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 static BasicBlock::iterator findInsertPointAfter(Instruction *I,
90 BasicBlock *MustDominate) {
91 BasicBlock::iterator IP = ++I->getIterator();
92 if (auto *II = dyn_cast<InvokeInst>(I))
93 IP = II->getNormalDest()->begin();
95 while (isa<PHINode>(IP))
98 while (IP->isEHPad()) {
99 if (isa<FuncletPadInst>(IP) || isa<LandingPadInst>(IP)) {
101 } else if (isa<CatchSwitchInst>(IP)) {
102 IP = MustDominate->getFirstInsertionPt();
104 llvm_unreachable("unexpected eh pad!");
111 /// InsertNoopCastOfTo - Insert a cast of V to the specified type,
112 /// which must be possible with a noop cast, doing what we can to share
114 Value *SCEVExpander::InsertNoopCastOfTo(Value *V, Type *Ty) {
115 Instruction::CastOps Op = CastInst::getCastOpcode(V, false, Ty, false);
116 assert((Op == Instruction::BitCast ||
117 Op == Instruction::PtrToInt ||
118 Op == Instruction::IntToPtr) &&
119 "InsertNoopCastOfTo cannot perform non-noop casts!");
120 assert(SE.getTypeSizeInBits(V->getType()) == SE.getTypeSizeInBits(Ty) &&
121 "InsertNoopCastOfTo cannot change sizes!");
123 // Short-circuit unnecessary bitcasts.
124 if (Op == Instruction::BitCast) {
125 if (V->getType() == Ty)
127 if (CastInst *CI = dyn_cast<CastInst>(V)) {
128 if (CI->getOperand(0)->getType() == Ty)
129 return CI->getOperand(0);
132 // Short-circuit unnecessary inttoptr<->ptrtoint casts.
133 if ((Op == Instruction::PtrToInt || Op == Instruction::IntToPtr) &&
134 SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(V->getType())) {
135 if (CastInst *CI = dyn_cast<CastInst>(V))
136 if ((CI->getOpcode() == Instruction::PtrToInt ||
137 CI->getOpcode() == Instruction::IntToPtr) &&
138 SE.getTypeSizeInBits(CI->getType()) ==
139 SE.getTypeSizeInBits(CI->getOperand(0)->getType()))
140 return CI->getOperand(0);
141 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
142 if ((CE->getOpcode() == Instruction::PtrToInt ||
143 CE->getOpcode() == Instruction::IntToPtr) &&
144 SE.getTypeSizeInBits(CE->getType()) ==
145 SE.getTypeSizeInBits(CE->getOperand(0)->getType()))
146 return CE->getOperand(0);
149 // Fold a cast of a constant.
150 if (Constant *C = dyn_cast<Constant>(V))
151 return ConstantExpr::getCast(Op, C, Ty);
153 // Cast the argument at the beginning of the entry block, after
154 // any bitcasts of other arguments.
155 if (Argument *A = dyn_cast<Argument>(V)) {
156 BasicBlock::iterator IP = A->getParent()->getEntryBlock().begin();
157 while ((isa<BitCastInst>(IP) &&
158 isa<Argument>(cast<BitCastInst>(IP)->getOperand(0)) &&
159 cast<BitCastInst>(IP)->getOperand(0) != A) ||
160 isa<DbgInfoIntrinsic>(IP))
162 return ReuseOrCreateCast(A, Ty, Op, IP);
165 // Cast the instruction immediately after the instruction.
166 Instruction *I = cast<Instruction>(V);
167 BasicBlock::iterator IP = findInsertPointAfter(I, Builder.GetInsertBlock());
168 return ReuseOrCreateCast(I, Ty, Op, IP);
171 /// InsertBinop - Insert the specified binary operator, doing a small amount
172 /// of work to avoid inserting an obviously redundant operation.
173 Value *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode,
174 Value *LHS, Value *RHS) {
175 // Fold a binop with constant operands.
176 if (Constant *CLHS = dyn_cast<Constant>(LHS))
177 if (Constant *CRHS = dyn_cast<Constant>(RHS))
178 return ConstantExpr::get(Opcode, CLHS, CRHS);
180 // Do a quick scan to see if we have this binop nearby. If so, reuse it.
181 unsigned ScanLimit = 6;
182 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
183 // Scanning starts from the last instruction before the insertion point.
184 BasicBlock::iterator IP = Builder.GetInsertPoint();
185 if (IP != BlockBegin) {
187 for (; ScanLimit; --IP, --ScanLimit) {
188 // Don't count dbg.value against the ScanLimit, to avoid perturbing the
190 if (isa<DbgInfoIntrinsic>(IP))
192 if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(0) == LHS &&
193 IP->getOperand(1) == RHS)
195 if (IP == BlockBegin) break;
199 // Save the original insertion point so we can restore it when we're done.
200 DebugLoc Loc = Builder.GetInsertPoint()->getDebugLoc();
201 BuilderType::InsertPointGuard Guard(Builder);
203 // Move the insertion point out of as many loops as we can.
204 while (const Loop *L = SE.LI.getLoopFor(Builder.GetInsertBlock())) {
205 if (!L->isLoopInvariant(LHS) || !L->isLoopInvariant(RHS)) break;
206 BasicBlock *Preheader = L->getLoopPreheader();
207 if (!Preheader) break;
209 // Ok, move up a level.
210 Builder.SetInsertPoint(Preheader->getTerminator());
213 // If we haven't found this binop, insert it.
214 Instruction *BO = cast<Instruction>(Builder.CreateBinOp(Opcode, LHS, RHS));
215 BO->setDebugLoc(Loc);
216 rememberInstruction(BO);
221 /// FactorOutConstant - Test if S is divisible by Factor, using signed
222 /// division. If so, update S with Factor divided out and return true.
223 /// S need not be evenly divisible if a reasonable remainder can be
225 /// TODO: When ScalarEvolution gets a SCEVSDivExpr, this can be made
226 /// unnecessary; in its place, just signed-divide Ops[i] by the scale and
227 /// check to see if the divide was folded.
228 static bool FactorOutConstant(const SCEV *&S, const SCEV *&Remainder,
229 const SCEV *Factor, ScalarEvolution &SE,
230 const DataLayout &DL) {
231 // Everything is divisible by one.
237 S = SE.getConstant(S->getType(), 1);
241 // For a Constant, check for a multiple of the given factor.
242 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
246 // Check for divisibility.
247 if (const SCEVConstant *FC = dyn_cast<SCEVConstant>(Factor)) {
249 ConstantInt::get(SE.getContext(), C->getAPInt().sdiv(FC->getAPInt()));
250 // If the quotient is zero and the remainder is non-zero, reject
251 // the value at this scale. It will be considered for subsequent
254 const SCEV *Div = SE.getConstant(CI);
256 Remainder = SE.getAddExpr(
257 Remainder, SE.getConstant(C->getAPInt().srem(FC->getAPInt())));
263 // In a Mul, check if there is a constant operand which is a multiple
264 // of the given factor.
265 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
266 // Size is known, check if there is a constant operand which is a multiple
267 // of the given factor. If so, we can factor it.
268 const SCEVConstant *FC = cast<SCEVConstant>(Factor);
269 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0)))
270 if (!C->getAPInt().srem(FC->getAPInt())) {
271 SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end());
272 NewMulOps[0] = SE.getConstant(C->getAPInt().sdiv(FC->getAPInt()));
273 S = SE.getMulExpr(NewMulOps);
278 // In an AddRec, check if both start and step are divisible.
279 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
280 const SCEV *Step = A->getStepRecurrence(SE);
281 const SCEV *StepRem = SE.getConstant(Step->getType(), 0);
282 if (!FactorOutConstant(Step, StepRem, Factor, SE, DL))
284 if (!StepRem->isZero())
286 const SCEV *Start = A->getStart();
287 if (!FactorOutConstant(Start, Remainder, Factor, SE, DL))
289 S = SE.getAddRecExpr(Start, Step, A->getLoop(),
290 A->getNoWrapFlags(SCEV::FlagNW));
297 /// SimplifyAddOperands - Sort and simplify a list of add operands. NumAddRecs
298 /// is the number of SCEVAddRecExprs present, which are kept at the end of
301 static void SimplifyAddOperands(SmallVectorImpl<const SCEV *> &Ops,
303 ScalarEvolution &SE) {
304 unsigned NumAddRecs = 0;
305 for (unsigned i = Ops.size(); i > 0 && isa<SCEVAddRecExpr>(Ops[i-1]); --i)
307 // Group Ops into non-addrecs and addrecs.
308 SmallVector<const SCEV *, 8> NoAddRecs(Ops.begin(), Ops.end() - NumAddRecs);
309 SmallVector<const SCEV *, 8> AddRecs(Ops.end() - NumAddRecs, Ops.end());
310 // Let ScalarEvolution sort and simplify the non-addrecs list.
311 const SCEV *Sum = NoAddRecs.empty() ?
312 SE.getConstant(Ty, 0) :
313 SE.getAddExpr(NoAddRecs);
314 // If it returned an add, use the operands. Otherwise it simplified
315 // the sum into a single value, so just use that.
