1 //===- LoopStrengthReduce.cpp - Strength Reduce IVs in Loops --------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This transformation analyzes and transforms the induction variables (and
11 // computations derived from them) into forms suitable for efficient execution
14 // This pass performs a strength reduction on array references inside loops that
15 // have as one or more of their components the loop induction variable, it
16 // rewrites expressions to take advantage of scaled-index addressing modes
17 // available on the target, and it performs a variety of other optimizations
18 // related to loop induction variables.
20 //===----------------------------------------------------------------------===//
22 #define DEBUG_TYPE "loop-reduce"
23 #include "llvm/Transforms/Scalar.h"
24 #include "llvm/Constants.h"
25 #include "llvm/Instructions.h"
26 #include "llvm/IntrinsicInst.h"
27 #include "llvm/Type.h"
28 #include "llvm/DerivedTypes.h"
29 #include "llvm/Analysis/Dominators.h"
30 #include "llvm/Analysis/IVUsers.h"
31 #include "llvm/Analysis/LoopInfo.h"
32 #include "llvm/Analysis/LoopPass.h"
33 #include "llvm/Analysis/ScalarEvolutionExpander.h"
34 #include "llvm/Transforms/Utils/AddrModeMatcher.h"
35 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
36 #include "llvm/Transforms/Utils/Local.h"
37 #include "llvm/ADT/Statistic.h"
38 #include "llvm/Support/CFG.h"
39 #include "llvm/Support/Debug.h"
40 #include "llvm/Support/CommandLine.h"
41 #include "llvm/Support/ValueHandle.h"
42 #include "llvm/Support/raw_ostream.h"
43 #include "llvm/Target/TargetLowering.h"
47 STATISTIC(NumReduced , "Number of IV uses strength reduced");
48 STATISTIC(NumInserted, "Number of PHIs inserted");
49 STATISTIC(NumVariable, "Number of PHIs with variable strides");
50 STATISTIC(NumEliminated, "Number of strides eliminated");
51 STATISTIC(NumShadow, "Number of Shadow IVs optimized");
52 STATISTIC(NumImmSunk, "Number of common expr immediates sunk into uses");
53 STATISTIC(NumLoopCond, "Number of loop terminating conds optimized");
54 STATISTIC(NumCountZero, "Number of count iv optimized to count toward zero");
56 static cl::opt<bool> EnableFullLSRMode("enable-full-lsr",
64 /// IVInfo - This structure keeps track of one IV expression inserted during
65 /// StrengthReduceStridedIVUsers. It contains the stride, the common base, as
66 /// well as the PHI node and increment value created for rewrite.
72 IVExpr(const SCEV *const stride, const SCEV *const base, PHINode *phi)
73 : Stride(stride), Base(base), PHI(phi) {}
76 /// IVsOfOneStride - This structure keeps track of all IV expression inserted
77 /// during StrengthReduceStridedIVUsers for a particular stride of the IV.
78 struct IVsOfOneStride {
79 std::vector<IVExpr> IVs;
81 void addIV(const SCEV *const Stride, const SCEV *const Base, PHINode *PHI) {
82 IVs.push_back(IVExpr(Stride, Base, PHI));
86 class LoopStrengthReduce : public LoopPass {
93 /// IVsByStride - Keep track of all IVs that have been inserted for a
94 /// particular stride.
95 std::map<const SCEV *, IVsOfOneStride> IVsByStride;
97 /// StrideNoReuse - Keep track of all the strides whose ivs cannot be
98 /// reused (nor should they be rewritten to reuse other strides).
99 SmallSet<const SCEV *, 4> StrideNoReuse;
101 /// DeadInsts - Keep track of instructions we may have made dead, so that
102 /// we can remove them after we are done working.
103 SmallVector<WeakVH, 16> DeadInsts;
105 /// TLI - Keep a pointer of a TargetLowering to consult for determining
106 /// transformation profitability.
107 const TargetLowering *TLI;
110 static char ID; // Pass ID, replacement for typeid
111 explicit LoopStrengthReduce(const TargetLowering *tli = NULL) :
112 LoopPass(&ID), TLI(tli) {
115 bool runOnLoop(Loop *L, LPPassManager &LPM);
117 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
118 // We split critical edges, so we change the CFG. However, we do update
119 // many analyses if they are around.
120 AU.addPreservedID(LoopSimplifyID);
121 AU.addPreserved<LoopInfo>();
122 AU.addPreserved<DominanceFrontier>();
123 AU.addPreserved<DominatorTree>();
125 AU.addRequiredID(LoopSimplifyID);
126 AU.addRequired<LoopInfo>();
127 AU.addRequired<DominatorTree>();
128 AU.addRequired<ScalarEvolution>();
129 AU.addPreserved<ScalarEvolution>();
130 AU.addRequired<IVUsers>();
131 AU.addPreserved<IVUsers>();
135 ICmpInst *ChangeCompareStride(Loop *L, ICmpInst *Cond,
136 IVStrideUse* &CondUse,
137 const SCEV *const * &CondStride);
139 void OptimizeIndvars(Loop *L);
140 void OptimizeLoopCountIV(const SCEV *Stride,
141 IVUsersOfOneStride &Uses, Loop *L);
142 void OptimizeLoopTermCond(Loop *L);
144 /// OptimizeShadowIV - If IV is used in a int-to-float cast
145 /// inside the loop then try to eliminate the cast opeation.
146 void OptimizeShadowIV(Loop *L);
148 /// OptimizeMax - Rewrite the loop's terminating condition
149 /// if it uses a max computation.
150 ICmpInst *OptimizeMax(Loop *L, ICmpInst *Cond,
151 IVStrideUse* &CondUse);
153 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse,
154 const SCEV *const * &CondStride);
155 bool RequiresTypeConversion(const Type *Ty, const Type *NewTy);
156 const SCEV *CheckForIVReuse(bool, bool, bool, const SCEV *const&,
157 IVExpr&, const Type*,
158 const std::vector<BasedUser>& UsersToProcess);
159 bool ValidScale(bool, int64_t,
160 const std::vector<BasedUser>& UsersToProcess);
161 bool ValidOffset(bool, int64_t, int64_t,
162 const std::vector<BasedUser>& UsersToProcess);
163 const SCEV *CollectIVUsers(const SCEV *const &Stride,
164 IVUsersOfOneStride &Uses,
166 bool &AllUsesAreAddresses,
167 bool &AllUsesAreOutsideLoop,
168 std::vector<BasedUser> &UsersToProcess);
169 bool ShouldUseFullStrengthReductionMode(
170 const std::vector<BasedUser> &UsersToProcess,
172 bool AllUsesAreAddresses,
174 void PrepareToStrengthReduceFully(
175 std::vector<BasedUser> &UsersToProcess,
177 const SCEV *CommonExprs,
179 SCEVExpander &PreheaderRewriter);
180 void PrepareToStrengthReduceFromSmallerStride(
181 std::vector<BasedUser> &UsersToProcess,
183 const IVExpr &ReuseIV,
184 Instruction *PreInsertPt);
185 void PrepareToStrengthReduceWithNewPhi(
186 std::vector<BasedUser> &UsersToProcess,
188 const SCEV *CommonExprs,
190 Instruction *IVIncInsertPt,
192 SCEVExpander &PreheaderRewriter);
193 void StrengthReduceStridedIVUsers(const SCEV *const &Stride,
194 IVUsersOfOneStride &Uses,
196 void DeleteTriviallyDeadInstructions();
200 char LoopStrengthReduce::ID = 0;
201 static RegisterPass<LoopStrengthReduce>
202 X("loop-reduce", "Loop Strength Reduction");
204 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
205 return new LoopStrengthReduce(TLI);
208 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
209 /// specified set are trivially dead, delete them and see if this makes any of
210 /// their operands subsequently dead.
211 void LoopStrengthReduce::DeleteTriviallyDeadInstructions() {
212 if (DeadInsts.empty()) return;
214 while (!DeadInsts.empty()) {
215 Instruction *I = dyn_cast_or_null<Instruction>(DeadInsts.back());
216 DeadInsts.pop_back();
218 if (I == 0 || !isInstructionTriviallyDead(I))
221 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) {
222 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
225 DeadInsts.push_back(U);
229 I->eraseFromParent();
234 /// containsAddRecFromDifferentLoop - Determine whether expression S involves a
235 /// subexpression that is an AddRec from a loop other than L. An outer loop
236 /// of L is OK, but not an inner loop nor a disjoint loop.
237 static bool containsAddRecFromDifferentLoop(const SCEV *S, Loop *L) {
238 // This is very common, put it first.
239 if (isa<SCEVConstant>(S))
241 if (const SCEVCommutativeExpr *AE = dyn_cast<SCEVCommutativeExpr>(S)) {
242 for (unsigned int i=0; i< AE->getNumOperands(); i++)
243 if (containsAddRecFromDifferentLoop(AE->getOperand(i), L))
247 if (const SCEVAddRecExpr *AE = dyn_cast<SCEVAddRecExpr>(S)) {
248 if (const Loop *newLoop = AE->getLoop()) {
251 // if newLoop is an outer loop of L, this is OK.
252 if (!LoopInfo::isNotAlreadyContainedIn(L, newLoop))
257 if (const SCEVUDivExpr *DE = dyn_cast<SCEVUDivExpr>(S))
258 return containsAddRecFromDifferentLoop(DE->getLHS(), L) ||
259 containsAddRecFromDifferentLoop(DE->getRHS(), L);
261 // SCEVSDivExpr has been backed out temporarily, but will be back; we'll
262 // need this when it is.
263 if (const SCEVSDivExpr *DE = dyn_cast<SCEVSDivExpr>(S))
264 return containsAddRecFromDifferentLoop(DE->getLHS(), L) ||
265 containsAddRecFromDifferentLoop(DE->getRHS(), L);
267 if (const SCEVCastExpr *CE = dyn_cast<SCEVCastExpr>(S))
268 return containsAddRecFromDifferentLoop(CE->getOperand(), L);
272 /// isAddressUse - Returns true if the specified instruction is using the
273 /// specified value as an address.
274 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
275 bool isAddress = isa<LoadInst>(Inst);
276 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
277 if (SI->getOperand(1) == OperandVal)
279 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
280 // Addressing modes can also be folded into prefetches and a variety
282 switch (II->getIntrinsicID()) {
284 case Intrinsic::prefetch:
285 case Intrinsic::x86_sse2_loadu_dq:
286 case Intrinsic::x86_sse2_loadu_pd:
287 case Intrinsic::x86_sse_loadu_ps:
288 case Intrinsic::x86_sse_storeu_ps:
289 case Intrinsic::x86_sse2_storeu_pd:
290 case Intrinsic::x86_sse2_storeu_dq:
291 case Intrinsic::x86_sse2_storel_dq:
292 if (II->getOperand(1) == OperandVal)
300 /// getAccessType - Return the type of the memory being accessed.
301 static const Type *getAccessType(const Instruction *Inst) {
302 const Type *AccessTy = Inst->getType();
303 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
304 AccessTy = SI->getOperand(0)->getType();
305 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
306 // Addressing modes can also be folded into prefetches and a variety
308 switch (II->getIntrinsicID()) {
310 case Intrinsic::x86_sse_storeu_ps:
311 case Intrinsic::x86_sse2_storeu_pd:
312 case Intrinsic::x86_sse2_storeu_dq:
313 case Intrinsic::x86_sse2_storel_dq:
314 AccessTy = II->getOperand(1)->getType();
322 /// BasedUser - For a particular base value, keep information about how we've
323 /// partitioned the expression so far.
325 /// SE - The current ScalarEvolution object.
328 /// Base - The Base value for the PHI node that needs to be inserted for
329 /// this use. As the use is processed, information gets moved from this
330 /// field to the Imm field (below). BasedUser values are sorted by this
334 /// Inst - The instruction using the induction variable.
337 /// OperandValToReplace - The operand value of Inst to replace with the
339 Value *OperandValToReplace;
341 /// Imm - The immediate value that should be added to the base immediately
342 /// before Inst, because it will be folded into the imm field of the
343 /// instruction. This is also sometimes used for loop-variant values that
344 /// must be added inside the loop.
347 /// Phi - The induction variable that performs the striding that
348 /// should be used for this user.
351 // isUseOfPostIncrementedValue - True if this should use the
352 // post-incremented version of this IV, not the preincremented version.
