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 // If the trip count of a loop is computable, this pass also makes the following
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.
25 //===----------------------------------------------------------------------===//
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"
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");
61 static cl::opt<bool> EnableIVRewrite(
62 "enable-iv-rewrite", cl::Hidden,
63 cl::desc("Enable canonical induction variable rewriting"));
65 // Trip count verification can be enabled by default under NDEBUG if we
66 // implement a strong expression equivalence checker in SCEV. Until then, we
67 // use the verify-indvars flag, which may assert in some cases.
68 static cl::opt<bool> VerifyIndvars(
69 "verify-indvars", cl::Hidden,
70 cl::desc("Verify the ScalarEvolution result after running indvars"));
73 class IndVarSimplify : public LoopPass {
80 SmallVector<WeakVH, 16> DeadInsts;
84 static char ID; // Pass identification, replacement for typeid
85 IndVarSimplify() : LoopPass(ID), IU(0), LI(0), SE(0), DT(0), TD(0),
87 initializeIndVarSimplifyPass(*PassRegistry::getPassRegistry());
90 virtual bool runOnLoop(Loop *L, LPPassManager &LPM);
92 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
93 AU.addRequired<DominatorTree>();
94 AU.addRequired<LoopInfo>();
95 AU.addRequired<ScalarEvolution>();
96 AU.addRequiredID(LoopSimplifyID);
97 AU.addRequiredID(LCSSAID);
99 AU.addRequired<IVUsers>();
100 AU.addPreserved<ScalarEvolution>();
101 AU.addPreservedID(LoopSimplifyID);
102 AU.addPreservedID(LCSSAID);
104 AU.addPreserved<IVUsers>();
105 AU.setPreservesCFG();
109 virtual void releaseMemory() {
113 bool isValidRewrite(Value *FromVal, Value *ToVal);
115 void HandleFloatingPointIV(Loop *L, PHINode *PH);
116 void RewriteNonIntegerIVs(Loop *L);
118 void SimplifyAndExtend(Loop *L, SCEVExpander &Rewriter, LPPassManager &LPM);
120 void RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter);
122 void RewriteIVExpressions(Loop *L, SCEVExpander &Rewriter);
124 Value *LinearFunctionTestReplace(Loop *L, const SCEV *BackedgeTakenCount,
125 PHINode *IndVar, SCEVExpander &Rewriter);
127 void SinkUnusedInvariants(Loop *L);
131 char IndVarSimplify::ID = 0;
132 INITIALIZE_PASS_BEGIN(IndVarSimplify, "indvars",
133 "Induction Variable Simplification", false, false)
134 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
135 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
136 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
137 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
138 INITIALIZE_PASS_DEPENDENCY(LCSSA)
139 INITIALIZE_PASS_DEPENDENCY(IVUsers)
140 INITIALIZE_PASS_END(IndVarSimplify, "indvars",
141 "Induction Variable Simplification", false, false)
143 Pass *llvm::createIndVarSimplifyPass() {
144 return new IndVarSimplify();
147 /// isValidRewrite - Return true if the SCEV expansion generated by the
148 /// rewriter can replace the original value. SCEV guarantees that it
149 /// produces the same value, but the way it is produced may be illegal IR.
150 /// Ideally, this function will only be called for verification.
151 bool IndVarSimplify::isValidRewrite(Value *FromVal, Value *ToVal) {
152 // If an SCEV expression subsumed multiple pointers, its expansion could
153 // reassociate the GEP changing the base pointer. This is illegal because the
154 // final address produced by a GEP chain must be inbounds relative to its
155 // underlying object. Otherwise basic alias analysis, among other things,
156 // could fail in a dangerous way. Ultimately, SCEV will be improved to avoid
157 // producing an expression involving multiple pointers. Until then, we must
160 // Retrieve the pointer operand of the GEP. Don't use GetUnderlyingObject
161 // because it understands lcssa phis while SCEV does not.
162 Value *FromPtr = FromVal;
163 Value *ToPtr = ToVal;
164 if (GEPOperator *GEP = dyn_cast<GEPOperator>(FromVal)) {
165 FromPtr = GEP->getPointerOperand();
167 if (GEPOperator *GEP = dyn_cast<GEPOperator>(ToVal)) {
168 ToPtr = GEP->getPointerOperand();
170 if (FromPtr != FromVal || ToPtr != ToVal) {
171 // Quickly check the common case
172 if (FromPtr == ToPtr)
175 // SCEV may have rewritten an expression that produces the GEP's pointer
176 // operand. That's ok as long as the pointer operand has the same base
177 // pointer. Unlike GetUnderlyingObject(), getPointerBase() will find the
178 // base of a recurrence. This handles the case in which SCEV expansion
179 // converts a pointer type recurrence into a nonrecurrent pointer base
180 // indexed by an integer recurrence.
182 // If the GEP base pointer is a vector of pointers, abort.
183 if (!FromPtr->getType()->isPointerTy() || !ToPtr->getType()->isPointerTy())
186 const SCEV *FromBase = SE->getPointerBase(SE->getSCEV(FromPtr));
187 const SCEV *ToBase = SE->getPointerBase(SE->getSCEV(ToPtr));
188 if (FromBase == ToBase)
191 DEBUG(dbgs() << "INDVARS: GEP rewrite bail out "
192 << *FromBase << " != " << *ToBase << "\n");
199 /// Determine the insertion point for this user. By default, insert immediately
200 /// before the user. SCEVExpander or LICM will hoist loop invariants out of the
201 /// loop. For PHI nodes, there may be multiple uses, so compute the nearest
202 /// common dominator for the incoming blocks.
203 static Instruction *getInsertPointForUses(Instruction *User, Value *Def,
205 PHINode *PHI = dyn_cast<PHINode>(User);
209 Instruction *InsertPt = 0;
210 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i) {
211 if (PHI->getIncomingValue(i) != Def)
214 BasicBlock *InsertBB = PHI->getIncomingBlock(i);
216 InsertPt = InsertBB->getTerminator();
219 InsertBB = DT->findNearestCommonDominator(InsertPt->getParent(), InsertBB);
220 InsertPt = InsertBB->getTerminator();
222 assert(InsertPt && "Missing phi operand");
223 assert((!isa<Instruction>(Def) ||
224 DT->dominates(cast<Instruction>(Def), InsertPt)) &&
225 "def does not dominate all uses");
229 //===----------------------------------------------------------------------===//
230 // RewriteNonIntegerIVs and helpers. Prefer integer IVs.
231 //===----------------------------------------------------------------------===//
233 /// ConvertToSInt - Convert APF to an integer, if possible.
234 static bool ConvertToSInt(const APFloat &APF, int64_t &IntVal) {
235 bool isExact = false;
236 if (&APF.getSemantics() == &APFloat::PPCDoubleDouble)
238 // See if we can convert this to an int64_t
240 if (APF.convertToInteger(&UIntVal, 64, true, APFloat::rmTowardZero,
241 &isExact) != APFloat::opOK || !isExact)
247 /// HandleFloatingPointIV - If the loop has floating induction variable
248 /// then insert corresponding integer induction variable if possible.
250 /// for(double i = 0; i < 10000; ++i)
252 /// is converted into
253 /// for(int i = 0; i < 10000; ++i)
256 void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PN) {
257 unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
258 unsigned BackEdge = IncomingEdge^1;
260 // Check incoming value.
261 ConstantFP *InitValueVal =
262 dyn_cast<ConstantFP>(PN->getIncomingValue(IncomingEdge));
265 if (!InitValueVal || !ConvertToSInt(InitValueVal->getValueAPF(), InitValue))
268 // Check IV increment. Reject this PN if increment operation is not
269 // an add or increment value can not be represented by an integer.
270 BinaryOperator *Incr =
271 dyn_cast<BinaryOperator>(PN->getIncomingValue(BackEdge));
272 if (Incr == 0 || Incr->getOpcode() != Instruction::FAdd) return;
274 // If this is not an add of the PHI with a constantfp, or if the constant fp
275 // is not an integer, bail out.
276 ConstantFP *IncValueVal = dyn_cast<ConstantFP>(Incr->getOperand(1));
278 if (IncValueVal == 0 || Incr->getOperand(0) != PN ||
279 !ConvertToSInt(IncValueVal->getValueAPF(), IncValue))
282 // Check Incr uses. One user is PN and the other user is an exit condition
283 // used by the conditional terminator.
284 Value::use_iterator IncrUse = Incr->use_begin();
285 Instruction *U1 = cast<Instruction>(*IncrUse++);
286 if (IncrUse == Incr->use_end()) return;
287 Instruction *U2 = cast<Instruction>(*IncrUse++);
288 if (IncrUse != Incr->use_end()) return;
290 // Find exit condition, which is an fcmp. If it doesn't exist, or if it isn't
291 // only used by a branch, we can't transform it.
292 FCmpInst *Compare = dyn_cast<FCmpInst>(U1);
294 Compare = dyn_cast<FCmpInst>(U2);
295 if (Compare == 0 || !Compare->hasOneUse() ||
296 !isa<BranchInst>(Compare->use_back()))
299 BranchInst *TheBr = cast<BranchInst>(Compare->use_back());
301 // We need to verify that the branch actually controls the iteration count
302 // of the loop. If not, the new IV can overflow and no one will notice.
