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/Dominators.h"
30 #include "llvm/Analysis/IVUsers.h"
31 #include "llvm/Analysis/LoopInfo.h"
32 #include "llvm/Analysis/LoopPass.h"
33 #include "llvm/Analysis/ScalarEvolutionExpander.h"
34 #include "llvm/Transforms/Utils/AddrModeMatcher.h"
35 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
36 #include "llvm/Transforms/Utils/Local.h"
37 #include "llvm/ADT/Statistic.h"
38 #include "llvm/Support/CFG.h"
39 #include "llvm/Support/Debug.h"
40 #include "llvm/Support/CommandLine.h"
41 #include "llvm/Support/ValueHandle.h"
42 #include "llvm/Support/raw_ostream.h"
43 #include "llvm/Target/TargetLowering.h"
47 STATISTIC(NumReduced , "Number of IV uses strength reduced");
48 STATISTIC(NumInserted, "Number of PHIs inserted");
49 STATISTIC(NumVariable, "Number of PHIs with variable strides");
50 STATISTIC(NumEliminated, "Number of strides eliminated");
51 STATISTIC(NumShadow, "Number of Shadow IVs optimized");
52 STATISTIC(NumImmSunk, "Number of common expr immediates sunk into uses");
53 STATISTIC(NumLoopCond, "Number of loop terminating conds optimized");
54 STATISTIC(NumCountZero, "Number of count iv optimized to count toward zero");
56 static cl::opt<bool> EnableFullLSRMode("enable-full-lsr",
64 /// IVInfo - This structure keeps track of one IV expression inserted during
65 /// StrengthReduceStridedIVUsers. It contains the stride, the common base, as
66 /// well as the PHI node and increment value created for rewrite.
72 IVExpr(const SCEV *const stride, const SCEV *const base, PHINode *phi)
73 : Stride(stride), Base(base), PHI(phi) {}
76 /// IVsOfOneStride - This structure keeps track of all IV expression inserted
77 /// during StrengthReduceStridedIVUsers for a particular stride of the IV.
78 struct IVsOfOneStride {
79 std::vector<IVExpr> IVs;
81 void addIV(const SCEV *const Stride, const SCEV *const Base, PHINode *PHI) {
82 IVs.push_back(IVExpr(Stride, Base, PHI));
86 class LoopStrengthReduce : public LoopPass {
93 /// IVsByStride - Keep track of all IVs that have been inserted for a
94 /// particular stride.
95 std::map<const SCEV *, IVsOfOneStride> IVsByStride;
97 /// StrideNoReuse - Keep track of all the strides whose ivs cannot be
98 /// reused (nor should they be rewritten to reuse other strides).
99 SmallSet<const SCEV *, 4> StrideNoReuse;
101 /// DeadInsts - Keep track of instructions we may have made dead, so that
102 /// we can remove them after we are done working.
103 SmallVector<WeakVH, 16> DeadInsts;
105 /// TLI - Keep a pointer of a TargetLowering to consult for determining
106 /// transformation profitability.
107 const TargetLowering *TLI;
110 static char ID; // Pass ID, replacement for typeid
111 explicit LoopStrengthReduce(const TargetLowering *tli = NULL) :
112 LoopPass(&ID), TLI(tli) {}
114 bool runOnLoop(Loop *L, LPPassManager &LPM);
116 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
117 // We split critical edges, so we change the CFG. However, we do update
118 // many analyses if they are around.
119 AU.addPreservedID(LoopSimplifyID);
120 AU.addPreserved<LoopInfo>();
121 AU.addPreserved<DominanceFrontier>();
122 AU.addPreserved<DominatorTree>();
124 AU.addRequiredID(LoopSimplifyID);
125 AU.addRequired<LoopInfo>();
126 AU.addRequired<DominatorTree>();
127 AU.addRequired<ScalarEvolution>();
128 AU.addPreserved<ScalarEvolution>();
129 AU.addRequired<IVUsers>();
130 AU.addPreserved<IVUsers>();
134 void OptimizeIndvars(Loop *L);
136 /// OptimizeLoopTermCond - Change loop terminating condition to use the
137 /// postinc iv when possible.
138 void OptimizeLoopTermCond(Loop *L);
140 /// OptimizeShadowIV - If IV is used in a int-to-float cast
141 /// inside the loop then try to eliminate the cast opeation.
142 void OptimizeShadowIV(Loop *L);
144 /// OptimizeMax - Rewrite the loop's terminating condition
145 /// if it uses a max computation.
146 ICmpInst *OptimizeMax(Loop *L, ICmpInst *Cond,
147 IVStrideUse* &CondUse);
149 /// OptimizeLoopCountIV - If, after all sharing of IVs, the IV used for
150 /// deciding when to exit the loop is used only for that purpose, try to
151 /// rearrange things so it counts down to a test against zero.
152 bool OptimizeLoopCountIV(Loop *L);
153 bool OptimizeLoopCountIVOfStride(const SCEV* &Stride,
154 IVStrideUse* &CondUse, Loop *L);
156 /// StrengthReduceIVUsersOfStride - Strength reduce all of the users of a
157 /// single stride of IV. All of the users may have different starting
158 /// values, and this may not be the only stride.
159 void StrengthReduceIVUsersOfStride(const SCEV *const &Stride,
160 IVUsersOfOneStride &Uses,
162 void StrengthReduceIVUsers(Loop *L);
164 ICmpInst *ChangeCompareStride(Loop *L, ICmpInst *Cond,
165 IVStrideUse* &CondUse,
166 const SCEV* &CondStride,
167 bool PostPass = false);
169 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse,
170 const SCEV* &CondStride);
171 bool RequiresTypeConversion(const Type *Ty, const Type *NewTy);
172 const SCEV *CheckForIVReuse(bool, bool, bool, const SCEV *const&,
173 IVExpr&, const Type*,
174 const std::vector<BasedUser>& UsersToProcess);
175 bool ValidScale(bool, int64_t,
176 const std::vector<BasedUser>& UsersToProcess);
177 bool ValidOffset(bool, int64_t, int64_t,
178 const std::vector<BasedUser>& UsersToProcess);
179 const SCEV *CollectIVUsers(const SCEV *const &Stride,
180 IVUsersOfOneStride &Uses,
182 bool &AllUsesAreAddresses,
183 bool &AllUsesAreOutsideLoop,
184 std::vector<BasedUser> &UsersToProcess);
185 bool StrideMightBeShared(const SCEV *Stride, Loop *L, bool CheckPreInc);
186 bool ShouldUseFullStrengthReductionMode(
187 const std::vector<BasedUser> &UsersToProcess,
189 bool AllUsesAreAddresses,
191 void PrepareToStrengthReduceFully(
192 std::vector<BasedUser> &UsersToProcess,
194 const SCEV *CommonExprs,
196 SCEVExpander &PreheaderRewriter);
197 void PrepareToStrengthReduceFromSmallerStride(
198 std::vector<BasedUser> &UsersToProcess,
200 const IVExpr &ReuseIV,
201 Instruction *PreInsertPt);
202 void PrepareToStrengthReduceWithNewPhi(
203 std::vector<BasedUser> &UsersToProcess,
205 const SCEV *CommonExprs,
207 Instruction *IVIncInsertPt,
209 SCEVExpander &PreheaderRewriter);
211 void DeleteTriviallyDeadInstructions();
215 char LoopStrengthReduce::ID = 0;
216 static RegisterPass<LoopStrengthReduce>
217 X("loop-reduce", "Loop Strength Reduction");
219 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
220 return new LoopStrengthReduce(TLI);
223 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
224 /// specified set are trivially dead, delete them and see if this makes any of
225 /// their operands subsequently dead.
226 void LoopStrengthReduce::DeleteTriviallyDeadInstructions() {
227 if (DeadInsts.empty()) return;
229 while (!DeadInsts.empty()) {
230 Instruction *I = dyn_cast_or_null<Instruction>(DeadInsts.pop_back_val());
232 if (I == 0 || !isInstructionTriviallyDead(I))
235 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
236 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
239 DeadInsts.push_back(U);
242 I->eraseFromParent();
247 /// containsAddRecFromDifferentLoop - Determine whether expression S involves a
248 /// subexpression that is an AddRec from a loop other than L. An outer loop
249 /// of L is OK, but not an inner loop nor a disjoint loop.
250 static bool containsAddRecFromDifferentLoop(const SCEV *S, Loop *L) {
251 // This is very common, put it first.
252 if (isa<SCEVConstant>(S))
254 if (const SCEVCommutativeExpr *AE = dyn_cast<SCEVCommutativeExpr>(S)) {
255 for (unsigned int i=0; i< AE->getNumOperands(); i++)
256 if (containsAddRecFromDifferentLoop(AE->getOperand(i), L))
260 if (const SCEVAddRecExpr *AE = dyn_cast<SCEVAddRecExpr>(S)) {
261 if (const Loop *newLoop = AE->getLoop()) {
264 // if newLoop is an outer loop of L, this is OK.
265 if (!LoopInfo::isNotAlreadyContainedIn(L, newLoop))
270 if (const SCEVUDivExpr *DE = dyn_cast<SCEVUDivExpr>(S))
271 return containsAddRecFromDifferentLoop(DE->getLHS(), L) ||
272 containsAddRecFromDifferentLoop(DE->getRHS(), L);
274 // SCEVSDivExpr has been backed out temporarily, but will be back; we'll
275 // need this when it is.
276 if (const SCEVSDivExpr *DE = dyn_cast<SCEVSDivExpr>(S))
277 return containsAddRecFromDifferentLoop(DE->getLHS(), L) ||
278 containsAddRecFromDifferentLoop(DE->getRHS(), L);
280 if (const SCEVCastExpr *CE = dyn_cast<SCEVCastExpr>(S))
281 return containsAddRecFromDifferentLoop(CE->getOperand(), L);
285 /// isAddressUse - Returns true if the specified instruction is using the
286 /// specified value as an address.
287 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
288 bool isAddress = isa<LoadInst>(Inst);
289 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
290 if (SI->getOperand(1) == OperandVal)
292 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
293 // Addressing modes can also be folded into prefetches and a variety
295 switch (II->getIntrinsicID()) {
297 case Intrinsic::prefetch:
298 case Intrinsic::x86_sse2_loadu_dq:
299 case Intrinsic::x86_sse2_loadu_pd:
300 case Intrinsic::x86_sse_loadu_ps:
301 case Intrinsic::x86_sse_storeu_ps:
302 case Intrinsic::x86_sse2_storeu_pd:
303 case Intrinsic::x86_sse2_storeu_dq:
304 case Intrinsic::x86_sse2_storel_dq:
305 if (II->getOperand(1) == OperandVal)
313 /// getAccessType - Return the type of the memory being accessed.
314 static const Type *getAccessType(const Instruction *Inst) {
315 const Type *AccessTy = Inst->getType();
316 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
317 AccessTy = SI->getOperand(0)->getType();
318 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
319 // Addressing modes can also be folded into prefetches and a variety
321 switch (II->getIntrinsicID()) {
323 case Intrinsic::x86_sse_storeu_ps:
324 case Intrinsic::x86_sse2_storeu_pd:
325 case Intrinsic::x86_sse2_storeu_dq:
326 case Intrinsic::x86_sse2_storel_dq:
327 AccessTy = II->getOperand(1)->getType();
335 /// BasedUser - For a particular base value, keep information about how we've
336 /// partitioned the expression so far.
338 /// SE - The current ScalarEvolution object.
341 /// Base - The Base value for the PHI node that needs to be inserted for
342 /// this use. As the use is processed, information gets moved from this
343 /// field to the Imm field (below). BasedUser values are sorted by this
347 /// Inst - The instruction using the induction variable.
350 /// OperandValToReplace - The operand value of Inst to replace with the
352 Value *OperandValToReplace;
354 /// Imm - The immediate value that should be added to the base immediately
355 /// before Inst, because it will be folded into the imm field of the
356 /// instruction. This is also sometimes used for loop-variant values that
357 /// must be added inside the loop.
360 /// Phi - The induction variable that performs the striding that
361 /// should be used for this user.
364 // isUseOfPostIncrementedValue - True if this should use the
365 // post-incremented version of this IV, not the preincremented version.
366 // This can only be set in special cases, such as the terminating setcc
367 // instruction for a loop and uses outside the loop that are dominated by
369 bool isUseOfPostIncrementedValue;
371 BasedUser(IVStrideUse &IVSU, ScalarEvolution *se)
372 : SE(se), Base(IVSU.getOffset()), Inst(IVSU.getUser()),
373 OperandValToReplace(IVSU.getOperandValToReplace()),
374 Imm(SE->getIntegerSCEV(0, Base->getType())),
375 isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue()) {}
377 // Once we rewrite the code to insert the new IVs we want, update the
378 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
380 void RewriteInstructionToUseNewBase(const SCEV *const &NewBase,
381 Instruction *InsertPt,
382 SCEVExpander &Rewriter, Loop *L, Pass *P,
383 SmallVectorImpl<WeakVH> &DeadInsts);
385 Value *InsertCodeForBaseAtPosition(const SCEV *const &NewBase,
387 SCEVExpander &Rewriter,
393 void BasedUser::dump() const {
394 errs() << " Base=" << *Base;
395 errs() << " Imm=" << *Imm;
396 errs() << " Inst: " << *Inst;
399 Value *BasedUser::InsertCodeForBaseAtPosition(const SCEV *const &NewBase,
401 SCEVExpander &Rewriter,
403 Value *Base = Rewriter.expandCodeFor(NewBase, 0, IP);
405 // Wrap the base in a SCEVUnknown so that ScalarEvolution doesn't try to
407 const SCEV *NewValSCEV = SE->getUnknown(Base);
409 // Always emit the immediate into the same block as the user.
