1 //===- IndVarSimplify.cpp - Induction Variable Elimination ----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
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.
14 // This transformation makes the following changes to each loop with an
15 // identifiable induction variable:
16 // 1. All loops are transformed to have a SINGLE canonical induction variable
17 // which starts at zero and steps by one.
18 // 2. The canonical induction variable is guaranteed to be the first PHI node
19 // in the loop header block.
20 // 3. Any pointer arithmetic recurrences are raised to use array subscripts.
22 // If the trip count of a loop is computable, this pass also makes the following
24 // 1. The exit condition for the loop is canonicalized to compare the
25 // induction value against the exit value. This turns loops like:
26 // 'for (i = 7; i*i < 1000; ++i)' into 'for (i = 0; i != 25; ++i)'
27 // 2. Any use outside of the loop of an expression derived from the indvar
28 // is changed to compute the derived value outside of the loop, eliminating
29 // the dependence on the exit value of the induction variable. If the only
30 // purpose of the loop is to compute the exit value of some derived
31 // expression, this transformation will make the loop dead.
33 // This transformation should be followed by strength reduction after all of the
34 // desired loop transformations have been performed. Additionally, on targets
35 // where it is profitable, the loop could be transformed to count down to zero
36 // (the "do loop" optimization).
38 //===----------------------------------------------------------------------===//
40 #define DEBUG_TYPE "indvars"
41 #include "llvm/Transforms/Scalar.h"
42 #include "llvm/BasicBlock.h"
43 #include "llvm/Constants.h"
44 #include "llvm/Instructions.h"
45 #include "llvm/Type.h"
46 #include "llvm/Analysis/ScalarEvolutionExpander.h"
47 #include "llvm/Analysis/LoopInfo.h"
48 #include "llvm/Analysis/LoopPass.h"
49 #include "llvm/Support/CFG.h"
50 #include "llvm/Support/Compiler.h"
51 #include "llvm/Support/Debug.h"
52 #include "llvm/Support/GetElementPtrTypeIterator.h"
53 #include "llvm/Transforms/Utils/Local.h"
54 #include "llvm/Support/CommandLine.h"
55 #include "llvm/ADT/SmallVector.h"
56 #include "llvm/ADT/SetVector.h"
57 #include "llvm/ADT/SmallPtrSet.h"
58 #include "llvm/ADT/Statistic.h"
61 STATISTIC(NumRemoved , "Number of aux indvars removed");
62 STATISTIC(NumInserted, "Number of canonical indvars added");
63 STATISTIC(NumReplaced, "Number of exit values replaced");
64 STATISTIC(NumLFTR , "Number of loop exit tests replaced");
67 class VISIBILITY_HIDDEN IndVarSimplify : public LoopPass {
73 static char ID; // Pass identification, replacement for typeid
74 IndVarSimplify() : LoopPass(&ID) {}
76 virtual bool runOnLoop(Loop *L, LPPassManager &LPM);
78 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
79 AU.addRequired<ScalarEvolution>();
80 AU.addRequiredID(LCSSAID);
81 AU.addRequiredID(LoopSimplifyID);
82 AU.addRequired<LoopInfo>();
83 AU.addPreserved<ScalarEvolution>();
84 AU.addPreservedID(LoopSimplifyID);
85 AU.addPreservedID(LCSSAID);
91 void RewriteNonIntegerIVs(Loop *L);
93 void LinearFunctionTestReplace(Loop *L, SCEVHandle BackedgeTakenCount,
95 BasicBlock *ExitingBlock,
97 SCEVExpander &Rewriter);
98 void RewriteLoopExitValues(Loop *L, const SCEV *BackedgeTakenCount);
100 void DeleteTriviallyDeadInstructions(SmallPtrSet<Instruction*, 16> &Insts);
102 void HandleFloatingPointIV(Loop *L, PHINode *PH,
103 SmallPtrSet<Instruction*, 16> &DeadInsts);
107 char IndVarSimplify::ID = 0;
108 static RegisterPass<IndVarSimplify>
109 X("indvars", "Canonicalize Induction Variables");
111 Pass *llvm::createIndVarSimplifyPass() {
112 return new IndVarSimplify();
115 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
116 /// specified set are trivially dead, delete them and see if this makes any of
117 /// their operands subsequently dead.
118 void IndVarSimplify::
119 DeleteTriviallyDeadInstructions(SmallPtrSet<Instruction*, 16> &Insts) {
120 while (!Insts.empty()) {
121 Instruction *I = *Insts.begin();
123 if (isInstructionTriviallyDead(I)) {
124 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
125 if (Instruction *U = dyn_cast<Instruction>(I->getOperand(i)))
127 SE->deleteValueFromRecords(I);
128 DOUT << "INDVARS: Deleting: " << *I;
129 I->eraseFromParent();
135 /// LinearFunctionTestReplace - This method rewrites the exit condition of the
136 /// loop to be a canonical != comparison against the incremented loop induction
137 /// variable. This pass is able to rewrite the exit tests of any loop where the
138 /// SCEV analysis can determine a loop-invariant trip count of the loop, which
139 /// is actually a much broader range than just linear tests.
140 void IndVarSimplify::LinearFunctionTestReplace(Loop *L,
141 SCEVHandle BackedgeTakenCount,
143 BasicBlock *ExitingBlock,
145 SCEVExpander &Rewriter) {
146 // If the exiting block is not the same as the backedge block, we must compare
147 // against the preincremented value, otherwise we prefer to compare against
148 // the post-incremented value.
150 SCEVHandle RHS = BackedgeTakenCount;
151 if (ExitingBlock == L->getLoopLatch()) {
152 // Add one to the "backedge-taken" count to get the trip count.
153 // If this addition may overflow, we have to be more pessimistic and
154 // cast the induction variable before doing the add.
