• 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 #define DEBUG_TYPE "indvars"
28 #include "llvm/Transforms/Scalar.h"
29 #include "llvm/BasicBlock.h"
30 #include "llvm/Constants.h"
31 #include "llvm/Instructions.h"
32 #include "llvm/IntrinsicInst.h"
33 #include "llvm/LLVMContext.h"
34 #include "llvm/Type.h"
35 #include "llvm/Analysis/Dominators.h"
36 #include "llvm/Analysis/IVUsers.h"
37 #include "llvm/Analysis/ScalarEvolutionExpander.h"
38 #include "llvm/Analysis/LoopInfo.h"
39 #include "llvm/Analysis/LoopPass.h"
40 #include "llvm/Support/CFG.h"
41 #include "llvm/Support/CommandLine.h"
42 #include "llvm/Support/Debug.h"
43 #include "llvm/Support/raw_ostream.h"
44 #include "llvm/Transforms/Utils/Local.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/SimplifyIndVar.h"
47 #include "llvm/Target/TargetData.h"
48 #include "llvm/ADT/DenseMap.h"
49 #include "llvm/ADT/SmallVector.h"
50 #include "llvm/ADT/Statistic.h"
51 using namespace llvm;
52 
53 STATISTIC(NumRemoved     , "Number of aux indvars removed");
54 STATISTIC(NumWidened     , "Number of indvars widened");
55 STATISTIC(NumInserted    , "Number of canonical indvars added");
56 STATISTIC(NumReplaced    , "Number of exit values replaced");
57 STATISTIC(NumLFTR        , "Number of loop exit tests replaced");
58 STATISTIC(NumElimExt     , "Number of IV sign/zero extends eliminated");
59 STATISTIC(NumElimIV      , "Number of congruent IVs eliminated");
60 
61 namespace llvm {
62   cl::opt<bool> EnableIVRewrite(
63     "enable-iv-rewrite", cl::Hidden,
64     cl::desc("Enable canonical induction variable rewriting"));
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   cl::opt<bool> VerifyIndvars(
70     "verify-indvars", cl::Hidden,
71     cl::desc("Verify the ScalarEvolution result after running indvars"));
72 }
73 
74 namespace {
75   class IndVarSimplify : public LoopPass {
76     IVUsers         *IU;
77     LoopInfo        *LI;
78     ScalarEvolution *SE;
79     DominatorTree   *DT;
80     TargetData      *TD;
81 
82     SmallVector<WeakVH, 16> DeadInsts;
83     bool Changed;
84   public:
85 
86     static char ID; // Pass identification, replacement for typeid
IndVarSimplify()87     IndVarSimplify() : LoopPass(ID), IU(0), LI(0), SE(0), DT(0), TD(0),
88                        Changed(false) {
89       initializeIndVarSimplifyPass(*PassRegistry::getPassRegistry());
90     }
91 
92     virtual bool runOnLoop(Loop *L, LPPassManager &LPM);
93 
getAnalysisUsage(AnalysisUsage & AU) const94     virtual void getAnalysisUsage(AnalysisUsage &AU) const {
95       AU.addRequired<DominatorTree>();
96       AU.addRequired<LoopInfo>();
97       AU.addRequired<ScalarEvolution>();
98       AU.addRequiredID(LoopSimplifyID);
99       AU.addRequiredID(LCSSAID);
100       if (EnableIVRewrite)
101         AU.addRequired<IVUsers>();
102       AU.addPreserved<ScalarEvolution>();
103       AU.addPreservedID(LoopSimplifyID);
104       AU.addPreservedID(LCSSAID);
105       if (EnableIVRewrite)
106         AU.addPreserved<IVUsers>();
107       AU.setPreservesCFG();
108     }
109 
110   private:
releaseMemory()111     virtual void releaseMemory() {
112       DeadInsts.clear();
113     }
114 
115     bool isValidRewrite(Value *FromVal, Value *ToVal);
116 
117     void HandleFloatingPointIV(Loop *L, PHINode *PH);
118     void RewriteNonIntegerIVs(Loop *L);
119 
120     void SimplifyAndExtend(Loop *L, SCEVExpander &Rewriter, LPPassManager &LPM);
121 
122     void RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter);
123 
124     void RewriteIVExpressions(Loop *L, SCEVExpander &Rewriter);
125 
126     Value *LinearFunctionTestReplace(Loop *L, const SCEV *BackedgeTakenCount,
127                                      PHINode *IndVar, SCEVExpander &Rewriter);
128 
129     void SinkUnusedInvariants(Loop *L);
130   };
131 }
132 
133 char IndVarSimplify::ID = 0;
134 INITIALIZE_PASS_BEGIN(IndVarSimplify, "indvars",
135                 "Induction Variable Simplification", false, false)
INITIALIZE_PASS_DEPENDENCY(DominatorTree)136 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
137 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
138 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
139 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
140 INITIALIZE_PASS_DEPENDENCY(LCSSA)
141 INITIALIZE_PASS_DEPENDENCY(IVUsers)
142 INITIALIZE_PASS_END(IndVarSimplify, "indvars",
143                 "Induction Variable Simplification", false, false)
144 
145 Pass *llvm::createIndVarSimplifyPass() {
146   return new IndVarSimplify();
147 }
148 
149 /// isValidRewrite - Return true if the SCEV expansion generated by the
150 /// rewriter can replace the original value. SCEV guarantees that it
151 /// produces the same value, but the way it is produced may be illegal IR.
152 /// Ideally, this function will only be called for verification.
isValidRewrite(Value * FromVal,Value * ToVal)153 bool IndVarSimplify::isValidRewrite(Value *FromVal, Value *ToVal) {
154   // If an SCEV expression subsumed multiple pointers, its expansion could
155   // reassociate the GEP changing the base pointer. This is illegal because the
156   // final address produced by a GEP chain must be inbounds relative to its
157   // underlying object. Otherwise basic alias analysis, among other things,
158   // could fail in a dangerous way. Ultimately, SCEV will be improved to avoid
159   // producing an expression involving multiple pointers. Until then, we must
160   // bail out here.
161   //
162   // Retrieve the pointer operand of the GEP. Don't use GetUnderlyingObject
163   // because it understands lcssa phis while SCEV does not.
164   Value *FromPtr = FromVal;
165   Value *ToPtr = ToVal;
166   if (GEPOperator *GEP = dyn_cast<GEPOperator>(FromVal)) {
167     FromPtr = GEP->getPointerOperand();
168   }
169   if (GEPOperator *GEP = dyn_cast<GEPOperator>(ToVal)) {
170     ToPtr = GEP->getPointerOperand();
171   }
172   if (FromPtr != FromVal || ToPtr != ToVal) {
173     // Quickly check the common case
174     if (FromPtr == ToPtr)
175       return true;
176 
177     // SCEV may have rewritten an expression that produces the GEP's pointer
178     // operand. That's ok as long as the pointer operand has the same base
179     // pointer. Unlike GetUnderlyingObject(), getPointerBase() will find the
180     // base of a recurrence. This handles the case in which SCEV expansion
181     // converts a pointer type recurrence into a nonrecurrent pointer base
182     // indexed by an integer recurrence.
183     const SCEV *FromBase = SE->getPointerBase(SE->getSCEV(FromPtr));
184     const SCEV *ToBase = SE->getPointerBase(SE->getSCEV(ToPtr));
185     if (FromBase == ToBase)
186       return true;
187 
188     DEBUG(dbgs() << "INDVARS: GEP rewrite bail out "
189           << *FromBase << " != " << *ToBase << "\n");
190 
191     return false;
192   }
193   return true;
194 }
195 
196 /// Determine the insertion point for this user. By default, insert immediately
197 /// before the user. SCEVExpander or LICM will hoist loop invariants out of the
198 /// loop. For PHI nodes, there may be multiple uses, so compute the nearest
199 /// common dominator for the incoming blocks.
getInsertPointForUses(Instruction * User,Value * Def,DominatorTree * DT)200 static Instruction *getInsertPointForUses(Instruction *User, Value *Def,
201                                           DominatorTree *DT) {
202   PHINode *PHI = dyn_cast<PHINode>(User);
203   if (!PHI)
204     return User;
205 
206   Instruction *InsertPt = 0;
207   for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i) {
208     if (PHI->getIncomingValue(i) != Def)
209       continue;
210 
211     BasicBlock *InsertBB = PHI->getIncomingBlock(i);
212     if (!InsertPt) {
213       InsertPt = InsertBB->getTerminator();
214       continue;
215     }
216     InsertBB = DT->findNearestCommonDominator(InsertPt->getParent(), InsertBB);
217     InsertPt = InsertBB->getTerminator();
218   }
219   assert(InsertPt && "Missing phi operand");
220   assert((!isa<Instruction>(Def) ||
221           DT->dominates(cast<Instruction>(Def), InsertPt)) &&
222          "def does not dominate all uses");
223   return InsertPt;
224 }
225 
226 //===----------------------------------------------------------------------===//
227 // RewriteNonIntegerIVs and helpers. Prefer integer IVs.
228 //===----------------------------------------------------------------------===//
229 
230 /// ConvertToSInt - Convert APF to an integer, if possible.
ConvertToSInt(const APFloat & APF,int64_t & IntVal)231 static bool ConvertToSInt(const APFloat &APF, int64_t &IntVal) {
232   bool isExact = false;
233   if (&APF.getSemantics() == &APFloat::PPCDoubleDouble)
234     return false;
235   // See if we can convert this to an int64_t
236   uint64_t UIntVal;
237   if (APF.convertToInteger(&UIntVal, 64, true, APFloat::rmTowardZero,
238                            &isExact) != APFloat::opOK || !isExact)
239     return false;
240   IntVal = UIntVal;
241   return true;
242 }
243 
244 /// HandleFloatingPointIV - If the loop has floating induction variable
245 /// then insert corresponding integer induction variable if possible.
246 /// For example,
247 /// for(double i = 0; i < 10000; ++i)
248 ///   bar(i)
249 /// is converted into
250 /// for(int i = 0; i < 10000; ++i)
251 ///   bar((double)i);
252 ///
HandleFloatingPointIV(Loop * L,PHINode * PN)253 void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PN) {
254   unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
255   unsigned BackEdge     = IncomingEdge^1;
256 
257   // Check incoming value.
258   ConstantFP *InitValueVal =
259     dyn_cast<ConstantFP>(PN->getIncomingValue(IncomingEdge));
260 
261   int64_t InitValue;
262   if (!InitValueVal || !ConvertToSInt(InitValueVal->getValueAPF(), InitValue))
263     return;
264 
265   // Check IV increment. Reject this PN if increment operation is not
266   // an add or increment value can not be represented by an integer.
267   BinaryOperator *Incr =
268     dyn_cast<BinaryOperator>(PN->getIncomingValue(BackEdge));
269   if (Incr == 0 || Incr->getOpcode() != Instruction::FAdd) return;
270 
271   // If this is not an add of the PHI with a constantfp, or if the constant fp
272   // is not an integer, bail out.
