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.pop_back_val());
233 if (I == 0 || !isInstructionTriviallyDead(I))
236 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
237 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
240 DeadInsts.push_back(U);
243 I->eraseFromParent();
248 /// containsAddRecFromDifferentLoop - Determine whether expression S involves a
249 /// subexpression that is an AddRec from a loop other than L. An outer loop
250 /// of L is OK, but not an inner loop nor a disjoint loop.
251 static bool containsAddRecFromDifferentLoop(const SCEV *S, Loop *L) {
252 // This is very common, put it first.
253 if (isa<SCEVConstant>(S))
255 if (const SCEVCommutativeExpr *AE = dyn_cast<SCEVCommutativeExpr>(S)) {
256 for (unsigned int i=0; i< AE->getNumOperands(); i++)
257 if (containsAddRecFromDifferentLoop(AE->getOperand(i), L))
261 if (const SCEVAddRecExpr *AE = dyn_cast<SCEVAddRecExpr>(S)) {
262 if (const Loop *newLoop = AE->getLoop()) {
265 // if newLoop is an outer loop of L, this is OK.
266 if (!LoopInfo::isNotAlreadyContainedIn(L, newLoop))
271 if (const SCEVUDivExpr *DE = dyn_cast<SCEVUDivExpr>(S))
272 return containsAddRecFromDifferentLoop(DE->getLHS(), L) ||
273 containsAddRecFromDifferentLoop(DE->getRHS(), L);
275 // SCEVSDivExpr has been backed out temporarily, but will be back; we'll
276 // need this when it is.
277 if (const SCEVSDivExpr *DE = dyn_cast<SCEVSDivExpr>(S))
278 return containsAddRecFromDifferentLoop(DE->getLHS(), L) ||
279 containsAddRecFromDifferentLoop(DE->getRHS(), L);
281 if (const SCEVCastExpr *CE = dyn_cast<SCEVCastExpr>(S))
282 return containsAddRecFromDifferentLoop(CE->getOperand(), L);
286 /// isAddressUse - Returns true if the specified instruction is using the
287 /// specified value as an address.
288 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
289 bool isAddress = isa<LoadInst>(Inst);
290 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
291 if (SI->getOperand(1) == OperandVal)
293 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
294 // Addressing modes can also be folded into prefetches and a variety
296 switch (II->getIntrinsicID()) {
298 case Intrinsic::prefetch:
299 case Intrinsic::x86_sse2_loadu_dq:
300 case Intrinsic::x86_sse2_loadu_pd:
301 case Intrinsic::x86_sse_loadu_ps:
302 case Intrinsic::x86_sse_storeu_ps:
303 case Intrinsic::x86_sse2_storeu_pd:
304 case Intrinsic::x86_sse2_storeu_dq:
305 case Intrinsic::x86_sse2_storel_dq:
306 if (II->getOperand(1) == OperandVal)
314 /// getAccessType - Return the type of the memory being accessed.
315 static const Type *getAccessType(const Instruction *Inst) {
316 const Type *AccessTy = Inst->getType();
317 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
318 AccessTy = SI->getOperand(0)->getType();
319 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
320 // Addressing modes can also be folded into prefetches and a variety
322 switch (II->getIntrinsicID()) {
324 case Intrinsic::x86_sse_storeu_ps:
325 case Intrinsic::x86_sse2_storeu_pd:
326 case Intrinsic::x86_sse2_storeu_dq:
327 case Intrinsic::x86_sse2_storel_dq:
328 AccessTy = II->getOperand(1)->getType();
336 /// BasedUser - For a particular base value, keep information about how we've
337 /// partitioned the expression so far.
339 /// SE - The current ScalarEvolution object.
342 /// Base - The Base value for the PHI node that needs to be inserted for
343 /// this use. As the use is processed, information gets moved from this
344 /// field to the Imm field (below). BasedUser values are sorted by this
348 /// Inst - The instruction using the induction variable.
351 /// OperandValToReplace - The operand value of Inst to replace with the
353 Value *OperandValToReplace;
355 /// Imm - The immediate value that should be added to the base immediately
356 /// before Inst, because it will be folded into the imm field of the
357 /// instruction. This is also sometimes used for loop-variant values that
358 /// must be added inside the loop.
361 /// Phi - The induction variable that performs the striding that
362 /// should be used for this user.
365 // isUseOfPostIncrementedValue - True if this should use the
366 // post-incremented version of this IV, not the preincremented version.
367 // This can only be set in special cases, such as the terminating setcc
368 // instruction for a loop and uses outside the loop that are dominated by
370 bool isUseOfPostIncrementedValue;
372 BasedUser(IVStrideUse &IVSU, ScalarEvolution *se)
373 : SE(se), Base(IVSU.getOffset()), Inst(IVSU.getUser()),
374 OperandValToReplace(IVSU.getOperandValToReplace()),
375 Imm(SE->getIntegerSCEV(0, Base->getType())),
376 isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue()) {}
378 // Once we rewrite the code to insert the new IVs we want, update the
379 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
381 void RewriteInstructionToUseNewBase(const SCEV *const &NewBase,
382 Instruction *InsertPt,
383 SCEVExpander &Rewriter, Loop *L, Pass *P,
385 SmallVectorImpl<WeakVH> &DeadInsts);
387 Value *InsertCodeForBaseAtPosition(const SCEV *const &NewBase,
389 SCEVExpander &Rewriter,
390 Instruction *IP, Loop *L,
396 void BasedUser::dump() const {
397 errs() << " Base=" << *Base;
398 errs() << " Imm=" << *Imm;
399 errs() << " Inst: " << *Inst;
402 Value *BasedUser::InsertCodeForBaseAtPosition(const SCEV *const &NewBase,
404 SCEVExpander &Rewriter,
405 Instruction *IP, Loop *L,
407 // Figure out where we *really* want to insert this code. In particular, if
408 // the user is inside of a loop that is nested inside of L, we really don't
409 // want to insert this expression before the user, we'd rather pull it out as
410 // many loops as possible.
411 Instruction *BaseInsertPt = IP;
413 // Figure out the most-nested loop that IP is in.
414 Loop *InsertLoop = LI.getLoopFor(IP->getParent());
416 // If InsertLoop is not L, and InsertLoop is nested inside of L, figure out
417 // the preheader of the outer-most loop where NewBase is not loop invariant.
418 if (L->contains(IP->getParent()))
419 while (InsertLoop && NewBase->isLoopInvariant(InsertLoop)) {
420 BaseInsertPt = InsertLoop->getLoopPreheader()->getTerminator();
421 InsertLoop = InsertLoop->getParentLoop();
424 Value *Base = Rewriter.expandCodeFor(NewBase, 0, BaseInsertPt);
426 const SCEV *NewValSCEV = SE->getUnknown(Base);
428 // Always emit the immediate into the same block as the user.
429 NewValSCEV = SE->getAddExpr(NewValSCEV, Imm);
431 return Rewriter.expandCodeFor(NewValSCEV, Ty, IP);
435 // Once we rewrite the code to insert the new IVs we want, update the
436 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
437 // to it. NewBasePt is the last instruction which contributes to the
438 // value of NewBase in the case that it's a diffferent instruction from
439 // the PHI that NewBase is computed from, or null otherwise.
441 void BasedUser::RewriteInstructionToUseNewBase(const SCEV *const &NewBase,
442 Instruction *NewBasePt,
443 SCEVExpander &Rewriter, Loop *L, Pass *P,
445 SmallVectorImpl<WeakVH> &DeadInsts) {
446 if (!isa<PHINode>(Inst)) {
447 // By default, insert code at the user instruction.
448 BasicBlock::iterator InsertPt = Inst;
450 // However, if the Operand is itself an instruction, the (potentially
451 // complex) inserted code may be shared by many users. Because of this, we
452 // want to emit code for the computation of the operand right before its old
453 // computation. This is usually safe, because we obviously used to use the
454 // computation when it was computed in its current block. However, in some
455 // cases (e.g. use of a post-incremented induction variable) the NewBase
456 // value will be pinned to live somewhere after the original computation.
457 // In this case, we have to back off.
459 // If this is a use outside the loop (which means after, since it is based
460 // on a loop indvar) we use the post-incremented value, so that we don't
461 // artificially make the preinc value live out the bottom of the loop.
462 if (!isUseOfPostIncrementedValue && L->contains(Inst->getParent())) {
463 if (NewBasePt && isa<PHINode>(OperandValToReplace)) {
464 InsertPt = NewBasePt;
466 } else if (Instruction *OpInst
467 = dyn_cast<Instruction>(OperandValToReplace)) {
469 while (isa<PHINode>(InsertPt)) ++InsertPt;
472 Value *NewVal = InsertCodeForBaseAtPosition(NewBase,
473 OperandValToReplace->getType(),
474 Rewriter, InsertPt, L, LI);
475 // Replace the use of the operand Value with the new Phi we just created.
476 Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
478 DEBUG(errs() << " Replacing with ");
479 DEBUG(WriteAsOperand(errs(), NewVal, /*PrintType=*/false));
480 DEBUG(errs() << ", which has value " << *NewBase << " plus IMM "
485 // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm
486 // expression into each operand block that uses it. Note that PHI nodes can
487 // have multiple entries for the same predecessor. We use a map to make sure
488 // that a PHI node only has a single Value* for each predecessor (which also
489 // prevents us from inserting duplicate code in some blocks).
490 DenseMap<BasicBlock*, Value*> InsertedCode;
491 PHINode *PN = cast<PHINode>(Inst);
492 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
493 if (PN->getIncomingValue(i) == OperandValToReplace) {
494 // If the original expression is outside the loop, put the replacement
495 // code in the same place as the original expression,
496 // which need not be an immediate predecessor of this PHI. This way we
497 // need only one copy of it even if it is referenced multiple times in
498 // the PHI. We don't do this when the original expression is inside the
499 // loop because multiple copies sometimes do useful sinking of code in
501 Instruction *OldLoc = dyn_cast<Instruction>(OperandValToReplace);
502 BasicBlock *PHIPred = PN->getIncomingBlock(i);
503 if (L->contains(OldLoc->getParent())) {
504 // If this is a critical edge, split the edge so that we do not insert
505 // the code on all predecessor/successor paths. We do this unless this
506 // is the canonical backedge for this loop, as this can make some
507 // inserted code be in an illegal position.
508 if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 &&
509 !isa<IndirectBrInst>(PHIPred->getTerminator()) &&
510 (PN->getParent() != L->getHeader() || !L->contains(PHIPred))) {
512 // First step, split the critical edge.
513 BasicBlock *NewBB = SplitCriticalEdge(PHIPred, PN->getParent(),
516 // Next step: move the basic block. In particular, if the PHI node
517 // is outside of the loop, and PredTI is in the loop, we want to
518 // move the block to be immediately before the PHI block, not
519 // immediately after PredTI.
520 if (L->contains(PHIPred) && !L->contains(PN->getParent()))
521 NewBB->moveBefore(PN->getParent());
523 // Splitting the edge can reduce the number of PHI entries we have.
524 e = PN->getNumIncomingValues();
526 i = PN->getBasicBlockIndex(PHIPred);
529 Value *&Code = InsertedCode[PHIPred];
531 // Insert the code into the end of the predecessor block.
532 Instruction *InsertPt = (L->contains(OldLoc->getParent())) ?
533 PHIPred->getTerminator() :
534 OldLoc->getParent()->getTerminator();
535 Code = InsertCodeForBaseAtPosition(NewBase, PN->getType(),
536 Rewriter, InsertPt, L, LI);
538 DEBUG(errs() << " Changing PHI use to ");
539 DEBUG(WriteAsOperand(errs(), Code, /*PrintType=*/false));
540 DEBUG(errs() << ", which has value " << *NewBase << " plus IMM "
544 // Replace the use of the operand Value with the new Phi we just created.
545 PN->setIncomingValue(i, Code);
550 // PHI node might have become a constant value after SplitCriticalEdge.
551 DeadInsts.push_back(Inst);
555 /// fitsInAddressMode - Return true if V can be subsumed within an addressing
556 /// mode, and does not need to be put in a register first.
557 static bool fitsInAddressMode(const SCEV *const &V, const Type *AccessTy,
558 const TargetLowering *TLI, bool HasBaseReg) {
559 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
560 int64_t VC = SC->getValue()->getSExtValue();
562 TargetLowering::AddrMode AM;
564 AM.HasBaseReg = HasBaseReg;
565 return TLI->isLegalAddressingMode(AM, AccessTy);
567 // Defaults to PPC. PPC allows a sign-extended 16-bit immediate field.