353 // This can only be set in special cases, such as the terminating setcc
354 // instruction for a loop and uses outside the loop that are dominated by
356 bool isUseOfPostIncrementedValue;
358 BasedUser(IVStrideUse &IVSU, ScalarEvolution *se)
359 : SE(se), Base(IVSU.getOffset()), Inst(IVSU.getUser()),
360 OperandValToReplace(IVSU.getOperandValToReplace()),
361 Imm(SE->getIntegerSCEV(0, Base->getType())),
362 isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue()) {}
364 // Once we rewrite the code to insert the new IVs we want, update the
365 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
367 void RewriteInstructionToUseNewBase(const SCEV *const &NewBase,
368 Instruction *InsertPt,
369 SCEVExpander &Rewriter, Loop *L, Pass *P,
371 SmallVectorImpl<WeakVH> &DeadInsts);
373 Value *InsertCodeForBaseAtPosition(const SCEV *const &NewBase,
375 SCEVExpander &Rewriter,
376 Instruction *IP, Loop *L,
382 void BasedUser::dump() const {
383 errs() << " Base=" << *Base;
384 errs() << " Imm=" << *Imm;
385 errs() << " Inst: " << *Inst;
388 Value *BasedUser::InsertCodeForBaseAtPosition(const SCEV *const &NewBase,
390 SCEVExpander &Rewriter,
391 Instruction *IP, Loop *L,
393 // Figure out where we *really* want to insert this code. In particular, if
394 // the user is inside of a loop that is nested inside of L, we really don't
395 // want to insert this expression before the user, we'd rather pull it out as
396 // many loops as possible.
397 Instruction *BaseInsertPt = IP;
399 // Figure out the most-nested loop that IP is in.
400 Loop *InsertLoop = LI.getLoopFor(IP->getParent());
402 // If InsertLoop is not L, and InsertLoop is nested inside of L, figure out
403 // the preheader of the outer-most loop where NewBase is not loop invariant.
404 if (L->contains(IP->getParent()))
405 while (InsertLoop && NewBase->isLoopInvariant(InsertLoop)) {
406 BaseInsertPt = InsertLoop->getLoopPreheader()->getTerminator();
407 InsertLoop = InsertLoop->getParentLoop();
410 Value *Base = Rewriter.expandCodeFor(NewBase, 0, BaseInsertPt);
412 const SCEV *NewValSCEV = SE->getUnknown(Base);
414 // Always emit the immediate into the same block as the user.
415 NewValSCEV = SE->getAddExpr(NewValSCEV, Imm);
417 return Rewriter.expandCodeFor(NewValSCEV, Ty, IP);
421 // Once we rewrite the code to insert the new IVs we want, update the
422 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
423 // to it. NewBasePt is the last instruction which contributes to the
424 // value of NewBase in the case that it's a diffferent instruction from
425 // the PHI that NewBase is computed from, or null otherwise.
427 void BasedUser::RewriteInstructionToUseNewBase(const SCEV *const &NewBase,
428 Instruction *NewBasePt,
429 SCEVExpander &Rewriter, Loop *L, Pass *P,
431 SmallVectorImpl<WeakVH> &DeadInsts) {
432 if (!isa<PHINode>(Inst)) {
433 // By default, insert code at the user instruction.
434 BasicBlock::iterator InsertPt = Inst;
436 // However, if the Operand is itself an instruction, the (potentially
437 // complex) inserted code may be shared by many users. Because of this, we
438 // want to emit code for the computation of the operand right before its old
439 // computation. This is usually safe, because we obviously used to use the
440 // computation when it was computed in its current block. However, in some
441 // cases (e.g. use of a post-incremented induction variable) the NewBase
442 // value will be pinned to live somewhere after the original computation.
443 // In this case, we have to back off.
445 // If this is a use outside the loop (which means after, since it is based
446 // on a loop indvar) we use the post-incremented value, so that we don't
447 // artificially make the preinc value live out the bottom of the loop.
448 if (!isUseOfPostIncrementedValue && L->contains(Inst->getParent())) {
449 if (NewBasePt && isa<PHINode>(OperandValToReplace)) {
450 InsertPt = NewBasePt;
452 } else if (Instruction *OpInst
453 = dyn_cast<Instruction>(OperandValToReplace)) {
455 while (isa<PHINode>(InsertPt)) ++InsertPt;
458 Value *NewVal = InsertCodeForBaseAtPosition(NewBase,
459 OperandValToReplace->getType(),
460 Rewriter, InsertPt, L, LI);
461 // Replace the use of the operand Value with the new Phi we just created.
462 Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
464 DEBUG(errs() << " Replacing with ");
465 DEBUG(WriteAsOperand(errs(), NewVal, /*PrintType=*/false));
466 DEBUG(errs() << ", which has value " << *NewBase << " plus IMM "
471 // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm
472 // expression into each operand block that uses it. Note that PHI nodes can
473 // have multiple entries for the same predecessor. We use a map to make sure
474 // that a PHI node only has a single Value* for each predecessor (which also
475 // prevents us from inserting duplicate code in some blocks).
476 DenseMap<BasicBlock*, Value*> InsertedCode;
477 PHINode *PN = cast<PHINode>(Inst);
478 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
479 if (PN->getIncomingValue(i) == OperandValToReplace) {
480 // If the original expression is outside the loop, put the replacement
481 // code in the same place as the original expression,
482 // which need not be an immediate predecessor of this PHI. This way we
483 // need only one copy of it even if it is referenced multiple times in
484 // the PHI. We don't do this when the original expression is inside the
485 // loop because multiple copies sometimes do useful sinking of code in
487 Instruction *OldLoc = dyn_cast<Instruction>(OperandValToReplace);
488 BasicBlock *PHIPred = PN->getIncomingBlock(i);
489 if (L->contains(OldLoc->getParent())) {
490 // If this is a critical edge, split the edge so that we do not insert
491 // the code on all predecessor/successor paths. We do this unless this
492 // is the canonical backedge for this loop, as this can make some
493 // inserted code be in an illegal position.
494 if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 &&
495 !isa<IndirectBrInst>(PHIPred->getTerminator()) &&
496 (PN->getParent() != L->getHeader() || !L->contains(PHIPred))) {
498 // First step, split the critical edge.
499 BasicBlock *NewBB = SplitCriticalEdge(PHIPred, PN->getParent(),
502 // Next step: move the basic block. In particular, if the PHI node
503 // is outside of the loop, and PredTI is in the loop, we want to
504 // move the block to be immediately before the PHI block, not
505 // immediately after PredTI.
506 if (L->contains(PHIPred) && !L->contains(PN->getParent()))
507 NewBB->moveBefore(PN->getParent());
509 // Splitting the edge can reduce the number of PHI entries we have.
510 e = PN->getNumIncomingValues();
512 i = PN->getBasicBlockIndex(PHIPred);
515 Value *&Code = InsertedCode[PHIPred];
517 // Insert the code into the end of the predecessor block.
518 Instruction *InsertPt = (L->contains(OldLoc->getParent())) ?
519 PHIPred->getTerminator() :
520 OldLoc->getParent()->getTerminator();
521 Code = InsertCodeForBaseAtPosition(NewBase, PN->getType(),
522 Rewriter, InsertPt, L, LI);
524 DEBUG(errs() << " Changing PHI use to ");
525 DEBUG(WriteAsOperand(errs(), Code, /*PrintType=*/false));
526 DEBUG(errs() << ", which has value " << *NewBase << " plus IMM "
530 // Replace the use of the operand Value with the new Phi we just created.
531 PN->setIncomingValue(i, Code);
536 // PHI node might have become a constant value after SplitCriticalEdge.
537 DeadInsts.push_back(Inst);
541 /// fitsInAddressMode - Return true if V can be subsumed within an addressing
542 /// mode, and does not need to be put in a register first.
543 static bool fitsInAddressMode(const SCEV *const &V, const Type *AccessTy,
544 const TargetLowering *TLI, bool HasBaseReg) {
545 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
546 int64_t VC = SC->getValue()->getSExtValue();
548 TargetLowering::AddrMode AM;
550 AM.HasBaseReg = HasBaseReg;
551 return TLI->isLegalAddressingMode(AM, AccessTy);
553 // Defaults to PPC. PPC allows a sign-extended 16-bit immediate field.
554 return (VC > -(1 << 16) && VC < (1 << 16)-1);
558 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
559 if (GlobalValue *GV = dyn_cast<GlobalValue>(SU->getValue())) {
561 TargetLowering::AddrMode AM;
563 AM.HasBaseReg = HasBaseReg;
564 return TLI->isLegalAddressingMode(AM, AccessTy);
566 // Default: assume global addresses are not legal.
573 /// MoveLoopVariantsToImmediateField - Move any subexpressions from Val that are
574 /// loop varying to the Imm operand.
575 static void MoveLoopVariantsToImmediateField(const SCEV *&Val, const SCEV *&Imm,
576 Loop *L, ScalarEvolution *SE) {
577 if (Val->isLoopInvariant(L)) return; // Nothing to do.
579 if (const SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
580 SmallVector<const SCEV *, 4> NewOps;
581 NewOps.reserve(SAE->getNumOperands());
583 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
584 if (!SAE->getOperand(i)->isLoopInvariant(L)) {
585 // If this is a loop-variant expression, it must stay in the immediate
586 // field of the expression.
587 Imm = SE->getAddExpr(Imm, SAE->getOperand(i));
589 NewOps.push_back(SAE->getOperand(i));
593 Val = SE->getIntegerSCEV(0, Val->getType());
595 Val = SE->getAddExpr(NewOps);
596 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
597 // Try to pull immediates out of the start value of nested addrec's.
598 const SCEV *Start = SARE->getStart();
599 MoveLoopVariantsToImmediateField(Start, Imm, L, SE);
601 SmallVector<const SCEV *, 4> Ops(SARE->op_begin(), SARE->op_end());
603 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
605 // Otherwise, all of Val is variant, move the whole thing over.
606 Imm = SE->getAddExpr(Imm, Val);
607 Val = SE->getIntegerSCEV(0, Val->getType());
612 /// MoveImmediateValues - Look at Val, and pull out any additions of constants
613 /// that can fit into the immediate field of instructions in the target.
614 /// Accumulate these immediate values into the Imm value.
615 static void MoveImmediateValues(const TargetLowering *TLI,
616 const Type *AccessTy,
617 const SCEV *&Val, const SCEV *&Imm,
618 bool isAddress, Loop *L,
619 ScalarEvolution *SE) {
620 if (const SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
621 SmallVector<const SCEV *, 4> NewOps;
622 NewOps.reserve(SAE->getNumOperands());
624 for (unsigned i = 0; i != SAE->getNumOperands(); ++i) {
625 const SCEV *NewOp = SAE->getOperand(i);
626 MoveImmediateValues(TLI, AccessTy, NewOp, Imm, isAddress, L, SE);
628 if (!NewOp->isLoopInvariant(L)) {
629 // If this is a loop-variant expression, it must stay in the immediate
630 // field of the expression.
631 Imm = SE->getAddExpr(Imm, NewOp);
633 NewOps.push_back(NewOp);
638 Val = SE->getIntegerSCEV(0, Val->getType());
640 Val = SE->getAddExpr(NewOps);
642 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
643 // Try to pull immediates out of the start value of nested addrec's.
644 const SCEV *Start = SARE->getStart();
645 MoveImmediateValues(TLI, AccessTy, Start, Imm, isAddress, L, SE);
647 if (Start != SARE->getStart()) {
648 SmallVector<const SCEV *, 4> Ops(SARE->op_begin(), SARE->op_end());
650 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
653 } else if (const SCEVMulExpr *SME = dyn_cast<SCEVMulExpr>(Val)) {
654 // Transform "8 * (4 + v)" -> "32 + 8*V" if "32" fits in the immed field.
656 fitsInAddressMode(SME->getOperand(0), AccessTy, TLI, false) &&
657 SME->getNumOperands() == 2 && SME->isLoopInvariant(L)) {
659 const SCEV *SubImm = SE->getIntegerSCEV(0, Val->getType());
660 const SCEV *NewOp = SME->getOperand(1);
661 MoveImmediateValues(TLI, AccessTy, NewOp, SubImm, isAddress, L, SE);
663 // If we extracted something out of the subexpressions, see if we can
665 if (NewOp != SME->getOperand(1)) {
666 // Scale SubImm up by "8". If the result is a target constant, we are
668 SubImm = SE->getMulExpr(SubImm, SME->getOperand(0));
669 if (fitsInAddressMode(SubImm, AccessTy, TLI, false)) {
670 // Accumulate the immediate.
671 Imm = SE->getAddExpr(Imm, SubImm);
673 // Update what is left of 'Val'.
674 Val = SE->getMulExpr(SME->getOperand(0), NewOp);
681 // Loop-variant expressions must stay in the immediate field of the
683 if ((isAddress && fitsInAddressMode(Val, AccessTy, TLI, false)) ||
684 !Val->isLoopInvariant(L)) {
685 Imm = SE->getAddExpr(Imm, Val);
686 Val = SE->getIntegerSCEV(0, Val->getType());
690 // Otherwise, no immediates to move.
693 static void MoveImmediateValues(const TargetLowering *TLI,
695 const SCEV *&Val, const SCEV *&Imm,
696 bool isAddress, Loop *L,
697 ScalarEvolution *SE) {
698 const Type *AccessTy = getAccessType(User);
699 MoveImmediateValues(TLI, AccessTy, Val, Imm, isAddress, L, SE);
702 /// SeparateSubExprs - Decompose Expr into all of the subexpressions that are
703 /// added together. This is used to reassociate common addition subexprs
704 /// together for maximal sharing when rewriting bases.
