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, 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, 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) ||
107 isa<LandingPadInst>(IP))
109 return ReuseOrCreateCast(A, Ty, Op, IP);
112 // Cast the instruction immediately after the instruction.
113 Instruction *I = cast<Instruction>(V);
114 BasicBlock::iterator IP = I; ++IP;
115 if (InvokeInst *II = dyn_cast<InvokeInst>(I))
116 IP = II->getNormalDest()->begin();
117 while (isa<PHINode>(IP) || isa<DbgInfoIntrinsic>(IP) ||
118 isa<LandingPadInst>(IP))
120 return ReuseOrCreateCast(I, Ty, Op, IP);
123 /// InsertBinop - Insert the specified binary operator, doing a small amount
124 /// of work to avoid inserting an obviously redundant operation.
125 Value *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode,
126 Value *LHS, Value *RHS) {
127 // Fold a binop with constant operands.
128 if (Constant *CLHS = dyn_cast<Constant>(LHS))
129 if (Constant *CRHS = dyn_cast<Constant>(RHS))
130 return ConstantExpr::get(Opcode, CLHS, CRHS);
132 // Do a quick scan to see if we have this binop nearby. If so, reuse it.
133 unsigned ScanLimit = 6;
134 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
135 // Scanning starts from the last instruction before the insertion point.
136 BasicBlock::iterator IP = Builder.GetInsertPoint();
137 if (IP != BlockBegin) {
139 for (; ScanLimit; --IP, --ScanLimit) {
140 // Don't count dbg.value against the ScanLimit, to avoid perturbing the
142 if (isa<DbgInfoIntrinsic>(IP))
144 if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(0) == LHS &&
145 IP->getOperand(1) == RHS)
147 if (IP == BlockBegin) break;
151 // Save the original insertion point so we can restore it when we're done.
152 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
153 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
155 // Move the insertion point out of as many loops as we can.
156 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
157 if (!L->isLoopInvariant(LHS) || !L->isLoopInvariant(RHS)) break;
158 BasicBlock *Preheader = L->getLoopPreheader();
159 if (!Preheader) break;
161 // Ok, move up a level.
162 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
165 // If we haven't found this binop, insert it.
166 Instruction *BO = cast<Instruction>(Builder.CreateBinOp(Opcode, LHS, RHS, "tmp"));
167 BO->setDebugLoc(SaveInsertPt->getDebugLoc());
168 rememberInstruction(BO);
170 // Restore the original insert point.
172 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
177 /// FactorOutConstant - Test if S is divisible by Factor, using signed
178 /// division. If so, update S with Factor divided out and return true.
179 /// S need not be evenly divisible if a reasonable remainder can be
181 /// TODO: When ScalarEvolution gets a SCEVSDivExpr, this can be made
182 /// unnecessary; in its place, just signed-divide Ops[i] by the scale and
183 /// check to see if the divide was folded.
184 static bool FactorOutConstant(const SCEV *&S,
185 const SCEV *&Remainder,
188 const TargetData *TD) {
189 // Everything is divisible by one.
195 S = SE.getConstant(S->getType(), 1);
199 // For a Constant, check for a multiple of the given factor.
200 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
204 // Check for divisibility.
205 if (const SCEVConstant *FC = dyn_cast<SCEVConstant>(Factor)) {
207 ConstantInt::get(SE.getContext(),
208 C->getValue()->getValue().sdiv(
209 FC->getValue()->getValue()));
210 // If the quotient is zero and the remainder is non-zero, reject
211 // the value at this scale. It will be considered for subsequent
214 const SCEV *Div = SE.getConstant(CI);
217 SE.getAddExpr(Remainder,
218 SE.getConstant(C->getValue()->getValue().srem(
219 FC->getValue()->getValue())));
225 // In a Mul, check if there is a constant operand which is a multiple
226 // of the given factor.
227 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
229 // With TargetData, the size is known. Check if there is a constant
230 // operand which is a multiple of the given factor. If so, we can
232 const SCEVConstant *FC = cast<SCEVConstant>(Factor);
233 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0)))
234 if (!C->getValue()->getValue().srem(FC->getValue()->getValue())) {
235 SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end());
237 SE.getConstant(C->getValue()->getValue().sdiv(
238 FC->getValue()->getValue()));
239 S = SE.getMulExpr(NewMulOps);
243 // Without TargetData, check if Factor can be factored out of any of the
244 // Mul's operands. If so, we can just remove it.
245 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) {
246 const SCEV *SOp = M->getOperand(i);
247 const SCEV *Remainder = SE.getConstant(SOp->getType(), 0);
248 if (FactorOutConstant(SOp, Remainder, Factor, SE, TD) &&
249 Remainder->isZero()) {
250 SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end());
252 S = SE.getMulExpr(NewMulOps);
259 // In an AddRec, check if both start and step are divisible.
260 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
261 const SCEV *Step = A->getStepRecurrence(SE);
262 const SCEV *StepRem = SE.getConstant(Step->getType(), 0);
263 if (!FactorOutConstant(Step, StepRem, Factor, SE, TD))
265 if (!StepRem->isZero())
267 const SCEV *Start = A->getStart();
268 if (!FactorOutConstant(Start, Remainder, Factor, SE, TD))
270 // FIXME: can use A->getNoWrapFlags(FlagNW)
271 S = SE.getAddRecExpr(Start, Step, A->getLoop(), SCEV::FlagAnyWrap);
278 /// SimplifyAddOperands - Sort and simplify a list of add operands. NumAddRecs
279 /// is the number of SCEVAddRecExprs present, which are kept at the end of
282 static void SimplifyAddOperands(SmallVectorImpl<const SCEV *> &Ops,
284 ScalarEvolution &SE) {
285 unsigned NumAddRecs = 0;
286 for (unsigned i = Ops.size(); i > 0 && isa<SCEVAddRecExpr>(Ops[i-1]); --i)
288 // Group Ops into non-addrecs and addrecs.
