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(NumPointer , "Number of pointer indvars promoted");
63 STATISTIC(NumInserted, "Number of canonical indvars added");
64 STATISTIC(NumReplaced, "Number of exit values replaced");
65 STATISTIC(NumLFTR , "Number of loop exit tests replaced");
68 class VISIBILITY_HIDDEN IndVarSimplify : public LoopPass {
74 static char ID; // Pass identification, replacement for typeid
75 IndVarSimplify() : LoopPass(&ID) {}
77 virtual bool runOnLoop(Loop *L, LPPassManager &LPM);
79 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
80 AU.addRequired<ScalarEvolution>();
81 AU.addRequiredID(LCSSAID);
82 AU.addRequiredID(LoopSimplifyID);
83 AU.addRequired<LoopInfo>();
84 AU.addPreserved<ScalarEvolution>();
85 AU.addPreservedID(LoopSimplifyID);
86 AU.addPreservedID(LCSSAID);
92 void RewriteNonIntegerIVs(Loop *L);
94 void EliminatePointerRecurrence(PHINode *PN, BasicBlock *Preheader,
95 SmallPtrSet<Instruction*, 16> &DeadInsts);
96 void LinearFunctionTestReplace(Loop *L, SCEVHandle IterationCount,
98 BasicBlock *ExitingBlock,
100 SCEVExpander &Rewriter,
101 bool SignExtendTripCount);
102 void RewriteLoopExitValues(Loop *L, SCEV *IterationCount);
104 void DeleteTriviallyDeadInstructions(SmallPtrSet<Instruction*, 16> &Insts);
106 void HandleFloatingPointIV(Loop *L, PHINode *PH,
107 SmallPtrSet<Instruction*, 16> &DeadInsts);
111 char IndVarSimplify::ID = 0;
112 static RegisterPass<IndVarSimplify>
113 X("indvars", "Canonicalize Induction Variables");
115 Pass *llvm::createIndVarSimplifyPass() {
116 return new IndVarSimplify();
119 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
120 /// specified set are trivially dead, delete them and see if this makes any of
121 /// their operands subsequently dead.
122 void IndVarSimplify::
123 DeleteTriviallyDeadInstructions(SmallPtrSet<Instruction*, 16> &Insts) {
124 while (!Insts.empty()) {
125 Instruction *I = *Insts.begin();
127 if (isInstructionTriviallyDead(I)) {
128 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
129 if (Instruction *U = dyn_cast<Instruction>(I->getOperand(i)))
131 SE->deleteValueFromRecords(I);
132 DOUT << "INDVARS: Deleting: " << *I;
133 I->eraseFromParent();
140 /// EliminatePointerRecurrence - Check to see if this is a trivial GEP pointer
141 /// recurrence. If so, change it into an integer recurrence, permitting
142 /// analysis by the SCEV routines.
143 void IndVarSimplify::EliminatePointerRecurrence(PHINode *PN,
144 BasicBlock *Preheader,
145 SmallPtrSet<Instruction*, 16> &DeadInsts) {
146 assert(PN->getNumIncomingValues() == 2 && "Noncanonicalized loop!");
147 unsigned PreheaderIdx = PN->getBasicBlockIndex(Preheader);
148 unsigned BackedgeIdx = PreheaderIdx^1;
149 if (GetElementPtrInst *GEPI =
150 dyn_cast<GetElementPtrInst>(PN->getIncomingValue(BackedgeIdx)))
151 if (GEPI->getOperand(0) == PN) {
152 assert(GEPI->getNumOperands() == 2 && "GEP types must match!");
153 DOUT << "INDVARS: Eliminating pointer recurrence: " << *GEPI;
155 // Okay, we found a pointer recurrence. Transform this pointer
156 // recurrence into an integer recurrence. Compute the value that gets
157 // added to the pointer at every iteration.
158 Value *AddedVal = GEPI->getOperand(1);
160 // Insert a new integer PHI node into the top of the block.
161 PHINode *NewPhi = PHINode::Create(AddedVal->getType(),
162 PN->getName()+".rec", PN);
163 NewPhi->addIncoming(Constant::getNullValue(NewPhi->getType()), Preheader);
165 // Create the new add instruction.
166 Value *NewAdd = BinaryOperator::CreateAdd(NewPhi, AddedVal,
167 GEPI->getName()+".rec", GEPI);
168 NewPhi->addIncoming(NewAdd, PN->getIncomingBlock(BackedgeIdx));
170 // Update the existing GEP to use the recurrence.
171 GEPI->setOperand(0, PN->getIncomingValue(PreheaderIdx));
173 // Update the GEP to use the new recurrence we just inserted.
174 GEPI->setOperand(1, NewAdd);
176 // If the incoming value is a constant expr GEP, try peeling out the array
177 // 0 index if possible to make things simpler.
