1 //===- IndVarSimplify.cpp - Induction Variable Elimination ----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This transformation analyzes and transforms the induction variables (and
11 // computations derived from them) into simpler forms suitable for subsequent
12 // analysis and transformation.
14 // If the trip count of a loop is computable, this pass also makes the following
16 // 1. The exit condition for the loop is canonicalized to compare the
17 // induction value against the exit value. This turns loops like:
18 // 'for (i = 7; i*i < 1000; ++i)' into 'for (i = 0; i != 25; ++i)'
19 // 2. Any use outside of the loop of an expression derived from the indvar
20 // is changed to compute the derived value outside of the loop, eliminating
21 // the dependence on the exit value of the induction variable. If the only
22 // purpose of the loop is to compute the exit value of some derived
23 // expression, this transformation will make the loop dead.
25 //===----------------------------------------------------------------------===//
27 #include "llvm/Transforms/Scalar.h"
28 #include "llvm/ADT/DenseMap.h"
29 #include "llvm/ADT/SmallVector.h"
30 #include "llvm/ADT/Statistic.h"
31 #include "llvm/Analysis/GlobalsModRef.h"
32 #include "llvm/Analysis/LoopInfo.h"
33 #include "llvm/Analysis/LoopPass.h"
34 #include "llvm/Analysis/ScalarEvolutionExpander.h"
35 #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h"
36 #include "llvm/Analysis/TargetLibraryInfo.h"
37 #include "llvm/Analysis/TargetTransformInfo.h"
38 #include "llvm/IR/BasicBlock.h"
39 #include "llvm/IR/CFG.h"
40 #include "llvm/IR/Constants.h"
41 #include "llvm/IR/DataLayout.h"
42 #include "llvm/IR/Dominators.h"
43 #include "llvm/IR/Instructions.h"
44 #include "llvm/IR/IntrinsicInst.h"
45 #include "llvm/IR/LLVMContext.h"
46 #include "llvm/IR/PatternMatch.h"
47 #include "llvm/IR/Type.h"
48 #include "llvm/Support/CommandLine.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/raw_ostream.h"
51 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
52 #include "llvm/Transforms/Utils/Local.h"
53 #include "llvm/Transforms/Utils/SimplifyIndVar.h"
56 #define DEBUG_TYPE "indvars"
58 STATISTIC(NumWidened , "Number of indvars widened");
59 STATISTIC(NumReplaced , "Number of exit values replaced");
60 STATISTIC(NumLFTR , "Number of loop exit tests replaced");
61 STATISTIC(NumElimExt , "Number of IV sign/zero extends eliminated");
62 STATISTIC(NumElimIV , "Number of congruent IVs eliminated");
64 // Trip count verification can be enabled by default under NDEBUG if we
65 // implement a strong expression equivalence checker in SCEV. Until then, we
66 // use the verify-indvars flag, which may assert in some cases.
67 static cl::opt<bool> VerifyIndvars(
68 "verify-indvars", cl::Hidden,
69 cl::desc("Verify the ScalarEvolution result after running indvars"));
71 static cl::opt<bool> ReduceLiveIVs("liv-reduce", cl::Hidden,
72 cl::desc("Reduce live induction variables."));
74 enum ReplaceExitVal { NeverRepl, OnlyCheapRepl, AlwaysRepl };
76 static cl::opt<ReplaceExitVal> ReplaceExitValue(
77 "replexitval", cl::Hidden, cl::init(OnlyCheapRepl),
78 cl::desc("Choose the strategy to replace exit value in IndVarSimplify"),
79 cl::values(clEnumValN(NeverRepl, "never", "never replace exit value"),
80 clEnumValN(OnlyCheapRepl, "cheap",
81 "only replace exit value when the cost is cheap"),
82 clEnumValN(AlwaysRepl, "always",
83 "always replace exit value whenever possible"),
91 class IndVarSimplify : public LoopPass {
95 TargetLibraryInfo *TLI;
96 const TargetTransformInfo *TTI;
98 SmallVector<WeakVH, 16> DeadInsts;
102 static char ID; // Pass identification, replacement for typeid
104 : LoopPass(ID), LI(nullptr), SE(nullptr), DT(nullptr), Changed(false) {
105 initializeIndVarSimplifyPass(*PassRegistry::getPassRegistry());
108 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
110 void getAnalysisUsage(AnalysisUsage &AU) const override {
111 AU.addRequired<DominatorTreeWrapperPass>();
112 AU.addRequired<LoopInfoWrapperPass>();
113 AU.addRequired<ScalarEvolutionWrapperPass>();
114 AU.addRequiredID(LoopSimplifyID);
115 AU.addRequiredID(LCSSAID);
116 AU.addPreserved<GlobalsAAWrapperPass>();
117 AU.addPreserved<ScalarEvolutionWrapperPass>();
118 AU.addPreservedID(LoopSimplifyID);
119 AU.addPreservedID(LCSSAID);
120 AU.setPreservesCFG();
124 void releaseMemory() override {
128 bool isValidRewrite(Value *FromVal, Value *ToVal);
130 void HandleFloatingPointIV(Loop *L, PHINode *PH);
131 void RewriteNonIntegerIVs(Loop *L);
133 void SimplifyAndExtend(Loop *L, SCEVExpander &Rewriter, LPPassManager &LPM);
135 bool CanLoopBeDeleted(Loop *L, SmallVector<RewritePhi, 8> &RewritePhiSet);
136 void RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter);
138 Value *LinearFunctionTestReplace(Loop *L, const SCEV *BackedgeTakenCount,
139 PHINode *IndVar, SCEVExpander &Rewriter);
141 void SinkUnusedInvariants(Loop *L);
143 Value *ExpandSCEVIfNeeded(SCEVExpander &Rewriter, const SCEV *S, Loop *L,
144 Instruction *InsertPt, Type *Ty);
148 char IndVarSimplify::ID = 0;
149 INITIALIZE_PASS_BEGIN(IndVarSimplify, "indvars",
150 "Induction Variable Simplification", false, false)
151 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
152 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
153 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
154 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
155 INITIALIZE_PASS_DEPENDENCY(LCSSA)
156 INITIALIZE_PASS_END(IndVarSimplify, "indvars",
157 "Induction Variable Simplification", false, false)
159 Pass *llvm::createIndVarSimplifyPass() {
160 return new IndVarSimplify();
163 /// Return true if the SCEV expansion generated by the rewriter can replace the
164 /// original value. SCEV guarantees that it produces the same value, but the way
165 /// it is produced may be illegal IR. Ideally, this function will only be
166 /// called for verification.
167 bool IndVarSimplify::isValidRewrite(Value *FromVal, Value *ToVal) {
168 // If an SCEV expression subsumed multiple pointers, its expansion could
169 // reassociate the GEP changing the base pointer. This is illegal because the
170 // final address produced by a GEP chain must be inbounds relative to its
171 // underlying object. Otherwise basic alias analysis, among other things,
172 // could fail in a dangerous way. Ultimately, SCEV will be improved to avoid
173 // producing an expression involving multiple pointers. Until then, we must
176 // Retrieve the pointer operand of the GEP. Don't use GetUnderlyingObject
177 // because it understands lcssa phis while SCEV does not.
178 Value *FromPtr = FromVal;
179 Value *ToPtr = ToVal;
180 if (GEPOperator *GEP = dyn_cast<GEPOperator>(FromVal)) {
181 FromPtr = GEP->getPointerOperand();
183 if (GEPOperator *GEP = dyn_cast<GEPOperator>(ToVal)) {
184 ToPtr = GEP->getPointerOperand();
186 if (FromPtr != FromVal || ToPtr != ToVal) {
187 // Quickly check the common case
188 if (FromPtr == ToPtr)
191 // SCEV may have rewritten an expression that produces the GEP's pointer
192 // operand. That's ok as long as the pointer operand has the same base
193 // pointer. Unlike GetUnderlyingObject(), getPointerBase() will find the
194 // base of a recurrence. This handles the case in which SCEV expansion
195 // converts a pointer type recurrence into a nonrecurrent pointer base
196 // indexed by an integer recurrence.
198 // If the GEP base pointer is a vector of pointers, abort.
199 if (!FromPtr->getType()->isPointerTy() || !ToPtr->getType()->isPointerTy())
202 const SCEV *FromBase = SE->getPointerBase(SE->getSCEV(FromPtr));
203 const SCEV *ToBase = SE->getPointerBase(SE->getSCEV(ToPtr));
204 if (FromBase == ToBase)
207 DEBUG(dbgs() << "INDVARS: GEP rewrite bail out "
208 << *FromBase << " != " << *ToBase << "\n");
215 /// Determine the insertion point for this user. By default, insert immediately
216 /// before the user. SCEVExpander or LICM will hoist loop invariants out of the
217 /// loop. For PHI nodes, there may be multiple uses, so compute the nearest
218 /// common dominator for the incoming blocks.
219 static Instruction *getInsertPointForUses(Instruction *User, Value *Def,
221 PHINode *PHI = dyn_cast<PHINode>(User);
225 Instruction *InsertPt = nullptr;
226 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i) {
227 if (PHI->getIncomingValue(i) != Def)
230 BasicBlock *InsertBB = PHI->getIncomingBlock(i);
232 InsertPt = InsertBB->getTerminator();
235 InsertBB = DT->findNearestCommonDominator(InsertPt->getParent(), InsertBB);
236 InsertPt = InsertBB->getTerminator();
238 assert(InsertPt && "Missing phi operand");
239 assert((!isa<Instruction>(Def) ||
240 DT->dominates(cast<Instruction>(Def), InsertPt)) &&
241 "def does not dominate all uses");
245 //===----------------------------------------------------------------------===//
246 // RewriteNonIntegerIVs and helpers. Prefer integer IVs.
247 //===----------------------------------------------------------------------===//
249 /// Convert APF to an integer, if possible.
250 static bool ConvertToSInt(const APFloat &APF, int64_t &IntVal) {
251 bool isExact = false;
252 // See if we can convert this to an int64_t
254 if (APF.convertToInteger(&UIntVal, 64, true, APFloat::rmTowardZero,
255 &isExact) != APFloat::opOK || !isExact)
261 /// If the loop has floating induction variable then insert corresponding
262 /// integer induction variable if possible.
264 /// for(double i = 0; i < 10000; ++i)
266 /// is converted into
267 /// for(int i = 0; i < 10000; ++i)
270 void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PN) {
271 unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
272 unsigned BackEdge = IncomingEdge^1;
274 // Check incoming value.
