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/SmallSet.h"
18 #include "llvm/ADT/STLExtras.h"
19 #include "llvm/Analysis/LoopInfo.h"
20 #include "llvm/Analysis/TargetTransformInfo.h"
21 #include "llvm/IR/DataLayout.h"
22 #include "llvm/IR/IntrinsicInst.h"
23 #include "llvm/IR/LLVMContext.h"
24 #include "llvm/Support/Debug.h"
28 /// ReuseOrCreateCast - Arrange for there to be a cast of V to Ty at IP,
29 /// reusing an existing cast if a suitable one exists, moving an existing
30 /// cast if a suitable one exists but isn't in the right place, or
31 /// creating a new one.
32 Value *SCEVExpander::ReuseOrCreateCast(Value *V, Type *Ty,
33 Instruction::CastOps Op,
34 BasicBlock::iterator IP) {
35 // This function must be called with the builder having a valid insertion
36 // point. It doesn't need to be the actual IP where the uses of the returned
37 // cast will be added, but it must dominate such IP.
38 // We use this precondition to produce a cast that will dominate all its
39 // uses. In particular, this is crucial for the case where the builder's
40 // insertion point *is* the point where we were asked to put the cast.
41 // Since we don't know the builder's insertion point is actually
42 // where the uses will be added (only that it dominates it), we are
43 // not allowed to move it.
44 BasicBlock::iterator BIP = Builder.GetInsertPoint();
46 Instruction *Ret = NULL;
48 // Check to see if there is already a cast!
49 for (Value::use_iterator UI = V->use_begin(), E = V->use_end();
52 if (U->getType() == Ty)
53 if (CastInst *CI = dyn_cast<CastInst>(U))
54 if (CI->getOpcode() == Op) {
55 // If the cast isn't where we want it, create a new cast at IP.
56 // Likewise, do not reuse a cast at BIP because it must dominate
57 // instructions that might be inserted before BIP.
58 if (BasicBlock::iterator(CI) != IP || BIP == IP) {
59 // Create a new cast, and leave the old cast in place in case
60 // it is being used as an insert point. Clear its operand
61 // so that it doesn't hold anything live.
62 Ret = CastInst::Create(Op, V, Ty, "", IP);
64 CI->replaceAllUsesWith(Ret);
65 CI->setOperand(0, UndefValue::get(V->getType()));
75 Ret = CastInst::Create(Op, V, Ty, V->getName(), IP);
77 // We assert at the end of the function since IP might point to an
78 // instruction with different dominance properties than a cast
79 // (an invoke for example) and not dominate BIP (but the cast does).
80 assert(SE.DT->dominates(Ret, BIP));
82 rememberInstruction(Ret);
86 /// InsertNoopCastOfTo - Insert a cast of V to the specified type,
87 /// which must be possible with a noop cast, doing what we can to share
89 Value *SCEVExpander::InsertNoopCastOfTo(Value *V, Type *Ty) {
90 Instruction::CastOps Op = CastInst::getCastOpcode(V, false, Ty, false);
91 assert((Op == Instruction::BitCast ||
92 Op == Instruction::PtrToInt ||
93 Op == Instruction::IntToPtr) &&
94 "InsertNoopCastOfTo cannot perform non-noop casts!");
95 assert(SE.getTypeSizeInBits(V->getType()) == SE.getTypeSizeInBits(Ty) &&
96 "InsertNoopCastOfTo cannot change sizes!");
98 // Short-circuit unnecessary bitcasts.
99 if (Op == Instruction::BitCast) {
100 if (V->getType() == Ty)
102 if (CastInst *CI = dyn_cast<CastInst>(V)) {
103 if (CI->getOperand(0)->getType() == Ty)
104 return CI->getOperand(0);
107 // Short-circuit unnecessary inttoptr<->ptrtoint casts.
108 if ((Op == Instruction::PtrToInt || Op == Instruction::IntToPtr) &&
109 SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(V->getType())) {
110 if (CastInst *CI = dyn_cast<CastInst>(V))
111 if ((CI->getOpcode() == Instruction::PtrToInt ||
112 CI->getOpcode() == Instruction::IntToPtr) &&
113 SE.getTypeSizeInBits(CI->getType()) ==
114 SE.getTypeSizeInBits(CI->getOperand(0)->getType()))
115 return CI->getOperand(0);
116 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
117 if ((CE->getOpcode() == Instruction::PtrToInt ||
118 CE->getOpcode() == Instruction::IntToPtr) &&
119 SE.getTypeSizeInBits(CE->getType()) ==
120 SE.getTypeSizeInBits(CE->getOperand(0)->getType()))
121 return CE->getOperand(0);
124 // Fold a cast of a constant.
125 if (Constant *C = dyn_cast<Constant>(V))
126 return ConstantExpr::getCast(Op, C, Ty);
128 // Cast the argument at the beginning of the entry block, after
129 // any bitcasts of other arguments.
130 if (Argument *A = dyn_cast<Argument>(V)) {
131 BasicBlock::iterator IP = A->getParent()->getEntryBlock().begin();
132 while ((isa<BitCastInst>(IP) &&
133 isa<Argument>(cast<BitCastInst>(IP)->getOperand(0)) &&
134 cast<BitCastInst>(IP)->getOperand(0) != A) ||
135 isa<DbgInfoIntrinsic>(IP) ||
136 isa<LandingPadInst>(IP))
138 return ReuseOrCreateCast(A, Ty, Op, IP);
141 // Cast the instruction immediately after the instruction.
142 Instruction *I = cast<Instruction>(V);
143 BasicBlock::iterator IP = I; ++IP;
144 if (InvokeInst *II = dyn_cast<InvokeInst>(I))
145 IP = II->getNormalDest()->begin();
146 while (isa<PHINode>(IP) || isa<LandingPadInst>(IP))
148 return ReuseOrCreateCast(I, Ty, Op, IP);
151 /// InsertBinop - Insert the specified binary operator, doing a small amount
152 /// of work to avoid inserting an obviously redundant operation.
153 Value *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode,
154 Value *LHS, Value *RHS) {
155 // Fold a binop with constant operands.
156 if (Constant *CLHS = dyn_cast<Constant>(LHS))
157 if (Constant *CRHS = dyn_cast<Constant>(RHS))
158 return ConstantExpr::get(Opcode, CLHS, CRHS);
160 // Do a quick scan to see if we have this binop nearby. If so, reuse it.
161 unsigned ScanLimit = 6;
162 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
163 // Scanning starts from the last instruction before the insertion point.
164 BasicBlock::iterator IP = Builder.GetInsertPoint();
165 if (IP != BlockBegin) {
167 for (; ScanLimit; --IP, --ScanLimit) {
168 // Don't count dbg.value against the ScanLimit, to avoid perturbing the
170 if (isa<DbgInfoIntrinsic>(IP))
172 if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(0) == LHS &&
173 IP->getOperand(1) == RHS)
175 if (IP == BlockBegin) break;
179 // Save the original insertion point so we can restore it when we're done.
180 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
181 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
183 // Move the insertion point out of as many loops as we can.
184 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
185 if (!L->isLoopInvariant(LHS) || !L->isLoopInvariant(RHS)) break;
186 BasicBlock *Preheader = L->getLoopPreheader();
187 if (!Preheader) break;
189 // Ok, move up a level.
190 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
193 // If we haven't found this binop, insert it.
194 Instruction *BO = cast<Instruction>(Builder.CreateBinOp(Opcode, LHS, RHS));
195 BO->setDebugLoc(SaveInsertPt->getDebugLoc());
196 rememberInstruction(BO);
198 // Restore the original insert point.
200 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
205 /// FactorOutConstant - Test if S is divisible by Factor, using signed
206 /// division. If so, update S with Factor divided out and return true.
207 /// S need not be evenly divisible if a reasonable remainder can be
209 /// TODO: When ScalarEvolution gets a SCEVSDivExpr, this can be made
210 /// unnecessary; in its place, just signed-divide Ops[i] by the scale and
211 /// check to see if the divide was folded.
212 static bool FactorOutConstant(const SCEV *&S,
213 const SCEV *&Remainder,
216 const DataLayout *TD) {
217 // Everything is divisible by one.
223 S = SE.getConstant(S->getType(), 1);
227 // For a Constant, check for a multiple of the given factor.
228 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
232 // Check for divisibility.
