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/DerivedTypes.h"
28 #include "llvm/Analysis/IVUsers.h"
29 #include "llvm/Analysis/LoopPass.h"
30 #include "llvm/Analysis/ScalarEvolutionExpander.h"
31 #include "llvm/Transforms/Utils/AddrModeMatcher.h"
32 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
33 #include "llvm/Transforms/Utils/Local.h"
34 #include "llvm/ADT/Statistic.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/CommandLine.h"
37 #include "llvm/Support/ValueHandle.h"
38 #include "llvm/Support/raw_ostream.h"
39 #include "llvm/Target/TargetLowering.h"
43 STATISTIC(NumReduced , "Number of IV uses strength reduced");
44 STATISTIC(NumInserted, "Number of PHIs inserted");
45 STATISTIC(NumVariable, "Number of PHIs with variable strides");
46 STATISTIC(NumEliminated, "Number of strides eliminated");
47 STATISTIC(NumShadow, "Number of Shadow IVs optimized");
48 STATISTIC(NumImmSunk, "Number of common expr immediates sunk into uses");
49 STATISTIC(NumLoopCond, "Number of loop terminating conds optimized");
50 STATISTIC(NumCountZero, "Number of count iv optimized to count toward zero");
52 static cl::opt<bool> EnableFullLSRMode("enable-full-lsr",
60 /// IVInfo - This structure keeps track of one IV expression inserted during
61 /// StrengthReduceStridedIVUsers. It contains the stride, the common base, as
62 /// well as the PHI node and increment value created for rewrite.
68 IVExpr(const SCEV *const stride, const SCEV *const base, PHINode *phi)
69 : Stride(stride), Base(base), PHI(phi) {}
72 /// IVsOfOneStride - This structure keeps track of all IV expression inserted
73 /// during StrengthReduceStridedIVUsers for a particular stride of the IV.
74 struct IVsOfOneStride {
75 std::vector<IVExpr> IVs;
77 void addIV(const SCEV *const Stride, const SCEV *const Base, PHINode *PHI) {
78 IVs.push_back(IVExpr(Stride, Base, PHI));
82 class LoopStrengthReduce : public LoopPass {
87 /// IVsByStride - Keep track of all IVs that have been inserted for a
88 /// particular stride.
89 std::map<const SCEV *, IVsOfOneStride> IVsByStride;
91 /// DeadInsts - Keep track of instructions we may have made dead, so that
92 /// we can remove them after we are done working.
93 SmallVector<WeakVH, 16> DeadInsts;
95 /// TLI - Keep a pointer of a TargetLowering to consult for determining
96 /// transformation profitability.
97 const TargetLowering *TLI;
100 static char ID; // Pass ID, replacement for typeid
101 explicit LoopStrengthReduce(const TargetLowering *tli = NULL) :
102 LoopPass(&ID), TLI(tli) {}
104 bool runOnLoop(Loop *L, LPPassManager &LPM);
106 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
107 // We split critical edges, so we change the CFG. However, we do update
108 // many analyses if they are around.
109 AU.addPreservedID(LoopSimplifyID);
110 AU.addPreserved("loops");
111 AU.addPreserved("domfrontier");
112 AU.addPreserved("domtree");
114 AU.addRequiredID(LoopSimplifyID);
115 AU.addRequired<ScalarEvolution>();
116 AU.addPreserved<ScalarEvolution>();
117 AU.addRequired<IVUsers>();
118 AU.addPreserved<IVUsers>();
122 void OptimizeIndvars(Loop *L);
124 /// OptimizeLoopTermCond - Change loop terminating condition to use the
125 /// postinc iv when possible.
126 void OptimizeLoopTermCond(Loop *L);
128 /// OptimizeShadowIV - If IV is used in a int-to-float cast
129 /// inside the loop then try to eliminate the cast opeation.
130 void OptimizeShadowIV(Loop *L);
132 /// OptimizeMax - Rewrite the loop's terminating condition
133 /// if it uses a max computation.
134 ICmpInst *OptimizeMax(Loop *L, ICmpInst *Cond,
135 IVStrideUse* &CondUse);
137 /// OptimizeLoopCountIV - If, after all sharing of IVs, the IV used for
138 /// deciding when to exit the loop is used only for that purpose, try to
139 /// rearrange things so it counts down to a test against zero.
140 bool OptimizeLoopCountIV(Loop *L);
141 bool OptimizeLoopCountIVOfStride(const SCEV* &Stride,
142 IVStrideUse* &CondUse, Loop *L);
144 /// StrengthReduceIVUsersOfStride - Strength reduce all of the users of a
145 /// single stride of IV. All of the users may have different starting
146 /// values, and this may not be the only stride.
147 void StrengthReduceIVUsersOfStride(const SCEV *Stride,
148 IVUsersOfOneStride &Uses,
150 void StrengthReduceIVUsers(Loop *L);
152 ICmpInst *ChangeCompareStride(Loop *L, ICmpInst *Cond,
153 IVStrideUse* &CondUse,
154 const SCEV* &CondStride,
155 bool PostPass = false);
157 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse,
158 const SCEV* &CondStride);
159 bool RequiresTypeConversion(const Type *Ty, const Type *NewTy);
160 const SCEV *CheckForIVReuse(bool, bool, bool, const SCEV *,
161 IVExpr&, const Type*,
162 const std::vector<BasedUser>& UsersToProcess);
163 bool ValidScale(bool, int64_t,
164 const std::vector<BasedUser>& UsersToProcess);
165 bool ValidOffset(bool, int64_t, int64_t,
166 const std::vector<BasedUser>& UsersToProcess);
167 const SCEV *CollectIVUsers(const SCEV *Stride,
168 IVUsersOfOneStride &Uses,
170 bool &AllUsesAreAddresses,
171 bool &AllUsesAreOutsideLoop,
172 std::vector<BasedUser> &UsersToProcess);
173 bool StrideMightBeShared(const SCEV *Stride, Loop *L, bool CheckPreInc);
174 bool ShouldUseFullStrengthReductionMode(
175 const std::vector<BasedUser> &UsersToProcess,
177 bool AllUsesAreAddresses,
179 void PrepareToStrengthReduceFully(
180 std::vector<BasedUser> &UsersToProcess,
182 const SCEV *CommonExprs,
184 SCEVExpander &PreheaderRewriter);
185 void PrepareToStrengthReduceFromSmallerStride(
186 std::vector<BasedUser> &UsersToProcess,
188 const IVExpr &ReuseIV,
189 Instruction *PreInsertPt);
190 void PrepareToStrengthReduceWithNewPhi(
191 std::vector<BasedUser> &UsersToProcess,
193 const SCEV *CommonExprs,
195 Instruction *IVIncInsertPt,
197 SCEVExpander &PreheaderRewriter);
199 void DeleteTriviallyDeadInstructions();
203 char LoopStrengthReduce::ID = 0;
204 static RegisterPass<LoopStrengthReduce>
205 X("loop-reduce", "Loop Strength Reduction");
207 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
208 return new LoopStrengthReduce(TLI);
211 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
212 /// specified set are trivially dead, delete them and see if this makes any of
213 /// their operands subsequently dead.
214 void LoopStrengthReduce::DeleteTriviallyDeadInstructions() {
215 while (!DeadInsts.empty()) {
216 Instruction *I = dyn_cast_or_null<Instruction>(DeadInsts.pop_back_val());
218 if (I == 0 || !isInstructionTriviallyDead(I))
221 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
222 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
225 DeadInsts.push_back(U);
228 I->eraseFromParent();
233 /// isAddressUse - Returns true if the specified instruction is using the
234 /// specified value as an address.
235 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
236 bool isAddress = isa<LoadInst>(Inst);
237 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
238 if (SI->getOperand(1) == OperandVal)
240 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
241 // Addressing modes can also be folded into prefetches and a variety
243 switch (II->getIntrinsicID()) {
245 case Intrinsic::prefetch:
246 case Intrinsic::x86_sse2_loadu_dq:
247 case Intrinsic::x86_sse2_loadu_pd:
248 case Intrinsic::x86_sse_loadu_ps:
249 case Intrinsic::x86_sse_storeu_ps:
250 case Intrinsic::x86_sse2_storeu_pd:
251 case Intrinsic::x86_sse2_storeu_dq:
252 case Intrinsic::x86_sse2_storel_dq:
253 if (II->getOperand(1) == OperandVal)
261 /// getAccessType - Return the type of the memory being accessed.
262 static const Type *getAccessType(const Instruction *Inst) {
263 const Type *AccessTy = Inst->getType();
264 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
265 AccessTy = SI->getOperand(0)->getType();
266 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
267 // Addressing modes can also be folded into prefetches and a variety
269 switch (II->getIntrinsicID()) {
271 case Intrinsic::x86_sse_storeu_ps:
272 case Intrinsic::x86_sse2_storeu_pd:
273 case Intrinsic::x86_sse2_storeu_dq:
274 case Intrinsic::x86_sse2_storel_dq:
275 AccessTy = II->getOperand(1)->getType();
283 /// BasedUser - For a particular base value, keep information about how we've
284 /// partitioned the expression so far.
286 /// Base - The Base value for the PHI node that needs to be inserted for
287 /// this use. As the use is processed, information gets moved from this
288 /// field to the Imm field (below). BasedUser values are sorted by this
292 /// Inst - The instruction using the induction variable.
295 /// OperandValToReplace - The operand value of Inst to replace with the
297 Value *OperandValToReplace;
299 /// Imm - The immediate value that should be added to the base immediately
300 /// before Inst, because it will be folded into the imm field of the
301 /// instruction. This is also sometimes used for loop-variant values that
302 /// must be added inside the loop.
305 /// Phi - The induction variable that performs the striding that
306 /// should be used for this user.
309 // isUseOfPostIncrementedValue - True if this should use the
310 // post-incremented version of this IV, not the preincremented version.
311 // This can only be set in special cases, such as the terminating setcc
312 // instruction for a loop and uses outside the loop that are dominated by
314 bool isUseOfPostIncrementedValue;
316 BasedUser(IVStrideUse &IVSU, ScalarEvolution *se)
317 : Base(IVSU.getOffset()), Inst(IVSU.getUser()),
318 OperandValToReplace(IVSU.getOperandValToReplace()),
319 Imm(se->getIntegerSCEV(0, Base->getType())),
320 isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue()) {}
322 // Once we rewrite the code to insert the new IVs we want, update the
323 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
325 void RewriteInstructionToUseNewBase(const SCEV *NewBase,
326 Instruction *InsertPt,
327 SCEVExpander &Rewriter, Loop *L, Pass *P,
328 SmallVectorImpl<WeakVH> &DeadInsts,
329 ScalarEvolution *SE);
331 Value *InsertCodeForBaseAtPosition(const SCEV *NewBase,
333 SCEVExpander &Rewriter,
335 ScalarEvolution *SE);
340 void BasedUser::dump() const {
341 dbgs() << " Base=" << *Base;
342 dbgs() << " Imm=" << *Imm;
343 dbgs() << " Inst: " << *Inst;
346 Value *BasedUser::InsertCodeForBaseAtPosition(const SCEV *NewBase,
348 SCEVExpander &Rewriter,
350 ScalarEvolution *SE) {
351 Value *Base = Rewriter.expandCodeFor(NewBase, 0, IP);
353 // Wrap the base in a SCEVUnknown so that ScalarEvolution doesn't try to
355 const SCEV *NewValSCEV = SE->getUnknown(Base);
357 // Always emit the immediate into the same block as the user.
358 NewValSCEV = SE->getAddExpr(NewValSCEV, Imm);
360 return Rewriter.expandCodeFor(NewValSCEV, Ty, IP);
364 // Once we rewrite the code to insert the new IVs we want, update the
365 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
366 // to it. NewBasePt is the last instruction which contributes to the
367 // value of NewBase in the case that it's a diffferent instruction from
368 // the PHI that NewBase is computed from, or null otherwise.
370 void BasedUser::RewriteInstructionToUseNewBase(const SCEV *NewBase,
371 Instruction *NewBasePt,
372 SCEVExpander &Rewriter, Loop *L, Pass *P,
373 SmallVectorImpl<WeakVH> &DeadInsts,
374 ScalarEvolution *SE) {
375 if (!isa<PHINode>(Inst)) {
376 // By default, insert code at the user instruction.
377 BasicBlock::iterator InsertPt = Inst;
379 // However, if the Operand is itself an instruction, the (potentially
380 // complex) inserted code may be shared by many users. Because of this, we
381 // want to emit code for the computation of the operand right before its old
382 // computation. This is usually safe, because we obviously used to use the
383 // computation when it was computed in its current block. However, in some
384 // cases (e.g. use of a post-incremented induction variable) the NewBase
385 // value will be pinned to live somewhere after the original computation.
386 // In this case, we have to back off.
388 // If this is a use outside the loop (which means after, since it is based
389 // on a loop indvar) we use the post-incremented value, so that we don't
390 // artificially make the preinc value live out the bottom of the loop.
