1 //===- LoopStrengthReduce.cpp - Strength Reduce IVs in Loops --------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This transformation analyzes and transforms the induction variables (and
11 // computations derived from them) into forms suitable for efficient execution
14 // This pass performs a strength reduction on array references inside loops that
15 // have as one or more of their components the loop induction variable, it
16 // rewrites expressions to take advantage of scaled-index addressing modes
17 // available on the target, and it performs a variety of other optimizations
18 // related to loop induction variables.
20 //===----------------------------------------------------------------------===//
22 #define DEBUG_TYPE "loop-reduce"
23 #include "llvm/Transforms/Scalar.h"
24 #include "llvm/Constants.h"
25 #include "llvm/Instructions.h"
26 #include "llvm/IntrinsicInst.h"
27 #include "llvm/DerivedTypes.h"
28 #include "llvm/Analysis/IVUsers.h"
29 #include "llvm/Analysis/LoopPass.h"
30 #include "llvm/Analysis/ScalarEvolutionExpander.h"
31 #include "llvm/Transforms/Utils/AddrModeMatcher.h"
32 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
33 #include "llvm/Transforms/Utils/Local.h"
34 #include "llvm/ADT/Statistic.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/CommandLine.h"
37 #include "llvm/Support/ValueHandle.h"
38 #include "llvm/Support/raw_ostream.h"
39 #include "llvm/Target/TargetLowering.h"
43 STATISTIC(NumReduced , "Number of IV uses strength reduced");
44 STATISTIC(NumInserted, "Number of PHIs inserted");
45 STATISTIC(NumVariable, "Number of PHIs with variable strides");
46 STATISTIC(NumEliminated, "Number of strides eliminated");
47 STATISTIC(NumShadow, "Number of Shadow IVs optimized");
48 STATISTIC(NumImmSunk, "Number of common expr immediates sunk into uses");
49 STATISTIC(NumLoopCond, "Number of loop terminating conds optimized");
50 STATISTIC(NumCountZero, "Number of count iv optimized to count toward zero");
52 static cl::opt<bool> EnableFullLSRMode("enable-full-lsr",
60 /// IVInfo - This structure keeps track of one IV expression inserted during
61 /// StrengthReduceStridedIVUsers. It contains the stride, the common base, as
62 /// well as the PHI node and increment value created for rewrite.
68 IVExpr(const SCEV *const stride, const SCEV *const base, PHINode *phi)
69 : Stride(stride), Base(base), PHI(phi) {}
72 /// IVsOfOneStride - This structure keeps track of all IV expression inserted
73 /// during StrengthReduceStridedIVUsers for a particular stride of the IV.
74 struct IVsOfOneStride {
75 std::vector<IVExpr> IVs;
77 void addIV(const SCEV *const Stride, const SCEV *const Base, PHINode *PHI) {
78 IVs.push_back(IVExpr(Stride, Base, PHI));
82 class LoopStrengthReduce : public LoopPass {
87 /// IVsByStride - Keep track of all IVs that have been inserted for a
88 /// particular stride.
89 std::map<const SCEV *, IVsOfOneStride> IVsByStride;
91 /// DeadInsts - Keep track of instructions we may have made dead, so that
92 /// we can remove them after we are done working.
93 SmallVector<WeakVH, 16> DeadInsts;
95 /// TLI - Keep a pointer of a TargetLowering to consult for determining
96 /// transformation profitability.
97 const TargetLowering *TLI;
100 static char ID; // Pass ID, replacement for typeid
101 explicit LoopStrengthReduce(const TargetLowering *tli = NULL) :
102 LoopPass(&ID), TLI(tli) {}
104 bool runOnLoop(Loop *L, LPPassManager &LPM);
106 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
107 // We split critical edges, so we change the CFG. However, we do update
108 // many analyses if they are around.
109 AU.addPreservedID(LoopSimplifyID);
110 AU.addPreserved("loops");
111 AU.addPreserved("domfrontier");
112 AU.addPreserved("domtree");
114 AU.addRequiredID(LoopSimplifyID);
115 AU.addRequired<ScalarEvolution>();
116 AU.addPreserved<ScalarEvolution>();
117 AU.addRequired<IVUsers>();
118 AU.addPreserved<IVUsers>();
122 void OptimizeIndvars(Loop *L);
124 /// OptimizeLoopTermCond - Change loop terminating condition to use the
125 /// postinc iv when possible.
126 void OptimizeLoopTermCond(Loop *L);
128 /// OptimizeShadowIV - If IV is used in a int-to-float cast
129 /// inside the loop then try to eliminate the cast opeation.
130 void OptimizeShadowIV(Loop *L);
132 /// OptimizeMax - Rewrite the loop's terminating condition
133 /// if it uses a max computation.
134 ICmpInst *OptimizeMax(Loop *L, ICmpInst *Cond,
135 IVStrideUse* &CondUse);
137 /// OptimizeLoopCountIV - If, after all sharing of IVs, the IV used for
138 /// deciding when to exit the loop is used only for that purpose, try to
139 /// rearrange things so it counts down to a test against zero.
140 bool OptimizeLoopCountIV(Loop *L);
141 bool OptimizeLoopCountIVOfStride(const SCEV* &Stride,
142 IVStrideUse* &CondUse, Loop *L);
144 /// StrengthReduceIVUsersOfStride - Strength reduce all of the users of a
145 /// single stride of IV. All of the users may have different starting
146 /// values, and this may not be the only stride.
147 void StrengthReduceIVUsersOfStride(const SCEV *const &Stride,
148 IVUsersOfOneStride &Uses,
150 void StrengthReduceIVUsers(Loop *L);
152 ICmpInst *ChangeCompareStride(Loop *L, ICmpInst *Cond,
153 IVStrideUse* &CondUse,
154 const SCEV* &CondStride,
155 bool PostPass = false);
157 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse,
158 const SCEV* &CondStride);
159 bool RequiresTypeConversion(const Type *Ty, const Type *NewTy);
160 const SCEV *CheckForIVReuse(bool, bool, bool, const SCEV *const&,
161 IVExpr&, const Type*,
162 const std::vector<BasedUser>& UsersToProcess);
163 bool ValidScale(bool, int64_t,
164 const std::vector<BasedUser>& UsersToProcess);
165 bool ValidOffset(bool, int64_t, int64_t,
166 const std::vector<BasedUser>& UsersToProcess);
167 const SCEV *CollectIVUsers(const SCEV *const &Stride,
168 IVUsersOfOneStride &Uses,
170 bool &AllUsesAreAddresses,
171 bool &AllUsesAreOutsideLoop,
172 std::vector<BasedUser> &UsersToProcess);
173 bool StrideMightBeShared(const SCEV *Stride, Loop *L, bool CheckPreInc);
174 bool ShouldUseFullStrengthReductionMode(
175 const std::vector<BasedUser> &UsersToProcess,
177 bool AllUsesAreAddresses,
179 void PrepareToStrengthReduceFully(
180 std::vector<BasedUser> &UsersToProcess,
182 const SCEV *CommonExprs,
184 SCEVExpander &PreheaderRewriter);
185 void PrepareToStrengthReduceFromSmallerStride(
186 std::vector<BasedUser> &UsersToProcess,
188 const IVExpr &ReuseIV,
189 Instruction *PreInsertPt);
190 void PrepareToStrengthReduceWithNewPhi(
191 std::vector<BasedUser> &UsersToProcess,
193 const SCEV *CommonExprs,
195 Instruction *IVIncInsertPt,
197 SCEVExpander &PreheaderRewriter);
199 void DeleteTriviallyDeadInstructions();
203 char LoopStrengthReduce::ID = 0;
204 static RegisterPass<LoopStrengthReduce>
205 X("loop-reduce", "Loop Strength Reduction");
207 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
208 return new LoopStrengthReduce(TLI);
211 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
212 /// specified set are trivially dead, delete them and see if this makes any of
213 /// their operands subsequently dead.
214 void LoopStrengthReduce::DeleteTriviallyDeadInstructions() {
215 if (DeadInsts.empty()) return;
217 while (!DeadInsts.empty()) {
218 Instruction *I = dyn_cast_or_null<Instruction>(DeadInsts.pop_back_val());
220 if (I == 0 || !isInstructionTriviallyDead(I))
223 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
224 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
227 DeadInsts.push_back(U);
230 I->eraseFromParent();
235 /// isAddressUse - Returns true if the specified instruction is using the
236 /// specified value as an address.
237 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
238 bool isAddress = isa<LoadInst>(Inst);
239 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
240 if (SI->getOperand(1) == OperandVal)
242 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
243 // Addressing modes can also be folded into prefetches and a variety
245 switch (II->getIntrinsicID()) {
247 case Intrinsic::prefetch:
248 case Intrinsic::x86_sse2_loadu_dq:
249 case Intrinsic::x86_sse2_loadu_pd:
250 case Intrinsic::x86_sse_loadu_ps:
251 case Intrinsic::x86_sse_storeu_ps:
252 case Intrinsic::x86_sse2_storeu_pd:
253 case Intrinsic::x86_sse2_storeu_dq:
254 case Intrinsic::x86_sse2_storel_dq:
255 if (II->getOperand(1) == OperandVal)
263 /// getAccessType - Return the type of the memory being accessed.
264 static const Type *getAccessType(const Instruction *Inst) {
265 const Type *AccessTy = Inst->getType();
266 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
267 AccessTy = SI->getOperand(0)->getType();
268 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
269 // Addressing modes can also be folded into prefetches and a variety
271 switch (II->getIntrinsicID()) {
273 case Intrinsic::x86_sse_storeu_ps:
274 case Intrinsic::x86_sse2_storeu_pd:
275 case Intrinsic::x86_sse2_storeu_dq:
276 case Intrinsic::x86_sse2_storel_dq:
277 AccessTy = II->getOperand(1)->getType();
285 /// BasedUser - For a particular base value, keep information about how we've
286 /// partitioned the expression so far.
288 /// Base - The Base value for the PHI node that needs to be inserted for
289 /// this use. As the use is processed, information gets moved from this
290 /// field to the Imm field (below). BasedUser values are sorted by this
294 /// Inst - The instruction using the induction variable.
297 /// OperandValToReplace - The operand value of Inst to replace with the
299 Value *OperandValToReplace;
301 /// Imm - The immediate value that should be added to the base immediately
302 /// before Inst, because it will be folded into the imm field of the
303 /// instruction. This is also sometimes used for loop-variant values that
304 /// must be added inside the loop.
307 /// Phi - The induction variable that performs the striding that
308 /// should be used for this user.
311 // isUseOfPostIncrementedValue - True if this should use the
312 // post-incremented version of this IV, not the preincremented version.
313 // This can only be set in special cases, such as the terminating setcc
314 // instruction for a loop and uses outside the loop that are dominated by
316 bool isUseOfPostIncrementedValue;
318 BasedUser(IVStrideUse &IVSU, ScalarEvolution *se)
319 : Base(IVSU.getOffset()), Inst(IVSU.getUser()),
320 OperandValToReplace(IVSU.getOperandValToReplace()),
321 Imm(se->getIntegerSCEV(0, Base->getType())),
322 isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue()) {}
324 // Once we rewrite the code to insert the new IVs we want, update the
325 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
327 void RewriteInstructionToUseNewBase(const SCEV *const &NewBase,
328 Instruction *InsertPt,
329 SCEVExpander &Rewriter, Loop *L, Pass *P,
330 SmallVectorImpl<WeakVH> &DeadInsts,
331 ScalarEvolution *SE);
333 Value *InsertCodeForBaseAtPosition(const SCEV *const &NewBase,
335 SCEVExpander &Rewriter,
337 ScalarEvolution *SE);
342 void BasedUser::dump() const {
343 errs() << " Base=" << *Base;
344 errs() << " Imm=" << *Imm;
345 errs() << " Inst: " << *Inst;
348 Value *BasedUser::InsertCodeForBaseAtPosition(const SCEV *const &NewBase,
350 SCEVExpander &Rewriter,
352 ScalarEvolution *SE) {
353 Value *Base = Rewriter.expandCodeFor(NewBase, 0, IP);
355 // Wrap the base in a SCEVUnknown so that ScalarEvolution doesn't try to
357 const SCEV *NewValSCEV = SE->getUnknown(Base);
359 // Always emit the immediate into the same block as the user.
360 NewValSCEV = SE->getAddExpr(NewValSCEV, Imm);
362 return Rewriter.expandCodeFor(NewValSCEV, Ty, IP);
366 // Once we rewrite the code to insert the new IVs we want, update the
367 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
368 // to it. NewBasePt is the last instruction which contributes to the
369 // value of NewBase in the case that it's a diffferent instruction from
370 // the PHI that NewBase is computed from, or null otherwise.
372 void BasedUser::RewriteInstructionToUseNewBase(const SCEV *const &NewBase,
373 Instruction *NewBasePt,
374 SCEVExpander &Rewriter, Loop *L, Pass *P,
375 SmallVectorImpl<WeakVH> &DeadInsts,
376 ScalarEvolution *SE) {
377 if (!isa<PHINode>(Inst)) {
378 // By default, insert code at the user instruction.