233 if (const SCEVConstant *FC = dyn_cast<SCEVConstant>(Factor)) {
235 ConstantInt::get(SE.getContext(),
236 C->getValue()->getValue().sdiv(
237 FC->getValue()->getValue()));
238 // If the quotient is zero and the remainder is non-zero, reject
239 // the value at this scale. It will be considered for subsequent
242 const SCEV *Div = SE.getConstant(CI);
245 SE.getAddExpr(Remainder,
246 SE.getConstant(C->getValue()->getValue().srem(
247 FC->getValue()->getValue())));
253 // In a Mul, check if there is a constant operand which is a multiple
254 // of the given factor.
255 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
257 // With DataLayout, the size is known. Check if there is a constant
258 // operand which is a multiple of the given factor. If so, we can
260 const SCEVConstant *FC = cast<SCEVConstant>(Factor);
261 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0)))
262 if (!C->getValue()->getValue().srem(FC->getValue()->getValue())) {
263 SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end());
265 SE.getConstant(C->getValue()->getValue().sdiv(
266 FC->getValue()->getValue()));
267 S = SE.getMulExpr(NewMulOps);
271 // Without DataLayout, check if Factor can be factored out of any of the
272 // Mul's operands. If so, we can just remove it.
273 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) {
274 const SCEV *SOp = M->getOperand(i);
275 const SCEV *Remainder = SE.getConstant(SOp->getType(), 0);
276 if (FactorOutConstant(SOp, Remainder, Factor, SE, TD) &&
277 Remainder->isZero()) {
278 SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end());
280 S = SE.getMulExpr(NewMulOps);
287 // In an AddRec, check if both start and step are divisible.
288 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
289 const SCEV *Step = A->getStepRecurrence(SE);
290 const SCEV *StepRem = SE.getConstant(Step->getType(), 0);
291 if (!FactorOutConstant(Step, StepRem, Factor, SE, TD))
293 if (!StepRem->isZero())
295 const SCEV *Start = A->getStart();
296 if (!FactorOutConstant(Start, Remainder, Factor, SE, TD))
298 S = SE.getAddRecExpr(Start, Step, A->getLoop(),
299 A->getNoWrapFlags(SCEV::FlagNW));
306 /// SimplifyAddOperands - Sort and simplify a list of add operands. NumAddRecs
307 /// is the number of SCEVAddRecExprs present, which are kept at the end of
310 static void SimplifyAddOperands(SmallVectorImpl<const SCEV *> &Ops,
312 ScalarEvolution &SE) {
313 unsigned NumAddRecs = 0;
314 for (unsigned i = Ops.size(); i > 0 && isa<SCEVAddRecExpr>(Ops[i-1]); --i)
316 // Group Ops into non-addrecs and addrecs.
317 SmallVector<const SCEV *, 8> NoAddRecs(Ops.begin(), Ops.end() - NumAddRecs);
318 SmallVector<const SCEV *, 8> AddRecs(Ops.end() - NumAddRecs, Ops.end());
319 // Let ScalarEvolution sort and simplify the non-addrecs list.
320 const SCEV *Sum = NoAddRecs.empty() ?
321 SE.getConstant(Ty, 0) :
322 SE.getAddExpr(NoAddRecs);
323 // If it returned an add, use the operands. Otherwise it simplified
324 // the sum into a single value, so just use that.
326 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Sum))
327 Ops.append(Add->op_begin(), Add->op_end());
328 else if (!Sum->isZero())
330 // Then append the addrecs.
331 Ops.append(AddRecs.begin(), AddRecs.end());
334 /// SplitAddRecs - Flatten a list of add operands, moving addrec start values
335 /// out to the top level. For example, convert {a + b,+,c} to a, b, {0,+,d}.
336 /// This helps expose more opportunities for folding parts of the expressions
337 /// into GEP indices.
339 static void SplitAddRecs(SmallVectorImpl<const SCEV *> &Ops,
341 ScalarEvolution &SE) {
343 SmallVector<const SCEV *, 8> AddRecs;
344 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
345 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Ops[i])) {
346 const SCEV *Start = A->getStart();
347 if (Start->isZero()) break;
348 const SCEV *Zero = SE.getConstant(Ty, 0);
349 AddRecs.push_back(SE.getAddRecExpr(Zero,
350 A->getStepRecurrence(SE),
352 A->getNoWrapFlags(SCEV::FlagNW)));
353 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Start)) {
355 Ops.append(Add->op_begin(), Add->op_end());
356 e += Add->getNumOperands();
361 if (!AddRecs.empty()) {
362 // Add the addrecs onto the end of the list.
363 Ops.append(AddRecs.begin(), AddRecs.end());
364 // Resort the operand list, moving any constants to the front.
365 SimplifyAddOperands(Ops, Ty, SE);
369 /// expandAddToGEP - Expand an addition expression with a pointer type into
370 /// a GEP instead of using ptrtoint+arithmetic+inttoptr. This helps
371 /// BasicAliasAnalysis and other passes analyze the result. See the rules
372 /// for getelementptr vs. inttoptr in
373 /// http://llvm.org/docs/LangRef.html#pointeraliasing
376 /// Design note: The correctness of using getelementptr here depends on
377 /// ScalarEvolution not recognizing inttoptr and ptrtoint operators, as
378 /// they may introduce pointer arithmetic which may not be safely converted
379 /// into getelementptr.
381 /// Design note: It might seem desirable for this function to be more
382 /// loop-aware. If some of the indices are loop-invariant while others
383 /// aren't, it might seem desirable to emit multiple GEPs, keeping the
384 /// loop-invariant portions of the overall computation outside the loop.
385 /// However, there are a few reasons this is not done here. Hoisting simple
386 /// arithmetic is a low-level optimization that often isn't very
387 /// important until late in the optimization process. In fact, passes
388 /// like InstructionCombining will combine GEPs, even if it means
389 /// pushing loop-invariant computation down into loops, so even if the
390 /// GEPs were split here, the work would quickly be undone. The
391 /// LoopStrengthReduction pass, which is usually run quite late (and
392 /// after the last InstructionCombining pass), takes care of hoisting
393 /// loop-invariant portions of expressions, after considering what
394 /// can be folded using target addressing modes.
396 Value *SCEVExpander::expandAddToGEP(const SCEV *const *op_begin,
397 const SCEV *const *op_end,
401 Type *ElTy = PTy->getElementType();
402 SmallVector<Value *, 4> GepIndices;
403 SmallVector<const SCEV *, 8> Ops(op_begin, op_end);
404 bool AnyNonZeroIndices = false;
406 // Split AddRecs up into parts as either of the parts may be usable
407 // without the other.
408 SplitAddRecs(Ops, Ty, SE);
410 // Descend down the pointer's type and attempt to convert the other
411 // operands into GEP indices, at each level. The first index in a GEP
412 // indexes into the array implied by the pointer operand; the rest of
413 // the indices index into the element or field type selected by the
416 // If the scale size is not 0, attempt to factor out a scale for
418 SmallVector<const SCEV *, 8> ScaledOps;
419 if (ElTy->isSized()) {
420 const SCEV *ElSize = SE.getSizeOfExpr(ElTy);
421 if (!ElSize->isZero()) {
422 SmallVector<const SCEV *, 8> NewOps;
423 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
424 const SCEV *Op = Ops[i];
425 const SCEV *Remainder = SE.getConstant(Ty, 0);
426 if (FactorOutConstant(Op, Remainder, ElSize, SE, SE.TD)) {
427 // Op now has ElSize factored out.
428 ScaledOps.push_back(Op);
429 if (!Remainder->isZero())
430 NewOps.push_back(Remainder);
431 AnyNonZeroIndices = true;
433 // The operand was not divisible, so add it to the list of operands
434 // we'll scan next iteration.
435 NewOps.push_back(Ops[i]);
438 // If we made any changes, update Ops.
439 if (!ScaledOps.empty()) {
441 SimplifyAddOperands(Ops, Ty, SE);
446 // Record the scaled array index for this level of the type. If
447 // we didn't find any operands that could be factored, tentatively
448 // assume that element zero was selected (since the zero offset
449 // would obviously be folded away).
450 Value *Scaled = ScaledOps.empty() ?
451 Constant::getNullValue(Ty) :
452 expandCodeFor(SE.getAddExpr(ScaledOps), Ty);
453 GepIndices.push_back(Scaled);
455 // Collect struct field index operands.
456 while (StructType *STy = dyn_cast<StructType>(ElTy)) {
457 bool FoundFieldNo = false;
458 // An empty struct has no fields.