155 SCEVHandle Zero = SE->getIntegerSCEV(0, BackedgeTakenCount->getType());
157 SE->getAddExpr(BackedgeTakenCount,
158 SE->getIntegerSCEV(1, BackedgeTakenCount->getType()));
159 if ((isa<SCEVConstant>(N) && !N->isZero()) ||
160 SE->isLoopGuardedByCond(L, ICmpInst::ICMP_NE, N, Zero)) {
161 // No overflow. Cast the sum.
162 RHS = SE->getTruncateOrZeroExtend(N, IndVar->getType());
164 // Potential overflow. Cast before doing the add.
165 RHS = SE->getTruncateOrZeroExtend(BackedgeTakenCount,
167 RHS = SE->getAddExpr(RHS,
168 SE->getIntegerSCEV(1, IndVar->getType()));
171 // The BackedgeTaken expression contains the number of times that the
172 // backedge branches to the loop header. This is one less than the
173 // number of times the loop executes, so use the incremented indvar.
174 CmpIndVar = L->getCanonicalInductionVariableIncrement();
176 // We have to use the preincremented value...
177 RHS = SE->getTruncateOrZeroExtend(BackedgeTakenCount,
182 // Expand the code for the iteration count into the preheader of the loop.
183 BasicBlock *Preheader = L->getLoopPreheader();
184 Value *ExitCnt = Rewriter.expandCodeFor(RHS, IndVar->getType(),
185 Preheader->getTerminator());
187 // Insert a new icmp_ne or icmp_eq instruction before the branch.
188 ICmpInst::Predicate Opcode;
189 if (L->contains(BI->getSuccessor(0)))
190 Opcode = ICmpInst::ICMP_NE;
192 Opcode = ICmpInst::ICMP_EQ;
194 DOUT << "INDVARS: Rewriting loop exit condition to:\n"
195 << " LHS:" << *CmpIndVar // includes a newline
197 << (Opcode == ICmpInst::ICMP_NE ? "!=" : "==") << "\n"
198 << " RHS:\t" << *RHS << "\n";
200 Value *Cond = new ICmpInst(Opcode, CmpIndVar, ExitCnt, "exitcond", BI);
201 BI->setCondition(Cond);
206 /// RewriteLoopExitValues - Check to see if this loop has a computable
207 /// loop-invariant execution count. If so, this means that we can compute the
208 /// final value of any expressions that are recurrent in the loop, and
209 /// substitute the exit values from the loop into any instructions outside of
210 /// the loop that use the final values of the current expressions.
211 void IndVarSimplify::RewriteLoopExitValues(Loop *L,
212 const SCEV *BackedgeTakenCount) {
213 BasicBlock *Preheader = L->getLoopPreheader();
215 // Scan all of the instructions in the loop, looking at those that have
216 // extra-loop users and which are recurrences.
217 SCEVExpander Rewriter(*SE, *LI);
219 // We insert the code into the preheader of the loop if the loop contains
220 // multiple exit blocks, or in the exit block if there is exactly one.
221 BasicBlock *BlockToInsertInto;
222 SmallVector<BasicBlock*, 8> ExitBlocks;
223 L->getUniqueExitBlocks(ExitBlocks);
224 if (ExitBlocks.size() == 1)
225 BlockToInsertInto = ExitBlocks[0];
227 BlockToInsertInto = Preheader;
228 BasicBlock::iterator InsertPt = BlockToInsertInto->getFirstNonPHI();
230 bool HasConstantItCount = isa<SCEVConstant>(BackedgeTakenCount);
232 SmallPtrSet<Instruction*, 16> InstructionsToDelete;
233 std::map<Instruction*, Value*> ExitValues;
235 // Find all values that are computed inside the loop, but used outside of it.
236 // Because of LCSSA, these values will only occur in LCSSA PHI Nodes. Scan
237 // the exit blocks of the loop to find them.
238 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
239 BasicBlock *ExitBB = ExitBlocks[i];
241 // If there are no PHI nodes in this exit block, then no values defined
242 // inside the loop are used on this path, skip it.
243 PHINode *PN = dyn_cast<PHINode>(ExitBB->begin());
246 unsigned NumPreds = PN->getNumIncomingValues();
248 // Iterate over all of the PHI nodes.
249 BasicBlock::iterator BBI = ExitBB->begin();
250 while ((PN = dyn_cast<PHINode>(BBI++))) {
252 // Iterate over all of the values in all the PHI nodes.
253 for (unsigned i = 0; i != NumPreds; ++i) {
254 // If the value being merged in is not integer or is not defined
255 // in the loop, skip it.
256 Value *InVal = PN->getIncomingValue(i);
257 if (!isa<Instruction>(InVal) ||
258 // SCEV only supports integer expressions for now.
259 (!isa<IntegerType>(InVal->getType()) &&
260 !isa<PointerType>(InVal->getType())))
263 // If this pred is for a subloop, not L itself, skip it.
264 if (LI->getLoopFor(PN->getIncomingBlock(i)) != L)
265 continue; // The Block is in a subloop, skip it.
267 // Check that InVal is defined in the loop.
268 Instruction *Inst = cast<Instruction>(InVal);
269 if (!L->contains(Inst->getParent()))
272 // We require that this value either have a computable evolution or that
273 // the loop have a constant iteration count. In the case where the loop
274 // has a constant iteration count, we can sometimes force evaluation of
275 // the exit value through brute force.
276 SCEVHandle SH = SE->getSCEV(Inst);
277 if (!SH->hasComputableLoopEvolution(L) && !HasConstantItCount)
278 continue; // Cannot get exit evolution for the loop value.
280 // Okay, this instruction has a user outside of the current loop
281 // and varies predictably *inside* the loop. Evaluate the value it
282 // contains when the loop exits, if possible.
