1 //===- LoopStrengthReduce.cpp - Strength Reduce GEPs in Loops -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by Nate Begeman and is distributed under the
6 // University of Illinois Open Source 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. This is
12 // accomplished by creating a new Value to hold the initial value of the array
13 // access for the first iteration, and then creating a new GEP instruction in
14 // the loop to increment the value by the appropriate amount.
16 //===----------------------------------------------------------------------===//
18 #define DEBUG_TYPE "loop-reduce"
19 #include "llvm/Transforms/Scalar.h"
20 #include "llvm/Constants.h"
21 #include "llvm/Instructions.h"
22 #include "llvm/Type.h"
23 #include "llvm/DerivedTypes.h"
24 #include "llvm/Analysis/Dominators.h"
25 #include "llvm/Analysis/LoopInfo.h"
26 #include "llvm/Analysis/LoopPass.h"
27 #include "llvm/Analysis/ScalarEvolutionExpander.h"
28 #include "llvm/Support/CFG.h"
29 #include "llvm/Support/GetElementPtrTypeIterator.h"
30 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
31 #include "llvm/Transforms/Utils/Local.h"
32 #include "llvm/Target/TargetData.h"
33 #include "llvm/ADT/Statistic.h"
34 #include "llvm/Support/Debug.h"
35 #include "llvm/Support/Compiler.h"
36 #include "llvm/Target/TargetLowering.h"
41 STATISTIC(NumReduced , "Number of GEPs strength reduced");
42 STATISTIC(NumInserted, "Number of PHIs inserted");
43 STATISTIC(NumVariable, "Number of PHIs with variable strides");
49 /// IVStrideUse - Keep track of one use of a strided induction variable, where
50 /// the stride is stored externally. The Offset member keeps track of the
51 /// offset from the IV, User is the actual user of the operand, and 'Operand'
52 /// is the operand # of the User that is the use.
53 struct VISIBILITY_HIDDEN IVStrideUse {
56 Value *OperandValToReplace;
58 // isUseOfPostIncrementedValue - True if this should use the
59 // post-incremented version of this IV, not the preincremented version.
60 // This can only be set in special cases, such as the terminating setcc
61 // instruction for a loop or uses dominated by the loop.
62 bool isUseOfPostIncrementedValue;
64 IVStrideUse(const SCEVHandle &Offs, Instruction *U, Value *O)
65 : Offset(Offs), User(U), OperandValToReplace(O),
66 isUseOfPostIncrementedValue(false) {}
69 /// IVUsersOfOneStride - This structure keeps track of all instructions that
70 /// have an operand that is based on the trip count multiplied by some stride.
71 /// The stride for all of these users is common and kept external to this
73 struct VISIBILITY_HIDDEN IVUsersOfOneStride {
74 /// Users - Keep track of all of the users of this stride as well as the
75 /// initial value and the operand that uses the IV.
76 std::vector<IVStrideUse> Users;
78 void addUser(const SCEVHandle &Offset,Instruction *User, Value *Operand) {
79 Users.push_back(IVStrideUse(Offset, User, Operand));
83 /// IVInfo - This structure keeps track of one IV expression inserted during
84 /// StrengthReduceStridedIVUsers. It contains the stride, the common base, as
85 /// well as the PHI node and increment value created for rewrite.
86 struct VISIBILITY_HIDDEN IVExpr {
93 : Stride(SCEVUnknown::getIntegerSCEV(0, Type::Int32Ty)),
94 Base (SCEVUnknown::getIntegerSCEV(0, Type::Int32Ty)) {}
95 IVExpr(const SCEVHandle &stride, const SCEVHandle &base, PHINode *phi,
97 : Stride(stride), Base(base), PHI(phi), IncV(incv) {}
100 /// IVsOfOneStride - This structure keeps track of all IV expression inserted
101 /// during StrengthReduceStridedIVUsers for a particular stride of the IV.
102 struct VISIBILITY_HIDDEN IVsOfOneStride {
103 std::vector<IVExpr> IVs;
105 void addIV(const SCEVHandle &Stride, const SCEVHandle &Base, PHINode *PHI,
107 IVs.push_back(IVExpr(Stride, Base, PHI, IncV));
111 class VISIBILITY_HIDDEN LoopStrengthReduce : public LoopPass {
115 const TargetData *TD;
116 const Type *UIntPtrTy;
119 /// IVUsesByStride - Keep track of all uses of induction variables that we
120 /// are interested in. The key of the map is the stride of the access.
121 std::map<SCEVHandle, IVUsersOfOneStride> IVUsesByStride;
123 /// IVsByStride - Keep track of all IVs that have been inserted for a
124 /// particular stride.
