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 pass performs a strength reduction on array references inside loops that
11 // have as one or more of their components the loop induction variable.
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "loop-reduce"
16 #include "llvm/Transforms/Scalar.h"
17 #include "llvm/Constants.h"
18 #include "llvm/Instructions.h"
19 #include "llvm/IntrinsicInst.h"
20 #include "llvm/Type.h"
21 #include "llvm/DerivedTypes.h"
22 #include "llvm/Analysis/Dominators.h"
23 #include "llvm/Analysis/LoopInfo.h"
24 #include "llvm/Analysis/LoopPass.h"
25 #include "llvm/Analysis/ScalarEvolutionExpander.h"
26 #include "llvm/Transforms/Utils/AddrModeMatcher.h"
27 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
28 #include "llvm/Transforms/Utils/Local.h"
29 #include "llvm/ADT/SmallPtrSet.h"
30 #include "llvm/ADT/Statistic.h"
31 #include "llvm/Support/CFG.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/Compiler.h"
34 #include "llvm/Support/CommandLine.h"
35 #include "llvm/Support/ValueHandle.h"
36 #include "llvm/Target/TargetLowering.h"
40 STATISTIC(NumReduced , "Number of IV uses strength reduced");
41 STATISTIC(NumInserted, "Number of PHIs inserted");
42 STATISTIC(NumVariable, "Number of PHIs with variable strides");
43 STATISTIC(NumEliminated, "Number of strides eliminated");
44 STATISTIC(NumShadow, "Number of Shadow IVs optimized");
45 STATISTIC(NumImmSunk, "Number of common expr immediates sunk into uses");
47 static cl::opt<bool> EnableFullLSRMode("enable-full-lsr",
55 /// IVStrideUse - Keep track of one use of a strided induction variable, where
56 /// the stride is stored externally. The Offset member keeps track of the
57 /// offset from the IV, User is the actual user of the operand, and
58 /// 'OperandValToReplace' is the operand of the User that is the use.
59 struct VISIBILITY_HIDDEN IVStrideUse {
62 Value *OperandValToReplace;
64 // isUseOfPostIncrementedValue - True if this should use the
65 // post-incremented version of this IV, not the preincremented version.
66 // This can only be set in special cases, such as the terminating setcc
67 // instruction for a loop or uses dominated by the loop.
68 bool isUseOfPostIncrementedValue;
70 IVStrideUse(const SCEVHandle &Offs, Instruction *U, Value *O)
71 : Offset(Offs), User(U), OperandValToReplace(O),
72 isUseOfPostIncrementedValue(false) {}
75 /// IVUsersOfOneStride - This structure keeps track of all instructions that
76 /// have an operand that is based on the trip count multiplied by some stride.
77 /// The stride for all of these users is common and kept external to this
79 struct VISIBILITY_HIDDEN IVUsersOfOneStride {
80 /// Users - Keep track of all of the users of this stride as well as the
81 /// initial value and the operand that uses the IV.
82 std::vector<IVStrideUse> Users;
84 void addUser(const SCEVHandle &Offset,Instruction *User, Value *Operand) {
85 Users.push_back(IVStrideUse(Offset, User, Operand));
89 /// IVInfo - This structure keeps track of one IV expression inserted during
90 /// StrengthReduceStridedIVUsers. It contains the stride, the common base, as
91 /// well as the PHI node and increment value created for rewrite.
92 struct VISIBILITY_HIDDEN IVExpr {
97 IVExpr(const SCEVHandle &stride, const SCEVHandle &base, PHINode *phi)
98 : Stride(stride), Base(base), PHI(phi) {}
101 /// IVsOfOneStride - This structure keeps track of all IV expression inserted
102 /// during StrengthReduceStridedIVUsers for a particular stride of the IV.
103 struct VISIBILITY_HIDDEN IVsOfOneStride {
104 std::vector<IVExpr> IVs;
106 void addIV(const SCEVHandle &Stride, const SCEVHandle &Base, PHINode *PHI) {
107 IVs.push_back(IVExpr(Stride, Base, PHI));
111 class VISIBILITY_HIDDEN LoopStrengthReduce : public LoopPass {
117 /// IVUsesByStride - Keep track of all uses of induction variables that we
118 /// are interested in. The key of the map is the stride of the access.
119 std::map<SCEVHandle, IVUsersOfOneStride> IVUsesByStride;
121 /// IVsByStride - Keep track of all IVs that have been inserted for a
122 /// particular stride.
123 std::map<SCEVHandle, IVsOfOneStride> IVsByStride;
125 /// StrideOrder - An ordering of the keys in IVUsesByStride that is stable:
126 /// We use this to iterate over the IVUsesByStride collection without being
127 /// dependent on random ordering of pointers in the process.
128 SmallVector<SCEVHandle, 16> StrideOrder;
130 /// DeadInsts - Keep track of instructions we may have made dead, so that
131 /// we can remove them after we are done working.
132 SmallVector<Instruction*, 16> DeadInsts;
134 /// TLI - Keep a pointer of a TargetLowering to consult for determining
135 /// transformation profitability.
136 const TargetLowering *TLI;
139 static char ID; // Pass ID, replacement for typeid
140 explicit LoopStrengthReduce(const TargetLowering *tli = NULL) :
141 LoopPass(&ID), TLI(tli) {
144 bool runOnLoop(Loop *L, LPPassManager &LPM);
146 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
147 // We split critical edges, so we change the CFG. However, we do update
148 // many analyses if they are around.
149 AU.addPreservedID(LoopSimplifyID);
150 AU.addPreserved<LoopInfo>();
151 AU.addPreserved<DominanceFrontier>();
152 AU.addPreserved<DominatorTree>();
154 AU.addRequiredID(LoopSimplifyID);
155 AU.addRequired<LoopInfo>();
156 AU.addRequired<DominatorTree>();
157 AU.addRequired<ScalarEvolution>();
158 AU.addPreserved<ScalarEvolution>();
162 bool AddUsersIfInteresting(Instruction *I, Loop *L,
163 SmallPtrSet<Instruction*,16> &Processed);
164 ICmpInst *ChangeCompareStride(Loop *L, ICmpInst *Cond,
165 IVStrideUse* &CondUse,
166 const SCEVHandle* &CondStride);
167 void OptimizeIndvars(Loop *L);
169 /// OptimizeShadowIV - If IV is used in a int-to-float cast
170 /// inside the loop then try to eliminate the cast opeation.
171 void OptimizeShadowIV(Loop *L);
173 /// OptimizeSMax - Rewrite the loop's terminating condition
174 /// if it uses an smax computation.
175 ICmpInst *OptimizeSMax(Loop *L, ICmpInst *Cond,
176 IVStrideUse* &CondUse);
178 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse,
179 const SCEVHandle *&CondStride);
180 bool RequiresTypeConversion(const Type *Ty, const Type *NewTy);
181 SCEVHandle CheckForIVReuse(bool, bool, bool, const SCEVHandle&,
182 IVExpr&, const Type*,
183 const std::vector<BasedUser>& UsersToProcess);
184 bool ValidStride(bool, int64_t,
185 const std::vector<BasedUser>& UsersToProcess);
186 SCEVHandle CollectIVUsers(const SCEVHandle &Stride,
187 IVUsersOfOneStride &Uses,
189 bool &AllUsesAreAddresses,
190 bool &AllUsesAreOutsideLoop,
191 std::vector<BasedUser> &UsersToProcess);
192 bool ShouldUseFullStrengthReductionMode(
193 const std::vector<BasedUser> &UsersToProcess,
195 bool AllUsesAreAddresses,
197 void PrepareToStrengthReduceFully(
198 std::vector<BasedUser> &UsersToProcess,
200 SCEVHandle CommonExprs,
202 SCEVExpander &PreheaderRewriter);
203 void PrepareToStrengthReduceFromSmallerStride(
204 std::vector<BasedUser> &UsersToProcess,
206 const IVExpr &ReuseIV,
207 Instruction *PreInsertPt);
208 void PrepareToStrengthReduceWithNewPhi(
209 std::vector<BasedUser> &UsersToProcess,
211 SCEVHandle CommonExprs,
214 SCEVExpander &PreheaderRewriter);
215 void StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
216 IVUsersOfOneStride &Uses,
218 void DeleteTriviallyDeadInstructions();
222 char LoopStrengthReduce::ID = 0;
223 static RegisterPass<LoopStrengthReduce>
224 X("loop-reduce", "Loop Strength Reduction");
226 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
227 return new LoopStrengthReduce(TLI);
230 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
231 /// specified set are trivially dead, delete them and see if this makes any of
232 /// their operands subsequently dead.
233 void LoopStrengthReduce::DeleteTriviallyDeadInstructions() {
234 if (DeadInsts.empty()) return;
236 // Sort the deadinsts list so that we can trivially eliminate duplicates as we
237 // go. The code below never adds a non-dead instruction to the worklist, but
238 // callers may not be so careful.
239 array_pod_sort(DeadInsts.begin(), DeadInsts.end());
241 // Drop duplicate instructions and those with uses.
242 for (unsigned i = 0, e = DeadInsts.size()-1; i < e; ++i) {
243 Instruction *I = DeadInsts[i];
244 if (!I->use_empty()) DeadInsts[i] = 0;
245 while (i != e && DeadInsts[i+1] == I)
249 while (!DeadInsts.empty()) {
250 Instruction *I = DeadInsts.back();
251 DeadInsts.pop_back();
253 if (I == 0 || !isInstructionTriviallyDead(I))
256 SE->deleteValueFromRecords(I);
258 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) {
259 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
262 DeadInsts.push_back(U);
266 I->eraseFromParent();
271 /// containsAddRecFromDifferentLoop - Determine whether expression S involves a
272 /// subexpression that is an AddRec from a loop other than L. An outer loop
273 /// of L is OK, but not an inner loop nor a disjoint loop.
274 static bool containsAddRecFromDifferentLoop(SCEVHandle S, Loop *L) {
275 // This is very common, put it first.
276 if (isa<SCEVConstant>(S))
278 if (const SCEVCommutativeExpr *AE = dyn_cast<SCEVCommutativeExpr>(S)) {
279 for (unsigned int i=0; i< AE->getNumOperands(); i++)
280 if (containsAddRecFromDifferentLoop(AE->getOperand(i), L))
284 if (const SCEVAddRecExpr *AE = dyn_cast<SCEVAddRecExpr>(S)) {
285 if (const Loop *newLoop = AE->getLoop()) {
288 // if newLoop is an outer loop of L, this is OK.
289 if (!LoopInfoBase<BasicBlock>::isNotAlreadyContainedIn(L, newLoop))
294 if (const SCEVUDivExpr *DE = dyn_cast<SCEVUDivExpr>(S))
295 return containsAddRecFromDifferentLoop(DE->getLHS(), L) ||
296 containsAddRecFromDifferentLoop(DE->getRHS(), L);
298 // SCEVSDivExpr has been backed out temporarily, but will be back; we'll
299 // need this when it is.
300 if (const SCEVSDivExpr *DE = dyn_cast<SCEVSDivExpr>(S))
301 return containsAddRecFromDifferentLoop(DE->getLHS(), L) ||
302 containsAddRecFromDifferentLoop(DE->getRHS(), L);
304 if (const SCEVCastExpr *CE = dyn_cast<SCEVCastExpr>(S))
305 return containsAddRecFromDifferentLoop(CE->getOperand(), L);
309 /// getSCEVStartAndStride - Compute the start and stride of this expression,
310 /// returning false if the expression is not a start/stride pair, or true if it
311 /// is. The stride must be a loop invariant expression, but the start may be
312 /// a mix of loop invariant and loop variant expressions. The start cannot,
313 /// however, contain an AddRec from a different loop, unless that loop is an
314 /// outer loop of the current loop.
315 static bool getSCEVStartAndStride(const SCEVHandle &SH, Loop *L,
316 SCEVHandle &Start, SCEVHandle &Stride,
317 ScalarEvolution *SE, DominatorTree *DT) {
318 SCEVHandle TheAddRec = Start; // Initialize to zero.
320 // If the outer level is an AddExpr, the operands are all start values except
321 // for a nested AddRecExpr.