178 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEPI->getOperand(0)))
179 if (CE->getOpcode() == Instruction::GetElementPtr) {
180 unsigned NumOps = CE->getNumOperands();
181 assert(NumOps > 1 && "CE folding didn't work!");
182 if (CE->getOperand(NumOps-1)->isNullValue()) {
183 // Check to make sure the last index really is an array index.
184 gep_type_iterator GTI = gep_type_begin(CE);
185 for (unsigned i = 1, e = CE->getNumOperands()-1;
188 if (isa<SequentialType>(*GTI)) {
189 // Pull the last index out of the constant expr GEP.
190 SmallVector<Value*, 8> CEIdxs(CE->op_begin()+1, CE->op_end()-1);
191 Constant *NCE = ConstantExpr::getGetElementPtr(CE->getOperand(0),
195 Idx[0] = Constant::getNullValue(Type::Int32Ty);
197 GetElementPtrInst *NGEPI = GetElementPtrInst::Create(
199 GEPI->getName(), GEPI);
200 SE->deleteValueFromRecords(GEPI);
201 GEPI->replaceAllUsesWith(NGEPI);
202 GEPI->eraseFromParent();
209 // Finally, if there are any other users of the PHI node, we must
210 // insert a new GEP instruction that uses the pre-incremented version
211 // of the induction amount.
212 if (!PN->use_empty()) {
213 BasicBlock::iterator InsertPos = PN; ++InsertPos;
214 while (isa<PHINode>(InsertPos)) ++InsertPos;
216 GetElementPtrInst::Create(PN->getIncomingValue(PreheaderIdx),
217 NewPhi, "", InsertPos);
218 PreInc->takeName(PN);
219 PN->replaceAllUsesWith(PreInc);
222 // Delete the old PHI for sure, and the GEP if its otherwise unused.
223 DeadInsts.insert(PN);
230 /// LinearFunctionTestReplace - This method rewrites the exit condition of the
231 /// loop to be a canonical != comparison against the incremented loop induction
232 /// variable. This pass is able to rewrite the exit tests of any loop where the
233 /// SCEV analysis can determine a loop-invariant trip count of the loop, which
234 /// is actually a much broader range than just linear tests.
235 void IndVarSimplify::LinearFunctionTestReplace(Loop *L,
236 SCEVHandle IterationCount,
238 BasicBlock *ExitingBlock,
240 SCEVExpander &Rewriter,
241 bool SignExtendTripCount) {
242 // If the exiting block is not the same as the backedge block, we must compare
243 // against the preincremented value, otherwise we prefer to compare against
244 // the post-incremented value.
246 if (ExitingBlock == L->getLoopLatch()) {
247 // What ScalarEvolution calls the "iteration count" is actually the
248 // number of times the branch is taken. Add one to get the number
249 // of times the branch is executed. If this addition may overflow,
250 // we have to be more pessimistic and cast the induction variable
251 // before doing the add.
252 SCEVHandle Zero = SE->getIntegerSCEV(0, IterationCount->getType());
254 SE->getAddExpr(IterationCount,
255 SE->getIntegerSCEV(1, IterationCount->getType()));
256 if ((isa<SCEVConstant>(N) && !N->isZero()) ||
257 SE->isLoopGuardedByCond(L, ICmpInst::ICMP_NE, N, Zero)) {
258 // No overflow. Cast the sum.
259 if (SignExtendTripCount)
260 IterationCount = SE->getTruncateOrSignExtend(N, IndVar->getType());
262 IterationCount = SE->getTruncateOrZeroExtend(N, IndVar->getType());
264 // Potential overflow. Cast before doing the add.
265 if (SignExtendTripCount)
266 IterationCount = SE->getTruncateOrSignExtend(IterationCount,
269 IterationCount = SE->getTruncateOrZeroExtend(IterationCount,
272 SE->getAddExpr(IterationCount,
273 SE->getIntegerSCEV(1, IndVar->getType()));
276 // The IterationCount expression contains the number of times that the
277 // backedge actually branches to the loop header. This is one less than the
278 // number of times the loop executes, so add one to it.
279 CmpIndVar = L->getCanonicalInductionVariableIncrement();
281 // We have to use the preincremented value...
282 if (SignExtendTripCount)
283 IterationCount = SE->getTruncateOrSignExtend(IterationCount,
286 IterationCount = SE->getTruncateOrZeroExtend(IterationCount,
291 // Expand the code for the iteration count into the preheader of the loop.
292 BasicBlock *Preheader = L->getLoopPreheader();
293 Value *ExitCnt = Rewriter.expandCodeFor(IterationCount,
294 Preheader->getTerminator());
296 // Insert a new icmp_ne or icmp_eq instruction before the branch.
