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"
24 /// ReuseOrCreateCast - Arrange for there to be a cast of V to Ty at IP,
25 /// reusing an existing cast if a suitable one exists, moving an existing
26 /// cast if a suitable one exists but isn't in the right place, or
27 /// creating a new one.
28 Value *SCEVExpander::ReuseOrCreateCast(Value *V, const Type *Ty,
29 Instruction::CastOps Op,
30 BasicBlock::iterator IP) {
31 // Check to see if there is already a cast!
32 for (Value::use_iterator UI = V->use_begin(), E = V->use_end();
35 if (U->getType() == Ty)
36 if (CastInst *CI = dyn_cast<CastInst>(U))
37 if (CI->getOpcode() == Op) {
38 // If the cast isn't where we want it, fix it.
39 if (BasicBlock::iterator(CI) != IP) {
40 // Create a new cast, and leave the old cast in place in case
41 // it is being used as an insert point. Clear its operand
42 // so that it doesn't hold anything live.
43 Instruction *NewCI = CastInst::Create(Op, V, Ty, "", IP);
45 CI->replaceAllUsesWith(NewCI);
46 CI->setOperand(0, UndefValue::get(V->getType()));
47 rememberInstruction(NewCI);
50 rememberInstruction(CI);
56 Instruction *I = CastInst::Create(Op, V, Ty, V->getName(), IP);
57 rememberInstruction(I);
61 /// InsertNoopCastOfTo - Insert a cast of V to the specified type,
62 /// which must be possible with a noop cast, doing what we can to share
64 Value *SCEVExpander::InsertNoopCastOfTo(Value *V, const Type *Ty) {
65 Instruction::CastOps Op = CastInst::getCastOpcode(V, false, Ty, false);
66 assert((Op == Instruction::BitCast ||
67 Op == Instruction::PtrToInt ||
68 Op == Instruction::IntToPtr) &&
69 "InsertNoopCastOfTo cannot perform non-noop casts!");
70 assert(SE.getTypeSizeInBits(V->getType()) == SE.getTypeSizeInBits(Ty) &&
71 "InsertNoopCastOfTo cannot change sizes!");
73 // Short-circuit unnecessary bitcasts.
74 if (Op == Instruction::BitCast && V->getType() == Ty)
77 // Short-circuit unnecessary inttoptr<->ptrtoint casts.
78 if ((Op == Instruction::PtrToInt || Op == Instruction::IntToPtr) &&
79 SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(V->getType())) {
80 if (CastInst *CI = dyn_cast<CastInst>(V))
81 if ((CI->getOpcode() == Instruction::PtrToInt ||
82 CI->getOpcode() == Instruction::IntToPtr) &&
83 SE.getTypeSizeInBits(CI->getType()) ==
84 SE.getTypeSizeInBits(CI->getOperand(0)->getType()))
85 return CI->getOperand(0);
86 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
87 if ((CE->getOpcode() == Instruction::PtrToInt ||
88 CE->getOpcode() == Instruction::IntToPtr) &&
89 SE.getTypeSizeInBits(CE->getType()) ==
90 SE.getTypeSizeInBits(CE->getOperand(0)->getType()))
91 return CE->getOperand(0);
94 // Fold a cast of a constant.
95 if (Constant *C = dyn_cast<Constant>(V))
96 return ConstantExpr::getCast(Op, C, Ty);
98 // Cast the argument at the beginning of the entry block, after
99 // any bitcasts of other arguments.
100 if (Argument *A = dyn_cast<Argument>(V)) {
101 BasicBlock::iterator IP = A->getParent()->getEntryBlock().begin();
102 while ((isa<BitCastInst>(IP) &&
103 isa<Argument>(cast<BitCastInst>(IP)->getOperand(0)) &&
104 cast<BitCastInst>(IP)->getOperand(0) != A) ||
105 isa<DbgInfoIntrinsic>(IP))
107 return ReuseOrCreateCast(A, Ty, Op, IP);
110 // Cast the instruction immediately after the instruction.
111 Instruction *I = cast<Instruction>(V);
112 BasicBlock::iterator IP = I; ++IP;
113 if (InvokeInst *II = dyn_cast<InvokeInst>(I))
114 IP = II->getNormalDest()->begin();
115 while (isa<PHINode>(IP) || isa<DbgInfoIntrinsic>(IP)) ++IP;
116 return ReuseOrCreateCast(I, Ty, Op, IP);
119 /// InsertBinop - Insert the specified binary operator, doing a small amount
120 /// of work to avoid inserting an obviously redundant operation.
121 Value *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode,
122 Value *LHS, Value *RHS) {
123 // Fold a binop with constant operands.
124 if (Constant *CLHS = dyn_cast<Constant>(LHS))
125 if (Constant *CRHS = dyn_cast<Constant>(RHS))
126 return ConstantExpr::get(Opcode, CLHS, CRHS);
128 // Do a quick scan to see if we have this binop nearby. If so, reuse it.
129 unsigned ScanLimit = 6;
130 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
131 // Scanning starts from the last instruction before the insertion point.
132 BasicBlock::iterator IP = Builder.GetInsertPoint();
133 if (IP != BlockBegin) {
135 for (; ScanLimit; --IP, --ScanLimit) {
136 // Don't count dbg.value against the ScanLimit, to avoid perturbing the
138 if (isa<DbgInfoIntrinsic>(IP))
140 if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(0) == LHS &&
141 IP->getOperand(1) == RHS)
143 if (IP == BlockBegin) break;
147 // Save the original insertion point so we can restore it when we're done.
148 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
149 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
151 // Move the insertion point out of as many loops as we can.
152 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
153 if (!L->isLoopInvariant(LHS) || !L->isLoopInvariant(RHS)) break;
154 BasicBlock *Preheader = L->getLoopPreheader();
155 if (!Preheader) break;
157 // Ok, move up a level.
158 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
161 // If we haven't found this binop, insert it.
162 Value *BO = Builder.CreateBinOp(Opcode, LHS, RHS, "tmp");
163 rememberInstruction(BO);
165 // Restore the original insert point.
167 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
172 /// FactorOutConstant - Test if S is divisible by Factor, using signed
173 /// division. If so, update S with Factor divided out and return true.
174 /// S need not be evenly divisible if a reasonable remainder can be
176 /// TODO: When ScalarEvolution gets a SCEVSDivExpr, this can be made
177 /// unnecessary; in its place, just signed-divide Ops[i] by the scale and
178 /// check to see if the divide was folded.