379 BasicBlock::iterator InsertPt = Inst;
381 // However, if the Operand is itself an instruction, the (potentially
382 // complex) inserted code may be shared by many users. Because of this, we
383 // want to emit code for the computation of the operand right before its old
384 // computation. This is usually safe, because we obviously used to use the
385 // computation when it was computed in its current block. However, in some
386 // cases (e.g. use of a post-incremented induction variable) the NewBase
387 // value will be pinned to live somewhere after the original computation.
388 // In this case, we have to back off.
390 // If this is a use outside the loop (which means after, since it is based
391 // on a loop indvar) we use the post-incremented value, so that we don't
392 // artificially make the preinc value live out the bottom of the loop.
393 if (!isUseOfPostIncrementedValue && L->contains(Inst->getParent())) {
394 if (NewBasePt && isa<PHINode>(OperandValToReplace)) {
395 InsertPt = NewBasePt;
397 } else if (Instruction *OpInst
398 = dyn_cast<Instruction>(OperandValToReplace)) {
400 while (isa<PHINode>(InsertPt)) ++InsertPt;
403 Value *NewVal = InsertCodeForBaseAtPosition(NewBase,
404 OperandValToReplace->getType(),
405 Rewriter, InsertPt, SE);
406 // Replace the use of the operand Value with the new Phi we just created.
407 Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
409 DEBUG(errs() << " Replacing with ");
410 DEBUG(WriteAsOperand(errs(), NewVal, /*PrintType=*/false));
411 DEBUG(errs() << ", which has value " << *NewBase << " plus IMM "
416 // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm
417 // expression into each operand block that uses it. Note that PHI nodes can
418 // have multiple entries for the same predecessor. We use a map to make sure
419 // that a PHI node only has a single Value* for each predecessor (which also
420 // prevents us from inserting duplicate code in some blocks).
421 DenseMap<BasicBlock*, Value*> InsertedCode;
422 PHINode *PN = cast<PHINode>(Inst);
423 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
424 if (PN->getIncomingValue(i) == OperandValToReplace) {
425 // If the original expression is outside the loop, put the replacement
426 // code in the same place as the original expression,
427 // which need not be an immediate predecessor of this PHI. This way we
428 // need only one copy of it even if it is referenced multiple times in
429 // the PHI. We don't do this when the original expression is inside the
430 // loop because multiple copies sometimes do useful sinking of code in
432 Instruction *OldLoc = dyn_cast<Instruction>(OperandValToReplace);
433 BasicBlock *PHIPred = PN->getIncomingBlock(i);
434 if (L->contains(OldLoc->getParent())) {
435 // If this is a critical edge, split the edge so that we do not insert
436 // the code on all predecessor/successor paths. We do this unless this
437 // is the canonical backedge for this loop, as this can make some
438 // inserted code be in an illegal position.
439 if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 &&
440 !isa<IndirectBrInst>(PHIPred->getTerminator()) &&
441 (PN->getParent() != L->getHeader() || !L->contains(PHIPred))) {
443 // First step, split the critical edge.
444 BasicBlock *NewBB = SplitCriticalEdge(PHIPred, PN->getParent(),
447 // Next step: move the basic block. In particular, if the PHI node
448 // is outside of the loop, and PredTI is in the loop, we want to
449 // move the block to be immediately before the PHI block, not
450 // immediately after PredTI.
451 if (L->contains(PHIPred) && !L->contains(PN->getParent()))
452 NewBB->moveBefore(PN->getParent());
454 // Splitting the edge can reduce the number of PHI entries we have.
455 e = PN->getNumIncomingValues();
457 i = PN->getBasicBlockIndex(PHIPred);
460 Value *&Code = InsertedCode[PHIPred];
462 // Insert the code into the end of the predecessor block.
463 Instruction *InsertPt = (L->contains(OldLoc->getParent())) ?
464 PHIPred->getTerminator() :
465 OldLoc->getParent()->getTerminator();
466 Code = InsertCodeForBaseAtPosition(NewBase, PN->getType(),
467 Rewriter, InsertPt, SE);
469 DEBUG(errs() << " Changing PHI use to ");
470 DEBUG(WriteAsOperand(errs(), Code, /*PrintType=*/false));
471 DEBUG(errs() << ", which has value " << *NewBase << " plus IMM "
475 // Replace the use of the operand Value with the new Phi we just created.
476 PN->setIncomingValue(i, Code);
481 // PHI node might have become a constant value after SplitCriticalEdge.
482 DeadInsts.push_back(Inst);
486 /// fitsInAddressMode - Return true if V can be subsumed within an addressing
487 /// mode, and does not need to be put in a register first.
488 static bool fitsInAddressMode(const SCEV *const &V, const Type *AccessTy,
489 const TargetLowering *TLI, bool HasBaseReg) {
490 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
491 int64_t VC = SC->getValue()->getSExtValue();
493 TargetLowering::AddrMode AM;
495 AM.HasBaseReg = HasBaseReg;
496 return TLI->isLegalAddressingMode(AM, AccessTy);
498 // Defaults to PPC. PPC allows a sign-extended 16-bit immediate field.
499 return (VC > -(1 << 16) && VC < (1 << 16)-1);
503 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
504 if (GlobalValue *GV = dyn_cast<GlobalValue>(SU->getValue())) {
506 TargetLowering::AddrMode AM;
508 AM.HasBaseReg = HasBaseReg;
509 return TLI->isLegalAddressingMode(AM, AccessTy);
511 // Default: assume global addresses are not legal.
518 /// MoveLoopVariantsToImmediateField - Move any subexpressions from Val that are
519 /// loop varying to the Imm operand.
520 static void MoveLoopVariantsToImmediateField(const SCEV *&Val, const SCEV *&Imm,
521 Loop *L, ScalarEvolution *SE) {
522 if (Val->isLoopInvariant(L)) return; // Nothing to do.
524 if (const SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
525 SmallVector<const SCEV *, 4> NewOps;
526 NewOps.reserve(SAE->getNumOperands());
528 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
529 if (!SAE->getOperand(i)->isLoopInvariant(L)) {
530 // If this is a loop-variant expression, it must stay in the immediate
531 // field of the expression.
532 Imm = SE->getAddExpr(Imm, SAE->getOperand(i));
534 NewOps.push_back(SAE->getOperand(i));
538 Val = SE->getIntegerSCEV(0, Val->getType());
540 Val = SE->getAddExpr(NewOps);
541 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
542 // Try to pull immediates out of the start value of nested addrec's.
543 const SCEV *Start = SARE->getStart();
544 MoveLoopVariantsToImmediateField(Start, Imm, L, SE);
546 SmallVector<const SCEV *, 4> Ops(SARE->op_begin(), SARE->op_end());
548 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
550 // Otherwise, all of Val is variant, move the whole thing over.
551 Imm = SE->getAddExpr(Imm, Val);
552 Val = SE->getIntegerSCEV(0, Val->getType());
557 /// MoveImmediateValues - Look at Val, and pull out any additions of constants
558 /// that can fit into the immediate field of instructions in the target.
559 /// Accumulate these immediate values into the Imm value.
560 static void MoveImmediateValues(const TargetLowering *TLI,
561 const Type *AccessTy,
562 const SCEV *&Val, const SCEV *&Imm,
563 bool isAddress, Loop *L,
564 ScalarEvolution *SE) {
565 if (const SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
566 SmallVector<const SCEV *, 4> NewOps;
567 NewOps.reserve(SAE->getNumOperands());
569 for (unsigned i = 0; i != SAE->getNumOperands(); ++i) {
570 const SCEV *NewOp = SAE->getOperand(i);
571 MoveImmediateValues(TLI, AccessTy, NewOp, Imm, isAddress, L, SE);
573 if (!NewOp->isLoopInvariant(L)) {
574 // If this is a loop-variant expression, it must stay in the immediate
575 // field of the expression.
576 Imm = SE->getAddExpr(Imm, NewOp);
578 NewOps.push_back(NewOp);
583 Val = SE->getIntegerSCEV(0, Val->getType());
585 Val = SE->getAddExpr(NewOps);
587 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
588 // Try to pull immediates out of the start value of nested addrec's.
589 const SCEV *Start = SARE->getStart();
590 MoveImmediateValues(TLI, AccessTy, Start, Imm, isAddress, L, SE);
592 if (Start != SARE->getStart()) {
593 SmallVector<const SCEV *, 4> Ops(SARE->op_begin(), SARE->op_end());
595 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
598 } else if (const SCEVMulExpr *SME = dyn_cast<SCEVMulExpr>(Val)) {
599 // Transform "8 * (4 + v)" -> "32 + 8*V" if "32" fits in the immed field.
601 fitsInAddressMode(SME->getOperand(0), AccessTy, TLI, false) &&
602 SME->getNumOperands() == 2 && SME->isLoopInvariant(L)) {
604 const SCEV *SubImm = SE->getIntegerSCEV(0, Val->getType());
605 const SCEV *NewOp = SME->getOperand(1);
606 MoveImmediateValues(TLI, AccessTy, NewOp, SubImm, isAddress, L, SE);
608 // If we extracted something out of the subexpressions, see if we can
610 if (NewOp != SME->getOperand(1)) {
611 // Scale SubImm up by "8". If the result is a target constant, we are
613 SubImm = SE->getMulExpr(SubImm, SME->getOperand(0));
614 if (fitsInAddressMode(SubImm, AccessTy, TLI, false)) {
615 // Accumulate the immediate.
616 Imm = SE->getAddExpr(Imm, SubImm);
618 // Update what is left of 'Val'.
619 Val = SE->getMulExpr(SME->getOperand(0), NewOp);
626 // Loop-variant expressions must stay in the immediate field of the
628 if ((isAddress && fitsInAddressMode(Val, AccessTy, TLI, false)) ||
629 !Val->isLoopInvariant(L)) {
630 Imm = SE->getAddExpr(Imm, Val);
631 Val = SE->getIntegerSCEV(0, Val->getType());
635 // Otherwise, no immediates to move.
638 static void MoveImmediateValues(const TargetLowering *TLI,
640 const SCEV *&Val, const SCEV *&Imm,
641 bool isAddress, Loop *L,
642 ScalarEvolution *SE) {
643 const Type *AccessTy = getAccessType(User);
644 MoveImmediateValues(TLI, AccessTy, Val, Imm, isAddress, L, SE);
647 /// SeparateSubExprs - Decompose Expr into all of the subexpressions that are
648 /// added together. This is used to reassociate common addition subexprs
649 /// together for maximal sharing when rewriting bases.
650 static void SeparateSubExprs(SmallVector<const SCEV *, 16> &SubExprs,
652 ScalarEvolution *SE) {
653 if (const SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) {
654 for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
655 SeparateSubExprs(SubExprs, AE->getOperand(j), SE);
656 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) {
657 const SCEV *Zero = SE->getIntegerSCEV(0, Expr->getType());
658 if (SARE->getOperand(0) == Zero) {
659 SubExprs.push_back(Expr);
661 // Compute the addrec with zero as its base.
662 SmallVector<const SCEV *, 4> Ops(SARE->op_begin(), SARE->op_end());
663 Ops[0] = Zero; // Start with zero base.
664 SubExprs.push_back(SE->getAddRecExpr(Ops, SARE->getLoop()));
667 SeparateSubExprs(SubExprs, SARE->getOperand(0), SE);
669 } else if (!Expr->isZero()) {
671 SubExprs.push_back(Expr);
675 // This is logically local to the following function, but C++ says we have
676 // to make it file scope.
677 struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; };
679 /// RemoveCommonExpressionsFromUseBases - Look through all of the Bases of all
680 /// the Uses, removing any common subexpressions, except that if all such
681 /// subexpressions can be folded into an addressing mode for all uses inside
682 /// the loop (this case is referred to as "free" in comments herein) we do
683 /// not remove anything. This looks for things like (a+b+c) and
684 /// (a+c+d) and computes the common (a+c) subexpression. The common expression
685 /// is *removed* from the Bases and returned.
687 RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses,
688 ScalarEvolution *SE, Loop *L,
689 const TargetLowering *TLI) {
690 unsigned NumUses = Uses.size();
692 // Only one use? This is a very common case, so we handle it specially and
694 const SCEV *Zero = SE->getIntegerSCEV(0, Uses[0].Base->getType());
695 const SCEV *Result = Zero;
696 const SCEV *FreeResult = Zero;
698 // If the use is inside the loop, use its base, regardless of what it is:
699 // it is clearly shared across all the IV's. If the use is outside the loop
700 // (which means after it) we don't want to factor anything *into* the loop,
701 // so just use 0 as the base.
702 if (L->contains(Uses[0].Inst->getParent()))
703 std::swap(Result, Uses[0].Base);
707 // To find common subexpressions, count how many of Uses use each expression.
708 // If any subexpressions are used Uses.size() times, they are common.
709 // Also track whether all uses of each expression can be moved into an
710 // an addressing mode "for free"; such expressions are left within the loop.
711 // struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; };
712 std::map<const SCEV *, SubExprUseData> SubExpressionUseData;
714 // UniqueSubExprs - Keep track of all of the subexpressions we see in the
715 // order we see them.