289 SmallVector<const SCEV *, 8> NoAddRecs(Ops.begin(), Ops.end() - NumAddRecs);
290 SmallVector<const SCEV *, 8> AddRecs(Ops.end() - NumAddRecs, Ops.end());
291 // Let ScalarEvolution sort and simplify the non-addrecs list.
292 const SCEV *Sum = NoAddRecs.empty() ?
293 SE.getConstant(Ty, 0) :
294 SE.getAddExpr(NoAddRecs);
295 // If it returned an add, use the operands. Otherwise it simplified
296 // the sum into a single value, so just use that.
298 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Sum))
299 Ops.append(Add->op_begin(), Add->op_end());
300 else if (!Sum->isZero())
302 // Then append the addrecs.
303 Ops.append(AddRecs.begin(), AddRecs.end());
306 /// SplitAddRecs - Flatten a list of add operands, moving addrec start values
307 /// out to the top level. For example, convert {a + b,+,c} to a, b, {0,+,d}.
308 /// This helps expose more opportunities for folding parts of the expressions
309 /// into GEP indices.
311 static void SplitAddRecs(SmallVectorImpl<const SCEV *> &Ops,
313 ScalarEvolution &SE) {
315 SmallVector<const SCEV *, 8> AddRecs;
316 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
317 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Ops[i])) {
318 const SCEV *Start = A->getStart();
319 if (Start->isZero()) break;
320 const SCEV *Zero = SE.getConstant(Ty, 0);
321 AddRecs.push_back(SE.getAddRecExpr(Zero,
322 A->getStepRecurrence(SE),
324 // FIXME: A->getNoWrapFlags(FlagNW)
326 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Start)) {
328 Ops.append(Add->op_begin(), Add->op_end());
329 e += Add->getNumOperands();
334 if (!AddRecs.empty()) {
335 // Add the addrecs onto the end of the list.
336 Ops.append(AddRecs.begin(), AddRecs.end());
337 // Resort the operand list, moving any constants to the front.
338 SimplifyAddOperands(Ops, Ty, SE);
342 /// expandAddToGEP - Expand an addition expression with a pointer type into
343 /// a GEP instead of using ptrtoint+arithmetic+inttoptr. This helps
344 /// BasicAliasAnalysis and other passes analyze the result. See the rules
345 /// for getelementptr vs. inttoptr in
346 /// http://llvm.org/docs/LangRef.html#pointeraliasing
349 /// Design note: The correctness of using getelementptr here depends on
350 /// ScalarEvolution not recognizing inttoptr and ptrtoint operators, as
351 /// they may introduce pointer arithmetic which may not be safely converted
352 /// into getelementptr.
354 /// Design note: It might seem desirable for this function to be more
355 /// loop-aware. If some of the indices are loop-invariant while others
356 /// aren't, it might seem desirable to emit multiple GEPs, keeping the
357 /// loop-invariant portions of the overall computation outside the loop.
358 /// However, there are a few reasons this is not done here. Hoisting simple
359 /// arithmetic is a low-level optimization that often isn't very
360 /// important until late in the optimization process. In fact, passes
361 /// like InstructionCombining will combine GEPs, even if it means
362 /// pushing loop-invariant computation down into loops, so even if the
363 /// GEPs were split here, the work would quickly be undone. The
364 /// LoopStrengthReduction pass, which is usually run quite late (and
365 /// after the last InstructionCombining pass), takes care of hoisting
366 /// loop-invariant portions of expressions, after considering what
367 /// can be folded using target addressing modes.
369 Value *SCEVExpander::expandAddToGEP(const SCEV *const *op_begin,
370 const SCEV *const *op_end,
374 Type *ElTy = PTy->getElementType();
375 SmallVector<Value *, 4> GepIndices;
376 SmallVector<const SCEV *, 8> Ops(op_begin, op_end);
377 bool AnyNonZeroIndices = false;
379 // Split AddRecs up into parts as either of the parts may be usable
380 // without the other.
381 SplitAddRecs(Ops, Ty, SE);
383 // Descend down the pointer's type and attempt to convert the other
384 // operands into GEP indices, at each level. The first index in a GEP
385 // indexes into the array implied by the pointer operand; the rest of
386 // the indices index into the element or field type selected by the
389 // If the scale size is not 0, attempt to factor out a scale for
391 SmallVector<const SCEV *, 8> ScaledOps;
392 if (ElTy->isSized()) {
393 const SCEV *ElSize = SE.getSizeOfExpr(ElTy);
394 if (!ElSize->isZero()) {
395 SmallVector<const SCEV *, 8> NewOps;
396 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
397 const SCEV *Op = Ops[i];
398 const SCEV *Remainder = SE.getConstant(Ty, 0);
399 if (FactorOutConstant(Op, Remainder, ElSize, SE, SE.TD)) {
400 // Op now has ElSize factored out.
401 ScaledOps.push_back(Op);
402 if (!Remainder->isZero())
403 NewOps.push_back(Remainder);
404 AnyNonZeroIndices = true;
406 // The operand was not divisible, so add it to the list of operands
407 // we'll scan next iteration.
408 NewOps.push_back(Ops[i]);
411 // If we made any changes, update Ops.
412 if (!ScaledOps.empty()) {
414 SimplifyAddOperands(Ops, Ty, SE);
419 // Record the scaled array index for this level of the type. If
420 // we didn't find any operands that could be factored, tentatively
421 // assume that element zero was selected (since the zero offset
422 // would obviously be folded away).
423 Value *Scaled = ScaledOps.empty() ?
424 Constant::getNullValue(Ty) :
425 expandCodeFor(SE.getAddExpr(ScaledOps), Ty);
426 GepIndices.push_back(Scaled);
428 // Collect struct field index operands.
429 while (StructType *STy = dyn_cast<StructType>(ElTy)) {
430 bool FoundFieldNo = false;
431 // An empty struct has no fields.
432 if (STy->getNumElements() == 0) break;
434 // With TargetData, field offsets are known. See if a constant offset
435 // falls within any of the struct fields.
