1 //===- IndVarSimplify.cpp - Induction Variable Elimination ----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This transformation analyzes and transforms the induction variables (and
11 // computations derived from them) into simpler forms suitable for subsequent
12 // analysis and transformation.
14 // If the trip count of a loop is computable, this pass also makes the following
16 // 1. The exit condition for the loop is canonicalized to compare the
17 // induction value against the exit value. This turns loops like:
18 // 'for (i = 7; i*i < 1000; ++i)' into 'for (i = 0; i != 25; ++i)'
19 // 2. Any use outside of the loop of an expression derived from the indvar
20 // is changed to compute the derived value outside of the loop, eliminating
21 // the dependence on the exit value of the induction variable. If the only
22 // purpose of the loop is to compute the exit value of some derived
23 // expression, this transformation will make the loop dead.
25 //===----------------------------------------------------------------------===//
27 #include "llvm/Transforms/Scalar.h"
28 #include "llvm/ADT/DenseMap.h"
29 #include "llvm/ADT/SmallVector.h"
30 #include "llvm/ADT/Statistic.h"
31 #include "llvm/Analysis/LoopInfo.h"
32 #include "llvm/Analysis/LoopPass.h"
33 #include "llvm/Analysis/ScalarEvolutionExpander.h"
34 #include "llvm/Analysis/TargetLibraryInfo.h"
35 #include "llvm/Analysis/TargetTransformInfo.h"
36 #include "llvm/IR/BasicBlock.h"
37 #include "llvm/IR/CFG.h"
38 #include "llvm/IR/Constants.h"
39 #include "llvm/IR/DataLayout.h"
40 #include "llvm/IR/Dominators.h"
41 #include "llvm/IR/Instructions.h"
42 #include "llvm/IR/IntrinsicInst.h"
43 #include "llvm/IR/LLVMContext.h"
44 #include "llvm/IR/PatternMatch.h"
45 #include "llvm/IR/Type.h"
46 #include "llvm/Support/CommandLine.h"
47 #include "llvm/Support/Debug.h"
48 #include "llvm/Support/raw_ostream.h"
49 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
50 #include "llvm/Transforms/Utils/Local.h"
51 #include "llvm/Transforms/Utils/SimplifyIndVar.h"
54 #define DEBUG_TYPE "indvars"
56 STATISTIC(NumWidened , "Number of indvars widened");
57 STATISTIC(NumReplaced , "Number of exit values replaced");
58 STATISTIC(NumLFTR , "Number of loop exit tests replaced");
59 STATISTIC(NumElimExt , "Number of IV sign/zero extends eliminated");
60 STATISTIC(NumElimIV , "Number of congruent IVs eliminated");
62 // Trip count verification can be enabled by default under NDEBUG if we
63 // implement a strong expression equivalence checker in SCEV. Until then, we
64 // use the verify-indvars flag, which may assert in some cases.
65 static cl::opt<bool> VerifyIndvars(
66 "verify-indvars", cl::Hidden,
67 cl::desc("Verify the ScalarEvolution result after running indvars"));
69 static cl::opt<bool> ReduceLiveIVs("liv-reduce", cl::Hidden,
70 cl::desc("Reduce live induction variables."));
72 enum ReplaceExitVal { NeverRepl, OnlyCheapRepl, AlwaysRepl };
74 static cl::opt<ReplaceExitVal> ReplaceExitValue(
75 "replexitval", cl::Hidden, cl::init(OnlyCheapRepl),
76 cl::desc("Choose the strategy to replace exit value in IndVarSimplify"),
77 cl::values(clEnumValN(NeverRepl, "never", "never replace exit value"),
78 clEnumValN(OnlyCheapRepl, "cheap",
79 "only replace exit value when the cost is cheap"),
80 clEnumValN(AlwaysRepl, "always",
81 "always replace exit value whenever possible"),
89 class IndVarSimplify : public LoopPass {
93 TargetLibraryInfo *TLI;
94 const TargetTransformInfo *TTI;
96 SmallVector<WeakVH, 16> DeadInsts;
100 static char ID; // Pass identification, replacement for typeid
102 : LoopPass(ID), LI(nullptr), SE(nullptr), DT(nullptr), Changed(false) {
103 initializeIndVarSimplifyPass(*PassRegistry::getPassRegistry());
106 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
108 void getAnalysisUsage(AnalysisUsage &AU) const override {
109 AU.addRequired<DominatorTreeWrapperPass>();
110 AU.addRequired<LoopInfoWrapperPass>();
111 AU.addRequired<ScalarEvolution>();
112 AU.addRequiredID(LoopSimplifyID);
113 AU.addRequiredID(LCSSAID);
114 AU.addPreserved<ScalarEvolution>();
115 AU.addPreservedID(LoopSimplifyID);
116 AU.addPreservedID(LCSSAID);
117 AU.setPreservesCFG();
121 void releaseMemory() override {
125 bool isValidRewrite(Value *FromVal, Value *ToVal);
127 void HandleFloatingPointIV(Loop *L, PHINode *PH);
128 void RewriteNonIntegerIVs(Loop *L);
130 void SimplifyAndExtend(Loop *L, SCEVExpander &Rewriter, LPPassManager &LPM);
132 bool CanLoopBeDeleted(Loop *L, SmallVector<RewritePhi, 8> &RewritePhiSet);
133 void RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter);
135 Value *LinearFunctionTestReplace(Loop *L, const SCEV *BackedgeTakenCount,
136 PHINode *IndVar, SCEVExpander &Rewriter);
138 void SinkUnusedInvariants(Loop *L);
140 Value *ExpandSCEVIfNeeded(SCEVExpander &Rewriter, const SCEV *S, Loop *L,
141 Instruction *InsertPt, Type *Ty);
145 char IndVarSimplify::ID = 0;
146 INITIALIZE_PASS_BEGIN(IndVarSimplify, "indvars",
147 "Induction Variable Simplification", false, false)
148 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
149 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
150 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
151 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
152 INITIALIZE_PASS_DEPENDENCY(LCSSA)
153 INITIALIZE_PASS_END(IndVarSimplify, "indvars",
154 "Induction Variable Simplification", false, false)
156 Pass *llvm::createIndVarSimplifyPass() {
157 return new IndVarSimplify();
160 /// isValidRewrite - Return true if the SCEV expansion generated by the
161 /// rewriter can replace the original value. SCEV guarantees that it
162 /// produces the same value, but the way it is produced may be illegal IR.
163 /// Ideally, this function will only be called for verification.
164 bool IndVarSimplify::isValidRewrite(Value *FromVal, Value *ToVal) {
165 // If an SCEV expression subsumed multiple pointers, its expansion could
166 // reassociate the GEP changing the base pointer. This is illegal because the
167 // final address produced by a GEP chain must be inbounds relative to its
168 // underlying object. Otherwise basic alias analysis, among other things,
169 // could fail in a dangerous way. Ultimately, SCEV will be improved to avoid
170 // producing an expression involving multiple pointers. Until then, we must
173 // Retrieve the pointer operand of the GEP. Don't use GetUnderlyingObject
174 // because it understands lcssa phis while SCEV does not.
175 Value *FromPtr = FromVal;
176 Value *ToPtr = ToVal;
177 if (GEPOperator *GEP = dyn_cast<GEPOperator>(FromVal)) {
178 FromPtr = GEP->getPointerOperand();
180 if (GEPOperator *GEP = dyn_cast<GEPOperator>(ToVal)) {
181 ToPtr = GEP->getPointerOperand();
183 if (FromPtr != FromVal || ToPtr != ToVal) {
184 // Quickly check the common case
185 if (FromPtr == ToPtr)
188 // SCEV may have rewritten an expression that produces the GEP's pointer
189 // operand. That's ok as long as the pointer operand has the same base
190 // pointer. Unlike GetUnderlyingObject(), getPointerBase() will find the
191 // base of a recurrence. This handles the case in which SCEV expansion
192 // converts a pointer type recurrence into a nonrecurrent pointer base
193 // indexed by an integer recurrence.
195 // If the GEP base pointer is a vector of pointers, abort.
196 if (!FromPtr->getType()->isPointerTy() || !ToPtr->getType()->isPointerTy())
199 const SCEV *FromBase = SE->getPointerBase(SE->getSCEV(FromPtr));
200 const SCEV *ToBase = SE->getPointerBase(SE->getSCEV(ToPtr));
201 if (FromBase == ToBase)
204 DEBUG(dbgs() << "INDVARS: GEP rewrite bail out "
205 << *FromBase << " != " << *ToBase << "\n");
212 /// Determine the insertion point for this user. By default, insert immediately
213 /// before the user. SCEVExpander or LICM will hoist loop invariants out of the
214 /// loop. For PHI nodes, there may be multiple uses, so compute the nearest
215 /// common dominator for the incoming blocks.
216 static Instruction *getInsertPointForUses(Instruction *User, Value *Def,
218 PHINode *PHI = dyn_cast<PHINode>(User);
222 Instruction *InsertPt = nullptr;
223 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i) {
224 if (PHI->getIncomingValue(i) != Def)
227 BasicBlock *InsertBB = PHI->getIncomingBlock(i);
229 InsertPt = InsertBB->getTerminator();
232 InsertBB = DT->findNearestCommonDominator(InsertPt->getParent(), InsertBB);
233 InsertPt = InsertBB->getTerminator();
235 assert(InsertPt && "Missing phi operand");
236 assert((!isa<Instruction>(Def) ||
237 DT->dominates(cast<Instruction>(Def), InsertPt)) &&
238 "def does not dominate all uses");
242 //===----------------------------------------------------------------------===//
243 // RewriteNonIntegerIVs and helpers. Prefer integer IVs.
244 //===----------------------------------------------------------------------===//
246 /// ConvertToSInt - Convert APF to an integer, if possible.
247 static bool ConvertToSInt(const APFloat &APF, int64_t &IntVal) {
248 bool isExact = false;
249 // See if we can convert this to an int64_t
251 if (APF.convertToInteger(&UIntVal, 64, true, APFloat::rmTowardZero,
252 &isExact) != APFloat::opOK || !isExact)
258 /// HandleFloatingPointIV - If the loop has floating induction variable
259 /// then insert corresponding integer induction variable if possible.
261 /// for(double i = 0; i < 10000; ++i)
263 /// is converted into
264 /// for(int i = 0; i < 10000; ++i)
267 void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PN) {
268 unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
269 unsigned BackEdge = IncomingEdge^1;
271 // Check incoming value.