716 SmallVector<const SCEV *, 16> UniqueSubExprs;
718 SmallVector<const SCEV *, 16> SubExprs;
719 unsigned NumUsesInsideLoop = 0;
720 for (unsigned i = 0; i != NumUses; ++i) {
721 // If the user is outside the loop, just ignore it for base computation.
722 // Since the user is outside the loop, it must be *after* the loop (if it
723 // were before, it could not be based on the loop IV). We don't want users
724 // after the loop to affect base computation of values *inside* the loop,
725 // because we can always add their offsets to the result IV after the loop
726 // is done, ensuring we get good code inside the loop.
727 if (!L->contains(Uses[i].Inst->getParent()))
731 // If the base is zero (which is common), return zero now, there are no
733 if (Uses[i].Base == Zero) return Zero;
735 // If this use is as an address we may be able to put CSEs in the addressing
736 // mode rather than hoisting them.
737 bool isAddrUse = isAddressUse(Uses[i].Inst, Uses[i].OperandValToReplace);
738 // We may need the AccessTy below, but only when isAddrUse, so compute it
739 // only in that case.
740 const Type *AccessTy = 0;
742 AccessTy = getAccessType(Uses[i].Inst);
744 // Split the expression into subexprs.
745 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
746 // Add one to SubExpressionUseData.Count for each subexpr present, and
747 // if the subexpr is not a valid immediate within an addressing mode use,
748 // set SubExpressionUseData.notAllUsesAreFree. We definitely want to
749 // hoist these out of the loop (if they are common to all uses).
750 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) {
751 if (++SubExpressionUseData[SubExprs[j]].Count == 1)
752 UniqueSubExprs.push_back(SubExprs[j]);
753 if (!isAddrUse || !fitsInAddressMode(SubExprs[j], AccessTy, TLI, false))
754 SubExpressionUseData[SubExprs[j]].notAllUsesAreFree = true;
759 // Now that we know how many times each is used, build Result. Iterate over
760 // UniqueSubexprs so that we have a stable ordering.
761 for (unsigned i = 0, e = UniqueSubExprs.size(); i != e; ++i) {
762 std::map<const SCEV *, SubExprUseData>::iterator I =
763 SubExpressionUseData.find(UniqueSubExprs[i]);
764 assert(I != SubExpressionUseData.end() && "Entry not found?");
765 if (I->second.Count == NumUsesInsideLoop) { // Found CSE!
766 if (I->second.notAllUsesAreFree)
767 Result = SE->getAddExpr(Result, I->first);
769 FreeResult = SE->getAddExpr(FreeResult, I->first);
771 // Remove non-cse's from SubExpressionUseData.
772 SubExpressionUseData.erase(I);
775 if (FreeResult != Zero) {
776 // We have some subexpressions that can be subsumed into addressing
777 // modes in every use inside the loop. However, it's possible that
778 // there are so many of them that the combined FreeResult cannot
779 // be subsumed, or that the target cannot handle both a FreeResult
780 // and a Result in the same instruction (for example because it would
781 // require too many registers). Check this.
782 for (unsigned i=0; i<NumUses; ++i) {
783 if (!L->contains(Uses[i].Inst->getParent()))
785 // We know this is an addressing mode use; if there are any uses that
786 // are not, FreeResult would be Zero.
787 const Type *AccessTy = getAccessType(Uses[i].Inst);
788 if (!fitsInAddressMode(FreeResult, AccessTy, TLI, Result!=Zero)) {
789 // FIXME: could split up FreeResult into pieces here, some hoisted
790 // and some not. There is no obvious advantage to this.
791 Result = SE->getAddExpr(Result, FreeResult);
798 // If we found no CSE's, return now.
799 if (Result == Zero) return Result;
801 // If we still have a FreeResult, remove its subexpressions from
802 // SubExpressionUseData. This means they will remain in the use Bases.
803 if (FreeResult != Zero) {
804 SeparateSubExprs(SubExprs, FreeResult, SE);
805 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) {
806 std::map<const SCEV *, SubExprUseData>::iterator I =
807 SubExpressionUseData.find(SubExprs[j]);
808 SubExpressionUseData.erase(I);
813 // Otherwise, remove all of the CSE's we found from each of the base values.
814 for (unsigned i = 0; i != NumUses; ++i) {
815 // Uses outside the loop don't necessarily include the common base, but
816 // the final IV value coming into those uses does. Instead of trying to
817 // remove the pieces of the common base, which might not be there,
818 // subtract off the base to compensate for this.
819 if (!L->contains(Uses[i].Inst->getParent())) {
820 Uses[i].Base = SE->getMinusSCEV(Uses[i].Base, Result);
824 // Split the expression into subexprs.
825 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
827 // Remove any common subexpressions.
828 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
829 if (SubExpressionUseData.count(SubExprs[j])) {
830 SubExprs.erase(SubExprs.begin()+j);
834 // Finally, add the non-shared expressions together.
835 if (SubExprs.empty())
838 Uses[i].Base = SE->getAddExpr(SubExprs);
845 /// ValidScale - Check whether the given Scale is valid for all loads and
846 /// stores in UsersToProcess.
848 bool LoopStrengthReduce::ValidScale(bool HasBaseReg, int64_t Scale,
849 const std::vector<BasedUser>& UsersToProcess) {
853 for (unsigned i = 0, e = UsersToProcess.size(); i!=e; ++i) {
854 // If this is a load or other access, pass the type of the access in.
855 const Type *AccessTy =
856 Type::getVoidTy(UsersToProcess[i].Inst->getContext());
857 if (isAddressUse(UsersToProcess[i].Inst,
858 UsersToProcess[i].OperandValToReplace))
859 AccessTy = getAccessType(UsersToProcess[i].Inst);
860 else if (isa<PHINode>(UsersToProcess[i].Inst))
863 TargetLowering::AddrMode AM;
864 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
865 AM.BaseOffs = SC->getValue()->getSExtValue();
866 AM.HasBaseReg = HasBaseReg || !UsersToProcess[i].Base->isZero();
869 // If load[imm+r*scale] is illegal, bail out.
870 if (!TLI->isLegalAddressingMode(AM, AccessTy))
876 /// ValidOffset - Check whether the given Offset is valid for all loads and
877 /// stores in UsersToProcess.
879 bool LoopStrengthReduce::ValidOffset(bool HasBaseReg,
882 const std::vector<BasedUser>& UsersToProcess) {
886 for (unsigned i=0, e = UsersToProcess.size(); i!=e; ++i) {
887 // If this is a load or other access, pass the type of the access in.
888 const Type *AccessTy =
889 Type::getVoidTy(UsersToProcess[i].Inst->getContext());
890 if (isAddressUse(UsersToProcess[i].Inst,
891 UsersToProcess[i].OperandValToReplace))
892 AccessTy = getAccessType(UsersToProcess[i].Inst);
893 else if (isa<PHINode>(UsersToProcess[i].Inst))
896 TargetLowering::AddrMode AM;
897 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
898 AM.BaseOffs = SC->getValue()->getSExtValue();
899 AM.BaseOffs = (uint64_t)AM.BaseOffs + (uint64_t)Offset;
900 AM.HasBaseReg = HasBaseReg || !UsersToProcess[i].Base->isZero();
903 // If load[imm+r*scale] is illegal, bail out.
904 if (!TLI->isLegalAddressingMode(AM, AccessTy))
910 /// RequiresTypeConversion - Returns true if converting Ty1 to Ty2 is not
912 bool LoopStrengthReduce::RequiresTypeConversion(const Type *Ty1,
916 Ty1 = SE->getEffectiveSCEVType(Ty1);
917 Ty2 = SE->getEffectiveSCEVType(Ty2);
920 if (Ty1->canLosslesslyBitCastTo(Ty2))
922 if (TLI && TLI->isTruncateFree(Ty1, Ty2))
927 /// CheckForIVReuse - Returns the multiple if the stride is the multiple
928 /// of a previous stride and it is a legal value for the target addressing
929 /// mode scale component and optional base reg. This allows the users of
930 /// this stride to be rewritten as prev iv * factor. It returns 0 if no
931 /// reuse is possible. Factors can be negative on same targets, e.g. ARM.
933 /// If all uses are outside the loop, we don't require that all multiplies
934 /// be folded into the addressing mode, nor even that the factor be constant;
935 /// a multiply (executed once) outside the loop is better than another IV
936 /// within. Well, usually.
937 const SCEV *LoopStrengthReduce::CheckForIVReuse(bool HasBaseReg,
938 bool AllUsesAreAddresses,
939 bool AllUsesAreOutsideLoop,
940 const SCEV *const &Stride,
941 IVExpr &IV, const Type *Ty,
942 const std::vector<BasedUser>& UsersToProcess) {
943 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Stride)) {
944 int64_t SInt = SC->getValue()->getSExtValue();
945 for (unsigned NewStride = 0, e = IU->StrideOrder.size();
946 NewStride != e; ++NewStride) {
947 std::map<const SCEV *, IVsOfOneStride>::iterator SI =
948 IVsByStride.find(IU->StrideOrder[NewStride]);
949 if (SI == IVsByStride.end() || !isa<SCEVConstant>(SI->first))
951 // The other stride has no uses, don't reuse it.
952 std::map<const SCEV *, IVUsersOfOneStride *>::iterator UI =
953 IU->IVUsesByStride.find(IU->StrideOrder[NewStride]);
954 if (UI->second->Users.empty())
956 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
957 if (SI->first != Stride &&
958 (unsigned(abs64(SInt)) < SSInt || (SInt % SSInt) != 0))
960 int64_t Scale = SInt / SSInt;
961 // Check that this stride is valid for all the types used for loads and
962 // stores; if it can be used for some and not others, we might as well use
963 // the original stride everywhere, since we have to create the IV for it
964 // anyway. If the scale is 1, then we don't need to worry about folding
967 (AllUsesAreAddresses &&
968 ValidScale(HasBaseReg, Scale, UsersToProcess))) {
969 // Prefer to reuse an IV with a base of zero.
970 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
971 IE = SI->second.IVs.end(); II != IE; ++II)
972 // Only reuse previous IV if it would not require a type conversion
973 // and if the base difference can be folded.
974 if (II->Base->isZero() &&
975 !RequiresTypeConversion(II->Base->getType(), Ty)) {
977 return SE->getIntegerSCEV(Scale, Stride->getType());
979 // Otherwise, settle for an IV with a foldable base.
980 if (AllUsesAreAddresses)
981 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
982 IE = SI->second.IVs.end(); II != IE; ++II)
983 // Only reuse previous IV if it would not require a type conversion
984 // and if the base difference can be folded.
985 if (SE->getEffectiveSCEVType(II->Base->getType()) ==
986 SE->getEffectiveSCEVType(Ty) &&
987 isa<SCEVConstant>(II->Base)) {
989 cast<SCEVConstant>(II->Base)->getValue()->getSExtValue();
990 if (Base > INT32_MIN && Base <= INT32_MAX &&
991 ValidOffset(HasBaseReg, -Base * Scale,
992 Scale, UsersToProcess)) {
994 return SE->getIntegerSCEV(Scale, Stride->getType());
999 } else if (AllUsesAreOutsideLoop) {
1000 // Accept nonconstant strides here; it is really really right to substitute
1001 // an existing IV if we can.
1002 for (unsigned NewStride = 0, e = IU->StrideOrder.size();
1003 NewStride != e; ++NewStride) {
1004 std::map<const SCEV *, IVsOfOneStride>::iterator SI =
1005 IVsByStride.find(IU->StrideOrder[NewStride]);
1006 if (SI == IVsByStride.end() || !isa<SCEVConstant>(SI->first))
1008 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1009 if (SI->first != Stride && SSInt != 1)
1011 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1012 IE = SI->second.IVs.end(); II != IE; ++II)
1013 // Accept nonzero base here.
1014 // Only reuse previous IV if it would not require a type conversion.
1015 if (!RequiresTypeConversion(II->Base->getType(), Ty)) {
1020 // Special case, old IV is -1*x and this one is x. Can treat this one as
1022 for (unsigned NewStride = 0, e = IU->StrideOrder.size();
1023 NewStride != e; ++NewStride) {
1024 std::map<const SCEV *, IVsOfOneStride>::iterator SI =
1025 IVsByStride.find(IU->StrideOrder[NewStride]);
1026 if (SI == IVsByStride.end())
1028 if (const SCEVMulExpr *ME = dyn_cast<SCEVMulExpr>(SI->first))
1029 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(ME->getOperand(0)))
1030 if (Stride == ME->getOperand(1) &&
1031 SC->getValue()->getSExtValue() == -1LL)
1032 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1033 IE = SI->second.IVs.end(); II != IE; ++II)
1034 // Accept nonzero base here.
1035 // Only reuse previous IV if it would not require type conversion.
1036 if (!RequiresTypeConversion(II->Base->getType(), Ty)) {
1038 return SE->getIntegerSCEV(-1LL, Stride->getType());
1042 return SE->getIntegerSCEV(0, Stride->getType());
1045 /// PartitionByIsUseOfPostIncrementedValue - Simple boolean predicate that
1046 /// returns true if Val's isUseOfPostIncrementedValue is true.
