1 //===- LoopStrengthReduce.cpp - Strength Reduce IVs in Loops --------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This transformation analyzes and transforms the induction variables (and
11 // computations derived from them) into forms suitable for efficient execution
14 // This pass performs a strength reduction on array references inside loops that
15 // have as one or more of their components the loop induction variable, it
16 // rewrites expressions to take advantage of scaled-index addressing modes
17 // available on the target, and it performs a variety of other optimizations
18 // related to loop induction variables.
20 //===----------------------------------------------------------------------===//
22 #define DEBUG_TYPE "loop-reduce"
23 #include "llvm/Transforms/Scalar.h"
24 #include "llvm/Constants.h"
25 #include "llvm/Instructions.h"
26 #include "llvm/IntrinsicInst.h"
27 #include "llvm/Type.h"
28 #include "llvm/DerivedTypes.h"
29 #include "llvm/Analysis/Dominators.h"
30 #include "llvm/Analysis/IVUsers.h"
31 #include "llvm/Analysis/LoopInfo.h"
32 #include "llvm/Analysis/LoopPass.h"
33 #include "llvm/Analysis/ScalarEvolutionExpander.h"
34 #include "llvm/Transforms/Utils/AddrModeMatcher.h"
35 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
36 #include "llvm/Transforms/Utils/Local.h"
37 #include "llvm/ADT/Statistic.h"
38 #include "llvm/Support/CFG.h"
39 #include "llvm/Support/Debug.h"
40 #include "llvm/Support/CommandLine.h"
41 #include "llvm/Support/ValueHandle.h"
42 #include "llvm/Support/raw_ostream.h"
43 #include "llvm/Target/TargetLowering.h"
47 STATISTIC(NumReduced , "Number of IV uses strength reduced");
48 STATISTIC(NumInserted, "Number of PHIs inserted");
49 STATISTIC(NumVariable, "Number of PHIs with variable strides");
50 STATISTIC(NumEliminated, "Number of strides eliminated");
51 STATISTIC(NumShadow, "Number of Shadow IVs optimized");
52 STATISTIC(NumImmSunk, "Number of common expr immediates sunk into uses");
53 STATISTIC(NumLoopCond, "Number of loop terminating conds optimized");
54 STATISTIC(NumCountZero, "Number of count iv optimized to count toward zero");
56 static cl::opt<bool> EnableFullLSRMode("enable-full-lsr",
64 /// IVInfo - This structure keeps track of one IV expression inserted during
65 /// StrengthReduceStridedIVUsers. It contains the stride, the common base, as
66 /// well as the PHI node and increment value created for rewrite.
72 IVExpr(const SCEV *const stride, const SCEV *const base, PHINode *phi)
73 : Stride(stride), Base(base), PHI(phi) {}
76 /// IVsOfOneStride - This structure keeps track of all IV expression inserted
77 /// during StrengthReduceStridedIVUsers for a particular stride of the IV.
78 struct IVsOfOneStride {
79 std::vector<IVExpr> IVs;
81 void addIV(const SCEV *const Stride, const SCEV *const Base, PHINode *PHI) {
82 IVs.push_back(IVExpr(Stride, Base, PHI));
86 class LoopStrengthReduce : public LoopPass {
93 /// IVsByStride - Keep track of all IVs that have been inserted for a
94 /// particular stride.
95 std::map<const SCEV *, IVsOfOneStride> IVsByStride;
97 /// StrideNoReuse - Keep track of all the strides whose ivs cannot be
98 /// reused (nor should they be rewritten to reuse other strides).
99 SmallSet<const SCEV *, 4> StrideNoReuse;
101 /// DeadInsts - Keep track of instructions we may have made dead, so that
102 /// we can remove them after we are done working.
103 SmallVector<WeakVH, 16> DeadInsts;
105 /// TLI - Keep a pointer of a TargetLowering to consult for determining
106 /// transformation profitability.
107 const TargetLowering *TLI;
110 static char ID; // Pass ID, replacement for typeid
111 explicit LoopStrengthReduce(const TargetLowering *tli = NULL) :
112 LoopPass(&ID), TLI(tli) {
115 bool runOnLoop(Loop *L, LPPassManager &LPM);
117 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
118 // We split critical edges, so we change the CFG. However, we do update
119 // many analyses if they are around.
120 AU.addPreservedID(LoopSimplifyID);
121 AU.addPreserved<LoopInfo>();
122 AU.addPreserved<DominanceFrontier>();
123 AU.addPreserved<DominatorTree>();
125 AU.addRequiredID(LoopSimplifyID);
126 AU.addRequired<LoopInfo>();
127 AU.addRequired<DominatorTree>();
128 AU.addRequired<ScalarEvolution>();
129 AU.addPreserved<ScalarEvolution>();
130 AU.addRequired<IVUsers>();
131 AU.addPreserved<IVUsers>();
135 void OptimizeIndvars(Loop *L);
137 /// OptimizeLoopTermCond - Change loop terminating condition to use the
138 /// postinc iv when possible.
139 void OptimizeLoopTermCond(Loop *L);
141 /// OptimizeShadowIV - If IV is used in a int-to-float cast
142 /// inside the loop then try to eliminate the cast opeation.
143 void OptimizeShadowIV(Loop *L);
145 /// OptimizeMax - Rewrite the loop's terminating condition
146 /// if it uses a max computation.
147 ICmpInst *OptimizeMax(Loop *L, ICmpInst *Cond,
148 IVStrideUse* &CondUse);
150 /// OptimizeLoopCountIV - If, after all sharing of IVs, the IV used for
151 /// deciding when to exit the loop is used only for that purpose, try to
152 /// rearrange things so it counts down to a test against zero.
153 bool OptimizeLoopCountIV(Loop *L);
154 bool OptimizeLoopCountIVOfStride(const SCEV* &Stride,
155 IVStrideUse* &CondUse, Loop *L);
157 /// StrengthReduceIVUsersOfStride - Strength reduce all of the users of a
158 /// single stride of IV. All of the users may have different starting
159 /// values, and this may not be the only stride.
160 void StrengthReduceIVUsersOfStride(const SCEV *const &Stride,
161 IVUsersOfOneStride &Uses,
163 void StrengthReduceIVUsers(Loop *L);
165 ICmpInst *ChangeCompareStride(Loop *L, ICmpInst *Cond,
166 IVStrideUse* &CondUse,
167 const SCEV* &CondStride,
168 bool PostPass = false);
170 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse,
171 const SCEV* &CondStride);
172 bool RequiresTypeConversion(const Type *Ty, const Type *NewTy);
173 const SCEV *CheckForIVReuse(bool, bool, bool, const SCEV *const&,
174 IVExpr&, const Type*,
175 const std::vector<BasedUser>& UsersToProcess);
176 bool ValidScale(bool, int64_t,
177 const std::vector<BasedUser>& UsersToProcess);
178 bool ValidOffset(bool, int64_t, int64_t,
179 const std::vector<BasedUser>& UsersToProcess);
180 const SCEV *CollectIVUsers(const SCEV *const &Stride,
181 IVUsersOfOneStride &Uses,
183 bool &AllUsesAreAddresses,
184 bool &AllUsesAreOutsideLoop,
185 std::vector<BasedUser> &UsersToProcess);
186 bool StrideMightBeShared(const SCEV *Stride, Loop *L, bool CheckPreInc);
187 bool ShouldUseFullStrengthReductionMode(
188 const std::vector<BasedUser> &UsersToProcess,
190 bool AllUsesAreAddresses,
192 void PrepareToStrengthReduceFully(
193 std::vector<BasedUser> &UsersToProcess,
195 const SCEV *CommonExprs,
197 SCEVExpander &PreheaderRewriter);
198 void PrepareToStrengthReduceFromSmallerStride(
199 std::vector<BasedUser> &UsersToProcess,
201 const IVExpr &ReuseIV,
202 Instruction *PreInsertPt);
203 void PrepareToStrengthReduceWithNewPhi(
204 std::vector<BasedUser> &UsersToProcess,
206 const SCEV *CommonExprs,
208 Instruction *IVIncInsertPt,
210 SCEVExpander &PreheaderRewriter);
212 void DeleteTriviallyDeadInstructions();
216 char LoopStrengthReduce::ID = 0;
217 static RegisterPass<LoopStrengthReduce>
218 X("loop-reduce", "Loop Strength Reduction");
220 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
221 return new LoopStrengthReduce(TLI);
224 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
225 /// specified set are trivially dead, delete them and see if this makes any of
226 /// their operands subsequently dead.
227 void LoopStrengthReduce::DeleteTriviallyDeadInstructions() {
228 if (DeadInsts.empty()) return;
230 while (!DeadInsts.empty()) {
231 Instruction *I = dyn_cast_or_null<Instruction>(DeadInsts.back());
232 DeadInsts.pop_back();
234 if (I == 0 || !isInstructionTriviallyDead(I))
237 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) {
238 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
241 DeadInsts.push_back(U);
245 I->eraseFromParent();
250 /// isAddressUse - Returns true if the specified instruction is using the
251 /// specified value as an address.
252 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
253 bool isAddress = isa<LoadInst>(Inst);
254 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
255 if (SI->getOperand(1) == OperandVal)
257 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
258 // Addressing modes can also be folded into prefetches and a variety
260 switch (II->getIntrinsicID()) {
262 case Intrinsic::prefetch:
263 case Intrinsic::x86_sse2_loadu_dq:
264 case Intrinsic::x86_sse2_loadu_pd:
265 case Intrinsic::x86_sse_loadu_ps:
266 case Intrinsic::x86_sse_storeu_ps:
267 case Intrinsic::x86_sse2_storeu_pd:
268 case Intrinsic::x86_sse2_storeu_dq:
269 case Intrinsic::x86_sse2_storel_dq:
270 if (II->getOperand(1) == OperandVal)
278 /// getAccessType - Return the type of the memory being accessed.
279 static const Type *getAccessType(const Instruction *Inst) {
280 const Type *AccessTy = Inst->getType();
281 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
282 AccessTy = SI->getOperand(0)->getType();
283 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
284 // Addressing modes can also be folded into prefetches and a variety
286 switch (II->getIntrinsicID()) {
288 case Intrinsic::x86_sse_storeu_ps:
289 case Intrinsic::x86_sse2_storeu_pd:
290 case Intrinsic::x86_sse2_storeu_dq:
291 case Intrinsic::x86_sse2_storel_dq:
292 AccessTy = II->getOperand(1)->getType();
300 /// BasedUser - For a particular base value, keep information about how we've
301 /// partitioned the expression so far.
303 /// SE - The current ScalarEvolution object.
306 /// Base - The Base value for the PHI node that needs to be inserted for
307 /// this use. As the use is processed, information gets moved from this
308 /// field to the Imm field (below). BasedUser values are sorted by this
312 /// Inst - The instruction using the induction variable.
315 /// OperandValToReplace - The operand value of Inst to replace with the
317 Value *OperandValToReplace;
319 /// Imm - The immediate value that should be added to the base immediately
320 /// before Inst, because it will be folded into the imm field of the
321 /// instruction. This is also sometimes used for loop-variant values that
322 /// must be added inside the loop.
325 /// Phi - The induction variable that performs the striding that
326 /// should be used for this user.
329 // isUseOfPostIncrementedValue - True if this should use the
330 // post-incremented version of this IV, not the preincremented version.
331 // This can only be set in special cases, such as the terminating setcc
332 // instruction for a loop and uses outside the loop that are dominated by
334 bool isUseOfPostIncrementedValue;
336 BasedUser(IVStrideUse &IVSU, ScalarEvolution *se)
337 : SE(se), Base(IVSU.getOffset()), Inst(IVSU.getUser()),
338 OperandValToReplace(IVSU.getOperandValToReplace()),
339 Imm(SE->getIntegerSCEV(0, Base->getType())),
340 isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue()) {}
342 // Once we rewrite the code to insert the new IVs we want, update the
343 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
345 void RewriteInstructionToUseNewBase(const SCEV *const &NewBase,
346 Instruction *InsertPt,
347 SCEVExpander &Rewriter, Loop *L, Pass *P,
349 SmallVectorImpl<WeakVH> &DeadInsts);
351 Value *InsertCodeForBaseAtPosition(const SCEV *const &NewBase,
353 SCEVExpander &Rewriter,
354 Instruction *IP, Loop *L,
360 void BasedUser::dump() const {
361 errs() << " Base=" << *Base;
362 errs() << " Imm=" << *Imm;
363 errs() << " Inst: " << *Inst;
366 Value *BasedUser::InsertCodeForBaseAtPosition(const SCEV *const &NewBase,
368 SCEVExpander &Rewriter,
369 Instruction *IP, Loop *L,
371 // Figure out where we *really* want to insert this code. In particular, if
372 // the user is inside of a loop that is nested inside of L, we really don't
373 // want to insert this expression before the user, we'd rather pull it out as
374 // many loops as possible.
375 Instruction *BaseInsertPt = IP;
377 // Figure out the most-nested loop that IP is in.
378 Loop *InsertLoop = LI.getLoopFor(IP->getParent());
380 // If InsertLoop is not L, and InsertLoop is nested inside of L, figure out
381 // the preheader of the outer-most loop where NewBase is not loop invariant.
