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
/// InsertNoopCastOfTo - Insert a cast of V to the specified type,
25
/// which must be possible with a noop cast, doing what we can to share
27
Value *SCEVExpander::InsertNoopCastOfTo(Value *V, const Type *Ty) {
28
Instruction::CastOps Op = CastInst::getCastOpcode(V, false, Ty, false);
29
assert((Op == Instruction::BitCast ||
30
Op == Instruction::PtrToInt ||
31
Op == Instruction::IntToPtr) &&
32
"InsertNoopCastOfTo cannot perform non-noop casts!");
33
assert(SE.getTypeSizeInBits(V->getType()) == SE.getTypeSizeInBits(Ty) &&
34
"InsertNoopCastOfTo cannot change sizes!");
36
// Short-circuit unnecessary bitcasts.
37
if (Op == Instruction::BitCast && V->getType() == Ty)
40
// Short-circuit unnecessary inttoptr<->ptrtoint casts.
41
if ((Op == Instruction::PtrToInt || Op == Instruction::IntToPtr) &&
42
SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(V->getType())) {
43
if (CastInst *CI = dyn_cast<CastInst>(V))
44
if ((CI->getOpcode() == Instruction::PtrToInt ||
45
CI->getOpcode() == Instruction::IntToPtr) &&
46
SE.getTypeSizeInBits(CI->getType()) ==
47
SE.getTypeSizeInBits(CI->getOperand(0)->getType()))
48
return CI->getOperand(0);
49
if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
50
if ((CE->getOpcode() == Instruction::PtrToInt ||
51
CE->getOpcode() == Instruction::IntToPtr) &&
52
SE.getTypeSizeInBits(CE->getType()) ==
53
SE.getTypeSizeInBits(CE->getOperand(0)->getType()))
54
return CE->getOperand(0);
57
if (Constant *C = dyn_cast<Constant>(V))
58
return ConstantExpr::getCast(Op, C, Ty);
60
if (Argument *A = dyn_cast<Argument>(V)) {
61
// Check to see if there is already a cast!
62
for (Value::use_iterator UI = A->use_begin(), E = A->use_end();
64
if ((*UI)->getType() == Ty)
65
if (CastInst *CI = dyn_cast<CastInst>(cast<Instruction>(*UI)))
66
if (CI->getOpcode() == Op) {
67
// If the cast isn't the first instruction of the function, move it.
68
if (BasicBlock::iterator(CI) !=
69
A->getParent()->getEntryBlock().begin()) {
70
// Recreate the cast at the beginning of the entry block.
71
// The old cast is left in place in case it is being used
72
// as an insert point.
74
CastInst::Create(Op, V, Ty, "",
75
A->getParent()->getEntryBlock().begin());
77
CI->replaceAllUsesWith(NewCI);
83
Instruction *I = CastInst::Create(Op, V, Ty, V->getName(),
84
A->getParent()->getEntryBlock().begin());
85
rememberInstruction(I);
89
Instruction *I = cast<Instruction>(V);
91
// Check to see if there is already a cast. If there is, use it.
92
for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
94
if ((*UI)->getType() == Ty)
95
if (CastInst *CI = dyn_cast<CastInst>(cast<Instruction>(*UI)))
96
if (CI->getOpcode() == Op) {
97
BasicBlock::iterator It = I; ++It;
98
if (isa<InvokeInst>(I))
99
It = cast<InvokeInst>(I)->getNormalDest()->begin();
100
while (isa<PHINode>(It)) ++It;
101
if (It != BasicBlock::iterator(CI)) {
102
// Recreate the cast after the user.
103
// The old cast is left in place in case it is being used
104
// as an insert point.
105
Instruction *NewCI = CastInst::Create(Op, V, Ty, "", It);
107
CI->replaceAllUsesWith(NewCI);
108
rememberInstruction(NewCI);
111
rememberInstruction(CI);
115
BasicBlock::iterator IP = I; ++IP;
116
if (InvokeInst *II = dyn_cast<InvokeInst>(I))
117
IP = II->getNormalDest()->begin();
118
while (isa<PHINode>(IP)) ++IP;
119
Instruction *CI = CastInst::Create(Op, V, Ty, V->getName(), IP);
120
rememberInstruction(CI);
124
/// InsertBinop - Insert the specified binary operator, doing a small amount
125
/// of work to avoid inserting an obviously redundant operation.
126
Value *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode,
127
Value *LHS, Value *RHS) {
128
// Fold a binop with constant operands.
129
if (Constant *CLHS = dyn_cast<Constant>(LHS))
130
if (Constant *CRHS = dyn_cast<Constant>(RHS))
131
return ConstantExpr::get(Opcode, CLHS, CRHS);
133
// Do a quick scan to see if we have this binop nearby. If so, reuse it.
134
unsigned ScanLimit = 6;
135
BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
136
// Scanning starts from the last instruction before the insertion point.
137
BasicBlock::iterator IP = Builder.GetInsertPoint();
138
if (IP != BlockBegin) {
140
for (; ScanLimit; --IP, --ScanLimit) {
141
// Don't count dbg.value against the ScanLimit, to avoid perturbing the
143
if (isa<DbgInfoIntrinsic>(IP))
145
if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(0) == LHS &&
146
IP->getOperand(1) == RHS)
148
if (IP == BlockBegin) break;
152
// Save the original insertion point so we can restore it when we're done.
153
BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
154
BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
156
// Move the insertion point out of as many loops as we can.
157
while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
158
if (!L->isLoopInvariant(LHS) || !L->isLoopInvariant(RHS)) break;
159
BasicBlock *Preheader = L->getLoopPreheader();
160
if (!Preheader) break;
162
// Ok, move up a level.
163
Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
166
// If we haven't found this binop, insert it.
167
Value *BO = Builder.CreateBinOp(Opcode, LHS, RHS, "tmp");
168
rememberInstruction(BO);
170
// Restore the original insert point.
172
restoreInsertPoint(SaveInsertBB, SaveInsertPt);
177
/// FactorOutConstant - Test if S is divisible by Factor, using signed
178
/// division. If so, update S with Factor divided out and return true.
179
/// S need not be evenly divisible if a reasonable remainder can be
181
/// TODO: When ScalarEvolution gets a SCEVSDivExpr, this can be made
182
/// unnecessary; in its place, just signed-divide Ops[i] by the scale and
183
/// check to see if the divide was folded.