1047 static bool PartitionByIsUseOfPostIncrementedValue(const BasedUser &Val) {
1048 return Val.isUseOfPostIncrementedValue;
1051 /// isNonConstantNegative - Return true if the specified scev is negated, but
1053 static bool isNonConstantNegative(const SCEV *const &Expr) {
1054 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Expr);
1055 if (!Mul) return false;
1057 // If there is a constant factor, it will be first.
1058 const SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
1059 if (!SC) return false;
1061 // Return true if the value is negative, this matches things like (-42 * V).
1062 return SC->getValue()->getValue().isNegative();
1065 /// CollectIVUsers - Transform our list of users and offsets to a bit more
1066 /// complex table. In this new vector, each 'BasedUser' contains 'Base', the
1067 /// base of the strided accesses, as well as the old information from Uses. We
1068 /// progressively move information from the Base field to the Imm field, until
1069 /// we eventually have the full access expression to rewrite the use.
1070 const SCEV *LoopStrengthReduce::CollectIVUsers(const SCEV *const &Stride,
1071 IVUsersOfOneStride &Uses,
1073 bool &AllUsesAreAddresses,
1074 bool &AllUsesAreOutsideLoop,
1075 std::vector<BasedUser> &UsersToProcess) {
1076 // FIXME: Generalize to non-affine IV's.
1077 if (!Stride->isLoopInvariant(L))
1078 return SE->getIntegerSCEV(0, Stride->getType());
1080 UsersToProcess.reserve(Uses.Users.size());
1081 for (ilist<IVStrideUse>::iterator I = Uses.Users.begin(),
1082 E = Uses.Users.end(); I != E; ++I) {
1083 UsersToProcess.push_back(BasedUser(*I, SE));
1085 // Move any loop variant operands from the offset field to the immediate
1086 // field of the use, so that we don't try to use something before it is
1088 MoveLoopVariantsToImmediateField(UsersToProcess.back().Base,
1089 UsersToProcess.back().Imm, L, SE);
1090 assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
1091 "Base value is not loop invariant!");
1094 // We now have a whole bunch of uses of like-strided induction variables, but
1095 // they might all have different bases. We want to emit one PHI node for this
1096 // stride which we fold as many common expressions (between the IVs) into as
1097 // possible. Start by identifying the common expressions in the base values
1098 // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
1099 // "A+B"), emit it to the preheader, then remove the expression from the
1100 // UsersToProcess base values.
1101 const SCEV *CommonExprs =
1102 RemoveCommonExpressionsFromUseBases(UsersToProcess, SE, L, TLI);
1104 // Next, figure out what we can represent in the immediate fields of
1105 // instructions. If we can represent anything there, move it to the imm
1106 // fields of the BasedUsers. We do this so that it increases the commonality
1107 // of the remaining uses.
1108 unsigned NumPHI = 0;
1109 bool HasAddress = false;
1110 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1111 // If the user is not in the current loop, this means it is using the exit
1112 // value of the IV. Do not put anything in the base, make sure it's all in
1113 // the immediate field to allow as much factoring as possible.
1114 if (!L->contains(UsersToProcess[i].Inst->getParent())) {
1115 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm,
1116 UsersToProcess[i].Base);
1117 UsersToProcess[i].Base =
1118 SE->getIntegerSCEV(0, UsersToProcess[i].Base->getType());
1120 // Not all uses are outside the loop.
1121 AllUsesAreOutsideLoop = false;
1123 // Addressing modes can be folded into loads and stores. Be careful that
1124 // the store is through the expression, not of the expression though.
1126 bool isAddress = isAddressUse(UsersToProcess[i].Inst,
1127 UsersToProcess[i].OperandValToReplace);
1128 if (isa<PHINode>(UsersToProcess[i].Inst)) {
1136 // If this use isn't an address, then not all uses are addresses.
1137 if (!isAddress && !isPHI)
1138 AllUsesAreAddresses = false;
1140 MoveImmediateValues(TLI, UsersToProcess[i].Inst, UsersToProcess[i].Base,
1141 UsersToProcess[i].Imm, isAddress, L, SE);
1145 // If one of the use is a PHI node and all other uses are addresses, still
1146 // allow iv reuse. Essentially we are trading one constant multiplication
1147 // for one fewer iv.
1149 AllUsesAreAddresses = false;
1151 // There are no in-loop address uses.
1152 if (AllUsesAreAddresses && (!HasAddress && !AllUsesAreOutsideLoop))
1153 AllUsesAreAddresses = false;
1158 /// ShouldUseFullStrengthReductionMode - Test whether full strength-reduction
1159 /// is valid and profitable for the given set of users of a stride. In
1160 /// full strength-reduction mode, all addresses at the current stride are
1161 /// strength-reduced all the way down to pointer arithmetic.
1163 bool LoopStrengthReduce::ShouldUseFullStrengthReductionMode(
1164 const std::vector<BasedUser> &UsersToProcess,
1166 bool AllUsesAreAddresses,
1167 const SCEV *Stride) {
1168 if (!EnableFullLSRMode)
1171 // The heuristics below aim to avoid increasing register pressure, but
1172 // fully strength-reducing all the addresses increases the number of
1173 // add instructions, so don't do this when optimizing for size.
1174 // TODO: If the loop is large, the savings due to simpler addresses
1175 // may oughtweight the costs of the extra increment instructions.
1176 if (L->getHeader()->getParent()->hasFnAttr(Attribute::OptimizeForSize))
1179 // TODO: For now, don't do full strength reduction if there could
1180 // potentially be greater-stride multiples of the current stride
1181 // which could reuse the current stride IV.
1182 if (IU->StrideOrder.back() != Stride)
1185 // Iterate through the uses to find conditions that automatically rule out
1187 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ) {
1188 const SCEV *Base = UsersToProcess[i].Base;
1189 const SCEV *Imm = UsersToProcess[i].Imm;
1190 // If any users have a loop-variant component, they can't be fully
1191 // strength-reduced.
1192 if (Imm && !Imm->isLoopInvariant(L))
1194 // If there are to users with the same base and the difference between
1195 // the two Imm values can't be folded into the address, full
1196 // strength reduction would increase register pressure.
1198 const SCEV *CurImm = UsersToProcess[i].Imm;
1199 if ((CurImm || Imm) && CurImm != Imm) {
1200 if (!CurImm) CurImm = SE->getIntegerSCEV(0, Stride->getType());
1201 if (!Imm) Imm = SE->getIntegerSCEV(0, Stride->getType());
1202 const Instruction *Inst = UsersToProcess[i].Inst;
1203 const Type *AccessTy = getAccessType(Inst);
1204 const SCEV *Diff = SE->getMinusSCEV(UsersToProcess[i].Imm, Imm);
1205 if (!Diff->isZero() &&
1206 (!AllUsesAreAddresses ||
1207 !fitsInAddressMode(Diff, AccessTy, TLI, /*HasBaseReg=*/true)))
1210 } while (++i != e && Base == UsersToProcess[i].Base);
1213 // If there's exactly one user in this stride, fully strength-reducing it
1214 // won't increase register pressure. If it's starting from a non-zero base,
1215 // it'll be simpler this way.
1216 if (UsersToProcess.size() == 1 && !UsersToProcess[0].Base->isZero())
1219 // Otherwise, if there are any users in this stride that don't require
1220 // a register for their base, full strength-reduction will increase
1221 // register pressure.
1222 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1223 if (UsersToProcess[i].Base->isZero())
1226 // Otherwise, go for it.
1230 /// InsertAffinePhi Create and insert a PHI node for an induction variable
1231 /// with the specified start and step values in the specified loop.
1233 /// If NegateStride is true, the stride should be negated by using a
1234 /// subtract instead of an add.
1236 /// Return the created phi node.
1238 static PHINode *InsertAffinePhi(const SCEV *Start, const SCEV *Step,
1239 Instruction *IVIncInsertPt,
1241 SCEVExpander &Rewriter) {
1242 assert(Start->isLoopInvariant(L) && "New PHI start is not loop invariant!");
1243 assert(Step->isLoopInvariant(L) && "New PHI stride is not loop invariant!");
1245 BasicBlock *Header = L->getHeader();
1246 BasicBlock *Preheader = L->getLoopPreheader();
1247 BasicBlock *LatchBlock = L->getLoopLatch();
1248 const Type *Ty = Start->getType();
1249 Ty = Rewriter.SE.getEffectiveSCEVType(Ty);
1251 PHINode *PN = PHINode::Create(Ty, "lsr.iv", Header->begin());
1252 PN->addIncoming(Rewriter.expandCodeFor(Start, Ty, Preheader->getTerminator()),
1255 // If the stride is negative, insert a sub instead of an add for the
1257 bool isNegative = isNonConstantNegative(Step);
1258 const SCEV *IncAmount = Step;
1260 IncAmount = Rewriter.SE.getNegativeSCEV(Step);
1262 // Insert an add instruction right before the terminator corresponding
1263 // to the back-edge or just before the only use. The location is determined
1264 // by the caller and passed in as IVIncInsertPt.
1265 Value *StepV = Rewriter.expandCodeFor(IncAmount, Ty,
1266 Preheader->getTerminator());
1269 IncV = BinaryOperator::CreateSub(PN, StepV, "lsr.iv.next",
1272 IncV = BinaryOperator::CreateAdd(PN, StepV, "lsr.iv.next",
1275 if (!isa<ConstantInt>(StepV)) ++NumVariable;
1277 PN->addIncoming(IncV, LatchBlock);
1283 static void SortUsersToProcess(std::vector<BasedUser> &UsersToProcess) {
1284 // We want to emit code for users inside the loop first. To do this, we
1285 // rearrange BasedUser so that the entries at the end have
1286 // isUseOfPostIncrementedValue = false, because we pop off the end of the
1287 // vector (so we handle them first).
1288 std::partition(UsersToProcess.begin(), UsersToProcess.end(),
1289 PartitionByIsUseOfPostIncrementedValue);
1291 // Sort this by base, so that things with the same base are handled
1292 // together. By partitioning first and stable-sorting later, we are
1293 // guaranteed that within each base we will pop off users from within the
1294 // loop before users outside of the loop with a particular base.
1296 // We would like to use stable_sort here, but we can't. The problem is that
1297 // const SCEV *'s don't have a deterministic ordering w.r.t to each other, so
1298 // we don't have anything to do a '<' comparison on. Because we think the
1299 // number of uses is small, do a horrible bubble sort which just relies on
1301 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1302 // Get a base value.
1303 const SCEV *Base = UsersToProcess[i].Base;
1305 // Compact everything with this base to be consecutive with this one.
1306 for (unsigned j = i+1; j != e; ++j) {
1307 if (UsersToProcess[j].Base == Base) {
1308 std::swap(UsersToProcess[i+1], UsersToProcess[j]);
1315 /// PrepareToStrengthReduceFully - Prepare to fully strength-reduce
1316 /// UsersToProcess, meaning lowering addresses all the way down to direct
1317 /// pointer arithmetic.
1320 LoopStrengthReduce::PrepareToStrengthReduceFully(
1321 std::vector<BasedUser> &UsersToProcess,
1323 const SCEV *CommonExprs,
1325 SCEVExpander &PreheaderRewriter) {
1326 DEBUG(errs() << " Fully reducing all users\n");
1328 // Rewrite the UsersToProcess records, creating a separate PHI for each
1329 // unique Base value.
1330 Instruction *IVIncInsertPt = L->getLoopLatch()->getTerminator();
1331 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ) {
1332 // TODO: The uses are grouped by base, but not sorted. We arbitrarily
1333 // pick the first Imm value here to start with, and adjust it for the
1335 const SCEV *Imm = UsersToProcess[i].Imm;
1336 const SCEV *Base = UsersToProcess[i].Base;
1337 const SCEV *Start = SE->getAddExpr(CommonExprs, Base, Imm);
1338 PHINode *Phi = InsertAffinePhi(Start, Stride, IVIncInsertPt, L,
1340 // Loop over all the users with the same base.
1342 UsersToProcess[i].Base = SE->getIntegerSCEV(0, Stride->getType());
1343 UsersToProcess[i].Imm = SE->getMinusSCEV(UsersToProcess[i].Imm, Imm);
1344 UsersToProcess[i].Phi = Phi;
1345 assert(UsersToProcess[i].Imm->isLoopInvariant(L) &&
1346 "ShouldUseFullStrengthReductionMode should reject this!");
1347 } while (++i != e && Base == UsersToProcess[i].Base);
1351 /// FindIVIncInsertPt - Return the location to insert the increment instruction.
1352 /// If the only use if a use of postinc value, (must be the loop termination
1353 /// condition), then insert it just before the use.
1354 static Instruction *FindIVIncInsertPt(std::vector<BasedUser> &UsersToProcess,
1356 if (UsersToProcess.size() == 1 &&
1357 UsersToProcess[0].isUseOfPostIncrementedValue &&
1358 L->contains(UsersToProcess[0].Inst->getParent()))
1359 return UsersToProcess[0].Inst;
1360 return L->getLoopLatch()->getTerminator();
1363 /// PrepareToStrengthReduceWithNewPhi - Insert a new induction variable for the
1364 /// given users to share.
