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 - Arange 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 /// or 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();
34 if ((*UI)->getType() == Ty)
35 if (CastInst *CI = dyn_cast<CastInst>(cast<Instruction>(*UI)))
36 if (CI->getOpcode() == Op) {
37 // If the cast isn't where we want it, fix it.
38 if (BasicBlock::iterator(CI) != IP) {
39 // Create a new cast, and leave the old cast in place in case
40 // it is being used as an insert point. Clear its operand
41 // so that it doesn't hold anything live.
42 Instruction *NewCI = CastInst::Create(Op, V, Ty, "", IP);
44 CI->replaceAllUsesWith(NewCI);
45 CI->setOperand(0, UndefValue::get(V->getType()));
46 rememberInstruction(NewCI);
53 Instruction *I = CastInst::Create(Op, V, Ty, V->getName(), IP);
54 rememberInstruction(I);
58 /// InsertNoopCastOfTo - Insert a cast of V to the specified type,
59 /// which must be possible with a noop cast, doing what we can to share
61 Value *SCEVExpander::InsertNoopCastOfTo(Value *V, const Type *Ty) {
62 Instruction::CastOps Op = CastInst::getCastOpcode(V, false, Ty, false);
63 assert((Op == Instruction::BitCast ||
64 Op == Instruction::PtrToInt ||
65 Op == Instruction::IntToPtr) &&
66 "InsertNoopCastOfTo cannot perform non-noop casts!");
67 assert(SE.getTypeSizeInBits(V->getType()) == SE.getTypeSizeInBits(Ty) &&
68 "InsertNoopCastOfTo cannot change sizes!");
70 // Short-circuit unnecessary bitcasts.
71 if (Op == Instruction::BitCast && V->getType() == Ty)
74 // Short-circuit unnecessary inttoptr<->ptrtoint casts.
75 if ((Op == Instruction::PtrToInt || Op == Instruction::IntToPtr) &&
76 SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(V->getType())) {
77 if (CastInst *CI = dyn_cast<CastInst>(V))
78 if ((CI->getOpcode() == Instruction::PtrToInt ||
79 CI->getOpcode() == Instruction::IntToPtr) &&
80 SE.getTypeSizeInBits(CI->getType()) ==
81 SE.getTypeSizeInBits(CI->getOperand(0)->getType()))
82 return CI->getOperand(0);
83 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
84 if ((CE->getOpcode() == Instruction::PtrToInt ||
85 CE->getOpcode() == Instruction::IntToPtr) &&
86 SE.getTypeSizeInBits(CE->getType()) ==
87 SE.getTypeSizeInBits(CE->getOperand(0)->getType()))
88 return CE->getOperand(0);
91 // Fold a cast of a constant.
92 if (Constant *C = dyn_cast<Constant>(V))
93 return ConstantExpr::getCast(Op, C, Ty);
95 // Cast the argument at the beginning of the entry block, after
96 // any bitcasts of other arguments.
97 if (Argument *A = dyn_cast<Argument>(V)) {
98 BasicBlock::iterator IP = A->getParent()->getEntryBlock().begin();
99 while ((isa<BitCastInst>(IP) &&
100 isa<Argument>(cast<BitCastInst>(IP)->getOperand(0)) &&
101 cast<BitCastInst>(IP)->getOperand(0) != A) ||
102 isa<DbgInfoIntrinsic>(IP))
104 return ReuseOrCreateCast(A, Ty, Op, IP);
107 // Cast the instruction immediately after the instruction.
108 Instruction *I = cast<Instruction>(V);
109 BasicBlock::iterator IP = I; ++IP;
110 if (InvokeInst *II = dyn_cast<InvokeInst>(I))
111 IP = II->getNormalDest()->begin();
112 while (isa<PHINode>(IP) || isa<DbgInfoIntrinsic>(IP)) ++IP;
113 return ReuseOrCreateCast(I, Ty, Op, IP);
116 /// InsertBinop - Insert the specified binary operator, doing a small amount
117 /// of work to avoid inserting an obviously redundant operation.
118 Value *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode,
119 Value *LHS, Value *RHS) {
120 // Fold a binop with constant operands.
121 if (Constant *CLHS = dyn_cast<Constant>(LHS))
122 if (Constant *CRHS = dyn_cast<Constant>(RHS))
123 return ConstantExpr::get(Opcode, CLHS, CRHS);
125 // Do a quick scan to see if we have this binop nearby. If so, reuse it.
126 unsigned ScanLimit = 6;
127 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
128 // Scanning starts from the last instruction before the insertion point.
129 BasicBlock::iterator IP = Builder.GetInsertPoint();
130 if (IP != BlockBegin) {
132 for (; ScanLimit; --IP, --ScanLimit) {
133 // Don't count dbg.value against the ScanLimit, to avoid perturbing the
135 if (isa<DbgInfoIntrinsic>(IP))
137 if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(0) == LHS &&
138 IP->getOperand(1) == RHS)
140 if (IP == BlockBegin) break;
144 // Save the original insertion point so we can restore it when we're done.
145 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
146 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
148 // Move the insertion point out of as many loops as we can.
149 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
150 if (!L->isLoopInvariant(LHS) || !L->isLoopInvariant(RHS)) break;
151 BasicBlock *Preheader = L->getLoopPreheader();
152 if (!Preheader) break;
154 // Ok, move up a level.
155 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
158 // If we haven't found this binop, insert it.
159 Value *BO = Builder.CreateBinOp(Opcode, LHS, RHS, "tmp");
160 rememberInstruction(BO);
162 // Restore the original insert point.
164 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
169 /// FactorOutConstant - Test if S is divisible by Factor, using signed
170 /// division. If so, update S with Factor divided out and return true.
171 /// S need not be evenly divisible if a reasonable remainder can be
173 /// TODO: When ScalarEvolution gets a SCEVSDivExpr, this can be made
174 /// unnecessary; in its place, just signed-divide Ops[i] by the scale and
175 /// check to see if the divide was folded.
176 static bool FactorOutConstant(const SCEV *&S,
177 const SCEV *&Remainder,
180 const TargetData *TD) {
181 // Everything is divisible by one.
