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/GlobalsModRef.h"
32 #include "llvm/Analysis/LoopInfo.h"
33 #include "llvm/Analysis/LoopPass.h"
34 #include "llvm/Analysis/ScalarEvolutionExpander.h"
35 #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h"
36 #include "llvm/Analysis/TargetLibraryInfo.h"
37 #include "llvm/Analysis/TargetTransformInfo.h"
38 #include "llvm/IR/BasicBlock.h"
39 #include "llvm/IR/CFG.h"
40 #include "llvm/IR/Constants.h"
41 #include "llvm/IR/DataLayout.h"
42 #include "llvm/IR/Dominators.h"
43 #include "llvm/IR/Instructions.h"
44 #include "llvm/IR/IntrinsicInst.h"
45 #include "llvm/IR/LLVMContext.h"
46 #include "llvm/IR/PatternMatch.h"
47 #include "llvm/IR/Type.h"
48 #include "llvm/Support/CommandLine.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/raw_ostream.h"
51 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
52 #include "llvm/Transforms/Utils/Local.h"
53 #include "llvm/Transforms/Utils/SimplifyIndVar.h"
56 #define DEBUG_TYPE "indvars"
58 STATISTIC(NumWidened , "Number of indvars widened");
59 STATISTIC(NumReplaced , "Number of exit values replaced");
60 STATISTIC(NumLFTR , "Number of loop exit tests replaced");
61 STATISTIC(NumElimExt , "Number of IV sign/zero extends eliminated");
62 STATISTIC(NumElimIV , "Number of congruent IVs eliminated");
64 // Trip count verification can be enabled by default under NDEBUG if we
65 // implement a strong expression equivalence checker in SCEV. Until then, we
66 // use the verify-indvars flag, which may assert in some cases.
67 static cl::opt<bool> VerifyIndvars(
68 "verify-indvars", cl::Hidden,
69 cl::desc("Verify the ScalarEvolution result after running indvars"));
71 static cl::opt<bool> ReduceLiveIVs("liv-reduce", cl::Hidden,
72 cl::desc("Reduce live induction variables."));
74 enum ReplaceExitVal { NeverRepl, OnlyCheapRepl, AlwaysRepl };
76 static cl::opt<ReplaceExitVal> ReplaceExitValue(
77 "replexitval", cl::Hidden, cl::init(OnlyCheapRepl),
78 cl::desc("Choose the strategy to replace exit value in IndVarSimplify"),
79 cl::values(clEnumValN(NeverRepl, "never", "never replace exit value"),
80 clEnumValN(OnlyCheapRepl, "cheap",
81 "only replace exit value when the cost is cheap"),
82 clEnumValN(AlwaysRepl, "always",
83 "always replace exit value whenever possible"),
91 class IndVarSimplify : public LoopPass {
95 TargetLibraryInfo *TLI;
96 const TargetTransformInfo *TTI;
98 SmallVector<WeakVH, 16> DeadInsts;
102 static char ID; // Pass identification, replacement for typeid
104 : LoopPass(ID), LI(nullptr), SE(nullptr), DT(nullptr), Changed(false) {
105 initializeIndVarSimplifyPass(*PassRegistry::getPassRegistry());
108 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
110 void getAnalysisUsage(AnalysisUsage &AU) const override {
111 AU.addRequired<DominatorTreeWrapperPass>();
112 AU.addRequired<LoopInfoWrapperPass>();
113 AU.addRequired<ScalarEvolutionWrapperPass>();
114 AU.addRequiredID(LoopSimplifyID);
115 AU.addRequiredID(LCSSAID);
116 AU.addPreserved<GlobalsAAWrapperPass>();
117 AU.addPreserved<ScalarEvolutionWrapperPass>();
118 AU.addPreservedID(LoopSimplifyID);
119 AU.addPreservedID(LCSSAID);
120 AU.setPreservesCFG();
124 void releaseMemory() override {
128 bool isValidRewrite(Value *FromVal, Value *ToVal);
130 void HandleFloatingPointIV(Loop *L, PHINode *PH);
131 void RewriteNonIntegerIVs(Loop *L);
133 void SimplifyAndExtend(Loop *L, SCEVExpander &Rewriter, LPPassManager &LPM);
135 bool CanLoopBeDeleted(Loop *L, SmallVector<RewritePhi, 8> &RewritePhiSet);
136 void RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter);
138 Value *LinearFunctionTestReplace(Loop *L, const SCEV *BackedgeTakenCount,
139 PHINode *IndVar, SCEVExpander &Rewriter);
141 void SinkUnusedInvariants(Loop *L);
143 Value *ExpandSCEVIfNeeded(SCEVExpander &Rewriter, const SCEV *S, Loop *L,
144 Instruction *InsertPt, Type *Ty);
148 char IndVarSimplify::ID = 0;
149 INITIALIZE_PASS_BEGIN(IndVarSimplify, "indvars",
150 "Induction Variable Simplification", false, false)
151 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
152 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
153 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
154 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
155 INITIALIZE_PASS_DEPENDENCY(LCSSA)
156 INITIALIZE_PASS_END(IndVarSimplify, "indvars",
157 "Induction Variable Simplification", false, false)
159 Pass *llvm::createIndVarSimplifyPass() {
160 return new IndVarSimplify();
163 /// isValidRewrite - Return true if the SCEV expansion generated by the
164 /// rewriter can replace the original value. SCEV guarantees that it
165 /// produces the same value, but the way it is produced may be illegal IR.
166 /// Ideally, this function will only be called for verification.
167 bool IndVarSimplify::isValidRewrite(Value *FromVal, Value *ToVal) {
168 // If an SCEV expression subsumed multiple pointers, its expansion could
169 // reassociate the GEP changing the base pointer. This is illegal because the
170 // final address produced by a GEP chain must be inbounds relative to its
171 // underlying object. Otherwise basic alias analysis, among other things,
172 // could fail in a dangerous way. Ultimately, SCEV will be improved to avoid
173 // producing an expression involving multiple pointers. Until then, we must
176 // Retrieve the pointer operand of the GEP. Don't use GetUnderlyingObject
177 // because it understands lcssa phis while SCEV does not.
178 Value *FromPtr = FromVal;
179 Value *ToPtr = ToVal;
180 if (GEPOperator *GEP = dyn_cast<GEPOperator>(FromVal)) {
181 FromPtr = GEP->getPointerOperand();
183 if (GEPOperator *GEP = dyn_cast<GEPOperator>(ToVal)) {
184 ToPtr = GEP->getPointerOperand();
186 if (FromPtr != FromVal || ToPtr != ToVal) {
187 // Quickly check the common case
188 if (FromPtr == ToPtr)
191 // SCEV may have rewritten an expression that produces the GEP's pointer
192 // operand. That's ok as long as the pointer operand has the same base
193 // pointer. Unlike GetUnderlyingObject(), getPointerBase() will find the
194 // base of a recurrence. This handles the case in which SCEV expansion
195 // converts a pointer type recurrence into a nonrecurrent pointer base
196 // indexed by an integer recurrence.
198 // If the GEP base pointer is a vector of pointers, abort.
199 if (!FromPtr->getType()->isPointerTy() || !ToPtr->getType()->isPointerTy())
202 const SCEV *FromBase = SE->getPointerBase(SE->getSCEV(FromPtr));
203 const SCEV *ToBase = SE->getPointerBase(SE->getSCEV(ToPtr));
204 if (FromBase == ToBase)
207 DEBUG(dbgs() << "INDVARS: GEP rewrite bail out "
208 << *FromBase << " != " << *ToBase << "\n");
215 /// Determine the insertion point for this user. By default, insert immediately
216 /// before the user. SCEVExpander or LICM will hoist loop invariants out of the
217 /// loop. For PHI nodes, there may be multiple uses, so compute the nearest
218 /// common dominator for the incoming blocks.
219 static Instruction *getInsertPointForUses(Instruction *User, Value *Def,
221 PHINode *PHI = dyn_cast<PHINode>(User);
225 Instruction *InsertPt = nullptr;
226 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i) {
227 if (PHI->getIncomingValue(i) != Def)
230 BasicBlock *InsertBB = PHI->getIncomingBlock(i);
232 InsertPt = InsertBB->getTerminator();
235 InsertBB = DT->findNearestCommonDominator(InsertPt->getParent(), InsertBB);
236 InsertPt = InsertBB->getTerminator();
238 assert(InsertPt && "Missing phi operand");
239 assert((!isa<Instruction>(Def) ||
240 DT->dominates(cast<Instruction>(Def), InsertPt)) &&
241 "def does not dominate all uses");
245 //===----------------------------------------------------------------------===//
246 // RewriteNonIntegerIVs and helpers. Prefer integer IVs.
247 //===----------------------------------------------------------------------===//
249 /// ConvertToSInt - Convert APF to an integer, if possible.
250 static bool ConvertToSInt(const APFloat &APF, int64_t &IntVal) {
251 bool isExact = false;
252 // See if we can convert this to an int64_t
254 if (APF.convertToInteger(&UIntVal, 64, true, APFloat::rmTowardZero,
255 &isExact) != APFloat::opOK || !isExact)
261 /// HandleFloatingPointIV - If the loop has floating induction variable
262 /// then insert corresponding integer induction variable if possible.
264 /// for(double i = 0; i < 10000; ++i)
266 /// is converted into
267 /// for(int i = 0; i < 10000; ++i)
270 void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PN) {
271 unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
272 unsigned BackEdge = IncomingEdge^1;
274 // Check incoming value.
275 ConstantFP *InitValueVal =
276 dyn_cast<ConstantFP>(PN->getIncomingValue(IncomingEdge));
279 if (!InitValueVal || !ConvertToSInt(InitValueVal->getValueAPF(), InitValue))
282 // Check IV increment. Reject this PN if increment operation is not
283 // an add or increment value can not be represented by an integer.
284 BinaryOperator *Incr =
285 dyn_cast<BinaryOperator>(PN->getIncomingValue(BackEdge));
286 if (Incr == nullptr || Incr->getOpcode() != Instruction::FAdd) return;
288 // If this is not an add of the PHI with a constantfp, or if the constant fp
289 // is not an integer, bail out.
