1 //===- ScalarEvolutionExpander.cpp - Scalar Evolution Analysis --*- C++ -*-===//
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 file contains the implementation of the scalar evolution expander,
11 // which is used to generate the code corresponding to a given scalar evolution
14 //===----------------------------------------------------------------------===//
16 #include "llvm/Analysis/ScalarEvolutionExpander.h"
17 #include "llvm/ADT/STLExtras.h"
18 #include "llvm/ADT/SmallSet.h"
19 #include "llvm/Analysis/InstructionSimplify.h"
20 #include "llvm/Analysis/LoopInfo.h"
21 #include "llvm/Analysis/TargetTransformInfo.h"
22 #include "llvm/IR/DataLayout.h"
23 #include "llvm/IR/Dominators.h"
24 #include "llvm/IR/IntrinsicInst.h"
25 #include "llvm/IR/LLVMContext.h"
26 #include "llvm/IR/Module.h"
27 #include "llvm/IR/PatternMatch.h"
28 #include "llvm/Support/Debug.h"
29 #include "llvm/Support/raw_ostream.h"
32 using namespace PatternMatch;
34 /// ReuseOrCreateCast - Arrange for there to be a cast of V to Ty at IP,
35 /// reusing an existing cast if a suitable one exists, moving an existing
36 /// cast if a suitable one exists but isn't in the right place, or
37 /// creating a new one.
38 Value *SCEVExpander::ReuseOrCreateCast(Value *V, Type *Ty,
39 Instruction::CastOps Op,
40 BasicBlock::iterator IP) {
41 // This function must be called with the builder having a valid insertion
42 // point. It doesn't need to be the actual IP where the uses of the returned
43 // cast will be added, but it must dominate such IP.
44 // We use this precondition to produce a cast that will dominate all its
45 // uses. In particular, this is crucial for the case where the builder's
46 // insertion point *is* the point where we were asked to put the cast.
47 // Since we don't know the builder's insertion point is actually
48 // where the uses will be added (only that it dominates it), we are
49 // not allowed to move it.
50 BasicBlock::iterator BIP = Builder.GetInsertPoint();
52 Instruction *Ret = nullptr;
54 // Check to see if there is already a cast!
55 for (User *U : V->users())
56 if (U->getType() == Ty)
57 if (CastInst *CI = dyn_cast<CastInst>(U))
58 if (CI->getOpcode() == Op) {
59 // If the cast isn't where we want it, create a new cast at IP.
60 // Likewise, do not reuse a cast at BIP because it must dominate
61 // instructions that might be inserted before BIP.
62 if (BasicBlock::iterator(CI) != IP || BIP == IP) {
63 // Create a new cast, and leave the old cast in place in case
64 // it is being used as an insert point. Clear its operand
65 // so that it doesn't hold anything live.
66 Ret = CastInst::Create(Op, V, Ty, "", IP);
68 CI->replaceAllUsesWith(Ret);
69 CI->setOperand(0, UndefValue::get(V->getType()));
78 Ret = CastInst::Create(Op, V, Ty, V->getName(), IP);
80 // We assert at the end of the function since IP might point to an
81 // instruction with different dominance properties than a cast
82 // (an invoke for example) and not dominate BIP (but the cast does).
83 assert(SE.DT.dominates(Ret, BIP));
85 rememberInstruction(Ret);
89 /// InsertNoopCastOfTo - Insert a cast of V to the specified type,
90 /// which must be possible with a noop cast, doing what we can to share
92 Value *SCEVExpander::InsertNoopCastOfTo(Value *V, Type *Ty) {
93 Instruction::CastOps Op = CastInst::getCastOpcode(V, false, Ty, false);
94 assert((Op == Instruction::BitCast ||
95 Op == Instruction::PtrToInt ||
96 Op == Instruction::IntToPtr) &&
97 "InsertNoopCastOfTo cannot perform non-noop casts!");
98 assert(SE.getTypeSizeInBits(V->getType()) == SE.getTypeSizeInBits(Ty) &&
99 "InsertNoopCastOfTo cannot change sizes!");
101 // Short-circuit unnecessary bitcasts.
102 if (Op == Instruction::BitCast) {
103 if (V->getType() == Ty)
105 if (CastInst *CI = dyn_cast<CastInst>(V)) {
106 if (CI->getOperand(0)->getType() == Ty)
107 return CI->getOperand(0);
110 // Short-circuit unnecessary inttoptr<->ptrtoint casts.
111 if ((Op == Instruction::PtrToInt || Op == Instruction::IntToPtr) &&
112 SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(V->getType())) {
113 if (CastInst *CI = dyn_cast<CastInst>(V))
114 if ((CI->getOpcode() == Instruction::PtrToInt ||
115 CI->getOpcode() == Instruction::IntToPtr) &&
116 SE.getTypeSizeInBits(CI->getType()) ==
117 SE.getTypeSizeInBits(CI->getOperand(0)->getType()))
118 return CI->getOperand(0);
119 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
120 if ((CE->getOpcode() == Instruction::PtrToInt ||
121 CE->getOpcode() == Instruction::IntToPtr) &&
122 SE.getTypeSizeInBits(CE->getType()) ==
123 SE.getTypeSizeInBits(CE->getOperand(0)->getType()))
124 return CE->getOperand(0);
127 // Fold a cast of a constant.
128 if (Constant *C = dyn_cast<Constant>(V))
129 return ConstantExpr::getCast(Op, C, Ty);
131 // Cast the argument at the beginning of the entry block, after
132 // any bitcasts of other arguments.
133 if (Argument *A = dyn_cast<Argument>(V)) {
134 BasicBlock::iterator IP = A->getParent()->getEntryBlock().begin();
135 while ((isa<BitCastInst>(IP) &&
136 isa<Argument>(cast<BitCastInst>(IP)->getOperand(0)) &&
137 cast<BitCastInst>(IP)->getOperand(0) != A) ||
138 isa<DbgInfoIntrinsic>(IP) ||
139 isa<LandingPadInst>(IP))
141 return ReuseOrCreateCast(A, Ty, Op, IP);
144 // Cast the instruction immediately after the instruction.
145 Instruction *I = cast<Instruction>(V);
146 BasicBlock::iterator IP = I; ++IP;
147 if (InvokeInst *II = dyn_cast<InvokeInst>(I))
148 IP = II->getNormalDest()->begin();
149 if (CatchPadInst *CPI = dyn_cast<CatchPadInst>(I))
150 IP = CPI->getNormalDest()->begin();
151 while (isa<PHINode>(IP) || isa<LandingPadInst>(IP))
153 return ReuseOrCreateCast(I, Ty, Op, IP);
156 /// InsertBinop - Insert the specified binary operator, doing a small amount
157 /// of work to avoid inserting an obviously redundant operation.
158 Value *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode,
159 Value *LHS, Value *RHS) {
160 // Fold a binop with constant operands.
161 if (Constant *CLHS = dyn_cast<Constant>(LHS))
162 if (Constant *CRHS = dyn_cast<Constant>(RHS))
163 return ConstantExpr::get(Opcode, CLHS, CRHS);
165 // Do a quick scan to see if we have this binop nearby. If so, reuse it.
166 unsigned ScanLimit = 6;
167 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
168 // Scanning starts from the last instruction before the insertion point.
169 BasicBlock::iterator IP = Builder.GetInsertPoint();
170 if (IP != BlockBegin) {
172 for (; ScanLimit; --IP, --ScanLimit) {
173 // Don't count dbg.value against the ScanLimit, to avoid perturbing the
175 if (isa<DbgInfoIntrinsic>(IP))
177 if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(0) == LHS &&
178 IP->getOperand(1) == RHS)
180 if (IP == BlockBegin) break;
184 // Save the original insertion point so we can restore it when we're done.
185 DebugLoc Loc = Builder.GetInsertPoint()->getDebugLoc();
186 BuilderType::InsertPointGuard Guard(Builder);
188 // Move the insertion point out of as many loops as we can.
189 while (const Loop *L = SE.LI.getLoopFor(Builder.GetInsertBlock())) {
190 if (!L->isLoopInvariant(LHS) || !L->isLoopInvariant(RHS)) break;
191 BasicBlock *Preheader = L->getLoopPreheader();
192 if (!Preheader) break;
194 // Ok, move up a level.
195 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
198 // If we haven't found this binop, insert it.
199 Instruction *BO = cast<Instruction>(Builder.CreateBinOp(Opcode, LHS, RHS));
200 BO->setDebugLoc(Loc);
201 rememberInstruction(BO);
206 /// FactorOutConstant - Test if S is divisible by Factor, using signed
207 /// division. If so, update S with Factor divided out and return true.
208 /// S need not be evenly divisible if a reasonable remainder can be
210 /// TODO: When ScalarEvolution gets a SCEVSDivExpr, this can be made
211 /// unnecessary; in its place, just signed-divide Ops[i] by the scale and
212 /// check to see if the divide was folded.
213 static bool FactorOutConstant(const SCEV *&S, const SCEV *&Remainder,
214 const SCEV *Factor, ScalarEvolution &SE,
215 const DataLayout &DL) {
216 // Everything is divisible by one.
222 S = SE.getConstant(S->getType(), 1);
226 // For a Constant, check for a multiple of the given factor.
227 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
231 // Check for divisibility.
232 if (const SCEVConstant *FC = dyn_cast<SCEVConstant>(Factor)) {
234 ConstantInt::get(SE.getContext(),
235 C->getValue()->getValue().sdiv(
236 FC->getValue()->getValue()));
237 // If the quotient is zero and the remainder is non-zero, reject
238 // the value at this scale. It will be considered for subsequent
241 const SCEV *Div = SE.getConstant(CI);
244 SE.getAddExpr(Remainder,
245 SE.getConstant(C->getValue()->getValue().srem(
246 FC->getValue()->getValue())));
252 // In a Mul, check if there is a constant operand which is a multiple
253 // of the given factor.
