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 void OptimizeIndvars(Loop *L);
137 /// OptimizeLoopTermCond - Change loop terminating condition to use the
138 /// postinc iv when possible.
139 void OptimizeLoopTermCond(Loop *L);
141 /// OptimizeShadowIV - If IV is used in a int-to-float cast
142 /// inside the loop then try to eliminate the cast opeation.
143 void OptimizeShadowIV(Loop *L);
145 /// OptimizeMax - Rewrite the loop's terminating condition
146 /// if it uses a max computation.
147 ICmpInst *OptimizeMax(Loop *L, ICmpInst *Cond,
148 IVStrideUse* &CondUse);
150 /// OptimizeLoopCountIV - If, after all sharing of IVs, the IV used for
151 /// deciding when to exit the loop is used only for that purpose, try to
152 /// rearrange things so it counts down to a test against zero.
153 bool OptimizeLoopCountIV(Loop *L);
154 bool OptimizeLoopCountIVOfStride(const SCEV* &Stride,
155 IVStrideUse* &CondUse, Loop *L);
157 /// StrengthReduceIVUsersOfStride - Strength reduce all of the users of a
158 /// single stride of IV. All of the users may have different starting
159 /// values, and this may not be the only stride.
160 void StrengthReduceIVUsersOfStride(const SCEV *const &Stride,
161 IVUsersOfOneStride &Uses,
163 void StrengthReduceIVUsers(Loop *L);
165 ICmpInst *ChangeCompareStride(Loop *L, ICmpInst *Cond,
166 IVStrideUse* &CondUse,
167 const SCEV* &CondStride,
168 bool PostPass = false);
170 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse,
171 const SCEV* &CondStride);
172 bool RequiresTypeConversion(const Type *Ty, const Type *NewTy);
173 const SCEV *CheckForIVReuse(bool, bool, bool, const SCEV *const&,
174 IVExpr&, const Type*,
175 const std::vector<BasedUser>& UsersToProcess);
176 bool ValidScale(bool, int64_t,
177 const std::vector<BasedUser>& UsersToProcess);
178 bool ValidOffset(bool, int64_t, int64_t,
179 const std::vector<BasedUser>& UsersToProcess);
180 const SCEV *CollectIVUsers(const SCEV *const &Stride,
181 IVUsersOfOneStride &Uses,
183 bool &AllUsesAreAddresses,
184 bool &AllUsesAreOutsideLoop,
185 std::vector<BasedUser> &UsersToProcess);
186 bool StrideMightBeShared(const SCEV *Stride, Loop *L, bool CheckPreInc);
187 bool ShouldUseFullStrengthReductionMode(
188 const std::vector<BasedUser> &UsersToProcess,
190 bool AllUsesAreAddresses,
192 void PrepareToStrengthReduceFully(
193 std::vector<BasedUser> &UsersToProcess,
195 const SCEV *CommonExprs,
197 SCEVExpander &PreheaderRewriter);
198 void PrepareToStrengthReduceFromSmallerStride(
199 std::vector<BasedUser> &UsersToProcess,
201 const IVExpr &ReuseIV,
202 Instruction *PreInsertPt);
203 void PrepareToStrengthReduceWithNewPhi(
204 std::vector<BasedUser> &UsersToProcess,
206 const SCEV *CommonExprs,
208 Instruction *IVIncInsertPt,
210 SCEVExpander &PreheaderRewriter);
212 void DeleteTriviallyDeadInstructions();
216 char LoopStrengthReduce::ID = 0;
217 static RegisterPass<LoopStrengthReduce>
218 X("loop-reduce", "Loop Strength Reduction");
220 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
221 return new LoopStrengthReduce(TLI);
224 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
225 /// specified set are trivially dead, delete them and see if this makes any of
226 /// their operands subsequently dead.
227 void LoopStrengthReduce::DeleteTriviallyDeadInstructions() {
228 if (DeadInsts.empty()) return;
230 while (!DeadInsts.empty()) {
231 Instruction *I = dyn_cast_or_null<Instruction>(DeadInsts.back());
232 DeadInsts.pop_back();
234 if (I == 0 || !isInstructionTriviallyDead(I))
237 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) {
238 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
241 DeadInsts.push_back(U);
245 I->eraseFromParent();
250 /// containsAddRecFromDifferentLoop - Determine whether expression S involves a
251 /// subexpression that is an AddRec from a loop other than L. An outer loop
252 /// of L is OK, but not an inner loop nor a disjoint loop.
253 static bool containsAddRecFromDifferentLoop(const SCEV *S, Loop *L) {
254 // This is very common, put it first.
255 if (isa<SCEVConstant>(S))
257 if (const SCEVCommutativeExpr *AE = dyn_cast<SCEVCommutativeExpr>(S)) {
258 for (unsigned int i=0; i< AE->getNumOperands(); i++)
259 if (containsAddRecFromDifferentLoop(AE->getOperand(i), L))
263 if (const SCEVAddRecExpr *AE = dyn_cast<SCEVAddRecExpr>(S)) {
264 if (const Loop *newLoop = AE->getLoop()) {
267 // if newLoop is an outer loop of L, this is OK.
268 if (!LoopInfo::isNotAlreadyContainedIn(L, newLoop))
273 if (const SCEVUDivExpr *DE = dyn_cast<SCEVUDivExpr>(S))
274 return containsAddRecFromDifferentLoop(DE->getLHS(), L) ||
275 containsAddRecFromDifferentLoop(DE->getRHS(), L);
277 // SCEVSDivExpr has been backed out temporarily, but will be back; we'll
278 // need this when it is.
279 if (const SCEVSDivExpr *DE = dyn_cast<SCEVSDivExpr>(S))
280 return containsAddRecFromDifferentLoop(DE->getLHS(), L) ||
281 containsAddRecFromDifferentLoop(DE->getRHS(), L);
283 if (const SCEVCastExpr *CE = dyn_cast<SCEVCastExpr>(S))
284 return containsAddRecFromDifferentLoop(CE->getOperand(), L);
288 /// isAddressUse - Returns true if the specified instruction is using the
289 /// specified value as an address.
290 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
291 bool isAddress = isa<LoadInst>(Inst);
292 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
293 if (SI->getOperand(1) == OperandVal)
295 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
296 // Addressing modes can also be folded into prefetches and a variety
298 switch (II->getIntrinsicID()) {
300 case Intrinsic::prefetch:
301 case Intrinsic::x86_sse2_loadu_dq:
302 case Intrinsic::x86_sse2_loadu_pd:
303 case Intrinsic::x86_sse_loadu_ps:
304 case Intrinsic::x86_sse_storeu_ps:
305 case Intrinsic::x86_sse2_storeu_pd:
306 case Intrinsic::x86_sse2_storeu_dq:
307 case Intrinsic::x86_sse2_storel_dq:
308 if (II->getOperand(1) == OperandVal)
316 /// getAccessType - Return the type of the memory being accessed.
317 static const Type *getAccessType(const Instruction *Inst) {
318 const Type *AccessTy = Inst->getType();
319 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
320 AccessTy = SI->getOperand(0)->getType();
321 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
322 // Addressing modes can also be folded into prefetches and a variety
324 switch (II->getIntrinsicID()) {
326 case Intrinsic::x86_sse_storeu_ps:
327 case Intrinsic::x86_sse2_storeu_pd:
328 case Intrinsic::x86_sse2_storeu_dq:
329 case Intrinsic::x86_sse2_storel_dq:
330 AccessTy = II->getOperand(1)->getType();
338 /// BasedUser - For a particular base value, keep information about how we've
339 /// partitioned the expression so far.
341 /// SE - The current ScalarEvolution object.
344 /// Base - The Base value for the PHI node that needs to be inserted for
345 /// this use. As the use is processed, information gets moved from this
346 /// field to the Imm field (below). BasedUser values are sorted by this
350 /// Inst - The instruction using the induction variable.
353 /// OperandValToReplace - The operand value of Inst to replace with the
355 Value *OperandValToReplace;
357 /// Imm - The immediate value that should be added to the base immediately
358 /// before Inst, because it will be folded into the imm field of the
359 /// instruction. This is also sometimes used for loop-variant values that
360 /// must be added inside the loop.
363 /// Phi - The induction variable that performs the striding that
364 /// should be used for this user.
367 // isUseOfPostIncrementedValue - True if this should use the
368 // post-incremented version of this IV, not the preincremented version.
369 // This can only be set in special cases, such as the terminating setcc
370 // instruction for a loop and uses outside the loop that are dominated by
372 bool isUseOfPostIncrementedValue;
374 BasedUser(IVStrideUse &IVSU, ScalarEvolution *se)
375 : SE(se), Base(IVSU.getOffset()), Inst(IVSU.getUser()),
376 OperandValToReplace(IVSU.getOperandValToReplace()),
377 Imm(SE->getIntegerSCEV(0, Base->getType())),
378 isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue()) {}
380 // Once we rewrite the code to insert the new IVs we want, update the
381 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
383 void RewriteInstructionToUseNewBase(const SCEV *const &NewBase,
384 Instruction *InsertPt,
385 SCEVExpander &Rewriter, Loop *L, Pass *P,
387 SmallVectorImpl<WeakVH> &DeadInsts);
389 Value *InsertCodeForBaseAtPosition(const SCEV *const &NewBase,
391 SCEVExpander &Rewriter,
392 Instruction *IP, Loop *L,
398 void BasedUser::dump() const {
399 errs() << " Base=" << *Base;
400 errs() << " Imm=" << *Imm;
401 errs() << " Inst: " << *Inst;
404 Value *BasedUser::InsertCodeForBaseAtPosition(const SCEV *const &NewBase,
406 SCEVExpander &Rewriter,
407 Instruction *IP, Loop *L,
409 // Figure out where we *really* want to insert this code. In particular, if
410 // the user is inside of a loop that is nested inside of L, we really don't
411 // want to insert this expression before the user, we'd rather pull it out as
412 // many loops as possible.
413 Instruction *BaseInsertPt = IP;
415 // Figure out the most-nested loop that IP is in.
416 Loop *InsertLoop = LI.getLoopFor(IP->getParent());
418 // If InsertLoop is not L, and InsertLoop is nested inside of L, figure out
419 // the preheader of the outer-most loop where NewBase is not loop invariant.
420 if (L->contains(IP->getParent()))
421 while (InsertLoop && NewBase->isLoopInvariant(InsertLoop)) {
422 BaseInsertPt = InsertLoop->getLoopPreheader()->getTerminator();
423 InsertLoop = InsertLoop->getParentLoop();
426 Value *Base = Rewriter.expandCodeFor(NewBase, 0, BaseInsertPt);
428 const SCEV *NewValSCEV = SE->getUnknown(Base);
430 // Always emit the immediate into the same block as the user.
431 NewValSCEV = SE->getAddExpr(NewValSCEV, Imm);
433 return Rewriter.expandCodeFor(NewValSCEV, Ty, IP);
437 // Once we rewrite the code to insert the new IVs we want, update the
438 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
439 // to it. NewBasePt is the last instruction which contributes to the
440 // value of NewBase in the case that it's a diffferent instruction from
441 // the PHI that NewBase is computed from, or null otherwise.
443 void BasedUser::RewriteInstructionToUseNewBase(const SCEV *const &NewBase,
444 Instruction *NewBasePt,
445 SCEVExpander &Rewriter, Loop *L, Pass *P,
447 SmallVectorImpl<WeakVH> &DeadInsts) {
448 if (!isa<PHINode>(Inst)) {
449 // By default, insert code at the user instruction.
450 BasicBlock::iterator InsertPt = Inst;
452 // However, if the Operand is itself an instruction, the (potentially
453 // complex) inserted code may be shared by many users. Because of this, we
454 // want to emit code for the computation of the operand right before its old
455 // computation. This is usually safe, because we obviously used to use the
456 // computation when it was computed in its current block. However, in some
457 // cases (e.g. use of a post-incremented induction variable) the NewBase
458 // value will be pinned to live somewhere after the original computation.
459 // In this case, we have to back off.
461 // If this is a use outside the loop (which means after, since it is based
462 // on a loop indvar) we use the post-incremented value, so that we don't
463 // artificially make the preinc value live out the bottom of the loop.
464 if (!isUseOfPostIncrementedValue && L->contains(Inst->getParent())) {
465 if (NewBasePt && isa<PHINode>(OperandValToReplace)) {
466 InsertPt = NewBasePt;
468 } else if (Instruction *OpInst
469 = dyn_cast<Instruction>(OperandValToReplace)) {
471 while (isa<PHINode>(InsertPt)) ++InsertPt;
474 Value *NewVal = InsertCodeForBaseAtPosition(NewBase,
475 OperandValToReplace->getType(),
476 Rewriter, InsertPt, L, LI);
477 // Replace the use of the operand Value with the new Phi we just created.
