1 //===- LoopStrengthReduce.cpp - Strength Reduce IVs in Loops --------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This transformation analyzes and transforms the induction variables (and
11 // computations derived from them) into forms suitable for efficient execution
14 // This pass performs a strength reduction on array references inside loops that
15 // have as one or more of their components the loop induction variable, it
16 // rewrites expressions to take advantage of scaled-index addressing modes
17 // available on the target, and it performs a variety of other optimizations
18 // related to loop induction variables.
20 //===----------------------------------------------------------------------===//
22 #define DEBUG_TYPE "loop-reduce"
23 #include "llvm/Transforms/Scalar.h"
24 #include "llvm/Constants.h"
25 #include "llvm/Instructions.h"
26 #include "llvm/IntrinsicInst.h"
27 #include "llvm/DerivedTypes.h"
28 #include "llvm/Analysis/IVUsers.h"
29 #include "llvm/Analysis/LoopPass.h"
30 #include "llvm/Analysis/ScalarEvolutionExpander.h"
31 #include "llvm/Transforms/Utils/AddrModeMatcher.h"
32 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
33 #include "llvm/Transforms/Utils/Local.h"
34 #include "llvm/ADT/Statistic.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/CommandLine.h"
37 #include "llvm/Support/ValueHandle.h"
38 #include "llvm/Support/raw_ostream.h"
39 #include "llvm/Target/TargetLowering.h"
43 STATISTIC(NumReduced , "Number of IV uses strength reduced");
44 STATISTIC(NumInserted, "Number of PHIs inserted");
45 STATISTIC(NumVariable, "Number of PHIs with variable strides");
46 STATISTIC(NumEliminated, "Number of strides eliminated");
47 STATISTIC(NumShadow, "Number of Shadow IVs optimized");
48 STATISTIC(NumImmSunk, "Number of common expr immediates sunk into uses");
49 STATISTIC(NumLoopCond, "Number of loop terminating conds optimized");
50 STATISTIC(NumCountZero, "Number of count iv optimized to count toward zero");
52 static cl::opt<bool> EnableFullLSRMode("enable-full-lsr",
60 /// IVInfo - This structure keeps track of one IV expression inserted during
61 /// StrengthReduceStridedIVUsers. It contains the stride, the common base, as
62 /// well as the PHI node and increment value created for rewrite.
68 IVExpr(const SCEV *const stride, const SCEV *const base, PHINode *phi)
69 : Stride(stride), Base(base), PHI(phi) {}
72 /// IVsOfOneStride - This structure keeps track of all IV expression inserted
73 /// during StrengthReduceStridedIVUsers for a particular stride of the IV.
74 struct IVsOfOneStride {
75 std::vector<IVExpr> IVs;
77 void addIV(const SCEV *const Stride, const SCEV *const Base, PHINode *PHI) {
78 IVs.push_back(IVExpr(Stride, Base, PHI));
82 class LoopStrengthReduce : public LoopPass {
87 /// IVsByStride - Keep track of all IVs that have been inserted for a
88 /// particular stride.
89 std::map<const SCEV *, IVsOfOneStride> IVsByStride;
91 /// DeadInsts - Keep track of instructions we may have made dead, so that
92 /// we can remove them after we are done working.
93 SmallVector<WeakVH, 16> DeadInsts;
95 /// TLI - Keep a pointer of a TargetLowering to consult for determining
96 /// transformation profitability.
97 const TargetLowering *TLI;
100 static char ID; // Pass ID, replacement for typeid
101 explicit LoopStrengthReduce(const TargetLowering *tli = NULL) :
102 LoopPass(&ID), TLI(tli) {}
104 bool runOnLoop(Loop *L, LPPassManager &LPM);
106 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
107 // We split critical edges, so we change the CFG. However, we do update
108 // many analyses if they are around.
109 AU.addPreservedID(LoopSimplifyID);
110 AU.addPreserved("loops");
111 AU.addPreserved("domfrontier");
112 AU.addPreserved("domtree");
114 AU.addRequiredID(LoopSimplifyID);
115 AU.addRequired<ScalarEvolution>();
116 AU.addPreserved<ScalarEvolution>();
117 AU.addRequired<IVUsers>();
118 AU.addPreserved<IVUsers>();
122 void OptimizeIndvars(Loop *L);
124 /// OptimizeLoopTermCond - Change loop terminating condition to use the
125 /// postinc iv when possible.
126 void OptimizeLoopTermCond(Loop *L);
128 /// OptimizeShadowIV - If IV is used in a int-to-float cast
129 /// inside the loop then try to eliminate the cast opeation.
130 void OptimizeShadowIV(Loop *L);
132 /// OptimizeMax - Rewrite the loop's terminating condition
133 /// if it uses a max computation.
134 ICmpInst *OptimizeMax(Loop *L, ICmpInst *Cond,
135 IVStrideUse* &CondUse);
137 /// OptimizeLoopCountIV - If, after all sharing of IVs, the IV used for
138 /// deciding when to exit the loop is used only for that purpose, try to
139 /// rearrange things so it counts down to a test against zero.
140 bool OptimizeLoopCountIV(Loop *L);
141 bool OptimizeLoopCountIVOfStride(const SCEV* &Stride,
142 IVStrideUse* &CondUse, Loop *L);
144 /// StrengthReduceIVUsersOfStride - Strength reduce all of the users of a
145 /// single stride of IV. All of the users may have different starting
146 /// values, and this may not be the only stride.
147 void StrengthReduceIVUsersOfStride(const SCEV *const &Stride,
148 IVUsersOfOneStride &Uses,
150 void StrengthReduceIVUsers(Loop *L);
152 ICmpInst *ChangeCompareStride(Loop *L, ICmpInst *Cond,
153 IVStrideUse* &CondUse,
154 const SCEV* &CondStride,
155 bool PostPass = false);
157 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse,
158 const SCEV* &CondStride);
159 bool RequiresTypeConversion(const Type *Ty, const Type *NewTy);
160 const SCEV *CheckForIVReuse(bool, bool, bool, const SCEV *const&,
161 IVExpr&, const Type*,
162 const std::vector<BasedUser>& UsersToProcess);
163 bool ValidScale(bool, int64_t,
164 const std::vector<BasedUser>& UsersToProcess);
165 bool ValidOffset(bool, int64_t, int64_t,
166 const std::vector<BasedUser>& UsersToProcess);
167 const SCEV *CollectIVUsers(const SCEV *const &Stride,
168 IVUsersOfOneStride &Uses,
170 bool &AllUsesAreAddresses,
171 bool &AllUsesAreOutsideLoop,
172 std::vector<BasedUser> &UsersToProcess);
173 bool StrideMightBeShared(const SCEV *Stride, Loop *L, bool CheckPreInc);
174 bool ShouldUseFullStrengthReductionMode(
175 const std::vector<BasedUser> &UsersToProcess,
177 bool AllUsesAreAddresses,
179 void PrepareToStrengthReduceFully(
180 std::vector<BasedUser> &UsersToProcess,
182 const SCEV *CommonExprs,
184 SCEVExpander &PreheaderRewriter);
185 void PrepareToStrengthReduceFromSmallerStride(
186 std::vector<BasedUser> &UsersToProcess,
188 const IVExpr &ReuseIV,
189 Instruction *PreInsertPt);
190 void PrepareToStrengthReduceWithNewPhi(
191 std::vector<BasedUser> &UsersToProcess,
193 const SCEV *CommonExprs,
195 Instruction *IVIncInsertPt,
197 SCEVExpander &PreheaderRewriter);
199 void DeleteTriviallyDeadInstructions();
203 char LoopStrengthReduce::ID = 0;
204 static RegisterPass<LoopStrengthReduce>
205 X("loop-reduce", "Loop Strength Reduction");
207 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
208 return new LoopStrengthReduce(TLI);
211 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
212 /// specified set are trivially dead, delete them and see if this makes any of
213 /// their operands subsequently dead.
214 void LoopStrengthReduce::DeleteTriviallyDeadInstructions() {
215 if (DeadInsts.empty()) return;
217 while (!DeadInsts.empty()) {
218 Instruction *I = dyn_cast_or_null<Instruction>(DeadInsts.pop_back_val());
220 if (I == 0 || !isInstructionTriviallyDead(I))
223 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
224 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
227 DeadInsts.push_back(U);
230 I->eraseFromParent();
235 /// containsAddRecFromDifferentLoop - Determine whether expression S involves a
236 /// subexpression that is an AddRec from a loop other than L. An outer loop
237 /// of L is OK, but not an inner loop nor a disjoint loop.
238 static bool containsAddRecFromDifferentLoop(const SCEV *S, Loop *L) {
239 // This is very common, put it first.
240 if (isa<SCEVConstant>(S))
242 if (const SCEVCommutativeExpr *AE = dyn_cast<SCEVCommutativeExpr>(S)) {
243 for (unsigned int i=0; i< AE->getNumOperands(); i++)
244 if (containsAddRecFromDifferentLoop(AE->getOperand(i), L))
248 if (const SCEVAddRecExpr *AE = dyn_cast<SCEVAddRecExpr>(S)) {
249 if (const Loop *newLoop = AE->getLoop()) {
252 // if newLoop is an outer loop of L, this is OK.
253 if (newLoop->contains(L->getHeader()))
258 if (const SCEVUDivExpr *DE = dyn_cast<SCEVUDivExpr>(S))
259 return containsAddRecFromDifferentLoop(DE->getLHS(), L) ||
260 containsAddRecFromDifferentLoop(DE->getRHS(), L);
262 // SCEVSDivExpr has been backed out temporarily, but will be back; we'll
263 // need this when it is.
264 if (const SCEVSDivExpr *DE = dyn_cast<SCEVSDivExpr>(S))
265 return containsAddRecFromDifferentLoop(DE->getLHS(), L) ||
266 containsAddRecFromDifferentLoop(DE->getRHS(), L);
268 if (const SCEVCastExpr *CE = dyn_cast<SCEVCastExpr>(S))
269 return containsAddRecFromDifferentLoop(CE->getOperand(), L);
273 /// isAddressUse - Returns true if the specified instruction is using the
274 /// specified value as an address.
275 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
276 bool isAddress = isa<LoadInst>(Inst);
277 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
278 if (SI->getOperand(1) == OperandVal)
280 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
281 // Addressing modes can also be folded into prefetches and a variety
283 switch (II->getIntrinsicID()) {
285 case Intrinsic::prefetch:
286 case Intrinsic::x86_sse2_loadu_dq:
287 case Intrinsic::x86_sse2_loadu_pd:
288 case Intrinsic::x86_sse_loadu_ps:
289 case Intrinsic::x86_sse_storeu_ps:
290 case Intrinsic::x86_sse2_storeu_pd:
291 case Intrinsic::x86_sse2_storeu_dq:
292 case Intrinsic::x86_sse2_storel_dq:
293 if (II->getOperand(1) == OperandVal)
301 /// getAccessType - Return the type of the memory being accessed.
302 static const Type *getAccessType(const Instruction *Inst) {
303 const Type *AccessTy = Inst->getType();
304 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
305 AccessTy = SI->getOperand(0)->getType();
306 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
307 // Addressing modes can also be folded into prefetches and a variety
309 switch (II->getIntrinsicID()) {
311 case Intrinsic::x86_sse_storeu_ps:
312 case Intrinsic::x86_sse2_storeu_pd:
313 case Intrinsic::x86_sse2_storeu_dq:
314 case Intrinsic::x86_sse2_storel_dq:
315 AccessTy = II->getOperand(1)->getType();
323 /// BasedUser - For a particular base value, keep information about how we've
324 /// partitioned the expression so far.
326 /// Base - The Base value for the PHI node that needs to be inserted for
327 /// this use. As the use is processed, information gets moved from this
328 /// field to the Imm field (below). BasedUser values are sorted by this
332 /// Inst - The instruction using the induction variable.
335 /// OperandValToReplace - The operand value of Inst to replace with the
337 Value *OperandValToReplace;
339 /// Imm - The immediate value that should be added to the base immediately
340 /// before Inst, because it will be folded into the imm field of the
341 /// instruction. This is also sometimes used for loop-variant values that
342 /// must be added inside the loop.
345 /// Phi - The induction variable that performs the striding that
346 /// should be used for this user.
349 // isUseOfPostIncrementedValue - True if this should use the
350 // post-incremented version of this IV, not the preincremented version.
351 // This can only be set in special cases, such as the terminating setcc
352 // instruction for a loop and uses outside the loop that are dominated by
354 bool isUseOfPostIncrementedValue;
356 BasedUser(IVStrideUse &IVSU, ScalarEvolution *se)
357 : Base(IVSU.getOffset()), Inst(IVSU.getUser()),
358 OperandValToReplace(IVSU.getOperandValToReplace()),
359 Imm(se->getIntegerSCEV(0, Base->getType())),
360 isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue()) {}
362 // Once we rewrite the code to insert the new IVs we want, update the
363 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
365 void RewriteInstructionToUseNewBase(const SCEV *const &NewBase,
366 Instruction *InsertPt,
367 SCEVExpander &Rewriter, Loop *L, Pass *P,
368 SmallVectorImpl<WeakVH> &DeadInsts,
369 ScalarEvolution *SE);
371 Value *InsertCodeForBaseAtPosition(const SCEV *const &NewBase,
373 SCEVExpander &Rewriter,
375 ScalarEvolution *SE);
380 void BasedUser::dump() const {
381 errs() << " Base=" << *Base;
382 errs() << " Imm=" << *Imm;
383 errs() << " Inst: " << *Inst;
386 Value *BasedUser::InsertCodeForBaseAtPosition(const SCEV *const &NewBase,
388 SCEVExpander &Rewriter,
390 ScalarEvolution *SE) {
391 Value *Base = Rewriter.expandCodeFor(NewBase, 0, IP);
393 // Wrap the base in a SCEVUnknown so that ScalarEvolution doesn't try to
395 const SCEV *NewValSCEV = SE->getUnknown(Base);
397 // Always emit the immediate into the same block as the user.
