet;
515   // Find all values that are computed inside the loop, but used outside of it.
516   // Because of LCSSA, these values will only occur in LCSSA PHI Nodes.  Scan
517   // the exit blocks of the loop to find them.
518   for (BasicBlock *ExitBB : ExitBlocks) {
519     // If there are no PHI nodes in this exit block, then no values defined
520     // inside the loop are used on this path, skip it.
521     PHINode *PN = dyn_cast<PHINode>(ExitBB->begin());
522     if (!PN) continue;
523 
524     unsigned NumPreds = PN->getNumIncomingValues();
525 
526     // Iterate over all of the PHI nodes.
527     BasicBlock::iterator BBI = ExitBB->begin();
528     while ((PN = dyn_cast<PHINode>(BBI++))) {
529       if (PN->use_empty())
530         continue; // dead use, don't replace it
531 
532       if (!SE->isSCEVable(PN->getType()))
533         continue;
534 
535       // It's necessary to tell ScalarEvolution about this explicitly so that
536       // it can walk the def-use list and forget all SCEVs, as it may not be
537       // watching the PHI itself. Once the new exit value is in place, there
538       // may not be a def-use connection between the loop and every instruction
539       // which got a SCEVAddRecExpr for that loop.
540       SE->forgetValue(PN);
541 
542       // Iterate over all of the values in all the PHI nodes.
543       for (unsigned i = 0; i != NumPreds; ++i) {
544         // If the value being merged in is not integer or is not defined
545         // in the loop, skip it.
546         Value *InVal = PN->getIncomingValue(i);
547         if (!isa<Instruction>(InVal))
548           continue;
549 
550         // If this pred is for a subloop, not L itself, skip it.
551         if (LI->getLoopFor(PN->getIncomingBlock(i)) != L)
552           continue; // The Block is in a subloop, skip it.
553 
554         // Check that InVal is defined in the loop.
555         Instruction *Inst = cast<Instruction>(InVal);
556         if (!L->contains(Inst))
557           continue;
558 
559         // Okay, this instruction has a user outside of the current loop
560         // and varies predictably *inside* the loop.  Evaluate the value it
561         // contains when the loop exits, if possible.
562         const SCEV *ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop());
563         if (!SE->isLoopInvariant(ExitValue, L) ||
564             !isSafeToExpand(ExitValue, *SE))
565           continue;
566 
567         // Computing the value outside of the loop brings no benefit if :
568         //  - it is definitely used inside the loop in a way which can not be
569         //    optimized away.
570         //  - no use outside of the loop can take advantage of hoisting the
571         //    computation out of the loop
572         if (ExitValue->getSCEVType()>=scMulExpr) {
573           unsigned NumHardInternalUses = 0;
574           unsigned NumSoftExternalUses = 0;
575           unsigned NumUses = 0;
576           for (auto IB = Inst->user_begin(), IE = Inst->user_end();
577                IB != IE && NumUses <= 6; ++IB) {
578             Instruction *UseInstr = cast<Instruction>(*IB);
579             unsigned Opc = UseInstr->getOpcode();
580             NumUses++;
581             if (L->contains(UseInstr)) {
582               if (Opc == Instruction::Call || Opc == Instruction::Ret)
583                 NumHardInternalUses++;
584             } else {
585               if (Opc == Instruction::PHI) {
586                 // Do not count the Phi as a use. LCSSA may have inserted
587                 // plenty of trivial ones.
588                 NumUses--;
589                 for (auto PB = UseInstr->user_begin(),
590                           PE = UseInstr->user_end();
591                      PB != PE && NumUses <= 6; ++PB, ++NumUses) {
592                   unsigned PhiOpc = cast<Instruction>(*PB)->getOpcode();
593                   if (PhiOpc != Instruction::Call && PhiOpc != Instruction::Ret)
594                     NumSoftExternalUses++;
595                 }
596                 continue;
597               }
598               if (Opc != Instruction::Call && Opc != Instruction::Ret)
599                 NumSoftExternalUses++;
600             }
601           }
602           if (NumUses <= 6 && NumHardInternalUses && !NumSoftExternalUses)
603             continue;
604         }
605 
606         bool HighCost = Rewriter.isHighCostExpansion(ExitValue, L, Inst);
607         Value *ExitVal =
608             expandSCEVIfNeeded(Rewriter, ExitValue, L, Inst, PN->getType());
609 
610         DEBUG(dbgs() << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal << '\n'
611                      << "  LoopVal = " << *Inst << "\n");
612 
613         if (!isValidRewrite(Inst, ExitVal)) {
614           DeadInsts.push_back(ExitVal);
615           continue;
616         }
617 
618         // Collect all the candidate PHINodes to be rewritten.
619         RewritePhiSet.emplace_back(PN, i, ExitVal, HighCost);
620       }
621     }
622   }
623 
624   bool LoopCanBeDel = canLoopBeDeleted(L, RewritePhiSet);
625 
626   // Transformation.
627   for (const RewritePhi &Phi : RewritePhiSet) {
628     PHINode *PN = Phi.PN;
629     Value *ExitVal = Phi.Val;
630 
631     // Only do the rewrite when the ExitValue can be expanded cheaply.
632     // If LoopCanBeDel is true, rewrite exit value aggressively.
633     if (ReplaceExitValue == OnlyCheapRepl && !LoopCanBeDel && Phi.HighCost) {
634       DeadInsts.push_back(ExitVal);
635       continue;
636     }
637 
638     Changed = true;
639     ++NumReplaced;
640     Instruction *Inst = cast<Instruction>(PN->getIncomingValue(Phi.Ith));
641     PN->setIncomingValue(Phi.Ith, ExitVal);
642 
643     // If this instruction is dead now, delete it. Don't do it now to avoid
644     // invalidating iterators.
645     if (isInstructionTriviallyDead(Inst, TLI))
646       DeadInsts.push_back(Inst);
647 
648     // Replace PN with ExitVal if that is legal and does not break LCSSA.
649     if (PN->getNumIncomingValues() == 1 &&
650         LI->replacementPreservesLCSSAForm(PN, ExitVal)) {
651       PN->replaceAllUsesWith(ExitVal);
652       PN->eraseFromParent();
653     }
654   }
655 
656   // The insertion point instruction may have been deleted; clear it out
657   // so that the rewriter doesn't trip over it later.
658   Rewriter.clearInsertPoint();
659 }
660 
661 //===---------------------------------------------------------------------===//
662 // rewriteFirstIterationLoopExitValues: Rewrite loop exit values if we know
663 // they will exit at the first iteration.
664 //===---------------------------------------------------------------------===//
665 
666 /// Check to see if this loop has loop invariant conditions which lead to loop
667 /// exits. If so, we know that if the exit path is taken, it is at the first
668 /// loop iteration. This lets us predict exit values of PHI nodes that live in
669 /// loop header.
rewriteFirstIterationLoopExitValues(Loop * L)670 void IndVarSimplify::rewriteFirstIterationLoopExitValues(Loop *L) {
671   // Verify the input to the pass is already in LCSSA form.
672   assert(L->isLCSSAForm(*DT));
673 
674   SmallVector<BasicBlock *, 8> ExitBlocks;
675   L->getUniqueExitBlocks(ExitBlocks);
676   auto *LoopHeader = L->getHeader();
677   assert(LoopHeader && "Invalid loop");
678 
679   for (auto *ExitBB : ExitBlocks) {
680     BasicBlock::iterator BBI = ExitBB->begin();
681     // If there are no more PHI nodes in this exit block, then no more
682     // values defined inside the loop are used on this path.
683     while (auto *PN = dyn_cast<PHINode>(BBI++)) {
684       for (unsigned IncomingValIdx = 0, E = PN->getNumIncomingValues();
685           IncomingValIdx != E; ++IncomingValIdx) {
686         auto *IncomingBB = PN->getIncomingBlock(IncomingValIdx);
687 
688         // We currently only support loop exits from loop header. If the
689         // incoming block is not loop header, we need to recursively check
690         // all conditions starting from loop header are loop invariants.
691         // Additional support might be added in the future.
692         if (IncomingBB != LoopHeader)
693           continue;
694 
695         // Get condition that leads to the exit path.
696         auto *TermInst = IncomingBB->getTerminator();
697 
698         Value *Cond = nullptr;
699         if (auto *BI = dyn_cast<BranchInst>(TermInst)) {
700           // Must be a conditional branch, otherwise the block
701           // should not be in the loop.
702           Cond = BI->getCondition();
703         } else if (auto *SI = dyn_cast<SwitchInst>(TermInst))
704           Cond = SI->getCondition();
705         else
706           continue;
707 
708         if (!L->isLoopInvariant(Cond))
709           continue;
710 
711         auto *ExitVal =
712             dyn_cast<PHINode>(PN->getIncomingValue(IncomingValIdx));
713 
714         // Only deal with PHIs.
715         if (!ExitVal)
716           continue;
717 
718         // If ExitVal is a PHI on the loop header, then we know its
719         // value along this exit because the exit can only be taken
720         // on the first iteration.
721         auto *LoopPreheader = L->getLoopPreheader();
722         assert(LoopPreheader && "Invalid loop");
723         int PreheaderIdx = ExitVal->getBasicBlockIndex(LoopPreheader);
724         if (PreheaderIdx != -1) {
725           assert(ExitVal->getParent() == LoopHeader &&
726                  "ExitVal must be in loop header");
727           PN->setIncomingValue(IncomingValIdx,
728               ExitVal->getIncomingValue(PreheaderIdx));
729         }
730       }
731     }
732   }
733 }
734 
735 /// Check whether it is possible to delete the loop after rewriting exit
736 /// value. If it is possible, ignore ReplaceExitValue and do rewriting
737 /// aggressively.
canLoopBeDeleted(Loop * L,SmallVector<RewritePhi,8> & RewritePhiSet)738 bool IndVarSimplify::canLoopBeDeleted(
739     Loop *L, SmallVector<RewritePhi, 8> &RewritePhiSet) {
740 
741   BasicBlock *Preheader = L->getLoopPreheader();
742   // If there is no preheader, the loop will not be deleted.
743   if (!Preheader)
744     return false;
745 
746   // In LoopDeletion pass Loop can be deleted when ExitingBlocks.size() > 1.
747   // We obviate multiple ExitingBlocks case for simplicity.
748   // TODO: If we see testcase with multiple ExitingBlocks can be deleted
749   // after exit value rewriting, we can enhance the logic here.