382 if (L->contains(IP->getParent()))
383 while (InsertLoop && NewBase->isLoopInvariant(InsertLoop)) {
384 BaseInsertPt = InsertLoop->getLoopPreheader()->getTerminator();
385 InsertLoop = InsertLoop->getParentLoop();
388 Value *Base = Rewriter.expandCodeFor(NewBase, 0, BaseInsertPt);
390 const SCEV *NewValSCEV = SE->getUnknown(Base);
392 // Always emit the immediate into the same block as the user.
393 NewValSCEV = SE->getAddExpr(NewValSCEV, Imm);
395 return Rewriter.expandCodeFor(NewValSCEV, Ty, IP);
399 // Once we rewrite the code to insert the new IVs we want, update the
400 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
401 // to it. NewBasePt is the last instruction which contributes to the
402 // value of NewBase in the case that it's a diffferent instruction from
403 // the PHI that NewBase is computed from, or null otherwise.
405 void BasedUser::RewriteInstructionToUseNewBase(const SCEV *const &NewBase,
406 Instruction *NewBasePt,
407 SCEVExpander &Rewriter, Loop *L, Pass *P,
409 SmallVectorImpl<WeakVH> &DeadInsts) {
410 if (!isa<PHINode>(Inst)) {
411 // By default, insert code at the user instruction.
412 BasicBlock::iterator InsertPt = Inst;
414 // However, if the Operand is itself an instruction, the (potentially
415 // complex) inserted code may be shared by many users. Because of this, we
416 // want to emit code for the computation of the operand right before its old
417 // computation. This is usually safe, because we obviously used to use the
418 // computation when it was computed in its current block. However, in some
419 // cases (e.g. use of a post-incremented induction variable) the NewBase
420 // value will be pinned to live somewhere after the original computation.
421 // In this case, we have to back off.
423 // If this is a use outside the loop (which means after, since it is based
424 // on a loop indvar) we use the post-incremented value, so that we don't
425 // artificially make the preinc value live out the bottom of the loop.
426 if (!isUseOfPostIncrementedValue && L->contains(Inst->getParent())) {
427 if (NewBasePt && isa<PHINode>(OperandValToReplace)) {
428 InsertPt = NewBasePt;
430 } else if (Instruction *OpInst
431 = dyn_cast<Instruction>(OperandValToReplace)) {
433 while (isa<PHINode>(InsertPt)) ++InsertPt;
436 Value *NewVal = InsertCodeForBaseAtPosition(NewBase,
437 OperandValToReplace->getType(),
438 Rewriter, InsertPt, L, LI);
439 // Replace the use of the operand Value with the new Phi we just created.
440 Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
442 DEBUG(errs() << " Replacing with ");
443 DEBUG(WriteAsOperand(errs(), NewVal, /*PrintType=*/false));
444 DEBUG(errs() << ", which has value " << *NewBase << " plus IMM "
449 // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm
450 // expression into each operand block that uses it. Note that PHI nodes can
451 // have multiple entries for the same predecessor. We use a map to make sure
452 // that a PHI node only has a single Value* for each predecessor (which also
453 // prevents us from inserting duplicate code in some blocks).
454 DenseMap<BasicBlock*, Value*> InsertedCode;
455 PHINode *PN = cast<PHINode>(Inst);
456 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
457 if (PN->getIncomingValue(i) == OperandValToReplace) {
458 // If the original expression is outside the loop, put the replacement
459 // code in the same place as the original expression,
460 // which need not be an immediate predecessor of this PHI. This way we
461 // need only one copy of it even if it is referenced multiple times in
462 // the PHI. We don't do this when the original expression is inside the
463 // loop because multiple copies sometimes do useful sinking of code in
465 Instruction *OldLoc = dyn_cast<Instruction>(OperandValToReplace);
466 BasicBlock *PHIPred = PN->getIncomingBlock(i);
467 if (L->contains(OldLoc->getParent())) {
468 // If this is a critical edge, split the edge so that we do not insert
469 // the code on all predecessor/successor paths. We do this unless this
470 // is the canonical backedge for this loop, as this can make some
471 // inserted code be in an illegal position.
472 if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 &&
473 !isa<IndirectBrInst>(PHIPred->getTerminator()) &&
474 (PN->getParent() != L->getHeader() || !L->contains(PHIPred))) {
476 // First step, split the critical edge.
477 BasicBlock *NewBB = SplitCriticalEdge(PHIPred, PN->getParent(),
480 // Next step: move the basic block. In particular, if the PHI node
481 // is outside of the loop, and PredTI is in the loop, we want to
482 // move the block to be immediately before the PHI block, not
483 // immediately after PredTI.
484 if (L->contains(PHIPred) && !L->contains(PN->getParent()))
485 NewBB->moveBefore(PN->getParent());
487 // Splitting the edge can reduce the number of PHI entries we have.
488 e = PN->getNumIncomingValues();
490 i = PN->getBasicBlockIndex(PHIPred);
493 Value *&Code = InsertedCode[PHIPred];
495 // Insert the code into the end of the predecessor block.
496 Instruction *InsertPt = (L->contains(OldLoc->getParent())) ?
497 PHIPred->getTerminator() :
498 OldLoc->getParent()->getTerminator();
499 Code = InsertCodeForBaseAtPosition(NewBase, PN->getType(),
500 Rewriter, InsertPt, L, LI);
502 DEBUG(errs() << " Changing PHI use to ");
503 DEBUG(WriteAsOperand(errs(), Code, /*PrintType=*/false));
504 DEBUG(errs() << ", which has value " << *NewBase << " plus IMM "
508 // Replace the use of the operand Value with the new Phi we just created.
509 PN->setIncomingValue(i, Code);
514 // PHI node might have become a constant value after SplitCriticalEdge.
515 DeadInsts.push_back(Inst);
519 /// fitsInAddressMode - Return true if V can be subsumed within an addressing
520 /// mode, and does not need to be put in a register first.
521 static bool fitsInAddressMode(const SCEV *const &V, const Type *AccessTy,
522 const TargetLowering *TLI, bool HasBaseReg) {
523 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
524 int64_t VC = SC->getValue()->getSExtValue();
526 TargetLowering::AddrMode AM;
528 AM.HasBaseReg = HasBaseReg;
529 return TLI->isLegalAddressingMode(AM, AccessTy);
531 // Defaults to PPC. PPC allows a sign-extended 16-bit immediate field.
532 return (VC > -(1 << 16) && VC < (1 << 16)-1);
536 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
537 if (GlobalValue *GV = dyn_cast<GlobalValue>(SU->getValue())) {
539 TargetLowering::AddrMode AM;
541 AM.HasBaseReg = HasBaseReg;
542 return TLI->isLegalAddressingMode(AM, AccessTy);
544 // Default: assume global addresses are not legal.
551 /// MoveLoopVariantsToImmediateField - Move any subexpressions from Val that are
552 /// loop varying to the Imm operand.
553 static void MoveLoopVariantsToImmediateField(const SCEV *&Val, const SCEV *&Imm,
554 Loop *L, ScalarEvolution *SE) {
555 if (Val->isLoopInvariant(L)) return; // Nothing to do.
557 if (const SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
558 SmallVector<const SCEV *, 4> NewOps;
559 NewOps.reserve(SAE->getNumOperands());
561 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
562 if (!SAE->getOperand(i)->isLoopInvariant(L)) {
563 // If this is a loop-variant expression, it must stay in the immediate
564 // field of the expression.
565 Imm = SE->getAddExpr(Imm, SAE->getOperand(i));
567 NewOps.push_back(SAE->getOperand(i));
571 Val = SE->getIntegerSCEV(0, Val->getType());
573 Val = SE->getAddExpr(NewOps);
574 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
575 // Try to pull immediates out of the start value of nested addrec's.
576 const SCEV *Start = SARE->getStart();
577 MoveLoopVariantsToImmediateField(Start, Imm, L, SE);
579 SmallVector<const SCEV *, 4> Ops(SARE->op_begin(), SARE->op_end());
581 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
583 // Otherwise, all of Val is variant, move the whole thing over.
584 Imm = SE->getAddExpr(Imm, Val);
585 Val = SE->getIntegerSCEV(0, Val->getType());
590 /// MoveImmediateValues - Look at Val, and pull out any additions of constants
591 /// that can fit into the immediate field of instructions in the target.
592 /// Accumulate these immediate values into the Imm value.
593 static void MoveImmediateValues(const TargetLowering *TLI,
594 const Type *AccessTy,
595 const SCEV *&Val, const SCEV *&Imm,
596 bool isAddress, Loop *L,
597 ScalarEvolution *SE) {
598 if (const SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
599 SmallVector<const SCEV *, 4> NewOps;
600 NewOps.reserve(SAE->getNumOperands());
602 for (unsigned i = 0; i != SAE->getNumOperands(); ++i) {
603 const SCEV *NewOp = SAE->getOperand(i);
604 MoveImmediateValues(TLI, AccessTy, NewOp, Imm, isAddress, L, SE);
606 if (!NewOp->isLoopInvariant(L)) {
607 // If this is a loop-variant expression, it must stay in the immediate
608 // field of the expression.
609 Imm = SE->getAddExpr(Imm, NewOp);
611 NewOps.push_back(NewOp);
616 Val = SE->getIntegerSCEV(0, Val->getType());
618 Val = SE->getAddExpr(NewOps);
620 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
621 // Try to pull immediates out of the start value of nested addrec's.
622 const SCEV *Start = SARE->getStart();
623 MoveImmediateValues(TLI, AccessTy, Start, Imm, isAddress, L, SE);
625 if (Start != SARE->getStart()) {
626 SmallVector<const SCEV *, 4> Ops(SARE->op_begin(), SARE->op_end());
628 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
631 } else if (const SCEVMulExpr *SME = dyn_cast<SCEVMulExpr>(Val)) {
632 // Transform "8 * (4 + v)" -> "32 + 8*V" if "32" fits in the immed field.
634 fitsInAddressMode(SME->getOperand(0), AccessTy, TLI, false) &&
635 SME->getNumOperands() == 2 && SME->isLoopInvariant(L)) {
637 const SCEV *SubImm = SE->getIntegerSCEV(0, Val->getType());
638 const SCEV *NewOp = SME->getOperand(1);
639 MoveImmediateValues(TLI, AccessTy, NewOp, SubImm, isAddress, L, SE);
641 // If we extracted something out of the subexpressions, see if we can
643 if (NewOp != SME->getOperand(1)) {
644 // Scale SubImm up by "8". If the result is a target constant, we are
646 SubImm = SE->getMulExpr(SubImm, SME->getOperand(0));
647 if (fitsInAddressMode(SubImm, AccessTy, TLI, false)) {
648 // Accumulate the immediate.
649 Imm = SE->getAddExpr(Imm, SubImm);
651 // Update what is left of 'Val'.
652 Val = SE->getMulExpr(SME->getOperand(0), NewOp);
659 // Loop-variant expressions must stay in the immediate field of the
661 if ((isAddress && fitsInAddressMode(Val, AccessTy, TLI, false)) ||
662 !Val->isLoopInvariant(L)) {
663 Imm = SE->getAddExpr(Imm, Val);
664 Val = SE->getIntegerSCEV(0, Val->getType());
668 // Otherwise, no immediates to move.
671 static void MoveImmediateValues(const TargetLowering *TLI,
673 const SCEV *&Val, const SCEV *&Imm,
674 bool isAddress, Loop *L,
675 ScalarEvolution *SE) {
676 const Type *AccessTy = getAccessType(User);
677 MoveImmediateValues(TLI, AccessTy, Val, Imm, isAddress, L, SE);
680 /// SeparateSubExprs - Decompose Expr into all of the subexpressions that are
681 /// added together. This is used to reassociate common addition subexprs
682 /// together for maximal sharing when rewriting bases.
683 static void SeparateSubExprs(SmallVector<const SCEV *, 16> &SubExprs,
685 ScalarEvolution *SE) {
686 if (const SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) {
687 for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
688 SeparateSubExprs(SubExprs, AE->getOperand(j), SE);
689 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) {
690 const SCEV *Zero = SE->getIntegerSCEV(0, Expr->getType());
691 if (SARE->getOperand(0) == Zero) {
692 SubExprs.push_back(Expr);
694 // Compute the addrec with zero as its base.
695 SmallVector<const SCEV *, 4> Ops(SARE->op_begin(), SARE->op_end());
696 Ops[0] = Zero; // Start with zero base.
697 SubExprs.push_back(SE->getAddRecExpr(Ops, SARE->getLoop()));
700 SeparateSubExprs(SubExprs, SARE->getOperand(0), SE);
702 } else if (!Expr->isZero()) {
704 SubExprs.push_back(Expr);
708 // This is logically local to the following function, but C++ says we have
709 // to make it file scope.
710 struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; };
712 /// RemoveCommonExpressionsFromUseBases - Look through all of the Bases of all
713 /// the Uses, removing any common subexpressions, except that if all such
714 /// subexpressions can be folded into an addressing mode for all uses inside
715 /// the loop (this case is referred to as "free" in comments herein) we do
716 /// not remove anything. This looks for things like (a+b+c) and
717 /// (a+c+d) and computes the common (a+c) subexpression. The common expression
718 /// is *removed* from the Bases and returned.
