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/ScalarEvolutionAliasAnalysis.h"
35 #include "llvm/Analysis/TargetLibraryInfo.h"
36 #include "llvm/Analysis/TargetTransformInfo.h"
37 #include "llvm/IR/BasicBlock.h"
38 #include "llvm/IR/CFG.h"
39 #include "llvm/IR/Constants.h"
40 #include "llvm/IR/DataLayout.h"
41 #include "llvm/IR/Dominators.h"
42 #include "llvm/IR/Instructions.h"
43 #include "llvm/IR/IntrinsicInst.h"
44 #include "llvm/IR/LLVMContext.h"
45 #include "llvm/IR/PatternMatch.h"
46 #include "llvm/IR/Type.h"
47 #include "llvm/Support/CommandLine.h"
48 #include "llvm/Support/Debug.h"
49 #include "llvm/Support/raw_ostream.h"
50 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
51 #include "llvm/Transforms/Utils/Local.h"
52 #include "llvm/Transforms/Utils/SimplifyIndVar.h"
55 #define DEBUG_TYPE "indvars"
57 STATISTIC(NumWidened , "Number of indvars widened");
58 STATISTIC(NumReplaced , "Number of exit values replaced");
59 STATISTIC(NumLFTR , "Number of loop exit tests replaced");
60 STATISTIC(NumElimExt , "Number of IV sign/zero extends eliminated");
61 STATISTIC(NumElimIV , "Number of congruent IVs eliminated");
63 // Trip count verification can be enabled by default under NDEBUG if we
64 // implement a strong expression equivalence checker in SCEV. Until then, we
65 // use the verify-indvars flag, which may assert in some cases.
66 static cl::opt<bool> VerifyIndvars(
67 "verify-indvars", cl::Hidden,
68 cl::desc("Verify the ScalarEvolution result after running indvars"));
70 static cl::opt<bool> ReduceLiveIVs("liv-reduce", cl::Hidden,
71 cl::desc("Reduce live induction variables."));
73 enum ReplaceExitVal { NeverRepl, OnlyCheapRepl, AlwaysRepl };
75 static cl::opt<ReplaceExitVal> ReplaceExitValue(
76 "replexitval", cl::Hidden, cl::init(OnlyCheapRepl),
77 cl::desc("Choose the strategy to replace exit value in IndVarSimplify"),
78 cl::values(clEnumValN(NeverRepl, "never", "never replace exit value"),
79 clEnumValN(OnlyCheapRepl, "cheap",
80 "only replace exit value when the cost is cheap"),
81 clEnumValN(AlwaysRepl, "always",
82 "always replace exit value whenever possible"),
90 class IndVarSimplify : public LoopPass {
94 TargetLibraryInfo *TLI;
95 const TargetTransformInfo *TTI;
97 SmallVector<WeakVH, 16> DeadInsts;
101 static char ID; // Pass identification, replacement for typeid
103 : LoopPass(ID), LI(nullptr), SE(nullptr), DT(nullptr), Changed(false) {
104 initializeIndVarSimplifyPass(*PassRegistry::getPassRegistry());
107 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
109 void getAnalysisUsage(AnalysisUsage &AU) const override {
110 AU.addRequired<DominatorTreeWrapperPass>();
111 AU.addRequired<LoopInfoWrapperPass>();
112 AU.addRequired<ScalarEvolutionWrapperPass>();
113 AU.addRequiredID(LoopSimplifyID);
114 AU.addRequiredID(LCSSAID);
115 AU.addPreserved<ScalarEvolutionWrapperPass>();
116 AU.addPreservedID(LoopSimplifyID);
117 AU.addPreservedID(LCSSAID);
118 AU.setPreservesCFG();
122 void releaseMemory() override {
126 bool isValidRewrite(Value *FromVal, Value *ToVal);
128 void HandleFloatingPointIV(Loop *L, PHINode *PH);
129 void RewriteNonIntegerIVs(Loop *L);
131 void SimplifyAndExtend(Loop *L, SCEVExpander &Rewriter, LPPassManager &LPM);
133 bool CanLoopBeDeleted(Loop *L, SmallVector<RewritePhi, 8> &RewritePhiSet);
134 void RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter);
136 Value *LinearFunctionTestReplace(Loop *L, const SCEV *BackedgeTakenCount,
137 PHINode *IndVar, SCEVExpander &Rewriter);
139 void SinkUnusedInvariants(Loop *L);
141 Value *ExpandSCEVIfNeeded(SCEVExpander &Rewriter, const SCEV *S, Loop *L,
142 Instruction *InsertPt, Type *Ty);
146 char IndVarSimplify::ID = 0;
147 INITIALIZE_PASS_BEGIN(IndVarSimplify, "indvars",
148 "Induction Variable Simplification", false, false)
149 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
150 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
151 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
152 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
153 INITIALIZE_PASS_DEPENDENCY(LCSSA)
154 INITIALIZE_PASS_END(IndVarSimplify, "indvars",
155 "Induction Variable Simplification", false, false)
157 Pass *llvm::createIndVarSimplifyPass() {
158 return new IndVarSimplify();
161 /// isValidRewrite - Return true if the SCEV expansion generated by the
162 /// rewriter can replace the original value. SCEV guarantees that it
163 /// produces the same value, but the way it is produced may be illegal IR.
164 /// Ideally, this function will only be called for verification.
165 bool IndVarSimplify::isValidRewrite(Value *FromVal, Value *ToVal) {
166 // If an SCEV expression subsumed multiple pointers, its expansion could
167 // reassociate the GEP changing the base pointer. This is illegal because the
168 // final address produced by a GEP chain must be inbounds relative to its
169 // underlying object. Otherwise basic alias analysis, among other things,
170 // could fail in a dangerous way. Ultimately, SCEV will be improved to avoid
171 // producing an expression involving multiple pointers. Until then, we must
174 // Retrieve the pointer operand of the GEP. Don't use GetUnderlyingObject
175 // because it understands lcssa phis while SCEV does not.
176 Value *FromPtr = FromVal;
177 Value *ToPtr = ToVal;
178 if (GEPOperator *GEP = dyn_cast<GEPOperator>(FromVal)) {
179 FromPtr = GEP->getPointerOperand();
181 if (GEPOperator *GEP = dyn_cast<GEPOperator>(ToVal)) {
182 ToPtr = GEP->getPointerOperand();
184 if (FromPtr != FromVal || ToPtr != ToVal) {
185 // Quickly check the common case
186 if (FromPtr == ToPtr)
189 // SCEV may have rewritten an expression that produces the GEP's pointer
190 // operand. That's ok as long as the pointer operand has the same base
191 // pointer. Unlike GetUnderlyingObject(), getPointerBase() will find the
192 // base of a recurrence. This handles the case in which SCEV expansion
193 // converts a pointer type recurrence into a nonrecurrent pointer base
194 // indexed by an integer recurrence.
196 // If the GEP base pointer is a vector of pointers, abort.
197 if (!FromPtr->getType()->isPointerTy() || !ToPtr->getType()->isPointerTy())
200 const SCEV *FromBase = SE->getPointerBase(SE->getSCEV(FromPtr));
201 const SCEV *ToBase = SE->getPointerBase(SE->getSCEV(ToPtr));
202 if (FromBase == ToBase)
205 DEBUG(dbgs() << "INDVARS: GEP rewrite bail out "
206 << *FromBase << " != " << *ToBase << "\n");
213 /// Determine the insertion point for this user. By default, insert immediately
214 /// before the user. SCEVExpander or LICM will hoist loop invariants out of the
215 /// loop. For PHI nodes, there may be multiple uses, so compute the nearest
216 /// common dominator for the incoming blocks.
217 static Instruction *getInsertPointForUses(Instruction *User, Value *Def,
219 PHINode *PHI = dyn_cast<PHINode>(User);
223 Instruction *InsertPt = nullptr;
224 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i) {
225 if (PHI->getIncomingValue(i) != Def)
228 BasicBlock *InsertBB = PHI->getIncomingBlock(i);
230 InsertPt = InsertBB->getTerminator();
233 InsertBB = DT->findNearestCommonDominator(InsertPt->getParent(), InsertBB);
234 InsertPt = InsertBB->getTerminator();
236 assert(InsertPt && "Missing phi operand");
237 assert((!isa<Instruction>(Def) ||
238 DT->dominates(cast<Instruction>(Def), InsertPt)) &&
239 "def does not dominate all uses");
243 //===----------------------------------------------------------------------===//
244 // RewriteNonIntegerIVs and helpers. Prefer integer IVs.
245 //===----------------------------------------------------------------------===//
247 /// ConvertToSInt - Convert APF to an integer, if possible.
248 static bool ConvertToSInt(const APFloat &APF, int64_t &IntVal) {
249 bool isExact = false;
250 // See if we can convert this to an int64_t
252 if (APF.convertToInteger(&UIntVal, 64, true, APFloat::rmTowardZero,
253 &isExact) != APFloat::opOK || !isExact)
259 /// HandleFloatingPointIV - If the loop has floating induction variable
260 /// then insert corresponding integer induction variable if possible.
262 /// for(double i = 0; i < 10000; ++i)
264 /// is converted into
265 /// for(int i = 0; i < 10000; ++i)
268 void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PN) {
269 unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
270 unsigned BackEdge = IncomingEdge^1;
272 // Check incoming value.
273 ConstantFP *InitValueVal =
274 dyn_cast<ConstantFP>(PN->getIncomingValue(IncomingEdge));
277 if (!InitValueVal || !ConvertToSInt(InitValueVal->getValueAPF(), InitValue))
280 // Check IV increment. Reject this PN if increment operation is not
281 // an add or increment value can not be represented by an integer.
282 BinaryOperator *Incr =
283 dyn_cast<BinaryOperator>(PN->getIncomingValue(BackEdge));
284 if (Incr == nullptr || Incr->getOpcode() != Instruction::FAdd) return;
286 // If this is not an add of the PHI with a constantfp, or if the constant fp
287 // is not an integer, bail out.
288 ConstantFP *IncValueVal = dyn_cast<ConstantFP>(Incr->getOperand(1));
290 if (IncValueVal == nullptr || Incr->getOperand(0) != PN ||
291 !ConvertToSInt(IncValueVal->getValueAPF(), IncValue))
294 // Check Incr uses. One user is PN and the other user is an exit condition
295 // used by the conditional terminator.