750   SmallVector<BasicBlock *, 4> ExitingBlocks;
751   L->getExitingBlocks(ExitingBlocks);
752   SmallVector<BasicBlock *, 8> ExitBlocks;
753   L->getUniqueExitBlocks(ExitBlocks);
754   if (ExitBlocks.size() > 1 || ExitingBlocks.size() > 1)
755     return false;
756 
757   BasicBlock *ExitBlock = ExitBlocks[0];
758   BasicBlock::iterator BI = ExitBlock->begin();
759   while (PHINode *P = dyn_cast<PHINode>(BI)) {
760     Value *Incoming = P->getIncomingValueForBlock(ExitingBlocks[0]);
761 
762     // If the Incoming value of P is found in RewritePhiSet, we know it
763     // could be rewritten to use a loop invariant value in transformation
764     // phase later. Skip it in the loop invariant check below.
765     bool found = false;
766     for (const RewritePhi &Phi : RewritePhiSet) {
767       unsigned i = Phi.Ith;
768       if (Phi.PN == P && (Phi.PN)->getIncomingValue(i) == Incoming) {
769         found = true;
770         break;
771       }
772     }
773 
774     Instruction *I;
775     if (!found && (I = dyn_cast<Instruction>(Incoming)))
776       if (!L->hasLoopInvariantOperands(I))
777         return false;
778 
779     ++BI;
780   }
781 
782   for (auto *BB : L->blocks())
783     if (any_of(*BB, [](Instruction &I) { return I.mayHaveSideEffects(); }))
784       return false;
785 
786   return true;
787 }
788 
789 //===----------------------------------------------------------------------===//
790 //  IV Widening - Extend the width of an IV to cover its widest uses.
791 //===----------------------------------------------------------------------===//
792 
793 namespace {
794 // Collect information about induction variables that are used by sign/zero
795 // extend operations. This information is recorded by CollectExtend and provides
796 // the input to WidenIV.
797 struct WideIVInfo {
798   PHINode *NarrowIV = nullptr;
799   Type *WidestNativeType = nullptr; // Widest integer type created [sz]ext
800   bool IsSigned = false;            // Was a sext user seen before a zext?
801 };
802 }
803 
804 /// Update information about the induction variable that is extended by this
805 /// sign or zero extend operation. This is used to determine the final width of
806 /// the IV before actually widening it.
visitIVCast(CastInst * Cast,WideIVInfo & WI,ScalarEvolution * SE,const TargetTransformInfo * TTI)807 static void visitIVCast(CastInst *Cast, WideIVInfo &WI, ScalarEvolution *SE,
808                         const TargetTransformInfo *TTI) {
809   bool IsSigned = Cast->getOpcode() == Instruction::SExt;
810   if (!IsSigned && Cast->getOpcode() != Instruction::ZExt)
811     return;
812 
813   Type *Ty = Cast->getType();
814   uint64_t Width = SE->getTypeSizeInBits(Ty);
815   if (!Cast->getModule()->getDataLayout().isLegalInteger(Width))
816     return;
817 
818   // Cast is either an sext or zext up to this point.
819   // We should not widen an indvar if arithmetics on the wider indvar are more
820   // expensive than those on the narrower indvar. We check only the cost of ADD
821   // because at least an ADD is required to increment the induction variable. We
822   // could compute more comprehensively the cost of all instructions on the
823   // induction variable when necessary.
824   if (TTI &&
825       TTI->getArithmeticInstrCost(Instruction::Add, Ty) >
826           TTI->getArithmeticInstrCost(Instruction::Add,
827                                       Cast->getOperand(0)->getType())) {
828     return;
829   }
830 
831   if (!WI.WidestNativeType) {
832     WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
833     WI.IsSigned = IsSigned;
834     return;
835   }
836 
837   // We extend the IV to satisfy the sign of its first user, arbitrarily.
838   if (WI.IsSigned != IsSigned)
839     return;
840 
841   if (Width > SE->getTypeSizeInBits(WI.WidestNativeType))
842     WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
843 }
844 
845 namespace {
846 
847 /// Record a link in the Narrow IV def-use chain along with the WideIV that
848 /// computes the same value as the Narrow IV def.  This avoids caching Use*
849 /// pointers.
850 struct NarrowIVDefUse {
851   Instruction *NarrowDef = nullptr;
852   Instruction *NarrowUse = nullptr;
853   Instruction *WideDef = nullptr;
854 
855   // True if the narrow def is never negative.  Tracking this information lets
856   // us use a sign extension instead of a zero extension or vice versa, when
857   // profitable and legal.
858   bool NeverNegative = false;
859 
NarrowIVDefUse__anon49122e510511::NarrowIVDefUse860   NarrowIVDefUse(Instruction *ND, Instruction *NU, Instruction *WD,
861                  bool NeverNegative)
862       : NarrowDef(ND), NarrowUse(NU), WideDef(WD),
863         NeverNegative(NeverNegative) {}
864 };
865 
866 /// The goal of this transform is to remove sign and zero extends without
867 /// creating any new induction variables. To do this, it creates a new phi of
868 /// the wider type and redirects all users, either removing extends or inserting
869 /// truncs whenever we stop propagating the type.
870 ///
871 class WidenIV {
872   // Parameters
873   PHINode *OrigPhi;
874   Type *WideType;
875   bool IsSigned;
876 
877   // Context
878   LoopInfo        *LI;
879   Loop            *L;
880   ScalarEvolution *SE;
881   DominatorTree   *DT;
882 
883   // Result
884   PHINode *WidePhi;
885   Instruction *WideInc;
886   const SCEV *WideIncExpr;
887   SmallVectorImpl<WeakVH> &DeadInsts;
888 
889   SmallPtrSet<Instruction*,16> Widened;
890   SmallVector<NarrowIVDefUse, 8> NarrowIVUsers;
891 
892 public:
WidenIV(const WideIVInfo & WI,LoopInfo * LInfo,ScalarEvolution * SEv,DominatorTree * DTree,SmallVectorImpl<WeakVH> & DI)893   WidenIV(const WideIVInfo &WI, LoopInfo *LInfo,
894           ScalarEvolution *SEv, DominatorTree *DTree,
895           SmallVectorImpl<WeakVH> &DI) :
896     OrigPhi(WI.NarrowIV),
897     WideType(WI.WidestNativeType),
898     IsSigned(WI.IsSigned),
899     LI(LInfo),
900     L(LI->getLoopFor(OrigPhi->getParent())),
901     SE(SEv),
902     DT(DTree),
903     WidePhi(nullptr),
904     WideInc(nullptr),
905     WideIncExpr(nullptr),
906     DeadInsts(DI) {
907     assert(L->getHeader() == OrigPhi->getParent() && "Phi must be an IV");
908   }
909 
910   PHINode *createWideIV(SCEVExpander &Rewriter);
911 
912 protected:
913   Value *createExtendInst(Value *NarrowOper, Type *WideType, bool IsSigned,
914                           Instruction *Use);
915 
916   Instruction *cloneIVUser(NarrowIVDefUse DU, const SCEVAddRecExpr *WideAR);
917   Instruction *cloneArithmeticIVUser(NarrowIVDefUse DU,
918                                      const SCEVAddRecExpr *WideAR);
919   Instruction *cloneBitwiseIVUser(NarrowIVDefUse DU);
920 
921   const SCEVAddRecExpr *getWideRecurrence(Instruction *NarrowUse);
922 
923   const SCEVAddRecExpr* getExtendedOperandRecurrence(NarrowIVDefUse DU);
924 
925   const SCEV *getSCEVByOpCode(const SCEV *LHS, const SCEV *RHS,
926                               unsigned OpCode) const;
927 
928   Instruction *widenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter);
929 
930   bool widenLoopCompare(NarrowIVDefUse DU);
931 
932   void pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef);
933 };
934 } // anonymous namespace
935 
936 /// Perform a quick domtree based check for loop invariance assuming that V is
937 /// used within the loop. LoopInfo::isLoopInvariant() seems gratuitous for this
938 /// purpose.
isLoopInvariant(Value * V,const Loop * L,const DominatorTree * DT)939 static bool isLoopInvariant(Value *V, const Loop *L, const DominatorTree *DT) {
940   Instruction *Inst = dyn_cast<Instruction>(V);
941   if (!Inst)
942     return true;
943 
944   return DT->properlyDominates(Inst->getParent(), L->getHeader());
945 }
946 
createExtendInst(Value * NarrowOper,Type * WideType,bool IsSigned,Instruction * Use)947 Value *WidenIV::createExtendInst(Value *NarrowOper, Type *WideType,
948                                  bool IsSigned, Instruction *Use) {
949   // Set the debug location and conservative insertion point.
950   IRBuilder<> Builder(Use);
951   // Hoist the insertion point into loop preheaders as far as possible.
952   for (const Loop *L = LI->getLoopFor(Use->getParent());
953        L && L->getLoopPreheader() && isLoopInvariant(NarrowOper, L, DT);
954        L = L->getParentLoop())
955     Builder.SetInsertPoint(L->getLoopPreheader()->getTerminator());
956 
957   return IsSigned ? Builder.CreateSExt(NarrowOper, WideType) :
958                     Builder.CreateZExt(NarrowOper, WideType);
959 }
960 
961 /// Instantiate a wide operation to replace a narrow operation. This only needs
962 /// to handle operations that can evaluation to SCEVAddRec. It can safely return
963 /// 0 for any operation we decide not to clone.
cloneIVUser(NarrowIVDefUse DU,const SCEVAddRecExpr * WideAR)964 Instruction *WidenIV::cloneIVUser(NarrowIVDefUse DU,
965                                   const SCEVAddRecExpr *WideAR) {
966   unsigned Opcode = DU.NarrowUse->getOpcode();
967   switch (Opcode) {
968   default:
969     return nullptr;
970   case Instruction::Add:
971   case Instruction::Mul:
972   case Instruction::UDiv:
973   case Instruction::Sub:
974     return cloneArithmeticIVUser(DU, WideAR);
975 
976   case Instruction::And:
977   case Instruction::Or:
978   case Instruction::Xor:
979   case Instruction::Shl:
980   case Instruction::LShr:
981   case Instruction::AShr:
982     return cloneBitwiseIVUser(DU);
983   }
984 }
985 
cloneBitwiseIVUser(NarrowIVDefUse DU)986 Instruction *WidenIV::cloneBitwiseIVUser(NarrowIVDefUse DU) {
987   Instruction *NarrowUse = DU.NarrowUse;
988   Instruction *NarrowDef = DU.NarrowDef;
989   Instruction *WideDef = DU.WideDef;
990 
991   DEBUG(dbgs() << "Cloning bitwise IVUser: " << *NarrowUse << "\n");
992 
993   // Replace NarrowDef operands with WideDef. Otherwise, we don't know anything
994   // about the narrow operand yet so must insert a [sz]ext. It is probably loop
995   // invariant and will be folded or hoisted. If it actually comes from a
996   // widened IV, it should be removed during a future call to widenIVUse.