410 NewValSCEV = SE->getAddExpr(NewValSCEV, Imm);
412 return Rewriter.expandCodeFor(NewValSCEV, Ty, IP);
416 // Once we rewrite the code to insert the new IVs we want, update the
417 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
418 // to it. NewBasePt is the last instruction which contributes to the
419 // value of NewBase in the case that it's a diffferent instruction from
420 // the PHI that NewBase is computed from, or null otherwise.
422 void BasedUser::RewriteInstructionToUseNewBase(const SCEV *const &NewBase,
423 Instruction *NewBasePt,
424 SCEVExpander &Rewriter, Loop *L, Pass *P,
425 SmallVectorImpl<WeakVH> &DeadInsts) {
426 if (!isa<PHINode>(Inst)) {
427 // By default, insert code at the user instruction.
428 BasicBlock::iterator InsertPt = Inst;
430 // However, if the Operand is itself an instruction, the (potentially
431 // complex) inserted code may be shared by many users. Because of this, we
432 // want to emit code for the computation of the operand right before its old
433 // computation. This is usually safe, because we obviously used to use the
434 // computation when it was computed in its current block. However, in some
435 // cases (e.g. use of a post-incremented induction variable) the NewBase
436 // value will be pinned to live somewhere after the original computation.
437 // In this case, we have to back off.
439 // If this is a use outside the loop (which means after, since it is based
440 // on a loop indvar) we use the post-incremented value, so that we don't
441 // artificially make the preinc value live out the bottom of the loop.
442 if (!isUseOfPostIncrementedValue && L->contains(Inst->getParent())) {
443 if (NewBasePt && isa<PHINode>(OperandValToReplace)) {
444 InsertPt = NewBasePt;
446 } else if (Instruction *OpInst
447 = dyn_cast<Instruction>(OperandValToReplace)) {
449 while (isa<PHINode>(InsertPt)) ++InsertPt;
452 Value *NewVal = InsertCodeForBaseAtPosition(NewBase,
453 OperandValToReplace->getType(),
455 // Replace the use of the operand Value with the new Phi we just created.
456 Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
458 DEBUG(errs() << " Replacing with ");
459 DEBUG(WriteAsOperand(errs(), NewVal, /*PrintType=*/false));
460 DEBUG(errs() << ", which has value " << *NewBase << " plus IMM "
465 // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm
466 // expression into each operand block that uses it. Note that PHI nodes can
467 // have multiple entries for the same predecessor. We use a map to make sure
468 // that a PHI node only has a single Value* for each predecessor (which also
469 // prevents us from inserting duplicate code in some blocks).
470 DenseMap<BasicBlock*, Value*> InsertedCode;
471 PHINode *PN = cast<PHINode>(Inst);
472 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
473 if (PN->getIncomingValue(i) == OperandValToReplace) {
474 // If the original expression is outside the loop, put the replacement
475 // code in the same place as the original expression,
476 // which need not be an immediate predecessor of this PHI. This way we
477 // need only one copy of it even if it is referenced multiple times in
478 // the PHI. We don't do this when the original expression is inside the
479 // loop because multiple copies sometimes do useful sinking of code in
481 Instruction *OldLoc = dyn_cast<Instruction>(OperandValToReplace);
482 BasicBlock *PHIPred = PN->getIncomingBlock(i);
483 if (L->contains(OldLoc->getParent())) {
484 // If this is a critical edge, split the edge so that we do not insert
485 // the code on all predecessor/successor paths. We do this unless this
486 // is the canonical backedge for this loop, as this can make some
487 // inserted code be in an illegal position.
488 if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 &&
489 !isa<IndirectBrInst>(PHIPred->getTerminator()) &&
490 (PN->getParent() != L->getHeader() || !L->contains(PHIPred))) {
492 // First step, split the critical edge.
493 BasicBlock *NewBB = SplitCriticalEdge(PHIPred, PN->getParent(),
496 // Next step: move the basic block. In particular, if the PHI node
497 // is outside of the loop, and PredTI is in the loop, we want to
498 // move the block to be immediately before the PHI block, not
499 // immediately after PredTI.
500 if (L->contains(PHIPred) && !L->contains(PN->getParent()))
501 NewBB->moveBefore(PN->getParent());
503 // Splitting the edge can reduce the number of PHI entries we have.
504 e = PN->getNumIncomingValues();
506 i = PN->getBasicBlockIndex(PHIPred);
509 Value *&Code = InsertedCode[PHIPred];
511 // Insert the code into the end of the predecessor block.
512 Instruction *InsertPt = (L->contains(OldLoc->getParent())) ?
513 PHIPred->getTerminator() :
514 OldLoc->getParent()->getTerminator();
515 Code = InsertCodeForBaseAtPosition(NewBase, PN->getType(),
518 DEBUG(errs() << " Changing PHI use to ");
519 DEBUG(WriteAsOperand(errs(), Code, /*PrintType=*/false));
520 DEBUG(errs() << ", which has value " << *NewBase << " plus IMM "
524 // Replace the use of the operand Value with the new Phi we just created.
525 PN->setIncomingValue(i, Code);
530 // PHI node might have become a constant value after SplitCriticalEdge.
531 DeadInsts.push_back(Inst);
535 /// fitsInAddressMode - Return true if V can be subsumed within an addressing
536 /// mode, and does not need to be put in a register first.
537 static bool fitsInAddressMode(const SCEV *const &V, const Type *AccessTy,
538 const TargetLowering *TLI, bool HasBaseReg) {
539 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
540 int64_t VC = SC->getValue()->getSExtValue();
542 TargetLowering::AddrMode AM;
544 AM.HasBaseReg = HasBaseReg;
545 return TLI->isLegalAddressingMode(AM, AccessTy);
547 // Defaults to PPC. PPC allows a sign-extended 16-bit immediate field.
548 return (VC > -(1 << 16) && VC < (1 << 16)-1);
552 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
553 if (GlobalValue *GV = dyn_cast<GlobalValue>(SU->getValue())) {
555 TargetLowering::AddrMode AM;
557 AM.HasBaseReg = HasBaseReg;
558 return TLI->isLegalAddressingMode(AM, AccessTy);
560 // Default: assume global addresses are not legal.
567 /// MoveLoopVariantsToImmediateField - Move any subexpressions from Val that are
568 /// loop varying to the Imm operand.
569 static void MoveLoopVariantsToImmediateField(const SCEV *&Val, const SCEV *&Imm,
570 Loop *L, ScalarEvolution *SE) {
571 if (Val->isLoopInvariant(L)) return; // Nothing to do.
573 if (const SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
574 SmallVector<const SCEV *, 4> NewOps;
575 NewOps.reserve(SAE->getNumOperands());
577 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
578 if (!SAE->getOperand(i)->isLoopInvariant(L)) {
579 // If this is a loop-variant expression, it must stay in the immediate
580 // field of the expression.
581 Imm = SE->getAddExpr(Imm, SAE->getOperand(i));
583 NewOps.push_back(SAE->getOperand(i));
587 Val = SE->getIntegerSCEV(0, Val->getType());
589 Val = SE->getAddExpr(NewOps);
590 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
591 // Try to pull immediates out of the start value of nested addrec's.
592 const SCEV *Start = SARE->getStart();
593 MoveLoopVariantsToImmediateField(Start, Imm, L, SE);
595 SmallVector<const SCEV *, 4> Ops(SARE->op_begin(), SARE->op_end());
597 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
599 // Otherwise, all of Val is variant, move the whole thing over.
600 Imm = SE->getAddExpr(Imm, Val);
601 Val = SE->getIntegerSCEV(0, Val->getType());
606 /// MoveImmediateValues - Look at Val, and pull out any additions of constants
607 /// that can fit into the immediate field of instructions in the target.
608 /// Accumulate these immediate values into the Imm value.
609 static void MoveImmediateValues(const TargetLowering *TLI,
610 const Type *AccessTy,
611 const SCEV *&Val, const SCEV *&Imm,
612 bool isAddress, Loop *L,
613 ScalarEvolution *SE) {
614 if (const SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
615 SmallVector<const SCEV *, 4> NewOps;
616 NewOps.reserve(SAE->getNumOperands());
618 for (unsigned i = 0; i != SAE->getNumOperands(); ++i) {
619 const SCEV *NewOp = SAE->getOperand(i);
620 MoveImmediateValues(TLI, AccessTy, NewOp, Imm, isAddress, L, SE);
622 if (!NewOp->isLoopInvariant(L)) {
623 // If this is a loop-variant expression, it must stay in the immediate
624 // field of the expression.
625 Imm = SE->getAddExpr(Imm, NewOp);
627 NewOps.push_back(NewOp);
632 Val = SE->getIntegerSCEV(0, Val->getType());
634 Val = SE->getAddExpr(NewOps);
636 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
637 // Try to pull immediates out of the start value of nested addrec's.
638 const SCEV *Start = SARE->getStart();
639 MoveImmediateValues(TLI, AccessTy, Start, Imm, isAddress, L, SE);
641 if (Start != SARE->getStart()) {
642 SmallVector<const SCEV *, 4> Ops(SARE->op_begin(), SARE->op_end());
644 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
647 } else if (const SCEVMulExpr *SME = dyn_cast<SCEVMulExpr>(Val)) {
648 // Transform "8 * (4 + v)" -> "32 + 8*V" if "32" fits in the immed field.
650 fitsInAddressMode(SME->getOperand(0), AccessTy, TLI, false) &&
651 SME->getNumOperands() == 2 && SME->isLoopInvariant(L)) {
653 const SCEV *SubImm = SE->getIntegerSCEV(0, Val->getType());
654 const SCEV *NewOp = SME->getOperand(1);
655 MoveImmediateValues(TLI, AccessTy, NewOp, SubImm, isAddress, L, SE);
657 // If we extracted something out of the subexpressions, see if we can
659 if (NewOp != SME->getOperand(1)) {
660 // Scale SubImm up by "8". If the result is a target constant, we are
662 SubImm = SE->getMulExpr(SubImm, SME->getOperand(0));
663 if (fitsInAddressMode(SubImm, AccessTy, TLI, false)) {
664 // Accumulate the immediate.
665 Imm = SE->getAddExpr(Imm, SubImm);
667 // Update what is left of 'Val'.
668 Val = SE->getMulExpr(SME->getOperand(0), NewOp);
675 // Loop-variant expressions must stay in the immediate field of the
677 if ((isAddress && fitsInAddressMode(Val, AccessTy, TLI, false)) ||
678 !Val->isLoopInvariant(L)) {
679 Imm = SE->getAddExpr(Imm, Val);
680 Val = SE->getIntegerSCEV(0, Val->getType());
684 // Otherwise, no immediates to move.
687 static void MoveImmediateValues(const TargetLowering *TLI,
689 const SCEV *&Val, const SCEV *&Imm,
690 bool isAddress, Loop *L,
691 ScalarEvolution *SE) {
692 const Type *AccessTy = getAccessType(User);
693 MoveImmediateValues(TLI, AccessTy, Val, Imm, isAddress, L, SE);
696 /// SeparateSubExprs - Decompose Expr into all of the subexpressions that are
697 /// added together. This is used to reassociate common addition subexprs
698 /// together for maximal sharing when rewriting bases.
699 static void SeparateSubExprs(SmallVector<const SCEV *, 16> &SubExprs,
701 ScalarEvolution *SE) {
702 if (const SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) {
703 for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
704 SeparateSubExprs(SubExprs, AE->getOperand(j), SE);
705 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) {
706 const SCEV *Zero = SE->getIntegerSCEV(0, Expr->getType());
707 if (SARE->getOperand(0) == Zero) {
708 SubExprs.push_back(Expr);
710 // Compute the addrec with zero as its base.
711 SmallVector<const SCEV *, 4> Ops(SARE->op_begin(), SARE->op_end());
712 Ops[0] = Zero; // Start with zero base.
713 SubExprs.push_back(SE->getAddRecExpr(Ops, SARE->getLoop()));
716 SeparateSubExprs(SubExprs, SARE->getOperand(0), SE);
718 } else if (!Expr->isZero()) {
720 SubExprs.push_back(Expr);
724 // This is logically local to the following function, but C++ says we have
725 // to make it file scope.
726 struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; };
728 /// RemoveCommonExpressionsFromUseBases - Look through all of the Bases of all
729 /// the Uses, removing any common subexpressions, except that if all such
730 /// subexpressions can be folded into an addressing mode for all uses inside
731 /// the loop (this case is referred to as "free" in comments herein) we do
732 /// not remove anything. This looks for things like (a+b+c) and
733 /// (a+c+d) and computes the common (a+c) subexpression. The common expression
734 /// is *removed* from the Bases and returned.
736 RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses,
737 ScalarEvolution *SE, Loop *L,
738 const TargetLowering *TLI) {
739 unsigned NumUses = Uses.size();
741 // Only one use? This is a very common case, so we handle it specially and
743 const SCEV *Zero = SE->getIntegerSCEV(0, Uses[0].Base->getType());
744 const SCEV *Result = Zero;
745 const SCEV *FreeResult = Zero;
747 // If the use is inside the loop, use its base, regardless of what it is:
748 // it is clearly shared across all the IV's. If the use is outside the loop
749 // (which means after it) we don't want to factor anything *into* the loop,
750 // so just use 0 as the base.
