1 //===- LoopStrengthReduce.cpp - Strength Reduce IVs in Loops --------------===//
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 transformation analyzes and transforms the induction variables (and
11 // computations derived from them) into forms suitable for efficient execution
14 // This pass performs a strength reduction on array references inside loops that
15 // have as one or more of their components the loop induction variable, it
16 // rewrites expressions to take advantage of scaled-index addressing modes
17 // available on the target, and it performs a variety of other optimizations
18 // related to loop induction variables.
20 //===----------------------------------------------------------------------===//
22 #define DEBUG_TYPE "loop-reduce"
23 #include "llvm/Transforms/Scalar.h"
24 #include "llvm/Constants.h"
25 #include "llvm/Instructions.h"
26 #include "llvm/IntrinsicInst.h"
27 #include "llvm/Type.h"
28 #include "llvm/DerivedTypes.h"
29 #include "llvm/Analysis/IVUsers.h"
30 #include "llvm/Analysis/LoopPass.h"
31 #include "llvm/Analysis/ScalarEvolutionExpander.h"
32 #include "llvm/Transforms/Utils/AddrModeMatcher.h"
33 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
34 #include "llvm/Transforms/Utils/Local.h"
35 #include "llvm/ADT/Statistic.h"
36 #include "llvm/Support/CFG.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Support/CommandLine.h"
39 #include "llvm/Support/ValueHandle.h"
40 #include "llvm/Support/raw_ostream.h"
41 #include "llvm/Target/TargetLowering.h"
45 STATISTIC(NumReduced , "Number of IV uses strength reduced");
46 STATISTIC(NumInserted, "Number of PHIs inserted");
47 STATISTIC(NumVariable, "Number of PHIs with variable strides");
48 STATISTIC(NumEliminated, "Number of strides eliminated");
49 STATISTIC(NumShadow, "Number of Shadow IVs optimized");
50 STATISTIC(NumImmSunk, "Number of common expr immediates sunk into uses");
51 STATISTIC(NumLoopCond, "Number of loop terminating conds optimized");
52 STATISTIC(NumCountZero, "Number of count iv optimized to count toward zero");
54 static cl::opt<bool> EnableFullLSRMode("enable-full-lsr",
62 /// IVInfo - This structure keeps track of one IV expression inserted during
63 /// StrengthReduceStridedIVUsers. It contains the stride, the common base, as
64 /// well as the PHI node and increment value created for rewrite.
70 IVExpr(const SCEV *const stride, const SCEV *const base, PHINode *phi)
71 : Stride(stride), Base(base), PHI(phi) {}
74 /// IVsOfOneStride - This structure keeps track of all IV expression inserted
75 /// during StrengthReduceStridedIVUsers for a particular stride of the IV.
76 struct IVsOfOneStride {
77 std::vector<IVExpr> IVs;
79 void addIV(const SCEV *const Stride, const SCEV *const Base, PHINode *PHI) {
80 IVs.push_back(IVExpr(Stride, Base, PHI));
84 class LoopStrengthReduce : public LoopPass {
89 /// IVsByStride - Keep track of all IVs that have been inserted for a
90 /// particular stride.
91 std::map<const SCEV *, IVsOfOneStride> IVsByStride;
93 /// DeadInsts - Keep track of instructions we may have made dead, so that
94 /// we can remove them after we are done working.
95 SmallVector<WeakVH, 16> DeadInsts;
97 /// TLI - Keep a pointer of a TargetLowering to consult for determining
98 /// transformation profitability.
99 const TargetLowering *TLI;
102 static char ID; // Pass ID, replacement for typeid
103 explicit LoopStrengthReduce(const TargetLowering *tli = NULL) :
104 LoopPass(&ID), TLI(tli) {}
106 bool runOnLoop(Loop *L, LPPassManager &LPM);
108 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
109 // We split critical edges, so we change the CFG. However, we do update
110 // many analyses if they are around.
111 AU.addPreservedID(LoopSimplifyID);
112 AU.addPreserved("loops");
113 AU.addPreserved("domfrontier");
114 AU.addPreserved("domtree");
116 AU.addRequiredID(LoopSimplifyID);
117 AU.addRequired<ScalarEvolution>();
118 AU.addPreserved<ScalarEvolution>();
119 AU.addRequired<IVUsers>();
120 AU.addPreserved<IVUsers>();
124 void OptimizeIndvars(Loop *L);
126 /// OptimizeLoopTermCond - Change loop terminating condition to use the
127 /// postinc iv when possible.
128 void OptimizeLoopTermCond(Loop *L);
130 /// OptimizeShadowIV - If IV is used in a int-to-float cast
131 /// inside the loop then try to eliminate the cast opeation.
132 void OptimizeShadowIV(Loop *L);
134 /// OptimizeMax - Rewrite the loop's terminating condition
135 /// if it uses a max computation.
136 ICmpInst *OptimizeMax(Loop *L, ICmpInst *Cond,
137 IVStrideUse* &CondUse);
139 /// OptimizeLoopCountIV - If, after all sharing of IVs, the IV used for
140 /// deciding when to exit the loop is used only for that purpose, try to
141 /// rearrange things so it counts down to a test against zero.
142 bool OptimizeLoopCountIV(Loop *L);
143 bool OptimizeLoopCountIVOfStride(const SCEV* &Stride,
144 IVStrideUse* &CondUse, Loop *L);
146 /// StrengthReduceIVUsersOfStride - Strength reduce all of the users of a
147 /// single stride of IV. All of the users may have different starting
148 /// values, and this may not be the only stride.
149 void StrengthReduceIVUsersOfStride(const SCEV *const &Stride,
150 IVUsersOfOneStride &Uses,
152 void StrengthReduceIVUsers(Loop *L);
154 ICmpInst *ChangeCompareStride(Loop *L, ICmpInst *Cond,
155 IVStrideUse* &CondUse,
156 const SCEV* &CondStride,
157 bool PostPass = false);
159 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse,
160 const SCEV* &CondStride);
161 bool RequiresTypeConversion(const Type *Ty, const Type *NewTy);
162 const SCEV *CheckForIVReuse(bool, bool, bool, const SCEV *const&,
163 IVExpr&, const Type*,
164 const std::vector<BasedUser>& UsersToProcess);
165 bool ValidScale(bool, int64_t,
166 const std::vector<BasedUser>& UsersToProcess);
167 bool ValidOffset(bool, int64_t, int64_t,
168 const std::vector<BasedUser>& UsersToProcess);
169 const SCEV *CollectIVUsers(const SCEV *const &Stride,
170 IVUsersOfOneStride &Uses,
172 bool &AllUsesAreAddresses,
173 bool &AllUsesAreOutsideLoop,
174 std::vector<BasedUser> &UsersToProcess);
175 bool StrideMightBeShared(const SCEV *Stride, Loop *L, bool CheckPreInc);
176 bool ShouldUseFullStrengthReductionMode(
177 const std::vector<BasedUser> &UsersToProcess,
179 bool AllUsesAreAddresses,
181 void PrepareToStrengthReduceFully(
182 std::vector<BasedUser> &UsersToProcess,
184 const SCEV *CommonExprs,
186 SCEVExpander &PreheaderRewriter);
187 void PrepareToStrengthReduceFromSmallerStride(
188 std::vector<BasedUser> &UsersToProcess,
190 const IVExpr &ReuseIV,
191 Instruction *PreInsertPt);
192 void PrepareToStrengthReduceWithNewPhi(
193 std::vector<BasedUser> &UsersToProcess,
195 const SCEV *CommonExprs,
197 Instruction *IVIncInsertPt,
199 SCEVExpander &PreheaderRewriter);
201 void DeleteTriviallyDeadInstructions();
205 char LoopStrengthReduce::ID = 0;
206 static RegisterPass<LoopStrengthReduce>
207 X("loop-reduce", "Loop Strength Reduction");
209 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
210 return new LoopStrengthReduce(TLI);
213 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
214 /// specified set are trivially dead, delete them and see if this makes any of
215 /// their operands subsequently dead.
216 void LoopStrengthReduce::DeleteTriviallyDeadInstructions() {
217 if (DeadInsts.empty()) return;
219 while (!DeadInsts.empty()) {
220 Instruction *I = dyn_cast_or_null<Instruction>(DeadInsts.pop_back_val());
222 if (I == 0 || !isInstructionTriviallyDead(I))
225 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
226 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
229 DeadInsts.push_back(U);
232 I->eraseFromParent();
237 /// containsAddRecFromDifferentLoop - Determine whether expression S involves a
238 /// subexpression that is an AddRec from a loop other than L. An outer loop
239 /// of L is OK, but not an inner loop nor a disjoint loop.
240 static bool containsAddRecFromDifferentLoop(const SCEV *S, Loop *L) {
241 // This is very common, put it first.
242 if (isa<SCEVConstant>(S))
244 if (const SCEVCommutativeExpr *AE = dyn_cast<SCEVCommutativeExpr>(S)) {
245 for (unsigned int i=0; i< AE->getNumOperands(); i++)
246 if (containsAddRecFromDifferentLoop(AE->getOperand(i), L))
250 if (const SCEVAddRecExpr *AE = dyn_cast<SCEVAddRecExpr>(S)) {
251 if (const Loop *newLoop = AE->getLoop()) {
254 // if newLoop is an outer loop of L, this is OK.
255 if (!newLoop->contains(L->getHeader()))
260 if (const SCEVUDivExpr *DE = dyn_cast<SCEVUDivExpr>(S))
261 return containsAddRecFromDifferentLoop(DE->getLHS(), L) ||
262 containsAddRecFromDifferentLoop(DE->getRHS(), L);
264 // SCEVSDivExpr has been backed out temporarily, but will be back; we'll
265 // need this when it is.
266 if (const SCEVSDivExpr *DE = dyn_cast<SCEVSDivExpr>(S))
267 return containsAddRecFromDifferentLoop(DE->getLHS(), L) ||
268 containsAddRecFromDifferentLoop(DE->getRHS(), L);
270 if (const SCEVCastExpr *CE = dyn_cast<SCEVCastExpr>(S))
271 return containsAddRecFromDifferentLoop(CE->getOperand(), L);
275 /// isAddressUse - Returns true if the specified instruction is using the
276 /// specified value as an address.
277 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
278 bool isAddress = isa<LoadInst>(Inst);
279 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
280 if (SI->getOperand(1) == OperandVal)
282 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
283 // Addressing modes can also be folded into prefetches and a variety
285 switch (II->getIntrinsicID()) {
287 case Intrinsic::prefetch:
288 case Intrinsic::x86_sse2_loadu_dq:
289 case Intrinsic::x86_sse2_loadu_pd:
290 case Intrinsic::x86_sse_loadu_ps:
291 case Intrinsic::x86_sse_storeu_ps:
292 case Intrinsic::x86_sse2_storeu_pd:
293 case Intrinsic::x86_sse2_storeu_dq:
294 case Intrinsic::x86_sse2_storel_dq:
295 if (II->getOperand(1) == OperandVal)
303 /// getAccessType - Return the type of the memory being accessed.
304 static const Type *getAccessType(const Instruction *Inst) {
305 const Type *AccessTy = Inst->getType();
306 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
307 AccessTy = SI->getOperand(0)->getType();
308 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
309 // Addressing modes can also be folded into prefetches and a variety
311 switch (II->getIntrinsicID()) {
313 case Intrinsic::x86_sse_storeu_ps:
314 case Intrinsic::x86_sse2_storeu_pd:
315 case Intrinsic::x86_sse2_storeu_dq:
316 case Intrinsic::x86_sse2_storel_dq:
317 AccessTy = II->getOperand(1)->getType();
325 /// BasedUser - For a particular base value, keep information about how we've
326 /// partitioned the expression so far.
328 /// Base - The Base value for the PHI node that needs to be inserted for
329 /// this use. As the use is processed, information gets moved from this
330 /// field to the Imm field (below). BasedUser values are sorted by this
334 /// Inst - The instruction using the induction variable.
337 /// OperandValToReplace - The operand value of Inst to replace with the
339 Value *OperandValToReplace;
341 /// Imm - The immediate value that should be added to the base immediately
342 /// before Inst, because it will be folded into the imm field of the
343 /// instruction. This is also sometimes used for loop-variant values that
344 /// must be added inside the loop.
347 /// Phi - The induction variable that performs the striding that
348 /// should be used for this user.
351 // isUseOfPostIncrementedValue - True if this should use the
352 // post-incremented version of this IV, not the preincremented version.
353 // This can only be set in special cases, such as the terminating setcc
354 // instruction for a loop and uses outside the loop that are dominated by
356 bool isUseOfPostIncrementedValue;
358 BasedUser(IVStrideUse &IVSU, ScalarEvolution *se)
359 : Base(IVSU.getOffset()), Inst(IVSU.getUser()),
360 OperandValToReplace(IVSU.getOperandValToReplace()),
361 Imm(se->getIntegerSCEV(0, Base->getType())),
362 isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue()) {}
364 // Once we rewrite the code to insert the new IVs we want, update the
365 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
367 void RewriteInstructionToUseNewBase(const SCEV *const &NewBase,
368 Instruction *InsertPt,
369 SCEVExpander &Rewriter, Loop *L, Pass *P,
370 SmallVectorImpl<WeakVH> &DeadInsts,
371 ScalarEvolution *SE);
373 Value *InsertCodeForBaseAtPosition(const SCEV *const &NewBase,
375 SCEVExpander &Rewriter,
377 ScalarEvolution *SE);
382 void BasedUser::dump() const {
383 errs() << " Base=" << *Base;
384 errs() << " Imm=" << *Imm;
385 errs() << " Inst: " << *Inst;
388 Value *BasedUser::InsertCodeForBaseAtPosition(const SCEV *const &NewBase,
390 SCEVExpander &Rewriter,
392 ScalarEvolution *SE) {
393 Value *Base = Rewriter.expandCodeFor(NewBase, 0, IP);
395 // Wrap the base in a SCEVUnknown so that ScalarEvolution doesn't try to
397 const SCEV *NewValSCEV = SE->getUnknown(Base);
399 // Always emit the immediate into the same block as the user.