275 ConstantFP *InitValueVal =
276 dyn_cast<ConstantFP>(PN->getIncomingValue(IncomingEdge));
279 if (!InitValueVal || !ConvertToSInt(InitValueVal->getValueAPF(), InitValue))
282 // Check IV increment. Reject this PN if increment operation is not
283 // an add or increment value can not be represented by an integer.
284 BinaryOperator *Incr =
285 dyn_cast<BinaryOperator>(PN->getIncomingValue(BackEdge));
286 if (Incr == nullptr || Incr->getOpcode() != Instruction::FAdd) return;
288 // If this is not an add of the PHI with a constantfp, or if the constant fp
289 // is not an integer, bail out.
290 ConstantFP *IncValueVal = dyn_cast<ConstantFP>(Incr->getOperand(1));
292 if (IncValueVal == nullptr || Incr->getOperand(0) != PN ||
293 !ConvertToSInt(IncValueVal->getValueAPF(), IncValue))
296 // Check Incr uses. One user is PN and the other user is an exit condition
297 // used by the conditional terminator.
298 Value::user_iterator IncrUse = Incr->user_begin();
299 Instruction *U1 = cast<Instruction>(*IncrUse++);
300 if (IncrUse == Incr->user_end()) return;
301 Instruction *U2 = cast<Instruction>(*IncrUse++);
302 if (IncrUse != Incr->user_end()) return;
304 // Find exit condition, which is an fcmp. If it doesn't exist, or if it isn't
305 // only used by a branch, we can't transform it.
306 FCmpInst *Compare = dyn_cast<FCmpInst>(U1);
308 Compare = dyn_cast<FCmpInst>(U2);
309 if (!Compare || !Compare->hasOneUse() ||
310 !isa<BranchInst>(Compare->user_back()))
313 BranchInst *TheBr = cast<BranchInst>(Compare->user_back());
315 // We need to verify that the branch actually controls the iteration count
316 // of the loop. If not, the new IV can overflow and no one will notice.
317 // The branch block must be in the loop and one of the successors must be out
319 assert(TheBr->isConditional() && "Can't use fcmp if not conditional");
320 if (!L->contains(TheBr->getParent()) ||
321 (L->contains(TheBr->getSuccessor(0)) &&
322 L->contains(TheBr->getSuccessor(1))))
326 // If it isn't a comparison with an integer-as-fp (the exit value), we can't
328 ConstantFP *ExitValueVal = dyn_cast<ConstantFP>(Compare->getOperand(1));
330 if (ExitValueVal == nullptr ||
331 !ConvertToSInt(ExitValueVal->getValueAPF(), ExitValue))
334 // Find new predicate for integer comparison.
335 CmpInst::Predicate NewPred = CmpInst::BAD_ICMP_PREDICATE;
336 switch (Compare->getPredicate()) {
337 default: return; // Unknown comparison.
338 case CmpInst::FCMP_OEQ:
339 case CmpInst::FCMP_UEQ: NewPred = CmpInst::ICMP_EQ; break;
340 case CmpInst::FCMP_ONE:
341 case CmpInst::FCMP_UNE: NewPred = CmpInst::ICMP_NE; break;
342 case CmpInst::FCMP_OGT:
343 case CmpInst::FCMP_UGT: NewPred = CmpInst::ICMP_SGT; break;
344 case CmpInst::FCMP_OGE:
345 case CmpInst::FCMP_UGE: NewPred = CmpInst::ICMP_SGE; break;
346 case CmpInst::FCMP_OLT:
347 case CmpInst::FCMP_ULT: NewPred = CmpInst::ICMP_SLT; break;
348 case CmpInst::FCMP_OLE:
349 case CmpInst::FCMP_ULE: NewPred = CmpInst::ICMP_SLE; break;
352 // We convert the floating point induction variable to a signed i32 value if
353 // we can. This is only safe if the comparison will not overflow in a way
354 // that won't be trapped by the integer equivalent operations. Check for this
356 // TODO: We could use i64 if it is native and the range requires it.
358 // The start/stride/exit values must all fit in signed i32.
359 if (!isInt<32>(InitValue) || !isInt<32>(IncValue) || !isInt<32>(ExitValue))
362 // If not actually striding (add x, 0.0), avoid touching the code.
366 // Positive and negative strides have different safety conditions.
368 // If we have a positive stride, we require the init to be less than the
370 if (InitValue >= ExitValue)
373 uint32_t Range = uint32_t(ExitValue-InitValue);
374 // Check for infinite loop, either:
375 // while (i <= Exit) or until (i > Exit)
376 if (NewPred == CmpInst::ICMP_SLE || NewPred == CmpInst::ICMP_SGT) {
377 if (++Range == 0) return; // Range overflows.
380 unsigned Leftover = Range % uint32_t(IncValue);
382 // If this is an equality comparison, we require that the strided value
383 // exactly land on the exit value, otherwise the IV condition will wrap
384 // around and do things the fp IV wouldn't.
385 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
389 // If the stride would wrap around the i32 before exiting, we can't
391 if (Leftover != 0 && int32_t(ExitValue+IncValue) < ExitValue)
395 // If we have a negative stride, we require the init to be greater than the
397 if (InitValue <= ExitValue)
400 uint32_t Range = uint32_t(InitValue-ExitValue);
401 // Check for infinite loop, either:
402 // while (i >= Exit) or until (i < Exit)
403 if (NewPred == CmpInst::ICMP_SGE || NewPred == CmpInst::ICMP_SLT) {
404 if (++Range == 0) return; // Range overflows.
407 unsigned Leftover = Range % uint32_t(-IncValue);
409 // If this is an equality comparison, we require that the strided value
410 // exactly land on the exit value, otherwise the IV condition will wrap
411 // around and do things the fp IV wouldn't.
412 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
416 // If the stride would wrap around the i32 before exiting, we can't
418 if (Leftover != 0 && int32_t(ExitValue+IncValue) > ExitValue)
422 IntegerType *Int32Ty = Type::getInt32Ty(PN->getContext());
424 // Insert new integer induction variable.
425 PHINode *NewPHI = PHINode::Create(Int32Ty, 2, PN->getName()+".int", PN);
426 NewPHI->addIncoming(ConstantInt::get(Int32Ty, InitValue),
427 PN->getIncomingBlock(IncomingEdge));
430 BinaryOperator::CreateAdd(NewPHI, ConstantInt::get(Int32Ty, IncValue),
431 Incr->getName()+".int", Incr);
432 NewPHI->addIncoming(NewAdd, PN->getIncomingBlock(BackEdge));
434 ICmpInst *NewCompare = new ICmpInst(TheBr, NewPred, NewAdd,
435 ConstantInt::get(Int32Ty, ExitValue),
438 // In the following deletions, PN may become dead and may be deleted.
439 // Use a WeakVH to observe whether this happens.
442 // Delete the old floating point exit comparison. The branch starts using the
444 NewCompare->takeName(Compare);
445 Compare->replaceAllUsesWith(NewCompare);
446 RecursivelyDeleteTriviallyDeadInstructions(Compare, TLI);
448 // Delete the old floating point increment.
449 Incr->replaceAllUsesWith(UndefValue::get(Incr->getType()));
450 RecursivelyDeleteTriviallyDeadInstructions(Incr, TLI);
452 // If the FP induction variable still has uses, this is because something else
453 // in the loop uses its value. In order to canonicalize the induction
454 // variable, we chose to eliminate the IV and rewrite it in terms of an
457 // We give preference to sitofp over uitofp because it is faster on most
460 Value *Conv = new SIToFPInst(NewPHI, PN->getType(), "indvar.conv",
461 PN->getParent()->getFirstInsertionPt());
462 PN->replaceAllUsesWith(Conv);
463 RecursivelyDeleteTriviallyDeadInstructions(PN, TLI);
468 void IndVarSimplify::RewriteNonIntegerIVs(Loop *L) {
469 // First step. Check to see if there are any floating-point recurrences.
470 // If there are, change them into integer recurrences, permitting analysis by
471 // the SCEV routines.
473 BasicBlock *Header = L->getHeader();
475 SmallVector<WeakVH, 8> PHIs;
476 for (BasicBlock::iterator I = Header->begin();
477 PHINode *PN = dyn_cast<PHINode>(I); ++I)
480 for (unsigned i = 0, e = PHIs.size(); i != e; ++i)
481 if (PHINode *PN = dyn_cast_or_null<PHINode>(&*PHIs[i]))
482 HandleFloatingPointIV(L, PN);
484 // If the loop previously had floating-point IV, ScalarEvolution
485 // may not have been able to compute a trip count. Now that we've done some
486 // re-writing, the trip count may be computable.
492 // Collect information about PHI nodes which can be transformed in
493 // RewriteLoopExitValues.
496 unsigned Ith; // Ith incoming value.
497 Value *Val; // Exit value after expansion.
498 bool HighCost; // High Cost when expansion.
499 bool SafePhi; // LCSSASafePhiForRAUW.
501 RewritePhi(PHINode *P, unsigned I, Value *V, bool H, bool S)
502 : PN(P), Ith(I), Val(V), HighCost(H), SafePhi(S) {}
506 Value *IndVarSimplify::ExpandSCEVIfNeeded(SCEVExpander &Rewriter, const SCEV *S,
507 Loop *L, Instruction *InsertPt,
509 // Before expanding S into an expensive LLVM expression, see if we can use an
510 // already existing value as the expansion for S.
511 if (Value *ExistingValue = Rewriter.findExistingExpansion(S, InsertPt, L))
512 if (ExistingValue->getType() == ResultTy)
513 return ExistingValue;
515 // We didn't find anything, fall back to using SCEVExpander.
516 return Rewriter.expandCodeFor(S, ResultTy, InsertPt);
519 //===----------------------------------------------------------------------===//
520 // RewriteLoopExitValues - Optimize IV users outside the loop.
521 // As a side effect, reduces the amount of IV processing within the loop.
522 //===----------------------------------------------------------------------===//
524 /// Check to see if this loop has a computable loop-invariant execution count.
525 /// If so, this means that we can compute the final value of any expressions
526 /// that are recurrent in the loop, and substitute the exit values from the loop
527 /// into any instructions outside of the loop that use the final values of the
528 /// current expressions.
530 /// This is mostly redundant with the regular IndVarSimplify activities that
531 /// happen later, except that it's more powerful in some cases, because it's
532 /// able to brute-force evaluate arbitrary instructions as long as they have
533 /// constant operands at the beginning of the loop.
