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/Support/Debug.h"
21 #include "llvm/Target/TargetData.h"
22 #include "llvm/Target/TargetLowering.h"
23 #include "llvm/ADT/STLExtras.h"
27 /// ReuseOrCreateCast - Arrange for there to be a cast of V to Ty at IP,
28 /// reusing an existing cast if a suitable one exists, moving an existing
29 /// cast if a suitable one exists but isn't in the right place, or
30 /// creating a new one.
31 Value *SCEVExpander::ReuseOrCreateCast(Value *V, Type *Ty,
32 Instruction::CastOps Op,
33 BasicBlock::iterator IP) {
34 // All new or reused instructions must strictly dominate the Builder's
35 // InsertPt to ensure that the expression's expansion dominates its uses.
36 // Assert that the requested insertion point works at least for new
38 assert(SE.DT->dominates(IP, Builder.GetInsertPoint()));
40 // Check to see if there is already a cast!
41 for (Value::use_iterator UI = V->use_begin(), E = V->use_end();
44 if (U->getType() == Ty)
45 if (CastInst *CI = dyn_cast<CastInst>(U))
46 if (CI->getOpcode() == Op) {
47 // If the cast isn't where we want it, fix it.
48 if (BasicBlock::iterator(CI) != IP
49 || IP == Builder.GetInsertPoint()) {
50 // Create a new cast, and leave the old cast in place in case
51 // it is being used as an insert point. Clear its operand
52 // so that it doesn't hold anything live.
53 Instruction *NewCI = CastInst::Create(Op, V, Ty, "", IP);
55 CI->replaceAllUsesWith(NewCI);
56 CI->setOperand(0, UndefValue::get(V->getType()));
57 rememberInstruction(NewCI);
60 rememberInstruction(CI);
66 Instruction *I = CastInst::Create(Op, V, Ty, V->getName(), IP);
67 rememberInstruction(I);
71 /// InsertNoopCastOfTo - Insert a cast of V to the specified type,
72 /// which must be possible with a noop cast, doing what we can to share
74 Value *SCEVExpander::InsertNoopCastOfTo(Value *V, Type *Ty) {
75 Instruction::CastOps Op = CastInst::getCastOpcode(V, false, Ty, false);
76 assert((Op == Instruction::BitCast ||
77 Op == Instruction::PtrToInt ||
78 Op == Instruction::IntToPtr) &&
79 "InsertNoopCastOfTo cannot perform non-noop casts!");
80 assert(SE.getTypeSizeInBits(V->getType()) == SE.getTypeSizeInBits(Ty) &&
81 "InsertNoopCastOfTo cannot change sizes!");
83 // Short-circuit unnecessary bitcasts.
84 if (Op == Instruction::BitCast) {
85 if (V->getType() == Ty)
87 if (CastInst *CI = dyn_cast<CastInst>(V)) {
88 if (CI->getOperand(0)->getType() == Ty)
89 return CI->getOperand(0);
92 // Short-circuit unnecessary inttoptr<->ptrtoint casts.
93 if ((Op == Instruction::PtrToInt || Op == Instruction::IntToPtr) &&
94 SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(V->getType())) {
95 if (CastInst *CI = dyn_cast<CastInst>(V))
96 if ((CI->getOpcode() == Instruction::PtrToInt ||
97 CI->getOpcode() == Instruction::IntToPtr) &&
98 SE.getTypeSizeInBits(CI->getType()) ==
99 SE.getTypeSizeInBits(CI->getOperand(0)->getType()))
100 return CI->getOperand(0);
101 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
102 if ((CE->getOpcode() == Instruction::PtrToInt ||
103 CE->getOpcode() == Instruction::IntToPtr) &&
104 SE.getTypeSizeInBits(CE->getType()) ==
105 SE.getTypeSizeInBits(CE->getOperand(0)->getType()))
106 return CE->getOperand(0);
109 // Fold a cast of a constant.
110 if (Constant *C = dyn_cast<Constant>(V))
111 return ConstantExpr::getCast(Op, C, Ty);
113 // Cast the argument at the beginning of the entry block, after
114 // any bitcasts of other arguments.
115 if (Argument *A = dyn_cast<Argument>(V)) {
116 BasicBlock::iterator IP = A->getParent()->getEntryBlock().begin();
117 while ((isa<BitCastInst>(IP) &&
118 isa<Argument>(cast<BitCastInst>(IP)->getOperand(0)) &&
119 cast<BitCastInst>(IP)->getOperand(0) != A) ||
120 isa<DbgInfoIntrinsic>(IP) ||
121 isa<LandingPadInst>(IP))
123 return ReuseOrCreateCast(A, Ty, Op, IP);
126 // Cast the instruction immediately after the instruction.
127 Instruction *I = cast<Instruction>(V);
128 BasicBlock::iterator IP = I; ++IP;
129 if (InvokeInst *II = dyn_cast<InvokeInst>(I))
130 IP = II->getNormalDest()->begin();
131 while (isa<PHINode>(IP) || isa<LandingPadInst>(IP))
133 return ReuseOrCreateCast(I, Ty, Op, IP);
136 /// InsertBinop - Insert the specified binary operator, doing a small amount
137 /// of work to avoid inserting an obviously redundant operation.
138 Value *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode,
139 Value *LHS, Value *RHS) {
140 // Fold a binop with constant operands.
141 if (Constant *CLHS = dyn_cast<Constant>(LHS))
142 if (Constant *CRHS = dyn_cast<Constant>(RHS))
143 return ConstantExpr::get(Opcode, CLHS, CRHS);
145 // Do a quick scan to see if we have this binop nearby. If so, reuse it.
146 unsigned ScanLimit = 6;
147 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
148 // Scanning starts from the last instruction before the insertion point.
149 BasicBlock::iterator IP = Builder.GetInsertPoint();
150 if (IP != BlockBegin) {
152 for (; ScanLimit; --IP, --ScanLimit) {
153 // Don't count dbg.value against the ScanLimit, to avoid perturbing the
155 if (isa<DbgInfoIntrinsic>(IP))
157 if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(0) == LHS &&
158 IP->getOperand(1) == RHS)
160 if (IP == BlockBegin) break;
164 // Save the original insertion point so we can restore it when we're done.
165 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
166 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
168 // Move the insertion point out of as many loops as we can.
169 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
170 if (!L->isLoopInvariant(LHS) || !L->isLoopInvariant(RHS)) break;
171 BasicBlock *Preheader = L->getLoopPreheader();
172 if (!Preheader) break;
174 // Ok, move up a level.
175 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
178 // If we haven't found this binop, insert it.
179 Instruction *BO = cast<Instruction>(Builder.CreateBinOp(Opcode, LHS, RHS));
180 BO->setDebugLoc(SaveInsertPt->getDebugLoc());
181 rememberInstruction(BO);
183 // Restore the original insert point.
185 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
190 /// FactorOutConstant - Test if S is divisible by Factor, using signed
191 /// division. If so, update S with Factor divided out and return true.
192 /// S need not be evenly divisible if a reasonable remainder can be
194 /// TODO: When ScalarEvolution gets a SCEVSDivExpr, this can be made
195 /// unnecessary; in its place, just signed-divide Ops[i] by the scale and
196 /// check to see if the divide was folded.
197 static bool FactorOutConstant(const SCEV *&S,
198 const SCEV *&Remainder,
201 const TargetData *TD) {
202 // Everything is divisible by one.
208 S = SE.getConstant(S->getType(), 1);
212 // For a Constant, check for a multiple of the given factor.
213 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
217 // Check for divisibility.
218 if (const SCEVConstant *FC = dyn_cast<SCEVConstant>(Factor)) {
220 ConstantInt::get(SE.getContext(),
221 C->getValue()->getValue().sdiv(
222 FC->getValue()->getValue()));
223 // If the quotient is zero and the remainder is non-zero, reject
224 // the value at this scale. It will be considered for subsequent
227 const SCEV *Div = SE.getConstant(CI);
230 SE.getAddExpr(Remainder,
231 SE.getConstant(C->getValue()->getValue().srem(
232 FC->getValue()->getValue())));
238 // In a Mul, check if there is a constant operand which is a multiple
239 // of the given factor.