297 ICmpInst::Predicate Opcode;
298 if (L->contains(BI->getSuccessor(0)))
299 Opcode = ICmpInst::ICMP_NE;
301 Opcode = ICmpInst::ICMP_EQ;
303 DOUT << "INDVARS: Rewriting loop exit condition to:\n"
304 << " LHS:" << *CmpIndVar // includes a newline
306 << (Opcode == ICmpInst::ICMP_NE ? "!=" : "==") << "\n"
307 << " RHS:\t" << *IterationCount << "\n";
309 Value *Cond = new ICmpInst(Opcode, CmpIndVar, ExitCnt, "exitcond", BI);
310 BI->setCondition(Cond);
315 /// RewriteLoopExitValues - Check to see if this loop has a computable
316 /// loop-invariant execution count. If so, this means that we can compute the
317 /// final value of any expressions that are recurrent in the loop, and
318 /// substitute the exit values from the loop into any instructions outside of
319 /// the loop that use the final values of the current expressions.
320 void IndVarSimplify::RewriteLoopExitValues(Loop *L, SCEV *IterationCount) {
321 BasicBlock *Preheader = L->getLoopPreheader();
323 // Scan all of the instructions in the loop, looking at those that have
324 // extra-loop users and which are recurrences.
325 SCEVExpander Rewriter(*SE, *LI);
327 // We insert the code into the preheader of the loop if the loop contains
328 // multiple exit blocks, or in the exit block if there is exactly one.
329 BasicBlock *BlockToInsertInto;
330 SmallVector<BasicBlock*, 8> ExitBlocks;
331 L->getUniqueExitBlocks(ExitBlocks);
332 if (ExitBlocks.size() == 1)
333 BlockToInsertInto = ExitBlocks[0];
335 BlockToInsertInto = Preheader;
336 BasicBlock::iterator InsertPt = BlockToInsertInto->getFirstNonPHI();
338 bool HasConstantItCount = isa<SCEVConstant>(IterationCount);
340 SmallPtrSet<Instruction*, 16> InstructionsToDelete;
341 std::map<Instruction*, Value*> ExitValues;
343 // Find all values that are computed inside the loop, but used outside of it.
344 // Because of LCSSA, these values will only occur in LCSSA PHI Nodes. Scan
345 // the exit blocks of the loop to find them.
346 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
347 BasicBlock *ExitBB = ExitBlocks[i];
349 // If there are no PHI nodes in this exit block, then no values defined
350 // inside the loop are used on this path, skip it.
351 PHINode *PN = dyn_cast<PHINode>(ExitBB->begin());
354 unsigned NumPreds = PN->getNumIncomingValues();
356 // Iterate over all of the PHI nodes.
357 BasicBlock::iterator BBI = ExitBB->begin();
358 while ((PN = dyn_cast<PHINode>(BBI++))) {
360 // Iterate over all of the values in all the PHI nodes.
361 for (unsigned i = 0; i != NumPreds; ++i) {
362 // If the value being merged in is not integer or is not defined
363 // in the loop, skip it.
364 Value *InVal = PN->getIncomingValue(i);
365 if (!isa<Instruction>(InVal) ||
366 // SCEV only supports integer expressions for now.
367 !isa<IntegerType>(InVal->getType()))
370 // If this pred is for a subloop, not L itself, skip it.
371 if (LI->getLoopFor(PN->getIncomingBlock(i)) != L)
372 continue; // The Block is in a subloop, skip it.
374 // Check that InVal is defined in the loop.
375 Instruction *Inst = cast<Instruction>(InVal);
376 if (!L->contains(Inst->getParent()))
379 // We require that this value either have a computable evolution or that
380 // the loop have a constant iteration count. In the case where the loop
381 // has a constant iteration count, we can sometimes force evaluation of
382 // the exit value through brute force.
383 SCEVHandle SH = SE->getSCEV(Inst);
384 if (!SH->hasComputableLoopEvolution(L) && !HasConstantItCount)
385 continue; // Cannot get exit evolution for the loop value.
387 // Okay, this instruction has a user outside of the current loop
388 // and varies predictably *inside* the loop. Evaluate the value it
389 // contains when the loop exits, if possible.
390 SCEVHandle ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop());
391 if (isa<SCEVCouldNotCompute>(ExitValue) ||
392 !ExitValue->isLoopInvariant(L))
398 // See if we already computed the exit value for the instruction, if so,
400 Value *&ExitVal = ExitValues[Inst];
402 ExitVal = Rewriter.expandCodeFor(ExitValue, InsertPt);
404 DOUT << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal
405 << " LoopVal = " << *Inst << "\n";
407 PN->setIncomingValue(i, ExitVal);
409 // If this instruction is dead now, schedule it to be removed.