534 void IndVarSimplify::RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter) {
535 // Verify the input to the pass in already in LCSSA form.
536 assert(L->isLCSSAForm(*DT));
538 SmallVector<BasicBlock*, 8> ExitBlocks;
539 L->getUniqueExitBlocks(ExitBlocks);
541 SmallVector<RewritePhi, 8> RewritePhiSet;
542 // Find all values that are computed inside the loop, but used outside of it.
543 // Because of LCSSA, these values will only occur in LCSSA PHI Nodes. Scan
544 // the exit blocks of the loop to find them.
545 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
546 BasicBlock *ExitBB = ExitBlocks[i];
548 // If there are no PHI nodes in this exit block, then no values defined
549 // inside the loop are used on this path, skip it.
550 PHINode *PN = dyn_cast<PHINode>(ExitBB->begin());
553 unsigned NumPreds = PN->getNumIncomingValues();
555 // We would like to be able to RAUW single-incoming value PHI nodes. We
556 // have to be certain this is safe even when this is an LCSSA PHI node.
557 // While the computed exit value is no longer varying in *this* loop, the
558 // exit block may be an exit block for an outer containing loop as well,
559 // the exit value may be varying in the outer loop, and thus it may still
560 // require an LCSSA PHI node. The safe case is when this is
561 // single-predecessor PHI node (LCSSA) and the exit block containing it is
562 // part of the enclosing loop, or this is the outer most loop of the nest.
563 // In either case the exit value could (at most) be varying in the same
564 // loop body as the phi node itself. Thus if it is in turn used outside of
565 // an enclosing loop it will only be via a separate LCSSA node.
566 bool LCSSASafePhiForRAUW =
568 (!L->getParentLoop() || L->getParentLoop() == LI->getLoopFor(ExitBB));
570 // Iterate over all of the PHI nodes.
571 BasicBlock::iterator BBI = ExitBB->begin();
572 while ((PN = dyn_cast<PHINode>(BBI++))) {
574 continue; // dead use, don't replace it
576 // SCEV only supports integer expressions for now.
577 if (!PN->getType()->isIntegerTy() && !PN->getType()->isPointerTy())
580 // It's necessary to tell ScalarEvolution about this explicitly so that
581 // it can walk the def-use list and forget all SCEVs, as it may not be
582 // watching the PHI itself. Once the new exit value is in place, there
583 // may not be a def-use connection between the loop and every instruction
584 // which got a SCEVAddRecExpr for that loop.
587 // Iterate over all of the values in all the PHI nodes.
588 for (unsigned i = 0; i != NumPreds; ++i) {
589 // If the value being merged in is not integer or is not defined
590 // in the loop, skip it.
591 Value *InVal = PN->getIncomingValue(i);
592 if (!isa<Instruction>(InVal))
595 // If this pred is for a subloop, not L itself, skip it.
596 if (LI->getLoopFor(PN->getIncomingBlock(i)) != L)
597 continue; // The Block is in a subloop, skip it.
599 // Check that InVal is defined in the loop.
600 Instruction *Inst = cast<Instruction>(InVal);
601 if (!L->contains(Inst))
604 // Okay, this instruction has a user outside of the current loop
605 // and varies predictably *inside* the loop. Evaluate the value it
606 // contains when the loop exits, if possible.
607 const SCEV *ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop());
608 if (!SE->isLoopInvariant(ExitValue, L) ||
609 !isSafeToExpand(ExitValue, *SE))
612 // Computing the value outside of the loop brings no benefit if :
613 // - it is definitely used inside the loop in a way which can not be
615 // - no use outside of the loop can take advantage of hoisting the
616 // computation out of the loop
617 if (ExitValue->getSCEVType()>=scMulExpr) {
618 unsigned NumHardInternalUses = 0;
619 unsigned NumSoftExternalUses = 0;
620 unsigned NumUses = 0;
621 for (auto IB = Inst->user_begin(), IE = Inst->user_end();
622 IB != IE && NumUses <= 6; ++IB) {
623 Instruction *UseInstr = cast<Instruction>(*IB);
624 unsigned Opc = UseInstr->getOpcode();
626 if (L->contains(UseInstr)) {
627 if (Opc == Instruction::Call || Opc == Instruction::Ret)
628 NumHardInternalUses++;
630 if (Opc == Instruction::PHI) {
631 // Do not count the Phi as a use. LCSSA may have inserted
632 // plenty of trivial ones.
634 for (auto PB = UseInstr->user_begin(),
635 PE = UseInstr->user_end();
636 PB != PE && NumUses <= 6; ++PB, ++NumUses) {
637 unsigned PhiOpc = cast<Instruction>(*PB)->getOpcode();
638 if (PhiOpc != Instruction::Call && PhiOpc != Instruction::Ret)
639 NumSoftExternalUses++;
643 if (Opc != Instruction::Call && Opc != Instruction::Ret)
644 NumSoftExternalUses++;
647 if (NumUses <= 6 && NumHardInternalUses && !NumSoftExternalUses)
651 bool HighCost = Rewriter.isHighCostExpansion(ExitValue, L, Inst);
653 ExpandSCEVIfNeeded(Rewriter, ExitValue, L, Inst, PN->getType());
655 DEBUG(dbgs() << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal << '\n'
656 << " LoopVal = " << *Inst << "\n");
658 if (!isValidRewrite(Inst, ExitVal)) {
659 DeadInsts.push_back(ExitVal);
663 // Collect all the candidate PHINodes to be rewritten.
664 RewritePhiSet.push_back(
665 RewritePhi(PN, i, ExitVal, HighCost, LCSSASafePhiForRAUW));
670 bool LoopCanBeDel = CanLoopBeDeleted(L, RewritePhiSet);
673 for (const RewritePhi &Phi : RewritePhiSet) {
674 PHINode *PN = Phi.PN;
675 Value *ExitVal = Phi.Val;
677 // Only do the rewrite when the ExitValue can be expanded cheaply.
678 // If LoopCanBeDel is true, rewrite exit value aggressively.
679 if (ReplaceExitValue == OnlyCheapRepl && !LoopCanBeDel && Phi.HighCost) {
680 DeadInsts.push_back(ExitVal);
686 Instruction *Inst = cast<Instruction>(PN->getIncomingValue(Phi.Ith));
687 PN->setIncomingValue(Phi.Ith, ExitVal);
689 // If this instruction is dead now, delete it. Don't do it now to avoid
690 // invalidating iterators.
691 if (isInstructionTriviallyDead(Inst, TLI))
692 DeadInsts.push_back(Inst);
694 // If we determined that this PHI is safe to replace even if an LCSSA
697 PN->replaceAllUsesWith(ExitVal);
698 PN->eraseFromParent();
702 // The insertion point instruction may have been deleted; clear it out
703 // so that the rewriter doesn't trip over it later.
704 Rewriter.clearInsertPoint();
707 /// Check whether it is possible to delete the loop after rewriting exit
708 /// value. If it is possible, ignore ReplaceExitValue and do rewriting
710 bool IndVarSimplify::CanLoopBeDeleted(
711 Loop *L, SmallVector<RewritePhi, 8> &RewritePhiSet) {
713 BasicBlock *Preheader = L->getLoopPreheader();
714 // If there is no preheader, the loop will not be deleted.
718 // In LoopDeletion pass Loop can be deleted when ExitingBlocks.size() > 1.
719 // We obviate multiple ExitingBlocks case for simplicity.
720 // TODO: If we see testcase with multiple ExitingBlocks can be deleted
721 // after exit value rewriting, we can enhance the logic here.
722 SmallVector<BasicBlock *, 4> ExitingBlocks;
723 L->getExitingBlocks(ExitingBlocks);
724 SmallVector<BasicBlock *, 8> ExitBlocks;
725 L->getUniqueExitBlocks(ExitBlocks);
726 if (ExitBlocks.size() > 1 || ExitingBlocks.size() > 1)
729 BasicBlock *ExitBlock = ExitBlocks[0];
730 BasicBlock::iterator BI = ExitBlock->begin();
731 while (PHINode *P = dyn_cast<PHINode>(BI)) {
732 Value *Incoming = P->getIncomingValueForBlock(ExitingBlocks[0]);
734 // If the Incoming value of P is found in RewritePhiSet, we know it
735 // could be rewritten to use a loop invariant value in transformation
736 // phase later. Skip it in the loop invariant check below.
738 for (const RewritePhi &Phi : RewritePhiSet) {
739 unsigned i = Phi.Ith;
740 if (Phi.PN == P && (Phi.PN)->getIncomingValue(i) == Incoming) {
747 if (!found && (I = dyn_cast<Instruction>(Incoming)))
748 if (!L->hasLoopInvariantOperands(I))
754 for (Loop::block_iterator LI = L->block_begin(), LE = L->block_end();
756 for (BasicBlock::iterator BI = (*LI)->begin(), BE = (*LI)->end(); BI != BE;
758 if (BI->mayHaveSideEffects())
766 //===----------------------------------------------------------------------===//
767 // IV Widening - Extend the width of an IV to cover its widest uses.
768 //===----------------------------------------------------------------------===//
771 // Collect information about induction variables that are used by sign/zero
772 // extend operations. This information is recorded by CollectExtend and provides
773 // the input to WidenIV.
775 PHINode *NarrowIV = nullptr;
776 Type *WidestNativeType = nullptr; // Widest integer type created [sz]ext
777 bool IsSigned = false; // Was a sext user seen before a zext?
781 /// Update information about the induction variable that is extended by this
782 /// sign or zero extend operation. This is used to determine the final width of
783 /// the IV before actually widening it.
784 static void visitIVCast(CastInst *Cast, WideIVInfo &WI, ScalarEvolution *SE,
785 const TargetTransformInfo *TTI) {
786 bool IsSigned = Cast->getOpcode() == Instruction::SExt;
787 if (!IsSigned && Cast->getOpcode() != Instruction::ZExt)
790 Type *Ty = Cast->getType();
791 uint64_t Width = SE->getTypeSizeInBits(Ty);
792 if (!Cast->getModule()->getDataLayout().isLegalInteger(Width))
795 // Cast is either an sext or zext up to this point.
796 // We should not widen an indvar if arithmetics on the wider indvar are more
797 // expensive than those on the narrower indvar. We check only the cost of ADD
798 // because at least an ADD is required to increment the induction variable. We
799 // could compute more comprehensively the cost of all instructions on the
800 // induction variable when necessary.