125 std::map<SCEVHandle, IVsOfOneStride> IVsByStride;
127 /// StrideOrder - An ordering of the keys in IVUsesByStride that is stable:
128 /// We use this to iterate over the IVUsesByStride collection without being
129 /// dependent on random ordering of pointers in the process.
130 std::vector<SCEVHandle> StrideOrder;
132 /// CastedValues - As we need to cast values to uintptr_t, this keeps track
133 /// of the casted version of each value. This is accessed by
134 /// getCastedVersionOf.
135 std::map<Value*, Value*> CastedPointers;
137 /// DeadInsts - Keep track of instructions we may have made dead, so that
138 /// we can remove them after we are done working.
139 std::set<Instruction*> DeadInsts;
141 /// TLI - Keep a pointer of a TargetLowering to consult for determining
142 /// transformation profitability.
143 const TargetLowering *TLI;
146 static char ID; // Pass ID, replacement for typeid
147 LoopStrengthReduce(const TargetLowering *tli = NULL) :
148 LoopPass((intptr_t)&ID), TLI(tli) {
151 bool runOnLoop(Loop *L, LPPassManager &LPM);
153 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
154 // We split critical edges, so we change the CFG. However, we do update
155 // many analyses if they are around.
156 AU.addPreservedID(LoopSimplifyID);
157 AU.addPreserved<LoopInfo>();
158 AU.addPreserved<ETForest>();
159 AU.addPreserved<DominanceFrontier>();
160 AU.addPreserved<DominatorTree>();
162 AU.addRequiredID(LoopSimplifyID);
163 AU.addRequired<LoopInfo>();
164 AU.addRequired<ETForest>();
165 AU.addRequired<TargetData>();
166 AU.addRequired<ScalarEvolution>();
169 /// getCastedVersionOf - Return the specified value casted to uintptr_t.
171 Value *getCastedVersionOf(Instruction::CastOps opcode, Value *V);
173 bool AddUsersIfInteresting(Instruction *I, Loop *L,
174 std::set<Instruction*> &Processed);
175 SCEVHandle GetExpressionSCEV(Instruction *E, Loop *L);
177 void OptimizeIndvars(Loop *L);
178 bool FindIVForUser(ICmpInst *Cond, IVStrideUse *&CondUse,
179 const SCEVHandle *&CondStride);
181 unsigned CheckForIVReuse(const SCEVHandle&, IVExpr&, const Type*,
182 const std::vector<BasedUser>& UsersToProcess);
184 bool ValidStride(int64_t, const std::vector<BasedUser>& UsersToProcess);
186 void StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
187 IVUsersOfOneStride &Uses,
188 Loop *L, bool isOnlyStride);
189 void DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts);
191 char LoopStrengthReduce::ID = 0;
192 RegisterPass<LoopStrengthReduce> X("loop-reduce", "Loop Strength Reduction");
195 LoopPass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
196 return new LoopStrengthReduce(TLI);
199 /// getCastedVersionOf - Return the specified value casted to uintptr_t. This
200 /// assumes that the Value* V is of integer or pointer type only.
202 Value *LoopStrengthReduce::getCastedVersionOf(Instruction::CastOps opcode,
204 if (V->getType() == UIntPtrTy) return V;
205 if (Constant *CB = dyn_cast<Constant>(V))
206 return ConstantExpr::getCast(opcode, CB, UIntPtrTy);
208 Value *&New = CastedPointers[V];
211 New = SCEVExpander::InsertCastOfTo(opcode, V, UIntPtrTy);
212 DeadInsts.insert(cast<Instruction>(New));
217 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
218 /// specified set are trivially dead, delete them and see if this makes any of
219 /// their operands subsequently dead.
220 void LoopStrengthReduce::
221 DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts) {
222 while (!Insts.empty()) {
223 Instruction *I = *Insts.begin();
224 Insts.erase(Insts.begin());
225 if (isInstructionTriviallyDead(I)) {
226 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
227 if (Instruction *U = dyn_cast<Instruction>(I->getOperand(i)))
229 SE->deleteInstructionFromRecords(I);
230 I->eraseFromParent();
237 /// GetExpressionSCEV - Compute and return the SCEV for the specified
239 SCEVHandle LoopStrengthReduce::GetExpressionSCEV(Instruction *Exp, Loop *L) {
240 // Pointer to pointer bitcast instructions return the same value as their
242 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Exp)) {
243 if (SE->hasSCEV(BCI) || !isa<Instruction>(BCI->getOperand(0)))
244 return SE->getSCEV(BCI);
245 SCEVHandle R = GetExpressionSCEV(cast<Instruction>(BCI->getOperand(0)), L);
250 // Scalar Evolutions doesn't know how to compute SCEV's for GEP instructions.