322 if (const SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(SH)) {
323 for (unsigned i = 0, e = AE->getNumOperands(); i != e; ++i)
324 if (SCEVAddRecExpr *AddRec =
325 dyn_cast<SCEVAddRecExpr>(AE->getOperand(i))) {
326 if (AddRec->getLoop() == L)
327 TheAddRec = SE->getAddExpr(AddRec, TheAddRec);
329 return false; // Nested IV of some sort?
331 Start = SE->getAddExpr(Start, AE->getOperand(i));
334 } else if (isa<SCEVAddRecExpr>(SH)) {
337 return false; // not analyzable.
340 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(TheAddRec);
341 if (!AddRec || AddRec->getLoop() != L) return false;
343 // FIXME: Generalize to non-affine IV's.
344 if (!AddRec->isAffine()) return false;
346 // If Start contains an SCEVAddRecExpr from a different loop, other than an
347 // outer loop of the current loop, reject it. SCEV has no concept of
348 // operating on more than one loop at a time so don't confuse it with such
350 if (containsAddRecFromDifferentLoop(AddRec->getOperand(0), L))
353 Start = SE->getAddExpr(Start, AddRec->getOperand(0));
355 if (!isa<SCEVConstant>(AddRec->getOperand(1))) {
356 // If stride is an instruction, make sure it dominates the loop preheader.
357 // Otherwise we could end up with a use before def situation.
358 BasicBlock *Preheader = L->getLoopPreheader();
359 if (!AddRec->getOperand(1)->dominates(Preheader, DT))
362 DOUT << "[" << L->getHeader()->getName()
363 << "] Variable stride: " << *AddRec << "\n";
366 Stride = AddRec->getOperand(1);
370 /// IVUseShouldUsePostIncValue - We have discovered a "User" of an IV expression
371 /// and now we need to decide whether the user should use the preinc or post-inc
372 /// value. If this user should use the post-inc version of the IV, return true.
374 /// Choosing wrong here can break dominance properties (if we choose to use the
375 /// post-inc value when we cannot) or it can end up adding extra live-ranges to
376 /// the loop, resulting in reg-reg copies (if we use the pre-inc value when we
377 /// should use the post-inc value).
378 static bool IVUseShouldUsePostIncValue(Instruction *User, Instruction *IV,
379 Loop *L, DominatorTree *DT, Pass *P,
380 SmallVectorImpl<Instruction*> &DeadInsts){
381 // If the user is in the loop, use the preinc value.
382 if (L->contains(User->getParent())) return false;
384 BasicBlock *LatchBlock = L->getLoopLatch();
386 // Ok, the user is outside of the loop. If it is dominated by the latch
387 // block, use the post-inc value.
388 if (DT->dominates(LatchBlock, User->getParent()))
391 // There is one case we have to be careful of: PHI nodes. These little guys
392 // can live in blocks that do not dominate the latch block, but (since their
393 // uses occur in the predecessor block, not the block the PHI lives in) should
394 // still use the post-inc value. Check for this case now.
395 PHINode *PN = dyn_cast<PHINode>(User);
396 if (!PN) return false; // not a phi, not dominated by latch block.
398 // Look at all of the uses of IV by the PHI node. If any use corresponds to
399 // a block that is not dominated by the latch block, give up and use the
400 // preincremented value.
401 unsigned NumUses = 0;
402 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
403 if (PN->getIncomingValue(i) == IV) {
405 if (!DT->dominates(LatchBlock, PN->getIncomingBlock(i)))
409 // Okay, all uses of IV by PN are in predecessor blocks that really are
410 // dominated by the latch block. Use the post-incremented value.
414 /// isAddressUse - Returns true if the specified instruction is using the
415 /// specified value as an address.
416 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
417 bool isAddress = isa<LoadInst>(Inst);
418 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
419 if (SI->getOperand(1) == OperandVal)
421 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
422 // Addressing modes can also be folded into prefetches and a variety
424 switch (II->getIntrinsicID()) {
426 case Intrinsic::prefetch:
427 case Intrinsic::x86_sse2_loadu_dq:
428 case Intrinsic::x86_sse2_loadu_pd:
429 case Intrinsic::x86_sse_loadu_ps:
430 case Intrinsic::x86_sse_storeu_ps:
431 case Intrinsic::x86_sse2_storeu_pd:
432 case Intrinsic::x86_sse2_storeu_dq:
433 case Intrinsic::x86_sse2_storel_dq:
434 if (II->getOperand(1) == OperandVal)
442 /// getAccessType - Return the type of the memory being accessed.
443 static const Type *getAccessType(const Instruction *Inst) {
444 const Type *UseTy = Inst->getType();
445 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
446 UseTy = SI->getOperand(0)->getType();
447 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
448 // Addressing modes can also be folded into prefetches and a variety
450 switch (II->getIntrinsicID()) {
452 case Intrinsic::x86_sse_storeu_ps:
453 case Intrinsic::x86_sse2_storeu_pd:
454 case Intrinsic::x86_sse2_storeu_dq:
455 case Intrinsic::x86_sse2_storel_dq:
456 UseTy = II->getOperand(1)->getType();
463 /// AddUsersIfInteresting - Inspect the specified instruction. If it is a
464 /// reducible SCEV, recursively add its users to the IVUsesByStride set and
465 /// return true. Otherwise, return false.
466 bool LoopStrengthReduce::AddUsersIfInteresting(Instruction *I, Loop *L,
467 SmallPtrSet<Instruction*,16> &Processed) {
468 if (!SE->isSCEVable(I->getType()))
469 return false; // Void and FP expressions cannot be reduced.
471 // LSR is not APInt clean, do not touch integers bigger than 64-bits.
472 if (SE->getTypeSizeInBits(I->getType()) > 64)
475 if (!Processed.insert(I))
476 return true; // Instruction already handled.
478 // Get the symbolic expression for this instruction.
479 SCEVHandle ISE = SE->getSCEV(I);
480 if (isa<SCEVCouldNotCompute>(ISE)) return false;
482 // Get the start and stride for this expression.
483 SCEVHandle Start = SE->getIntegerSCEV(0, ISE->getType());
484 SCEVHandle Stride = Start;
485 if (!getSCEVStartAndStride(ISE, L, Start, Stride, SE, DT))
486 return false; // Non-reducible symbolic expression, bail out.
488 std::vector<Instruction *> IUsers;
489 // Collect all I uses now because IVUseShouldUsePostIncValue may
490 // invalidate use_iterator.
491 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; ++UI)
492 IUsers.push_back(cast<Instruction>(*UI));
494 for (unsigned iused_index = 0, iused_size = IUsers.size();
495 iused_index != iused_size; ++iused_index) {
497 Instruction *User = IUsers[iused_index];
499 // Do not infinitely recurse on PHI nodes.
500 if (isa<PHINode>(User) && Processed.count(User))
503 // Descend recursively, but not into PHI nodes outside the current loop.
504 // It's important to see the entire expression outside the loop to get
505 // choices that depend on addressing mode use right, although we won't
506 // consider references ouside the loop in all cases.
507 // If User is already in Processed, we don't want to recurse into it again,
508 // but do want to record a second reference in the same instruction.
509 bool AddUserToIVUsers = false;
510 if (LI->getLoopFor(User->getParent()) != L) {
511 if (isa<PHINode>(User) || Processed.count(User) ||
512 !AddUsersIfInteresting(User, L, Processed)) {
513 DOUT << "FOUND USER in other loop: " << *User
514 << " OF SCEV: " << *ISE << "\n";
515 AddUserToIVUsers = true;
517 } else if (Processed.count(User) ||
518 !AddUsersIfInteresting(User, L, Processed)) {
519 DOUT << "FOUND USER: " << *User
520 << " OF SCEV: " << *ISE << "\n";
521 AddUserToIVUsers = true;
524 if (AddUserToIVUsers) {
525 IVUsersOfOneStride &StrideUses = IVUsesByStride[Stride];
526 if (StrideUses.Users.empty()) // First occurrence of this stride?
527 StrideOrder.push_back(Stride);
529 // Okay, we found a user that we cannot reduce. Analyze the instruction
530 // and decide what to do with it. If we are a use inside of the loop, use
531 // the value before incrementation, otherwise use it after incrementation.
532 if (IVUseShouldUsePostIncValue(User, I, L, DT, this, DeadInsts)) {
533 // The value used will be incremented by the stride more than we are
534 // expecting, so subtract this off.
535 SCEVHandle NewStart = SE->getMinusSCEV(Start, Stride);
536 StrideUses.addUser(NewStart, User, I);
537 StrideUses.Users.back().isUseOfPostIncrementedValue = true;
538 DOUT << " USING POSTINC SCEV, START=" << *NewStart<< "\n";
540 StrideUses.addUser(Start, User, I);
548 /// BasedUser - For a particular base value, keep information about how we've
549 /// partitioned the expression so far.
551 /// SE - The current ScalarEvolution object.
554 /// Base - The Base value for the PHI node that needs to be inserted for
555 /// this use. As the use is processed, information gets moved from this
556 /// field to the Imm field (below). BasedUser values are sorted by this
560 /// Inst - The instruction using the induction variable.
563 /// OperandValToReplace - The operand value of Inst to replace with the
565 Value *OperandValToReplace;
567 /// Imm - The immediate value that should be added to the base immediately
568 /// before Inst, because it will be folded into the imm field of the
569 /// instruction. This is also sometimes used for loop-variant values that
570 /// must be added inside the loop.
573 /// Phi - The induction variable that performs the striding that
574 /// should be used for this user.
577 // isUseOfPostIncrementedValue - True if this should use the
578 // post-incremented version of this IV, not the preincremented version.
579 // This can only be set in special cases, such as the terminating setcc
580 // instruction for a loop and uses outside the loop that are dominated by
582 bool isUseOfPostIncrementedValue;
584 BasedUser(IVStrideUse &IVSU, ScalarEvolution *se)
585 : SE(se), Base(IVSU.Offset), Inst(IVSU.User),
586 OperandValToReplace(IVSU.OperandValToReplace),
587 Imm(SE->getIntegerSCEV(0, Base->getType())),
588 isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue) {}
590 // Once we rewrite the code to insert the new IVs we want, update the
591 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
593 void RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
594 Instruction *InsertPt,
595 SCEVExpander &Rewriter, Loop *L, Pass *P,
596 SmallVectorImpl<Instruction*> &DeadInsts);
598 Value *InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
600 SCEVExpander &Rewriter,
601 Instruction *IP, Loop *L);
606 void BasedUser::dump() const {
607 cerr << " Base=" << *Base;
608 cerr << " Imm=" << *Imm;
609 cerr << " Inst: " << *Inst;
612 Value *BasedUser::InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
614 SCEVExpander &Rewriter,
615 Instruction *IP, Loop *L) {
616 // Figure out where we *really* want to insert this code. In particular, if
617 // the user is inside of a loop that is nested inside of L, we really don't
618 // want to insert this expression before the user, we'd rather pull it out as
619 // many loops as possible.
620 LoopInfo &LI = Rewriter.getLoopInfo();
621 Instruction *BaseInsertPt = IP;
623 // Figure out the most-nested loop that IP is in.
624 Loop *InsertLoop = LI.getLoopFor(IP->getParent());
626 // If InsertLoop is not L, and InsertLoop is nested inside of L, figure out
627 // the preheader of the outer-most loop where NewBase is not loop invariant.
628 if (L->contains(IP->getParent()))
629 while (InsertLoop && NewBase->isLoopInvariant(InsertLoop)) {
630 BaseInsertPt = InsertLoop->getLoopPreheader()->getTerminator();
631 InsertLoop = InsertLoop->getParentLoop();
634 Value *Base = Rewriter.expandCodeFor(NewBase, Ty, BaseInsertPt);
636 // If there is no immediate value, skip the next part.
640 // If we are inserting the base and imm values in the same block, make sure to
641 // adjust the IP position if insertion reused a result.
642 if (IP == BaseInsertPt)
643 IP = Rewriter.getInsertionPoint();
645 // Always emit the immediate (if non-zero) into the same block as the user.
