1 //===- IndVarSimplify.cpp - Induction Variable Elimination ----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This transformation analyzes and transforms the induction variables (and
11 // computations derived from them) into simpler forms suitable for subsequent
12 // analysis and transformation.
14 // If the trip count of a loop is computable, this pass also makes the following
16 // 1. The exit condition for the loop is canonicalized to compare the
17 // induction value against the exit value. This turns loops like:
18 // 'for (i = 7; i*i < 1000; ++i)' into 'for (i = 0; i != 25; ++i)'
19 // 2. Any use outside of the loop of an expression derived from the indvar
20 // is changed to compute the derived value outside of the loop, eliminating
21 // the dependence on the exit value of the induction variable. If the only
22 // purpose of the loop is to compute the exit value of some derived
23 // expression, this transformation will make the loop dead.
25 //===----------------------------------------------------------------------===//
27 #include "llvm/Transforms/Scalar.h"
28 #include "llvm/ADT/DenseMap.h"
29 #include "llvm/ADT/SmallVector.h"
30 #include "llvm/ADT/Statistic.h"
31 #include "llvm/Analysis/GlobalsModRef.h"
32 #include "llvm/Analysis/LoopInfo.h"
33 #include "llvm/Analysis/LoopPass.h"
34 #include "llvm/Analysis/ScalarEvolutionExpander.h"
35 #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h"
36 #include "llvm/Analysis/TargetLibraryInfo.h"
37 #include "llvm/Analysis/TargetTransformInfo.h"
38 #include "llvm/IR/BasicBlock.h"
39 #include "llvm/IR/CFG.h"
40 #include "llvm/IR/Constants.h"
41 #include "llvm/IR/DataLayout.h"
42 #include "llvm/IR/Dominators.h"
43 #include "llvm/IR/Instructions.h"
44 #include "llvm/IR/IntrinsicInst.h"
45 #include "llvm/IR/LLVMContext.h"
46 #include "llvm/IR/PatternMatch.h"
47 #include "llvm/IR/Type.h"
48 #include "llvm/Support/CommandLine.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/raw_ostream.h"
51 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
52 #include "llvm/Transforms/Utils/Local.h"
53 #include "llvm/Transforms/Utils/SimplifyIndVar.h"
56 #define DEBUG_TYPE "indvars"
58 STATISTIC(NumWidened , "Number of indvars widened");
59 STATISTIC(NumReplaced , "Number of exit values replaced");
60 STATISTIC(NumLFTR , "Number of loop exit tests replaced");
61 STATISTIC(NumElimExt , "Number of IV sign/zero extends eliminated");
62 STATISTIC(NumElimIV , "Number of congruent IVs eliminated");
64 // Trip count verification can be enabled by default under NDEBUG if we
65 // implement a strong expression equivalence checker in SCEV. Until then, we
66 // use the verify-indvars flag, which may assert in some cases.
67 static cl::opt<bool> VerifyIndvars(
68 "verify-indvars", cl::Hidden,
69 cl::desc("Verify the ScalarEvolution result after running indvars"));
71 static cl::opt<bool> ReduceLiveIVs("liv-reduce", cl::Hidden,
72 cl::desc("Reduce live induction variables."));
74 enum ReplaceExitVal { NeverRepl, OnlyCheapRepl, AlwaysRepl };
76 static cl::opt<ReplaceExitVal> ReplaceExitValue(
77 "replexitval", cl::Hidden, cl::init(OnlyCheapRepl),
78 cl::desc("Choose the strategy to replace exit value in IndVarSimplify"),
79 cl::values(clEnumValN(NeverRepl, "never", "never replace exit value"),
80 clEnumValN(OnlyCheapRepl, "cheap",
81 "only replace exit value when the cost is cheap"),
82 clEnumValN(AlwaysRepl, "always",
83 "always replace exit value whenever possible"),
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<ScalarEvolutionWrapperPass>();
112 AU.addRequiredID(LoopSimplifyID);
113 AU.addRequiredID(LCSSAID);
114 AU.addPreserved<GlobalsAAWrapperPass>();
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 /// Return true if the SCEV expansion generated by the rewriter can replace the
162 /// original value. SCEV guarantees that it produces the same value, but the way
163 /// it is produced may be illegal IR. Ideally, this function will only be
164 /// 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 (auto *GEP = dyn_cast<GEPOperator>(FromVal)) {
179 FromPtr = GEP->getPointerOperand();
181 if (auto *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 /// 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 /// If the loop has floating induction variable then insert corresponding
260 /// 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 auto *InitValueVal = 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 auto *Incr = dyn_cast<BinaryOperator>(PN->getIncomingValue(BackEdge));
282 if (Incr == nullptr || Incr->getOpcode() != Instruction::FAdd) return;
284 // If this is not an add of the PHI with a constantfp, or if the constant fp
285 // is not an integer, bail out.
286 ConstantFP *IncValueVal = dyn_cast<ConstantFP>(Incr->getOperand(1));
288 if (IncValueVal == nullptr || Incr->getOperand(0) != PN ||
289 !ConvertToSInt(IncValueVal->getValueAPF(), IncValue))
292 // Check Incr uses. One user is PN and the other user is an exit condition
293 // used by the conditional terminator.
294 Value::user_iterator IncrUse = Incr->user_begin();
295 Instruction *U1 = cast<Instruction>(*IncrUse++);
296 if (IncrUse == Incr->user_end()) return;
297 Instruction *U2 = cast<Instruction>(*IncrUse++);
298 if (IncrUse != Incr->user_end()) return;
300 // Find exit condition, which is an fcmp. If it doesn't exist, or if it isn't
301 // only used by a branch, we can't transform it.
302 FCmpInst *Compare = dyn_cast<FCmpInst>(U1);
304 Compare = dyn_cast<FCmpInst>(U2);
305 if (!Compare || !Compare->hasOneUse() ||
306 !isa<BranchInst>(Compare->user_back()))
309 BranchInst *TheBr = cast<BranchInst>(Compare->user_back());
311 // We need to verify that the branch actually controls the iteration count
312 // of the loop. If not, the new IV can overflow and no one will notice.
313 // The branch block must be in the loop and one of the successors must be out
315 assert(TheBr->isConditional() && "Can't use fcmp if not conditional");
316 if (!L->contains(TheBr->getParent()) ||
317 (L->contains(TheBr->getSuccessor(0)) &&
318 L->contains(TheBr->getSuccessor(1))))
322 // If it isn't a comparison with an integer-as-fp (the exit value), we can't
324 ConstantFP *ExitValueVal = dyn_cast<ConstantFP>(Compare->getOperand(1));
326 if (ExitValueVal == nullptr ||
327 !ConvertToSInt(ExitValueVal->getValueAPF(), ExitValue))
330 // Find new predicate for integer comparison.
331 CmpInst::Predicate NewPred = CmpInst::BAD_ICMP_PREDICATE;
332 switch (Compare->getPredicate()) {
333 default: return; // Unknown comparison.
334 case CmpInst::FCMP_OEQ:
335 case CmpInst::FCMP_UEQ: NewPred = CmpInst::ICMP_EQ; break;
336 case CmpInst::FCMP_ONE:
337 case CmpInst::FCMP_UNE: NewPred = CmpInst::ICMP_NE; break;
338 case CmpInst::FCMP_OGT:
339 case CmpInst::FCMP_UGT: NewPred = CmpInst::ICMP_SGT; break;
340 case CmpInst::FCMP_OGE:
341 case CmpInst::FCMP_UGE: NewPred = CmpInst::ICMP_SGE; break;
342 case CmpInst::FCMP_OLT:
343 case CmpInst::FCMP_ULT: NewPred = CmpInst::ICMP_SLT; break;
344 case CmpInst::FCMP_OLE:
345 case CmpInst::FCMP_ULE: NewPred = CmpInst::ICMP_SLE; break;
348 // We convert the floating point induction variable to a signed i32 value if
349 // we can. This is only safe if the comparison will not overflow in a way
350 // that won't be trapped by the integer equivalent operations. Check for this
352 // TODO: We could use i64 if it is native and the range requires it.
354 // The start/stride/exit values must all fit in signed i32.
355 if (!isInt<32>(InitValue) || !isInt<32>(IncValue) || !isInt<32>(ExitValue))
358 // If not actually striding (add x, 0.0), avoid touching the code.
362 // Positive and negative strides have different safety conditions.
364 // If we have a positive stride, we require the init to be less than the
366 if (InitValue >= ExitValue)
369 uint32_t Range = uint32_t(ExitValue-InitValue);
370 // Check for infinite loop, either:
371 // while (i <= Exit) or until (i > Exit)
372 if (NewPred == CmpInst::ICMP_SLE || NewPred == CmpInst::ICMP_SGT) {
373 if (++Range == 0) return; // Range overflows.