459 if (STy->getNumElements() == 0) break;
461 // With DataLayout, field offsets are known. See if a constant offset
462 // falls within any of the struct fields.
463 if (Ops.empty()) break;
464 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[0]))
465 if (SE.getTypeSizeInBits(C->getType()) <= 64) {
466 const StructLayout &SL = *SE.TD->getStructLayout(STy);
467 uint64_t FullOffset = C->getValue()->getZExtValue();
468 if (FullOffset < SL.getSizeInBytes()) {
469 unsigned ElIdx = SL.getElementContainingOffset(FullOffset);
470 GepIndices.push_back(
471 ConstantInt::get(Type::getInt32Ty(Ty->getContext()), ElIdx));
472 ElTy = STy->getTypeAtIndex(ElIdx);
474 SE.getConstant(Ty, FullOffset - SL.getElementOffset(ElIdx));
475 AnyNonZeroIndices = true;
480 // Without DataLayout, just check for an offsetof expression of the
481 // appropriate struct type.
482 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
483 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Ops[i])) {
486 if (U->isOffsetOf(CTy, FieldNo) && CTy == STy) {
487 GepIndices.push_back(FieldNo);
489 STy->getTypeAtIndex(cast<ConstantInt>(FieldNo)->getZExtValue());
490 Ops[i] = SE.getConstant(Ty, 0);
491 AnyNonZeroIndices = true;
497 // If no struct field offsets were found, tentatively assume that
498 // field zero was selected (since the zero offset would obviously
501 ElTy = STy->getTypeAtIndex(0u);
502 GepIndices.push_back(
503 Constant::getNullValue(Type::getInt32Ty(Ty->getContext())));
507 if (ArrayType *ATy = dyn_cast<ArrayType>(ElTy))
508 ElTy = ATy->getElementType();
513 // If none of the operands were convertible to proper GEP indices, cast
514 // the base to i8* and do an ugly getelementptr with that. It's still
515 // better than ptrtoint+arithmetic+inttoptr at least.
516 if (!AnyNonZeroIndices) {
517 // Cast the base to i8*.
518 V = InsertNoopCastOfTo(V,
519 Type::getInt8PtrTy(Ty->getContext(), PTy->getAddressSpace()));
521 assert(!isa<Instruction>(V) ||
522 SE.DT->dominates(cast<Instruction>(V), Builder.GetInsertPoint()));
524 // Expand the operands for a plain byte offset.
525 Value *Idx = expandCodeFor(SE.getAddExpr(Ops), Ty);
527 // Fold a GEP with constant operands.
528 if (Constant *CLHS = dyn_cast<Constant>(V))
529 if (Constant *CRHS = dyn_cast<Constant>(Idx))
530 return ConstantExpr::getGetElementPtr(CLHS, CRHS);
532 // Do a quick scan to see if we have this GEP nearby. If so, reuse it.
533 unsigned ScanLimit = 6;
534 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
535 // Scanning starts from the last instruction before the insertion point.
536 BasicBlock::iterator IP = Builder.GetInsertPoint();
537 if (IP != BlockBegin) {
539 for (; ScanLimit; --IP, --ScanLimit) {
540 // Don't count dbg.value against the ScanLimit, to avoid perturbing the
542 if (isa<DbgInfoIntrinsic>(IP))
544 if (IP->getOpcode() == Instruction::GetElementPtr &&
545 IP->getOperand(0) == V && IP->getOperand(1) == Idx)
547 if (IP == BlockBegin) break;
551 // Save the original insertion point so we can restore it when we're done.
552 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
553 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
555 // Move the insertion point out of as many loops as we can.
556 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
557 if (!L->isLoopInvariant(V) || !L->isLoopInvariant(Idx)) break;
558 BasicBlock *Preheader = L->getLoopPreheader();
559 if (!Preheader) break;
561 // Ok, move up a level.
562 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
566 Value *GEP = Builder.CreateGEP(V, Idx, "uglygep");
567 rememberInstruction(GEP);
569 // Restore the original insert point.
571 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
576 // Save the original insertion point so we can restore it when we're done.
577 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
578 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
580 // Move the insertion point out of as many loops as we can.
581 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
582 if (!L->isLoopInvariant(V)) break;
584 bool AnyIndexNotLoopInvariant = false;
585 for (SmallVectorImpl<Value *>::const_iterator I = GepIndices.begin(),
586 E = GepIndices.end(); I != E; ++I)
587 if (!L->isLoopInvariant(*I)) {
588 AnyIndexNotLoopInvariant = true;
591 if (AnyIndexNotLoopInvariant)
594 BasicBlock *Preheader = L->getLoopPreheader();
595 if (!Preheader) break;
597 // Ok, move up a level.
598 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
601 // Insert a pretty getelementptr. Note that this GEP is not marked inbounds,
602 // because ScalarEvolution may have changed the address arithmetic to
603 // compute a value which is beyond the end of the allocated object.
605 if (V->getType() != PTy)
606 Casted = InsertNoopCastOfTo(Casted, PTy);
607 Value *GEP = Builder.CreateGEP(Casted,
610 Ops.push_back(SE.getUnknown(GEP));
611 rememberInstruction(GEP);
613 // Restore the original insert point.
615 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
617 return expand(SE.getAddExpr(Ops));
620 /// PickMostRelevantLoop - Given two loops pick the one that's most relevant for
621 /// SCEV expansion. If they are nested, this is the most nested. If they are
622 /// neighboring, pick the later.
623 static const Loop *PickMostRelevantLoop(const Loop *A, const Loop *B,
627 if (A->contains(B)) return B;
628 if (B->contains(A)) return A;
629 if (DT.dominates(A->getHeader(), B->getHeader())) return B;
630 if (DT.dominates(B->getHeader(), A->getHeader())) return A;
631 return A; // Arbitrarily break the tie.
634 /// getRelevantLoop - Get the most relevant loop associated with the given
635 /// expression, according to PickMostRelevantLoop.
636 const Loop *SCEVExpander::getRelevantLoop(const SCEV *S) {
637 // Test whether we've already computed the most relevant loop for this SCEV.
638 std::pair<DenseMap<const SCEV *, const Loop *>::iterator, bool> Pair =
639 RelevantLoops.insert(std::make_pair(S, static_cast<const Loop *>(0)));
641 return Pair.first->second;
643 if (isa<SCEVConstant>(S))
644 // A constant has no relevant loops.
646 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
647 if (const Instruction *I = dyn_cast<Instruction>(U->getValue()))
648 return Pair.first->second = SE.LI->getLoopFor(I->getParent());
649 // A non-instruction has no relevant loops.
652 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S)) {
654 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
656 for (SCEVNAryExpr::op_iterator I = N->op_begin(), E = N->op_end();
658 L = PickMostRelevantLoop(L, getRelevantLoop(*I), *SE.DT);
659 return RelevantLoops[N] = L;
661 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S)) {
662 const Loop *Result = getRelevantLoop(C->getOperand());
663 return RelevantLoops[C] = Result;
665 if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
667 PickMostRelevantLoop(getRelevantLoop(D->getLHS()),
668 getRelevantLoop(D->getRHS()),
670 return RelevantLoops[D] = Result;
672 llvm_unreachable("Unexpected SCEV type!");
677 /// LoopCompare - Compare loops by PickMostRelevantLoop.
681 explicit LoopCompare(DominatorTree &dt) : DT(dt) {}
683 bool operator()(std::pair<const Loop *, const SCEV *> LHS,
684 std::pair<const Loop *, const SCEV *> RHS) const {
685 // Keep pointer operands sorted at the end.
686 if (LHS.second->getType()->isPointerTy() !=
687 RHS.second->getType()->isPointerTy())
688 return LHS.second->getType()->isPointerTy();
690 // Compare loops with PickMostRelevantLoop.
691 if (LHS.first != RHS.first)
692 return PickMostRelevantLoop(LHS.first, RHS.first, DT) != LHS.first;
694 // If one operand is a non-constant negative and the other is not,
695 // put the non-constant negative on the right so that a sub can
696 // be used instead of a negate and add.
697 if (LHS.second->isNonConstantNegative()) {
698 if (!RHS.second->isNonConstantNegative())
700 } else if (RHS.second->isNonConstantNegative())
703 // Otherwise they are equivalent according to this comparison.
