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/Analysis/TargetLibraryInfo.h"
35 #include "llvm/Analysis/TargetTransformInfo.h"
36 #include "llvm/IR/BasicBlock.h"
37 #include "llvm/IR/CFG.h"
38 #include "llvm/IR/Constants.h"
39 #include "llvm/IR/DataLayout.h"
40 #include "llvm/IR/Dominators.h"
41 #include "llvm/IR/Instructions.h"
42 #include "llvm/IR/IntrinsicInst.h"
43 #include "llvm/IR/LLVMContext.h"
44 #include "llvm/IR/Type.h"
45 #include "llvm/Support/CommandLine.h"
46 #include "llvm/Support/Debug.h"
47 #include "llvm/Support/raw_ostream.h"
48 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
49 #include "llvm/Transforms/Utils/Local.h"
50 #include "llvm/Transforms/Utils/SimplifyIndVar.h"
53 #define DEBUG_TYPE "indvars"
55 STATISTIC(NumWidened , "Number of indvars widened");
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 // Trip count verification can be enabled by default under NDEBUG if we
62 // implement a strong expression equivalence checker in SCEV. Until then, we
63 // use the verify-indvars flag, which may assert in some cases.
64 static cl::opt<bool> VerifyIndvars(
65 "verify-indvars", cl::Hidden,
66 cl::desc("Verify the ScalarEvolution result after running indvars"));
68 static cl::opt<bool> ReduceLiveIVs("liv-reduce", cl::Hidden,
69 cl::desc("Reduce live induction variables."));
71 enum ReplaceExitVal { NeverRepl, OnlyCheapRepl, AlwaysRepl };
73 static cl::opt<ReplaceExitVal> ReplaceExitValue(
74 "replexitval", cl::Hidden, cl::init(OnlyCheapRepl),
75 cl::desc("Choose the strategy to replace exit value in IndVarSimplify"),
76 cl::values(clEnumValN(NeverRepl, "never", "never replace exit value"),
77 clEnumValN(OnlyCheapRepl, "cheap",
78 "only replace exit value when the cost is cheap"),
79 clEnumValN(AlwaysRepl, "always",
80 "always replace exit value whenever possible"),
88 class IndVarSimplify : public LoopPass {
92 TargetLibraryInfo *TLI;
93 const TargetTransformInfo *TTI;
95 SmallVector<WeakVH, 16> DeadInsts;
99 static char ID; // Pass identification, replacement for typeid
101 : LoopPass(ID), LI(nullptr), SE(nullptr), DT(nullptr), Changed(false) {
102 initializeIndVarSimplifyPass(*PassRegistry::getPassRegistry());
105 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
107 void getAnalysisUsage(AnalysisUsage &AU) const override {
108 AU.addRequired<DominatorTreeWrapperPass>();
109 AU.addRequired<LoopInfoWrapperPass>();
110 AU.addRequired<ScalarEvolution>();
111 AU.addRequiredID(LoopSimplifyID);
112 AU.addRequiredID(LCSSAID);
113 AU.addPreserved<ScalarEvolution>();
114 AU.addPreservedID(LoopSimplifyID);
115 AU.addPreservedID(LCSSAID);
116 AU.setPreservesCFG();
120 void releaseMemory() override {
124 bool isValidRewrite(Value *FromVal, Value *ToVal);
126 void HandleFloatingPointIV(Loop *L, PHINode *PH);
127 void RewriteNonIntegerIVs(Loop *L);
129 void SimplifyAndExtend(Loop *L, SCEVExpander &Rewriter, LPPassManager &LPM);
131 bool CanLoopBeDeleted(Loop *L, SmallVector<RewritePhi, 8> &RewritePhiSet);
132 void RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter);
134 Value *LinearFunctionTestReplace(Loop *L, const SCEV *BackedgeTakenCount,
135 PHINode *IndVar, SCEVExpander &Rewriter);
137 void SinkUnusedInvariants(Loop *L);
141 char IndVarSimplify::ID = 0;
142 INITIALIZE_PASS_BEGIN(IndVarSimplify, "indvars",
143 "Induction Variable Simplification", false, false)
144 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
145 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
146 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
147 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
148 INITIALIZE_PASS_DEPENDENCY(LCSSA)
149 INITIALIZE_PASS_END(IndVarSimplify, "indvars",
150 "Induction Variable Simplification", false, false)
152 Pass *llvm::createIndVarSimplifyPass() {
153 return new IndVarSimplify();
156 /// isValidRewrite - Return true if the SCEV expansion generated by the
157 /// rewriter can replace the original value. SCEV guarantees that it
158 /// produces the same value, but the way it is produced may be illegal IR.
159 /// Ideally, this function will only be called for verification.
160 bool IndVarSimplify::isValidRewrite(Value *FromVal, Value *ToVal) {
161 // If an SCEV expression subsumed multiple pointers, its expansion could
162 // reassociate the GEP changing the base pointer. This is illegal because the
163 // final address produced by a GEP chain must be inbounds relative to its
164 // underlying object. Otherwise basic alias analysis, among other things,
165 // could fail in a dangerous way. Ultimately, SCEV will be improved to avoid
166 // producing an expression involving multiple pointers. Until then, we must
169 // Retrieve the pointer operand of the GEP. Don't use GetUnderlyingObject
170 // because it understands lcssa phis while SCEV does not.
171 Value *FromPtr = FromVal;
172 Value *ToPtr = ToVal;
173 if (GEPOperator *GEP = dyn_cast<GEPOperator>(FromVal)) {
174 FromPtr = GEP->getPointerOperand();
176 if (GEPOperator *GEP = dyn_cast<GEPOperator>(ToVal)) {
177 ToPtr = GEP->getPointerOperand();
179 if (FromPtr != FromVal || ToPtr != ToVal) {
180 // Quickly check the common case
181 if (FromPtr == ToPtr)
184 // SCEV may have rewritten an expression that produces the GEP's pointer
185 // operand. That's ok as long as the pointer operand has the same base
186 // pointer. Unlike GetUnderlyingObject(), getPointerBase() will find the
187 // base of a recurrence. This handles the case in which SCEV expansion
188 // converts a pointer type recurrence into a nonrecurrent pointer base
189 // indexed by an integer recurrence.
191 // If the GEP base pointer is a vector of pointers, abort.
192 if (!FromPtr->getType()->isPointerTy() || !ToPtr->getType()->isPointerTy())
195 const SCEV *FromBase = SE->getPointerBase(SE->getSCEV(FromPtr));
196 const SCEV *ToBase = SE->getPointerBase(SE->getSCEV(ToPtr));
197 if (FromBase == ToBase)
200 DEBUG(dbgs() << "INDVARS: GEP rewrite bail out "
201 << *FromBase << " != " << *ToBase << "\n");
208 /// Determine the insertion point for this user. By default, insert immediately
209 /// before the user. SCEVExpander or LICM will hoist loop invariants out of the
210 /// loop. For PHI nodes, there may be multiple uses, so compute the nearest
211 /// common dominator for the incoming blocks.
212 static Instruction *getInsertPointForUses(Instruction *User, Value *Def,
214 PHINode *PHI = dyn_cast<PHINode>(User);
218 Instruction *InsertPt = nullptr;
219 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i) {
220 if (PHI->getIncomingValue(i) != Def)
223 BasicBlock *InsertBB = PHI->getIncomingBlock(i);
225 InsertPt = InsertBB->getTerminator();
228 InsertBB = DT->findNearestCommonDominator(InsertPt->getParent(), InsertBB);
229 InsertPt = InsertBB->getTerminator();
231 assert(InsertPt && "Missing phi operand");
232 assert((!isa<Instruction>(Def) ||
233 DT->dominates(cast<Instruction>(Def), InsertPt)) &&
234 "def does not dominate all uses");
238 //===----------------------------------------------------------------------===//
239 // RewriteNonIntegerIVs and helpers. Prefer integer IVs.
240 //===----------------------------------------------------------------------===//
242 /// ConvertToSInt - Convert APF to an integer, if possible.
243 static bool ConvertToSInt(const APFloat &APF, int64_t &IntVal) {
244 bool isExact = false;
245 // See if we can convert this to an int64_t
247 if (APF.convertToInteger(&UIntVal, 64, true, APFloat::rmTowardZero,
248 &isExact) != APFloat::opOK || !isExact)
254 /// HandleFloatingPointIV - If the loop has floating induction variable
255 /// then insert corresponding integer induction variable if possible.
257 /// for(double i = 0; i < 10000; ++i)
259 /// is converted into
260 /// for(int i = 0; i < 10000; ++i)
263 void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PN) {
264 unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
265 unsigned BackEdge = IncomingEdge^1;
267 // Check incoming value.
268 ConstantFP *InitValueVal =
269 dyn_cast<ConstantFP>(PN->getIncomingValue(IncomingEdge));
272 if (!InitValueVal || !ConvertToSInt(InitValueVal->getValueAPF(), InitValue))
275 // Check IV increment. Reject this PN if increment operation is not
276 // an add or increment value can not be represented by an integer.
277 BinaryOperator *Incr =
278 dyn_cast<BinaryOperator>(PN->getIncomingValue(BackEdge));
279 if (Incr == nullptr || Incr->getOpcode() != Instruction::FAdd) return;
281 // If this is not an add of the PHI with a constantfp, or if the constant fp
282 // is not an integer, bail out.
283 ConstantFP *IncValueVal = dyn_cast<ConstantFP>(Incr->getOperand(1));
285 if (IncValueVal == nullptr || Incr->getOperand(0) != PN ||
286 !ConvertToSInt(IncValueVal->getValueAPF(), IncValue))
289 // Check Incr uses. One user is PN and the other user is an exit condition
290 // used by the conditional terminator.
291 Value::user_iterator IncrUse = Incr->user_begin();
292 Instruction *U1 = cast<Instruction>(*IncrUse++);
293 if (IncrUse == Incr->user_end()) return;
294 Instruction *U2 = cast<Instruction>(*IncrUse++);
295 if (IncrUse != Incr->user_end()) return;
297 // Find exit condition, which is an fcmp. If it doesn't exist, or if it isn't
298 // only used by a branch, we can't transform it.