376 unsigned Leftover = Range % uint32_t(IncValue);
378 // If this is an equality comparison, we require that the strided value
379 // exactly land on the exit value, otherwise the IV condition will wrap
380 // around and do things the fp IV wouldn't.
381 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
385 // If the stride would wrap around the i32 before exiting, we can't
387 if (Leftover != 0 && int32_t(ExitValue+IncValue) < ExitValue)
391 // If we have a negative stride, we require the init to be greater than the
393 if (InitValue <= ExitValue)
396 uint32_t Range = uint32_t(InitValue-ExitValue);
397 // Check for infinite loop, either:
398 // while (i >= Exit) or until (i < Exit)
399 if (NewPred == CmpInst::ICMP_SGE || NewPred == CmpInst::ICMP_SLT) {
400 if (++Range == 0) return; // Range overflows.
403 unsigned Leftover = Range % uint32_t(-IncValue);
405 // If this is an equality comparison, we require that the strided value
406 // exactly land on the exit value, otherwise the IV condition will wrap
407 // around and do things the fp IV wouldn't.
408 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
412 // If the stride would wrap around the i32 before exiting, we can't
414 if (Leftover != 0 && int32_t(ExitValue+IncValue) > ExitValue)
418 IntegerType *Int32Ty = Type::getInt32Ty(PN->getContext());
420 // Insert new integer induction variable.
421 PHINode *NewPHI = PHINode::Create(Int32Ty, 2, PN->getName()+".int", PN);
422 NewPHI->addIncoming(ConstantInt::get(Int32Ty, InitValue),
423 PN->getIncomingBlock(IncomingEdge));
426 BinaryOperator::CreateAdd(NewPHI, ConstantInt::get(Int32Ty, IncValue),
427 Incr->getName()+".int", Incr);
428 NewPHI->addIncoming(NewAdd, PN->getIncomingBlock(BackEdge));
430 ICmpInst *NewCompare = new ICmpInst(TheBr, NewPred, NewAdd,
431 ConstantInt::get(Int32Ty, ExitValue),
434 // In the following deletions, PN may become dead and may be deleted.
435 // Use a WeakVH to observe whether this happens.
438 // Delete the old floating point exit comparison. The branch starts using the
440 NewCompare->takeName(Compare);
441 Compare->replaceAllUsesWith(NewCompare);
442 RecursivelyDeleteTriviallyDeadInstructions(Compare, TLI);
444 // Delete the old floating point increment.
445 Incr->replaceAllUsesWith(UndefValue::get(Incr->getType()));
446 RecursivelyDeleteTriviallyDeadInstructions(Incr, TLI);
448 // If the FP induction variable still has uses, this is because something else
449 // in the loop uses its value. In order to canonicalize the induction
450 // variable, we chose to eliminate the IV and rewrite it in terms of an
453 // We give preference to sitofp over uitofp because it is faster on most
456 Value *Conv = new SIToFPInst(NewPHI, PN->getType(), "indvar.conv",
457 &*PN->getParent()->getFirstInsertionPt());
458 PN->replaceAllUsesWith(Conv);
459 RecursivelyDeleteTriviallyDeadInstructions(PN, TLI);
464 void IndVarSimplify::rewriteNonIntegerIVs(Loop *L) {
465 // First step. Check to see if there are any floating-point recurrences.
466 // If there are, change them into integer recurrences, permitting analysis by
467 // the SCEV routines.
469 BasicBlock *Header = L->getHeader();
471 SmallVector<WeakVH, 8> PHIs;
472 for (BasicBlock::iterator I = Header->begin();
473 PHINode *PN = dyn_cast<PHINode>(I); ++I)
476 for (unsigned i = 0, e = PHIs.size(); i != e; ++i)
477 if (PHINode *PN = dyn_cast_or_null<PHINode>(&*PHIs[i]))
478 handleFloatingPointIV(L, PN);
480 // If the loop previously had floating-point IV, ScalarEvolution
481 // may not have been able to compute a trip count. Now that we've done some
482 // re-writing, the trip count may be computable.
488 // Collect information about PHI nodes which can be transformed in
489 // rewriteLoopExitValues.
492 unsigned Ith; // Ith incoming value.
493 Value *Val; // Exit value after expansion.
494 bool HighCost; // High Cost when expansion.
495 bool SafePhi; // LCSSASafePhiForRAUW.
497 RewritePhi(PHINode *P, unsigned I, Value *V, bool H, bool S)
498 : PN(P), Ith(I), Val(V), HighCost(H), SafePhi(S) {}
502 Value *IndVarSimplify::expandSCEVIfNeeded(SCEVExpander &Rewriter, const SCEV *S,
503 Loop *L, Instruction *InsertPt,
505 // Before expanding S into an expensive LLVM expression, see if we can use an
506 // already existing value as the expansion for S.
507 if (Value *ExistingValue = Rewriter.findExistingExpansion(S, InsertPt, L))
508 if (ExistingValue->getType() == ResultTy)
509 return ExistingValue;
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 /// Check to see if this loop has a computable loop-invariant execution count.
521 /// If so, this means that we can compute the final value of any expressions
522 /// that are recurrent in the loop, and substitute the exit values from the loop
523 /// into any instructions outside of the loop that use the final values of the
524 /// 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 /// Check whether it is possible to delete the loop after rewriting exit
704 /// value. If it is possible, ignore ReplaceExitValue and do rewriting
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 provides
769 // the input to WidenIV.
771 PHINode *NarrowIV = nullptr;
772 Type *WidestNativeType = nullptr; // Widest integer type created [sz]ext
773 bool IsSigned = false; // Was a sext user seen before a zext?
777 /// Update information about the induction variable that is extended by this
778 /// sign or zero extend operation. This is used to determine the final width of
779 /// the IV before actually widening it.
780 static void visitIVCast(CastInst *Cast, WideIVInfo &WI, ScalarEvolution *SE,
781 const TargetTransformInfo *TTI) {
782 bool IsSigned = Cast->getOpcode() == Instruction::SExt;
783 if (!IsSigned && Cast->getOpcode() != Instruction::ZExt)
786 Type *Ty = Cast->getType();
787 uint64_t Width = SE->getTypeSizeInBits(Ty);
788 if (!Cast->getModule()->getDataLayout().isLegalInteger(Width))
791 // Cast is either an sext or zext up to this point.
792 // We should not widen an indvar if arithmetics on the wider indvar are more
793 // expensive than those on the narrower indvar. We check only the cost of ADD
794 // because at least an ADD is required to increment the induction variable. We
795 // could compute more comprehensively the cost of all instructions on the
796 // induction variable when necessary.
798 TTI->getArithmeticInstrCost(Instruction::Add, Ty) >
799 TTI->getArithmeticInstrCost(Instruction::Add,
800 Cast->getOperand(0)->getType())) {
804 if (!WI.WidestNativeType) {
805 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
806 WI.IsSigned = IsSigned;
810 // We extend the IV to satisfy the sign of its first user, arbitrarily.
811 if (WI.IsSigned != IsSigned)
814 if (Width > SE->getTypeSizeInBits(WI.WidestNativeType))
815 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
820 /// Record a link in the Narrow IV def-use chain along with the WideIV that
821 /// computes the same value as the Narrow IV def. This avoids caching Use*
823 struct NarrowIVDefUse {
824 Instruction *NarrowDef = nullptr;
825 Instruction *NarrowUse = nullptr;
826 Instruction *WideDef = nullptr;
828 // True if the narrow def is never negative. Tracking this information lets
829 // us use a sign extension instead of a zero extension or vice versa, when
830 // profitable and legal.
831 bool NeverNegative = false;
833 NarrowIVDefUse(Instruction *ND, Instruction *NU, Instruction *WD,
835 : NarrowDef(ND), NarrowUse(NU), WideDef(WD),
836 NeverNegative(NeverNegative) {}
839 /// The goal of this transform is to remove sign and zero extends without
840 /// creating any new induction variables. To do this, it creates a new phi of
841 /// the wider type and redirects all users, either removing extends or inserting
842 /// truncs whenever we stop propagating the type.