317 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Sum))
318 Ops.append(Add->op_begin(), Add->op_end());
319 else if (!Sum->isZero())
321 // Then append the addrecs.
322 Ops.append(AddRecs.begin(), AddRecs.end());
325 /// SplitAddRecs - Flatten a list of add operands, moving addrec start values
326 /// out to the top level. For example, convert {a + b,+,c} to a, b, {0,+,d}.
327 /// This helps expose more opportunities for folding parts of the expressions
328 /// into GEP indices.
330 static void SplitAddRecs(SmallVectorImpl<const SCEV *> &Ops,
332 ScalarEvolution &SE) {
334 SmallVector<const SCEV *, 8> AddRecs;
335 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
336 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Ops[i])) {
337 const SCEV *Start = A->getStart();
338 if (Start->isZero()) break;
339 const SCEV *Zero = SE.getConstant(Ty, 0);
340 AddRecs.push_back(SE.getAddRecExpr(Zero,
341 A->getStepRecurrence(SE),
343 A->getNoWrapFlags(SCEV::FlagNW)));
344 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Start)) {
346 Ops.append(Add->op_begin(), Add->op_end());
347 e += Add->getNumOperands();
352 if (!AddRecs.empty()) {
353 // Add the addrecs onto the end of the list.
354 Ops.append(AddRecs.begin(), AddRecs.end());
355 // Resort the operand list, moving any constants to the front.
356 SimplifyAddOperands(Ops, Ty, SE);
360 /// expandAddToGEP - Expand an addition expression with a pointer type into
361 /// a GEP instead of using ptrtoint+arithmetic+inttoptr. This helps
362 /// BasicAliasAnalysis and other passes analyze the result. See the rules
363 /// for getelementptr vs. inttoptr in
364 /// http://llvm.org/docs/LangRef.html#pointeraliasing
367 /// Design note: The correctness of using getelementptr here depends on
368 /// ScalarEvolution not recognizing inttoptr and ptrtoint operators, as
369 /// they may introduce pointer arithmetic which may not be safely converted
370 /// into getelementptr.
372 /// Design note: It might seem desirable for this function to be more
373 /// loop-aware. If some of the indices are loop-invariant while others
374 /// aren't, it might seem desirable to emit multiple GEPs, keeping the
375 /// loop-invariant portions of the overall computation outside the loop.
376 /// However, there are a few reasons this is not done here. Hoisting simple
377 /// arithmetic is a low-level optimization that often isn't very
378 /// important until late in the optimization process. In fact, passes
379 /// like InstructionCombining will combine GEPs, even if it means
380 /// pushing loop-invariant computation down into loops, so even if the
381 /// GEPs were split here, the work would quickly be undone. The
382 /// LoopStrengthReduction pass, which is usually run quite late (and
383 /// after the last InstructionCombining pass), takes care of hoisting
384 /// loop-invariant portions of expressions, after considering what
385 /// can be folded using target addressing modes.
387 Value *SCEVExpander::expandAddToGEP(const SCEV *const *op_begin,
388 const SCEV *const *op_end,
392 Type *OriginalElTy = PTy->getElementType();
393 Type *ElTy = OriginalElTy;
394 SmallVector<Value *, 4> GepIndices;
395 SmallVector<const SCEV *, 8> Ops(op_begin, op_end);
396 bool AnyNonZeroIndices = false;
398 // Split AddRecs up into parts as either of the parts may be usable
399 // without the other.
400 SplitAddRecs(Ops, Ty, SE);
402 Type *IntPtrTy = DL.getIntPtrType(PTy);
404 // Descend down the pointer's type and attempt to convert the other
405 // operands into GEP indices, at each level. The first index in a GEP
406 // indexes into the array implied by the pointer operand; the rest of
407 // the indices index into the element or field type selected by the
410 // If the scale size is not 0, attempt to factor out a scale for
412 SmallVector<const SCEV *, 8> ScaledOps;
413 if (ElTy->isSized()) {
414 const SCEV *ElSize = SE.getSizeOfExpr(IntPtrTy, ElTy);
415 if (!ElSize->isZero()) {
416 SmallVector<const SCEV *, 8> NewOps;
417 for (const SCEV *Op : Ops) {
418 const SCEV *Remainder = SE.getConstant(Ty, 0);
419 if (FactorOutConstant(Op, Remainder, ElSize, SE, DL)) {
420 // Op now has ElSize factored out.
421 ScaledOps.push_back(Op);
422 if (!Remainder->isZero())
423 NewOps.push_back(Remainder);
424 AnyNonZeroIndices = true;
426 // The operand was not divisible, so add it to the list of operands
427 // we'll scan next iteration.
428 NewOps.push_back(Op);
431 // If we made any changes, update Ops.
432 if (!ScaledOps.empty()) {
434 SimplifyAddOperands(Ops, Ty, SE);
439 // Record the scaled array index for this level of the type. If
440 // we didn't find any operands that could be factored, tentatively
441 // assume that element zero was selected (since the zero offset
442 // would obviously be folded away).
443 Value *Scaled = ScaledOps.empty() ?
444 Constant::getNullValue(Ty) :
445 expandCodeFor(SE.getAddExpr(ScaledOps), Ty);
446 GepIndices.push_back(Scaled);
448 // Collect struct field index operands.
449 while (StructType *STy = dyn_cast<StructType>(ElTy)) {
450 bool FoundFieldNo = false;
451 // An empty struct has no fields.
452 if (STy->getNumElements() == 0) break;
453 // Field offsets are known. See if a constant offset falls within any of
454 // the struct fields.
457 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[0]))
458 if (SE.getTypeSizeInBits(C->getType()) <= 64) {
459 const StructLayout &SL = *DL.getStructLayout(STy);
460 uint64_t FullOffset = C->getValue()->getZExtValue();
461 if (FullOffset < SL.getSizeInBytes()) {
462 unsigned ElIdx = SL.getElementContainingOffset(FullOffset);
463 GepIndices.push_back(
464 ConstantInt::get(Type::getInt32Ty(Ty->getContext()), ElIdx));
465 ElTy = STy->getTypeAtIndex(ElIdx);
467 SE.getConstant(Ty, FullOffset - SL.getElementOffset(ElIdx));
468 AnyNonZeroIndices = true;
472 // If no struct field offsets were found, tentatively assume that
473 // field zero was selected (since the zero offset would obviously
476 ElTy = STy->getTypeAtIndex(0u);
477 GepIndices.push_back(
478 Constant::getNullValue(Type::getInt32Ty(Ty->getContext())));
482 if (ArrayType *ATy = dyn_cast<ArrayType>(ElTy))
483 ElTy = ATy->getElementType();
488 // If none of the operands were convertible to proper GEP indices, cast
489 // the base to i8* and do an ugly getelementptr with that. It's still
490 // better than ptrtoint+arithmetic+inttoptr at least.
491 if (!AnyNonZeroIndices) {
492 // Cast the base to i8*.
493 V = InsertNoopCastOfTo(V,
494 Type::getInt8PtrTy(Ty->getContext(), PTy->getAddressSpace()));
496 assert(!isa<Instruction>(V) ||
497 SE.DT.dominates(cast<Instruction>(V), &*Builder.GetInsertPoint()));
499 // Expand the operands for a plain byte offset.
500 Value *Idx = expandCodeFor(SE.getAddExpr(Ops), Ty);
502 // Fold a GEP with constant operands.
503 if (Constant *CLHS = dyn_cast<Constant>(V))
504 if (Constant *CRHS = dyn_cast<Constant>(Idx))
505 return ConstantExpr::getGetElementPtr(Type::getInt8Ty(Ty->getContext()),
508 // Do a quick scan to see if we have this GEP nearby. If so, reuse it.
509 unsigned ScanLimit = 6;
510 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
511 // Scanning starts from the last instruction before the insertion point.
512 BasicBlock::iterator IP = Builder.GetInsertPoint();
513 if (IP != BlockBegin) {
515 for (; ScanLimit; --IP, --ScanLimit) {
516 // Don't count dbg.value against the ScanLimit, to avoid perturbing the
518 if (isa<DbgInfoIntrinsic>(IP))
520 if (IP->getOpcode() == Instruction::GetElementPtr &&
521 IP->getOperand(0) == V && IP->getOperand(1) == Idx)
523 if (IP == BlockBegin) break;
527 // Save the original insertion point so we can restore it when we're done.
528 BuilderType::InsertPointGuard Guard(Builder);
530 // Move the insertion point out of as many loops as we can.
531 while (const Loop *L = SE.LI.getLoopFor(Builder.GetInsertBlock())) {
532 if (!L->isLoopInvariant(V) || !L->isLoopInvariant(Idx)) break;
533 BasicBlock *Preheader = L->getLoopPreheader();
534 if (!Preheader) break;
536 // Ok, move up a level.
537 Builder.SetInsertPoint(Preheader->getTerminator());
541 Value *GEP = Builder.CreateGEP(Builder.getInt8Ty(), V, Idx, "uglygep");
542 rememberInstruction(GEP);
547 // Save the original insertion point so we can restore it when we're done.
