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 #include "llvm/Transforms/Scalar.h"
28 #include "llvm/ADT/DenseMap.h"
29 #include "llvm/ADT/SmallVector.h"
30 #include "llvm/ADT/Statistic.h"
31 #include "llvm/Analysis/LoopInfo.h"
32 #include "llvm/Analysis/LoopPass.h"
33 #include "llvm/Analysis/ScalarEvolutionExpander.h"
34 #include "llvm/IR/BasicBlock.h"
35 #include "llvm/IR/CFG.h"
36 #include "llvm/IR/Constants.h"
37 #include "llvm/IR/DataLayout.h"
38 #include "llvm/IR/Dominators.h"
39 #include "llvm/IR/Instructions.h"
40 #include "llvm/IR/IntrinsicInst.h"
41 #include "llvm/IR/LLVMContext.h"
42 #include "llvm/IR/Type.h"
43 #include "llvm/Support/CommandLine.h"
44 #include "llvm/Support/Debug.h"
45 #include "llvm/Support/raw_ostream.h"
46 #include "llvm/Target/TargetLibraryInfo.h"
47 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
48 #include "llvm/Transforms/Utils/Local.h"
49 #include "llvm/Transforms/Utils/SimplifyIndVar.h"
52 #define DEBUG_TYPE "indvars"
54 STATISTIC(NumWidened , "Number of indvars widened");
55 STATISTIC(NumReplaced , "Number of exit values replaced");
56 STATISTIC(NumLFTR , "Number of loop exit tests replaced");
57 STATISTIC(NumElimExt , "Number of IV sign/zero extends eliminated");
58 STATISTIC(NumElimIV , "Number of congruent IVs eliminated");
60 // Trip count verification can be enabled by default under NDEBUG if we
61 // implement a strong expression equivalence checker in SCEV. Until then, we
62 // use the verify-indvars flag, which may assert in some cases.
63 static cl::opt<bool> VerifyIndvars(
64 "verify-indvars", cl::Hidden,
65 cl::desc("Verify the ScalarEvolution result after running indvars"));
67 static cl::opt<bool> ReduceLiveIVs("liv-reduce", cl::Hidden,
68 cl::desc("Reduce live induction variables."));
71 class IndVarSimplify : public LoopPass {
76 TargetLibraryInfo *TLI;
78 SmallVector<WeakVH, 16> DeadInsts;
82 static char ID; // Pass identification, replacement for typeid
83 IndVarSimplify() : LoopPass(ID), LI(nullptr), SE(nullptr), DT(nullptr),
84 DL(nullptr), Changed(false) {
85 initializeIndVarSimplifyPass(*PassRegistry::getPassRegistry());
88 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
90 void getAnalysisUsage(AnalysisUsage &AU) const override {
91 AU.addRequired<DominatorTreeWrapperPass>();
92 AU.addRequired<LoopInfo>();
93 AU.addRequired<ScalarEvolution>();
94 AU.addRequiredID(LoopSimplifyID);
95 AU.addRequiredID(LCSSAID);
96 AU.addPreserved<ScalarEvolution>();
97 AU.addPreservedID(LoopSimplifyID);
98 AU.addPreservedID(LCSSAID);
103 void releaseMemory() override {
107 bool isValidRewrite(Value *FromVal, Value *ToVal);
109 void HandleFloatingPointIV(Loop *L, PHINode *PH);
110 void RewriteNonIntegerIVs(Loop *L);
112 void SimplifyAndExtend(Loop *L, SCEVExpander &Rewriter, LPPassManager &LPM);
114 void RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter);
116 Value *LinearFunctionTestReplace(Loop *L, const SCEV *BackedgeTakenCount,
117 PHINode *IndVar, SCEVExpander &Rewriter);
119 void SinkUnusedInvariants(Loop *L);
123 char IndVarSimplify::ID = 0;
124 INITIALIZE_PASS_BEGIN(IndVarSimplify, "indvars",
125 "Induction Variable Simplification", false, false)
126 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
127 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
128 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
129 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
130 INITIALIZE_PASS_DEPENDENCY(LCSSA)
131 INITIALIZE_PASS_END(IndVarSimplify, "indvars",
132 "Induction Variable Simplification", false, false)
134 Pass *llvm::createIndVarSimplifyPass() {
135 return new IndVarSimplify();
138 /// isValidRewrite - Return true if the SCEV expansion generated by the
139 /// rewriter can replace the original value. SCEV guarantees that it
140 /// produces the same value, but the way it is produced may be illegal IR.
141 /// Ideally, this function will only be called for verification.
142 bool IndVarSimplify::isValidRewrite(Value *FromVal, Value *ToVal) {
143 // If an SCEV expression subsumed multiple pointers, its expansion could
144 // reassociate the GEP changing the base pointer. This is illegal because the
145 // final address produced by a GEP chain must be inbounds relative to its
146 // underlying object. Otherwise basic alias analysis, among other things,
147 // could fail in a dangerous way. Ultimately, SCEV will be improved to avoid
148 // producing an expression involving multiple pointers. Until then, we must
151 // Retrieve the pointer operand of the GEP. Don't use GetUnderlyingObject
152 // because it understands lcssa phis while SCEV does not.
153 Value *FromPtr = FromVal;
154 Value *ToPtr = ToVal;
155 if (GEPOperator *GEP = dyn_cast<GEPOperator>(FromVal)) {
156 FromPtr = GEP->getPointerOperand();
158 if (GEPOperator *GEP = dyn_cast<GEPOperator>(ToVal)) {
159 ToPtr = GEP->getPointerOperand();
161 if (FromPtr != FromVal || ToPtr != ToVal) {
162 // Quickly check the common case
163 if (FromPtr == ToPtr)
166 // SCEV may have rewritten an expression that produces the GEP's pointer
167 // operand. That's ok as long as the pointer operand has the same base
168 // pointer. Unlike GetUnderlyingObject(), getPointerBase() will find the
169 // base of a recurrence. This handles the case in which SCEV expansion
170 // converts a pointer type recurrence into a nonrecurrent pointer base
171 // indexed by an integer recurrence.
173 // If the GEP base pointer is a vector of pointers, abort.
174 if (!FromPtr->getType()->isPointerTy() || !ToPtr->getType()->isPointerTy())
177 const SCEV *FromBase = SE->getPointerBase(SE->getSCEV(FromPtr));
178 const SCEV *ToBase = SE->getPointerBase(SE->getSCEV(ToPtr));
179 if (FromBase == ToBase)
182 DEBUG(dbgs() << "INDVARS: GEP rewrite bail out "
183 << *FromBase << " != " << *ToBase << "\n");
190 /// Determine the insertion point for this user. By default, insert immediately
191 /// before the user. SCEVExpander or LICM will hoist loop invariants out of the
192 /// loop. For PHI nodes, there may be multiple uses, so compute the nearest
193 /// common dominator for the incoming blocks.
194 static Instruction *getInsertPointForUses(Instruction *User, Value *Def,
196 PHINode *PHI = dyn_cast<PHINode>(User);
200 Instruction *InsertPt = nullptr;
201 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i) {
202 if (PHI->getIncomingValue(i) != Def)
205 BasicBlock *InsertBB = PHI->getIncomingBlock(i);
207 InsertPt = InsertBB->getTerminator();
210 InsertBB = DT->findNearestCommonDominator(InsertPt->getParent(), InsertBB);
211 InsertPt = InsertBB->getTerminator();
213 assert(InsertPt && "Missing phi operand");
214 assert((!isa<Instruction>(Def) ||
215 DT->dominates(cast<Instruction>(Def), InsertPt)) &&
216 "def does not dominate all uses");
220 //===----------------------------------------------------------------------===//
221 // RewriteNonIntegerIVs and helpers. Prefer integer IVs.
222 //===----------------------------------------------------------------------===//
224 /// ConvertToSInt - Convert APF to an integer, if possible.
225 static bool ConvertToSInt(const APFloat &APF, int64_t &IntVal) {
226 bool isExact = false;
227 // See if we can convert this to an int64_t
229 if (APF.convertToInteger(&UIntVal, 64, true, APFloat::rmTowardZero,
230 &isExact) != APFloat::opOK || !isExact)
236 /// HandleFloatingPointIV - If the loop has floating induction variable
237 /// then insert corresponding integer induction variable if possible.
239 /// for(double i = 0; i < 10000; ++i)
241 /// is converted into
242 /// for(int i = 0; i < 10000; ++i)
245 void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PN) {
246 unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
247 unsigned BackEdge = IncomingEdge^1;
249 // Check incoming value.
250 ConstantFP *InitValueVal =
251 dyn_cast<ConstantFP>(PN->getIncomingValue(IncomingEdge));
254 if (!InitValueVal || !ConvertToSInt(InitValueVal->getValueAPF(), InitValue))
257 // Check IV increment. Reject this PN if increment operation is not
258 // an add or increment value can not be represented by an integer.
259 BinaryOperator *Incr =
260 dyn_cast<BinaryOperator>(PN->getIncomingValue(BackEdge));
261 if (Incr == nullptr || Incr->getOpcode() != Instruction::FAdd) return;
263 // If this is not an add of the PHI with a constantfp, or if the constant fp
264 // is not an integer, bail out.
265 ConstantFP *IncValueVal = dyn_cast<ConstantFP>(Incr->getOperand(1));
267 if (IncValueVal == nullptr || Incr->getOperand(0) != PN ||
268 !ConvertToSInt(IncValueVal->getValueAPF(), IncValue))
271 // Check Incr uses. One user is PN and the other user is an exit condition
272 // used by the conditional terminator.