184
static bool FactorOutConstant(const SCEV *&S,
185
const SCEV *&Remainder,
188
const TargetData *TD) {
189
// Everything is divisible by one.
195
S = SE.getIntegerSCEV(1, S->getType());
199
// For a Constant, check for a multiple of the given factor.
200
if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
204
// Check for divisibility.
205
if (const SCEVConstant *FC = dyn_cast<SCEVConstant>(Factor)) {
207
ConstantInt::get(SE.getContext(),
208
C->getValue()->getValue().sdiv(
209
FC->getValue()->getValue()));
210
// If the quotient is zero and the remainder is non-zero, reject
211
// the value at this scale. It will be considered for subsequent
214
const SCEV *Div = SE.getConstant(CI);
217
SE.getAddExpr(Remainder,
218
SE.getConstant(C->getValue()->getValue().srem(
219
FC->getValue()->getValue())));
225
// In a Mul, check if there is a constant operand which is a multiple
226
// of the given factor.
227
if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
229
// With TargetData, the size is known. Check if there is a constant
230
// operand which is a multiple of the given factor. If so, we can
232
const SCEVConstant *FC = cast<SCEVConstant>(Factor);
233
if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0)))
234
if (!C->getValue()->getValue().srem(FC->getValue()->getValue())) {
235
const SmallVectorImpl<const SCEV *> &MOperands = M->getOperands();
236
SmallVector<const SCEV *, 4> NewMulOps(MOperands.begin(),
239
SE.getConstant(C->getValue()->getValue().sdiv(
240
FC->getValue()->getValue()));
241
S = SE.getMulExpr(NewMulOps);
245
// Without TargetData, check if Factor can be factored out of any of the
246
// Mul's operands. If so, we can just remove it.
247
for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) {
248
const SCEV *SOp = M->getOperand(i);
249
const SCEV *Remainder = SE.getIntegerSCEV(0, SOp->getType());
250
if (FactorOutConstant(SOp, Remainder, Factor, SE, TD) &&
251
Remainder->isZero()) {
252
const SmallVectorImpl<const SCEV *> &MOperands = M->getOperands();
253
SmallVector<const SCEV *, 4> NewMulOps(MOperands.begin(),
256
S = SE.getMulExpr(NewMulOps);
263
// In an AddRec, check if both start and step are divisible.
264
if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
265
const SCEV *Step = A->getStepRecurrence(SE);
266
const SCEV *StepRem = SE.getIntegerSCEV(0, Step->getType());
267
if (!FactorOutConstant(Step, StepRem, Factor, SE, TD))
269
if (!StepRem->isZero())
271
const SCEV *Start = A->getStart();
272
if (!FactorOutConstant(Start, Remainder, Factor, SE, TD))
274
S = SE.getAddRecExpr(Start, Step, A->getLoop());
281
/// SimplifyAddOperands - Sort and simplify a list of add operands. NumAddRecs
282
/// is the number of SCEVAddRecExprs present, which are kept at the end of
285
static void SimplifyAddOperands(SmallVectorImpl<const SCEV *> &Ops,
287
ScalarEvolution &SE) {
288
unsigned NumAddRecs = 0;
289
for (unsigned i = Ops.size(); i > 0 && isa<SCEVAddRecExpr>(Ops[i-1]); --i)
291
// Group Ops into non-addrecs and addrecs.
292
SmallVector<const SCEV *, 8> NoAddRecs(Ops.begin(), Ops.end() - NumAddRecs);
293
SmallVector<const SCEV *, 8> AddRecs(Ops.end() - NumAddRecs, Ops.end());
294
// Let ScalarEvolution sort and simplify the non-addrecs list.
295
const SCEV *Sum = NoAddRecs.empty() ?
296
SE.getIntegerSCEV(0, Ty) :
297
SE.getAddExpr(NoAddRecs);
298
// If it returned an add, use the operands. Otherwise it simplified
299
// the sum into a single value, so just use that.
300
if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Sum))
301
Ops = Add->getOperands();
307
// Then append the addrecs.
308
Ops.insert(Ops.end(), AddRecs.begin(), AddRecs.end());
311
/// SplitAddRecs - Flatten a list of add operands, moving addrec start values
312
/// out to the top level. For example, convert {a + b,+,c} to a, b, {0,+,d}.
313
/// This helps expose more opportunities for folding parts of the expressions
314
/// into GEP indices.
316
static void SplitAddRecs(SmallVectorImpl<const SCEV *> &Ops,
318
ScalarEvolution &SE) {
320
SmallVector<const SCEV *, 8> AddRecs;
321
for (unsigned i = 0, e = Ops.size(); i != e; ++i)
322
while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Ops[i])) {
323
const SCEV *Start = A->getStart();
324
if (Start->isZero()) break;
325
const SCEV *Zero = SE.getIntegerSCEV(0, Ty);
326
AddRecs.push_back(SE.getAddRecExpr(Zero,
327
A->getStepRecurrence(SE),
329
if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Start)) {
331
Ops.insert(Ops.end(), Add->op_begin(), Add->op_end());
332
e += Add->getNumOperands();
337
if (!AddRecs.empty()) {
338
// Add the addrecs onto the end of the list.
339
Ops.insert(Ops.end(), AddRecs.begin(), AddRecs.end());
340
// Resort the operand list, moving any constants to the front.
341
SimplifyAddOperands(Ops, Ty, SE);
345
/// expandAddToGEP - Expand an addition expression with a pointer type into
346
/// a GEP instead of using ptrtoint+arithmetic+inttoptr. This helps
347
/// BasicAliasAnalysis and other passes analyze the result. See the rules
348
/// for getelementptr vs. inttoptr in
349
/// http://llvm.org/docs/LangRef.html#pointeraliasing
352
/// Design note: The correctness of using getelementptr here depends on
353
/// ScalarEvolution not recognizing inttoptr and ptrtoint operators, as
354
/// they may introduce pointer arithmetic which may not be safely converted
355
/// into getelementptr.
357
/// Design note: It might seem desirable for this function to be more
358
/// loop-aware. If some of the indices are loop-invariant while others
359
/// aren't, it might seem desirable to emit multiple GEPs, keeping the
360
/// loop-invariant portions of the overall computation outside the loop.