410 if (Inst->use_empty())
411 InstructionsToDelete.insert(Inst);
413 // See if this is a single-entry LCSSA PHI node. If so, we can (and
415 // the PHI entirely. This is safe, because the NewVal won't be variant
416 // in the loop, so we don't need an LCSSA phi node anymore.
418 SE->deleteValueFromRecords(PN);
419 PN->replaceAllUsesWith(ExitVal);
420 PN->eraseFromParent();
427 DeleteTriviallyDeadInstructions(InstructionsToDelete);
430 void IndVarSimplify::RewriteNonIntegerIVs(Loop *L) {
431 // First step. Check to see if there are any trivial GEP pointer recurrences.
432 // If there are, change them into integer recurrences, permitting analysis by
433 // the SCEV routines.
435 BasicBlock *Header = L->getHeader();
436 BasicBlock *Preheader = L->getLoopPreheader();
438 SmallPtrSet<Instruction*, 16> DeadInsts;
439 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
440 PHINode *PN = cast<PHINode>(I);
441 if (isa<PointerType>(PN->getType()))
442 EliminatePointerRecurrence(PN, Preheader, DeadInsts);
444 HandleFloatingPointIV(L, PN, DeadInsts);
447 // If the loop previously had a pointer or floating-point IV, ScalarEvolution
448 // may not have been able to compute a trip count. Now that we've done some
449 // re-writing, the trip count may be computable.
451 SE->forgetLoopIterationCount(L);
453 if (!DeadInsts.empty())
454 DeleteTriviallyDeadInstructions(DeadInsts);
457 /// getEffectiveIndvarType - Determine the widest type that the
458 /// induction-variable PHINode Phi is cast to.
460 static const Type *getEffectiveIndvarType(const PHINode *Phi) {
461 const Type *Ty = Phi->getType();
463 for (Value::use_const_iterator UI = Phi->use_begin(), UE = Phi->use_end();
465 const Type *CandidateType = NULL;
466 if (const ZExtInst *ZI = dyn_cast<ZExtInst>(UI))
467 CandidateType = ZI->getDestTy();
468 else if (const SExtInst *SI = dyn_cast<SExtInst>(UI))
469 CandidateType = SI->getDestTy();
471 CandidateType->getPrimitiveSizeInBits() >
472 Ty->getPrimitiveSizeInBits())
479 /// TestOrigIVForWrap - Analyze the original induction variable
480 /// that controls the loop's iteration to determine whether it
481 /// would ever undergo signed or unsigned overflow. Also, check
482 /// whether an induction variable in the same type that starts
483 /// at 0 would undergo signed overflow.
485 /// In addition to setting the NoSignedWrap, NoUnsignedWrap, and
486 /// SignExtendTripCount variables, return the PHI for this induction
489 /// TODO: This duplicates a fair amount of ScalarEvolution logic.
490 /// Perhaps this can be merged with ScalarEvolution::getIterationCount
491 /// and/or ScalarEvolution::get{Sign,Zero}ExtendExpr.
493 static const PHINode *TestOrigIVForWrap(const Loop *L,
494 const BranchInst *BI,
495 const Instruction *OrigCond,
497 bool &NoUnsignedWrap,
498 bool &SignExtendTripCount) {
499 // Verify that the loop is sane and find the exit condition.
500 const ICmpInst *Cmp = dyn_cast<ICmpInst>(OrigCond);
503 const Value *CmpLHS = Cmp->getOperand(0);
504 const Value *CmpRHS = Cmp->getOperand(1);
505 const BasicBlock *TrueBB = BI->getSuccessor(0);
506 const BasicBlock *FalseBB = BI->getSuccessor(1);
507 ICmpInst::Predicate Pred = Cmp->getPredicate();
509 // Canonicalize a constant to the RHS.
510 if (isa<ConstantInt>(CmpLHS)) {
511 Pred = ICmpInst::getSwappedPredicate(Pred);
512 std::swap(CmpLHS, CmpRHS);
514 // Canonicalize SLE to SLT.
515 if (Pred == ICmpInst::ICMP_SLE)
516 if (const ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS))
517 if (!CI->getValue().isMaxSignedValue()) {
518 CmpRHS = ConstantInt::get(CI->getValue() + 1);
519 Pred = ICmpInst::ICMP_SLT;
521 // Canonicalize SGT to SGE.
522 if (Pred == ICmpInst::ICMP_SGT)
523 if (const ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS))
524 if (!CI->getValue().isMaxSignedValue()) {
525 CmpRHS = ConstantInt::get(CI->getValue() + 1);
526 Pred = ICmpInst::ICMP_SGE;
528 // Canonicalize SGE to SLT.