391 if (!isUseOfPostIncrementedValue && L->contains(Inst)) {
392 if (NewBasePt && isa<PHINode>(OperandValToReplace)) {
393 InsertPt = NewBasePt;
395 } else if (Instruction *OpInst
396 = dyn_cast<Instruction>(OperandValToReplace)) {
398 while (isa<PHINode>(InsertPt)) ++InsertPt;
401 Value *NewVal = InsertCodeForBaseAtPosition(NewBase,
402 OperandValToReplace->getType(),
403 Rewriter, InsertPt, SE);
404 // Replace the use of the operand Value with the new Phi we just created.
405 Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
407 DEBUG(dbgs() << " Replacing with ");
408 DEBUG(WriteAsOperand(dbgs(), NewVal, /*PrintType=*/false));
409 DEBUG(dbgs() << ", which has value " << *NewBase << " plus IMM "
414 // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm
415 // expression into each operand block that uses it. Note that PHI nodes can
416 // have multiple entries for the same predecessor. We use a map to make sure
417 // that a PHI node only has a single Value* for each predecessor (which also
418 // prevents us from inserting duplicate code in some blocks).
419 DenseMap<BasicBlock*, Value*> InsertedCode;
420 PHINode *PN = cast<PHINode>(Inst);
421 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
422 if (PN->getIncomingValue(i) == OperandValToReplace) {
423 // If the original expression is outside the loop, put the replacement
424 // code in the same place as the original expression,
425 // which need not be an immediate predecessor of this PHI. This way we
426 // need only one copy of it even if it is referenced multiple times in
427 // the PHI. We don't do this when the original expression is inside the
428 // loop because multiple copies sometimes do useful sinking of code in
430 Instruction *OldLoc = dyn_cast<Instruction>(OperandValToReplace);
431 BasicBlock *PHIPred = PN->getIncomingBlock(i);
432 if (L->contains(OldLoc)) {
433 // If this is a critical edge, split the edge so that we do not insert
434 // the code on all predecessor/successor paths. We do this unless this
435 // is the canonical backedge for this loop, as this can make some
436 // inserted code be in an illegal position.
437 if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 &&
438 !isa<IndirectBrInst>(PHIPred->getTerminator()) &&
439 (PN->getParent() != L->getHeader() || !L->contains(PHIPred))) {
441 // First step, split the critical edge.
442 BasicBlock *NewBB = SplitCriticalEdge(PHIPred, PN->getParent(),
445 // Next step: move the basic block. In particular, if the PHI node
446 // is outside of the loop, and PredTI is in the loop, we want to
447 // move the block to be immediately before the PHI block, not
448 // immediately after PredTI.
449 if (L->contains(PHIPred) && !L->contains(PN))
450 NewBB->moveBefore(PN->getParent());
452 // Splitting the edge can reduce the number of PHI entries we have.
453 e = PN->getNumIncomingValues();
455 i = PN->getBasicBlockIndex(PHIPred);
458 Value *&Code = InsertedCode[PHIPred];
460 // Insert the code into the end of the predecessor block.
461 Instruction *InsertPt = (L->contains(OldLoc)) ?
462 PHIPred->getTerminator() :
463 OldLoc->getParent()->getTerminator();
464 Code = InsertCodeForBaseAtPosition(NewBase, PN->getType(),
465 Rewriter, InsertPt, SE);
467 DEBUG(dbgs() << " Changing PHI use to ");
468 DEBUG(WriteAsOperand(dbgs(), Code, /*PrintType=*/false));
469 DEBUG(dbgs() << ", which has value " << *NewBase << " plus IMM "
473 // Replace the use of the operand Value with the new Phi we just created.
474 PN->setIncomingValue(i, Code);
479 // PHI node might have become a constant value after SplitCriticalEdge.
480 DeadInsts.push_back(Inst);
484 /// fitsInAddressMode - Return true if V can be subsumed within an addressing
485 /// mode, and does not need to be put in a register first.
486 static bool fitsInAddressMode(const SCEV *V, const Type *AccessTy,
487 const TargetLowering *TLI, bool HasBaseReg) {
488 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
489 int64_t VC = SC->getValue()->getSExtValue();
491 TargetLowering::AddrMode AM;
493 AM.HasBaseReg = HasBaseReg;
494 return TLI->isLegalAddressingMode(AM, AccessTy);
496 // Defaults to PPC. PPC allows a sign-extended 16-bit immediate field.
497 return (VC > -(1 << 16) && VC < (1 << 16)-1);
501 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
502 if (GlobalValue *GV = dyn_cast<GlobalValue>(SU->getValue())) {
504 TargetLowering::AddrMode AM;
506 AM.HasBaseReg = HasBaseReg;
507 return TLI->isLegalAddressingMode(AM, AccessTy);
509 // Default: assume global addresses are not legal.
516 /// MoveLoopVariantsToImmediateField - Move any subexpressions from Val that are
517 /// loop varying to the Imm operand.
518 static void MoveLoopVariantsToImmediateField(const SCEV *&Val, const SCEV *&Imm,
519 Loop *L, ScalarEvolution *SE) {
520 if (Val->isLoopInvariant(L)) return; // Nothing to do.
522 if (const SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
523 SmallVector<const SCEV *, 4> NewOps;
524 NewOps.reserve(SAE->getNumOperands());
526 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
527 if (!SAE->getOperand(i)->isLoopInvariant(L)) {
528 // If this is a loop-variant expression, it must stay in the immediate
529 // field of the expression.
530 Imm = SE->getAddExpr(Imm, SAE->getOperand(i));
532 NewOps.push_back(SAE->getOperand(i));
536 Val = SE->getIntegerSCEV(0, Val->getType());
538 Val = SE->getAddExpr(NewOps);
539 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
540 // Try to pull immediates out of the start value of nested addrec's.
541 const SCEV *Start = SARE->getStart();
542 MoveLoopVariantsToImmediateField(Start, Imm, L, SE);
544 SmallVector<const SCEV *, 4> Ops(SARE->op_begin(), SARE->op_end());
546 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
548 // Otherwise, all of Val is variant, move the whole thing over.
549 Imm = SE->getAddExpr(Imm, Val);
550 Val = SE->getIntegerSCEV(0, Val->getType());
555 /// MoveImmediateValues - Look at Val, and pull out any additions of constants
556 /// that can fit into the immediate field of instructions in the target.
557 /// Accumulate these immediate values into the Imm value.
558 static void MoveImmediateValues(const TargetLowering *TLI,
559 const Type *AccessTy,
560 const SCEV *&Val, const SCEV *&Imm,
561 bool isAddress, Loop *L,
562 ScalarEvolution *SE) {
563 if (const SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
564 SmallVector<const SCEV *, 4> NewOps;
565 NewOps.reserve(SAE->getNumOperands());
567 for (unsigned i = 0; i != SAE->getNumOperands(); ++i) {
568 const SCEV *NewOp = SAE->getOperand(i);
569 MoveImmediateValues(TLI, AccessTy, NewOp, Imm, isAddress, L, SE);
571 if (!NewOp->isLoopInvariant(L)) {
572 // If this is a loop-variant expression, it must stay in the immediate
573 // field of the expression.
574 Imm = SE->getAddExpr(Imm, NewOp);
576 NewOps.push_back(NewOp);
581 Val = SE->getIntegerSCEV(0, Val->getType());
583 Val = SE->getAddExpr(NewOps);
585 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
586 // Try to pull immediates out of the start value of nested addrec's.
587 const SCEV *Start = SARE->getStart();
588 MoveImmediateValues(TLI, AccessTy, Start, Imm, isAddress, L, SE);
590 if (Start != SARE->getStart()) {
591 SmallVector<const SCEV *, 4> Ops(SARE->op_begin(), SARE->op_end());
593 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
596 } else if (const SCEVMulExpr *SME = dyn_cast<SCEVMulExpr>(Val)) {
597 // Transform "8 * (4 + v)" -> "32 + 8*V" if "32" fits in the immed field.
599 fitsInAddressMode(SME->getOperand(0), AccessTy, TLI, false) &&
600 SME->getNumOperands() == 2 && SME->isLoopInvariant(L)) {
602 const SCEV *SubImm = SE->getIntegerSCEV(0, Val->getType());
603 const SCEV *NewOp = SME->getOperand(1);
604 MoveImmediateValues(TLI, AccessTy, NewOp, SubImm, isAddress, L, SE);
606 // If we extracted something out of the subexpressions, see if we can
608 if (NewOp != SME->getOperand(1)) {
609 // Scale SubImm up by "8". If the result is a target constant, we are
611 SubImm = SE->getMulExpr(SubImm, SME->getOperand(0));
612 if (fitsInAddressMode(SubImm, AccessTy, TLI, false)) {
613 // Accumulate the immediate.
614 Imm = SE->getAddExpr(Imm, SubImm);
616 // Update what is left of 'Val'.
617 Val = SE->getMulExpr(SME->getOperand(0), NewOp);
624 // Loop-variant expressions must stay in the immediate field of the
626 if ((isAddress && fitsInAddressMode(Val, AccessTy, TLI, false)) ||
627 !Val->isLoopInvariant(L)) {
628 Imm = SE->getAddExpr(Imm, Val);
629 Val = SE->getIntegerSCEV(0, Val->getType());
633 // Otherwise, no immediates to move.
636 static void MoveImmediateValues(const TargetLowering *TLI,
638 const SCEV *&Val, const SCEV *&Imm,
639 bool isAddress, Loop *L,
640 ScalarEvolution *SE) {
641 const Type *AccessTy = getAccessType(User);
642 MoveImmediateValues(TLI, AccessTy, Val, Imm, isAddress, L, SE);
645 /// SeparateSubExprs - Decompose Expr into all of the subexpressions that are
646 /// added together. This is used to reassociate common addition subexprs
647 /// together for maximal sharing when rewriting bases.
648 static void SeparateSubExprs(SmallVector<const SCEV *, 16> &SubExprs,
650 ScalarEvolution *SE) {
651 if (const SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) {
652 for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
653 SeparateSubExprs(SubExprs, AE->getOperand(j), SE);
654 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) {
655 const SCEV *Zero = SE->getIntegerSCEV(0, Expr->getType());
656 if (SARE->getOperand(0) == Zero) {
657 SubExprs.push_back(Expr);
659 // Compute the addrec with zero as its base.
660 SmallVector<const SCEV *, 4> Ops(SARE->op_begin(), SARE->op_end());
661 Ops[0] = Zero; // Start with zero base.
662 SubExprs.push_back(SE->getAddRecExpr(Ops, SARE->getLoop()));
665 SeparateSubExprs(SubExprs, SARE->getOperand(0), SE);
667 } else if (!Expr->isZero()) {
669 SubExprs.push_back(Expr);
673 // This is logically local to the following function, but C++ says we have
674 // to make it file scope.
675 struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; };
677 /// RemoveCommonExpressionsFromUseBases - Look through all of the Bases of all
678 /// the Uses, removing any common subexpressions, except that if all such
679 /// subexpressions can be folded into an addressing mode for all uses inside
680 /// the loop (this case is referred to as "free" in comments herein) we do
681 /// not remove anything. This looks for things like (a+b+c) and
682 /// (a+c+d) and computes the common (a+c) subexpression. The common expression
683 /// is *removed* from the Bases and returned.
685 RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses,
686 ScalarEvolution *SE, Loop *L,
687 const TargetLowering *TLI) {
688 unsigned NumUses = Uses.size();
690 // Only one use? This is a very common case, so we handle it specially and
692 const SCEV *Zero = SE->getIntegerSCEV(0, Uses[0].Base->getType());
693 const SCEV *Result = Zero;
694 const SCEV *FreeResult = Zero;
696 // If the use is inside the loop, use its base, regardless of what it is:
697 // it is clearly shared across all the IV's. If the use is outside the loop
698 // (which means after it) we don't want to factor anything *into* the loop,
699 // so just use 0 as the base.
700 if (L->contains(Uses[0].Inst))
701 std::swap(Result, Uses[0].Base);
705 // To find common subexpressions, count how many of Uses use each expression.
706 // If any subexpressions are used Uses.size() times, they are common.
707 // Also track whether all uses of each expression can be moved into an
708 // an addressing mode "for free"; such expressions are left within the loop.
709 // struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; };
710 std::map<const SCEV *, SubExprUseData> SubExpressionUseData;
712 // UniqueSubExprs - Keep track of all of the subexpressions we see in the
713 // order we see them.
714 SmallVector<const SCEV *, 16> UniqueSubExprs;
716 SmallVector<const SCEV *, 16> SubExprs;
717 unsigned NumUsesInsideLoop = 0;
718 for (unsigned i = 0; i != NumUses; ++i) {
719 // If the user is outside the loop, just ignore it for base computation.
720 // Since the user is outside the loop, it must be *after* the loop (if it
721 // were before, it could not be based on the loop IV). We don't want users
722 // after the loop to affect base computation of values *inside* the loop,
723 // because we can always add their offsets to the result IV after the loop
724 // is done, ensuring we get good code inside the loop.