361
/// However, there are a few reasons this is not done here. Hoisting simple
362
/// arithmetic is a low-level optimization that often isn't very
363
/// important until late in the optimization process. In fact, passes
364
/// like InstructionCombining will combine GEPs, even if it means
365
/// pushing loop-invariant computation down into loops, so even if the
366
/// GEPs were split here, the work would quickly be undone. The
367
/// LoopStrengthReduction pass, which is usually run quite late (and
368
/// after the last InstructionCombining pass), takes care of hoisting
369
/// loop-invariant portions of expressions, after considering what
370
/// can be folded using target addressing modes.
372
Value *SCEVExpander::expandAddToGEP(const SCEV *const *op_begin,
373
const SCEV *const *op_end,
374
const PointerType *PTy,
377
const Type *ElTy = PTy->getElementType();
378
SmallVector<Value *, 4> GepIndices;
379
SmallVector<const SCEV *, 8> Ops(op_begin, op_end);
380
bool AnyNonZeroIndices = false;
382
// Split AddRecs up into parts as either of the parts may be usable
383
// without the other.
384
SplitAddRecs(Ops, Ty, SE);
386
// Descend down the pointer's type and attempt to convert the other
387
// operands into GEP indices, at each level. The first index in a GEP
388
// indexes into the array implied by the pointer operand; the rest of
389
// the indices index into the element or field type selected by the
392
// If the scale size is not 0, attempt to factor out a scale for
394
SmallVector<const SCEV *, 8> ScaledOps;
395
if (ElTy->isSized()) {
396
const SCEV *ElSize = SE.getSizeOfExpr(ElTy);
397
if (!ElSize->isZero()) {
398
SmallVector<const SCEV *, 8> NewOps;
399
for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
400
const SCEV *Op = Ops[i];
401
const SCEV *Remainder = SE.getIntegerSCEV(0, Ty);
402
if (FactorOutConstant(Op, Remainder, ElSize, SE, SE.TD)) {
403
// Op now has ElSize factored out.
404
ScaledOps.push_back(Op);
405
if (!Remainder->isZero())
406
NewOps.push_back(Remainder);
407
AnyNonZeroIndices = true;
409
// The operand was not divisible, so add it to the list of operands
410
// we'll scan next iteration.
411
NewOps.push_back(Ops[i]);
414
// If we made any changes, update Ops.
415
if (!ScaledOps.empty()) {
417
SimplifyAddOperands(Ops, Ty, SE);
422
// Record the scaled array index for this level of the type. If
423
// we didn't find any operands that could be factored, tentatively
424
// assume that element zero was selected (since the zero offset
425
// would obviously be folded away).
426
Value *Scaled = ScaledOps.empty() ?
427
Constant::getNullValue(Ty) :
428
expandCodeFor(SE.getAddExpr(ScaledOps), Ty);
429
GepIndices.push_back(Scaled);
431
// Collect struct field index operands.
432
while (const StructType *STy = dyn_cast<StructType>(ElTy)) {
433
bool FoundFieldNo = false;
434
// An empty struct has no fields.
435
if (STy->getNumElements() == 0) break;
437
// With TargetData, field offsets are known. See if a constant offset
438
// falls within any of the struct fields.
439
if (Ops.empty()) break;
440
if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[0]))
441
if (SE.getTypeSizeInBits(C->getType()) <= 64) {
442
const StructLayout &SL = *SE.TD->getStructLayout(STy);
443
uint64_t FullOffset = C->getValue()->getZExtValue();
444
if (FullOffset < SL.getSizeInBytes()) {
445
unsigned ElIdx = SL.getElementContainingOffset(FullOffset);
446
GepIndices.push_back(
447
ConstantInt::get(Type::getInt32Ty(Ty->getContext()), ElIdx));
448
ElTy = STy->getTypeAtIndex(ElIdx);
450
SE.getConstant(Ty, FullOffset - SL.getElementOffset(ElIdx));
451
AnyNonZeroIndices = true;
456
// Without TargetData, just check for an offsetof expression of the
457
// appropriate struct type.
458
for (unsigned i = 0, e = Ops.size(); i != e; ++i)
459
if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Ops[i])) {
462
if (U->isOffsetOf(CTy, FieldNo) && CTy == STy) {
463
GepIndices.push_back(FieldNo);
465
STy->getTypeAtIndex(cast<ConstantInt>(FieldNo)->getZExtValue());
466
Ops[i] = SE.getConstant(Ty, 0);
467
AnyNonZeroIndices = true;
473
// If no struct field offsets were found, tentatively assume that
474
// field zero was selected (since the zero offset would obviously
477
ElTy = STy->getTypeAtIndex(0u);
478
GepIndices.push_back(
479
Constant::getNullValue(Type::getInt32Ty(Ty->getContext())));
483
if (const ArrayType *ATy = dyn_cast<ArrayType>(ElTy))
484
ElTy = ATy->getElementType();
489
// If none of the operands were convertible to proper GEP indices, cast
490
// the base to i8* and do an ugly getelementptr with that. It's still
491
// better than ptrtoint+arithmetic+inttoptr at least.
492
if (!AnyNonZeroIndices) {
493
// Cast the base to i8*.
494
V = InsertNoopCastOfTo(V,
495
Type::getInt8PtrTy(Ty->getContext(), PTy->getAddressSpace()));
497
// Expand the operands for a plain byte offset.
498
Value *Idx = expandCodeFor(SE.getAddExpr(Ops), Ty);
500
// Fold a GEP with constant operands.
501
if (Constant *CLHS = dyn_cast<Constant>(V))
502
if (Constant *CRHS = dyn_cast<Constant>(Idx))
503
return ConstantExpr::getGetElementPtr(CLHS, &CRHS, 1);
505
// Do a quick scan to see if we have this GEP nearby. If so, reuse it.
506
unsigned ScanLimit = 6;
507
BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
508
// Scanning starts from the last instruction before the insertion point.
509
BasicBlock::iterator IP = Builder.GetInsertPoint();
510
if (IP != BlockBegin) {
512
for (; ScanLimit; --IP, --ScanLimit) {
513
// Don't count dbg.value against the ScanLimit, to avoid perturbing the
515
if (isa<DbgInfoIntrinsic>(IP))
517
if (IP->getOpcode() == Instruction::GetElementPtr &&
518
IP->getOperand(0) == V && IP->getOperand(1) == Idx)
520
if (IP == BlockBegin) break;
524
// Save the original insertion point so we can restore it when we're done.