568 return (VC > -(1 << 16) && VC < (1 << 16)-1);
572 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
573 if (GlobalValue *GV = dyn_cast<GlobalValue>(SU->getValue())) {
575 TargetLowering::AddrMode AM;
577 AM.HasBaseReg = HasBaseReg;
578 return TLI->isLegalAddressingMode(AM, AccessTy);
580 // Default: assume global addresses are not legal.
587 /// MoveLoopVariantsToImmediateField - Move any subexpressions from Val that are
588 /// loop varying to the Imm operand.
589 static void MoveLoopVariantsToImmediateField(const SCEV *&Val, const SCEV *&Imm,
590 Loop *L, ScalarEvolution *SE) {
591 if (Val->isLoopInvariant(L)) return; // Nothing to do.
593 if (const SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
594 SmallVector<const SCEV *, 4> NewOps;
595 NewOps.reserve(SAE->getNumOperands());
597 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
598 if (!SAE->getOperand(i)->isLoopInvariant(L)) {
599 // If this is a loop-variant expression, it must stay in the immediate
600 // field of the expression.
601 Imm = SE->getAddExpr(Imm, SAE->getOperand(i));
603 NewOps.push_back(SAE->getOperand(i));
607 Val = SE->getIntegerSCEV(0, Val->getType());
609 Val = SE->getAddExpr(NewOps);
610 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
611 // Try to pull immediates out of the start value of nested addrec's.
612 const SCEV *Start = SARE->getStart();
613 MoveLoopVariantsToImmediateField(Start, Imm, L, SE);
615 SmallVector<const SCEV *, 4> Ops(SARE->op_begin(), SARE->op_end());
617 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
619 // Otherwise, all of Val is variant, move the whole thing over.
620 Imm = SE->getAddExpr(Imm, Val);
621 Val = SE->getIntegerSCEV(0, Val->getType());
626 /// MoveImmediateValues - Look at Val, and pull out any additions of constants
627 /// that can fit into the immediate field of instructions in the target.
628 /// Accumulate these immediate values into the Imm value.
629 static void MoveImmediateValues(const TargetLowering *TLI,
630 const Type *AccessTy,
631 const SCEV *&Val, const SCEV *&Imm,
632 bool isAddress, Loop *L,
633 ScalarEvolution *SE) {
634 if (const SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
635 SmallVector<const SCEV *, 4> NewOps;
636 NewOps.reserve(SAE->getNumOperands());
638 for (unsigned i = 0; i != SAE->getNumOperands(); ++i) {
639 const SCEV *NewOp = SAE->getOperand(i);
640 MoveImmediateValues(TLI, AccessTy, NewOp, Imm, isAddress, L, SE);
642 if (!NewOp->isLoopInvariant(L)) {
643 // If this is a loop-variant expression, it must stay in the immediate
644 // field of the expression.
645 Imm = SE->getAddExpr(Imm, NewOp);
647 NewOps.push_back(NewOp);
652 Val = SE->getIntegerSCEV(0, Val->getType());
654 Val = SE->getAddExpr(NewOps);
656 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
657 // Try to pull immediates out of the start value of nested addrec's.
658 const SCEV *Start = SARE->getStart();
659 MoveImmediateValues(TLI, AccessTy, Start, Imm, isAddress, L, SE);
661 if (Start != SARE->getStart()) {
662 SmallVector<const SCEV *, 4> Ops(SARE->op_begin(), SARE->op_end());
664 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
667 } else if (const SCEVMulExpr *SME = dyn_cast<SCEVMulExpr>(Val)) {
668 // Transform "8 * (4 + v)" -> "32 + 8*V" if "32" fits in the immed field.
670 fitsInAddressMode(SME->getOperand(0), AccessTy, TLI, false) &&
671 SME->getNumOperands() == 2 && SME->isLoopInvariant(L)) {
673 const SCEV *SubImm = SE->getIntegerSCEV(0, Val->getType());
674 const SCEV *NewOp = SME->getOperand(1);
675 MoveImmediateValues(TLI, AccessTy, NewOp, SubImm, isAddress, L, SE);
677 // If we extracted something out of the subexpressions, see if we can
679 if (NewOp != SME->getOperand(1)) {
680 // Scale SubImm up by "8". If the result is a target constant, we are
682 SubImm = SE->getMulExpr(SubImm, SME->getOperand(0));
683 if (fitsInAddressMode(SubImm, AccessTy, TLI, false)) {
684 // Accumulate the immediate.
685 Imm = SE->getAddExpr(Imm, SubImm);
687 // Update what is left of 'Val'.
688 Val = SE->getMulExpr(SME->getOperand(0), NewOp);
695 // Loop-variant expressions must stay in the immediate field of the
697 if ((isAddress && fitsInAddressMode(Val, AccessTy, TLI, false)) ||
698 !Val->isLoopInvariant(L)) {
699 Imm = SE->getAddExpr(Imm, Val);
700 Val = SE->getIntegerSCEV(0, Val->getType());
704 // Otherwise, no immediates to move.
707 static void MoveImmediateValues(const TargetLowering *TLI,
709 const SCEV *&Val, const SCEV *&Imm,
710 bool isAddress, Loop *L,
711 ScalarEvolution *SE) {
712 const Type *AccessTy = getAccessType(User);
713 MoveImmediateValues(TLI, AccessTy, Val, Imm, isAddress, L, SE);
716 /// SeparateSubExprs - Decompose Expr into all of the subexpressions that are
717 /// added together. This is used to reassociate common addition subexprs
718 /// together for maximal sharing when rewriting bases.
719 static void SeparateSubExprs(SmallVector<const SCEV *, 16> &SubExprs,
721 ScalarEvolution *SE) {
722 if (const SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) {
723 for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
724 SeparateSubExprs(SubExprs, AE->getOperand(j), SE);
725 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) {
726 const SCEV *Zero = SE->getIntegerSCEV(0, Expr->getType());
727 if (SARE->getOperand(0) == Zero) {
728 SubExprs.push_back(Expr);
730 // Compute the addrec with zero as its base.
731 SmallVector<const SCEV *, 4> Ops(SARE->op_begin(), SARE->op_end());
732 Ops[0] = Zero; // Start with zero base.
733 SubExprs.push_back(SE->getAddRecExpr(Ops, SARE->getLoop()));
736 SeparateSubExprs(SubExprs, SARE->getOperand(0), SE);
738 } else if (!Expr->isZero()) {
740 SubExprs.push_back(Expr);
744 // This is logically local to the following function, but C++ says we have
745 // to make it file scope.
746 struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; };
748 /// RemoveCommonExpressionsFromUseBases - Look through all of the Bases of all
749 /// the Uses, removing any common subexpressions, except that if all such
750 /// subexpressions can be folded into an addressing mode for all uses inside
751 /// the loop (this case is referred to as "free" in comments herein) we do
752 /// not remove anything. This looks for things like (a+b+c) and
753 /// (a+c+d) and computes the common (a+c) subexpression. The common expression
754 /// is *removed* from the Bases and returned.
756 RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses,
757 ScalarEvolution *SE, Loop *L,
758 const TargetLowering *TLI) {
759 unsigned NumUses = Uses.size();
761 // Only one use? This is a very common case, so we handle it specially and
763 const SCEV *Zero = SE->getIntegerSCEV(0, Uses[0].Base->getType());
764 const SCEV *Result = Zero;
765 const SCEV *FreeResult = Zero;
767 // If the use is inside the loop, use its base, regardless of what it is:
768 // it is clearly shared across all the IV's. If the use is outside the loop
769 // (which means after it) we don't want to factor anything *into* the loop,
770 // so just use 0 as the base.
771 if (L->contains(Uses[0].Inst->getParent()))
772 std::swap(Result, Uses[0].Base);
776 // To find common subexpressions, count how many of Uses use each expression.
777 // If any subexpressions are used Uses.size() times, they are common.
778 // Also track whether all uses of each expression can be moved into an
779 // an addressing mode "for free"; such expressions are left within the loop.
780 // struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; };
781 std::map<const SCEV *, SubExprUseData> SubExpressionUseData;
783 // UniqueSubExprs - Keep track of all of the subexpressions we see in the
784 // order we see them.
785 SmallVector<const SCEV *, 16> UniqueSubExprs;
787 SmallVector<const SCEV *, 16> SubExprs;
788 unsigned NumUsesInsideLoop = 0;
789 for (unsigned i = 0; i != NumUses; ++i) {
790 // If the user is outside the loop, just ignore it for base computation.
791 // Since the user is outside the loop, it must be *after* the loop (if it
792 // were before, it could not be based on the loop IV). We don't want users
793 // after the loop to affect base computation of values *inside* the loop,
794 // because we can always add their offsets to the result IV after the loop
795 // is done, ensuring we get good code inside the loop.
796 if (!L->contains(Uses[i].Inst->getParent()))
800 // If the base is zero (which is common), return zero now, there are no
802 if (Uses[i].Base == Zero) return Zero;
804 // If this use is as an address we may be able to put CSEs in the addressing
805 // mode rather than hoisting them.
806 bool isAddrUse = isAddressUse(Uses[i].Inst, Uses[i].OperandValToReplace);
807 // We may need the AccessTy below, but only when isAddrUse, so compute it
808 // only in that case.
809 const Type *AccessTy = 0;
811 AccessTy = getAccessType(Uses[i].Inst);
813 // Split the expression into subexprs.
814 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
815 // Add one to SubExpressionUseData.Count for each subexpr present, and
816 // if the subexpr is not a valid immediate within an addressing mode use,
817 // set SubExpressionUseData.notAllUsesAreFree. We definitely want to
818 // hoist these out of the loop (if they are common to all uses).
819 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) {
820 if (++SubExpressionUseData[SubExprs[j]].Count == 1)
821 UniqueSubExprs.push_back(SubExprs[j]);
822 if (!isAddrUse || !fitsInAddressMode(SubExprs[j], AccessTy, TLI, false))
823 SubExpressionUseData[SubExprs[j]].notAllUsesAreFree = true;
828 // Now that we know how many times each is used, build Result. Iterate over
829 // UniqueSubexprs so that we have a stable ordering.
830 for (unsigned i = 0, e = UniqueSubExprs.size(); i != e; ++i) {
831 std::map<const SCEV *, SubExprUseData>::iterator I =
832 SubExpressionUseData.find(UniqueSubExprs[i]);
833 assert(I != SubExpressionUseData.end() && "Entry not found?");
834 if (I->second.Count == NumUsesInsideLoop) { // Found CSE!
835 if (I->second.notAllUsesAreFree)
836 Result = SE->getAddExpr(Result, I->first);
838 FreeResult = SE->getAddExpr(FreeResult, I->first);
840 // Remove non-cse's from SubExpressionUseData.
841 SubExpressionUseData.erase(I);
844 if (FreeResult != Zero) {
845 // We have some subexpressions that can be subsumed into addressing
846 // modes in every use inside the loop. However, it's possible that
847 // there are so many of them that the combined FreeResult cannot
848 // be subsumed, or that the target cannot handle both a FreeResult
849 // and a Result in the same instruction (for example because it would
850 // require too many registers). Check this.
851 for (unsigned i=0; i<NumUses; ++i) {
852 if (!L->contains(Uses[i].Inst->getParent()))
854 // We know this is an addressing mode use; if there are any uses that
855 // are not, FreeResult would be Zero.
856 const Type *AccessTy = getAccessType(Uses[i].Inst);
857 if (!fitsInAddressMode(FreeResult, AccessTy, TLI, Result!=Zero)) {
858 // FIXME: could split up FreeResult into pieces here, some hoisted
859 // and some not. There is no obvious advantage to this.
860 Result = SE->getAddExpr(Result, FreeResult);
867 // If we found no CSE's, return now.
868 if (Result == Zero) return Result;
870 // If we still have a FreeResult, remove its subexpressions from
871 // SubExpressionUseData. This means they will remain in the use Bases.
872 if (FreeResult != Zero) {
873 SeparateSubExprs(SubExprs, FreeResult, SE);
874 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) {
875 std::map<const SCEV *, SubExprUseData>::iterator I =
876 SubExpressionUseData.find(SubExprs[j]);
877 SubExpressionUseData.erase(I);
882 // Otherwise, remove all of the CSE's we found from each of the base values.
883 for (unsigned i = 0; i != NumUses; ++i) {
884 // Uses outside the loop don't necessarily include the common base, but
885 // the final IV value coming into those uses does. Instead of trying to
886 // remove the pieces of the common base, which might not be there,
887 // subtract off the base to compensate for this.