997   Value *LHS = (NarrowUse->getOperand(0) == NarrowDef)
998                    ? WideDef
999                    : createExtendInst(NarrowUse->getOperand(0), WideType,
1000                                       IsSigned, NarrowUse);
1001   Value *RHS = (NarrowUse->getOperand(1) == NarrowDef)
1002                    ? WideDef
1003                    : createExtendInst(NarrowUse->getOperand(1), WideType,
1004                                       IsSigned, NarrowUse);
1005 
1006   auto *NarrowBO = cast<BinaryOperator>(NarrowUse);
1007   auto *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(), LHS, RHS,
1008                                         NarrowBO->getName());
1009   IRBuilder<> Builder(NarrowUse);
1010   Builder.Insert(WideBO);
1011   WideBO->copyIRFlags(NarrowBO);
1012   return WideBO;
1013 }
1014 
cloneArithmeticIVUser(NarrowIVDefUse DU,const SCEVAddRecExpr * WideAR)1015 Instruction *WidenIV::cloneArithmeticIVUser(NarrowIVDefUse DU,
1016                                             const SCEVAddRecExpr *WideAR) {
1017   Instruction *NarrowUse = DU.NarrowUse;
1018   Instruction *NarrowDef = DU.NarrowDef;
1019   Instruction *WideDef = DU.WideDef;
1020 
1021   DEBUG(dbgs() << "Cloning arithmetic IVUser: " << *NarrowUse << "\n");
1022 
1023   unsigned IVOpIdx = (NarrowUse->getOperand(0) == NarrowDef) ? 0 : 1;
1024 
1025   // We're trying to find X such that
1026   //
1027   //  Widen(NarrowDef `op` NonIVNarrowDef) == WideAR == WideDef `op.wide` X
1028   //
1029   // We guess two solutions to X, sext(NonIVNarrowDef) and zext(NonIVNarrowDef),
1030   // and check using SCEV if any of them are correct.
1031 
1032   // Returns true if extending NonIVNarrowDef according to `SignExt` is a
1033   // correct solution to X.
1034   auto GuessNonIVOperand = [&](bool SignExt) {
1035     const SCEV *WideLHS;
1036     const SCEV *WideRHS;
1037 
1038     auto GetExtend = [this, SignExt](const SCEV *S, Type *Ty) {
1039       if (SignExt)
1040         return SE->getSignExtendExpr(S, Ty);
1041       return SE->getZeroExtendExpr(S, Ty);
1042     };
1043 
1044     if (IVOpIdx == 0) {
1045       WideLHS = SE->getSCEV(WideDef);
1046       const SCEV *NarrowRHS = SE->getSCEV(NarrowUse->getOperand(1));
1047       WideRHS = GetExtend(NarrowRHS, WideType);
1048     } else {
1049       const SCEV *NarrowLHS = SE->getSCEV(NarrowUse->getOperand(0));
1050       WideLHS = GetExtend(NarrowLHS, WideType);
1051       WideRHS = SE->getSCEV(WideDef);
1052     }
1053 
1054     // WideUse is "WideDef `op.wide` X" as described in the comment.
1055     const SCEV *WideUse = nullptr;
1056 
1057     switch (NarrowUse->getOpcode()) {
1058     default:
1059       llvm_unreachable("No other possibility!");
1060 
1061     case Instruction::Add:
1062       WideUse = SE->getAddExpr(WideLHS, WideRHS);
1063       break;
1064 
1065     case Instruction::Mul:
1066       WideUse = SE->getMulExpr(WideLHS, WideRHS);
1067       break;
1068 
1069     case Instruction::UDiv:
1070       WideUse = SE->getUDivExpr(WideLHS, WideRHS);
1071       break;
1072 
1073     case Instruction::Sub:
1074       WideUse = SE->getMinusSCEV(WideLHS, WideRHS);
1075       break;
1076     }
1077 
1078     return WideUse == WideAR;
1079   };
1080 
1081   bool SignExtend = IsSigned;
1082   if (!GuessNonIVOperand(SignExtend)) {
1083     SignExtend = !SignExtend;
1084     if (!GuessNonIVOperand(SignExtend))
1085       return nullptr;
1086   }
1087 
1088   Value *LHS = (NarrowUse->getOperand(0) == NarrowDef)
1089                    ? WideDef
1090                    : createExtendInst(NarrowUse->getOperand(0), WideType,
1091                                       SignExtend, NarrowUse);
1092   Value *RHS = (NarrowUse->getOperand(1) == NarrowDef)
1093                    ? WideDef
1094                    : createExtendInst(NarrowUse->getOperand(1), WideType,
1095                                       SignExtend, NarrowUse);
1096 
1097   auto *NarrowBO = cast<BinaryOperator>(NarrowUse);
1098   auto *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(), LHS, RHS,
1099                                         NarrowBO->getName());
1100 
1101   IRBuilder<> Builder(NarrowUse);
1102   Builder.Insert(WideBO);
1103   WideBO->copyIRFlags(NarrowBO);
1104   return WideBO;
1105 }
1106 
getSCEVByOpCode(const SCEV * LHS,const SCEV * RHS,unsigned OpCode) const1107 const SCEV *WidenIV::getSCEVByOpCode(const SCEV *LHS, const SCEV *RHS,
1108                                      unsigned OpCode) const {
1109   if (OpCode == Instruction::Add)
1110     return SE->getAddExpr(LHS, RHS);
1111   if (OpCode == Instruction::Sub)
1112     return SE->getMinusSCEV(LHS, RHS);
1113   if (OpCode == Instruction::Mul)
1114     return SE->getMulExpr(LHS, RHS);
1115 
1116   llvm_unreachable("Unsupported opcode.");
1117 }
1118 
1119 /// No-wrap operations can transfer sign extension of their result to their
1120 /// operands. Generate the SCEV value for the widened operation without
1121 /// actually modifying the IR yet. If the expression after extending the
1122 /// operands is an AddRec for this loop, return it.
getExtendedOperandRecurrence(NarrowIVDefUse DU)1123 const SCEVAddRecExpr* WidenIV::getExtendedOperandRecurrence(NarrowIVDefUse DU) {
1124 
1125   // Handle the common case of add<nsw/nuw>
1126   const unsigned OpCode = DU.NarrowUse->getOpcode();
1127   // Only Add/Sub/Mul instructions supported yet.
1128   if (OpCode != Instruction::Add && OpCode != Instruction::Sub &&
1129       OpCode != Instruction::Mul)
1130     return nullptr;
1131 
1132   // One operand (NarrowDef) has already been extended to WideDef. Now determine
1133   // if extending the other will lead to a recurrence.
1134   const unsigned ExtendOperIdx =
1135       DU.NarrowUse->getOperand(0) == DU.NarrowDef ? 1 : 0;
1136   assert(DU.NarrowUse->getOperand(1-ExtendOperIdx) == DU.NarrowDef && "bad DU");
1137 
1138   const SCEV *ExtendOperExpr = nullptr;
1139   const OverflowingBinaryOperator *OBO =
1140     cast<OverflowingBinaryOperator>(DU.NarrowUse);
1141   if (IsSigned && OBO->hasNoSignedWrap())
1142     ExtendOperExpr = SE->getSignExtendExpr(
1143       SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType);
1144   else if(!IsSigned && OBO->hasNoUnsignedWrap())
1145     ExtendOperExpr = SE->getZeroExtendExpr(
1146       SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType);
1147   else
1148     return nullptr;
1149 
1150   // When creating this SCEV expr, don't apply the current operations NSW or NUW
1151   // flags. This instruction may be guarded by control flow that the no-wrap
1152   // behavior depends on. Non-control-equivalent instructions can be mapped to
1153   // the same SCEV expression, and it would be incorrect to transfer NSW/NUW
1154   // semantics to those operations.
1155   const SCEV *lhs = SE->getSCEV(DU.WideDef);
1156   const SCEV *rhs = ExtendOperExpr;
1157 
1158   // Let's swap operands to the initial order for the case of non-commutative
1159   // operations, like SUB. See PR21014.
1160   if (ExtendOperIdx == 0)
1161     std::swap(lhs, rhs);
1162   const SCEVAddRecExpr *AddRec =
1163       dyn_cast<SCEVAddRecExpr>(getSCEVByOpCode(lhs, rhs, OpCode));
1164 
1165   if (!AddRec || AddRec->getLoop() != L)
1166     return nullptr;
1167   return AddRec;
1168 }
1169 
1170 /// Is this instruction potentially interesting for further simplification after
1171 /// widening it's type? In other words, can the extend be safely hoisted out of
1172 /// the loop with SCEV reducing the value to a recurrence on the same loop. If
1173 /// so, return the sign or zero extended recurrence. Otherwise return NULL.
getWideRecurrence(Instruction * NarrowUse)1174 const SCEVAddRecExpr *WidenIV::getWideRecurrence(Instruction *NarrowUse) {
1175   if (!SE->isSCEVable(NarrowUse->getType()))
1176     return nullptr;
1177 
1178   const SCEV *NarrowExpr = SE->getSCEV(NarrowUse);
1179   if (SE->getTypeSizeInBits(NarrowExpr->getType())
1180       >= SE->getTypeSizeInBits(WideType)) {
1181     // NarrowUse implicitly widens its operand. e.g. a gep with a narrow
1182     // index. So don't follow this use.
1183     return nullptr;
1184   }
1185 
1186   const SCEV *WideExpr = IsSigned ?