296 Value::user_iterator IncrUse = Incr->user_begin();
297 Instruction *U1 = cast<Instruction>(*IncrUse++);
298 if (IncrUse == Incr->user_end()) return;
299 Instruction *U2 = cast<Instruction>(*IncrUse++);
300 if (IncrUse != Incr->user_end()) return;
302 // Find exit condition, which is an fcmp. If it doesn't exist, or if it isn't
303 // only used by a branch, we can't transform it.
304 FCmpInst *Compare = dyn_cast<FCmpInst>(U1);
306 Compare = dyn_cast<FCmpInst>(U2);
307 if (!Compare || !Compare->hasOneUse() ||
308 !isa<BranchInst>(Compare->user_back()))
311 BranchInst *TheBr = cast<BranchInst>(Compare->user_back());
313 // We need to verify that the branch actually controls the iteration count
314 // of the loop. If not, the new IV can overflow and no one will notice.
315 // The branch block must be in the loop and one of the successors must be out
317 assert(TheBr->isConditional() && "Can't use fcmp if not conditional");
318 if (!L->contains(TheBr->getParent()) ||
319 (L->contains(TheBr->getSuccessor(0)) &&
320 L->contains(TheBr->getSuccessor(1))))
324 // If it isn't a comparison with an integer-as-fp (the exit value), we can't
326 ConstantFP *ExitValueVal = dyn_cast<ConstantFP>(Compare->getOperand(1));
328 if (ExitValueVal == nullptr ||
329 !ConvertToSInt(ExitValueVal->getValueAPF(), ExitValue))
332 // Find new predicate for integer comparison.
333 CmpInst::Predicate NewPred = CmpInst::BAD_ICMP_PREDICATE;
334 switch (Compare->getPredicate()) {
335 default: return; // Unknown comparison.
336 case CmpInst::FCMP_OEQ:
337 case CmpInst::FCMP_UEQ: NewPred = CmpInst::ICMP_EQ; break;
338 case CmpInst::FCMP_ONE:
339 case CmpInst::FCMP_UNE: NewPred = CmpInst::ICMP_NE; break;
340 case CmpInst::FCMP_OGT:
341 case CmpInst::FCMP_UGT: NewPred = CmpInst::ICMP_SGT; break;
342 case CmpInst::FCMP_OGE:
343 case CmpInst::FCMP_UGE: NewPred = CmpInst::ICMP_SGE; break;
344 case CmpInst::FCMP_OLT:
345 case CmpInst::FCMP_ULT: NewPred = CmpInst::ICMP_SLT; break;
346 case CmpInst::FCMP_OLE:
347 case CmpInst::FCMP_ULE: NewPred = CmpInst::ICMP_SLE; break;
350 // We convert the floating point induction variable to a signed i32 value if
351 // we can. This is only safe if the comparison will not overflow in a way
352 // that won't be trapped by the integer equivalent operations. Check for this
354 // TODO: We could use i64 if it is native and the range requires it.
356 // The start/stride/exit values must all fit in signed i32.
357 if (!isInt<32>(InitValue) || !isInt<32>(IncValue) || !isInt<32>(ExitValue))
360 // If not actually striding (add x, 0.0), avoid touching the code.
364 // Positive and negative strides have different safety conditions.
366 // If we have a positive stride, we require the init to be less than the
368 if (InitValue >= ExitValue)
371 uint32_t Range = uint32_t(ExitValue-InitValue);
372 // Check for infinite loop, either:
373 // while (i <= Exit) or until (i > Exit)
374 if (NewPred == CmpInst::ICMP_SLE || NewPred == CmpInst::ICMP_SGT) {
375 if (++Range == 0) return; // Range overflows.
378 unsigned Leftover = Range % uint32_t(IncValue);
380 // If this is an equality comparison, we require that the strided value
381 // exactly land on the exit value, otherwise the IV condition will wrap
382 // around and do things the fp IV wouldn't.
383 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
387 // If the stride would wrap around the i32 before exiting, we can't
389 if (Leftover != 0 && int32_t(ExitValue+IncValue) < ExitValue)
393 // If we have a negative stride, we require the init to be greater than the
395 if (InitValue <= ExitValue)
398 uint32_t Range = uint32_t(InitValue-ExitValue);
399 // Check for infinite loop, either:
400 // while (i >= Exit) or until (i < Exit)
401 if (NewPred == CmpInst::ICMP_SGE || NewPred == CmpInst::ICMP_SLT) {
402 if (++Range == 0) return; // Range overflows.
405 unsigned Leftover = Range % uint32_t(-IncValue);
407 // If this is an equality comparison, we require that the strided value
408 // exactly land on the exit value, otherwise the IV condition will wrap
409 // around and do things the fp IV wouldn't.
410 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
414 // If the stride would wrap around the i32 before exiting, we can't
416 if (Leftover != 0 && int32_t(ExitValue+IncValue) > ExitValue)
420 IntegerType *Int32Ty = Type::getInt32Ty(PN->getContext());
422 // Insert new integer induction variable.
423 PHINode *NewPHI = PHINode::Create(Int32Ty, 2, PN->getName()+".int", PN);
424 NewPHI->addIncoming(ConstantInt::get(Int32Ty, InitValue),
425 PN->getIncomingBlock(IncomingEdge));
428 BinaryOperator::CreateAdd(NewPHI, ConstantInt::get(Int32Ty, IncValue),
429 Incr->getName()+".int", Incr);
430 NewPHI->addIncoming(NewAdd, PN->getIncomingBlock(BackEdge));
432 ICmpInst *NewCompare = new ICmpInst(TheBr, NewPred, NewAdd,
433 ConstantInt::get(Int32Ty, ExitValue),
436 // In the following deletions, PN may become dead and may be deleted.
437 // Use a WeakVH to observe whether this happens.
440 // Delete the old floating point exit comparison. The branch starts using the
442 NewCompare->takeName(Compare);
443 Compare->replaceAllUsesWith(NewCompare);
444 RecursivelyDeleteTriviallyDeadInstructions(Compare, TLI);
446 // Delete the old floating point increment.
447 Incr->replaceAllUsesWith(UndefValue::get(Incr->getType()));
448 RecursivelyDeleteTriviallyDeadInstructions(Incr, TLI);
450 // If the FP induction variable still has uses, this is because something else
451 // in the loop uses its value. In order to canonicalize the induction
452 // variable, we chose to eliminate the IV and rewrite it in terms of an
455 // We give preference to sitofp over uitofp because it is faster on most
458 Value *Conv = new SIToFPInst(NewPHI, PN->getType(), "indvar.conv",
459 PN->getParent()->getFirstInsertionPt());
460 PN->replaceAllUsesWith(Conv);
461 RecursivelyDeleteTriviallyDeadInstructions(PN, TLI);
466 void IndVarSimplify::RewriteNonIntegerIVs(Loop *L) {
467 // First step. Check to see if there are any floating-point recurrences.
468 // If there are, change them into integer recurrences, permitting analysis by
469 // the SCEV routines.
471 BasicBlock *Header = L->getHeader();
473 SmallVector<WeakVH, 8> PHIs;
474 for (BasicBlock::iterator I = Header->begin();
475 PHINode *PN = dyn_cast<PHINode>(I); ++I)
478 for (unsigned i = 0, e = PHIs.size(); i != e; ++i)
479 if (PHINode *PN = dyn_cast_or_null<PHINode>(&*PHIs[i]))
480 HandleFloatingPointIV(L, PN);
482 // If the loop previously had floating-point IV, ScalarEvolution
483 // may not have been able to compute a trip count. Now that we've done some
484 // re-writing, the trip count may be computable.
490 // Collect information about PHI nodes which can be transformed in
491 // RewriteLoopExitValues.
494 unsigned Ith; // Ith incoming value.
495 Value *Val; // Exit value after expansion.
496 bool HighCost; // High Cost when expansion.
497 bool SafePhi; // LCSSASafePhiForRAUW.
499 RewritePhi(PHINode *P, unsigned I, Value *V, bool H, bool S)
500 : PN(P), Ith(I), Val(V), HighCost(H), SafePhi(S) {}
504 Value *IndVarSimplify::ExpandSCEVIfNeeded(SCEVExpander &Rewriter, const SCEV *S,
505 Loop *L, Instruction *InsertPt,
507 // Before expanding S into an expensive LLVM expression, see if we can use an
508 // already existing value as the expansion for S.
509 if (Value *RetValue = Rewriter.findExistingExpansion(S, InsertPt, L))
512 // We didn't find anything, fall back to using SCEVExpander.
513 return Rewriter.expandCodeFor(S, ResultTy, InsertPt);
516 //===----------------------------------------------------------------------===//
517 // RewriteLoopExitValues - Optimize IV users outside the loop.
518 // As a side effect, reduces the amount of IV processing within the loop.
519 //===----------------------------------------------------------------------===//
521 /// RewriteLoopExitValues - Check to see if this loop has a computable
522 /// loop-invariant execution count. If so, this means that we can compute the
523 /// final value of any expressions that are recurrent in the loop, and
524 /// substitute the exit values from the loop into any instructions outside of
525 /// the loop that use the final values of the current expressions.
527 /// This is mostly redundant with the regular IndVarSimplify activities that
528 /// happen later, except that it's more powerful in some cases, because it's
529 /// able to brute-force evaluate arbitrary instructions as long as they have
530 /// constant operands at the beginning of the loop.
531 void IndVarSimplify::RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter) {
532 // Verify the input to the pass in already in LCSSA form.
533 assert(L->isLCSSAForm(*DT));
535 SmallVector<BasicBlock*, 8> ExitBlocks;
536 L->getUniqueExitBlocks(ExitBlocks);
538 SmallVector<RewritePhi, 8> RewritePhiSet;
539 // Find all values that are computed inside the loop, but used outside of it.
540 // Because of LCSSA, these values will only occur in LCSSA PHI Nodes. Scan
541 // the exit blocks of the loop to find them.
542 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
543 BasicBlock *ExitBB = ExitBlocks[i];
545 // If there are no PHI nodes in this exit block, then no values defined
546 // inside the loop are used on this path, skip it.