400 NewValSCEV = SE->getAddExpr(NewValSCEV, Imm);
402 return Rewriter.expandCodeFor(NewValSCEV, Ty, IP);
406 // Once we rewrite the code to insert the new IVs we want, update the
407 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
408 // to it. NewBasePt is the last instruction which contributes to the
409 // value of NewBase in the case that it's a diffferent instruction from
410 // the PHI that NewBase is computed from, or null otherwise.
412 void BasedUser::RewriteInstructionToUseNewBase(const SCEV *const &NewBase,
413 Instruction *NewBasePt,
414 SCEVExpander &Rewriter, Loop *L, Pass *P,
415 SmallVectorImpl<WeakVH> &DeadInsts,
416 ScalarEvolution *SE) {
417 if (!isa<PHINode>(Inst)) {
418 // By default, insert code at the user instruction.
419 BasicBlock::iterator InsertPt = Inst;
421 // However, if the Operand is itself an instruction, the (potentially
422 // complex) inserted code may be shared by many users. Because of this, we
423 // want to emit code for the computation of the operand right before its old
424 // computation. This is usually safe, because we obviously used to use the
425 // computation when it was computed in its current block. However, in some
426 // cases (e.g. use of a post-incremented induction variable) the NewBase
427 // value will be pinned to live somewhere after the original computation.
428 // In this case, we have to back off.
430 // If this is a use outside the loop (which means after, since it is based
431 // on a loop indvar) we use the post-incremented value, so that we don't
432 // artificially make the preinc value live out the bottom of the loop.
433 if (!isUseOfPostIncrementedValue && L->contains(Inst->getParent())) {
434 if (NewBasePt && isa<PHINode>(OperandValToReplace)) {
435 InsertPt = NewBasePt;
437 } else if (Instruction *OpInst
438 = dyn_cast<Instruction>(OperandValToReplace)) {
440 while (isa<PHINode>(InsertPt)) ++InsertPt;
443 Value *NewVal = InsertCodeForBaseAtPosition(NewBase,
444 OperandValToReplace->getType(),
445 Rewriter, InsertPt, SE);
446 // Replace the use of the operand Value with the new Phi we just created.
447 Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
449 DEBUG(errs() << " Replacing with ");
450 DEBUG(WriteAsOperand(errs(), NewVal, /*PrintType=*/false));
451 DEBUG(errs() << ", which has value " << *NewBase << " plus IMM "
456 // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm
457 // expression into each operand block that uses it. Note that PHI nodes can
458 // have multiple entries for the same predecessor. We use a map to make sure
459 // that a PHI node only has a single Value* for each predecessor (which also
460 // prevents us from inserting duplicate code in some blocks).
461 DenseMap<BasicBlock*, Value*> InsertedCode;
462 PHINode *PN = cast<PHINode>(Inst);
463 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
464 if (PN->getIncomingValue(i) == OperandValToReplace) {
465 // If the original expression is outside the loop, put the replacement
466 // code in the same place as the original expression,
467 // which need not be an immediate predecessor of this PHI. This way we
468 // need only one copy of it even if it is referenced multiple times in
469 // the PHI. We don't do this when the original expression is inside the
470 // loop because multiple copies sometimes do useful sinking of code in
472 Instruction *OldLoc = dyn_cast<Instruction>(OperandValToReplace);
473 BasicBlock *PHIPred = PN->getIncomingBlock(i);
474 if (L->contains(OldLoc->getParent())) {
475 // If this is a critical edge, split the edge so that we do not insert
476 // the code on all predecessor/successor paths. We do this unless this
477 // is the canonical backedge for this loop, as this can make some
478 // inserted code be in an illegal position.
479 if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 &&
480 !isa<IndirectBrInst>(PHIPred->getTerminator()) &&
481 (PN->getParent() != L->getHeader() || !L->contains(PHIPred))) {
483 // First step, split the critical edge.
484 BasicBlock *NewBB = SplitCriticalEdge(PHIPred, PN->getParent(),
487 // Next step: move the basic block. In particular, if the PHI node
488 // is outside of the loop, and PredTI is in the loop, we want to
489 // move the block to be immediately before the PHI block, not
490 // immediately after PredTI.
491 if (L->contains(PHIPred) && !L->contains(PN->getParent()))
492 NewBB->moveBefore(PN->getParent());
494 // Splitting the edge can reduce the number of PHI entries we have.
495 e = PN->getNumIncomingValues();
497 i = PN->getBasicBlockIndex(PHIPred);
500 Value *&Code = InsertedCode[PHIPred];
502 // Insert the code into the end of the predecessor block.
503 Instruction *InsertPt = (L->contains(OldLoc->getParent())) ?
504 PHIPred->getTerminator() :
505 OldLoc->getParent()->getTerminator();
506 Code = InsertCodeForBaseAtPosition(NewBase, PN->getType(),
507 Rewriter, InsertPt, SE);
509 DEBUG(errs() << " Changing PHI use to ");
510 DEBUG(WriteAsOperand(errs(), Code, /*PrintType=*/false));
511 DEBUG(errs() << ", which has value " << *NewBase << " plus IMM "
515 // Replace the use of the operand Value with the new Phi we just created.
516 PN->setIncomingValue(i, Code);
521 // PHI node might have become a constant value after SplitCriticalEdge.
522 DeadInsts.push_back(Inst);
526 /// fitsInAddressMode - Return true if V can be subsumed within an addressing
527 /// mode, and does not need to be put in a register first.
528 static bool fitsInAddressMode(const SCEV *const &V, const Type *AccessTy,
529 const TargetLowering *TLI, bool HasBaseReg) {
530 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
531 int64_t VC = SC->getValue()->getSExtValue();
533 TargetLowering::AddrMode AM;
535 AM.HasBaseReg = HasBaseReg;
536 return TLI->isLegalAddressingMode(AM, AccessTy);
538 // Defaults to PPC. PPC allows a sign-extended 16-bit immediate field.
539 return (VC > -(1 << 16) && VC < (1 << 16)-1);
543 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
544 if (GlobalValue *GV = dyn_cast<GlobalValue>(SU->getValue())) {
546 TargetLowering::AddrMode AM;
548 AM.HasBaseReg = HasBaseReg;
549 return TLI->isLegalAddressingMode(AM, AccessTy);
551 // Default: assume global addresses are not legal.
558 /// MoveLoopVariantsToImmediateField - Move any subexpressions from Val that are
559 /// loop varying to the Imm operand.
560 static void MoveLoopVariantsToImmediateField(const SCEV *&Val, const SCEV *&Imm,
561 Loop *L, ScalarEvolution *SE) {
562 if (Val->isLoopInvariant(L)) return; // Nothing to do.
564 if (const SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
565 SmallVector<const SCEV *, 4> NewOps;
566 NewOps.reserve(SAE->getNumOperands());
568 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
569 if (!SAE->getOperand(i)->isLoopInvariant(L)) {
570 // If this is a loop-variant expression, it must stay in the immediate
571 // field of the expression.
572 Imm = SE->getAddExpr(Imm, SAE->getOperand(i));
574 NewOps.push_back(SAE->getOperand(i));
578 Val = SE->getIntegerSCEV(0, Val->getType());
580 Val = SE->getAddExpr(NewOps);
581 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
582 // Try to pull immediates out of the start value of nested addrec's.
583 const SCEV *Start = SARE->getStart();
584 MoveLoopVariantsToImmediateField(Start, Imm, L, SE);
586 SmallVector<const SCEV *, 4> Ops(SARE->op_begin(), SARE->op_end());
588 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
590 // Otherwise, all of Val is variant, move the whole thing over.
591 Imm = SE->getAddExpr(Imm, Val);
592 Val = SE->getIntegerSCEV(0, Val->getType());
597 /// MoveImmediateValues - Look at Val, and pull out any additions of constants
598 /// that can fit into the immediate field of instructions in the target.
599 /// Accumulate these immediate values into the Imm value.
600 static void MoveImmediateValues(const TargetLowering *TLI,
601 const Type *AccessTy,
602 const SCEV *&Val, const SCEV *&Imm,
603 bool isAddress, Loop *L,
604 ScalarEvolution *SE) {
605 if (const SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
606 SmallVector<const SCEV *, 4> NewOps;
607 NewOps.reserve(SAE->getNumOperands());
609 for (unsigned i = 0; i != SAE->getNumOperands(); ++i) {
610 const SCEV *NewOp = SAE->getOperand(i);
611 MoveImmediateValues(TLI, AccessTy, NewOp, Imm, isAddress, L, SE);
613 if (!NewOp->isLoopInvariant(L)) {
614 // If this is a loop-variant expression, it must stay in the immediate
615 // field of the expression.
616 Imm = SE->getAddExpr(Imm, NewOp);
618 NewOps.push_back(NewOp);
623 Val = SE->getIntegerSCEV(0, Val->getType());
625 Val = SE->getAddExpr(NewOps);
627 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
628 // Try to pull immediates out of the start value of nested addrec's.
629 const SCEV *Start = SARE->getStart();
630 MoveImmediateValues(TLI, AccessTy, Start, Imm, isAddress, L, SE);
632 if (Start != SARE->getStart()) {
633 SmallVector<const SCEV *, 4> Ops(SARE->op_begin(), SARE->op_end());
635 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
638 } else if (const SCEVMulExpr *SME = dyn_cast<SCEVMulExpr>(Val)) {
639 // Transform "8 * (4 + v)" -> "32 + 8*V" if "32" fits in the immed field.
641 fitsInAddressMode(SME->getOperand(0), AccessTy, TLI, false) &&
642 SME->getNumOperands() == 2 && SME->isLoopInvariant(L)) {
644 const SCEV *SubImm = SE->getIntegerSCEV(0, Val->getType());
645 const SCEV *NewOp = SME->getOperand(1);
646 MoveImmediateValues(TLI, AccessTy, NewOp, SubImm, isAddress, L, SE);
648 // If we extracted something out of the subexpressions, see if we can
650 if (NewOp != SME->getOperand(1)) {
651 // Scale SubImm up by "8". If the result is a target constant, we are
653 SubImm = SE->getMulExpr(SubImm, SME->getOperand(0));
654 if (fitsInAddressMode(SubImm, AccessTy, TLI, false)) {
655 // Accumulate the immediate.
656 Imm = SE->getAddExpr(Imm, SubImm);
658 // Update what is left of 'Val'.
659 Val = SE->getMulExpr(SME->getOperand(0), NewOp);
666 // Loop-variant expressions must stay in the immediate field of the
668 if ((isAddress && fitsInAddressMode(Val, AccessTy, TLI, false)) ||
669 !Val->isLoopInvariant(L)) {
670 Imm = SE->getAddExpr(Imm, Val);
671 Val = SE->getIntegerSCEV(0, Val->getType());
675 // Otherwise, no immediates to move.
678 static void MoveImmediateValues(const TargetLowering *TLI,
680 const SCEV *&Val, const SCEV *&Imm,
681 bool isAddress, Loop *L,
682 ScalarEvolution *SE) {
683 const Type *AccessTy = getAccessType(User);
684 MoveImmediateValues(TLI, AccessTy, Val, Imm, isAddress, L, SE);
687 /// SeparateSubExprs - Decompose Expr into all of the subexpressions that are
688 /// added together. This is used to reassociate common addition subexprs
689 /// together for maximal sharing when rewriting bases.
690 static void SeparateSubExprs(SmallVector<const SCEV *, 16> &SubExprs,
692 ScalarEvolution *SE) {
693 if (const SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) {
694 for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
695 SeparateSubExprs(SubExprs, AE->getOperand(j), SE);
696 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) {
697 const SCEV *Zero = SE->getIntegerSCEV(0, Expr->getType());
698 if (SARE->getOperand(0) == Zero) {
699 SubExprs.push_back(Expr);
701 // Compute the addrec with zero as its base.
702 SmallVector<const SCEV *, 4> Ops(SARE->op_begin(), SARE->op_end());
703 Ops[0] = Zero; // Start with zero base.
704 SubExprs.push_back(SE->getAddRecExpr(Ops, SARE->getLoop()));
707 SeparateSubExprs(SubExprs, SARE->getOperand(0), SE);
709 } else if (!Expr->isZero()) {
711 SubExprs.push_back(Expr);
715 // This is logically local to the following function, but C++ says we have
716 // to make it file scope.
717 struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; };
719 /// RemoveCommonExpressionsFromUseBases - Look through all of the Bases of all
720 /// the Uses, removing any common subexpressions, except that if all such
721 /// subexpressions can be folded into an addressing mode for all uses inside
722 /// the loop (this case is referred to as "free" in comments herein) we do
723 /// not remove anything. This looks for things like (a+b+c) and
724 /// (a+c+d) and computes the common (a+c) subexpression. The common expression
725 /// is *removed* from the Bases and returned.