888 if (!L->contains(Uses[i].Inst->getParent())) {
889 Uses[i].Base = SE->getMinusSCEV(Uses[i].Base, Result);
893 // Split the expression into subexprs.
894 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
896 // Remove any common subexpressions.
897 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
898 if (SubExpressionUseData.count(SubExprs[j])) {
899 SubExprs.erase(SubExprs.begin()+j);
903 // Finally, add the non-shared expressions together.
904 if (SubExprs.empty())
907 Uses[i].Base = SE->getAddExpr(SubExprs);
914 /// ValidScale - Check whether the given Scale is valid for all loads and
915 /// stores in UsersToProcess.
917 bool LoopStrengthReduce::ValidScale(bool HasBaseReg, int64_t Scale,
918 const std::vector<BasedUser>& UsersToProcess) {
922 for (unsigned i = 0, e = UsersToProcess.size(); i!=e; ++i) {
923 // If this is a load or other access, pass the type of the access in.
924 const Type *AccessTy =
925 Type::getVoidTy(UsersToProcess[i].Inst->getContext());
926 if (isAddressUse(UsersToProcess[i].Inst,
927 UsersToProcess[i].OperandValToReplace))
928 AccessTy = getAccessType(UsersToProcess[i].Inst);
929 else if (isa<PHINode>(UsersToProcess[i].Inst))
932 TargetLowering::AddrMode AM;
933 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
934 AM.BaseOffs = SC->getValue()->getSExtValue();
935 AM.HasBaseReg = HasBaseReg || !UsersToProcess[i].Base->isZero();
938 // If load[imm+r*scale] is illegal, bail out.
939 if (!TLI->isLegalAddressingMode(AM, AccessTy))
945 /// ValidOffset - Check whether the given Offset is valid for all loads and
946 /// stores in UsersToProcess.
948 bool LoopStrengthReduce::ValidOffset(bool HasBaseReg,
951 const std::vector<BasedUser>& UsersToProcess) {
955 for (unsigned i=0, e = UsersToProcess.size(); i!=e; ++i) {
956 // If this is a load or other access, pass the type of the access in.
957 const Type *AccessTy =
958 Type::getVoidTy(UsersToProcess[i].Inst->getContext());
959 if (isAddressUse(UsersToProcess[i].Inst,
960 UsersToProcess[i].OperandValToReplace))
961 AccessTy = getAccessType(UsersToProcess[i].Inst);
962 else if (isa<PHINode>(UsersToProcess[i].Inst))
965 TargetLowering::AddrMode AM;
966 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
967 AM.BaseOffs = SC->getValue()->getSExtValue();
968 AM.BaseOffs = (uint64_t)AM.BaseOffs + (uint64_t)Offset;
969 AM.HasBaseReg = HasBaseReg || !UsersToProcess[i].Base->isZero();
972 // If load[imm+r*scale] is illegal, bail out.
973 if (!TLI->isLegalAddressingMode(AM, AccessTy))
979 /// RequiresTypeConversion - Returns true if converting Ty1 to Ty2 is not
981 bool LoopStrengthReduce::RequiresTypeConversion(const Type *Ty1,
985 Ty1 = SE->getEffectiveSCEVType(Ty1);
986 Ty2 = SE->getEffectiveSCEVType(Ty2);
989 if (Ty1->canLosslesslyBitCastTo(Ty2))
991 if (TLI && TLI->isTruncateFree(Ty1, Ty2))
996 /// CheckForIVReuse - Returns the multiple if the stride is the multiple
997 /// of a previous stride and it is a legal value for the target addressing
998 /// mode scale component and optional base reg. This allows the users of
999 /// this stride to be rewritten as prev iv * factor. It returns 0 if no
1000 /// reuse is possible. Factors can be negative on same targets, e.g. ARM.
1002 /// If all uses are outside the loop, we don't require that all multiplies
1003 /// be folded into the addressing mode, nor even that the factor be constant;
1004 /// a multiply (executed once) outside the loop is better than another IV
1005 /// within. Well, usually.
1006 const SCEV *LoopStrengthReduce::CheckForIVReuse(bool HasBaseReg,
1007 bool AllUsesAreAddresses,
1008 bool AllUsesAreOutsideLoop,
1009 const SCEV *const &Stride,
1010 IVExpr &IV, const Type *Ty,
1011 const std::vector<BasedUser>& UsersToProcess) {
1012 if (StrideNoReuse.count(Stride))
1013 return SE->getIntegerSCEV(0, Stride->getType());
1015 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Stride)) {
1016 int64_t SInt = SC->getValue()->getSExtValue();
1017 for (unsigned NewStride = 0, e = IU->StrideOrder.size();
1018 NewStride != e; ++NewStride) {
1019 std::map<const SCEV *, IVsOfOneStride>::iterator SI =
1020 IVsByStride.find(IU->StrideOrder[NewStride]);
1021 if (SI == IVsByStride.end() || !isa<SCEVConstant>(SI->first) ||
1022 StrideNoReuse.count(SI->first))
1024 // The other stride has no uses, don't reuse it.
1025 std::map<const SCEV *, IVUsersOfOneStride *>::iterator UI =
1026 IU->IVUsesByStride.find(IU->StrideOrder[NewStride]);
1027 if (UI->second->Users.empty())
1029 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1030 if (SI->first != Stride &&
1031 (unsigned(abs64(SInt)) < SSInt || (SInt % SSInt) != 0))
1033 int64_t Scale = SInt / SSInt;
1034 // Check that this stride is valid for all the types used for loads and
1035 // stores; if it can be used for some and not others, we might as well use
1036 // the original stride everywhere, since we have to create the IV for it
1037 // anyway. If the scale is 1, then we don't need to worry about folding
1040 (AllUsesAreAddresses &&
1041 ValidScale(HasBaseReg, Scale, UsersToProcess))) {
1042 // Prefer to reuse an IV with a base of zero.
1043 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1044 IE = SI->second.IVs.end(); II != IE; ++II)
1045 // Only reuse previous IV if it would not require a type conversion
1046 // and if the base difference can be folded.
1047 if (II->Base->isZero() &&
1048 !RequiresTypeConversion(II->Base->getType(), Ty)) {
1050 return SE->getIntegerSCEV(Scale, Stride->getType());
1052 // Otherwise, settle for an IV with a foldable base.
1053 if (AllUsesAreAddresses)
1054 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1055 IE = SI->second.IVs.end(); II != IE; ++II)
1056 // Only reuse previous IV if it would not require a type conversion
1057 // and if the base difference can be folded.
1058 if (SE->getEffectiveSCEVType(II->Base->getType()) ==
1059 SE->getEffectiveSCEVType(Ty) &&
1060 isa<SCEVConstant>(II->Base)) {
1062 cast<SCEVConstant>(II->Base)->getValue()->getSExtValue();
1063 if (Base > INT32_MIN && Base <= INT32_MAX &&
1064 ValidOffset(HasBaseReg, -Base * Scale,
1065 Scale, UsersToProcess)) {
1067 return SE->getIntegerSCEV(Scale, Stride->getType());
1072 } else if (AllUsesAreOutsideLoop) {
1073 // Accept nonconstant strides here; it is really really right to substitute
1074 // an existing IV if we can.
1075 for (unsigned NewStride = 0, e = IU->StrideOrder.size();
1076 NewStride != e; ++NewStride) {
1077 std::map<const SCEV *, IVsOfOneStride>::iterator SI =
1078 IVsByStride.find(IU->StrideOrder[NewStride]);
1079 if (SI == IVsByStride.end() || !isa<SCEVConstant>(SI->first))
1081 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1082 if (SI->first != Stride && SSInt != 1)
1084 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1085 IE = SI->second.IVs.end(); II != IE; ++II)
1086 // Accept nonzero base here.
1087 // Only reuse previous IV if it would not require a type conversion.
1088 if (!RequiresTypeConversion(II->Base->getType(), Ty)) {
1093 // Special case, old IV is -1*x and this one is x. Can treat this one as
1095 for (unsigned NewStride = 0, e = IU->StrideOrder.size();
1096 NewStride != e; ++NewStride) {
1097 std::map<const SCEV *, IVsOfOneStride>::iterator SI =
1098 IVsByStride.find(IU->StrideOrder[NewStride]);
1099 if (SI == IVsByStride.end())
1101 if (const SCEVMulExpr *ME = dyn_cast<SCEVMulExpr>(SI->first))
1102 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(ME->getOperand(0)))
1103 if (Stride == ME->getOperand(1) &&
1104 SC->getValue()->getSExtValue() == -1LL)
1105 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1106 IE = SI->second.IVs.end(); II != IE; ++II)
1107 // Accept nonzero base here.
1108 // Only reuse previous IV if it would not require type conversion.
1109 if (!RequiresTypeConversion(II->Base->getType(), Ty)) {
1111 return SE->getIntegerSCEV(-1LL, Stride->getType());
1115 return SE->getIntegerSCEV(0, Stride->getType());
1118 /// PartitionByIsUseOfPostIncrementedValue - Simple boolean predicate that
1119 /// returns true if Val's isUseOfPostIncrementedValue is true.
1120 static bool PartitionByIsUseOfPostIncrementedValue(const BasedUser &Val) {
1121 return Val.isUseOfPostIncrementedValue;
1124 /// isNonConstantNegative - Return true if the specified scev is negated, but
1126 static bool isNonConstantNegative(const SCEV *const &Expr) {
1127 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Expr);
1128 if (!Mul) return false;
1130 // If there is a constant factor, it will be first.
1131 const SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
1132 if (!SC) return false;
1134 // Return true if the value is negative, this matches things like (-42 * V).
1135 return SC->getValue()->getValue().isNegative();
1138 /// CollectIVUsers - Transform our list of users and offsets to a bit more
1139 /// complex table. In this new vector, each 'BasedUser' contains 'Base', the
1140 /// base of the strided accesses, as well as the old information from Uses. We
1141 /// progressively move information from the Base field to the Imm field, until
1142 /// we eventually have the full access expression to rewrite the use.
1143 const SCEV *LoopStrengthReduce::CollectIVUsers(const SCEV *const &Stride,
1144 IVUsersOfOneStride &Uses,
1146 bool &AllUsesAreAddresses,
1147 bool &AllUsesAreOutsideLoop,
1148 std::vector<BasedUser> &UsersToProcess) {
1149 // FIXME: Generalize to non-affine IV's.
1150 if (!Stride->isLoopInvariant(L))
1151 return SE->getIntegerSCEV(0, Stride->getType());
1153 UsersToProcess.reserve(Uses.Users.size());
1154 for (ilist<IVStrideUse>::iterator I = Uses.Users.begin(),
1155 E = Uses.Users.end(); I != E; ++I) {
1156 UsersToProcess.push_back(BasedUser(*I, SE));
1158 // Move any loop variant operands from the offset field to the immediate
1159 // field of the use, so that we don't try to use something before it is
1161 MoveLoopVariantsToImmediateField(UsersToProcess.back().Base,
1162 UsersToProcess.back().Imm, L, SE);
1163 assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
1164 "Base value is not loop invariant!");
1167 // We now have a whole bunch of uses of like-strided induction variables, but
1168 // they might all have different bases. We want to emit one PHI node for this
1169 // stride which we fold as many common expressions (between the IVs) into as
1170 // possible. Start by identifying the common expressions in the base values
1171 // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
1172 // "A+B"), emit it to the preheader, then remove the expression from the
1173 // UsersToProcess base values.
1174 const SCEV *CommonExprs =
1175 RemoveCommonExpressionsFromUseBases(UsersToProcess, SE, L, TLI);
1177 // Next, figure out what we can represent in the immediate fields of
1178 // instructions. If we can represent anything there, move it to the imm
1179 // fields of the BasedUsers. We do this so that it increases the commonality
1180 // of the remaining uses.
1181 unsigned NumPHI = 0;
1182 bool HasAddress = false;
1183 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1184 // If the user is not in the current loop, this means it is using the exit
1185 // value of the IV. Do not put anything in the base, make sure it's all in
1186 // the immediate field to allow as much factoring as possible.
1187 if (!L->contains(UsersToProcess[i].Inst->getParent())) {
1188 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm,
1189 UsersToProcess[i].Base);
1190 UsersToProcess[i].Base =
1191 SE->getIntegerSCEV(0, UsersToProcess[i].Base->getType());
1193 // Not all uses are outside the loop.
1194 AllUsesAreOutsideLoop = false;
1196 // Addressing modes can be folded into loads and stores. Be careful that
1197 // the store is through the expression, not of the expression though.