290 ConstantFP *IncValueVal = dyn_cast<ConstantFP>(Incr->getOperand(1));
292 if (IncValueVal == nullptr || Incr->getOperand(0) != PN ||
293 !ConvertToSInt(IncValueVal->getValueAPF(), IncValue))
296 // Check Incr uses. One user is PN and the other user is an exit condition
297 // used by the conditional terminator.
298 Value::user_iterator IncrUse = Incr->user_begin();
299 Instruction *U1 = cast<Instruction>(*IncrUse++);
300 if (IncrUse == Incr->user_end()) return;
301 Instruction *U2 = cast<Instruction>(*IncrUse++);
302 if (IncrUse != Incr->user_end()) return;
304 // Find exit condition, which is an fcmp. If it doesn't exist, or if it isn't
305 // only used by a branch, we can't transform it.
306 FCmpInst *Compare = dyn_cast<FCmpInst>(U1);
308 Compare = dyn_cast<FCmpInst>(U2);
309 if (!Compare || !Compare->hasOneUse() ||
310 !isa<BranchInst>(Compare->user_back()))
313 BranchInst *TheBr = cast<BranchInst>(Compare->user_back());
315 // We need to verify that the branch actually controls the iteration count
316 // of the loop. If not, the new IV can overflow and no one will notice.
317 // The branch block must be in the loop and one of the successors must be out
319 assert(TheBr->isConditional() && "Can't use fcmp if not conditional");
320 if (!L->contains(TheBr->getParent()) ||
321 (L->contains(TheBr->getSuccessor(0)) &&
322 L->contains(TheBr->getSuccessor(1))))
326 // If it isn't a comparison with an integer-as-fp (the exit value), we can't
328 ConstantFP *ExitValueVal = dyn_cast<ConstantFP>(Compare->getOperand(1));
330 if (ExitValueVal == nullptr ||
331 !ConvertToSInt(ExitValueVal->getValueAPF(), ExitValue))
334 // Find new predicate for integer comparison.
335 CmpInst::Predicate NewPred = CmpInst::BAD_ICMP_PREDICATE;
336 switch (Compare->getPredicate()) {
337 default: return; // Unknown comparison.
338 case CmpInst::FCMP_OEQ:
339 case CmpInst::FCMP_UEQ: NewPred = CmpInst::ICMP_EQ; break;
340 case CmpInst::FCMP_ONE:
341 case CmpInst::FCMP_UNE: NewPred = CmpInst::ICMP_NE; break;
342 case CmpInst::FCMP_OGT:
343 case CmpInst::FCMP_UGT: NewPred = CmpInst::ICMP_SGT; break;
344 case CmpInst::FCMP_OGE:
345 case CmpInst::FCMP_UGE: NewPred = CmpInst::ICMP_SGE; break;
346 case CmpInst::FCMP_OLT:
347 case CmpInst::FCMP_ULT: NewPred = CmpInst::ICMP_SLT; break;
348 case CmpInst::FCMP_OLE:
349 case CmpInst::FCMP_ULE: NewPred = CmpInst::ICMP_SLE; break;
352 // We convert the floating point induction variable to a signed i32 value if
353 // we can. This is only safe if the comparison will not overflow in a way
354 // that won't be trapped by the integer equivalent operations. Check for this
356 // TODO: We could use i64 if it is native and the range requires it.
358 // The start/stride/exit values must all fit in signed i32.
359 if (!isInt<32>(InitValue) || !isInt<32>(IncValue) || !isInt<32>(ExitValue))
362 // If not actually striding (add x, 0.0), avoid touching the code.
366 // Positive and negative strides have different safety conditions.
368 // If we have a positive stride, we require the init to be less than the
370 if (InitValue >= ExitValue)
373 uint32_t Range = uint32_t(ExitValue-InitValue);
374 // Check for infinite loop, either:
375 // while (i <= Exit) or until (i > Exit)
376 if (NewPred == CmpInst::ICMP_SLE || NewPred == CmpInst::ICMP_SGT) {
377 if (++Range == 0) return; // Range overflows.
380 unsigned Leftover = Range % uint32_t(IncValue);
382 // If this is an equality comparison, we require that the strided value
383 // exactly land on the exit value, otherwise the IV condition will wrap
384 // around and do things the fp IV wouldn't.
385 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
389 // If the stride would wrap around the i32 before exiting, we can't
391 if (Leftover != 0 && int32_t(ExitValue+IncValue) < ExitValue)
395 // If we have a negative stride, we require the init to be greater than the
397 if (InitValue <= ExitValue)
400 uint32_t Range = uint32_t(InitValue-ExitValue);
401 // Check for infinite loop, either:
402 // while (i >= Exit) or until (i < Exit)
403 if (NewPred == CmpInst::ICMP_SGE || NewPred == CmpInst::ICMP_SLT) {
404 if (++Range == 0) return; // Range overflows.
407 unsigned Leftover = Range % uint32_t(-IncValue);
409 // If this is an equality comparison, we require that the strided value
410 // exactly land on the exit value, otherwise the IV condition will wrap
411 // around and do things the fp IV wouldn't.
412 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
416 // If the stride would wrap around the i32 before exiting, we can't
418 if (Leftover != 0 && int32_t(ExitValue+IncValue) > ExitValue)
422 IntegerType *Int32Ty = Type::getInt32Ty(PN->getContext());
424 // Insert new integer induction variable.
425 PHINode *NewPHI = PHINode::Create(Int32Ty, 2, PN->getName()+".int", PN);
426 NewPHI->addIncoming(ConstantInt::get(Int32Ty, InitValue),
427 PN->getIncomingBlock(IncomingEdge));
430 BinaryOperator::CreateAdd(NewPHI, ConstantInt::get(Int32Ty, IncValue),
431 Incr->getName()+".int", Incr);
432 NewPHI->addIncoming(NewAdd, PN->getIncomingBlock(BackEdge));
434 ICmpInst *NewCompare = new ICmpInst(TheBr, NewPred, NewAdd,
435 ConstantInt::get(Int32Ty, ExitValue),
438 // In the following deletions, PN may become dead and may be deleted.
439 // Use a WeakVH to observe whether this happens.
442 // Delete the old floating point exit comparison. The branch starts using the
444 NewCompare->takeName(Compare);
445 Compare->replaceAllUsesWith(NewCompare);
446 RecursivelyDeleteTriviallyDeadInstructions(Compare, TLI);
448 // Delete the old floating point increment.
449 Incr->replaceAllUsesWith(UndefValue::get(Incr->getType()));
450 RecursivelyDeleteTriviallyDeadInstructions(Incr, TLI);
452 // If the FP induction variable still has uses, this is because something else
453 // in the loop uses its value. In order to canonicalize the induction
454 // variable, we chose to eliminate the IV and rewrite it in terms of an
457 // We give preference to sitofp over uitofp because it is faster on most
460 Value *Conv = new SIToFPInst(NewPHI, PN->getType(), "indvar.conv",
461 PN->getParent()->getFirstInsertionPt());
462 PN->replaceAllUsesWith(Conv);
463 RecursivelyDeleteTriviallyDeadInstructions(PN, TLI);
468 void IndVarSimplify::RewriteNonIntegerIVs(Loop *L) {
469 // First step. Check to see if there are any floating-point recurrences.
470 // If there are, change them into integer recurrences, permitting analysis by
471 // the SCEV routines.
473 BasicBlock *Header = L->getHeader();
475 SmallVector<WeakVH, 8> PHIs;
476 for (BasicBlock::iterator I = Header->begin();
477 PHINode *PN = dyn_cast<PHINode>(I); ++I)
480 for (unsigned i = 0, e = PHIs.size(); i != e; ++i)
481 if (PHINode *PN = dyn_cast_or_null<PHINode>(&*PHIs[i]))
482 HandleFloatingPointIV(L, PN);
484 // If the loop previously had floating-point IV, ScalarEvolution
485 // may not have been able to compute a trip count. Now that we've done some
486 // re-writing, the trip count may be computable.
492 // Collect information about PHI nodes which can be transformed in
493 // RewriteLoopExitValues.
496 unsigned Ith; // Ith incoming value.
497 Value *Val; // Exit value after expansion.
498 bool HighCost; // High Cost when expansion.
499 bool SafePhi; // LCSSASafePhiForRAUW.
501 RewritePhi(PHINode *P, unsigned I, Value *V, bool H, bool S)
502 : PN(P), Ith(I), Val(V), HighCost(H), SafePhi(S) {}
506 Value *IndVarSimplify::ExpandSCEVIfNeeded(SCEVExpander &Rewriter, const SCEV *S,
507 Loop *L, Instruction *InsertPt,
509 // Before expanding S into an expensive LLVM expression, see if we can use an
510 // already existing value as the expansion for S.
511 if (Value *ExistingValue = Rewriter.findExistingExpansion(S, InsertPt, L))
512 return ExistingValue;
514 // We didn't find anything, fall back to using SCEVExpander.
515 return Rewriter.expandCodeFor(S, ResultTy, InsertPt);
518 //===----------------------------------------------------------------------===//
519 // RewriteLoopExitValues - Optimize IV users outside the loop.
520 // As a side effect, reduces the amount of IV processing within the loop.
521 //===----------------------------------------------------------------------===//
523 /// RewriteLoopExitValues - Check to see if this loop has a computable
524 /// loop-invariant execution count. If so, this means that we can compute the
525 /// final value of any expressions that are recurrent in the loop, and
526 /// substitute the exit values from the loop into any instructions outside of
527 /// the loop that use the final values of the current expressions.
529 /// This is mostly redundant with the regular IndVarSimplify activities that
530 /// happen later, except that it's more powerful in some cases, because it's
531 /// able to brute-force evaluate arbitrary instructions as long as they have
532 /// constant operands at the beginning of the loop.
533 void IndVarSimplify::RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter) {
534 // Verify the input to the pass in already in LCSSA form.
535 assert(L->isLCSSAForm(*DT));
537 SmallVector<BasicBlock*, 8> ExitBlocks;
538 L->getUniqueExitBlocks(ExitBlocks);
540 SmallVector<RewritePhi, 8> RewritePhiSet;
541 // Find all values that are computed inside the loop, but used outside of it.
542 // Because of LCSSA, these values will only occur in LCSSA PHI Nodes. Scan
543 // the exit blocks of the loop to find them.
544 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
545 BasicBlock *ExitBB = ExitBlocks[i];
547 // If there are no PHI nodes in this exit block, then no values defined
548 // inside the loop are used on this path, skip it.