802 TTI->getArithmeticInstrCost(Instruction::Add, Ty) >
803 TTI->getArithmeticInstrCost(Instruction::Add,
804 Cast->getOperand(0)->getType())) {
808 if (!WI.WidestNativeType) {
809 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
810 WI.IsSigned = IsSigned;
814 // We extend the IV to satisfy the sign of its first user, arbitrarily.
815 if (WI.IsSigned != IsSigned)
818 if (Width > SE->getTypeSizeInBits(WI.WidestNativeType))
819 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
824 /// Record a link in the Narrow IV def-use chain along with the WideIV that
825 /// computes the same value as the Narrow IV def. This avoids caching Use*
827 struct NarrowIVDefUse {
828 Instruction *NarrowDef = nullptr;
829 Instruction *NarrowUse = nullptr;
830 Instruction *WideDef = nullptr;
832 // True if the narrow def is never negative. Tracking this information lets
833 // us use a sign extension instead of a zero extension or vice versa, when
834 // profitable and legal.
835 bool NeverNegative = false;
837 NarrowIVDefUse(Instruction *ND, Instruction *NU, Instruction *WD,
839 : NarrowDef(ND), NarrowUse(NU), WideDef(WD),
840 NeverNegative(NeverNegative) {}
843 /// The goal of this transform is to remove sign and zero extends without
844 /// creating any new induction variables. To do this, it creates a new phi of
845 /// the wider type and redirects all users, either removing extends or inserting
846 /// truncs whenever we stop propagating the type.
862 Instruction *WideInc;
863 const SCEV *WideIncExpr;
864 SmallVectorImpl<WeakVH> &DeadInsts;
866 SmallPtrSet<Instruction*,16> Widened;
867 SmallVector<NarrowIVDefUse, 8> NarrowIVUsers;
870 WidenIV(const WideIVInfo &WI, LoopInfo *LInfo,
871 ScalarEvolution *SEv, DominatorTree *DTree,
872 SmallVectorImpl<WeakVH> &DI) :
873 OrigPhi(WI.NarrowIV),
874 WideType(WI.WidestNativeType),
875 IsSigned(WI.IsSigned),
877 L(LI->getLoopFor(OrigPhi->getParent())),
882 WideIncExpr(nullptr),
884 assert(L->getHeader() == OrigPhi->getParent() && "Phi must be an IV");
887 PHINode *CreateWideIV(SCEVExpander &Rewriter);
890 Value *getExtend(Value *NarrowOper, Type *WideType, bool IsSigned,
893 Instruction *CloneIVUser(NarrowIVDefUse DU);
895 const SCEVAddRecExpr *GetWideRecurrence(Instruction *NarrowUse);
897 const SCEVAddRecExpr* GetExtendedOperandRecurrence(NarrowIVDefUse DU);
899 const SCEV *GetSCEVByOpCode(const SCEV *LHS, const SCEV *RHS,
900 unsigned OpCode) const;
902 Instruction *WidenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter);
904 bool WidenLoopCompare(NarrowIVDefUse DU);
906 void pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef);
908 } // anonymous namespace
910 /// Perform a quick domtree based check for loop invariance assuming that V is
911 /// used within the loop. LoopInfo::isLoopInvariant() seems gratuitous for this
913 static bool isLoopInvariant(Value *V, const Loop *L, const DominatorTree *DT) {
914 Instruction *Inst = dyn_cast<Instruction>(V);
918 return DT->properlyDominates(Inst->getParent(), L->getHeader());
921 Value *WidenIV::getExtend(Value *NarrowOper, Type *WideType, bool IsSigned,
923 // Set the debug location and conservative insertion point.
924 IRBuilder<> Builder(Use);
925 // Hoist the insertion point into loop preheaders as far as possible.
926 for (const Loop *L = LI->getLoopFor(Use->getParent());
927 L && L->getLoopPreheader() && isLoopInvariant(NarrowOper, L, DT);
928 L = L->getParentLoop())
929 Builder.SetInsertPoint(L->getLoopPreheader()->getTerminator());
931 return IsSigned ? Builder.CreateSExt(NarrowOper, WideType) :
932 Builder.CreateZExt(NarrowOper, WideType);
935 /// Instantiate a wide operation to replace a narrow operation. This only needs
936 /// to handle operations that can evaluation to SCEVAddRec. It can safely return
937 /// 0 for any operation we decide not to clone.
938 Instruction *WidenIV::CloneIVUser(NarrowIVDefUse DU) {
939 unsigned Opcode = DU.NarrowUse->getOpcode();
943 case Instruction::Add:
944 case Instruction::Mul:
945 case Instruction::UDiv:
946 case Instruction::Sub:
947 case Instruction::And:
948 case Instruction::Or:
949 case Instruction::Xor:
950 case Instruction::Shl:
951 case Instruction::LShr:
952 case Instruction::AShr:
953 DEBUG(dbgs() << "Cloning IVUser: " << *DU.NarrowUse << "\n");
955 // Replace NarrowDef operands with WideDef. Otherwise, we don't know
956 // anything about the narrow operand yet so must insert a [sz]ext. It is
957 // probably loop invariant and will be folded or hoisted. If it actually
958 // comes from a widened IV, it should be removed during a future call to
960 Value *LHS = (DU.NarrowUse->getOperand(0) == DU.NarrowDef) ? DU.WideDef :
961 getExtend(DU.NarrowUse->getOperand(0), WideType, IsSigned, DU.NarrowUse);
962 Value *RHS = (DU.NarrowUse->getOperand(1) == DU.NarrowDef) ? DU.WideDef :
963 getExtend(DU.NarrowUse->getOperand(1), WideType, IsSigned, DU.NarrowUse);
965 auto *NarrowBO = cast<BinaryOperator>(DU.NarrowUse);
966 auto *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(), LHS, RHS,
967 NarrowBO->getName());
968 IRBuilder<> Builder(DU.NarrowUse);
969 Builder.Insert(WideBO);
970 if (const auto *OBO = dyn_cast<OverflowingBinaryOperator>(NarrowBO)) {
971 if (OBO->hasNoUnsignedWrap()) WideBO->setHasNoUnsignedWrap();
972 if (OBO->hasNoSignedWrap()) WideBO->setHasNoSignedWrap();
978 const SCEV *WidenIV::GetSCEVByOpCode(const SCEV *LHS, const SCEV *RHS,
979 unsigned OpCode) const {
980 if (OpCode == Instruction::Add)
981 return SE->getAddExpr(LHS, RHS);
982 if (OpCode == Instruction::Sub)
983 return SE->getMinusSCEV(LHS, RHS);
984 if (OpCode == Instruction::Mul)
985 return SE->getMulExpr(LHS, RHS);
987 llvm_unreachable("Unsupported opcode.");
990 /// No-wrap operations can transfer sign extension of their result to their
991 /// operands. Generate the SCEV value for the widened operation without
992 /// actually modifying the IR yet. If the expression after extending the
993 /// operands is an AddRec for this loop, return it.
994 const SCEVAddRecExpr* WidenIV::GetExtendedOperandRecurrence(NarrowIVDefUse DU) {
996 // Handle the common case of add<nsw/nuw>
997 const unsigned OpCode = DU.NarrowUse->getOpcode();
998 // Only Add/Sub/Mul instructions supported yet.
999 if (OpCode != Instruction::Add && OpCode != Instruction::Sub &&
1000 OpCode != Instruction::Mul)
1003 // One operand (NarrowDef) has already been extended to WideDef. Now determine
1004 // if extending the other will lead to a recurrence.
1005 const unsigned ExtendOperIdx =
1006 DU.NarrowUse->getOperand(0) == DU.NarrowDef ? 1 : 0;
1007 assert(DU.NarrowUse->getOperand(1-ExtendOperIdx) == DU.NarrowDef && "bad DU");
1009 const SCEV *ExtendOperExpr = nullptr;
1010 const OverflowingBinaryOperator *OBO =
1011 cast<OverflowingBinaryOperator>(DU.NarrowUse);
1012 if (IsSigned && OBO->hasNoSignedWrap())
1013 ExtendOperExpr = SE->getSignExtendExpr(
1014 SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType);
1015 else if(!IsSigned && OBO->hasNoUnsignedWrap())
1016 ExtendOperExpr = SE->getZeroExtendExpr(
1017 SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType);
1021 // When creating this SCEV expr, don't apply the current operations NSW or NUW
1022 // flags. This instruction may be guarded by control flow that the no-wrap
1023 // behavior depends on. Non-control-equivalent instructions can be mapped to
1024 // the same SCEV expression, and it would be incorrect to transfer NSW/NUW
1025 // semantics to those operations.
1026 const SCEV *lhs = SE->getSCEV(DU.WideDef);
1027 const SCEV *rhs = ExtendOperExpr;
1029 // Let's swap operands to the initial order for the case of non-commutative
1030 // operations, like SUB. See PR21014.
1031 if (ExtendOperIdx == 0)
1032 std::swap(lhs, rhs);
1033 const SCEVAddRecExpr *AddRec =
1034 dyn_cast<SCEVAddRecExpr>(GetSCEVByOpCode(lhs, rhs, OpCode));
1036 if (!AddRec || AddRec->getLoop() != L)
1041 /// Is this instruction potentially interesting for further simplification after
1042 /// widening it's type? In other words, can the extend be safely hoisted out of
1043 /// the loop with SCEV reducing the value to a recurrence on the same loop. If
1044 /// so, return the sign or zero extended recurrence. Otherwise return NULL.
1045 const SCEVAddRecExpr *WidenIV::GetWideRecurrence(Instruction *NarrowUse) {
1046 if (!SE->isSCEVable(NarrowUse->getType()))
1049 const SCEV *NarrowExpr = SE->getSCEV(NarrowUse);
1050 if (SE->getTypeSizeInBits(NarrowExpr->getType())
1051 >= SE->getTypeSizeInBits(WideType)) {
1052 // NarrowUse implicitly widens its operand. e.g. a gep with a narrow
1053 // index. So don't follow this use.
1057 const SCEV *WideExpr = IsSigned ?
1058 SE->getSignExtendExpr(NarrowExpr, WideType) :
1059 SE->getZeroExtendExpr(NarrowExpr, WideType);
1060 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(WideExpr);
1061 if (!AddRec || AddRec->getLoop() != L)
1066 /// This IV user cannot be widen. Replace this use of the original narrow IV
1067 /// with a truncation of the new wide IV to isolate and eliminate the narrow IV.