1367 LoopStrengthReduce::PrepareToStrengthReduceWithNewPhi(
1368 std::vector<BasedUser> &UsersToProcess,
1370 const SCEV *CommonExprs,
1372 Instruction *IVIncInsertPt,
1374 SCEVExpander &PreheaderRewriter) {
1375 DEBUG(errs() << " Inserting new PHI:\n");
1377 PHINode *Phi = InsertAffinePhi(SE->getUnknown(CommonBaseV),
1378 Stride, IVIncInsertPt, L,
1381 // Remember this in case a later stride is multiple of this.
1382 IVsByStride[Stride].addIV(Stride, CommonExprs, Phi);
1384 // All the users will share this new IV.
1385 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1386 UsersToProcess[i].Phi = Phi;
1388 DEBUG(errs() << " IV=");
1389 DEBUG(WriteAsOperand(errs(), Phi, /*PrintType=*/false));
1390 DEBUG(errs() << "\n");
1393 /// PrepareToStrengthReduceFromSmallerStride - Prepare for the given users to
1394 /// reuse an induction variable with a stride that is a factor of the current
1395 /// induction variable.
1398 LoopStrengthReduce::PrepareToStrengthReduceFromSmallerStride(
1399 std::vector<BasedUser> &UsersToProcess,
1401 const IVExpr &ReuseIV,
1402 Instruction *PreInsertPt) {
1403 DEBUG(errs() << " Rewriting in terms of existing IV of STRIDE "
1404 << *ReuseIV.Stride << " and BASE " << *ReuseIV.Base << "\n");
1406 // All the users will share the reused IV.
1407 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1408 UsersToProcess[i].Phi = ReuseIV.PHI;
1410 Constant *C = dyn_cast<Constant>(CommonBaseV);
1412 (!C->isNullValue() &&
1413 !fitsInAddressMode(SE->getUnknown(CommonBaseV), CommonBaseV->getType(),
1415 // We want the common base emitted into the preheader! This is just
1416 // using cast as a copy so BitCast (no-op cast) is appropriate
1417 CommonBaseV = new BitCastInst(CommonBaseV, CommonBaseV->getType(),
1418 "commonbase", PreInsertPt);
1421 static bool IsImmFoldedIntoAddrMode(GlobalValue *GV, int64_t Offset,
1422 const Type *AccessTy,
1423 std::vector<BasedUser> &UsersToProcess,
1424 const TargetLowering *TLI) {
1425 SmallVector<Instruction*, 16> AddrModeInsts;
1426 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1427 if (UsersToProcess[i].isUseOfPostIncrementedValue)
1429 ExtAddrMode AddrMode =
1430 AddressingModeMatcher::Match(UsersToProcess[i].OperandValToReplace,
1431 AccessTy, UsersToProcess[i].Inst,
1432 AddrModeInsts, *TLI);
1433 if (GV && GV != AddrMode.BaseGV)
1435 if (Offset && !AddrMode.BaseOffs)
1436 // FIXME: How to accurate check it's immediate offset is folded.
1438 AddrModeInsts.clear();
1443 /// StrengthReduceIVUsersOfStride - Strength reduce all of the users of a single
1444 /// stride of IV. All of the users may have different starting values, and this
1445 /// may not be the only stride.
1447 LoopStrengthReduce::StrengthReduceIVUsersOfStride(const SCEV *const &Stride,
1448 IVUsersOfOneStride &Uses,
1450 // If all the users are moved to another stride, then there is nothing to do.
1451 if (Uses.Users.empty())
1454 // Keep track if every use in UsersToProcess is an address. If they all are,
1455 // we may be able to rewrite the entire collection of them in terms of a
1456 // smaller-stride IV.
1457 bool AllUsesAreAddresses = true;
1459 // Keep track if every use of a single stride is outside the loop. If so,
1460 // we want to be more aggressive about reusing a smaller-stride IV; a
1461 // multiply outside the loop is better than another IV inside. Well, usually.
1462 bool AllUsesAreOutsideLoop = true;
1464 // Transform our list of users and offsets to a bit more complex table. In
1465 // this new vector, each 'BasedUser' contains 'Base' the base of the
1466 // strided accessas well as the old information from Uses. We progressively
1467 // move information from the Base field to the Imm field, until we eventually
1468 // have the full access expression to rewrite the use.
1469 std::vector<BasedUser> UsersToProcess;
1470 const SCEV *CommonExprs = CollectIVUsers(Stride, Uses, L, AllUsesAreAddresses,
1471 AllUsesAreOutsideLoop,
1474 // Sort the UsersToProcess array so that users with common bases are
1475 // next to each other.
1476 SortUsersToProcess(UsersToProcess);
1478 // If we managed to find some expressions in common, we'll need to carry
1479 // their value in a register and add it in for each use. This will take up
1480 // a register operand, which potentially restricts what stride values are
1482 bool HaveCommonExprs = !CommonExprs->isZero();
1483 const Type *ReplacedTy = CommonExprs->getType();
1485 // If all uses are addresses, consider sinking the immediate part of the
1486 // common expression back into uses if they can fit in the immediate fields.
1487 if (TLI && HaveCommonExprs && AllUsesAreAddresses) {
1488 const SCEV *NewCommon = CommonExprs;
1489 const SCEV *Imm = SE->getIntegerSCEV(0, ReplacedTy);
1490 MoveImmediateValues(TLI, Type::getVoidTy(
1491 L->getLoopPreheader()->getContext()),
1492 NewCommon, Imm, true, L, SE);
1493 if (!Imm->isZero()) {
1496 // If the immediate part of the common expression is a GV, check if it's
1497 // possible to fold it into the target addressing mode.
1498 GlobalValue *GV = 0;
1499 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(Imm))
1500 GV = dyn_cast<GlobalValue>(SU->getValue());
1502 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Imm))
1503 Offset = SC->getValue()->getSExtValue();
1505 // Pass VoidTy as the AccessTy to be conservative, because
1506 // there could be multiple access types among all the uses.
1507 DoSink = IsImmFoldedIntoAddrMode(GV, Offset,
1508 Type::getVoidTy(L->getLoopPreheader()->getContext()),
1509 UsersToProcess, TLI);
1512 DEBUG(errs() << " Sinking " << *Imm << " back down into uses\n");
1513 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1514 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm, Imm);
1515 CommonExprs = NewCommon;
1516 HaveCommonExprs = !CommonExprs->isZero();
1522 // Now that we know what we need to do, insert the PHI node itself.
1524 DEBUG(errs() << "LSR: Examining IVs of TYPE " << *ReplacedTy << " of STRIDE "
1526 << " Common base: " << *CommonExprs << "\n");
1528 SCEVExpander Rewriter(*SE);
1529 SCEVExpander PreheaderRewriter(*SE);
1531 BasicBlock *Preheader = L->getLoopPreheader();
1532 Instruction *PreInsertPt = Preheader->getTerminator();
1533 BasicBlock *LatchBlock = L->getLoopLatch();
1534 Instruction *IVIncInsertPt = LatchBlock->getTerminator();
1536 Value *CommonBaseV = Constant::getNullValue(ReplacedTy);
1538 const SCEV *RewriteFactor = SE->getIntegerSCEV(0, ReplacedTy);
1539 IVExpr ReuseIV(SE->getIntegerSCEV(0,
1540 Type::getInt32Ty(Preheader->getContext())),
1541 SE->getIntegerSCEV(0,
1542 Type::getInt32Ty(Preheader->getContext())),
1545 // Choose a strength-reduction strategy and prepare for it by creating
1546 // the necessary PHIs and adjusting the bookkeeping.
1547 if (ShouldUseFullStrengthReductionMode(UsersToProcess, L,
1548 AllUsesAreAddresses, Stride)) {
1549 PrepareToStrengthReduceFully(UsersToProcess, Stride, CommonExprs, L,
1552 // Emit the initial base value into the loop preheader.
1553 CommonBaseV = PreheaderRewriter.expandCodeFor(CommonExprs, ReplacedTy,
1556 // If all uses are addresses, check if it is possible to reuse an IV. The
1557 // new IV must have a stride that is a multiple of the old stride; the
1558 // multiple must be a number that can be encoded in the scale field of the
1559 // target addressing mode; and we must have a valid instruction after this
1560 // substitution, including the immediate field, if any.
1561 RewriteFactor = CheckForIVReuse(HaveCommonExprs, AllUsesAreAddresses,
1562 AllUsesAreOutsideLoop,
1563 Stride, ReuseIV, ReplacedTy,
1565 if (!RewriteFactor->isZero())
1566 PrepareToStrengthReduceFromSmallerStride(UsersToProcess, CommonBaseV,
1567 ReuseIV, PreInsertPt);
1569 IVIncInsertPt = FindIVIncInsertPt(UsersToProcess, L);
1570 PrepareToStrengthReduceWithNewPhi(UsersToProcess, Stride, CommonExprs,
1571 CommonBaseV, IVIncInsertPt,
1572 L, PreheaderRewriter);
1576 // Process all the users now, replacing their strided uses with
1577 // strength-reduced forms. This outer loop handles all bases, the inner
1578 // loop handles all users of a particular base.
1579 while (!UsersToProcess.empty()) {
1580 const SCEV *Base = UsersToProcess.back().Base;
1581 Instruction *Inst = UsersToProcess.back().Inst;
1583 // Emit the code for Base into the preheader.
1585 if (!Base->isZero()) {
1586 BaseV = PreheaderRewriter.expandCodeFor(Base, 0, PreInsertPt);
1588 DEBUG(errs() << " INSERTING code for BASE = " << *Base << ":");
1589 if (BaseV->hasName())
1590 DEBUG(errs() << " Result value name = %" << BaseV->getName());
1591 DEBUG(errs() << "\n");
1593 // If BaseV is a non-zero constant, make sure that it gets inserted into
1594 // the preheader, instead of being forward substituted into the uses. We
1595 // do this by forcing a BitCast (noop cast) to be inserted into the
1596 // preheader in this case.
1597 if (!fitsInAddressMode(Base, getAccessType(Inst), TLI, false) &&
1598 isa<Constant>(BaseV)) {
1599 // We want this constant emitted into the preheader! This is just
1600 // using cast as a copy so BitCast (no-op cast) is appropriate
1601 BaseV = new BitCastInst(BaseV, BaseV->getType(), "preheaderinsert",
1606 // Emit the code to add the immediate offset to the Phi value, just before
1607 // the instructions that we identified as using this stride and base.
1609 // FIXME: Use emitted users to emit other users.
1610 BasedUser &User = UsersToProcess.back();
1612 DEBUG(errs() << " Examining ");
1613 if (User.isUseOfPostIncrementedValue)
1614 DEBUG(errs() << "postinc");
1616 DEBUG(errs() << "preinc");
1617 DEBUG(errs() << " use ");
1618 DEBUG(WriteAsOperand(errs(), UsersToProcess.back().OperandValToReplace,
1619 /*PrintType=*/false));
1620 DEBUG(errs() << " in Inst: " << *User.Inst);
1622 // If this instruction wants to use the post-incremented value, move it
1623 // after the post-inc and use its value instead of the PHI.
1624 Value *RewriteOp = User.Phi;
1625 if (User.isUseOfPostIncrementedValue) {
1626 RewriteOp = User.Phi->getIncomingValueForBlock(LatchBlock);
1627 // If this user is in the loop, make sure it is the last thing in the
1628 // loop to ensure it is dominated by the increment. In case it's the
1629 // only use of the iv, the increment instruction is already before the
1631 if (L->contains(User.Inst->getParent()) && User.Inst != IVIncInsertPt)
1632 User.Inst->moveBefore(IVIncInsertPt);
1635 const SCEV *RewriteExpr = SE->getUnknown(RewriteOp);
1637 if (SE->getEffectiveSCEVType(RewriteOp->getType()) !=
1638 SE->getEffectiveSCEVType(ReplacedTy)) {
1639 assert(SE->getTypeSizeInBits(RewriteOp->getType()) >
1640 SE->getTypeSizeInBits(ReplacedTy) &&
1641 "Unexpected widening cast!");
1642 RewriteExpr = SE->getTruncateExpr(RewriteExpr, ReplacedTy);
1645 // If we had to insert new instructions for RewriteOp, we have to
1646 // consider that they may not have been able to end up immediately
1647 // next to RewriteOp, because non-PHI instructions may never precede
1648 // PHI instructions in a block. In this case, remember where the last
1649 // instruction was inserted so that if we're replacing a different
1650 // PHI node, we can use the later point to expand the final
1652 Instruction *NewBasePt = dyn_cast<Instruction>(RewriteOp);
1653 if (RewriteOp == User.Phi) NewBasePt = 0;
1655 // Clear the SCEVExpander's expression map so that we are guaranteed
1656 // to have the code emitted where we expect it.
1659 // If we are reusing the iv, then it must be multiplied by a constant
1660 // factor to take advantage of the addressing mode scale component.
1661 if (!RewriteFactor->isZero()) {
1662 // If we're reusing an IV with a nonzero base (currently this happens
1663 // only when all reuses are outside the loop) subtract that base here.