1187     SE->getSignExtendExpr(NarrowExpr, WideType) :
1188     SE->getZeroExtendExpr(NarrowExpr, WideType);
1189   const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(WideExpr);
1190   if (!AddRec || AddRec->getLoop() != L)
1191     return nullptr;
1192   return AddRec;
1193 }
1194 
1195 /// This IV user cannot be widen. Replace this use of the original narrow IV
1196 /// with a truncation of the new wide IV to isolate and eliminate the narrow IV.
truncateIVUse(NarrowIVDefUse DU,DominatorTree * DT,LoopInfo * LI)1197 static void truncateIVUse(NarrowIVDefUse DU, DominatorTree *DT, LoopInfo *LI) {
1198   DEBUG(dbgs() << "INDVARS: Truncate IV " << *DU.WideDef
1199         << " for user " << *DU.NarrowUse << "\n");
1200   IRBuilder<> Builder(
1201       getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT, LI));
1202   Value *Trunc = Builder.CreateTrunc(DU.WideDef, DU.NarrowDef->getType());
1203   DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, Trunc);
1204 }
1205 
1206 /// If the narrow use is a compare instruction, then widen the compare
1207 //  (and possibly the other operand).  The extend operation is hoisted into the
1208 // loop preheader as far as possible.
widenLoopCompare(NarrowIVDefUse DU)1209 bool WidenIV::widenLoopCompare(NarrowIVDefUse DU) {
1210   ICmpInst *Cmp = dyn_cast<ICmpInst>(DU.NarrowUse);
1211   if (!Cmp)
1212     return false;
1213 
1214   // We can legally widen the comparison in the following two cases:
1215   //
1216   //  - The signedness of the IV extension and comparison match
1217   //
1218   //  - The narrow IV is always positive (and thus its sign extension is equal
1219   //    to its zero extension).  For instance, let's say we're zero extending
1220   //    %narrow for the following use
1221   //
1222   //      icmp slt i32 %narrow, %val   ... (A)
1223   //
1224   //    and %narrow is always positive.  Then
1225   //
1226   //      (A) == icmp slt i32 sext(%narrow), sext(%val)
1227   //          == icmp slt i32 zext(%narrow), sext(%val)
1228 
1229   if (!(DU.NeverNegative || IsSigned == Cmp->isSigned()))
1230     return false;
1231 
1232   Value *Op = Cmp->getOperand(Cmp->getOperand(0) == DU.NarrowDef ? 1 : 0);
1233   unsigned CastWidth = SE->getTypeSizeInBits(Op->getType());
1234   unsigned IVWidth = SE->getTypeSizeInBits(WideType);
1235   assert (CastWidth <= IVWidth && "Unexpected width while widening compare.");
1236 
1237   // Widen the compare instruction.
1238   IRBuilder<> Builder(
1239       getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT, LI));
1240   DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef);
1241 
1242   // Widen the other operand of the compare, if necessary.
1243   if (CastWidth < IVWidth) {
1244     Value *ExtOp = createExtendInst(Op, WideType, Cmp->isSigned(), Cmp);
1245     DU.NarrowUse->replaceUsesOfWith(Op, ExtOp);
1246   }
1247   return true;
1248 }
1249 
1250 /// Determine whether an individual user of the narrow IV can be widened. If so,
1251 /// return the wide clone of the user.
widenIVUse(NarrowIVDefUse DU,SCEVExpander & Rewriter)1252 Instruction *WidenIV::widenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter) {
1253 
1254   // Stop traversing the def-use chain at inner-loop phis or post-loop phis.
1255   if (PHINode *UsePhi = dyn_cast<PHINode>(DU.NarrowUse)) {
1256     if (LI->getLoopFor(UsePhi->getParent()) != L) {
1257       // For LCSSA phis, sink the truncate outside the loop.
1258       // After SimplifyCFG most loop exit targets have a single predecessor.
1259       // Otherwise fall back to a truncate within the loop.
1260       if (UsePhi->getNumOperands() != 1)
1261         truncateIVUse(DU, DT, LI);
1262       else {
1263         // Widening the PHI requires us to insert a trunc.  The logical place
1264         // for this trunc is in the same BB as the PHI.  This is not possible if
1265         // the BB is terminated by a catchswitch.
1266         if (isa<CatchSwitchInst>(UsePhi->getParent()->getTerminator()))
1267           return nullptr;
1268 
1269         PHINode *WidePhi =
1270           PHINode::Create(DU.WideDef->getType(), 1, UsePhi->getName() + ".wide",
1271                           UsePhi);
1272         WidePhi->addIncoming(DU.WideDef, UsePhi->getIncomingBlock(0));
1273         IRBuilder<> Builder(&*WidePhi->getParent()->getFirstInsertionPt());
1274         Value *Trunc = Builder.CreateTrunc(WidePhi, DU.NarrowDef->getType());
1275         UsePhi->replaceAllUsesWith(Trunc);
1276         DeadInsts.emplace_back(UsePhi);
1277         DEBUG(dbgs() << "INDVARS: Widen lcssa phi " << *UsePhi
1278               << " to " << *WidePhi << "\n");
1279       }
1280       return nullptr;
1281     }
1282   }
1283   // Our raison d'etre! Eliminate sign and zero extension.
1284   if (IsSigned ? isa<SExtInst>(DU.NarrowUse) : isa<ZExtInst>(DU.NarrowUse)) {
1285     Value *NewDef = DU.WideDef;
1286     if (DU.NarrowUse->getType() != WideType) {
1287       unsigned CastWidth = SE->getTypeSizeInBits(DU.NarrowUse->getType());
1288       unsigned IVWidth = SE->getTypeSizeInBits(WideType);
1289       if (CastWidth < IVWidth) {
1290         // The cast isn't as wide as the IV, so insert a Trunc.
1291         IRBuilder<> Builder(DU.NarrowUse);
1292         NewDef = Builder.CreateTrunc(DU.WideDef, DU.NarrowUse->getType());
1293       }
1294       else {
1295         // A wider extend was hidden behind a narrower one. This may induce
1296         // another round of IV widening in which the intermediate IV becomes
1297         // dead. It should be very rare.
1298         DEBUG(dbgs() << "INDVARS: New IV " << *WidePhi
1299               << " not wide enough to subsume " << *DU.NarrowUse << "\n");
1300         DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef);
1301         NewDef = DU.NarrowUse;
1302       }
1303     }
1304     if (NewDef != DU.NarrowUse) {
1305       DEBUG(dbgs() << "INDVARS: eliminating " << *DU.NarrowUse
1306             << " replaced by " << *DU.WideDef << "\n");
1307       ++NumElimExt;
1308       DU.NarrowUse->replaceAllUsesWith(NewDef);
1309       DeadInsts.emplace_back(DU.NarrowUse);
1310     }
1311     // Now that the extend is gone, we want to expose it's uses for potential
1312     // further simplification. We don't need to directly inform SimplifyIVUsers
1313     // of the new users, because their parent IV will be processed later as a
1314     // new loop phi. If we preserved IVUsers analysis, we would also want to
1315     // push the uses of WideDef here.
1316 
1317     // No further widening is needed. The deceased [sz]ext had done it for us.
1318     return nullptr;
1319   }
1320 
1321   // Does this user itself evaluate to a recurrence after widening?
1322   const SCEVAddRecExpr *WideAddRec = getWideRecurrence(DU.NarrowUse);
1323   if (!WideAddRec)
1324     WideAddRec = getExtendedOperandRecurrence(DU);
1325 
1326   if (!WideAddRec) {
1327     // If use is a loop condition, try to promote the condition instead of
1328     // truncating the IV first.
1329     if (widenLoopCompare(DU))
1330       return nullptr;
1331 
1332     // This user does not evaluate to a recurence after widening, so don't
1333     // follow it. Instead insert a Trunc to kill off the original use,
1334     // eventually isolating the original narrow IV so it can be removed.
1335     truncateIVUse(DU, DT, LI);
1336     return nullptr;
1337   }
1338   // Assume block terminators cannot evaluate to a recurrence. We can't to
1339   // insert a Trunc after a terminator if there happens to be a critical edge.
1340   assert(DU.NarrowUse != DU.NarrowUse->getParent()->getTerminator() &&
1341          "SCEV is not expected to evaluate a block terminator");
1342 
1343   // Reuse the IV increment that SCEVExpander created as long as it dominates
1344   // NarrowUse.
1345   Instruction *WideUse = nullptr;
1346   if (WideAddRec == WideIncExpr && Rewriter.hoistIVInc(WideInc, DU.NarrowUse))
1347     WideUse = WideInc;
1348   else {
1349     WideUse = cloneIVUser(DU, WideAddRec);
1350     if (!WideUse)
1351       return nullptr;
1352   }
1353   // Evaluation of WideAddRec ensured that the narrow expression could be
1354   // extended outside the loop without overflow. This suggests that the wide use
1355   // evaluates to the same expression as the extended narrow use, but doesn't
1356   // absolutely guarantee it. Hence the following failsafe check. In rare cases
1357   // where it fails, we simply throw away the newly created wide use.
1358   if (WideAddRec != SE->getSCEV(WideUse)) {
1359     DEBUG(dbgs() << "Wide use expression mismatch: " << *WideUse
1360           << ": " << *SE->getSCEV(WideUse) << " != " << *WideAddRec << "\n");
1361     DeadInsts.emplace_back(WideUse);
1362     return nullptr;
1363   }
1364 
1365   // Returning WideUse pushes it on the worklist.
1366   return WideUse;
1367 }
1368 
1369 /// Add eligible users of NarrowDef to NarrowIVUsers.
1370 ///
pushNarrowIVUsers(Instruction * NarrowDef,Instruction * WideDef)1371 void WidenIV::pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef) {
1372   const SCEV *NarrowSCEV = SE->getSCEV(NarrowDef);
1373   bool NeverNegative =
1374       SE->isKnownPredicate(ICmpInst::ICMP_SGE, NarrowSCEV,
1375                            SE->getConstant(NarrowSCEV->getType(), 0));
1376   for (User *U : NarrowDef->users()) {
1377     Instruction *NarrowUser = cast<Instruction>(U);
1378 
1379     // Handle data flow merges and bizarre phi cycles.
1380     if (!Widened.insert(NarrowUser).second)
1381       continue;
1382 
1383     NarrowIVUsers.emplace_back(NarrowDef, NarrowUser, WideDef, NeverNegative);
1384   }
1385 }
1386 
1387 /// Process a single induction variable. First use the SCEVExpander to create a
1388 /// wide induction variable that evaluates to the same recurrence as the
1389 /// original narrow IV. Then use a worklist to forward traverse the narrow IV's
1390 /// def-use chain. After widenIVUse has processed all interesting IV users, the
1391 /// narrow IV will be isolated for removal by DeleteDeadPHIs.
1392 ///
1393 /// It would be simpler to delete uses as they are processed, but we must avoid
1394 /// invalidating SCEV expressions.
1395 ///
createWideIV(SCEVExpander & Rewriter)1396 PHINode *WidenIV::createWideIV(SCEVExpander &Rewriter) {
1397   // Is this phi an induction variable?
1398   const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(OrigPhi));
1399   if (!AddRec)
1400     return nullptr;
1401 
1402   // Widen the induction variable expression.
1403   const SCEV *WideIVExpr = IsSigned ?
1404     SE->getSignExtendExpr(AddRec, WideType) :
1405     SE->getZeroExtendExpr(AddRec, WideType);
1406 
1407   assert(SE->getEffectiveSCEVType(WideIVExpr->getType()) == WideType &&
1408          "Expect the new IV expression to preserve its type");
1409 
1410   // Can the IV be extended outside the loop without overflow?