725 if (!L->contains(Uses[i].Inst))
729 // If the base is zero (which is common), return zero now, there are no
731 if (Uses[i].Base == Zero) return Zero;
733 // If this use is as an address we may be able to put CSEs in the addressing
734 // mode rather than hoisting them.
735 bool isAddrUse = isAddressUse(Uses[i].Inst, Uses[i].OperandValToReplace);
736 // We may need the AccessTy below, but only when isAddrUse, so compute it
737 // only in that case.
738 const Type *AccessTy = 0;
740 AccessTy = getAccessType(Uses[i].Inst);
742 // Split the expression into subexprs.
743 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
744 // Add one to SubExpressionUseData.Count for each subexpr present, and
745 // if the subexpr is not a valid immediate within an addressing mode use,
746 // set SubExpressionUseData.notAllUsesAreFree. We definitely want to
747 // hoist these out of the loop (if they are common to all uses).
748 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) {
749 if (++SubExpressionUseData[SubExprs[j]].Count == 1)
750 UniqueSubExprs.push_back(SubExprs[j]);
751 if (!isAddrUse || !fitsInAddressMode(SubExprs[j], AccessTy, TLI, false))
752 SubExpressionUseData[SubExprs[j]].notAllUsesAreFree = true;
757 // Now that we know how many times each is used, build Result. Iterate over
758 // UniqueSubexprs so that we have a stable ordering.
759 for (unsigned i = 0, e = UniqueSubExprs.size(); i != e; ++i) {
760 std::map<const SCEV *, SubExprUseData>::iterator I =
761 SubExpressionUseData.find(UniqueSubExprs[i]);
762 assert(I != SubExpressionUseData.end() && "Entry not found?");
763 if (I->second.Count == NumUsesInsideLoop) { // Found CSE!
764 if (I->second.notAllUsesAreFree)
765 Result = SE->getAddExpr(Result, I->first);
767 FreeResult = SE->getAddExpr(FreeResult, I->first);
769 // Remove non-cse's from SubExpressionUseData.
770 SubExpressionUseData.erase(I);
773 if (FreeResult != Zero) {
774 // We have some subexpressions that can be subsumed into addressing
775 // modes in every use inside the loop. However, it's possible that
776 // there are so many of them that the combined FreeResult cannot
777 // be subsumed, or that the target cannot handle both a FreeResult
778 // and a Result in the same instruction (for example because it would
779 // require too many registers). Check this.
780 for (unsigned i=0; i<NumUses; ++i) {
781 if (!L->contains(Uses[i].Inst))
783 // We know this is an addressing mode use; if there are any uses that
784 // are not, FreeResult would be Zero.
785 const Type *AccessTy = getAccessType(Uses[i].Inst);
786 if (!fitsInAddressMode(FreeResult, AccessTy, TLI, Result!=Zero)) {
787 // FIXME: could split up FreeResult into pieces here, some hoisted
788 // and some not. There is no obvious advantage to this.
789 Result = SE->getAddExpr(Result, FreeResult);
796 // If we found no CSE's, return now.
797 if (Result == Zero) return Result;
799 // If we still have a FreeResult, remove its subexpressions from
800 // SubExpressionUseData. This means they will remain in the use Bases.
801 if (FreeResult != Zero) {
802 SeparateSubExprs(SubExprs, FreeResult, SE);
803 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) {
804 std::map<const SCEV *, SubExprUseData>::iterator I =
805 SubExpressionUseData.find(SubExprs[j]);
806 SubExpressionUseData.erase(I);
811 // Otherwise, remove all of the CSE's we found from each of the base values.
812 for (unsigned i = 0; i != NumUses; ++i) {
813 // Uses outside the loop don't necessarily include the common base, but
814 // the final IV value coming into those uses does. Instead of trying to
815 // remove the pieces of the common base, which might not be there,
816 // subtract off the base to compensate for this.
817 if (!L->contains(Uses[i].Inst)) {
818 Uses[i].Base = SE->getMinusSCEV(Uses[i].Base, Result);
822 // Split the expression into subexprs.
823 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
825 // Remove any common subexpressions.
826 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
827 if (SubExpressionUseData.count(SubExprs[j])) {
828 SubExprs.erase(SubExprs.begin()+j);
832 // Finally, add the non-shared expressions together.
833 if (SubExprs.empty())
836 Uses[i].Base = SE->getAddExpr(SubExprs);
843 /// ValidScale - Check whether the given Scale is valid for all loads and
844 /// stores in UsersToProcess.
846 bool LoopStrengthReduce::ValidScale(bool HasBaseReg, int64_t Scale,
847 const std::vector<BasedUser>& UsersToProcess) {
851 for (unsigned i = 0, e = UsersToProcess.size(); i!=e; ++i) {
852 // If this is a load or other access, pass the type of the access in.
853 const Type *AccessTy =
854 Type::getVoidTy(UsersToProcess[i].Inst->getContext());
855 if (isAddressUse(UsersToProcess[i].Inst,
856 UsersToProcess[i].OperandValToReplace))
857 AccessTy = getAccessType(UsersToProcess[i].Inst);
858 else if (isa<PHINode>(UsersToProcess[i].Inst))
861 TargetLowering::AddrMode AM;
862 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
863 AM.BaseOffs = SC->getValue()->getSExtValue();
864 AM.HasBaseReg = HasBaseReg || !UsersToProcess[i].Base->isZero();
867 // If load[imm+r*scale] is illegal, bail out.
868 if (!TLI->isLegalAddressingMode(AM, AccessTy))
874 /// ValidOffset - Check whether the given Offset is valid for all loads and
875 /// stores in UsersToProcess.
877 bool LoopStrengthReduce::ValidOffset(bool HasBaseReg,
880 const std::vector<BasedUser>& UsersToProcess) {
884 for (unsigned i=0, e = UsersToProcess.size(); i!=e; ++i) {
885 // If this is a load or other access, pass the type of the access in.
886 const Type *AccessTy =
887 Type::getVoidTy(UsersToProcess[i].Inst->getContext());
888 if (isAddressUse(UsersToProcess[i].Inst,
889 UsersToProcess[i].OperandValToReplace))
890 AccessTy = getAccessType(UsersToProcess[i].Inst);
891 else if (isa<PHINode>(UsersToProcess[i].Inst))
894 TargetLowering::AddrMode AM;
895 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
896 AM.BaseOffs = SC->getValue()->getSExtValue();
897 AM.BaseOffs = (uint64_t)AM.BaseOffs + (uint64_t)Offset;
898 AM.HasBaseReg = HasBaseReg || !UsersToProcess[i].Base->isZero();
901 // If load[imm+r*scale] is illegal, bail out.
902 if (!TLI->isLegalAddressingMode(AM, AccessTy))
908 /// RequiresTypeConversion - Returns true if converting Ty1 to Ty2 is not
910 bool LoopStrengthReduce::RequiresTypeConversion(const Type *Ty1,
914 Ty1 = SE->getEffectiveSCEVType(Ty1);
915 Ty2 = SE->getEffectiveSCEVType(Ty2);
918 if (Ty1->canLosslesslyBitCastTo(Ty2))
920 if (TLI && TLI->isTruncateFree(Ty1, Ty2))
925 /// CheckForIVReuse - Returns the multiple if the stride is the multiple
926 /// of a previous stride and it is a legal value for the target addressing
927 /// mode scale component and optional base reg. This allows the users of
928 /// this stride to be rewritten as prev iv * factor. It returns 0 if no
929 /// reuse is possible. Factors can be negative on same targets, e.g. ARM.
931 /// If all uses are outside the loop, we don't require that all multiplies
932 /// be folded into the addressing mode, nor even that the factor be constant;
933 /// a multiply (executed once) outside the loop is better than another IV
934 /// within. Well, usually.
935 const SCEV *LoopStrengthReduce::CheckForIVReuse(bool HasBaseReg,
936 bool AllUsesAreAddresses,
937 bool AllUsesAreOutsideLoop,
939 IVExpr &IV, const Type *Ty,
940 const std::vector<BasedUser>& UsersToProcess) {
941 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Stride)) {
942 int64_t SInt = SC->getValue()->getSExtValue();
943 for (unsigned NewStride = 0, e = IU->StrideOrder.size();
944 NewStride != e; ++NewStride) {
945 std::map<const SCEV *, IVsOfOneStride>::iterator SI =
946 IVsByStride.find(IU->StrideOrder[NewStride]);
947 if (SI == IVsByStride.end() || !isa<SCEVConstant>(SI->first))
949 // The other stride has no uses, don't reuse it.
950 std::map<const SCEV *, IVUsersOfOneStride *>::iterator UI =
951 IU->IVUsesByStride.find(IU->StrideOrder[NewStride]);
952 if (UI->second->Users.empty())
954 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
955 if (SI->first != Stride &&
956 (unsigned(abs64(SInt)) < SSInt || (SInt % SSInt) != 0))
958 int64_t Scale = SInt / SSInt;
959 // Check that this stride is valid for all the types used for loads and
960 // stores; if it can be used for some and not others, we might as well use
961 // the original stride everywhere, since we have to create the IV for it
962 // anyway. If the scale is 1, then we don't need to worry about folding
965 (AllUsesAreAddresses &&
966 ValidScale(HasBaseReg, Scale, UsersToProcess))) {
967 // Prefer to reuse an IV with a base of zero.
968 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
969 IE = SI->second.IVs.end(); II != IE; ++II)
970 // Only reuse previous IV if it would not require a type conversion
971 // and if the base difference can be folded.
972 if (II->Base->isZero() &&
973 !RequiresTypeConversion(II->Base->getType(), Ty)) {
975 return SE->getIntegerSCEV(Scale, Stride->getType());
977 // Otherwise, settle for an IV with a foldable base.
978 if (AllUsesAreAddresses)
979 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
980 IE = SI->second.IVs.end(); II != IE; ++II)
981 // Only reuse previous IV if it would not require a type conversion
982 // and if the base difference can be folded.
983 if (SE->getEffectiveSCEVType(II->Base->getType()) ==
984 SE->getEffectiveSCEVType(Ty) &&
985 isa<SCEVConstant>(II->Base)) {
987 cast<SCEVConstant>(II->Base)->getValue()->getSExtValue();
988 if (Base > INT32_MIN && Base <= INT32_MAX &&
989 ValidOffset(HasBaseReg, -Base * Scale,
990 Scale, UsersToProcess)) {
992 return SE->getIntegerSCEV(Scale, Stride->getType());
997 } else if (AllUsesAreOutsideLoop) {
998 // Accept nonconstant strides here; it is really really right to substitute
999 // an existing IV if we can.
1000 for (unsigned NewStride = 0, e = IU->StrideOrder.size();
1001 NewStride != e; ++NewStride) {
1002 std::map<const SCEV *, IVsOfOneStride>::iterator SI =
1003 IVsByStride.find(IU->StrideOrder[NewStride]);
1004 if (SI == IVsByStride.end() || !isa<SCEVConstant>(SI->first))
1006 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1007 if (SI->first != Stride && SSInt != 1)
1009 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1010 IE = SI->second.IVs.end(); II != IE; ++II)
1011 // Accept nonzero base here.
1012 // Only reuse previous IV if it would not require a type conversion.
1013 if (!RequiresTypeConversion(II->Base->getType(), Ty)) {
1018 // Special case, old IV is -1*x and this one is x. Can treat this one as
1020 for (unsigned NewStride = 0, e = IU->StrideOrder.size();
1021 NewStride != e; ++NewStride) {
1022 std::map<const SCEV *, IVsOfOneStride>::iterator SI =
1023 IVsByStride.find(IU->StrideOrder[NewStride]);
1024 if (SI == IVsByStride.end())
1026 if (const SCEVMulExpr *ME = dyn_cast<SCEVMulExpr>(SI->first))
1027 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(ME->getOperand(0)))
1028 if (Stride == ME->getOperand(1) &&
1029 SC->getValue()->getSExtValue() == -1LL)
1030 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1031 IE = SI->second.IVs.end(); II != IE; ++II)
1032 // Accept nonzero base here.
1033 // Only reuse previous IV if it would not require type conversion.
1034 if (!RequiresTypeConversion(II->Base->getType(), Ty)) {
1036 return SE->getIntegerSCEV(-1LL, Stride->getType());
1040 return SE->getIntegerSCEV(0, Stride->getType());
1043 /// PartitionByIsUseOfPostIncrementedValue - Simple boolean predicate that
1044 /// returns true if Val's isUseOfPostIncrementedValue is true.
1045 static bool PartitionByIsUseOfPostIncrementedValue(const BasedUser &Val) {
1046 return Val.isUseOfPostIncrementedValue;
1049 /// isNonConstantNegative - Return true if the specified scev is negated, but
1051 static bool isNonConstantNegative(const SCEV *Expr) {
1052 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Expr);
1053 if (!Mul) return false;
1055 // If there is a constant factor, it will be first.
1056 const SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
1057 if (!SC) return false;
1059 // Return true if the value is negative, this matches things like (-42 * V).