727 RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses,
728 ScalarEvolution *SE, Loop *L,
729 const TargetLowering *TLI) {
730 unsigned NumUses = Uses.size();
732 // Only one use? This is a very common case, so we handle it specially and
734 const SCEV *Zero = SE->getIntegerSCEV(0, Uses[0].Base->getType());
735 const SCEV *Result = Zero;
736 const SCEV *FreeResult = Zero;
738 // If the use is inside the loop, use its base, regardless of what it is:
739 // it is clearly shared across all the IV's. If the use is outside the loop
740 // (which means after it) we don't want to factor anything *into* the loop,
741 // so just use 0 as the base.
742 if (L->contains(Uses[0].Inst->getParent()))
743 std::swap(Result, Uses[0].Base);
747 // To find common subexpressions, count how many of Uses use each expression.
748 // If any subexpressions are used Uses.size() times, they are common.
749 // Also track whether all uses of each expression can be moved into an
750 // an addressing mode "for free"; such expressions are left within the loop.
751 // struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; };
752 std::map<const SCEV *, SubExprUseData> SubExpressionUseData;
754 // UniqueSubExprs - Keep track of all of the subexpressions we see in the
755 // order we see them.
756 SmallVector<const SCEV *, 16> UniqueSubExprs;
758 SmallVector<const SCEV *, 16> SubExprs;
759 unsigned NumUsesInsideLoop = 0;
760 for (unsigned i = 0; i != NumUses; ++i) {
761 // If the user is outside the loop, just ignore it for base computation.
762 // Since the user is outside the loop, it must be *after* the loop (if it
763 // were before, it could not be based on the loop IV). We don't want users
764 // after the loop to affect base computation of values *inside* the loop,
765 // because we can always add their offsets to the result IV after the loop
766 // is done, ensuring we get good code inside the loop.
767 if (!L->contains(Uses[i].Inst->getParent()))
771 // If the base is zero (which is common), return zero now, there are no
773 if (Uses[i].Base == Zero) return Zero;
775 // If this use is as an address we may be able to put CSEs in the addressing
776 // mode rather than hoisting them.
777 bool isAddrUse = isAddressUse(Uses[i].Inst, Uses[i].OperandValToReplace);
778 // We may need the AccessTy below, but only when isAddrUse, so compute it
779 // only in that case.
780 const Type *AccessTy = 0;
782 AccessTy = getAccessType(Uses[i].Inst);
784 // Split the expression into subexprs.
785 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
786 // Add one to SubExpressionUseData.Count for each subexpr present, and
787 // if the subexpr is not a valid immediate within an addressing mode use,
788 // set SubExpressionUseData.notAllUsesAreFree. We definitely want to
789 // hoist these out of the loop (if they are common to all uses).
790 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) {
791 if (++SubExpressionUseData[SubExprs[j]].Count == 1)
792 UniqueSubExprs.push_back(SubExprs[j]);
793 if (!isAddrUse || !fitsInAddressMode(SubExprs[j], AccessTy, TLI, false))
794 SubExpressionUseData[SubExprs[j]].notAllUsesAreFree = true;
799 // Now that we know how many times each is used, build Result. Iterate over
800 // UniqueSubexprs so that we have a stable ordering.
801 for (unsigned i = 0, e = UniqueSubExprs.size(); i != e; ++i) {
802 std::map<const SCEV *, SubExprUseData>::iterator I =
803 SubExpressionUseData.find(UniqueSubExprs[i]);
804 assert(I != SubExpressionUseData.end() && "Entry not found?");
805 if (I->second.Count == NumUsesInsideLoop) { // Found CSE!
806 if (I->second.notAllUsesAreFree)
807 Result = SE->getAddExpr(Result, I->first);
809 FreeResult = SE->getAddExpr(FreeResult, I->first);
811 // Remove non-cse's from SubExpressionUseData.
812 SubExpressionUseData.erase(I);
815 if (FreeResult != Zero) {
816 // We have some subexpressions that can be subsumed into addressing
817 // modes in every use inside the loop. However, it's possible that
818 // there are so many of them that the combined FreeResult cannot
819 // be subsumed, or that the target cannot handle both a FreeResult
820 // and a Result in the same instruction (for example because it would
821 // require too many registers). Check this.
822 for (unsigned i=0; i<NumUses; ++i) {
823 if (!L->contains(Uses[i].Inst->getParent()))
825 // We know this is an addressing mode use; if there are any uses that
826 // are not, FreeResult would be Zero.
827 const Type *AccessTy = getAccessType(Uses[i].Inst);
828 if (!fitsInAddressMode(FreeResult, AccessTy, TLI, Result!=Zero)) {
829 // FIXME: could split up FreeResult into pieces here, some hoisted
830 // and some not. There is no obvious advantage to this.
831 Result = SE->getAddExpr(Result, FreeResult);
838 // If we found no CSE's, return now.
839 if (Result == Zero) return Result;
841 // If we still have a FreeResult, remove its subexpressions from
842 // SubExpressionUseData. This means they will remain in the use Bases.
843 if (FreeResult != Zero) {
844 SeparateSubExprs(SubExprs, FreeResult, SE);
845 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) {
846 std::map<const SCEV *, SubExprUseData>::iterator I =
847 SubExpressionUseData.find(SubExprs[j]);
848 SubExpressionUseData.erase(I);
853 // Otherwise, remove all of the CSE's we found from each of the base values.
854 for (unsigned i = 0; i != NumUses; ++i) {
855 // Uses outside the loop don't necessarily include the common base, but
856 // the final IV value coming into those uses does. Instead of trying to
857 // remove the pieces of the common base, which might not be there,
858 // subtract off the base to compensate for this.
859 if (!L->contains(Uses[i].Inst->getParent())) {
860 Uses[i].Base = SE->getMinusSCEV(Uses[i].Base, Result);
864 // Split the expression into subexprs.
865 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
867 // Remove any common subexpressions.
868 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
869 if (SubExpressionUseData.count(SubExprs[j])) {
870 SubExprs.erase(SubExprs.begin()+j);
874 // Finally, add the non-shared expressions together.
875 if (SubExprs.empty())
878 Uses[i].Base = SE->getAddExpr(SubExprs);
885 /// ValidScale - Check whether the given Scale is valid for all loads and
886 /// stores in UsersToProcess.
888 bool LoopStrengthReduce::ValidScale(bool HasBaseReg, int64_t Scale,
889 const std::vector<BasedUser>& UsersToProcess) {
893 for (unsigned i = 0, e = UsersToProcess.size(); i!=e; ++i) {
894 // If this is a load or other access, pass the type of the access in.
895 const Type *AccessTy =
896 Type::getVoidTy(UsersToProcess[i].Inst->getContext());
897 if (isAddressUse(UsersToProcess[i].Inst,
898 UsersToProcess[i].OperandValToReplace))
899 AccessTy = getAccessType(UsersToProcess[i].Inst);
900 else if (isa<PHINode>(UsersToProcess[i].Inst))
903 TargetLowering::AddrMode AM;
904 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
905 AM.BaseOffs = SC->getValue()->getSExtValue();
906 AM.HasBaseReg = HasBaseReg || !UsersToProcess[i].Base->isZero();
909 // If load[imm+r*scale] is illegal, bail out.
910 if (!TLI->isLegalAddressingMode(AM, AccessTy))
916 /// ValidOffset - Check whether the given Offset is valid for all loads and
917 /// stores in UsersToProcess.
919 bool LoopStrengthReduce::ValidOffset(bool HasBaseReg,
922 const std::vector<BasedUser>& UsersToProcess) {
926 for (unsigned i=0, e = UsersToProcess.size(); i!=e; ++i) {
927 // If this is a load or other access, pass the type of the access in.
928 const Type *AccessTy =
929 Type::getVoidTy(UsersToProcess[i].Inst->getContext());
930 if (isAddressUse(UsersToProcess[i].Inst,
931 UsersToProcess[i].OperandValToReplace))
932 AccessTy = getAccessType(UsersToProcess[i].Inst);
933 else if (isa<PHINode>(UsersToProcess[i].Inst))
936 TargetLowering::AddrMode AM;
937 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
938 AM.BaseOffs = SC->getValue()->getSExtValue();
939 AM.BaseOffs = (uint64_t)AM.BaseOffs + (uint64_t)Offset;
940 AM.HasBaseReg = HasBaseReg || !UsersToProcess[i].Base->isZero();
943 // If load[imm+r*scale] is illegal, bail out.
944 if (!TLI->isLegalAddressingMode(AM, AccessTy))
950 /// RequiresTypeConversion - Returns true if converting Ty1 to Ty2 is not
952 bool LoopStrengthReduce::RequiresTypeConversion(const Type *Ty1,
956 Ty1 = SE->getEffectiveSCEVType(Ty1);
957 Ty2 = SE->getEffectiveSCEVType(Ty2);
960 if (Ty1->canLosslesslyBitCastTo(Ty2))
962 if (TLI && TLI->isTruncateFree(Ty1, Ty2))
967 /// CheckForIVReuse - Returns the multiple if the stride is the multiple
968 /// of a previous stride and it is a legal value for the target addressing
969 /// mode scale component and optional base reg. This allows the users of
970 /// this stride to be rewritten as prev iv * factor. It returns 0 if no
971 /// reuse is possible. Factors can be negative on same targets, e.g. ARM.
973 /// If all uses are outside the loop, we don't require that all multiplies
974 /// be folded into the addressing mode, nor even that the factor be constant;
975 /// a multiply (executed once) outside the loop is better than another IV
976 /// within. Well, usually.
977 const SCEV *LoopStrengthReduce::CheckForIVReuse(bool HasBaseReg,
978 bool AllUsesAreAddresses,
979 bool AllUsesAreOutsideLoop,
980 const SCEV *const &Stride,
981 IVExpr &IV, const Type *Ty,
982 const std::vector<BasedUser>& UsersToProcess) {
983 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Stride)) {
984 int64_t SInt = SC->getValue()->getSExtValue();
985 for (unsigned NewStride = 0, e = IU->StrideOrder.size();
986 NewStride != e; ++NewStride) {
987 std::map<const SCEV *, IVsOfOneStride>::iterator SI =
988 IVsByStride.find(IU->StrideOrder[NewStride]);
989 if (SI == IVsByStride.end() || !isa<SCEVConstant>(SI->first))
991 // The other stride has no uses, don't reuse it.
992 std::map<const SCEV *, IVUsersOfOneStride *>::iterator UI =
993 IU->IVUsesByStride.find(IU->StrideOrder[NewStride]);
994 if (UI->second->Users.empty())
996 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
997 if (SI->first != Stride &&
998 (unsigned(abs64(SInt)) < SSInt || (SInt % SSInt) != 0))
1000 int64_t Scale = SInt / SSInt;
1001 // Check that this stride is valid for all the types used for loads and
1002 // stores; if it can be used for some and not others, we might as well use
1003 // the original stride everywhere, since we have to create the IV for it
1004 // anyway. If the scale is 1, then we don't need to worry about folding
1007 (AllUsesAreAddresses &&
1008 ValidScale(HasBaseReg, Scale, UsersToProcess))) {
1009 // Prefer to reuse an IV with a base of zero.
1010 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1011 IE = SI->second.IVs.end(); II != IE; ++II)
1012 // Only reuse previous IV if it would not require a type conversion
1013 // and if the base difference can be folded.
1014 if (II->Base->isZero() &&
1015 !RequiresTypeConversion(II->Base->getType(), Ty)) {
1017 return SE->getIntegerSCEV(Scale, Stride->getType());
1019 // Otherwise, settle for an IV with a foldable base.
1020 if (AllUsesAreAddresses)
1021 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1022 IE = SI->second.IVs.end(); II != IE; ++II)
1023 // Only reuse previous IV if it would not require a type conversion
1024 // and if the base difference can be folded.
1025 if (SE->getEffectiveSCEVType(II->Base->getType()) ==
1026 SE->getEffectiveSCEVType(Ty) &&
1027 isa<SCEVConstant>(II->Base)) {
1029 cast<SCEVConstant>(II->Base)->getValue()->getSExtValue();
1030 if (Base > INT32_MIN && Base <= INT32_MAX &&
1031 ValidOffset(HasBaseReg, -Base * Scale,
1032 Scale, UsersToProcess)) {
1034 return SE->getIntegerSCEV(Scale, Stride->getType());
1039 } else if (AllUsesAreOutsideLoop) {
1040 // Accept nonconstant strides here; it is really really right to substitute
1041 // an existing IV if we can.
1042 for (unsigned NewStride = 0, e = IU->StrideOrder.size();
1043 NewStride != e; ++NewStride) {
1044 std::map<const SCEV *, IVsOfOneStride>::iterator SI =
1045 IVsByStride.find(IU->StrideOrder[NewStride]);
1046 if (SI == IVsByStride.end() || !isa<SCEVConstant>(SI->first))
1048 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1049 if (SI->first != Stride && SSInt != 1)
1051 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1052 IE = SI->second.IVs.end(); II != IE; ++II)
1053 // Accept nonzero base here.
1054 // Only reuse previous IV if it would not require a type conversion.