1199 bool isAddress = isAddressUse(UsersToProcess[i].Inst,
1200 UsersToProcess[i].OperandValToReplace);
1201 if (isa<PHINode>(UsersToProcess[i].Inst)) {
1209 // If this use isn't an address, then not all uses are addresses.
1210 if (!isAddress && !isPHI)
1211 AllUsesAreAddresses = false;
1213 MoveImmediateValues(TLI, UsersToProcess[i].Inst, UsersToProcess[i].Base,
1214 UsersToProcess[i].Imm, isAddress, L, SE);
1218 // If one of the use is a PHI node and all other uses are addresses, still
1219 // allow iv reuse. Essentially we are trading one constant multiplication
1220 // for one fewer iv.
1222 AllUsesAreAddresses = false;
1224 // There are no in-loop address uses.
1225 if (AllUsesAreAddresses && (!HasAddress && !AllUsesAreOutsideLoop))
1226 AllUsesAreAddresses = false;
1231 /// ShouldUseFullStrengthReductionMode - Test whether full strength-reduction
1232 /// is valid and profitable for the given set of users of a stride. In
1233 /// full strength-reduction mode, all addresses at the current stride are
1234 /// strength-reduced all the way down to pointer arithmetic.
1236 bool LoopStrengthReduce::ShouldUseFullStrengthReductionMode(
1237 const std::vector<BasedUser> &UsersToProcess,
1239 bool AllUsesAreAddresses,
1240 const SCEV *Stride) {
1241 if (!EnableFullLSRMode)
1244 // The heuristics below aim to avoid increasing register pressure, but
1245 // fully strength-reducing all the addresses increases the number of
1246 // add instructions, so don't do this when optimizing for size.
1247 // TODO: If the loop is large, the savings due to simpler addresses
1248 // may oughtweight the costs of the extra increment instructions.
1249 if (L->getHeader()->getParent()->hasFnAttr(Attribute::OptimizeForSize))
1252 // TODO: For now, don't do full strength reduction if there could
1253 // potentially be greater-stride multiples of the current stride
1254 // which could reuse the current stride IV.
1255 if (IU->StrideOrder.back() != Stride)
1258 // Iterate through the uses to find conditions that automatically rule out
1260 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ) {
1261 const SCEV *Base = UsersToProcess[i].Base;
1262 const SCEV *Imm = UsersToProcess[i].Imm;
1263 // If any users have a loop-variant component, they can't be fully
1264 // strength-reduced.
1265 if (Imm && !Imm->isLoopInvariant(L))
1267 // If there are to users with the same base and the difference between
1268 // the two Imm values can't be folded into the address, full
1269 // strength reduction would increase register pressure.
1271 const SCEV *CurImm = UsersToProcess[i].Imm;
1272 if ((CurImm || Imm) && CurImm != Imm) {
1273 if (!CurImm) CurImm = SE->getIntegerSCEV(0, Stride->getType());
1274 if (!Imm) Imm = SE->getIntegerSCEV(0, Stride->getType());
1275 const Instruction *Inst = UsersToProcess[i].Inst;
1276 const Type *AccessTy = getAccessType(Inst);
1277 const SCEV *Diff = SE->getMinusSCEV(UsersToProcess[i].Imm, Imm);
1278 if (!Diff->isZero() &&
1279 (!AllUsesAreAddresses ||
1280 !fitsInAddressMode(Diff, AccessTy, TLI, /*HasBaseReg=*/true)))
1283 } while (++i != e && Base == UsersToProcess[i].Base);
1286 // If there's exactly one user in this stride, fully strength-reducing it
1287 // won't increase register pressure. If it's starting from a non-zero base,
1288 // it'll be simpler this way.
1289 if (UsersToProcess.size() == 1 && !UsersToProcess[0].Base->isZero())
1292 // Otherwise, if there are any users in this stride that don't require
1293 // a register for their base, full strength-reduction will increase
1294 // register pressure.
1295 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1296 if (UsersToProcess[i].Base->isZero())
1299 // Otherwise, go for it.
1303 /// InsertAffinePhi Create and insert a PHI node for an induction variable
1304 /// with the specified start and step values in the specified loop.
1306 /// If NegateStride is true, the stride should be negated by using a
1307 /// subtract instead of an add.
1309 /// Return the created phi node.
1311 static PHINode *InsertAffinePhi(const SCEV *Start, const SCEV *Step,
1312 Instruction *IVIncInsertPt,
1314 SCEVExpander &Rewriter) {
1315 assert(Start->isLoopInvariant(L) && "New PHI start is not loop invariant!");
1316 assert(Step->isLoopInvariant(L) && "New PHI stride is not loop invariant!");
1318 BasicBlock *Header = L->getHeader();
1319 BasicBlock *Preheader = L->getLoopPreheader();
1320 BasicBlock *LatchBlock = L->getLoopLatch();
1321 const Type *Ty = Start->getType();
1322 Ty = Rewriter.SE.getEffectiveSCEVType(Ty);
1324 PHINode *PN = PHINode::Create(Ty, "lsr.iv", Header->begin());
1325 PN->addIncoming(Rewriter.expandCodeFor(Start, Ty, Preheader->getTerminator()),
1328 // If the stride is negative, insert a sub instead of an add for the
1330 bool isNegative = isNonConstantNegative(Step);
1331 const SCEV *IncAmount = Step;
1333 IncAmount = Rewriter.SE.getNegativeSCEV(Step);
1335 // Insert an add instruction right before the terminator corresponding
1336 // to the back-edge or just before the only use. The location is determined
1337 // by the caller and passed in as IVIncInsertPt.
1338 Value *StepV = Rewriter.expandCodeFor(IncAmount, Ty,
1339 Preheader->getTerminator());
1342 IncV = BinaryOperator::CreateSub(PN, StepV, "lsr.iv.next",
1345 IncV = BinaryOperator::CreateAdd(PN, StepV, "lsr.iv.next",
1348 if (!isa<ConstantInt>(StepV)) ++NumVariable;
1350 PN->addIncoming(IncV, LatchBlock);
1356 static void SortUsersToProcess(std::vector<BasedUser> &UsersToProcess) {
1357 // We want to emit code for users inside the loop first. To do this, we
1358 // rearrange BasedUser so that the entries at the end have
1359 // isUseOfPostIncrementedValue = false, because we pop off the end of the
1360 // vector (so we handle them first).
1361 std::partition(UsersToProcess.begin(), UsersToProcess.end(),
1362 PartitionByIsUseOfPostIncrementedValue);
1364 // Sort this by base, so that things with the same base are handled
1365 // together. By partitioning first and stable-sorting later, we are
1366 // guaranteed that within each base we will pop off users from within the
1367 // loop before users outside of the loop with a particular base.
1369 // We would like to use stable_sort here, but we can't. The problem is that
1370 // const SCEV *'s don't have a deterministic ordering w.r.t to each other, so
1371 // we don't have anything to do a '<' comparison on. Because we think the
1372 // number of uses is small, do a horrible bubble sort which just relies on
1374 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1375 // Get a base value.
1376 const SCEV *Base = UsersToProcess[i].Base;
1378 // Compact everything with this base to be consecutive with this one.
1379 for (unsigned j = i+1; j != e; ++j) {
1380 if (UsersToProcess[j].Base == Base) {
1381 std::swap(UsersToProcess[i+1], UsersToProcess[j]);
1388 /// PrepareToStrengthReduceFully - Prepare to fully strength-reduce
1389 /// UsersToProcess, meaning lowering addresses all the way down to direct
1390 /// pointer arithmetic.
1393 LoopStrengthReduce::PrepareToStrengthReduceFully(
1394 std::vector<BasedUser> &UsersToProcess,
1396 const SCEV *CommonExprs,
1398 SCEVExpander &PreheaderRewriter) {
1399 DEBUG(errs() << " Fully reducing all users\n");
1401 // Rewrite the UsersToProcess records, creating a separate PHI for each
1402 // unique Base value.
1403 Instruction *IVIncInsertPt = L->getLoopLatch()->getTerminator();
1404 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ) {
1405 // TODO: The uses are grouped by base, but not sorted. We arbitrarily
1406 // pick the first Imm value here to start with, and adjust it for the
1408 const SCEV *Imm = UsersToProcess[i].Imm;
1409 const SCEV *Base = UsersToProcess[i].Base;
1410 const SCEV *Start = SE->getAddExpr(CommonExprs, Base, Imm);
1411 PHINode *Phi = InsertAffinePhi(Start, Stride, IVIncInsertPt, L,
1413 // Loop over all the users with the same base.
1415 UsersToProcess[i].Base = SE->getIntegerSCEV(0, Stride->getType());
1416 UsersToProcess[i].Imm = SE->getMinusSCEV(UsersToProcess[i].Imm, Imm);
1417 UsersToProcess[i].Phi = Phi;
1418 assert(UsersToProcess[i].Imm->isLoopInvariant(L) &&
1419 "ShouldUseFullStrengthReductionMode should reject this!");
1420 } while (++i != e && Base == UsersToProcess[i].Base);
1424 /// FindIVIncInsertPt - Return the location to insert the increment instruction.
1425 /// If the only use if a use of postinc value, (must be the loop termination
1426 /// condition), then insert it just before the use.
1427 static Instruction *FindIVIncInsertPt(std::vector<BasedUser> &UsersToProcess,
1429 if (UsersToProcess.size() == 1 &&
1430 UsersToProcess[0].isUseOfPostIncrementedValue &&
1431 L->contains(UsersToProcess[0].Inst->getParent()))
1432 return UsersToProcess[0].Inst;
1433 return L->getLoopLatch()->getTerminator();
1436 /// PrepareToStrengthReduceWithNewPhi - Insert a new induction variable for the
1437 /// given users to share.
1440 LoopStrengthReduce::PrepareToStrengthReduceWithNewPhi(
1441 std::vector<BasedUser> &UsersToProcess,
1443 const SCEV *CommonExprs,
1445 Instruction *IVIncInsertPt,
1447 SCEVExpander &PreheaderRewriter) {
1448 DEBUG(errs() << " Inserting new PHI:\n");
1450 PHINode *Phi = InsertAffinePhi(SE->getUnknown(CommonBaseV),
1451 Stride, IVIncInsertPt, L,
1454 // Remember this in case a later stride is multiple of this.
1455 IVsByStride[Stride].addIV(Stride, CommonExprs, Phi);
1457 // All the users will share this new IV.
1458 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1459 UsersToProcess[i].Phi = Phi;
1461 DEBUG(errs() << " IV=");
1462 DEBUG(WriteAsOperand(errs(), Phi, /*PrintType=*/false));
1463 DEBUG(errs() << "\n");
1466 /// PrepareToStrengthReduceFromSmallerStride - Prepare for the given users to
1467 /// reuse an induction variable with a stride that is a factor of the current
1468 /// induction variable.
1471 LoopStrengthReduce::PrepareToStrengthReduceFromSmallerStride(
1472 std::vector<BasedUser> &UsersToProcess,
1474 const IVExpr &ReuseIV,
1475 Instruction *PreInsertPt) {
1476 DEBUG(errs() << " Rewriting in terms of existing IV of STRIDE "
1477 << *ReuseIV.Stride << " and BASE " << *ReuseIV.Base << "\n");
1479 // All the users will share the reused IV.
1480 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1481 UsersToProcess[i].Phi = ReuseIV.PHI;
1483 Constant *C = dyn_cast<Constant>(CommonBaseV);
1485 (!C->isNullValue() &&
1486 !fitsInAddressMode(SE->getUnknown(CommonBaseV), CommonBaseV->getType(),
1488 // We want the common base emitted into the preheader! This is just
1489 // using cast as a copy so BitCast (no-op cast) is appropriate
1490 CommonBaseV = new BitCastInst(CommonBaseV, CommonBaseV->getType(),
1491 "commonbase", PreInsertPt);
1494 static bool IsImmFoldedIntoAddrMode(GlobalValue *GV, int64_t Offset,
1495 const Type *AccessTy,
1496 std::vector<BasedUser> &UsersToProcess,
1497 const TargetLowering *TLI) {
1498 SmallVector<Instruction*, 16> AddrModeInsts;
1499 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1500 if (UsersToProcess[i].isUseOfPostIncrementedValue)
1502 ExtAddrMode AddrMode =
1503 AddressingModeMatcher::Match(UsersToProcess[i].OperandValToReplace,
1504 AccessTy, UsersToProcess[i].Inst,
1505 AddrModeInsts, *TLI);
1506 if (GV && GV != AddrMode.BaseGV)
1508 if (Offset && !AddrMode.BaseOffs)
1509 // FIXME: How to accurate check it's immediate offset is folded.