858 Instruction *WideInc;
859 const SCEV *WideIncExpr;
860 SmallVectorImpl<WeakVH> &DeadInsts;
862 SmallPtrSet<Instruction*,16> Widened;
863 SmallVector<NarrowIVDefUse, 8> NarrowIVUsers;
866 WidenIV(const WideIVInfo &WI, LoopInfo *LInfo,
867 ScalarEvolution *SEv, DominatorTree *DTree,
868 SmallVectorImpl<WeakVH> &DI) :
869 OrigPhi(WI.NarrowIV),
870 WideType(WI.WidestNativeType),
871 IsSigned(WI.IsSigned),
873 L(LI->getLoopFor(OrigPhi->getParent())),
878 WideIncExpr(nullptr),
880 assert(L->getHeader() == OrigPhi->getParent() && "Phi must be an IV");
883 PHINode *createWideIV(SCEVExpander &Rewriter);
886 Value *getExtend(Value *NarrowOper, Type *WideType, bool IsSigned,
889 Instruction *cloneIVUser(NarrowIVDefUse DU, const SCEVAddRecExpr *WideAR);
890 Instruction *cloneArithmeticIVUser(NarrowIVDefUse DU,
891 const SCEVAddRecExpr *WideAR);
892 Instruction *cloneBitwiseIVUser(NarrowIVDefUse DU);
894 const SCEVAddRecExpr *getWideRecurrence(Instruction *NarrowUse);
896 const SCEVAddRecExpr* getExtendedOperandRecurrence(NarrowIVDefUse DU);
898 const SCEV *getSCEVByOpCode(const SCEV *LHS, const SCEV *RHS,
899 unsigned OpCode) const;
901 Instruction *widenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter);
903 bool widenLoopCompare(NarrowIVDefUse DU);
905 void pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef);
907 } // anonymous namespace
909 /// Perform a quick domtree based check for loop invariance assuming that V is
910 /// used within the loop. LoopInfo::isLoopInvariant() seems gratuitous for this
912 static bool isLoopInvariant(Value *V, const Loop *L, const DominatorTree *DT) {
913 Instruction *Inst = dyn_cast<Instruction>(V);
917 return DT->properlyDominates(Inst->getParent(), L->getHeader());
920 Value *WidenIV::getExtend(Value *NarrowOper, Type *WideType, bool IsSigned,
922 // Set the debug location and conservative insertion point.
923 IRBuilder<> Builder(Use);
924 // Hoist the insertion point into loop preheaders as far as possible.
925 for (const Loop *L = LI->getLoopFor(Use->getParent());
926 L && L->getLoopPreheader() && isLoopInvariant(NarrowOper, L, DT);
927 L = L->getParentLoop())
928 Builder.SetInsertPoint(L->getLoopPreheader()->getTerminator());
930 return IsSigned ? Builder.CreateSExt(NarrowOper, WideType) :
931 Builder.CreateZExt(NarrowOper, WideType);
934 /// Instantiate a wide operation to replace a narrow operation. This only needs
935 /// to handle operations that can evaluation to SCEVAddRec. It can safely return
936 /// 0 for any operation we decide not to clone.
937 Instruction *WidenIV::cloneIVUser(NarrowIVDefUse DU,
938 const SCEVAddRecExpr *WideAR) {
939 unsigned Opcode = DU.NarrowUse->getOpcode();
943 case Instruction::Add:
944 case Instruction::Mul:
945 case Instruction::UDiv:
946 case Instruction::Sub:
947 return cloneArithmeticIVUser(DU, WideAR);
949 case Instruction::And:
950 case Instruction::Or:
951 case Instruction::Xor:
952 case Instruction::Shl:
953 case Instruction::LShr:
954 case Instruction::AShr:
955 return cloneBitwiseIVUser(DU);
959 Instruction *WidenIV::cloneBitwiseIVUser(NarrowIVDefUse DU) {
960 DEBUG(dbgs() << "Cloning bitwise IVUser: " << *DU.NarrowUse << "\n");
962 // Replace NarrowDef operands with WideDef. Otherwise, we don't know anything
963 // about the narrow operand yet so must insert a [sz]ext. It is probably loop
964 // invariant and will be folded or hoisted. If it actually comes from a
965 // widened IV, it should be removed during a future call to widenIVUse.
966 Value *LHS = (DU.NarrowUse->getOperand(0) == DU.NarrowDef)
968 : getExtend(DU.NarrowUse->getOperand(0), WideType, IsSigned,
970 Value *RHS = (DU.NarrowUse->getOperand(1) == DU.NarrowDef)
972 : getExtend(DU.NarrowUse->getOperand(1), WideType, IsSigned,
975 auto *NarrowBO = cast<BinaryOperator>(DU.NarrowUse);
976 auto *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(), LHS, RHS,
977 NarrowBO->getName());
978 IRBuilder<> Builder(DU.NarrowUse);
979 Builder.Insert(WideBO);
980 if (const auto *OBO = dyn_cast<OverflowingBinaryOperator>(NarrowBO)) {
981 if (OBO->hasNoUnsignedWrap())
982 WideBO->setHasNoUnsignedWrap();
983 if (OBO->hasNoSignedWrap())
984 WideBO->setHasNoSignedWrap();
989 Instruction *WidenIV::cloneArithmeticIVUser(NarrowIVDefUse DU,
990 const SCEVAddRecExpr *WideAR) {
991 DEBUG(dbgs() << "Cloning arithmetic IVUser: " << *DU.NarrowUse << "\n");
993 // Replace NarrowDef operands with WideDef. Otherwise, we don't know anything
994 // about the narrow operand yet so must insert a [sz]ext. It is probably loop
995 // invariant and will be folded or hoisted. If it actually comes from a
996 // widened IV, it should be removed during a future call to widenIVUse.
997 Value *LHS = (DU.NarrowUse->getOperand(0) == DU.NarrowDef)
999 : getExtend(DU.NarrowUse->getOperand(0), WideType, IsSigned,
1001 Value *RHS = (DU.NarrowUse->getOperand(1) == DU.NarrowDef)
1003 : getExtend(DU.NarrowUse->getOperand(1), WideType, IsSigned,
1006 auto *NarrowBO = cast<BinaryOperator>(DU.NarrowUse);
1007 auto *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(), LHS, RHS,
1008 NarrowBO->getName());
1009 IRBuilder<> Builder(DU.NarrowUse);
1010 Builder.Insert(WideBO);
1011 if (const auto *OBO = dyn_cast<OverflowingBinaryOperator>(NarrowBO)) {
1012 if (OBO->hasNoUnsignedWrap())
1013 WideBO->setHasNoUnsignedWrap();
1014 if (OBO->hasNoSignedWrap())
1015 WideBO->setHasNoSignedWrap();
1020 const SCEV *WidenIV::getSCEVByOpCode(const SCEV *LHS, const SCEV *RHS,
1021 unsigned OpCode) const {
1022 if (OpCode == Instruction::Add)
1023 return SE->getAddExpr(LHS, RHS);
1024 if (OpCode == Instruction::Sub)
1025 return SE->getMinusSCEV(LHS, RHS);
1026 if (OpCode == Instruction::Mul)
1027 return SE->getMulExpr(LHS, RHS);
1029 llvm_unreachable("Unsupported opcode.");
1032 /// No-wrap operations can transfer sign extension of their result to their
1033 /// operands. Generate the SCEV value for the widened operation without
1034 /// actually modifying the IR yet. If the expression after extending the
1035 /// operands is an AddRec for this loop, return it.
1036 const SCEVAddRecExpr* WidenIV::getExtendedOperandRecurrence(NarrowIVDefUse DU) {
1038 // Handle the common case of add<nsw/nuw>
1039 const unsigned OpCode = DU.NarrowUse->getOpcode();
1040 // Only Add/Sub/Mul instructions supported yet.
1041 if (OpCode != Instruction::Add && OpCode != Instruction::Sub &&
1042 OpCode != Instruction::Mul)
1045 // One operand (NarrowDef) has already been extended to WideDef. Now determine
1046 // if extending the other will lead to a recurrence.
1047 const unsigned ExtendOperIdx =
1048 DU.NarrowUse->getOperand(0) == DU.NarrowDef ? 1 : 0;
1049 assert(DU.NarrowUse->getOperand(1-ExtendOperIdx) == DU.NarrowDef && "bad DU");
1051 const SCEV *ExtendOperExpr = nullptr;
1052 const OverflowingBinaryOperator *OBO =
1053 cast<OverflowingBinaryOperator>(DU.NarrowUse);
1054 if (IsSigned && OBO->hasNoSignedWrap())
1055 ExtendOperExpr = SE->getSignExtendExpr(
1056 SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType);
1057 else if(!IsSigned && OBO->hasNoUnsignedWrap())
1058 ExtendOperExpr = SE->getZeroExtendExpr(
1059 SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType);
1063 // When creating this SCEV expr, don't apply the current operations NSW or NUW
1064 // flags. This instruction may be guarded by control flow that the no-wrap
1065 // behavior depends on. Non-control-equivalent instructions can be mapped to
1066 // the same SCEV expression, and it would be incorrect to transfer NSW/NUW
1067 // semantics to those operations.
1068 const SCEV *lhs = SE->getSCEV(DU.WideDef);
1069 const SCEV *rhs = ExtendOperExpr;
1071 // Let's swap operands to the initial order for the case of non-commutative
1072 // operations, like SUB. See PR21014.