187 S = SE.getConstant(S->getType(), 1);
191 // For a Constant, check for a multiple of the given factor.
192 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
196 // Check for divisibility.
197 if (const SCEVConstant *FC = dyn_cast<SCEVConstant>(Factor)) {
199 ConstantInt::get(SE.getContext(),
200 C->getValue()->getValue().sdiv(
201 FC->getValue()->getValue()));
202 // If the quotient is zero and the remainder is non-zero, reject
203 // the value at this scale. It will be considered for subsequent
206 const SCEV *Div = SE.getConstant(CI);
209 SE.getAddExpr(Remainder,
210 SE.getConstant(C->getValue()->getValue().srem(
211 FC->getValue()->getValue())));
217 // In a Mul, check if there is a constant operand which is a multiple
218 // of the given factor.
219 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
221 // With TargetData, the size is known. Check if there is a constant
222 // operand which is a multiple of the given factor. If so, we can
224 const SCEVConstant *FC = cast<SCEVConstant>(Factor);
225 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0)))
226 if (!C->getValue()->getValue().srem(FC->getValue()->getValue())) {
227 SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end());
229 SE.getConstant(C->getValue()->getValue().sdiv(
230 FC->getValue()->getValue()));
231 S = SE.getMulExpr(NewMulOps);
235 // Without TargetData, check if Factor can be factored out of any of the
236 // Mul's operands. If so, we can just remove it.
237 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) {
238 const SCEV *SOp = M->getOperand(i);
239 const SCEV *Remainder = SE.getConstant(SOp->getType(), 0);
240 if (FactorOutConstant(SOp, Remainder, Factor, SE, TD) &&
241 Remainder->isZero()) {
242 SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end());
244 S = SE.getMulExpr(NewMulOps);
251 // In an AddRec, check if both start and step are divisible.
252 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
253 const SCEV *Step = A->getStepRecurrence(SE);
254 const SCEV *StepRem = SE.getConstant(Step->getType(), 0);
255 if (!FactorOutConstant(Step, StepRem, Factor, SE, TD))
257 if (!StepRem->isZero())
259 const SCEV *Start = A->getStart();
260 if (!FactorOutConstant(Start, Remainder, Factor, SE, TD))
262 S = SE.getAddRecExpr(Start, Step, A->getLoop());
269 /// SimplifyAddOperands - Sort and simplify a list of add operands. NumAddRecs
270 /// is the number of SCEVAddRecExprs present, which are kept at the end of
273 static void SimplifyAddOperands(SmallVectorImpl<const SCEV *> &Ops,
275 ScalarEvolution &SE) {
276 unsigned NumAddRecs = 0;
277 for (unsigned i = Ops.size(); i > 0 && isa<SCEVAddRecExpr>(Ops[i-1]); --i)
279 // Group Ops into non-addrecs and addrecs.
280 SmallVector<const SCEV *, 8> NoAddRecs(Ops.begin(), Ops.end() - NumAddRecs);
281 SmallVector<const SCEV *, 8> AddRecs(Ops.end() - NumAddRecs, Ops.end());
282 // Let ScalarEvolution sort and simplify the non-addrecs list.
283 const SCEV *Sum = NoAddRecs.empty() ?
284 SE.getConstant(Ty, 0) :
285 SE.getAddExpr(NoAddRecs);
286 // If it returned an add, use the operands. Otherwise it simplified
287 // the sum into a single value, so just use that.
289 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Sum))
290 Ops.insert(Ops.end(), Add->op_begin(), Add->op_end());
291 else if (!Sum->isZero())
293 // Then append the addrecs.
294 Ops.insert(Ops.end(), AddRecs.begin(), AddRecs.end());
297 /// SplitAddRecs - Flatten a list of add operands, moving addrec start values
298 /// out to the top level. For example, convert {a + b,+,c} to a, b, {0,+,d}.
299 /// This helps expose more opportunities for folding parts of the expressions
300 /// into GEP indices.
302 static void SplitAddRecs(SmallVectorImpl<const SCEV *> &Ops,
304 ScalarEvolution &SE) {
306 SmallVector<const SCEV *, 8> AddRecs;
307 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
308 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Ops[i])) {
309 const SCEV *Start = A->getStart();
310 if (Start->isZero()) break;
311 const SCEV *Zero = SE.getConstant(Ty, 0);
312 AddRecs.push_back(SE.getAddRecExpr(Zero,
313 A->getStepRecurrence(SE),
315 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Start)) {
317 Ops.insert(Ops.end(), Add->op_begin(), Add->op_end());
318 e += Add->getNumOperands();
323 if (!AddRecs.empty()) {
324 // Add the addrecs onto the end of the list.
325 Ops.insert(Ops.end(), AddRecs.begin(), AddRecs.end());
326 // Resort the operand list, moving any constants to the front.
327 SimplifyAddOperands(Ops, Ty, SE);
331 /// expandAddToGEP - Expand an addition expression with a pointer type into
332 /// a GEP instead of using ptrtoint+arithmetic+inttoptr. This helps
333 /// BasicAliasAnalysis and other passes analyze the result. See the rules
334 /// for getelementptr vs. inttoptr in
335 /// http://llvm.org/docs/LangRef.html#pointeraliasing
338 /// Design note: The correctness of using getelementptr here depends on
339 /// ScalarEvolution not recognizing inttoptr and ptrtoint operators, as
340 /// they may introduce pointer arithmetic which may not be safely converted
341 /// into getelementptr.
343 /// Design note: It might seem desirable for this function to be more
344 /// loop-aware. If some of the indices are loop-invariant while others
345 /// aren't, it might seem desirable to emit multiple GEPs, keeping the
346 /// loop-invariant portions of the overall computation outside the loop.
347 /// However, there are a few reasons this is not done here. Hoisting simple
348 /// arithmetic is a low-level optimization that often isn't very
349 /// important until late in the optimization process. In fact, passes
350 /// like InstructionCombining will combine GEPs, even if it means
351 /// pushing loop-invariant computation down into loops, so even if the
352 /// GEPs were split here, the work would quickly be undone. The
353 /// LoopStrengthReduction pass, which is usually run quite late (and
354 /// after the last InstructionCombining pass), takes care of hoisting
355 /// loop-invariant portions of expressions, after considering what
356 /// can be folded using target addressing modes.