720 RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses,
721 ScalarEvolution *SE, Loop *L,
722 const TargetLowering *TLI) {
723 unsigned NumUses = Uses.size();
725 // Only one use? This is a very common case, so we handle it specially and
727 const SCEV *Zero = SE->getIntegerSCEV(0, Uses[0].Base->getType());
728 const SCEV *Result = Zero;
729 const SCEV *FreeResult = Zero;
731 // If the use is inside the loop, use its base, regardless of what it is:
732 // it is clearly shared across all the IV's. If the use is outside the loop
733 // (which means after it) we don't want to factor anything *into* the loop,
734 // so just use 0 as the base.
735 if (L->contains(Uses[0].Inst->getParent()))
736 std::swap(Result, Uses[0].Base);
740 // To find common subexpressions, count how many of Uses use each expression.
741 // If any subexpressions are used Uses.size() times, they are common.
742 // Also track whether all uses of each expression can be moved into an
743 // an addressing mode "for free"; such expressions are left within the loop.
744 // struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; };
745 std::map<const SCEV *, SubExprUseData> SubExpressionUseData;
747 // UniqueSubExprs - Keep track of all of the subexpressions we see in the
748 // order we see them.
749 SmallVector<const SCEV *, 16> UniqueSubExprs;
751 SmallVector<const SCEV *, 16> SubExprs;
752 unsigned NumUsesInsideLoop = 0;
753 for (unsigned i = 0; i != NumUses; ++i) {
754 // If the user is outside the loop, just ignore it for base computation.
755 // Since the user is outside the loop, it must be *after* the loop (if it
756 // were before, it could not be based on the loop IV). We don't want users
757 // after the loop to affect base computation of values *inside* the loop,
758 // because we can always add their offsets to the result IV after the loop
759 // is done, ensuring we get good code inside the loop.
760 if (!L->contains(Uses[i].Inst->getParent()))
764 // If the base is zero (which is common), return zero now, there are no
766 if (Uses[i].Base == Zero) return Zero;
768 // If this use is as an address we may be able to put CSEs in the addressing
769 // mode rather than hoisting them.
770 bool isAddrUse = isAddressUse(Uses[i].Inst, Uses[i].OperandValToReplace);
771 // We may need the AccessTy below, but only when isAddrUse, so compute it
772 // only in that case.
773 const Type *AccessTy = 0;
775 AccessTy = getAccessType(Uses[i].Inst);
777 // Split the expression into subexprs.
778 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
779 // Add one to SubExpressionUseData.Count for each subexpr present, and
780 // if the subexpr is not a valid immediate within an addressing mode use,
781 // set SubExpressionUseData.notAllUsesAreFree. We definitely want to
782 // hoist these out of the loop (if they are common to all uses).
783 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) {
784 if (++SubExpressionUseData[SubExprs[j]].Count == 1)
785 UniqueSubExprs.push_back(SubExprs[j]);
786 if (!isAddrUse || !fitsInAddressMode(SubExprs[j], AccessTy, TLI, false))
787 SubExpressionUseData[SubExprs[j]].notAllUsesAreFree = true;
792 // Now that we know how many times each is used, build Result. Iterate over
793 // UniqueSubexprs so that we have a stable ordering.
794 for (unsigned i = 0, e = UniqueSubExprs.size(); i != e; ++i) {
795 std::map<const SCEV *, SubExprUseData>::iterator I =
796 SubExpressionUseData.find(UniqueSubExprs[i]);
797 assert(I != SubExpressionUseData.end() && "Entry not found?");
798 if (I->second.Count == NumUsesInsideLoop) { // Found CSE!
799 if (I->second.notAllUsesAreFree)
800 Result = SE->getAddExpr(Result, I->first);
802 FreeResult = SE->getAddExpr(FreeResult, I->first);
804 // Remove non-cse's from SubExpressionUseData.
805 SubExpressionUseData.erase(I);
808 if (FreeResult != Zero) {
809 // We have some subexpressions that can be subsumed into addressing
810 // modes in every use inside the loop. However, it's possible that
811 // there are so many of them that the combined FreeResult cannot
812 // be subsumed, or that the target cannot handle both a FreeResult
813 // and a Result in the same instruction (for example because it would
814 // require too many registers). Check this.
815 for (unsigned i=0; i<NumUses; ++i) {
816 if (!L->contains(Uses[i].Inst->getParent()))
818 // We know this is an addressing mode use; if there are any uses that
819 // are not, FreeResult would be Zero.
820 const Type *AccessTy = getAccessType(Uses[i].Inst);
821 if (!fitsInAddressMode(FreeResult, AccessTy, TLI, Result!=Zero)) {
822 // FIXME: could split up FreeResult into pieces here, some hoisted
823 // and some not. There is no obvious advantage to this.
824 Result = SE->getAddExpr(Result, FreeResult);
831 // If we found no CSE's, return now.
832 if (Result == Zero) return Result;
834 // If we still have a FreeResult, remove its subexpressions from
835 // SubExpressionUseData. This means they will remain in the use Bases.
836 if (FreeResult != Zero) {
837 SeparateSubExprs(SubExprs, FreeResult, SE);
838 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) {
839 std::map<const SCEV *, SubExprUseData>::iterator I =
840 SubExpressionUseData.find(SubExprs[j]);
841 SubExpressionUseData.erase(I);
846 // Otherwise, remove all of the CSE's we found from each of the base values.
847 for (unsigned i = 0; i != NumUses; ++i) {
848 // Uses outside the loop don't necessarily include the common base, but
849 // the final IV value coming into those uses does. Instead of trying to
850 // remove the pieces of the common base, which might not be there,
851 // subtract off the base to compensate for this.
852 if (!L->contains(Uses[i].Inst->getParent())) {
853 Uses[i].Base = SE->getMinusSCEV(Uses[i].Base, Result);
857 // Split the expression into subexprs.
858 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
860 // Remove any common subexpressions.
861 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
862 if (SubExpressionUseData.count(SubExprs[j])) {
863 SubExprs.erase(SubExprs.begin()+j);
867 // Finally, add the non-shared expressions together.
868 if (SubExprs.empty())
871 Uses[i].Base = SE->getAddExpr(SubExprs);
878 /// ValidScale - Check whether the given Scale is valid for all loads and
879 /// stores in UsersToProcess.
881 bool LoopStrengthReduce::ValidScale(bool HasBaseReg, int64_t Scale,
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.HasBaseReg = HasBaseReg || !UsersToProcess[i].Base->isZero();
902 // If load[imm+r*scale] is illegal, bail out.
903 if (!TLI->isLegalAddressingMode(AM, AccessTy))
909 /// ValidOffset - Check whether the given Offset is valid for all loads and
910 /// stores in UsersToProcess.
912 bool LoopStrengthReduce::ValidOffset(bool HasBaseReg,
915 const std::vector<BasedUser>& UsersToProcess) {
919 for (unsigned i=0, e = UsersToProcess.size(); i!=e; ++i) {
920 // If this is a load or other access, pass the type of the access in.
921 const Type *AccessTy =
922 Type::getVoidTy(UsersToProcess[i].Inst->getContext());
923 if (isAddressUse(UsersToProcess[i].Inst,
924 UsersToProcess[i].OperandValToReplace))
925 AccessTy = getAccessType(UsersToProcess[i].Inst);
926 else if (isa<PHINode>(UsersToProcess[i].Inst))
929 TargetLowering::AddrMode AM;
930 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
931 AM.BaseOffs = SC->getValue()->getSExtValue();
932 AM.BaseOffs = (uint64_t)AM.BaseOffs + (uint64_t)Offset;
933 AM.HasBaseReg = HasBaseReg || !UsersToProcess[i].Base->isZero();
936 // If load[imm+r*scale] is illegal, bail out.
937 if (!TLI->isLegalAddressingMode(AM, AccessTy))
943 /// RequiresTypeConversion - Returns true if converting Ty1 to Ty2 is not
945 bool LoopStrengthReduce::RequiresTypeConversion(const Type *Ty1,
949 Ty1 = SE->getEffectiveSCEVType(Ty1);
950 Ty2 = SE->getEffectiveSCEVType(Ty2);
953 if (Ty1->canLosslesslyBitCastTo(Ty2))
955 if (TLI && TLI->isTruncateFree(Ty1, Ty2))
960 /// CheckForIVReuse - Returns the multiple if the stride is the multiple
961 /// of a previous stride and it is a legal value for the target addressing
962 /// mode scale component and optional base reg. This allows the users of
963 /// this stride to be rewritten as prev iv * factor. It returns 0 if no
964 /// reuse is possible. Factors can be negative on same targets, e.g. ARM.
966 /// If all uses are outside the loop, we don't require that all multiplies
967 /// be folded into the addressing mode, nor even that the factor be constant;
968 /// a multiply (executed once) outside the loop is better than another IV
969 /// within. Well, usually.
970 const SCEV *LoopStrengthReduce::CheckForIVReuse(bool HasBaseReg,
971 bool AllUsesAreAddresses,
972 bool AllUsesAreOutsideLoop,
973 const SCEV *const &Stride,
974 IVExpr &IV, const Type *Ty,
975 const std::vector<BasedUser>& UsersToProcess) {
976 if (StrideNoReuse.count(Stride))
977 return SE->getIntegerSCEV(0, Stride->getType());
979 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Stride)) {
980 int64_t SInt = SC->getValue()->getSExtValue();
981 for (unsigned NewStride = 0, e = IU->StrideOrder.size();
982 NewStride != e; ++NewStride) {
983 std::map<const SCEV *, IVsOfOneStride>::iterator SI =
984 IVsByStride.find(IU->StrideOrder[NewStride]);
985 if (SI == IVsByStride.end() || !isa<SCEVConstant>(SI->first) ||
986 StrideNoReuse.count(SI->first))
988 // The other stride has no uses, don't reuse it.
989 std::map<const SCEV *, IVUsersOfOneStride *>::iterator UI =
990 IU->IVUsesByStride.find(IU->StrideOrder[NewStride]);
991 if (UI->second->Users.empty())
993 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
994 if (SI->first != Stride &&
995 (unsigned(abs64(SInt)) < SSInt || (SInt % SSInt) != 0))
997 int64_t Scale = SInt / SSInt;
998 // Check that this stride is valid for all the types used for loads and
999 // stores; if it can be used for some and not others, we might as well use
1000 // the original stride everywhere, since we have to create the IV for it
1001 // anyway. If the scale is 1, then we don't need to worry about folding
1004 (AllUsesAreAddresses &&
1005 ValidScale(HasBaseReg, Scale, UsersToProcess))) {
1006 // Prefer to reuse an IV with a base of zero.
1007 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1008 IE = SI->second.IVs.end(); II != IE; ++II)
1009 // Only reuse previous IV if it would not require a type conversion
1010 // and if the base difference can be folded.
1011 if (II->Base->isZero() &&
1012 !RequiresTypeConversion(II->Base->getType(), Ty)) {
1014 return SE->getIntegerSCEV(Scale, Stride->getType());
1016 // Otherwise, settle for an IV with a foldable base.
1017 if (AllUsesAreAddresses)
1018 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1019 IE = SI->second.IVs.end(); II != IE; ++II)
1020 // Only reuse previous IV if it would not require a type conversion
1021 // and if the base difference can be folded.
1022 if (SE->getEffectiveSCEVType(II->Base->getType()) ==
1023 SE->getEffectiveSCEVType(Ty) &&
1024 isa<SCEVConstant>(II->Base)) {
1026 cast<SCEVConstant>(II->Base)->getValue()->getSExtValue();
1027 if (Base > INT32_MIN && Base <= INT32_MAX &&
1028 ValidOffset(HasBaseReg, -Base * Scale,
1029 Scale, UsersToProcess)) {
1031 return SE->getIntegerSCEV(Scale, Stride->getType());
1036 } else if (AllUsesAreOutsideLoop) {
1037 // Accept nonconstant strides here; it is really really right to substitute
1038 // an existing IV if we can.
1039 for (unsigned NewStride = 0, e = IU->StrideOrder.size();
1040 NewStride != e; ++NewStride) {
1041 std::map<const SCEV *, IVsOfOneStride>::iterator SI =
1042 IVsByStride.find(IU->StrideOrder[NewStride]);
1043 if (SI == IVsByStride.end() || !isa<SCEVConstant>(SI->first))
1045 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1046 if (SI->first != Stride && SSInt != 1)
1048 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1049 IE = SI->second.IVs.end(); II != IE; ++II)
1050 // Accept nonzero base here.
1051 // Only reuse previous IV if it would not require a type conversion.
1052 if (!RequiresTypeConversion(II->Base->getType(), Ty)) {
1057 // Special case, old IV is -1*x and this one is x. Can treat this one as
1059 for (unsigned NewStride = 0, e = IU->StrideOrder.size();
1060 NewStride != e; ++NewStride) {
1061 std::map<const SCEV *, IVsOfOneStride>::iterator SI =
1062 IVsByStride.find(IU->StrideOrder[NewStride]);
1063 if (SI == IVsByStride.end())
1065 if (const SCEVMulExpr *ME = dyn_cast<SCEVMulExpr>(SI->first))
1066 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(ME->getOperand(0)))
1067 if (Stride == ME->getOperand(1) &&
1068 SC->getValue()->getSExtValue() == -1LL)
1069 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1070 IE = SI->second.IVs.end(); II != IE; ++II)
1071 // Accept nonzero base here.
1072 // Only reuse previous IV if it would not require type conversion.
