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/LLVMContext.h"
28 #include "llvm/Type.h"
29 #include "llvm/DerivedTypes.h"
30 #include "llvm/Analysis/Dominators.h"
31 #include "llvm/Analysis/IVUsers.h"
32 #include "llvm/Analysis/LoopInfo.h"
33 #include "llvm/Analysis/LoopPass.h"
34 #include "llvm/Analysis/ScalarEvolutionExpander.h"
35 #include "llvm/Transforms/Utils/AddrModeMatcher.h"
36 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
37 #include "llvm/Transforms/Utils/Local.h"
38 #include "llvm/ADT/Statistic.h"
39 #include "llvm/Support/CFG.h"
40 #include "llvm/Support/Debug.h"
41 #include "llvm/Support/Compiler.h"
42 #include "llvm/Support/CommandLine.h"
43 #include "llvm/Support/ValueHandle.h"
44 #include "llvm/Target/TargetLowering.h"
48 STATISTIC(NumReduced , "Number of IV uses strength reduced");
49 STATISTIC(NumInserted, "Number of PHIs inserted");
50 STATISTIC(NumVariable, "Number of PHIs with variable strides");
51 STATISTIC(NumEliminated, "Number of strides eliminated");
52 STATISTIC(NumShadow, "Number of Shadow IVs optimized");
53 STATISTIC(NumImmSunk, "Number of common expr immediates sunk into uses");
54 STATISTIC(NumLoopCond, "Number of loop terminating conds optimized");
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.
67 struct VISIBILITY_HIDDEN IVExpr {
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 VISIBILITY_HIDDEN 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 VISIBILITY_HIDDEN 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 ICmpInst *ChangeCompareStride(Loop *L, ICmpInst *Cond,
136 IVStrideUse* &CondUse,
137 const SCEV *const * &CondStride);
139 void OptimizeIndvars(Loop *L);
140 void OptimizeLoopCountIV(Loop *L);
141 void OptimizeLoopTermCond(Loop *L);
143 /// OptimizeShadowIV - If IV is used in a int-to-float cast
144 /// inside the loop then try to eliminate the cast opeation.
145 void OptimizeShadowIV(Loop *L);
147 /// OptimizeMax - Rewrite the loop's terminating condition
148 /// if it uses a max computation.
149 ICmpInst *OptimizeMax(Loop *L, ICmpInst *Cond,
150 IVStrideUse* &CondUse);
152 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse,
153 const SCEV *const * &CondStride);
154 bool RequiresTypeConversion(const Type *Ty, const Type *NewTy);
155 const SCEV *CheckForIVReuse(bool, bool, bool, const SCEV *const&,
156 IVExpr&, const Type*,
157 const std::vector<BasedUser>& UsersToProcess);
158 bool ValidScale(bool, int64_t,
159 const std::vector<BasedUser>& UsersToProcess);
160 bool ValidOffset(bool, int64_t, int64_t,
161 const std::vector<BasedUser>& UsersToProcess);
162 const SCEV *CollectIVUsers(const SCEV *const &Stride,
163 IVUsersOfOneStride &Uses,
165 bool &AllUsesAreAddresses,
166 bool &AllUsesAreOutsideLoop,
167 std::vector<BasedUser> &UsersToProcess);
168 bool ShouldUseFullStrengthReductionMode(
169 const std::vector<BasedUser> &UsersToProcess,
171 bool AllUsesAreAddresses,
173 void PrepareToStrengthReduceFully(
174 std::vector<BasedUser> &UsersToProcess,
176 const SCEV *CommonExprs,
178 SCEVExpander &PreheaderRewriter);
179 void PrepareToStrengthReduceFromSmallerStride(
180 std::vector<BasedUser> &UsersToProcess,
182 const IVExpr &ReuseIV,
183 Instruction *PreInsertPt);
184 void PrepareToStrengthReduceWithNewPhi(
185 std::vector<BasedUser> &UsersToProcess,
187 const SCEV *CommonExprs,
189 Instruction *IVIncInsertPt,
191 SCEVExpander &PreheaderRewriter);
192 void StrengthReduceStridedIVUsers(const SCEV *const &Stride,
193 IVUsersOfOneStride &Uses,
195 void DeleteTriviallyDeadInstructions();
199 char LoopStrengthReduce::ID = 0;
200 static RegisterPass<LoopStrengthReduce>
201 X("loop-reduce", "Loop Strength Reduction");
203 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
204 return new LoopStrengthReduce(TLI);
207 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
208 /// specified set are trivially dead, delete them and see if this makes any of
209 /// their operands subsequently dead.
210 void LoopStrengthReduce::DeleteTriviallyDeadInstructions() {
211 if (DeadInsts.empty()) return;
213 while (!DeadInsts.empty()) {
214 Instruction *I = dyn_cast_or_null<Instruction>(DeadInsts.back());
215 DeadInsts.pop_back();
217 if (I == 0 || !isInstructionTriviallyDead(I))
220 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) {
221 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
224 DeadInsts.push_back(U);
228 I->eraseFromParent();
233 /// containsAddRecFromDifferentLoop - Determine whether expression S involves a
234 /// subexpression that is an AddRec from a loop other than L. An outer loop
235 /// of L is OK, but not an inner loop nor a disjoint loop.
236 static bool containsAddRecFromDifferentLoop(const SCEV *S, Loop *L) {
237 // This is very common, put it first.
238 if (isa<SCEVConstant>(S))
240 if (const SCEVCommutativeExpr *AE = dyn_cast<SCEVCommutativeExpr>(S)) {
241 for (unsigned int i=0; i< AE->getNumOperands(); i++)
242 if (containsAddRecFromDifferentLoop(AE->getOperand(i), L))
246 if (const SCEVAddRecExpr *AE = dyn_cast<SCEVAddRecExpr>(S)) {
247 if (const Loop *newLoop = AE->getLoop()) {
250 // if newLoop is an outer loop of L, this is OK.
251 if (!LoopInfoBase<BasicBlock>::isNotAlreadyContainedIn(L, newLoop))
256 if (const SCEVUDivExpr *DE = dyn_cast<SCEVUDivExpr>(S))
257 return containsAddRecFromDifferentLoop(DE->getLHS(), L) ||
258 containsAddRecFromDifferentLoop(DE->getRHS(), L);
260 // SCEVSDivExpr has been backed out temporarily, but will be back; we'll
261 // need this when it is.
262 if (const SCEVSDivExpr *DE = dyn_cast<SCEVSDivExpr>(S))
263 return containsAddRecFromDifferentLoop(DE->getLHS(), L) ||
264 containsAddRecFromDifferentLoop(DE->getRHS(), L);
266 if (const SCEVCastExpr *CE = dyn_cast<SCEVCastExpr>(S))
267 return containsAddRecFromDifferentLoop(CE->getOperand(), L);
271 /// isAddressUse - Returns true if the specified instruction is using the
272 /// specified value as an address.
273 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
274 bool isAddress = isa<LoadInst>(Inst);
275 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
276 if (SI->getOperand(1) == OperandVal)
278 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
279 // Addressing modes can also be folded into prefetches and a variety
281 switch (II->getIntrinsicID()) {
283 case Intrinsic::prefetch:
284 case Intrinsic::x86_sse2_loadu_dq:
285 case Intrinsic::x86_sse2_loadu_pd:
286 case Intrinsic::x86_sse_loadu_ps:
287 case Intrinsic::x86_sse_storeu_ps:
288 case Intrinsic::x86_sse2_storeu_pd:
289 case Intrinsic::x86_sse2_storeu_dq:
290 case Intrinsic::x86_sse2_storel_dq:
291 if (II->getOperand(1) == OperandVal)
299 /// getAccessType - Return the type of the memory being accessed.
300 static const Type *getAccessType(const Instruction *Inst) {
301 const Type *AccessTy = Inst->getType();
302 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
303 AccessTy = SI->getOperand(0)->getType();
304 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
305 // Addressing modes can also be folded into prefetches and a variety
307 switch (II->getIntrinsicID()) {
309 case Intrinsic::x86_sse_storeu_ps:
310 case Intrinsic::x86_sse2_storeu_pd:
311 case Intrinsic::x86_sse2_storeu_dq:
312 case Intrinsic::x86_sse2_storel_dq:
313 AccessTy = II->getOperand(1)->getType();
321 /// BasedUser - For a particular base value, keep information about how we've
322 /// partitioned the expression so far.
324 /// SE - The current ScalarEvolution object.
327 /// Base - The Base value for the PHI node that needs to be inserted for
328 /// this use. As the use is processed, information gets moved from this
329 /// field to the Imm field (below). BasedUser values are sorted by this
333 /// Inst - The instruction using the induction variable.
336 /// OperandValToReplace - The operand value of Inst to replace with the
338 Value *OperandValToReplace;
340 /// Imm - The immediate value that should be added to the base immediately
341 /// before Inst, because it will be folded into the imm field of the
342 /// instruction. This is also sometimes used for loop-variant values that
343 /// must be added inside the loop.
346 /// Phi - The induction variable that performs the striding that
347 /// should be used for this user.
350 // isUseOfPostIncrementedValue - True if this should use the
351 // post-incremented version of this IV, not the preincremented version.
352 // This can only be set in special cases, such as the terminating setcc
353 // instruction for a loop and uses outside the loop that are dominated by
355 bool isUseOfPostIncrementedValue;
357 BasedUser(IVStrideUse &IVSU, ScalarEvolution *se)
358 : SE(se), Base(IVSU.getOffset()), Inst(IVSU.getUser()),
359 OperandValToReplace(IVSU.getOperandValToReplace()),
360 Imm(SE->getIntegerSCEV(0, Base->getType())),
361 isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue()) {}
363 // Once we rewrite the code to insert the new IVs we want, update the
364 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
366 void RewriteInstructionToUseNewBase(const SCEV *const &NewBase,
367 Instruction *InsertPt,
368 SCEVExpander &Rewriter, Loop *L, Pass *P,
370 SmallVectorImpl<WeakVH> &DeadInsts);
372 Value *InsertCodeForBaseAtPosition(const SCEV *const &NewBase,
374 SCEVExpander &Rewriter,
375 Instruction *IP, Loop *L,
381 void BasedUser::dump() const {
382 cerr << " Base=" << *Base;
383 cerr << " Imm=" << *Imm;
384 cerr << " Inst: " << *Inst;
387 Value *BasedUser::InsertCodeForBaseAtPosition(const SCEV *const &NewBase,
389 SCEVExpander &Rewriter,
390 Instruction *IP, Loop *L,
392 // Figure out where we *really* want to insert this code. In particular, if
393 // the user is inside of a loop that is nested inside of L, we really don't
394 // want to insert this expression before the user, we'd rather pull it out as
395 // many loops as possible.
396 Instruction *BaseInsertPt = IP;
398 // Figure out the most-nested loop that IP is in.
399 Loop *InsertLoop = LI.getLoopFor(IP->getParent());
401 // If InsertLoop is not L, and InsertLoop is nested inside of L, figure out
402 // the preheader of the outer-most loop where NewBase is not loop invariant.
403 if (L->contains(IP->getParent()))
404 while (InsertLoop && NewBase->isLoopInvariant(InsertLoop)) {
405 BaseInsertPt = InsertLoop->getLoopPreheader()->getTerminator();
406 InsertLoop = InsertLoop->getParentLoop();
409 Value *Base = Rewriter.expandCodeFor(NewBase, 0, BaseInsertPt);
411 const SCEV *NewValSCEV = SE->getUnknown(Base);
413 // Always emit the immediate into the same block as the user.
414 NewValSCEV = SE->getAddExpr(NewValSCEV, Imm);
416 return Rewriter.expandCodeFor(NewValSCEV, Ty, IP);
420 // Once we rewrite the code to insert the new IVs we want, update the
421 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
422 // to it. NewBasePt is the last instruction which contributes to the
423 // value of NewBase in the case that it's a diffferent instruction from
424 // the PHI that NewBase is computed from, or null otherwise.