240 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
242 // With TargetData, the size is known. Check if there is a constant
243 // operand which is a multiple of the given factor. If so, we can
245 const SCEVConstant *FC = cast<SCEVConstant>(Factor);
246 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0)))
247 if (!C->getValue()->getValue().srem(FC->getValue()->getValue())) {
248 SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end());
250 SE.getConstant(C->getValue()->getValue().sdiv(
251 FC->getValue()->getValue()));
252 S = SE.getMulExpr(NewMulOps);
256 // Without TargetData, check if Factor can be factored out of any of the
257 // Mul's operands. If so, we can just remove it.
258 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) {
259 const SCEV *SOp = M->getOperand(i);
260 const SCEV *Remainder = SE.getConstant(SOp->getType(), 0);
261 if (FactorOutConstant(SOp, Remainder, Factor, SE, TD) &&
262 Remainder->isZero()) {
263 SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end());
265 S = SE.getMulExpr(NewMulOps);
272 // In an AddRec, check if both start and step are divisible.
273 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
274 const SCEV *Step = A->getStepRecurrence(SE);
275 const SCEV *StepRem = SE.getConstant(Step->getType(), 0);
276 if (!FactorOutConstant(Step, StepRem, Factor, SE, TD))
278 if (!StepRem->isZero())
280 const SCEV *Start = A->getStart();
281 if (!FactorOutConstant(Start, Remainder, Factor, SE, TD))
283 // FIXME: can use A->getNoWrapFlags(FlagNW)
284 S = SE.getAddRecExpr(Start, Step, A->getLoop(), SCEV::FlagAnyWrap);
291 /// SimplifyAddOperands - Sort and simplify a list of add operands. NumAddRecs
292 /// is the number of SCEVAddRecExprs present, which are kept at the end of
295 static void SimplifyAddOperands(SmallVectorImpl<const SCEV *> &Ops,
297 ScalarEvolution &SE) {
298 unsigned NumAddRecs = 0;
299 for (unsigned i = Ops.size(); i > 0 && isa<SCEVAddRecExpr>(Ops[i-1]); --i)
301 // Group Ops into non-addrecs and addrecs.
302 SmallVector<const SCEV *, 8> NoAddRecs(Ops.begin(), Ops.end() - NumAddRecs);
303 SmallVector<const SCEV *, 8> AddRecs(Ops.end() - NumAddRecs, Ops.end());
304 // Let ScalarEvolution sort and simplify the non-addrecs list.
305 const SCEV *Sum = NoAddRecs.empty() ?
306 SE.getConstant(Ty, 0) :
307 SE.getAddExpr(NoAddRecs);
308 // If it returned an add, use the operands. Otherwise it simplified
309 // the sum into a single value, so just use that.
311 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Sum))
312 Ops.append(Add->op_begin(), Add->op_end());
313 else if (!Sum->isZero())
315 // Then append the addrecs.
316 Ops.append(AddRecs.begin(), AddRecs.end());
319 /// SplitAddRecs - Flatten a list of add operands, moving addrec start values
320 /// out to the top level. For example, convert {a + b,+,c} to a, b, {0,+,d}.
321 /// This helps expose more opportunities for folding parts of the expressions
322 /// into GEP indices.
324 static void SplitAddRecs(SmallVectorImpl<const SCEV *> &Ops,
326 ScalarEvolution &SE) {
328 SmallVector<const SCEV *, 8> AddRecs;
329 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
330 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Ops[i])) {
331 const SCEV *Start = A->getStart();
332 if (Start->isZero()) break;
333 const SCEV *Zero = SE.getConstant(Ty, 0);
334 AddRecs.push_back(SE.getAddRecExpr(Zero,
335 A->getStepRecurrence(SE),
337 // FIXME: A->getNoWrapFlags(FlagNW)
339 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Start)) {
341 Ops.append(Add->op_begin(), Add->op_end());
342 e += Add->getNumOperands();
347 if (!AddRecs.empty()) {
348 // Add the addrecs onto the end of the list.
349 Ops.append(AddRecs.begin(), AddRecs.end());
350 // Resort the operand list, moving any constants to the front.
351 SimplifyAddOperands(Ops, Ty, SE);
355 /// expandAddToGEP - Expand an addition expression with a pointer type into
356 /// a GEP instead of using ptrtoint+arithmetic+inttoptr. This helps
357 /// BasicAliasAnalysis and other passes analyze the result. See the rules
358 /// for getelementptr vs. inttoptr in
359 /// http://llvm.org/docs/LangRef.html#pointeraliasing
362 /// Design note: The correctness of using getelementptr here depends on
363 /// ScalarEvolution not recognizing inttoptr and ptrtoint operators, as
364 /// they may introduce pointer arithmetic which may not be safely converted
365 /// into getelementptr.
367 /// Design note: It might seem desirable for this function to be more
368 /// loop-aware. If some of the indices are loop-invariant while others
369 /// aren't, it might seem desirable to emit multiple GEPs, keeping the
370 /// loop-invariant portions of the overall computation outside the loop.
371 /// However, there are a few reasons this is not done here. Hoisting simple
372 /// arithmetic is a low-level optimization that often isn't very
373 /// important until late in the optimization process. In fact, passes
374 /// like InstructionCombining will combine GEPs, even if it means
375 /// pushing loop-invariant computation down into loops, so even if the
376 /// GEPs were split here, the work would quickly be undone. The
377 /// LoopStrengthReduction pass, which is usually run quite late (and
378 /// after the last InstructionCombining pass), takes care of hoisting
379 /// loop-invariant portions of expressions, after considering what
380 /// can be folded using target addressing modes.
382 Value *SCEVExpander::expandAddToGEP(const SCEV *const *op_begin,
383 const SCEV *const *op_end,
387 Type *ElTy = PTy->getElementType();
388 SmallVector<Value *, 4> GepIndices;
389 SmallVector<const SCEV *, 8> Ops(op_begin, op_end);
390 bool AnyNonZeroIndices = false;
392 // Split AddRecs up into parts as either of the parts may be usable
393 // without the other.
394 SplitAddRecs(Ops, Ty, SE);
396 // Descend down the pointer's type and attempt to convert the other
397 // operands into GEP indices, at each level. The first index in a GEP
398 // indexes into the array implied by the pointer operand; the rest of
399 // the indices index into the element or field type selected by the
402 // If the scale size is not 0, attempt to factor out a scale for
404 SmallVector<const SCEV *, 8> ScaledOps;
405 if (ElTy->isSized()) {
406 const SCEV *ElSize = SE.getSizeOfExpr(ElTy);
407 if (!ElSize->isZero()) {
408 SmallVector<const SCEV *, 8> NewOps;
409 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
410 const SCEV *Op = Ops[i];
411 const SCEV *Remainder = SE.getConstant(Ty, 0);
412 if (FactorOutConstant(Op, Remainder, ElSize, SE, SE.TD)) {
413 // Op now has ElSize factored out.
414 ScaledOps.push_back(Op);
415 if (!Remainder->isZero())
416 NewOps.push_back(Remainder);
417 AnyNonZeroIndices = true;
419 // The operand was not divisible, so add it to the list of operands
420 // we'll scan next iteration.
421 NewOps.push_back(Ops[i]);
424 // If we made any changes, update Ops.
425 if (!ScaledOps.empty()) {
427 SimplifyAddOperands(Ops, Ty, SE);
432 // Record the scaled array index for this level of the type. If
433 // we didn't find any operands that could be factored, tentatively
434 // assume that element zero was selected (since the zero offset
435 // would obviously be folded away).
436 Value *Scaled = ScaledOps.empty() ?
437 Constant::getNullValue(Ty) :
438 expandCodeFor(SE.getAddExpr(ScaledOps), Ty);
439 GepIndices.push_back(Scaled);
441 // Collect struct field index operands.
442 while (StructType *STy = dyn_cast<StructType>(ElTy)) {
443 bool FoundFieldNo = false;
444 // An empty struct has no fields.