1060 return SC->getValue()->getValue().isNegative();
1063 /// CollectIVUsers - Transform our list of users and offsets to a bit more
1064 /// complex table. In this new vector, each 'BasedUser' contains 'Base', the
1065 /// base of the strided accesses, as well as the old information from Uses. We
1066 /// progressively move information from the Base field to the Imm field, until
1067 /// we eventually have the full access expression to rewrite the use.
1068 const SCEV *LoopStrengthReduce::CollectIVUsers(const SCEV *Stride,
1069 IVUsersOfOneStride &Uses,
1071 bool &AllUsesAreAddresses,
1072 bool &AllUsesAreOutsideLoop,
1073 std::vector<BasedUser> &UsersToProcess) {
1074 // FIXME: Generalize to non-affine IV's.
1075 if (!Stride->isLoopInvariant(L))
1076 return SE->getIntegerSCEV(0, Stride->getType());
1078 UsersToProcess.reserve(Uses.Users.size());
1079 for (ilist<IVStrideUse>::iterator I = Uses.Users.begin(),
1080 E = Uses.Users.end(); I != E; ++I) {
1081 UsersToProcess.push_back(BasedUser(*I, SE));
1083 // Move any loop variant operands from the offset field to the immediate
1084 // field of the use, so that we don't try to use something before it is
1086 MoveLoopVariantsToImmediateField(UsersToProcess.back().Base,
1087 UsersToProcess.back().Imm, L, SE);
1088 assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
1089 "Base value is not loop invariant!");
1092 // We now have a whole bunch of uses of like-strided induction variables, but
1093 // they might all have different bases. We want to emit one PHI node for this
1094 // stride which we fold as many common expressions (between the IVs) into as
1095 // possible. Start by identifying the common expressions in the base values
1096 // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
1097 // "A+B"), emit it to the preheader, then remove the expression from the
1098 // UsersToProcess base values.
1099 const SCEV *CommonExprs =
1100 RemoveCommonExpressionsFromUseBases(UsersToProcess, SE, L, TLI);
1102 // Next, figure out what we can represent in the immediate fields of
1103 // instructions. If we can represent anything there, move it to the imm
1104 // fields of the BasedUsers. We do this so that it increases the commonality
1105 // of the remaining uses.
1106 unsigned NumPHI = 0;
1107 bool HasAddress = false;
1108 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1109 // If the user is not in the current loop, this means it is using the exit
1110 // value of the IV. Do not put anything in the base, make sure it's all in
1111 // the immediate field to allow as much factoring as possible.
1112 if (!L->contains(UsersToProcess[i].Inst)) {
1113 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm,
1114 UsersToProcess[i].Base);
1115 UsersToProcess[i].Base =
1116 SE->getIntegerSCEV(0, UsersToProcess[i].Base->getType());
1118 // Not all uses are outside the loop.
1119 AllUsesAreOutsideLoop = false;
1121 // Addressing modes can be folded into loads and stores. Be careful that
1122 // the store is through the expression, not of the expression though.
1124 bool isAddress = isAddressUse(UsersToProcess[i].Inst,
1125 UsersToProcess[i].OperandValToReplace);
1126 if (isa<PHINode>(UsersToProcess[i].Inst)) {
1134 // If this use isn't an address, then not all uses are addresses.
1135 if (!isAddress && !isPHI)
1136 AllUsesAreAddresses = false;
1138 MoveImmediateValues(TLI, UsersToProcess[i].Inst, UsersToProcess[i].Base,
1139 UsersToProcess[i].Imm, isAddress, L, SE);
1143 // If one of the use is a PHI node and all other uses are addresses, still
1144 // allow iv reuse. Essentially we are trading one constant multiplication
1145 // for one fewer iv.
1147 AllUsesAreAddresses = false;
1149 // There are no in-loop address uses.
1150 if (AllUsesAreAddresses && (!HasAddress && !AllUsesAreOutsideLoop))
1151 AllUsesAreAddresses = false;
1156 /// ShouldUseFullStrengthReductionMode - Test whether full strength-reduction
1157 /// is valid and profitable for the given set of users of a stride. In
1158 /// full strength-reduction mode, all addresses at the current stride are
1159 /// strength-reduced all the way down to pointer arithmetic.
1161 bool LoopStrengthReduce::ShouldUseFullStrengthReductionMode(
1162 const std::vector<BasedUser> &UsersToProcess,
1164 bool AllUsesAreAddresses,
1165 const SCEV *Stride) {
1166 if (!EnableFullLSRMode)
1169 // The heuristics below aim to avoid increasing register pressure, but
1170 // fully strength-reducing all the addresses increases the number of
1171 // add instructions, so don't do this when optimizing for size.
1172 // TODO: If the loop is large, the savings due to simpler addresses
1173 // may oughtweight the costs of the extra increment instructions.
1174 if (L->getHeader()->getParent()->hasFnAttr(Attribute::OptimizeForSize))
1177 // TODO: For now, don't do full strength reduction if there could
1178 // potentially be greater-stride multiples of the current stride
1179 // which could reuse the current stride IV.
1180 if (IU->StrideOrder.back() != Stride)
1183 // Iterate through the uses to find conditions that automatically rule out
1185 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ) {
1186 const SCEV *Base = UsersToProcess[i].Base;
1187 const SCEV *Imm = UsersToProcess[i].Imm;
1188 // If any users have a loop-variant component, they can't be fully
1189 // strength-reduced.
1190 if (Imm && !Imm->isLoopInvariant(L))
1192 // If there are to users with the same base and the difference between
1193 // the two Imm values can't be folded into the address, full
1194 // strength reduction would increase register pressure.
1196 const SCEV *CurImm = UsersToProcess[i].Imm;
1197 if ((CurImm || Imm) && CurImm != Imm) {
1198 if (!CurImm) CurImm = SE->getIntegerSCEV(0, Stride->getType());
1199 if (!Imm) Imm = SE->getIntegerSCEV(0, Stride->getType());
1200 const Instruction *Inst = UsersToProcess[i].Inst;
1201 const Type *AccessTy = getAccessType(Inst);
1202 const SCEV *Diff = SE->getMinusSCEV(UsersToProcess[i].Imm, Imm);
1203 if (!Diff->isZero() &&
1204 (!AllUsesAreAddresses ||
1205 !fitsInAddressMode(Diff, AccessTy, TLI, /*HasBaseReg=*/true)))
1208 } while (++i != e && Base == UsersToProcess[i].Base);
1211 // If there's exactly one user in this stride, fully strength-reducing it
1212 // won't increase register pressure. If it's starting from a non-zero base,
1213 // it'll be simpler this way.
1214 if (UsersToProcess.size() == 1 && !UsersToProcess[0].Base->isZero())
1217 // Otherwise, if there are any users in this stride that don't require
1218 // a register for their base, full strength-reduction will increase
1219 // register pressure.
1220 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1221 if (UsersToProcess[i].Base->isZero())
1224 // Otherwise, go for it.
1228 /// InsertAffinePhi Create and insert a PHI node for an induction variable
1229 /// with the specified start and step values in the specified loop.
1231 /// If NegateStride is true, the stride should be negated by using a
1232 /// subtract instead of an add.
1234 /// Return the created phi node.
1236 static PHINode *InsertAffinePhi(const SCEV *Start, const SCEV *Step,
1237 Instruction *IVIncInsertPt,
1239 SCEVExpander &Rewriter) {
1240 assert(Start->isLoopInvariant(L) && "New PHI start is not loop invariant!");
1241 assert(Step->isLoopInvariant(L) && "New PHI stride is not loop invariant!");
1243 BasicBlock *Header = L->getHeader();
1244 BasicBlock *Preheader = L->getLoopPreheader();
1245 BasicBlock *LatchBlock = L->getLoopLatch();
1246 const Type *Ty = Start->getType();
1247 Ty = Rewriter.SE.getEffectiveSCEVType(Ty);
1249 PHINode *PN = PHINode::Create(Ty, "lsr.iv", Header->begin());
1250 PN->addIncoming(Rewriter.expandCodeFor(Start, Ty, Preheader->getTerminator()),
1253 // If the stride is negative, insert a sub instead of an add for the
1255 bool isNegative = isNonConstantNegative(Step);
1256 const SCEV *IncAmount = Step;
1258 IncAmount = Rewriter.SE.getNegativeSCEV(Step);
1260 // Insert an add instruction right before the terminator corresponding
1261 // to the back-edge or just before the only use. The location is determined
1262 // by the caller and passed in as IVIncInsertPt.
1263 Value *StepV = Rewriter.expandCodeFor(IncAmount, Ty,
1264 Preheader->getTerminator());
1267 IncV = BinaryOperator::CreateSub(PN, StepV, "lsr.iv.next",
1270 IncV = BinaryOperator::CreateAdd(PN, StepV, "lsr.iv.next",
1273 if (!isa<ConstantInt>(StepV)) ++NumVariable;
1275 PN->addIncoming(IncV, LatchBlock);
1281 static void SortUsersToProcess(std::vector<BasedUser> &UsersToProcess) {
1282 // We want to emit code for users inside the loop first. To do this, we
1283 // rearrange BasedUser so that the entries at the end have
1284 // isUseOfPostIncrementedValue = false, because we pop off the end of the
1285 // vector (so we handle them first).
1286 std::partition(UsersToProcess.begin(), UsersToProcess.end(),
1287 PartitionByIsUseOfPostIncrementedValue);
1289 // Sort this by base, so that things with the same base are handled
1290 // together. By partitioning first and stable-sorting later, we are
1291 // guaranteed that within each base we will pop off users from within the
1292 // loop before users outside of the loop with a particular base.
1294 // We would like to use stable_sort here, but we can't. The problem is that
1295 // const SCEV *'s don't have a deterministic ordering w.r.t to each other, so
1296 // we don't have anything to do a '<' comparison on. Because we think the
1297 // number of uses is small, do a horrible bubble sort which just relies on
1299 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1300 // Get a base value.
1301 const SCEV *Base = UsersToProcess[i].Base;
1303 // Compact everything with this base to be consecutive with this one.
1304 for (unsigned j = i+1; j != e; ++j) {
1305 if (UsersToProcess[j].Base == Base) {
1306 std::swap(UsersToProcess[i+1], UsersToProcess[j]);
1313 /// PrepareToStrengthReduceFully - Prepare to fully strength-reduce
1314 /// UsersToProcess, meaning lowering addresses all the way down to direct
1315 /// pointer arithmetic.
1318 LoopStrengthReduce::PrepareToStrengthReduceFully(
1319 std::vector<BasedUser> &UsersToProcess,
1321 const SCEV *CommonExprs,
1323 SCEVExpander &PreheaderRewriter) {
1324 DEBUG(dbgs() << " Fully reducing all users\n");
1326 // Rewrite the UsersToProcess records, creating a separate PHI for each
1327 // unique Base value.
1328 Instruction *IVIncInsertPt = L->getLoopLatch()->getTerminator();
1329 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ) {
1330 // TODO: The uses are grouped by base, but not sorted. We arbitrarily
1331 // pick the first Imm value here to start with, and adjust it for the
1333 const SCEV *Imm = UsersToProcess[i].Imm;
1334 const SCEV *Base = UsersToProcess[i].Base;
1335 const SCEV *Start = SE->getAddExpr(CommonExprs, Base, Imm);
1336 PHINode *Phi = InsertAffinePhi(Start, Stride, IVIncInsertPt, L,
1338 // Loop over all the users with the same base.
1340 UsersToProcess[i].Base = SE->getIntegerSCEV(0, Stride->getType());
1341 UsersToProcess[i].Imm = SE->getMinusSCEV(UsersToProcess[i].Imm, Imm);
1342 UsersToProcess[i].Phi = Phi;
1343 assert(UsersToProcess[i].Imm->isLoopInvariant(L) &&
1344 "ShouldUseFullStrengthReductionMode should reject this!");
1345 } while (++i != e && Base == UsersToProcess[i].Base);
1349 /// FindIVIncInsertPt - Return the location to insert the increment instruction.
1350 /// If the only use if a use of postinc value, (must be the loop termination
1351 /// condition), then insert it just before the use.
1352 static Instruction *FindIVIncInsertPt(std::vector<BasedUser> &UsersToProcess,
1354 if (UsersToProcess.size() == 1 &&
1355 UsersToProcess[0].isUseOfPostIncrementedValue &&
1356 L->contains(UsersToProcess[0].Inst))
1357 return UsersToProcess[0].Inst;
1358 return L->getLoopLatch()->getTerminator();
1361 /// PrepareToStrengthReduceWithNewPhi - Insert a new induction variable for the
1362 /// given users to share.
1365 LoopStrengthReduce::PrepareToStrengthReduceWithNewPhi(
1366 std::vector<BasedUser> &UsersToProcess,
1368 const SCEV *CommonExprs,
1370 Instruction *IVIncInsertPt,
1372 SCEVExpander &PreheaderRewriter) {
1373 DEBUG(dbgs() << " Inserting new PHI:\n");
1375 PHINode *Phi = InsertAffinePhi(SE->getUnknown(CommonBaseV),
1376 Stride, IVIncInsertPt, L,
1379 // Remember this in case a later stride is multiple of this.