548 BuilderType::InsertPoint SaveInsertPt = Builder.saveIP();
550 // Move the insertion point out of as many loops as we can.
551 while (const Loop *L = SE.LI.getLoopFor(Builder.GetInsertBlock())) {
552 if (!L->isLoopInvariant(V)) break;
554 bool AnyIndexNotLoopInvariant =
555 std::any_of(GepIndices.begin(), GepIndices.end(),
556 [L](Value *Op) { return !L->isLoopInvariant(Op); });
558 if (AnyIndexNotLoopInvariant)
561 BasicBlock *Preheader = L->getLoopPreheader();
562 if (!Preheader) break;
564 // Ok, move up a level.
565 Builder.SetInsertPoint(Preheader->getTerminator());
568 // Insert a pretty getelementptr. Note that this GEP is not marked inbounds,
569 // because ScalarEvolution may have changed the address arithmetic to
570 // compute a value which is beyond the end of the allocated object.
572 if (V->getType() != PTy)
573 Casted = InsertNoopCastOfTo(Casted, PTy);
574 Value *GEP = Builder.CreateGEP(OriginalElTy, Casted, GepIndices, "scevgep");
575 Ops.push_back(SE.getUnknown(GEP));
576 rememberInstruction(GEP);
578 // Restore the original insert point.
579 Builder.restoreIP(SaveInsertPt);
581 return expand(SE.getAddExpr(Ops));
584 /// PickMostRelevantLoop - Given two loops pick the one that's most relevant for
585 /// SCEV expansion. If they are nested, this is the most nested. If they are
586 /// neighboring, pick the later.
587 static const Loop *PickMostRelevantLoop(const Loop *A, const Loop *B,
591 if (A->contains(B)) return B;
592 if (B->contains(A)) return A;
593 if (DT.dominates(A->getHeader(), B->getHeader())) return B;
594 if (DT.dominates(B->getHeader(), A->getHeader())) return A;
595 return A; // Arbitrarily break the tie.
598 /// getRelevantLoop - Get the most relevant loop associated with the given
599 /// expression, according to PickMostRelevantLoop.
600 const Loop *SCEVExpander::getRelevantLoop(const SCEV *S) {
601 // Test whether we've already computed the most relevant loop for this SCEV.
602 auto Pair = RelevantLoops.insert(std::make_pair(S, nullptr));
604 return Pair.first->second;
606 if (isa<SCEVConstant>(S))
607 // A constant has no relevant loops.
609 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
610 if (const Instruction *I = dyn_cast<Instruction>(U->getValue()))
611 return Pair.first->second = SE.LI.getLoopFor(I->getParent());
612 // A non-instruction has no relevant loops.
615 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S)) {
616 const Loop *L = nullptr;
617 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
619 for (const SCEV *Op : N->operands())
620 L = PickMostRelevantLoop(L, getRelevantLoop(Op), SE.DT);
621 return RelevantLoops[N] = L;
623 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S)) {
624 const Loop *Result = getRelevantLoop(C->getOperand());
625 return RelevantLoops[C] = Result;
627 if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
628 const Loop *Result = PickMostRelevantLoop(
629 getRelevantLoop(D->getLHS()), getRelevantLoop(D->getRHS()), SE.DT);
630 return RelevantLoops[D] = Result;
632 llvm_unreachable("Unexpected SCEV type!");
637 /// LoopCompare - Compare loops by PickMostRelevantLoop.
641 explicit LoopCompare(DominatorTree &dt) : DT(dt) {}
643 bool operator()(std::pair<const Loop *, const SCEV *> LHS,
644 std::pair<const Loop *, const SCEV *> RHS) const {
645 // Keep pointer operands sorted at the end.
646 if (LHS.second->getType()->isPointerTy() !=
647 RHS.second->getType()->isPointerTy())
648 return LHS.second->getType()->isPointerTy();
650 // Compare loops with PickMostRelevantLoop.
651 if (LHS.first != RHS.first)
652 return PickMostRelevantLoop(LHS.first, RHS.first, DT) != LHS.first;
654 // If one operand is a non-constant negative and the other is not,
655 // put the non-constant negative on the right so that a sub can
656 // be used instead of a negate and add.
657 if (LHS.second->isNonConstantNegative()) {
658 if (!RHS.second->isNonConstantNegative())
660 } else if (RHS.second->isNonConstantNegative())
663 // Otherwise they are equivalent according to this comparison.
670 Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) {
671 Type *Ty = SE.getEffectiveSCEVType(S->getType());
673 // Collect all the add operands in a loop, along with their associated loops.
674 // Iterate in reverse so that constants are emitted last, all else equal, and
675 // so that pointer operands are inserted first, which the code below relies on
676 // to form more involved GEPs.
677 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops;
678 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(S->op_end()),
679 E(S->op_begin()); I != E; ++I)
680 OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I));
682 // Sort by loop. Use a stable sort so that constants follow non-constants and
683 // pointer operands precede non-pointer operands.
684 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(SE.DT));
686 // Emit instructions to add all the operands. Hoist as much as possible
687 // out of loops, and form meaningful getelementptrs where possible.
688 Value *Sum = nullptr;
689 for (auto I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E;) {
690 const Loop *CurLoop = I->first;
691 const SCEV *Op = I->second;
693 // This is the first operand. Just expand it.
696 } else if (PointerType *PTy = dyn_cast<PointerType>(Sum->getType())) {
697 // The running sum expression is a pointer. Try to form a getelementptr
698 // at this level with that as the base.
699 SmallVector<const SCEV *, 4> NewOps;
700 for (; I != E && I->first == CurLoop; ++I) {
701 // If the operand is SCEVUnknown and not instructions, peek through
702 // it, to enable more of it to be folded into the GEP.
703 const SCEV *X = I->second;
704 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(X))
705 if (!isa<Instruction>(U->getValue()))
706 X = SE.getSCEV(U->getValue());
709 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, Sum);
710 } else if (PointerType *PTy = dyn_cast<PointerType>(Op->getType())) {
711 // The running sum is an integer, and there's a pointer at this level.
712 // Try to form a getelementptr. If the running sum is instructions,
713 // use a SCEVUnknown to avoid re-analyzing them.
714 SmallVector<const SCEV *, 4> NewOps;
715 NewOps.push_back(isa<Instruction>(Sum) ? SE.getUnknown(Sum) :
717 for (++I; I != E && I->first == CurLoop; ++I)
718 NewOps.push_back(I->second);
719 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, expand(Op));
720 } else if (Op->isNonConstantNegative()) {
721 // Instead of doing a negate and add, just do a subtract.
722 Value *W = expandCodeFor(SE.getNegativeSCEV(Op), Ty);
723 Sum = InsertNoopCastOfTo(Sum, Ty);
724 Sum = InsertBinop(Instruction::Sub, Sum, W);
728 Value *W = expandCodeFor(Op, Ty);
729 Sum = InsertNoopCastOfTo(Sum, Ty);
730 // Canonicalize a constant to the RHS.
731 if (isa<Constant>(Sum)) std::swap(Sum, W);
732 Sum = InsertBinop(Instruction::Add, Sum, W);
740 Value *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) {
741 Type *Ty = SE.getEffectiveSCEVType(S->getType());
743 // Collect all the mul operands in a loop, along with their associated loops.
744 // Iterate in reverse so that constants are emitted last, all else equal.
745 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops;
746 for (std::reverse_iterator<SCEVMulExpr::op_iterator> I(S->op_end()),
747 E(S->op_begin()); I != E; ++I)
748 OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I));
750 // Sort by loop. Use a stable sort so that constants follow non-constants.
751 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(SE.DT));
753 // Emit instructions to mul all the operands. Hoist as much as possible
755 Value *Prod = nullptr;
756 for (const auto &I : OpsAndLoops) {
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->getAPInt();
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 (auto I = IncV->op_begin() + 1, E = IncV->op_end(); I != E; ++I) {
884 if (isa<Constant>(*I))
886 if (Instruction *OInst = dyn_cast<Instruction>(*I)) {
887 if (!SE.DT.dominates(OInst, InsertPos))
891 // allow any kind of GEP as long as it can be hoisted.
894 // This must be a pointer addition of constants (pretty), which is already
895 // handled, or some number of address-size elements (ugly). Ugly geps
896 // have 2 operands. i1* is used by the expander to represent an
897 // address-size element.
898 if (IncV->getNumOperands() != 2)
900 unsigned AS = cast<PointerType>(IncV->getType())->getAddressSpace();
901 if (IncV->getType() != Type::getInt1PtrTy(SE.getContext(), AS)
902 && IncV->getType() != Type::getInt8PtrTy(SE.getContext(), AS))
906 return dyn_cast<Instruction>(IncV->getOperand(0));
910 /// hoistStep - Attempt to hoist a simple IV increment above InsertPos to make
911 /// it available to other uses in this loop. Recursively hoist any operands,
912 /// until we reach a value that dominates InsertPos.