1073 if (ExtendOperIdx == 0)
1074 std::swap(lhs, rhs);
1075 const SCEVAddRecExpr *AddRec =
1076 dyn_cast<SCEVAddRecExpr>(getSCEVByOpCode(lhs, rhs, OpCode));
1078 if (!AddRec || AddRec->getLoop() != L)
1083 /// Is this instruction potentially interesting for further simplification after
1084 /// widening it's type? In other words, can the extend be safely hoisted out of
1085 /// the loop with SCEV reducing the value to a recurrence on the same loop. If
1086 /// so, return the sign or zero extended recurrence. Otherwise return NULL.
1087 const SCEVAddRecExpr *WidenIV::getWideRecurrence(Instruction *NarrowUse) {
1088 if (!SE->isSCEVable(NarrowUse->getType()))
1091 const SCEV *NarrowExpr = SE->getSCEV(NarrowUse);
1092 if (SE->getTypeSizeInBits(NarrowExpr->getType())
1093 >= SE->getTypeSizeInBits(WideType)) {
1094 // NarrowUse implicitly widens its operand. e.g. a gep with a narrow
1095 // index. So don't follow this use.
1099 const SCEV *WideExpr = IsSigned ?
1100 SE->getSignExtendExpr(NarrowExpr, WideType) :
1101 SE->getZeroExtendExpr(NarrowExpr, WideType);
1102 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(WideExpr);
1103 if (!AddRec || AddRec->getLoop() != L)
1108 /// This IV user cannot be widen. Replace this use of the original narrow IV
1109 /// with a truncation of the new wide IV to isolate and eliminate the narrow IV.
1110 static void truncateIVUse(NarrowIVDefUse DU, DominatorTree *DT) {
1111 DEBUG(dbgs() << "INDVARS: Truncate IV " << *DU.WideDef
1112 << " for user " << *DU.NarrowUse << "\n");
1113 IRBuilder<> Builder(getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT));
1114 Value *Trunc = Builder.CreateTrunc(DU.WideDef, DU.NarrowDef->getType());
1115 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, Trunc);
1118 /// If the narrow use is a compare instruction, then widen the compare
1119 // (and possibly the other operand). The extend operation is hoisted into the
1120 // loop preheader as far as possible.
1121 bool WidenIV::widenLoopCompare(NarrowIVDefUse DU) {
1122 ICmpInst *Cmp = dyn_cast<ICmpInst>(DU.NarrowUse);
1126 // We can legally widen the comparison in the following two cases:
1128 // - The signedness of the IV extension and comparison match
1130 // - The narrow IV is always positive (and thus its sign extension is equal
1131 // to its zero extension). For instance, let's say we're zero extending
1132 // %narrow for the following use
1134 // icmp slt i32 %narrow, %val ... (A)
1136 // and %narrow is always positive. Then
1138 // (A) == icmp slt i32 sext(%narrow), sext(%val)
1139 // == icmp slt i32 zext(%narrow), sext(%val)
1141 if (!(DU.NeverNegative || IsSigned == Cmp->isSigned()))
1144 Value *Op = Cmp->getOperand(Cmp->getOperand(0) == DU.NarrowDef ? 1 : 0);
1145 unsigned CastWidth = SE->getTypeSizeInBits(Op->getType());
1146 unsigned IVWidth = SE->getTypeSizeInBits(WideType);
1147 assert (CastWidth <= IVWidth && "Unexpected width while widening compare.");
1149 // Widen the compare instruction.
1150 IRBuilder<> Builder(getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT));
1151 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef);
1153 // Widen the other operand of the compare, if necessary.
1154 if (CastWidth < IVWidth) {
1155 Value *ExtOp = getExtend(Op, WideType, Cmp->isSigned(), Cmp);
1156 DU.NarrowUse->replaceUsesOfWith(Op, ExtOp);
1161 /// Determine whether an individual user of the narrow IV can be widened. If so,
1162 /// return the wide clone of the user.
1163 Instruction *WidenIV::widenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter) {
1165 // Stop traversing the def-use chain at inner-loop phis or post-loop phis.
1166 if (PHINode *UsePhi = dyn_cast<PHINode>(DU.NarrowUse)) {
1167 if (LI->getLoopFor(UsePhi->getParent()) != L) {
1168 // For LCSSA phis, sink the truncate outside the loop.
1169 // After SimplifyCFG most loop exit targets have a single predecessor.
1170 // Otherwise fall back to a truncate within the loop.
1171 if (UsePhi->getNumOperands() != 1)
1172 truncateIVUse(DU, DT);
1175 PHINode::Create(DU.WideDef->getType(), 1, UsePhi->getName() + ".wide",
1177 WidePhi->addIncoming(DU.WideDef, UsePhi->getIncomingBlock(0));
1178 IRBuilder<> Builder(&*WidePhi->getParent()->getFirstInsertionPt());
1179 Value *Trunc = Builder.CreateTrunc(WidePhi, DU.NarrowDef->getType());
1180 UsePhi->replaceAllUsesWith(Trunc);
1181 DeadInsts.emplace_back(UsePhi);
1182 DEBUG(dbgs() << "INDVARS: Widen lcssa phi " << *UsePhi
1183 << " to " << *WidePhi << "\n");
1188 // Our raison d'etre! Eliminate sign and zero extension.
1189 if (IsSigned ? isa<SExtInst>(DU.NarrowUse) : isa<ZExtInst>(DU.NarrowUse)) {
1190 Value *NewDef = DU.WideDef;
1191 if (DU.NarrowUse->getType() != WideType) {
1192 unsigned CastWidth = SE->getTypeSizeInBits(DU.NarrowUse->getType());
1193 unsigned IVWidth = SE->getTypeSizeInBits(WideType);
1194 if (CastWidth < IVWidth) {
1195 // The cast isn't as wide as the IV, so insert a Trunc.
1196 IRBuilder<> Builder(DU.NarrowUse);
1197 NewDef = Builder.CreateTrunc(DU.WideDef, DU.NarrowUse->getType());
1200 // A wider extend was hidden behind a narrower one. This may induce
1201 // another round of IV widening in which the intermediate IV becomes
1202 // dead. It should be very rare.
1203 DEBUG(dbgs() << "INDVARS: New IV " << *WidePhi
1204 << " not wide enough to subsume " << *DU.NarrowUse << "\n");
1205 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef);
1206 NewDef = DU.NarrowUse;
1209 if (NewDef != DU.NarrowUse) {
1210 DEBUG(dbgs() << "INDVARS: eliminating " << *DU.NarrowUse
1211 << " replaced by " << *DU.WideDef << "\n");
1213 DU.NarrowUse->replaceAllUsesWith(NewDef);
1214 DeadInsts.emplace_back(DU.NarrowUse);
1216 // Now that the extend is gone, we want to expose it's uses for potential
1217 // further simplification. We don't need to directly inform SimplifyIVUsers
1218 // of the new users, because their parent IV will be processed later as a
1219 // new loop phi. If we preserved IVUsers analysis, we would also want to
1220 // push the uses of WideDef here.
1222 // No further widening is needed. The deceased [sz]ext had done it for us.
1226 // Does this user itself evaluate to a recurrence after widening?
1227 const SCEVAddRecExpr *WideAddRec = getWideRecurrence(DU.NarrowUse);
1229 WideAddRec = getExtendedOperandRecurrence(DU);
1232 // If use is a loop condition, try to promote the condition instead of
1233 // truncating the IV first.
1234 if (widenLoopCompare(DU))
1237 // This user does not evaluate to a recurence after widening, so don't
1238 // follow it. Instead insert a Trunc to kill off the original use,
1239 // eventually isolating the original narrow IV so it can be removed.
1240 truncateIVUse(DU, DT);
1243 // Assume block terminators cannot evaluate to a recurrence. We can't to
1244 // insert a Trunc after a terminator if there happens to be a critical edge.
1245 assert(DU.NarrowUse != DU.NarrowUse->getParent()->getTerminator() &&
1246 "SCEV is not expected to evaluate a block terminator");
1248 // Reuse the IV increment that SCEVExpander created as long as it dominates
1250 Instruction *WideUse = nullptr;
1251 if (WideAddRec == WideIncExpr
1252 && Rewriter.hoistIVInc(WideInc, DU.NarrowUse))
1255 WideUse = cloneIVUser(DU, WideAddRec);
1259 // Evaluation of WideAddRec ensured that the narrow expression could be
1260 // extended outside the loop without overflow. This suggests that the wide use
1261 // evaluates to the same expression as the extended narrow use, but doesn't
1262 // absolutely guarantee it. Hence the following failsafe check. In rare cases
1263 // where it fails, we simply throw away the newly created wide use.
