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/Analysis/LoopInfo.h"
18 #include "llvm/IntrinsicInst.h"
19 #include "llvm/LLVMContext.h"
20 #include "llvm/Target/TargetData.h"
21 #include "llvm/ADT/STLExtras.h"
25 /// ReuseOrCreateCast - Arrange for there to be a cast of V to Ty at IP,
26 /// reusing an existing cast if a suitable one exists, moving an existing
27 /// cast if a suitable one exists but isn't in the right place, or
28 /// creating a new one.
29 Value *SCEVExpander::ReuseOrCreateCast(Value *V, const Type *Ty,
30 Instruction::CastOps Op,
31 BasicBlock::iterator IP) {
32 // Check to see if there is already a cast!
33 for (Value::use_iterator UI = V->use_begin(), E = V->use_end();
36 if (U->getType() == Ty)
37 if (CastInst *CI = dyn_cast<CastInst>(U))
38 if (CI->getOpcode() == Op) {
39 // If the cast isn't where we want it, fix it.
40 if (BasicBlock::iterator(CI) != IP) {
41 // Create a new cast, and leave the old cast in place in case
42 // it is being used as an insert point. Clear its operand
43 // so that it doesn't hold anything live.
44 Instruction *NewCI = CastInst::Create(Op, V, Ty, "", IP);
46 CI->replaceAllUsesWith(NewCI);
47 CI->setOperand(0, UndefValue::get(V->getType()));
48 rememberInstruction(NewCI);
51 rememberInstruction(CI);
57 Instruction *I = CastInst::Create(Op, V, Ty, V->getName(), IP);
58 rememberInstruction(I);
62 /// InsertNoopCastOfTo - Insert a cast of V to the specified type,
63 /// which must be possible with a noop cast, doing what we can to share
65 Value *SCEVExpander::InsertNoopCastOfTo(Value *V, const Type *Ty) {
66 Instruction::CastOps Op = CastInst::getCastOpcode(V, false, Ty, false);
67 assert((Op == Instruction::BitCast ||
68 Op == Instruction::PtrToInt ||
69 Op == Instruction::IntToPtr) &&
70 "InsertNoopCastOfTo cannot perform non-noop casts!");
71 assert(SE.getTypeSizeInBits(V->getType()) == SE.getTypeSizeInBits(Ty) &&
72 "InsertNoopCastOfTo cannot change sizes!");
74 // Short-circuit unnecessary bitcasts.
75 if (Op == Instruction::BitCast && V->getType() == Ty)
78 // Short-circuit unnecessary inttoptr<->ptrtoint casts.
79 if ((Op == Instruction::PtrToInt || Op == Instruction::IntToPtr) &&
80 SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(V->getType())) {
81 if (CastInst *CI = dyn_cast<CastInst>(V))
82 if ((CI->getOpcode() == Instruction::PtrToInt ||
83 CI->getOpcode() == Instruction::IntToPtr) &&
84 SE.getTypeSizeInBits(CI->getType()) ==
85 SE.getTypeSizeInBits(CI->getOperand(0)->getType()))
86 return CI->getOperand(0);
87 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
88 if ((CE->getOpcode() == Instruction::PtrToInt ||
89 CE->getOpcode() == Instruction::IntToPtr) &&
90 SE.getTypeSizeInBits(CE->getType()) ==
91 SE.getTypeSizeInBits(CE->getOperand(0)->getType()))
92 return CE->getOperand(0);
95 // Fold a cast of a constant.
96 if (Constant *C = dyn_cast<Constant>(V))
97 return ConstantExpr::getCast(Op, C, Ty);
99 // Cast the argument at the beginning of the entry block, after
100 // any bitcasts of other arguments.
101 if (Argument *A = dyn_cast<Argument>(V)) {
102 BasicBlock::iterator IP = A->getParent()->getEntryBlock().begin();
103 while ((isa<BitCastInst>(IP) &&
104 isa<Argument>(cast<BitCastInst>(IP)->getOperand(0)) &&
105 cast<BitCastInst>(IP)->getOperand(0) != A) ||
106 isa<DbgInfoIntrinsic>(IP))
108 return ReuseOrCreateCast(A, Ty, Op, IP);
111 // Cast the instruction immediately after the instruction.
112 Instruction *I = cast<Instruction>(V);
113 BasicBlock::iterator IP = I; ++IP;
114 if (InvokeInst *II = dyn_cast<InvokeInst>(I))
115 IP = II->getNormalDest()->begin();
116 while (isa<PHINode>(IP) || isa<DbgInfoIntrinsic>(IP)) ++IP;
117 return ReuseOrCreateCast(I, Ty, Op, IP);
120 /// InsertBinop - Insert the specified binary operator, doing a small amount
121 /// of work to avoid inserting an obviously redundant operation.
122 Value *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode,
123 Value *LHS, Value *RHS) {
124 // Fold a binop with constant operands.
125 if (Constant *CLHS = dyn_cast<Constant>(LHS))
126 if (Constant *CRHS = dyn_cast<Constant>(RHS))
127 return ConstantExpr::get(Opcode, CLHS, CRHS);
129 // Do a quick scan to see if we have this binop nearby. If so, reuse it.
130 unsigned ScanLimit = 6;
131 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
132 // Scanning starts from the last instruction before the insertion point.
133 BasicBlock::iterator IP = Builder.GetInsertPoint();
134 if (IP != BlockBegin) {
136 for (; ScanLimit; --IP, --ScanLimit) {
137 // Don't count dbg.value against the ScanLimit, to avoid perturbing the
139 if (isa<DbgInfoIntrinsic>(IP))
141 if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(0) == LHS &&
142 IP->getOperand(1) == RHS)
144 if (IP == BlockBegin) break;
148 // Save the original insertion point so we can restore it when we're done.
149 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
150 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
152 // Move the insertion point out of as many loops as we can.
153 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
154 if (!L->isLoopInvariant(LHS) || !L->isLoopInvariant(RHS)) break;
155 BasicBlock *Preheader = L->getLoopPreheader();
156 if (!Preheader) break;
158 // Ok, move up a level.
159 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
162 // If we haven't found this binop, insert it.
163 Instruction *BO = cast<Instruction>(Builder.CreateBinOp(Opcode, LHS, RHS, "tmp"));
164 BO->setDebugLoc(SaveInsertPt->getDebugLoc());
165 rememberInstruction(BO);
167 // Restore the original insert point.