913 bool SCEVExpander::hoistIVInc(Instruction *IncV, Instruction *InsertPos) {
914 if (SE.DT.dominates(IncV, InsertPos))
917 // InsertPos must itself dominate IncV so that IncV's new position satisfies
918 // its existing users.
919 if (isa<PHINode>(InsertPos) ||
920 !SE.DT.dominates(InsertPos->getParent(), IncV->getParent()))
923 if (!SE.LI.movementPreservesLCSSAForm(IncV, InsertPos))
926 // Check that the chain of IV operands leading back to Phi can be hoisted.
927 SmallVector<Instruction*, 4> IVIncs;
929 Instruction *Oper = getIVIncOperand(IncV, InsertPos, /*allowScale*/true);
932 // IncV is safe to hoist.
933 IVIncs.push_back(IncV);
935 if (SE.DT.dominates(IncV, InsertPos))
938 for (auto I = IVIncs.rbegin(), E = IVIncs.rend(); I != E; ++I) {
939 (*I)->moveBefore(InsertPos);
944 /// Determine if this cyclic phi is in a form that would have been generated by
945 /// LSR. We don't care if the phi was actually expanded in this pass, as long
946 /// as it is in a low-cost form, for example, no implied multiplication. This
947 /// should match any patterns generated by getAddRecExprPHILiterally and
949 bool SCEVExpander::isExpandedAddRecExprPHI(PHINode *PN, Instruction *IncV,
951 for(Instruction *IVOper = IncV;
952 (IVOper = getIVIncOperand(IVOper, L->getLoopPreheader()->getTerminator(),
953 /*allowScale=*/false));) {
960 /// expandIVInc - Expand an IV increment at Builder's current InsertPos.
961 /// Typically this is the LatchBlock terminator or IVIncInsertPos, but we may
962 /// need to materialize IV increments elsewhere to handle difficult situations.
963 Value *SCEVExpander::expandIVInc(PHINode *PN, Value *StepV, const Loop *L,
964 Type *ExpandTy, Type *IntTy,
967 // If the PHI is a pointer, use a GEP, otherwise use an add or sub.
968 if (ExpandTy->isPointerTy()) {
969 PointerType *GEPPtrTy = cast<PointerType>(ExpandTy);
970 // If the step isn't constant, don't use an implicitly scaled GEP, because
971 // that would require a multiply inside the loop.
972 if (!isa<ConstantInt>(StepV))
973 GEPPtrTy = PointerType::get(Type::getInt1Ty(SE.getContext()),
974 GEPPtrTy->getAddressSpace());
975 const SCEV *const StepArray[1] = { SE.getSCEV(StepV) };
976 IncV = expandAddToGEP(StepArray, StepArray+1, GEPPtrTy, IntTy, PN);
977 if (IncV->getType() != PN->getType()) {
978 IncV = Builder.CreateBitCast(IncV, PN->getType());
979 rememberInstruction(IncV);
983 Builder.CreateSub(PN, StepV, Twine(IVName) + ".iv.next") :
984 Builder.CreateAdd(PN, StepV, Twine(IVName) + ".iv.next");
985 rememberInstruction(IncV);
990 /// \brief Hoist the addrec instruction chain rooted in the loop phi above the
991 /// position. This routine assumes that this is possible (has been checked).
992 static void hoistBeforePos(DominatorTree *DT, Instruction *InstToHoist,
993 Instruction *Pos, PHINode *LoopPhi) {
995 if (DT->dominates(InstToHoist, Pos))
997 // Make sure the increment is where we want it. But don't move it
998 // down past a potential existing post-inc user.
999 InstToHoist->moveBefore(Pos);
1001 InstToHoist = cast<Instruction>(InstToHoist->getOperand(0));
1002 } while (InstToHoist != LoopPhi);
1005 /// \brief Check whether we can cheaply express the requested SCEV in terms of
1006 /// the available PHI SCEV by truncation and/or inversion of the step.
1007 static bool canBeCheaplyTransformed(ScalarEvolution &SE,
1008 const SCEVAddRecExpr *Phi,
1009 const SCEVAddRecExpr *Requested,
1011 Type *PhiTy = SE.getEffectiveSCEVType(Phi->getType());
1012 Type *RequestedTy = SE.getEffectiveSCEVType(Requested->getType());
1014 if (RequestedTy->getIntegerBitWidth() > PhiTy->getIntegerBitWidth())
1017 // Try truncate it if necessary.
1018 Phi = dyn_cast<SCEVAddRecExpr>(SE.getTruncateOrNoop(Phi, RequestedTy));
1022 // Check whether truncation will help.
1023 if (Phi == Requested) {
1028 // Check whether inverting will help: {R,+,-1} == R - {0,+,1}.
1029 if (SE.getAddExpr(Requested->getStart(),
1030 SE.getNegativeSCEV(Requested)) == Phi) {
1038 static bool IsIncrementNSW(ScalarEvolution &SE, const SCEVAddRecExpr *AR) {
1039 if (!isa<IntegerType>(AR->getType()))
1042 unsigned BitWidth = cast<IntegerType>(AR->getType())->getBitWidth();
1043 Type *WideTy = IntegerType::get(AR->getType()->getContext(), BitWidth * 2);
1044 const SCEV *Step = AR->getStepRecurrence(SE);
1045 const SCEV *OpAfterExtend = SE.getAddExpr(SE.getSignExtendExpr(Step, WideTy),
1046 SE.getSignExtendExpr(AR, WideTy));
1047 const SCEV *ExtendAfterOp =
1048 SE.getSignExtendExpr(SE.getAddExpr(AR, Step), WideTy);
1049 return ExtendAfterOp == OpAfterExtend;
1052 static bool IsIncrementNUW(ScalarEvolution &SE, const SCEVAddRecExpr *AR) {
1053 if (!isa<IntegerType>(AR->getType()))
1056 unsigned BitWidth = cast<IntegerType>(AR->getType())->getBitWidth();
1057 Type *WideTy = IntegerType::get(AR->getType()->getContext(), BitWidth * 2);
1058 const SCEV *Step = AR->getStepRecurrence(SE);
1059 const SCEV *OpAfterExtend = SE.getAddExpr(SE.getZeroExtendExpr(Step, WideTy),
1060 SE.getZeroExtendExpr(AR, WideTy));
1061 const SCEV *ExtendAfterOp =
1062 SE.getZeroExtendExpr(SE.getAddExpr(AR, Step), WideTy);
1063 return ExtendAfterOp == OpAfterExtend;
1066 /// getAddRecExprPHILiterally - Helper for expandAddRecExprLiterally. Expand
1067 /// the base addrec, which is the addrec without any non-loop-dominating
1068 /// values, and return the PHI.
1070 SCEVExpander::getAddRecExprPHILiterally(const SCEVAddRecExpr *Normalized,
1076 assert((!IVIncInsertLoop||IVIncInsertPos) && "Uninitialized insert position");
1078 // Reuse a previously-inserted PHI, if present.
1079 BasicBlock *LatchBlock = L->getLoopLatch();
1081 PHINode *AddRecPhiMatch = nullptr;
1082 Instruction *IncV = nullptr;
1086 // Only try partially matching scevs that need truncation and/or
1087 // step-inversion if we know this loop is outside the current loop.
1088 bool TryNonMatchingSCEV =
1090 SE.DT.properlyDominates(LatchBlock, IVIncInsertLoop->getHeader());
1092 for (auto &I : *L->getHeader()) {
1093 auto *PN = dyn_cast<PHINode>(&I);
1094 if (!PN || !SE.isSCEVable(PN->getType()))
1097 const SCEVAddRecExpr *PhiSCEV = dyn_cast<SCEVAddRecExpr>(SE.getSCEV(PN));
1101 bool IsMatchingSCEV = PhiSCEV == Normalized;
1102 // We only handle truncation and inversion of phi recurrences for the
1103 // expanded expression if the expanded expression's loop dominates the
1104 // loop we insert to. Check now, so we can bail out early.
1105 if (!IsMatchingSCEV && !TryNonMatchingSCEV)
1108 Instruction *TempIncV =
1109 cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock));
1111 // Check whether we can reuse this PHI node.
1113 if (!isExpandedAddRecExprPHI(PN, TempIncV, L))
1115 if (L == IVIncInsertLoop && !hoistIVInc(TempIncV, IVIncInsertPos))
1118 if (!isNormalAddRecExprPHI(PN, TempIncV, L))
1122 // Stop if we have found an exact match SCEV.
1123 if (IsMatchingSCEV) {
1127 AddRecPhiMatch = PN;
1131 // Try whether the phi can be translated into the requested form
1132 // (truncated and/or offset by a constant).
1133 if ((!TruncTy || InvertStep) &&
1134 canBeCheaplyTransformed(SE, PhiSCEV, Normalized, InvertStep)) {
1135 // Record the phi node. But don't stop we might find an exact match
1137 AddRecPhiMatch = PN;
1139 TruncTy = SE.getEffectiveSCEVType(Normalized->getType());
1143 if (AddRecPhiMatch) {
1144 // Potentially, move the increment. We have made sure in
1145 // isExpandedAddRecExprPHI or hoistIVInc that this is possible.