549 PHINode *PN = dyn_cast<PHINode>(ExitBB->begin());
552 unsigned NumPreds = PN->getNumIncomingValues();
554 // We would like to be able to RAUW single-incoming value PHI nodes. We
555 // have to be certain this is safe even when this is an LCSSA PHI node.
556 // While the computed exit value is no longer varying in *this* loop, the
557 // exit block may be an exit block for an outer containing loop as well,
558 // the exit value may be varying in the outer loop, and thus it may still
559 // require an LCSSA PHI node. The safe case is when this is
560 // single-predecessor PHI node (LCSSA) and the exit block containing it is
561 // part of the enclosing loop, or this is the outer most loop of the nest.
562 // In either case the exit value could (at most) be varying in the same
563 // loop body as the phi node itself. Thus if it is in turn used outside of
564 // an enclosing loop it will only be via a separate LCSSA node.
565 bool LCSSASafePhiForRAUW =
567 (!L->getParentLoop() || L->getParentLoop() == LI->getLoopFor(ExitBB));
569 // Iterate over all of the PHI nodes.
570 BasicBlock::iterator BBI = ExitBB->begin();
571 while ((PN = dyn_cast<PHINode>(BBI++))) {
573 continue; // dead use, don't replace it
575 // SCEV only supports integer expressions for now.
576 if (!PN->getType()->isIntegerTy() && !PN->getType()->isPointerTy())
579 // It's necessary to tell ScalarEvolution about this explicitly so that
580 // it can walk the def-use list and forget all SCEVs, as it may not be
581 // watching the PHI itself. Once the new exit value is in place, there
582 // may not be a def-use connection between the loop and every instruction
583 // which got a SCEVAddRecExpr for that loop.
586 // Iterate over all of the values in all the PHI nodes.
587 for (unsigned i = 0; i != NumPreds; ++i) {
588 // If the value being merged in is not integer or is not defined
589 // in the loop, skip it.
590 Value *InVal = PN->getIncomingValue(i);
591 if (!isa<Instruction>(InVal))
594 // If this pred is for a subloop, not L itself, skip it.
595 if (LI->getLoopFor(PN->getIncomingBlock(i)) != L)
596 continue; // The Block is in a subloop, skip it.
598 // Check that InVal is defined in the loop.
599 Instruction *Inst = cast<Instruction>(InVal);
600 if (!L->contains(Inst))
603 // Okay, this instruction has a user outside of the current loop
604 // and varies predictably *inside* the loop. Evaluate the value it
605 // contains when the loop exits, if possible.
606 const SCEV *ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop());
607 if (!SE->isLoopInvariant(ExitValue, L) ||
608 !isSafeToExpand(ExitValue, *SE))
611 // Computing the value outside of the loop brings no benefit if :
612 // - it is definitely used inside the loop in a way which can not be
614 // - no use outside of the loop can take advantage of hoisting the
615 // computation out of the loop
616 if (ExitValue->getSCEVType()>=scMulExpr) {
617 unsigned NumHardInternalUses = 0;
618 unsigned NumSoftExternalUses = 0;
619 unsigned NumUses = 0;
620 for (auto IB = Inst->user_begin(), IE = Inst->user_end();
621 IB != IE && NumUses <= 6; ++IB) {
622 Instruction *UseInstr = cast<Instruction>(*IB);
623 unsigned Opc = UseInstr->getOpcode();
625 if (L->contains(UseInstr)) {
626 if (Opc == Instruction::Call || Opc == Instruction::Ret)
627 NumHardInternalUses++;
629 if (Opc == Instruction::PHI) {
630 // Do not count the Phi as a use. LCSSA may have inserted
631 // plenty of trivial ones.
633 for (auto PB = UseInstr->user_begin(),
634 PE = UseInstr->user_end();
635 PB != PE && NumUses <= 6; ++PB, ++NumUses) {
636 unsigned PhiOpc = cast<Instruction>(*PB)->getOpcode();
637 if (PhiOpc != Instruction::Call && PhiOpc != Instruction::Ret)
638 NumSoftExternalUses++;
642 if (Opc != Instruction::Call && Opc != Instruction::Ret)
643 NumSoftExternalUses++;
646 if (NumUses <= 6 && NumHardInternalUses && !NumSoftExternalUses)
650 bool HighCost = Rewriter.isHighCostExpansion(ExitValue, L, Inst);
652 ExpandSCEVIfNeeded(Rewriter, ExitValue, L, Inst, PN->getType());
654 DEBUG(dbgs() << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal << '\n'
655 << " LoopVal = " << *Inst << "\n");
657 if (!isValidRewrite(Inst, ExitVal)) {
658 DeadInsts.push_back(ExitVal);
662 // Collect all the candidate PHINodes to be rewritten.
663 RewritePhiSet.push_back(
664 RewritePhi(PN, i, ExitVal, HighCost, LCSSASafePhiForRAUW));
669 bool LoopCanBeDel = CanLoopBeDeleted(L, RewritePhiSet);
672 for (const RewritePhi &Phi : RewritePhiSet) {
673 PHINode *PN = Phi.PN;
674 Value *ExitVal = Phi.Val;
676 // Only do the rewrite when the ExitValue can be expanded cheaply.
677 // If LoopCanBeDel is true, rewrite exit value aggressively.
678 if (ReplaceExitValue == OnlyCheapRepl && !LoopCanBeDel && Phi.HighCost) {
679 DeadInsts.push_back(ExitVal);
685 Instruction *Inst = cast<Instruction>(PN->getIncomingValue(Phi.Ith));
686 PN->setIncomingValue(Phi.Ith, ExitVal);
688 // If this instruction is dead now, delete it. Don't do it now to avoid
689 // invalidating iterators.
690 if (isInstructionTriviallyDead(Inst, TLI))
691 DeadInsts.push_back(Inst);
693 // If we determined that this PHI is safe to replace even if an LCSSA
696 PN->replaceAllUsesWith(ExitVal);
697 PN->eraseFromParent();
701 // The insertion point instruction may have been deleted; clear it out
702 // so that the rewriter doesn't trip over it later.
703 Rewriter.clearInsertPoint();
706 /// CanLoopBeDeleted - Check whether it is possible to delete the loop after
707 /// rewriting exit value. If it is possible, ignore ReplaceExitValue and
708 /// do rewriting aggressively.
709 bool IndVarSimplify::CanLoopBeDeleted(
710 Loop *L, SmallVector<RewritePhi, 8> &RewritePhiSet) {
712 BasicBlock *Preheader = L->getLoopPreheader();
713 // If there is no preheader, the loop will not be deleted.
717 // In LoopDeletion pass Loop can be deleted when ExitingBlocks.size() > 1.
718 // We obviate multiple ExitingBlocks case for simplicity.
719 // TODO: If we see testcase with multiple ExitingBlocks can be deleted
720 // after exit value rewriting, we can enhance the logic here.
721 SmallVector<BasicBlock *, 4> ExitingBlocks;
722 L->getExitingBlocks(ExitingBlocks);
723 SmallVector<BasicBlock *, 8> ExitBlocks;
724 L->getUniqueExitBlocks(ExitBlocks);
725 if (ExitBlocks.size() > 1 || ExitingBlocks.size() > 1)
728 BasicBlock *ExitBlock = ExitBlocks[0];
729 BasicBlock::iterator BI = ExitBlock->begin();
730 while (PHINode *P = dyn_cast<PHINode>(BI)) {
731 Value *Incoming = P->getIncomingValueForBlock(ExitingBlocks[0]);
733 // If the Incoming value of P is found in RewritePhiSet, we know it
734 // could be rewritten to use a loop invariant value in transformation
735 // phase later. Skip it in the loop invariant check below.
737 for (const RewritePhi &Phi : RewritePhiSet) {
738 unsigned i = Phi.Ith;
739 if (Phi.PN == P && (Phi.PN)->getIncomingValue(i) == Incoming) {
746 if (!found && (I = dyn_cast<Instruction>(Incoming)))
747 if (!L->hasLoopInvariantOperands(I))
753 for (Loop::block_iterator LI = L->block_begin(), LE = L->block_end();
755 for (BasicBlock::iterator BI = (*LI)->begin(), BE = (*LI)->end(); BI != BE;
757 if (BI->mayHaveSideEffects())
765 //===----------------------------------------------------------------------===//
766 // IV Widening - Extend the width of an IV to cover its widest uses.
767 //===----------------------------------------------------------------------===//
770 // Collect information about induction variables that are used by sign/zero
771 // extend operations. This information is recorded by CollectExtend and
772 // provides the input to WidenIV.
775 Type *WidestNativeType; // Widest integer type created [sz]ext
776 bool IsSigned; // Was a sext user seen before a zext?
778 WideIVInfo() : NarrowIV(nullptr), WidestNativeType(nullptr),
783 /// visitCast - Update information about the induction variable that is
784 /// extended by this sign or zero extend operation. This is used to determine
785 /// the final width of the IV before actually widening it.
786 static void visitIVCast(CastInst *Cast, WideIVInfo &WI, ScalarEvolution *SE,
787 const TargetTransformInfo *TTI) {
788 bool IsSigned = Cast->getOpcode() == Instruction::SExt;
789 if (!IsSigned && Cast->getOpcode() != Instruction::ZExt)
792 Type *Ty = Cast->getType();
793 uint64_t Width = SE->getTypeSizeInBits(Ty);
794 if (!Cast->getModule()->getDataLayout().isLegalInteger(Width))
797 // Cast is either an sext or zext up to this point.
798 // We should not widen an indvar if arithmetics on the wider indvar are more
799 // expensive than those on the narrower indvar. We check only the cost of ADD
800 // because at least an ADD is required to increment the induction variable. We
801 // could compute more comprehensively the cost of all instructions on the
802 // induction variable when necessary.
804 TTI->getArithmeticInstrCost(Instruction::Add, Ty) >
805 TTI->getArithmeticInstrCost(Instruction::Add,
806 Cast->getOperand(0)->getType())) {
810 if (!WI.WidestNativeType) {
811 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
812 WI.IsSigned = IsSigned;
816 // We extend the IV to satisfy the sign of its first user, arbitrarily.