299 FCmpInst *Compare = dyn_cast<FCmpInst>(U1);
301 Compare = dyn_cast<FCmpInst>(U2);
302 if (!Compare || !Compare->hasOneUse() ||
303 !isa<BranchInst>(Compare->user_back()))
306 BranchInst *TheBr = cast<BranchInst>(Compare->user_back());
308 // We need to verify that the branch actually controls the iteration count
309 // of the loop. If not, the new IV can overflow and no one will notice.
310 // The branch block must be in the loop and one of the successors must be out
312 assert(TheBr->isConditional() && "Can't use fcmp if not conditional");
313 if (!L->contains(TheBr->getParent()) ||
314 (L->contains(TheBr->getSuccessor(0)) &&
315 L->contains(TheBr->getSuccessor(1))))
319 // If it isn't a comparison with an integer-as-fp (the exit value), we can't
321 ConstantFP *ExitValueVal = dyn_cast<ConstantFP>(Compare->getOperand(1));
323 if (ExitValueVal == nullptr ||
324 !ConvertToSInt(ExitValueVal->getValueAPF(), ExitValue))
327 // Find new predicate for integer comparison.
328 CmpInst::Predicate NewPred = CmpInst::BAD_ICMP_PREDICATE;
329 switch (Compare->getPredicate()) {
330 default: return; // Unknown comparison.
331 case CmpInst::FCMP_OEQ:
332 case CmpInst::FCMP_UEQ: NewPred = CmpInst::ICMP_EQ; break;
333 case CmpInst::FCMP_ONE:
334 case CmpInst::FCMP_UNE: NewPred = CmpInst::ICMP_NE; break;
335 case CmpInst::FCMP_OGT:
336 case CmpInst::FCMP_UGT: NewPred = CmpInst::ICMP_SGT; break;
337 case CmpInst::FCMP_OGE:
338 case CmpInst::FCMP_UGE: NewPred = CmpInst::ICMP_SGE; break;
339 case CmpInst::FCMP_OLT:
340 case CmpInst::FCMP_ULT: NewPred = CmpInst::ICMP_SLT; break;
341 case CmpInst::FCMP_OLE:
342 case CmpInst::FCMP_ULE: NewPred = CmpInst::ICMP_SLE; break;
345 // We convert the floating point induction variable to a signed i32 value if
346 // we can. This is only safe if the comparison will not overflow in a way
347 // that won't be trapped by the integer equivalent operations. Check for this
349 // TODO: We could use i64 if it is native and the range requires it.
351 // The start/stride/exit values must all fit in signed i32.
352 if (!isInt<32>(InitValue) || !isInt<32>(IncValue) || !isInt<32>(ExitValue))
355 // If not actually striding (add x, 0.0), avoid touching the code.
359 // Positive and negative strides have different safety conditions.
361 // If we have a positive stride, we require the init to be less than the
363 if (InitValue >= ExitValue)
366 uint32_t Range = uint32_t(ExitValue-InitValue);
367 // Check for infinite loop, either:
368 // while (i <= Exit) or until (i > Exit)
369 if (NewPred == CmpInst::ICMP_SLE || NewPred == CmpInst::ICMP_SGT) {
370 if (++Range == 0) return; // Range overflows.
373 unsigned Leftover = Range % uint32_t(IncValue);
375 // If this is an equality comparison, we require that the strided value
376 // exactly land on the exit value, otherwise the IV condition will wrap
377 // around and do things the fp IV wouldn't.
378 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
382 // If the stride would wrap around the i32 before exiting, we can't
384 if (Leftover != 0 && int32_t(ExitValue+IncValue) < ExitValue)
388 // If we have a negative stride, we require the init to be greater than the
390 if (InitValue <= ExitValue)
393 uint32_t Range = uint32_t(InitValue-ExitValue);
394 // Check for infinite loop, either:
395 // while (i >= Exit) or until (i < Exit)
396 if (NewPred == CmpInst::ICMP_SGE || NewPred == CmpInst::ICMP_SLT) {
397 if (++Range == 0) return; // Range overflows.
400 unsigned Leftover = Range % uint32_t(-IncValue);
402 // If this is an equality comparison, we require that the strided value
403 // exactly land on the exit value, otherwise the IV condition will wrap
404 // around and do things the fp IV wouldn't.
405 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
409 // If the stride would wrap around the i32 before exiting, we can't
411 if (Leftover != 0 && int32_t(ExitValue+IncValue) > ExitValue)
415 IntegerType *Int32Ty = Type::getInt32Ty(PN->getContext());
417 // Insert new integer induction variable.
418 PHINode *NewPHI = PHINode::Create(Int32Ty, 2, PN->getName()+".int", PN);
419 NewPHI->addIncoming(ConstantInt::get(Int32Ty, InitValue),
420 PN->getIncomingBlock(IncomingEdge));
423 BinaryOperator::CreateAdd(NewPHI, ConstantInt::get(Int32Ty, IncValue),
424 Incr->getName()+".int", Incr);
425 NewPHI->addIncoming(NewAdd, PN->getIncomingBlock(BackEdge));
427 ICmpInst *NewCompare = new ICmpInst(TheBr, NewPred, NewAdd,
428 ConstantInt::get(Int32Ty, ExitValue),
431 // In the following deletions, PN may become dead and may be deleted.
432 // Use a WeakVH to observe whether this happens.
435 // Delete the old floating point exit comparison. The branch starts using the
437 NewCompare->takeName(Compare);
438 Compare->replaceAllUsesWith(NewCompare);
439 RecursivelyDeleteTriviallyDeadInstructions(Compare, TLI);
441 // Delete the old floating point increment.
442 Incr->replaceAllUsesWith(UndefValue::get(Incr->getType()));
443 RecursivelyDeleteTriviallyDeadInstructions(Incr, TLI);
445 // If the FP induction variable still has uses, this is because something else
446 // in the loop uses its value. In order to canonicalize the induction
447 // variable, we chose to eliminate the IV and rewrite it in terms of an
450 // We give preference to sitofp over uitofp because it is faster on most
453 Value *Conv = new SIToFPInst(NewPHI, PN->getType(), "indvar.conv",
454 PN->getParent()->getFirstInsertionPt());
455 PN->replaceAllUsesWith(Conv);
456 RecursivelyDeleteTriviallyDeadInstructions(PN, TLI);
461 void IndVarSimplify::RewriteNonIntegerIVs(Loop *L) {
462 // First step. Check to see if there are any floating-point recurrences.
463 // If there are, change them into integer recurrences, permitting analysis by
464 // the SCEV routines.
466 BasicBlock *Header = L->getHeader();
468 SmallVector<WeakVH, 8> PHIs;
469 for (BasicBlock::iterator I = Header->begin();
470 PHINode *PN = dyn_cast<PHINode>(I); ++I)
473 for (unsigned i = 0, e = PHIs.size(); i != e; ++i)
474 if (PHINode *PN = dyn_cast_or_null<PHINode>(&*PHIs[i]))
475 HandleFloatingPointIV(L, PN);
477 // If the loop previously had floating-point IV, ScalarEvolution
478 // may not have been able to compute a trip count. Now that we've done some
479 // re-writing, the trip count may be computable.
485 // Collect information about PHI nodes which can be transformed in
486 // RewriteLoopExitValues.
489 unsigned Ith; // Ith incoming value.
490 Value *Val; // Exit value after expansion.
491 bool HighCost; // High Cost when expansion.
492 bool SafePhi; // LCSSASafePhiForRAUW.
494 RewritePhi(PHINode *P, unsigned I, Value *V, bool H, bool S)
495 : PN(P), Ith(I), Val(V), HighCost(H), SafePhi(S) {}
499 //===----------------------------------------------------------------------===//
500 // RewriteLoopExitValues - Optimize IV users outside the loop.
501 // As a side effect, reduces the amount of IV processing within the loop.
502 //===----------------------------------------------------------------------===//
504 /// RewriteLoopExitValues - Check to see if this loop has a computable
505 /// loop-invariant execution count. If so, this means that we can compute the
506 /// final value of any expressions that are recurrent in the loop, and
507 /// substitute the exit values from the loop into any instructions outside of
508 /// the loop that use the final values of the current expressions.
510 /// This is mostly redundant with the regular IndVarSimplify activities that
511 /// happen later, except that it's more powerful in some cases, because it's
512 /// able to brute-force evaluate arbitrary instructions as long as they have
513 /// constant operands at the beginning of the loop.
514 void IndVarSimplify::RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter) {
515 // Verify the input to the pass in already in LCSSA form.
516 assert(L->isLCSSAForm(*DT));
518 SmallVector<BasicBlock*, 8> ExitBlocks;
519 L->getUniqueExitBlocks(ExitBlocks);
521 SmallVector<RewritePhi, 8> RewritePhiSet;
522 // Find all values that are computed inside the loop, but used outside of it.
523 // Because of LCSSA, these values will only occur in LCSSA PHI Nodes. Scan
524 // the exit blocks of the loop to find them.
525 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
526 BasicBlock *ExitBB = ExitBlocks[i];
528 // If there are no PHI nodes in this exit block, then no values defined
529 // inside the loop are used on this path, skip it.
530 PHINode *PN = dyn_cast<PHINode>(ExitBB->begin());
533 unsigned NumPreds = PN->getNumIncomingValues();
535 // We would like to be able to RAUW single-incoming value PHI nodes. We
536 // have to be certain this is safe even when this is an LCSSA PHI node.
537 // While the computed exit value is no longer varying in *this* loop, the
538 // exit block may be an exit block for an outer containing loop as well,
539 // the exit value may be varying in the outer loop, and thus it may still
540 // require an LCSSA PHI node. The safe case is when this is
541 // single-predecessor PHI node (LCSSA) and the exit block containing it is
542 // part of the enclosing loop, or this is the outer most loop of the nest.
543 // In either case the exit value could (at most) be varying in the same
544 // loop body as the phi node itself. Thus if it is in turn used outside of
545 // an enclosing loop it will only be via a separate LCSSA node.
546 bool LCSSASafePhiForRAUW =
548 (!L->getParentLoop() || L->getParentLoop() == LI->getLoopFor(ExitBB));
550 // Iterate over all of the PHI nodes.