646 SCEVHandle NewValSCEV = SE->getAddExpr(SE->getUnknown(Base), Imm);
647 return Rewriter.expandCodeFor(NewValSCEV, Ty, IP);
651 // Once we rewrite the code to insert the new IVs we want, update the
652 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
653 // to it. NewBasePt is the last instruction which contributes to the
654 // value of NewBase in the case that it's a diffferent instruction from
655 // the PHI that NewBase is computed from, or null otherwise.
657 void BasedUser::RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
658 Instruction *NewBasePt,
659 SCEVExpander &Rewriter, Loop *L, Pass *P,
660 SmallVectorImpl<Instruction*> &DeadInsts){
661 if (!isa<PHINode>(Inst)) {
662 // By default, insert code at the user instruction.
663 BasicBlock::iterator InsertPt = Inst;
665 // However, if the Operand is itself an instruction, the (potentially
666 // complex) inserted code may be shared by many users. Because of this, we
667 // want to emit code for the computation of the operand right before its old
668 // computation. This is usually safe, because we obviously used to use the
669 // computation when it was computed in its current block. However, in some
670 // cases (e.g. use of a post-incremented induction variable) the NewBase
671 // value will be pinned to live somewhere after the original computation.
672 // In this case, we have to back off.
674 // If this is a use outside the loop (which means after, since it is based
675 // on a loop indvar) we use the post-incremented value, so that we don't
676 // artificially make the preinc value live out the bottom of the loop.
677 if (!isUseOfPostIncrementedValue && L->contains(Inst->getParent())) {
678 if (NewBasePt && isa<PHINode>(OperandValToReplace)) {
679 InsertPt = NewBasePt;
681 } else if (Instruction *OpInst
682 = dyn_cast<Instruction>(OperandValToReplace)) {
684 while (isa<PHINode>(InsertPt)) ++InsertPt;
687 Value *NewVal = InsertCodeForBaseAtPosition(NewBase,
688 OperandValToReplace->getType(),
689 Rewriter, InsertPt, L);
690 // Replace the use of the operand Value with the new Phi we just created.
691 Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
693 DOUT << " Replacing with ";
694 DEBUG(WriteAsOperand(*DOUT, NewVal, /*PrintType=*/false));
695 DOUT << ", which has value " << *NewBase << " plus IMM " << *Imm << "\n";
699 // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm
700 // expression into each operand block that uses it. Note that PHI nodes can
701 // have multiple entries for the same predecessor. We use a map to make sure
702 // that a PHI node only has a single Value* for each predecessor (which also
703 // prevents us from inserting duplicate code in some blocks).
704 DenseMap<BasicBlock*, Value*> InsertedCode;
705 PHINode *PN = cast<PHINode>(Inst);
706 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
707 if (PN->getIncomingValue(i) == OperandValToReplace) {
708 // If the original expression is outside the loop, put the replacement
709 // code in the same place as the original expression,
710 // which need not be an immediate predecessor of this PHI. This way we
711 // need only one copy of it even if it is referenced multiple times in
712 // the PHI. We don't do this when the original expression is inside the
713 // loop because multiple copies sometimes do useful sinking of code in
715 Instruction *OldLoc = dyn_cast<Instruction>(OperandValToReplace);
716 if (L->contains(OldLoc->getParent())) {
717 // If this is a critical edge, split the edge so that we do not insert
718 // the code on all predecessor/successor paths. We do this unless this
719 // is the canonical backedge for this loop, as this can make some
720 // inserted code be in an illegal position.
721 BasicBlock *PHIPred = PN->getIncomingBlock(i);
722 if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 &&
723 (PN->getParent() != L->getHeader() || !L->contains(PHIPred))) {
725 // First step, split the critical edge.
726 SplitCriticalEdge(PHIPred, PN->getParent(), P, false);
728 // Next step: move the basic block. In particular, if the PHI node
729 // is outside of the loop, and PredTI is in the loop, we want to
730 // move the block to be immediately before the PHI block, not
731 // immediately after PredTI.
732 if (L->contains(PHIPred) && !L->contains(PN->getParent())) {
733 BasicBlock *NewBB = PN->getIncomingBlock(i);
734 NewBB->moveBefore(PN->getParent());
737 // Splitting the edge can reduce the number of PHI entries we have.
738 e = PN->getNumIncomingValues();
741 Value *&Code = InsertedCode[PN->getIncomingBlock(i)];
743 // Insert the code into the end of the predecessor block.
744 Instruction *InsertPt = (L->contains(OldLoc->getParent())) ?
745 PN->getIncomingBlock(i)->getTerminator() :
746 OldLoc->getParent()->getTerminator();
747 Code = InsertCodeForBaseAtPosition(NewBase, PN->getType(),
748 Rewriter, InsertPt, L);
750 DOUT << " Changing PHI use to ";
751 DEBUG(WriteAsOperand(*DOUT, Code, /*PrintType=*/false));
752 DOUT << ", which has value " << *NewBase << " plus IMM " << *Imm << "\n";
755 // Replace the use of the operand Value with the new Phi we just created.
756 PN->setIncomingValue(i, Code);
761 // PHI node might have become a constant value after SplitCriticalEdge.
762 DeadInsts.push_back(Inst);
766 /// fitsInAddressMode - Return true if V can be subsumed within an addressing
767 /// mode, and does not need to be put in a register first.
768 static bool fitsInAddressMode(const SCEVHandle &V, const Type *UseTy,
769 const TargetLowering *TLI, bool HasBaseReg) {
770 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
771 int64_t VC = SC->getValue()->getSExtValue();
773 TargetLowering::AddrMode AM;
775 AM.HasBaseReg = HasBaseReg;
776 return TLI->isLegalAddressingMode(AM, UseTy);
778 // Defaults to PPC. PPC allows a sign-extended 16-bit immediate field.
779 return (VC > -(1 << 16) && VC < (1 << 16)-1);
783 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
784 if (GlobalValue *GV = dyn_cast<GlobalValue>(SU->getValue())) {
786 TargetLowering::AddrMode AM;
788 AM.HasBaseReg = HasBaseReg;
789 return TLI->isLegalAddressingMode(AM, UseTy);
791 // Default: assume global addresses are not legal.
798 /// MoveLoopVariantsToImmediateField - Move any subexpressions from Val that are
799 /// loop varying to the Imm operand.
800 static void MoveLoopVariantsToImmediateField(SCEVHandle &Val, SCEVHandle &Imm,
801 Loop *L, ScalarEvolution *SE) {
802 if (Val->isLoopInvariant(L)) return; // Nothing to do.
804 if (const SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
805 std::vector<SCEVHandle> NewOps;
806 NewOps.reserve(SAE->getNumOperands());
808 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
809 if (!SAE->getOperand(i)->isLoopInvariant(L)) {
810 // If this is a loop-variant expression, it must stay in the immediate
811 // field of the expression.
812 Imm = SE->getAddExpr(Imm, SAE->getOperand(i));
814 NewOps.push_back(SAE->getOperand(i));
818 Val = SE->getIntegerSCEV(0, Val->getType());
820 Val = SE->getAddExpr(NewOps);
821 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
822 // Try to pull immediates out of the start value of nested addrec's.
823 SCEVHandle Start = SARE->getStart();
824 MoveLoopVariantsToImmediateField(Start, Imm, L, SE);
826 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
828 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
830 // Otherwise, all of Val is variant, move the whole thing over.
831 Imm = SE->getAddExpr(Imm, Val);
832 Val = SE->getIntegerSCEV(0, Val->getType());
837 /// MoveImmediateValues - Look at Val, and pull out any additions of constants
838 /// that can fit into the immediate field of instructions in the target.
839 /// Accumulate these immediate values into the Imm value.
840 static void MoveImmediateValues(const TargetLowering *TLI,
842 SCEVHandle &Val, SCEVHandle &Imm,
843 bool isAddress, Loop *L,
844 ScalarEvolution *SE) {
845 if (const SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
846 std::vector<SCEVHandle> NewOps;
847 NewOps.reserve(SAE->getNumOperands());
849 for (unsigned i = 0; i != SAE->getNumOperands(); ++i) {
850 SCEVHandle NewOp = SAE->getOperand(i);
851 MoveImmediateValues(TLI, UseTy, NewOp, Imm, isAddress, L, SE);
853 if (!NewOp->isLoopInvariant(L)) {
854 // If this is a loop-variant expression, it must stay in the immediate
855 // field of the expression.
856 Imm = SE->getAddExpr(Imm, NewOp);
858 NewOps.push_back(NewOp);
863 Val = SE->getIntegerSCEV(0, Val->getType());
865 Val = SE->getAddExpr(NewOps);
867 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
868 // Try to pull immediates out of the start value of nested addrec's.
869 SCEVHandle Start = SARE->getStart();
870 MoveImmediateValues(TLI, UseTy, Start, Imm, isAddress, L, SE);
872 if (Start != SARE->getStart()) {
873 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
875 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
878 } else if (const SCEVMulExpr *SME = dyn_cast<SCEVMulExpr>(Val)) {
879 // Transform "8 * (4 + v)" -> "32 + 8*V" if "32" fits in the immed field.
880 if (isAddress && fitsInAddressMode(SME->getOperand(0), UseTy, TLI, false) &&
881 SME->getNumOperands() == 2 && SME->isLoopInvariant(L)) {
883 SCEVHandle SubImm = SE->getIntegerSCEV(0, Val->getType());
884 SCEVHandle NewOp = SME->getOperand(1);
885 MoveImmediateValues(TLI, UseTy, NewOp, SubImm, isAddress, L, SE);
887 // If we extracted something out of the subexpressions, see if we can
889 if (NewOp != SME->getOperand(1)) {
890 // Scale SubImm up by "8". If the result is a target constant, we are
892 SubImm = SE->getMulExpr(SubImm, SME->getOperand(0));
893 if (fitsInAddressMode(SubImm, UseTy, TLI, false)) {
894 // Accumulate the immediate.
895 Imm = SE->getAddExpr(Imm, SubImm);
897 // Update what is left of 'Val'.
898 Val = SE->getMulExpr(SME->getOperand(0), NewOp);
905 // Loop-variant expressions must stay in the immediate field of the
907 if ((isAddress && fitsInAddressMode(Val, UseTy, TLI, false)) ||
908 !Val->isLoopInvariant(L)) {
909 Imm = SE->getAddExpr(Imm, Val);
910 Val = SE->getIntegerSCEV(0, Val->getType());
914 // Otherwise, no immediates to move.
917 static void MoveImmediateValues(const TargetLowering *TLI,
919 SCEVHandle &Val, SCEVHandle &Imm,
920 bool isAddress, Loop *L,
921 ScalarEvolution *SE) {
922 const Type *UseTy = getAccessType(User);
923 MoveImmediateValues(TLI, UseTy, Val, Imm, isAddress, L, SE);
926 /// SeparateSubExprs - Decompose Expr into all of the subexpressions that are
927 /// added together. This is used to reassociate common addition subexprs
928 /// together for maximal sharing when rewriting bases.
929 static void SeparateSubExprs(std::vector<SCEVHandle> &SubExprs,
931 ScalarEvolution *SE) {
932 if (const SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) {
933 for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
934 SeparateSubExprs(SubExprs, AE->getOperand(j), SE);
935 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) {
936 SCEVHandle Zero = SE->getIntegerSCEV(0, Expr->getType());
937 if (SARE->getOperand(0) == Zero) {
938 SubExprs.push_back(Expr);
940 // Compute the addrec with zero as its base.
941 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
942 Ops[0] = Zero; // Start with zero base.
943 SubExprs.push_back(SE->getAddRecExpr(Ops, SARE->getLoop()));
946 SeparateSubExprs(SubExprs, SARE->getOperand(0), SE);
948 } else if (!Expr->isZero()) {
950 SubExprs.push_back(Expr);
954 // This is logically local to the following function, but C++ says we have
955 // to make it file scope.
956 struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; };
958 /// RemoveCommonExpressionsFromUseBases - Look through all of the Bases of all
959 /// the Uses, removing any common subexpressions, except that if all such
960 /// subexpressions can be folded into an addressing mode for all uses inside
961 /// the loop (this case is referred to as "free" in comments herein) we do
962 /// not remove anything. This looks for things like (a+b+c) and
963 /// (a+c+d) and computes the common (a+c) subexpression. The common expression
964 /// is *removed* from the Bases and returned.