272 ConstantFP *InitValueVal =
273 dyn_cast<ConstantFP>(PN->getIncomingValue(IncomingEdge));
276 if (!InitValueVal || !ConvertToSInt(InitValueVal->getValueAPF(), InitValue))
279 // Check IV increment. Reject this PN if increment operation is not
280 // an add or increment value can not be represented by an integer.
281 BinaryOperator *Incr =
282 dyn_cast<BinaryOperator>(PN->getIncomingValue(BackEdge));
283 if (Incr == nullptr || Incr->getOpcode() != Instruction::FAdd) return;
285 // If this is not an add of the PHI with a constantfp, or if the constant fp
286 // is not an integer, bail out.
287 ConstantFP *IncValueVal = dyn_cast<ConstantFP>(Incr->getOperand(1));
289 if (IncValueVal == nullptr || Incr->getOperand(0) != PN ||
290 !ConvertToSInt(IncValueVal->getValueAPF(), IncValue))
293 // Check Incr uses. One user is PN and the other user is an exit condition
294 // used by the conditional terminator.
295 Value::user_iterator IncrUse = Incr->user_begin();
296 Instruction *U1 = cast<Instruction>(*IncrUse++);
297 if (IncrUse == Incr->user_end()) return;
298 Instruction *U2 = cast<Instruction>(*IncrUse++);
299 if (IncrUse != Incr->user_end()) return;
301 // Find exit condition, which is an fcmp. If it doesn't exist, or if it isn't
302 // only used by a branch, we can't transform it.
303 FCmpInst *Compare = dyn_cast<FCmpInst>(U1);
305 Compare = dyn_cast<FCmpInst>(U2);
306 if (!Compare || !Compare->hasOneUse() ||
307 !isa<BranchInst>(Compare->user_back()))
310 BranchInst *TheBr = cast<BranchInst>(Compare->user_back());
312 // We need to verify that the branch actually controls the iteration count
313 // of the loop. If not, the new IV can overflow and no one will notice.
314 // The branch block must be in the loop and one of the successors must be out
316 assert(TheBr->isConditional() && "Can't use fcmp if not conditional");
317 if (!L->contains(TheBr->getParent()) ||
318 (L->contains(TheBr->getSuccessor(0)) &&
319 L->contains(TheBr->getSuccessor(1))))
323 // If it isn't a comparison with an integer-as-fp (the exit value), we can't
325 ConstantFP *ExitValueVal = dyn_cast<ConstantFP>(Compare->getOperand(1));
327 if (ExitValueVal == nullptr ||
328 !ConvertToSInt(ExitValueVal->getValueAPF(), ExitValue))
331 // Find new predicate for integer comparison.
332 CmpInst::Predicate NewPred = CmpInst::BAD_ICMP_PREDICATE;
333 switch (Compare->getPredicate()) {
334 default: return; // Unknown comparison.
335 case CmpInst::FCMP_OEQ:
336 case CmpInst::FCMP_UEQ: NewPred = CmpInst::ICMP_EQ; break;
337 case CmpInst::FCMP_ONE:
338 case CmpInst::FCMP_UNE: NewPred = CmpInst::ICMP_NE; break;
339 case CmpInst::FCMP_OGT:
340 case CmpInst::FCMP_UGT: NewPred = CmpInst::ICMP_SGT; break;
341 case CmpInst::FCMP_OGE:
342 case CmpInst::FCMP_UGE: NewPred = CmpInst::ICMP_SGE; break;
343 case CmpInst::FCMP_OLT:
344 case CmpInst::FCMP_ULT: NewPred = CmpInst::ICMP_SLT; break;
345 case CmpInst::FCMP_OLE:
346 case CmpInst::FCMP_ULE: NewPred = CmpInst::ICMP_SLE; break;
349 // We convert the floating point induction variable to a signed i32 value if
350 // we can. This is only safe if the comparison will not overflow in a way
351 // that won't be trapped by the integer equivalent operations. Check for this
353 // TODO: We could use i64 if it is native and the range requires it.
355 // The start/stride/exit values must all fit in signed i32.
356 if (!isInt<32>(InitValue) || !isInt<32>(IncValue) || !isInt<32>(ExitValue))
359 // If not actually striding (add x, 0.0), avoid touching the code.
363 // Positive and negative strides have different safety conditions.
365 // If we have a positive stride, we require the init to be less than the
367 if (InitValue >= ExitValue)
370 uint32_t Range = uint32_t(ExitValue-InitValue);
371 // Check for infinite loop, either:
372 // while (i <= Exit) or until (i > Exit)
373 if (NewPred == CmpInst::ICMP_SLE || NewPred == CmpInst::ICMP_SGT) {
374 if (++Range == 0) return; // Range overflows.
377 unsigned Leftover = Range % uint32_t(IncValue);
379 // If this is an equality comparison, we require that the strided value
380 // exactly land on the exit value, otherwise the IV condition will wrap
381 // around and do things the fp IV wouldn't.
382 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
386 // If the stride would wrap around the i32 before exiting, we can't
388 if (Leftover != 0 && int32_t(ExitValue+IncValue) < ExitValue)
392 // If we have a negative stride, we require the init to be greater than the
394 if (InitValue <= ExitValue)
397 uint32_t Range = uint32_t(InitValue-ExitValue);
398 // Check for infinite loop, either:
399 // while (i >= Exit) or until (i < Exit)
400 if (NewPred == CmpInst::ICMP_SGE || NewPred == CmpInst::ICMP_SLT) {
401 if (++Range == 0) return; // Range overflows.
404 unsigned Leftover = Range % uint32_t(-IncValue);
406 // If this is an equality comparison, we require that the strided value
407 // exactly land on the exit value, otherwise the IV condition will wrap
408 // around and do things the fp IV wouldn't.
409 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
413 // If the stride would wrap around the i32 before exiting, we can't
415 if (Leftover != 0 && int32_t(ExitValue+IncValue) > ExitValue)
419 IntegerType *Int32Ty = Type::getInt32Ty(PN->getContext());
421 // Insert new integer induction variable.
422 PHINode *NewPHI = PHINode::Create(Int32Ty, 2, PN->getName()+".int", PN);
423 NewPHI->addIncoming(ConstantInt::get(Int32Ty, InitValue),
424 PN->getIncomingBlock(IncomingEdge));
427 BinaryOperator::CreateAdd(NewPHI, ConstantInt::get(Int32Ty, IncValue),
428 Incr->getName()+".int", Incr);
429 NewPHI->addIncoming(NewAdd, PN->getIncomingBlock(BackEdge));
431 ICmpInst *NewCompare = new ICmpInst(TheBr, NewPred, NewAdd,
432 ConstantInt::get(Int32Ty, ExitValue),
435 // In the following deletions, PN may become dead and may be deleted.
436 // Use a WeakVH to observe whether this happens.
439 // Delete the old floating point exit comparison. The branch starts using the
441 NewCompare->takeName(Compare);
442 Compare->replaceAllUsesWith(NewCompare);
443 RecursivelyDeleteTriviallyDeadInstructions(Compare, TLI);
445 // Delete the old floating point increment.
446 Incr->replaceAllUsesWith(UndefValue::get(Incr->getType()));
447 RecursivelyDeleteTriviallyDeadInstructions(Incr, TLI);
449 // If the FP induction variable still has uses, this is because something else
450 // in the loop uses its value. In order to canonicalize the induction
451 // variable, we chose to eliminate the IV and rewrite it in terms of an
454 // We give preference to sitofp over uitofp because it is faster on most
457 Value *Conv = new SIToFPInst(NewPHI, PN->getType(), "indvar.conv",
458 PN->getParent()->getFirstInsertionPt());
459 PN->replaceAllUsesWith(Conv);
460 RecursivelyDeleteTriviallyDeadInstructions(PN, TLI);
465 void IndVarSimplify::RewriteNonIntegerIVs(Loop *L) {
466 // First step. Check to see if there are any floating-point recurrences.
467 // If there are, change them into integer recurrences, permitting analysis by
468 // the SCEV routines.
470 BasicBlock *Header = L->getHeader();
472 SmallVector<WeakVH, 8> PHIs;
473 for (BasicBlock::iterator I = Header->begin();
474 PHINode *PN = dyn_cast<PHINode>(I); ++I)
477 for (unsigned i = 0, e = PHIs.size(); i != e; ++i)
478 if (PHINode *PN = dyn_cast_or_null<PHINode>(&*PHIs[i]))
479 HandleFloatingPointIV(L, PN);
481 // If the loop previously had floating-point IV, ScalarEvolution
482 // may not have been able to compute a trip count. Now that we've done some
483 // re-writing, the trip count may be computable.
489 // Collect information about PHI nodes which can be transformed in
490 // RewriteLoopExitValues.
493 unsigned Ith; // Ith incoming value.
494 Value *Val; // Exit value after expansion.
495 bool HighCost; // High Cost when expansion.
496 bool SafePhi; // LCSSASafePhiForRAUW.
498 RewritePhi(PHINode *P, unsigned I, Value *V, bool H, bool S)
499 : PN(P), Ith(I), Val(V), HighCost(H), SafePhi(S) {}
503 Value *IndVarSimplify::ExpandSCEVIfNeeded(SCEVExpander &Rewriter, const SCEV *S,
504 Loop *L, Instruction *InsertPt,
506 // Before expanding S into an expensive LLVM expression, see if we can use an
507 // already existing value as the expansion for S.
508 if (Value *RetValue = Rewriter.findExistingExpansion(S, InsertPt, L))
511 // We didn't find anything, fall back to using SCEVExpander.
512 return Rewriter.expandCodeFor(S, ResultTy, InsertPt);
515 //===----------------------------------------------------------------------===//
516 // RewriteLoopExitValues - Optimize IV users outside the loop.
517 // As a side effect, reduces the amount of IV processing within the loop.
518 //===----------------------------------------------------------------------===//
520 /// RewriteLoopExitValues - Check to see if this loop has a computable
521 /// loop-invariant execution count. If so, this means that we can compute the
522 /// final value of any expressions that are recurrent in the loop, and
523 /// substitute the exit values from the loop into any instructions outside of
524 /// the loop that use the final values of the current expressions.
526 /// This is mostly redundant with the regular IndVarSimplify activities that
527 /// happen later, except that it's more powerful in some cases, because it's
528 /// able to brute-force evaluate arbitrary instructions as long as they have
529 /// constant operands at the beginning of the loop.