169 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
174 /// FactorOutConstant - Test if S is divisible by Factor, using signed
175 /// division. If so, update S with Factor divided out and return true.
176 /// S need not be evenly divisible if a reasonable remainder can be
178 /// TODO: When ScalarEvolution gets a SCEVSDivExpr, this can be made
179 /// unnecessary; in its place, just signed-divide Ops[i] by the scale and
180 /// check to see if the divide was folded.
181 static bool FactorOutConstant(const SCEV *&S,
182 const SCEV *&Remainder,
185 const TargetData *TD) {
186 // Everything is divisible by one.
192 S = SE.getConstant(S->getType(), 1);
196 // For a Constant, check for a multiple of the given factor.
197 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
201 // Check for divisibility.
202 if (const SCEVConstant *FC = dyn_cast<SCEVConstant>(Factor)) {
204 ConstantInt::get(SE.getContext(),
205 C->getValue()->getValue().sdiv(
206 FC->getValue()->getValue()));
207 // If the quotient is zero and the remainder is non-zero, reject
208 // the value at this scale. It will be considered for subsequent
211 const SCEV *Div = SE.getConstant(CI);
214 SE.getAddExpr(Remainder,
215 SE.getConstant(C->getValue()->getValue().srem(
216 FC->getValue()->getValue())));
222 // In a Mul, check if there is a constant operand which is a multiple
223 // of the given factor.
224 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
226 // With TargetData, the size is known. Check if there is a constant
227 // operand which is a multiple of the given factor. If so, we can
229 const SCEVConstant *FC = cast<SCEVConstant>(Factor);
230 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0)))
231 if (!C->getValue()->getValue().srem(FC->getValue()->getValue())) {
232 SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end());
234 SE.getConstant(C->getValue()->getValue().sdiv(
235 FC->getValue()->getValue()));
236 S = SE.getMulExpr(NewMulOps);
240 // Without TargetData, check if Factor can be factored out of any of the
241 // Mul's operands. If so, we can just remove it.
242 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) {
243 const SCEV *SOp = M->getOperand(i);
244 const SCEV *Remainder = SE.getConstant(SOp->getType(), 0);
245 if (FactorOutConstant(SOp, Remainder, Factor, SE, TD) &&
246 Remainder->isZero()) {
247 SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end());
249 S = SE.getMulExpr(NewMulOps);
256 // In an AddRec, check if both start and step are divisible.
257 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
258 const SCEV *Step = A->getStepRecurrence(SE);
259 const SCEV *StepRem = SE.getConstant(Step->getType(), 0);
260 if (!FactorOutConstant(Step, StepRem, Factor, SE, TD))
262 if (!StepRem->isZero())
264 const SCEV *Start = A->getStart();
265 if (!FactorOutConstant(Start, Remainder, Factor, SE, TD))
267 // FIXME: can use A->getNoWrapFlags(FlagNW)
268 S = SE.getAddRecExpr(Start, Step, A->getLoop(), SCEV::FlagAnyWrap);
275 /// SimplifyAddOperands - Sort and simplify a list of add operands. NumAddRecs
276 /// is the number of SCEVAddRecExprs present, which are kept at the end of
279 static void SimplifyAddOperands(SmallVectorImpl<const SCEV *> &Ops,
281 ScalarEvolution &SE) {
282 unsigned NumAddRecs = 0;
283 for (unsigned i = Ops.size(); i > 0 && isa<SCEVAddRecExpr>(Ops[i-1]); --i)
285 // Group Ops into non-addrecs and addrecs.
286 SmallVector<const SCEV *, 8> NoAddRecs(Ops.begin(), Ops.end() - NumAddRecs);
287 SmallVector<const SCEV *, 8> AddRecs(Ops.end() - NumAddRecs, Ops.end());
288 // Let ScalarEvolution sort and simplify the non-addrecs list.
289 const SCEV *Sum = NoAddRecs.empty() ?
290 SE.getConstant(Ty, 0) :
291 SE.getAddExpr(NoAddRecs);
292 // If it returned an add, use the operands. Otherwise it simplified
293 // the sum into a single value, so just use that.
295 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Sum))
296 Ops.append(Add->op_begin(), Add->op_end());
297 else if (!Sum->isZero())
299 // Then append the addrecs.
300 Ops.append(AddRecs.begin(), AddRecs.end());
303 /// SplitAddRecs - Flatten a list of add operands, moving addrec start values
304 /// out to the top level. For example, convert {a + b,+,c} to a, b, {0,+,d}.
305 /// This helps expose more opportunities for folding parts of the expressions
306 /// into GEP indices.
308 static void SplitAddRecs(SmallVectorImpl<const SCEV *> &Ops,
310 ScalarEvolution &SE) {
312 SmallVector<const SCEV *, 8> AddRecs;
313 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
314 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Ops[i])) {
315 const SCEV *Start = A->getStart();
316 if (Start->isZero()) break;
317 const SCEV *Zero = SE.getConstant(Ty, 0);
318 AddRecs.push_back(SE.getAddRecExpr(Zero,
319 A->getStepRecurrence(SE),
321 // FIXME: A->getNoWrapFlags(FlagNW)
323 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Start)) {
325 Ops.append(Add->op_begin(), Add->op_end());
326 e += Add->getNumOperands();
331 if (!AddRecs.empty()) {
332 // Add the addrecs onto the end of the list.
333 Ops.append(AddRecs.begin(), AddRecs.end());
334 // Resort the operand list, moving any constants to the front.
335 SimplifyAddOperands(Ops, Ty, SE);
339 /// expandAddToGEP - Expand an addition expression with a pointer type into
340 /// a GEP instead of using ptrtoint+arithmetic+inttoptr. This helps
341 /// BasicAliasAnalysis and other passes analyze the result. See the rules
342 /// for getelementptr vs. inttoptr in
343 /// http://llvm.org/docs/LangRef.html#pointeraliasing
346 /// Design note: The correctness of using getelementptr here depends on
347 /// ScalarEvolution not recognizing inttoptr and ptrtoint operators, as
348 /// they may introduce pointer arithmetic which may not be safely converted
349 /// into getelementptr.
351 /// Design note: It might seem desirable for this function to be more
352 /// loop-aware. If some of the indices are loop-invariant while others
353 /// aren't, it might seem desirable to emit multiple GEPs, keeping the
354 /// loop-invariant portions of the overall computation outside the loop.