1380 IVsByStride[Stride].addIV(Stride, CommonExprs, Phi);
1382 // All the users will share this new IV.
1383 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1384 UsersToProcess[i].Phi = Phi;
1386 DEBUG(dbgs() << " IV=");
1387 DEBUG(WriteAsOperand(dbgs(), Phi, /*PrintType=*/false));
1388 DEBUG(dbgs() << "\n");
1391 /// PrepareToStrengthReduceFromSmallerStride - Prepare for the given users to
1392 /// reuse an induction variable with a stride that is a factor of the current
1393 /// induction variable.
1396 LoopStrengthReduce::PrepareToStrengthReduceFromSmallerStride(
1397 std::vector<BasedUser> &UsersToProcess,
1399 const IVExpr &ReuseIV,
1400 Instruction *PreInsertPt) {
1401 DEBUG(dbgs() << " Rewriting in terms of existing IV of STRIDE "
1402 << *ReuseIV.Stride << " and BASE " << *ReuseIV.Base << "\n");
1404 // All the users will share the reused IV.
1405 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1406 UsersToProcess[i].Phi = ReuseIV.PHI;
1408 Constant *C = dyn_cast<Constant>(CommonBaseV);
1410 (!C->isNullValue() &&
1411 !fitsInAddressMode(SE->getUnknown(CommonBaseV), CommonBaseV->getType(),
1413 // We want the common base emitted into the preheader! This is just
1414 // using cast as a copy so BitCast (no-op cast) is appropriate
1415 CommonBaseV = new BitCastInst(CommonBaseV, CommonBaseV->getType(),
1416 "commonbase", PreInsertPt);
1419 static bool IsImmFoldedIntoAddrMode(GlobalValue *GV, int64_t Offset,
1420 const Type *AccessTy,
1421 std::vector<BasedUser> &UsersToProcess,
1422 const TargetLowering *TLI) {
1423 SmallVector<Instruction*, 16> AddrModeInsts;
1424 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1425 if (UsersToProcess[i].isUseOfPostIncrementedValue)
1427 ExtAddrMode AddrMode =
1428 AddressingModeMatcher::Match(UsersToProcess[i].OperandValToReplace,
1429 AccessTy, UsersToProcess[i].Inst,
1430 AddrModeInsts, *TLI);
1431 if (GV && GV != AddrMode.BaseGV)
1433 if (Offset && !AddrMode.BaseOffs)
1434 // FIXME: How to accurate check it's immediate offset is folded.
1436 AddrModeInsts.clear();
1441 /// StrengthReduceIVUsersOfStride - Strength reduce all of the users of a single
1442 /// stride of IV. All of the users may have different starting values, and this
1443 /// may not be the only stride.
1445 LoopStrengthReduce::StrengthReduceIVUsersOfStride(const SCEV *Stride,
1446 IVUsersOfOneStride &Uses,
1448 // If all the users are moved to another stride, then there is nothing to do.
1449 if (Uses.Users.empty())
1452 // Keep track if every use in UsersToProcess is an address. If they all are,
1453 // we may be able to rewrite the entire collection of them in terms of a
1454 // smaller-stride IV.
1455 bool AllUsesAreAddresses = true;
1457 // Keep track if every use of a single stride is outside the loop. If so,
1458 // we want to be more aggressive about reusing a smaller-stride IV; a
1459 // multiply outside the loop is better than another IV inside. Well, usually.
1460 bool AllUsesAreOutsideLoop = true;
1462 // Transform our list of users and offsets to a bit more complex table. In
1463 // this new vector, each 'BasedUser' contains 'Base' the base of the
1464 // strided accessas well as the old information from Uses. We progressively
1465 // move information from the Base field to the Imm field, until we eventually
1466 // have the full access expression to rewrite the use.
1467 std::vector<BasedUser> UsersToProcess;
1468 const SCEV *CommonExprs = CollectIVUsers(Stride, Uses, L, AllUsesAreAddresses,
1469 AllUsesAreOutsideLoop,
1472 // Sort the UsersToProcess array so that users with common bases are
1473 // next to each other.
1474 SortUsersToProcess(UsersToProcess);
1476 // If we managed to find some expressions in common, we'll need to carry
1477 // their value in a register and add it in for each use. This will take up
1478 // a register operand, which potentially restricts what stride values are
1480 bool HaveCommonExprs = !CommonExprs->isZero();
1481 const Type *ReplacedTy = CommonExprs->getType();
1483 // If all uses are addresses, consider sinking the immediate part of the
1484 // common expression back into uses if they can fit in the immediate fields.
1485 if (TLI && HaveCommonExprs && AllUsesAreAddresses) {
1486 const SCEV *NewCommon = CommonExprs;
1487 const SCEV *Imm = SE->getIntegerSCEV(0, ReplacedTy);
1488 MoveImmediateValues(TLI, Type::getVoidTy(
1489 L->getLoopPreheader()->getContext()),
1490 NewCommon, Imm, true, L, SE);
1491 if (!Imm->isZero()) {
1494 // If the immediate part of the common expression is a GV, check if it's
1495 // possible to fold it into the target addressing mode.
1496 GlobalValue *GV = 0;
1497 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(Imm))
1498 GV = dyn_cast<GlobalValue>(SU->getValue());
1500 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Imm))
1501 Offset = SC->getValue()->getSExtValue();
1503 // Pass VoidTy as the AccessTy to be conservative, because
1504 // there could be multiple access types among all the uses.
1505 DoSink = IsImmFoldedIntoAddrMode(GV, Offset,
1506 Type::getVoidTy(L->getLoopPreheader()->getContext()),
1507 UsersToProcess, TLI);
1510 DEBUG(dbgs() << " Sinking " << *Imm << " back down into uses\n");
1511 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1512 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm, Imm);
1513 CommonExprs = NewCommon;
1514 HaveCommonExprs = !CommonExprs->isZero();
1520 // Now that we know what we need to do, insert the PHI node itself.
1522 DEBUG(dbgs() << "LSR: Examining IVs of TYPE " << *ReplacedTy << " of STRIDE "
1524 << " Common base: " << *CommonExprs << "\n");
1526 SCEVExpander Rewriter(*SE);
1527 SCEVExpander PreheaderRewriter(*SE);
1529 BasicBlock *Preheader = L->getLoopPreheader();
1530 Instruction *PreInsertPt = Preheader->getTerminator();
1531 BasicBlock *LatchBlock = L->getLoopLatch();
1532 Instruction *IVIncInsertPt = LatchBlock->getTerminator();
1534 Value *CommonBaseV = Constant::getNullValue(ReplacedTy);
1536 const SCEV *RewriteFactor = SE->getIntegerSCEV(0, ReplacedTy);
1537 IVExpr ReuseIV(SE->getIntegerSCEV(0,
1538 Type::getInt32Ty(Preheader->getContext())),
1539 SE->getIntegerSCEV(0,
1540 Type::getInt32Ty(Preheader->getContext())),
1543 // Choose a strength-reduction strategy and prepare for it by creating
1544 // the necessary PHIs and adjusting the bookkeeping.
1545 if (ShouldUseFullStrengthReductionMode(UsersToProcess, L,
1546 AllUsesAreAddresses, Stride)) {
1547 PrepareToStrengthReduceFully(UsersToProcess, Stride, CommonExprs, L,
1550 // Emit the initial base value into the loop preheader.
1551 CommonBaseV = PreheaderRewriter.expandCodeFor(CommonExprs, ReplacedTy,
1554 // If all uses are addresses, check if it is possible to reuse an IV. The
1555 // new IV must have a stride that is a multiple of the old stride; the
1556 // multiple must be a number that can be encoded in the scale field of the
1557 // target addressing mode; and we must have a valid instruction after this
1558 // substitution, including the immediate field, if any.
1559 RewriteFactor = CheckForIVReuse(HaveCommonExprs, AllUsesAreAddresses,
1560 AllUsesAreOutsideLoop,
1561 Stride, ReuseIV, ReplacedTy,
1563 if (!RewriteFactor->isZero())
1564 PrepareToStrengthReduceFromSmallerStride(UsersToProcess, CommonBaseV,
1565 ReuseIV, PreInsertPt);
1567 IVIncInsertPt = FindIVIncInsertPt(UsersToProcess, L);
1568 PrepareToStrengthReduceWithNewPhi(UsersToProcess, Stride, CommonExprs,
1569 CommonBaseV, IVIncInsertPt,
1570 L, PreheaderRewriter);
1574 // Process all the users now, replacing their strided uses with
1575 // strength-reduced forms. This outer loop handles all bases, the inner
1576 // loop handles all users of a particular base.
1577 while (!UsersToProcess.empty()) {
1578 const SCEV *Base = UsersToProcess.back().Base;
1579 Instruction *Inst = UsersToProcess.back().Inst;
1581 // Emit the code for Base into the preheader.
1583 if (!Base->isZero()) {
1584 BaseV = PreheaderRewriter.expandCodeFor(Base, 0, PreInsertPt);
1586 DEBUG(dbgs() << " INSERTING code for BASE = " << *Base << ":");
1587 if (BaseV->hasName())
1588 DEBUG(dbgs() << " Result value name = %" << BaseV->getName());
1589 DEBUG(dbgs() << "\n");
1591 // If BaseV is a non-zero constant, make sure that it gets inserted into
1592 // the preheader, instead of being forward substituted into the uses. We
1593 // do this by forcing a BitCast (noop cast) to be inserted into the
1594 // preheader in this case.
1595 if (!fitsInAddressMode(Base, getAccessType(Inst), TLI, false) &&
1596 isa<Constant>(BaseV)) {
1597 // We want this constant emitted into the preheader! This is just
1598 // using cast as a copy so BitCast (no-op cast) is appropriate
1599 BaseV = new BitCastInst(BaseV, BaseV->getType(), "preheaderinsert",
1604 // Emit the code to add the immediate offset to the Phi value, just before
1605 // the instructions that we identified as using this stride and base.
1607 // FIXME: Use emitted users to emit other users.
1608 BasedUser &User = UsersToProcess.back();
1610 DEBUG(dbgs() << " Examining ");
1611 if (User.isUseOfPostIncrementedValue)
1612 DEBUG(dbgs() << "postinc");
1614 DEBUG(dbgs() << "preinc");
1615 DEBUG(dbgs() << " use ");
1616 DEBUG(WriteAsOperand(dbgs(), UsersToProcess.back().OperandValToReplace,
1617 /*PrintType=*/false));
1618 DEBUG(dbgs() << " in Inst: " << *User.Inst << '\n');
1620 // If this instruction wants to use the post-incremented value, move it
1621 // after the post-inc and use its value instead of the PHI.
1622 Value *RewriteOp = User.Phi;
1623 if (User.isUseOfPostIncrementedValue) {
1624 RewriteOp = User.Phi->getIncomingValueForBlock(LatchBlock);
1625 // If this user is in the loop, make sure it is the last thing in the
1626 // loop to ensure it is dominated by the increment. In case it's the
1627 // only use of the iv, the increment instruction is already before the
1629 if (L->contains(User.Inst) && User.Inst != IVIncInsertPt)
1630 User.Inst->moveBefore(IVIncInsertPt);
1633 const SCEV *RewriteExpr = SE->getUnknown(RewriteOp);
1635 if (SE->getEffectiveSCEVType(RewriteOp->getType()) !=
1636 SE->getEffectiveSCEVType(ReplacedTy)) {
1637 assert(SE->getTypeSizeInBits(RewriteOp->getType()) >
1638 SE->getTypeSizeInBits(ReplacedTy) &&
1639 "Unexpected widening cast!");
1640 RewriteExpr = SE->getTruncateExpr(RewriteExpr, ReplacedTy);
1643 // If we had to insert new instructions for RewriteOp, we have to
1644 // consider that they may not have been able to end up immediately
1645 // next to RewriteOp, because non-PHI instructions may never precede
1646 // PHI instructions in a block. In this case, remember where the last
1647 // instruction was inserted so that if we're replacing a different
1648 // PHI node, we can use the later point to expand the final
1650 Instruction *NewBasePt = dyn_cast<Instruction>(RewriteOp);
1651 if (RewriteOp == User.Phi) NewBasePt = 0;
1653 // Clear the SCEVExpander's expression map so that we are guaranteed
1654 // to have the code emitted where we expect it.
1657 // If we are reusing the iv, then it must be multiplied by a constant
1658 // factor to take advantage of the addressing mode scale component.
1659 if (!RewriteFactor->isZero()) {
1660 // If we're reusing an IV with a nonzero base (currently this happens
1661 // only when all reuses are outside the loop) subtract that base here.
1662 // The base has been used to initialize the PHI node but we don't want
1664 if (!ReuseIV.Base->isZero()) {
1665 const SCEV *typedBase = ReuseIV.Base;
1666 if (SE->getEffectiveSCEVType(RewriteExpr->getType()) !=
1667 SE->getEffectiveSCEVType(ReuseIV.Base->getType())) {
1668 // It's possible the original IV is a larger type than the new IV,
1669 // in which case we have to truncate the Base. We checked in
1670 // RequiresTypeConversion that this is valid.