530 void IndVarSimplify::RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter) {
531 // Verify the input to the pass in already in LCSSA form.
532 assert(L->isLCSSAForm(*DT));
534 SmallVector<BasicBlock*, 8> ExitBlocks;
535 L->getUniqueExitBlocks(ExitBlocks);
537 SmallVector<RewritePhi, 8> RewritePhiSet;
538 // Find all values that are computed inside the loop, but used outside of it.
539 // Because of LCSSA, these values will only occur in LCSSA PHI Nodes. Scan
540 // the exit blocks of the loop to find them.
541 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
542 BasicBlock *ExitBB = ExitBlocks[i];
544 // If there are no PHI nodes in this exit block, then no values defined
545 // inside the loop are used on this path, skip it.
546 PHINode *PN = dyn_cast<PHINode>(ExitBB->begin());
549 unsigned NumPreds = PN->getNumIncomingValues();
551 // We would like to be able to RAUW single-incoming value PHI nodes. We
552 // have to be certain this is safe even when this is an LCSSA PHI node.
553 // While the computed exit value is no longer varying in *this* loop, the
554 // exit block may be an exit block for an outer containing loop as well,
555 // the exit value may be varying in the outer loop, and thus it may still
556 // require an LCSSA PHI node. The safe case is when this is
557 // single-predecessor PHI node (LCSSA) and the exit block containing it is
558 // part of the enclosing loop, or this is the outer most loop of the nest.
559 // In either case the exit value could (at most) be varying in the same
560 // loop body as the phi node itself. Thus if it is in turn used outside of
561 // an enclosing loop it will only be via a separate LCSSA node.
562 bool LCSSASafePhiForRAUW =
564 (!L->getParentLoop() || L->getParentLoop() == LI->getLoopFor(ExitBB));
566 // Iterate over all of the PHI nodes.
567 BasicBlock::iterator BBI = ExitBB->begin();
568 while ((PN = dyn_cast<PHINode>(BBI++))) {
570 continue; // dead use, don't replace it
572 // SCEV only supports integer expressions for now.
573 if (!PN->getType()->isIntegerTy() && !PN->getType()->isPointerTy())
576 // It's necessary to tell ScalarEvolution about this explicitly so that
577 // it can walk the def-use list and forget all SCEVs, as it may not be
578 // watching the PHI itself. Once the new exit value is in place, there
579 // may not be a def-use connection between the loop and every instruction
580 // which got a SCEVAddRecExpr for that loop.
583 // Iterate over all of the values in all the PHI nodes.
584 for (unsigned i = 0; i != NumPreds; ++i) {
585 // If the value being merged in is not integer or is not defined
586 // in the loop, skip it.
587 Value *InVal = PN->getIncomingValue(i);
588 if (!isa<Instruction>(InVal))
591 // If this pred is for a subloop, not L itself, skip it.
592 if (LI->getLoopFor(PN->getIncomingBlock(i)) != L)
593 continue; // The Block is in a subloop, skip it.
595 // Check that InVal is defined in the loop.
596 Instruction *Inst = cast<Instruction>(InVal);
597 if (!L->contains(Inst))
600 // Okay, this instruction has a user outside of the current loop
601 // and varies predictably *inside* the loop. Evaluate the value it
602 // contains when the loop exits, if possible.
603 const SCEV *ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop());
604 if (!SE->isLoopInvariant(ExitValue, L) ||
605 !isSafeToExpand(ExitValue, *SE))
608 // Computing the value outside of the loop brings no benefit if :
609 // - it is definitely used inside the loop in a way which can not be
611 // - no use outside of the loop can take advantage of hoisting the
612 // computation out of the loop
613 if (ExitValue->getSCEVType()>=scMulExpr) {
614 unsigned NumHardInternalUses = 0;
615 unsigned NumSoftExternalUses = 0;
616 unsigned NumUses = 0;
617 for (auto IB = Inst->user_begin(), IE = Inst->user_end();
618 IB != IE && NumUses <= 6; ++IB) {
619 Instruction *UseInstr = cast<Instruction>(*IB);
620 unsigned Opc = UseInstr->getOpcode();
622 if (L->contains(UseInstr)) {
623 if (Opc == Instruction::Call || Opc == Instruction::Ret)
624 NumHardInternalUses++;
626 if (Opc == Instruction::PHI) {
627 // Do not count the Phi as a use. LCSSA may have inserted
628 // plenty of trivial ones.
630 for (auto PB = UseInstr->user_begin(),
631 PE = UseInstr->user_end();
632 PB != PE && NumUses <= 6; ++PB, ++NumUses) {
633 unsigned PhiOpc = cast<Instruction>(*PB)->getOpcode();
634 if (PhiOpc != Instruction::Call && PhiOpc != Instruction::Ret)
635 NumSoftExternalUses++;
639 if (Opc != Instruction::Call && Opc != Instruction::Ret)
640 NumSoftExternalUses++;
643 if (NumUses <= 6 && NumHardInternalUses && !NumSoftExternalUses)
647 bool HighCost = Rewriter.isHighCostExpansion(ExitValue, L, Inst);
649 ExpandSCEVIfNeeded(Rewriter, ExitValue, L, Inst, PN->getType());
651 DEBUG(dbgs() << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal << '\n'
652 << " LoopVal = " << *Inst << "\n");
654 if (!isValidRewrite(Inst, ExitVal)) {
655 DeadInsts.push_back(ExitVal);
659 // Collect all the candidate PHINodes to be rewritten.
660 RewritePhiSet.push_back(
661 RewritePhi(PN, i, ExitVal, HighCost, LCSSASafePhiForRAUW));
666 bool LoopCanBeDel = CanLoopBeDeleted(L, RewritePhiSet);
669 for (const RewritePhi &Phi : RewritePhiSet) {
670 PHINode *PN = Phi.PN;
671 Value *ExitVal = Phi.Val;
673 // Only do the rewrite when the ExitValue can be expanded cheaply.
674 // If LoopCanBeDel is true, rewrite exit value aggressively.
675 if (ReplaceExitValue == OnlyCheapRepl && !LoopCanBeDel && Phi.HighCost) {
676 DeadInsts.push_back(ExitVal);
682 Instruction *Inst = cast<Instruction>(PN->getIncomingValue(Phi.Ith));
683 PN->setIncomingValue(Phi.Ith, ExitVal);
685 // If this instruction is dead now, delete it. Don't do it now to avoid
686 // invalidating iterators.
687 if (isInstructionTriviallyDead(Inst, TLI))
688 DeadInsts.push_back(Inst);
690 // If we determined that this PHI is safe to replace even if an LCSSA
693 PN->replaceAllUsesWith(ExitVal);
694 PN->eraseFromParent();
698 // The insertion point instruction may have been deleted; clear it out
699 // so that the rewriter doesn't trip over it later.
700 Rewriter.clearInsertPoint();
703 /// CanLoopBeDeleted - Check whether it is possible to delete the loop after
704 /// rewriting exit value. If it is possible, ignore ReplaceExitValue and
705 /// do rewriting aggressively.
706 bool IndVarSimplify::CanLoopBeDeleted(
707 Loop *L, SmallVector<RewritePhi, 8> &RewritePhiSet) {
709 BasicBlock *Preheader = L->getLoopPreheader();
710 // If there is no preheader, the loop will not be deleted.
714 // In LoopDeletion pass Loop can be deleted when ExitingBlocks.size() > 1.
715 // We obviate multiple ExitingBlocks case for simplicity.
716 // TODO: If we see testcase with multiple ExitingBlocks can be deleted
717 // after exit value rewriting, we can enhance the logic here.
718 SmallVector<BasicBlock *, 4> ExitingBlocks;
719 L->getExitingBlocks(ExitingBlocks);
720 SmallVector<BasicBlock *, 8> ExitBlocks;
721 L->getUniqueExitBlocks(ExitBlocks);
722 if (ExitBlocks.size() > 1 || ExitingBlocks.size() > 1)
725 BasicBlock *ExitBlock = ExitBlocks[0];
726 BasicBlock::iterator BI = ExitBlock->begin();
727 while (PHINode *P = dyn_cast<PHINode>(BI)) {
728 Value *Incoming = P->getIncomingValueForBlock(ExitingBlocks[0]);
730 // If the Incoming value of P is found in RewritePhiSet, we know it
731 // could be rewritten to use a loop invariant value in transformation
732 // phase later. Skip it in the loop invariant check below.
734 for (const RewritePhi &Phi : RewritePhiSet) {
735 unsigned i = Phi.Ith;
736 if (Phi.PN == P && (Phi.PN)->getIncomingValue(i) == Incoming) {
743 if (!found && (I = dyn_cast<Instruction>(Incoming)))
744 if (!L->hasLoopInvariantOperands(I))
750 for (Loop::block_iterator LI = L->block_begin(), LE = L->block_end();
752 for (BasicBlock::iterator BI = (*LI)->begin(), BE = (*LI)->end(); BI != BE;
754 if (BI->mayHaveSideEffects())
762 //===----------------------------------------------------------------------===//
763 // IV Widening - Extend the width of an IV to cover its widest uses.
764 //===----------------------------------------------------------------------===//
767 // Collect information about induction variables that are used by sign/zero
768 // extend operations. This information is recorded by CollectExtend and
769 // provides the input to WidenIV.
772 Type *WidestNativeType; // Widest integer type created [sz]ext
773 bool IsSigned; // Was a sext user seen before a zext?
775 WideIVInfo() : NarrowIV(nullptr), WidestNativeType(nullptr),
780 /// visitCast - Update information about the induction variable that is
781 /// extended by this sign or zero extend operation. This is used to determine
782 /// the final width of the IV before actually widening it.
783 static void visitIVCast(CastInst *Cast, WideIVInfo &WI, ScalarEvolution *SE,
784 const TargetTransformInfo *TTI) {
785 bool IsSigned = Cast->getOpcode() == Instruction::SExt;
786 if (!IsSigned && Cast->getOpcode() != Instruction::ZExt)
789 Type *Ty = Cast->getType();
790 uint64_t Width = SE->getTypeSizeInBits(Ty);
791 if (!Cast->getModule()->getDataLayout().isLegalInteger(Width))
794 // Cast is either an sext or zext up to this point.
795 // We should not widen an indvar if arithmetics on the wider indvar are more
796 // expensive than those on the narrower indvar. We check only the cost of ADD
797 // because at least an ADD is required to increment the induction variable. We
798 // could compute more comprehensively the cost of all instructions on the
799 // induction variable when necessary.