525
BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
526
BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
528
// Move the insertion point out of as many loops as we can.
529
while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
530
if (!L->isLoopInvariant(V) || !L->isLoopInvariant(Idx)) break;
531
BasicBlock *Preheader = L->getLoopPreheader();
532
if (!Preheader) break;
534
// Ok, move up a level.
535
Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
539
Value *GEP = Builder.CreateGEP(V, Idx, "uglygep");
540
rememberInstruction(GEP);
542
// Restore the original insert point.
544
restoreInsertPoint(SaveInsertBB, SaveInsertPt);
549
// Save the original insertion point so we can restore it when we're done.
550
BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
551
BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
553
// Move the insertion point out of as many loops as we can.
554
while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
555
if (!L->isLoopInvariant(V)) break;
557
bool AnyIndexNotLoopInvariant = false;
558
for (SmallVectorImpl<Value *>::const_iterator I = GepIndices.begin(),
559
E = GepIndices.end(); I != E; ++I)
560
if (!L->isLoopInvariant(*I)) {
561
AnyIndexNotLoopInvariant = true;
564
if (AnyIndexNotLoopInvariant)
567
BasicBlock *Preheader = L->getLoopPreheader();
568
if (!Preheader) break;
570
// Ok, move up a level.
571
Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
574
// Insert a pretty getelementptr. Note that this GEP is not marked inbounds,
575
// because ScalarEvolution may have changed the address arithmetic to
576
// compute a value which is beyond the end of the allocated object.
578
if (V->getType() != PTy)
579
Casted = InsertNoopCastOfTo(Casted, PTy);
580
Value *GEP = Builder.CreateGEP(Casted,
584
Ops.push_back(SE.getUnknown(GEP));
585
rememberInstruction(GEP);
587
// Restore the original insert point.
589
restoreInsertPoint(SaveInsertBB, SaveInsertPt);
591
return expand(SE.getAddExpr(Ops));
594
/// isNonConstantNegative - Return true if the specified scev is negated, but
596
static bool isNonConstantNegative(const SCEV *F) {
597
const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(F);
598
if (!Mul) return false;
600
// If there is a constant factor, it will be first.
601
const SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
602
if (!SC) return false;
604
// Return true if the value is negative, this matches things like (-42 * V).
605
return SC->getValue()->getValue().isNegative();
608
/// PickMostRelevantLoop - Given two loops pick the one that's most relevant for
609
/// SCEV expansion. If they are nested, this is the most nested. If they are
610
/// neighboring, pick the later.
611
static const Loop *PickMostRelevantLoop(const Loop *A, const Loop *B,
615
if (A->contains(B)) return B;
616
if (B->contains(A)) return A;
617
if (DT.dominates(A->getHeader(), B->getHeader())) return B;
618
if (DT.dominates(B->getHeader(), A->getHeader())) return A;
619
return A; // Arbitrarily break the tie.
622
/// GetRelevantLoop - Get the most relevant loop associated with the given
623
/// expression, according to PickMostRelevantLoop.
624
static const Loop *GetRelevantLoop(const SCEV *S, LoopInfo &LI,
626
if (isa<SCEVConstant>(S))
628
if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
629
if (const Instruction *I = dyn_cast<Instruction>(U->getValue()))
630
return LI.getLoopFor(I->getParent());
633
if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S)) {
635
if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
637
for (SCEVNAryExpr::op_iterator I = N->op_begin(), E = N->op_end();
639
L = PickMostRelevantLoop(L, GetRelevantLoop(*I, LI, DT), DT);
642
if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
643
return GetRelevantLoop(C->getOperand(), LI, DT);
644
if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S))
645
return PickMostRelevantLoop(GetRelevantLoop(D->getLHS(), LI, DT),
646
GetRelevantLoop(D->getRHS(), LI, DT),
648
llvm_unreachable("Unexpected SCEV type!");
651
/// LoopCompare - Compare loops by PickMostRelevantLoop.
655
explicit LoopCompare(DominatorTree &dt) : DT(dt) {}
657
bool operator()(std::pair<const Loop *, const SCEV *> LHS,
658
std::pair<const Loop *, const SCEV *> RHS) const {
659
// Compare loops with PickMostRelevantLoop.
660
if (LHS.first != RHS.first)
661
return PickMostRelevantLoop(LHS.first, RHS.first, DT) != LHS.first;
663
// If one operand is a non-constant negative and the other is not,
664
// put the non-constant negative on the right so that a sub can
665
// be used instead of a negate and add.
666
if (isNonConstantNegative(LHS.second)) {
667
if (!isNonConstantNegative(RHS.second))
669
} else if (isNonConstantNegative(RHS.second))
672
// Otherwise they are equivalent according to this comparison.
677
Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) {
678
const Type *Ty = SE.getEffectiveSCEVType(S->getType());
680
// Collect all the add operands in a loop, along with their associated loops.
681
// Iterate in reverse so that constants are emitted last, all else equal, and
682
// so that pointer operands are inserted first, which the code below relies on
683
// to form more involved GEPs.
684
SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops;
685
for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(S->op_end()),
686
E(S->op_begin()); I != E; ++I)
687
OpsAndLoops.push_back(std::make_pair(GetRelevantLoop(*I, *SE.LI, *SE.DT),
690
// Sort by loop. Use a stable sort so that constants follow non-constants and
691
// pointer operands precede non-pointer operands.
692
std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT));
694
// Emit instructions to add all the operands. Hoist as much as possible
695
// out of loops, and form meaningful getelementptrs where possible.
697
for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator
698
I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) {
699
const Loop *CurLoop = I->first;
700
const SCEV *Op = I->second;
702
// This is the first operand. Just expand it.
705
} else if (const PointerType *PTy = dyn_cast<PointerType>(Sum->getType())) {
706
// The running sum expression is a pointer. Try to form a getelementptr
707
// at this level with that as the base.
708
SmallVector<const SCEV *, 4> NewOps;
709
for (; I != E && I->first == CurLoop; ++I)
710
NewOps.push_back(I->second);
711
Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, Sum);
712
} else if (const PointerType *PTy = dyn_cast<PointerType>(Op->getType())) {
713
// The running sum is an integer, and there's a pointer at this level.
714
// Try to form a getelementptr.