1671 assert(SE->getTypeSizeInBits(RewriteExpr->getType()) <
1672 SE->getTypeSizeInBits(ReuseIV.Base->getType()) &&
1673 "Unexpected lengthening conversion!");
1674 typedBase = SE->getTruncateExpr(ReuseIV.Base,
1675 RewriteExpr->getType());
1677 RewriteExpr = SE->getMinusSCEV(RewriteExpr, typedBase);
1680 // Multiply old variable, with base removed, by new scale factor.
1681 RewriteExpr = SE->getMulExpr(RewriteFactor,
1684 // The common base is emitted in the loop preheader. But since we
1685 // are reusing an IV, it has not been used to initialize the PHI node.
1686 // Add it to the expression used to rewrite the uses.
1687 // When this use is outside the loop, we earlier subtracted the
1688 // common base, and are adding it back here. Use the same expression
1689 // as before, rather than CommonBaseV, so DAGCombiner will zap it.
1690 if (!CommonExprs->isZero()) {
1691 if (L->contains(User.Inst))
1692 RewriteExpr = SE->getAddExpr(RewriteExpr,
1693 SE->getUnknown(CommonBaseV));
1695 RewriteExpr = SE->getAddExpr(RewriteExpr, CommonExprs);
1699 // Now that we know what we need to do, insert code before User for the
1700 // immediate and any loop-variant expressions.
1702 // Add BaseV to the PHI value if needed.
1703 RewriteExpr = SE->getAddExpr(RewriteExpr, SE->getUnknown(BaseV));
1705 User.RewriteInstructionToUseNewBase(RewriteExpr, NewBasePt,
1709 // Mark old value we replaced as possibly dead, so that it is eliminated
1710 // if we just replaced the last use of that value.
1711 DeadInsts.push_back(User.OperandValToReplace);
1713 UsersToProcess.pop_back();
1716 // If there are any more users to process with the same base, process them
1717 // now. We sorted by base above, so we just have to check the last elt.
1718 } while (!UsersToProcess.empty() && UsersToProcess.back().Base == Base);
1719 // TODO: Next, find out which base index is the most common, pull it out.
1722 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
1723 // different starting values, into different PHIs.
1726 void LoopStrengthReduce::StrengthReduceIVUsers(Loop *L) {
1727 // Note: this processes each stride/type pair individually. All users
1728 // passed into StrengthReduceIVUsersOfStride have the same type AND stride.
1729 // Also, note that we iterate over IVUsesByStride indirectly by using
1730 // StrideOrder. This extra layer of indirection makes the ordering of
1731 // strides deterministic - not dependent on map order.
1732 for (unsigned Stride = 0, e = IU->StrideOrder.size(); Stride != e; ++Stride) {
1733 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
1734 IU->IVUsesByStride.find(IU->StrideOrder[Stride]);
1735 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
1736 // FIXME: Generalize to non-affine IV's.
1737 if (!SI->first->isLoopInvariant(L))
1739 StrengthReduceIVUsersOfStride(SI->first, *SI->second, L);
1743 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1744 /// set the IV user and stride information and return true, otherwise return
1746 bool LoopStrengthReduce::FindIVUserForCond(ICmpInst *Cond,
1747 IVStrideUse *&CondUse,
1748 const SCEV* &CondStride) {
1749 for (unsigned Stride = 0, e = IU->StrideOrder.size();
1750 Stride != e && !CondUse; ++Stride) {
1751 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
1752 IU->IVUsesByStride.find(IU->StrideOrder[Stride]);
1753 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
1755 for (ilist<IVStrideUse>::iterator UI = SI->second->Users.begin(),
1756 E = SI->second->Users.end(); UI != E; ++UI)
1757 if (UI->getUser() == Cond) {
1758 // NOTE: we could handle setcc instructions with multiple uses here, but
1759 // InstCombine does it as well for simple uses, it's not clear that it
1760 // occurs enough in real life to handle.
1762 CondStride = SI->first;
1770 // Constant strides come first which in turns are sorted by their absolute
1771 // values. If absolute values are the same, then positive strides comes first.
1773 // 4, -1, X, 1, 2 ==> 1, -1, 2, 4, X
1774 struct StrideCompare {
1775 const ScalarEvolution *SE;
1776 explicit StrideCompare(const ScalarEvolution *se) : SE(se) {}
1778 bool operator()(const SCEV *LHS, const SCEV *RHS) {
1779 const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS);
1780 const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS);
1782 int64_t LV = LHSC->getValue()->getSExtValue();
1783 int64_t RV = RHSC->getValue()->getSExtValue();
1784 uint64_t ALV = (LV < 0) ? -LV : LV;
1785 uint64_t ARV = (RV < 0) ? -RV : RV;
1793 // If it's the same value but different type, sort by bit width so
1794 // that we emit larger induction variables before smaller
1795 // ones, letting the smaller be re-written in terms of larger ones.
1796 return SE->getTypeSizeInBits(RHS->getType()) <
1797 SE->getTypeSizeInBits(LHS->getType());
1799 return LHSC && !RHSC;
1804 /// ChangeCompareStride - If a loop termination compare instruction is the only
1805 /// use of its stride, and the comparison is against a constant value, try to
1806 /// eliminate the stride by moving the compare instruction to another stride and
1807 /// changing its constant operand accordingly. E.g.
1813 /// if (v2 < 10) goto loop
1818 /// if (v1 < 30) goto loop
1819 ICmpInst *LoopStrengthReduce::ChangeCompareStride(Loop *L, ICmpInst *Cond,
1820 IVStrideUse* &CondUse,
1821 const SCEV* &CondStride,
1823 // If there's only one stride in the loop, there's nothing to do here.
1824 if (IU->StrideOrder.size() < 2)
1826 // If there are other users of the condition's stride, don't bother
1827 // trying to change the condition because the stride will still
1829 std::map<const SCEV *, IVUsersOfOneStride *>::iterator I =
1830 IU->IVUsesByStride.find(CondStride);
1831 if (I == IU->IVUsesByStride.end())
1833 if (I->second->Users.size() > 1) {
1834 for (ilist<IVStrideUse>::iterator II = I->second->Users.begin(),
1835 EE = I->second->Users.end(); II != EE; ++II) {
1836 if (II->getUser() == Cond)
1838 if (!isInstructionTriviallyDead(II->getUser()))
1842 // Only handle constant strides for now.
1843 const SCEVConstant *SC = dyn_cast<SCEVConstant>(CondStride);
1844 if (!SC) return Cond;
1846 ICmpInst::Predicate Predicate = Cond->getPredicate();
1847 int64_t CmpSSInt = SC->getValue()->getSExtValue();
1848 unsigned BitWidth = SE->getTypeSizeInBits(CondStride->getType());
1849 uint64_t SignBit = 1ULL << (BitWidth-1);
1850 const Type *CmpTy = Cond->getOperand(0)->getType();
1851 const Type *NewCmpTy = NULL;
1852 unsigned TyBits = SE->getTypeSizeInBits(CmpTy);
1853 unsigned NewTyBits = 0;
1854 const SCEV *NewStride = NULL;
1855 Value *NewCmpLHS = NULL;
1856 Value *NewCmpRHS = NULL;
1858 const SCEV *NewOffset = SE->getIntegerSCEV(0, CmpTy);
1860 if (ConstantInt *C = dyn_cast<ConstantInt>(Cond->getOperand(1))) {
1861 int64_t CmpVal = C->getValue().getSExtValue();
1863 // Check the relevant induction variable for conformance to
1865 const SCEV *IV = SE->getSCEV(Cond->getOperand(0));
1866 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1867 if (!AR || !AR->isAffine())
1870 const SCEVConstant *StartC = dyn_cast<SCEVConstant>(AR->getStart());
1871 // Check stride constant and the comparision constant signs to detect
1874 if ((StartC->getValue()->getSExtValue() < CmpVal && CmpSSInt < 0) ||
1875 (StartC->getValue()->getSExtValue() > CmpVal && CmpSSInt > 0))
1878 // More restrictive check for the other cases.
1879 if ((CmpVal & SignBit) != (CmpSSInt & SignBit))
1883 // Look for a suitable stride / iv as replacement.
1884 for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
1885 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
1886 IU->IVUsesByStride.find(IU->StrideOrder[i]);
1887 if (!isa<SCEVConstant>(SI->first) || SI->second->Users.empty())
1889 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1890 if (SSInt == CmpSSInt ||
1891 abs64(SSInt) < abs64(CmpSSInt) ||
1892 (SSInt % CmpSSInt) != 0)
1895 Scale = SSInt / CmpSSInt;
1896 int64_t NewCmpVal = CmpVal * Scale;
1898 // If old icmp value fits in icmp immediate field, but the new one doesn't
1899 // try something else.
1901 TLI->isLegalICmpImmediate(CmpVal) &&
1902 !TLI->isLegalICmpImmediate(NewCmpVal))
1905 APInt Mul = APInt(BitWidth*2, CmpVal, true);
1906 Mul = Mul * APInt(BitWidth*2, Scale, true);
1907 // Check for overflow.
1908 if (!Mul.isSignedIntN(BitWidth))
1910 // Check for overflow in the stride's type too.
1911 if (!Mul.isSignedIntN(SE->getTypeSizeInBits(SI->first->getType())))
1914 // Watch out for overflow.
1915 if (ICmpInst::isSigned(Predicate) &&
1916 (CmpVal & SignBit) != (NewCmpVal & SignBit))
1919 // Pick the best iv to use trying to avoid a cast.
1921 for (ilist<IVStrideUse>::iterator UI = SI->second->Users.begin(),
1922 E = SI->second->Users.end(); UI != E; ++UI) {
1923 Value *Op = UI->getOperandValToReplace();
1925 // If the IVStrideUse implies a cast, check for an actual cast which
1926 // can be used to find the original IV expression.
1927 if (SE->getEffectiveSCEVType(Op->getType()) !=
1928 SE->getEffectiveSCEVType(SI->first->getType())) {
1929 CastInst *CI = dyn_cast<CastInst>(Op);
1930 // If it's not a simple cast, it's complicated.
1933 // If it's a cast from a type other than the stride type,
1934 // it's complicated.
1935 if (CI->getOperand(0)->getType() != SI->first->getType())
1937 // Ok, we found the IV expression in the stride's type.
1938 Op = CI->getOperand(0);
1942 if (NewCmpLHS->getType() == CmpTy)
1948 NewCmpTy = NewCmpLHS->getType();
1949 NewTyBits = SE->getTypeSizeInBits(NewCmpTy);
1950 const Type *NewCmpIntTy = IntegerType::get(Cond->getContext(), NewTyBits);
1951 if (RequiresTypeConversion(NewCmpTy, CmpTy)) {
1952 // Check if it is possible to rewrite it using
1953 // an iv / stride of a smaller integer type.
1954 unsigned Bits = NewTyBits;
1955 if (ICmpInst::isSigned(Predicate))
1957 uint64_t Mask = (1ULL << Bits) - 1;
1958 if (((uint64_t)NewCmpVal & Mask) != (uint64_t)NewCmpVal)
1962 // Don't rewrite if use offset is non-constant and the new type is
1963 // of a different type.
1964 // FIXME: too conservative?
1965 if (NewTyBits != TyBits && !isa<SCEVConstant>(CondUse->getOffset()))
1969 bool AllUsesAreAddresses = true;
1970 bool AllUsesAreOutsideLoop = true;
1971 std::vector<BasedUser> UsersToProcess;
1972 const SCEV *CommonExprs = CollectIVUsers(SI->first, *SI->second, L,
1973 AllUsesAreAddresses,
1974 AllUsesAreOutsideLoop,
1976 // Avoid rewriting the compare instruction with an iv of new stride
1977 // if it's likely the new stride uses will be rewritten using the
1978 // stride of the compare instruction.
1979 if (AllUsesAreAddresses &&
1980 ValidScale(!CommonExprs->isZero(), Scale, UsersToProcess))
1984 // Avoid rewriting the compare instruction with an iv which has
1985 // implicit extension or truncation built into it.
1986 // TODO: This is over-conservative.
1987 if (SE->getTypeSizeInBits(CondUse->getOffset()->getType()) != TyBits)
1990 // If scale is negative, use swapped predicate unless it's testing
1992 if (Scale < 0 && !Cond->isEquality())
1993 Predicate = ICmpInst::getSwappedPredicate(Predicate);
1995 NewStride = IU->StrideOrder[i];
1996 if (!isa<PointerType>(NewCmpTy))
1997 NewCmpRHS = ConstantInt::get(NewCmpTy, NewCmpVal);
1999 Constant *CI = ConstantInt::get(NewCmpIntTy, NewCmpVal);
2000 NewCmpRHS = ConstantExpr::getIntToPtr(CI, NewCmpTy);
2002 NewOffset = TyBits == NewTyBits
2003 ? SE->getMulExpr(CondUse->getOffset(),
2004 SE->getConstant(CmpTy, Scale))
2005 : SE->getConstant(NewCmpIntTy,
2006 cast<SCEVConstant>(CondUse->getOffset())->getValue()
2007 ->getSExtValue()*Scale);
2012 // Forgo this transformation if it the increment happens to be
2013 // unfortunately positioned after the condition, and the condition
2014 // has multiple uses which prevent it from being moved immediately
2015 // before the branch. See
2016 // test/Transforms/LoopStrengthReduce/change-compare-stride-trickiness-*.ll
2017 // for an example of this situation.