478 Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
480 DEBUG(errs() << " Replacing with ");
481 DEBUG(WriteAsOperand(errs(), NewVal, /*PrintType=*/false));
482 DEBUG(errs() << ", which has value " << *NewBase << " plus IMM "
487 // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm
488 // expression into each operand block that uses it. Note that PHI nodes can
489 // have multiple entries for the same predecessor. We use a map to make sure
490 // that a PHI node only has a single Value* for each predecessor (which also
491 // prevents us from inserting duplicate code in some blocks).
492 DenseMap<BasicBlock*, Value*> InsertedCode;
493 PHINode *PN = cast<PHINode>(Inst);
494 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
495 if (PN->getIncomingValue(i) == OperandValToReplace) {
496 // If the original expression is outside the loop, put the replacement
497 // code in the same place as the original expression,
498 // which need not be an immediate predecessor of this PHI. This way we
499 // need only one copy of it even if it is referenced multiple times in
500 // the PHI. We don't do this when the original expression is inside the
501 // loop because multiple copies sometimes do useful sinking of code in
503 Instruction *OldLoc = dyn_cast<Instruction>(OperandValToReplace);
504 BasicBlock *PHIPred = PN->getIncomingBlock(i);
505 if (L->contains(OldLoc->getParent())) {
506 // If this is a critical edge, split the edge so that we do not insert
507 // the code on all predecessor/successor paths. We do this unless this
508 // is the canonical backedge for this loop, as this can make some
509 // inserted code be in an illegal position.
510 if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 &&
511 !isa<IndirectBrInst>(PHIPred->getTerminator()) &&
512 (PN->getParent() != L->getHeader() || !L->contains(PHIPred))) {
514 // First step, split the critical edge.
515 BasicBlock *NewBB = SplitCriticalEdge(PHIPred, PN->getParent(),
518 // Next step: move the basic block. In particular, if the PHI node
519 // is outside of the loop, and PredTI is in the loop, we want to
520 // move the block to be immediately before the PHI block, not
521 // immediately after PredTI.
522 if (L->contains(PHIPred) && !L->contains(PN->getParent()))
523 NewBB->moveBefore(PN->getParent());
525 // Splitting the edge can reduce the number of PHI entries we have.
526 e = PN->getNumIncomingValues();
528 i = PN->getBasicBlockIndex(PHIPred);
531 Value *&Code = InsertedCode[PHIPred];
533 // Insert the code into the end of the predecessor block.
534 Instruction *InsertPt = (L->contains(OldLoc->getParent())) ?
535 PHIPred->getTerminator() :
536 OldLoc->getParent()->getTerminator();
537 Code = InsertCodeForBaseAtPosition(NewBase, PN->getType(),
538 Rewriter, InsertPt, L, LI);
540 DEBUG(errs() << " Changing PHI use to ");
541 DEBUG(WriteAsOperand(errs(), Code, /*PrintType=*/false));
542 DEBUG(errs() << ", which has value " << *NewBase << " plus IMM "
546 // Replace the use of the operand Value with the new Phi we just created.
547 PN->setIncomingValue(i, Code);
552 // PHI node might have become a constant value after SplitCriticalEdge.
553 DeadInsts.push_back(Inst);
557 /// fitsInAddressMode - Return true if V can be subsumed within an addressing
558 /// mode, and does not need to be put in a register first.
559 static bool fitsInAddressMode(const SCEV *const &V, const Type *AccessTy,
560 const TargetLowering *TLI, bool HasBaseReg) {
561 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
562 int64_t VC = SC->getValue()->getSExtValue();
564 TargetLowering::AddrMode AM;
566 AM.HasBaseReg = HasBaseReg;
567 return TLI->isLegalAddressingMode(AM, AccessTy);
569 // Defaults to PPC. PPC allows a sign-extended 16-bit immediate field.
570 return (VC > -(1 << 16) && VC < (1 << 16)-1);
574 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
575 if (GlobalValue *GV = dyn_cast<GlobalValue>(SU->getValue())) {
577 TargetLowering::AddrMode AM;
579 AM.HasBaseReg = HasBaseReg;
580 return TLI->isLegalAddressingMode(AM, AccessTy);
582 // Default: assume global addresses are not legal.
589 /// MoveLoopVariantsToImmediateField - Move any subexpressions from Val that are
590 /// loop varying to the Imm operand.
591 static void MoveLoopVariantsToImmediateField(const SCEV *&Val, const SCEV *&Imm,
592 Loop *L, ScalarEvolution *SE) {
593 if (Val->isLoopInvariant(L)) return; // Nothing to do.
595 if (const SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
596 SmallVector<const SCEV *, 4> NewOps;
597 NewOps.reserve(SAE->getNumOperands());
599 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
600 if (!SAE->getOperand(i)->isLoopInvariant(L)) {
601 // If this is a loop-variant expression, it must stay in the immediate
602 // field of the expression.
603 Imm = SE->getAddExpr(Imm, SAE->getOperand(i));
605 NewOps.push_back(SAE->getOperand(i));
609 Val = SE->getIntegerSCEV(0, Val->getType());
611 Val = SE->getAddExpr(NewOps);
612 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
613 // Try to pull immediates out of the start value of nested addrec's.
614 const SCEV *Start = SARE->getStart();
615 MoveLoopVariantsToImmediateField(Start, Imm, L, SE);
617 SmallVector<const SCEV *, 4> Ops(SARE->op_begin(), SARE->op_end());
619 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
621 // Otherwise, all of Val is variant, move the whole thing over.
622 Imm = SE->getAddExpr(Imm, Val);
623 Val = SE->getIntegerSCEV(0, Val->getType());
628 /// MoveImmediateValues - Look at Val, and pull out any additions of constants
629 /// that can fit into the immediate field of instructions in the target.
630 /// Accumulate these immediate values into the Imm value.
631 static void MoveImmediateValues(const TargetLowering *TLI,
632 const Type *AccessTy,
633 const SCEV *&Val, const SCEV *&Imm,
634 bool isAddress, Loop *L,
635 ScalarEvolution *SE) {
636 if (const SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
637 SmallVector<const SCEV *, 4> NewOps;
638 NewOps.reserve(SAE->getNumOperands());
640 for (unsigned i = 0; i != SAE->getNumOperands(); ++i) {
641 const SCEV *NewOp = SAE->getOperand(i);
642 MoveImmediateValues(TLI, AccessTy, NewOp, Imm, isAddress, L, SE);
644 if (!NewOp->isLoopInvariant(L)) {
645 // If this is a loop-variant expression, it must stay in the immediate
646 // field of the expression.
647 Imm = SE->getAddExpr(Imm, NewOp);
649 NewOps.push_back(NewOp);
654 Val = SE->getIntegerSCEV(0, Val->getType());
656 Val = SE->getAddExpr(NewOps);
658 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
659 // Try to pull immediates out of the start value of nested addrec's.
660 const SCEV *Start = SARE->getStart();
661 MoveImmediateValues(TLI, AccessTy, Start, Imm, isAddress, L, SE);
663 if (Start != SARE->getStart()) {
664 SmallVector<const SCEV *, 4> Ops(SARE->op_begin(), SARE->op_end());
666 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
669 } else if (const SCEVMulExpr *SME = dyn_cast<SCEVMulExpr>(Val)) {
670 // Transform "8 * (4 + v)" -> "32 + 8*V" if "32" fits in the immed field.
672 fitsInAddressMode(SME->getOperand(0), AccessTy, TLI, false) &&
673 SME->getNumOperands() == 2 && SME->isLoopInvariant(L)) {
675 const SCEV *SubImm = SE->getIntegerSCEV(0, Val->getType());
676 const SCEV *NewOp = SME->getOperand(1);
677 MoveImmediateValues(TLI, AccessTy, NewOp, SubImm, isAddress, L, SE);
679 // If we extracted something out of the subexpressions, see if we can
681 if (NewOp != SME->getOperand(1)) {
682 // Scale SubImm up by "8". If the result is a target constant, we are
684 SubImm = SE->getMulExpr(SubImm, SME->getOperand(0));
685 if (fitsInAddressMode(SubImm, AccessTy, TLI, false)) {
686 // Accumulate the immediate.
687 Imm = SE->getAddExpr(Imm, SubImm);
689 // Update what is left of 'Val'.
690 Val = SE->getMulExpr(SME->getOperand(0), NewOp);
697 // Loop-variant expressions must stay in the immediate field of the
699 if ((isAddress && fitsInAddressMode(Val, AccessTy, TLI, false)) ||
700 !Val->isLoopInvariant(L)) {
701 Imm = SE->getAddExpr(Imm, Val);
702 Val = SE->getIntegerSCEV(0, Val->getType());
706 // Otherwise, no immediates to move.
709 static void MoveImmediateValues(const TargetLowering *TLI,
711 const SCEV *&Val, const SCEV *&Imm,
712 bool isAddress, Loop *L,
713 ScalarEvolution *SE) {
714 const Type *AccessTy = getAccessType(User);
715 MoveImmediateValues(TLI, AccessTy, Val, Imm, isAddress, L, SE);
718 /// SeparateSubExprs - Decompose Expr into all of the subexpressions that are
719 /// added together. This is used to reassociate common addition subexprs
720 /// together for maximal sharing when rewriting bases.
721 static void SeparateSubExprs(SmallVector<const SCEV *, 16> &SubExprs,
723 ScalarEvolution *SE) {
724 if (const SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) {
725 for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
726 SeparateSubExprs(SubExprs, AE->getOperand(j), SE);
727 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) {
728 const SCEV *Zero = SE->getIntegerSCEV(0, Expr->getType());
729 if (SARE->getOperand(0) == Zero) {
730 SubExprs.push_back(Expr);
732 // Compute the addrec with zero as its base.
733 SmallVector<const SCEV *, 4> Ops(SARE->op_begin(), SARE->op_end());
734 Ops[0] = Zero; // Start with zero base.
735 SubExprs.push_back(SE->getAddRecExpr(Ops, SARE->getLoop()));
738 SeparateSubExprs(SubExprs, SARE->getOperand(0), SE);
740 } else if (!Expr->isZero()) {
742 SubExprs.push_back(Expr);
746 // This is logically local to the following function, but C++ says we have
747 // to make it file scope.
748 struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; };
750 /// RemoveCommonExpressionsFromUseBases - Look through all of the Bases of all
751 /// the Uses, removing any common subexpressions, except that if all such
752 /// subexpressions can be folded into an addressing mode for all uses inside
753 /// the loop (this case is referred to as "free" in comments herein) we do
754 /// not remove anything. This looks for things like (a+b+c) and
755 /// (a+c+d) and computes the common (a+c) subexpression. The common expression
756 /// is *removed* from the Bases and returned.
758 RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses,
759 ScalarEvolution *SE, Loop *L,
760 const TargetLowering *TLI) {
761 unsigned NumUses = Uses.size();
763 // Only one use? This is a very common case, so we handle it specially and
765 const SCEV *Zero = SE->getIntegerSCEV(0, Uses[0].Base->getType());
766 const SCEV *Result = Zero;
767 const SCEV *FreeResult = Zero;
769 // If the use is inside the loop, use its base, regardless of what it is:
770 // it is clearly shared across all the IV's. If the use is outside the loop
771 // (which means after it) we don't want to factor anything *into* the loop,
772 // so just use 0 as the base.
773 if (L->contains(Uses[0].Inst->getParent()))
774 std::swap(Result, Uses[0].Base);
778 // To find common subexpressions, count how many of Uses use each expression.
779 // If any subexpressions are used Uses.size() times, they are common.
780 // Also track whether all uses of each expression can be moved into an
781 // an addressing mode "for free"; such expressions are left within the loop.
782 // struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; };
783 std::map<const SCEV *, SubExprUseData> SubExpressionUseData;
785 // UniqueSubExprs - Keep track of all of the subexpressions we see in the
786 // order we see them.
787 SmallVector<const SCEV *, 16> UniqueSubExprs;
789 SmallVector<const SCEV *, 16> SubExprs;
790 unsigned NumUsesInsideLoop = 0;
791 for (unsigned i = 0; i != NumUses; ++i) {
792 // If the user is outside the loop, just ignore it for base computation.
793 // Since the user is outside the loop, it must be *after* the loop (if it
794 // were before, it could not be based on the loop IV). We don't want users
795 // after the loop to affect base computation of values *inside* the loop,
796 // because we can always add their offsets to the result IV after the loop
797 // is done, ensuring we get good code inside the loop.
798 if (!L->contains(Uses[i].Inst->getParent()))
802 // If the base is zero (which is common), return zero now, there are no
804 if (Uses[i].Base == Zero) return Zero;
806 // If this use is as an address we may be able to put CSEs in the addressing
807 // mode rather than hoisting them.
808 bool isAddrUse = isAddressUse(Uses[i].Inst, Uses[i].OperandValToReplace);
809 // We may need the AccessTy below, but only when isAddrUse, so compute it
810 // only in that case.
811 const Type *AccessTy = 0;
813 AccessTy = getAccessType(Uses[i].Inst);
815 // Split the expression into subexprs.
816 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
817 // Add one to SubExpressionUseData.Count for each subexpr present, and
818 // if the subexpr is not a valid immediate within an addressing mode use,
819 // set SubExpressionUseData.notAllUsesAreFree. We definitely want to
820 // hoist these out of the loop (if they are common to all uses).