551 BasicBlock::iterator BBI = ExitBB->begin();
552 while ((PN = dyn_cast<PHINode>(BBI++))) {
554 continue; // dead use, don't replace it
556 // SCEV only supports integer expressions for now.
557 if (!PN->getType()->isIntegerTy() && !PN->getType()->isPointerTy())
560 // It's necessary to tell ScalarEvolution about this explicitly so that
561 // it can walk the def-use list and forget all SCEVs, as it may not be
562 // watching the PHI itself. Once the new exit value is in place, there
563 // may not be a def-use connection between the loop and every instruction
564 // which got a SCEVAddRecExpr for that loop.
567 // Iterate over all of the values in all the PHI nodes.
568 for (unsigned i = 0; i != NumPreds; ++i) {
569 // If the value being merged in is not integer or is not defined
570 // in the loop, skip it.
571 Value *InVal = PN->getIncomingValue(i);
572 if (!isa<Instruction>(InVal))
575 // If this pred is for a subloop, not L itself, skip it.
576 if (LI->getLoopFor(PN->getIncomingBlock(i)) != L)
577 continue; // The Block is in a subloop, skip it.
579 // Check that InVal is defined in the loop.
580 Instruction *Inst = cast<Instruction>(InVal);
581 if (!L->contains(Inst))
584 // Okay, this instruction has a user outside of the current loop
585 // and varies predictably *inside* the loop. Evaluate the value it
586 // contains when the loop exits, if possible.
587 const SCEV *ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop());
588 if (!SE->isLoopInvariant(ExitValue, L) ||
589 !isSafeToExpand(ExitValue, *SE))
592 // Computing the value outside of the loop brings no benefit if :
593 // - it is definitely used inside the loop in a way which can not be
595 // - no use outside of the loop can take advantage of hoisting the
596 // computation out of the loop
597 if (ExitValue->getSCEVType()>=scMulExpr) {
598 unsigned NumHardInternalUses = 0;
599 unsigned NumSoftExternalUses = 0;
600 unsigned NumUses = 0;
601 for (auto IB = Inst->user_begin(), IE = Inst->user_end();
602 IB != IE && NumUses <= 6; ++IB) {
603 Instruction *UseInstr = cast<Instruction>(*IB);
604 unsigned Opc = UseInstr->getOpcode();
606 if (L->contains(UseInstr)) {
607 if (Opc == Instruction::Call || Opc == Instruction::Ret)
608 NumHardInternalUses++;
610 if (Opc == Instruction::PHI) {
611 // Do not count the Phi as a use. LCSSA may have inserted
612 // plenty of trivial ones.
614 for (auto PB = UseInstr->user_begin(),
615 PE = UseInstr->user_end();
616 PB != PE && NumUses <= 6; ++PB, ++NumUses) {
617 unsigned PhiOpc = cast<Instruction>(*PB)->getOpcode();
618 if (PhiOpc != Instruction::Call && PhiOpc != Instruction::Ret)
619 NumSoftExternalUses++;
623 if (Opc != Instruction::Call && Opc != Instruction::Ret)
624 NumSoftExternalUses++;
627 if (NumUses <= 6 && NumHardInternalUses && !NumSoftExternalUses)
631 Value *ExitVal = Rewriter.expandCodeFor(ExitValue, PN->getType(), Inst);
633 DEBUG(dbgs() << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal << '\n'
634 << " LoopVal = " << *Inst << "\n");
636 if (!isValidRewrite(Inst, ExitVal)) {
637 DeadInsts.push_back(ExitVal);
640 bool HighCost = Rewriter.isHighCostExpansion(ExitValue, L);
642 // Collect all the candidate PHINodes to be rewritten.
643 RewritePhiSet.push_back(
644 RewritePhi(PN, i, ExitVal, HighCost, LCSSASafePhiForRAUW));
649 bool LoopCanBeDel = CanLoopBeDeleted(L, RewritePhiSet);
652 for (const RewritePhi &Phi : RewritePhiSet) {
653 PHINode *PN = Phi.PN;
654 Value *ExitVal = Phi.Val;
656 // Only do the rewrite when the ExitValue can be expanded cheaply.
657 // If LoopCanBeDel is true, rewrite exit value aggressively.
658 if (ReplaceExitValue == OnlyCheapRepl && !LoopCanBeDel && Phi.HighCost) {
659 DeadInsts.push_back(ExitVal);
665 Instruction *Inst = cast<Instruction>(PN->getIncomingValue(Phi.Ith));
666 PN->setIncomingValue(Phi.Ith, ExitVal);
668 // If this instruction is dead now, delete it. Don't do it now to avoid
669 // invalidating iterators.
670 if (isInstructionTriviallyDead(Inst, TLI))
671 DeadInsts.push_back(Inst);
673 // If we determined that this PHI is safe to replace even if an LCSSA
676 PN->replaceAllUsesWith(ExitVal);
677 PN->eraseFromParent();
681 // The insertion point instruction may have been deleted; clear it out
682 // so that the rewriter doesn't trip over it later.
683 Rewriter.clearInsertPoint();
686 /// CanLoopBeDeleted - Check whether it is possible to delete the loop after
687 /// rewriting exit value. If it is possible, ignore ReplaceExitValue and
688 /// do rewriting aggressively.
689 bool IndVarSimplify::CanLoopBeDeleted(
690 Loop *L, SmallVector<RewritePhi, 8> &RewritePhiSet) {
692 BasicBlock *Preheader = L->getLoopPreheader();
693 // If there is no preheader, the loop will not be deleted.
697 // In LoopDeletion pass Loop can be deleted when ExitingBlocks.size() > 1.
698 // We obviate multiple ExitingBlocks case for simplicity.
699 // TODO: If we see testcase with multiple ExitingBlocks can be deleted
700 // after exit value rewriting, we can enhance the logic here.
701 SmallVector<BasicBlock *, 4> ExitingBlocks;
702 L->getExitingBlocks(ExitingBlocks);
703 SmallVector<BasicBlock *, 8> ExitBlocks;
704 L->getUniqueExitBlocks(ExitBlocks);
705 if (ExitBlocks.size() > 1 || ExitingBlocks.size() > 1)
708 BasicBlock *ExitBlock = ExitBlocks[0];
709 BasicBlock::iterator BI = ExitBlock->begin();
710 while (PHINode *P = dyn_cast<PHINode>(BI)) {
711 Value *Incoming = P->getIncomingValueForBlock(ExitingBlocks[0]);
713 // If the Incoming value of P is found in RewritePhiSet, we know it
714 // could be rewritten to use a loop invariant value in transformation
715 // phase later. Skip it in the loop invariant check below.
717 for (const RewritePhi &Phi : RewritePhiSet) {
718 unsigned i = Phi.Ith;
719 if (Phi.PN == P && (Phi.PN)->getIncomingValue(i) == Incoming) {
726 if (!found && (I = dyn_cast<Instruction>(Incoming)))
727 if (!L->hasLoopInvariantOperands(I))
733 for (Loop::block_iterator LI = L->block_begin(), LE = L->block_end();
735 for (BasicBlock::iterator BI = (*LI)->begin(), BE = (*LI)->end(); BI != BE;
737 if (BI->mayHaveSideEffects())
745 //===----------------------------------------------------------------------===//
746 // IV Widening - Extend the width of an IV to cover its widest uses.
747 //===----------------------------------------------------------------------===//
750 // Collect information about induction variables that are used by sign/zero
751 // extend operations. This information is recorded by CollectExtend and
752 // provides the input to WidenIV.
755 Type *WidestNativeType; // Widest integer type created [sz]ext
756 bool IsSigned; // Was a sext user seen before a zext?
758 WideIVInfo() : NarrowIV(nullptr), WidestNativeType(nullptr),
763 /// visitCast - Update information about the induction variable that is
764 /// extended by this sign or zero extend operation. This is used to determine
765 /// the final width of the IV before actually widening it.
766 static void visitIVCast(CastInst *Cast, WideIVInfo &WI, ScalarEvolution *SE,
767 const TargetTransformInfo *TTI) {
768 bool IsSigned = Cast->getOpcode() == Instruction::SExt;
769 if (!IsSigned && Cast->getOpcode() != Instruction::ZExt)
772 Type *Ty = Cast->getType();
773 uint64_t Width = SE->getTypeSizeInBits(Ty);
774 if (!Cast->getModule()->getDataLayout().isLegalInteger(Width))
777 // Cast is either an sext or zext up to this point.
778 // We should not widen an indvar if arithmetics on the wider indvar are more
779 // expensive than those on the narrower indvar. We check only the cost of ADD
780 // because at least an ADD is required to increment the induction variable. We
781 // could compute more comprehensively the cost of all instructions on the
782 // induction variable when necessary.
784 TTI->getArithmeticInstrCost(Instruction::Add, Ty) >
785 TTI->getArithmeticInstrCost(Instruction::Add,
786 Cast->getOperand(0)->getType())) {
790 if (!WI.WidestNativeType) {
791 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
792 WI.IsSigned = IsSigned;
796 // We extend the IV to satisfy the sign of its first user, arbitrarily.
797 if (WI.IsSigned != IsSigned)
800 if (Width > SE->getTypeSizeInBits(WI.WidestNativeType))
801 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
806 /// NarrowIVDefUse - Record a link in the Narrow IV def-use chain along with the
807 /// WideIV that computes the same value as the Narrow IV def. This avoids
808 /// caching Use* pointers.
809 struct NarrowIVDefUse {
810 Instruction *NarrowDef;
811 Instruction *NarrowUse;
812 Instruction *WideDef;
814 NarrowIVDefUse(): NarrowDef(nullptr), NarrowUse(nullptr), WideDef(nullptr) {}
816 NarrowIVDefUse(Instruction *ND, Instruction *NU, Instruction *WD):
817 NarrowDef(ND), NarrowUse(NU), WideDef(WD) {}
820 /// WidenIV - The goal of this transform is to remove sign and zero extends
821 /// without creating any new induction variables. To do this, it creates a new
822 /// phi of the wider type and redirects all users, either removing extends or
823 /// inserting truncs whenever we stop propagating the type.