303 // The branch block must be in the loop and one of the successors must be out
305 assert(TheBr->isConditional() && "Can't use fcmp if not conditional");
306 if (!L->contains(TheBr->getParent()) ||
307 (L->contains(TheBr->getSuccessor(0)) &&
308 L->contains(TheBr->getSuccessor(1))))
312 // If it isn't a comparison with an integer-as-fp (the exit value), we can't
314 ConstantFP *ExitValueVal = dyn_cast<ConstantFP>(Compare->getOperand(1));
316 if (ExitValueVal == 0 ||
317 !ConvertToSInt(ExitValueVal->getValueAPF(), ExitValue))
320 // Find new predicate for integer comparison.
321 CmpInst::Predicate NewPred = CmpInst::BAD_ICMP_PREDICATE;
322 switch (Compare->getPredicate()) {
323 default: return; // Unknown comparison.
324 case CmpInst::FCMP_OEQ:
325 case CmpInst::FCMP_UEQ: NewPred = CmpInst::ICMP_EQ; break;
326 case CmpInst::FCMP_ONE:
327 case CmpInst::FCMP_UNE: NewPred = CmpInst::ICMP_NE; break;
328 case CmpInst::FCMP_OGT:
329 case CmpInst::FCMP_UGT: NewPred = CmpInst::ICMP_SGT; break;
330 case CmpInst::FCMP_OGE:
331 case CmpInst::FCMP_UGE: NewPred = CmpInst::ICMP_SGE; break;
332 case CmpInst::FCMP_OLT:
333 case CmpInst::FCMP_ULT: NewPred = CmpInst::ICMP_SLT; break;
334 case CmpInst::FCMP_OLE:
335 case CmpInst::FCMP_ULE: NewPred = CmpInst::ICMP_SLE; break;
338 // We convert the floating point induction variable to a signed i32 value if
339 // we can. This is only safe if the comparison will not overflow in a way
340 // that won't be trapped by the integer equivalent operations. Check for this
342 // TODO: We could use i64 if it is native and the range requires it.
344 // The start/stride/exit values must all fit in signed i32.
345 if (!isInt<32>(InitValue) || !isInt<32>(IncValue) || !isInt<32>(ExitValue))
348 // If not actually striding (add x, 0.0), avoid touching the code.
352 // Positive and negative strides have different safety conditions.
354 // If we have a positive stride, we require the init to be less than the
356 if (InitValue >= ExitValue)
359 uint32_t Range = uint32_t(ExitValue-InitValue);
360 // Check for infinite loop, either:
361 // while (i <= Exit) or until (i > Exit)
362 if (NewPred == CmpInst::ICMP_SLE || NewPred == CmpInst::ICMP_SGT) {
363 if (++Range == 0) return; // Range overflows.
366 unsigned Leftover = Range % uint32_t(IncValue);
368 // If this is an equality comparison, we require that the strided value
369 // exactly land on the exit value, otherwise the IV condition will wrap
370 // around and do things the fp IV wouldn't.
371 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
375 // If the stride would wrap around the i32 before exiting, we can't
377 if (Leftover != 0 && int32_t(ExitValue+IncValue) < ExitValue)
381 // If we have a negative stride, we require the init to be greater than the
383 if (InitValue <= ExitValue)
386 uint32_t Range = uint32_t(InitValue-ExitValue);
387 // Check for infinite loop, either:
388 // while (i >= Exit) or until (i < Exit)
389 if (NewPred == CmpInst::ICMP_SGE || NewPred == CmpInst::ICMP_SLT) {
390 if (++Range == 0) return; // Range overflows.
393 unsigned Leftover = Range % uint32_t(-IncValue);
395 // If this is an equality comparison, we require that the strided value
396 // exactly land on the exit value, otherwise the IV condition will wrap
397 // around and do things the fp IV wouldn't.
398 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
402 // If the stride would wrap around the i32 before exiting, we can't
404 if (Leftover != 0 && int32_t(ExitValue+IncValue) > ExitValue)
408 IntegerType *Int32Ty = Type::getInt32Ty(PN->getContext());
410 // Insert new integer induction variable.
411 PHINode *NewPHI = PHINode::Create(Int32Ty, 2, PN->getName()+".int", PN);
412 NewPHI->addIncoming(ConstantInt::get(Int32Ty, InitValue),
413 PN->getIncomingBlock(IncomingEdge));
416 BinaryOperator::CreateAdd(NewPHI, ConstantInt::get(Int32Ty, IncValue),
417 Incr->getName()+".int", Incr);
418 NewPHI->addIncoming(NewAdd, PN->getIncomingBlock(BackEdge));
420 ICmpInst *NewCompare = new ICmpInst(TheBr, NewPred, NewAdd,
421 ConstantInt::get(Int32Ty, ExitValue),
424 // In the following deletions, PN may become dead and may be deleted.
425 // Use a WeakVH to observe whether this happens.
428 // Delete the old floating point exit comparison. The branch starts using the
430 NewCompare->takeName(Compare);
431 Compare->replaceAllUsesWith(NewCompare);
432 RecursivelyDeleteTriviallyDeadInstructions(Compare);
434 // Delete the old floating point increment.
435 Incr->replaceAllUsesWith(UndefValue::get(Incr->getType()));
436 RecursivelyDeleteTriviallyDeadInstructions(Incr);
438 // If the FP induction variable still has uses, this is because something else
439 // in the loop uses its value. In order to canonicalize the induction
440 // variable, we chose to eliminate the IV and rewrite it in terms of an
443 // We give preference to sitofp over uitofp because it is faster on most
446 Value *Conv = new SIToFPInst(NewPHI, PN->getType(), "indvar.conv",
447 PN->getParent()->getFirstInsertionPt());
448 PN->replaceAllUsesWith(Conv);
449 RecursivelyDeleteTriviallyDeadInstructions(PN);
452 // Add a new IVUsers entry for the newly-created integer PHI.
454 IU->AddUsersIfInteresting(NewPHI);
459 void IndVarSimplify::RewriteNonIntegerIVs(Loop *L) {
460 // First step. Check to see if there are any floating-point recurrences.
461 // If there are, change them into integer recurrences, permitting analysis by
462 // the SCEV routines.
464 BasicBlock *Header = L->getHeader();
466 SmallVector<WeakVH, 8> PHIs;
467 for (BasicBlock::iterator I = Header->begin();
468 PHINode *PN = dyn_cast<PHINode>(I); ++I)
471 for (unsigned i = 0, e = PHIs.size(); i != e; ++i)
472 if (PHINode *PN = dyn_cast_or_null<PHINode>(&*PHIs[i]))
473 HandleFloatingPointIV(L, PN);
475 // If the loop previously had floating-point IV, ScalarEvolution
476 // may not have been able to compute a trip count. Now that we've done some
477 // re-writing, the trip count may be computable.
482 //===----------------------------------------------------------------------===//
483 // RewriteLoopExitValues - Optimize IV users outside the loop.
484 // As a side effect, reduces the amount of IV processing within the loop.
485 //===----------------------------------------------------------------------===//
487 /// RewriteLoopExitValues - Check to see if this loop has a computable
488 /// loop-invariant execution count. If so, this means that we can compute the
489 /// final value of any expressions that are recurrent in the loop, and
490 /// substitute the exit values from the loop into any instructions outside of
491 /// the loop that use the final values of the current expressions.
493 /// This is mostly redundant with the regular IndVarSimplify activities that
494 /// happen later, except that it's more powerful in some cases, because it's
495 /// able to brute-force evaluate arbitrary instructions as long as they have
496 /// constant operands at the beginning of the loop.
497 void IndVarSimplify::RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter) {
498 // Verify the input to the pass in already in LCSSA form.
499 assert(L->isLCSSAForm(*DT));
501 SmallVector<BasicBlock*, 8> ExitBlocks;
502 L->getUniqueExitBlocks(ExitBlocks);
504 // Find all values that are computed inside the loop, but used outside of it.
505 // Because of LCSSA, these values will only occur in LCSSA PHI Nodes. Scan
506 // the exit blocks of the loop to find them.
507 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
508 BasicBlock *ExitBB = ExitBlocks[i];
510 // If there are no PHI nodes in this exit block, then no values defined
511 // inside the loop are used on this path, skip it.
512 PHINode *PN = dyn_cast<PHINode>(ExitBB->begin());
515 unsigned NumPreds = PN->getNumIncomingValues();
517 // Iterate over all of the PHI nodes.
518 BasicBlock::iterator BBI = ExitBB->begin();
519 while ((PN = dyn_cast<PHINode>(BBI++))) {
521 continue; // dead use, don't replace it
523 // SCEV only supports integer expressions for now.
524 if (!PN->getType()->isIntegerTy() && !PN->getType()->isPointerTy())
527 // It's necessary to tell ScalarEvolution about this explicitly so that
528 // it can walk the def-use list and forget all SCEVs, as it may not be
529 // watching the PHI itself. Once the new exit value is in place, there
530 // may not be a def-use connection between the loop and every instruction
531 // which got a SCEVAddRecExpr for that loop.
534 // Iterate over all of the values in all the PHI nodes.
535 for (unsigned i = 0; i != NumPreds; ++i) {
536 // If the value being merged in is not integer or is not defined
537 // in the loop, skip it.
538 Value *InVal = PN->getIncomingValue(i);
539 if (!isa<Instruction>(InVal))
542 // If this pred is for a subloop, not L itself, skip it.
543 if (LI->getLoopFor(PN->getIncomingBlock(i)) != L)
544 continue; // The Block is in a subloop, skip it.
546 // Check that InVal is defined in the loop.
547 Instruction *Inst = cast<Instruction>(InVal);
548 if (!L->contains(Inst))
551 // Okay, this instruction has a user outside of the current loop
552 // and varies predictably *inside* the loop. Evaluate the value it
553 // contains when the loop exits, if possible.