1264 if (WideAddRec != SE->getSCEV(WideUse)) {
1265 DEBUG(dbgs() << "Wide use expression mismatch: " << *WideUse
1266 << ": " << *SE->getSCEV(WideUse) << " != " << *WideAddRec << "\n");
1267 DeadInsts.emplace_back(WideUse);
1271 // Returning WideUse pushes it on the worklist.
1275 /// Add eligible users of NarrowDef to NarrowIVUsers.
1277 void WidenIV::pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef) {
1278 const SCEV *NarrowSCEV = SE->getSCEV(NarrowDef);
1279 bool NeverNegative =
1280 SE->isKnownPredicate(ICmpInst::ICMP_SGE, NarrowSCEV,
1281 SE->getConstant(NarrowSCEV->getType(), 0));
1282 for (User *U : NarrowDef->users()) {
1283 Instruction *NarrowUser = cast<Instruction>(U);
1285 // Handle data flow merges and bizarre phi cycles.
1286 if (!Widened.insert(NarrowUser).second)
1289 NarrowIVUsers.push_back(
1290 NarrowIVDefUse(NarrowDef, NarrowUser, WideDef, NeverNegative));
1294 /// Process a single induction variable. First use the SCEVExpander to create a
1295 /// wide induction variable that evaluates to the same recurrence as the
1296 /// original narrow IV. Then use a worklist to forward traverse the narrow IV's
1297 /// def-use chain. After widenIVUse has processed all interesting IV users, the
1298 /// narrow IV will be isolated for removal by DeleteDeadPHIs.
1300 /// It would be simpler to delete uses as they are processed, but we must avoid
1301 /// invalidating SCEV expressions.
1303 PHINode *WidenIV::createWideIV(SCEVExpander &Rewriter) {
1304 // Is this phi an induction variable?
1305 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(OrigPhi));
1309 // Widen the induction variable expression.
1310 const SCEV *WideIVExpr = IsSigned ?
1311 SE->getSignExtendExpr(AddRec, WideType) :
1312 SE->getZeroExtendExpr(AddRec, WideType);
1314 assert(SE->getEffectiveSCEVType(WideIVExpr->getType()) == WideType &&
1315 "Expect the new IV expression to preserve its type");
1317 // Can the IV be extended outside the loop without overflow?
1318 AddRec = dyn_cast<SCEVAddRecExpr>(WideIVExpr);
1319 if (!AddRec || AddRec->getLoop() != L)
1322 // An AddRec must have loop-invariant operands. Since this AddRec is
1323 // materialized by a loop header phi, the expression cannot have any post-loop
1324 // operands, so they must dominate the loop header.
1325 assert(SE->properlyDominates(AddRec->getStart(), L->getHeader()) &&
1326 SE->properlyDominates(AddRec->getStepRecurrence(*SE), L->getHeader())
1327 && "Loop header phi recurrence inputs do not dominate the loop");
1329 // The rewriter provides a value for the desired IV expression. This may
1330 // either find an existing phi or materialize a new one. Either way, we
1331 // expect a well-formed cyclic phi-with-increments. i.e. any operand not part
1332 // of the phi-SCC dominates the loop entry.
1333 Instruction *InsertPt = &L->getHeader()->front();
1334 WidePhi = cast<PHINode>(Rewriter.expandCodeFor(AddRec, WideType, InsertPt));
1336 // Remembering the WideIV increment generated by SCEVExpander allows
1337 // widenIVUse to reuse it when widening the narrow IV's increment. We don't
1338 // employ a general reuse mechanism because the call above is the only call to
1339 // SCEVExpander. Henceforth, we produce 1-to-1 narrow to wide uses.
1340 if (BasicBlock *LatchBlock = L->getLoopLatch()) {
1342 cast<Instruction>(WidePhi->getIncomingValueForBlock(LatchBlock));
1343 WideIncExpr = SE->getSCEV(WideInc);
1346 DEBUG(dbgs() << "Wide IV: " << *WidePhi << "\n");
1349 // Traverse the def-use chain using a worklist starting at the original IV.
1350 assert(Widened.empty() && NarrowIVUsers.empty() && "expect initial state" );
1352 Widened.insert(OrigPhi);
1353 pushNarrowIVUsers(OrigPhi, WidePhi);
1355 while (!NarrowIVUsers.empty()) {
1356 NarrowIVDefUse DU = NarrowIVUsers.pop_back_val();
1358 // Process a def-use edge. This may replace the use, so don't hold a
1359 // use_iterator across it.
1360 Instruction *WideUse = widenIVUse(DU, Rewriter);
1362 // Follow all def-use edges from the previous narrow use.
1364 pushNarrowIVUsers(DU.NarrowUse, WideUse);
1366 // widenIVUse may have removed the def-use edge.
1367 if (DU.NarrowDef->use_empty())
1368 DeadInsts.emplace_back(DU.NarrowDef);
1373 //===----------------------------------------------------------------------===//
1374 // Live IV Reduction - Minimize IVs live across the loop.
1375 //===----------------------------------------------------------------------===//
1378 //===----------------------------------------------------------------------===//
1379 // Simplification of IV users based on SCEV evaluation.
1380 //===----------------------------------------------------------------------===//
1383 class IndVarSimplifyVisitor : public IVVisitor {
1384 ScalarEvolution *SE;
1385 const TargetTransformInfo *TTI;
1391 IndVarSimplifyVisitor(PHINode *IV, ScalarEvolution *SCEV,
1392 const TargetTransformInfo *TTI,
1393 const DominatorTree *DTree)
1394 : SE(SCEV), TTI(TTI), IVPhi(IV) {
1396 WI.NarrowIV = IVPhi;
1398 setSplitOverflowIntrinsics();
1401 // Implement the interface used by simplifyUsersOfIV.
1402 void visitCast(CastInst *Cast) override { visitIVCast(Cast, WI, SE, TTI); }
1406 /// Iteratively perform simplification on a worklist of IV users. Each
1407 /// successive simplification may push more users which may themselves be
1408 /// candidates for simplification.
1410 /// Sign/Zero extend elimination is interleaved with IV simplification.
1412 void IndVarSimplify::simplifyAndExtend(Loop *L,
1413 SCEVExpander &Rewriter,
1414 LPPassManager &LPM) {
1415 SmallVector<WideIVInfo, 8> WideIVs;
1417 SmallVector<PHINode*, 8> LoopPhis;
1418 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
1419 LoopPhis.push_back(cast<PHINode>(I));
1421 // Each round of simplification iterates through the SimplifyIVUsers worklist
1422 // for all current phis, then determines whether any IVs can be
1423 // widened. Widening adds new phis to LoopPhis, inducing another round of
1424 // simplification on the wide IVs.
1425 while (!LoopPhis.empty()) {
1426 // Evaluate as many IV expressions as possible before widening any IVs. This
1427 // forces SCEV to set no-wrap flags before evaluating sign/zero
1428 // extension. The first time SCEV attempts to normalize sign/zero extension,
1429 // the result becomes final. So for the most predictable results, we delay
1430 // evaluation of sign/zero extend evaluation until needed, and avoid running
1431 // other SCEV based analysis prior to simplifyAndExtend.
1433 PHINode *CurrIV = LoopPhis.pop_back_val();
1435 // Information about sign/zero extensions of CurrIV.
1436 IndVarSimplifyVisitor Visitor(CurrIV, SE, TTI, DT);
1438 Changed |= simplifyUsersOfIV(CurrIV, SE, DT, &LPM, DeadInsts, &Visitor);
1440 if (Visitor.WI.WidestNativeType) {
1441 WideIVs.push_back(Visitor.WI);
1443 } while(!LoopPhis.empty());
1445 for (; !WideIVs.empty(); WideIVs.pop_back()) {
1446 WidenIV Widener(WideIVs.back(), LI, SE, DT, DeadInsts);
1447 if (PHINode *WidePhi = Widener.createWideIV(Rewriter)) {
1449 LoopPhis.push_back(WidePhi);
1455 //===----------------------------------------------------------------------===//
1456 // linearFunctionTestReplace and its kin. Rewrite the loop exit condition.
1457 //===----------------------------------------------------------------------===//
1459 /// Return true if this loop's backedge taken count expression can be safely and
1460 /// cheaply expanded into an instruction sequence that can be used by
1461 /// linearFunctionTestReplace.
1463 /// TODO: This fails for pointer-type loop counters with greater than one byte
1464 /// strides, consequently preventing LFTR from running. For the purpose of LFTR
1465 /// we could skip this check in the case that the LFTR loop counter (chosen by
1466 /// FindLoopCounter) is also pointer type. Instead, we could directly convert
1467 /// the loop test to an inequality test by checking the target data's alignment
1468 /// of element types (given that the initial pointer value originates from or is
1469 /// used by ABI constrained operation, as opposed to inttoptr/ptrtoint).