445 if (STy->getNumElements() == 0) break;
447 // With TargetData, field offsets are known. See if a constant offset
448 // falls within any of the struct fields.
449 if (Ops.empty()) break;
450 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[0]))
451 if (SE.getTypeSizeInBits(C->getType()) <= 64) {
452 const StructLayout &SL = *SE.TD->getStructLayout(STy);
453 uint64_t FullOffset = C->getValue()->getZExtValue();
454 if (FullOffset < SL.getSizeInBytes()) {
455 unsigned ElIdx = SL.getElementContainingOffset(FullOffset);
456 GepIndices.push_back(
457 ConstantInt::get(Type::getInt32Ty(Ty->getContext()), ElIdx));
458 ElTy = STy->getTypeAtIndex(ElIdx);
460 SE.getConstant(Ty, FullOffset - SL.getElementOffset(ElIdx));
461 AnyNonZeroIndices = true;
466 // Without TargetData, just check for an offsetof expression of the
467 // appropriate struct type.
468 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
469 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Ops[i])) {
472 if (U->isOffsetOf(CTy, FieldNo) && CTy == STy) {
473 GepIndices.push_back(FieldNo);
475 STy->getTypeAtIndex(cast<ConstantInt>(FieldNo)->getZExtValue());
476 Ops[i] = SE.getConstant(Ty, 0);
477 AnyNonZeroIndices = true;
483 // If no struct field offsets were found, tentatively assume that
484 // field zero was selected (since the zero offset would obviously
487 ElTy = STy->getTypeAtIndex(0u);
488 GepIndices.push_back(
489 Constant::getNullValue(Type::getInt32Ty(Ty->getContext())));
493 if (ArrayType *ATy = dyn_cast<ArrayType>(ElTy))
494 ElTy = ATy->getElementType();
499 // If none of the operands were convertible to proper GEP indices, cast
500 // the base to i8* and do an ugly getelementptr with that. It's still
501 // better than ptrtoint+arithmetic+inttoptr at least.
502 if (!AnyNonZeroIndices) {
503 // Cast the base to i8*.
504 V = InsertNoopCastOfTo(V,
505 Type::getInt8PtrTy(Ty->getContext(), PTy->getAddressSpace()));
507 // Expand the operands for a plain byte offset.
508 Value *Idx = expandCodeFor(SE.getAddExpr(Ops), Ty);
510 // Fold a GEP with constant operands.
511 if (Constant *CLHS = dyn_cast<Constant>(V))
512 if (Constant *CRHS = dyn_cast<Constant>(Idx))
513 return ConstantExpr::getGetElementPtr(CLHS, CRHS);
515 // Do a quick scan to see if we have this GEP nearby. If so, reuse it.
516 unsigned ScanLimit = 6;
517 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
518 // Scanning starts from the last instruction before the insertion point.
519 BasicBlock::iterator IP = Builder.GetInsertPoint();
520 if (IP != BlockBegin) {
522 for (; ScanLimit; --IP, --ScanLimit) {
523 // Don't count dbg.value against the ScanLimit, to avoid perturbing the
525 if (isa<DbgInfoIntrinsic>(IP))
527 if (IP->getOpcode() == Instruction::GetElementPtr &&
528 IP->getOperand(0) == V && IP->getOperand(1) == Idx)
530 if (IP == BlockBegin) break;
534 // Save the original insertion point so we can restore it when we're done.
535 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
536 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
538 // Move the insertion point out of as many loops as we can.
539 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
540 if (!L->isLoopInvariant(V) || !L->isLoopInvariant(Idx)) break;
541 BasicBlock *Preheader = L->getLoopPreheader();
542 if (!Preheader) break;
544 // Ok, move up a level.
545 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
549 Value *GEP = Builder.CreateGEP(V, Idx, "uglygep");
550 rememberInstruction(GEP);
552 // Restore the original insert point.
554 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
559 // Save the original insertion point so we can restore it when we're done.
560 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
561 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
563 // Move the insertion point out of as many loops as we can.
564 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
565 if (!L->isLoopInvariant(V)) break;
567 bool AnyIndexNotLoopInvariant = false;
568 for (SmallVectorImpl<Value *>::const_iterator I = GepIndices.begin(),
569 E = GepIndices.end(); I != E; ++I)
570 if (!L->isLoopInvariant(*I)) {
571 AnyIndexNotLoopInvariant = true;
574 if (AnyIndexNotLoopInvariant)
577 BasicBlock *Preheader = L->getLoopPreheader();
578 if (!Preheader) break;
580 // Ok, move up a level.
581 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
584 // Insert a pretty getelementptr. Note that this GEP is not marked inbounds,
585 // because ScalarEvolution may have changed the address arithmetic to
586 // compute a value which is beyond the end of the allocated object.
588 if (V->getType() != PTy)
589 Casted = InsertNoopCastOfTo(Casted, PTy);
590 Value *GEP = Builder.CreateGEP(Casted,
593 Ops.push_back(SE.getUnknown(GEP));
594 rememberInstruction(GEP);
596 // Restore the original insert point.
598 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
600 return expand(SE.getAddExpr(Ops));
603 /// PickMostRelevantLoop - Given two loops pick the one that's most relevant for
604 /// SCEV expansion. If they are nested, this is the most nested. If they are
605 /// neighboring, pick the later.
606 static const Loop *PickMostRelevantLoop(const Loop *A, const Loop *B,
610 if (A->contains(B)) return B;
611 if (B->contains(A)) return A;
612 if (DT.dominates(A->getHeader(), B->getHeader())) return B;
613 if (DT.dominates(B->getHeader(), A->getHeader())) return A;
614 return A; // Arbitrarily break the tie.
617 /// getRelevantLoop - Get the most relevant loop associated with the given
618 /// expression, according to PickMostRelevantLoop.
619 const Loop *SCEVExpander::getRelevantLoop(const SCEV *S) {
620 // Test whether we've already computed the most relevant loop for this SCEV.
621 std::pair<DenseMap<const SCEV *, const Loop *>::iterator, bool> Pair =
622 RelevantLoops.insert(std::make_pair(S, static_cast<const Loop *>(0)));
624 return Pair.first->second;
626 if (isa<SCEVConstant>(S))
627 // A constant has no relevant loops.
629 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
630 if (const Instruction *I = dyn_cast<Instruction>(U->getValue()))
631 return Pair.first->second = SE.LI->getLoopFor(I->getParent());
632 // A non-instruction has no relevant loops.
635 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S)) {
637 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
639 for (SCEVNAryExpr::op_iterator I = N->op_begin(), E = N->op_end();
641 L = PickMostRelevantLoop(L, getRelevantLoop(*I), *SE.DT);
642 return RelevantLoops[N] = L;
644 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S)) {
645 const Loop *Result = getRelevantLoop(C->getOperand());
646 return RelevantLoops[C] = Result;
648 if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
650 PickMostRelevantLoop(getRelevantLoop(D->getLHS()),
651 getRelevantLoop(D->getRHS()),
653 return RelevantLoops[D] = Result;
655 llvm_unreachable("Unexpected SCEV type!");
660 /// LoopCompare - Compare loops by PickMostRelevantLoop.
664 explicit LoopCompare(DominatorTree &dt) : DT(dt) {}
666 bool operator()(std::pair<const Loop *, const SCEV *> LHS,
667 std::pair<const Loop *, const SCEV *> RHS) const {
668 // Keep pointer operands sorted at the end.
669 if (LHS.second->getType()->isPointerTy() !=
670 RHS.second->getType()->isPointerTy())
671 return LHS.second->getType()->isPointerTy();
673 // Compare loops with PickMostRelevantLoop.
674 if (LHS.first != RHS.first)
675 return PickMostRelevantLoop(LHS.first, RHS.first, DT) != LHS.first;
677 // If one operand is a non-constant negative and the other is not,
678 // put the non-constant negative on the right so that a sub can
679 // be used instead of a negate and add.