547 PHINode *PN = dyn_cast<PHINode>(ExitBB->begin());
550 unsigned NumPreds = PN->getNumIncomingValues();
552 // We would like to be able to RAUW single-incoming value PHI nodes. We
553 // have to be certain this is safe even when this is an LCSSA PHI node.
554 // While the computed exit value is no longer varying in *this* loop, the
555 // exit block may be an exit block for an outer containing loop as well,
556 // the exit value may be varying in the outer loop, and thus it may still
557 // require an LCSSA PHI node. The safe case is when this is
558 // single-predecessor PHI node (LCSSA) and the exit block containing it is
559 // part of the enclosing loop, or this is the outer most loop of the nest.
560 // In either case the exit value could (at most) be varying in the same
561 // loop body as the phi node itself. Thus if it is in turn used outside of
562 // an enclosing loop it will only be via a separate LCSSA node.
563 bool LCSSASafePhiForRAUW =
565 (!L->getParentLoop() || L->getParentLoop() == LI->getLoopFor(ExitBB));
567 // Iterate over all of the PHI nodes.
568 BasicBlock::iterator BBI = ExitBB->begin();
569 while ((PN = dyn_cast<PHINode>(BBI++))) {
571 continue; // dead use, don't replace it
573 // SCEV only supports integer expressions for now.
574 if (!PN->getType()->isIntegerTy() && !PN->getType()->isPointerTy())
577 // It's necessary to tell ScalarEvolution about this explicitly so that
578 // it can walk the def-use list and forget all SCEVs, as it may not be
579 // watching the PHI itself. Once the new exit value is in place, there
580 // may not be a def-use connection between the loop and every instruction
581 // which got a SCEVAddRecExpr for that loop.
584 // Iterate over all of the values in all the PHI nodes.
585 for (unsigned i = 0; i != NumPreds; ++i) {
586 // If the value being merged in is not integer or is not defined
587 // in the loop, skip it.
588 Value *InVal = PN->getIncomingValue(i);
589 if (!isa<Instruction>(InVal))
592 // If this pred is for a subloop, not L itself, skip it.
593 if (LI->getLoopFor(PN->getIncomingBlock(i)) != L)
594 continue; // The Block is in a subloop, skip it.
596 // Check that InVal is defined in the loop.
597 Instruction *Inst = cast<Instruction>(InVal);
598 if (!L->contains(Inst))
601 // Okay, this instruction has a user outside of the current loop
602 // and varies predictably *inside* the loop. Evaluate the value it
603 // contains when the loop exits, if possible.
604 const SCEV *ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop());
605 if (!SE->isLoopInvariant(ExitValue, L) ||
606 !isSafeToExpand(ExitValue, *SE))
609 // Computing the value outside of the loop brings no benefit if :
610 // - it is definitely used inside the loop in a way which can not be
612 // - no use outside of the loop can take advantage of hoisting the
613 // computation out of the loop
614 if (ExitValue->getSCEVType()>=scMulExpr) {
615 unsigned NumHardInternalUses = 0;
616 unsigned NumSoftExternalUses = 0;
617 unsigned NumUses = 0;
618 for (auto IB = Inst->user_begin(), IE = Inst->user_end();
619 IB != IE && NumUses <= 6; ++IB) {
620 Instruction *UseInstr = cast<Instruction>(*IB);
621 unsigned Opc = UseInstr->getOpcode();
623 if (L->contains(UseInstr)) {
624 if (Opc == Instruction::Call || Opc == Instruction::Ret)
625 NumHardInternalUses++;
627 if (Opc == Instruction::PHI) {
628 // Do not count the Phi as a use. LCSSA may have inserted
629 // plenty of trivial ones.
631 for (auto PB = UseInstr->user_begin(),
632 PE = UseInstr->user_end();
633 PB != PE && NumUses <= 6; ++PB, ++NumUses) {
634 unsigned PhiOpc = cast<Instruction>(*PB)->getOpcode();
635 if (PhiOpc != Instruction::Call && PhiOpc != Instruction::Ret)
636 NumSoftExternalUses++;
640 if (Opc != Instruction::Call && Opc != Instruction::Ret)
641 NumSoftExternalUses++;
644 if (NumUses <= 6 && NumHardInternalUses && !NumSoftExternalUses)
648 bool HighCost = Rewriter.isHighCostExpansion(ExitValue, L, Inst);
650 ExpandSCEVIfNeeded(Rewriter, ExitValue, L, Inst, PN->getType());
652 DEBUG(dbgs() << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal << '\n'
653 << " LoopVal = " << *Inst << "\n");
655 if (!isValidRewrite(Inst, ExitVal)) {
656 DeadInsts.push_back(ExitVal);
660 // Collect all the candidate PHINodes to be rewritten.
661 RewritePhiSet.push_back(
662 RewritePhi(PN, i, ExitVal, HighCost, LCSSASafePhiForRAUW));
667 bool LoopCanBeDel = CanLoopBeDeleted(L, RewritePhiSet);
670 for (const RewritePhi &Phi : RewritePhiSet) {
671 PHINode *PN = Phi.PN;
672 Value *ExitVal = Phi.Val;
674 // Only do the rewrite when the ExitValue can be expanded cheaply.
675 // If LoopCanBeDel is true, rewrite exit value aggressively.
676 if (ReplaceExitValue == OnlyCheapRepl && !LoopCanBeDel && Phi.HighCost) {
677 DeadInsts.push_back(ExitVal);
683 Instruction *Inst = cast<Instruction>(PN->getIncomingValue(Phi.Ith));
684 PN->setIncomingValue(Phi.Ith, ExitVal);
686 // If this instruction is dead now, delete it. Don't do it now to avoid
687 // invalidating iterators.
688 if (isInstructionTriviallyDead(Inst, TLI))
689 DeadInsts.push_back(Inst);
691 // If we determined that this PHI is safe to replace even if an LCSSA
694 PN->replaceAllUsesWith(ExitVal);
695 PN->eraseFromParent();
699 // The insertion point instruction may have been deleted; clear it out
700 // so that the rewriter doesn't trip over it later.
701 Rewriter.clearInsertPoint();
704 /// CanLoopBeDeleted - Check whether it is possible to delete the loop after
705 /// rewriting exit value. If it is possible, ignore ReplaceExitValue and
706 /// do rewriting aggressively.
707 bool IndVarSimplify::CanLoopBeDeleted(
708 Loop *L, SmallVector<RewritePhi, 8> &RewritePhiSet) {
710 BasicBlock *Preheader = L->getLoopPreheader();
711 // If there is no preheader, the loop will not be deleted.
715 // In LoopDeletion pass Loop can be deleted when ExitingBlocks.size() > 1.
716 // We obviate multiple ExitingBlocks case for simplicity.
717 // TODO: If we see testcase with multiple ExitingBlocks can be deleted
718 // after exit value rewriting, we can enhance the logic here.
719 SmallVector<BasicBlock *, 4> ExitingBlocks;
720 L->getExitingBlocks(ExitingBlocks);
721 SmallVector<BasicBlock *, 8> ExitBlocks;
722 L->getUniqueExitBlocks(ExitBlocks);
723 if (ExitBlocks.size() > 1 || ExitingBlocks.size() > 1)
726 BasicBlock *ExitBlock = ExitBlocks[0];
727 BasicBlock::iterator BI = ExitBlock->begin();
728 while (PHINode *P = dyn_cast<PHINode>(BI)) {
729 Value *Incoming = P->getIncomingValueForBlock(ExitingBlocks[0]);
731 // If the Incoming value of P is found in RewritePhiSet, we know it
732 // could be rewritten to use a loop invariant value in transformation
733 // phase later. Skip it in the loop invariant check below.
735 for (const RewritePhi &Phi : RewritePhiSet) {
736 unsigned i = Phi.Ith;
737 if (Phi.PN == P && (Phi.PN)->getIncomingValue(i) == Incoming) {
744 if (!found && (I = dyn_cast<Instruction>(Incoming)))
745 if (!L->hasLoopInvariantOperands(I))
751 for (Loop::block_iterator LI = L->block_begin(), LE = L->block_end();
753 for (BasicBlock::iterator BI = (*LI)->begin(), BE = (*LI)->end(); BI != BE;
755 if (BI->mayHaveSideEffects())
763 //===----------------------------------------------------------------------===//
764 // IV Widening - Extend the width of an IV to cover its widest uses.
765 //===----------------------------------------------------------------------===//
768 // Collect information about induction variables that are used by sign/zero
769 // extend operations. This information is recorded by CollectExtend and
770 // provides the input to WidenIV.
773 Type *WidestNativeType; // Widest integer type created [sz]ext
774 bool IsSigned; // Was a sext user seen before a zext?
776 WideIVInfo() : NarrowIV(nullptr), WidestNativeType(nullptr),
781 /// visitCast - Update information about the induction variable that is
782 /// extended by this sign or zero extend operation. This is used to determine
783 /// the final width of the IV before actually widening it.
784 static void visitIVCast(CastInst *Cast, WideIVInfo &WI, ScalarEvolution *SE,
785 const TargetTransformInfo *TTI) {
786 bool IsSigned = Cast->getOpcode() == Instruction::SExt;
787 if (!IsSigned && Cast->getOpcode() != Instruction::ZExt)
790 Type *Ty = Cast->getType();
791 uint64_t Width = SE->getTypeSizeInBits(Ty);
792 if (!Cast->getModule()->getDataLayout().isLegalInteger(Width))
795 // Cast is either an sext or zext up to this point.
796 // We should not widen an indvar if arithmetics on the wider indvar are more
797 // expensive than those on the narrower indvar. We check only the cost of ADD
798 // because at least an ADD is required to increment the induction variable. We
799 // could compute more comprehensively the cost of all instructions on the
800 // induction variable when necessary.
802 TTI->getArithmeticInstrCost(Instruction::Add, Ty) >
803 TTI->getArithmeticInstrCost(Instruction::Add,
804 Cast->getOperand(0)->getType())) {
808 if (!WI.WidestNativeType) {
809 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
810 WI.IsSigned = IsSigned;
814 // We extend the IV to satisfy the sign of its first user, arbitrarily.
815 if (WI.IsSigned != IsSigned)
818 if (Width > SE->getTypeSizeInBits(WI.WidestNativeType))
819 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
824 /// NarrowIVDefUse - Record a link in the Narrow IV def-use chain along with the
825 /// WideIV that computes the same value as the Narrow IV def. This avoids
826 /// caching Use* pointers.