715
SmallVector<const SCEV *, 4> NewOps;
716
NewOps.push_back(SE.getUnknown(Sum));
717
for (++I; I != E && I->first == CurLoop; ++I)
718
NewOps.push_back(I->second);
719
Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, expand(Op));
720
} else if (isNonConstantNegative(Op)) {
721
// Instead of doing a negate and add, just do a subtract.
722
Value *W = expandCodeFor(SE.getNegativeSCEV(Op), Ty);
723
Sum = InsertNoopCastOfTo(Sum, Ty);
724
Sum = InsertBinop(Instruction::Sub, Sum, W);
728
Value *W = expandCodeFor(Op, Ty);
729
Sum = InsertNoopCastOfTo(Sum, Ty);
730
// Canonicalize a constant to the RHS.
731
if (isa<Constant>(Sum)) std::swap(Sum, W);
732
Sum = InsertBinop(Instruction::Add, Sum, W);
740
Value *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) {
741
const Type *Ty = SE.getEffectiveSCEVType(S->getType());
743
// Collect all the mul operands in a loop, along with their associated loops.
744
// Iterate in reverse so that constants are emitted last, all else equal.
745
SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops;
746
for (std::reverse_iterator<SCEVMulExpr::op_iterator> I(S->op_end()),
747
E(S->op_begin()); I != E; ++I)
748
OpsAndLoops.push_back(std::make_pair(GetRelevantLoop(*I, *SE.LI, *SE.DT),
751
// Sort by loop. Use a stable sort so that constants follow non-constants.
752
std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT));
754
// Emit instructions to mul all the operands. Hoist as much as possible
757
for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator
758
I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) {
759
const SCEV *Op = I->second;
761
// This is the first operand. Just expand it.
764
} else if (Op->isAllOnesValue()) {
765
// Instead of doing a multiply by negative one, just do a negate.
766
Prod = InsertNoopCastOfTo(Prod, Ty);
767
Prod = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), Prod);
771
Value *W = expandCodeFor(Op, Ty);
772
Prod = InsertNoopCastOfTo(Prod, Ty);
773
// Canonicalize a constant to the RHS.
774
if (isa<Constant>(Prod)) std::swap(Prod, W);
775
Prod = InsertBinop(Instruction::Mul, Prod, W);
783
Value *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) {
784
const Type *Ty = SE.getEffectiveSCEVType(S->getType());
786
Value *LHS = expandCodeFor(S->getLHS(), Ty);
787
if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getRHS())) {
788
const APInt &RHS = SC->getValue()->getValue();
789
if (RHS.isPowerOf2())
790
return InsertBinop(Instruction::LShr, LHS,
791
ConstantInt::get(Ty, RHS.logBase2()));
794
Value *RHS = expandCodeFor(S->getRHS(), Ty);
795
return InsertBinop(Instruction::UDiv, LHS, RHS);
798
/// Move parts of Base into Rest to leave Base with the minimal
799
/// expression that provides a pointer operand suitable for a
801
static void ExposePointerBase(const SCEV *&Base, const SCEV *&Rest,
802
ScalarEvolution &SE) {
803
while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Base)) {
804
Base = A->getStart();
805
Rest = SE.getAddExpr(Rest,
806
SE.getAddRecExpr(SE.getIntegerSCEV(0, A->getType()),
807
A->getStepRecurrence(SE),
810
if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(Base)) {
811
Base = A->getOperand(A->getNumOperands()-1);
812
SmallVector<const SCEV *, 8> NewAddOps(A->op_begin(), A->op_end());
813
NewAddOps.back() = Rest;
814
Rest = SE.getAddExpr(NewAddOps);
815
ExposePointerBase(Base, Rest, SE);
819
/// getAddRecExprPHILiterally - Helper for expandAddRecExprLiterally. Expand
820
/// the base addrec, which is the addrec without any non-loop-dominating
821
/// values, and return the PHI.
823
SCEVExpander::getAddRecExprPHILiterally(const SCEVAddRecExpr *Normalized,
825
const Type *ExpandTy,
827
// Reuse a previously-inserted PHI, if present.
828
for (BasicBlock::iterator I = L->getHeader()->begin();
829
PHINode *PN = dyn_cast<PHINode>(I); ++I)
830
if (SE.isSCEVable(PN->getType()) &&
831
(SE.getEffectiveSCEVType(PN->getType()) ==
832
SE.getEffectiveSCEVType(Normalized->getType())) &&
833
SE.getSCEV(PN) == Normalized)
834
if (BasicBlock *LatchBlock = L->getLoopLatch()) {
836
cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock));
838
// Determine if this is a well-behaved chain of instructions leading
839
// back to the PHI. It probably will be, if we're scanning an inner
840
// loop already visited by LSR for example, but it wouldn't have
843
if (IncV->getNumOperands() == 0 || isa<PHINode>(IncV)) {
847
// If any of the operands don't dominate the insert position, bail.
848
// Addrec operands are always loop-invariant, so this can only happen
849
// if there are instructions which haven't been hoisted.
850
for (User::op_iterator OI = IncV->op_begin()+1,
851
OE = IncV->op_end(); OI != OE; ++OI)
852
if (Instruction *OInst = dyn_cast<Instruction>(OI))
853
if (!SE.DT->dominates(OInst, IVIncInsertPos)) {
859
// Advance to the next instruction.
860
IncV = dyn_cast<Instruction>(IncV->getOperand(0));
863
if (IncV->mayHaveSideEffects()) {
867
} while (IncV != PN);
870
// Ok, the add recurrence looks usable.
871
// Remember this PHI, even in post-inc mode.
872
InsertedValues.insert(PN);
873
// Remember the increment.
874
IncV = cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock));
875
rememberInstruction(IncV);
876
if (L == IVIncInsertLoop)
878
if (SE.DT->dominates(IncV, IVIncInsertPos))
880
// Make sure the increment is where we want it. But don't move it
881
// down past a potential existing post-inc user.
882
IncV->moveBefore(IVIncInsertPos);
883
IVIncInsertPos = IncV;
884
IncV = cast<Instruction>(IncV->getOperand(0));
885
} while (IncV != PN);
890
// Save the original insertion point so we can restore it when we're done.
891
BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
892
BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
894
// Expand code for the start value.
895
Value *StartV = expandCodeFor(Normalized->getStart(), ExpandTy,
896
L->getHeader()->begin());
898
// Expand code for the step value. Insert instructions right before the
899
// terminator corresponding to the back-edge. Do this before creating the PHI
900
// so that PHI reuse code doesn't see an incomplete PHI. If the stride is
901
// negative, insert a sub instead of an add for the increment (unless it's a
902
// constant, because subtracts of constants are canonicalized to adds).