251 // If this is a GEP that SE doesn't know about, compute it now and insert it.
252 // If this is not a GEP, or if we have already done this computation, just let
254 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Exp);
255 if (!GEP || SE->hasSCEV(GEP))
256 return SE->getSCEV(Exp);
258 // Analyze all of the subscripts of this getelementptr instruction, looking
259 // for uses that are determined by the trip count of L. First, skip all
260 // operands the are not dependent on the IV.
262 // Build up the base expression. Insert an LLVM cast of the pointer to
264 SCEVHandle GEPVal = SCEVUnknown::get(
265 getCastedVersionOf(Instruction::PtrToInt, GEP->getOperand(0)));
267 gep_type_iterator GTI = gep_type_begin(GEP);
269 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
270 // If this is a use of a recurrence that we can analyze, and it comes before
271 // Op does in the GEP operand list, we will handle this when we process this
273 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
274 const StructLayout *SL = TD->getStructLayout(STy);
275 unsigned Idx = cast<ConstantInt>(GEP->getOperand(i))->getZExtValue();
276 uint64_t Offset = SL->getElementOffset(Idx);
277 GEPVal = SCEVAddExpr::get(GEPVal,
278 SCEVUnknown::getIntegerSCEV(Offset, UIntPtrTy));
280 unsigned GEPOpiBits =
281 GEP->getOperand(i)->getType()->getPrimitiveSizeInBits();
282 unsigned IntPtrBits = UIntPtrTy->getPrimitiveSizeInBits();
283 Instruction::CastOps opcode = (GEPOpiBits < IntPtrBits ?
284 Instruction::SExt : (GEPOpiBits > IntPtrBits ? Instruction::Trunc :
285 Instruction::BitCast));
286 Value *OpVal = getCastedVersionOf(opcode, GEP->getOperand(i));
287 SCEVHandle Idx = SE->getSCEV(OpVal);
289 uint64_t TypeSize = TD->getTypeSize(GTI.getIndexedType());
291 Idx = SCEVMulExpr::get(Idx,
292 SCEVConstant::get(ConstantInt::get(UIntPtrTy,
294 GEPVal = SCEVAddExpr::get(GEPVal, Idx);
298 SE->setSCEV(GEP, GEPVal);
302 /// getSCEVStartAndStride - Compute the start and stride of this expression,
303 /// returning false if the expression is not a start/stride pair, or true if it
304 /// is. The stride must be a loop invariant expression, but the start may be
305 /// a mix of loop invariant and loop variant expressions.
306 static bool getSCEVStartAndStride(const SCEVHandle &SH, Loop *L,
307 SCEVHandle &Start, SCEVHandle &Stride) {
308 SCEVHandle TheAddRec = Start; // Initialize to zero.
310 // If the outer level is an AddExpr, the operands are all start values except
311 // for a nested AddRecExpr.
312 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(SH)) {
313 for (unsigned i = 0, e = AE->getNumOperands(); i != e; ++i)
314 if (SCEVAddRecExpr *AddRec =
315 dyn_cast<SCEVAddRecExpr>(AE->getOperand(i))) {
316 if (AddRec->getLoop() == L)
317 TheAddRec = SCEVAddExpr::get(AddRec, TheAddRec);
319 return false; // Nested IV of some sort?
321 Start = SCEVAddExpr::get(Start, AE->getOperand(i));
324 } else if (isa<SCEVAddRecExpr>(SH)) {
327 return false; // not analyzable.
330 SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(TheAddRec);
331 if (!AddRec || AddRec->getLoop() != L) return false;
333 // FIXME: Generalize to non-affine IV's.
334 if (!AddRec->isAffine()) return false;
336 Start = SCEVAddExpr::get(Start, AddRec->getOperand(0));
338 if (!isa<SCEVConstant>(AddRec->getOperand(1)))
339 DOUT << "[" << L->getHeader()->getName()
340 << "] Variable stride: " << *AddRec << "\n";
342 Stride = AddRec->getOperand(1);
346 /// IVUseShouldUsePostIncValue - We have discovered a "User" of an IV expression
347 /// and now we need to decide whether the user should use the preinc or post-inc
348 /// value. If this user should use the post-inc version of the IV, return true.
350 /// Choosing wrong here can break dominance properties (if we choose to use the
351 /// post-inc value when we cannot) or it can end up adding extra live-ranges to
352 /// the loop, resulting in reg-reg copies (if we use the pre-inc value when we
353 /// should use the post-inc value).