254 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
255 // Size is known, check if there is a constant operand which is a multiple
256 // of the given factor. If so, we can factor it.
257 const SCEVConstant *FC = cast<SCEVConstant>(Factor);
258 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0)))
259 if (!C->getValue()->getValue().srem(FC->getValue()->getValue())) {
260 SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end());
261 NewMulOps[0] = SE.getConstant(
262 C->getValue()->getValue().sdiv(FC->getValue()->getValue()));
263 S = SE.getMulExpr(NewMulOps);
268 // In an AddRec, check if both start and step are divisible.
269 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
270 const SCEV *Step = A->getStepRecurrence(SE);
271 const SCEV *StepRem = SE.getConstant(Step->getType(), 0);
272 if (!FactorOutConstant(Step, StepRem, Factor, SE, DL))
274 if (!StepRem->isZero())
276 const SCEV *Start = A->getStart();
277 if (!FactorOutConstant(Start, Remainder, Factor, SE, DL))
279 S = SE.getAddRecExpr(Start, Step, A->getLoop(),
280 A->getNoWrapFlags(SCEV::FlagNW));
287 /// SimplifyAddOperands - Sort and simplify a list of add operands. NumAddRecs
288 /// is the number of SCEVAddRecExprs present, which are kept at the end of
291 static void SimplifyAddOperands(SmallVectorImpl<const SCEV *> &Ops,
293 ScalarEvolution &SE) {
294 unsigned NumAddRecs = 0;
295 for (unsigned i = Ops.size(); i > 0 && isa<SCEVAddRecExpr>(Ops[i-1]); --i)
297 // Group Ops into non-addrecs and addrecs.
298 SmallVector<const SCEV *, 8> NoAddRecs(Ops.begin(), Ops.end() - NumAddRecs);
299 SmallVector<const SCEV *, 8> AddRecs(Ops.end() - NumAddRecs, Ops.end());
300 // Let ScalarEvolution sort and simplify the non-addrecs list.
301 const SCEV *Sum = NoAddRecs.empty() ?
302 SE.getConstant(Ty, 0) :
303 SE.getAddExpr(NoAddRecs);
304 // If it returned an add, use the operands. Otherwise it simplified
305 // the sum into a single value, so just use that.
307 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Sum))
308 Ops.append(Add->op_begin(), Add->op_end());
309 else if (!Sum->isZero())
311 // Then append the addrecs.
312 Ops.append(AddRecs.begin(), AddRecs.end());
315 /// SplitAddRecs - Flatten a list of add operands, moving addrec start values
316 /// out to the top level. For example, convert {a + b,+,c} to a, b, {0,+,d}.
317 /// This helps expose more opportunities for folding parts of the expressions
318 /// into GEP indices.
320 static void SplitAddRecs(SmallVectorImpl<const SCEV *> &Ops,
322 ScalarEvolution &SE) {
324 SmallVector<const SCEV *, 8> AddRecs;
325 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
326 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Ops[i])) {
327 const SCEV *Start = A->getStart();
328 if (Start->isZero()) break;
329 const SCEV *Zero = SE.getConstant(Ty, 0);
330 AddRecs.push_back(SE.getAddRecExpr(Zero,
331 A->getStepRecurrence(SE),
333 A->getNoWrapFlags(SCEV::FlagNW)));
334 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Start)) {
336 Ops.append(Add->op_begin(), Add->op_end());
337 e += Add->getNumOperands();
342 if (!AddRecs.empty()) {
343 // Add the addrecs onto the end of the list.
344 Ops.append(AddRecs.begin(), AddRecs.end());
345 // Resort the operand list, moving any constants to the front.
346 SimplifyAddOperands(Ops, Ty, SE);
350 /// expandAddToGEP - Expand an addition expression with a pointer type into
351 /// a GEP instead of using ptrtoint+arithmetic+inttoptr. This helps
352 /// BasicAliasAnalysis and other passes analyze the result. See the rules
353 /// for getelementptr vs. inttoptr in
354 /// http://llvm.org/docs/LangRef.html#pointeraliasing
357 /// Design note: The correctness of using getelementptr here depends on
358 /// ScalarEvolution not recognizing inttoptr and ptrtoint operators, as
359 /// they may introduce pointer arithmetic which may not be safely converted
360 /// into getelementptr.
362 /// Design note: It might seem desirable for this function to be more
363 /// loop-aware. If some of the indices are loop-invariant while others
364 /// aren't, it might seem desirable to emit multiple GEPs, keeping the
365 /// loop-invariant portions of the overall computation outside the loop.
366 /// However, there are a few reasons this is not done here. Hoisting simple
367 /// arithmetic is a low-level optimization that often isn't very
368 /// important until late in the optimization process. In fact, passes
369 /// like InstructionCombining will combine GEPs, even if it means
370 /// pushing loop-invariant computation down into loops, so even if the
371 /// GEPs were split here, the work would quickly be undone. The
372 /// LoopStrengthReduction pass, which is usually run quite late (and
373 /// after the last InstructionCombining pass), takes care of hoisting
374 /// loop-invariant portions of expressions, after considering what
375 /// can be folded using target addressing modes.
377 Value *SCEVExpander::expandAddToGEP(const SCEV *const *op_begin,
378 const SCEV *const *op_end,
382 Type *OriginalElTy = PTy->getElementType();
383 Type *ElTy = OriginalElTy;
384 SmallVector<Value *, 4> GepIndices;
385 SmallVector<const SCEV *, 8> Ops(op_begin, op_end);
386 bool AnyNonZeroIndices = false;
388 // Split AddRecs up into parts as either of the parts may be usable
389 // without the other.
390 SplitAddRecs(Ops, Ty, SE);
392 Type *IntPtrTy = DL.getIntPtrType(PTy);
394 // Descend down the pointer's type and attempt to convert the other
395 // operands into GEP indices, at each level. The first index in a GEP
396 // indexes into the array implied by the pointer operand; the rest of
397 // the indices index into the element or field type selected by the
400 // If the scale size is not 0, attempt to factor out a scale for
402 SmallVector<const SCEV *, 8> ScaledOps;
403 if (ElTy->isSized()) {
404 const SCEV *ElSize = SE.getSizeOfExpr(IntPtrTy, ElTy);
405 if (!ElSize->isZero()) {
406 SmallVector<const SCEV *, 8> NewOps;
407 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
408 const SCEV *Op = Ops[i];
409 const SCEV *Remainder = SE.getConstant(Ty, 0);
410 if (FactorOutConstant(Op, Remainder, ElSize, SE, DL)) {
411 // Op now has ElSize factored out.
412 ScaledOps.push_back(Op);
413 if (!Remainder->isZero())
414 NewOps.push_back(Remainder);
415 AnyNonZeroIndices = true;
417 // The operand was not divisible, so add it to the list of operands
418 // we'll scan next iteration.
419 NewOps.push_back(Ops[i]);
422 // If we made any changes, update Ops.
423 if (!ScaledOps.empty()) {
425 SimplifyAddOperands(Ops, Ty, SE);
430 // Record the scaled array index for this level of the type. If
431 // we didn't find any operands that could be factored, tentatively
432 // assume that element zero was selected (since the zero offset
433 // would obviously be folded away).
434 Value *Scaled = ScaledOps.empty() ?
435 Constant::getNullValue(Ty) :
436 expandCodeFor(SE.getAddExpr(ScaledOps), Ty);
437 GepIndices.push_back(Scaled);
439 // Collect struct field index operands.
440 while (StructType *STy = dyn_cast<StructType>(ElTy)) {
441 bool FoundFieldNo = false;
442 // An empty struct has no fields.
443 if (STy->getNumElements() == 0) break;
444 // Field offsets are known. See if a constant offset falls within any of
445 // the struct fields.
448 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[0]))
449 if (SE.getTypeSizeInBits(C->getType()) <= 64) {
450 const StructLayout &SL = *DL.getStructLayout(STy);
451 uint64_t FullOffset = C->getValue()->getZExtValue();
452 if (FullOffset < SL.getSizeInBytes()) {
453 unsigned ElIdx = SL.getElementContainingOffset(FullOffset);
454 GepIndices.push_back(
455 ConstantInt::get(Type::getInt32Ty(Ty->getContext()), ElIdx));
456 ElTy = STy->getTypeAtIndex(ElIdx);
458 SE.getConstant(Ty, FullOffset - SL.getElementOffset(ElIdx));
459 AnyNonZeroIndices = true;
463 // If no struct field offsets were found, tentatively assume that
464 // field zero was selected (since the zero offset would obviously
467 ElTy = STy->getTypeAtIndex(0u);
468 GepIndices.push_back(
469 Constant::getNullValue(Type::getInt32Ty(Ty->getContext())));
473 if (ArrayType *ATy = dyn_cast<ArrayType>(ElTy))
474 ElTy = ATy->getElementType();
479 // If none of the operands were convertible to proper GEP indices, cast
480 // the base to i8* and do an ugly getelementptr with that. It's still
481 // better than ptrtoint+arithmetic+inttoptr at least.
482 if (!AnyNonZeroIndices) {
483 // Cast the base to i8*.
484 V = InsertNoopCastOfTo(V,
485 Type::getInt8PtrTy(Ty->getContext(), PTy->getAddressSpace()));
487 assert(!isa<Instruction>(V) ||
488 SE.DT.dominates(cast<Instruction>(V), Builder.GetInsertPoint()));
490 // Expand the operands for a plain byte offset.
491 Value *Idx = expandCodeFor(SE.getAddExpr(Ops), Ty);
493 // Fold a GEP with constant operands.
494 if (Constant *CLHS = dyn_cast<Constant>(V))
495 if (Constant *CRHS = dyn_cast<Constant>(Idx))
496 return ConstantExpr::getGetElementPtr(Type::getInt8Ty(Ty->getContext()),
499 // Do a quick scan to see if we have this GEP nearby. If so, reuse it.