705 static void SeparateSubExprs(SmallVector<const SCEV *, 16> &SubExprs,
707 ScalarEvolution *SE) {
708 if (const SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) {
709 for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
710 SeparateSubExprs(SubExprs, AE->getOperand(j), SE);
711 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) {
712 const SCEV *Zero = SE->getIntegerSCEV(0, Expr->getType());
713 if (SARE->getOperand(0) == Zero) {
714 SubExprs.push_back(Expr);
716 // Compute the addrec with zero as its base.
717 SmallVector<const SCEV *, 4> Ops(SARE->op_begin(), SARE->op_end());
718 Ops[0] = Zero; // Start with zero base.
719 SubExprs.push_back(SE->getAddRecExpr(Ops, SARE->getLoop()));
722 SeparateSubExprs(SubExprs, SARE->getOperand(0), SE);
724 } else if (!Expr->isZero()) {
726 SubExprs.push_back(Expr);
730 // This is logically local to the following function, but C++ says we have
731 // to make it file scope.
732 struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; };
734 /// RemoveCommonExpressionsFromUseBases - Look through all of the Bases of all
735 /// the Uses, removing any common subexpressions, except that if all such
736 /// subexpressions can be folded into an addressing mode for all uses inside
737 /// the loop (this case is referred to as "free" in comments herein) we do
738 /// not remove anything. This looks for things like (a+b+c) and
739 /// (a+c+d) and computes the common (a+c) subexpression. The common expression
740 /// is *removed* from the Bases and returned.
742 RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses,
743 ScalarEvolution *SE, Loop *L,
744 const TargetLowering *TLI) {
745 unsigned NumUses = Uses.size();
747 // Only one use? This is a very common case, so we handle it specially and
749 const SCEV *Zero = SE->getIntegerSCEV(0, Uses[0].Base->getType());
750 const SCEV *Result = Zero;
751 const SCEV *FreeResult = Zero;
753 // If the use is inside the loop, use its base, regardless of what it is:
754 // it is clearly shared across all the IV's. If the use is outside the loop
755 // (which means after it) we don't want to factor anything *into* the loop,
756 // so just use 0 as the base.
757 if (L->contains(Uses[0].Inst->getParent()))
758 std::swap(Result, Uses[0].Base);
762 // To find common subexpressions, count how many of Uses use each expression.
763 // If any subexpressions are used Uses.size() times, they are common.
764 // Also track whether all uses of each expression can be moved into an
765 // an addressing mode "for free"; such expressions are left within the loop.
766 // struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; };
767 std::map<const SCEV *, SubExprUseData> SubExpressionUseData;
769 // UniqueSubExprs - Keep track of all of the subexpressions we see in the
770 // order we see them.
771 SmallVector<const SCEV *, 16> UniqueSubExprs;
773 SmallVector<const SCEV *, 16> SubExprs;
774 unsigned NumUsesInsideLoop = 0;
775 for (unsigned i = 0; i != NumUses; ++i) {
776 // If the user is outside the loop, just ignore it for base computation.
777 // Since the user is outside the loop, it must be *after* the loop (if it
778 // were before, it could not be based on the loop IV). We don't want users
779 // after the loop to affect base computation of values *inside* the loop,
780 // because we can always add their offsets to the result IV after the loop
781 // is done, ensuring we get good code inside the loop.
782 if (!L->contains(Uses[i].Inst->getParent()))
786 // If the base is zero (which is common), return zero now, there are no
788 if (Uses[i].Base == Zero) return Zero;
790 // If this use is as an address we may be able to put CSEs in the addressing
791 // mode rather than hoisting them.
792 bool isAddrUse = isAddressUse(Uses[i].Inst, Uses[i].OperandValToReplace);
793 // We may need the AccessTy below, but only when isAddrUse, so compute it
794 // only in that case.
795 const Type *AccessTy = 0;
797 AccessTy = getAccessType(Uses[i].Inst);
799 // Split the expression into subexprs.
800 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
801 // Add one to SubExpressionUseData.Count for each subexpr present, and
802 // if the subexpr is not a valid immediate within an addressing mode use,
803 // set SubExpressionUseData.notAllUsesAreFree. We definitely want to
804 // hoist these out of the loop (if they are common to all uses).
805 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) {
806 if (++SubExpressionUseData[SubExprs[j]].Count == 1)
807 UniqueSubExprs.push_back(SubExprs[j]);
808 if (!isAddrUse || !fitsInAddressMode(SubExprs[j], AccessTy, TLI, false))
809 SubExpressionUseData[SubExprs[j]].notAllUsesAreFree = true;
814 // Now that we know how many times each is used, build Result. Iterate over
815 // UniqueSubexprs so that we have a stable ordering.
816 for (unsigned i = 0, e = UniqueSubExprs.size(); i != e; ++i) {
817 std::map<const SCEV *, SubExprUseData>::iterator I =
818 SubExpressionUseData.find(UniqueSubExprs[i]);
819 assert(I != SubExpressionUseData.end() && "Entry not found?");
820 if (I->second.Count == NumUsesInsideLoop) { // Found CSE!
821 if (I->second.notAllUsesAreFree)
822 Result = SE->getAddExpr(Result, I->first);
824 FreeResult = SE->getAddExpr(FreeResult, I->first);
826 // Remove non-cse's from SubExpressionUseData.
827 SubExpressionUseData.erase(I);
830 if (FreeResult != Zero) {
831 // We have some subexpressions that can be subsumed into addressing
832 // modes in every use inside the loop. However, it's possible that
833 // there are so many of them that the combined FreeResult cannot
834 // be subsumed, or that the target cannot handle both a FreeResult
835 // and a Result in the same instruction (for example because it would
836 // require too many registers). Check this.
837 for (unsigned i=0; i<NumUses; ++i) {
838 if (!L->contains(Uses[i].Inst->getParent()))
840 // We know this is an addressing mode use; if there are any uses that
841 // are not, FreeResult would be Zero.
842 const Type *AccessTy = getAccessType(Uses[i].Inst);
843 if (!fitsInAddressMode(FreeResult, AccessTy, TLI, Result!=Zero)) {
844 // FIXME: could split up FreeResult into pieces here, some hoisted
845 // and some not. There is no obvious advantage to this.
846 Result = SE->getAddExpr(Result, FreeResult);
853 // If we found no CSE's, return now.
854 if (Result == Zero) return Result;
856 // If we still have a FreeResult, remove its subexpressions from
857 // SubExpressionUseData. This means they will remain in the use Bases.
858 if (FreeResult != Zero) {
859 SeparateSubExprs(SubExprs, FreeResult, SE);
860 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) {
861 std::map<const SCEV *, SubExprUseData>::iterator I =
862 SubExpressionUseData.find(SubExprs[j]);
863 SubExpressionUseData.erase(I);
868 // Otherwise, remove all of the CSE's we found from each of the base values.
869 for (unsigned i = 0; i != NumUses; ++i) {
870 // Uses outside the loop don't necessarily include the common base, but
871 // the final IV value coming into those uses does. Instead of trying to
872 // remove the pieces of the common base, which might not be there,
873 // subtract off the base to compensate for this.
874 if (!L->contains(Uses[i].Inst->getParent())) {
875 Uses[i].Base = SE->getMinusSCEV(Uses[i].Base, Result);
879 // Split the expression into subexprs.
880 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
882 // Remove any common subexpressions.
883 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
884 if (SubExpressionUseData.count(SubExprs[j])) {
885 SubExprs.erase(SubExprs.begin()+j);
889 // Finally, add the non-shared expressions together.
890 if (SubExprs.empty())
893 Uses[i].Base = SE->getAddExpr(SubExprs);
900 /// ValidScale - Check whether the given Scale is valid for all loads and
901 /// stores in UsersToProcess.
903 bool LoopStrengthReduce::ValidScale(bool HasBaseReg, int64_t Scale,
904 const std::vector<BasedUser>& UsersToProcess) {
908 for (unsigned i = 0, e = UsersToProcess.size(); i!=e; ++i) {
909 // If this is a load or other access, pass the type of the access in.
910 const Type *AccessTy =
911 Type::getVoidTy(UsersToProcess[i].Inst->getContext());
912 if (isAddressUse(UsersToProcess[i].Inst,
913 UsersToProcess[i].OperandValToReplace))
914 AccessTy = getAccessType(UsersToProcess[i].Inst);
915 else if (isa<PHINode>(UsersToProcess[i].Inst))
918 TargetLowering::AddrMode AM;
919 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
920 AM.BaseOffs = SC->getValue()->getSExtValue();
921 AM.HasBaseReg = HasBaseReg || !UsersToProcess[i].Base->isZero();
924 // If load[imm+r*scale] is illegal, bail out.
925 if (!TLI->isLegalAddressingMode(AM, AccessTy))
931 /// ValidOffset - Check whether the given Offset is valid for all loads and
932 /// stores in UsersToProcess.
934 bool LoopStrengthReduce::ValidOffset(bool HasBaseReg,
937 const std::vector<BasedUser>& UsersToProcess) {
941 for (unsigned i=0, e = UsersToProcess.size(); i!=e; ++i) {
942 // If this is a load or other access, pass the type of the access in.
943 const Type *AccessTy =
944 Type::getVoidTy(UsersToProcess[i].Inst->getContext());
945 if (isAddressUse(UsersToProcess[i].Inst,
946 UsersToProcess[i].OperandValToReplace))
947 AccessTy = getAccessType(UsersToProcess[i].Inst);
948 else if (isa<PHINode>(UsersToProcess[i].Inst))
951 TargetLowering::AddrMode AM;
952 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
953 AM.BaseOffs = SC->getValue()->getSExtValue();
954 AM.BaseOffs = (uint64_t)AM.BaseOffs + (uint64_t)Offset;
955 AM.HasBaseReg = HasBaseReg || !UsersToProcess[i].Base->isZero();
958 // If load[imm+r*scale] is illegal, bail out.
959 if (!TLI->isLegalAddressingMode(AM, AccessTy))
965 /// RequiresTypeConversion - Returns true if converting Ty1 to Ty2 is not
967 bool LoopStrengthReduce::RequiresTypeConversion(const Type *Ty1,
971 Ty1 = SE->getEffectiveSCEVType(Ty1);
972 Ty2 = SE->getEffectiveSCEVType(Ty2);
975 if (Ty1->canLosslesslyBitCastTo(Ty2))
977 if (TLI && TLI->isTruncateFree(Ty1, Ty2))
982 /// CheckForIVReuse - Returns the multiple if the stride is the multiple
983 /// of a previous stride and it is a legal value for the target addressing
984 /// mode scale component and optional base reg. This allows the users of
985 /// this stride to be rewritten as prev iv * factor. It returns 0 if no
986 /// reuse is possible. Factors can be negative on same targets, e.g. ARM.
988 /// If all uses are outside the loop, we don't require that all multiplies
989 /// be folded into the addressing mode, nor even that the factor be constant;
990 /// a multiply (executed once) outside the loop is better than another IV
991 /// within. Well, usually.
992 const SCEV *LoopStrengthReduce::CheckForIVReuse(bool HasBaseReg,
993 bool AllUsesAreAddresses,
994 bool AllUsesAreOutsideLoop,
995 const SCEV *const &Stride,
996 IVExpr &IV, const Type *Ty,
997 const std::vector<BasedUser>& UsersToProcess) {
998 if (StrideNoReuse.count(Stride))
999 return SE->getIntegerSCEV(0, Stride->getType());
1001 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Stride)) {
1002 int64_t SInt = SC->getValue()->getSExtValue();
1003 for (unsigned NewStride = 0, e = IU->StrideOrder.size();
1004 NewStride != e; ++NewStride) {
1005 std::map<const SCEV *, IVsOfOneStride>::iterator SI =
1006 IVsByStride.find(IU->StrideOrder[NewStride]);
1007 if (SI == IVsByStride.end() || !isa<SCEVConstant>(SI->first) ||
1008 StrideNoReuse.count(SI->first))
1010 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1011 if (SI->first != Stride &&
1012 (unsigned(abs64(SInt)) < SSInt || (SInt % SSInt) != 0))
1014 int64_t Scale = SInt / SSInt;
1015 // Check that this stride is valid for all the types used for loads and
1016 // stores; if it can be used for some and not others, we might as well use
1017 // the original stride everywhere, since we have to create the IV for it
1018 // anyway. If the scale is 1, then we don't need to worry about folding
1021 (AllUsesAreAddresses &&
1022 ValidScale(HasBaseReg, Scale, UsersToProcess))) {
1023 // Prefer to reuse an IV with a base of zero.