354 static bool IVUseShouldUsePostIncValue(Instruction *User, Instruction *IV,
355 Loop *L, ETForest *EF, Pass *P) {
356 // If the user is in the loop, use the preinc value.
357 if (L->contains(User->getParent())) return false;
359 BasicBlock *LatchBlock = L->getLoopLatch();
361 // Ok, the user is outside of the loop. If it is dominated by the latch
362 // block, use the post-inc value.
363 if (EF->dominates(LatchBlock, User->getParent()))
366 // There is one case we have to be careful of: PHI nodes. These little guys
367 // can live in blocks that do not dominate the latch block, but (since their
368 // uses occur in the predecessor block, not the block the PHI lives in) should
369 // still use the post-inc value. Check for this case now.
370 PHINode *PN = dyn_cast<PHINode>(User);
371 if (!PN) return false; // not a phi, not dominated by latch block.
373 // Look at all of the uses of IV by the PHI node. If any use corresponds to
374 // a block that is not dominated by the latch block, give up and use the
375 // preincremented value.
376 unsigned NumUses = 0;
377 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
378 if (PN->getIncomingValue(i) == IV) {
380 if (!EF->dominates(LatchBlock, PN->getIncomingBlock(i)))
384 // Okay, all uses of IV by PN are in predecessor blocks that really are
385 // dominated by the latch block. Split the critical edges and use the
386 // post-incremented value.
387 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
388 if (PN->getIncomingValue(i) == IV) {
389 SplitCriticalEdge(PN->getIncomingBlock(i), PN->getParent(), P,
391 // Splitting the critical edge can reduce the number of entries in this
393 e = PN->getNumIncomingValues();
394 if (--NumUses == 0) break;
402 /// AddUsersIfInteresting - Inspect the specified instruction. If it is a
403 /// reducible SCEV, recursively add its users to the IVUsesByStride set and
404 /// return true. Otherwise, return false.
405 bool LoopStrengthReduce::AddUsersIfInteresting(Instruction *I, Loop *L,
406 std::set<Instruction*> &Processed) {
407 if (!I->getType()->isInteger() && !isa<PointerType>(I->getType()))
408 return false; // Void and FP expressions cannot be reduced.
409 if (!Processed.insert(I).second)
410 return true; // Instruction already handled.
412 // Get the symbolic expression for this instruction.
413 SCEVHandle ISE = GetExpressionSCEV(I, L);
414 if (isa<SCEVCouldNotCompute>(ISE)) return false;
416 // Get the start and stride for this expression.
417 SCEVHandle Start = SCEVUnknown::getIntegerSCEV(0, ISE->getType());
418 SCEVHandle Stride = Start;
419 if (!getSCEVStartAndStride(ISE, L, Start, Stride))
420 return false; // Non-reducible symbolic expression, bail out.
422 std::vector<Instruction *> IUsers;
423 // Collect all I uses now because IVUseShouldUsePostIncValue may
424 // invalidate use_iterator.
425 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; ++UI)
426 IUsers.push_back(cast<Instruction>(*UI));
428 for (unsigned iused_index = 0, iused_size = IUsers.size();
429 iused_index != iused_size; ++iused_index) {
431 Instruction *User = IUsers[iused_index];
433 // Do not infinitely recurse on PHI nodes.
434 if (isa<PHINode>(User) && Processed.count(User))
437 // If this is an instruction defined in a nested loop, or outside this loop,
438 // don't recurse into it.
439 bool AddUserToIVUsers = false;
440 if (LI->getLoopFor(User->getParent()) != L) {
441 DOUT << "FOUND USER in other loop: " << *User
442 << " OF SCEV: " << *ISE << "\n";
443 AddUserToIVUsers = true;
444 } else if (!AddUsersIfInteresting(User, L, Processed)) {
445 DOUT << "FOUND USER: " << *User
446 << " OF SCEV: " << *ISE << "\n";
447 AddUserToIVUsers = true;
450 if (AddUserToIVUsers) {
451 IVUsersOfOneStride &StrideUses = IVUsesByStride[Stride];
452 if (StrideUses.Users.empty()) // First occurance of this stride?
453 StrideOrder.push_back(Stride);
455 // Okay, we found a user that we cannot reduce. Analyze the instruction
456 // and decide what to do with it. If we are a use inside of the loop, use
457 // the value before incrementation, otherwise use it after incrementation.
458 if (IVUseShouldUsePostIncValue(User, I, L, EF, this)) {
459 // The value used will be incremented by the stride more than we are
460 // expecting, so subtract this off.