283 SCEVHandle ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop());
284 if (isa<SCEVCouldNotCompute>(ExitValue) ||
285 !ExitValue->isLoopInvariant(L))
291 // See if we already computed the exit value for the instruction, if so,
293 Value *&ExitVal = ExitValues[Inst];
295 ExitVal = Rewriter.expandCodeFor(ExitValue, PN->getType(), InsertPt);
297 DOUT << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal
298 << " LoopVal = " << *Inst << "\n";
300 PN->setIncomingValue(i, ExitVal);
302 // If this instruction is dead now, schedule it to be removed.
303 if (Inst->use_empty())
304 InstructionsToDelete.insert(Inst);
306 // See if this is a single-entry LCSSA PHI node. If so, we can (and
308 // the PHI entirely. This is safe, because the NewVal won't be variant
309 // in the loop, so we don't need an LCSSA phi node anymore.
311 SE->deleteValueFromRecords(PN);
312 PN->replaceAllUsesWith(ExitVal);
313 PN->eraseFromParent();
320 DeleteTriviallyDeadInstructions(InstructionsToDelete);
323 void IndVarSimplify::RewriteNonIntegerIVs(Loop *L) {
324 // First step. Check to see if there are any floating-point recurrences.
325 // If there are, change them into integer recurrences, permitting analysis by
326 // the SCEV routines.
328 BasicBlock *Header = L->getHeader();
330 SmallPtrSet<Instruction*, 16> DeadInsts;
331 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
332 PHINode *PN = cast<PHINode>(I);
333 HandleFloatingPointIV(L, PN, DeadInsts);
336 // If the loop previously had floating-point IV, ScalarEvolution
337 // may not have been able to compute a trip count. Now that we've done some
338 // re-writing, the trip count may be computable.
340 SE->forgetLoopBackedgeTakenCount(L);
342 if (!DeadInsts.empty())
343 DeleteTriviallyDeadInstructions(DeadInsts);
346 /// getEffectiveIndvarType - Determine the widest type that the
347 /// induction-variable PHINode Phi is cast to.
349 static const Type *getEffectiveIndvarType(const PHINode *Phi,
350 const ScalarEvolution *SE) {
351 const Type *Ty = Phi->getType();
353 for (Value::use_const_iterator UI = Phi->use_begin(), UE = Phi->use_end();
355 const Type *CandidateType = NULL;
356 if (const ZExtInst *ZI = dyn_cast<ZExtInst>(UI))
357 CandidateType = ZI->getDestTy();
358 else if (const SExtInst *SI = dyn_cast<SExtInst>(UI))
359 CandidateType = SI->getDestTy();
360 else if (const IntToPtrInst *IP = dyn_cast<IntToPtrInst>(UI))
361 CandidateType = IP->getDestTy();
362 else if (const PtrToIntInst *PI = dyn_cast<PtrToIntInst>(UI))
363 CandidateType = PI->getDestTy();
365 SE->isSCEVable(CandidateType) &&
366 SE->getTypeSizeInBits(CandidateType) > SE->getTypeSizeInBits(Ty))
373 /// TestOrigIVForWrap - Analyze the original induction variable
374 /// that controls the loop's iteration to determine whether it
375 /// would ever undergo signed or unsigned overflow. Also, check
376 /// whether an induction variable in the same type that starts
377 /// at 0 would undergo signed overflow.
379 /// In addition to setting the NoSignedWrap and NoUnsignedWrap
380 /// variables to true when appropriate (they are not set to false here),
381 /// return the PHI for this induction variable. Also record the initial
382 /// and final values and the increment; these are not meaningful unless
383 /// either NoSignedWrap or NoUnsignedWrap is true, and are always meaningful
384 /// in that case, although the final value may be 0 indicating a nonconstant.
386 /// TODO: This duplicates a fair amount of ScalarEvolution logic.
387 /// Perhaps this can be merged with
388 /// ScalarEvolution::getBackedgeTakenCount
389 /// and/or ScalarEvolution::get{Sign,Zero}ExtendExpr.
391 static const PHINode *TestOrigIVForWrap(const Loop *L,
392 const BranchInst *BI,
393 const Instruction *OrigCond,
394 const ScalarEvolution &SE,
396 bool &NoUnsignedWrap,
397 const ConstantInt* &InitialVal,
398 const ConstantInt* &IncrVal,
399 const ConstantInt* &LimitVal) {
400 // Verify that the loop is sane and find the exit condition.
401 const ICmpInst *Cmp = dyn_cast<ICmpInst>(OrigCond);
404 const Value *CmpLHS = Cmp->getOperand(0);
405 const Value *CmpRHS = Cmp->getOperand(1);
406 const BasicBlock *TrueBB = BI->getSuccessor(0);
407 const BasicBlock *FalseBB = BI->getSuccessor(1);
408 ICmpInst::Predicate Pred = Cmp->getPredicate();
410 // Canonicalize a constant to the RHS.
411 if (isa<ConstantInt>(CmpLHS)) {
412 Pred = ICmpInst::getSwappedPredicate(Pred);
413 std::swap(CmpLHS, CmpRHS);
415 // Canonicalize SLE to SLT.
416 if (Pred == ICmpInst::ICMP_SLE)
417 if (const ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS))
418 if (!CI->getValue().isMaxSignedValue()) {
419 CmpRHS = ConstantInt::get(CI->getValue() + 1);
420 Pred = ICmpInst::ICMP_SLT;
422 // Canonicalize SGT to SGE.
423 if (Pred == ICmpInst::ICMP_SGT)
424 if (const ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS))
425 if (!CI->getValue().isMaxSignedValue()) {
426 CmpRHS = ConstantInt::get(CI->getValue() + 1);
427 Pred = ICmpInst::ICMP_SGE;
429 // Canonicalize SGE to SLT.
430 if (Pred == ICmpInst::ICMP_SGE) {
431 std::swap(TrueBB, FalseBB);
432 Pred = ICmpInst::ICMP_SLT;
434 // Canonicalize ULE to ULT.