1073 if (!RequiresTypeConversion(II->Base->getType(), Ty)) {
1075 return SE->getIntegerSCEV(-1LL, Stride->getType());
1079 return SE->getIntegerSCEV(0, Stride->getType());
1082 /// PartitionByIsUseOfPostIncrementedValue - Simple boolean predicate that
1083 /// returns true if Val's isUseOfPostIncrementedValue is true.
1084 static bool PartitionByIsUseOfPostIncrementedValue(const BasedUser &Val) {
1085 return Val.isUseOfPostIncrementedValue;
1088 /// isNonConstantNegative - Return true if the specified scev is negated, but
1090 static bool isNonConstantNegative(const SCEV *const &Expr) {
1091 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Expr);
1092 if (!Mul) return false;
1094 // If there is a constant factor, it will be first.
1095 const SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
1096 if (!SC) return false;
1098 // Return true if the value is negative, this matches things like (-42 * V).
1099 return SC->getValue()->getValue().isNegative();
1102 /// CollectIVUsers - Transform our list of users and offsets to a bit more
1103 /// complex table. In this new vector, each 'BasedUser' contains 'Base', the
1104 /// base of the strided accesses, as well as the old information from Uses. We
1105 /// progressively move information from the Base field to the Imm field, until
1106 /// we eventually have the full access expression to rewrite the use.
1107 const SCEV *LoopStrengthReduce::CollectIVUsers(const SCEV *const &Stride,
1108 IVUsersOfOneStride &Uses,
1110 bool &AllUsesAreAddresses,
1111 bool &AllUsesAreOutsideLoop,
1112 std::vector<BasedUser> &UsersToProcess) {
1113 // FIXME: Generalize to non-affine IV's.
1114 if (!Stride->isLoopInvariant(L))
1115 return SE->getIntegerSCEV(0, Stride->getType());
1117 UsersToProcess.reserve(Uses.Users.size());
1118 for (ilist<IVStrideUse>::iterator I = Uses.Users.begin(),
1119 E = Uses.Users.end(); I != E; ++I) {
1120 UsersToProcess.push_back(BasedUser(*I, SE));
1122 // Move any loop variant operands from the offset field to the immediate
1123 // field of the use, so that we don't try to use something before it is
1125 MoveLoopVariantsToImmediateField(UsersToProcess.back().Base,
1126 UsersToProcess.back().Imm, L, SE);
1127 assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
1128 "Base value is not loop invariant!");
1131 // We now have a whole bunch of uses of like-strided induction variables, but
1132 // they might all have different bases. We want to emit one PHI node for this
1133 // stride which we fold as many common expressions (between the IVs) into as
1134 // possible. Start by identifying the common expressions in the base values
1135 // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
1136 // "A+B"), emit it to the preheader, then remove the expression from the
1137 // UsersToProcess base values.
1138 const SCEV *CommonExprs =
1139 RemoveCommonExpressionsFromUseBases(UsersToProcess, SE, L, TLI);
1141 // Next, figure out what we can represent in the immediate fields of
1142 // instructions. If we can represent anything there, move it to the imm
1143 // fields of the BasedUsers. We do this so that it increases the commonality
1144 // of the remaining uses.
1145 unsigned NumPHI = 0;
1146 bool HasAddress = false;
1147 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1148 // If the user is not in the current loop, this means it is using the exit
1149 // value of the IV. Do not put anything in the base, make sure it's all in
1150 // the immediate field to allow as much factoring as possible.
1151 if (!L->contains(UsersToProcess[i].Inst->getParent())) {
1152 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm,
1153 UsersToProcess[i].Base);
1154 UsersToProcess[i].Base =
1155 SE->getIntegerSCEV(0, UsersToProcess[i].Base->getType());
1157 // Not all uses are outside the loop.
1158 AllUsesAreOutsideLoop = false;
1160 // Addressing modes can be folded into loads and stores. Be careful that
1161 // the store is through the expression, not of the expression though.
1163 bool isAddress = isAddressUse(UsersToProcess[i].Inst,
1164 UsersToProcess[i].OperandValToReplace);
1165 if (isa<PHINode>(UsersToProcess[i].Inst)) {
1173 // If this use isn't an address, then not all uses are addresses.
1174 if (!isAddress && !isPHI)
1175 AllUsesAreAddresses = false;
1177 MoveImmediateValues(TLI, UsersToProcess[i].Inst, UsersToProcess[i].Base,
1178 UsersToProcess[i].Imm, isAddress, L, SE);
1182 // If one of the use is a PHI node and all other uses are addresses, still
1183 // allow iv reuse. Essentially we are trading one constant multiplication
1184 // for one fewer iv.
1186 AllUsesAreAddresses = false;
1188 // There are no in-loop address uses.
1189 if (AllUsesAreAddresses && (!HasAddress && !AllUsesAreOutsideLoop))
1190 AllUsesAreAddresses = false;
1195 /// ShouldUseFullStrengthReductionMode - Test whether full strength-reduction
1196 /// is valid and profitable for the given set of users of a stride. In
1197 /// full strength-reduction mode, all addresses at the current stride are
1198 /// strength-reduced all the way down to pointer arithmetic.
1200 bool LoopStrengthReduce::ShouldUseFullStrengthReductionMode(
1201 const std::vector<BasedUser> &UsersToProcess,
1203 bool AllUsesAreAddresses,
1204 const SCEV *Stride) {
1205 if (!EnableFullLSRMode)
1208 // The heuristics below aim to avoid increasing register pressure, but
1209 // fully strength-reducing all the addresses increases the number of
1210 // add instructions, so don't do this when optimizing for size.
1211 // TODO: If the loop is large, the savings due to simpler addresses
1212 // may oughtweight the costs of the extra increment instructions.
1213 if (L->getHeader()->getParent()->hasFnAttr(Attribute::OptimizeForSize))
1216 // TODO: For now, don't do full strength reduction if there could
1217 // potentially be greater-stride multiples of the current stride
1218 // which could reuse the current stride IV.
1219 if (IU->StrideOrder.back() != Stride)
1222 // Iterate through the uses to find conditions that automatically rule out
1224 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ) {
1225 const SCEV *Base = UsersToProcess[i].Base;
1226 const SCEV *Imm = UsersToProcess[i].Imm;
1227 // If any users have a loop-variant component, they can't be fully
1228 // strength-reduced.
1229 if (Imm && !Imm->isLoopInvariant(L))
1231 // If there are to users with the same base and the difference between
1232 // the two Imm values can't be folded into the address, full
1233 // strength reduction would increase register pressure.
1235 const SCEV *CurImm = UsersToProcess[i].Imm;
1236 if ((CurImm || Imm) && CurImm != Imm) {
1237 if (!CurImm) CurImm = SE->getIntegerSCEV(0, Stride->getType());
1238 if (!Imm) Imm = SE->getIntegerSCEV(0, Stride->getType());
1239 const Instruction *Inst = UsersToProcess[i].Inst;
1240 const Type *AccessTy = getAccessType(Inst);
1241 const SCEV *Diff = SE->getMinusSCEV(UsersToProcess[i].Imm, Imm);
1242 if (!Diff->isZero() &&
1243 (!AllUsesAreAddresses ||
1244 !fitsInAddressMode(Diff, AccessTy, TLI, /*HasBaseReg=*/true)))
1247 } while (++i != e && Base == UsersToProcess[i].Base);
1250 // If there's exactly one user in this stride, fully strength-reducing it
1251 // won't increase register pressure. If it's starting from a non-zero base,
1252 // it'll be simpler this way.
1253 if (UsersToProcess.size() == 1 && !UsersToProcess[0].Base->isZero())
1256 // Otherwise, if there are any users in this stride that don't require
1257 // a register for their base, full strength-reduction will increase
1258 // register pressure.
1259 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1260 if (UsersToProcess[i].Base->isZero())
1263 // Otherwise, go for it.
1267 /// InsertAffinePhi Create and insert a PHI node for an induction variable
1268 /// with the specified start and step values in the specified loop.
1270 /// If NegateStride is true, the stride should be negated by using a
1271 /// subtract instead of an add.
1273 /// Return the created phi node.
1275 static PHINode *InsertAffinePhi(const SCEV *Start, const SCEV *Step,
1276 Instruction *IVIncInsertPt,
1278 SCEVExpander &Rewriter) {
1279 assert(Start->isLoopInvariant(L) && "New PHI start is not loop invariant!");
1280 assert(Step->isLoopInvariant(L) && "New PHI stride is not loop invariant!");
1282 BasicBlock *Header = L->getHeader();
1283 BasicBlock *Preheader = L->getLoopPreheader();
1284 BasicBlock *LatchBlock = L->getLoopLatch();
1285 const Type *Ty = Start->getType();
1286 Ty = Rewriter.SE.getEffectiveSCEVType(Ty);
1288 PHINode *PN = PHINode::Create(Ty, "lsr.iv", Header->begin());
1289 PN->addIncoming(Rewriter.expandCodeFor(Start, Ty, Preheader->getTerminator()),
1292 // If the stride is negative, insert a sub instead of an add for the
1294 bool isNegative = isNonConstantNegative(Step);
1295 const SCEV *IncAmount = Step;
1297 IncAmount = Rewriter.SE.getNegativeSCEV(Step);
1299 // Insert an add instruction right before the terminator corresponding
1300 // to the back-edge or just before the only use. The location is determined
1301 // by the caller and passed in as IVIncInsertPt.
1302 Value *StepV = Rewriter.expandCodeFor(IncAmount, Ty,
1303 Preheader->getTerminator());
1306 IncV = BinaryOperator::CreateSub(PN, StepV, "lsr.iv.next",
1309 IncV = BinaryOperator::CreateAdd(PN, StepV, "lsr.iv.next",
1312 if (!isa<ConstantInt>(StepV)) ++NumVariable;
1314 PN->addIncoming(IncV, LatchBlock);
1320 static void SortUsersToProcess(std::vector<BasedUser> &UsersToProcess) {
1321 // We want to emit code for users inside the loop first. To do this, we
1322 // rearrange BasedUser so that the entries at the end have
1323 // isUseOfPostIncrementedValue = false, because we pop off the end of the
1324 // vector (so we handle them first).
1325 std::partition(UsersToProcess.begin(), UsersToProcess.end(),
1326 PartitionByIsUseOfPostIncrementedValue);
1328 // Sort this by base, so that things with the same base are handled
1329 // together. By partitioning first and stable-sorting later, we are
1330 // guaranteed that within each base we will pop off users from within the
1331 // loop before users outside of the loop with a particular base.
1333 // We would like to use stable_sort here, but we can't. The problem is that
1334 // const SCEV *'s don't have a deterministic ordering w.r.t to each other, so
1335 // we don't have anything to do a '<' comparison on. Because we think the
1336 // number of uses is small, do a horrible bubble sort which just relies on
1338 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1339 // Get a base value.
1340 const SCEV *Base = UsersToProcess[i].Base;
1342 // Compact everything with this base to be consecutive with this one.
1343 for (unsigned j = i+1; j != e; ++j) {
1344 if (UsersToProcess[j].Base == Base) {
1345 std::swap(UsersToProcess[i+1], UsersToProcess[j]);
1352 /// PrepareToStrengthReduceFully - Prepare to fully strength-reduce
1353 /// UsersToProcess, meaning lowering addresses all the way down to direct
1354 /// pointer arithmetic.
1357 LoopStrengthReduce::PrepareToStrengthReduceFully(
1358 std::vector<BasedUser> &UsersToProcess,
1360 const SCEV *CommonExprs,
1362 SCEVExpander &PreheaderRewriter) {
1363 DEBUG(errs() << " Fully reducing all users\n");
1365 // Rewrite the UsersToProcess records, creating a separate PHI for each
1366 // unique Base value.
1367 Instruction *IVIncInsertPt = L->getLoopLatch()->getTerminator();
1368 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ) {
1369 // TODO: The uses are grouped by base, but not sorted. We arbitrarily
1370 // pick the first Imm value here to start with, and adjust it for the
1372 const SCEV *Imm = UsersToProcess[i].Imm;
1373 const SCEV *Base = UsersToProcess[i].Base;
1374 const SCEV *Start = SE->getAddExpr(CommonExprs, Base, Imm);
1375 PHINode *Phi = InsertAffinePhi(Start, Stride, IVIncInsertPt, L,
1377 // Loop over all the users with the same base.
1379 UsersToProcess[i].Base = SE->getIntegerSCEV(0, Stride->getType());
1380 UsersToProcess[i].Imm = SE->getMinusSCEV(UsersToProcess[i].Imm, Imm);
1381 UsersToProcess[i].Phi = Phi;
1382 assert(UsersToProcess[i].Imm->isLoopInvariant(L) &&
1383 "ShouldUseFullStrengthReductionMode should reject this!");
1384 } while (++i != e && Base == UsersToProcess[i].Base);
1388 /// FindIVIncInsertPt - Return the location to insert the increment instruction.
1389 /// If the only use if a use of postinc value, (must be the loop termination
1390 /// condition), then insert it just before the use.
1391 static Instruction *FindIVIncInsertPt(std::vector<BasedUser> &UsersToProcess,
1393 if (UsersToProcess.size() == 1 &&
1394 UsersToProcess[0].isUseOfPostIncrementedValue &&
1395 L->contains(UsersToProcess[0].Inst->getParent()))
1396 return UsersToProcess[0].Inst;
1397 return L->getLoopLatch()->getTerminator();
1400 /// PrepareToStrengthReduceWithNewPhi - Insert a new induction variable for the
1401 /// given users to share.