398 NewValSCEV = SE->getAddExpr(NewValSCEV, Imm);
400 return Rewriter.expandCodeFor(NewValSCEV, Ty, IP);
404 // Once we rewrite the code to insert the new IVs we want, update the
405 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
406 // to it. NewBasePt is the last instruction which contributes to the
407 // value of NewBase in the case that it's a diffferent instruction from
408 // the PHI that NewBase is computed from, or null otherwise.
410 void BasedUser::RewriteInstructionToUseNewBase(const SCEV *const &NewBase,
411 Instruction *NewBasePt,
412 SCEVExpander &Rewriter, Loop *L, Pass *P,
413 SmallVectorImpl<WeakVH> &DeadInsts,
414 ScalarEvolution *SE) {
415 if (!isa<PHINode>(Inst)) {
416 // By default, insert code at the user instruction.
417 BasicBlock::iterator InsertPt = Inst;
419 // However, if the Operand is itself an instruction, the (potentially
420 // complex) inserted code may be shared by many users. Because of this, we
421 // want to emit code for the computation of the operand right before its old
422 // computation. This is usually safe, because we obviously used to use the
423 // computation when it was computed in its current block. However, in some
424 // cases (e.g. use of a post-incremented induction variable) the NewBase
425 // value will be pinned to live somewhere after the original computation.
426 // In this case, we have to back off.
428 // If this is a use outside the loop (which means after, since it is based
429 // on a loop indvar) we use the post-incremented value, so that we don't
430 // artificially make the preinc value live out the bottom of the loop.
431 if (!isUseOfPostIncrementedValue && L->contains(Inst->getParent())) {
432 if (NewBasePt && isa<PHINode>(OperandValToReplace)) {
433 InsertPt = NewBasePt;
435 } else if (Instruction *OpInst
436 = dyn_cast<Instruction>(OperandValToReplace)) {
438 while (isa<PHINode>(InsertPt)) ++InsertPt;
441 Value *NewVal = InsertCodeForBaseAtPosition(NewBase,
442 OperandValToReplace->getType(),
443 Rewriter, InsertPt, SE);
444 // Replace the use of the operand Value with the new Phi we just created.
445 Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
447 DEBUG(errs() << " Replacing with ");
448 DEBUG(WriteAsOperand(errs(), NewVal, /*PrintType=*/false));
449 DEBUG(errs() << ", which has value " << *NewBase << " plus IMM "
454 // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm
455 // expression into each operand block that uses it. Note that PHI nodes can
456 // have multiple entries for the same predecessor. We use a map to make sure
457 // that a PHI node only has a single Value* for each predecessor (which also
458 // prevents us from inserting duplicate code in some blocks).
459 DenseMap<BasicBlock*, Value*> InsertedCode;
460 PHINode *PN = cast<PHINode>(Inst);
461 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
462 if (PN->getIncomingValue(i) == OperandValToReplace) {
463 // If the original expression is outside the loop, put the replacement
464 // code in the same place as the original expression,
465 // which need not be an immediate predecessor of this PHI. This way we
466 // need only one copy of it even if it is referenced multiple times in
467 // the PHI. We don't do this when the original expression is inside the
468 // loop because multiple copies sometimes do useful sinking of code in
470 Instruction *OldLoc = dyn_cast<Instruction>(OperandValToReplace);
471 BasicBlock *PHIPred = PN->getIncomingBlock(i);
472 if (L->contains(OldLoc->getParent())) {
473 // If this is a critical edge, split the edge so that we do not insert
474 // the code on all predecessor/successor paths. We do this unless this
475 // is the canonical backedge for this loop, as this can make some
476 // inserted code be in an illegal position.
477 if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 &&
478 !isa<IndirectBrInst>(PHIPred->getTerminator()) &&
479 (PN->getParent() != L->getHeader() || !L->contains(PHIPred))) {
481 // First step, split the critical edge.
482 BasicBlock *NewBB = SplitCriticalEdge(PHIPred, PN->getParent(),
485 // Next step: move the basic block. In particular, if the PHI node
486 // is outside of the loop, and PredTI is in the loop, we want to
487 // move the block to be immediately before the PHI block, not
488 // immediately after PredTI.
489 if (L->contains(PHIPred) && !L->contains(PN->getParent()))
490 NewBB->moveBefore(PN->getParent());
492 // Splitting the edge can reduce the number of PHI entries we have.
493 e = PN->getNumIncomingValues();
495 i = PN->getBasicBlockIndex(PHIPred);
498 Value *&Code = InsertedCode[PHIPred];
500 // Insert the code into the end of the predecessor block.
501 Instruction *InsertPt = (L->contains(OldLoc->getParent())) ?
502 PHIPred->getTerminator() :
503 OldLoc->getParent()->getTerminator();
504 Code = InsertCodeForBaseAtPosition(NewBase, PN->getType(),
505 Rewriter, InsertPt, SE);
507 DEBUG(errs() << " Changing PHI use to ");
508 DEBUG(WriteAsOperand(errs(), Code, /*PrintType=*/false));
509 DEBUG(errs() << ", which has value " << *NewBase << " plus IMM "
513 // Replace the use of the operand Value with the new Phi we just created.
514 PN->setIncomingValue(i, Code);
519 // PHI node might have become a constant value after SplitCriticalEdge.
520 DeadInsts.push_back(Inst);
524 /// fitsInAddressMode - Return true if V can be subsumed within an addressing
525 /// mode, and does not need to be put in a register first.
526 static bool fitsInAddressMode(const SCEV *const &V, const Type *AccessTy,
527 const TargetLowering *TLI, bool HasBaseReg) {
528 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
529 int64_t VC = SC->getValue()->getSExtValue();
531 TargetLowering::AddrMode AM;
533 AM.HasBaseReg = HasBaseReg;
534 return TLI->isLegalAddressingMode(AM, AccessTy);
536 // Defaults to PPC. PPC allows a sign-extended 16-bit immediate field.
537 return (VC > -(1 << 16) && VC < (1 << 16)-1);
541 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
542 if (GlobalValue *GV = dyn_cast<GlobalValue>(SU->getValue())) {
544 TargetLowering::AddrMode AM;
546 AM.HasBaseReg = HasBaseReg;
547 return TLI->isLegalAddressingMode(AM, AccessTy);
549 // Default: assume global addresses are not legal.
556 /// MoveLoopVariantsToImmediateField - Move any subexpressions from Val that are
557 /// loop varying to the Imm operand.
558 static void MoveLoopVariantsToImmediateField(const SCEV *&Val, const SCEV *&Imm,
559 Loop *L, ScalarEvolution *SE) {
560 if (Val->isLoopInvariant(L)) return; // Nothing to do.
562 if (const SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
563 SmallVector<const SCEV *, 4> NewOps;
564 NewOps.reserve(SAE->getNumOperands());
566 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
567 if (!SAE->getOperand(i)->isLoopInvariant(L)) {
568 // If this is a loop-variant expression, it must stay in the immediate
569 // field of the expression.
570 Imm = SE->getAddExpr(Imm, SAE->getOperand(i));
572 NewOps.push_back(SAE->getOperand(i));
576 Val = SE->getIntegerSCEV(0, Val->getType());
578 Val = SE->getAddExpr(NewOps);
579 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
580 // Try to pull immediates out of the start value of nested addrec's.
581 const SCEV *Start = SARE->getStart();
582 MoveLoopVariantsToImmediateField(Start, Imm, L, SE);
584 SmallVector<const SCEV *, 4> Ops(SARE->op_begin(), SARE->op_end());
586 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
588 // Otherwise, all of Val is variant, move the whole thing over.
589 Imm = SE->getAddExpr(Imm, Val);
590 Val = SE->getIntegerSCEV(0, Val->getType());
595 /// MoveImmediateValues - Look at Val, and pull out any additions of constants
596 /// that can fit into the immediate field of instructions in the target.
597 /// Accumulate these immediate values into the Imm value.
598 static void MoveImmediateValues(const TargetLowering *TLI,
599 const Type *AccessTy,
600 const SCEV *&Val, const SCEV *&Imm,
601 bool isAddress, Loop *L,
602 ScalarEvolution *SE) {
603 if (const SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
604 SmallVector<const SCEV *, 4> NewOps;
605 NewOps.reserve(SAE->getNumOperands());
607 for (unsigned i = 0; i != SAE->getNumOperands(); ++i) {
608 const SCEV *NewOp = SAE->getOperand(i);
609 MoveImmediateValues(TLI, AccessTy, NewOp, Imm, isAddress, L, SE);
611 if (!NewOp->isLoopInvariant(L)) {
612 // If this is a loop-variant expression, it must stay in the immediate
613 // field of the expression.
614 Imm = SE->getAddExpr(Imm, NewOp);
616 NewOps.push_back(NewOp);
621 Val = SE->getIntegerSCEV(0, Val->getType());
623 Val = SE->getAddExpr(NewOps);
625 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
626 // Try to pull immediates out of the start value of nested addrec's.
627 const SCEV *Start = SARE->getStart();
628 MoveImmediateValues(TLI, AccessTy, Start, Imm, isAddress, L, SE);
630 if (Start != SARE->getStart()) {
631 SmallVector<const SCEV *, 4> Ops(SARE->op_begin(), SARE->op_end());
633 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
636 } else if (const SCEVMulExpr *SME = dyn_cast<SCEVMulExpr>(Val)) {
637 // Transform "8 * (4 + v)" -> "32 + 8*V" if "32" fits in the immed field.
639 fitsInAddressMode(SME->getOperand(0), AccessTy, TLI, false) &&
640 SME->getNumOperands() == 2 && SME->isLoopInvariant(L)) {
642 const SCEV *SubImm = SE->getIntegerSCEV(0, Val->getType());
643 const SCEV *NewOp = SME->getOperand(1);
644 MoveImmediateValues(TLI, AccessTy, NewOp, SubImm, isAddress, L, SE);
646 // If we extracted something out of the subexpressions, see if we can
648 if (NewOp != SME->getOperand(1)) {
649 // Scale SubImm up by "8". If the result is a target constant, we are
651 SubImm = SE->getMulExpr(SubImm, SME->getOperand(0));
652 if (fitsInAddressMode(SubImm, AccessTy, TLI, false)) {
653 // Accumulate the immediate.
654 Imm = SE->getAddExpr(Imm, SubImm);
656 // Update what is left of 'Val'.
657 Val = SE->getMulExpr(SME->getOperand(0), NewOp);
664 // Loop-variant expressions must stay in the immediate field of the
666 if ((isAddress && fitsInAddressMode(Val, AccessTy, TLI, false)) ||
667 !Val->isLoopInvariant(L)) {
668 Imm = SE->getAddExpr(Imm, Val);
669 Val = SE->getIntegerSCEV(0, Val->getType());
673 // Otherwise, no immediates to move.
676 static void MoveImmediateValues(const TargetLowering *TLI,
678 const SCEV *&Val, const SCEV *&Imm,
679 bool isAddress, Loop *L,
680 ScalarEvolution *SE) {
681 const Type *AccessTy = getAccessType(User);
682 MoveImmediateValues(TLI, AccessTy, Val, Imm, isAddress, L, SE);
685 /// SeparateSubExprs - Decompose Expr into all of the subexpressions that are
686 /// added together. This is used to reassociate common addition subexprs
687 /// together for maximal sharing when rewriting bases.
688 static void SeparateSubExprs(SmallVector<const SCEV *, 16> &SubExprs,
690 ScalarEvolution *SE) {
691 if (const SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) {
692 for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
693 SeparateSubExprs(SubExprs, AE->getOperand(j), SE);
694 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) {
695 const SCEV *Zero = SE->getIntegerSCEV(0, Expr->getType());
696 if (SARE->getOperand(0) == Zero) {
697 SubExprs.push_back(Expr);
699 // Compute the addrec with zero as its base.
700 SmallVector<const SCEV *, 4> Ops(SARE->op_begin(), SARE->op_end());
701 Ops[0] = Zero; // Start with zero base.
702 SubExprs.push_back(SE->getAddRecExpr(Ops, SARE->getLoop()));
705 SeparateSubExprs(SubExprs, SARE->getOperand(0), SE);
707 } else if (!Expr->isZero()) {
709 SubExprs.push_back(Expr);
713 // This is logically local to the following function, but C++ says we have
714 // to make it file scope.
715 struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; };
717 /// RemoveCommonExpressionsFromUseBases - Look through all of the Bases of all
718 /// the Uses, removing any common subexpressions, except that if all such
719 /// subexpressions can be folded into an addressing mode for all uses inside
720 /// the loop (this case is referred to as "free" in comments herein) we do
721 /// not remove anything. This looks for things like (a+b+c) and
722 /// (a+c+d) and computes the common (a+c) subexpression. The common expression
723 /// is *removed* from the Bases and returned.