1511 AddrModeInsts.clear();
1516 /// StrengthReduceIVUsersOfStride - Strength reduce all of the users of a single
1517 /// stride of IV. All of the users may have different starting values, and this
1518 /// may not be the only stride.
1520 LoopStrengthReduce::StrengthReduceIVUsersOfStride(const SCEV *const &Stride,
1521 IVUsersOfOneStride &Uses,
1523 // If all the users are moved to another stride, then there is nothing to do.
1524 if (Uses.Users.empty())
1527 // Keep track if every use in UsersToProcess is an address. If they all are,
1528 // we may be able to rewrite the entire collection of them in terms of a
1529 // smaller-stride IV.
1530 bool AllUsesAreAddresses = true;
1532 // Keep track if every use of a single stride is outside the loop. If so,
1533 // we want to be more aggressive about reusing a smaller-stride IV; a
1534 // multiply outside the loop is better than another IV inside. Well, usually.
1535 bool AllUsesAreOutsideLoop = true;
1537 // Transform our list of users and offsets to a bit more complex table. In
1538 // this new vector, each 'BasedUser' contains 'Base' the base of the
1539 // strided accessas well as the old information from Uses. We progressively
1540 // move information from the Base field to the Imm field, until we eventually
1541 // have the full access expression to rewrite the use.
1542 std::vector<BasedUser> UsersToProcess;
1543 const SCEV *CommonExprs = CollectIVUsers(Stride, Uses, L, AllUsesAreAddresses,
1544 AllUsesAreOutsideLoop,
1547 // Sort the UsersToProcess array so that users with common bases are
1548 // next to each other.
1549 SortUsersToProcess(UsersToProcess);
1551 // If we managed to find some expressions in common, we'll need to carry
1552 // their value in a register and add it in for each use. This will take up
1553 // a register operand, which potentially restricts what stride values are
1555 bool HaveCommonExprs = !CommonExprs->isZero();
1556 const Type *ReplacedTy = CommonExprs->getType();
1558 // If all uses are addresses, consider sinking the immediate part of the
1559 // common expression back into uses if they can fit in the immediate fields.
1560 if (TLI && HaveCommonExprs && AllUsesAreAddresses) {
1561 const SCEV *NewCommon = CommonExprs;
1562 const SCEV *Imm = SE->getIntegerSCEV(0, ReplacedTy);
1563 MoveImmediateValues(TLI, Type::getVoidTy(
1564 L->getLoopPreheader()->getContext()),
1565 NewCommon, Imm, true, L, SE);
1566 if (!Imm->isZero()) {
1569 // If the immediate part of the common expression is a GV, check if it's
1570 // possible to fold it into the target addressing mode.
1571 GlobalValue *GV = 0;
1572 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(Imm))
1573 GV = dyn_cast<GlobalValue>(SU->getValue());
1575 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Imm))
1576 Offset = SC->getValue()->getSExtValue();
1578 // Pass VoidTy as the AccessTy to be conservative, because
1579 // there could be multiple access types among all the uses.
1580 DoSink = IsImmFoldedIntoAddrMode(GV, Offset,
1581 Type::getVoidTy(L->getLoopPreheader()->getContext()),
1582 UsersToProcess, TLI);
1585 DEBUG(errs() << " Sinking " << *Imm << " back down into uses\n");
1586 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1587 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm, Imm);
1588 CommonExprs = NewCommon;
1589 HaveCommonExprs = !CommonExprs->isZero();
1595 // Now that we know what we need to do, insert the PHI node itself.
1597 DEBUG(errs() << "LSR: Examining IVs of TYPE " << *ReplacedTy << " of STRIDE "
1599 << " Common base: " << *CommonExprs << "\n");
1601 SCEVExpander Rewriter(*SE);
1602 SCEVExpander PreheaderRewriter(*SE);
1604 BasicBlock *Preheader = L->getLoopPreheader();
1605 Instruction *PreInsertPt = Preheader->getTerminator();
1606 BasicBlock *LatchBlock = L->getLoopLatch();
1607 Instruction *IVIncInsertPt = LatchBlock->getTerminator();
1609 Value *CommonBaseV = Constant::getNullValue(ReplacedTy);
1611 const SCEV *RewriteFactor = SE->getIntegerSCEV(0, ReplacedTy);
1612 IVExpr ReuseIV(SE->getIntegerSCEV(0,
1613 Type::getInt32Ty(Preheader->getContext())),
1614 SE->getIntegerSCEV(0,
1615 Type::getInt32Ty(Preheader->getContext())),
1618 // Choose a strength-reduction strategy and prepare for it by creating
1619 // the necessary PHIs and adjusting the bookkeeping.
1620 if (ShouldUseFullStrengthReductionMode(UsersToProcess, L,
1621 AllUsesAreAddresses, Stride)) {
1622 PrepareToStrengthReduceFully(UsersToProcess, Stride, CommonExprs, L,
1625 // Emit the initial base value into the loop preheader.
1626 CommonBaseV = PreheaderRewriter.expandCodeFor(CommonExprs, ReplacedTy,
1629 // If all uses are addresses, check if it is possible to reuse an IV. The
1630 // new IV must have a stride that is a multiple of the old stride; the
1631 // multiple must be a number that can be encoded in the scale field of the
1632 // target addressing mode; and we must have a valid instruction after this
1633 // substitution, including the immediate field, if any.
1634 RewriteFactor = CheckForIVReuse(HaveCommonExprs, AllUsesAreAddresses,
1635 AllUsesAreOutsideLoop,
1636 Stride, ReuseIV, ReplacedTy,
1638 if (!RewriteFactor->isZero())
1639 PrepareToStrengthReduceFromSmallerStride(UsersToProcess, CommonBaseV,
1640 ReuseIV, PreInsertPt);
1642 IVIncInsertPt = FindIVIncInsertPt(UsersToProcess, L);
1643 PrepareToStrengthReduceWithNewPhi(UsersToProcess, Stride, CommonExprs,
1644 CommonBaseV, IVIncInsertPt,
1645 L, PreheaderRewriter);
1649 // Process all the users now, replacing their strided uses with
1650 // strength-reduced forms. This outer loop handles all bases, the inner
1651 // loop handles all users of a particular base.
1652 while (!UsersToProcess.empty()) {
1653 const SCEV *Base = UsersToProcess.back().Base;
1654 Instruction *Inst = UsersToProcess.back().Inst;
1656 // Emit the code for Base into the preheader.
1658 if (!Base->isZero()) {
1659 BaseV = PreheaderRewriter.expandCodeFor(Base, 0, PreInsertPt);
1661 DEBUG(errs() << " INSERTING code for BASE = " << *Base << ":");
1662 if (BaseV->hasName())
1663 DEBUG(errs() << " Result value name = %" << BaseV->getName());
1664 DEBUG(errs() << "\n");
1666 // If BaseV is a non-zero constant, make sure that it gets inserted into
1667 // the preheader, instead of being forward substituted into the uses. We
1668 // do this by forcing a BitCast (noop cast) to be inserted into the
1669 // preheader in this case.
1670 if (!fitsInAddressMode(Base, getAccessType(Inst), TLI, false) &&
1671 isa<Constant>(BaseV)) {
1672 // We want this constant emitted into the preheader! This is just
1673 // using cast as a copy so BitCast (no-op cast) is appropriate
1674 BaseV = new BitCastInst(BaseV, BaseV->getType(), "preheaderinsert",
1679 // Emit the code to add the immediate offset to the Phi value, just before
1680 // the instructions that we identified as using this stride and base.
1682 // FIXME: Use emitted users to emit other users.
1683 BasedUser &User = UsersToProcess.back();
1685 DEBUG(errs() << " Examining ");
1686 if (User.isUseOfPostIncrementedValue)
1687 DEBUG(errs() << "postinc");
1689 DEBUG(errs() << "preinc");
1690 DEBUG(errs() << " use ");
1691 DEBUG(WriteAsOperand(errs(), UsersToProcess.back().OperandValToReplace,
1692 /*PrintType=*/false));
1693 DEBUG(errs() << " in Inst: " << *User.Inst);
1695 // If this instruction wants to use the post-incremented value, move it
1696 // after the post-inc and use its value instead of the PHI.
1697 Value *RewriteOp = User.Phi;
1698 if (User.isUseOfPostIncrementedValue) {
1699 RewriteOp = User.Phi->getIncomingValueForBlock(LatchBlock);
1700 // If this user is in the loop, make sure it is the last thing in the
1701 // loop to ensure it is dominated by the increment. In case it's the
1702 // only use of the iv, the increment instruction is already before the
1704 if (L->contains(User.Inst->getParent()) && User.Inst != IVIncInsertPt)
1705 User.Inst->moveBefore(IVIncInsertPt);
1708 const SCEV *RewriteExpr = SE->getUnknown(RewriteOp);
1710 if (SE->getEffectiveSCEVType(RewriteOp->getType()) !=
1711 SE->getEffectiveSCEVType(ReplacedTy)) {
1712 assert(SE->getTypeSizeInBits(RewriteOp->getType()) >
1713 SE->getTypeSizeInBits(ReplacedTy) &&
1714 "Unexpected widening cast!");
1715 RewriteExpr = SE->getTruncateExpr(RewriteExpr, ReplacedTy);
1718 // If we had to insert new instructions for RewriteOp, we have to
1719 // consider that they may not have been able to end up immediately
1720 // next to RewriteOp, because non-PHI instructions may never precede
1721 // PHI instructions in a block. In this case, remember where the last
1722 // instruction was inserted so that if we're replacing a different
1723 // PHI node, we can use the later point to expand the final
1725 Instruction *NewBasePt = dyn_cast<Instruction>(RewriteOp);
1726 if (RewriteOp == User.Phi) NewBasePt = 0;
1728 // Clear the SCEVExpander's expression map so that we are guaranteed
1729 // to have the code emitted where we expect it.
1732 // If we are reusing the iv, then it must be multiplied by a constant
1733 // factor to take advantage of the addressing mode scale component.
1734 if (!RewriteFactor->isZero()) {
1735 // If we're reusing an IV with a nonzero base (currently this happens
1736 // only when all reuses are outside the loop) subtract that base here.
1737 // The base has been used to initialize the PHI node but we don't want
1739 if (!ReuseIV.Base->isZero()) {
1740 const SCEV *typedBase = ReuseIV.Base;
1741 if (SE->getEffectiveSCEVType(RewriteExpr->getType()) !=
1742 SE->getEffectiveSCEVType(ReuseIV.Base->getType())) {
1743 // It's possible the original IV is a larger type than the new IV,
1744 // in which case we have to truncate the Base. We checked in
1745 // RequiresTypeConversion that this is valid.
1746 assert(SE->getTypeSizeInBits(RewriteExpr->getType()) <
1747 SE->getTypeSizeInBits(ReuseIV.Base->getType()) &&
1748 "Unexpected lengthening conversion!");
1749 typedBase = SE->getTruncateExpr(ReuseIV.Base,
1750 RewriteExpr->getType());
1752 RewriteExpr = SE->getMinusSCEV(RewriteExpr, typedBase);
1755 // Multiply old variable, with base removed, by new scale factor.
1756 RewriteExpr = SE->getMulExpr(RewriteFactor,
1759 // The common base is emitted in the loop preheader. But since we
1760 // are reusing an IV, it has not been used to initialize the PHI node.
1761 // Add it to the expression used to rewrite the uses.
1762 // When this use is outside the loop, we earlier subtracted the
1763 // common base, and are adding it back here. Use the same expression
1764 // as before, rather than CommonBaseV, so DAGCombiner will zap it.
1765 if (!CommonExprs->isZero()) {
1766 if (L->contains(User.Inst->getParent()))
1767 RewriteExpr = SE->getAddExpr(RewriteExpr,
1768 SE->getUnknown(CommonBaseV));
1770 RewriteExpr = SE->getAddExpr(RewriteExpr, CommonExprs);
1774 // Now that we know what we need to do, insert code before User for the
1775 // immediate and any loop-variant expressions.
1777 // Add BaseV to the PHI value if needed.
1778 RewriteExpr = SE->getAddExpr(RewriteExpr, SE->getUnknown(BaseV));
1780 User.RewriteInstructionToUseNewBase(RewriteExpr, NewBasePt,
1781 Rewriter, L, this, *LI,
1784 // Mark old value we replaced as possibly dead, so that it is eliminated
1785 // if we just replaced the last use of that value.