1024 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1025 IE = SI->second.IVs.end(); II != IE; ++II)
1026 // Only reuse previous IV if it would not require a type conversion
1027 // and if the base difference can be folded.
1028 if (II->Base->isZero() &&
1029 !RequiresTypeConversion(II->Base->getType(), Ty)) {
1031 return SE->getIntegerSCEV(Scale, Stride->getType());
1033 // Otherwise, settle for an IV with a foldable base.
1034 if (AllUsesAreAddresses)
1035 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1036 IE = SI->second.IVs.end(); II != IE; ++II)
1037 // Only reuse previous IV if it would not require a type conversion
1038 // and if the base difference can be folded.
1039 if (SE->getEffectiveSCEVType(II->Base->getType()) ==
1040 SE->getEffectiveSCEVType(Ty) &&
1041 isa<SCEVConstant>(II->Base)) {
1043 cast<SCEVConstant>(II->Base)->getValue()->getSExtValue();
1044 if (Base > INT32_MIN && Base <= INT32_MAX &&
1045 ValidOffset(HasBaseReg, -Base * Scale,
1046 Scale, UsersToProcess)) {
1048 return SE->getIntegerSCEV(Scale, Stride->getType());
1053 } else if (AllUsesAreOutsideLoop) {
1054 // Accept nonconstant strides here; it is really really right to substitute
1055 // an existing IV if we can.
1056 for (unsigned NewStride = 0, e = IU->StrideOrder.size();
1057 NewStride != e; ++NewStride) {
1058 std::map<const SCEV *, IVsOfOneStride>::iterator SI =
1059 IVsByStride.find(IU->StrideOrder[NewStride]);
1060 if (SI == IVsByStride.end() || !isa<SCEVConstant>(SI->first))
1062 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1063 if (SI->first != Stride && SSInt != 1)
1065 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1066 IE = SI->second.IVs.end(); II != IE; ++II)
1067 // Accept nonzero base here.
1068 // Only reuse previous IV if it would not require a type conversion.
1069 if (!RequiresTypeConversion(II->Base->getType(), Ty)) {
1074 // Special case, old IV is -1*x and this one is x. Can treat this one as
1076 for (unsigned NewStride = 0, e = IU->StrideOrder.size();
1077 NewStride != e; ++NewStride) {
1078 std::map<const SCEV *, IVsOfOneStride>::iterator SI =
1079 IVsByStride.find(IU->StrideOrder[NewStride]);
1080 if (SI == IVsByStride.end())
1082 if (const SCEVMulExpr *ME = dyn_cast<SCEVMulExpr>(SI->first))
1083 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(ME->getOperand(0)))
1084 if (Stride == ME->getOperand(1) &&
1085 SC->getValue()->getSExtValue() == -1LL)
1086 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1087 IE = SI->second.IVs.end(); II != IE; ++II)
1088 // Accept nonzero base here.
1089 // Only reuse previous IV if it would not require type conversion.
1090 if (!RequiresTypeConversion(II->Base->getType(), Ty)) {
1092 return SE->getIntegerSCEV(-1LL, Stride->getType());
1096 return SE->getIntegerSCEV(0, Stride->getType());
1099 /// PartitionByIsUseOfPostIncrementedValue - Simple boolean predicate that
1100 /// returns true if Val's isUseOfPostIncrementedValue is true.
1101 static bool PartitionByIsUseOfPostIncrementedValue(const BasedUser &Val) {
1102 return Val.isUseOfPostIncrementedValue;
1105 /// isNonConstantNegative - Return true if the specified scev is negated, but
1107 static bool isNonConstantNegative(const SCEV *const &Expr) {
1108 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Expr);
1109 if (!Mul) return false;
1111 // If there is a constant factor, it will be first.
1112 const SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
1113 if (!SC) return false;
1115 // Return true if the value is negative, this matches things like (-42 * V).
1116 return SC->getValue()->getValue().isNegative();
1119 /// CollectIVUsers - Transform our list of users and offsets to a bit more
1120 /// complex table. In this new vector, each 'BasedUser' contains 'Base', the base
1121 /// of the strided accesses, as well as the old information from Uses. We
1122 /// progressively move information from the Base field to the Imm field, until
1123 /// we eventually have the full access expression to rewrite the use.
1124 const SCEV *LoopStrengthReduce::CollectIVUsers(const SCEV *const &Stride,
1125 IVUsersOfOneStride &Uses,
1127 bool &AllUsesAreAddresses,
1128 bool &AllUsesAreOutsideLoop,
1129 std::vector<BasedUser> &UsersToProcess) {
1130 // FIXME: Generalize to non-affine IV's.
1131 if (!Stride->isLoopInvariant(L))
1132 return SE->getIntegerSCEV(0, Stride->getType());
1134 UsersToProcess.reserve(Uses.Users.size());
1135 for (ilist<IVStrideUse>::iterator I = Uses.Users.begin(),
1136 E = Uses.Users.end(); I != E; ++I) {
1137 UsersToProcess.push_back(BasedUser(*I, SE));
1139 // Move any loop variant operands from the offset field to the immediate
1140 // field of the use, so that we don't try to use something before it is
1142 MoveLoopVariantsToImmediateField(UsersToProcess.back().Base,
1143 UsersToProcess.back().Imm, L, SE);
1144 assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
1145 "Base value is not loop invariant!");
1148 // We now have a whole bunch of uses of like-strided induction variables, but
1149 // they might all have different bases. We want to emit one PHI node for this
1150 // stride which we fold as many common expressions (between the IVs) into as
1151 // possible. Start by identifying the common expressions in the base values
1152 // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
1153 // "A+B"), emit it to the preheader, then remove the expression from the
1154 // UsersToProcess base values.
1155 const SCEV *CommonExprs =
1156 RemoveCommonExpressionsFromUseBases(UsersToProcess, SE, L, TLI);
1158 // Next, figure out what we can represent in the immediate fields of
1159 // instructions. If we can represent anything there, move it to the imm
1160 // fields of the BasedUsers. We do this so that it increases the commonality
1161 // of the remaining uses.
1162 unsigned NumPHI = 0;
1163 bool HasAddress = false;
1164 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1165 // If the user is not in the current loop, this means it is using the exit
1166 // value of the IV. Do not put anything in the base, make sure it's all in
1167 // the immediate field to allow as much factoring as possible.
1168 if (!L->contains(UsersToProcess[i].Inst->getParent())) {
1169 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm,
1170 UsersToProcess[i].Base);
1171 UsersToProcess[i].Base =
1172 SE->getIntegerSCEV(0, UsersToProcess[i].Base->getType());
1174 // Not all uses are outside the loop.
1175 AllUsesAreOutsideLoop = false;
1177 // Addressing modes can be folded into loads and stores. Be careful that
1178 // the store is through the expression, not of the expression though.
1180 bool isAddress = isAddressUse(UsersToProcess[i].Inst,
1181 UsersToProcess[i].OperandValToReplace);
1182 if (isa<PHINode>(UsersToProcess[i].Inst)) {
1190 // If this use isn't an address, then not all uses are addresses.
1191 if (!isAddress && !isPHI)
1192 AllUsesAreAddresses = false;
1194 MoveImmediateValues(TLI, UsersToProcess[i].Inst, UsersToProcess[i].Base,
1195 UsersToProcess[i].Imm, isAddress, L, SE);
1199 // If one of the use is a PHI node and all other uses are addresses, still
1200 // allow iv reuse. Essentially we are trading one constant multiplication
1201 // for one fewer iv.
1203 AllUsesAreAddresses = false;
1205 // There are no in-loop address uses.
1206 if (AllUsesAreAddresses && (!HasAddress && !AllUsesAreOutsideLoop))
1207 AllUsesAreAddresses = false;
1212 /// ShouldUseFullStrengthReductionMode - Test whether full strength-reduction
1213 /// is valid and profitable for the given set of users of a stride. In
1214 /// full strength-reduction mode, all addresses at the current stride are
1215 /// strength-reduced all the way down to pointer arithmetic.
1217 bool LoopStrengthReduce::ShouldUseFullStrengthReductionMode(
1218 const std::vector<BasedUser> &UsersToProcess,
1220 bool AllUsesAreAddresses,
1221 const SCEV *Stride) {
1222 if (!EnableFullLSRMode)
1225 // The heuristics below aim to avoid increasing register pressure, but
1226 // fully strength-reducing all the addresses increases the number of
1227 // add instructions, so don't do this when optimizing for size.
1228 // TODO: If the loop is large, the savings due to simpler addresses
1229 // may oughtweight the costs of the extra increment instructions.
1230 if (L->getHeader()->getParent()->hasFnAttr(Attribute::OptimizeForSize))
1233 // TODO: For now, don't do full strength reduction if there could
1234 // potentially be greater-stride multiples of the current stride
1235 // which could reuse the current stride IV.
1236 if (IU->StrideOrder.back() != Stride)
1239 // Iterate through the uses to find conditions that automatically rule out
1241 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ) {
1242 const SCEV *Base = UsersToProcess[i].Base;
1243 const SCEV *Imm = UsersToProcess[i].Imm;
1244 // If any users have a loop-variant component, they can't be fully
1245 // strength-reduced.
1246 if (Imm && !Imm->isLoopInvariant(L))
1248 // If there are to users with the same base and the difference between
1249 // the two Imm values can't be folded into the address, full
1250 // strength reduction would increase register pressure.
1252 const SCEV *CurImm = UsersToProcess[i].Imm;
1253 if ((CurImm || Imm) && CurImm != Imm) {
1254 if (!CurImm) CurImm = SE->getIntegerSCEV(0, Stride->getType());
1255 if (!Imm) Imm = SE->getIntegerSCEV(0, Stride->getType());
1256 const Instruction *Inst = UsersToProcess[i].Inst;
1257 const Type *AccessTy = getAccessType(Inst);
1258 const SCEV *Diff = SE->getMinusSCEV(UsersToProcess[i].Imm, Imm);
1259 if (!Diff->isZero() &&
1260 (!AllUsesAreAddresses ||
1261 !fitsInAddressMode(Diff, AccessTy, TLI, /*HasBaseReg=*/true)))
1264 } while (++i != e && Base == UsersToProcess[i].Base);
1267 // If there's exactly one user in this stride, fully strength-reducing it
1268 // won't increase register pressure. If it's starting from a non-zero base,
1269 // it'll be simpler this way.
1270 if (UsersToProcess.size() == 1 && !UsersToProcess[0].Base->isZero())
1273 // Otherwise, if there are any users in this stride that don't require
1274 // a register for their base, full strength-reduction will increase
1275 // register pressure.
1276 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1277 if (UsersToProcess[i].Base->isZero())
1280 // Otherwise, go for it.
1284 /// InsertAffinePhi Create and insert a PHI node for an induction variable
1285 /// with the specified start and step values in the specified loop.
1287 /// If NegateStride is true, the stride should be negated by using a
1288 /// subtract instead of an add.
1290 /// Return the created phi node.
1292 static PHINode *InsertAffinePhi(const SCEV *Start, const SCEV *Step,
1293 Instruction *IVIncInsertPt,
1295 SCEVExpander &Rewriter) {
1296 assert(Start->isLoopInvariant(L) && "New PHI start is not loop invariant!");
1297 assert(Step->isLoopInvariant(L) && "New PHI stride is not loop invariant!");
1299 BasicBlock *Header = L->getHeader();
1300 BasicBlock *Preheader = L->getLoopPreheader();
1301 BasicBlock *LatchBlock = L->getLoopLatch();
1302 const Type *Ty = Start->getType();
1303 Ty = Rewriter.SE.getEffectiveSCEVType(Ty);
1305 PHINode *PN = PHINode::Create(Ty, "lsr.iv", Header->begin());
1306 PN->addIncoming(Rewriter.expandCodeFor(Start, Ty, Preheader->getTerminator()),
1309 // If the stride is negative, insert a sub instead of an add for the
1311 bool isNegative = isNonConstantNegative(Step);
1312 const SCEV *IncAmount = Step;
1314 IncAmount = Rewriter.SE.getNegativeSCEV(Step);
1316 // Insert an add instruction right before the terminator corresponding
1317 // to the back-edge or just before the only use. The location is determined
1318 // by the caller and passed in as IVIncInsertPt.
1319 Value *StepV = Rewriter.expandCodeFor(IncAmount, Ty,
1320 Preheader->getTerminator());
1323 IncV = BinaryOperator::CreateSub(PN, StepV, "lsr.iv.next",
1326 IncV = BinaryOperator::CreateAdd(PN, StepV, "lsr.iv.next",
1329 if (!isa<ConstantInt>(StepV)) ++NumVariable;
1331 PN->addIncoming(IncV, LatchBlock);
1337 static void SortUsersToProcess(std::vector<BasedUser> &UsersToProcess) {
1338 // We want to emit code for users inside the loop first. To do this, we
1339 // rearrange BasedUser so that the entries at the end have
1340 // isUseOfPostIncrementedValue = false, because we pop off the end of the
1341 // vector (so we handle them first).