680 if (LHS.second->isNonConstantNegative()) {
681 if (!RHS.second->isNonConstantNegative())
683 } else if (RHS.second->isNonConstantNegative())
686 // Otherwise they are equivalent according to this comparison.
693 Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) {
694 Type *Ty = SE.getEffectiveSCEVType(S->getType());
696 // Collect all the add operands in a loop, along with their associated loops.
697 // Iterate in reverse so that constants are emitted last, all else equal, and
698 // so that pointer operands are inserted first, which the code below relies on
699 // to form more involved GEPs.
700 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops;
701 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(S->op_end()),
702 E(S->op_begin()); I != E; ++I)
703 OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I));
705 // Sort by loop. Use a stable sort so that constants follow non-constants and
706 // pointer operands precede non-pointer operands.
707 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT));
709 // Emit instructions to add all the operands. Hoist as much as possible
710 // out of loops, and form meaningful getelementptrs where possible.
712 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator
713 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) {
714 const Loop *CurLoop = I->first;
715 const SCEV *Op = I->second;
717 // This is the first operand. Just expand it.
720 } else if (PointerType *PTy = dyn_cast<PointerType>(Sum->getType())) {
721 // The running sum expression is a pointer. Try to form a getelementptr
722 // at this level with that as the base.
723 SmallVector<const SCEV *, 4> NewOps;
724 for (; I != E && I->first == CurLoop; ++I) {
725 // If the operand is SCEVUnknown and not instructions, peek through
726 // it, to enable more of it to be folded into the GEP.
727 const SCEV *X = I->second;
728 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(X))
729 if (!isa<Instruction>(U->getValue()))
730 X = SE.getSCEV(U->getValue());
733 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, Sum);
734 } else if (PointerType *PTy = dyn_cast<PointerType>(Op->getType())) {
735 // The running sum is an integer, and there's a pointer at this level.
736 // Try to form a getelementptr. If the running sum is instructions,
737 // use a SCEVUnknown to avoid re-analyzing them.
738 SmallVector<const SCEV *, 4> NewOps;
739 NewOps.push_back(isa<Instruction>(Sum) ? SE.getUnknown(Sum) :
741 for (++I; I != E && I->first == CurLoop; ++I)
742 NewOps.push_back(I->second);
743 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, expand(Op));
744 } else if (Op->isNonConstantNegative()) {
745 // Instead of doing a negate and add, just do a subtract.
746 Value *W = expandCodeFor(SE.getNegativeSCEV(Op), Ty);
747 Sum = InsertNoopCastOfTo(Sum, Ty);
748 Sum = InsertBinop(Instruction::Sub, Sum, W);
752 Value *W = expandCodeFor(Op, Ty);
753 Sum = InsertNoopCastOfTo(Sum, Ty);
754 // Canonicalize a constant to the RHS.
755 if (isa<Constant>(Sum)) std::swap(Sum, W);
756 Sum = InsertBinop(Instruction::Add, Sum, W);
764 Value *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) {
765 Type *Ty = SE.getEffectiveSCEVType(S->getType());
767 // Collect all the mul operands in a loop, along with their associated loops.
768 // Iterate in reverse so that constants are emitted last, all else equal.
769 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops;
770 for (std::reverse_iterator<SCEVMulExpr::op_iterator> I(S->op_end()),
771 E(S->op_begin()); I != E; ++I)
772 OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I));
774 // Sort by loop. Use a stable sort so that constants follow non-constants.
775 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT));
777 // Emit instructions to mul all the operands. Hoist as much as possible
780 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator
781 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) {
782 const SCEV *Op = I->second;
784 // This is the first operand. Just expand it.
787 } else if (Op->isAllOnesValue()) {
788 // Instead of doing a multiply by negative one, just do a negate.
789 Prod = InsertNoopCastOfTo(Prod, Ty);
790 Prod = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), Prod);
794 Value *W = expandCodeFor(Op, Ty);
795 Prod = InsertNoopCastOfTo(Prod, Ty);
796 // Canonicalize a constant to the RHS.
797 if (isa<Constant>(Prod)) std::swap(Prod, W);
798 Prod = InsertBinop(Instruction::Mul, Prod, W);
806 Value *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) {
807 Type *Ty = SE.getEffectiveSCEVType(S->getType());
809 Value *LHS = expandCodeFor(S->getLHS(), Ty);
810 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getRHS())) {
811 const APInt &RHS = SC->getValue()->getValue();
812 if (RHS.isPowerOf2())
813 return InsertBinop(Instruction::LShr, LHS,
814 ConstantInt::get(Ty, RHS.logBase2()));
817 Value *RHS = expandCodeFor(S->getRHS(), Ty);
818 return InsertBinop(Instruction::UDiv, LHS, RHS);
821 /// Move parts of Base into Rest to leave Base with the minimal
822 /// expression that provides a pointer operand suitable for a
824 static void ExposePointerBase(const SCEV *&Base, const SCEV *&Rest,
825 ScalarEvolution &SE) {
826 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Base)) {
827 Base = A->getStart();
828 Rest = SE.getAddExpr(Rest,
829 SE.getAddRecExpr(SE.getConstant(A->getType(), 0),
830 A->getStepRecurrence(SE),
832 // FIXME: A->getNoWrapFlags(FlagNW)
835 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(Base)) {
836 Base = A->getOperand(A->getNumOperands()-1);
837 SmallVector<const SCEV *, 8> NewAddOps(A->op_begin(), A->op_end());
838 NewAddOps.back() = Rest;
839 Rest = SE.getAddExpr(NewAddOps);
840 ExposePointerBase(Base, Rest, SE);
844 /// Determine if this is a well-behaved chain of instructions leading back to
845 /// the PHI. If so, it may be reused by expanded expressions.
846 bool SCEVExpander::isNormalAddRecExprPHI(PHINode *PN, Instruction *IncV,
848 if (IncV->getNumOperands() == 0 || isa<PHINode>(IncV) ||
849 (isa<CastInst>(IncV) && !isa<BitCastInst>(IncV)))
851 // If any of the operands don't dominate the insert position, bail.
852 // Addrec operands are always loop-invariant, so this can only happen
853 // if there are instructions which haven't been hoisted.
854 if (L == IVIncInsertLoop) {
855 for (User::op_iterator OI = IncV->op_begin()+1,
856 OE = IncV->op_end(); OI != OE; ++OI)
857 if (Instruction *OInst = dyn_cast<Instruction>(OI))
858 if (!SE.DT->dominates(OInst, IVIncInsertPos))
861 // Advance to the next instruction.
862 IncV = dyn_cast<Instruction>(IncV->getOperand(0));
866 if (IncV->mayHaveSideEffects())
872 return isNormalAddRecExprPHI(PN, IncV, L);
875 /// getIVIncOperand returns an induction variable increment's induction
876 /// variable operand.
878 /// If allowScale is set, any type of GEP is allowed as long as the nonIV
879 /// operands dominate InsertPos.
881 /// If allowScale is not set, ensure that a GEP increment conforms to one of the
882 /// simple patterns generated by getAddRecExprPHILiterally and
883 /// expandAddtoGEP. If the pattern isn't recognized, return NULL.
884 Instruction *SCEVExpander::getIVIncOperand(Instruction *IncV,
885 Instruction *InsertPos,
887 if (IncV == InsertPos)
890 switch (IncV->getOpcode()) {
893 // Check for a simple Add/Sub or GEP of a loop invariant step.
894 case Instruction::Add:
895 case Instruction::Sub: {
896 Instruction *OInst = dyn_cast<Instruction>(IncV->getOperand(1));
897 if (!OInst || SE.DT->properlyDominates(OInst, InsertPos))
898 return dyn_cast<Instruction>(IncV->getOperand(0));
901 case Instruction::BitCast:
902 return dyn_cast<Instruction>(IncV->getOperand(0));
903 case Instruction::GetElementPtr:
904 for (Instruction::op_iterator I = IncV->op_begin()+1, E = IncV->op_end();
906 if (isa<Constant>(*I))
908 if (Instruction *OInst = dyn_cast<Instruction>(*I)) {
909 if (!SE.DT->properlyDominates(OInst, InsertPos))
913 // allow any kind of GEP as long as it can be hoisted.