273 Value::user_iterator IncrUse = Incr->user_begin();
274 Instruction *U1 = cast<Instruction>(*IncrUse++);
275 if (IncrUse == Incr->user_end()) return;
276 Instruction *U2 = cast<Instruction>(*IncrUse++);
277 if (IncrUse != Incr->user_end()) return;
279 // Find exit condition, which is an fcmp. If it doesn't exist, or if it isn't
280 // only used by a branch, we can't transform it.
281 FCmpInst *Compare = dyn_cast<FCmpInst>(U1);
283 Compare = dyn_cast<FCmpInst>(U2);
284 if (!Compare || !Compare->hasOneUse() ||
285 !isa<BranchInst>(Compare->user_back()))
288 BranchInst *TheBr = cast<BranchInst>(Compare->user_back());
290 // We need to verify that the branch actually controls the iteration count
291 // of the loop. If not, the new IV can overflow and no one will notice.
292 // The branch block must be in the loop and one of the successors must be out
294 assert(TheBr->isConditional() && "Can't use fcmp if not conditional");
295 if (!L->contains(TheBr->getParent()) ||
296 (L->contains(TheBr->getSuccessor(0)) &&
297 L->contains(TheBr->getSuccessor(1))))
301 // If it isn't a comparison with an integer-as-fp (the exit value), we can't
303 ConstantFP *ExitValueVal = dyn_cast<ConstantFP>(Compare->getOperand(1));
305 if (ExitValueVal == nullptr ||
306 !ConvertToSInt(ExitValueVal->getValueAPF(), ExitValue))
309 // Find new predicate for integer comparison.
310 CmpInst::Predicate NewPred = CmpInst::BAD_ICMP_PREDICATE;
311 switch (Compare->getPredicate()) {
312 default: return; // Unknown comparison.
313 case CmpInst::FCMP_OEQ:
314 case CmpInst::FCMP_UEQ: NewPred = CmpInst::ICMP_EQ; break;
315 case CmpInst::FCMP_ONE:
316 case CmpInst::FCMP_UNE: NewPred = CmpInst::ICMP_NE; break;
317 case CmpInst::FCMP_OGT:
318 case CmpInst::FCMP_UGT: NewPred = CmpInst::ICMP_SGT; break;
319 case CmpInst::FCMP_OGE:
320 case CmpInst::FCMP_UGE: NewPred = CmpInst::ICMP_SGE; break;
321 case CmpInst::FCMP_OLT:
322 case CmpInst::FCMP_ULT: NewPred = CmpInst::ICMP_SLT; break;
323 case CmpInst::FCMP_OLE:
324 case CmpInst::FCMP_ULE: NewPred = CmpInst::ICMP_SLE; break;
327 // We convert the floating point induction variable to a signed i32 value if
328 // we can. This is only safe if the comparison will not overflow in a way
329 // that won't be trapped by the integer equivalent operations. Check for this
331 // TODO: We could use i64 if it is native and the range requires it.
333 // The start/stride/exit values must all fit in signed i32.
334 if (!isInt<32>(InitValue) || !isInt<32>(IncValue) || !isInt<32>(ExitValue))
337 // If not actually striding (add x, 0.0), avoid touching the code.
341 // Positive and negative strides have different safety conditions.
343 // If we have a positive stride, we require the init to be less than the
345 if (InitValue >= ExitValue)
348 uint32_t Range = uint32_t(ExitValue-InitValue);
349 // Check for infinite loop, either:
350 // while (i <= Exit) or until (i > Exit)
351 if (NewPred == CmpInst::ICMP_SLE || NewPred == CmpInst::ICMP_SGT) {
352 if (++Range == 0) return; // Range overflows.
355 unsigned Leftover = Range % uint32_t(IncValue);
357 // If this is an equality comparison, we require that the strided value
358 // exactly land on the exit value, otherwise the IV condition will wrap
359 // around and do things the fp IV wouldn't.
360 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
364 // If the stride would wrap around the i32 before exiting, we can't
366 if (Leftover != 0 && int32_t(ExitValue+IncValue) < ExitValue)
370 // If we have a negative stride, we require the init to be greater than the
372 if (InitValue <= ExitValue)
375 uint32_t Range = uint32_t(InitValue-ExitValue);
376 // Check for infinite loop, either:
377 // while (i >= Exit) or until (i < Exit)
378 if (NewPred == CmpInst::ICMP_SGE || NewPred == CmpInst::ICMP_SLT) {
379 if (++Range == 0) return; // Range overflows.
382 unsigned Leftover = Range % uint32_t(-IncValue);
384 // If this is an equality comparison, we require that the strided value
385 // exactly land on the exit value, otherwise the IV condition will wrap
386 // around and do things the fp IV wouldn't.
387 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
391 // If the stride would wrap around the i32 before exiting, we can't
393 if (Leftover != 0 && int32_t(ExitValue+IncValue) > ExitValue)
397 IntegerType *Int32Ty = Type::getInt32Ty(PN->getContext());
399 // Insert new integer induction variable.
400 PHINode *NewPHI = PHINode::Create(Int32Ty, 2, PN->getName()+".int", PN);
401 NewPHI->addIncoming(ConstantInt::get(Int32Ty, InitValue),
402 PN->getIncomingBlock(IncomingEdge));
405 BinaryOperator::CreateAdd(NewPHI, ConstantInt::get(Int32Ty, IncValue),
406 Incr->getName()+".int", Incr);
407 NewPHI->addIncoming(NewAdd, PN->getIncomingBlock(BackEdge));
409 ICmpInst *NewCompare = new ICmpInst(TheBr, NewPred, NewAdd,
410 ConstantInt::get(Int32Ty, ExitValue),
413 // In the following deletions, PN may become dead and may be deleted.
414 // Use a WeakVH to observe whether this happens.
417 // Delete the old floating point exit comparison. The branch starts using the
419 NewCompare->takeName(Compare);
420 Compare->replaceAllUsesWith(NewCompare);
421 RecursivelyDeleteTriviallyDeadInstructions(Compare, TLI);
423 // Delete the old floating point increment.
424 Incr->replaceAllUsesWith(UndefValue::get(Incr->getType()));
425 RecursivelyDeleteTriviallyDeadInstructions(Incr, TLI);
427 // If the FP induction variable still has uses, this is because something else
428 // in the loop uses its value. In order to canonicalize the induction
429 // variable, we chose to eliminate the IV and rewrite it in terms of an
432 // We give preference to sitofp over uitofp because it is faster on most
435 Value *Conv = new SIToFPInst(NewPHI, PN->getType(), "indvar.conv",
436 PN->getParent()->getFirstInsertionPt());
437 PN->replaceAllUsesWith(Conv);
438 RecursivelyDeleteTriviallyDeadInstructions(PN, TLI);
443 void IndVarSimplify::RewriteNonIntegerIVs(Loop *L) {
444 // First step. Check to see if there are any floating-point recurrences.
445 // If there are, change them into integer recurrences, permitting analysis by
446 // the SCEV routines.
448 BasicBlock *Header = L->getHeader();
450 SmallVector<WeakVH, 8> PHIs;
451 for (BasicBlock::iterator I = Header->begin();
452 PHINode *PN = dyn_cast<PHINode>(I); ++I)
455 for (unsigned i = 0, e = PHIs.size(); i != e; ++i)
456 if (PHINode *PN = dyn_cast_or_null<PHINode>(&*PHIs[i]))
457 HandleFloatingPointIV(L, PN);
459 // If the loop previously had floating-point IV, ScalarEvolution
460 // may not have been able to compute a trip count. Now that we've done some
461 // re-writing, the trip count may be computable.
466 //===----------------------------------------------------------------------===//
467 // RewriteLoopExitValues - Optimize IV users outside the loop.
468 // As a side effect, reduces the amount of IV processing within the loop.
469 //===----------------------------------------------------------------------===//
471 /// RewriteLoopExitValues - Check to see if this loop has a computable
472 /// loop-invariant execution count. If so, this means that we can compute the
473 /// final value of any expressions that are recurrent in the loop, and
474 /// substitute the exit values from the loop into any instructions outside of
475 /// the loop that use the final values of the current expressions.
477 /// This is mostly redundant with the regular IndVarSimplify activities that
478 /// happen later, except that it's more powerful in some cases, because it's
479 /// able to brute-force evaluate arbitrary instructions as long as they have
480 /// constant operands at the beginning of the loop.
481 void IndVarSimplify::RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter) {
482 // Verify the input to the pass in already in LCSSA form.
483 assert(L->isLCSSAForm(*DT));
485 SmallVector<BasicBlock*, 8> ExitBlocks;
486 L->getUniqueExitBlocks(ExitBlocks);
488 // Find all values that are computed inside the loop, but used outside of it.
489 // Because of LCSSA, these values will only occur in LCSSA PHI Nodes. Scan
490 // the exit blocks of the loop to find them.
491 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
492 BasicBlock *ExitBB = ExitBlocks[i];
494 // If there are no PHI nodes in this exit block, then no values defined
495 // inside the loop are used on this path, skip it.
496 PHINode *PN = dyn_cast<PHINode>(ExitBB->begin());
499 unsigned NumPreds = PN->getNumIncomingValues();
501 // We would like to be able to RAUW single-incoming value PHI nodes. We
502 // have to be certain this is safe even when this is an LCSSA PHI node.
503 // While the computed exit value is no longer varying in *this* loop, the
504 // exit block may be an exit block for an outer containing loop as well,
505 // the exit value may be varying in the outer loop, and thus it may still
506 // require an LCSSA PHI node. The safe case is when this is
507 // single-predecessor PHI node (LCSSA) and the exit block containing it is
508 // part of the enclosing loop, or this is the outer most loop of the nest.