710 Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) {
711 Type *Ty = SE.getEffectiveSCEVType(S->getType());
713 // Collect all the add operands in a loop, along with their associated loops.
714 // Iterate in reverse so that constants are emitted last, all else equal, and
715 // so that pointer operands are inserted first, which the code below relies on
716 // to form more involved GEPs.
717 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops;
718 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(S->op_end()),
719 E(S->op_begin()); I != E; ++I)
720 OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I));
722 // Sort by loop. Use a stable sort so that constants follow non-constants and
723 // pointer operands precede non-pointer operands.
724 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT));
726 // Emit instructions to add all the operands. Hoist as much as possible
727 // out of loops, and form meaningful getelementptrs where possible.
729 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator
730 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) {
731 const Loop *CurLoop = I->first;
732 const SCEV *Op = I->second;
734 // This is the first operand. Just expand it.
737 } else if (PointerType *PTy = dyn_cast<PointerType>(Sum->getType())) {
738 // The running sum expression is a pointer. Try to form a getelementptr
739 // at this level with that as the base.
740 SmallVector<const SCEV *, 4> NewOps;
741 for (; I != E && I->first == CurLoop; ++I) {
742 // If the operand is SCEVUnknown and not instructions, peek through
743 // it, to enable more of it to be folded into the GEP.
744 const SCEV *X = I->second;
745 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(X))
746 if (!isa<Instruction>(U->getValue()))
747 X = SE.getSCEV(U->getValue());
750 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, Sum);
751 } else if (PointerType *PTy = dyn_cast<PointerType>(Op->getType())) {
752 // The running sum is an integer, and there's a pointer at this level.
753 // Try to form a getelementptr. If the running sum is instructions,
754 // use a SCEVUnknown to avoid re-analyzing them.
755 SmallVector<const SCEV *, 4> NewOps;
756 NewOps.push_back(isa<Instruction>(Sum) ? SE.getUnknown(Sum) :
758 for (++I; I != E && I->first == CurLoop; ++I)
759 NewOps.push_back(I->second);
760 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, expand(Op));
761 } else if (Op->isNonConstantNegative()) {
762 // Instead of doing a negate and add, just do a subtract.
763 Value *W = expandCodeFor(SE.getNegativeSCEV(Op), Ty);
764 Sum = InsertNoopCastOfTo(Sum, Ty);
765 Sum = InsertBinop(Instruction::Sub, Sum, W);
769 Value *W = expandCodeFor(Op, Ty);
770 Sum = InsertNoopCastOfTo(Sum, Ty);
771 // Canonicalize a constant to the RHS.
772 if (isa<Constant>(Sum)) std::swap(Sum, W);
773 Sum = InsertBinop(Instruction::Add, Sum, W);
781 Value *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) {
782 Type *Ty = SE.getEffectiveSCEVType(S->getType());
784 // Collect all the mul operands in a loop, along with their associated loops.
785 // Iterate in reverse so that constants are emitted last, all else equal.
786 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops;
787 for (std::reverse_iterator<SCEVMulExpr::op_iterator> I(S->op_end()),
788 E(S->op_begin()); I != E; ++I)
789 OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I));
791 // Sort by loop. Use a stable sort so that constants follow non-constants.
792 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT));
794 // Emit instructions to mul all the operands. Hoist as much as possible
797 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator
798 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) {
799 const SCEV *Op = I->second;
801 // This is the first operand. Just expand it.
804 } else if (Op->isAllOnesValue()) {
805 // Instead of doing a multiply by negative one, just do a negate.
806 Prod = InsertNoopCastOfTo(Prod, Ty);
807 Prod = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), Prod);
811 Value *W = expandCodeFor(Op, Ty);
812 Prod = InsertNoopCastOfTo(Prod, Ty);
813 // Canonicalize a constant to the RHS.
814 if (isa<Constant>(Prod)) std::swap(Prod, W);
815 Prod = InsertBinop(Instruction::Mul, Prod, W);
823 Value *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) {
824 Type *Ty = SE.getEffectiveSCEVType(S->getType());
826 Value *LHS = expandCodeFor(S->getLHS(), Ty);
827 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getRHS())) {
828 const APInt &RHS = SC->getValue()->getValue();
829 if (RHS.isPowerOf2())
830 return InsertBinop(Instruction::LShr, LHS,
831 ConstantInt::get(Ty, RHS.logBase2()));
834 Value *RHS = expandCodeFor(S->getRHS(), Ty);
835 return InsertBinop(Instruction::UDiv, LHS, RHS);
838 /// Move parts of Base into Rest to leave Base with the minimal
839 /// expression that provides a pointer operand suitable for a
841 static void ExposePointerBase(const SCEV *&Base, const SCEV *&Rest,
842 ScalarEvolution &SE) {
843 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Base)) {
844 Base = A->getStart();
845 Rest = SE.getAddExpr(Rest,
846 SE.getAddRecExpr(SE.getConstant(A->getType(), 0),
847 A->getStepRecurrence(SE),
849 A->getNoWrapFlags(SCEV::FlagNW)));
851 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(Base)) {
852 Base = A->getOperand(A->getNumOperands()-1);
853 SmallVector<const SCEV *, 8> NewAddOps(A->op_begin(), A->op_end());
854 NewAddOps.back() = Rest;
855 Rest = SE.getAddExpr(NewAddOps);
856 ExposePointerBase(Base, Rest, SE);
860 /// Determine if this is a well-behaved chain of instructions leading back to
861 /// the PHI. If so, it may be reused by expanded expressions.
862 bool SCEVExpander::isNormalAddRecExprPHI(PHINode *PN, Instruction *IncV,
864 if (IncV->getNumOperands() == 0 || isa<PHINode>(IncV) ||
865 (isa<CastInst>(IncV) && !isa<BitCastInst>(IncV)))
867 // If any of the operands don't dominate the insert position, bail.
868 // Addrec operands are always loop-invariant, so this can only happen
869 // if there are instructions which haven't been hoisted.
870 if (L == IVIncInsertLoop) {
871 for (User::op_iterator OI = IncV->op_begin()+1,
872 OE = IncV->op_end(); OI != OE; ++OI)
873 if (Instruction *OInst = dyn_cast<Instruction>(OI))
874 if (!SE.DT->dominates(OInst, IVIncInsertPos))
877 // Advance to the next instruction.
878 IncV = dyn_cast<Instruction>(IncV->getOperand(0));
882 if (IncV->mayHaveSideEffects())
888 return isNormalAddRecExprPHI(PN, IncV, L);
891 /// getIVIncOperand returns an induction variable increment's induction
892 /// variable operand.
894 /// If allowScale is set, any type of GEP is allowed as long as the nonIV
895 /// operands dominate InsertPos.
897 /// If allowScale is not set, ensure that a GEP increment conforms to one of the
898 /// simple patterns generated by getAddRecExprPHILiterally and
899 /// expandAddtoGEP. If the pattern isn't recognized, return NULL.
900 Instruction *SCEVExpander::getIVIncOperand(Instruction *IncV,
901 Instruction *InsertPos,
903 if (IncV == InsertPos)
906 switch (IncV->getOpcode()) {
909 // Check for a simple Add/Sub or GEP of a loop invariant step.
910 case Instruction::Add:
911 case Instruction::Sub: {
912 Instruction *OInst = dyn_cast<Instruction>(IncV->getOperand(1));
913 if (!OInst || SE.DT->dominates(OInst, InsertPos))
914 return dyn_cast<Instruction>(IncV->getOperand(0));
917 case Instruction::BitCast:
918 return dyn_cast<Instruction>(IncV->getOperand(0));
919 case Instruction::GetElementPtr:
920 for (Instruction::op_iterator I = IncV->op_begin()+1, E = IncV->op_end();
922 if (isa<Constant>(*I))
924 if (Instruction *OInst = dyn_cast<Instruction>(*I)) {
925 if (!SE.DT->dominates(OInst, InsertPos))
929 // allow any kind of GEP as long as it can be hoisted.
932 // This must be a pointer addition of constants (pretty), which is already
933 // handled, or some number of address-size elements (ugly). Ugly geps
934 // have 2 operands. i1* is used by the expander to represent an
935 // address-size element.