179 static bool FactorOutConstant(const SCEV *&S,
180 const SCEV *&Remainder,
183 const TargetData *TD) {
184 // Everything is divisible by one.
190 S = SE.getConstant(S->getType(), 1);
194 // For a Constant, check for a multiple of the given factor.
195 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
199 // Check for divisibility.
200 if (const SCEVConstant *FC = dyn_cast<SCEVConstant>(Factor)) {
202 ConstantInt::get(SE.getContext(),
203 C->getValue()->getValue().sdiv(
204 FC->getValue()->getValue()));
205 // If the quotient is zero and the remainder is non-zero, reject
206 // the value at this scale. It will be considered for subsequent
209 const SCEV *Div = SE.getConstant(CI);
212 SE.getAddExpr(Remainder,
213 SE.getConstant(C->getValue()->getValue().srem(
214 FC->getValue()->getValue())));
220 // In a Mul, check if there is a constant operand which is a multiple
221 // of the given factor.
222 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
224 // With TargetData, the size is known. Check if there is a constant
225 // operand which is a multiple of the given factor. If so, we can
227 const SCEVConstant *FC = cast<SCEVConstant>(Factor);
228 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0)))
229 if (!C->getValue()->getValue().srem(FC->getValue()->getValue())) {
230 SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end());
232 SE.getConstant(C->getValue()->getValue().sdiv(
233 FC->getValue()->getValue()));
234 S = SE.getMulExpr(NewMulOps);
238 // Without TargetData, check if Factor can be factored out of any of the
239 // Mul's operands. If so, we can just remove it.
240 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) {
241 const SCEV *SOp = M->getOperand(i);
242 const SCEV *Remainder = SE.getConstant(SOp->getType(), 0);
243 if (FactorOutConstant(SOp, Remainder, Factor, SE, TD) &&
244 Remainder->isZero()) {
245 SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end());
247 S = SE.getMulExpr(NewMulOps);
254 // In an AddRec, check if both start and step are divisible.
255 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
256 const SCEV *Step = A->getStepRecurrence(SE);
257 const SCEV *StepRem = SE.getConstant(Step->getType(), 0);
258 if (!FactorOutConstant(Step, StepRem, Factor, SE, TD))
260 if (!StepRem->isZero())
262 const SCEV *Start = A->getStart();
263 if (!FactorOutConstant(Start, Remainder, Factor, SE, TD))
265 // FIXME: can use A->getNoWrapFlags(FlagNW)
266 S = SE.getAddRecExpr(Start, Step, A->getLoop(), SCEV::FlagAnyWrap);
273 /// SimplifyAddOperands - Sort and simplify a list of add operands. NumAddRecs
274 /// is the number of SCEVAddRecExprs present, which are kept at the end of
277 static void SimplifyAddOperands(SmallVectorImpl<const SCEV *> &Ops,
279 ScalarEvolution &SE) {
280 unsigned NumAddRecs = 0;
281 for (unsigned i = Ops.size(); i > 0 && isa<SCEVAddRecExpr>(Ops[i-1]); --i)
283 // Group Ops into non-addrecs and addrecs.
284 SmallVector<const SCEV *, 8> NoAddRecs(Ops.begin(), Ops.end() - NumAddRecs);
285 SmallVector<const SCEV *, 8> AddRecs(Ops.end() - NumAddRecs, Ops.end());
286 // Let ScalarEvolution sort and simplify the non-addrecs list.
287 const SCEV *Sum = NoAddRecs.empty() ?
288 SE.getConstant(Ty, 0) :
289 SE.getAddExpr(NoAddRecs);
290 // If it returned an add, use the operands. Otherwise it simplified
291 // the sum into a single value, so just use that.
293 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Sum))
294 Ops.append(Add->op_begin(), Add->op_end());
295 else if (!Sum->isZero())
297 // Then append the addrecs.
298 Ops.append(AddRecs.begin(), AddRecs.end());
301 /// SplitAddRecs - Flatten a list of add operands, moving addrec start values
302 /// out to the top level. For example, convert {a + b,+,c} to a, b, {0,+,d}.
303 /// This helps expose more opportunities for folding parts of the expressions
304 /// into GEP indices.
306 static void SplitAddRecs(SmallVectorImpl<const SCEV *> &Ops,
308 ScalarEvolution &SE) {
310 SmallVector<const SCEV *, 8> AddRecs;
311 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
312 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Ops[i])) {
313 const SCEV *Start = A->getStart();
314 if (Start->isZero()) break;
315 const SCEV *Zero = SE.getConstant(Ty, 0);
316 AddRecs.push_back(SE.getAddRecExpr(Zero,
317 A->getStepRecurrence(SE),
319 // FIXME: A->getNoWrapFlags(FlagNW)
321 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Start)) {
323 Ops.append(Add->op_begin(), Add->op_end());
324 e += Add->getNumOperands();
329 if (!AddRecs.empty()) {
330 // Add the addrecs onto the end of the list.
331 Ops.append(AddRecs.begin(), AddRecs.end());
332 // Resort the operand list, moving any constants to the front.
333 SimplifyAddOperands(Ops, Ty, SE);
337 /// expandAddToGEP - Expand an addition expression with a pointer type into
338 /// a GEP instead of using ptrtoint+arithmetic+inttoptr. This helps
339 /// BasicAliasAnalysis and other passes analyze the result. See the rules
340 /// for getelementptr vs. inttoptr in
341 /// http://llvm.org/docs/LangRef.html#pointeraliasing
344 /// Design note: The correctness of using getelementptr here depends on
345 /// ScalarEvolution not recognizing inttoptr and ptrtoint operators, as
346 /// they may introduce pointer arithmetic which may not be safely converted
347 /// into getelementptr.
349 /// Design note: It might seem desirable for this function to be more
350 /// loop-aware. If some of the indices are loop-invariant while others
351 /// aren't, it might seem desirable to emit multiple GEPs, keeping the
352 /// loop-invariant portions of the overall computation outside the loop.
353 /// However, there are a few reasons this is not done here. Hoisting simple
354 /// arithmetic is a low-level optimization that often isn't very
355 /// important until late in the optimization process. In fact, passes
356 /// like InstructionCombining will combine GEPs, even if it means
357 /// pushing loop-invariant computation down into loops, so even if the
358 /// GEPs were split here, the work would quickly be undone. The
359 /// LoopStrengthReduction pass, which is usually run quite late (and
360 /// after the last InstructionCombining pass), takes care of hoisting
361 /// loop-invariant portions of expressions, after considering what
362 /// can be folded using target addressing modes.