839 Instruction *WideInc;
840 const SCEV *WideIncExpr;
841 SmallVectorImpl<WeakVH> &DeadInsts;
843 SmallPtrSet<Instruction*,16> Widened;
844 SmallVector<NarrowIVDefUse, 8> NarrowIVUsers;
847 WidenIV(const WideIVInfo &WI, LoopInfo *LInfo,
848 ScalarEvolution *SEv, DominatorTree *DTree,
849 SmallVectorImpl<WeakVH> &DI) :
850 OrigPhi(WI.NarrowIV),
851 WideType(WI.WidestNativeType),
852 IsSigned(WI.IsSigned),
854 L(LI->getLoopFor(OrigPhi->getParent())),
859 WideIncExpr(nullptr),
861 assert(L->getHeader() == OrigPhi->getParent() && "Phi must be an IV");
864 PHINode *CreateWideIV(SCEVExpander &Rewriter);
867 Value *getExtend(Value *NarrowOper, Type *WideType, bool IsSigned,
870 Instruction *CloneIVUser(NarrowIVDefUse DU);
872 const SCEVAddRecExpr *GetWideRecurrence(Instruction *NarrowUse);
874 const SCEVAddRecExpr* GetExtendedOperandRecurrence(NarrowIVDefUse DU);
876 const SCEV *GetSCEVByOpCode(const SCEV *LHS, const SCEV *RHS,
877 unsigned OpCode) const;
879 Instruction *WidenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter);
881 bool WidenLoopCompare(NarrowIVDefUse DU);
883 void pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef);
885 } // anonymous namespace
887 /// isLoopInvariant - Perform a quick domtree based check for loop invariance
888 /// assuming that V is used within the loop. LoopInfo::isLoopInvariant() seems
889 /// gratuitous for this purpose.
890 static bool isLoopInvariant(Value *V, const Loop *L, const DominatorTree *DT) {
891 Instruction *Inst = dyn_cast<Instruction>(V);
895 return DT->properlyDominates(Inst->getParent(), L->getHeader());
898 Value *WidenIV::getExtend(Value *NarrowOper, Type *WideType, bool IsSigned,
900 // Set the debug location and conservative insertion point.
901 IRBuilder<> Builder(Use);
902 // Hoist the insertion point into loop preheaders as far as possible.
903 for (const Loop *L = LI->getLoopFor(Use->getParent());
904 L && L->getLoopPreheader() && isLoopInvariant(NarrowOper, L, DT);
905 L = L->getParentLoop())
906 Builder.SetInsertPoint(L->getLoopPreheader()->getTerminator());
908 return IsSigned ? Builder.CreateSExt(NarrowOper, WideType) :
909 Builder.CreateZExt(NarrowOper, WideType);
912 /// CloneIVUser - Instantiate a wide operation to replace a narrow
913 /// operation. This only needs to handle operations that can evaluation to
914 /// SCEVAddRec. It can safely return 0 for any operation we decide not to clone.
915 Instruction *WidenIV::CloneIVUser(NarrowIVDefUse DU) {
916 unsigned Opcode = DU.NarrowUse->getOpcode();
920 case Instruction::Add:
921 case Instruction::Mul:
922 case Instruction::UDiv:
923 case Instruction::Sub:
924 case Instruction::And:
925 case Instruction::Or:
926 case Instruction::Xor:
927 case Instruction::Shl:
928 case Instruction::LShr:
929 case Instruction::AShr:
930 DEBUG(dbgs() << "Cloning IVUser: " << *DU.NarrowUse << "\n");
932 // Replace NarrowDef operands with WideDef. Otherwise, we don't know
933 // anything about the narrow operand yet so must insert a [sz]ext. It is
934 // probably loop invariant and will be folded or hoisted. If it actually
935 // comes from a widened IV, it should be removed during a future call to
937 Value *LHS = (DU.NarrowUse->getOperand(0) == DU.NarrowDef) ? DU.WideDef :
938 getExtend(DU.NarrowUse->getOperand(0), WideType, IsSigned, DU.NarrowUse);
939 Value *RHS = (DU.NarrowUse->getOperand(1) == DU.NarrowDef) ? DU.WideDef :
940 getExtend(DU.NarrowUse->getOperand(1), WideType, IsSigned, DU.NarrowUse);
942 BinaryOperator *NarrowBO = cast<BinaryOperator>(DU.NarrowUse);
943 BinaryOperator *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(),
945 NarrowBO->getName());
946 IRBuilder<> Builder(DU.NarrowUse);
947 Builder.Insert(WideBO);
948 if (const OverflowingBinaryOperator *OBO =
949 dyn_cast<OverflowingBinaryOperator>(NarrowBO)) {
950 if (OBO->hasNoUnsignedWrap()) WideBO->setHasNoUnsignedWrap();
951 if (OBO->hasNoSignedWrap()) WideBO->setHasNoSignedWrap();
957 const SCEV *WidenIV::GetSCEVByOpCode(const SCEV *LHS, const SCEV *RHS,
958 unsigned OpCode) const {
959 if (OpCode == Instruction::Add)
960 return SE->getAddExpr(LHS, RHS);
961 if (OpCode == Instruction::Sub)
962 return SE->getMinusSCEV(LHS, RHS);
963 if (OpCode == Instruction::Mul)
964 return SE->getMulExpr(LHS, RHS);
966 llvm_unreachable("Unsupported opcode.");
969 /// No-wrap operations can transfer sign extension of their result to their
970 /// operands. Generate the SCEV value for the widened operation without
971 /// actually modifying the IR yet. If the expression after extending the
972 /// operands is an AddRec for this loop, return it.
973 const SCEVAddRecExpr* WidenIV::GetExtendedOperandRecurrence(NarrowIVDefUse DU) {
975 // Handle the common case of add<nsw/nuw>
976 const unsigned OpCode = DU.NarrowUse->getOpcode();
977 // Only Add/Sub/Mul instructions supported yet.
978 if (OpCode != Instruction::Add && OpCode != Instruction::Sub &&
979 OpCode != Instruction::Mul)
982 // One operand (NarrowDef) has already been extended to WideDef. Now determine
983 // if extending the other will lead to a recurrence.
984 const unsigned ExtendOperIdx =
985 DU.NarrowUse->getOperand(0) == DU.NarrowDef ? 1 : 0;
986 assert(DU.NarrowUse->getOperand(1-ExtendOperIdx) == DU.NarrowDef && "bad DU");
988 const SCEV *ExtendOperExpr = nullptr;
989 const OverflowingBinaryOperator *OBO =
990 cast<OverflowingBinaryOperator>(DU.NarrowUse);
991 if (IsSigned && OBO->hasNoSignedWrap())
992 ExtendOperExpr = SE->getSignExtendExpr(
993 SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType);
994 else if(!IsSigned && OBO->hasNoUnsignedWrap())
995 ExtendOperExpr = SE->getZeroExtendExpr(
996 SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType);
1000 // When creating this SCEV expr, don't apply the current operations NSW or NUW
1001 // flags. This instruction may be guarded by control flow that the no-wrap
1002 // behavior depends on. Non-control-equivalent instructions can be mapped to
1003 // the same SCEV expression, and it would be incorrect to transfer NSW/NUW
1004 // semantics to those operations.
1005 const SCEV *lhs = SE->getSCEV(DU.WideDef);
1006 const SCEV *rhs = ExtendOperExpr;
1008 // Let's swap operands to the initial order for the case of non-commutative
1009 // operations, like SUB. See PR21014.
1010 if (ExtendOperIdx == 0)
1011 std::swap(lhs, rhs);
1012 const SCEVAddRecExpr *AddRec =
1013 dyn_cast<SCEVAddRecExpr>(GetSCEVByOpCode(lhs, rhs, OpCode));
1015 if (!AddRec || AddRec->getLoop() != L)
1020 /// GetWideRecurrence - Is this instruction potentially interesting for further
1021 /// simplification after widening it's type? In other words, can the
1022 /// extend be safely hoisted out of the loop with SCEV reducing the value to a
1023 /// recurrence on the same loop. If so, return the sign or zero extended
1024 /// recurrence. Otherwise return NULL.
1025 const SCEVAddRecExpr *WidenIV::GetWideRecurrence(Instruction *NarrowUse) {
1026 if (!SE->isSCEVable(NarrowUse->getType()))
1029 const SCEV *NarrowExpr = SE->getSCEV(NarrowUse);
1030 if (SE->getTypeSizeInBits(NarrowExpr->getType())
1031 >= SE->getTypeSizeInBits(WideType)) {
1032 // NarrowUse implicitly widens its operand. e.g. a gep with a narrow
1033 // index. So don't follow this use.
1037 const SCEV *WideExpr = IsSigned ?
1038 SE->getSignExtendExpr(NarrowExpr, WideType) :
1039 SE->getZeroExtendExpr(NarrowExpr, WideType);
1040 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(WideExpr);
1041 if (!AddRec || AddRec->getLoop() != L)
1046 /// This IV user cannot be widen. Replace this use of the original narrow IV
1047 /// with a truncation of the new wide IV to isolate and eliminate the narrow IV.
1048 static void truncateIVUse(NarrowIVDefUse DU, DominatorTree *DT) {
1049 DEBUG(dbgs() << "INDVARS: Truncate IV " << *DU.WideDef
1050 << " for user " << *DU.NarrowUse << "\n");
1051 IRBuilder<> Builder(getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT));
1052 Value *Trunc = Builder.CreateTrunc(DU.WideDef, DU.NarrowDef->getType());
1053 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, Trunc);
1056 /// If the narrow use is a compare instruction, then widen the compare
1057 // (and possibly the other operand). The extend operation is hoisted into the
1058 // loop preheader as far as possible.
1059 bool WidenIV::WidenLoopCompare(NarrowIVDefUse DU) {
1060 ICmpInst *Cmp = dyn_cast<ICmpInst>(DU.NarrowUse);
1064 // Sign of IV user and compare must match.
