1 //===- LoopStrengthReduce.cpp - Strength Reduce IVs in Loops --------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This transformation analyzes and transforms the induction variables (and
11 // computations derived from them) into forms suitable for efficient execution
14 // This pass performs a strength reduction on array references inside loops that
15 // have as one or more of their components the loop induction variable, it
16 // rewrites expressions to take advantage of scaled-index addressing modes
17 // available on the target, and it performs a variety of other optimizations
18 // related to loop induction variables.
20 //===----------------------------------------------------------------------===//
22 #define DEBUG_TYPE "loop-reduce"
23 #include "llvm/Transforms/Scalar.h"
24 #include "llvm/Constants.h"
25 #include "llvm/Instructions.h"
26 #include "llvm/IntrinsicInst.h"
27 #include "llvm/Type.h"
28 #include "llvm/DerivedTypes.h"
29 #include "llvm/Analysis/IVUsers.h"
30 #include "llvm/Analysis/LoopPass.h"
31 #include "llvm/Analysis/ScalarEvolutionExpander.h"
32 #include "llvm/Transforms/Utils/AddrModeMatcher.h"
33 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
34 #include "llvm/Transforms/Utils/Local.h"
35 #include "llvm/ADT/Statistic.h"
36 #include "llvm/Support/CFG.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Support/CommandLine.h"
39 #include "llvm/Support/ValueHandle.h"
40 #include "llvm/Support/raw_ostream.h"
41 #include "llvm/Target/TargetLowering.h"
45 STATISTIC(NumReduced , "Number of IV uses strength reduced");
46 STATISTIC(NumInserted, "Number of PHIs inserted");
47 STATISTIC(NumVariable, "Number of PHIs with variable strides");
48 STATISTIC(NumEliminated, "Number of strides eliminated");
49 STATISTIC(NumShadow, "Number of Shadow IVs optimized");
50 STATISTIC(NumImmSunk, "Number of common expr immediates sunk into uses");
51 STATISTIC(NumLoopCond, "Number of loop terminating conds optimized");
52 STATISTIC(NumCountZero, "Number of count iv optimized to count toward zero");
54 static cl::opt<bool> EnableFullLSRMode("enable-full-lsr",
62 /// IVInfo - This structure keeps track of one IV expression inserted during
63 /// StrengthReduceStridedIVUsers. It contains the stride, the common base, as
64 /// well as the PHI node and increment value created for rewrite.
70 IVExpr(const SCEV *const stride, const SCEV *const base, PHINode *phi)
71 : Stride(stride), Base(base), PHI(phi) {}
74 /// IVsOfOneStride - This structure keeps track of all IV expression inserted
75 /// during StrengthReduceStridedIVUsers for a particular stride of the IV.
76 struct IVsOfOneStride {
77 std::vector<IVExpr> IVs;
79 void addIV(const SCEV *const Stride, const SCEV *const Base, PHINode *PHI) {
80 IVs.push_back(IVExpr(Stride, Base, PHI));
84 class LoopStrengthReduce : public LoopPass {
89 /// IVsByStride - Keep track of all IVs that have been inserted for a
90 /// particular stride.
91 std::map<const SCEV *, IVsOfOneStride> IVsByStride;
93 /// DeadInsts - Keep track of instructions we may have made dead, so that
94 /// we can remove them after we are done working.
95 SmallVector<WeakVH, 16> DeadInsts;
97 /// TLI - Keep a pointer of a TargetLowering to consult for determining
98 /// transformation profitability.
99 const TargetLowering *TLI;
102 static char ID; // Pass ID, replacement for typeid
103 explicit LoopStrengthReduce(const TargetLowering *tli = NULL) :
104 LoopPass(&ID), TLI(tli) {}
106 bool runOnLoop(Loop *L, LPPassManager &LPM);
108 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
109 // We split critical edges, so we change the CFG. However, we do update
110 // many analyses if they are around.
111 AU.addPreservedID(LoopSimplifyID);
112 AU.addPreserved("loops");
113 AU.addPreserved("domfrontier");
114 AU.addPreserved("domtree");
116 AU.addRequiredID(LoopSimplifyID);
117 AU.addRequired<ScalarEvolution>();
118 AU.addPreserved<ScalarEvolution>();
119 AU.addRequired<IVUsers>();
120 AU.addPreserved<IVUsers>();
124 void OptimizeIndvars(Loop *L);
126 /// OptimizeLoopTermCond - Change loop terminating condition to use the
127 /// postinc iv when possible.
128 void OptimizeLoopTermCond(Loop *L);
130 /// OptimizeShadowIV - If IV is used in a int-to-float cast
131 /// inside the loop then try to eliminate the cast opeation.
132 void OptimizeShadowIV(Loop *L);
134 /// OptimizeMax - Rewrite the loop's terminating condition
135 /// if it uses a max computation.
136 ICmpInst *OptimizeMax(Loop *L, ICmpInst *Cond,
137 IVStrideUse* &CondUse);
139 /// OptimizeLoopCountIV - If, after all sharing of IVs, the IV used for
140 /// deciding when to exit the loop is used only for that purpose, try to
141 /// rearrange things so it counts down to a test against zero.
142 bool OptimizeLoopCountIV(Loop *L);
143 bool OptimizeLoopCountIVOfStride(const SCEV* &Stride,
144 IVStrideUse* &CondUse, Loop *L);
146 /// StrengthReduceIVUsersOfStride - Strength reduce all of the users of a
147 /// single stride of IV. All of the users may have different starting
148 /// values, and this may not be the only stride.
149 void StrengthReduceIVUsersOfStride(const SCEV *const &Stride,
150 IVUsersOfOneStride &Uses,
152 void StrengthReduceIVUsers(Loop *L);
154 ICmpInst *ChangeCompareStride(Loop *L, ICmpInst *Cond,
155 IVStrideUse* &CondUse,
156 const SCEV* &CondStride,
157 bool PostPass = false);
159 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse,
160 const SCEV* &CondStride);
161 bool RequiresTypeConversion(const Type *Ty, const Type *NewTy);
162 const SCEV *CheckForIVReuse(bool, bool, bool, const SCEV *const&,
163 IVExpr&, const Type*,
164 const std::vector<BasedUser>& UsersToProcess);
165 bool ValidScale(bool, int64_t,
166 const std::vector<BasedUser>& UsersToProcess);
167 bool ValidOffset(bool, int64_t, int64_t,
168 const std::vector<BasedUser>& UsersToProcess);
169 const SCEV *CollectIVUsers(const SCEV *const &Stride,
170 IVUsersOfOneStride &Uses,
172 bool &AllUsesAreAddresses,
173 bool &AllUsesAreOutsideLoop,
174 std::vector<BasedUser> &UsersToProcess);
175 bool StrideMightBeShared(const SCEV *Stride, Loop *L, bool CheckPreInc);
176 bool ShouldUseFullStrengthReductionMode(
177 const std::vector<BasedUser> &UsersToProcess,
179 bool AllUsesAreAddresses,
181 void PrepareToStrengthReduceFully(
182 std::vector<BasedUser> &UsersToProcess,
184 const SCEV *CommonExprs,
186 SCEVExpander &PreheaderRewriter);
187 void PrepareToStrengthReduceFromSmallerStride(
188 std::vector<BasedUser> &UsersToProcess,
190 const IVExpr &ReuseIV,
191 Instruction *PreInsertPt);
192 void PrepareToStrengthReduceWithNewPhi(
193 std::vector<BasedUser> &UsersToProcess,
195 const SCEV *CommonExprs,
197 Instruction *IVIncInsertPt,
199 SCEVExpander &PreheaderRewriter);
201 void DeleteTriviallyDeadInstructions();
205 char LoopStrengthReduce::ID = 0;
206 static RegisterPass<LoopStrengthReduce>
207 X("loop-reduce", "Loop Strength Reduction");
209 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
210 return new LoopStrengthReduce(TLI);
213 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
214 /// specified set are trivially dead, delete them and see if this makes any of
215 /// their operands subsequently dead.
216 void LoopStrengthReduce::DeleteTriviallyDeadInstructions() {
217 if (DeadInsts.empty()) return;
219 while (!DeadInsts.empty()) {
220 Instruction *I = dyn_cast_or_null<Instruction>(DeadInsts.pop_back_val());
222 if (I == 0 || !isInstructionTriviallyDead(I))
225 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
226 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
229 DeadInsts.push_back(U);
232 I->eraseFromParent();
237 /// containsAddRecFromDifferentLoop - Determine whether expression S involves a
238 /// subexpression that is an AddRec from a loop other than L. An outer loop
239 /// of L is OK, but not an inner loop nor a disjoint loop.
240 static bool containsAddRecFromDifferentLoop(const SCEV *S, Loop *L) {
241 // This is very common, put it first.
242 if (isa<SCEVConstant>(S))
244 if (const SCEVCommutativeExpr *AE = dyn_cast<SCEVCommutativeExpr>(S)) {
245 for (unsigned int i=0; i< AE->getNumOperands(); i++)
246 if (containsAddRecFromDifferentLoop(AE->getOperand(i), L))
250 if (const SCEVAddRecExpr *AE = dyn_cast<SCEVAddRecExpr>(S)) {
251 if (const Loop *newLoop = AE->getLoop()) {
254 // if newLoop is an outer loop of L, this is OK.
255 if (!newLoop->contains(L->getHeader()))
260 if (const SCEVUDivExpr *DE = dyn_cast<SCEVUDivExpr>(S))
261 return containsAddRecFromDifferentLoop(DE->getLHS(), L) ||
262 containsAddRecFromDifferentLoop(DE->getRHS(), L);
264 // SCEVSDivExpr has been backed out temporarily, but will be back; we'll
265 // need this when it is.
266 if (const SCEVSDivExpr *DE = dyn_cast<SCEVSDivExpr>(S))
267 return containsAddRecFromDifferentLoop(DE->getLHS(), L) ||
268 containsAddRecFromDifferentLoop(DE->getRHS(), L);
270 if (const SCEVCastExpr *CE = dyn_cast<SCEVCastExpr>(S))
271 return containsAddRecFromDifferentLoop(CE->getOperand(), L);
275 /// isAddressUse - Returns true if the specified instruction is using the
276 /// specified value as an address.
277 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
278 bool isAddress = isa<LoadInst>(Inst);
279 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
280 if (SI->getOperand(1) == OperandVal)
282 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
283 // Addressing modes can also be folded into prefetches and a variety
285 switch (II->getIntrinsicID()) {
287 case Intrinsic::prefetch:
288 case Intrinsic::x86_sse2_loadu_dq:
289 case Intrinsic::x86_sse2_loadu_pd:
290 case Intrinsic::x86_sse_loadu_ps:
291 case Intrinsic::x86_sse_storeu_ps:
292 case Intrinsic::x86_sse2_storeu_pd:
293 case Intrinsic::x86_sse2_storeu_dq:
294 case Intrinsic::x86_sse2_storel_dq:
295 if (II->getOperand(1) == OperandVal)
303 /// getAccessType - Return the type of the memory being accessed.
304 static const Type *getAccessType(const Instruction *Inst) {
305 const Type *AccessTy = Inst->getType();
306 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
307 AccessTy = SI->getOperand(0)->getType();
308 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
309 // Addressing modes can also be folded into prefetches and a variety
311 switch (II->getIntrinsicID()) {
313 case Intrinsic::x86_sse_storeu_ps:
314 case Intrinsic::x86_sse2_storeu_pd:
315 case Intrinsic::x86_sse2_storeu_dq:
316 case Intrinsic::x86_sse2_storel_dq:
317 AccessTy = II->getOperand(1)->getType();
325 /// BasedUser - For a particular base value, keep information about how we've
326 /// partitioned the expression so far.
328 /// SE - The current ScalarEvolution object.
331 /// Base - The Base value for the PHI node that needs to be inserted for
332 /// this use. As the use is processed, information gets moved from this
333 /// field to the Imm field (below). BasedUser values are sorted by this
337 /// Inst - The instruction using the induction variable.
340 /// OperandValToReplace - The operand value of Inst to replace with the
342 Value *OperandValToReplace;
344 /// Imm - The immediate value that should be added to the base immediately
345 /// before Inst, because it will be folded into the imm field of the
346 /// instruction. This is also sometimes used for loop-variant values that
347 /// must be added inside the loop.
350 /// Phi - The induction variable that performs the striding that
351 /// should be used for this user.
354 // isUseOfPostIncrementedValue - True if this should use the
355 // post-incremented version of this IV, not the preincremented version.
356 // This can only be set in special cases, such as the terminating setcc
357 // instruction for a loop and uses outside the loop that are dominated by
359 bool isUseOfPostIncrementedValue;
361 BasedUser(IVStrideUse &IVSU, ScalarEvolution *se)
362 : SE(se), Base(IVSU.getOffset()), Inst(IVSU.getUser()),
363 OperandValToReplace(IVSU.getOperandValToReplace()),
364 Imm(SE->getIntegerSCEV(0, Base->getType())),
365 isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue()) {}
367 // Once we rewrite the code to insert the new IVs we want, update the
368 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
370 void RewriteInstructionToUseNewBase(const SCEV *const &NewBase,
371 Instruction *InsertPt,
372 SCEVExpander &Rewriter, Loop *L, Pass *P,
373 SmallVectorImpl<WeakVH> &DeadInsts);
375 Value *InsertCodeForBaseAtPosition(const SCEV *const &NewBase,
377 SCEVExpander &Rewriter,
383 void BasedUser::dump() const {
384 errs() << " Base=" << *Base;
385 errs() << " Imm=" << *Imm;
386 errs() << " Inst: " << *Inst;
389 Value *BasedUser::InsertCodeForBaseAtPosition(const SCEV *const &NewBase,
391 SCEVExpander &Rewriter,
393 Value *Base = Rewriter.expandCodeFor(NewBase, 0, IP);
395 // Wrap the base in a SCEVUnknown so that ScalarEvolution doesn't try to
397 const SCEV *NewValSCEV = SE->getUnknown(Base);
399 // Always emit the immediate into the same block as the user.
