1 //===- LoopStrengthReduce.cpp - Strength Reduce IVs in Loops --------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This transformation analyzes and transforms the induction variables (and
11 // computations derived from them) into forms suitable for efficient execution
14 // This pass performs a strength reduction on array references inside loops that
15 // have as one or more of their components the loop induction variable, it
16 // rewrites expressions to take advantage of scaled-index addressing modes
17 // available on the target, and it performs a variety of other optimizations
18 // related to loop induction variables.
20 //===----------------------------------------------------------------------===//
22 #define DEBUG_TYPE "loop-reduce"
23 #include "llvm/Transforms/Scalar.h"
24 #include "llvm/Constants.h"
25 #include "llvm/Instructions.h"
26 #include "llvm/IntrinsicInst.h"
27 #include "llvm/Type.h"
28 #include "llvm/DerivedTypes.h"
29 #include "llvm/Analysis/IVUsers.h"
30 #include "llvm/Analysis/LoopPass.h"
31 #include "llvm/Analysis/ScalarEvolutionExpander.h"
32 #include "llvm/Transforms/Utils/AddrModeMatcher.h"
33 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
34 #include "llvm/Transforms/Utils/Local.h"
35 #include "llvm/ADT/Statistic.h"
36 #include "llvm/Support/CFG.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Support/CommandLine.h"
39 #include "llvm/Support/ValueHandle.h"
40 #include "llvm/Support/raw_ostream.h"
41 #include "llvm/Target/TargetLowering.h"
45 STATISTIC(NumReduced , "Number of IV uses strength reduced");
46 STATISTIC(NumInserted, "Number of PHIs inserted");
47 STATISTIC(NumVariable, "Number of PHIs with variable strides");
48 STATISTIC(NumEliminated, "Number of strides eliminated");
49 STATISTIC(NumShadow, "Number of Shadow IVs optimized");
50 STATISTIC(NumImmSunk, "Number of common expr immediates sunk into uses");
51 STATISTIC(NumLoopCond, "Number of loop terminating conds optimized");
52 STATISTIC(NumCountZero, "Number of count iv optimized to count toward zero");
54 static cl::opt<bool> EnableFullLSRMode("enable-full-lsr",
62 /// IVInfo - This structure keeps track of one IV expression inserted during
63 /// StrengthReduceStridedIVUsers. It contains the stride, the common base, as
64 /// well as the PHI node and increment value created for rewrite.
70 IVExpr(const SCEV *const stride, const SCEV *const base, PHINode *phi)
71 : Stride(stride), Base(base), PHI(phi) {}
74 /// IVsOfOneStride - This structure keeps track of all IV expression inserted
75 /// during StrengthReduceStridedIVUsers for a particular stride of the IV.
76 struct IVsOfOneStride {
77 std::vector<IVExpr> IVs;
79 void addIV(const SCEV *const Stride, const SCEV *const Base, PHINode *PHI) {
80 IVs.push_back(IVExpr(Stride, Base, PHI));
84 class LoopStrengthReduce : public LoopPass {
89 /// IVsByStride - Keep track of all IVs that have been inserted for a
90 /// particular stride.
91 std::map<const SCEV *, IVsOfOneStride> IVsByStride;
93 /// StrideNoReuse - Keep track of all the strides whose ivs cannot be
94 /// reused (nor should they be rewritten to reuse other strides).
95 SmallSet<const SCEV *, 4> StrideNoReuse;
97 /// DeadInsts - Keep track of instructions we may have made dead, so that
98 /// we can remove them after we are done working.
99 SmallVector<WeakVH, 16> DeadInsts;
101 /// TLI - Keep a pointer of a TargetLowering to consult for determining
102 /// transformation profitability.
103 const TargetLowering *TLI;
106 static char ID; // Pass ID, replacement for typeid
107 explicit LoopStrengthReduce(const TargetLowering *tli = NULL) :
108 LoopPass(&ID), TLI(tli) {}
110 bool runOnLoop(Loop *L, LPPassManager &LPM);
112 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
113 // We split critical edges, so we change the CFG. However, we do update
114 // many analyses if they are around.
115 AU.addPreservedID(LoopSimplifyID);
116 AU.addPreserved("loops");
117 AU.addPreserved("domfrontier");
118 AU.addPreserved("domtree");
120 AU.addRequiredID(LoopSimplifyID);
121 AU.addRequired<ScalarEvolution>();
122 AU.addPreserved<ScalarEvolution>();
123 AU.addRequired<IVUsers>();
124 AU.addPreserved<IVUsers>();
128 void OptimizeIndvars(Loop *L);
130 /// OptimizeLoopTermCond - Change loop terminating condition to use the
131 /// postinc iv when possible.
132 void OptimizeLoopTermCond(Loop *L);
134 /// OptimizeShadowIV - If IV is used in a int-to-float cast
135 /// inside the loop then try to eliminate the cast opeation.
136 void OptimizeShadowIV(Loop *L);
138 /// OptimizeMax - Rewrite the loop's terminating condition
139 /// if it uses a max computation.
140 ICmpInst *OptimizeMax(Loop *L, ICmpInst *Cond,
141 IVStrideUse* &CondUse);
143 /// OptimizeLoopCountIV - If, after all sharing of IVs, the IV used for
144 /// deciding when to exit the loop is used only for that purpose, try to
145 /// rearrange things so it counts down to a test against zero.
146 bool OptimizeLoopCountIV(Loop *L);
147 bool OptimizeLoopCountIVOfStride(const SCEV* &Stride,
148 IVStrideUse* &CondUse, Loop *L);
150 /// StrengthReduceIVUsersOfStride - Strength reduce all of the users of a
151 /// single stride of IV. All of the users may have different starting
152 /// values, and this may not be the only stride.
153 void StrengthReduceIVUsersOfStride(const SCEV *const &Stride,
154 IVUsersOfOneStride &Uses,
156 void StrengthReduceIVUsers(Loop *L);
158 ICmpInst *ChangeCompareStride(Loop *L, ICmpInst *Cond,
159 IVStrideUse* &CondUse,
160 const SCEV* &CondStride,
161 bool PostPass = false);
163 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse,
164 const SCEV* &CondStride);
165 bool RequiresTypeConversion(const Type *Ty, const Type *NewTy);
166 const SCEV *CheckForIVReuse(bool, bool, bool, const SCEV *const&,
167 IVExpr&, const Type*,
168 const std::vector<BasedUser>& UsersToProcess);
169 bool ValidScale(bool, int64_t,
170 const std::vector<BasedUser>& UsersToProcess);
171 bool ValidOffset(bool, int64_t, int64_t,
172 const std::vector<BasedUser>& UsersToProcess);
173 const SCEV *CollectIVUsers(const SCEV *const &Stride,
174 IVUsersOfOneStride &Uses,
176 bool &AllUsesAreAddresses,
177 bool &AllUsesAreOutsideLoop,
178 std::vector<BasedUser> &UsersToProcess);
179 bool StrideMightBeShared(const SCEV *Stride, Loop *L, bool CheckPreInc);
180 bool ShouldUseFullStrengthReductionMode(
181 const std::vector<BasedUser> &UsersToProcess,
183 bool AllUsesAreAddresses,
185 void PrepareToStrengthReduceFully(
186 std::vector<BasedUser> &UsersToProcess,
188 const SCEV *CommonExprs,
190 SCEVExpander &PreheaderRewriter);
191 void PrepareToStrengthReduceFromSmallerStride(
192 std::vector<BasedUser> &UsersToProcess,
194 const IVExpr &ReuseIV,
195 Instruction *PreInsertPt);
196 void PrepareToStrengthReduceWithNewPhi(
197 std::vector<BasedUser> &UsersToProcess,
199 const SCEV *CommonExprs,
201 Instruction *IVIncInsertPt,
203 SCEVExpander &PreheaderRewriter);
205 void DeleteTriviallyDeadInstructions();
209 char LoopStrengthReduce::ID = 0;
210 static RegisterPass<LoopStrengthReduce>
211 X("loop-reduce", "Loop Strength Reduction");
213 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
214 return new LoopStrengthReduce(TLI);
217 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
218 /// specified set are trivially dead, delete them and see if this makes any of
219 /// their operands subsequently dead.
220 void LoopStrengthReduce::DeleteTriviallyDeadInstructions() {
221 if (DeadInsts.empty()) return;
223 while (!DeadInsts.empty()) {
224 Instruction *I = dyn_cast_or_null<Instruction>(DeadInsts.pop_back_val());
226 if (I == 0 || !isInstructionTriviallyDead(I))
229 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
230 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
233 DeadInsts.push_back(U);
236 I->eraseFromParent();
241 /// containsAddRecFromDifferentLoop - Determine whether expression S involves a
242 /// subexpression that is an AddRec from a loop other than L. An outer loop
243 /// of L is OK, but not an inner loop nor a disjoint loop.
244 static bool containsAddRecFromDifferentLoop(const SCEV *S, Loop *L) {
245 // This is very common, put it first.
246 if (isa<SCEVConstant>(S))
248 if (const SCEVCommutativeExpr *AE = dyn_cast<SCEVCommutativeExpr>(S)) {
249 for (unsigned int i=0; i< AE->getNumOperands(); i++)
250 if (containsAddRecFromDifferentLoop(AE->getOperand(i), L))
254 if (const SCEVAddRecExpr *AE = dyn_cast<SCEVAddRecExpr>(S)) {
255 if (const Loop *newLoop = AE->getLoop()) {
258 // if newLoop is an outer loop of L, this is OK.
259 if (!newLoop->contains(L->getHeader()))
264 if (const SCEVUDivExpr *DE = dyn_cast<SCEVUDivExpr>(S))
265 return containsAddRecFromDifferentLoop(DE->getLHS(), L) ||
266 containsAddRecFromDifferentLoop(DE->getRHS(), L);
268 // SCEVSDivExpr has been backed out temporarily, but will be back; we'll
269 // need this when it is.
270 if (const SCEVSDivExpr *DE = dyn_cast<SCEVSDivExpr>(S))
271 return containsAddRecFromDifferentLoop(DE->getLHS(), L) ||
272 containsAddRecFromDifferentLoop(DE->getRHS(), L);
274 if (const SCEVCastExpr *CE = dyn_cast<SCEVCastExpr>(S))
275 return containsAddRecFromDifferentLoop(CE->getOperand(), L);
279 /// isAddressUse - Returns true if the specified instruction is using the
280 /// specified value as an address.
281 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
282 bool isAddress = isa<LoadInst>(Inst);
283 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
284 if (SI->getOperand(1) == OperandVal)
286 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
287 // Addressing modes can also be folded into prefetches and a variety
289 switch (II->getIntrinsicID()) {
291 case Intrinsic::prefetch:
292 case Intrinsic::x86_sse2_loadu_dq:
293 case Intrinsic::x86_sse2_loadu_pd:
294 case Intrinsic::x86_sse_loadu_ps:
295 case Intrinsic::x86_sse_storeu_ps:
296 case Intrinsic::x86_sse2_storeu_pd:
297 case Intrinsic::x86_sse2_storeu_dq:
298 case Intrinsic::x86_sse2_storel_dq:
299 if (II->getOperand(1) == OperandVal)
307 /// getAccessType - Return the type of the memory being accessed.
308 static const Type *getAccessType(const Instruction *Inst) {
309 const Type *AccessTy = Inst->getType();
310 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
311 AccessTy = SI->getOperand(0)->getType();
312 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
313 // Addressing modes can also be folded into prefetches and a variety
315 switch (II->getIntrinsicID()) {
317 case Intrinsic::x86_sse_storeu_ps:
318 case Intrinsic::x86_sse2_storeu_pd:
319 case Intrinsic::x86_sse2_storeu_dq:
320 case Intrinsic::x86_sse2_storel_dq:
321 AccessTy = II->getOperand(1)->getType();
329 /// BasedUser - For a particular base value, keep information about how we've
330 /// partitioned the expression so far.
332 /// SE - The current ScalarEvolution object.
335 /// Base - The Base value for the PHI node that needs to be inserted for
336 /// this use. As the use is processed, information gets moved from this
337 /// field to the Imm field (below). BasedUser values are sorted by this
341 /// Inst - The instruction using the induction variable.
344 /// OperandValToReplace - The operand value of Inst to replace with the
346 Value *OperandValToReplace;
348 /// Imm - The immediate value that should be added to the base immediately
349 /// before Inst, because it will be folded into the imm field of the
350 /// instruction. This is also sometimes used for loop-variant values that
351 /// must be added inside the loop.
354 /// Phi - The induction variable that performs the striding that
355 /// should be used for this user.
358 // isUseOfPostIncrementedValue - True if this should use the
359 // post-incremented version of this IV, not the preincremented version.