1068 static void truncateIVUse(NarrowIVDefUse DU, DominatorTree *DT) {
1069 DEBUG(dbgs() << "INDVARS: Truncate IV " << *DU.WideDef
1070 << " for user " << *DU.NarrowUse << "\n");
1071 IRBuilder<> Builder(getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT));
1072 Value *Trunc = Builder.CreateTrunc(DU.WideDef, DU.NarrowDef->getType());
1073 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, Trunc);
1076 /// If the narrow use is a compare instruction, then widen the compare
1077 // (and possibly the other operand). The extend operation is hoisted into the
1078 // loop preheader as far as possible.
1079 bool WidenIV::WidenLoopCompare(NarrowIVDefUse DU) {
1080 ICmpInst *Cmp = dyn_cast<ICmpInst>(DU.NarrowUse);
1084 // We can legally widen the comparison in the following two cases:
1086 // - The signedness of the IV extension and comparison match
1088 // - The narrow IV is always positive (and thus its sign extension is equal
1089 // to its zero extension). For instance, let's say we're zero extending
1090 // %narrow for the following use
1092 // icmp slt i32 %narrow, %val ... (A)
1094 // and %narrow is always positive. Then
1096 // (A) == icmp slt i32 sext(%narrow), sext(%val)
1097 // == icmp slt i32 zext(%narrow), sext(%val)
1099 if (!(DU.NeverNegative || IsSigned == Cmp->isSigned()))
1102 Value *Op = Cmp->getOperand(Cmp->getOperand(0) == DU.NarrowDef ? 1 : 0);
1103 unsigned CastWidth = SE->getTypeSizeInBits(Op->getType());
1104 unsigned IVWidth = SE->getTypeSizeInBits(WideType);
1105 assert (CastWidth <= IVWidth && "Unexpected width while widening compare.");
1107 // Widen the compare instruction.
1108 IRBuilder<> Builder(getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT));
1109 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef);
1111 // Widen the other operand of the compare, if necessary.
1112 if (CastWidth < IVWidth) {
1113 Value *ExtOp = getExtend(Op, WideType, Cmp->isSigned(), Cmp);
1114 DU.NarrowUse->replaceUsesOfWith(Op, ExtOp);
1119 /// Determine whether an individual user of the narrow IV can be widened. If so,
1120 /// return the wide clone of the user.
1121 Instruction *WidenIV::WidenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter) {
1123 // Stop traversing the def-use chain at inner-loop phis or post-loop phis.
1124 if (PHINode *UsePhi = dyn_cast<PHINode>(DU.NarrowUse)) {
1125 if (LI->getLoopFor(UsePhi->getParent()) != L) {
1126 // For LCSSA phis, sink the truncate outside the loop.
1127 // After SimplifyCFG most loop exit targets have a single predecessor.
1128 // Otherwise fall back to a truncate within the loop.
1129 if (UsePhi->getNumOperands() != 1)
1130 truncateIVUse(DU, DT);
1133 PHINode::Create(DU.WideDef->getType(), 1, UsePhi->getName() + ".wide",
1135 WidePhi->addIncoming(DU.WideDef, UsePhi->getIncomingBlock(0));
1136 IRBuilder<> Builder(WidePhi->getParent()->getFirstInsertionPt());
1137 Value *Trunc = Builder.CreateTrunc(WidePhi, DU.NarrowDef->getType());
1138 UsePhi->replaceAllUsesWith(Trunc);
1139 DeadInsts.emplace_back(UsePhi);
1140 DEBUG(dbgs() << "INDVARS: Widen lcssa phi " << *UsePhi
1141 << " to " << *WidePhi << "\n");
1146 // Our raison d'etre! Eliminate sign and zero extension.
1147 if (IsSigned ? isa<SExtInst>(DU.NarrowUse) : isa<ZExtInst>(DU.NarrowUse)) {
1148 Value *NewDef = DU.WideDef;
1149 if (DU.NarrowUse->getType() != WideType) {
1150 unsigned CastWidth = SE->getTypeSizeInBits(DU.NarrowUse->getType());
1151 unsigned IVWidth = SE->getTypeSizeInBits(WideType);
1152 if (CastWidth < IVWidth) {
1153 // The cast isn't as wide as the IV, so insert a Trunc.
1154 IRBuilder<> Builder(DU.NarrowUse);
1155 NewDef = Builder.CreateTrunc(DU.WideDef, DU.NarrowUse->getType());
1158 // A wider extend was hidden behind a narrower one. This may induce
1159 // another round of IV widening in which the intermediate IV becomes
1160 // dead. It should be very rare.
1161 DEBUG(dbgs() << "INDVARS: New IV " << *WidePhi
1162 << " not wide enough to subsume " << *DU.NarrowUse << "\n");
1163 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef);
1164 NewDef = DU.NarrowUse;
1167 if (NewDef != DU.NarrowUse) {
1168 DEBUG(dbgs() << "INDVARS: eliminating " << *DU.NarrowUse
1169 << " replaced by " << *DU.WideDef << "\n");
1171 DU.NarrowUse->replaceAllUsesWith(NewDef);
1172 DeadInsts.emplace_back(DU.NarrowUse);
1174 // Now that the extend is gone, we want to expose it's uses for potential
1175 // further simplification. We don't need to directly inform SimplifyIVUsers
1176 // of the new users, because their parent IV will be processed later as a
1177 // new loop phi. If we preserved IVUsers analysis, we would also want to
1178 // push the uses of WideDef here.
1180 // No further widening is needed. The deceased [sz]ext had done it for us.
1184 // Does this user itself evaluate to a recurrence after widening?
1185 const SCEVAddRecExpr *WideAddRec = GetWideRecurrence(DU.NarrowUse);
1187 WideAddRec = GetExtendedOperandRecurrence(DU);
1190 // If use is a loop condition, try to promote the condition instead of
1191 // truncating the IV first.
1192 if (WidenLoopCompare(DU))
1195 // This user does not evaluate to a recurence after widening, so don't
1196 // follow it. Instead insert a Trunc to kill off the original use,
1197 // eventually isolating the original narrow IV so it can be removed.
1198 truncateIVUse(DU, DT);
1201 // Assume block terminators cannot evaluate to a recurrence. We can't to
1202 // insert a Trunc after a terminator if there happens to be a critical edge.
1203 assert(DU.NarrowUse != DU.NarrowUse->getParent()->getTerminator() &&
1204 "SCEV is not expected to evaluate a block terminator");
1206 // Reuse the IV increment that SCEVExpander created as long as it dominates
1208 Instruction *WideUse = nullptr;
1209 if (WideAddRec == WideIncExpr
1210 && Rewriter.hoistIVInc(WideInc, DU.NarrowUse))
1213 WideUse = CloneIVUser(DU);
1217 // Evaluation of WideAddRec ensured that the narrow expression could be
1218 // extended outside the loop without overflow. This suggests that the wide use
1219 // evaluates to the same expression as the extended narrow use, but doesn't
1220 // absolutely guarantee it. Hence the following failsafe check. In rare cases
1221 // where it fails, we simply throw away the newly created wide use.
1222 if (WideAddRec != SE->getSCEV(WideUse)) {
1223 DEBUG(dbgs() << "Wide use expression mismatch: " << *WideUse
1224 << ": " << *SE->getSCEV(WideUse) << " != " << *WideAddRec << "\n");
1225 DeadInsts.emplace_back(WideUse);
1229 // Returning WideUse pushes it on the worklist.
1233 /// Add eligible users of NarrowDef to NarrowIVUsers.
1235 void WidenIV::pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef) {
1236 const SCEV *NarrowSCEV = SE->getSCEV(NarrowDef);
1237 bool NeverNegative =
1238 SE->isKnownPredicate(ICmpInst::ICMP_SGE, NarrowSCEV,
1239 SE->getConstant(NarrowSCEV->getType(), 0));
1240 for (User *U : NarrowDef->users()) {
1241 Instruction *NarrowUser = cast<Instruction>(U);
1243 // Handle data flow merges and bizarre phi cycles.
1244 if (!Widened.insert(NarrowUser).second)
1247 NarrowIVUsers.push_back(
1248 NarrowIVDefUse(NarrowDef, NarrowUser, WideDef, NeverNegative));
1252 /// Process a single induction variable. First use the SCEVExpander to create a
1253 /// wide induction variable that evaluates to the same recurrence as the
1254 /// original narrow IV. Then use a worklist to forward traverse the narrow IV's
1255 /// def-use chain. After WidenIVUse has processed all interesting IV users, the
1256 /// narrow IV will be isolated for removal by DeleteDeadPHIs.
1258 /// It would be simpler to delete uses as they are processed, but we must avoid
1259 /// invalidating SCEV expressions.
1261 PHINode *WidenIV::CreateWideIV(SCEVExpander &Rewriter) {
1262 // Is this phi an induction variable?
1263 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(OrigPhi));
1267 // Widen the induction variable expression.
1268 const SCEV *WideIVExpr = IsSigned ?
1269 SE->getSignExtendExpr(AddRec, WideType) :
1270 SE->getZeroExtendExpr(AddRec, WideType);
1272 assert(SE->getEffectiveSCEVType(WideIVExpr->getType()) == WideType &&
1273 "Expect the new IV expression to preserve its type");
1275 // Can the IV be extended outside the loop without overflow?
1276 AddRec = dyn_cast<SCEVAddRecExpr>(WideIVExpr);
1277 if (!AddRec || AddRec->getLoop() != L)
1280 // An AddRec must have loop-invariant operands. Since this AddRec is
1281 // materialized by a loop header phi, the expression cannot have any post-loop
1282 // operands, so they must dominate the loop header.
1283 assert(SE->properlyDominates(AddRec->getStart(), L->getHeader()) &&
1284 SE->properlyDominates(AddRec->getStepRecurrence(*SE), L->getHeader())
1285 && "Loop header phi recurrence inputs do not dominate the loop");
1287 // The rewriter provides a value for the desired IV expression. This may
1288 // either find an existing phi or materialize a new one. Either way, we
1289 // expect a well-formed cyclic phi-with-increments. i.e. any operand not part
1290 // of the phi-SCC dominates the loop entry.
1291 Instruction *InsertPt = L->getHeader()->begin();
1292 WidePhi = cast<PHINode>(Rewriter.expandCodeFor(AddRec, WideType, InsertPt));
1294 // Remembering the WideIV increment generated by SCEVExpander allows
1295 // WidenIVUse to reuse it when widening the narrow IV's increment. We don't
1296 // employ a general reuse mechanism because the call above is the only call to
1297 // SCEVExpander. Henceforth, we produce 1-to-1 narrow to wide uses.