364 Value *SCEVExpander::expandAddToGEP(const SCEV *const *op_begin,
365 const SCEV *const *op_end,
366 const PointerType *PTy,
369 const Type *ElTy = PTy->getElementType();
370 SmallVector<Value *, 4> GepIndices;
371 SmallVector<const SCEV *, 8> Ops(op_begin, op_end);
372 bool AnyNonZeroIndices = false;
374 // Split AddRecs up into parts as either of the parts may be usable
375 // without the other.
376 SplitAddRecs(Ops, Ty, SE);
378 // Descend down the pointer's type and attempt to convert the other
379 // operands into GEP indices, at each level. The first index in a GEP
380 // indexes into the array implied by the pointer operand; the rest of
381 // the indices index into the element or field type selected by the
384 // If the scale size is not 0, attempt to factor out a scale for
386 SmallVector<const SCEV *, 8> ScaledOps;
387 if (ElTy->isSized()) {
388 const SCEV *ElSize = SE.getSizeOfExpr(ElTy);
389 if (!ElSize->isZero()) {
390 SmallVector<const SCEV *, 8> NewOps;
391 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
392 const SCEV *Op = Ops[i];
393 const SCEV *Remainder = SE.getConstant(Ty, 0);
394 if (FactorOutConstant(Op, Remainder, ElSize, SE, SE.TD)) {
395 // Op now has ElSize factored out.
396 ScaledOps.push_back(Op);
397 if (!Remainder->isZero())
398 NewOps.push_back(Remainder);
399 AnyNonZeroIndices = true;
401 // The operand was not divisible, so add it to the list of operands
402 // we'll scan next iteration.
403 NewOps.push_back(Ops[i]);
406 // If we made any changes, update Ops.
407 if (!ScaledOps.empty()) {
409 SimplifyAddOperands(Ops, Ty, SE);
414 // Record the scaled array index for this level of the type. If
415 // we didn't find any operands that could be factored, tentatively
416 // assume that element zero was selected (since the zero offset
417 // would obviously be folded away).
418 Value *Scaled = ScaledOps.empty() ?
419 Constant::getNullValue(Ty) :
420 expandCodeFor(SE.getAddExpr(ScaledOps), Ty);
421 GepIndices.push_back(Scaled);
423 // Collect struct field index operands.
424 while (const StructType *STy = dyn_cast<StructType>(ElTy)) {
425 bool FoundFieldNo = false;
426 // An empty struct has no fields.
427 if (STy->getNumElements() == 0) break;
429 // With TargetData, field offsets are known. See if a constant offset
430 // falls within any of the struct fields.
431 if (Ops.empty()) break;
432 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[0]))
433 if (SE.getTypeSizeInBits(C->getType()) <= 64) {
434 const StructLayout &SL = *SE.TD->getStructLayout(STy);
435 uint64_t FullOffset = C->getValue()->getZExtValue();
436 if (FullOffset < SL.getSizeInBytes()) {
437 unsigned ElIdx = SL.getElementContainingOffset(FullOffset);
438 GepIndices.push_back(
439 ConstantInt::get(Type::getInt32Ty(Ty->getContext()), ElIdx));
440 ElTy = STy->getTypeAtIndex(ElIdx);
442 SE.getConstant(Ty, FullOffset - SL.getElementOffset(ElIdx));
443 AnyNonZeroIndices = true;
448 // Without TargetData, just check for an offsetof expression of the
449 // appropriate struct type.
450 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
451 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Ops[i])) {
454 if (U->isOffsetOf(CTy, FieldNo) && CTy == STy) {
455 GepIndices.push_back(FieldNo);
457 STy->getTypeAtIndex(cast<ConstantInt>(FieldNo)->getZExtValue());
458 Ops[i] = SE.getConstant(Ty, 0);
459 AnyNonZeroIndices = true;
465 // If no struct field offsets were found, tentatively assume that
466 // field zero was selected (since the zero offset would obviously
469 ElTy = STy->getTypeAtIndex(0u);
470 GepIndices.push_back(
471 Constant::getNullValue(Type::getInt32Ty(Ty->getContext())));
475 if (const ArrayType *ATy = dyn_cast<ArrayType>(ElTy))
476 ElTy = ATy->getElementType();
481 // If none of the operands were convertible to proper GEP indices, cast
482 // the base to i8* and do an ugly getelementptr with that. It's still
483 // better than ptrtoint+arithmetic+inttoptr at least.
484 if (!AnyNonZeroIndices) {
485 // Cast the base to i8*.
486 V = InsertNoopCastOfTo(V,
487 Type::getInt8PtrTy(Ty->getContext(), PTy->getAddressSpace()));
489 // Expand the operands for a plain byte offset.
490 Value *Idx = expandCodeFor(SE.getAddExpr(Ops), Ty);
492 // Fold a GEP with constant operands.
493 if (Constant *CLHS = dyn_cast<Constant>(V))
494 if (Constant *CRHS = dyn_cast<Constant>(Idx))
495 return ConstantExpr::getGetElementPtr(CLHS, &CRHS, 1);
497 // Do a quick scan to see if we have this GEP nearby. If so, reuse it.
498 unsigned ScanLimit = 6;
499 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
500 // Scanning starts from the last instruction before the insertion point.
501 BasicBlock::iterator IP = Builder.GetInsertPoint();
502 if (IP != BlockBegin) {
504 for (; ScanLimit; --IP, --ScanLimit) {
505 // Don't count dbg.value against the ScanLimit, to avoid perturbing the
507 if (isa<DbgInfoIntrinsic>(IP))
509 if (IP->getOpcode() == Instruction::GetElementPtr &&
510 IP->getOperand(0) == V && IP->getOperand(1) == Idx)
512 if (IP == BlockBegin) break;
516 // Save the original insertion point so we can restore it when we're done.
517 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
518 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
520 // Move the insertion point out of as many loops as we can.
521 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
522 if (!L->isLoopInvariant(V) || !L->isLoopInvariant(Idx)) break;
523 BasicBlock *Preheader = L->getLoopPreheader();
524 if (!Preheader) break;
526 // Ok, move up a level.
527 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
531 Value *GEP = Builder.CreateGEP(V, Idx, "uglygep");
532 rememberInstruction(GEP);
534 // Restore the original insert point.