1342 std::partition(UsersToProcess.begin(), UsersToProcess.end(),
1343 PartitionByIsUseOfPostIncrementedValue);
1345 // Sort this by base, so that things with the same base are handled
1346 // together. By partitioning first and stable-sorting later, we are
1347 // guaranteed that within each base we will pop off users from within the
1348 // loop before users outside of the loop with a particular base.
1350 // We would like to use stable_sort here, but we can't. The problem is that
1351 // const SCEV *'s don't have a deterministic ordering w.r.t to each other, so
1352 // we don't have anything to do a '<' comparison on. Because we think the
1353 // number of uses is small, do a horrible bubble sort which just relies on
1355 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1356 // Get a base value.
1357 const SCEV *Base = UsersToProcess[i].Base;
1359 // Compact everything with this base to be consecutive with this one.
1360 for (unsigned j = i+1; j != e; ++j) {
1361 if (UsersToProcess[j].Base == Base) {
1362 std::swap(UsersToProcess[i+1], UsersToProcess[j]);
1369 /// PrepareToStrengthReduceFully - Prepare to fully strength-reduce
1370 /// UsersToProcess, meaning lowering addresses all the way down to direct
1371 /// pointer arithmetic.
1374 LoopStrengthReduce::PrepareToStrengthReduceFully(
1375 std::vector<BasedUser> &UsersToProcess,
1377 const SCEV *CommonExprs,
1379 SCEVExpander &PreheaderRewriter) {
1380 DEBUG(errs() << " Fully reducing all users\n");
1382 // Rewrite the UsersToProcess records, creating a separate PHI for each
1383 // unique Base value.
1384 Instruction *IVIncInsertPt = L->getLoopLatch()->getTerminator();
1385 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ) {
1386 // TODO: The uses are grouped by base, but not sorted. We arbitrarily
1387 // pick the first Imm value here to start with, and adjust it for the
1389 const SCEV *Imm = UsersToProcess[i].Imm;
1390 const SCEV *Base = UsersToProcess[i].Base;
1391 const SCEV *Start = SE->getAddExpr(CommonExprs, Base, Imm);
1392 PHINode *Phi = InsertAffinePhi(Start, Stride, IVIncInsertPt, L,
1394 // Loop over all the users with the same base.
1396 UsersToProcess[i].Base = SE->getIntegerSCEV(0, Stride->getType());
1397 UsersToProcess[i].Imm = SE->getMinusSCEV(UsersToProcess[i].Imm, Imm);
1398 UsersToProcess[i].Phi = Phi;
1399 assert(UsersToProcess[i].Imm->isLoopInvariant(L) &&
1400 "ShouldUseFullStrengthReductionMode should reject this!");
1401 } while (++i != e && Base == UsersToProcess[i].Base);
1405 /// FindIVIncInsertPt - Return the location to insert the increment instruction.
1406 /// If the only use if a use of postinc value, (must be the loop termination
1407 /// condition), then insert it just before the use.
1408 static Instruction *FindIVIncInsertPt(std::vector<BasedUser> &UsersToProcess,
1410 if (UsersToProcess.size() == 1 &&
1411 UsersToProcess[0].isUseOfPostIncrementedValue &&
1412 L->contains(UsersToProcess[0].Inst->getParent()))
1413 return UsersToProcess[0].Inst;
1414 return L->getLoopLatch()->getTerminator();
1417 /// PrepareToStrengthReduceWithNewPhi - Insert a new induction variable for the
1418 /// given users to share.
1421 LoopStrengthReduce::PrepareToStrengthReduceWithNewPhi(
1422 std::vector<BasedUser> &UsersToProcess,
1424 const SCEV *CommonExprs,
1426 Instruction *IVIncInsertPt,
1428 SCEVExpander &PreheaderRewriter) {
1429 DEBUG(errs() << " Inserting new PHI:\n");
1431 PHINode *Phi = InsertAffinePhi(SE->getUnknown(CommonBaseV),
1432 Stride, IVIncInsertPt, L,
1435 // Remember this in case a later stride is multiple of this.
1436 IVsByStride[Stride].addIV(Stride, CommonExprs, Phi);
1438 // All the users will share this new IV.
1439 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1440 UsersToProcess[i].Phi = Phi;
1442 DEBUG(errs() << " IV=");
1443 DEBUG(WriteAsOperand(errs(), Phi, /*PrintType=*/false));
1444 DEBUG(errs() << "\n");
1447 /// PrepareToStrengthReduceFromSmallerStride - Prepare for the given users to
1448 /// reuse an induction variable with a stride that is a factor of the current
1449 /// induction variable.
1452 LoopStrengthReduce::PrepareToStrengthReduceFromSmallerStride(
1453 std::vector<BasedUser> &UsersToProcess,
1455 const IVExpr &ReuseIV,
1456 Instruction *PreInsertPt) {
1457 DEBUG(errs() << " Rewriting in terms of existing IV of STRIDE "
1458 << *ReuseIV.Stride << " and BASE " << *ReuseIV.Base << "\n");
1460 // All the users will share the reused IV.
1461 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1462 UsersToProcess[i].Phi = ReuseIV.PHI;
1464 Constant *C = dyn_cast<Constant>(CommonBaseV);
1466 (!C->isNullValue() &&
1467 !fitsInAddressMode(SE->getUnknown(CommonBaseV), CommonBaseV->getType(),
1469 // We want the common base emitted into the preheader! This is just
1470 // using cast as a copy so BitCast (no-op cast) is appropriate
1471 CommonBaseV = new BitCastInst(CommonBaseV, CommonBaseV->getType(),
1472 "commonbase", PreInsertPt);
1475 static bool IsImmFoldedIntoAddrMode(GlobalValue *GV, int64_t Offset,
1476 const Type *AccessTy,
1477 std::vector<BasedUser> &UsersToProcess,
1478 const TargetLowering *TLI) {
1479 SmallVector<Instruction*, 16> AddrModeInsts;
1480 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1481 if (UsersToProcess[i].isUseOfPostIncrementedValue)
1483 ExtAddrMode AddrMode =
1484 AddressingModeMatcher::Match(UsersToProcess[i].OperandValToReplace,
1485 AccessTy, UsersToProcess[i].Inst,
1486 AddrModeInsts, *TLI);
1487 if (GV && GV != AddrMode.BaseGV)
1489 if (Offset && !AddrMode.BaseOffs)
1490 // FIXME: How to accurate check it's immediate offset is folded.
1492 AddrModeInsts.clear();
1497 /// StrengthReduceStridedIVUsers - Strength reduce all of the users of a single
1498 /// stride of IV. All of the users may have different starting values, and this
1499 /// may not be the only stride.
1500 void LoopStrengthReduce::StrengthReduceStridedIVUsers(const SCEV *const &Stride,
1501 IVUsersOfOneStride &Uses,
1503 // If all the users are moved to another stride, then there is nothing to do.
1504 if (Uses.Users.empty())
1507 // Keep track if every use in UsersToProcess is an address. If they all are,
1508 // we may be able to rewrite the entire collection of them in terms of a
1509 // smaller-stride IV.
1510 bool AllUsesAreAddresses = true;
1512 // Keep track if every use of a single stride is outside the loop. If so,
1513 // we want to be more aggressive about reusing a smaller-stride IV; a
1514 // multiply outside the loop is better than another IV inside. Well, usually.
1515 bool AllUsesAreOutsideLoop = true;
1517 // Transform our list of users and offsets to a bit more complex table. In
1518 // this new vector, each 'BasedUser' contains 'Base' the base of the
1519 // strided accessas well as the old information from Uses. We progressively
1520 // move information from the Base field to the Imm field, until we eventually
1521 // have the full access expression to rewrite the use.
1522 std::vector<BasedUser> UsersToProcess;
1523 const SCEV *CommonExprs = CollectIVUsers(Stride, Uses, L, AllUsesAreAddresses,
1524 AllUsesAreOutsideLoop,
1527 // Sort the UsersToProcess array so that users with common bases are
1528 // next to each other.
1529 SortUsersToProcess(UsersToProcess);
1531 // If we managed to find some expressions in common, we'll need to carry
1532 // their value in a register and add it in for each use. This will take up
1533 // a register operand, which potentially restricts what stride values are
1535 bool HaveCommonExprs = !CommonExprs->isZero();
1536 const Type *ReplacedTy = CommonExprs->getType();
1538 // If all uses are addresses, consider sinking the immediate part of the
1539 // common expression back into uses if they can fit in the immediate fields.
1540 if (TLI && HaveCommonExprs && AllUsesAreAddresses) {
1541 const SCEV *NewCommon = CommonExprs;
1542 const SCEV *Imm = SE->getIntegerSCEV(0, ReplacedTy);
1543 MoveImmediateValues(TLI, Type::getVoidTy(
1544 L->getLoopPreheader()->getContext()),
1545 NewCommon, Imm, true, L, SE);
1546 if (!Imm->isZero()) {
1549 // If the immediate part of the common expression is a GV, check if it's
1550 // possible to fold it into the target addressing mode.
1551 GlobalValue *GV = 0;
1552 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(Imm))
1553 GV = dyn_cast<GlobalValue>(SU->getValue());
1555 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Imm))
1556 Offset = SC->getValue()->getSExtValue();
1558 // Pass VoidTy as the AccessTy to be conservative, because
1559 // there could be multiple access types among all the uses.
1560 DoSink = IsImmFoldedIntoAddrMode(GV, Offset,
1561 Type::getVoidTy(L->getLoopPreheader()->getContext()),
1562 UsersToProcess, TLI);
1565 DEBUG(errs() << " Sinking " << *Imm << " back down into uses\n");
1566 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1567 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm, Imm);
1568 CommonExprs = NewCommon;
1569 HaveCommonExprs = !CommonExprs->isZero();
1575 // Now that we know what we need to do, insert the PHI node itself.
1577 DEBUG(errs() << "LSR: Examining IVs of TYPE " << *ReplacedTy << " of STRIDE "
1579 << " Common base: " << *CommonExprs << "\n");
1581 SCEVExpander Rewriter(*SE);
1582 SCEVExpander PreheaderRewriter(*SE);
1584 BasicBlock *Preheader = L->getLoopPreheader();
1585 Instruction *PreInsertPt = Preheader->getTerminator();
1586 BasicBlock *LatchBlock = L->getLoopLatch();
1587 Instruction *IVIncInsertPt = LatchBlock->getTerminator();
1589 Value *CommonBaseV = Constant::getNullValue(ReplacedTy);
1591 const SCEV *RewriteFactor = SE->getIntegerSCEV(0, ReplacedTy);
1592 IVExpr ReuseIV(SE->getIntegerSCEV(0,
1593 Type::getInt32Ty(Preheader->getContext())),
1594 SE->getIntegerSCEV(0,
1595 Type::getInt32Ty(Preheader->getContext())),
1598 // Choose a strength-reduction strategy and prepare for it by creating
1599 // the necessary PHIs and adjusting the bookkeeping.
1600 if (ShouldUseFullStrengthReductionMode(UsersToProcess, L,
1601 AllUsesAreAddresses, Stride)) {
1602 PrepareToStrengthReduceFully(UsersToProcess, Stride, CommonExprs, L,
1605 // Emit the initial base value into the loop preheader.
1606 CommonBaseV = PreheaderRewriter.expandCodeFor(CommonExprs, ReplacedTy,
1609 // If all uses are addresses, check if it is possible to reuse an IV. The
1610 // new IV must have a stride that is a multiple of the old stride; the
1611 // multiple must be a number that can be encoded in the scale field of the
1612 // target addressing mode; and we must have a valid instruction after this
1613 // substitution, including the immediate field, if any.
1614 RewriteFactor = CheckForIVReuse(HaveCommonExprs, AllUsesAreAddresses,
1615 AllUsesAreOutsideLoop,
1616 Stride, ReuseIV, ReplacedTy,
1618 if (!RewriteFactor->isZero())
1619 PrepareToStrengthReduceFromSmallerStride(UsersToProcess, CommonBaseV,
1620 ReuseIV, PreInsertPt);
1622 IVIncInsertPt = FindIVIncInsertPt(UsersToProcess, L);
1623 PrepareToStrengthReduceWithNewPhi(UsersToProcess, Stride, CommonExprs,
1624 CommonBaseV, IVIncInsertPt,
1625 L, PreheaderRewriter);
1629 // Process all the users now, replacing their strided uses with
1630 // strength-reduced forms. This outer loop handles all bases, the inner
1631 // loop handles all users of a particular base.
1632 while (!UsersToProcess.empty()) {
1633 const SCEV *Base = UsersToProcess.back().Base;
1634 Instruction *Inst = UsersToProcess.back().Inst;
1636 // Emit the code for Base into the preheader.