554 const SCEV *ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop());
555 if (!SE->isLoopInvariant(ExitValue, L))
558 Value *ExitVal = Rewriter.expandCodeFor(ExitValue, PN->getType(), Inst);
560 DEBUG(dbgs() << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal << '\n'
561 << " LoopVal = " << *Inst << "\n");
563 if (!isValidRewrite(Inst, ExitVal)) {
564 DeadInsts.push_back(ExitVal);
570 PN->setIncomingValue(i, ExitVal);
572 // If this instruction is dead now, delete it.
573 RecursivelyDeleteTriviallyDeadInstructions(Inst);
576 // Completely replace a single-pred PHI. This is safe, because the
577 // NewVal won't be variant in the loop, so we don't need an LCSSA phi
579 PN->replaceAllUsesWith(ExitVal);
580 RecursivelyDeleteTriviallyDeadInstructions(PN);
584 // Clone the PHI and delete the original one. This lets IVUsers and
585 // any other maps purge the original user from their records.
586 PHINode *NewPN = cast<PHINode>(PN->clone());
588 NewPN->insertBefore(PN);
589 PN->replaceAllUsesWith(NewPN);
590 PN->eraseFromParent();
595 // The insertion point instruction may have been deleted; clear it out
596 // so that the rewriter doesn't trip over it later.
597 Rewriter.clearInsertPoint();
600 //===----------------------------------------------------------------------===//
601 // Rewrite IV users based on a canonical IV.
602 // Only for use with -enable-iv-rewrite.
603 //===----------------------------------------------------------------------===//
605 /// FIXME: It is an extremely bad idea to indvar substitute anything more
606 /// complex than affine induction variables. Doing so will put expensive
607 /// polynomial evaluations inside of the loop, and the str reduction pass
608 /// currently can only reduce affine polynomials. For now just disable
609 /// indvar subst on anything more complex than an affine addrec, unless
610 /// it can be expanded to a trivial value.
611 static bool isSafe(const SCEV *S, const Loop *L, ScalarEvolution *SE) {
612 // Loop-invariant values are safe.
613 if (SE->isLoopInvariant(S, L)) return true;
615 // Affine addrecs are safe. Non-affine are not, because LSR doesn't know how
616 // to transform them into efficient code.
617 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
618 return AR->isAffine();
620 // An add is safe it all its operands are safe.
621 if (const SCEVCommutativeExpr *Commutative
622 = dyn_cast<SCEVCommutativeExpr>(S)) {
623 for (SCEVCommutativeExpr::op_iterator I = Commutative->op_begin(),
624 E = Commutative->op_end(); I != E; ++I)
625 if (!isSafe(*I, L, SE)) return false;
629 // A cast is safe if its operand is.
630 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
631 return isSafe(C->getOperand(), L, SE);
633 // A udiv is safe if its operands are.
634 if (const SCEVUDivExpr *UD = dyn_cast<SCEVUDivExpr>(S))
635 return isSafe(UD->getLHS(), L, SE) &&
636 isSafe(UD->getRHS(), L, SE);
638 // SCEVUnknown is always safe.
639 if (isa<SCEVUnknown>(S))
642 // Nothing else is safe.
646 void IndVarSimplify::RewriteIVExpressions(Loop *L, SCEVExpander &Rewriter) {
647 // Rewrite all induction variable expressions in terms of the canonical
648 // induction variable.
650 // If there were induction variables of other sizes or offsets, manually
651 // add the offsets to the primary induction variable and cast, avoiding
652 // the need for the code evaluation methods to insert induction variables
653 // of different sizes.
654 for (IVUsers::iterator UI = IU->begin(), E = IU->end(); UI != E; ++UI) {
655 Value *Op = UI->getOperandValToReplace();
656 Type *UseTy = Op->getType();
657 Instruction *User = UI->getUser();
659 // Compute the final addrec to expand into code.
660 const SCEV *AR = IU->getReplacementExpr(*UI);
662 // Evaluate the expression out of the loop, if possible.
663 if (!L->contains(UI->getUser())) {
664 const SCEV *ExitVal = SE->getSCEVAtScope(AR, L->getParentLoop());
665 if (SE->isLoopInvariant(ExitVal, L))
669 // FIXME: It is an extremely bad idea to indvar substitute anything more
670 // complex than affine induction variables. Doing so will put expensive
671 // polynomial evaluations inside of the loop, and the str reduction pass
672 // currently can only reduce affine polynomials. For now just disable
673 // indvar subst on anything more complex than an affine addrec, unless
674 // it can be expanded to a trivial value.
675 if (!isSafe(AR, L, SE))
678 // Determine the insertion point for this user. By default, insert
679 // immediately before the user. The SCEVExpander class will automatically
680 // hoist loop invariants out of the loop. For PHI nodes, there may be
681 // multiple uses, so compute the nearest common dominator for the
683 Instruction *InsertPt = getInsertPointForUses(User, Op, DT);
685 // Now expand it into actual Instructions and patch it into place.
686 Value *NewVal = Rewriter.expandCodeFor(AR, UseTy, InsertPt);
688 DEBUG(dbgs() << "INDVARS: Rewrote IV '" << *AR << "' " << *Op << '\n'
689 << " into = " << *NewVal << "\n");
691 if (!isValidRewrite(Op, NewVal)) {
692 DeadInsts.push_back(NewVal);
695 // Inform ScalarEvolution that this value is changing. The change doesn't
696 // affect its value, but it does potentially affect which use lists the
697 // value will be on after the replacement, which affects ScalarEvolution's
698 // ability to walk use lists and drop dangling pointers when a value is
700 SE->forgetValue(User);
702 // Patch the new value into place.
704 NewVal->takeName(Op);
705 if (Instruction *NewValI = dyn_cast<Instruction>(NewVal))
706 NewValI->setDebugLoc(User->getDebugLoc());
707 User->replaceUsesOfWith(Op, NewVal);
708 UI->setOperandValToReplace(NewVal);
713 // The old value may be dead now.
714 DeadInsts.push_back(Op);
718 //===----------------------------------------------------------------------===//
719 // IV Widening - Extend the width of an IV to cover its widest uses.
720 //===----------------------------------------------------------------------===//
723 // Collect information about induction variables that are used by sign/zero
724 // extend operations. This information is recorded by CollectExtend and
725 // provides the input to WidenIV.
728 Type *WidestNativeType; // Widest integer type created [sz]ext
729 bool IsSigned; // Was an sext user seen before a zext?
731 WideIVInfo() : NarrowIV(0), WidestNativeType(0), IsSigned(false) {}
734 class WideIVVisitor : public IVVisitor {
736 const TargetData *TD;
741 WideIVVisitor(PHINode *NarrowIV, ScalarEvolution *SCEV,
742 const TargetData *TData) :
743 SE(SCEV), TD(TData) { WI.NarrowIV = NarrowIV; }
745 // Implement the interface used by simplifyUsersOfIV.
746 virtual void visitCast(CastInst *Cast);
750 /// visitCast - Update information about the induction variable that is
751 /// extended by this sign or zero extend operation. This is used to determine
752 /// the final width of the IV before actually widening it.
753 void WideIVVisitor::visitCast(CastInst *Cast) {
754 bool IsSigned = Cast->getOpcode() == Instruction::SExt;
755 if (!IsSigned && Cast->getOpcode() != Instruction::ZExt)
758 Type *Ty = Cast->getType();
759 uint64_t Width = SE->getTypeSizeInBits(Ty);
760 if (TD && !TD->isLegalInteger(Width))
763 if (!WI.WidestNativeType) {
764 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
765 WI.IsSigned = IsSigned;
769 // We extend the IV to satisfy the sign of its first user, arbitrarily.
770 if (WI.IsSigned != IsSigned)
773 if (Width > SE->getTypeSizeInBits(WI.WidestNativeType))
774 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
779 /// NarrowIVDefUse - Record a link in the Narrow IV def-use chain along with the
780 /// WideIV that computes the same value as the Narrow IV def. This avoids
781 /// caching Use* pointers.
782 struct NarrowIVDefUse {
783 Instruction *NarrowDef;
784 Instruction *NarrowUse;
785 Instruction *WideDef;
787 NarrowIVDefUse(): NarrowDef(0), NarrowUse(0), WideDef(0) {}
789 NarrowIVDefUse(Instruction *ND, Instruction *NU, Instruction *WD):
790 NarrowDef(ND), NarrowUse(NU), WideDef(WD) {}
793 /// WidenIV - The goal of this transform is to remove sign and zero extends
794 /// without creating any new induction variables. To do this, it creates a new
795 /// phi of the wider type and redirects all users, either removing extends or
796 /// inserting truncs whenever we stop propagating the type.
812 Instruction *WideInc;
813 const SCEV *WideIncExpr;
814 SmallVectorImpl<WeakVH> &DeadInsts;
816 SmallPtrSet<Instruction*,16> Widened;
817 SmallVector<NarrowIVDefUse, 8> NarrowIVUsers;
820 WidenIV(const WideIVInfo &WI, LoopInfo *LInfo,
821 ScalarEvolution *SEv, DominatorTree *DTree,
822 SmallVectorImpl<WeakVH> &DI) :
823 OrigPhi(WI.NarrowIV),
824 WideType(WI.WidestNativeType),
825 IsSigned(WI.IsSigned),
827 L(LI->getLoopFor(OrigPhi->getParent())),
834 assert(L->getHeader() == OrigPhi->getParent() && "Phi must be an IV");
837 PHINode *CreateWideIV(SCEVExpander &Rewriter);
840 Value *getExtend(Value *NarrowOper, Type *WideType, bool IsSigned,
843 Instruction *CloneIVUser(NarrowIVDefUse DU);
845 const SCEVAddRecExpr *GetWideRecurrence(Instruction *NarrowUse);
847 const SCEVAddRecExpr* GetExtendedOperandRecurrence(NarrowIVDefUse DU);
849 Instruction *WidenIVUse(NarrowIVDefUse DU);
851 void pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef);
853 } // anonymous namespace
855 /// isLoopInvariant - Perform a quick domtree based check for loop invariance
856 /// assuming that V is used within the loop. LoopInfo::isLoopInvariant() seems
857 /// gratuitous for this purpose.