355 /// However, there are a few reasons this is not done here. Hoisting simple
356 /// arithmetic is a low-level optimization that often isn't very
357 /// important until late in the optimization process. In fact, passes
358 /// like InstructionCombining will combine GEPs, even if it means
359 /// pushing loop-invariant computation down into loops, so even if the
360 /// GEPs were split here, the work would quickly be undone. The
361 /// LoopStrengthReduction pass, which is usually run quite late (and
362 /// after the last InstructionCombining pass), takes care of hoisting
363 /// loop-invariant portions of expressions, after considering what
364 /// can be folded using target addressing modes.
366 Value *SCEVExpander::expandAddToGEP(const SCEV *const *op_begin,
367 const SCEV *const *op_end,
368 const PointerType *PTy,
371 const Type *ElTy = PTy->getElementType();
372 SmallVector<Value *, 4> GepIndices;
373 SmallVector<const SCEV *, 8> Ops(op_begin, op_end);
374 bool AnyNonZeroIndices = false;
376 // Split AddRecs up into parts as either of the parts may be usable
377 // without the other.
378 SplitAddRecs(Ops, Ty, SE);
380 // Descend down the pointer's type and attempt to convert the other
381 // operands into GEP indices, at each level. The first index in a GEP
382 // indexes into the array implied by the pointer operand; the rest of
383 // the indices index into the element or field type selected by the
386 // If the scale size is not 0, attempt to factor out a scale for
388 SmallVector<const SCEV *, 8> ScaledOps;
389 if (ElTy->isSized()) {
390 const SCEV *ElSize = SE.getSizeOfExpr(ElTy);
391 if (!ElSize->isZero()) {
392 SmallVector<const SCEV *, 8> NewOps;
393 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
394 const SCEV *Op = Ops[i];
395 const SCEV *Remainder = SE.getConstant(Ty, 0);
396 if (FactorOutConstant(Op, Remainder, ElSize, SE, SE.TD)) {
397 // Op now has ElSize factored out.
398 ScaledOps.push_back(Op);
399 if (!Remainder->isZero())
400 NewOps.push_back(Remainder);
401 AnyNonZeroIndices = true;
403 // The operand was not divisible, so add it to the list of operands
404 // we'll scan next iteration.
405 NewOps.push_back(Ops[i]);
408 // If we made any changes, update Ops.
409 if (!ScaledOps.empty()) {
411 SimplifyAddOperands(Ops, Ty, SE);
416 // Record the scaled array index for this level of the type. If
417 // we didn't find any operands that could be factored, tentatively
418 // assume that element zero was selected (since the zero offset
419 // would obviously be folded away).
420 Value *Scaled = ScaledOps.empty() ?
421 Constant::getNullValue(Ty) :
422 expandCodeFor(SE.getAddExpr(ScaledOps), Ty);
423 GepIndices.push_back(Scaled);
425 // Collect struct field index operands.
426 while (const StructType *STy = dyn_cast<StructType>(ElTy)) {
427 bool FoundFieldNo = false;
428 // An empty struct has no fields.
429 if (STy->getNumElements() == 0) break;
431 // With TargetData, field offsets are known. See if a constant offset
432 // falls within any of the struct fields.
433 if (Ops.empty()) break;
434 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[0]))
435 if (SE.getTypeSizeInBits(C->getType()) <= 64) {
436 const StructLayout &SL = *SE.TD->getStructLayout(STy);
437 uint64_t FullOffset = C->getValue()->getZExtValue();
438 if (FullOffset < SL.getSizeInBytes()) {
439 unsigned ElIdx = SL.getElementContainingOffset(FullOffset);
440 GepIndices.push_back(
441 ConstantInt::get(Type::getInt32Ty(Ty->getContext()), ElIdx));
442 ElTy = STy->getTypeAtIndex(ElIdx);
444 SE.getConstant(Ty, FullOffset - SL.getElementOffset(ElIdx));
445 AnyNonZeroIndices = true;
450 // Without TargetData, just check for an offsetof expression of the
451 // appropriate struct type.
452 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
453 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Ops[i])) {
456 if (U->isOffsetOf(CTy, FieldNo) && CTy == STy) {
457 GepIndices.push_back(FieldNo);
459 STy->getTypeAtIndex(cast<ConstantInt>(FieldNo)->getZExtValue());
460 Ops[i] = SE.getConstant(Ty, 0);
461 AnyNonZeroIndices = true;
467 // If no struct field offsets were found, tentatively assume that
468 // field zero was selected (since the zero offset would obviously
471 ElTy = STy->getTypeAtIndex(0u);
472 GepIndices.push_back(
473 Constant::getNullValue(Type::getInt32Ty(Ty->getContext())));
477 if (const ArrayType *ATy = dyn_cast<ArrayType>(ElTy))
478 ElTy = ATy->getElementType();
483 // If none of the operands were convertible to proper GEP indices, cast
484 // the base to i8* and do an ugly getelementptr with that. It's still
485 // better than ptrtoint+arithmetic+inttoptr at least.
486 if (!AnyNonZeroIndices) {
487 // Cast the base to i8*.
488 V = InsertNoopCastOfTo(V,
489 Type::getInt8PtrTy(Ty->getContext(), PTy->getAddressSpace()));
491 // Expand the operands for a plain byte offset.
492 Value *Idx = expandCodeFor(SE.getAddExpr(Ops), Ty);
494 // Fold a GEP with constant operands.
495 if (Constant *CLHS = dyn_cast<Constant>(V))
496 if (Constant *CRHS = dyn_cast<Constant>(Idx))
497 return ConstantExpr::getGetElementPtr(CLHS, &CRHS, 1);
499 // Do a quick scan to see if we have this GEP nearby. If so, reuse it.
500 unsigned ScanLimit = 6;
501 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
502 // Scanning starts from the last instruction before the insertion point.
503 BasicBlock::iterator IP = Builder.GetInsertPoint();
504 if (IP != BlockBegin) {
506 for (; ScanLimit; --IP, --ScanLimit) {
507 // Don't count dbg.value against the ScanLimit, to avoid perturbing the
509 if (isa<DbgInfoIntrinsic>(IP))
511 if (IP->getOpcode() == Instruction::GetElementPtr &&
512 IP->getOperand(0) == V && IP->getOperand(1) == Idx)
514 if (IP == BlockBegin) break;
518 // Save the original insertion point so we can restore it when we're done.
519 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
520 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
522 // Move the insertion point out of as many loops as we can.
523 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
524 if (!L->isLoopInvariant(V) || !L->isLoopInvariant(Idx)) break;
525 BasicBlock *Preheader = L->getLoopPreheader();
526 if (!Preheader) break;
528 // Ok, move up a level.