1664 // The base has been used to initialize the PHI node but we don't want
1666 if (!ReuseIV.Base->isZero()) {
1667 const SCEV *typedBase = ReuseIV.Base;
1668 if (SE->getEffectiveSCEVType(RewriteExpr->getType()) !=
1669 SE->getEffectiveSCEVType(ReuseIV.Base->getType())) {
1670 // It's possible the original IV is a larger type than the new IV,
1671 // in which case we have to truncate the Base. We checked in
1672 // RequiresTypeConversion that this is valid.
1673 assert(SE->getTypeSizeInBits(RewriteExpr->getType()) <
1674 SE->getTypeSizeInBits(ReuseIV.Base->getType()) &&
1675 "Unexpected lengthening conversion!");
1676 typedBase = SE->getTruncateExpr(ReuseIV.Base,
1677 RewriteExpr->getType());
1679 RewriteExpr = SE->getMinusSCEV(RewriteExpr, typedBase);
1682 // Multiply old variable, with base removed, by new scale factor.
1683 RewriteExpr = SE->getMulExpr(RewriteFactor,
1686 // The common base is emitted in the loop preheader. But since we
1687 // are reusing an IV, it has not been used to initialize the PHI node.
1688 // Add it to the expression used to rewrite the uses.
1689 // When this use is outside the loop, we earlier subtracted the
1690 // common base, and are adding it back here. Use the same expression
1691 // as before, rather than CommonBaseV, so DAGCombiner will zap it.
1692 if (!CommonExprs->isZero()) {
1693 if (L->contains(User.Inst->getParent()))
1694 RewriteExpr = SE->getAddExpr(RewriteExpr,
1695 SE->getUnknown(CommonBaseV));
1697 RewriteExpr = SE->getAddExpr(RewriteExpr, CommonExprs);
1701 // Now that we know what we need to do, insert code before User for the
1702 // immediate and any loop-variant expressions.
1704 // Add BaseV to the PHI value if needed.
1705 RewriteExpr = SE->getAddExpr(RewriteExpr, SE->getUnknown(BaseV));
1707 User.RewriteInstructionToUseNewBase(RewriteExpr, NewBasePt,
1711 // Mark old value we replaced as possibly dead, so that it is eliminated
1712 // if we just replaced the last use of that value.
1713 DeadInsts.push_back(User.OperandValToReplace);
1715 UsersToProcess.pop_back();
1718 // If there are any more users to process with the same base, process them
1719 // now. We sorted by base above, so we just have to check the last elt.
1720 } while (!UsersToProcess.empty() && UsersToProcess.back().Base == Base);
1721 // TODO: Next, find out which base index is the most common, pull it out.
1724 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
1725 // different starting values, into different PHIs.
1728 void LoopStrengthReduce::StrengthReduceIVUsers(Loop *L) {
1729 // Note: this processes each stride/type pair individually. All users
1730 // passed into StrengthReduceIVUsersOfStride have the same type AND stride.
1731 // Also, note that we iterate over IVUsesByStride indirectly by using
1732 // StrideOrder. This extra layer of indirection makes the ordering of
1733 // strides deterministic - not dependent on map order.
1734 for (unsigned Stride = 0, e = IU->StrideOrder.size(); Stride != e; ++Stride) {
1735 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
1736 IU->IVUsesByStride.find(IU->StrideOrder[Stride]);
1737 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
1738 // FIXME: Generalize to non-affine IV's.
1739 if (!SI->first->isLoopInvariant(L))
1741 StrengthReduceIVUsersOfStride(SI->first, *SI->second, L);
1745 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1746 /// set the IV user and stride information and return true, otherwise return
1748 bool LoopStrengthReduce::FindIVUserForCond(ICmpInst *Cond,
1749 IVStrideUse *&CondUse,
1750 const SCEV* &CondStride) {
1751 for (unsigned Stride = 0, e = IU->StrideOrder.size();
1752 Stride != e && !CondUse; ++Stride) {
1753 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
1754 IU->IVUsesByStride.find(IU->StrideOrder[Stride]);
1755 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
1757 for (ilist<IVStrideUse>::iterator UI = SI->second->Users.begin(),
1758 E = SI->second->Users.end(); UI != E; ++UI)
1759 if (UI->getUser() == Cond) {
1760 // NOTE: we could handle setcc instructions with multiple uses here, but
1761 // InstCombine does it as well for simple uses, it's not clear that it
1762 // occurs enough in real life to handle.
1764 CondStride = SI->first;
1772 // Constant strides come first which in turns are sorted by their absolute
1773 // values. If absolute values are the same, then positive strides comes first.
1775 // 4, -1, X, 1, 2 ==> 1, -1, 2, 4, X
1776 struct StrideCompare {
1777 const ScalarEvolution *SE;
1778 explicit StrideCompare(const ScalarEvolution *se) : SE(se) {}
1780 bool operator()(const SCEV *const &LHS, const SCEV *const &RHS) {
1781 const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS);
1782 const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS);
1784 int64_t LV = LHSC->getValue()->getSExtValue();
1785 int64_t RV = RHSC->getValue()->getSExtValue();
1786 uint64_t ALV = (LV < 0) ? -LV : LV;
1787 uint64_t ARV = (RV < 0) ? -RV : RV;
1795 // If it's the same value but different type, sort by bit width so
1796 // that we emit larger induction variables before smaller
1797 // ones, letting the smaller be re-written in terms of larger ones.
1798 return SE->getTypeSizeInBits(RHS->getType()) <
1799 SE->getTypeSizeInBits(LHS->getType());
1801 return LHSC && !RHSC;
1806 /// ChangeCompareStride - If a loop termination compare instruction is the
1807 /// only use of its stride, and the compaison is against a constant value,
1808 /// try eliminate the stride by moving the compare instruction to another
1809 /// stride and change its constant operand accordingly. e.g.
1815 /// if (v2 < 10) goto loop
1820 /// if (v1 < 30) goto loop
1821 ICmpInst *LoopStrengthReduce::ChangeCompareStride(Loop *L, ICmpInst *Cond,
1822 IVStrideUse* &CondUse,
1823 const SCEV* &CondStride,
1825 // If there's only one stride in the loop, there's nothing to do here.
1826 if (IU->StrideOrder.size() < 2)
1828 // If there are other users of the condition's stride, don't bother
1829 // trying to change the condition because the stride will still
1831 std::map<const SCEV *, IVUsersOfOneStride *>::iterator I =
1832 IU->IVUsesByStride.find(CondStride);
1833 if (I == IU->IVUsesByStride.end())
1835 if (I->second->Users.size() > 1) {
1836 for (ilist<IVStrideUse>::iterator II = I->second->Users.begin(),
1837 EE = I->second->Users.end(); II != EE; ++II) {
1838 if (II->getUser() == Cond)
1840 if (!isInstructionTriviallyDead(II->getUser()))
1844 // Only handle constant strides for now.
1845 const SCEVConstant *SC = dyn_cast<SCEVConstant>(CondStride);
1846 if (!SC) return Cond;
1848 ICmpInst::Predicate Predicate = Cond->getPredicate();
1849 int64_t CmpSSInt = SC->getValue()->getSExtValue();
1850 unsigned BitWidth = SE->getTypeSizeInBits(CondStride->getType());
1851 uint64_t SignBit = 1ULL << (BitWidth-1);
1852 const Type *CmpTy = Cond->getOperand(0)->getType();
1853 const Type *NewCmpTy = NULL;
1854 unsigned TyBits = SE->getTypeSizeInBits(CmpTy);
1855 unsigned NewTyBits = 0;
1856 const SCEV *NewStride = NULL;
1857 Value *NewCmpLHS = NULL;
1858 Value *NewCmpRHS = NULL;
1860 const SCEV *NewOffset = SE->getIntegerSCEV(0, CmpTy);
1862 if (ConstantInt *C = dyn_cast<ConstantInt>(Cond->getOperand(1))) {
1863 int64_t CmpVal = C->getValue().getSExtValue();
1865 // Check the relevant induction variable for conformance to
1867 const SCEV *IV = SE->getSCEV(Cond->getOperand(0));
1868 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1869 if (!AR || !AR->isAffine())
1872 const SCEVConstant *StartC = dyn_cast<SCEVConstant>(AR->getStart());
1873 // Check stride constant and the comparision constant signs to detect
1876 if ((StartC->getValue()->getSExtValue() < CmpVal && CmpSSInt < 0) ||
1877 (StartC->getValue()->getSExtValue() > CmpVal && CmpSSInt > 0))
1880 // More restrictive check for the other cases.
1881 if ((CmpVal & SignBit) != (CmpSSInt & SignBit))
1885 // Look for a suitable stride / iv as replacement.
1886 for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
1887 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
1888 IU->IVUsesByStride.find(IU->StrideOrder[i]);
1889 if (!isa<SCEVConstant>(SI->first) || SI->second->Users.empty())
1891 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1892 if (SSInt == CmpSSInt ||
1893 abs64(SSInt) < abs64(CmpSSInt) ||
1894 (SSInt % CmpSSInt) != 0)
1897 Scale = SSInt / CmpSSInt;
1898 int64_t NewCmpVal = CmpVal * Scale;
1900 // If old icmp value fits in icmp immediate field, but the new one doesn't
1901 // try something else.
1903 TLI->isLegalICmpImmediate(CmpVal) &&
1904 !TLI->isLegalICmpImmediate(NewCmpVal))
1907 APInt Mul = APInt(BitWidth*2, CmpVal, true);
1908 Mul = Mul * APInt(BitWidth*2, Scale, true);
1909 // Check for overflow.
1910 if (!Mul.isSignedIntN(BitWidth))
1912 // Check for overflow in the stride's type too.
1913 if (!Mul.isSignedIntN(SE->getTypeSizeInBits(SI->first->getType())))
1916 // Watch out for overflow.
1917 if (ICmpInst::isSigned(Predicate) &&
1918 (CmpVal & SignBit) != (NewCmpVal & SignBit))
1921 // Pick the best iv to use trying to avoid a cast.
1923 for (ilist<IVStrideUse>::iterator UI = SI->second->Users.begin(),
1924 E = SI->second->Users.end(); UI != E; ++UI) {
1925 Value *Op = UI->getOperandValToReplace();
1927 // If the IVStrideUse implies a cast, check for an actual cast which
1928 // can be used to find the original IV expression.
1929 if (SE->getEffectiveSCEVType(Op->getType()) !=
1930 SE->getEffectiveSCEVType(SI->first->getType())) {
1931 CastInst *CI = dyn_cast<CastInst>(Op);
1932 // If it's not a simple cast, it's complicated.
1935 // If it's a cast from a type other than the stride type,
1936 // it's complicated.
1937 if (CI->getOperand(0)->getType() != SI->first->getType())
1939 // Ok, we found the IV expression in the stride's type.
1940 Op = CI->getOperand(0);
1944 if (NewCmpLHS->getType() == CmpTy)
1950 NewCmpTy = NewCmpLHS->getType();
1951 NewTyBits = SE->getTypeSizeInBits(NewCmpTy);
1952 const Type *NewCmpIntTy = IntegerType::get(Cond->getContext(), NewTyBits);
1953 if (RequiresTypeConversion(NewCmpTy, CmpTy)) {
1954 // Check if it is possible to rewrite it using
1955 // an iv / stride of a smaller integer type.
1956 unsigned Bits = NewTyBits;
1957 if (ICmpInst::isSigned(Predicate))
1959 uint64_t Mask = (1ULL << Bits) - 1;
1960 if (((uint64_t)NewCmpVal & Mask) != (uint64_t)NewCmpVal)
1964 // Don't rewrite if use offset is non-constant and the new type is
1965 // of a different type.
1966 // FIXME: too conservative?
1967 if (NewTyBits != TyBits && !isa<SCEVConstant>(CondUse->getOffset()))
1971 bool AllUsesAreAddresses = true;
1972 bool AllUsesAreOutsideLoop = true;
1973 std::vector<BasedUser> UsersToProcess;
1974 const SCEV *CommonExprs = CollectIVUsers(SI->first, *SI->second, L,
1975 AllUsesAreAddresses,
1976 AllUsesAreOutsideLoop,
1978 // Avoid rewriting the compare instruction with an iv of new stride
1979 // if it's likely the new stride uses will be rewritten using the
1980 // stride of the compare instruction.
1981 if (AllUsesAreAddresses &&
1982 ValidScale(!CommonExprs->isZero(), Scale, UsersToProcess))
1986 // Avoid rewriting the compare instruction with an iv which has
1987 // implicit extension or truncation built into it.
1988 // TODO: This is over-conservative.
1989 if (SE->getTypeSizeInBits(CondUse->getOffset()->getType()) != TyBits)
1992 // If scale is negative, use swapped predicate unless it's testing
1994 if (Scale < 0 && !Cond->isEquality())
1995 Predicate = ICmpInst::getSwappedPredicate(Predicate);
1997 NewStride = IU->StrideOrder[i];
1998 if (!isa<PointerType>(NewCmpTy))
1999 NewCmpRHS = ConstantInt::get(NewCmpTy, NewCmpVal);
2001 Constant *CI = ConstantInt::get(NewCmpIntTy, NewCmpVal);
2002 NewCmpRHS = ConstantExpr::getIntToPtr(CI, NewCmpTy);
2004 NewOffset = TyBits == NewTyBits
2005 ? SE->getMulExpr(CondUse->getOffset(),
2006 SE->getConstant(CmpTy, Scale))
2007 : SE->getConstant(NewCmpIntTy,
2008 cast<SCEVConstant>(CondUse->getOffset())->getValue()
2009 ->getSExtValue()*Scale);
2014 // Forgo this transformation if it the increment happens to be
2015 // unfortunately positioned after the condition, and the condition
2016 // has multiple uses which prevent it from being moved immediately
2017 // before the branch. See
2018 // test/Transforms/LoopStrengthReduce/change-compare-stride-trickiness-*.ll
2019 // for an example of this situation.