400 NewValSCEV = SE->getAddExpr(NewValSCEV, Imm);
402 return Rewriter.expandCodeFor(NewValSCEV, Ty, IP);
406 // Once we rewrite the code to insert the new IVs we want, update the
407 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
408 // to it. NewBasePt is the last instruction which contributes to the
409 // value of NewBase in the case that it's a diffferent instruction from
410 // the PHI that NewBase is computed from, or null otherwise.
412 void BasedUser::RewriteInstructionToUseNewBase(const SCEV *const &NewBase,
413 Instruction *NewBasePt,
414 SCEVExpander &Rewriter, Loop *L, Pass *P,
415 SmallVectorImpl<WeakVH> &DeadInsts) {
416 if (!isa<PHINode>(Inst)) {
417 // By default, insert code at the user instruction.
418 BasicBlock::iterator InsertPt = Inst;
420 // However, if the Operand is itself an instruction, the (potentially
421 // complex) inserted code may be shared by many users. Because of this, we
422 // want to emit code for the computation of the operand right before its old
423 // computation. This is usually safe, because we obviously used to use the
424 // computation when it was computed in its current block. However, in some
425 // cases (e.g. use of a post-incremented induction variable) the NewBase
426 // value will be pinned to live somewhere after the original computation.
427 // In this case, we have to back off.
429 // If this is a use outside the loop (which means after, since it is based
430 // on a loop indvar) we use the post-incremented value, so that we don't
431 // artificially make the preinc value live out the bottom of the loop.
432 if (!isUseOfPostIncrementedValue && L->contains(Inst->getParent())) {
433 if (NewBasePt && isa<PHINode>(OperandValToReplace)) {
434 InsertPt = NewBasePt;
436 } else if (Instruction *OpInst
437 = dyn_cast<Instruction>(OperandValToReplace)) {
439 while (isa<PHINode>(InsertPt)) ++InsertPt;
442 Value *NewVal = InsertCodeForBaseAtPosition(NewBase,
443 OperandValToReplace->getType(),
445 // Replace the use of the operand Value with the new Phi we just created.
446 Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
448 DEBUG(errs() << " Replacing with ");
449 DEBUG(WriteAsOperand(errs(), NewVal, /*PrintType=*/false));
450 DEBUG(errs() << ", which has value " << *NewBase << " plus IMM "
455 // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm
456 // expression into each operand block that uses it. Note that PHI nodes can
457 // have multiple entries for the same predecessor. We use a map to make sure
458 // that a PHI node only has a single Value* for each predecessor (which also
459 // prevents us from inserting duplicate code in some blocks).
460 DenseMap<BasicBlock*, Value*> InsertedCode;
461 PHINode *PN = cast<PHINode>(Inst);
462 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
463 if (PN->getIncomingValue(i) == OperandValToReplace) {
464 // If the original expression is outside the loop, put the replacement
465 // code in the same place as the original expression,
466 // which need not be an immediate predecessor of this PHI. This way we
467 // need only one copy of it even if it is referenced multiple times in
468 // the PHI. We don't do this when the original expression is inside the
469 // loop because multiple copies sometimes do useful sinking of code in
471 Instruction *OldLoc = dyn_cast<Instruction>(OperandValToReplace);
472 BasicBlock *PHIPred = PN->getIncomingBlock(i);
473 if (L->contains(OldLoc->getParent())) {
474 // If this is a critical edge, split the edge so that we do not insert
475 // the code on all predecessor/successor paths. We do this unless this
476 // is the canonical backedge for this loop, as this can make some
477 // inserted code be in an illegal position.
478 if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 &&
479 !isa<IndirectBrInst>(PHIPred->getTerminator()) &&
480 (PN->getParent() != L->getHeader() || !L->contains(PHIPred))) {
482 // First step, split the critical edge.
483 BasicBlock *NewBB = SplitCriticalEdge(PHIPred, PN->getParent(),
486 // Next step: move the basic block. In particular, if the PHI node
487 // is outside of the loop, and PredTI is in the loop, we want to
488 // move the block to be immediately before the PHI block, not
489 // immediately after PredTI.
490 if (L->contains(PHIPred) && !L->contains(PN->getParent()))
491 NewBB->moveBefore(PN->getParent());
493 // Splitting the edge can reduce the number of PHI entries we have.
494 e = PN->getNumIncomingValues();
496 i = PN->getBasicBlockIndex(PHIPred);
499 Value *&Code = InsertedCode[PHIPred];
501 // Insert the code into the end of the predecessor block.
502 Instruction *InsertPt = (L->contains(OldLoc->getParent())) ?
503 PHIPred->getTerminator() :
504 OldLoc->getParent()->getTerminator();
505 Code = InsertCodeForBaseAtPosition(NewBase, PN->getType(),
508 DEBUG(errs() << " Changing PHI use to ");
509 DEBUG(WriteAsOperand(errs(), Code, /*PrintType=*/false));
510 DEBUG(errs() << ", which has value " << *NewBase << " plus IMM "
514 // Replace the use of the operand Value with the new Phi we just created.
515 PN->setIncomingValue(i, Code);
520 // PHI node might have become a constant value after SplitCriticalEdge.
521 DeadInsts.push_back(Inst);
525 /// fitsInAddressMode - Return true if V can be subsumed within an addressing
526 /// mode, and does not need to be put in a register first.
527 static bool fitsInAddressMode(const SCEV *const &V, const Type *AccessTy,
528 const TargetLowering *TLI, bool HasBaseReg) {
529 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
530 int64_t VC = SC->getValue()->getSExtValue();
532 TargetLowering::AddrMode AM;
534 AM.HasBaseReg = HasBaseReg;
535 return TLI->isLegalAddressingMode(AM, AccessTy);
537 // Defaults to PPC. PPC allows a sign-extended 16-bit immediate field.
538 return (VC > -(1 << 16) && VC < (1 << 16)-1);
542 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
543 if (GlobalValue *GV = dyn_cast<GlobalValue>(SU->getValue())) {
545 TargetLowering::AddrMode AM;
547 AM.HasBaseReg = HasBaseReg;
548 return TLI->isLegalAddressingMode(AM, AccessTy);
550 // Default: assume global addresses are not legal.
557 /// MoveLoopVariantsToImmediateField - Move any subexpressions from Val that are
558 /// loop varying to the Imm operand.
559 static void MoveLoopVariantsToImmediateField(const SCEV *&Val, const SCEV *&Imm,
560 Loop *L, ScalarEvolution *SE) {
561 if (Val->isLoopInvariant(L)) return; // Nothing to do.
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 if (!SAE->getOperand(i)->isLoopInvariant(L)) {
569 // If this is a loop-variant expression, it must stay in the immediate
570 // field of the expression.
571 Imm = SE->getAddExpr(Imm, SAE->getOperand(i));
573 NewOps.push_back(SAE->getOperand(i));
577 Val = SE->getIntegerSCEV(0, Val->getType());
579 Val = SE->getAddExpr(NewOps);
580 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
581 // Try to pull immediates out of the start value of nested addrec's.
582 const SCEV *Start = SARE->getStart();
583 MoveLoopVariantsToImmediateField(Start, Imm, L, SE);
585 SmallVector<const SCEV *, 4> Ops(SARE->op_begin(), SARE->op_end());
587 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
589 // Otherwise, all of Val is variant, move the whole thing over.
590 Imm = SE->getAddExpr(Imm, Val);
591 Val = SE->getIntegerSCEV(0, Val->getType());
596 /// MoveImmediateValues - Look at Val, and pull out any additions of constants
597 /// that can fit into the immediate field of instructions in the target.
598 /// Accumulate these immediate values into the Imm value.
599 static void MoveImmediateValues(const TargetLowering *TLI,
600 const Type *AccessTy,
601 const SCEV *&Val, const SCEV *&Imm,
602 bool isAddress, Loop *L,
603 ScalarEvolution *SE) {
604 if (const SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
605 SmallVector<const SCEV *, 4> NewOps;
606 NewOps.reserve(SAE->getNumOperands());
608 for (unsigned i = 0; i != SAE->getNumOperands(); ++i) {
609 const SCEV *NewOp = SAE->getOperand(i);
610 MoveImmediateValues(TLI, AccessTy, NewOp, Imm, isAddress, L, SE);
612 if (!NewOp->isLoopInvariant(L)) {
613 // If this is a loop-variant expression, it must stay in the immediate
614 // field of the expression.
615 Imm = SE->getAddExpr(Imm, NewOp);
617 NewOps.push_back(NewOp);
622 Val = SE->getIntegerSCEV(0, Val->getType());
624 Val = SE->getAddExpr(NewOps);
626 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
627 // Try to pull immediates out of the start value of nested addrec's.
628 const SCEV *Start = SARE->getStart();
629 MoveImmediateValues(TLI, AccessTy, Start, Imm, isAddress, L, SE);
631 if (Start != SARE->getStart()) {
632 SmallVector<const SCEV *, 4> Ops(SARE->op_begin(), SARE->op_end());
634 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
637 } else if (const SCEVMulExpr *SME = dyn_cast<SCEVMulExpr>(Val)) {
638 // Transform "8 * (4 + v)" -> "32 + 8*V" if "32" fits in the immed field.
640 fitsInAddressMode(SME->getOperand(0), AccessTy, TLI, false) &&
641 SME->getNumOperands() == 2 && SME->isLoopInvariant(L)) {
643 const SCEV *SubImm = SE->getIntegerSCEV(0, Val->getType());
644 const SCEV *NewOp = SME->getOperand(1);
645 MoveImmediateValues(TLI, AccessTy, NewOp, SubImm, isAddress, L, SE);
647 // If we extracted something out of the subexpressions, see if we can
649 if (NewOp != SME->getOperand(1)) {
650 // Scale SubImm up by "8". If the result is a target constant, we are
652 SubImm = SE->getMulExpr(SubImm, SME->getOperand(0));
653 if (fitsInAddressMode(SubImm, AccessTy, TLI, false)) {
654 // Accumulate the immediate.
655 Imm = SE->getAddExpr(Imm, SubImm);
657 // Update what is left of 'Val'.
658 Val = SE->getMulExpr(SME->getOperand(0), NewOp);
665 // Loop-variant expressions must stay in the immediate field of the
667 if ((isAddress && fitsInAddressMode(Val, AccessTy, TLI, false)) ||
668 !Val->isLoopInvariant(L)) {
669 Imm = SE->getAddExpr(Imm, Val);
670 Val = SE->getIntegerSCEV(0, Val->getType());
674 // Otherwise, no immediates to move.
677 static void MoveImmediateValues(const TargetLowering *TLI,
679 const SCEV *&Val, const SCEV *&Imm,
680 bool isAddress, Loop *L,
681 ScalarEvolution *SE) {
682 const Type *AccessTy = getAccessType(User);
683 MoveImmediateValues(TLI, AccessTy, Val, Imm, isAddress, L, SE);
686 /// SeparateSubExprs - Decompose Expr into all of the subexpressions that are
687 /// added together. This is used to reassociate common addition subexprs
688 /// together for maximal sharing when rewriting bases.
689 static void SeparateSubExprs(SmallVector<const SCEV *, 16> &SubExprs,
691 ScalarEvolution *SE) {
692 if (const SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) {
693 for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
694 SeparateSubExprs(SubExprs, AE->getOperand(j), SE);
695 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) {
696 const SCEV *Zero = SE->getIntegerSCEV(0, Expr->getType());
697 if (SARE->getOperand(0) == Zero) {
698 SubExprs.push_back(Expr);
700 // Compute the addrec with zero as its base.
701 SmallVector<const SCEV *, 4> Ops(SARE->op_begin(), SARE->op_end());
702 Ops[0] = Zero; // Start with zero base.
703 SubExprs.push_back(SE->getAddRecExpr(Ops, SARE->getLoop()));
706 SeparateSubExprs(SubExprs, SARE->getOperand(0), SE);
708 } else if (!Expr->isZero()) {
710 SubExprs.push_back(Expr);
714 // This is logically local to the following function, but C++ says we have
715 // to make it file scope.
716 struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; };
718 /// RemoveCommonExpressionsFromUseBases - Look through all of the Bases of all
719 /// the Uses, removing any common subexpressions, except that if all such
720 /// subexpressions can be folded into an addressing mode for all uses inside
721 /// the loop (this case is referred to as "free" in comments herein) we do
722 /// not remove anything. This looks for things like (a+b+c) and
723 /// (a+c+d) and computes the common (a+c) subexpression. The common expression
724 /// is *removed* from the Bases and returned.