801 TTI->getArithmeticInstrCost(Instruction::Add, Ty) >
802 TTI->getArithmeticInstrCost(Instruction::Add,
803 Cast->getOperand(0)->getType())) {
807 if (!WI.WidestNativeType) {
808 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
809 WI.IsSigned = IsSigned;
813 // We extend the IV to satisfy the sign of its first user, arbitrarily.
814 if (WI.IsSigned != IsSigned)
817 if (Width > SE->getTypeSizeInBits(WI.WidestNativeType))
818 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
823 /// NarrowIVDefUse - Record a link in the Narrow IV def-use chain along with the
824 /// WideIV that computes the same value as the Narrow IV def. This avoids
825 /// caching Use* pointers.
826 struct NarrowIVDefUse {
827 Instruction *NarrowDef;
828 Instruction *NarrowUse;
829 Instruction *WideDef;
831 NarrowIVDefUse(): NarrowDef(nullptr), NarrowUse(nullptr), WideDef(nullptr) {}
833 NarrowIVDefUse(Instruction *ND, Instruction *NU, Instruction *WD):
834 NarrowDef(ND), NarrowUse(NU), WideDef(WD) {}
837 /// WidenIV - The goal of this transform is to remove sign and zero extends
838 /// without creating any new induction variables. To do this, it creates a new
839 /// phi of the wider type and redirects all users, either removing extends or
840 /// inserting truncs whenever we stop propagating the type.
856 Instruction *WideInc;
857 const SCEV *WideIncExpr;
858 SmallVectorImpl<WeakVH> &DeadInsts;
860 SmallPtrSet<Instruction*,16> Widened;
861 SmallVector<NarrowIVDefUse, 8> NarrowIVUsers;
864 WidenIV(const WideIVInfo &WI, LoopInfo *LInfo,
865 ScalarEvolution *SEv, DominatorTree *DTree,
866 SmallVectorImpl<WeakVH> &DI) :
867 OrigPhi(WI.NarrowIV),
868 WideType(WI.WidestNativeType),
869 IsSigned(WI.IsSigned),
871 L(LI->getLoopFor(OrigPhi->getParent())),
876 WideIncExpr(nullptr),
878 assert(L->getHeader() == OrigPhi->getParent() && "Phi must be an IV");
881 PHINode *CreateWideIV(SCEVExpander &Rewriter);
884 Value *getExtend(Value *NarrowOper, Type *WideType, bool IsSigned,
887 Instruction *CloneIVUser(NarrowIVDefUse DU);
889 const SCEVAddRecExpr *GetWideRecurrence(Instruction *NarrowUse);
891 const SCEVAddRecExpr* GetExtendedOperandRecurrence(NarrowIVDefUse DU);
893 const SCEV *GetSCEVByOpCode(const SCEV *LHS, const SCEV *RHS,
894 unsigned OpCode) const;
896 Instruction *WidenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter);
898 bool WidenLoopCompare(NarrowIVDefUse DU);
900 void pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef);
902 } // anonymous namespace
904 /// isLoopInvariant - Perform a quick domtree based check for loop invariance
905 /// assuming that V is used within the loop. LoopInfo::isLoopInvariant() seems
906 /// gratuitous for this purpose.
907 static bool isLoopInvariant(Value *V, const Loop *L, const DominatorTree *DT) {
908 Instruction *Inst = dyn_cast<Instruction>(V);
912 return DT->properlyDominates(Inst->getParent(), L->getHeader());
915 Value *WidenIV::getExtend(Value *NarrowOper, Type *WideType, bool IsSigned,
917 // Set the debug location and conservative insertion point.
918 IRBuilder<> Builder(Use);
919 // Hoist the insertion point into loop preheaders as far as possible.
920 for (const Loop *L = LI->getLoopFor(Use->getParent());
921 L && L->getLoopPreheader() && isLoopInvariant(NarrowOper, L, DT);
922 L = L->getParentLoop())
923 Builder.SetInsertPoint(L->getLoopPreheader()->getTerminator());
925 return IsSigned ? Builder.CreateSExt(NarrowOper, WideType) :
926 Builder.CreateZExt(NarrowOper, WideType);
929 /// CloneIVUser - Instantiate a wide operation to replace a narrow
930 /// operation. This only needs to handle operations that can evaluation to
931 /// SCEVAddRec. It can safely return 0 for any operation we decide not to clone.
932 Instruction *WidenIV::CloneIVUser(NarrowIVDefUse DU) {
933 unsigned Opcode = DU.NarrowUse->getOpcode();
937 case Instruction::Add:
938 case Instruction::Mul:
939 case Instruction::UDiv:
940 case Instruction::Sub:
941 case Instruction::And:
942 case Instruction::Or:
943 case Instruction::Xor:
944 case Instruction::Shl:
945 case Instruction::LShr:
946 case Instruction::AShr:
947 DEBUG(dbgs() << "Cloning IVUser: " << *DU.NarrowUse << "\n");
949 // Replace NarrowDef operands with WideDef. Otherwise, we don't know
950 // anything about the narrow operand yet so must insert a [sz]ext. It is
951 // probably loop invariant and will be folded or hoisted. If it actually
952 // comes from a widened IV, it should be removed during a future call to
954 Value *LHS = (DU.NarrowUse->getOperand(0) == DU.NarrowDef) ? DU.WideDef :
955 getExtend(DU.NarrowUse->getOperand(0), WideType, IsSigned, DU.NarrowUse);
956 Value *RHS = (DU.NarrowUse->getOperand(1) == DU.NarrowDef) ? DU.WideDef :
957 getExtend(DU.NarrowUse->getOperand(1), WideType, IsSigned, DU.NarrowUse);
959 BinaryOperator *NarrowBO = cast<BinaryOperator>(DU.NarrowUse);
960 BinaryOperator *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(),
962 NarrowBO->getName());
963 IRBuilder<> Builder(DU.NarrowUse);
964 Builder.Insert(WideBO);
965 if (const OverflowingBinaryOperator *OBO =
966 dyn_cast<OverflowingBinaryOperator>(NarrowBO)) {
967 if (OBO->hasNoUnsignedWrap()) WideBO->setHasNoUnsignedWrap();
968 if (OBO->hasNoSignedWrap()) WideBO->setHasNoSignedWrap();
974 const SCEV *WidenIV::GetSCEVByOpCode(const SCEV *LHS, const SCEV *RHS,
975 unsigned OpCode) const {
976 if (OpCode == Instruction::Add)
977 return SE->getAddExpr(LHS, RHS);
978 if (OpCode == Instruction::Sub)
979 return SE->getMinusSCEV(LHS, RHS);
980 if (OpCode == Instruction::Mul)
981 return SE->getMulExpr(LHS, RHS);
983 llvm_unreachable("Unsupported opcode.");
986 /// No-wrap operations can transfer sign extension of their result to their
987 /// operands. Generate the SCEV value for the widened operation without
988 /// actually modifying the IR yet. If the expression after extending the
989 /// operands is an AddRec for this loop, return it.
990 const SCEVAddRecExpr* WidenIV::GetExtendedOperandRecurrence(NarrowIVDefUse DU) {
992 // Handle the common case of add<nsw/nuw>
993 const unsigned OpCode = DU.NarrowUse->getOpcode();
994 // Only Add/Sub/Mul instructions supported yet.
995 if (OpCode != Instruction::Add && OpCode != Instruction::Sub &&
996 OpCode != Instruction::Mul)
999 // One operand (NarrowDef) has already been extended to WideDef. Now determine
1000 // if extending the other will lead to a recurrence.
1001 const unsigned ExtendOperIdx =
1002 DU.NarrowUse->getOperand(0) == DU.NarrowDef ? 1 : 0;
1003 assert(DU.NarrowUse->getOperand(1-ExtendOperIdx) == DU.NarrowDef && "bad DU");
1005 const SCEV *ExtendOperExpr = nullptr;
1006 const OverflowingBinaryOperator *OBO =
1007 cast<OverflowingBinaryOperator>(DU.NarrowUse);
1008 if (IsSigned && OBO->hasNoSignedWrap())
1009 ExtendOperExpr = SE->getSignExtendExpr(
1010 SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType);
1011 else if(!IsSigned && OBO->hasNoUnsignedWrap())
1012 ExtendOperExpr = SE->getZeroExtendExpr(
1013 SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType);
1017 // When creating this SCEV expr, don't apply the current operations NSW or NUW
1018 // flags. This instruction may be guarded by control flow that the no-wrap
1019 // behavior depends on. Non-control-equivalent instructions can be mapped to
1020 // the same SCEV expression, and it would be incorrect to transfer NSW/NUW
1021 // semantics to those operations.
1022 const SCEV *lhs = SE->getSCEV(DU.WideDef);
1023 const SCEV *rhs = ExtendOperExpr;
1025 // Let's swap operands to the initial order for the case of non-commutative
1026 // operations, like SUB. See PR21014.
1027 if (ExtendOperIdx == 0)
1028 std::swap(lhs, rhs);
1029 const SCEVAddRecExpr *AddRec =
1030 dyn_cast<SCEVAddRecExpr>(GetSCEVByOpCode(lhs, rhs, OpCode));
1032 if (!AddRec || AddRec->getLoop() != L)
1037 /// GetWideRecurrence - Is this instruction potentially interesting for further
1038 /// simplification after widening it's type? In other words, can the
1039 /// extend be safely hoisted out of the loop with SCEV reducing the value to a
1040 /// recurrence on the same loop. If so, return the sign or zero extended
1041 /// recurrence. Otherwise return NULL.
1042 const SCEVAddRecExpr *WidenIV::GetWideRecurrence(Instruction *NarrowUse) {
1043 if (!SE->isSCEVable(NarrowUse->getType()))
1046 const SCEV *NarrowExpr = SE->getSCEV(NarrowUse);
1047 if (SE->getTypeSizeInBits(NarrowExpr->getType())
1048 >= SE->getTypeSizeInBits(WideType)) {
1049 // NarrowUse implicitly widens its operand. e.g. a gep with a narrow
1050 // index. So don't follow this use.
1054 const SCEV *WideExpr = IsSigned ?
1055 SE->getSignExtendExpr(NarrowExpr, WideType) :
1056 SE->getZeroExtendExpr(NarrowExpr, WideType);
1057 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(WideExpr);
1058 if (!AddRec || AddRec->getLoop() != L)
1063 /// This IV user cannot be widen. Replace this use of the original narrow IV
1064 /// with a truncation of the new wide IV to isolate and eliminate the narrow IV.