1298 if (BasicBlock *LatchBlock = L->getLoopLatch()) {
1300 cast<Instruction>(WidePhi->getIncomingValueForBlock(LatchBlock));
1301 WideIncExpr = SE->getSCEV(WideInc);
1304 DEBUG(dbgs() << "Wide IV: " << *WidePhi << "\n");
1307 // Traverse the def-use chain using a worklist starting at the original IV.
1308 assert(Widened.empty() && NarrowIVUsers.empty() && "expect initial state" );
1310 Widened.insert(OrigPhi);
1311 pushNarrowIVUsers(OrigPhi, WidePhi);
1313 while (!NarrowIVUsers.empty()) {
1314 NarrowIVDefUse DU = NarrowIVUsers.pop_back_val();
1316 // Process a def-use edge. This may replace the use, so don't hold a
1317 // use_iterator across it.
1318 Instruction *WideUse = WidenIVUse(DU, Rewriter);
1320 // Follow all def-use edges from the previous narrow use.
1322 pushNarrowIVUsers(DU.NarrowUse, WideUse);
1324 // WidenIVUse may have removed the def-use edge.
1325 if (DU.NarrowDef->use_empty())
1326 DeadInsts.emplace_back(DU.NarrowDef);
1331 //===----------------------------------------------------------------------===//
1332 // Live IV Reduction - Minimize IVs live across the loop.
1333 //===----------------------------------------------------------------------===//
1336 //===----------------------------------------------------------------------===//
1337 // Simplification of IV users based on SCEV evaluation.
1338 //===----------------------------------------------------------------------===//
1341 class IndVarSimplifyVisitor : public IVVisitor {
1342 ScalarEvolution *SE;
1343 const TargetTransformInfo *TTI;
1349 IndVarSimplifyVisitor(PHINode *IV, ScalarEvolution *SCEV,
1350 const TargetTransformInfo *TTI,
1351 const DominatorTree *DTree)
1352 : SE(SCEV), TTI(TTI), IVPhi(IV) {
1354 WI.NarrowIV = IVPhi;
1356 setSplitOverflowIntrinsics();
1359 // Implement the interface used by simplifyUsersOfIV.
1360 void visitCast(CastInst *Cast) override { visitIVCast(Cast, WI, SE, TTI); }
1364 /// Iteratively perform simplification on a worklist of IV users. Each
1365 /// successive simplification may push more users which may themselves be
1366 /// candidates for simplification.
1368 /// Sign/Zero extend elimination is interleaved with IV simplification.
1370 void IndVarSimplify::SimplifyAndExtend(Loop *L,
1371 SCEVExpander &Rewriter,
1372 LPPassManager &LPM) {
1373 SmallVector<WideIVInfo, 8> WideIVs;
1375 SmallVector<PHINode*, 8> LoopPhis;
1376 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
1377 LoopPhis.push_back(cast<PHINode>(I));
1379 // Each round of simplification iterates through the SimplifyIVUsers worklist
1380 // for all current phis, then determines whether any IVs can be
1381 // widened. Widening adds new phis to LoopPhis, inducing another round of
1382 // simplification on the wide IVs.
1383 while (!LoopPhis.empty()) {
1384 // Evaluate as many IV expressions as possible before widening any IVs. This
1385 // forces SCEV to set no-wrap flags before evaluating sign/zero
1386 // extension. The first time SCEV attempts to normalize sign/zero extension,
1387 // the result becomes final. So for the most predictable results, we delay
1388 // evaluation of sign/zero extend evaluation until needed, and avoid running
1389 // other SCEV based analysis prior to SimplifyAndExtend.
1391 PHINode *CurrIV = LoopPhis.pop_back_val();
1393 // Information about sign/zero extensions of CurrIV.
1394 IndVarSimplifyVisitor Visitor(CurrIV, SE, TTI, DT);
1396 Changed |= simplifyUsersOfIV(CurrIV, SE, DT, &LPM, DeadInsts, &Visitor);
1398 if (Visitor.WI.WidestNativeType) {
1399 WideIVs.push_back(Visitor.WI);
1401 } while(!LoopPhis.empty());
1403 for (; !WideIVs.empty(); WideIVs.pop_back()) {
1404 WidenIV Widener(WideIVs.back(), LI, SE, DT, DeadInsts);
1405 if (PHINode *WidePhi = Widener.CreateWideIV(Rewriter)) {
1407 LoopPhis.push_back(WidePhi);
1413 //===----------------------------------------------------------------------===//
1414 // LinearFunctionTestReplace and its kin. Rewrite the loop exit condition.
1415 //===----------------------------------------------------------------------===//
1417 /// Return true if this loop's backedge taken count expression can be safely and
1418 /// cheaply expanded into an instruction sequence that can be used by
1419 /// LinearFunctionTestReplace.
1421 /// TODO: This fails for pointer-type loop counters with greater than one byte
1422 /// strides, consequently preventing LFTR from running. For the purpose of LFTR
1423 /// we could skip this check in the case that the LFTR loop counter (chosen by
1424 /// FindLoopCounter) is also pointer type. Instead, we could directly convert
1425 /// the loop test to an inequality test by checking the target data's alignment
1426 /// of element types (given that the initial pointer value originates from or is
1427 /// used by ABI constrained operation, as opposed to inttoptr/ptrtoint).
1428 /// However, we don't yet have a strong motivation for converting loop tests
1429 /// into inequality tests.
1430 static bool canExpandBackedgeTakenCount(Loop *L, ScalarEvolution *SE,
1431 SCEVExpander &Rewriter) {
1432 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
1433 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount) ||
1434 BackedgeTakenCount->isZero())
1437 if (!L->getExitingBlock())
1440 // Can't rewrite non-branch yet.
1441 if (!isa<BranchInst>(L->getExitingBlock()->getTerminator()))
1444 if (Rewriter.isHighCostExpansion(BackedgeTakenCount, L))
1450 /// Return the loop header phi IFF IncV adds a loop invariant value to the phi.
1451 static PHINode *getLoopPhiForCounter(Value *IncV, Loop *L, DominatorTree *DT) {
1452 Instruction *IncI = dyn_cast<Instruction>(IncV);
1456 switch (IncI->getOpcode()) {
1457 case Instruction::Add:
1458 case Instruction::Sub:
1460 case Instruction::GetElementPtr:
1461 // An IV counter must preserve its type.
1462 if (IncI->getNumOperands() == 2)
1468 PHINode *Phi = dyn_cast<PHINode>(IncI->getOperand(0));
1469 if (Phi && Phi->getParent() == L->getHeader()) {
1470 if (isLoopInvariant(IncI->getOperand(1), L, DT))
1474 if (IncI->getOpcode() == Instruction::GetElementPtr)
1477 // Allow add/sub to be commuted.
1478 Phi = dyn_cast<PHINode>(IncI->getOperand(1));
1479 if (Phi && Phi->getParent() == L->getHeader()) {
1480 if (isLoopInvariant(IncI->getOperand(0), L, DT))
1486 /// Return the compare guarding the loop latch, or NULL for unrecognized tests.
1487 static ICmpInst *getLoopTest(Loop *L) {
1488 assert(L->getExitingBlock() && "expected loop exit");
1490 BasicBlock *LatchBlock = L->getLoopLatch();
1491 // Don't bother with LFTR if the loop is not properly simplified.
1495 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
1496 assert(BI && "expected exit branch");
1498 return dyn_cast<ICmpInst>(BI->getCondition());
1501 /// LinearFunctionTestReplace policy. Return true unless we can show that the
1502 /// current exit test is already sufficiently canonical.
1503 static bool needsLFTR(Loop *L, DominatorTree *DT) {
1504 // Do LFTR to simplify the exit condition to an ICMP.
1505 ICmpInst *Cond = getLoopTest(L);
1509 // Do LFTR to simplify the exit ICMP to EQ/NE
1510 ICmpInst::Predicate Pred = Cond->getPredicate();
1511 if (Pred != ICmpInst::ICMP_NE && Pred != ICmpInst::ICMP_EQ)
1514 // Look for a loop invariant RHS
1515 Value *LHS = Cond->getOperand(0);
1516 Value *RHS = Cond->getOperand(1);
1517 if (!isLoopInvariant(RHS, L, DT)) {
1518 if (!isLoopInvariant(LHS, L, DT))
1520 std::swap(LHS, RHS);
1522 // Look for a simple IV counter LHS
1523 PHINode *Phi = dyn_cast<PHINode>(LHS);
1525 Phi = getLoopPhiForCounter(LHS, L, DT);
1530 // Do LFTR if PHI node is defined in the loop, but is *not* a counter.
1531 int Idx = Phi->getBasicBlockIndex(L->getLoopLatch());
1535 // Do LFTR if the exit condition's IV is *not* a simple counter.
1536 Value *IncV = Phi->getIncomingValue(Idx);
1537 return Phi != getLoopPhiForCounter(IncV, L, DT);
1540 /// Recursive helper for hasConcreteDef(). Unfortunately, this currently boils
1541 /// down to checking that all operands are constant and listing instructions
1542 /// that may hide undef.
1543 static bool hasConcreteDefImpl(Value *V, SmallPtrSetImpl<Value*> &Visited,
1545 if (isa<Constant>(V))
1546 return !isa<UndefValue>(V);
1551 // Conservatively handle non-constant non-instructions. For example, Arguments
1553 Instruction *I = dyn_cast<Instruction>(V);
1557 // Load and return values may be undef.
1558 if(I->mayReadFromMemory() || isa<CallInst>(I) || isa<InvokeInst>(I))
1561 // Optimistically handle other instructions.
1562 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) {
1563 if (!Visited.insert(*OI).second)
1565 if (!hasConcreteDefImpl(*OI, Visited, Depth+1))
1571 /// Return true if the given value is concrete. We must prove that undef can
1574 /// TODO: If we decide that this is a good approach to checking for undef, we
1575 /// may factor it into a common location.
1576 static bool hasConcreteDef(Value *V) {
1577 SmallPtrSet<Value*, 8> Visited;
1579 return hasConcreteDefImpl(V, Visited, 0);
1582 /// Return true if this IV has any uses other than the (soon to be rewritten)
1584 static bool AlmostDeadIV(PHINode *Phi, BasicBlock *LatchBlock, Value *Cond) {
1585 int LatchIdx = Phi->getBasicBlockIndex(LatchBlock);
1586 Value *IncV = Phi->getIncomingValue(LatchIdx);
1588 for (User *U : Phi->users())
1589 if (U != Cond && U != IncV) return false;
1591 for (User *U : IncV->users())
1592 if (U != Cond && U != Phi) return false;
1596 /// Find an affine IV in canonical form.