273   ConstantFP *IncValueVal = dyn_cast<ConstantFP>(Incr->getOperand(1));
274   int64_t IncValue;
275   if (IncValueVal == 0 || Incr->getOperand(0) != PN ||
276       !ConvertToSInt(IncValueVal->getValueAPF(), IncValue))
277     return;
278 
279   // Check Incr uses. One user is PN and the other user is an exit condition
280   // used by the conditional terminator.
281   Value::use_iterator IncrUse = Incr->use_begin();
282   Instruction *U1 = cast<Instruction>(*IncrUse++);
283   if (IncrUse == Incr->use_end()) return;
284   Instruction *U2 = cast<Instruction>(*IncrUse++);
285   if (IncrUse != Incr->use_end()) return;
286 
287   // Find exit condition, which is an fcmp.  If it doesn't exist, or if it isn't
288   // only used by a branch, we can't transform it.
289   FCmpInst *Compare = dyn_cast<FCmpInst>(U1);
290   if (!Compare)
291     Compare = dyn_cast<FCmpInst>(U2);
292   if (Compare == 0 || !Compare->hasOneUse() ||
293       !isa<BranchInst>(Compare->use_back()))
294     return;
295 
296   BranchInst *TheBr = cast<BranchInst>(Compare->use_back());
297 
298   // We need to verify that the branch actually controls the iteration count
299   // of the loop.  If not, the new IV can overflow and no one will notice.
300   // The branch block must be in the loop and one of the successors must be out
301   // of the loop.
302   assert(TheBr->isConditional() && "Can't use fcmp if not conditional");
303   if (!L->contains(TheBr->getParent()) ||
304       (L->contains(TheBr->getSuccessor(0)) &&
305        L->contains(TheBr->getSuccessor(1))))
306     return;
307 
308 
309   // If it isn't a comparison with an integer-as-fp (the exit value), we can't
310   // transform it.
311   ConstantFP *ExitValueVal = dyn_cast<ConstantFP>(Compare->getOperand(1));
312   int64_t ExitValue;
313   if (ExitValueVal == 0 ||
314       !ConvertToSInt(ExitValueVal->getValueAPF(), ExitValue))
315     return;
316 
317   // Find new predicate for integer comparison.
318   CmpInst::Predicate NewPred = CmpInst::BAD_ICMP_PREDICATE;
319   switch (Compare->getPredicate()) {
320   default: return;  // Unknown comparison.
321   case CmpInst::FCMP_OEQ:
322   case CmpInst::FCMP_UEQ: NewPred = CmpInst::ICMP_EQ; break;
323   case CmpInst::FCMP_ONE:
324   case CmpInst::FCMP_UNE: NewPred = CmpInst::ICMP_NE; break;
325   case CmpInst::FCMP_OGT:
326   case CmpInst::FCMP_UGT: NewPred = CmpInst::ICMP_SGT; break;
327   case CmpInst::FCMP_OGE:
328   case CmpInst::FCMP_UGE: NewPred = CmpInst::ICMP_SGE; break;
329   case CmpInst::FCMP_OLT:
330   case CmpInst::FCMP_ULT: NewPred = CmpInst::ICMP_SLT; break;
331   case CmpInst::FCMP_OLE:
332   case CmpInst::FCMP_ULE: NewPred = CmpInst::ICMP_SLE; break;
333   }
334 
335   // We convert the floating point induction variable to a signed i32 value if
336   // we can.  This is only safe if the comparison will not overflow in a way
337   // that won't be trapped by the integer equivalent operations.  Check for this
338   // now.
339   // TODO: We could use i64 if it is native and the range requires it.
340 
341   // The start/stride/exit values must all fit in signed i32.
342   if (!isInt<32>(InitValue) || !isInt<32>(IncValue) || !isInt<32>(ExitValue))
343     return;
344 
345   // If not actually striding (add x, 0.0), avoid touching the code.
346   if (IncValue == 0)
347     return;
348 
349   // Positive and negative strides have different safety conditions.
350   if (IncValue > 0) {
351     // If we have a positive stride, we require the init to be less than the
352     // exit value.
353     if (InitValue >= ExitValue)
354       return;
355 
356     uint32_t Range = uint32_t(ExitValue-InitValue);
357     // Check for infinite loop, either:
358     // while (i <= Exit) or until (i > Exit)
359     if (NewPred == CmpInst::ICMP_SLE || NewPred == CmpInst::ICMP_SGT) {
360       if (++Range == 0) return;  // Range overflows.
361     }
362 
363     unsigned Leftover = Range % uint32_t(IncValue);
364 
365     // If this is an equality comparison, we require that the strided value
366     // exactly land on the exit value, otherwise the IV condition will wrap
367     // around and do things the fp IV wouldn't.
368     if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
369         Leftover != 0)
370       return;
371 
372     // If the stride would wrap around the i32 before exiting, we can't
373     // transform the IV.
374     if (Leftover != 0 && int32_t(ExitValue+IncValue) < ExitValue)
375       return;
376 
377   } else {
378     // If we have a negative stride, we require the init to be greater than the
379     // exit value.
380     if (InitValue <= ExitValue)
381       return;
382 
383     uint32_t Range = uint32_t(InitValue-ExitValue);
384     // Check for infinite loop, either:
385     // while (i >= Exit) or until (i < Exit)
386     if (NewPred == CmpInst::ICMP_SGE || NewPred == CmpInst::ICMP_SLT) {
387       if (++Range == 0) return;  // Range overflows.
388     }
389 
390     unsigned Leftover = Range % uint32_t(-IncValue);
391 
392     // If this is an equality comparison, we require that the strided value
393     // exactly land on the exit value, otherwise the IV condition will wrap
394     // around and do things the fp IV wouldn't.
395     if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
396         Leftover != 0)
397       return;
398 
399     // If the stride would wrap around the i32 before exiting, we can't
400     // transform the IV.
401     if (Leftover != 0 && int32_t(ExitValue+IncValue) > ExitValue)
402       return;
403   }
404 
405   IntegerType *Int32Ty = Type::getInt32Ty(PN->getContext());
406 
407   // Insert new integer induction variable.
408   PHINode *NewPHI = PHINode::Create(Int32Ty, 2, PN->getName()+".int", PN);
409   NewPHI->addIncoming(ConstantInt::get(Int32Ty, InitValue),
410                       PN->getIncomingBlock(IncomingEdge));
411 
412   Value *NewAdd =
413     BinaryOperator::CreateAdd(NewPHI, ConstantInt::get(Int32Ty, IncValue),
414                               Incr->getName()+".int", Incr);
415   NewPHI->addIncoming(NewAdd, PN->getIncomingBlock(BackEdge));
416 
417   ICmpInst *NewCompare = new ICmpInst(TheBr, NewPred, NewAdd,
418                                       ConstantInt::get(Int32Ty, ExitValue),
419                                       Compare->getName());
420 
421   // In the following deletions, PN may become dead and may be deleted.
422   // Use a WeakVH to observe whether this happens.
423   WeakVH WeakPH = PN;
424 
425   // Delete the old floating point exit comparison.  The branch starts using the
426   // new comparison.
427   NewCompare->takeName(Compare);
428   Compare->replaceAllUsesWith(NewCompare);
429   RecursivelyDeleteTriviallyDeadInstructions(Compare);
430 
431   // Delete the old floating point increment.
432   Incr->replaceAllUsesWith(UndefValue::get(Incr->getType()));
433   RecursivelyDeleteTriviallyDeadInstructions(Incr);
434 
435   // If the FP induction variable still has uses, this is because something else
436   // in the loop uses its value.  In order to canonicalize the induction
437   // variable, we chose to eliminate the IV and rewrite it in terms of an
438   // int->fp cast.
439   //
440   // We give preference to sitofp over uitofp because it is faster on most
441   // platforms.
442   if (WeakPH) {
443     Value *Conv = new SIToFPInst(NewPHI, PN->getType(), "indvar.conv",
444                                  PN->getParent()->getFirstInsertionPt());
445     PN->replaceAllUsesWith(Conv);
446     RecursivelyDeleteTriviallyDeadInstructions(PN);
447   }
448 
449   // Add a new IVUsers entry for the newly-created integer PHI.
450   if (IU)
451     IU->AddUsersIfInteresting(NewPHI);
452 
453   Changed = true;
454 }
455 
RewriteNonIntegerIVs(Loop * L)456 void IndVarSimplify::RewriteNonIntegerIVs(Loop *L) {
457   // First step.  Check to see if there are any floating-point recurrences.
458   // If there are, change them into integer recurrences, permitting analysis by
459   // the SCEV routines.
460   //
461   BasicBlock *Header = L->getHeader();
462 
463   SmallVector<WeakVH, 8> PHIs;
464   for (BasicBlock::iterator I = Header->begin();
465        PHINode *PN = dyn_cast<PHINode>(I); ++I)
466     PHIs.push_back(PN);
467 
468   for (unsigned i = 0, e = PHIs.size(); i != e; ++i)
469     if (PHINode *PN = dyn_cast_or_null<PHINode>(&*PHIs[i]))
470       HandleFloatingPointIV(L, PN);
471 
472   // If the loop previously had floating-point IV, ScalarEvolution
473   // may not have been able to compute a trip count. Now that we've done some
474   // re-writing, the trip count may be computable.
475   if (Changed)
476     SE->forgetLoop(L);
477 }
478 
479 //===----------------------------------------------------------------------===//
480 // RewriteLoopExitValues - Optimize IV users outside the loop.
481 // As a side effect, reduces the amount of IV processing within the loop.
482 //===----------------------------------------------------------------------===//
483 
484 /// RewriteLoopExitValues - Check to see if this loop has a computable
485 /// loop-invariant execution count.  If so, this means that we can compute the
486 /// final value of any expressions that are recurrent in the loop, and
487 /// substitute the exit values from the loop into any instructions outside of
488 /// the loop that use the final values of the current expressions.
489 ///
490 /// This is mostly redundant with the regular IndVarSimplify activities that
491 /// happen later, except that it's more powerful in some cases, because it's
492 /// able to brute-force evaluate arbitrary instructions as long as they have
493 /// constant operands at the beginning of the loop.
RewriteLoopExitValues(Loop * L,SCEVExpander & Rewriter)494 void IndVarSimplify::RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter) {
495   // Verify the input to the pass in already in LCSSA form.
496   assert(L->isLCSSAForm(*DT));
497 
498   SmallVector<BasicBlock*, 8> ExitBlocks;
499   L->getUniqueExitBlocks(ExitBlocks);
500 
501   // Find all values that are computed inside the loop, but used outside of it.
502   // Because of LCSSA, these values will only occur in LCSSA PHI Nodes.  Scan
503   // the exit blocks of the loop to find them.
504   for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
505     BasicBlock *ExitBB = ExitBlocks[i];
506 
507     // If there are no PHI nodes in this exit block, then no values defined
508     // inside the loop are used on this path, skip it.
509     PHINode *PN = dyn_cast<PHINode>(ExitBB->begin());
510     if (!PN) continue;
511 
512     unsigned NumPreds = PN->getNumIncomingValues();
513 
514     // Iterate over all of the PHI nodes.