725 RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses,
726 ScalarEvolution *SE, Loop *L,
727 const TargetLowering *TLI) {
728 unsigned NumUses = Uses.size();
730 // Only one use? This is a very common case, so we handle it specially and
732 const SCEV *Zero = SE->getIntegerSCEV(0, Uses[0].Base->getType());
733 const SCEV *Result = Zero;
734 const SCEV *FreeResult = Zero;
736 // If the use is inside the loop, use its base, regardless of what it is:
737 // it is clearly shared across all the IV's. If the use is outside the loop
738 // (which means after it) we don't want to factor anything *into* the loop,
739 // so just use 0 as the base.
740 if (L->contains(Uses[0].Inst->getParent()))
741 std::swap(Result, Uses[0].Base);
745 // To find common subexpressions, count how many of Uses use each expression.
746 // If any subexpressions are used Uses.size() times, they are common.
747 // Also track whether all uses of each expression can be moved into an
748 // an addressing mode "for free"; such expressions are left within the loop.
749 // struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; };
750 std::map<const SCEV *, SubExprUseData> SubExpressionUseData;
752 // UniqueSubExprs - Keep track of all of the subexpressions we see in the
753 // order we see them.
754 SmallVector<const SCEV *, 16> UniqueSubExprs;
756 SmallVector<const SCEV *, 16> SubExprs;
757 unsigned NumUsesInsideLoop = 0;
758 for (unsigned i = 0; i != NumUses; ++i) {
759 // If the user is outside the loop, just ignore it for base computation.
760 // Since the user is outside the loop, it must be *after* the loop (if it
761 // were before, it could not be based on the loop IV). We don't want users
762 // after the loop to affect base computation of values *inside* the loop,
763 // because we can always add their offsets to the result IV after the loop
764 // is done, ensuring we get good code inside the loop.
765 if (!L->contains(Uses[i].Inst->getParent()))
769 // If the base is zero (which is common), return zero now, there are no
771 if (Uses[i].Base == Zero) return Zero;
773 // If this use is as an address we may be able to put CSEs in the addressing
774 // mode rather than hoisting them.
775 bool isAddrUse = isAddressUse(Uses[i].Inst, Uses[i].OperandValToReplace);
776 // We may need the AccessTy below, but only when isAddrUse, so compute it
777 // only in that case.
778 const Type *AccessTy = 0;
780 AccessTy = getAccessType(Uses[i].Inst);
782 // Split the expression into subexprs.
783 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
784 // Add one to SubExpressionUseData.Count for each subexpr present, and
785 // if the subexpr is not a valid immediate within an addressing mode use,
786 // set SubExpressionUseData.notAllUsesAreFree. We definitely want to
787 // hoist these out of the loop (if they are common to all uses).
788 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) {
789 if (++SubExpressionUseData[SubExprs[j]].Count == 1)
790 UniqueSubExprs.push_back(SubExprs[j]);
791 if (!isAddrUse || !fitsInAddressMode(SubExprs[j], AccessTy, TLI, false))
792 SubExpressionUseData[SubExprs[j]].notAllUsesAreFree = true;
797 // Now that we know how many times each is used, build Result. Iterate over
798 // UniqueSubexprs so that we have a stable ordering.
799 for (unsigned i = 0, e = UniqueSubExprs.size(); i != e; ++i) {
800 std::map<const SCEV *, SubExprUseData>::iterator I =
801 SubExpressionUseData.find(UniqueSubExprs[i]);
802 assert(I != SubExpressionUseData.end() && "Entry not found?");
803 if (I->second.Count == NumUsesInsideLoop) { // Found CSE!
804 if (I->second.notAllUsesAreFree)
805 Result = SE->getAddExpr(Result, I->first);
807 FreeResult = SE->getAddExpr(FreeResult, I->first);
809 // Remove non-cse's from SubExpressionUseData.
810 SubExpressionUseData.erase(I);
813 if (FreeResult != Zero) {
814 // We have some subexpressions that can be subsumed into addressing
815 // modes in every use inside the loop. However, it's possible that
816 // there are so many of them that the combined FreeResult cannot
817 // be subsumed, or that the target cannot handle both a FreeResult
818 // and a Result in the same instruction (for example because it would
819 // require too many registers). Check this.
820 for (unsigned i=0; i<NumUses; ++i) {
821 if (!L->contains(Uses[i].Inst->getParent()))
823 // We know this is an addressing mode use; if there are any uses that
824 // are not, FreeResult would be Zero.
825 const Type *AccessTy = getAccessType(Uses[i].Inst);
826 if (!fitsInAddressMode(FreeResult, AccessTy, TLI, Result!=Zero)) {
827 // FIXME: could split up FreeResult into pieces here, some hoisted
828 // and some not. There is no obvious advantage to this.
829 Result = SE->getAddExpr(Result, FreeResult);
836 // If we found no CSE's, return now.
837 if (Result == Zero) return Result;
839 // If we still have a FreeResult, remove its subexpressions from
840 // SubExpressionUseData. This means they will remain in the use Bases.
841 if (FreeResult != Zero) {
842 SeparateSubExprs(SubExprs, FreeResult, SE);
843 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) {
844 std::map<const SCEV *, SubExprUseData>::iterator I =
845 SubExpressionUseData.find(SubExprs[j]);
846 SubExpressionUseData.erase(I);
851 // Otherwise, remove all of the CSE's we found from each of the base values.
852 for (unsigned i = 0; i != NumUses; ++i) {
853 // Uses outside the loop don't necessarily include the common base, but
854 // the final IV value coming into those uses does. Instead of trying to
855 // remove the pieces of the common base, which might not be there,
856 // subtract off the base to compensate for this.
857 if (!L->contains(Uses[i].Inst->getParent())) {
858 Uses[i].Base = SE->getMinusSCEV(Uses[i].Base, Result);
862 // Split the expression into subexprs.
863 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
865 // Remove any common subexpressions.
866 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
867 if (SubExpressionUseData.count(SubExprs[j])) {
868 SubExprs.erase(SubExprs.begin()+j);
872 // Finally, add the non-shared expressions together.
873 if (SubExprs.empty())
876 Uses[i].Base = SE->getAddExpr(SubExprs);
883 /// ValidScale - Check whether the given Scale is valid for all loads and
884 /// stores in UsersToProcess.
886 bool LoopStrengthReduce::ValidScale(bool HasBaseReg, int64_t Scale,
887 const std::vector<BasedUser>& UsersToProcess) {
891 for (unsigned i = 0, e = UsersToProcess.size(); i!=e; ++i) {
892 // If this is a load or other access, pass the type of the access in.
893 const Type *AccessTy =
894 Type::getVoidTy(UsersToProcess[i].Inst->getContext());
895 if (isAddressUse(UsersToProcess[i].Inst,
896 UsersToProcess[i].OperandValToReplace))
897 AccessTy = getAccessType(UsersToProcess[i].Inst);
898 else if (isa<PHINode>(UsersToProcess[i].Inst))
901 TargetLowering::AddrMode AM;
902 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
903 AM.BaseOffs = SC->getValue()->getSExtValue();
904 AM.HasBaseReg = HasBaseReg || !UsersToProcess[i].Base->isZero();
907 // If load[imm+r*scale] is illegal, bail out.
908 if (!TLI->isLegalAddressingMode(AM, AccessTy))
914 /// ValidOffset - Check whether the given Offset is valid for all loads and
915 /// stores in UsersToProcess.
917 bool LoopStrengthReduce::ValidOffset(bool HasBaseReg,
920 const std::vector<BasedUser>& UsersToProcess) {
924 for (unsigned i=0, e = UsersToProcess.size(); i!=e; ++i) {
925 // If this is a load or other access, pass the type of the access in.
926 const Type *AccessTy =
927 Type::getVoidTy(UsersToProcess[i].Inst->getContext());
928 if (isAddressUse(UsersToProcess[i].Inst,
929 UsersToProcess[i].OperandValToReplace))
930 AccessTy = getAccessType(UsersToProcess[i].Inst);
931 else if (isa<PHINode>(UsersToProcess[i].Inst))
934 TargetLowering::AddrMode AM;
935 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
936 AM.BaseOffs = SC->getValue()->getSExtValue();
937 AM.BaseOffs = (uint64_t)AM.BaseOffs + (uint64_t)Offset;
938 AM.HasBaseReg = HasBaseReg || !UsersToProcess[i].Base->isZero();
941 // If load[imm+r*scale] is illegal, bail out.
942 if (!TLI->isLegalAddressingMode(AM, AccessTy))
948 /// RequiresTypeConversion - Returns true if converting Ty1 to Ty2 is not
950 bool LoopStrengthReduce::RequiresTypeConversion(const Type *Ty1,
954 Ty1 = SE->getEffectiveSCEVType(Ty1);
955 Ty2 = SE->getEffectiveSCEVType(Ty2);
958 if (Ty1->canLosslesslyBitCastTo(Ty2))
960 if (TLI && TLI->isTruncateFree(Ty1, Ty2))
965 /// CheckForIVReuse - Returns the multiple if the stride is the multiple
966 /// of a previous stride and it is a legal value for the target addressing
967 /// mode scale component and optional base reg. This allows the users of
968 /// this stride to be rewritten as prev iv * factor. It returns 0 if no
969 /// reuse is possible. Factors can be negative on same targets, e.g. ARM.
971 /// If all uses are outside the loop, we don't require that all multiplies
972 /// be folded into the addressing mode, nor even that the factor be constant;
973 /// a multiply (executed once) outside the loop is better than another IV
974 /// within. Well, usually.
975 const SCEV *LoopStrengthReduce::CheckForIVReuse(bool HasBaseReg,
976 bool AllUsesAreAddresses,
977 bool AllUsesAreOutsideLoop,
978 const SCEV *const &Stride,
979 IVExpr &IV, const Type *Ty,
980 const std::vector<BasedUser>& UsersToProcess) {
981 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Stride)) {
982 int64_t SInt = SC->getValue()->getSExtValue();
983 for (unsigned NewStride = 0, e = IU->StrideOrder.size();
984 NewStride != e; ++NewStride) {
985 std::map<const SCEV *, IVsOfOneStride>::iterator SI =
986 IVsByStride.find(IU->StrideOrder[NewStride]);
987 if (SI == IVsByStride.end() || !isa<SCEVConstant>(SI->first))
989 // The other stride has no uses, don't reuse it.
990 std::map<const SCEV *, IVUsersOfOneStride *>::iterator UI =
991 IU->IVUsesByStride.find(IU->StrideOrder[NewStride]);
992 if (UI->second->Users.empty())
994 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
995 if (SI->first != Stride &&
996 (unsigned(abs64(SInt)) < SSInt || (SInt % SSInt) != 0))
998 int64_t Scale = SInt / SSInt;
999 // Check that this stride is valid for all the types used for loads and
1000 // stores; if it can be used for some and not others, we might as well use
1001 // the original stride everywhere, since we have to create the IV for it
1002 // anyway. If the scale is 1, then we don't need to worry about folding
1005 (AllUsesAreAddresses &&
1006 ValidScale(HasBaseReg, Scale, UsersToProcess))) {
1007 // Prefer to reuse an IV with a base of zero.
1008 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1009 IE = SI->second.IVs.end(); II != IE; ++II)
1010 // Only reuse previous IV if it would not require a type conversion
1011 // and if the base difference can be folded.
1012 if (II->Base->isZero() &&
1013 !RequiresTypeConversion(II->Base->getType(), Ty)) {
1015 return SE->getIntegerSCEV(Scale, Stride->getType());
1017 // Otherwise, settle for an IV with a foldable base.
1018 if (AllUsesAreAddresses)
1019 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1020 IE = SI->second.IVs.end(); II != IE; ++II)
1021 // Only reuse previous IV if it would not require a type conversion
1022 // and if the base difference can be folded.
1023 if (SE->getEffectiveSCEVType(II->Base->getType()) ==
1024 SE->getEffectiveSCEVType(Ty) &&
1025 isa<SCEVConstant>(II->Base)) {
1027 cast<SCEVConstant>(II->Base)->getValue()->getSExtValue();
1028 if (Base > INT32_MIN && Base <= INT32_MAX &&
1029 ValidOffset(HasBaseReg, -Base * Scale,
1030 Scale, UsersToProcess)) {
1032 return SE->getIntegerSCEV(Scale, Stride->getType());
1037 } else if (AllUsesAreOutsideLoop) {
1038 // Accept nonconstant strides here; it is really really right to substitute
1039 // an existing IV if we can.
1040 for (unsigned NewStride = 0, e = IU->StrideOrder.size();
1041 NewStride != e; ++NewStride) {
1042 std::map<const SCEV *, IVsOfOneStride>::iterator SI =
1043 IVsByStride.find(IU->StrideOrder[NewStride]);
1044 if (SI == IVsByStride.end() || !isa<SCEVConstant>(SI->first))
1046 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1047 if (SI->first != Stride && SSInt != 1)
1049 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1050 IE = SI->second.IVs.end(); II != IE; ++II)
1051 // Accept nonzero base here.
1052 // Only reuse previous IV if it would not require a type conversion.