726 RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses,
727 ScalarEvolution *SE, Loop *L,
728 const TargetLowering *TLI) {
729 unsigned NumUses = Uses.size();
731 // Only one use? This is a very common case, so we handle it specially and
733 const SCEV *Zero = SE->getIntegerSCEV(0, Uses[0].Base->getType());
734 const SCEV *Result = Zero;
735 const SCEV *FreeResult = Zero;
737 // If the use is inside the loop, use its base, regardless of what it is:
738 // it is clearly shared across all the IV's. If the use is outside the loop
739 // (which means after it) we don't want to factor anything *into* the loop,
740 // so just use 0 as the base.
741 if (L->contains(Uses[0].Inst->getParent()))
742 std::swap(Result, Uses[0].Base);
746 // To find common subexpressions, count how many of Uses use each expression.
747 // If any subexpressions are used Uses.size() times, they are common.
748 // Also track whether all uses of each expression can be moved into an
749 // an addressing mode "for free"; such expressions are left within the loop.
750 // struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; };
751 std::map<const SCEV *, SubExprUseData> SubExpressionUseData;
753 // UniqueSubExprs - Keep track of all of the subexpressions we see in the
754 // order we see them.
755 SmallVector<const SCEV *, 16> UniqueSubExprs;
757 SmallVector<const SCEV *, 16> SubExprs;
758 unsigned NumUsesInsideLoop = 0;
759 for (unsigned i = 0; i != NumUses; ++i) {
760 // If the user is outside the loop, just ignore it for base computation.
761 // Since the user is outside the loop, it must be *after* the loop (if it
762 // were before, it could not be based on the loop IV). We don't want users
763 // after the loop to affect base computation of values *inside* the loop,
764 // because we can always add their offsets to the result IV after the loop
765 // is done, ensuring we get good code inside the loop.
766 if (!L->contains(Uses[i].Inst->getParent()))
770 // If the base is zero (which is common), return zero now, there are no
772 if (Uses[i].Base == Zero) return Zero;
774 // If this use is as an address we may be able to put CSEs in the addressing
775 // mode rather than hoisting them.
776 bool isAddrUse = isAddressUse(Uses[i].Inst, Uses[i].OperandValToReplace);
777 // We may need the AccessTy below, but only when isAddrUse, so compute it
778 // only in that case.
779 const Type *AccessTy = 0;
781 AccessTy = getAccessType(Uses[i].Inst);
783 // Split the expression into subexprs.
784 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
785 // Add one to SubExpressionUseData.Count for each subexpr present, and
786 // if the subexpr is not a valid immediate within an addressing mode use,
787 // set SubExpressionUseData.notAllUsesAreFree. We definitely want to
788 // hoist these out of the loop (if they are common to all uses).
789 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) {
790 if (++SubExpressionUseData[SubExprs[j]].Count == 1)
791 UniqueSubExprs.push_back(SubExprs[j]);
792 if (!isAddrUse || !fitsInAddressMode(SubExprs[j], AccessTy, TLI, false))
793 SubExpressionUseData[SubExprs[j]].notAllUsesAreFree = true;
798 // Now that we know how many times each is used, build Result. Iterate over
799 // UniqueSubexprs so that we have a stable ordering.
800 for (unsigned i = 0, e = UniqueSubExprs.size(); i != e; ++i) {
801 std::map<const SCEV *, SubExprUseData>::iterator I =
802 SubExpressionUseData.find(UniqueSubExprs[i]);
803 assert(I != SubExpressionUseData.end() && "Entry not found?");
804 if (I->second.Count == NumUsesInsideLoop) { // Found CSE!
805 if (I->second.notAllUsesAreFree)
806 Result = SE->getAddExpr(Result, I->first);
808 FreeResult = SE->getAddExpr(FreeResult, I->first);
810 // Remove non-cse's from SubExpressionUseData.
811 SubExpressionUseData.erase(I);
814 if (FreeResult != Zero) {
815 // We have some subexpressions that can be subsumed into addressing
816 // modes in every use inside the loop. However, it's possible that
817 // there are so many of them that the combined FreeResult cannot
818 // be subsumed, or that the target cannot handle both a FreeResult
819 // and a Result in the same instruction (for example because it would
820 // require too many registers). Check this.
821 for (unsigned i=0; i<NumUses; ++i) {
822 if (!L->contains(Uses[i].Inst->getParent()))
824 // We know this is an addressing mode use; if there are any uses that
825 // are not, FreeResult would be Zero.
826 const Type *AccessTy = getAccessType(Uses[i].Inst);
827 if (!fitsInAddressMode(FreeResult, AccessTy, TLI, Result!=Zero)) {
828 // FIXME: could split up FreeResult into pieces here, some hoisted
829 // and some not. There is no obvious advantage to this.
830 Result = SE->getAddExpr(Result, FreeResult);
837 // If we found no CSE's, return now.
838 if (Result == Zero) return Result;
840 // If we still have a FreeResult, remove its subexpressions from
841 // SubExpressionUseData. This means they will remain in the use Bases.
842 if (FreeResult != Zero) {
843 SeparateSubExprs(SubExprs, FreeResult, SE);
844 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) {
845 std::map<const SCEV *, SubExprUseData>::iterator I =
846 SubExpressionUseData.find(SubExprs[j]);
847 SubExpressionUseData.erase(I);
852 // Otherwise, remove all of the CSE's we found from each of the base values.
853 for (unsigned i = 0; i != NumUses; ++i) {
854 // Uses outside the loop don't necessarily include the common base, but
855 // the final IV value coming into those uses does. Instead of trying to
856 // remove the pieces of the common base, which might not be there,
857 // subtract off the base to compensate for this.
858 if (!L->contains(Uses[i].Inst->getParent())) {
859 Uses[i].Base = SE->getMinusSCEV(Uses[i].Base, Result);
863 // Split the expression into subexprs.
864 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
866 // Remove any common subexpressions.
867 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
868 if (SubExpressionUseData.count(SubExprs[j])) {
869 SubExprs.erase(SubExprs.begin()+j);
873 // Finally, add the non-shared expressions together.
874 if (SubExprs.empty())
877 Uses[i].Base = SE->getAddExpr(SubExprs);
884 /// ValidScale - Check whether the given Scale is valid for all loads and
885 /// stores in UsersToProcess.
887 bool LoopStrengthReduce::ValidScale(bool HasBaseReg, int64_t Scale,
888 const std::vector<BasedUser>& UsersToProcess) {
892 for (unsigned i = 0, e = UsersToProcess.size(); i!=e; ++i) {
893 // If this is a load or other access, pass the type of the access in.
894 const Type *AccessTy =
895 Type::getVoidTy(UsersToProcess[i].Inst->getContext());
896 if (isAddressUse(UsersToProcess[i].Inst,
897 UsersToProcess[i].OperandValToReplace))
898 AccessTy = getAccessType(UsersToProcess[i].Inst);
899 else if (isa<PHINode>(UsersToProcess[i].Inst))
902 TargetLowering::AddrMode AM;
903 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
904 AM.BaseOffs = SC->getValue()->getSExtValue();
905 AM.HasBaseReg = HasBaseReg || !UsersToProcess[i].Base->isZero();
908 // If load[imm+r*scale] is illegal, bail out.
909 if (!TLI->isLegalAddressingMode(AM, AccessTy))
915 /// ValidOffset - Check whether the given Offset is valid for all loads and
916 /// stores in UsersToProcess.
918 bool LoopStrengthReduce::ValidOffset(bool HasBaseReg,
921 const std::vector<BasedUser>& UsersToProcess) {
925 for (unsigned i=0, e = UsersToProcess.size(); i!=e; ++i) {
926 // If this is a load or other access, pass the type of the access in.
927 const Type *AccessTy =
928 Type::getVoidTy(UsersToProcess[i].Inst->getContext());
929 if (isAddressUse(UsersToProcess[i].Inst,
930 UsersToProcess[i].OperandValToReplace))
931 AccessTy = getAccessType(UsersToProcess[i].Inst);
932 else if (isa<PHINode>(UsersToProcess[i].Inst))
935 TargetLowering::AddrMode AM;
936 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
937 AM.BaseOffs = SC->getValue()->getSExtValue();
938 AM.BaseOffs = (uint64_t)AM.BaseOffs + (uint64_t)Offset;
939 AM.HasBaseReg = HasBaseReg || !UsersToProcess[i].Base->isZero();
942 // If load[imm+r*scale] is illegal, bail out.
943 if (!TLI->isLegalAddressingMode(AM, AccessTy))
949 /// RequiresTypeConversion - Returns true if converting Ty1 to Ty2 is not
951 bool LoopStrengthReduce::RequiresTypeConversion(const Type *Ty1,
955 Ty1 = SE->getEffectiveSCEVType(Ty1);
956 Ty2 = SE->getEffectiveSCEVType(Ty2);
959 if (Ty1->canLosslesslyBitCastTo(Ty2))
961 if (TLI && TLI->isTruncateFree(Ty1, Ty2))
966 /// CheckForIVReuse - Returns the multiple if the stride is the multiple
967 /// of a previous stride and it is a legal value for the target addressing
968 /// mode scale component and optional base reg. This allows the users of
969 /// this stride to be rewritten as prev iv * factor. It returns 0 if no
970 /// reuse is possible. Factors can be negative on same targets, e.g. ARM.
972 /// If all uses are outside the loop, we don't require that all multiplies
973 /// be folded into the addressing mode, nor even that the factor be constant;
974 /// a multiply (executed once) outside the loop is better than another IV
975 /// within. Well, usually.
976 const SCEV *LoopStrengthReduce::CheckForIVReuse(bool HasBaseReg,
977 bool AllUsesAreAddresses,
978 bool AllUsesAreOutsideLoop,
979 const SCEV *const &Stride,
980 IVExpr &IV, const Type *Ty,
981 const std::vector<BasedUser>& UsersToProcess) {
982 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Stride)) {
983 int64_t SInt = SC->getValue()->getSExtValue();
984 for (unsigned NewStride = 0, e = IU->StrideOrder.size();
985 NewStride != e; ++NewStride) {
986 std::map<const SCEV *, IVsOfOneStride>::iterator SI =
987 IVsByStride.find(IU->StrideOrder[NewStride]);
988 if (SI == IVsByStride.end() || !isa<SCEVConstant>(SI->first))
990 // The other stride has no uses, don't reuse it.
991 std::map<const SCEV *, IVUsersOfOneStride *>::iterator UI =
992 IU->IVUsesByStride.find(IU->StrideOrder[NewStride]);
993 if (UI->second->Users.empty())
995 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
996 if (SI->first != Stride &&
997 (unsigned(abs64(SInt)) < SSInt || (SInt % SSInt) != 0))
999 int64_t Scale = SInt / SSInt;
1000 // Check that this stride is valid for all the types used for loads and
1001 // stores; if it can be used for some and not others, we might as well use
1002 // the original stride everywhere, since we have to create the IV for it
1003 // anyway. If the scale is 1, then we don't need to worry about folding
1006 (AllUsesAreAddresses &&
1007 ValidScale(HasBaseReg, Scale, UsersToProcess))) {
1008 // Prefer to reuse an IV with a base of zero.
1009 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1010 IE = SI->second.IVs.end(); II != IE; ++II)
1011 // Only reuse previous IV if it would not require a type conversion
1012 // and if the base difference can be folded.
1013 if (II->Base->isZero() &&
1014 !RequiresTypeConversion(II->Base->getType(), Ty)) {
1016 return SE->getIntegerSCEV(Scale, Stride->getType());
1018 // Otherwise, settle for an IV with a foldable base.
1019 if (AllUsesAreAddresses)
1020 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1021 IE = SI->second.IVs.end(); II != IE; ++II)
1022 // Only reuse previous IV if it would not require a type conversion
1023 // and if the base difference can be folded.
1024 if (SE->getEffectiveSCEVType(II->Base->getType()) ==
1025 SE->getEffectiveSCEVType(Ty) &&
1026 isa<SCEVConstant>(II->Base)) {
1028 cast<SCEVConstant>(II->Base)->getValue()->getSExtValue();
1029 if (Base > INT32_MIN && Base <= INT32_MAX &&
1030 ValidOffset(HasBaseReg, -Base * Scale,
1031 Scale, UsersToProcess)) {
1033 return SE->getIntegerSCEV(Scale, Stride->getType());
1038 } else if (AllUsesAreOutsideLoop) {
1039 // Accept nonconstant strides here; it is really really right to substitute
1040 // an existing IV if we can.
1041 for (unsigned NewStride = 0, e = IU->StrideOrder.size();
1042 NewStride != e; ++NewStride) {
1043 std::map<const SCEV *, IVsOfOneStride>::iterator SI =
1044 IVsByStride.find(IU->StrideOrder[NewStride]);
1045 if (SI == IVsByStride.end() || !isa<SCEVConstant>(SI->first))
1047 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1048 if (SI->first != Stride && SSInt != 1)
1050 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1051 IE = SI->second.IVs.end(); II != IE; ++II)
1052 // Accept nonzero base here.
1053 // Only reuse previous IV if it would not require a type conversion.