821 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) {
822 if (++SubExpressionUseData[SubExprs[j]].Count == 1)
823 UniqueSubExprs.push_back(SubExprs[j]);
824 if (!isAddrUse || !fitsInAddressMode(SubExprs[j], AccessTy, TLI, false))
825 SubExpressionUseData[SubExprs[j]].notAllUsesAreFree = true;
830 // Now that we know how many times each is used, build Result. Iterate over
831 // UniqueSubexprs so that we have a stable ordering.
832 for (unsigned i = 0, e = UniqueSubExprs.size(); i != e; ++i) {
833 std::map<const SCEV *, SubExprUseData>::iterator I =
834 SubExpressionUseData.find(UniqueSubExprs[i]);
835 assert(I != SubExpressionUseData.end() && "Entry not found?");
836 if (I->second.Count == NumUsesInsideLoop) { // Found CSE!
837 if (I->second.notAllUsesAreFree)
838 Result = SE->getAddExpr(Result, I->first);
840 FreeResult = SE->getAddExpr(FreeResult, I->first);
842 // Remove non-cse's from SubExpressionUseData.
843 SubExpressionUseData.erase(I);
846 if (FreeResult != Zero) {
847 // We have some subexpressions that can be subsumed into addressing
848 // modes in every use inside the loop. However, it's possible that
849 // there are so many of them that the combined FreeResult cannot
850 // be subsumed, or that the target cannot handle both a FreeResult
851 // and a Result in the same instruction (for example because it would
852 // require too many registers). Check this.
853 for (unsigned i=0; i<NumUses; ++i) {
854 if (!L->contains(Uses[i].Inst->getParent()))
856 // We know this is an addressing mode use; if there are any uses that
857 // are not, FreeResult would be Zero.
858 const Type *AccessTy = getAccessType(Uses[i].Inst);
859 if (!fitsInAddressMode(FreeResult, AccessTy, TLI, Result!=Zero)) {
860 // FIXME: could split up FreeResult into pieces here, some hoisted
861 // and some not. There is no obvious advantage to this.
862 Result = SE->getAddExpr(Result, FreeResult);
869 // If we found no CSE's, return now.
870 if (Result == Zero) return Result;
872 // If we still have a FreeResult, remove its subexpressions from
873 // SubExpressionUseData. This means they will remain in the use Bases.
874 if (FreeResult != Zero) {
875 SeparateSubExprs(SubExprs, FreeResult, SE);
876 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) {
877 std::map<const SCEV *, SubExprUseData>::iterator I =
878 SubExpressionUseData.find(SubExprs[j]);
879 SubExpressionUseData.erase(I);
884 // Otherwise, remove all of the CSE's we found from each of the base values.
885 for (unsigned i = 0; i != NumUses; ++i) {
886 // Uses outside the loop don't necessarily include the common base, but
887 // the final IV value coming into those uses does. Instead of trying to
888 // remove the pieces of the common base, which might not be there,
889 // subtract off the base to compensate for this.
890 if (!L->contains(Uses[i].Inst->getParent())) {
891 Uses[i].Base = SE->getMinusSCEV(Uses[i].Base, Result);
895 // Split the expression into subexprs.
896 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
898 // Remove any common subexpressions.
899 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
900 if (SubExpressionUseData.count(SubExprs[j])) {
901 SubExprs.erase(SubExprs.begin()+j);
905 // Finally, add the non-shared expressions together.
906 if (SubExprs.empty())
909 Uses[i].Base = SE->getAddExpr(SubExprs);
916 /// ValidScale - Check whether the given Scale is valid for all loads and
917 /// stores in UsersToProcess.
919 bool LoopStrengthReduce::ValidScale(bool HasBaseReg, int64_t Scale,
920 const std::vector<BasedUser>& UsersToProcess) {
924 for (unsigned i = 0, e = UsersToProcess.size(); i!=e; ++i) {
925 // If this is a load or other access, pass the type of the access in.
926 const Type *AccessTy =
927 Type::getVoidTy(UsersToProcess[i].Inst->getContext());
928 if (isAddressUse(UsersToProcess[i].Inst,
929 UsersToProcess[i].OperandValToReplace))
930 AccessTy = getAccessType(UsersToProcess[i].Inst);
931 else if (isa<PHINode>(UsersToProcess[i].Inst))
934 TargetLowering::AddrMode AM;
935 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
936 AM.BaseOffs = SC->getValue()->getSExtValue();
937 AM.HasBaseReg = HasBaseReg || !UsersToProcess[i].Base->isZero();
940 // If load[imm+r*scale] is illegal, bail out.
941 if (!TLI->isLegalAddressingMode(AM, AccessTy))
947 /// ValidOffset - Check whether the given Offset is valid for all loads and
948 /// stores in UsersToProcess.
950 bool LoopStrengthReduce::ValidOffset(bool HasBaseReg,
953 const std::vector<BasedUser>& UsersToProcess) {
957 for (unsigned i=0, e = UsersToProcess.size(); i!=e; ++i) {
958 // If this is a load or other access, pass the type of the access in.
959 const Type *AccessTy =
960 Type::getVoidTy(UsersToProcess[i].Inst->getContext());
961 if (isAddressUse(UsersToProcess[i].Inst,
962 UsersToProcess[i].OperandValToReplace))
963 AccessTy = getAccessType(UsersToProcess[i].Inst);
964 else if (isa<PHINode>(UsersToProcess[i].Inst))
967 TargetLowering::AddrMode AM;
968 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
969 AM.BaseOffs = SC->getValue()->getSExtValue();
970 AM.BaseOffs = (uint64_t)AM.BaseOffs + (uint64_t)Offset;
971 AM.HasBaseReg = HasBaseReg || !UsersToProcess[i].Base->isZero();
974 // If load[imm+r*scale] is illegal, bail out.
975 if (!TLI->isLegalAddressingMode(AM, AccessTy))
981 /// RequiresTypeConversion - Returns true if converting Ty1 to Ty2 is not
983 bool LoopStrengthReduce::RequiresTypeConversion(const Type *Ty1,
987 Ty1 = SE->getEffectiveSCEVType(Ty1);
988 Ty2 = SE->getEffectiveSCEVType(Ty2);
991 if (Ty1->canLosslesslyBitCastTo(Ty2))
993 if (TLI && TLI->isTruncateFree(Ty1, Ty2))
998 /// CheckForIVReuse - Returns the multiple if the stride is the multiple
999 /// of a previous stride and it is a legal value for the target addressing
1000 /// mode scale component and optional base reg. This allows the users of
1001 /// this stride to be rewritten as prev iv * factor. It returns 0 if no
1002 /// reuse is possible. Factors can be negative on same targets, e.g. ARM.
1004 /// If all uses are outside the loop, we don't require that all multiplies
1005 /// be folded into the addressing mode, nor even that the factor be constant;
1006 /// a multiply (executed once) outside the loop is better than another IV
1007 /// within. Well, usually.
1008 const SCEV *LoopStrengthReduce::CheckForIVReuse(bool HasBaseReg,
1009 bool AllUsesAreAddresses,
1010 bool AllUsesAreOutsideLoop,
1011 const SCEV *const &Stride,
1012 IVExpr &IV, const Type *Ty,
1013 const std::vector<BasedUser>& UsersToProcess) {
1014 if (StrideNoReuse.count(Stride))
1015 return SE->getIntegerSCEV(0, Stride->getType());
1017 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Stride)) {
1018 int64_t SInt = SC->getValue()->getSExtValue();
1019 for (unsigned NewStride = 0, e = IU->StrideOrder.size();
1020 NewStride != e; ++NewStride) {
1021 std::map<const SCEV *, IVsOfOneStride>::iterator SI =
1022 IVsByStride.find(IU->StrideOrder[NewStride]);
1023 if (SI == IVsByStride.end() || !isa<SCEVConstant>(SI->first) ||
1024 StrideNoReuse.count(SI->first))
1026 // The other stride has no uses, don't reuse it.
1027 std::map<const SCEV *, IVUsersOfOneStride *>::iterator UI =
1028 IU->IVUsesByStride.find(IU->StrideOrder[NewStride]);
1029 if (UI->second->Users.empty())
1031 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1032 if (SI->first != Stride &&
1033 (unsigned(abs64(SInt)) < SSInt || (SInt % SSInt) != 0))
1035 int64_t Scale = SInt / SSInt;
1036 // Check that this stride is valid for all the types used for loads and
1037 // stores; if it can be used for some and not others, we might as well use
1038 // the original stride everywhere, since we have to create the IV for it
1039 // anyway. If the scale is 1, then we don't need to worry about folding
1042 (AllUsesAreAddresses &&
1043 ValidScale(HasBaseReg, Scale, UsersToProcess))) {
1044 // Prefer to reuse an IV with a base of zero.
1045 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1046 IE = SI->second.IVs.end(); II != IE; ++II)
1047 // Only reuse previous IV if it would not require a type conversion
1048 // and if the base difference can be folded.
1049 if (II->Base->isZero() &&
1050 !RequiresTypeConversion(II->Base->getType(), Ty)) {
1052 return SE->getIntegerSCEV(Scale, Stride->getType());
1054 // Otherwise, settle for an IV with a foldable base.
1055 if (AllUsesAreAddresses)
1056 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1057 IE = SI->second.IVs.end(); II != IE; ++II)
1058 // Only reuse previous IV if it would not require a type conversion
1059 // and if the base difference can be folded.
1060 if (SE->getEffectiveSCEVType(II->Base->getType()) ==
1061 SE->getEffectiveSCEVType(Ty) &&
1062 isa<SCEVConstant>(II->Base)) {
1064 cast<SCEVConstant>(II->Base)->getValue()->getSExtValue();
1065 if (Base > INT32_MIN && Base <= INT32_MAX &&
1066 ValidOffset(HasBaseReg, -Base * Scale,
1067 Scale, UsersToProcess)) {
1069 return SE->getIntegerSCEV(Scale, Stride->getType());
1074 } else if (AllUsesAreOutsideLoop) {
1075 // Accept nonconstant strides here; it is really really right to substitute
1076 // an existing IV if we can.
1077 for (unsigned NewStride = 0, e = IU->StrideOrder.size();
1078 NewStride != e; ++NewStride) {
1079 std::map<const SCEV *, IVsOfOneStride>::iterator SI =
1080 IVsByStride.find(IU->StrideOrder[NewStride]);
1081 if (SI == IVsByStride.end() || !isa<SCEVConstant>(SI->first))
1083 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1084 if (SI->first != Stride && SSInt != 1)
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 a type conversion.
1090 if (!RequiresTypeConversion(II->Base->getType(), Ty)) {
1095 // Special case, old IV is -1*x and this one is x. Can treat this one as
1097 for (unsigned NewStride = 0, e = IU->StrideOrder.size();
1098 NewStride != e; ++NewStride) {
1099 std::map<const SCEV *, IVsOfOneStride>::iterator SI =
1100 IVsByStride.find(IU->StrideOrder[NewStride]);
1101 if (SI == IVsByStride.end())
1103 if (const SCEVMulExpr *ME = dyn_cast<SCEVMulExpr>(SI->first))
1104 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(ME->getOperand(0)))
1105 if (Stride == ME->getOperand(1) &&
1106 SC->getValue()->getSExtValue() == -1LL)
1107 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1108 IE = SI->second.IVs.end(); II != IE; ++II)
1109 // Accept nonzero base here.
1110 // Only reuse previous IV if it would not require type conversion.
1111 if (!RequiresTypeConversion(II->Base->getType(), Ty)) {
1113 return SE->getIntegerSCEV(-1LL, Stride->getType());
1117 return SE->getIntegerSCEV(0, Stride->getType());
1120 /// PartitionByIsUseOfPostIncrementedValue - Simple boolean predicate that
1121 /// returns true if Val's isUseOfPostIncrementedValue is true.
1122 static bool PartitionByIsUseOfPostIncrementedValue(const BasedUser &Val) {
1123 return Val.isUseOfPostIncrementedValue;
1126 /// isNonConstantNegative - Return true if the specified scev is negated, but
1128 static bool isNonConstantNegative(const SCEV *const &Expr) {
1129 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Expr);
1130 if (!Mul) return false;
1132 // If there is a constant factor, it will be first.
1133 const SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
1134 if (!SC) return false;
1136 // Return true if the value is negative, this matches things like (-42 * V).
1137 return SC->getValue()->getValue().isNegative();
1140 /// CollectIVUsers - Transform our list of users and offsets to a bit more
1141 /// complex table. In this new vector, each 'BasedUser' contains 'Base', the
1142 /// base of the strided accesses, as well as the old information from Uses. We
1143 /// progressively move information from the Base field to the Imm field, until
1144 /// we eventually have the full access expression to rewrite the use.