435 if (Pred == ICmpInst::ICMP_ULE)
436 if (const ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS))
437 if (!CI->getValue().isMaxValue()) {
438 CmpRHS = ConstantInt::get(CI->getValue() + 1);
439 Pred = ICmpInst::ICMP_ULT;
441 // Canonicalize UGT to UGE.
442 if (Pred == ICmpInst::ICMP_UGT)
443 if (const ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS))
444 if (!CI->getValue().isMaxValue()) {
445 CmpRHS = ConstantInt::get(CI->getValue() + 1);
446 Pred = ICmpInst::ICMP_UGE;
448 // Canonicalize UGE to ULT.
449 if (Pred == ICmpInst::ICMP_UGE) {
450 std::swap(TrueBB, FalseBB);
451 Pred = ICmpInst::ICMP_ULT;
453 // For now, analyze only LT loops for signed overflow.
454 if (Pred != ICmpInst::ICMP_SLT && Pred != ICmpInst::ICMP_ULT)
457 bool isSigned = Pred == ICmpInst::ICMP_SLT;
459 // Get the increment instruction. Look past casts if we will
460 // be able to prove that the original induction variable doesn't
461 // undergo signed or unsigned overflow, respectively.
462 const Value *IncrInst = CmpLHS;
464 if (const SExtInst *SI = dyn_cast<SExtInst>(CmpLHS)) {
465 if (!isa<ConstantInt>(CmpRHS) ||
466 !cast<ConstantInt>(CmpRHS)->getValue()
467 .isSignedIntN(SE.getTypeSizeInBits(IncrInst->getType())))
469 IncrInst = SI->getOperand(0);
472 if (const ZExtInst *ZI = dyn_cast<ZExtInst>(CmpLHS)) {
473 if (!isa<ConstantInt>(CmpRHS) ||
474 !cast<ConstantInt>(CmpRHS)->getValue()
475 .isIntN(SE.getTypeSizeInBits(IncrInst->getType())))
477 IncrInst = ZI->getOperand(0);
481 // For now, only analyze induction variables that have simple increments.
482 const BinaryOperator *IncrOp = dyn_cast<BinaryOperator>(IncrInst);
483 if (!IncrOp || IncrOp->getOpcode() != Instruction::Add)
485 IncrVal = dyn_cast<ConstantInt>(IncrOp->getOperand(1));
489 // Make sure the PHI looks like a normal IV.
490 const PHINode *PN = dyn_cast<PHINode>(IncrOp->getOperand(0));
491 if (!PN || PN->getNumIncomingValues() != 2)
493 unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
494 unsigned BackEdge = !IncomingEdge;
495 if (!L->contains(PN->getIncomingBlock(BackEdge)) ||
496 PN->getIncomingValue(BackEdge) != IncrOp)
498 if (!L->contains(TrueBB))
501 // For now, only analyze loops with a constant start value, so that
502 // we can easily determine if the start value is not a maximum value
503 // which would wrap on the first iteration.
504 InitialVal = dyn_cast<ConstantInt>(PN->getIncomingValue(IncomingEdge));
508 // The upper limit need not be a constant; we'll check later.
509 LimitVal = dyn_cast<ConstantInt>(CmpRHS);
511 // We detect the impossibility of wrapping in two cases, both of
512 // which require starting with a non-max value:
513 // - The IV counts up by one, and the loop iterates only while it remains
514 // less than a limiting value (any) in the same type.
515 // - The IV counts up by a positive increment other than 1, and the
516 // constant limiting value + the increment is less than the max value
517 // (computed as max-increment to avoid overflow)
518 if (isSigned && !InitialVal->getValue().isMaxSignedValue()) {
519 if (IncrVal->equalsInt(1))
520 NoSignedWrap = true; // LimitVal need not be constant
522 uint64_t numBits = LimitVal->getValue().getBitWidth();
523 if (IncrVal->getValue().sgt(APInt::getNullValue(numBits)) &&
524 (APInt::getSignedMaxValue(numBits) - IncrVal->getValue())
525 .sgt(LimitVal->getValue()))
528 } else if (!isSigned && !InitialVal->getValue().isMaxValue()) {
529 if (IncrVal->equalsInt(1))
530 NoUnsignedWrap = true; // LimitVal need not be constant
532 uint64_t numBits = LimitVal->getValue().getBitWidth();
533 if (IncrVal->getValue().ugt(APInt::getNullValue(numBits)) &&
534 (APInt::getMaxValue(numBits) - IncrVal->getValue())
535 .ugt(LimitVal->getValue()))
536 NoUnsignedWrap = true;
542 static Value *getSignExtendedTruncVar(const SCEVAddRecExpr *AR,
544 const Type *LargestType, Loop *L,
546 SCEVExpander &Rewriter,
547 BasicBlock::iterator InsertPt) {
548 SCEVHandle ExtendedStart =
549 SE->getSignExtendExpr(AR->getStart(), LargestType);
550 SCEVHandle ExtendedStep =
551 SE->getSignExtendExpr(AR->getStepRecurrence(*SE), LargestType);
552 SCEVHandle ExtendedAddRec =
553 SE->getAddRecExpr(ExtendedStart, ExtendedStep, L);
554 if (LargestType != myType)
555 ExtendedAddRec = SE->getTruncateExpr(ExtendedAddRec, myType);
556 return Rewriter.expandCodeFor(ExtendedAddRec, myType, InsertPt);
559 static Value *getZeroExtendedTruncVar(const SCEVAddRecExpr *AR,
561 const Type *LargestType, Loop *L,
563 SCEVExpander &Rewriter,
564 BasicBlock::iterator InsertPt) {
565 SCEVHandle ExtendedStart =
566 SE->getZeroExtendExpr(AR->getStart(), LargestType);
567 SCEVHandle ExtendedStep =
568 SE->getZeroExtendExpr(AR->getStepRecurrence(*SE), LargestType);
569 SCEVHandle ExtendedAddRec =
570 SE->getAddRecExpr(ExtendedStart, ExtendedStep, L);
571 if (LargestType != myType)
572 ExtendedAddRec = SE->getTruncateExpr(ExtendedAddRec, myType);
573 return Rewriter.expandCodeFor(ExtendedAddRec, myType, InsertPt);
576 /// allUsesAreSameTyped - See whether all Uses of I are instructions
577 /// with the same Opcode and the same type.