515     BasicBlock::iterator BBI = ExitBB->begin();
516     while ((PN = dyn_cast<PHINode>(BBI++))) {
517       if (PN->use_empty())
518         continue; // dead use, don't replace it
519 
520       // SCEV only supports integer expressions for now.
521       if (!PN->getType()->isIntegerTy() && !PN->getType()->isPointerTy())
522         continue;
523 
524       // It's necessary to tell ScalarEvolution about this explicitly so that
525       // it can walk the def-use list and forget all SCEVs, as it may not be
526       // watching the PHI itself. Once the new exit value is in place, there
527       // may not be a def-use connection between the loop and every instruction
528       // which got a SCEVAddRecExpr for that loop.
529       SE->forgetValue(PN);
530 
531       // Iterate over all of the values in all the PHI nodes.
532       for (unsigned i = 0; i != NumPreds; ++i) {
533         // If the value being merged in is not integer or is not defined
534         // in the loop, skip it.
535         Value *InVal = PN->getIncomingValue(i);
536         if (!isa<Instruction>(InVal))
537           continue;
538 
539         // If this pred is for a subloop, not L itself, skip it.
540         if (LI->getLoopFor(PN->getIncomingBlock(i)) != L)
541           continue; // The Block is in a subloop, skip it.
542 
543         // Check that InVal is defined in the loop.
544         Instruction *Inst = cast<Instruction>(InVal);
545         if (!L->contains(Inst))
546           continue;
547 
548         // Okay, this instruction has a user outside of the current loop
549         // and varies predictably *inside* the loop.  Evaluate the value it
550         // contains when the loop exits, if possible.
551         const SCEV *ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop());
552         if (!SE->isLoopInvariant(ExitValue, L))
553           continue;
554 
555         Value *ExitVal = Rewriter.expandCodeFor(ExitValue, PN->getType(), Inst);
556 
557         DEBUG(dbgs() << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal << '\n'
558                      << "  LoopVal = " << *Inst << "\n");
559 
560         if (!isValidRewrite(Inst, ExitVal)) {
561           DeadInsts.push_back(ExitVal);
562           continue;
563         }
564         Changed = true;
565         ++NumReplaced;
566 
567         PN->setIncomingValue(i, ExitVal);
568 
569         // If this instruction is dead now, delete it.
570         RecursivelyDeleteTriviallyDeadInstructions(Inst);
571 
572         if (NumPreds == 1) {
573           // Completely replace a single-pred PHI. This is safe, because the
574           // NewVal won't be variant in the loop, so we don't need an LCSSA phi
575           // node anymore.
576           PN->replaceAllUsesWith(ExitVal);
577           RecursivelyDeleteTriviallyDeadInstructions(PN);
578         }
579       }
580       if (NumPreds != 1) {
581         // Clone the PHI and delete the original one. This lets IVUsers and
582         // any other maps purge the original user from their records.
583         PHINode *NewPN = cast<PHINode>(PN->clone());
584         NewPN->takeName(PN);
585         NewPN->insertBefore(PN);
586         PN->replaceAllUsesWith(NewPN);
587         PN->eraseFromParent();
588       }
589     }
590   }
591 
592   // The insertion point instruction may have been deleted; clear it out
593   // so that the rewriter doesn't trip over it later.
594   Rewriter.clearInsertPoint();
595 }
596 
597 //===----------------------------------------------------------------------===//
598 //  Rewrite IV users based on a canonical IV.
599 //  Only for use with -enable-iv-rewrite.
600 //===----------------------------------------------------------------------===//
601 
602 /// FIXME: It is an extremely bad idea to indvar substitute anything more
603 /// complex than affine induction variables.  Doing so will put expensive
604 /// polynomial evaluations inside of the loop, and the str reduction pass
605 /// currently can only reduce affine polynomials.  For now just disable
606 /// indvar subst on anything more complex than an affine addrec, unless
607 /// it can be expanded to a trivial value.
isSafe(const SCEV * S,const Loop * L,ScalarEvolution * SE)608 static bool isSafe(const SCEV *S, const Loop *L, ScalarEvolution *SE) {
609   // Loop-invariant values are safe.
610   if (SE->isLoopInvariant(S, L)) return true;
611 
612   // Affine addrecs are safe. Non-affine are not, because LSR doesn't know how
613   // to transform them into efficient code.
614   if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
615     return AR->isAffine();
616 
617   // An add is safe it all its operands are safe.
618   if (const SCEVCommutativeExpr *Commutative
619       = dyn_cast<SCEVCommutativeExpr>(S)) {
620     for (SCEVCommutativeExpr::op_iterator I = Commutative->op_begin(),
621          E = Commutative->op_end(); I != E; ++I)
622       if (!isSafe(*I, L, SE)) return false;
623     return true;
624   }
625 
626   // A cast is safe if its operand is.
627   if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
628     return isSafe(C->getOperand(), L, SE);
629 
630   // A udiv is safe if its operands are.
631   if (const SCEVUDivExpr *UD = dyn_cast<SCEVUDivExpr>(S))
632     return isSafe(UD->getLHS(), L, SE) &&
633            isSafe(UD->getRHS(), L, SE);
634 
635   // SCEVUnknown is always safe.
636   if (isa<SCEVUnknown>(S))
637     return true;
638 
639   // Nothing else is safe.
640   return false;
641 }
642 
RewriteIVExpressions(Loop * L,SCEVExpander & Rewriter)643 void IndVarSimplify::RewriteIVExpressions(Loop *L, SCEVExpander &Rewriter) {
644   // Rewrite all induction variable expressions in terms of the canonical
645   // induction variable.
646   //
647   // If there were induction variables of other sizes or offsets, manually
648   // add the offsets to the primary induction variable and cast, avoiding
649   // the need for the code evaluation methods to insert induction variables
650   // of different sizes.
651   for (IVUsers::iterator UI = IU->begin(), E = IU->end(); UI != E; ++UI) {
652     Value *Op = UI->getOperandValToReplace();
653     Type *UseTy = Op->getType();
654     Instruction *User = UI->getUser();
655 
656     // Compute the final addrec to expand into code.
657     const SCEV *AR = IU->getReplacementExpr(*UI);
658 
659     // Evaluate the expression out of the loop, if possible.
660     if (!L->contains(UI->getUser())) {
661       const SCEV *ExitVal = SE->getSCEVAtScope(AR, L->getParentLoop());
662       if (SE->isLoopInvariant(ExitVal, L))
663         AR = ExitVal;
664     }
665 
666     // FIXME: It is an extremely bad idea to indvar substitute anything more
667     // complex than affine induction variables.  Doing so will put expensive
668     // polynomial evaluations inside of the loop, and the str reduction pass
669     // currently can only reduce affine polynomials.  For now just disable
670     // indvar subst on anything more complex than an affine addrec, unless
671     // it can be expanded to a trivial value.
672     if (!isSafe(AR, L, SE))
673       continue;
674 
675     // Determine the insertion point for this user. By default, insert
676     // immediately before the user. The SCEVExpander class will automatically
677     // hoist loop invariants out of the loop. For PHI nodes, there may be
678     // multiple uses, so compute the nearest common dominator for the
679     // incoming blocks.
680     Instruction *InsertPt = getInsertPointForUses(User, Op, DT);
681 
682     // Now expand it into actual Instructions and patch it into place.
683     Value *NewVal = Rewriter.expandCodeFor(AR, UseTy, InsertPt);
684 
685     DEBUG(dbgs() << "INDVARS: Rewrote IV '" << *AR << "' " << *Op << '\n'
686                  << "   into = " << *NewVal << "\n");
687 
688     if (!isValidRewrite(Op, NewVal)) {
689       DeadInsts.push_back(NewVal);
690       continue;
691     }
692     // Inform ScalarEvolution that this value is changing. The change doesn't
693     // affect its value, but it does potentially affect which use lists the
694     // value will be on after the replacement, which affects ScalarEvolution's
695     // ability to walk use lists and drop dangling pointers when a value is
696     // deleted.
697     SE->forgetValue(User);
698 
699     // Patch the new value into place.
700     if (Op->hasName())
701       NewVal->takeName(Op);
702     if (Instruction *NewValI = dyn_cast<Instruction>(NewVal))
703       NewValI->setDebugLoc(User->getDebugLoc());
704     User->replaceUsesOfWith(Op, NewVal);
705     UI->setOperandValToReplace(NewVal);
706 
707     ++NumRemoved;
708     Changed = true;
709 
710     // The old value may be dead now.
711     DeadInsts.push_back(Op);
712   }
713 }
714 
715 //===----------------------------------------------------------------------===//
716 //  IV Widening - Extend the width of an IV to cover its widest uses.
717 //===----------------------------------------------------------------------===//
718 
719 namespace {
720   // Collect information about induction variables that are used by sign/zero
721   // extend operations. This information is recorded by CollectExtend and
722   // provides the input to WidenIV.
723   struct WideIVInfo {
724     PHINode *NarrowIV;
725     Type *WidestNativeType; // Widest integer type created [sz]ext
726     bool IsSigned;          // Was an sext user seen before a zext?
727 
WideIVInfo__anonf7e225bb0211::WideIVInfo728     WideIVInfo() : NarrowIV(0), WidestNativeType(0), IsSigned(false) {}
729   };
730 
731   class WideIVVisitor : public IVVisitor {
732     ScalarEvolution *SE;
733     const TargetData *TD;
734 
735   public:
736     WideIVInfo WI;
737 
WideIVVisitor(PHINode * NarrowIV,ScalarEvolution * SCEV,const TargetData * TData)738     WideIVVisitor(PHINode *NarrowIV, ScalarEvolution *SCEV,
739                   const TargetData *TData) :
740       SE(SCEV), TD(TData) { WI.NarrowIV = NarrowIV; }
741 
742     // Implement the interface used by simplifyUsersOfIV.
743     virtual void visitCast(CastInst *Cast);
744   };
745 }
746 
747 /// visitCast - Update information about the induction variable that is
748 /// extended by this sign or zero extend operation. This is used to determine
749 /// the final width of the IV before actually widening it.
visitCast(CastInst * Cast)750 void WideIVVisitor::visitCast(CastInst *Cast) {
751   bool IsSigned = Cast->getOpcode() == Instruction::SExt;
752   if (!IsSigned && Cast->getOpcode() != Instruction::ZExt)
753     return;
754 
755   Type *Ty = Cast->getType();
756   uint64_t Width = SE->getTypeSizeInBits(Ty);
757   if (TD && !TD->isLegalInteger(Width))
758     return;
759 
760   if (!WI.WidestNativeType) {
761     WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
762     WI.IsSigned = IsSigned;
763     return;
764   }
765 
766   // We extend the IV to satisfy the sign of its first user, arbitrarily.
767   if (WI.IsSigned != IsSigned)
768     return;
769 
770   if (Width > SE->getTypeSizeInBits(WI.WidestNativeType))
771     WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
772 }
773 
774 namespace {
775 
776 /// NarrowIVDefUse - Record a link in the Narrow IV def-use chain along with the
777 /// WideIV that computes the same value as the Narrow IV def.  This avoids
778 /// caching Use* pointers.