529 if (Pred == ICmpInst::ICMP_SGE) {
530 std::swap(TrueBB, FalseBB);
531 Pred = ICmpInst::ICMP_SLT;
533 // Canonicalize ULE to ULT.
534 if (Pred == ICmpInst::ICMP_ULE)
535 if (const ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS))
536 if (!CI->getValue().isMaxValue()) {
537 CmpRHS = ConstantInt::get(CI->getValue() + 1);
538 Pred = ICmpInst::ICMP_ULT;
540 // Canonicalize UGT to UGE.
541 if (Pred == ICmpInst::ICMP_UGT)
542 if (const ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS))
543 if (!CI->getValue().isMaxValue()) {
544 CmpRHS = ConstantInt::get(CI->getValue() + 1);
545 Pred = ICmpInst::ICMP_UGE;
547 // Canonicalize UGE to ULT.
548 if (Pred == ICmpInst::ICMP_UGE) {
549 std::swap(TrueBB, FalseBB);
550 Pred = ICmpInst::ICMP_ULT;
552 // For now, analyze only LT loops for signed overflow.
553 if (Pred != ICmpInst::ICMP_SLT && Pred != ICmpInst::ICMP_ULT)
556 bool isSigned = Pred == ICmpInst::ICMP_SLT;
558 // Get the increment instruction. Look past casts if we will
559 // be able to prove that the original induction variable doesn't
560 // undergo signed or unsigned overflow, respectively.
561 const Value *IncrVal = CmpLHS;
563 if (const SExtInst *SI = dyn_cast<SExtInst>(CmpLHS)) {
564 if (!isa<ConstantInt>(CmpRHS) ||
565 !cast<ConstantInt>(CmpRHS)->getValue()
566 .isSignedIntN(IncrVal->getType()->getPrimitiveSizeInBits()))
568 IncrVal = SI->getOperand(0);
571 if (const ZExtInst *ZI = dyn_cast<ZExtInst>(CmpLHS)) {
572 if (!isa<ConstantInt>(CmpRHS) ||
573 !cast<ConstantInt>(CmpRHS)->getValue()
574 .isIntN(IncrVal->getType()->getPrimitiveSizeInBits()))
576 IncrVal = ZI->getOperand(0);
580 // For now, only analyze induction variables that have simple increments.
581 const BinaryOperator *IncrOp = dyn_cast<BinaryOperator>(IncrVal);
583 IncrOp->getOpcode() != Instruction::Add ||
584 !isa<ConstantInt>(IncrOp->getOperand(1)) ||
585 !cast<ConstantInt>(IncrOp->getOperand(1))->equalsInt(1))
588 // Make sure the PHI looks like a normal IV.
589 const PHINode *PN = dyn_cast<PHINode>(IncrOp->getOperand(0));
590 if (!PN || PN->getNumIncomingValues() != 2)
592 unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
593 unsigned BackEdge = !IncomingEdge;
594 if (!L->contains(PN->getIncomingBlock(BackEdge)) ||
595 PN->getIncomingValue(BackEdge) != IncrOp)
597 if (!L->contains(TrueBB))
600 // For now, only analyze loops with a constant start value, so that
601 // we can easily determine if the start value is not a maximum value
602 // which would wrap on the first iteration.
603 const ConstantInt *InitialVal =
604 dyn_cast<ConstantInt>(PN->getIncomingValue(IncomingEdge));
608 // The original induction variable will start at some non-max value,
609 // it counts up by one, and the loop iterates only while it remans
610 // less than some value in the same type. As such, it will never wrap.
611 if (isSigned && !InitialVal->getValue().isMaxSignedValue()) {
613 // If the original induction variable starts at zero or greater,
614 // the trip count can be considered signed.
615 if (InitialVal->getValue().isNonNegative())
616 SignExtendTripCount = true;
617 } else if (!isSigned && !InitialVal->getValue().isMaxValue())
618 NoUnsignedWrap = true;
622 bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) {
623 LI = &getAnalysis<LoopInfo>();
624 SE = &getAnalysis<ScalarEvolution>();
627 // If there are any floating-point or pointer recurrences, attempt to
628 // transform them to use integer recurrences.
629 RewriteNonIntegerIVs(L);
631 BasicBlock *Header = L->getHeader();
632 BasicBlock *ExitingBlock = L->getExitingBlock();
633 SmallPtrSet<Instruction*, 16> DeadInsts;
635 // Verify the input to the pass in already in LCSSA form.
636 assert(L->isLCSSAForm());
638 // Check to see if this loop has a computable loop-invariant execution count.
639 // If so, this means that we can compute the final value of any expressions
640 // that are recurrent in the loop, and substitute the exit values from the
641 // loop into any instructions outside of the loop that use the final values of
642 // the current expressions.