529 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
533 Value *GEP = Builder.CreateGEP(V, Idx, "uglygep");
534 rememberInstruction(GEP);
536 // Restore the original insert point.
538 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
543 // Save the original insertion point so we can restore it when we're done.
544 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
545 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
547 // Move the insertion point out of as many loops as we can.
548 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
549 if (!L->isLoopInvariant(V)) break;
551 bool AnyIndexNotLoopInvariant = false;
552 for (SmallVectorImpl<Value *>::const_iterator I = GepIndices.begin(),
553 E = GepIndices.end(); I != E; ++I)
554 if (!L->isLoopInvariant(*I)) {
555 AnyIndexNotLoopInvariant = true;
558 if (AnyIndexNotLoopInvariant)
561 BasicBlock *Preheader = L->getLoopPreheader();
562 if (!Preheader) break;
564 // Ok, move up a level.
565 Builder.SetInsertPoint(Preheader, 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(Casted,
578 Ops.push_back(SE.getUnknown(GEP));
579 rememberInstruction(GEP);
581 // Restore the original insert point.
583 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
585 return expand(SE.getAddExpr(Ops));
588 /// isNonConstantNegative - Return true if the specified scev is negated, but
590 static bool isNonConstantNegative(const SCEV *F) {
591 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(F);
592 if (!Mul) return false;
594 // If there is a constant factor, it will be first.
595 const SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
596 if (!SC) return false;
598 // Return true if the value is negative, this matches things like (-42 * V).
599 return SC->getValue()->getValue().isNegative();
602 /// PickMostRelevantLoop - Given two loops pick the one that's most relevant for
603 /// SCEV expansion. If they are nested, this is the most nested. If they are
604 /// neighboring, pick the later.
605 static const Loop *PickMostRelevantLoop(const Loop *A, const Loop *B,
609 if (A->contains(B)) return B;
610 if (B->contains(A)) return A;
611 if (DT.dominates(A->getHeader(), B->getHeader())) return B;
612 if (DT.dominates(B->getHeader(), A->getHeader())) return A;
613 return A; // Arbitrarily break the tie.
616 /// getRelevantLoop - Get the most relevant loop associated with the given
617 /// expression, according to PickMostRelevantLoop.
618 const Loop *SCEVExpander::getRelevantLoop(const SCEV *S) {
619 // Test whether we've already computed the most relevant loop for this SCEV.
620 std::pair<DenseMap<const SCEV *, const Loop *>::iterator, bool> Pair =
621 RelevantLoops.insert(std::make_pair(S, static_cast<const Loop *>(0)));
623 return Pair.first->second;
625 if (isa<SCEVConstant>(S))
626 // A constant has no relevant loops.
628 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
629 if (const Instruction *I = dyn_cast<Instruction>(U->getValue()))
630 return Pair.first->second = SE.LI->getLoopFor(I->getParent());
631 // A non-instruction has no relevant loops.
634 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S)) {
636 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
638 for (SCEVNAryExpr::op_iterator I = N->op_begin(), E = N->op_end();
640 L = PickMostRelevantLoop(L, getRelevantLoop(*I), *SE.DT);
641 return RelevantLoops[N] = L;
643 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S)) {
644 const Loop *Result = getRelevantLoop(C->getOperand());
645 return RelevantLoops[C] = Result;
647 if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
649 PickMostRelevantLoop(getRelevantLoop(D->getLHS()),
650 getRelevantLoop(D->getRHS()),
652 return RelevantLoops[D] = Result;
654 llvm_unreachable("Unexpected SCEV type!");
660 /// LoopCompare - Compare loops by PickMostRelevantLoop.
664 explicit LoopCompare(DominatorTree &dt) : DT(dt) {}
666 bool operator()(std::pair<const Loop *, const SCEV *> LHS,
667 std::pair<const Loop *, const SCEV *> RHS) const {
668 // Keep pointer operands sorted at the end.
669 if (LHS.second->getType()->isPointerTy() !=
670 RHS.second->getType()->isPointerTy())
671 return LHS.second->getType()->isPointerTy();
673 // Compare loops with PickMostRelevantLoop.
674 if (LHS.first != RHS.first)
675 return PickMostRelevantLoop(LHS.first, RHS.first, DT) != LHS.first;
677 // If one operand is a non-constant negative and the other is not,
678 // put the non-constant negative on the right so that a sub can
679 // be used instead of a negate and add.
680 if (isNonConstantNegative(LHS.second)) {
681 if (!isNonConstantNegative(RHS.second))
683 } else if (isNonConstantNegative(RHS.second))
686 // Otherwise they are equivalent according to this comparison.
693 Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) {
694 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
696 // Collect all the add operands in a loop, along with their associated loops.
697 // Iterate in reverse so that constants are emitted last, all else equal, and
698 // so that pointer operands are inserted first, which the code below relies on
699 // to form more involved GEPs.
700 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops;
701 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(S->op_end()),
702 E(S->op_begin()); I != E; ++I)
703 OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I));
705 // Sort by loop. Use a stable sort so that constants follow non-constants and
706 // pointer operands precede non-pointer operands.
707 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT));
709 // Emit instructions to add all the operands. Hoist as much as possible
710 // out of loops, and form meaningful getelementptrs where possible.
712 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator
713 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) {
714 const Loop *CurLoop = I->first;
715 const SCEV *Op = I->second;
717 // This is the first operand. Just expand it.
720 } else if (const PointerType *PTy = dyn_cast<PointerType>(Sum->getType())) {
721 // The running sum expression is a pointer. Try to form a getelementptr
722 // at this level with that as the base.
723 SmallVector<const SCEV *, 4> NewOps;
724 for (; I != E && I->first == CurLoop; ++I) {
725 // If the operand is SCEVUnknown and not instructions, peek through
726 // it, to enable more of it to be folded into the GEP.
727 const SCEV *X = I->second;
728 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(X))
729 if (!isa<Instruction>(U->getValue()))
730 X = SE.getSCEV(U->getValue());
733 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, Sum);
734 } else if (const PointerType *PTy = dyn_cast<PointerType>(Op->getType())) {
735 // The running sum is an integer, and there's a pointer at this level.
736 // Try to form a getelementptr. If the running sum is instructions,
737 // use a SCEVUnknown to avoid re-analyzing them.