2018 if (!Cond->hasOneUse()) {
2019 for (BasicBlock::iterator I = Cond, E = Cond->getParent()->end();
2026 // Create a new compare instruction using new stride / iv.
2027 ICmpInst *OldCond = Cond;
2028 // Insert new compare instruction.
2029 Cond = new ICmpInst(OldCond, Predicate, NewCmpLHS, NewCmpRHS,
2030 L->getHeader()->getName() + ".termcond");
2032 DEBUG(dbgs() << " Change compare stride in Inst " << *OldCond);
2033 DEBUG(dbgs() << " to " << *Cond << '\n');
2035 // Remove the old compare instruction. The old indvar is probably dead too.
2036 DeadInsts.push_back(CondUse->getOperandValToReplace());
2037 OldCond->replaceAllUsesWith(Cond);
2038 OldCond->eraseFromParent();
2040 IU->IVUsesByStride[NewStride]->addUser(NewOffset, Cond, NewCmpLHS);
2041 CondUse = &IU->IVUsesByStride[NewStride]->Users.back();
2042 CondStride = NewStride;
2050 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
2051 /// a max computation.
2053 /// This is a narrow solution to a specific, but acute, problem. For loops
2059 /// } while (++i < n);
2061 /// the trip count isn't just 'n', because 'n' might not be positive. And
2062 /// unfortunately this can come up even for loops where the user didn't use
2063 /// a C do-while loop. For example, seemingly well-behaved top-test loops
2064 /// will commonly be lowered like this:
2070 /// } while (++i < n);
2073 /// and then it's possible for subsequent optimization to obscure the if
2074 /// test in such a way that indvars can't find it.
2076 /// When indvars can't find the if test in loops like this, it creates a
2077 /// max expression, which allows it to give the loop a canonical
2078 /// induction variable:
2081 /// max = n < 1 ? 1 : n;
2084 /// } while (++i != max);
2086 /// Canonical induction variables are necessary because the loop passes
2087 /// are designed around them. The most obvious example of this is the
2088 /// LoopInfo analysis, which doesn't remember trip count values. It
2089 /// expects to be able to rediscover the trip count each time it is
2090 /// needed, and it does this using a simple analyis that only succeeds if
2091 /// the loop has a canonical induction variable.
2093 /// However, when it comes time to generate code, the maximum operation
2094 /// can be quite costly, especially if it's inside of an outer loop.
2096 /// This function solves this problem by detecting this type of loop and
2097 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
2098 /// the instructions for the maximum computation.
2100 ICmpInst *LoopStrengthReduce::OptimizeMax(Loop *L, ICmpInst *Cond,
2101 IVStrideUse* &CondUse) {
2102 // Check that the loop matches the pattern we're looking for.
2103 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
2104 Cond->getPredicate() != CmpInst::ICMP_NE)
2107 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
2108 if (!Sel || !Sel->hasOneUse()) return Cond;
2110 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
2111 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2113 const SCEV *One = SE->getIntegerSCEV(1, BackedgeTakenCount->getType());
2115 // Add one to the backedge-taken count to get the trip count.
2116 const SCEV *IterationCount = SE->getAddExpr(BackedgeTakenCount, One);
2118 // Check for a max calculation that matches the pattern.
2119 if (!isa<SCEVSMaxExpr>(IterationCount) && !isa<SCEVUMaxExpr>(IterationCount))
2121 const SCEVNAryExpr *Max = cast<SCEVNAryExpr>(IterationCount);
2122 if (Max != SE->getSCEV(Sel)) return Cond;
2124 // To handle a max with more than two operands, this optimization would
2125 // require additional checking and setup.
2126 if (Max->getNumOperands() != 2)
2129 const SCEV *MaxLHS = Max->getOperand(0);
2130 const SCEV *MaxRHS = Max->getOperand(1);
2131 if (!MaxLHS || MaxLHS != One) return Cond;
2133 // Check the relevant induction variable for conformance to
2135 const SCEV *IV = SE->getSCEV(Cond->getOperand(0));
2136 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2137 if (!AR || !AR->isAffine() ||
2138 AR->getStart() != One ||
2139 AR->getStepRecurrence(*SE) != One)
2142 assert(AR->getLoop() == L &&
2143 "Loop condition operand is an addrec in a different loop!");
2145 // Check the right operand of the select, and remember it, as it will
2146 // be used in the new comparison instruction.
2148 if (SE->getSCEV(Sel->getOperand(1)) == MaxRHS)
2149 NewRHS = Sel->getOperand(1);
2150 else if (SE->getSCEV(Sel->getOperand(2)) == MaxRHS)
2151 NewRHS = Sel->getOperand(2);
2152 if (!NewRHS) return Cond;
2154 // Determine the new comparison opcode. It may be signed or unsigned,
2155 // and the original comparison may be either equality or inequality.
2156 CmpInst::Predicate Pred =
2157 isa<SCEVSMaxExpr>(Max) ? CmpInst::ICMP_SLT : CmpInst::ICMP_ULT;
2158 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
2159 Pred = CmpInst::getInversePredicate(Pred);
2161 // Ok, everything looks ok to change the condition into an SLT or SGE and
2162 // delete the max calculation.
2164 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
2166 // Delete the max calculation instructions.
2167 Cond->replaceAllUsesWith(NewCond);
2168 CondUse->setUser(NewCond);
2169 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2170 Cond->eraseFromParent();
2171 Sel->eraseFromParent();
2172 if (Cmp->use_empty())
2173 Cmp->eraseFromParent();
2177 /// OptimizeShadowIV - If IV is used in a int-to-float cast
2178 /// inside the loop then try to eliminate the cast opeation.
2179 void LoopStrengthReduce::OptimizeShadowIV(Loop *L) {
2181 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
2182 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2185 for (unsigned Stride = 0, e = IU->StrideOrder.size(); Stride != e;
2187 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
2188 IU->IVUsesByStride.find(IU->StrideOrder[Stride]);
2189 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
2190 if (!isa<SCEVConstant>(SI->first))
2193 for (ilist<IVStrideUse>::iterator UI = SI->second->Users.begin(),
2194 E = SI->second->Users.end(); UI != E; /* empty */) {
2195 ilist<IVStrideUse>::iterator CandidateUI = UI;
2197 Instruction *ShadowUse = CandidateUI->getUser();
2198 const Type *DestTy = NULL;
2200 /* If shadow use is a int->float cast then insert a second IV
2201 to eliminate this cast.
2203 for (unsigned i = 0; i < n; ++i)
2209 for (unsigned i = 0; i < n; ++i, ++d)
2212 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser()))
2213 DestTy = UCast->getDestTy();
2214 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser()))
2215 DestTy = SCast->getDestTy();
2216 if (!DestTy) continue;
2219 // If target does not support DestTy natively then do not apply
2220 // this transformation.
2221 EVT DVT = TLI->getValueType(DestTy);
2222 if (!TLI->isTypeLegal(DVT)) continue;
2225 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
2227 if (PH->getNumIncomingValues() != 2) continue;
2229 const Type *SrcTy = PH->getType();
2230 int Mantissa = DestTy->getFPMantissaWidth();
2231 if (Mantissa == -1) continue;
2232 if ((int)SE->getTypeSizeInBits(SrcTy) > Mantissa)
2235 unsigned Entry, Latch;
2236 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
2244 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
2245 if (!Init) continue;
2246 Constant *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());
2248 BinaryOperator *Incr =
2249 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
2250 if (!Incr) continue;
2251 if (Incr->getOpcode() != Instruction::Add
2252 && Incr->getOpcode() != Instruction::Sub)
2255 /* Initialize new IV, double d = 0.0 in above example. */
2256 ConstantInt *C = NULL;
2257 if (Incr->getOperand(0) == PH)
2258 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
2259 else if (Incr->getOperand(1) == PH)
2260 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
2266 // Ignore negative constants, as the code below doesn't handle them
2267 // correctly. TODO: Remove this restriction.
2268 if (!C->getValue().isStrictlyPositive()) continue;
2270 /* Add new PHINode. */
2271 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);
2273 /* create new increment. '++d' in above example. */
2274 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
2275 BinaryOperator *NewIncr =
2276 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
2277 Instruction::FAdd : Instruction::FSub,
2278 NewPH, CFP, "IV.S.next.", Incr);
2280 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
2281 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
2283 /* Remove cast operation */
2284 ShadowUse->replaceAllUsesWith(NewPH);
2285 ShadowUse->eraseFromParent();
2292 /// OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
2293 /// uses in the loop, look to see if we can eliminate some, in favor of using
2294 /// common indvars for the different uses.
2295 void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
2296 // TODO: implement optzns here.
2298 OptimizeShadowIV(L);
2301 bool LoopStrengthReduce::StrideMightBeShared(const SCEV* Stride, Loop *L,
2303 int64_t SInt = cast<SCEVConstant>(Stride)->getValue()->getSExtValue();
2304 for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
2305 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
2306 IU->IVUsesByStride.find(IU->StrideOrder[i]);
2307 const SCEV *Share = SI->first;
2308 if (!isa<SCEVConstant>(SI->first) || Share == Stride)
2310 int64_t SSInt = cast<SCEVConstant>(Share)->getValue()->getSExtValue();
2312 return true; // This can definitely be reused.
2313 if (unsigned(abs64(SSInt)) < SInt || (SSInt % SInt) != 0)
2315 int64_t Scale = SSInt / SInt;
2316 bool AllUsesAreAddresses = true;
2317 bool AllUsesAreOutsideLoop = true;
2318 std::vector<BasedUser> UsersToProcess;
2319 const SCEV *CommonExprs = CollectIVUsers(SI->first, *SI->second, L,
2320 AllUsesAreAddresses,
2321 AllUsesAreOutsideLoop,
2323 if (AllUsesAreAddresses &&
2324 ValidScale(!CommonExprs->isZero(), Scale, UsersToProcess)) {
2327 // Any pre-inc iv use?
2328 IVUsersOfOneStride &StrideUses = *IU->IVUsesByStride[Share];
2329 for (ilist<IVStrideUse>::iterator I = StrideUses.Users.begin(),
2330 E = StrideUses.Users.end(); I != E; ++I) {
2331 if (!I->isUseOfPostIncrementedValue())
2339 /// isUsedByExitBranch - Return true if icmp is used by a loop terminating
2340 /// conditional branch or it's and / or with other conditions before being used
2341 /// as the condition.
2342 static bool isUsedByExitBranch(ICmpInst *Cond, Loop *L) {
2343 BasicBlock *CondBB = Cond->getParent();
2344 if (!L->isLoopExiting(CondBB))
2346 BranchInst *TermBr = dyn_cast<BranchInst>(CondBB->getTerminator());
2347 if (!TermBr || !TermBr->isConditional())
2350 Value *User = *Cond->use_begin();
2351 Instruction *UserInst = dyn_cast<Instruction>(User);
2353 (UserInst->getOpcode() == Instruction::And ||
2354 UserInst->getOpcode() == Instruction::Or)) {
2355 if (!UserInst->hasOneUse() || UserInst->getParent() != CondBB)
2357 User = *User->use_begin();
2358 UserInst = dyn_cast<Instruction>(User);
2360 return User == TermBr;
2363 static bool ShouldCountToZero(ICmpInst *Cond, IVStrideUse* &CondUse,
2364 ScalarEvolution *SE, Loop *L,
2365 const TargetLowering *TLI = 0) {
2366 if (!L->contains(Cond))
2369 if (!isa<SCEVConstant>(CondUse->getOffset()))
2372 // Handle only tests for equality for the moment.
2373 if (!Cond->isEquality() || !Cond->hasOneUse())
2375 if (!isUsedByExitBranch(Cond, L))
2378 Value *CondOp0 = Cond->getOperand(0);
2379 const SCEV *IV = SE->getSCEV(CondOp0);
2380 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2381 if (!AR || !AR->isAffine())
2384 const SCEVConstant *SC = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE));
2385 if (!SC || SC->getValue()->getSExtValue() < 0)
2386 // If it's already counting down, don't do anything.
2389 // If the RHS of the comparison is not an loop invariant, the rewrite
2390 // cannot be done. Also bail out if it's already comparing against a zero.
2391 // If we are checking this before cmp stride optimization, check if it's
2392 // comparing against a already legal immediate.
2393 Value *RHS = Cond->getOperand(1);
2394 ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS);
2395 if (!L->isLoopInvariant(RHS) ||
2396 (RHSC && RHSC->isZero()) ||
2397 (RHSC && TLI && TLI->isLegalICmpImmediate(RHSC->getSExtValue())))
2400 // Make sure the IV is only used for counting. Value may be preinc or
2401 // postinc; 2 uses in either case.