2020 if (!Cond->hasOneUse()) {
2021 for (BasicBlock::iterator I = Cond, E = Cond->getParent()->end();
2028 // Create a new compare instruction using new stride / iv.
2029 ICmpInst *OldCond = Cond;
2030 // Insert new compare instruction.
2031 Cond = new ICmpInst(OldCond, Predicate, NewCmpLHS, NewCmpRHS,
2032 L->getHeader()->getName() + ".termcond");
2034 DEBUG(errs() << " Change compare stride in Inst " << *OldCond);
2035 DEBUG(errs() << " to " << *Cond << '\n');
2037 // Remove the old compare instruction. The old indvar is probably dead too.
2038 DeadInsts.push_back(CondUse->getOperandValToReplace());
2039 OldCond->replaceAllUsesWith(Cond);
2040 OldCond->eraseFromParent();
2042 IU->IVUsesByStride[NewStride]->addUser(NewOffset, Cond, NewCmpLHS);
2043 CondUse = &IU->IVUsesByStride[NewStride]->Users.back();
2044 CondStride = NewStride;
2052 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
2053 /// a max computation.
2055 /// This is a narrow solution to a specific, but acute, problem. For loops
2061 /// } while (++i < n);
2063 /// the trip count isn't just 'n', because 'n' might not be positive. And
2064 /// unfortunately this can come up even for loops where the user didn't use
2065 /// a C do-while loop. For example, seemingly well-behaved top-test loops
2066 /// will commonly be lowered like this:
2072 /// } while (++i < n);
2075 /// and then it's possible for subsequent optimization to obscure the if
2076 /// test in such a way that indvars can't find it.
2078 /// When indvars can't find the if test in loops like this, it creates a
2079 /// max expression, which allows it to give the loop a canonical
2080 /// induction variable:
2083 /// max = n < 1 ? 1 : n;
2086 /// } while (++i != max);
2088 /// Canonical induction variables are necessary because the loop passes
2089 /// are designed around them. The most obvious example of this is the
2090 /// LoopInfo analysis, which doesn't remember trip count values. It
2091 /// expects to be able to rediscover the trip count each time it is
2092 /// needed, and it does this using a simple analyis that only succeeds if
2093 /// the loop has a canonical induction variable.
2095 /// However, when it comes time to generate code, the maximum operation
2096 /// can be quite costly, especially if it's inside of an outer loop.
2098 /// This function solves this problem by detecting this type of loop and
2099 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
2100 /// the instructions for the maximum computation.
2102 ICmpInst *LoopStrengthReduce::OptimizeMax(Loop *L, ICmpInst *Cond,
2103 IVStrideUse* &CondUse) {
2104 // Check that the loop matches the pattern we're looking for.
2105 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
2106 Cond->getPredicate() != CmpInst::ICMP_NE)
2109 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
2110 if (!Sel || !Sel->hasOneUse()) return Cond;
2112 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
2113 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2115 const SCEV *One = SE->getIntegerSCEV(1, BackedgeTakenCount->getType());
2117 // Add one to the backedge-taken count to get the trip count.
2118 const SCEV *IterationCount = SE->getAddExpr(BackedgeTakenCount, One);
2120 // Check for a max calculation that matches the pattern.
2121 if (!isa<SCEVSMaxExpr>(IterationCount) && !isa<SCEVUMaxExpr>(IterationCount))
2123 const SCEVNAryExpr *Max = cast<SCEVNAryExpr>(IterationCount);
2124 if (Max != SE->getSCEV(Sel)) return Cond;
2126 // To handle a max with more than two operands, this optimization would
2127 // require additional checking and setup.
2128 if (Max->getNumOperands() != 2)
2131 const SCEV *MaxLHS = Max->getOperand(0);
2132 const SCEV *MaxRHS = Max->getOperand(1);
2133 if (!MaxLHS || MaxLHS != One) return Cond;
2135 // Check the relevant induction variable for conformance to
2137 const SCEV *IV = SE->getSCEV(Cond->getOperand(0));
2138 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2139 if (!AR || !AR->isAffine() ||
2140 AR->getStart() != One ||
2141 AR->getStepRecurrence(*SE) != One)
2144 assert(AR->getLoop() == L &&
2145 "Loop condition operand is an addrec in a different loop!");
2147 // Check the right operand of the select, and remember it, as it will
2148 // be used in the new comparison instruction.
2150 if (SE->getSCEV(Sel->getOperand(1)) == MaxRHS)
2151 NewRHS = Sel->getOperand(1);
2152 else if (SE->getSCEV(Sel->getOperand(2)) == MaxRHS)
2153 NewRHS = Sel->getOperand(2);
2154 if (!NewRHS) return Cond;
2156 // Determine the new comparison opcode. It may be signed or unsigned,
2157 // and the original comparison may be either equality or inequality.
2158 CmpInst::Predicate Pred =
2159 isa<SCEVSMaxExpr>(Max) ? CmpInst::ICMP_SLT : CmpInst::ICMP_ULT;
2160 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
2161 Pred = CmpInst::getInversePredicate(Pred);
2163 // Ok, everything looks ok to change the condition into an SLT or SGE and
2164 // delete the max calculation.
2166 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
2168 // Delete the max calculation instructions.
2169 Cond->replaceAllUsesWith(NewCond);
2170 CondUse->setUser(NewCond);
2171 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2172 Cond->eraseFromParent();
2173 Sel->eraseFromParent();
2174 if (Cmp->use_empty())
2175 Cmp->eraseFromParent();
2179 /// OptimizeShadowIV - If IV is used in a int-to-float cast
2180 /// inside the loop then try to eliminate the cast opeation.
2181 void LoopStrengthReduce::OptimizeShadowIV(Loop *L) {
2183 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
2184 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2187 for (unsigned Stride = 0, e = IU->StrideOrder.size(); Stride != e;
2189 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
2190 IU->IVUsesByStride.find(IU->StrideOrder[Stride]);
2191 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
2192 if (!isa<SCEVConstant>(SI->first))
2195 for (ilist<IVStrideUse>::iterator UI = SI->second->Users.begin(),
2196 E = SI->second->Users.end(); UI != E; /* empty */) {
2197 ilist<IVStrideUse>::iterator CandidateUI = UI;
2199 Instruction *ShadowUse = CandidateUI->getUser();
2200 const Type *DestTy = NULL;
2202 /* If shadow use is a int->float cast then insert a second IV
2203 to eliminate this cast.
2205 for (unsigned i = 0; i < n; ++i)
2211 for (unsigned i = 0; i < n; ++i, ++d)
2214 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser()))
2215 DestTy = UCast->getDestTy();
2216 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser()))
2217 DestTy = SCast->getDestTy();
2218 if (!DestTy) continue;
2221 // If target does not support DestTy natively then do not apply
2222 // this transformation.
2223 EVT DVT = TLI->getValueType(DestTy);
2224 if (!TLI->isTypeLegal(DVT)) continue;
2227 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
2229 if (PH->getNumIncomingValues() != 2) continue;
2231 const Type *SrcTy = PH->getType();
2232 int Mantissa = DestTy->getFPMantissaWidth();
2233 if (Mantissa == -1) continue;
2234 if ((int)SE->getTypeSizeInBits(SrcTy) > Mantissa)
2237 unsigned Entry, Latch;
2238 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
2246 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
2247 if (!Init) continue;
2248 Constant *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());
2250 BinaryOperator *Incr =
2251 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
2252 if (!Incr) continue;
2253 if (Incr->getOpcode() != Instruction::Add
2254 && Incr->getOpcode() != Instruction::Sub)
2257 /* Initialize new IV, double d = 0.0 in above example. */
2258 ConstantInt *C = NULL;
2259 if (Incr->getOperand(0) == PH)
2260 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
2261 else if (Incr->getOperand(1) == PH)
2262 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
2268 // Ignore negative constants, as the code below doesn't handle them
2269 // correctly. TODO: Remove this restriction.
2270 if (!C->getValue().isStrictlyPositive()) continue;
2272 /* Add new PHINode. */
2273 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);
2275 /* create new increment. '++d' in above example. */
2276 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
2277 BinaryOperator *NewIncr =
2278 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
2279 Instruction::FAdd : Instruction::FSub,
2280 NewPH, CFP, "IV.S.next.", Incr);
2282 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
2283 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
2285 /* Remove cast operation */
2286 ShadowUse->replaceAllUsesWith(NewPH);
2287 ShadowUse->eraseFromParent();
2294 /// OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
2295 /// uses in the loop, look to see if we can eliminate some, in favor of using
2296 /// common indvars for the different uses.
2297 void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
2298 // TODO: implement optzns here.
2300 OptimizeShadowIV(L);
2303 bool LoopStrengthReduce::StrideMightBeShared(const SCEV* Stride, Loop *L,
2305 int64_t SInt = cast<SCEVConstant>(Stride)->getValue()->getSExtValue();
2306 for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
2307 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
2308 IU->IVUsesByStride.find(IU->StrideOrder[i]);
2309 const SCEV *Share = SI->first;
2310 if (!isa<SCEVConstant>(SI->first) || Share == Stride)
2312 int64_t SSInt = cast<SCEVConstant>(Share)->getValue()->getSExtValue();
2314 return true; // This can definitely be reused.
2315 if (unsigned(abs64(SSInt)) < SInt || (SSInt % SInt) != 0)
2317 int64_t Scale = SSInt / SInt;
2318 bool AllUsesAreAddresses = true;
2319 bool AllUsesAreOutsideLoop = true;
2320 std::vector<BasedUser> UsersToProcess;
2321 const SCEV *CommonExprs = CollectIVUsers(SI->first, *SI->second, L,
2322 AllUsesAreAddresses,
2323 AllUsesAreOutsideLoop,
2325 if (AllUsesAreAddresses &&
2326 ValidScale(!CommonExprs->isZero(), Scale, UsersToProcess)) {
2329 // Any pre-inc iv use?
2330 IVUsersOfOneStride &StrideUses = *IU->IVUsesByStride[Share];
2331 for (ilist<IVStrideUse>::iterator I = StrideUses.Users.begin(),
2332 E = StrideUses.Users.end(); I != E; ++I) {
2333 if (!I->isUseOfPostIncrementedValue())
2341 /// isUsedByExitBranch - Return true if icmp is used by a loop terminating
2342 /// conditional branch or it's and / or with other conditions before being used
2343 /// as the condition.
2344 static bool isUsedByExitBranch(ICmpInst *Cond, Loop *L) {
2345 BasicBlock *CondBB = Cond->getParent();
2346 if (!L->isLoopExiting(CondBB))
2348 BranchInst *TermBr = dyn_cast<BranchInst>(CondBB->getTerminator());
2349 if (!TermBr || !TermBr->isConditional())
2352 Value *User = *Cond->use_begin();
2353 Instruction *UserInst = dyn_cast<Instruction>(User);
2355 (UserInst->getOpcode() == Instruction::And ||
2356 UserInst->getOpcode() == Instruction::Or)) {
2357 if (!UserInst->hasOneUse() || UserInst->getParent() != CondBB)
2359 User = *User->use_begin();
2360 UserInst = dyn_cast<Instruction>(User);
2362 return User == TermBr;
2365 static bool ShouldCountToZero(ICmpInst *Cond, IVStrideUse* &CondUse,
2366 ScalarEvolution *SE, Loop *L,
2367 const TargetLowering *TLI = 0) {
2368 if (!L->contains(Cond->getParent()))
2371 if (!isa<SCEVConstant>(CondUse->getOffset()))
2374 // Handle only tests for equality for the moment.
2375 if (!Cond->isEquality() || !Cond->hasOneUse())
2377 if (!isUsedByExitBranch(Cond, L))
2380 Value *CondOp0 = Cond->getOperand(0);
2381 const SCEV *IV = SE->getSCEV(CondOp0);
2382 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2383 if (!AR || !AR->isAffine())
2386 const SCEVConstant *SC = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE));
2387 if (!SC || SC->getValue()->getSExtValue() < 0)
2388 // If it's already counting down, don't do anything.
2391 // If the RHS of the comparison is not an loop invariant, the rewrite
2392 // cannot be done. Also bail out if it's already comparing against a zero.
2393 // If we are checking this before cmp stride optimization, check if it's
2394 // comparing against a already legal immediate.
2395 Value *RHS = Cond->getOperand(1);
2396 ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS);
2397 if (!L->isLoopInvariant(RHS) ||
2398 (RHSC && RHSC->isZero()) ||
2399 (RHSC && TLI && TLI->isLegalICmpImmediate(RHSC->getSExtValue())))
2402 // Make sure the IV is only used for counting. Value may be preinc or
2403 // postinc; 2 uses in either case.