779 struct NarrowIVDefUse {
780   Instruction *NarrowDef;
781   Instruction *NarrowUse;
782   Instruction *WideDef;
783 
NarrowIVDefUse__anonf7e225bb0311::NarrowIVDefUse784   NarrowIVDefUse(): NarrowDef(0), NarrowUse(0), WideDef(0) {}
785 
NarrowIVDefUse__anonf7e225bb0311::NarrowIVDefUse786   NarrowIVDefUse(Instruction *ND, Instruction *NU, Instruction *WD):
787     NarrowDef(ND), NarrowUse(NU), WideDef(WD) {}
788 };
789 
790 /// WidenIV - The goal of this transform is to remove sign and zero extends
791 /// without creating any new induction variables. To do this, it creates a new
792 /// phi of the wider type and redirects all users, either removing extends or
793 /// inserting truncs whenever we stop propagating the type.
794 ///
795 class WidenIV {
796   // Parameters
797   PHINode *OrigPhi;
798   Type *WideType;
799   bool IsSigned;
800 
801   // Context
802   LoopInfo        *LI;
803   Loop            *L;
804   ScalarEvolution *SE;
805   DominatorTree   *DT;
806 
807   // Result
808   PHINode *WidePhi;
809   Instruction *WideInc;
810   const SCEV *WideIncExpr;
811   SmallVectorImpl<WeakVH> &DeadInsts;
812 
813   SmallPtrSet<Instruction*,16> Widened;
814   SmallVector<NarrowIVDefUse, 8> NarrowIVUsers;
815 
816 public:
WidenIV(const WideIVInfo & WI,LoopInfo * LInfo,ScalarEvolution * SEv,DominatorTree * DTree,SmallVectorImpl<WeakVH> & DI)817   WidenIV(const WideIVInfo &WI, LoopInfo *LInfo,
818           ScalarEvolution *SEv, DominatorTree *DTree,
819           SmallVectorImpl<WeakVH> &DI) :
820     OrigPhi(WI.NarrowIV),
821     WideType(WI.WidestNativeType),
822     IsSigned(WI.IsSigned),
823     LI(LInfo),
824     L(LI->getLoopFor(OrigPhi->getParent())),
825     SE(SEv),
826     DT(DTree),
827     WidePhi(0),
828     WideInc(0),
829     WideIncExpr(0),
830     DeadInsts(DI) {
831     assert(L->getHeader() == OrigPhi->getParent() && "Phi must be an IV");
832   }
833 
834   PHINode *CreateWideIV(SCEVExpander &Rewriter);
835 
836 protected:
837   Value *getExtend(Value *NarrowOper, Type *WideType, bool IsSigned,
838                    Instruction *Use);
839 
840   Instruction *CloneIVUser(NarrowIVDefUse DU);
841 
842   const SCEVAddRecExpr *GetWideRecurrence(Instruction *NarrowUse);
843 
844   const SCEVAddRecExpr* GetExtendedOperandRecurrence(NarrowIVDefUse DU);
845 
846   Instruction *WidenIVUse(NarrowIVDefUse DU);
847 
848   void pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef);
849 };
850 } // anonymous namespace
851 
852 /// isLoopInvariant - Perform a quick domtree based check for loop invariance
853 /// assuming that V is used within the loop. LoopInfo::isLoopInvariant() seems
854 /// gratuitous for this purpose.
isLoopInvariant(Value * V,const Loop * L,const DominatorTree * DT)855 static bool isLoopInvariant(Value *V, const Loop *L, const DominatorTree *DT) {
856   Instruction *Inst = dyn_cast<Instruction>(V);
857   if (!Inst)
858     return true;
859 
860   return DT->properlyDominates(Inst->getParent(), L->getHeader());
861 }
862 
getExtend(Value * NarrowOper,Type * WideType,bool IsSigned,Instruction * Use)863 Value *WidenIV::getExtend(Value *NarrowOper, Type *WideType, bool IsSigned,
864                           Instruction *Use) {
865   // Set the debug location and conservative insertion point.
866   IRBuilder<> Builder(Use);
867   // Hoist the insertion point into loop preheaders as far as possible.
868   for (const Loop *L = LI->getLoopFor(Use->getParent());
869        L && L->getLoopPreheader() && isLoopInvariant(NarrowOper, L, DT);
870        L = L->getParentLoop())
871     Builder.SetInsertPoint(L->getLoopPreheader()->getTerminator());
872 
873   return IsSigned ? Builder.CreateSExt(NarrowOper, WideType) :
874                     Builder.CreateZExt(NarrowOper, WideType);
875 }
876 
877 /// CloneIVUser - Instantiate a wide operation to replace a narrow
878 /// operation. This only needs to handle operations that can evaluation to
879 /// SCEVAddRec. It can safely return 0 for any operation we decide not to clone.
CloneIVUser(NarrowIVDefUse DU)880 Instruction *WidenIV::CloneIVUser(NarrowIVDefUse DU) {
881   unsigned Opcode = DU.NarrowUse->getOpcode();
882   switch (Opcode) {
883   default:
884     return 0;
885   case Instruction::Add:
886   case Instruction::Mul:
887   case Instruction::UDiv:
888   case Instruction::Sub:
889   case Instruction::And:
890   case Instruction::Or:
891   case Instruction::Xor:
892   case Instruction::Shl:
893   case Instruction::LShr:
894   case Instruction::AShr:
895     DEBUG(dbgs() << "Cloning IVUser: " << *DU.NarrowUse << "\n");
896 
897     // Replace NarrowDef operands with WideDef. Otherwise, we don't know
898     // anything about the narrow operand yet so must insert a [sz]ext. It is
899     // probably loop invariant and will be folded or hoisted. If it actually
900     // comes from a widened IV, it should be removed during a future call to
901     // WidenIVUse.
902     Value *LHS = (DU.NarrowUse->getOperand(0) == DU.NarrowDef) ? DU.WideDef :
903       getExtend(DU.NarrowUse->getOperand(0), WideType, IsSigned, DU.NarrowUse);
904     Value *RHS = (DU.NarrowUse->getOperand(1) == DU.NarrowDef) ? DU.WideDef :
905       getExtend(DU.NarrowUse->getOperand(1), WideType, IsSigned, DU.NarrowUse);
906 
907     BinaryOperator *NarrowBO = cast<BinaryOperator>(DU.NarrowUse);
908     BinaryOperator *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(),
909                                                     LHS, RHS,
910                                                     NarrowBO->getName());
911     IRBuilder<> Builder(DU.NarrowUse);
912     Builder.Insert(WideBO);
913     if (const OverflowingBinaryOperator *OBO =
914         dyn_cast<OverflowingBinaryOperator>(NarrowBO)) {
915       if (OBO->hasNoUnsignedWrap()) WideBO->setHasNoUnsignedWrap();
916       if (OBO->hasNoSignedWrap()) WideBO->setHasNoSignedWrap();
917     }
918     return WideBO;
919   }
920   llvm_unreachable(0);
921 }
922 
923 /// No-wrap operations can transfer sign extension of their result to their
924 /// operands. Generate the SCEV value for the widened operation without
925 /// actually modifying the IR yet. If the expression after extending the
926 /// operands is an AddRec for this loop, return it.
GetExtendedOperandRecurrence(NarrowIVDefUse DU)927 const SCEVAddRecExpr* WidenIV::GetExtendedOperandRecurrence(NarrowIVDefUse DU) {
928   // Handle the common case of add<nsw/nuw>
929   if (DU.NarrowUse->getOpcode() != Instruction::Add)
930     return 0;
931 
932   // One operand (NarrowDef) has already been extended to WideDef. Now determine
933   // if extending the other will lead to a recurrence.
934   unsigned ExtendOperIdx = DU.NarrowUse->getOperand(0) == DU.NarrowDef ? 1 : 0;
935   assert(DU.NarrowUse->getOperand(1-ExtendOperIdx) == DU.NarrowDef && "bad DU");
936 
937   const SCEV *ExtendOperExpr = 0;
938   const OverflowingBinaryOperator *OBO =
939     cast<OverflowingBinaryOperator>(DU.NarrowUse);
940   if (IsSigned && OBO->hasNoSignedWrap())
941     ExtendOperExpr = SE->getSignExtendExpr(
942       SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType);
943   else if(!IsSigned && OBO->hasNoUnsignedWrap())
944     ExtendOperExpr = SE->getZeroExtendExpr(
945       SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType);
946   else
947     return 0;
948 
949   const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(
950     SE->getAddExpr(SE->getSCEV(DU.WideDef), ExtendOperExpr,
951                    IsSigned ? SCEV::FlagNSW : SCEV::FlagNUW));
952 
953   if (!AddRec || AddRec->getLoop() != L)
954     return 0;
955   return AddRec;
956 }
957 
958 /// GetWideRecurrence - Is this instruction potentially interesting from
959 /// IVUsers' perspective after widening it's type? In other words, can the
960 /// extend be safely hoisted out of the loop with SCEV reducing the value to a
961 /// recurrence on the same loop. If so, return the sign or zero extended
962 /// recurrence. Otherwise return NULL.
GetWideRecurrence(Instruction * NarrowUse)963 const SCEVAddRecExpr *WidenIV::GetWideRecurrence(Instruction *NarrowUse) {
964   if (!SE->isSCEVable(NarrowUse->getType()))
965     return 0;
966 
967   const SCEV *NarrowExpr = SE->getSCEV(NarrowUse);
968   if (SE->getTypeSizeInBits(NarrowExpr->getType())
969       >= SE->getTypeSizeInBits(WideType)) {
970     // NarrowUse implicitly widens its operand. e.g. a gep with a narrow
971     // index. So don't follow this use.
972     return 0;
973   }
974 
975   const SCEV *WideExpr = IsSigned ?
976     SE->getSignExtendExpr(NarrowExpr, WideType) :
977     SE->getZeroExtendExpr(NarrowExpr, WideType);
978   const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(WideExpr);
979   if (!AddRec || AddRec->getLoop() != L)
980     return 0;
981   return AddRec;
982 }
983 
984 /// WidenIVUse - Determine whether an individual user of the narrow IV can be
985 /// widened. If so, return the wide clone of the user.
WidenIVUse(NarrowIVDefUse DU)986 Instruction *WidenIV::WidenIVUse(NarrowIVDefUse DU) {
987 
988   // Stop traversing the def-use chain at inner-loop phis or post-loop phis.
989   if (isa<PHINode>(DU.NarrowUse) &&
990       LI->getLoopFor(DU.NarrowUse->getParent()) != L)
991     return 0;
992 
993   // Our raison d'etre! Eliminate sign and zero extension.
994   if (IsSigned ? isa<SExtInst>(DU.NarrowUse) : isa<ZExtInst>(DU.NarrowUse)) {
995     Value *NewDef = DU.WideDef;
996     if (DU.NarrowUse->getType() != WideType) {
997       unsigned CastWidth = SE->getTypeSizeInBits(DU.NarrowUse->getType());
998       unsigned IVWidth = SE->getTypeSizeInBits(WideType);
999       if (CastWidth < IVWidth) {
1000         // The cast isn't as wide as the IV, so insert a Trunc.