1145 const SCEV *LoopStrengthReduce::CollectIVUsers(const SCEV *const &Stride,
1146 IVUsersOfOneStride &Uses,
1148 bool &AllUsesAreAddresses,
1149 bool &AllUsesAreOutsideLoop,
1150 std::vector<BasedUser> &UsersToProcess) {
1151 // FIXME: Generalize to non-affine IV's.
1152 if (!Stride->isLoopInvariant(L))
1153 return SE->getIntegerSCEV(0, Stride->getType());
1155 UsersToProcess.reserve(Uses.Users.size());
1156 for (ilist<IVStrideUse>::iterator I = Uses.Users.begin(),
1157 E = Uses.Users.end(); I != E; ++I) {
1158 UsersToProcess.push_back(BasedUser(*I, SE));
1160 // Move any loop variant operands from the offset field to the immediate
1161 // field of the use, so that we don't try to use something before it is
1163 MoveLoopVariantsToImmediateField(UsersToProcess.back().Base,
1164 UsersToProcess.back().Imm, L, SE);
1165 assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
1166 "Base value is not loop invariant!");
1169 // We now have a whole bunch of uses of like-strided induction variables, but
1170 // they might all have different bases. We want to emit one PHI node for this
1171 // stride which we fold as many common expressions (between the IVs) into as
1172 // possible. Start by identifying the common expressions in the base values
1173 // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
1174 // "A+B"), emit it to the preheader, then remove the expression from the
1175 // UsersToProcess base values.
1176 const SCEV *CommonExprs =
1177 RemoveCommonExpressionsFromUseBases(UsersToProcess, SE, L, TLI);
1179 // Next, figure out what we can represent in the immediate fields of
1180 // instructions. If we can represent anything there, move it to the imm
1181 // fields of the BasedUsers. We do this so that it increases the commonality
1182 // of the remaining uses.
1183 unsigned NumPHI = 0;
1184 bool HasAddress = false;
1185 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1186 // If the user is not in the current loop, this means it is using the exit
1187 // value of the IV. Do not put anything in the base, make sure it's all in
1188 // the immediate field to allow as much factoring as possible.
1189 if (!L->contains(UsersToProcess[i].Inst->getParent())) {
1190 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm,
1191 UsersToProcess[i].Base);
1192 UsersToProcess[i].Base =
1193 SE->getIntegerSCEV(0, UsersToProcess[i].Base->getType());
1195 // Not all uses are outside the loop.
1196 AllUsesAreOutsideLoop = false;
1198 // Addressing modes can be folded into loads and stores. Be careful that
1199 // the store is through the expression, not of the expression though.
1201 bool isAddress = isAddressUse(UsersToProcess[i].Inst,
1202 UsersToProcess[i].OperandValToReplace);
1203 if (isa<PHINode>(UsersToProcess[i].Inst)) {
1211 // If this use isn't an address, then not all uses are addresses.
1212 if (!isAddress && !isPHI)
1213 AllUsesAreAddresses = false;
1215 MoveImmediateValues(TLI, UsersToProcess[i].Inst, UsersToProcess[i].Base,
1216 UsersToProcess[i].Imm, isAddress, L, SE);
1220 // If one of the use is a PHI node and all other uses are addresses, still
1221 // allow iv reuse. Essentially we are trading one constant multiplication
1222 // for one fewer iv.
1224 AllUsesAreAddresses = false;
1226 // There are no in-loop address uses.
1227 if (AllUsesAreAddresses && (!HasAddress && !AllUsesAreOutsideLoop))
1228 AllUsesAreAddresses = false;
1233 /// ShouldUseFullStrengthReductionMode - Test whether full strength-reduction
1234 /// is valid and profitable for the given set of users of a stride. In
1235 /// full strength-reduction mode, all addresses at the current stride are
1236 /// strength-reduced all the way down to pointer arithmetic.
1238 bool LoopStrengthReduce::ShouldUseFullStrengthReductionMode(
1239 const std::vector<BasedUser> &UsersToProcess,
1241 bool AllUsesAreAddresses,
1242 const SCEV *Stride) {
1243 if (!EnableFullLSRMode)
1246 // The heuristics below aim to avoid increasing register pressure, but
1247 // fully strength-reducing all the addresses increases the number of
1248 // add instructions, so don't do this when optimizing for size.
1249 // TODO: If the loop is large, the savings due to simpler addresses
1250 // may oughtweight the costs of the extra increment instructions.
1251 if (L->getHeader()->getParent()->hasFnAttr(Attribute::OptimizeForSize))
1254 // TODO: For now, don't do full strength reduction if there could
1255 // potentially be greater-stride multiples of the current stride
1256 // which could reuse the current stride IV.
1257 if (IU->StrideOrder.back() != Stride)
1260 // Iterate through the uses to find conditions that automatically rule out
1262 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ) {
1263 const SCEV *Base = UsersToProcess[i].Base;
1264 const SCEV *Imm = UsersToProcess[i].Imm;
1265 // If any users have a loop-variant component, they can't be fully
1266 // strength-reduced.
1267 if (Imm && !Imm->isLoopInvariant(L))
1269 // If there are to users with the same base and the difference between
1270 // the two Imm values can't be folded into the address, full
1271 // strength reduction would increase register pressure.
1273 const SCEV *CurImm = UsersToProcess[i].Imm;
1274 if ((CurImm || Imm) && CurImm != Imm) {
1275 if (!CurImm) CurImm = SE->getIntegerSCEV(0, Stride->getType());
1276 if (!Imm) Imm = SE->getIntegerSCEV(0, Stride->getType());
1277 const Instruction *Inst = UsersToProcess[i].Inst;
1278 const Type *AccessTy = getAccessType(Inst);
1279 const SCEV *Diff = SE->getMinusSCEV(UsersToProcess[i].Imm, Imm);
1280 if (!Diff->isZero() &&
1281 (!AllUsesAreAddresses ||
1282 !fitsInAddressMode(Diff, AccessTy, TLI, /*HasBaseReg=*/true)))
1285 } while (++i != e && Base == UsersToProcess[i].Base);
1288 // If there's exactly one user in this stride, fully strength-reducing it
1289 // won't increase register pressure. If it's starting from a non-zero base,
1290 // it'll be simpler this way.
1291 if (UsersToProcess.size() == 1 && !UsersToProcess[0].Base->isZero())
1294 // Otherwise, if there are any users in this stride that don't require
1295 // a register for their base, full strength-reduction will increase
1296 // register pressure.
1297 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1298 if (UsersToProcess[i].Base->isZero())
1301 // Otherwise, go for it.
1305 /// InsertAffinePhi Create and insert a PHI node for an induction variable
1306 /// with the specified start and step values in the specified loop.
1308 /// If NegateStride is true, the stride should be negated by using a
1309 /// subtract instead of an add.
1311 /// Return the created phi node.
1313 static PHINode *InsertAffinePhi(const SCEV *Start, const SCEV *Step,
1314 Instruction *IVIncInsertPt,
1316 SCEVExpander &Rewriter) {
1317 assert(Start->isLoopInvariant(L) && "New PHI start is not loop invariant!");
1318 assert(Step->isLoopInvariant(L) && "New PHI stride is not loop invariant!");
1320 BasicBlock *Header = L->getHeader();
1321 BasicBlock *Preheader = L->getLoopPreheader();
1322 BasicBlock *LatchBlock = L->getLoopLatch();
1323 const Type *Ty = Start->getType();
1324 Ty = Rewriter.SE.getEffectiveSCEVType(Ty);
1326 PHINode *PN = PHINode::Create(Ty, "lsr.iv", Header->begin());
1327 PN->addIncoming(Rewriter.expandCodeFor(Start, Ty, Preheader->getTerminator()),
1330 // If the stride is negative, insert a sub instead of an add for the
1332 bool isNegative = isNonConstantNegative(Step);
1333 const SCEV *IncAmount = Step;
1335 IncAmount = Rewriter.SE.getNegativeSCEV(Step);
1337 // Insert an add instruction right before the terminator corresponding
1338 // to the back-edge or just before the only use. The location is determined
1339 // by the caller and passed in as IVIncInsertPt.
1340 Value *StepV = Rewriter.expandCodeFor(IncAmount, Ty,
1341 Preheader->getTerminator());
1344 IncV = BinaryOperator::CreateSub(PN, StepV, "lsr.iv.next",
1347 IncV = BinaryOperator::CreateAdd(PN, StepV, "lsr.iv.next",
1350 if (!isa<ConstantInt>(StepV)) ++NumVariable;
1352 PN->addIncoming(IncV, LatchBlock);
1358 static void SortUsersToProcess(std::vector<BasedUser> &UsersToProcess) {
1359 // We want to emit code for users inside the loop first. To do this, we
1360 // rearrange BasedUser so that the entries at the end have
1361 // isUseOfPostIncrementedValue = false, because we pop off the end of the
1362 // vector (so we handle them first).
1363 std::partition(UsersToProcess.begin(), UsersToProcess.end(),
1364 PartitionByIsUseOfPostIncrementedValue);
1366 // Sort this by base, so that things with the same base are handled
1367 // together. By partitioning first and stable-sorting later, we are
1368 // guaranteed that within each base we will pop off users from within the
1369 // loop before users outside of the loop with a particular base.
1371 // We would like to use stable_sort here, but we can't. The problem is that
1372 // const SCEV *'s don't have a deterministic ordering w.r.t to each other, so
1373 // we don't have anything to do a '<' comparison on. Because we think the
1374 // number of uses is small, do a horrible bubble sort which just relies on
1376 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1377 // Get a base value.
1378 const SCEV *Base = UsersToProcess[i].Base;
1380 // Compact everything with this base to be consecutive with this one.
1381 for (unsigned j = i+1; j != e; ++j) {
1382 if (UsersToProcess[j].Base == Base) {
1383 std::swap(UsersToProcess[i+1], UsersToProcess[j]);
1390 /// PrepareToStrengthReduceFully - Prepare to fully strength-reduce
1391 /// UsersToProcess, meaning lowering addresses all the way down to direct
1392 /// pointer arithmetic.
1395 LoopStrengthReduce::PrepareToStrengthReduceFully(
1396 std::vector<BasedUser> &UsersToProcess,
1398 const SCEV *CommonExprs,
1400 SCEVExpander &PreheaderRewriter) {
1401 DEBUG(errs() << " Fully reducing all users\n");
1403 // Rewrite the UsersToProcess records, creating a separate PHI for each
1404 // unique Base value.
1405 Instruction *IVIncInsertPt = L->getLoopLatch()->getTerminator();
1406 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ) {
1407 // TODO: The uses are grouped by base, but not sorted. We arbitrarily
1408 // pick the first Imm value here to start with, and adjust it for the
1410 const SCEV *Imm = UsersToProcess[i].Imm;
1411 const SCEV *Base = UsersToProcess[i].Base;
1412 const SCEV *Start = SE->getAddExpr(CommonExprs, Base, Imm);
1413 PHINode *Phi = InsertAffinePhi(Start, Stride, IVIncInsertPt, L,
1415 // Loop over all the users with the same base.
1417 UsersToProcess[i].Base = SE->getIntegerSCEV(0, Stride->getType());
1418 UsersToProcess[i].Imm = SE->getMinusSCEV(UsersToProcess[i].Imm, Imm);
1419 UsersToProcess[i].Phi = Phi;
1420 assert(UsersToProcess[i].Imm->isLoopInvariant(L) &&
1421 "ShouldUseFullStrengthReductionMode should reject this!");
1422 } while (++i != e && Base == UsersToProcess[i].Base);
1426 /// FindIVIncInsertPt - Return the location to insert the increment instruction.
1427 /// If the only use if a use of postinc value, (must be the loop termination
1428 /// condition), then insert it just before the use.
1429 static Instruction *FindIVIncInsertPt(std::vector<BasedUser> &UsersToProcess,
1431 if (UsersToProcess.size() == 1 &&
1432 UsersToProcess[0].isUseOfPostIncrementedValue &&
1433 L->contains(UsersToProcess[0].Inst->getParent()))
1434 return UsersToProcess[0].Inst;
1435 return L->getLoopLatch()->getTerminator();
1438 /// PrepareToStrengthReduceWithNewPhi - Insert a new induction variable for the
1439 /// given users to share.
1442 LoopStrengthReduce::PrepareToStrengthReduceWithNewPhi(
1443 std::vector<BasedUser> &UsersToProcess,
1445 const SCEV *CommonExprs,
1447 Instruction *IVIncInsertPt,
1449 SCEVExpander &PreheaderRewriter) {
1450 DEBUG(errs() << " Inserting new PHI:\n");
1452 PHINode *Phi = InsertAffinePhi(SE->getUnknown(CommonBaseV),
1453 Stride, IVIncInsertPt, L,
1456 // Remember this in case a later stride is multiple of this.
1457 IVsByStride[Stride].addIV(Stride, CommonExprs, Phi);
1459 // All the users will share this new IV.
1460 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1461 UsersToProcess[i].Phi = Phi;
1463 DEBUG(errs() << " IV=");
1464 DEBUG(WriteAsOperand(errs(), Phi, /*PrintType=*/false));
1465 DEBUG(errs() << "\n");
1468 /// PrepareToStrengthReduceFromSmallerStride - Prepare for the given users to
1469 /// reuse an induction variable with a stride that is a factor of the current
1470 /// induction variable.