500 unsigned ScanLimit = 6;
501 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
502 // Scanning starts from the last instruction before the insertion point.
503 BasicBlock::iterator IP = Builder.GetInsertPoint();
504 if (IP != BlockBegin) {
506 for (; ScanLimit; --IP, --ScanLimit) {
507 // Don't count dbg.value against the ScanLimit, to avoid perturbing the
509 if (isa<DbgInfoIntrinsic>(IP))
511 if (IP->getOpcode() == Instruction::GetElementPtr &&
512 IP->getOperand(0) == V && IP->getOperand(1) == Idx)
514 if (IP == BlockBegin) break;
518 // Save the original insertion point so we can restore it when we're done.
519 BuilderType::InsertPointGuard Guard(Builder);
521 // Move the insertion point out of as many loops as we can.
522 while (const Loop *L = SE.LI.getLoopFor(Builder.GetInsertBlock())) {
523 if (!L->isLoopInvariant(V) || !L->isLoopInvariant(Idx)) break;
524 BasicBlock *Preheader = L->getLoopPreheader();
525 if (!Preheader) break;
527 // Ok, move up a level.
528 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
532 Value *GEP = Builder.CreateGEP(Builder.getInt8Ty(), V, Idx, "uglygep");
533 rememberInstruction(GEP);
538 // Save the original insertion point so we can restore it when we're done.
539 BuilderType::InsertPoint SaveInsertPt = Builder.saveIP();
541 // Move the insertion point out of as many loops as we can.
542 while (const Loop *L = SE.LI.getLoopFor(Builder.GetInsertBlock())) {
543 if (!L->isLoopInvariant(V)) break;
545 bool AnyIndexNotLoopInvariant = false;
546 for (SmallVectorImpl<Value *>::const_iterator I = GepIndices.begin(),
547 E = GepIndices.end(); I != E; ++I)
548 if (!L->isLoopInvariant(*I)) {
549 AnyIndexNotLoopInvariant = true;
552 if (AnyIndexNotLoopInvariant)
555 BasicBlock *Preheader = L->getLoopPreheader();
556 if (!Preheader) break;
558 // Ok, move up a level.
559 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
562 // Insert a pretty getelementptr. Note that this GEP is not marked inbounds,
563 // because ScalarEvolution may have changed the address arithmetic to
564 // compute a value which is beyond the end of the allocated object.
566 if (V->getType() != PTy)
567 Casted = InsertNoopCastOfTo(Casted, PTy);
568 Value *GEP = Builder.CreateGEP(OriginalElTy, Casted,
571 Ops.push_back(SE.getUnknown(GEP));
572 rememberInstruction(GEP);
574 // Restore the original insert point.
575 Builder.restoreIP(SaveInsertPt);
577 return expand(SE.getAddExpr(Ops));
580 /// PickMostRelevantLoop - Given two loops pick the one that's most relevant for
581 /// SCEV expansion. If they are nested, this is the most nested. If they are
582 /// neighboring, pick the later.
583 static const Loop *PickMostRelevantLoop(const Loop *A, const Loop *B,
587 if (A->contains(B)) return B;
588 if (B->contains(A)) return A;
589 if (DT.dominates(A->getHeader(), B->getHeader())) return B;
590 if (DT.dominates(B->getHeader(), A->getHeader())) return A;
591 return A; // Arbitrarily break the tie.
594 /// getRelevantLoop - Get the most relevant loop associated with the given
595 /// expression, according to PickMostRelevantLoop.
596 const Loop *SCEVExpander::getRelevantLoop(const SCEV *S) {
597 // Test whether we've already computed the most relevant loop for this SCEV.
598 std::pair<DenseMap<const SCEV *, const Loop *>::iterator, bool> Pair =
599 RelevantLoops.insert(std::make_pair(S, nullptr));
601 return Pair.first->second;
603 if (isa<SCEVConstant>(S))
604 // A constant has no relevant loops.
606 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
607 if (const Instruction *I = dyn_cast<Instruction>(U->getValue()))
608 return Pair.first->second = SE.LI.getLoopFor(I->getParent());
609 // A non-instruction has no relevant loops.
612 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S)) {
613 const Loop *L = nullptr;
614 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
616 for (SCEVNAryExpr::op_iterator I = N->op_begin(), E = N->op_end();
618 L = PickMostRelevantLoop(L, getRelevantLoop(*I), SE.DT);
619 return RelevantLoops[N] = L;
621 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S)) {
622 const Loop *Result = getRelevantLoop(C->getOperand());
623 return RelevantLoops[C] = Result;
625 if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
626 const Loop *Result = PickMostRelevantLoop(
627 getRelevantLoop(D->getLHS()), getRelevantLoop(D->getRHS()), SE.DT);
628 return RelevantLoops[D] = Result;
630 llvm_unreachable("Unexpected SCEV type!");
635 /// LoopCompare - Compare loops by PickMostRelevantLoop.
639 explicit LoopCompare(DominatorTree &dt) : DT(dt) {}
641 bool operator()(std::pair<const Loop *, const SCEV *> LHS,
642 std::pair<const Loop *, const SCEV *> RHS) const {
643 // Keep pointer operands sorted at the end.
644 if (LHS.second->getType()->isPointerTy() !=
645 RHS.second->getType()->isPointerTy())
646 return LHS.second->getType()->isPointerTy();
648 // Compare loops with PickMostRelevantLoop.
649 if (LHS.first != RHS.first)
650 return PickMostRelevantLoop(LHS.first, RHS.first, DT) != LHS.first;
652 // If one operand is a non-constant negative and the other is not,
653 // put the non-constant negative on the right so that a sub can
654 // be used instead of a negate and add.
655 if (LHS.second->isNonConstantNegative()) {
656 if (!RHS.second->isNonConstantNegative())
658 } else if (RHS.second->isNonConstantNegative())
661 // Otherwise they are equivalent according to this comparison.
668 Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) {
669 Type *Ty = SE.getEffectiveSCEVType(S->getType());
671 // Collect all the add operands in a loop, along with their associated loops.
672 // Iterate in reverse so that constants are emitted last, all else equal, and
673 // so that pointer operands are inserted first, which the code below relies on
674 // to form more involved GEPs.
675 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops;
676 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(S->op_end()),
677 E(S->op_begin()); I != E; ++I)
678 OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I));
680 // Sort by loop. Use a stable sort so that constants follow non-constants and
681 // pointer operands precede non-pointer operands.
682 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(SE.DT));
684 // Emit instructions to add all the operands. Hoist as much as possible
685 // out of loops, and form meaningful getelementptrs where possible.
686 Value *Sum = nullptr;
687 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator
688 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) {
689 const Loop *CurLoop = I->first;
690 const SCEV *Op = I->second;
692 // This is the first operand. Just expand it.
695 } else if (PointerType *PTy = dyn_cast<PointerType>(Sum->getType())) {
696 // The running sum expression is a pointer. Try to form a getelementptr
697 // at this level with that as the base.
698 SmallVector<const SCEV *, 4> NewOps;
699 for (; I != E && I->first == CurLoop; ++I) {
700 // If the operand is SCEVUnknown and not instructions, peek through
701 // it, to enable more of it to be folded into the GEP.
702 const SCEV *X = I->second;
703 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(X))
704 if (!isa<Instruction>(U->getValue()))
705 X = SE.getSCEV(U->getValue());
708 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, Sum);
709 } else if (PointerType *PTy = dyn_cast<PointerType>(Op->getType())) {
710 // The running sum is an integer, and there's a pointer at this level.
711 // Try to form a getelementptr. If the running sum is instructions,
712 // use a SCEVUnknown to avoid re-analyzing them.
713 SmallVector<const SCEV *, 4> NewOps;
714 NewOps.push_back(isa<Instruction>(Sum) ? SE.getUnknown(Sum) :
716 for (++I; I != E && I->first == CurLoop; ++I)
717 NewOps.push_back(I->second);
718 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, expand(Op));
719 } else if (Op->isNonConstantNegative()) {
720 // Instead of doing a negate and add, just do a subtract.
721 Value *W = expandCodeFor(SE.getNegativeSCEV(Op), Ty);
722 Sum = InsertNoopCastOfTo(Sum, Ty);
723 Sum = InsertBinop(Instruction::Sub, Sum, W);
727 Value *W = expandCodeFor(Op, Ty);
728 Sum = InsertNoopCastOfTo(Sum, Ty);
729 // Canonicalize a constant to the RHS.
730 if (isa<Constant>(Sum)) std::swap(Sum, W);
731 Sum = InsertBinop(Instruction::Add, Sum, W);
739 Value *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) {
740 Type *Ty = SE.getEffectiveSCEVType(S->getType());
742 // Collect all the mul operands in a loop, along with their associated loops.
743 // Iterate in reverse so that constants are emitted last, all else equal.
744 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops;
745 for (std::reverse_iterator<SCEVMulExpr::op_iterator> I(S->op_end()),
746 E(S->op_begin()); I != E; ++I)
747 OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I));
749 // Sort by loop. Use a stable sort so that constants follow non-constants.
750 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(SE.DT));
752 // Emit instructions to mul all the operands. Hoist as much as possible
754 Value *Prod = nullptr;
755 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator
756 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ++I) {
757 const SCEV *Op = I->second;
759 // This is the first operand. Just expand it.
761 } else if (Op->isAllOnesValue()) {
762 // Instead of doing a multiply by negative one, just do a negate.
763 Prod = InsertNoopCastOfTo(Prod, Ty);
764 Prod = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), Prod);
767 Value *W = expandCodeFor(Op, Ty);
768 Prod = InsertNoopCastOfTo(Prod, Ty);
769 // Canonicalize a constant to the RHS.