436 if (Ops.empty()) break;
437 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[0]))
438 if (SE.getTypeSizeInBits(C->getType()) <= 64) {
439 const StructLayout &SL = *SE.TD->getStructLayout(STy);
440 uint64_t FullOffset = C->getValue()->getZExtValue();
441 if (FullOffset < SL.getSizeInBytes()) {
442 unsigned ElIdx = SL.getElementContainingOffset(FullOffset);
443 GepIndices.push_back(
444 ConstantInt::get(Type::getInt32Ty(Ty->getContext()), ElIdx));
445 ElTy = STy->getTypeAtIndex(ElIdx);
447 SE.getConstant(Ty, FullOffset - SL.getElementOffset(ElIdx));
448 AnyNonZeroIndices = true;
453 // Without TargetData, just check for an offsetof expression of the
454 // appropriate struct type.
455 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
456 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Ops[i])) {
459 if (U->isOffsetOf(CTy, FieldNo) && CTy == STy) {
460 GepIndices.push_back(FieldNo);
462 STy->getTypeAtIndex(cast<ConstantInt>(FieldNo)->getZExtValue());
463 Ops[i] = SE.getConstant(Ty, 0);
464 AnyNonZeroIndices = true;
470 // If no struct field offsets were found, tentatively assume that
471 // field zero was selected (since the zero offset would obviously
474 ElTy = STy->getTypeAtIndex(0u);
475 GepIndices.push_back(
476 Constant::getNullValue(Type::getInt32Ty(Ty->getContext())));
480 if (ArrayType *ATy = dyn_cast<ArrayType>(ElTy))
481 ElTy = ATy->getElementType();
486 // If none of the operands were convertible to proper GEP indices, cast
487 // the base to i8* and do an ugly getelementptr with that. It's still
488 // better than ptrtoint+arithmetic+inttoptr at least.
489 if (!AnyNonZeroIndices) {
490 // Cast the base to i8*.
491 V = InsertNoopCastOfTo(V,
492 Type::getInt8PtrTy(Ty->getContext(), PTy->getAddressSpace()));
494 // Expand the operands for a plain byte offset.
495 Value *Idx = expandCodeFor(SE.getAddExpr(Ops), Ty);
497 // Fold a GEP with constant operands.
498 if (Constant *CLHS = dyn_cast<Constant>(V))
499 if (Constant *CRHS = dyn_cast<Constant>(Idx))
500 return ConstantExpr::getGetElementPtr(CLHS, CRHS);
502 // Do a quick scan to see if we have this GEP nearby. If so, reuse it.
503 unsigned ScanLimit = 6;
504 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
505 // Scanning starts from the last instruction before the insertion point.
506 BasicBlock::iterator IP = Builder.GetInsertPoint();
507 if (IP != BlockBegin) {
509 for (; ScanLimit; --IP, --ScanLimit) {
510 // Don't count dbg.value against the ScanLimit, to avoid perturbing the
512 if (isa<DbgInfoIntrinsic>(IP))
514 if (IP->getOpcode() == Instruction::GetElementPtr &&
515 IP->getOperand(0) == V && IP->getOperand(1) == Idx)
517 if (IP == BlockBegin) break;
521 // Save the original insertion point so we can restore it when we're done.
522 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
523 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
525 // Move the insertion point out of as many loops as we can.
526 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
527 if (!L->isLoopInvariant(V) || !L->isLoopInvariant(Idx)) break;
528 BasicBlock *Preheader = L->getLoopPreheader();
529 if (!Preheader) break;
531 // Ok, move up a level.
532 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
536 Value *GEP = Builder.CreateGEP(V, Idx, "uglygep");
537 rememberInstruction(GEP);
539 // Restore the original insert point.
541 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
546 // Save the original insertion point so we can restore it when we're done.
547 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
548 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
550 // Move the insertion point out of as many loops as we can.
551 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
552 if (!L->isLoopInvariant(V)) break;
554 bool AnyIndexNotLoopInvariant = false;
555 for (SmallVectorImpl<Value *>::const_iterator I = GepIndices.begin(),
556 E = GepIndices.end(); I != E; ++I)
557 if (!L->isLoopInvariant(*I)) {
558 AnyIndexNotLoopInvariant = true;
561 if (AnyIndexNotLoopInvariant)
564 BasicBlock *Preheader = L->getLoopPreheader();
565 if (!Preheader) break;
567 // Ok, move up a level.
568 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
571 // Insert a pretty getelementptr. Note that this GEP is not marked inbounds,
572 // because ScalarEvolution may have changed the address arithmetic to
573 // compute a value which is beyond the end of the allocated object.
575 if (V->getType() != PTy)
576 Casted = InsertNoopCastOfTo(Casted, PTy);
577 Value *GEP = Builder.CreateGEP(Casted,
580 Ops.push_back(SE.getUnknown(GEP));
581 rememberInstruction(GEP);
583 // Restore the original insert point.
585 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
587 return expand(SE.getAddExpr(Ops));
590 /// isNonConstantNegative - Return true if the specified scev is negated, but
592 static bool isNonConstantNegative(const SCEV *F) {
593 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(F);
594 if (!Mul) return false;
596 // If there is a constant factor, it will be first.
597 const SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
598 if (!SC) return false;
600 // Return true if the value is negative, this matches things like (-42 * V).
601 return SC->getValue()->getValue().isNegative();
604 /// PickMostRelevantLoop - Given two loops pick the one that's most relevant for
605 /// SCEV expansion. If they are nested, this is the most nested. If they are
606 /// neighboring, pick the later.
607 static const Loop *PickMostRelevantLoop(const Loop *A, const Loop *B,
611 if (A->contains(B)) return B;
612 if (B->contains(A)) return A;
613 if (DT.dominates(A->getHeader(), B->getHeader())) return B;
614 if (DT.dominates(B->getHeader(), A->getHeader())) return A;
615 return A; // Arbitrarily break the tie.
618 /// getRelevantLoop - Get the most relevant loop associated with the given
619 /// expression, according to PickMostRelevantLoop.