858 static bool isLoopInvariant(Value *V, const Loop *L, const DominatorTree *DT) {
859 Instruction *Inst = dyn_cast<Instruction>(V);
863 return DT->properlyDominates(Inst->getParent(), L->getHeader());
866 Value *WidenIV::getExtend(Value *NarrowOper, Type *WideType, bool IsSigned,
868 // Set the debug location and conservative insertion point.
869 IRBuilder<> Builder(Use);
870 // Hoist the insertion point into loop preheaders as far as possible.
871 for (const Loop *L = LI->getLoopFor(Use->getParent());
872 L && L->getLoopPreheader() && isLoopInvariant(NarrowOper, L, DT);
873 L = L->getParentLoop())
874 Builder.SetInsertPoint(L->getLoopPreheader()->getTerminator());
876 return IsSigned ? Builder.CreateSExt(NarrowOper, WideType) :
877 Builder.CreateZExt(NarrowOper, WideType);
880 /// CloneIVUser - Instantiate a wide operation to replace a narrow
881 /// operation. This only needs to handle operations that can evaluation to
882 /// SCEVAddRec. It can safely return 0 for any operation we decide not to clone.
883 Instruction *WidenIV::CloneIVUser(NarrowIVDefUse DU) {
884 unsigned Opcode = DU.NarrowUse->getOpcode();
888 case Instruction::Add:
889 case Instruction::Mul:
890 case Instruction::UDiv:
891 case Instruction::Sub:
892 case Instruction::And:
893 case Instruction::Or:
894 case Instruction::Xor:
895 case Instruction::Shl:
896 case Instruction::LShr:
897 case Instruction::AShr:
898 DEBUG(dbgs() << "Cloning IVUser: " << *DU.NarrowUse << "\n");
900 // Replace NarrowDef operands with WideDef. Otherwise, we don't know
901 // anything about the narrow operand yet so must insert a [sz]ext. It is
902 // probably loop invariant and will be folded or hoisted. If it actually
903 // comes from a widened IV, it should be removed during a future call to
905 Value *LHS = (DU.NarrowUse->getOperand(0) == DU.NarrowDef) ? DU.WideDef :
906 getExtend(DU.NarrowUse->getOperand(0), WideType, IsSigned, DU.NarrowUse);
907 Value *RHS = (DU.NarrowUse->getOperand(1) == DU.NarrowDef) ? DU.WideDef :
908 getExtend(DU.NarrowUse->getOperand(1), WideType, IsSigned, DU.NarrowUse);
910 BinaryOperator *NarrowBO = cast<BinaryOperator>(DU.NarrowUse);
911 BinaryOperator *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(),
913 NarrowBO->getName());
914 IRBuilder<> Builder(DU.NarrowUse);
915 Builder.Insert(WideBO);
916 if (const OverflowingBinaryOperator *OBO =
917 dyn_cast<OverflowingBinaryOperator>(NarrowBO)) {
918 if (OBO->hasNoUnsignedWrap()) WideBO->setHasNoUnsignedWrap();
919 if (OBO->hasNoSignedWrap()) WideBO->setHasNoSignedWrap();
926 /// No-wrap operations can transfer sign extension of their result to their
927 /// operands. Generate the SCEV value for the widened operation without
928 /// actually modifying the IR yet. If the expression after extending the
929 /// operands is an AddRec for this loop, return it.
930 const SCEVAddRecExpr* WidenIV::GetExtendedOperandRecurrence(NarrowIVDefUse DU) {
931 // Handle the common case of add<nsw/nuw>
932 if (DU.NarrowUse->getOpcode() != Instruction::Add)
935 // One operand (NarrowDef) has already been extended to WideDef. Now determine
936 // if extending the other will lead to a recurrence.
937 unsigned ExtendOperIdx = DU.NarrowUse->getOperand(0) == DU.NarrowDef ? 1 : 0;
938 assert(DU.NarrowUse->getOperand(1-ExtendOperIdx) == DU.NarrowDef && "bad DU");
940 const SCEV *ExtendOperExpr = 0;
941 const OverflowingBinaryOperator *OBO =
942 cast<OverflowingBinaryOperator>(DU.NarrowUse);
943 if (IsSigned && OBO->hasNoSignedWrap())
944 ExtendOperExpr = SE->getSignExtendExpr(
945 SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType);
946 else if(!IsSigned && OBO->hasNoUnsignedWrap())
947 ExtendOperExpr = SE->getZeroExtendExpr(
948 SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType);
952 // When creating this AddExpr, don't apply the current operations NSW or NUW
953 // flags. This instruction may be guarded by control flow that the no-wrap
954 // behavior depends on. Non-control-equivalent instructions can be mapped to
955 // the same SCEV expression, and it would be incorrect to transfer NSW/NUW
956 // semantics to those operations.
957 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(
958 SE->getAddExpr(SE->getSCEV(DU.WideDef), ExtendOperExpr));
960 if (!AddRec || AddRec->getLoop() != L)
965 /// GetWideRecurrence - Is this instruction potentially interesting from
966 /// IVUsers' perspective after widening it's type? In other words, can the
967 /// extend be safely hoisted out of the loop with SCEV reducing the value to a
968 /// recurrence on the same loop. If so, return the sign or zero extended
969 /// recurrence. Otherwise return NULL.
970 const SCEVAddRecExpr *WidenIV::GetWideRecurrence(Instruction *NarrowUse) {
971 if (!SE->isSCEVable(NarrowUse->getType()))
974 const SCEV *NarrowExpr = SE->getSCEV(NarrowUse);
975 if (SE->getTypeSizeInBits(NarrowExpr->getType())
976 >= SE->getTypeSizeInBits(WideType)) {
977 // NarrowUse implicitly widens its operand. e.g. a gep with a narrow
978 // index. So don't follow this use.
982 const SCEV *WideExpr = IsSigned ?
983 SE->getSignExtendExpr(NarrowExpr, WideType) :
984 SE->getZeroExtendExpr(NarrowExpr, WideType);
985 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(WideExpr);
986 if (!AddRec || AddRec->getLoop() != L)
991 /// WidenIVUse - Determine whether an individual user of the narrow IV can be
992 /// widened. If so, return the wide clone of the user.
993 Instruction *WidenIV::WidenIVUse(NarrowIVDefUse DU) {
995 // Stop traversing the def-use chain at inner-loop phis or post-loop phis.
996 if (isa<PHINode>(DU.NarrowUse) &&
997 LI->getLoopFor(DU.NarrowUse->getParent()) != L)
1000 // Our raison d'etre! Eliminate sign and zero extension.
1001 if (IsSigned ? isa<SExtInst>(DU.NarrowUse) : isa<ZExtInst>(DU.NarrowUse)) {
1002 Value *NewDef = DU.WideDef;
1003 if (DU.NarrowUse->getType() != WideType) {
1004 unsigned CastWidth = SE->getTypeSizeInBits(DU.NarrowUse->getType());
1005 unsigned IVWidth = SE->getTypeSizeInBits(WideType);
1006 if (CastWidth < IVWidth) {
1007 // The cast isn't as wide as the IV, so insert a Trunc.
1008 IRBuilder<> Builder(DU.NarrowUse);
1009 NewDef = Builder.CreateTrunc(DU.WideDef, DU.NarrowUse->getType());
1012 // A wider extend was hidden behind a narrower one. This may induce
1013 // another round of IV widening in which the intermediate IV becomes
1014 // dead. It should be very rare.
1015 DEBUG(dbgs() << "INDVARS: New IV " << *WidePhi
1016 << " not wide enough to subsume " << *DU.NarrowUse << "\n");
1017 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef);
1018 NewDef = DU.NarrowUse;
1021 if (NewDef != DU.NarrowUse) {
1022 DEBUG(dbgs() << "INDVARS: eliminating " << *DU.NarrowUse
1023 << " replaced by " << *DU.WideDef << "\n");
1025 DU.NarrowUse->replaceAllUsesWith(NewDef);
1026 DeadInsts.push_back(DU.NarrowUse);
1028 // Now that the extend is gone, we want to expose it's uses for potential
1029 // further simplification. We don't need to directly inform SimplifyIVUsers
1030 // of the new users, because their parent IV will be processed later as a
1031 // new loop phi. If we preserved IVUsers analysis, we would also want to
1032 // push the uses of WideDef here.
1034 // No further widening is needed. The deceased [sz]ext had done it for us.
1038 // Does this user itself evaluate to a recurrence after widening?
1039 const SCEVAddRecExpr *WideAddRec = GetWideRecurrence(DU.NarrowUse);
1041 WideAddRec = GetExtendedOperandRecurrence(DU);
1044 // This user does not evaluate to a recurence after widening, so don't
1045 // follow it. Instead insert a Trunc to kill off the original use,
1046 // eventually isolating the original narrow IV so it can be removed.