770 if (isa<Constant>(Prod)) std::swap(Prod, W);
772 if (match(W, m_Power2(RHS))) {
773 // Canonicalize Prod*(1<<C) to Prod<<C.
774 assert(!Ty->isVectorTy() && "vector types are not SCEVable");
775 Prod = InsertBinop(Instruction::Shl, Prod,
776 ConstantInt::get(Ty, RHS->logBase2()));
778 Prod = InsertBinop(Instruction::Mul, Prod, W);
786 Value *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) {
787 Type *Ty = SE.getEffectiveSCEVType(S->getType());
789 Value *LHS = expandCodeFor(S->getLHS(), Ty);
790 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getRHS())) {
791 const APInt &RHS = SC->getValue()->getValue();
792 if (RHS.isPowerOf2())
793 return InsertBinop(Instruction::LShr, LHS,
794 ConstantInt::get(Ty, RHS.logBase2()));
797 Value *RHS = expandCodeFor(S->getRHS(), Ty);
798 return InsertBinop(Instruction::UDiv, LHS, RHS);
801 /// Move parts of Base into Rest to leave Base with the minimal
802 /// expression that provides a pointer operand suitable for a
804 static void ExposePointerBase(const SCEV *&Base, const SCEV *&Rest,
805 ScalarEvolution &SE) {
806 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Base)) {
807 Base = A->getStart();
808 Rest = SE.getAddExpr(Rest,
809 SE.getAddRecExpr(SE.getConstant(A->getType(), 0),
810 A->getStepRecurrence(SE),
812 A->getNoWrapFlags(SCEV::FlagNW)));
814 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(Base)) {
815 Base = A->getOperand(A->getNumOperands()-1);
816 SmallVector<const SCEV *, 8> NewAddOps(A->op_begin(), A->op_end());
817 NewAddOps.back() = Rest;
818 Rest = SE.getAddExpr(NewAddOps);
819 ExposePointerBase(Base, Rest, SE);
823 /// Determine if this is a well-behaved chain of instructions leading back to
824 /// the PHI. If so, it may be reused by expanded expressions.
825 bool SCEVExpander::isNormalAddRecExprPHI(PHINode *PN, Instruction *IncV,
827 if (IncV->getNumOperands() == 0 || isa<PHINode>(IncV) ||
828 (isa<CastInst>(IncV) && !isa<BitCastInst>(IncV)))
830 // If any of the operands don't dominate the insert position, bail.
831 // Addrec operands are always loop-invariant, so this can only happen
832 // if there are instructions which haven't been hoisted.
833 if (L == IVIncInsertLoop) {
834 for (User::op_iterator OI = IncV->op_begin()+1,
835 OE = IncV->op_end(); OI != OE; ++OI)
836 if (Instruction *OInst = dyn_cast<Instruction>(OI))
837 if (!SE.DT.dominates(OInst, IVIncInsertPos))
840 // Advance to the next instruction.
841 IncV = dyn_cast<Instruction>(IncV->getOperand(0));
845 if (IncV->mayHaveSideEffects())
851 return isNormalAddRecExprPHI(PN, IncV, L);
854 /// getIVIncOperand returns an induction variable increment's induction
855 /// variable operand.
857 /// If allowScale is set, any type of GEP is allowed as long as the nonIV
858 /// operands dominate InsertPos.
860 /// If allowScale is not set, ensure that a GEP increment conforms to one of the
861 /// simple patterns generated by getAddRecExprPHILiterally and
862 /// expandAddtoGEP. If the pattern isn't recognized, return NULL.
863 Instruction *SCEVExpander::getIVIncOperand(Instruction *IncV,
864 Instruction *InsertPos,
866 if (IncV == InsertPos)
869 switch (IncV->getOpcode()) {
872 // Check for a simple Add/Sub or GEP of a loop invariant step.
873 case Instruction::Add:
874 case Instruction::Sub: {
875 Instruction *OInst = dyn_cast<Instruction>(IncV->getOperand(1));
876 if (!OInst || SE.DT.dominates(OInst, InsertPos))
877 return dyn_cast<Instruction>(IncV->getOperand(0));
880 case Instruction::BitCast:
881 return dyn_cast<Instruction>(IncV->getOperand(0));
882 case Instruction::GetElementPtr:
883 for (Instruction::op_iterator I = IncV->op_begin()+1, E = IncV->op_end();
885 if (isa<Constant>(*I))
887 if (Instruction *OInst = dyn_cast<Instruction>(*I)) {
888 if (!SE.DT.dominates(OInst, InsertPos))
892 // allow any kind of GEP as long as it can be hoisted.
895 // This must be a pointer addition of constants (pretty), which is already
896 // handled, or some number of address-size elements (ugly). Ugly geps
897 // have 2 operands. i1* is used by the expander to represent an
898 // address-size element.
899 if (IncV->getNumOperands() != 2)
901 unsigned AS = cast<PointerType>(IncV->getType())->getAddressSpace();
902 if (IncV->getType() != Type::getInt1PtrTy(SE.getContext(), AS)
903 && IncV->getType() != Type::getInt8PtrTy(SE.getContext(), AS))
907 return dyn_cast<Instruction>(IncV->getOperand(0));
911 /// hoistStep - Attempt to hoist a simple IV increment above InsertPos to make
912 /// it available to other uses in this loop. Recursively hoist any operands,
913 /// until we reach a value that dominates InsertPos.
914 bool SCEVExpander::hoistIVInc(Instruction *IncV, Instruction *InsertPos) {
915 if (SE.DT.dominates(IncV, InsertPos))
918 // InsertPos must itself dominate IncV so that IncV's new position satisfies
919 // its existing users.
920 if (isa<PHINode>(InsertPos) ||
921 !SE.DT.dominates(InsertPos->getParent(), IncV->getParent()))
924 // Check that the chain of IV operands leading back to Phi can be hoisted.
925 SmallVector<Instruction*, 4> IVIncs;
927 Instruction *Oper = getIVIncOperand(IncV, InsertPos, /*allowScale*/true);
930 // IncV is safe to hoist.
931 IVIncs.push_back(IncV);
933 if (SE.DT.dominates(IncV, InsertPos))
936 for (SmallVectorImpl<Instruction*>::reverse_iterator I = IVIncs.rbegin(),
937 E = IVIncs.rend(); I != E; ++I) {
938 (*I)->moveBefore(InsertPos);
943 /// Determine if this cyclic phi is in a form that would have been generated by
944 /// LSR. We don't care if the phi was actually expanded in this pass, as long
945 /// as it is in a low-cost form, for example, no implied multiplication. This
946 /// should match any patterns generated by getAddRecExprPHILiterally and
948 bool SCEVExpander::isExpandedAddRecExprPHI(PHINode *PN, Instruction *IncV,
950 for(Instruction *IVOper = IncV;
951 (IVOper = getIVIncOperand(IVOper, L->getLoopPreheader()->getTerminator(),
952 /*allowScale=*/false));) {
959 /// expandIVInc - Expand an IV increment at Builder's current InsertPos.
960 /// Typically this is the LatchBlock terminator or IVIncInsertPos, but we may
961 /// need to materialize IV increments elsewhere to handle difficult situations.
962 Value *SCEVExpander::expandIVInc(PHINode *PN, Value *StepV, const Loop *L,
963 Type *ExpandTy, Type *IntTy,
966 // If the PHI is a pointer, use a GEP, otherwise use an add or sub.
967 if (ExpandTy->isPointerTy()) {
968 PointerType *GEPPtrTy = cast<PointerType>(ExpandTy);
969 // If the step isn't constant, don't use an implicitly scaled GEP, because
970 // that would require a multiply inside the loop.
971 if (!isa<ConstantInt>(StepV))
972 GEPPtrTy = PointerType::get(Type::getInt1Ty(SE.getContext()),
973 GEPPtrTy->getAddressSpace());
974 const SCEV *const StepArray[1] = { SE.getSCEV(StepV) };
975 IncV = expandAddToGEP(StepArray, StepArray+1, GEPPtrTy, IntTy, PN);
976 if (IncV->getType() != PN->getType()) {
977 IncV = Builder.CreateBitCast(IncV, PN->getType());
978 rememberInstruction(IncV);
982 Builder.CreateSub(PN, StepV, Twine(IVName) + ".iv.next") :
983 Builder.CreateAdd(PN, StepV, Twine(IVName) + ".iv.next");
984 rememberInstruction(IncV);
989 /// \brief Hoist the addrec instruction chain rooted in the loop phi above the
990 /// position. This routine assumes that this is possible (has been checked).
991 static void hoistBeforePos(DominatorTree *DT, Instruction *InstToHoist,
992 Instruction *Pos, PHINode *LoopPhi) {
994 if (DT->dominates(InstToHoist, Pos))
996 // Make sure the increment is where we want it. But don't move it
997 // down past a potential existing post-inc user.
998 InstToHoist->moveBefore(Pos);
1000 InstToHoist = cast<Instruction>(InstToHoist->getOperand(0));
1001 } while (InstToHoist != LoopPhi);
1004 /// \brief Check whether we can cheaply express the requested SCEV in terms of
1005 /// the available PHI SCEV by truncation and/or inversion of the step.
1006 static bool canBeCheaplyTransformed(ScalarEvolution &SE,
1007 const SCEVAddRecExpr *Phi,
1008 const SCEVAddRecExpr *Requested,
1010 Type *PhiTy = SE.getEffectiveSCEVType(Phi->getType());
1011 Type *RequestedTy = SE.getEffectiveSCEVType(Requested->getType());
1013 if (RequestedTy->getIntegerBitWidth() > PhiTy->getIntegerBitWidth())
1016 // Try truncate it if necessary.
1017 Phi = dyn_cast<SCEVAddRecExpr>(SE.getTruncateOrNoop(Phi, RequestedTy));
1021 // Check whether truncation will help.