1055 if (!RequiresTypeConversion(II->Base->getType(), Ty)) {
1060 // Special case, old IV is -1*x and this one is x. Can treat this one as
1062 for (unsigned NewStride = 0, e = IU->StrideOrder.size();
1063 NewStride != e; ++NewStride) {
1064 std::map<const SCEV *, IVsOfOneStride>::iterator SI =
1065 IVsByStride.find(IU->StrideOrder[NewStride]);
1066 if (SI == IVsByStride.end())
1068 if (const SCEVMulExpr *ME = dyn_cast<SCEVMulExpr>(SI->first))
1069 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(ME->getOperand(0)))
1070 if (Stride == ME->getOperand(1) &&
1071 SC->getValue()->getSExtValue() == -1LL)
1072 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1073 IE = SI->second.IVs.end(); II != IE; ++II)
1074 // Accept nonzero base here.
1075 // Only reuse previous IV if it would not require type conversion.
1076 if (!RequiresTypeConversion(II->Base->getType(), Ty)) {
1078 return SE->getIntegerSCEV(-1LL, Stride->getType());
1082 return SE->getIntegerSCEV(0, Stride->getType());
1085 /// PartitionByIsUseOfPostIncrementedValue - Simple boolean predicate that
1086 /// returns true if Val's isUseOfPostIncrementedValue is true.
1087 static bool PartitionByIsUseOfPostIncrementedValue(const BasedUser &Val) {
1088 return Val.isUseOfPostIncrementedValue;
1091 /// isNonConstantNegative - Return true if the specified scev is negated, but
1093 static bool isNonConstantNegative(const SCEV *const &Expr) {
1094 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Expr);
1095 if (!Mul) return false;
1097 // If there is a constant factor, it will be first.
1098 const SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
1099 if (!SC) return false;
1101 // Return true if the value is negative, this matches things like (-42 * V).
1102 return SC->getValue()->getValue().isNegative();
1105 /// CollectIVUsers - Transform our list of users and offsets to a bit more
1106 /// complex table. In this new vector, each 'BasedUser' contains 'Base', the
1107 /// base of the strided accesses, as well as the old information from Uses. We
1108 /// progressively move information from the Base field to the Imm field, until
1109 /// we eventually have the full access expression to rewrite the use.
1110 const SCEV *LoopStrengthReduce::CollectIVUsers(const SCEV *const &Stride,
1111 IVUsersOfOneStride &Uses,
1113 bool &AllUsesAreAddresses,
1114 bool &AllUsesAreOutsideLoop,
1115 std::vector<BasedUser> &UsersToProcess) {
1116 // FIXME: Generalize to non-affine IV's.
1117 if (!Stride->isLoopInvariant(L))
1118 return SE->getIntegerSCEV(0, Stride->getType());
1120 UsersToProcess.reserve(Uses.Users.size());
1121 for (ilist<IVStrideUse>::iterator I = Uses.Users.begin(),
1122 E = Uses.Users.end(); I != E; ++I) {
1123 UsersToProcess.push_back(BasedUser(*I, SE));
1125 // Move any loop variant operands from the offset field to the immediate
1126 // field of the use, so that we don't try to use something before it is
1128 MoveLoopVariantsToImmediateField(UsersToProcess.back().Base,
1129 UsersToProcess.back().Imm, L, SE);
1130 assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
1131 "Base value is not loop invariant!");
1134 // We now have a whole bunch of uses of like-strided induction variables, but
1135 // they might all have different bases. We want to emit one PHI node for this
1136 // stride which we fold as many common expressions (between the IVs) into as
1137 // possible. Start by identifying the common expressions in the base values
1138 // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
1139 // "A+B"), emit it to the preheader, then remove the expression from the
1140 // UsersToProcess base values.
1141 const SCEV *CommonExprs =
1142 RemoveCommonExpressionsFromUseBases(UsersToProcess, SE, L, TLI);
1144 // Next, figure out what we can represent in the immediate fields of
1145 // instructions. If we can represent anything there, move it to the imm
1146 // fields of the BasedUsers. We do this so that it increases the commonality
1147 // of the remaining uses.
1148 unsigned NumPHI = 0;
1149 bool HasAddress = false;
1150 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1151 // If the user is not in the current loop, this means it is using the exit
1152 // value of the IV. Do not put anything in the base, make sure it's all in
1153 // the immediate field to allow as much factoring as possible.
1154 if (!L->contains(UsersToProcess[i].Inst->getParent())) {
1155 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm,
1156 UsersToProcess[i].Base);
1157 UsersToProcess[i].Base =
1158 SE->getIntegerSCEV(0, UsersToProcess[i].Base->getType());
1160 // Not all uses are outside the loop.
1161 AllUsesAreOutsideLoop = false;
1163 // Addressing modes can be folded into loads and stores. Be careful that
1164 // the store is through the expression, not of the expression though.
1166 bool isAddress = isAddressUse(UsersToProcess[i].Inst,
1167 UsersToProcess[i].OperandValToReplace);
1168 if (isa<PHINode>(UsersToProcess[i].Inst)) {
1176 // If this use isn't an address, then not all uses are addresses.
1177 if (!isAddress && !isPHI)
1178 AllUsesAreAddresses = false;
1180 MoveImmediateValues(TLI, UsersToProcess[i].Inst, UsersToProcess[i].Base,
1181 UsersToProcess[i].Imm, isAddress, L, SE);
1185 // If one of the use is a PHI node and all other uses are addresses, still
1186 // allow iv reuse. Essentially we are trading one constant multiplication
1187 // for one fewer iv.
1189 AllUsesAreAddresses = false;
1191 // There are no in-loop address uses.
1192 if (AllUsesAreAddresses && (!HasAddress && !AllUsesAreOutsideLoop))
1193 AllUsesAreAddresses = false;
1198 /// ShouldUseFullStrengthReductionMode - Test whether full strength-reduction
1199 /// is valid and profitable for the given set of users of a stride. In
1200 /// full strength-reduction mode, all addresses at the current stride are
1201 /// strength-reduced all the way down to pointer arithmetic.
1203 bool LoopStrengthReduce::ShouldUseFullStrengthReductionMode(
1204 const std::vector<BasedUser> &UsersToProcess,
1206 bool AllUsesAreAddresses,
1207 const SCEV *Stride) {
1208 if (!EnableFullLSRMode)
1211 // The heuristics below aim to avoid increasing register pressure, but
1212 // fully strength-reducing all the addresses increases the number of
1213 // add instructions, so don't do this when optimizing for size.
1214 // TODO: If the loop is large, the savings due to simpler addresses
1215 // may oughtweight the costs of the extra increment instructions.
1216 if (L->getHeader()->getParent()->hasFnAttr(Attribute::OptimizeForSize))
1219 // TODO: For now, don't do full strength reduction if there could
1220 // potentially be greater-stride multiples of the current stride
1221 // which could reuse the current stride IV.
1222 if (IU->StrideOrder.back() != Stride)
1225 // Iterate through the uses to find conditions that automatically rule out
1227 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ) {
1228 const SCEV *Base = UsersToProcess[i].Base;
1229 const SCEV *Imm = UsersToProcess[i].Imm;
1230 // If any users have a loop-variant component, they can't be fully
1231 // strength-reduced.
1232 if (Imm && !Imm->isLoopInvariant(L))
1234 // If there are to users with the same base and the difference between
1235 // the two Imm values can't be folded into the address, full
1236 // strength reduction would increase register pressure.
1238 const SCEV *CurImm = UsersToProcess[i].Imm;
1239 if ((CurImm || Imm) && CurImm != Imm) {
1240 if (!CurImm) CurImm = SE->getIntegerSCEV(0, Stride->getType());
1241 if (!Imm) Imm = SE->getIntegerSCEV(0, Stride->getType());
1242 const Instruction *Inst = UsersToProcess[i].Inst;
1243 const Type *AccessTy = getAccessType(Inst);
1244 const SCEV *Diff = SE->getMinusSCEV(UsersToProcess[i].Imm, Imm);
1245 if (!Diff->isZero() &&
1246 (!AllUsesAreAddresses ||
1247 !fitsInAddressMode(Diff, AccessTy, TLI, /*HasBaseReg=*/true)))
1250 } while (++i != e && Base == UsersToProcess[i].Base);
1253 // If there's exactly one user in this stride, fully strength-reducing it
1254 // won't increase register pressure. If it's starting from a non-zero base,
1255 // it'll be simpler this way.
1256 if (UsersToProcess.size() == 1 && !UsersToProcess[0].Base->isZero())
1259 // Otherwise, if there are any users in this stride that don't require
1260 // a register for their base, full strength-reduction will increase
1261 // register pressure.
1262 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1263 if (UsersToProcess[i].Base->isZero())
1266 // Otherwise, go for it.
1270 /// InsertAffinePhi Create and insert a PHI node for an induction variable
1271 /// with the specified start and step values in the specified loop.
1273 /// If NegateStride is true, the stride should be negated by using a
1274 /// subtract instead of an add.
1276 /// Return the created phi node.
1278 static PHINode *InsertAffinePhi(const SCEV *Start, const SCEV *Step,
1279 Instruction *IVIncInsertPt,
1281 SCEVExpander &Rewriter) {
1282 assert(Start->isLoopInvariant(L) && "New PHI start is not loop invariant!");
1283 assert(Step->isLoopInvariant(L) && "New PHI stride is not loop invariant!");
1285 BasicBlock *Header = L->getHeader();
1286 BasicBlock *Preheader = L->getLoopPreheader();
1287 BasicBlock *LatchBlock = L->getLoopLatch();
1288 const Type *Ty = Start->getType();
1289 Ty = Rewriter.SE.getEffectiveSCEVType(Ty);
1291 PHINode *PN = PHINode::Create(Ty, "lsr.iv", Header->begin());
1292 PN->addIncoming(Rewriter.expandCodeFor(Start, Ty, Preheader->getTerminator()),
1295 // If the stride is negative, insert a sub instead of an add for the
1297 bool isNegative = isNonConstantNegative(Step);
1298 const SCEV *IncAmount = Step;
1300 IncAmount = Rewriter.SE.getNegativeSCEV(Step);
1302 // Insert an add instruction right before the terminator corresponding
1303 // to the back-edge or just before the only use. The location is determined
1304 // by the caller and passed in as IVIncInsertPt.
1305 Value *StepV = Rewriter.expandCodeFor(IncAmount, Ty,
1306 Preheader->getTerminator());
1309 IncV = BinaryOperator::CreateSub(PN, StepV, "lsr.iv.next",
1312 IncV = BinaryOperator::CreateAdd(PN, StepV, "lsr.iv.next",
1315 if (!isa<ConstantInt>(StepV)) ++NumVariable;
1317 PN->addIncoming(IncV, LatchBlock);
1323 static void SortUsersToProcess(std::vector<BasedUser> &UsersToProcess) {
1324 // We want to emit code for users inside the loop first. To do this, we
1325 // rearrange BasedUser so that the entries at the end have
1326 // isUseOfPostIncrementedValue = false, because we pop off the end of the
1327 // vector (so we handle them first).
1328 std::partition(UsersToProcess.begin(), UsersToProcess.end(),
1329 PartitionByIsUseOfPostIncrementedValue);
1331 // Sort this by base, so that things with the same base are handled
1332 // together. By partitioning first and stable-sorting later, we are
1333 // guaranteed that within each base we will pop off users from within the
1334 // loop before users outside of the loop with a particular base.
1336 // We would like to use stable_sort here, but we can't. The problem is that
1337 // const SCEV *'s don't have a deterministic ordering w.r.t to each other, so
1338 // we don't have anything to do a '<' comparison on. Because we think the
1339 // number of uses is small, do a horrible bubble sort which just relies on
1341 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1342 // Get a base value.
1343 const SCEV *Base = UsersToProcess[i].Base;
1345 // Compact everything with this base to be consecutive with this one.
1346 for (unsigned j = i+1; j != e; ++j) {
1347 if (UsersToProcess[j].Base == Base) {
1348 std::swap(UsersToProcess[i+1], UsersToProcess[j]);
1355 /// PrepareToStrengthReduceFully - Prepare to fully strength-reduce
1356 /// UsersToProcess, meaning lowering addresses all the way down to direct
1357 /// pointer arithmetic.
1360 LoopStrengthReduce::PrepareToStrengthReduceFully(
1361 std::vector<BasedUser> &UsersToProcess,
1363 const SCEV *CommonExprs,
1365 SCEVExpander &PreheaderRewriter) {
1366 DEBUG(errs() << " Fully reducing all users\n");
1368 // Rewrite the UsersToProcess records, creating a separate PHI for each
1369 // unique Base value.
1370 Instruction *IVIncInsertPt = L->getLoopLatch()->getTerminator();
1371 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ) {
1372 // TODO: The uses are grouped by base, but not sorted. We arbitrarily
1373 // pick the first Imm value here to start with, and adjust it for the
1375 const SCEV *Imm = UsersToProcess[i].Imm;
1376 const SCEV *Base = UsersToProcess[i].Base;
1377 const SCEV *Start = SE->getAddExpr(CommonExprs, Base, Imm);
1378 PHINode *Phi = InsertAffinePhi(Start, Stride, IVIncInsertPt, L,
1380 // Loop over all the users with the same base.
1382 UsersToProcess[i].Base = SE->getIntegerSCEV(0, Stride->getType());
1383 UsersToProcess[i].Imm = SE->getMinusSCEV(UsersToProcess[i].Imm, Imm);
1384 UsersToProcess[i].Phi = Phi;
1385 assert(UsersToProcess[i].Imm->isLoopInvariant(L) &&
1386 "ShouldUseFullStrengthReductionMode should reject this!");
1387 } while (++i != e && Base == UsersToProcess[i].Base);
1391 /// FindIVIncInsertPt - Return the location to insert the increment instruction.