509 // In either case the exit value could (at most) be varying in the same
510 // loop body as the phi node itself. Thus if it is in turn used outside of
511 // an enclosing loop it will only be via a separate LCSSA node.
512 bool LCSSASafePhiForRAUW =
514 (!L->getParentLoop() || L->getParentLoop() == LI->getLoopFor(ExitBB));
516 // Iterate over all of the PHI nodes.
517 BasicBlock::iterator BBI = ExitBB->begin();
518 while ((PN = dyn_cast<PHINode>(BBI++))) {
520 continue; // dead use, don't replace it
522 // SCEV only supports integer expressions for now.
523 if (!PN->getType()->isIntegerTy() && !PN->getType()->isPointerTy())
526 // It's necessary to tell ScalarEvolution about this explicitly so that
527 // it can walk the def-use list and forget all SCEVs, as it may not be
528 // watching the PHI itself. Once the new exit value is in place, there
529 // may not be a def-use connection between the loop and every instruction
530 // which got a SCEVAddRecExpr for that loop.
533 // Iterate over all of the values in all the PHI nodes.
534 for (unsigned i = 0; i != NumPreds; ++i) {
535 // If the value being merged in is not integer or is not defined
536 // in the loop, skip it.
537 Value *InVal = PN->getIncomingValue(i);
538 if (!isa<Instruction>(InVal))
541 // If this pred is for a subloop, not L itself, skip it.
542 if (LI->getLoopFor(PN->getIncomingBlock(i)) != L)
543 continue; // The Block is in a subloop, skip it.
545 // Check that InVal is defined in the loop.
546 Instruction *Inst = cast<Instruction>(InVal);
547 if (!L->contains(Inst))
550 // Okay, this instruction has a user outside of the current loop
551 // and varies predictably *inside* the loop. Evaluate the value it
552 // contains when the loop exits, if possible.
553 const SCEV *ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop());
554 if (!SE->isLoopInvariant(ExitValue, L) ||
555 !isSafeToExpand(ExitValue, *SE))
558 // Computing the value outside of the loop brings no benefit if :
559 // - it is definitely used inside the loop in a way which can not be
561 // - no use outside of the loop can take advantage of hoisting the
562 // computation out of the loop
563 if (ExitValue->getSCEVType()>=scMulExpr) {
564 unsigned NumHardInternalUses = 0;
565 unsigned NumSoftExternalUses = 0;
566 unsigned NumUses = 0;
567 for (auto IB = Inst->user_begin(), IE = Inst->user_end();
568 IB != IE && NumUses <= 6; ++IB) {
569 Instruction *UseInstr = cast<Instruction>(*IB);
570 unsigned Opc = UseInstr->getOpcode();
572 if (L->contains(UseInstr)) {
573 if (Opc == Instruction::Call || Opc == Instruction::Ret)
574 NumHardInternalUses++;
576 if (Opc == Instruction::PHI) {
577 // Do not count the Phi as a use. LCSSA may have inserted
578 // plenty of trivial ones.
580 for (auto PB = UseInstr->user_begin(),
581 PE = UseInstr->user_end();
582 PB != PE && NumUses <= 6; ++PB, ++NumUses) {
583 unsigned PhiOpc = cast<Instruction>(*PB)->getOpcode();
584 if (PhiOpc != Instruction::Call && PhiOpc != Instruction::Ret)
585 NumSoftExternalUses++;
589 if (Opc != Instruction::Call && Opc != Instruction::Ret)
590 NumSoftExternalUses++;
593 if (NumUses <= 6 && NumHardInternalUses && !NumSoftExternalUses)
597 Value *ExitVal = Rewriter.expandCodeFor(ExitValue, PN->getType(), Inst);
599 DEBUG(dbgs() << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal << '\n'
600 << " LoopVal = " << *Inst << "\n");
602 if (!isValidRewrite(Inst, ExitVal)) {
603 DeadInsts.push_back(ExitVal);
609 PN->setIncomingValue(i, ExitVal);
611 // If this instruction is dead now, delete it. Don't do it now to avoid
612 // invalidating iterators.
613 if (isInstructionTriviallyDead(Inst, TLI))
614 DeadInsts.push_back(Inst);
616 // If we determined that this PHI is safe to replace even if an LCSSA
618 if (LCSSASafePhiForRAUW) {
619 PN->replaceAllUsesWith(ExitVal);
620 PN->eraseFromParent();
624 // If we were unable to completely replace the PHI node, clone the PHI
625 // and delete the original one. This lets IVUsers and any other maps
626 // purge the original user from their records.
627 if (!LCSSASafePhiForRAUW) {
628 PHINode *NewPN = cast<PHINode>(PN->clone());
630 NewPN->insertBefore(PN);
631 PN->replaceAllUsesWith(NewPN);
632 PN->eraseFromParent();
637 // The insertion point instruction may have been deleted; clear it out
638 // so that the rewriter doesn't trip over it later.
639 Rewriter.clearInsertPoint();
642 //===----------------------------------------------------------------------===//
643 // IV Widening - Extend the width of an IV to cover its widest uses.
644 //===----------------------------------------------------------------------===//
647 // Collect information about induction variables that are used by sign/zero
648 // extend operations. This information is recorded by CollectExtend and
649 // provides the input to WidenIV.
652 Type *WidestNativeType; // Widest integer type created [sz]ext
653 bool IsSigned; // Was an sext user seen before a zext?
655 WideIVInfo() : NarrowIV(nullptr), WidestNativeType(nullptr),
660 /// visitCast - Update information about the induction variable that is
661 /// extended by this sign or zero extend operation. This is used to determine
662 /// the final width of the IV before actually widening it.
663 static void visitIVCast(CastInst *Cast, WideIVInfo &WI, ScalarEvolution *SE,
664 const DataLayout *DL) {
665 bool IsSigned = Cast->getOpcode() == Instruction::SExt;
666 if (!IsSigned && Cast->getOpcode() != Instruction::ZExt)
669 Type *Ty = Cast->getType();
670 uint64_t Width = SE->getTypeSizeInBits(Ty);
671 if (DL && !DL->isLegalInteger(Width))
674 if (!WI.WidestNativeType) {
675 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
676 WI.IsSigned = IsSigned;
680 // We extend the IV to satisfy the sign of its first user, arbitrarily.
681 if (WI.IsSigned != IsSigned)
684 if (Width > SE->getTypeSizeInBits(WI.WidestNativeType))
685 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
690 /// NarrowIVDefUse - Record a link in the Narrow IV def-use chain along with the
691 /// WideIV that computes the same value as the Narrow IV def. This avoids
692 /// caching Use* pointers.
693 struct NarrowIVDefUse {
694 Instruction *NarrowDef;
695 Instruction *NarrowUse;
696 Instruction *WideDef;
698 NarrowIVDefUse(): NarrowDef(nullptr), NarrowUse(nullptr), WideDef(nullptr) {}
700 NarrowIVDefUse(Instruction *ND, Instruction *NU, Instruction *WD):
701 NarrowDef(ND), NarrowUse(NU), WideDef(WD) {}
704 /// WidenIV - The goal of this transform is to remove sign and zero extends
705 /// without creating any new induction variables. To do this, it creates a new
706 /// phi of the wider type and redirects all users, either removing extends or
707 /// inserting truncs whenever we stop propagating the type.
723 Instruction *WideInc;
724 const SCEV *WideIncExpr;
725 SmallVectorImpl<WeakVH> &DeadInsts;
727 SmallPtrSet<Instruction*,16> Widened;
728 SmallVector<NarrowIVDefUse, 8> NarrowIVUsers;
731 WidenIV(const WideIVInfo &WI, LoopInfo *LInfo,
732 ScalarEvolution *SEv, DominatorTree *DTree,
733 SmallVectorImpl<WeakVH> &DI) :
734 OrigPhi(WI.NarrowIV),
735 WideType(WI.WidestNativeType),
736 IsSigned(WI.IsSigned),
738 L(LI->getLoopFor(OrigPhi->getParent())),
743 WideIncExpr(nullptr),
745 assert(L->getHeader() == OrigPhi->getParent() && "Phi must be an IV");
748 PHINode *CreateWideIV(SCEVExpander &Rewriter);
751 Value *getExtend(Value *NarrowOper, Type *WideType, bool IsSigned,
754 Instruction *CloneIVUser(NarrowIVDefUse DU);
756 const SCEVAddRecExpr *GetWideRecurrence(Instruction *NarrowUse);
758 const SCEVAddRecExpr* GetExtendedOperandRecurrence(NarrowIVDefUse DU);
760 const SCEV *GetSCEVByOpCode(const SCEV *LHS, const SCEV *RHS,
761 unsigned OpCode) const;
763 Instruction *WidenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter);
765 void pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef);
767 } // anonymous namespace
769 /// isLoopInvariant - Perform a quick domtree based check for loop invariance
770 /// assuming that V is used within the loop. LoopInfo::isLoopInvariant() seems
771 /// gratuitous for this purpose.
772 static bool isLoopInvariant(Value *V, const Loop *L, const DominatorTree *DT) {
773 Instruction *Inst = dyn_cast<Instruction>(V);
777 return DT->properlyDominates(Inst->getParent(), L->getHeader());
780 Value *WidenIV::getExtend(Value *NarrowOper, Type *WideType, bool IsSigned,
782 // Set the debug location and conservative insertion point.
783 IRBuilder<> Builder(Use);
784 // Hoist the insertion point into loop preheaders as far as possible.