827 struct NarrowIVDefUse {
828 Instruction *NarrowDef;
829 Instruction *NarrowUse;
830 Instruction *WideDef;
832 NarrowIVDefUse(): NarrowDef(nullptr), NarrowUse(nullptr), WideDef(nullptr) {}
834 NarrowIVDefUse(Instruction *ND, Instruction *NU, Instruction *WD):
835 NarrowDef(ND), NarrowUse(NU), WideDef(WD) {}
838 /// WidenIV - The goal of this transform is to remove sign and zero extends
839 /// without creating any new induction variables. To do this, it creates a new
840 /// phi of the wider type and redirects all users, either removing extends or
841 /// inserting truncs whenever we stop propagating the type.
857 Instruction *WideInc;
858 const SCEV *WideIncExpr;
859 SmallVectorImpl<WeakVH> &DeadInsts;
861 SmallPtrSet<Instruction*,16> Widened;
862 SmallVector<NarrowIVDefUse, 8> NarrowIVUsers;
865 WidenIV(const WideIVInfo &WI, LoopInfo *LInfo,
866 ScalarEvolution *SEv, DominatorTree *DTree,
867 SmallVectorImpl<WeakVH> &DI) :
868 OrigPhi(WI.NarrowIV),
869 WideType(WI.WidestNativeType),
870 IsSigned(WI.IsSigned),
872 L(LI->getLoopFor(OrigPhi->getParent())),
877 WideIncExpr(nullptr),
879 assert(L->getHeader() == OrigPhi->getParent() && "Phi must be an IV");
882 PHINode *CreateWideIV(SCEVExpander &Rewriter);
885 Value *getExtend(Value *NarrowOper, Type *WideType, bool IsSigned,
888 Instruction *CloneIVUser(NarrowIVDefUse DU);
890 const SCEVAddRecExpr *GetWideRecurrence(Instruction *NarrowUse);
892 const SCEVAddRecExpr* GetExtendedOperandRecurrence(NarrowIVDefUse DU);
894 const SCEV *GetSCEVByOpCode(const SCEV *LHS, const SCEV *RHS,
895 unsigned OpCode) const;
897 Instruction *WidenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter);
899 bool WidenLoopCompare(NarrowIVDefUse DU);
901 void pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef);
903 } // anonymous namespace
905 /// isLoopInvariant - Perform a quick domtree based check for loop invariance
906 /// assuming that V is used within the loop. LoopInfo::isLoopInvariant() seems
907 /// gratuitous for this purpose.
908 static bool isLoopInvariant(Value *V, const Loop *L, const DominatorTree *DT) {
909 Instruction *Inst = dyn_cast<Instruction>(V);
913 return DT->properlyDominates(Inst->getParent(), L->getHeader());
916 Value *WidenIV::getExtend(Value *NarrowOper, Type *WideType, bool IsSigned,
918 // Set the debug location and conservative insertion point.
919 IRBuilder<> Builder(Use);
920 // Hoist the insertion point into loop preheaders as far as possible.
921 for (const Loop *L = LI->getLoopFor(Use->getParent());
922 L && L->getLoopPreheader() && isLoopInvariant(NarrowOper, L, DT);
923 L = L->getParentLoop())
924 Builder.SetInsertPoint(L->getLoopPreheader()->getTerminator());
926 return IsSigned ? Builder.CreateSExt(NarrowOper, WideType) :
927 Builder.CreateZExt(NarrowOper, WideType);
930 /// CloneIVUser - Instantiate a wide operation to replace a narrow
931 /// operation. This only needs to handle operations that can evaluation to
932 /// SCEVAddRec. It can safely return 0 for any operation we decide not to clone.
933 Instruction *WidenIV::CloneIVUser(NarrowIVDefUse DU) {
934 unsigned Opcode = DU.NarrowUse->getOpcode();
938 case Instruction::Add:
939 case Instruction::Mul:
940 case Instruction::UDiv:
941 case Instruction::Sub:
942 case Instruction::And:
943 case Instruction::Or:
944 case Instruction::Xor:
945 case Instruction::Shl:
946 case Instruction::LShr:
947 case Instruction::AShr:
948 DEBUG(dbgs() << "Cloning IVUser: " << *DU.NarrowUse << "\n");
950 // Replace NarrowDef operands with WideDef. Otherwise, we don't know
951 // anything about the narrow operand yet so must insert a [sz]ext. It is
952 // probably loop invariant and will be folded or hoisted. If it actually
953 // comes from a widened IV, it should be removed during a future call to
955 Value *LHS = (DU.NarrowUse->getOperand(0) == DU.NarrowDef) ? DU.WideDef :
956 getExtend(DU.NarrowUse->getOperand(0), WideType, IsSigned, DU.NarrowUse);
957 Value *RHS = (DU.NarrowUse->getOperand(1) == DU.NarrowDef) ? DU.WideDef :
958 getExtend(DU.NarrowUse->getOperand(1), WideType, IsSigned, DU.NarrowUse);
960 BinaryOperator *NarrowBO = cast<BinaryOperator>(DU.NarrowUse);
961 BinaryOperator *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(),
963 NarrowBO->getName());
964 IRBuilder<> Builder(DU.NarrowUse);
965 Builder.Insert(WideBO);
966 if (const OverflowingBinaryOperator *OBO =
967 dyn_cast<OverflowingBinaryOperator>(NarrowBO)) {
968 if (OBO->hasNoUnsignedWrap()) WideBO->setHasNoUnsignedWrap();
969 if (OBO->hasNoSignedWrap()) WideBO->setHasNoSignedWrap();
975 const SCEV *WidenIV::GetSCEVByOpCode(const SCEV *LHS, const SCEV *RHS,
976 unsigned OpCode) const {
977 if (OpCode == Instruction::Add)
978 return SE->getAddExpr(LHS, RHS);
979 if (OpCode == Instruction::Sub)
980 return SE->getMinusSCEV(LHS, RHS);
981 if (OpCode == Instruction::Mul)
982 return SE->getMulExpr(LHS, RHS);
984 llvm_unreachable("Unsupported opcode.");
987 /// No-wrap operations can transfer sign extension of their result to their
988 /// operands. Generate the SCEV value for the widened operation without
989 /// actually modifying the IR yet. If the expression after extending the
990 /// operands is an AddRec for this loop, return it.
991 const SCEVAddRecExpr* WidenIV::GetExtendedOperandRecurrence(NarrowIVDefUse DU) {
993 // Handle the common case of add<nsw/nuw>
994 const unsigned OpCode = DU.NarrowUse->getOpcode();
995 // Only Add/Sub/Mul instructions supported yet.
996 if (OpCode != Instruction::Add && OpCode != Instruction::Sub &&
997 OpCode != Instruction::Mul)
1000 // One operand (NarrowDef) has already been extended to WideDef. Now determine
1001 // if extending the other will lead to a recurrence.
1002 const unsigned ExtendOperIdx =
1003 DU.NarrowUse->getOperand(0) == DU.NarrowDef ? 1 : 0;
1004 assert(DU.NarrowUse->getOperand(1-ExtendOperIdx) == DU.NarrowDef && "bad DU");
1006 const SCEV *ExtendOperExpr = nullptr;
1007 const OverflowingBinaryOperator *OBO =
1008 cast<OverflowingBinaryOperator>(DU.NarrowUse);
1009 if (IsSigned && OBO->hasNoSignedWrap())
1010 ExtendOperExpr = SE->getSignExtendExpr(
1011 SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType);
1012 else if(!IsSigned && OBO->hasNoUnsignedWrap())
1013 ExtendOperExpr = SE->getZeroExtendExpr(
1014 SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType);
1018 // When creating this SCEV expr, don't apply the current operations NSW or NUW
1019 // flags. This instruction may be guarded by control flow that the no-wrap
1020 // behavior depends on. Non-control-equivalent instructions can be mapped to
1021 // the same SCEV expression, and it would be incorrect to transfer NSW/NUW
1022 // semantics to those operations.
1023 const SCEV *lhs = SE->getSCEV(DU.WideDef);
1024 const SCEV *rhs = ExtendOperExpr;
1026 // Let's swap operands to the initial order for the case of non-commutative
1027 // operations, like SUB. See PR21014.
1028 if (ExtendOperIdx == 0)
1029 std::swap(lhs, rhs);
1030 const SCEVAddRecExpr *AddRec =
1031 dyn_cast<SCEVAddRecExpr>(GetSCEVByOpCode(lhs, rhs, OpCode));
1033 if (!AddRec || AddRec->getLoop() != L)
1038 /// GetWideRecurrence - Is this instruction potentially interesting for further
1039 /// simplification after widening it's type? In other words, can the
1040 /// extend be safely hoisted out of the loop with SCEV reducing the value to a
1041 /// recurrence on the same loop. If so, return the sign or zero extended
1042 /// recurrence. Otherwise return NULL.
1043 const SCEVAddRecExpr *WidenIV::GetWideRecurrence(Instruction *NarrowUse) {
1044 if (!SE->isSCEVable(NarrowUse->getType()))
1047 const SCEV *NarrowExpr = SE->getSCEV(NarrowUse);
1048 if (SE->getTypeSizeInBits(NarrowExpr->getType())
1049 >= SE->getTypeSizeInBits(WideType)) {
1050 // NarrowUse implicitly widens its operand. e.g. a gep with a narrow
1051 // index. So don't follow this use.
1055 const SCEV *WideExpr = IsSigned ?
1056 SE->getSignExtendExpr(NarrowExpr, WideType) :
1057 SE->getZeroExtendExpr(NarrowExpr, WideType);
1058 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(WideExpr);
1059 if (!AddRec || AddRec->getLoop() != L)
1064 /// This IV user cannot be widen. Replace this use of the original narrow IV
1065 /// with a truncation of the new wide IV to isolate and eliminate the narrow IV.
1066 static void truncateIVUse(NarrowIVDefUse DU, DominatorTree *DT) {
1067 DEBUG(dbgs() << "INDVARS: Truncate IV " << *DU.WideDef
1068 << " for user " << *DU.NarrowUse << "\n");
1069 IRBuilder<> Builder(getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT));
1070 Value *Trunc = Builder.CreateTrunc(DU.WideDef, DU.NarrowDef->getType());
1071 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, Trunc);
1074 /// If the narrow use is a compare instruction, then widen the compare
1075 // (and possibly the other operand). The extend operation is hoisted into the
1076 // loop preheader as far as possible.