1392 /// If the only use if a use of postinc value, (must be the loop termination
1393 /// condition), then insert it just before the use.
1394 static Instruction *FindIVIncInsertPt(std::vector<BasedUser> &UsersToProcess,
1396 if (UsersToProcess.size() == 1 &&
1397 UsersToProcess[0].isUseOfPostIncrementedValue &&
1398 L->contains(UsersToProcess[0].Inst->getParent()))
1399 return UsersToProcess[0].Inst;
1400 return L->getLoopLatch()->getTerminator();
1403 /// PrepareToStrengthReduceWithNewPhi - Insert a new induction variable for the
1404 /// given users to share.
1407 LoopStrengthReduce::PrepareToStrengthReduceWithNewPhi(
1408 std::vector<BasedUser> &UsersToProcess,
1410 const SCEV *CommonExprs,
1412 Instruction *IVIncInsertPt,
1414 SCEVExpander &PreheaderRewriter) {
1415 DEBUG(errs() << " Inserting new PHI:\n");
1417 PHINode *Phi = InsertAffinePhi(SE->getUnknown(CommonBaseV),
1418 Stride, IVIncInsertPt, L,
1421 // Remember this in case a later stride is multiple of this.
1422 IVsByStride[Stride].addIV(Stride, CommonExprs, Phi);
1424 // All the users will share this new IV.
1425 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1426 UsersToProcess[i].Phi = Phi;
1428 DEBUG(errs() << " IV=");
1429 DEBUG(WriteAsOperand(errs(), Phi, /*PrintType=*/false));
1430 DEBUG(errs() << "\n");
1433 /// PrepareToStrengthReduceFromSmallerStride - Prepare for the given users to
1434 /// reuse an induction variable with a stride that is a factor of the current
1435 /// induction variable.
1438 LoopStrengthReduce::PrepareToStrengthReduceFromSmallerStride(
1439 std::vector<BasedUser> &UsersToProcess,
1441 const IVExpr &ReuseIV,
1442 Instruction *PreInsertPt) {
1443 DEBUG(errs() << " Rewriting in terms of existing IV of STRIDE "
1444 << *ReuseIV.Stride << " and BASE " << *ReuseIV.Base << "\n");
1446 // All the users will share the reused IV.
1447 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1448 UsersToProcess[i].Phi = ReuseIV.PHI;
1450 Constant *C = dyn_cast<Constant>(CommonBaseV);
1452 (!C->isNullValue() &&
1453 !fitsInAddressMode(SE->getUnknown(CommonBaseV), CommonBaseV->getType(),
1455 // We want the common base emitted into the preheader! This is just
1456 // using cast as a copy so BitCast (no-op cast) is appropriate
1457 CommonBaseV = new BitCastInst(CommonBaseV, CommonBaseV->getType(),
1458 "commonbase", PreInsertPt);
1461 static bool IsImmFoldedIntoAddrMode(GlobalValue *GV, int64_t Offset,
1462 const Type *AccessTy,
1463 std::vector<BasedUser> &UsersToProcess,
1464 const TargetLowering *TLI) {
1465 SmallVector<Instruction*, 16> AddrModeInsts;
1466 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1467 if (UsersToProcess[i].isUseOfPostIncrementedValue)
1469 ExtAddrMode AddrMode =
1470 AddressingModeMatcher::Match(UsersToProcess[i].OperandValToReplace,
1471 AccessTy, UsersToProcess[i].Inst,
1472 AddrModeInsts, *TLI);
1473 if (GV && GV != AddrMode.BaseGV)
1475 if (Offset && !AddrMode.BaseOffs)
1476 // FIXME: How to accurate check it's immediate offset is folded.
1478 AddrModeInsts.clear();
1483 /// StrengthReduceIVUsersOfStride - Strength reduce all of the users of a single
1484 /// stride of IV. All of the users may have different starting values, and this
1485 /// may not be the only stride.
1487 LoopStrengthReduce::StrengthReduceIVUsersOfStride(const SCEV *const &Stride,
1488 IVUsersOfOneStride &Uses,
1490 // If all the users are moved to another stride, then there is nothing to do.
1491 if (Uses.Users.empty())
1494 // Keep track if every use in UsersToProcess is an address. If they all are,
1495 // we may be able to rewrite the entire collection of them in terms of a
1496 // smaller-stride IV.
1497 bool AllUsesAreAddresses = true;
1499 // Keep track if every use of a single stride is outside the loop. If so,
1500 // we want to be more aggressive about reusing a smaller-stride IV; a
1501 // multiply outside the loop is better than another IV inside. Well, usually.
1502 bool AllUsesAreOutsideLoop = true;
1504 // Transform our list of users and offsets to a bit more complex table. In
1505 // this new vector, each 'BasedUser' contains 'Base' the base of the
1506 // strided accessas well as the old information from Uses. We progressively
1507 // move information from the Base field to the Imm field, until we eventually
1508 // have the full access expression to rewrite the use.
1509 std::vector<BasedUser> UsersToProcess;
1510 const SCEV *CommonExprs = CollectIVUsers(Stride, Uses, L, AllUsesAreAddresses,
1511 AllUsesAreOutsideLoop,
1514 // Sort the UsersToProcess array so that users with common bases are
1515 // next to each other.
1516 SortUsersToProcess(UsersToProcess);
1518 // If we managed to find some expressions in common, we'll need to carry
1519 // their value in a register and add it in for each use. This will take up
1520 // a register operand, which potentially restricts what stride values are
1522 bool HaveCommonExprs = !CommonExprs->isZero();
1523 const Type *ReplacedTy = CommonExprs->getType();
1525 // If all uses are addresses, consider sinking the immediate part of the
1526 // common expression back into uses if they can fit in the immediate fields.
1527 if (TLI && HaveCommonExprs && AllUsesAreAddresses) {
1528 const SCEV *NewCommon = CommonExprs;
1529 const SCEV *Imm = SE->getIntegerSCEV(0, ReplacedTy);
1530 MoveImmediateValues(TLI, Type::getVoidTy(
1531 L->getLoopPreheader()->getContext()),
1532 NewCommon, Imm, true, L, SE);
1533 if (!Imm->isZero()) {
1536 // If the immediate part of the common expression is a GV, check if it's
1537 // possible to fold it into the target addressing mode.
1538 GlobalValue *GV = 0;
1539 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(Imm))
1540 GV = dyn_cast<GlobalValue>(SU->getValue());
1542 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Imm))
1543 Offset = SC->getValue()->getSExtValue();
1545 // Pass VoidTy as the AccessTy to be conservative, because
1546 // there could be multiple access types among all the uses.
1547 DoSink = IsImmFoldedIntoAddrMode(GV, Offset,
1548 Type::getVoidTy(L->getLoopPreheader()->getContext()),
1549 UsersToProcess, TLI);
1552 DEBUG(errs() << " Sinking " << *Imm << " back down into uses\n");
1553 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1554 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm, Imm);
1555 CommonExprs = NewCommon;
1556 HaveCommonExprs = !CommonExprs->isZero();
1562 // Now that we know what we need to do, insert the PHI node itself.
1564 DEBUG(errs() << "LSR: Examining IVs of TYPE " << *ReplacedTy << " of STRIDE "
1566 << " Common base: " << *CommonExprs << "\n");
1568 SCEVExpander Rewriter(*SE);
1569 SCEVExpander PreheaderRewriter(*SE);
1571 BasicBlock *Preheader = L->getLoopPreheader();
1572 Instruction *PreInsertPt = Preheader->getTerminator();
1573 BasicBlock *LatchBlock = L->getLoopLatch();
1574 Instruction *IVIncInsertPt = LatchBlock->getTerminator();
1576 Value *CommonBaseV = Constant::getNullValue(ReplacedTy);
1578 const SCEV *RewriteFactor = SE->getIntegerSCEV(0, ReplacedTy);
1579 IVExpr ReuseIV(SE->getIntegerSCEV(0,
1580 Type::getInt32Ty(Preheader->getContext())),
1581 SE->getIntegerSCEV(0,
1582 Type::getInt32Ty(Preheader->getContext())),
1585 // Choose a strength-reduction strategy and prepare for it by creating
1586 // the necessary PHIs and adjusting the bookkeeping.
1587 if (ShouldUseFullStrengthReductionMode(UsersToProcess, L,
1588 AllUsesAreAddresses, Stride)) {
1589 PrepareToStrengthReduceFully(UsersToProcess, Stride, CommonExprs, L,
1592 // Emit the initial base value into the loop preheader.
1593 CommonBaseV = PreheaderRewriter.expandCodeFor(CommonExprs, ReplacedTy,
1596 // If all uses are addresses, check if it is possible to reuse an IV. The
1597 // new IV must have a stride that is a multiple of the old stride; the
1598 // multiple must be a number that can be encoded in the scale field of the
1599 // target addressing mode; and we must have a valid instruction after this
1600 // substitution, including the immediate field, if any.
1601 RewriteFactor = CheckForIVReuse(HaveCommonExprs, AllUsesAreAddresses,
1602 AllUsesAreOutsideLoop,
1603 Stride, ReuseIV, ReplacedTy,
1605 if (!RewriteFactor->isZero())
1606 PrepareToStrengthReduceFromSmallerStride(UsersToProcess, CommonBaseV,
1607 ReuseIV, PreInsertPt);
1609 IVIncInsertPt = FindIVIncInsertPt(UsersToProcess, L);
1610 PrepareToStrengthReduceWithNewPhi(UsersToProcess, Stride, CommonExprs,
1611 CommonBaseV, IVIncInsertPt,
1612 L, PreheaderRewriter);
1616 // Process all the users now, replacing their strided uses with
1617 // strength-reduced forms. This outer loop handles all bases, the inner
1618 // loop handles all users of a particular base.
1619 while (!UsersToProcess.empty()) {
1620 const SCEV *Base = UsersToProcess.back().Base;
1621 Instruction *Inst = UsersToProcess.back().Inst;
1623 // Emit the code for Base into the preheader.
1625 if (!Base->isZero()) {
1626 BaseV = PreheaderRewriter.expandCodeFor(Base, 0, PreInsertPt);
1628 DEBUG(errs() << " INSERTING code for BASE = " << *Base << ":");
1629 if (BaseV->hasName())
1630 DEBUG(errs() << " Result value name = %" << BaseV->getName());
1631 DEBUG(errs() << "\n");
1633 // If BaseV is a non-zero constant, make sure that it gets inserted into
1634 // the preheader, instead of being forward substituted into the uses. We
1635 // do this by forcing a BitCast (noop cast) to be inserted into the
1636 // preheader in this case.
1637 if (!fitsInAddressMode(Base, getAccessType(Inst), TLI, false) &&
1638 isa<Constant>(BaseV)) {
1639 // We want this constant emitted into the preheader! This is just
1640 // using cast as a copy so BitCast (no-op cast) is appropriate
1641 BaseV = new BitCastInst(BaseV, BaseV->getType(), "preheaderinsert",
1646 // Emit the code to add the immediate offset to the Phi value, just before
1647 // the instructions that we identified as using this stride and base.
1649 // FIXME: Use emitted users to emit other users.
1650 BasedUser &User = UsersToProcess.back();
1652 DEBUG(errs() << " Examining ");
1653 if (User.isUseOfPostIncrementedValue)
1654 DEBUG(errs() << "postinc");
1656 DEBUG(errs() << "preinc");
1657 DEBUG(errs() << " use ");
1658 DEBUG(WriteAsOperand(errs(), UsersToProcess.back().OperandValToReplace,
1659 /*PrintType=*/false));
1660 DEBUG(errs() << " in Inst: " << *User.Inst);
1662 // If this instruction wants to use the post-incremented value, move it
1663 // after the post-inc and use its value instead of the PHI.
1664 Value *RewriteOp = User.Phi;
1665 if (User.isUseOfPostIncrementedValue) {
1666 RewriteOp = User.Phi->getIncomingValueForBlock(LatchBlock);
1667 // If this user is in the loop, make sure it is the last thing in the
1668 // loop to ensure it is dominated by the increment. In case it's the
1669 // only use of the iv, the increment instruction is already before the
1671 if (L->contains(User.Inst->getParent()) && User.Inst != IVIncInsertPt)
1672 User.Inst->moveBefore(IVIncInsertPt);
1675 const SCEV *RewriteExpr = SE->getUnknown(RewriteOp);
1677 if (SE->getEffectiveSCEVType(RewriteOp->getType()) !=
1678 SE->getEffectiveSCEVType(ReplacedTy)) {
1679 assert(SE->getTypeSizeInBits(RewriteOp->getType()) >
1680 SE->getTypeSizeInBits(ReplacedTy) &&
1681 "Unexpected widening cast!");
1682 RewriteExpr = SE->getTruncateExpr(RewriteExpr, ReplacedTy);
1685 // If we had to insert new instructions for RewriteOp, we have to
1686 // consider that they may not have been able to end up immediately
1687 // next to RewriteOp, because non-PHI instructions may never precede
1688 // PHI instructions in a block. In this case, remember where the last
1689 // instruction was inserted so that if we're replacing a different
1690 // PHI node, we can use the later point to expand the final
1692 Instruction *NewBasePt = dyn_cast<Instruction>(RewriteOp);
1693 if (RewriteOp == User.Phi) NewBasePt = 0;
1695 // Clear the SCEVExpander's expression map so that we are guaranteed
1696 // to have the code emitted where we expect it.