785 for (const Loop *L = LI->getLoopFor(Use->getParent());
786 L && L->getLoopPreheader() && isLoopInvariant(NarrowOper, L, DT);
787 L = L->getParentLoop())
788 Builder.SetInsertPoint(L->getLoopPreheader()->getTerminator());
790 return IsSigned ? Builder.CreateSExt(NarrowOper, WideType) :
791 Builder.CreateZExt(NarrowOper, WideType);
794 /// CloneIVUser - Instantiate a wide operation to replace a narrow
795 /// operation. This only needs to handle operations that can evaluation to
796 /// SCEVAddRec. It can safely return 0 for any operation we decide not to clone.
797 Instruction *WidenIV::CloneIVUser(NarrowIVDefUse DU) {
798 unsigned Opcode = DU.NarrowUse->getOpcode();
802 case Instruction::Add:
803 case Instruction::Mul:
804 case Instruction::UDiv:
805 case Instruction::Sub:
806 case Instruction::And:
807 case Instruction::Or:
808 case Instruction::Xor:
809 case Instruction::Shl:
810 case Instruction::LShr:
811 case Instruction::AShr:
812 DEBUG(dbgs() << "Cloning IVUser: " << *DU.NarrowUse << "\n");
814 // Replace NarrowDef operands with WideDef. Otherwise, we don't know
815 // anything about the narrow operand yet so must insert a [sz]ext. It is
816 // probably loop invariant and will be folded or hoisted. If it actually
817 // comes from a widened IV, it should be removed during a future call to
819 Value *LHS = (DU.NarrowUse->getOperand(0) == DU.NarrowDef) ? DU.WideDef :
820 getExtend(DU.NarrowUse->getOperand(0), WideType, IsSigned, DU.NarrowUse);
821 Value *RHS = (DU.NarrowUse->getOperand(1) == DU.NarrowDef) ? DU.WideDef :
822 getExtend(DU.NarrowUse->getOperand(1), WideType, IsSigned, DU.NarrowUse);
824 BinaryOperator *NarrowBO = cast<BinaryOperator>(DU.NarrowUse);
825 BinaryOperator *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(),
827 NarrowBO->getName());
828 IRBuilder<> Builder(DU.NarrowUse);
829 Builder.Insert(WideBO);
830 if (const OverflowingBinaryOperator *OBO =
831 dyn_cast<OverflowingBinaryOperator>(NarrowBO)) {
832 if (OBO->hasNoUnsignedWrap()) WideBO->setHasNoUnsignedWrap();
833 if (OBO->hasNoSignedWrap()) WideBO->setHasNoSignedWrap();
839 const SCEV *WidenIV::GetSCEVByOpCode(const SCEV *LHS, const SCEV *RHS,
840 unsigned OpCode) const {
841 if (OpCode == Instruction::Add)
842 return SE->getAddExpr(LHS, RHS);
843 if (OpCode == Instruction::Sub)
844 return SE->getMinusSCEV(LHS, RHS);
845 if (OpCode == Instruction::Mul)
846 return SE->getMulExpr(LHS, RHS);
848 llvm_unreachable("Unsupported opcode.");
851 /// No-wrap operations can transfer sign extension of their result to their
852 /// operands. Generate the SCEV value for the widened operation without
853 /// actually modifying the IR yet. If the expression after extending the
854 /// operands is an AddRec for this loop, return it.
855 const SCEVAddRecExpr* WidenIV::GetExtendedOperandRecurrence(NarrowIVDefUse DU) {
857 // Handle the common case of add<nsw/nuw>
858 const unsigned OpCode = DU.NarrowUse->getOpcode();
859 // Only Add/Sub/Mul instructions supported yet.
860 if (OpCode != Instruction::Add && OpCode != Instruction::Sub &&
861 OpCode != Instruction::Mul)
864 // One operand (NarrowDef) has already been extended to WideDef. Now determine
865 // if extending the other will lead to a recurrence.
866 unsigned ExtendOperIdx = DU.NarrowUse->getOperand(0) == DU.NarrowDef ? 1 : 0;
867 assert(DU.NarrowUse->getOperand(1-ExtendOperIdx) == DU.NarrowDef && "bad DU");
869 const SCEV *ExtendOperExpr = nullptr;
870 const OverflowingBinaryOperator *OBO =
871 cast<OverflowingBinaryOperator>(DU.NarrowUse);
872 if (IsSigned && OBO->hasNoSignedWrap())
873 ExtendOperExpr = SE->getSignExtendExpr(
874 SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType);
875 else if(!IsSigned && OBO->hasNoUnsignedWrap())
876 ExtendOperExpr = SE->getZeroExtendExpr(
877 SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType);
881 // When creating this SCEV expr, don't apply the current operations NSW or NUW
882 // flags. This instruction may be guarded by control flow that the no-wrap
883 // behavior depends on. Non-control-equivalent instructions can be mapped to
884 // the same SCEV expression, and it would be incorrect to transfer NSW/NUW
885 // semantics to those operations.
886 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(
887 GetSCEVByOpCode(SE->getSCEV(DU.WideDef), ExtendOperExpr, OpCode));
888 if (!AddRec || AddRec->getLoop() != L)
893 /// GetWideRecurrence - Is this instruction potentially interesting from
894 /// IVUsers' perspective after widening it's type? In other words, can the
895 /// extend be safely hoisted out of the loop with SCEV reducing the value to a
896 /// recurrence on the same loop. If so, return the sign or zero extended
897 /// recurrence. Otherwise return NULL.
898 const SCEVAddRecExpr *WidenIV::GetWideRecurrence(Instruction *NarrowUse) {
899 if (!SE->isSCEVable(NarrowUse->getType()))
902 const SCEV *NarrowExpr = SE->getSCEV(NarrowUse);
903 if (SE->getTypeSizeInBits(NarrowExpr->getType())
904 >= SE->getTypeSizeInBits(WideType)) {
905 // NarrowUse implicitly widens its operand. e.g. a gep with a narrow
906 // index. So don't follow this use.
910 const SCEV *WideExpr = IsSigned ?
911 SE->getSignExtendExpr(NarrowExpr, WideType) :
912 SE->getZeroExtendExpr(NarrowExpr, WideType);
913 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(WideExpr);
914 if (!AddRec || AddRec->getLoop() != L)
919 /// This IV user cannot be widen. Replace this use of the original narrow IV
920 /// with a truncation of the new wide IV to isolate and eliminate the narrow IV.
921 static void truncateIVUse(NarrowIVDefUse DU, DominatorTree *DT) {
922 DEBUG(dbgs() << "INDVARS: Truncate IV " << *DU.WideDef
923 << " for user " << *DU.NarrowUse << "\n");
924 IRBuilder<> Builder(getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT));
925 Value *Trunc = Builder.CreateTrunc(DU.WideDef, DU.NarrowDef->getType());
926 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, Trunc);
929 /// WidenIVUse - Determine whether an individual user of the narrow IV can be
930 /// widened. If so, return the wide clone of the user.
931 Instruction *WidenIV::WidenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter) {
933 // Stop traversing the def-use chain at inner-loop phis or post-loop phis.
934 if (PHINode *UsePhi = dyn_cast<PHINode>(DU.NarrowUse)) {
935 if (LI->getLoopFor(UsePhi->getParent()) != L) {
936 // For LCSSA phis, sink the truncate outside the loop.
937 // After SimplifyCFG most loop exit targets have a single predecessor.
938 // Otherwise fall back to a truncate within the loop.
939 if (UsePhi->getNumOperands() != 1)
940 truncateIVUse(DU, DT);
943 PHINode::Create(DU.WideDef->getType(), 1, UsePhi->getName() + ".wide",
945 WidePhi->addIncoming(DU.WideDef, UsePhi->getIncomingBlock(0));
946 IRBuilder<> Builder(WidePhi->getParent()->getFirstInsertionPt());
947 Value *Trunc = Builder.CreateTrunc(WidePhi, DU.NarrowDef->getType());
948 UsePhi->replaceAllUsesWith(Trunc);
949 DeadInsts.push_back(UsePhi);
950 DEBUG(dbgs() << "INDVARS: Widen lcssa phi " << *UsePhi
951 << " to " << *WidePhi << "\n");
956 // Our raison d'etre! Eliminate sign and zero extension.
957 if (IsSigned ? isa<SExtInst>(DU.NarrowUse) : isa<ZExtInst>(DU.NarrowUse)) {
958 Value *NewDef = DU.WideDef;
959 if (DU.NarrowUse->getType() != WideType) {
960 unsigned CastWidth = SE->getTypeSizeInBits(DU.NarrowUse->getType());
961 unsigned IVWidth = SE->getTypeSizeInBits(WideType);
962 if (CastWidth < IVWidth) {
963 // The cast isn't as wide as the IV, so insert a Trunc.
964 IRBuilder<> Builder(DU.NarrowUse);
965 NewDef = Builder.CreateTrunc(DU.WideDef, DU.NarrowUse->getType());
968 // A wider extend was hidden behind a narrower one. This may induce
969 // another round of IV widening in which the intermediate IV becomes
970 // dead. It should be very rare.
971 DEBUG(dbgs() << "INDVARS: New IV " << *WidePhi
972 << " not wide enough to subsume " << *DU.NarrowUse << "\n");
973 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef);
974 NewDef = DU.NarrowUse;
977 if (NewDef != DU.NarrowUse) {
978 DEBUG(dbgs() << "INDVARS: eliminating " << *DU.NarrowUse
979 << " replaced by " << *DU.WideDef << "\n");
981 DU.NarrowUse->replaceAllUsesWith(NewDef);
982 DeadInsts.push_back(DU.NarrowUse);
984 // Now that the extend is gone, we want to expose it's uses for potential
985 // further simplification. We don't need to directly inform SimplifyIVUsers
986 // of the new users, because their parent IV will be processed later as a
987 // new loop phi. If we preserved IVUsers analysis, we would also want to
988 // push the uses of WideDef here.