536 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
541 // Save the original insertion point so we can restore it when we're done.
542 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
543 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
545 // Move the insertion point out of as many loops as we can.
546 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
547 if (!L->isLoopInvariant(V)) break;
549 bool AnyIndexNotLoopInvariant = false;
550 for (SmallVectorImpl<Value *>::const_iterator I = GepIndices.begin(),
551 E = GepIndices.end(); I != E; ++I)
552 if (!L->isLoopInvariant(*I)) {
553 AnyIndexNotLoopInvariant = true;
556 if (AnyIndexNotLoopInvariant)
559 BasicBlock *Preheader = L->getLoopPreheader();
560 if (!Preheader) break;
562 // Ok, move up a level.
563 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
566 // Insert a pretty getelementptr. Note that this GEP is not marked inbounds,
567 // because ScalarEvolution may have changed the address arithmetic to
568 // compute a value which is beyond the end of the allocated object.
570 if (V->getType() != PTy)
571 Casted = InsertNoopCastOfTo(Casted, PTy);
572 Value *GEP = Builder.CreateGEP(Casted,
576 Ops.push_back(SE.getUnknown(GEP));
577 rememberInstruction(GEP);
579 // Restore the original insert point.
581 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
583 return expand(SE.getAddExpr(Ops));
586 /// isNonConstantNegative - Return true if the specified scev is negated, but
588 static bool isNonConstantNegative(const SCEV *F) {
589 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(F);
590 if (!Mul) return false;
592 // If there is a constant factor, it will be first.
593 const SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
594 if (!SC) return false;
596 // Return true if the value is negative, this matches things like (-42 * V).
597 return SC->getValue()->getValue().isNegative();
600 /// PickMostRelevantLoop - Given two loops pick the one that's most relevant for
601 /// SCEV expansion. If they are nested, this is the most nested. If they are
602 /// neighboring, pick the later.
603 static const Loop *PickMostRelevantLoop(const Loop *A, const Loop *B,
607 if (A->contains(B)) return B;
608 if (B->contains(A)) return A;
609 if (DT.dominates(A->getHeader(), B->getHeader())) return B;
610 if (DT.dominates(B->getHeader(), A->getHeader())) return A;
611 return A; // Arbitrarily break the tie.
614 /// getRelevantLoop - Get the most relevant loop associated with the given
615 /// expression, according to PickMostRelevantLoop.
616 const Loop *SCEVExpander::getRelevantLoop(const SCEV *S) {
617 // Test whether we've already computed the most relevant loop for this SCEV.
618 std::pair<DenseMap<const SCEV *, const Loop *>::iterator, bool> Pair =
619 RelevantLoops.insert(std::make_pair(S, static_cast<const Loop *>(0)));
621 return Pair.first->second;
623 if (isa<SCEVConstant>(S))
624 // A constant has no relevant loops.
626 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
627 if (const Instruction *I = dyn_cast<Instruction>(U->getValue()))
628 return Pair.first->second = SE.LI->getLoopFor(I->getParent());
629 // A non-instruction has no relevant loops.
632 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S)) {
634 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
636 for (SCEVNAryExpr::op_iterator I = N->op_begin(), E = N->op_end();
638 L = PickMostRelevantLoop(L, getRelevantLoop(*I), *SE.DT);
639 return RelevantLoops[N] = L;
641 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S)) {
642 const Loop *Result = getRelevantLoop(C->getOperand());
643 return RelevantLoops[C] = Result;
645 if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
647 PickMostRelevantLoop(getRelevantLoop(D->getLHS()),
648 getRelevantLoop(D->getRHS()),
650 return RelevantLoops[D] = Result;
652 llvm_unreachable("Unexpected SCEV type!");
658 /// LoopCompare - Compare loops by PickMostRelevantLoop.
662 explicit LoopCompare(DominatorTree &dt) : DT(dt) {}
664 bool operator()(std::pair<const Loop *, const SCEV *> LHS,
665 std::pair<const Loop *, const SCEV *> RHS) const {
666 // Keep pointer operands sorted at the end.
667 if (LHS.second->getType()->isPointerTy() !=
668 RHS.second->getType()->isPointerTy())
669 return LHS.second->getType()->isPointerTy();
671 // Compare loops with PickMostRelevantLoop.
672 if (LHS.first != RHS.first)
673 return PickMostRelevantLoop(LHS.first, RHS.first, DT) != LHS.first;
675 // If one operand is a non-constant negative and the other is not,
676 // put the non-constant negative on the right so that a sub can
677 // be used instead of a negate and add.
678 if (isNonConstantNegative(LHS.second)) {
679 if (!isNonConstantNegative(RHS.second))
681 } else if (isNonConstantNegative(RHS.second))
684 // Otherwise they are equivalent according to this comparison.
691 Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) {
692 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
694 // Collect all the add operands in a loop, along with their associated loops.
695 // Iterate in reverse so that constants are emitted last, all else equal, and
696 // so that pointer operands are inserted first, which the code below relies on
697 // to form more involved GEPs.
698 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops;
699 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(S->op_end()),
700 E(S->op_begin()); I != E; ++I)
701 OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I));
703 // Sort by loop. Use a stable sort so that constants follow non-constants and
704 // pointer operands precede non-pointer operands.
705 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT));
707 // Emit instructions to add all the operands. Hoist as much as possible
708 // out of loops, and form meaningful getelementptrs where possible.
710 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator
711 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) {
712 const Loop *CurLoop = I->first;
713 const SCEV *Op = I->second;
715 // This is the first operand. Just expand it.
718 } else if (const PointerType *PTy = dyn_cast<PointerType>(Sum->getType())) {
719 // The running sum expression is a pointer. Try to form a getelementptr
720 // at this level with that as the base.
721 SmallVector<const SCEV *, 4> NewOps;
722 for (; I != E && I->first == CurLoop; ++I) {
723 // If the operand is SCEVUnknown and not instructions, peek through
724 // it, to enable more of it to be folded into the GEP.
725 const SCEV *X = I->second;
726 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(X))
727 if (!isa<Instruction>(U->getValue()))
728 X = SE.getSCEV(U->getValue());
731 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, Sum);
732 } else if (const PointerType *PTy = dyn_cast<PointerType>(Op->getType())) {
733 // The running sum is an integer, and there's a pointer at this level.