1065 if (IsSigned != CmpInst::isSigned(Cmp->getPredicate()))
1068 Value *Op = Cmp->getOperand(Cmp->getOperand(0) == DU.NarrowDef ? 1 : 0);
1069 unsigned CastWidth = SE->getTypeSizeInBits(Op->getType());
1070 unsigned IVWidth = SE->getTypeSizeInBits(WideType);
1071 assert (CastWidth <= IVWidth && "Unexpected width while widening compare.");
1073 // Widen the compare instruction.
1074 IRBuilder<> Builder(getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT));
1075 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef);
1077 // Widen the other operand of the compare, if necessary.
1078 if (CastWidth < IVWidth) {
1079 Value *ExtOp = getExtend(Op, WideType, IsSigned, Cmp);
1080 DU.NarrowUse->replaceUsesOfWith(Op, ExtOp);
1085 /// WidenIVUse - Determine whether an individual user of the narrow IV can be
1086 /// widened. If so, return the wide clone of the user.
1087 Instruction *WidenIV::WidenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter) {
1089 // Stop traversing the def-use chain at inner-loop phis or post-loop phis.
1090 if (PHINode *UsePhi = dyn_cast<PHINode>(DU.NarrowUse)) {
1091 if (LI->getLoopFor(UsePhi->getParent()) != L) {
1092 // For LCSSA phis, sink the truncate outside the loop.
1093 // After SimplifyCFG most loop exit targets have a single predecessor.
1094 // Otherwise fall back to a truncate within the loop.
1095 if (UsePhi->getNumOperands() != 1)
1096 truncateIVUse(DU, DT);
1099 PHINode::Create(DU.WideDef->getType(), 1, UsePhi->getName() + ".wide",
1101 WidePhi->addIncoming(DU.WideDef, UsePhi->getIncomingBlock(0));
1102 IRBuilder<> Builder(WidePhi->getParent()->getFirstInsertionPt());
1103 Value *Trunc = Builder.CreateTrunc(WidePhi, DU.NarrowDef->getType());
1104 UsePhi->replaceAllUsesWith(Trunc);
1105 DeadInsts.emplace_back(UsePhi);
1106 DEBUG(dbgs() << "INDVARS: Widen lcssa phi " << *UsePhi
1107 << " to " << *WidePhi << "\n");
1112 // Our raison d'etre! Eliminate sign and zero extension.
1113 if (IsSigned ? isa<SExtInst>(DU.NarrowUse) : isa<ZExtInst>(DU.NarrowUse)) {
1114 Value *NewDef = DU.WideDef;
1115 if (DU.NarrowUse->getType() != WideType) {
1116 unsigned CastWidth = SE->getTypeSizeInBits(DU.NarrowUse->getType());
1117 unsigned IVWidth = SE->getTypeSizeInBits(WideType);
1118 if (CastWidth < IVWidth) {
1119 // The cast isn't as wide as the IV, so insert a Trunc.
1120 IRBuilder<> Builder(DU.NarrowUse);
1121 NewDef = Builder.CreateTrunc(DU.WideDef, DU.NarrowUse->getType());
1124 // A wider extend was hidden behind a narrower one. This may induce
1125 // another round of IV widening in which the intermediate IV becomes
1126 // dead. It should be very rare.
1127 DEBUG(dbgs() << "INDVARS: New IV " << *WidePhi
1128 << " not wide enough to subsume " << *DU.NarrowUse << "\n");
1129 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef);
1130 NewDef = DU.NarrowUse;
1133 if (NewDef != DU.NarrowUse) {
1134 DEBUG(dbgs() << "INDVARS: eliminating " << *DU.NarrowUse
1135 << " replaced by " << *DU.WideDef << "\n");
1137 DU.NarrowUse->replaceAllUsesWith(NewDef);
1138 DeadInsts.emplace_back(DU.NarrowUse);
1140 // Now that the extend is gone, we want to expose it's uses for potential
1141 // further simplification. We don't need to directly inform SimplifyIVUsers
1142 // of the new users, because their parent IV will be processed later as a
1143 // new loop phi. If we preserved IVUsers analysis, we would also want to
1144 // push the uses of WideDef here.
1146 // No further widening is needed. The deceased [sz]ext had done it for us.
1150 // Does this user itself evaluate to a recurrence after widening?
1151 const SCEVAddRecExpr *WideAddRec = GetWideRecurrence(DU.NarrowUse);
1153 WideAddRec = GetExtendedOperandRecurrence(DU);
1156 // If use is a loop condition, try to promote the condition instead of
1157 // truncating the IV first.
1158 if (WidenLoopCompare(DU))
1161 // This user does not evaluate to a recurence after widening, so don't
1162 // follow it. Instead insert a Trunc to kill off the original use,
1163 // eventually isolating the original narrow IV so it can be removed.
1164 truncateIVUse(DU, DT);
1167 // Assume block terminators cannot evaluate to a recurrence. We can't to
1168 // insert a Trunc after a terminator if there happens to be a critical edge.
1169 assert(DU.NarrowUse != DU.NarrowUse->getParent()->getTerminator() &&
1170 "SCEV is not expected to evaluate a block terminator");
1172 // Reuse the IV increment that SCEVExpander created as long as it dominates
1174 Instruction *WideUse = nullptr;
1175 if (WideAddRec == WideIncExpr
1176 && Rewriter.hoistIVInc(WideInc, DU.NarrowUse))
1179 WideUse = CloneIVUser(DU);
1183 // Evaluation of WideAddRec ensured that the narrow expression could be
1184 // extended outside the loop without overflow. This suggests that the wide use
1185 // evaluates to the same expression as the extended narrow use, but doesn't
1186 // absolutely guarantee it. Hence the following failsafe check. In rare cases
1187 // where it fails, we simply throw away the newly created wide use.
1188 if (WideAddRec != SE->getSCEV(WideUse)) {
1189 DEBUG(dbgs() << "Wide use expression mismatch: " << *WideUse
1190 << ": " << *SE->getSCEV(WideUse) << " != " << *WideAddRec << "\n");
1191 DeadInsts.emplace_back(WideUse);
1195 // Returning WideUse pushes it on the worklist.
1199 /// pushNarrowIVUsers - Add eligible users of NarrowDef to NarrowIVUsers.
1201 void WidenIV::pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef) {
1202 for (User *U : NarrowDef->users()) {
1203 Instruction *NarrowUser = cast<Instruction>(U);
1205 // Handle data flow merges and bizarre phi cycles.
1206 if (!Widened.insert(NarrowUser).second)
1209 NarrowIVUsers.push_back(NarrowIVDefUse(NarrowDef, NarrowUser, WideDef));
1213 /// CreateWideIV - Process a single induction variable. First use the
1214 /// SCEVExpander to create a wide induction variable that evaluates to the same
1215 /// recurrence as the original narrow IV. Then use a worklist to forward
1216 /// traverse the narrow IV's def-use chain. After WidenIVUse has processed all
1217 /// interesting IV users, the narrow IV will be isolated for removal by
1220 /// It would be simpler to delete uses as they are processed, but we must avoid
1221 /// invalidating SCEV expressions.
1223 PHINode *WidenIV::CreateWideIV(SCEVExpander &Rewriter) {
1224 // Is this phi an induction variable?
1225 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(OrigPhi));
1229 // Widen the induction variable expression.
1230 const SCEV *WideIVExpr = IsSigned ?
1231 SE->getSignExtendExpr(AddRec, WideType) :
1232 SE->getZeroExtendExpr(AddRec, WideType);
1234 assert(SE->getEffectiveSCEVType(WideIVExpr->getType()) == WideType &&
1235 "Expect the new IV expression to preserve its type");
1237 // Can the IV be extended outside the loop without overflow?
1238 AddRec = dyn_cast<SCEVAddRecExpr>(WideIVExpr);
1239 if (!AddRec || AddRec->getLoop() != L)
1242 // An AddRec must have loop-invariant operands. Since this AddRec is
1243 // materialized by a loop header phi, the expression cannot have any post-loop
1244 // operands, so they must dominate the loop header.
1245 assert(SE->properlyDominates(AddRec->getStart(), L->getHeader()) &&
1246 SE->properlyDominates(AddRec->getStepRecurrence(*SE), L->getHeader())
1247 && "Loop header phi recurrence inputs do not dominate the loop");
1249 // The rewriter provides a value for the desired IV expression. This may
1250 // either find an existing phi or materialize a new one. Either way, we
1251 // expect a well-formed cyclic phi-with-increments. i.e. any operand not part
1252 // of the phi-SCC dominates the loop entry.
1253 Instruction *InsertPt = L->getHeader()->begin();
1254 WidePhi = cast<PHINode>(Rewriter.expandCodeFor(AddRec, WideType, InsertPt));
1256 // Remembering the WideIV increment generated by SCEVExpander allows
1257 // WidenIVUse to reuse it when widening the narrow IV's increment. We don't
1258 // employ a general reuse mechanism because the call above is the only call to
1259 // SCEVExpander. Henceforth, we produce 1-to-1 narrow to wide uses.
1260 if (BasicBlock *LatchBlock = L->getLoopLatch()) {
1262 cast<Instruction>(WidePhi->getIncomingValueForBlock(LatchBlock));
1263 WideIncExpr = SE->getSCEV(WideInc);
1266 DEBUG(dbgs() << "Wide IV: " << *WidePhi << "\n");
1269 // Traverse the def-use chain using a worklist starting at the original IV.
1270 assert(Widened.empty() && NarrowIVUsers.empty() && "expect initial state" );
1272 Widened.insert(OrigPhi);
1273 pushNarrowIVUsers(OrigPhi, WidePhi);
1275 while (!NarrowIVUsers.empty()) {
1276 NarrowIVDefUse DU = NarrowIVUsers.pop_back_val();
1278 // Process a def-use edge. This may replace the use, so don't hold a
1279 // use_iterator across it.
1280 Instruction *WideUse = WidenIVUse(DU, Rewriter);
1282 // Follow all def-use edges from the previous narrow use.
1284 pushNarrowIVUsers(DU.NarrowUse, WideUse);
1286 // WidenIVUse may have removed the def-use edge.
1287 if (DU.NarrowDef->use_empty())
1288 DeadInsts.emplace_back(DU.NarrowDef);
1293 //===----------------------------------------------------------------------===//
1294 // Live IV Reduction - Minimize IVs live across the loop.