1001         IRBuilder<> Builder(DU.NarrowUse);
1002         NewDef = Builder.CreateTrunc(DU.WideDef, DU.NarrowUse->getType());
1003       }
1004       else {
1005         // A wider extend was hidden behind a narrower one. This may induce
1006         // another round of IV widening in which the intermediate IV becomes
1007         // dead. It should be very rare.
1008         DEBUG(dbgs() << "INDVARS: New IV " << *WidePhi
1009               << " not wide enough to subsume " << *DU.NarrowUse << "\n");
1010         DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef);
1011         NewDef = DU.NarrowUse;
1012       }
1013     }
1014     if (NewDef != DU.NarrowUse) {
1015       DEBUG(dbgs() << "INDVARS: eliminating " << *DU.NarrowUse
1016             << " replaced by " << *DU.WideDef << "\n");
1017       ++NumElimExt;
1018       DU.NarrowUse->replaceAllUsesWith(NewDef);
1019       DeadInsts.push_back(DU.NarrowUse);
1020     }
1021     // Now that the extend is gone, we want to expose it's uses for potential
1022     // further simplification. We don't need to directly inform SimplifyIVUsers
1023     // of the new users, because their parent IV will be processed later as a
1024     // new loop phi. If we preserved IVUsers analysis, we would also want to
1025     // push the uses of WideDef here.
1026 
1027     // No further widening is needed. The deceased [sz]ext had done it for us.
1028     return 0;
1029   }
1030 
1031   // Does this user itself evaluate to a recurrence after widening?
1032   const SCEVAddRecExpr *WideAddRec = GetWideRecurrence(DU.NarrowUse);
1033   if (!WideAddRec) {
1034       WideAddRec = GetExtendedOperandRecurrence(DU);
1035   }
1036   if (!WideAddRec) {
1037     // This user does not evaluate to a recurence after widening, so don't
1038     // follow it. Instead insert a Trunc to kill off the original use,
1039     // eventually isolating the original narrow IV so it can be removed.
1040     IRBuilder<> Builder(getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT));
1041     Value *Trunc = Builder.CreateTrunc(DU.WideDef, DU.NarrowDef->getType());
1042     DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, Trunc);
1043     return 0;
1044   }
1045   // Assume block terminators cannot evaluate to a recurrence. We can't to
1046   // insert a Trunc after a terminator if there happens to be a critical edge.
1047   assert(DU.NarrowUse != DU.NarrowUse->getParent()->getTerminator() &&
1048          "SCEV is not expected to evaluate a block terminator");
1049 
1050   // Reuse the IV increment that SCEVExpander created as long as it dominates
1051   // NarrowUse.
1052   Instruction *WideUse = 0;
1053   if (WideAddRec == WideIncExpr
1054       && SCEVExpander::hoistStep(WideInc, DU.NarrowUse, DT))
1055     WideUse = WideInc;
1056   else {
1057     WideUse = CloneIVUser(DU);
1058     if (!WideUse)
1059       return 0;
1060   }
1061   // Evaluation of WideAddRec ensured that the narrow expression could be
1062   // extended outside the loop without overflow. This suggests that the wide use
1063   // evaluates to the same expression as the extended narrow use, but doesn't
1064   // absolutely guarantee it. Hence the following failsafe check. In rare cases
1065   // where it fails, we simply throw away the newly created wide use.
1066   if (WideAddRec != SE->getSCEV(WideUse)) {
1067     DEBUG(dbgs() << "Wide use expression mismatch: " << *WideUse
1068           << ": " << *SE->getSCEV(WideUse) << " != " << *WideAddRec << "\n");
1069     DeadInsts.push_back(WideUse);
1070     return 0;
1071   }
1072 
1073   // Returning WideUse pushes it on the worklist.
1074   return WideUse;
1075 }
1076 
1077 /// pushNarrowIVUsers - Add eligible users of NarrowDef to NarrowIVUsers.
1078 ///
pushNarrowIVUsers(Instruction * NarrowDef,Instruction * WideDef)1079 void WidenIV::pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef) {
1080   for (Value::use_iterator UI = NarrowDef->use_begin(),
1081          UE = NarrowDef->use_end(); UI != UE; ++UI) {
1082     Instruction *NarrowUse = cast<Instruction>(*UI);
1083 
1084     // Handle data flow merges and bizarre phi cycles.
1085     if (!Widened.insert(NarrowUse))
1086       continue;
1087 
1088     NarrowIVUsers.push_back(NarrowIVDefUse(NarrowDef, NarrowUse, WideDef));
1089   }
1090 }
1091 
1092 /// CreateWideIV - Process a single induction variable. First use the
1093 /// SCEVExpander to create a wide induction variable that evaluates to the same
1094 /// recurrence as the original narrow IV. Then use a worklist to forward
1095 /// traverse the narrow IV's def-use chain. After WidenIVUse has processed all
1096 /// interesting IV users, the narrow IV will be isolated for removal by
1097 /// DeleteDeadPHIs.
1098 ///
1099 /// It would be simpler to delete uses as they are processed, but we must avoid
1100 /// invalidating SCEV expressions.
1101 ///
CreateWideIV(SCEVExpander & Rewriter)1102 PHINode *WidenIV::CreateWideIV(SCEVExpander &Rewriter) {
1103   // Is this phi an induction variable?
1104   const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(OrigPhi));
1105   if (!AddRec)
1106     return NULL;
1107 
1108   // Widen the induction variable expression.
1109   const SCEV *WideIVExpr = IsSigned ?
1110     SE->getSignExtendExpr(AddRec, WideType) :
1111     SE->getZeroExtendExpr(AddRec, WideType);
1112 
1113   assert(SE->getEffectiveSCEVType(WideIVExpr->getType()) == WideType &&
1114          "Expect the new IV expression to preserve its type");
1115 
1116   // Can the IV be extended outside the loop without overflow?
1117   AddRec = dyn_cast<SCEVAddRecExpr>(WideIVExpr);
1118   if (!AddRec || AddRec->getLoop() != L)
1119     return NULL;
1120 
1121   // An AddRec must have loop-invariant operands. Since this AddRec is
1122   // materialized by a loop header phi, the expression cannot have any post-loop
1123   // operands, so they must dominate the loop header.
1124   assert(SE->properlyDominates(AddRec->getStart(), L->getHeader()) &&
1125          SE->properlyDominates(AddRec->getStepRecurrence(*SE), L->getHeader())
1126          && "Loop header phi recurrence inputs do not dominate the loop");
1127 
1128   // The rewriter provides a value for the desired IV expression. This may
1129   // either find an existing phi or materialize a new one. Either way, we
1130   // expect a well-formed cyclic phi-with-increments. i.e. any operand not part
1131   // of the phi-SCC dominates the loop entry.
1132   Instruction *InsertPt = L->getHeader()->begin();
1133   WidePhi = cast<PHINode>(Rewriter.expandCodeFor(AddRec, WideType, InsertPt));
1134 
1135   // Remembering the WideIV increment generated by SCEVExpander allows
1136   // WidenIVUse to reuse it when widening the narrow IV's increment. We don't
1137   // employ a general reuse mechanism because the call above is the only call to
1138   // SCEVExpander. Henceforth, we produce 1-to-1 narrow to wide uses.
1139   if (BasicBlock *LatchBlock = L->getLoopLatch()) {
1140     WideInc =
1141       cast<Instruction>(WidePhi->getIncomingValueForBlock(LatchBlock));
1142     WideIncExpr = SE->getSCEV(WideInc);
1143   }
1144 
1145   DEBUG(dbgs() << "Wide IV: " << *WidePhi << "\n");
1146   ++NumWidened;
1147 
1148   // Traverse the def-use chain using a worklist starting at the original IV.
1149   assert(Widened.empty() && NarrowIVUsers.empty() && "expect initial state" );
1150 
1151   Widened.insert(OrigPhi);
1152   pushNarrowIVUsers(OrigPhi, WidePhi);
1153 
1154   while (!NarrowIVUsers.empty()) {
1155     NarrowIVDefUse DU = NarrowIVUsers.pop_back_val();
1156 
1157     // Process a def-use edge. This may replace the use, so don't hold a
1158     // use_iterator across it.
1159     Instruction *WideUse = WidenIVUse(DU);
1160 
1161     // Follow all def-use edges from the previous narrow use.
1162     if (WideUse)
1163       pushNarrowIVUsers(DU.NarrowUse, WideUse);
1164 
1165     // WidenIVUse may have removed the def-use edge.
1166     if (DU.NarrowDef->use_empty())
1167       DeadInsts.push_back(DU.NarrowDef);
1168   }
1169   return WidePhi;
1170 }
1171 
1172 //===----------------------------------------------------------------------===//
1173 //  Simplification of IV users based on SCEV evaluation.
1174 //===----------------------------------------------------------------------===//
1175 
1176 
1177 /// SimplifyAndExtend - Iteratively perform simplification on a worklist of IV
1178 /// users. Each successive simplification may push more users which may
1179 /// themselves be candidates for simplification.
1180 ///
1181 /// Sign/Zero extend elimination is interleaved with IV simplification.
1182 ///
SimplifyAndExtend(Loop * L,SCEVExpander & Rewriter,LPPassManager & LPM)1183 void IndVarSimplify::SimplifyAndExtend(Loop *L,
1184                                        SCEVExpander &Rewriter,
1185                                        LPPassManager &LPM) {
1186   SmallVector<WideIVInfo, 8> WideIVs;
1187 
1188   SmallVector<PHINode*, 8> LoopPhis;
1189   for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
1190     LoopPhis.push_back(cast<PHINode>(I));
1191   }
1192   // Each round of simplification iterates through the SimplifyIVUsers worklist
1193   // for all current phis, then determines whether any IVs can be
1194   // widened. Widening adds new phis to LoopPhis, inducing another round of
1195   // simplification on the wide IVs.
1196   while (!LoopPhis.empty()) {
1197     // Evaluate as many IV expressions as possible before widening any IVs. This
1198     // forces SCEV to set no-wrap flags before evaluating sign/zero
1199     // extension. The first time SCEV attempts to normalize sign/zero extension,
1200     // the result becomes final. So for the most predictable results, we delay
1201     // evaluation of sign/zero extend evaluation until needed, and avoid running
1202     // other SCEV based analysis prior to SimplifyAndExtend.
1203     do {
1204       PHINode *CurrIV = LoopPhis.pop_back_val();
1205 
1206       // Information about sign/zero extensions of CurrIV.
1207       WideIVVisitor WIV(CurrIV, SE, TD);
1208 
1209       Changed |= simplifyUsersOfIV(CurrIV, SE, &LPM, DeadInsts, &WIV);
1210 
1211       if (WIV.WI.WidestNativeType) {
1212         WideIVs.push_back(WIV.WI);
1213       }
1214     } while(!LoopPhis.empty());
1215 
1216     for (; !WideIVs.empty(); WideIVs.pop_back()) {
1217       WidenIV Widener(WideIVs.back(), LI, SE, DT, DeadInsts);
1218       if (PHINode *WidePhi = Widener.CreateWideIV(Rewriter)) {
1219         Changed = true;
1220         LoopPhis.push_back(WidePhi);
1221       }
1222     }
1223   }
1224 }
1225 
1226 //===----------------------------------------------------------------------===//
1227 //  LinearFunctionTestReplace and its kin. Rewrite the loop exit condition.