751 if (L->contains(Uses[0].Inst->getParent()))
752 std::swap(Result, Uses[0].Base);
756 // To find common subexpressions, count how many of Uses use each expression.
757 // If any subexpressions are used Uses.size() times, they are common.
758 // Also track whether all uses of each expression can be moved into an
759 // an addressing mode "for free"; such expressions are left within the loop.
760 // struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; };
761 std::map<const SCEV *, SubExprUseData> SubExpressionUseData;
763 // UniqueSubExprs - Keep track of all of the subexpressions we see in the
764 // order we see them.
765 SmallVector<const SCEV *, 16> UniqueSubExprs;
767 SmallVector<const SCEV *, 16> SubExprs;
768 unsigned NumUsesInsideLoop = 0;
769 for (unsigned i = 0; i != NumUses; ++i) {
770 // If the user is outside the loop, just ignore it for base computation.
771 // Since the user is outside the loop, it must be *after* the loop (if it
772 // were before, it could not be based on the loop IV). We don't want users
773 // after the loop to affect base computation of values *inside* the loop,
774 // because we can always add their offsets to the result IV after the loop
775 // is done, ensuring we get good code inside the loop.
776 if (!L->contains(Uses[i].Inst->getParent()))
780 // If the base is zero (which is common), return zero now, there are no
782 if (Uses[i].Base == Zero) return Zero;
784 // If this use is as an address we may be able to put CSEs in the addressing
785 // mode rather than hoisting them.
786 bool isAddrUse = isAddressUse(Uses[i].Inst, Uses[i].OperandValToReplace);
787 // We may need the AccessTy below, but only when isAddrUse, so compute it
788 // only in that case.
789 const Type *AccessTy = 0;
791 AccessTy = getAccessType(Uses[i].Inst);
793 // Split the expression into subexprs.
794 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
795 // Add one to SubExpressionUseData.Count for each subexpr present, and
796 // if the subexpr is not a valid immediate within an addressing mode use,
797 // set SubExpressionUseData.notAllUsesAreFree. We definitely want to
798 // hoist these out of the loop (if they are common to all uses).
799 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) {
800 if (++SubExpressionUseData[SubExprs[j]].Count == 1)
801 UniqueSubExprs.push_back(SubExprs[j]);
802 if (!isAddrUse || !fitsInAddressMode(SubExprs[j], AccessTy, TLI, false))
803 SubExpressionUseData[SubExprs[j]].notAllUsesAreFree = true;
808 // Now that we know how many times each is used, build Result. Iterate over
809 // UniqueSubexprs so that we have a stable ordering.
810 for (unsigned i = 0, e = UniqueSubExprs.size(); i != e; ++i) {
811 std::map<const SCEV *, SubExprUseData>::iterator I =
812 SubExpressionUseData.find(UniqueSubExprs[i]);
813 assert(I != SubExpressionUseData.end() && "Entry not found?");
814 if (I->second.Count == NumUsesInsideLoop) { // Found CSE!
815 if (I->second.notAllUsesAreFree)
816 Result = SE->getAddExpr(Result, I->first);
818 FreeResult = SE->getAddExpr(FreeResult, I->first);
820 // Remove non-cse's from SubExpressionUseData.
821 SubExpressionUseData.erase(I);
824 if (FreeResult != Zero) {
825 // We have some subexpressions that can be subsumed into addressing
826 // modes in every use inside the loop. However, it's possible that
827 // there are so many of them that the combined FreeResult cannot
828 // be subsumed, or that the target cannot handle both a FreeResult
829 // and a Result in the same instruction (for example because it would
830 // require too many registers). Check this.
831 for (unsigned i=0; i<NumUses; ++i) {
832 if (!L->contains(Uses[i].Inst->getParent()))
834 // We know this is an addressing mode use; if there are any uses that
835 // are not, FreeResult would be Zero.
836 const Type *AccessTy = getAccessType(Uses[i].Inst);
837 if (!fitsInAddressMode(FreeResult, AccessTy, TLI, Result!=Zero)) {
838 // FIXME: could split up FreeResult into pieces here, some hoisted
839 // and some not. There is no obvious advantage to this.
840 Result = SE->getAddExpr(Result, FreeResult);
847 // If we found no CSE's, return now.
848 if (Result == Zero) return Result;
850 // If we still have a FreeResult, remove its subexpressions from
851 // SubExpressionUseData. This means they will remain in the use Bases.
852 if (FreeResult != Zero) {
853 SeparateSubExprs(SubExprs, FreeResult, SE);
854 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) {
855 std::map<const SCEV *, SubExprUseData>::iterator I =
856 SubExpressionUseData.find(SubExprs[j]);
857 SubExpressionUseData.erase(I);
862 // Otherwise, remove all of the CSE's we found from each of the base values.
863 for (unsigned i = 0; i != NumUses; ++i) {
864 // Uses outside the loop don't necessarily include the common base, but
865 // the final IV value coming into those uses does. Instead of trying to
866 // remove the pieces of the common base, which might not be there,
867 // subtract off the base to compensate for this.
868 if (!L->contains(Uses[i].Inst->getParent())) {
869 Uses[i].Base = SE->getMinusSCEV(Uses[i].Base, Result);
873 // Split the expression into subexprs.
874 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
876 // Remove any common subexpressions.
877 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
878 if (SubExpressionUseData.count(SubExprs[j])) {
879 SubExprs.erase(SubExprs.begin()+j);
883 // Finally, add the non-shared expressions together.
884 if (SubExprs.empty())
887 Uses[i].Base = SE->getAddExpr(SubExprs);
894 /// ValidScale - Check whether the given Scale is valid for all loads and
895 /// stores in UsersToProcess.
897 bool LoopStrengthReduce::ValidScale(bool HasBaseReg, int64_t Scale,
898 const std::vector<BasedUser>& UsersToProcess) {
902 for (unsigned i = 0, e = UsersToProcess.size(); i!=e; ++i) {
903 // If this is a load or other access, pass the type of the access in.
904 const Type *AccessTy =
905 Type::getVoidTy(UsersToProcess[i].Inst->getContext());
906 if (isAddressUse(UsersToProcess[i].Inst,
907 UsersToProcess[i].OperandValToReplace))
908 AccessTy = getAccessType(UsersToProcess[i].Inst);
909 else if (isa<PHINode>(UsersToProcess[i].Inst))
912 TargetLowering::AddrMode AM;
913 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
914 AM.BaseOffs = SC->getValue()->getSExtValue();
915 AM.HasBaseReg = HasBaseReg || !UsersToProcess[i].Base->isZero();
918 // If load[imm+r*scale] is illegal, bail out.
919 if (!TLI->isLegalAddressingMode(AM, AccessTy))
925 /// ValidOffset - Check whether the given Offset is valid for all loads and
926 /// stores in UsersToProcess.
928 bool LoopStrengthReduce::ValidOffset(bool HasBaseReg,
931 const std::vector<BasedUser>& UsersToProcess) {
935 for (unsigned i=0, e = UsersToProcess.size(); i!=e; ++i) {
936 // If this is a load or other access, pass the type of the access in.
937 const Type *AccessTy =
938 Type::getVoidTy(UsersToProcess[i].Inst->getContext());
939 if (isAddressUse(UsersToProcess[i].Inst,
940 UsersToProcess[i].OperandValToReplace))
941 AccessTy = getAccessType(UsersToProcess[i].Inst);
942 else if (isa<PHINode>(UsersToProcess[i].Inst))
945 TargetLowering::AddrMode AM;
946 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
947 AM.BaseOffs = SC->getValue()->getSExtValue();
948 AM.BaseOffs = (uint64_t)AM.BaseOffs + (uint64_t)Offset;
949 AM.HasBaseReg = HasBaseReg || !UsersToProcess[i].Base->isZero();
952 // If load[imm+r*scale] is illegal, bail out.
953 if (!TLI->isLegalAddressingMode(AM, AccessTy))
959 /// RequiresTypeConversion - Returns true if converting Ty1 to Ty2 is not
961 bool LoopStrengthReduce::RequiresTypeConversion(const Type *Ty1,
965 Ty1 = SE->getEffectiveSCEVType(Ty1);
966 Ty2 = SE->getEffectiveSCEVType(Ty2);
969 if (Ty1->canLosslesslyBitCastTo(Ty2))
971 if (TLI && TLI->isTruncateFree(Ty1, Ty2))
976 /// CheckForIVReuse - Returns the multiple if the stride is the multiple
977 /// of a previous stride and it is a legal value for the target addressing
978 /// mode scale component and optional base reg. This allows the users of
979 /// this stride to be rewritten as prev iv * factor. It returns 0 if no
980 /// reuse is possible. Factors can be negative on same targets, e.g. ARM.
982 /// If all uses are outside the loop, we don't require that all multiplies
983 /// be folded into the addressing mode, nor even that the factor be constant;
984 /// a multiply (executed once) outside the loop is better than another IV
985 /// within. Well, usually.
986 const SCEV *LoopStrengthReduce::CheckForIVReuse(bool HasBaseReg,
987 bool AllUsesAreAddresses,
988 bool AllUsesAreOutsideLoop,
989 const SCEV *const &Stride,
990 IVExpr &IV, const Type *Ty,
991 const std::vector<BasedUser>& UsersToProcess) {
992 if (StrideNoReuse.count(Stride))
993 return SE->getIntegerSCEV(0, Stride->getType());
995 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Stride)) {
996 int64_t SInt = SC->getValue()->getSExtValue();
997 for (unsigned NewStride = 0, e = IU->StrideOrder.size();
998 NewStride != e; ++NewStride) {
999 std::map<const SCEV *, IVsOfOneStride>::iterator SI =
1000 IVsByStride.find(IU->StrideOrder[NewStride]);
1001 if (SI == IVsByStride.end() || !isa<SCEVConstant>(SI->first) ||
1002 StrideNoReuse.count(SI->first))
1004 // The other stride has no uses, don't reuse it.
1005 std::map<const SCEV *, IVUsersOfOneStride *>::iterator UI =
1006 IU->IVUsesByStride.find(IU->StrideOrder[NewStride]);
1007 if (UI->second->Users.empty())
1009 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1010 if (SI->first != Stride &&
1011 (unsigned(abs64(SInt)) < SSInt || (SInt % SSInt) != 0))
1013 int64_t Scale = SInt / SSInt;
1014 // Check that this stride is valid for all the types used for loads and
1015 // stores; if it can be used for some and not others, we might as well use
1016 // the original stride everywhere, since we have to create the IV for it
1017 // anyway. If the scale is 1, then we don't need to worry about folding
1020 (AllUsesAreAddresses &&
1021 ValidScale(HasBaseReg, Scale, UsersToProcess))) {
1022 // Prefer to reuse an IV with a base of zero.
1023 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1024 IE = SI->second.IVs.end(); II != IE; ++II)
1025 // Only reuse previous IV if it would not require a type conversion
1026 // and if the base difference can be folded.
1027 if (II->Base->isZero() &&
1028 !RequiresTypeConversion(II->Base->getType(), Ty)) {
1030 return SE->getIntegerSCEV(Scale, Stride->getType());
1032 // Otherwise, settle for an IV with a foldable base.
1033 if (AllUsesAreAddresses)
1034 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1035 IE = SI->second.IVs.end(); II != IE; ++II)
1036 // Only reuse previous IV if it would not require a type conversion
1037 // and if the base difference can be folded.
1038 if (SE->getEffectiveSCEVType(II->Base->getType()) ==
1039 SE->getEffectiveSCEVType(Ty) &&
1040 isa<SCEVConstant>(II->Base)) {
1042 cast<SCEVConstant>(II->Base)->getValue()->getSExtValue();
1043 if (Base > INT32_MIN && Base <= INT32_MAX &&
1044 ValidOffset(HasBaseReg, -Base * Scale,
1045 Scale, UsersToProcess)) {
1047 return SE->getIntegerSCEV(Scale, Stride->getType());
1052 } else if (AllUsesAreOutsideLoop) {
1053 // Accept nonconstant strides here; it is really really right to substitute
1054 // an existing IV if we can.
1055 for (unsigned NewStride = 0, e = IU->StrideOrder.size();
1056 NewStride != e; ++NewStride) {
1057 std::map<const SCEV *, IVsOfOneStride>::iterator SI =
1058 IVsByStride.find(IU->StrideOrder[NewStride]);
1059 if (SI == IVsByStride.end() || !isa<SCEVConstant>(SI->first))
1061 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1062 if (SI->first != Stride && SSInt != 1)
1064 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1065 IE = SI->second.IVs.end(); II != IE; ++II)
1066 // Accept nonzero base here.
1067 // Only reuse previous IV if it would not require a type conversion.
1068 if (!RequiresTypeConversion(II->Base->getType(), Ty)) {
1073 // Special case, old IV is -1*x and this one is x. Can treat this one as
1075 for (unsigned NewStride = 0, e = IU->StrideOrder.size();
1076 NewStride != e; ++NewStride) {
1077 std::map<const SCEV *, IVsOfOneStride>::iterator SI =
1078 IVsByStride.find(IU->StrideOrder[NewStride]);
1079 if (SI == IVsByStride.end())
1081 if (const SCEVMulExpr *ME = dyn_cast<SCEVMulExpr>(SI->first))
1082 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(ME->getOperand(0)))
1083 if (Stride == ME->getOperand(1) &&
1084 SC->getValue()->getSExtValue() == -1LL)
1085 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1086 IE = SI->second.IVs.end(); II != IE; ++II)
1087 // Accept nonzero base here.