620 const Loop *SCEVExpander::getRelevantLoop(const SCEV *S) {
621 // Test whether we've already computed the most relevant loop for this SCEV.
622 std::pair<DenseMap<const SCEV *, const Loop *>::iterator, bool> Pair =
623 RelevantLoops.insert(std::make_pair(S, static_cast<const Loop *>(0)));
625 return Pair.first->second;
627 if (isa<SCEVConstant>(S))
628 // A constant has no relevant loops.
630 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
631 if (const Instruction *I = dyn_cast<Instruction>(U->getValue()))
632 return Pair.first->second = SE.LI->getLoopFor(I->getParent());
633 // A non-instruction has no relevant loops.
636 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S)) {
638 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
640 for (SCEVNAryExpr::op_iterator I = N->op_begin(), E = N->op_end();
642 L = PickMostRelevantLoop(L, getRelevantLoop(*I), *SE.DT);
643 return RelevantLoops[N] = L;
645 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S)) {
646 const Loop *Result = getRelevantLoop(C->getOperand());
647 return RelevantLoops[C] = Result;
649 if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
651 PickMostRelevantLoop(getRelevantLoop(D->getLHS()),
652 getRelevantLoop(D->getRHS()),
654 return RelevantLoops[D] = Result;
656 llvm_unreachable("Unexpected SCEV type!");
662 /// LoopCompare - Compare loops by PickMostRelevantLoop.
666 explicit LoopCompare(DominatorTree &dt) : DT(dt) {}
668 bool operator()(std::pair<const Loop *, const SCEV *> LHS,
669 std::pair<const Loop *, const SCEV *> RHS) const {
670 // Keep pointer operands sorted at the end.
671 if (LHS.second->getType()->isPointerTy() !=
672 RHS.second->getType()->isPointerTy())
673 return LHS.second->getType()->isPointerTy();
675 // Compare loops with PickMostRelevantLoop.
676 if (LHS.first != RHS.first)
677 return PickMostRelevantLoop(LHS.first, RHS.first, DT) != LHS.first;
679 // If one operand is a non-constant negative and the other is not,
680 // put the non-constant negative on the right so that a sub can
681 // be used instead of a negate and add.
682 if (isNonConstantNegative(LHS.second)) {
683 if (!isNonConstantNegative(RHS.second))
685 } else if (isNonConstantNegative(RHS.second))
688 // Otherwise they are equivalent according to this comparison.
695 Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) {
696 Type *Ty = SE.getEffectiveSCEVType(S->getType());
698 // Collect all the add operands in a loop, along with their associated loops.
699 // Iterate in reverse so that constants are emitted last, all else equal, and
700 // so that pointer operands are inserted first, which the code below relies on
701 // to form more involved GEPs.
702 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops;
703 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(S->op_end()),
704 E(S->op_begin()); I != E; ++I)
705 OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I));
707 // Sort by loop. Use a stable sort so that constants follow non-constants and
708 // pointer operands precede non-pointer operands.
709 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT));
711 // Emit instructions to add all the operands. Hoist as much as possible
712 // out of loops, and form meaningful getelementptrs where possible.
714 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator
715 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) {
716 const Loop *CurLoop = I->first;
717 const SCEV *Op = I->second;
719 // This is the first operand. Just expand it.
722 } else if (PointerType *PTy = dyn_cast<PointerType>(Sum->getType())) {
723 // The running sum expression is a pointer. Try to form a getelementptr
724 // at this level with that as the base.
725 SmallVector<const SCEV *, 4> NewOps;
726 for (; I != E && I->first == CurLoop; ++I) {
727 // If the operand is SCEVUnknown and not instructions, peek through
728 // it, to enable more of it to be folded into the GEP.
729 const SCEV *X = I->second;
730 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(X))
731 if (!isa<Instruction>(U->getValue()))
732 X = SE.getSCEV(U->getValue());
735 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, Sum);
736 } else if (PointerType *PTy = dyn_cast<PointerType>(Op->getType())) {
737 // The running sum is an integer, and there's a pointer at this level.
738 // Try to form a getelementptr. If the running sum is instructions,
739 // use a SCEVUnknown to avoid re-analyzing them.
740 SmallVector<const SCEV *, 4> NewOps;
741 NewOps.push_back(isa<Instruction>(Sum) ? SE.getUnknown(Sum) :
743 for (++I; I != E && I->first == CurLoop; ++I)
744 NewOps.push_back(I->second);
745 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, expand(Op));
746 } else if (isNonConstantNegative(Op)) {
747 // Instead of doing a negate and add, just do a subtract.
748 Value *W = expandCodeFor(SE.getNegativeSCEV(Op), Ty);
749 Sum = InsertNoopCastOfTo(Sum, Ty);
750 Sum = InsertBinop(Instruction::Sub, Sum, W);
754 Value *W = expandCodeFor(Op, Ty);
755 Sum = InsertNoopCastOfTo(Sum, Ty);
756 // Canonicalize a constant to the RHS.
757 if (isa<Constant>(Sum)) std::swap(Sum, W);
758 Sum = InsertBinop(Instruction::Add, Sum, W);
766 Value *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) {
767 Type *Ty = SE.getEffectiveSCEVType(S->getType());
769 // Collect all the mul operands in a loop, along with their associated loops.
770 // Iterate in reverse so that constants are emitted last, all else equal.
771 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops;
772 for (std::reverse_iterator<SCEVMulExpr::op_iterator> I(S->op_end()),
773 E(S->op_begin()); I != E; ++I)
774 OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I));
776 // Sort by loop. Use a stable sort so that constants follow non-constants.
777 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT));
779 // Emit instructions to mul all the operands. Hoist as much as possible
782 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator
783 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) {
784 const SCEV *Op = I->second;
786 // This is the first operand. Just expand it.
789 } else if (Op->isAllOnesValue()) {
790 // Instead of doing a multiply by negative one, just do a negate.
791 Prod = InsertNoopCastOfTo(Prod, Ty);
792 Prod = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), Prod);
796 Value *W = expandCodeFor(Op, Ty);
797 Prod = InsertNoopCastOfTo(Prod, Ty);
798 // Canonicalize a constant to the RHS.