360 // This can only be set in special cases, such as the terminating setcc
361 // instruction for a loop and uses outside the loop that are dominated by
363 bool isUseOfPostIncrementedValue;
365 BasedUser(IVStrideUse &IVSU, ScalarEvolution *se)
366 : SE(se), Base(IVSU.getOffset()), Inst(IVSU.getUser()),
367 OperandValToReplace(IVSU.getOperandValToReplace()),
368 Imm(SE->getIntegerSCEV(0, Base->getType())),
369 isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue()) {}
371 // Once we rewrite the code to insert the new IVs we want, update the
372 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
374 void RewriteInstructionToUseNewBase(const SCEV *const &NewBase,
375 Instruction *InsertPt,
376 SCEVExpander &Rewriter, Loop *L, Pass *P,
377 SmallVectorImpl<WeakVH> &DeadInsts);
379 Value *InsertCodeForBaseAtPosition(const SCEV *const &NewBase,
381 SCEVExpander &Rewriter,
387 void BasedUser::dump() const {
388 errs() << " Base=" << *Base;
389 errs() << " Imm=" << *Imm;
390 errs() << " Inst: " << *Inst;
393 Value *BasedUser::InsertCodeForBaseAtPosition(const SCEV *const &NewBase,
395 SCEVExpander &Rewriter,
397 Value *Base = Rewriter.expandCodeFor(NewBase, 0, IP);
399 // Wrap the base in a SCEVUnknown so that ScalarEvolution doesn't try to
401 const SCEV *NewValSCEV = SE->getUnknown(Base);
403 // Always emit the immediate into the same block as the user.
404 NewValSCEV = SE->getAddExpr(NewValSCEV, Imm);
406 return Rewriter.expandCodeFor(NewValSCEV, Ty, IP);
410 // Once we rewrite the code to insert the new IVs we want, update the
411 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
412 // to it. NewBasePt is the last instruction which contributes to the
413 // value of NewBase in the case that it's a diffferent instruction from
414 // the PHI that NewBase is computed from, or null otherwise.
416 void BasedUser::RewriteInstructionToUseNewBase(const SCEV *const &NewBase,
417 Instruction *NewBasePt,
418 SCEVExpander &Rewriter, Loop *L, Pass *P,
419 SmallVectorImpl<WeakVH> &DeadInsts) {
420 if (!isa<PHINode>(Inst)) {
421 // By default, insert code at the user instruction.
422 BasicBlock::iterator InsertPt = Inst;
424 // However, if the Operand is itself an instruction, the (potentially
425 // complex) inserted code may be shared by many users. Because of this, we
426 // want to emit code for the computation of the operand right before its old
427 // computation. This is usually safe, because we obviously used to use the
428 // computation when it was computed in its current block. However, in some
429 // cases (e.g. use of a post-incremented induction variable) the NewBase
430 // value will be pinned to live somewhere after the original computation.
431 // In this case, we have to back off.
433 // If this is a use outside the loop (which means after, since it is based
434 // on a loop indvar) we use the post-incremented value, so that we don't
435 // artificially make the preinc value live out the bottom of the loop.
436 if (!isUseOfPostIncrementedValue && L->contains(Inst->getParent())) {
437 if (NewBasePt && isa<PHINode>(OperandValToReplace)) {
438 InsertPt = NewBasePt;
440 } else if (Instruction *OpInst
441 = dyn_cast<Instruction>(OperandValToReplace)) {
443 while (isa<PHINode>(InsertPt)) ++InsertPt;
446 Value *NewVal = InsertCodeForBaseAtPosition(NewBase,
447 OperandValToReplace->getType(),
449 // Replace the use of the operand Value with the new Phi we just created.
450 Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
452 DEBUG(errs() << " Replacing with ");
453 DEBUG(WriteAsOperand(errs(), NewVal, /*PrintType=*/false));
454 DEBUG(errs() << ", which has value " << *NewBase << " plus IMM "
459 // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm
460 // expression into each operand block that uses it. Note that PHI nodes can
461 // have multiple entries for the same predecessor. We use a map to make sure
462 // that a PHI node only has a single Value* for each predecessor (which also
463 // prevents us from inserting duplicate code in some blocks).
464 DenseMap<BasicBlock*, Value*> InsertedCode;
465 PHINode *PN = cast<PHINode>(Inst);
466 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
467 if (PN->getIncomingValue(i) == OperandValToReplace) {
468 // If the original expression is outside the loop, put the replacement
469 // code in the same place as the original expression,
470 // which need not be an immediate predecessor of this PHI. This way we
471 // need only one copy of it even if it is referenced multiple times in
472 // the PHI. We don't do this when the original expression is inside the
473 // loop because multiple copies sometimes do useful sinking of code in
475 Instruction *OldLoc = dyn_cast<Instruction>(OperandValToReplace);
476 BasicBlock *PHIPred = PN->getIncomingBlock(i);
477 if (L->contains(OldLoc->getParent())) {
478 // If this is a critical edge, split the edge so that we do not insert
479 // the code on all predecessor/successor paths. We do this unless this
480 // is the canonical backedge for this loop, as this can make some
481 // inserted code be in an illegal position.
482 if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 &&
483 !isa<IndirectBrInst>(PHIPred->getTerminator()) &&
484 (PN->getParent() != L->getHeader() || !L->contains(PHIPred))) {
486 // First step, split the critical edge.
487 BasicBlock *NewBB = SplitCriticalEdge(PHIPred, PN->getParent(),
490 // Next step: move the basic block. In particular, if the PHI node
491 // is outside of the loop, and PredTI is in the loop, we want to
492 // move the block to be immediately before the PHI block, not
493 // immediately after PredTI.
494 if (L->contains(PHIPred) && !L->contains(PN->getParent()))
495 NewBB->moveBefore(PN->getParent());
497 // Splitting the edge can reduce the number of PHI entries we have.
498 e = PN->getNumIncomingValues();
500 i = PN->getBasicBlockIndex(PHIPred);
503 Value *&Code = InsertedCode[PHIPred];
505 // Insert the code into the end of the predecessor block.
506 Instruction *InsertPt = (L->contains(OldLoc->getParent())) ?
507 PHIPred->getTerminator() :
508 OldLoc->getParent()->getTerminator();
509 Code = InsertCodeForBaseAtPosition(NewBase, PN->getType(),
512 DEBUG(errs() << " Changing PHI use to ");
513 DEBUG(WriteAsOperand(errs(), Code, /*PrintType=*/false));
514 DEBUG(errs() << ", which has value " << *NewBase << " plus IMM "
518 // Replace the use of the operand Value with the new Phi we just created.
519 PN->setIncomingValue(i, Code);
524 // PHI node might have become a constant value after SplitCriticalEdge.
525 DeadInsts.push_back(Inst);
529 /// fitsInAddressMode - Return true if V can be subsumed within an addressing
530 /// mode, and does not need to be put in a register first.
531 static bool fitsInAddressMode(const SCEV *const &V, const Type *AccessTy,
532 const TargetLowering *TLI, bool HasBaseReg) {
533 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
534 int64_t VC = SC->getValue()->getSExtValue();
536 TargetLowering::AddrMode AM;
538 AM.HasBaseReg = HasBaseReg;
539 return TLI->isLegalAddressingMode(AM, AccessTy);
541 // Defaults to PPC. PPC allows a sign-extended 16-bit immediate field.
542 return (VC > -(1 << 16) && VC < (1 << 16)-1);
546 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
547 if (GlobalValue *GV = dyn_cast<GlobalValue>(SU->getValue())) {
549 TargetLowering::AddrMode AM;
551 AM.HasBaseReg = HasBaseReg;
552 return TLI->isLegalAddressingMode(AM, AccessTy);
554 // Default: assume global addresses are not legal.
561 /// MoveLoopVariantsToImmediateField - Move any subexpressions from Val that are
562 /// loop varying to the Imm operand.
563 static void MoveLoopVariantsToImmediateField(const SCEV *&Val, const SCEV *&Imm,
564 Loop *L, ScalarEvolution *SE) {
565 if (Val->isLoopInvariant(L)) return; // Nothing to do.
567 if (const SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
568 SmallVector<const SCEV *, 4> NewOps;
569 NewOps.reserve(SAE->getNumOperands());
571 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
572 if (!SAE->getOperand(i)->isLoopInvariant(L)) {
573 // If this is a loop-variant expression, it must stay in the immediate
574 // field of the expression.
575 Imm = SE->getAddExpr(Imm, SAE->getOperand(i));
577 NewOps.push_back(SAE->getOperand(i));
581 Val = SE->getIntegerSCEV(0, Val->getType());
583 Val = SE->getAddExpr(NewOps);
584 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
585 // Try to pull immediates out of the start value of nested addrec's.
586 const SCEV *Start = SARE->getStart();
587 MoveLoopVariantsToImmediateField(Start, Imm, L, SE);
589 SmallVector<const SCEV *, 4> Ops(SARE->op_begin(), SARE->op_end());
591 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
593 // Otherwise, all of Val is variant, move the whole thing over.
594 Imm = SE->getAddExpr(Imm, Val);
595 Val = SE->getIntegerSCEV(0, Val->getType());
600 /// MoveImmediateValues - Look at Val, and pull out any additions of constants
601 /// that can fit into the immediate field of instructions in the target.
602 /// Accumulate these immediate values into the Imm value.
603 static void MoveImmediateValues(const TargetLowering *TLI,
604 const Type *AccessTy,
605 const SCEV *&Val, const SCEV *&Imm,
606 bool isAddress, Loop *L,
607 ScalarEvolution *SE) {
608 if (const SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
609 SmallVector<const SCEV *, 4> NewOps;
610 NewOps.reserve(SAE->getNumOperands());
612 for (unsigned i = 0; i != SAE->getNumOperands(); ++i) {
613 const SCEV *NewOp = SAE->getOperand(i);
614 MoveImmediateValues(TLI, AccessTy, NewOp, Imm, isAddress, L, SE);
616 if (!NewOp->isLoopInvariant(L)) {
617 // If this is a loop-variant expression, it must stay in the immediate
618 // field of the expression.
619 Imm = SE->getAddExpr(Imm, NewOp);
621 NewOps.push_back(NewOp);
626 Val = SE->getIntegerSCEV(0, Val->getType());
628 Val = SE->getAddExpr(NewOps);
630 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
631 // Try to pull immediates out of the start value of nested addrec's.
632 const SCEV *Start = SARE->getStart();
633 MoveImmediateValues(TLI, AccessTy, Start, Imm, isAddress, L, SE);
635 if (Start != SARE->getStart()) {
636 SmallVector<const SCEV *, 4> Ops(SARE->op_begin(), SARE->op_end());
638 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
641 } else if (const SCEVMulExpr *SME = dyn_cast<SCEVMulExpr>(Val)) {
642 // Transform "8 * (4 + v)" -> "32 + 8*V" if "32" fits in the immed field.
644 fitsInAddressMode(SME->getOperand(0), AccessTy, TLI, false) &&
645 SME->getNumOperands() == 2 && SME->isLoopInvariant(L)) {
647 const SCEV *SubImm = SE->getIntegerSCEV(0, Val->getType());
648 const SCEV *NewOp = SME->getOperand(1);
649 MoveImmediateValues(TLI, AccessTy, NewOp, SubImm, isAddress, L, SE);
651 // If we extracted something out of the subexpressions, see if we can
653 if (NewOp != SME->getOperand(1)) {
654 // Scale SubImm up by "8". If the result is a target constant, we are
656 SubImm = SE->getMulExpr(SubImm, SME->getOperand(0));
657 if (fitsInAddressMode(SubImm, AccessTy, TLI, false)) {
658 // Accumulate the immediate.
659 Imm = SE->getAddExpr(Imm, SubImm);
661 // Update what is left of 'Val'.
662 Val = SE->getMulExpr(SME->getOperand(0), NewOp);
669 // Loop-variant expressions must stay in the immediate field of the
671 if ((isAddress && fitsInAddressMode(Val, AccessTy, TLI, false)) ||
672 !Val->isLoopInvariant(L)) {
673 Imm = SE->getAddExpr(Imm, Val);
674 Val = SE->getIntegerSCEV(0, Val->getType());
678 // Otherwise, no immediates to move.
681 static void MoveImmediateValues(const TargetLowering *TLI,
683 const SCEV *&Val, const SCEV *&Imm,
684 bool isAddress, Loop *L,
685 ScalarEvolution *SE) {
686 const Type *AccessTy = getAccessType(User);
687 MoveImmediateValues(TLI, AccessTy, Val, Imm, isAddress, L, SE);
690 /// SeparateSubExprs - Decompose Expr into all of the subexpressions that are
691 /// added together. This is used to reassociate common addition subexprs
692 /// together for maximal sharing when rewriting bases.
693 static void SeparateSubExprs(SmallVector<const SCEV *, 16> &SubExprs,
695 ScalarEvolution *SE) {
696 if (const SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) {
697 for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
698 SeparateSubExprs(SubExprs, AE->getOperand(j), SE);
699 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) {
700 const SCEV *Zero = SE->getIntegerSCEV(0, Expr->getType());
701 if (SARE->getOperand(0) == Zero) {
702 SubExprs.push_back(Expr);
704 // Compute the addrec with zero as its base.
705 SmallVector<const SCEV *, 4> Ops(SARE->op_begin(), SARE->op_end());
706 Ops[0] = Zero; // Start with zero base.
707 SubExprs.push_back(SE->getAddRecExpr(Ops, SARE->getLoop()));
710 SeparateSubExprs(SubExprs, SARE->getOperand(0), SE);
712 } else if (!Expr->isZero()) {
714 SubExprs.push_back(Expr);
718 // This is logically local to the following function, but C++ says we have
719 // to make it file scope.
720 struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; };
722 /// RemoveCommonExpressionsFromUseBases - Look through all of the Bases of all
723 /// the Uses, removing any common subexpressions, except that if all such
724 /// subexpressions can be folded into an addressing mode for all uses inside
725 /// the loop (this case is referred to as "free" in comments herein) we do
726 /// not remove anything. This looks for things like (a+b+c) and
727 /// (a+c+d) and computes the common (a+c) subexpression. The common expression
728 /// is *removed* from the Bases and returned.