1077 bool WidenIV::WidenLoopCompare(NarrowIVDefUse DU) {
1078 ICmpInst *Cmp = dyn_cast<ICmpInst>(DU.NarrowUse);
1082 // Sign of IV user and compare must match.
1083 if (IsSigned != CmpInst::isSigned(Cmp->getPredicate()))
1086 Value *Op = Cmp->getOperand(Cmp->getOperand(0) == DU.NarrowDef ? 1 : 0);
1087 unsigned CastWidth = SE->getTypeSizeInBits(Op->getType());
1088 unsigned IVWidth = SE->getTypeSizeInBits(WideType);
1089 assert (CastWidth <= IVWidth && "Unexpected width while widening compare.");
1091 // Widen the compare instruction.
1092 IRBuilder<> Builder(getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT));
1093 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef);
1095 // Widen the other operand of the compare, if necessary.
1096 if (CastWidth < IVWidth) {
1097 Value *ExtOp = getExtend(Op, WideType, IsSigned, Cmp);
1098 DU.NarrowUse->replaceUsesOfWith(Op, ExtOp);
1103 /// WidenIVUse - Determine whether an individual user of the narrow IV can be
1104 /// widened. If so, return the wide clone of the user.
1105 Instruction *WidenIV::WidenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter) {
1107 // Stop traversing the def-use chain at inner-loop phis or post-loop phis.
1108 if (PHINode *UsePhi = dyn_cast<PHINode>(DU.NarrowUse)) {
1109 if (LI->getLoopFor(UsePhi->getParent()) != L) {
1110 // For LCSSA phis, sink the truncate outside the loop.
1111 // After SimplifyCFG most loop exit targets have a single predecessor.
1112 // Otherwise fall back to a truncate within the loop.
1113 if (UsePhi->getNumOperands() != 1)
1114 truncateIVUse(DU, DT);
1117 PHINode::Create(DU.WideDef->getType(), 1, UsePhi->getName() + ".wide",
1119 WidePhi->addIncoming(DU.WideDef, UsePhi->getIncomingBlock(0));
1120 IRBuilder<> Builder(WidePhi->getParent()->getFirstInsertionPt());
1121 Value *Trunc = Builder.CreateTrunc(WidePhi, DU.NarrowDef->getType());
1122 UsePhi->replaceAllUsesWith(Trunc);
1123 DeadInsts.emplace_back(UsePhi);
1124 DEBUG(dbgs() << "INDVARS: Widen lcssa phi " << *UsePhi
1125 << " to " << *WidePhi << "\n");
1130 // Our raison d'etre! Eliminate sign and zero extension.
1131 if (IsSigned ? isa<SExtInst>(DU.NarrowUse) : isa<ZExtInst>(DU.NarrowUse)) {
1132 Value *NewDef = DU.WideDef;
1133 if (DU.NarrowUse->getType() != WideType) {
1134 unsigned CastWidth = SE->getTypeSizeInBits(DU.NarrowUse->getType());
1135 unsigned IVWidth = SE->getTypeSizeInBits(WideType);
1136 if (CastWidth < IVWidth) {
1137 // The cast isn't as wide as the IV, so insert a Trunc.
1138 IRBuilder<> Builder(DU.NarrowUse);
1139 NewDef = Builder.CreateTrunc(DU.WideDef, DU.NarrowUse->getType());
1142 // A wider extend was hidden behind a narrower one. This may induce
1143 // another round of IV widening in which the intermediate IV becomes
1144 // dead. It should be very rare.
1145 DEBUG(dbgs() << "INDVARS: New IV " << *WidePhi
1146 << " not wide enough to subsume " << *DU.NarrowUse << "\n");
1147 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef);
1148 NewDef = DU.NarrowUse;
1151 if (NewDef != DU.NarrowUse) {
1152 DEBUG(dbgs() << "INDVARS: eliminating " << *DU.NarrowUse
1153 << " replaced by " << *DU.WideDef << "\n");
1155 DU.NarrowUse->replaceAllUsesWith(NewDef);
1156 DeadInsts.emplace_back(DU.NarrowUse);
1158 // Now that the extend is gone, we want to expose it's uses for potential
1159 // further simplification. We don't need to directly inform SimplifyIVUsers
1160 // of the new users, because their parent IV will be processed later as a
1161 // new loop phi. If we preserved IVUsers analysis, we would also want to
1162 // push the uses of WideDef here.
1164 // No further widening is needed. The deceased [sz]ext had done it for us.
1168 // Does this user itself evaluate to a recurrence after widening?
1169 const SCEVAddRecExpr *WideAddRec = GetWideRecurrence(DU.NarrowUse);
1171 WideAddRec = GetExtendedOperandRecurrence(DU);
1174 // If use is a loop condition, try to promote the condition instead of
1175 // truncating the IV first.
1176 if (WidenLoopCompare(DU))
1179 // This user does not evaluate to a recurence after widening, so don't
1180 // follow it. Instead insert a Trunc to kill off the original use,
1181 // eventually isolating the original narrow IV so it can be removed.
1182 truncateIVUse(DU, DT);
1185 // Assume block terminators cannot evaluate to a recurrence. We can't to
1186 // insert a Trunc after a terminator if there happens to be a critical edge.
1187 assert(DU.NarrowUse != DU.NarrowUse->getParent()->getTerminator() &&
1188 "SCEV is not expected to evaluate a block terminator");
1190 // Reuse the IV increment that SCEVExpander created as long as it dominates
1192 Instruction *WideUse = nullptr;
1193 if (WideAddRec == WideIncExpr
1194 && Rewriter.hoistIVInc(WideInc, DU.NarrowUse))
1197 WideUse = CloneIVUser(DU);
1201 // Evaluation of WideAddRec ensured that the narrow expression could be
1202 // extended outside the loop without overflow. This suggests that the wide use
1203 // evaluates to the same expression as the extended narrow use, but doesn't
1204 // absolutely guarantee it. Hence the following failsafe check. In rare cases
1205 // where it fails, we simply throw away the newly created wide use.
1206 if (WideAddRec != SE->getSCEV(WideUse)) {
1207 DEBUG(dbgs() << "Wide use expression mismatch: " << *WideUse
1208 << ": " << *SE->getSCEV(WideUse) << " != " << *WideAddRec << "\n");
1209 DeadInsts.emplace_back(WideUse);
1213 // Returning WideUse pushes it on the worklist.
1217 /// pushNarrowIVUsers - Add eligible users of NarrowDef to NarrowIVUsers.
1219 void WidenIV::pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef) {
1220 for (User *U : NarrowDef->users()) {
1221 Instruction *NarrowUser = cast<Instruction>(U);
1223 // Handle data flow merges and bizarre phi cycles.
1224 if (!Widened.insert(NarrowUser).second)
1227 NarrowIVUsers.push_back(NarrowIVDefUse(NarrowDef, NarrowUser, WideDef));
1231 /// CreateWideIV - Process a single induction variable. First use the
1232 /// SCEVExpander to create a wide induction variable that evaluates to the same
1233 /// recurrence as the original narrow IV. Then use a worklist to forward
1234 /// traverse the narrow IV's def-use chain. After WidenIVUse has processed all
1235 /// interesting IV users, the narrow IV will be isolated for removal by
1238 /// It would be simpler to delete uses as they are processed, but we must avoid
1239 /// invalidating SCEV expressions.
1241 PHINode *WidenIV::CreateWideIV(SCEVExpander &Rewriter) {
1242 // Is this phi an induction variable?
1243 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(OrigPhi));
1247 // Widen the induction variable expression.
1248 const SCEV *WideIVExpr = IsSigned ?
1249 SE->getSignExtendExpr(AddRec, WideType) :
1250 SE->getZeroExtendExpr(AddRec, WideType);
1252 assert(SE->getEffectiveSCEVType(WideIVExpr->getType()) == WideType &&
1253 "Expect the new IV expression to preserve its type");
1255 // Can the IV be extended outside the loop without overflow?
1256 AddRec = dyn_cast<SCEVAddRecExpr>(WideIVExpr);
1257 if (!AddRec || AddRec->getLoop() != L)
1260 // An AddRec must have loop-invariant operands. Since this AddRec is
1261 // materialized by a loop header phi, the expression cannot have any post-loop
1262 // operands, so they must dominate the loop header.
1263 assert(SE->properlyDominates(AddRec->getStart(), L->getHeader()) &&
1264 SE->properlyDominates(AddRec->getStepRecurrence(*SE), L->getHeader())
1265 && "Loop header phi recurrence inputs do not dominate the loop");
1267 // The rewriter provides a value for the desired IV expression. This may
1268 // either find an existing phi or materialize a new one. Either way, we
1269 // expect a well-formed cyclic phi-with-increments. i.e. any operand not part
1270 // of the phi-SCC dominates the loop entry.
1271 Instruction *InsertPt = L->getHeader()->begin();
1272 WidePhi = cast<PHINode>(Rewriter.expandCodeFor(AddRec, WideType, InsertPt));
1274 // Remembering the WideIV increment generated by SCEVExpander allows
1275 // WidenIVUse to reuse it when widening the narrow IV's increment. We don't
1276 // employ a general reuse mechanism because the call above is the only call to
1277 // SCEVExpander. Henceforth, we produce 1-to-1 narrow to wide uses.
1278 if (BasicBlock *LatchBlock = L->getLoopLatch()) {
1280 cast<Instruction>(WidePhi->getIncomingValueForBlock(LatchBlock));
1281 WideIncExpr = SE->getSCEV(WideInc);
1284 DEBUG(dbgs() << "Wide IV: " << *WidePhi << "\n");
1287 // Traverse the def-use chain using a worklist starting at the original IV.
1288 assert(Widened.empty() && NarrowIVUsers.empty() && "expect initial state" );
1290 Widened.insert(OrigPhi);
1291 pushNarrowIVUsers(OrigPhi, WidePhi);
1293 while (!NarrowIVUsers.empty()) {
1294 NarrowIVDefUse DU = NarrowIVUsers.pop_back_val();
1296 // Process a def-use edge. This may replace the use, so don't hold a
1297 // use_iterator across it.