1088 // Only reuse previous IV if it would not require type conversion.
1089 if (!RequiresTypeConversion(II->Base->getType(), Ty)) {
1091 return SE->getIntegerSCEV(-1LL, Stride->getType());
1095 return SE->getIntegerSCEV(0, Stride->getType());
1098 /// PartitionByIsUseOfPostIncrementedValue - Simple boolean predicate that
1099 /// returns true if Val's isUseOfPostIncrementedValue is true.
1100 static bool PartitionByIsUseOfPostIncrementedValue(const BasedUser &Val) {
1101 return Val.isUseOfPostIncrementedValue;
1104 /// isNonConstantNegative - Return true if the specified scev is negated, but
1106 static bool isNonConstantNegative(const SCEV *const &Expr) {
1107 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Expr);
1108 if (!Mul) return false;
1110 // If there is a constant factor, it will be first.
1111 const SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
1112 if (!SC) return false;
1114 // Return true if the value is negative, this matches things like (-42 * V).
1115 return SC->getValue()->getValue().isNegative();
1118 /// CollectIVUsers - Transform our list of users and offsets to a bit more
1119 /// complex table. In this new vector, each 'BasedUser' contains 'Base', the
1120 /// base of the strided accesses, as well as the old information from Uses. We
1121 /// progressively move information from the Base field to the Imm field, until
1122 /// we eventually have the full access expression to rewrite the use.
1123 const SCEV *LoopStrengthReduce::CollectIVUsers(const SCEV *const &Stride,
1124 IVUsersOfOneStride &Uses,
1126 bool &AllUsesAreAddresses,
1127 bool &AllUsesAreOutsideLoop,
1128 std::vector<BasedUser> &UsersToProcess) {
1129 // FIXME: Generalize to non-affine IV's.
1130 if (!Stride->isLoopInvariant(L))
1131 return SE->getIntegerSCEV(0, Stride->getType());
1133 UsersToProcess.reserve(Uses.Users.size());
1134 for (ilist<IVStrideUse>::iterator I = Uses.Users.begin(),
1135 E = Uses.Users.end(); I != E; ++I) {
1136 UsersToProcess.push_back(BasedUser(*I, SE));
1138 // Move any loop variant operands from the offset field to the immediate
1139 // field of the use, so that we don't try to use something before it is
1141 MoveLoopVariantsToImmediateField(UsersToProcess.back().Base,
1142 UsersToProcess.back().Imm, L, SE);
1143 assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
1144 "Base value is not loop invariant!");
1147 // We now have a whole bunch of uses of like-strided induction variables, but
1148 // they might all have different bases. We want to emit one PHI node for this
1149 // stride which we fold as many common expressions (between the IVs) into as
1150 // possible. Start by identifying the common expressions in the base values
1151 // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
1152 // "A+B"), emit it to the preheader, then remove the expression from the
1153 // UsersToProcess base values.
1154 const SCEV *CommonExprs =
1155 RemoveCommonExpressionsFromUseBases(UsersToProcess, SE, L, TLI);
1157 // Next, figure out what we can represent in the immediate fields of
1158 // instructions. If we can represent anything there, move it to the imm
1159 // fields of the BasedUsers. We do this so that it increases the commonality
1160 // of the remaining uses.
1161 unsigned NumPHI = 0;
1162 bool HasAddress = false;
1163 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1164 // If the user is not in the current loop, this means it is using the exit
1165 // value of the IV. Do not put anything in the base, make sure it's all in
1166 // the immediate field to allow as much factoring as possible.
1167 if (!L->contains(UsersToProcess[i].Inst->getParent())) {
1168 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm,
1169 UsersToProcess[i].Base);
1170 UsersToProcess[i].Base =
1171 SE->getIntegerSCEV(0, UsersToProcess[i].Base->getType());
1173 // Not all uses are outside the loop.
1174 AllUsesAreOutsideLoop = false;
1176 // Addressing modes can be folded into loads and stores. Be careful that
1177 // the store is through the expression, not of the expression though.
1179 bool isAddress = isAddressUse(UsersToProcess[i].Inst,
1180 UsersToProcess[i].OperandValToReplace);
1181 if (isa<PHINode>(UsersToProcess[i].Inst)) {
1189 // If this use isn't an address, then not all uses are addresses.
1190 if (!isAddress && !isPHI)
1191 AllUsesAreAddresses = false;
1193 MoveImmediateValues(TLI, UsersToProcess[i].Inst, UsersToProcess[i].Base,
1194 UsersToProcess[i].Imm, isAddress, L, SE);
1198 // If one of the use is a PHI node and all other uses are addresses, still
1199 // allow iv reuse. Essentially we are trading one constant multiplication
1200 // for one fewer iv.
1202 AllUsesAreAddresses = false;
1204 // There are no in-loop address uses.
1205 if (AllUsesAreAddresses && (!HasAddress && !AllUsesAreOutsideLoop))
1206 AllUsesAreAddresses = false;
1211 /// ShouldUseFullStrengthReductionMode - Test whether full strength-reduction
1212 /// is valid and profitable for the given set of users of a stride. In
1213 /// full strength-reduction mode, all addresses at the current stride are
1214 /// strength-reduced all the way down to pointer arithmetic.
1216 bool LoopStrengthReduce::ShouldUseFullStrengthReductionMode(
1217 const std::vector<BasedUser> &UsersToProcess,
1219 bool AllUsesAreAddresses,
1220 const SCEV *Stride) {
1221 if (!EnableFullLSRMode)
1224 // The heuristics below aim to avoid increasing register pressure, but
1225 // fully strength-reducing all the addresses increases the number of
1226 // add instructions, so don't do this when optimizing for size.
1227 // TODO: If the loop is large, the savings due to simpler addresses
1228 // may oughtweight the costs of the extra increment instructions.
1229 if (L->getHeader()->getParent()->hasFnAttr(Attribute::OptimizeForSize))
1232 // TODO: For now, don't do full strength reduction if there could
1233 // potentially be greater-stride multiples of the current stride
1234 // which could reuse the current stride IV.
1235 if (IU->StrideOrder.back() != Stride)
1238 // Iterate through the uses to find conditions that automatically rule out
1240 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ) {
1241 const SCEV *Base = UsersToProcess[i].Base;
1242 const SCEV *Imm = UsersToProcess[i].Imm;
1243 // If any users have a loop-variant component, they can't be fully
1244 // strength-reduced.
1245 if (Imm && !Imm->isLoopInvariant(L))
1247 // If there are to users with the same base and the difference between
1248 // the two Imm values can't be folded into the address, full
1249 // strength reduction would increase register pressure.
1251 const SCEV *CurImm = UsersToProcess[i].Imm;
1252 if ((CurImm || Imm) && CurImm != Imm) {
1253 if (!CurImm) CurImm = SE->getIntegerSCEV(0, Stride->getType());
1254 if (!Imm) Imm = SE->getIntegerSCEV(0, Stride->getType());
1255 const Instruction *Inst = UsersToProcess[i].Inst;
1256 const Type *AccessTy = getAccessType(Inst);
1257 const SCEV *Diff = SE->getMinusSCEV(UsersToProcess[i].Imm, Imm);
1258 if (!Diff->isZero() &&
1259 (!AllUsesAreAddresses ||
1260 !fitsInAddressMode(Diff, AccessTy, TLI, /*HasBaseReg=*/true)))
1263 } while (++i != e && Base == UsersToProcess[i].Base);
1266 // If there's exactly one user in this stride, fully strength-reducing it
1267 // won't increase register pressure. If it's starting from a non-zero base,
1268 // it'll be simpler this way.
1269 if (UsersToProcess.size() == 1 && !UsersToProcess[0].Base->isZero())
1272 // Otherwise, if there are any users in this stride that don't require
1273 // a register for their base, full strength-reduction will increase
1274 // register pressure.
1275 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1276 if (UsersToProcess[i].Base->isZero())
1279 // Otherwise, go for it.
1283 /// InsertAffinePhi Create and insert a PHI node for an induction variable
1284 /// with the specified start and step values in the specified loop.
1286 /// If NegateStride is true, the stride should be negated by using a
1287 /// subtract instead of an add.
1289 /// Return the created phi node.
1291 static PHINode *InsertAffinePhi(const SCEV *Start, const SCEV *Step,
1292 Instruction *IVIncInsertPt,
1294 SCEVExpander &Rewriter) {
1295 assert(Start->isLoopInvariant(L) && "New PHI start is not loop invariant!");
1296 assert(Step->isLoopInvariant(L) && "New PHI stride is not loop invariant!");
1298 BasicBlock *Header = L->getHeader();
1299 BasicBlock *Preheader = L->getLoopPreheader();
1300 BasicBlock *LatchBlock = L->getLoopLatch();
1301 const Type *Ty = Start->getType();
1302 Ty = Rewriter.SE.getEffectiveSCEVType(Ty);
1304 PHINode *PN = PHINode::Create(Ty, "lsr.iv", Header->begin());
1305 PN->addIncoming(Rewriter.expandCodeFor(Start, Ty, Preheader->getTerminator()),
1308 // If the stride is negative, insert a sub instead of an add for the
1310 bool isNegative = isNonConstantNegative(Step);
1311 const SCEV *IncAmount = Step;
1313 IncAmount = Rewriter.SE.getNegativeSCEV(Step);
1315 // Insert an add instruction right before the terminator corresponding
1316 // to the back-edge or just before the only use. The location is determined
1317 // by the caller and passed in as IVIncInsertPt.
1318 Value *StepV = Rewriter.expandCodeFor(IncAmount, Ty,
1319 Preheader->getTerminator());
1322 IncV = BinaryOperator::CreateSub(PN, StepV, "lsr.iv.next",
1325 IncV = BinaryOperator::CreateAdd(PN, StepV, "lsr.iv.next",
1328 if (!isa<ConstantInt>(StepV)) ++NumVariable;
1330 PN->addIncoming(IncV, LatchBlock);
1336 static void SortUsersToProcess(std::vector<BasedUser> &UsersToProcess) {
1337 // We want to emit code for users inside the loop first. To do this, we
1338 // rearrange BasedUser so that the entries at the end have
1339 // isUseOfPostIncrementedValue = false, because we pop off the end of the
1340 // vector (so we handle them first).
1341 std::partition(UsersToProcess.begin(), UsersToProcess.end(),
1342 PartitionByIsUseOfPostIncrementedValue);
1344 // Sort this by base, so that things with the same base are handled
1345 // together. By partitioning first and stable-sorting later, we are
1346 // guaranteed that within each base we will pop off users from within the
1347 // loop before users outside of the loop with a particular base.
1349 // We would like to use stable_sort here, but we can't. The problem is that
1350 // const SCEV *'s don't have a deterministic ordering w.r.t to each other, so
1351 // we don't have anything to do a '<' comparison on. Because we think the
1352 // number of uses is small, do a horrible bubble sort which just relies on
1354 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1355 // Get a base value.
1356 const SCEV *Base = UsersToProcess[i].Base;
1358 // Compact everything with this base to be consecutive with this one.
1359 for (unsigned j = i+1; j != e; ++j) {
1360 if (UsersToProcess[j].Base == Base) {
1361 std::swap(UsersToProcess[i+1], UsersToProcess[j]);
1368 /// PrepareToStrengthReduceFully - Prepare to fully strength-reduce
1369 /// UsersToProcess, meaning lowering addresses all the way down to direct
1370 /// pointer arithmetic.
1373 LoopStrengthReduce::PrepareToStrengthReduceFully(
1374 std::vector<BasedUser> &UsersToProcess,
1376 const SCEV *CommonExprs,
1378 SCEVExpander &PreheaderRewriter) {
1379 DEBUG(errs() << " Fully reducing all users\n");
1381 // Rewrite the UsersToProcess records, creating a separate PHI for each
1382 // unique Base value.
1383 Instruction *IVIncInsertPt = L->getLoopLatch()->getTerminator();
1384 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ) {
1385 // TODO: The uses are grouped by base, but not sorted. We arbitrarily
1386 // pick the first Imm value here to start with, and adjust it for the
1388 const SCEV *Imm = UsersToProcess[i].Imm;
1389 const SCEV *Base = UsersToProcess[i].Base;
1390 const SCEV *Start = SE->getAddExpr(CommonExprs, Base, Imm);
1391 PHINode *Phi = InsertAffinePhi(Start, Stride, IVIncInsertPt, L,
1393 // Loop over all the users with the same base.
1395 UsersToProcess[i].Base = SE->getIntegerSCEV(0, Stride->getType());
1396 UsersToProcess[i].Imm = SE->getMinusSCEV(UsersToProcess[i].Imm, Imm);
1397 UsersToProcess[i].Phi = Phi;
1398 assert(UsersToProcess[i].Imm->isLoopInvariant(L) &&
1399 "ShouldUseFullStrengthReductionMode should reject this!");
1400 } while (++i != e && Base == UsersToProcess[i].Base);
1404 /// FindIVIncInsertPt - Return the location to insert the increment instruction.