426 void BasedUser::RewriteInstructionToUseNewBase(const SCEV *const &NewBase,
427 Instruction *NewBasePt,
428 SCEVExpander &Rewriter, Loop *L, Pass *P,
430 SmallVectorImpl<WeakVH> &DeadInsts) {
431 if (!isa<PHINode>(Inst)) {
432 // By default, insert code at the user instruction.
433 BasicBlock::iterator InsertPt = Inst;
435 // However, if the Operand is itself an instruction, the (potentially
436 // complex) inserted code may be shared by many users. Because of this, we
437 // want to emit code for the computation of the operand right before its old
438 // computation. This is usually safe, because we obviously used to use the
439 // computation when it was computed in its current block. However, in some
440 // cases (e.g. use of a post-incremented induction variable) the NewBase
441 // value will be pinned to live somewhere after the original computation.
442 // In this case, we have to back off.
444 // If this is a use outside the loop (which means after, since it is based
445 // on a loop indvar) we use the post-incremented value, so that we don't
446 // artificially make the preinc value live out the bottom of the loop.
447 if (!isUseOfPostIncrementedValue && L->contains(Inst->getParent())) {
448 if (NewBasePt && isa<PHINode>(OperandValToReplace)) {
449 InsertPt = NewBasePt;
451 } else if (Instruction *OpInst
452 = dyn_cast<Instruction>(OperandValToReplace)) {
454 while (isa<PHINode>(InsertPt)) ++InsertPt;
457 Value *NewVal = InsertCodeForBaseAtPosition(NewBase,
458 OperandValToReplace->getType(),
459 Rewriter, InsertPt, L, LI);
460 // Replace the use of the operand Value with the new Phi we just created.
461 Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
463 DOUT << " Replacing with ";
464 DEBUG(WriteAsOperand(*DOUT, NewVal, /*PrintType=*/false));
465 DOUT << ", which has value " << *NewBase << " plus IMM " << *Imm << "\n";
469 // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm
470 // expression into each operand block that uses it. Note that PHI nodes can
471 // have multiple entries for the same predecessor. We use a map to make sure
472 // that a PHI node only has a single Value* for each predecessor (which also
473 // prevents us from inserting duplicate code in some blocks).
474 DenseMap<BasicBlock*, Value*> InsertedCode;
475 PHINode *PN = cast<PHINode>(Inst);
476 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
477 if (PN->getIncomingValue(i) == OperandValToReplace) {
478 // If the original expression is outside the loop, put the replacement
479 // code in the same place as the original expression,
480 // which need not be an immediate predecessor of this PHI. This way we
481 // need only one copy of it even if it is referenced multiple times in
482 // the PHI. We don't do this when the original expression is inside the
483 // loop because multiple copies sometimes do useful sinking of code in
485 Instruction *OldLoc = dyn_cast<Instruction>(OperandValToReplace);
486 if (L->contains(OldLoc->getParent())) {
487 // If this is a critical edge, split the edge so that we do not insert
488 // the code on all predecessor/successor paths. We do this unless this
489 // is the canonical backedge for this loop, as this can make some
490 // inserted code be in an illegal position.
491 BasicBlock *PHIPred = PN->getIncomingBlock(i);
492 if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 &&
493 (PN->getParent() != L->getHeader() || !L->contains(PHIPred))) {
495 // First step, split the critical edge.
496 SplitCriticalEdge(PHIPred, PN->getParent(), P, false);
498 // Next step: move the basic block. In particular, if the PHI node
499 // is outside of the loop, and PredTI is in the loop, we want to
500 // move the block to be immediately before the PHI block, not
501 // immediately after PredTI.
502 if (L->contains(PHIPred) && !L->contains(PN->getParent())) {
503 BasicBlock *NewBB = PN->getIncomingBlock(i);
504 NewBB->moveBefore(PN->getParent());
507 // Splitting the edge can reduce the number of PHI entries we have.
508 e = PN->getNumIncomingValues();
511 Value *&Code = InsertedCode[PN->getIncomingBlock(i)];
513 // Insert the code into the end of the predecessor block.
514 Instruction *InsertPt = (L->contains(OldLoc->getParent())) ?
515 PN->getIncomingBlock(i)->getTerminator() :
516 OldLoc->getParent()->getTerminator();
517 Code = InsertCodeForBaseAtPosition(NewBase, PN->getType(),
518 Rewriter, InsertPt, L, LI);
520 DOUT << " Changing PHI use to ";
521 DEBUG(WriteAsOperand(*DOUT, Code, /*PrintType=*/false));
522 DOUT << ", which has value " << *NewBase << " plus IMM " << *Imm << "\n";
525 // Replace the use of the operand Value with the new Phi we just created.
526 PN->setIncomingValue(i, Code);
531 // PHI node might have become a constant value after SplitCriticalEdge.
532 DeadInsts.push_back(Inst);
536 /// fitsInAddressMode - Return true if V can be subsumed within an addressing
537 /// mode, and does not need to be put in a register first.
538 static bool fitsInAddressMode(const SCEV *const &V, const Type *AccessTy,
539 const TargetLowering *TLI, bool HasBaseReg) {
540 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
541 int64_t VC = SC->getValue()->getSExtValue();
543 TargetLowering::AddrMode AM;
545 AM.HasBaseReg = HasBaseReg;
546 return TLI->isLegalAddressingMode(AM, AccessTy);
548 // Defaults to PPC. PPC allows a sign-extended 16-bit immediate field.
549 return (VC > -(1 << 16) && VC < (1 << 16)-1);
553 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
554 if (GlobalValue *GV = dyn_cast<GlobalValue>(SU->getValue())) {
556 TargetLowering::AddrMode AM;
558 AM.HasBaseReg = HasBaseReg;
559 return TLI->isLegalAddressingMode(AM, AccessTy);
561 // Default: assume global addresses are not legal.
568 /// MoveLoopVariantsToImmediateField - Move any subexpressions from Val that are
569 /// loop varying to the Imm operand.
570 static void MoveLoopVariantsToImmediateField(const SCEV *&Val, const SCEV *&Imm,
571 Loop *L, ScalarEvolution *SE) {
572 if (Val->isLoopInvariant(L)) return; // Nothing to do.
574 if (const SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
575 SmallVector<const SCEV *, 4> NewOps;
576 NewOps.reserve(SAE->getNumOperands());
578 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
579 if (!SAE->getOperand(i)->isLoopInvariant(L)) {
580 // If this is a loop-variant expression, it must stay in the immediate
581 // field of the expression.
582 Imm = SE->getAddExpr(Imm, SAE->getOperand(i));
584 NewOps.push_back(SAE->getOperand(i));
588 Val = SE->getIntegerSCEV(0, Val->getType());
590 Val = SE->getAddExpr(NewOps);
591 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
592 // Try to pull immediates out of the start value of nested addrec's.
593 const SCEV *Start = SARE->getStart();
594 MoveLoopVariantsToImmediateField(Start, Imm, L, SE);
596 SmallVector<const SCEV *, 4> Ops(SARE->op_begin(), SARE->op_end());
598 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
600 // Otherwise, all of Val is variant, move the whole thing over.
601 Imm = SE->getAddExpr(Imm, Val);
602 Val = SE->getIntegerSCEV(0, Val->getType());
607 /// MoveImmediateValues - Look at Val, and pull out any additions of constants
608 /// that can fit into the immediate field of instructions in the target.
609 /// Accumulate these immediate values into the Imm value.
610 static void MoveImmediateValues(const TargetLowering *TLI,
611 const Type *AccessTy,
612 const SCEV *&Val, const SCEV *&Imm,
613 bool isAddress, Loop *L,
614 ScalarEvolution *SE) {
615 if (const SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
616 SmallVector<const SCEV *, 4> NewOps;
617 NewOps.reserve(SAE->getNumOperands());
619 for (unsigned i = 0; i != SAE->getNumOperands(); ++i) {
620 const SCEV *NewOp = SAE->getOperand(i);
621 MoveImmediateValues(TLI, AccessTy, NewOp, Imm, isAddress, L, SE);
623 if (!NewOp->isLoopInvariant(L)) {
624 // If this is a loop-variant expression, it must stay in the immediate
625 // field of the expression.
626 Imm = SE->getAddExpr(Imm, NewOp);
628 NewOps.push_back(NewOp);
633 Val = SE->getIntegerSCEV(0, Val->getType());
635 Val = SE->getAddExpr(NewOps);
637 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
638 // Try to pull immediates out of the start value of nested addrec's.
639 const SCEV *Start = SARE->getStart();
640 MoveImmediateValues(TLI, AccessTy, Start, Imm, isAddress, L, SE);
642 if (Start != SARE->getStart()) {
643 SmallVector<const SCEV *, 4> Ops(SARE->op_begin(), SARE->op_end());
645 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
648 } else if (const SCEVMulExpr *SME = dyn_cast<SCEVMulExpr>(Val)) {
649 // Transform "8 * (4 + v)" -> "32 + 8*V" if "32" fits in the immed field.
651 fitsInAddressMode(SME->getOperand(0), AccessTy, TLI, false) &&
652 SME->getNumOperands() == 2 && SME->isLoopInvariant(L)) {
654 const SCEV *SubImm = SE->getIntegerSCEV(0, Val->getType());
655 const SCEV *NewOp = SME->getOperand(1);
656 MoveImmediateValues(TLI, AccessTy, NewOp, SubImm, isAddress, L, SE);
658 // If we extracted something out of the subexpressions, see if we can
660 if (NewOp != SME->getOperand(1)) {
661 // Scale SubImm up by "8". If the result is a target constant, we are
663 SubImm = SE->getMulExpr(SubImm, SME->getOperand(0));
664 if (fitsInAddressMode(SubImm, AccessTy, TLI, false)) {
665 // Accumulate the immediate.
666 Imm = SE->getAddExpr(Imm, SubImm);
668 // Update what is left of 'Val'.
669 Val = SE->getMulExpr(SME->getOperand(0), NewOp);
676 // Loop-variant expressions must stay in the immediate field of the
678 if ((isAddress && fitsInAddressMode(Val, AccessTy, TLI, false)) ||
679 !Val->isLoopInvariant(L)) {
680 Imm = SE->getAddExpr(Imm, Val);
681 Val = SE->getIntegerSCEV(0, Val->getType());
685 // Otherwise, no immediates to move.
688 static void MoveImmediateValues(const TargetLowering *TLI,
690 const SCEV *&Val, const SCEV *&Imm,
691 bool isAddress, Loop *L,
692 ScalarEvolution *SE) {
693 const Type *AccessTy = getAccessType(User);
694 MoveImmediateValues(TLI, AccessTy, Val, Imm, isAddress, L, SE);
697 /// SeparateSubExprs - Decompose Expr into all of the subexpressions that are
698 /// added together. This is used to reassociate common addition subexprs
699 /// together for maximal sharing when rewriting bases.
700 static void SeparateSubExprs(SmallVector<const SCEV *, 16> &SubExprs,
702 ScalarEvolution *SE) {
703 if (const SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) {
704 for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
705 SeparateSubExprs(SubExprs, AE->getOperand(j), SE);
706 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) {
707 const SCEV *Zero = SE->getIntegerSCEV(0, Expr->getType());
708 if (SARE->getOperand(0) == Zero) {
709 SubExprs.push_back(Expr);
711 // Compute the addrec with zero as its base.
712 SmallVector<const SCEV *, 4> Ops(SARE->op_begin(), SARE->op_end());
713 Ops[0] = Zero; // Start with zero base.
714 SubExprs.push_back(SE->getAddRecExpr(Ops, SARE->getLoop()));
717 SeparateSubExprs(SubExprs, SARE->getOperand(0), SE);
719 } else if (!Expr->isZero()) {
721 SubExprs.push_back(Expr);
725 // This is logically local to the following function, but C++ says we have
726 // to make it file scope.
727 struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; };
729 /// RemoveCommonExpressionsFromUseBases - Look through all of the Bases of all
730 /// the Uses, removing any common subexpressions, except that if all such
731 /// subexpressions can be folded into an addressing mode for all uses inside
732 /// the loop (this case is referred to as "free" in comments herein) we do
733 /// not remove anything. This looks for things like (a+b+c) and
734 /// (a+c+d) and computes the common (a+c) subexpression. The common expression
735 /// is *removed* from the Bases and returned.