734 // Try to form a getelementptr. If the running sum is instructions,
735 // use a SCEVUnknown to avoid re-analyzing them.
736 SmallVector<const SCEV *, 4> NewOps;
737 NewOps.push_back(isa<Instruction>(Sum) ? SE.getUnknown(Sum) :
739 for (++I; I != E && I->first == CurLoop; ++I)
740 NewOps.push_back(I->second);
741 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, expand(Op));
742 } else if (isNonConstantNegative(Op)) {
743 // Instead of doing a negate and add, just do a subtract.
744 Value *W = expandCodeFor(SE.getNegativeSCEV(Op), Ty);
745 Sum = InsertNoopCastOfTo(Sum, Ty);
746 Sum = InsertBinop(Instruction::Sub, Sum, W);
750 Value *W = expandCodeFor(Op, Ty);
751 Sum = InsertNoopCastOfTo(Sum, Ty);
752 // Canonicalize a constant to the RHS.
753 if (isa<Constant>(Sum)) std::swap(Sum, W);
754 Sum = InsertBinop(Instruction::Add, Sum, W);
762 Value *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) {
763 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
765 // Collect all the mul operands in a loop, along with their associated loops.
766 // Iterate in reverse so that constants are emitted last, all else equal.
767 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops;
768 for (std::reverse_iterator<SCEVMulExpr::op_iterator> I(S->op_end()),
769 E(S->op_begin()); I != E; ++I)
770 OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I));
772 // Sort by loop. Use a stable sort so that constants follow non-constants.
773 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT));
775 // Emit instructions to mul all the operands. Hoist as much as possible
778 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator
779 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) {
780 const SCEV *Op = I->second;
782 // This is the first operand. Just expand it.
785 } else if (Op->isAllOnesValue()) {
786 // Instead of doing a multiply by negative one, just do a negate.
787 Prod = InsertNoopCastOfTo(Prod, Ty);
788 Prod = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), Prod);
792 Value *W = expandCodeFor(Op, Ty);
793 Prod = InsertNoopCastOfTo(Prod, Ty);
794 // Canonicalize a constant to the RHS.
795 if (isa<Constant>(Prod)) std::swap(Prod, W);
796 Prod = InsertBinop(Instruction::Mul, Prod, W);
804 Value *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) {
805 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
807 Value *LHS = expandCodeFor(S->getLHS(), Ty);
808 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getRHS())) {
809 const APInt &RHS = SC->getValue()->getValue();
810 if (RHS.isPowerOf2())
811 return InsertBinop(Instruction::LShr, LHS,
812 ConstantInt::get(Ty, RHS.logBase2()));
815 Value *RHS = expandCodeFor(S->getRHS(), Ty);
816 return InsertBinop(Instruction::UDiv, LHS, RHS);
819 /// Move parts of Base into Rest to leave Base with the minimal
820 /// expression that provides a pointer operand suitable for a
822 static void ExposePointerBase(const SCEV *&Base, const SCEV *&Rest,
823 ScalarEvolution &SE) {
824 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Base)) {
825 Base = A->getStart();
826 Rest = SE.getAddExpr(Rest,
827 SE.getAddRecExpr(SE.getConstant(A->getType(), 0),
828 A->getStepRecurrence(SE),
830 // FIXME: A->getNoWrapFlags(FlagNW)
833 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(Base)) {
834 Base = A->getOperand(A->getNumOperands()-1);
835 SmallVector<const SCEV *, 8> NewAddOps(A->op_begin(), A->op_end());
836 NewAddOps.back() = Rest;
837 Rest = SE.getAddExpr(NewAddOps);
838 ExposePointerBase(Base, Rest, SE);
842 /// getAddRecExprPHILiterally - Helper for expandAddRecExprLiterally. Expand
843 /// the base addrec, which is the addrec without any non-loop-dominating
844 /// values, and return the PHI.
846 SCEVExpander::getAddRecExprPHILiterally(const SCEVAddRecExpr *Normalized,
848 const Type *ExpandTy,
850 // Reuse a previously-inserted PHI, if present.
851 for (BasicBlock::iterator I = L->getHeader()->begin();
852 PHINode *PN = dyn_cast<PHINode>(I); ++I)
853 if (SE.isSCEVable(PN->getType()) &&
854 (SE.getEffectiveSCEVType(PN->getType()) ==
855 SE.getEffectiveSCEVType(Normalized->getType())) &&
856 SE.getSCEV(PN) == Normalized)
857 if (BasicBlock *LatchBlock = L->getLoopLatch()) {
859 cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock));
861 // Determine if this is a well-behaved chain of instructions leading
862 // back to the PHI. It probably will be, if we're scanning an inner
863 // loop already visited by LSR for example, but it wouldn't have
866 if (IncV->getNumOperands() == 0 || isa<PHINode>(IncV) ||
867 (isa<CastInst>(IncV) && !isa<BitCastInst>(IncV))) {
871 // If any of the operands don't dominate the insert position, bail.
872 // Addrec operands are always loop-invariant, so this can only happen
873 // if there are instructions which haven't been hoisted.
874 for (User::op_iterator OI = IncV->op_begin()+1,
875 OE = IncV->op_end(); OI != OE; ++OI)
876 if (Instruction *OInst = dyn_cast<Instruction>(OI))
877 if (!SE.DT->dominates(OInst, IVIncInsertPos)) {
883 // Advance to the next instruction.
884 IncV = dyn_cast<Instruction>(IncV->getOperand(0));
887 if (IncV->mayHaveSideEffects()) {
891 } while (IncV != PN);
894 // Ok, the add recurrence looks usable.
895 // Remember this PHI, even in post-inc mode.
896 InsertedValues.insert(PN);
897 // Remember the increment.
898 IncV = cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock));
899 rememberInstruction(IncV);
900 if (L == IVIncInsertLoop)
902 if (SE.DT->dominates(IncV, IVIncInsertPos))
904 // Make sure the increment is where we want it. But don't move it
905 // down past a potential existing post-inc user.
906 IncV->moveBefore(IVIncInsertPos);
907 IVIncInsertPos = IncV;
908 IncV = cast<Instruction>(IncV->getOperand(0));
909 } while (IncV != PN);
914 // Save the original insertion point so we can restore it when we're done.
915 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
916 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
918 // Expand code for the start value.