1054 if (!RequiresTypeConversion(II->Base->getType(), Ty)) {
1059 // Special case, old IV is -1*x and this one is x. Can treat this one as
1061 for (unsigned NewStride = 0, e = IU->StrideOrder.size();
1062 NewStride != e; ++NewStride) {
1063 std::map<const SCEV *, IVsOfOneStride>::iterator SI =
1064 IVsByStride.find(IU->StrideOrder[NewStride]);
1065 if (SI == IVsByStride.end())
1067 if (const SCEVMulExpr *ME = dyn_cast<SCEVMulExpr>(SI->first))
1068 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(ME->getOperand(0)))
1069 if (Stride == ME->getOperand(1) &&
1070 SC->getValue()->getSExtValue() == -1LL)
1071 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1072 IE = SI->second.IVs.end(); II != IE; ++II)
1073 // Accept nonzero base here.
1074 // Only reuse previous IV if it would not require type conversion.
1075 if (!RequiresTypeConversion(II->Base->getType(), Ty)) {
1077 return SE->getIntegerSCEV(-1LL, Stride->getType());
1081 return SE->getIntegerSCEV(0, Stride->getType());
1084 /// PartitionByIsUseOfPostIncrementedValue - Simple boolean predicate that
1085 /// returns true if Val's isUseOfPostIncrementedValue is true.
1086 static bool PartitionByIsUseOfPostIncrementedValue(const BasedUser &Val) {
1087 return Val.isUseOfPostIncrementedValue;
1090 /// isNonConstantNegative - Return true if the specified scev is negated, but
1092 static bool isNonConstantNegative(const SCEV *const &Expr) {
1093 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Expr);
1094 if (!Mul) return false;
1096 // If there is a constant factor, it will be first.
1097 const SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
1098 if (!SC) return false;
1100 // Return true if the value is negative, this matches things like (-42 * V).
1101 return SC->getValue()->getValue().isNegative();
1104 /// CollectIVUsers - Transform our list of users and offsets to a bit more
1105 /// complex table. In this new vector, each 'BasedUser' contains 'Base', the
1106 /// base of the strided accesses, as well as the old information from Uses. We
1107 /// progressively move information from the Base field to the Imm field, until
1108 /// we eventually have the full access expression to rewrite the use.
1109 const SCEV *LoopStrengthReduce::CollectIVUsers(const SCEV *const &Stride,
1110 IVUsersOfOneStride &Uses,
1112 bool &AllUsesAreAddresses,
1113 bool &AllUsesAreOutsideLoop,
1114 std::vector<BasedUser> &UsersToProcess) {
1115 // FIXME: Generalize to non-affine IV's.
1116 if (!Stride->isLoopInvariant(L))
1117 return SE->getIntegerSCEV(0, Stride->getType());
1119 UsersToProcess.reserve(Uses.Users.size());
1120 for (ilist<IVStrideUse>::iterator I = Uses.Users.begin(),
1121 E = Uses.Users.end(); I != E; ++I) {
1122 UsersToProcess.push_back(BasedUser(*I, SE));
1124 // Move any loop variant operands from the offset field to the immediate
1125 // field of the use, so that we don't try to use something before it is
1127 MoveLoopVariantsToImmediateField(UsersToProcess.back().Base,
1128 UsersToProcess.back().Imm, L, SE);
1129 assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
1130 "Base value is not loop invariant!");
1133 // We now have a whole bunch of uses of like-strided induction variables, but
1134 // they might all have different bases. We want to emit one PHI node for this
1135 // stride which we fold as many common expressions (between the IVs) into as
1136 // possible. Start by identifying the common expressions in the base values
1137 // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
1138 // "A+B"), emit it to the preheader, then remove the expression from the
1139 // UsersToProcess base values.
1140 const SCEV *CommonExprs =
1141 RemoveCommonExpressionsFromUseBases(UsersToProcess, SE, L, TLI);
1143 // Next, figure out what we can represent in the immediate fields of
1144 // instructions. If we can represent anything there, move it to the imm
1145 // fields of the BasedUsers. We do this so that it increases the commonality
1146 // of the remaining uses.
1147 unsigned NumPHI = 0;
1148 bool HasAddress = false;
1149 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1150 // If the user is not in the current loop, this means it is using the exit
1151 // value of the IV. Do not put anything in the base, make sure it's all in
1152 // the immediate field to allow as much factoring as possible.
1153 if (!L->contains(UsersToProcess[i].Inst->getParent())) {
1154 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm,
1155 UsersToProcess[i].Base);
1156 UsersToProcess[i].Base =
1157 SE->getIntegerSCEV(0, UsersToProcess[i].Base->getType());
1159 // Not all uses are outside the loop.
1160 AllUsesAreOutsideLoop = false;
1162 // Addressing modes can be folded into loads and stores. Be careful that
1163 // the store is through the expression, not of the expression though.
1165 bool isAddress = isAddressUse(UsersToProcess[i].Inst,
1166 UsersToProcess[i].OperandValToReplace);
1167 if (isa<PHINode>(UsersToProcess[i].Inst)) {
1175 // If this use isn't an address, then not all uses are addresses.
1176 if (!isAddress && !isPHI)
1177 AllUsesAreAddresses = false;
1179 MoveImmediateValues(TLI, UsersToProcess[i].Inst, UsersToProcess[i].Base,
1180 UsersToProcess[i].Imm, isAddress, L, SE);
1184 // If one of the use is a PHI node and all other uses are addresses, still
1185 // allow iv reuse. Essentially we are trading one constant multiplication
1186 // for one fewer iv.
1188 AllUsesAreAddresses = false;
1190 // There are no in-loop address uses.
1191 if (AllUsesAreAddresses && (!HasAddress && !AllUsesAreOutsideLoop))
1192 AllUsesAreAddresses = false;
1197 /// ShouldUseFullStrengthReductionMode - Test whether full strength-reduction
1198 /// is valid and profitable for the given set of users of a stride. In
1199 /// full strength-reduction mode, all addresses at the current stride are
1200 /// strength-reduced all the way down to pointer arithmetic.
1202 bool LoopStrengthReduce::ShouldUseFullStrengthReductionMode(
1203 const std::vector<BasedUser> &UsersToProcess,
1205 bool AllUsesAreAddresses,
1206 const SCEV *Stride) {
1207 if (!EnableFullLSRMode)
1210 // The heuristics below aim to avoid increasing register pressure, but
1211 // fully strength-reducing all the addresses increases the number of
1212 // add instructions, so don't do this when optimizing for size.
1213 // TODO: If the loop is large, the savings due to simpler addresses
1214 // may oughtweight the costs of the extra increment instructions.
1215 if (L->getHeader()->getParent()->hasFnAttr(Attribute::OptimizeForSize))
1218 // TODO: For now, don't do full strength reduction if there could
1219 // potentially be greater-stride multiples of the current stride
1220 // which could reuse the current stride IV.
1221 if (IU->StrideOrder.back() != Stride)
1224 // Iterate through the uses to find conditions that automatically rule out
1226 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ) {
1227 const SCEV *Base = UsersToProcess[i].Base;
1228 const SCEV *Imm = UsersToProcess[i].Imm;
1229 // If any users have a loop-variant component, they can't be fully
1230 // strength-reduced.
1231 if (Imm && !Imm->isLoopInvariant(L))
1233 // If there are to users with the same base and the difference between
1234 // the two Imm values can't be folded into the address, full
1235 // strength reduction would increase register pressure.
1237 const SCEV *CurImm = UsersToProcess[i].Imm;
1238 if ((CurImm || Imm) && CurImm != Imm) {
1239 if (!CurImm) CurImm = SE->getIntegerSCEV(0, Stride->getType());
1240 if (!Imm) Imm = SE->getIntegerSCEV(0, Stride->getType());
1241 const Instruction *Inst = UsersToProcess[i].Inst;
1242 const Type *AccessTy = getAccessType(Inst);
1243 const SCEV *Diff = SE->getMinusSCEV(UsersToProcess[i].Imm, Imm);
1244 if (!Diff->isZero() &&
1245 (!AllUsesAreAddresses ||
1246 !fitsInAddressMode(Diff, AccessTy, TLI, /*HasBaseReg=*/true)))
1249 } while (++i != e && Base == UsersToProcess[i].Base);
1252 // If there's exactly one user in this stride, fully strength-reducing it
1253 // won't increase register pressure. If it's starting from a non-zero base,
1254 // it'll be simpler this way.
1255 if (UsersToProcess.size() == 1 && !UsersToProcess[0].Base->isZero())
1258 // Otherwise, if there are any users in this stride that don't require
1259 // a register for their base, full strength-reduction will increase
1260 // register pressure.
1261 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1262 if (UsersToProcess[i].Base->isZero())
1265 // Otherwise, go for it.
1269 /// InsertAffinePhi Create and insert a PHI node for an induction variable
1270 /// with the specified start and step values in the specified loop.
1272 /// If NegateStride is true, the stride should be negated by using a
1273 /// subtract instead of an add.
1275 /// Return the created phi node.
1277 static PHINode *InsertAffinePhi(const SCEV *Start, const SCEV *Step,
1278 Instruction *IVIncInsertPt,
1280 SCEVExpander &Rewriter) {
1281 assert(Start->isLoopInvariant(L) && "New PHI start is not loop invariant!");
1282 assert(Step->isLoopInvariant(L) && "New PHI stride is not loop invariant!");
1284 BasicBlock *Header = L->getHeader();
1285 BasicBlock *Preheader = L->getLoopPreheader();
1286 BasicBlock *LatchBlock = L->getLoopLatch();
1287 const Type *Ty = Start->getType();
1288 Ty = Rewriter.SE.getEffectiveSCEVType(Ty);
1290 PHINode *PN = PHINode::Create(Ty, "lsr.iv", Header->begin());
1291 PN->addIncoming(Rewriter.expandCodeFor(Start, Ty, Preheader->getTerminator()),
1294 // If the stride is negative, insert a sub instead of an add for the
1296 bool isNegative = isNonConstantNegative(Step);
1297 const SCEV *IncAmount = Step;
1299 IncAmount = Rewriter.SE.getNegativeSCEV(Step);
1301 // Insert an add instruction right before the terminator corresponding
1302 // to the back-edge or just before the only use. The location is determined
1303 // by the caller and passed in as IVIncInsertPt.
1304 Value *StepV = Rewriter.expandCodeFor(IncAmount, Ty,
1305 Preheader->getTerminator());
1308 IncV = BinaryOperator::CreateSub(PN, StepV, "lsr.iv.next",
1311 IncV = BinaryOperator::CreateAdd(PN, StepV, "lsr.iv.next",
1314 if (!isa<ConstantInt>(StepV)) ++NumVariable;
1316 PN->addIncoming(IncV, LatchBlock);
1322 static void SortUsersToProcess(std::vector<BasedUser> &UsersToProcess) {
1323 // We want to emit code for users inside the loop first. To do this, we
1324 // rearrange BasedUser so that the entries at the end have
1325 // isUseOfPostIncrementedValue = false, because we pop off the end of the
1326 // vector (so we handle them first).
1327 std::partition(UsersToProcess.begin(), UsersToProcess.end(),
1328 PartitionByIsUseOfPostIncrementedValue);
1330 // Sort this by base, so that things with the same base are handled
1331 // together. By partitioning first and stable-sorting later, we are
1332 // guaranteed that within each base we will pop off users from within the
1333 // loop before users outside of the loop with a particular base.
1335 // We would like to use stable_sort here, but we can't. The problem is that
1336 // const SCEV *'s don't have a deterministic ordering w.r.t to each other, so
1337 // we don't have anything to do a '<' comparison on. Because we think the
1338 // number of uses is small, do a horrible bubble sort which just relies on
1340 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1341 // Get a base value.
1342 const SCEV *Base = UsersToProcess[i].Base;
1344 // Compact everything with this base to be consecutive with this one.
1345 for (unsigned j = i+1; j != e; ++j) {
1346 if (UsersToProcess[j].Base == Base) {
1347 std::swap(UsersToProcess[i+1], UsersToProcess[j]);
1354 /// PrepareToStrengthReduceFully - Prepare to fully strength-reduce
1355 /// UsersToProcess, meaning lowering addresses all the way down to direct
1356 /// pointer arithmetic.
1359 LoopStrengthReduce::PrepareToStrengthReduceFully(
1360 std::vector<BasedUser> &UsersToProcess,
1362 const SCEV *CommonExprs,
1364 SCEVExpander &PreheaderRewriter) {
1365 DEBUG(errs() << " Fully reducing all users\n");
1367 // Rewrite the UsersToProcess records, creating a separate PHI for each
1368 // unique Base value.
1369 Instruction *IVIncInsertPt = L->getLoopLatch()->getTerminator();
1370 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ) {
1371 // TODO: The uses are grouped by base, but not sorted. We arbitrarily
1372 // pick the first Imm value here to start with, and adjust it for the
1374 const SCEV *Imm = UsersToProcess[i].Imm;
1375 const SCEV *Base = UsersToProcess[i].Base;
1376 const SCEV *Start = SE->getAddExpr(CommonExprs, Base, Imm);
1377 PHINode *Phi = InsertAffinePhi(Start, Stride, IVIncInsertPt, L,
1379 // Loop over all the users with the same base.
1381 UsersToProcess[i].Base = SE->getIntegerSCEV(0, Stride->getType());
1382 UsersToProcess[i].Imm = SE->getMinusSCEV(UsersToProcess[i].Imm, Imm);
1383 UsersToProcess[i].Phi = Phi;
1384 assert(UsersToProcess[i].Imm->isLoopInvariant(L) &&
1385 "ShouldUseFullStrengthReductionMode should reject this!");
1386 } while (++i != e && Base == UsersToProcess[i].Base);
1390 /// FindIVIncInsertPt - Return the location to insert the increment instruction.