1047 IRBuilder<> Builder(getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT));
1048 Value *Trunc = Builder.CreateTrunc(DU.WideDef, DU.NarrowDef->getType());
1049 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, Trunc);
1052 // Assume block terminators cannot evaluate to a recurrence. We can't to
1053 // insert a Trunc after a terminator if there happens to be a critical edge.
1054 assert(DU.NarrowUse != DU.NarrowUse->getParent()->getTerminator() &&
1055 "SCEV is not expected to evaluate a block terminator");
1057 // Reuse the IV increment that SCEVExpander created as long as it dominates
1059 Instruction *WideUse = 0;
1060 if (WideAddRec == WideIncExpr
1061 && SCEVExpander::hoistStep(WideInc, DU.NarrowUse, DT))
1064 WideUse = CloneIVUser(DU);
1068 // Evaluation of WideAddRec ensured that the narrow expression could be
1069 // extended outside the loop without overflow. This suggests that the wide use
1070 // evaluates to the same expression as the extended narrow use, but doesn't
1071 // absolutely guarantee it. Hence the following failsafe check. In rare cases
1072 // where it fails, we simply throw away the newly created wide use.
1073 if (WideAddRec != SE->getSCEV(WideUse)) {
1074 DEBUG(dbgs() << "Wide use expression mismatch: " << *WideUse
1075 << ": " << *SE->getSCEV(WideUse) << " != " << *WideAddRec << "\n");
1076 DeadInsts.push_back(WideUse);
1080 // Returning WideUse pushes it on the worklist.
1084 /// pushNarrowIVUsers - Add eligible users of NarrowDef to NarrowIVUsers.
1086 void WidenIV::pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef) {
1087 for (Value::use_iterator UI = NarrowDef->use_begin(),
1088 UE = NarrowDef->use_end(); UI != UE; ++UI) {
1089 Instruction *NarrowUse = cast<Instruction>(*UI);
1091 // Handle data flow merges and bizarre phi cycles.
1092 if (!Widened.insert(NarrowUse))
1095 NarrowIVUsers.push_back(NarrowIVDefUse(NarrowDef, NarrowUse, WideDef));
1099 /// CreateWideIV - Process a single induction variable. First use the
1100 /// SCEVExpander to create a wide induction variable that evaluates to the same
1101 /// recurrence as the original narrow IV. Then use a worklist to forward
1102 /// traverse the narrow IV's def-use chain. After WidenIVUse has processed all
1103 /// interesting IV users, the narrow IV will be isolated for removal by
1106 /// It would be simpler to delete uses as they are processed, but we must avoid
1107 /// invalidating SCEV expressions.
1109 PHINode *WidenIV::CreateWideIV(SCEVExpander &Rewriter) {
1110 // Is this phi an induction variable?
1111 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(OrigPhi));
1115 // Widen the induction variable expression.
1116 const SCEV *WideIVExpr = IsSigned ?
1117 SE->getSignExtendExpr(AddRec, WideType) :
1118 SE->getZeroExtendExpr(AddRec, WideType);
1120 assert(SE->getEffectiveSCEVType(WideIVExpr->getType()) == WideType &&
1121 "Expect the new IV expression to preserve its type");
1123 // Can the IV be extended outside the loop without overflow?
1124 AddRec = dyn_cast<SCEVAddRecExpr>(WideIVExpr);
1125 if (!AddRec || AddRec->getLoop() != L)
1128 // An AddRec must have loop-invariant operands. Since this AddRec is
1129 // materialized by a loop header phi, the expression cannot have any post-loop
1130 // operands, so they must dominate the loop header.
1131 assert(SE->properlyDominates(AddRec->getStart(), L->getHeader()) &&
1132 SE->properlyDominates(AddRec->getStepRecurrence(*SE), L->getHeader())
1133 && "Loop header phi recurrence inputs do not dominate the loop");
1135 // The rewriter provides a value for the desired IV expression. This may
1136 // either find an existing phi or materialize a new one. Either way, we
1137 // expect a well-formed cyclic phi-with-increments. i.e. any operand not part
1138 // of the phi-SCC dominates the loop entry.
1139 Instruction *InsertPt = L->getHeader()->begin();
1140 WidePhi = cast<PHINode>(Rewriter.expandCodeFor(AddRec, WideType, InsertPt));
1142 // Remembering the WideIV increment generated by SCEVExpander allows
1143 // WidenIVUse to reuse it when widening the narrow IV's increment. We don't
1144 // employ a general reuse mechanism because the call above is the only call to
1145 // SCEVExpander. Henceforth, we produce 1-to-1 narrow to wide uses.
1146 if (BasicBlock *LatchBlock = L->getLoopLatch()) {
1148 cast<Instruction>(WidePhi->getIncomingValueForBlock(LatchBlock));
1149 WideIncExpr = SE->getSCEV(WideInc);
1152 DEBUG(dbgs() << "Wide IV: " << *WidePhi << "\n");
1155 // Traverse the def-use chain using a worklist starting at the original IV.
1156 assert(Widened.empty() && NarrowIVUsers.empty() && "expect initial state" );
1158 Widened.insert(OrigPhi);
1159 pushNarrowIVUsers(OrigPhi, WidePhi);
1161 while (!NarrowIVUsers.empty()) {
1162 NarrowIVDefUse DU = NarrowIVUsers.pop_back_val();
1164 // Process a def-use edge. This may replace the use, so don't hold a
1165 // use_iterator across it.
1166 Instruction *WideUse = WidenIVUse(DU);
1168 // Follow all def-use edges from the previous narrow use.
1170 pushNarrowIVUsers(DU.NarrowUse, WideUse);
1172 // WidenIVUse may have removed the def-use edge.
1173 if (DU.NarrowDef->use_empty())
1174 DeadInsts.push_back(DU.NarrowDef);
1179 //===----------------------------------------------------------------------===//
1180 // Simplification of IV users based on SCEV evaluation.
1181 //===----------------------------------------------------------------------===//
1184 /// SimplifyAndExtend - Iteratively perform simplification on a worklist of IV
1185 /// users. Each successive simplification may push more users which may
1186 /// themselves be candidates for simplification.
1188 /// Sign/Zero extend elimination is interleaved with IV simplification.
1190 void IndVarSimplify::SimplifyAndExtend(Loop *L,
1191 SCEVExpander &Rewriter,
1192 LPPassManager &LPM) {
1193 SmallVector<WideIVInfo, 8> WideIVs;
1195 SmallVector<PHINode*, 8> LoopPhis;
1196 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
1197 LoopPhis.push_back(cast<PHINode>(I));
1199 // Each round of simplification iterates through the SimplifyIVUsers worklist
1200 // for all current phis, then determines whether any IVs can be
1201 // widened. Widening adds new phis to LoopPhis, inducing another round of
1202 // simplification on the wide IVs.
1203 while (!LoopPhis.empty()) {
1204 // Evaluate as many IV expressions as possible before widening any IVs. This
1205 // forces SCEV to set no-wrap flags before evaluating sign/zero
1206 // extension. The first time SCEV attempts to normalize sign/zero extension,
1207 // the result becomes final. So for the most predictable results, we delay
1208 // evaluation of sign/zero extend evaluation until needed, and avoid running
1209 // other SCEV based analysis prior to SimplifyAndExtend.
1211 PHINode *CurrIV = LoopPhis.pop_back_val();
1213 // Information about sign/zero extensions of CurrIV.
1214 WideIVVisitor WIV(CurrIV, SE, TD);
1216 Changed |= simplifyUsersOfIV(CurrIV, SE, &LPM, DeadInsts, &WIV);
1218 if (WIV.WI.WidestNativeType) {
1219 WideIVs.push_back(WIV.WI);
1221 } while(!LoopPhis.empty());
1223 for (; !WideIVs.empty(); WideIVs.pop_back()) {
1224 WidenIV Widener(WideIVs.back(), LI, SE, DT, DeadInsts);
1225 if (PHINode *WidePhi = Widener.CreateWideIV(Rewriter)) {
1227 LoopPhis.push_back(WidePhi);
1233 //===----------------------------------------------------------------------===//
1234 // LinearFunctionTestReplace and its kin. Rewrite the loop exit condition.
1235 //===----------------------------------------------------------------------===//
1237 /// Check for expressions that ScalarEvolution generates to compute
1238 /// BackedgeTakenInfo. If these expressions have not been reduced, then
1239 /// expanding them may incur additional cost (albeit in the loop preheader).
1240 static bool isHighCostExpansion(const SCEV *S, BranchInst *BI,
1241 ScalarEvolution *SE) {
1242 // If the backedge-taken count is a UDiv, it's very likely a UDiv that
1243 // ScalarEvolution's HowFarToZero or HowManyLessThans produced to compute a
1244 // precise expression, rather than a UDiv from the user's code. If we can't
1245 // find a UDiv in the code with some simple searching, assume the former and
1246 // forego rewriting the loop.
1247 if (isa<SCEVUDivExpr>(S)) {
1248 ICmpInst *OrigCond = dyn_cast<ICmpInst>(BI->getCondition());
1249 if (!OrigCond) return true;
1250 const SCEV *R = SE->getSCEV(OrigCond->getOperand(1));
1251 R = SE->getMinusSCEV(R, SE->getConstant(R->getType(), 1));
1253 const SCEV *L = SE->getSCEV(OrigCond->getOperand(0));
1254 L = SE->getMinusSCEV(L, SE->getConstant(L->getType(), 1));
1260 if (EnableIVRewrite)
1263 // Recurse past add expressions, which commonly occur in the
1264 // BackedgeTakenCount. They may already exist in program code, and if not,
1265 // they are not too expensive rematerialize.