358 Value *SCEVExpander::expandAddToGEP(const SCEV *const *op_begin,
359 const SCEV *const *op_end,
360 const PointerType *PTy,
363 const Type *ElTy = PTy->getElementType();
364 SmallVector<Value *, 4> GepIndices;
365 SmallVector<const SCEV *, 8> Ops(op_begin, op_end);
366 bool AnyNonZeroIndices = false;
368 // Split AddRecs up into parts as either of the parts may be usable
369 // without the other.
370 SplitAddRecs(Ops, Ty, SE);
372 // Descend down the pointer's type and attempt to convert the other
373 // operands into GEP indices, at each level. The first index in a GEP
374 // indexes into the array implied by the pointer operand; the rest of
375 // the indices index into the element or field type selected by the
378 // If the scale size is not 0, attempt to factor out a scale for
380 SmallVector<const SCEV *, 8> ScaledOps;
381 if (ElTy->isSized()) {
382 const SCEV *ElSize = SE.getSizeOfExpr(ElTy);
383 if (!ElSize->isZero()) {
384 SmallVector<const SCEV *, 8> NewOps;
385 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
386 const SCEV *Op = Ops[i];
387 const SCEV *Remainder = SE.getConstant(Ty, 0);
388 if (FactorOutConstant(Op, Remainder, ElSize, SE, SE.TD)) {
389 // Op now has ElSize factored out.
390 ScaledOps.push_back(Op);
391 if (!Remainder->isZero())
392 NewOps.push_back(Remainder);
393 AnyNonZeroIndices = true;
395 // The operand was not divisible, so add it to the list of operands
396 // we'll scan next iteration.
397 NewOps.push_back(Ops[i]);
400 // If we made any changes, update Ops.
401 if (!ScaledOps.empty()) {
403 SimplifyAddOperands(Ops, Ty, SE);
408 // Record the scaled array index for this level of the type. If
409 // we didn't find any operands that could be factored, tentatively
410 // assume that element zero was selected (since the zero offset
411 // would obviously be folded away).
412 Value *Scaled = ScaledOps.empty() ?
413 Constant::getNullValue(Ty) :
414 expandCodeFor(SE.getAddExpr(ScaledOps), Ty);
415 GepIndices.push_back(Scaled);
417 // Collect struct field index operands.
418 while (const StructType *STy = dyn_cast<StructType>(ElTy)) {
419 bool FoundFieldNo = false;
420 // An empty struct has no fields.
421 if (STy->getNumElements() == 0) break;
423 // With TargetData, field offsets are known. See if a constant offset
424 // falls within any of the struct fields.
425 if (Ops.empty()) break;
426 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[0]))
427 if (SE.getTypeSizeInBits(C->getType()) <= 64) {
428 const StructLayout &SL = *SE.TD->getStructLayout(STy);
429 uint64_t FullOffset = C->getValue()->getZExtValue();
430 if (FullOffset < SL.getSizeInBytes()) {
431 unsigned ElIdx = SL.getElementContainingOffset(FullOffset);
432 GepIndices.push_back(
433 ConstantInt::get(Type::getInt32Ty(Ty->getContext()), ElIdx));
434 ElTy = STy->getTypeAtIndex(ElIdx);
436 SE.getConstant(Ty, FullOffset - SL.getElementOffset(ElIdx));
437 AnyNonZeroIndices = true;
442 // Without TargetData, just check for an offsetof expression of the
443 // appropriate struct type.
444 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
445 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Ops[i])) {
448 if (U->isOffsetOf(CTy, FieldNo) && CTy == STy) {
449 GepIndices.push_back(FieldNo);
451 STy->getTypeAtIndex(cast<ConstantInt>(FieldNo)->getZExtValue());
452 Ops[i] = SE.getConstant(Ty, 0);
453 AnyNonZeroIndices = true;
459 // If no struct field offsets were found, tentatively assume that
460 // field zero was selected (since the zero offset would obviously
463 ElTy = STy->getTypeAtIndex(0u);
464 GepIndices.push_back(
465 Constant::getNullValue(Type::getInt32Ty(Ty->getContext())));
469 if (const ArrayType *ATy = dyn_cast<ArrayType>(ElTy))
470 ElTy = ATy->getElementType();
475 // If none of the operands were convertible to proper GEP indices, cast
476 // the base to i8* and do an ugly getelementptr with that. It's still
477 // better than ptrtoint+arithmetic+inttoptr at least.
478 if (!AnyNonZeroIndices) {
479 // Cast the base to i8*.
480 V = InsertNoopCastOfTo(V,
481 Type::getInt8PtrTy(Ty->getContext(), PTy->getAddressSpace()));
483 // Expand the operands for a plain byte offset.
484 Value *Idx = expandCodeFor(SE.getAddExpr(Ops), Ty);
486 // Fold a GEP with constant operands.
487 if (Constant *CLHS = dyn_cast<Constant>(V))
488 if (Constant *CRHS = dyn_cast<Constant>(Idx))
489 return ConstantExpr::getGetElementPtr(CLHS, &CRHS, 1);
491 // Do a quick scan to see if we have this GEP nearby. If so, reuse it.
492 unsigned ScanLimit = 6;
493 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
494 // Scanning starts from the last instruction before the insertion point.
495 BasicBlock::iterator IP = Builder.GetInsertPoint();
496 if (IP != BlockBegin) {
498 for (; ScanLimit; --IP, --ScanLimit) {
499 // Don't count dbg.value against the ScanLimit, to avoid perturbing the
501 if (isa<DbgInfoIntrinsic>(IP))
503 if (IP->getOpcode() == Instruction::GetElementPtr &&
504 IP->getOperand(0) == V && IP->getOperand(1) == Idx)
506 if (IP == BlockBegin) break;
510 // Save the original insertion point so we can restore it when we're done.
511 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
512 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
514 // Move the insertion point out of as many loops as we can.
515 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
516 if (!L->isLoopInvariant(V) || !L->isLoopInvariant(Idx)) break;
517 BasicBlock *Preheader = L->getLoopPreheader();
518 if (!Preheader) break;
520 // Ok, move up a level.