1228 //===----------------------------------------------------------------------===//
1229 
1230 /// Check for expressions that ScalarEvolution generates to compute
1231 /// BackedgeTakenInfo. If these expressions have not been reduced, then
1232 /// expanding them may incur additional cost (albeit in the loop preheader).
isHighCostExpansion(const SCEV * S,BranchInst * BI,ScalarEvolution * SE)1233 static bool isHighCostExpansion(const SCEV *S, BranchInst *BI,
1234                                 ScalarEvolution *SE) {
1235   // If the backedge-taken count is a UDiv, it's very likely a UDiv that
1236   // ScalarEvolution's HowFarToZero or HowManyLessThans produced to compute a
1237   // precise expression, rather than a UDiv from the user's code. If we can't
1238   // find a UDiv in the code with some simple searching, assume the former and
1239   // forego rewriting the loop.
1240   if (isa<SCEVUDivExpr>(S)) {
1241     ICmpInst *OrigCond = dyn_cast<ICmpInst>(BI->getCondition());
1242     if (!OrigCond) return true;
1243     const SCEV *R = SE->getSCEV(OrigCond->getOperand(1));
1244     R = SE->getMinusSCEV(R, SE->getConstant(R->getType(), 1));
1245     if (R != S) {
1246       const SCEV *L = SE->getSCEV(OrigCond->getOperand(0));
1247       L = SE->getMinusSCEV(L, SE->getConstant(L->getType(), 1));
1248       if (L != S)
1249         return true;
1250     }
1251   }
1252 
1253   if (EnableIVRewrite)
1254     return false;
1255 
1256   // Recurse past add expressions, which commonly occur in the
1257   // BackedgeTakenCount. They may already exist in program code, and if not,
1258   // they are not too expensive rematerialize.
1259   if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
1260     for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
1261          I != E; ++I) {
1262       if (isHighCostExpansion(*I, BI, SE))
1263         return true;
1264     }
1265     return false;
1266   }
1267 
1268   // HowManyLessThans uses a Max expression whenever the loop is not guarded by
1269   // the exit condition.
1270   if (isa<SCEVSMaxExpr>(S) || isa<SCEVUMaxExpr>(S))
1271     return true;
1272 
1273   // If we haven't recognized an expensive SCEV patter, assume its an expression
1274   // produced by program code.
1275   return false;
1276 }
1277 
1278 /// canExpandBackedgeTakenCount - Return true if this loop's backedge taken
1279 /// count expression can be safely and cheaply expanded into an instruction
1280 /// sequence that can be used by LinearFunctionTestReplace.
canExpandBackedgeTakenCount(Loop * L,ScalarEvolution * SE)1281 static bool canExpandBackedgeTakenCount(Loop *L, ScalarEvolution *SE) {
1282   const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
1283   if (isa<SCEVCouldNotCompute>(BackedgeTakenCount) ||
1284       BackedgeTakenCount->isZero())
1285     return false;
1286 
1287   if (!L->getExitingBlock())
1288     return false;
1289 
1290   // Can't rewrite non-branch yet.
1291   BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
1292   if (!BI)
1293     return false;
1294 
1295   if (isHighCostExpansion(BackedgeTakenCount, BI, SE))
1296     return false;
1297 
1298   return true;
1299 }
1300 
1301 /// getBackedgeIVType - Get the widest type used by the loop test after peeking
1302 /// through Truncs.
1303 ///
1304 /// TODO: Unnecessary when ForceLFTR is removed.
getBackedgeIVType(Loop * L)1305 static Type *getBackedgeIVType(Loop *L) {
1306   if (!L->getExitingBlock())
1307     return 0;
1308 
1309   // Can't rewrite non-branch yet.
1310   BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
1311   if (!BI)
1312     return 0;
1313 
1314   ICmpInst *Cond = dyn_cast<ICmpInst>(BI->getCondition());
1315   if (!Cond)
1316     return 0;
1317 
1318   Type *Ty = 0;
1319   for(User::op_iterator OI = Cond->op_begin(), OE = Cond->op_end();
1320       OI != OE; ++OI) {
1321     assert((!Ty || Ty == (*OI)->getType()) && "bad icmp operand types");
1322     TruncInst *Trunc = dyn_cast<TruncInst>(*OI);
1323     if (!Trunc)
1324       continue;
1325 
1326     return Trunc->getSrcTy();
1327   }
1328   return Ty;
1329 }
1330 
1331 /// getLoopPhiForCounter - Return the loop header phi IFF IncV adds a loop
1332 /// invariant value to the phi.
getLoopPhiForCounter(Value * IncV,Loop * L,DominatorTree * DT)1333 static PHINode *getLoopPhiForCounter(Value *IncV, Loop *L, DominatorTree *DT) {
1334   Instruction *IncI = dyn_cast<Instruction>(IncV);
1335   if (!IncI)
1336     return 0;
1337 
1338   switch (IncI->getOpcode()) {
1339   case Instruction::Add:
1340   case Instruction::Sub:
1341     break;
1342   case Instruction::GetElementPtr:
1343     // An IV counter must preserve its type.
1344     if (IncI->getNumOperands() == 2)
1345       break;
1346   default:
1347     return 0;
1348   }
1349 
1350   PHINode *Phi = dyn_cast<PHINode>(IncI->getOperand(0));
1351   if (Phi && Phi->getParent() == L->getHeader()) {
1352     if (isLoopInvariant(IncI->getOperand(1), L, DT))
1353       return Phi;
1354     return 0;
1355   }
1356   if (IncI->getOpcode() == Instruction::GetElementPtr)
1357     return 0;
1358 
1359   // Allow add/sub to be commuted.
1360   Phi = dyn_cast<PHINode>(IncI->getOperand(1));
1361   if (Phi && Phi->getParent() == L->getHeader()) {
1362     if (isLoopInvariant(IncI->getOperand(0), L, DT))
1363       return Phi;
1364   }
1365   return 0;
1366 }
1367 
1368 /// needsLFTR - LinearFunctionTestReplace policy. Return true unless we can show
1369 /// that the current exit test is already sufficiently canonical.
needsLFTR(Loop * L,DominatorTree * DT)1370 static bool needsLFTR(Loop *L, DominatorTree *DT) {
1371   assert(L->getExitingBlock() && "expected loop exit");
1372 
1373   BasicBlock *LatchBlock = L->getLoopLatch();
1374   // Don't bother with LFTR if the loop is not properly simplified.
1375   if (!LatchBlock)
1376     return false;
1377 
1378   BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
1379   assert(BI && "expected exit branch");
1380 
1381   // Do LFTR to simplify the exit condition to an ICMP.
1382   ICmpInst *Cond = dyn_cast<ICmpInst>(BI->getCondition());
1383   if (!Cond)
1384     return true;
1385 
1386   // Do LFTR to simplify the exit ICMP to EQ/NE
1387   ICmpInst::Predicate Pred = Cond->getPredicate();
1388   if (Pred != ICmpInst::ICMP_NE && Pred != ICmpInst::ICMP_EQ)
1389     return true;
1390 
1391   // Look for a loop invariant RHS
1392   Value *LHS = Cond->getOperand(0);
1393   Value *RHS = Cond->getOperand(1);
1394   if (!isLoopInvariant(RHS, L, DT)) {
1395     if (!isLoopInvariant(LHS, L, DT))
1396       return true;
1397     std::swap(LHS, RHS);
1398   }
1399   // Look for a simple IV counter LHS
1400   PHINode *Phi = dyn_cast<PHINode>(LHS);
1401   if (!Phi)
1402     Phi = getLoopPhiForCounter(LHS, L, DT);
1403 
1404   if (!Phi)
1405     return true;
1406 
1407   // Do LFTR if the exit condition's IV is *not* a simple counter.
1408   Value *IncV = Phi->getIncomingValueForBlock(L->getLoopLatch());
1409   return Phi != getLoopPhiForCounter(IncV, L, DT);
1410 }
1411 
1412 /// AlmostDeadIV - Return true if this IV has any uses other than the (soon to
1413 /// be rewritten) loop exit test.
AlmostDeadIV(PHINode * Phi,BasicBlock * LatchBlock,Value * Cond)1414 static bool AlmostDeadIV(PHINode *Phi, BasicBlock *LatchBlock, Value *Cond) {
1415   int LatchIdx = Phi->getBasicBlockIndex(LatchBlock);
1416   Value *IncV = Phi->getIncomingValue(LatchIdx);
1417 
1418   for (Value::use_iterator UI = Phi->use_begin(), UE = Phi->use_end();
1419        UI != UE; ++UI) {
1420     if (*UI != Cond && *UI != IncV) return false;
1421   }
1422 
1423   for (Value::use_iterator UI = IncV->use_begin(), UE = IncV->use_end();
1424        UI != UE; ++UI) {
1425     if (*UI != Cond && *UI != Phi) return false;
1426   }
1427   return true;
1428 }
1429 
1430 /// FindLoopCounter - Find an affine IV in canonical form.
1431 ///
1432 /// FIXME: Accept -1 stride and set IVLimit = IVInit - BECount
1433 ///
1434 /// FIXME: Accept non-unit stride as long as SCEV can reduce BECount * Stride.
1435 /// This is difficult in general for SCEV because of potential overflow. But we
1436 /// could at least handle constant BECounts.
1437 static PHINode *
FindLoopCounter(Loop * L,const SCEV * BECount,ScalarEvolution * SE,DominatorTree * DT,const TargetData * TD)1438 FindLoopCounter(Loop *L, const SCEV *BECount,
1439                 ScalarEvolution *SE, DominatorTree *DT, const TargetData *TD) {
1440   // I'm not sure how BECount could be a pointer type, but we definitely don't
1441   // want to LFTR that.
1442   if (BECount->getType()->isPointerTy())
1443     return 0;
1444 
1445   uint64_t BCWidth = SE->getTypeSizeInBits(BECount->getType());
1446 
1447   Value *Cond =
1448     cast<BranchInst>(L->getExitingBlock()->getTerminator())->getCondition();
1449 
1450   // Loop over all of the PHI nodes, looking for a simple counter.
1451   PHINode *BestPhi = 0;
1452   const SCEV *BestInit = 0;
1453   BasicBlock *LatchBlock = L->getLoopLatch();
1454   assert(LatchBlock && "needsLFTR should guarantee a loop latch");
1455 
1456   for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
1457     PHINode *Phi = cast<PHINode>(I);
1458     if (!SE->isSCEVable(Phi->getType()))
1459       continue;
1460 
1461     const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Phi));
1462     if (!AR || AR->getLoop() != L || !AR->isAffine())
1463       continue;
1464 
1465     // AR may be a pointer type, while BECount is an integer type.
1466     // AR may be wider than BECount. With eq/ne tests overflow is immaterial.
1467     // AR may not be a narrower type, or we may never exit.