1391 /// If the only use if a use of postinc value, (must be the loop termination
1392 /// condition), then insert it just before the use.
1393 static Instruction *FindIVIncInsertPt(std::vector<BasedUser> &UsersToProcess,
1395 if (UsersToProcess.size() == 1 &&
1396 UsersToProcess[0].isUseOfPostIncrementedValue &&
1397 L->contains(UsersToProcess[0].Inst->getParent()))
1398 return UsersToProcess[0].Inst;
1399 return L->getLoopLatch()->getTerminator();
1402 /// PrepareToStrengthReduceWithNewPhi - Insert a new induction variable for the
1403 /// given users to share.
1406 LoopStrengthReduce::PrepareToStrengthReduceWithNewPhi(
1407 std::vector<BasedUser> &UsersToProcess,
1409 const SCEV *CommonExprs,
1411 Instruction *IVIncInsertPt,
1413 SCEVExpander &PreheaderRewriter) {
1414 DEBUG(errs() << " Inserting new PHI:\n");
1416 PHINode *Phi = InsertAffinePhi(SE->getUnknown(CommonBaseV),
1417 Stride, IVIncInsertPt, L,
1420 // Remember this in case a later stride is multiple of this.
1421 IVsByStride[Stride].addIV(Stride, CommonExprs, Phi);
1423 // All the users will share this new IV.
1424 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1425 UsersToProcess[i].Phi = Phi;
1427 DEBUG(errs() << " IV=");
1428 DEBUG(WriteAsOperand(errs(), Phi, /*PrintType=*/false));
1429 DEBUG(errs() << "\n");
1432 /// PrepareToStrengthReduceFromSmallerStride - Prepare for the given users to
1433 /// reuse an induction variable with a stride that is a factor of the current
1434 /// induction variable.
1437 LoopStrengthReduce::PrepareToStrengthReduceFromSmallerStride(
1438 std::vector<BasedUser> &UsersToProcess,
1440 const IVExpr &ReuseIV,
1441 Instruction *PreInsertPt) {
1442 DEBUG(errs() << " Rewriting in terms of existing IV of STRIDE "
1443 << *ReuseIV.Stride << " and BASE " << *ReuseIV.Base << "\n");
1445 // All the users will share the reused IV.
1446 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1447 UsersToProcess[i].Phi = ReuseIV.PHI;
1449 Constant *C = dyn_cast<Constant>(CommonBaseV);
1451 (!C->isNullValue() &&
1452 !fitsInAddressMode(SE->getUnknown(CommonBaseV), CommonBaseV->getType(),
1454 // We want the common base emitted into the preheader! This is just
1455 // using cast as a copy so BitCast (no-op cast) is appropriate
1456 CommonBaseV = new BitCastInst(CommonBaseV, CommonBaseV->getType(),
1457 "commonbase", PreInsertPt);
1460 static bool IsImmFoldedIntoAddrMode(GlobalValue *GV, int64_t Offset,
1461 const Type *AccessTy,
1462 std::vector<BasedUser> &UsersToProcess,
1463 const TargetLowering *TLI) {
1464 SmallVector<Instruction*, 16> AddrModeInsts;
1465 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1466 if (UsersToProcess[i].isUseOfPostIncrementedValue)
1468 ExtAddrMode AddrMode =
1469 AddressingModeMatcher::Match(UsersToProcess[i].OperandValToReplace,
1470 AccessTy, UsersToProcess[i].Inst,
1471 AddrModeInsts, *TLI);
1472 if (GV && GV != AddrMode.BaseGV)
1474 if (Offset && !AddrMode.BaseOffs)
1475 // FIXME: How to accurate check it's immediate offset is folded.
1477 AddrModeInsts.clear();
1482 /// StrengthReduceIVUsersOfStride - Strength reduce all of the users of a single
1483 /// stride of IV. All of the users may have different starting values, and this
1484 /// may not be the only stride.
1486 LoopStrengthReduce::StrengthReduceIVUsersOfStride(const SCEV *const &Stride,
1487 IVUsersOfOneStride &Uses,
1489 // If all the users are moved to another stride, then there is nothing to do.
1490 if (Uses.Users.empty())
1493 // Keep track if every use in UsersToProcess is an address. If they all are,
1494 // we may be able to rewrite the entire collection of them in terms of a
1495 // smaller-stride IV.
1496 bool AllUsesAreAddresses = true;
1498 // Keep track if every use of a single stride is outside the loop. If so,
1499 // we want to be more aggressive about reusing a smaller-stride IV; a
1500 // multiply outside the loop is better than another IV inside. Well, usually.
1501 bool AllUsesAreOutsideLoop = true;
1503 // Transform our list of users and offsets to a bit more complex table. In
1504 // this new vector, each 'BasedUser' contains 'Base' the base of the
1505 // strided accessas well as the old information from Uses. We progressively
1506 // move information from the Base field to the Imm field, until we eventually
1507 // have the full access expression to rewrite the use.
1508 std::vector<BasedUser> UsersToProcess;
1509 const SCEV *CommonExprs = CollectIVUsers(Stride, Uses, L, AllUsesAreAddresses,
1510 AllUsesAreOutsideLoop,
1513 // Sort the UsersToProcess array so that users with common bases are
1514 // next to each other.
1515 SortUsersToProcess(UsersToProcess);
1517 // If we managed to find some expressions in common, we'll need to carry
1518 // their value in a register and add it in for each use. This will take up
1519 // a register operand, which potentially restricts what stride values are
1521 bool HaveCommonExprs = !CommonExprs->isZero();
1522 const Type *ReplacedTy = CommonExprs->getType();
1524 // If all uses are addresses, consider sinking the immediate part of the
1525 // common expression back into uses if they can fit in the immediate fields.
1526 if (TLI && HaveCommonExprs && AllUsesAreAddresses) {
1527 const SCEV *NewCommon = CommonExprs;
1528 const SCEV *Imm = SE->getIntegerSCEV(0, ReplacedTy);
1529 MoveImmediateValues(TLI, Type::getVoidTy(
1530 L->getLoopPreheader()->getContext()),
1531 NewCommon, Imm, true, L, SE);
1532 if (!Imm->isZero()) {
1535 // If the immediate part of the common expression is a GV, check if it's
1536 // possible to fold it into the target addressing mode.
1537 GlobalValue *GV = 0;
1538 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(Imm))
1539 GV = dyn_cast<GlobalValue>(SU->getValue());
1541 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Imm))
1542 Offset = SC->getValue()->getSExtValue();
1544 // Pass VoidTy as the AccessTy to be conservative, because
1545 // there could be multiple access types among all the uses.
1546 DoSink = IsImmFoldedIntoAddrMode(GV, Offset,
1547 Type::getVoidTy(L->getLoopPreheader()->getContext()),
1548 UsersToProcess, TLI);
1551 DEBUG(errs() << " Sinking " << *Imm << " back down into uses\n");
1552 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1553 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm, Imm);
1554 CommonExprs = NewCommon;
1555 HaveCommonExprs = !CommonExprs->isZero();
1561 // Now that we know what we need to do, insert the PHI node itself.
1563 DEBUG(errs() << "LSR: Examining IVs of TYPE " << *ReplacedTy << " of STRIDE "
1565 << " Common base: " << *CommonExprs << "\n");
1567 SCEVExpander Rewriter(*SE);
1568 SCEVExpander PreheaderRewriter(*SE);
1570 BasicBlock *Preheader = L->getLoopPreheader();
1571 Instruction *PreInsertPt = Preheader->getTerminator();
1572 BasicBlock *LatchBlock = L->getLoopLatch();
1573 Instruction *IVIncInsertPt = LatchBlock->getTerminator();
1575 Value *CommonBaseV = Constant::getNullValue(ReplacedTy);
1577 const SCEV *RewriteFactor = SE->getIntegerSCEV(0, ReplacedTy);
1578 IVExpr ReuseIV(SE->getIntegerSCEV(0,
1579 Type::getInt32Ty(Preheader->getContext())),
1580 SE->getIntegerSCEV(0,
1581 Type::getInt32Ty(Preheader->getContext())),
1584 // Choose a strength-reduction strategy and prepare for it by creating
1585 // the necessary PHIs and adjusting the bookkeeping.
1586 if (ShouldUseFullStrengthReductionMode(UsersToProcess, L,
1587 AllUsesAreAddresses, Stride)) {
1588 PrepareToStrengthReduceFully(UsersToProcess, Stride, CommonExprs, L,
1591 // Emit the initial base value into the loop preheader.
1592 CommonBaseV = PreheaderRewriter.expandCodeFor(CommonExprs, ReplacedTy,
1595 // If all uses are addresses, check if it is possible to reuse an IV. The
1596 // new IV must have a stride that is a multiple of the old stride; the
1597 // multiple must be a number that can be encoded in the scale field of the
1598 // target addressing mode; and we must have a valid instruction after this
1599 // substitution, including the immediate field, if any.
1600 RewriteFactor = CheckForIVReuse(HaveCommonExprs, AllUsesAreAddresses,
1601 AllUsesAreOutsideLoop,
1602 Stride, ReuseIV, ReplacedTy,
1604 if (!RewriteFactor->isZero())
1605 PrepareToStrengthReduceFromSmallerStride(UsersToProcess, CommonBaseV,
1606 ReuseIV, PreInsertPt);
1608 IVIncInsertPt = FindIVIncInsertPt(UsersToProcess, L);
1609 PrepareToStrengthReduceWithNewPhi(UsersToProcess, Stride, CommonExprs,
1610 CommonBaseV, IVIncInsertPt,
1611 L, PreheaderRewriter);
1615 // Process all the users now, replacing their strided uses with
1616 // strength-reduced forms. This outer loop handles all bases, the inner
1617 // loop handles all users of a particular base.
1618 while (!UsersToProcess.empty()) {
1619 const SCEV *Base = UsersToProcess.back().Base;
1620 Instruction *Inst = UsersToProcess.back().Inst;
1622 // Emit the code for Base into the preheader.
1624 if (!Base->isZero()) {
1625 BaseV = PreheaderRewriter.expandCodeFor(Base, 0, PreInsertPt);
1627 DEBUG(errs() << " INSERTING code for BASE = " << *Base << ":");
1628 if (BaseV->hasName())
1629 DEBUG(errs() << " Result value name = %" << BaseV->getName());
1630 DEBUG(errs() << "\n");
1632 // If BaseV is a non-zero constant, make sure that it gets inserted into
1633 // the preheader, instead of being forward substituted into the uses. We
1634 // do this by forcing a BitCast (noop cast) to be inserted into the
1635 // preheader in this case.
1636 if (!fitsInAddressMode(Base, getAccessType(Inst), TLI, false) &&
1637 isa<Constant>(BaseV)) {
1638 // We want this constant emitted into the preheader! This is just
1639 // using cast as a copy so BitCast (no-op cast) is appropriate
1640 BaseV = new BitCastInst(BaseV, BaseV->getType(), "preheaderinsert",
1645 // Emit the code to add the immediate offset to the Phi value, just before
1646 // the instructions that we identified as using this stride and base.
1648 // FIXME: Use emitted users to emit other users.
1649 BasedUser &User = UsersToProcess.back();
1651 DEBUG(errs() << " Examining ");
1652 if (User.isUseOfPostIncrementedValue)
1653 DEBUG(errs() << "postinc");
1655 DEBUG(errs() << "preinc");
1656 DEBUG(errs() << " use ");
1657 DEBUG(WriteAsOperand(errs(), UsersToProcess.back().OperandValToReplace,
1658 /*PrintType=*/false));
1659 DEBUG(errs() << " in Inst: " << *User.Inst);
1661 // If this instruction wants to use the post-incremented value, move it
1662 // after the post-inc and use its value instead of the PHI.
1663 Value *RewriteOp = User.Phi;
1664 if (User.isUseOfPostIncrementedValue) {
1665 RewriteOp = User.Phi->getIncomingValueForBlock(LatchBlock);
1666 // If this user is in the loop, make sure it is the last thing in the
1667 // loop to ensure it is dominated by the increment. In case it's the
1668 // only use of the iv, the increment instruction is already before the
1670 if (L->contains(User.Inst->getParent()) && User.Inst != IVIncInsertPt)
1671 User.Inst->moveBefore(IVIncInsertPt);
1674 const SCEV *RewriteExpr = SE->getUnknown(RewriteOp);
1676 if (SE->getEffectiveSCEVType(RewriteOp->getType()) !=
1677 SE->getEffectiveSCEVType(ReplacedTy)) {
1678 assert(SE->getTypeSizeInBits(RewriteOp->getType()) >
1679 SE->getTypeSizeInBits(ReplacedTy) &&
1680 "Unexpected widening cast!");
1681 RewriteExpr = SE->getTruncateExpr(RewriteExpr, ReplacedTy);
1684 // If we had to insert new instructions for RewriteOp, we have to
1685 // consider that they may not have been able to end up immediately
1686 // next to RewriteOp, because non-PHI instructions may never precede
1687 // PHI instructions in a block. In this case, remember where the last
1688 // instruction was inserted so that if we're replacing a different
1689 // PHI node, we can use the later point to expand the final
1691 Instruction *NewBasePt = dyn_cast<Instruction>(RewriteOp);
1692 if (RewriteOp == User.Phi) NewBasePt = 0;
1694 // Clear the SCEVExpander's expression map so that we are guaranteed
1695 // to have the code emitted where we expect it.