1404 LoopStrengthReduce::PrepareToStrengthReduceWithNewPhi(
1405 std::vector<BasedUser> &UsersToProcess,
1407 const SCEV *CommonExprs,
1409 Instruction *IVIncInsertPt,
1411 SCEVExpander &PreheaderRewriter) {
1412 DEBUG(errs() << " Inserting new PHI:\n");
1414 PHINode *Phi = InsertAffinePhi(SE->getUnknown(CommonBaseV),
1415 Stride, IVIncInsertPt, L,
1418 // Remember this in case a later stride is multiple of this.
1419 IVsByStride[Stride].addIV(Stride, CommonExprs, Phi);
1421 // All the users will share this new IV.
1422 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1423 UsersToProcess[i].Phi = Phi;
1425 DEBUG(errs() << " IV=");
1426 DEBUG(WriteAsOperand(errs(), Phi, /*PrintType=*/false));
1427 DEBUG(errs() << "\n");
1430 /// PrepareToStrengthReduceFromSmallerStride - Prepare for the given users to
1431 /// reuse an induction variable with a stride that is a factor of the current
1432 /// induction variable.
1435 LoopStrengthReduce::PrepareToStrengthReduceFromSmallerStride(
1436 std::vector<BasedUser> &UsersToProcess,
1438 const IVExpr &ReuseIV,
1439 Instruction *PreInsertPt) {
1440 DEBUG(errs() << " Rewriting in terms of existing IV of STRIDE "
1441 << *ReuseIV.Stride << " and BASE " << *ReuseIV.Base << "\n");
1443 // All the users will share the reused IV.
1444 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1445 UsersToProcess[i].Phi = ReuseIV.PHI;
1447 Constant *C = dyn_cast<Constant>(CommonBaseV);
1449 (!C->isNullValue() &&
1450 !fitsInAddressMode(SE->getUnknown(CommonBaseV), CommonBaseV->getType(),
1452 // We want the common base emitted into the preheader! This is just
1453 // using cast as a copy so BitCast (no-op cast) is appropriate
1454 CommonBaseV = new BitCastInst(CommonBaseV, CommonBaseV->getType(),
1455 "commonbase", PreInsertPt);
1458 static bool IsImmFoldedIntoAddrMode(GlobalValue *GV, int64_t Offset,
1459 const Type *AccessTy,
1460 std::vector<BasedUser> &UsersToProcess,
1461 const TargetLowering *TLI) {
1462 SmallVector<Instruction*, 16> AddrModeInsts;
1463 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1464 if (UsersToProcess[i].isUseOfPostIncrementedValue)
1466 ExtAddrMode AddrMode =
1467 AddressingModeMatcher::Match(UsersToProcess[i].OperandValToReplace,
1468 AccessTy, UsersToProcess[i].Inst,
1469 AddrModeInsts, *TLI);
1470 if (GV && GV != AddrMode.BaseGV)
1472 if (Offset && !AddrMode.BaseOffs)
1473 // FIXME: How to accurate check it's immediate offset is folded.
1475 AddrModeInsts.clear();
1480 /// StrengthReduceIVUsersOfStride - Strength reduce all of the users of a single
1481 /// stride of IV. All of the users may have different starting values, and this
1482 /// may not be the only stride.
1484 LoopStrengthReduce::StrengthReduceIVUsersOfStride(const SCEV *const &Stride,
1485 IVUsersOfOneStride &Uses,
1487 // If all the users are moved to another stride, then there is nothing to do.
1488 if (Uses.Users.empty())
1491 // Keep track if every use in UsersToProcess is an address. If they all are,
1492 // we may be able to rewrite the entire collection of them in terms of a
1493 // smaller-stride IV.
1494 bool AllUsesAreAddresses = true;
1496 // Keep track if every use of a single stride is outside the loop. If so,
1497 // we want to be more aggressive about reusing a smaller-stride IV; a
1498 // multiply outside the loop is better than another IV inside. Well, usually.
1499 bool AllUsesAreOutsideLoop = true;
1501 // Transform our list of users and offsets to a bit more complex table. In
1502 // this new vector, each 'BasedUser' contains 'Base' the base of the
1503 // strided accessas well as the old information from Uses. We progressively
1504 // move information from the Base field to the Imm field, until we eventually
1505 // have the full access expression to rewrite the use.
1506 std::vector<BasedUser> UsersToProcess;
1507 const SCEV *CommonExprs = CollectIVUsers(Stride, Uses, L, AllUsesAreAddresses,
1508 AllUsesAreOutsideLoop,
1511 // Sort the UsersToProcess array so that users with common bases are
1512 // next to each other.
1513 SortUsersToProcess(UsersToProcess);
1515 // If we managed to find some expressions in common, we'll need to carry
1516 // their value in a register and add it in for each use. This will take up
1517 // a register operand, which potentially restricts what stride values are
1519 bool HaveCommonExprs = !CommonExprs->isZero();
1520 const Type *ReplacedTy = CommonExprs->getType();
1522 // If all uses are addresses, consider sinking the immediate part of the
1523 // common expression back into uses if they can fit in the immediate fields.
1524 if (TLI && HaveCommonExprs && AllUsesAreAddresses) {
1525 const SCEV *NewCommon = CommonExprs;
1526 const SCEV *Imm = SE->getIntegerSCEV(0, ReplacedTy);
1527 MoveImmediateValues(TLI, Type::getVoidTy(
1528 L->getLoopPreheader()->getContext()),
1529 NewCommon, Imm, true, L, SE);
1530 if (!Imm->isZero()) {
1533 // If the immediate part of the common expression is a GV, check if it's
1534 // possible to fold it into the target addressing mode.
1535 GlobalValue *GV = 0;
1536 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(Imm))
1537 GV = dyn_cast<GlobalValue>(SU->getValue());
1539 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Imm))
1540 Offset = SC->getValue()->getSExtValue();
1542 // Pass VoidTy as the AccessTy to be conservative, because
1543 // there could be multiple access types among all the uses.
1544 DoSink = IsImmFoldedIntoAddrMode(GV, Offset,
1545 Type::getVoidTy(L->getLoopPreheader()->getContext()),
1546 UsersToProcess, TLI);
1549 DEBUG(errs() << " Sinking " << *Imm << " back down into uses\n");
1550 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1551 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm, Imm);
1552 CommonExprs = NewCommon;
1553 HaveCommonExprs = !CommonExprs->isZero();
1559 // Now that we know what we need to do, insert the PHI node itself.
1561 DEBUG(errs() << "LSR: Examining IVs of TYPE " << *ReplacedTy << " of STRIDE "
1563 << " Common base: " << *CommonExprs << "\n");
1565 SCEVExpander Rewriter(*SE);
1566 SCEVExpander PreheaderRewriter(*SE);
1568 BasicBlock *Preheader = L->getLoopPreheader();
1569 Instruction *PreInsertPt = Preheader->getTerminator();
1570 BasicBlock *LatchBlock = L->getLoopLatch();
1571 Instruction *IVIncInsertPt = LatchBlock->getTerminator();
1573 Value *CommonBaseV = Constant::getNullValue(ReplacedTy);
1575 const SCEV *RewriteFactor = SE->getIntegerSCEV(0, ReplacedTy);
1576 IVExpr ReuseIV(SE->getIntegerSCEV(0,
1577 Type::getInt32Ty(Preheader->getContext())),
1578 SE->getIntegerSCEV(0,
1579 Type::getInt32Ty(Preheader->getContext())),
1582 // Choose a strength-reduction strategy and prepare for it by creating
1583 // the necessary PHIs and adjusting the bookkeeping.
1584 if (ShouldUseFullStrengthReductionMode(UsersToProcess, L,
1585 AllUsesAreAddresses, Stride)) {
1586 PrepareToStrengthReduceFully(UsersToProcess, Stride, CommonExprs, L,
1589 // Emit the initial base value into the loop preheader.
1590 CommonBaseV = PreheaderRewriter.expandCodeFor(CommonExprs, ReplacedTy,
1593 // If all uses are addresses, check if it is possible to reuse an IV. The
1594 // new IV must have a stride that is a multiple of the old stride; the
1595 // multiple must be a number that can be encoded in the scale field of the
1596 // target addressing mode; and we must have a valid instruction after this
1597 // substitution, including the immediate field, if any.
1598 RewriteFactor = CheckForIVReuse(HaveCommonExprs, AllUsesAreAddresses,
1599 AllUsesAreOutsideLoop,
1600 Stride, ReuseIV, ReplacedTy,
1602 if (!RewriteFactor->isZero())
1603 PrepareToStrengthReduceFromSmallerStride(UsersToProcess, CommonBaseV,
1604 ReuseIV, PreInsertPt);
1606 IVIncInsertPt = FindIVIncInsertPt(UsersToProcess, L);
1607 PrepareToStrengthReduceWithNewPhi(UsersToProcess, Stride, CommonExprs,
1608 CommonBaseV, IVIncInsertPt,
1609 L, PreheaderRewriter);
1613 // Process all the users now, replacing their strided uses with
1614 // strength-reduced forms. This outer loop handles all bases, the inner
1615 // loop handles all users of a particular base.
1616 while (!UsersToProcess.empty()) {
1617 const SCEV *Base = UsersToProcess.back().Base;
1618 Instruction *Inst = UsersToProcess.back().Inst;
1620 // Emit the code for Base into the preheader.
1622 if (!Base->isZero()) {
1623 BaseV = PreheaderRewriter.expandCodeFor(Base, 0, PreInsertPt);
1625 DEBUG(errs() << " INSERTING code for BASE = " << *Base << ":");
1626 if (BaseV->hasName())
1627 DEBUG(errs() << " Result value name = %" << BaseV->getName());
1628 DEBUG(errs() << "\n");
1630 // If BaseV is a non-zero constant, make sure that it gets inserted into
1631 // the preheader, instead of being forward substituted into the uses. We
1632 // do this by forcing a BitCast (noop cast) to be inserted into the
1633 // preheader in this case.
1634 if (!fitsInAddressMode(Base, getAccessType(Inst), TLI, false) &&
1635 isa<Constant>(BaseV)) {
1636 // We want this constant emitted into the preheader! This is just
1637 // using cast as a copy so BitCast (no-op cast) is appropriate
1638 BaseV = new BitCastInst(BaseV, BaseV->getType(), "preheaderinsert",
1643 // Emit the code to add the immediate offset to the Phi value, just before
1644 // the instructions that we identified as using this stride and base.
1646 // FIXME: Use emitted users to emit other users.
1647 BasedUser &User = UsersToProcess.back();
1649 DEBUG(errs() << " Examining ");
1650 if (User.isUseOfPostIncrementedValue)
1651 DEBUG(errs() << "postinc");
1653 DEBUG(errs() << "preinc");
1654 DEBUG(errs() << " use ");
1655 DEBUG(WriteAsOperand(errs(), UsersToProcess.back().OperandValToReplace,
1656 /*PrintType=*/false));
1657 DEBUG(errs() << " in Inst: " << *User.Inst);
1659 // If this instruction wants to use the post-incremented value, move it
1660 // after the post-inc and use its value instead of the PHI.
1661 Value *RewriteOp = User.Phi;
1662 if (User.isUseOfPostIncrementedValue) {
1663 RewriteOp = User.Phi->getIncomingValueForBlock(LatchBlock);
1664 // If this user is in the loop, make sure it is the last thing in the
1665 // loop to ensure it is dominated by the increment. In case it's the
1666 // only use of the iv, the increment instruction is already before the
1668 if (L->contains(User.Inst->getParent()) && User.Inst != IVIncInsertPt)
1669 User.Inst->moveBefore(IVIncInsertPt);
1672 const SCEV *RewriteExpr = SE->getUnknown(RewriteOp);
1674 if (SE->getEffectiveSCEVType(RewriteOp->getType()) !=
1675 SE->getEffectiveSCEVType(ReplacedTy)) {
1676 assert(SE->getTypeSizeInBits(RewriteOp->getType()) >
1677 SE->getTypeSizeInBits(ReplacedTy) &&
1678 "Unexpected widening cast!");
1679 RewriteExpr = SE->getTruncateExpr(RewriteExpr, ReplacedTy);
1682 // If we had to insert new instructions for RewriteOp, we have to
1683 // consider that they may not have been able to end up immediately
1684 // next to RewriteOp, because non-PHI instructions may never precede
1685 // PHI instructions in a block. In this case, remember where the last
1686 // instruction was inserted so that if we're replacing a different
1687 // PHI node, we can use the later point to expand the final
1689 Instruction *NewBasePt = dyn_cast<Instruction>(RewriteOp);
1690 if (RewriteOp == User.Phi) NewBasePt = 0;
1692 // Clear the SCEVExpander's expression map so that we are guaranteed
1693 // to have the code emitted where we expect it.
1696 // If we are reusing the iv, then it must be multiplied by a constant
1697 // factor to take advantage of the addressing mode scale component.
1698 if (!RewriteFactor->isZero()) {
1699 // If we're reusing an IV with a nonzero base (currently this happens
1700 // only when all reuses are outside the loop) subtract that base here.