916 // This must be a pointer addition of constants (pretty), which is already
917 // handled, or some number of address-size elements (ugly). Ugly geps
918 // have 2 operands. i1* is used by the expander to represent an
919 // address-size element.
920 if (IncV->getNumOperands() != 2)
922 unsigned AS = cast<PointerType>(IncV->getType())->getAddressSpace();
923 if (IncV->getType() != Type::getInt1PtrTy(SE.getContext(), AS)
924 && IncV->getType() != Type::getInt8PtrTy(SE.getContext(), AS))
928 return dyn_cast<Instruction>(IncV->getOperand(0));
932 /// hoistStep - Attempt to hoist a simple IV increment above InsertPos to make
933 /// it available to other uses in this loop. Recursively hoist any operands,
934 /// until we reach a value that dominates InsertPos.
935 bool SCEVExpander::hoistIVInc(Instruction *IncV, Instruction *InsertPos) {
936 if (SE.DT->properlyDominates(IncV, InsertPos))
939 // InsertPos must itself dominate IncV so that IncV's new position satisfies
940 // its existing users.
941 if (!SE.DT->dominates(InsertPos->getParent(), IncV->getParent()))
944 // Check that the chain of IV operands leading back to Phi can be hoisted.
945 SmallVector<Instruction*, 4> IVIncs;
947 Instruction *Oper = getIVIncOperand(IncV, InsertPos, /*allowScale*/true);
950 // IncV is safe to hoist.
951 IVIncs.push_back(IncV);
953 if (SE.DT->properlyDominates(IncV, InsertPos))
956 for (SmallVectorImpl<Instruction*>::reverse_iterator I = IVIncs.rbegin(),
957 E = IVIncs.rend(); I != E; ++I) {
958 (*I)->moveBefore(InsertPos);
963 /// Determine if this cyclic phi is in a form that would have been generated by
964 /// LSR. We don't care if the phi was actually expanded in this pass, as long
965 /// as it is in a low-cost form, for example, no implied multiplication. This
966 /// should match any patterns generated by getAddRecExprPHILiterally and
968 bool SCEVExpander::isExpandedAddRecExprPHI(PHINode *PN, Instruction *IncV,
970 for(Instruction *IVOper = IncV;
971 (IVOper = getIVIncOperand(IVOper, L->getLoopPreheader()->getTerminator(),
972 /*allowScale=*/false));) {
979 /// expandIVInc - Expand an IV increment at Builder's current InsertPos.
980 /// Typically this is the LatchBlock terminator or IVIncInsertPos, but we may
981 /// need to materialize IV increments elsewhere to handle difficult situations.
982 Value *SCEVExpander::expandIVInc(PHINode *PN, Value *StepV, const Loop *L,
983 Type *ExpandTy, Type *IntTy,
986 // If the PHI is a pointer, use a GEP, otherwise use an add or sub.
987 if (ExpandTy->isPointerTy()) {
988 PointerType *GEPPtrTy = cast<PointerType>(ExpandTy);
989 // If the step isn't constant, don't use an implicitly scaled GEP, because
990 // that would require a multiply inside the loop.
991 if (!isa<ConstantInt>(StepV))
992 GEPPtrTy = PointerType::get(Type::getInt1Ty(SE.getContext()),
993 GEPPtrTy->getAddressSpace());
994 const SCEV *const StepArray[1] = { SE.getSCEV(StepV) };
995 IncV = expandAddToGEP(StepArray, StepArray+1, GEPPtrTy, IntTy, PN);
996 if (IncV->getType() != PN->getType()) {
997 IncV = Builder.CreateBitCast(IncV, PN->getType());
998 rememberInstruction(IncV);
1001 IncV = useSubtract ?
1002 Builder.CreateSub(PN, StepV, Twine(IVName) + ".iv.next") :
1003 Builder.CreateAdd(PN, StepV, Twine(IVName) + ".iv.next");
1004 rememberInstruction(IncV);
1009 /// getAddRecExprPHILiterally - Helper for expandAddRecExprLiterally. Expand
1010 /// the base addrec, which is the addrec without any non-loop-dominating
1011 /// values, and return the PHI.
1013 SCEVExpander::getAddRecExprPHILiterally(const SCEVAddRecExpr *Normalized,
1017 assert((!IVIncInsertLoop||IVIncInsertPos) && "Uninitialized insert position");
1019 // Reuse a previously-inserted PHI, if present.
1020 BasicBlock *LatchBlock = L->getLoopLatch();
1022 for (BasicBlock::iterator I = L->getHeader()->begin();
1023 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1024 if (!SE.isSCEVable(PN->getType()) ||
1025 (SE.getEffectiveSCEVType(PN->getType()) !=
1026 SE.getEffectiveSCEVType(Normalized->getType())) ||
1027 SE.getSCEV(PN) != Normalized)
1031 cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock));
1034 if (!isExpandedAddRecExprPHI(PN, IncV, L))
1036 if (L == IVIncInsertLoop && !hoistIVInc(IncV, IVIncInsertPos))
1040 if (!isNormalAddRecExprPHI(PN, IncV, L))
1042 if (L == IVIncInsertLoop)
1044 if (SE.DT->dominates(IncV, IVIncInsertPos))
1046 // Make sure the increment is where we want it. But don't move it
1047 // down past a potential existing post-inc user.
1048 IncV->moveBefore(IVIncInsertPos);
1049 IVIncInsertPos = IncV;
1050 IncV = cast<Instruction>(IncV->getOperand(0));
1051 } while (IncV != PN);
1053 // Ok, the add recurrence looks usable.
1054 // Remember this PHI, even in post-inc mode.
1055 InsertedValues.insert(PN);
1056 // Remember the increment.
1057 rememberInstruction(IncV);
1062 // Save the original insertion point so we can restore it when we're done.
1063 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1064 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1066 // Another AddRec may need to be recursively expanded below. For example, if
1067 // this AddRec is quadratic, the StepV may itself be an AddRec in this
1068 // loop. Remove this loop from the PostIncLoops set before expanding such
1069 // AddRecs. Otherwise, we cannot find a valid position for the step
1070 // (i.e. StepV can never dominate its loop header). Ideally, we could do
1071 // SavedIncLoops.swap(PostIncLoops), but we generally have a single element,
1072 // so it's not worth implementing SmallPtrSet::swap.
1073 PostIncLoopSet SavedPostIncLoops = PostIncLoops;
1074 PostIncLoops.clear();
1076 // Expand code for the start value.
1077 Value *StartV = expandCodeFor(Normalized->getStart(), ExpandTy,
1078 L->getHeader()->begin());
1080 // StartV must be hoisted into L's preheader to dominate the new phi.
1081 assert(!isa<Instruction>(StartV) ||
1082 SE.DT->properlyDominates(cast<Instruction>(StartV)->getParent(),
1085 // Expand code for the step value. Do this before creating the PHI so that PHI
1086 // reuse code doesn't see an incomplete PHI.
1087 const SCEV *Step = Normalized->getStepRecurrence(SE);
1088 // If the stride is negative, insert a sub instead of an add for the increment
1089 // (unless it's a constant, because subtracts of constants are canonicalized
1091 bool useSubtract = !ExpandTy->isPointerTy() && Step->isNonConstantNegative();
1093 Step = SE.getNegativeSCEV(Step);
1094 // Expand the step somewhere that dominates the loop header.
1095 Value *StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin());
1098 BasicBlock *Header = L->getHeader();
1099 Builder.SetInsertPoint(Header, Header->begin());
1100 pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header);
1101 PHINode *PN = Builder.CreatePHI(ExpandTy, std::distance(HPB, HPE),
1102 Twine(IVName) + ".iv");
1103 rememberInstruction(PN);
1105 // Create the step instructions and populate the PHI.
1106 for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) {
1107 BasicBlock *Pred = *HPI;
1109 // Add a start value.
1110 if (!L->contains(Pred)) {
1111 PN->addIncoming(StartV, Pred);
1115 // Create a step value and add it to the PHI.