1473 LoopStrengthReduce::PrepareToStrengthReduceFromSmallerStride(
1474 std::vector<BasedUser> &UsersToProcess,
1476 const IVExpr &ReuseIV,
1477 Instruction *PreInsertPt) {
1478 DEBUG(errs() << " Rewriting in terms of existing IV of STRIDE "
1479 << *ReuseIV.Stride << " and BASE " << *ReuseIV.Base << "\n");
1481 // All the users will share the reused IV.
1482 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1483 UsersToProcess[i].Phi = ReuseIV.PHI;
1485 Constant *C = dyn_cast<Constant>(CommonBaseV);
1487 (!C->isNullValue() &&
1488 !fitsInAddressMode(SE->getUnknown(CommonBaseV), CommonBaseV->getType(),
1490 // We want the common base emitted into the preheader! This is just
1491 // using cast as a copy so BitCast (no-op cast) is appropriate
1492 CommonBaseV = new BitCastInst(CommonBaseV, CommonBaseV->getType(),
1493 "commonbase", PreInsertPt);
1496 static bool IsImmFoldedIntoAddrMode(GlobalValue *GV, int64_t Offset,
1497 const Type *AccessTy,
1498 std::vector<BasedUser> &UsersToProcess,
1499 const TargetLowering *TLI) {
1500 SmallVector<Instruction*, 16> AddrModeInsts;
1501 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1502 if (UsersToProcess[i].isUseOfPostIncrementedValue)
1504 ExtAddrMode AddrMode =
1505 AddressingModeMatcher::Match(UsersToProcess[i].OperandValToReplace,
1506 AccessTy, UsersToProcess[i].Inst,
1507 AddrModeInsts, *TLI);
1508 if (GV && GV != AddrMode.BaseGV)
1510 if (Offset && !AddrMode.BaseOffs)
1511 // FIXME: How to accurate check it's immediate offset is folded.
1513 AddrModeInsts.clear();
1518 /// StrengthReduceIVUsersOfStride - Strength reduce all of the users of a single
1519 /// stride of IV. All of the users may have different starting values, and this
1520 /// may not be the only stride.
1522 LoopStrengthReduce::StrengthReduceIVUsersOfStride(const SCEV *const &Stride,
1523 IVUsersOfOneStride &Uses,
1525 // If all the users are moved to another stride, then there is nothing to do.
1526 if (Uses.Users.empty())
1529 // Keep track if every use in UsersToProcess is an address. If they all are,
1530 // we may be able to rewrite the entire collection of them in terms of a
1531 // smaller-stride IV.
1532 bool AllUsesAreAddresses = true;
1534 // Keep track if every use of a single stride is outside the loop. If so,
1535 // we want to be more aggressive about reusing a smaller-stride IV; a
1536 // multiply outside the loop is better than another IV inside. Well, usually.
1537 bool AllUsesAreOutsideLoop = true;
1539 // Transform our list of users and offsets to a bit more complex table. In
1540 // this new vector, each 'BasedUser' contains 'Base' the base of the
1541 // strided accessas well as the old information from Uses. We progressively
1542 // move information from the Base field to the Imm field, until we eventually
1543 // have the full access expression to rewrite the use.
1544 std::vector<BasedUser> UsersToProcess;
1545 const SCEV *CommonExprs = CollectIVUsers(Stride, Uses, L, AllUsesAreAddresses,
1546 AllUsesAreOutsideLoop,
1549 // Sort the UsersToProcess array so that users with common bases are
1550 // next to each other.
1551 SortUsersToProcess(UsersToProcess);
1553 // If we managed to find some expressions in common, we'll need to carry
1554 // their value in a register and add it in for each use. This will take up
1555 // a register operand, which potentially restricts what stride values are
1557 bool HaveCommonExprs = !CommonExprs->isZero();
1558 const Type *ReplacedTy = CommonExprs->getType();
1560 // If all uses are addresses, consider sinking the immediate part of the
1561 // common expression back into uses if they can fit in the immediate fields.
1562 if (TLI && HaveCommonExprs && AllUsesAreAddresses) {
1563 const SCEV *NewCommon = CommonExprs;
1564 const SCEV *Imm = SE->getIntegerSCEV(0, ReplacedTy);
1565 MoveImmediateValues(TLI, Type::getVoidTy(
1566 L->getLoopPreheader()->getContext()),
1567 NewCommon, Imm, true, L, SE);
1568 if (!Imm->isZero()) {
1571 // If the immediate part of the common expression is a GV, check if it's
1572 // possible to fold it into the target addressing mode.
1573 GlobalValue *GV = 0;
1574 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(Imm))
1575 GV = dyn_cast<GlobalValue>(SU->getValue());
1577 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Imm))
1578 Offset = SC->getValue()->getSExtValue();
1580 // Pass VoidTy as the AccessTy to be conservative, because
1581 // there could be multiple access types among all the uses.
1582 DoSink = IsImmFoldedIntoAddrMode(GV, Offset,
1583 Type::getVoidTy(L->getLoopPreheader()->getContext()),
1584 UsersToProcess, TLI);
1587 DEBUG(errs() << " Sinking " << *Imm << " back down into uses\n");
1588 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1589 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm, Imm);
1590 CommonExprs = NewCommon;
1591 HaveCommonExprs = !CommonExprs->isZero();
1597 // Now that we know what we need to do, insert the PHI node itself.
1599 DEBUG(errs() << "LSR: Examining IVs of TYPE " << *ReplacedTy << " of STRIDE "
1601 << " Common base: " << *CommonExprs << "\n");
1603 SCEVExpander Rewriter(*SE);
1604 SCEVExpander PreheaderRewriter(*SE);
1606 BasicBlock *Preheader = L->getLoopPreheader();
1607 Instruction *PreInsertPt = Preheader->getTerminator();
1608 BasicBlock *LatchBlock = L->getLoopLatch();
1609 Instruction *IVIncInsertPt = LatchBlock->getTerminator();
1611 Value *CommonBaseV = Constant::getNullValue(ReplacedTy);
1613 const SCEV *RewriteFactor = SE->getIntegerSCEV(0, ReplacedTy);
1614 IVExpr ReuseIV(SE->getIntegerSCEV(0,
1615 Type::getInt32Ty(Preheader->getContext())),
1616 SE->getIntegerSCEV(0,
1617 Type::getInt32Ty(Preheader->getContext())),
1620 // Choose a strength-reduction strategy and prepare for it by creating
1621 // the necessary PHIs and adjusting the bookkeeping.
1622 if (ShouldUseFullStrengthReductionMode(UsersToProcess, L,
1623 AllUsesAreAddresses, Stride)) {
1624 PrepareToStrengthReduceFully(UsersToProcess, Stride, CommonExprs, L,
1627 // Emit the initial base value into the loop preheader.
1628 CommonBaseV = PreheaderRewriter.expandCodeFor(CommonExprs, ReplacedTy,
1631 // If all uses are addresses, check if it is possible to reuse an IV. The
1632 // new IV must have a stride that is a multiple of the old stride; the
1633 // multiple must be a number that can be encoded in the scale field of the
1634 // target addressing mode; and we must have a valid instruction after this
1635 // substitution, including the immediate field, if any.
1636 RewriteFactor = CheckForIVReuse(HaveCommonExprs, AllUsesAreAddresses,
1637 AllUsesAreOutsideLoop,
1638 Stride, ReuseIV, ReplacedTy,
1640 if (!RewriteFactor->isZero())
1641 PrepareToStrengthReduceFromSmallerStride(UsersToProcess, CommonBaseV,
1642 ReuseIV, PreInsertPt);
1644 IVIncInsertPt = FindIVIncInsertPt(UsersToProcess, L);
1645 PrepareToStrengthReduceWithNewPhi(UsersToProcess, Stride, CommonExprs,
1646 CommonBaseV, IVIncInsertPt,
1647 L, PreheaderRewriter);
1651 // Process all the users now, replacing their strided uses with
1652 // strength-reduced forms. This outer loop handles all bases, the inner
1653 // loop handles all users of a particular base.
1654 while (!UsersToProcess.empty()) {
1655 const SCEV *Base = UsersToProcess.back().Base;
1656 Instruction *Inst = UsersToProcess.back().Inst;
1658 // Emit the code for Base into the preheader.
1660 if (!Base->isZero()) {
1661 BaseV = PreheaderRewriter.expandCodeFor(Base, 0, PreInsertPt);
1663 DEBUG(errs() << " INSERTING code for BASE = " << *Base << ":");
1664 if (BaseV->hasName())
1665 DEBUG(errs() << " Result value name = %" << BaseV->getName());
1666 DEBUG(errs() << "\n");
1668 // If BaseV is a non-zero constant, make sure that it gets inserted into
1669 // the preheader, instead of being forward substituted into the uses. We
1670 // do this by forcing a BitCast (noop cast) to be inserted into the
1671 // preheader in this case.
1672 if (!fitsInAddressMode(Base, getAccessType(Inst), TLI, false) &&
1673 isa<Constant>(BaseV)) {
1674 // We want this constant emitted into the preheader! This is just
1675 // using cast as a copy so BitCast (no-op cast) is appropriate
1676 BaseV = new BitCastInst(BaseV, BaseV->getType(), "preheaderinsert",
1681 // Emit the code to add the immediate offset to the Phi value, just before
1682 // the instructions that we identified as using this stride and base.
1684 // FIXME: Use emitted users to emit other users.
1685 BasedUser &User = UsersToProcess.back();
1687 DEBUG(errs() << " Examining ");
1688 if (User.isUseOfPostIncrementedValue)
1689 DEBUG(errs() << "postinc");
1691 DEBUG(errs() << "preinc");
1692 DEBUG(errs() << " use ");
1693 DEBUG(WriteAsOperand(errs(), UsersToProcess.back().OperandValToReplace,
1694 /*PrintType=*/false));
1695 DEBUG(errs() << " in Inst: " << *User.Inst);
1697 // If this instruction wants to use the post-incremented value, move it
1698 // after the post-inc and use its value instead of the PHI.
1699 Value *RewriteOp = User.Phi;
1700 if (User.isUseOfPostIncrementedValue) {
1701 RewriteOp = User.Phi->getIncomingValueForBlock(LatchBlock);
1702 // If this user is in the loop, make sure it is the last thing in the
1703 // loop to ensure it is dominated by the increment. In case it's the
1704 // only use of the iv, the increment instruction is already before the
1706 if (L->contains(User.Inst->getParent()) && User.Inst != IVIncInsertPt)
1707 User.Inst->moveBefore(IVIncInsertPt);
1710 const SCEV *RewriteExpr = SE->getUnknown(RewriteOp);
1712 if (SE->getEffectiveSCEVType(RewriteOp->getType()) !=
1713 SE->getEffectiveSCEVType(ReplacedTy)) {
1714 assert(SE->getTypeSizeInBits(RewriteOp->getType()) >
1715 SE->getTypeSizeInBits(ReplacedTy) &&
1716 "Unexpected widening cast!");
1717 RewriteExpr = SE->getTruncateExpr(RewriteExpr, ReplacedTy);
1720 // If we had to insert new instructions for RewriteOp, we have to
1721 // consider that they may not have been able to end up immediately
1722 // next to RewriteOp, because non-PHI instructions may never precede
1723 // PHI instructions in a block. In this case, remember where the last
1724 // instruction was inserted so that if we're replacing a different
1725 // PHI node, we can use the later point to expand the final
1727 Instruction *NewBasePt = dyn_cast<Instruction>(RewriteOp);
1728 if (RewriteOp == User.Phi) NewBasePt = 0;
1730 // Clear the SCEVExpander's expression map so that we are guaranteed
1731 // to have the code emitted where we expect it.
1734 // If we are reusing the iv, then it must be multiplied by a constant
1735 // factor to take advantage of the addressing mode scale component.
1736 if (!RewriteFactor->isZero()) {
1737 // If we're reusing an IV with a nonzero base (currently this happens
1738 // only when all reuses are outside the loop) subtract that base here.
1739 // The base has been used to initialize the PHI node but we don't want
1741 if (!ReuseIV.Base->isZero()) {
1742 const SCEV *typedBase = ReuseIV.Base;
1743 if (SE->getEffectiveSCEVType(RewriteExpr->getType()) !=
1744 SE->getEffectiveSCEVType(ReuseIV.Base->getType())) {
1745 // It's possible the original IV is a larger type than the new IV,
1746 // in which case we have to truncate the Base. We checked in
1747 // RequiresTypeConversion that this is valid.
1748 assert(SE->getTypeSizeInBits(RewriteExpr->getType()) <
1749 SE->getTypeSizeInBits(ReuseIV.Base->getType()) &&
1750 "Unexpected lengthening conversion!");
1751 typedBase = SE->getTruncateExpr(ReuseIV.Base,
1752 RewriteExpr->getType());
1754 RewriteExpr = SE->getMinusSCEV(RewriteExpr, typedBase);
1757 // Multiply old variable, with base removed, by new scale factor.