1405 /// If the only use if a use of postinc value, (must be the loop termination
1406 /// condition), then insert it just before the use.
1407 static Instruction *FindIVIncInsertPt(std::vector<BasedUser> &UsersToProcess,
1409 if (UsersToProcess.size() == 1 &&
1410 UsersToProcess[0].isUseOfPostIncrementedValue &&
1411 L->contains(UsersToProcess[0].Inst->getParent()))
1412 return UsersToProcess[0].Inst;
1413 return L->getLoopLatch()->getTerminator();
1416 /// PrepareToStrengthReduceWithNewPhi - Insert a new induction variable for the
1417 /// given users to share.
1420 LoopStrengthReduce::PrepareToStrengthReduceWithNewPhi(
1421 std::vector<BasedUser> &UsersToProcess,
1423 const SCEV *CommonExprs,
1425 Instruction *IVIncInsertPt,
1427 SCEVExpander &PreheaderRewriter) {
1428 DEBUG(errs() << " Inserting new PHI:\n");
1430 PHINode *Phi = InsertAffinePhi(SE->getUnknown(CommonBaseV),
1431 Stride, IVIncInsertPt, L,
1434 // Remember this in case a later stride is multiple of this.
1435 IVsByStride[Stride].addIV(Stride, CommonExprs, Phi);
1437 // All the users will share this new IV.
1438 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1439 UsersToProcess[i].Phi = Phi;
1441 DEBUG(errs() << " IV=");
1442 DEBUG(WriteAsOperand(errs(), Phi, /*PrintType=*/false));
1443 DEBUG(errs() << "\n");
1446 /// PrepareToStrengthReduceFromSmallerStride - Prepare for the given users to
1447 /// reuse an induction variable with a stride that is a factor of the current
1448 /// induction variable.
1451 LoopStrengthReduce::PrepareToStrengthReduceFromSmallerStride(
1452 std::vector<BasedUser> &UsersToProcess,
1454 const IVExpr &ReuseIV,
1455 Instruction *PreInsertPt) {
1456 DEBUG(errs() << " Rewriting in terms of existing IV of STRIDE "
1457 << *ReuseIV.Stride << " and BASE " << *ReuseIV.Base << "\n");
1459 // All the users will share the reused IV.
1460 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1461 UsersToProcess[i].Phi = ReuseIV.PHI;
1463 Constant *C = dyn_cast<Constant>(CommonBaseV);
1465 (!C->isNullValue() &&
1466 !fitsInAddressMode(SE->getUnknown(CommonBaseV), CommonBaseV->getType(),
1468 // We want the common base emitted into the preheader! This is just
1469 // using cast as a copy so BitCast (no-op cast) is appropriate
1470 CommonBaseV = new BitCastInst(CommonBaseV, CommonBaseV->getType(),
1471 "commonbase", PreInsertPt);
1474 static bool IsImmFoldedIntoAddrMode(GlobalValue *GV, int64_t Offset,
1475 const Type *AccessTy,
1476 std::vector<BasedUser> &UsersToProcess,
1477 const TargetLowering *TLI) {
1478 SmallVector<Instruction*, 16> AddrModeInsts;
1479 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1480 if (UsersToProcess[i].isUseOfPostIncrementedValue)
1482 ExtAddrMode AddrMode =
1483 AddressingModeMatcher::Match(UsersToProcess[i].OperandValToReplace,
1484 AccessTy, UsersToProcess[i].Inst,
1485 AddrModeInsts, *TLI);
1486 if (GV && GV != AddrMode.BaseGV)
1488 if (Offset && !AddrMode.BaseOffs)
1489 // FIXME: How to accurate check it's immediate offset is folded.
1491 AddrModeInsts.clear();
1496 /// StrengthReduceIVUsersOfStride - Strength reduce all of the users of a single
1497 /// stride of IV. All of the users may have different starting values, and this
1498 /// may not be the only stride.
1500 LoopStrengthReduce::StrengthReduceIVUsersOfStride(const SCEV *const &Stride,
1501 IVUsersOfOneStride &Uses,
1503 // If all the users are moved to another stride, then there is nothing to do.
1504 if (Uses.Users.empty())
1507 // Keep track if every use in UsersToProcess is an address. If they all are,
1508 // we may be able to rewrite the entire collection of them in terms of a
1509 // smaller-stride IV.
1510 bool AllUsesAreAddresses = true;
1512 // Keep track if every use of a single stride is outside the loop. If so,
1513 // we want to be more aggressive about reusing a smaller-stride IV; a
1514 // multiply outside the loop is better than another IV inside. Well, usually.
1515 bool AllUsesAreOutsideLoop = true;
1517 // Transform our list of users and offsets to a bit more complex table. In
1518 // this new vector, each 'BasedUser' contains 'Base' the base of the
1519 // strided accessas well as the old information from Uses. We progressively
1520 // move information from the Base field to the Imm field, until we eventually
1521 // have the full access expression to rewrite the use.
1522 std::vector<BasedUser> UsersToProcess;
1523 const SCEV *CommonExprs = CollectIVUsers(Stride, Uses, L, AllUsesAreAddresses,
1524 AllUsesAreOutsideLoop,
1527 // Sort the UsersToProcess array so that users with common bases are
1528 // next to each other.
1529 SortUsersToProcess(UsersToProcess);
1531 // If we managed to find some expressions in common, we'll need to carry
1532 // their value in a register and add it in for each use. This will take up
1533 // a register operand, which potentially restricts what stride values are
1535 bool HaveCommonExprs = !CommonExprs->isZero();
1536 const Type *ReplacedTy = CommonExprs->getType();
1538 // If all uses are addresses, consider sinking the immediate part of the
1539 // common expression back into uses if they can fit in the immediate fields.
1540 if (TLI && HaveCommonExprs && AllUsesAreAddresses) {
1541 const SCEV *NewCommon = CommonExprs;
1542 const SCEV *Imm = SE->getIntegerSCEV(0, ReplacedTy);
1543 MoveImmediateValues(TLI, Type::getVoidTy(
1544 L->getLoopPreheader()->getContext()),
1545 NewCommon, Imm, true, L, SE);
1546 if (!Imm->isZero()) {
1549 // If the immediate part of the common expression is a GV, check if it's
1550 // possible to fold it into the target addressing mode.
1551 GlobalValue *GV = 0;
1552 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(Imm))
1553 GV = dyn_cast<GlobalValue>(SU->getValue());
1555 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Imm))
1556 Offset = SC->getValue()->getSExtValue();
1558 // Pass VoidTy as the AccessTy to be conservative, because
1559 // there could be multiple access types among all the uses.
1560 DoSink = IsImmFoldedIntoAddrMode(GV, Offset,
1561 Type::getVoidTy(L->getLoopPreheader()->getContext()),
1562 UsersToProcess, TLI);
1565 DEBUG(errs() << " Sinking " << *Imm << " back down into uses\n");
1566 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1567 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm, Imm);
1568 CommonExprs = NewCommon;
1569 HaveCommonExprs = !CommonExprs->isZero();
1575 // Now that we know what we need to do, insert the PHI node itself.
1577 DEBUG(errs() << "LSR: Examining IVs of TYPE " << *ReplacedTy << " of STRIDE "
1579 << " Common base: " << *CommonExprs << "\n");
1581 SCEVExpander Rewriter(*SE);
1582 SCEVExpander PreheaderRewriter(*SE);
1584 BasicBlock *Preheader = L->getLoopPreheader();
1585 Instruction *PreInsertPt = Preheader->getTerminator();
1586 BasicBlock *LatchBlock = L->getLoopLatch();
1587 Instruction *IVIncInsertPt = LatchBlock->getTerminator();
1589 Value *CommonBaseV = Constant::getNullValue(ReplacedTy);
1591 const SCEV *RewriteFactor = SE->getIntegerSCEV(0, ReplacedTy);
1592 IVExpr ReuseIV(SE->getIntegerSCEV(0,
1593 Type::getInt32Ty(Preheader->getContext())),
1594 SE->getIntegerSCEV(0,
1595 Type::getInt32Ty(Preheader->getContext())),
1598 // Choose a strength-reduction strategy and prepare for it by creating
1599 // the necessary PHIs and adjusting the bookkeeping.
1600 if (ShouldUseFullStrengthReductionMode(UsersToProcess, L,
1601 AllUsesAreAddresses, Stride)) {
1602 PrepareToStrengthReduceFully(UsersToProcess, Stride, CommonExprs, L,
1605 // Emit the initial base value into the loop preheader.
1606 CommonBaseV = PreheaderRewriter.expandCodeFor(CommonExprs, ReplacedTy,
1609 // If all uses are addresses, check if it is possible to reuse an IV. The
1610 // new IV must have a stride that is a multiple of the old stride; the
1611 // multiple must be a number that can be encoded in the scale field of the
1612 // target addressing mode; and we must have a valid instruction after this
1613 // substitution, including the immediate field, if any.
1614 RewriteFactor = CheckForIVReuse(HaveCommonExprs, AllUsesAreAddresses,
1615 AllUsesAreOutsideLoop,
1616 Stride, ReuseIV, ReplacedTy,
1618 if (!RewriteFactor->isZero())
1619 PrepareToStrengthReduceFromSmallerStride(UsersToProcess, CommonBaseV,
1620 ReuseIV, PreInsertPt);
1622 IVIncInsertPt = FindIVIncInsertPt(UsersToProcess, L);
1623 PrepareToStrengthReduceWithNewPhi(UsersToProcess, Stride, CommonExprs,
1624 CommonBaseV, IVIncInsertPt,
1625 L, PreheaderRewriter);
1629 // Process all the users now, replacing their strided uses with
1630 // strength-reduced forms. This outer loop handles all bases, the inner
1631 // loop handles all users of a particular base.
1632 while (!UsersToProcess.empty()) {
1633 const SCEV *Base = UsersToProcess.back().Base;
1634 Instruction *Inst = UsersToProcess.back().Inst;
1636 // Emit the code for Base into the preheader.
1638 if (!Base->isZero()) {
1639 BaseV = PreheaderRewriter.expandCodeFor(Base, 0, PreInsertPt);
1641 DEBUG(errs() << " INSERTING code for BASE = " << *Base << ":");
1642 if (BaseV->hasName())
1643 DEBUG(errs() << " Result value name = %" << BaseV->getName());
1644 DEBUG(errs() << "\n");
1646 // If BaseV is a non-zero constant, make sure that it gets inserted into
1647 // the preheader, instead of being forward substituted into the uses. We
1648 // do this by forcing a BitCast (noop cast) to be inserted into the
1649 // preheader in this case.
1650 if (!fitsInAddressMode(Base, getAccessType(Inst), TLI, false) &&
1651 isa<Constant>(BaseV)) {
1652 // We want this constant emitted into the preheader! This is just
1653 // using cast as a copy so BitCast (no-op cast) is appropriate
1654 BaseV = new BitCastInst(BaseV, BaseV->getType(), "preheaderinsert",
1659 // Emit the code to add the immediate offset to the Phi value, just before
1660 // the instructions that we identified as using this stride and base.
1662 // FIXME: Use emitted users to emit other users.
1663 BasedUser &User = UsersToProcess.back();
1665 DEBUG(errs() << " Examining ");
1666 if (User.isUseOfPostIncrementedValue)
1667 DEBUG(errs() << "postinc");
1669 DEBUG(errs() << "preinc");
1670 DEBUG(errs() << " use ");
1671 DEBUG(WriteAsOperand(errs(), UsersToProcess.back().OperandValToReplace,
1672 /*PrintType=*/false));
1673 DEBUG(errs() << " in Inst: " << *User.Inst);
1675 // If this instruction wants to use the post-incremented value, move it
1676 // after the post-inc and use its value instead of the PHI.
1677 Value *RewriteOp = User.Phi;
1678 if (User.isUseOfPostIncrementedValue) {
1679 RewriteOp = User.Phi->getIncomingValueForBlock(LatchBlock);
1680 // If this user is in the loop, make sure it is the last thing in the
1681 // loop to ensure it is dominated by the increment. In case it's the
1682 // only use of the iv, the increment instruction is already before the
1684 if (L->contains(User.Inst->getParent()) && User.Inst != IVIncInsertPt)
1685 User.Inst->moveBefore(IVIncInsertPt);
1688 const SCEV *RewriteExpr = SE->getUnknown(RewriteOp);
1690 if (SE->getEffectiveSCEVType(RewriteOp->getType()) !=
1691 SE->getEffectiveSCEVType(ReplacedTy)) {
1692 assert(SE->getTypeSizeInBits(RewriteOp->getType()) >
1693 SE->getTypeSizeInBits(ReplacedTy) &&
1694 "Unexpected widening cast!");
1695 RewriteExpr = SE->getTruncateExpr(RewriteExpr, ReplacedTy);
1698 // If we had to insert new instructions for RewriteOp, we have to
1699 // consider that they may not have been able to end up immediately
1700 // next to RewriteOp, because non-PHI instructions may never precede
1701 // PHI instructions in a block. In this case, remember where the last
1702 // instruction was inserted so that if we're replacing a different
1703 // PHI node, we can use the later point to expand the final
1705 Instruction *NewBasePt = dyn_cast<Instruction>(RewriteOp);
1706 if (RewriteOp == User.Phi) NewBasePt = 0;
1708 // Clear the SCEVExpander's expression map so that we are guaranteed
1709 // to have the code emitted where we expect it.