817 if (WI.IsSigned != IsSigned)
820 if (Width > SE->getTypeSizeInBits(WI.WidestNativeType))
821 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
826 /// NarrowIVDefUse - Record a link in the Narrow IV def-use chain along with the
827 /// WideIV that computes the same value as the Narrow IV def. This avoids
828 /// caching Use* pointers.
829 struct NarrowIVDefUse {
830 Instruction *NarrowDef;
831 Instruction *NarrowUse;
832 Instruction *WideDef;
834 NarrowIVDefUse(): NarrowDef(nullptr), NarrowUse(nullptr), WideDef(nullptr) {}
836 NarrowIVDefUse(Instruction *ND, Instruction *NU, Instruction *WD):
837 NarrowDef(ND), NarrowUse(NU), WideDef(WD) {}
840 /// WidenIV - The goal of this transform is to remove sign and zero extends
841 /// without creating any new induction variables. To do this, it creates a new
842 /// phi of the wider type and redirects all users, either removing extends or
843 /// inserting truncs whenever we stop propagating the type.
859 Instruction *WideInc;
860 const SCEV *WideIncExpr;
861 SmallVectorImpl<WeakVH> &DeadInsts;
863 SmallPtrSet<Instruction*,16> Widened;
864 SmallVector<NarrowIVDefUse, 8> NarrowIVUsers;
867 WidenIV(const WideIVInfo &WI, LoopInfo *LInfo,
868 ScalarEvolution *SEv, DominatorTree *DTree,
869 SmallVectorImpl<WeakVH> &DI) :
870 OrigPhi(WI.NarrowIV),
871 WideType(WI.WidestNativeType),
872 IsSigned(WI.IsSigned),
874 L(LI->getLoopFor(OrigPhi->getParent())),
879 WideIncExpr(nullptr),
881 assert(L->getHeader() == OrigPhi->getParent() && "Phi must be an IV");
884 PHINode *CreateWideIV(SCEVExpander &Rewriter);
887 Value *getExtend(Value *NarrowOper, Type *WideType, bool IsSigned,
890 Instruction *CloneIVUser(NarrowIVDefUse DU);
892 const SCEVAddRecExpr *GetWideRecurrence(Instruction *NarrowUse);
894 const SCEVAddRecExpr* GetExtendedOperandRecurrence(NarrowIVDefUse DU);
896 const SCEV *GetSCEVByOpCode(const SCEV *LHS, const SCEV *RHS,
897 unsigned OpCode) const;
899 Instruction *WidenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter);
901 bool WidenLoopCompare(NarrowIVDefUse DU);
903 void pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef);
905 } // anonymous namespace
907 /// isLoopInvariant - Perform a quick domtree based check for loop invariance
908 /// assuming that V is used within the loop. LoopInfo::isLoopInvariant() seems
909 /// gratuitous for this purpose.
910 static bool isLoopInvariant(Value *V, const Loop *L, const DominatorTree *DT) {
911 Instruction *Inst = dyn_cast<Instruction>(V);
915 return DT->properlyDominates(Inst->getParent(), L->getHeader());
918 Value *WidenIV::getExtend(Value *NarrowOper, Type *WideType, bool IsSigned,
920 // Set the debug location and conservative insertion point.
921 IRBuilder<> Builder(Use);
922 // Hoist the insertion point into loop preheaders as far as possible.
923 for (const Loop *L = LI->getLoopFor(Use->getParent());
924 L && L->getLoopPreheader() && isLoopInvariant(NarrowOper, L, DT);
925 L = L->getParentLoop())
926 Builder.SetInsertPoint(L->getLoopPreheader()->getTerminator());
928 return IsSigned ? Builder.CreateSExt(NarrowOper, WideType) :
929 Builder.CreateZExt(NarrowOper, WideType);
932 /// CloneIVUser - Instantiate a wide operation to replace a narrow
933 /// operation. This only needs to handle operations that can evaluation to
934 /// SCEVAddRec. It can safely return 0 for any operation we decide not to clone.
935 Instruction *WidenIV::CloneIVUser(NarrowIVDefUse DU) {
936 unsigned Opcode = DU.NarrowUse->getOpcode();
940 case Instruction::Add:
941 case Instruction::Mul:
942 case Instruction::UDiv:
943 case Instruction::Sub:
944 case Instruction::And:
945 case Instruction::Or:
946 case Instruction::Xor:
947 case Instruction::Shl:
948 case Instruction::LShr:
949 case Instruction::AShr:
950 DEBUG(dbgs() << "Cloning IVUser: " << *DU.NarrowUse << "\n");
952 // Replace NarrowDef operands with WideDef. Otherwise, we don't know
953 // anything about the narrow operand yet so must insert a [sz]ext. It is
954 // probably loop invariant and will be folded or hoisted. If it actually
955 // comes from a widened IV, it should be removed during a future call to
957 Value *LHS = (DU.NarrowUse->getOperand(0) == DU.NarrowDef) ? DU.WideDef :
958 getExtend(DU.NarrowUse->getOperand(0), WideType, IsSigned, DU.NarrowUse);
959 Value *RHS = (DU.NarrowUse->getOperand(1) == DU.NarrowDef) ? DU.WideDef :
960 getExtend(DU.NarrowUse->getOperand(1), WideType, IsSigned, DU.NarrowUse);
962 BinaryOperator *NarrowBO = cast<BinaryOperator>(DU.NarrowUse);
963 BinaryOperator *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(),
965 NarrowBO->getName());
966 IRBuilder<> Builder(DU.NarrowUse);
967 Builder.Insert(WideBO);
968 if (const OverflowingBinaryOperator *OBO =
969 dyn_cast<OverflowingBinaryOperator>(NarrowBO)) {
970 if (OBO->hasNoUnsignedWrap()) WideBO->setHasNoUnsignedWrap();
971 if (OBO->hasNoSignedWrap()) WideBO->setHasNoSignedWrap();
977 const SCEV *WidenIV::GetSCEVByOpCode(const SCEV *LHS, const SCEV *RHS,
978 unsigned OpCode) const {
979 if (OpCode == Instruction::Add)
980 return SE->getAddExpr(LHS, RHS);
981 if (OpCode == Instruction::Sub)
982 return SE->getMinusSCEV(LHS, RHS);
983 if (OpCode == Instruction::Mul)
984 return SE->getMulExpr(LHS, RHS);
986 llvm_unreachable("Unsupported opcode.");
989 /// No-wrap operations can transfer sign extension of their result to their
990 /// operands. Generate the SCEV value for the widened operation without
991 /// actually modifying the IR yet. If the expression after extending the
992 /// operands is an AddRec for this loop, return it.
993 const SCEVAddRecExpr* WidenIV::GetExtendedOperandRecurrence(NarrowIVDefUse DU) {
995 // Handle the common case of add<nsw/nuw>
996 const unsigned OpCode = DU.NarrowUse->getOpcode();
997 // Only Add/Sub/Mul instructions supported yet.
998 if (OpCode != Instruction::Add && OpCode != Instruction::Sub &&
999 OpCode != Instruction::Mul)
1002 // One operand (NarrowDef) has already been extended to WideDef. Now determine
1003 // if extending the other will lead to a recurrence.
1004 const unsigned ExtendOperIdx =
1005 DU.NarrowUse->getOperand(0) == DU.NarrowDef ? 1 : 0;
1006 assert(DU.NarrowUse->getOperand(1-ExtendOperIdx) == DU.NarrowDef && "bad DU");
1008 const SCEV *ExtendOperExpr = nullptr;
1009 const OverflowingBinaryOperator *OBO =
1010 cast<OverflowingBinaryOperator>(DU.NarrowUse);
1011 if (IsSigned && OBO->hasNoSignedWrap())
1012 ExtendOperExpr = SE->getSignExtendExpr(
1013 SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType);
1014 else if(!IsSigned && OBO->hasNoUnsignedWrap())
1015 ExtendOperExpr = SE->getZeroExtendExpr(
1016 SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType);
1020 // When creating this SCEV expr, don't apply the current operations NSW or NUW
1021 // flags. This instruction may be guarded by control flow that the no-wrap
1022 // behavior depends on. Non-control-equivalent instructions can be mapped to
1023 // the same SCEV expression, and it would be incorrect to transfer NSW/NUW
1024 // semantics to those operations.
1025 const SCEV *lhs = SE->getSCEV(DU.WideDef);
1026 const SCEV *rhs = ExtendOperExpr;
1028 // Let's swap operands to the initial order for the case of non-commutative
1029 // operations, like SUB. See PR21014.
1030 if (ExtendOperIdx == 0)
1031 std::swap(lhs, rhs);
1032 const SCEVAddRecExpr *AddRec =
1033 dyn_cast<SCEVAddRecExpr>(GetSCEVByOpCode(lhs, rhs, OpCode));
1035 if (!AddRec || AddRec->getLoop() != L)
1040 /// GetWideRecurrence - Is this instruction potentially interesting for further
1041 /// simplification after widening it's type? In other words, can the
1042 /// extend be safely hoisted out of the loop with SCEV reducing the value to a
1043 /// recurrence on the same loop. If so, return the sign or zero extended
1044 /// recurrence. Otherwise return NULL.
1045 const SCEVAddRecExpr *WidenIV::GetWideRecurrence(Instruction *NarrowUse) {
1046 if (!SE->isSCEVable(NarrowUse->getType()))
1049 const SCEV *NarrowExpr = SE->getSCEV(NarrowUse);
1050 if (SE->getTypeSizeInBits(NarrowExpr->getType())
1051 >= SE->getTypeSizeInBits(WideType)) {
1052 // NarrowUse implicitly widens its operand. e.g. a gep with a narrow
1053 // index. So don't follow this use.
1057 const SCEV *WideExpr = IsSigned ?
1058 SE->getSignExtendExpr(NarrowExpr, WideType) :
1059 SE->getZeroExtendExpr(NarrowExpr, WideType);
1060 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(WideExpr);
1061 if (!AddRec || AddRec->getLoop() != L)
1066 /// This IV user cannot be widen. Replace this use of the original narrow IV
1067 /// with a truncation of the new wide IV to isolate and eliminate the narrow IV.