644 SCEVHandle IterationCount = SE->getIterationCount(L);
645 if (!isa<SCEVCouldNotCompute>(IterationCount))
646 RewriteLoopExitValues(L, IterationCount);
648 // Next, analyze all of the induction variables in the loop, canonicalizing
649 // auxillary induction variables.
650 std::vector<std::pair<PHINode*, SCEVHandle> > IndVars;
652 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
653 PHINode *PN = cast<PHINode>(I);
654 if (PN->getType()->isInteger()) { // FIXME: when we have fast-math, enable!
655 SCEVHandle SCEV = SE->getSCEV(PN);
656 // FIXME: It is an extremely bad idea to indvar substitute anything more
657 // complex than affine induction variables. Doing so will put expensive
658 // polynomial evaluations inside of the loop, and the str reduction pass
659 // currently can only reduce affine polynomials. For now just disable
660 // indvar subst on anything more complex than an affine addrec.
661 if (SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SCEV))
662 if (AR->getLoop() == L && AR->isAffine())
663 IndVars.push_back(std::make_pair(PN, SCEV));
667 // Compute the type of the largest recurrence expression, and collect
668 // the set of the types of the other recurrence expressions.
669 const Type *LargestType = 0;
670 SmallSetVector<const Type *, 4> SizesToInsert;
671 if (!isa<SCEVCouldNotCompute>(IterationCount)) {
672 LargestType = IterationCount->getType();
673 SizesToInsert.insert(IterationCount->getType());
675 for (unsigned i = 0, e = IndVars.size(); i != e; ++i) {
676 const PHINode *PN = IndVars[i].first;
677 SizesToInsert.insert(PN->getType());
678 const Type *EffTy = getEffectiveIndvarType(PN);
679 SizesToInsert.insert(EffTy);
681 EffTy->getPrimitiveSizeInBits() >
682 LargestType->getPrimitiveSizeInBits())
686 // Create a rewriter object which we'll use to transform the code with.
687 SCEVExpander Rewriter(*SE, *LI);
689 // Now that we know the largest of of the induction variables in this loop,
690 // insert a canonical induction variable of the largest size.
692 if (!SizesToInsert.empty()) {
693 IndVar = Rewriter.getOrInsertCanonicalInductionVariable(L,LargestType);
696 DOUT << "INDVARS: New CanIV: " << *IndVar;
699 // If we have a trip count expression, rewrite the loop's exit condition
700 // using it. We can currently only handle loops with a single exit.
701 bool NoSignedWrap = false;
702 bool NoUnsignedWrap = false;
703 bool SignExtendTripCount = false;
704 const PHINode *OrigControllingPHI = 0;
705 if (!isa<SCEVCouldNotCompute>(IterationCount) && ExitingBlock)
706 // Can't rewrite non-branch yet.
707 if (BranchInst *BI = dyn_cast<BranchInst>(ExitingBlock->getTerminator())) {
708 if (Instruction *OrigCond = dyn_cast<Instruction>(BI->getCondition())) {
709 // Determine if the OrigIV will ever undergo overflow.
711 TestOrigIVForWrap(L, BI, OrigCond,
712 NoSignedWrap, NoUnsignedWrap,
713 SignExtendTripCount);
715 // We'll be replacing the original condition, so it'll be dead.
716 DeadInsts.insert(OrigCond);
719 LinearFunctionTestReplace(L, IterationCount, IndVar,
720 ExitingBlock, BI, Rewriter,
721 SignExtendTripCount);
724 // Now that we have a canonical induction variable, we can rewrite any
725 // recurrences in terms of the induction variable. Start with the auxillary
726 // induction variables, and recursively rewrite any of their uses.
727 BasicBlock::iterator InsertPt = Header->getFirstNonPHI();
729 // If there were induction variables of other sizes, cast the primary
730 // induction variable to the right size for them, avoiding the need for the
731 // code evaluation methods to insert induction variables of different sizes.
732 for (unsigned i = 0, e = SizesToInsert.size(); i != e; ++i) {
733 const Type *Ty = SizesToInsert[i];
734 if (Ty != LargestType) {
735 Instruction *New = new TruncInst(IndVar, Ty, "indvar", InsertPt);
736 Rewriter.addInsertedValue(New, SE->getSCEV(New));
737 DOUT << "INDVARS: Made trunc IV for type " << *Ty << ": "
742 // Rewrite all induction variables in terms of the canonical induction
744 while (!IndVars.empty()) {
745 PHINode *PN = IndVars.back().first;
746 SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(IndVars.back().second);
747 Value *NewVal = Rewriter.expandCodeFor(AR, InsertPt);
748 DOUT << "INDVARS: Rewrote IV '" << *AR << "' " << *PN
749 << " into = " << *NewVal << "\n";
750 NewVal->takeName(PN);
752 /// If the new canonical induction variable is wider than the original,
753 /// and the original has uses that are casts to wider types, see if the
754 /// truncate and extend can be omitted.