737 RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses,
738 ScalarEvolution *SE, Loop *L,
739 const TargetLowering *TLI) {
740 unsigned NumUses = Uses.size();
742 // Only one use? This is a very common case, so we handle it specially and
744 const SCEV *Zero = SE->getIntegerSCEV(0, Uses[0].Base->getType());
745 const SCEV *Result = Zero;
746 const SCEV *FreeResult = Zero;
748 // If the use is inside the loop, use its base, regardless of what it is:
749 // it is clearly shared across all the IV's. If the use is outside the loop
750 // (which means after it) we don't want to factor anything *into* the loop,
751 // so just use 0 as the base.
752 if (L->contains(Uses[0].Inst->getParent()))
753 std::swap(Result, Uses[0].Base);
757 // To find common subexpressions, count how many of Uses use each expression.
758 // If any subexpressions are used Uses.size() times, they are common.
759 // Also track whether all uses of each expression can be moved into an
760 // an addressing mode "for free"; such expressions are left within the loop.
761 // struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; };
762 std::map<const SCEV *, SubExprUseData> SubExpressionUseData;
764 // UniqueSubExprs - Keep track of all of the subexpressions we see in the
765 // order we see them.
766 SmallVector<const SCEV *, 16> UniqueSubExprs;
768 SmallVector<const SCEV *, 16> SubExprs;
769 unsigned NumUsesInsideLoop = 0;
770 for (unsigned i = 0; i != NumUses; ++i) {
771 // If the user is outside the loop, just ignore it for base computation.
772 // Since the user is outside the loop, it must be *after* the loop (if it
773 // were before, it could not be based on the loop IV). We don't want users
774 // after the loop to affect base computation of values *inside* the loop,
775 // because we can always add their offsets to the result IV after the loop
776 // is done, ensuring we get good code inside the loop.
777 if (!L->contains(Uses[i].Inst->getParent()))
781 // If the base is zero (which is common), return zero now, there are no
783 if (Uses[i].Base == Zero) return Zero;
785 // If this use is as an address we may be able to put CSEs in the addressing
786 // mode rather than hoisting them.
787 bool isAddrUse = isAddressUse(Uses[i].Inst, Uses[i].OperandValToReplace);
788 // We may need the AccessTy below, but only when isAddrUse, so compute it
789 // only in that case.
790 const Type *AccessTy = 0;
792 AccessTy = getAccessType(Uses[i].Inst);
794 // Split the expression into subexprs.
795 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
796 // Add one to SubExpressionUseData.Count for each subexpr present, and
797 // if the subexpr is not a valid immediate within an addressing mode use,
798 // set SubExpressionUseData.notAllUsesAreFree. We definitely want to
799 // hoist these out of the loop (if they are common to all uses).
800 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) {
801 if (++SubExpressionUseData[SubExprs[j]].Count == 1)
802 UniqueSubExprs.push_back(SubExprs[j]);
803 if (!isAddrUse || !fitsInAddressMode(SubExprs[j], AccessTy, TLI, false))
804 SubExpressionUseData[SubExprs[j]].notAllUsesAreFree = true;
809 // Now that we know how many times each is used, build Result. Iterate over
810 // UniqueSubexprs so that we have a stable ordering.
811 for (unsigned i = 0, e = UniqueSubExprs.size(); i != e; ++i) {
812 std::map<const SCEV *, SubExprUseData>::iterator I =
813 SubExpressionUseData.find(UniqueSubExprs[i]);
814 assert(I != SubExpressionUseData.end() && "Entry not found?");
815 if (I->second.Count == NumUsesInsideLoop) { // Found CSE!
816 if (I->second.notAllUsesAreFree)
817 Result = SE->getAddExpr(Result, I->first);
819 FreeResult = SE->getAddExpr(FreeResult, I->first);
821 // Remove non-cse's from SubExpressionUseData.
822 SubExpressionUseData.erase(I);
825 if (FreeResult != Zero) {
826 // We have some subexpressions that can be subsumed into addressing
827 // modes in every use inside the loop. However, it's possible that
828 // there are so many of them that the combined FreeResult cannot
829 // be subsumed, or that the target cannot handle both a FreeResult
830 // and a Result in the same instruction (for example because it would
831 // require too many registers). Check this.
832 for (unsigned i=0; i<NumUses; ++i) {
833 if (!L->contains(Uses[i].Inst->getParent()))
835 // We know this is an addressing mode use; if there are any uses that
836 // are not, FreeResult would be Zero.
837 const Type *AccessTy = getAccessType(Uses[i].Inst);
838 if (!fitsInAddressMode(FreeResult, AccessTy, TLI, Result!=Zero)) {
839 // FIXME: could split up FreeResult into pieces here, some hoisted
840 // and some not. There is no obvious advantage to this.
841 Result = SE->getAddExpr(Result, FreeResult);
848 // If we found no CSE's, return now.
849 if (Result == Zero) return Result;
851 // If we still have a FreeResult, remove its subexpressions from
852 // SubExpressionUseData. This means they will remain in the use Bases.
853 if (FreeResult != Zero) {
854 SeparateSubExprs(SubExprs, FreeResult, SE);
855 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) {
856 std::map<const SCEV *, SubExprUseData>::iterator I =
857 SubExpressionUseData.find(SubExprs[j]);
858 SubExpressionUseData.erase(I);
863 // Otherwise, remove all of the CSE's we found from each of the base values.
864 for (unsigned i = 0; i != NumUses; ++i) {
865 // Uses outside the loop don't necessarily include the common base, but
866 // the final IV value coming into those uses does. Instead of trying to
867 // remove the pieces of the common base, which might not be there,
868 // subtract off the base to compensate for this.
869 if (!L->contains(Uses[i].Inst->getParent())) {
870 Uses[i].Base = SE->getMinusSCEV(Uses[i].Base, Result);
874 // Split the expression into subexprs.
875 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
877 // Remove any common subexpressions.
878 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
879 if (SubExpressionUseData.count(SubExprs[j])) {
880 SubExprs.erase(SubExprs.begin()+j);
884 // Finally, add the non-shared expressions together.
885 if (SubExprs.empty())
888 Uses[i].Base = SE->getAddExpr(SubExprs);
895 /// ValidScale - Check whether the given Scale is valid for all loads and
896 /// stores in UsersToProcess.
898 bool LoopStrengthReduce::ValidScale(bool HasBaseReg, int64_t Scale,
899 const std::vector<BasedUser>& UsersToProcess) {
903 for (unsigned i = 0, e = UsersToProcess.size(); i!=e; ++i) {
904 // If this is a load or other access, pass the type of the access in.
905 const Type *AccessTy = Type::VoidTy;
906 if (isAddressUse(UsersToProcess[i].Inst,
907 UsersToProcess[i].OperandValToReplace))
908 AccessTy = getAccessType(UsersToProcess[i].Inst);
909 else if (isa<PHINode>(UsersToProcess[i].Inst))
912 TargetLowering::AddrMode AM;
913 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
914 AM.BaseOffs = SC->getValue()->getSExtValue();
915 AM.HasBaseReg = HasBaseReg || !UsersToProcess[i].Base->isZero();
918 // If load[imm+r*scale] is illegal, bail out.
919 if (!TLI->isLegalAddressingMode(AM, AccessTy))
925 /// ValidOffset - Check whether the given Offset is valid for all loads and
926 /// stores in UsersToProcess.
928 bool LoopStrengthReduce::ValidOffset(bool HasBaseReg,
931 const std::vector<BasedUser>& UsersToProcess) {
935 for (unsigned i=0, e = UsersToProcess.size(); i!=e; ++i) {
936 // If this is a load or other access, pass the type of the access in.
937 const Type *AccessTy = Type::VoidTy;
938 if (isAddressUse(UsersToProcess[i].Inst,
939 UsersToProcess[i].OperandValToReplace))
940 AccessTy = getAccessType(UsersToProcess[i].Inst);
941 else if (isa<PHINode>(UsersToProcess[i].Inst))
944 TargetLowering::AddrMode AM;
945 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
946 AM.BaseOffs = SC->getValue()->getSExtValue();
947 AM.BaseOffs = (uint64_t)AM.BaseOffs + (uint64_t)Offset;
948 AM.HasBaseReg = HasBaseReg || !UsersToProcess[i].Base->isZero();
951 // If load[imm+r*scale] is illegal, bail out.
952 if (!TLI->isLegalAddressingMode(AM, AccessTy))
958 /// RequiresTypeConversion - Returns true if converting Ty1 to Ty2 is not
960 bool LoopStrengthReduce::RequiresTypeConversion(const Type *Ty1,
964 Ty1 = SE->getEffectiveSCEVType(Ty1);
965 Ty2 = SE->getEffectiveSCEVType(Ty2);
968 if (Ty1->canLosslesslyBitCastTo(Ty2))
970 if (TLI && TLI->isTruncateFree(Ty1, Ty2))
975 /// CheckForIVReuse - Returns the multiple if the stride is the multiple
976 /// of a previous stride and it is a legal value for the target addressing
977 /// mode scale component and optional base reg. This allows the users of
978 /// this stride to be rewritten as prev iv * factor. It returns 0 if no
979 /// reuse is possible. Factors can be negative on same targets, e.g. ARM.
981 /// If all uses are outside the loop, we don't require that all multiplies
982 /// be folded into the addressing mode, nor even that the factor be constant;
983 /// a multiply (executed once) outside the loop is better than another IV
984 /// within. Well, usually.
985 const SCEV *LoopStrengthReduce::CheckForIVReuse(bool HasBaseReg,
986 bool AllUsesAreAddresses,
987 bool AllUsesAreOutsideLoop,
988 const SCEV *const &Stride,
989 IVExpr &IV, const Type *Ty,
990 const std::vector<BasedUser>& UsersToProcess) {
991 if (StrideNoReuse.count(Stride))
992 return SE->getIntegerSCEV(0, Stride->getType());
994 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Stride)) {
995 int64_t SInt = SC->getValue()->getSExtValue();
996 for (unsigned NewStride = 0, e = IU->StrideOrder.size();
997 NewStride != e; ++NewStride) {
998 std::map<const SCEV *, IVsOfOneStride>::iterator SI =
999 IVsByStride.find(IU->StrideOrder[NewStride]);
1000 if (SI == IVsByStride.end() || !isa<SCEVConstant>(SI->first) ||
1001 StrideNoReuse.count(SI->first))
1003 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1004 if (SI->first != Stride &&
1005 (unsigned(abs64(SInt)) < SSInt || (SInt % SSInt) != 0))
1007 int64_t Scale = SInt / SSInt;
1008 // Check that this stride is valid for all the types used for loads and
1009 // stores; if it can be used for some and not others, we might as well use
1010 // the original stride everywhere, since we have to create the IV for it
1011 // anyway. If the scale is 1, then we don't need to worry about folding
1014 (AllUsesAreAddresses &&
1015 ValidScale(HasBaseReg, Scale, UsersToProcess))) {
1016 // Prefer to reuse an IV with a base of zero.
1017 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1018 IE = SI->second.IVs.end(); II != IE; ++II)
1019 // Only reuse previous IV if it would not require a type conversion
1020 // and if the base difference can be folded.
1021 if (II->Base->isZero() &&
1022 !RequiresTypeConversion(II->Base->getType(), Ty)) {
1024 return SE->getIntegerSCEV(Scale, Stride->getType());
1026 // Otherwise, settle for an IV with a foldable base.
1027 if (AllUsesAreAddresses)
1028 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1029 IE = SI->second.IVs.end(); II != IE; ++II)
1030 // Only reuse previous IV if it would not require a type conversion
1031 // and if the base difference can be folded.
1032 if (SE->getEffectiveSCEVType(II->Base->getType()) ==
1033 SE->getEffectiveSCEVType(Ty) &&
1034 isa<SCEVConstant>(II->Base)) {
1036 cast<SCEVConstant>(II->Base)->getValue()->getSExtValue();
1037 if (Base > INT32_MIN && Base <= INT32_MAX &&
1038 ValidOffset(HasBaseReg, -Base * Scale,
1039 Scale, UsersToProcess)) {
1041 return SE->getIntegerSCEV(Scale, Stride->getType());
1046 } else if (AllUsesAreOutsideLoop) {
1047 // Accept nonconstant strides here; it is really really right to substitute
1048 // an existing IV if we can.