799 if (isa<Constant>(Prod)) std::swap(Prod, W);
800 Prod = InsertBinop(Instruction::Mul, Prod, W);
808 Value *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) {
809 Type *Ty = SE.getEffectiveSCEVType(S->getType());
811 Value *LHS = expandCodeFor(S->getLHS(), Ty);
812 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getRHS())) {
813 const APInt &RHS = SC->getValue()->getValue();
814 if (RHS.isPowerOf2())
815 return InsertBinop(Instruction::LShr, LHS,
816 ConstantInt::get(Ty, RHS.logBase2()));
819 Value *RHS = expandCodeFor(S->getRHS(), Ty);
820 return InsertBinop(Instruction::UDiv, LHS, RHS);
823 /// Move parts of Base into Rest to leave Base with the minimal
824 /// expression that provides a pointer operand suitable for a
826 static void ExposePointerBase(const SCEV *&Base, const SCEV *&Rest,
827 ScalarEvolution &SE) {
828 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Base)) {
829 Base = A->getStart();
830 Rest = SE.getAddExpr(Rest,
831 SE.getAddRecExpr(SE.getConstant(A->getType(), 0),
832 A->getStepRecurrence(SE),
834 // FIXME: A->getNoWrapFlags(FlagNW)
837 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(Base)) {
838 Base = A->getOperand(A->getNumOperands()-1);
839 SmallVector<const SCEV *, 8> NewAddOps(A->op_begin(), A->op_end());
840 NewAddOps.back() = Rest;
841 Rest = SE.getAddExpr(NewAddOps);
842 ExposePointerBase(Base, Rest, SE);
846 /// getAddRecExprPHILiterally - Helper for expandAddRecExprLiterally. Expand
847 /// the base addrec, which is the addrec without any non-loop-dominating
848 /// values, and return the PHI.
850 SCEVExpander::getAddRecExprPHILiterally(const SCEVAddRecExpr *Normalized,
854 assert((!IVIncInsertLoop||IVIncInsertPos) && "Uninitialized insert position");
856 // Reuse a previously-inserted PHI, if present.
857 for (BasicBlock::iterator I = L->getHeader()->begin();
858 PHINode *PN = dyn_cast<PHINode>(I); ++I)
859 if (SE.isSCEVable(PN->getType()) &&
860 (SE.getEffectiveSCEVType(PN->getType()) ==
861 SE.getEffectiveSCEVType(Normalized->getType())) &&
862 SE.getSCEV(PN) == Normalized)
863 if (BasicBlock *LatchBlock = L->getLoopLatch()) {
865 cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock));
867 // Determine if this is a well-behaved chain of instructions leading
868 // back to the PHI. It probably will be, if we're scanning an inner
869 // loop already visited by LSR for example, but it wouldn't have
872 if (IncV->getNumOperands() == 0 || isa<PHINode>(IncV) ||
873 (isa<CastInst>(IncV) && !isa<BitCastInst>(IncV))) {
877 // If any of the operands don't dominate the insert position, bail.
878 // Addrec operands are always loop-invariant, so this can only happen
879 // if there are instructions which haven't been hoisted.
880 if (L == IVIncInsertLoop) {
881 for (User::op_iterator OI = IncV->op_begin()+1,
882 OE = IncV->op_end(); OI != OE; ++OI)
883 if (Instruction *OInst = dyn_cast<Instruction>(OI))
884 if (!SE.DT->dominates(OInst, IVIncInsertPos)) {
891 // Advance to the next instruction.
892 IncV = dyn_cast<Instruction>(IncV->getOperand(0));
895 if (IncV->mayHaveSideEffects()) {
899 } while (IncV != PN);
902 // Ok, the add recurrence looks usable.
903 // Remember this PHI, even in post-inc mode.
904 InsertedValues.insert(PN);
905 // Remember the increment.
906 IncV = cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock));
907 rememberInstruction(IncV);
908 if (L == IVIncInsertLoop)
910 if (SE.DT->dominates(IncV, IVIncInsertPos))
912 // Make sure the increment is where we want it. But don't move it
913 // down past a potential existing post-inc user.
914 IncV->moveBefore(IVIncInsertPos);
915 IVIncInsertPos = IncV;
916 IncV = cast<Instruction>(IncV->getOperand(0));
917 } while (IncV != PN);
922 // Save the original insertion point so we can restore it when we're done.
923 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
924 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
926 // Expand code for the start value.
927 Value *StartV = expandCodeFor(Normalized->getStart(), ExpandTy,
928 L->getHeader()->begin());
930 // StartV must be hoisted into L's preheader to dominate the new phi.
931 assert(!isa<Instruction>(StartV) ||
932 SE.DT->properlyDominates(cast<Instruction>(StartV)->getParent(),
935 // Expand code for the step value. Insert instructions right before the
936 // terminator corresponding to the back-edge. Do this before creating the PHI
937 // so that PHI reuse code doesn't see an incomplete PHI. If the stride is
938 // negative, insert a sub instead of an add for the increment (unless it's a
939 // constant, because subtracts of constants are canonicalized to adds).
940 const SCEV *Step = Normalized->getStepRecurrence(SE);
941 bool isPointer = ExpandTy->isPointerTy();
942 bool isNegative = !isPointer && isNonConstantNegative(Step);
944 Step = SE.getNegativeSCEV(Step);
945 Value *StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin());
948 BasicBlock *Header = L->getHeader();
949 Builder.SetInsertPoint(Header, Header->begin());
950 pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header);
951 PHINode *PN = Builder.CreatePHI(ExpandTy, std::distance(HPB, HPE),
952 Twine(IVName) + ".iv");
953 rememberInstruction(PN);
955 // Create the step instructions and populate the PHI.
956 for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) {
957 BasicBlock *Pred = *HPI;
959 // Add a start value.