1701 // The base has been used to initialize the PHI node but we don't want
1703 if (!ReuseIV.Base->isZero()) {
1704 const SCEV *typedBase = ReuseIV.Base;
1705 if (SE->getEffectiveSCEVType(RewriteExpr->getType()) !=
1706 SE->getEffectiveSCEVType(ReuseIV.Base->getType())) {
1707 // It's possible the original IV is a larger type than the new IV,
1708 // in which case we have to truncate the Base. We checked in
1709 // RequiresTypeConversion that this is valid.
1710 assert(SE->getTypeSizeInBits(RewriteExpr->getType()) <
1711 SE->getTypeSizeInBits(ReuseIV.Base->getType()) &&
1712 "Unexpected lengthening conversion!");
1713 typedBase = SE->getTruncateExpr(ReuseIV.Base,
1714 RewriteExpr->getType());
1716 RewriteExpr = SE->getMinusSCEV(RewriteExpr, typedBase);
1719 // Multiply old variable, with base removed, by new scale factor.
1720 RewriteExpr = SE->getMulExpr(RewriteFactor,
1723 // The common base is emitted in the loop preheader. But since we
1724 // are reusing an IV, it has not been used to initialize the PHI node.
1725 // Add it to the expression used to rewrite the uses.
1726 // When this use is outside the loop, we earlier subtracted the
1727 // common base, and are adding it back here. Use the same expression
1728 // as before, rather than CommonBaseV, so DAGCombiner will zap it.
1729 if (!CommonExprs->isZero()) {
1730 if (L->contains(User.Inst->getParent()))
1731 RewriteExpr = SE->getAddExpr(RewriteExpr,
1732 SE->getUnknown(CommonBaseV));
1734 RewriteExpr = SE->getAddExpr(RewriteExpr, CommonExprs);
1738 // Now that we know what we need to do, insert code before User for the
1739 // immediate and any loop-variant expressions.
1741 // Add BaseV to the PHI value if needed.
1742 RewriteExpr = SE->getAddExpr(RewriteExpr, SE->getUnknown(BaseV));
1744 User.RewriteInstructionToUseNewBase(RewriteExpr, NewBasePt,
1745 Rewriter, L, this, *LI,
1748 // Mark old value we replaced as possibly dead, so that it is eliminated
1749 // if we just replaced the last use of that value.
1750 DeadInsts.push_back(User.OperandValToReplace);
1752 UsersToProcess.pop_back();
1755 // If there are any more users to process with the same base, process them
1756 // now. We sorted by base above, so we just have to check the last elt.
1757 } while (!UsersToProcess.empty() && UsersToProcess.back().Base == Base);
1758 // TODO: Next, find out which base index is the most common, pull it out.
1761 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
1762 // different starting values, into different PHIs.
1765 void LoopStrengthReduce::StrengthReduceIVUsers(Loop *L) {
1766 // Note: this processes each stride/type pair individually. All users
1767 // passed into StrengthReduceIVUsersOfStride have the same type AND stride.
1768 // Also, note that we iterate over IVUsesByStride indirectly by using
1769 // StrideOrder. This extra layer of indirection makes the ordering of
1770 // strides deterministic - not dependent on map order.
1771 for (unsigned Stride = 0, e = IU->StrideOrder.size(); Stride != e; ++Stride) {
1772 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
1773 IU->IVUsesByStride.find(IU->StrideOrder[Stride]);
1774 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
1775 // FIXME: Generalize to non-affine IV's.
1776 if (!SI->first->isLoopInvariant(L))
1778 StrengthReduceIVUsersOfStride(SI->first, *SI->second, L);
1782 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1783 /// set the IV user and stride information and return true, otherwise return
1785 bool LoopStrengthReduce::FindIVUserForCond(ICmpInst *Cond,
1786 IVStrideUse *&CondUse,
1787 const SCEV* &CondStride) {
1788 for (unsigned Stride = 0, e = IU->StrideOrder.size();
1789 Stride != e && !CondUse; ++Stride) {
1790 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
1791 IU->IVUsesByStride.find(IU->StrideOrder[Stride]);
1792 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
1794 for (ilist<IVStrideUse>::iterator UI = SI->second->Users.begin(),
1795 E = SI->second->Users.end(); UI != E; ++UI)
1796 if (UI->getUser() == Cond) {
1797 // NOTE: we could handle setcc instructions with multiple uses here, but
1798 // InstCombine does it as well for simple uses, it's not clear that it
1799 // occurs enough in real life to handle.
1801 CondStride = SI->first;
1809 // Constant strides come first which in turns are sorted by their absolute
1810 // values. If absolute values are the same, then positive strides comes first.
1812 // 4, -1, X, 1, 2 ==> 1, -1, 2, 4, X
1813 struct StrideCompare {
1814 const ScalarEvolution *SE;
1815 explicit StrideCompare(const ScalarEvolution *se) : SE(se) {}
1817 bool operator()(const SCEV *const &LHS, const SCEV *const &RHS) {
1818 const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS);
1819 const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS);
1821 int64_t LV = LHSC->getValue()->getSExtValue();
1822 int64_t RV = RHSC->getValue()->getSExtValue();
1823 uint64_t ALV = (LV < 0) ? -LV : LV;
1824 uint64_t ARV = (RV < 0) ? -RV : RV;
1832 // If it's the same value but different type, sort by bit width so
1833 // that we emit larger induction variables before smaller
1834 // ones, letting the smaller be re-written in terms of larger ones.
1835 return SE->getTypeSizeInBits(RHS->getType()) <
1836 SE->getTypeSizeInBits(LHS->getType());
1838 return LHSC && !RHSC;
1843 /// ChangeCompareStride - If a loop termination compare instruction is the
1844 /// only use of its stride, and the compaison is against a constant value,
1845 /// try eliminate the stride by moving the compare instruction to another
1846 /// stride and change its constant operand accordingly. e.g.
1852 /// if (v2 < 10) goto loop
1857 /// if (v1 < 30) goto loop
1858 ICmpInst *LoopStrengthReduce::ChangeCompareStride(Loop *L, ICmpInst *Cond,
1859 IVStrideUse* &CondUse,
1860 const SCEV* &CondStride,
1862 // If there's only one stride in the loop, there's nothing to do here.
1863 if (IU->StrideOrder.size() < 2)
1865 // If there are other users of the condition's stride, don't bother
1866 // trying to change the condition because the stride will still
1868 std::map<const SCEV *, IVUsersOfOneStride *>::iterator I =
1869 IU->IVUsesByStride.find(CondStride);
1870 if (I == IU->IVUsesByStride.end())
1872 if (I->second->Users.size() > 1) {
1873 for (ilist<IVStrideUse>::iterator II = I->second->Users.begin(),
1874 EE = I->second->Users.end(); II != EE; ++II) {
1875 if (II->getUser() == Cond)
1877 if (!isInstructionTriviallyDead(II->getUser()))
1881 // Only handle constant strides for now.
1882 const SCEVConstant *SC = dyn_cast<SCEVConstant>(CondStride);
1883 if (!SC) return Cond;
1885 ICmpInst::Predicate Predicate = Cond->getPredicate();
1886 int64_t CmpSSInt = SC->getValue()->getSExtValue();
1887 unsigned BitWidth = SE->getTypeSizeInBits(CondStride->getType());
1888 uint64_t SignBit = 1ULL << (BitWidth-1);
1889 const Type *CmpTy = Cond->getOperand(0)->getType();
1890 const Type *NewCmpTy = NULL;
1891 unsigned TyBits = SE->getTypeSizeInBits(CmpTy);
1892 unsigned NewTyBits = 0;
1893 const SCEV *NewStride = NULL;
1894 Value *NewCmpLHS = NULL;
1895 Value *NewCmpRHS = NULL;
1897 const SCEV *NewOffset = SE->getIntegerSCEV(0, CmpTy);
1899 if (ConstantInt *C = dyn_cast<ConstantInt>(Cond->getOperand(1))) {
1900 int64_t CmpVal = C->getValue().getSExtValue();
1902 // Check the relevant induction variable for conformance to
1904 const SCEV *IV = SE->getSCEV(Cond->getOperand(0));
1905 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1906 if (!AR || !AR->isAffine())
1909 const SCEVConstant *StartC = dyn_cast<SCEVConstant>(AR->getStart());
1910 // Check stride constant and the comparision constant signs to detect
1913 if ((StartC->getValue()->getSExtValue() < CmpVal && CmpSSInt < 0) ||
1914 (StartC->getValue()->getSExtValue() > CmpVal && CmpSSInt > 0))
1917 // More restrictive check for the other cases.
1918 if ((CmpVal & SignBit) != (CmpSSInt & SignBit))
1922 // Look for a suitable stride / iv as replacement.
1923 for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
1924 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
1925 IU->IVUsesByStride.find(IU->StrideOrder[i]);
1926 if (!isa<SCEVConstant>(SI->first) || SI->second->Users.empty())
1928 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1929 if (SSInt == CmpSSInt ||
1930 abs64(SSInt) < abs64(CmpSSInt) ||
1931 (SSInt % CmpSSInt) != 0)
1934 Scale = SSInt / CmpSSInt;
1935 int64_t NewCmpVal = CmpVal * Scale;
1937 // If old icmp value fits in icmp immediate field, but the new one doesn't
1938 // try something else.
1940 TLI->isLegalICmpImmediate(CmpVal) &&
1941 !TLI->isLegalICmpImmediate(NewCmpVal))
1944 APInt Mul = APInt(BitWidth*2, CmpVal, true);
1945 Mul = Mul * APInt(BitWidth*2, Scale, true);
1946 // Check for overflow.
1947 if (!Mul.isSignedIntN(BitWidth))
1949 // Check for overflow in the stride's type too.
1950 if (!Mul.isSignedIntN(SE->getTypeSizeInBits(SI->first->getType())))
1953 // Watch out for overflow.
1954 if (ICmpInst::isSigned(Predicate) &&
1955 (CmpVal & SignBit) != (NewCmpVal & SignBit))
1958 // Pick the best iv to use trying to avoid a cast.
1960 for (ilist<IVStrideUse>::iterator UI = SI->second->Users.begin(),
1961 E = SI->second->Users.end(); UI != E; ++UI) {
1962 Value *Op = UI->getOperandValToReplace();
1964 // If the IVStrideUse implies a cast, check for an actual cast which
1965 // can be used to find the original IV expression.
1966 if (SE->getEffectiveSCEVType(Op->getType()) !=
1967 SE->getEffectiveSCEVType(SI->first->getType())) {
1968 CastInst *CI = dyn_cast<CastInst>(Op);
1969 // If it's not a simple cast, it's complicated.
1972 // If it's a cast from a type other than the stride type,
1973 // it's complicated.
1974 if (CI->getOperand(0)->getType() != SI->first->getType())
1976 // Ok, we found the IV expression in the stride's type.
1977 Op = CI->getOperand(0);
1981 if (NewCmpLHS->getType() == CmpTy)
1987 NewCmpTy = NewCmpLHS->getType();
1988 NewTyBits = SE->getTypeSizeInBits(NewCmpTy);
1989 const Type *NewCmpIntTy = IntegerType::get(Cond->getContext(), NewTyBits);
1990 if (RequiresTypeConversion(NewCmpTy, CmpTy)) {
1991 // Check if it is possible to rewrite it using
1992 // an iv / stride of a smaller integer type.
1993 unsigned Bits = NewTyBits;
1994 if (ICmpInst::isSigned(Predicate))
1996 uint64_t Mask = (1ULL << Bits) - 1;
1997 if (((uint64_t)NewCmpVal & Mask) != (uint64_t)NewCmpVal)
2001 // Don't rewrite if use offset is non-constant and the new type is
2002 // of a different type.
2003 // FIXME: too conservative?
2004 if (NewTyBits != TyBits && !isa<SCEVConstant>(CondUse->getOffset()))
2008 bool AllUsesAreAddresses = true;
2009 bool AllUsesAreOutsideLoop = true;
2010 std::vector<BasedUser> UsersToProcess;
2011 const SCEV *CommonExprs = CollectIVUsers(SI->first, *SI->second, L,
2012 AllUsesAreAddresses,
2013 AllUsesAreOutsideLoop,
2015 // Avoid rewriting the compare instruction with an iv of new stride
2016 // if it's likely the new stride uses will be rewritten using the
2017 // stride of the compare instruction.
2018 if (AllUsesAreAddresses &&
2019 ValidScale(!CommonExprs->isZero(), Scale, UsersToProcess))
2023 // Avoid rewriting the compare instruction with an iv which has
2024 // implicit extension or truncation built into it.
2025 // TODO: This is over-conservative.