730 RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses,
731 ScalarEvolution *SE, Loop *L,
732 const TargetLowering *TLI) {
733 unsigned NumUses = Uses.size();
735 // Only one use? This is a very common case, so we handle it specially and
737 const SCEV *Zero = SE->getIntegerSCEV(0, Uses[0].Base->getType());
738 const SCEV *Result = Zero;
739 const SCEV *FreeResult = Zero;
741 // If the use is inside the loop, use its base, regardless of what it is:
742 // it is clearly shared across all the IV's. If the use is outside the loop
743 // (which means after it) we don't want to factor anything *into* the loop,
744 // so just use 0 as the base.
745 if (L->contains(Uses[0].Inst->getParent()))
746 std::swap(Result, Uses[0].Base);
750 // To find common subexpressions, count how many of Uses use each expression.
751 // If any subexpressions are used Uses.size() times, they are common.
752 // Also track whether all uses of each expression can be moved into an
753 // an addressing mode "for free"; such expressions are left within the loop.
754 // struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; };
755 std::map<const SCEV *, SubExprUseData> SubExpressionUseData;
757 // UniqueSubExprs - Keep track of all of the subexpressions we see in the
758 // order we see them.
759 SmallVector<const SCEV *, 16> UniqueSubExprs;
761 SmallVector<const SCEV *, 16> SubExprs;
762 unsigned NumUsesInsideLoop = 0;
763 for (unsigned i = 0; i != NumUses; ++i) {
764 // If the user is outside the loop, just ignore it for base computation.
765 // Since the user is outside the loop, it must be *after* the loop (if it
766 // were before, it could not be based on the loop IV). We don't want users
767 // after the loop to affect base computation of values *inside* the loop,
768 // because we can always add their offsets to the result IV after the loop
769 // is done, ensuring we get good code inside the loop.
770 if (!L->contains(Uses[i].Inst->getParent()))
774 // If the base is zero (which is common), return zero now, there are no
776 if (Uses[i].Base == Zero) return Zero;
778 // If this use is as an address we may be able to put CSEs in the addressing
779 // mode rather than hoisting them.
780 bool isAddrUse = isAddressUse(Uses[i].Inst, Uses[i].OperandValToReplace);
781 // We may need the AccessTy below, but only when isAddrUse, so compute it
782 // only in that case.
783 const Type *AccessTy = 0;
785 AccessTy = getAccessType(Uses[i].Inst);
787 // Split the expression into subexprs.
788 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
789 // Add one to SubExpressionUseData.Count for each subexpr present, and
790 // if the subexpr is not a valid immediate within an addressing mode use,
791 // set SubExpressionUseData.notAllUsesAreFree. We definitely want to
792 // hoist these out of the loop (if they are common to all uses).
793 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) {
794 if (++SubExpressionUseData[SubExprs[j]].Count == 1)
795 UniqueSubExprs.push_back(SubExprs[j]);
796 if (!isAddrUse || !fitsInAddressMode(SubExprs[j], AccessTy, TLI, false))
797 SubExpressionUseData[SubExprs[j]].notAllUsesAreFree = true;
802 // Now that we know how many times each is used, build Result. Iterate over
803 // UniqueSubexprs so that we have a stable ordering.
804 for (unsigned i = 0, e = UniqueSubExprs.size(); i != e; ++i) {
805 std::map<const SCEV *, SubExprUseData>::iterator I =
806 SubExpressionUseData.find(UniqueSubExprs[i]);
807 assert(I != SubExpressionUseData.end() && "Entry not found?");
808 if (I->second.Count == NumUsesInsideLoop) { // Found CSE!
809 if (I->second.notAllUsesAreFree)
810 Result = SE->getAddExpr(Result, I->first);
812 FreeResult = SE->getAddExpr(FreeResult, I->first);
814 // Remove non-cse's from SubExpressionUseData.
815 SubExpressionUseData.erase(I);
818 if (FreeResult != Zero) {
819 // We have some subexpressions that can be subsumed into addressing
820 // modes in every use inside the loop. However, it's possible that
821 // there are so many of them that the combined FreeResult cannot
822 // be subsumed, or that the target cannot handle both a FreeResult
823 // and a Result in the same instruction (for example because it would
824 // require too many registers). Check this.
825 for (unsigned i=0; i<NumUses; ++i) {
826 if (!L->contains(Uses[i].Inst->getParent()))
828 // We know this is an addressing mode use; if there are any uses that
829 // are not, FreeResult would be Zero.
830 const Type *AccessTy = getAccessType(Uses[i].Inst);
831 if (!fitsInAddressMode(FreeResult, AccessTy, TLI, Result!=Zero)) {
832 // FIXME: could split up FreeResult into pieces here, some hoisted
833 // and some not. There is no obvious advantage to this.
834 Result = SE->getAddExpr(Result, FreeResult);
841 // If we found no CSE's, return now.
842 if (Result == Zero) return Result;
844 // If we still have a FreeResult, remove its subexpressions from
845 // SubExpressionUseData. This means they will remain in the use Bases.
846 if (FreeResult != Zero) {
847 SeparateSubExprs(SubExprs, FreeResult, SE);
848 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) {
849 std::map<const SCEV *, SubExprUseData>::iterator I =
850 SubExpressionUseData.find(SubExprs[j]);
851 SubExpressionUseData.erase(I);
856 // Otherwise, remove all of the CSE's we found from each of the base values.
857 for (unsigned i = 0; i != NumUses; ++i) {
858 // Uses outside the loop don't necessarily include the common base, but
859 // the final IV value coming into those uses does. Instead of trying to
860 // remove the pieces of the common base, which might not be there,
861 // subtract off the base to compensate for this.
862 if (!L->contains(Uses[i].Inst->getParent())) {
863 Uses[i].Base = SE->getMinusSCEV(Uses[i].Base, Result);
867 // Split the expression into subexprs.
868 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
870 // Remove any common subexpressions.
871 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
872 if (SubExpressionUseData.count(SubExprs[j])) {
873 SubExprs.erase(SubExprs.begin()+j);
877 // Finally, add the non-shared expressions together.
878 if (SubExprs.empty())
881 Uses[i].Base = SE->getAddExpr(SubExprs);
888 /// ValidScale - Check whether the given Scale is valid for all loads and
889 /// stores in UsersToProcess.
891 bool LoopStrengthReduce::ValidScale(bool HasBaseReg, int64_t Scale,
892 const std::vector<BasedUser>& UsersToProcess) {
896 for (unsigned i = 0, e = UsersToProcess.size(); i!=e; ++i) {
897 // If this is a load or other access, pass the type of the access in.
898 const Type *AccessTy =
899 Type::getVoidTy(UsersToProcess[i].Inst->getContext());
900 if (isAddressUse(UsersToProcess[i].Inst,
901 UsersToProcess[i].OperandValToReplace))
902 AccessTy = getAccessType(UsersToProcess[i].Inst);
903 else if (isa<PHINode>(UsersToProcess[i].Inst))
906 TargetLowering::AddrMode AM;
907 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
908 AM.BaseOffs = SC->getValue()->getSExtValue();
909 AM.HasBaseReg = HasBaseReg || !UsersToProcess[i].Base->isZero();
912 // If load[imm+r*scale] is illegal, bail out.
913 if (!TLI->isLegalAddressingMode(AM, AccessTy))
919 /// ValidOffset - Check whether the given Offset is valid for all loads and
920 /// stores in UsersToProcess.
922 bool LoopStrengthReduce::ValidOffset(bool HasBaseReg,
925 const std::vector<BasedUser>& UsersToProcess) {
929 for (unsigned i=0, e = UsersToProcess.size(); i!=e; ++i) {
930 // If this is a load or other access, pass the type of the access in.
931 const Type *AccessTy =
932 Type::getVoidTy(UsersToProcess[i].Inst->getContext());
933 if (isAddressUse(UsersToProcess[i].Inst,
934 UsersToProcess[i].OperandValToReplace))
935 AccessTy = getAccessType(UsersToProcess[i].Inst);
936 else if (isa<PHINode>(UsersToProcess[i].Inst))
939 TargetLowering::AddrMode AM;
940 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
941 AM.BaseOffs = SC->getValue()->getSExtValue();
942 AM.BaseOffs = (uint64_t)AM.BaseOffs + (uint64_t)Offset;
943 AM.HasBaseReg = HasBaseReg || !UsersToProcess[i].Base->isZero();
946 // If load[imm+r*scale] is illegal, bail out.
947 if (!TLI->isLegalAddressingMode(AM, AccessTy))
953 /// RequiresTypeConversion - Returns true if converting Ty1 to Ty2 is not
955 bool LoopStrengthReduce::RequiresTypeConversion(const Type *Ty1,
959 Ty1 = SE->getEffectiveSCEVType(Ty1);
960 Ty2 = SE->getEffectiveSCEVType(Ty2);
963 if (Ty1->canLosslesslyBitCastTo(Ty2))
965 if (TLI && TLI->isTruncateFree(Ty1, Ty2))
970 /// CheckForIVReuse - Returns the multiple if the stride is the multiple
971 /// of a previous stride and it is a legal value for the target addressing
972 /// mode scale component and optional base reg. This allows the users of
973 /// this stride to be rewritten as prev iv * factor. It returns 0 if no
974 /// reuse is possible. Factors can be negative on same targets, e.g. ARM.
976 /// If all uses are outside the loop, we don't require that all multiplies
977 /// be folded into the addressing mode, nor even that the factor be constant;
978 /// a multiply (executed once) outside the loop is better than another IV
979 /// within. Well, usually.
980 const SCEV *LoopStrengthReduce::CheckForIVReuse(bool HasBaseReg,
981 bool AllUsesAreAddresses,
982 bool AllUsesAreOutsideLoop,
983 const SCEV *const &Stride,
984 IVExpr &IV, const Type *Ty,
985 const std::vector<BasedUser>& UsersToProcess) {
986 if (StrideNoReuse.count(Stride))
987 return SE->getIntegerSCEV(0, Stride->getType());
989 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Stride)) {
990 int64_t SInt = SC->getValue()->getSExtValue();
991 for (unsigned NewStride = 0, e = IU->StrideOrder.size();
992 NewStride != e; ++NewStride) {
993 std::map<const SCEV *, IVsOfOneStride>::iterator SI =
994 IVsByStride.find(IU->StrideOrder[NewStride]);
995 if (SI == IVsByStride.end() || !isa<SCEVConstant>(SI->first) ||
996 StrideNoReuse.count(SI->first))
998 // The other stride has no uses, don't reuse it.
999 std::map<const SCEV *, IVUsersOfOneStride *>::iterator UI =
1000 IU->IVUsesByStride.find(IU->StrideOrder[NewStride]);
1001 if (UI->second->Users.empty())
1003 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1004 if (SI->first != Stride &&
1005 (unsigned(abs64(SInt)) < SSInt || (SInt % SSInt) != 0))
1007 int64_t Scale = SInt / SSInt;
1008 // Check that this stride is valid for all the types used for loads and
1009 // stores; if it can be used for some and not others, we might as well use
1010 // the original stride everywhere, since we have to create the IV for it
1011 // anyway. If the scale is 1, then we don't need to worry about folding
1014 (AllUsesAreAddresses &&
1015 ValidScale(HasBaseReg, Scale, UsersToProcess))) {
1016 // Prefer to reuse an IV with a base of zero.
1017 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1018 IE = SI->second.IVs.end(); II != IE; ++II)
1019 // Only reuse previous IV if it would not require a type conversion
1020 // and if the base difference can be folded.
1021 if (II->Base->isZero() &&
1022 !RequiresTypeConversion(II->Base->getType(), Ty)) {
1024 return SE->getIntegerSCEV(Scale, Stride->getType());
1026 // Otherwise, settle for an IV with a foldable base.
1027 if (AllUsesAreAddresses)
1028 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1029 IE = SI->second.IVs.end(); II != IE; ++II)
1030 // Only reuse previous IV if it would not require a type conversion
1031 // and if the base difference can be folded.
1032 if (SE->getEffectiveSCEVType(II->Base->getType()) ==
1033 SE->getEffectiveSCEVType(Ty) &&
1034 isa<SCEVConstant>(II->Base)) {
1036 cast<SCEVConstant>(II->Base)->getValue()->getSExtValue();
1037 if (Base > INT32_MIN && Base <= INT32_MAX &&
1038 ValidOffset(HasBaseReg, -Base * Scale,
1039 Scale, UsersToProcess)) {
1041 return SE->getIntegerSCEV(Scale, Stride->getType());
1046 } else if (AllUsesAreOutsideLoop) {
1047 // Accept nonconstant strides here; it is really really right to substitute
1048 // an existing IV if we can.
1049 for (unsigned NewStride = 0, e = IU->StrideOrder.size();
1050 NewStride != e; ++NewStride) {
1051 std::map<const SCEV *, IVsOfOneStride>::iterator SI =
1052 IVsByStride.find(IU->StrideOrder[NewStride]);
1053 if (SI == IVsByStride.end() || !isa<SCEVConstant>(SI->first))
1055 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1056 if (SI->first != Stride && SSInt != 1)
1058 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1059 IE = SI->second.IVs.end(); II != IE; ++II)
1060 // Accept nonzero base here.
1061 // Only reuse previous IV if it would not require a type conversion.