960 if (!L->contains(Pred)) {
961 PN->addIncoming(StartV, Pred);
965 // Create a step value and add it to the PHI. If IVIncInsertLoop is
966 // non-null and equal to the addrec's loop, insert the instructions
967 // at IVIncInsertPos.
968 Instruction *InsertPos = L == IVIncInsertLoop ?
969 IVIncInsertPos : Pred->getTerminator();
970 Builder.SetInsertPoint(InsertPos);
972 // If the PHI is a pointer, use a GEP, otherwise use an add or sub.
974 PointerType *GEPPtrTy = cast<PointerType>(ExpandTy);
975 // If the step isn't constant, don't use an implicitly scaled GEP, because
976 // that would require a multiply inside the loop.
977 if (!isa<ConstantInt>(StepV))
978 GEPPtrTy = PointerType::get(Type::getInt1Ty(SE.getContext()),
979 GEPPtrTy->getAddressSpace());
980 const SCEV *const StepArray[1] = { SE.getSCEV(StepV) };
981 IncV = expandAddToGEP(StepArray, StepArray+1, GEPPtrTy, IntTy, PN);
982 if (IncV->getType() != PN->getType()) {
983 IncV = Builder.CreateBitCast(IncV, PN->getType(), "tmp");
984 rememberInstruction(IncV);
988 Builder.CreateSub(PN, StepV, Twine(IVName) + ".iv.next") :
989 Builder.CreateAdd(PN, StepV, Twine(IVName) + ".iv.next");
990 rememberInstruction(IncV);
992 PN->addIncoming(IncV, Pred);
995 // Restore the original insert point.
997 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
999 // Remember this PHI, even in post-inc mode.
1000 InsertedValues.insert(PN);
1005 Value *SCEVExpander::expandAddRecExprLiterally(const SCEVAddRecExpr *S) {
1006 Type *STy = S->getType();
1007 Type *IntTy = SE.getEffectiveSCEVType(STy);
1008 const Loop *L = S->getLoop();
1010 // Determine a normalized form of this expression, which is the expression
1011 // before any post-inc adjustment is made.
1012 const SCEVAddRecExpr *Normalized = S;
1013 if (PostIncLoops.count(L)) {
1014 PostIncLoopSet Loops;
1017 cast<SCEVAddRecExpr>(TransformForPostIncUse(Normalize, S, 0, 0,
1018 Loops, SE, *SE.DT));
1021 // Strip off any non-loop-dominating component from the addrec start.
1022 const SCEV *Start = Normalized->getStart();
1023 const SCEV *PostLoopOffset = 0;
1024 if (!SE.properlyDominates(Start, L->getHeader())) {
1025 PostLoopOffset = Start;
1026 Start = SE.getConstant(Normalized->getType(), 0);
1027 Normalized = cast<SCEVAddRecExpr>(
1028 SE.getAddRecExpr(Start, Normalized->getStepRecurrence(SE),
1029 Normalized->getLoop(),
1030 // FIXME: Normalized->getNoWrapFlags(FlagNW)
1031 SCEV::FlagAnyWrap));
1034 // Strip off any non-loop-dominating component from the addrec step.
1035 const SCEV *Step = Normalized->getStepRecurrence(SE);
1036 const SCEV *PostLoopScale = 0;
1037 if (!SE.dominates(Step, L->getHeader())) {
1038 PostLoopScale = Step;
1039 Step = SE.getConstant(Normalized->getType(), 1);
1041 cast<SCEVAddRecExpr>(SE.getAddRecExpr(Start, Step,
1042 Normalized->getLoop(),
1043 // FIXME: Normalized
1044 // ->getNoWrapFlags(FlagNW)
1045 SCEV::FlagAnyWrap));
1048 // Expand the core addrec. If we need post-loop scaling, force it to
1049 // expand to an integer type to avoid the need for additional casting.
1050 Type *ExpandTy = PostLoopScale ? IntTy : STy;
1051 PHINode *PN = getAddRecExprPHILiterally(Normalized, L, ExpandTy, IntTy);
1053 // Accommodate post-inc mode, if necessary.
1055 if (!PostIncLoops.count(L))
1058 // In PostInc mode, use the post-incremented value.
1059 BasicBlock *LatchBlock = L->getLoopLatch();
1060 assert(LatchBlock && "PostInc mode requires a unique loop latch!");
1061 Result = PN->getIncomingValueForBlock(LatchBlock);
1064 // Re-apply any non-loop-dominating scale.
1065 if (PostLoopScale) {
1066 Result = InsertNoopCastOfTo(Result, IntTy);
1067 Result = Builder.CreateMul(Result,
1068 expandCodeFor(PostLoopScale, IntTy));
1069 rememberInstruction(Result);
1072 // Re-apply any non-loop-dominating offset.
1073 if (PostLoopOffset) {
1074 if (PointerType *PTy = dyn_cast<PointerType>(ExpandTy)) {
1075 const SCEV *const OffsetArray[1] = { PostLoopOffset };
1076 Result = expandAddToGEP(OffsetArray, OffsetArray+1, PTy, IntTy, Result);
1078 Result = InsertNoopCastOfTo(Result, IntTy);
1079 Result = Builder.CreateAdd(Result,
1080 expandCodeFor(PostLoopOffset, IntTy));
1081 rememberInstruction(Result);
1088 Value *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) {
1089 if (!CanonicalMode) return expandAddRecExprLiterally(S);
1091 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1092 const Loop *L = S->getLoop();
1094 // First check for an existing canonical IV in a suitable type.
1095 PHINode *CanonicalIV = 0;
1096 if (PHINode *PN = L->getCanonicalInductionVariable())
1097 if (SE.getTypeSizeInBits(PN->getType()) >= SE.getTypeSizeInBits(Ty))
1100 // Rewrite an AddRec in terms of the canonical induction variable, if
1101 // its type is more narrow.