1116 // If IVIncInsertLoop is non-null and equal to the addrec's loop, insert the
1117 // instructions at IVIncInsertPos.
1118 Instruction *InsertPos = L == IVIncInsertLoop ?
1119 IVIncInsertPos : Pred->getTerminator();
1120 Builder.SetInsertPoint(InsertPos);
1121 Value *IncV = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract);
1123 PN->addIncoming(IncV, Pred);
1126 // Restore the original insert point.
1128 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
1130 // After expanding subexpressions, restore the PostIncLoops set so the caller
1131 // can ensure that IVIncrement dominates the current uses.
1132 PostIncLoops = SavedPostIncLoops;
1134 // Remember this PHI, even in post-inc mode.
1135 InsertedValues.insert(PN);
1140 Value *SCEVExpander::expandAddRecExprLiterally(const SCEVAddRecExpr *S) {
1141 Type *STy = S->getType();
1142 Type *IntTy = SE.getEffectiveSCEVType(STy);
1143 const Loop *L = S->getLoop();
1145 // Determine a normalized form of this expression, which is the expression
1146 // before any post-inc adjustment is made.
1147 const SCEVAddRecExpr *Normalized = S;
1148 if (PostIncLoops.count(L)) {
1149 PostIncLoopSet Loops;
1152 cast<SCEVAddRecExpr>(TransformForPostIncUse(Normalize, S, 0, 0,
1153 Loops, SE, *SE.DT));
1156 // Strip off any non-loop-dominating component from the addrec start.
1157 const SCEV *Start = Normalized->getStart();
1158 const SCEV *PostLoopOffset = 0;
1159 if (!SE.properlyDominates(Start, L->getHeader())) {
1160 PostLoopOffset = Start;
1161 Start = SE.getConstant(Normalized->getType(), 0);
1162 Normalized = cast<SCEVAddRecExpr>(
1163 SE.getAddRecExpr(Start, Normalized->getStepRecurrence(SE),
1164 Normalized->getLoop(),
1165 // FIXME: Normalized->getNoWrapFlags(FlagNW)
1166 SCEV::FlagAnyWrap));
1169 // Strip off any non-loop-dominating component from the addrec step.
1170 const SCEV *Step = Normalized->getStepRecurrence(SE);
1171 const SCEV *PostLoopScale = 0;
1172 if (!SE.dominates(Step, L->getHeader())) {
1173 PostLoopScale = Step;
1174 Step = SE.getConstant(Normalized->getType(), 1);
1176 cast<SCEVAddRecExpr>(SE.getAddRecExpr(Start, Step,
1177 Normalized->getLoop(),
1178 // FIXME: Normalized
1179 // ->getNoWrapFlags(FlagNW)
1180 SCEV::FlagAnyWrap));
1183 // Expand the core addrec. If we need post-loop scaling, force it to
1184 // expand to an integer type to avoid the need for additional casting.
1185 Type *ExpandTy = PostLoopScale ? IntTy : STy;
1186 PHINode *PN = getAddRecExprPHILiterally(Normalized, L, ExpandTy, IntTy);
1188 // Accommodate post-inc mode, if necessary.
1190 if (!PostIncLoops.count(L))
1193 // In PostInc mode, use the post-incremented value.
1194 BasicBlock *LatchBlock = L->getLoopLatch();
1195 assert(LatchBlock && "PostInc mode requires a unique loop latch!");
1196 Result = PN->getIncomingValueForBlock(LatchBlock);
1198 // For an expansion to use the postinc form, the client must call
1199 // expandCodeFor with an InsertPoint that is either outside the PostIncLoop
1200 // or dominated by IVIncInsertPos.
1201 if (isa<Instruction>(Result)
1202 && !SE.DT->dominates(cast<Instruction>(Result),
1203 Builder.GetInsertPoint())) {
1204 // The induction variable's postinc expansion does not dominate this use.
1205 // IVUsers tries to prevent this case, so it is rare. However, it can
1206 // happen when an IVUser outside the loop is not dominated by the latch
1207 // block. Adjusting IVIncInsertPos before expansion begins cannot handle
1208 // all cases. Consider a phi outide whose operand is replaced during
1209 // expansion with the value of the postinc user. Without fundamentally
1210 // changing the way postinc users are tracked, the only remedy is
1211 // inserting an extra IV increment. StepV might fold into PostLoopOffset,
1212 // but hopefully expandCodeFor handles that.
1214 !ExpandTy->isPointerTy() && Step->isNonConstantNegative();
1216 Step = SE.getNegativeSCEV(Step);
1217 // Expand the step somewhere that dominates the loop header.
1218 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1219 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1220 Value *StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin());
1221 // Restore the insertion point to the place where the caller has
1222 // determined dominates all uses.
1223 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
1224 Result = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract);
1228 // Re-apply any non-loop-dominating scale.
1229 if (PostLoopScale) {
1230 Result = InsertNoopCastOfTo(Result, IntTy);
1231 Result = Builder.CreateMul(Result,
1232 expandCodeFor(PostLoopScale, IntTy));
1233 rememberInstruction(Result);
1236 // Re-apply any non-loop-dominating offset.
1237 if (PostLoopOffset) {
1238 if (PointerType *PTy = dyn_cast<PointerType>(ExpandTy)) {
1239 const SCEV *const OffsetArray[1] = { PostLoopOffset };
1240 Result = expandAddToGEP(OffsetArray, OffsetArray+1, PTy, IntTy, Result);
1242 Result = InsertNoopCastOfTo(Result, IntTy);
1243 Result = Builder.CreateAdd(Result,
1244 expandCodeFor(PostLoopOffset, IntTy));
1245 rememberInstruction(Result);
1252 Value *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) {
1253 if (!CanonicalMode) return expandAddRecExprLiterally(S);
1255 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1256 const Loop *L = S->getLoop();
1258 // First check for an existing canonical IV in a suitable type.
1259 PHINode *CanonicalIV = 0;
1260 if (PHINode *PN = L->getCanonicalInductionVariable())
1261 if (SE.getTypeSizeInBits(PN->getType()) >= SE.getTypeSizeInBits(Ty))
1264 // Rewrite an AddRec in terms of the canonical induction variable, if
1265 // its type is more narrow.
1267 SE.getTypeSizeInBits(CanonicalIV->getType()) >
1268 SE.getTypeSizeInBits(Ty)) {
1269 SmallVector<const SCEV *, 4> NewOps(S->getNumOperands());
1270 for (unsigned i = 0, e = S->getNumOperands(); i != e; ++i)
1271 NewOps[i] = SE.getAnyExtendExpr(S->op_begin()[i], CanonicalIV->getType());
1272 Value *V = expand(SE.getAddRecExpr(NewOps, S->getLoop(),
1273 // FIXME: S->getNoWrapFlags(FlagNW)
1274 SCEV::FlagAnyWrap));
1275 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1276 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1277 BasicBlock::iterator NewInsertPt =
1278 llvm::next(BasicBlock::iterator(cast<Instruction>(V)));
1279 while (isa<PHINode>(NewInsertPt) || isa<DbgInfoIntrinsic>(NewInsertPt) ||
1280 isa<LandingPadInst>(NewInsertPt))
1282 V = expandCodeFor(SE.getTruncateExpr(SE.getUnknown(V), Ty), 0,
1284 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
1288 // {X,+,F} --> X + {0,+,F}
1289 if (!S->getStart()->isZero()) {
1290 SmallVector<const SCEV *, 4> NewOps(S->op_begin(), S->op_end());
1291 NewOps[0] = SE.getConstant(Ty, 0);
1292 // FIXME: can use S->getNoWrapFlags()
1293 const SCEV *Rest = SE.getAddRecExpr(NewOps, L, SCEV::FlagAnyWrap);
1295 // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the
1296 // comments on expandAddToGEP for details.
1297 const SCEV *Base = S->getStart();
1298 const SCEV *RestArray[1] = { Rest };
1299 // Dig into the expression to find the pointer base for a GEP.