1266 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
1267 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
1269 if (isHighCostExpansion(*I, BI, SE))
1275 // HowManyLessThans uses a Max expression whenever the loop is not guarded by
1276 // the exit condition.
1277 if (isa<SCEVSMaxExpr>(S) || isa<SCEVUMaxExpr>(S))
1280 // If we haven't recognized an expensive SCEV patter, assume its an expression
1281 // produced by program code.
1285 /// canExpandBackedgeTakenCount - Return true if this loop's backedge taken
1286 /// count expression can be safely and cheaply expanded into an instruction
1287 /// sequence that can be used by LinearFunctionTestReplace.
1289 /// TODO: This fails for pointer-type loop counters with greater than one byte
1290 /// strides, consequently preventing LFTR from running. For the purpose of LFTR
1291 /// we could skip this check in the case that the LFTR loop counter (chosen by
1292 /// FindLoopCounter) is also pointer type. Instead, we could directly convert
1293 /// the loop test to an inequality test by checking the target data's alignment
1294 /// of element types (given that the initial pointer value originates from or is
1295 /// used by ABI constrained operation, as opposed to inttoptr/ptrtoint).
1296 /// However, we don't yet have a strong motivation for converting loop tests
1297 /// into inequality tests.
1298 static bool canExpandBackedgeTakenCount(Loop *L, ScalarEvolution *SE) {
1299 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
1300 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount) ||
1301 BackedgeTakenCount->isZero())
1304 if (!L->getExitingBlock())
1307 // Can't rewrite non-branch yet.
1308 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
1312 if (isHighCostExpansion(BackedgeTakenCount, BI, SE))
1318 /// getBackedgeIVType - Get the widest type used by the loop test after peeking
1321 /// TODO: Unnecessary when ForceLFTR is removed.
1322 static Type *getBackedgeIVType(Loop *L) {
1323 if (!L->getExitingBlock())
1326 // Can't rewrite non-branch yet.
1327 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
1331 ICmpInst *Cond = dyn_cast<ICmpInst>(BI->getCondition());
1336 for(User::op_iterator OI = Cond->op_begin(), OE = Cond->op_end();
1338 assert((!Ty || Ty == (*OI)->getType()) && "bad icmp operand types");
1339 TruncInst *Trunc = dyn_cast<TruncInst>(*OI);
1343 return Trunc->getSrcTy();
1348 /// getLoopPhiForCounter - Return the loop header phi IFF IncV adds a loop
1349 /// invariant value to the phi.
1350 static PHINode *getLoopPhiForCounter(Value *IncV, Loop *L, DominatorTree *DT) {
1351 Instruction *IncI = dyn_cast<Instruction>(IncV);
1355 switch (IncI->getOpcode()) {
1356 case Instruction::Add:
1357 case Instruction::Sub:
1359 case Instruction::GetElementPtr:
1360 // An IV counter must preserve its type.
1361 if (IncI->getNumOperands() == 2)
1367 PHINode *Phi = dyn_cast<PHINode>(IncI->getOperand(0));
1368 if (Phi && Phi->getParent() == L->getHeader()) {
1369 if (isLoopInvariant(IncI->getOperand(1), L, DT))
1373 if (IncI->getOpcode() == Instruction::GetElementPtr)
1376 // Allow add/sub to be commuted.
1377 Phi = dyn_cast<PHINode>(IncI->getOperand(1));
1378 if (Phi && Phi->getParent() == L->getHeader()) {
1379 if (isLoopInvariant(IncI->getOperand(0), L, DT))
1385 /// needsLFTR - LinearFunctionTestReplace policy. Return true unless we can show
1386 /// that the current exit test is already sufficiently canonical.
1387 static bool needsLFTR(Loop *L, DominatorTree *DT) {
1388 assert(L->getExitingBlock() && "expected loop exit");
1390 BasicBlock *LatchBlock = L->getLoopLatch();
1391 // Don't bother with LFTR if the loop is not properly simplified.
1395 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
1396 assert(BI && "expected exit branch");
1398 // Do LFTR to simplify the exit condition to an ICMP.
1399 ICmpInst *Cond = dyn_cast<ICmpInst>(BI->getCondition());
1403 // Do LFTR to simplify the exit ICMP to EQ/NE
1404 ICmpInst::Predicate Pred = Cond->getPredicate();
1405 if (Pred != ICmpInst::ICMP_NE && Pred != ICmpInst::ICMP_EQ)
1408 // Look for a loop invariant RHS
1409 Value *LHS = Cond->getOperand(0);
1410 Value *RHS = Cond->getOperand(1);
1411 if (!isLoopInvariant(RHS, L, DT)) {
1412 if (!isLoopInvariant(LHS, L, DT))
1414 std::swap(LHS, RHS);
1416 // Look for a simple IV counter LHS
1417 PHINode *Phi = dyn_cast<PHINode>(LHS);
1419 Phi = getLoopPhiForCounter(LHS, L, DT);
1424 // Do LFTR if the exit condition's IV is *not* a simple counter.
1425 Value *IncV = Phi->getIncomingValueForBlock(L->getLoopLatch());
1426 return Phi != getLoopPhiForCounter(IncV, L, DT);
1429 /// AlmostDeadIV - Return true if this IV has any uses other than the (soon to
1430 /// be rewritten) loop exit test.
1431 static bool AlmostDeadIV(PHINode *Phi, BasicBlock *LatchBlock, Value *Cond) {
1432 int LatchIdx = Phi->getBasicBlockIndex(LatchBlock);
1433 Value *IncV = Phi->getIncomingValue(LatchIdx);
1435 for (Value::use_iterator UI = Phi->use_begin(), UE = Phi->use_end();
1437 if (*UI != Cond && *UI != IncV) return false;
1440 for (Value::use_iterator UI = IncV->use_begin(), UE = IncV->use_end();
1442 if (*UI != Cond && *UI != Phi) return false;
1447 /// FindLoopCounter - Find an affine IV in canonical form.
1449 /// BECount may be an i8* pointer type. The pointer difference is already
1450 /// valid count without scaling the address stride, so it remains a pointer
1451 /// expression as far as SCEV is concerned.
1453 /// FIXME: Accept -1 stride and set IVLimit = IVInit - BECount
1455 /// FIXME: Accept non-unit stride as long as SCEV can reduce BECount * Stride.
1456 /// This is difficult in general for SCEV because of potential overflow. But we
1457 /// could at least handle constant BECounts.
1459 FindLoopCounter(Loop *L, const SCEV *BECount,
1460 ScalarEvolution *SE, DominatorTree *DT, const TargetData *TD) {
1461 uint64_t BCWidth = SE->getTypeSizeInBits(BECount->getType());
1464 cast<BranchInst>(L->getExitingBlock()->getTerminator())->getCondition();
1466 // Loop over all of the PHI nodes, looking for a simple counter.
1467 PHINode *BestPhi = 0;
1468 const SCEV *BestInit = 0;
1469 BasicBlock *LatchBlock = L->getLoopLatch();
1470 assert(LatchBlock && "needsLFTR should guarantee a loop latch");
1472 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
1473 PHINode *Phi = cast<PHINode>(I);
1474 if (!SE->isSCEVable(Phi->getType()))
1477 // Avoid comparing an integer IV against a pointer Limit.
1478 if (BECount->getType()->isPointerTy() && !Phi->getType()->isPointerTy())
1481 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Phi));
1482 if (!AR || AR->getLoop() != L || !AR->isAffine())
1485 // AR may be a pointer type, while BECount is an integer type.
1486 // AR may be wider than BECount. With eq/ne tests overflow is immaterial.
1487 // AR may not be a narrower type, or we may never exit.
1488 uint64_t PhiWidth = SE->getTypeSizeInBits(AR->getType());
1489 if (PhiWidth < BCWidth || (TD && !TD->isLegalInteger(PhiWidth)))
1492 const SCEV *Step = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE));
1493 if (!Step || !Step->isOne())
1496 int LatchIdx = Phi->getBasicBlockIndex(LatchBlock);
1497 Value *IncV = Phi->getIncomingValue(LatchIdx);
1498 if (getLoopPhiForCounter(IncV, L, DT) != Phi)
1501 const SCEV *Init = AR->getStart();
1503 if (BestPhi && !AlmostDeadIV(BestPhi, LatchBlock, Cond)) {
1504 // Don't force a live loop counter if another IV can be used.
1505 if (AlmostDeadIV(Phi, LatchBlock, Cond))
1508 // Prefer to count-from-zero. This is a more "canonical" counter form. It
1509 // also prefers integer to pointer IVs.
1510 if (BestInit->isZero() != Init->isZero()) {
1511 if (BestInit->isZero())
1514 // If two IVs both count from zero or both count from nonzero then the
1515 // narrower is likely a dead phi that has been widened. Use the wider phi
1516 // to allow the other to be eliminated.
1517 if (PhiWidth <= SE->getTypeSizeInBits(BestPhi->getType()))
1526 /// genLoopLimit - Help LinearFunctionTestReplace by generating a value that
1527 /// holds the RHS of the new loop test.
1528 static Value *genLoopLimit(PHINode *IndVar, const SCEV *IVCount, Loop *L,
1529 SCEVExpander &Rewriter, ScalarEvolution *SE) {
1530 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(IndVar));
1531 assert(AR && AR->getLoop() == L && AR->isAffine() && "bad loop counter");
1532 const SCEV *IVInit = AR->getStart();
1534 // IVInit may be a pointer while IVCount is an integer when FindLoopCounter
1535 // finds a valid pointer IV. Sign extend BECount in order to materialize a
1536 // GEP. Avoid running SCEVExpander on a new pointer value, instead reusing
1537 // the existing GEPs whenever possible.