919 Value *StartV = expandCodeFor(Normalized->getStart(), ExpandTy,
920 L->getHeader()->begin());
922 // Expand code for the step value. Insert instructions right before the
923 // terminator corresponding to the back-edge. Do this before creating the PHI
924 // so that PHI reuse code doesn't see an incomplete PHI. If the stride is
925 // negative, insert a sub instead of an add for the increment (unless it's a
926 // constant, because subtracts of constants are canonicalized to adds).
927 const SCEV *Step = Normalized->getStepRecurrence(SE);
928 bool isPointer = ExpandTy->isPointerTy();
929 bool isNegative = !isPointer && isNonConstantNegative(Step);
931 Step = SE.getNegativeSCEV(Step);
932 Value *StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin());
935 Builder.SetInsertPoint(L->getHeader(), L->getHeader()->begin());
936 PHINode *PN = Builder.CreatePHI(ExpandTy, "lsr.iv");
937 rememberInstruction(PN);
939 // Create the step instructions and populate the PHI.
940 BasicBlock *Header = L->getHeader();
941 for (pred_iterator HPI = pred_begin(Header), HPE = pred_end(Header);
943 BasicBlock *Pred = *HPI;
945 // Add a start value.
946 if (!L->contains(Pred)) {
947 PN->addIncoming(StartV, Pred);
951 // Create a step value and add it to the PHI. If IVIncInsertLoop is
952 // non-null and equal to the addrec's loop, insert the instructions
953 // at IVIncInsertPos.
954 Instruction *InsertPos = L == IVIncInsertLoop ?
955 IVIncInsertPos : Pred->getTerminator();
956 Builder.SetInsertPoint(InsertPos->getParent(), InsertPos);
958 // If the PHI is a pointer, use a GEP, otherwise use an add or sub.
960 const PointerType *GEPPtrTy = cast<PointerType>(ExpandTy);
961 // If the step isn't constant, don't use an implicitly scaled GEP, because
962 // that would require a multiply inside the loop.
963 if (!isa<ConstantInt>(StepV))
964 GEPPtrTy = PointerType::get(Type::getInt1Ty(SE.getContext()),
965 GEPPtrTy->getAddressSpace());
966 const SCEV *const StepArray[1] = { SE.getSCEV(StepV) };
967 IncV = expandAddToGEP(StepArray, StepArray+1, GEPPtrTy, IntTy, PN);
968 if (IncV->getType() != PN->getType()) {
969 IncV = Builder.CreateBitCast(IncV, PN->getType(), "tmp");
970 rememberInstruction(IncV);
974 Builder.CreateSub(PN, StepV, "lsr.iv.next") :
975 Builder.CreateAdd(PN, StepV, "lsr.iv.next");
976 rememberInstruction(IncV);
978 PN->addIncoming(IncV, Pred);
981 // Restore the original insert point.
983 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
985 // Remember this PHI, even in post-inc mode.
986 InsertedValues.insert(PN);
991 Value *SCEVExpander::expandAddRecExprLiterally(const SCEVAddRecExpr *S) {
992 const Type *STy = S->getType();
993 const Type *IntTy = SE.getEffectiveSCEVType(STy);
994 const Loop *L = S->getLoop();
996 // Determine a normalized form of this expression, which is the expression
997 // before any post-inc adjustment is made.
998 const SCEVAddRecExpr *Normalized = S;
999 if (PostIncLoops.count(L)) {
1000 PostIncLoopSet Loops;
1003 cast<SCEVAddRecExpr>(TransformForPostIncUse(Normalize, S, 0, 0,
1004 Loops, SE, *SE.DT));
1007 // Strip off any non-loop-dominating component from the addrec start.
1008 const SCEV *Start = Normalized->getStart();
1009 const SCEV *PostLoopOffset = 0;
1010 if (!SE.properlyDominates(Start, L->getHeader())) {
1011 PostLoopOffset = Start;
1012 Start = SE.getConstant(Normalized->getType(), 0);
1013 Normalized = cast<SCEVAddRecExpr>(
1014 SE.getAddRecExpr(Start, Normalized->getStepRecurrence(SE),
1015 Normalized->getLoop(),
1016 // FIXME: Normalized->getNoWrapFlags(FlagNW)
1017 SCEV::FlagAnyWrap));
1020 // Strip off any non-loop-dominating component from the addrec step.
1021 const SCEV *Step = Normalized->getStepRecurrence(SE);
1022 const SCEV *PostLoopScale = 0;
1023 if (!SE.dominates(Step, L->getHeader())) {
1024 PostLoopScale = Step;
1025 Step = SE.getConstant(Normalized->getType(), 1);
1027 cast<SCEVAddRecExpr>(SE.getAddRecExpr(Start, Step,
1028 Normalized->getLoop(),
1029 // FIXME: Normalized
1030 // ->getNoWrapFlags(FlagNW)
1031 SCEV::FlagAnyWrap));
1034 // Expand the core addrec. If we need post-loop scaling, force it to
1035 // expand to an integer type to avoid the need for additional casting.
1036 const Type *ExpandTy = PostLoopScale ? IntTy : STy;
1037 PHINode *PN = getAddRecExprPHILiterally(Normalized, L, ExpandTy, IntTy);
1039 // Accommodate post-inc mode, if necessary.
1041 if (!PostIncLoops.count(L))
1044 // In PostInc mode, use the post-incremented value.
1045 BasicBlock *LatchBlock = L->getLoopLatch();
1046 assert(LatchBlock && "PostInc mode requires a unique loop latch!");
1047 Result = PN->getIncomingValueForBlock(LatchBlock);
1050 // Re-apply any non-loop-dominating scale.
1051 if (PostLoopScale) {
1052 Result = InsertNoopCastOfTo(Result, IntTy);
1053 Result = Builder.CreateMul(Result,
1054 expandCodeFor(PostLoopScale, IntTy));
1055 rememberInstruction(Result);
1058 // Re-apply any non-loop-dominating offset.
1059 if (PostLoopOffset) {
1060 if (const PointerType *PTy = dyn_cast<PointerType>(ExpandTy)) {
1061 const SCEV *const OffsetArray[1] = { PostLoopOffset };
1062 Result = expandAddToGEP(OffsetArray, OffsetArray+1, PTy, IntTy, Result);
1064 Result = InsertNoopCastOfTo(Result, IntTy);
1065 Result = Builder.CreateAdd(Result,
1066 expandCodeFor(PostLoopOffset, IntTy));
1067 rememberInstruction(Result);
1074 Value *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) {
1075 if (!CanonicalMode) return expandAddRecExprLiterally(S);
1077 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
1078 const Loop *L = S->getLoop();
1080 // First check for an existing canonical IV in a suitable type.