1022 if (Phi == Requested) {
1027 // Check whether inverting will help: {R,+,-1} == R - {0,+,1}.
1028 if (SE.getAddExpr(Requested->getStart(),
1029 SE.getNegativeSCEV(Requested)) == Phi) {
1037 static bool IsIncrementNSW(ScalarEvolution &SE, const SCEVAddRecExpr *AR) {
1038 if (!isa<IntegerType>(AR->getType()))
1041 unsigned BitWidth = cast<IntegerType>(AR->getType())->getBitWidth();
1042 Type *WideTy = IntegerType::get(AR->getType()->getContext(), BitWidth * 2);
1043 const SCEV *Step = AR->getStepRecurrence(SE);
1044 const SCEV *OpAfterExtend = SE.getAddExpr(SE.getSignExtendExpr(Step, WideTy),
1045 SE.getSignExtendExpr(AR, WideTy));
1046 const SCEV *ExtendAfterOp =
1047 SE.getSignExtendExpr(SE.getAddExpr(AR, Step), WideTy);
1048 return ExtendAfterOp == OpAfterExtend;
1051 static bool IsIncrementNUW(ScalarEvolution &SE, const SCEVAddRecExpr *AR) {
1052 if (!isa<IntegerType>(AR->getType()))
1055 unsigned BitWidth = cast<IntegerType>(AR->getType())->getBitWidth();
1056 Type *WideTy = IntegerType::get(AR->getType()->getContext(), BitWidth * 2);
1057 const SCEV *Step = AR->getStepRecurrence(SE);
1058 const SCEV *OpAfterExtend = SE.getAddExpr(SE.getZeroExtendExpr(Step, WideTy),
1059 SE.getZeroExtendExpr(AR, WideTy));
1060 const SCEV *ExtendAfterOp =
1061 SE.getZeroExtendExpr(SE.getAddExpr(AR, Step), WideTy);
1062 return ExtendAfterOp == OpAfterExtend;
1065 /// getAddRecExprPHILiterally - Helper for expandAddRecExprLiterally. Expand
1066 /// the base addrec, which is the addrec without any non-loop-dominating
1067 /// values, and return the PHI.
1069 SCEVExpander::getAddRecExprPHILiterally(const SCEVAddRecExpr *Normalized,
1075 assert((!IVIncInsertLoop||IVIncInsertPos) && "Uninitialized insert position");
1077 // Reuse a previously-inserted PHI, if present.
1078 BasicBlock *LatchBlock = L->getLoopLatch();
1080 PHINode *AddRecPhiMatch = nullptr;
1081 Instruction *IncV = nullptr;
1085 // Only try partially matching scevs that need truncation and/or
1086 // step-inversion if we know this loop is outside the current loop.
1087 bool TryNonMatchingSCEV =
1089 SE.DT.properlyDominates(LatchBlock, IVIncInsertLoop->getHeader());
1091 for (BasicBlock::iterator I = L->getHeader()->begin();
1092 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1093 if (!SE.isSCEVable(PN->getType()))
1096 const SCEVAddRecExpr *PhiSCEV = dyn_cast<SCEVAddRecExpr>(SE.getSCEV(PN));
1100 bool IsMatchingSCEV = PhiSCEV == Normalized;
1101 // We only handle truncation and inversion of phi recurrences for the
1102 // expanded expression if the expanded expression's loop dominates the
1103 // loop we insert to. Check now, so we can bail out early.
1104 if (!IsMatchingSCEV && !TryNonMatchingSCEV)
1107 Instruction *TempIncV =
1108 cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock));
1110 // Check whether we can reuse this PHI node.
1112 if (!isExpandedAddRecExprPHI(PN, TempIncV, L))
1114 if (L == IVIncInsertLoop && !hoistIVInc(TempIncV, IVIncInsertPos))
1117 if (!isNormalAddRecExprPHI(PN, TempIncV, L))
1121 // Stop if we have found an exact match SCEV.
1122 if (IsMatchingSCEV) {
1126 AddRecPhiMatch = PN;
1130 // Try whether the phi can be translated into the requested form
1131 // (truncated and/or offset by a constant).
1132 if ((!TruncTy || InvertStep) &&
1133 canBeCheaplyTransformed(SE, PhiSCEV, Normalized, InvertStep)) {
1134 // Record the phi node. But don't stop we might find an exact match
1136 AddRecPhiMatch = PN;
1138 TruncTy = SE.getEffectiveSCEVType(Normalized->getType());
1142 if (AddRecPhiMatch) {
1143 // Potentially, move the increment. We have made sure in
1144 // isExpandedAddRecExprPHI or hoistIVInc that this is possible.
1145 if (L == IVIncInsertLoop)
1146 hoistBeforePos(&SE.DT, IncV, IVIncInsertPos, AddRecPhiMatch);
1148 // Ok, the add recurrence looks usable.
1149 // Remember this PHI, even in post-inc mode.
1150 InsertedValues.insert(AddRecPhiMatch);
1151 // Remember the increment.
1152 rememberInstruction(IncV);
1153 return AddRecPhiMatch;
1157 // Save the original insertion point so we can restore it when we're done.
1158 BuilderType::InsertPointGuard Guard(Builder);
1160 // Another AddRec may need to be recursively expanded below. For example, if
1161 // this AddRec is quadratic, the StepV may itself be an AddRec in this
1162 // loop. Remove this loop from the PostIncLoops set before expanding such
1163 // AddRecs. Otherwise, we cannot find a valid position for the step
1164 // (i.e. StepV can never dominate its loop header). Ideally, we could do
1165 // SavedIncLoops.swap(PostIncLoops), but we generally have a single element,
1166 // so it's not worth implementing SmallPtrSet::swap.
1167 PostIncLoopSet SavedPostIncLoops = PostIncLoops;
1168 PostIncLoops.clear();
1170 // Expand code for the start value.
1171 Value *StartV = expandCodeFor(Normalized->getStart(), ExpandTy,
1172 L->getHeader()->begin());
1174 // StartV must be hoisted into L's preheader to dominate the new phi.
1175 assert(!isa<Instruction>(StartV) ||
1176 SE.DT.properlyDominates(cast<Instruction>(StartV)->getParent(),
1179 // Expand code for the step value. Do this before creating the PHI so that PHI
1180 // reuse code doesn't see an incomplete PHI.
1181 const SCEV *Step = Normalized->getStepRecurrence(SE);
1182 // If the stride is negative, insert a sub instead of an add for the increment
1183 // (unless it's a constant, because subtracts of constants are canonicalized
1185 bool useSubtract = !ExpandTy->isPointerTy() && Step->isNonConstantNegative();
1187 Step = SE.getNegativeSCEV(Step);
1188 // Expand the step somewhere that dominates the loop header.
1189 Value *StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin());
1191 // The no-wrap behavior proved by IsIncrement(NUW|NSW) is only applicable if
1192 // we actually do emit an addition. It does not apply if we emit a
1194 bool IncrementIsNUW = !useSubtract && IsIncrementNUW(SE, Normalized);
1195 bool IncrementIsNSW = !useSubtract && IsIncrementNSW(SE, Normalized);
1198 BasicBlock *Header = L->getHeader();
1199 Builder.SetInsertPoint(Header, Header->begin());
1200 pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header);
1201 PHINode *PN = Builder.CreatePHI(ExpandTy, std::distance(HPB, HPE),
1202 Twine(IVName) + ".iv");
1203 rememberInstruction(PN);
1205 // Create the step instructions and populate the PHI.
1206 for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) {
1207 BasicBlock *Pred = *HPI;
1209 // Add a start value.
1210 if (!L->contains(Pred)) {
1211 PN->addIncoming(StartV, Pred);
1215 // Create a step value and add it to the PHI.
1216 // If IVIncInsertLoop is non-null and equal to the addrec's loop, insert the
1217 // instructions at IVIncInsertPos.
1218 Instruction *InsertPos = L == IVIncInsertLoop ?
1219 IVIncInsertPos : Pred->getTerminator();
1220 Builder.SetInsertPoint(InsertPos);
1221 Value *IncV = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract);
1223 if (isa<OverflowingBinaryOperator>(IncV)) {
1225 cast<BinaryOperator>(IncV)->setHasNoUnsignedWrap();
1227 cast<BinaryOperator>(IncV)->setHasNoSignedWrap();
1229 PN->addIncoming(IncV, Pred);
1232 // After expanding subexpressions, restore the PostIncLoops set so the caller
1233 // can ensure that IVIncrement dominates the current uses.
1234 PostIncLoops = SavedPostIncLoops;
1236 // Remember this PHI, even in post-inc mode.
1237 InsertedValues.insert(PN);
1242 Value *SCEVExpander::expandAddRecExprLiterally(const SCEVAddRecExpr *S) {
1243 Type *STy = S->getType();
1244 Type *IntTy = SE.getEffectiveSCEVType(STy);
1245 const Loop *L = S->getLoop();
1247 // Determine a normalized form of this expression, which is the expression
1248 // before any post-inc adjustment is made.
1249 const SCEVAddRecExpr *Normalized = S;
1250 if (PostIncLoops.count(L)) {
1251 PostIncLoopSet Loops;
1253 Normalized = cast<SCEVAddRecExpr>(TransformForPostIncUse(
1254 Normalize, S, nullptr, nullptr, Loops, SE, SE.DT));
1257 // Strip off any non-loop-dominating component from the addrec start.
1258 const SCEV *Start = Normalized->getStart();
1259 const SCEV *PostLoopOffset = nullptr;
1260 if (!SE.properlyDominates(Start, L->getHeader())) {
1261 PostLoopOffset = Start;
1262 Start = SE.getConstant(Normalized->getType(), 0);
1263 Normalized = cast<SCEVAddRecExpr>(
1264 SE.getAddRecExpr(Start, Normalized->getStepRecurrence(SE),
1265 Normalized->getLoop(),
1266 Normalized->getNoWrapFlags(SCEV::FlagNW)));
1269 // Strip off any non-loop-dominating component from the addrec step.