1065 static void truncateIVUse(NarrowIVDefUse DU, DominatorTree *DT) {
1066 DEBUG(dbgs() << "INDVARS: Truncate IV " << *DU.WideDef
1067 << " for user " << *DU.NarrowUse << "\n");
1068 IRBuilder<> Builder(getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT));
1069 Value *Trunc = Builder.CreateTrunc(DU.WideDef, DU.NarrowDef->getType());
1070 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, Trunc);
1073 /// If the narrow use is a compare instruction, then widen the compare
1074 // (and possibly the other operand). The extend operation is hoisted into the
1075 // loop preheader as far as possible.
1076 bool WidenIV::WidenLoopCompare(NarrowIVDefUse DU) {
1077 ICmpInst *Cmp = dyn_cast<ICmpInst>(DU.NarrowUse);
1081 // Sign of IV user and compare must match.
1082 if (IsSigned != CmpInst::isSigned(Cmp->getPredicate()))
1085 Value *Op = Cmp->getOperand(Cmp->getOperand(0) == DU.NarrowDef ? 1 : 0);
1086 unsigned CastWidth = SE->getTypeSizeInBits(Op->getType());
1087 unsigned IVWidth = SE->getTypeSizeInBits(WideType);
1088 assert (CastWidth <= IVWidth && "Unexpected width while widening compare.");
1090 // Widen the compare instruction.
1091 IRBuilder<> Builder(getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT));
1092 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef);
1094 // Widen the other operand of the compare, if necessary.
1095 if (CastWidth < IVWidth) {
1096 Value *ExtOp = getExtend(Op, WideType, IsSigned, Cmp);
1097 DU.NarrowUse->replaceUsesOfWith(Op, ExtOp);
1102 /// WidenIVUse - Determine whether an individual user of the narrow IV can be
1103 /// widened. If so, return the wide clone of the user.
1104 Instruction *WidenIV::WidenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter) {
1106 // Stop traversing the def-use chain at inner-loop phis or post-loop phis.
1107 if (PHINode *UsePhi = dyn_cast<PHINode>(DU.NarrowUse)) {
1108 if (LI->getLoopFor(UsePhi->getParent()) != L) {
1109 // For LCSSA phis, sink the truncate outside the loop.
1110 // After SimplifyCFG most loop exit targets have a single predecessor.
1111 // Otherwise fall back to a truncate within the loop.
1112 if (UsePhi->getNumOperands() != 1)
1113 truncateIVUse(DU, DT);
1116 PHINode::Create(DU.WideDef->getType(), 1, UsePhi->getName() + ".wide",
1118 WidePhi->addIncoming(DU.WideDef, UsePhi->getIncomingBlock(0));
1119 IRBuilder<> Builder(WidePhi->getParent()->getFirstInsertionPt());
1120 Value *Trunc = Builder.CreateTrunc(WidePhi, DU.NarrowDef->getType());
1121 UsePhi->replaceAllUsesWith(Trunc);
1122 DeadInsts.emplace_back(UsePhi);
1123 DEBUG(dbgs() << "INDVARS: Widen lcssa phi " << *UsePhi
1124 << " to " << *WidePhi << "\n");
1129 // Our raison d'etre! Eliminate sign and zero extension.
1130 if (IsSigned ? isa<SExtInst>(DU.NarrowUse) : isa<ZExtInst>(DU.NarrowUse)) {
1131 Value *NewDef = DU.WideDef;
1132 if (DU.NarrowUse->getType() != WideType) {
1133 unsigned CastWidth = SE->getTypeSizeInBits(DU.NarrowUse->getType());
1134 unsigned IVWidth = SE->getTypeSizeInBits(WideType);
1135 if (CastWidth < IVWidth) {
1136 // The cast isn't as wide as the IV, so insert a Trunc.
1137 IRBuilder<> Builder(DU.NarrowUse);
1138 NewDef = Builder.CreateTrunc(DU.WideDef, DU.NarrowUse->getType());
1141 // A wider extend was hidden behind a narrower one. This may induce
1142 // another round of IV widening in which the intermediate IV becomes
1143 // dead. It should be very rare.
1144 DEBUG(dbgs() << "INDVARS: New IV " << *WidePhi
1145 << " not wide enough to subsume " << *DU.NarrowUse << "\n");
1146 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef);
1147 NewDef = DU.NarrowUse;
1150 if (NewDef != DU.NarrowUse) {
1151 DEBUG(dbgs() << "INDVARS: eliminating " << *DU.NarrowUse
1152 << " replaced by " << *DU.WideDef << "\n");
1154 DU.NarrowUse->replaceAllUsesWith(NewDef);
1155 DeadInsts.emplace_back(DU.NarrowUse);
1157 // Now that the extend is gone, we want to expose it's uses for potential
1158 // further simplification. We don't need to directly inform SimplifyIVUsers
1159 // of the new users, because their parent IV will be processed later as a
1160 // new loop phi. If we preserved IVUsers analysis, we would also want to
1161 // push the uses of WideDef here.
1163 // No further widening is needed. The deceased [sz]ext had done it for us.
1167 // Does this user itself evaluate to a recurrence after widening?
1168 const SCEVAddRecExpr *WideAddRec = GetWideRecurrence(DU.NarrowUse);
1170 WideAddRec = GetExtendedOperandRecurrence(DU);
1173 // If use is a loop condition, try to promote the condition instead of
1174 // truncating the IV first.
1175 if (WidenLoopCompare(DU))
1178 // This user does not evaluate to a recurence after widening, so don't
1179 // follow it. Instead insert a Trunc to kill off the original use,
1180 // eventually isolating the original narrow IV so it can be removed.
1181 truncateIVUse(DU, DT);
1184 // Assume block terminators cannot evaluate to a recurrence. We can't to
1185 // insert a Trunc after a terminator if there happens to be a critical edge.
1186 assert(DU.NarrowUse != DU.NarrowUse->getParent()->getTerminator() &&
1187 "SCEV is not expected to evaluate a block terminator");
1189 // Reuse the IV increment that SCEVExpander created as long as it dominates
1191 Instruction *WideUse = nullptr;
1192 if (WideAddRec == WideIncExpr
1193 && Rewriter.hoistIVInc(WideInc, DU.NarrowUse))
1196 WideUse = CloneIVUser(DU);
1200 // Evaluation of WideAddRec ensured that the narrow expression could be
1201 // extended outside the loop without overflow. This suggests that the wide use
1202 // evaluates to the same expression as the extended narrow use, but doesn't
1203 // absolutely guarantee it. Hence the following failsafe check. In rare cases
1204 // where it fails, we simply throw away the newly created wide use.
1205 if (WideAddRec != SE->getSCEV(WideUse)) {
1206 DEBUG(dbgs() << "Wide use expression mismatch: " << *WideUse
1207 << ": " << *SE->getSCEV(WideUse) << " != " << *WideAddRec << "\n");
1208 DeadInsts.emplace_back(WideUse);
1212 // Returning WideUse pushes it on the worklist.
1216 /// pushNarrowIVUsers - Add eligible users of NarrowDef to NarrowIVUsers.
1218 void WidenIV::pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef) {
1219 for (User *U : NarrowDef->users()) {
1220 Instruction *NarrowUser = cast<Instruction>(U);
1222 // Handle data flow merges and bizarre phi cycles.
1223 if (!Widened.insert(NarrowUser).second)
1226 NarrowIVUsers.push_back(NarrowIVDefUse(NarrowDef, NarrowUser, WideDef));
1230 /// CreateWideIV - Process a single induction variable. First use the
1231 /// SCEVExpander to create a wide induction variable that evaluates to the same
1232 /// recurrence as the original narrow IV. Then use a worklist to forward
1233 /// traverse the narrow IV's def-use chain. After WidenIVUse has processed all
1234 /// interesting IV users, the narrow IV will be isolated for removal by
1237 /// It would be simpler to delete uses as they are processed, but we must avoid
1238 /// invalidating SCEV expressions.
1240 PHINode *WidenIV::CreateWideIV(SCEVExpander &Rewriter) {
1241 // Is this phi an induction variable?
1242 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(OrigPhi));
1246 // Widen the induction variable expression.
1247 const SCEV *WideIVExpr = IsSigned ?
1248 SE->getSignExtendExpr(AddRec, WideType) :
1249 SE->getZeroExtendExpr(AddRec, WideType);
1251 assert(SE->getEffectiveSCEVType(WideIVExpr->getType()) == WideType &&
1252 "Expect the new IV expression to preserve its type");
1254 // Can the IV be extended outside the loop without overflow?
1255 AddRec = dyn_cast<SCEVAddRecExpr>(WideIVExpr);
1256 if (!AddRec || AddRec->getLoop() != L)
1259 // An AddRec must have loop-invariant operands. Since this AddRec is
1260 // materialized by a loop header phi, the expression cannot have any post-loop
1261 // operands, so they must dominate the loop header.
1262 assert(SE->properlyDominates(AddRec->getStart(), L->getHeader()) &&
1263 SE->properlyDominates(AddRec->getStepRecurrence(*SE), L->getHeader())
1264 && "Loop header phi recurrence inputs do not dominate the loop");
1266 // The rewriter provides a value for the desired IV expression. This may
1267 // either find an existing phi or materialize a new one. Either way, we
1268 // expect a well-formed cyclic phi-with-increments. i.e. any operand not part
1269 // of the phi-SCC dominates the loop entry.
1270 Instruction *InsertPt = L->getHeader()->begin();
1271 WidePhi = cast<PHINode>(Rewriter.expandCodeFor(AddRec, WideType, InsertPt));
1273 // Remembering the WideIV increment generated by SCEVExpander allows
1274 // WidenIVUse to reuse it when widening the narrow IV's increment. We don't
1275 // employ a general reuse mechanism because the call above is the only call to
1276 // SCEVExpander. Henceforth, we produce 1-to-1 narrow to wide uses.
1277 if (BasicBlock *LatchBlock = L->getLoopLatch()) {
1279 cast<Instruction>(WidePhi->getIncomingValueForBlock(LatchBlock));
1280 WideIncExpr = SE->getSCEV(WideInc);
1283 DEBUG(dbgs() << "Wide IV: " << *WidePhi << "\n");
1286 // Traverse the def-use chain using a worklist starting at the original IV.