1053 if (!RequiresTypeConversion(II->Base->getType(), Ty)) {
1058 // Special case, old IV is -1*x and this one is x. Can treat this one as
1060 for (unsigned NewStride = 0, e = IU->StrideOrder.size();
1061 NewStride != e; ++NewStride) {
1062 std::map<const SCEV *, IVsOfOneStride>::iterator SI =
1063 IVsByStride.find(IU->StrideOrder[NewStride]);
1064 if (SI == IVsByStride.end())
1066 if (const SCEVMulExpr *ME = dyn_cast<SCEVMulExpr>(SI->first))
1067 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(ME->getOperand(0)))
1068 if (Stride == ME->getOperand(1) &&
1069 SC->getValue()->getSExtValue() == -1LL)
1070 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1071 IE = SI->second.IVs.end(); II != IE; ++II)
1072 // Accept nonzero base here.
1073 // Only reuse previous IV if it would not require type conversion.
1074 if (!RequiresTypeConversion(II->Base->getType(), Ty)) {
1076 return SE->getIntegerSCEV(-1LL, Stride->getType());
1080 return SE->getIntegerSCEV(0, Stride->getType());
1083 /// PartitionByIsUseOfPostIncrementedValue - Simple boolean predicate that
1084 /// returns true if Val's isUseOfPostIncrementedValue is true.
1085 static bool PartitionByIsUseOfPostIncrementedValue(const BasedUser &Val) {
1086 return Val.isUseOfPostIncrementedValue;
1089 /// isNonConstantNegative - Return true if the specified scev is negated, but
1091 static bool isNonConstantNegative(const SCEV *const &Expr) {
1092 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Expr);
1093 if (!Mul) return false;
1095 // If there is a constant factor, it will be first.
1096 const SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
1097 if (!SC) return false;
1099 // Return true if the value is negative, this matches things like (-42 * V).
1100 return SC->getValue()->getValue().isNegative();
1103 /// CollectIVUsers - Transform our list of users and offsets to a bit more
1104 /// complex table. In this new vector, each 'BasedUser' contains 'Base', the
1105 /// base of the strided accesses, as well as the old information from Uses. We
1106 /// progressively move information from the Base field to the Imm field, until
1107 /// we eventually have the full access expression to rewrite the use.
1108 const SCEV *LoopStrengthReduce::CollectIVUsers(const SCEV *const &Stride,
1109 IVUsersOfOneStride &Uses,
1111 bool &AllUsesAreAddresses,
1112 bool &AllUsesAreOutsideLoop,
1113 std::vector<BasedUser> &UsersToProcess) {
1114 // FIXME: Generalize to non-affine IV's.
1115 if (!Stride->isLoopInvariant(L))
1116 return SE->getIntegerSCEV(0, Stride->getType());
1118 UsersToProcess.reserve(Uses.Users.size());
1119 for (ilist<IVStrideUse>::iterator I = Uses.Users.begin(),
1120 E = Uses.Users.end(); I != E; ++I) {
1121 UsersToProcess.push_back(BasedUser(*I, SE));
1123 // Move any loop variant operands from the offset field to the immediate
1124 // field of the use, so that we don't try to use something before it is
1126 MoveLoopVariantsToImmediateField(UsersToProcess.back().Base,
1127 UsersToProcess.back().Imm, L, SE);
1128 assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
1129 "Base value is not loop invariant!");
1132 // We now have a whole bunch of uses of like-strided induction variables, but
1133 // they might all have different bases. We want to emit one PHI node for this
1134 // stride which we fold as many common expressions (between the IVs) into as
1135 // possible. Start by identifying the common expressions in the base values
1136 // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
1137 // "A+B"), emit it to the preheader, then remove the expression from the
1138 // UsersToProcess base values.
1139 const SCEV *CommonExprs =
1140 RemoveCommonExpressionsFromUseBases(UsersToProcess, SE, L, TLI);
1142 // Next, figure out what we can represent in the immediate fields of
1143 // instructions. If we can represent anything there, move it to the imm
1144 // fields of the BasedUsers. We do this so that it increases the commonality
1145 // of the remaining uses.
1146 unsigned NumPHI = 0;
1147 bool HasAddress = false;
1148 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1149 // If the user is not in the current loop, this means it is using the exit
1150 // value of the IV. Do not put anything in the base, make sure it's all in
1151 // the immediate field to allow as much factoring as possible.
1152 if (!L->contains(UsersToProcess[i].Inst->getParent())) {
1153 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm,
1154 UsersToProcess[i].Base);
1155 UsersToProcess[i].Base =
1156 SE->getIntegerSCEV(0, UsersToProcess[i].Base->getType());
1158 // Not all uses are outside the loop.
1159 AllUsesAreOutsideLoop = false;
1161 // Addressing modes can be folded into loads and stores. Be careful that
1162 // the store is through the expression, not of the expression though.
1164 bool isAddress = isAddressUse(UsersToProcess[i].Inst,
1165 UsersToProcess[i].OperandValToReplace);
1166 if (isa<PHINode>(UsersToProcess[i].Inst)) {
1174 // If this use isn't an address, then not all uses are addresses.
1175 if (!isAddress && !isPHI)
1176 AllUsesAreAddresses = false;
1178 MoveImmediateValues(TLI, UsersToProcess[i].Inst, UsersToProcess[i].Base,
1179 UsersToProcess[i].Imm, isAddress, L, SE);
1183 // If one of the use is a PHI node and all other uses are addresses, still
1184 // allow iv reuse. Essentially we are trading one constant multiplication
1185 // for one fewer iv.
1187 AllUsesAreAddresses = false;
1189 // There are no in-loop address uses.
1190 if (AllUsesAreAddresses && (!HasAddress && !AllUsesAreOutsideLoop))
1191 AllUsesAreAddresses = false;
1196 /// ShouldUseFullStrengthReductionMode - Test whether full strength-reduction
1197 /// is valid and profitable for the given set of users of a stride. In
1198 /// full strength-reduction mode, all addresses at the current stride are
1199 /// strength-reduced all the way down to pointer arithmetic.
1201 bool LoopStrengthReduce::ShouldUseFullStrengthReductionMode(
1202 const std::vector<BasedUser> &UsersToProcess,
1204 bool AllUsesAreAddresses,
1205 const SCEV *Stride) {
1206 if (!EnableFullLSRMode)
1209 // The heuristics below aim to avoid increasing register pressure, but
1210 // fully strength-reducing all the addresses increases the number of
1211 // add instructions, so don't do this when optimizing for size.
1212 // TODO: If the loop is large, the savings due to simpler addresses
1213 // may oughtweight the costs of the extra increment instructions.
1214 if (L->getHeader()->getParent()->hasFnAttr(Attribute::OptimizeForSize))
1217 // TODO: For now, don't do full strength reduction if there could
1218 // potentially be greater-stride multiples of the current stride
1219 // which could reuse the current stride IV.
1220 if (IU->StrideOrder.back() != Stride)
1223 // Iterate through the uses to find conditions that automatically rule out
1225 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ) {
1226 const SCEV *Base = UsersToProcess[i].Base;
1227 const SCEV *Imm = UsersToProcess[i].Imm;
1228 // If any users have a loop-variant component, they can't be fully
1229 // strength-reduced.
1230 if (Imm && !Imm->isLoopInvariant(L))
1232 // If there are to users with the same base and the difference between
1233 // the two Imm values can't be folded into the address, full
1234 // strength reduction would increase register pressure.
1236 const SCEV *CurImm = UsersToProcess[i].Imm;
1237 if ((CurImm || Imm) && CurImm != Imm) {
1238 if (!CurImm) CurImm = SE->getIntegerSCEV(0, Stride->getType());
1239 if (!Imm) Imm = SE->getIntegerSCEV(0, Stride->getType());
1240 const Instruction *Inst = UsersToProcess[i].Inst;
1241 const Type *AccessTy = getAccessType(Inst);
1242 const SCEV *Diff = SE->getMinusSCEV(UsersToProcess[i].Imm, Imm);
1243 if (!Diff->isZero() &&
1244 (!AllUsesAreAddresses ||
1245 !fitsInAddressMode(Diff, AccessTy, TLI, /*HasBaseReg=*/true)))
1248 } while (++i != e && Base == UsersToProcess[i].Base);
1251 // If there's exactly one user in this stride, fully strength-reducing it
1252 // won't increase register pressure. If it's starting from a non-zero base,
1253 // it'll be simpler this way.
1254 if (UsersToProcess.size() == 1 && !UsersToProcess[0].Base->isZero())
1257 // Otherwise, if there are any users in this stride that don't require
1258 // a register for their base, full strength-reduction will increase
1259 // register pressure.
1260 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1261 if (UsersToProcess[i].Base->isZero())
1264 // Otherwise, go for it.
1268 /// InsertAffinePhi Create and insert a PHI node for an induction variable
1269 /// with the specified start and step values in the specified loop.
1271 /// If NegateStride is true, the stride should be negated by using a
1272 /// subtract instead of an add.
1274 /// Return the created phi node.
1276 static PHINode *InsertAffinePhi(const SCEV *Start, const SCEV *Step,
1277 Instruction *IVIncInsertPt,
1279 SCEVExpander &Rewriter) {
1280 assert(Start->isLoopInvariant(L) && "New PHI start is not loop invariant!");
1281 assert(Step->isLoopInvariant(L) && "New PHI stride is not loop invariant!");
1283 BasicBlock *Header = L->getHeader();
1284 BasicBlock *Preheader = L->getLoopPreheader();
1285 BasicBlock *LatchBlock = L->getLoopLatch();
1286 const Type *Ty = Start->getType();
1287 Ty = Rewriter.SE.getEffectiveSCEVType(Ty);
1289 PHINode *PN = PHINode::Create(Ty, "lsr.iv", Header->begin());
1290 PN->addIncoming(Rewriter.expandCodeFor(Start, Ty, Preheader->getTerminator()),
1293 // If the stride is negative, insert a sub instead of an add for the
1295 bool isNegative = isNonConstantNegative(Step);
1296 const SCEV *IncAmount = Step;
1298 IncAmount = Rewriter.SE.getNegativeSCEV(Step);
1300 // Insert an add instruction right before the terminator corresponding
1301 // to the back-edge or just before the only use. The location is determined
1302 // by the caller and passed in as IVIncInsertPt.
1303 Value *StepV = Rewriter.expandCodeFor(IncAmount, Ty,
1304 Preheader->getTerminator());
1307 IncV = BinaryOperator::CreateSub(PN, StepV, "lsr.iv.next",
1310 IncV = BinaryOperator::CreateAdd(PN, StepV, "lsr.iv.next",
1313 if (!isa<ConstantInt>(StepV)) ++NumVariable;
1315 PN->addIncoming(IncV, LatchBlock);
1321 static void SortUsersToProcess(std::vector<BasedUser> &UsersToProcess) {
1322 // We want to emit code for users inside the loop first. To do this, we
1323 // rearrange BasedUser so that the entries at the end have
1324 // isUseOfPostIncrementedValue = false, because we pop off the end of the
1325 // vector (so we handle them first).
1326 std::partition(UsersToProcess.begin(), UsersToProcess.end(),
1327 PartitionByIsUseOfPostIncrementedValue);
1329 // Sort this by base, so that things with the same base are handled
1330 // together. By partitioning first and stable-sorting later, we are
1331 // guaranteed that within each base we will pop off users from within the
1332 // loop before users outside of the loop with a particular base.
1334 // We would like to use stable_sort here, but we can't. The problem is that
1335 // const SCEV *'s don't have a deterministic ordering w.r.t to each other, so
1336 // we don't have anything to do a '<' comparison on. Because we think the
1337 // number of uses is small, do a horrible bubble sort which just relies on
1339 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1340 // Get a base value.
1341 const SCEV *Base = UsersToProcess[i].Base;
1343 // Compact everything with this base to be consecutive with this one.
1344 for (unsigned j = i+1; j != e; ++j) {
1345 if (UsersToProcess[j].Base == Base) {
1346 std::swap(UsersToProcess[i+1], UsersToProcess[j]);
1353 /// PrepareToStrengthReduceFully - Prepare to fully strength-reduce
1354 /// UsersToProcess, meaning lowering addresses all the way down to direct
1355 /// pointer arithmetic.
1358 LoopStrengthReduce::PrepareToStrengthReduceFully(
1359 std::vector<BasedUser> &UsersToProcess,
1361 const SCEV *CommonExprs,
1363 SCEVExpander &PreheaderRewriter) {
1364 DEBUG(errs() << " Fully reducing all users\n");
1366 // Rewrite the UsersToProcess records, creating a separate PHI for each
1367 // unique Base value.
1368 Instruction *IVIncInsertPt = L->getLoopLatch()->getTerminator();
1369 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ) {
1370 // TODO: The uses are grouped by base, but not sorted. We arbitrarily
1371 // pick the first Imm value here to start with, and adjust it for the
1373 const SCEV *Imm = UsersToProcess[i].Imm;
1374 const SCEV *Base = UsersToProcess[i].Base;
1375 const SCEV *Start = SE->getAddExpr(CommonExprs, Base, Imm);
1376 PHINode *Phi = InsertAffinePhi(Start, Stride, IVIncInsertPt, L,
1378 // Loop over all the users with the same base.
1380 UsersToProcess[i].Base = SE->getIntegerSCEV(0, Stride->getType());
1381 UsersToProcess[i].Imm = SE->getMinusSCEV(UsersToProcess[i].Imm, Imm);
1382 UsersToProcess[i].Phi = Phi;
1383 assert(UsersToProcess[i].Imm->isLoopInvariant(L) &&
1384 "ShouldUseFullStrengthReductionMode should reject this!");
1385 } while (++i != e && Base == UsersToProcess[i].Base);
1389 /// FindIVIncInsertPt - Return the location to insert the increment instruction.