1068 static void truncateIVUse(NarrowIVDefUse DU, DominatorTree *DT) {
1069 DEBUG(dbgs() << "INDVARS: Truncate IV " << *DU.WideDef
1070 << " for user " << *DU.NarrowUse << "\n");
1071 IRBuilder<> Builder(getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT));
1072 Value *Trunc = Builder.CreateTrunc(DU.WideDef, DU.NarrowDef->getType());
1073 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, Trunc);
1076 /// If the narrow use is a compare instruction, then widen the compare
1077 // (and possibly the other operand). The extend operation is hoisted into the
1078 // loop preheader as far as possible.
1079 bool WidenIV::WidenLoopCompare(NarrowIVDefUse DU) {
1080 ICmpInst *Cmp = dyn_cast<ICmpInst>(DU.NarrowUse);
1084 // Sign of IV user and compare must match.
1085 if (IsSigned != CmpInst::isSigned(Cmp->getPredicate()))
1088 Value *Op = Cmp->getOperand(Cmp->getOperand(0) == DU.NarrowDef ? 1 : 0);
1089 unsigned CastWidth = SE->getTypeSizeInBits(Op->getType());
1090 unsigned IVWidth = SE->getTypeSizeInBits(WideType);
1091 assert (CastWidth <= IVWidth && "Unexpected width while widening compare.");
1093 // Widen the compare instruction.
1094 IRBuilder<> Builder(getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT));
1095 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef);
1097 // Widen the other operand of the compare, if necessary.
1098 if (CastWidth < IVWidth) {
1099 Value *ExtOp = getExtend(Op, WideType, IsSigned, Cmp);
1100 DU.NarrowUse->replaceUsesOfWith(Op, ExtOp);
1105 /// WidenIVUse - Determine whether an individual user of the narrow IV can be
1106 /// widened. If so, return the wide clone of the user.
1107 Instruction *WidenIV::WidenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter) {
1109 // Stop traversing the def-use chain at inner-loop phis or post-loop phis.
1110 if (PHINode *UsePhi = dyn_cast<PHINode>(DU.NarrowUse)) {
1111 if (LI->getLoopFor(UsePhi->getParent()) != L) {
1112 // For LCSSA phis, sink the truncate outside the loop.
1113 // After SimplifyCFG most loop exit targets have a single predecessor.
1114 // Otherwise fall back to a truncate within the loop.
1115 if (UsePhi->getNumOperands() != 1)
1116 truncateIVUse(DU, DT);
1119 PHINode::Create(DU.WideDef->getType(), 1, UsePhi->getName() + ".wide",
1121 WidePhi->addIncoming(DU.WideDef, UsePhi->getIncomingBlock(0));
1122 IRBuilder<> Builder(WidePhi->getParent()->getFirstInsertionPt());
1123 Value *Trunc = Builder.CreateTrunc(WidePhi, DU.NarrowDef->getType());
1124 UsePhi->replaceAllUsesWith(Trunc);
1125 DeadInsts.emplace_back(UsePhi);
1126 DEBUG(dbgs() << "INDVARS: Widen lcssa phi " << *UsePhi
1127 << " to " << *WidePhi << "\n");
1132 // Our raison d'etre! Eliminate sign and zero extension.
1133 if (IsSigned ? isa<SExtInst>(DU.NarrowUse) : isa<ZExtInst>(DU.NarrowUse)) {
1134 Value *NewDef = DU.WideDef;
1135 if (DU.NarrowUse->getType() != WideType) {
1136 unsigned CastWidth = SE->getTypeSizeInBits(DU.NarrowUse->getType());
1137 unsigned IVWidth = SE->getTypeSizeInBits(WideType);
1138 if (CastWidth < IVWidth) {
1139 // The cast isn't as wide as the IV, so insert a Trunc.
1140 IRBuilder<> Builder(DU.NarrowUse);
1141 NewDef = Builder.CreateTrunc(DU.WideDef, DU.NarrowUse->getType());
1144 // A wider extend was hidden behind a narrower one. This may induce
1145 // another round of IV widening in which the intermediate IV becomes
1146 // dead. It should be very rare.
1147 DEBUG(dbgs() << "INDVARS: New IV " << *WidePhi
1148 << " not wide enough to subsume " << *DU.NarrowUse << "\n");
1149 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef);
1150 NewDef = DU.NarrowUse;
1153 if (NewDef != DU.NarrowUse) {
1154 DEBUG(dbgs() << "INDVARS: eliminating " << *DU.NarrowUse
1155 << " replaced by " << *DU.WideDef << "\n");
1157 DU.NarrowUse->replaceAllUsesWith(NewDef);
1158 DeadInsts.emplace_back(DU.NarrowUse);
1160 // Now that the extend is gone, we want to expose it's uses for potential
1161 // further simplification. We don't need to directly inform SimplifyIVUsers
1162 // of the new users, because their parent IV will be processed later as a
1163 // new loop phi. If we preserved IVUsers analysis, we would also want to
1164 // push the uses of WideDef here.
1166 // No further widening is needed. The deceased [sz]ext had done it for us.
1170 // Does this user itself evaluate to a recurrence after widening?
1171 const SCEVAddRecExpr *WideAddRec = GetWideRecurrence(DU.NarrowUse);
1173 WideAddRec = GetExtendedOperandRecurrence(DU);
1176 // If use is a loop condition, try to promote the condition instead of
1177 // truncating the IV first.
1178 if (WidenLoopCompare(DU))
1181 // This user does not evaluate to a recurence after widening, so don't
1182 // follow it. Instead insert a Trunc to kill off the original use,
1183 // eventually isolating the original narrow IV so it can be removed.
1184 truncateIVUse(DU, DT);
1187 // Assume block terminators cannot evaluate to a recurrence. We can't to
1188 // insert a Trunc after a terminator if there happens to be a critical edge.
1189 assert(DU.NarrowUse != DU.NarrowUse->getParent()->getTerminator() &&
1190 "SCEV is not expected to evaluate a block terminator");
1192 // Reuse the IV increment that SCEVExpander created as long as it dominates
1194 Instruction *WideUse = nullptr;
1195 if (WideAddRec == WideIncExpr
1196 && Rewriter.hoistIVInc(WideInc, DU.NarrowUse))
1199 WideUse = CloneIVUser(DU);
1203 // Evaluation of WideAddRec ensured that the narrow expression could be
1204 // extended outside the loop without overflow. This suggests that the wide use
1205 // evaluates to the same expression as the extended narrow use, but doesn't
1206 // absolutely guarantee it. Hence the following failsafe check. In rare cases
1207 // where it fails, we simply throw away the newly created wide use.
1208 if (WideAddRec != SE->getSCEV(WideUse)) {
1209 DEBUG(dbgs() << "Wide use expression mismatch: " << *WideUse
1210 << ": " << *SE->getSCEV(WideUse) << " != " << *WideAddRec << "\n");
1211 DeadInsts.emplace_back(WideUse);
1215 // Returning WideUse pushes it on the worklist.
1219 /// pushNarrowIVUsers - Add eligible users of NarrowDef to NarrowIVUsers.
1221 void WidenIV::pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef) {
1222 for (User *U : NarrowDef->users()) {
1223 Instruction *NarrowUser = cast<Instruction>(U);
1225 // Handle data flow merges and bizarre phi cycles.
1226 if (!Widened.insert(NarrowUser).second)
1229 NarrowIVUsers.push_back(NarrowIVDefUse(NarrowDef, NarrowUser, WideDef));
1233 /// CreateWideIV - Process a single induction variable. First use the
1234 /// SCEVExpander to create a wide induction variable that evaluates to the same
1235 /// recurrence as the original narrow IV. Then use a worklist to forward
1236 /// traverse the narrow IV's def-use chain. After WidenIVUse has processed all
1237 /// interesting IV users, the narrow IV will be isolated for removal by
1240 /// It would be simpler to delete uses as they are processed, but we must avoid
1241 /// invalidating SCEV expressions.
1243 PHINode *WidenIV::CreateWideIV(SCEVExpander &Rewriter) {
1244 // Is this phi an induction variable?
1245 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(OrigPhi));
1249 // Widen the induction variable expression.
1250 const SCEV *WideIVExpr = IsSigned ?
1251 SE->getSignExtendExpr(AddRec, WideType) :
1252 SE->getZeroExtendExpr(AddRec, WideType);
1254 assert(SE->getEffectiveSCEVType(WideIVExpr->getType()) == WideType &&
1255 "Expect the new IV expression to preserve its type");
1257 // Can the IV be extended outside the loop without overflow?
1258 AddRec = dyn_cast<SCEVAddRecExpr>(WideIVExpr);
1259 if (!AddRec || AddRec->getLoop() != L)
1262 // An AddRec must have loop-invariant operands. Since this AddRec is
1263 // materialized by a loop header phi, the expression cannot have any post-loop
1264 // operands, so they must dominate the loop header.
1265 assert(SE->properlyDominates(AddRec->getStart(), L->getHeader()) &&
1266 SE->properlyDominates(AddRec->getStepRecurrence(*SE), L->getHeader())
1267 && "Loop header phi recurrence inputs do not dominate the loop");
1269 // The rewriter provides a value for the desired IV expression. This may
1270 // either find an existing phi or materialize a new one. Either way, we
1271 // expect a well-formed cyclic phi-with-increments. i.e. any operand not part
1272 // of the phi-SCC dominates the loop entry.
1273 Instruction *InsertPt = L->getHeader()->begin();
1274 WidePhi = cast<PHINode>(Rewriter.expandCodeFor(AddRec, WideType, InsertPt));
1276 // Remembering the WideIV increment generated by SCEVExpander allows
1277 // WidenIVUse to reuse it when widening the narrow IV's increment. We don't
1278 // employ a general reuse mechanism because the call above is the only call to
1279 // SCEVExpander. Henceforth, we produce 1-to-1 narrow to wide uses.
1280 if (BasicBlock *LatchBlock = L->getLoopLatch()) {
1282 cast<Instruction>(WidePhi->getIncomingValueForBlock(LatchBlock));
1283 WideIncExpr = SE->getSCEV(WideInc);
1286 DEBUG(dbgs() << "Wide IV: " << *WidePhi << "\n");
1289 // Traverse the def-use chain using a worklist starting at the original IV.
1290 assert(Widened.empty() && NarrowIVUsers.empty() && "expect initial state" );
1292 Widened.insert(OrigPhi);
1293 pushNarrowIVUsers(OrigPhi, WidePhi);
1295 while (!NarrowIVUsers.empty()) {
1296 NarrowIVDefUse DU = NarrowIVUsers.pop_back_val();
1298 // Process a def-use edge. This may replace the use, so don't hold a
1299 // use_iterator across it.