1062 if (!RequiresTypeConversion(II->Base->getType(), Ty)) {
1067 // Special case, old IV is -1*x and this one is x. Can treat this one as
1069 for (unsigned NewStride = 0, e = IU->StrideOrder.size();
1070 NewStride != e; ++NewStride) {
1071 std::map<const SCEV *, IVsOfOneStride>::iterator SI =
1072 IVsByStride.find(IU->StrideOrder[NewStride]);
1073 if (SI == IVsByStride.end())
1075 if (const SCEVMulExpr *ME = dyn_cast<SCEVMulExpr>(SI->first))
1076 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(ME->getOperand(0)))
1077 if (Stride == ME->getOperand(1) &&
1078 SC->getValue()->getSExtValue() == -1LL)
1079 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1080 IE = SI->second.IVs.end(); II != IE; ++II)
1081 // Accept nonzero base here.
1082 // Only reuse previous IV if it would not require type conversion.
1083 if (!RequiresTypeConversion(II->Base->getType(), Ty)) {
1085 return SE->getIntegerSCEV(-1LL, Stride->getType());
1089 return SE->getIntegerSCEV(0, Stride->getType());
1092 /// PartitionByIsUseOfPostIncrementedValue - Simple boolean predicate that
1093 /// returns true if Val's isUseOfPostIncrementedValue is true.
1094 static bool PartitionByIsUseOfPostIncrementedValue(const BasedUser &Val) {
1095 return Val.isUseOfPostIncrementedValue;
1098 /// isNonConstantNegative - Return true if the specified scev is negated, but
1100 static bool isNonConstantNegative(const SCEV *const &Expr) {
1101 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Expr);
1102 if (!Mul) return false;
1104 // If there is a constant factor, it will be first.
1105 const SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
1106 if (!SC) return false;
1108 // Return true if the value is negative, this matches things like (-42 * V).
1109 return SC->getValue()->getValue().isNegative();
1112 /// CollectIVUsers - Transform our list of users and offsets to a bit more
1113 /// complex table. In this new vector, each 'BasedUser' contains 'Base', the
1114 /// base of the strided accesses, as well as the old information from Uses. We
1115 /// progressively move information from the Base field to the Imm field, until
1116 /// we eventually have the full access expression to rewrite the use.
1117 const SCEV *LoopStrengthReduce::CollectIVUsers(const SCEV *const &Stride,
1118 IVUsersOfOneStride &Uses,
1120 bool &AllUsesAreAddresses,
1121 bool &AllUsesAreOutsideLoop,
1122 std::vector<BasedUser> &UsersToProcess) {
1123 // FIXME: Generalize to non-affine IV's.
1124 if (!Stride->isLoopInvariant(L))
1125 return SE->getIntegerSCEV(0, Stride->getType());
1127 UsersToProcess.reserve(Uses.Users.size());
1128 for (ilist<IVStrideUse>::iterator I = Uses.Users.begin(),
1129 E = Uses.Users.end(); I != E; ++I) {
1130 UsersToProcess.push_back(BasedUser(*I, SE));
1132 // Move any loop variant operands from the offset field to the immediate
1133 // field of the use, so that we don't try to use something before it is
1135 MoveLoopVariantsToImmediateField(UsersToProcess.back().Base,
1136 UsersToProcess.back().Imm, L, SE);
1137 assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
1138 "Base value is not loop invariant!");
1141 // We now have a whole bunch of uses of like-strided induction variables, but
1142 // they might all have different bases. We want to emit one PHI node for this
1143 // stride which we fold as many common expressions (between the IVs) into as
1144 // possible. Start by identifying the common expressions in the base values
1145 // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
1146 // "A+B"), emit it to the preheader, then remove the expression from the
1147 // UsersToProcess base values.
1148 const SCEV *CommonExprs =
1149 RemoveCommonExpressionsFromUseBases(UsersToProcess, SE, L, TLI);
1151 // Next, figure out what we can represent in the immediate fields of
1152 // instructions. If we can represent anything there, move it to the imm
1153 // fields of the BasedUsers. We do this so that it increases the commonality
1154 // of the remaining uses.
1155 unsigned NumPHI = 0;
1156 bool HasAddress = false;
1157 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1158 // If the user is not in the current loop, this means it is using the exit
1159 // value of the IV. Do not put anything in the base, make sure it's all in
1160 // the immediate field to allow as much factoring as possible.
1161 if (!L->contains(UsersToProcess[i].Inst->getParent())) {
1162 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm,
1163 UsersToProcess[i].Base);
1164 UsersToProcess[i].Base =
1165 SE->getIntegerSCEV(0, UsersToProcess[i].Base->getType());
1167 // Not all uses are outside the loop.
1168 AllUsesAreOutsideLoop = false;
1170 // Addressing modes can be folded into loads and stores. Be careful that
1171 // the store is through the expression, not of the expression though.
1173 bool isAddress = isAddressUse(UsersToProcess[i].Inst,
1174 UsersToProcess[i].OperandValToReplace);
1175 if (isa<PHINode>(UsersToProcess[i].Inst)) {
1183 // If this use isn't an address, then not all uses are addresses.
1184 if (!isAddress && !isPHI)
1185 AllUsesAreAddresses = false;
1187 MoveImmediateValues(TLI, UsersToProcess[i].Inst, UsersToProcess[i].Base,
1188 UsersToProcess[i].Imm, isAddress, L, SE);
1192 // If one of the use is a PHI node and all other uses are addresses, still
1193 // allow iv reuse. Essentially we are trading one constant multiplication
1194 // for one fewer iv.
1196 AllUsesAreAddresses = false;
1198 // There are no in-loop address uses.
1199 if (AllUsesAreAddresses && (!HasAddress && !AllUsesAreOutsideLoop))
1200 AllUsesAreAddresses = false;
1205 /// ShouldUseFullStrengthReductionMode - Test whether full strength-reduction
1206 /// is valid and profitable for the given set of users of a stride. In
1207 /// full strength-reduction mode, all addresses at the current stride are
1208 /// strength-reduced all the way down to pointer arithmetic.
1210 bool LoopStrengthReduce::ShouldUseFullStrengthReductionMode(
1211 const std::vector<BasedUser> &UsersToProcess,
1213 bool AllUsesAreAddresses,
1214 const SCEV *Stride) {
1215 if (!EnableFullLSRMode)
1218 // The heuristics below aim to avoid increasing register pressure, but
1219 // fully strength-reducing all the addresses increases the number of
1220 // add instructions, so don't do this when optimizing for size.
1221 // TODO: If the loop is large, the savings due to simpler addresses
1222 // may oughtweight the costs of the extra increment instructions.
1223 if (L->getHeader()->getParent()->hasFnAttr(Attribute::OptimizeForSize))
1226 // TODO: For now, don't do full strength reduction if there could
1227 // potentially be greater-stride multiples of the current stride
1228 // which could reuse the current stride IV.
1229 if (IU->StrideOrder.back() != Stride)
1232 // Iterate through the uses to find conditions that automatically rule out
1234 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ) {
1235 const SCEV *Base = UsersToProcess[i].Base;
1236 const SCEV *Imm = UsersToProcess[i].Imm;
1237 // If any users have a loop-variant component, they can't be fully
1238 // strength-reduced.
1239 if (Imm && !Imm->isLoopInvariant(L))
1241 // If there are to users with the same base and the difference between
1242 // the two Imm values can't be folded into the address, full
1243 // strength reduction would increase register pressure.
1245 const SCEV *CurImm = UsersToProcess[i].Imm;
1246 if ((CurImm || Imm) && CurImm != Imm) {
1247 if (!CurImm) CurImm = SE->getIntegerSCEV(0, Stride->getType());
1248 if (!Imm) Imm = SE->getIntegerSCEV(0, Stride->getType());
1249 const Instruction *Inst = UsersToProcess[i].Inst;
1250 const Type *AccessTy = getAccessType(Inst);
1251 const SCEV *Diff = SE->getMinusSCEV(UsersToProcess[i].Imm, Imm);
1252 if (!Diff->isZero() &&
1253 (!AllUsesAreAddresses ||
1254 !fitsInAddressMode(Diff, AccessTy, TLI, /*HasBaseReg=*/true)))
1257 } while (++i != e && Base == UsersToProcess[i].Base);
1260 // If there's exactly one user in this stride, fully strength-reducing it
1261 // won't increase register pressure. If it's starting from a non-zero base,
1262 // it'll be simpler this way.
1263 if (UsersToProcess.size() == 1 && !UsersToProcess[0].Base->isZero())
1266 // Otherwise, if there are any users in this stride that don't require
1267 // a register for their base, full strength-reduction will increase
1268 // register pressure.
1269 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1270 if (UsersToProcess[i].Base->isZero())
1273 // Otherwise, go for it.
1277 /// InsertAffinePhi Create and insert a PHI node for an induction variable
1278 /// with the specified start and step values in the specified loop.
1280 /// If NegateStride is true, the stride should be negated by using a
1281 /// subtract instead of an add.
1283 /// Return the created phi node.
1285 static PHINode *InsertAffinePhi(const SCEV *Start, const SCEV *Step,
1286 Instruction *IVIncInsertPt,
1288 SCEVExpander &Rewriter) {
1289 assert(Start->isLoopInvariant(L) && "New PHI start is not loop invariant!");
1290 assert(Step->isLoopInvariant(L) && "New PHI stride is not loop invariant!");
1292 BasicBlock *Header = L->getHeader();
1293 BasicBlock *Preheader = L->getLoopPreheader();
1294 BasicBlock *LatchBlock = L->getLoopLatch();
1295 const Type *Ty = Start->getType();
1296 Ty = Rewriter.SE.getEffectiveSCEVType(Ty);
1298 PHINode *PN = PHINode::Create(Ty, "lsr.iv", Header->begin());
1299 PN->addIncoming(Rewriter.expandCodeFor(Start, Ty, Preheader->getTerminator()),
1302 // If the stride is negative, insert a sub instead of an add for the
1304 bool isNegative = isNonConstantNegative(Step);
1305 const SCEV *IncAmount = Step;
1307 IncAmount = Rewriter.SE.getNegativeSCEV(Step);
1309 // Insert an add instruction right before the terminator corresponding
1310 // to the back-edge or just before the only use. The location is determined
1311 // by the caller and passed in as IVIncInsertPt.
1312 Value *StepV = Rewriter.expandCodeFor(IncAmount, Ty,
1313 Preheader->getTerminator());
1316 IncV = BinaryOperator::CreateSub(PN, StepV, "lsr.iv.next",
1319 IncV = BinaryOperator::CreateAdd(PN, StepV, "lsr.iv.next",
1322 if (!isa<ConstantInt>(StepV)) ++NumVariable;
1324 PN->addIncoming(IncV, LatchBlock);
1330 static void SortUsersToProcess(std::vector<BasedUser> &UsersToProcess) {
1331 // We want to emit code for users inside the loop first. To do this, we
1332 // rearrange BasedUser so that the entries at the end have
1333 // isUseOfPostIncrementedValue = false, because we pop off the end of the
1334 // vector (so we handle them first).
1335 std::partition(UsersToProcess.begin(), UsersToProcess.end(),
1336 PartitionByIsUseOfPostIncrementedValue);
1338 // Sort this by base, so that things with the same base are handled
1339 // together. By partitioning first and stable-sorting later, we are
1340 // guaranteed that within each base we will pop off users from within the
1341 // loop before users outside of the loop with a particular base.
1343 // We would like to use stable_sort here, but we can't. The problem is that
1344 // const SCEV *'s don't have a deterministic ordering w.r.t to each other, so
1345 // we don't have anything to do a '<' comparison on. Because we think the
1346 // number of uses is small, do a horrible bubble sort which just relies on
1348 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1349 // Get a base value.
1350 const SCEV *Base = UsersToProcess[i].Base;
1352 // Compact everything with this base to be consecutive with this one.
1353 for (unsigned j = i+1; j != e; ++j) {
1354 if (UsersToProcess[j].Base == Base) {
1355 std::swap(UsersToProcess[i+1], UsersToProcess[j]);
1362 /// PrepareToStrengthReduceFully - Prepare to fully strength-reduce
1363 /// UsersToProcess, meaning lowering addresses all the way down to direct
1364 /// pointer arithmetic.
1367 LoopStrengthReduce::PrepareToStrengthReduceFully(
1368 std::vector<BasedUser> &UsersToProcess,
1370 const SCEV *CommonExprs,
1372 SCEVExpander &PreheaderRewriter) {
1373 DEBUG(errs() << " Fully reducing all users\n");
1375 // Rewrite the UsersToProcess records, creating a separate PHI for each
1376 // unique Base value.
1377 Instruction *IVIncInsertPt = L->getLoopLatch()->getTerminator();
1378 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ) {
1379 // TODO: The uses are grouped by base, but not sorted. We arbitrarily
1380 // pick the first Imm value here to start with, and adjust it for the
1382 const SCEV *Imm = UsersToProcess[i].Imm;
1383 const SCEV *Base = UsersToProcess[i].Base;
1384 const SCEV *Start = SE->getAddExpr(CommonExprs, Base, Imm);
1385 PHINode *Phi = InsertAffinePhi(Start, Stride, IVIncInsertPt, L,
1387 // Loop over all the users with the same base.