936 if (IncV->getNumOperands() != 2)
938 unsigned AS = cast<PointerType>(IncV->getType())->getAddressSpace();
939 if (IncV->getType() != Type::getInt1PtrTy(SE.getContext(), AS)
940 && IncV->getType() != Type::getInt8PtrTy(SE.getContext(), AS))
944 return dyn_cast<Instruction>(IncV->getOperand(0));
948 /// hoistStep - Attempt to hoist a simple IV increment above InsertPos to make
949 /// it available to other uses in this loop. Recursively hoist any operands,
950 /// until we reach a value that dominates InsertPos.
951 bool SCEVExpander::hoistIVInc(Instruction *IncV, Instruction *InsertPos) {
952 if (SE.DT->dominates(IncV, InsertPos))
955 // InsertPos must itself dominate IncV so that IncV's new position satisfies
956 // its existing users.
957 if (isa<PHINode>(InsertPos)
958 || !SE.DT->dominates(InsertPos->getParent(), IncV->getParent()))
961 // Check that the chain of IV operands leading back to Phi can be hoisted.
962 SmallVector<Instruction*, 4> IVIncs;
964 Instruction *Oper = getIVIncOperand(IncV, InsertPos, /*allowScale*/true);
967 // IncV is safe to hoist.
968 IVIncs.push_back(IncV);
970 if (SE.DT->dominates(IncV, InsertPos))
973 for (SmallVectorImpl<Instruction*>::reverse_iterator I = IVIncs.rbegin(),
974 E = IVIncs.rend(); I != E; ++I) {
975 (*I)->moveBefore(InsertPos);
980 /// Determine if this cyclic phi is in a form that would have been generated by
981 /// LSR. We don't care if the phi was actually expanded in this pass, as long
982 /// as it is in a low-cost form, for example, no implied multiplication. This
983 /// should match any patterns generated by getAddRecExprPHILiterally and
985 bool SCEVExpander::isExpandedAddRecExprPHI(PHINode *PN, Instruction *IncV,
987 for(Instruction *IVOper = IncV;
988 (IVOper = getIVIncOperand(IVOper, L->getLoopPreheader()->getTerminator(),
989 /*allowScale=*/false));) {
996 /// expandIVInc - Expand an IV increment at Builder's current InsertPos.
997 /// Typically this is the LatchBlock terminator or IVIncInsertPos, but we may
998 /// need to materialize IV increments elsewhere to handle difficult situations.
999 Value *SCEVExpander::expandIVInc(PHINode *PN, Value *StepV, const Loop *L,
1000 Type *ExpandTy, Type *IntTy,
1003 // If the PHI is a pointer, use a GEP, otherwise use an add or sub.
1004 if (ExpandTy->isPointerTy()) {
1005 PointerType *GEPPtrTy = cast<PointerType>(ExpandTy);
1006 // If the step isn't constant, don't use an implicitly scaled GEP, because
1007 // that would require a multiply inside the loop.
1008 if (!isa<ConstantInt>(StepV))
1009 GEPPtrTy = PointerType::get(Type::getInt1Ty(SE.getContext()),
1010 GEPPtrTy->getAddressSpace());
1011 const SCEV *const StepArray[1] = { SE.getSCEV(StepV) };
1012 IncV = expandAddToGEP(StepArray, StepArray+1, GEPPtrTy, IntTy, PN);
1013 if (IncV->getType() != PN->getType()) {
1014 IncV = Builder.CreateBitCast(IncV, PN->getType());
1015 rememberInstruction(IncV);
1018 IncV = useSubtract ?
1019 Builder.CreateSub(PN, StepV, Twine(IVName) + ".iv.next") :
1020 Builder.CreateAdd(PN, StepV, Twine(IVName) + ".iv.next");
1021 rememberInstruction(IncV);
1026 /// getAddRecExprPHILiterally - Helper for expandAddRecExprLiterally. Expand
1027 /// the base addrec, which is the addrec without any non-loop-dominating
1028 /// values, and return the PHI.
1030 SCEVExpander::getAddRecExprPHILiterally(const SCEVAddRecExpr *Normalized,
1034 assert((!IVIncInsertLoop||IVIncInsertPos) && "Uninitialized insert position");
1036 // Reuse a previously-inserted PHI, if present.
1037 BasicBlock *LatchBlock = L->getLoopLatch();
1039 for (BasicBlock::iterator I = L->getHeader()->begin();
1040 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1041 if (!SE.isSCEVable(PN->getType()) ||
1042 (SE.getEffectiveSCEVType(PN->getType()) !=
1043 SE.getEffectiveSCEVType(Normalized->getType())) ||
1044 SE.getSCEV(PN) != Normalized)
1048 cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock));
1051 if (!isExpandedAddRecExprPHI(PN, IncV, L))
1053 if (L == IVIncInsertLoop && !hoistIVInc(IncV, IVIncInsertPos))
1057 if (!isNormalAddRecExprPHI(PN, IncV, L))
1059 if (L == IVIncInsertLoop)
1061 if (SE.DT->dominates(IncV, IVIncInsertPos))
1063 // Make sure the increment is where we want it. But don't move it
1064 // down past a potential existing post-inc user.
1065 IncV->moveBefore(IVIncInsertPos);
1066 IVIncInsertPos = IncV;
1067 IncV = cast<Instruction>(IncV->getOperand(0));
1068 } while (IncV != PN);
1070 // Ok, the add recurrence looks usable.
1071 // Remember this PHI, even in post-inc mode.
1072 InsertedValues.insert(PN);
1073 // Remember the increment.
1074 rememberInstruction(IncV);
1079 // Save the original insertion point so we can restore it when we're done.
1080 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1081 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1083 // Another AddRec may need to be recursively expanded below. For example, if
1084 // this AddRec is quadratic, the StepV may itself be an AddRec in this
1085 // loop. Remove this loop from the PostIncLoops set before expanding such
1086 // AddRecs. Otherwise, we cannot find a valid position for the step
1087 // (i.e. StepV can never dominate its loop header). Ideally, we could do
1088 // SavedIncLoops.swap(PostIncLoops), but we generally have a single element,
1089 // so it's not worth implementing SmallPtrSet::swap.
1090 PostIncLoopSet SavedPostIncLoops = PostIncLoops;
1091 PostIncLoops.clear();
1093 // Expand code for the start value.
1094 Value *StartV = expandCodeFor(Normalized->getStart(), ExpandTy,
1095 L->getHeader()->begin());
1097 // StartV must be hoisted into L's preheader to dominate the new phi.
1098 assert(!isa<Instruction>(StartV) ||
1099 SE.DT->properlyDominates(cast<Instruction>(StartV)->getParent(),
1102 // Expand code for the step value. Do this before creating the PHI so that PHI
1103 // reuse code doesn't see an incomplete PHI.
1104 const SCEV *Step = Normalized->getStepRecurrence(SE);
1105 // If the stride is negative, insert a sub instead of an add for the increment
1106 // (unless it's a constant, because subtracts of constants are canonicalized
1108 bool useSubtract = !ExpandTy->isPointerTy() && Step->isNonConstantNegative();
1110 Step = SE.getNegativeSCEV(Step);
1111 // Expand the step somewhere that dominates the loop header.
1112 Value *StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin());
1115 BasicBlock *Header = L->getHeader();
1116 Builder.SetInsertPoint(Header, Header->begin());
1117 pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header);
1118 PHINode *PN = Builder.CreatePHI(ExpandTy, std::distance(HPB, HPE),
1119 Twine(IVName) + ".iv");
1120 rememberInstruction(PN);
1122 // Create the step instructions and populate the PHI.
1123 for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) {
1124 BasicBlock *Pred = *HPI;
1126 // Add a start value.
1127 if (!L->contains(Pred)) {
1128 PN->addIncoming(StartV, Pred);
1132 // Create a step value and add it to the PHI.
1133 // If IVIncInsertLoop is non-null and equal to the addrec's loop, insert the
1134 // instructions at IVIncInsertPos.
1135 Instruction *InsertPos = L == IVIncInsertLoop ?
1136 IVIncInsertPos : Pred->getTerminator();
1137 Builder.SetInsertPoint(InsertPos);
1138 Value *IncV = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract);
1139 if (isa<OverflowingBinaryOperator>(IncV)) {
1140 if (Normalized->getNoWrapFlags(SCEV::FlagNUW))
1141 cast<BinaryOperator>(IncV)->setHasNoUnsignedWrap();
1142 if (Normalized->getNoWrapFlags(SCEV::FlagNSW))
1143 cast<BinaryOperator>(IncV)->setHasNoSignedWrap();
1145 PN->addIncoming(IncV, Pred);
1148 // Restore the original insert point.