1638 if (!Base->isZero()) {
1639 BaseV = PreheaderRewriter.expandCodeFor(Base, 0, PreInsertPt);
1641 DEBUG(errs() << " INSERTING code for BASE = " << *Base << ":");
1642 if (BaseV->hasName())
1643 DEBUG(errs() << " Result value name = %" << BaseV->getName());
1644 DEBUG(errs() << "\n");
1646 // If BaseV is a non-zero constant, make sure that it gets inserted into
1647 // the preheader, instead of being forward substituted into the uses. We
1648 // do this by forcing a BitCast (noop cast) to be inserted into the
1649 // preheader in this case.
1650 if (!fitsInAddressMode(Base, getAccessType(Inst), TLI, false) &&
1651 isa<Constant>(BaseV)) {
1652 // We want this constant emitted into the preheader! This is just
1653 // using cast as a copy so BitCast (no-op cast) is appropriate
1654 BaseV = new BitCastInst(BaseV, BaseV->getType(), "preheaderinsert",
1659 // Emit the code to add the immediate offset to the Phi value, just before
1660 // the instructions that we identified as using this stride and base.
1662 // FIXME: Use emitted users to emit other users.
1663 BasedUser &User = UsersToProcess.back();
1665 DEBUG(errs() << " Examining ");
1666 if (User.isUseOfPostIncrementedValue)
1667 DEBUG(errs() << "postinc");
1669 DEBUG(errs() << "preinc");
1670 DEBUG(errs() << " use ");
1671 DEBUG(WriteAsOperand(errs(), UsersToProcess.back().OperandValToReplace,
1672 /*PrintType=*/false));
1673 DEBUG(errs() << " in Inst: " << *User.Inst);
1675 // If this instruction wants to use the post-incremented value, move it
1676 // after the post-inc and use its value instead of the PHI.
1677 Value *RewriteOp = User.Phi;
1678 if (User.isUseOfPostIncrementedValue) {
1679 RewriteOp = User.Phi->getIncomingValueForBlock(LatchBlock);
1680 // If this user is in the loop, make sure it is the last thing in the
1681 // loop to ensure it is dominated by the increment. In case it's the
1682 // only use of the iv, the increment instruction is already before the
1684 if (L->contains(User.Inst->getParent()) && User.Inst != IVIncInsertPt)
1685 User.Inst->moveBefore(IVIncInsertPt);
1688 const SCEV *RewriteExpr = SE->getUnknown(RewriteOp);
1690 if (SE->getEffectiveSCEVType(RewriteOp->getType()) !=
1691 SE->getEffectiveSCEVType(ReplacedTy)) {
1692 assert(SE->getTypeSizeInBits(RewriteOp->getType()) >
1693 SE->getTypeSizeInBits(ReplacedTy) &&
1694 "Unexpected widening cast!");
1695 RewriteExpr = SE->getTruncateExpr(RewriteExpr, ReplacedTy);
1698 // If we had to insert new instructions for RewriteOp, we have to
1699 // consider that they may not have been able to end up immediately
1700 // next to RewriteOp, because non-PHI instructions may never precede
1701 // PHI instructions in a block. In this case, remember where the last
1702 // instruction was inserted so that if we're replacing a different
1703 // PHI node, we can use the later point to expand the final
1705 Instruction *NewBasePt = dyn_cast<Instruction>(RewriteOp);
1706 if (RewriteOp == User.Phi) NewBasePt = 0;
1708 // Clear the SCEVExpander's expression map so that we are guaranteed
1709 // to have the code emitted where we expect it.
1712 // If we are reusing the iv, then it must be multiplied by a constant
1713 // factor to take advantage of the addressing mode scale component.
1714 if (!RewriteFactor->isZero()) {
1715 // If we're reusing an IV with a nonzero base (currently this happens
1716 // only when all reuses are outside the loop) subtract that base here.
1717 // The base has been used to initialize the PHI node but we don't want
1719 if (!ReuseIV.Base->isZero()) {
1720 const SCEV *typedBase = ReuseIV.Base;
1721 if (SE->getEffectiveSCEVType(RewriteExpr->getType()) !=
1722 SE->getEffectiveSCEVType(ReuseIV.Base->getType())) {
1723 // It's possible the original IV is a larger type than the new IV,
1724 // in which case we have to truncate the Base. We checked in
1725 // RequiresTypeConversion that this is valid.
1726 assert(SE->getTypeSizeInBits(RewriteExpr->getType()) <
1727 SE->getTypeSizeInBits(ReuseIV.Base->getType()) &&
1728 "Unexpected lengthening conversion!");
1729 typedBase = SE->getTruncateExpr(ReuseIV.Base,
1730 RewriteExpr->getType());
1732 RewriteExpr = SE->getMinusSCEV(RewriteExpr, typedBase);
1735 // Multiply old variable, with base removed, by new scale factor.
1736 RewriteExpr = SE->getMulExpr(RewriteFactor,
1739 // The common base is emitted in the loop preheader. But since we
1740 // are reusing an IV, it has not been used to initialize the PHI node.
1741 // Add it to the expression used to rewrite the uses.
1742 // When this use is outside the loop, we earlier subtracted the
1743 // common base, and are adding it back here. Use the same expression
1744 // as before, rather than CommonBaseV, so DAGCombiner will zap it.
1745 if (!CommonExprs->isZero()) {
1746 if (L->contains(User.Inst->getParent()))
1747 RewriteExpr = SE->getAddExpr(RewriteExpr,
1748 SE->getUnknown(CommonBaseV));
1750 RewriteExpr = SE->getAddExpr(RewriteExpr, CommonExprs);
1754 // Now that we know what we need to do, insert code before User for the
1755 // immediate and any loop-variant expressions.
1757 // Add BaseV to the PHI value if needed.
1758 RewriteExpr = SE->getAddExpr(RewriteExpr, SE->getUnknown(BaseV));
1760 User.RewriteInstructionToUseNewBase(RewriteExpr, NewBasePt,
1761 Rewriter, L, this, *LI,
1764 // Mark old value we replaced as possibly dead, so that it is eliminated
1765 // if we just replaced the last use of that value.
1766 DeadInsts.push_back(User.OperandValToReplace);
1768 UsersToProcess.pop_back();
1771 // If there are any more users to process with the same base, process them
1772 // now. We sorted by base above, so we just have to check the last elt.
1773 } while (!UsersToProcess.empty() && UsersToProcess.back().Base == Base);
1774 // TODO: Next, find out which base index is the most common, pull it out.
1777 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
1778 // different starting values, into different PHIs.
1781 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1782 /// set the IV user and stride information and return true, otherwise return
1784 bool LoopStrengthReduce::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse,
1785 const SCEV *const * &CondStride) {
1786 for (unsigned Stride = 0, e = IU->StrideOrder.size();
1787 Stride != e && !CondUse; ++Stride) {
1788 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
1789 IU->IVUsesByStride.find(IU->StrideOrder[Stride]);
1790 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
1792 for (ilist<IVStrideUse>::iterator UI = SI->second->Users.begin(),
1793 E = SI->second->Users.end(); UI != E; ++UI)
1794 if (UI->getUser() == Cond) {
1795 // NOTE: we could handle setcc instructions with multiple uses here, but
1796 // InstCombine does it as well for simple uses, it's not clear that it
1797 // occurs enough in real life to handle.
1799 CondStride = &SI->first;
1807 // Constant strides come first which in turns are sorted by their absolute
1808 // values. If absolute values are the same, then positive strides comes first.
1810 // 4, -1, X, 1, 2 ==> 1, -1, 2, 4, X
1811 struct StrideCompare {
1812 const ScalarEvolution *SE;
1813 explicit StrideCompare(const ScalarEvolution *se) : SE(se) {}
1815 bool operator()(const SCEV *const &LHS, const SCEV *const &RHS) {
1816 const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS);
1817 const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS);
1819 int64_t LV = LHSC->getValue()->getSExtValue();
1820 int64_t RV = RHSC->getValue()->getSExtValue();
1821 uint64_t ALV = (LV < 0) ? -LV : LV;
1822 uint64_t ARV = (RV < 0) ? -RV : RV;
1830 // If it's the same value but different type, sort by bit width so
1831 // that we emit larger induction variables before smaller
1832 // ones, letting the smaller be re-written in terms of larger ones.
1833 return SE->getTypeSizeInBits(RHS->getType()) <
1834 SE->getTypeSizeInBits(LHS->getType());
1836 return LHSC && !RHSC;
1841 /// ChangeCompareStride - If a loop termination compare instruction is the
1842 /// only use of its stride, and the compaison is against a constant value,
1843 /// try eliminate the stride by moving the compare instruction to another
1844 /// stride and change its constant operand accordingly. e.g.
1850 /// if (v2 < 10) goto loop
1855 /// if (v1 < 30) goto loop
1856 ICmpInst *LoopStrengthReduce::ChangeCompareStride(Loop *L, ICmpInst *Cond,
1857 IVStrideUse* &CondUse,
1858 const SCEV *const* &CondStride) {
1859 // If there's only one stride in the loop, there's nothing to do here.
1860 if (IU->StrideOrder.size() < 2)
1862 // If there are other users of the condition's stride, don't bother
1863 // trying to change the condition because the stride will still
1865 std::map<const SCEV *, IVUsersOfOneStride *>::iterator I =
1866 IU->IVUsesByStride.find(*CondStride);
1867 if (I == IU->IVUsesByStride.end() ||
1868 I->second->Users.size() != 1)
1870 // Only handle constant strides for now.
1871 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*CondStride);
1872 if (!SC) return Cond;
1874 ICmpInst::Predicate Predicate = Cond->getPredicate();
1875 int64_t CmpSSInt = SC->getValue()->getSExtValue();
1876 unsigned BitWidth = SE->getTypeSizeInBits((*CondStride)->getType());
1877 uint64_t SignBit = 1ULL << (BitWidth-1);
1878 const Type *CmpTy = Cond->getOperand(0)->getType();
1879 const Type *NewCmpTy = NULL;
1880 unsigned TyBits = SE->getTypeSizeInBits(CmpTy);
1881 unsigned NewTyBits = 0;
1882 const SCEV **NewStride = NULL;
1883 Value *NewCmpLHS = NULL;
1884 Value *NewCmpRHS = NULL;
1886 const SCEV *NewOffset = SE->getIntegerSCEV(0, CmpTy);
1888 if (ConstantInt *C = dyn_cast<ConstantInt>(Cond->getOperand(1))) {
1889 int64_t CmpVal = C->getValue().getSExtValue();
1891 // Check stride constant and the comparision constant signs to detect
1893 if ((CmpVal & SignBit) != (CmpSSInt & SignBit))
1896 // Look for a suitable stride / iv as replacement.
1897 for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
1898 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
1899 IU->IVUsesByStride.find(IU->StrideOrder[i]);
1900 if (!isa<SCEVConstant>(SI->first))
1902 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1903 if (SSInt == CmpSSInt ||
1904 abs64(SSInt) < abs64(CmpSSInt) ||
1905 (SSInt % CmpSSInt) != 0)
1908 Scale = SSInt / CmpSSInt;
1909 int64_t NewCmpVal = CmpVal * Scale;
1910 APInt Mul = APInt(BitWidth*2, CmpVal, true);
1911 Mul = Mul * APInt(BitWidth*2, Scale, true);
1912 // Check for overflow.
1913 if (!Mul.isSignedIntN(BitWidth))
1915 // Check for overflow in the stride's type too.
1916 if (!Mul.isSignedIntN(SE->getTypeSizeInBits(SI->first->getType())))
1919 // Watch out for overflow.
1920 if (ICmpInst::isSigned(Predicate) &&
1921 (CmpVal & SignBit) != (NewCmpVal & SignBit))
1924 if (NewCmpVal == CmpVal)
1926 // Pick the best iv to use trying to avoid a cast.
1928 for (ilist<IVStrideUse>::iterator UI = SI->second->Users.begin(),
1929 E = SI->second->Users.end(); UI != E; ++UI) {
1930 Value *Op = UI->getOperandValToReplace();
1932 // If the IVStrideUse implies a cast, check for an actual cast which
1933 // can be used to find the original IV expression.
1934 if (SE->getEffectiveSCEVType(Op->getType()) !=
1935 SE->getEffectiveSCEVType(SI->first->getType())) {
1936 CastInst *CI = dyn_cast<CastInst>(Op);
1937 // If it's not a simple cast, it's complicated.
1940 // If it's a cast from a type other than the stride type,
1941 // it's complicated.
1942 if (CI->getOperand(0)->getType() != SI->first->getType())
1944 // Ok, we found the IV expression in the stride's type.
1945 Op = CI->getOperand(0);
1949 if (NewCmpLHS->getType() == CmpTy)
1955 NewCmpTy = NewCmpLHS->getType();
1956 NewTyBits = SE->getTypeSizeInBits(NewCmpTy);
1957 const Type *NewCmpIntTy = IntegerType::get(Cond->getContext(), NewTyBits);
1958 if (RequiresTypeConversion(NewCmpTy, CmpTy)) {
1959 // Check if it is possible to rewrite it using
1960 // an iv / stride of a smaller integer type.