578 static bool allUsesAreSameTyped(unsigned int Opcode, Instruction *I) {
579 const Type* firstType = NULL;
580 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
582 Instruction *II = dyn_cast<Instruction>(*UI);
583 if (!II || II->getOpcode() != Opcode)
586 firstType = II->getType();
587 else if (firstType != II->getType())
593 bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) {
594 LI = &getAnalysis<LoopInfo>();
595 SE = &getAnalysis<ScalarEvolution>();
598 // If there are any floating-point recurrences, attempt to
599 // transform them to use integer recurrences.
600 RewriteNonIntegerIVs(L);
602 BasicBlock *Header = L->getHeader();
603 BasicBlock *ExitingBlock = L->getExitingBlock();
604 SmallPtrSet<Instruction*, 16> DeadInsts;
606 // Verify the input to the pass in already in LCSSA form.
607 assert(L->isLCSSAForm());
609 // Check to see if this loop has a computable loop-invariant execution count.
610 // If so, this means that we can compute the final value of any expressions
611 // that are recurrent in the loop, and substitute the exit values from the
612 // loop into any instructions outside of the loop that use the final values of
613 // the current expressions.
615 SCEVHandle BackedgeTakenCount = SE->getBackedgeTakenCount(L);
616 if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount))
617 RewriteLoopExitValues(L, BackedgeTakenCount);
619 // Next, analyze all of the induction variables in the loop, canonicalizing
620 // auxillary induction variables.
621 std::vector<std::pair<PHINode*, SCEVHandle> > IndVars;
623 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
624 PHINode *PN = cast<PHINode>(I);
625 if (SE->isSCEVable(PN->getType())) {
626 SCEVHandle SCEV = SE->getSCEV(PN);
627 // FIXME: It is an extremely bad idea to indvar substitute anything more
628 // complex than affine induction variables. Doing so will put expensive
629 // polynomial evaluations inside of the loop, and the str reduction pass
630 // currently can only reduce affine polynomials. For now just disable
631 // indvar subst on anything more complex than an affine addrec.
632 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SCEV))
633 if (AR->getLoop() == L && AR->isAffine())
634 IndVars.push_back(std::make_pair(PN, SCEV));
638 // Compute the type of the largest recurrence expression, and collect
639 // the set of the types of the other recurrence expressions.
640 const Type *LargestType = 0;
641 SmallSetVector<const Type *, 4> SizesToInsert;
642 if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount)) {
643 LargestType = BackedgeTakenCount->getType();
644 LargestType = SE->getEffectiveSCEVType(LargestType);
645 SizesToInsert.insert(LargestType);
647 for (unsigned i = 0, e = IndVars.size(); i != e; ++i) {
648 const PHINode *PN = IndVars[i].first;
649 const Type *PNTy = PN->getType();
650 PNTy = SE->getEffectiveSCEVType(PNTy);
651 SizesToInsert.insert(PNTy);
652 const Type *EffTy = getEffectiveIndvarType(PN, SE);
653 EffTy = SE->getEffectiveSCEVType(EffTy);
654 SizesToInsert.insert(EffTy);
656 SE->getTypeSizeInBits(EffTy) >
657 SE->getTypeSizeInBits(LargestType))
661 // Create a rewriter object which we'll use to transform the code with.
662 SCEVExpander Rewriter(*SE, *LI);
664 // Now that we know the largest of of the induction variables in this loop,
665 // insert a canonical induction variable of the largest size.
667 if (!SizesToInsert.empty()) {
668 IndVar = Rewriter.getOrInsertCanonicalInductionVariable(L,LargestType);
671 DOUT << "INDVARS: New CanIV: " << *IndVar;
674 // If we have a trip count expression, rewrite the loop's exit condition
675 // using it. We can currently only handle loops with a single exit.
676 bool NoSignedWrap = false;
677 bool NoUnsignedWrap = false;
678 const ConstantInt* InitialVal, * IncrVal, * LimitVal;
679 const PHINode *OrigControllingPHI = 0;
680 if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount) && ExitingBlock)
681 // Can't rewrite non-branch yet.
682 if (BranchInst *BI = dyn_cast<BranchInst>(ExitingBlock->getTerminator())) {
683 if (Instruction *OrigCond = dyn_cast<Instruction>(BI->getCondition())) {
684 // Determine if the OrigIV will ever undergo overflow.
686 TestOrigIVForWrap(L, BI, OrigCond, *SE,
687 NoSignedWrap, NoUnsignedWrap,
688 InitialVal, IncrVal, LimitVal);
690 // We'll be replacing the original condition, so it'll be dead.
691 DeadInsts.insert(OrigCond);
694 LinearFunctionTestReplace(L, BackedgeTakenCount, IndVar,
695 ExitingBlock, BI, Rewriter);
698 // Now that we have a canonical induction variable, we can rewrite any
699 // recurrences in terms of the induction variable. Start with the auxillary
700 // induction variables, and recursively rewrite any of their uses.
701 BasicBlock::iterator InsertPt = Header->getFirstNonPHI();
703 // If there were induction variables of other sizes, cast the primary
704 // induction variable to the right size for them, avoiding the need for the
705 // code evaluation methods to insert induction variables of different sizes.