1146 if (L == IVIncInsertLoop)
1147 hoistBeforePos(&SE.DT, IncV, IVIncInsertPos, AddRecPhiMatch);
1149 // Ok, the add recurrence looks usable.
1150 // Remember this PHI, even in post-inc mode.
1151 InsertedValues.insert(AddRecPhiMatch);
1152 // Remember the increment.
1153 rememberInstruction(IncV);
1154 return AddRecPhiMatch;
1158 // Save the original insertion point so we can restore it when we're done.
1159 BuilderType::InsertPointGuard Guard(Builder);
1161 // Another AddRec may need to be recursively expanded below. For example, if
1162 // this AddRec is quadratic, the StepV may itself be an AddRec in this
1163 // loop. Remove this loop from the PostIncLoops set before expanding such
1164 // AddRecs. Otherwise, we cannot find a valid position for the step
1165 // (i.e. StepV can never dominate its loop header). Ideally, we could do
1166 // SavedIncLoops.swap(PostIncLoops), but we generally have a single element,
1167 // so it's not worth implementing SmallPtrSet::swap.
1168 PostIncLoopSet SavedPostIncLoops = PostIncLoops;
1169 PostIncLoops.clear();
1171 // Expand code for the start value.
1173 expandCodeFor(Normalized->getStart(), ExpandTy, &L->getHeader()->front());
1175 // StartV must be hoisted into L's preheader to dominate the new phi.
1176 assert(!isa<Instruction>(StartV) ||
1177 SE.DT.properlyDominates(cast<Instruction>(StartV)->getParent(),
1180 // Expand code for the step value. Do this before creating the PHI so that PHI
1181 // reuse code doesn't see an incomplete PHI.
1182 const SCEV *Step = Normalized->getStepRecurrence(SE);
1183 // If the stride is negative, insert a sub instead of an add for the increment
1184 // (unless it's a constant, because subtracts of constants are canonicalized
1186 bool useSubtract = !ExpandTy->isPointerTy() && Step->isNonConstantNegative();
1188 Step = SE.getNegativeSCEV(Step);
1189 // Expand the step somewhere that dominates the loop header.
1190 Value *StepV = expandCodeFor(Step, IntTy, &L->getHeader()->front());
1192 // The no-wrap behavior proved by IsIncrement(NUW|NSW) is only applicable if
1193 // we actually do emit an addition. It does not apply if we emit a
1195 bool IncrementIsNUW = !useSubtract && IsIncrementNUW(SE, Normalized);
1196 bool IncrementIsNSW = !useSubtract && IsIncrementNSW(SE, Normalized);
1199 BasicBlock *Header = L->getHeader();
1200 Builder.SetInsertPoint(Header, Header->begin());
1201 pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header);
1202 PHINode *PN = Builder.CreatePHI(ExpandTy, std::distance(HPB, HPE),
1203 Twine(IVName) + ".iv");
1204 rememberInstruction(PN);
1206 // Create the step instructions and populate the PHI.
1207 for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) {
1208 BasicBlock *Pred = *HPI;
1210 // Add a start value.
1211 if (!L->contains(Pred)) {
1212 PN->addIncoming(StartV, Pred);
1216 // Create a step value and add it to the PHI.
1217 // If IVIncInsertLoop is non-null and equal to the addrec's loop, insert the
1218 // instructions at IVIncInsertPos.
1219 Instruction *InsertPos = L == IVIncInsertLoop ?
1220 IVIncInsertPos : Pred->getTerminator();
1221 Builder.SetInsertPoint(InsertPos);
1222 Value *IncV = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract);
1224 if (isa<OverflowingBinaryOperator>(IncV)) {
1226 cast<BinaryOperator>(IncV)->setHasNoUnsignedWrap();
1228 cast<BinaryOperator>(IncV)->setHasNoSignedWrap();
1230 PN->addIncoming(IncV, Pred);
1233 // After expanding subexpressions, restore the PostIncLoops set so the caller
1234 // can ensure that IVIncrement dominates the current uses.
1235 PostIncLoops = SavedPostIncLoops;
1237 // Remember this PHI, even in post-inc mode.
1238 InsertedValues.insert(PN);
1243 Value *SCEVExpander::expandAddRecExprLiterally(const SCEVAddRecExpr *S) {
1244 Type *STy = S->getType();
1245 Type *IntTy = SE.getEffectiveSCEVType(STy);
1246 const Loop *L = S->getLoop();
1248 // Determine a normalized form of this expression, which is the expression
1249 // before any post-inc adjustment is made.
1250 const SCEVAddRecExpr *Normalized = S;
1251 if (PostIncLoops.count(L)) {
1252 PostIncLoopSet Loops;
1254 Normalized = cast<SCEVAddRecExpr>(TransformForPostIncUse(
1255 Normalize, S, nullptr, nullptr, Loops, SE, SE.DT));
1258 // Strip off any non-loop-dominating component from the addrec start.
1259 const SCEV *Start = Normalized->getStart();
1260 const SCEV *PostLoopOffset = nullptr;
1261 if (!SE.properlyDominates(Start, L->getHeader())) {
1262 PostLoopOffset = Start;
1263 Start = SE.getConstant(Normalized->getType(), 0);
1264 Normalized = cast<SCEVAddRecExpr>(
1265 SE.getAddRecExpr(Start, Normalized->getStepRecurrence(SE),
1266 Normalized->getLoop(),
1267 Normalized->getNoWrapFlags(SCEV::FlagNW)));
1270 // Strip off any non-loop-dominating component from the addrec step.
1271 const SCEV *Step = Normalized->getStepRecurrence(SE);
1272 const SCEV *PostLoopScale = nullptr;
1273 if (!SE.dominates(Step, L->getHeader())) {
1274 PostLoopScale = Step;
1275 Step = SE.getConstant(Normalized->getType(), 1);
1277 cast<SCEVAddRecExpr>(SE.getAddRecExpr(
1278 Start, Step, Normalized->getLoop(),
1279 Normalized->getNoWrapFlags(SCEV::FlagNW)));
1282 // Expand the core addrec. If we need post-loop scaling, force it to
1283 // expand to an integer type to avoid the need for additional casting.
1284 Type *ExpandTy = PostLoopScale ? IntTy : STy;
1285 // In some cases, we decide to reuse an existing phi node but need to truncate
1286 // it and/or invert the step.
1287 Type *TruncTy = nullptr;
1288 bool InvertStep = false;
1289 PHINode *PN = getAddRecExprPHILiterally(Normalized, L, ExpandTy, IntTy,
1290 TruncTy, InvertStep);
1292 // Accommodate post-inc mode, if necessary.
1294 if (!PostIncLoops.count(L))
1297 // In PostInc mode, use the post-incremented value.
1298 BasicBlock *LatchBlock = L->getLoopLatch();
1299 assert(LatchBlock && "PostInc mode requires a unique loop latch!");
1300 Result = PN->getIncomingValueForBlock(LatchBlock);
1302 // For an expansion to use the postinc form, the client must call
1303 // expandCodeFor with an InsertPoint that is either outside the PostIncLoop
1304 // or dominated by IVIncInsertPos.
1305 if (isa<Instruction>(Result) &&
1306 !SE.DT.dominates(cast<Instruction>(Result),
1307 &*Builder.GetInsertPoint())) {
1308 // The induction variable's postinc expansion does not dominate this use.
1309 // IVUsers tries to prevent this case, so it is rare. However, it can
1310 // happen when an IVUser outside the loop is not dominated by the latch
1311 // block. Adjusting IVIncInsertPos before expansion begins cannot handle
1312 // all cases. Consider a phi outide whose operand is replaced during
1313 // expansion with the value of the postinc user. Without fundamentally
1314 // changing the way postinc users are tracked, the only remedy is
1315 // inserting an extra IV increment. StepV might fold into PostLoopOffset,
1316 // but hopefully expandCodeFor handles that.
1318 !ExpandTy->isPointerTy() && Step->isNonConstantNegative();
1320 Step = SE.getNegativeSCEV(Step);
1323 // Expand the step somewhere that dominates the loop header.
1324 BuilderType::InsertPointGuard Guard(Builder);
1325 StepV = expandCodeFor(Step, IntTy, &L->getHeader()->front());
1327 Result = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract);
1331 // We have decided to reuse an induction variable of a dominating loop. Apply
1332 // truncation and/or invertion of the step.
1334 Type *ResTy = Result->getType();
1335 // Normalize the result type.
1336 if (ResTy != SE.getEffectiveSCEVType(ResTy))
1337 Result = InsertNoopCastOfTo(Result, SE.getEffectiveSCEVType(ResTy));
1338 // Truncate the result.
1339 if (TruncTy != Result->getType()) {
1340 Result = Builder.CreateTrunc(Result, TruncTy);
1341 rememberInstruction(Result);
1343 // Invert the result.
1345 Result = Builder.CreateSub(expandCodeFor(Normalized->getStart(), TruncTy),
1347 rememberInstruction(Result);
1351 // Re-apply any non-loop-dominating scale.