521 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
525 Value *GEP = Builder.CreateGEP(V, Idx, "uglygep");
526 rememberInstruction(GEP);
528 // Restore the original insert point.
530 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
535 // Save the original insertion point so we can restore it when we're done.
536 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
537 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
539 // Move the insertion point out of as many loops as we can.
540 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
541 if (!L->isLoopInvariant(V)) break;
543 bool AnyIndexNotLoopInvariant = false;
544 for (SmallVectorImpl<Value *>::const_iterator I = GepIndices.begin(),
545 E = GepIndices.end(); I != E; ++I)
546 if (!L->isLoopInvariant(*I)) {
547 AnyIndexNotLoopInvariant = true;
550 if (AnyIndexNotLoopInvariant)
553 BasicBlock *Preheader = L->getLoopPreheader();
554 if (!Preheader) break;
556 // Ok, move up a level.
557 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
560 // Insert a pretty getelementptr. Note that this GEP is not marked inbounds,
561 // because ScalarEvolution may have changed the address arithmetic to
562 // compute a value which is beyond the end of the allocated object.
564 if (V->getType() != PTy)
565 Casted = InsertNoopCastOfTo(Casted, PTy);
566 Value *GEP = Builder.CreateGEP(Casted,
570 Ops.push_back(SE.getUnknown(GEP));
571 rememberInstruction(GEP);
573 // Restore the original insert point.
575 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
577 return expand(SE.getAddExpr(Ops));
580 /// isNonConstantNegative - Return true if the specified scev is negated, but
582 static bool isNonConstantNegative(const SCEV *F) {
583 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(F);
584 if (!Mul) return false;
586 // If there is a constant factor, it will be first.
587 const SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
588 if (!SC) return false;
590 // Return true if the value is negative, this matches things like (-42 * V).
591 return SC->getValue()->getValue().isNegative();
594 /// PickMostRelevantLoop - Given two loops pick the one that's most relevant for
595 /// SCEV expansion. If they are nested, this is the most nested. If they are
596 /// neighboring, pick the later.
597 static const Loop *PickMostRelevantLoop(const Loop *A, const Loop *B,
601 if (A->contains(B)) return B;
602 if (B->contains(A)) return A;
603 if (DT.dominates(A->getHeader(), B->getHeader())) return B;
604 if (DT.dominates(B->getHeader(), A->getHeader())) return A;
605 return A; // Arbitrarily break the tie.
608 /// GetRelevantLoop - Get the most relevant loop associated with the given
609 /// expression, according to PickMostRelevantLoop.
610 static const Loop *GetRelevantLoop(const SCEV *S, LoopInfo &LI,
612 if (isa<SCEVConstant>(S))
614 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
615 if (const Instruction *I = dyn_cast<Instruction>(U->getValue()))
616 return LI.getLoopFor(I->getParent());
619 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S)) {
621 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
623 for (SCEVNAryExpr::op_iterator I = N->op_begin(), E = N->op_end();
625 L = PickMostRelevantLoop(L, GetRelevantLoop(*I, LI, DT), DT);
628 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
629 return GetRelevantLoop(C->getOperand(), LI, DT);
630 if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S))
631 return PickMostRelevantLoop(GetRelevantLoop(D->getLHS(), LI, DT),
632 GetRelevantLoop(D->getRHS(), LI, DT),
634 llvm_unreachable("Unexpected SCEV type!");
639 /// LoopCompare - Compare loops by PickMostRelevantLoop.
643 explicit LoopCompare(DominatorTree &dt) : DT(dt) {}
645 bool operator()(std::pair<const Loop *, const SCEV *> LHS,
646 std::pair<const Loop *, const SCEV *> RHS) const {
647 // Compare loops with PickMostRelevantLoop.
648 if (LHS.first != RHS.first)
649 return PickMostRelevantLoop(LHS.first, RHS.first, DT) != LHS.first;
651 // If one operand is a non-constant negative and the other is not,
652 // put the non-constant negative on the right so that a sub can
653 // be used instead of a negate and add.
654 if (isNonConstantNegative(LHS.second)) {
655 if (!isNonConstantNegative(RHS.second))
657 } else if (isNonConstantNegative(RHS.second))
660 // Otherwise they are equivalent according to this comparison.
667 Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) {
668 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
670 // Collect all the add operands in a loop, along with their associated loops.
671 // Iterate in reverse so that constants are emitted last, all else equal, and
672 // so that pointer operands are inserted first, which the code below relies on
673 // to form more involved GEPs.
674 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops;
675 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(S->op_end()),
676 E(S->op_begin()); I != E; ++I)
677 OpsAndLoops.push_back(std::make_pair(GetRelevantLoop(*I, *SE.LI, *SE.DT),
680 // Sort by loop. Use a stable sort so that constants follow non-constants and
681 // pointer operands precede non-pointer operands.
682 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT));
684 // Emit instructions to add all the operands. Hoist as much as possible
685 // out of loops, and form meaningful getelementptrs where possible.
687 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator
688 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) {
689 const Loop *CurLoop = I->first;
690 const SCEV *Op = I->second;
692 // This is the first operand. Just expand it.
695 } else if (const PointerType *PTy = dyn_cast<PointerType>(Sum->getType())) {
696 // The running sum expression is a pointer. Try to form a getelementptr
697 // at this level with that as the base.
698 SmallVector<const SCEV *, 4> NewOps;
699 for (; I != E && I->first == CurLoop; ++I)
700 NewOps.push_back(I->second);
701 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, Sum);
702 } else if (const PointerType *PTy = dyn_cast<PointerType>(Op->getType())) {
703 // The running sum is an integer, and there's a pointer at this level.
704 // Try to form a getelementptr. If the running sum is instructions,
705 // use a SCEVUnknown to avoid re-analyzing them.
706 SmallVector<const SCEV *, 4> NewOps;
707 NewOps.push_back(isa<Instruction>(Sum) ? SE.getUnknown(Sum) :
709 for (++I; I != E && I->first == CurLoop; ++I)
710 NewOps.push_back(I->second);
711 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, expand(Op));
712 } else if (isNonConstantNegative(Op)) {
713 // Instead of doing a negate and add, just do a subtract.