1712 // If we are reusing the iv, then it must be multiplied by a constant
1713 // factor to take advantage of the addressing mode scale component.
1714 if (!RewriteFactor->isZero()) {
1715 // If we're reusing an IV with a nonzero base (currently this happens
1716 // only when all reuses are outside the loop) subtract that base here.
1717 // The base has been used to initialize the PHI node but we don't want
1719 if (!ReuseIV.Base->isZero()) {
1720 const SCEV *typedBase = ReuseIV.Base;
1721 if (SE->getEffectiveSCEVType(RewriteExpr->getType()) !=
1722 SE->getEffectiveSCEVType(ReuseIV.Base->getType())) {
1723 // It's possible the original IV is a larger type than the new IV,
1724 // in which case we have to truncate the Base. We checked in
1725 // RequiresTypeConversion that this is valid.
1726 assert(SE->getTypeSizeInBits(RewriteExpr->getType()) <
1727 SE->getTypeSizeInBits(ReuseIV.Base->getType()) &&
1728 "Unexpected lengthening conversion!");
1729 typedBase = SE->getTruncateExpr(ReuseIV.Base,
1730 RewriteExpr->getType());
1732 RewriteExpr = SE->getMinusSCEV(RewriteExpr, typedBase);
1735 // Multiply old variable, with base removed, by new scale factor.
1736 RewriteExpr = SE->getMulExpr(RewriteFactor,
1739 // The common base is emitted in the loop preheader. But since we
1740 // are reusing an IV, it has not been used to initialize the PHI node.
1741 // Add it to the expression used to rewrite the uses.
1742 // When this use is outside the loop, we earlier subtracted the
1743 // common base, and are adding it back here. Use the same expression
1744 // as before, rather than CommonBaseV, so DAGCombiner will zap it.
1745 if (!CommonExprs->isZero()) {
1746 if (L->contains(User.Inst->getParent()))
1747 RewriteExpr = SE->getAddExpr(RewriteExpr,
1748 SE->getUnknown(CommonBaseV));
1750 RewriteExpr = SE->getAddExpr(RewriteExpr, CommonExprs);
1754 // Now that we know what we need to do, insert code before User for the
1755 // immediate and any loop-variant expressions.
1757 // Add BaseV to the PHI value if needed.
1758 RewriteExpr = SE->getAddExpr(RewriteExpr, SE->getUnknown(BaseV));
1760 User.RewriteInstructionToUseNewBase(RewriteExpr, NewBasePt,
1764 // Mark old value we replaced as possibly dead, so that it is eliminated
1765 // if we just replaced the last use of that value.
1766 DeadInsts.push_back(User.OperandValToReplace);
1768 UsersToProcess.pop_back();
1771 // If there are any more users to process with the same base, process them
1772 // now. We sorted by base above, so we just have to check the last elt.
1773 } while (!UsersToProcess.empty() && UsersToProcess.back().Base == Base);
1774 // TODO: Next, find out which base index is the most common, pull it out.
1777 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
1778 // different starting values, into different PHIs.
1781 void LoopStrengthReduce::StrengthReduceIVUsers(Loop *L) {
1782 // Note: this processes each stride/type pair individually. All users
1783 // passed into StrengthReduceIVUsersOfStride have the same type AND stride.
1784 // Also, note that we iterate over IVUsesByStride indirectly by using
1785 // StrideOrder. This extra layer of indirection makes the ordering of
1786 // strides deterministic - not dependent on map order.
1787 for (unsigned Stride = 0, e = IU->StrideOrder.size(); Stride != e; ++Stride) {
1788 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
1789 IU->IVUsesByStride.find(IU->StrideOrder[Stride]);
1790 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
1791 // FIXME: Generalize to non-affine IV's.
1792 if (!SI->first->isLoopInvariant(L))
1794 StrengthReduceIVUsersOfStride(SI->first, *SI->second, L);
1798 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1799 /// set the IV user and stride information and return true, otherwise return
1801 bool LoopStrengthReduce::FindIVUserForCond(ICmpInst *Cond,
1802 IVStrideUse *&CondUse,
1803 const SCEV* &CondStride) {
1804 for (unsigned Stride = 0, e = IU->StrideOrder.size();
1805 Stride != e && !CondUse; ++Stride) {
1806 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
1807 IU->IVUsesByStride.find(IU->StrideOrder[Stride]);
1808 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
1810 for (ilist<IVStrideUse>::iterator UI = SI->second->Users.begin(),
1811 E = SI->second->Users.end(); UI != E; ++UI)
1812 if (UI->getUser() == Cond) {
1813 // NOTE: we could handle setcc instructions with multiple uses here, but
1814 // InstCombine does it as well for simple uses, it's not clear that it
1815 // occurs enough in real life to handle.
1817 CondStride = SI->first;
1825 // Constant strides come first which in turns are sorted by their absolute
1826 // values. If absolute values are the same, then positive strides comes first.
1828 // 4, -1, X, 1, 2 ==> 1, -1, 2, 4, X
1829 struct StrideCompare {
1830 const ScalarEvolution *SE;
1831 explicit StrideCompare(const ScalarEvolution *se) : SE(se) {}
1833 bool operator()(const SCEV *const &LHS, const SCEV *const &RHS) {
1834 const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS);
1835 const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS);
1837 int64_t LV = LHSC->getValue()->getSExtValue();
1838 int64_t RV = RHSC->getValue()->getSExtValue();
1839 uint64_t ALV = (LV < 0) ? -LV : LV;
1840 uint64_t ARV = (RV < 0) ? -RV : RV;
1848 // If it's the same value but different type, sort by bit width so
1849 // that we emit larger induction variables before smaller
1850 // ones, letting the smaller be re-written in terms of larger ones.
1851 return SE->getTypeSizeInBits(RHS->getType()) <
1852 SE->getTypeSizeInBits(LHS->getType());
1854 return LHSC && !RHSC;
1859 /// ChangeCompareStride - If a loop termination compare instruction is the
1860 /// only use of its stride, and the compaison is against a constant value,
1861 /// try eliminate the stride by moving the compare instruction to another
1862 /// stride and change its constant operand accordingly. e.g.
1868 /// if (v2 < 10) goto loop
1873 /// if (v1 < 30) goto loop
1874 ICmpInst *LoopStrengthReduce::ChangeCompareStride(Loop *L, ICmpInst *Cond,
1875 IVStrideUse* &CondUse,
1876 const SCEV* &CondStride,
1878 // If there's only one stride in the loop, there's nothing to do here.
1879 if (IU->StrideOrder.size() < 2)
1881 // If there are other users of the condition's stride, don't bother
1882 // trying to change the condition because the stride will still
1884 std::map<const SCEV *, IVUsersOfOneStride *>::iterator I =
1885 IU->IVUsesByStride.find(CondStride);
1886 if (I == IU->IVUsesByStride.end())
1888 if (I->second->Users.size() > 1) {
1889 for (ilist<IVStrideUse>::iterator II = I->second->Users.begin(),
1890 EE = I->second->Users.end(); II != EE; ++II) {
1891 if (II->getUser() == Cond)
1893 if (!isInstructionTriviallyDead(II->getUser()))
1897 // Only handle constant strides for now.
1898 const SCEVConstant *SC = dyn_cast<SCEVConstant>(CondStride);
1899 if (!SC) return Cond;
1901 ICmpInst::Predicate Predicate = Cond->getPredicate();
1902 int64_t CmpSSInt = SC->getValue()->getSExtValue();
1903 unsigned BitWidth = SE->getTypeSizeInBits(CondStride->getType());
1904 uint64_t SignBit = 1ULL << (BitWidth-1);
1905 const Type *CmpTy = Cond->getOperand(0)->getType();
1906 const Type *NewCmpTy = NULL;
1907 unsigned TyBits = SE->getTypeSizeInBits(CmpTy);
1908 unsigned NewTyBits = 0;
1909 const SCEV *NewStride = NULL;
1910 Value *NewCmpLHS = NULL;
1911 Value *NewCmpRHS = NULL;
1913 const SCEV *NewOffset = SE->getIntegerSCEV(0, CmpTy);
1915 if (ConstantInt *C = dyn_cast<ConstantInt>(Cond->getOperand(1))) {
1916 int64_t CmpVal = C->getValue().getSExtValue();
1918 // Check the relevant induction variable for conformance to
1920 const SCEV *IV = SE->getSCEV(Cond->getOperand(0));
1921 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1922 if (!AR || !AR->isAffine())
1925 const SCEVConstant *StartC = dyn_cast<SCEVConstant>(AR->getStart());
1926 // Check stride constant and the comparision constant signs to detect
1929 if ((StartC->getValue()->getSExtValue() < CmpVal && CmpSSInt < 0) ||
1930 (StartC->getValue()->getSExtValue() > CmpVal && CmpSSInt > 0))
1933 // More restrictive check for the other cases.
1934 if ((CmpVal & SignBit) != (CmpSSInt & SignBit))
1938 // Look for a suitable stride / iv as replacement.
1939 for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
1940 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
1941 IU->IVUsesByStride.find(IU->StrideOrder[i]);
1942 if (!isa<SCEVConstant>(SI->first) || SI->second->Users.empty())
1944 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1945 if (SSInt == CmpSSInt ||
1946 abs64(SSInt) < abs64(CmpSSInt) ||
1947 (SSInt % CmpSSInt) != 0)
1950 Scale = SSInt / CmpSSInt;
1951 int64_t NewCmpVal = CmpVal * Scale;
1953 // If old icmp value fits in icmp immediate field, but the new one doesn't
1954 // try something else.
1956 TLI->isLegalICmpImmediate(CmpVal) &&
1957 !TLI->isLegalICmpImmediate(NewCmpVal))
1960 APInt Mul = APInt(BitWidth*2, CmpVal, true);
1961 Mul = Mul * APInt(BitWidth*2, Scale, true);
1962 // Check for overflow.
1963 if (!Mul.isSignedIntN(BitWidth))
1965 // Check for overflow in the stride's type too.
1966 if (!Mul.isSignedIntN(SE->getTypeSizeInBits(SI->first->getType())))
1969 // Watch out for overflow.
1970 if (ICmpInst::isSigned(Predicate) &&
1971 (CmpVal & SignBit) != (NewCmpVal & SignBit))
1974 // Pick the best iv to use trying to avoid a cast.
1976 for (ilist<IVStrideUse>::iterator UI = SI->second->Users.begin(),
1977 E = SI->second->Users.end(); UI != E; ++UI) {
1978 Value *Op = UI->getOperandValToReplace();
1980 // If the IVStrideUse implies a cast, check for an actual cast which
1981 // can be used to find the original IV expression.
1982 if (SE->getEffectiveSCEVType(Op->getType()) !=
1983 SE->getEffectiveSCEVType(SI->first->getType())) {
1984 CastInst *CI = dyn_cast<CastInst>(Op);
1985 // If it's not a simple cast, it's complicated.
1988 // If it's a cast from a type other than the stride type,
1989 // it's complicated.
1990 if (CI->getOperand(0)->getType() != SI->first->getType())
1992 // Ok, we found the IV expression in the stride's type.
1993 Op = CI->getOperand(0);
1997 if (NewCmpLHS->getType() == CmpTy)
2003 NewCmpTy = NewCmpLHS->getType();
2004 NewTyBits = SE->getTypeSizeInBits(NewCmpTy);
2005 const Type *NewCmpIntTy = IntegerType::get(Cond->getContext(), NewTyBits);
2006 if (RequiresTypeConversion(NewCmpTy, CmpTy)) {
2007 // Check if it is possible to rewrite it using
2008 // an iv / stride of a smaller integer type.
2009 unsigned Bits = NewTyBits;
2010 if (ICmpInst::isSigned(Predicate))
2012 uint64_t Mask = (1ULL << Bits) - 1;
2013 if (((uint64_t)NewCmpVal & Mask) != (uint64_t)NewCmpVal)
2017 // Don't rewrite if use offset is non-constant and the new type is
2018 // of a different type.
2019 // FIXME: too conservative?
2020 if (NewTyBits != TyBits && !isa<SCEVConstant>(CondUse->getOffset()))
2024 bool AllUsesAreAddresses = true;
2025 bool AllUsesAreOutsideLoop = true;
2026 std::vector<BasedUser> UsersToProcess;
2027 const SCEV *CommonExprs = CollectIVUsers(SI->first, *SI->second, L,
2028 AllUsesAreAddresses,
2029 AllUsesAreOutsideLoop,
2031 // Avoid rewriting the compare instruction with an iv of new stride
2032 // if it's likely the new stride uses will be rewritten using the
2033 // stride of the compare instruction.
2034 if (AllUsesAreAddresses &&
2035 ValidScale(!CommonExprs->isZero(), Scale, UsersToProcess))
2039 // Avoid rewriting the compare instruction with an iv which has
2040 // implicit extension or truncation built into it.
2041 // TODO: This is over-conservative.