1287 assert(Widened.empty() && NarrowIVUsers.empty() && "expect initial state" );
1289 Widened.insert(OrigPhi);
1290 pushNarrowIVUsers(OrigPhi, WidePhi);
1292 while (!NarrowIVUsers.empty()) {
1293 NarrowIVDefUse DU = NarrowIVUsers.pop_back_val();
1295 // Process a def-use edge. This may replace the use, so don't hold a
1296 // use_iterator across it.
1297 Instruction *WideUse = WidenIVUse(DU, Rewriter);
1299 // Follow all def-use edges from the previous narrow use.
1301 pushNarrowIVUsers(DU.NarrowUse, WideUse);
1303 // WidenIVUse may have removed the def-use edge.
1304 if (DU.NarrowDef->use_empty())
1305 DeadInsts.emplace_back(DU.NarrowDef);
1310 //===----------------------------------------------------------------------===//
1311 // Live IV Reduction - Minimize IVs live across the loop.
1312 //===----------------------------------------------------------------------===//
1315 //===----------------------------------------------------------------------===//
1316 // Simplification of IV users based on SCEV evaluation.
1317 //===----------------------------------------------------------------------===//
1320 class IndVarSimplifyVisitor : public IVVisitor {
1321 ScalarEvolution *SE;
1322 const TargetTransformInfo *TTI;
1328 IndVarSimplifyVisitor(PHINode *IV, ScalarEvolution *SCEV,
1329 const TargetTransformInfo *TTI,
1330 const DominatorTree *DTree)
1331 : SE(SCEV), TTI(TTI), IVPhi(IV) {
1333 WI.NarrowIV = IVPhi;
1335 setSplitOverflowIntrinsics();
1338 // Implement the interface used by simplifyUsersOfIV.
1339 void visitCast(CastInst *Cast) override { visitIVCast(Cast, WI, SE, TTI); }
1343 /// SimplifyAndExtend - Iteratively perform simplification on a worklist of IV
1344 /// users. Each successive simplification may push more users which may
1345 /// themselves be candidates for simplification.
1347 /// Sign/Zero extend elimination is interleaved with IV simplification.
1349 void IndVarSimplify::SimplifyAndExtend(Loop *L,
1350 SCEVExpander &Rewriter,
1351 LPPassManager &LPM) {
1352 SmallVector<WideIVInfo, 8> WideIVs;
1354 SmallVector<PHINode*, 8> LoopPhis;
1355 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
1356 LoopPhis.push_back(cast<PHINode>(I));
1358 // Each round of simplification iterates through the SimplifyIVUsers worklist
1359 // for all current phis, then determines whether any IVs can be
1360 // widened. Widening adds new phis to LoopPhis, inducing another round of
1361 // simplification on the wide IVs.
1362 while (!LoopPhis.empty()) {
1363 // Evaluate as many IV expressions as possible before widening any IVs. This
1364 // forces SCEV to set no-wrap flags before evaluating sign/zero
1365 // extension. The first time SCEV attempts to normalize sign/zero extension,
1366 // the result becomes final. So for the most predictable results, we delay
1367 // evaluation of sign/zero extend evaluation until needed, and avoid running
1368 // other SCEV based analysis prior to SimplifyAndExtend.
1370 PHINode *CurrIV = LoopPhis.pop_back_val();
1372 // Information about sign/zero extensions of CurrIV.
1373 IndVarSimplifyVisitor Visitor(CurrIV, SE, TTI, DT);
1375 Changed |= simplifyUsersOfIV(CurrIV, SE, &LPM, DeadInsts, &Visitor);
1377 if (Visitor.WI.WidestNativeType) {
1378 WideIVs.push_back(Visitor.WI);
1380 } while(!LoopPhis.empty());
1382 for (; !WideIVs.empty(); WideIVs.pop_back()) {
1383 WidenIV Widener(WideIVs.back(), LI, SE, DT, DeadInsts);
1384 if (PHINode *WidePhi = Widener.CreateWideIV(Rewriter)) {
1386 LoopPhis.push_back(WidePhi);
1392 //===----------------------------------------------------------------------===//
1393 // LinearFunctionTestReplace and its kin. Rewrite the loop exit condition.
1394 //===----------------------------------------------------------------------===//
1396 /// canExpandBackedgeTakenCount - Return true if this loop's backedge taken
1397 /// count expression can be safely and cheaply expanded into an instruction
1398 /// sequence that can be used by LinearFunctionTestReplace.
1400 /// TODO: This fails for pointer-type loop counters with greater than one byte
1401 /// strides, consequently preventing LFTR from running. For the purpose of LFTR
1402 /// we could skip this check in the case that the LFTR loop counter (chosen by
1403 /// FindLoopCounter) is also pointer type. Instead, we could directly convert
1404 /// the loop test to an inequality test by checking the target data's alignment
1405 /// of element types (given that the initial pointer value originates from or is
1406 /// used by ABI constrained operation, as opposed to inttoptr/ptrtoint).
1407 /// However, we don't yet have a strong motivation for converting loop tests
1408 /// into inequality tests.
1409 static bool canExpandBackedgeTakenCount(Loop *L, ScalarEvolution *SE,
1410 SCEVExpander &Rewriter) {
1411 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
1412 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount) ||
1413 BackedgeTakenCount->isZero())
1416 if (!L->getExitingBlock())
1419 // Can't rewrite non-branch yet.
1420 if (!isa<BranchInst>(L->getExitingBlock()->getTerminator()))
1423 if (Rewriter.isHighCostExpansion(BackedgeTakenCount, L))
1429 /// getLoopPhiForCounter - Return the loop header phi IFF IncV adds a loop
1430 /// invariant value to the phi.
1431 static PHINode *getLoopPhiForCounter(Value *IncV, Loop *L, DominatorTree *DT) {
1432 Instruction *IncI = dyn_cast<Instruction>(IncV);
1436 switch (IncI->getOpcode()) {
1437 case Instruction::Add:
1438 case Instruction::Sub:
1440 case Instruction::GetElementPtr:
1441 // An IV counter must preserve its type.
1442 if (IncI->getNumOperands() == 2)
1448 PHINode *Phi = dyn_cast<PHINode>(IncI->getOperand(0));
1449 if (Phi && Phi->getParent() == L->getHeader()) {
1450 if (isLoopInvariant(IncI->getOperand(1), L, DT))
1454 if (IncI->getOpcode() == Instruction::GetElementPtr)
1457 // Allow add/sub to be commuted.
1458 Phi = dyn_cast<PHINode>(IncI->getOperand(1));
1459 if (Phi && Phi->getParent() == L->getHeader()) {
1460 if (isLoopInvariant(IncI->getOperand(0), L, DT))
1466 /// Return the compare guarding the loop latch, or NULL for unrecognized tests.
1467 static ICmpInst *getLoopTest(Loop *L) {
1468 assert(L->getExitingBlock() && "expected loop exit");
1470 BasicBlock *LatchBlock = L->getLoopLatch();
1471 // Don't bother with LFTR if the loop is not properly simplified.
1475 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
1476 assert(BI && "expected exit branch");
1478 return dyn_cast<ICmpInst>(BI->getCondition());
1481 /// needsLFTR - LinearFunctionTestReplace policy. Return true unless we can show
1482 /// that the current exit test is already sufficiently canonical.
1483 static bool needsLFTR(Loop *L, DominatorTree *DT) {
1484 // Do LFTR to simplify the exit condition to an ICMP.
1485 ICmpInst *Cond = getLoopTest(L);
1489 // Do LFTR to simplify the exit ICMP to EQ/NE
1490 ICmpInst::Predicate Pred = Cond->getPredicate();
1491 if (Pred != ICmpInst::ICMP_NE && Pred != ICmpInst::ICMP_EQ)
1494 // Look for a loop invariant RHS
1495 Value *LHS = Cond->getOperand(0);
1496 Value *RHS = Cond->getOperand(1);
1497 if (!isLoopInvariant(RHS, L, DT)) {
1498 if (!isLoopInvariant(LHS, L, DT))
1500 std::swap(LHS, RHS);
1502 // Look for a simple IV counter LHS
1503 PHINode *Phi = dyn_cast<PHINode>(LHS);
1505 Phi = getLoopPhiForCounter(LHS, L, DT);
1510 // Do LFTR if PHI node is defined in the loop, but is *not* a counter.
1511 int Idx = Phi->getBasicBlockIndex(L->getLoopLatch());
1515 // Do LFTR if the exit condition's IV is *not* a simple counter.
1516 Value *IncV = Phi->getIncomingValue(Idx);
1517 return Phi != getLoopPhiForCounter(IncV, L, DT);
1520 /// Recursive helper for hasConcreteDef(). Unfortunately, this currently boils
1521 /// down to checking that all operands are constant and listing instructions
1522 /// that may hide undef.
1523 static bool hasConcreteDefImpl(Value *V, SmallPtrSetImpl<Value*> &Visited,
1525 if (isa<Constant>(V))
1526 return !isa<UndefValue>(V);
1531 // Conservatively handle non-constant non-instructions. For example, Arguments
1533 Instruction *I = dyn_cast<Instruction>(V);
1537 // Load and return values may be undef.
1538 if(I->mayReadFromMemory() || isa<CallInst>(I) || isa<InvokeInst>(I))
1541 // Optimistically handle other instructions.
1542 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) {
1543 if (!Visited.insert(*OI).second)
1545 if (!hasConcreteDefImpl(*OI, Visited, Depth+1))
1551 /// Return true if the given value is concrete. We must prove that undef can
1554 /// TODO: If we decide that this is a good approach to checking for undef, we
1555 /// may factor it into a common location.
1556 static bool hasConcreteDef(Value *V) {
1557 SmallPtrSet<Value*, 8> Visited;
1559 return hasConcreteDefImpl(V, Visited, 0);
1562 /// AlmostDeadIV - Return true if this IV has any uses other than the (soon to
1563 /// be rewritten) loop exit test.
1564 static bool AlmostDeadIV(PHINode *Phi, BasicBlock *LatchBlock, Value *Cond) {
1565 int LatchIdx = Phi->getBasicBlockIndex(LatchBlock);
1566 Value *IncV = Phi->getIncomingValue(LatchIdx);
1568 for (User *U : Phi->users())
1569 if (U != Cond && U != IncV) return false;
1571 for (User *U : IncV->users())
1572 if (U != Cond && U != Phi) return false;
1576 /// FindLoopCounter - Find an affine IV in canonical form.