714 Value *W = expandCodeFor(SE.getNegativeSCEV(Op), Ty);
715 Sum = InsertNoopCastOfTo(Sum, Ty);
716 Sum = InsertBinop(Instruction::Sub, Sum, W);
720 Value *W = expandCodeFor(Op, Ty);
721 Sum = InsertNoopCastOfTo(Sum, Ty);
722 // Canonicalize a constant to the RHS.
723 if (isa<Constant>(Sum)) std::swap(Sum, W);
724 Sum = InsertBinop(Instruction::Add, Sum, W);
732 Value *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) {
733 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
735 // Collect all the mul operands in a loop, along with their associated loops.
736 // Iterate in reverse so that constants are emitted last, all else equal.
737 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops;
738 for (std::reverse_iterator<SCEVMulExpr::op_iterator> I(S->op_end()),
739 E(S->op_begin()); I != E; ++I)
740 OpsAndLoops.push_back(std::make_pair(GetRelevantLoop(*I, *SE.LI, *SE.DT),
743 // Sort by loop. Use a stable sort so that constants follow non-constants.
744 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT));
746 // Emit instructions to mul all the operands. Hoist as much as possible
749 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator
750 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) {
751 const SCEV *Op = I->second;
753 // This is the first operand. Just expand it.
756 } else if (Op->isAllOnesValue()) {
757 // Instead of doing a multiply by negative one, just do a negate.
758 Prod = InsertNoopCastOfTo(Prod, Ty);
759 Prod = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), Prod);
763 Value *W = expandCodeFor(Op, Ty);
764 Prod = InsertNoopCastOfTo(Prod, Ty);
765 // Canonicalize a constant to the RHS.
766 if (isa<Constant>(Prod)) std::swap(Prod, W);
767 Prod = InsertBinop(Instruction::Mul, Prod, W);
775 Value *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) {
776 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
778 Value *LHS = expandCodeFor(S->getLHS(), Ty);
779 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getRHS())) {
780 const APInt &RHS = SC->getValue()->getValue();
781 if (RHS.isPowerOf2())
782 return InsertBinop(Instruction::LShr, LHS,
783 ConstantInt::get(Ty, RHS.logBase2()));
786 Value *RHS = expandCodeFor(S->getRHS(), Ty);
787 return InsertBinop(Instruction::UDiv, LHS, RHS);
790 /// Move parts of Base into Rest to leave Base with the minimal
791 /// expression that provides a pointer operand suitable for a
793 static void ExposePointerBase(const SCEV *&Base, const SCEV *&Rest,
794 ScalarEvolution &SE) {
795 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Base)) {
796 Base = A->getStart();
797 Rest = SE.getAddExpr(Rest,
798 SE.getAddRecExpr(SE.getConstant(A->getType(), 0),
799 A->getStepRecurrence(SE),
802 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(Base)) {
803 Base = A->getOperand(A->getNumOperands()-1);
804 SmallVector<const SCEV *, 8> NewAddOps(A->op_begin(), A->op_end());
805 NewAddOps.back() = Rest;
806 Rest = SE.getAddExpr(NewAddOps);
807 ExposePointerBase(Base, Rest, SE);
811 /// getAddRecExprPHILiterally - Helper for expandAddRecExprLiterally. Expand
812 /// the base addrec, which is the addrec without any non-loop-dominating
813 /// values, and return the PHI.
815 SCEVExpander::getAddRecExprPHILiterally(const SCEVAddRecExpr *Normalized,
817 const Type *ExpandTy,
819 // Reuse a previously-inserted PHI, if present.
820 for (BasicBlock::iterator I = L->getHeader()->begin();
821 PHINode *PN = dyn_cast<PHINode>(I); ++I)
822 if (SE.isSCEVable(PN->getType()) &&
823 (SE.getEffectiveSCEVType(PN->getType()) ==
824 SE.getEffectiveSCEVType(Normalized->getType())) &&
825 SE.getSCEV(PN) == Normalized)
826 if (BasicBlock *LatchBlock = L->getLoopLatch()) {
828 cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock));
830 // Determine if this is a well-behaved chain of instructions leading
831 // back to the PHI. It probably will be, if we're scanning an inner
832 // loop already visited by LSR for example, but it wouldn't have
835 if (IncV->getNumOperands() == 0 || isa<PHINode>(IncV)) {
839 // If any of the operands don't dominate the insert position, bail.
840 // Addrec operands are always loop-invariant, so this can only happen
841 // if there are instructions which haven't been hoisted.
842 for (User::op_iterator OI = IncV->op_begin()+1,
843 OE = IncV->op_end(); OI != OE; ++OI)
844 if (Instruction *OInst = dyn_cast<Instruction>(OI))
845 if (!SE.DT->dominates(OInst, IVIncInsertPos)) {
851 // Advance to the next instruction.
852 IncV = dyn_cast<Instruction>(IncV->getOperand(0));
855 if (IncV->mayHaveSideEffects()) {
859 } while (IncV != PN);
862 // Ok, the add recurrence looks usable.
863 // Remember this PHI, even in post-inc mode.
864 InsertedValues.insert(PN);
865 // Remember the increment.
866 IncV = cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock));
867 rememberInstruction(IncV);
868 if (L == IVIncInsertLoop)
870 if (SE.DT->dominates(IncV, IVIncInsertPos))
872 // Make sure the increment is where we want it. But don't move it
873 // down past a potential existing post-inc user.
874 IncV->moveBefore(IVIncInsertPos);
875 IVIncInsertPos = IncV;
876 IncV = cast<Instruction>(IncV->getOperand(0));
877 } while (IncV != PN);
882 // Save the original insertion point so we can restore it when we're done.
883 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
884 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
886 // Expand code for the start value.
887 Value *StartV = expandCodeFor(Normalized->getStart(), ExpandTy,
888 L->getHeader()->begin());
890 // Expand code for the step value. Insert instructions right before the
891 // terminator corresponding to the back-edge. Do this before creating the PHI
892 // so that PHI reuse code doesn't see an incomplete PHI. If the stride is
893 // negative, insert a sub instead of an add for the increment (unless it's a
894 // constant, because subtracts of constants are canonicalized to adds).