2042 if (SE->getTypeSizeInBits(CondUse->getOffset()->getType()) != TyBits)
2045 // If scale is negative, use swapped predicate unless it's testing
2047 if (Scale < 0 && !Cond->isEquality())
2048 Predicate = ICmpInst::getSwappedPredicate(Predicate);
2050 NewStride = IU->StrideOrder[i];
2051 if (!isa<PointerType>(NewCmpTy))
2052 NewCmpRHS = ConstantInt::get(NewCmpTy, NewCmpVal);
2054 Constant *CI = ConstantInt::get(NewCmpIntTy, NewCmpVal);
2055 NewCmpRHS = ConstantExpr::getIntToPtr(CI, NewCmpTy);
2057 NewOffset = TyBits == NewTyBits
2058 ? SE->getMulExpr(CondUse->getOffset(),
2059 SE->getConstant(CmpTy, Scale))
2060 : SE->getConstant(NewCmpIntTy,
2061 cast<SCEVConstant>(CondUse->getOffset())->getValue()
2062 ->getSExtValue()*Scale);
2067 // Forgo this transformation if it the increment happens to be
2068 // unfortunately positioned after the condition, and the condition
2069 // has multiple uses which prevent it from being moved immediately
2070 // before the branch. See
2071 // test/Transforms/LoopStrengthReduce/change-compare-stride-trickiness-*.ll
2072 // for an example of this situation.
2073 if (!Cond->hasOneUse()) {
2074 for (BasicBlock::iterator I = Cond, E = Cond->getParent()->end();
2081 // Create a new compare instruction using new stride / iv.
2082 ICmpInst *OldCond = Cond;
2083 // Insert new compare instruction.
2084 Cond = new ICmpInst(OldCond, Predicate, NewCmpLHS, NewCmpRHS,
2085 L->getHeader()->getName() + ".termcond");
2087 DEBUG(errs() << " Change compare stride in Inst " << *OldCond);
2088 DEBUG(errs() << " to " << *Cond << '\n');
2090 // Remove the old compare instruction. The old indvar is probably dead too.
2091 DeadInsts.push_back(CondUse->getOperandValToReplace());
2092 OldCond->replaceAllUsesWith(Cond);
2093 OldCond->eraseFromParent();
2095 IU->IVUsesByStride[NewStride]->addUser(NewOffset, Cond, NewCmpLHS);
2096 CondUse = &IU->IVUsesByStride[NewStride]->Users.back();
2097 CondStride = NewStride;
2105 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
2106 /// a max computation.
2108 /// This is a narrow solution to a specific, but acute, problem. For loops
2114 /// } while (++i < n);
2116 /// the trip count isn't just 'n', because 'n' might not be positive. And
2117 /// unfortunately this can come up even for loops where the user didn't use
2118 /// a C do-while loop. For example, seemingly well-behaved top-test loops
2119 /// will commonly be lowered like this:
2125 /// } while (++i < n);
2128 /// and then it's possible for subsequent optimization to obscure the if
2129 /// test in such a way that indvars can't find it.
2131 /// When indvars can't find the if test in loops like this, it creates a
2132 /// max expression, which allows it to give the loop a canonical
2133 /// induction variable:
2136 /// max = n < 1 ? 1 : n;
2139 /// } while (++i != max);
2141 /// Canonical induction variables are necessary because the loop passes
2142 /// are designed around them. The most obvious example of this is the
2143 /// LoopInfo analysis, which doesn't remember trip count values. It
2144 /// expects to be able to rediscover the trip count each time it is
2145 /// needed, and it does this using a simple analyis that only succeeds if
2146 /// the loop has a canonical induction variable.
2148 /// However, when it comes time to generate code, the maximum operation
2149 /// can be quite costly, especially if it's inside of an outer loop.
2151 /// This function solves this problem by detecting this type of loop and
2152 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
2153 /// the instructions for the maximum computation.
2155 ICmpInst *LoopStrengthReduce::OptimizeMax(Loop *L, ICmpInst *Cond,
2156 IVStrideUse* &CondUse) {
2157 // Check that the loop matches the pattern we're looking for.
2158 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
2159 Cond->getPredicate() != CmpInst::ICMP_NE)
2162 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
2163 if (!Sel || !Sel->hasOneUse()) return Cond;
2165 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
2166 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2168 const SCEV *One = SE->getIntegerSCEV(1, BackedgeTakenCount->getType());
2170 // Add one to the backedge-taken count to get the trip count.
2171 const SCEV *IterationCount = SE->getAddExpr(BackedgeTakenCount, One);
2173 // Check for a max calculation that matches the pattern.
2174 if (!isa<SCEVSMaxExpr>(IterationCount) && !isa<SCEVUMaxExpr>(IterationCount))
2176 const SCEVNAryExpr *Max = cast<SCEVNAryExpr>(IterationCount);
2177 if (Max != SE->getSCEV(Sel)) return Cond;
2179 // To handle a max with more than two operands, this optimization would
2180 // require additional checking and setup.
2181 if (Max->getNumOperands() != 2)
2184 const SCEV *MaxLHS = Max->getOperand(0);
2185 const SCEV *MaxRHS = Max->getOperand(1);
2186 if (!MaxLHS || MaxLHS != One) return Cond;
2188 // Check the relevant induction variable for conformance to
2190 const SCEV *IV = SE->getSCEV(Cond->getOperand(0));
2191 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2192 if (!AR || !AR->isAffine() ||
2193 AR->getStart() != One ||
2194 AR->getStepRecurrence(*SE) != One)
2197 assert(AR->getLoop() == L &&
2198 "Loop condition operand is an addrec in a different loop!");
2200 // Check the right operand of the select, and remember it, as it will
2201 // be used in the new comparison instruction.
2203 if (SE->getSCEV(Sel->getOperand(1)) == MaxRHS)
2204 NewRHS = Sel->getOperand(1);
2205 else if (SE->getSCEV(Sel->getOperand(2)) == MaxRHS)
2206 NewRHS = Sel->getOperand(2);
2207 if (!NewRHS) return Cond;
2209 // Determine the new comparison opcode. It may be signed or unsigned,
2210 // and the original comparison may be either equality or inequality.
2211 CmpInst::Predicate Pred =
2212 isa<SCEVSMaxExpr>(Max) ? CmpInst::ICMP_SLT : CmpInst::ICMP_ULT;
2213 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
2214 Pred = CmpInst::getInversePredicate(Pred);
2216 // Ok, everything looks ok to change the condition into an SLT or SGE and
2217 // delete the max calculation.
2219 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
2221 // Delete the max calculation instructions.
2222 Cond->replaceAllUsesWith(NewCond);
2223 CondUse->setUser(NewCond);
2224 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2225 Cond->eraseFromParent();
2226 Sel->eraseFromParent();
2227 if (Cmp->use_empty())
2228 Cmp->eraseFromParent();
2232 /// OptimizeShadowIV - If IV is used in a int-to-float cast
2233 /// inside the loop then try to eliminate the cast opeation.
2234 void LoopStrengthReduce::OptimizeShadowIV(Loop *L) {
2236 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
2237 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2240 for (unsigned Stride = 0, e = IU->StrideOrder.size(); Stride != e;
2242 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
2243 IU->IVUsesByStride.find(IU->StrideOrder[Stride]);
2244 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
2245 if (!isa<SCEVConstant>(SI->first))
2248 for (ilist<IVStrideUse>::iterator UI = SI->second->Users.begin(),
2249 E = SI->second->Users.end(); UI != E; /* empty */) {
2250 ilist<IVStrideUse>::iterator CandidateUI = UI;
2252 Instruction *ShadowUse = CandidateUI->getUser();
2253 const Type *DestTy = NULL;
2255 /* If shadow use is a int->float cast then insert a second IV
2256 to eliminate this cast.
2258 for (unsigned i = 0; i < n; ++i)
2264 for (unsigned i = 0; i < n; ++i, ++d)
2267 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser()))
2268 DestTy = UCast->getDestTy();
2269 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser()))
2270 DestTy = SCast->getDestTy();
2271 if (!DestTy) continue;
2274 // If target does not support DestTy natively then do not apply
2275 // this transformation.
2276 EVT DVT = TLI->getValueType(DestTy);
2277 if (!TLI->isTypeLegal(DVT)) continue;
2280 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
2282 if (PH->getNumIncomingValues() != 2) continue;
2284 const Type *SrcTy = PH->getType();
2285 int Mantissa = DestTy->getFPMantissaWidth();
2286 if (Mantissa == -1) continue;
2287 if ((int)SE->getTypeSizeInBits(SrcTy) > Mantissa)
2290 unsigned Entry, Latch;
2291 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
2299 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
2300 if (!Init) continue;
2301 Constant *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());
2303 BinaryOperator *Incr =
2304 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
2305 if (!Incr) continue;
2306 if (Incr->getOpcode() != Instruction::Add
2307 && Incr->getOpcode() != Instruction::Sub)
2310 /* Initialize new IV, double d = 0.0 in above example. */
2311 ConstantInt *C = NULL;
2312 if (Incr->getOperand(0) == PH)
2313 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
2314 else if (Incr->getOperand(1) == PH)
2315 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
2321 // Ignore negative constants, as the code below doesn't handle them
2322 // correctly. TODO: Remove this restriction.
2323 if (!C->getValue().isStrictlyPositive()) continue;
2325 /* Add new PHINode. */
2326 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);
2328 /* create new increment. '++d' in above example. */
2329 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
2330 BinaryOperator *NewIncr =
2331 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
2332 Instruction::FAdd : Instruction::FSub,
2333 NewPH, CFP, "IV.S.next.", Incr);
2335 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
2336 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
2338 /* Remove cast operation */
2339 ShadowUse->replaceAllUsesWith(NewPH);
2340 ShadowUse->eraseFromParent();
2347 /// OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
2348 /// uses in the loop, look to see if we can eliminate some, in favor of using
2349 /// common indvars for the different uses.
2350 void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
2351 // TODO: implement optzns here.
2353 OptimizeShadowIV(L);
2356 bool LoopStrengthReduce::StrideMightBeShared(const SCEV* Stride, Loop *L,
2358 int64_t SInt = cast<SCEVConstant>(Stride)->getValue()->getSExtValue();
2359 for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
2360 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
2361 IU->IVUsesByStride.find(IU->StrideOrder[i]);
2362 const SCEV *Share = SI->first;
2363 if (!isa<SCEVConstant>(SI->first) || Share == Stride)
2365 int64_t SSInt = cast<SCEVConstant>(Share)->getValue()->getSExtValue();
2367 return true; // This can definitely be reused.
2368 if (unsigned(abs64(SSInt)) < SInt || (SSInt % SInt) != 0)
2370 int64_t Scale = SSInt / SInt;
2371 bool AllUsesAreAddresses = true;
2372 bool AllUsesAreOutsideLoop = true;
2373 std::vector<BasedUser> UsersToProcess;
2374 const SCEV *CommonExprs = CollectIVUsers(SI->first, *SI->second, L,
2375 AllUsesAreAddresses,
2376 AllUsesAreOutsideLoop,
2378 if (AllUsesAreAddresses &&
2379 ValidScale(!CommonExprs->isZero(), Scale, UsersToProcess)) {
2382 // Any pre-inc iv use?
2383 IVUsersOfOneStride &StrideUses = *IU->IVUsesByStride[Share];
2384 for (ilist<IVStrideUse>::iterator I = StrideUses.Users.begin(),
2385 E = StrideUses.Users.end(); I != E; ++I) {
2386 if (!I->isUseOfPostIncrementedValue())
2394 /// isUsedByExitBranch - Return true if icmp is used by a loop terminating
2395 /// conditional branch or it's and / or with other conditions before being used
2396 /// as the condition.
2397 static bool isUsedByExitBranch(ICmpInst *Cond, Loop *L) {
2398 BasicBlock *CondBB = Cond->getParent();
2399 if (!L->isLoopExiting(CondBB))
2401 BranchInst *TermBr = dyn_cast<BranchInst>(CondBB->getTerminator());
2402 if (!TermBr || !TermBr->isConditional())
2405 Value *User = *Cond->use_begin();
2406 Instruction *UserInst = dyn_cast<Instruction>(User);
2408 (UserInst->getOpcode() == Instruction::And ||
2409 UserInst->getOpcode() == Instruction::Or)) {
2410 if (!UserInst->hasOneUse() || UserInst->getParent() != CondBB)
2412 User = *User->use_begin();
2413 UserInst = dyn_cast<Instruction>(User);
2415 return User == TermBr;
2418 static bool ShouldCountToZero(ICmpInst *Cond, IVStrideUse* &CondUse,
2419 ScalarEvolution *SE, Loop *L,
2420 const TargetLowering *TLI = 0) {
2421 if (!L->contains(Cond->getParent()))
2424 if (!isa<SCEVConstant>(CondUse->getOffset()))
2427 // Handle only tests for equality for the moment.
2428 if (!Cond->isEquality() || !Cond->hasOneUse())
2430 if (!isUsedByExitBranch(Cond, L))
2433 Value *CondOp0 = Cond->getOperand(0);
2434 const SCEV *IV = SE->getSCEV(CondOp0);
2435 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2436 if (!AR || !AR->isAffine())
2439 const SCEVConstant *SC = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE));
2440 if (!SC || SC->getValue()->getSExtValue() < 0)
2441 // If it's already counting down, don't do anything.
2444 // If the RHS of the comparison is not an loop invariant, the rewrite
2445 // cannot be done. Also bail out if it's already comparing against a zero.
2446 // If we are checking this before cmp stride optimization, check if it's
2447 // comparing against a already legal immediate.
2448 Value *RHS = Cond->getOperand(1);
2449 ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS);
2450 if (!L->isLoopInvariant(RHS) ||
2451 (RHSC && RHSC->isZero()) ||
2452 (RHSC && TLI && TLI->isLegalICmpImmediate(RHSC->getSExtValue())))
2455 // Make sure the IV is only used for counting. Value may be preinc or
2456 // postinc; 2 uses in either case.