738 SmallVector<const SCEV *, 4> NewOps;
739 NewOps.push_back(isa<Instruction>(Sum) ? SE.getUnknown(Sum) :
741 for (++I; I != E && I->first == CurLoop; ++I)
742 NewOps.push_back(I->second);
743 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, expand(Op));
744 } else if (isNonConstantNegative(Op)) {
745 // Instead of doing a negate and add, just do a subtract.
746 Value *W = expandCodeFor(SE.getNegativeSCEV(Op), Ty);
747 Sum = InsertNoopCastOfTo(Sum, Ty);
748 Sum = InsertBinop(Instruction::Sub, Sum, W);
752 Value *W = expandCodeFor(Op, Ty);
753 Sum = InsertNoopCastOfTo(Sum, Ty);
754 // Canonicalize a constant to the RHS.
755 if (isa<Constant>(Sum)) std::swap(Sum, W);
756 Sum = InsertBinop(Instruction::Add, Sum, W);
764 Value *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) {
765 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
767 // Collect all the mul operands in a loop, along with their associated loops.
768 // Iterate in reverse so that constants are emitted last, all else equal.
769 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops;
770 for (std::reverse_iterator<SCEVMulExpr::op_iterator> I(S->op_end()),
771 E(S->op_begin()); I != E; ++I)
772 OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I));
774 // Sort by loop. Use a stable sort so that constants follow non-constants.
775 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT));
777 // Emit instructions to mul all the operands. Hoist as much as possible
780 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator
781 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) {
782 const SCEV *Op = I->second;
784 // This is the first operand. Just expand it.
787 } else if (Op->isAllOnesValue()) {
788 // Instead of doing a multiply by negative one, just do a negate.
789 Prod = InsertNoopCastOfTo(Prod, Ty);
790 Prod = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), Prod);
794 Value *W = expandCodeFor(Op, Ty);
795 Prod = InsertNoopCastOfTo(Prod, Ty);
796 // Canonicalize a constant to the RHS.
797 if (isa<Constant>(Prod)) std::swap(Prod, W);
798 Prod = InsertBinop(Instruction::Mul, Prod, W);
806 Value *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) {
807 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
809 Value *LHS = expandCodeFor(S->getLHS(), Ty);
810 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getRHS())) {
811 const APInt &RHS = SC->getValue()->getValue();
812 if (RHS.isPowerOf2())
813 return InsertBinop(Instruction::LShr, LHS,
814 ConstantInt::get(Ty, RHS.logBase2()));
817 Value *RHS = expandCodeFor(S->getRHS(), Ty);
818 return InsertBinop(Instruction::UDiv, LHS, RHS);
821 /// Move parts of Base into Rest to leave Base with the minimal
822 /// expression that provides a pointer operand suitable for a
824 static void ExposePointerBase(const SCEV *&Base, const SCEV *&Rest,
825 ScalarEvolution &SE) {
826 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Base)) {
827 Base = A->getStart();
828 Rest = SE.getAddExpr(Rest,
829 SE.getAddRecExpr(SE.getConstant(A->getType(), 0),
830 A->getStepRecurrence(SE),
832 // FIXME: A->getNoWrapFlags(FlagNW)
835 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(Base)) {
836 Base = A->getOperand(A->getNumOperands()-1);
837 SmallVector<const SCEV *, 8> NewAddOps(A->op_begin(), A->op_end());
838 NewAddOps.back() = Rest;
839 Rest = SE.getAddExpr(NewAddOps);
840 ExposePointerBase(Base, Rest, SE);
844 /// getAddRecExprPHILiterally - Helper for expandAddRecExprLiterally. Expand
845 /// the base addrec, which is the addrec without any non-loop-dominating
846 /// values, and return the PHI.
848 SCEVExpander::getAddRecExprPHILiterally(const SCEVAddRecExpr *Normalized,
850 const Type *ExpandTy,
852 assert(!IVIncInsertLoop || IVIncInsertPos && "Uninitialized insert position");
854 // Reuse a previously-inserted PHI, if present.
855 for (BasicBlock::iterator I = L->getHeader()->begin();
856 PHINode *PN = dyn_cast<PHINode>(I); ++I)
857 if (SE.isSCEVable(PN->getType()) &&
858 (SE.getEffectiveSCEVType(PN->getType()) ==
859 SE.getEffectiveSCEVType(Normalized->getType())) &&
860 SE.getSCEV(PN) == Normalized)
861 if (BasicBlock *LatchBlock = L->getLoopLatch()) {
863 cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock));
865 // Determine if this is a well-behaved chain of instructions leading
866 // back to the PHI. It probably will be, if we're scanning an inner
867 // loop already visited by LSR for example, but it wouldn't have
870 if (IncV->getNumOperands() == 0 || isa<PHINode>(IncV) ||
871 (isa<CastInst>(IncV) && !isa<BitCastInst>(IncV))) {
875 // If any of the operands don't dominate the insert position, bail.
876 // Addrec operands are always loop-invariant, so this can only happen
877 // if there are instructions which haven't been hoisted.
878 if (L == IVIncInsertLoop) {
879 for (User::op_iterator OI = IncV->op_begin()+1,
880 OE = IncV->op_end(); OI != OE; ++OI)
881 if (Instruction *OInst = dyn_cast<Instruction>(OI))
882 if (!SE.DT->dominates(OInst, IVIncInsertPos)) {
889 // Advance to the next instruction.
890 IncV = dyn_cast<Instruction>(IncV->getOperand(0));
893 if (IncV->mayHaveSideEffects()) {
897 } while (IncV != PN);
900 // Ok, the add recurrence looks usable.
901 // Remember this PHI, even in post-inc mode.
902 InsertedValues.insert(PN);
903 // Remember the increment.
904 IncV = cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock));
905 rememberInstruction(IncV);
906 if (L == IVIncInsertLoop)
908 if (SE.DT->dominates(IncV, IVIncInsertPos))
910 // Make sure the increment is where we want it. But don't move it
911 // down past a potential existing post-inc user.
912 IncV->moveBefore(IVIncInsertPos);
913 IVIncInsertPos = IncV;
914 IncV = cast<Instruction>(IncV->getOperand(0));
915 } while (IncV != PN);
920 // Save the original insertion point so we can restore it when we're done.
921 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
922 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
924 // Expand code for the start value.
925 Value *StartV = expandCodeFor(Normalized->getStart(), ExpandTy,
926 L->getHeader()->begin());
928 // StartV must be hoisted into L's preheader to dominate the new phi.