1300 Instruction *WideUse = WidenIVUse(DU, Rewriter);
1302 // Follow all def-use edges from the previous narrow use.
1304 pushNarrowIVUsers(DU.NarrowUse, WideUse);
1306 // WidenIVUse may have removed the def-use edge.
1307 if (DU.NarrowDef->use_empty())
1308 DeadInsts.emplace_back(DU.NarrowDef);
1313 //===----------------------------------------------------------------------===//
1314 // Live IV Reduction - Minimize IVs live across the loop.
1315 //===----------------------------------------------------------------------===//
1318 //===----------------------------------------------------------------------===//
1319 // Simplification of IV users based on SCEV evaluation.
1320 //===----------------------------------------------------------------------===//
1323 class IndVarSimplifyVisitor : public IVVisitor {
1324 ScalarEvolution *SE;
1325 const TargetTransformInfo *TTI;
1331 IndVarSimplifyVisitor(PHINode *IV, ScalarEvolution *SCEV,
1332 const TargetTransformInfo *TTI,
1333 const DominatorTree *DTree)
1334 : SE(SCEV), TTI(TTI), IVPhi(IV) {
1336 WI.NarrowIV = IVPhi;
1338 setSplitOverflowIntrinsics();
1341 // Implement the interface used by simplifyUsersOfIV.
1342 void visitCast(CastInst *Cast) override { visitIVCast(Cast, WI, SE, TTI); }
1346 /// SimplifyAndExtend - Iteratively perform simplification on a worklist of IV
1347 /// users. Each successive simplification may push more users which may
1348 /// themselves be candidates for simplification.
1350 /// Sign/Zero extend elimination is interleaved with IV simplification.
1352 void IndVarSimplify::SimplifyAndExtend(Loop *L,
1353 SCEVExpander &Rewriter,
1354 LPPassManager &LPM) {
1355 SmallVector<WideIVInfo, 8> WideIVs;
1357 SmallVector<PHINode*, 8> LoopPhis;
1358 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
1359 LoopPhis.push_back(cast<PHINode>(I));
1361 // Each round of simplification iterates through the SimplifyIVUsers worklist
1362 // for all current phis, then determines whether any IVs can be
1363 // widened. Widening adds new phis to LoopPhis, inducing another round of
1364 // simplification on the wide IVs.
1365 while (!LoopPhis.empty()) {
1366 // Evaluate as many IV expressions as possible before widening any IVs. This
1367 // forces SCEV to set no-wrap flags before evaluating sign/zero
1368 // extension. The first time SCEV attempts to normalize sign/zero extension,
1369 // the result becomes final. So for the most predictable results, we delay
1370 // evaluation of sign/zero extend evaluation until needed, and avoid running
1371 // other SCEV based analysis prior to SimplifyAndExtend.
1373 PHINode *CurrIV = LoopPhis.pop_back_val();
1375 // Information about sign/zero extensions of CurrIV.
1376 IndVarSimplifyVisitor Visitor(CurrIV, SE, TTI, DT);
1378 Changed |= simplifyUsersOfIV(CurrIV, SE, &LPM, DeadInsts, &Visitor);
1380 if (Visitor.WI.WidestNativeType) {
1381 WideIVs.push_back(Visitor.WI);
1383 } while(!LoopPhis.empty());
1385 for (; !WideIVs.empty(); WideIVs.pop_back()) {
1386 WidenIV Widener(WideIVs.back(), LI, SE, DT, DeadInsts);
1387 if (PHINode *WidePhi = Widener.CreateWideIV(Rewriter)) {
1389 LoopPhis.push_back(WidePhi);
1395 //===----------------------------------------------------------------------===//
1396 // LinearFunctionTestReplace and its kin. Rewrite the loop exit condition.
1397 //===----------------------------------------------------------------------===//
1399 /// canExpandBackedgeTakenCount - Return true if this loop's backedge taken
1400 /// count expression can be safely and cheaply expanded into an instruction
1401 /// sequence that can be used by LinearFunctionTestReplace.
1403 /// TODO: This fails for pointer-type loop counters with greater than one byte
1404 /// strides, consequently preventing LFTR from running. For the purpose of LFTR
1405 /// we could skip this check in the case that the LFTR loop counter (chosen by
1406 /// FindLoopCounter) is also pointer type. Instead, we could directly convert
1407 /// the loop test to an inequality test by checking the target data's alignment
1408 /// of element types (given that the initial pointer value originates from or is
1409 /// used by ABI constrained operation, as opposed to inttoptr/ptrtoint).
1410 /// However, we don't yet have a strong motivation for converting loop tests
1411 /// into inequality tests.
1412 static bool canExpandBackedgeTakenCount(Loop *L, ScalarEvolution *SE,
1413 SCEVExpander &Rewriter) {
1414 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
1415 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount) ||
1416 BackedgeTakenCount->isZero())
1419 if (!L->getExitingBlock())
1422 // Can't rewrite non-branch yet.
1423 if (!isa<BranchInst>(L->getExitingBlock()->getTerminator()))
1426 if (Rewriter.isHighCostExpansion(BackedgeTakenCount, L))
1432 /// getLoopPhiForCounter - Return the loop header phi IFF IncV adds a loop
1433 /// invariant value to the phi.
1434 static PHINode *getLoopPhiForCounter(Value *IncV, Loop *L, DominatorTree *DT) {
1435 Instruction *IncI = dyn_cast<Instruction>(IncV);
1439 switch (IncI->getOpcode()) {
1440 case Instruction::Add:
1441 case Instruction::Sub:
1443 case Instruction::GetElementPtr:
1444 // An IV counter must preserve its type.
1445 if (IncI->getNumOperands() == 2)
1451 PHINode *Phi = dyn_cast<PHINode>(IncI->getOperand(0));
1452 if (Phi && Phi->getParent() == L->getHeader()) {
1453 if (isLoopInvariant(IncI->getOperand(1), L, DT))
1457 if (IncI->getOpcode() == Instruction::GetElementPtr)
1460 // Allow add/sub to be commuted.
1461 Phi = dyn_cast<PHINode>(IncI->getOperand(1));
1462 if (Phi && Phi->getParent() == L->getHeader()) {
1463 if (isLoopInvariant(IncI->getOperand(0), L, DT))
1469 /// Return the compare guarding the loop latch, or NULL for unrecognized tests.
1470 static ICmpInst *getLoopTest(Loop *L) {
1471 assert(L->getExitingBlock() && "expected loop exit");
1473 BasicBlock *LatchBlock = L->getLoopLatch();
1474 // Don't bother with LFTR if the loop is not properly simplified.
1478 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
1479 assert(BI && "expected exit branch");
1481 return dyn_cast<ICmpInst>(BI->getCondition());
1484 /// needsLFTR - LinearFunctionTestReplace policy. Return true unless we can show
1485 /// that the current exit test is already sufficiently canonical.
1486 static bool needsLFTR(Loop *L, DominatorTree *DT) {
1487 // Do LFTR to simplify the exit condition to an ICMP.
1488 ICmpInst *Cond = getLoopTest(L);
1492 // Do LFTR to simplify the exit ICMP to EQ/NE
1493 ICmpInst::Predicate Pred = Cond->getPredicate();
1494 if (Pred != ICmpInst::ICMP_NE && Pred != ICmpInst::ICMP_EQ)
1497 // Look for a loop invariant RHS
1498 Value *LHS = Cond->getOperand(0);
1499 Value *RHS = Cond->getOperand(1);
1500 if (!isLoopInvariant(RHS, L, DT)) {
1501 if (!isLoopInvariant(LHS, L, DT))
1503 std::swap(LHS, RHS);
1505 // Look for a simple IV counter LHS
1506 PHINode *Phi = dyn_cast<PHINode>(LHS);
1508 Phi = getLoopPhiForCounter(LHS, L, DT);
1513 // Do LFTR if PHI node is defined in the loop, but is *not* a counter.
1514 int Idx = Phi->getBasicBlockIndex(L->getLoopLatch());
1518 // Do LFTR if the exit condition's IV is *not* a simple counter.
1519 Value *IncV = Phi->getIncomingValue(Idx);
1520 return Phi != getLoopPhiForCounter(IncV, L, DT);
1523 /// Recursive helper for hasConcreteDef(). Unfortunately, this currently boils
1524 /// down to checking that all operands are constant and listing instructions
1525 /// that may hide undef.
1526 static bool hasConcreteDefImpl(Value *V, SmallPtrSetImpl<Value*> &Visited,
1528 if (isa<Constant>(V))
1529 return !isa<UndefValue>(V);
1534 // Conservatively handle non-constant non-instructions. For example, Arguments
1536 Instruction *I = dyn_cast<Instruction>(V);
1540 // Load and return values may be undef.
1541 if(I->mayReadFromMemory() || isa<CallInst>(I) || isa<InvokeInst>(I))
1544 // Optimistically handle other instructions.
1545 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) {
1546 if (!Visited.insert(*OI).second)
1548 if (!hasConcreteDefImpl(*OI, Visited, Depth+1))
1554 /// Return true if the given value is concrete. We must prove that undef can
1557 /// TODO: If we decide that this is a good approach to checking for undef, we
1558 /// may factor it into a common location.
1559 static bool hasConcreteDef(Value *V) {
1560 SmallPtrSet<Value*, 8> Visited;
1562 return hasConcreteDefImpl(V, Visited, 0);
1565 /// AlmostDeadIV - Return true if this IV has any uses other than the (soon to
1566 /// be rewritten) loop exit test.
1567 static bool AlmostDeadIV(PHINode *Phi, BasicBlock *LatchBlock, Value *Cond) {
1568 int LatchIdx = Phi->getBasicBlockIndex(LatchBlock);
1569 Value *IncV = Phi->getIncomingValue(LatchIdx);
1571 for (User *U : Phi->users())
1572 if (U != Cond && U != IncV) return false;
1574 for (User *U : IncV->users())
1575 if (U != Cond && U != Phi) return false;
1579 /// FindLoopCounter - Find an affine IV in canonical form.
1581 /// BECount may be an i8* pointer type. The pointer difference is already
1582 /// valid count without scaling the address stride, so it remains a pointer
1583 /// expression as far as SCEV is concerned.