2402 if (!CondOp0->hasNUses(2))
2408 /// OptimizeLoopTermCond - Change loop terminating condition to use the
2409 /// postinc iv when possible.
2410 void LoopStrengthReduce::OptimizeLoopTermCond(Loop *L) {
2411 BasicBlock *LatchBlock = L->getLoopLatch();
2412 bool LatchExit = L->isLoopExiting(LatchBlock);
2413 SmallVector<BasicBlock*, 8> ExitingBlocks;
2414 L->getExitingBlocks(ExitingBlocks);
2416 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
2417 BasicBlock *ExitingBlock = ExitingBlocks[i];
2419 // Finally, get the terminating condition for the loop if possible. If we
2420 // can, we want to change it to use a post-incremented version of its
2421 // induction variable, to allow coalescing the live ranges for the IV into
2422 // one register value.
2424 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2427 // FIXME: Overly conservative, termination condition could be an 'or' etc..
2428 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
2431 // Search IVUsesByStride to find Cond's IVUse if there is one.
2432 IVStrideUse *CondUse = 0;
2433 const SCEV *CondStride = 0;
2434 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2435 if (!FindIVUserForCond(Cond, CondUse, CondStride))
2438 // If the latch block is exiting and it's not a single block loop, it's
2439 // not safe to use postinc iv in other exiting blocks. FIXME: overly
2440 // conservative? How about icmp stride optimization?
2441 bool UsePostInc = !(e > 1 && LatchExit && ExitingBlock != LatchBlock);
2442 if (UsePostInc && ExitingBlock != LatchBlock) {
2443 if (!Cond->hasOneUse())
2444 // See below, we don't want the condition to be cloned.
2447 // If exiting block is the latch block, we know it's safe and profitable
2448 // to transform the icmp to use post-inc iv. Otherwise do so only if it
2449 // would not reuse another iv and its iv would be reused by other uses.
2450 // We are optimizing for the case where the icmp is the only use of the
2452 IVUsersOfOneStride &StrideUses = *IU->IVUsesByStride[CondStride];
2453 for (ilist<IVStrideUse>::iterator I = StrideUses.Users.begin(),
2454 E = StrideUses.Users.end(); I != E; ++I) {
2455 if (I->getUser() == Cond)
2457 if (!I->isUseOfPostIncrementedValue()) {
2464 // If iv for the stride might be shared and any of the users use pre-inc
2465 // iv might be used, then it's not safe to use post-inc iv.
2467 isa<SCEVConstant>(CondStride) &&
2468 StrideMightBeShared(CondStride, L, true))
2472 // If the trip count is computed in terms of a max (due to ScalarEvolution
2473 // being unable to find a sufficient guard, for example), change the loop
2474 // comparison to use SLT or ULT instead of NE.
2475 Cond = OptimizeMax(L, Cond, CondUse);
2477 // If possible, change stride and operands of the compare instruction to
2478 // eliminate one stride. However, avoid rewriting the compare instruction
2479 // with an iv of new stride if it's likely the new stride uses will be
2480 // rewritten using the stride of the compare instruction.
2481 if (ExitingBlock == LatchBlock && isa<SCEVConstant>(CondStride)) {
2482 // If the condition stride is a constant and it's the only use, we might
2483 // want to optimize it first by turning it to count toward zero.
2484 if (!StrideMightBeShared(CondStride, L, false) &&
2485 !ShouldCountToZero(Cond, CondUse, SE, L, TLI))
2486 Cond = ChangeCompareStride(L, Cond, CondUse, CondStride);
2492 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
2495 // It's possible for the setcc instruction to be anywhere in the loop, and
2496 // possible for it to have multiple users. If it is not immediately before
2497 // the exiting block branch, move it.
2498 if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
2499 if (Cond->hasOneUse()) { // Condition has a single use, just move it.
2500 Cond->moveBefore(TermBr);
2502 // Otherwise, clone the terminating condition and insert into the
2504 Cond = cast<ICmpInst>(Cond->clone());
2505 Cond->setName(L->getHeader()->getName() + ".termcond");
2506 ExitingBlock->getInstList().insert(TermBr, Cond);
2508 // Clone the IVUse, as the old use still exists!
2509 IU->IVUsesByStride[CondStride]->addUser(CondUse->getOffset(), Cond,
2510 CondUse->getOperandValToReplace());
2511 CondUse = &IU->IVUsesByStride[CondStride]->Users.back();
2515 // If we get to here, we know that we can transform the setcc instruction to
2516 // use the post-incremented version of the IV, allowing us to coalesce the
2517 // live ranges for the IV correctly.
2518 CondUse->setOffset(SE->getMinusSCEV(CondUse->getOffset(), CondStride));
2519 CondUse->setIsUseOfPostIncrementedValue(true);
2526 bool LoopStrengthReduce::OptimizeLoopCountIVOfStride(const SCEV* &Stride,
2527 IVStrideUse* &CondUse,
2529 // If the only use is an icmp of a loop exiting conditional branch, then
2530 // attempt the optimization.
2531 BasedUser User = BasedUser(*CondUse, SE);
2532 assert(isa<ICmpInst>(User.Inst) && "Expecting an ICMPInst!");
2533 ICmpInst *Cond = cast<ICmpInst>(User.Inst);
2535 // Less strict check now that compare stride optimization is done.
2536 if (!ShouldCountToZero(Cond, CondUse, SE, L))
2539 Value *CondOp0 = Cond->getOperand(0);
2540 PHINode *PHIExpr = dyn_cast<PHINode>(CondOp0);
2543 // Value tested is postinc. Find the phi node.
2544 Incr = dyn_cast<BinaryOperator>(CondOp0);
2545 // FIXME: Just use User.OperandValToReplace here?
2546 if (!Incr || Incr->getOpcode() != Instruction::Add)
2549 PHIExpr = dyn_cast<PHINode>(Incr->getOperand(0));
2552 // 1 use for preinc value, the increment.
2553 if (!PHIExpr->hasOneUse())
2556 assert(isa<PHINode>(CondOp0) &&
2557 "Unexpected loop exiting counting instruction sequence!");
2558 PHIExpr = cast<PHINode>(CondOp0);
2559 // Value tested is preinc. Find the increment.
2560 // A CmpInst is not a BinaryOperator; we depend on this.
2561 Instruction::use_iterator UI = PHIExpr->use_begin();
2562 Incr = dyn_cast<BinaryOperator>(UI);
2564 Incr = dyn_cast<BinaryOperator>(++UI);
2565 // One use for postinc value, the phi. Unnecessarily conservative?
2566 if (!Incr || !Incr->hasOneUse() || Incr->getOpcode() != Instruction::Add)
2570 // Replace the increment with a decrement.
2571 DEBUG(dbgs() << "LSR: Examining use ");
2572 DEBUG(WriteAsOperand(dbgs(), CondOp0, /*PrintType=*/false));
2573 DEBUG(dbgs() << " in Inst: " << *Cond << '\n');
2574 BinaryOperator *Decr = BinaryOperator::Create(Instruction::Sub,
2575 Incr->getOperand(0), Incr->getOperand(1), "tmp", Incr);
2576 Incr->replaceAllUsesWith(Decr);
2577 Incr->eraseFromParent();
2579 // Substitute endval-startval for the original startval, and 0 for the
2580 // original endval. Since we're only testing for equality this is OK even
2581 // if the computation wraps around.
2582 BasicBlock *Preheader = L->getLoopPreheader();
2583 Instruction *PreInsertPt = Preheader->getTerminator();
2584 unsigned InBlock = L->contains(PHIExpr->getIncomingBlock(0)) ? 1 : 0;
2585 Value *StartVal = PHIExpr->getIncomingValue(InBlock);
2586 Value *EndVal = Cond->getOperand(1);
2587 DEBUG(dbgs() << " Optimize loop counting iv to count down ["
2588 << *EndVal << " .. " << *StartVal << "]\n");
2590 // FIXME: check for case where both are constant.
2591 Constant* Zero = ConstantInt::get(Cond->getOperand(1)->getType(), 0);
2592 BinaryOperator *NewStartVal = BinaryOperator::Create(Instruction::Sub,
2593 EndVal, StartVal, "tmp", PreInsertPt);
2594 PHIExpr->setIncomingValue(InBlock, NewStartVal);
2595 Cond->setOperand(1, Zero);
2596 DEBUG(dbgs() << " New icmp: " << *Cond << "\n");
2598 int64_t SInt = cast<SCEVConstant>(Stride)->getValue()->getSExtValue();
2599 const SCEV *NewStride = 0;
2601 for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
2602 const SCEV *OldStride = IU->StrideOrder[i];
2603 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(OldStride))
2604 if (SC->getValue()->getSExtValue() == -SInt) {
2606 NewStride = OldStride;
2612 NewStride = SE->getIntegerSCEV(-SInt, Stride->getType());
2613 IU->AddUser(NewStride, CondUse->getOffset(), Cond, Cond->getOperand(0));
2614 IU->IVUsesByStride[Stride]->removeUser(CondUse);
2616 CondUse = &IU->IVUsesByStride[NewStride]->Users.back();
2624 /// OptimizeLoopCountIV - If, after all sharing of IVs, the IV used for deciding
2625 /// when to exit the loop is used only for that purpose, try to rearrange things
2626 /// so it counts down to a test against zero.
2627 bool LoopStrengthReduce::OptimizeLoopCountIV(Loop *L) {
2628 bool ThisChanged = false;
2629 for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
2630 const SCEV *Stride = IU->StrideOrder[i];
2631 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
2632 IU->IVUsesByStride.find(Stride);
2633 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
2634 // FIXME: Generalize to non-affine IV's.
2635 if (!SI->first->isLoopInvariant(L))
2637 // If stride is a constant and it has an icmpinst use, check if we can
2638 // optimize the loop to count down.
2639 if (isa<SCEVConstant>(Stride) && SI->second->Users.size() == 1) {
2640 Instruction *User = SI->second->Users.begin()->getUser();
2641 if (!isa<ICmpInst>(User))
2643 const SCEV *CondStride = Stride;
2644 IVStrideUse *Use = &*SI->second->Users.begin();
2645 if (!OptimizeLoopCountIVOfStride(CondStride, Use, L))
2649 // Now check if it's possible to reuse this iv for other stride uses.
2650 for (unsigned j = 0, ee = IU->StrideOrder.size(); j != ee; ++j) {
2651 const SCEV *SStride = IU->StrideOrder[j];
2652 if (SStride == CondStride)
2654 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SII =
2655 IU->IVUsesByStride.find(SStride);
2656 assert(SII != IU->IVUsesByStride.end() && "Stride doesn't exist!");
2657 // FIXME: Generalize to non-affine IV's.
2658 if (!SII->first->isLoopInvariant(L))
2660 // FIXME: Rewrite other stride using CondStride.
2665 Changed |= ThisChanged;
2669 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager &LPM) {
2670 IU = &getAnalysis<IVUsers>();
2671 SE = &getAnalysis<ScalarEvolution>();
2674 // If LoopSimplify form is not available, stay out of trouble.
2675 if (!L->getLoopPreheader() || !L->getLoopLatch())
2678 if (!IU->IVUsesByStride.empty()) {
2679 DEBUG(dbgs() << "\nLSR on \"" << L->getHeader()->getParent()->getName()
2683 // Sort the StrideOrder so we process larger strides first.
2684 std::stable_sort(IU->StrideOrder.begin(), IU->StrideOrder.end(),
2687 // Optimize induction variables. Some indvar uses can be transformed to use
2688 // strides that will be needed for other purposes. A common example of this
2689 // is the exit test for the loop, which can often be rewritten to use the
2690 // computation of some other indvar to decide when to terminate the loop.
2693 // Change loop terminating condition to use the postinc iv when possible
2694 // and optimize loop terminating compare. FIXME: Move this after
2695 // StrengthReduceIVUsersOfStride?
2696 OptimizeLoopTermCond(L);
2698 // FIXME: We can shrink overlarge IV's here. e.g. if the code has
2699 // computation in i64 values and the target doesn't support i64, demote
2700 // the computation to 32-bit if safe.
2702 // FIXME: Attempt to reuse values across multiple IV's. In particular, we
2703 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should
2704 // be codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC.
2705 // Need to be careful that IV's are all the same type. Only works for
2706 // intptr_t indvars.
2708 // IVsByStride keeps IVs for one particular loop.
2709 assert(IVsByStride.empty() && "Stale entries in IVsByStride?");
2711 StrengthReduceIVUsers(L);
2713 // After all sharing is done, see if we can adjust the loop to test against
2714 // zero instead of counting up to a maximum. This is usually faster.
2715 OptimizeLoopCountIV(L);
2717 // We're done analyzing this loop; release all the state we built up for it.
2718 IVsByStride.clear();
2720 // Clean up after ourselves
2721 DeleteTriviallyDeadInstructions();
2724 // At this point, it is worth checking to see if any recurrence PHIs are also
2725 // dead, so that we can remove them as well.
2726 Changed |= DeleteDeadPHIs(L->getHeader());