1389 UsersToProcess[i].Base = SE->getIntegerSCEV(0, Stride->getType());
1390 UsersToProcess[i].Imm = SE->getMinusSCEV(UsersToProcess[i].Imm, Imm);
1391 UsersToProcess[i].Phi = Phi;
1392 assert(UsersToProcess[i].Imm->isLoopInvariant(L) &&
1393 "ShouldUseFullStrengthReductionMode should reject this!");
1394 } while (++i != e && Base == UsersToProcess[i].Base);
1398 /// FindIVIncInsertPt - Return the location to insert the increment instruction.
1399 /// If the only use if a use of postinc value, (must be the loop termination
1400 /// condition), then insert it just before the use.
1401 static Instruction *FindIVIncInsertPt(std::vector<BasedUser> &UsersToProcess,
1403 if (UsersToProcess.size() == 1 &&
1404 UsersToProcess[0].isUseOfPostIncrementedValue &&
1405 L->contains(UsersToProcess[0].Inst->getParent()))
1406 return UsersToProcess[0].Inst;
1407 return L->getLoopLatch()->getTerminator();
1410 /// PrepareToStrengthReduceWithNewPhi - Insert a new induction variable for the
1411 /// given users to share.
1414 LoopStrengthReduce::PrepareToStrengthReduceWithNewPhi(
1415 std::vector<BasedUser> &UsersToProcess,
1417 const SCEV *CommonExprs,
1419 Instruction *IVIncInsertPt,
1421 SCEVExpander &PreheaderRewriter) {
1422 DEBUG(errs() << " Inserting new PHI:\n");
1424 PHINode *Phi = InsertAffinePhi(SE->getUnknown(CommonBaseV),
1425 Stride, IVIncInsertPt, L,
1428 // Remember this in case a later stride is multiple of this.
1429 IVsByStride[Stride].addIV(Stride, CommonExprs, Phi);
1431 // All the users will share this new IV.
1432 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1433 UsersToProcess[i].Phi = Phi;
1435 DEBUG(errs() << " IV=");
1436 DEBUG(WriteAsOperand(errs(), Phi, /*PrintType=*/false));
1437 DEBUG(errs() << "\n");
1440 /// PrepareToStrengthReduceFromSmallerStride - Prepare for the given users to
1441 /// reuse an induction variable with a stride that is a factor of the current
1442 /// induction variable.
1445 LoopStrengthReduce::PrepareToStrengthReduceFromSmallerStride(
1446 std::vector<BasedUser> &UsersToProcess,
1448 const IVExpr &ReuseIV,
1449 Instruction *PreInsertPt) {
1450 DEBUG(errs() << " Rewriting in terms of existing IV of STRIDE "
1451 << *ReuseIV.Stride << " and BASE " << *ReuseIV.Base << "\n");
1453 // All the users will share the reused IV.
1454 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1455 UsersToProcess[i].Phi = ReuseIV.PHI;
1457 Constant *C = dyn_cast<Constant>(CommonBaseV);
1459 (!C->isNullValue() &&
1460 !fitsInAddressMode(SE->getUnknown(CommonBaseV), CommonBaseV->getType(),
1462 // We want the common base emitted into the preheader! This is just
1463 // using cast as a copy so BitCast (no-op cast) is appropriate
1464 CommonBaseV = new BitCastInst(CommonBaseV, CommonBaseV->getType(),
1465 "commonbase", PreInsertPt);
1468 static bool IsImmFoldedIntoAddrMode(GlobalValue *GV, int64_t Offset,
1469 const Type *AccessTy,
1470 std::vector<BasedUser> &UsersToProcess,
1471 const TargetLowering *TLI) {
1472 SmallVector<Instruction*, 16> AddrModeInsts;
1473 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1474 if (UsersToProcess[i].isUseOfPostIncrementedValue)
1476 ExtAddrMode AddrMode =
1477 AddressingModeMatcher::Match(UsersToProcess[i].OperandValToReplace,
1478 AccessTy, UsersToProcess[i].Inst,
1479 AddrModeInsts, *TLI);
1480 if (GV && GV != AddrMode.BaseGV)
1482 if (Offset && !AddrMode.BaseOffs)
1483 // FIXME: How to accurate check it's immediate offset is folded.
1485 AddrModeInsts.clear();
1490 /// StrengthReduceIVUsersOfStride - Strength reduce all of the users of a single
1491 /// stride of IV. All of the users may have different starting values, and this
1492 /// may not be the only stride.
1494 LoopStrengthReduce::StrengthReduceIVUsersOfStride(const SCEV *const &Stride,
1495 IVUsersOfOneStride &Uses,
1497 // If all the users are moved to another stride, then there is nothing to do.
1498 if (Uses.Users.empty())
1501 // Keep track if every use in UsersToProcess is an address. If they all are,
1502 // we may be able to rewrite the entire collection of them in terms of a
1503 // smaller-stride IV.
1504 bool AllUsesAreAddresses = true;
1506 // Keep track if every use of a single stride is outside the loop. If so,
1507 // we want to be more aggressive about reusing a smaller-stride IV; a
1508 // multiply outside the loop is better than another IV inside. Well, usually.
1509 bool AllUsesAreOutsideLoop = true;
1511 // Transform our list of users and offsets to a bit more complex table. In
1512 // this new vector, each 'BasedUser' contains 'Base' the base of the
1513 // strided accessas well as the old information from Uses. We progressively
1514 // move information from the Base field to the Imm field, until we eventually
1515 // have the full access expression to rewrite the use.
1516 std::vector<BasedUser> UsersToProcess;
1517 const SCEV *CommonExprs = CollectIVUsers(Stride, Uses, L, AllUsesAreAddresses,
1518 AllUsesAreOutsideLoop,
1521 // Sort the UsersToProcess array so that users with common bases are
1522 // next to each other.
1523 SortUsersToProcess(UsersToProcess);
1525 // If we managed to find some expressions in common, we'll need to carry
1526 // their value in a register and add it in for each use. This will take up
1527 // a register operand, which potentially restricts what stride values are
1529 bool HaveCommonExprs = !CommonExprs->isZero();
1530 const Type *ReplacedTy = CommonExprs->getType();
1532 // If all uses are addresses, consider sinking the immediate part of the
1533 // common expression back into uses if they can fit in the immediate fields.
1534 if (TLI && HaveCommonExprs && AllUsesAreAddresses) {
1535 const SCEV *NewCommon = CommonExprs;
1536 const SCEV *Imm = SE->getIntegerSCEV(0, ReplacedTy);
1537 MoveImmediateValues(TLI, Type::getVoidTy(
1538 L->getLoopPreheader()->getContext()),
1539 NewCommon, Imm, true, L, SE);
1540 if (!Imm->isZero()) {
1543 // If the immediate part of the common expression is a GV, check if it's
1544 // possible to fold it into the target addressing mode.
1545 GlobalValue *GV = 0;
1546 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(Imm))
1547 GV = dyn_cast<GlobalValue>(SU->getValue());
1549 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Imm))
1550 Offset = SC->getValue()->getSExtValue();
1552 // Pass VoidTy as the AccessTy to be conservative, because
1553 // there could be multiple access types among all the uses.
1554 DoSink = IsImmFoldedIntoAddrMode(GV, Offset,
1555 Type::getVoidTy(L->getLoopPreheader()->getContext()),
1556 UsersToProcess, TLI);
1559 DEBUG(errs() << " Sinking " << *Imm << " back down into uses\n");
1560 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1561 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm, Imm);
1562 CommonExprs = NewCommon;
1563 HaveCommonExprs = !CommonExprs->isZero();
1569 // Now that we know what we need to do, insert the PHI node itself.
1571 DEBUG(errs() << "LSR: Examining IVs of TYPE " << *ReplacedTy << " of STRIDE "
1573 << " Common base: " << *CommonExprs << "\n");
1575 SCEVExpander Rewriter(*SE);
1576 SCEVExpander PreheaderRewriter(*SE);
1578 BasicBlock *Preheader = L->getLoopPreheader();
1579 Instruction *PreInsertPt = Preheader->getTerminator();
1580 BasicBlock *LatchBlock = L->getLoopLatch();
1581 Instruction *IVIncInsertPt = LatchBlock->getTerminator();
1583 Value *CommonBaseV = Constant::getNullValue(ReplacedTy);
1585 const SCEV *RewriteFactor = SE->getIntegerSCEV(0, ReplacedTy);
1586 IVExpr ReuseIV(SE->getIntegerSCEV(0,
1587 Type::getInt32Ty(Preheader->getContext())),
1588 SE->getIntegerSCEV(0,
1589 Type::getInt32Ty(Preheader->getContext())),
1592 // Choose a strength-reduction strategy and prepare for it by creating
1593 // the necessary PHIs and adjusting the bookkeeping.
1594 if (ShouldUseFullStrengthReductionMode(UsersToProcess, L,
1595 AllUsesAreAddresses, Stride)) {
1596 PrepareToStrengthReduceFully(UsersToProcess, Stride, CommonExprs, L,
1599 // Emit the initial base value into the loop preheader.
1600 CommonBaseV = PreheaderRewriter.expandCodeFor(CommonExprs, ReplacedTy,
1603 // If all uses are addresses, check if it is possible to reuse an IV. The
1604 // new IV must have a stride that is a multiple of the old stride; the
1605 // multiple must be a number that can be encoded in the scale field of the
1606 // target addressing mode; and we must have a valid instruction after this
1607 // substitution, including the immediate field, if any.
1608 RewriteFactor = CheckForIVReuse(HaveCommonExprs, AllUsesAreAddresses,
1609 AllUsesAreOutsideLoop,
1610 Stride, ReuseIV, ReplacedTy,
1612 if (!RewriteFactor->isZero())
1613 PrepareToStrengthReduceFromSmallerStride(UsersToProcess, CommonBaseV,
1614 ReuseIV, PreInsertPt);
1616 IVIncInsertPt = FindIVIncInsertPt(UsersToProcess, L);
1617 PrepareToStrengthReduceWithNewPhi(UsersToProcess, Stride, CommonExprs,
1618 CommonBaseV, IVIncInsertPt,
1619 L, PreheaderRewriter);
1623 // Process all the users now, replacing their strided uses with
1624 // strength-reduced forms. This outer loop handles all bases, the inner
1625 // loop handles all users of a particular base.
1626 while (!UsersToProcess.empty()) {
1627 const SCEV *Base = UsersToProcess.back().Base;
1628 Instruction *Inst = UsersToProcess.back().Inst;
1630 // Emit the code for Base into the preheader.
1632 if (!Base->isZero()) {
1633 BaseV = PreheaderRewriter.expandCodeFor(Base, 0, PreInsertPt);
1635 DEBUG(errs() << " INSERTING code for BASE = " << *Base << ":");
1636 if (BaseV->hasName())
1637 DEBUG(errs() << " Result value name = %" << BaseV->getName());
1638 DEBUG(errs() << "\n");
1640 // If BaseV is a non-zero constant, make sure that it gets inserted into
1641 // the preheader, instead of being forward substituted into the uses. We
1642 // do this by forcing a BitCast (noop cast) to be inserted into the
1643 // preheader in this case.
1644 if (!fitsInAddressMode(Base, getAccessType(Inst), TLI, false) &&
1645 isa<Constant>(BaseV)) {
1646 // We want this constant emitted into the preheader! This is just
1647 // using cast as a copy so BitCast (no-op cast) is appropriate
1648 BaseV = new BitCastInst(BaseV, BaseV->getType(), "preheaderinsert",
1653 // Emit the code to add the immediate offset to the Phi value, just before
1654 // the instructions that we identified as using this stride and base.
1656 // FIXME: Use emitted users to emit other users.
1657 BasedUser &User = UsersToProcess.back();
1659 DEBUG(errs() << " Examining ");
1660 if (User.isUseOfPostIncrementedValue)
1661 DEBUG(errs() << "postinc");
1663 DEBUG(errs() << "preinc");
1664 DEBUG(errs() << " use ");
1665 DEBUG(WriteAsOperand(errs(), UsersToProcess.back().OperandValToReplace,
1666 /*PrintType=*/false));
1667 DEBUG(errs() << " in Inst: " << *User.Inst);
1669 // If this instruction wants to use the post-incremented value, move it
1670 // after the post-inc and use its value instead of the PHI.
1671 Value *RewriteOp = User.Phi;
1672 if (User.isUseOfPostIncrementedValue) {
1673 RewriteOp = User.Phi->getIncomingValueForBlock(LatchBlock);
1674 // If this user is in the loop, make sure it is the last thing in the
1675 // loop to ensure it is dominated by the increment. In case it's the
1676 // only use of the iv, the increment instruction is already before the
1678 if (L->contains(User.Inst->getParent()) && User.Inst != IVIncInsertPt)
1679 User.Inst->moveBefore(IVIncInsertPt);
1682 const SCEV *RewriteExpr = SE->getUnknown(RewriteOp);
1684 if (SE->getEffectiveSCEVType(RewriteOp->getType()) !=
1685 SE->getEffectiveSCEVType(ReplacedTy)) {
1686 assert(SE->getTypeSizeInBits(RewriteOp->getType()) >
1687 SE->getTypeSizeInBits(ReplacedTy) &&
1688 "Unexpected widening cast!");
1689 RewriteExpr = SE->getTruncateExpr(RewriteExpr, ReplacedTy);
1692 // If we had to insert new instructions for RewriteOp, we have to
1693 // consider that they may not have been able to end up immediately
1694 // next to RewriteOp, because non-PHI instructions may never precede
1695 // PHI instructions in a block. In this case, remember where the last
1696 // instruction was inserted so that if we're replacing a different
1697 // PHI node, we can use the later point to expand the final
1699 Instruction *NewBasePt = dyn_cast<Instruction>(RewriteOp);
1700 if (RewriteOp == User.Phi) NewBasePt = 0;
1702 // Clear the SCEVExpander's expression map so that we are guaranteed
1703 // to have the code emitted where we expect it.