1411   AddRec = dyn_cast<SCEVAddRecExpr>(WideIVExpr);
1412   if (!AddRec || AddRec->getLoop() != L)
1413     return nullptr;
1414 
1415   // An AddRec must have loop-invariant operands. Since this AddRec is
1416   // materialized by a loop header phi, the expression cannot have any post-loop
1417   // operands, so they must dominate the loop header.
1418   assert(
1419       SE->properlyDominates(AddRec->getStart(), L->getHeader()) &&
1420       SE->properlyDominates(AddRec->getStepRecurrence(*SE), L->getHeader()) &&
1421       "Loop header phi recurrence inputs do not dominate the loop");
1422 
1423   // The rewriter provides a value for the desired IV expression. This may
1424   // either find an existing phi or materialize a new one. Either way, we
1425   // expect a well-formed cyclic phi-with-increments. i.e. any operand not part
1426   // of the phi-SCC dominates the loop entry.
1427   Instruction *InsertPt = &L->getHeader()->front();
1428   WidePhi = cast<PHINode>(Rewriter.expandCodeFor(AddRec, WideType, InsertPt));
1429 
1430   // Remembering the WideIV increment generated by SCEVExpander allows
1431   // widenIVUse to reuse it when widening the narrow IV's increment. We don't
1432   // employ a general reuse mechanism because the call above is the only call to
1433   // SCEVExpander. Henceforth, we produce 1-to-1 narrow to wide uses.
1434   if (BasicBlock *LatchBlock = L->getLoopLatch()) {
1435     WideInc =
1436       cast<Instruction>(WidePhi->getIncomingValueForBlock(LatchBlock));
1437     WideIncExpr = SE->getSCEV(WideInc);
1438   }
1439 
1440   DEBUG(dbgs() << "Wide IV: " << *WidePhi << "\n");
1441   ++NumWidened;
1442 
1443   // Traverse the def-use chain using a worklist starting at the original IV.
1444   assert(Widened.empty() && NarrowIVUsers.empty() && "expect initial state" );
1445 
1446   Widened.insert(OrigPhi);
1447   pushNarrowIVUsers(OrigPhi, WidePhi);
1448 
1449   while (!NarrowIVUsers.empty()) {
1450     NarrowIVDefUse DU = NarrowIVUsers.pop_back_val();
1451 
1452     // Process a def-use edge. This may replace the use, so don't hold a
1453     // use_iterator across it.
1454     Instruction *WideUse = widenIVUse(DU, Rewriter);
1455 
1456     // Follow all def-use edges from the previous narrow use.
1457     if (WideUse)
1458       pushNarrowIVUsers(DU.NarrowUse, WideUse);
1459 
1460     // widenIVUse may have removed the def-use edge.
1461     if (DU.NarrowDef->use_empty())
1462       DeadInsts.emplace_back(DU.NarrowDef);
1463   }
1464   return WidePhi;
1465 }
1466 
1467 //===----------------------------------------------------------------------===//
1468 //  Live IV Reduction - Minimize IVs live across the loop.
1469 //===----------------------------------------------------------------------===//
1470 
1471 
1472 //===----------------------------------------------------------------------===//
1473 //  Simplification of IV users based on SCEV evaluation.
1474 //===----------------------------------------------------------------------===//
1475 
1476 namespace {
1477 class IndVarSimplifyVisitor : public IVVisitor {
1478   ScalarEvolution *SE;
1479   const TargetTransformInfo *TTI;
1480   PHINode *IVPhi;
1481 
1482 public:
1483   WideIVInfo WI;
1484 
IndVarSimplifyVisitor(PHINode * IV,ScalarEvolution * SCEV,const TargetTransformInfo * TTI,const DominatorTree * DTree)1485   IndVarSimplifyVisitor(PHINode *IV, ScalarEvolution *SCEV,
1486                         const TargetTransformInfo *TTI,
1487                         const DominatorTree *DTree)
1488     : SE(SCEV), TTI(TTI), IVPhi(IV) {
1489     DT = DTree;
1490     WI.NarrowIV = IVPhi;
1491   }
1492 
1493   // Implement the interface used by simplifyUsersOfIV.
visitCast(CastInst * Cast)1494   void visitCast(CastInst *Cast) override { visitIVCast(Cast, WI, SE, TTI); }
1495 };
1496 }
1497 
1498 /// Iteratively perform simplification on a worklist of IV users. Each
1499 /// successive simplification may push more users which may themselves be
1500 /// candidates for simplification.
1501 ///
1502 /// Sign/Zero extend elimination is interleaved with IV simplification.
1503 ///
simplifyAndExtend(Loop * L,SCEVExpander & Rewriter,LoopInfo * LI)1504 void IndVarSimplify::simplifyAndExtend(Loop *L,
1505                                        SCEVExpander &Rewriter,
1506                                        LoopInfo *LI) {
1507   SmallVector<WideIVInfo, 8> WideIVs;
1508 
1509   SmallVector<PHINode*, 8> LoopPhis;
1510   for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
1511     LoopPhis.push_back(cast<PHINode>(I));
1512   }
1513   // Each round of simplification iterates through the SimplifyIVUsers worklist
1514   // for all current phis, then determines whether any IVs can be
1515   // widened. Widening adds new phis to LoopPhis, inducing another round of
1516   // simplification on the wide IVs.
1517   while (!LoopPhis.empty()) {
1518     // Evaluate as many IV expressions as possible before widening any IVs. This
1519     // forces SCEV to set no-wrap flags before evaluating sign/zero
1520     // extension. The first time SCEV attempts to normalize sign/zero extension,
1521     // the result becomes final. So for the most predictable results, we delay
1522     // evaluation of sign/zero extend evaluation until needed, and avoid running
1523     // other SCEV based analysis prior to simplifyAndExtend.
1524     do {
1525       PHINode *CurrIV = LoopPhis.pop_back_val();
1526 
1527       // Information about sign/zero extensions of CurrIV.
1528       IndVarSimplifyVisitor Visitor(CurrIV, SE, TTI, DT);
1529 
1530       Changed |= simplifyUsersOfIV(CurrIV, SE, DT, LI, DeadInsts, &Visitor);
1531 
1532       if (Visitor.WI.WidestNativeType) {
1533         WideIVs.push_back(Visitor.WI);
1534       }
1535     } while(!LoopPhis.empty());
1536 
1537     for (; !WideIVs.empty(); WideIVs.pop_back()) {
1538       WidenIV Widener(WideIVs.back(), LI, SE, DT, DeadInsts);
1539       if (PHINode *WidePhi = Widener.createWideIV(Rewriter)) {
1540         Changed = true;
1541         LoopPhis.push_back(WidePhi);
1542       }
1543     }
1544   }
1545 }
1546 
1547 //===----------------------------------------------------------------------===//
1548 //  linearFunctionTestReplace and its kin. Rewrite the loop exit condition.
1549 //===----------------------------------------------------------------------===//
1550 
1551 /// Return true if this loop's backedge taken count expression can be safely and
1552 /// cheaply expanded into an instruction sequence that can be used by
1553 /// linearFunctionTestReplace.
1554 ///
1555 /// TODO: This fails for pointer-type loop counters with greater than one byte
1556 /// strides, consequently preventing LFTR from running. For the purpose of LFTR
1557 /// we could skip this check in the case that the LFTR loop counter (chosen by
1558 /// FindLoopCounter) is also pointer type. Instead, we could directly convert
1559 /// the loop test to an inequality test by checking the target data's alignment
1560 /// of element types (given that the initial pointer value originates from or is
1561 /// used by ABI constrained operation, as opposed to inttoptr/ptrtoint).
1562 /// However, we don't yet have a strong motivation for converting loop tests
1563 /// into inequality tests.
canExpandBackedgeTakenCount(Loop * L,ScalarEvolution * SE,SCEVExpander & Rewriter)1564 static bool canExpandBackedgeTakenCount(Loop *L, ScalarEvolution *SE,
1565                                         SCEVExpander &Rewriter) {
1566   const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
1567   if (isa<SCEVCouldNotCompute>(BackedgeTakenCount) ||
1568       BackedgeTakenCount->isZero())
1569     return false;
1570 
1571   if (!L->getExitingBlock())
1572     return false;
1573 
1574   // Can't rewrite non-branch yet.
1575   if (!isa<BranchInst>(L->getExitingBlock()->getTerminator()))
1576     return false;
1577 
1578   if (Rewriter.isHighCostExpansion(BackedgeTakenCount, L))
1579     return false;
1580 
1581   return true;
1582 }
1583 
1584 /// Return the loop header phi IFF IncV adds a loop invariant value to the phi.
getLoopPhiForCounter(Value * IncV,Loop * L,DominatorTree * DT)1585 static PHINode *getLoopPhiForCounter(Value *IncV, Loop *L, DominatorTree *DT) {
1586   Instruction *IncI = dyn_cast<Instruction>(IncV);
1587   if (!IncI)
1588     return nullptr;
1589 
1590   switch (IncI->getOpcode()) {
1591   case Instruction::Add:
1592   case Instruction::Sub:
1593     break;
1594   case Instruction::GetElementPtr:
1595     // An IV counter must preserve its type.
1596     if (IncI->getNumOperands() == 2)
1597       break;
1598   default:
1599     return nullptr;
1600   }
1601 
1602   PHINode *Phi = dyn_cast<PHINode>(IncI->getOperand(0));
1603   if (Phi && Phi->getParent() == L->getHeader()) {
1604     if (isLoopInvariant(IncI->getOperand(1), L, DT))
1605       return Phi;
1606     return nullptr;
1607   }
1608   if (IncI->getOpcode() == Instruction::GetElementPtr)
1609     return nullptr;
1610 
1611   // Allow add/sub to be commuted.
1612   Phi = dyn_cast<PHINode>(IncI->getOperand(1));
1613   if (Phi && Phi->getParent() == L->getHeader()) {
1614     if (isLoopInvariant(IncI->getOperand(0), L, DT))
1615       return Phi;
1616   }
1617   return nullptr;
1618 }
1619 
1620 /// Return the compare guarding the loop latch, or NULL for unrecognized tests.
getLoopTest(Loop * L)1621 static ICmpInst *getLoopTest(Loop *L) {
1622   assert(L->getExitingBlock() && "expected loop exit");
1623 
1624   BasicBlock *LatchBlock = L->getLoopLatch();
1625   // Don't bother with LFTR if the loop is not properly simplified.