755 if (PN == OrigControllingPHI && PN->getType() != LargestType)
756 for (Value::use_iterator UI = PN->use_begin(), UE = PN->use_end();
758 if (isa<SExtInst>(UI) && NoSignedWrap) {
759 SCEVHandle ExtendedStart =
760 SE->getSignExtendExpr(AR->getStart(), LargestType);
761 SCEVHandle ExtendedStep =
762 SE->getSignExtendExpr(AR->getStepRecurrence(*SE), LargestType);
763 SCEVHandle ExtendedAddRec =
764 SE->getAddRecExpr(ExtendedStart, ExtendedStep, L);
765 if (LargestType != UI->getType())
766 ExtendedAddRec = SE->getTruncateExpr(ExtendedAddRec, UI->getType());
767 Value *TruncIndVar = Rewriter.expandCodeFor(ExtendedAddRec, InsertPt);
768 UI->replaceAllUsesWith(TruncIndVar);
769 if (Instruction *DeadUse = dyn_cast<Instruction>(*UI))
770 DeadInsts.insert(DeadUse);
772 if (isa<ZExtInst>(UI) && NoUnsignedWrap) {
773 SCEVHandle ExtendedStart =
774 SE->getZeroExtendExpr(AR->getStart(), LargestType);
775 SCEVHandle ExtendedStep =
776 SE->getZeroExtendExpr(AR->getStepRecurrence(*SE), LargestType);
777 SCEVHandle ExtendedAddRec =
778 SE->getAddRecExpr(ExtendedStart, ExtendedStep, L);
779 if (LargestType != UI->getType())
780 ExtendedAddRec = SE->getTruncateExpr(ExtendedAddRec, UI->getType());
781 Value *TruncIndVar = Rewriter.expandCodeFor(ExtendedAddRec, InsertPt);
782 UI->replaceAllUsesWith(TruncIndVar);
783 if (Instruction *DeadUse = dyn_cast<Instruction>(*UI))
784 DeadInsts.insert(DeadUse);
788 // Replace the old PHI Node with the inserted computation.
789 PN->replaceAllUsesWith(NewVal);
790 DeadInsts.insert(PN);
796 DeleteTriviallyDeadInstructions(DeadInsts);
797 assert(L->isLCSSAForm());
801 /// Return true if it is OK to use SIToFPInst for an inducation variable
802 /// with given inital and exit values.
803 static bool useSIToFPInst(ConstantFP &InitV, ConstantFP &ExitV,
804 uint64_t intIV, uint64_t intEV) {
806 if (InitV.getValueAPF().isNegative() || ExitV.getValueAPF().isNegative())
809 // If the iteration range can be handled by SIToFPInst then use it.
810 APInt Max = APInt::getSignedMaxValue(32);
811 if (Max.getZExtValue() > static_cast<uint64_t>(abs(intEV - intIV)))
817 /// convertToInt - Convert APF to an integer, if possible.
818 static bool convertToInt(const APFloat &APF, uint64_t *intVal) {
820 bool isExact = false;
821 if (&APF.getSemantics() == &APFloat::PPCDoubleDouble)
823 if (APF.convertToInteger(intVal, 32, APF.isNegative(),
824 APFloat::rmTowardZero, &isExact)
833 /// HandleFloatingPointIV - If the loop has floating induction variable
834 /// then insert corresponding integer induction variable if possible.
836 /// for(double i = 0; i < 10000; ++i)
838 /// is converted into
839 /// for(int i = 0; i < 10000; ++i)
842 void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PH,
843 SmallPtrSet<Instruction*, 16> &DeadInsts) {
845 unsigned IncomingEdge = L->contains(PH->getIncomingBlock(0));
846 unsigned BackEdge = IncomingEdge^1;
848 // Check incoming value.
849 ConstantFP *InitValue = dyn_cast<ConstantFP>(PH->getIncomingValue(IncomingEdge));
850 if (!InitValue) return;
851 uint64_t newInitValue = Type::Int32Ty->getPrimitiveSizeInBits();
852 if (!convertToInt(InitValue->getValueAPF(), &newInitValue))
855 // Check IV increment. Reject this PH if increement operation is not
856 // an add or increment value can not be represented by an integer.