966 RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses,
967 ScalarEvolution *SE, Loop *L,
968 const TargetLowering *TLI) {
969 unsigned NumUses = Uses.size();
971 // Only one use? This is a very common case, so we handle it specially and
973 SCEVHandle Zero = SE->getIntegerSCEV(0, Uses[0].Base->getType());
974 SCEVHandle Result = Zero;
975 SCEVHandle FreeResult = Zero;
977 // If the use is inside the loop, use its base, regardless of what it is:
978 // it is clearly shared across all the IV's. If the use is outside the loop
979 // (which means after it) we don't want to factor anything *into* the loop,
980 // so just use 0 as the base.
981 if (L->contains(Uses[0].Inst->getParent()))
982 std::swap(Result, Uses[0].Base);
986 // To find common subexpressions, count how many of Uses use each expression.
987 // If any subexpressions are used Uses.size() times, they are common.
988 // Also track whether all uses of each expression can be moved into an
989 // an addressing mode "for free"; such expressions are left within the loop.
990 // struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; };
991 std::map<SCEVHandle, SubExprUseData> SubExpressionUseData;
993 // UniqueSubExprs - Keep track of all of the subexpressions we see in the
994 // order we see them.
995 std::vector<SCEVHandle> UniqueSubExprs;
997 std::vector<SCEVHandle> SubExprs;
998 unsigned NumUsesInsideLoop = 0;
999 for (unsigned i = 0; i != NumUses; ++i) {
1000 // If the user is outside the loop, just ignore it for base computation.
1001 // Since the user is outside the loop, it must be *after* the loop (if it
1002 // were before, it could not be based on the loop IV). We don't want users
1003 // after the loop to affect base computation of values *inside* the loop,
1004 // because we can always add their offsets to the result IV after the loop
1005 // is done, ensuring we get good code inside the loop.
1006 if (!L->contains(Uses[i].Inst->getParent()))
1008 NumUsesInsideLoop++;
1010 // If the base is zero (which is common), return zero now, there are no
1011 // CSEs we can find.
1012 if (Uses[i].Base == Zero) return Zero;
1014 // If this use is as an address we may be able to put CSEs in the addressing
1015 // mode rather than hoisting them.
1016 bool isAddrUse = isAddressUse(Uses[i].Inst, Uses[i].OperandValToReplace);
1017 // We may need the UseTy below, but only when isAddrUse, so compute it
1018 // only in that case.
1019 const Type *UseTy = 0;
1021 UseTy = getAccessType(Uses[i].Inst);
1023 // Split the expression into subexprs.
1024 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
1025 // Add one to SubExpressionUseData.Count for each subexpr present, and
1026 // if the subexpr is not a valid immediate within an addressing mode use,
1027 // set SubExpressionUseData.notAllUsesAreFree. We definitely want to
1028 // hoist these out of the loop (if they are common to all uses).
1029 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) {
1030 if (++SubExpressionUseData[SubExprs[j]].Count == 1)
1031 UniqueSubExprs.push_back(SubExprs[j]);
1032 if (!isAddrUse || !fitsInAddressMode(SubExprs[j], UseTy, TLI, false))
1033 SubExpressionUseData[SubExprs[j]].notAllUsesAreFree = true;
1038 // Now that we know how many times each is used, build Result. Iterate over
1039 // UniqueSubexprs so that we have a stable ordering.
1040 for (unsigned i = 0, e = UniqueSubExprs.size(); i != e; ++i) {
1041 std::map<SCEVHandle, SubExprUseData>::iterator I =
1042 SubExpressionUseData.find(UniqueSubExprs[i]);
1043 assert(I != SubExpressionUseData.end() && "Entry not found?");
1044 if (I->second.Count == NumUsesInsideLoop) { // Found CSE!
1045 if (I->second.notAllUsesAreFree)
1046 Result = SE->getAddExpr(Result, I->first);
1048 FreeResult = SE->getAddExpr(FreeResult, I->first);
1050 // Remove non-cse's from SubExpressionUseData.
1051 SubExpressionUseData.erase(I);
1054 if (FreeResult != Zero) {
1055 // We have some subexpressions that can be subsumed into addressing
1056 // modes in every use inside the loop. However, it's possible that
1057 // there are so many of them that the combined FreeResult cannot
1058 // be subsumed, or that the target cannot handle both a FreeResult
1059 // and a Result in the same instruction (for example because it would
1060 // require too many registers). Check this.
1061 for (unsigned i=0; i<NumUses; ++i) {
1062 if (!L->contains(Uses[i].Inst->getParent()))
1064 // We know this is an addressing mode use; if there are any uses that
1065 // are not, FreeResult would be Zero.
1066 const Type *UseTy = getAccessType(Uses[i].Inst);
1067 if (!fitsInAddressMode(FreeResult, UseTy, TLI, Result!=Zero)) {
1068 // FIXME: could split up FreeResult into pieces here, some hoisted
1069 // and some not. There is no obvious advantage to this.
1070 Result = SE->getAddExpr(Result, FreeResult);
1077 // If we found no CSE's, return now.
1078 if (Result == Zero) return Result;
1080 // If we still have a FreeResult, remove its subexpressions from
1081 // SubExpressionUseData. This means they will remain in the use Bases.
1082 if (FreeResult != Zero) {
1083 SeparateSubExprs(SubExprs, FreeResult, SE);
1084 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) {
1085 std::map<SCEVHandle, SubExprUseData>::iterator I =
1086 SubExpressionUseData.find(SubExprs[j]);
1087 SubExpressionUseData.erase(I);
1092 // Otherwise, remove all of the CSE's we found from each of the base values.
1093 for (unsigned i = 0; i != NumUses; ++i) {
1094 // Uses outside the loop don't necessarily include the common base, but
1095 // the final IV value coming into those uses does. Instead of trying to
1096 // remove the pieces of the common base, which might not be there,
1097 // subtract off the base to compensate for this.
1098 if (!L->contains(Uses[i].Inst->getParent())) {
1099 Uses[i].Base = SE->getMinusSCEV(Uses[i].Base, Result);
1103 // Split the expression into subexprs.
1104 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
1106 // Remove any common subexpressions.
1107 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
1108 if (SubExpressionUseData.count(SubExprs[j])) {
1109 SubExprs.erase(SubExprs.begin()+j);
1113 // Finally, add the non-shared expressions together.
1114 if (SubExprs.empty())
1115 Uses[i].Base = Zero;
1117 Uses[i].Base = SE->getAddExpr(SubExprs);
1124 /// ValidStride - Check whether the given Scale is valid for all loads and
1125 /// stores in UsersToProcess.
1127 bool LoopStrengthReduce::ValidStride(bool HasBaseReg,
1129 const std::vector<BasedUser>& UsersToProcess) {
1133 for (unsigned i=0, e = UsersToProcess.size(); i!=e; ++i) {
1134 // If this is a load or other access, pass the type of the access in.
1135 const Type *AccessTy = Type::VoidTy;
1136 if (isAddressUse(UsersToProcess[i].Inst,
1137 UsersToProcess[i].OperandValToReplace))
1138 AccessTy = getAccessType(UsersToProcess[i].Inst);
1139 else if (isa<PHINode>(UsersToProcess[i].Inst))
1142 TargetLowering::AddrMode AM;
1143 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
1144 AM.BaseOffs = SC->getValue()->getSExtValue();
1145 AM.HasBaseReg = HasBaseReg || !UsersToProcess[i].Base->isZero();
1148 // If load[imm+r*scale] is illegal, bail out.
1149 if (!TLI->isLegalAddressingMode(AM, AccessTy))
1155 /// RequiresTypeConversion - Returns true if converting Ty1 to Ty2 is not
1157 bool LoopStrengthReduce::RequiresTypeConversion(const Type *Ty1,
1161 Ty1 = SE->getEffectiveSCEVType(Ty1);
1162 Ty2 = SE->getEffectiveSCEVType(Ty2);
1165 if (Ty1->canLosslesslyBitCastTo(Ty2))
1167 if (TLI && TLI->isTruncateFree(Ty1, Ty2))
1172 /// CheckForIVReuse - Returns the multiple if the stride is the multiple
1173 /// of a previous stride and it is a legal value for the target addressing
1174 /// mode scale component and optional base reg. This allows the users of
1175 /// this stride to be rewritten as prev iv * factor. It returns 0 if no
1176 /// reuse is possible. Factors can be negative on same targets, e.g. ARM.
1178 /// If all uses are outside the loop, we don't require that all multiplies
1179 /// be folded into the addressing mode, nor even that the factor be constant;
1180 /// a multiply (executed once) outside the loop is better than another IV
1181 /// within. Well, usually.
1182 SCEVHandle LoopStrengthReduce::CheckForIVReuse(bool HasBaseReg,
1183 bool AllUsesAreAddresses,
1184 bool AllUsesAreOutsideLoop,
1185 const SCEVHandle &Stride,
1186 IVExpr &IV, const Type *Ty,
1187 const std::vector<BasedUser>& UsersToProcess) {
1188 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Stride)) {
1189 int64_t SInt = SC->getValue()->getSExtValue();
1190 for (unsigned NewStride = 0, e = StrideOrder.size(); NewStride != e;
1192 std::map<SCEVHandle, IVsOfOneStride>::iterator SI =
1193 IVsByStride.find(StrideOrder[NewStride]);
1194 if (SI == IVsByStride.end() || !isa<SCEVConstant>(SI->first))
1196 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1197 if (SI->first != Stride &&
1198 (unsigned(abs(SInt)) < SSInt || (SInt % SSInt) != 0))
1200 int64_t Scale = SInt / SSInt;
1201 // Check that this stride is valid for all the types used for loads and
1202 // stores; if it can be used for some and not others, we might as well use
1203 // the original stride everywhere, since we have to create the IV for it
1204 // anyway. If the scale is 1, then we don't need to worry about folding
1207 (AllUsesAreAddresses &&
1208 ValidStride(HasBaseReg, Scale, UsersToProcess)))
1209 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1210 IE = SI->second.IVs.end(); II != IE; ++II)
1211 // FIXME: Only handle base == 0 for now.
1212 // Only reuse previous IV if it would not require a type conversion.
1213 if (II->Base->isZero() &&
1214 !RequiresTypeConversion(II->Base->getType(), Ty)) {
1216 return SE->getIntegerSCEV(Scale, Stride->getType());
1219 } else if (AllUsesAreOutsideLoop) {
1220 // Accept nonconstant strides here; it is really really right to substitute
1221 // an existing IV if we can.
1222 for (unsigned NewStride = 0, e = StrideOrder.size(); NewStride != e;
1224 std::map<SCEVHandle, IVsOfOneStride>::iterator SI =
1225 IVsByStride.find(StrideOrder[NewStride]);
1226 if (SI == IVsByStride.end() || !isa<SCEVConstant>(SI->first))
1228 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1229 if (SI->first != Stride && SSInt != 1)
1231 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1232 IE = SI->second.IVs.end(); II != IE; ++II)
1233 // Accept nonzero base here.
1234 // Only reuse previous IV if it would not require a type conversion.
1235 if (!RequiresTypeConversion(II->Base->getType(), Ty)) {
1240 // Special case, old IV is -1*x and this one is x. Can treat this one as
1242 for (unsigned NewStride = 0, e = StrideOrder.size(); NewStride != e;
1244 std::map<SCEVHandle, IVsOfOneStride>::iterator SI =
1245 IVsByStride.find(StrideOrder[NewStride]);
1246 if (SI == IVsByStride.end())
1248 if (const SCEVMulExpr *ME = dyn_cast<SCEVMulExpr>(SI->first))
1249 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(ME->getOperand(0)))
1250 if (Stride == ME->getOperand(1) &&
1251 SC->getValue()->getSExtValue() == -1LL)
1252 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1253 IE = SI->second.IVs.end(); II != IE; ++II)
1254 // Accept nonzero base here.
1255 // Only reuse previous IV if it would not require type conversion.
1256 if (!RequiresTypeConversion(II->Base->getType(), Ty)) {
1258 return SE->getIntegerSCEV(-1LL, Stride->getType());
1262 return SE->getIntegerSCEV(0, Stride->getType());
1265 /// PartitionByIsUseOfPostIncrementedValue - Simple boolean predicate that
1266 /// returns true if Val's isUseOfPostIncrementedValue is true.