1626   if (!LatchBlock)
1627     return nullptr;
1628 
1629   BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
1630   assert(BI && "expected exit branch");
1631 
1632   return dyn_cast<ICmpInst>(BI->getCondition());
1633 }
1634 
1635 /// linearFunctionTestReplace policy. Return true unless we can show that the
1636 /// current exit test is already sufficiently canonical.
needsLFTR(Loop * L,DominatorTree * DT)1637 static bool needsLFTR(Loop *L, DominatorTree *DT) {
1638   // Do LFTR to simplify the exit condition to an ICMP.
1639   ICmpInst *Cond = getLoopTest(L);
1640   if (!Cond)
1641     return true;
1642 
1643   // Do LFTR to simplify the exit ICMP to EQ/NE
1644   ICmpInst::Predicate Pred = Cond->getPredicate();
1645   if (Pred != ICmpInst::ICMP_NE && Pred != ICmpInst::ICMP_EQ)
1646     return true;
1647 
1648   // Look for a loop invariant RHS
1649   Value *LHS = Cond->getOperand(0);
1650   Value *RHS = Cond->getOperand(1);
1651   if (!isLoopInvariant(RHS, L, DT)) {
1652     if (!isLoopInvariant(LHS, L, DT))
1653       return true;
1654     std::swap(LHS, RHS);
1655   }
1656   // Look for a simple IV counter LHS
1657   PHINode *Phi = dyn_cast<PHINode>(LHS);
1658   if (!Phi)
1659     Phi = getLoopPhiForCounter(LHS, L, DT);
1660 
1661   if (!Phi)
1662     return true;
1663 
1664   // Do LFTR if PHI node is defined in the loop, but is *not* a counter.
1665   int Idx = Phi->getBasicBlockIndex(L->getLoopLatch());
1666   if (Idx < 0)
1667     return true;
1668 
1669   // Do LFTR if the exit condition's IV is *not* a simple counter.
1670   Value *IncV = Phi->getIncomingValue(Idx);
1671   return Phi != getLoopPhiForCounter(IncV, L, DT);
1672 }
1673 
1674 /// Recursive helper for hasConcreteDef(). Unfortunately, this currently boils
1675 /// down to checking that all operands are constant and listing instructions
1676 /// that may hide undef.
hasConcreteDefImpl(Value * V,SmallPtrSetImpl<Value * > & Visited,unsigned Depth)1677 static bool hasConcreteDefImpl(Value *V, SmallPtrSetImpl<Value*> &Visited,
1678                                unsigned Depth) {
1679   if (isa<Constant>(V))
1680     return !isa<UndefValue>(V);
1681 
1682   if (Depth >= 6)
1683     return false;
1684 
1685   // Conservatively handle non-constant non-instructions. For example, Arguments
1686   // may be undef.
1687   Instruction *I = dyn_cast<Instruction>(V);
1688   if (!I)
1689     return false;
1690 
1691   // Load and return values may be undef.
1692   if(I->mayReadFromMemory() || isa<CallInst>(I) || isa<InvokeInst>(I))
1693     return false;
1694 
1695   // Optimistically handle other instructions.
1696   for (Value *Op : I->operands()) {
1697     if (!Visited.insert(Op).second)
1698       continue;
1699     if (!hasConcreteDefImpl(Op, Visited, Depth+1))
1700       return false;
1701   }
1702   return true;
1703 }
1704 
1705 /// Return true if the given value is concrete. We must prove that undef can
1706 /// never reach it.
1707 ///
1708 /// TODO: If we decide that this is a good approach to checking for undef, we
1709 /// may factor it into a common location.
hasConcreteDef(Value * V)1710 static bool hasConcreteDef(Value *V) {
1711   SmallPtrSet<Value*, 8> Visited;
1712   Visited.insert(V);
1713   return hasConcreteDefImpl(V, Visited, 0);
1714 }
1715 
1716 /// Return true if this IV has any uses other than the (soon to be rewritten)
1717 /// loop exit test.
AlmostDeadIV(PHINode * Phi,BasicBlock * LatchBlock,Value * Cond)1718 static bool AlmostDeadIV(PHINode *Phi, BasicBlock *LatchBlock, Value *Cond) {
1719   int LatchIdx = Phi->getBasicBlockIndex(LatchBlock);
1720   Value *IncV = Phi->getIncomingValue(LatchIdx);
1721 
1722   for (User *U : Phi->users())
1723     if (U != Cond && U != IncV) return false;
1724 
1725   for (User *U : IncV->users())
1726     if (U != Cond && U != Phi) return false;
1727   return true;
1728 }
1729 
1730 /// Find an affine IV in canonical form.
1731 ///
1732 /// BECount may be an i8* pointer type. The pointer difference is already
1733 /// valid count without scaling the address stride, so it remains a pointer
1734 /// expression as far as SCEV is concerned.
1735 ///
1736 /// Currently only valid for LFTR. See the comments on hasConcreteDef below.
1737 ///
1738 /// FIXME: Accept -1 stride and set IVLimit = IVInit - BECount
1739 ///
1740 /// FIXME: Accept non-unit stride as long as SCEV can reduce BECount * Stride.
1741 /// This is difficult in general for SCEV because of potential overflow. But we
1742 /// could at least handle constant BECounts.
FindLoopCounter(Loop * L,const SCEV * BECount,ScalarEvolution * SE,DominatorTree * DT)1743 static PHINode *FindLoopCounter(Loop *L, const SCEV *BECount,
1744                                 ScalarEvolution *SE, DominatorTree *DT) {
1745   uint64_t BCWidth = SE->getTypeSizeInBits(BECount->getType());
1746 
1747   Value *Cond =
1748     cast<BranchInst>(L->getExitingBlock()->getTerminator())->getCondition();
1749 
1750   // Loop over all of the PHI nodes, looking for a simple counter.
1751   PHINode *BestPhi = nullptr;
1752   const SCEV *BestInit = nullptr;
1753   BasicBlock *LatchBlock = L->getLoopLatch();
1754   assert(LatchBlock && "needsLFTR should guarantee a loop latch");
1755   const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
1756 
1757   for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
1758     PHINode *Phi = cast<PHINode>(I);
1759     if (!SE->isSCEVable(Phi->getType()))
1760       continue;
1761 
1762     // Avoid comparing an integer IV against a pointer Limit.
1763     if (BECount->getType()->isPointerTy() && !Phi->getType()->isPointerTy())
1764       continue;
1765 
1766     const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Phi));
1767     if (!AR || AR->getLoop() != L || !AR->isAffine())
1768       continue;
1769 
1770     // AR may be a pointer type, while BECount is an integer type.
1771     // AR may be wider than BECount. With eq/ne tests overflow is immaterial.
1772     // AR may not be a narrower type, or we may never exit.
1773     uint64_t PhiWidth = SE->getTypeSizeInBits(AR->getType());
1774     if (PhiWidth < BCWidth || !DL.isLegalInteger(PhiWidth))
1775       continue;
1776 
1777     const SCEV *Step = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE));
1778     if (!Step || !Step->isOne())
1779       continue;
1780 
1781     int LatchIdx = Phi->getBasicBlockIndex(LatchBlock);
1782     Value *IncV = Phi->getIncomingValue(LatchIdx);
1783     if (getLoopPhiForCounter(IncV, L, DT) != Phi)
1784       continue;
1785 
1786     // Avoid reusing a potentially undef value to compute other values that may
1787     // have originally had a concrete definition.
1788     if (!hasConcreteDef(Phi)) {
1789       // We explicitly allow unknown phis as long as they are already used by
1790       // the loop test. In this case we assume that performing LFTR could not
1791       // increase the number of undef users.
1792       if (ICmpInst *Cond = getLoopTest(L)) {
1793         if (Phi != getLoopPhiForCounter(Cond->getOperand(0), L, DT) &&
1794             Phi != getLoopPhiForCounter(Cond->getOperand(1), L, DT)) {
1795           continue;
1796         }
1797       }
1798     }
1799     const SCEV *Init = AR->getStart();
1800 
1801     if (BestPhi && !AlmostDeadIV(BestPhi, LatchBlock, Cond)) {
1802       // Don't force a live loop counter if another IV can be used.
1803       if (AlmostDeadIV(Phi, LatchBlock, Cond))
1804         continue;
1805 
1806       // Prefer to count-from-zero. This is a more "canonical" counter form. It
1807       // also prefers integer to pointer IVs.
1808       if (BestInit->isZero() != Init->isZero()) {
1809         if (BestInit->isZero())
1810           continue;
1811       }
1812       // If two IVs both count from zero or both count from nonzero then the
1813       // narrower is likely a dead phi that has been widened. Use the wider phi
1814       // to allow the other to be eliminated.
1815       else if (PhiWidth <= SE->getTypeSizeInBits(BestPhi->getType()))
1816         continue;
1817     }
1818     BestPhi = Phi;
1819     BestInit = Init;
1820   }
1821   return BestPhi;
1822 }
1823 
1824 /// Help linearFunctionTestReplace by generating a value that holds the RHS of
1825 /// the new loop test.
genLoopLimit(PHINode * IndVar,const SCEV * IVCount,Loop * L,SCEVExpander & Rewriter,ScalarEvolution * SE)1826 static Value *genLoopLimit(PHINode *IndVar, const SCEV *IVCount, Loop *L,
1827                            SCEVExpander &Rewriter, ScalarEvolution *SE) {
1828   const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(IndVar));
1829   assert(AR && AR->getLoop() == L && AR->isAffine() && "bad loop counter");
1830   const SCEV *IVInit = AR->getStart();
1831 
1832   // IVInit may be a pointer while IVCount is an integer when FindLoopCounter
1833   // finds a valid pointer IV. Sign extend BECount in order to materialize a
1834   // GEP. Avoid running SCEVExpander on a new pointer value, instead reusing
1835   // the existing GEPs whenever possible.
1836   if (IndVar->getType()->isPointerTy() && !IVCount->getType()->isPointerTy()) {
1837     // IVOffset will be the new GEP offset that is interpreted by GEP as a
1838     // signed value. IVCount on the other hand represents the loop trip count,
1839     // which is an unsigned value. FindLoopCounter only allows induction
1840     // variables that have a positive unit stride of one. This means we don't
1841     // have to handle the case of negative offsets (yet) and just need to zero
1842     // extend IVCount.
1843     Type *OfsTy = SE->getEffectiveSCEVType(IVInit->getType());
1844     const SCEV *IVOffset = SE->getTruncateOrZeroExtend(IVCount, OfsTy);
1845 
1846     // Expand the code for the iteration count.