1699 // If we are reusing the iv, then it must be multiplied by a constant
1700 // factor to take advantage of the addressing mode scale component.
1701 if (!RewriteFactor->isZero()) {
1702 // If we're reusing an IV with a nonzero base (currently this happens
1703 // only when all reuses are outside the loop) subtract that base here.
1704 // The base has been used to initialize the PHI node but we don't want
1706 if (!ReuseIV.Base->isZero()) {
1707 const SCEV *typedBase = ReuseIV.Base;
1708 if (SE->getEffectiveSCEVType(RewriteExpr->getType()) !=
1709 SE->getEffectiveSCEVType(ReuseIV.Base->getType())) {
1710 // It's possible the original IV is a larger type than the new IV,
1711 // in which case we have to truncate the Base. We checked in
1712 // RequiresTypeConversion that this is valid.
1713 assert(SE->getTypeSizeInBits(RewriteExpr->getType()) <
1714 SE->getTypeSizeInBits(ReuseIV.Base->getType()) &&
1715 "Unexpected lengthening conversion!");
1716 typedBase = SE->getTruncateExpr(ReuseIV.Base,
1717 RewriteExpr->getType());
1719 RewriteExpr = SE->getMinusSCEV(RewriteExpr, typedBase);
1722 // Multiply old variable, with base removed, by new scale factor.
1723 RewriteExpr = SE->getMulExpr(RewriteFactor,
1726 // The common base is emitted in the loop preheader. But since we
1727 // are reusing an IV, it has not been used to initialize the PHI node.
1728 // Add it to the expression used to rewrite the uses.
1729 // When this use is outside the loop, we earlier subtracted the
1730 // common base, and are adding it back here. Use the same expression
1731 // as before, rather than CommonBaseV, so DAGCombiner will zap it.
1732 if (!CommonExprs->isZero()) {
1733 if (L->contains(User.Inst->getParent()))
1734 RewriteExpr = SE->getAddExpr(RewriteExpr,
1735 SE->getUnknown(CommonBaseV));
1737 RewriteExpr = SE->getAddExpr(RewriteExpr, CommonExprs);
1741 // Now that we know what we need to do, insert code before User for the
1742 // immediate and any loop-variant expressions.
1744 // Add BaseV to the PHI value if needed.
1745 RewriteExpr = SE->getAddExpr(RewriteExpr, SE->getUnknown(BaseV));
1747 User.RewriteInstructionToUseNewBase(RewriteExpr, NewBasePt,
1751 // Mark old value we replaced as possibly dead, so that it is eliminated
1752 // if we just replaced the last use of that value.
1753 DeadInsts.push_back(User.OperandValToReplace);
1755 UsersToProcess.pop_back();
1758 // If there are any more users to process with the same base, process them
1759 // now. We sorted by base above, so we just have to check the last elt.
1760 } while (!UsersToProcess.empty() && UsersToProcess.back().Base == Base);
1761 // TODO: Next, find out which base index is the most common, pull it out.
1764 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
1765 // different starting values, into different PHIs.
1768 void LoopStrengthReduce::StrengthReduceIVUsers(Loop *L) {
1769 // Note: this processes each stride/type pair individually. All users
1770 // passed into StrengthReduceIVUsersOfStride have the same type AND stride.
1771 // Also, note that we iterate over IVUsesByStride indirectly by using
1772 // StrideOrder. This extra layer of indirection makes the ordering of
1773 // strides deterministic - not dependent on map order.
1774 for (unsigned Stride = 0, e = IU->StrideOrder.size(); Stride != e; ++Stride) {
1775 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
1776 IU->IVUsesByStride.find(IU->StrideOrder[Stride]);
1777 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
1778 // FIXME: Generalize to non-affine IV's.
1779 if (!SI->first->isLoopInvariant(L))
1781 StrengthReduceIVUsersOfStride(SI->first, *SI->second, L);
1785 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1786 /// set the IV user and stride information and return true, otherwise return
1788 bool LoopStrengthReduce::FindIVUserForCond(ICmpInst *Cond,
1789 IVStrideUse *&CondUse,
1790 const SCEV* &CondStride) {
1791 for (unsigned Stride = 0, e = IU->StrideOrder.size();
1792 Stride != e && !CondUse; ++Stride) {
1793 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
1794 IU->IVUsesByStride.find(IU->StrideOrder[Stride]);
1795 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
1797 for (ilist<IVStrideUse>::iterator UI = SI->second->Users.begin(),
1798 E = SI->second->Users.end(); UI != E; ++UI)
1799 if (UI->getUser() == Cond) {
1800 // NOTE: we could handle setcc instructions with multiple uses here, but
1801 // InstCombine does it as well for simple uses, it's not clear that it
1802 // occurs enough in real life to handle.
1804 CondStride = SI->first;
1812 // Constant strides come first which in turns are sorted by their absolute
1813 // values. If absolute values are the same, then positive strides comes first.
1815 // 4, -1, X, 1, 2 ==> 1, -1, 2, 4, X
1816 struct StrideCompare {
1817 const ScalarEvolution *SE;
1818 explicit StrideCompare(const ScalarEvolution *se) : SE(se) {}
1820 bool operator()(const SCEV *const &LHS, const SCEV *const &RHS) {
1821 const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS);
1822 const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS);
1824 int64_t LV = LHSC->getValue()->getSExtValue();
1825 int64_t RV = RHSC->getValue()->getSExtValue();
1826 uint64_t ALV = (LV < 0) ? -LV : LV;
1827 uint64_t ARV = (RV < 0) ? -RV : RV;
1835 // If it's the same value but different type, sort by bit width so
1836 // that we emit larger induction variables before smaller
1837 // ones, letting the smaller be re-written in terms of larger ones.
1838 return SE->getTypeSizeInBits(RHS->getType()) <
1839 SE->getTypeSizeInBits(LHS->getType());
1841 return LHSC && !RHSC;
1846 /// ChangeCompareStride - If a loop termination compare instruction is the
1847 /// only use of its stride, and the compaison is against a constant value,
1848 /// try eliminate the stride by moving the compare instruction to another
1849 /// stride and change its constant operand accordingly. e.g.
1855 /// if (v2 < 10) goto loop
1860 /// if (v1 < 30) goto loop
1861 ICmpInst *LoopStrengthReduce::ChangeCompareStride(Loop *L, ICmpInst *Cond,
1862 IVStrideUse* &CondUse,
1863 const SCEV* &CondStride,
1865 // If there's only one stride in the loop, there's nothing to do here.
1866 if (IU->StrideOrder.size() < 2)
1868 // If there are other users of the condition's stride, don't bother
1869 // trying to change the condition because the stride will still
1871 std::map<const SCEV *, IVUsersOfOneStride *>::iterator I =
1872 IU->IVUsesByStride.find(CondStride);
1873 if (I == IU->IVUsesByStride.end())
1875 if (I->second->Users.size() > 1) {
1876 for (ilist<IVStrideUse>::iterator II = I->second->Users.begin(),
1877 EE = I->second->Users.end(); II != EE; ++II) {
1878 if (II->getUser() == Cond)
1880 if (!isInstructionTriviallyDead(II->getUser()))
1884 // Only handle constant strides for now.
1885 const SCEVConstant *SC = dyn_cast<SCEVConstant>(CondStride);
1886 if (!SC) return Cond;
1888 ICmpInst::Predicate Predicate = Cond->getPredicate();
1889 int64_t CmpSSInt = SC->getValue()->getSExtValue();
1890 unsigned BitWidth = SE->getTypeSizeInBits(CondStride->getType());
1891 uint64_t SignBit = 1ULL << (BitWidth-1);
1892 const Type *CmpTy = Cond->getOperand(0)->getType();
1893 const Type *NewCmpTy = NULL;
1894 unsigned TyBits = SE->getTypeSizeInBits(CmpTy);
1895 unsigned NewTyBits = 0;
1896 const SCEV *NewStride = NULL;
1897 Value *NewCmpLHS = NULL;
1898 Value *NewCmpRHS = NULL;
1900 const SCEV *NewOffset = SE->getIntegerSCEV(0, CmpTy);
1902 if (ConstantInt *C = dyn_cast<ConstantInt>(Cond->getOperand(1))) {
1903 int64_t CmpVal = C->getValue().getSExtValue();
1905 // Check the relevant induction variable for conformance to
1907 const SCEV *IV = SE->getSCEV(Cond->getOperand(0));
1908 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1909 if (!AR || !AR->isAffine())
1912 const SCEVConstant *StartC = dyn_cast<SCEVConstant>(AR->getStart());
1913 // Check stride constant and the comparision constant signs to detect
1916 if ((StartC->getValue()->getSExtValue() < CmpVal && CmpSSInt < 0) ||
1917 (StartC->getValue()->getSExtValue() > CmpVal && CmpSSInt > 0))
1920 // More restrictive check for the other cases.
1921 if ((CmpVal & SignBit) != (CmpSSInt & SignBit))
1925 // Look for a suitable stride / iv as replacement.
1926 for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
1927 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
1928 IU->IVUsesByStride.find(IU->StrideOrder[i]);
1929 if (!isa<SCEVConstant>(SI->first) || SI->second->Users.empty())
1931 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1932 if (SSInt == CmpSSInt ||
1933 abs64(SSInt) < abs64(CmpSSInt) ||
1934 (SSInt % CmpSSInt) != 0)
1937 Scale = SSInt / CmpSSInt;
1938 int64_t NewCmpVal = CmpVal * Scale;
1940 // If old icmp value fits in icmp immediate field, but the new one doesn't
1941 // try something else.
1943 TLI->isLegalICmpImmediate(CmpVal) &&
1944 !TLI->isLegalICmpImmediate(NewCmpVal))
1947 APInt Mul = APInt(BitWidth*2, CmpVal, true);
1948 Mul = Mul * APInt(BitWidth*2, Scale, true);
1949 // Check for overflow.
1950 if (!Mul.isSignedIntN(BitWidth))
1952 // Check for overflow in the stride's type too.
1953 if (!Mul.isSignedIntN(SE->getTypeSizeInBits(SI->first->getType())))
1956 // Watch out for overflow.
1957 if (ICmpInst::isSigned(Predicate) &&
1958 (CmpVal & SignBit) != (NewCmpVal & SignBit))
1961 // Pick the best iv to use trying to avoid a cast.
1963 for (ilist<IVStrideUse>::iterator UI = SI->second->Users.begin(),
1964 E = SI->second->Users.end(); UI != E; ++UI) {
1965 Value *Op = UI->getOperandValToReplace();
1967 // If the IVStrideUse implies a cast, check for an actual cast which
1968 // can be used to find the original IV expression.
1969 if (SE->getEffectiveSCEVType(Op->getType()) !=
1970 SE->getEffectiveSCEVType(SI->first->getType())) {
1971 CastInst *CI = dyn_cast<CastInst>(Op);
1972 // If it's not a simple cast, it's complicated.
1975 // If it's a cast from a type other than the stride type,
1976 // it's complicated.
1977 if (CI->getOperand(0)->getType() != SI->first->getType())
1979 // Ok, we found the IV expression in the stride's type.
1980 Op = CI->getOperand(0);
1984 if (NewCmpLHS->getType() == CmpTy)
1990 NewCmpTy = NewCmpLHS->getType();
1991 NewTyBits = SE->getTypeSizeInBits(NewCmpTy);
1992 const Type *NewCmpIntTy = IntegerType::get(Cond->getContext(), NewTyBits);
1993 if (RequiresTypeConversion(NewCmpTy, CmpTy)) {
1994 // Check if it is possible to rewrite it using
1995 // an iv / stride of a smaller integer type.
1996 unsigned Bits = NewTyBits;
1997 if (ICmpInst::isSigned(Predicate))
1999 uint64_t Mask = (1ULL << Bits) - 1;
2000 if (((uint64_t)NewCmpVal & Mask) != (uint64_t)NewCmpVal)
2004 // Don't rewrite if use offset is non-constant and the new type is
2005 // of a different type.
2006 // FIXME: too conservative?
2007 if (NewTyBits != TyBits && !isa<SCEVConstant>(CondUse->getOffset()))
2011 bool AllUsesAreAddresses = true;
2012 bool AllUsesAreOutsideLoop = true;
2013 std::vector<BasedUser> UsersToProcess;
2014 const SCEV *CommonExprs = CollectIVUsers(SI->first, *SI->second, L,
2015 AllUsesAreAddresses,
2016 AllUsesAreOutsideLoop,
2018 // Avoid rewriting the compare instruction with an iv of new stride
2019 // if it's likely the new stride uses will be rewritten using the
2020 // stride of the compare instruction.
2021 if (AllUsesAreAddresses &&
2022 ValidScale(!CommonExprs->isZero(), Scale, UsersToProcess))
2026 // Avoid rewriting the compare instruction with an iv which has
2027 // implicit extension or truncation built into it.
2028 // TODO: This is over-conservative.