1352 if (PostLoopScale) {
1353 assert(S->isAffine() && "Can't linearly scale non-affine recurrences.");
1354 Result = InsertNoopCastOfTo(Result, IntTy);
1355 Result = Builder.CreateMul(Result,
1356 expandCodeFor(PostLoopScale, IntTy));
1357 rememberInstruction(Result);
1360 // Re-apply any non-loop-dominating offset.
1361 if (PostLoopOffset) {
1362 if (PointerType *PTy = dyn_cast<PointerType>(ExpandTy)) {
1363 const SCEV *const OffsetArray[1] = { PostLoopOffset };
1364 Result = expandAddToGEP(OffsetArray, OffsetArray+1, PTy, IntTy, Result);
1366 Result = InsertNoopCastOfTo(Result, IntTy);
1367 Result = Builder.CreateAdd(Result,
1368 expandCodeFor(PostLoopOffset, IntTy));
1369 rememberInstruction(Result);
1376 Value *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) {
1377 if (!CanonicalMode) return expandAddRecExprLiterally(S);
1379 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1380 const Loop *L = S->getLoop();
1382 // First check for an existing canonical IV in a suitable type.
1383 PHINode *CanonicalIV = nullptr;
1384 if (PHINode *PN = L->getCanonicalInductionVariable())
1385 if (SE.getTypeSizeInBits(PN->getType()) >= SE.getTypeSizeInBits(Ty))
1388 // Rewrite an AddRec in terms of the canonical induction variable, if
1389 // its type is more narrow.
1391 SE.getTypeSizeInBits(CanonicalIV->getType()) >
1392 SE.getTypeSizeInBits(Ty)) {
1393 SmallVector<const SCEV *, 4> NewOps(S->getNumOperands());
1394 for (unsigned i = 0, e = S->getNumOperands(); i != e; ++i)
1395 NewOps[i] = SE.getAnyExtendExpr(S->op_begin()[i], CanonicalIV->getType());
1396 Value *V = expand(SE.getAddRecExpr(NewOps, S->getLoop(),
1397 S->getNoWrapFlags(SCEV::FlagNW)));
1398 BasicBlock::iterator NewInsertPt =
1399 findInsertPointAfter(cast<Instruction>(V), Builder.GetInsertBlock());
1400 V = expandCodeFor(SE.getTruncateExpr(SE.getUnknown(V), Ty), nullptr,
1405 // {X,+,F} --> X + {0,+,F}
1406 if (!S->getStart()->isZero()) {
1407 SmallVector<const SCEV *, 4> NewOps(S->op_begin(), S->op_end());
1408 NewOps[0] = SE.getConstant(Ty, 0);
1409 const SCEV *Rest = SE.getAddRecExpr(NewOps, L,
1410 S->getNoWrapFlags(SCEV::FlagNW));
1412 // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the
1413 // comments on expandAddToGEP for details.
1414 const SCEV *Base = S->getStart();
1415 const SCEV *RestArray[1] = { Rest };
1416 // Dig into the expression to find the pointer base for a GEP.
1417 ExposePointerBase(Base, RestArray[0], SE);
1418 // If we found a pointer, expand the AddRec with a GEP.
1419 if (PointerType *PTy = dyn_cast<PointerType>(Base->getType())) {
1420 // Make sure the Base isn't something exotic, such as a multiplied
1421 // or divided pointer value. In those cases, the result type isn't
1422 // actually a pointer type.
1423 if (!isa<SCEVMulExpr>(Base) && !isa<SCEVUDivExpr>(Base)) {
1424 Value *StartV = expand(Base);
1425 assert(StartV->getType() == PTy && "Pointer type mismatch for GEP!");
1426 return expandAddToGEP(RestArray, RestArray+1, PTy, Ty, StartV);
1430 // Just do a normal add. Pre-expand the operands to suppress folding.
1431 return expand(SE.getAddExpr(SE.getUnknown(expand(S->getStart())),
1432 SE.getUnknown(expand(Rest))));
1435 // If we don't yet have a canonical IV, create one.
1437 // Create and insert the PHI node for the induction variable in the
1439 BasicBlock *Header = L->getHeader();
1440 pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header);
1441 CanonicalIV = PHINode::Create(Ty, std::distance(HPB, HPE), "indvar",
1443 rememberInstruction(CanonicalIV);
1445 SmallSet<BasicBlock *, 4> PredSeen;
1446 Constant *One = ConstantInt::get(Ty, 1);
1447 for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) {
1448 BasicBlock *HP = *HPI;
1449 if (!PredSeen.insert(HP).second) {
1450 // There must be an incoming value for each predecessor, even the
1452 CanonicalIV->addIncoming(CanonicalIV->getIncomingValueForBlock(HP), HP);
1456 if (L->contains(HP)) {
1457 // Insert a unit add instruction right before the terminator
1458 // corresponding to the back-edge.
1459 Instruction *Add = BinaryOperator::CreateAdd(CanonicalIV, One,
1461 HP->getTerminator());
1462 Add->setDebugLoc(HP->getTerminator()->getDebugLoc());
1463 rememberInstruction(Add);
1464 CanonicalIV->addIncoming(Add, HP);
1466 CanonicalIV->addIncoming(Constant::getNullValue(Ty), HP);
1471 // {0,+,1} --> Insert a canonical induction variable into the loop!
1472 if (S->isAffine() && S->getOperand(1)->isOne()) {
1473 assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) &&
1474 "IVs with types different from the canonical IV should "
1475 "already have been handled!");
1479 // {0,+,F} --> {0,+,1} * F
1481 // If this is a simple linear addrec, emit it now as a special case.
1482 if (S->isAffine()) // {0,+,F} --> i*F
1484 expand(SE.getTruncateOrNoop(
1485 SE.getMulExpr(SE.getUnknown(CanonicalIV),
1486 SE.getNoopOrAnyExtend(S->getOperand(1),
1487 CanonicalIV->getType())),
1490 // If this is a chain of recurrences, turn it into a closed form, using the
1491 // folders, then expandCodeFor the closed form. This allows the folders to
1492 // simplify the expression without having to build a bunch of special code
1493 // into this folder.
1494 const SCEV *IH = SE.getUnknown(CanonicalIV); // Get I as a "symbolic" SCEV.
1496 // Promote S up to the canonical IV type, if the cast is foldable.
1497 const SCEV *NewS = S;
1498 const SCEV *Ext = SE.getNoopOrAnyExtend(S, CanonicalIV->getType());
1499 if (isa<SCEVAddRecExpr>(Ext))
1502 const SCEV *V = cast<SCEVAddRecExpr>(NewS)->evaluateAtIteration(IH, SE);
1503 //cerr << "Evaluated: " << *this << "\n to: " << *V << "\n";
1505 // Truncate the result down to the original type, if needed.
1506 const SCEV *T = SE.getTruncateOrNoop(V, Ty);
1510 Value *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) {
1511 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1512 Value *V = expandCodeFor(S->getOperand(),
1513 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1514 Value *I = Builder.CreateTrunc(V, Ty);
1515 rememberInstruction(I);
1519 Value *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) {
1520 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1521 Value *V = expandCodeFor(S->getOperand(),
1522 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1523 Value *I = Builder.CreateZExt(V, Ty);
1524 rememberInstruction(I);
1528 Value *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) {
1529 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1530 Value *V = expandCodeFor(S->getOperand(),
1531 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1532 Value *I = Builder.CreateSExt(V, Ty);
1533 rememberInstruction(I);
1537 Value *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) {
1538 Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
1539 Type *Ty = LHS->getType();
1540 for (int i = S->getNumOperands()-2; i >= 0; --i) {
1541 // In the case of mixed integer and pointer types, do the
1542 // rest of the comparisons as integer.
1543 if (S->getOperand(i)->getType() != Ty) {
1544 Ty = SE.getEffectiveSCEVType(Ty);
1545 LHS = InsertNoopCastOfTo(LHS, Ty);
1547 Value *RHS = expandCodeFor(S->getOperand(i), Ty);
1548 Value *ICmp = Builder.CreateICmpSGT(LHS, RHS);
1549 rememberInstruction(ICmp);
1550 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "smax");
1551 rememberInstruction(Sel);
1554 // In the case of mixed integer and pointer types, cast the
1555 // final result back to the pointer type.
1556 if (LHS->getType() != S->getType())
1557 LHS = InsertNoopCastOfTo(LHS, S->getType());
1561 Value *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) {
1562 Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
1563 Type *Ty = LHS->getType();
1564 for (int i = S->getNumOperands()-2; i >= 0; --i) {
1565 // In the case of mixed integer and pointer types, do the
1566 // rest of the comparisons as integer.
1567 if (S->getOperand(i)->getType() != Ty) {
1568 Ty = SE.getEffectiveSCEVType(Ty);
1569 LHS = InsertNoopCastOfTo(LHS, Ty);
1571 Value *RHS = expandCodeFor(S->getOperand(i), Ty);
1572 Value *ICmp = Builder.CreateICmpUGT(LHS, RHS);
1573 rememberInstruction(ICmp);
1574 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "umax");
1575 rememberInstruction(Sel);
1578 // In the case of mixed integer and pointer types, cast the
1579 // final result back to the pointer type.