1706 // If we are reusing the iv, then it must be multiplied by a constant
1707 // factor to take advantage of the addressing mode scale component.
1708 if (!RewriteFactor->isZero()) {
1709 // If we're reusing an IV with a nonzero base (currently this happens
1710 // only when all reuses are outside the loop) subtract that base here.
1711 // The base has been used to initialize the PHI node but we don't want
1713 if (!ReuseIV.Base->isZero()) {
1714 const SCEV *typedBase = ReuseIV.Base;
1715 if (SE->getEffectiveSCEVType(RewriteExpr->getType()) !=
1716 SE->getEffectiveSCEVType(ReuseIV.Base->getType())) {
1717 // It's possible the original IV is a larger type than the new IV,
1718 // in which case we have to truncate the Base. We checked in
1719 // RequiresTypeConversion that this is valid.
1720 assert(SE->getTypeSizeInBits(RewriteExpr->getType()) <
1721 SE->getTypeSizeInBits(ReuseIV.Base->getType()) &&
1722 "Unexpected lengthening conversion!");
1723 typedBase = SE->getTruncateExpr(ReuseIV.Base,
1724 RewriteExpr->getType());
1726 RewriteExpr = SE->getMinusSCEV(RewriteExpr, typedBase);
1729 // Multiply old variable, with base removed, by new scale factor.
1730 RewriteExpr = SE->getMulExpr(RewriteFactor,
1733 // The common base is emitted in the loop preheader. But since we
1734 // are reusing an IV, it has not been used to initialize the PHI node.
1735 // Add it to the expression used to rewrite the uses.
1736 // When this use is outside the loop, we earlier subtracted the
1737 // common base, and are adding it back here. Use the same expression
1738 // as before, rather than CommonBaseV, so DAGCombiner will zap it.
1739 if (!CommonExprs->isZero()) {
1740 if (L->contains(User.Inst->getParent()))
1741 RewriteExpr = SE->getAddExpr(RewriteExpr,
1742 SE->getUnknown(CommonBaseV));
1744 RewriteExpr = SE->getAddExpr(RewriteExpr, CommonExprs);
1748 // Now that we know what we need to do, insert code before User for the
1749 // immediate and any loop-variant expressions.
1751 // Add BaseV to the PHI value if needed.
1752 RewriteExpr = SE->getAddExpr(RewriteExpr, SE->getUnknown(BaseV));
1754 User.RewriteInstructionToUseNewBase(RewriteExpr, NewBasePt,
1758 // Mark old value we replaced as possibly dead, so that it is eliminated
1759 // if we just replaced the last use of that value.
1760 DeadInsts.push_back(User.OperandValToReplace);
1762 UsersToProcess.pop_back();
1765 // If there are any more users to process with the same base, process them
1766 // now. We sorted by base above, so we just have to check the last elt.
1767 } while (!UsersToProcess.empty() && UsersToProcess.back().Base == Base);
1768 // TODO: Next, find out which base index is the most common, pull it out.
1771 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
1772 // different starting values, into different PHIs.
1775 void LoopStrengthReduce::StrengthReduceIVUsers(Loop *L) {
1776 // Note: this processes each stride/type pair individually. All users
1777 // passed into StrengthReduceIVUsersOfStride have the same type AND stride.
1778 // Also, note that we iterate over IVUsesByStride indirectly by using
1779 // StrideOrder. This extra layer of indirection makes the ordering of
1780 // strides deterministic - not dependent on map order.
1781 for (unsigned Stride = 0, e = IU->StrideOrder.size(); Stride != e; ++Stride) {
1782 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
1783 IU->IVUsesByStride.find(IU->StrideOrder[Stride]);
1784 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
1785 // FIXME: Generalize to non-affine IV's.
1786 if (!SI->first->isLoopInvariant(L))
1788 StrengthReduceIVUsersOfStride(SI->first, *SI->second, L);
1792 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1793 /// set the IV user and stride information and return true, otherwise return
1795 bool LoopStrengthReduce::FindIVUserForCond(ICmpInst *Cond,
1796 IVStrideUse *&CondUse,
1797 const SCEV* &CondStride) {
1798 for (unsigned Stride = 0, e = IU->StrideOrder.size();
1799 Stride != e && !CondUse; ++Stride) {
1800 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
1801 IU->IVUsesByStride.find(IU->StrideOrder[Stride]);
1802 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
1804 for (ilist<IVStrideUse>::iterator UI = SI->second->Users.begin(),
1805 E = SI->second->Users.end(); UI != E; ++UI)
1806 if (UI->getUser() == Cond) {
1807 // NOTE: we could handle setcc instructions with multiple uses here, but
1808 // InstCombine does it as well for simple uses, it's not clear that it
1809 // occurs enough in real life to handle.
1811 CondStride = SI->first;
1819 // Constant strides come first which in turns are sorted by their absolute
1820 // values. If absolute values are the same, then positive strides comes first.
1822 // 4, -1, X, 1, 2 ==> 1, -1, 2, 4, X
1823 struct StrideCompare {
1824 const ScalarEvolution *SE;
1825 explicit StrideCompare(const ScalarEvolution *se) : SE(se) {}
1827 bool operator()(const SCEV *const &LHS, const SCEV *const &RHS) {
1828 const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS);
1829 const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS);
1831 int64_t LV = LHSC->getValue()->getSExtValue();
1832 int64_t RV = RHSC->getValue()->getSExtValue();
1833 uint64_t ALV = (LV < 0) ? -LV : LV;
1834 uint64_t ARV = (RV < 0) ? -RV : RV;
1842 // If it's the same value but different type, sort by bit width so
1843 // that we emit larger induction variables before smaller
1844 // ones, letting the smaller be re-written in terms of larger ones.
1845 return SE->getTypeSizeInBits(RHS->getType()) <
1846 SE->getTypeSizeInBits(LHS->getType());
1848 return LHSC && !RHSC;
1853 /// ChangeCompareStride - If a loop termination compare instruction is the
1854 /// only use of its stride, and the compaison is against a constant value,
1855 /// try eliminate the stride by moving the compare instruction to another
1856 /// stride and change its constant operand accordingly. e.g.
1862 /// if (v2 < 10) goto loop
1867 /// if (v1 < 30) goto loop
1868 ICmpInst *LoopStrengthReduce::ChangeCompareStride(Loop *L, ICmpInst *Cond,
1869 IVStrideUse* &CondUse,
1870 const SCEV* &CondStride,
1872 // If there's only one stride in the loop, there's nothing to do here.
1873 if (IU->StrideOrder.size() < 2)
1875 // If there are other users of the condition's stride, don't bother
1876 // trying to change the condition because the stride will still
1878 std::map<const SCEV *, IVUsersOfOneStride *>::iterator I =
1879 IU->IVUsesByStride.find(CondStride);
1880 if (I == IU->IVUsesByStride.end())
1882 if (I->second->Users.size() > 1) {
1883 for (ilist<IVStrideUse>::iterator II = I->second->Users.begin(),
1884 EE = I->second->Users.end(); II != EE; ++II) {
1885 if (II->getUser() == Cond)
1887 if (!isInstructionTriviallyDead(II->getUser()))
1891 // Only handle constant strides for now.
1892 const SCEVConstant *SC = dyn_cast<SCEVConstant>(CondStride);
1893 if (!SC) return Cond;
1895 ICmpInst::Predicate Predicate = Cond->getPredicate();
1896 int64_t CmpSSInt = SC->getValue()->getSExtValue();
1897 unsigned BitWidth = SE->getTypeSizeInBits(CondStride->getType());
1898 uint64_t SignBit = 1ULL << (BitWidth-1);
1899 const Type *CmpTy = Cond->getOperand(0)->getType();
1900 const Type *NewCmpTy = NULL;
1901 unsigned TyBits = SE->getTypeSizeInBits(CmpTy);
1902 unsigned NewTyBits = 0;
1903 const SCEV *NewStride = NULL;
1904 Value *NewCmpLHS = NULL;
1905 Value *NewCmpRHS = NULL;
1907 const SCEV *NewOffset = SE->getIntegerSCEV(0, CmpTy);
1909 if (ConstantInt *C = dyn_cast<ConstantInt>(Cond->getOperand(1))) {
1910 int64_t CmpVal = C->getValue().getSExtValue();
1912 // Check the relevant induction variable for conformance to
1914 const SCEV *IV = SE->getSCEV(Cond->getOperand(0));
1915 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1916 if (!AR || !AR->isAffine())
1919 const SCEVConstant *StartC = dyn_cast<SCEVConstant>(AR->getStart());
1920 // Check stride constant and the comparision constant signs to detect
1923 if ((StartC->getValue()->getSExtValue() < CmpVal && CmpSSInt < 0) ||
1924 (StartC->getValue()->getSExtValue() > CmpVal && CmpSSInt > 0))
1927 // More restrictive check for the other cases.
1928 if ((CmpVal & SignBit) != (CmpSSInt & SignBit))
1932 // Look for a suitable stride / iv as replacement.
1933 for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
1934 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
1935 IU->IVUsesByStride.find(IU->StrideOrder[i]);
1936 if (!isa<SCEVConstant>(SI->first) || SI->second->Users.empty())
1938 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1939 if (SSInt == CmpSSInt ||
1940 abs64(SSInt) < abs64(CmpSSInt) ||
1941 (SSInt % CmpSSInt) != 0)
1944 Scale = SSInt / CmpSSInt;
1945 int64_t NewCmpVal = CmpVal * Scale;
1947 // If old icmp value fits in icmp immediate field, but the new one doesn't
1948 // try something else.
1950 TLI->isLegalICmpImmediate(CmpVal) &&
1951 !TLI->isLegalICmpImmediate(NewCmpVal))
1954 APInt Mul = APInt(BitWidth*2, CmpVal, true);
1955 Mul = Mul * APInt(BitWidth*2, Scale, true);
1956 // Check for overflow.
1957 if (!Mul.isSignedIntN(BitWidth))
1959 // Check for overflow in the stride's type too.
1960 if (!Mul.isSignedIntN(SE->getTypeSizeInBits(SI->first->getType())))
1963 // Watch out for overflow.
1964 if (ICmpInst::isSigned(Predicate) &&
1965 (CmpVal & SignBit) != (NewCmpVal & SignBit))
1968 // Pick the best iv to use trying to avoid a cast.
1970 for (ilist<IVStrideUse>::iterator UI = SI->second->Users.begin(),
1971 E = SI->second->Users.end(); UI != E; ++UI) {
1972 Value *Op = UI->getOperandValToReplace();
1974 // If the IVStrideUse implies a cast, check for an actual cast which
1975 // can be used to find the original IV expression.
1976 if (SE->getEffectiveSCEVType(Op->getType()) !=
1977 SE->getEffectiveSCEVType(SI->first->getType())) {
1978 CastInst *CI = dyn_cast<CastInst>(Op);
1979 // If it's not a simple cast, it's complicated.
1982 // If it's a cast from a type other than the stride type,
1983 // it's complicated.
1984 if (CI->getOperand(0)->getType() != SI->first->getType())
1986 // Ok, we found the IV expression in the stride's type.
1987 Op = CI->getOperand(0);
1991 if (NewCmpLHS->getType() == CmpTy)
1997 NewCmpTy = NewCmpLHS->getType();
1998 NewTyBits = SE->getTypeSizeInBits(NewCmpTy);
1999 const Type *NewCmpIntTy = IntegerType::get(Cond->getContext(), NewTyBits);
2000 if (RequiresTypeConversion(NewCmpTy, CmpTy)) {
2001 // Check if it is possible to rewrite it using
2002 // an iv / stride of a smaller integer type.
2003 unsigned Bits = NewTyBits;
2004 if (ICmpInst::isSigned(Predicate))
2006 uint64_t Mask = (1ULL << Bits) - 1;
2007 if (((uint64_t)NewCmpVal & Mask) != (uint64_t)NewCmpVal)
2011 // Don't rewrite if use offset is non-constant and the new type is
2012 // of a different type.
2013 // FIXME: too conservative?
2014 if (NewTyBits != TyBits && !isa<SCEVConstant>(CondUse->getOffset()))
2018 bool AllUsesAreAddresses = true;
2019 bool AllUsesAreOutsideLoop = true;
2020 std::vector<BasedUser> UsersToProcess;
2021 const SCEV *CommonExprs = CollectIVUsers(SI->first, *SI->second, L,
2022 AllUsesAreAddresses,
2023 AllUsesAreOutsideLoop,
2025 // Avoid rewriting the compare instruction with an iv of new stride
2026 // if it's likely the new stride uses will be rewritten using the
2027 // stride of the compare instruction.
2028 if (AllUsesAreAddresses &&
2029 ValidScale(!CommonExprs->isZero(), Scale, UsersToProcess))
2033 // Avoid rewriting the compare instruction with an iv which has
2034 // implicit extension or truncation built into it.
2035 // TODO: This is over-conservative.