990 // No further widening is needed. The deceased [sz]ext had done it for us.
994 // Does this user itself evaluate to a recurrence after widening?
995 const SCEVAddRecExpr *WideAddRec = GetWideRecurrence(DU.NarrowUse);
997 WideAddRec = GetExtendedOperandRecurrence(DU);
1000 // This user does not evaluate to a recurence after widening, so don't
1001 // follow it. Instead insert a Trunc to kill off the original use,
1002 // eventually isolating the original narrow IV so it can be removed.
1003 truncateIVUse(DU, DT);
1006 // Assume block terminators cannot evaluate to a recurrence. We can't to
1007 // insert a Trunc after a terminator if there happens to be a critical edge.
1008 assert(DU.NarrowUse != DU.NarrowUse->getParent()->getTerminator() &&
1009 "SCEV is not expected to evaluate a block terminator");
1011 // Reuse the IV increment that SCEVExpander created as long as it dominates
1013 Instruction *WideUse = nullptr;
1014 if (WideAddRec == WideIncExpr
1015 && Rewriter.hoistIVInc(WideInc, DU.NarrowUse))
1018 WideUse = CloneIVUser(DU);
1022 // Evaluation of WideAddRec ensured that the narrow expression could be
1023 // extended outside the loop without overflow. This suggests that the wide use
1024 // evaluates to the same expression as the extended narrow use, but doesn't
1025 // absolutely guarantee it. Hence the following failsafe check. In rare cases
1026 // where it fails, we simply throw away the newly created wide use.
1027 if (WideAddRec != SE->getSCEV(WideUse)) {
1028 DEBUG(dbgs() << "Wide use expression mismatch: " << *WideUse
1029 << ": " << *SE->getSCEV(WideUse) << " != " << *WideAddRec << "\n");
1030 DeadInsts.push_back(WideUse);
1034 // Returning WideUse pushes it on the worklist.
1038 /// pushNarrowIVUsers - Add eligible users of NarrowDef to NarrowIVUsers.
1040 void WidenIV::pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef) {
1041 for (User *U : NarrowDef->users()) {
1042 Instruction *NarrowUser = cast<Instruction>(U);
1044 // Handle data flow merges and bizarre phi cycles.
1045 if (!Widened.insert(NarrowUser))
1048 NarrowIVUsers.push_back(NarrowIVDefUse(NarrowDef, NarrowUser, WideDef));
1052 /// CreateWideIV - Process a single induction variable. First use the
1053 /// SCEVExpander to create a wide induction variable that evaluates to the same
1054 /// recurrence as the original narrow IV. Then use a worklist to forward
1055 /// traverse the narrow IV's def-use chain. After WidenIVUse has processed all
1056 /// interesting IV users, the narrow IV will be isolated for removal by
1059 /// It would be simpler to delete uses as they are processed, but we must avoid
1060 /// invalidating SCEV expressions.
1062 PHINode *WidenIV::CreateWideIV(SCEVExpander &Rewriter) {
1063 // Is this phi an induction variable?
1064 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(OrigPhi));
1068 // Widen the induction variable expression.
1069 const SCEV *WideIVExpr = IsSigned ?
1070 SE->getSignExtendExpr(AddRec, WideType) :
1071 SE->getZeroExtendExpr(AddRec, WideType);
1073 assert(SE->getEffectiveSCEVType(WideIVExpr->getType()) == WideType &&
1074 "Expect the new IV expression to preserve its type");
1076 // Can the IV be extended outside the loop without overflow?
1077 AddRec = dyn_cast<SCEVAddRecExpr>(WideIVExpr);
1078 if (!AddRec || AddRec->getLoop() != L)
1081 // An AddRec must have loop-invariant operands. Since this AddRec is
1082 // materialized by a loop header phi, the expression cannot have any post-loop
1083 // operands, so they must dominate the loop header.
1084 assert(SE->properlyDominates(AddRec->getStart(), L->getHeader()) &&
1085 SE->properlyDominates(AddRec->getStepRecurrence(*SE), L->getHeader())
1086 && "Loop header phi recurrence inputs do not dominate the loop");
1088 // The rewriter provides a value for the desired IV expression. This may
1089 // either find an existing phi or materialize a new one. Either way, we
1090 // expect a well-formed cyclic phi-with-increments. i.e. any operand not part
1091 // of the phi-SCC dominates the loop entry.
1092 Instruction *InsertPt = L->getHeader()->begin();
1093 WidePhi = cast<PHINode>(Rewriter.expandCodeFor(AddRec, WideType, InsertPt));
1095 // Remembering the WideIV increment generated by SCEVExpander allows
1096 // WidenIVUse to reuse it when widening the narrow IV's increment. We don't
1097 // employ a general reuse mechanism because the call above is the only call to
1098 // SCEVExpander. Henceforth, we produce 1-to-1 narrow to wide uses.
1099 if (BasicBlock *LatchBlock = L->getLoopLatch()) {
1101 cast<Instruction>(WidePhi->getIncomingValueForBlock(LatchBlock));
1102 WideIncExpr = SE->getSCEV(WideInc);
1105 DEBUG(dbgs() << "Wide IV: " << *WidePhi << "\n");
1108 // Traverse the def-use chain using a worklist starting at the original IV.
1109 assert(Widened.empty() && NarrowIVUsers.empty() && "expect initial state" );
1111 Widened.insert(OrigPhi);
1112 pushNarrowIVUsers(OrigPhi, WidePhi);
1114 while (!NarrowIVUsers.empty()) {
1115 NarrowIVDefUse DU = NarrowIVUsers.pop_back_val();
1117 // Process a def-use edge. This may replace the use, so don't hold a
1118 // use_iterator across it.
1119 Instruction *WideUse = WidenIVUse(DU, Rewriter);
1121 // Follow all def-use edges from the previous narrow use.
1123 pushNarrowIVUsers(DU.NarrowUse, WideUse);
1125 // WidenIVUse may have removed the def-use edge.
1126 if (DU.NarrowDef->use_empty())
1127 DeadInsts.push_back(DU.NarrowDef);
1132 //===----------------------------------------------------------------------===//
1133 // Live IV Reduction - Minimize IVs live across the loop.
1134 //===----------------------------------------------------------------------===//
1137 //===----------------------------------------------------------------------===//
1138 // Simplification of IV users based on SCEV evaluation.
1139 //===----------------------------------------------------------------------===//
1142 class IndVarSimplifyVisitor : public IVVisitor {
1143 ScalarEvolution *SE;
1144 const DataLayout *DL;
1150 IndVarSimplifyVisitor(PHINode *IV, ScalarEvolution *SCEV,
1151 const DataLayout *DL, const DominatorTree *DTree):
1152 SE(SCEV), DL(DL), IVPhi(IV) {
1154 WI.NarrowIV = IVPhi;
1156 setSplitOverflowIntrinsics();
1159 // Implement the interface used by simplifyUsersOfIV.
1160 void visitCast(CastInst *Cast) override { visitIVCast(Cast, WI, SE, DL); }
1164 /// SimplifyAndExtend - Iteratively perform simplification on a worklist of IV
1165 /// users. Each successive simplification may push more users which may
1166 /// themselves be candidates for simplification.
1168 /// Sign/Zero extend elimination is interleaved with IV simplification.
1170 void IndVarSimplify::SimplifyAndExtend(Loop *L,
1171 SCEVExpander &Rewriter,
1172 LPPassManager &LPM) {
1173 SmallVector<WideIVInfo, 8> WideIVs;
1175 SmallVector<PHINode*, 8> LoopPhis;
1176 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
1177 LoopPhis.push_back(cast<PHINode>(I));
1179 // Each round of simplification iterates through the SimplifyIVUsers worklist
1180 // for all current phis, then determines whether any IVs can be
1181 // widened. Widening adds new phis to LoopPhis, inducing another round of
1182 // simplification on the wide IVs.
1183 while (!LoopPhis.empty()) {
1184 // Evaluate as many IV expressions as possible before widening any IVs. This
1185 // forces SCEV to set no-wrap flags before evaluating sign/zero
1186 // extension. The first time SCEV attempts to normalize sign/zero extension,
1187 // the result becomes final. So for the most predictable results, we delay
1188 // evaluation of sign/zero extend evaluation until needed, and avoid running
1189 // other SCEV based analysis prior to SimplifyAndExtend.
1191 PHINode *CurrIV = LoopPhis.pop_back_val();
1193 // Information about sign/zero extensions of CurrIV.
1194 IndVarSimplifyVisitor Visitor(CurrIV, SE, DL, DT);
1196 Changed |= simplifyUsersOfIV(CurrIV, SE, &LPM, DeadInsts, &Visitor);
1198 if (Visitor.WI.WidestNativeType) {
1199 WideIVs.push_back(Visitor.WI);
1201 } while(!LoopPhis.empty());
1203 for (; !WideIVs.empty(); WideIVs.pop_back()) {
1204 WidenIV Widener(WideIVs.back(), LI, SE, DT, DeadInsts);
1205 if (PHINode *WidePhi = Widener.CreateWideIV(Rewriter)) {
1207 LoopPhis.push_back(WidePhi);
1213 //===----------------------------------------------------------------------===//
1214 // LinearFunctionTestReplace and its kin. Rewrite the loop exit condition.
1215 //===----------------------------------------------------------------------===//
1217 /// Check for expressions that ScalarEvolution generates to compute
1218 /// BackedgeTakenInfo. If these expressions have not been reduced, then
1219 /// expanding them may incur additional cost (albeit in the loop preheader).