1578 /// BECount may be an i8* pointer type. The pointer difference is already
1579 /// valid count without scaling the address stride, so it remains a pointer
1580 /// expression as far as SCEV is concerned.
1582 /// Currently only valid for LFTR. See the comments on hasConcreteDef below.
1584 /// FIXME: Accept -1 stride and set IVLimit = IVInit - BECount
1586 /// FIXME: Accept non-unit stride as long as SCEV can reduce BECount * Stride.
1587 /// This is difficult in general for SCEV because of potential overflow. But we
1588 /// could at least handle constant BECounts.
1589 static PHINode *FindLoopCounter(Loop *L, const SCEV *BECount,
1590 ScalarEvolution *SE, DominatorTree *DT) {
1591 uint64_t BCWidth = SE->getTypeSizeInBits(BECount->getType());
1594 cast<BranchInst>(L->getExitingBlock()->getTerminator())->getCondition();
1596 // Loop over all of the PHI nodes, looking for a simple counter.
1597 PHINode *BestPhi = nullptr;
1598 const SCEV *BestInit = nullptr;
1599 BasicBlock *LatchBlock = L->getLoopLatch();
1600 assert(LatchBlock && "needsLFTR should guarantee a loop latch");
1602 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
1603 PHINode *Phi = cast<PHINode>(I);
1604 if (!SE->isSCEVable(Phi->getType()))
1607 // Avoid comparing an integer IV against a pointer Limit.
1608 if (BECount->getType()->isPointerTy() && !Phi->getType()->isPointerTy())
1611 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Phi));
1612 if (!AR || AR->getLoop() != L || !AR->isAffine())
1615 // AR may be a pointer type, while BECount is an integer type.
1616 // AR may be wider than BECount. With eq/ne tests overflow is immaterial.
1617 // AR may not be a narrower type, or we may never exit.
1618 uint64_t PhiWidth = SE->getTypeSizeInBits(AR->getType());
1619 if (PhiWidth < BCWidth ||
1620 !L->getHeader()->getModule()->getDataLayout().isLegalInteger(PhiWidth))
1623 const SCEV *Step = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE));
1624 if (!Step || !Step->isOne())
1627 int LatchIdx = Phi->getBasicBlockIndex(LatchBlock);
1628 Value *IncV = Phi->getIncomingValue(LatchIdx);
1629 if (getLoopPhiForCounter(IncV, L, DT) != Phi)
1632 // Avoid reusing a potentially undef value to compute other values that may
1633 // have originally had a concrete definition.
1634 if (!hasConcreteDef(Phi)) {
1635 // We explicitly allow unknown phis as long as they are already used by
1636 // the loop test. In this case we assume that performing LFTR could not
1637 // increase the number of undef users.
1638 if (ICmpInst *Cond = getLoopTest(L)) {
1639 if (Phi != getLoopPhiForCounter(Cond->getOperand(0), L, DT)
1640 && Phi != getLoopPhiForCounter(Cond->getOperand(1), L, DT)) {
1645 const SCEV *Init = AR->getStart();
1647 if (BestPhi && !AlmostDeadIV(BestPhi, LatchBlock, Cond)) {
1648 // Don't force a live loop counter if another IV can be used.
1649 if (AlmostDeadIV(Phi, LatchBlock, Cond))
1652 // Prefer to count-from-zero. This is a more "canonical" counter form. It
1653 // also prefers integer to pointer IVs.
1654 if (BestInit->isZero() != Init->isZero()) {
1655 if (BestInit->isZero())
1658 // If two IVs both count from zero or both count from nonzero then the
1659 // narrower is likely a dead phi that has been widened. Use the wider phi
1660 // to allow the other to be eliminated.
1661 else if (PhiWidth <= SE->getTypeSizeInBits(BestPhi->getType()))
1670 /// genLoopLimit - Help LinearFunctionTestReplace by generating a value that
1671 /// holds the RHS of the new loop test.
1672 static Value *genLoopLimit(PHINode *IndVar, const SCEV *IVCount, Loop *L,
1673 SCEVExpander &Rewriter, ScalarEvolution *SE) {
1674 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(IndVar));
1675 assert(AR && AR->getLoop() == L && AR->isAffine() && "bad loop counter");
1676 const SCEV *IVInit = AR->getStart();
1678 // IVInit may be a pointer while IVCount is an integer when FindLoopCounter
1679 // finds a valid pointer IV. Sign extend BECount in order to materialize a
1680 // GEP. Avoid running SCEVExpander on a new pointer value, instead reusing
1681 // the existing GEPs whenever possible.
1682 if (IndVar->getType()->isPointerTy()
1683 && !IVCount->getType()->isPointerTy()) {
1685 // IVOffset will be the new GEP offset that is interpreted by GEP as a
1686 // signed value. IVCount on the other hand represents the loop trip count,
1687 // which is an unsigned value. FindLoopCounter only allows induction
1688 // variables that have a positive unit stride of one. This means we don't
1689 // have to handle the case of negative offsets (yet) and just need to zero
1691 Type *OfsTy = SE->getEffectiveSCEVType(IVInit->getType());
1692 const SCEV *IVOffset = SE->getTruncateOrZeroExtend(IVCount, OfsTy);
1694 // Expand the code for the iteration count.
1695 assert(SE->isLoopInvariant(IVOffset, L) &&
1696 "Computed iteration count is not loop invariant!");
1697 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1698 Value *GEPOffset = Rewriter.expandCodeFor(IVOffset, OfsTy, BI);
1700 Value *GEPBase = IndVar->getIncomingValueForBlock(L->getLoopPreheader());
1701 assert(AR->getStart() == SE->getSCEV(GEPBase) && "bad loop counter");
1702 // We could handle pointer IVs other than i8*, but we need to compensate for
1703 // gep index scaling. See canExpandBackedgeTakenCount comments.
1704 assert(SE->getSizeOfExpr(IntegerType::getInt64Ty(IndVar->getContext()),
1705 cast<PointerType>(GEPBase->getType())->getElementType())->isOne()
1706 && "unit stride pointer IV must be i8*");
1708 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
1709 return Builder.CreateGEP(nullptr, GEPBase, GEPOffset, "lftr.limit");
1712 // In any other case, convert both IVInit and IVCount to integers before
1713 // comparing. This may result in SCEV expension of pointers, but in practice
1714 // SCEV will fold the pointer arithmetic away as such:
1715 // BECount = (IVEnd - IVInit - 1) => IVLimit = IVInit (postinc).
1717 // Valid Cases: (1) both integers is most common; (2) both may be pointers
1718 // for simple memset-style loops.
1720 // IVInit integer and IVCount pointer would only occur if a canonical IV
1721 // were generated on top of case #2, which is not expected.
1723 const SCEV *IVLimit = nullptr;
1724 // For unit stride, IVCount = Start + BECount with 2's complement overflow.
1725 // For non-zero Start, compute IVCount here.
1726 if (AR->getStart()->isZero())
1729 assert(AR->getStepRecurrence(*SE)->isOne() && "only handles unit stride");
1730 const SCEV *IVInit = AR->getStart();
1732 // For integer IVs, truncate the IV before computing IVInit + BECount.
1733 if (SE->getTypeSizeInBits(IVInit->getType())
1734 > SE->getTypeSizeInBits(IVCount->getType()))
1735 IVInit = SE->getTruncateExpr(IVInit, IVCount->getType());
1737 IVLimit = SE->getAddExpr(IVInit, IVCount);
1739 // Expand the code for the iteration count.
1740 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1741 IRBuilder<> Builder(BI);
1742 assert(SE->isLoopInvariant(IVLimit, L) &&
1743 "Computed iteration count is not loop invariant!");
1744 // Ensure that we generate the same type as IndVar, or a smaller integer
1745 // type. In the presence of null pointer values, we have an integer type
1746 // SCEV expression (IVInit) for a pointer type IV value (IndVar).
1747 Type *LimitTy = IVCount->getType()->isPointerTy() ?
1748 IndVar->getType() : IVCount->getType();
1749 return Rewriter.expandCodeFor(IVLimit, LimitTy, BI);
1753 /// LinearFunctionTestReplace - This method rewrites the exit condition of the
1754 /// loop to be a canonical != comparison against the incremented loop induction
1755 /// variable. This pass is able to rewrite the exit tests of any loop where the
1756 /// SCEV analysis can determine a loop-invariant trip count of the loop, which
1757 /// is actually a much broader range than just linear tests.
1758 Value *IndVarSimplify::
1759 LinearFunctionTestReplace(Loop *L,
1760 const SCEV *BackedgeTakenCount,
1762 SCEVExpander &Rewriter) {
1763 assert(canExpandBackedgeTakenCount(L, SE, Rewriter) && "precondition");
1765 // Initialize CmpIndVar and IVCount to their preincremented values.
1766 Value *CmpIndVar = IndVar;
1767 const SCEV *IVCount = BackedgeTakenCount;
1769 // If the exiting block is the same as the backedge block, we prefer to
1770 // compare against the post-incremented value, otherwise we must compare
1771 // against the preincremented value.
1772 if (L->getExitingBlock() == L->getLoopLatch()) {
1773 // Add one to the "backedge-taken" count to get the trip count.
1774 // This addition may overflow, which is valid as long as the comparison is
1775 // truncated to BackedgeTakenCount->getType().
1776 IVCount = SE->getAddExpr(BackedgeTakenCount,
1777 SE->getConstant(BackedgeTakenCount->getType(), 1));
1778 // The BackedgeTaken expression contains the number of times that the
1779 // backedge branches to the loop header. This is one less than the
1780 // number of times the loop executes, so use the incremented indvar.
1781 CmpIndVar = IndVar->getIncomingValueForBlock(L->getExitingBlock());
1784 Value *ExitCnt = genLoopLimit(IndVar, IVCount, L, Rewriter, SE);
1785 assert(ExitCnt->getType()->isPointerTy() == IndVar->getType()->isPointerTy()
1786 && "genLoopLimit missed a cast");
1788 // Insert a new icmp_ne or icmp_eq instruction before the branch.
1789 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1790 ICmpInst::Predicate P;
1791 if (L->contains(BI->getSuccessor(0)))
1792 P = ICmpInst::ICMP_NE;
1794 P = ICmpInst::ICMP_EQ;
1796 DEBUG(dbgs() << "INDVARS: Rewriting loop exit condition to:\n"
1797 << " LHS:" << *CmpIndVar << '\n'
1799 << (P == ICmpInst::ICMP_NE ? "!=" : "==") << "\n"
1800 << " RHS:\t" << *ExitCnt << "\n"
1801 << " IVCount:\t" << *IVCount << "\n");
1803 IRBuilder<> Builder(BI);
1805 // LFTR can ignore IV overflow and truncate to the width of
1806 // BECount. This avoids materializing the add(zext(add)) expression.