2029 if (SE->getTypeSizeInBits(CondUse->getOffset()->getType()) != TyBits)
2032 // If scale is negative, use swapped predicate unless it's testing
2034 if (Scale < 0 && !Cond->isEquality())
2035 Predicate = ICmpInst::getSwappedPredicate(Predicate);
2037 NewStride = IU->StrideOrder[i];
2038 if (!isa<PointerType>(NewCmpTy))
2039 NewCmpRHS = ConstantInt::get(NewCmpTy, NewCmpVal);
2041 Constant *CI = ConstantInt::get(NewCmpIntTy, NewCmpVal);
2042 NewCmpRHS = ConstantExpr::getIntToPtr(CI, NewCmpTy);
2044 NewOffset = TyBits == NewTyBits
2045 ? SE->getMulExpr(CondUse->getOffset(),
2046 SE->getConstant(CmpTy, Scale))
2047 : SE->getConstant(NewCmpIntTy,
2048 cast<SCEVConstant>(CondUse->getOffset())->getValue()
2049 ->getSExtValue()*Scale);
2054 // Forgo this transformation if it the increment happens to be
2055 // unfortunately positioned after the condition, and the condition
2056 // has multiple uses which prevent it from being moved immediately
2057 // before the branch. See
2058 // test/Transforms/LoopStrengthReduce/change-compare-stride-trickiness-*.ll
2059 // for an example of this situation.
2060 if (!Cond->hasOneUse()) {
2061 for (BasicBlock::iterator I = Cond, E = Cond->getParent()->end();
2068 // Create a new compare instruction using new stride / iv.
2069 ICmpInst *OldCond = Cond;
2070 // Insert new compare instruction.
2071 Cond = new ICmpInst(OldCond, Predicate, NewCmpLHS, NewCmpRHS,
2072 L->getHeader()->getName() + ".termcond");
2074 DEBUG(errs() << " Change compare stride in Inst " << *OldCond);
2075 DEBUG(errs() << " to " << *Cond << '\n');
2077 // Remove the old compare instruction. The old indvar is probably dead too.
2078 DeadInsts.push_back(CondUse->getOperandValToReplace());
2079 OldCond->replaceAllUsesWith(Cond);
2080 OldCond->eraseFromParent();
2082 IU->IVUsesByStride[NewStride]->addUser(NewOffset, Cond, NewCmpLHS);
2083 CondUse = &IU->IVUsesByStride[NewStride]->Users.back();
2084 CondStride = NewStride;
2092 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
2093 /// a max computation.
2095 /// This is a narrow solution to a specific, but acute, problem. For loops
2101 /// } while (++i < n);
2103 /// the trip count isn't just 'n', because 'n' might not be positive. And
2104 /// unfortunately this can come up even for loops where the user didn't use
2105 /// a C do-while loop. For example, seemingly well-behaved top-test loops
2106 /// will commonly be lowered like this:
2112 /// } while (++i < n);
2115 /// and then it's possible for subsequent optimization to obscure the if
2116 /// test in such a way that indvars can't find it.
2118 /// When indvars can't find the if test in loops like this, it creates a
2119 /// max expression, which allows it to give the loop a canonical
2120 /// induction variable:
2123 /// max = n < 1 ? 1 : n;
2126 /// } while (++i != max);
2128 /// Canonical induction variables are necessary because the loop passes
2129 /// are designed around them. The most obvious example of this is the
2130 /// LoopInfo analysis, which doesn't remember trip count values. It
2131 /// expects to be able to rediscover the trip count each time it is
2132 /// needed, and it does this using a simple analyis that only succeeds if
2133 /// the loop has a canonical induction variable.
2135 /// However, when it comes time to generate code, the maximum operation
2136 /// can be quite costly, especially if it's inside of an outer loop.
2138 /// This function solves this problem by detecting this type of loop and
2139 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
2140 /// the instructions for the maximum computation.
2142 ICmpInst *LoopStrengthReduce::OptimizeMax(Loop *L, ICmpInst *Cond,
2143 IVStrideUse* &CondUse) {
2144 // Check that the loop matches the pattern we're looking for.
2145 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
2146 Cond->getPredicate() != CmpInst::ICMP_NE)
2149 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
2150 if (!Sel || !Sel->hasOneUse()) return Cond;
2152 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
2153 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2155 const SCEV *One = SE->getIntegerSCEV(1, BackedgeTakenCount->getType());
2157 // Add one to the backedge-taken count to get the trip count.
2158 const SCEV *IterationCount = SE->getAddExpr(BackedgeTakenCount, One);
2160 // Check for a max calculation that matches the pattern.
2161 if (!isa<SCEVSMaxExpr>(IterationCount) && !isa<SCEVUMaxExpr>(IterationCount))
2163 const SCEVNAryExpr *Max = cast<SCEVNAryExpr>(IterationCount);
2164 if (Max != SE->getSCEV(Sel)) return Cond;
2166 // To handle a max with more than two operands, this optimization would
2167 // require additional checking and setup.
2168 if (Max->getNumOperands() != 2)
2171 const SCEV *MaxLHS = Max->getOperand(0);
2172 const SCEV *MaxRHS = Max->getOperand(1);
2173 if (!MaxLHS || MaxLHS != One) return Cond;
2175 // Check the relevant induction variable for conformance to
2177 const SCEV *IV = SE->getSCEV(Cond->getOperand(0));
2178 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2179 if (!AR || !AR->isAffine() ||
2180 AR->getStart() != One ||
2181 AR->getStepRecurrence(*SE) != One)
2184 assert(AR->getLoop() == L &&
2185 "Loop condition operand is an addrec in a different loop!");
2187 // Check the right operand of the select, and remember it, as it will
2188 // be used in the new comparison instruction.
2190 if (SE->getSCEV(Sel->getOperand(1)) == MaxRHS)
2191 NewRHS = Sel->getOperand(1);
2192 else if (SE->getSCEV(Sel->getOperand(2)) == MaxRHS)
2193 NewRHS = Sel->getOperand(2);
2194 if (!NewRHS) return Cond;
2196 // Determine the new comparison opcode. It may be signed or unsigned,
2197 // and the original comparison may be either equality or inequality.
2198 CmpInst::Predicate Pred =
2199 isa<SCEVSMaxExpr>(Max) ? CmpInst::ICMP_SLT : CmpInst::ICMP_ULT;
2200 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
2201 Pred = CmpInst::getInversePredicate(Pred);
2203 // Ok, everything looks ok to change the condition into an SLT or SGE and
2204 // delete the max calculation.
2206 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
2208 // Delete the max calculation instructions.
2209 Cond->replaceAllUsesWith(NewCond);
2210 CondUse->setUser(NewCond);
2211 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2212 Cond->eraseFromParent();
2213 Sel->eraseFromParent();
2214 if (Cmp->use_empty())
2215 Cmp->eraseFromParent();
2219 /// OptimizeShadowIV - If IV is used in a int-to-float cast
2220 /// inside the loop then try to eliminate the cast opeation.
2221 void LoopStrengthReduce::OptimizeShadowIV(Loop *L) {
2223 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
2224 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2227 for (unsigned Stride = 0, e = IU->StrideOrder.size(); Stride != e;
2229 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
2230 IU->IVUsesByStride.find(IU->StrideOrder[Stride]);
2231 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
2232 if (!isa<SCEVConstant>(SI->first))
2235 for (ilist<IVStrideUse>::iterator UI = SI->second->Users.begin(),
2236 E = SI->second->Users.end(); UI != E; /* empty */) {
2237 ilist<IVStrideUse>::iterator CandidateUI = UI;
2239 Instruction *ShadowUse = CandidateUI->getUser();
2240 const Type *DestTy = NULL;
2242 /* If shadow use is a int->float cast then insert a second IV
2243 to eliminate this cast.
2245 for (unsigned i = 0; i < n; ++i)
2251 for (unsigned i = 0; i < n; ++i, ++d)
2254 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser()))
2255 DestTy = UCast->getDestTy();
2256 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser()))
2257 DestTy = SCast->getDestTy();
2258 if (!DestTy) continue;
2261 // If target does not support DestTy natively then do not apply
2262 // this transformation.
2263 EVT DVT = TLI->getValueType(DestTy);
2264 if (!TLI->isTypeLegal(DVT)) continue;
2267 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
2269 if (PH->getNumIncomingValues() != 2) continue;
2271 const Type *SrcTy = PH->getType();
2272 int Mantissa = DestTy->getFPMantissaWidth();
2273 if (Mantissa == -1) continue;
2274 if ((int)SE->getTypeSizeInBits(SrcTy) > Mantissa)
2277 unsigned Entry, Latch;
2278 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
2286 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
2287 if (!Init) continue;
2288 Constant *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());
2290 BinaryOperator *Incr =
2291 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
2292 if (!Incr) continue;
2293 if (Incr->getOpcode() != Instruction::Add
2294 && Incr->getOpcode() != Instruction::Sub)
2297 /* Initialize new IV, double d = 0.0 in above example. */
2298 ConstantInt *C = NULL;
2299 if (Incr->getOperand(0) == PH)
2300 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
2301 else if (Incr->getOperand(1) == PH)
2302 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
2308 // Ignore negative constants, as the code below doesn't handle them
2309 // correctly. TODO: Remove this restriction.
2310 if (!C->getValue().isStrictlyPositive()) continue;
2312 /* Add new PHINode. */
2313 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);
2315 /* create new increment. '++d' in above example. */
2316 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
2317 BinaryOperator *NewIncr =
2318 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
2319 Instruction::FAdd : Instruction::FSub,
2320 NewPH, CFP, "IV.S.next.", Incr);
2322 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
2323 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
2325 /* Remove cast operation */
2326 ShadowUse->replaceAllUsesWith(NewPH);
2327 ShadowUse->eraseFromParent();
2334 /// OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
2335 /// uses in the loop, look to see if we can eliminate some, in favor of using
2336 /// common indvars for the different uses.
2337 void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
2338 // TODO: implement optzns here.
2340 OptimizeShadowIV(L);
2343 bool LoopStrengthReduce::StrideMightBeShared(const SCEV* Stride, Loop *L,
2345 int64_t SInt = cast<SCEVConstant>(Stride)->getValue()->getSExtValue();
2346 for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
2347 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
2348 IU->IVUsesByStride.find(IU->StrideOrder[i]);
2349 const SCEV *Share = SI->first;
2350 if (!isa<SCEVConstant>(SI->first) || Share == Stride)
2352 int64_t SSInt = cast<SCEVConstant>(Share)->getValue()->getSExtValue();
2354 return true; // This can definitely be reused.
2355 if (unsigned(abs64(SSInt)) < SInt || (SSInt % SInt) != 0)
2357 int64_t Scale = SSInt / SInt;
2358 bool AllUsesAreAddresses = true;
2359 bool AllUsesAreOutsideLoop = true;
2360 std::vector<BasedUser> UsersToProcess;
2361 const SCEV *CommonExprs = CollectIVUsers(SI->first, *SI->second, L,
2362 AllUsesAreAddresses,
2363 AllUsesAreOutsideLoop,
2365 if (AllUsesAreAddresses &&
2366 ValidScale(!CommonExprs->isZero(), Scale, UsersToProcess)) {
2369 // Any pre-inc iv use?
2370 IVUsersOfOneStride &StrideUses = *IU->IVUsesByStride[Share];
2371 for (ilist<IVStrideUse>::iterator I = StrideUses.Users.begin(),
2372 E = StrideUses.Users.end(); I != E; ++I) {
2373 if (!I->isUseOfPostIncrementedValue())
2381 /// isUsedByExitBranch - Return true if icmp is used by a loop terminating
2382 /// conditional branch or it's and / or with other conditions before being used
2383 /// as the condition.
2384 static bool isUsedByExitBranch(ICmpInst *Cond, Loop *L) {
2385 BasicBlock *CondBB = Cond->getParent();
2386 if (!L->isLoopExiting(CondBB))
2388 BranchInst *TermBr = dyn_cast<BranchInst>(CondBB->getTerminator());
2389 if (!TermBr || !TermBr->isConditional())
2392 Value *User = *Cond->use_begin();
2393 Instruction *UserInst = dyn_cast<Instruction>(User);
2395 (UserInst->getOpcode() == Instruction::And ||
2396 UserInst->getOpcode() == Instruction::Or)) {
2397 if (!UserInst->hasOneUse() || UserInst->getParent() != CondBB)
2399 User = *User->use_begin();
2400 UserInst = dyn_cast<Instruction>(User);
2402 return User == TermBr;
2405 static bool ShouldCountToZero(ICmpInst *Cond, IVStrideUse* &CondUse,
2406 ScalarEvolution *SE, Loop *L,
2407 const TargetLowering *TLI = 0) {
2408 if (!L->contains(Cond->getParent()))
2411 if (!isa<SCEVConstant>(CondUse->getOffset()))
2414 // Handle only tests for equality for the moment.
2415 if (!Cond->isEquality() || !Cond->hasOneUse())
2417 if (!isUsedByExitBranch(Cond, L))
2420 Value *CondOp0 = Cond->getOperand(0);
2421 const SCEV *IV = SE->getSCEV(CondOp0);
2422 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2423 if (!AR || !AR->isAffine())
2426 const SCEVConstant *SC = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE));
2427 if (!SC || SC->getValue()->getSExtValue() < 0)
2428 // If it's already counting down, don't do anything.
2431 // If the RHS of the comparison is not an loop invariant, the rewrite
2432 // cannot be done. Also bail out if it's already comparing against a zero.
2433 // If we are checking this before cmp stride optimization, check if it's
2434 // comparing against a already legal immediate.