2026 if (SE->getTypeSizeInBits(CondUse->getOffset()->getType()) != TyBits)
2029 // If scale is negative, use swapped predicate unless it's testing
2031 if (Scale < 0 && !Cond->isEquality())
2032 Predicate = ICmpInst::getSwappedPredicate(Predicate);
2034 NewStride = IU->StrideOrder[i];
2035 if (!isa<PointerType>(NewCmpTy))
2036 NewCmpRHS = ConstantInt::get(NewCmpTy, NewCmpVal);
2038 Constant *CI = ConstantInt::get(NewCmpIntTy, NewCmpVal);
2039 NewCmpRHS = ConstantExpr::getIntToPtr(CI, NewCmpTy);
2041 NewOffset = TyBits == NewTyBits
2042 ? SE->getMulExpr(CondUse->getOffset(),
2043 SE->getConstant(CmpTy, Scale))
2044 : SE->getConstant(NewCmpIntTy,
2045 cast<SCEVConstant>(CondUse->getOffset())->getValue()
2046 ->getSExtValue()*Scale);
2051 // Forgo this transformation if it the increment happens to be
2052 // unfortunately positioned after the condition, and the condition
2053 // has multiple uses which prevent it from being moved immediately
2054 // before the branch. See
2055 // test/Transforms/LoopStrengthReduce/change-compare-stride-trickiness-*.ll
2056 // for an example of this situation.
2057 if (!Cond->hasOneUse()) {
2058 for (BasicBlock::iterator I = Cond, E = Cond->getParent()->end();
2065 // Create a new compare instruction using new stride / iv.
2066 ICmpInst *OldCond = Cond;
2067 // Insert new compare instruction.
2068 Cond = new ICmpInst(OldCond, Predicate, NewCmpLHS, NewCmpRHS,
2069 L->getHeader()->getName() + ".termcond");
2071 DEBUG(errs() << " Change compare stride in Inst " << *OldCond);
2072 DEBUG(errs() << " to " << *Cond << '\n');
2074 // Remove the old compare instruction. The old indvar is probably dead too.
2075 DeadInsts.push_back(CondUse->getOperandValToReplace());
2076 OldCond->replaceAllUsesWith(Cond);
2077 OldCond->eraseFromParent();
2079 IU->IVUsesByStride[NewStride]->addUser(NewOffset, Cond, NewCmpLHS);
2080 CondUse = &IU->IVUsesByStride[NewStride]->Users.back();
2081 CondStride = NewStride;
2089 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
2090 /// a max computation.
2092 /// This is a narrow solution to a specific, but acute, problem. For loops
2098 /// } while (++i < n);
2100 /// the trip count isn't just 'n', because 'n' might not be positive. And
2101 /// unfortunately this can come up even for loops where the user didn't use
2102 /// a C do-while loop. For example, seemingly well-behaved top-test loops
2103 /// will commonly be lowered like this:
2109 /// } while (++i < n);
2112 /// and then it's possible for subsequent optimization to obscure the if
2113 /// test in such a way that indvars can't find it.
2115 /// When indvars can't find the if test in loops like this, it creates a
2116 /// max expression, which allows it to give the loop a canonical
2117 /// induction variable:
2120 /// max = n < 1 ? 1 : n;
2123 /// } while (++i != max);
2125 /// Canonical induction variables are necessary because the loop passes
2126 /// are designed around them. The most obvious example of this is the
2127 /// LoopInfo analysis, which doesn't remember trip count values. It
2128 /// expects to be able to rediscover the trip count each time it is
2129 /// needed, and it does this using a simple analyis that only succeeds if
2130 /// the loop has a canonical induction variable.
2132 /// However, when it comes time to generate code, the maximum operation
2133 /// can be quite costly, especially if it's inside of an outer loop.
2135 /// This function solves this problem by detecting this type of loop and
2136 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
2137 /// the instructions for the maximum computation.
2139 ICmpInst *LoopStrengthReduce::OptimizeMax(Loop *L, ICmpInst *Cond,
2140 IVStrideUse* &CondUse) {
2141 // Check that the loop matches the pattern we're looking for.
2142 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
2143 Cond->getPredicate() != CmpInst::ICMP_NE)
2146 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
2147 if (!Sel || !Sel->hasOneUse()) return Cond;
2149 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
2150 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2152 const SCEV *One = SE->getIntegerSCEV(1, BackedgeTakenCount->getType());
2154 // Add one to the backedge-taken count to get the trip count.
2155 const SCEV *IterationCount = SE->getAddExpr(BackedgeTakenCount, One);
2157 // Check for a max calculation that matches the pattern.
2158 if (!isa<SCEVSMaxExpr>(IterationCount) && !isa<SCEVUMaxExpr>(IterationCount))
2160 const SCEVNAryExpr *Max = cast<SCEVNAryExpr>(IterationCount);
2161 if (Max != SE->getSCEV(Sel)) return Cond;
2163 // To handle a max with more than two operands, this optimization would
2164 // require additional checking and setup.
2165 if (Max->getNumOperands() != 2)
2168 const SCEV *MaxLHS = Max->getOperand(0);
2169 const SCEV *MaxRHS = Max->getOperand(1);
2170 if (!MaxLHS || MaxLHS != One) return Cond;
2172 // Check the relevant induction variable for conformance to
2174 const SCEV *IV = SE->getSCEV(Cond->getOperand(0));
2175 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2176 if (!AR || !AR->isAffine() ||
2177 AR->getStart() != One ||
2178 AR->getStepRecurrence(*SE) != One)
2181 assert(AR->getLoop() == L &&
2182 "Loop condition operand is an addrec in a different loop!");
2184 // Check the right operand of the select, and remember it, as it will
2185 // be used in the new comparison instruction.
2187 if (SE->getSCEV(Sel->getOperand(1)) == MaxRHS)
2188 NewRHS = Sel->getOperand(1);
2189 else if (SE->getSCEV(Sel->getOperand(2)) == MaxRHS)
2190 NewRHS = Sel->getOperand(2);
2191 if (!NewRHS) return Cond;
2193 // Determine the new comparison opcode. It may be signed or unsigned,
2194 // and the original comparison may be either equality or inequality.
2195 CmpInst::Predicate Pred =
2196 isa<SCEVSMaxExpr>(Max) ? CmpInst::ICMP_SLT : CmpInst::ICMP_ULT;
2197 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
2198 Pred = CmpInst::getInversePredicate(Pred);
2200 // Ok, everything looks ok to change the condition into an SLT or SGE and
2201 // delete the max calculation.
2203 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
2205 // Delete the max calculation instructions.
2206 Cond->replaceAllUsesWith(NewCond);
2207 CondUse->setUser(NewCond);
2208 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2209 Cond->eraseFromParent();
2210 Sel->eraseFromParent();
2211 if (Cmp->use_empty())
2212 Cmp->eraseFromParent();
2216 /// OptimizeShadowIV - If IV is used in a int-to-float cast
2217 /// inside the loop then try to eliminate the cast opeation.
2218 void LoopStrengthReduce::OptimizeShadowIV(Loop *L) {
2220 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
2221 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2224 for (unsigned Stride = 0, e = IU->StrideOrder.size(); Stride != e;
2226 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
2227 IU->IVUsesByStride.find(IU->StrideOrder[Stride]);
2228 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
2229 if (!isa<SCEVConstant>(SI->first))
2232 for (ilist<IVStrideUse>::iterator UI = SI->second->Users.begin(),
2233 E = SI->second->Users.end(); UI != E; /* empty */) {
2234 ilist<IVStrideUse>::iterator CandidateUI = UI;
2236 Instruction *ShadowUse = CandidateUI->getUser();
2237 const Type *DestTy = NULL;
2239 /* If shadow use is a int->float cast then insert a second IV
2240 to eliminate this cast.
2242 for (unsigned i = 0; i < n; ++i)
2248 for (unsigned i = 0; i < n; ++i, ++d)
2251 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser()))
2252 DestTy = UCast->getDestTy();
2253 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser()))
2254 DestTy = SCast->getDestTy();
2255 if (!DestTy) continue;
2258 // If target does not support DestTy natively then do not apply
2259 // this transformation.
2260 EVT DVT = TLI->getValueType(DestTy);
2261 if (!TLI->isTypeLegal(DVT)) continue;
2264 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
2266 if (PH->getNumIncomingValues() != 2) continue;
2268 const Type *SrcTy = PH->getType();
2269 int Mantissa = DestTy->getFPMantissaWidth();
2270 if (Mantissa == -1) continue;
2271 if ((int)SE->getTypeSizeInBits(SrcTy) > Mantissa)
2274 unsigned Entry, Latch;
2275 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
2283 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
2284 if (!Init) continue;
2285 Constant *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());
2287 BinaryOperator *Incr =
2288 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
2289 if (!Incr) continue;
2290 if (Incr->getOpcode() != Instruction::Add
2291 && Incr->getOpcode() != Instruction::Sub)
2294 /* Initialize new IV, double d = 0.0 in above example. */
2295 ConstantInt *C = NULL;
2296 if (Incr->getOperand(0) == PH)
2297 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
2298 else if (Incr->getOperand(1) == PH)
2299 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
2305 // Ignore negative constants, as the code below doesn't handle them
2306 // correctly. TODO: Remove this restriction.
2307 if (!C->getValue().isStrictlyPositive()) continue;
2309 /* Add new PHINode. */
2310 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);
2312 /* create new increment. '++d' in above example. */
2313 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
2314 BinaryOperator *NewIncr =
2315 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
2316 Instruction::FAdd : Instruction::FSub,
2317 NewPH, CFP, "IV.S.next.", Incr);
2319 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
2320 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
2322 /* Remove cast operation */
2323 ShadowUse->replaceAllUsesWith(NewPH);
2324 ShadowUse->eraseFromParent();
2331 /// OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
2332 /// uses in the loop, look to see if we can eliminate some, in favor of using
2333 /// common indvars for the different uses.
2334 void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
2335 // TODO: implement optzns here.
2337 OptimizeShadowIV(L);
2340 bool LoopStrengthReduce::StrideMightBeShared(const SCEV* Stride, Loop *L,
2342 int64_t SInt = cast<SCEVConstant>(Stride)->getValue()->getSExtValue();
2343 for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
2344 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
2345 IU->IVUsesByStride.find(IU->StrideOrder[i]);
2346 const SCEV *Share = SI->first;
2347 if (!isa<SCEVConstant>(SI->first) || Share == Stride)
2349 int64_t SSInt = cast<SCEVConstant>(Share)->getValue()->getSExtValue();
2351 return true; // This can definitely be reused.
2352 if (unsigned(abs64(SSInt)) < SInt || (SSInt % SInt) != 0)
2354 int64_t Scale = SSInt / SInt;
2355 bool AllUsesAreAddresses = true;
2356 bool AllUsesAreOutsideLoop = true;
2357 std::vector<BasedUser> UsersToProcess;
2358 const SCEV *CommonExprs = CollectIVUsers(SI->first, *SI->second, L,
2359 AllUsesAreAddresses,
2360 AllUsesAreOutsideLoop,
2362 if (AllUsesAreAddresses &&
2363 ValidScale(!CommonExprs->isZero(), Scale, UsersToProcess)) {
2366 // Any pre-inc iv use?
2367 IVUsersOfOneStride &StrideUses = *IU->IVUsesByStride[Share];
2368 for (ilist<IVStrideUse>::iterator I = StrideUses.Users.begin(),
2369 E = StrideUses.Users.end(); I != E; ++I) {
2370 if (!I->isUseOfPostIncrementedValue())
2378 /// isUsedByExitBranch - Return true if icmp is used by a loop terminating
2379 /// conditional branch or it's and / or with other conditions before being used
2380 /// as the condition.
2381 static bool isUsedByExitBranch(ICmpInst *Cond, Loop *L) {
2382 BasicBlock *CondBB = Cond->getParent();
2383 if (!L->isLoopExiting(CondBB))
2385 BranchInst *TermBr = dyn_cast<BranchInst>(CondBB->getTerminator());
2386 if (!TermBr || !TermBr->isConditional())
2389 Value *User = *Cond->use_begin();
2390 Instruction *UserInst = dyn_cast<Instruction>(User);
2392 (UserInst->getOpcode() == Instruction::And ||
2393 UserInst->getOpcode() == Instruction::Or)) {
2394 if (!UserInst->hasOneUse() || UserInst->getParent() != CondBB)
2396 User = *User->use_begin();
2397 UserInst = dyn_cast<Instruction>(User);
2399 return User == TermBr;
2402 static bool ShouldCountToZero(ICmpInst *Cond, IVStrideUse* &CondUse,
2403 ScalarEvolution *SE, Loop *L,
2404 const TargetLowering *TLI = 0) {
2405 if (!L->contains(Cond->getParent()))
2408 if (!isa<SCEVConstant>(CondUse->getOffset()))
2411 // Handle only tests for equality for the moment.
2412 if (!Cond->isEquality() || !Cond->hasOneUse())
2414 if (!isUsedByExitBranch(Cond, L))
2417 Value *CondOp0 = Cond->getOperand(0);
2418 const SCEV *IV = SE->getSCEV(CondOp0);
2419 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2420 if (!AR || !AR->isAffine())
2423 const SCEVConstant *SC = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE));
2424 if (!SC || SC->getValue()->getSExtValue() < 0)
2425 // If it's already counting down, don't do anything.
2428 // If the RHS of the comparison is not an loop invariant, the rewrite
2429 // cannot be done. Also bail out if it's already comparing against a zero.
2430 // If we are checking this before cmp stride optimization, check if it's
2431 // comparing against a already legal immediate.
2432 Value *RHS = Cond->getOperand(1);
2433 ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS);
2434 if (!L->isLoopInvariant(RHS) ||
2435 (RHSC && RHSC->isZero()) ||
2436 (RHSC && TLI && TLI->isLegalICmpImmediate(RHSC->getSExtValue())))
2439 // Make sure the IV is only used for counting. Value may be preinc or
2440 // postinc; 2 uses in either case.