1807 unsigned CmpIndVarSize = SE->getTypeSizeInBits(CmpIndVar->getType());
1808 unsigned ExitCntSize = SE->getTypeSizeInBits(ExitCnt->getType());
1809 if (CmpIndVarSize > ExitCntSize) {
1810 const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(SE->getSCEV(IndVar));
1811 const SCEV *ARStart = AR->getStart();
1812 const SCEV *ARStep = AR->getStepRecurrence(*SE);
1813 // For constant IVCount, avoid truncation.
1814 if (isa<SCEVConstant>(ARStart) && isa<SCEVConstant>(IVCount)) {
1815 const APInt &Start = cast<SCEVConstant>(ARStart)->getValue()->getValue();
1816 APInt Count = cast<SCEVConstant>(IVCount)->getValue()->getValue();
1817 // Note that the post-inc value of BackedgeTakenCount may have overflowed
1818 // above such that IVCount is now zero.
1819 if (IVCount != BackedgeTakenCount && Count == 0) {
1820 Count = APInt::getMaxValue(Count.getBitWidth()).zext(CmpIndVarSize);
1824 Count = Count.zext(CmpIndVarSize);
1826 if (cast<SCEVConstant>(ARStep)->getValue()->isNegative())
1827 NewLimit = Start - Count;
1829 NewLimit = Start + Count;
1830 ExitCnt = ConstantInt::get(CmpIndVar->getType(), NewLimit);
1832 DEBUG(dbgs() << " Widen RHS:\t" << *ExitCnt << "\n");
1834 CmpIndVar = Builder.CreateTrunc(CmpIndVar, ExitCnt->getType(),
1838 Value *Cond = Builder.CreateICmp(P, CmpIndVar, ExitCnt, "exitcond");
1839 Value *OrigCond = BI->getCondition();
1840 // It's tempting to use replaceAllUsesWith here to fully replace the old
1841 // comparison, but that's not immediately safe, since users of the old
1842 // comparison may not be dominated by the new comparison. Instead, just
1843 // update the branch to use the new comparison; in the common case this
1844 // will make old comparison dead.
1845 BI->setCondition(Cond);
1846 DeadInsts.push_back(OrigCond);
1853 //===----------------------------------------------------------------------===//
1854 // SinkUnusedInvariants. A late subpass to cleanup loop preheaders.
1855 //===----------------------------------------------------------------------===//
1857 /// If there's a single exit block, sink any loop-invariant values that
1858 /// were defined in the preheader but not used inside the loop into the
1859 /// exit block to reduce register pressure in the loop.
1860 void IndVarSimplify::SinkUnusedInvariants(Loop *L) {
1861 BasicBlock *ExitBlock = L->getExitBlock();
1862 if (!ExitBlock) return;
1864 BasicBlock *Preheader = L->getLoopPreheader();
1865 if (!Preheader) return;
1867 Instruction *InsertPt = ExitBlock->getFirstInsertionPt();
1868 BasicBlock::iterator I = Preheader->getTerminator();
1869 while (I != Preheader->begin()) {
1871 // New instructions were inserted at the end of the preheader.
1872 if (isa<PHINode>(I))
1875 // Don't move instructions which might have side effects, since the side
1876 // effects need to complete before instructions inside the loop. Also don't
1877 // move instructions which might read memory, since the loop may modify
1878 // memory. Note that it's okay if the instruction might have undefined
1879 // behavior: LoopSimplify guarantees that the preheader dominates the exit
1881 if (I->mayHaveSideEffects() || I->mayReadFromMemory())
1884 // Skip debug info intrinsics.
1885 if (isa<DbgInfoIntrinsic>(I))
1888 // Skip landingpad instructions.
1889 if (isa<LandingPadInst>(I))
1892 // Don't sink alloca: we never want to sink static alloca's out of the
1893 // entry block, and correctly sinking dynamic alloca's requires
1894 // checks for stacksave/stackrestore intrinsics.
1895 // FIXME: Refactor this check somehow?
1896 if (isa<AllocaInst>(I))
1899 // Determine if there is a use in or before the loop (direct or
1901 bool UsedInLoop = false;
1902 for (Use &U : I->uses()) {
1903 Instruction *User = cast<Instruction>(U.getUser());
1904 BasicBlock *UseBB = User->getParent();
1905 if (PHINode *P = dyn_cast<PHINode>(User)) {
1907 PHINode::getIncomingValueNumForOperand(U.getOperandNo());
1908 UseBB = P->getIncomingBlock(i);
1910 if (UseBB == Preheader || L->contains(UseBB)) {
1916 // If there is, the def must remain in the preheader.
1920 // Otherwise, sink it to the exit block.
1921 Instruction *ToMove = I;
1924 if (I != Preheader->begin()) {
1925 // Skip debug info intrinsics.
1928 } while (isa<DbgInfoIntrinsic>(I) && I != Preheader->begin());
1930 if (isa<DbgInfoIntrinsic>(I) && I == Preheader->begin())
1936 ToMove->moveBefore(InsertPt);
1942 //===----------------------------------------------------------------------===//
1943 // IndVarSimplify driver. Manage several subpasses of IV simplification.
1944 //===----------------------------------------------------------------------===//
1946 bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) {
1947 if (skipOptnoneFunction(L))
1950 // If LoopSimplify form is not available, stay out of trouble. Some notes:
1951 // - LSR currently only supports LoopSimplify-form loops. Indvars'
1952 // canonicalization can be a pessimization without LSR to "clean up"
1954 // - We depend on having a preheader; in particular,
1955 // Loop::getCanonicalInductionVariable only supports loops with preheaders,
1956 // and we're in trouble if we can't find the induction variable even when
1957 // we've manually inserted one.
1958 if (!L->isLoopSimplifyForm())
1961 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1962 SE = &getAnalysis<ScalarEvolution>();
1963 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1964 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1965 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1966 auto *TTIP = getAnalysisIfAvailable<TargetTransformInfoWrapperPass>();
1967 TTI = TTIP ? &TTIP->getTTI(*L->getHeader()->getParent()) : nullptr;
1968 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
1973 // If there are any floating-point recurrences, attempt to
1974 // transform them to use integer recurrences.
1975 RewriteNonIntegerIVs(L);
1977 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
1979 // Create a rewriter object which we'll use to transform the code with.
1980 SCEVExpander Rewriter(*SE, DL, "indvars");
1982 Rewriter.setDebugType(DEBUG_TYPE);
1985 // Eliminate redundant IV users.
1987 // Simplification works best when run before other consumers of SCEV. We
1988 // attempt to avoid evaluating SCEVs for sign/zero extend operations until
1989 // other expressions involving loop IVs have been evaluated. This helps SCEV
1990 // set no-wrap flags before normalizing sign/zero extension.
1991 Rewriter.disableCanonicalMode();
1992 SimplifyAndExtend(L, Rewriter, LPM);
1994 // Check to see if this loop has a computable loop-invariant execution count.
1995 // If so, this means that we can compute the final value of any expressions
1996 // that are recurrent in the loop, and substitute the exit values from the
1997 // loop into any instructions outside of the loop that use the final values of
1998 // the current expressions.
2000 if (ReplaceExitValue != NeverRepl &&
2001 !isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2002 RewriteLoopExitValues(L, Rewriter);
2004 // Eliminate redundant IV cycles.
2005 NumElimIV += Rewriter.replaceCongruentIVs(L, DT, DeadInsts);
2007 // If we have a trip count expression, rewrite the loop's exit condition
2008 // using it. We can currently only handle loops with a single exit.
2009 if (canExpandBackedgeTakenCount(L, SE, Rewriter) && needsLFTR(L, DT)) {
2010 PHINode *IndVar = FindLoopCounter(L, BackedgeTakenCount, SE, DT);
2012 // Check preconditions for proper SCEVExpander operation. SCEV does not
2013 // express SCEVExpander's dependencies, such as LoopSimplify. Instead any
2014 // pass that uses the SCEVExpander must do it. This does not work well for
2015 // loop passes because SCEVExpander makes assumptions about all loops,
2016 // while LoopPassManager only forces the current loop to be simplified.
2018 // FIXME: SCEV expansion has no way to bail out, so the caller must
2019 // explicitly check any assumptions made by SCEV. Brittle.
2020 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(BackedgeTakenCount);
2021 if (!AR || AR->getLoop()->getLoopPreheader())
2022 (void)LinearFunctionTestReplace(L, BackedgeTakenCount, IndVar,
2026 // Clear the rewriter cache, because values that are in the rewriter's cache
2027 // can be deleted in the loop below, causing the AssertingVH in the cache to
2031 // Now that we're done iterating through lists, clean up any instructions
2032 // which are now dead.
2033 while (!DeadInsts.empty())
2034 if (Instruction *Inst =
2035 dyn_cast_or_null<Instruction>(DeadInsts.pop_back_val()))
2036 RecursivelyDeleteTriviallyDeadInstructions(Inst, TLI);
2038 // The Rewriter may not be used from this point on.
2040 // Loop-invariant instructions in the preheader that aren't used in the
2041 // loop may be sunk below the loop to reduce register pressure.
2042 SinkUnusedInvariants(L);
2044 // Clean up dead instructions.
2045 Changed |= DeleteDeadPHIs(L->getHeader(), TLI);
2046 // Check a post-condition.
2047 assert(L->isLCSSAForm(*DT) &&
2048 "Indvars did not leave the loop in lcssa form!");
2050 // Verify that LFTR, and any other change have not interfered with SCEV's
2051 // ability to compute trip count.
2053 if (VerifyIndvars && !isa<SCEVCouldNotCompute>(BackedgeTakenCount)) {
2055 const SCEV *NewBECount = SE->getBackedgeTakenCount(L);
2056 if (SE->getTypeSizeInBits(BackedgeTakenCount->getType()) <
2057 SE->getTypeSizeInBits(NewBECount->getType()))
2058 NewBECount = SE->getTruncateOrNoop(NewBECount,
2059 BackedgeTakenCount->getType());
2061 BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount,
2062 NewBECount->getType());
2063 assert(BackedgeTakenCount == NewBECount && "indvars must preserve SCEV");