1390 /// If the only use if a use of postinc value, (must be the loop termination
1391 /// condition), then insert it just before the use.
1392 static Instruction *FindIVIncInsertPt(std::vector<BasedUser> &UsersToProcess,
1394 if (UsersToProcess.size() == 1 &&
1395 UsersToProcess[0].isUseOfPostIncrementedValue &&
1396 L->contains(UsersToProcess[0].Inst->getParent()))
1397 return UsersToProcess[0].Inst;
1398 return L->getLoopLatch()->getTerminator();
1401 /// PrepareToStrengthReduceWithNewPhi - Insert a new induction variable for the
1402 /// given users to share.
1405 LoopStrengthReduce::PrepareToStrengthReduceWithNewPhi(
1406 std::vector<BasedUser> &UsersToProcess,
1408 const SCEV *CommonExprs,
1410 Instruction *IVIncInsertPt,
1412 SCEVExpander &PreheaderRewriter) {
1413 DEBUG(errs() << " Inserting new PHI:\n");
1415 PHINode *Phi = InsertAffinePhi(SE->getUnknown(CommonBaseV),
1416 Stride, IVIncInsertPt, L,
1419 // Remember this in case a later stride is multiple of this.
1420 IVsByStride[Stride].addIV(Stride, CommonExprs, Phi);
1422 // All the users will share this new IV.
1423 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1424 UsersToProcess[i].Phi = Phi;
1426 DEBUG(errs() << " IV=");
1427 DEBUG(WriteAsOperand(errs(), Phi, /*PrintType=*/false));
1428 DEBUG(errs() << "\n");
1431 /// PrepareToStrengthReduceFromSmallerStride - Prepare for the given users to
1432 /// reuse an induction variable with a stride that is a factor of the current
1433 /// induction variable.
1436 LoopStrengthReduce::PrepareToStrengthReduceFromSmallerStride(
1437 std::vector<BasedUser> &UsersToProcess,
1439 const IVExpr &ReuseIV,
1440 Instruction *PreInsertPt) {
1441 DEBUG(errs() << " Rewriting in terms of existing IV of STRIDE "
1442 << *ReuseIV.Stride << " and BASE " << *ReuseIV.Base << "\n");
1444 // All the users will share the reused IV.
1445 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1446 UsersToProcess[i].Phi = ReuseIV.PHI;
1448 Constant *C = dyn_cast<Constant>(CommonBaseV);
1450 (!C->isNullValue() &&
1451 !fitsInAddressMode(SE->getUnknown(CommonBaseV), CommonBaseV->getType(),
1453 // We want the common base emitted into the preheader! This is just
1454 // using cast as a copy so BitCast (no-op cast) is appropriate
1455 CommonBaseV = new BitCastInst(CommonBaseV, CommonBaseV->getType(),
1456 "commonbase", PreInsertPt);
1459 static bool IsImmFoldedIntoAddrMode(GlobalValue *GV, int64_t Offset,
1460 const Type *AccessTy,
1461 std::vector<BasedUser> &UsersToProcess,
1462 const TargetLowering *TLI) {
1463 SmallVector<Instruction*, 16> AddrModeInsts;
1464 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1465 if (UsersToProcess[i].isUseOfPostIncrementedValue)
1467 ExtAddrMode AddrMode =
1468 AddressingModeMatcher::Match(UsersToProcess[i].OperandValToReplace,
1469 AccessTy, UsersToProcess[i].Inst,
1470 AddrModeInsts, *TLI);
1471 if (GV && GV != AddrMode.BaseGV)
1473 if (Offset && !AddrMode.BaseOffs)
1474 // FIXME: How to accurate check it's immediate offset is folded.
1476 AddrModeInsts.clear();
1481 /// StrengthReduceIVUsersOfStride - Strength reduce all of the users of a single
1482 /// stride of IV. All of the users may have different starting values, and this
1483 /// may not be the only stride.
1485 LoopStrengthReduce::StrengthReduceIVUsersOfStride(const SCEV *const &Stride,
1486 IVUsersOfOneStride &Uses,
1488 // If all the users are moved to another stride, then there is nothing to do.
1489 if (Uses.Users.empty())
1492 // Keep track if every use in UsersToProcess is an address. If they all are,
1493 // we may be able to rewrite the entire collection of them in terms of a
1494 // smaller-stride IV.
1495 bool AllUsesAreAddresses = true;
1497 // Keep track if every use of a single stride is outside the loop. If so,
1498 // we want to be more aggressive about reusing a smaller-stride IV; a
1499 // multiply outside the loop is better than another IV inside. Well, usually.
1500 bool AllUsesAreOutsideLoop = true;
1502 // Transform our list of users and offsets to a bit more complex table. In
1503 // this new vector, each 'BasedUser' contains 'Base' the base of the
1504 // strided accessas well as the old information from Uses. We progressively
1505 // move information from the Base field to the Imm field, until we eventually
1506 // have the full access expression to rewrite the use.
1507 std::vector<BasedUser> UsersToProcess;
1508 const SCEV *CommonExprs = CollectIVUsers(Stride, Uses, L, AllUsesAreAddresses,
1509 AllUsesAreOutsideLoop,
1512 // Sort the UsersToProcess array so that users with common bases are
1513 // next to each other.
1514 SortUsersToProcess(UsersToProcess);
1516 // If we managed to find some expressions in common, we'll need to carry
1517 // their value in a register and add it in for each use. This will take up
1518 // a register operand, which potentially restricts what stride values are
1520 bool HaveCommonExprs = !CommonExprs->isZero();
1521 const Type *ReplacedTy = CommonExprs->getType();
1523 // If all uses are addresses, consider sinking the immediate part of the
1524 // common expression back into uses if they can fit in the immediate fields.
1525 if (TLI && HaveCommonExprs && AllUsesAreAddresses) {
1526 const SCEV *NewCommon = CommonExprs;
1527 const SCEV *Imm = SE->getIntegerSCEV(0, ReplacedTy);
1528 MoveImmediateValues(TLI, Type::getVoidTy(
1529 L->getLoopPreheader()->getContext()),
1530 NewCommon, Imm, true, L, SE);
1531 if (!Imm->isZero()) {
1534 // If the immediate part of the common expression is a GV, check if it's
1535 // possible to fold it into the target addressing mode.
1536 GlobalValue *GV = 0;
1537 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(Imm))
1538 GV = dyn_cast<GlobalValue>(SU->getValue());
1540 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Imm))
1541 Offset = SC->getValue()->getSExtValue();
1543 // Pass VoidTy as the AccessTy to be conservative, because
1544 // there could be multiple access types among all the uses.
1545 DoSink = IsImmFoldedIntoAddrMode(GV, Offset,
1546 Type::getVoidTy(L->getLoopPreheader()->getContext()),
1547 UsersToProcess, TLI);
1550 DEBUG(errs() << " Sinking " << *Imm << " back down into uses\n");
1551 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1552 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm, Imm);
1553 CommonExprs = NewCommon;
1554 HaveCommonExprs = !CommonExprs->isZero();
1560 // Now that we know what we need to do, insert the PHI node itself.
1562 DEBUG(errs() << "LSR: Examining IVs of TYPE " << *ReplacedTy << " of STRIDE "
1564 << " Common base: " << *CommonExprs << "\n");
1566 SCEVExpander Rewriter(*SE);
1567 SCEVExpander PreheaderRewriter(*SE);
1569 BasicBlock *Preheader = L->getLoopPreheader();
1570 Instruction *PreInsertPt = Preheader->getTerminator();
1571 BasicBlock *LatchBlock = L->getLoopLatch();
1572 Instruction *IVIncInsertPt = LatchBlock->getTerminator();
1574 Value *CommonBaseV = Constant::getNullValue(ReplacedTy);
1576 const SCEV *RewriteFactor = SE->getIntegerSCEV(0, ReplacedTy);
1577 IVExpr ReuseIV(SE->getIntegerSCEV(0,
1578 Type::getInt32Ty(Preheader->getContext())),
1579 SE->getIntegerSCEV(0,
1580 Type::getInt32Ty(Preheader->getContext())),
1583 // Choose a strength-reduction strategy and prepare for it by creating
1584 // the necessary PHIs and adjusting the bookkeeping.
1585 if (ShouldUseFullStrengthReductionMode(UsersToProcess, L,
1586 AllUsesAreAddresses, Stride)) {
1587 PrepareToStrengthReduceFully(UsersToProcess, Stride, CommonExprs, L,
1590 // Emit the initial base value into the loop preheader.
1591 CommonBaseV = PreheaderRewriter.expandCodeFor(CommonExprs, ReplacedTy,
1594 // If all uses are addresses, check if it is possible to reuse an IV. The
1595 // new IV must have a stride that is a multiple of the old stride; the
1596 // multiple must be a number that can be encoded in the scale field of the
1597 // target addressing mode; and we must have a valid instruction after this
1598 // substitution, including the immediate field, if any.
1599 RewriteFactor = CheckForIVReuse(HaveCommonExprs, AllUsesAreAddresses,
1600 AllUsesAreOutsideLoop,
1601 Stride, ReuseIV, ReplacedTy,
1603 if (!RewriteFactor->isZero())
1604 PrepareToStrengthReduceFromSmallerStride(UsersToProcess, CommonBaseV,
1605 ReuseIV, PreInsertPt);
1607 IVIncInsertPt = FindIVIncInsertPt(UsersToProcess, L);
1608 PrepareToStrengthReduceWithNewPhi(UsersToProcess, Stride, CommonExprs,
1609 CommonBaseV, IVIncInsertPt,
1610 L, PreheaderRewriter);
1614 // Process all the users now, replacing their strided uses with
1615 // strength-reduced forms. This outer loop handles all bases, the inner
1616 // loop handles all users of a particular base.
1617 while (!UsersToProcess.empty()) {
1618 const SCEV *Base = UsersToProcess.back().Base;
1619 Instruction *Inst = UsersToProcess.back().Inst;
1621 // Emit the code for Base into the preheader.
1623 if (!Base->isZero()) {
1624 BaseV = PreheaderRewriter.expandCodeFor(Base, 0, PreInsertPt);
1626 DEBUG(errs() << " INSERTING code for BASE = " << *Base << ":");
1627 if (BaseV->hasName())
1628 DEBUG(errs() << " Result value name = %" << BaseV->getName());
1629 DEBUG(errs() << "\n");
1631 // If BaseV is a non-zero constant, make sure that it gets inserted into
1632 // the preheader, instead of being forward substituted into the uses. We
1633 // do this by forcing a BitCast (noop cast) to be inserted into the
1634 // preheader in this case.
1635 if (!fitsInAddressMode(Base, getAccessType(Inst), TLI, false) &&
1636 isa<Constant>(BaseV)) {
1637 // We want this constant emitted into the preheader! This is just
1638 // using cast as a copy so BitCast (no-op cast) is appropriate
1639 BaseV = new BitCastInst(BaseV, BaseV->getType(), "preheaderinsert",
1644 // Emit the code to add the immediate offset to the Phi value, just before
1645 // the instructions that we identified as using this stride and base.
1647 // FIXME: Use emitted users to emit other users.
1648 BasedUser &User = UsersToProcess.back();
1650 DEBUG(errs() << " Examining ");
1651 if (User.isUseOfPostIncrementedValue)
1652 DEBUG(errs() << "postinc");
1654 DEBUG(errs() << "preinc");
1655 DEBUG(errs() << " use ");
1656 DEBUG(WriteAsOperand(errs(), UsersToProcess.back().OperandValToReplace,
1657 /*PrintType=*/false));
1658 DEBUG(errs() << " in Inst: " << *User.Inst);
1660 // If this instruction wants to use the post-incremented value, move it
1661 // after the post-inc and use its value instead of the PHI.
1662 Value *RewriteOp = User.Phi;
1663 if (User.isUseOfPostIncrementedValue) {
1664 RewriteOp = User.Phi->getIncomingValueForBlock(LatchBlock);
1665 // If this user is in the loop, make sure it is the last thing in the
1666 // loop to ensure it is dominated by the increment. In case it's the
1667 // only use of the iv, the increment instruction is already before the
1669 if (L->contains(User.Inst->getParent()) && User.Inst != IVIncInsertPt)
1670 User.Inst->moveBefore(IVIncInsertPt);
1673 const SCEV *RewriteExpr = SE->getUnknown(RewriteOp);
1675 if (SE->getEffectiveSCEVType(RewriteOp->getType()) !=
1676 SE->getEffectiveSCEVType(ReplacedTy)) {
1677 assert(SE->getTypeSizeInBits(RewriteOp->getType()) >
1678 SE->getTypeSizeInBits(ReplacedTy) &&
1679 "Unexpected widening cast!");
1680 RewriteExpr = SE->getTruncateExpr(RewriteExpr, ReplacedTy);
1683 // If we had to insert new instructions for RewriteOp, we have to
1684 // consider that they may not have been able to end up immediately
1685 // next to RewriteOp, because non-PHI instructions may never precede
1686 // PHI instructions in a block. In this case, remember where the last
1687 // instruction was inserted so that if we're replacing a different
1688 // PHI node, we can use the later point to expand the final
1690 Instruction *NewBasePt = dyn_cast<Instruction>(RewriteOp);
1691 if (RewriteOp == User.Phi) NewBasePt = 0;
1693 // Clear the SCEVExpander's expression map so that we are guaranteed
1694 // to have the code emitted where we expect it.