929 Instruction *StartI = dyn_cast<Instruction>(StartV);
930 assert(!StartI || SE.DT->properlyDominates(StartI->getParent(),
931 L->getHeader()) && "");
934 // Expand code for the step value. Insert instructions right before the
935 // terminator corresponding to the back-edge. Do this before creating the PHI
936 // so that PHI reuse code doesn't see an incomplete PHI. If the stride is
937 // negative, insert a sub instead of an add for the increment (unless it's a
938 // constant, because subtracts of constants are canonicalized to adds).
939 const SCEV *Step = Normalized->getStepRecurrence(SE);
940 bool isPointer = ExpandTy->isPointerTy();
941 bool isNegative = !isPointer && isNonConstantNegative(Step);
943 Step = SE.getNegativeSCEV(Step);
944 Value *StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin());
947 BasicBlock *Header = L->getHeader();
948 Builder.SetInsertPoint(Header, Header->begin());
949 pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header);
950 PHINode *PN = Builder.CreatePHI(ExpandTy, std::distance(HPB, HPE),
951 Twine(IVName) + ".iv");
952 rememberInstruction(PN);
954 // Create the step instructions and populate the PHI.
955 for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) {
956 BasicBlock *Pred = *HPI;
958 // Add a start value.
959 if (!L->contains(Pred)) {
960 PN->addIncoming(StartV, Pred);
964 // Create a step value and add it to the PHI. If IVIncInsertLoop is
965 // non-null and equal to the addrec's loop, insert the instructions
966 // at IVIncInsertPos.
967 Instruction *InsertPos = L == IVIncInsertLoop ?
968 IVIncInsertPos : Pred->getTerminator();
969 Builder.SetInsertPoint(InsertPos);
971 // If the PHI is a pointer, use a GEP, otherwise use an add or sub.
973 const PointerType *GEPPtrTy = cast<PointerType>(ExpandTy);
974 // If the step isn't constant, don't use an implicitly scaled GEP, because
975 // that would require a multiply inside the loop.
976 if (!isa<ConstantInt>(StepV))
977 GEPPtrTy = PointerType::get(Type::getInt1Ty(SE.getContext()),
978 GEPPtrTy->getAddressSpace());
979 const SCEV *const StepArray[1] = { SE.getSCEV(StepV) };
980 IncV = expandAddToGEP(StepArray, StepArray+1, GEPPtrTy, IntTy, PN);
981 if (IncV->getType() != PN->getType()) {
982 IncV = Builder.CreateBitCast(IncV, PN->getType(), "tmp");
983 rememberInstruction(IncV);
987 Builder.CreateSub(PN, StepV, Twine(IVName) + ".iv.next") :
988 Builder.CreateAdd(PN, StepV, Twine(IVName) + ".iv.next");
989 rememberInstruction(IncV);
991 PN->addIncoming(IncV, Pred);
994 // Restore the original insert point.
996 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
998 // Remember this PHI, even in post-inc mode.
999 InsertedValues.insert(PN);
1004 Value *SCEVExpander::expandAddRecExprLiterally(const SCEVAddRecExpr *S) {
1005 const Type *STy = S->getType();
1006 const Type *IntTy = SE.getEffectiveSCEVType(STy);
1007 const Loop *L = S->getLoop();
1009 // Determine a normalized form of this expression, which is the expression
1010 // before any post-inc adjustment is made.
1011 const SCEVAddRecExpr *Normalized = S;
1012 if (PostIncLoops.count(L)) {
1013 PostIncLoopSet Loops;
1016 cast<SCEVAddRecExpr>(TransformForPostIncUse(Normalize, S, 0, 0,
1017 Loops, SE, *SE.DT));
1020 // Strip off any non-loop-dominating component from the addrec start.
1021 const SCEV *Start = Normalized->getStart();
1022 const SCEV *PostLoopOffset = 0;
1023 if (!SE.properlyDominates(Start, L->getHeader())) {
1024 PostLoopOffset = Start;
1025 Start = SE.getConstant(Normalized->getType(), 0);
1026 Normalized = cast<SCEVAddRecExpr>(
1027 SE.getAddRecExpr(Start, Normalized->getStepRecurrence(SE),
1028 Normalized->getLoop(),
1029 // FIXME: Normalized->getNoWrapFlags(FlagNW)
1030 SCEV::FlagAnyWrap));
1033 // Strip off any non-loop-dominating component from the addrec step.
1034 const SCEV *Step = Normalized->getStepRecurrence(SE);
1035 const SCEV *PostLoopScale = 0;
1036 if (!SE.dominates(Step, L->getHeader())) {
1037 PostLoopScale = Step;
1038 Step = SE.getConstant(Normalized->getType(), 1);
1040 cast<SCEVAddRecExpr>(SE.getAddRecExpr(Start, Step,
1041 Normalized->getLoop(),
1042 // FIXME: Normalized
1043 // ->getNoWrapFlags(FlagNW)
1044 SCEV::FlagAnyWrap));
1047 // Expand the core addrec. If we need post-loop scaling, force it to
1048 // expand to an integer type to avoid the need for additional casting.
1049 const Type *ExpandTy = PostLoopScale ? IntTy : STy;
1050 PHINode *PN = getAddRecExprPHILiterally(Normalized, L, ExpandTy, IntTy);
1052 // Accommodate post-inc mode, if necessary.
1054 if (!PostIncLoops.count(L))
1057 // In PostInc mode, use the post-incremented value.
1058 BasicBlock *LatchBlock = L->getLoopLatch();
1059 assert(LatchBlock && "PostInc mode requires a unique loop latch!");
1060 Result = PN->getIncomingValueForBlock(LatchBlock);
1063 // Re-apply any non-loop-dominating scale.
1064 if (PostLoopScale) {
1065 Result = InsertNoopCastOfTo(Result, IntTy);
1066 Result = Builder.CreateMul(Result,
1067 expandCodeFor(PostLoopScale, IntTy));
1068 rememberInstruction(Result);
1071 // Re-apply any non-loop-dominating offset.
1072 if (PostLoopOffset) {
1073 if (const PointerType *PTy = dyn_cast<PointerType>(ExpandTy)) {
1074 const SCEV *const OffsetArray[1] = { PostLoopOffset };
1075 Result = expandAddToGEP(OffsetArray, OffsetArray+1, PTy, IntTy, Result);
1077 Result = InsertNoopCastOfTo(Result, IntTy);
1078 Result = Builder.CreateAdd(Result,
1079 expandCodeFor(PostLoopOffset, IntTy));
1080 rememberInstruction(Result);
1087 Value *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) {
1088 if (!CanonicalMode) return expandAddRecExprLiterally(S);
1090 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
1091 const Loop *L = S->getLoop();
1093 // First check for an existing canonical IV in a suitable type.