1470 /// However, we don't yet have a strong motivation for converting loop tests
1471 /// into inequality tests.
1472 static bool canExpandBackedgeTakenCount(Loop *L, ScalarEvolution *SE,
1473 SCEVExpander &Rewriter) {
1474 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
1475 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount) ||
1476 BackedgeTakenCount->isZero())
1479 if (!L->getExitingBlock())
1482 // Can't rewrite non-branch yet.
1483 if (!isa<BranchInst>(L->getExitingBlock()->getTerminator()))
1486 if (Rewriter.isHighCostExpansion(BackedgeTakenCount, L))
1492 /// Return the loop header phi IFF IncV adds a loop invariant value to the phi.
1493 static PHINode *getLoopPhiForCounter(Value *IncV, Loop *L, DominatorTree *DT) {
1494 Instruction *IncI = dyn_cast<Instruction>(IncV);
1498 switch (IncI->getOpcode()) {
1499 case Instruction::Add:
1500 case Instruction::Sub:
1502 case Instruction::GetElementPtr:
1503 // An IV counter must preserve its type.
1504 if (IncI->getNumOperands() == 2)
1510 PHINode *Phi = dyn_cast<PHINode>(IncI->getOperand(0));
1511 if (Phi && Phi->getParent() == L->getHeader()) {
1512 if (isLoopInvariant(IncI->getOperand(1), L, DT))
1516 if (IncI->getOpcode() == Instruction::GetElementPtr)
1519 // Allow add/sub to be commuted.
1520 Phi = dyn_cast<PHINode>(IncI->getOperand(1));
1521 if (Phi && Phi->getParent() == L->getHeader()) {
1522 if (isLoopInvariant(IncI->getOperand(0), L, DT))
1528 /// Return the compare guarding the loop latch, or NULL for unrecognized tests.
1529 static ICmpInst *getLoopTest(Loop *L) {
1530 assert(L->getExitingBlock() && "expected loop exit");
1532 BasicBlock *LatchBlock = L->getLoopLatch();
1533 // Don't bother with LFTR if the loop is not properly simplified.
1537 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
1538 assert(BI && "expected exit branch");
1540 return dyn_cast<ICmpInst>(BI->getCondition());
1543 /// linearFunctionTestReplace policy. Return true unless we can show that the
1544 /// current exit test is already sufficiently canonical.
1545 static bool needsLFTR(Loop *L, DominatorTree *DT) {
1546 // Do LFTR to simplify the exit condition to an ICMP.
1547 ICmpInst *Cond = getLoopTest(L);
1551 // Do LFTR to simplify the exit ICMP to EQ/NE
1552 ICmpInst::Predicate Pred = Cond->getPredicate();
1553 if (Pred != ICmpInst::ICMP_NE && Pred != ICmpInst::ICMP_EQ)
1556 // Look for a loop invariant RHS
1557 Value *LHS = Cond->getOperand(0);
1558 Value *RHS = Cond->getOperand(1);
1559 if (!isLoopInvariant(RHS, L, DT)) {
1560 if (!isLoopInvariant(LHS, L, DT))
1562 std::swap(LHS, RHS);
1564 // Look for a simple IV counter LHS
1565 PHINode *Phi = dyn_cast<PHINode>(LHS);
1567 Phi = getLoopPhiForCounter(LHS, L, DT);
1572 // Do LFTR if PHI node is defined in the loop, but is *not* a counter.
1573 int Idx = Phi->getBasicBlockIndex(L->getLoopLatch());
1577 // Do LFTR if the exit condition's IV is *not* a simple counter.
1578 Value *IncV = Phi->getIncomingValue(Idx);
1579 return Phi != getLoopPhiForCounter(IncV, L, DT);
1582 /// Recursive helper for hasConcreteDef(). Unfortunately, this currently boils
1583 /// down to checking that all operands are constant and listing instructions
1584 /// that may hide undef.
1585 static bool hasConcreteDefImpl(Value *V, SmallPtrSetImpl<Value*> &Visited,
1587 if (isa<Constant>(V))
1588 return !isa<UndefValue>(V);
1593 // Conservatively handle non-constant non-instructions. For example, Arguments
1595 Instruction *I = dyn_cast<Instruction>(V);
1599 // Load and return values may be undef.
1600 if(I->mayReadFromMemory() || isa<CallInst>(I) || isa<InvokeInst>(I))
1603 // Optimistically handle other instructions.
1604 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) {
1605 if (!Visited.insert(*OI).second)
1607 if (!hasConcreteDefImpl(*OI, Visited, Depth+1))
1613 /// Return true if the given value is concrete. We must prove that undef can
1616 /// TODO: If we decide that this is a good approach to checking for undef, we
1617 /// may factor it into a common location.
1618 static bool hasConcreteDef(Value *V) {
1619 SmallPtrSet<Value*, 8> Visited;
1621 return hasConcreteDefImpl(V, Visited, 0);
1624 /// Return true if this IV has any uses other than the (soon to be rewritten)
1626 static bool AlmostDeadIV(PHINode *Phi, BasicBlock *LatchBlock, Value *Cond) {
1627 int LatchIdx = Phi->getBasicBlockIndex(LatchBlock);
1628 Value *IncV = Phi->getIncomingValue(LatchIdx);
1630 for (User *U : Phi->users())
1631 if (U != Cond && U != IncV) return false;
1633 for (User *U : IncV->users())
1634 if (U != Cond && U != Phi) return false;
1638 /// Find an affine IV in canonical form.
1640 /// BECount may be an i8* pointer type. The pointer difference is already
1641 /// valid count without scaling the address stride, so it remains a pointer
1642 /// expression as far as SCEV is concerned.
1644 /// Currently only valid for LFTR. See the comments on hasConcreteDef below.
1646 /// FIXME: Accept -1 stride and set IVLimit = IVInit - BECount
1648 /// FIXME: Accept non-unit stride as long as SCEV can reduce BECount * Stride.
1649 /// This is difficult in general for SCEV because of potential overflow. But we
1650 /// could at least handle constant BECounts.
1651 static PHINode *FindLoopCounter(Loop *L, const SCEV *BECount,
1652 ScalarEvolution *SE, DominatorTree *DT) {
1653 uint64_t BCWidth = SE->getTypeSizeInBits(BECount->getType());
1656 cast<BranchInst>(L->getExitingBlock()->getTerminator())->getCondition();
1658 // Loop over all of the PHI nodes, looking for a simple counter.
1659 PHINode *BestPhi = nullptr;
1660 const SCEV *BestInit = nullptr;
1661 BasicBlock *LatchBlock = L->getLoopLatch();
1662 assert(LatchBlock && "needsLFTR should guarantee a loop latch");
1664 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
1665 PHINode *Phi = cast<PHINode>(I);
1666 if (!SE->isSCEVable(Phi->getType()))
1669 // Avoid comparing an integer IV against a pointer Limit.
1670 if (BECount->getType()->isPointerTy() && !Phi->getType()->isPointerTy())
1673 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Phi));
1674 if (!AR || AR->getLoop() != L || !AR->isAffine())
1677 // AR may be a pointer type, while BECount is an integer type.
1678 // AR may be wider than BECount. With eq/ne tests overflow is immaterial.
1679 // AR may not be a narrower type, or we may never exit.
1680 uint64_t PhiWidth = SE->getTypeSizeInBits(AR->getType());
1681 if (PhiWidth < BCWidth ||
1682 !L->getHeader()->getModule()->getDataLayout().isLegalInteger(PhiWidth))
1685 const SCEV *Step = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE));
1686 if (!Step || !Step->isOne())
1689 int LatchIdx = Phi->getBasicBlockIndex(LatchBlock);
1690 Value *IncV = Phi->getIncomingValue(LatchIdx);
1691 if (getLoopPhiForCounter(IncV, L, DT) != Phi)
1694 // Avoid reusing a potentially undef value to compute other values that may
1695 // have originally had a concrete definition.
1696 if (!hasConcreteDef(Phi)) {
1697 // We explicitly allow unknown phis as long as they are already used by
1698 // the loop test. In this case we assume that performing LFTR could not
1699 // increase the number of undef users.
1700 if (ICmpInst *Cond = getLoopTest(L)) {
1701 if (Phi != getLoopPhiForCounter(Cond->getOperand(0), L, DT)
1702 && Phi != getLoopPhiForCounter(Cond->getOperand(1), L, DT)) {
1707 const SCEV *Init = AR->getStart();
1709 if (BestPhi && !AlmostDeadIV(BestPhi, LatchBlock, Cond)) {
1710 // Don't force a live loop counter if another IV can be used.
1711 if (AlmostDeadIV(Phi, LatchBlock, Cond))
1714 // Prefer to count-from-zero. This is a more "canonical" counter form. It
1715 // also prefers integer to pointer IVs.