1697 // If we are reusing the iv, then it must be multiplied by a constant
1698 // factor to take advantage of the addressing mode scale component.
1699 if (!RewriteFactor->isZero()) {
1700 // If we're reusing an IV with a nonzero base (currently this happens
1701 // only when all reuses are outside the loop) subtract that base here.
1702 // The base has been used to initialize the PHI node but we don't want
1704 if (!ReuseIV.Base->isZero()) {
1705 const SCEV *typedBase = ReuseIV.Base;
1706 if (SE->getEffectiveSCEVType(RewriteExpr->getType()) !=
1707 SE->getEffectiveSCEVType(ReuseIV.Base->getType())) {
1708 // It's possible the original IV is a larger type than the new IV,
1709 // in which case we have to truncate the Base. We checked in
1710 // RequiresTypeConversion that this is valid.
1711 assert(SE->getTypeSizeInBits(RewriteExpr->getType()) <
1712 SE->getTypeSizeInBits(ReuseIV.Base->getType()) &&
1713 "Unexpected lengthening conversion!");
1714 typedBase = SE->getTruncateExpr(ReuseIV.Base,
1715 RewriteExpr->getType());
1717 RewriteExpr = SE->getMinusSCEV(RewriteExpr, typedBase);
1720 // Multiply old variable, with base removed, by new scale factor.
1721 RewriteExpr = SE->getMulExpr(RewriteFactor,
1724 // The common base is emitted in the loop preheader. But since we
1725 // are reusing an IV, it has not been used to initialize the PHI node.
1726 // Add it to the expression used to rewrite the uses.
1727 // When this use is outside the loop, we earlier subtracted the
1728 // common base, and are adding it back here. Use the same expression
1729 // as before, rather than CommonBaseV, so DAGCombiner will zap it.
1730 if (!CommonExprs->isZero()) {
1731 if (L->contains(User.Inst->getParent()))
1732 RewriteExpr = SE->getAddExpr(RewriteExpr,
1733 SE->getUnknown(CommonBaseV));
1735 RewriteExpr = SE->getAddExpr(RewriteExpr, CommonExprs);
1739 // Now that we know what we need to do, insert code before User for the
1740 // immediate and any loop-variant expressions.
1742 // Add BaseV to the PHI value if needed.
1743 RewriteExpr = SE->getAddExpr(RewriteExpr, SE->getUnknown(BaseV));
1745 User.RewriteInstructionToUseNewBase(RewriteExpr, NewBasePt,
1749 // Mark old value we replaced as possibly dead, so that it is eliminated
1750 // if we just replaced the last use of that value.
1751 DeadInsts.push_back(User.OperandValToReplace);
1753 UsersToProcess.pop_back();
1756 // If there are any more users to process with the same base, process them
1757 // now. We sorted by base above, so we just have to check the last elt.
1758 } while (!UsersToProcess.empty() && UsersToProcess.back().Base == Base);
1759 // TODO: Next, find out which base index is the most common, pull it out.
1762 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
1763 // different starting values, into different PHIs.
1766 void LoopStrengthReduce::StrengthReduceIVUsers(Loop *L) {
1767 // Note: this processes each stride/type pair individually. All users
1768 // passed into StrengthReduceIVUsersOfStride have the same type AND stride.
1769 // Also, note that we iterate over IVUsesByStride indirectly by using
1770 // StrideOrder. This extra layer of indirection makes the ordering of
1771 // strides deterministic - not dependent on map order.
1772 for (unsigned Stride = 0, e = IU->StrideOrder.size(); Stride != e; ++Stride) {
1773 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
1774 IU->IVUsesByStride.find(IU->StrideOrder[Stride]);
1775 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
1776 // FIXME: Generalize to non-affine IV's.
1777 if (!SI->first->isLoopInvariant(L))
1779 StrengthReduceIVUsersOfStride(SI->first, *SI->second, L);
1783 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1784 /// set the IV user and stride information and return true, otherwise return
1786 bool LoopStrengthReduce::FindIVUserForCond(ICmpInst *Cond,
1787 IVStrideUse *&CondUse,
1788 const SCEV* &CondStride) {
1789 for (unsigned Stride = 0, e = IU->StrideOrder.size();
1790 Stride != e && !CondUse; ++Stride) {
1791 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
1792 IU->IVUsesByStride.find(IU->StrideOrder[Stride]);
1793 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
1795 for (ilist<IVStrideUse>::iterator UI = SI->second->Users.begin(),
1796 E = SI->second->Users.end(); UI != E; ++UI)
1797 if (UI->getUser() == Cond) {
1798 // NOTE: we could handle setcc instructions with multiple uses here, but
1799 // InstCombine does it as well for simple uses, it's not clear that it
1800 // occurs enough in real life to handle.
1802 CondStride = SI->first;
1810 // Constant strides come first which in turns are sorted by their absolute
1811 // values. If absolute values are the same, then positive strides comes first.
1813 // 4, -1, X, 1, 2 ==> 1, -1, 2, 4, X
1814 struct StrideCompare {
1815 const ScalarEvolution *SE;
1816 explicit StrideCompare(const ScalarEvolution *se) : SE(se) {}
1818 bool operator()(const SCEV *const &LHS, const SCEV *const &RHS) {
1819 const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS);
1820 const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS);
1822 int64_t LV = LHSC->getValue()->getSExtValue();
1823 int64_t RV = RHSC->getValue()->getSExtValue();
1824 uint64_t ALV = (LV < 0) ? -LV : LV;
1825 uint64_t ARV = (RV < 0) ? -RV : RV;
1833 // If it's the same value but different type, sort by bit width so
1834 // that we emit larger induction variables before smaller
1835 // ones, letting the smaller be re-written in terms of larger ones.
1836 return SE->getTypeSizeInBits(RHS->getType()) <
1837 SE->getTypeSizeInBits(LHS->getType());
1839 return LHSC && !RHSC;
1844 /// ChangeCompareStride - If a loop termination compare instruction is the
1845 /// only use of its stride, and the compaison is against a constant value,
1846 /// try eliminate the stride by moving the compare instruction to another
1847 /// stride and change its constant operand accordingly. e.g.
1853 /// if (v2 < 10) goto loop
1858 /// if (v1 < 30) goto loop
1859 ICmpInst *LoopStrengthReduce::ChangeCompareStride(Loop *L, ICmpInst *Cond,
1860 IVStrideUse* &CondUse,
1861 const SCEV* &CondStride,
1863 // If there's only one stride in the loop, there's nothing to do here.
1864 if (IU->StrideOrder.size() < 2)
1866 // If there are other users of the condition's stride, don't bother
1867 // trying to change the condition because the stride will still
1869 std::map<const SCEV *, IVUsersOfOneStride *>::iterator I =
1870 IU->IVUsesByStride.find(CondStride);
1871 if (I == IU->IVUsesByStride.end())
1873 if (I->second->Users.size() > 1) {
1874 for (ilist<IVStrideUse>::iterator II = I->second->Users.begin(),
1875 EE = I->second->Users.end(); II != EE; ++II) {
1876 if (II->getUser() == Cond)
1878 if (!isInstructionTriviallyDead(II->getUser()))
1882 // Only handle constant strides for now.
1883 const SCEVConstant *SC = dyn_cast<SCEVConstant>(CondStride);
1884 if (!SC) return Cond;
1886 ICmpInst::Predicate Predicate = Cond->getPredicate();
1887 int64_t CmpSSInt = SC->getValue()->getSExtValue();
1888 unsigned BitWidth = SE->getTypeSizeInBits(CondStride->getType());
1889 uint64_t SignBit = 1ULL << (BitWidth-1);
1890 const Type *CmpTy = Cond->getOperand(0)->getType();
1891 const Type *NewCmpTy = NULL;
1892 unsigned TyBits = SE->getTypeSizeInBits(CmpTy);
1893 unsigned NewTyBits = 0;
1894 const SCEV *NewStride = NULL;
1895 Value *NewCmpLHS = NULL;
1896 Value *NewCmpRHS = NULL;
1898 const SCEV *NewOffset = SE->getIntegerSCEV(0, CmpTy);
1900 if (ConstantInt *C = dyn_cast<ConstantInt>(Cond->getOperand(1))) {
1901 int64_t CmpVal = C->getValue().getSExtValue();
1903 // Check the relevant induction variable for conformance to
1905 const SCEV *IV = SE->getSCEV(Cond->getOperand(0));
1906 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1907 if (!AR || !AR->isAffine())
1910 const SCEVConstant *StartC = dyn_cast<SCEVConstant>(AR->getStart());
1911 // Check stride constant and the comparision constant signs to detect
1914 if ((StartC->getValue()->getSExtValue() < CmpVal && CmpSSInt < 0) ||
1915 (StartC->getValue()->getSExtValue() > CmpVal && CmpSSInt > 0))
1918 // More restrictive check for the other cases.
1919 if ((CmpVal & SignBit) != (CmpSSInt & SignBit))
1923 // Look for a suitable stride / iv as replacement.
1924 for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
1925 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
1926 IU->IVUsesByStride.find(IU->StrideOrder[i]);
1927 if (!isa<SCEVConstant>(SI->first) || SI->second->Users.empty())
1929 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1930 if (SSInt == CmpSSInt ||
1931 abs64(SSInt) < abs64(CmpSSInt) ||
1932 (SSInt % CmpSSInt) != 0)
1935 Scale = SSInt / CmpSSInt;
1936 int64_t NewCmpVal = CmpVal * Scale;
1938 // If old icmp value fits in icmp immediate field, but the new one doesn't
1939 // try something else.
1941 TLI->isLegalICmpImmediate(CmpVal) &&
1942 !TLI->isLegalICmpImmediate(NewCmpVal))
1945 APInt Mul = APInt(BitWidth*2, CmpVal, true);
1946 Mul = Mul * APInt(BitWidth*2, Scale, true);
1947 // Check for overflow.
1948 if (!Mul.isSignedIntN(BitWidth))
1950 // Check for overflow in the stride's type too.
1951 if (!Mul.isSignedIntN(SE->getTypeSizeInBits(SI->first->getType())))
1954 // Watch out for overflow.
1955 if (ICmpInst::isSigned(Predicate) &&
1956 (CmpVal & SignBit) != (NewCmpVal & SignBit))
1959 // Pick the best iv to use trying to avoid a cast.
1961 for (ilist<IVStrideUse>::iterator UI = SI->second->Users.begin(),
1962 E = SI->second->Users.end(); UI != E; ++UI) {
1963 Value *Op = UI->getOperandValToReplace();
1965 // If the IVStrideUse implies a cast, check for an actual cast which
1966 // can be used to find the original IV expression.
1967 if (SE->getEffectiveSCEVType(Op->getType()) !=
1968 SE->getEffectiveSCEVType(SI->first->getType())) {
1969 CastInst *CI = dyn_cast<CastInst>(Op);
1970 // If it's not a simple cast, it's complicated.
1973 // If it's a cast from a type other than the stride type,
1974 // it's complicated.
1975 if (CI->getOperand(0)->getType() != SI->first->getType())
1977 // Ok, we found the IV expression in the stride's type.
1978 Op = CI->getOperand(0);
1982 if (NewCmpLHS->getType() == CmpTy)
1988 NewCmpTy = NewCmpLHS->getType();
1989 NewTyBits = SE->getTypeSizeInBits(NewCmpTy);
1990 const Type *NewCmpIntTy = IntegerType::get(Cond->getContext(), NewTyBits);
1991 if (RequiresTypeConversion(NewCmpTy, CmpTy)) {
1992 // Check if it is possible to rewrite it using
1993 // an iv / stride of a smaller integer type.
1994 unsigned Bits = NewTyBits;
1995 if (ICmpInst::isSigned(Predicate))
1997 uint64_t Mask = (1ULL << Bits) - 1;
1998 if (((uint64_t)NewCmpVal & Mask) != (uint64_t)NewCmpVal)
2002 // Don't rewrite if use offset is non-constant and the new type is
2003 // of a different type.
2004 // FIXME: too conservative?
2005 if (NewTyBits != TyBits && !isa<SCEVConstant>(CondUse->getOffset()))
2009 bool AllUsesAreAddresses = true;
2010 bool AllUsesAreOutsideLoop = true;
2011 std::vector<BasedUser> UsersToProcess;
2012 const SCEV *CommonExprs = CollectIVUsers(SI->first, *SI->second, L,
2013 AllUsesAreAddresses,
2014 AllUsesAreOutsideLoop,
2016 // Avoid rewriting the compare instruction with an iv of new stride
2017 // if it's likely the new stride uses will be rewritten using the
2018 // stride of the compare instruction.
2019 if (AllUsesAreAddresses &&
2020 ValidScale(!CommonExprs->isZero(), Scale, UsersToProcess))
2024 // Avoid rewriting the compare instruction with an iv which has
2025 // implicit extension or truncation built into it.
2026 // TODO: This is over-conservative.