1295 //===----------------------------------------------------------------------===//
1298 //===----------------------------------------------------------------------===//
1299 // Simplification of IV users based on SCEV evaluation.
1300 //===----------------------------------------------------------------------===//
1303 class IndVarSimplifyVisitor : public IVVisitor {
1304 ScalarEvolution *SE;
1305 const TargetTransformInfo *TTI;
1311 IndVarSimplifyVisitor(PHINode *IV, ScalarEvolution *SCEV,
1312 const TargetTransformInfo *TTI,
1313 const DominatorTree *DTree)
1314 : SE(SCEV), TTI(TTI), IVPhi(IV) {
1316 WI.NarrowIV = IVPhi;
1318 setSplitOverflowIntrinsics();
1321 // Implement the interface used by simplifyUsersOfIV.
1322 void visitCast(CastInst *Cast) override { visitIVCast(Cast, WI, SE, TTI); }
1326 /// SimplifyAndExtend - Iteratively perform simplification on a worklist of IV
1327 /// users. Each successive simplification may push more users which may
1328 /// themselves be candidates for simplification.
1330 /// Sign/Zero extend elimination is interleaved with IV simplification.
1332 void IndVarSimplify::SimplifyAndExtend(Loop *L,
1333 SCEVExpander &Rewriter,
1334 LPPassManager &LPM) {
1335 SmallVector<WideIVInfo, 8> WideIVs;
1337 SmallVector<PHINode*, 8> LoopPhis;
1338 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
1339 LoopPhis.push_back(cast<PHINode>(I));
1341 // Each round of simplification iterates through the SimplifyIVUsers worklist
1342 // for all current phis, then determines whether any IVs can be
1343 // widened. Widening adds new phis to LoopPhis, inducing another round of
1344 // simplification on the wide IVs.
1345 while (!LoopPhis.empty()) {
1346 // Evaluate as many IV expressions as possible before widening any IVs. This
1347 // forces SCEV to set no-wrap flags before evaluating sign/zero
1348 // extension. The first time SCEV attempts to normalize sign/zero extension,
1349 // the result becomes final. So for the most predictable results, we delay
1350 // evaluation of sign/zero extend evaluation until needed, and avoid running
1351 // other SCEV based analysis prior to SimplifyAndExtend.
1353 PHINode *CurrIV = LoopPhis.pop_back_val();
1355 // Information about sign/zero extensions of CurrIV.
1356 IndVarSimplifyVisitor Visitor(CurrIV, SE, TTI, DT);
1358 Changed |= simplifyUsersOfIV(CurrIV, SE, &LPM, DeadInsts, &Visitor);
1360 if (Visitor.WI.WidestNativeType) {
1361 WideIVs.push_back(Visitor.WI);
1363 } while(!LoopPhis.empty());
1365 for (; !WideIVs.empty(); WideIVs.pop_back()) {
1366 WidenIV Widener(WideIVs.back(), LI, SE, DT, DeadInsts);
1367 if (PHINode *WidePhi = Widener.CreateWideIV(Rewriter)) {
1369 LoopPhis.push_back(WidePhi);
1375 //===----------------------------------------------------------------------===//
1376 // LinearFunctionTestReplace and its kin. Rewrite the loop exit condition.
1377 //===----------------------------------------------------------------------===//
1379 /// canExpandBackedgeTakenCount - Return true if this loop's backedge taken
1380 /// count expression can be safely and cheaply expanded into an instruction
1381 /// sequence that can be used by LinearFunctionTestReplace.
1383 /// TODO: This fails for pointer-type loop counters with greater than one byte
1384 /// strides, consequently preventing LFTR from running. For the purpose of LFTR
1385 /// we could skip this check in the case that the LFTR loop counter (chosen by
1386 /// FindLoopCounter) is also pointer type. Instead, we could directly convert
1387 /// the loop test to an inequality test by checking the target data's alignment
1388 /// of element types (given that the initial pointer value originates from or is
1389 /// used by ABI constrained operation, as opposed to inttoptr/ptrtoint).
1390 /// However, we don't yet have a strong motivation for converting loop tests
1391 /// into inequality tests.
1392 static bool canExpandBackedgeTakenCount(Loop *L, ScalarEvolution *SE,
1393 SCEVExpander &Rewriter) {
1394 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
1395 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount) ||
1396 BackedgeTakenCount->isZero())
1399 if (!L->getExitingBlock())
1402 // Can't rewrite non-branch yet.
1403 if (!isa<BranchInst>(L->getExitingBlock()->getTerminator()))
1406 if (Rewriter.isHighCostExpansion(BackedgeTakenCount, L))
1412 /// getLoopPhiForCounter - Return the loop header phi IFF IncV adds a loop
1413 /// invariant value to the phi.
1414 static PHINode *getLoopPhiForCounter(Value *IncV, Loop *L, DominatorTree *DT) {
1415 Instruction *IncI = dyn_cast<Instruction>(IncV);
1419 switch (IncI->getOpcode()) {
1420 case Instruction::Add:
1421 case Instruction::Sub:
1423 case Instruction::GetElementPtr:
1424 // An IV counter must preserve its type.
1425 if (IncI->getNumOperands() == 2)
1431 PHINode *Phi = dyn_cast<PHINode>(IncI->getOperand(0));
1432 if (Phi && Phi->getParent() == L->getHeader()) {
1433 if (isLoopInvariant(IncI->getOperand(1), L, DT))
1437 if (IncI->getOpcode() == Instruction::GetElementPtr)
1440 // Allow add/sub to be commuted.
1441 Phi = dyn_cast<PHINode>(IncI->getOperand(1));
1442 if (Phi && Phi->getParent() == L->getHeader()) {
1443 if (isLoopInvariant(IncI->getOperand(0), L, DT))
1449 /// Return the compare guarding the loop latch, or NULL for unrecognized tests.
1450 static ICmpInst *getLoopTest(Loop *L) {
1451 assert(L->getExitingBlock() && "expected loop exit");
1453 BasicBlock *LatchBlock = L->getLoopLatch();
1454 // Don't bother with LFTR if the loop is not properly simplified.
1458 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
1459 assert(BI && "expected exit branch");
1461 return dyn_cast<ICmpInst>(BI->getCondition());
1464 /// needsLFTR - LinearFunctionTestReplace policy. Return true unless we can show
1465 /// that the current exit test is already sufficiently canonical.
1466 static bool needsLFTR(Loop *L, DominatorTree *DT) {
1467 // Do LFTR to simplify the exit condition to an ICMP.
1468 ICmpInst *Cond = getLoopTest(L);
1472 // Do LFTR to simplify the exit ICMP to EQ/NE
1473 ICmpInst::Predicate Pred = Cond->getPredicate();
1474 if (Pred != ICmpInst::ICMP_NE && Pred != ICmpInst::ICMP_EQ)
1477 // Look for a loop invariant RHS
1478 Value *LHS = Cond->getOperand(0);
1479 Value *RHS = Cond->getOperand(1);
1480 if (!isLoopInvariant(RHS, L, DT)) {
1481 if (!isLoopInvariant(LHS, L, DT))
1483 std::swap(LHS, RHS);
1485 // Look for a simple IV counter LHS
1486 PHINode *Phi = dyn_cast<PHINode>(LHS);
1488 Phi = getLoopPhiForCounter(LHS, L, DT);
1493 // Do LFTR if PHI node is defined in the loop, but is *not* a counter.
1494 int Idx = Phi->getBasicBlockIndex(L->getLoopLatch());
1498 // Do LFTR if the exit condition's IV is *not* a simple counter.
1499 Value *IncV = Phi->getIncomingValue(Idx);
1500 return Phi != getLoopPhiForCounter(IncV, L, DT);
1503 /// Recursive helper for hasConcreteDef(). Unfortunately, this currently boils
1504 /// down to checking that all operands are constant and listing instructions
1505 /// that may hide undef.
1506 static bool hasConcreteDefImpl(Value *V, SmallPtrSetImpl<Value*> &Visited,
1508 if (isa<Constant>(V))
1509 return !isa<UndefValue>(V);
1514 // Conservatively handle non-constant non-instructions. For example, Arguments
1516 Instruction *I = dyn_cast<Instruction>(V);
1520 // Load and return values may be undef.
1521 if(I->mayReadFromMemory() || isa<CallInst>(I) || isa<InvokeInst>(I))
1524 // Optimistically handle other instructions.
1525 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) {
1526 if (!Visited.insert(*OI).second)
1528 if (!hasConcreteDefImpl(*OI, Visited, Depth+1))
1534 /// Return true if the given value is concrete. We must prove that undef can
1537 /// TODO: If we decide that this is a good approach to checking for undef, we
1538 /// may factor it into a common location.
1539 static bool hasConcreteDef(Value *V) {
1540 SmallPtrSet<Value*, 8> Visited;
1542 return hasConcreteDefImpl(V, Visited, 0);
1545 /// AlmostDeadIV - Return true if this IV has any uses other than the (soon to
1546 /// be rewritten) loop exit test.
1547 static bool AlmostDeadIV(PHINode *Phi, BasicBlock *LatchBlock, Value *Cond) {
1548 int LatchIdx = Phi->getBasicBlockIndex(LatchBlock);
1549 Value *IncV = Phi->getIncomingValue(LatchIdx);
1551 for (User *U : Phi->users())
1552 if (U != Cond && U != IncV) return false;
1554 for (User *U : IncV->users())
1555 if (U != Cond && U != Phi) return false;
1559 /// FindLoopCounter - Find an affine IV in canonical form.
1561 /// BECount may be an i8* pointer type. The pointer difference is already
1562 /// valid count without scaling the address stride, so it remains a pointer
1563 /// expression as far as SCEV is concerned.
1565 /// Currently only valid for LFTR. See the comments on hasConcreteDef below.
1567 /// FIXME: Accept -1 stride and set IVLimit = IVInit - BECount
1569 /// FIXME: Accept non-unit stride as long as SCEV can reduce BECount * Stride.
1570 /// This is difficult in general for SCEV because of potential overflow. But we
1571 /// could at least handle constant BECounts.