1716 if (BestInit->isZero() != Init->isZero()) {
1717 if (BestInit->isZero())
1720 // If two IVs both count from zero or both count from nonzero then the
1721 // narrower is likely a dead phi that has been widened. Use the wider phi
1722 // to allow the other to be eliminated.
1723 else if (PhiWidth <= SE->getTypeSizeInBits(BestPhi->getType()))
1732 /// Help linearFunctionTestReplace by generating a value that holds the RHS of
1733 /// the new loop test.
1734 static Value *genLoopLimit(PHINode *IndVar, const SCEV *IVCount, Loop *L,
1735 SCEVExpander &Rewriter, ScalarEvolution *SE) {
1736 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(IndVar));
1737 assert(AR && AR->getLoop() == L && AR->isAffine() && "bad loop counter");
1738 const SCEV *IVInit = AR->getStart();
1740 // IVInit may be a pointer while IVCount is an integer when FindLoopCounter
1741 // finds a valid pointer IV. Sign extend BECount in order to materialize a
1742 // GEP. Avoid running SCEVExpander on a new pointer value, instead reusing
1743 // the existing GEPs whenever possible.
1744 if (IndVar->getType()->isPointerTy()
1745 && !IVCount->getType()->isPointerTy()) {
1747 // IVOffset will be the new GEP offset that is interpreted by GEP as a
1748 // signed value. IVCount on the other hand represents the loop trip count,
1749 // which is an unsigned value. FindLoopCounter only allows induction
1750 // variables that have a positive unit stride of one. This means we don't
1751 // have to handle the case of negative offsets (yet) and just need to zero
1753 Type *OfsTy = SE->getEffectiveSCEVType(IVInit->getType());
1754 const SCEV *IVOffset = SE->getTruncateOrZeroExtend(IVCount, OfsTy);
1756 // Expand the code for the iteration count.
1757 assert(SE->isLoopInvariant(IVOffset, L) &&
1758 "Computed iteration count is not loop invariant!");
1759 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1760 Value *GEPOffset = Rewriter.expandCodeFor(IVOffset, OfsTy, BI);
1762 Value *GEPBase = IndVar->getIncomingValueForBlock(L->getLoopPreheader());
1763 assert(AR->getStart() == SE->getSCEV(GEPBase) && "bad loop counter");
1764 // We could handle pointer IVs other than i8*, but we need to compensate for
1765 // gep index scaling. See canExpandBackedgeTakenCount comments.
1766 assert(SE->getSizeOfExpr(IntegerType::getInt64Ty(IndVar->getContext()),
1767 cast<PointerType>(GEPBase->getType())->getElementType())->isOne()
1768 && "unit stride pointer IV must be i8*");
1770 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
1771 return Builder.CreateGEP(nullptr, GEPBase, GEPOffset, "lftr.limit");
1774 // In any other case, convert both IVInit and IVCount to integers before
1775 // comparing. This may result in SCEV expension of pointers, but in practice
1776 // SCEV will fold the pointer arithmetic away as such:
1777 // BECount = (IVEnd - IVInit - 1) => IVLimit = IVInit (postinc).
1779 // Valid Cases: (1) both integers is most common; (2) both may be pointers
1780 // for simple memset-style loops.
1782 // IVInit integer and IVCount pointer would only occur if a canonical IV
1783 // were generated on top of case #2, which is not expected.
1785 const SCEV *IVLimit = nullptr;
1786 // For unit stride, IVCount = Start + BECount with 2's complement overflow.
1787 // For non-zero Start, compute IVCount here.
1788 if (AR->getStart()->isZero())
1791 assert(AR->getStepRecurrence(*SE)->isOne() && "only handles unit stride");
1792 const SCEV *IVInit = AR->getStart();
1794 // For integer IVs, truncate the IV before computing IVInit + BECount.
1795 if (SE->getTypeSizeInBits(IVInit->getType())
1796 > SE->getTypeSizeInBits(IVCount->getType()))
1797 IVInit = SE->getTruncateExpr(IVInit, IVCount->getType());
1799 IVLimit = SE->getAddExpr(IVInit, IVCount);
1801 // Expand the code for the iteration count.
1802 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1803 IRBuilder<> Builder(BI);
1804 assert(SE->isLoopInvariant(IVLimit, L) &&
1805 "Computed iteration count is not loop invariant!");
1806 // Ensure that we generate the same type as IndVar, or a smaller integer
1807 // type. In the presence of null pointer values, we have an integer type
1808 // SCEV expression (IVInit) for a pointer type IV value (IndVar).
1809 Type *LimitTy = IVCount->getType()->isPointerTy() ?
1810 IndVar->getType() : IVCount->getType();
1811 return Rewriter.expandCodeFor(IVLimit, LimitTy, BI);
1815 /// This method rewrites the exit condition of the loop to be a canonical !=
1816 /// comparison against the incremented loop induction variable. This pass is
1817 /// able to rewrite the exit tests of any loop where the SCEV analysis can
1818 /// determine a loop-invariant trip count of the loop, which is actually a much
1819 /// broader range than just linear tests.
1820 Value *IndVarSimplify::
1821 linearFunctionTestReplace(Loop *L,
1822 const SCEV *BackedgeTakenCount,
1824 SCEVExpander &Rewriter) {
1825 assert(canExpandBackedgeTakenCount(L, SE, Rewriter) && "precondition");
1827 // Initialize CmpIndVar and IVCount to their preincremented values.
1828 Value *CmpIndVar = IndVar;
1829 const SCEV *IVCount = BackedgeTakenCount;
1831 // If the exiting block is the same as the backedge block, we prefer to
1832 // compare against the post-incremented value, otherwise we must compare
1833 // against the preincremented value.
1834 if (L->getExitingBlock() == L->getLoopLatch()) {
1835 // Add one to the "backedge-taken" count to get the trip count.
1836 // This addition may overflow, which is valid as long as the comparison is
1837 // truncated to BackedgeTakenCount->getType().
1838 IVCount = SE->getAddExpr(BackedgeTakenCount,
1839 SE->getOne(BackedgeTakenCount->getType()));
1840 // The BackedgeTaken expression contains the number of times that the
1841 // backedge branches to the loop header. This is one less than the
1842 // number of times the loop executes, so use the incremented indvar.
1843 CmpIndVar = IndVar->getIncomingValueForBlock(L->getExitingBlock());
1846 Value *ExitCnt = genLoopLimit(IndVar, IVCount, L, Rewriter, SE);
1847 assert(ExitCnt->getType()->isPointerTy() == IndVar->getType()->isPointerTy()
1848 && "genLoopLimit missed a cast");
1850 // Insert a new icmp_ne or icmp_eq instruction before the branch.
1851 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1852 ICmpInst::Predicate P;
1853 if (L->contains(BI->getSuccessor(0)))
1854 P = ICmpInst::ICMP_NE;
1856 P = ICmpInst::ICMP_EQ;
1858 DEBUG(dbgs() << "INDVARS: Rewriting loop exit condition to:\n"
1859 << " LHS:" << *CmpIndVar << '\n'
1861 << (P == ICmpInst::ICMP_NE ? "!=" : "==") << "\n"
1862 << " RHS:\t" << *ExitCnt << "\n"
1863 << " IVCount:\t" << *IVCount << "\n");
1865 IRBuilder<> Builder(BI);
1867 // LFTR can ignore IV overflow and truncate to the width of
1868 // BECount. This avoids materializing the add(zext(add)) expression.
1869 unsigned CmpIndVarSize = SE->getTypeSizeInBits(CmpIndVar->getType());
1870 unsigned ExitCntSize = SE->getTypeSizeInBits(ExitCnt->getType());
1871 if (CmpIndVarSize > ExitCntSize) {
1872 const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(SE->getSCEV(IndVar));
1873 const SCEV *ARStart = AR->getStart();
1874 const SCEV *ARStep = AR->getStepRecurrence(*SE);
1875 // For constant IVCount, avoid truncation.
1876 if (isa<SCEVConstant>(ARStart) && isa<SCEVConstant>(IVCount)) {
1877 const APInt &Start = cast<SCEVConstant>(ARStart)->getValue()->getValue();
1878 APInt Count = cast<SCEVConstant>(IVCount)->getValue()->getValue();
1879 // Note that the post-inc value of BackedgeTakenCount may have overflowed
1880 // above such that IVCount is now zero.