1538 if (IndVar->getType()->isPointerTy()
1539 && !IVCount->getType()->isPointerTy()) {
1541 Type *OfsTy = SE->getEffectiveSCEVType(IVInit->getType());
1542 const SCEV *IVOffset = SE->getTruncateOrSignExtend(IVCount, OfsTy);
1544 // Expand the code for the iteration count.
1545 assert(SE->isLoopInvariant(IVOffset, L) &&
1546 "Computed iteration count is not loop invariant!");
1547 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1548 Value *GEPOffset = Rewriter.expandCodeFor(IVOffset, OfsTy, BI);
1550 Value *GEPBase = IndVar->getIncomingValueForBlock(L->getLoopPreheader());
1551 assert(AR->getStart() == SE->getSCEV(GEPBase) && "bad loop counter");
1552 // We could handle pointer IVs other than i8*, but we need to compensate for
1553 // gep index scaling. See canExpandBackedgeTakenCount comments.
1554 assert(SE->getSizeOfExpr(
1555 cast<PointerType>(GEPBase->getType())->getElementType())->isOne()
1556 && "unit stride pointer IV must be i8*");
1558 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
1559 return Builder.CreateGEP(GEPBase, GEPOffset, "lftr.limit");
1562 // In any other case, convert both IVInit and IVCount to integers before
1563 // comparing. This may result in SCEV expension of pointers, but in practice
1564 // SCEV will fold the pointer arithmetic away as such:
1565 // BECount = (IVEnd - IVInit - 1) => IVLimit = IVInit (postinc).
1567 // Valid Cases: (1) both integers is most common; (2) both may be pointers
1568 // for simple memset-style loops; (3) IVInit is an integer and IVCount is a
1569 // pointer may occur when enable-iv-rewrite generates a canonical IV on top
1572 const SCEV *IVLimit = 0;
1573 // For unit stride, IVCount = Start + BECount with 2's complement overflow.
1574 // For non-zero Start, compute IVCount here.
1575 if (AR->getStart()->isZero())
1578 assert(AR->getStepRecurrence(*SE)->isOne() && "only handles unit stride");
1579 const SCEV *IVInit = AR->getStart();
1581 // For integer IVs, truncate the IV before computing IVInit + BECount.
1582 if (SE->getTypeSizeInBits(IVInit->getType())
1583 > SE->getTypeSizeInBits(IVCount->getType()))
1584 IVInit = SE->getTruncateExpr(IVInit, IVCount->getType());
1586 IVLimit = SE->getAddExpr(IVInit, IVCount);
1588 // Expand the code for the iteration count.
1589 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1590 IRBuilder<> Builder(BI);
1591 assert(SE->isLoopInvariant(IVLimit, L) &&
1592 "Computed iteration count is not loop invariant!");
1593 // Ensure that we generate the same type as IndVar, or a smaller integer
1594 // type. In the presence of null pointer values, we have an integer type
1595 // SCEV expression (IVInit) for a pointer type IV value (IndVar).
1596 Type *LimitTy = IVCount->getType()->isPointerTy() ?
1597 IndVar->getType() : IVCount->getType();
1598 return Rewriter.expandCodeFor(IVLimit, LimitTy, BI);
1602 /// LinearFunctionTestReplace - This method rewrites the exit condition of the
1603 /// loop to be a canonical != comparison against the incremented loop induction
1604 /// variable. This pass is able to rewrite the exit tests of any loop where the
1605 /// SCEV analysis can determine a loop-invariant trip count of the loop, which
1606 /// is actually a much broader range than just linear tests.
1607 Value *IndVarSimplify::
1608 LinearFunctionTestReplace(Loop *L,
1609 const SCEV *BackedgeTakenCount,
1611 SCEVExpander &Rewriter) {
1612 assert(canExpandBackedgeTakenCount(L, SE) && "precondition");
1614 // LFTR can ignore IV overflow and truncate to the width of
1615 // BECount. This avoids materializing the add(zext(add)) expression.
1616 Type *CntTy = !EnableIVRewrite ?
1617 BackedgeTakenCount->getType() : IndVar->getType();
1619 const SCEV *IVCount = BackedgeTakenCount;
1621 // If the exiting block is the same as the backedge block, we prefer to
1622 // compare against the post-incremented value, otherwise we must compare
1623 // against the preincremented value.
1625 if (L->getExitingBlock() == L->getLoopLatch()) {
1626 // Add one to the "backedge-taken" count to get the trip count.
1627 // If this addition may overflow, we have to be more pessimistic and
1628 // cast the induction variable before doing the add.
1630 SE->getAddExpr(IVCount, SE->getConstant(IVCount->getType(), 1));
1631 if (CntTy == IVCount->getType())
1634 const SCEV *Zero = SE->getConstant(IVCount->getType(), 0);
1635 if ((isa<SCEVConstant>(N) && !N->isZero()) ||
1636 SE->isLoopEntryGuardedByCond(L, ICmpInst::ICMP_NE, N, Zero)) {
1637 // No overflow. Cast the sum.
1638 IVCount = SE->getTruncateOrZeroExtend(N, CntTy);
1640 // Potential overflow. Cast before doing the add.
1641 IVCount = SE->getTruncateOrZeroExtend(IVCount, CntTy);
1642 IVCount = SE->getAddExpr(IVCount, SE->getConstant(CntTy, 1));
1645 // The BackedgeTaken expression contains the number of times that the
1646 // backedge branches to the loop header. This is one less than the
1647 // number of times the loop executes, so use the incremented indvar.
1648 CmpIndVar = IndVar->getIncomingValueForBlock(L->getExitingBlock());
1650 // We must use the preincremented value...
1651 IVCount = SE->getTruncateOrZeroExtend(IVCount, CntTy);
1655 Value *ExitCnt = genLoopLimit(IndVar, IVCount, L, Rewriter, SE);
1656 assert(ExitCnt->getType()->isPointerTy() == IndVar->getType()->isPointerTy()
1657 && "genLoopLimit missed a cast");
1659 // Insert a new icmp_ne or icmp_eq instruction before the branch.
1660 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1661 ICmpInst::Predicate P;
1662 if (L->contains(BI->getSuccessor(0)))
1663 P = ICmpInst::ICMP_NE;
1665 P = ICmpInst::ICMP_EQ;
1667 DEBUG(dbgs() << "INDVARS: Rewriting loop exit condition to:\n"
1668 << " LHS:" << *CmpIndVar << '\n'
1670 << (P == ICmpInst::ICMP_NE ? "!=" : "==") << "\n"
1671 << " RHS:\t" << *ExitCnt << "\n"
1672 << " IVCount:\t" << *IVCount << "\n");
1674 IRBuilder<> Builder(BI);
1675 if (SE->getTypeSizeInBits(CmpIndVar->getType())
1676 > SE->getTypeSizeInBits(ExitCnt->getType())) {
1677 CmpIndVar = Builder.CreateTrunc(CmpIndVar, ExitCnt->getType(),
1681 Value *Cond = Builder.CreateICmp(P, CmpIndVar, ExitCnt, "exitcond");
1682 Value *OrigCond = BI->getCondition();
1683 // It's tempting to use replaceAllUsesWith here to fully replace the old
1684 // comparison, but that's not immediately safe, since users of the old
1685 // comparison may not be dominated by the new comparison. Instead, just
1686 // update the branch to use the new comparison; in the common case this
1687 // will make old comparison dead.
1688 BI->setCondition(Cond);
1689 DeadInsts.push_back(OrigCond);
1696 //===----------------------------------------------------------------------===//
1697 // SinkUnusedInvariants. A late subpass to cleanup loop preheaders.
1698 //===----------------------------------------------------------------------===//
1700 /// If there's a single exit block, sink any loop-invariant values that
1701 /// were defined in the preheader but not used inside the loop into the
1702 /// exit block to reduce register pressure in the loop.
1703 void IndVarSimplify::SinkUnusedInvariants(Loop *L) {
1704 BasicBlock *ExitBlock = L->getExitBlock();
1705 if (!ExitBlock) return;
1707 BasicBlock *Preheader = L->getLoopPreheader();
1708 if (!Preheader) return;
1710 Instruction *InsertPt = ExitBlock->getFirstInsertionPt();
1711 BasicBlock::iterator I = Preheader->getTerminator();
1712 while (I != Preheader->begin()) {
1714 // New instructions were inserted at the end of the preheader.
1715 if (isa<PHINode>(I))
1718 // Don't move instructions which might have side effects, since the side
1719 // effects need to complete before instructions inside the loop. Also don't
1720 // move instructions which might read memory, since the loop may modify
1721 // memory. Note that it's okay if the instruction might have undefined
1722 // behavior: LoopSimplify guarantees that the preheader dominates the exit
1724 if (I->mayHaveSideEffects() || I->mayReadFromMemory())
1727 // Skip debug info intrinsics.
1728 if (isa<DbgInfoIntrinsic>(I))
1731 // Skip landingpad instructions.
1732 if (isa<LandingPadInst>(I))
1735 // Don't sink alloca: we never want to sink static alloca's out of the
1736 // entry block, and correctly sinking dynamic alloca's requires
1737 // checks for stacksave/stackrestore intrinsics.
1738 // FIXME: Refactor this check somehow?