1698 // If we are reusing the iv, then it must be multiplied by a constant
1699 // factor to take advantage of the addressing mode scale component.
1700 if (!RewriteFactor->isZero()) {
1701 // If we're reusing an IV with a nonzero base (currently this happens
1702 // only when all reuses are outside the loop) subtract that base here.
1703 // The base has been used to initialize the PHI node but we don't want
1705 if (!ReuseIV.Base->isZero()) {
1706 const SCEV *typedBase = ReuseIV.Base;
1707 if (SE->getEffectiveSCEVType(RewriteExpr->getType()) !=
1708 SE->getEffectiveSCEVType(ReuseIV.Base->getType())) {
1709 // It's possible the original IV is a larger type than the new IV,
1710 // in which case we have to truncate the Base. We checked in
1711 // RequiresTypeConversion that this is valid.
1712 assert(SE->getTypeSizeInBits(RewriteExpr->getType()) <
1713 SE->getTypeSizeInBits(ReuseIV.Base->getType()) &&
1714 "Unexpected lengthening conversion!");
1715 typedBase = SE->getTruncateExpr(ReuseIV.Base,
1716 RewriteExpr->getType());
1718 RewriteExpr = SE->getMinusSCEV(RewriteExpr, typedBase);
1721 // Multiply old variable, with base removed, by new scale factor.
1722 RewriteExpr = SE->getMulExpr(RewriteFactor,
1725 // The common base is emitted in the loop preheader. But since we
1726 // are reusing an IV, it has not been used to initialize the PHI node.
1727 // Add it to the expression used to rewrite the uses.
1728 // When this use is outside the loop, we earlier subtracted the
1729 // common base, and are adding it back here. Use the same expression
1730 // as before, rather than CommonBaseV, so DAGCombiner will zap it.
1731 if (!CommonExprs->isZero()) {
1732 if (L->contains(User.Inst->getParent()))
1733 RewriteExpr = SE->getAddExpr(RewriteExpr,
1734 SE->getUnknown(CommonBaseV));
1736 RewriteExpr = SE->getAddExpr(RewriteExpr, CommonExprs);
1740 // Now that we know what we need to do, insert code before User for the
1741 // immediate and any loop-variant expressions.
1743 // Add BaseV to the PHI value if needed.
1744 RewriteExpr = SE->getAddExpr(RewriteExpr, SE->getUnknown(BaseV));
1746 User.RewriteInstructionToUseNewBase(RewriteExpr, NewBasePt,
1750 // Mark old value we replaced as possibly dead, so that it is eliminated
1751 // if we just replaced the last use of that value.
1752 DeadInsts.push_back(User.OperandValToReplace);
1754 UsersToProcess.pop_back();
1757 // If there are any more users to process with the same base, process them
1758 // now. We sorted by base above, so we just have to check the last elt.
1759 } while (!UsersToProcess.empty() && UsersToProcess.back().Base == Base);
1760 // TODO: Next, find out which base index is the most common, pull it out.
1763 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
1764 // different starting values, into different PHIs.
1767 void LoopStrengthReduce::StrengthReduceIVUsers(Loop *L) {
1768 // Note: this processes each stride/type pair individually. All users
1769 // passed into StrengthReduceIVUsersOfStride have the same type AND stride.
1770 // Also, note that we iterate over IVUsesByStride indirectly by using
1771 // StrideOrder. This extra layer of indirection makes the ordering of
1772 // strides deterministic - not dependent on map order.
1773 for (unsigned Stride = 0, e = IU->StrideOrder.size(); Stride != e; ++Stride) {
1774 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
1775 IU->IVUsesByStride.find(IU->StrideOrder[Stride]);
1776 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
1777 // FIXME: Generalize to non-affine IV's.
1778 if (!SI->first->isLoopInvariant(L))
1780 StrengthReduceIVUsersOfStride(SI->first, *SI->second, L);
1784 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1785 /// set the IV user and stride information and return true, otherwise return
1787 bool LoopStrengthReduce::FindIVUserForCond(ICmpInst *Cond,
1788 IVStrideUse *&CondUse,
1789 const SCEV* &CondStride) {
1790 for (unsigned Stride = 0, e = IU->StrideOrder.size();
1791 Stride != e && !CondUse; ++Stride) {
1792 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
1793 IU->IVUsesByStride.find(IU->StrideOrder[Stride]);
1794 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
1796 for (ilist<IVStrideUse>::iterator UI = SI->second->Users.begin(),
1797 E = SI->second->Users.end(); UI != E; ++UI)
1798 if (UI->getUser() == Cond) {
1799 // NOTE: we could handle setcc instructions with multiple uses here, but
1800 // InstCombine does it as well for simple uses, it's not clear that it
1801 // occurs enough in real life to handle.
1803 CondStride = SI->first;
1811 // Constant strides come first which in turns are sorted by their absolute
1812 // values. If absolute values are the same, then positive strides comes first.
1814 // 4, -1, X, 1, 2 ==> 1, -1, 2, 4, X
1815 struct StrideCompare {
1816 const ScalarEvolution *SE;
1817 explicit StrideCompare(const ScalarEvolution *se) : SE(se) {}
1819 bool operator()(const SCEV *const &LHS, const SCEV *const &RHS) {
1820 const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS);
1821 const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS);
1823 int64_t LV = LHSC->getValue()->getSExtValue();
1824 int64_t RV = RHSC->getValue()->getSExtValue();
1825 uint64_t ALV = (LV < 0) ? -LV : LV;
1826 uint64_t ARV = (RV < 0) ? -RV : RV;
1834 // If it's the same value but different type, sort by bit width so
1835 // that we emit larger induction variables before smaller
1836 // ones, letting the smaller be re-written in terms of larger ones.
1837 return SE->getTypeSizeInBits(RHS->getType()) <
1838 SE->getTypeSizeInBits(LHS->getType());
1840 return LHSC && !RHSC;
1845 /// ChangeCompareStride - If a loop termination compare instruction is the
1846 /// only use of its stride, and the compaison is against a constant value,
1847 /// try eliminate the stride by moving the compare instruction to another
1848 /// stride and change its constant operand accordingly. e.g.
1854 /// if (v2 < 10) goto loop
1859 /// if (v1 < 30) goto loop
1860 ICmpInst *LoopStrengthReduce::ChangeCompareStride(Loop *L, ICmpInst *Cond,
1861 IVStrideUse* &CondUse,
1862 const SCEV* &CondStride,
1864 // If there's only one stride in the loop, there's nothing to do here.
1865 if (IU->StrideOrder.size() < 2)
1867 // If there are other users of the condition's stride, don't bother
1868 // trying to change the condition because the stride will still
1870 std::map<const SCEV *, IVUsersOfOneStride *>::iterator I =
1871 IU->IVUsesByStride.find(CondStride);
1872 if (I == IU->IVUsesByStride.end())
1874 if (I->second->Users.size() > 1) {
1875 for (ilist<IVStrideUse>::iterator II = I->second->Users.begin(),
1876 EE = I->second->Users.end(); II != EE; ++II) {
1877 if (II->getUser() == Cond)
1879 if (!isInstructionTriviallyDead(II->getUser()))
1883 // Only handle constant strides for now.
1884 const SCEVConstant *SC = dyn_cast<SCEVConstant>(CondStride);
1885 if (!SC) return Cond;
1887 ICmpInst::Predicate Predicate = Cond->getPredicate();
1888 int64_t CmpSSInt = SC->getValue()->getSExtValue();
1889 unsigned BitWidth = SE->getTypeSizeInBits(CondStride->getType());
1890 uint64_t SignBit = 1ULL << (BitWidth-1);
1891 const Type *CmpTy = Cond->getOperand(0)->getType();
1892 const Type *NewCmpTy = NULL;
1893 unsigned TyBits = SE->getTypeSizeInBits(CmpTy);
1894 unsigned NewTyBits = 0;
1895 const SCEV *NewStride = NULL;
1896 Value *NewCmpLHS = NULL;
1897 Value *NewCmpRHS = NULL;
1899 const SCEV *NewOffset = SE->getIntegerSCEV(0, CmpTy);
1901 if (ConstantInt *C = dyn_cast<ConstantInt>(Cond->getOperand(1))) {
1902 int64_t CmpVal = C->getValue().getSExtValue();
1904 // Check the relevant induction variable for conformance to
1906 const SCEV *IV = SE->getSCEV(Cond->getOperand(0));
1907 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1908 if (!AR || !AR->isAffine())
1911 const SCEVConstant *StartC = dyn_cast<SCEVConstant>(AR->getStart());
1912 // Check stride constant and the comparision constant signs to detect
1915 if ((StartC->getValue()->getSExtValue() < CmpVal && CmpSSInt < 0) ||
1916 (StartC->getValue()->getSExtValue() > CmpVal && CmpSSInt > 0))
1919 // More restrictive check for the other cases.
1920 if ((CmpVal & SignBit) != (CmpSSInt & SignBit))
1924 // Look for a suitable stride / iv as replacement.
1925 for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
1926 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
1927 IU->IVUsesByStride.find(IU->StrideOrder[i]);
1928 if (!isa<SCEVConstant>(SI->first) || SI->second->Users.empty())
1930 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1931 if (SSInt == CmpSSInt ||
1932 abs64(SSInt) < abs64(CmpSSInt) ||
1933 (SSInt % CmpSSInt) != 0)
1936 Scale = SSInt / CmpSSInt;
1937 int64_t NewCmpVal = CmpVal * Scale;
1939 // If old icmp value fits in icmp immediate field, but the new one doesn't
1940 // try something else.
1942 TLI->isLegalICmpImmediate(CmpVal) &&
1943 !TLI->isLegalICmpImmediate(NewCmpVal))
1946 APInt Mul = APInt(BitWidth*2, CmpVal, true);
1947 Mul = Mul * APInt(BitWidth*2, Scale, true);
1948 // Check for overflow.
1949 if (!Mul.isSignedIntN(BitWidth))
1951 // Check for overflow in the stride's type too.
1952 if (!Mul.isSignedIntN(SE->getTypeSizeInBits(SI->first->getType())))
1955 // Watch out for overflow.
1956 if (ICmpInst::isSigned(Predicate) &&
1957 (CmpVal & SignBit) != (NewCmpVal & SignBit))
1960 // Pick the best iv to use trying to avoid a cast.
1962 for (ilist<IVStrideUse>::iterator UI = SI->second->Users.begin(),
1963 E = SI->second->Users.end(); UI != E; ++UI) {
1964 Value *Op = UI->getOperandValToReplace();
1966 // If the IVStrideUse implies a cast, check for an actual cast which
1967 // can be used to find the original IV expression.
1968 if (SE->getEffectiveSCEVType(Op->getType()) !=
1969 SE->getEffectiveSCEVType(SI->first->getType())) {
1970 CastInst *CI = dyn_cast<CastInst>(Op);
1971 // If it's not a simple cast, it's complicated.
1974 // If it's a cast from a type other than the stride type,
1975 // it's complicated.
1976 if (CI->getOperand(0)->getType() != SI->first->getType())
1978 // Ok, we found the IV expression in the stride's type.
1979 Op = CI->getOperand(0);
1983 if (NewCmpLHS->getType() == CmpTy)
1989 NewCmpTy = NewCmpLHS->getType();
1990 NewTyBits = SE->getTypeSizeInBits(NewCmpTy);
1991 const Type *NewCmpIntTy = IntegerType::get(Cond->getContext(), NewTyBits);
1992 if (RequiresTypeConversion(NewCmpTy, CmpTy)) {
1993 // Check if it is possible to rewrite it using
1994 // an iv / stride of a smaller integer type.
1995 unsigned Bits = NewTyBits;
1996 if (ICmpInst::isSigned(Predicate))
1998 uint64_t Mask = (1ULL << Bits) - 1;
1999 if (((uint64_t)NewCmpVal & Mask) != (uint64_t)NewCmpVal)
2003 // Don't rewrite if use offset is non-constant and the new type is
2004 // of a different type.
2005 // FIXME: too conservative?
2006 if (NewTyBits != TyBits && !isa<SCEVConstant>(CondUse->getOffset()))
2010 bool AllUsesAreAddresses = true;
2011 bool AllUsesAreOutsideLoop = true;
2012 std::vector<BasedUser> UsersToProcess;
2013 const SCEV *CommonExprs = CollectIVUsers(SI->first, *SI->second, L,
2014 AllUsesAreAddresses,
2015 AllUsesAreOutsideLoop,
2017 // Avoid rewriting the compare instruction with an iv of new stride
2018 // if it's likely the new stride uses will be rewritten using the
2019 // stride of the compare instruction.
2020 if (AllUsesAreAddresses &&
2021 ValidScale(!CommonExprs->isZero(), Scale, UsersToProcess))
2025 // Avoid rewriting the compare instruction with an iv which has
2026 // implicit extension or truncation built into it.
2027 // TODO: This is over-conservative.