1298 Instruction *WideUse = WidenIVUse(DU, Rewriter);
1300 // Follow all def-use edges from the previous narrow use.
1302 pushNarrowIVUsers(DU.NarrowUse, WideUse);
1304 // WidenIVUse may have removed the def-use edge.
1305 if (DU.NarrowDef->use_empty())
1306 DeadInsts.emplace_back(DU.NarrowDef);
1311 //===----------------------------------------------------------------------===//
1312 // Live IV Reduction - Minimize IVs live across the loop.
1313 //===----------------------------------------------------------------------===//
1316 //===----------------------------------------------------------------------===//
1317 // Simplification of IV users based on SCEV evaluation.
1318 //===----------------------------------------------------------------------===//
1321 class IndVarSimplifyVisitor : public IVVisitor {
1322 ScalarEvolution *SE;
1323 const TargetTransformInfo *TTI;
1329 IndVarSimplifyVisitor(PHINode *IV, ScalarEvolution *SCEV,
1330 const TargetTransformInfo *TTI,
1331 const DominatorTree *DTree)
1332 : SE(SCEV), TTI(TTI), IVPhi(IV) {
1334 WI.NarrowIV = IVPhi;
1336 setSplitOverflowIntrinsics();
1339 // Implement the interface used by simplifyUsersOfIV.
1340 void visitCast(CastInst *Cast) override { visitIVCast(Cast, WI, SE, TTI); }
1344 /// SimplifyAndExtend - Iteratively perform simplification on a worklist of IV
1345 /// users. Each successive simplification may push more users which may
1346 /// themselves be candidates for simplification.
1348 /// Sign/Zero extend elimination is interleaved with IV simplification.
1350 void IndVarSimplify::SimplifyAndExtend(Loop *L,
1351 SCEVExpander &Rewriter,
1352 LPPassManager &LPM) {
1353 SmallVector<WideIVInfo, 8> WideIVs;
1355 SmallVector<PHINode*, 8> LoopPhis;
1356 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
1357 LoopPhis.push_back(cast<PHINode>(I));
1359 // Each round of simplification iterates through the SimplifyIVUsers worklist
1360 // for all current phis, then determines whether any IVs can be
1361 // widened. Widening adds new phis to LoopPhis, inducing another round of
1362 // simplification on the wide IVs.
1363 while (!LoopPhis.empty()) {
1364 // Evaluate as many IV expressions as possible before widening any IVs. This
1365 // forces SCEV to set no-wrap flags before evaluating sign/zero
1366 // extension. The first time SCEV attempts to normalize sign/zero extension,
1367 // the result becomes final. So for the most predictable results, we delay
1368 // evaluation of sign/zero extend evaluation until needed, and avoid running
1369 // other SCEV based analysis prior to SimplifyAndExtend.
1371 PHINode *CurrIV = LoopPhis.pop_back_val();
1373 // Information about sign/zero extensions of CurrIV.
1374 IndVarSimplifyVisitor Visitor(CurrIV, SE, TTI, DT);
1376 Changed |= simplifyUsersOfIV(CurrIV, SE, &LPM, DeadInsts, &Visitor);
1378 if (Visitor.WI.WidestNativeType) {
1379 WideIVs.push_back(Visitor.WI);
1381 } while(!LoopPhis.empty());
1383 for (; !WideIVs.empty(); WideIVs.pop_back()) {
1384 WidenIV Widener(WideIVs.back(), LI, SE, DT, DeadInsts);
1385 if (PHINode *WidePhi = Widener.CreateWideIV(Rewriter)) {
1387 LoopPhis.push_back(WidePhi);
1393 //===----------------------------------------------------------------------===//
1394 // LinearFunctionTestReplace and its kin. Rewrite the loop exit condition.
1395 //===----------------------------------------------------------------------===//
1397 /// canExpandBackedgeTakenCount - Return true if this loop's backedge taken
1398 /// count expression can be safely and cheaply expanded into an instruction
1399 /// sequence that can be used by LinearFunctionTestReplace.
1401 /// TODO: This fails for pointer-type loop counters with greater than one byte
1402 /// strides, consequently preventing LFTR from running. For the purpose of LFTR
1403 /// we could skip this check in the case that the LFTR loop counter (chosen by
1404 /// FindLoopCounter) is also pointer type. Instead, we could directly convert
1405 /// the loop test to an inequality test by checking the target data's alignment
1406 /// of element types (given that the initial pointer value originates from or is
1407 /// used by ABI constrained operation, as opposed to inttoptr/ptrtoint).
1408 /// However, we don't yet have a strong motivation for converting loop tests
1409 /// into inequality tests.
1410 static bool canExpandBackedgeTakenCount(Loop *L, ScalarEvolution *SE,
1411 SCEVExpander &Rewriter) {
1412 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
1413 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount) ||
1414 BackedgeTakenCount->isZero())
1417 if (!L->getExitingBlock())
1420 // Can't rewrite non-branch yet.
1421 if (!isa<BranchInst>(L->getExitingBlock()->getTerminator()))
1424 if (Rewriter.isHighCostExpansion(BackedgeTakenCount, L))
1430 /// getLoopPhiForCounter - Return the loop header phi IFF IncV adds a loop
1431 /// invariant value to the phi.
1432 static PHINode *getLoopPhiForCounter(Value *IncV, Loop *L, DominatorTree *DT) {
1433 Instruction *IncI = dyn_cast<Instruction>(IncV);
1437 switch (IncI->getOpcode()) {
1438 case Instruction::Add:
1439 case Instruction::Sub:
1441 case Instruction::GetElementPtr:
1442 // An IV counter must preserve its type.
1443 if (IncI->getNumOperands() == 2)
1449 PHINode *Phi = dyn_cast<PHINode>(IncI->getOperand(0));
1450 if (Phi && Phi->getParent() == L->getHeader()) {
1451 if (isLoopInvariant(IncI->getOperand(1), L, DT))
1455 if (IncI->getOpcode() == Instruction::GetElementPtr)
1458 // Allow add/sub to be commuted.
1459 Phi = dyn_cast<PHINode>(IncI->getOperand(1));
1460 if (Phi && Phi->getParent() == L->getHeader()) {
1461 if (isLoopInvariant(IncI->getOperand(0), L, DT))
1467 /// Return the compare guarding the loop latch, or NULL for unrecognized tests.
1468 static ICmpInst *getLoopTest(Loop *L) {
1469 assert(L->getExitingBlock() && "expected loop exit");
1471 BasicBlock *LatchBlock = L->getLoopLatch();
1472 // Don't bother with LFTR if the loop is not properly simplified.
1476 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
1477 assert(BI && "expected exit branch");
1479 return dyn_cast<ICmpInst>(BI->getCondition());
1482 /// needsLFTR - LinearFunctionTestReplace policy. Return true unless we can show
1483 /// that the current exit test is already sufficiently canonical.
1484 static bool needsLFTR(Loop *L, DominatorTree *DT) {
1485 // Do LFTR to simplify the exit condition to an ICMP.
1486 ICmpInst *Cond = getLoopTest(L);
1490 // Do LFTR to simplify the exit ICMP to EQ/NE
1491 ICmpInst::Predicate Pred = Cond->getPredicate();
1492 if (Pred != ICmpInst::ICMP_NE && Pred != ICmpInst::ICMP_EQ)
1495 // Look for a loop invariant RHS
1496 Value *LHS = Cond->getOperand(0);
1497 Value *RHS = Cond->getOperand(1);
1498 if (!isLoopInvariant(RHS, L, DT)) {
1499 if (!isLoopInvariant(LHS, L, DT))
1501 std::swap(LHS, RHS);
1503 // Look for a simple IV counter LHS
1504 PHINode *Phi = dyn_cast<PHINode>(LHS);
1506 Phi = getLoopPhiForCounter(LHS, L, DT);
1511 // Do LFTR if PHI node is defined in the loop, but is *not* a counter.
1512 int Idx = Phi->getBasicBlockIndex(L->getLoopLatch());
1516 // Do LFTR if the exit condition's IV is *not* a simple counter.
1517 Value *IncV = Phi->getIncomingValue(Idx);
1518 return Phi != getLoopPhiForCounter(IncV, L, DT);
1521 /// Recursive helper for hasConcreteDef(). Unfortunately, this currently boils
1522 /// down to checking that all operands are constant and listing instructions
1523 /// that may hide undef.
1524 static bool hasConcreteDefImpl(Value *V, SmallPtrSetImpl<Value*> &Visited,
1526 if (isa<Constant>(V))
1527 return !isa<UndefValue>(V);
1532 // Conservatively handle non-constant non-instructions. For example, Arguments
1534 Instruction *I = dyn_cast<Instruction>(V);
1538 // Load and return values may be undef.
1539 if(I->mayReadFromMemory() || isa<CallInst>(I) || isa<InvokeInst>(I))
1542 // Optimistically handle other instructions.
1543 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) {
1544 if (!Visited.insert(*OI).second)
1546 if (!hasConcreteDefImpl(*OI, Visited, Depth+1))
1552 /// Return true if the given value is concrete. We must prove that undef can
1555 /// TODO: If we decide that this is a good approach to checking for undef, we
1556 /// may factor it into a common location.
1557 static bool hasConcreteDef(Value *V) {
1558 SmallPtrSet<Value*, 8> Visited;
1560 return hasConcreteDefImpl(V, Visited, 0);
1563 /// AlmostDeadIV - Return true if this IV has any uses other than the (soon to
1564 /// be rewritten) loop exit test.
1565 static bool AlmostDeadIV(PHINode *Phi, BasicBlock *LatchBlock, Value *Cond) {
1566 int LatchIdx = Phi->getBasicBlockIndex(LatchBlock);
1567 Value *IncV = Phi->getIncomingValue(LatchIdx);
1569 for (User *U : Phi->users())
1570 if (U != Cond && U != IncV) return false;
1572 for (User *U : IncV->users())
1573 if (U != Cond && U != Phi) return false;
1577 /// FindLoopCounter - Find an affine IV in canonical form.
1579 /// BECount may be an i8* pointer type. The pointer difference is already
1580 /// valid count without scaling the address stride, so it remains a pointer
1581 /// expression as far as SCEV is concerned.
1583 /// Currently only valid for LFTR. See the comments on hasConcreteDef below.