895 const SCEV *Step = Normalized->getStepRecurrence(SE);
896 bool isPointer = ExpandTy->isPointerTy();
897 bool isNegative = !isPointer && isNonConstantNegative(Step);
899 Step = SE.getNegativeSCEV(Step);
900 Value *StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin());
903 Builder.SetInsertPoint(L->getHeader(), L->getHeader()->begin());
904 PHINode *PN = Builder.CreatePHI(ExpandTy, "lsr.iv");
905 rememberInstruction(PN);
907 // Create the step instructions and populate the PHI.
908 BasicBlock *Header = L->getHeader();
909 for (pred_iterator HPI = pred_begin(Header), HPE = pred_end(Header);
911 BasicBlock *Pred = *HPI;
913 // Add a start value.
914 if (!L->contains(Pred)) {
915 PN->addIncoming(StartV, Pred);
919 // Create a step value and add it to the PHI. If IVIncInsertLoop is
920 // non-null and equal to the addrec's loop, insert the instructions
921 // at IVIncInsertPos.
922 Instruction *InsertPos = L == IVIncInsertLoop ?
923 IVIncInsertPos : Pred->getTerminator();
924 Builder.SetInsertPoint(InsertPos->getParent(), InsertPos);
926 // If the PHI is a pointer, use a GEP, otherwise use an add or sub.
928 const PointerType *GEPPtrTy = cast<PointerType>(ExpandTy);
929 // If the step isn't constant, don't use an implicitly scaled GEP, because
930 // that would require a multiply inside the loop.
931 if (!isa<ConstantInt>(StepV))
932 GEPPtrTy = PointerType::get(Type::getInt1Ty(SE.getContext()),
933 GEPPtrTy->getAddressSpace());
934 const SCEV *const StepArray[1] = { SE.getSCEV(StepV) };
935 IncV = expandAddToGEP(StepArray, StepArray+1, GEPPtrTy, IntTy, PN);
936 if (IncV->getType() != PN->getType()) {
937 IncV = Builder.CreateBitCast(IncV, PN->getType(), "tmp");
938 rememberInstruction(IncV);
942 Builder.CreateSub(PN, StepV, "lsr.iv.next") :
943 Builder.CreateAdd(PN, StepV, "lsr.iv.next");
944 rememberInstruction(IncV);
946 PN->addIncoming(IncV, Pred);
949 // Restore the original insert point.
951 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
953 // Remember this PHI, even in post-inc mode.
954 InsertedValues.insert(PN);
959 Value *SCEVExpander::expandAddRecExprLiterally(const SCEVAddRecExpr *S) {
960 const Type *STy = S->getType();
961 const Type *IntTy = SE.getEffectiveSCEVType(STy);
962 const Loop *L = S->getLoop();
964 // Determine a normalized form of this expression, which is the expression
965 // before any post-inc adjustment is made.
966 const SCEVAddRecExpr *Normalized = S;
967 if (PostIncLoops.count(L)) {
968 PostIncLoopSet Loops;
971 cast<SCEVAddRecExpr>(TransformForPostIncUse(Normalize, S, 0, 0,
975 // Strip off any non-loop-dominating component from the addrec start.
976 const SCEV *Start = Normalized->getStart();
977 const SCEV *PostLoopOffset = 0;
978 if (!Start->properlyDominates(L->getHeader(), SE.DT)) {
979 PostLoopOffset = Start;
980 Start = SE.getConstant(Normalized->getType(), 0);
982 cast<SCEVAddRecExpr>(SE.getAddRecExpr(Start,
983 Normalized->getStepRecurrence(SE),
984 Normalized->getLoop()));
987 // Strip off any non-loop-dominating component from the addrec step.
988 const SCEV *Step = Normalized->getStepRecurrence(SE);
989 const SCEV *PostLoopScale = 0;
990 if (!Step->dominates(L->getHeader(), SE.DT)) {
991 PostLoopScale = Step;
992 Step = SE.getConstant(Normalized->getType(), 1);
994 cast<SCEVAddRecExpr>(SE.getAddRecExpr(Start, Step,
995 Normalized->getLoop()));
998 // Expand the core addrec. If we need post-loop scaling, force it to
999 // expand to an integer type to avoid the need for additional casting.
1000 const Type *ExpandTy = PostLoopScale ? IntTy : STy;
1001 PHINode *PN = getAddRecExprPHILiterally(Normalized, L, ExpandTy, IntTy);
1003 // Accommodate post-inc mode, if necessary.
1005 if (!PostIncLoops.count(L))
1008 // In PostInc mode, use the post-incremented value.
1009 BasicBlock *LatchBlock = L->getLoopLatch();
1010 assert(LatchBlock && "PostInc mode requires a unique loop latch!");
1011 Result = PN->getIncomingValueForBlock(LatchBlock);
1014 // Re-apply any non-loop-dominating scale.
1015 if (PostLoopScale) {
1016 Result = InsertNoopCastOfTo(Result, IntTy);
1017 Result = Builder.CreateMul(Result,
1018 expandCodeFor(PostLoopScale, IntTy));
1019 rememberInstruction(Result);
1022 // Re-apply any non-loop-dominating offset.
1023 if (PostLoopOffset) {
1024 if (const PointerType *PTy = dyn_cast<PointerType>(ExpandTy)) {
1025 const SCEV *const OffsetArray[1] = { PostLoopOffset };
1026 Result = expandAddToGEP(OffsetArray, OffsetArray+1, PTy, IntTy, Result);
1028 Result = InsertNoopCastOfTo(Result, IntTy);
1029 Result = Builder.CreateAdd(Result,
1030 expandCodeFor(PostLoopOffset, IntTy));
1031 rememberInstruction(Result);
1038 Value *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) {
1039 if (!CanonicalMode) return expandAddRecExprLiterally(S);
1041 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
1042 const Loop *L = S->getLoop();
1044 // First check for an existing canonical IV in a suitable type.