1049 for (unsigned NewStride = 0, e = IU->StrideOrder.size();
1050 NewStride != e; ++NewStride) {
1051 std::map<const SCEV *, IVsOfOneStride>::iterator SI =
1052 IVsByStride.find(IU->StrideOrder[NewStride]);
1053 if (SI == IVsByStride.end() || !isa<SCEVConstant>(SI->first))
1055 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1056 if (SI->first != Stride && SSInt != 1)
1058 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1059 IE = SI->second.IVs.end(); II != IE; ++II)
1060 // Accept nonzero base here.
1061 // Only reuse previous IV if it would not require a type conversion.
1062 if (!RequiresTypeConversion(II->Base->getType(), Ty)) {
1067 // Special case, old IV is -1*x and this one is x. Can treat this one as
1069 for (unsigned NewStride = 0, e = IU->StrideOrder.size();
1070 NewStride != e; ++NewStride) {
1071 std::map<const SCEV *, IVsOfOneStride>::iterator SI =
1072 IVsByStride.find(IU->StrideOrder[NewStride]);
1073 if (SI == IVsByStride.end())
1075 if (const SCEVMulExpr *ME = dyn_cast<SCEVMulExpr>(SI->first))
1076 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(ME->getOperand(0)))
1077 if (Stride == ME->getOperand(1) &&
1078 SC->getValue()->getSExtValue() == -1LL)
1079 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1080 IE = SI->second.IVs.end(); II != IE; ++II)
1081 // Accept nonzero base here.
1082 // Only reuse previous IV if it would not require type conversion.
1083 if (!RequiresTypeConversion(II->Base->getType(), Ty)) {
1085 return SE->getIntegerSCEV(-1LL, Stride->getType());
1089 return SE->getIntegerSCEV(0, Stride->getType());
1092 /// PartitionByIsUseOfPostIncrementedValue - Simple boolean predicate that
1093 /// returns true if Val's isUseOfPostIncrementedValue is true.
1094 static bool PartitionByIsUseOfPostIncrementedValue(const BasedUser &Val) {
1095 return Val.isUseOfPostIncrementedValue;
1098 /// isNonConstantNegative - Return true if the specified scev is negated, but
1100 static bool isNonConstantNegative(const SCEV *const &Expr) {
1101 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Expr);
1102 if (!Mul) return false;
1104 // If there is a constant factor, it will be first.
1105 const SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
1106 if (!SC) return false;
1108 // Return true if the value is negative, this matches things like (-42 * V).
1109 return SC->getValue()->getValue().isNegative();
1112 /// CollectIVUsers - Transform our list of users and offsets to a bit more
1113 /// complex table. In this new vector, each 'BasedUser' contains 'Base', the base
1114 /// of the strided accesses, as well as the old information from Uses. We
1115 /// progressively move information from the Base field to the Imm field, until
1116 /// we eventually have the full access expression to rewrite the use.
1117 const SCEV *LoopStrengthReduce::CollectIVUsers(const SCEV *const &Stride,
1118 IVUsersOfOneStride &Uses,
1120 bool &AllUsesAreAddresses,
1121 bool &AllUsesAreOutsideLoop,
1122 std::vector<BasedUser> &UsersToProcess) {
1123 // FIXME: Generalize to non-affine IV's.
1124 if (!Stride->isLoopInvariant(L))
1125 return SE->getIntegerSCEV(0, Stride->getType());
1127 UsersToProcess.reserve(Uses.Users.size());
1128 for (ilist<IVStrideUse>::iterator I = Uses.Users.begin(),
1129 E = Uses.Users.end(); I != E; ++I) {
1130 UsersToProcess.push_back(BasedUser(*I, SE));
1132 // Move any loop variant operands from the offset field to the immediate
1133 // field of the use, so that we don't try to use something before it is
1135 MoveLoopVariantsToImmediateField(UsersToProcess.back().Base,
1136 UsersToProcess.back().Imm, L, SE);
1137 assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
1138 "Base value is not loop invariant!");
1141 // We now have a whole bunch of uses of like-strided induction variables, but
1142 // they might all have different bases. We want to emit one PHI node for this
1143 // stride which we fold as many common expressions (between the IVs) into as
1144 // possible. Start by identifying the common expressions in the base values
1145 // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
1146 // "A+B"), emit it to the preheader, then remove the expression from the
1147 // UsersToProcess base values.
1148 const SCEV *CommonExprs =
1149 RemoveCommonExpressionsFromUseBases(UsersToProcess, SE, L, TLI);
1151 // Next, figure out what we can represent in the immediate fields of
1152 // instructions. If we can represent anything there, move it to the imm
1153 // fields of the BasedUsers. We do this so that it increases the commonality
1154 // of the remaining uses.
1155 unsigned NumPHI = 0;
1156 bool HasAddress = false;
1157 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1158 // If the user is not in the current loop, this means it is using the exit
1159 // value of the IV. Do not put anything in the base, make sure it's all in
1160 // the immediate field to allow as much factoring as possible.
1161 if (!L->contains(UsersToProcess[i].Inst->getParent())) {
1162 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm,
1163 UsersToProcess[i].Base);
1164 UsersToProcess[i].Base =
1165 SE->getIntegerSCEV(0, UsersToProcess[i].Base->getType());
1167 // Not all uses are outside the loop.
1168 AllUsesAreOutsideLoop = false;
1170 // Addressing modes can be folded into loads and stores. Be careful that
1171 // the store is through the expression, not of the expression though.
1173 bool isAddress = isAddressUse(UsersToProcess[i].Inst,
1174 UsersToProcess[i].OperandValToReplace);
1175 if (isa<PHINode>(UsersToProcess[i].Inst)) {
1183 // If this use isn't an address, then not all uses are addresses.
1184 if (!isAddress && !isPHI)
1185 AllUsesAreAddresses = false;
1187 MoveImmediateValues(TLI, UsersToProcess[i].Inst, UsersToProcess[i].Base,
1188 UsersToProcess[i].Imm, isAddress, L, SE);
1192 // If one of the use is a PHI node and all other uses are addresses, still
1193 // allow iv reuse. Essentially we are trading one constant multiplication
1194 // for one fewer iv.
1196 AllUsesAreAddresses = false;
1198 // There are no in-loop address uses.
1199 if (AllUsesAreAddresses && (!HasAddress && !AllUsesAreOutsideLoop))
1200 AllUsesAreAddresses = false;
1205 /// ShouldUseFullStrengthReductionMode - Test whether full strength-reduction
1206 /// is valid and profitable for the given set of users of a stride. In
1207 /// full strength-reduction mode, all addresses at the current stride are
1208 /// strength-reduced all the way down to pointer arithmetic.
1210 bool LoopStrengthReduce::ShouldUseFullStrengthReductionMode(
1211 const std::vector<BasedUser> &UsersToProcess,
1213 bool AllUsesAreAddresses,
1214 const SCEV *Stride) {
1215 if (!EnableFullLSRMode)
1218 // The heuristics below aim to avoid increasing register pressure, but
1219 // fully strength-reducing all the addresses increases the number of
1220 // add instructions, so don't do this when optimizing for size.
1221 // TODO: If the loop is large, the savings due to simpler addresses
1222 // may oughtweight the costs of the extra increment instructions.
1223 if (L->getHeader()->getParent()->hasFnAttr(Attribute::OptimizeForSize))
1226 // TODO: For now, don't do full strength reduction if there could
1227 // potentially be greater-stride multiples of the current stride
1228 // which could reuse the current stride IV.
1229 if (IU->StrideOrder.back() != Stride)
1232 // Iterate through the uses to find conditions that automatically rule out
1234 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ) {
1235 const SCEV *Base = UsersToProcess[i].Base;
1236 const SCEV *Imm = UsersToProcess[i].Imm;
1237 // If any users have a loop-variant component, they can't be fully
1238 // strength-reduced.
1239 if (Imm && !Imm->isLoopInvariant(L))
1241 // If there are to users with the same base and the difference between
1242 // the two Imm values can't be folded into the address, full
1243 // strength reduction would increase register pressure.
1245 const SCEV *CurImm = UsersToProcess[i].Imm;
1246 if ((CurImm || Imm) && CurImm != Imm) {
1247 if (!CurImm) CurImm = SE->getIntegerSCEV(0, Stride->getType());
1248 if (!Imm) Imm = SE->getIntegerSCEV(0, Stride->getType());
1249 const Instruction *Inst = UsersToProcess[i].Inst;
1250 const Type *AccessTy = getAccessType(Inst);
1251 const SCEV *Diff = SE->getMinusSCEV(UsersToProcess[i].Imm, Imm);
1252 if (!Diff->isZero() &&
1253 (!AllUsesAreAddresses ||
1254 !fitsInAddressMode(Diff, AccessTy, TLI, /*HasBaseReg=*/true)))
1257 } while (++i != e && Base == UsersToProcess[i].Base);
1260 // If there's exactly one user in this stride, fully strength-reducing it
1261 // won't increase register pressure. If it's starting from a non-zero base,
1262 // it'll be simpler this way.
1263 if (UsersToProcess.size() == 1 && !UsersToProcess[0].Base->isZero())
1266 // Otherwise, if there are any users in this stride that don't require
1267 // a register for their base, full strength-reduction will increase
1268 // register pressure.
1269 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1270 if (UsersToProcess[i].Base->isZero())
1273 // Otherwise, go for it.
1277 /// InsertAffinePhi Create and insert a PHI node for an induction variable
1278 /// with the specified start and step values in the specified loop.
1280 /// If NegateStride is true, the stride should be negated by using a
1281 /// subtract instead of an add.
1283 /// Return the created phi node.
1285 static PHINode *InsertAffinePhi(const SCEV *Start, const SCEV *Step,
1286 Instruction *IVIncInsertPt,
1288 SCEVExpander &Rewriter) {
1289 assert(Start->isLoopInvariant(L) && "New PHI start is not loop invariant!");
1290 assert(Step->isLoopInvariant(L) && "New PHI stride is not loop invariant!");
1292 BasicBlock *Header = L->getHeader();
1293 BasicBlock *Preheader = L->getLoopPreheader();
1294 BasicBlock *LatchBlock = L->getLoopLatch();
1295 const Type *Ty = Start->getType();
1296 Ty = Rewriter.SE.getEffectiveSCEVType(Ty);
1298 PHINode *PN = PHINode::Create(Ty, "lsr.iv", Header->begin());
1299 PN->addIncoming(Rewriter.expandCodeFor(Start, Ty, Preheader->getTerminator()),
1302 // If the stride is negative, insert a sub instead of an add for the
1304 bool isNegative = isNonConstantNegative(Step);
1305 const SCEV *IncAmount = Step;
1307 IncAmount = Rewriter.SE.getNegativeSCEV(Step);
1309 // Insert an add instruction right before the terminator corresponding
1310 // to the back-edge or just before the only use. The location is determined
1311 // by the caller and passed in as IVIncInsertPt.
1312 Value *StepV = Rewriter.expandCodeFor(IncAmount, Ty,
1313 Preheader->getTerminator());
1316 IncV = BinaryOperator::CreateSub(PN, StepV, "lsr.iv.next",
1319 IncV = BinaryOperator::CreateAdd(PN, StepV, "lsr.iv.next",
1322 if (!isa<ConstantInt>(StepV)) ++NumVariable;
1324 PN->addIncoming(IncV, LatchBlock);
1330 static void SortUsersToProcess(std::vector<BasedUser> &UsersToProcess) {
1331 // We want to emit code for users inside the loop first. To do this, we
1332 // rearrange BasedUser so that the entries at the end have
1333 // isUseOfPostIncrementedValue = false, because we pop off the end of the
1334 // vector (so we handle them first).
1335 std::partition(UsersToProcess.begin(), UsersToProcess.end(),
1336 PartitionByIsUseOfPostIncrementedValue);
1338 // Sort this by base, so that things with the same base are handled
1339 // together. By partitioning first and stable-sorting later, we are
1340 // guaranteed that within each base we will pop off users from within the
1341 // loop before users outside of the loop with a particular base.