1961 unsigned Bits = NewTyBits;
1962 if (ICmpInst::isSigned(Predicate))
1964 uint64_t Mask = (1ULL << Bits) - 1;
1965 if (((uint64_t)NewCmpVal & Mask) != (uint64_t)NewCmpVal)
1969 // Don't rewrite if use offset is non-constant and the new type is
1970 // of a different type.
1971 // FIXME: too conservative?
1972 if (NewTyBits != TyBits && !isa<SCEVConstant>(CondUse->getOffset()))
1975 bool AllUsesAreAddresses = true;
1976 bool AllUsesAreOutsideLoop = true;
1977 std::vector<BasedUser> UsersToProcess;
1978 const SCEV *CommonExprs = CollectIVUsers(SI->first, *SI->second, L,
1979 AllUsesAreAddresses,
1980 AllUsesAreOutsideLoop,
1982 // Avoid rewriting the compare instruction with an iv of new stride
1983 // if it's likely the new stride uses will be rewritten using the
1984 // stride of the compare instruction.
1985 if (AllUsesAreAddresses &&
1986 ValidScale(!CommonExprs->isZero(), Scale, UsersToProcess))
1989 // Avoid rewriting the compare instruction with an iv which has
1990 // implicit extension or truncation built into it.
1991 // TODO: This is over-conservative.
1992 if (SE->getTypeSizeInBits(CondUse->getOffset()->getType()) != TyBits)
1995 // If scale is negative, use swapped predicate unless it's testing
1997 if (Scale < 0 && !Cond->isEquality())
1998 Predicate = ICmpInst::getSwappedPredicate(Predicate);
2000 NewStride = &IU->StrideOrder[i];
2001 if (!isa<PointerType>(NewCmpTy))
2002 NewCmpRHS = ConstantInt::get(NewCmpTy, NewCmpVal);
2004 Constant *CI = ConstantInt::get(NewCmpIntTy, NewCmpVal);
2005 NewCmpRHS = ConstantExpr::getIntToPtr(CI, NewCmpTy);
2007 NewOffset = TyBits == NewTyBits
2008 ? SE->getMulExpr(CondUse->getOffset(),
2009 SE->getConstant(CmpTy, Scale))
2010 : SE->getConstant(NewCmpIntTy,
2011 cast<SCEVConstant>(CondUse->getOffset())->getValue()
2012 ->getSExtValue()*Scale);
2017 // Forgo this transformation if it the increment happens to be
2018 // unfortunately positioned after the condition, and the condition
2019 // has multiple uses which prevent it from being moved immediately
2020 // before the branch. See
2021 // test/Transforms/LoopStrengthReduce/change-compare-stride-trickiness-*.ll
2022 // for an example of this situation.
2023 if (!Cond->hasOneUse()) {
2024 for (BasicBlock::iterator I = Cond, E = Cond->getParent()->end();
2031 // Create a new compare instruction using new stride / iv.
2032 ICmpInst *OldCond = Cond;
2033 // Insert new compare instruction.
2034 Cond = new ICmpInst(OldCond, Predicate, NewCmpLHS, NewCmpRHS,
2035 L->getHeader()->getName() + ".termcond");
2037 // Remove the old compare instruction. The old indvar is probably dead too.
2038 DeadInsts.push_back(CondUse->getOperandValToReplace());
2039 OldCond->replaceAllUsesWith(Cond);
2040 OldCond->eraseFromParent();
2042 IU->IVUsesByStride[*NewStride]->addUser(NewOffset, Cond, NewCmpLHS);
2043 CondUse = &IU->IVUsesByStride[*NewStride]->Users.back();
2044 CondStride = NewStride;
2052 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
2053 /// a max computation.
2055 /// This is a narrow solution to a specific, but acute, problem. For loops
2061 /// } while (++i < n);
2063 /// the trip count isn't just 'n', because 'n' might not be positive. And
2064 /// unfortunately this can come up even for loops where the user didn't use
2065 /// a C do-while loop. For example, seemingly well-behaved top-test loops
2066 /// will commonly be lowered like this:
2072 /// } while (++i < n);
2075 /// and then it's possible for subsequent optimization to obscure the if
2076 /// test in such a way that indvars can't find it.
2078 /// When indvars can't find the if test in loops like this, it creates a
2079 /// max expression, which allows it to give the loop a canonical
2080 /// induction variable:
2083 /// max = n < 1 ? 1 : n;
2086 /// } while (++i != max);
2088 /// Canonical induction variables are necessary because the loop passes
2089 /// are designed around them. The most obvious example of this is the
2090 /// LoopInfo analysis, which doesn't remember trip count values. It
2091 /// expects to be able to rediscover the trip count each time it is
2092 /// needed, and it does this using a simple analyis that only succeeds if
2093 /// the loop has a canonical induction variable.
2095 /// However, when it comes time to generate code, the maximum operation
2096 /// can be quite costly, especially if it's inside of an outer loop.
2098 /// This function solves this problem by detecting this type of loop and
2099 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
2100 /// the instructions for the maximum computation.
2102 ICmpInst *LoopStrengthReduce::OptimizeMax(Loop *L, ICmpInst *Cond,
2103 IVStrideUse* &CondUse) {
2104 // Check that the loop matches the pattern we're looking for.
2105 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
2106 Cond->getPredicate() != CmpInst::ICMP_NE)
2109 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
2110 if (!Sel || !Sel->hasOneUse()) return Cond;
2112 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
2113 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2115 const SCEV *One = SE->getIntegerSCEV(1, BackedgeTakenCount->getType());
2117 // Add one to the backedge-taken count to get the trip count.
2118 const SCEV *IterationCount = SE->getAddExpr(BackedgeTakenCount, One);
2120 // Check for a max calculation that matches the pattern.
2121 if (!isa<SCEVSMaxExpr>(IterationCount) && !isa<SCEVUMaxExpr>(IterationCount))
2123 const SCEVNAryExpr *Max = cast<SCEVNAryExpr>(IterationCount);
2124 if (Max != SE->getSCEV(Sel)) return Cond;
2126 // To handle a max with more than two operands, this optimization would
2127 // require additional checking and setup.
2128 if (Max->getNumOperands() != 2)
2131 const SCEV *MaxLHS = Max->getOperand(0);
2132 const SCEV *MaxRHS = Max->getOperand(1);
2133 if (!MaxLHS || MaxLHS != One) return Cond;
2135 // Check the relevant induction variable for conformance to
2137 const SCEV *IV = SE->getSCEV(Cond->getOperand(0));
2138 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2139 if (!AR || !AR->isAffine() ||
2140 AR->getStart() != One ||
2141 AR->getStepRecurrence(*SE) != One)
2144 assert(AR->getLoop() == L &&
2145 "Loop condition operand is an addrec in a different loop!");
2147 // Check the right operand of the select, and remember it, as it will
2148 // be used in the new comparison instruction.
2150 if (SE->getSCEV(Sel->getOperand(1)) == MaxRHS)
2151 NewRHS = Sel->getOperand(1);
2152 else if (SE->getSCEV(Sel->getOperand(2)) == MaxRHS)
2153 NewRHS = Sel->getOperand(2);
2154 if (!NewRHS) return Cond;
2156 // Determine the new comparison opcode. It may be signed or unsigned,
2157 // and the original comparison may be either equality or inequality.
2158 CmpInst::Predicate Pred =
2159 isa<SCEVSMaxExpr>(Max) ? CmpInst::ICMP_SLT : CmpInst::ICMP_ULT;
2160 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
2161 Pred = CmpInst::getInversePredicate(Pred);
2163 // Ok, everything looks ok to change the condition into an SLT or SGE and
2164 // delete the max calculation.
2166 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
2168 // Delete the max calculation instructions.
2169 Cond->replaceAllUsesWith(NewCond);
2170 CondUse->setUser(NewCond);
2171 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2172 Cond->eraseFromParent();
2173 Sel->eraseFromParent();
2174 if (Cmp->use_empty())
2175 Cmp->eraseFromParent();
2179 /// OptimizeShadowIV - If IV is used in a int-to-float cast
2180 /// inside the loop then try to eliminate the cast opeation.
2181 void LoopStrengthReduce::OptimizeShadowIV(Loop *L) {
2183 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
2184 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2187 for (unsigned Stride = 0, e = IU->StrideOrder.size(); Stride != e;
2189 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
2190 IU->IVUsesByStride.find(IU->StrideOrder[Stride]);
2191 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
2192 if (!isa<SCEVConstant>(SI->first))
2195 for (ilist<IVStrideUse>::iterator UI = SI->second->Users.begin(),
2196 E = SI->second->Users.end(); UI != E; /* empty */) {
2197 ilist<IVStrideUse>::iterator CandidateUI = UI;
2199 Instruction *ShadowUse = CandidateUI->getUser();
2200 const Type *DestTy = NULL;
2202 /* If shadow use is a int->float cast then insert a second IV
2203 to eliminate this cast.
2205 for (unsigned i = 0; i < n; ++i)
2211 for (unsigned i = 0; i < n; ++i, ++d)
2214 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser()))
2215 DestTy = UCast->getDestTy();
2216 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser()))
2217 DestTy = SCast->getDestTy();
2218 if (!DestTy) continue;
2221 // If target does not support DestTy natively then do not apply
2222 // this transformation.
2223 EVT DVT = TLI->getValueType(DestTy);
2224 if (!TLI->isTypeLegal(DVT)) continue;
2227 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
2229 if (PH->getNumIncomingValues() != 2) continue;
2231 const Type *SrcTy = PH->getType();
2232 int Mantissa = DestTy->getFPMantissaWidth();
2233 if (Mantissa == -1) continue;
2234 if ((int)SE->getTypeSizeInBits(SrcTy) > Mantissa)
2237 unsigned Entry, Latch;
2238 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
2246 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
2247 if (!Init) continue;
2248 Constant *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());
2250 BinaryOperator *Incr =
2251 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
2252 if (!Incr) continue;
2253 if (Incr->getOpcode() != Instruction::Add
2254 && Incr->getOpcode() != Instruction::Sub)
2257 /* Initialize new IV, double d = 0.0 in above example. */
2258 ConstantInt *C = NULL;
2259 if (Incr->getOperand(0) == PH)
2260 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
2261 else if (Incr->getOperand(1) == PH)
2262 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
2268 // Ignore negative constants, as the code below doesn't handle them
2269 // correctly. TODO: Remove this restriction.
2270 if (!C->getValue().isStrictlyPositive()) continue;
2272 /* Add new PHINode. */
2273 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);
2275 /* create new increment. '++d' in above example. */
2276 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
2277 BinaryOperator *NewIncr =
2278 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
2279 Instruction::FAdd : Instruction::FSub,
2280 NewPH, CFP, "IV.S.next.", Incr);
2282 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
2283 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
2285 /* Remove cast operation */
2286 ShadowUse->replaceAllUsesWith(NewPH);
2287 ShadowUse->eraseFromParent();
2294 /// OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
2295 /// uses in the loop, look to see if we can eliminate some, in favor of using
2296 /// common indvars for the different uses.
2297 void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
2298 // TODO: implement optzns here.
2300 OptimizeShadowIV(L);
2303 /// OptimizeLoopTermCond - Change loop terminating condition to use the
2304 /// postinc iv when possible.
2305 void LoopStrengthReduce::OptimizeLoopTermCond(Loop *L) {
2306 // Finally, get the terminating condition for the loop if possible. If we
2307 // can, we want to change it to use a post-incremented version of its
2308 // induction variable, to allow coalescing the live ranges for the IV into
2309 // one register value.
2310 BasicBlock *LatchBlock = L->getLoopLatch();
2311 BasicBlock *ExitingBlock = L->getExitingBlock();
2314 // Multiple exits, just look at the exit in the latch block if there is one.
2315 ExitingBlock = LatchBlock;
2316 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2319 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
2322 // Search IVUsesByStride to find Cond's IVUse if there is one.
2323 IVStrideUse *CondUse = 0;
2324 const SCEV *const *CondStride = 0;
2325 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2326 if (!FindIVUserForCond(Cond, CondUse, CondStride))
2327 return; // setcc doesn't use the IV.
2329 if (ExitingBlock != LatchBlock) {
2330 if (!Cond->hasOneUse())
2331 // See below, we don't want the condition to be cloned.
2334 // If exiting block is the latch block, we know it's safe and profitable to
2335 // transform the icmp to use post-inc iv. Otherwise do so only if it would
2336 // not reuse another iv and its iv would be reused by other uses. We are
2337 // optimizing for the case where the icmp is the only use of the iv.