1847     assert(SE->isLoopInvariant(IVOffset, L) &&
1848            "Computed iteration count is not loop invariant!");
1849     BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1850     Value *GEPOffset = Rewriter.expandCodeFor(IVOffset, OfsTy, BI);
1851 
1852     Value *GEPBase = IndVar->getIncomingValueForBlock(L->getLoopPreheader());
1853     assert(AR->getStart() == SE->getSCEV(GEPBase) && "bad loop counter");
1854     // We could handle pointer IVs other than i8*, but we need to compensate for
1855     // gep index scaling. See canExpandBackedgeTakenCount comments.
1856     assert(SE->getSizeOfExpr(IntegerType::getInt64Ty(IndVar->getContext()),
1857                              cast<PointerType>(GEPBase->getType())
1858                                  ->getElementType())->isOne() &&
1859            "unit stride pointer IV must be i8*");
1860 
1861     IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
1862     return Builder.CreateGEP(nullptr, GEPBase, GEPOffset, "lftr.limit");
1863   } else {
1864     // In any other case, convert both IVInit and IVCount to integers before
1865     // comparing. This may result in SCEV expension of pointers, but in practice
1866     // SCEV will fold the pointer arithmetic away as such:
1867     // BECount = (IVEnd - IVInit - 1) => IVLimit = IVInit (postinc).
1868     //
1869     // Valid Cases: (1) both integers is most common; (2) both may be pointers
1870     // for simple memset-style loops.
1871     //
1872     // IVInit integer and IVCount pointer would only occur if a canonical IV
1873     // were generated on top of case #2, which is not expected.
1874 
1875     const SCEV *IVLimit = nullptr;
1876     // For unit stride, IVCount = Start + BECount with 2's complement overflow.
1877     // For non-zero Start, compute IVCount here.
1878     if (AR->getStart()->isZero())
1879       IVLimit = IVCount;
1880     else {
1881       assert(AR->getStepRecurrence(*SE)->isOne() && "only handles unit stride");
1882       const SCEV *IVInit = AR->getStart();
1883 
1884       // For integer IVs, truncate the IV before computing IVInit + BECount.
1885       if (SE->getTypeSizeInBits(IVInit->getType())
1886           > SE->getTypeSizeInBits(IVCount->getType()))
1887         IVInit = SE->getTruncateExpr(IVInit, IVCount->getType());
1888 
1889       IVLimit = SE->getAddExpr(IVInit, IVCount);
1890     }
1891     // Expand the code for the iteration count.
1892     BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1893     IRBuilder<> Builder(BI);
1894     assert(SE->isLoopInvariant(IVLimit, L) &&
1895            "Computed iteration count is not loop invariant!");
1896     // Ensure that we generate the same type as IndVar, or a smaller integer
1897     // type. In the presence of null pointer values, we have an integer type
1898     // SCEV expression (IVInit) for a pointer type IV value (IndVar).
1899     Type *LimitTy = IVCount->getType()->isPointerTy() ?
1900       IndVar->getType() : IVCount->getType();
1901     return Rewriter.expandCodeFor(IVLimit, LimitTy, BI);
1902   }
1903 }
1904 
1905 /// This method rewrites the exit condition of the loop to be a canonical !=
1906 /// comparison against the incremented loop induction variable.  This pass is
1907 /// able to rewrite the exit tests of any loop where the SCEV analysis can
1908 /// determine a loop-invariant trip count of the loop, which is actually a much
1909 /// broader range than just linear tests.
1910 Value *IndVarSimplify::
linearFunctionTestReplace(Loop * L,const SCEV * BackedgeTakenCount,PHINode * IndVar,SCEVExpander & Rewriter)1911 linearFunctionTestReplace(Loop *L,
1912                           const SCEV *BackedgeTakenCount,
1913                           PHINode *IndVar,
1914                           SCEVExpander &Rewriter) {
1915   assert(canExpandBackedgeTakenCount(L, SE, Rewriter) && "precondition");
1916 
1917   // Initialize CmpIndVar and IVCount to their preincremented values.
1918   Value *CmpIndVar = IndVar;
1919   const SCEV *IVCount = BackedgeTakenCount;
1920 
1921   // If the exiting block is the same as the backedge block, we prefer to
1922   // compare against the post-incremented value, otherwise we must compare
1923   // against the preincremented value.
1924   if (L->getExitingBlock() == L->getLoopLatch()) {
1925     // Add one to the "backedge-taken" count to get the trip count.
1926     // This addition may overflow, which is valid as long as the comparison is
1927     // truncated to BackedgeTakenCount->getType().
1928     IVCount = SE->getAddExpr(BackedgeTakenCount,
1929                              SE->getOne(BackedgeTakenCount->getType()));
1930     // The BackedgeTaken expression contains the number of times that the
1931     // backedge branches to the loop header.  This is one less than the
1932     // number of times the loop executes, so use the incremented indvar.
1933     CmpIndVar = IndVar->getIncomingValueForBlock(L->getExitingBlock());
1934   }
1935 
1936   Value *ExitCnt = genLoopLimit(IndVar, IVCount, L, Rewriter, SE);
1937   assert(ExitCnt->getType()->isPointerTy() ==
1938              IndVar->getType()->isPointerTy() &&
1939          "genLoopLimit missed a cast");
1940 
1941   // Insert a new icmp_ne or icmp_eq instruction before the branch.
1942   BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1943   ICmpInst::Predicate P;
1944   if (L->contains(BI->getSuccessor(0)))
1945     P = ICmpInst::ICMP_NE;
1946   else
1947     P = ICmpInst::ICMP_EQ;
1948 
1949   DEBUG(dbgs() << "INDVARS: Rewriting loop exit condition to:\n"
1950                << "      LHS:" << *CmpIndVar << '\n'
1951                << "       op:\t"
1952                << (P == ICmpInst::ICMP_NE ? "!=" : "==") << "\n"
1953                << "      RHS:\t" << *ExitCnt << "\n"
1954                << "  IVCount:\t" << *IVCount << "\n");
1955 
1956   IRBuilder<> Builder(BI);
1957 
1958   // LFTR can ignore IV overflow and truncate to the width of
1959   // BECount. This avoids materializing the add(zext(add)) expression.
1960   unsigned CmpIndVarSize = SE->getTypeSizeInBits(CmpIndVar->getType());
1961   unsigned ExitCntSize = SE->getTypeSizeInBits(ExitCnt->getType());
1962   if (CmpIndVarSize > ExitCntSize) {
1963     const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(SE->getSCEV(IndVar));
1964     const SCEV *ARStart = AR->getStart();
1965     const SCEV *ARStep = AR->getStepRecurrence(*SE);
1966     // For constant IVCount, avoid truncation.
1967     if (isa<SCEVConstant>(ARStart) && isa<SCEVConstant>(IVCount)) {
1968       const APInt &Start = cast<SCEVConstant>(ARStart)->getAPInt();
1969       APInt Count = cast<SCEVConstant>(IVCount)->getAPInt();
1970       // Note that the post-inc value of BackedgeTakenCount may have overflowed
1971       // above such that IVCount is now zero.
1972       if (IVCount != BackedgeTakenCount && Count == 0) {
1973         Count = APInt::getMaxValue(Count.getBitWidth()).zext(CmpIndVarSize);
1974         ++Count;
1975       }
1976       else
1977         Count = Count.zext(CmpIndVarSize);
1978       APInt NewLimit;
1979       if (cast<SCEVConstant>(ARStep)->getValue()->isNegative())
1980         NewLimit = Start - Count;
1981       else
1982         NewLimit = Start + Count;
1983       ExitCnt = ConstantInt::get(CmpIndVar->getType(), NewLimit);
1984 
1985       DEBUG(dbgs() << "  Widen RHS:\t" << *ExitCnt << "\n");
1986     } else {
1987       CmpIndVar = Builder.CreateTrunc(CmpIndVar, ExitCnt->getType(),
1988                                       "lftr.wideiv");
1989     }
1990   }
1991   Value *Cond = Builder.CreateICmp(P, CmpIndVar, ExitCnt, "exitcond");
1992   Value *OrigCond = BI->getCondition();
1993   // It's tempting to use replaceAllUsesWith here to fully replace the old
1994   // comparison, but that's not immediately safe, since users of the old
1995   // comparison may not be dominated by the new comparison. Instead, just
1996   // update the branch to use the new comparison; in the common case this
1997   // will make old comparison dead.
1998   BI->setCondition(Cond);
1999   DeadInsts.push_back(OrigCond);
2000 
2001   ++NumLFTR;
2002   Changed = true;
2003   return Cond;
2004 }
2005 
2006 //===----------------------------------------------------------------------===//
2007 //  sinkUnusedInvariants. A late subpass to cleanup loop preheaders.
2008 //===----------------------------------------------------------------------===//
2009 
2010 /// If there's a single exit block, sink any loop-invariant values that
2011 /// were defined in the preheader but not used inside the loop into the
2012 /// exit block to reduce register pressure in the loop.
sinkUnusedInvariants(Loop * L)2013 void IndVarSimplify::sinkUnusedInvariants(Loop *L) {
2014   BasicBlock *ExitBlock = L->getExitBlock();
2015   if (!ExitBlock) return;
2016 
2017   BasicBlock *Preheader = L->getLoopPreheader();
2018   if (!Preheader) return;
2019 
2020   Instruction *InsertPt = &*ExitBlock->getFirstInsertionPt();
2021   BasicBlock::iterator I(Preheader->getTerminator());
2022   while (I != Preheader->begin()) {
2023     --I;
2024     // New instructions were inserted at the end of the preheader.
2025     if (isa<PHINode>(I))
2026       break;
2027 
2028     // Don't move instructions which might have side effects, since the side
2029     // effects need to complete before instructions inside the loop.  Also don't
2030     // move instructions which might read memory, since the loop may modify
2031     // memory. Note that it's okay if the instruction might have undefined
2032     // behavior: LoopSimplify guarantees that the preheader dominates the exit
2033     // block.
2034     if (I->mayHaveSideEffects() || I->mayReadFromMemory())
2035       continue;
2036 
2037     // Skip debug info intrinsics.
2038     if (isa<DbgInfoIntrinsic>(I))
2039       continue;
2040 
2041     // Skip eh pad instructions.
2042     if (I->isEHPad())
2043       continue;
2044 
2045     // Don't sink alloca: we never want to sink static alloca's out of the
2046     // entry block, and correctly sinking dynamic alloca's requires
2047     // checks for stacksave/stackrestore intrinsics.
2048     // FIXME: Refactor this check somehow?
2049     if (isa<AllocaInst>(I))
2050       continue;
2051 
2052     // Determine if there is a use in or before the loop (direct or
2053     // otherwise).