1343 // We would like to use stable_sort here, but we can't. The problem is that
1344 // const SCEV *'s don't have a deterministic ordering w.r.t to each other, so
1345 // we don't have anything to do a '<' comparison on. Because we think the
1346 // number of uses is small, do a horrible bubble sort which just relies on
1348 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1349 // Get a base value.
1350 const SCEV *Base = UsersToProcess[i].Base;
1352 // Compact everything with this base to be consecutive with this one.
1353 for (unsigned j = i+1; j != e; ++j) {
1354 if (UsersToProcess[j].Base == Base) {
1355 std::swap(UsersToProcess[i+1], UsersToProcess[j]);
1362 /// PrepareToStrengthReduceFully - Prepare to fully strength-reduce
1363 /// UsersToProcess, meaning lowering addresses all the way down to direct
1364 /// pointer arithmetic.
1367 LoopStrengthReduce::PrepareToStrengthReduceFully(
1368 std::vector<BasedUser> &UsersToProcess,
1370 const SCEV *CommonExprs,
1372 SCEVExpander &PreheaderRewriter) {
1373 DOUT << " Fully reducing all users\n";
1375 // Rewrite the UsersToProcess records, creating a separate PHI for each
1376 // unique Base value.
1377 Instruction *IVIncInsertPt = L->getLoopLatch()->getTerminator();
1378 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ) {
1379 // TODO: The uses are grouped by base, but not sorted. We arbitrarily
1380 // pick the first Imm value here to start with, and adjust it for the
1382 const SCEV *Imm = UsersToProcess[i].Imm;
1383 const SCEV *Base = UsersToProcess[i].Base;
1384 const SCEV *Start = SE->getAddExpr(CommonExprs, Base, Imm);
1385 PHINode *Phi = InsertAffinePhi(Start, Stride, IVIncInsertPt, L,
1387 // Loop over all the users with the same base.
1389 UsersToProcess[i].Base = SE->getIntegerSCEV(0, Stride->getType());
1390 UsersToProcess[i].Imm = SE->getMinusSCEV(UsersToProcess[i].Imm, Imm);
1391 UsersToProcess[i].Phi = Phi;
1392 assert(UsersToProcess[i].Imm->isLoopInvariant(L) &&
1393 "ShouldUseFullStrengthReductionMode should reject this!");
1394 } while (++i != e && Base == UsersToProcess[i].Base);
1398 /// FindIVIncInsertPt - Return the location to insert the increment instruction.
1399 /// If the only use if a use of postinc value, (must be the loop termination
1400 /// condition), then insert it just before the use.
1401 static Instruction *FindIVIncInsertPt(std::vector<BasedUser> &UsersToProcess,
1403 if (UsersToProcess.size() == 1 &&
1404 UsersToProcess[0].isUseOfPostIncrementedValue &&
1405 L->contains(UsersToProcess[0].Inst->getParent()))
1406 return UsersToProcess[0].Inst;
1407 return L->getLoopLatch()->getTerminator();
1410 /// PrepareToStrengthReduceWithNewPhi - Insert a new induction variable for the
1411 /// given users to share.
1414 LoopStrengthReduce::PrepareToStrengthReduceWithNewPhi(
1415 std::vector<BasedUser> &UsersToProcess,
1417 const SCEV *CommonExprs,
1419 Instruction *IVIncInsertPt,
1421 SCEVExpander &PreheaderRewriter) {
1422 DOUT << " Inserting new PHI:\n";
1424 PHINode *Phi = InsertAffinePhi(SE->getUnknown(CommonBaseV),
1425 Stride, IVIncInsertPt, L,
1428 // Remember this in case a later stride is multiple of this.
1429 IVsByStride[Stride].addIV(Stride, CommonExprs, Phi);
1431 // All the users will share this new IV.
1432 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1433 UsersToProcess[i].Phi = Phi;
1436 DEBUG(WriteAsOperand(*DOUT, Phi, /*PrintType=*/false));
1440 /// PrepareToStrengthReduceFromSmallerStride - Prepare for the given users to
1441 /// reuse an induction variable with a stride that is a factor of the current
1442 /// induction variable.
1445 LoopStrengthReduce::PrepareToStrengthReduceFromSmallerStride(
1446 std::vector<BasedUser> &UsersToProcess,
1448 const IVExpr &ReuseIV,
1449 Instruction *PreInsertPt) {
1450 DOUT << " Rewriting in terms of existing IV of STRIDE " << *ReuseIV.Stride
1451 << " and BASE " << *ReuseIV.Base << "\n";
1453 // All the users will share the reused IV.
1454 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1455 UsersToProcess[i].Phi = ReuseIV.PHI;
1457 Constant *C = dyn_cast<Constant>(CommonBaseV);
1459 (!C->isNullValue() &&
1460 !fitsInAddressMode(SE->getUnknown(CommonBaseV), CommonBaseV->getType(),
1462 // We want the common base emitted into the preheader! This is just
1463 // using cast as a copy so BitCast (no-op cast) is appropriate
1464 CommonBaseV = new BitCastInst(CommonBaseV, CommonBaseV->getType(),
1465 "commonbase", PreInsertPt);
1468 static bool IsImmFoldedIntoAddrMode(GlobalValue *GV, int64_t Offset,
1469 const Type *AccessTy,
1470 std::vector<BasedUser> &UsersToProcess,
1471 const TargetLowering *TLI) {
1472 SmallVector<Instruction*, 16> AddrModeInsts;
1473 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1474 if (UsersToProcess[i].isUseOfPostIncrementedValue)
1476 ExtAddrMode AddrMode =
1477 AddressingModeMatcher::Match(UsersToProcess[i].OperandValToReplace,
1478 AccessTy, UsersToProcess[i].Inst,
1479 AddrModeInsts, *TLI);
1480 if (GV && GV != AddrMode.BaseGV)
1482 if (Offset && !AddrMode.BaseOffs)
1483 // FIXME: How to accurate check it's immediate offset is folded.
1485 AddrModeInsts.clear();
1490 /// StrengthReduceStridedIVUsers - Strength reduce all of the users of a single
1491 /// stride of IV. All of the users may have different starting values, and this
1492 /// may not be the only stride.
1493 void LoopStrengthReduce::StrengthReduceStridedIVUsers(const SCEV *const &Stride,
1494 IVUsersOfOneStride &Uses,
1496 // If all the users are moved to another stride, then there is nothing to do.
1497 if (Uses.Users.empty())
1500 // Keep track if every use in UsersToProcess is an address. If they all are,
1501 // we may be able to rewrite the entire collection of them in terms of a
1502 // smaller-stride IV.
1503 bool AllUsesAreAddresses = true;
1505 // Keep track if every use of a single stride is outside the loop. If so,
1506 // we want to be more aggressive about reusing a smaller-stride IV; a
1507 // multiply outside the loop is better than another IV inside. Well, usually.
1508 bool AllUsesAreOutsideLoop = true;
1510 // Transform our list of users and offsets to a bit more complex table. In
1511 // this new vector, each 'BasedUser' contains 'Base' the base of the
1512 // strided accessas well as the old information from Uses. We progressively
1513 // move information from the Base field to the Imm field, until we eventually
1514 // have the full access expression to rewrite the use.
1515 std::vector<BasedUser> UsersToProcess;
1516 const SCEV *CommonExprs = CollectIVUsers(Stride, Uses, L, AllUsesAreAddresses,
1517 AllUsesAreOutsideLoop,
1520 // Sort the UsersToProcess array so that users with common bases are
1521 // next to each other.
1522 SortUsersToProcess(UsersToProcess);
1524 // If we managed to find some expressions in common, we'll need to carry
1525 // their value in a register and add it in for each use. This will take up
1526 // a register operand, which potentially restricts what stride values are
1528 bool HaveCommonExprs = !CommonExprs->isZero();
1529 const Type *ReplacedTy = CommonExprs->getType();
1531 // If all uses are addresses, consider sinking the immediate part of the
1532 // common expression back into uses if they can fit in the immediate fields.
1533 if (TLI && HaveCommonExprs && AllUsesAreAddresses) {
1534 const SCEV *NewCommon = CommonExprs;
1535 const SCEV *Imm = SE->getIntegerSCEV(0, ReplacedTy);
1536 MoveImmediateValues(TLI, Type::VoidTy, NewCommon, Imm, true, L, SE);
1537 if (!Imm->isZero()) {
1540 // If the immediate part of the common expression is a GV, check if it's
1541 // possible to fold it into the target addressing mode.
1542 GlobalValue *GV = 0;
1543 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(Imm))
1544 GV = dyn_cast<GlobalValue>(SU->getValue());
1546 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Imm))
1547 Offset = SC->getValue()->getSExtValue();
1549 // Pass VoidTy as the AccessTy to be conservative, because
1550 // there could be multiple access types among all the uses.
1551 DoSink = IsImmFoldedIntoAddrMode(GV, Offset, Type::VoidTy,
1552 UsersToProcess, TLI);
1555 DOUT << " Sinking " << *Imm << " back down into uses\n";
1556 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1557 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm, Imm);
1558 CommonExprs = NewCommon;
1559 HaveCommonExprs = !CommonExprs->isZero();
1565 // Now that we know what we need to do, insert the PHI node itself.
1567 DOUT << "LSR: Examining IVs of TYPE " << *ReplacedTy << " of STRIDE "
1569 << " Common base: " << *CommonExprs << "\n";
1571 SCEVExpander Rewriter(*SE);
1572 SCEVExpander PreheaderRewriter(*SE);
1574 BasicBlock *Preheader = L->getLoopPreheader();
1575 Instruction *PreInsertPt = Preheader->getTerminator();
1576 BasicBlock *LatchBlock = L->getLoopLatch();
1577 Instruction *IVIncInsertPt = LatchBlock->getTerminator();
1579 Value *CommonBaseV = Context->getNullValue(ReplacedTy);
1581 const SCEV *RewriteFactor = SE->getIntegerSCEV(0, ReplacedTy);
1582 IVExpr ReuseIV(SE->getIntegerSCEV(0, Type::Int32Ty),
1583 SE->getIntegerSCEV(0, Type::Int32Ty),
1586 /// Choose a strength-reduction strategy and prepare for it by creating
1587 /// the necessary PHIs and adjusting the bookkeeping.
1588 if (ShouldUseFullStrengthReductionMode(UsersToProcess, L,
1589 AllUsesAreAddresses, Stride)) {
1590 PrepareToStrengthReduceFully(UsersToProcess, Stride, CommonExprs, L,
1593 // Emit the initial base value into the loop preheader.
1594 CommonBaseV = PreheaderRewriter.expandCodeFor(CommonExprs, ReplacedTy,
1597 // If all uses are addresses, check if it is possible to reuse an IV. The
1598 // new IV must have a stride that is a multiple of the old stride; the
1599 // multiple must be a number that can be encoded in the scale field of the
1600 // target addressing mode; and we must have a valid instruction after this
1601 // substitution, including the immediate field, if any.
1602 RewriteFactor = CheckForIVReuse(HaveCommonExprs, AllUsesAreAddresses,
1603 AllUsesAreOutsideLoop,
1604 Stride, ReuseIV, ReplacedTy,
1606 if (!RewriteFactor->isZero())
1607 PrepareToStrengthReduceFromSmallerStride(UsersToProcess, CommonBaseV,
1608 ReuseIV, PreInsertPt);
1610 IVIncInsertPt = FindIVIncInsertPt(UsersToProcess, L);
1611 PrepareToStrengthReduceWithNewPhi(UsersToProcess, Stride, CommonExprs,
1612 CommonBaseV, IVIncInsertPt,
1613 L, PreheaderRewriter);
1617 // Process all the users now, replacing their strided uses with
1618 // strength-reduced forms. This outer loop handles all bases, the inner
1619 // loop handles all users of a particular base.
1620 while (!UsersToProcess.empty()) {
1621 const SCEV *Base = UsersToProcess.back().Base;
1622 Instruction *Inst = UsersToProcess.back().Inst;
1624 // Emit the code for Base into the preheader.