1150 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
1152 // After expanding subexpressions, restore the PostIncLoops set so the caller
1153 // can ensure that IVIncrement dominates the current uses.
1154 PostIncLoops = SavedPostIncLoops;
1156 // Remember this PHI, even in post-inc mode.
1157 InsertedValues.insert(PN);
1162 Value *SCEVExpander::expandAddRecExprLiterally(const SCEVAddRecExpr *S) {
1163 Type *STy = S->getType();
1164 Type *IntTy = SE.getEffectiveSCEVType(STy);
1165 const Loop *L = S->getLoop();
1167 // Determine a normalized form of this expression, which is the expression
1168 // before any post-inc adjustment is made.
1169 const SCEVAddRecExpr *Normalized = S;
1170 if (PostIncLoops.count(L)) {
1171 PostIncLoopSet Loops;
1174 cast<SCEVAddRecExpr>(TransformForPostIncUse(Normalize, S, 0, 0,
1175 Loops, SE, *SE.DT));
1178 // Strip off any non-loop-dominating component from the addrec start.
1179 const SCEV *Start = Normalized->getStart();
1180 const SCEV *PostLoopOffset = 0;
1181 if (!SE.properlyDominates(Start, L->getHeader())) {
1182 PostLoopOffset = Start;
1183 Start = SE.getConstant(Normalized->getType(), 0);
1184 Normalized = cast<SCEVAddRecExpr>(
1185 SE.getAddRecExpr(Start, Normalized->getStepRecurrence(SE),
1186 Normalized->getLoop(),
1187 Normalized->getNoWrapFlags(SCEV::FlagNW)));
1190 // Strip off any non-loop-dominating component from the addrec step.
1191 const SCEV *Step = Normalized->getStepRecurrence(SE);
1192 const SCEV *PostLoopScale = 0;
1193 if (!SE.dominates(Step, L->getHeader())) {
1194 PostLoopScale = Step;
1195 Step = SE.getConstant(Normalized->getType(), 1);
1197 cast<SCEVAddRecExpr>(SE.getAddRecExpr(
1198 Start, Step, Normalized->getLoop(),
1199 Normalized->getNoWrapFlags(SCEV::FlagNW)));
1202 // Expand the core addrec. If we need post-loop scaling, force it to
1203 // expand to an integer type to avoid the need for additional casting.
1204 Type *ExpandTy = PostLoopScale ? IntTy : STy;
1205 PHINode *PN = getAddRecExprPHILiterally(Normalized, L, ExpandTy, IntTy);
1207 // Accommodate post-inc mode, if necessary.
1209 if (!PostIncLoops.count(L))
1212 // In PostInc mode, use the post-incremented value.
1213 BasicBlock *LatchBlock = L->getLoopLatch();
1214 assert(LatchBlock && "PostInc mode requires a unique loop latch!");
1215 Result = PN->getIncomingValueForBlock(LatchBlock);
1217 // For an expansion to use the postinc form, the client must call
1218 // expandCodeFor with an InsertPoint that is either outside the PostIncLoop
1219 // or dominated by IVIncInsertPos.
1220 if (isa<Instruction>(Result)
1221 && !SE.DT->dominates(cast<Instruction>(Result),
1222 Builder.GetInsertPoint())) {
1223 // The induction variable's postinc expansion does not dominate this use.
1224 // IVUsers tries to prevent this case, so it is rare. However, it can
1225 // happen when an IVUser outside the loop is not dominated by the latch
1226 // block. Adjusting IVIncInsertPos before expansion begins cannot handle
1227 // all cases. Consider a phi outide whose operand is replaced during
1228 // expansion with the value of the postinc user. Without fundamentally
1229 // changing the way postinc users are tracked, the only remedy is
1230 // inserting an extra IV increment. StepV might fold into PostLoopOffset,
1231 // but hopefully expandCodeFor handles that.
1233 !ExpandTy->isPointerTy() && Step->isNonConstantNegative();
1235 Step = SE.getNegativeSCEV(Step);
1236 // Expand the step somewhere that dominates the loop header.
1237 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1238 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1239 Value *StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin());
1240 // Restore the insertion point to the place where the caller has
1241 // determined dominates all uses.
1242 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
1243 Result = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract);
1247 // Re-apply any non-loop-dominating scale.
1248 if (PostLoopScale) {
1249 Result = InsertNoopCastOfTo(Result, IntTy);
1250 Result = Builder.CreateMul(Result,
1251 expandCodeFor(PostLoopScale, IntTy));
1252 rememberInstruction(Result);
1255 // Re-apply any non-loop-dominating offset.
1256 if (PostLoopOffset) {
1257 if (PointerType *PTy = dyn_cast<PointerType>(ExpandTy)) {
1258 const SCEV *const OffsetArray[1] = { PostLoopOffset };
1259 Result = expandAddToGEP(OffsetArray, OffsetArray+1, PTy, IntTy, Result);
1261 Result = InsertNoopCastOfTo(Result, IntTy);
1262 Result = Builder.CreateAdd(Result,
1263 expandCodeFor(PostLoopOffset, IntTy));
1264 rememberInstruction(Result);
1271 Value *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) {
1272 if (!CanonicalMode) return expandAddRecExprLiterally(S);
1274 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1275 const Loop *L = S->getLoop();
1277 // First check for an existing canonical IV in a suitable type.
1278 PHINode *CanonicalIV = 0;
1279 if (PHINode *PN = L->getCanonicalInductionVariable())
1280 if (SE.getTypeSizeInBits(PN->getType()) >= SE.getTypeSizeInBits(Ty))
1283 // Rewrite an AddRec in terms of the canonical induction variable, if
1284 // its type is more narrow.
1286 SE.getTypeSizeInBits(CanonicalIV->getType()) >
1287 SE.getTypeSizeInBits(Ty)) {
1288 SmallVector<const SCEV *, 4> NewOps(S->getNumOperands());
1289 for (unsigned i = 0, e = S->getNumOperands(); i != e; ++i)
1290 NewOps[i] = SE.getAnyExtendExpr(S->op_begin()[i], CanonicalIV->getType());
1291 Value *V = expand(SE.getAddRecExpr(NewOps, S->getLoop(),
1292 S->getNoWrapFlags(SCEV::FlagNW)));
1293 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1294 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1295 BasicBlock::iterator NewInsertPt =
1296 llvm::next(BasicBlock::iterator(cast<Instruction>(V)));
1297 while (isa<PHINode>(NewInsertPt) || isa<DbgInfoIntrinsic>(NewInsertPt) ||
1298 isa<LandingPadInst>(NewInsertPt))
1300 V = expandCodeFor(SE.getTruncateExpr(SE.getUnknown(V), Ty), 0,
1302 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
1306 // {X,+,F} --> X + {0,+,F}
1307 if (!S->getStart()->isZero()) {
1308 SmallVector<const SCEV *, 4> NewOps(S->op_begin(), S->op_end());
1309 NewOps[0] = SE.getConstant(Ty, 0);
1310 const SCEV *Rest = SE.getAddRecExpr(NewOps, L,
1311 S->getNoWrapFlags(SCEV::FlagNW));
1313 // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the
1314 // comments on expandAddToGEP for details.
1315 const SCEV *Base = S->getStart();
1316 const SCEV *RestArray[1] = { Rest };
1317 // Dig into the expression to find the pointer base for a GEP.
1318 ExposePointerBase(Base, RestArray[0], SE);
1319 // If we found a pointer, expand the AddRec with a GEP.
1320 if (PointerType *PTy = dyn_cast<PointerType>(Base->getType())) {
1321 // Make sure the Base isn't something exotic, such as a multiplied
1322 // or divided pointer value. In those cases, the result type isn't
1323 // actually a pointer type.
1324 if (!isa<SCEVMulExpr>(Base) && !isa<SCEVUDivExpr>(Base)) {
1325 Value *StartV = expand(Base);
1326 assert(StartV->getType() == PTy && "Pointer type mismatch for GEP!");
1327 return expandAddToGEP(RestArray, RestArray+1, PTy, Ty, StartV);
1331 // Just do a normal add. Pre-expand the operands to suppress folding.
1332 return expand(SE.getAddExpr(SE.getUnknown(expand(S->getStart())),
1333 SE.getUnknown(expand(Rest))));
1336 // If we don't yet have a canonical IV, create one.