2027 if (SE->getTypeSizeInBits(CondUse->getOffset()->getType()) != TyBits)
2030 // If scale is negative, use swapped predicate unless it's testing
2032 if (Scale < 0 && !Cond->isEquality())
2033 Predicate = ICmpInst::getSwappedPredicate(Predicate);
2035 NewStride = IU->StrideOrder[i];
2036 if (!isa<PointerType>(NewCmpTy))
2037 NewCmpRHS = ConstantInt::get(NewCmpTy, NewCmpVal);
2039 Constant *CI = ConstantInt::get(NewCmpIntTy, NewCmpVal);
2040 NewCmpRHS = ConstantExpr::getIntToPtr(CI, NewCmpTy);
2042 NewOffset = TyBits == NewTyBits
2043 ? SE->getMulExpr(CondUse->getOffset(),
2044 SE->getConstant(CmpTy, Scale))
2045 : SE->getConstant(NewCmpIntTy,
2046 cast<SCEVConstant>(CondUse->getOffset())->getValue()
2047 ->getSExtValue()*Scale);
2052 // Forgo this transformation if it the increment happens to be
2053 // unfortunately positioned after the condition, and the condition
2054 // has multiple uses which prevent it from being moved immediately
2055 // before the branch. See
2056 // test/Transforms/LoopStrengthReduce/change-compare-stride-trickiness-*.ll
2057 // for an example of this situation.
2058 if (!Cond->hasOneUse()) {
2059 for (BasicBlock::iterator I = Cond, E = Cond->getParent()->end();
2066 // Create a new compare instruction using new stride / iv.
2067 ICmpInst *OldCond = Cond;
2068 // Insert new compare instruction.
2069 Cond = new ICmpInst(OldCond, Predicate, NewCmpLHS, NewCmpRHS,
2070 L->getHeader()->getName() + ".termcond");
2072 DEBUG(errs() << " Change compare stride in Inst " << *OldCond);
2073 DEBUG(errs() << " to " << *Cond << '\n');
2075 // Remove the old compare instruction. The old indvar is probably dead too.
2076 DeadInsts.push_back(CondUse->getOperandValToReplace());
2077 OldCond->replaceAllUsesWith(Cond);
2078 OldCond->eraseFromParent();
2080 IU->IVUsesByStride[NewStride]->addUser(NewOffset, Cond, NewCmpLHS);
2081 CondUse = &IU->IVUsesByStride[NewStride]->Users.back();
2082 CondStride = NewStride;
2090 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
2091 /// a max computation.
2093 /// This is a narrow solution to a specific, but acute, problem. For loops
2099 /// } while (++i < n);
2101 /// the trip count isn't just 'n', because 'n' might not be positive. And
2102 /// unfortunately this can come up even for loops where the user didn't use
2103 /// a C do-while loop. For example, seemingly well-behaved top-test loops
2104 /// will commonly be lowered like this:
2110 /// } while (++i < n);
2113 /// and then it's possible for subsequent optimization to obscure the if
2114 /// test in such a way that indvars can't find it.
2116 /// When indvars can't find the if test in loops like this, it creates a
2117 /// max expression, which allows it to give the loop a canonical
2118 /// induction variable:
2121 /// max = n < 1 ? 1 : n;
2124 /// } while (++i != max);
2126 /// Canonical induction variables are necessary because the loop passes
2127 /// are designed around them. The most obvious example of this is the
2128 /// LoopInfo analysis, which doesn't remember trip count values. It
2129 /// expects to be able to rediscover the trip count each time it is
2130 /// needed, and it does this using a simple analyis that only succeeds if
2131 /// the loop has a canonical induction variable.
2133 /// However, when it comes time to generate code, the maximum operation
2134 /// can be quite costly, especially if it's inside of an outer loop.
2136 /// This function solves this problem by detecting this type of loop and
2137 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
2138 /// the instructions for the maximum computation.
2140 ICmpInst *LoopStrengthReduce::OptimizeMax(Loop *L, ICmpInst *Cond,
2141 IVStrideUse* &CondUse) {
2142 // Check that the loop matches the pattern we're looking for.
2143 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
2144 Cond->getPredicate() != CmpInst::ICMP_NE)
2147 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
2148 if (!Sel || !Sel->hasOneUse()) return Cond;
2150 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
2151 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2153 const SCEV *One = SE->getIntegerSCEV(1, BackedgeTakenCount->getType());
2155 // Add one to the backedge-taken count to get the trip count.
2156 const SCEV *IterationCount = SE->getAddExpr(BackedgeTakenCount, One);
2158 // Check for a max calculation that matches the pattern.
2159 if (!isa<SCEVSMaxExpr>(IterationCount) && !isa<SCEVUMaxExpr>(IterationCount))
2161 const SCEVNAryExpr *Max = cast<SCEVNAryExpr>(IterationCount);
2162 if (Max != SE->getSCEV(Sel)) return Cond;
2164 // To handle a max with more than two operands, this optimization would
2165 // require additional checking and setup.
2166 if (Max->getNumOperands() != 2)
2169 const SCEV *MaxLHS = Max->getOperand(0);
2170 const SCEV *MaxRHS = Max->getOperand(1);
2171 if (!MaxLHS || MaxLHS != One) return Cond;
2173 // Check the relevant induction variable for conformance to
2175 const SCEV *IV = SE->getSCEV(Cond->getOperand(0));
2176 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2177 if (!AR || !AR->isAffine() ||
2178 AR->getStart() != One ||
2179 AR->getStepRecurrence(*SE) != One)
2182 assert(AR->getLoop() == L &&
2183 "Loop condition operand is an addrec in a different loop!");
2185 // Check the right operand of the select, and remember it, as it will
2186 // be used in the new comparison instruction.
2188 if (SE->getSCEV(Sel->getOperand(1)) == MaxRHS)
2189 NewRHS = Sel->getOperand(1);
2190 else if (SE->getSCEV(Sel->getOperand(2)) == MaxRHS)
2191 NewRHS = Sel->getOperand(2);
2192 if (!NewRHS) return Cond;
2194 // Determine the new comparison opcode. It may be signed or unsigned,
2195 // and the original comparison may be either equality or inequality.
2196 CmpInst::Predicate Pred =
2197 isa<SCEVSMaxExpr>(Max) ? CmpInst::ICMP_SLT : CmpInst::ICMP_ULT;
2198 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
2199 Pred = CmpInst::getInversePredicate(Pred);
2201 // Ok, everything looks ok to change the condition into an SLT or SGE and
2202 // delete the max calculation.
2204 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
2206 // Delete the max calculation instructions.
2207 Cond->replaceAllUsesWith(NewCond);
2208 CondUse->setUser(NewCond);
2209 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2210 Cond->eraseFromParent();
2211 Sel->eraseFromParent();
2212 if (Cmp->use_empty())
2213 Cmp->eraseFromParent();
2217 /// OptimizeShadowIV - If IV is used in a int-to-float cast
2218 /// inside the loop then try to eliminate the cast opeation.
2219 void LoopStrengthReduce::OptimizeShadowIV(Loop *L) {
2221 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
2222 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2225 for (unsigned Stride = 0, e = IU->StrideOrder.size(); Stride != e;
2227 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
2228 IU->IVUsesByStride.find(IU->StrideOrder[Stride]);
2229 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
2230 if (!isa<SCEVConstant>(SI->first))
2233 for (ilist<IVStrideUse>::iterator UI = SI->second->Users.begin(),
2234 E = SI->second->Users.end(); UI != E; /* empty */) {
2235 ilist<IVStrideUse>::iterator CandidateUI = UI;
2237 Instruction *ShadowUse = CandidateUI->getUser();
2238 const Type *DestTy = NULL;
2240 /* If shadow use is a int->float cast then insert a second IV
2241 to eliminate this cast.
2243 for (unsigned i = 0; i < n; ++i)
2249 for (unsigned i = 0; i < n; ++i, ++d)
2252 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser()))
2253 DestTy = UCast->getDestTy();
2254 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser()))
2255 DestTy = SCast->getDestTy();
2256 if (!DestTy) continue;
2259 // If target does not support DestTy natively then do not apply
2260 // this transformation.
2261 EVT DVT = TLI->getValueType(DestTy);
2262 if (!TLI->isTypeLegal(DVT)) continue;
2265 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
2267 if (PH->getNumIncomingValues() != 2) continue;
2269 const Type *SrcTy = PH->getType();
2270 int Mantissa = DestTy->getFPMantissaWidth();
2271 if (Mantissa == -1) continue;
2272 if ((int)SE->getTypeSizeInBits(SrcTy) > Mantissa)
2275 unsigned Entry, Latch;
2276 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
2284 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
2285 if (!Init) continue;
2286 Constant *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());
2288 BinaryOperator *Incr =
2289 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
2290 if (!Incr) continue;
2291 if (Incr->getOpcode() != Instruction::Add
2292 && Incr->getOpcode() != Instruction::Sub)
2295 /* Initialize new IV, double d = 0.0 in above example. */
2296 ConstantInt *C = NULL;
2297 if (Incr->getOperand(0) == PH)
2298 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
2299 else if (Incr->getOperand(1) == PH)
2300 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
2306 // Ignore negative constants, as the code below doesn't handle them
2307 // correctly. TODO: Remove this restriction.
2308 if (!C->getValue().isStrictlyPositive()) continue;
2310 /* Add new PHINode. */
2311 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);
2313 /* create new increment. '++d' in above example. */
2314 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
2315 BinaryOperator *NewIncr =
2316 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
2317 Instruction::FAdd : Instruction::FSub,
2318 NewPH, CFP, "IV.S.next.", Incr);
2320 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
2321 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
2323 /* Remove cast operation */
2324 ShadowUse->replaceAllUsesWith(NewPH);
2325 ShadowUse->eraseFromParent();
2332 /// OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
2333 /// uses in the loop, look to see if we can eliminate some, in favor of using
2334 /// common indvars for the different uses.
2335 void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
2336 // TODO: implement optzns here.
2338 OptimizeShadowIV(L);
2341 bool LoopStrengthReduce::StrideMightBeShared(const SCEV* Stride, Loop *L,
2343 int64_t SInt = cast<SCEVConstant>(Stride)->getValue()->getSExtValue();
2344 for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
2345 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
2346 IU->IVUsesByStride.find(IU->StrideOrder[i]);
2347 const SCEV *Share = SI->first;
2348 if (!isa<SCEVConstant>(SI->first) || Share == Stride)
2350 int64_t SSInt = cast<SCEVConstant>(Share)->getValue()->getSExtValue();
2352 return true; // This can definitely be reused.
2353 if (unsigned(abs64(SSInt)) < SInt || (SSInt % SInt) != 0)
2355 int64_t Scale = SSInt / SInt;
2356 bool AllUsesAreAddresses = true;
2357 bool AllUsesAreOutsideLoop = true;
2358 std::vector<BasedUser> UsersToProcess;
2359 const SCEV *CommonExprs = CollectIVUsers(SI->first, *SI->second, L,
2360 AllUsesAreAddresses,
2361 AllUsesAreOutsideLoop,
2363 if (AllUsesAreAddresses &&
2364 ValidScale(!CommonExprs->isZero(), Scale, UsersToProcess)) {
2367 // Any pre-inc iv use?
2368 IVUsersOfOneStride &StrideUses = *IU->IVUsesByStride[Share];
2369 for (ilist<IVStrideUse>::iterator I = StrideUses.Users.begin(),
2370 E = StrideUses.Users.end(); I != E; ++I) {
2371 if (!I->isUseOfPostIncrementedValue())
2379 /// isUsedByExitBranch - Return true if icmp is used by a loop terminating
2380 /// conditional branch or it's and / or with other conditions before being used
2381 /// as the condition.
2382 static bool isUsedByExitBranch(ICmpInst *Cond, Loop *L) {
2383 BasicBlock *CondBB = Cond->getParent();
2384 if (!L->isLoopExiting(CondBB))
2386 BranchInst *TermBr = dyn_cast<BranchInst>(CondBB->getTerminator());
2387 if (!TermBr || !TermBr->isConditional())
2390 Value *User = *Cond->use_begin();
2391 Instruction *UserInst = dyn_cast<Instruction>(User);
2393 (UserInst->getOpcode() == Instruction::And ||
2394 UserInst->getOpcode() == Instruction::Or)) {
2395 if (!UserInst->hasOneUse() || UserInst->getParent() != CondBB)
2397 User = *User->use_begin();
2398 UserInst = dyn_cast<Instruction>(User);
2400 return User == TermBr;
2403 static bool ShouldCountToZero(ICmpInst *Cond, IVStrideUse* &CondUse,
2404 ScalarEvolution *SE, Loop *L,
2405 const TargetLowering *TLI = 0) {
2406 if (!L->contains(Cond->getParent()))
2409 if (!isa<SCEVConstant>(CondUse->getOffset()))
2412 // Handle only tests for equality for the moment.
2413 if (!Cond->isEquality() || !Cond->hasOneUse())
2415 if (!isUsedByExitBranch(Cond, L))
2418 Value *CondOp0 = Cond->getOperand(0);
2419 const SCEV *IV = SE->getSCEV(CondOp0);
2420 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2421 if (!AR || !AR->isAffine())
2424 const SCEVConstant *SC = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE));
2425 if (!SC || SC->getValue()->getSExtValue() < 0)
2426 // If it's already counting down, don't do anything.
2429 // If the RHS of the comparison is not an loop invariant, the rewrite
2430 // cannot be done. Also bail out if it's already comparing against a zero.
2431 // If we are checking this before cmp stride optimization, check if it's
2432 // comparing against a already legal immediate.