1626 if (!Base->isZero()) {
1627 BaseV = PreheaderRewriter.expandCodeFor(Base, 0, PreInsertPt);
1629 DOUT << " INSERTING code for BASE = " << *Base << ":";
1630 if (BaseV->hasName())
1631 DOUT << " Result value name = %" << BaseV->getNameStr();
1634 // If BaseV is a non-zero constant, make sure that it gets inserted into
1635 // the preheader, instead of being forward substituted into the uses. We
1636 // do this by forcing a BitCast (noop cast) to be inserted into the
1637 // preheader in this case.
1638 if (!fitsInAddressMode(Base, getAccessType(Inst), TLI, false) &&
1639 !isa<Instruction>(BaseV)) {
1640 // We want this constant emitted into the preheader! This is just
1641 // using cast as a copy so BitCast (no-op cast) is appropriate
1642 BaseV = new BitCastInst(BaseV, BaseV->getType(), "preheaderinsert",
1647 // Emit the code to add the immediate offset to the Phi value, just before
1648 // the instructions that we identified as using this stride and base.
1650 // FIXME: Use emitted users to emit other users.
1651 BasedUser &User = UsersToProcess.back();
1653 DOUT << " Examining ";
1654 if (User.isUseOfPostIncrementedValue)
1659 DEBUG(WriteAsOperand(*DOUT, UsersToProcess.back().OperandValToReplace,
1660 /*PrintType=*/false));
1661 DOUT << " in Inst: " << *(User.Inst);
1663 // If this instruction wants to use the post-incremented value, move it
1664 // after the post-inc and use its value instead of the PHI.
1665 Value *RewriteOp = User.Phi;
1666 if (User.isUseOfPostIncrementedValue) {
1667 RewriteOp = User.Phi->getIncomingValueForBlock(LatchBlock);
1668 // If this user is in the loop, make sure it is the last thing in the
1669 // loop to ensure it is dominated by the increment. In case it's the
1670 // only use of the iv, the increment instruction is already before the
1672 if (L->contains(User.Inst->getParent()) && User.Inst != IVIncInsertPt)
1673 User.Inst->moveBefore(IVIncInsertPt);
1676 const SCEV *RewriteExpr = SE->getUnknown(RewriteOp);
1678 if (SE->getEffectiveSCEVType(RewriteOp->getType()) !=
1679 SE->getEffectiveSCEVType(ReplacedTy)) {
1680 assert(SE->getTypeSizeInBits(RewriteOp->getType()) >
1681 SE->getTypeSizeInBits(ReplacedTy) &&
1682 "Unexpected widening cast!");
1683 RewriteExpr = SE->getTruncateExpr(RewriteExpr, ReplacedTy);
1686 // If we had to insert new instructions for RewriteOp, we have to
1687 // consider that they may not have been able to end up immediately
1688 // next to RewriteOp, because non-PHI instructions may never precede
1689 // PHI instructions in a block. In this case, remember where the last
1690 // instruction was inserted so that if we're replacing a different
1691 // PHI node, we can use the later point to expand the final
1693 Instruction *NewBasePt = dyn_cast<Instruction>(RewriteOp);
1694 if (RewriteOp == User.Phi) NewBasePt = 0;
1696 // Clear the SCEVExpander's expression map so that we are guaranteed
1697 // to have the code emitted where we expect it.
1700 // If we are reusing the iv, then it must be multiplied by a constant
1701 // factor to take advantage of the addressing mode scale component.
1702 if (!RewriteFactor->isZero()) {
1703 // If we're reusing an IV with a nonzero base (currently this happens
1704 // only when all reuses are outside the loop) subtract that base here.
1705 // The base has been used to initialize the PHI node but we don't want
1707 if (!ReuseIV.Base->isZero()) {
1708 const SCEV *typedBase = ReuseIV.Base;
1709 if (SE->getEffectiveSCEVType(RewriteExpr->getType()) !=
1710 SE->getEffectiveSCEVType(ReuseIV.Base->getType())) {
1711 // It's possible the original IV is a larger type than the new IV,
1712 // in which case we have to truncate the Base. We checked in
1713 // RequiresTypeConversion that this is valid.
1714 assert(SE->getTypeSizeInBits(RewriteExpr->getType()) <
1715 SE->getTypeSizeInBits(ReuseIV.Base->getType()) &&
1716 "Unexpected lengthening conversion!");
1717 typedBase = SE->getTruncateExpr(ReuseIV.Base,
1718 RewriteExpr->getType());
1720 RewriteExpr = SE->getMinusSCEV(RewriteExpr, typedBase);
1723 // Multiply old variable, with base removed, by new scale factor.
1724 RewriteExpr = SE->getMulExpr(RewriteFactor,
1727 // The common base is emitted in the loop preheader. But since we
1728 // are reusing an IV, it has not been used to initialize the PHI node.
1729 // Add it to the expression used to rewrite the uses.
1730 // When this use is outside the loop, we earlier subtracted the
1731 // common base, and are adding it back here. Use the same expression
1732 // as before, rather than CommonBaseV, so DAGCombiner will zap it.
1733 if (!CommonExprs->isZero()) {
1734 if (L->contains(User.Inst->getParent()))
1735 RewriteExpr = SE->getAddExpr(RewriteExpr,
1736 SE->getUnknown(CommonBaseV));
1738 RewriteExpr = SE->getAddExpr(RewriteExpr, CommonExprs);
1742 // Now that we know what we need to do, insert code before User for the
1743 // immediate and any loop-variant expressions.
1745 // Add BaseV to the PHI value if needed.
1746 RewriteExpr = SE->getAddExpr(RewriteExpr, SE->getUnknown(BaseV));
1748 User.RewriteInstructionToUseNewBase(RewriteExpr, NewBasePt,
1749 Rewriter, L, this, *LI,
1752 // Mark old value we replaced as possibly dead, so that it is eliminated
1753 // if we just replaced the last use of that value.
1754 DeadInsts.push_back(User.OperandValToReplace);
1756 UsersToProcess.pop_back();
1759 // If there are any more users to process with the same base, process them
1760 // now. We sorted by base above, so we just have to check the last elt.
1761 } while (!UsersToProcess.empty() && UsersToProcess.back().Base == Base);
1762 // TODO: Next, find out which base index is the most common, pull it out.
1765 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
1766 // different starting values, into different PHIs.
1769 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1770 /// set the IV user and stride information and return true, otherwise return
1772 bool LoopStrengthReduce::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse,
1773 const SCEV *const * &CondStride) {
1774 for (unsigned Stride = 0, e = IU->StrideOrder.size();
1775 Stride != e && !CondUse; ++Stride) {
1776 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
1777 IU->IVUsesByStride.find(IU->StrideOrder[Stride]);
1778 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
1780 for (ilist<IVStrideUse>::iterator UI = SI->second->Users.begin(),
1781 E = SI->second->Users.end(); UI != E; ++UI)
1782 if (UI->getUser() == Cond) {
1783 // NOTE: we could handle setcc instructions with multiple uses here, but
1784 // InstCombine does it as well for simple uses, it's not clear that it
1785 // occurs enough in real life to handle.
1787 CondStride = &SI->first;
1795 // Constant strides come first which in turns are sorted by their absolute
1796 // values. If absolute values are the same, then positive strides comes first.
1798 // 4, -1, X, 1, 2 ==> 1, -1, 2, 4, X
1799 struct StrideCompare {
1800 const ScalarEvolution *SE;
1801 explicit StrideCompare(const ScalarEvolution *se) : SE(se) {}
1803 bool operator()(const SCEV *const &LHS, const SCEV *const &RHS) {
1804 const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS);
1805 const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS);
1807 int64_t LV = LHSC->getValue()->getSExtValue();
1808 int64_t RV = RHSC->getValue()->getSExtValue();
1809 uint64_t ALV = (LV < 0) ? -LV : LV;
1810 uint64_t ARV = (RV < 0) ? -RV : RV;
1818 // If it's the same value but different type, sort by bit width so
1819 // that we emit larger induction variables before smaller
1820 // ones, letting the smaller be re-written in terms of larger ones.
1821 return SE->getTypeSizeInBits(RHS->getType()) <
1822 SE->getTypeSizeInBits(LHS->getType());
1824 return LHSC && !RHSC;
1829 /// ChangeCompareStride - If a loop termination compare instruction is the
1830 /// only use of its stride, and the compaison is against a constant value,
1831 /// try eliminate the stride by moving the compare instruction to another
1832 /// stride and change its constant operand accordingly. e.g.
1838 /// if (v2 < 10) goto loop
1843 /// if (v1 < 30) goto loop
1844 ICmpInst *LoopStrengthReduce::ChangeCompareStride(Loop *L, ICmpInst *Cond,
1845 IVStrideUse* &CondUse,
1846 const SCEV *const* &CondStride) {
1847 // If there's only one stride in the loop, there's nothing to do here.
1848 if (IU->StrideOrder.size() < 2)
1850 // If there are other users of the condition's stride, don't bother
1851 // trying to change the condition because the stride will still
1853 std::map<const SCEV *, IVUsersOfOneStride *>::iterator I =
1854 IU->IVUsesByStride.find(*CondStride);
1855 if (I == IU->IVUsesByStride.end() ||
1856 I->second->Users.size() != 1)
1858 // Only handle constant strides for now.
1859 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*CondStride);
1860 if (!SC) return Cond;
1862 ICmpInst::Predicate Predicate = Cond->getPredicate();
1863 int64_t CmpSSInt = SC->getValue()->getSExtValue();
1864 unsigned BitWidth = SE->getTypeSizeInBits((*CondStride)->getType());
1865 uint64_t SignBit = 1ULL << (BitWidth-1);
1866 const Type *CmpTy = Cond->getOperand(0)->getType();
1867 const Type *NewCmpTy = NULL;
1868 unsigned TyBits = SE->getTypeSizeInBits(CmpTy);
1869 unsigned NewTyBits = 0;
1870 const SCEV **NewStride = NULL;
1871 Value *NewCmpLHS = NULL;
1872 Value *NewCmpRHS = NULL;
1874 const SCEV *NewOffset = SE->getIntegerSCEV(0, CmpTy);
1876 if (ConstantInt *C = dyn_cast<ConstantInt>(Cond->getOperand(1))) {
1877 int64_t CmpVal = C->getValue().getSExtValue();
1879 // Check stride constant and the comparision constant signs to detect
1881 if ((CmpVal & SignBit) != (CmpSSInt & SignBit))
1884 // Look for a suitable stride / iv as replacement.
1885 for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
1886 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
1887 IU->IVUsesByStride.find(IU->StrideOrder[i]);
1888 if (!isa<SCEVConstant>(SI->first))
1890 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1891 if (SSInt == CmpSSInt ||
1892 abs64(SSInt) < abs64(CmpSSInt) ||
1893 (SSInt % CmpSSInt) != 0)
1896 Scale = SSInt / CmpSSInt;
1897 int64_t NewCmpVal = CmpVal * Scale;
1898 APInt Mul = APInt(BitWidth*2, CmpVal, true);
1899 Mul = Mul * APInt(BitWidth*2, Scale, true);
1900 // Check for overflow.
1901 if (!Mul.isSignedIntN(BitWidth))
1903 // Check for overflow in the stride's type too.
1904 if (!Mul.isSignedIntN(SE->getTypeSizeInBits(SI->first->getType())))
1907 // Watch out for overflow.
1908 if (ICmpInst::isSignedPredicate(Predicate) &&
1909 (CmpVal & SignBit) != (NewCmpVal & SignBit))
1912 if (NewCmpVal == CmpVal)
1914 // Pick the best iv to use trying to avoid a cast.
1916 for (ilist<IVStrideUse>::iterator UI = SI->second->Users.begin(),
1917 E = SI->second->Users.end(); UI != E; ++UI) {
1918 Value *Op = UI->getOperandValToReplace();
1920 // If the IVStrideUse implies a cast, check for an actual cast which
1921 // can be used to find the original IV expression.
1922 if (SE->getEffectiveSCEVType(Op->getType()) !=
1923 SE->getEffectiveSCEVType(SI->first->getType())) {
1924 CastInst *CI = dyn_cast<CastInst>(Op);
1925 // If it's not a simple cast, it's complicated.