1267 static bool PartitionByIsUseOfPostIncrementedValue(const BasedUser &Val) {
1268 return Val.isUseOfPostIncrementedValue;
1271 /// isNonConstantNegative - Return true if the specified scev is negated, but
1273 static bool isNonConstantNegative(const SCEVHandle &Expr) {
1274 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Expr);
1275 if (!Mul) return false;
1277 // If there is a constant factor, it will be first.
1278 const SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
1279 if (!SC) return false;
1281 // Return true if the value is negative, this matches things like (-42 * V).
1282 return SC->getValue()->getValue().isNegative();
1285 // CollectIVUsers - Transform our list of users and offsets to a bit more
1286 // complex table. In this new vector, each 'BasedUser' contains 'Base', the base
1287 // of the strided accesses, as well as the old information from Uses. We
1288 // progressively move information from the Base field to the Imm field, until
1289 // we eventually have the full access expression to rewrite the use.
1290 SCEVHandle LoopStrengthReduce::CollectIVUsers(const SCEVHandle &Stride,
1291 IVUsersOfOneStride &Uses,
1293 bool &AllUsesAreAddresses,
1294 bool &AllUsesAreOutsideLoop,
1295 std::vector<BasedUser> &UsersToProcess) {
1296 UsersToProcess.reserve(Uses.Users.size());
1297 for (unsigned i = 0, e = Uses.Users.size(); i != e; ++i) {
1298 UsersToProcess.push_back(BasedUser(Uses.Users[i], SE));
1300 // Move any loop variant operands from the offset field to the immediate
1301 // field of the use, so that we don't try to use something before it is
1303 MoveLoopVariantsToImmediateField(UsersToProcess.back().Base,
1304 UsersToProcess.back().Imm, L, SE);
1305 assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
1306 "Base value is not loop invariant!");
1309 // We now have a whole bunch of uses of like-strided induction variables, but
1310 // they might all have different bases. We want to emit one PHI node for this
1311 // stride which we fold as many common expressions (between the IVs) into as
1312 // possible. Start by identifying the common expressions in the base values
1313 // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
1314 // "A+B"), emit it to the preheader, then remove the expression from the
1315 // UsersToProcess base values.
1316 SCEVHandle CommonExprs =
1317 RemoveCommonExpressionsFromUseBases(UsersToProcess, SE, L, TLI);
1319 // Next, figure out what we can represent in the immediate fields of
1320 // instructions. If we can represent anything there, move it to the imm
1321 // fields of the BasedUsers. We do this so that it increases the commonality
1322 // of the remaining uses.
1323 unsigned NumPHI = 0;
1324 bool HasAddress = false;
1325 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1326 // If the user is not in the current loop, this means it is using the exit
1327 // value of the IV. Do not put anything in the base, make sure it's all in
1328 // the immediate field to allow as much factoring as possible.
1329 if (!L->contains(UsersToProcess[i].Inst->getParent())) {
1330 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm,
1331 UsersToProcess[i].Base);
1332 UsersToProcess[i].Base =
1333 SE->getIntegerSCEV(0, UsersToProcess[i].Base->getType());
1335 // Not all uses are outside the loop.
1336 AllUsesAreOutsideLoop = false;
1338 // Addressing modes can be folded into loads and stores. Be careful that
1339 // the store is through the expression, not of the expression though.
1341 bool isAddress = isAddressUse(UsersToProcess[i].Inst,
1342 UsersToProcess[i].OperandValToReplace);
1343 if (isa<PHINode>(UsersToProcess[i].Inst)) {
1351 // If this use isn't an address, then not all uses are addresses.
1352 if (!isAddress && !isPHI)
1353 AllUsesAreAddresses = false;
1355 MoveImmediateValues(TLI, UsersToProcess[i].Inst, UsersToProcess[i].Base,
1356 UsersToProcess[i].Imm, isAddress, L, SE);
1360 // If one of the use is a PHI node and all other uses are addresses, still
1361 // allow iv reuse. Essentially we are trading one constant multiplication
1362 // for one fewer iv.
1364 AllUsesAreAddresses = false;
1366 // There are no in-loop address uses.
1367 if (AllUsesAreAddresses && (!HasAddress && !AllUsesAreOutsideLoop))
1368 AllUsesAreAddresses = false;
1373 /// ShouldUseFullStrengthReductionMode - Test whether full strength-reduction
1374 /// is valid and profitable for the given set of users of a stride. In
1375 /// full strength-reduction mode, all addresses at the current stride are
1376 /// strength-reduced all the way down to pointer arithmetic.
1378 bool LoopStrengthReduce::ShouldUseFullStrengthReductionMode(
1379 const std::vector<BasedUser> &UsersToProcess,
1381 bool AllUsesAreAddresses,
1382 SCEVHandle Stride) {
1383 if (!EnableFullLSRMode)
1386 // The heuristics below aim to avoid increasing register pressure, but
1387 // fully strength-reducing all the addresses increases the number of
1388 // add instructions, so don't do this when optimizing for size.
1389 // TODO: If the loop is large, the savings due to simpler addresses
1390 // may oughtweight the costs of the extra increment instructions.
1391 if (L->getHeader()->getParent()->hasFnAttr(Attribute::OptimizeForSize))
1394 // TODO: For now, don't do full strength reduction if there could
1395 // potentially be greater-stride multiples of the current stride
1396 // which could reuse the current stride IV.
1397 if (StrideOrder.back() != Stride)
1400 // Iterate through the uses to find conditions that automatically rule out
1402 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ) {
1403 SCEV *Base = UsersToProcess[i].Base;
1404 SCEV *Imm = UsersToProcess[i].Imm;
1405 // If any users have a loop-variant component, they can't be fully
1406 // strength-reduced.
1407 if (Imm && !Imm->isLoopInvariant(L))
1409 // If there are to users with the same base and the difference between
1410 // the two Imm values can't be folded into the address, full
1411 // strength reduction would increase register pressure.
1413 SCEV *CurImm = UsersToProcess[i].Imm;
1414 if ((CurImm || Imm) && CurImm != Imm) {
1415 if (!CurImm) CurImm = SE->getIntegerSCEV(0, Stride->getType());
1416 if (!Imm) Imm = SE->getIntegerSCEV(0, Stride->getType());
1417 const Instruction *Inst = UsersToProcess[i].Inst;
1418 const Type *UseTy = getAccessType(Inst);
1419 SCEVHandle Diff = SE->getMinusSCEV(UsersToProcess[i].Imm, Imm);
1420 if (!Diff->isZero() &&
1421 (!AllUsesAreAddresses ||
1422 !fitsInAddressMode(Diff, UseTy, TLI, /*HasBaseReg=*/true)))
1425 } while (++i != e && Base == UsersToProcess[i].Base);
1428 // If there's exactly one user in this stride, fully strength-reducing it
1429 // won't increase register pressure. If it's starting from a non-zero base,
1430 // it'll be simpler this way.
1431 if (UsersToProcess.size() == 1 && !UsersToProcess[0].Base->isZero())
1434 // Otherwise, if there are any users in this stride that don't require
1435 // a register for their base, full strength-reduction will increase
1436 // register pressure.
1437 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1438 if (UsersToProcess[i].Base->isZero())
1441 // Otherwise, go for it.
1445 /// InsertAffinePhi Create and insert a PHI node for an induction variable
1446 /// with the specified start and step values in the specified loop.
1448 /// If NegateStride is true, the stride should be negated by using a
1449 /// subtract instead of an add.
1451 /// Return the created phi node.
1453 static PHINode *InsertAffinePhi(SCEVHandle Start, SCEVHandle Step,
1455 SCEVExpander &Rewriter) {
1456 assert(Start->isLoopInvariant(L) && "New PHI start is not loop invariant!");
1457 assert(Step->isLoopInvariant(L) && "New PHI stride is not loop invariant!");
1459 BasicBlock *Header = L->getHeader();
1460 BasicBlock *Preheader = L->getLoopPreheader();
1461 BasicBlock *LatchBlock = L->getLoopLatch();
1462 const Type *Ty = Start->getType();
1463 Ty = Rewriter.SE.getEffectiveSCEVType(Ty);
1465 PHINode *PN = PHINode::Create(Ty, "lsr.iv", Header->begin());
1466 PN->addIncoming(Rewriter.expandCodeFor(Start, Ty, Preheader->getTerminator()),
1469 // If the stride is negative, insert a sub instead of an add for the
1471 bool isNegative = isNonConstantNegative(Step);
1472 SCEVHandle IncAmount = Step;
1474 IncAmount = Rewriter.SE.getNegativeSCEV(Step);
1476 // Insert an add instruction right before the terminator corresponding
1477 // to the back-edge.
1478 Value *StepV = Rewriter.expandCodeFor(IncAmount, Ty,
1479 Preheader->getTerminator());
1482 IncV = BinaryOperator::CreateSub(PN, StepV, "lsr.iv.next",
1483 LatchBlock->getTerminator());
1485 IncV = BinaryOperator::CreateAdd(PN, StepV, "lsr.iv.next",
1486 LatchBlock->getTerminator());
1488 if (!isa<ConstantInt>(StepV)) ++NumVariable;
1490 PN->addIncoming(IncV, LatchBlock);
1496 static void SortUsersToProcess(std::vector<BasedUser> &UsersToProcess) {
1497 // We want to emit code for users inside the loop first. To do this, we
1498 // rearrange BasedUser so that the entries at the end have
1499 // isUseOfPostIncrementedValue = false, because we pop off the end of the
1500 // vector (so we handle them first).
1501 std::partition(UsersToProcess.begin(), UsersToProcess.end(),
1502 PartitionByIsUseOfPostIncrementedValue);
1504 // Sort this by base, so that things with the same base are handled
1505 // together. By partitioning first and stable-sorting later, we are
1506 // guaranteed that within each base we will pop off users from within the
1507 // loop before users outside of the loop with a particular base.
1509 // We would like to use stable_sort here, but we can't. The problem is that
1510 // SCEVHandle's don't have a deterministic ordering w.r.t to each other, so
1511 // we don't have anything to do a '<' comparison on. Because we think the
1512 // number of uses is small, do a horrible bubble sort which just relies on
1514 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1515 // Get a base value.
1516 SCEVHandle Base = UsersToProcess[i].Base;
1518 // Compact everything with this base to be consecutive with this one.
1519 for (unsigned j = i+1; j != e; ++j) {
1520 if (UsersToProcess[j].Base == Base) {
1521 std::swap(UsersToProcess[i+1], UsersToProcess[j]);
1528 /// PrepareToStrengthReduceFully - Prepare to fully strength-reduce
1529 /// UsersToProcess, meaning lowering addresses all the way down to direct
1530 /// pointer arithmetic.
1533 LoopStrengthReduce::PrepareToStrengthReduceFully(
1534 std::vector<BasedUser> &UsersToProcess,
1536 SCEVHandle CommonExprs,
1538 SCEVExpander &PreheaderRewriter) {
1539 DOUT << " Fully reducing all users\n";
1541 // Rewrite the UsersToProcess records, creating a separate PHI for each
1542 // unique Base value.
1543 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ) {
1544 // TODO: The uses are grouped by base, but not sorted. We arbitrarily
1545 // pick the first Imm value here to start with, and adjust it for the
1547 SCEVHandle Imm = UsersToProcess[i].Imm;
1548 SCEVHandle Base = UsersToProcess[i].Base;
1549 SCEVHandle Start = SE->getAddExpr(CommonExprs, Base, Imm);
1550 PHINode *Phi = InsertAffinePhi(Start, Stride, L,
1552 // Loop over all the users with the same base.
1554 UsersToProcess[i].Base = SE->getIntegerSCEV(0, Stride->getType());
1555 UsersToProcess[i].Imm = SE->getMinusSCEV(UsersToProcess[i].Imm, Imm);
1556 UsersToProcess[i].Phi = Phi;
1557 assert(UsersToProcess[i].Imm->isLoopInvariant(L) &&
1558 "ShouldUseFullStrengthReductionMode should reject this!");
1559 } while (++i != e && Base == UsersToProcess[i].Base);
1563 /// PrepareToStrengthReduceWithNewPhi - Insert a new induction variable for the
1564 /// given users to share.