1081 PHINode *CanonicalIV = 0;
1082 if (PHINode *PN = L->getCanonicalInductionVariable())
1083 if (SE.getTypeSizeInBits(PN->getType()) >= SE.getTypeSizeInBits(Ty))
1086 // Rewrite an AddRec in terms of the canonical induction variable, if
1087 // its type is more narrow.
1089 SE.getTypeSizeInBits(CanonicalIV->getType()) >
1090 SE.getTypeSizeInBits(Ty)) {
1091 SmallVector<const SCEV *, 4> NewOps(S->getNumOperands());
1092 for (unsigned i = 0, e = S->getNumOperands(); i != e; ++i)
1093 NewOps[i] = SE.getAnyExtendExpr(S->op_begin()[i], CanonicalIV->getType());
1094 Value *V = expand(SE.getAddRecExpr(NewOps, S->getLoop(),
1095 // FIXME: S->getNoWrapFlags(FlagNW)
1096 SCEV::FlagAnyWrap));
1097 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1098 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1099 BasicBlock::iterator NewInsertPt =
1100 llvm::next(BasicBlock::iterator(cast<Instruction>(V)));
1101 while (isa<PHINode>(NewInsertPt) || isa<DbgInfoIntrinsic>(NewInsertPt))
1103 V = expandCodeFor(SE.getTruncateExpr(SE.getUnknown(V), Ty), 0,
1105 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
1109 // {X,+,F} --> X + {0,+,F}
1110 if (!S->getStart()->isZero()) {
1111 SmallVector<const SCEV *, 4> NewOps(S->op_begin(), S->op_end());
1112 NewOps[0] = SE.getConstant(Ty, 0);
1113 // FIXME: can use S->getNoWrapFlags()
1114 const SCEV *Rest = SE.getAddRecExpr(NewOps, L, SCEV::FlagAnyWrap);
1116 // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the
1117 // comments on expandAddToGEP for details.
1118 const SCEV *Base = S->getStart();
1119 const SCEV *RestArray[1] = { Rest };
1120 // Dig into the expression to find the pointer base for a GEP.
1121 ExposePointerBase(Base, RestArray[0], SE);
1122 // If we found a pointer, expand the AddRec with a GEP.
1123 if (const PointerType *PTy = dyn_cast<PointerType>(Base->getType())) {
1124 // Make sure the Base isn't something exotic, such as a multiplied
1125 // or divided pointer value. In those cases, the result type isn't
1126 // actually a pointer type.
1127 if (!isa<SCEVMulExpr>(Base) && !isa<SCEVUDivExpr>(Base)) {
1128 Value *StartV = expand(Base);
1129 assert(StartV->getType() == PTy && "Pointer type mismatch for GEP!");
1130 return expandAddToGEP(RestArray, RestArray+1, PTy, Ty, StartV);
1134 // Just do a normal add. Pre-expand the operands to suppress folding.
1135 return expand(SE.getAddExpr(SE.getUnknown(expand(S->getStart())),
1136 SE.getUnknown(expand(Rest))));
1139 // If we don't yet have a canonical IV, create one.
1141 // Create and insert the PHI node for the induction variable in the
1143 BasicBlock *Header = L->getHeader();
1144 CanonicalIV = PHINode::Create(Ty, "indvar", Header->begin());
1145 rememberInstruction(CanonicalIV);
1147 Constant *One = ConstantInt::get(Ty, 1);
1148 for (pred_iterator HPI = pred_begin(Header), HPE = pred_end(Header);
1149 HPI != HPE; ++HPI) {
1150 BasicBlock *HP = *HPI;
1151 if (L->contains(HP)) {
1152 // Insert a unit add instruction right before the terminator
1153 // corresponding to the back-edge.
1154 Instruction *Add = BinaryOperator::CreateAdd(CanonicalIV, One,
1156 HP->getTerminator());
1157 rememberInstruction(Add);
1158 CanonicalIV->addIncoming(Add, HP);
1160 CanonicalIV->addIncoming(Constant::getNullValue(Ty), HP);
1165 // {0,+,1} --> Insert a canonical induction variable into the loop!
1166 if (S->isAffine() && S->getOperand(1)->isOne()) {
1167 assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) &&
1168 "IVs with types different from the canonical IV should "
1169 "already have been handled!");
1173 // {0,+,F} --> {0,+,1} * F
1175 // If this is a simple linear addrec, emit it now as a special case.
1176 if (S->isAffine()) // {0,+,F} --> i*F
1178 expand(SE.getTruncateOrNoop(
1179 SE.getMulExpr(SE.getUnknown(CanonicalIV),
1180 SE.getNoopOrAnyExtend(S->getOperand(1),
1181 CanonicalIV->getType())),
1184 // If this is a chain of recurrences, turn it into a closed form, using the
1185 // folders, then expandCodeFor the closed form. This allows the folders to
1186 // simplify the expression without having to build a bunch of special code
1187 // into this folder.
1188 const SCEV *IH = SE.getUnknown(CanonicalIV); // Get I as a "symbolic" SCEV.
1190 // Promote S up to the canonical IV type, if the cast is foldable.
1191 const SCEV *NewS = S;
1192 const SCEV *Ext = SE.getNoopOrAnyExtend(S, CanonicalIV->getType());
1193 if (isa<SCEVAddRecExpr>(Ext))
1196 const SCEV *V = cast<SCEVAddRecExpr>(NewS)->evaluateAtIteration(IH, SE);
1197 //cerr << "Evaluated: " << *this << "\n to: " << *V << "\n";
1199 // Truncate the result down to the original type, if needed.
1200 const SCEV *T = SE.getTruncateOrNoop(V, Ty);
1204 Value *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) {
1205 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
1206 Value *V = expandCodeFor(S->getOperand(),
1207 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1208 Value *I = Builder.CreateTrunc(V, Ty, "tmp");
1209 rememberInstruction(I);
1213 Value *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) {
1214 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
1215 Value *V = expandCodeFor(S->getOperand(),
1216 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1217 Value *I = Builder.CreateZExt(V, Ty, "tmp");
1218 rememberInstruction(I);
1222 Value *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) {
1223 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
1224 Value *V = expandCodeFor(S->getOperand(),
1225 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1226 Value *I = Builder.CreateSExt(V, Ty, "tmp");
1227 rememberInstruction(I);
1231 Value *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) {
1232 Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
1233 const Type *Ty = LHS->getType();
1234 for (int i = S->getNumOperands()-2; i >= 0; --i) {
1235 // In the case of mixed integer and pointer types, do the
1236 // rest of the comparisons as integer.