1270 const SCEV *Step = Normalized->getStepRecurrence(SE);
1271 const SCEV *PostLoopScale = nullptr;
1272 if (!SE.dominates(Step, L->getHeader())) {
1273 PostLoopScale = Step;
1274 Step = SE.getConstant(Normalized->getType(), 1);
1276 cast<SCEVAddRecExpr>(SE.getAddRecExpr(
1277 Start, Step, Normalized->getLoop(),
1278 Normalized->getNoWrapFlags(SCEV::FlagNW)));
1281 // Expand the core addrec. If we need post-loop scaling, force it to
1282 // expand to an integer type to avoid the need for additional casting.
1283 Type *ExpandTy = PostLoopScale ? IntTy : STy;
1284 // In some cases, we decide to reuse an existing phi node but need to truncate
1285 // it and/or invert the step.
1286 Type *TruncTy = nullptr;
1287 bool InvertStep = false;
1288 PHINode *PN = getAddRecExprPHILiterally(Normalized, L, ExpandTy, IntTy,
1289 TruncTy, InvertStep);
1291 // Accommodate post-inc mode, if necessary.
1293 if (!PostIncLoops.count(L))
1296 // In PostInc mode, use the post-incremented value.
1297 BasicBlock *LatchBlock = L->getLoopLatch();
1298 assert(LatchBlock && "PostInc mode requires a unique loop latch!");
1299 Result = PN->getIncomingValueForBlock(LatchBlock);
1301 // For an expansion to use the postinc form, the client must call
1302 // expandCodeFor with an InsertPoint that is either outside the PostIncLoop
1303 // or dominated by IVIncInsertPos.
1304 if (isa<Instruction>(Result) &&
1305 !SE.DT.dominates(cast<Instruction>(Result), Builder.GetInsertPoint())) {
1306 // The induction variable's postinc expansion does not dominate this use.
1307 // IVUsers tries to prevent this case, so it is rare. However, it can
1308 // happen when an IVUser outside the loop is not dominated by the latch
1309 // block. Adjusting IVIncInsertPos before expansion begins cannot handle
1310 // all cases. Consider a phi outide whose operand is replaced during
1311 // expansion with the value of the postinc user. Without fundamentally
1312 // changing the way postinc users are tracked, the only remedy is
1313 // inserting an extra IV increment. StepV might fold into PostLoopOffset,
1314 // but hopefully expandCodeFor handles that.
1316 !ExpandTy->isPointerTy() && Step->isNonConstantNegative();
1318 Step = SE.getNegativeSCEV(Step);
1321 // Expand the step somewhere that dominates the loop header.
1322 BuilderType::InsertPointGuard Guard(Builder);
1323 StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin());
1325 Result = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract);
1329 // We have decided to reuse an induction variable of a dominating loop. Apply
1330 // truncation and/or invertion of the step.
1332 Type *ResTy = Result->getType();
1333 // Normalize the result type.
1334 if (ResTy != SE.getEffectiveSCEVType(ResTy))
1335 Result = InsertNoopCastOfTo(Result, SE.getEffectiveSCEVType(ResTy));
1336 // Truncate the result.
1337 if (TruncTy != Result->getType()) {
1338 Result = Builder.CreateTrunc(Result, TruncTy);
1339 rememberInstruction(Result);
1341 // Invert the result.
1343 Result = Builder.CreateSub(expandCodeFor(Normalized->getStart(), TruncTy),
1345 rememberInstruction(Result);
1349 // Re-apply any non-loop-dominating scale.
1350 if (PostLoopScale) {
1351 assert(S->isAffine() && "Can't linearly scale non-affine recurrences.");
1352 Result = InsertNoopCastOfTo(Result, IntTy);
1353 Result = Builder.CreateMul(Result,
1354 expandCodeFor(PostLoopScale, IntTy));
1355 rememberInstruction(Result);
1358 // Re-apply any non-loop-dominating offset.
1359 if (PostLoopOffset) {
1360 if (PointerType *PTy = dyn_cast<PointerType>(ExpandTy)) {
1361 const SCEV *const OffsetArray[1] = { PostLoopOffset };
1362 Result = expandAddToGEP(OffsetArray, OffsetArray+1, PTy, IntTy, Result);
1364 Result = InsertNoopCastOfTo(Result, IntTy);
1365 Result = Builder.CreateAdd(Result,
1366 expandCodeFor(PostLoopOffset, IntTy));
1367 rememberInstruction(Result);
1374 Value *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) {
1375 if (!CanonicalMode) return expandAddRecExprLiterally(S);
1377 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1378 const Loop *L = S->getLoop();
1380 // First check for an existing canonical IV in a suitable type.
1381 PHINode *CanonicalIV = nullptr;
1382 if (PHINode *PN = L->getCanonicalInductionVariable())
1383 if (SE.getTypeSizeInBits(PN->getType()) >= SE.getTypeSizeInBits(Ty))
1386 // Rewrite an AddRec in terms of the canonical induction variable, if
1387 // its type is more narrow.
1389 SE.getTypeSizeInBits(CanonicalIV->getType()) >
1390 SE.getTypeSizeInBits(Ty)) {
1391 SmallVector<const SCEV *, 4> NewOps(S->getNumOperands());
1392 for (unsigned i = 0, e = S->getNumOperands(); i != e; ++i)
1393 NewOps[i] = SE.getAnyExtendExpr(S->op_begin()[i], CanonicalIV->getType());
1394 Value *V = expand(SE.getAddRecExpr(NewOps, S->getLoop(),
1395 S->getNoWrapFlags(SCEV::FlagNW)));
1396 BasicBlock::iterator NewInsertPt =
1397 std::next(BasicBlock::iterator(cast<Instruction>(V)));
1398 BuilderType::InsertPointGuard Guard(Builder);
1399 while (isa<PHINode>(NewInsertPt) || isa<DbgInfoIntrinsic>(NewInsertPt) ||
1400 isa<LandingPadInst>(NewInsertPt))
1402 V = expandCodeFor(SE.getTruncateExpr(SE.getUnknown(V), Ty), nullptr,
1407 // {X,+,F} --> X + {0,+,F}
1408 if (!S->getStart()->isZero()) {
1409 SmallVector<const SCEV *, 4> NewOps(S->op_begin(), S->op_end());
1410 NewOps[0] = SE.getConstant(Ty, 0);
1411 const SCEV *Rest = SE.getAddRecExpr(NewOps, L,
1412 S->getNoWrapFlags(SCEV::FlagNW));
1414 // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the
1415 // comments on expandAddToGEP for details.
1416 const SCEV *Base = S->getStart();
1417 const SCEV *RestArray[1] = { Rest };
1418 // Dig into the expression to find the pointer base for a GEP.
1419 ExposePointerBase(Base, RestArray[0], SE);
1420 // If we found a pointer, expand the AddRec with a GEP.
1421 if (PointerType *PTy = dyn_cast<PointerType>(Base->getType())) {
1422 // Make sure the Base isn't something exotic, such as a multiplied
1423 // or divided pointer value. In those cases, the result type isn't
1424 // actually a pointer type.
1425 if (!isa<SCEVMulExpr>(Base) && !isa<SCEVUDivExpr>(Base)) {
1426 Value *StartV = expand(Base);
1427 assert(StartV->getType() == PTy && "Pointer type mismatch for GEP!");
1428 return expandAddToGEP(RestArray, RestArray+1, PTy, Ty, StartV);
1432 // Just do a normal add. Pre-expand the operands to suppress folding.
1433 return expand(SE.getAddExpr(SE.getUnknown(expand(S->getStart())),
1434 SE.getUnknown(expand(Rest))));
1437 // If we don't yet have a canonical IV, create one.
1439 // Create and insert the PHI node for the induction variable in the
1441 BasicBlock *Header = L->getHeader();
1442 pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header);
1443 CanonicalIV = PHINode::Create(Ty, std::distance(HPB, HPE), "indvar",
1445 rememberInstruction(CanonicalIV);
1447 SmallSet<BasicBlock *, 4> PredSeen;
1448 Constant *One = ConstantInt::get(Ty, 1);
1449 for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) {
1450 BasicBlock *HP = *HPI;
1451 if (!PredSeen.insert(HP).second) {
1452 // There must be an incoming value for each predecessor, even the
1454 CanonicalIV->addIncoming(CanonicalIV->getIncomingValueForBlock(HP), HP);
1458 if (L->contains(HP)) {
1459 // Insert a unit add instruction right before the terminator
1460 // corresponding to the back-edge.
1461 Instruction *Add = BinaryOperator::CreateAdd(CanonicalIV, One,
1463 HP->getTerminator());
1464 Add->setDebugLoc(HP->getTerminator()->getDebugLoc());
1465 rememberInstruction(Add);
1466 CanonicalIV->addIncoming(Add, HP);
1468 CanonicalIV->addIncoming(Constant::getNullValue(Ty), HP);
1473 // {0,+,1} --> Insert a canonical induction variable into the loop!
1474 if (S->isAffine() && S->getOperand(1)->isOne()) {
1475 assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) &&
1476 "IVs with types different from the canonical IV should "
1477 "already have been handled!");
1481 // {0,+,F} --> {0,+,1} * F
1483 // If this is a simple linear addrec, emit it now as a special case.
1484 if (S->isAffine()) // {0,+,F} --> i*F
1486 expand(SE.getTruncateOrNoop(
1487 SE.getMulExpr(SE.getUnknown(CanonicalIV),
1488 SE.getNoopOrAnyExtend(S->getOperand(1),
1489 CanonicalIV->getType())),
1492 // If this is a chain of recurrences, turn it into a closed form, using the
1493 // folders, then expandCodeFor the closed form. This allows the folders to
1494 // simplify the expression without having to build a bunch of special code
1495 // into this folder.