1881 if (IVCount != BackedgeTakenCount && Count == 0) {
1882 Count = APInt::getMaxValue(Count.getBitWidth()).zext(CmpIndVarSize);
1886 Count = Count.zext(CmpIndVarSize);
1888 if (cast<SCEVConstant>(ARStep)->getValue()->isNegative())
1889 NewLimit = Start - Count;
1891 NewLimit = Start + Count;
1892 ExitCnt = ConstantInt::get(CmpIndVar->getType(), NewLimit);
1894 DEBUG(dbgs() << " Widen RHS:\t" << *ExitCnt << "\n");
1896 CmpIndVar = Builder.CreateTrunc(CmpIndVar, ExitCnt->getType(),
1900 Value *Cond = Builder.CreateICmp(P, CmpIndVar, ExitCnt, "exitcond");
1901 Value *OrigCond = BI->getCondition();
1902 // It's tempting to use replaceAllUsesWith here to fully replace the old
1903 // comparison, but that's not immediately safe, since users of the old
1904 // comparison may not be dominated by the new comparison. Instead, just
1905 // update the branch to use the new comparison; in the common case this
1906 // will make old comparison dead.
1907 BI->setCondition(Cond);
1908 DeadInsts.push_back(OrigCond);
1915 //===----------------------------------------------------------------------===//
1916 // sinkUnusedInvariants. A late subpass to cleanup loop preheaders.
1917 //===----------------------------------------------------------------------===//
1919 /// If there's a single exit block, sink any loop-invariant values that
1920 /// were defined in the preheader but not used inside the loop into the
1921 /// exit block to reduce register pressure in the loop.
1922 void IndVarSimplify::sinkUnusedInvariants(Loop *L) {
1923 BasicBlock *ExitBlock = L->getExitBlock();
1924 if (!ExitBlock) return;
1926 BasicBlock *Preheader = L->getLoopPreheader();
1927 if (!Preheader) return;
1929 Instruction *InsertPt = &*ExitBlock->getFirstInsertionPt();
1930 BasicBlock::iterator I(Preheader->getTerminator());
1931 while (I != Preheader->begin()) {
1933 // New instructions were inserted at the end of the preheader.
1934 if (isa<PHINode>(I))
1937 // Don't move instructions which might have side effects, since the side
1938 // effects need to complete before instructions inside the loop. Also don't
1939 // move instructions which might read memory, since the loop may modify
1940 // memory. Note that it's okay if the instruction might have undefined
1941 // behavior: LoopSimplify guarantees that the preheader dominates the exit
1943 if (I->mayHaveSideEffects() || I->mayReadFromMemory())
1946 // Skip debug info intrinsics.
1947 if (isa<DbgInfoIntrinsic>(I))
1950 // Skip eh pad instructions.
1954 // Don't sink alloca: we never want to sink static alloca's out of the
1955 // entry block, and correctly sinking dynamic alloca's requires
1956 // checks for stacksave/stackrestore intrinsics.
1957 // FIXME: Refactor this check somehow?
1958 if (isa<AllocaInst>(I))
1961 // Determine if there is a use in or before the loop (direct or
1963 bool UsedInLoop = false;
1964 for (Use &U : I->uses()) {
1965 Instruction *User = cast<Instruction>(U.getUser());
1966 BasicBlock *UseBB = User->getParent();
1967 if (PHINode *P = dyn_cast<PHINode>(User)) {
1969 PHINode::getIncomingValueNumForOperand(U.getOperandNo());
1970 UseBB = P->getIncomingBlock(i);
1972 if (UseBB == Preheader || L->contains(UseBB)) {
1978 // If there is, the def must remain in the preheader.
1982 // Otherwise, sink it to the exit block.
1983 Instruction *ToMove = &*I;
1986 if (I != Preheader->begin()) {
1987 // Skip debug info intrinsics.
1990 } while (isa<DbgInfoIntrinsic>(I) && I != Preheader->begin());
1992 if (isa<DbgInfoIntrinsic>(I) && I == Preheader->begin())
1998 ToMove->moveBefore(InsertPt);
2004 //===----------------------------------------------------------------------===//
2005 // IndVarSimplify driver. Manage several subpasses of IV simplification.
2006 //===----------------------------------------------------------------------===//
2008 bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) {
2009 if (skipOptnoneFunction(L))
2012 // If LoopSimplify form is not available, stay out of trouble. Some notes:
2013 // - LSR currently only supports LoopSimplify-form loops. Indvars'
2014 // canonicalization can be a pessimization without LSR to "clean up"
2016 // - We depend on having a preheader; in particular,
2017 // Loop::getCanonicalInductionVariable only supports loops with preheaders,
2018 // and we're in trouble if we can't find the induction variable even when
2019 // we've manually inserted one.
2020 if (!L->isLoopSimplifyForm())
2023 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
2024 SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
2025 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2026 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
2027 TLI = TLIP ? &TLIP->getTLI() : nullptr;
2028 auto *TTIP = getAnalysisIfAvailable<TargetTransformInfoWrapperPass>();
2029 TTI = TTIP ? &TTIP->getTTI(*L->getHeader()->getParent()) : nullptr;
2030 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
2035 // If there are any floating-point recurrences, attempt to
2036 // transform them to use integer recurrences.
2037 rewriteNonIntegerIVs(L);
2039 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
2041 // Create a rewriter object which we'll use to transform the code with.
2042 SCEVExpander Rewriter(*SE, DL, "indvars");
2044 Rewriter.setDebugType(DEBUG_TYPE);
2047 // Eliminate redundant IV users.
2049 // Simplification works best when run before other consumers of SCEV. We
2050 // attempt to avoid evaluating SCEVs for sign/zero extend operations until
2051 // other expressions involving loop IVs have been evaluated. This helps SCEV
2052 // set no-wrap flags before normalizing sign/zero extension.
2053 Rewriter.disableCanonicalMode();
2054 simplifyAndExtend(L, Rewriter, LPM);
2056 // Check to see if this loop has a computable loop-invariant execution count.
2057 // If so, this means that we can compute the final value of any expressions
2058 // that are recurrent in the loop, and substitute the exit values from the
2059 // loop into any instructions outside of the loop that use the final values of
2060 // the current expressions.
2062 if (ReplaceExitValue != NeverRepl &&
2063 !isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2064 rewriteLoopExitValues(L, Rewriter);
2066 // Eliminate redundant IV cycles.
2067 NumElimIV += Rewriter.replaceCongruentIVs(L, DT, DeadInsts);
2069 // If we have a trip count expression, rewrite the loop's exit condition
2070 // using it. We can currently only handle loops with a single exit.
2071 if (canExpandBackedgeTakenCount(L, SE, Rewriter) && needsLFTR(L, DT)) {
2072 PHINode *IndVar = FindLoopCounter(L, BackedgeTakenCount, SE, DT);
2074 // Check preconditions for proper SCEVExpander operation. SCEV does not
2075 // express SCEVExpander's dependencies, such as LoopSimplify. Instead any
2076 // pass that uses the SCEVExpander must do it. This does not work well for
2077 // loop passes because SCEVExpander makes assumptions about all loops,
2078 // while LoopPassManager only forces the current loop to be simplified.
2080 // FIXME: SCEV expansion has no way to bail out, so the caller must
2081 // explicitly check any assumptions made by SCEV. Brittle.
2082 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(BackedgeTakenCount);
2083 if (!AR || AR->getLoop()->getLoopPreheader())
2084 (void)linearFunctionTestReplace(L, BackedgeTakenCount, IndVar,
2088 // Clear the rewriter cache, because values that are in the rewriter's cache
2089 // can be deleted in the loop below, causing the AssertingVH in the cache to
2093 // Now that we're done iterating through lists, clean up any instructions
2094 // which are now dead.
2095 while (!DeadInsts.empty())
2096 if (Instruction *Inst =
2097 dyn_cast_or_null<Instruction>(DeadInsts.pop_back_val()))
2098 RecursivelyDeleteTriviallyDeadInstructions(Inst, TLI);
2100 // The Rewriter may not be used from this point on.
2102 // Loop-invariant instructions in the preheader that aren't used in the
2103 // loop may be sunk below the loop to reduce register pressure.
2104 sinkUnusedInvariants(L);
2106 // Clean up dead instructions.
2107 Changed |= DeleteDeadPHIs(L->getHeader(), TLI);
2108 // Check a post-condition.
2109 assert(L->isLCSSAForm(*DT) &&
2110 "Indvars did not leave the loop in lcssa form!");
2112 // Verify that LFTR, and any other change have not interfered with SCEV's
2113 // ability to compute trip count.
2115 if (VerifyIndvars && !isa<SCEVCouldNotCompute>(BackedgeTakenCount)) {
2117 const SCEV *NewBECount = SE->getBackedgeTakenCount(L);
2118 if (SE->getTypeSizeInBits(BackedgeTakenCount->getType()) <
2119 SE->getTypeSizeInBits(NewBECount->getType()))
2120 NewBECount = SE->getTruncateOrNoop(NewBECount,
2121 BackedgeTakenCount->getType());
2123 BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount,
2124 NewBECount->getType());
2125 assert(BackedgeTakenCount == NewBECount && "indvars must preserve SCEV");