1585 /// FIXME: Accept -1 stride and set IVLimit = IVInit - BECount
1587 /// FIXME: Accept non-unit stride as long as SCEV can reduce BECount * Stride.
1588 /// This is difficult in general for SCEV because of potential overflow. But we
1589 /// could at least handle constant BECounts.
1590 static PHINode *FindLoopCounter(Loop *L, const SCEV *BECount,
1591 ScalarEvolution *SE, DominatorTree *DT) {
1592 uint64_t BCWidth = SE->getTypeSizeInBits(BECount->getType());
1595 cast<BranchInst>(L->getExitingBlock()->getTerminator())->getCondition();
1597 // Loop over all of the PHI nodes, looking for a simple counter.
1598 PHINode *BestPhi = nullptr;
1599 const SCEV *BestInit = nullptr;
1600 BasicBlock *LatchBlock = L->getLoopLatch();
1601 assert(LatchBlock && "needsLFTR should guarantee a loop latch");
1603 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
1604 PHINode *Phi = cast<PHINode>(I);
1605 if (!SE->isSCEVable(Phi->getType()))
1608 // Avoid comparing an integer IV against a pointer Limit.
1609 if (BECount->getType()->isPointerTy() && !Phi->getType()->isPointerTy())
1612 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Phi));
1613 if (!AR || AR->getLoop() != L || !AR->isAffine())
1616 // AR may be a pointer type, while BECount is an integer type.
1617 // AR may be wider than BECount. With eq/ne tests overflow is immaterial.
1618 // AR may not be a narrower type, or we may never exit.
1619 uint64_t PhiWidth = SE->getTypeSizeInBits(AR->getType());
1620 if (PhiWidth < BCWidth ||
1621 !L->getHeader()->getModule()->getDataLayout().isLegalInteger(PhiWidth))
1624 const SCEV *Step = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE));
1625 if (!Step || !Step->isOne())
1628 int LatchIdx = Phi->getBasicBlockIndex(LatchBlock);
1629 Value *IncV = Phi->getIncomingValue(LatchIdx);
1630 if (getLoopPhiForCounter(IncV, L, DT) != Phi)
1633 // Avoid reusing a potentially undef value to compute other values that may
1634 // have originally had a concrete definition.
1635 if (!hasConcreteDef(Phi)) {
1636 // We explicitly allow unknown phis as long as they are already used by
1637 // the loop test. In this case we assume that performing LFTR could not
1638 // increase the number of undef users.
1639 if (ICmpInst *Cond = getLoopTest(L)) {
1640 if (Phi != getLoopPhiForCounter(Cond->getOperand(0), L, DT)
1641 && Phi != getLoopPhiForCounter(Cond->getOperand(1), L, DT)) {
1646 const SCEV *Init = AR->getStart();
1648 if (BestPhi && !AlmostDeadIV(BestPhi, LatchBlock, Cond)) {
1649 // Don't force a live loop counter if another IV can be used.
1650 if (AlmostDeadIV(Phi, LatchBlock, Cond))
1653 // Prefer to count-from-zero. This is a more "canonical" counter form. It
1654 // also prefers integer to pointer IVs.
1655 if (BestInit->isZero() != Init->isZero()) {
1656 if (BestInit->isZero())
1659 // If two IVs both count from zero or both count from nonzero then the
1660 // narrower is likely a dead phi that has been widened. Use the wider phi
1661 // to allow the other to be eliminated.
1662 else if (PhiWidth <= SE->getTypeSizeInBits(BestPhi->getType()))
1671 /// genLoopLimit - Help LinearFunctionTestReplace by generating a value that
1672 /// holds the RHS of the new loop test.
1673 static Value *genLoopLimit(PHINode *IndVar, const SCEV *IVCount, Loop *L,
1674 SCEVExpander &Rewriter, ScalarEvolution *SE) {
1675 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(IndVar));
1676 assert(AR && AR->getLoop() == L && AR->isAffine() && "bad loop counter");
1677 const SCEV *IVInit = AR->getStart();
1679 // IVInit may be a pointer while IVCount is an integer when FindLoopCounter
1680 // finds a valid pointer IV. Sign extend BECount in order to materialize a
1681 // GEP. Avoid running SCEVExpander on a new pointer value, instead reusing
1682 // the existing GEPs whenever possible.
1683 if (IndVar->getType()->isPointerTy()
1684 && !IVCount->getType()->isPointerTy()) {
1686 // IVOffset will be the new GEP offset that is interpreted by GEP as a
1687 // signed value. IVCount on the other hand represents the loop trip count,
1688 // which is an unsigned value. FindLoopCounter only allows induction
1689 // variables that have a positive unit stride of one. This means we don't
1690 // have to handle the case of negative offsets (yet) and just need to zero
1692 Type *OfsTy = SE->getEffectiveSCEVType(IVInit->getType());
1693 const SCEV *IVOffset = SE->getTruncateOrZeroExtend(IVCount, OfsTy);
1695 // Expand the code for the iteration count.
1696 assert(SE->isLoopInvariant(IVOffset, L) &&
1697 "Computed iteration count is not loop invariant!");
1698 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1699 Value *GEPOffset = Rewriter.expandCodeFor(IVOffset, OfsTy, BI);
1701 Value *GEPBase = IndVar->getIncomingValueForBlock(L->getLoopPreheader());
1702 assert(AR->getStart() == SE->getSCEV(GEPBase) && "bad loop counter");
1703 // We could handle pointer IVs other than i8*, but we need to compensate for
1704 // gep index scaling. See canExpandBackedgeTakenCount comments.
1705 assert(SE->getSizeOfExpr(IntegerType::getInt64Ty(IndVar->getContext()),
1706 cast<PointerType>(GEPBase->getType())->getElementType())->isOne()
1707 && "unit stride pointer IV must be i8*");
1709 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
1710 return Builder.CreateGEP(nullptr, GEPBase, GEPOffset, "lftr.limit");
1713 // In any other case, convert both IVInit and IVCount to integers before
1714 // comparing. This may result in SCEV expension of pointers, but in practice
1715 // SCEV will fold the pointer arithmetic away as such:
1716 // BECount = (IVEnd - IVInit - 1) => IVLimit = IVInit (postinc).
1718 // Valid Cases: (1) both integers is most common; (2) both may be pointers
1719 // for simple memset-style loops.
1721 // IVInit integer and IVCount pointer would only occur if a canonical IV
1722 // were generated on top of case #2, which is not expected.
1724 const SCEV *IVLimit = nullptr;
1725 // For unit stride, IVCount = Start + BECount with 2's complement overflow.
1726 // For non-zero Start, compute IVCount here.
1727 if (AR->getStart()->isZero())
1730 assert(AR->getStepRecurrence(*SE)->isOne() && "only handles unit stride");
1731 const SCEV *IVInit = AR->getStart();
1733 // For integer IVs, truncate the IV before computing IVInit + BECount.
1734 if (SE->getTypeSizeInBits(IVInit->getType())
1735 > SE->getTypeSizeInBits(IVCount->getType()))
1736 IVInit = SE->getTruncateExpr(IVInit, IVCount->getType());
1738 IVLimit = SE->getAddExpr(IVInit, IVCount);
1740 // Expand the code for the iteration count.
1741 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1742 IRBuilder<> Builder(BI);
1743 assert(SE->isLoopInvariant(IVLimit, L) &&
1744 "Computed iteration count is not loop invariant!");
1745 // Ensure that we generate the same type as IndVar, or a smaller integer
1746 // type. In the presence of null pointer values, we have an integer type
1747 // SCEV expression (IVInit) for a pointer type IV value (IndVar).
1748 Type *LimitTy = IVCount->getType()->isPointerTy() ?
1749 IndVar->getType() : IVCount->getType();
1750 return Rewriter.expandCodeFor(IVLimit, LimitTy, BI);
1754 /// LinearFunctionTestReplace - This method rewrites the exit condition of the
1755 /// loop to be a canonical != comparison against the incremented loop induction
1756 /// variable. This pass is able to rewrite the exit tests of any loop where the
1757 /// SCEV analysis can determine a loop-invariant trip count of the loop, which
1758 /// is actually a much broader range than just linear tests.
1759 Value *IndVarSimplify::
1760 LinearFunctionTestReplace(Loop *L,
1761 const SCEV *BackedgeTakenCount,
1763 SCEVExpander &Rewriter) {
1764 assert(canExpandBackedgeTakenCount(L, SE, Rewriter) && "precondition");
1766 // Initialize CmpIndVar and IVCount to their preincremented values.
1767 Value *CmpIndVar = IndVar;
1768 const SCEV *IVCount = BackedgeTakenCount;
1770 // If the exiting block is the same as the backedge block, we prefer to
1771 // compare against the post-incremented value, otherwise we must compare
1772 // against the preincremented value.
1773 if (L->getExitingBlock() == L->getLoopLatch()) {
1774 // Add one to the "backedge-taken" count to get the trip count.
1775 // This addition may overflow, which is valid as long as the comparison is
1776 // truncated to BackedgeTakenCount->getType().
1777 IVCount = SE->getAddExpr(BackedgeTakenCount,
1778 SE->getConstant(BackedgeTakenCount->getType(), 1));
1779 // The BackedgeTaken expression contains the number of times that the
1780 // backedge branches to the loop header. This is one less than the
1781 // number of times the loop executes, so use the incremented indvar.
1782 CmpIndVar = IndVar->getIncomingValueForBlock(L->getExitingBlock());
1785 Value *ExitCnt = genLoopLimit(IndVar, IVCount, L, Rewriter, SE);
1786 assert(ExitCnt->getType()->isPointerTy() == IndVar->getType()->isPointerTy()
1787 && "genLoopLimit missed a cast");
1789 // Insert a new icmp_ne or icmp_eq instruction before the branch.
1790 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1791 ICmpInst::Predicate P;
1792 if (L->contains(BI->getSuccessor(0)))
1793 P = ICmpInst::ICMP_NE;
1795 P = ICmpInst::ICMP_EQ;
1797 DEBUG(dbgs() << "INDVARS: Rewriting loop exit condition to:\n"
1798 << " LHS:" << *CmpIndVar << '\n'
1800 << (P == ICmpInst::ICMP_NE ? "!=" : "==") << "\n"
1801 << " RHS:\t" << *ExitCnt << "\n"
1802 << " IVCount:\t" << *IVCount << "\n");
1804 IRBuilder<> Builder(BI);
1806 // LFTR can ignore IV overflow and truncate to the width of
1807 // BECount. This avoids materializing the add(zext(add)) expression.