1338 // Create and insert the PHI node for the induction variable in the
1340 BasicBlock *Header = L->getHeader();
1341 pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header);
1342 CanonicalIV = PHINode::Create(Ty, std::distance(HPB, HPE), "indvar",
1344 rememberInstruction(CanonicalIV);
1346 SmallSet<BasicBlock *, 4> PredSeen;
1347 Constant *One = ConstantInt::get(Ty, 1);
1348 for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) {
1349 BasicBlock *HP = *HPI;
1350 if (!PredSeen.insert(HP))
1353 if (L->contains(HP)) {
1354 // Insert a unit add instruction right before the terminator
1355 // corresponding to the back-edge.
1356 Instruction *Add = BinaryOperator::CreateAdd(CanonicalIV, One,
1358 HP->getTerminator());
1359 Add->setDebugLoc(HP->getTerminator()->getDebugLoc());
1360 rememberInstruction(Add);
1361 CanonicalIV->addIncoming(Add, HP);
1363 CanonicalIV->addIncoming(Constant::getNullValue(Ty), HP);
1368 // {0,+,1} --> Insert a canonical induction variable into the loop!
1369 if (S->isAffine() && S->getOperand(1)->isOne()) {
1370 assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) &&
1371 "IVs with types different from the canonical IV should "
1372 "already have been handled!");
1376 // {0,+,F} --> {0,+,1} * F
1378 // If this is a simple linear addrec, emit it now as a special case.
1379 if (S->isAffine()) // {0,+,F} --> i*F
1381 expand(SE.getTruncateOrNoop(
1382 SE.getMulExpr(SE.getUnknown(CanonicalIV),
1383 SE.getNoopOrAnyExtend(S->getOperand(1),
1384 CanonicalIV->getType())),
1387 // If this is a chain of recurrences, turn it into a closed form, using the
1388 // folders, then expandCodeFor the closed form. This allows the folders to
1389 // simplify the expression without having to build a bunch of special code
1390 // into this folder.
1391 const SCEV *IH = SE.getUnknown(CanonicalIV); // Get I as a "symbolic" SCEV.
1393 // Promote S up to the canonical IV type, if the cast is foldable.
1394 const SCEV *NewS = S;
1395 const SCEV *Ext = SE.getNoopOrAnyExtend(S, CanonicalIV->getType());
1396 if (isa<SCEVAddRecExpr>(Ext))
1399 const SCEV *V = cast<SCEVAddRecExpr>(NewS)->evaluateAtIteration(IH, SE);
1400 //cerr << "Evaluated: " << *this << "\n to: " << *V << "\n";
1402 // Truncate the result down to the original type, if needed.
1403 const SCEV *T = SE.getTruncateOrNoop(V, Ty);
1407 Value *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) {
1408 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1409 Value *V = expandCodeFor(S->getOperand(),
1410 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1411 Value *I = Builder.CreateTrunc(V, Ty);
1412 rememberInstruction(I);
1416 Value *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) {
1417 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1418 Value *V = expandCodeFor(S->getOperand(),
1419 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1420 Value *I = Builder.CreateZExt(V, Ty);
1421 rememberInstruction(I);
1425 Value *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) {
1426 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1427 Value *V = expandCodeFor(S->getOperand(),
1428 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1429 Value *I = Builder.CreateSExt(V, Ty);
1430 rememberInstruction(I);
1434 Value *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) {
1435 Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
1436 Type *Ty = LHS->getType();
1437 for (int i = S->getNumOperands()-2; i >= 0; --i) {
1438 // In the case of mixed integer and pointer types, do the
1439 // rest of the comparisons as integer.
1440 if (S->getOperand(i)->getType() != Ty) {
1441 Ty = SE.getEffectiveSCEVType(Ty);
1442 LHS = InsertNoopCastOfTo(LHS, Ty);
1444 Value *RHS = expandCodeFor(S->getOperand(i), Ty);
1445 Value *ICmp = Builder.CreateICmpSGT(LHS, RHS);
1446 rememberInstruction(ICmp);
1447 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "smax");
1448 rememberInstruction(Sel);
1451 // In the case of mixed integer and pointer types, cast the
1452 // final result back to the pointer type.
1453 if (LHS->getType() != S->getType())
1454 LHS = InsertNoopCastOfTo(LHS, S->getType());
1458 Value *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) {
1459 Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
1460 Type *Ty = LHS->getType();
1461 for (int i = S->getNumOperands()-2; i >= 0; --i) {
1462 // In the case of mixed integer and pointer types, do the
1463 // rest of the comparisons as integer.
1464 if (S->getOperand(i)->getType() != Ty) {
1465 Ty = SE.getEffectiveSCEVType(Ty);
1466 LHS = InsertNoopCastOfTo(LHS, Ty);
1468 Value *RHS = expandCodeFor(S->getOperand(i), Ty);
1469 Value *ICmp = Builder.CreateICmpUGT(LHS, RHS);
1470 rememberInstruction(ICmp);
1471 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "umax");
1472 rememberInstruction(Sel);
1475 // In the case of mixed integer and pointer types, cast the
1476 // final result back to the pointer type.
1477 if (LHS->getType() != S->getType())
1478 LHS = InsertNoopCastOfTo(LHS, S->getType());
1482 Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty,
1484 Builder.SetInsertPoint(IP->getParent(), IP);
1485 return expandCodeFor(SH, Ty);
1488 Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty) {
1489 // Expand the code for this SCEV.
1490 Value *V = expand(SH);
1492 assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) &&
1493 "non-trivial casts should be done with the SCEVs directly!");
1494 V = InsertNoopCastOfTo(V, Ty);
1499 Value *SCEVExpander::expand(const SCEV *S) {
1500 // Compute an insertion point for this SCEV object. Hoist the instructions
1501 // as far out in the loop nest as possible.
1502 Instruction *InsertPt = Builder.GetInsertPoint();
1503 for (Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock()); ;
1504 L = L->getParentLoop())
1505 if (SE.isLoopInvariant(S, L)) {
1507 if (BasicBlock *Preheader = L->getLoopPreheader())
1508 InsertPt = Preheader->getTerminator();
1510 // LSR sets the insertion point for AddRec start/step values to the
1511 // block start to simplify value reuse, even though it's an invalid
1512 // position. SCEVExpander must correct for this in all cases.
1513 InsertPt = L->getHeader()->getFirstInsertionPt();
1516 // If the SCEV is computable at this level, insert it into the header
1517 // after the PHIs (and after any other instructions that we've inserted
1518 // there) so that it is guaranteed to dominate any user inside the loop.
1519 if (L && SE.hasComputableLoopEvolution(S, L) && !PostIncLoops.count(L))
1520 InsertPt = L->getHeader()->getFirstInsertionPt();
1521 while (InsertPt != Builder.GetInsertPoint()
1522 && (isInsertedInstruction(InsertPt)
1523 || isa<DbgInfoIntrinsic>(InsertPt))) {
1524 InsertPt = llvm::next(BasicBlock::iterator(InsertPt));
1529 // Check to see if we already expanded this here.
1530 std::map<std::pair<const SCEV *, Instruction *>, TrackingVH<Value> >::iterator
1531 I = InsertedExpressions.find(std::make_pair(S, InsertPt));
1532 if (I != InsertedExpressions.end())
1535 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1536 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1537 Builder.SetInsertPoint(InsertPt->getParent(), InsertPt);
1539 // Expand the expression into instructions.
1540 Value *V = visit(S);
1542 // Remember the expanded value for this SCEV at this location.
1544 // This is independent of PostIncLoops. The mapped value simply materializes
1545 // the expression at this insertion point. If the mapped value happened to be
1546 // a postinc expansion, it could be reused by a non postinc user, but only if
1547 // its insertion point was already at the head of the loop.
1548 InsertedExpressions[std::make_pair(S, InsertPt)] = V;
1550 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
1554 void SCEVExpander::rememberInstruction(Value *I) {
1555 if (!PostIncLoops.empty())
1556 InsertedPostIncValues.insert(I);
1558 InsertedValues.insert(I);
1561 void SCEVExpander::restoreInsertPoint(BasicBlock *BB, BasicBlock::iterator I) {
1562 Builder.SetInsertPoint(BB, I);
1565 /// getOrInsertCanonicalInductionVariable - This method returns the
1566 /// canonical induction variable of the specified type for the specified
1567 /// loop (inserting one if there is none). A canonical induction variable
1568 /// starts at zero and steps by one on each iteration.