1300 ExposePointerBase(Base, RestArray[0], SE);
1301 // If we found a pointer, expand the AddRec with a GEP.
1302 if (PointerType *PTy = dyn_cast<PointerType>(Base->getType())) {
1303 // Make sure the Base isn't something exotic, such as a multiplied
1304 // or divided pointer value. In those cases, the result type isn't
1305 // actually a pointer type.
1306 if (!isa<SCEVMulExpr>(Base) && !isa<SCEVUDivExpr>(Base)) {
1307 Value *StartV = expand(Base);
1308 assert(StartV->getType() == PTy && "Pointer type mismatch for GEP!");
1309 return expandAddToGEP(RestArray, RestArray+1, PTy, Ty, StartV);
1313 // Just do a normal add. Pre-expand the operands to suppress folding.
1314 return expand(SE.getAddExpr(SE.getUnknown(expand(S->getStart())),
1315 SE.getUnknown(expand(Rest))));
1318 // If we don't yet have a canonical IV, create one.
1320 // Create and insert the PHI node for the induction variable in the
1322 BasicBlock *Header = L->getHeader();
1323 pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header);
1324 CanonicalIV = PHINode::Create(Ty, std::distance(HPB, HPE), "indvar",
1326 rememberInstruction(CanonicalIV);
1328 Constant *One = ConstantInt::get(Ty, 1);
1329 for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) {
1330 BasicBlock *HP = *HPI;
1331 if (L->contains(HP)) {
1332 // Insert a unit add instruction right before the terminator
1333 // corresponding to the back-edge.
1334 Instruction *Add = BinaryOperator::CreateAdd(CanonicalIV, One,
1336 HP->getTerminator());
1337 Add->setDebugLoc(HP->getTerminator()->getDebugLoc());
1338 rememberInstruction(Add);
1339 CanonicalIV->addIncoming(Add, HP);
1341 CanonicalIV->addIncoming(Constant::getNullValue(Ty), HP);
1346 // {0,+,1} --> Insert a canonical induction variable into the loop!
1347 if (S->isAffine() && S->getOperand(1)->isOne()) {
1348 assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) &&
1349 "IVs with types different from the canonical IV should "
1350 "already have been handled!");
1354 // {0,+,F} --> {0,+,1} * F
1356 // If this is a simple linear addrec, emit it now as a special case.
1357 if (S->isAffine()) // {0,+,F} --> i*F
1359 expand(SE.getTruncateOrNoop(
1360 SE.getMulExpr(SE.getUnknown(CanonicalIV),
1361 SE.getNoopOrAnyExtend(S->getOperand(1),
1362 CanonicalIV->getType())),
1365 // If this is a chain of recurrences, turn it into a closed form, using the
1366 // folders, then expandCodeFor the closed form. This allows the folders to
1367 // simplify the expression without having to build a bunch of special code
1368 // into this folder.
1369 const SCEV *IH = SE.getUnknown(CanonicalIV); // Get I as a "symbolic" SCEV.
1371 // Promote S up to the canonical IV type, if the cast is foldable.
1372 const SCEV *NewS = S;
1373 const SCEV *Ext = SE.getNoopOrAnyExtend(S, CanonicalIV->getType());
1374 if (isa<SCEVAddRecExpr>(Ext))
1377 const SCEV *V = cast<SCEVAddRecExpr>(NewS)->evaluateAtIteration(IH, SE);
1378 //cerr << "Evaluated: " << *this << "\n to: " << *V << "\n";
1380 // Truncate the result down to the original type, if needed.
1381 const SCEV *T = SE.getTruncateOrNoop(V, Ty);
1385 Value *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) {
1386 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1387 Value *V = expandCodeFor(S->getOperand(),
1388 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1389 Value *I = Builder.CreateTrunc(V, Ty);
1390 rememberInstruction(I);
1394 Value *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) {
1395 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1396 Value *V = expandCodeFor(S->getOperand(),
1397 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1398 Value *I = Builder.CreateZExt(V, Ty);
1399 rememberInstruction(I);
1403 Value *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) {
1404 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1405 Value *V = expandCodeFor(S->getOperand(),
1406 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1407 Value *I = Builder.CreateSExt(V, Ty);
1408 rememberInstruction(I);
1412 Value *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) {
1413 Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
1414 Type *Ty = LHS->getType();
1415 for (int i = S->getNumOperands()-2; i >= 0; --i) {
1416 // In the case of mixed integer and pointer types, do the
1417 // rest of the comparisons as integer.
1418 if (S->getOperand(i)->getType() != Ty) {
1419 Ty = SE.getEffectiveSCEVType(Ty);
1420 LHS = InsertNoopCastOfTo(LHS, Ty);
1422 Value *RHS = expandCodeFor(S->getOperand(i), Ty);
1423 Value *ICmp = Builder.CreateICmpSGT(LHS, RHS);
1424 rememberInstruction(ICmp);
1425 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "smax");
1426 rememberInstruction(Sel);
1429 // In the case of mixed integer and pointer types, cast the
1430 // final result back to the pointer type.
1431 if (LHS->getType() != S->getType())
1432 LHS = InsertNoopCastOfTo(LHS, S->getType());
1436 Value *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) {
1437 Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
1438 Type *Ty = LHS->getType();
1439 for (int i = S->getNumOperands()-2; i >= 0; --i) {
1440 // In the case of mixed integer and pointer types, do the
1441 // rest of the comparisons as integer.
1442 if (S->getOperand(i)->getType() != Ty) {
1443 Ty = SE.getEffectiveSCEVType(Ty);
1444 LHS = InsertNoopCastOfTo(LHS, Ty);
1446 Value *RHS = expandCodeFor(S->getOperand(i), Ty);
1447 Value *ICmp = Builder.CreateICmpUGT(LHS, RHS);
1448 rememberInstruction(ICmp);
1449 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "umax");
1450 rememberInstruction(Sel);
1453 // In the case of mixed integer and pointer types, cast the
1454 // final result back to the pointer type.
1455 if (LHS->getType() != S->getType())
1456 LHS = InsertNoopCastOfTo(LHS, S->getType());
1460 Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty,
1462 Builder.SetInsertPoint(IP->getParent(), IP);
1463 return expandCodeFor(SH, Ty);
1466 Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty) {
1467 // Expand the code for this SCEV.
1468 Value *V = expand(SH);
1470 assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) &&
1471 "non-trivial casts should be done with the SCEVs directly!");
1472 V = InsertNoopCastOfTo(V, Ty);
1477 Value *SCEVExpander::expand(const SCEV *S) {
1478 // Compute an insertion point for this SCEV object. Hoist the instructions
1479 // as far out in the loop nest as possible.
1480 Instruction *InsertPt = Builder.GetInsertPoint();
1481 for (Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock()); ;
1482 L = L->getParentLoop())
1483 if (SE.isLoopInvariant(S, L)) {
1485 if (BasicBlock *Preheader = L->getLoopPreheader())
1486 InsertPt = Preheader->getTerminator();
1488 // LSR sets the insertion point for AddRec start/step values to the
1489 // block start to simplify value reuse, even though it's an invalid
1490 // position. SCEVExpander must correct for this in all cases.
1491 InsertPt = L->getHeader()->getFirstInsertionPt();
1494 // If the SCEV is computable at this level, insert it into the header
1495 // after the PHIs (and after any other instructions that we've inserted
1496 // there) so that it is guaranteed to dominate any user inside the loop.
1497 if (L && SE.hasComputableLoopEvolution(S, L) && !PostIncLoops.count(L))
1498 InsertPt = L->getHeader()->getFirstInsertionPt();
1499 while (InsertPt != Builder.GetInsertPoint()
1500 && (isInsertedInstruction(InsertPt)
1501 || isa<DbgInfoIntrinsic>(InsertPt))) {
1502 InsertPt = llvm::next(BasicBlock::iterator(InsertPt));
1507 // Check to see if we already expanded this here.