1468     uint64_t PhiWidth = SE->getTypeSizeInBits(AR->getType());
1469     if (PhiWidth < BCWidth || (TD && !TD->isLegalInteger(PhiWidth)))
1470       continue;
1471 
1472     const SCEV *Step = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE));
1473     if (!Step || !Step->isOne())
1474       continue;
1475 
1476     int LatchIdx = Phi->getBasicBlockIndex(LatchBlock);
1477     Value *IncV = Phi->getIncomingValue(LatchIdx);
1478     if (getLoopPhiForCounter(IncV, L, DT) != Phi)
1479       continue;
1480 
1481     const SCEV *Init = AR->getStart();
1482 
1483     if (BestPhi && !AlmostDeadIV(BestPhi, LatchBlock, Cond)) {
1484       // Don't force a live loop counter if another IV can be used.
1485       if (AlmostDeadIV(Phi, LatchBlock, Cond))
1486         continue;
1487 
1488       // Prefer to count-from-zero. This is a more "canonical" counter form. It
1489       // also prefers integer to pointer IVs.
1490       if (BestInit->isZero() != Init->isZero()) {
1491         if (BestInit->isZero())
1492           continue;
1493       }
1494       // If two IVs both count from zero or both count from nonzero then the
1495       // narrower is likely a dead phi that has been widened. Use the wider phi
1496       // to allow the other to be eliminated.
1497       if (PhiWidth <= SE->getTypeSizeInBits(BestPhi->getType()))
1498         continue;
1499     }
1500     BestPhi = Phi;
1501     BestInit = Init;
1502   }
1503   return BestPhi;
1504 }
1505 
1506 /// LinearFunctionTestReplace - This method rewrites the exit condition of the
1507 /// loop to be a canonical != comparison against the incremented loop induction
1508 /// variable.  This pass is able to rewrite the exit tests of any loop where the
1509 /// SCEV analysis can determine a loop-invariant trip count of the loop, which
1510 /// is actually a much broader range than just linear tests.
1511 Value *IndVarSimplify::
LinearFunctionTestReplace(Loop * L,const SCEV * BackedgeTakenCount,PHINode * IndVar,SCEVExpander & Rewriter)1512 LinearFunctionTestReplace(Loop *L,
1513                           const SCEV *BackedgeTakenCount,
1514                           PHINode *IndVar,
1515                           SCEVExpander &Rewriter) {
1516   assert(canExpandBackedgeTakenCount(L, SE) && "precondition");
1517   BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1518 
1519   // LFTR can ignore IV overflow and truncate to the width of
1520   // BECount. This avoids materializing the add(zext(add)) expression.
1521   Type *CntTy = !EnableIVRewrite ?
1522     BackedgeTakenCount->getType() : IndVar->getType();
1523 
1524   const SCEV *IVLimit = BackedgeTakenCount;
1525 
1526   // If the exiting block is not the same as the backedge block, we must compare
1527   // against the preincremented value, otherwise we prefer to compare against
1528   // the post-incremented value.
1529   Value *CmpIndVar;
1530   if (L->getExitingBlock() == L->getLoopLatch()) {
1531     // Add one to the "backedge-taken" count to get the trip count.
1532     // If this addition may overflow, we have to be more pessimistic and
1533     // cast the induction variable before doing the add.
1534     const SCEV *N =
1535       SE->getAddExpr(IVLimit, SE->getConstant(IVLimit->getType(), 1));
1536     if (CntTy == IVLimit->getType())
1537       IVLimit = N;
1538     else {
1539       const SCEV *Zero = SE->getConstant(IVLimit->getType(), 0);
1540       if ((isa<SCEVConstant>(N) && !N->isZero()) ||
1541           SE->isLoopEntryGuardedByCond(L, ICmpInst::ICMP_NE, N, Zero)) {
1542         // No overflow. Cast the sum.
1543         IVLimit = SE->getTruncateOrZeroExtend(N, CntTy);
1544       } else {
1545         // Potential overflow. Cast before doing the add.
1546         IVLimit = SE->getTruncateOrZeroExtend(IVLimit, CntTy);
1547         IVLimit = SE->getAddExpr(IVLimit, SE->getConstant(CntTy, 1));
1548       }
1549     }
1550     // The BackedgeTaken expression contains the number of times that the
1551     // backedge branches to the loop header.  This is one less than the
1552     // number of times the loop executes, so use the incremented indvar.
1553     CmpIndVar = IndVar->getIncomingValueForBlock(L->getExitingBlock());
1554   } else {
1555     // We have to use the preincremented value...
1556     IVLimit = SE->getTruncateOrZeroExtend(IVLimit, CntTy);
1557     CmpIndVar = IndVar;
1558   }
1559 
1560   // For unit stride, IVLimit = Start + BECount with 2's complement overflow.
1561   // So for, non-zero start compute the IVLimit here.
1562   bool isPtrIV = false;
1563   Type *CmpTy = CntTy;
1564   const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(IndVar));
1565   assert(AR && AR->getLoop() == L && AR->isAffine() && "bad loop counter");
1566   if (!AR->getStart()->isZero()) {
1567     assert(AR->getStepRecurrence(*SE)->isOne() && "only handles unit stride");
1568     const SCEV *IVInit = AR->getStart();
1569 
1570     // For pointer types, sign extend BECount in order to materialize a GEP.
1571     // Note that for without EnableIVRewrite, we never run SCEVExpander on a
1572     // pointer type, because we must preserve the existing GEPs. Instead we
1573     // directly generate a GEP later.
1574     if (IVInit->getType()->isPointerTy()) {
1575       isPtrIV = true;
1576       CmpTy = SE->getEffectiveSCEVType(IVInit->getType());
1577       IVLimit = SE->getTruncateOrSignExtend(IVLimit, CmpTy);
1578     }
1579     // For integer types, truncate the IV before computing IVInit + BECount.
1580     else {
1581       if (SE->getTypeSizeInBits(IVInit->getType())
1582           > SE->getTypeSizeInBits(CmpTy))
1583         IVInit = SE->getTruncateExpr(IVInit, CmpTy);
1584 
1585       IVLimit = SE->getAddExpr(IVInit, IVLimit);
1586     }
1587   }
1588   // Expand the code for the iteration count.
1589   IRBuilder<> Builder(BI);
1590 
1591   assert(SE->isLoopInvariant(IVLimit, L) &&
1592          "Computed iteration count is not loop invariant!");
1593   Value *ExitCnt = Rewriter.expandCodeFor(IVLimit, CmpTy, BI);
1594 
1595   // Create a gep for IVInit + IVLimit from on an existing pointer base.
1596   assert(isPtrIV == IndVar->getType()->isPointerTy() &&
1597          "IndVar type must match IVInit type");
1598   if (isPtrIV) {
1599       Value *IVStart = IndVar->getIncomingValueForBlock(L->getLoopPreheader());
1600       assert(AR->getStart() == SE->getSCEV(IVStart) && "bad loop counter");
1601       assert(SE->getSizeOfExpr(
1602                cast<PointerType>(IVStart->getType())->getElementType())->isOne()
1603              && "unit stride pointer IV must be i8*");
1604 
1605       Builder.SetInsertPoint(L->getLoopPreheader()->getTerminator());
1606       ExitCnt = Builder.CreateGEP(IVStart, ExitCnt, "lftr.limit");
1607       Builder.SetInsertPoint(BI);
1608   }
1609 
1610   // Insert a new icmp_ne or icmp_eq instruction before the branch.
1611   ICmpInst::Predicate P;
1612   if (L->contains(BI->getSuccessor(0)))
1613     P = ICmpInst::ICMP_NE;
1614   else
1615     P = ICmpInst::ICMP_EQ;
1616 
1617   DEBUG(dbgs() << "INDVARS: Rewriting loop exit condition to:\n"
1618                << "      LHS:" << *CmpIndVar << '\n'
1619                << "       op:\t"
1620                << (P == ICmpInst::ICMP_NE ? "!=" : "==") << "\n"
1621                << "      RHS:\t" << *ExitCnt << "\n"
1622                << "     Expr:\t" << *IVLimit << "\n");
1623 
1624   if (SE->getTypeSizeInBits(CmpIndVar->getType())
1625       > SE->getTypeSizeInBits(CmpTy)) {
1626     CmpIndVar = Builder.CreateTrunc(CmpIndVar, CmpTy, "lftr.wideiv");
1627   }
1628 
1629   Value *Cond = Builder.CreateICmp(P, CmpIndVar, ExitCnt, "exitcond");
1630   Value *OrigCond = BI->getCondition();
1631   // It's tempting to use replaceAllUsesWith here to fully replace the old
1632   // comparison, but that's not immediately safe, since users of the old
1633   // comparison may not be dominated by the new comparison. Instead, just
1634   // update the branch to use the new comparison; in the common case this
1635   // will make old comparison dead.
1636   BI->setCondition(Cond);
1637   DeadInsts.push_back(OrigCond);
1638 
1639   ++NumLFTR;
1640   Changed = true;
1641   return Cond;
1642 }
1643 
1644 //===----------------------------------------------------------------------===//
1645 //  SinkUnusedInvariants. A late subpass to cleanup loop preheaders.
1646 //===----------------------------------------------------------------------===//
1647 
1648 /// If there's a single exit block, sink any loop-invariant values that
1649 /// were defined in the preheader but not used inside the loop into the
1650 /// exit block to reduce register pressure in the loop.
SinkUnusedInvariants(Loop * L)1651 void IndVarSimplify::SinkUnusedInvariants(Loop *L) {
1652   BasicBlock *ExitBlock = L->getExitBlock();
1653   if (!ExitBlock) return;
1654 
1655   BasicBlock *Preheader = L->getLoopPreheader();
1656   if (!Preheader) return;
1657 
1658   Instruction *InsertPt = ExitBlock->getFirstInsertionPt();
1659   BasicBlock::iterator I = Preheader->getTerminator();
1660   while (I != Preheader->begin()) {
1661     --I;
1662     // New instructions were inserted at the end of the preheader.
1663     if (isa<PHINode>(I))
1664       break;
1665 
1666     // Don't move instructions which might have side effects, since the side
1667     // effects need to complete before instructions inside the loop.  Also don't
1668     // move instructions which might read memory, since the loop may modify
1669     // memory. Note that it's okay if the instruction might have undefined
1670     // behavior: LoopSimplify guarantees that the preheader dominates the exit
1671     // block.
1672     if (I->mayHaveSideEffects() || I->mayReadFromMemory())
1673       continue;
1674 
1675     // Skip debug info intrinsics.
1676     if (isa<DbgInfoIntrinsic>(I))
1677       continue;
1678 
1679     // Skip landingpad instructions.
1680     if (isa<LandingPadInst>(I))
1681       continue;
1682 
1683     // Don't sink static AllocaInsts out of the entry block, which would
1684     // turn them into dynamic allocas!
1685     if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
1686       if (AI->isStaticAlloca())
1687         continue;
1688 
1689     // Determine if there is a use in or before the loop (direct or
1690     // otherwise).