1758 RewriteExpr = SE->getMulExpr(RewriteFactor,
1761 // The common base is emitted in the loop preheader. But since we
1762 // are reusing an IV, it has not been used to initialize the PHI node.
1763 // Add it to the expression used to rewrite the uses.
1764 // When this use is outside the loop, we earlier subtracted the
1765 // common base, and are adding it back here. Use the same expression
1766 // as before, rather than CommonBaseV, so DAGCombiner will zap it.
1767 if (!CommonExprs->isZero()) {
1768 if (L->contains(User.Inst->getParent()))
1769 RewriteExpr = SE->getAddExpr(RewriteExpr,
1770 SE->getUnknown(CommonBaseV));
1772 RewriteExpr = SE->getAddExpr(RewriteExpr, CommonExprs);
1776 // Now that we know what we need to do, insert code before User for the
1777 // immediate and any loop-variant expressions.
1779 // Add BaseV to the PHI value if needed.
1780 RewriteExpr = SE->getAddExpr(RewriteExpr, SE->getUnknown(BaseV));
1782 User.RewriteInstructionToUseNewBase(RewriteExpr, NewBasePt,
1783 Rewriter, L, this, *LI,
1786 // Mark old value we replaced as possibly dead, so that it is eliminated
1787 // if we just replaced the last use of that value.
1788 DeadInsts.push_back(User.OperandValToReplace);
1790 UsersToProcess.pop_back();
1793 // If there are any more users to process with the same base, process them
1794 // now. We sorted by base above, so we just have to check the last elt.
1795 } while (!UsersToProcess.empty() && UsersToProcess.back().Base == Base);
1796 // TODO: Next, find out which base index is the most common, pull it out.
1799 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
1800 // different starting values, into different PHIs.
1803 void LoopStrengthReduce::StrengthReduceIVUsers(Loop *L) {
1804 // Note: this processes each stride/type pair individually. All users
1805 // passed into StrengthReduceIVUsersOfStride have the same type AND stride.
1806 // Also, note that we iterate over IVUsesByStride indirectly by using
1807 // StrideOrder. This extra layer of indirection makes the ordering of
1808 // strides deterministic - not dependent on map order.
1809 for (unsigned Stride = 0, e = IU->StrideOrder.size(); Stride != e; ++Stride) {
1810 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
1811 IU->IVUsesByStride.find(IU->StrideOrder[Stride]);
1812 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
1813 // FIXME: Generalize to non-affine IV's.
1814 if (!SI->first->isLoopInvariant(L))
1816 StrengthReduceIVUsersOfStride(SI->first, *SI->second, L);
1820 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1821 /// set the IV user and stride information and return true, otherwise return
1823 bool LoopStrengthReduce::FindIVUserForCond(ICmpInst *Cond,
1824 IVStrideUse *&CondUse,
1825 const SCEV* &CondStride) {
1826 for (unsigned Stride = 0, e = IU->StrideOrder.size();
1827 Stride != e && !CondUse; ++Stride) {
1828 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
1829 IU->IVUsesByStride.find(IU->StrideOrder[Stride]);
1830 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
1832 for (ilist<IVStrideUse>::iterator UI = SI->second->Users.begin(),
1833 E = SI->second->Users.end(); UI != E; ++UI)
1834 if (UI->getUser() == Cond) {
1835 // NOTE: we could handle setcc instructions with multiple uses here, but
1836 // InstCombine does it as well for simple uses, it's not clear that it
1837 // occurs enough in real life to handle.
1839 CondStride = SI->first;
1847 // Constant strides come first which in turns are sorted by their absolute
1848 // values. If absolute values are the same, then positive strides comes first.
1850 // 4, -1, X, 1, 2 ==> 1, -1, 2, 4, X
1851 struct StrideCompare {
1852 const ScalarEvolution *SE;
1853 explicit StrideCompare(const ScalarEvolution *se) : SE(se) {}
1855 bool operator()(const SCEV *const &LHS, const SCEV *const &RHS) {
1856 const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS);
1857 const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS);
1859 int64_t LV = LHSC->getValue()->getSExtValue();
1860 int64_t RV = RHSC->getValue()->getSExtValue();
1861 uint64_t ALV = (LV < 0) ? -LV : LV;
1862 uint64_t ARV = (RV < 0) ? -RV : RV;
1870 // If it's the same value but different type, sort by bit width so
1871 // that we emit larger induction variables before smaller
1872 // ones, letting the smaller be re-written in terms of larger ones.
1873 return SE->getTypeSizeInBits(RHS->getType()) <
1874 SE->getTypeSizeInBits(LHS->getType());
1876 return LHSC && !RHSC;
1881 /// ChangeCompareStride - If a loop termination compare instruction is the
1882 /// only use of its stride, and the compaison is against a constant value,
1883 /// try eliminate the stride by moving the compare instruction to another
1884 /// stride and change its constant operand accordingly. e.g.
1890 /// if (v2 < 10) goto loop
1895 /// if (v1 < 30) goto loop
1896 ICmpInst *LoopStrengthReduce::ChangeCompareStride(Loop *L, ICmpInst *Cond,
1897 IVStrideUse* &CondUse,
1898 const SCEV* &CondStride,
1900 // If there's only one stride in the loop, there's nothing to do here.
1901 if (IU->StrideOrder.size() < 2)
1903 // If there are other users of the condition's stride, don't bother
1904 // trying to change the condition because the stride will still
1906 std::map<const SCEV *, IVUsersOfOneStride *>::iterator I =
1907 IU->IVUsesByStride.find(CondStride);
1908 if (I == IU->IVUsesByStride.end())
1910 if (I->second->Users.size() > 1) {
1911 for (ilist<IVStrideUse>::iterator II = I->second->Users.begin(),
1912 EE = I->second->Users.end(); II != EE; ++II) {
1913 if (II->getUser() == Cond)
1915 if (!isInstructionTriviallyDead(II->getUser()))
1919 // Only handle constant strides for now.
1920 const SCEVConstant *SC = dyn_cast<SCEVConstant>(CondStride);
1921 if (!SC) return Cond;
1923 ICmpInst::Predicate Predicate = Cond->getPredicate();
1924 int64_t CmpSSInt = SC->getValue()->getSExtValue();
1925 unsigned BitWidth = SE->getTypeSizeInBits(CondStride->getType());
1926 uint64_t SignBit = 1ULL << (BitWidth-1);
1927 const Type *CmpTy = Cond->getOperand(0)->getType();
1928 const Type *NewCmpTy = NULL;
1929 unsigned TyBits = SE->getTypeSizeInBits(CmpTy);
1930 unsigned NewTyBits = 0;
1931 const SCEV *NewStride = NULL;
1932 Value *NewCmpLHS = NULL;
1933 Value *NewCmpRHS = NULL;
1935 const SCEV *NewOffset = SE->getIntegerSCEV(0, CmpTy);
1937 if (ConstantInt *C = dyn_cast<ConstantInt>(Cond->getOperand(1))) {
1938 int64_t CmpVal = C->getValue().getSExtValue();
1940 // Check the relevant induction variable for conformance to
1942 const SCEV *IV = SE->getSCEV(Cond->getOperand(0));
1943 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1944 if (!AR || !AR->isAffine())
1947 const SCEVConstant *StartC = dyn_cast<SCEVConstant>(AR->getStart());
1948 // Check stride constant and the comparision constant signs to detect
1951 if ((StartC->getValue()->getSExtValue() < CmpVal && CmpSSInt < 0) ||
1952 (StartC->getValue()->getSExtValue() > CmpVal && CmpSSInt > 0))
1955 // More restrictive check for the other cases.
1956 if ((CmpVal & SignBit) != (CmpSSInt & SignBit))
1960 // Look for a suitable stride / iv as replacement.
1961 for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
1962 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
1963 IU->IVUsesByStride.find(IU->StrideOrder[i]);
1964 if (!isa<SCEVConstant>(SI->first) || SI->second->Users.empty())
1966 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1967 if (SSInt == CmpSSInt ||
1968 abs64(SSInt) < abs64(CmpSSInt) ||
1969 (SSInt % CmpSSInt) != 0)
1972 Scale = SSInt / CmpSSInt;
1973 int64_t NewCmpVal = CmpVal * Scale;
1975 // If old icmp value fits in icmp immediate field, but the new one doesn't
1976 // try something else.
1978 TLI->isLegalICmpImmediate(CmpVal) &&
1979 !TLI->isLegalICmpImmediate(NewCmpVal))
1982 APInt Mul = APInt(BitWidth*2, CmpVal, true);
1983 Mul = Mul * APInt(BitWidth*2, Scale, true);
1984 // Check for overflow.
1985 if (!Mul.isSignedIntN(BitWidth))
1987 // Check for overflow in the stride's type too.
1988 if (!Mul.isSignedIntN(SE->getTypeSizeInBits(SI->first->getType())))
1991 // Watch out for overflow.
1992 if (ICmpInst::isSigned(Predicate) &&
1993 (CmpVal & SignBit) != (NewCmpVal & SignBit))
1996 // Pick the best iv to use trying to avoid a cast.
1998 for (ilist<IVStrideUse>::iterator UI = SI->second->Users.begin(),
1999 E = SI->second->Users.end(); UI != E; ++UI) {
2000 Value *Op = UI->getOperandValToReplace();
2002 // If the IVStrideUse implies a cast, check for an actual cast which
2003 // can be used to find the original IV expression.
2004 if (SE->getEffectiveSCEVType(Op->getType()) !=
2005 SE->getEffectiveSCEVType(SI->first->getType())) {
2006 CastInst *CI = dyn_cast<CastInst>(Op);
2007 // If it's not a simple cast, it's complicated.
2010 // If it's a cast from a type other than the stride type,
2011 // it's complicated.
2012 if (CI->getOperand(0)->getType() != SI->first->getType())
2014 // Ok, we found the IV expression in the stride's type.
2015 Op = CI->getOperand(0);
2019 if (NewCmpLHS->getType() == CmpTy)
2025 NewCmpTy = NewCmpLHS->getType();
2026 NewTyBits = SE->getTypeSizeInBits(NewCmpTy);
2027 const Type *NewCmpIntTy = IntegerType::get(Cond->getContext(), NewTyBits);
2028 if (RequiresTypeConversion(NewCmpTy, CmpTy)) {
2029 // Check if it is possible to rewrite it using
2030 // an iv / stride of a smaller integer type.
2031 unsigned Bits = NewTyBits;
2032 if (ICmpInst::isSigned(Predicate))
2034 uint64_t Mask = (1ULL << Bits) - 1;
2035 if (((uint64_t)NewCmpVal & Mask) != (uint64_t)NewCmpVal)
2039 // Don't rewrite if use offset is non-constant and the new type is
2040 // of a different type.
2041 // FIXME: too conservative?
2042 if (NewTyBits != TyBits && !isa<SCEVConstant>(CondUse->getOffset()))
2046 bool AllUsesAreAddresses = true;
2047 bool AllUsesAreOutsideLoop = true;
2048 std::vector<BasedUser> UsersToProcess;
2049 const SCEV *CommonExprs = CollectIVUsers(SI->first, *SI->second, L,
2050 AllUsesAreAddresses,
2051 AllUsesAreOutsideLoop,
2053 // Avoid rewriting the compare instruction with an iv of new stride
2054 // if it's likely the new stride uses will be rewritten using the
2055 // stride of the compare instruction.
2056 if (AllUsesAreAddresses &&
2057 ValidScale(!CommonExprs->isZero(), Scale, UsersToProcess))
2061 // Avoid rewriting the compare instruction with an iv which has
2062 // implicit extension or truncation built into it.
2063 // TODO: This is over-conservative.
2064 if (SE->getTypeSizeInBits(CondUse->getOffset()->getType()) != TyBits)
2067 // If scale is negative, use swapped predicate unless it's testing
2069 if (Scale < 0 && !Cond->isEquality())
2070 Predicate = ICmpInst::getSwappedPredicate(Predicate);
2072 NewStride = IU->StrideOrder[i];
2073 if (!isa<PointerType>(NewCmpTy))
2074 NewCmpRHS = ConstantInt::get(NewCmpTy, NewCmpVal);
2076 Constant *CI = ConstantInt::get(NewCmpIntTy, NewCmpVal);
2077 NewCmpRHS = ConstantExpr::getIntToPtr(CI, NewCmpTy);
2079 NewOffset = TyBits == NewTyBits
2080 ? SE->getMulExpr(CondUse->getOffset(),
2081 SE->getConstant(CmpTy, Scale))
2082 : SE->getConstant(NewCmpIntTy,
2083 cast<SCEVConstant>(CondUse->getOffset())->getValue()
2084 ->getSExtValue()*Scale);
2089 // Forgo this transformation if it the increment happens to be
2090 // unfortunately positioned after the condition, and the condition
2091 // has multiple uses which prevent it from being moved immediately
2092 // before the branch. See
2093 // test/Transforms/LoopStrengthReduce/change-compare-stride-trickiness-*.ll
2094 // for an example of this situation.