706 for (unsigned i = 0, e = SizesToInsert.size(); i != e; ++i) {
707 const Type *Ty = SizesToInsert[i];
708 if (Ty != LargestType) {
709 Instruction *New = new TruncInst(IndVar, Ty, "indvar", InsertPt);
710 Rewriter.addInsertedValue(New, SE->getSCEV(New));
711 DOUT << "INDVARS: Made trunc IV for type " << *Ty << ": "
716 // Rewrite all induction variables in terms of the canonical induction
718 while (!IndVars.empty()) {
719 PHINode *PN = IndVars.back().first;
720 const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(IndVars.back().second);
721 Value *NewVal = Rewriter.expandCodeFor(AR, PN->getType(), InsertPt);
722 DOUT << "INDVARS: Rewrote IV '" << *AR << "' " << *PN
723 << " into = " << *NewVal << "\n";
724 NewVal->takeName(PN);
726 /// If the new canonical induction variable is wider than the original,
727 /// and the original has uses that are casts to wider types, see if the
728 /// truncate and extend can be omitted.
729 if (PN == OrigControllingPHI && PN->getType() != LargestType)
730 for (Value::use_iterator UI = PN->use_begin(), UE = PN->use_end();
732 Instruction *UInst = dyn_cast<Instruction>(*UI);
733 if (UInst && isa<SExtInst>(UInst) && NoSignedWrap) {
734 Value *TruncIndVar = getSignExtendedTruncVar(AR, SE, LargestType, L,
735 UInst->getType(), Rewriter, InsertPt);
736 UInst->replaceAllUsesWith(TruncIndVar);
737 DeadInsts.insert(UInst);
739 // See if we can figure out sext(i+constant) doesn't wrap, so we can
740 // use a larger add. This is common in subscripting.
741 if (UInst && UInst->getOpcode()==Instruction::Add &&
742 !UInst->use_empty() &&
743 allUsesAreSameTyped(Instruction::SExt, UInst) &&
744 isa<ConstantInt>(UInst->getOperand(1)) &&
745 NoSignedWrap && LimitVal) {
746 uint64_t oldBitSize = LimitVal->getValue().getBitWidth();
747 uint64_t newBitSize = LargestType->getPrimitiveSizeInBits();
748 ConstantInt* AddRHS = dyn_cast<ConstantInt>(UInst->getOperand(1));
749 if (((APInt::getSignedMaxValue(oldBitSize) - IncrVal->getValue()) -
750 AddRHS->getValue()).sgt(LimitVal->getValue())) {
751 // We've determined this is (i+constant) and it won't overflow.
752 if (isa<SExtInst>(UInst->use_begin())) {
753 SExtInst* oldSext = dyn_cast<SExtInst>(UInst->use_begin());
754 uint64_t truncSize = oldSext->getType()->getPrimitiveSizeInBits();
755 Value *TruncIndVar = getSignExtendedTruncVar(AR, SE, LargestType,
756 L, oldSext->getType(), Rewriter,
758 APInt APnewAddRHS = APInt(AddRHS->getValue()).sext(newBitSize);
759 if (newBitSize > truncSize)
760 APnewAddRHS = APnewAddRHS.trunc(truncSize);
761 ConstantInt* newAddRHS =ConstantInt::get(APnewAddRHS);
763 BinaryOperator::CreateAdd(TruncIndVar, newAddRHS,
764 UInst->getName()+".nosex", UInst);
765 for (Value::use_iterator UI2 = UInst->use_begin(),
766 UE2 = UInst->use_end(); UI2 != UE2; ++UI2) {
767 Instruction *II = dyn_cast<Instruction>(UI2);
768 II->replaceAllUsesWith(NewAdd);
769 DeadInsts.insert(II);
771 DeadInsts.insert(UInst);
775 // Try for sext(i | constant). This is safe as long as the
776 // high bit of the constant is not set.
777 if (UInst && UInst->getOpcode()==Instruction::Or &&
778 !UInst->use_empty() &&
779 allUsesAreSameTyped(Instruction::SExt, UInst) && NoSignedWrap &&
780 isa<ConstantInt>(UInst->getOperand(1))) {
781 ConstantInt* RHS = dyn_cast<ConstantInt>(UInst->getOperand(1));
782 if (!RHS->getValue().isNegative()) {
783 uint64_t newBitSize = LargestType->getPrimitiveSizeInBits();
784 SExtInst* oldSext = dyn_cast<SExtInst>(UInst->use_begin());
785 uint64_t truncSize = oldSext->getType()->getPrimitiveSizeInBits();
786 Value *TruncIndVar = getSignExtendedTruncVar(AR, SE, LargestType,
787 L, oldSext->getType(), Rewriter,
789 APInt APnewOrRHS = APInt(RHS->getValue()).sext(newBitSize);
790 if (newBitSize > truncSize)
791 APnewOrRHS = APnewOrRHS.trunc(truncSize);
792 ConstantInt* newOrRHS =ConstantInt::get(APnewOrRHS);
794 BinaryOperator::CreateOr(TruncIndVar, newOrRHS,
795 UInst->getName()+".nosex", UInst);
796 for (Value::use_iterator UI2 = UInst->use_begin(),
797 UE2 = UInst->use_end(); UI2 != UE2; ++UI2) {
798 Instruction *II = dyn_cast<Instruction>(UI2);
799 II->replaceAllUsesWith(NewOr);
800 DeadInsts.insert(II);
802 DeadInsts.insert(UInst);
805 // A zext of a signed variable known not to overflow is still safe.
806 if (UInst && isa<ZExtInst>(UInst) && (NoUnsignedWrap || NoSignedWrap)) {
807 Value *TruncIndVar = getZeroExtendedTruncVar(AR, SE, LargestType, L,
808 UInst->getType(), Rewriter, InsertPt);
809 UInst->replaceAllUsesWith(TruncIndVar);
810 DeadInsts.insert(UInst);
812 // If we have zext(i&constant), it's always safe to use the larger
813 // variable. This is not common but is a bottleneck in Openssl.
814 // (RHS doesn't have to be constant. There should be a better approach
815 // than bottom-up pattern matching for this...)