1508 std::map<std::pair<const SCEV *, Instruction *>,
1509 AssertingVH<Value> >::iterator I =
1510 InsertedExpressions.find(std::make_pair(S, InsertPt));
1511 if (I != InsertedExpressions.end())
1514 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1515 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1516 Builder.SetInsertPoint(InsertPt->getParent(), InsertPt);
1518 // Expand the expression into instructions.
1519 Value *V = visit(S);
1521 // Remember the expanded value for this SCEV at this location.
1523 // This is independent of PostIncLoops. The mapped value simply materializes
1524 // the expression at this insertion point. If the mapped value happened to be
1525 // a postinc expansion, it could be reused by a non postinc user, but only if
1526 // its insertion point was already at the head of the loop.
1527 InsertedExpressions[std::make_pair(S, InsertPt)] = V;
1529 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
1533 void SCEVExpander::rememberInstruction(Value *I) {
1534 if (!PostIncLoops.empty())
1535 InsertedPostIncValues.insert(I);
1537 InsertedValues.insert(I);
1540 void SCEVExpander::restoreInsertPoint(BasicBlock *BB, BasicBlock::iterator I) {
1541 Builder.SetInsertPoint(BB, I);
1544 /// getOrInsertCanonicalInductionVariable - This method returns the
1545 /// canonical induction variable of the specified type for the specified
1546 /// loop (inserting one if there is none). A canonical induction variable
1547 /// starts at zero and steps by one on each iteration.
1549 SCEVExpander::getOrInsertCanonicalInductionVariable(const Loop *L,
1551 assert(Ty->isIntegerTy() && "Can only insert integer induction variables!");
1553 // Build a SCEV for {0,+,1}<L>.
1554 // Conservatively use FlagAnyWrap for now.
1555 const SCEV *H = SE.getAddRecExpr(SE.getConstant(Ty, 0),
1556 SE.getConstant(Ty, 1), L, SCEV::FlagAnyWrap);
1558 // Emit code for it.
1559 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1560 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1561 PHINode *V = cast<PHINode>(expandCodeFor(H, 0, L->getHeader()->begin()));
1563 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
1568 /// Sort values by integer width for replaceCongruentIVs.
1569 static bool width_descending(Value *lhs, Value *rhs) {
1570 // Put pointers at the back and make sure pointer < pointer = false.
1571 if (!lhs->getType()->isIntegerTy() || !rhs->getType()->isIntegerTy())
1572 return rhs->getType()->isIntegerTy() && !lhs->getType()->isIntegerTy();
1573 return rhs->getType()->getPrimitiveSizeInBits()
1574 < lhs->getType()->getPrimitiveSizeInBits();
1577 /// replaceCongruentIVs - Check for congruent phis in this loop header and
1578 /// replace them with their most canonical representative. Return the number of
1579 /// phis eliminated.
1581 /// This does not depend on any SCEVExpander state but should be used in
1582 /// the same context that SCEVExpander is used.
1583 unsigned SCEVExpander::replaceCongruentIVs(Loop *L, const DominatorTree *DT,
1584 SmallVectorImpl<WeakVH> &DeadInsts,
1585 const TargetLowering *TLI) {
1586 // Find integer phis in order of increasing width.
1587 SmallVector<PHINode*, 8> Phis;
1588 for (BasicBlock::iterator I = L->getHeader()->begin();
1589 PHINode *Phi = dyn_cast<PHINode>(I); ++I) {
1590 Phis.push_back(Phi);
1593 std::sort(Phis.begin(), Phis.end(), width_descending);
1595 unsigned NumElim = 0;
1596 DenseMap<const SCEV *, PHINode *> ExprToIVMap;
1597 // Process phis from wide to narrow. Mapping wide phis to the their truncation
1598 // so narrow phis can reuse them.
1599 for (SmallVectorImpl<PHINode*>::const_iterator PIter = Phis.begin(),
1600 PEnd = Phis.end(); PIter != PEnd; ++PIter) {
1601 PHINode *Phi = *PIter;
1603 if (!SE.isSCEVable(Phi->getType()))
1606 PHINode *&OrigPhiRef = ExprToIVMap[SE.getSCEV(Phi)];
1609 if (Phi->getType()->isIntegerTy() && TLI
1610 && TLI->isTruncateFree(Phi->getType(), Phis.back()->getType())) {
1611 // This phi can be freely truncated to the narrowest phi type. Map the
1612 // truncated expression to it so it will be reused for narrow types.
1613 const SCEV *TruncExpr =
1614 SE.getTruncateExpr(SE.getSCEV(Phi), Phis.back()->getType());
1615 ExprToIVMap[TruncExpr] = Phi;
1620 // Replacing a pointer phi with an integer phi or vice-versa doesn't make
1622 if (OrigPhiRef->getType()->isPointerTy() != Phi->getType()->isPointerTy())
1625 if (BasicBlock *LatchBlock = L->getLoopLatch()) {
1626 Instruction *OrigInc =
1627 cast<Instruction>(OrigPhiRef->getIncomingValueForBlock(LatchBlock));
1628 Instruction *IsomorphicInc =
1629 cast<Instruction>(Phi->getIncomingValueForBlock(LatchBlock));
1631 // If this phi has the same width but is more canonical, replace the
1632 // original with it. As part of the "more canonical" determination,
1633 // respect a prior decision to use an IV chain.
1634 if (OrigPhiRef->getType() == Phi->getType()
1635 && !(ChainedPhis.count(Phi)
1636 || isExpandedAddRecExprPHI(OrigPhiRef, OrigInc, L))
1637 && (ChainedPhis.count(Phi)
1638 || isExpandedAddRecExprPHI(Phi, IsomorphicInc, L))) {
1639 std::swap(OrigPhiRef, Phi);
1640 std::swap(OrigInc, IsomorphicInc);
1642 // Replacing the congruent phi is sufficient because acyclic redundancy
1643 // elimination, CSE/GVN, should handle the rest. However, once SCEV proves
1644 // that a phi is congruent, it's often the head of an IV user cycle that
1645 // is isomorphic with the original phi. It's worth eagerly cleaning up the
1646 // common case of a single IV increment so that DeleteDeadPHIs can remove
1647 // cycles that had postinc uses.
1648 const SCEV *TruncExpr = SE.getTruncateOrNoop(SE.getSCEV(OrigInc),
1649 IsomorphicInc->getType());
1650 if (OrigInc != IsomorphicInc
1651 && TruncExpr == SE.getSCEV(IsomorphicInc)
1652 && ((isa<PHINode>(OrigInc) && isa<PHINode>(IsomorphicInc))
1653 || hoistIVInc(OrigInc, IsomorphicInc))) {
1654 DEBUG_WITH_TYPE(DebugType, dbgs()
1655 << "INDVARS: Eliminated congruent iv.inc: "
1656 << *IsomorphicInc << '\n');
1657 Value *NewInc = OrigInc;
1658 if (OrigInc->getType() != IsomorphicInc->getType()) {
1659 Instruction *IP = isa<PHINode>(OrigInc)
1660 ? (Instruction*)L->getHeader()->getFirstInsertionPt()
1661 : OrigInc->getNextNode();
1662 IRBuilder<> Builder(IP);
1663 Builder.SetCurrentDebugLocation(IsomorphicInc->getDebugLoc());
1665 CreateTruncOrBitCast(OrigInc, IsomorphicInc->getType(), IVName);
1667 IsomorphicInc->replaceAllUsesWith(NewInc);
1668 DeadInsts.push_back(IsomorphicInc);
1671 DEBUG_WITH_TYPE(DebugType, dbgs()
1672 << "INDVARS: Eliminated congruent iv: " << *Phi << '\n');
1674 Value *NewIV = OrigPhiRef;
1675 if (OrigPhiRef->getType() != Phi->getType()) {
1676 IRBuilder<> Builder(L->getHeader()->getFirstInsertionPt());
1677 Builder.SetCurrentDebugLocation(Phi->getDebugLoc());
1678 NewIV = Builder.CreateTruncOrBitCast(OrigPhiRef, Phi->getType(), IVName);
1680 Phi->replaceAllUsesWith(NewIV);
1681 DeadInsts.push_back(Phi);