1786 DeadInsts.push_back(User.OperandValToReplace);
1788 UsersToProcess.pop_back();
1791 // If there are any more users to process with the same base, process them
1792 // now. We sorted by base above, so we just have to check the last elt.
1793 } while (!UsersToProcess.empty() && UsersToProcess.back().Base == Base);
1794 // TODO: Next, find out which base index is the most common, pull it out.
1797 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
1798 // different starting values, into different PHIs.
1801 void LoopStrengthReduce::StrengthReduceIVUsers(Loop *L) {
1802 // Note: this processes each stride/type pair individually. All users
1803 // passed into StrengthReduceIVUsersOfStride have the same type AND stride.
1804 // Also, note that we iterate over IVUsesByStride indirectly by using
1805 // StrideOrder. This extra layer of indirection makes the ordering of
1806 // strides deterministic - not dependent on map order.
1807 for (unsigned Stride = 0, e = IU->StrideOrder.size(); Stride != e; ++Stride) {
1808 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
1809 IU->IVUsesByStride.find(IU->StrideOrder[Stride]);
1810 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
1811 // FIXME: Generalize to non-affine IV's.
1812 if (!SI->first->isLoopInvariant(L))
1814 StrengthReduceIVUsersOfStride(SI->first, *SI->second, L);
1818 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1819 /// set the IV user and stride information and return true, otherwise return
1821 bool LoopStrengthReduce::FindIVUserForCond(ICmpInst *Cond,
1822 IVStrideUse *&CondUse,
1823 const SCEV* &CondStride) {
1824 for (unsigned Stride = 0, e = IU->StrideOrder.size();
1825 Stride != e && !CondUse; ++Stride) {
1826 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
1827 IU->IVUsesByStride.find(IU->StrideOrder[Stride]);
1828 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
1830 for (ilist<IVStrideUse>::iterator UI = SI->second->Users.begin(),
1831 E = SI->second->Users.end(); UI != E; ++UI)
1832 if (UI->getUser() == Cond) {
1833 // NOTE: we could handle setcc instructions with multiple uses here, but
1834 // InstCombine does it as well for simple uses, it's not clear that it
1835 // occurs enough in real life to handle.
1837 CondStride = SI->first;
1845 // Constant strides come first which in turns are sorted by their absolute
1846 // values. If absolute values are the same, then positive strides comes first.
1848 // 4, -1, X, 1, 2 ==> 1, -1, 2, 4, X
1849 struct StrideCompare {
1850 const ScalarEvolution *SE;
1851 explicit StrideCompare(const ScalarEvolution *se) : SE(se) {}
1853 bool operator()(const SCEV *const &LHS, const SCEV *const &RHS) {
1854 const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS);
1855 const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS);
1857 int64_t LV = LHSC->getValue()->getSExtValue();
1858 int64_t RV = RHSC->getValue()->getSExtValue();
1859 uint64_t ALV = (LV < 0) ? -LV : LV;
1860 uint64_t ARV = (RV < 0) ? -RV : RV;
1868 // If it's the same value but different type, sort by bit width so
1869 // that we emit larger induction variables before smaller
1870 // ones, letting the smaller be re-written in terms of larger ones.
1871 return SE->getTypeSizeInBits(RHS->getType()) <
1872 SE->getTypeSizeInBits(LHS->getType());
1874 return LHSC && !RHSC;
1879 /// ChangeCompareStride - If a loop termination compare instruction is the
1880 /// only use of its stride, and the compaison is against a constant value,
1881 /// try eliminate the stride by moving the compare instruction to another
1882 /// stride and change its constant operand accordingly. e.g.
1888 /// if (v2 < 10) goto loop
1893 /// if (v1 < 30) goto loop
1894 ICmpInst *LoopStrengthReduce::ChangeCompareStride(Loop *L, ICmpInst *Cond,
1895 IVStrideUse* &CondUse,
1896 const SCEV* &CondStride,
1898 // If there's only one stride in the loop, there's nothing to do here.
1899 if (IU->StrideOrder.size() < 2)
1901 // If there are other users of the condition's stride, don't bother
1902 // trying to change the condition because the stride will still
1904 std::map<const SCEV *, IVUsersOfOneStride *>::iterator I =
1905 IU->IVUsesByStride.find(CondStride);
1906 if (I == IU->IVUsesByStride.end())
1908 if (I->second->Users.size() > 1) {
1909 for (ilist<IVStrideUse>::iterator II = I->second->Users.begin(),
1910 EE = I->second->Users.end(); II != EE; ++II) {
1911 if (II->getUser() == Cond)
1913 if (!isInstructionTriviallyDead(II->getUser()))
1917 // Only handle constant strides for now.
1918 const SCEVConstant *SC = dyn_cast<SCEVConstant>(CondStride);
1919 if (!SC) return Cond;
1921 ICmpInst::Predicate Predicate = Cond->getPredicate();
1922 int64_t CmpSSInt = SC->getValue()->getSExtValue();
1923 unsigned BitWidth = SE->getTypeSizeInBits(CondStride->getType());
1924 uint64_t SignBit = 1ULL << (BitWidth-1);
1925 const Type *CmpTy = Cond->getOperand(0)->getType();
1926 const Type *NewCmpTy = NULL;
1927 unsigned TyBits = SE->getTypeSizeInBits(CmpTy);
1928 unsigned NewTyBits = 0;
1929 const SCEV *NewStride = NULL;
1930 Value *NewCmpLHS = NULL;
1931 Value *NewCmpRHS = NULL;
1933 const SCEV *NewOffset = SE->getIntegerSCEV(0, CmpTy);
1935 if (ConstantInt *C = dyn_cast<ConstantInt>(Cond->getOperand(1))) {
1936 int64_t CmpVal = C->getValue().getSExtValue();
1938 // Check the relevant induction variable for conformance to
1940 const SCEV *IV = SE->getSCEV(Cond->getOperand(0));
1941 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1942 if (!AR || !AR->isAffine())
1945 const SCEVConstant *StartC = dyn_cast<SCEVConstant>(AR->getStart());
1946 // Check stride constant and the comparision constant signs to detect
1949 if ((StartC->getValue()->getSExtValue() < CmpVal && CmpSSInt < 0) ||
1950 (StartC->getValue()->getSExtValue() > CmpVal && CmpSSInt > 0))
1953 // More restrictive check for the other cases.
1954 if ((CmpVal & SignBit) != (CmpSSInt & SignBit))
1958 // Look for a suitable stride / iv as replacement.
1959 for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
1960 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
1961 IU->IVUsesByStride.find(IU->StrideOrder[i]);
1962 if (!isa<SCEVConstant>(SI->first) || SI->second->Users.empty())
1964 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1965 if (SSInt == CmpSSInt ||
1966 abs64(SSInt) < abs64(CmpSSInt) ||
1967 (SSInt % CmpSSInt) != 0)
1970 Scale = SSInt / CmpSSInt;
1971 int64_t NewCmpVal = CmpVal * Scale;
1973 // If old icmp value fits in icmp immediate field, but the new one doesn't
1974 // try something else.
1976 TLI->isLegalICmpImmediate(CmpVal) &&
1977 !TLI->isLegalICmpImmediate(NewCmpVal))
1980 APInt Mul = APInt(BitWidth*2, CmpVal, true);
1981 Mul = Mul * APInt(BitWidth*2, Scale, true);
1982 // Check for overflow.
1983 if (!Mul.isSignedIntN(BitWidth))
1985 // Check for overflow in the stride's type too.
1986 if (!Mul.isSignedIntN(SE->getTypeSizeInBits(SI->first->getType())))
1989 // Watch out for overflow.
1990 if (ICmpInst::isSigned(Predicate) &&
1991 (CmpVal & SignBit) != (NewCmpVal & SignBit))
1994 // Pick the best iv to use trying to avoid a cast.
1996 for (ilist<IVStrideUse>::iterator UI = SI->second->Users.begin(),
1997 E = SI->second->Users.end(); UI != E; ++UI) {
1998 Value *Op = UI->getOperandValToReplace();
2000 // If the IVStrideUse implies a cast, check for an actual cast which
2001 // can be used to find the original IV expression.
2002 if (SE->getEffectiveSCEVType(Op->getType()) !=
2003 SE->getEffectiveSCEVType(SI->first->getType())) {
2004 CastInst *CI = dyn_cast<CastInst>(Op);
2005 // If it's not a simple cast, it's complicated.
2008 // If it's a cast from a type other than the stride type,
2009 // it's complicated.
2010 if (CI->getOperand(0)->getType() != SI->first->getType())
2012 // Ok, we found the IV expression in the stride's type.
2013 Op = CI->getOperand(0);
2017 if (NewCmpLHS->getType() == CmpTy)
2023 NewCmpTy = NewCmpLHS->getType();
2024 NewTyBits = SE->getTypeSizeInBits(NewCmpTy);
2025 const Type *NewCmpIntTy = IntegerType::get(Cond->getContext(), NewTyBits);
2026 if (RequiresTypeConversion(NewCmpTy, CmpTy)) {
2027 // Check if it is possible to rewrite it using
2028 // an iv / stride of a smaller integer type.
2029 unsigned Bits = NewTyBits;
2030 if (ICmpInst::isSigned(Predicate))
2032 uint64_t Mask = (1ULL << Bits) - 1;
2033 if (((uint64_t)NewCmpVal & Mask) != (uint64_t)NewCmpVal)
2037 // Don't rewrite if use offset is non-constant and the new type is
2038 // of a different type.
2039 // FIXME: too conservative?
2040 if (NewTyBits != TyBits && !isa<SCEVConstant>(CondUse->getOffset()))
2044 bool AllUsesAreAddresses = true;
2045 bool AllUsesAreOutsideLoop = true;
2046 std::vector<BasedUser> UsersToProcess;
2047 const SCEV *CommonExprs = CollectIVUsers(SI->first, *SI->second, L,
2048 AllUsesAreAddresses,
2049 AllUsesAreOutsideLoop,
2051 // Avoid rewriting the compare instruction with an iv of new stride
2052 // if it's likely the new stride uses will be rewritten using the
2053 // stride of the compare instruction.
2054 if (AllUsesAreAddresses &&
2055 ValidScale(!CommonExprs->isZero(), Scale, UsersToProcess))
2059 // Avoid rewriting the compare instruction with an iv which has
2060 // implicit extension or truncation built into it.
2061 // TODO: This is over-conservative.
2062 if (SE->getTypeSizeInBits(CondUse->getOffset()->getType()) != TyBits)
2065 // If scale is negative, use swapped predicate unless it's testing
2067 if (Scale < 0 && !Cond->isEquality())
2068 Predicate = ICmpInst::getSwappedPredicate(Predicate);
2070 NewStride = IU->StrideOrder[i];
2071 if (!isa<PointerType>(NewCmpTy))
2072 NewCmpRHS = ConstantInt::get(NewCmpTy, NewCmpVal);
2074 Constant *CI = ConstantInt::get(NewCmpIntTy, NewCmpVal);
2075 NewCmpRHS = ConstantExpr::getIntToPtr(CI, NewCmpTy);
2077 NewOffset = TyBits == NewTyBits
2078 ? SE->getMulExpr(CondUse->getOffset(),
2079 SE->getConstant(CmpTy, Scale))
2080 : SE->getConstant(NewCmpIntTy,
2081 cast<SCEVConstant>(CondUse->getOffset())->getValue()
2082 ->getSExtValue()*Scale);
2087 // Forgo this transformation if it the increment happens to be
2088 // unfortunately positioned after the condition, and the condition
2089 // has multiple uses which prevent it from being moved immediately
2090 // before the branch. See
2091 // test/Transforms/LoopStrengthReduce/change-compare-stride-trickiness-*.ll
2092 // for an example of this situation.
2093 if (!Cond->hasOneUse()) {
2094 for (BasicBlock::iterator I = Cond, E = Cond->getParent()->end();
2101 // Create a new compare instruction using new stride / iv.
2102 ICmpInst *OldCond = Cond;
2103 // Insert new compare instruction.
2104 Cond = new ICmpInst(OldCond, Predicate, NewCmpLHS, NewCmpRHS,
2105 L->getHeader()->getName() + ".termcond");
2107 DEBUG(errs() << " Change compare stride in Inst " << *OldCond);
2108 DEBUG(errs() << " to " << *Cond << '\n');
2110 // Remove the old compare instruction. The old indvar is probably dead too.