1094 PHINode *CanonicalIV = 0;
1095 if (PHINode *PN = L->getCanonicalInductionVariable())
1096 if (SE.getTypeSizeInBits(PN->getType()) >= SE.getTypeSizeInBits(Ty))
1099 // Rewrite an AddRec in terms of the canonical induction variable, if
1100 // its type is more narrow.
1102 SE.getTypeSizeInBits(CanonicalIV->getType()) >
1103 SE.getTypeSizeInBits(Ty)) {
1104 SmallVector<const SCEV *, 4> NewOps(S->getNumOperands());
1105 for (unsigned i = 0, e = S->getNumOperands(); i != e; ++i)
1106 NewOps[i] = SE.getAnyExtendExpr(S->op_begin()[i], CanonicalIV->getType());
1107 Value *V = expand(SE.getAddRecExpr(NewOps, S->getLoop(),
1108 // FIXME: S->getNoWrapFlags(FlagNW)
1109 SCEV::FlagAnyWrap));
1110 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1111 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1112 BasicBlock::iterator NewInsertPt =
1113 llvm::next(BasicBlock::iterator(cast<Instruction>(V)));
1114 while (isa<PHINode>(NewInsertPt) || isa<DbgInfoIntrinsic>(NewInsertPt))
1116 V = expandCodeFor(SE.getTruncateExpr(SE.getUnknown(V), Ty), 0,
1118 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
1122 // {X,+,F} --> X + {0,+,F}
1123 if (!S->getStart()->isZero()) {
1124 SmallVector<const SCEV *, 4> NewOps(S->op_begin(), S->op_end());
1125 NewOps[0] = SE.getConstant(Ty, 0);
1126 // FIXME: can use S->getNoWrapFlags()
1127 const SCEV *Rest = SE.getAddRecExpr(NewOps, L, SCEV::FlagAnyWrap);
1129 // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the
1130 // comments on expandAddToGEP for details.
1131 const SCEV *Base = S->getStart();
1132 const SCEV *RestArray[1] = { Rest };
1133 // Dig into the expression to find the pointer base for a GEP.
1134 ExposePointerBase(Base, RestArray[0], SE);
1135 // If we found a pointer, expand the AddRec with a GEP.
1136 if (const PointerType *PTy = dyn_cast<PointerType>(Base->getType())) {
1137 // Make sure the Base isn't something exotic, such as a multiplied
1138 // or divided pointer value. In those cases, the result type isn't
1139 // actually a pointer type.
1140 if (!isa<SCEVMulExpr>(Base) && !isa<SCEVUDivExpr>(Base)) {
1141 Value *StartV = expand(Base);
1142 assert(StartV->getType() == PTy && "Pointer type mismatch for GEP!");
1143 return expandAddToGEP(RestArray, RestArray+1, PTy, Ty, StartV);
1147 // Just do a normal add. Pre-expand the operands to suppress folding.
1148 return expand(SE.getAddExpr(SE.getUnknown(expand(S->getStart())),
1149 SE.getUnknown(expand(Rest))));
1152 // If we don't yet have a canonical IV, create one.
1154 // Create and insert the PHI node for the induction variable in the
1156 BasicBlock *Header = L->getHeader();
1157 pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header);
1158 CanonicalIV = PHINode::Create(Ty, std::distance(HPB, HPE), "indvar",
1160 rememberInstruction(CanonicalIV);
1162 Constant *One = ConstantInt::get(Ty, 1);
1163 for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) {
1164 BasicBlock *HP = *HPI;
1165 if (L->contains(HP)) {
1166 // Insert a unit add instruction right before the terminator
1167 // corresponding to the back-edge.
1168 Instruction *Add = BinaryOperator::CreateAdd(CanonicalIV, One,
1170 HP->getTerminator());
1171 Add->setDebugLoc(HP->getTerminator()->getDebugLoc());
1172 rememberInstruction(Add);
1173 CanonicalIV->addIncoming(Add, HP);
1175 CanonicalIV->addIncoming(Constant::getNullValue(Ty), HP);
1180 // {0,+,1} --> Insert a canonical induction variable into the loop!
1181 if (S->isAffine() && S->getOperand(1)->isOne()) {
1182 assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) &&
1183 "IVs with types different from the canonical IV should "
1184 "already have been handled!");
1188 // {0,+,F} --> {0,+,1} * F
1190 // If this is a simple linear addrec, emit it now as a special case.
1191 if (S->isAffine()) // {0,+,F} --> i*F
1193 expand(SE.getTruncateOrNoop(
1194 SE.getMulExpr(SE.getUnknown(CanonicalIV),
1195 SE.getNoopOrAnyExtend(S->getOperand(1),
1196 CanonicalIV->getType())),
1199 // If this is a chain of recurrences, turn it into a closed form, using the
1200 // folders, then expandCodeFor the closed form. This allows the folders to
1201 // simplify the expression without having to build a bunch of special code
1202 // into this folder.
1203 const SCEV *IH = SE.getUnknown(CanonicalIV); // Get I as a "symbolic" SCEV.
1205 // Promote S up to the canonical IV type, if the cast is foldable.
1206 const SCEV *NewS = S;
1207 const SCEV *Ext = SE.getNoopOrAnyExtend(S, CanonicalIV->getType());
1208 if (isa<SCEVAddRecExpr>(Ext))
1211 const SCEV *V = cast<SCEVAddRecExpr>(NewS)->evaluateAtIteration(IH, SE);
1212 //cerr << "Evaluated: " << *this << "\n to: " << *V << "\n";
1214 // Truncate the result down to the original type, if needed.
1215 const SCEV *T = SE.getTruncateOrNoop(V, Ty);
1219 Value *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) {
1220 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
1221 Value *V = expandCodeFor(S->getOperand(),
1222 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1223 Value *I = Builder.CreateTrunc(V, Ty, "tmp");
1224 rememberInstruction(I);
1228 Value *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) {
1229 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
1230 Value *V = expandCodeFor(S->getOperand(),
1231 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1232 Value *I = Builder.CreateZExt(V, Ty, "tmp");
1233 rememberInstruction(I);
1237 Value *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) {
1238 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
1239 Value *V = expandCodeFor(S->getOperand(),
1240 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1241 Value *I = Builder.CreateSExt(V, Ty, "tmp");
1242 rememberInstruction(I);
1246 Value *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) {
1247 Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
1248 const Type *Ty = LHS->getType();
1249 for (int i = S->getNumOperands()-2; i >= 0; --i) {
1250 // In the case of mixed integer and pointer types, do the
1251 // rest of the comparisons as integer.