1585 /// Currently only valid for LFTR. See the comments on hasConcreteDef below.
1587 /// FIXME: Accept -1 stride and set IVLimit = IVInit - BECount
1589 /// FIXME: Accept non-unit stride as long as SCEV can reduce BECount * Stride.
1590 /// This is difficult in general for SCEV because of potential overflow. But we
1591 /// could at least handle constant BECounts.
1592 static PHINode *FindLoopCounter(Loop *L, const SCEV *BECount,
1593 ScalarEvolution *SE, DominatorTree *DT) {
1594 uint64_t BCWidth = SE->getTypeSizeInBits(BECount->getType());
1597 cast<BranchInst>(L->getExitingBlock()->getTerminator())->getCondition();
1599 // Loop over all of the PHI nodes, looking for a simple counter.
1600 PHINode *BestPhi = nullptr;
1601 const SCEV *BestInit = nullptr;
1602 BasicBlock *LatchBlock = L->getLoopLatch();
1603 assert(LatchBlock && "needsLFTR should guarantee a loop latch");
1605 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
1606 PHINode *Phi = cast<PHINode>(I);
1607 if (!SE->isSCEVable(Phi->getType()))
1610 // Avoid comparing an integer IV against a pointer Limit.
1611 if (BECount->getType()->isPointerTy() && !Phi->getType()->isPointerTy())
1614 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Phi));
1615 if (!AR || AR->getLoop() != L || !AR->isAffine())
1618 // AR may be a pointer type, while BECount is an integer type.
1619 // AR may be wider than BECount. With eq/ne tests overflow is immaterial.
1620 // AR may not be a narrower type, or we may never exit.
1621 uint64_t PhiWidth = SE->getTypeSizeInBits(AR->getType());
1622 if (PhiWidth < BCWidth ||
1623 !L->getHeader()->getModule()->getDataLayout().isLegalInteger(PhiWidth))
1626 const SCEV *Step = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE));
1627 if (!Step || !Step->isOne())
1630 int LatchIdx = Phi->getBasicBlockIndex(LatchBlock);
1631 Value *IncV = Phi->getIncomingValue(LatchIdx);
1632 if (getLoopPhiForCounter(IncV, L, DT) != Phi)
1635 // Avoid reusing a potentially undef value to compute other values that may
1636 // have originally had a concrete definition.
1637 if (!hasConcreteDef(Phi)) {
1638 // We explicitly allow unknown phis as long as they are already used by
1639 // the loop test. In this case we assume that performing LFTR could not
1640 // increase the number of undef users.
1641 if (ICmpInst *Cond = getLoopTest(L)) {
1642 if (Phi != getLoopPhiForCounter(Cond->getOperand(0), L, DT)
1643 && Phi != getLoopPhiForCounter(Cond->getOperand(1), L, DT)) {
1648 const SCEV *Init = AR->getStart();
1650 if (BestPhi && !AlmostDeadIV(BestPhi, LatchBlock, Cond)) {
1651 // Don't force a live loop counter if another IV can be used.
1652 if (AlmostDeadIV(Phi, LatchBlock, Cond))
1655 // Prefer to count-from-zero. This is a more "canonical" counter form. It
1656 // also prefers integer to pointer IVs.
1657 if (BestInit->isZero() != Init->isZero()) {
1658 if (BestInit->isZero())
1661 // If two IVs both count from zero or both count from nonzero then the
1662 // narrower is likely a dead phi that has been widened. Use the wider phi
1663 // to allow the other to be eliminated.
1664 else if (PhiWidth <= SE->getTypeSizeInBits(BestPhi->getType()))
1673 /// genLoopLimit - Help LinearFunctionTestReplace by generating a value that
1674 /// holds the RHS of the new loop test.
1675 static Value *genLoopLimit(PHINode *IndVar, const SCEV *IVCount, Loop *L,
1676 SCEVExpander &Rewriter, ScalarEvolution *SE) {
1677 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(IndVar));
1678 assert(AR && AR->getLoop() == L && AR->isAffine() && "bad loop counter");
1679 const SCEV *IVInit = AR->getStart();
1681 // IVInit may be a pointer while IVCount is an integer when FindLoopCounter
1682 // finds a valid pointer IV. Sign extend BECount in order to materialize a
1683 // GEP. Avoid running SCEVExpander on a new pointer value, instead reusing
1684 // the existing GEPs whenever possible.
1685 if (IndVar->getType()->isPointerTy()
1686 && !IVCount->getType()->isPointerTy()) {
1688 // IVOffset will be the new GEP offset that is interpreted by GEP as a
1689 // signed value. IVCount on the other hand represents the loop trip count,
1690 // which is an unsigned value. FindLoopCounter only allows induction
1691 // variables that have a positive unit stride of one. This means we don't
1692 // have to handle the case of negative offsets (yet) and just need to zero
1694 Type *OfsTy = SE->getEffectiveSCEVType(IVInit->getType());
1695 const SCEV *IVOffset = SE->getTruncateOrZeroExtend(IVCount, OfsTy);
1697 // Expand the code for the iteration count.
1698 assert(SE->isLoopInvariant(IVOffset, L) &&
1699 "Computed iteration count is not loop invariant!");
1700 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1701 Value *GEPOffset = Rewriter.expandCodeFor(IVOffset, OfsTy, BI);
1703 Value *GEPBase = IndVar->getIncomingValueForBlock(L->getLoopPreheader());
1704 assert(AR->getStart() == SE->getSCEV(GEPBase) && "bad loop counter");
1705 // We could handle pointer IVs other than i8*, but we need to compensate for
1706 // gep index scaling. See canExpandBackedgeTakenCount comments.
1707 assert(SE->getSizeOfExpr(IntegerType::getInt64Ty(IndVar->getContext()),
1708 cast<PointerType>(GEPBase->getType())->getElementType())->isOne()
1709 && "unit stride pointer IV must be i8*");
1711 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
1712 return Builder.CreateGEP(nullptr, GEPBase, GEPOffset, "lftr.limit");
1715 // In any other case, convert both IVInit and IVCount to integers before
1716 // comparing. This may result in SCEV expension of pointers, but in practice
1717 // SCEV will fold the pointer arithmetic away as such:
1718 // BECount = (IVEnd - IVInit - 1) => IVLimit = IVInit (postinc).
1720 // Valid Cases: (1) both integers is most common; (2) both may be pointers
1721 // for simple memset-style loops.
1723 // IVInit integer and IVCount pointer would only occur if a canonical IV
1724 // were generated on top of case #2, which is not expected.
1726 const SCEV *IVLimit = nullptr;
1727 // For unit stride, IVCount = Start + BECount with 2's complement overflow.
1728 // For non-zero Start, compute IVCount here.
1729 if (AR->getStart()->isZero())
1732 assert(AR->getStepRecurrence(*SE)->isOne() && "only handles unit stride");
1733 const SCEV *IVInit = AR->getStart();
1735 // For integer IVs, truncate the IV before computing IVInit + BECount.
1736 if (SE->getTypeSizeInBits(IVInit->getType())
1737 > SE->getTypeSizeInBits(IVCount->getType()))
1738 IVInit = SE->getTruncateExpr(IVInit, IVCount->getType());
1740 IVLimit = SE->getAddExpr(IVInit, IVCount);
1742 // Expand the code for the iteration count.
1743 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1744 IRBuilder<> Builder(BI);
1745 assert(SE->isLoopInvariant(IVLimit, L) &&
1746 "Computed iteration count is not loop invariant!");
1747 // Ensure that we generate the same type as IndVar, or a smaller integer
1748 // type. In the presence of null pointer values, we have an integer type
1749 // SCEV expression (IVInit) for a pointer type IV value (IndVar).
1750 Type *LimitTy = IVCount->getType()->isPointerTy() ?
1751 IndVar->getType() : IVCount->getType();
1752 return Rewriter.expandCodeFor(IVLimit, LimitTy, BI);
1756 /// LinearFunctionTestReplace - This method rewrites the exit condition of the
1757 /// loop to be a canonical != comparison against the incremented loop induction
1758 /// variable. This pass is able to rewrite the exit tests of any loop where the
1759 /// SCEV analysis can determine a loop-invariant trip count of the loop, which
1760 /// is actually a much broader range than just linear tests.
1761 Value *IndVarSimplify::
1762 LinearFunctionTestReplace(Loop *L,
1763 const SCEV *BackedgeTakenCount,
1765 SCEVExpander &Rewriter) {
1766 assert(canExpandBackedgeTakenCount(L, SE, Rewriter) && "precondition");
1768 // Initialize CmpIndVar and IVCount to their preincremented values.
1769 Value *CmpIndVar = IndVar;
1770 const SCEV *IVCount = BackedgeTakenCount;
1772 // If the exiting block is the same as the backedge block, we prefer to
1773 // compare against the post-incremented value, otherwise we must compare
1774 // against the preincremented value.
1775 if (L->getExitingBlock() == L->getLoopLatch()) {
1776 // Add one to the "backedge-taken" count to get the trip count.
1777 // This addition may overflow, which is valid as long as the comparison is
1778 // truncated to BackedgeTakenCount->getType().
1779 IVCount = SE->getAddExpr(BackedgeTakenCount,
1780 SE->getConstant(BackedgeTakenCount->getType(), 1));
1781 // The BackedgeTaken expression contains the number of times that the
1782 // backedge branches to the loop header. This is one less than the
1783 // number of times the loop executes, so use the incremented indvar.
1784 CmpIndVar = IndVar->getIncomingValueForBlock(L->getExitingBlock());
1787 Value *ExitCnt = genLoopLimit(IndVar, IVCount, L, Rewriter, SE);
1788 assert(ExitCnt->getType()->isPointerTy() == IndVar->getType()->isPointerTy()
1789 && "genLoopLimit missed a cast");
1791 // Insert a new icmp_ne or icmp_eq instruction before the branch.
1792 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1793 ICmpInst::Predicate P;
1794 if (L->contains(BI->getSuccessor(0)))
1795 P = ICmpInst::ICMP_NE;
1797 P = ICmpInst::ICMP_EQ;
1799 DEBUG(dbgs() << "INDVARS: Rewriting loop exit condition to:\n"
1800 << " LHS:" << *CmpIndVar << '\n'
1802 << (P == ICmpInst::ICMP_NE ? "!=" : "==") << "\n"
1803 << " RHS:\t" << *ExitCnt << "\n"
1804 << " IVCount:\t" << *IVCount << "\n");
1806 IRBuilder<> Builder(BI);
1808 // LFTR can ignore IV overflow and truncate to the width of
1809 // BECount. This avoids materializing the add(zext(add)) expression.