1567 LoopStrengthReduce::PrepareToStrengthReduceWithNewPhi(
1568 std::vector<BasedUser> &UsersToProcess,
1570 SCEVHandle CommonExprs,
1573 SCEVExpander &PreheaderRewriter) {
1574 DOUT << " Inserting new PHI:\n";
1576 PHINode *Phi = InsertAffinePhi(SE->getUnknown(CommonBaseV),
1580 // Remember this in case a later stride is multiple of this.
1581 IVsByStride[Stride].addIV(Stride, CommonExprs, Phi);
1583 // All the users will share this new IV.
1584 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1585 UsersToProcess[i].Phi = Phi;
1588 DEBUG(WriteAsOperand(*DOUT, Phi, /*PrintType=*/false));
1592 /// PrepareToStrengthReduceWithNewPhi - Prepare for the given users to reuse
1593 /// an induction variable with a stride that is a factor of the current
1594 /// induction variable.
1597 LoopStrengthReduce::PrepareToStrengthReduceFromSmallerStride(
1598 std::vector<BasedUser> &UsersToProcess,
1600 const IVExpr &ReuseIV,
1601 Instruction *PreInsertPt) {
1602 DOUT << " Rewriting in terms of existing IV of STRIDE " << *ReuseIV.Stride
1603 << " and BASE " << *ReuseIV.Base << "\n";
1605 // All the users will share the reused IV.
1606 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1607 UsersToProcess[i].Phi = ReuseIV.PHI;
1609 Constant *C = dyn_cast<Constant>(CommonBaseV);
1611 (!C->isNullValue() &&
1612 !fitsInAddressMode(SE->getUnknown(CommonBaseV), CommonBaseV->getType(),
1614 // We want the common base emitted into the preheader! This is just
1615 // using cast as a copy so BitCast (no-op cast) is appropriate
1616 CommonBaseV = new BitCastInst(CommonBaseV, CommonBaseV->getType(),
1617 "commonbase", PreInsertPt);
1620 static bool IsImmFoldedIntoAddrMode(GlobalValue *GV, int64_t Offset,
1621 const Type *AccessTy,
1622 std::vector<BasedUser> &UsersToProcess,
1623 const TargetLowering *TLI) {
1624 SmallVector<Instruction*, 16> AddrModeInsts;
1625 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1626 if (UsersToProcess[i].isUseOfPostIncrementedValue)
1628 ExtAddrMode AddrMode =
1629 AddressingModeMatcher::Match(UsersToProcess[i].OperandValToReplace,
1630 AccessTy, UsersToProcess[i].Inst,
1631 AddrModeInsts, *TLI);
1632 if (GV && GV != AddrMode.BaseGV)
1634 if (Offset && !AddrMode.BaseOffs)
1635 // FIXME: How to accurate check it's immediate offset is folded.
1637 AddrModeInsts.clear();
1642 /// StrengthReduceStridedIVUsers - Strength reduce all of the users of a single
1643 /// stride of IV. All of the users may have different starting values, and this
1644 /// may not be the only stride.
1645 void LoopStrengthReduce::StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
1646 IVUsersOfOneStride &Uses,
1648 // If all the users are moved to another stride, then there is nothing to do.
1649 if (Uses.Users.empty())
1652 // Keep track if every use in UsersToProcess is an address. If they all are,
1653 // we may be able to rewrite the entire collection of them in terms of a
1654 // smaller-stride IV.
1655 bool AllUsesAreAddresses = true;
1657 // Keep track if every use of a single stride is outside the loop. If so,
1658 // we want to be more aggressive about reusing a smaller-stride IV; a
1659 // multiply outside the loop is better than another IV inside. Well, usually.
1660 bool AllUsesAreOutsideLoop = true;
1662 // Transform our list of users and offsets to a bit more complex table. In
1663 // this new vector, each 'BasedUser' contains 'Base' the base of the
1664 // strided accessas well as the old information from Uses. We progressively
1665 // move information from the Base field to the Imm field, until we eventually
1666 // have the full access expression to rewrite the use.
1667 std::vector<BasedUser> UsersToProcess;
1668 SCEVHandle CommonExprs = CollectIVUsers(Stride, Uses, L, AllUsesAreAddresses,
1669 AllUsesAreOutsideLoop,
1672 // Sort the UsersToProcess array so that users with common bases are
1673 // next to each other.
1674 SortUsersToProcess(UsersToProcess);
1676 // If we managed to find some expressions in common, we'll need to carry
1677 // their value in a register and add it in for each use. This will take up
1678 // a register operand, which potentially restricts what stride values are
1680 bool HaveCommonExprs = !CommonExprs->isZero();
1681 const Type *ReplacedTy = CommonExprs->getType();
1683 // If all uses are addresses, consider sinking the immediate part of the
1684 // common expression back into uses if they can fit in the immediate fields.
1685 if (TLI && HaveCommonExprs && AllUsesAreAddresses) {
1686 SCEVHandle NewCommon = CommonExprs;
1687 SCEVHandle Imm = SE->getIntegerSCEV(0, ReplacedTy);
1688 MoveImmediateValues(TLI, Type::VoidTy, NewCommon, Imm, true, L, SE);
1689 if (!Imm->isZero()) {
1692 // If the immediate part of the common expression is a GV, check if it's
1693 // possible to fold it into the target addressing mode.
1694 GlobalValue *GV = 0;
1695 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(Imm))
1696 GV = dyn_cast<GlobalValue>(SU->getValue());
1698 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Imm))
1699 Offset = SC->getValue()->getSExtValue();
1701 // Pass VoidTy as the AccessTy to be conservative, because
1702 // there could be multiple access types among all the uses.
1703 DoSink = IsImmFoldedIntoAddrMode(GV, Offset, Type::VoidTy,
1704 UsersToProcess, TLI);
1707 DOUT << " Sinking " << *Imm << " back down into uses\n";
1708 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1709 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm, Imm);
1710 CommonExprs = NewCommon;
1711 HaveCommonExprs = !CommonExprs->isZero();
1717 // Now that we know what we need to do, insert the PHI node itself.
1719 DOUT << "LSR: Examining IVs of TYPE " << *ReplacedTy << " of STRIDE "
1721 << " Common base: " << *CommonExprs << "\n";
1723 SCEVExpander Rewriter(*SE, *LI);
1724 SCEVExpander PreheaderRewriter(*SE, *LI);
1726 BasicBlock *Preheader = L->getLoopPreheader();
1727 Instruction *PreInsertPt = Preheader->getTerminator();
1728 BasicBlock *LatchBlock = L->getLoopLatch();
1730 Value *CommonBaseV = Constant::getNullValue(ReplacedTy);
1732 SCEVHandle RewriteFactor = SE->getIntegerSCEV(0, ReplacedTy);
1733 IVExpr ReuseIV(SE->getIntegerSCEV(0, Type::Int32Ty),
1734 SE->getIntegerSCEV(0, Type::Int32Ty),
1737 /// Choose a strength-reduction strategy and prepare for it by creating
1738 /// the necessary PHIs and adjusting the bookkeeping.
1739 if (ShouldUseFullStrengthReductionMode(UsersToProcess, L,
1740 AllUsesAreAddresses, Stride)) {
1741 PrepareToStrengthReduceFully(UsersToProcess, Stride, CommonExprs, L,
1744 // Emit the initial base value into the loop preheader.
1745 CommonBaseV = PreheaderRewriter.expandCodeFor(CommonExprs, ReplacedTy,
1748 // If all uses are addresses, check if it is possible to reuse an IV with a
1749 // stride that is a factor of this stride. And that the multiple is a number
1750 // that can be encoded in the scale field of the target addressing mode. And
1751 // that we will have a valid instruction after this substition, including
1752 // the immediate field, if any.
1753 RewriteFactor = CheckForIVReuse(HaveCommonExprs, AllUsesAreAddresses,
1754 AllUsesAreOutsideLoop,
1755 Stride, ReuseIV, ReplacedTy,
1757 if (isa<SCEVConstant>(RewriteFactor) &&
1758 cast<SCEVConstant>(RewriteFactor)->isZero())
1759 PrepareToStrengthReduceWithNewPhi(UsersToProcess, Stride, CommonExprs,
1760 CommonBaseV, L, PreheaderRewriter);
1762 PrepareToStrengthReduceFromSmallerStride(UsersToProcess, CommonBaseV,
1763 ReuseIV, PreInsertPt);
1766 // Process all the users now, replacing their strided uses with
1767 // strength-reduced forms. This outer loop handles all bases, the inner
1768 // loop handles all users of a particular base.
1769 while (!UsersToProcess.empty()) {
1770 SCEVHandle Base = UsersToProcess.back().Base;
1771 Instruction *Inst = UsersToProcess.back().Inst;
1773 // Emit the code for Base into the preheader.
1775 if (!Base->isZero()) {
1776 BaseV = PreheaderRewriter.expandCodeFor(Base, Base->getType(),
1779 DOUT << " INSERTING code for BASE = " << *Base << ":";
1780 if (BaseV->hasName())
1781 DOUT << " Result value name = %" << BaseV->getNameStr();
1784 // If BaseV is a non-zero constant, make sure that it gets inserted into
1785 // the preheader, instead of being forward substituted into the uses. We
1786 // do this by forcing a BitCast (noop cast) to be inserted into the
1787 // preheader in this case.
1788 if (!fitsInAddressMode(Base, getAccessType(Inst), TLI, false)) {
1789 // We want this constant emitted into the preheader! This is just
1790 // using cast as a copy so BitCast (no-op cast) is appropriate
1791 BaseV = new BitCastInst(BaseV, BaseV->getType(), "preheaderinsert",
1796 // Emit the code to add the immediate offset to the Phi value, just before
1797 // the instructions that we identified as using this stride and base.
1799 // FIXME: Use emitted users to emit other users.
1800 BasedUser &User = UsersToProcess.back();
1802 DOUT << " Examining use ";
1803 DEBUG(WriteAsOperand(*DOUT, UsersToProcess.back().OperandValToReplace,
1804 /*PrintType=*/false));
1805 DOUT << " in Inst: " << *(User.Inst);
1807 // If this instruction wants to use the post-incremented value, move it
1808 // after the post-inc and use its value instead of the PHI.
1809 Value *RewriteOp = User.Phi;
1810 if (User.isUseOfPostIncrementedValue) {
1811 RewriteOp = User.Phi->getIncomingValueForBlock(LatchBlock);
1813 // If this user is in the loop, make sure it is the last thing in the
1814 // loop to ensure it is dominated by the increment.
1815 if (L->contains(User.Inst->getParent()))
1816 User.Inst->moveBefore(LatchBlock->getTerminator());
1819 SCEVHandle RewriteExpr = SE->getUnknown(RewriteOp);
1821 if (SE->getTypeSizeInBits(RewriteOp->getType()) !=
1822 SE->getTypeSizeInBits(ReplacedTy)) {
1823 assert(SE->getTypeSizeInBits(RewriteOp->getType()) >
1824 SE->getTypeSizeInBits(ReplacedTy) &&
1825 "Unexpected widening cast!");
1826 RewriteExpr = SE->getTruncateExpr(RewriteExpr, ReplacedTy);
1829 // If we had to insert new instructions for RewriteOp, we have to
1830 // consider that they may not have been able to end up immediately
1831 // next to RewriteOp, because non-PHI instructions may never precede
1832 // PHI instructions in a block. In this case, remember where the last
1833 // instruction was inserted so that if we're replacing a different
1834 // PHI node, we can use the later point to expand the final
1836 Instruction *NewBasePt = dyn_cast<Instruction>(RewriteOp);
1837 if (RewriteOp == User.Phi) NewBasePt = 0;
1839 // Clear the SCEVExpander's expression map so that we are guaranteed
1840 // to have the code emitted where we expect it.
1843 // If we are reusing the iv, then it must be multiplied by a constant
1844 // factor to take advantage of the addressing mode scale component.
1845 if (!RewriteFactor->isZero()) {
1846 // If we're reusing an IV with a nonzero base (currently this happens
1847 // only when all reuses are outside the loop) subtract that base here.