1045 PHINode *CanonicalIV = 0;
1046 if (PHINode *PN = L->getCanonicalInductionVariable())
1047 if (SE.isSCEVable(PN->getType()) &&
1048 SE.getEffectiveSCEVType(PN->getType())->isIntegerTy() &&
1049 SE.getTypeSizeInBits(PN->getType()) >= SE.getTypeSizeInBits(Ty))
1052 // Rewrite an AddRec in terms of the canonical induction variable, if
1053 // its type is more narrow.
1055 SE.getTypeSizeInBits(CanonicalIV->getType()) >
1056 SE.getTypeSizeInBits(Ty)) {
1057 SmallVector<const SCEV *, 4> NewOps(S->getNumOperands());
1058 for (unsigned i = 0, e = S->getNumOperands(); i != e; ++i)
1059 NewOps[i] = SE.getAnyExtendExpr(S->op_begin()[i], CanonicalIV->getType());
1060 Value *V = expand(SE.getAddRecExpr(NewOps, S->getLoop()));
1061 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1062 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1063 BasicBlock::iterator NewInsertPt =
1064 llvm::next(BasicBlock::iterator(cast<Instruction>(V)));
1065 while (isa<PHINode>(NewInsertPt) || isa<DbgInfoIntrinsic>(NewInsertPt))
1067 V = expandCodeFor(SE.getTruncateExpr(SE.getUnknown(V), Ty), 0,
1069 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
1073 // {X,+,F} --> X + {0,+,F}
1074 if (!S->getStart()->isZero()) {
1075 SmallVector<const SCEV *, 4> NewOps(S->op_begin(), S->op_end());
1076 NewOps[0] = SE.getConstant(Ty, 0);
1077 const SCEV *Rest = SE.getAddRecExpr(NewOps, L);
1079 // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the
1080 // comments on expandAddToGEP for details.
1081 const SCEV *Base = S->getStart();
1082 const SCEV *RestArray[1] = { Rest };
1083 // Dig into the expression to find the pointer base for a GEP.
1084 ExposePointerBase(Base, RestArray[0], SE);
1085 // If we found a pointer, expand the AddRec with a GEP.
1086 if (const PointerType *PTy = dyn_cast<PointerType>(Base->getType())) {
1087 // Make sure the Base isn't something exotic, such as a multiplied
1088 // or divided pointer value. In those cases, the result type isn't
1089 // actually a pointer type.
1090 if (!isa<SCEVMulExpr>(Base) && !isa<SCEVUDivExpr>(Base)) {
1091 Value *StartV = expand(Base);
1092 assert(StartV->getType() == PTy && "Pointer type mismatch for GEP!");
1093 return expandAddToGEP(RestArray, RestArray+1, PTy, Ty, StartV);
1097 // Just do a normal add. Pre-expand the operands to suppress folding.
1098 return expand(SE.getAddExpr(SE.getUnknown(expand(S->getStart())),
1099 SE.getUnknown(expand(Rest))));
1102 // {0,+,1} --> Insert a canonical induction variable into the loop!
1103 if (S->isAffine() &&
1104 S->getOperand(1) == SE.getConstant(Ty, 1)) {
1105 // If there's a canonical IV, just use it.
1107 assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) &&
1108 "IVs with types different from the canonical IV should "
1109 "already have been handled!");
1113 // Create and insert the PHI node for the induction variable in the
1115 BasicBlock *Header = L->getHeader();
1116 PHINode *PN = PHINode::Create(Ty, "indvar", Header->begin());
1117 rememberInstruction(PN);
1119 Constant *One = ConstantInt::get(Ty, 1);
1120 for (pred_iterator HPI = pred_begin(Header), HPE = pred_end(Header);
1122 if (L->contains(*HPI)) {
1123 // Insert a unit add instruction right before the terminator
1124 // corresponding to the back-edge.
1125 Instruction *Add = BinaryOperator::CreateAdd(PN, One, "indvar.next",
1126 (*HPI)->getTerminator());
1127 rememberInstruction(Add);
1128 PN->addIncoming(Add, *HPI);
1130 PN->addIncoming(Constant::getNullValue(Ty), *HPI);
1134 // {0,+,F} --> {0,+,1} * F
1135 // Get the canonical induction variable I for this loop.
1136 Value *I = CanonicalIV ?
1138 getOrInsertCanonicalInductionVariable(L, Ty);
1140 // If this is a simple linear addrec, emit it now as a special case.
1141 if (S->isAffine()) // {0,+,F} --> i*F
1143 expand(SE.getTruncateOrNoop(
1144 SE.getMulExpr(SE.getUnknown(I),
1145 SE.getNoopOrAnyExtend(S->getOperand(1),
1149 // If this is a chain of recurrences, turn it into a closed form, using the
1150 // folders, then expandCodeFor the closed form. This allows the folders to
1151 // simplify the expression without having to build a bunch of special code
1152 // into this folder.
1153 const SCEV *IH = SE.getUnknown(I); // Get I as a "symbolic" SCEV.
1155 // Promote S up to the canonical IV type, if the cast is foldable.
1156 const SCEV *NewS = S;
1157 const SCEV *Ext = SE.getNoopOrAnyExtend(S, I->getType());
1158 if (isa<SCEVAddRecExpr>(Ext))
1161 const SCEV *V = cast<SCEVAddRecExpr>(NewS)->evaluateAtIteration(IH, SE);
1162 //cerr << "Evaluated: " << *this << "\n to: " << *V << "\n";
1164 // Truncate the result down to the original type, if needed.
1165 const SCEV *T = SE.getTruncateOrNoop(V, Ty);
1169 Value *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) {
1170 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
1171 Value *V = expandCodeFor(S->getOperand(),
1172 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1173 Value *I = Builder.CreateTrunc(V, Ty, "tmp");
1174 rememberInstruction(I);
1178 Value *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) {
1179 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
1180 Value *V = expandCodeFor(S->getOperand(),
1181 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1182 Value *I = Builder.CreateZExt(V, Ty, "tmp");
1183 rememberInstruction(I);
1187 Value *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) {
1188 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
1189 Value *V = expandCodeFor(S->getOperand(),
1190 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1191 Value *I = Builder.CreateSExt(V, Ty, "tmp");
1192 rememberInstruction(I);
1196 Value *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) {
1197 Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
1198 const Type *Ty = LHS->getType();
1199 for (int i = S->getNumOperands()-2; i >= 0; --i) {
1200 // In the case of mixed integer and pointer types, do the
1201 // rest of the comparisons as integer.