1810 unsigned CmpIndVarSize = SE->getTypeSizeInBits(CmpIndVar->getType());
1811 unsigned ExitCntSize = SE->getTypeSizeInBits(ExitCnt->getType());
1812 if (CmpIndVarSize > ExitCntSize) {
1813 const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(SE->getSCEV(IndVar));
1814 const SCEV *ARStart = AR->getStart();
1815 const SCEV *ARStep = AR->getStepRecurrence(*SE);
1816 // For constant IVCount, avoid truncation.
1817 if (isa<SCEVConstant>(ARStart) && isa<SCEVConstant>(IVCount)) {
1818 const APInt &Start = cast<SCEVConstant>(ARStart)->getValue()->getValue();
1819 APInt Count = cast<SCEVConstant>(IVCount)->getValue()->getValue();
1820 // Note that the post-inc value of BackedgeTakenCount may have overflowed
1821 // above such that IVCount is now zero.
1822 if (IVCount != BackedgeTakenCount && Count == 0) {
1823 Count = APInt::getMaxValue(Count.getBitWidth()).zext(CmpIndVarSize);
1827 Count = Count.zext(CmpIndVarSize);
1829 if (cast<SCEVConstant>(ARStep)->getValue()->isNegative())
1830 NewLimit = Start - Count;
1832 NewLimit = Start + Count;
1833 ExitCnt = ConstantInt::get(CmpIndVar->getType(), NewLimit);
1835 DEBUG(dbgs() << " Widen RHS:\t" << *ExitCnt << "\n");
1837 CmpIndVar = Builder.CreateTrunc(CmpIndVar, ExitCnt->getType(),
1841 Value *Cond = Builder.CreateICmp(P, CmpIndVar, ExitCnt, "exitcond");
1842 Value *OrigCond = BI->getCondition();
1843 // It's tempting to use replaceAllUsesWith here to fully replace the old
1844 // comparison, but that's not immediately safe, since users of the old
1845 // comparison may not be dominated by the new comparison. Instead, just
1846 // update the branch to use the new comparison; in the common case this
1847 // will make old comparison dead.
1848 BI->setCondition(Cond);
1849 DeadInsts.push_back(OrigCond);
1856 //===----------------------------------------------------------------------===//
1857 // SinkUnusedInvariants. A late subpass to cleanup loop preheaders.
1858 //===----------------------------------------------------------------------===//
1860 /// If there's a single exit block, sink any loop-invariant values that
1861 /// were defined in the preheader but not used inside the loop into the
1862 /// exit block to reduce register pressure in the loop.
1863 void IndVarSimplify::SinkUnusedInvariants(Loop *L) {
1864 BasicBlock *ExitBlock = L->getExitBlock();
1865 if (!ExitBlock) return;
1867 BasicBlock *Preheader = L->getLoopPreheader();
1868 if (!Preheader) return;
1870 Instruction *InsertPt = ExitBlock->getFirstInsertionPt();
1871 BasicBlock::iterator I = Preheader->getTerminator();
1872 while (I != Preheader->begin()) {
1874 // New instructions were inserted at the end of the preheader.
1875 if (isa<PHINode>(I))
1878 // Don't move instructions which might have side effects, since the side
1879 // effects need to complete before instructions inside the loop. Also don't
1880 // move instructions which might read memory, since the loop may modify
1881 // memory. Note that it's okay if the instruction might have undefined
1882 // behavior: LoopSimplify guarantees that the preheader dominates the exit
1884 if (I->mayHaveSideEffects() || I->mayReadFromMemory())
1887 // Skip debug info intrinsics.
1888 if (isa<DbgInfoIntrinsic>(I))
1891 // Skip eh pad instructions.
1895 // Don't sink alloca: we never want to sink static alloca's out of the
1896 // entry block, and correctly sinking dynamic alloca's requires
1897 // checks for stacksave/stackrestore intrinsics.
1898 // FIXME: Refactor this check somehow?
1899 if (isa<AllocaInst>(I))
1902 // Determine if there is a use in or before the loop (direct or
1904 bool UsedInLoop = false;
1905 for (Use &U : I->uses()) {
1906 Instruction *User = cast<Instruction>(U.getUser());
1907 BasicBlock *UseBB = User->getParent();
1908 if (PHINode *P = dyn_cast<PHINode>(User)) {
1910 PHINode::getIncomingValueNumForOperand(U.getOperandNo());
1911 UseBB = P->getIncomingBlock(i);
1913 if (UseBB == Preheader || L->contains(UseBB)) {
1919 // If there is, the def must remain in the preheader.
1923 // Otherwise, sink it to the exit block.
1924 Instruction *ToMove = I;
1927 if (I != Preheader->begin()) {
1928 // Skip debug info intrinsics.
1931 } while (isa<DbgInfoIntrinsic>(I) && I != Preheader->begin());
1933 if (isa<DbgInfoIntrinsic>(I) && I == Preheader->begin())
1939 ToMove->moveBefore(InsertPt);
1945 //===----------------------------------------------------------------------===//
1946 // IndVarSimplify driver. Manage several subpasses of IV simplification.
1947 //===----------------------------------------------------------------------===//
1949 bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) {
1950 if (skipOptnoneFunction(L))
1953 // If LoopSimplify form is not available, stay out of trouble. Some notes:
1954 // - LSR currently only supports LoopSimplify-form loops. Indvars'
1955 // canonicalization can be a pessimization without LSR to "clean up"
1957 // - We depend on having a preheader; in particular,
1958 // Loop::getCanonicalInductionVariable only supports loops with preheaders,
1959 // and we're in trouble if we can't find the induction variable even when
1960 // we've manually inserted one.
1961 if (!L->isLoopSimplifyForm())
1964 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1965 SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
1966 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1967 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1968 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1969 auto *TTIP = getAnalysisIfAvailable<TargetTransformInfoWrapperPass>();
1970 TTI = TTIP ? &TTIP->getTTI(*L->getHeader()->getParent()) : nullptr;
1971 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
1976 // If there are any floating-point recurrences, attempt to
1977 // transform them to use integer recurrences.
1978 RewriteNonIntegerIVs(L);
1980 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
1982 // Create a rewriter object which we'll use to transform the code with.
1983 SCEVExpander Rewriter(*SE, DL, "indvars");
1985 Rewriter.setDebugType(DEBUG_TYPE);
1988 // Eliminate redundant IV users.
1990 // Simplification works best when run before other consumers of SCEV. We
1991 // attempt to avoid evaluating SCEVs for sign/zero extend operations until
1992 // other expressions involving loop IVs have been evaluated. This helps SCEV
1993 // set no-wrap flags before normalizing sign/zero extension.
1994 Rewriter.disableCanonicalMode();
1995 SimplifyAndExtend(L, Rewriter, LPM);
1997 // Check to see if this loop has a computable loop-invariant execution count.
1998 // If so, this means that we can compute the final value of any expressions
1999 // that are recurrent in the loop, and substitute the exit values from the
2000 // loop into any instructions outside of the loop that use the final values of
2001 // the current expressions.
2003 if (ReplaceExitValue != NeverRepl &&
2004 !isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2005 RewriteLoopExitValues(L, Rewriter);
2007 // Eliminate redundant IV cycles.
2008 NumElimIV += Rewriter.replaceCongruentIVs(L, DT, DeadInsts);
2010 // If we have a trip count expression, rewrite the loop's exit condition
2011 // using it. We can currently only handle loops with a single exit.
2012 if (canExpandBackedgeTakenCount(L, SE, Rewriter) && needsLFTR(L, DT)) {
2013 PHINode *IndVar = FindLoopCounter(L, BackedgeTakenCount, SE, DT);
2015 // Check preconditions for proper SCEVExpander operation. SCEV does not
2016 // express SCEVExpander's dependencies, such as LoopSimplify. Instead any
2017 // pass that uses the SCEVExpander must do it. This does not work well for
2018 // loop passes because SCEVExpander makes assumptions about all loops,
2019 // while LoopPassManager only forces the current loop to be simplified.
2021 // FIXME: SCEV expansion has no way to bail out, so the caller must
2022 // explicitly check any assumptions made by SCEV. Brittle.
2023 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(BackedgeTakenCount);
2024 if (!AR || AR->getLoop()->getLoopPreheader())
2025 (void)LinearFunctionTestReplace(L, BackedgeTakenCount, IndVar,
2029 // Clear the rewriter cache, because values that are in the rewriter's cache
2030 // can be deleted in the loop below, causing the AssertingVH in the cache to
2034 // Now that we're done iterating through lists, clean up any instructions
2035 // which are now dead.
2036 while (!DeadInsts.empty())
2037 if (Instruction *Inst =
2038 dyn_cast_or_null<Instruction>(DeadInsts.pop_back_val()))
2039 RecursivelyDeleteTriviallyDeadInstructions(Inst, TLI);
2041 // The Rewriter may not be used from this point on.
2043 // Loop-invariant instructions in the preheader that aren't used in the
2044 // loop may be sunk below the loop to reduce register pressure.
2045 SinkUnusedInvariants(L);
2047 // Clean up dead instructions.
2048 Changed |= DeleteDeadPHIs(L->getHeader(), TLI);
2049 // Check a post-condition.
2050 assert(L->isLCSSAForm(*DT) &&
2051 "Indvars did not leave the loop in lcssa form!");
2053 // Verify that LFTR, and any other change have not interfered with SCEV's
2054 // ability to compute trip count.
2056 if (VerifyIndvars && !isa<SCEVCouldNotCompute>(BackedgeTakenCount)) {
2058 const SCEV *NewBECount = SE->getBackedgeTakenCount(L);
2059 if (SE->getTypeSizeInBits(BackedgeTakenCount->getType()) <
2060 SE->getTypeSizeInBits(NewBECount->getType()))
2061 NewBECount = SE->getTruncateOrNoop(NewBECount,
2062 BackedgeTakenCount->getType());
2064 BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount,
2065 NewBECount->getType());
2066 assert(BackedgeTakenCount == NewBECount && "indvars must preserve SCEV");