1691     bool UsedInLoop = false;
1692     for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
1693          UI != UE; ++UI) {
1694       User *U = *UI;
1695       BasicBlock *UseBB = cast<Instruction>(U)->getParent();
1696       if (PHINode *P = dyn_cast<PHINode>(U)) {
1697         unsigned i =
1698           PHINode::getIncomingValueNumForOperand(UI.getOperandNo());
1699         UseBB = P->getIncomingBlock(i);
1700       }
1701       if (UseBB == Preheader || L->contains(UseBB)) {
1702         UsedInLoop = true;
1703         break;
1704       }
1705     }
1706 
1707     // If there is, the def must remain in the preheader.
1708     if (UsedInLoop)
1709       continue;
1710 
1711     // Otherwise, sink it to the exit block.
1712     Instruction *ToMove = I;
1713     bool Done = false;
1714 
1715     if (I != Preheader->begin()) {
1716       // Skip debug info intrinsics.
1717       do {
1718         --I;
1719       } while (isa<DbgInfoIntrinsic>(I) && I != Preheader->begin());
1720 
1721       if (isa<DbgInfoIntrinsic>(I) && I == Preheader->begin())
1722         Done = true;
1723     } else {
1724       Done = true;
1725     }
1726 
1727     ToMove->moveBefore(InsertPt);
1728     if (Done) break;
1729     InsertPt = ToMove;
1730   }
1731 }
1732 
1733 //===----------------------------------------------------------------------===//
1734 //  IndVarSimplify driver. Manage several subpasses of IV simplification.
1735 //===----------------------------------------------------------------------===//
1736 
runOnLoop(Loop * L,LPPassManager & LPM)1737 bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) {
1738   // If LoopSimplify form is not available, stay out of trouble. Some notes:
1739   //  - LSR currently only supports LoopSimplify-form loops. Indvars'
1740   //    canonicalization can be a pessimization without LSR to "clean up"
1741   //    afterwards.
1742   //  - We depend on having a preheader; in particular,
1743   //    Loop::getCanonicalInductionVariable only supports loops with preheaders,
1744   //    and we're in trouble if we can't find the induction variable even when
1745   //    we've manually inserted one.
1746   if (!L->isLoopSimplifyForm())
1747     return false;
1748 
1749   if (EnableIVRewrite)
1750     IU = &getAnalysis<IVUsers>();
1751   LI = &getAnalysis<LoopInfo>();
1752   SE = &getAnalysis<ScalarEvolution>();
1753   DT = &getAnalysis<DominatorTree>();
1754   TD = getAnalysisIfAvailable<TargetData>();
1755 
1756   DeadInsts.clear();
1757   Changed = false;
1758 
1759   // If there are any floating-point recurrences, attempt to
1760   // transform them to use integer recurrences.
1761   RewriteNonIntegerIVs(L);
1762 
1763   const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
1764 
1765   // Create a rewriter object which we'll use to transform the code with.
1766   SCEVExpander Rewriter(*SE, "indvars");
1767 #ifndef NDEBUG
1768   Rewriter.setDebugType(DEBUG_TYPE);
1769 #endif
1770 
1771   // Eliminate redundant IV users.
1772   //
1773   // Simplification works best when run before other consumers of SCEV. We
1774   // attempt to avoid evaluating SCEVs for sign/zero extend operations until
1775   // other expressions involving loop IVs have been evaluated. This helps SCEV
1776   // set no-wrap flags before normalizing sign/zero extension.
1777   if (!EnableIVRewrite) {
1778     Rewriter.disableCanonicalMode();
1779     SimplifyAndExtend(L, Rewriter, LPM);
1780   }
1781 
1782   // Check to see if this loop has a computable loop-invariant execution count.
1783   // If so, this means that we can compute the final value of any expressions
1784   // that are recurrent in the loop, and substitute the exit values from the
1785   // loop into any instructions outside of the loop that use the final values of
1786   // the current expressions.
1787   //
1788   if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1789     RewriteLoopExitValues(L, Rewriter);
1790 
1791   // Eliminate redundant IV users.
1792   if (EnableIVRewrite)
1793     Changed |= simplifyIVUsers(IU, SE, &LPM, DeadInsts);
1794 
1795   // Eliminate redundant IV cycles.
1796   if (!EnableIVRewrite)
1797     NumElimIV += Rewriter.replaceCongruentIVs(L, DT, DeadInsts);
1798 
1799   // Compute the type of the largest recurrence expression, and decide whether
1800   // a canonical induction variable should be inserted.
1801   Type *LargestType = 0;
1802   bool NeedCannIV = false;
1803   bool ExpandBECount = canExpandBackedgeTakenCount(L, SE);
1804   if (EnableIVRewrite && ExpandBECount) {
1805     // If we have a known trip count and a single exit block, we'll be
1806     // rewriting the loop exit test condition below, which requires a
1807     // canonical induction variable.
1808     NeedCannIV = true;
1809     Type *Ty = BackedgeTakenCount->getType();
1810     if (!EnableIVRewrite) {
1811       // In this mode, SimplifyIVUsers may have already widened the IV used by
1812       // the backedge test and inserted a Trunc on the compare's operand. Get
1813       // the wider type to avoid creating a redundant narrow IV only used by the
1814       // loop test.
1815       LargestType = getBackedgeIVType(L);
1816     }
1817     if (!LargestType ||
1818         SE->getTypeSizeInBits(Ty) >
1819         SE->getTypeSizeInBits(LargestType))
1820       LargestType = SE->getEffectiveSCEVType(Ty);
1821   }
1822   if (EnableIVRewrite) {
1823     for (IVUsers::const_iterator I = IU->begin(), E = IU->end(); I != E; ++I) {
1824       NeedCannIV = true;
1825       Type *Ty =
1826         SE->getEffectiveSCEVType(I->getOperandValToReplace()->getType());
1827       if (!LargestType ||
1828           SE->getTypeSizeInBits(Ty) >
1829           SE->getTypeSizeInBits(LargestType))
1830         LargestType = Ty;
1831     }
1832   }
1833 
1834   // Now that we know the largest of the induction variable expressions
1835   // in this loop, insert a canonical induction variable of the largest size.
1836   PHINode *IndVar = 0;
1837   if (NeedCannIV) {
1838     // Check to see if the loop already has any canonical-looking induction
1839     // variables. If any are present and wider than the planned canonical
1840     // induction variable, temporarily remove them, so that the Rewriter
1841     // doesn't attempt to reuse them.
1842     SmallVector<PHINode *, 2> OldCannIVs;
1843     while (PHINode *OldCannIV = L->getCanonicalInductionVariable()) {
1844       if (SE->getTypeSizeInBits(OldCannIV->getType()) >
1845           SE->getTypeSizeInBits(LargestType))
1846         OldCannIV->removeFromParent();
1847       else
1848         break;
1849       OldCannIVs.push_back(OldCannIV);
1850     }
1851 
1852     IndVar = Rewriter.getOrInsertCanonicalInductionVariable(L, LargestType);
1853 
1854     ++NumInserted;
1855     Changed = true;
1856     DEBUG(dbgs() << "INDVARS: New CanIV: " << *IndVar << '\n');
1857 
1858     // Now that the official induction variable is established, reinsert
1859     // any old canonical-looking variables after it so that the IR remains
1860     // consistent. They will be deleted as part of the dead-PHI deletion at
1861     // the end of the pass.
1862     while (!OldCannIVs.empty()) {
1863       PHINode *OldCannIV = OldCannIVs.pop_back_val();
1864       OldCannIV->insertBefore(L->getHeader()->getFirstInsertionPt());
1865     }
1866   }
1867   else if (!EnableIVRewrite && ExpandBECount && needsLFTR(L, DT)) {
1868     IndVar = FindLoopCounter(L, BackedgeTakenCount, SE, DT, TD);
1869   }
1870   // If we have a trip count expression, rewrite the loop's exit condition
1871   // using it.  We can currently only handle loops with a single exit.
1872   Value *NewICmp = 0;
1873   if (ExpandBECount && IndVar) {
1874     // Check preconditions for proper SCEVExpander operation. SCEV does not
1875     // express SCEVExpander's dependencies, such as LoopSimplify. Instead any
1876     // pass that uses the SCEVExpander must do it. This does not work well for
1877     // loop passes because SCEVExpander makes assumptions about all loops, while
1878     // LoopPassManager only forces the current loop to be simplified.
1879     //
1880     // FIXME: SCEV expansion has no way to bail out, so the caller must
1881     // explicitly check any assumptions made by SCEV. Brittle.
1882     const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(BackedgeTakenCount);
1883     if (!AR || AR->getLoop()->getLoopPreheader())
1884       NewICmp =
1885         LinearFunctionTestReplace(L, BackedgeTakenCount, IndVar, Rewriter);
1886   }
1887   // Rewrite IV-derived expressions.
1888   if (EnableIVRewrite)
1889     RewriteIVExpressions(L, Rewriter);
1890 
1891   // Clear the rewriter cache, because values that are in the rewriter's cache
1892   // can be deleted in the loop below, causing the AssertingVH in the cache to
1893   // trigger.
1894   Rewriter.clear();
1895 
1896   // Now that we're done iterating through lists, clean up any instructions
1897   // which are now dead.
1898   while (!DeadInsts.empty())
1899     if (Instruction *Inst =
1900           dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val()))
1901       RecursivelyDeleteTriviallyDeadInstructions(Inst);
1902 
1903   // The Rewriter may not be used from this point on.
1904 
1905   // Loop-invariant instructions in the preheader that aren't used in the
1906   // loop may be sunk below the loop to reduce register pressure.
1907   SinkUnusedInvariants(L);
1908 
1909   // For completeness, inform IVUsers of the IV use in the newly-created
1910   // loop exit test instruction.
1911   if (IU && NewICmp) {
1912     ICmpInst *NewICmpInst = dyn_cast<ICmpInst>(NewICmp);
1913     if (NewICmpInst)
1914       IU->AddUsersIfInteresting(cast<Instruction>(NewICmpInst->getOperand(0)));
1915   }
1916   // Clean up dead instructions.
1917   Changed |= DeleteDeadPHIs(L->getHeader());
1918   // Check a post-condition.
1919   assert(L->isLCSSAForm(*DT) &&
1920          "Indvars did not leave the loop in lcssa form!");
1921 
1922   // Verify that LFTR, and any other change have not interfered with SCEV's
1923   // ability to compute trip count.
1924 #ifndef NDEBUG
1925   if (!EnableIVRewrite && VerifyIndvars &&
1926       !isa<SCEVCouldNotCompute>(BackedgeTakenCount)) {
1927     SE->forgetLoop(L);
1928     const SCEV *NewBECount = SE->getBackedgeTakenCount(L);
1929     if (SE->getTypeSizeInBits(BackedgeTakenCount->getType()) <
1930         SE->getTypeSizeInBits(NewBECount->getType()))
1931       NewBECount = SE->getTruncateOrNoop(NewBECount,
1932                                          BackedgeTakenCount->getType());
1933     else
1934       BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount,
1935                                                  NewBECount->getType());
1936     assert(BackedgeTakenCount == NewBECount && "indvars must preserve SCEV");
1937   }
1938 #endif
1939 
1940   return Changed;
1941 }
1942