2095 if (!Cond->hasOneUse()) {
2096 for (BasicBlock::iterator I = Cond, E = Cond->getParent()->end();
2103 // Create a new compare instruction using new stride / iv.
2104 ICmpInst *OldCond = Cond;
2105 // Insert new compare instruction.
2106 Cond = new ICmpInst(OldCond, Predicate, NewCmpLHS, NewCmpRHS,
2107 L->getHeader()->getName() + ".termcond");
2109 DEBUG(errs() << " Change compare stride in Inst " << *OldCond);
2110 DEBUG(errs() << " to " << *Cond << '\n');
2112 // Remove the old compare instruction. The old indvar is probably dead too.
2113 DeadInsts.push_back(CondUse->getOperandValToReplace());
2114 OldCond->replaceAllUsesWith(Cond);
2115 OldCond->eraseFromParent();
2117 IU->IVUsesByStride[NewStride]->addUser(NewOffset, Cond, NewCmpLHS);
2118 CondUse = &IU->IVUsesByStride[NewStride]->Users.back();
2119 CondStride = NewStride;
2127 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
2128 /// a max computation.
2130 /// This is a narrow solution to a specific, but acute, problem. For loops
2136 /// } while (++i < n);
2138 /// the trip count isn't just 'n', because 'n' might not be positive. And
2139 /// unfortunately this can come up even for loops where the user didn't use
2140 /// a C do-while loop. For example, seemingly well-behaved top-test loops
2141 /// will commonly be lowered like this:
2147 /// } while (++i < n);
2150 /// and then it's possible for subsequent optimization to obscure the if
2151 /// test in such a way that indvars can't find it.
2153 /// When indvars can't find the if test in loops like this, it creates a
2154 /// max expression, which allows it to give the loop a canonical
2155 /// induction variable:
2158 /// max = n < 1 ? 1 : n;
2161 /// } while (++i != max);
2163 /// Canonical induction variables are necessary because the loop passes
2164 /// are designed around them. The most obvious example of this is the
2165 /// LoopInfo analysis, which doesn't remember trip count values. It
2166 /// expects to be able to rediscover the trip count each time it is
2167 /// needed, and it does this using a simple analyis that only succeeds if
2168 /// the loop has a canonical induction variable.
2170 /// However, when it comes time to generate code, the maximum operation
2171 /// can be quite costly, especially if it's inside of an outer loop.
2173 /// This function solves this problem by detecting this type of loop and
2174 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
2175 /// the instructions for the maximum computation.
2177 ICmpInst *LoopStrengthReduce::OptimizeMax(Loop *L, ICmpInst *Cond,
2178 IVStrideUse* &CondUse) {
2179 // Check that the loop matches the pattern we're looking for.
2180 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
2181 Cond->getPredicate() != CmpInst::ICMP_NE)
2184 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
2185 if (!Sel || !Sel->hasOneUse()) return Cond;
2187 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
2188 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2190 const SCEV *One = SE->getIntegerSCEV(1, BackedgeTakenCount->getType());
2192 // Add one to the backedge-taken count to get the trip count.
2193 const SCEV *IterationCount = SE->getAddExpr(BackedgeTakenCount, One);
2195 // Check for a max calculation that matches the pattern.
2196 if (!isa<SCEVSMaxExpr>(IterationCount) && !isa<SCEVUMaxExpr>(IterationCount))
2198 const SCEVNAryExpr *Max = cast<SCEVNAryExpr>(IterationCount);
2199 if (Max != SE->getSCEV(Sel)) return Cond;
2201 // To handle a max with more than two operands, this optimization would
2202 // require additional checking and setup.
2203 if (Max->getNumOperands() != 2)
2206 const SCEV *MaxLHS = Max->getOperand(0);
2207 const SCEV *MaxRHS = Max->getOperand(1);
2208 if (!MaxLHS || MaxLHS != One) return Cond;
2210 // Check the relevant induction variable for conformance to
2212 const SCEV *IV = SE->getSCEV(Cond->getOperand(0));
2213 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2214 if (!AR || !AR->isAffine() ||
2215 AR->getStart() != One ||
2216 AR->getStepRecurrence(*SE) != One)
2219 assert(AR->getLoop() == L &&
2220 "Loop condition operand is an addrec in a different loop!");
2222 // Check the right operand of the select, and remember it, as it will
2223 // be used in the new comparison instruction.
2225 if (SE->getSCEV(Sel->getOperand(1)) == MaxRHS)
2226 NewRHS = Sel->getOperand(1);
2227 else if (SE->getSCEV(Sel->getOperand(2)) == MaxRHS)
2228 NewRHS = Sel->getOperand(2);
2229 if (!NewRHS) return Cond;
2231 // Determine the new comparison opcode. It may be signed or unsigned,
2232 // and the original comparison may be either equality or inequality.
2233 CmpInst::Predicate Pred =
2234 isa<SCEVSMaxExpr>(Max) ? CmpInst::ICMP_SLT : CmpInst::ICMP_ULT;
2235 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
2236 Pred = CmpInst::getInversePredicate(Pred);
2238 // Ok, everything looks ok to change the condition into an SLT or SGE and
2239 // delete the max calculation.
2241 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
2243 // Delete the max calculation instructions.
2244 Cond->replaceAllUsesWith(NewCond);
2245 CondUse->setUser(NewCond);
2246 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2247 Cond->eraseFromParent();
2248 Sel->eraseFromParent();
2249 if (Cmp->use_empty())
2250 Cmp->eraseFromParent();
2254 /// OptimizeShadowIV - If IV is used in a int-to-float cast
2255 /// inside the loop then try to eliminate the cast opeation.
2256 void LoopStrengthReduce::OptimizeShadowIV(Loop *L) {
2258 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
2259 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2262 for (unsigned Stride = 0, e = IU->StrideOrder.size(); Stride != e;
2264 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
2265 IU->IVUsesByStride.find(IU->StrideOrder[Stride]);
2266 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
2267 if (!isa<SCEVConstant>(SI->first))
2270 for (ilist<IVStrideUse>::iterator UI = SI->second->Users.begin(),
2271 E = SI->second->Users.end(); UI != E; /* empty */) {
2272 ilist<IVStrideUse>::iterator CandidateUI = UI;
2274 Instruction *ShadowUse = CandidateUI->getUser();
2275 const Type *DestTy = NULL;
2277 /* If shadow use is a int->float cast then insert a second IV
2278 to eliminate this cast.
2280 for (unsigned i = 0; i < n; ++i)
2286 for (unsigned i = 0; i < n; ++i, ++d)
2289 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser()))
2290 DestTy = UCast->getDestTy();
2291 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser()))
2292 DestTy = SCast->getDestTy();
2293 if (!DestTy) continue;
2296 // If target does not support DestTy natively then do not apply
2297 // this transformation.
2298 EVT DVT = TLI->getValueType(DestTy);
2299 if (!TLI->isTypeLegal(DVT)) continue;
2302 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
2304 if (PH->getNumIncomingValues() != 2) continue;
2306 const Type *SrcTy = PH->getType();
2307 int Mantissa = DestTy->getFPMantissaWidth();
2308 if (Mantissa == -1) continue;
2309 if ((int)SE->getTypeSizeInBits(SrcTy) > Mantissa)
2312 unsigned Entry, Latch;
2313 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
2321 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
2322 if (!Init) continue;
2323 Constant *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());
2325 BinaryOperator *Incr =
2326 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
2327 if (!Incr) continue;
2328 if (Incr->getOpcode() != Instruction::Add
2329 && Incr->getOpcode() != Instruction::Sub)
2332 /* Initialize new IV, double d = 0.0 in above example. */
2333 ConstantInt *C = NULL;
2334 if (Incr->getOperand(0) == PH)
2335 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
2336 else if (Incr->getOperand(1) == PH)
2337 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
2343 // Ignore negative constants, as the code below doesn't handle them
2344 // correctly. TODO: Remove this restriction.
2345 if (!C->getValue().isStrictlyPositive()) continue;
2347 /* Add new PHINode. */
2348 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);
2350 /* create new increment. '++d' in above example. */
2351 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
2352 BinaryOperator *NewIncr =
2353 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
2354 Instruction::FAdd : Instruction::FSub,
2355 NewPH, CFP, "IV.S.next.", Incr);
2357 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
2358 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
2360 /* Remove cast operation */
2361 ShadowUse->replaceAllUsesWith(NewPH);
2362 ShadowUse->eraseFromParent();
2369 /// OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
2370 /// uses in the loop, look to see if we can eliminate some, in favor of using
2371 /// common indvars for the different uses.
2372 void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
2373 // TODO: implement optzns here.
2375 OptimizeShadowIV(L);
2378 bool LoopStrengthReduce::StrideMightBeShared(const SCEV* Stride, Loop *L,
2380 int64_t SInt = cast<SCEVConstant>(Stride)->getValue()->getSExtValue();
2381 for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
2382 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
2383 IU->IVUsesByStride.find(IU->StrideOrder[i]);
2384 const SCEV *Share = SI->first;
2385 if (!isa<SCEVConstant>(SI->first) || Share == Stride)
2387 int64_t SSInt = cast<SCEVConstant>(Share)->getValue()->getSExtValue();
2389 return true; // This can definitely be reused.
2390 if (unsigned(abs64(SSInt)) < SInt || (SSInt % SInt) != 0)
2392 int64_t Scale = SSInt / SInt;
2393 bool AllUsesAreAddresses = true;
2394 bool AllUsesAreOutsideLoop = true;
2395 std::vector<BasedUser> UsersToProcess;
2396 const SCEV *CommonExprs = CollectIVUsers(SI->first, *SI->second, L,
2397 AllUsesAreAddresses,
2398 AllUsesAreOutsideLoop,
2400 if (AllUsesAreAddresses &&
2401 ValidScale(!CommonExprs->isZero(), Scale, UsersToProcess)) {
2404 // Any pre-inc iv use?
2405 IVUsersOfOneStride &StrideUses = *IU->IVUsesByStride[Share];
2406 for (ilist<IVStrideUse>::iterator I = StrideUses.Users.begin(),
2407 E = StrideUses.Users.end(); I != E; ++I) {
2408 if (!I->isUseOfPostIncrementedValue())
2416 /// isUsedByExitBranch - Return true if icmp is used by a loop terminating
2417 /// conditional branch or it's and / or with other conditions before being used
2418 /// as the condition.
2419 static bool isUsedByExitBranch(ICmpInst *Cond, Loop *L) {
2420 BasicBlock *CondBB = Cond->getParent();
2421 if (!L->isLoopExiting(CondBB))
2423 BranchInst *TermBr = dyn_cast<BranchInst>(CondBB->getTerminator());
2424 if (!TermBr || !TermBr->isConditional())
2427 Value *User = *Cond->use_begin();
2428 Instruction *UserInst = dyn_cast<Instruction>(User);
2430 (UserInst->getOpcode() == Instruction::And ||
2431 UserInst->getOpcode() == Instruction::Or)) {
2432 if (!UserInst->hasOneUse() || UserInst->getParent() != CondBB)
2434 User = *User->use_begin();
2435 UserInst = dyn_cast<Instruction>(User);
2437 return User == TermBr;
2440 static bool ShouldCountToZero(ICmpInst *Cond, IVStrideUse* &CondUse,
2441 ScalarEvolution *SE, Loop *L,
2442 const TargetLowering *TLI = 0) {
2443 if (!L->contains(Cond->getParent()))
2446 if (!isa<SCEVConstant>(CondUse->getOffset()))
2449 // Handle only tests for equality for the moment.
2450 if (!Cond->isEquality() || !Cond->hasOneUse())
2452 if (!isUsedByExitBranch(Cond, L))
2455 Value *CondOp0 = Cond->getOperand(0);
2456 const SCEV *IV = SE->getSCEV(CondOp0);
2457 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2458 if (!AR || !AR->isAffine())
2461 const SCEVConstant *SC = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE));
2462 if (!SC || SC->getValue()->getSExtValue() < 0)
2463 // If it's already counting down, don't do anything.
2466 // If the RHS of the comparison is not an loop invariant, the rewrite
2467 // cannot be done. Also bail out if it's already comparing against a zero.
2468 // If we are checking this before cmp stride optimization, check if it's
2469 // comparing against a already legal immediate.
2470 Value *RHS = Cond->getOperand(1);
2471 ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS);
2472 if (!L->isLoopInvariant(RHS) ||
2473 (RHSC && RHSC->isZero()) ||
2474 (RHSC && TLI && TLI->isLegalICmpImmediate(RHSC->getSExtValue())))
2477 // Make sure the IV is only used for counting. Value may be preinc or
2478 // postinc; 2 uses in either case.
2479 if (!CondOp0->hasNUses(2))
2485 /// OptimizeLoopTermCond - Change loop terminating condition to use the
2486 /// postinc iv when possible.