1598 /// BECount may be an i8* pointer type. The pointer difference is already
1599 /// valid count without scaling the address stride, so it remains a pointer
1600 /// expression as far as SCEV is concerned.
1602 /// Currently only valid for LFTR. See the comments on hasConcreteDef below.
1604 /// FIXME: Accept -1 stride and set IVLimit = IVInit - BECount
1606 /// FIXME: Accept non-unit stride as long as SCEV can reduce BECount * Stride.
1607 /// This is difficult in general for SCEV because of potential overflow. But we
1608 /// could at least handle constant BECounts.
1609 static PHINode *FindLoopCounter(Loop *L, const SCEV *BECount,
1610 ScalarEvolution *SE, DominatorTree *DT) {
1611 uint64_t BCWidth = SE->getTypeSizeInBits(BECount->getType());
1614 cast<BranchInst>(L->getExitingBlock()->getTerminator())->getCondition();
1616 // Loop over all of the PHI nodes, looking for a simple counter.
1617 PHINode *BestPhi = nullptr;
1618 const SCEV *BestInit = nullptr;
1619 BasicBlock *LatchBlock = L->getLoopLatch();
1620 assert(LatchBlock && "needsLFTR should guarantee a loop latch");
1622 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
1623 PHINode *Phi = cast<PHINode>(I);
1624 if (!SE->isSCEVable(Phi->getType()))
1627 // Avoid comparing an integer IV against a pointer Limit.
1628 if (BECount->getType()->isPointerTy() && !Phi->getType()->isPointerTy())
1631 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Phi));
1632 if (!AR || AR->getLoop() != L || !AR->isAffine())
1635 // AR may be a pointer type, while BECount is an integer type.
1636 // AR may be wider than BECount. With eq/ne tests overflow is immaterial.
1637 // AR may not be a narrower type, or we may never exit.
1638 uint64_t PhiWidth = SE->getTypeSizeInBits(AR->getType());
1639 if (PhiWidth < BCWidth ||
1640 !L->getHeader()->getModule()->getDataLayout().isLegalInteger(PhiWidth))
1643 const SCEV *Step = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE));
1644 if (!Step || !Step->isOne())
1647 int LatchIdx = Phi->getBasicBlockIndex(LatchBlock);
1648 Value *IncV = Phi->getIncomingValue(LatchIdx);
1649 if (getLoopPhiForCounter(IncV, L, DT) != Phi)
1652 // Avoid reusing a potentially undef value to compute other values that may
1653 // have originally had a concrete definition.
1654 if (!hasConcreteDef(Phi)) {
1655 // We explicitly allow unknown phis as long as they are already used by
1656 // the loop test. In this case we assume that performing LFTR could not
1657 // increase the number of undef users.
1658 if (ICmpInst *Cond = getLoopTest(L)) {
1659 if (Phi != getLoopPhiForCounter(Cond->getOperand(0), L, DT)
1660 && Phi != getLoopPhiForCounter(Cond->getOperand(1), L, DT)) {
1665 const SCEV *Init = AR->getStart();
1667 if (BestPhi && !AlmostDeadIV(BestPhi, LatchBlock, Cond)) {
1668 // Don't force a live loop counter if another IV can be used.
1669 if (AlmostDeadIV(Phi, LatchBlock, Cond))
1672 // Prefer to count-from-zero. This is a more "canonical" counter form. It
1673 // also prefers integer to pointer IVs.
1674 if (BestInit->isZero() != Init->isZero()) {
1675 if (BestInit->isZero())
1678 // If two IVs both count from zero or both count from nonzero then the
1679 // narrower is likely a dead phi that has been widened. Use the wider phi
1680 // to allow the other to be eliminated.
1681 else if (PhiWidth <= SE->getTypeSizeInBits(BestPhi->getType()))
1690 /// Help LinearFunctionTestReplace by generating a value that holds the RHS of
1691 /// the new loop test.
1692 static Value *genLoopLimit(PHINode *IndVar, const SCEV *IVCount, Loop *L,
1693 SCEVExpander &Rewriter, ScalarEvolution *SE) {
1694 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(IndVar));
1695 assert(AR && AR->getLoop() == L && AR->isAffine() && "bad loop counter");
1696 const SCEV *IVInit = AR->getStart();
1698 // IVInit may be a pointer while IVCount is an integer when FindLoopCounter
1699 // finds a valid pointer IV. Sign extend BECount in order to materialize a
1700 // GEP. Avoid running SCEVExpander on a new pointer value, instead reusing
1701 // the existing GEPs whenever possible.
1702 if (IndVar->getType()->isPointerTy()
1703 && !IVCount->getType()->isPointerTy()) {
1705 // IVOffset will be the new GEP offset that is interpreted by GEP as a
1706 // signed value. IVCount on the other hand represents the loop trip count,
1707 // which is an unsigned value. FindLoopCounter only allows induction
1708 // variables that have a positive unit stride of one. This means we don't
1709 // have to handle the case of negative offsets (yet) and just need to zero
1711 Type *OfsTy = SE->getEffectiveSCEVType(IVInit->getType());
1712 const SCEV *IVOffset = SE->getTruncateOrZeroExtend(IVCount, OfsTy);
1714 // Expand the code for the iteration count.
1715 assert(SE->isLoopInvariant(IVOffset, L) &&
1716 "Computed iteration count is not loop invariant!");
1717 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1718 Value *GEPOffset = Rewriter.expandCodeFor(IVOffset, OfsTy, BI);
1720 Value *GEPBase = IndVar->getIncomingValueForBlock(L->getLoopPreheader());
1721 assert(AR->getStart() == SE->getSCEV(GEPBase) && "bad loop counter");
1722 // We could handle pointer IVs other than i8*, but we need to compensate for
1723 // gep index scaling. See canExpandBackedgeTakenCount comments.
1724 assert(SE->getSizeOfExpr(IntegerType::getInt64Ty(IndVar->getContext()),
1725 cast<PointerType>(GEPBase->getType())->getElementType())->isOne()
1726 && "unit stride pointer IV must be i8*");
1728 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
1729 return Builder.CreateGEP(nullptr, GEPBase, GEPOffset, "lftr.limit");
1732 // In any other case, convert both IVInit and IVCount to integers before
1733 // comparing. This may result in SCEV expension of pointers, but in practice
1734 // SCEV will fold the pointer arithmetic away as such:
1735 // BECount = (IVEnd - IVInit - 1) => IVLimit = IVInit (postinc).
1737 // Valid Cases: (1) both integers is most common; (2) both may be pointers
1738 // for simple memset-style loops.
1740 // IVInit integer and IVCount pointer would only occur if a canonical IV
1741 // were generated on top of case #2, which is not expected.
1743 const SCEV *IVLimit = nullptr;
1744 // For unit stride, IVCount = Start + BECount with 2's complement overflow.
1745 // For non-zero Start, compute IVCount here.
1746 if (AR->getStart()->isZero())
1749 assert(AR->getStepRecurrence(*SE)->isOne() && "only handles unit stride");
1750 const SCEV *IVInit = AR->getStart();
1752 // For integer IVs, truncate the IV before computing IVInit + BECount.
1753 if (SE->getTypeSizeInBits(IVInit->getType())
1754 > SE->getTypeSizeInBits(IVCount->getType()))
1755 IVInit = SE->getTruncateExpr(IVInit, IVCount->getType());
1757 IVLimit = SE->getAddExpr(IVInit, IVCount);
1759 // Expand the code for the iteration count.
1760 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1761 IRBuilder<> Builder(BI);
1762 assert(SE->isLoopInvariant(IVLimit, L) &&
1763 "Computed iteration count is not loop invariant!");
1764 // Ensure that we generate the same type as IndVar, or a smaller integer
1765 // type. In the presence of null pointer values, we have an integer type
1766 // SCEV expression (IVInit) for a pointer type IV value (IndVar).
1767 Type *LimitTy = IVCount->getType()->isPointerTy() ?
1768 IndVar->getType() : IVCount->getType();
1769 return Rewriter.expandCodeFor(IVLimit, LimitTy, BI);
1773 /// This method rewrites the exit condition of the loop to be a canonical !=
1774 /// comparison against the incremented loop induction variable. This pass is
1775 /// able to rewrite the exit tests of any loop where the SCEV analysis can
1776 /// determine a loop-invariant trip count of the loop, which is actually a much
1777 /// broader range than just linear tests.
1778 Value *IndVarSimplify::
1779 LinearFunctionTestReplace(Loop *L,
1780 const SCEV *BackedgeTakenCount,
1782 SCEVExpander &Rewriter) {
1783 assert(canExpandBackedgeTakenCount(L, SE, Rewriter) && "precondition");
1785 // Initialize CmpIndVar and IVCount to their preincremented values.
1786 Value *CmpIndVar = IndVar;
1787 const SCEV *IVCount = BackedgeTakenCount;
1789 // If the exiting block is the same as the backedge block, we prefer to
1790 // compare against the post-incremented value, otherwise we must compare
1791 // against the preincremented value.
1792 if (L->getExitingBlock() == L->getLoopLatch()) {
1793 // Add one to the "backedge-taken" count to get the trip count.
1794 // This addition may overflow, which is valid as long as the comparison is
1795 // truncated to BackedgeTakenCount->getType().
1796 IVCount = SE->getAddExpr(BackedgeTakenCount,
1797 SE->getOne(BackedgeTakenCount->getType()));
1798 // The BackedgeTaken expression contains the number of times that the
1799 // backedge branches to the loop header. This is one less than the
1800 // number of times the loop executes, so use the incremented indvar.
1801 CmpIndVar = IndVar->getIncomingValueForBlock(L->getExitingBlock());
1804 Value *ExitCnt = genLoopLimit(IndVar, IVCount, L, Rewriter, SE);
1805 assert(ExitCnt->getType()->isPointerTy() == IndVar->getType()->isPointerTy()
1806 && "genLoopLimit missed a cast");
1808 // Insert a new icmp_ne or icmp_eq instruction before the branch.