857 BinaryOperator *Incr =
858 dyn_cast<BinaryOperator>(PH->getIncomingValue(BackEdge));
860 if (Incr->getOpcode() != Instruction::Add) return;
861 ConstantFP *IncrValue = NULL;
862 unsigned IncrVIndex = 1;
863 if (Incr->getOperand(1) == PH)
865 IncrValue = dyn_cast<ConstantFP>(Incr->getOperand(IncrVIndex));
866 if (!IncrValue) return;
867 uint64_t newIncrValue = Type::Int32Ty->getPrimitiveSizeInBits();
868 if (!convertToInt(IncrValue->getValueAPF(), &newIncrValue))
871 // Check Incr uses. One user is PH and the other users is exit condition used
872 // by the conditional terminator.
873 Value::use_iterator IncrUse = Incr->use_begin();
874 Instruction *U1 = cast<Instruction>(IncrUse++);
875 if (IncrUse == Incr->use_end()) return;
876 Instruction *U2 = cast<Instruction>(IncrUse++);
877 if (IncrUse != Incr->use_end()) return;
879 // Find exit condition.
880 FCmpInst *EC = dyn_cast<FCmpInst>(U1);
882 EC = dyn_cast<FCmpInst>(U2);
885 if (BranchInst *BI = dyn_cast<BranchInst>(EC->getParent()->getTerminator())) {
886 if (!BI->isConditional()) return;
887 if (BI->getCondition() != EC) return;
890 // Find exit value. If exit value can not be represented as an interger then
891 // do not handle this floating point PH.
892 ConstantFP *EV = NULL;
893 unsigned EVIndex = 1;
894 if (EC->getOperand(1) == Incr)
896 EV = dyn_cast<ConstantFP>(EC->getOperand(EVIndex));
898 uint64_t intEV = Type::Int32Ty->getPrimitiveSizeInBits();
899 if (!convertToInt(EV->getValueAPF(), &intEV))
902 // Find new predicate for integer comparison.
903 CmpInst::Predicate NewPred = CmpInst::BAD_ICMP_PREDICATE;
904 switch (EC->getPredicate()) {
905 case CmpInst::FCMP_OEQ:
906 case CmpInst::FCMP_UEQ:
907 NewPred = CmpInst::ICMP_EQ;
909 case CmpInst::FCMP_OGT:
910 case CmpInst::FCMP_UGT:
911 NewPred = CmpInst::ICMP_UGT;
913 case CmpInst::FCMP_OGE:
914 case CmpInst::FCMP_UGE:
915 NewPred = CmpInst::ICMP_UGE;
917 case CmpInst::FCMP_OLT:
918 case CmpInst::FCMP_ULT:
919 NewPred = CmpInst::ICMP_ULT;
921 case CmpInst::FCMP_OLE:
922 case CmpInst::FCMP_ULE:
923 NewPred = CmpInst::ICMP_ULE;
928 if (NewPred == CmpInst::BAD_ICMP_PREDICATE) return;
930 // Insert new integer induction variable.
931 PHINode *NewPHI = PHINode::Create(Type::Int32Ty,
932 PH->getName()+".int", PH);
933 NewPHI->addIncoming(ConstantInt::get(Type::Int32Ty, newInitValue),
934 PH->getIncomingBlock(IncomingEdge));
936 Value *NewAdd = BinaryOperator::CreateAdd(NewPHI,
937 ConstantInt::get(Type::Int32Ty,
939 Incr->getName()+".int", Incr);
940 NewPHI->addIncoming(NewAdd, PH->getIncomingBlock(BackEdge));
942 ConstantInt *NewEV = ConstantInt::get(Type::Int32Ty, intEV);
943 Value *LHS = (EVIndex == 1 ? NewPHI->getIncomingValue(BackEdge) : NewEV);
944 Value *RHS = (EVIndex == 1 ? NewEV : NewPHI->getIncomingValue(BackEdge));
945 ICmpInst *NewEC = new ICmpInst(NewPred, LHS, RHS, EC->getNameStart(),
946 EC->getParent()->getTerminator());
948 // Delete old, floating point, exit comparision instruction.
949 EC->replaceAllUsesWith(NewEC);
950 DeadInsts.insert(EC);
952 // Delete old, floating point, increment instruction.
953 Incr->replaceAllUsesWith(UndefValue::get(Incr->getType()));
954 DeadInsts.insert(Incr);
956 // Replace floating induction variable. Give SIToFPInst preference over
957 // UIToFPInst because it is faster on platforms that are widely used.
958 if (useSIToFPInst(*InitValue, *EV, newInitValue, intEV)) {
959 SIToFPInst *Conv = new SIToFPInst(NewPHI, PH->getType(), "indvar.conv",
960 PH->getParent()->getFirstNonPHI());
961 PH->replaceAllUsesWith(Conv);
963 UIToFPInst *Conv = new UIToFPInst(NewPHI, PH->getType(), "indvar.conv",
964 PH->getParent()->getFirstNonPHI());
965 PH->replaceAllUsesWith(Conv);
967 DeadInsts.insert(PH);