1202 if (S->getOperand(i)->getType() != Ty) {
1203 Ty = SE.getEffectiveSCEVType(Ty);
1204 LHS = InsertNoopCastOfTo(LHS, Ty);
1206 Value *RHS = expandCodeFor(S->getOperand(i), Ty);
1207 Value *ICmp = Builder.CreateICmpSGT(LHS, RHS, "tmp");
1208 rememberInstruction(ICmp);
1209 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "smax");
1210 rememberInstruction(Sel);
1213 // In the case of mixed integer and pointer types, cast the
1214 // final result back to the pointer type.
1215 if (LHS->getType() != S->getType())
1216 LHS = InsertNoopCastOfTo(LHS, S->getType());
1220 Value *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) {
1221 Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
1222 const Type *Ty = LHS->getType();
1223 for (int i = S->getNumOperands()-2; i >= 0; --i) {
1224 // In the case of mixed integer and pointer types, do the
1225 // rest of the comparisons as integer.
1226 if (S->getOperand(i)->getType() != Ty) {
1227 Ty = SE.getEffectiveSCEVType(Ty);
1228 LHS = InsertNoopCastOfTo(LHS, Ty);
1230 Value *RHS = expandCodeFor(S->getOperand(i), Ty);
1231 Value *ICmp = Builder.CreateICmpUGT(LHS, RHS, "tmp");
1232 rememberInstruction(ICmp);
1233 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "umax");
1234 rememberInstruction(Sel);
1237 // In the case of mixed integer and pointer types, cast the
1238 // final result back to the pointer type.
1239 if (LHS->getType() != S->getType())
1240 LHS = InsertNoopCastOfTo(LHS, S->getType());
1244 Value *SCEVExpander::expandCodeFor(const SCEV *SH, const Type *Ty,
1246 BasicBlock::iterator IP = I;
1247 while (isInsertedInstruction(IP) || isa<DbgInfoIntrinsic>(IP))
1249 Builder.SetInsertPoint(IP->getParent(), IP);
1250 return expandCodeFor(SH, Ty);
1253 Value *SCEVExpander::expandCodeFor(const SCEV *SH, const Type *Ty) {
1254 // Expand the code for this SCEV.
1255 Value *V = expand(SH);
1257 assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) &&
1258 "non-trivial casts should be done with the SCEVs directly!");
1259 V = InsertNoopCastOfTo(V, Ty);
1264 Value *SCEVExpander::expand(const SCEV *S) {
1265 // Compute an insertion point for this SCEV object. Hoist the instructions
1266 // as far out in the loop nest as possible.
1267 Instruction *InsertPt = Builder.GetInsertPoint();
1268 for (Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock()); ;
1269 L = L->getParentLoop())
1270 if (S->isLoopInvariant(L)) {
1272 if (BasicBlock *Preheader = L->getLoopPreheader())
1273 InsertPt = Preheader->getTerminator();
1275 // If the SCEV is computable at this level, insert it into the header
1276 // after the PHIs (and after any other instructions that we've inserted
1277 // there) so that it is guaranteed to dominate any user inside the loop.
1278 if (L && S->hasComputableLoopEvolution(L) && !PostIncLoops.count(L))
1279 InsertPt = L->getHeader()->getFirstNonPHI();
1280 while (isInsertedInstruction(InsertPt) || isa<DbgInfoIntrinsic>(InsertPt))
1281 InsertPt = llvm::next(BasicBlock::iterator(InsertPt));
1285 // Check to see if we already expanded this here.
1286 std::map<std::pair<const SCEV *, Instruction *>,
1287 AssertingVH<Value> >::iterator I =
1288 InsertedExpressions.find(std::make_pair(S, InsertPt));
1289 if (I != InsertedExpressions.end())
1292 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1293 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1294 Builder.SetInsertPoint(InsertPt->getParent(), InsertPt);
1296 // Expand the expression into instructions.
1297 Value *V = visit(S);
1299 // Remember the expanded value for this SCEV at this location.
1300 if (PostIncLoops.empty())
1301 InsertedExpressions[std::make_pair(S, InsertPt)] = V;
1303 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
1307 void SCEVExpander::rememberInstruction(Value *I) {
1308 if (!PostIncLoops.empty())
1309 InsertedPostIncValues.insert(I);
1311 InsertedValues.insert(I);
1313 // If we just claimed an existing instruction and that instruction had
1314 // been the insert point, adjust the insert point forward so that
1315 // subsequently inserted code will be dominated.
1316 if (Builder.GetInsertPoint() == I) {
1317 BasicBlock::iterator It = cast<Instruction>(I);
1318 do { ++It; } while (isInsertedInstruction(It) ||
1319 isa<DbgInfoIntrinsic>(It));
1320 Builder.SetInsertPoint(Builder.GetInsertBlock(), It);
1324 void SCEVExpander::restoreInsertPoint(BasicBlock *BB, BasicBlock::iterator I) {
1325 // If we acquired more instructions since the old insert point was saved,
1326 // advance past them.
1327 while (isInsertedInstruction(I) || isa<DbgInfoIntrinsic>(I)) ++I;
1329 Builder.SetInsertPoint(BB, I);
1332 /// getOrInsertCanonicalInductionVariable - This method returns the
1333 /// canonical induction variable of the specified type for the specified
1334 /// loop (inserting one if there is none). A canonical induction variable
1335 /// starts at zero and steps by one on each iteration.
1337 SCEVExpander::getOrInsertCanonicalInductionVariable(const Loop *L,
1339 assert(Ty->isIntegerTy() && "Can only insert integer induction variables!");
1340 const SCEV *H = SE.getAddRecExpr(SE.getConstant(Ty, 0),
1341 SE.getConstant(Ty, 1), L);
1342 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1343 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1344 Value *V = expandCodeFor(H, 0, L->getHeader()->begin());
1346 restoreInsertPoint(SaveInsertBB, SaveInsertPt);