1103 SE.getTypeSizeInBits(CanonicalIV->getType()) >
1104 SE.getTypeSizeInBits(Ty)) {
1105 SmallVector<const SCEV *, 4> NewOps(S->getNumOperands());
1106 for (unsigned i = 0, e = S->getNumOperands(); i != e; ++i)
1107 NewOps[i] = SE.getAnyExtendExpr(S->op_begin()[i], CanonicalIV->getType());
1108 Value *V = expand(SE.getAddRecExpr(NewOps, S->getLoop(),
1109 // FIXME: S->getNoWrapFlags(FlagNW)
1110 SCEV::FlagAnyWrap));
1111 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1112 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1113 BasicBlock::iterator NewInsertPt =
1114 llvm::next(BasicBlock::iterator(cast<Instruction>(V)));
1115 while (isa<PHINode>(NewInsertPt) || isa<DbgInfoIntrinsic>(NewInsertPt) ||
1116 isa<LandingPadInst>(NewInsertPt))
1118 V = expandCodeFor(SE.getTruncateExpr(SE.getUnknown(V), Ty), 0,
1120 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
1124 // {X,+,F} --> X + {0,+,F}
1125 if (!S->getStart()->isZero()) {
1126 SmallVector<const SCEV *, 4> NewOps(S->op_begin(), S->op_end());
1127 NewOps[0] = SE.getConstant(Ty, 0);
1128 // FIXME: can use S->getNoWrapFlags()
1129 const SCEV *Rest = SE.getAddRecExpr(NewOps, L, SCEV::FlagAnyWrap);
1131 // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the
1132 // comments on expandAddToGEP for details.
1133 const SCEV *Base = S->getStart();
1134 const SCEV *RestArray[1] = { Rest };
1135 // Dig into the expression to find the pointer base for a GEP.
1136 ExposePointerBase(Base, RestArray[0], SE);
1137 // If we found a pointer, expand the AddRec with a GEP.
1138 if (PointerType *PTy = dyn_cast<PointerType>(Base->getType())) {
1139 // Make sure the Base isn't something exotic, such as a multiplied
1140 // or divided pointer value. In those cases, the result type isn't
1141 // actually a pointer type.
1142 if (!isa<SCEVMulExpr>(Base) && !isa<SCEVUDivExpr>(Base)) {
1143 Value *StartV = expand(Base);
1144 assert(StartV->getType() == PTy && "Pointer type mismatch for GEP!");
1145 return expandAddToGEP(RestArray, RestArray+1, PTy, Ty, StartV);
1149 // Just do a normal add. Pre-expand the operands to suppress folding.
1150 return expand(SE.getAddExpr(SE.getUnknown(expand(S->getStart())),
1151 SE.getUnknown(expand(Rest))));
1154 // If we don't yet have a canonical IV, create one.
1156 // Create and insert the PHI node for the induction variable in the
1158 BasicBlock *Header = L->getHeader();
1159 pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header);
1160 CanonicalIV = PHINode::Create(Ty, std::distance(HPB, HPE), "indvar",
1162 rememberInstruction(CanonicalIV);
1164 Constant *One = ConstantInt::get(Ty, 1);
1165 for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) {
1166 BasicBlock *HP = *HPI;
1167 if (L->contains(HP)) {
1168 // Insert a unit add instruction right before the terminator
1169 // corresponding to the back-edge.
1170 Instruction *Add = BinaryOperator::CreateAdd(CanonicalIV, One,
1172 HP->getTerminator());
1173 Add->setDebugLoc(HP->getTerminator()->getDebugLoc());
1174 rememberInstruction(Add);
1175 CanonicalIV->addIncoming(Add, HP);
1177 CanonicalIV->addIncoming(Constant::getNullValue(Ty), HP);
1182 // {0,+,1} --> Insert a canonical induction variable into the loop!
1183 if (S->isAffine() && S->getOperand(1)->isOne()) {
1184 assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) &&
1185 "IVs with types different from the canonical IV should "
1186 "already have been handled!");
1190 // {0,+,F} --> {0,+,1} * F
1192 // If this is a simple linear addrec, emit it now as a special case.
1193 if (S->isAffine()) // {0,+,F} --> i*F
1195 expand(SE.getTruncateOrNoop(
1196 SE.getMulExpr(SE.getUnknown(CanonicalIV),
1197 SE.getNoopOrAnyExtend(S->getOperand(1),
1198 CanonicalIV->getType())),
1201 // If this is a chain of recurrences, turn it into a closed form, using the
1202 // folders, then expandCodeFor the closed form. This allows the folders to
1203 // simplify the expression without having to build a bunch of special code
1204 // into this folder.
1205 const SCEV *IH = SE.getUnknown(CanonicalIV); // Get I as a "symbolic" SCEV.
1207 // Promote S up to the canonical IV type, if the cast is foldable.
1208 const SCEV *NewS = S;
1209 const SCEV *Ext = SE.getNoopOrAnyExtend(S, CanonicalIV->getType());
1210 if (isa<SCEVAddRecExpr>(Ext))
1213 const SCEV *V = cast<SCEVAddRecExpr>(NewS)->evaluateAtIteration(IH, SE);
1214 //cerr << "Evaluated: " << *this << "\n to: " << *V << "\n";
1216 // Truncate the result down to the original type, if needed.
1217 const SCEV *T = SE.getTruncateOrNoop(V, Ty);
1221 Value *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) {
1222 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1223 Value *V = expandCodeFor(S->getOperand(),
1224 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1225 Value *I = Builder.CreateTrunc(V, Ty, "tmp");
1226 rememberInstruction(I);
1230 Value *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) {
1231 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1232 Value *V = expandCodeFor(S->getOperand(),
1233 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1234 Value *I = Builder.CreateZExt(V, Ty, "tmp");
1235 rememberInstruction(I);
1239 Value *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) {
1240 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1241 Value *V = expandCodeFor(S->getOperand(),
1242 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1243 Value *I = Builder.CreateSExt(V, Ty, "tmp");
1244 rememberInstruction(I);
1248 Value *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) {
1249 Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
1250 Type *Ty = LHS->getType();
1251 for (int i = S->getNumOperands()-2; i >= 0; --i) {
1252 // In the case of mixed integer and pointer types, do the
1253 // rest of the comparisons as integer.