1928 // If it's a cast from a type other than the stride type,
1929 // it's complicated.
1930 if (CI->getOperand(0)->getType() != SI->first->getType())
1932 // Ok, we found the IV expression in the stride's type.
1933 Op = CI->getOperand(0);
1937 if (NewCmpLHS->getType() == CmpTy)
1943 NewCmpTy = NewCmpLHS->getType();
1944 NewTyBits = SE->getTypeSizeInBits(NewCmpTy);
1945 const Type *NewCmpIntTy = Context->getIntegerType(NewTyBits);
1946 if (RequiresTypeConversion(NewCmpTy, CmpTy)) {
1947 // Check if it is possible to rewrite it using
1948 // an iv / stride of a smaller integer type.
1949 unsigned Bits = NewTyBits;
1950 if (ICmpInst::isSignedPredicate(Predicate))
1952 uint64_t Mask = (1ULL << Bits) - 1;
1953 if (((uint64_t)NewCmpVal & Mask) != (uint64_t)NewCmpVal)
1957 // Don't rewrite if use offset is non-constant and the new type is
1958 // of a different type.
1959 // FIXME: too conservative?
1960 if (NewTyBits != TyBits && !isa<SCEVConstant>(CondUse->getOffset()))
1963 bool AllUsesAreAddresses = true;
1964 bool AllUsesAreOutsideLoop = true;
1965 std::vector<BasedUser> UsersToProcess;
1966 const SCEV *CommonExprs = CollectIVUsers(SI->first, *SI->second, L,
1967 AllUsesAreAddresses,
1968 AllUsesAreOutsideLoop,
1970 // Avoid rewriting the compare instruction with an iv of new stride
1971 // if it's likely the new stride uses will be rewritten using the
1972 // stride of the compare instruction.
1973 if (AllUsesAreAddresses &&
1974 ValidScale(!CommonExprs->isZero(), Scale, UsersToProcess))
1977 // Avoid rewriting the compare instruction with an iv which has
1978 // implicit extension or truncation built into it.
1979 // TODO: This is over-conservative.
1980 if (SE->getTypeSizeInBits(CondUse->getOffset()->getType()) != TyBits)
1983 // If scale is negative, use swapped predicate unless it's testing
1985 if (Scale < 0 && !Cond->isEquality())
1986 Predicate = ICmpInst::getSwappedPredicate(Predicate);
1988 NewStride = &IU->StrideOrder[i];
1989 if (!isa<PointerType>(NewCmpTy))
1990 NewCmpRHS = Context->getConstantInt(NewCmpTy, NewCmpVal);
1992 Constant *CI = Context->getConstantInt(NewCmpIntTy, NewCmpVal);
1993 NewCmpRHS = Context->getConstantExprIntToPtr(CI, NewCmpTy);
1995 NewOffset = TyBits == NewTyBits
1996 ? SE->getMulExpr(CondUse->getOffset(),
1997 SE->getConstant(CmpTy, Scale))
1998 : SE->getConstant(NewCmpIntTy,
1999 cast<SCEVConstant>(CondUse->getOffset())->getValue()
2000 ->getSExtValue()*Scale);
2005 // Forgo this transformation if it the increment happens to be
2006 // unfortunately positioned after the condition, and the condition
2007 // has multiple uses which prevent it from being moved immediately
2008 // before the branch. See
2009 // test/Transforms/LoopStrengthReduce/change-compare-stride-trickiness-*.ll
2010 // for an example of this situation.
2011 if (!Cond->hasOneUse()) {
2012 for (BasicBlock::iterator I = Cond, E = Cond->getParent()->end();
2019 // Create a new compare instruction using new stride / iv.
2020 ICmpInst *OldCond = Cond;
2021 // Insert new compare instruction.
2022 Cond = new ICmpInst(OldCond, Predicate, NewCmpLHS, NewCmpRHS,
2023 L->getHeader()->getName() + ".termcond");
2025 // Remove the old compare instruction. The old indvar is probably dead too.
2026 DeadInsts.push_back(CondUse->getOperandValToReplace());
2027 OldCond->replaceAllUsesWith(Cond);
2028 OldCond->eraseFromParent();
2030 IU->IVUsesByStride[*NewStride]->addUser(NewOffset, Cond, NewCmpLHS);
2031 CondUse = &IU->IVUsesByStride[*NewStride]->Users.back();
2032 CondStride = NewStride;
2040 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
2041 /// a max computation.
2043 /// This is a narrow solution to a specific, but acute, problem. For loops
2049 /// } while (++i < n);
2051 /// the trip count isn't just 'n', because 'n' might not be positive. And
2052 /// unfortunately this can come up even for loops where the user didn't use
2053 /// a C do-while loop. For example, seemingly well-behaved top-test loops
2054 /// will commonly be lowered like this:
2060 /// } while (++i < n);
2063 /// and then it's possible for subsequent optimization to obscure the if
2064 /// test in such a way that indvars can't find it.
2066 /// When indvars can't find the if test in loops like this, it creates a
2067 /// max expression, which allows it to give the loop a canonical
2068 /// induction variable:
2071 /// max = n < 1 ? 1 : n;
2074 /// } while (++i != max);
2076 /// Canonical induction variables are necessary because the loop passes
2077 /// are designed around them. The most obvious example of this is the
2078 /// LoopInfo analysis, which doesn't remember trip count values. It
2079 /// expects to be able to rediscover the trip count each time it is
2080 /// needed, and it does this using a simple analyis that only succeeds if
2081 /// the loop has a canonical induction variable.
2083 /// However, when it comes time to generate code, the maximum operation
2084 /// can be quite costly, especially if it's inside of an outer loop.
2086 /// This function solves this problem by detecting this type of loop and
2087 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
2088 /// the instructions for the maximum computation.
2090 ICmpInst *LoopStrengthReduce::OptimizeMax(Loop *L, ICmpInst *Cond,
2091 IVStrideUse* &CondUse) {
2092 // Check that the loop matches the pattern we're looking for.
2093 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
2094 Cond->getPredicate() != CmpInst::ICMP_NE)
2097 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
2098 if (!Sel || !Sel->hasOneUse()) return Cond;
2100 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
2101 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2103 const SCEV *One = SE->getIntegerSCEV(1, BackedgeTakenCount->getType());
2105 // Add one to the backedge-taken count to get the trip count.
2106 const SCEV *IterationCount = SE->getAddExpr(BackedgeTakenCount, One);
2108 // Check for a max calculation that matches the pattern.
2109 if (!isa<SCEVSMaxExpr>(IterationCount) && !isa<SCEVUMaxExpr>(IterationCount))
2111 const SCEVNAryExpr *Max = cast<SCEVNAryExpr>(IterationCount);
2112 if (Max != SE->getSCEV(Sel)) return Cond;
2114 // To handle a max with more than two operands, this optimization would
2115 // require additional checking and setup.
2116 if (Max->getNumOperands() != 2)
2119 const SCEV *MaxLHS = Max->getOperand(0);
2120 const SCEV *MaxRHS = Max->getOperand(1);
2121 if (!MaxLHS || MaxLHS != One) return Cond;
2123 // Check the relevant induction variable for conformance to
2125 const SCEV *IV = SE->getSCEV(Cond->getOperand(0));
2126 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2127 if (!AR || !AR->isAffine() ||
2128 AR->getStart() != One ||
2129 AR->getStepRecurrence(*SE) != One)
2132 assert(AR->getLoop() == L &&
2133 "Loop condition operand is an addrec in a different loop!");
2135 // Check the right operand of the select, and remember it, as it will
2136 // be used in the new comparison instruction.
2138 if (SE->getSCEV(Sel->getOperand(1)) == MaxRHS)
2139 NewRHS = Sel->getOperand(1);
2140 else if (SE->getSCEV(Sel->getOperand(2)) == MaxRHS)
2141 NewRHS = Sel->getOperand(2);
2142 if (!NewRHS) return Cond;
2144 // Determine the new comparison opcode. It may be signed or unsigned,
2145 // and the original comparison may be either equality or inequality.
2146 CmpInst::Predicate Pred =
2147 isa<SCEVSMaxExpr>(Max) ? CmpInst::ICMP_SLT : CmpInst::ICMP_ULT;
2148 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
2149 Pred = CmpInst::getInversePredicate(Pred);
2151 // Ok, everything looks ok to change the condition into an SLT or SGE and
2152 // delete the max calculation.
2154 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
2156 // Delete the max calculation instructions.
2157 Cond->replaceAllUsesWith(NewCond);
2158 CondUse->setUser(NewCond);
2159 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2160 Cond->eraseFromParent();
2161 Sel->eraseFromParent();
2162 if (Cmp->use_empty())
2163 Cmp->eraseFromParent();
2167 /// OptimizeShadowIV - If IV is used in a int-to-float cast
2168 /// inside the loop then try to eliminate the cast opeation.
2169 void LoopStrengthReduce::OptimizeShadowIV(Loop *L) {
2171 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
2172 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2175 for (unsigned Stride = 0, e = IU->StrideOrder.size(); Stride != e;
2177 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
2178 IU->IVUsesByStride.find(IU->StrideOrder[Stride]);
2179 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
2180 if (!isa<SCEVConstant>(SI->first))
2183 for (ilist<IVStrideUse>::iterator UI = SI->second->Users.begin(),
2184 E = SI->second->Users.end(); UI != E; /* empty */) {
2185 ilist<IVStrideUse>::iterator CandidateUI = UI;
2187 Instruction *ShadowUse = CandidateUI->getUser();
2188 const Type *DestTy = NULL;
2190 /* If shadow use is a int->float cast then insert a second IV
2191 to eliminate this cast.
2193 for (unsigned i = 0; i < n; ++i)
2199 for (unsigned i = 0; i < n; ++i, ++d)
2202 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser()))
2203 DestTy = UCast->getDestTy();
2204 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser()))
2205 DestTy = SCast->getDestTy();
2206 if (!DestTy) continue;
2209 // If target does not support DestTy natively then do not apply
2210 // this transformation.
2211 MVT DVT = TLI->getValueType(DestTy);
2212 if (!TLI->isTypeLegal(DVT)) continue;
2215 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
2217 if (PH->getNumIncomingValues() != 2) continue;
2219 const Type *SrcTy = PH->getType();
2220 int Mantissa = DestTy->getFPMantissaWidth();
2221 if (Mantissa == -1) continue;
2222 if ((int)SE->getTypeSizeInBits(SrcTy) > Mantissa)
2225 unsigned Entry, Latch;
2226 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
2234 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
2235 if (!Init) continue;
2236 Constant *NewInit = Context->getConstantFP(DestTy, Init->getZExtValue());
2238 BinaryOperator *Incr =
2239 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
2240 if (!Incr) continue;
2241 if (Incr->getOpcode() != Instruction::Add
2242 && Incr->getOpcode() != Instruction::Sub)
2245 /* Initialize new IV, double d = 0.0 in above example. */
2246 ConstantInt *C = NULL;
2247 if (Incr->getOperand(0) == PH)
2248 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
2249 else if (Incr->getOperand(1) == PH)
2250 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
2256 /* Add new PHINode. */
2257 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);
2259 /* create new increment. '++d' in above example. */
2260 Constant *CFP = Context->getConstantFP(DestTy, C->getZExtValue());
2261 BinaryOperator *NewIncr =
2262 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
2263 Instruction::FAdd : Instruction::FSub,
2264 NewPH, CFP, "IV.S.next.", Incr);
2266 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
2267 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
2269 /* Remove cast operation */
2270 ShadowUse->replaceAllUsesWith(NewPH);
2271 ShadowUse->eraseFromParent();
2278 /// OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
2279 /// uses in the loop, look to see if we can eliminate some, in favor of using
2280 /// common indvars for the different uses.
2281 void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
2282 // TODO: implement optzns here.
2284 OptimizeShadowIV(L);
2287 /// OptimizeLoopTermCond - Change loop terminating condition to use the
2288 /// postinc iv when possible.