2111 DeadInsts.push_back(CondUse->getOperandValToReplace());
2112 OldCond->replaceAllUsesWith(Cond);
2113 OldCond->eraseFromParent();
2115 IU->IVUsesByStride[NewStride]->addUser(NewOffset, Cond, NewCmpLHS);
2116 CondUse = &IU->IVUsesByStride[NewStride]->Users.back();
2117 CondStride = NewStride;
2125 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
2126 /// a max computation.
2128 /// This is a narrow solution to a specific, but acute, problem. For loops
2134 /// } while (++i < n);
2136 /// the trip count isn't just 'n', because 'n' might not be positive. And
2137 /// unfortunately this can come up even for loops where the user didn't use
2138 /// a C do-while loop. For example, seemingly well-behaved top-test loops
2139 /// will commonly be lowered like this:
2145 /// } while (++i < n);
2148 /// and then it's possible for subsequent optimization to obscure the if
2149 /// test in such a way that indvars can't find it.
2151 /// When indvars can't find the if test in loops like this, it creates a
2152 /// max expression, which allows it to give the loop a canonical
2153 /// induction variable:
2156 /// max = n < 1 ? 1 : n;
2159 /// } while (++i != max);
2161 /// Canonical induction variables are necessary because the loop passes
2162 /// are designed around them. The most obvious example of this is the
2163 /// LoopInfo analysis, which doesn't remember trip count values. It
2164 /// expects to be able to rediscover the trip count each time it is
2165 /// needed, and it does this using a simple analyis that only succeeds if
2166 /// the loop has a canonical induction variable.
2168 /// However, when it comes time to generate code, the maximum operation
2169 /// can be quite costly, especially if it's inside of an outer loop.
2171 /// This function solves this problem by detecting this type of loop and
2172 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
2173 /// the instructions for the maximum computation.
2175 ICmpInst *LoopStrengthReduce::OptimizeMax(Loop *L, ICmpInst *Cond,
2176 IVStrideUse* &CondUse) {
2177 // Check that the loop matches the pattern we're looking for.
2178 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
2179 Cond->getPredicate() != CmpInst::ICMP_NE)
2182 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
2183 if (!Sel || !Sel->hasOneUse()) return Cond;
2185 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
2186 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2188 const SCEV *One = SE->getIntegerSCEV(1, BackedgeTakenCount->getType());
2190 // Add one to the backedge-taken count to get the trip count.
2191 const SCEV *IterationCount = SE->getAddExpr(BackedgeTakenCount, One);
2193 // Check for a max calculation that matches the pattern.
2194 if (!isa<SCEVSMaxExpr>(IterationCount) && !isa<SCEVUMaxExpr>(IterationCount))
2196 const SCEVNAryExpr *Max = cast<SCEVNAryExpr>(IterationCount);
2197 if (Max != SE->getSCEV(Sel)) return Cond;
2199 // To handle a max with more than two operands, this optimization would
2200 // require additional checking and setup.
2201 if (Max->getNumOperands() != 2)
2204 const SCEV *MaxLHS = Max->getOperand(0);
2205 const SCEV *MaxRHS = Max->getOperand(1);
2206 if (!MaxLHS || MaxLHS != One) return Cond;
2208 // Check the relevant induction variable for conformance to
2210 const SCEV *IV = SE->getSCEV(Cond->getOperand(0));
2211 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2212 if (!AR || !AR->isAffine() ||
2213 AR->getStart() != One ||
2214 AR->getStepRecurrence(*SE) != One)
2217 assert(AR->getLoop() == L &&
2218 "Loop condition operand is an addrec in a different loop!");
2220 // Check the right operand of the select, and remember it, as it will
2221 // be used in the new comparison instruction.
2223 if (SE->getSCEV(Sel->getOperand(1)) == MaxRHS)
2224 NewRHS = Sel->getOperand(1);
2225 else if (SE->getSCEV(Sel->getOperand(2)) == MaxRHS)
2226 NewRHS = Sel->getOperand(2);
2227 if (!NewRHS) return Cond;
2229 // Determine the new comparison opcode. It may be signed or unsigned,
2230 // and the original comparison may be either equality or inequality.
2231 CmpInst::Predicate Pred =
2232 isa<SCEVSMaxExpr>(Max) ? CmpInst::ICMP_SLT : CmpInst::ICMP_ULT;
2233 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
2234 Pred = CmpInst::getInversePredicate(Pred);
2236 // Ok, everything looks ok to change the condition into an SLT or SGE and
2237 // delete the max calculation.
2239 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
2241 // Delete the max calculation instructions.
2242 Cond->replaceAllUsesWith(NewCond);
2243 CondUse->setUser(NewCond);
2244 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2245 Cond->eraseFromParent();
2246 Sel->eraseFromParent();
2247 if (Cmp->use_empty())
2248 Cmp->eraseFromParent();
2252 /// OptimizeShadowIV - If IV is used in a int-to-float cast
2253 /// inside the loop then try to eliminate the cast opeation.
2254 void LoopStrengthReduce::OptimizeShadowIV(Loop *L) {
2256 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
2257 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2260 for (unsigned Stride = 0, e = IU->StrideOrder.size(); Stride != e;
2262 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
2263 IU->IVUsesByStride.find(IU->StrideOrder[Stride]);
2264 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
2265 if (!isa<SCEVConstant>(SI->first))
2268 for (ilist<IVStrideUse>::iterator UI = SI->second->Users.begin(),
2269 E = SI->second->Users.end(); UI != E; /* empty */) {
2270 ilist<IVStrideUse>::iterator CandidateUI = UI;
2272 Instruction *ShadowUse = CandidateUI->getUser();
2273 const Type *DestTy = NULL;
2275 /* If shadow use is a int->float cast then insert a second IV
2276 to eliminate this cast.
2278 for (unsigned i = 0; i < n; ++i)
2284 for (unsigned i = 0; i < n; ++i, ++d)
2287 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser()))
2288 DestTy = UCast->getDestTy();
2289 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser()))
2290 DestTy = SCast->getDestTy();
2291 if (!DestTy) continue;
2294 // If target does not support DestTy natively then do not apply
2295 // this transformation.
2296 EVT DVT = TLI->getValueType(DestTy);
2297 if (!TLI->isTypeLegal(DVT)) continue;
2300 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
2302 if (PH->getNumIncomingValues() != 2) continue;
2304 const Type *SrcTy = PH->getType();
2305 int Mantissa = DestTy->getFPMantissaWidth();
2306 if (Mantissa == -1) continue;
2307 if ((int)SE->getTypeSizeInBits(SrcTy) > Mantissa)
2310 unsigned Entry, Latch;
2311 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
2319 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
2320 if (!Init) continue;
2321 Constant *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());
2323 BinaryOperator *Incr =
2324 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
2325 if (!Incr) continue;
2326 if (Incr->getOpcode() != Instruction::Add
2327 && Incr->getOpcode() != Instruction::Sub)
2330 /* Initialize new IV, double d = 0.0 in above example. */
2331 ConstantInt *C = NULL;
2332 if (Incr->getOperand(0) == PH)
2333 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
2334 else if (Incr->getOperand(1) == PH)
2335 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
2341 // Ignore negative constants, as the code below doesn't handle them
2342 // correctly. TODO: Remove this restriction.
2343 if (!C->getValue().isStrictlyPositive()) continue;
2345 /* Add new PHINode. */
2346 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);
2348 /* create new increment. '++d' in above example. */
2349 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
2350 BinaryOperator *NewIncr =
2351 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
2352 Instruction::FAdd : Instruction::FSub,
2353 NewPH, CFP, "IV.S.next.", Incr);
2355 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
2356 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
2358 /* Remove cast operation */
2359 ShadowUse->replaceAllUsesWith(NewPH);
2360 ShadowUse->eraseFromParent();
2367 /// OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
2368 /// uses in the loop, look to see if we can eliminate some, in favor of using
2369 /// common indvars for the different uses.
2370 void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
2371 // TODO: implement optzns here.
2373 OptimizeShadowIV(L);
2376 bool LoopStrengthReduce::StrideMightBeShared(const SCEV* Stride, Loop *L,
2378 int64_t SInt = cast<SCEVConstant>(Stride)->getValue()->getSExtValue();
2379 for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
2380 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
2381 IU->IVUsesByStride.find(IU->StrideOrder[i]);
2382 const SCEV *Share = SI->first;
2383 if (!isa<SCEVConstant>(SI->first) || Share == Stride)
2385 int64_t SSInt = cast<SCEVConstant>(Share)->getValue()->getSExtValue();
2387 return true; // This can definitely be reused.
2388 if (unsigned(abs64(SSInt)) < SInt || (SSInt % SInt) != 0)
2390 int64_t Scale = SSInt / SInt;
2391 bool AllUsesAreAddresses = true;
2392 bool AllUsesAreOutsideLoop = true;
2393 std::vector<BasedUser> UsersToProcess;
2394 const SCEV *CommonExprs = CollectIVUsers(SI->first, *SI->second, L,
2395 AllUsesAreAddresses,
2396 AllUsesAreOutsideLoop,
2398 if (AllUsesAreAddresses &&
2399 ValidScale(!CommonExprs->isZero(), Scale, UsersToProcess)) {
2402 // Any pre-inc iv use?
2403 IVUsersOfOneStride &StrideUses = *IU->IVUsesByStride[Share];
2404 for (ilist<IVStrideUse>::iterator I = StrideUses.Users.begin(),
2405 E = StrideUses.Users.end(); I != E; ++I) {
2406 if (!I->isUseOfPostIncrementedValue())
2414 /// isUsedByExitBranch - Return true if icmp is used by a loop terminating
2415 /// conditional branch or it's and / or with other conditions before being used
2416 /// as the condition.
2417 static bool isUsedByExitBranch(ICmpInst *Cond, Loop *L) {
2418 BasicBlock *CondBB = Cond->getParent();
2419 if (!L->isLoopExiting(CondBB))
2421 BranchInst *TermBr = dyn_cast<BranchInst>(CondBB->getTerminator());
2422 if (!TermBr || !TermBr->isConditional())
2425 Value *User = *Cond->use_begin();
2426 Instruction *UserInst = dyn_cast<Instruction>(User);
2428 (UserInst->getOpcode() == Instruction::And ||
2429 UserInst->getOpcode() == Instruction::Or)) {
2430 if (!UserInst->hasOneUse() || UserInst->getParent() != CondBB)
2432 User = *User->use_begin();
2433 UserInst = dyn_cast<Instruction>(User);
2435 return User == TermBr;
2438 static bool ShouldCountToZero(ICmpInst *Cond, IVStrideUse* &CondUse,
2439 ScalarEvolution *SE, Loop *L,
2440 const TargetLowering *TLI = 0) {
2441 if (!L->contains(Cond->getParent()))
2444 if (!isa<SCEVConstant>(CondUse->getOffset()))
2447 // Handle only tests for equality for the moment.
2448 if (!Cond->isEquality() || !Cond->hasOneUse())
2450 if (!isUsedByExitBranch(Cond, L))
2453 Value *CondOp0 = Cond->getOperand(0);
2454 const SCEV *IV = SE->getSCEV(CondOp0);
2455 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2456 if (!AR || !AR->isAffine())
2459 const SCEVConstant *SC = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE));
2460 if (!SC || SC->getValue()->getSExtValue() < 0)
2461 // If it's already counting down, don't do anything.
2464 // If the RHS of the comparison is not an loop invariant, the rewrite
2465 // cannot be done. Also bail out if it's already comparing against a zero.
2466 // If we are checking this before cmp stride optimization, check if it's
2467 // comparing against a already legal immediate.
2468 Value *RHS = Cond->getOperand(1);
2469 ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS);
2470 if (!L->isLoopInvariant(RHS) ||
2471 (RHSC && RHSC->isZero()) ||
2472 (RHSC && TLI && TLI->isLegalICmpImmediate(RHSC->getSExtValue())))
2475 // Make sure the IV is only used for counting. Value may be preinc or
2476 // postinc; 2 uses in either case.
2477 if (!CondOp0->hasNUses(2))
2483 /// OptimizeLoopTermCond - Change loop terminating condition to use the
2484 /// postinc iv when possible.
2485 void LoopStrengthReduce::OptimizeLoopTermCond(Loop *L) {
2486 BasicBlock *LatchBlock = L->getLoopLatch();
2487 bool LatchExit = L->isLoopExiting(LatchBlock);
2488 SmallVector<BasicBlock*, 8> ExitingBlocks;
2489 L->getExitingBlocks(ExitingBlocks);
2491 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
2492 BasicBlock *ExitingBlock = ExitingBlocks[i];
2494 // Finally, get the terminating condition for the loop if possible. If we
2495 // can, we want to change it to use a post-incremented version of its
2496 // induction variable, to allow coalescing the live ranges for the IV into
2497 // one register value.