1572 static PHINode *FindLoopCounter(Loop *L, const SCEV *BECount,
1573 ScalarEvolution *SE, DominatorTree *DT) {
1574 uint64_t BCWidth = SE->getTypeSizeInBits(BECount->getType());
1577 cast<BranchInst>(L->getExitingBlock()->getTerminator())->getCondition();
1579 // Loop over all of the PHI nodes, looking for a simple counter.
1580 PHINode *BestPhi = nullptr;
1581 const SCEV *BestInit = nullptr;
1582 BasicBlock *LatchBlock = L->getLoopLatch();
1583 assert(LatchBlock && "needsLFTR should guarantee a loop latch");
1585 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
1586 PHINode *Phi = cast<PHINode>(I);
1587 if (!SE->isSCEVable(Phi->getType()))
1590 // Avoid comparing an integer IV against a pointer Limit.
1591 if (BECount->getType()->isPointerTy() && !Phi->getType()->isPointerTy())
1594 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Phi));
1595 if (!AR || AR->getLoop() != L || !AR->isAffine())
1598 // AR may be a pointer type, while BECount is an integer type.
1599 // AR may be wider than BECount. With eq/ne tests overflow is immaterial.
1600 // AR may not be a narrower type, or we may never exit.
1601 uint64_t PhiWidth = SE->getTypeSizeInBits(AR->getType());
1602 if (PhiWidth < BCWidth ||
1603 !L->getHeader()->getModule()->getDataLayout().isLegalInteger(PhiWidth))
1606 const SCEV *Step = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE));
1607 if (!Step || !Step->isOne())
1610 int LatchIdx = Phi->getBasicBlockIndex(LatchBlock);
1611 Value *IncV = Phi->getIncomingValue(LatchIdx);
1612 if (getLoopPhiForCounter(IncV, L, DT) != Phi)
1615 // Avoid reusing a potentially undef value to compute other values that may
1616 // have originally had a concrete definition.
1617 if (!hasConcreteDef(Phi)) {
1618 // We explicitly allow unknown phis as long as they are already used by
1619 // the loop test. In this case we assume that performing LFTR could not
1620 // increase the number of undef users.
1621 if (ICmpInst *Cond = getLoopTest(L)) {
1622 if (Phi != getLoopPhiForCounter(Cond->getOperand(0), L, DT)
1623 && Phi != getLoopPhiForCounter(Cond->getOperand(1), L, DT)) {
1628 const SCEV *Init = AR->getStart();
1630 if (BestPhi && !AlmostDeadIV(BestPhi, LatchBlock, Cond)) {
1631 // Don't force a live loop counter if another IV can be used.
1632 if (AlmostDeadIV(Phi, LatchBlock, Cond))
1635 // Prefer to count-from-zero. This is a more "canonical" counter form. It
1636 // also prefers integer to pointer IVs.
1637 if (BestInit->isZero() != Init->isZero()) {
1638 if (BestInit->isZero())
1641 // If two IVs both count from zero or both count from nonzero then the
1642 // narrower is likely a dead phi that has been widened. Use the wider phi
1643 // to allow the other to be eliminated.
1644 else if (PhiWidth <= SE->getTypeSizeInBits(BestPhi->getType()))
1653 /// genLoopLimit - Help LinearFunctionTestReplace by generating a value that
1654 /// holds the RHS of the new loop test.
1655 static Value *genLoopLimit(PHINode *IndVar, const SCEV *IVCount, Loop *L,
1656 SCEVExpander &Rewriter, ScalarEvolution *SE) {
1657 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(IndVar));
1658 assert(AR && AR->getLoop() == L && AR->isAffine() && "bad loop counter");
1659 const SCEV *IVInit = AR->getStart();
1661 // IVInit may be a pointer while IVCount is an integer when FindLoopCounter
1662 // finds a valid pointer IV. Sign extend BECount in order to materialize a
1663 // GEP. Avoid running SCEVExpander on a new pointer value, instead reusing
1664 // the existing GEPs whenever possible.
1665 if (IndVar->getType()->isPointerTy()
1666 && !IVCount->getType()->isPointerTy()) {
1668 // IVOffset will be the new GEP offset that is interpreted by GEP as a
1669 // signed value. IVCount on the other hand represents the loop trip count,
1670 // which is an unsigned value. FindLoopCounter only allows induction
1671 // variables that have a positive unit stride of one. This means we don't
1672 // have to handle the case of negative offsets (yet) and just need to zero
1674 Type *OfsTy = SE->getEffectiveSCEVType(IVInit->getType());
1675 const SCEV *IVOffset = SE->getTruncateOrZeroExtend(IVCount, OfsTy);
1677 // Expand the code for the iteration count.
1678 assert(SE->isLoopInvariant(IVOffset, L) &&
1679 "Computed iteration count is not loop invariant!");
1680 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1681 Value *GEPOffset = Rewriter.expandCodeFor(IVOffset, OfsTy, BI);
1683 Value *GEPBase = IndVar->getIncomingValueForBlock(L->getLoopPreheader());
1684 assert(AR->getStart() == SE->getSCEV(GEPBase) && "bad loop counter");
1685 // We could handle pointer IVs other than i8*, but we need to compensate for
1686 // gep index scaling. See canExpandBackedgeTakenCount comments.
1687 assert(SE->getSizeOfExpr(IntegerType::getInt64Ty(IndVar->getContext()),
1688 cast<PointerType>(GEPBase->getType())->getElementType())->isOne()
1689 && "unit stride pointer IV must be i8*");
1691 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
1692 return Builder.CreateGEP(nullptr, GEPBase, GEPOffset, "lftr.limit");
1695 // In any other case, convert both IVInit and IVCount to integers before
1696 // comparing. This may result in SCEV expension of pointers, but in practice
1697 // SCEV will fold the pointer arithmetic away as such:
1698 // BECount = (IVEnd - IVInit - 1) => IVLimit = IVInit (postinc).
1700 // Valid Cases: (1) both integers is most common; (2) both may be pointers
1701 // for simple memset-style loops.
1703 // IVInit integer and IVCount pointer would only occur if a canonical IV
1704 // were generated on top of case #2, which is not expected.
1706 const SCEV *IVLimit = nullptr;
1707 // For unit stride, IVCount = Start + BECount with 2's complement overflow.
1708 // For non-zero Start, compute IVCount here.
1709 if (AR->getStart()->isZero())
1712 assert(AR->getStepRecurrence(*SE)->isOne() && "only handles unit stride");
1713 const SCEV *IVInit = AR->getStart();
1715 // For integer IVs, truncate the IV before computing IVInit + BECount.
1716 if (SE->getTypeSizeInBits(IVInit->getType())
1717 > SE->getTypeSizeInBits(IVCount->getType()))
1718 IVInit = SE->getTruncateExpr(IVInit, IVCount->getType());
1720 IVLimit = SE->getAddExpr(IVInit, IVCount);
1722 // Expand the code for the iteration count.
1723 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1724 IRBuilder<> Builder(BI);
1725 assert(SE->isLoopInvariant(IVLimit, L) &&
1726 "Computed iteration count is not loop invariant!");
1727 // Ensure that we generate the same type as IndVar, or a smaller integer
1728 // type. In the presence of null pointer values, we have an integer type
1729 // SCEV expression (IVInit) for a pointer type IV value (IndVar).
1730 Type *LimitTy = IVCount->getType()->isPointerTy() ?
1731 IndVar->getType() : IVCount->getType();
1732 return Rewriter.expandCodeFor(IVLimit, LimitTy, BI);
1736 /// LinearFunctionTestReplace - This method rewrites the exit condition of the
1737 /// loop to be a canonical != comparison against the incremented loop induction
1738 /// variable. This pass is able to rewrite the exit tests of any loop where the
1739 /// SCEV analysis can determine a loop-invariant trip count of the loop, which
1740 /// is actually a much broader range than just linear tests.
1741 Value *IndVarSimplify::
1742 LinearFunctionTestReplace(Loop *L,
1743 const SCEV *BackedgeTakenCount,
1745 SCEVExpander &Rewriter) {
1746 assert(canExpandBackedgeTakenCount(L, SE, Rewriter) && "precondition");
1748 // Initialize CmpIndVar and IVCount to their preincremented values.
1749 Value *CmpIndVar = IndVar;
1750 const SCEV *IVCount = BackedgeTakenCount;
1752 // If the exiting block is the same as the backedge block, we prefer to
1753 // compare against the post-incremented value, otherwise we must compare
1754 // against the preincremented value.
1755 if (L->getExitingBlock() == L->getLoopLatch()) {
1756 // Add one to the "backedge-taken" count to get the trip count.
1757 // This addition may overflow, which is valid as long as the comparison is
1758 // truncated to BackedgeTakenCount->getType().
1759 IVCount = SE->getAddExpr(BackedgeTakenCount,
1760 SE->getConstant(BackedgeTakenCount->getType(), 1));
1761 // The BackedgeTaken expression contains the number of times that the
1762 // backedge branches to the loop header. This is one less than the
1763 // number of times the loop executes, so use the incremented indvar.
1764 CmpIndVar = IndVar->getIncomingValueForBlock(L->getExitingBlock());
1767 Value *ExitCnt = genLoopLimit(IndVar, IVCount, L, Rewriter, SE);
1768 assert(ExitCnt->getType()->isPointerTy() == IndVar->getType()->isPointerTy()
1769 && "genLoopLimit missed a cast");
1771 // Insert a new icmp_ne or icmp_eq instruction before the branch.
1772 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1773 ICmpInst::Predicate P;
1774 if (L->contains(BI->getSuccessor(0)))
1775 P = ICmpInst::ICMP_NE;
1777 P = ICmpInst::ICMP_EQ;
1779 DEBUG(dbgs() << "INDVARS: Rewriting loop exit condition to:\n"
1780 << " LHS:" << *CmpIndVar << '\n'
1782 << (P == ICmpInst::ICMP_NE ? "!=" : "==") << "\n"
1783 << " RHS:\t" << *ExitCnt << "\n"
1784 << " IVCount:\t" << *IVCount << "\n");
1786 IRBuilder<> Builder(BI);
1788 // LFTR can ignore IV overflow and truncate to the width of
1789 // BECount. This avoids materializing the add(zext(add)) expression.