2036 if (SE->getTypeSizeInBits(CondUse->getOffset()->getType()) != TyBits)
2039 // If scale is negative, use swapped predicate unless it's testing
2041 if (Scale < 0 && !Cond->isEquality())
2042 Predicate = ICmpInst::getSwappedPredicate(Predicate);
2044 NewStride = IU->StrideOrder[i];
2045 if (!isa<PointerType>(NewCmpTy))
2046 NewCmpRHS = ConstantInt::get(NewCmpTy, NewCmpVal);
2048 Constant *CI = ConstantInt::get(NewCmpIntTy, NewCmpVal);
2049 NewCmpRHS = ConstantExpr::getIntToPtr(CI, NewCmpTy);
2051 NewOffset = TyBits == NewTyBits
2052 ? SE->getMulExpr(CondUse->getOffset(),
2053 SE->getConstant(CmpTy, Scale))
2054 : SE->getConstant(NewCmpIntTy,
2055 cast<SCEVConstant>(CondUse->getOffset())->getValue()
2056 ->getSExtValue()*Scale);
2061 // Forgo this transformation if it the increment happens to be
2062 // unfortunately positioned after the condition, and the condition
2063 // has multiple uses which prevent it from being moved immediately
2064 // before the branch. See
2065 // test/Transforms/LoopStrengthReduce/change-compare-stride-trickiness-*.ll
2066 // for an example of this situation.
2067 if (!Cond->hasOneUse()) {
2068 for (BasicBlock::iterator I = Cond, E = Cond->getParent()->end();
2075 // Create a new compare instruction using new stride / iv.
2076 ICmpInst *OldCond = Cond;
2077 // Insert new compare instruction.
2078 Cond = new ICmpInst(OldCond, Predicate, NewCmpLHS, NewCmpRHS,
2079 L->getHeader()->getName() + ".termcond");
2081 DEBUG(errs() << " Change compare stride in Inst " << *OldCond);
2082 DEBUG(errs() << " to " << *Cond << '\n');
2084 // Remove the old compare instruction. The old indvar is probably dead too.
2085 DeadInsts.push_back(CondUse->getOperandValToReplace());
2086 OldCond->replaceAllUsesWith(Cond);
2087 OldCond->eraseFromParent();
2089 IU->IVUsesByStride[NewStride]->addUser(NewOffset, Cond, NewCmpLHS);
2090 CondUse = &IU->IVUsesByStride[NewStride]->Users.back();
2091 CondStride = NewStride;
2099 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
2100 /// a max computation.
2102 /// This is a narrow solution to a specific, but acute, problem. For loops
2108 /// } while (++i < n);
2110 /// the trip count isn't just 'n', because 'n' might not be positive. And
2111 /// unfortunately this can come up even for loops where the user didn't use
2112 /// a C do-while loop. For example, seemingly well-behaved top-test loops
2113 /// will commonly be lowered like this:
2119 /// } while (++i < n);
2122 /// and then it's possible for subsequent optimization to obscure the if
2123 /// test in such a way that indvars can't find it.
2125 /// When indvars can't find the if test in loops like this, it creates a
2126 /// max expression, which allows it to give the loop a canonical
2127 /// induction variable:
2130 /// max = n < 1 ? 1 : n;
2133 /// } while (++i != max);
2135 /// Canonical induction variables are necessary because the loop passes
2136 /// are designed around them. The most obvious example of this is the
2137 /// LoopInfo analysis, which doesn't remember trip count values. It
2138 /// expects to be able to rediscover the trip count each time it is
2139 /// needed, and it does this using a simple analyis that only succeeds if
2140 /// the loop has a canonical induction variable.
2142 /// However, when it comes time to generate code, the maximum operation
2143 /// can be quite costly, especially if it's inside of an outer loop.
2145 /// This function solves this problem by detecting this type of loop and
2146 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
2147 /// the instructions for the maximum computation.
2149 ICmpInst *LoopStrengthReduce::OptimizeMax(Loop *L, ICmpInst *Cond,
2150 IVStrideUse* &CondUse) {
2151 // Check that the loop matches the pattern we're looking for.
2152 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
2153 Cond->getPredicate() != CmpInst::ICMP_NE)
2156 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
2157 if (!Sel || !Sel->hasOneUse()) return Cond;
2159 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
2160 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2162 const SCEV *One = SE->getIntegerSCEV(1, BackedgeTakenCount->getType());
2164 // Add one to the backedge-taken count to get the trip count.
2165 const SCEV *IterationCount = SE->getAddExpr(BackedgeTakenCount, One);
2167 // Check for a max calculation that matches the pattern.
2168 if (!isa<SCEVSMaxExpr>(IterationCount) && !isa<SCEVUMaxExpr>(IterationCount))
2170 const SCEVNAryExpr *Max = cast<SCEVNAryExpr>(IterationCount);
2171 if (Max != SE->getSCEV(Sel)) return Cond;
2173 // To handle a max with more than two operands, this optimization would
2174 // require additional checking and setup.
2175 if (Max->getNumOperands() != 2)
2178 const SCEV *MaxLHS = Max->getOperand(0);
2179 const SCEV *MaxRHS = Max->getOperand(1);
2180 if (!MaxLHS || MaxLHS != One) return Cond;
2182 // Check the relevant induction variable for conformance to
2184 const SCEV *IV = SE->getSCEV(Cond->getOperand(0));
2185 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2186 if (!AR || !AR->isAffine() ||
2187 AR->getStart() != One ||
2188 AR->getStepRecurrence(*SE) != One)
2191 assert(AR->getLoop() == L &&
2192 "Loop condition operand is an addrec in a different loop!");
2194 // Check the right operand of the select, and remember it, as it will
2195 // be used in the new comparison instruction.
2197 if (SE->getSCEV(Sel->getOperand(1)) == MaxRHS)
2198 NewRHS = Sel->getOperand(1);
2199 else if (SE->getSCEV(Sel->getOperand(2)) == MaxRHS)
2200 NewRHS = Sel->getOperand(2);
2201 if (!NewRHS) return Cond;
2203 // Determine the new comparison opcode. It may be signed or unsigned,
2204 // and the original comparison may be either equality or inequality.
2205 CmpInst::Predicate Pred =
2206 isa<SCEVSMaxExpr>(Max) ? CmpInst::ICMP_SLT : CmpInst::ICMP_ULT;
2207 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
2208 Pred = CmpInst::getInversePredicate(Pred);
2210 // Ok, everything looks ok to change the condition into an SLT or SGE and
2211 // delete the max calculation.
2213 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
2215 // Delete the max calculation instructions.
2216 Cond->replaceAllUsesWith(NewCond);
2217 CondUse->setUser(NewCond);
2218 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2219 Cond->eraseFromParent();
2220 Sel->eraseFromParent();
2221 if (Cmp->use_empty())
2222 Cmp->eraseFromParent();
2226 /// OptimizeShadowIV - If IV is used in a int-to-float cast
2227 /// inside the loop then try to eliminate the cast opeation.
2228 void LoopStrengthReduce::OptimizeShadowIV(Loop *L) {
2230 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
2231 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2234 for (unsigned Stride = 0, e = IU->StrideOrder.size(); Stride != e;
2236 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
2237 IU->IVUsesByStride.find(IU->StrideOrder[Stride]);
2238 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
2239 if (!isa<SCEVConstant>(SI->first))
2242 for (ilist<IVStrideUse>::iterator UI = SI->second->Users.begin(),
2243 E = SI->second->Users.end(); UI != E; /* empty */) {
2244 ilist<IVStrideUse>::iterator CandidateUI = UI;
2246 Instruction *ShadowUse = CandidateUI->getUser();
2247 const Type *DestTy = NULL;
2249 /* If shadow use is a int->float cast then insert a second IV
2250 to eliminate this cast.
2252 for (unsigned i = 0; i < n; ++i)
2258 for (unsigned i = 0; i < n; ++i, ++d)
2261 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser()))
2262 DestTy = UCast->getDestTy();
2263 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser()))
2264 DestTy = SCast->getDestTy();
2265 if (!DestTy) continue;
2268 // If target does not support DestTy natively then do not apply
2269 // this transformation.
2270 EVT DVT = TLI->getValueType(DestTy);
2271 if (!TLI->isTypeLegal(DVT)) continue;
2274 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
2276 if (PH->getNumIncomingValues() != 2) continue;
2278 const Type *SrcTy = PH->getType();
2279 int Mantissa = DestTy->getFPMantissaWidth();
2280 if (Mantissa == -1) continue;
2281 if ((int)SE->getTypeSizeInBits(SrcTy) > Mantissa)
2284 unsigned Entry, Latch;
2285 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
2293 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
2294 if (!Init) continue;
2295 Constant *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());
2297 BinaryOperator *Incr =
2298 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
2299 if (!Incr) continue;
2300 if (Incr->getOpcode() != Instruction::Add
2301 && Incr->getOpcode() != Instruction::Sub)
2304 /* Initialize new IV, double d = 0.0 in above example. */
2305 ConstantInt *C = NULL;
2306 if (Incr->getOperand(0) == PH)
2307 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
2308 else if (Incr->getOperand(1) == PH)
2309 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
2315 // Ignore negative constants, as the code below doesn't handle them
2316 // correctly. TODO: Remove this restriction.
2317 if (!C->getValue().isStrictlyPositive()) continue;
2319 /* Add new PHINode. */
2320 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);
2322 /* create new increment. '++d' in above example. */
2323 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
2324 BinaryOperator *NewIncr =
2325 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
2326 Instruction::FAdd : Instruction::FSub,
2327 NewPH, CFP, "IV.S.next.", Incr);
2329 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
2330 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
2332 /* Remove cast operation */
2333 ShadowUse->replaceAllUsesWith(NewPH);
2334 ShadowUse->eraseFromParent();
2341 /// OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
2342 /// uses in the loop, look to see if we can eliminate some, in favor of using
2343 /// common indvars for the different uses.
2344 void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
2345 // TODO: implement optzns here.
2347 OptimizeShadowIV(L);
2350 bool LoopStrengthReduce::StrideMightBeShared(const SCEV* Stride, Loop *L,
2352 int64_t SInt = cast<SCEVConstant>(Stride)->getValue()->getSExtValue();
2353 for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
2354 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
2355 IU->IVUsesByStride.find(IU->StrideOrder[i]);
2356 const SCEV *Share = SI->first;
2357 if (!isa<SCEVConstant>(SI->first) || Share == Stride)
2359 int64_t SSInt = cast<SCEVConstant>(Share)->getValue()->getSExtValue();
2361 return true; // This can definitely be reused.
2362 if (unsigned(abs64(SSInt)) < SInt || (SSInt % SInt) != 0)
2364 int64_t Scale = SSInt / SInt;
2365 bool AllUsesAreAddresses = true;
2366 bool AllUsesAreOutsideLoop = true;
2367 std::vector<BasedUser> UsersToProcess;
2368 const SCEV *CommonExprs = CollectIVUsers(SI->first, *SI->second, L,
2369 AllUsesAreAddresses,
2370 AllUsesAreOutsideLoop,
2372 if (AllUsesAreAddresses &&
2373 ValidScale(!CommonExprs->isZero(), Scale, UsersToProcess)) {
2376 // Any pre-inc iv use?
2377 IVUsersOfOneStride &StrideUses = *IU->IVUsesByStride[Share];
2378 for (ilist<IVStrideUse>::iterator I = StrideUses.Users.begin(),
2379 E = StrideUses.Users.end(); I != E; ++I) {
2380 if (!I->isUseOfPostIncrementedValue())
2388 /// isUsedByExitBranch - Return true if icmp is used by a loop terminating
2389 /// conditional branch or it's and / or with other conditions before being used
2390 /// as the condition.
2391 static bool isUsedByExitBranch(ICmpInst *Cond, Loop *L) {
2392 BasicBlock *CondBB = Cond->getParent();
2393 if (!L->isLoopExiting(CondBB))
2395 BranchInst *TermBr = dyn_cast<BranchInst>(CondBB->getTerminator());
2396 if (!TermBr || !TermBr->isConditional())
2399 Value *User = *Cond->use_begin();
2400 Instruction *UserInst = dyn_cast<Instruction>(User);
2402 (UserInst->getOpcode() == Instruction::And ||
2403 UserInst->getOpcode() == Instruction::Or)) {
2404 if (!UserInst->hasOneUse() || UserInst->getParent() != CondBB)
2406 User = *User->use_begin();
2407 UserInst = dyn_cast<Instruction>(User);
2409 return User == TermBr;
2412 static bool ShouldCountToZero(ICmpInst *Cond, IVStrideUse* &CondUse,
2413 ScalarEvolution *SE, Loop *L,
2414 const TargetLowering *TLI = 0) {
2415 if (!L->contains(Cond->getParent()))
2418 if (!isa<SCEVConstant>(CondUse->getOffset()))
2421 // Handle only tests for equality for the moment.
2422 if (!Cond->isEquality() || !Cond->hasOneUse())
2424 if (!isUsedByExitBranch(Cond, L))
2427 Value *CondOp0 = Cond->getOperand(0);
2428 const SCEV *IV = SE->getSCEV(CondOp0);
2429 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2430 if (!AR || !AR->isAffine())
2433 const SCEVConstant *SC = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE));
2434 if (!SC || SC->getValue()->getSExtValue() < 0)
2435 // If it's already counting down, don't do anything.
2438 // If the RHS of the comparison is not an loop invariant, the rewrite
2439 // cannot be done. Also bail out if it's already comparing against a zero.
2440 // If we are checking this before cmp stride optimization, check if it's
2441 // comparing against a already legal immediate.