2404 if (!CondOp0->hasNUses(2))
2410 /// OptimizeLoopTermCond - Change loop terminating condition to use the
2411 /// postinc iv when possible.
2412 void LoopStrengthReduce::OptimizeLoopTermCond(Loop *L) {
2413 BasicBlock *LatchBlock = L->getLoopLatch();
2414 bool LatchExit = L->isLoopExiting(LatchBlock);
2415 SmallVector<BasicBlock*, 8> ExitingBlocks;
2416 L->getExitingBlocks(ExitingBlocks);
2418 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
2419 BasicBlock *ExitingBlock = ExitingBlocks[i];
2421 // Finally, get the terminating condition for the loop if possible. If we
2422 // can, we want to change it to use a post-incremented version of its
2423 // induction variable, to allow coalescing the live ranges for the IV into
2424 // one register value.
2426 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2429 // FIXME: Overly conservative, termination condition could be an 'or' etc..
2430 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
2433 // Search IVUsesByStride to find Cond's IVUse if there is one.
2434 IVStrideUse *CondUse = 0;
2435 const SCEV *CondStride = 0;
2436 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2437 if (!FindIVUserForCond(Cond, CondUse, CondStride))
2440 // If the latch block is exiting and it's not a single block loop, it's
2441 // not safe to use postinc iv in other exiting blocks. FIXME: overly
2442 // conservative? How about icmp stride optimization?
2443 bool UsePostInc = !(e > 1 && LatchExit && ExitingBlock != LatchBlock);
2444 if (UsePostInc && ExitingBlock != LatchBlock) {
2445 if (!Cond->hasOneUse())
2446 // See below, we don't want the condition to be cloned.
2449 // If exiting block is the latch block, we know it's safe and profitable
2450 // to transform the icmp to use post-inc iv. Otherwise do so only if it
2451 // would not reuse another iv and its iv would be reused by other uses.
2452 // We are optimizing for the case where the icmp is the only use of the
2454 IVUsersOfOneStride &StrideUses = *IU->IVUsesByStride[CondStride];
2455 for (ilist<IVStrideUse>::iterator I = StrideUses.Users.begin(),
2456 E = StrideUses.Users.end(); I != E; ++I) {
2457 if (I->getUser() == Cond)
2459 if (!I->isUseOfPostIncrementedValue()) {
2466 // If iv for the stride might be shared and any of the users use pre-inc
2467 // iv might be used, then it's not safe to use post-inc iv.
2469 isa<SCEVConstant>(CondStride) &&
2470 StrideMightBeShared(CondStride, L, true))
2474 // If the trip count is computed in terms of a max (due to ScalarEvolution
2475 // being unable to find a sufficient guard, for example), change the loop
2476 // comparison to use SLT or ULT instead of NE.
2477 Cond = OptimizeMax(L, Cond, CondUse);
2479 // If possible, change stride and operands of the compare instruction to
2480 // eliminate one stride. However, avoid rewriting the compare instruction
2481 // with an iv of new stride if it's likely the new stride uses will be
2482 // rewritten using the stride of the compare instruction.
2483 if (ExitingBlock == LatchBlock && isa<SCEVConstant>(CondStride)) {
2484 // If the condition stride is a constant and it's the only use, we might
2485 // want to optimize it first by turning it to count toward zero.
2486 if (!StrideMightBeShared(CondStride, L, false) &&
2487 !ShouldCountToZero(Cond, CondUse, SE, L, TLI))
2488 Cond = ChangeCompareStride(L, Cond, CondUse, CondStride);
2494 DEBUG(errs() << " Change loop exiting icmp to use postinc iv: "
2497 // It's possible for the setcc instruction to be anywhere in the loop, and
2498 // possible for it to have multiple users. If it is not immediately before
2499 // the exiting block branch, move it.
2500 if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
2501 if (Cond->hasOneUse()) { // Condition has a single use, just move it.
2502 Cond->moveBefore(TermBr);
2504 // Otherwise, clone the terminating condition and insert into the
2506 Cond = cast<ICmpInst>(Cond->clone());
2507 Cond->setName(L->getHeader()->getName() + ".termcond");
2508 ExitingBlock->getInstList().insert(TermBr, Cond);
2510 // Clone the IVUse, as the old use still exists!
2511 IU->IVUsesByStride[CondStride]->addUser(CondUse->getOffset(), Cond,
2512 CondUse->getOperandValToReplace());
2513 CondUse = &IU->IVUsesByStride[CondStride]->Users.back();
2517 // If we get to here, we know that we can transform the setcc instruction to
2518 // use the post-incremented version of the IV, allowing us to coalesce the
2519 // live ranges for the IV correctly.
2520 CondUse->setOffset(SE->getMinusSCEV(CondUse->getOffset(), CondStride));
2521 CondUse->setIsUseOfPostIncrementedValue(true);
2528 bool LoopStrengthReduce::OptimizeLoopCountIVOfStride(const SCEV* &Stride,
2529 IVStrideUse* &CondUse,
2531 // If the only use is an icmp of a loop exiting conditional branch, then
2532 // attempt the optimization.
2533 BasedUser User = BasedUser(*CondUse, SE);
2534 assert(isa<ICmpInst>(User.Inst) && "Expecting an ICMPInst!");
2535 ICmpInst *Cond = cast<ICmpInst>(User.Inst);
2537 // Less strict check now that compare stride optimization is done.
2538 if (!ShouldCountToZero(Cond, CondUse, SE, L))
2541 Value *CondOp0 = Cond->getOperand(0);
2542 PHINode *PHIExpr = dyn_cast<PHINode>(CondOp0);
2545 // Value tested is postinc. Find the phi node.
2546 Incr = dyn_cast<BinaryOperator>(CondOp0);
2547 // FIXME: Just use User.OperandValToReplace here?
2548 if (!Incr || Incr->getOpcode() != Instruction::Add)
2551 PHIExpr = dyn_cast<PHINode>(Incr->getOperand(0));
2554 // 1 use for preinc value, the increment.
2555 if (!PHIExpr->hasOneUse())
2558 assert(isa<PHINode>(CondOp0) &&
2559 "Unexpected loop exiting counting instruction sequence!");
2560 PHIExpr = cast<PHINode>(CondOp0);
2561 // Value tested is preinc. Find the increment.
2562 // A CmpInst is not a BinaryOperator; we depend on this.
2563 Instruction::use_iterator UI = PHIExpr->use_begin();
2564 Incr = dyn_cast<BinaryOperator>(UI);
2566 Incr = dyn_cast<BinaryOperator>(++UI);
2567 // One use for postinc value, the phi. Unnecessarily conservative?
2568 if (!Incr || !Incr->hasOneUse() || Incr->getOpcode() != Instruction::Add)
2572 // Replace the increment with a decrement.
2573 DEBUG(errs() << "LSR: Examining use ");
2574 DEBUG(WriteAsOperand(errs(), CondOp0, /*PrintType=*/false));
2575 DEBUG(errs() << " in Inst: " << *Cond << '\n');
2576 BinaryOperator *Decr = BinaryOperator::Create(Instruction::Sub,
2577 Incr->getOperand(0), Incr->getOperand(1), "tmp", Incr);
2578 Incr->replaceAllUsesWith(Decr);
2579 Incr->eraseFromParent();
2581 // Substitute endval-startval for the original startval, and 0 for the
2582 // original endval. Since we're only testing for equality this is OK even
2583 // if the computation wraps around.
2584 BasicBlock *Preheader = L->getLoopPreheader();
2585 Instruction *PreInsertPt = Preheader->getTerminator();
2586 unsigned InBlock = L->contains(PHIExpr->getIncomingBlock(0)) ? 1 : 0;
2587 Value *StartVal = PHIExpr->getIncomingValue(InBlock);
2588 Value *EndVal = Cond->getOperand(1);
2589 DEBUG(errs() << " Optimize loop counting iv to count down ["
2590 << *EndVal << " .. " << *StartVal << "]\n");
2592 // FIXME: check for case where both are constant.
2593 Constant* Zero = ConstantInt::get(Cond->getOperand(1)->getType(), 0);
2594 BinaryOperator *NewStartVal = BinaryOperator::Create(Instruction::Sub,
2595 EndVal, StartVal, "tmp", PreInsertPt);
2596 PHIExpr->setIncomingValue(InBlock, NewStartVal);
2597 Cond->setOperand(1, Zero);
2598 DEBUG(errs() << " New icmp: " << *Cond << "\n");
2600 int64_t SInt = cast<SCEVConstant>(Stride)->getValue()->getSExtValue();
2601 const SCEV *NewStride = 0;
2603 for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
2604 const SCEV *OldStride = IU->StrideOrder[i];
2605 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(OldStride))
2606 if (SC->getValue()->getSExtValue() == -SInt) {
2608 NewStride = OldStride;
2614 NewStride = SE->getIntegerSCEV(-SInt, Stride->getType());
2615 IU->AddUser(NewStride, CondUse->getOffset(), Cond, Cond->getOperand(0));
2616 IU->IVUsesByStride[Stride]->removeUser(CondUse);
2618 CondUse = &IU->IVUsesByStride[NewStride]->Users.back();
2626 /// OptimizeLoopCountIV - If, after all sharing of IVs, the IV used for deciding
2627 /// when to exit the loop is used only for that purpose, try to rearrange things
2628 /// so it counts down to a test against zero.
2629 bool LoopStrengthReduce::OptimizeLoopCountIV(Loop *L) {
2630 bool ThisChanged = false;
2631 for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
2632 const SCEV *Stride = IU->StrideOrder[i];
2633 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
2634 IU->IVUsesByStride.find(Stride);
2635 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
2636 // FIXME: Generalize to non-affine IV's.
2637 if (!SI->first->isLoopInvariant(L))
2639 // If stride is a constant and it has an icmpinst use, check if we can
2640 // optimize the loop to count down.
2641 if (isa<SCEVConstant>(Stride) && SI->second->Users.size() == 1) {
2642 Instruction *User = SI->second->Users.begin()->getUser();
2643 if (!isa<ICmpInst>(User))
2645 const SCEV *CondStride = Stride;
2646 IVStrideUse *Use = &*SI->second->Users.begin();
2647 if (!OptimizeLoopCountIVOfStride(CondStride, Use, L))
2651 // Now check if it's possible to reuse this iv for other stride uses.
2652 for (unsigned j = 0, ee = IU->StrideOrder.size(); j != ee; ++j) {
2653 const SCEV *SStride = IU->StrideOrder[j];
2654 if (SStride == CondStride)
2656 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SII =
2657 IU->IVUsesByStride.find(SStride);
2658 assert(SII != IU->IVUsesByStride.end() && "Stride doesn't exist!");
2659 // FIXME: Generalize to non-affine IV's.
2660 if (!SII->first->isLoopInvariant(L))
2662 // FIXME: Rewrite other stride using CondStride.
2667 Changed |= ThisChanged;
2671 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager &LPM) {
2672 IU = &getAnalysis<IVUsers>();
2673 SE = &getAnalysis<ScalarEvolution>();
2676 // If LoopSimplify form is not available, stay out of trouble.
2677 if (!L->getLoopPreheader() || !L->getLoopLatch())
2680 if (!IU->IVUsesByStride.empty()) {
2681 DEBUG(errs() << "\nLSR on \"" << L->getHeader()->getParent()->getName()
2685 // Sort the StrideOrder so we process larger strides first.
2686 std::stable_sort(IU->StrideOrder.begin(), IU->StrideOrder.end(),
2689 // Optimize induction variables. Some indvar uses can be transformed to use
2690 // strides that will be needed for other purposes. A common example of this
2691 // is the exit test for the loop, which can often be rewritten to use the
2692 // computation of some other indvar to decide when to terminate the loop.
2695 // Change loop terminating condition to use the postinc iv when possible
2696 // and optimize loop terminating compare. FIXME: Move this after
2697 // StrengthReduceIVUsersOfStride?
2698 OptimizeLoopTermCond(L);
2700 // FIXME: We can shrink overlarge IV's here. e.g. if the code has
2701 // computation in i64 values and the target doesn't support i64, demote
2702 // the computation to 32-bit if safe.
2704 // FIXME: Attempt to reuse values across multiple IV's. In particular, we
2705 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should
2706 // be codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC.
2707 // Need to be careful that IV's are all the same type. Only works for
2708 // intptr_t indvars.
2710 // IVsByStride keeps IVs for one particular loop.
2711 assert(IVsByStride.empty() && "Stale entries in IVsByStride?");
2713 StrengthReduceIVUsers(L);
2715 // After all sharing is done, see if we can adjust the loop to test against
2716 // zero instead of counting up to a maximum. This is usually faster.
2717 OptimizeLoopCountIV(L);
2719 // We're done analyzing this loop; release all the state we built up for it.
2720 IVsByStride.clear();
2722 // Clean up after ourselves
2723 if (!DeadInsts.empty())
2724 DeleteTriviallyDeadInstructions();
2727 // At this point, it is worth checking to see if any recurrence PHIs are also
2728 // dead, so that we can remove them as well.
2729 DeleteDeadPHIs(L->getHeader());