2487 void LoopStrengthReduce::OptimizeLoopTermCond(Loop *L) {
2488 BasicBlock *LatchBlock = L->getLoopLatch();
2489 bool LatchExit = L->isLoopExiting(LatchBlock);
2490 SmallVector<BasicBlock*, 8> ExitingBlocks;
2491 L->getExitingBlocks(ExitingBlocks);
2493 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
2494 BasicBlock *ExitingBlock = ExitingBlocks[i];
2496 // Finally, get the terminating condition for the loop if possible. If we
2497 // can, we want to change it to use a post-incremented version of its
2498 // induction variable, to allow coalescing the live ranges for the IV into
2499 // one register value.
2501 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2504 // FIXME: Overly conservative, termination condition could be an 'or' etc..
2505 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
2508 // Search IVUsesByStride to find Cond's IVUse if there is one.
2509 IVStrideUse *CondUse = 0;
2510 const SCEV *CondStride = 0;
2511 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2512 if (!FindIVUserForCond(Cond, CondUse, CondStride))
2515 // If the latch block is exiting and it's not a single block loop, it's
2516 // not safe to use postinc iv in other exiting blocks. FIXME: overly
2517 // conservative? How about icmp stride optimization?
2518 bool UsePostInc = !(e > 1 && LatchExit && ExitingBlock != LatchBlock);
2519 if (UsePostInc && ExitingBlock != LatchBlock) {
2520 if (!Cond->hasOneUse())
2521 // See below, we don't want the condition to be cloned.
2524 // If exiting block is the latch block, we know it's safe and profitable
2525 // to transform the icmp to use post-inc iv. Otherwise do so only if it
2526 // would not reuse another iv and its iv would be reused by other uses.
2527 // We are optimizing for the case where the icmp is the only use of the
2529 IVUsersOfOneStride &StrideUses = *IU->IVUsesByStride[CondStride];
2530 for (ilist<IVStrideUse>::iterator I = StrideUses.Users.begin(),
2531 E = StrideUses.Users.end(); I != E; ++I) {
2532 if (I->getUser() == Cond)
2534 if (!I->isUseOfPostIncrementedValue()) {
2541 // If iv for the stride might be shared and any of the users use pre-inc
2542 // iv might be used, then it's not safe to use post-inc iv.
2544 isa<SCEVConstant>(CondStride) &&
2545 StrideMightBeShared(CondStride, L, true))
2549 // If the trip count is computed in terms of a max (due to ScalarEvolution
2550 // being unable to find a sufficient guard, for example), change the loop
2551 // comparison to use SLT or ULT instead of NE.
2552 Cond = OptimizeMax(L, Cond, CondUse);
2554 // If possible, change stride and operands of the compare instruction to
2555 // eliminate one stride. However, avoid rewriting the compare instruction
2556 // with an iv of new stride if it's likely the new stride uses will be
2557 // rewritten using the stride of the compare instruction.
2558 if (ExitingBlock == LatchBlock && isa<SCEVConstant>(CondStride)) {
2559 // If the condition stride is a constant and it's the only use, we might
2560 // want to optimize it first by turning it to count toward zero.
2561 if (!StrideMightBeShared(CondStride, L, false) &&
2562 !ShouldCountToZero(Cond, CondUse, SE, L, TLI))
2563 Cond = ChangeCompareStride(L, Cond, CondUse, CondStride);
2569 DEBUG(errs() << " Change loop exiting icmp to use postinc iv: "
2572 // It's possible for the setcc instruction to be anywhere in the loop, and
2573 // possible for it to have multiple users. If it is not immediately before
2574 // the exiting block branch, move it.
2575 if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
2576 if (Cond->hasOneUse()) { // Condition has a single use, just move it.
2577 Cond->moveBefore(TermBr);
2579 // Otherwise, clone the terminating condition and insert into the
2581 Cond = cast<ICmpInst>(Cond->clone());
2582 Cond->setName(L->getHeader()->getName() + ".termcond");
2583 ExitingBlock->getInstList().insert(TermBr, Cond);
2585 // Clone the IVUse, as the old use still exists!
2586 IU->IVUsesByStride[CondStride]->addUser(CondUse->getOffset(), Cond,
2587 CondUse->getOperandValToReplace());
2588 CondUse = &IU->IVUsesByStride[CondStride]->Users.back();
2592 // If we get to here, we know that we can transform the setcc instruction to
2593 // use the post-incremented version of the IV, allowing us to coalesce the
2594 // live ranges for the IV correctly.
2595 CondUse->setOffset(SE->getMinusSCEV(CondUse->getOffset(), CondStride));
2596 CondUse->setIsUseOfPostIncrementedValue(true);
2603 bool LoopStrengthReduce::OptimizeLoopCountIVOfStride(const SCEV* &Stride,
2604 IVStrideUse* &CondUse,
2606 // If the only use is an icmp of a loop exiting conditional branch, then
2607 // attempt the optimization.
2608 BasedUser User = BasedUser(*CondUse, SE);
2609 assert(isa<ICmpInst>(User.Inst) && "Expecting an ICMPInst!");
2610 ICmpInst *Cond = cast<ICmpInst>(User.Inst);
2612 // Less strict check now that compare stride optimization is done.
2613 if (!ShouldCountToZero(Cond, CondUse, SE, L))
2616 Value *CondOp0 = Cond->getOperand(0);
2617 PHINode *PHIExpr = dyn_cast<PHINode>(CondOp0);
2620 // Value tested is postinc. Find the phi node.
2621 Incr = dyn_cast<BinaryOperator>(CondOp0);
2622 // FIXME: Just use User.OperandValToReplace here?
2623 if (!Incr || Incr->getOpcode() != Instruction::Add)
2626 PHIExpr = dyn_cast<PHINode>(Incr->getOperand(0));
2629 // 1 use for preinc value, the increment.
2630 if (!PHIExpr->hasOneUse())
2633 assert(isa<PHINode>(CondOp0) &&
2634 "Unexpected loop exiting counting instruction sequence!");
2635 PHIExpr = cast<PHINode>(CondOp0);
2636 // Value tested is preinc. Find the increment.
2637 // A CmpInst is not a BinaryOperator; we depend on this.
2638 Instruction::use_iterator UI = PHIExpr->use_begin();
2639 Incr = dyn_cast<BinaryOperator>(UI);
2641 Incr = dyn_cast<BinaryOperator>(++UI);
2642 // One use for postinc value, the phi. Unnecessarily conservative?
2643 if (!Incr || !Incr->hasOneUse() || Incr->getOpcode() != Instruction::Add)
2647 // Replace the increment with a decrement.
2648 DEBUG(errs() << "LSR: Examining use ");
2649 DEBUG(WriteAsOperand(errs(), CondOp0, /*PrintType=*/false));
2650 DEBUG(errs() << " in Inst: " << *Cond << '\n');
2651 BinaryOperator *Decr = BinaryOperator::Create(Instruction::Sub,
2652 Incr->getOperand(0), Incr->getOperand(1), "tmp", Incr);
2653 Incr->replaceAllUsesWith(Decr);
2654 Incr->eraseFromParent();
2656 // Substitute endval-startval for the original startval, and 0 for the
2657 // original endval. Since we're only testing for equality this is OK even
2658 // if the computation wraps around.
2659 BasicBlock *Preheader = L->getLoopPreheader();
2660 Instruction *PreInsertPt = Preheader->getTerminator();
2661 unsigned InBlock = L->contains(PHIExpr->getIncomingBlock(0)) ? 1 : 0;
2662 Value *StartVal = PHIExpr->getIncomingValue(InBlock);
2663 Value *EndVal = Cond->getOperand(1);
2664 DEBUG(errs() << " Optimize loop counting iv to count down ["
2665 << *EndVal << " .. " << *StartVal << "]\n");
2667 // FIXME: check for case where both are constant.
2668 Constant* Zero = ConstantInt::get(Cond->getOperand(1)->getType(), 0);
2669 BinaryOperator *NewStartVal = BinaryOperator::Create(Instruction::Sub,
2670 EndVal, StartVal, "tmp", PreInsertPt);
2671 PHIExpr->setIncomingValue(InBlock, NewStartVal);
2672 Cond->setOperand(1, Zero);
2673 DEBUG(errs() << " New icmp: " << *Cond << "\n");
2675 int64_t SInt = cast<SCEVConstant>(Stride)->getValue()->getSExtValue();
2676 const SCEV *NewStride = 0;
2678 for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
2679 const SCEV *OldStride = IU->StrideOrder[i];
2680 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(OldStride))
2681 if (SC->getValue()->getSExtValue() == -SInt) {
2683 NewStride = OldStride;
2689 NewStride = SE->getIntegerSCEV(-SInt, Stride->getType());
2690 IU->AddUser(NewStride, CondUse->getOffset(), Cond, Cond->getOperand(0));
2691 IU->IVUsesByStride[Stride]->removeUser(CondUse);
2693 CondUse = &IU->IVUsesByStride[NewStride]->Users.back();
2701 /// OptimizeLoopCountIV - If, after all sharing of IVs, the IV used for deciding
2702 /// when to exit the loop is used only for that purpose, try to rearrange things
2703 /// so it counts down to a test against zero.
2704 bool LoopStrengthReduce::OptimizeLoopCountIV(Loop *L) {
2705 bool ThisChanged = false;
2706 for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
2707 const SCEV *Stride = IU->StrideOrder[i];
2708 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
2709 IU->IVUsesByStride.find(Stride);
2710 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
2711 // FIXME: Generalize to non-affine IV's.
2712 if (!SI->first->isLoopInvariant(L))
2714 // If stride is a constant and it has an icmpinst use, check if we can
2715 // optimize the loop to count down.
2716 if (isa<SCEVConstant>(Stride) && SI->second->Users.size() == 1) {
2717 Instruction *User = SI->second->Users.begin()->getUser();
2718 if (!isa<ICmpInst>(User))
2720 const SCEV *CondStride = Stride;
2721 IVStrideUse *Use = &*SI->second->Users.begin();
2722 if (!OptimizeLoopCountIVOfStride(CondStride, Use, L))
2726 // Now check if it's possible to reuse this iv for other stride uses.
2727 for (unsigned j = 0, ee = IU->StrideOrder.size(); j != ee; ++j) {
2728 const SCEV *SStride = IU->StrideOrder[j];
2729 if (SStride == CondStride)
2731 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SII =
2732 IU->IVUsesByStride.find(SStride);
2733 assert(SII != IU->IVUsesByStride.end() && "Stride doesn't exist!");
2734 // FIXME: Generalize to non-affine IV's.
2735 if (!SII->first->isLoopInvariant(L))
2737 // FIXME: Rewrite other stride using CondStride.
2742 Changed |= ThisChanged;
2746 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager &LPM) {
2747 IU = &getAnalysis<IVUsers>();
2748 LI = &getAnalysis<LoopInfo>();
2749 DT = &getAnalysis<DominatorTree>();
2750 SE = &getAnalysis<ScalarEvolution>();
2753 // If LoopSimplify form is not available, stay out of trouble.
2754 if (!L->getLoopPreheader() || !L->getLoopLatch())
2757 if (!IU->IVUsesByStride.empty()) {
2758 DEBUG(errs() << "\nLSR on \"" << L->getHeader()->getParent()->getName()
2762 // Sort the StrideOrder so we process larger strides first.
2763 std::stable_sort(IU->StrideOrder.begin(), IU->StrideOrder.end(),
2766 // Optimize induction variables. Some indvar uses can be transformed to use
2767 // strides that will be needed for other purposes. A common example of this
2768 // is the exit test for the loop, which can often be rewritten to use the
2769 // computation of some other indvar to decide when to terminate the loop.
2772 // Change loop terminating condition to use the postinc iv when possible
2773 // and optimize loop terminating compare. FIXME: Move this after
2774 // StrengthReduceIVUsersOfStride?
2775 OptimizeLoopTermCond(L);
2777 // FIXME: We can shrink overlarge IV's here. e.g. if the code has
2778 // computation in i64 values and the target doesn't support i64, demote
2779 // the computation to 32-bit if safe.
2781 // FIXME: Attempt to reuse values across multiple IV's. In particular, we
2782 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should
2783 // be codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC.
2784 // Need to be careful that IV's are all the same type. Only works for
2785 // intptr_t indvars.
2787 // IVsByStride keeps IVs for one particular loop.
2788 assert(IVsByStride.empty() && "Stale entries in IVsByStride?");
2790 StrengthReduceIVUsers(L);
2792 // After all sharing is done, see if we can adjust the loop to test against
2793 // zero instead of counting up to a maximum. This is usually faster.
2794 OptimizeLoopCountIV(L);
2797 // We're done analyzing this loop; release all the state we built up for it.
2798 IVsByStride.clear();
2799 StrideNoReuse.clear();
2801 // Clean up after ourselves
2802 if (!DeadInsts.empty())
2803 DeleteTriviallyDeadInstructions();
2805 // At this point, it is worth checking to see if any recurrence PHIs are also
2806 // dead, so that we can remove them as well.
2807 DeleteDeadPHIs(L->getHeader());