1496 const SCEV *IH = SE.getUnknown(CanonicalIV); // Get I as a "symbolic" SCEV.
1498 // Promote S up to the canonical IV type, if the cast is foldable.
1499 const SCEV *NewS = S;
1500 const SCEV *Ext = SE.getNoopOrAnyExtend(S, CanonicalIV->getType());
1501 if (isa<SCEVAddRecExpr>(Ext))
1504 const SCEV *V = cast<SCEVAddRecExpr>(NewS)->evaluateAtIteration(IH, SE);
1505 //cerr << "Evaluated: " << *this << "\n to: " << *V << "\n";
1507 // Truncate the result down to the original type, if needed.
1508 const SCEV *T = SE.getTruncateOrNoop(V, Ty);
1512 Value *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) {
1513 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1514 Value *V = expandCodeFor(S->getOperand(),
1515 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1516 Value *I = Builder.CreateTrunc(V, Ty);
1517 rememberInstruction(I);
1521 Value *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) {
1522 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1523 Value *V = expandCodeFor(S->getOperand(),
1524 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1525 Value *I = Builder.CreateZExt(V, Ty);
1526 rememberInstruction(I);
1530 Value *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) {
1531 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1532 Value *V = expandCodeFor(S->getOperand(),
1533 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1534 Value *I = Builder.CreateSExt(V, Ty);
1535 rememberInstruction(I);
1539 Value *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) {
1540 Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
1541 Type *Ty = LHS->getType();
1542 for (int i = S->getNumOperands()-2; i >= 0; --i) {
1543 // In the case of mixed integer and pointer types, do the
1544 // rest of the comparisons as integer.
1545 if (S->getOperand(i)->getType() != Ty) {
1546 Ty = SE.getEffectiveSCEVType(Ty);
1547 LHS = InsertNoopCastOfTo(LHS, Ty);
1549 Value *RHS = expandCodeFor(S->getOperand(i), Ty);
1550 Value *ICmp = Builder.CreateICmpSGT(LHS, RHS);
1551 rememberInstruction(ICmp);
1552 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "smax");
1553 rememberInstruction(Sel);
1556 // In the case of mixed integer and pointer types, cast the
1557 // final result back to the pointer type.
1558 if (LHS->getType() != S->getType())
1559 LHS = InsertNoopCastOfTo(LHS, S->getType());
1563 Value *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) {
1564 Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
1565 Type *Ty = LHS->getType();
1566 for (int i = S->getNumOperands()-2; i >= 0; --i) {
1567 // In the case of mixed integer and pointer types, do the
1568 // rest of the comparisons as integer.
1569 if (S->getOperand(i)->getType() != Ty) {
1570 Ty = SE.getEffectiveSCEVType(Ty);
1571 LHS = InsertNoopCastOfTo(LHS, Ty);
1573 Value *RHS = expandCodeFor(S->getOperand(i), Ty);
1574 Value *ICmp = Builder.CreateICmpUGT(LHS, RHS);
1575 rememberInstruction(ICmp);
1576 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "umax");
1577 rememberInstruction(Sel);
1580 // In the case of mixed integer and pointer types, cast the
1581 // final result back to the pointer type.
1582 if (LHS->getType() != S->getType())
1583 LHS = InsertNoopCastOfTo(LHS, S->getType());
1587 Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty,
1589 Builder.SetInsertPoint(IP->getParent(), IP);
1590 return expandCodeFor(SH, Ty);
1593 Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty) {
1594 // Expand the code for this SCEV.
1595 Value *V = expand(SH);
1597 assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) &&
1598 "non-trivial casts should be done with the SCEVs directly!");
1599 V = InsertNoopCastOfTo(V, Ty);
1604 Value *SCEVExpander::expand(const SCEV *S) {
1605 // Compute an insertion point for this SCEV object. Hoist the instructions
1606 // as far out in the loop nest as possible.
1607 Instruction *InsertPt = Builder.GetInsertPoint();
1608 for (Loop *L = SE.LI.getLoopFor(Builder.GetInsertBlock());;
1609 L = L->getParentLoop())
1610 if (SE.isLoopInvariant(S, L)) {
1612 if (BasicBlock *Preheader = L->getLoopPreheader())
1613 InsertPt = Preheader->getTerminator();
1615 // LSR sets the insertion point for AddRec start/step values to the
1616 // block start to simplify value reuse, even though it's an invalid
1617 // position. SCEVExpander must correct for this in all cases.
1618 InsertPt = L->getHeader()->getFirstInsertionPt();
1621 // If the SCEV is computable at this level, insert it into the header
1622 // after the PHIs (and after any other instructions that we've inserted
1623 // there) so that it is guaranteed to dominate any user inside the loop.
1624 if (L && SE.hasComputableLoopEvolution(S, L) && !PostIncLoops.count(L))
1625 InsertPt = L->getHeader()->getFirstInsertionPt();
1626 while (InsertPt != Builder.GetInsertPoint()
1627 && (isInsertedInstruction(InsertPt)
1628 || isa<DbgInfoIntrinsic>(InsertPt))) {
1629 InsertPt = std::next(BasicBlock::iterator(InsertPt));
1634 // Check to see if we already expanded this here.
1635 std::map<std::pair<const SCEV *, Instruction *>, TrackingVH<Value> >::iterator
1636 I = InsertedExpressions.find(std::make_pair(S, InsertPt));
1637 if (I != InsertedExpressions.end())
1640 BuilderType::InsertPointGuard Guard(Builder);
1641 Builder.SetInsertPoint(InsertPt->getParent(), InsertPt);
1643 // Expand the expression into instructions.
1644 Value *V = visit(S);
1646 // Remember the expanded value for this SCEV at this location.
1648 // This is independent of PostIncLoops. The mapped value simply materializes
1649 // the expression at this insertion point. If the mapped value happened to be
1650 // a postinc expansion, it could be reused by a non-postinc user, but only if
1651 // its insertion point was already at the head of the loop.
1652 InsertedExpressions[std::make_pair(S, InsertPt)] = V;
1656 void SCEVExpander::rememberInstruction(Value *I) {
1657 if (!PostIncLoops.empty())
1658 InsertedPostIncValues.insert(I);
1660 InsertedValues.insert(I);
1663 /// getOrInsertCanonicalInductionVariable - This method returns the
1664 /// canonical induction variable of the specified type for the specified
1665 /// loop (inserting one if there is none). A canonical induction variable
1666 /// starts at zero and steps by one on each iteration.
1668 SCEVExpander::getOrInsertCanonicalInductionVariable(const Loop *L,
1670 assert(Ty->isIntegerTy() && "Can only insert integer induction variables!");
1672 // Build a SCEV for {0,+,1}<L>.
1673 // Conservatively use FlagAnyWrap for now.
1674 const SCEV *H = SE.getAddRecExpr(SE.getConstant(Ty, 0),
1675 SE.getConstant(Ty, 1), L, SCEV::FlagAnyWrap);
1677 // Emit code for it.
1678 BuilderType::InsertPointGuard Guard(Builder);
1679 PHINode *V = cast<PHINode>(expandCodeFor(H, nullptr,
1680 L->getHeader()->begin()));
1685 /// replaceCongruentIVs - Check for congruent phis in this loop header and
1686 /// replace them with their most canonical representative. Return the number of
1687 /// phis eliminated.
1689 /// This does not depend on any SCEVExpander state but should be used in
1690 /// the same context that SCEVExpander is used.
1691 unsigned SCEVExpander::replaceCongruentIVs(Loop *L, const DominatorTree *DT,
1692 SmallVectorImpl<WeakVH> &DeadInsts,
1693 const TargetTransformInfo *TTI) {
1694 // Find integer phis in order of increasing width.
1695 SmallVector<PHINode*, 8> Phis;
1696 for (BasicBlock::iterator I = L->getHeader()->begin();
1697 PHINode *Phi = dyn_cast<PHINode>(I); ++I) {
1698 Phis.push_back(Phi);
1701 std::sort(Phis.begin(), Phis.end(), [](Value *LHS, Value *RHS) {
1702 // Put pointers at the back and make sure pointer < pointer = false.
1703 if (!LHS->getType()->isIntegerTy() || !RHS->getType()->isIntegerTy())
1704 return RHS->getType()->isIntegerTy() && !LHS->getType()->isIntegerTy();
1705 return RHS->getType()->getPrimitiveSizeInBits() <
1706 LHS->getType()->getPrimitiveSizeInBits();
1709 unsigned NumElim = 0;
1710 DenseMap<const SCEV *, PHINode *> ExprToIVMap;
1711 // Process phis from wide to narrow. Map wide phis to their truncation
1712 // so narrow phis can reuse them.
1713 for (SmallVectorImpl<PHINode*>::const_iterator PIter = Phis.begin(),
1714 PEnd = Phis.end(); PIter != PEnd; ++PIter) {
1715 PHINode *Phi = *PIter;
1717 // Fold constant phis. They may be congruent to other constant phis and
1718 // would confuse the logic below that expects proper IVs.
1719 if (Value *V = SimplifyInstruction(Phi, DL, &SE.TLI, &SE.DT, &SE.AC)) {
1720 Phi->replaceAllUsesWith(V);
1721 DeadInsts.emplace_back(Phi);
1723 DEBUG_WITH_TYPE(DebugType, dbgs()
1724 << "INDVARS: Eliminated constant iv: " << *Phi << '\n');
1728 if (!SE.isSCEVable(Phi->getType()))
1731 PHINode *&OrigPhiRef = ExprToIVMap[SE.getSCEV(Phi)];
1734 if (Phi->getType()->isIntegerTy() && TTI
1735 && TTI->isTruncateFree(Phi->getType(), Phis.back()->getType())) {
1736 // This phi can be freely truncated to the narrowest phi type. Map the
1737 // truncated expression to it so it will be reused for narrow types.