903
const SCEV *Step = Normalized->getStepRecurrence(SE);
904
bool isPointer = ExpandTy->isPointerTy();
905
bool isNegative = !isPointer && isNonConstantNegative(Step);
907
Step = SE.getNegativeSCEV(Step);
908
Value *StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin());
911
Builder.SetInsertPoint(L->getHeader(), L->getHeader()->begin());
912
PHINode *PN = Builder.CreatePHI(ExpandTy, "lsr.iv");
913
rememberInstruction(PN);
915
// Create the step instructions and populate the PHI.
916
BasicBlock *Header = L->getHeader();
917
for (pred_iterator HPI = pred_begin(Header), HPE = pred_end(Header);
919
BasicBlock *Pred = *HPI;
921
// Add a start value.
922
if (!L->contains(Pred)) {
923
PN->addIncoming(StartV, Pred);
927
// Create a step value and add it to the PHI. If IVIncInsertLoop is
928
// non-null and equal to the addrec's loop, insert the instructions
929
// at IVIncInsertPos.
930
Instruction *InsertPos = L == IVIncInsertLoop ?
931
IVIncInsertPos : Pred->getTerminator();
932
Builder.SetInsertPoint(InsertPos->getParent(), InsertPos);
934
// If the PHI is a pointer, use a GEP, otherwise use an add or sub.
936
const PointerType *GEPPtrTy = cast<PointerType>(ExpandTy);
937
// If the step isn't constant, don't use an implicitly scaled GEP, because
938
// that would require a multiply inside the loop.
939
if (!isa<ConstantInt>(StepV))
940
GEPPtrTy = PointerType::get(Type::getInt1Ty(SE.getContext()),
941
GEPPtrTy->getAddressSpace());
942
const SCEV *const StepArray[1] = { SE.getSCEV(StepV) };
943
IncV = expandAddToGEP(StepArray, StepArray+1, GEPPtrTy, IntTy, PN);
944
if (IncV->getType() != PN->getType()) {
945
IncV = Builder.CreateBitCast(IncV, PN->getType(), "tmp");
946
rememberInstruction(IncV);
950
Builder.CreateSub(PN, StepV, "lsr.iv.next") :
951
Builder.CreateAdd(PN, StepV, "lsr.iv.next");
952
rememberInstruction(IncV);
954
PN->addIncoming(IncV, Pred);
957
// Restore the original insert point.
959
restoreInsertPoint(SaveInsertBB, SaveInsertPt);
961
// Remember this PHI, even in post-inc mode.
962
InsertedValues.insert(PN);
967
Value *SCEVExpander::expandAddRecExprLiterally(const SCEVAddRecExpr *S) {
968
const Type *STy = S->getType();
969
const Type *IntTy = SE.getEffectiveSCEVType(STy);
970
const Loop *L = S->getLoop();
972
// Determine a normalized form of this expression, which is the expression
973
// before any post-inc adjustment is made.
974
const SCEVAddRecExpr *Normalized = S;
975
if (L == PostIncLoop) {
976
const SCEV *Step = S->getStepRecurrence(SE);
977
Normalized = cast<SCEVAddRecExpr>(SE.getMinusSCEV(S, Step));
980
// Strip off any non-loop-dominating component from the addrec start.
981
const SCEV *Start = Normalized->getStart();
982
const SCEV *PostLoopOffset = 0;
983
if (!Start->properlyDominates(L->getHeader(), SE.DT)) {
984
PostLoopOffset = Start;
985
Start = SE.getIntegerSCEV(0, Normalized->getType());
987
cast<SCEVAddRecExpr>(SE.getAddRecExpr(Start,
988
Normalized->getStepRecurrence(SE),
989
Normalized->getLoop()));
992
// Strip off any non-loop-dominating component from the addrec step.
993
const SCEV *Step = Normalized->getStepRecurrence(SE);
994
const SCEV *PostLoopScale = 0;
995
if (!Step->hasComputableLoopEvolution(L) &&
996
!Step->dominates(L->getHeader(), SE.DT)) {
997
PostLoopScale = Step;
998
Step = SE.getIntegerSCEV(1, Normalized->getType());
1000
cast<SCEVAddRecExpr>(SE.getAddRecExpr(Start, Step,
1001
Normalized->getLoop()));
1004
// Expand the core addrec. If we need post-loop scaling, force it to
1005
// expand to an integer type to avoid the need for additional casting.
1006
const Type *ExpandTy = PostLoopScale ? IntTy : STy;
1007
PHINode *PN = getAddRecExprPHILiterally(Normalized, L, ExpandTy, IntTy);
1009
// Accommodate post-inc mode, if necessary.
1011
if (L != PostIncLoop)
1014
// In PostInc mode, use the post-incremented value.
1015
BasicBlock *LatchBlock = L->getLoopLatch();
1016
assert(LatchBlock && "PostInc mode requires a unique loop latch!");
1017
Result = PN->getIncomingValueForBlock(LatchBlock);
1020
// Re-apply any non-loop-dominating scale.
1021
if (PostLoopScale) {
1022
Result = InsertNoopCastOfTo(Result, IntTy);
1023
Result = Builder.CreateMul(Result,
1024
expandCodeFor(PostLoopScale, IntTy));
1025
rememberInstruction(Result);
1028
// Re-apply any non-loop-dominating offset.
1029
if (PostLoopOffset) {
1030
if (const PointerType *PTy = dyn_cast<PointerType>(ExpandTy)) {
1031
const SCEV *const OffsetArray[1] = { PostLoopOffset };
1032
Result = expandAddToGEP(OffsetArray, OffsetArray+1, PTy, IntTy, Result);
1034
Result = InsertNoopCastOfTo(Result, IntTy);
1035
Result = Builder.CreateAdd(Result,
1036
expandCodeFor(PostLoopOffset, IntTy));
1037
rememberInstruction(Result);
1044
Value *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) {
1045
if (!CanonicalMode) return expandAddRecExprLiterally(S);
1047
const Type *Ty = SE.getEffectiveSCEVType(S->getType());
1048
const Loop *L = S->getLoop();
1050
// First check for an existing canonical IV in a suitable type.