2457 if (!CondOp0->hasNUses(2))
2463 /// OptimizeLoopTermCond - Change loop terminating condition to use the
2464 /// postinc iv when possible.
2465 void LoopStrengthReduce::OptimizeLoopTermCond(Loop *L) {
2466 BasicBlock *LatchBlock = L->getLoopLatch();
2467 bool LatchExit = L->isLoopExiting(LatchBlock);
2468 SmallVector<BasicBlock*, 8> ExitingBlocks;
2469 L->getExitingBlocks(ExitingBlocks);
2471 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
2472 BasicBlock *ExitingBlock = ExitingBlocks[i];
2474 // Finally, get the terminating condition for the loop if possible. If we
2475 // can, we want to change it to use a post-incremented version of its
2476 // induction variable, to allow coalescing the live ranges for the IV into
2477 // one register value.
2479 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2482 // FIXME: Overly conservative, termination condition could be an 'or' etc..
2483 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
2486 // Search IVUsesByStride to find Cond's IVUse if there is one.
2487 IVStrideUse *CondUse = 0;
2488 const SCEV *CondStride = 0;
2489 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2490 if (!FindIVUserForCond(Cond, CondUse, CondStride))
2493 // If the latch block is exiting and it's not a single block loop, it's
2494 // not safe to use postinc iv in other exiting blocks. FIXME: overly
2495 // conservative? How about icmp stride optimization?
2496 bool UsePostInc = !(e > 1 && LatchExit && ExitingBlock != LatchBlock);
2497 if (UsePostInc && ExitingBlock != LatchBlock) {
2498 if (!Cond->hasOneUse())
2499 // See below, we don't want the condition to be cloned.
2502 // If exiting block is the latch block, we know it's safe and profitable
2503 // to transform the icmp to use post-inc iv. Otherwise do so only if it
2504 // would not reuse another iv and its iv would be reused by other uses.
2505 // We are optimizing for the case where the icmp is the only use of the
2507 IVUsersOfOneStride &StrideUses = *IU->IVUsesByStride[CondStride];
2508 for (ilist<IVStrideUse>::iterator I = StrideUses.Users.begin(),
2509 E = StrideUses.Users.end(); I != E; ++I) {
2510 if (I->getUser() == Cond)
2512 if (!I->isUseOfPostIncrementedValue()) {
2519 // If iv for the stride might be shared and any of the users use pre-inc
2520 // iv might be used, then it's not safe to use post-inc iv.
2522 isa<SCEVConstant>(CondStride) &&
2523 StrideMightBeShared(CondStride, L, true))
2527 // If the trip count is computed in terms of a max (due to ScalarEvolution
2528 // being unable to find a sufficient guard, for example), change the loop
2529 // comparison to use SLT or ULT instead of NE.
2530 Cond = OptimizeMax(L, Cond, CondUse);
2532 // If possible, change stride and operands of the compare instruction to
2533 // eliminate one stride. However, avoid rewriting the compare instruction
2534 // with an iv of new stride if it's likely the new stride uses will be
2535 // rewritten using the stride of the compare instruction.
2536 if (ExitingBlock == LatchBlock && isa<SCEVConstant>(CondStride)) {
2537 // If the condition stride is a constant and it's the only use, we might
2538 // want to optimize it first by turning it to count toward zero.
2539 if (!StrideMightBeShared(CondStride, L, false) &&
2540 !ShouldCountToZero(Cond, CondUse, SE, L, TLI))
2541 Cond = ChangeCompareStride(L, Cond, CondUse, CondStride);
2547 DEBUG(errs() << " Change loop exiting icmp to use postinc iv: "
2550 // It's possible for the setcc instruction to be anywhere in the loop, and
2551 // possible for it to have multiple users. If it is not immediately before
2552 // the exiting block branch, move it.
2553 if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
2554 if (Cond->hasOneUse()) { // Condition has a single use, just move it.
2555 Cond->moveBefore(TermBr);
2557 // Otherwise, clone the terminating condition and insert into the
2559 Cond = cast<ICmpInst>(Cond->clone());
2560 Cond->setName(L->getHeader()->getName() + ".termcond");
2561 ExitingBlock->getInstList().insert(TermBr, Cond);
2563 // Clone the IVUse, as the old use still exists!
2564 IU->IVUsesByStride[CondStride]->addUser(CondUse->getOffset(), Cond,
2565 CondUse->getOperandValToReplace());
2566 CondUse = &IU->IVUsesByStride[CondStride]->Users.back();
2570 // If we get to here, we know that we can transform the setcc instruction to
2571 // use the post-incremented version of the IV, allowing us to coalesce the
2572 // live ranges for the IV correctly.
2573 CondUse->setOffset(SE->getMinusSCEV(CondUse->getOffset(), CondStride));
2574 CondUse->setIsUseOfPostIncrementedValue(true);
2581 bool LoopStrengthReduce::OptimizeLoopCountIVOfStride(const SCEV* &Stride,
2582 IVStrideUse* &CondUse,
2584 // If the only use is an icmp of a loop exiting conditional branch, then
2585 // attempt the optimization.
2586 BasedUser User = BasedUser(*CondUse, SE);
2587 assert(isa<ICmpInst>(User.Inst) && "Expecting an ICMPInst!");
2588 ICmpInst *Cond = cast<ICmpInst>(User.Inst);
2590 // Less strict check now that compare stride optimization is done.
2591 if (!ShouldCountToZero(Cond, CondUse, SE, L))
2594 Value *CondOp0 = Cond->getOperand(0);
2595 PHINode *PHIExpr = dyn_cast<PHINode>(CondOp0);
2598 // Value tested is postinc. Find the phi node.
2599 Incr = dyn_cast<BinaryOperator>(CondOp0);
2600 // FIXME: Just use User.OperandValToReplace here?
2601 if (!Incr || Incr->getOpcode() != Instruction::Add)
2604 PHIExpr = dyn_cast<PHINode>(Incr->getOperand(0));
2607 // 1 use for preinc value, the increment.
2608 if (!PHIExpr->hasOneUse())
2611 assert(isa<PHINode>(CondOp0) &&
2612 "Unexpected loop exiting counting instruction sequence!");
2613 PHIExpr = cast<PHINode>(CondOp0);
2614 // Value tested is preinc. Find the increment.
2615 // A CmpInst is not a BinaryOperator; we depend on this.
2616 Instruction::use_iterator UI = PHIExpr->use_begin();
2617 Incr = dyn_cast<BinaryOperator>(UI);
2619 Incr = dyn_cast<BinaryOperator>(++UI);
2620 // One use for postinc value, the phi. Unnecessarily conservative?
2621 if (!Incr || !Incr->hasOneUse() || Incr->getOpcode() != Instruction::Add)
2625 // Replace the increment with a decrement.
2626 DEBUG(errs() << "LSR: Examining use ");
2627 DEBUG(WriteAsOperand(errs(), CondOp0, /*PrintType=*/false));
2628 DEBUG(errs() << " in Inst: " << *Cond << '\n');
2629 BinaryOperator *Decr = BinaryOperator::Create(Instruction::Sub,
2630 Incr->getOperand(0), Incr->getOperand(1), "tmp", Incr);
2631 Incr->replaceAllUsesWith(Decr);
2632 Incr->eraseFromParent();
2634 // Substitute endval-startval for the original startval, and 0 for the
2635 // original endval. Since we're only testing for equality this is OK even
2636 // if the computation wraps around.
2637 BasicBlock *Preheader = L->getLoopPreheader();
2638 Instruction *PreInsertPt = Preheader->getTerminator();
2639 unsigned InBlock = L->contains(PHIExpr->getIncomingBlock(0)) ? 1 : 0;
2640 Value *StartVal = PHIExpr->getIncomingValue(InBlock);
2641 Value *EndVal = Cond->getOperand(1);
2642 DEBUG(errs() << " Optimize loop counting iv to count down ["
2643 << *EndVal << " .. " << *StartVal << "]\n");
2645 // FIXME: check for case where both are constant.
2646 Constant* Zero = ConstantInt::get(Cond->getOperand(1)->getType(), 0);
2647 BinaryOperator *NewStartVal = BinaryOperator::Create(Instruction::Sub,
2648 EndVal, StartVal, "tmp", PreInsertPt);
2649 PHIExpr->setIncomingValue(InBlock, NewStartVal);
2650 Cond->setOperand(1, Zero);
2651 DEBUG(errs() << " New icmp: " << *Cond << "\n");
2653 int64_t SInt = cast<SCEVConstant>(Stride)->getValue()->getSExtValue();
2654 const SCEV *NewStride = 0;
2656 for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
2657 const SCEV *OldStride = IU->StrideOrder[i];
2658 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(OldStride))
2659 if (SC->getValue()->getSExtValue() == -SInt) {
2661 NewStride = OldStride;
2667 NewStride = SE->getIntegerSCEV(-SInt, Stride->getType());
2668 IU->AddUser(NewStride, CondUse->getOffset(), Cond, Cond->getOperand(0));
2669 IU->IVUsesByStride[Stride]->removeUser(CondUse);
2671 CondUse = &IU->IVUsesByStride[NewStride]->Users.back();
2679 /// OptimizeLoopCountIV - If, after all sharing of IVs, the IV used for deciding
2680 /// when to exit the loop is used only for that purpose, try to rearrange things
2681 /// so it counts down to a test against zero.
2682 bool LoopStrengthReduce::OptimizeLoopCountIV(Loop *L) {
2683 bool ThisChanged = false;
2684 for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
2685 const SCEV *Stride = IU->StrideOrder[i];
2686 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
2687 IU->IVUsesByStride.find(Stride);
2688 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
2689 // FIXME: Generalize to non-affine IV's.
2690 if (!SI->first->isLoopInvariant(L))
2692 // If stride is a constant and it has an icmpinst use, check if we can
2693 // optimize the loop to count down.
2694 if (isa<SCEVConstant>(Stride) && SI->second->Users.size() == 1) {
2695 Instruction *User = SI->second->Users.begin()->getUser();
2696 if (!isa<ICmpInst>(User))
2698 const SCEV *CondStride = Stride;
2699 IVStrideUse *Use = &*SI->second->Users.begin();
2700 if (!OptimizeLoopCountIVOfStride(CondStride, Use, L))
2704 // Now check if it's possible to reuse this iv for other stride uses.
2705 for (unsigned j = 0, ee = IU->StrideOrder.size(); j != ee; ++j) {
2706 const SCEV *SStride = IU->StrideOrder[j];
2707 if (SStride == CondStride)
2709 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SII =
2710 IU->IVUsesByStride.find(SStride);
2711 assert(SII != IU->IVUsesByStride.end() && "Stride doesn't exist!");
2712 // FIXME: Generalize to non-affine IV's.
2713 if (!SII->first->isLoopInvariant(L))
2715 // FIXME: Rewrite other stride using CondStride.
2720 Changed |= ThisChanged;
2724 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager &LPM) {
2725 IU = &getAnalysis<IVUsers>();
2726 LI = &getAnalysis<LoopInfo>();
2727 DT = &getAnalysis<DominatorTree>();
2728 SE = &getAnalysis<ScalarEvolution>();
2731 // If LoopSimplify form is not available, stay out of trouble.
2732 if (!L->getLoopPreheader() || !L->getLoopLatch())
2735 if (!IU->IVUsesByStride.empty()) {
2736 DEBUG(errs() << "\nLSR on \"" << L->getHeader()->getParent()->getName()
2740 // Sort the StrideOrder so we process larger strides first.
2741 std::stable_sort(IU->StrideOrder.begin(), IU->StrideOrder.end(),
2744 // Optimize induction variables. Some indvar uses can be transformed to use
2745 // strides that will be needed for other purposes. A common example of this
2746 // is the exit test for the loop, which can often be rewritten to use the
2747 // computation of some other indvar to decide when to terminate the loop.
2750 // Change loop terminating condition to use the postinc iv when possible
2751 // and optimize loop terminating compare. FIXME: Move this after
2752 // StrengthReduceIVUsersOfStride?
2753 OptimizeLoopTermCond(L);
2755 // FIXME: We can shrink overlarge IV's here. e.g. if the code has
2756 // computation in i64 values and the target doesn't support i64, demote
2757 // the computation to 32-bit if safe.
2759 // FIXME: Attempt to reuse values across multiple IV's. In particular, we
2760 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should
2761 // be codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC.
2762 // Need to be careful that IV's are all the same type. Only works for
2763 // intptr_t indvars.
2765 // IVsByStride keeps IVs for one particular loop.
2766 assert(IVsByStride.empty() && "Stale entries in IVsByStride?");
2768 StrengthReduceIVUsers(L);
2770 // After all sharing is done, see if we can adjust the loop to test against
2771 // zero instead of counting up to a maximum. This is usually faster.
2772 OptimizeLoopCountIV(L);
2775 // We're done analyzing this loop; release all the state we built up for it.
2776 IVsByStride.clear();
2777 StrideNoReuse.clear();
2779 // Clean up after ourselves
2780 if (!DeadInsts.empty())
2781 DeleteTriviallyDeadInstructions();
2783 // At this point, it is worth checking to see if any recurrence PHIs are also
2784 // dead, so that we can remove them as well.
2785 DeleteDeadPHIs(L->getHeader());