2028 if (SE->getTypeSizeInBits(CondUse->getOffset()->getType()) != TyBits)
2031 // If scale is negative, use swapped predicate unless it's testing
2033 if (Scale < 0 && !Cond->isEquality())
2034 Predicate = ICmpInst::getSwappedPredicate(Predicate);
2036 NewStride = IU->StrideOrder[i];
2037 if (!isa<PointerType>(NewCmpTy))
2038 NewCmpRHS = ConstantInt::get(NewCmpTy, NewCmpVal);
2040 Constant *CI = ConstantInt::get(NewCmpIntTy, NewCmpVal);
2041 NewCmpRHS = ConstantExpr::getIntToPtr(CI, NewCmpTy);
2043 NewOffset = TyBits == NewTyBits
2044 ? SE->getMulExpr(CondUse->getOffset(),
2045 SE->getConstant(CmpTy, Scale))
2046 : SE->getConstant(NewCmpIntTy,
2047 cast<SCEVConstant>(CondUse->getOffset())->getValue()
2048 ->getSExtValue()*Scale);
2053 // Forgo this transformation if it the increment happens to be
2054 // unfortunately positioned after the condition, and the condition
2055 // has multiple uses which prevent it from being moved immediately
2056 // before the branch. See
2057 // test/Transforms/LoopStrengthReduce/change-compare-stride-trickiness-*.ll
2058 // for an example of this situation.
2059 if (!Cond->hasOneUse()) {
2060 for (BasicBlock::iterator I = Cond, E = Cond->getParent()->end();
2067 // Create a new compare instruction using new stride / iv.
2068 ICmpInst *OldCond = Cond;
2069 // Insert new compare instruction.
2070 Cond = new ICmpInst(OldCond, Predicate, NewCmpLHS, NewCmpRHS,
2071 L->getHeader()->getName() + ".termcond");
2073 DEBUG(errs() << " Change compare stride in Inst " << *OldCond);
2074 DEBUG(errs() << " to " << *Cond << '\n');
2076 // Remove the old compare instruction. The old indvar is probably dead too.
2077 DeadInsts.push_back(CondUse->getOperandValToReplace());
2078 OldCond->replaceAllUsesWith(Cond);
2079 OldCond->eraseFromParent();
2081 IU->IVUsesByStride[NewStride]->addUser(NewOffset, Cond, NewCmpLHS);
2082 CondUse = &IU->IVUsesByStride[NewStride]->Users.back();
2083 CondStride = NewStride;
2091 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
2092 /// a max computation.
2094 /// This is a narrow solution to a specific, but acute, problem. For loops
2100 /// } while (++i < n);
2102 /// the trip count isn't just 'n', because 'n' might not be positive. And
2103 /// unfortunately this can come up even for loops where the user didn't use
2104 /// a C do-while loop. For example, seemingly well-behaved top-test loops
2105 /// will commonly be lowered like this:
2111 /// } while (++i < n);
2114 /// and then it's possible for subsequent optimization to obscure the if
2115 /// test in such a way that indvars can't find it.
2117 /// When indvars can't find the if test in loops like this, it creates a
2118 /// max expression, which allows it to give the loop a canonical
2119 /// induction variable:
2122 /// max = n < 1 ? 1 : n;
2125 /// } while (++i != max);
2127 /// Canonical induction variables are necessary because the loop passes
2128 /// are designed around them. The most obvious example of this is the
2129 /// LoopInfo analysis, which doesn't remember trip count values. It
2130 /// expects to be able to rediscover the trip count each time it is
2131 /// needed, and it does this using a simple analyis that only succeeds if
2132 /// the loop has a canonical induction variable.
2134 /// However, when it comes time to generate code, the maximum operation
2135 /// can be quite costly, especially if it's inside of an outer loop.
2137 /// This function solves this problem by detecting this type of loop and
2138 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
2139 /// the instructions for the maximum computation.
2141 ICmpInst *LoopStrengthReduce::OptimizeMax(Loop *L, ICmpInst *Cond,
2142 IVStrideUse* &CondUse) {
2143 // Check that the loop matches the pattern we're looking for.
2144 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
2145 Cond->getPredicate() != CmpInst::ICMP_NE)
2148 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
2149 if (!Sel || !Sel->hasOneUse()) return Cond;
2151 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
2152 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2154 const SCEV *One = SE->getIntegerSCEV(1, BackedgeTakenCount->getType());
2156 // Add one to the backedge-taken count to get the trip count.
2157 const SCEV *IterationCount = SE->getAddExpr(BackedgeTakenCount, One);
2159 // Check for a max calculation that matches the pattern.
2160 if (!isa<SCEVSMaxExpr>(IterationCount) && !isa<SCEVUMaxExpr>(IterationCount))
2162 const SCEVNAryExpr *Max = cast<SCEVNAryExpr>(IterationCount);
2163 if (Max != SE->getSCEV(Sel)) return Cond;
2165 // To handle a max with more than two operands, this optimization would
2166 // require additional checking and setup.
2167 if (Max->getNumOperands() != 2)
2170 const SCEV *MaxLHS = Max->getOperand(0);
2171 const SCEV *MaxRHS = Max->getOperand(1);
2172 if (!MaxLHS || MaxLHS != One) return Cond;
2174 // Check the relevant induction variable for conformance to
2176 const SCEV *IV = SE->getSCEV(Cond->getOperand(0));
2177 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2178 if (!AR || !AR->isAffine() ||
2179 AR->getStart() != One ||
2180 AR->getStepRecurrence(*SE) != One)
2183 assert(AR->getLoop() == L &&
2184 "Loop condition operand is an addrec in a different loop!");
2186 // Check the right operand of the select, and remember it, as it will
2187 // be used in the new comparison instruction.
2189 if (SE->getSCEV(Sel->getOperand(1)) == MaxRHS)
2190 NewRHS = Sel->getOperand(1);
2191 else if (SE->getSCEV(Sel->getOperand(2)) == MaxRHS)
2192 NewRHS = Sel->getOperand(2);
2193 if (!NewRHS) return Cond;
2195 // Determine the new comparison opcode. It may be signed or unsigned,
2196 // and the original comparison may be either equality or inequality.
2197 CmpInst::Predicate Pred =
2198 isa<SCEVSMaxExpr>(Max) ? CmpInst::ICMP_SLT : CmpInst::ICMP_ULT;
2199 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
2200 Pred = CmpInst::getInversePredicate(Pred);
2202 // Ok, everything looks ok to change the condition into an SLT or SGE and
2203 // delete the max calculation.
2205 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
2207 // Delete the max calculation instructions.
2208 Cond->replaceAllUsesWith(NewCond);
2209 CondUse->setUser(NewCond);
2210 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2211 Cond->eraseFromParent();
2212 Sel->eraseFromParent();
2213 if (Cmp->use_empty())
2214 Cmp->eraseFromParent();
2218 /// OptimizeShadowIV - If IV is used in a int-to-float cast
2219 /// inside the loop then try to eliminate the cast opeation.
2220 void LoopStrengthReduce::OptimizeShadowIV(Loop *L) {
2222 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
2223 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2226 for (unsigned Stride = 0, e = IU->StrideOrder.size(); Stride != e;
2228 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
2229 IU->IVUsesByStride.find(IU->StrideOrder[Stride]);
2230 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
2231 if (!isa<SCEVConstant>(SI->first))
2234 for (ilist<IVStrideUse>::iterator UI = SI->second->Users.begin(),
2235 E = SI->second->Users.end(); UI != E; /* empty */) {
2236 ilist<IVStrideUse>::iterator CandidateUI = UI;
2238 Instruction *ShadowUse = CandidateUI->getUser();
2239 const Type *DestTy = NULL;
2241 /* If shadow use is a int->float cast then insert a second IV
2242 to eliminate this cast.
2244 for (unsigned i = 0; i < n; ++i)
2250 for (unsigned i = 0; i < n; ++i, ++d)
2253 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser()))
2254 DestTy = UCast->getDestTy();
2255 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser()))
2256 DestTy = SCast->getDestTy();
2257 if (!DestTy) continue;
2260 // If target does not support DestTy natively then do not apply
2261 // this transformation.
2262 EVT DVT = TLI->getValueType(DestTy);
2263 if (!TLI->isTypeLegal(DVT)) continue;
2266 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
2268 if (PH->getNumIncomingValues() != 2) continue;
2270 const Type *SrcTy = PH->getType();
2271 int Mantissa = DestTy->getFPMantissaWidth();
2272 if (Mantissa == -1) continue;
2273 if ((int)SE->getTypeSizeInBits(SrcTy) > Mantissa)
2276 unsigned Entry, Latch;
2277 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
2285 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
2286 if (!Init) continue;
2287 Constant *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());
2289 BinaryOperator *Incr =
2290 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
2291 if (!Incr) continue;
2292 if (Incr->getOpcode() != Instruction::Add
2293 && Incr->getOpcode() != Instruction::Sub)
2296 /* Initialize new IV, double d = 0.0 in above example. */
2297 ConstantInt *C = NULL;
2298 if (Incr->getOperand(0) == PH)
2299 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
2300 else if (Incr->getOperand(1) == PH)
2301 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
2307 // Ignore negative constants, as the code below doesn't handle them
2308 // correctly. TODO: Remove this restriction.
2309 if (!C->getValue().isStrictlyPositive()) continue;
2311 /* Add new PHINode. */
2312 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);
2314 /* create new increment. '++d' in above example. */
2315 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
2316 BinaryOperator *NewIncr =
2317 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
2318 Instruction::FAdd : Instruction::FSub,
2319 NewPH, CFP, "IV.S.next.", Incr);
2321 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
2322 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
2324 /* Remove cast operation */
2325 ShadowUse->replaceAllUsesWith(NewPH);
2326 ShadowUse->eraseFromParent();
2333 /// OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
2334 /// uses in the loop, look to see if we can eliminate some, in favor of using
2335 /// common indvars for the different uses.
2336 void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
2337 // TODO: implement optzns here.
2339 OptimizeShadowIV(L);
2342 bool LoopStrengthReduce::StrideMightBeShared(const SCEV* Stride, Loop *L,
2344 int64_t SInt = cast<SCEVConstant>(Stride)->getValue()->getSExtValue();
2345 for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
2346 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
2347 IU->IVUsesByStride.find(IU->StrideOrder[i]);
2348 const SCEV *Share = SI->first;
2349 if (!isa<SCEVConstant>(SI->first) || Share == Stride)
2351 int64_t SSInt = cast<SCEVConstant>(Share)->getValue()->getSExtValue();
2353 return true; // This can definitely be reused.
2354 if (unsigned(abs64(SSInt)) < SInt || (SSInt % SInt) != 0)
2356 int64_t Scale = SSInt / SInt;
2357 bool AllUsesAreAddresses = true;
2358 bool AllUsesAreOutsideLoop = true;
2359 std::vector<BasedUser> UsersToProcess;
2360 const SCEV *CommonExprs = CollectIVUsers(SI->first, *SI->second, L,
2361 AllUsesAreAddresses,
2362 AllUsesAreOutsideLoop,
2364 if (AllUsesAreAddresses &&
2365 ValidScale(!CommonExprs->isZero(), Scale, UsersToProcess)) {
2368 // Any pre-inc iv use?
2369 IVUsersOfOneStride &StrideUses = *IU->IVUsesByStride[Share];
2370 for (ilist<IVStrideUse>::iterator I = StrideUses.Users.begin(),
2371 E = StrideUses.Users.end(); I != E; ++I) {
2372 if (!I->isUseOfPostIncrementedValue())
2380 /// isUsedByExitBranch - Return true if icmp is used by a loop terminating
2381 /// conditional branch or it's and / or with other conditions before being used
2382 /// as the condition.
2383 static bool isUsedByExitBranch(ICmpInst *Cond, Loop *L) {
2384 BasicBlock *CondBB = Cond->getParent();
2385 if (!L->isLoopExiting(CondBB))
2387 BranchInst *TermBr = dyn_cast<BranchInst>(CondBB->getTerminator());
2388 if (!TermBr || !TermBr->isConditional())
2391 Value *User = *Cond->use_begin();
2392 Instruction *UserInst = dyn_cast<Instruction>(User);
2394 (UserInst->getOpcode() == Instruction::And ||
2395 UserInst->getOpcode() == Instruction::Or)) {
2396 if (!UserInst->hasOneUse() || UserInst->getParent() != CondBB)
2398 User = *User->use_begin();
2399 UserInst = dyn_cast<Instruction>(User);
2401 return User == TermBr;
2404 static bool ShouldCountToZero(ICmpInst *Cond, IVStrideUse* &CondUse,
2405 ScalarEvolution *SE, Loop *L,
2406 const TargetLowering *TLI = 0) {
2407 if (!L->contains(Cond->getParent()))
2410 if (!isa<SCEVConstant>(CondUse->getOffset()))
2413 // Handle only tests for equality for the moment.
2414 if (!Cond->isEquality() || !Cond->hasOneUse())
2416 if (!isUsedByExitBranch(Cond, L))
2419 Value *CondOp0 = Cond->getOperand(0);
2420 const SCEV *IV = SE->getSCEV(CondOp0);
2421 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2422 if (!AR || !AR->isAffine())
2425 const SCEVConstant *SC = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE));
2426 if (!SC || SC->getValue()->getSExtValue() < 0)
2427 // If it's already counting down, don't do anything.
2430 // If the RHS of the comparison is not an loop invariant, the rewrite
2431 // cannot be done. Also bail out if it's already comparing against a zero.
2432 // If we are checking this before cmp stride optimization, check if it's
2433 // comparing against a already legal immediate.