2499 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2502 // FIXME: Overly conservative, termination condition could be an 'or' etc..
2503 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
2506 // Search IVUsesByStride to find Cond's IVUse if there is one.
2507 IVStrideUse *CondUse = 0;
2508 const SCEV *CondStride = 0;
2509 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2510 if (!FindIVUserForCond(Cond, CondUse, CondStride))
2513 // If the latch block is exiting and it's not a single block loop, it's
2514 // not safe to use postinc iv in other exiting blocks. FIXME: overly
2515 // conservative? How about icmp stride optimization?
2516 bool UsePostInc = !(e > 1 && LatchExit && ExitingBlock != LatchBlock);
2517 if (UsePostInc && ExitingBlock != LatchBlock) {
2518 if (!Cond->hasOneUse())
2519 // See below, we don't want the condition to be cloned.
2522 // If exiting block is the latch block, we know it's safe and profitable
2523 // to transform the icmp to use post-inc iv. Otherwise do so only if it
2524 // would not reuse another iv and its iv would be reused by other uses.
2525 // We are optimizing for the case where the icmp is the only use of the
2527 IVUsersOfOneStride &StrideUses = *IU->IVUsesByStride[CondStride];
2528 for (ilist<IVStrideUse>::iterator I = StrideUses.Users.begin(),
2529 E = StrideUses.Users.end(); I != E; ++I) {
2530 if (I->getUser() == Cond)
2532 if (!I->isUseOfPostIncrementedValue()) {
2539 // If iv for the stride might be shared and any of the users use pre-inc
2540 // iv might be used, then it's not safe to use post-inc iv.
2542 isa<SCEVConstant>(CondStride) &&
2543 StrideMightBeShared(CondStride, L, true))
2547 // If the trip count is computed in terms of a max (due to ScalarEvolution
2548 // being unable to find a sufficient guard, for example), change the loop
2549 // comparison to use SLT or ULT instead of NE.
2550 Cond = OptimizeMax(L, Cond, CondUse);
2552 // If possible, change stride and operands of the compare instruction to
2553 // eliminate one stride. However, avoid rewriting the compare instruction
2554 // with an iv of new stride if it's likely the new stride uses will be
2555 // rewritten using the stride of the compare instruction.
2556 if (ExitingBlock == LatchBlock && isa<SCEVConstant>(CondStride)) {
2557 // If the condition stride is a constant and it's the only use, we might
2558 // want to optimize it first by turning it to count toward zero.
2559 if (!StrideMightBeShared(CondStride, L, false) &&
2560 !ShouldCountToZero(Cond, CondUse, SE, L, TLI))
2561 Cond = ChangeCompareStride(L, Cond, CondUse, CondStride);
2567 DEBUG(errs() << " Change loop exiting icmp to use postinc iv: "
2570 // It's possible for the setcc instruction to be anywhere in the loop, and
2571 // possible for it to have multiple users. If it is not immediately before
2572 // the exiting block branch, move it.
2573 if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
2574 if (Cond->hasOneUse()) { // Condition has a single use, just move it.
2575 Cond->moveBefore(TermBr);
2577 // Otherwise, clone the terminating condition and insert into the
2579 Cond = cast<ICmpInst>(Cond->clone());
2580 Cond->setName(L->getHeader()->getName() + ".termcond");
2581 ExitingBlock->getInstList().insert(TermBr, Cond);
2583 // Clone the IVUse, as the old use still exists!
2584 IU->IVUsesByStride[CondStride]->addUser(CondUse->getOffset(), Cond,
2585 CondUse->getOperandValToReplace());
2586 CondUse = &IU->IVUsesByStride[CondStride]->Users.back();
2590 // If we get to here, we know that we can transform the setcc instruction to
2591 // use the post-incremented version of the IV, allowing us to coalesce the
2592 // live ranges for the IV correctly.
2593 CondUse->setOffset(SE->getMinusSCEV(CondUse->getOffset(), CondStride));
2594 CondUse->setIsUseOfPostIncrementedValue(true);
2601 bool LoopStrengthReduce::OptimizeLoopCountIVOfStride(const SCEV* &Stride,
2602 IVStrideUse* &CondUse,
2604 // If the only use is an icmp of a loop exiting conditional branch, then
2605 // attempt the optimization.
2606 BasedUser User = BasedUser(*CondUse, SE);
2607 assert(isa<ICmpInst>(User.Inst) && "Expecting an ICMPInst!");
2608 ICmpInst *Cond = cast<ICmpInst>(User.Inst);
2610 // Less strict check now that compare stride optimization is done.
2611 if (!ShouldCountToZero(Cond, CondUse, SE, L))
2614 Value *CondOp0 = Cond->getOperand(0);
2615 PHINode *PHIExpr = dyn_cast<PHINode>(CondOp0);
2618 // Value tested is postinc. Find the phi node.
2619 Incr = dyn_cast<BinaryOperator>(CondOp0);
2620 // FIXME: Just use User.OperandValToReplace here?
2621 if (!Incr || Incr->getOpcode() != Instruction::Add)
2624 PHIExpr = dyn_cast<PHINode>(Incr->getOperand(0));
2627 // 1 use for preinc value, the increment.
2628 if (!PHIExpr->hasOneUse())
2631 assert(isa<PHINode>(CondOp0) &&
2632 "Unexpected loop exiting counting instruction sequence!");
2633 PHIExpr = cast<PHINode>(CondOp0);
2634 // Value tested is preinc. Find the increment.
2635 // A CmpInst is not a BinaryOperator; we depend on this.
2636 Instruction::use_iterator UI = PHIExpr->use_begin();
2637 Incr = dyn_cast<BinaryOperator>(UI);
2639 Incr = dyn_cast<BinaryOperator>(++UI);
2640 // One use for postinc value, the phi. Unnecessarily conservative?
2641 if (!Incr || !Incr->hasOneUse() || Incr->getOpcode() != Instruction::Add)
2645 // Replace the increment with a decrement.
2646 DEBUG(errs() << "LSR: Examining use ");
2647 DEBUG(WriteAsOperand(errs(), CondOp0, /*PrintType=*/false));
2648 DEBUG(errs() << " in Inst: " << *Cond << '\n');
2649 BinaryOperator *Decr = BinaryOperator::Create(Instruction::Sub,
2650 Incr->getOperand(0), Incr->getOperand(1), "tmp", Incr);
2651 Incr->replaceAllUsesWith(Decr);
2652 Incr->eraseFromParent();
2654 // Substitute endval-startval for the original startval, and 0 for the
2655 // original endval. Since we're only testing for equality this is OK even
2656 // if the computation wraps around.
2657 BasicBlock *Preheader = L->getLoopPreheader();
2658 Instruction *PreInsertPt = Preheader->getTerminator();
2659 unsigned InBlock = L->contains(PHIExpr->getIncomingBlock(0)) ? 1 : 0;
2660 Value *StartVal = PHIExpr->getIncomingValue(InBlock);
2661 Value *EndVal = Cond->getOperand(1);
2662 DEBUG(errs() << " Optimize loop counting iv to count down ["
2663 << *EndVal << " .. " << *StartVal << "]\n");
2665 // FIXME: check for case where both are constant.
2666 Constant* Zero = ConstantInt::get(Cond->getOperand(1)->getType(), 0);
2667 BinaryOperator *NewStartVal = BinaryOperator::Create(Instruction::Sub,
2668 EndVal, StartVal, "tmp", PreInsertPt);
2669 PHIExpr->setIncomingValue(InBlock, NewStartVal);
2670 Cond->setOperand(1, Zero);
2671 DEBUG(errs() << " New icmp: " << *Cond << "\n");
2673 int64_t SInt = cast<SCEVConstant>(Stride)->getValue()->getSExtValue();
2674 const SCEV *NewStride = 0;
2676 for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
2677 const SCEV *OldStride = IU->StrideOrder[i];
2678 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(OldStride))
2679 if (SC->getValue()->getSExtValue() == -SInt) {
2681 NewStride = OldStride;
2687 NewStride = SE->getIntegerSCEV(-SInt, Stride->getType());
2688 IU->AddUser(NewStride, CondUse->getOffset(), Cond, Cond->getOperand(0));
2689 IU->IVUsesByStride[Stride]->removeUser(CondUse);
2691 CondUse = &IU->IVUsesByStride[NewStride]->Users.back();
2699 /// OptimizeLoopCountIV - If, after all sharing of IVs, the IV used for deciding
2700 /// when to exit the loop is used only for that purpose, try to rearrange things
2701 /// so it counts down to a test against zero.
2702 bool LoopStrengthReduce::OptimizeLoopCountIV(Loop *L) {
2703 bool ThisChanged = false;
2704 for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
2705 const SCEV *Stride = IU->StrideOrder[i];
2706 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
2707 IU->IVUsesByStride.find(Stride);
2708 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
2709 // FIXME: Generalize to non-affine IV's.
2710 if (!SI->first->isLoopInvariant(L))
2712 // If stride is a constant and it has an icmpinst use, check if we can
2713 // optimize the loop to count down.
2714 if (isa<SCEVConstant>(Stride) && SI->second->Users.size() == 1) {
2715 Instruction *User = SI->second->Users.begin()->getUser();
2716 if (!isa<ICmpInst>(User))
2718 const SCEV *CondStride = Stride;
2719 IVStrideUse *Use = &*SI->second->Users.begin();
2720 if (!OptimizeLoopCountIVOfStride(CondStride, Use, L))
2724 // Now check if it's possible to reuse this iv for other stride uses.
2725 for (unsigned j = 0, ee = IU->StrideOrder.size(); j != ee; ++j) {
2726 const SCEV *SStride = IU->StrideOrder[j];
2727 if (SStride == CondStride)
2729 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SII =
2730 IU->IVUsesByStride.find(SStride);
2731 assert(SII != IU->IVUsesByStride.end() && "Stride doesn't exist!");
2732 // FIXME: Generalize to non-affine IV's.
2733 if (!SII->first->isLoopInvariant(L))
2735 // FIXME: Rewrite other stride using CondStride.
2740 Changed |= ThisChanged;
2744 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager &LPM) {
2745 IU = &getAnalysis<IVUsers>();
2746 LI = &getAnalysis<LoopInfo>();
2747 DT = &getAnalysis<DominatorTree>();
2748 SE = &getAnalysis<ScalarEvolution>();
2751 // If LoopSimplify form is not available, stay out of trouble.
2752 if (!L->getLoopPreheader() || !L->getLoopLatch())
2755 if (!IU->IVUsesByStride.empty()) {
2756 DEBUG(errs() << "\nLSR on \"" << L->getHeader()->getParent()->getName()
2760 // Sort the StrideOrder so we process larger strides first.
2761 std::stable_sort(IU->StrideOrder.begin(), IU->StrideOrder.end(),
2764 // Optimize induction variables. Some indvar uses can be transformed to use
2765 // strides that will be needed for other purposes. A common example of this
2766 // is the exit test for the loop, which can often be rewritten to use the
2767 // computation of some other indvar to decide when to terminate the loop.
2770 // Change loop terminating condition to use the postinc iv when possible
2771 // and optimize loop terminating compare. FIXME: Move this after
2772 // StrengthReduceIVUsersOfStride?
2773 OptimizeLoopTermCond(L);
2775 // FIXME: We can shrink overlarge IV's here. e.g. if the code has
2776 // computation in i64 values and the target doesn't support i64, demote
2777 // the computation to 32-bit if safe.
2779 // FIXME: Attempt to reuse values across multiple IV's. In particular, we
2780 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should
2781 // be codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC.
2782 // Need to be careful that IV's are all the same type. Only works for
2783 // intptr_t indvars.
2785 // IVsByStride keeps IVs for one particular loop.
2786 assert(IVsByStride.empty() && "Stale entries in IVsByStride?");
2788 StrengthReduceIVUsers(L);
2790 // After all sharing is done, see if we can adjust the loop to test against
2791 // zero instead of counting up to a maximum. This is usually faster.
2792 OptimizeLoopCountIV(L);
2795 // We're done analyzing this loop; release all the state we built up for it.
2796 IVsByStride.clear();
2797 StrideNoReuse.clear();
2799 // Clean up after ourselves
2800 if (!DeadInsts.empty())
2801 DeleteTriviallyDeadInstructions();
2803 // At this point, it is worth checking to see if any recurrence PHIs are also
2804 // dead, so that we can remove them as well.
2805 DeleteDeadPHIs(L->getHeader());