2433 Value *RHS = Cond->getOperand(1);
2434 ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS);
2435 if (!L->isLoopInvariant(RHS) ||
2436 (RHSC && RHSC->isZero()) ||
2437 (RHSC && TLI && TLI->isLegalICmpImmediate(RHSC->getSExtValue())))
2440 // Make sure the IV is only used for counting. Value may be preinc or
2441 // postinc; 2 uses in either case.
2442 if (!CondOp0->hasNUses(2))
2448 /// OptimizeLoopTermCond - Change loop terminating condition to use the
2449 /// postinc iv when possible.
2450 void LoopStrengthReduce::OptimizeLoopTermCond(Loop *L) {
2451 BasicBlock *LatchBlock = L->getLoopLatch();
2452 bool LatchExit = L->isLoopExiting(LatchBlock);
2453 SmallVector<BasicBlock*, 8> ExitingBlocks;
2454 L->getExitingBlocks(ExitingBlocks);
2456 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
2457 BasicBlock *ExitingBlock = ExitingBlocks[i];
2459 // Finally, get the terminating condition for the loop if possible. If we
2460 // can, we want to change it to use a post-incremented version of its
2461 // induction variable, to allow coalescing the live ranges for the IV into
2462 // one register value.
2464 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2467 // FIXME: Overly conservative, termination condition could be an 'or' etc..
2468 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
2471 // Search IVUsesByStride to find Cond's IVUse if there is one.
2472 IVStrideUse *CondUse = 0;
2473 const SCEV *CondStride = 0;
2474 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2475 if (!FindIVUserForCond(Cond, CondUse, CondStride))
2478 // If the latch block is exiting and it's not a single block loop, it's
2479 // not safe to use postinc iv in other exiting blocks. FIXME: overly
2480 // conservative? How about icmp stride optimization?
2481 bool UsePostInc = !(e > 1 && LatchExit && ExitingBlock != LatchBlock);
2482 if (UsePostInc && ExitingBlock != LatchBlock) {
2483 if (!Cond->hasOneUse())
2484 // See below, we don't want the condition to be cloned.
2487 // If exiting block is the latch block, we know it's safe and profitable
2488 // to transform the icmp to use post-inc iv. Otherwise do so only if it
2489 // would not reuse another iv and its iv would be reused by other uses.
2490 // We are optimizing for the case where the icmp is the only use of the
2492 IVUsersOfOneStride &StrideUses = *IU->IVUsesByStride[CondStride];
2493 for (ilist<IVStrideUse>::iterator I = StrideUses.Users.begin(),
2494 E = StrideUses.Users.end(); I != E; ++I) {
2495 if (I->getUser() == Cond)
2497 if (!I->isUseOfPostIncrementedValue()) {
2504 // If iv for the stride might be shared and any of the users use pre-inc
2505 // iv might be used, then it's not safe to use post-inc iv.
2507 isa<SCEVConstant>(CondStride) &&
2508 StrideMightBeShared(CondStride, L, true))
2512 // If the trip count is computed in terms of a max (due to ScalarEvolution
2513 // being unable to find a sufficient guard, for example), change the loop
2514 // comparison to use SLT or ULT instead of NE.
2515 Cond = OptimizeMax(L, Cond, CondUse);
2517 // If possible, change stride and operands of the compare instruction to
2518 // eliminate one stride. However, avoid rewriting the compare instruction
2519 // with an iv of new stride if it's likely the new stride uses will be
2520 // rewritten using the stride of the compare instruction.
2521 if (ExitingBlock == LatchBlock && isa<SCEVConstant>(CondStride)) {
2522 // If the condition stride is a constant and it's the only use, we might
2523 // want to optimize it first by turning it to count toward zero.
2524 if (!StrideMightBeShared(CondStride, L, false) &&
2525 !ShouldCountToZero(Cond, CondUse, SE, L, TLI))
2526 Cond = ChangeCompareStride(L, Cond, CondUse, CondStride);
2532 DEBUG(errs() << " Change loop exiting icmp to use postinc iv: "
2535 // It's possible for the setcc instruction to be anywhere in the loop, and
2536 // possible for it to have multiple users. If it is not immediately before
2537 // the exiting block branch, move it.
2538 if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
2539 if (Cond->hasOneUse()) { // Condition has a single use, just move it.
2540 Cond->moveBefore(TermBr);
2542 // Otherwise, clone the terminating condition and insert into the
2544 Cond = cast<ICmpInst>(Cond->clone());
2545 Cond->setName(L->getHeader()->getName() + ".termcond");
2546 ExitingBlock->getInstList().insert(TermBr, Cond);
2548 // Clone the IVUse, as the old use still exists!
2549 IU->IVUsesByStride[CondStride]->addUser(CondUse->getOffset(), Cond,
2550 CondUse->getOperandValToReplace());
2551 CondUse = &IU->IVUsesByStride[CondStride]->Users.back();
2555 // If we get to here, we know that we can transform the setcc instruction to
2556 // use the post-incremented version of the IV, allowing us to coalesce the
2557 // live ranges for the IV correctly.
2558 CondUse->setOffset(SE->getMinusSCEV(CondUse->getOffset(), CondStride));
2559 CondUse->setIsUseOfPostIncrementedValue(true);
2566 bool LoopStrengthReduce::OptimizeLoopCountIVOfStride(const SCEV* &Stride,
2567 IVStrideUse* &CondUse,
2569 // If the only use is an icmp of a loop exiting conditional branch, then
2570 // attempt the optimization.
2571 BasedUser User = BasedUser(*CondUse, SE);
2572 assert(isa<ICmpInst>(User.Inst) && "Expecting an ICMPInst!");
2573 ICmpInst *Cond = cast<ICmpInst>(User.Inst);
2575 // Less strict check now that compare stride optimization is done.
2576 if (!ShouldCountToZero(Cond, CondUse, SE, L))
2579 Value *CondOp0 = Cond->getOperand(0);
2580 PHINode *PHIExpr = dyn_cast<PHINode>(CondOp0);
2583 // Value tested is postinc. Find the phi node.
2584 Incr = dyn_cast<BinaryOperator>(CondOp0);
2585 // FIXME: Just use User.OperandValToReplace here?
2586 if (!Incr || Incr->getOpcode() != Instruction::Add)
2589 PHIExpr = dyn_cast<PHINode>(Incr->getOperand(0));
2592 // 1 use for preinc value, the increment.
2593 if (!PHIExpr->hasOneUse())
2596 assert(isa<PHINode>(CondOp0) &&
2597 "Unexpected loop exiting counting instruction sequence!");
2598 PHIExpr = cast<PHINode>(CondOp0);
2599 // Value tested is preinc. Find the increment.
2600 // A CmpInst is not a BinaryOperator; we depend on this.
2601 Instruction::use_iterator UI = PHIExpr->use_begin();
2602 Incr = dyn_cast<BinaryOperator>(UI);
2604 Incr = dyn_cast<BinaryOperator>(++UI);
2605 // One use for postinc value, the phi. Unnecessarily conservative?
2606 if (!Incr || !Incr->hasOneUse() || Incr->getOpcode() != Instruction::Add)
2610 // Replace the increment with a decrement.
2611 DEBUG(errs() << "LSR: Examining use ");
2612 DEBUG(WriteAsOperand(errs(), CondOp0, /*PrintType=*/false));
2613 DEBUG(errs() << " in Inst: " << *Cond << '\n');
2614 BinaryOperator *Decr = BinaryOperator::Create(Instruction::Sub,
2615 Incr->getOperand(0), Incr->getOperand(1), "tmp", Incr);
2616 Incr->replaceAllUsesWith(Decr);
2617 Incr->eraseFromParent();
2619 // Substitute endval-startval for the original startval, and 0 for the
2620 // original endval. Since we're only testing for equality this is OK even
2621 // if the computation wraps around.
2622 BasicBlock *Preheader = L->getLoopPreheader();
2623 Instruction *PreInsertPt = Preheader->getTerminator();
2624 unsigned InBlock = L->contains(PHIExpr->getIncomingBlock(0)) ? 1 : 0;
2625 Value *StartVal = PHIExpr->getIncomingValue(InBlock);
2626 Value *EndVal = Cond->getOperand(1);
2627 DEBUG(errs() << " Optimize loop counting iv to count down ["
2628 << *EndVal << " .. " << *StartVal << "]\n");
2630 // FIXME: check for case where both are constant.
2631 Constant* Zero = ConstantInt::get(Cond->getOperand(1)->getType(), 0);
2632 BinaryOperator *NewStartVal = BinaryOperator::Create(Instruction::Sub,
2633 EndVal, StartVal, "tmp", PreInsertPt);
2634 PHIExpr->setIncomingValue(InBlock, NewStartVal);
2635 Cond->setOperand(1, Zero);
2636 DEBUG(errs() << " New icmp: " << *Cond << "\n");
2638 int64_t SInt = cast<SCEVConstant>(Stride)->getValue()->getSExtValue();
2639 const SCEV *NewStride = 0;
2641 for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
2642 const SCEV *OldStride = IU->StrideOrder[i];
2643 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(OldStride))
2644 if (SC->getValue()->getSExtValue() == -SInt) {
2646 NewStride = OldStride;
2652 NewStride = SE->getIntegerSCEV(-SInt, Stride->getType());
2653 IU->AddUser(NewStride, CondUse->getOffset(), Cond, Cond->getOperand(0));
2654 IU->IVUsesByStride[Stride]->removeUser(CondUse);
2656 CondUse = &IU->IVUsesByStride[NewStride]->Users.back();
2664 /// OptimizeLoopCountIV - If, after all sharing of IVs, the IV used for deciding
2665 /// when to exit the loop is used only for that purpose, try to rearrange things
2666 /// so it counts down to a test against zero.
2667 bool LoopStrengthReduce::OptimizeLoopCountIV(Loop *L) {
2668 bool ThisChanged = false;
2669 for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
2670 const SCEV *Stride = IU->StrideOrder[i];
2671 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
2672 IU->IVUsesByStride.find(Stride);
2673 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
2674 // FIXME: Generalize to non-affine IV's.
2675 if (!SI->first->isLoopInvariant(L))
2677 // If stride is a constant and it has an icmpinst use, check if we can
2678 // optimize the loop to count down.
2679 if (isa<SCEVConstant>(Stride) && SI->second->Users.size() == 1) {
2680 Instruction *User = SI->second->Users.begin()->getUser();
2681 if (!isa<ICmpInst>(User))
2683 const SCEV *CondStride = Stride;
2684 IVStrideUse *Use = &*SI->second->Users.begin();
2685 if (!OptimizeLoopCountIVOfStride(CondStride, Use, L))
2689 // Now check if it's possible to reuse this iv for other stride uses.
2690 for (unsigned j = 0, ee = IU->StrideOrder.size(); j != ee; ++j) {
2691 const SCEV *SStride = IU->StrideOrder[j];
2692 if (SStride == CondStride)
2694 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SII =
2695 IU->IVUsesByStride.find(SStride);
2696 assert(SII != IU->IVUsesByStride.end() && "Stride doesn't exist!");
2697 // FIXME: Generalize to non-affine IV's.
2698 if (!SII->first->isLoopInvariant(L))
2700 // FIXME: Rewrite other stride using CondStride.
2705 Changed |= ThisChanged;
2709 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager &LPM) {
2710 IU = &getAnalysis<IVUsers>();
2711 SE = &getAnalysis<ScalarEvolution>();
2714 // If LoopSimplify form is not available, stay out of trouble.
2715 if (!L->getLoopPreheader() || !L->getLoopLatch())
2718 if (!IU->IVUsesByStride.empty()) {
2719 DEBUG(errs() << "\nLSR on \"" << L->getHeader()->getParent()->getName()
2723 // Sort the StrideOrder so we process larger strides first.
2724 std::stable_sort(IU->StrideOrder.begin(), IU->StrideOrder.end(),
2727 // Optimize induction variables. Some indvar uses can be transformed to use
2728 // strides that will be needed for other purposes. A common example of this
2729 // is the exit test for the loop, which can often be rewritten to use the
2730 // computation of some other indvar to decide when to terminate the loop.
2733 // Change loop terminating condition to use the postinc iv when possible
2734 // and optimize loop terminating compare. FIXME: Move this after
2735 // StrengthReduceIVUsersOfStride?
2736 OptimizeLoopTermCond(L);
2738 // FIXME: We can shrink overlarge IV's here. e.g. if the code has
2739 // computation in i64 values and the target doesn't support i64, demote
2740 // the computation to 32-bit if safe.
2742 // FIXME: Attempt to reuse values across multiple IV's. In particular, we
2743 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should
2744 // be codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC.
2745 // Need to be careful that IV's are all the same type. Only works for
2746 // intptr_t indvars.
2748 // IVsByStride keeps IVs for one particular loop.
2749 assert(IVsByStride.empty() && "Stale entries in IVsByStride?");
2751 StrengthReduceIVUsers(L);
2753 // After all sharing is done, see if we can adjust the loop to test against
2754 // zero instead of counting up to a maximum. This is usually faster.
2755 OptimizeLoopCountIV(L);
2757 // We're done analyzing this loop; release all the state we built up for it.
2758 IVsByStride.clear();
2760 // Clean up after ourselves
2761 if (!DeadInsts.empty())
2762 DeleteTriviallyDeadInstructions();
2765 // At this point, it is worth checking to see if any recurrence PHIs are also
2766 // dead, so that we can remove them as well.
2767 DeleteDeadPHIs(L->getHeader());