1237 if (S->getOperand(i)->getType() != Ty) {
1238 Ty = SE.getEffectiveSCEVType(Ty);
1239 LHS = InsertNoopCastOfTo(LHS, Ty);
1241 Value *RHS = expandCodeFor(S->getOperand(i), Ty);
1242 Value *ICmp = Builder.CreateICmpSGT(LHS, RHS, "tmp");
1243 rememberInstruction(ICmp);
1244 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "smax");
1245 rememberInstruction(Sel);
1248 // In the case of mixed integer and pointer types, cast the
1249 // final result back to the pointer type.
1250 if (LHS->getType() != S->getType())
1251 LHS = InsertNoopCastOfTo(LHS, S->getType());
1255 Value *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) {
1256 Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
1257 const Type *Ty = LHS->getType();
1258 for (int i = S->getNumOperands()-2; i >= 0; --i) {
1259 // In the case of mixed integer and pointer types, do the
1260 // rest of the comparisons as integer.
1261 if (S->getOperand(i)->getType() != Ty) {
1262 Ty = SE.getEffectiveSCEVType(Ty);
1263 LHS = InsertNoopCastOfTo(LHS, Ty);
1265 Value *RHS = expandCodeFor(S->getOperand(i), Ty);
1266 Value *ICmp = Builder.CreateICmpUGT(LHS, RHS, "tmp");
1267 rememberInstruction(ICmp);
1268 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "umax");
1269 rememberInstruction(Sel);
1272 // In the case of mixed integer and pointer types, cast the
1273 // final result back to the pointer type.
1274 if (LHS->getType() != S->getType())
1275 LHS = InsertNoopCastOfTo(LHS, S->getType());
1279 Value *SCEVExpander::expandCodeFor(const SCEV *SH, const Type *Ty,
1281 BasicBlock::iterator IP = I;
1282 while (isInsertedInstruction(IP) || isa<DbgInfoIntrinsic>(IP))
1284 Builder.SetInsertPoint(IP->getParent(), IP);
1285 return expandCodeFor(SH, Ty);
1288 Value *SCEVExpander::expandCodeFor(const SCEV *SH, const Type *Ty) {
1289 // Expand the code for this SCEV.
1290 Value *V = expand(SH);
1292 assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) &&
1293 "non-trivial casts should be done with the SCEVs directly!");
1294 V = InsertNoopCastOfTo(V, Ty);
1299 Value *SCEVExpander::expand(const SCEV *S) {
1300 // Compute an insertion point for this SCEV object. Hoist the instructions
1301 // as far out in the loop nest as possible.
1302 Instruction *InsertPt = Builder.GetInsertPoint();
1303 for (Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock()); ;
1304 L = L->getParentLoop())
1305 if (SE.isLoopInvariant(S, L)) {
1307 if (BasicBlock *Preheader = L->getLoopPreheader())
1308 InsertPt = Preheader->getTerminator();
1310 // If the SCEV is computable at this level, insert it into the header
1311 // after the PHIs (and after any other instructions that we've inserted
1312 // there) so that it is guaranteed to dominate any user inside the loop.
1313 if (L && SE.hasComputableLoopEvolution(S, L) && !PostIncLoops.count(L))
1314 InsertPt = L->getHeader()->getFirstNonPHI();
1315 while (isInsertedInstruction(InsertPt) || isa<DbgInfoIntrinsic>(InsertPt))
1316 InsertPt = llvm::next(BasicBlock::iterator(InsertPt));
1320 // Check to see if we already expanded this here.
1321 std::map<std::pair<const SCEV *, Instruction *>,
1322 AssertingVH<Value> >::iterator I =
1323 InsertedExpressions.find(std::make_pair(S, InsertPt));
1324 if (I != InsertedExpressions.end())
1327 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1328 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1329 Builder.SetInsertPoint(InsertPt->getParent(), InsertPt);
1331 // Expand the expression into instructions.
1332 Value *V = visit(S);
1334 // Remember the expanded value for this SCEV at this location.
1335 if (PostIncLoops.empty())
1336 InsertedExpressions[std::make_pair(S, InsertPt)] = V;
1338 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
1342 void SCEVExpander::rememberInstruction(Value *I) {
1343 if (!PostIncLoops.empty())
1344 InsertedPostIncValues.insert(I);
1346 InsertedValues.insert(I);
1348 // If we just claimed an existing instruction and that instruction had
1349 // been the insert point, adjust the insert point forward so that
1350 // subsequently inserted code will be dominated.
1351 if (Builder.GetInsertPoint() == I) {
1352 BasicBlock::iterator It = cast<Instruction>(I);
1353 do { ++It; } while (isInsertedInstruction(It) ||
1354 isa<DbgInfoIntrinsic>(It));
1355 Builder.SetInsertPoint(Builder.GetInsertBlock(), It);
1359 void SCEVExpander::restoreInsertPoint(BasicBlock *BB, BasicBlock::iterator I) {
1360 // If we acquired more instructions since the old insert point was saved,
1361 // advance past them.
1362 while (isInsertedInstruction(I) || isa<DbgInfoIntrinsic>(I)) ++I;
1364 Builder.SetInsertPoint(BB, I);
1367 /// getOrInsertCanonicalInductionVariable - This method returns the
1368 /// canonical induction variable of the specified type for the specified
1369 /// loop (inserting one if there is none). A canonical induction variable
1370 /// starts at zero and steps by one on each iteration.
1372 SCEVExpander::getOrInsertCanonicalInductionVariable(const Loop *L,
1374 assert(Ty->isIntegerTy() && "Can only insert integer induction variables!");
1376 // Build a SCEV for {0,+,1}<L>.
1377 // Conservatively use FlagAnyWrap for now.
1378 const SCEV *H = SE.getAddRecExpr(SE.getConstant(Ty, 0),
1379 SE.getConstant(Ty, 1), L, SCEV::FlagAnyWrap);
1381 // Emit code for it.
1382 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1383 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1384 PHINode *V = cast<PHINode>(expandCodeFor(H, 0, L->getHeader()->begin()));
1386 restoreInsertPoint(SaveInsertBB, SaveInsertPt);