1220 static bool isHighCostExpansion(const SCEV *S, BranchInst *BI,
1221 SmallPtrSetImpl<const SCEV*> &Processed,
1222 ScalarEvolution *SE) {
1223 if (!Processed.insert(S))
1226 // If the backedge-taken count is a UDiv, it's very likely a UDiv that
1227 // ScalarEvolution's HowFarToZero or HowManyLessThans produced to compute a
1228 // precise expression, rather than a UDiv from the user's code. If we can't
1229 // find a UDiv in the code with some simple searching, assume the former and
1230 // forego rewriting the loop.
1231 if (isa<SCEVUDivExpr>(S)) {
1232 ICmpInst *OrigCond = dyn_cast<ICmpInst>(BI->getCondition());
1233 if (!OrigCond) return true;
1234 const SCEV *R = SE->getSCEV(OrigCond->getOperand(1));
1235 R = SE->getMinusSCEV(R, SE->getConstant(R->getType(), 1));
1237 const SCEV *L = SE->getSCEV(OrigCond->getOperand(0));
1238 L = SE->getMinusSCEV(L, SE->getConstant(L->getType(), 1));
1244 // Recurse past add expressions, which commonly occur in the
1245 // BackedgeTakenCount. They may already exist in program code, and if not,
1246 // they are not too expensive rematerialize.
1247 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
1248 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
1250 if (isHighCostExpansion(*I, BI, Processed, SE))
1256 // HowManyLessThans uses a Max expression whenever the loop is not guarded by
1257 // the exit condition.
1258 if (isa<SCEVSMaxExpr>(S) || isa<SCEVUMaxExpr>(S))
1261 // If we haven't recognized an expensive SCEV pattern, assume it's an
1262 // expression produced by program code.
1266 /// canExpandBackedgeTakenCount - Return true if this loop's backedge taken
1267 /// count expression can be safely and cheaply expanded into an instruction
1268 /// sequence that can be used by LinearFunctionTestReplace.
1270 /// TODO: This fails for pointer-type loop counters with greater than one byte
1271 /// strides, consequently preventing LFTR from running. For the purpose of LFTR
1272 /// we could skip this check in the case that the LFTR loop counter (chosen by
1273 /// FindLoopCounter) is also pointer type. Instead, we could directly convert
1274 /// the loop test to an inequality test by checking the target data's alignment
1275 /// of element types (given that the initial pointer value originates from or is
1276 /// used by ABI constrained operation, as opposed to inttoptr/ptrtoint).
1277 /// However, we don't yet have a strong motivation for converting loop tests
1278 /// into inequality tests.
1279 static bool canExpandBackedgeTakenCount(Loop *L, ScalarEvolution *SE) {
1280 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
1281 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount) ||
1282 BackedgeTakenCount->isZero())
1285 if (!L->getExitingBlock())
1288 // Can't rewrite non-branch yet.
1289 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
1293 SmallPtrSet<const SCEV*, 8> Processed;
1294 if (isHighCostExpansion(BackedgeTakenCount, BI, Processed, SE))
1300 /// getLoopPhiForCounter - Return the loop header phi IFF IncV adds a loop
1301 /// invariant value to the phi.
1302 static PHINode *getLoopPhiForCounter(Value *IncV, Loop *L, DominatorTree *DT) {
1303 Instruction *IncI = dyn_cast<Instruction>(IncV);
1307 switch (IncI->getOpcode()) {
1308 case Instruction::Add:
1309 case Instruction::Sub:
1311 case Instruction::GetElementPtr:
1312 // An IV counter must preserve its type.
1313 if (IncI->getNumOperands() == 2)
1319 PHINode *Phi = dyn_cast<PHINode>(IncI->getOperand(0));
1320 if (Phi && Phi->getParent() == L->getHeader()) {
1321 if (isLoopInvariant(IncI->getOperand(1), L, DT))
1325 if (IncI->getOpcode() == Instruction::GetElementPtr)
1328 // Allow add/sub to be commuted.
1329 Phi = dyn_cast<PHINode>(IncI->getOperand(1));
1330 if (Phi && Phi->getParent() == L->getHeader()) {
1331 if (isLoopInvariant(IncI->getOperand(0), L, DT))
1337 /// Return the compare guarding the loop latch, or NULL for unrecognized tests.
1338 static ICmpInst *getLoopTest(Loop *L) {
1339 assert(L->getExitingBlock() && "expected loop exit");
1341 BasicBlock *LatchBlock = L->getLoopLatch();
1342 // Don't bother with LFTR if the loop is not properly simplified.
1346 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
1347 assert(BI && "expected exit branch");
1349 return dyn_cast<ICmpInst>(BI->getCondition());
1352 /// needsLFTR - LinearFunctionTestReplace policy. Return true unless we can show
1353 /// that the current exit test is already sufficiently canonical.
1354 static bool needsLFTR(Loop *L, DominatorTree *DT) {
1355 // Do LFTR to simplify the exit condition to an ICMP.
1356 ICmpInst *Cond = getLoopTest(L);
1360 // Do LFTR to simplify the exit ICMP to EQ/NE
1361 ICmpInst::Predicate Pred = Cond->getPredicate();
1362 if (Pred != ICmpInst::ICMP_NE && Pred != ICmpInst::ICMP_EQ)
1365 // Look for a loop invariant RHS
1366 Value *LHS = Cond->getOperand(0);
1367 Value *RHS = Cond->getOperand(1);
1368 if (!isLoopInvariant(RHS, L, DT)) {
1369 if (!isLoopInvariant(LHS, L, DT))
1371 std::swap(LHS, RHS);
1373 // Look for a simple IV counter LHS
1374 PHINode *Phi = dyn_cast<PHINode>(LHS);
1376 Phi = getLoopPhiForCounter(LHS, L, DT);
1381 // Do LFTR if PHI node is defined in the loop, but is *not* a counter.
1382 int Idx = Phi->getBasicBlockIndex(L->getLoopLatch());
1386 // Do LFTR if the exit condition's IV is *not* a simple counter.
1387 Value *IncV = Phi->getIncomingValue(Idx);
1388 return Phi != getLoopPhiForCounter(IncV, L, DT);
1391 /// Recursive helper for hasConcreteDef(). Unfortunately, this currently boils
1392 /// down to checking that all operands are constant and listing instructions
1393 /// that may hide undef.
1394 static bool hasConcreteDefImpl(Value *V, SmallPtrSetImpl<Value*> &Visited,
1396 if (isa<Constant>(V))
1397 return !isa<UndefValue>(V);
1402 // Conservatively handle non-constant non-instructions. For example, Arguments
1404 Instruction *I = dyn_cast<Instruction>(V);
1408 // Load and return values may be undef.
1409 if(I->mayReadFromMemory() || isa<CallInst>(I) || isa<InvokeInst>(I))
1412 // Optimistically handle other instructions.
1413 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) {
1414 if (!Visited.insert(*OI))
1416 if (!hasConcreteDefImpl(*OI, Visited, Depth+1))
1422 /// Return true if the given value is concrete. We must prove that undef can
1425 /// TODO: If we decide that this is a good approach to checking for undef, we
1426 /// may factor it into a common location.
1427 static bool hasConcreteDef(Value *V) {
1428 SmallPtrSet<Value*, 8> Visited;
1430 return hasConcreteDefImpl(V, Visited, 0);
1433 /// AlmostDeadIV - Return true if this IV has any uses other than the (soon to
1434 /// be rewritten) loop exit test.
1435 static bool AlmostDeadIV(PHINode *Phi, BasicBlock *LatchBlock, Value *Cond) {
1436 int LatchIdx = Phi->getBasicBlockIndex(LatchBlock);
1437 Value *IncV = Phi->getIncomingValue(LatchIdx);
1439 for (User *U : Phi->users())
1440 if (U != Cond && U != IncV) return false;
1442 for (User *U : IncV->users())
1443 if (U != Cond && U != 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 /// Currently only valid for LFTR. See the comments on hasConcreteDef below.
1455 /// FIXME: Accept -1 stride and set IVLimit = IVInit - BECount
1457 /// FIXME: Accept non-unit stride as long as SCEV can reduce BECount * Stride.
1458 /// This is difficult in general for SCEV because of potential overflow. But we
1459 /// could at least handle constant BECounts.
1461 FindLoopCounter(Loop *L, const SCEV *BECount,
1462 ScalarEvolution *SE, DominatorTree *DT, const DataLayout *DL) {
1463 uint64_t BCWidth = SE->getTypeSizeInBits(BECount->getType());
1466 cast<BranchInst>(L->getExitingBlock()->getTerminator())->getCondition();
1468 // Loop over all of the PHI nodes, looking for a simple counter.
1469 PHINode *BestPhi = nullptr;
1470 const SCEV *BestInit = nullptr;
1471 BasicBlock *LatchBlock = L->getLoopLatch();
1472 assert(LatchBlock && "needsLFTR should guarantee a loop latch");
1474 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
1475 PHINode *Phi = cast<PHINode>(I);
1476 if (!SE->isSCEVable(Phi->getType()))
1479 // Avoid comparing an integer IV against a pointer Limit.
1480 if (BECount->getType()->isPointerTy() && !Phi->getType()->isPointerTy())
1483 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Phi));
1484 if (!AR || AR->getLoop() != L || !AR->isAffine())
1487 // AR may be a pointer type, while BECount is an integer type.
1488 // AR may be wider than BECount. With eq/ne tests overflow is immaterial.
1489 // AR may not be a narrower type, or we may never exit.
1490 uint64_t PhiWidth = SE->getTypeSizeInBits(AR->getType());
1491 if (PhiWidth < BCWidth || (DL && !DL->isLegalInteger(PhiWidth)))
1494 const SCEV *Step = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE));
1495 if (!Step || !Step->isOne())
1498 int LatchIdx = Phi->getBasicBlockIndex(LatchBlock);
1499 Value *IncV = Phi->getIncomingValue(LatchIdx);
1500 if (getLoopPhiForCounter(IncV, L, DT) != Phi)
1503 // Avoid reusing a potentially undef value to compute other values that may
1504 // have originally had a concrete definition.