2434 Value *RHS = Cond->getOperand(1);
2435 ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS);
2436 if (!L->isLoopInvariant(RHS) ||
2437 (RHSC && RHSC->isZero()) ||
2438 (RHSC && TLI && TLI->isLegalICmpImmediate(RHSC->getSExtValue())))
2441 // Make sure the IV is only used for counting. Value may be preinc or
2442 // postinc; 2 uses in either case.
2443 if (!CondOp0->hasNUses(2))
2449 /// OptimizeLoopTermCond - Change loop terminating condition to use the
2450 /// postinc iv when possible.
2451 void LoopStrengthReduce::OptimizeLoopTermCond(Loop *L) {
2452 BasicBlock *LatchBlock = L->getLoopLatch();
2453 bool LatchExit = L->isLoopExiting(LatchBlock);
2454 SmallVector<BasicBlock*, 8> ExitingBlocks;
2455 L->getExitingBlocks(ExitingBlocks);
2457 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
2458 BasicBlock *ExitingBlock = ExitingBlocks[i];
2460 // Finally, get the terminating condition for the loop if possible. If we
2461 // can, we want to change it to use a post-incremented version of its
2462 // induction variable, to allow coalescing the live ranges for the IV into
2463 // one register value.
2465 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2468 // FIXME: Overly conservative, termination condition could be an 'or' etc..
2469 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
2472 // Search IVUsesByStride to find Cond's IVUse if there is one.
2473 IVStrideUse *CondUse = 0;
2474 const SCEV *CondStride = 0;
2475 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2476 if (!FindIVUserForCond(Cond, CondUse, CondStride))
2479 // If the latch block is exiting and it's not a single block loop, it's
2480 // not safe to use postinc iv in other exiting blocks. FIXME: overly
2481 // conservative? How about icmp stride optimization?
2482 bool UsePostInc = !(e > 1 && LatchExit && ExitingBlock != LatchBlock);
2483 if (UsePostInc && ExitingBlock != LatchBlock) {
2484 if (!Cond->hasOneUse())
2485 // See below, we don't want the condition to be cloned.
2488 // If exiting block is the latch block, we know it's safe and profitable
2489 // to transform the icmp to use post-inc iv. Otherwise do so only if it
2490 // would not reuse another iv and its iv would be reused by other uses.
2491 // We are optimizing for the case where the icmp is the only use of the
2493 IVUsersOfOneStride &StrideUses = *IU->IVUsesByStride[CondStride];
2494 for (ilist<IVStrideUse>::iterator I = StrideUses.Users.begin(),
2495 E = StrideUses.Users.end(); I != E; ++I) {
2496 if (I->getUser() == Cond)
2498 if (!I->isUseOfPostIncrementedValue()) {
2505 // If iv for the stride might be shared and any of the users use pre-inc
2506 // iv might be used, then it's not safe to use post-inc iv.
2508 isa<SCEVConstant>(CondStride) &&
2509 StrideMightBeShared(CondStride, L, true))
2513 // If the trip count is computed in terms of a max (due to ScalarEvolution
2514 // being unable to find a sufficient guard, for example), change the loop
2515 // comparison to use SLT or ULT instead of NE.
2516 Cond = OptimizeMax(L, Cond, CondUse);
2518 // If possible, change stride and operands of the compare instruction to
2519 // eliminate one stride. However, avoid rewriting the compare instruction
2520 // with an iv of new stride if it's likely the new stride uses will be
2521 // rewritten using the stride of the compare instruction.
2522 if (ExitingBlock == LatchBlock && isa<SCEVConstant>(CondStride)) {
2523 // If the condition stride is a constant and it's the only use, we might
2524 // want to optimize it first by turning it to count toward zero.
2525 if (!StrideMightBeShared(CondStride, L, false) &&
2526 !ShouldCountToZero(Cond, CondUse, SE, L, TLI))
2527 Cond = ChangeCompareStride(L, Cond, CondUse, CondStride);
2533 DEBUG(errs() << " Change loop exiting icmp to use postinc iv: "
2536 // It's possible for the setcc instruction to be anywhere in the loop, and
2537 // possible for it to have multiple users. If it is not immediately before
2538 // the exiting block branch, move it.
2539 if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
2540 if (Cond->hasOneUse()) { // Condition has a single use, just move it.
2541 Cond->moveBefore(TermBr);
2543 // Otherwise, clone the terminating condition and insert into the
2545 Cond = cast<ICmpInst>(Cond->clone());
2546 Cond->setName(L->getHeader()->getName() + ".termcond");
2547 ExitingBlock->getInstList().insert(TermBr, Cond);
2549 // Clone the IVUse, as the old use still exists!
2550 IU->IVUsesByStride[CondStride]->addUser(CondUse->getOffset(), Cond,
2551 CondUse->getOperandValToReplace());
2552 CondUse = &IU->IVUsesByStride[CondStride]->Users.back();
2556 // If we get to here, we know that we can transform the setcc instruction to
2557 // use the post-incremented version of the IV, allowing us to coalesce the
2558 // live ranges for the IV correctly.
2559 CondUse->setOffset(SE->getMinusSCEV(CondUse->getOffset(), CondStride));
2560 CondUse->setIsUseOfPostIncrementedValue(true);
2567 bool LoopStrengthReduce::OptimizeLoopCountIVOfStride(const SCEV* &Stride,
2568 IVStrideUse* &CondUse,
2570 // If the only use is an icmp of a loop exiting conditional branch, then
2571 // attempt the optimization.
2572 BasedUser User = BasedUser(*CondUse, SE);
2573 assert(isa<ICmpInst>(User.Inst) && "Expecting an ICMPInst!");
2574 ICmpInst *Cond = cast<ICmpInst>(User.Inst);
2576 // Less strict check now that compare stride optimization is done.
2577 if (!ShouldCountToZero(Cond, CondUse, SE, L))
2580 Value *CondOp0 = Cond->getOperand(0);
2581 PHINode *PHIExpr = dyn_cast<PHINode>(CondOp0);
2584 // Value tested is postinc. Find the phi node.
2585 Incr = dyn_cast<BinaryOperator>(CondOp0);
2586 // FIXME: Just use User.OperandValToReplace here?
2587 if (!Incr || Incr->getOpcode() != Instruction::Add)
2590 PHIExpr = dyn_cast<PHINode>(Incr->getOperand(0));
2593 // 1 use for preinc value, the increment.
2594 if (!PHIExpr->hasOneUse())
2597 assert(isa<PHINode>(CondOp0) &&
2598 "Unexpected loop exiting counting instruction sequence!");
2599 PHIExpr = cast<PHINode>(CondOp0);
2600 // Value tested is preinc. Find the increment.
2601 // A CmpInst is not a BinaryOperator; we depend on this.
2602 Instruction::use_iterator UI = PHIExpr->use_begin();
2603 Incr = dyn_cast<BinaryOperator>(UI);
2605 Incr = dyn_cast<BinaryOperator>(++UI);
2606 // One use for postinc value, the phi. Unnecessarily conservative?
2607 if (!Incr || !Incr->hasOneUse() || Incr->getOpcode() != Instruction::Add)
2611 // Replace the increment with a decrement.
2612 DEBUG(errs() << "LSR: Examining use ");
2613 DEBUG(WriteAsOperand(errs(), CondOp0, /*PrintType=*/false));
2614 DEBUG(errs() << " in Inst: " << *Cond << '\n');
2615 BinaryOperator *Decr = BinaryOperator::Create(Instruction::Sub,
2616 Incr->getOperand(0), Incr->getOperand(1), "tmp", Incr);
2617 Incr->replaceAllUsesWith(Decr);
2618 Incr->eraseFromParent();
2620 // Substitute endval-startval for the original startval, and 0 for the
2621 // original endval. Since we're only testing for equality this is OK even
2622 // if the computation wraps around.
2623 BasicBlock *Preheader = L->getLoopPreheader();
2624 Instruction *PreInsertPt = Preheader->getTerminator();
2625 unsigned InBlock = L->contains(PHIExpr->getIncomingBlock(0)) ? 1 : 0;
2626 Value *StartVal = PHIExpr->getIncomingValue(InBlock);
2627 Value *EndVal = Cond->getOperand(1);
2628 DEBUG(errs() << " Optimize loop counting iv to count down ["
2629 << *EndVal << " .. " << *StartVal << "]\n");
2631 // FIXME: check for case where both are constant.
2632 Constant* Zero = ConstantInt::get(Cond->getOperand(1)->getType(), 0);
2633 BinaryOperator *NewStartVal = BinaryOperator::Create(Instruction::Sub,
2634 EndVal, StartVal, "tmp", PreInsertPt);
2635 PHIExpr->setIncomingValue(InBlock, NewStartVal);
2636 Cond->setOperand(1, Zero);
2637 DEBUG(errs() << " New icmp: " << *Cond << "\n");
2639 int64_t SInt = cast<SCEVConstant>(Stride)->getValue()->getSExtValue();
2640 const SCEV *NewStride = 0;
2642 for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
2643 const SCEV *OldStride = IU->StrideOrder[i];
2644 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(OldStride))
2645 if (SC->getValue()->getSExtValue() == -SInt) {
2647 NewStride = OldStride;
2653 NewStride = SE->getIntegerSCEV(-SInt, Stride->getType());
2654 IU->AddUser(NewStride, CondUse->getOffset(), Cond, Cond->getOperand(0));
2655 IU->IVUsesByStride[Stride]->removeUser(CondUse);
2657 CondUse = &IU->IVUsesByStride[NewStride]->Users.back();
2665 /// OptimizeLoopCountIV - If, after all sharing of IVs, the IV used for deciding
2666 /// when to exit the loop is used only for that purpose, try to rearrange things
2667 /// so it counts down to a test against zero.
2668 bool LoopStrengthReduce::OptimizeLoopCountIV(Loop *L) {
2669 bool ThisChanged = false;
2670 for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
2671 const SCEV *Stride = IU->StrideOrder[i];
2672 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
2673 IU->IVUsesByStride.find(Stride);
2674 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
2675 // FIXME: Generalize to non-affine IV's.
2676 if (!SI->first->isLoopInvariant(L))
2678 // If stride is a constant and it has an icmpinst use, check if we can
2679 // optimize the loop to count down.
2680 if (isa<SCEVConstant>(Stride) && SI->second->Users.size() == 1) {
2681 Instruction *User = SI->second->Users.begin()->getUser();
2682 if (!isa<ICmpInst>(User))
2684 const SCEV *CondStride = Stride;
2685 IVStrideUse *Use = &*SI->second->Users.begin();
2686 if (!OptimizeLoopCountIVOfStride(CondStride, Use, L))
2690 // Now check if it's possible to reuse this iv for other stride uses.
2691 for (unsigned j = 0, ee = IU->StrideOrder.size(); j != ee; ++j) {
2692 const SCEV *SStride = IU->StrideOrder[j];
2693 if (SStride == CondStride)
2695 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SII =
2696 IU->IVUsesByStride.find(SStride);
2697 assert(SII != IU->IVUsesByStride.end() && "Stride doesn't exist!");
2698 // FIXME: Generalize to non-affine IV's.
2699 if (!SII->first->isLoopInvariant(L))
2701 // FIXME: Rewrite other stride using CondStride.
2706 Changed |= ThisChanged;
2710 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager &LPM) {
2711 IU = &getAnalysis<IVUsers>();
2712 SE = &getAnalysis<ScalarEvolution>();
2715 // If LoopSimplify form is not available, stay out of trouble.
2716 if (!L->getLoopPreheader() || !L->getLoopLatch())
2719 if (!IU->IVUsesByStride.empty()) {
2720 DEBUG(errs() << "\nLSR on \"" << L->getHeader()->getParent()->getName()
2724 // Sort the StrideOrder so we process larger strides first.
2725 std::stable_sort(IU->StrideOrder.begin(), IU->StrideOrder.end(),
2728 // Optimize induction variables. Some indvar uses can be transformed to use
2729 // strides that will be needed for other purposes. A common example of this
2730 // is the exit test for the loop, which can often be rewritten to use the
2731 // computation of some other indvar to decide when to terminate the loop.
2734 // Change loop terminating condition to use the postinc iv when possible
2735 // and optimize loop terminating compare. FIXME: Move this after
2736 // StrengthReduceIVUsersOfStride?
2737 OptimizeLoopTermCond(L);
2739 // FIXME: We can shrink overlarge IV's here. e.g. if the code has
2740 // computation in i64 values and the target doesn't support i64, demote
2741 // the computation to 32-bit if safe.
2743 // FIXME: Attempt to reuse values across multiple IV's. In particular, we
2744 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should
2745 // be codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC.
2746 // Need to be careful that IV's are all the same type. Only works for
2747 // intptr_t indvars.
2749 // IVsByStride keeps IVs for one particular loop.
2750 assert(IVsByStride.empty() && "Stale entries in IVsByStride?");
2752 StrengthReduceIVUsers(L);
2754 // After all sharing is done, see if we can adjust the loop to test against
2755 // zero instead of counting up to a maximum. This is usually faster.
2756 OptimizeLoopCountIV(L);
2759 // We're done analyzing this loop; release all the state we built up for it.
2760 IVsByStride.clear();
2762 // Clean up after ourselves
2763 if (!DeadInsts.empty())
2764 DeleteTriviallyDeadInstructions();
2766 // At this point, it is worth checking to see if any recurrence PHIs are also
2767 // dead, so that we can remove them as well.
2768 DeleteDeadPHIs(L->getHeader());