816 if (UInst && UInst->getOpcode()==Instruction::And &&
817 !UInst->use_empty() &&
818 allUsesAreSameTyped(Instruction::ZExt, UInst) &&
819 isa<ConstantInt>(UInst->getOperand(1))) {
820 uint64_t newBitSize = LargestType->getPrimitiveSizeInBits();
821 ConstantInt* AndRHS = dyn_cast<ConstantInt>(UInst->getOperand(1));
822 ZExtInst* oldZext = dyn_cast<ZExtInst>(UInst->use_begin());
823 uint64_t truncSize = oldZext->getType()->getPrimitiveSizeInBits();
824 Value *TruncIndVar = getSignExtendedTruncVar(AR, SE, LargestType,
825 L, oldZext->getType(), Rewriter, InsertPt);
826 APInt APnewAndRHS = APInt(AndRHS->getValue()).zext(newBitSize);
827 if (newBitSize > truncSize)
828 APnewAndRHS = APnewAndRHS.trunc(truncSize);
829 ConstantInt* newAndRHS = ConstantInt::get(APnewAndRHS);
831 BinaryOperator::CreateAnd(TruncIndVar, newAndRHS,
832 UInst->getName()+".nozex", UInst);
833 for (Value::use_iterator UI2 = UInst->use_begin(),
834 UE2 = UInst->use_end(); UI2 != UE2; ++UI2) {
835 Instruction *II = dyn_cast<Instruction>(UI2);
836 II->replaceAllUsesWith(NewAnd);
837 DeadInsts.insert(II);
839 DeadInsts.insert(UInst);
841 // If we have zext((i+constant)&constant), we can use the larger
842 // variable even if the add does overflow. This works whenever the
843 // constant being ANDed is the same size as i, which it presumably is.
844 // We don't need to restrict the expression being and'ed to i+const,
845 // but we have to promote everything in it, so it's convenient.
846 // zext((i | constant)&constant) is also valid and accepted here.
847 if (UInst && (UInst->getOpcode()==Instruction::Add ||
848 UInst->getOpcode()==Instruction::Or) &&
849 UInst->hasOneUse() &&
850 isa<ConstantInt>(UInst->getOperand(1))) {
851 uint64_t newBitSize = LargestType->getPrimitiveSizeInBits();
852 ConstantInt* AddRHS = dyn_cast<ConstantInt>(UInst->getOperand(1));
853 Instruction *UInst2 = dyn_cast<Instruction>(UInst->use_begin());
854 if (UInst2 && UInst2->getOpcode() == Instruction::And &&
855 !UInst2->use_empty() &&
856 allUsesAreSameTyped(Instruction::ZExt, UInst2) &&
857 isa<ConstantInt>(UInst2->getOperand(1))) {
858 ZExtInst* oldZext = dyn_cast<ZExtInst>(UInst2->use_begin());
859 uint64_t truncSize = oldZext->getType()->getPrimitiveSizeInBits();
860 Value *TruncIndVar = getSignExtendedTruncVar(AR, SE, LargestType,
861 L, oldZext->getType(), Rewriter, InsertPt);
862 ConstantInt* AndRHS = dyn_cast<ConstantInt>(UInst2->getOperand(1));
863 APInt APnewAddRHS = APInt(AddRHS->getValue()).zext(newBitSize);
864 if (newBitSize > truncSize)
865 APnewAddRHS = APnewAddRHS.trunc(truncSize);
866 ConstantInt* newAddRHS = ConstantInt::get(APnewAddRHS);
867 Value *NewAdd = ((UInst->getOpcode()==Instruction::Add) ?
868 BinaryOperator::CreateAdd(TruncIndVar, newAddRHS,
869 UInst->getName()+".nozex", UInst2) :
870 BinaryOperator::CreateOr(TruncIndVar, newAddRHS,
871 UInst->getName()+".nozex", UInst2));
872 APInt APcopy2 = APInt(AndRHS->getValue());
873 ConstantInt* newAndRHS = ConstantInt::get(APcopy2.zext(newBitSize));
875 BinaryOperator::CreateAnd(NewAdd, newAndRHS,
876 UInst->getName()+".nozex", UInst2);
877 for (Value::use_iterator UI2 = UInst2->use_begin(),
878 UE2 = UInst2->use_end(); UI2 != UE2; ++UI2) {
879 Instruction *II = dyn_cast<Instruction>(UI2);
880 II->replaceAllUsesWith(NewAnd);
881 DeadInsts.insert(II);
883 DeadInsts.insert(UInst);
884 DeadInsts.insert(UInst2);
889 // Replace the old PHI Node with the inserted computation.
890 PN->replaceAllUsesWith(NewVal);
891 DeadInsts.insert(PN);
897 DeleteTriviallyDeadInstructions(DeadInsts);
898 assert(L->isLCSSAForm());
902 /// Return true if it is OK to use SIToFPInst for an inducation variable
903 /// with given inital and exit values.
904 static bool useSIToFPInst(ConstantFP &InitV, ConstantFP &ExitV,
905 uint64_t intIV, uint64_t intEV) {
907 if (InitV.getValueAPF().isNegative() || ExitV.getValueAPF().isNegative())
910 // If the iteration range can be handled by SIToFPInst then use it.
911 APInt Max = APInt::getSignedMaxValue(32);
912 if (Max.getZExtValue() > static_cast<uint64_t>(abs(intEV - intIV)))
918 /// convertToInt - Convert APF to an integer, if possible.
919 static bool convertToInt(const APFloat &APF, uint64_t *intVal) {
921 bool isExact = false;
922 if (&APF.getSemantics() == &APFloat::PPCDoubleDouble)
924 if (APF.convertToInteger(intVal, 32, APF.isNegative(),
925 APFloat::rmTowardZero, &isExact)
934 /// HandleFloatingPointIV - If the loop has floating induction variable
935 /// then insert corresponding integer induction variable if possible.