2289 void LoopStrengthReduce::OptimizeLoopTermCond(Loop *L) {
2290 // Finally, get the terminating condition for the loop if possible. If we
2291 // can, we want to change it to use a post-incremented version of its
2292 // induction variable, to allow coalescing the live ranges for the IV into
2293 // one register value.
2294 BasicBlock *LatchBlock = L->getLoopLatch();
2295 BasicBlock *ExitingBlock = L->getExitingBlock();
2297 // Multiple exits, just look at the exit in the latch block if there is one.
2298 ExitingBlock = LatchBlock;
2299 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2302 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
2305 // Search IVUsesByStride to find Cond's IVUse if there is one.
2306 IVStrideUse *CondUse = 0;
2307 const SCEV *const *CondStride = 0;
2308 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2309 if (!FindIVUserForCond(Cond, CondUse, CondStride))
2310 return; // setcc doesn't use the IV.
2312 if (ExitingBlock != LatchBlock) {
2313 if (!Cond->hasOneUse())
2314 // See below, we don't want the condition to be cloned.
2317 // If exiting block is the latch block, we know it's safe and profitable to
2318 // transform the icmp to use post-inc iv. Otherwise do so only if it would
2319 // not reuse another iv and its iv would be reused by other uses. We are
2320 // optimizing for the case where the icmp is the only use of the iv.
2321 IVUsersOfOneStride &StrideUses = *IU->IVUsesByStride[*CondStride];
2322 for (ilist<IVStrideUse>::iterator I = StrideUses.Users.begin(),
2323 E = StrideUses.Users.end(); I != E; ++I) {
2324 if (I->getUser() == Cond)
2326 if (!I->isUseOfPostIncrementedValue())
2330 // FIXME: This is expensive, and worse still ChangeCompareStride does a
2331 // similar check. Can we perform all the icmp related transformations after
2332 // StrengthReduceStridedIVUsers?
2333 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(*CondStride)) {
2334 int64_t SInt = SC->getValue()->getSExtValue();
2335 for (unsigned NewStride = 0, ee = IU->StrideOrder.size(); NewStride != ee;
2337 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
2338 IU->IVUsesByStride.find(IU->StrideOrder[NewStride]);
2339 if (!isa<SCEVConstant>(SI->first) || SI->first == *CondStride)
2342 cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
2344 return; // This can definitely be reused.
2345 if (unsigned(abs64(SSInt)) < SInt || (SSInt % SInt) != 0)
2347 int64_t Scale = SSInt / SInt;
2348 bool AllUsesAreAddresses = true;
2349 bool AllUsesAreOutsideLoop = true;
2350 std::vector<BasedUser> UsersToProcess;
2351 const SCEV *CommonExprs = CollectIVUsers(SI->first, *SI->second, L,
2352 AllUsesAreAddresses,
2353 AllUsesAreOutsideLoop,
2355 // Avoid rewriting the compare instruction with an iv of new stride
2356 // if it's likely the new stride uses will be rewritten using the
2357 // stride of the compare instruction.
2358 if (AllUsesAreAddresses &&
2359 ValidScale(!CommonExprs->isZero(), Scale, UsersToProcess))
2364 StrideNoReuse.insert(*CondStride);
2367 // If the trip count is computed in terms of a max (due to ScalarEvolution
2368 // being unable to find a sufficient guard, for example), change the loop
2369 // comparison to use SLT or ULT instead of NE.
2370 Cond = OptimizeMax(L, Cond, CondUse);
2372 // If possible, change stride and operands of the compare instruction to
2373 // eliminate one stride.
2374 if (ExitingBlock == LatchBlock)
2375 Cond = ChangeCompareStride(L, Cond, CondUse, CondStride);
2377 // It's possible for the setcc instruction to be anywhere in the loop, and
2378 // possible for it to have multiple users. If it is not immediately before
2379 // the latch block branch, move it.
2380 if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
2381 if (Cond->hasOneUse()) { // Condition has a single use, just move it.
2382 Cond->moveBefore(TermBr);
2384 // Otherwise, clone the terminating condition and insert into the loopend.
2385 Cond = cast<ICmpInst>(Cond->clone(*Context));
2386 Cond->setName(L->getHeader()->getName() + ".termcond");
2387 LatchBlock->getInstList().insert(TermBr, Cond);
2389 // Clone the IVUse, as the old use still exists!
2390 IU->IVUsesByStride[*CondStride]->addUser(CondUse->getOffset(), Cond,
2391 CondUse->getOperandValToReplace());
2392 CondUse = &IU->IVUsesByStride[*CondStride]->Users.back();
2396 // If we get to here, we know that we can transform the setcc instruction to
2397 // use the post-incremented version of the IV, allowing us to coalesce the
2398 // live ranges for the IV correctly.
2399 CondUse->setOffset(SE->getMinusSCEV(CondUse->getOffset(), *CondStride));
2400 CondUse->setIsUseOfPostIncrementedValue(true);
2406 /// OptimizeLoopCountIV - If, after all sharing of IVs, the IV used for deciding
2407 /// when to exit the loop is used only for that purpose, try to rearrange things
2408 /// so it counts down to a test against zero.
2409 void LoopStrengthReduce::OptimizeLoopCountIV(Loop *L) {
2411 // If the number of times the loop is executed isn't computable, give up.
2412 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
2413 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2416 // Get the terminating condition for the loop if possible (this isn't
2417 // necessarily in the latch, or a block that's a predecessor of the header).
2418 if (!L->getExitBlock())
2419 return; // More than one loop exit blocks.
2421 // Okay, there is one exit block. Try to find the condition that causes the
2422 // loop to be exited.
2423 BasicBlock *ExitingBlock = L->getExitingBlock();
2425 return; // More than one block exiting!
2427 // Okay, we've computed the exiting block. See what condition causes us to
2430 // FIXME: we should be able to handle switch instructions (with a single exit)
2431 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2432 if (TermBr == 0) return;
2433 assert(TermBr->isConditional() && "If unconditional, it can't be in loop!");
2434 if (!isa<ICmpInst>(TermBr->getCondition()))
2436 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2438 // Handle only tests for equality for the moment, and only stride 1.
2439 if (Cond->getPredicate() != CmpInst::ICMP_EQ)
2441 const SCEV *IV = SE->getSCEV(Cond->getOperand(0));
2442 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2443 const SCEV *One = SE->getIntegerSCEV(1, BackedgeTakenCount->getType());
2444 if (!AR || !AR->isAffine() || AR->getStepRecurrence(*SE) != One)
2446 // If the RHS of the comparison is defined inside the loop, the rewrite
2448 if (Instruction *CR = dyn_cast<Instruction>(Cond->getOperand(1)))
2449 if (L->contains(CR->getParent()))
2452 // Make sure the IV is only used for counting. Value may be preinc or
2453 // postinc; 2 uses in either case.
2454 if (!Cond->getOperand(0)->hasNUses(2))
2456 PHINode *phi = dyn_cast<PHINode>(Cond->getOperand(0));
2458 if (phi && phi->getParent()==L->getHeader()) {
2459 // value tested is preinc. Find the increment.
2460 // A CmpInst is not a BinaryOperator; we depend on this.
2461 Instruction::use_iterator UI = phi->use_begin();
2462 incr = dyn_cast<BinaryOperator>(UI);
2464 incr = dyn_cast<BinaryOperator>(++UI);
2465 // 1 use for postinc value, the phi. Unnecessarily conservative?
2466 if (!incr || !incr->hasOneUse() || incr->getOpcode()!=Instruction::Add)
2469 // Value tested is postinc. Find the phi node.
2470 incr = dyn_cast<BinaryOperator>(Cond->getOperand(0));
2471 if (!incr || incr->getOpcode()!=Instruction::Add)
2474 Instruction::use_iterator UI = Cond->getOperand(0)->use_begin();
2475 phi = dyn_cast<PHINode>(UI);
2477 phi = dyn_cast<PHINode>(++UI);
2478 // 1 use for preinc value, the increment.
2479 if (!phi || phi->getParent()!=L->getHeader() || !phi->hasOneUse())
2483 // Replace the increment with a decrement.
2484 BinaryOperator *decr =
2485 BinaryOperator::Create(Instruction::Sub, incr->getOperand(0),
2486 incr->getOperand(1), "tmp", incr);
2487 incr->replaceAllUsesWith(decr);
2488 incr->eraseFromParent();
2490 // Substitute endval-startval for the original startval, and 0 for the
2491 // original endval. Since we're only testing for equality this is OK even
2492 // if the computation wraps around.
2493 BasicBlock *Preheader = L->getLoopPreheader();
2494 Instruction *PreInsertPt = Preheader->getTerminator();
2495 int inBlock = L->contains(phi->getIncomingBlock(0)) ? 1 : 0;
2496 Value *startVal = phi->getIncomingValue(inBlock);
2497 Value *endVal = Cond->getOperand(1);
2498 // FIXME check for case where both are constant
2499 Constant* Zero = Context->getConstantInt(Cond->getOperand(1)->getType(), 0);
2500 BinaryOperator *NewStartVal =
2501 BinaryOperator::Create(Instruction::Sub, endVal, startVal,
2502 "tmp", PreInsertPt);
2503 phi->setIncomingValue(inBlock, NewStartVal);
2504 Cond->setOperand(1, Zero);
2509 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager &LPM) {
2511 IU = &getAnalysis<IVUsers>();
2512 LI = &getAnalysis<LoopInfo>();
2513 DT = &getAnalysis<DominatorTree>();
2514 SE = &getAnalysis<ScalarEvolution>();
2517 if (!IU->IVUsesByStride.empty()) {
2519 DOUT << "\nLSR on \"" << L->getHeader()->getParent()->getNameStart()
2524 // Sort the StrideOrder so we process larger strides first.
2525 std::stable_sort(IU->StrideOrder.begin(), IU->StrideOrder.end(),
2528 // Optimize induction variables. Some indvar uses can be transformed to use
2529 // strides that will be needed for other purposes. A common example of this
2530 // is the exit test for the loop, which can often be rewritten to use the
2531 // computation of some other indvar to decide when to terminate the loop.
2534 // Change loop terminating condition to use the postinc iv when possible
2535 // and optimize loop terminating compare. FIXME: Move this after
2536 // StrengthReduceStridedIVUsers?
2537 OptimizeLoopTermCond(L);
2539 // FIXME: We can shrink overlarge IV's here. e.g. if the code has
2540 // computation in i64 values and the target doesn't support i64, demote
2541 // the computation to 32-bit if safe.
2543 // FIXME: Attempt to reuse values across multiple IV's. In particular, we
2544 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should
2545 // be codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC.
2546 // Need to be careful that IV's are all the same type. Only works for
2547 // intptr_t indvars.
2549 // IVsByStride keeps IVs for one particular loop.
2550 assert(IVsByStride.empty() && "Stale entries in IVsByStride?");
2552 // Note: this processes each stride/type pair individually. All users
2553 // passed into StrengthReduceStridedIVUsers have the same type AND stride.
2554 // Also, note that we iterate over IVUsesByStride indirectly by using
2555 // StrideOrder. This extra layer of indirection makes the ordering of
2556 // strides deterministic - not dependent on map order.
2557 for (unsigned Stride = 0, e = IU->StrideOrder.size();
2558 Stride != e; ++Stride) {
2559 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
2560 IU->IVUsesByStride.find(IU->StrideOrder[Stride]);
2561 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
2562 // FIXME: Generalize to non-affine IV's.
2563 if (!SI->first->isLoopInvariant(L))
2565 StrengthReduceStridedIVUsers(SI->first, *SI->second, L);
2569 // After all sharing is done, see if we can adjust the loop to test against
2570 // zero instead of counting up to a maximum. This is usually faster.
2571 OptimizeLoopCountIV(L);
2573 // We're done analyzing this loop; release all the state we built up for it.
2574 IVsByStride.clear();
2575 StrideNoReuse.clear();
2577 // Clean up after ourselves
2578 if (!DeadInsts.empty())
2579 DeleteTriviallyDeadInstructions();
2581 // At this point, it is worth checking to see if any recurrence PHIs are also
2582 // dead, so that we can remove them as well.
2583 DeleteDeadPHIs(L->getHeader());