2435 Value *RHS = Cond->getOperand(1);
2436 ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS);
2437 if (!L->isLoopInvariant(RHS) ||
2438 (RHSC && RHSC->isZero()) ||
2439 (RHSC && TLI && TLI->isLegalICmpImmediate(RHSC->getSExtValue())))
2442 // Make sure the IV is only used for counting. Value may be preinc or
2443 // postinc; 2 uses in either case.
2444 if (!CondOp0->hasNUses(2))
2450 /// OptimizeLoopTermCond - Change loop terminating condition to use the
2451 /// postinc iv when possible.
2452 void LoopStrengthReduce::OptimizeLoopTermCond(Loop *L) {
2453 BasicBlock *LatchBlock = L->getLoopLatch();
2454 bool LatchExit = L->isLoopExiting(LatchBlock);
2455 SmallVector<BasicBlock*, 8> ExitingBlocks;
2456 L->getExitingBlocks(ExitingBlocks);
2458 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
2459 BasicBlock *ExitingBlock = ExitingBlocks[i];
2461 // Finally, get the terminating condition for the loop if possible. If we
2462 // can, we want to change it to use a post-incremented version of its
2463 // induction variable, to allow coalescing the live ranges for the IV into
2464 // one register value.
2466 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2469 // FIXME: Overly conservative, termination condition could be an 'or' etc..
2470 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
2473 // Search IVUsesByStride to find Cond's IVUse if there is one.
2474 IVStrideUse *CondUse = 0;
2475 const SCEV *CondStride = 0;
2476 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2477 if (!FindIVUserForCond(Cond, CondUse, CondStride))
2480 // If the latch block is exiting and it's not a single block loop, it's
2481 // not safe to use postinc iv in other exiting blocks. FIXME: overly
2482 // conservative? How about icmp stride optimization?
2483 bool UsePostInc = !(e > 1 && LatchExit && ExitingBlock != LatchBlock);
2484 if (UsePostInc && ExitingBlock != LatchBlock) {
2485 if (!Cond->hasOneUse())
2486 // See below, we don't want the condition to be cloned.
2489 // If exiting block is the latch block, we know it's safe and profitable
2490 // to transform the icmp to use post-inc iv. Otherwise do so only if it
2491 // would not reuse another iv and its iv would be reused by other uses.
2492 // We are optimizing for the case where the icmp is the only use of the
2494 IVUsersOfOneStride &StrideUses = *IU->IVUsesByStride[CondStride];
2495 for (ilist<IVStrideUse>::iterator I = StrideUses.Users.begin(),
2496 E = StrideUses.Users.end(); I != E; ++I) {
2497 if (I->getUser() == Cond)
2499 if (!I->isUseOfPostIncrementedValue()) {
2506 // If iv for the stride might be shared and any of the users use pre-inc
2507 // iv might be used, then it's not safe to use post-inc iv.
2509 isa<SCEVConstant>(CondStride) &&
2510 StrideMightBeShared(CondStride, L, true))
2514 // If the trip count is computed in terms of a max (due to ScalarEvolution
2515 // being unable to find a sufficient guard, for example), change the loop
2516 // comparison to use SLT or ULT instead of NE.
2517 Cond = OptimizeMax(L, Cond, CondUse);
2519 // If possible, change stride and operands of the compare instruction to
2520 // eliminate one stride. However, avoid rewriting the compare instruction
2521 // with an iv of new stride if it's likely the new stride uses will be
2522 // rewritten using the stride of the compare instruction.
2523 if (ExitingBlock == LatchBlock && isa<SCEVConstant>(CondStride)) {
2524 // If the condition stride is a constant and it's the only use, we might
2525 // want to optimize it first by turning it to count toward zero.
2526 if (!StrideMightBeShared(CondStride, L, false) &&
2527 !ShouldCountToZero(Cond, CondUse, SE, L, TLI))
2528 Cond = ChangeCompareStride(L, Cond, CondUse, CondStride);
2534 DEBUG(errs() << " Change loop exiting icmp to use postinc iv: "
2537 // It's possible for the setcc instruction to be anywhere in the loop, and
2538 // possible for it to have multiple users. If it is not immediately before
2539 // the exiting block branch, move it.
2540 if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
2541 if (Cond->hasOneUse()) { // Condition has a single use, just move it.
2542 Cond->moveBefore(TermBr);
2544 // Otherwise, clone the terminating condition and insert into the
2546 Cond = cast<ICmpInst>(Cond->clone());
2547 Cond->setName(L->getHeader()->getName() + ".termcond");
2548 ExitingBlock->getInstList().insert(TermBr, Cond);
2550 // Clone the IVUse, as the old use still exists!
2551 IU->IVUsesByStride[CondStride]->addUser(CondUse->getOffset(), Cond,
2552 CondUse->getOperandValToReplace());
2553 CondUse = &IU->IVUsesByStride[CondStride]->Users.back();
2557 // If we get to here, we know that we can transform the setcc instruction to
2558 // use the post-incremented version of the IV, allowing us to coalesce the
2559 // live ranges for the IV correctly.
2560 CondUse->setOffset(SE->getMinusSCEV(CondUse->getOffset(), CondStride));
2561 CondUse->setIsUseOfPostIncrementedValue(true);
2568 bool LoopStrengthReduce::OptimizeLoopCountIVOfStride(const SCEV* &Stride,
2569 IVStrideUse* &CondUse,
2571 // If the only use is an icmp of a loop exiting conditional branch, then
2572 // attempt the optimization.
2573 BasedUser User = BasedUser(*CondUse, SE);
2574 assert(isa<ICmpInst>(User.Inst) && "Expecting an ICMPInst!");
2575 ICmpInst *Cond = cast<ICmpInst>(User.Inst);
2577 // Less strict check now that compare stride optimization is done.
2578 if (!ShouldCountToZero(Cond, CondUse, SE, L))
2581 Value *CondOp0 = Cond->getOperand(0);
2582 PHINode *PHIExpr = dyn_cast<PHINode>(CondOp0);
2585 // Value tested is postinc. Find the phi node.
2586 Incr = dyn_cast<BinaryOperator>(CondOp0);
2587 // FIXME: Just use User.OperandValToReplace here?
2588 if (!Incr || Incr->getOpcode() != Instruction::Add)
2591 PHIExpr = dyn_cast<PHINode>(Incr->getOperand(0));
2594 // 1 use for preinc value, the increment.
2595 if (!PHIExpr->hasOneUse())
2598 assert(isa<PHINode>(CondOp0) &&
2599 "Unexpected loop exiting counting instruction sequence!");
2600 PHIExpr = cast<PHINode>(CondOp0);
2601 // Value tested is preinc. Find the increment.
2602 // A CmpInst is not a BinaryOperator; we depend on this.
2603 Instruction::use_iterator UI = PHIExpr->use_begin();
2604 Incr = dyn_cast<BinaryOperator>(UI);
2606 Incr = dyn_cast<BinaryOperator>(++UI);
2607 // One use for postinc value, the phi. Unnecessarily conservative?
2608 if (!Incr || !Incr->hasOneUse() || Incr->getOpcode() != Instruction::Add)
2612 // Replace the increment with a decrement.
2613 DEBUG(errs() << "LSR: Examining use ");
2614 DEBUG(WriteAsOperand(errs(), CondOp0, /*PrintType=*/false));
2615 DEBUG(errs() << " in Inst: " << *Cond << '\n');
2616 BinaryOperator *Decr = BinaryOperator::Create(Instruction::Sub,
2617 Incr->getOperand(0), Incr->getOperand(1), "tmp", Incr);
2618 Incr->replaceAllUsesWith(Decr);
2619 Incr->eraseFromParent();
2621 // Substitute endval-startval for the original startval, and 0 for the
2622 // original endval. Since we're only testing for equality this is OK even
2623 // if the computation wraps around.
2624 BasicBlock *Preheader = L->getLoopPreheader();
2625 Instruction *PreInsertPt = Preheader->getTerminator();
2626 unsigned InBlock = L->contains(PHIExpr->getIncomingBlock(0)) ? 1 : 0;
2627 Value *StartVal = PHIExpr->getIncomingValue(InBlock);
2628 Value *EndVal = Cond->getOperand(1);
2629 DEBUG(errs() << " Optimize loop counting iv to count down ["
2630 << *EndVal << " .. " << *StartVal << "]\n");
2632 // FIXME: check for case where both are constant.
2633 Constant* Zero = ConstantInt::get(Cond->getOperand(1)->getType(), 0);
2634 BinaryOperator *NewStartVal = BinaryOperator::Create(Instruction::Sub,
2635 EndVal, StartVal, "tmp", PreInsertPt);
2636 PHIExpr->setIncomingValue(InBlock, NewStartVal);
2637 Cond->setOperand(1, Zero);
2638 DEBUG(errs() << " New icmp: " << *Cond << "\n");
2640 int64_t SInt = cast<SCEVConstant>(Stride)->getValue()->getSExtValue();
2641 const SCEV *NewStride = 0;
2643 for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
2644 const SCEV *OldStride = IU->StrideOrder[i];
2645 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(OldStride))
2646 if (SC->getValue()->getSExtValue() == -SInt) {
2648 NewStride = OldStride;
2654 NewStride = SE->getIntegerSCEV(-SInt, Stride->getType());
2655 IU->AddUser(NewStride, CondUse->getOffset(), Cond, Cond->getOperand(0));
2656 IU->IVUsesByStride[Stride]->removeUser(CondUse);
2658 CondUse = &IU->IVUsesByStride[NewStride]->Users.back();
2666 /// OptimizeLoopCountIV - If, after all sharing of IVs, the IV used for deciding
2667 /// when to exit the loop is used only for that purpose, try to rearrange things
2668 /// so it counts down to a test against zero.
2669 bool LoopStrengthReduce::OptimizeLoopCountIV(Loop *L) {
2670 bool ThisChanged = false;
2671 for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
2672 const SCEV *Stride = IU->StrideOrder[i];
2673 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
2674 IU->IVUsesByStride.find(Stride);
2675 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
2676 // FIXME: Generalize to non-affine IV's.
2677 if (!SI->first->isLoopInvariant(L))
2679 // If stride is a constant and it has an icmpinst use, check if we can
2680 // optimize the loop to count down.
2681 if (isa<SCEVConstant>(Stride) && SI->second->Users.size() == 1) {
2682 Instruction *User = SI->second->Users.begin()->getUser();
2683 if (!isa<ICmpInst>(User))
2685 const SCEV *CondStride = Stride;
2686 IVStrideUse *Use = &*SI->second->Users.begin();
2687 if (!OptimizeLoopCountIVOfStride(CondStride, Use, L))
2691 // Now check if it's possible to reuse this iv for other stride uses.
2692 for (unsigned j = 0, ee = IU->StrideOrder.size(); j != ee; ++j) {
2693 const SCEV *SStride = IU->StrideOrder[j];
2694 if (SStride == CondStride)
2696 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SII =
2697 IU->IVUsesByStride.find(SStride);
2698 assert(SII != IU->IVUsesByStride.end() && "Stride doesn't exist!");
2699 // FIXME: Generalize to non-affine IV's.
2700 if (!SII->first->isLoopInvariant(L))
2702 // FIXME: Rewrite other stride using CondStride.
2707 Changed |= ThisChanged;
2711 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager &LPM) {
2712 IU = &getAnalysis<IVUsers>();
2713 SE = &getAnalysis<ScalarEvolution>();
2716 // If LoopSimplify form is not available, stay out of trouble.
2717 if (!L->getLoopPreheader() || !L->getLoopLatch())
2720 if (!IU->IVUsesByStride.empty()) {
2721 DEBUG(errs() << "\nLSR on \"" << L->getHeader()->getParent()->getName()
2725 // Sort the StrideOrder so we process larger strides first.
2726 std::stable_sort(IU->StrideOrder.begin(), IU->StrideOrder.end(),
2729 // Optimize induction variables. Some indvar uses can be transformed to use
2730 // strides that will be needed for other purposes. A common example of this
2731 // is the exit test for the loop, which can often be rewritten to use the
2732 // computation of some other indvar to decide when to terminate the loop.
2735 // Change loop terminating condition to use the postinc iv when possible
2736 // and optimize loop terminating compare. FIXME: Move this after
2737 // StrengthReduceIVUsersOfStride?
2738 OptimizeLoopTermCond(L);
2740 // FIXME: We can shrink overlarge IV's here. e.g. if the code has
2741 // computation in i64 values and the target doesn't support i64, demote
2742 // the computation to 32-bit if safe.
2744 // FIXME: Attempt to reuse values across multiple IV's. In particular, we
2745 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should
2746 // be codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC.
2747 // Need to be careful that IV's are all the same type. Only works for
2748 // intptr_t indvars.
2750 // IVsByStride keeps IVs for one particular loop.
2751 assert(IVsByStride.empty() && "Stale entries in IVsByStride?");
2753 StrengthReduceIVUsers(L);
2755 // After all sharing is done, see if we can adjust the loop to test against
2756 // zero instead of counting up to a maximum. This is usually faster.
2757 OptimizeLoopCountIV(L);
2759 // We're done analyzing this loop; release all the state we built up for it.
2760 IVsByStride.clear();
2762 // Clean up after ourselves
2763 if (!DeadInsts.empty())
2764 DeleteTriviallyDeadInstructions();
2767 // At this point, it is worth checking to see if any recurrence PHIs are also
2768 // dead, so that we can remove them as well.
2769 DeleteDeadPHIs(L->getHeader());