1570 SCEVExpander::getOrInsertCanonicalInductionVariable(const Loop *L,
1572 assert(Ty->isIntegerTy() && "Can only insert integer induction variables!");
1574 // Build a SCEV for {0,+,1}<L>.
1575 // Conservatively use FlagAnyWrap for now.
1576 const SCEV *H = SE.getAddRecExpr(SE.getConstant(Ty, 0),
1577 SE.getConstant(Ty, 1), L, SCEV::FlagAnyWrap);
1579 // Emit code for it.
1580 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1581 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1582 PHINode *V = cast<PHINode>(expandCodeFor(H, 0, L->getHeader()->begin()));
1584 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
1589 /// Sort values by integer width for replaceCongruentIVs.
1590 static bool width_descending(Value *lhs, Value *rhs) {
1591 // Put pointers at the back and make sure pointer < pointer = false.
1592 if (!lhs->getType()->isIntegerTy() || !rhs->getType()->isIntegerTy())
1593 return rhs->getType()->isIntegerTy() && !lhs->getType()->isIntegerTy();
1594 return rhs->getType()->getPrimitiveSizeInBits()
1595 < lhs->getType()->getPrimitiveSizeInBits();
1598 /// replaceCongruentIVs - Check for congruent phis in this loop header and
1599 /// replace them with their most canonical representative. Return the number of
1600 /// phis eliminated.
1602 /// This does not depend on any SCEVExpander state but should be used in
1603 /// the same context that SCEVExpander is used.
1604 unsigned SCEVExpander::replaceCongruentIVs(Loop *L, const DominatorTree *DT,
1605 SmallVectorImpl<WeakVH> &DeadInsts,
1606 const TargetTransformInfo *TTI) {
1607 // Find integer phis in order of increasing width.
1608 SmallVector<PHINode*, 8> Phis;
1609 for (BasicBlock::iterator I = L->getHeader()->begin();
1610 PHINode *Phi = dyn_cast<PHINode>(I); ++I) {
1611 Phis.push_back(Phi);
1614 std::sort(Phis.begin(), Phis.end(), width_descending);
1616 unsigned NumElim = 0;
1617 DenseMap<const SCEV *, PHINode *> ExprToIVMap;
1618 // Process phis from wide to narrow. Mapping wide phis to the their truncation
1619 // so narrow phis can reuse them.
1620 for (SmallVectorImpl<PHINode*>::const_iterator PIter = Phis.begin(),
1621 PEnd = Phis.end(); PIter != PEnd; ++PIter) {
1622 PHINode *Phi = *PIter;
1624 // Fold constant phis. They may be congruent to other constant phis and
1625 // would confuse the logic below that expects proper IVs.
1626 if (Value *V = Phi->hasConstantValue()) {
1627 Phi->replaceAllUsesWith(V);
1628 DeadInsts.push_back(Phi);
1630 DEBUG_WITH_TYPE(DebugType, dbgs()
1631 << "INDVARS: Eliminated constant iv: " << *Phi << '\n');
1635 if (!SE.isSCEVable(Phi->getType()))
1638 PHINode *&OrigPhiRef = ExprToIVMap[SE.getSCEV(Phi)];
1641 if (Phi->getType()->isIntegerTy() && TTI
1642 && TTI->isTruncateFree(Phi->getType(), Phis.back()->getType())) {
1643 // This phi can be freely truncated to the narrowest phi type. Map the
1644 // truncated expression to it so it will be reused for narrow types.
1645 const SCEV *TruncExpr =
1646 SE.getTruncateExpr(SE.getSCEV(Phi), Phis.back()->getType());
1647 ExprToIVMap[TruncExpr] = Phi;
1652 // Replacing a pointer phi with an integer phi or vice-versa doesn't make
1654 if (OrigPhiRef->getType()->isPointerTy() != Phi->getType()->isPointerTy())
1657 if (BasicBlock *LatchBlock = L->getLoopLatch()) {
1658 Instruction *OrigInc =
1659 cast<Instruction>(OrigPhiRef->getIncomingValueForBlock(LatchBlock));
1660 Instruction *IsomorphicInc =
1661 cast<Instruction>(Phi->getIncomingValueForBlock(LatchBlock));
1663 // If this phi has the same width but is more canonical, replace the
1664 // original with it. As part of the "more canonical" determination,
1665 // respect a prior decision to use an IV chain.
1666 if (OrigPhiRef->getType() == Phi->getType()
1667 && !(ChainedPhis.count(Phi)
1668 || isExpandedAddRecExprPHI(OrigPhiRef, OrigInc, L))
1669 && (ChainedPhis.count(Phi)
1670 || isExpandedAddRecExprPHI(Phi, IsomorphicInc, L))) {
1671 std::swap(OrigPhiRef, Phi);
1672 std::swap(OrigInc, IsomorphicInc);
1674 // Replacing the congruent phi is sufficient because acyclic redundancy
1675 // elimination, CSE/GVN, should handle the rest. However, once SCEV proves
1676 // that a phi is congruent, it's often the head of an IV user cycle that
1677 // is isomorphic with the original phi. It's worth eagerly cleaning up the
1678 // common case of a single IV increment so that DeleteDeadPHIs can remove
1679 // cycles that had postinc uses.
1680 const SCEV *TruncExpr = SE.getTruncateOrNoop(SE.getSCEV(OrigInc),
1681 IsomorphicInc->getType());
1682 if (OrigInc != IsomorphicInc
1683 && TruncExpr == SE.getSCEV(IsomorphicInc)
1684 && ((isa<PHINode>(OrigInc) && isa<PHINode>(IsomorphicInc))
1685 || hoistIVInc(OrigInc, IsomorphicInc))) {
1686 DEBUG_WITH_TYPE(DebugType, dbgs()
1687 << "INDVARS: Eliminated congruent iv.inc: "
1688 << *IsomorphicInc << '\n');
1689 Value *NewInc = OrigInc;
1690 if (OrigInc->getType() != IsomorphicInc->getType()) {
1691 Instruction *IP = isa<PHINode>(OrigInc)
1692 ? (Instruction*)L->getHeader()->getFirstInsertionPt()
1693 : OrigInc->getNextNode();
1694 IRBuilder<> Builder(IP);
1695 Builder.SetCurrentDebugLocation(IsomorphicInc->getDebugLoc());
1697 CreateTruncOrBitCast(OrigInc, IsomorphicInc->getType(), IVName);
1699 IsomorphicInc->replaceAllUsesWith(NewInc);
1700 DeadInsts.push_back(IsomorphicInc);
1703 DEBUG_WITH_TYPE(DebugType, dbgs()
1704 << "INDVARS: Eliminated congruent iv: " << *Phi << '\n');
1706 Value *NewIV = OrigPhiRef;
1707 if (OrigPhiRef->getType() != Phi->getType()) {
1708 IRBuilder<> Builder(L->getHeader()->getFirstInsertionPt());
1709 Builder.SetCurrentDebugLocation(Phi->getDebugLoc());
1710 NewIV = Builder.CreateTruncOrBitCast(OrigPhiRef, Phi->getType(), IVName);
1712 Phi->replaceAllUsesWith(NewIV);
1713 DeadInsts.push_back(Phi);
1719 // Search for a SCEV subexpression that is not safe to expand. Any expression
1720 // that may expand to a !isSafeToSpeculativelyExecute value is unsafe, namely
1721 // UDiv expressions. We don't know if the UDiv is derived from an IR divide
1722 // instruction, but the important thing is that we prove the denominator is
1723 // nonzero before expansion.
1725 // IVUsers already checks that IV-derived expressions are safe. So this check is
1726 // only needed when the expression includes some subexpression that is not IV
1729 // Currently, we only allow division by a nonzero constant here. If this is
1730 // inadequate, we could easily allow division by SCEVUnknown by using
1731 // ValueTracking to check isKnownNonZero().
1732 struct SCEVFindUnsafe {
1735 SCEVFindUnsafe(): IsUnsafe(false) {}
1737 bool follow(const SCEV *S) {
1738 const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S);
1741 const SCEVConstant *SC = dyn_cast<SCEVConstant>(D->getRHS());
1742 if (SC && !SC->getValue()->isZero())
1747 bool isDone() const { return IsUnsafe; }
1752 bool isSafeToExpand(const SCEV *S) {
1753 SCEVFindUnsafe Search;
1754 visitAll(S, Search);
1755 return !Search.IsUnsafe;