1848 // The base has been used to initialize the PHI node but we don't want
1850 if (!ReuseIV.Base->isZero()) {
1851 SCEVHandle typedBase = ReuseIV.Base;
1852 if (SE->getTypeSizeInBits(RewriteExpr->getType()) !=
1853 SE->getTypeSizeInBits(ReuseIV.Base->getType())) {
1854 // It's possible the original IV is a larger type than the new IV,
1855 // in which case we have to truncate the Base. We checked in
1856 // RequiresTypeConversion that this is valid.
1857 assert(SE->getTypeSizeInBits(RewriteExpr->getType()) <
1858 SE->getTypeSizeInBits(ReuseIV.Base->getType()) &&
1859 "Unexpected lengthening conversion!");
1860 typedBase = SE->getTruncateExpr(ReuseIV.Base,
1861 RewriteExpr->getType());
1863 RewriteExpr = SE->getMinusSCEV(RewriteExpr, typedBase);
1866 // Multiply old variable, with base removed, by new scale factor.
1867 RewriteExpr = SE->getMulExpr(RewriteFactor,
1870 // The common base is emitted in the loop preheader. But since we
1871 // are reusing an IV, it has not been used to initialize the PHI node.
1872 // Add it to the expression used to rewrite the uses.
1873 // When this use is outside the loop, we earlier subtracted the
1874 // common base, and are adding it back here. Use the same expression
1875 // as before, rather than CommonBaseV, so DAGCombiner will zap it.
1876 if (!CommonExprs->isZero()) {
1877 if (L->contains(User.Inst->getParent()))
1878 RewriteExpr = SE->getAddExpr(RewriteExpr,
1879 SE->getUnknown(CommonBaseV));
1881 RewriteExpr = SE->getAddExpr(RewriteExpr, CommonExprs);
1885 // Now that we know what we need to do, insert code before User for the
1886 // immediate and any loop-variant expressions.
1888 // Add BaseV to the PHI value if needed.
1889 RewriteExpr = SE->getAddExpr(RewriteExpr, SE->getUnknown(BaseV));
1891 User.RewriteInstructionToUseNewBase(RewriteExpr, NewBasePt,
1895 // Mark old value we replaced as possibly dead, so that it is eliminated
1896 // if we just replaced the last use of that value.
1897 DeadInsts.push_back(cast<Instruction>(User.OperandValToReplace));
1899 UsersToProcess.pop_back();
1902 // If there are any more users to process with the same base, process them
1903 // now. We sorted by base above, so we just have to check the last elt.
1904 } while (!UsersToProcess.empty() && UsersToProcess.back().Base == Base);
1905 // TODO: Next, find out which base index is the most common, pull it out.
1908 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
1909 // different starting values, into different PHIs.
1912 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1913 /// set the IV user and stride information and return true, otherwise return
1915 bool LoopStrengthReduce::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse,
1916 const SCEVHandle *&CondStride) {
1917 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e && !CondUse;
1919 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1920 IVUsesByStride.find(StrideOrder[Stride]);
1921 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1923 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
1924 E = SI->second.Users.end(); UI != E; ++UI)
1925 if (UI->User == Cond) {
1926 // NOTE: we could handle setcc instructions with multiple uses here, but
1927 // InstCombine does it as well for simple uses, it's not clear that it
1928 // occurs enough in real life to handle.
1930 CondStride = &SI->first;
1938 // Constant strides come first which in turns are sorted by their absolute
1939 // values. If absolute values are the same, then positive strides comes first.
1941 // 4, -1, X, 1, 2 ==> 1, -1, 2, 4, X
1942 struct StrideCompare {
1943 const ScalarEvolution *SE;
1944 explicit StrideCompare(const ScalarEvolution *se) : SE(se) {}
1946 bool operator()(const SCEVHandle &LHS, const SCEVHandle &RHS) {
1947 const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS);
1948 const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS);
1950 int64_t LV = LHSC->getValue()->getSExtValue();
1951 int64_t RV = RHSC->getValue()->getSExtValue();
1952 uint64_t ALV = (LV < 0) ? -LV : LV;
1953 uint64_t ARV = (RV < 0) ? -RV : RV;
1961 // If it's the same value but different type, sort by bit width so
1962 // that we emit larger induction variables before smaller
1963 // ones, letting the smaller be re-written in terms of larger ones.
1964 return SE->getTypeSizeInBits(RHS->getType()) <
1965 SE->getTypeSizeInBits(LHS->getType());
1967 return LHSC && !RHSC;
1972 /// ChangeCompareStride - If a loop termination compare instruction is the
1973 /// only use of its stride, and the compaison is against a constant value,
1974 /// try eliminate the stride by moving the compare instruction to another
1975 /// stride and change its constant operand accordingly. e.g.
1981 /// if (v2 < 10) goto loop
1986 /// if (v1 < 30) goto loop
1987 ICmpInst *LoopStrengthReduce::ChangeCompareStride(Loop *L, ICmpInst *Cond,
1988 IVStrideUse* &CondUse,
1989 const SCEVHandle* &CondStride) {
1990 if (StrideOrder.size() < 2 ||
1991 IVUsesByStride[*CondStride].Users.size() != 1)
1993 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*CondStride);
1994 if (!SC) return Cond;
1996 ICmpInst::Predicate Predicate = Cond->getPredicate();
1997 int64_t CmpSSInt = SC->getValue()->getSExtValue();
1998 unsigned BitWidth = SE->getTypeSizeInBits((*CondStride)->getType());
1999 uint64_t SignBit = 1ULL << (BitWidth-1);
2000 const Type *CmpTy = Cond->getOperand(0)->getType();
2001 const Type *NewCmpTy = NULL;
2002 unsigned TyBits = SE->getTypeSizeInBits(CmpTy);
2003 unsigned NewTyBits = 0;
2004 SCEVHandle *NewStride = NULL;
2005 Value *NewCmpLHS = NULL;
2006 Value *NewCmpRHS = NULL;
2008 SCEVHandle NewOffset = SE->getIntegerSCEV(0, CmpTy);
2010 if (ConstantInt *C = dyn_cast<ConstantInt>(Cond->getOperand(1))) {
2011 int64_t CmpVal = C->getValue().getSExtValue();
2013 // Check stride constant and the comparision constant signs to detect
2015 if ((CmpVal & SignBit) != (CmpSSInt & SignBit))
2018 // Look for a suitable stride / iv as replacement.
2019 for (unsigned i = 0, e = StrideOrder.size(); i != e; ++i) {
2020 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
2021 IVUsesByStride.find(StrideOrder[i]);
2022 if (!isa<SCEVConstant>(SI->first))
2024 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
2025 if (SSInt == CmpSSInt ||
2026 abs(SSInt) < abs(CmpSSInt) ||
2027 (SSInt % CmpSSInt) != 0)
2030 Scale = SSInt / CmpSSInt;
2031 int64_t NewCmpVal = CmpVal * Scale;
2032 APInt Mul = APInt(BitWidth, NewCmpVal);
2033 // Check for overflow.
2034 if (Mul.getSExtValue() != NewCmpVal)
2037 // Watch out for overflow.
2038 if (ICmpInst::isSignedPredicate(Predicate) &&
2039 (CmpVal & SignBit) != (NewCmpVal & SignBit))
2042 if (NewCmpVal == CmpVal)
2044 // Pick the best iv to use trying to avoid a cast.
2046 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
2047 E = SI->second.Users.end(); UI != E; ++UI) {
2048 NewCmpLHS = UI->OperandValToReplace;
2049 if (NewCmpLHS->getType() == CmpTy)
2055 NewCmpTy = NewCmpLHS->getType();
2056 NewTyBits = SE->getTypeSizeInBits(NewCmpTy);
2057 const Type *NewCmpIntTy = IntegerType::get(NewTyBits);
2058 if (RequiresTypeConversion(NewCmpTy, CmpTy)) {
2059 // Check if it is possible to rewrite it using
2060 // an iv / stride of a smaller integer type.
2061 unsigned Bits = NewTyBits;
2062 if (ICmpInst::isSignedPredicate(Predicate))
2064 uint64_t Mask = (1ULL << Bits) - 1;
2065 if (((uint64_t)NewCmpVal & Mask) != (uint64_t)NewCmpVal)
2069 // Don't rewrite if use offset is non-constant and the new type is
2070 // of a different type.
2071 // FIXME: too conservative?
2072 if (NewTyBits != TyBits && !isa<SCEVConstant>(CondUse->Offset))
2075 bool AllUsesAreAddresses = true;
2076 bool AllUsesAreOutsideLoop = true;
2077 std::vector<BasedUser> UsersToProcess;
2078 SCEVHandle CommonExprs = CollectIVUsers(SI->first, SI->second, L,
2079 AllUsesAreAddresses,
2080 AllUsesAreOutsideLoop,
2082 // Avoid rewriting the compare instruction with an iv of new stride
2083 // if it's likely the new stride uses will be rewritten using the
2084 // stride of the compare instruction.
2085 if (AllUsesAreAddresses &&
2086 ValidStride(!CommonExprs->isZero(), Scale, UsersToProcess))
2089 // If scale is negative, use swapped predicate unless it's testing
2091 if (Scale < 0 && !Cond->isEquality())
2092 Predicate = ICmpInst::getSwappedPredicate(Predicate);
2094 NewStride = &StrideOrder[i];
2095 if (!isa<PointerType>(NewCmpTy))
2096 NewCmpRHS = ConstantInt::get(NewCmpTy, NewCmpVal);
2098 ConstantInt *CI = ConstantInt::get(NewCmpIntTy, NewCmpVal);
2099 NewCmpRHS = ConstantExpr::getIntToPtr(CI, NewCmpTy);
2101 NewOffset = TyBits == NewTyBits
2102 ? SE->getMulExpr(CondUse->Offset,
2103 SE->getConstant(ConstantInt::get(CmpTy, Scale)))
2104 : SE->getConstant(ConstantInt::get(NewCmpIntTy,
2105 cast<SCEVConstant>(CondUse->Offset)->getValue()->getSExtValue()*Scale));
2110 // Forgo this transformation if it the increment happens to be
2111 // unfortunately positioned after the condition, and the condition
2112 // has multiple uses which prevent it from being moved immediately
2113 // before the branch. See
2114 // test/Transforms/LoopStrengthReduce/change-compare-stride-trickiness-*.ll
2115 // for an example of this situation.
2116 if (!Cond->hasOneUse()) {
2117 for (BasicBlock::iterator I = Cond, E = Cond->getParent()->end();
2124 // Create a new compare instruction using new stride / iv.
2125 ICmpInst *OldCond = Cond;
2126 // Insert new compare instruction.
2127 Cond = new ICmpInst(Predicate, NewCmpLHS, NewCmpRHS,
2128 L->getHeader()->getName() + ".termcond",
2131 // Remove the old compare instruction. The old indvar is probably dead too.
2132 DeadInsts.push_back(cast<Instruction>(CondUse->OperandValToReplace));
2133 SE->deleteValueFromRecords(OldCond);
2134 OldCond->replaceAllUsesWith(Cond);
2135 OldCond->eraseFromParent();
2137 IVUsesByStride[*CondStride].Users.pop_back();
2138 IVUsesByStride[*NewStride].addUser(NewOffset, Cond, NewCmpLHS);
2139 CondUse = &IVUsesByStride[*NewStride].Users.back();
2140 CondStride = NewStride;
2148 /// OptimizeSMax - Rewrite the loop's terminating condition if it uses
2149 /// an smax computation.
2151 /// This is a narrow solution to a specific, but acute, problem. For loops
2157 /// } while (++i < n);
2159 /// where the comparison is signed, the trip count isn't just 'n', because
2160 /// 'n' could be negative. And unfortunately this can come up even for loops
2161 /// where the user didn't use a C do-while loop. For example, seemingly
2162 /// well-behaved top-test loops will commonly be lowered like this:
2168 /// } while (++i < n);
2171 /// and then it's possible for subsequent optimization to obscure the if
2172 /// test in such a way that indvars can't find it.