1505 if (!hasConcreteDef(Phi)) {
1506 // We explicitly allow unknown phis as long as they are already used by
1507 // the loop test. In this case we assume that performing LFTR could not
1508 // increase the number of undef users.
1509 if (ICmpInst *Cond = getLoopTest(L)) {
1510 if (Phi != getLoopPhiForCounter(Cond->getOperand(0), L, DT)
1511 && Phi != getLoopPhiForCounter(Cond->getOperand(1), L, DT)) {
1516 const SCEV *Init = AR->getStart();
1518 if (BestPhi && !AlmostDeadIV(BestPhi, LatchBlock, Cond)) {
1519 // Don't force a live loop counter if another IV can be used.
1520 if (AlmostDeadIV(Phi, LatchBlock, Cond))
1523 // Prefer to count-from-zero. This is a more "canonical" counter form. It
1524 // also prefers integer to pointer IVs.
1525 if (BestInit->isZero() != Init->isZero()) {
1526 if (BestInit->isZero())
1529 // If two IVs both count from zero or both count from nonzero then the
1530 // narrower is likely a dead phi that has been widened. Use the wider phi
1531 // to allow the other to be eliminated.
1532 else if (PhiWidth <= SE->getTypeSizeInBits(BestPhi->getType()))
1541 /// genLoopLimit - Help LinearFunctionTestReplace by generating a value that
1542 /// holds the RHS of the new loop test.
1543 static Value *genLoopLimit(PHINode *IndVar, const SCEV *IVCount, Loop *L,
1544 SCEVExpander &Rewriter, ScalarEvolution *SE) {
1545 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(IndVar));
1546 assert(AR && AR->getLoop() == L && AR->isAffine() && "bad loop counter");
1547 const SCEV *IVInit = AR->getStart();
1549 // IVInit may be a pointer while IVCount is an integer when FindLoopCounter
1550 // finds a valid pointer IV. Sign extend BECount in order to materialize a
1551 // GEP. Avoid running SCEVExpander on a new pointer value, instead reusing
1552 // the existing GEPs whenever possible.
1553 if (IndVar->getType()->isPointerTy()
1554 && !IVCount->getType()->isPointerTy()) {
1556 // IVOffset will be the new GEP offset that is interpreted by GEP as a
1557 // signed value. IVCount on the other hand represents the loop trip count,
1558 // which is an unsigned value. FindLoopCounter only allows induction
1559 // variables that have a positive unit stride of one. This means we don't
1560 // have to handle the case of negative offsets (yet) and just need to zero
1562 Type *OfsTy = SE->getEffectiveSCEVType(IVInit->getType());
1563 const SCEV *IVOffset = SE->getTruncateOrZeroExtend(IVCount, OfsTy);
1565 // Expand the code for the iteration count.
1566 assert(SE->isLoopInvariant(IVOffset, L) &&
1567 "Computed iteration count is not loop invariant!");
1568 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1569 Value *GEPOffset = Rewriter.expandCodeFor(IVOffset, OfsTy, BI);
1571 Value *GEPBase = IndVar->getIncomingValueForBlock(L->getLoopPreheader());
1572 assert(AR->getStart() == SE->getSCEV(GEPBase) && "bad loop counter");
1573 // We could handle pointer IVs other than i8*, but we need to compensate for
1574 // gep index scaling. See canExpandBackedgeTakenCount comments.
1575 assert(SE->getSizeOfExpr(IntegerType::getInt64Ty(IndVar->getContext()),
1576 cast<PointerType>(GEPBase->getType())->getElementType())->isOne()
1577 && "unit stride pointer IV must be i8*");
1579 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
1580 return Builder.CreateGEP(GEPBase, GEPOffset, "lftr.limit");
1583 // In any other case, convert both IVInit and IVCount to integers before
1584 // comparing. This may result in SCEV expension of pointers, but in practice
1585 // SCEV will fold the pointer arithmetic away as such:
1586 // BECount = (IVEnd - IVInit - 1) => IVLimit = IVInit (postinc).
1588 // Valid Cases: (1) both integers is most common; (2) both may be pointers
1589 // for simple memset-style loops.
1591 // IVInit integer and IVCount pointer would only occur if a canonical IV
1592 // were generated on top of case #2, which is not expected.
1594 const SCEV *IVLimit = nullptr;
1595 // For unit stride, IVCount = Start + BECount with 2's complement overflow.
1596 // For non-zero Start, compute IVCount here.
1597 if (AR->getStart()->isZero())
1600 assert(AR->getStepRecurrence(*SE)->isOne() && "only handles unit stride");
1601 const SCEV *IVInit = AR->getStart();
1603 // For integer IVs, truncate the IV before computing IVInit + BECount.
1604 if (SE->getTypeSizeInBits(IVInit->getType())
1605 > SE->getTypeSizeInBits(IVCount->getType()))
1606 IVInit = SE->getTruncateExpr(IVInit, IVCount->getType());
1608 IVLimit = SE->getAddExpr(IVInit, IVCount);
1610 // Expand the code for the iteration count.
1611 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1612 IRBuilder<> Builder(BI);
1613 assert(SE->isLoopInvariant(IVLimit, L) &&
1614 "Computed iteration count is not loop invariant!");
1615 // Ensure that we generate the same type as IndVar, or a smaller integer
1616 // type. In the presence of null pointer values, we have an integer type
1617 // SCEV expression (IVInit) for a pointer type IV value (IndVar).
1618 Type *LimitTy = IVCount->getType()->isPointerTy() ?
1619 IndVar->getType() : IVCount->getType();
1620 return Rewriter.expandCodeFor(IVLimit, LimitTy, BI);
1624 /// LinearFunctionTestReplace - This method rewrites the exit condition of the
1625 /// loop to be a canonical != comparison against the incremented loop induction
1626 /// variable. This pass is able to rewrite the exit tests of any loop where the
1627 /// SCEV analysis can determine a loop-invariant trip count of the loop, which
1628 /// is actually a much broader range than just linear tests.
1629 Value *IndVarSimplify::
1630 LinearFunctionTestReplace(Loop *L,
1631 const SCEV *BackedgeTakenCount,
1633 SCEVExpander &Rewriter) {
1634 assert(canExpandBackedgeTakenCount(L, SE) && "precondition");
1636 // Initialize CmpIndVar and IVCount to their preincremented values.
1637 Value *CmpIndVar = IndVar;
1638 const SCEV *IVCount = BackedgeTakenCount;
1640 // If the exiting block is the same as the backedge block, we prefer to
1641 // compare against the post-incremented value, otherwise we must compare
1642 // against the preincremented value.
1643 if (L->getExitingBlock() == L->getLoopLatch()) {
1644 // The BackedgeTaken expression contains the number of times that the
1645 // backedge branches to the loop header. This is one less than the
1646 // number of times the loop executes, so use the incremented indvar.
1647 llvm::Value *IncrementedIndvar =
1648 IndVar->getIncomingValueForBlock(L->getExitingBlock());
1649 const auto *IncrementedIndvarSCEV =
1650 cast<SCEVAddRecExpr>(SE->getSCEV(IncrementedIndvar));
1651 // It is unsafe to use the incremented indvar if it has a wrapping flag, we
1652 // don't want to compare against a poison value. Check the SCEV that
1653 // corresponds to the incremented indvar, the SCEVExpander will only insert
1654 // flags in the IR if the SCEV originally had wrapping flags.
1655 // FIXME: In theory, SCEV could drop flags even though they exist in IR.
1656 // A more robust solution would involve getting a new expression for
1657 // CmpIndVar by applying non-NSW/NUW AddExprs.
1658 if (!ScalarEvolution::maskFlags(IncrementedIndvarSCEV->getNoWrapFlags(),
1659 SCEV::FlagNUW | SCEV::FlagNSW)) {
1660 // Add one to the "backedge-taken" count to get the trip count.
1661 // This addition may overflow, which is valid as long as the comparison is
1662 // truncated to BackedgeTakenCount->getType().
1664 SE->getAddExpr(BackedgeTakenCount,
1665 SE->getConstant(BackedgeTakenCount->getType(), 1));
1666 CmpIndVar = IncrementedIndvar;
1670 Value *ExitCnt = genLoopLimit(IndVar, IVCount, L, Rewriter, SE);
1671 assert(ExitCnt->getType()->isPointerTy() == IndVar->getType()->isPointerTy()
1672 && "genLoopLimit missed a cast");
1674 // Insert a new icmp_ne or icmp_eq instruction before the branch.
1675 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1676 ICmpInst::Predicate P;
1677 if (L->contains(BI->getSuccessor(0)))
1678 P = ICmpInst::ICMP_NE;
1680 P = ICmpInst::ICMP_EQ;
1682 DEBUG(dbgs() << "INDVARS: Rewriting loop exit condition to:\n"
1683 << " LHS:" << *CmpIndVar << '\n'
1685 << (P == ICmpInst::ICMP_NE ? "!=" : "==") << "\n"
1686 << " RHS:\t" << *ExitCnt << "\n"
1687 << " IVCount:\t" << *IVCount << "\n");
1689 IRBuilder<> Builder(BI);
1691 // LFTR can ignore IV overflow and truncate to the width of
1692 // BECount. This avoids materializing the add(zext(add)) expression.