2442 Value *RHS = Cond->getOperand(1);
2443 ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS);
2444 if (!L->isLoopInvariant(RHS) ||
2445 (RHSC && RHSC->isZero()) ||
2446 (RHSC && TLI && TLI->isLegalICmpImmediate(RHSC->getSExtValue())))
2449 // Make sure the IV is only used for counting. Value may be preinc or
2450 // postinc; 2 uses in either case.
2451 if (!CondOp0->hasNUses(2))
2457 /// OptimizeLoopTermCond - Change loop terminating condition to use the
2458 /// postinc iv when possible.
2459 void LoopStrengthReduce::OptimizeLoopTermCond(Loop *L) {
2460 BasicBlock *LatchBlock = L->getLoopLatch();
2461 bool LatchExit = L->isLoopExiting(LatchBlock);
2462 SmallVector<BasicBlock*, 8> ExitingBlocks;
2463 L->getExitingBlocks(ExitingBlocks);
2465 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
2466 BasicBlock *ExitingBlock = ExitingBlocks[i];
2468 // Finally, get the terminating condition for the loop if possible. If we
2469 // can, we want to change it to use a post-incremented version of its
2470 // induction variable, to allow coalescing the live ranges for the IV into
2471 // one register value.
2473 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2476 // FIXME: Overly conservative, termination condition could be an 'or' etc..
2477 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
2480 // Search IVUsesByStride to find Cond's IVUse if there is one.
2481 IVStrideUse *CondUse = 0;
2482 const SCEV *CondStride = 0;
2483 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2484 if (!FindIVUserForCond(Cond, CondUse, CondStride))
2487 // If the latch block is exiting and it's not a single block loop, it's
2488 // not safe to use postinc iv in other exiting blocks. FIXME: overly
2489 // conservative? How about icmp stride optimization?
2490 bool UsePostInc = !(e > 1 && LatchExit && ExitingBlock != LatchBlock);
2491 if (UsePostInc && ExitingBlock != LatchBlock) {
2492 if (!Cond->hasOneUse())
2493 // See below, we don't want the condition to be cloned.
2496 // If exiting block is the latch block, we know it's safe and profitable
2497 // to transform the icmp to use post-inc iv. Otherwise do so only if it
2498 // would not reuse another iv and its iv would be reused by other uses.
2499 // We are optimizing for the case where the icmp is the only use of the
2501 IVUsersOfOneStride &StrideUses = *IU->IVUsesByStride[CondStride];
2502 for (ilist<IVStrideUse>::iterator I = StrideUses.Users.begin(),
2503 E = StrideUses.Users.end(); I != E; ++I) {
2504 if (I->getUser() == Cond)
2506 if (!I->isUseOfPostIncrementedValue()) {
2513 // If iv for the stride might be shared and any of the users use pre-inc
2514 // iv might be used, then it's not safe to use post-inc iv.
2516 isa<SCEVConstant>(CondStride) &&
2517 StrideMightBeShared(CondStride, L, true))
2521 // If the trip count is computed in terms of a max (due to ScalarEvolution
2522 // being unable to find a sufficient guard, for example), change the loop
2523 // comparison to use SLT or ULT instead of NE.
2524 Cond = OptimizeMax(L, Cond, CondUse);
2526 // If possible, change stride and operands of the compare instruction to
2527 // eliminate one stride. However, avoid rewriting the compare instruction
2528 // with an iv of new stride if it's likely the new stride uses will be
2529 // rewritten using the stride of the compare instruction.
2530 if (ExitingBlock == LatchBlock && isa<SCEVConstant>(CondStride)) {
2531 // If the condition stride is a constant and it's the only use, we might
2532 // want to optimize it first by turning it to count toward zero.
2533 if (!StrideMightBeShared(CondStride, L, false) &&
2534 !ShouldCountToZero(Cond, CondUse, SE, L, TLI))
2535 Cond = ChangeCompareStride(L, Cond, CondUse, CondStride);
2541 DEBUG(errs() << " Change loop exiting icmp to use postinc iv: "
2544 // It's possible for the setcc instruction to be anywhere in the loop, and
2545 // possible for it to have multiple users. If it is not immediately before
2546 // the exiting block branch, move it.
2547 if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
2548 if (Cond->hasOneUse()) { // Condition has a single use, just move it.
2549 Cond->moveBefore(TermBr);
2551 // Otherwise, clone the terminating condition and insert into the
2553 Cond = cast<ICmpInst>(Cond->clone());
2554 Cond->setName(L->getHeader()->getName() + ".termcond");
2555 ExitingBlock->getInstList().insert(TermBr, Cond);
2557 // Clone the IVUse, as the old use still exists!
2558 IU->IVUsesByStride[CondStride]->addUser(CondUse->getOffset(), Cond,
2559 CondUse->getOperandValToReplace());
2560 CondUse = &IU->IVUsesByStride[CondStride]->Users.back();
2564 // If we get to here, we know that we can transform the setcc instruction to
2565 // use the post-incremented version of the IV, allowing us to coalesce the
2566 // live ranges for the IV correctly.
2567 CondUse->setOffset(SE->getMinusSCEV(CondUse->getOffset(), CondStride));
2568 CondUse->setIsUseOfPostIncrementedValue(true);
2575 bool LoopStrengthReduce::OptimizeLoopCountIVOfStride(const SCEV* &Stride,
2576 IVStrideUse* &CondUse,
2578 // If the only use is an icmp of a loop exiting conditional branch, then
2579 // attempt the optimization.
2580 BasedUser User = BasedUser(*CondUse, SE);
2581 assert(isa<ICmpInst>(User.Inst) && "Expecting an ICMPInst!");
2582 ICmpInst *Cond = cast<ICmpInst>(User.Inst);
2584 // Less strict check now that compare stride optimization is done.
2585 if (!ShouldCountToZero(Cond, CondUse, SE, L))
2588 Value *CondOp0 = Cond->getOperand(0);
2589 PHINode *PHIExpr = dyn_cast<PHINode>(CondOp0);
2592 // Value tested is postinc. Find the phi node.
2593 Incr = dyn_cast<BinaryOperator>(CondOp0);
2594 // FIXME: Just use User.OperandValToReplace here?
2595 if (!Incr || Incr->getOpcode() != Instruction::Add)
2598 PHIExpr = dyn_cast<PHINode>(Incr->getOperand(0));
2601 // 1 use for preinc value, the increment.
2602 if (!PHIExpr->hasOneUse())
2605 assert(isa<PHINode>(CondOp0) &&
2606 "Unexpected loop exiting counting instruction sequence!");
2607 PHIExpr = cast<PHINode>(CondOp0);
2608 // Value tested is preinc. Find the increment.
2609 // A CmpInst is not a BinaryOperator; we depend on this.
2610 Instruction::use_iterator UI = PHIExpr->use_begin();
2611 Incr = dyn_cast<BinaryOperator>(UI);
2613 Incr = dyn_cast<BinaryOperator>(++UI);
2614 // One use for postinc value, the phi. Unnecessarily conservative?
2615 if (!Incr || !Incr->hasOneUse() || Incr->getOpcode() != Instruction::Add)
2619 // Replace the increment with a decrement.
2620 DEBUG(errs() << "LSR: Examining use ");
2621 DEBUG(WriteAsOperand(errs(), CondOp0, /*PrintType=*/false));
2622 DEBUG(errs() << " in Inst: " << *Cond << '\n');
2623 BinaryOperator *Decr = BinaryOperator::Create(Instruction::Sub,
2624 Incr->getOperand(0), Incr->getOperand(1), "tmp", Incr);
2625 Incr->replaceAllUsesWith(Decr);
2626 Incr->eraseFromParent();
2628 // Substitute endval-startval for the original startval, and 0 for the
2629 // original endval. Since we're only testing for equality this is OK even
2630 // if the computation wraps around.
2631 BasicBlock *Preheader = L->getLoopPreheader();
2632 Instruction *PreInsertPt = Preheader->getTerminator();
2633 unsigned InBlock = L->contains(PHIExpr->getIncomingBlock(0)) ? 1 : 0;
2634 Value *StartVal = PHIExpr->getIncomingValue(InBlock);
2635 Value *EndVal = Cond->getOperand(1);
2636 DEBUG(errs() << " Optimize loop counting iv to count down ["
2637 << *EndVal << " .. " << *StartVal << "]\n");
2639 // FIXME: check for case where both are constant.
2640 Constant* Zero = ConstantInt::get(Cond->getOperand(1)->getType(), 0);
2641 BinaryOperator *NewStartVal = BinaryOperator::Create(Instruction::Sub,
2642 EndVal, StartVal, "tmp", PreInsertPt);
2643 PHIExpr->setIncomingValue(InBlock, NewStartVal);
2644 Cond->setOperand(1, Zero);
2645 DEBUG(errs() << " New icmp: " << *Cond << "\n");
2647 int64_t SInt = cast<SCEVConstant>(Stride)->getValue()->getSExtValue();
2648 const SCEV *NewStride = 0;
2650 for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
2651 const SCEV *OldStride = IU->StrideOrder[i];
2652 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(OldStride))
2653 if (SC->getValue()->getSExtValue() == -SInt) {
2655 NewStride = OldStride;
2661 NewStride = SE->getIntegerSCEV(-SInt, Stride->getType());
2662 IU->AddUser(NewStride, CondUse->getOffset(), Cond, Cond->getOperand(0));
2663 IU->IVUsesByStride[Stride]->removeUser(CondUse);
2665 CondUse = &IU->IVUsesByStride[NewStride]->Users.back();
2673 /// OptimizeLoopCountIV - If, after all sharing of IVs, the IV used for deciding
2674 /// when to exit the loop is used only for that purpose, try to rearrange things
2675 /// so it counts down to a test against zero.
2676 bool LoopStrengthReduce::OptimizeLoopCountIV(Loop *L) {
2677 bool ThisChanged = false;
2678 for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
2679 const SCEV *Stride = IU->StrideOrder[i];
2680 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
2681 IU->IVUsesByStride.find(Stride);
2682 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
2683 // FIXME: Generalize to non-affine IV's.
2684 if (!SI->first->isLoopInvariant(L))
2686 // If stride is a constant and it has an icmpinst use, check if we can
2687 // optimize the loop to count down.
2688 if (isa<SCEVConstant>(Stride) && SI->second->Users.size() == 1) {
2689 Instruction *User = SI->second->Users.begin()->getUser();
2690 if (!isa<ICmpInst>(User))
2692 const SCEV *CondStride = Stride;
2693 IVStrideUse *Use = &*SI->second->Users.begin();
2694 if (!OptimizeLoopCountIVOfStride(CondStride, Use, L))
2698 // Now check if it's possible to reuse this iv for other stride uses.
2699 for (unsigned j = 0, ee = IU->StrideOrder.size(); j != ee; ++j) {
2700 const SCEV *SStride = IU->StrideOrder[j];
2701 if (SStride == CondStride)
2703 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SII =
2704 IU->IVUsesByStride.find(SStride);
2705 assert(SII != IU->IVUsesByStride.end() && "Stride doesn't exist!");
2706 // FIXME: Generalize to non-affine IV's.
2707 if (!SII->first->isLoopInvariant(L))
2709 // FIXME: Rewrite other stride using CondStride.
2714 Changed |= ThisChanged;
2718 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager &LPM) {
2719 IU = &getAnalysis<IVUsers>();
2720 SE = &getAnalysis<ScalarEvolution>();
2723 // If LoopSimplify form is not available, stay out of trouble.
2724 if (!L->getLoopPreheader() || !L->getLoopLatch())
2727 if (!IU->IVUsesByStride.empty()) {
2728 DEBUG(errs() << "\nLSR on \"" << L->getHeader()->getParent()->getName()
2732 // Sort the StrideOrder so we process larger strides first.
2733 std::stable_sort(IU->StrideOrder.begin(), IU->StrideOrder.end(),
2736 // Optimize induction variables. Some indvar uses can be transformed to use
2737 // strides that will be needed for other purposes. A common example of this
2738 // is the exit test for the loop, which can often be rewritten to use the
2739 // computation of some other indvar to decide when to terminate the loop.
2742 // Change loop terminating condition to use the postinc iv when possible
2743 // and optimize loop terminating compare. FIXME: Move this after
2744 // StrengthReduceIVUsersOfStride?
2745 OptimizeLoopTermCond(L);
2747 // FIXME: We can shrink overlarge IV's here. e.g. if the code has
2748 // computation in i64 values and the target doesn't support i64, demote
2749 // the computation to 32-bit if safe.
2751 // FIXME: Attempt to reuse values across multiple IV's. In particular, we
2752 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should
2753 // be codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC.
2754 // Need to be careful that IV's are all the same type. Only works for
2755 // intptr_t indvars.
2757 // IVsByStride keeps IVs for one particular loop.
2758 assert(IVsByStride.empty() && "Stale entries in IVsByStride?");
2760 StrengthReduceIVUsers(L);
2762 // After all sharing is done, see if we can adjust the loop to test against
2763 // zero instead of counting up to a maximum. This is usually faster.
2764 OptimizeLoopCountIV(L);
2767 // We're done analyzing this loop; release all the state we built up for it.
2768 IVsByStride.clear();
2769 StrideNoReuse.clear();
2771 // Clean up after ourselves
2772 if (!DeadInsts.empty())
2773 DeleteTriviallyDeadInstructions();
2775 // At this point, it is worth checking to see if any recurrence PHIs are also
2776 // dead, so that we can remove them as well.
2777 DeleteDeadPHIs(L->getHeader());