2174 /// When indvars can't find the if test in loops like this, it creates a
2175 /// signed-max expression, which allows it to give the loop a canonical
2176 /// induction variable:
2179 /// smax = n < 1 ? 1 : n;
2182 /// } while (++i != smax);
2184 /// Canonical induction variables are necessary because the loop passes
2185 /// are designed around them. The most obvious example of this is the
2186 /// LoopInfo analysis, which doesn't remember trip count values. It
2187 /// expects to be able to rediscover the trip count each time it is
2188 /// needed, and it does this using a simple analyis that only succeeds if
2189 /// the loop has a canonical induction variable.
2191 /// However, when it comes time to generate code, the maximum operation
2192 /// can be quite costly, especially if it's inside of an outer loop.
2194 /// This function solves this problem by detecting this type of loop and
2195 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
2196 /// the instructions for the maximum computation.
2198 ICmpInst *LoopStrengthReduce::OptimizeSMax(Loop *L, ICmpInst *Cond,
2199 IVStrideUse* &CondUse) {
2200 // Check that the loop matches the pattern we're looking for.
2201 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
2202 Cond->getPredicate() != CmpInst::ICMP_NE)
2205 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
2206 if (!Sel || !Sel->hasOneUse()) return Cond;
2208 SCEVHandle BackedgeTakenCount = SE->getBackedgeTakenCount(L);
2209 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2211 SCEVHandle One = SE->getIntegerSCEV(1, BackedgeTakenCount->getType());
2213 // Add one to the backedge-taken count to get the trip count.
2214 SCEVHandle IterationCount = SE->getAddExpr(BackedgeTakenCount, One);
2216 // Check for a max calculation that matches the pattern.
2217 SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(IterationCount);
2218 if (!SMax || SMax != SE->getSCEV(Sel)) return Cond;
2220 SCEVHandle SMaxLHS = SMax->getOperand(0);
2221 SCEVHandle SMaxRHS = SMax->getOperand(1);
2222 if (!SMaxLHS || SMaxLHS != One) return Cond;
2224 // Check the relevant induction variable for conformance to
2226 SCEVHandle IV = SE->getSCEV(Cond->getOperand(0));
2227 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2228 if (!AR || !AR->isAffine() ||
2229 AR->getStart() != One ||
2230 AR->getStepRecurrence(*SE) != One)
2233 assert(AR->getLoop() == L &&
2234 "Loop condition operand is an addrec in a different loop!");
2236 // Check the right operand of the select, and remember it, as it will
2237 // be used in the new comparison instruction.
2239 if (SE->getSCEV(Sel->getOperand(1)) == SMaxRHS)
2240 NewRHS = Sel->getOperand(1);
2241 else if (SE->getSCEV(Sel->getOperand(2)) == SMaxRHS)
2242 NewRHS = Sel->getOperand(2);
2243 if (!NewRHS) return Cond;
2245 // Ok, everything looks ok to change the condition into an SLT or SGE and
2246 // delete the max calculation.
2248 new ICmpInst(Cond->getPredicate() == CmpInst::ICMP_NE ?
2251 Cond->getOperand(0), NewRHS, "scmp", Cond);
2253 // Delete the max calculation instructions.
2254 SE->deleteValueFromRecords(Cond);
2255 Cond->replaceAllUsesWith(NewCond);
2256 Cond->eraseFromParent();
2257 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2258 SE->deleteValueFromRecords(Sel);
2259 Sel->eraseFromParent();
2260 if (Cmp->use_empty()) {
2261 SE->deleteValueFromRecords(Cmp);
2262 Cmp->eraseFromParent();
2264 CondUse->User = NewCond;
2268 /// OptimizeShadowIV - If IV is used in a int-to-float cast
2269 /// inside the loop then try to eliminate the cast opeation.
2270 void LoopStrengthReduce::OptimizeShadowIV(Loop *L) {
2272 SCEVHandle BackedgeTakenCount = SE->getBackedgeTakenCount(L);
2273 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2276 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e;
2278 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
2279 IVUsesByStride.find(StrideOrder[Stride]);
2280 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
2281 if (!isa<SCEVConstant>(SI->first))
2284 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
2285 E = SI->second.Users.end(); UI != E; /* empty */) {
2286 std::vector<IVStrideUse>::iterator CandidateUI = UI;
2288 Instruction *ShadowUse = CandidateUI->User;
2289 const Type *DestTy = NULL;
2291 /* If shadow use is a int->float cast then insert a second IV
2292 to eliminate this cast.
2294 for (unsigned i = 0; i < n; ++i)
2300 for (unsigned i = 0; i < n; ++i, ++d)
2303 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->User))
2304 DestTy = UCast->getDestTy();
2305 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->User))
2306 DestTy = SCast->getDestTy();
2307 if (!DestTy) continue;
2310 /* If target does not support DestTy natively then do not apply
2311 this transformation. */
2312 MVT DVT = TLI->getValueType(DestTy);
2313 if (!TLI->isTypeLegal(DVT)) continue;
2316 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
2318 if (PH->getNumIncomingValues() != 2) continue;
2320 const Type *SrcTy = PH->getType();
2321 int Mantissa = DestTy->getFPMantissaWidth();
2322 if (Mantissa == -1) continue;
2323 if ((int)SE->getTypeSizeInBits(SrcTy) > Mantissa)
2326 unsigned Entry, Latch;
2327 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
2335 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
2336 if (!Init) continue;
2337 ConstantFP *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());
2339 BinaryOperator *Incr =
2340 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
2341 if (!Incr) continue;
2342 if (Incr->getOpcode() != Instruction::Add
2343 && Incr->getOpcode() != Instruction::Sub)
2346 /* Initialize new IV, double d = 0.0 in above example. */
2347 ConstantInt *C = NULL;
2348 if (Incr->getOperand(0) == PH)
2349 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
2350 else if (Incr->getOperand(1) == PH)
2351 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
2357 /* Add new PHINode. */
2358 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);
2360 /* create new increment. '++d' in above example. */
2361 ConstantFP *CFP = ConstantFP::get(DestTy, C->getZExtValue());
2362 BinaryOperator *NewIncr =
2363 BinaryOperator::Create(Incr->getOpcode(),
2364 NewPH, CFP, "IV.S.next.", Incr);
2366 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
2367 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
2369 /* Remove cast operation */
2370 SE->deleteValueFromRecords(ShadowUse);
2371 ShadowUse->replaceAllUsesWith(NewPH);
2372 ShadowUse->eraseFromParent();
2373 SI->second.Users.erase(CandidateUI);
2380 // OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
2381 // uses in the loop, look to see if we can eliminate some, in favor of using
2382 // common indvars for the different uses.
2383 void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
2384 // TODO: implement optzns here.
2386 OptimizeShadowIV(L);
2388 // Finally, get the terminating condition for the loop if possible. If we
2389 // can, we want to change it to use a post-incremented version of its
2390 // induction variable, to allow coalescing the live ranges for the IV into
2391 // one register value.
2392 PHINode *SomePHI = cast<PHINode>(L->getHeader()->begin());
2393 BasicBlock *Preheader = L->getLoopPreheader();
2394 BasicBlock *LatchBlock =
2395 SomePHI->getIncomingBlock(SomePHI->getIncomingBlock(0) == Preheader);
2396 BranchInst *TermBr = dyn_cast<BranchInst>(LatchBlock->getTerminator());
2397 if (!TermBr || TermBr->isUnconditional() ||
2398 !isa<ICmpInst>(TermBr->getCondition()))
2400 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2402 // Search IVUsesByStride to find Cond's IVUse if there is one.
2403 IVStrideUse *CondUse = 0;
2404 const SCEVHandle *CondStride = 0;
2406 if (!FindIVUserForCond(Cond, CondUse, CondStride))
2407 return; // setcc doesn't use the IV.
2409 // If the trip count is computed in terms of an smax (due to ScalarEvolution
2410 // being unable to find a sufficient guard, for example), change the loop
2411 // comparison to use SLT instead of NE.
2412 Cond = OptimizeSMax(L, Cond, CondUse);
2414 // If possible, change stride and operands of the compare instruction to
2415 // eliminate one stride.
2416 Cond = ChangeCompareStride(L, Cond, CondUse, CondStride);
2418 // It's possible for the setcc instruction to be anywhere in the loop, and
2419 // possible for it to have multiple users. If it is not immediately before
2420 // the latch block branch, move it.
2421 if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
2422 if (Cond->hasOneUse()) { // Condition has a single use, just move it.
2423 Cond->moveBefore(TermBr);
2425 // Otherwise, clone the terminating condition and insert into the loopend.
2426 Cond = cast<ICmpInst>(Cond->clone());
2427 Cond->setName(L->getHeader()->getName() + ".termcond");
2428 LatchBlock->getInstList().insert(TermBr, Cond);
2430 // Clone the IVUse, as the old use still exists!
2431 IVUsesByStride[*CondStride].addUser(CondUse->Offset, Cond,
2432 CondUse->OperandValToReplace);
2433 CondUse = &IVUsesByStride[*CondStride].Users.back();
2437 // If we get to here, we know that we can transform the setcc instruction to
2438 // use the post-incremented version of the IV, allowing us to coalesce the
2439 // live ranges for the IV correctly.
2440 CondUse->Offset = SE->getMinusSCEV(CondUse->Offset, *CondStride);
2441 CondUse->isUseOfPostIncrementedValue = true;
2445 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager &LPM) {
2447 LI = &getAnalysis<LoopInfo>();
2448 DT = &getAnalysis<DominatorTree>();
2449 SE = &getAnalysis<ScalarEvolution>();
2452 // Find all uses of induction variables in this loop, and categorize
2453 // them by stride. Start by finding all of the PHI nodes in the header for
2454 // this loop. If they are induction variables, inspect their uses.
2455 SmallPtrSet<Instruction*,16> Processed; // Don't reprocess instructions.
2456 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I)
2457 AddUsersIfInteresting(I, L, Processed);
2459 if (!IVUsesByStride.empty()) {
2461 DOUT << "\nLSR on \"" << L->getHeader()->getParent()->getNameStart()
2466 // Sort the StrideOrder so we process larger strides first.
2467 std::stable_sort(StrideOrder.begin(), StrideOrder.end(), StrideCompare(SE));
2469 // Optimize induction variables. Some indvar uses can be transformed to use
2470 // strides that will be needed for other purposes. A common example of this
2471 // is the exit test for the loop, which can often be rewritten to use the
2472 // computation of some other indvar to decide when to terminate the loop.
2475 // FIXME: We can widen subreg IV's here for RISC targets. e.g. instead of
2476 // doing computation in byte values, promote to 32-bit values if safe.
2478 // FIXME: Attempt to reuse values across multiple IV's. In particular, we
2479 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should
2480 // be codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC.
2481 // Need to be careful that IV's are all the same type. Only works for
2482 // intptr_t indvars.
2484 // IVsByStride keeps IVs for one particular loop.
2485 assert(IVsByStride.empty() && "Stale entries in IVsByStride?");
2487 // Note: this processes each stride/type pair individually. All users
2488 // passed into StrengthReduceStridedIVUsers have the same type AND stride.
2489 // Also, note that we iterate over IVUsesByStride indirectly by using
2490 // StrideOrder. This extra layer of indirection makes the ordering of
2491 // strides deterministic - not dependent on map order.
2492 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e; ++Stride) {
2493 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
2494 IVUsesByStride.find(StrideOrder[Stride]);
2495 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
2496 StrengthReduceStridedIVUsers(SI->first, SI->second, L);
2500 // We're done analyzing this loop; release all the state we built up for it.
2501 IVUsesByStride.clear();
2502 IVsByStride.clear();
2503 StrideOrder.clear();
2505 // Clean up after ourselves
2506 if (!DeadInsts.empty())
2507 DeleteTriviallyDeadInstructions();
2509 // At this point, it is worth checking to see if any recurrence PHIs are also
2510 // dead, so that we can remove them as well. To keep ScalarEvolution
2511 // current, use a ValueDeletionListener class.
2512 struct LSRListener : public ValueDeletionListener {
2513 ScalarEvolution &SE;
2514 explicit LSRListener(ScalarEvolution &se) : SE(se) {}
2516 virtual void ValueWillBeDeleted(Value *V) {
2517 SE.deleteValueFromRecords(V);
2520 DeleteDeadPHIs(L->getHeader(), &VDL);