1790 unsigned CmpIndVarSize = SE->getTypeSizeInBits(CmpIndVar->getType());
1791 unsigned ExitCntSize = SE->getTypeSizeInBits(ExitCnt->getType());
1792 if (CmpIndVarSize > ExitCntSize) {
1793 const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(SE->getSCEV(IndVar));
1794 const SCEV *ARStart = AR->getStart();
1795 const SCEV *ARStep = AR->getStepRecurrence(*SE);
1796 // For constant IVCount, avoid truncation.
1797 if (isa<SCEVConstant>(ARStart) && isa<SCEVConstant>(IVCount)) {
1798 const APInt &Start = cast<SCEVConstant>(ARStart)->getValue()->getValue();
1799 APInt Count = cast<SCEVConstant>(IVCount)->getValue()->getValue();
1800 // Note that the post-inc value of BackedgeTakenCount may have overflowed
1801 // above such that IVCount is now zero.
1802 if (IVCount != BackedgeTakenCount && Count == 0) {
1803 Count = APInt::getMaxValue(Count.getBitWidth()).zext(CmpIndVarSize);
1807 Count = Count.zext(CmpIndVarSize);
1809 if (cast<SCEVConstant>(ARStep)->getValue()->isNegative())
1810 NewLimit = Start - Count;
1812 NewLimit = Start + Count;
1813 ExitCnt = ConstantInt::get(CmpIndVar->getType(), NewLimit);
1815 DEBUG(dbgs() << " Widen RHS:\t" << *ExitCnt << "\n");
1817 CmpIndVar = Builder.CreateTrunc(CmpIndVar, ExitCnt->getType(),
1821 Value *Cond = Builder.CreateICmp(P, CmpIndVar, ExitCnt, "exitcond");
1822 Value *OrigCond = BI->getCondition();
1823 // It's tempting to use replaceAllUsesWith here to fully replace the old
1824 // comparison, but that's not immediately safe, since users of the old
1825 // comparison may not be dominated by the new comparison. Instead, just
1826 // update the branch to use the new comparison; in the common case this
1827 // will make old comparison dead.
1828 BI->setCondition(Cond);
1829 DeadInsts.push_back(OrigCond);
1836 //===----------------------------------------------------------------------===//
1837 // SinkUnusedInvariants. A late subpass to cleanup loop preheaders.
1838 //===----------------------------------------------------------------------===//
1840 /// If there's a single exit block, sink any loop-invariant values that
1841 /// were defined in the preheader but not used inside the loop into the
1842 /// exit block to reduce register pressure in the loop.
1843 void IndVarSimplify::SinkUnusedInvariants(Loop *L) {
1844 BasicBlock *ExitBlock = L->getExitBlock();
1845 if (!ExitBlock) return;
1847 BasicBlock *Preheader = L->getLoopPreheader();
1848 if (!Preheader) return;
1850 Instruction *InsertPt = ExitBlock->getFirstInsertionPt();
1851 BasicBlock::iterator I = Preheader->getTerminator();
1852 while (I != Preheader->begin()) {
1854 // New instructions were inserted at the end of the preheader.
1855 if (isa<PHINode>(I))
1858 // Don't move instructions which might have side effects, since the side
1859 // effects need to complete before instructions inside the loop. Also don't
1860 // move instructions which might read memory, since the loop may modify
1861 // memory. Note that it's okay if the instruction might have undefined
1862 // behavior: LoopSimplify guarantees that the preheader dominates the exit
1864 if (I->mayHaveSideEffects() || I->mayReadFromMemory())
1867 // Skip debug info intrinsics.
1868 if (isa<DbgInfoIntrinsic>(I))
1871 // Skip landingpad instructions.
1872 if (isa<LandingPadInst>(I))
1875 // Don't sink alloca: we never want to sink static alloca's out of the
1876 // entry block, and correctly sinking dynamic alloca's requires
1877 // checks for stacksave/stackrestore intrinsics.
1878 // FIXME: Refactor this check somehow?
1879 if (isa<AllocaInst>(I))
1882 // Determine if there is a use in or before the loop (direct or
1884 bool UsedInLoop = false;
1885 for (Use &U : I->uses()) {
1886 Instruction *User = cast<Instruction>(U.getUser());
1887 BasicBlock *UseBB = User->getParent();
1888 if (PHINode *P = dyn_cast<PHINode>(User)) {
1890 PHINode::getIncomingValueNumForOperand(U.getOperandNo());
1891 UseBB = P->getIncomingBlock(i);
1893 if (UseBB == Preheader || L->contains(UseBB)) {
1899 // If there is, the def must remain in the preheader.
1903 // Otherwise, sink it to the exit block.
1904 Instruction *ToMove = I;
1907 if (I != Preheader->begin()) {
1908 // Skip debug info intrinsics.
1911 } while (isa<DbgInfoIntrinsic>(I) && I != Preheader->begin());
1913 if (isa<DbgInfoIntrinsic>(I) && I == Preheader->begin())
1919 ToMove->moveBefore(InsertPt);
1925 //===----------------------------------------------------------------------===//
1926 // IndVarSimplify driver. Manage several subpasses of IV simplification.
1927 //===----------------------------------------------------------------------===//
1929 bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) {
1930 if (skipOptnoneFunction(L))
1933 // If LoopSimplify form is not available, stay out of trouble. Some notes:
1934 // - LSR currently only supports LoopSimplify-form loops. Indvars'
1935 // canonicalization can be a pessimization without LSR to "clean up"
1937 // - We depend on having a preheader; in particular,
1938 // Loop::getCanonicalInductionVariable only supports loops with preheaders,
1939 // and we're in trouble if we can't find the induction variable even when
1940 // we've manually inserted one.
1941 if (!L->isLoopSimplifyForm())
1944 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1945 SE = &getAnalysis<ScalarEvolution>();
1946 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1947 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1948 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1949 auto *TTIP = getAnalysisIfAvailable<TargetTransformInfoWrapperPass>();
1950 TTI = TTIP ? &TTIP->getTTI(*L->getHeader()->getParent()) : nullptr;
1951 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
1956 // If there are any floating-point recurrences, attempt to
1957 // transform them to use integer recurrences.
1958 RewriteNonIntegerIVs(L);
1960 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
1962 // Create a rewriter object which we'll use to transform the code with.
1963 SCEVExpander Rewriter(*SE, DL, "indvars");
1965 Rewriter.setDebugType(DEBUG_TYPE);
1968 // Eliminate redundant IV users.
1970 // Simplification works best when run before other consumers of SCEV. We
1971 // attempt to avoid evaluating SCEVs for sign/zero extend operations until
1972 // other expressions involving loop IVs have been evaluated. This helps SCEV
1973 // set no-wrap flags before normalizing sign/zero extension.
1974 Rewriter.disableCanonicalMode();
1975 SimplifyAndExtend(L, Rewriter, LPM);
1977 // Check to see if this loop has a computable loop-invariant execution count.
1978 // If so, this means that we can compute the final value of any expressions
1979 // that are recurrent in the loop, and substitute the exit values from the
1980 // loop into any instructions outside of the loop that use the final values of
1981 // the current expressions.
1983 if (ReplaceExitValue != NeverRepl &&
1984 !isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1985 RewriteLoopExitValues(L, Rewriter);
1987 // Eliminate redundant IV cycles.
1988 NumElimIV += Rewriter.replaceCongruentIVs(L, DT, DeadInsts);
1990 // If we have a trip count expression, rewrite the loop's exit condition
1991 // using it. We can currently only handle loops with a single exit.
1992 if (canExpandBackedgeTakenCount(L, SE, Rewriter) && needsLFTR(L, DT)) {
1993 PHINode *IndVar = FindLoopCounter(L, BackedgeTakenCount, SE, DT);
1995 // Check preconditions for proper SCEVExpander operation. SCEV does not
1996 // express SCEVExpander's dependencies, such as LoopSimplify. Instead any
1997 // pass that uses the SCEVExpander must do it. This does not work well for
1998 // loop passes because SCEVExpander makes assumptions about all loops,
1999 // while LoopPassManager only forces the current loop to be simplified.
2001 // FIXME: SCEV expansion has no way to bail out, so the caller must
2002 // explicitly check any assumptions made by SCEV. Brittle.
2003 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(BackedgeTakenCount);
2004 if (!AR || AR->getLoop()->getLoopPreheader())
2005 (void)LinearFunctionTestReplace(L, BackedgeTakenCount, IndVar,
2009 // Clear the rewriter cache, because values that are in the rewriter's cache
2010 // can be deleted in the loop below, causing the AssertingVH in the cache to
2014 // Now that we're done iterating through lists, clean up any instructions
2015 // which are now dead.
2016 while (!DeadInsts.empty())
2017 if (Instruction *Inst =
2018 dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val()))
2019 RecursivelyDeleteTriviallyDeadInstructions(Inst, TLI);
2021 // The Rewriter may not be used from this point on.
2023 // Loop-invariant instructions in the preheader that aren't used in the
2024 // loop may be sunk below the loop to reduce register pressure.
2025 SinkUnusedInvariants(L);
2027 // Clean up dead instructions.
2028 Changed |= DeleteDeadPHIs(L->getHeader(), TLI);
2029 // Check a post-condition.
2030 assert(L->isLCSSAForm(*DT) &&
2031 "Indvars did not leave the loop in lcssa form!");
2033 // Verify that LFTR, and any other change have not interfered with SCEV's
2034 // ability to compute trip count.
2036 if (VerifyIndvars && !isa<SCEVCouldNotCompute>(BackedgeTakenCount)) {
2038 const SCEV *NewBECount = SE->getBackedgeTakenCount(L);
2039 if (SE->getTypeSizeInBits(BackedgeTakenCount->getType()) <
2040 SE->getTypeSizeInBits(NewBECount->getType()))
2041 NewBECount = SE->getTruncateOrNoop(NewBECount,
2042 BackedgeTakenCount->getType());
2044 BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount,
2045 NewBECount->getType());
2046 assert(BackedgeTakenCount == NewBECount && "indvars must preserve SCEV");