1809 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1810 ICmpInst::Predicate P;
1811 if (L->contains(BI->getSuccessor(0)))
1812 P = ICmpInst::ICMP_NE;
1814 P = ICmpInst::ICMP_EQ;
1816 DEBUG(dbgs() << "INDVARS: Rewriting loop exit condition to:\n"
1817 << " LHS:" << *CmpIndVar << '\n'
1819 << (P == ICmpInst::ICMP_NE ? "!=" : "==") << "\n"
1820 << " RHS:\t" << *ExitCnt << "\n"
1821 << " IVCount:\t" << *IVCount << "\n");
1823 IRBuilder<> Builder(BI);
1825 // LFTR can ignore IV overflow and truncate to the width of
1826 // BECount. This avoids materializing the add(zext(add)) expression.
1827 unsigned CmpIndVarSize = SE->getTypeSizeInBits(CmpIndVar->getType());
1828 unsigned ExitCntSize = SE->getTypeSizeInBits(ExitCnt->getType());
1829 if (CmpIndVarSize > ExitCntSize) {
1830 const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(SE->getSCEV(IndVar));
1831 const SCEV *ARStart = AR->getStart();
1832 const SCEV *ARStep = AR->getStepRecurrence(*SE);
1833 // For constant IVCount, avoid truncation.
1834 if (isa<SCEVConstant>(ARStart) && isa<SCEVConstant>(IVCount)) {
1835 const APInt &Start = cast<SCEVConstant>(ARStart)->getValue()->getValue();
1836 APInt Count = cast<SCEVConstant>(IVCount)->getValue()->getValue();
1837 // Note that the post-inc value of BackedgeTakenCount may have overflowed
1838 // above such that IVCount is now zero.
1839 if (IVCount != BackedgeTakenCount && Count == 0) {
1840 Count = APInt::getMaxValue(Count.getBitWidth()).zext(CmpIndVarSize);
1844 Count = Count.zext(CmpIndVarSize);
1846 if (cast<SCEVConstant>(ARStep)->getValue()->isNegative())
1847 NewLimit = Start - Count;
1849 NewLimit = Start + Count;
1850 ExitCnt = ConstantInt::get(CmpIndVar->getType(), NewLimit);
1852 DEBUG(dbgs() << " Widen RHS:\t" << *ExitCnt << "\n");
1854 CmpIndVar = Builder.CreateTrunc(CmpIndVar, ExitCnt->getType(),
1858 Value *Cond = Builder.CreateICmp(P, CmpIndVar, ExitCnt, "exitcond");
1859 Value *OrigCond = BI->getCondition();
1860 // It's tempting to use replaceAllUsesWith here to fully replace the old
1861 // comparison, but that's not immediately safe, since users of the old
1862 // comparison may not be dominated by the new comparison. Instead, just
1863 // update the branch to use the new comparison; in the common case this
1864 // will make old comparison dead.
1865 BI->setCondition(Cond);
1866 DeadInsts.push_back(OrigCond);
1873 //===----------------------------------------------------------------------===//
1874 // SinkUnusedInvariants. A late subpass to cleanup loop preheaders.
1875 //===----------------------------------------------------------------------===//
1877 /// If there's a single exit block, sink any loop-invariant values that
1878 /// were defined in the preheader but not used inside the loop into the
1879 /// exit block to reduce register pressure in the loop.
1880 void IndVarSimplify::SinkUnusedInvariants(Loop *L) {
1881 BasicBlock *ExitBlock = L->getExitBlock();
1882 if (!ExitBlock) return;
1884 BasicBlock *Preheader = L->getLoopPreheader();
1885 if (!Preheader) return;
1887 Instruction *InsertPt = ExitBlock->getFirstInsertionPt();
1888 BasicBlock::iterator I = Preheader->getTerminator();
1889 while (I != Preheader->begin()) {
1891 // New instructions were inserted at the end of the preheader.
1892 if (isa<PHINode>(I))
1895 // Don't move instructions which might have side effects, since the side
1896 // effects need to complete before instructions inside the loop. Also don't
1897 // move instructions which might read memory, since the loop may modify
1898 // memory. Note that it's okay if the instruction might have undefined
1899 // behavior: LoopSimplify guarantees that the preheader dominates the exit
1901 if (I->mayHaveSideEffects() || I->mayReadFromMemory())
1904 // Skip debug info intrinsics.
1905 if (isa<DbgInfoIntrinsic>(I))
1908 // Skip eh pad instructions.
1912 // Don't sink alloca: we never want to sink static alloca's out of the
1913 // entry block, and correctly sinking dynamic alloca's requires
1914 // checks for stacksave/stackrestore intrinsics.
1915 // FIXME: Refactor this check somehow?
1916 if (isa<AllocaInst>(I))
1919 // Determine if there is a use in or before the loop (direct or
1921 bool UsedInLoop = false;
1922 for (Use &U : I->uses()) {
1923 Instruction *User = cast<Instruction>(U.getUser());
1924 BasicBlock *UseBB = User->getParent();
1925 if (PHINode *P = dyn_cast<PHINode>(User)) {
1927 PHINode::getIncomingValueNumForOperand(U.getOperandNo());
1928 UseBB = P->getIncomingBlock(i);
1930 if (UseBB == Preheader || L->contains(UseBB)) {
1936 // If there is, the def must remain in the preheader.
1940 // Otherwise, sink it to the exit block.
1941 Instruction *ToMove = I;
1944 if (I != Preheader->begin()) {
1945 // Skip debug info intrinsics.
1948 } while (isa<DbgInfoIntrinsic>(I) && I != Preheader->begin());
1950 if (isa<DbgInfoIntrinsic>(I) && I == Preheader->begin())
1956 ToMove->moveBefore(InsertPt);
1962 //===----------------------------------------------------------------------===//
1963 // IndVarSimplify driver. Manage several subpasses of IV simplification.
1964 //===----------------------------------------------------------------------===//
1966 bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) {
1967 if (skipOptnoneFunction(L))
1970 // If LoopSimplify form is not available, stay out of trouble. Some notes:
1971 // - LSR currently only supports LoopSimplify-form loops. Indvars'
1972 // canonicalization can be a pessimization without LSR to "clean up"
1974 // - We depend on having a preheader; in particular,
1975 // Loop::getCanonicalInductionVariable only supports loops with preheaders,
1976 // and we're in trouble if we can't find the induction variable even when
1977 // we've manually inserted one.
1978 if (!L->isLoopSimplifyForm())
1981 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1982 SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
1983 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1984 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1985 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1986 auto *TTIP = getAnalysisIfAvailable<TargetTransformInfoWrapperPass>();
1987 TTI = TTIP ? &TTIP->getTTI(*L->getHeader()->getParent()) : nullptr;
1988 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
1993 // If there are any floating-point recurrences, attempt to
1994 // transform them to use integer recurrences.
1995 RewriteNonIntegerIVs(L);
1997 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
1999 // Create a rewriter object which we'll use to transform the code with.
2000 SCEVExpander Rewriter(*SE, DL, "indvars");
2002 Rewriter.setDebugType(DEBUG_TYPE);
2005 // Eliminate redundant IV users.
2007 // Simplification works best when run before other consumers of SCEV. We
2008 // attempt to avoid evaluating SCEVs for sign/zero extend operations until
2009 // other expressions involving loop IVs have been evaluated. This helps SCEV
2010 // set no-wrap flags before normalizing sign/zero extension.
2011 Rewriter.disableCanonicalMode();
2012 SimplifyAndExtend(L, Rewriter, LPM);
2014 // Check to see if this loop has a computable loop-invariant execution count.
2015 // If so, this means that we can compute the final value of any expressions
2016 // that are recurrent in the loop, and substitute the exit values from the
2017 // loop into any instructions outside of the loop that use the final values of
2018 // the current expressions.
2020 if (ReplaceExitValue != NeverRepl &&
2021 !isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2022 RewriteLoopExitValues(L, Rewriter);
2024 // Eliminate redundant IV cycles.
2025 NumElimIV += Rewriter.replaceCongruentIVs(L, DT, DeadInsts);
2027 // If we have a trip count expression, rewrite the loop's exit condition
2028 // using it. We can currently only handle loops with a single exit.
2029 if (canExpandBackedgeTakenCount(L, SE, Rewriter) && needsLFTR(L, DT)) {
2030 PHINode *IndVar = FindLoopCounter(L, BackedgeTakenCount, SE, DT);
2032 // Check preconditions for proper SCEVExpander operation. SCEV does not
2033 // express SCEVExpander's dependencies, such as LoopSimplify. Instead any
2034 // pass that uses the SCEVExpander must do it. This does not work well for
2035 // loop passes because SCEVExpander makes assumptions about all loops,
2036 // while LoopPassManager only forces the current loop to be simplified.
2038 // FIXME: SCEV expansion has no way to bail out, so the caller must
2039 // explicitly check any assumptions made by SCEV. Brittle.
2040 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(BackedgeTakenCount);
2041 if (!AR || AR->getLoop()->getLoopPreheader())
2042 (void)LinearFunctionTestReplace(L, BackedgeTakenCount, IndVar,
2046 // Clear the rewriter cache, because values that are in the rewriter's cache
2047 // can be deleted in the loop below, causing the AssertingVH in the cache to
2051 // Now that we're done iterating through lists, clean up any instructions
2052 // which are now dead.
2053 while (!DeadInsts.empty())
2054 if (Instruction *Inst =
2055 dyn_cast_or_null<Instruction>(DeadInsts.pop_back_val()))
2056 RecursivelyDeleteTriviallyDeadInstructions(Inst, TLI);
2058 // The Rewriter may not be used from this point on.
2060 // Loop-invariant instructions in the preheader that aren't used in the
2061 // loop may be sunk below the loop to reduce register pressure.
2062 SinkUnusedInvariants(L);
2064 // Clean up dead instructions.
2065 Changed |= DeleteDeadPHIs(L->getHeader(), TLI);
2066 // Check a post-condition.
2067 assert(L->isLCSSAForm(*DT) &&
2068 "Indvars did not leave the loop in lcssa form!");
2070 // Verify that LFTR, and any other change have not interfered with SCEV's
2071 // ability to compute trip count.
2073 if (VerifyIndvars && !isa<SCEVCouldNotCompute>(BackedgeTakenCount)) {
2075 const SCEV *NewBECount = SE->getBackedgeTakenCount(L);
2076 if (SE->getTypeSizeInBits(BackedgeTakenCount->getType()) <
2077 SE->getTypeSizeInBits(NewBECount->getType()))
2078 NewBECount = SE->getTruncateOrNoop(NewBECount,
2079 BackedgeTakenCount->getType());
2081 BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount,
2082 NewBECount->getType());
2083 assert(BackedgeTakenCount == NewBECount && "indvars must preserve SCEV");