2054     bool UsedInLoop = false;
2055     for (Use &U : I->uses()) {
2056       Instruction *User = cast<Instruction>(U.getUser());
2057       BasicBlock *UseBB = User->getParent();
2058       if (PHINode *P = dyn_cast<PHINode>(User)) {
2059         unsigned i =
2060           PHINode::getIncomingValueNumForOperand(U.getOperandNo());
2061         UseBB = P->getIncomingBlock(i);
2062       }
2063       if (UseBB == Preheader || L->contains(UseBB)) {
2064         UsedInLoop = true;
2065         break;
2066       }
2067     }
2068 
2069     // If there is, the def must remain in the preheader.
2070     if (UsedInLoop)
2071       continue;
2072 
2073     // Otherwise, sink it to the exit block.
2074     Instruction *ToMove = &*I;
2075     bool Done = false;
2076 
2077     if (I != Preheader->begin()) {
2078       // Skip debug info intrinsics.
2079       do {
2080         --I;
2081       } while (isa<DbgInfoIntrinsic>(I) && I != Preheader->begin());
2082 
2083       if (isa<DbgInfoIntrinsic>(I) && I == Preheader->begin())
2084         Done = true;
2085     } else {
2086       Done = true;
2087     }
2088 
2089     ToMove->moveBefore(InsertPt);
2090     if (Done) break;
2091     InsertPt = ToMove;
2092   }
2093 }
2094 
2095 //===----------------------------------------------------------------------===//
2096 //  IndVarSimplify driver. Manage several subpasses of IV simplification.
2097 //===----------------------------------------------------------------------===//
2098 
run(Loop * L)2099 bool IndVarSimplify::run(Loop *L) {
2100   // We need (and expect!) the incoming loop to be in LCSSA.
2101   assert(L->isRecursivelyLCSSAForm(*DT) && "LCSSA required to run indvars!");
2102 
2103   // If LoopSimplify form is not available, stay out of trouble. Some notes:
2104   //  - LSR currently only supports LoopSimplify-form loops. Indvars'
2105   //    canonicalization can be a pessimization without LSR to "clean up"
2106   //    afterwards.
2107   //  - We depend on having a preheader; in particular,
2108   //    Loop::getCanonicalInductionVariable only supports loops with preheaders,
2109   //    and we're in trouble if we can't find the induction variable even when
2110   //    we've manually inserted one.
2111   if (!L->isLoopSimplifyForm())
2112     return false;
2113 
2114   // If there are any floating-point recurrences, attempt to
2115   // transform them to use integer recurrences.
2116   rewriteNonIntegerIVs(L);
2117 
2118   const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
2119 
2120   // Create a rewriter object which we'll use to transform the code with.
2121   SCEVExpander Rewriter(*SE, DL, "indvars");
2122 #ifndef NDEBUG
2123   Rewriter.setDebugType(DEBUG_TYPE);
2124 #endif
2125 
2126   // Eliminate redundant IV users.
2127   //
2128   // Simplification works best when run before other consumers of SCEV. We
2129   // attempt to avoid evaluating SCEVs for sign/zero extend operations until
2130   // other expressions involving loop IVs have been evaluated. This helps SCEV
2131   // set no-wrap flags before normalizing sign/zero extension.
2132   Rewriter.disableCanonicalMode();
2133   simplifyAndExtend(L, Rewriter, LI);
2134 
2135   // Check to see if this loop has a computable loop-invariant execution count.
2136   // If so, this means that we can compute the final value of any expressions
2137   // that are recurrent in the loop, and substitute the exit values from the
2138   // loop into any instructions outside of the loop that use the final values of
2139   // the current expressions.
2140   //
2141   if (ReplaceExitValue != NeverRepl &&
2142       !isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2143     rewriteLoopExitValues(L, Rewriter);
2144 
2145   // Eliminate redundant IV cycles.
2146   NumElimIV += Rewriter.replaceCongruentIVs(L, DT, DeadInsts);
2147 
2148   // If we have a trip count expression, rewrite the loop's exit condition
2149   // using it.  We can currently only handle loops with a single exit.
2150   if (canExpandBackedgeTakenCount(L, SE, Rewriter) && needsLFTR(L, DT)) {
2151     PHINode *IndVar = FindLoopCounter(L, BackedgeTakenCount, SE, DT);
2152     if (IndVar) {
2153       // Check preconditions for proper SCEVExpander operation. SCEV does not
2154       // express SCEVExpander's dependencies, such as LoopSimplify. Instead any
2155       // pass that uses the SCEVExpander must do it. This does not work well for
2156       // loop passes because SCEVExpander makes assumptions about all loops,
2157       // while LoopPassManager only forces the current loop to be simplified.
2158       //
2159       // FIXME: SCEV expansion has no way to bail out, so the caller must
2160       // explicitly check any assumptions made by SCEV. Brittle.
2161       const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(BackedgeTakenCount);
2162       if (!AR || AR->getLoop()->getLoopPreheader())
2163         (void)linearFunctionTestReplace(L, BackedgeTakenCount, IndVar,
2164                                         Rewriter);
2165     }
2166   }
2167   // Clear the rewriter cache, because values that are in the rewriter's cache
2168   // can be deleted in the loop below, causing the AssertingVH in the cache to
2169   // trigger.
2170   Rewriter.clear();
2171 
2172   // Now that we're done iterating through lists, clean up any instructions
2173   // which are now dead.
2174   while (!DeadInsts.empty())
2175     if (Instruction *Inst =
2176             dyn_cast_or_null<Instruction>(DeadInsts.pop_back_val()))
2177       RecursivelyDeleteTriviallyDeadInstructions(Inst, TLI);
2178 
2179   // The Rewriter may not be used from this point on.
2180 
2181   // Loop-invariant instructions in the preheader that aren't used in the
2182   // loop may be sunk below the loop to reduce register pressure.
2183   sinkUnusedInvariants(L);
2184 
2185   // rewriteFirstIterationLoopExitValues does not rely on the computation of
2186   // trip count and therefore can further simplify exit values in addition to
2187   // rewriteLoopExitValues.
2188   rewriteFirstIterationLoopExitValues(L);
2189 
2190   // Clean up dead instructions.
2191   Changed |= DeleteDeadPHIs(L->getHeader(), TLI);
2192 
2193   // Check a post-condition.
2194   assert(L->isRecursivelyLCSSAForm(*DT) && "Indvars did not preserve LCSSA!");
2195 
2196   // Verify that LFTR, and any other change have not interfered with SCEV's
2197   // ability to compute trip count.
2198 #ifndef NDEBUG
2199   if (VerifyIndvars && !isa<SCEVCouldNotCompute>(BackedgeTakenCount)) {
2200     SE->forgetLoop(L);
2201     const SCEV *NewBECount = SE->getBackedgeTakenCount(L);
2202     if (SE->getTypeSizeInBits(BackedgeTakenCount->getType()) <
2203         SE->getTypeSizeInBits(NewBECount->getType()))
2204       NewBECount = SE->getTruncateOrNoop(NewBECount,
2205                                          BackedgeTakenCount->getType());
2206     else
2207       BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount,
2208                                                  NewBECount->getType());
2209     assert(BackedgeTakenCount == NewBECount && "indvars must preserve SCEV");
2210   }
2211 #endif
2212 
2213   return Changed;
2214 }
2215 
run(Loop & L,AnalysisManager<Loop> & AM)2216 PreservedAnalyses IndVarSimplifyPass::run(Loop &L, AnalysisManager<Loop> &AM) {
2217   auto &FAM = AM.getResult<FunctionAnalysisManagerLoopProxy>(L).getManager();
2218   Function *F = L.getHeader()->getParent();
2219   const DataLayout &DL = F->getParent()->getDataLayout();
2220 
2221   auto *LI = FAM.getCachedResult<LoopAnalysis>(*F);
2222   auto *SE = FAM.getCachedResult<ScalarEvolutionAnalysis>(*F);
2223   auto *DT = FAM.getCachedResult<DominatorTreeAnalysis>(*F);
2224 
2225   assert((LI && SE && DT) &&
2226          "Analyses required for indvarsimplify not available!");
2227 
2228   // Optional analyses.
2229   auto *TTI = FAM.getCachedResult<TargetIRAnalysis>(*F);
2230   auto *TLI = FAM.getCachedResult<TargetLibraryAnalysis>(*F);
2231 
2232   IndVarSimplify IVS(LI, SE, DT, DL, TLI, TTI);
2233   if (!IVS.run(&L))
2234     return PreservedAnalyses::all();
2235 
2236   // FIXME: This should also 'preserve the CFG'.
2237   return getLoopPassPreservedAnalyses();
2238 }
2239 
2240 namespace {
2241 struct IndVarSimplifyLegacyPass : public LoopPass {
2242   static char ID; // Pass identification, replacement for typeid
IndVarSimplifyLegacyPass__anon49122e510911::IndVarSimplifyLegacyPass2243   IndVarSimplifyLegacyPass() : LoopPass(ID) {
2244     initializeIndVarSimplifyLegacyPassPass(*PassRegistry::getPassRegistry());
2245   }
2246 
runOnLoop__anon49122e510911::IndVarSimplifyLegacyPass2247   bool runOnLoop(Loop *L, LPPassManager &LPM) override {
2248     if (skipLoop(L))
2249       return false;
2250 
2251     auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
2252     auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
2253     auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2254     auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
2255     auto *TLI = TLIP ? &TLIP->getTLI() : nullptr;
2256     auto *TTIP = getAnalysisIfAvailable<TargetTransformInfoWrapperPass>();
2257     auto *TTI = TTIP ? &TTIP->getTTI(*L->getHeader()->getParent()) : nullptr;
2258     const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
2259 
2260     IndVarSimplify IVS(LI, SE, DT, DL, TLI, TTI);
2261     return IVS.run(L);
2262   }
2263 
getAnalysisUsage__anon49122e510911::IndVarSimplifyLegacyPass2264   void getAnalysisUsage(AnalysisUsage &AU) const override {
2265     AU.setPreservesCFG();
2266     getLoopAnalysisUsage(AU);
2267   }
2268 };
2269 }
2270 
2271 char IndVarSimplifyLegacyPass::ID = 0;
2272 INITIALIZE_PASS_BEGIN(IndVarSimplifyLegacyPass, "indvars",
2273                       "Induction Variable Simplification", false, false)
INITIALIZE_PASS_DEPENDENCY(LoopPass)2274 INITIALIZE_PASS_DEPENDENCY(LoopPass)
2275 INITIALIZE_PASS_END(IndVarSimplifyLegacyPass, "indvars",
2276                     "Induction Variable Simplification", false, false)
2277 
2278 Pass *llvm::createIndVarSimplifyPass() {
2279   return new IndVarSimplifyLegacyPass();
2280 }
2281