2338 IVUsersOfOneStride &StrideUses = *IU->IVUsesByStride[*CondStride];
2339 for (ilist<IVStrideUse>::iterator I = StrideUses.Users.begin(),
2340 E = StrideUses.Users.end(); I != E; ++I) {
2341 if (I->getUser() == Cond)
2343 if (!I->isUseOfPostIncrementedValue())
2347 // FIXME: This is expensive, and worse still ChangeCompareStride does a
2348 // similar check. Can we perform all the icmp related transformations after
2349 // StrengthReduceStridedIVUsers?
2350 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(*CondStride)) {
2351 int64_t SInt = SC->getValue()->getSExtValue();
2352 for (unsigned NewStride = 0, ee = IU->StrideOrder.size(); NewStride != ee;
2354 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
2355 IU->IVUsesByStride.find(IU->StrideOrder[NewStride]);
2356 if (!isa<SCEVConstant>(SI->first) || SI->first == *CondStride)
2359 cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
2361 return; // This can definitely be reused.
2362 if (unsigned(abs64(SSInt)) < SInt || (SSInt % SInt) != 0)
2364 int64_t Scale = SSInt / SInt;
2365 bool AllUsesAreAddresses = true;
2366 bool AllUsesAreOutsideLoop = true;
2367 std::vector<BasedUser> UsersToProcess;
2368 const SCEV *CommonExprs = CollectIVUsers(SI->first, *SI->second, L,
2369 AllUsesAreAddresses,
2370 AllUsesAreOutsideLoop,
2372 // Avoid rewriting the compare instruction with an iv of new stride
2373 // if it's likely the new stride uses will be rewritten using the
2374 // stride of the compare instruction.
2375 if (AllUsesAreAddresses &&
2376 ValidScale(!CommonExprs->isZero(), Scale, UsersToProcess))
2381 StrideNoReuse.insert(*CondStride);
2384 // If the trip count is computed in terms of a max (due to ScalarEvolution
2385 // being unable to find a sufficient guard, for example), change the loop
2386 // comparison to use SLT or ULT instead of NE.
2387 Cond = OptimizeMax(L, Cond, CondUse);
2389 // If possible, change stride and operands of the compare instruction to
2390 // eliminate one stride.
2391 if (ExitingBlock == LatchBlock)
2392 Cond = ChangeCompareStride(L, Cond, CondUse, CondStride);
2394 // It's possible for the setcc instruction to be anywhere in the loop, and
2395 // possible for it to have multiple users. If it is not immediately before
2396 // the latch block branch, move it.
2397 if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
2398 if (Cond->hasOneUse()) { // Condition has a single use, just move it.
2399 Cond->moveBefore(TermBr);
2401 // Otherwise, clone the terminating condition and insert into the loopend.
2402 Cond = cast<ICmpInst>(Cond->clone());
2403 Cond->setName(L->getHeader()->getName() + ".termcond");
2404 LatchBlock->getInstList().insert(TermBr, Cond);
2406 // Clone the IVUse, as the old use still exists!
2407 IU->IVUsesByStride[*CondStride]->addUser(CondUse->getOffset(), Cond,
2408 CondUse->getOperandValToReplace());
2409 CondUse = &IU->IVUsesByStride[*CondStride]->Users.back();
2413 // If we get to here, we know that we can transform the setcc instruction to
2414 // use the post-incremented version of the IV, allowing us to coalesce the
2415 // live ranges for the IV correctly.
2416 CondUse->setOffset(SE->getMinusSCEV(CondUse->getOffset(), *CondStride));
2417 CondUse->setIsUseOfPostIncrementedValue(true);
2423 /// isUsedByExitBranch - Return true if icmp is used by a loop terminating
2424 /// conditional branch or it's and / or with other conditions before being used
2425 /// as the condition.
2426 static bool isUsedByExitBranch(ICmpInst *Cond, Loop *L) {
2427 BasicBlock *CondBB = Cond->getParent();
2428 if (!L->isLoopExiting(CondBB))
2430 BranchInst *TermBr = dyn_cast<BranchInst>(CondBB->getTerminator());
2431 if (!TermBr || !TermBr->isConditional())
2434 Value *User = *Cond->use_begin();
2435 Instruction *UserInst = dyn_cast<Instruction>(User);
2437 (UserInst->getOpcode() == Instruction::And ||
2438 UserInst->getOpcode() == Instruction::Or)) {
2439 if (!UserInst->hasOneUse() || UserInst->getParent() != CondBB)
2441 User = *User->use_begin();
2442 UserInst = dyn_cast<Instruction>(User);
2444 return User == TermBr;
2447 /// OptimizeLoopCountIV - If, after all sharing of IVs, the IV used for deciding
2448 /// when to exit the loop is used only for that purpose, try to rearrange things
2449 /// so it counts down to a test against zero which tends to be cheaper.
2450 void LoopStrengthReduce::OptimizeLoopCountIV(const SCEV *Stride,
2451 IVUsersOfOneStride &Uses,
2453 if (Uses.Users.size() != 1)
2456 // If the only use is an icmp of an loop exiting conditional branch, then
2457 // attempts the optimization.
2458 BasedUser User = BasedUser(*Uses.Users.begin(), SE);
2459 Instruction *Inst = User.Inst;
2460 if (!L->contains(Inst->getParent()))
2463 ICmpInst *Cond = dyn_cast<ICmpInst>(Inst);
2466 // Handle only tests for equality for the moment.
2467 if (Cond->getPredicate() != CmpInst::ICMP_EQ || !Cond->hasOneUse())
2469 if (!isUsedByExitBranch(Cond, L))
2472 Value *CondOp0 = Cond->getOperand(0);
2473 const SCEV *IV = SE->getSCEV(CondOp0);
2474 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2475 if (!AR || !AR->isAffine())
2478 const SCEVConstant *SC = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE));
2479 if (!SC || SC->getValue()->getSExtValue() < 0)
2480 // If it's already counting down, don't do anything.
2483 // If the RHS of the comparison is not an loop invariant, the rewrite
2484 // cannot be done. Also bail out if it's already comparing against a zero.
2485 Value *RHS = Cond->getOperand(1);
2486 if (!L->isLoopInvariant(RHS) ||
2487 (isa<ConstantInt>(RHS) && cast<ConstantInt>(RHS)->isZero()))
2490 // Make sure the IV is only used for counting. Value may be preinc or
2491 // postinc; 2 uses in either case.
2492 if (!CondOp0->hasNUses(2))
2496 if (User.isUseOfPostIncrementedValue) {
2497 // Value tested is postinc. Find the phi node.
2498 Incr = dyn_cast<BinaryOperator>(CondOp0);
2499 if (!Incr || Incr->getOpcode() != Instruction::Add)
2502 Instruction::use_iterator UI = CondOp0->use_begin();
2503 PHIExpr = dyn_cast<PHINode>(UI);
2505 PHIExpr = dyn_cast<PHINode>(++UI);
2506 // 1 use for preinc value, the increment.
2507 if (!PHIExpr || !PHIExpr->hasOneUse())
2510 assert(isa<PHINode>(CondOp0) &&
2511 "Unexpected loop exiting counting instruction sequence!");
2512 PHIExpr = cast<PHINode>(CondOp0);
2513 // Value tested is preinc. Find the increment.
2514 // A CmpInst is not a BinaryOperator; we depend on this.
2515 Instruction::use_iterator UI = PHIExpr->use_begin();
2516 Incr = dyn_cast<BinaryOperator>(UI);
2518 Incr = dyn_cast<BinaryOperator>(++UI);
2519 // One use for postinc value, the phi. Unnecessarily conservative?
2520 if (!Incr || !Incr->hasOneUse() || Incr->getOpcode() != Instruction::Add)
2524 // Replace the increment with a decrement.
2525 DEBUG(errs() << " Examining ");
2526 if (User.isUseOfPostIncrementedValue)
2527 DEBUG(errs() << "postinc");
2529 DEBUG(errs() << "preinc");
2530 DEBUG(errs() << " use ");
2531 DEBUG(WriteAsOperand(errs(), CondOp0, /*PrintType=*/false));
2532 DEBUG(errs() << " in Inst: " << *Inst << '\n');
2533 BinaryOperator *Decr = BinaryOperator::Create(Instruction::Sub,
2534 Incr->getOperand(0), Incr->getOperand(1), "tmp", Incr);
2535 Incr->replaceAllUsesWith(Decr);
2536 Incr->eraseFromParent();
2538 // Substitute endval-startval for the original startval, and 0 for the
2539 // original endval. Since we're only testing for equality this is OK even
2540 // if the computation wraps around.
2541 BasicBlock *Preheader = L->getLoopPreheader();
2542 Instruction *PreInsertPt = Preheader->getTerminator();
2543 unsigned InBlock = L->contains(PHIExpr->getIncomingBlock(0)) ? 1 : 0;
2544 Value *StartVal = PHIExpr->getIncomingValue(InBlock);
2545 Value *EndVal = Cond->getOperand(1);
2546 DEBUG(errs() << " Optimize loop counting iv to count down ["
2547 << *EndVal << " .. " << *StartVal << "]\n");
2549 // FIXME: check for case where both are constant.
2550 Constant* Zero = ConstantInt::get(Cond->getOperand(1)->getType(), 0);
2551 BinaryOperator *NewStartVal = BinaryOperator::Create(Instruction::Sub,
2552 EndVal, StartVal, "tmp", PreInsertPt);
2553 PHIExpr->setIncomingValue(InBlock, NewStartVal);
2554 Cond->setOperand(1, Zero);
2555 DEBUG(errs() << " New icmp: " << *Cond << "\n");
2561 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager &LPM) {
2563 IU = &getAnalysis<IVUsers>();
2564 LI = &getAnalysis<LoopInfo>();
2565 DT = &getAnalysis<DominatorTree>();
2566 SE = &getAnalysis<ScalarEvolution>();
2569 // If LoopSimplify form is not available, stay out of trouble.
2570 if (!L->getLoopPreheader() || !L->getLoopLatch())
2573 if (!IU->IVUsesByStride.empty()) {
2574 DEBUG(errs() << "\nLSR on \"" << L->getHeader()->getParent()->getName()
2578 // Sort the StrideOrder so we process larger strides first.
2579 std::stable_sort(IU->StrideOrder.begin(), IU->StrideOrder.end(),
2582 // Optimize induction variables. Some indvar uses can be transformed to use
2583 // strides that will be needed for other purposes. A common example of this
2584 // is the exit test for the loop, which can often be rewritten to use the
2585 // computation of some other indvar to decide when to terminate the loop.
2588 // Change loop terminating condition to use the postinc iv when possible
2589 // and optimize loop terminating compare. FIXME: Move this after
2590 // StrengthReduceStridedIVUsers?
2591 OptimizeLoopTermCond(L);
2593 // FIXME: We can shrink overlarge IV's here. e.g. if the code has
2594 // computation in i64 values and the target doesn't support i64, demote
2595 // the computation to 32-bit if safe.
2597 // FIXME: Attempt to reuse values across multiple IV's. In particular, we
2598 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should
2599 // be codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC.
2600 // Need to be careful that IV's are all the same type. Only works for
2601 // intptr_t indvars.
2603 // IVsByStride keeps IVs for one particular loop.
2604 assert(IVsByStride.empty() && "Stale entries in IVsByStride?");
2606 // Note: this processes each stride/type pair individually. All users
2607 // passed into StrengthReduceStridedIVUsers have the same type AND stride.
2608 // Also, note that we iterate over IVUsesByStride indirectly by using
2609 // StrideOrder. This extra layer of indirection makes the ordering of
2610 // strides deterministic - not dependent on map order.
2611 for (unsigned Stride = 0, e = IU->StrideOrder.size();
2612 Stride != e; ++Stride) {
2613 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
2614 IU->IVUsesByStride.find(IU->StrideOrder[Stride]);
2615 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
2616 // FIXME: Generalize to non-affine IV's.
2617 if (!SI->first->isLoopInvariant(L))
2619 StrengthReduceStridedIVUsers(SI->first, *SI->second, L);
2622 // After all sharing is done, see if we can adjust the loop to test against
2623 // zero instead of counting up to a maximum. This is usually faster.
2624 for (unsigned Stride = 0, e = IU->StrideOrder.size();
2625 Stride != e; ++Stride) {
2626 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
2627 IU->IVUsesByStride.find(IU->StrideOrder[Stride]);
2628 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
2629 // FIXME: Generalize to non-affine IV's.
2630 if (!SI->first->isLoopInvariant(L))
2632 OptimizeLoopCountIV(SI->first, *SI->second, L);
2636 // We're done analyzing this loop; release all the state we built up for it.
2637 IVsByStride.clear();
2638 StrideNoReuse.clear();
2640 // Clean up after ourselves
2641 if (!DeadInsts.empty())
2642 DeleteTriviallyDeadInstructions();
2644 // At this point, it is worth checking to see if any recurrence PHIs are also
2645 // dead, so that we can remove them as well.
2646 DeleteDeadPHIs(L->getHeader());