1580 if (LHS->getType() != S->getType())
1581 LHS = InsertNoopCastOfTo(LHS, S->getType());
1585 Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty,
1588 Builder.SetInsertPoint(IP);
1589 return expandCodeFor(SH, Ty);
1592 Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty) {
1593 // Expand the code for this SCEV.
1594 Value *V = expand(SH);
1596 assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) &&
1597 "non-trivial casts should be done with the SCEVs directly!");
1598 V = InsertNoopCastOfTo(V, Ty);
1603 Value *SCEVExpander::expand(const SCEV *S) {
1604 // Compute an insertion point for this SCEV object. Hoist the instructions
1605 // as far out in the loop nest as possible.
1606 Instruction *InsertPt = &*Builder.GetInsertPoint();
1607 for (Loop *L = SE.LI.getLoopFor(Builder.GetInsertBlock());;
1608 L = L->getParentLoop())
1609 if (SE.isLoopInvariant(S, L)) {
1611 if (BasicBlock *Preheader = L->getLoopPreheader())
1612 InsertPt = Preheader->getTerminator();
1614 // LSR sets the insertion point for AddRec start/step values to the
1615 // block start to simplify value reuse, even though it's an invalid
1616 // position. SCEVExpander must correct for this in all cases.
1617 InsertPt = &*L->getHeader()->getFirstInsertionPt();
1620 // If the SCEV is computable at this level, insert it into the header
1621 // after the PHIs (and after any other instructions that we've inserted
1622 // there) so that it is guaranteed to dominate any user inside the loop.
1623 if (L && SE.hasComputableLoopEvolution(S, L) && !PostIncLoops.count(L))
1624 InsertPt = &*L->getHeader()->getFirstInsertionPt();
1625 while (InsertPt != Builder.GetInsertPoint()
1626 && (isInsertedInstruction(InsertPt)
1627 || isa<DbgInfoIntrinsic>(InsertPt))) {
1628 InsertPt = &*std::next(InsertPt->getIterator());
1633 // Check to see if we already expanded this here.
1634 auto I = InsertedExpressions.find(std::make_pair(S, InsertPt));
1635 if (I != InsertedExpressions.end())
1638 BuilderType::InsertPointGuard Guard(Builder);
1639 Builder.SetInsertPoint(InsertPt);
1641 // Expand the expression into instructions.
1642 Value *V = visit(S);
1644 // Remember the expanded value for this SCEV at this location.
1646 // This is independent of PostIncLoops. The mapped value simply materializes
1647 // the expression at this insertion point. If the mapped value happened to be
1648 // a postinc expansion, it could be reused by a non-postinc user, but only if
1649 // its insertion point was already at the head of the loop.
1650 InsertedExpressions[std::make_pair(S, InsertPt)] = V;
1654 void SCEVExpander::rememberInstruction(Value *I) {
1655 if (!PostIncLoops.empty())
1656 InsertedPostIncValues.insert(I);
1658 InsertedValues.insert(I);
1661 /// getOrInsertCanonicalInductionVariable - This method returns the
1662 /// canonical induction variable of the specified type for the specified
1663 /// loop (inserting one if there is none). A canonical induction variable
1664 /// starts at zero and steps by one on each iteration.
1666 SCEVExpander::getOrInsertCanonicalInductionVariable(const Loop *L,
1668 assert(Ty->isIntegerTy() && "Can only insert integer induction variables!");
1670 // Build a SCEV for {0,+,1}<L>.
1671 // Conservatively use FlagAnyWrap for now.
1672 const SCEV *H = SE.getAddRecExpr(SE.getConstant(Ty, 0),
1673 SE.getConstant(Ty, 1), L, SCEV::FlagAnyWrap);
1675 // Emit code for it.
1676 BuilderType::InsertPointGuard Guard(Builder);
1678 cast<PHINode>(expandCodeFor(H, nullptr, &L->getHeader()->front()));
1683 /// replaceCongruentIVs - Check for congruent phis in this loop header and
1684 /// replace them with their most canonical representative. Return the number of
1685 /// phis eliminated.
1687 /// This does not depend on any SCEVExpander state but should be used in
1688 /// the same context that SCEVExpander is used.
1689 unsigned SCEVExpander::replaceCongruentIVs(Loop *L, const DominatorTree *DT,
1690 SmallVectorImpl<WeakVH> &DeadInsts,
1691 const TargetTransformInfo *TTI) {
1692 // Find integer phis in order of increasing width.
1693 SmallVector<PHINode*, 8> Phis;
1694 for (auto &I : *L->getHeader()) {
1695 if (auto *PN = dyn_cast<PHINode>(&I))
1702 std::sort(Phis.begin(), Phis.end(), [](Value *LHS, Value *RHS) {
1703 // Put pointers at the back and make sure pointer < pointer = false.
1704 if (!LHS->getType()->isIntegerTy() || !RHS->getType()->isIntegerTy())
1705 return RHS->getType()->isIntegerTy() && !LHS->getType()->isIntegerTy();
1706 return RHS->getType()->getPrimitiveSizeInBits() <
1707 LHS->getType()->getPrimitiveSizeInBits();
1710 unsigned NumElim = 0;
1711 DenseMap<const SCEV *, PHINode *> ExprToIVMap;
1712 // Process phis from wide to narrow. Map wide phis to their truncation
1713 // so narrow phis can reuse them.
1714 for (PHINode *Phi : Phis) {
1715 auto SimplifyPHINode = [&](PHINode *PN) -> Value * {
1716 if (Value *V = SimplifyInstruction(PN, DL, &SE.TLI, &SE.DT, &SE.AC))
1718 if (!SE.isSCEVable(PN->getType()))
1720 auto *Const = dyn_cast<SCEVConstant>(SE.getSCEV(PN));
1723 return Const->getValue();
1726 // Fold constant phis. They may be congruent to other constant phis and
1727 // would confuse the logic below that expects proper IVs.
1728 if (Value *V = SimplifyPHINode(Phi)) {
1729 if (V->getType() != Phi->getType())
1731 Phi->replaceAllUsesWith(V);
1732 DeadInsts.emplace_back(Phi);
1734 DEBUG_WITH_TYPE(DebugType, dbgs()
1735 << "INDVARS: Eliminated constant iv: " << *Phi << '\n');
1739 if (!SE.isSCEVable(Phi->getType()))
1742 PHINode *&OrigPhiRef = ExprToIVMap[SE.getSCEV(Phi)];
1745 if (Phi->getType()->isIntegerTy() && TTI
1746 && TTI->isTruncateFree(Phi->getType(), Phis.back()->getType())) {
1747 // This phi can be freely truncated to the narrowest phi type. Map the
1748 // truncated expression to it so it will be reused for narrow types.
1749 const SCEV *TruncExpr =
1750 SE.getTruncateExpr(SE.getSCEV(Phi), Phis.back()->getType());
1751 ExprToIVMap[TruncExpr] = Phi;
1756 // Replacing a pointer phi with an integer phi or vice-versa doesn't make
1758 if (OrigPhiRef->getType()->isPointerTy() != Phi->getType()->isPointerTy())
1761 if (BasicBlock *LatchBlock = L->getLoopLatch()) {
1762 Instruction *OrigInc =
1763 cast<Instruction>(OrigPhiRef->getIncomingValueForBlock(LatchBlock));
1764 Instruction *IsomorphicInc =
1765 cast<Instruction>(Phi->getIncomingValueForBlock(LatchBlock));
1767 // If this phi has the same width but is more canonical, replace the
1768 // original with it. As part of the "more canonical" determination,
1769 // respect a prior decision to use an IV chain.
1770 if (OrigPhiRef->getType() == Phi->getType()
1771 && !(ChainedPhis.count(Phi)
1772 || isExpandedAddRecExprPHI(OrigPhiRef, OrigInc, L))
1773 && (ChainedPhis.count(Phi)
1774 || isExpandedAddRecExprPHI(Phi, IsomorphicInc, L))) {
1775 std::swap(OrigPhiRef, Phi);
1776 std::swap(OrigInc, IsomorphicInc);
1778 // Replacing the congruent phi is sufficient because acyclic redundancy
1779 // elimination, CSE/GVN, should handle the rest. However, once SCEV proves
1780 // that a phi is congruent, it's often the head of an IV user cycle that
1781 // is isomorphic with the original phi. It's worth eagerly cleaning up the
1782 // common case of a single IV increment so that DeleteDeadPHIs can remove
1783 // cycles that had postinc uses.