1051
PHINode *CanonicalIV = 0;
1052
if (PHINode *PN = L->getCanonicalInductionVariable())
1053
if (SE.isSCEVable(PN->getType()) &&
1054
SE.getEffectiveSCEVType(PN->getType())->isIntegerTy() &&
1055
SE.getTypeSizeInBits(PN->getType()) >= SE.getTypeSizeInBits(Ty))
1058
// Rewrite an AddRec in terms of the canonical induction variable, if
1059
// its type is more narrow.
1061
SE.getTypeSizeInBits(CanonicalIV->getType()) >
1062
SE.getTypeSizeInBits(Ty)) {
1063
const SmallVectorImpl<const SCEV *> &Ops = S->getOperands();
1064
SmallVector<const SCEV *, 4> NewOps(Ops.size());
1065
for (unsigned i = 0, e = Ops.size(); i != e; ++i)
1066
NewOps[i] = SE.getAnyExtendExpr(Ops[i], CanonicalIV->getType());
1067
Value *V = expand(SE.getAddRecExpr(NewOps, S->getLoop()));
1068
BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1069
BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1070
BasicBlock::iterator NewInsertPt =
1071
llvm::next(BasicBlock::iterator(cast<Instruction>(V)));
1072
while (isa<PHINode>(NewInsertPt)) ++NewInsertPt;
1073
V = expandCodeFor(SE.getTruncateExpr(SE.getUnknown(V), Ty), 0,
1075
restoreInsertPoint(SaveInsertBB, SaveInsertPt);
1079
// {X,+,F} --> X + {0,+,F}
1080
if (!S->getStart()->isZero()) {
1081
const SmallVectorImpl<const SCEV *> &SOperands = S->getOperands();
1082
SmallVector<const SCEV *, 4> NewOps(SOperands.begin(), SOperands.end());
1083
NewOps[0] = SE.getIntegerSCEV(0, Ty);
1084
const SCEV *Rest = SE.getAddRecExpr(NewOps, L);
1086
// Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the
1087
// comments on expandAddToGEP for details.
1088
const SCEV *Base = S->getStart();
1089
const SCEV *RestArray[1] = { Rest };
1090
// Dig into the expression to find the pointer base for a GEP.
1091
ExposePointerBase(Base, RestArray[0], SE);
1092
// If we found a pointer, expand the AddRec with a GEP.
1093
if (const PointerType *PTy = dyn_cast<PointerType>(Base->getType())) {
1094
// Make sure the Base isn't something exotic, such as a multiplied
1095
// or divided pointer value. In those cases, the result type isn't
1096
// actually a pointer type.
1097
if (!isa<SCEVMulExpr>(Base) && !isa<SCEVUDivExpr>(Base)) {
1098
Value *StartV = expand(Base);
1099
assert(StartV->getType() == PTy && "Pointer type mismatch for GEP!");
1100
return expandAddToGEP(RestArray, RestArray+1, PTy, Ty, StartV);
1104
// Just do a normal add. Pre-expand the operands to suppress folding.
1105
return expand(SE.getAddExpr(SE.getUnknown(expand(S->getStart())),
1106
SE.getUnknown(expand(Rest))));
1109
// {0,+,1} --> Insert a canonical induction variable into the loop!
1110
if (S->isAffine() &&
1111
S->getOperand(1) == SE.getIntegerSCEV(1, Ty)) {
1112
// If there's a canonical IV, just use it.
1114
assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) &&
1115
"IVs with types different from the canonical IV should "
1116
"already have been handled!");
1120
// Create and insert the PHI node for the induction variable in the
1122
BasicBlock *Header = L->getHeader();
1123
PHINode *PN = PHINode::Create(Ty, "indvar", Header->begin());
1124
rememberInstruction(PN);
1126
Constant *One = ConstantInt::get(Ty, 1);
1127
for (pred_iterator HPI = pred_begin(Header), HPE = pred_end(Header);
1129
if (L->contains(*HPI)) {
1130
// Insert a unit add instruction right before the terminator
1131
// corresponding to the back-edge.
1132
Instruction *Add = BinaryOperator::CreateAdd(PN, One, "indvar.next",
1133
(*HPI)->getTerminator());
1134
rememberInstruction(Add);
1135
PN->addIncoming(Add, *HPI);
1137
PN->addIncoming(Constant::getNullValue(Ty), *HPI);
1141
// {0,+,F} --> {0,+,1} * F
1142
// Get the canonical induction variable I for this loop.
1143
Value *I = CanonicalIV ?
1145
getOrInsertCanonicalInductionVariable(L, Ty);
1147
// If this is a simple linear addrec, emit it now as a special case.
1148
if (S->isAffine()) // {0,+,F} --> i*F
1150
expand(SE.getTruncateOrNoop(
1151
SE.getMulExpr(SE.getUnknown(I),
1152
SE.getNoopOrAnyExtend(S->getOperand(1),
1156
// If this is a chain of recurrences, turn it into a closed form, using the
1157
// folders, then expandCodeFor the closed form. This allows the folders to
1158
// simplify the expression without having to build a bunch of special code
1159
// into this folder.
1160
const SCEV *IH = SE.getUnknown(I); // Get I as a "symbolic" SCEV.
1162
// Promote S up to the canonical IV type, if the cast is foldable.
1163
const SCEV *NewS = S;
1164
const SCEV *Ext = SE.getNoopOrAnyExtend(S, I->getType());
1165
if (isa<SCEVAddRecExpr>(Ext))
1168
const SCEV *V = cast<SCEVAddRecExpr>(NewS)->evaluateAtIteration(IH, SE);
1169
//cerr << "Evaluated: " << *this << "\n to: " << *V << "\n";
1171
// Truncate the result down to the original type, if needed.