2441 if (!CondOp0->hasNUses(2))
2447 /// OptimizeLoopTermCond - Change loop terminating condition to use the
2448 /// postinc iv when possible.
2449 void LoopStrengthReduce::OptimizeLoopTermCond(Loop *L) {
2450 BasicBlock *LatchBlock = L->getLoopLatch();
2451 bool LatchExit = L->isLoopExiting(LatchBlock);
2452 SmallVector<BasicBlock*, 8> ExitingBlocks;
2453 L->getExitingBlocks(ExitingBlocks);
2455 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
2456 BasicBlock *ExitingBlock = ExitingBlocks[i];
2458 // Finally, get the terminating condition for the loop if possible. If we
2459 // can, we want to change it to use a post-incremented version of its
2460 // induction variable, to allow coalescing the live ranges for the IV into
2461 // one register value.
2463 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2466 // FIXME: Overly conservative, termination condition could be an 'or' etc..
2467 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
2470 // Search IVUsesByStride to find Cond's IVUse if there is one.
2471 IVStrideUse *CondUse = 0;
2472 const SCEV *CondStride = 0;
2473 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2474 if (!FindIVUserForCond(Cond, CondUse, CondStride))
2477 // If the latch block is exiting and it's not a single block loop, it's
2478 // not safe to use postinc iv in other exiting blocks. FIXME: overly
2479 // conservative? How about icmp stride optimization?
2480 bool UsePostInc = !(e > 1 && LatchExit && ExitingBlock != LatchBlock);
2481 if (UsePostInc && ExitingBlock != LatchBlock) {
2482 if (!Cond->hasOneUse())
2483 // See below, we don't want the condition to be cloned.
2486 // If exiting block is the latch block, we know it's safe and profitable
2487 // to transform the icmp to use post-inc iv. Otherwise do so only if it
2488 // would not reuse another iv and its iv would be reused by other uses.
2489 // We are optimizing for the case where the icmp is the only use of the
2491 IVUsersOfOneStride &StrideUses = *IU->IVUsesByStride[CondStride];
2492 for (ilist<IVStrideUse>::iterator I = StrideUses.Users.begin(),
2493 E = StrideUses.Users.end(); I != E; ++I) {
2494 if (I->getUser() == Cond)
2496 if (!I->isUseOfPostIncrementedValue()) {
2503 // If iv for the stride might be shared and any of the users use pre-inc
2504 // iv might be used, then it's not safe to use post-inc iv.
2506 isa<SCEVConstant>(CondStride) &&
2507 StrideMightBeShared(CondStride, L, true))
2511 // If the trip count is computed in terms of a max (due to ScalarEvolution
2512 // being unable to find a sufficient guard, for example), change the loop
2513 // comparison to use SLT or ULT instead of NE.
2514 Cond = OptimizeMax(L, Cond, CondUse);
2516 // If possible, change stride and operands of the compare instruction to
2517 // eliminate one stride. However, avoid rewriting the compare instruction
2518 // with an iv of new stride if it's likely the new stride uses will be
2519 // rewritten using the stride of the compare instruction.
2520 if (ExitingBlock == LatchBlock && isa<SCEVConstant>(CondStride)) {
2521 // If the condition stride is a constant and it's the only use, we might
2522 // want to optimize it first by turning it to count toward zero.
2523 if (!StrideMightBeShared(CondStride, L, false) &&
2524 !ShouldCountToZero(Cond, CondUse, SE, L, TLI))
2525 Cond = ChangeCompareStride(L, Cond, CondUse, CondStride);
2531 DEBUG(errs() << " Change loop exiting icmp to use postinc iv: "
2534 // It's possible for the setcc instruction to be anywhere in the loop, and
2535 // possible for it to have multiple users. If it is not immediately before
2536 // the exiting block branch, move it.
2537 if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
2538 if (Cond->hasOneUse()) { // Condition has a single use, just move it.
2539 Cond->moveBefore(TermBr);
2541 // Otherwise, clone the terminating condition and insert into the
2543 Cond = cast<ICmpInst>(Cond->clone());
2544 Cond->setName(L->getHeader()->getName() + ".termcond");
2545 ExitingBlock->getInstList().insert(TermBr, Cond);
2547 // Clone the IVUse, as the old use still exists!
2548 IU->IVUsesByStride[CondStride]->addUser(CondUse->getOffset(), Cond,
2549 CondUse->getOperandValToReplace());
2550 CondUse = &IU->IVUsesByStride[CondStride]->Users.back();
2554 // If we get to here, we know that we can transform the setcc instruction to
2555 // use the post-incremented version of the IV, allowing us to coalesce the
2556 // live ranges for the IV correctly.
2557 CondUse->setOffset(SE->getMinusSCEV(CondUse->getOffset(), CondStride));
2558 CondUse->setIsUseOfPostIncrementedValue(true);
2565 bool LoopStrengthReduce::OptimizeLoopCountIVOfStride(const SCEV* &Stride,
2566 IVStrideUse* &CondUse,
2568 // If the only use is an icmp of a loop exiting conditional branch, then
2569 // attempt the optimization.
2570 BasedUser User = BasedUser(*CondUse, SE);
2571 assert(isa<ICmpInst>(User.Inst) && "Expecting an ICMPInst!");
2572 ICmpInst *Cond = cast<ICmpInst>(User.Inst);
2574 // Less strict check now that compare stride optimization is done.
2575 if (!ShouldCountToZero(Cond, CondUse, SE, L))
2578 Value *CondOp0 = Cond->getOperand(0);
2579 PHINode *PHIExpr = dyn_cast<PHINode>(CondOp0);
2582 // Value tested is postinc. Find the phi node.
2583 Incr = dyn_cast<BinaryOperator>(CondOp0);
2584 // FIXME: Just use User.OperandValToReplace here?
2585 if (!Incr || Incr->getOpcode() != Instruction::Add)
2588 PHIExpr = dyn_cast<PHINode>(Incr->getOperand(0));
2591 // 1 use for preinc value, the increment.
2592 if (!PHIExpr->hasOneUse())
2595 assert(isa<PHINode>(CondOp0) &&
2596 "Unexpected loop exiting counting instruction sequence!");
2597 PHIExpr = cast<PHINode>(CondOp0);
2598 // Value tested is preinc. Find the increment.
2599 // A CmpInst is not a BinaryOperator; we depend on this.
2600 Instruction::use_iterator UI = PHIExpr->use_begin();
2601 Incr = dyn_cast<BinaryOperator>(UI);
2603 Incr = dyn_cast<BinaryOperator>(++UI);
2604 // One use for postinc value, the phi. Unnecessarily conservative?
2605 if (!Incr || !Incr->hasOneUse() || Incr->getOpcode() != Instruction::Add)
2609 // Replace the increment with a decrement.
2610 DEBUG(errs() << "LSR: Examining use ");
2611 DEBUG(WriteAsOperand(errs(), CondOp0, /*PrintType=*/false));
2612 DEBUG(errs() << " in Inst: " << *Cond << '\n');
2613 BinaryOperator *Decr = BinaryOperator::Create(Instruction::Sub,
2614 Incr->getOperand(0), Incr->getOperand(1), "tmp", Incr);
2615 Incr->replaceAllUsesWith(Decr);
2616 Incr->eraseFromParent();
2618 // Substitute endval-startval for the original startval, and 0 for the
2619 // original endval. Since we're only testing for equality this is OK even
2620 // if the computation wraps around.
2621 BasicBlock *Preheader = L->getLoopPreheader();
2622 Instruction *PreInsertPt = Preheader->getTerminator();
2623 unsigned InBlock = L->contains(PHIExpr->getIncomingBlock(0)) ? 1 : 0;
2624 Value *StartVal = PHIExpr->getIncomingValue(InBlock);
2625 Value *EndVal = Cond->getOperand(1);
2626 DEBUG(errs() << " Optimize loop counting iv to count down ["
2627 << *EndVal << " .. " << *StartVal << "]\n");
2629 // FIXME: check for case where both are constant.
2630 Constant* Zero = ConstantInt::get(Cond->getOperand(1)->getType(), 0);
2631 BinaryOperator *NewStartVal = BinaryOperator::Create(Instruction::Sub,
2632 EndVal, StartVal, "tmp", PreInsertPt);
2633 PHIExpr->setIncomingValue(InBlock, NewStartVal);
2634 Cond->setOperand(1, Zero);
2635 DEBUG(errs() << " New icmp: " << *Cond << "\n");
2637 int64_t SInt = cast<SCEVConstant>(Stride)->getValue()->getSExtValue();
2638 const SCEV *NewStride = 0;
2640 for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
2641 const SCEV *OldStride = IU->StrideOrder[i];
2642 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(OldStride))
2643 if (SC->getValue()->getSExtValue() == -SInt) {
2645 NewStride = OldStride;
2651 NewStride = SE->getIntegerSCEV(-SInt, Stride->getType());
2652 IU->AddUser(NewStride, CondUse->getOffset(), Cond, Cond->getOperand(0));
2653 IU->IVUsesByStride[Stride]->removeUser(CondUse);
2655 CondUse = &IU->IVUsesByStride[NewStride]->Users.back();
2663 /// OptimizeLoopCountIV - If, after all sharing of IVs, the IV used for deciding
2664 /// when to exit the loop is used only for that purpose, try to rearrange things
2665 /// so it counts down to a test against zero.
2666 bool LoopStrengthReduce::OptimizeLoopCountIV(Loop *L) {
2667 bool ThisChanged = false;
2668 for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
2669 const SCEV *Stride = IU->StrideOrder[i];
2670 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
2671 IU->IVUsesByStride.find(Stride);
2672 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
2673 // FIXME: Generalize to non-affine IV's.
2674 if (!SI->first->isLoopInvariant(L))
2676 // If stride is a constant and it has an icmpinst use, check if we can
2677 // optimize the loop to count down.
2678 if (isa<SCEVConstant>(Stride) && SI->second->Users.size() == 1) {
2679 Instruction *User = SI->second->Users.begin()->getUser();
2680 if (!isa<ICmpInst>(User))
2682 const SCEV *CondStride = Stride;
2683 IVStrideUse *Use = &*SI->second->Users.begin();
2684 if (!OptimizeLoopCountIVOfStride(CondStride, Use, L))
2688 // Now check if it's possible to reuse this iv for other stride uses.
2689 for (unsigned j = 0, ee = IU->StrideOrder.size(); j != ee; ++j) {
2690 const SCEV *SStride = IU->StrideOrder[j];
2691 if (SStride == CondStride)
2693 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SII =
2694 IU->IVUsesByStride.find(SStride);
2695 assert(SII != IU->IVUsesByStride.end() && "Stride doesn't exist!");
2696 // FIXME: Generalize to non-affine IV's.
2697 if (!SII->first->isLoopInvariant(L))
2699 // FIXME: Rewrite other stride using CondStride.
2704 Changed |= ThisChanged;
2708 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager &LPM) {
2709 IU = &getAnalysis<IVUsers>();
2710 LI = &getAnalysis<LoopInfo>();
2711 DT = &getAnalysis<DominatorTree>();
2712 SE = &getAnalysis<ScalarEvolution>();
2715 // If LoopSimplify form is not available, stay out of trouble.
2716 if (!L->getLoopPreheader() || !L->getLoopLatch())
2719 if (!IU->IVUsesByStride.empty()) {
2720 DEBUG(errs() << "\nLSR on \"" << L->getHeader()->getParent()->getName()
2724 // Sort the StrideOrder so we process larger strides first.
2725 std::stable_sort(IU->StrideOrder.begin(), IU->StrideOrder.end(),
2728 // Optimize induction variables. Some indvar uses can be transformed to use
2729 // strides that will be needed for other purposes. A common example of this
2730 // is the exit test for the loop, which can often be rewritten to use the
2731 // computation of some other indvar to decide when to terminate the loop.
2734 // Change loop terminating condition to use the postinc iv when possible
2735 // and optimize loop terminating compare. FIXME: Move this after
2736 // StrengthReduceIVUsersOfStride?
2737 OptimizeLoopTermCond(L);
2739 // FIXME: We can shrink overlarge IV's here. e.g. if the code has
2740 // computation in i64 values and the target doesn't support i64, demote
2741 // the computation to 32-bit if safe.
2743 // FIXME: Attempt to reuse values across multiple IV's. In particular, we
2744 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should
2745 // be codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC.
2746 // Need to be careful that IV's are all the same type. Only works for
2747 // intptr_t indvars.
2749 // IVsByStride keeps IVs for one particular loop.
2750 assert(IVsByStride.empty() && "Stale entries in IVsByStride?");
2752 StrengthReduceIVUsers(L);
2754 // After all sharing is done, see if we can adjust the loop to test against
2755 // zero instead of counting up to a maximum. This is usually faster.
2756 OptimizeLoopCountIV(L);
2759 // We're done analyzing this loop; release all the state we built up for it.
2760 IVsByStride.clear();
2761 StrideNoReuse.clear();
2763 // Clean up after ourselves
2764 if (!DeadInsts.empty())
2765 DeleteTriviallyDeadInstructions();
2767 // At this point, it is worth checking to see if any recurrence PHIs are also
2768 // dead, so that we can remove them as well.
2769 DeleteDeadPHIs(L->getHeader());