1693 unsigned CmpIndVarSize = SE->getTypeSizeInBits(CmpIndVar->getType());
1694 unsigned ExitCntSize = SE->getTypeSizeInBits(ExitCnt->getType());
1695 if (CmpIndVarSize > ExitCntSize) {
1696 const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(SE->getSCEV(IndVar));
1697 const SCEV *ARStart = AR->getStart();
1698 const SCEV *ARStep = AR->getStepRecurrence(*SE);
1699 // For constant IVCount, avoid truncation.
1700 if (isa<SCEVConstant>(ARStart) && isa<SCEVConstant>(IVCount)) {
1701 const APInt &Start = cast<SCEVConstant>(ARStart)->getValue()->getValue();
1702 APInt Count = cast<SCEVConstant>(IVCount)->getValue()->getValue();
1703 // Note that the post-inc value of BackedgeTakenCount may have overflowed
1704 // above such that IVCount is now zero.
1705 if (IVCount != BackedgeTakenCount && Count == 0) {
1706 Count = APInt::getMaxValue(Count.getBitWidth()).zext(CmpIndVarSize);
1710 Count = Count.zext(CmpIndVarSize);
1712 if (cast<SCEVConstant>(ARStep)->getValue()->isNegative())
1713 NewLimit = Start - Count;
1715 NewLimit = Start + Count;
1716 ExitCnt = ConstantInt::get(CmpIndVar->getType(), NewLimit);
1718 DEBUG(dbgs() << " Widen RHS:\t" << *ExitCnt << "\n");
1720 CmpIndVar = Builder.CreateTrunc(CmpIndVar, ExitCnt->getType(),
1724 Value *Cond = Builder.CreateICmp(P, CmpIndVar, ExitCnt, "exitcond");
1725 Value *OrigCond = BI->getCondition();
1726 // It's tempting to use replaceAllUsesWith here to fully replace the old
1727 // comparison, but that's not immediately safe, since users of the old
1728 // comparison may not be dominated by the new comparison. Instead, just
1729 // update the branch to use the new comparison; in the common case this
1730 // will make old comparison dead.
1731 BI->setCondition(Cond);
1732 DeadInsts.push_back(OrigCond);
1739 //===----------------------------------------------------------------------===//
1740 // SinkUnusedInvariants. A late subpass to cleanup loop preheaders.
1741 //===----------------------------------------------------------------------===//
1743 /// If there's a single exit block, sink any loop-invariant values that
1744 /// were defined in the preheader but not used inside the loop into the
1745 /// exit block to reduce register pressure in the loop.
1746 void IndVarSimplify::SinkUnusedInvariants(Loop *L) {
1747 BasicBlock *ExitBlock = L->getExitBlock();
1748 if (!ExitBlock) return;
1750 BasicBlock *Preheader = L->getLoopPreheader();
1751 if (!Preheader) return;
1753 Instruction *InsertPt = ExitBlock->getFirstInsertionPt();
1754 BasicBlock::iterator I = Preheader->getTerminator();
1755 while (I != Preheader->begin()) {
1757 // New instructions were inserted at the end of the preheader.
1758 if (isa<PHINode>(I))
1761 // Don't move instructions which might have side effects, since the side
1762 // effects need to complete before instructions inside the loop. Also don't
1763 // move instructions which might read memory, since the loop may modify
1764 // memory. Note that it's okay if the instruction might have undefined
1765 // behavior: LoopSimplify guarantees that the preheader dominates the exit
1767 if (I->mayHaveSideEffects() || I->mayReadFromMemory())
1770 // Skip debug info intrinsics.
1771 if (isa<DbgInfoIntrinsic>(I))
1774 // Skip landingpad instructions.
1775 if (isa<LandingPadInst>(I))
1778 // Don't sink alloca: we never want to sink static alloca's out of the
1779 // entry block, and correctly sinking dynamic alloca's requires
1780 // checks for stacksave/stackrestore intrinsics.
1781 // FIXME: Refactor this check somehow?
1782 if (isa<AllocaInst>(I))
1785 // Determine if there is a use in or before the loop (direct or
1787 bool UsedInLoop = false;
1788 for (Use &U : I->uses()) {
1789 Instruction *User = cast<Instruction>(U.getUser());
1790 BasicBlock *UseBB = User->getParent();
1791 if (PHINode *P = dyn_cast<PHINode>(User)) {
1793 PHINode::getIncomingValueNumForOperand(U.getOperandNo());
1794 UseBB = P->getIncomingBlock(i);
1796 if (UseBB == Preheader || L->contains(UseBB)) {
1802 // If there is, the def must remain in the preheader.
1806 // Otherwise, sink it to the exit block.
1807 Instruction *ToMove = I;
1810 if (I != Preheader->begin()) {
1811 // Skip debug info intrinsics.
1814 } while (isa<DbgInfoIntrinsic>(I) && I != Preheader->begin());
1816 if (isa<DbgInfoIntrinsic>(I) && I == Preheader->begin())
1822 ToMove->moveBefore(InsertPt);
1828 //===----------------------------------------------------------------------===//
1829 // IndVarSimplify driver. Manage several subpasses of IV simplification.
1830 //===----------------------------------------------------------------------===//
1832 bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) {
1833 if (skipOptnoneFunction(L))
1836 // If LoopSimplify form is not available, stay out of trouble. Some notes:
1837 // - LSR currently only supports LoopSimplify-form loops. Indvars'
1838 // canonicalization can be a pessimization without LSR to "clean up"
1840 // - We depend on having a preheader; in particular,
1841 // Loop::getCanonicalInductionVariable only supports loops with preheaders,
1842 // and we're in trouble if we can't find the induction variable even when
1843 // we've manually inserted one.
1844 if (!L->isLoopSimplifyForm())
1847 LI = &getAnalysis<LoopInfo>();
1848 SE = &getAnalysis<ScalarEvolution>();
1849 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1850 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1851 DL = DLP ? &DLP->getDataLayout() : nullptr;
1852 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
1857 // If there are any floating-point recurrences, attempt to
1858 // transform them to use integer recurrences.
1859 RewriteNonIntegerIVs(L);
1861 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
1863 // Create a rewriter object which we'll use to transform the code with.
1864 SCEVExpander Rewriter(*SE, "indvars");
1866 Rewriter.setDebugType(DEBUG_TYPE);
1869 // Eliminate redundant IV users.
1871 // Simplification works best when run before other consumers of SCEV. We
1872 // attempt to avoid evaluating SCEVs for sign/zero extend operations until
1873 // other expressions involving loop IVs have been evaluated. This helps SCEV
1874 // set no-wrap flags before normalizing sign/zero extension.
1875 Rewriter.disableCanonicalMode();
1876 SimplifyAndExtend(L, Rewriter, LPM);
1878 // Check to see if this loop has a computable loop-invariant execution count.
1879 // If so, this means that we can compute the final value of any expressions
1880 // that are recurrent in the loop, and substitute the exit values from the
1881 // loop into any instructions outside of the loop that use the final values of
1882 // the current expressions.
1884 if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1885 RewriteLoopExitValues(L, Rewriter);
1887 // Eliminate redundant IV cycles.
1888 NumElimIV += Rewriter.replaceCongruentIVs(L, DT, DeadInsts);
1890 // If we have a trip count expression, rewrite the loop's exit condition
1891 // using it. We can currently only handle loops with a single exit.
1892 if (canExpandBackedgeTakenCount(L, SE) && needsLFTR(L, DT)) {
1893 PHINode *IndVar = FindLoopCounter(L, BackedgeTakenCount, SE, DT, DL);
1895 // Check preconditions for proper SCEVExpander operation. SCEV does not
1896 // express SCEVExpander's dependencies, such as LoopSimplify. Instead any
1897 // pass that uses the SCEVExpander must do it. This does not work well for
1898 // loop passes because SCEVExpander makes assumptions about all loops,
1899 // while LoopPassManager only forces the current loop to be simplified.
1901 // FIXME: SCEV expansion has no way to bail out, so the caller must
1902 // explicitly check any assumptions made by SCEV. Brittle.
1903 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(BackedgeTakenCount);
1904 if (!AR || AR->getLoop()->getLoopPreheader())
1905 (void)LinearFunctionTestReplace(L, BackedgeTakenCount, IndVar,
1909 // Clear the rewriter cache, because values that are in the rewriter's cache
1910 // can be deleted in the loop below, causing the AssertingVH in the cache to
1914 // Now that we're done iterating through lists, clean up any instructions
1915 // which are now dead.
1916 while (!DeadInsts.empty())
1917 if (Instruction *Inst =
1918 dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val()))
1919 RecursivelyDeleteTriviallyDeadInstructions(Inst, TLI);
1921 // The Rewriter may not be used from this point on.
1923 // Loop-invariant instructions in the preheader that aren't used in the
1924 // loop may be sunk below the loop to reduce register pressure.
1925 SinkUnusedInvariants(L);
1927 // Clean up dead instructions.
1928 Changed |= DeleteDeadPHIs(L->getHeader(), TLI);
1929 // Check a post-condition.
1930 assert(L->isLCSSAForm(*DT) &&
1931 "Indvars did not leave the loop in lcssa form!");
1933 // Verify that LFTR, and any other change have not interfered with SCEV's
1934 // ability to compute trip count.
1936 if (VerifyIndvars && !isa<SCEVCouldNotCompute>(BackedgeTakenCount)) {
1938 const SCEV *NewBECount = SE->getBackedgeTakenCount(L);
1939 if (SE->getTypeSizeInBits(BackedgeTakenCount->getType()) <
1940 SE->getTypeSizeInBits(NewBECount->getType()))
1941 NewBECount = SE->getTruncateOrNoop(NewBECount,
1942 BackedgeTakenCount->getType());
1944 BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount,
1945 NewBECount->getType());
1946 assert(BackedgeTakenCount == NewBECount && "indvars must preserve SCEV");