1254 if (S->getOperand(i)->getType() != Ty) {
1255 Ty = SE.getEffectiveSCEVType(Ty);
1256 LHS = InsertNoopCastOfTo(LHS, Ty);
1258 Value *RHS = expandCodeFor(S->getOperand(i), Ty);
1259 Value *ICmp = Builder.CreateICmpSGT(LHS, RHS, "tmp");
1260 rememberInstruction(ICmp);
1261 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "smax");
1262 rememberInstruction(Sel);
1265 // In the case of mixed integer and pointer types, cast the
1266 // final result back to the pointer type.
1267 if (LHS->getType() != S->getType())
1268 LHS = InsertNoopCastOfTo(LHS, S->getType());
1272 Value *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) {
1273 Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
1274 Type *Ty = LHS->getType();
1275 for (int i = S->getNumOperands()-2; i >= 0; --i) {
1276 // In the case of mixed integer and pointer types, do the
1277 // rest of the comparisons as integer.
1278 if (S->getOperand(i)->getType() != Ty) {
1279 Ty = SE.getEffectiveSCEVType(Ty);
1280 LHS = InsertNoopCastOfTo(LHS, Ty);
1282 Value *RHS = expandCodeFor(S->getOperand(i), Ty);
1283 Value *ICmp = Builder.CreateICmpUGT(LHS, RHS, "tmp");
1284 rememberInstruction(ICmp);
1285 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "umax");
1286 rememberInstruction(Sel);
1289 // In the case of mixed integer and pointer types, cast the
1290 // final result back to the pointer type.
1291 if (LHS->getType() != S->getType())
1292 LHS = InsertNoopCastOfTo(LHS, S->getType());
1296 Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty,
1298 BasicBlock::iterator IP = I;
1299 while (isInsertedInstruction(IP) || isa<DbgInfoIntrinsic>(IP))
1301 Builder.SetInsertPoint(IP->getParent(), IP);
1302 return expandCodeFor(SH, Ty);
1305 Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty) {
1306 // Expand the code for this SCEV.
1307 Value *V = expand(SH);
1309 assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) &&
1310 "non-trivial casts should be done with the SCEVs directly!");
1311 V = InsertNoopCastOfTo(V, Ty);
1316 Value *SCEVExpander::expand(const SCEV *S) {
1317 // Compute an insertion point for this SCEV object. Hoist the instructions
1318 // as far out in the loop nest as possible.
1319 Instruction *InsertPt = Builder.GetInsertPoint();
1320 for (Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock()); ;
1321 L = L->getParentLoop())
1322 if (SE.isLoopInvariant(S, L)) {
1324 if (BasicBlock *Preheader = L->getLoopPreheader())
1325 InsertPt = Preheader->getTerminator();
1327 // If the SCEV is computable at this level, insert it into the header
1328 // after the PHIs (and after any other instructions that we've inserted
1329 // there) so that it is guaranteed to dominate any user inside the loop.
1330 if (L && SE.hasComputableLoopEvolution(S, L) && !PostIncLoops.count(L))
1331 InsertPt = L->getHeader()->getFirstInsertionPt();
1332 while (isInsertedInstruction(InsertPt) || isa<DbgInfoIntrinsic>(InsertPt))
1333 InsertPt = llvm::next(BasicBlock::iterator(InsertPt));
1337 // Check to see if we already expanded this here.
1338 std::map<std::pair<const SCEV *, Instruction *>,
1339 AssertingVH<Value> >::iterator I =
1340 InsertedExpressions.find(std::make_pair(S, InsertPt));
1341 if (I != InsertedExpressions.end())
1344 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1345 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1346 Builder.SetInsertPoint(InsertPt->getParent(), InsertPt);
1348 // Expand the expression into instructions.
1349 Value *V = visit(S);
1351 // Remember the expanded value for this SCEV at this location.
1352 if (PostIncLoops.empty())
1353 InsertedExpressions[std::make_pair(S, InsertPt)] = V;
1355 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
1359 void SCEVExpander::rememberInstruction(Value *I) {
1360 if (!PostIncLoops.empty())
1361 InsertedPostIncValues.insert(I);
1363 InsertedValues.insert(I);
1365 // If we just claimed an existing instruction and that instruction had
1366 // been the insert point, adjust the insert point forward so that
1367 // subsequently inserted code will be dominated.
1368 if (Builder.GetInsertPoint() == I) {
1369 BasicBlock::iterator It = cast<Instruction>(I);
1370 do { ++It; } while (isInsertedInstruction(It) ||
1371 isa<DbgInfoIntrinsic>(It));
1372 Builder.SetInsertPoint(Builder.GetInsertBlock(), It);
1376 void SCEVExpander::restoreInsertPoint(BasicBlock *BB, BasicBlock::iterator I) {
1377 // If we acquired more instructions since the old insert point was saved,
1378 // advance past them.
1379 while (isInsertedInstruction(I) || isa<DbgInfoIntrinsic>(I)) ++I;
1381 Builder.SetInsertPoint(BB, I);
1384 /// getOrInsertCanonicalInductionVariable - This method returns the
1385 /// canonical induction variable of the specified type for the specified
1386 /// loop (inserting one if there is none). A canonical induction variable
1387 /// starts at zero and steps by one on each iteration.
1389 SCEVExpander::getOrInsertCanonicalInductionVariable(const Loop *L,
1391 assert(Ty->isIntegerTy() && "Can only insert integer induction variables!");
1393 // Build a SCEV for {0,+,1}<L>.
1394 // Conservatively use FlagAnyWrap for now.
1395 const SCEV *H = SE.getAddRecExpr(SE.getConstant(Ty, 0),
1396 SE.getConstant(Ty, 1), L, SCEV::FlagAnyWrap);
1398 // Emit code for it.
1399 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1400 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1401 PHINode *V = cast<PHINode>(expandCodeFor(H, 0, L->getHeader()->begin()));
1403 restoreInsertPoint(SaveInsertBB, SaveInsertPt);