1738 const SCEV *TruncExpr =
1739 SE.getTruncateExpr(SE.getSCEV(Phi), Phis.back()->getType());
1740 ExprToIVMap[TruncExpr] = Phi;
1745 // Replacing a pointer phi with an integer phi or vice-versa doesn't make
1747 if (OrigPhiRef->getType()->isPointerTy() != Phi->getType()->isPointerTy())
1750 if (BasicBlock *LatchBlock = L->getLoopLatch()) {
1751 Instruction *OrigInc =
1752 cast<Instruction>(OrigPhiRef->getIncomingValueForBlock(LatchBlock));
1753 Instruction *IsomorphicInc =
1754 cast<Instruction>(Phi->getIncomingValueForBlock(LatchBlock));
1756 // If this phi has the same width but is more canonical, replace the
1757 // original with it. As part of the "more canonical" determination,
1758 // respect a prior decision to use an IV chain.
1759 if (OrigPhiRef->getType() == Phi->getType()
1760 && !(ChainedPhis.count(Phi)
1761 || isExpandedAddRecExprPHI(OrigPhiRef, OrigInc, L))
1762 && (ChainedPhis.count(Phi)
1763 || isExpandedAddRecExprPHI(Phi, IsomorphicInc, L))) {
1764 std::swap(OrigPhiRef, Phi);
1765 std::swap(OrigInc, IsomorphicInc);
1767 // Replacing the congruent phi is sufficient because acyclic redundancy
1768 // elimination, CSE/GVN, should handle the rest. However, once SCEV proves
1769 // that a phi is congruent, it's often the head of an IV user cycle that
1770 // is isomorphic with the original phi. It's worth eagerly cleaning up the
1771 // common case of a single IV increment so that DeleteDeadPHIs can remove
1772 // cycles that had postinc uses.
1773 const SCEV *TruncExpr = SE.getTruncateOrNoop(SE.getSCEV(OrigInc),
1774 IsomorphicInc->getType());
1775 if (OrigInc != IsomorphicInc
1776 && TruncExpr == SE.getSCEV(IsomorphicInc)
1777 && ((isa<PHINode>(OrigInc) && isa<PHINode>(IsomorphicInc))
1778 || hoistIVInc(OrigInc, IsomorphicInc))) {
1779 DEBUG_WITH_TYPE(DebugType, dbgs()
1780 << "INDVARS: Eliminated congruent iv.inc: "
1781 << *IsomorphicInc << '\n');
1782 Value *NewInc = OrigInc;
1783 if (OrigInc->getType() != IsomorphicInc->getType()) {
1784 Instruction *IP = nullptr;
1785 if (PHINode *PN = dyn_cast<PHINode>(OrigInc))
1786 IP = PN->getParent()->getFirstInsertionPt();
1788 IP = OrigInc->getNextNode();
1790 IRBuilder<> Builder(IP);
1791 Builder.SetCurrentDebugLocation(IsomorphicInc->getDebugLoc());
1793 CreateTruncOrBitCast(OrigInc, IsomorphicInc->getType(), IVName);
1795 IsomorphicInc->replaceAllUsesWith(NewInc);
1796 DeadInsts.emplace_back(IsomorphicInc);
1799 DEBUG_WITH_TYPE(DebugType, dbgs()
1800 << "INDVARS: Eliminated congruent iv: " << *Phi << '\n');
1802 Value *NewIV = OrigPhiRef;
1803 if (OrigPhiRef->getType() != Phi->getType()) {
1804 IRBuilder<> Builder(L->getHeader()->getFirstInsertionPt());
1805 Builder.SetCurrentDebugLocation(Phi->getDebugLoc());
1806 NewIV = Builder.CreateTruncOrBitCast(OrigPhiRef, Phi->getType(), IVName);
1808 Phi->replaceAllUsesWith(NewIV);
1809 DeadInsts.emplace_back(Phi);
1814 Value *SCEVExpander::findExistingExpansion(const SCEV *S,
1815 const Instruction *At, Loop *L) {
1816 using namespace llvm::PatternMatch;
1818 SmallVector<BasicBlock *, 4> ExitingBlocks;
1819 L->getExitingBlocks(ExitingBlocks);
1821 // Look for suitable value in simple conditions at the loop exits.
1822 for (BasicBlock *BB : ExitingBlocks) {
1823 ICmpInst::Predicate Pred;
1824 Instruction *LHS, *RHS;
1825 BasicBlock *TrueBB, *FalseBB;
1827 if (!match(BB->getTerminator(),
1828 m_Br(m_ICmp(Pred, m_Instruction(LHS), m_Instruction(RHS)),
1832 if (SE.getSCEV(LHS) == S && SE.DT.dominates(LHS, At))
1835 if (SE.getSCEV(RHS) == S && SE.DT.dominates(RHS, At))
1839 // There is potential to make this significantly smarter, but this simple
1840 // heuristic already gets some interesting cases.
1842 // Can not find suitable value.
1846 bool SCEVExpander::isHighCostExpansionHelper(
1847 const SCEV *S, Loop *L, const Instruction *At,
1848 SmallPtrSetImpl<const SCEV *> &Processed) {
1850 // If we can find an existing value for this scev avaliable at the point "At"
1851 // then consider the expression cheap.
1852 if (At && findExistingExpansion(S, At, L) != nullptr)
1855 // Zero/One operand expressions
1856 switch (S->getSCEVType()) {
1861 return isHighCostExpansionHelper(cast<SCEVTruncateExpr>(S)->getOperand(),
1864 return isHighCostExpansionHelper(cast<SCEVZeroExtendExpr>(S)->getOperand(),
1867 return isHighCostExpansionHelper(cast<SCEVSignExtendExpr>(S)->getOperand(),
1871 if (!Processed.insert(S).second)
1874 if (auto *UDivExpr = dyn_cast<SCEVUDivExpr>(S)) {
1875 // If the divisor is a power of two and the SCEV type fits in a native
1876 // integer, consider the division cheap irrespective of whether it occurs in
1877 // the user code since it can be lowered into a right shift.
1878 if (auto *SC = dyn_cast<SCEVConstant>(UDivExpr->getRHS()))
1879 if (SC->getValue()->getValue().isPowerOf2()) {
1880 const DataLayout &DL =
1881 L->getHeader()->getParent()->getParent()->getDataLayout();
1882 unsigned Width = cast<IntegerType>(UDivExpr->getType())->getBitWidth();
1883 return DL.isIllegalInteger(Width);
1886 // UDivExpr is very likely a UDiv that ScalarEvolution's HowFarToZero or
1887 // HowManyLessThans produced to compute a precise expression, rather than a
1888 // UDiv from the user's code. If we can't find a UDiv in the code with some
1889 // simple searching, assume the former consider UDivExpr expensive to
1891 BasicBlock *ExitingBB = L->getExitingBlock();
1895 // At the beginning of this function we already tried to find existing value
1896 // for plain 'S'. Now try to lookup 'S + 1' since it is common pattern
1897 // involving division. This is just a simple search heuristic.
1899 At = &ExitingBB->back();
1900 if (!findExistingExpansion(
1901 SE.getAddExpr(S, SE.getConstant(S->getType(), 1)), At, L))
1905 // HowManyLessThans uses a Max expression whenever the loop is not guarded by
1906 // the exit condition.
1907 if (isa<SCEVSMaxExpr>(S) || isa<SCEVUMaxExpr>(S))
1910 // Recurse past nary expressions, which commonly occur in the
1911 // BackedgeTakenCount. They may already exist in program code, and if not,
1912 // they are not too expensive rematerialize.
1913 if (const SCEVNAryExpr *NAry = dyn_cast<SCEVNAryExpr>(S)) {
1914 for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end();
1916 if (isHighCostExpansionHelper(*I, L, At, Processed))
1921 // If we haven't recognized an expensive SCEV pattern, assume it's an
1922 // expression produced by program code.
1927 // Search for a SCEV subexpression that is not safe to expand. Any expression
1928 // that may expand to a !isSafeToSpeculativelyExecute value is unsafe, namely
1929 // UDiv expressions. We don't know if the UDiv is derived from an IR divide
1930 // instruction, but the important thing is that we prove the denominator is
1931 // nonzero before expansion.
1933 // IVUsers already checks that IV-derived expressions are safe. So this check is
1934 // only needed when the expression includes some subexpression that is not IV
1937 // Currently, we only allow division by a nonzero constant here. If this is
1938 // inadequate, we could easily allow division by SCEVUnknown by using
1939 // ValueTracking to check isKnownNonZero().
1941 // We cannot generally expand recurrences unless the step dominates the loop
1942 // header. The expander handles the special case of affine recurrences by
1943 // scaling the recurrence outside the loop, but this technique isn't generally
1944 // applicable. Expanding a nested recurrence outside a loop requires computing
1945 // binomial coefficients. This could be done, but the recurrence has to be in a
1946 // perfectly reduced form, which can't be guaranteed.
1947 struct SCEVFindUnsafe {
1948 ScalarEvolution &SE;
1951 SCEVFindUnsafe(ScalarEvolution &se): SE(se), IsUnsafe(false) {}
1953 bool follow(const SCEV *S) {
1954 if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
1955 const SCEVConstant *SC = dyn_cast<SCEVConstant>(D->getRHS());
1956 if (!SC || SC->getValue()->isZero()) {
1961 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
1962 const SCEV *Step = AR->getStepRecurrence(SE);
1963 if (!AR->isAffine() && !SE.dominates(Step, AR->getLoop()->getHeader())) {
1970 bool isDone() const { return IsUnsafe; }
1975 bool isSafeToExpand(const SCEV *S, ScalarEvolution &SE) {
1976 SCEVFindUnsafe Search(SE);
1977 visitAll(S, Search);
1978 return !Search.IsUnsafe;