1172
const SCEV *T = SE.getTruncateOrNoop(V, Ty);
1176
Value *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) {
1177
const Type *Ty = SE.getEffectiveSCEVType(S->getType());
1178
Value *V = expandCodeFor(S->getOperand(),
1179
SE.getEffectiveSCEVType(S->getOperand()->getType()));
1180
Value *I = Builder.CreateTrunc(V, Ty, "tmp");
1181
rememberInstruction(I);
1185
Value *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) {
1186
const Type *Ty = SE.getEffectiveSCEVType(S->getType());
1187
Value *V = expandCodeFor(S->getOperand(),
1188
SE.getEffectiveSCEVType(S->getOperand()->getType()));
1189
Value *I = Builder.CreateZExt(V, Ty, "tmp");
1190
rememberInstruction(I);
1194
Value *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) {
1195
const Type *Ty = SE.getEffectiveSCEVType(S->getType());
1196
Value *V = expandCodeFor(S->getOperand(),
1197
SE.getEffectiveSCEVType(S->getOperand()->getType()));
1198
Value *I = Builder.CreateSExt(V, Ty, "tmp");
1199
rememberInstruction(I);
1203
Value *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) {
1204
Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
1205
const Type *Ty = LHS->getType();
1206
for (int i = S->getNumOperands()-2; i >= 0; --i) {
1207
// In the case of mixed integer and pointer types, do the
1208
// rest of the comparisons as integer.
1209
if (S->getOperand(i)->getType() != Ty) {
1210
Ty = SE.getEffectiveSCEVType(Ty);
1211
LHS = InsertNoopCastOfTo(LHS, Ty);
1213
Value *RHS = expandCodeFor(S->getOperand(i), Ty);
1214
Value *ICmp = Builder.CreateICmpSGT(LHS, RHS, "tmp");
1215
rememberInstruction(ICmp);
1216
Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "smax");
1217
rememberInstruction(Sel);
1220
// In the case of mixed integer and pointer types, cast the
1221
// final result back to the pointer type.
1222
if (LHS->getType() != S->getType())
1223
LHS = InsertNoopCastOfTo(LHS, S->getType());
1227
Value *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) {
1228
Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
1229
const Type *Ty = LHS->getType();
1230
for (int i = S->getNumOperands()-2; i >= 0; --i) {
1231
// In the case of mixed integer and pointer types, do the
1232
// rest of the comparisons as integer.
1233
if (S->getOperand(i)->getType() != Ty) {
1234
Ty = SE.getEffectiveSCEVType(Ty);
1235
LHS = InsertNoopCastOfTo(LHS, Ty);
1237
Value *RHS = expandCodeFor(S->getOperand(i), Ty);
1238
Value *ICmp = Builder.CreateICmpUGT(LHS, RHS, "tmp");
1239
rememberInstruction(ICmp);
1240
Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "umax");
1241
rememberInstruction(Sel);
1244
// In the case of mixed integer and pointer types, cast the
1245
// final result back to the pointer type.
1246
if (LHS->getType() != S->getType())
1247
LHS = InsertNoopCastOfTo(LHS, S->getType());
1251
Value *SCEVExpander::expandCodeFor(const SCEV *SH, const Type *Ty) {
1252
// Expand the code for this SCEV.
1253
Value *V = expand(SH);
1255
assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) &&
1256
"non-trivial casts should be done with the SCEVs directly!");
1257
V = InsertNoopCastOfTo(V, Ty);
1262
Value *SCEVExpander::expand(const SCEV *S) {
1263
// Compute an insertion point for this SCEV object. Hoist the instructions
1264
// as far out in the loop nest as possible.
1265
Instruction *InsertPt = Builder.GetInsertPoint();
1266
for (Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock()); ;
1267
L = L->getParentLoop())
1268
if (S->isLoopInvariant(L)) {
1270
if (BasicBlock *Preheader = L->getLoopPreheader()) {
1271
InsertPt = Preheader->getTerminator();
1272
BasicBlock::iterator IP = InsertPt;
1273
// Back past any debug info instructions. Sometimes we inserted
1274
// something earlier before debug info but after any real instructions.
1275
// This should behave the same as if debug info was not present.
1276
while (IP != Preheader->begin()) {
1278
if (!isa<DbgInfoIntrinsic>(IP))
1284
// If the SCEV is computable at this level, insert it into the header
1285
// after the PHIs (and after any other instructions that we've inserted
1286
// there) so that it is guaranteed to dominate any user inside the loop.
1287
if (L && S->hasComputableLoopEvolution(L) && L != PostIncLoop)
1288
InsertPt = L->getHeader()->getFirstNonPHI();
1289
while (isInsertedInstruction(InsertPt))
1290
InsertPt = llvm::next(BasicBlock::iterator(InsertPt));
1294
// Check to see if we already expanded this here.
1295
std::map<std::pair<const SCEV *, Instruction *>,
1296
AssertingVH<Value> >::iterator I =
1297
InsertedExpressions.find(std::make_pair(S, InsertPt));
1298
if (I != InsertedExpressions.end())
1301
BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1302
BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1303
Builder.SetInsertPoint(InsertPt->getParent(), InsertPt);
1305
// Expand the expression into instructions.
1306
Value *V = visit(S);
1308
// Remember the expanded value for this SCEV at this location.
1310
InsertedExpressions[std::make_pair(S, InsertPt)] = V;
1312
restoreInsertPoint(SaveInsertBB, SaveInsertPt);
1316
void SCEVExpander::rememberInstruction(Value *I) {
1318
InsertedValues.insert(I);
1320
// If we just claimed an existing instruction and that instruction had
1321
// been the insert point, adjust the insert point forward so that
1322
// subsequently inserted code will be dominated.
1323
if (Builder.GetInsertPoint() == I) {
1324
BasicBlock::iterator It = cast<Instruction>(I);
1325
do { ++It; } while (isInsertedInstruction(It));
1326
Builder.SetInsertPoint(Builder.GetInsertBlock(), It);
1330
void SCEVExpander::restoreInsertPoint(BasicBlock *BB, BasicBlock::iterator I) {
1331
// If we acquired more instructions since the old insert point was saved,
1332
// advance past them.
1333
while (isInsertedInstruction(I)) ++I;
1335
Builder.SetInsertPoint(BB, I);
1338
/// getOrInsertCanonicalInductionVariable - This method returns the
1339
/// canonical induction variable of the specified type for the specified
1340
/// loop (inserting one if there is none). A canonical induction variable
1341
/// starts at zero and steps by one on each iteration.
1343
SCEVExpander::getOrInsertCanonicalInductionVariable(const Loop *L,
1345
assert(Ty->isIntegerTy() && "Can only insert integer induction variables!");
1346
const SCEV *H = SE.getAddRecExpr(SE.getIntegerSCEV(0, Ty),
1347
SE.getIntegerSCEV(1, Ty), L);
1348
BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1349
BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1350
Value *V = expandCodeFor(H, 0, L->getHeader()->begin());
1352
restoreInsertPoint(SaveInsertBB, SaveInsertPt);