1784 const SCEV *TruncExpr = SE.getTruncateOrNoop(SE.getSCEV(OrigInc),
1785 IsomorphicInc->getType());
1786 if (OrigInc != IsomorphicInc
1787 && TruncExpr == SE.getSCEV(IsomorphicInc)
1788 && ((isa<PHINode>(OrigInc) && isa<PHINode>(IsomorphicInc))
1789 || hoistIVInc(OrigInc, IsomorphicInc))) {
1790 DEBUG_WITH_TYPE(DebugType, dbgs()
1791 << "INDVARS: Eliminated congruent iv.inc: "
1792 << *IsomorphicInc << '\n');
1793 Value *NewInc = OrigInc;
1794 if (OrigInc->getType() != IsomorphicInc->getType()) {
1795 Instruction *IP = nullptr;
1796 if (PHINode *PN = dyn_cast<PHINode>(OrigInc))
1797 IP = &*PN->getParent()->getFirstInsertionPt();
1799 IP = OrigInc->getNextNode();
1801 IRBuilder<> Builder(IP);
1802 Builder.SetCurrentDebugLocation(IsomorphicInc->getDebugLoc());
1804 CreateTruncOrBitCast(OrigInc, IsomorphicInc->getType(), IVName);
1806 IsomorphicInc->replaceAllUsesWith(NewInc);
1807 DeadInsts.emplace_back(IsomorphicInc);
1810 DEBUG_WITH_TYPE(DebugType, dbgs()
1811 << "INDVARS: Eliminated congruent iv: " << *Phi << '\n');
1813 Value *NewIV = OrigPhiRef;
1814 if (OrigPhiRef->getType() != Phi->getType()) {
1815 IRBuilder<> Builder(&*L->getHeader()->getFirstInsertionPt());
1816 Builder.SetCurrentDebugLocation(Phi->getDebugLoc());
1817 NewIV = Builder.CreateTruncOrBitCast(OrigPhiRef, Phi->getType(), IVName);
1819 Phi->replaceAllUsesWith(NewIV);
1820 DeadInsts.emplace_back(Phi);
1825 Value *SCEVExpander::findExistingExpansion(const SCEV *S,
1826 const Instruction *At, Loop *L) {
1827 using namespace llvm::PatternMatch;
1829 SmallVector<BasicBlock *, 4> ExitingBlocks;
1830 L->getExitingBlocks(ExitingBlocks);
1832 // Look for suitable value in simple conditions at the loop exits.
1833 for (BasicBlock *BB : ExitingBlocks) {
1834 ICmpInst::Predicate Pred;
1835 Instruction *LHS, *RHS;
1836 BasicBlock *TrueBB, *FalseBB;
1838 if (!match(BB->getTerminator(),
1839 m_Br(m_ICmp(Pred, m_Instruction(LHS), m_Instruction(RHS)),
1843 if (SE.getSCEV(LHS) == S && SE.DT.dominates(LHS, At))
1846 if (SE.getSCEV(RHS) == S && SE.DT.dominates(RHS, At))
1850 // There is potential to make this significantly smarter, but this simple
1851 // heuristic already gets some interesting cases.
1853 // Can not find suitable value.
1857 bool SCEVExpander::isHighCostExpansionHelper(
1858 const SCEV *S, Loop *L, const Instruction *At,
1859 SmallPtrSetImpl<const SCEV *> &Processed) {
1861 // If we can find an existing value for this scev avaliable at the point "At"
1862 // then consider the expression cheap.
1863 if (At && findExistingExpansion(S, At, L) != nullptr)
1866 // Zero/One operand expressions
1867 switch (S->getSCEVType()) {
1872 return isHighCostExpansionHelper(cast<SCEVTruncateExpr>(S)->getOperand(),
1875 return isHighCostExpansionHelper(cast<SCEVZeroExtendExpr>(S)->getOperand(),
1878 return isHighCostExpansionHelper(cast<SCEVSignExtendExpr>(S)->getOperand(),
1882 if (!Processed.insert(S).second)
1885 if (auto *UDivExpr = dyn_cast<SCEVUDivExpr>(S)) {
1886 // If the divisor is a power of two and the SCEV type fits in a native
1887 // integer, consider the division cheap irrespective of whether it occurs in
1888 // the user code since it can be lowered into a right shift.
1889 if (auto *SC = dyn_cast<SCEVConstant>(UDivExpr->getRHS()))
1890 if (SC->getAPInt().isPowerOf2()) {
1891 const DataLayout &DL =
1892 L->getHeader()->getParent()->getParent()->getDataLayout();
1893 unsigned Width = cast<IntegerType>(UDivExpr->getType())->getBitWidth();
1894 return DL.isIllegalInteger(Width);
1897 // UDivExpr is very likely a UDiv that ScalarEvolution's HowFarToZero or
1898 // HowManyLessThans produced to compute a precise expression, rather than a
1899 // UDiv from the user's code. If we can't find a UDiv in the code with some
1900 // simple searching, assume the former consider UDivExpr expensive to
1902 BasicBlock *ExitingBB = L->getExitingBlock();
1906 // At the beginning of this function we already tried to find existing value
1907 // for plain 'S'. Now try to lookup 'S + 1' since it is common pattern
1908 // involving division. This is just a simple search heuristic.
1910 At = &ExitingBB->back();
1911 if (!findExistingExpansion(
1912 SE.getAddExpr(S, SE.getConstant(S->getType(), 1)), At, L))
1916 // HowManyLessThans uses a Max expression whenever the loop is not guarded by
1917 // the exit condition.
1918 if (isa<SCEVSMaxExpr>(S) || isa<SCEVUMaxExpr>(S))
1921 // Recurse past nary expressions, which commonly occur in the
1922 // BackedgeTakenCount. They may already exist in program code, and if not,
1923 // they are not too expensive rematerialize.
1924 if (const SCEVNAryExpr *NAry = dyn_cast<SCEVNAryExpr>(S)) {
1925 for (auto *Op : NAry->operands())
1926 if (isHighCostExpansionHelper(Op, L, At, Processed))
1930 // If we haven't recognized an expensive SCEV pattern, assume it's an
1931 // expression produced by program code.
1935 Value *SCEVExpander::expandCodeForPredicate(const SCEVPredicate *Pred,
1938 switch (Pred->getKind()) {
1939 case SCEVPredicate::P_Union:
1940 return expandUnionPredicate(cast<SCEVUnionPredicate>(Pred), IP);
1941 case SCEVPredicate::P_Equal:
1942 return expandEqualPredicate(cast<SCEVEqualPredicate>(Pred), IP);
1944 llvm_unreachable("Unknown SCEV predicate type");
1947 Value *SCEVExpander::expandEqualPredicate(const SCEVEqualPredicate *Pred,
1949 Value *Expr0 = expandCodeFor(Pred->getLHS(), Pred->getLHS()->getType(), IP);
1950 Value *Expr1 = expandCodeFor(Pred->getRHS(), Pred->getRHS()->getType(), IP);
1952 Builder.SetInsertPoint(IP);
1953 auto *I = Builder.CreateICmpNE(Expr0, Expr1, "ident.check");
1957 Value *SCEVExpander::expandUnionPredicate(const SCEVUnionPredicate *Union,
1959 auto *BoolType = IntegerType::get(IP->getContext(), 1);
1960 Value *Check = ConstantInt::getNullValue(BoolType);
1962 // Loop over all checks in this set.
1963 for (auto Pred : Union->getPredicates()) {
1964 auto *NextCheck = expandCodeForPredicate(Pred, IP);
1965 Builder.SetInsertPoint(IP);
1966 Check = Builder.CreateOr(Check, NextCheck);
1973 // Search for a SCEV subexpression that is not safe to expand. Any expression
1974 // that may expand to a !isSafeToSpeculativelyExecute value is unsafe, namely
1975 // UDiv expressions. We don't know if the UDiv is derived from an IR divide
1976 // instruction, but the important thing is that we prove the denominator is
1977 // nonzero before expansion.
1979 // IVUsers already checks that IV-derived expressions are safe. So this check is
1980 // only needed when the expression includes some subexpression that is not IV
1983 // Currently, we only allow division by a nonzero constant here. If this is
1984 // inadequate, we could easily allow division by SCEVUnknown by using
1985 // ValueTracking to check isKnownNonZero().
1987 // We cannot generally expand recurrences unless the step dominates the loop
1988 // header. The expander handles the special case of affine recurrences by
1989 // scaling the recurrence outside the loop, but this technique isn't generally
1990 // applicable. Expanding a nested recurrence outside a loop requires computing
1991 // binomial coefficients. This could be done, but the recurrence has to be in a
1992 // perfectly reduced form, which can't be guaranteed.
1993 struct SCEVFindUnsafe {
1994 ScalarEvolution &SE;
1997 SCEVFindUnsafe(ScalarEvolution &se): SE(se), IsUnsafe(false) {}
1999 bool follow(const SCEV *S) {
2000 if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
2001 const SCEVConstant *SC = dyn_cast<SCEVConstant>(D->getRHS());
2002 if (!SC || SC->getValue()->isZero()) {
2007 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2008 const SCEV *Step = AR->getStepRecurrence(SE);
2009 if (!AR->isAffine() && !SE.dominates(Step, AR->getLoop()->getHeader())) {
2016 bool isDone() const { return IsUnsafe; }
2021 bool isSafeToExpand(const SCEV *S, ScalarEvolution &SE) {
2022 SCEVFindUnsafe Search(SE);
2023 visitAll(S, Search);
2024 return !Search.IsUnsafe;