1252 if (S->getOperand(i)->getType() != Ty) {
1253 Ty = SE.getEffectiveSCEVType(Ty);
1254 LHS = InsertNoopCastOfTo(LHS, Ty);
1256 Value *RHS = expandCodeFor(S->getOperand(i), Ty);
1257 Value *ICmp = Builder.CreateICmpSGT(LHS, RHS, "tmp");
1258 rememberInstruction(ICmp);
1259 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "smax");
1260 rememberInstruction(Sel);
1263 // In the case of mixed integer and pointer types, cast the
1264 // final result back to the pointer type.
1265 if (LHS->getType() != S->getType())
1266 LHS = InsertNoopCastOfTo(LHS, S->getType());
1270 Value *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) {
1271 Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
1272 const Type *Ty = LHS->getType();
1273 for (int i = S->getNumOperands()-2; i >= 0; --i) {
1274 // In the case of mixed integer and pointer types, do the
1275 // rest of the comparisons as integer.
1276 if (S->getOperand(i)->getType() != Ty) {
1277 Ty = SE.getEffectiveSCEVType(Ty);
1278 LHS = InsertNoopCastOfTo(LHS, Ty);
1280 Value *RHS = expandCodeFor(S->getOperand(i), Ty);
1281 Value *ICmp = Builder.CreateICmpUGT(LHS, RHS, "tmp");
1282 rememberInstruction(ICmp);
1283 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "umax");
1284 rememberInstruction(Sel);
1287 // In the case of mixed integer and pointer types, cast the
1288 // final result back to the pointer type.
1289 if (LHS->getType() != S->getType())
1290 LHS = InsertNoopCastOfTo(LHS, S->getType());
1294 Value *SCEVExpander::expandCodeFor(const SCEV *SH, const Type *Ty,
1296 BasicBlock::iterator IP = I;
1297 while (isInsertedInstruction(IP) || isa<DbgInfoIntrinsic>(IP))
1299 Builder.SetInsertPoint(IP->getParent(), IP);
1300 return expandCodeFor(SH, Ty);
1303 Value *SCEVExpander::expandCodeFor(const SCEV *SH, const Type *Ty) {
1304 // Expand the code for this SCEV.
1305 Value *V = expand(SH);
1307 assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) &&
1308 "non-trivial casts should be done with the SCEVs directly!");
1309 V = InsertNoopCastOfTo(V, Ty);
1314 Value *SCEVExpander::expand(const SCEV *S) {
1315 // Compute an insertion point for this SCEV object. Hoist the instructions
1316 // as far out in the loop nest as possible.
1317 Instruction *InsertPt = Builder.GetInsertPoint();
1318 for (Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock()); ;
1319 L = L->getParentLoop())
1320 if (SE.isLoopInvariant(S, L)) {
1322 if (BasicBlock *Preheader = L->getLoopPreheader())
1323 InsertPt = Preheader->getTerminator();
1325 // If the SCEV is computable at this level, insert it into the header
1326 // after the PHIs (and after any other instructions that we've inserted
1327 // there) so that it is guaranteed to dominate any user inside the loop.
1328 if (L && SE.hasComputableLoopEvolution(S, L) && !PostIncLoops.count(L))
1329 InsertPt = L->getHeader()->getFirstNonPHI();
1330 while (isInsertedInstruction(InsertPt) || isa<DbgInfoIntrinsic>(InsertPt))
1331 InsertPt = llvm::next(BasicBlock::iterator(InsertPt));
1335 // Check to see if we already expanded this here.
1336 std::map<std::pair<const SCEV *, Instruction *>,
1337 AssertingVH<Value> >::iterator I =
1338 InsertedExpressions.find(std::make_pair(S, InsertPt));
1339 if (I != InsertedExpressions.end())
1342 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1343 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1344 Builder.SetInsertPoint(InsertPt->getParent(), InsertPt);
1346 // Expand the expression into instructions.
1347 Value *V = visit(S);
1349 // Remember the expanded value for this SCEV at this location.
1350 if (PostIncLoops.empty())
1351 InsertedExpressions[std::make_pair(S, InsertPt)] = V;
1353 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
1357 void SCEVExpander::rememberInstruction(Value *I) {
1358 if (!PostIncLoops.empty())
1359 InsertedPostIncValues.insert(I);
1361 InsertedValues.insert(I);
1363 // If we just claimed an existing instruction and that instruction had
1364 // been the insert point, adjust the insert point forward so that
1365 // subsequently inserted code will be dominated.
1366 if (Builder.GetInsertPoint() == I) {
1367 BasicBlock::iterator It = cast<Instruction>(I);
1368 do { ++It; } while (isInsertedInstruction(It) ||
1369 isa<DbgInfoIntrinsic>(It));
1370 Builder.SetInsertPoint(Builder.GetInsertBlock(), It);
1374 void SCEVExpander::restoreInsertPoint(BasicBlock *BB, BasicBlock::iterator I) {
1375 // If we acquired more instructions since the old insert point was saved,
1376 // advance past them.
1377 while (isInsertedInstruction(I) || isa<DbgInfoIntrinsic>(I)) ++I;
1379 Builder.SetInsertPoint(BB, I);
1382 /// getOrInsertCanonicalInductionVariable - This method returns the
1383 /// canonical induction variable of the specified type for the specified
1384 /// loop (inserting one if there is none). A canonical induction variable
1385 /// starts at zero and steps by one on each iteration.
1387 SCEVExpander::getOrInsertCanonicalInductionVariable(const Loop *L,
1389 assert(Ty->isIntegerTy() && "Can only insert integer induction variables!");
1391 // Build a SCEV for {0,+,1}<L>.
1392 // Conservatively use FlagAnyWrap for now.
1393 const SCEV *H = SE.getAddRecExpr(SE.getConstant(Ty, 0),
1394 SE.getConstant(Ty, 1), L, SCEV::FlagAnyWrap);
1396 // Emit code for it.
1397 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1398 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1399 PHINode *V = cast<PHINode>(expandCodeFor(H, 0, L->getHeader()->begin()));
1401 restoreInsertPoint(SaveInsertBB, SaveInsertPt);