1808 unsigned CmpIndVarSize = SE->getTypeSizeInBits(CmpIndVar->getType());
1809 unsigned ExitCntSize = SE->getTypeSizeInBits(ExitCnt->getType());
1810 if (CmpIndVarSize > ExitCntSize) {
1811 const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(SE->getSCEV(IndVar));
1812 const SCEV *ARStart = AR->getStart();
1813 const SCEV *ARStep = AR->getStepRecurrence(*SE);
1814 // For constant IVCount, avoid truncation.
1815 if (isa<SCEVConstant>(ARStart) && isa<SCEVConstant>(IVCount)) {
1816 const APInt &Start = cast<SCEVConstant>(ARStart)->getValue()->getValue();
1817 APInt Count = cast<SCEVConstant>(IVCount)->getValue()->getValue();
1818 // Note that the post-inc value of BackedgeTakenCount may have overflowed
1819 // above such that IVCount is now zero.
1820 if (IVCount != BackedgeTakenCount && Count == 0) {
1821 Count = APInt::getMaxValue(Count.getBitWidth()).zext(CmpIndVarSize);
1825 Count = Count.zext(CmpIndVarSize);
1827 if (cast<SCEVConstant>(ARStep)->getValue()->isNegative())
1828 NewLimit = Start - Count;
1830 NewLimit = Start + Count;
1831 ExitCnt = ConstantInt::get(CmpIndVar->getType(), NewLimit);
1833 DEBUG(dbgs() << " Widen RHS:\t" << *ExitCnt << "\n");
1835 CmpIndVar = Builder.CreateTrunc(CmpIndVar, ExitCnt->getType(),
1839 Value *Cond = Builder.CreateICmp(P, CmpIndVar, ExitCnt, "exitcond");
1840 Value *OrigCond = BI->getCondition();
1841 // It's tempting to use replaceAllUsesWith here to fully replace the old
1842 // comparison, but that's not immediately safe, since users of the old
1843 // comparison may not be dominated by the new comparison. Instead, just
1844 // update the branch to use the new comparison; in the common case this
1845 // will make old comparison dead.
1846 BI->setCondition(Cond);
1847 DeadInsts.push_back(OrigCond);
1854 //===----------------------------------------------------------------------===//
1855 // SinkUnusedInvariants. A late subpass to cleanup loop preheaders.
1856 //===----------------------------------------------------------------------===//
1858 /// If there's a single exit block, sink any loop-invariant values that
1859 /// were defined in the preheader but not used inside the loop into the
1860 /// exit block to reduce register pressure in the loop.
1861 void IndVarSimplify::SinkUnusedInvariants(Loop *L) {
1862 BasicBlock *ExitBlock = L->getExitBlock();
1863 if (!ExitBlock) return;
1865 BasicBlock *Preheader = L->getLoopPreheader();
1866 if (!Preheader) return;
1868 Instruction *InsertPt = ExitBlock->getFirstInsertionPt();
1869 BasicBlock::iterator I = Preheader->getTerminator();
1870 while (I != Preheader->begin()) {
1872 // New instructions were inserted at the end of the preheader.
1873 if (isa<PHINode>(I))
1876 // Don't move instructions which might have side effects, since the side
1877 // effects need to complete before instructions inside the loop. Also don't
1878 // move instructions which might read memory, since the loop may modify
1879 // memory. Note that it's okay if the instruction might have undefined
1880 // behavior: LoopSimplify guarantees that the preheader dominates the exit
1882 if (I->mayHaveSideEffects() || I->mayReadFromMemory())
1885 // Skip debug info intrinsics.
1886 if (isa<DbgInfoIntrinsic>(I))
1889 // Skip eh pad instructions.
1893 // Don't sink alloca: we never want to sink static alloca's out of the
1894 // entry block, and correctly sinking dynamic alloca's requires
1895 // checks for stacksave/stackrestore intrinsics.
1896 // FIXME: Refactor this check somehow?
1897 if (isa<AllocaInst>(I))
1900 // Determine if there is a use in or before the loop (direct or
1902 bool UsedInLoop = false;
1903 for (Use &U : I->uses()) {
1904 Instruction *User = cast<Instruction>(U.getUser());
1905 BasicBlock *UseBB = User->getParent();
1906 if (PHINode *P = dyn_cast<PHINode>(User)) {
1908 PHINode::getIncomingValueNumForOperand(U.getOperandNo());
1909 UseBB = P->getIncomingBlock(i);
1911 if (UseBB == Preheader || L->contains(UseBB)) {
1917 // If there is, the def must remain in the preheader.
1921 // Otherwise, sink it to the exit block.
1922 Instruction *ToMove = I;
1925 if (I != Preheader->begin()) {
1926 // Skip debug info intrinsics.
1929 } while (isa<DbgInfoIntrinsic>(I) && I != Preheader->begin());
1931 if (isa<DbgInfoIntrinsic>(I) && I == Preheader->begin())
1937 ToMove->moveBefore(InsertPt);
1943 //===----------------------------------------------------------------------===//
1944 // IndVarSimplify driver. Manage several subpasses of IV simplification.
1945 //===----------------------------------------------------------------------===//
1947 bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) {
1948 if (skipOptnoneFunction(L))
1951 // If LoopSimplify form is not available, stay out of trouble. Some notes:
1952 // - LSR currently only supports LoopSimplify-form loops. Indvars'
1953 // canonicalization can be a pessimization without LSR to "clean up"
1955 // - We depend on having a preheader; in particular,
1956 // Loop::getCanonicalInductionVariable only supports loops with preheaders,
1957 // and we're in trouble if we can't find the induction variable even when
1958 // we've manually inserted one.
1959 if (!L->isLoopSimplifyForm())
1962 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1963 SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
1964 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1965 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1966 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1967 auto *TTIP = getAnalysisIfAvailable<TargetTransformInfoWrapperPass>();
1968 TTI = TTIP ? &TTIP->getTTI(*L->getHeader()->getParent()) : nullptr;
1969 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
1974 // If there are any floating-point recurrences, attempt to
1975 // transform them to use integer recurrences.
1976 RewriteNonIntegerIVs(L);
1978 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
1980 // Create a rewriter object which we'll use to transform the code with.
1981 SCEVExpander Rewriter(*SE, DL, "indvars");
1983 Rewriter.setDebugType(DEBUG_TYPE);
1986 // Eliminate redundant IV users.
1988 // Simplification works best when run before other consumers of SCEV. We
1989 // attempt to avoid evaluating SCEVs for sign/zero extend operations until
1990 // other expressions involving loop IVs have been evaluated. This helps SCEV
1991 // set no-wrap flags before normalizing sign/zero extension.
1992 Rewriter.disableCanonicalMode();
1993 SimplifyAndExtend(L, Rewriter, LPM);
1995 // Check to see if this loop has a computable loop-invariant execution count.
1996 // If so, this means that we can compute the final value of any expressions
1997 // that are recurrent in the loop, and substitute the exit values from the
1998 // loop into any instructions outside of the loop that use the final values of
1999 // the current expressions.
2001 if (ReplaceExitValue != NeverRepl &&
2002 !isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2003 RewriteLoopExitValues(L, Rewriter);
2005 // Eliminate redundant IV cycles.
2006 NumElimIV += Rewriter.replaceCongruentIVs(L, DT, DeadInsts);
2008 // If we have a trip count expression, rewrite the loop's exit condition
2009 // using it. We can currently only handle loops with a single exit.
2010 if (canExpandBackedgeTakenCount(L, SE, Rewriter) && needsLFTR(L, DT)) {
2011 PHINode *IndVar = FindLoopCounter(L, BackedgeTakenCount, SE, DT);
2013 // Check preconditions for proper SCEVExpander operation. SCEV does not
2014 // express SCEVExpander's dependencies, such as LoopSimplify. Instead any
2015 // pass that uses the SCEVExpander must do it. This does not work well for
2016 // loop passes because SCEVExpander makes assumptions about all loops,
2017 // while LoopPassManager only forces the current loop to be simplified.
2019 // FIXME: SCEV expansion has no way to bail out, so the caller must
2020 // explicitly check any assumptions made by SCEV. Brittle.
2021 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(BackedgeTakenCount);
2022 if (!AR || AR->getLoop()->getLoopPreheader())
2023 (void)LinearFunctionTestReplace(L, BackedgeTakenCount, IndVar,
2027 // Clear the rewriter cache, because values that are in the rewriter's cache
2028 // can be deleted in the loop below, causing the AssertingVH in the cache to
2032 // Now that we're done iterating through lists, clean up any instructions
2033 // which are now dead.
2034 while (!DeadInsts.empty())
2035 if (Instruction *Inst =
2036 dyn_cast_or_null<Instruction>(DeadInsts.pop_back_val()))
2037 RecursivelyDeleteTriviallyDeadInstructions(Inst, TLI);
2039 // The Rewriter may not be used from this point on.
2041 // Loop-invariant instructions in the preheader that aren't used in the
2042 // loop may be sunk below the loop to reduce register pressure.
2043 SinkUnusedInvariants(L);
2045 // Clean up dead instructions.
2046 Changed |= DeleteDeadPHIs(L->getHeader(), TLI);
2047 // Check a post-condition.
2048 assert(L->isLCSSAForm(*DT) &&
2049 "Indvars did not leave the loop in lcssa form!");
2051 // Verify that LFTR, and any other change have not interfered with SCEV's
2052 // ability to compute trip count.
2054 if (VerifyIndvars && !isa<SCEVCouldNotCompute>(BackedgeTakenCount)) {
2056 const SCEV *NewBECount = SE->getBackedgeTakenCount(L);
2057 if (SE->getTypeSizeInBits(BackedgeTakenCount->getType()) <
2058 SE->getTypeSizeInBits(NewBECount->getType()))
2059 NewBECount = SE->getTruncateOrNoop(NewBECount,
2060 BackedgeTakenCount->getType());
2062 BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount,
2063 NewBECount->getType());
2064 assert(BackedgeTakenCount == NewBECount && "indvars must preserve SCEV");