937 /// for(double i = 0; i < 10000; ++i)
939 /// is converted into
940 /// for(int i = 0; i < 10000; ++i)
943 void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PH,
944 SmallPtrSet<Instruction*, 16> &DeadInsts) {
946 unsigned IncomingEdge = L->contains(PH->getIncomingBlock(0));
947 unsigned BackEdge = IncomingEdge^1;
949 // Check incoming value.
950 ConstantFP *InitValue = dyn_cast<ConstantFP>(PH->getIncomingValue(IncomingEdge));
951 if (!InitValue) return;
952 uint64_t newInitValue = Type::Int32Ty->getPrimitiveSizeInBits();
953 if (!convertToInt(InitValue->getValueAPF(), &newInitValue))
956 // Check IV increment. Reject this PH if increement operation is not
957 // an add or increment value can not be represented by an integer.
958 BinaryOperator *Incr =
959 dyn_cast<BinaryOperator>(PH->getIncomingValue(BackEdge));
961 if (Incr->getOpcode() != Instruction::Add) return;
962 ConstantFP *IncrValue = NULL;
963 unsigned IncrVIndex = 1;
964 if (Incr->getOperand(1) == PH)
966 IncrValue = dyn_cast<ConstantFP>(Incr->getOperand(IncrVIndex));
967 if (!IncrValue) return;
968 uint64_t newIncrValue = Type::Int32Ty->getPrimitiveSizeInBits();
969 if (!convertToInt(IncrValue->getValueAPF(), &newIncrValue))
972 // Check Incr uses. One user is PH and the other users is exit condition used
973 // by the conditional terminator.
974 Value::use_iterator IncrUse = Incr->use_begin();
975 Instruction *U1 = cast<Instruction>(IncrUse++);
976 if (IncrUse == Incr->use_end()) return;
977 Instruction *U2 = cast<Instruction>(IncrUse++);
978 if (IncrUse != Incr->use_end()) return;
980 // Find exit condition.
981 FCmpInst *EC = dyn_cast<FCmpInst>(U1);
983 EC = dyn_cast<FCmpInst>(U2);
986 if (BranchInst *BI = dyn_cast<BranchInst>(EC->getParent()->getTerminator())) {
987 if (!BI->isConditional()) return;
988 if (BI->getCondition() != EC) return;
991 // Find exit value. If exit value can not be represented as an interger then
992 // do not handle this floating point PH.
993 ConstantFP *EV = NULL;
994 unsigned EVIndex = 1;
995 if (EC->getOperand(1) == Incr)
997 EV = dyn_cast<ConstantFP>(EC->getOperand(EVIndex));
999 uint64_t intEV = Type::Int32Ty->getPrimitiveSizeInBits();
1000 if (!convertToInt(EV->getValueAPF(), &intEV))
1003 // Find new predicate for integer comparison.
1004 CmpInst::Predicate NewPred = CmpInst::BAD_ICMP_PREDICATE;
1005 switch (EC->getPredicate()) {
1006 case CmpInst::FCMP_OEQ:
1007 case CmpInst::FCMP_UEQ:
1008 NewPred = CmpInst::ICMP_EQ;
1010 case CmpInst::FCMP_OGT:
1011 case CmpInst::FCMP_UGT:
1012 NewPred = CmpInst::ICMP_UGT;
1014 case CmpInst::FCMP_OGE:
1015 case CmpInst::FCMP_UGE:
1016 NewPred = CmpInst::ICMP_UGE;
1018 case CmpInst::FCMP_OLT:
1019 case CmpInst::FCMP_ULT:
1020 NewPred = CmpInst::ICMP_ULT;
1022 case CmpInst::FCMP_OLE:
1023 case CmpInst::FCMP_ULE:
1024 NewPred = CmpInst::ICMP_ULE;
1029 if (NewPred == CmpInst::BAD_ICMP_PREDICATE) return;
1031 // Insert new integer induction variable.
1032 PHINode *NewPHI = PHINode::Create(Type::Int32Ty,
1033 PH->getName()+".int", PH);
1034 NewPHI->addIncoming(ConstantInt::get(Type::Int32Ty, newInitValue),
1035 PH->getIncomingBlock(IncomingEdge));
1037 Value *NewAdd = BinaryOperator::CreateAdd(NewPHI,
1038 ConstantInt::get(Type::Int32Ty,
1040 Incr->getName()+".int", Incr);
1041 NewPHI->addIncoming(NewAdd, PH->getIncomingBlock(BackEdge));
1043 ConstantInt *NewEV = ConstantInt::get(Type::Int32Ty, intEV);
1044 Value *LHS = (EVIndex == 1 ? NewPHI->getIncomingValue(BackEdge) : NewEV);
1045 Value *RHS = (EVIndex == 1 ? NewEV : NewPHI->getIncomingValue(BackEdge));
1046 ICmpInst *NewEC = new ICmpInst(NewPred, LHS, RHS, EC->getNameStart(),
1047 EC->getParent()->getTerminator());
1049 // Delete old, floating point, exit comparision instruction.
1050 EC->replaceAllUsesWith(NewEC);
1051 DeadInsts.insert(EC);
1053 // Delete old, floating point, increment instruction.
1054 Incr->replaceAllUsesWith(UndefValue::get(Incr->getType()));
1055 DeadInsts.insert(Incr);
1057 // Replace floating induction variable. Give SIToFPInst preference over
1058 // UIToFPInst because it is faster on platforms that are widely used.
1059 if (useSIToFPInst(*InitValue, *EV, newInitValue, intEV)) {
1060 SIToFPInst *Conv = new SIToFPInst(NewPHI, PH->getType(), "indvar.conv",
1061 PH->getParent()->getFirstNonPHI());
1062 PH->replaceAllUsesWith(Conv);
1064 UIToFPInst *Conv = new UIToFPInst(NewPHI, PH->getType(), "indvar.conv",
1065 PH->getParent()->getFirstNonPHI());
1066 PH->replaceAllUsesWith(Conv);
1068 DeadInsts.insert(PH);