1739 if (isa<AllocaInst>(I))
1742 // Determine if there is a use in or before the loop (direct or
1744 bool UsedInLoop = false;
1745 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
1748 BasicBlock *UseBB = cast<Instruction>(U)->getParent();
1749 if (PHINode *P = dyn_cast<PHINode>(U)) {
1751 PHINode::getIncomingValueNumForOperand(UI.getOperandNo());
1752 UseBB = P->getIncomingBlock(i);
1754 if (UseBB == Preheader || L->contains(UseBB)) {
1760 // If there is, the def must remain in the preheader.
1764 // Otherwise, sink it to the exit block.
1765 Instruction *ToMove = I;
1768 if (I != Preheader->begin()) {
1769 // Skip debug info intrinsics.
1772 } while (isa<DbgInfoIntrinsic>(I) && I != Preheader->begin());
1774 if (isa<DbgInfoIntrinsic>(I) && I == Preheader->begin())
1780 ToMove->moveBefore(InsertPt);
1786 //===----------------------------------------------------------------------===//
1787 // IndVarSimplify driver. Manage several subpasses of IV simplification.
1788 //===----------------------------------------------------------------------===//
1790 bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) {
1791 // If LoopSimplify form is not available, stay out of trouble. Some notes:
1792 // - LSR currently only supports LoopSimplify-form loops. Indvars'
1793 // canonicalization can be a pessimization without LSR to "clean up"
1795 // - We depend on having a preheader; in particular,
1796 // Loop::getCanonicalInductionVariable only supports loops with preheaders,
1797 // and we're in trouble if we can't find the induction variable even when
1798 // we've manually inserted one.
1799 if (!L->isLoopSimplifyForm())
1802 if (EnableIVRewrite)
1803 IU = &getAnalysis<IVUsers>();
1804 LI = &getAnalysis<LoopInfo>();
1805 SE = &getAnalysis<ScalarEvolution>();
1806 DT = &getAnalysis<DominatorTree>();
1807 TD = getAnalysisIfAvailable<TargetData>();
1812 // If there are any floating-point recurrences, attempt to
1813 // transform them to use integer recurrences.
1814 RewriteNonIntegerIVs(L);
1816 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
1818 // Create a rewriter object which we'll use to transform the code with.
1819 SCEVExpander Rewriter(*SE, "indvars");
1821 Rewriter.setDebugType(DEBUG_TYPE);
1824 // Eliminate redundant IV users.
1826 // Simplification works best when run before other consumers of SCEV. We
1827 // attempt to avoid evaluating SCEVs for sign/zero extend operations until
1828 // other expressions involving loop IVs have been evaluated. This helps SCEV
1829 // set no-wrap flags before normalizing sign/zero extension.
1830 if (!EnableIVRewrite) {
1831 Rewriter.disableCanonicalMode();
1832 SimplifyAndExtend(L, Rewriter, LPM);
1835 // Check to see if this loop has a computable loop-invariant execution count.
1836 // If so, this means that we can compute the final value of any expressions
1837 // that are recurrent in the loop, and substitute the exit values from the
1838 // loop into any instructions outside of the loop that use the final values of
1839 // the current expressions.
1841 if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1842 RewriteLoopExitValues(L, Rewriter);
1844 // Eliminate redundant IV users.
1845 if (EnableIVRewrite)
1846 Changed |= simplifyIVUsers(IU, SE, &LPM, DeadInsts);
1848 // Eliminate redundant IV cycles.
1849 if (!EnableIVRewrite)
1850 NumElimIV += Rewriter.replaceCongruentIVs(L, DT, DeadInsts);
1852 // Compute the type of the largest recurrence expression, and decide whether
1853 // a canonical induction variable should be inserted.
1854 Type *LargestType = 0;
1855 bool NeedCannIV = false;
1856 bool ExpandBECount = canExpandBackedgeTakenCount(L, SE);
1857 if (EnableIVRewrite && ExpandBECount) {
1858 // If we have a known trip count and a single exit block, we'll be
1859 // rewriting the loop exit test condition below, which requires a
1860 // canonical induction variable.
1862 Type *Ty = BackedgeTakenCount->getType();
1863 if (!EnableIVRewrite) {
1864 // In this mode, SimplifyIVUsers may have already widened the IV used by
1865 // the backedge test and inserted a Trunc on the compare's operand. Get
1866 // the wider type to avoid creating a redundant narrow IV only used by the
1868 LargestType = getBackedgeIVType(L);
1871 SE->getTypeSizeInBits(Ty) >
1872 SE->getTypeSizeInBits(LargestType))
1873 LargestType = SE->getEffectiveSCEVType(Ty);
1875 if (EnableIVRewrite) {
1876 for (IVUsers::const_iterator I = IU->begin(), E = IU->end(); I != E; ++I) {
1879 SE->getEffectiveSCEVType(I->getOperandValToReplace()->getType());
1881 SE->getTypeSizeInBits(Ty) >
1882 SE->getTypeSizeInBits(LargestType))
1887 // Now that we know the largest of the induction variable expressions
1888 // in this loop, insert a canonical induction variable of the largest size.
1889 PHINode *IndVar = 0;
1891 // Check to see if the loop already has any canonical-looking induction
1892 // variables. If any are present and wider than the planned canonical
1893 // induction variable, temporarily remove them, so that the Rewriter
1894 // doesn't attempt to reuse them.
1895 SmallVector<PHINode *, 2> OldCannIVs;
1896 while (PHINode *OldCannIV = L->getCanonicalInductionVariable()) {
1897 if (SE->getTypeSizeInBits(OldCannIV->getType()) >
1898 SE->getTypeSizeInBits(LargestType))
1899 OldCannIV->removeFromParent();
1902 OldCannIVs.push_back(OldCannIV);
1905 IndVar = Rewriter.getOrInsertCanonicalInductionVariable(L, LargestType);
1909 DEBUG(dbgs() << "INDVARS: New CanIV: " << *IndVar << '\n');
1911 // Now that the official induction variable is established, reinsert
1912 // any old canonical-looking variables after it so that the IR remains
1913 // consistent. They will be deleted as part of the dead-PHI deletion at
1914 // the end of the pass.
1915 while (!OldCannIVs.empty()) {
1916 PHINode *OldCannIV = OldCannIVs.pop_back_val();
1917 OldCannIV->insertBefore(L->getHeader()->getFirstInsertionPt());
1920 else if (!EnableIVRewrite && ExpandBECount && needsLFTR(L, DT)) {
1921 IndVar = FindLoopCounter(L, BackedgeTakenCount, SE, DT, TD);
1923 // If we have a trip count expression, rewrite the loop's exit condition
1924 // using it. We can currently only handle loops with a single exit.
1926 if (ExpandBECount && IndVar) {
1927 // Check preconditions for proper SCEVExpander operation. SCEV does not
1928 // express SCEVExpander's dependencies, such as LoopSimplify. Instead any
1929 // pass that uses the SCEVExpander must do it. This does not work well for
1930 // loop passes because SCEVExpander makes assumptions about all loops, while
1931 // LoopPassManager only forces the current loop to be simplified.
1933 // FIXME: SCEV expansion has no way to bail out, so the caller must
1934 // explicitly check any assumptions made by SCEV. Brittle.
1935 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(BackedgeTakenCount);
1936 if (!AR || AR->getLoop()->getLoopPreheader())
1938 LinearFunctionTestReplace(L, BackedgeTakenCount, IndVar, Rewriter);
1940 // Rewrite IV-derived expressions.
1941 if (EnableIVRewrite)
1942 RewriteIVExpressions(L, Rewriter);
1944 // Clear the rewriter cache, because values that are in the rewriter's cache
1945 // can be deleted in the loop below, causing the AssertingVH in the cache to
1949 // Now that we're done iterating through lists, clean up any instructions
1950 // which are now dead.
1951 while (!DeadInsts.empty())
1952 if (Instruction *Inst =
1953 dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val()))
1954 RecursivelyDeleteTriviallyDeadInstructions(Inst);
1956 // The Rewriter may not be used from this point on.
1958 // Loop-invariant instructions in the preheader that aren't used in the
1959 // loop may be sunk below the loop to reduce register pressure.
1960 SinkUnusedInvariants(L);
1962 // For completeness, inform IVUsers of the IV use in the newly-created
1963 // loop exit test instruction.
1964 if (IU && NewICmp) {
1965 ICmpInst *NewICmpInst = dyn_cast<ICmpInst>(NewICmp);
1967 IU->AddUsersIfInteresting(cast<Instruction>(NewICmpInst->getOperand(0)));
1969 // Clean up dead instructions.
1970 Changed |= DeleteDeadPHIs(L->getHeader());
1971 // Check a post-condition.
1972 assert(L->isLCSSAForm(*DT) &&
1973 "Indvars did not leave the loop in lcssa form!");
1975 // Verify that LFTR, and any other change have not interfered with SCEV's
1976 // ability to compute trip count.
1978 if (!EnableIVRewrite && VerifyIndvars &&
1979 !isa<SCEVCouldNotCompute>(BackedgeTakenCount)) {
1981 const SCEV *NewBECount = SE->getBackedgeTakenCount(L);
1982 if (SE->getTypeSizeInBits(BackedgeTakenCount->getType()) <
1983 SE->getTypeSizeInBits(NewBECount->getType()))
1984 NewBECount = SE->getTruncateOrNoop(NewBECount,
1985 BackedgeTakenCount->getType());
1987 BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount,
1988 NewBECount->getType());
1989 assert(BackedgeTakenCount == NewBECount && "indvars must preserve SCEV");