461 SCEVHandle NewStart = SCEV::getMinusSCEV(Start, Stride);
462 StrideUses.addUser(NewStart, User, I);
463 StrideUses.Users.back().isUseOfPostIncrementedValue = true;
464 DOUT << " USING POSTINC SCEV, START=" << *NewStart<< "\n";
466 StrideUses.addUser(Start, User, I);
474 /// BasedUser - For a particular base value, keep information about how we've
475 /// partitioned the expression so far.
477 /// Base - The Base value for the PHI node that needs to be inserted for
478 /// this use. As the use is processed, information gets moved from this
479 /// field to the Imm field (below). BasedUser values are sorted by this
483 /// Inst - The instruction using the induction variable.
486 /// OperandValToReplace - The operand value of Inst to replace with the
488 Value *OperandValToReplace;
490 /// Imm - The immediate value that should be added to the base immediately
491 /// before Inst, because it will be folded into the imm field of the
495 /// EmittedBase - The actual value* to use for the base value of this
496 /// operation. This is null if we should just use zero so far.
499 // isUseOfPostIncrementedValue - True if this should use the
500 // post-incremented version of this IV, not the preincremented version.
501 // This can only be set in special cases, such as the terminating setcc
502 // instruction for a loop and uses outside the loop that are dominated by
504 bool isUseOfPostIncrementedValue;
506 BasedUser(IVStrideUse &IVSU)
507 : Base(IVSU.Offset), Inst(IVSU.User),
508 OperandValToReplace(IVSU.OperandValToReplace),
509 Imm(SCEVUnknown::getIntegerSCEV(0, Base->getType())), EmittedBase(0),
510 isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue) {}
512 // Once we rewrite the code to insert the new IVs we want, update the
513 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
515 void RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
516 SCEVExpander &Rewriter, Loop *L,
519 Value *InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
520 SCEVExpander &Rewriter,
521 Instruction *IP, Loop *L);
526 void BasedUser::dump() const {
527 cerr << " Base=" << *Base;
528 cerr << " Imm=" << *Imm;
530 cerr << " EB=" << *EmittedBase;
532 cerr << " Inst: " << *Inst;
535 Value *BasedUser::InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
536 SCEVExpander &Rewriter,
537 Instruction *IP, Loop *L) {
538 // Figure out where we *really* want to insert this code. In particular, if
539 // the user is inside of a loop that is nested inside of L, we really don't
540 // want to insert this expression before the user, we'd rather pull it out as
541 // many loops as possible.
542 LoopInfo &LI = Rewriter.getLoopInfo();
543 Instruction *BaseInsertPt = IP;
545 // Figure out the most-nested loop that IP is in.
546 Loop *InsertLoop = LI.getLoopFor(IP->getParent());
548 // If InsertLoop is not L, and InsertLoop is nested inside of L, figure out
549 // the preheader of the outer-most loop where NewBase is not loop invariant.
550 while (InsertLoop && NewBase->isLoopInvariant(InsertLoop)) {
551 BaseInsertPt = InsertLoop->getLoopPreheader()->getTerminator();
552 InsertLoop = InsertLoop->getParentLoop();
555 // If there is no immediate value, skip the next part.
556 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Imm))
557 if (SC->getValue()->isZero())
558 return Rewriter.expandCodeFor(NewBase, BaseInsertPt,
559 OperandValToReplace->getType());
561 Value *Base = Rewriter.expandCodeFor(NewBase, BaseInsertPt);
563 // Always emit the immediate (if non-zero) into the same block as the user.
564 SCEVHandle NewValSCEV = SCEVAddExpr::get(SCEVUnknown::get(Base), Imm);
565 return Rewriter.expandCodeFor(NewValSCEV, IP,
566 OperandValToReplace->getType());
570 // Once we rewrite the code to insert the new IVs we want, update the
571 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
573 void BasedUser::RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
574 SCEVExpander &Rewriter,
576 if (!isa<PHINode>(Inst)) {
577 // By default, insert code at the user instruction.
578 BasicBlock::iterator InsertPt = Inst;
580 // However, if the Operand is itself an instruction, the (potentially
581 // complex) inserted code may be shared by many users. Because of this, we
582 // want to emit code for the computation of the operand right before its old
583 // computation. This is usually safe, because we obviously used to use the
584 // computation when it was computed in its current block. However, in some
585 // cases (e.g. use of a post-incremented induction variable) the NewBase
586 // value will be pinned to live somewhere after the original computation.
587 // In this case, we have to back off.
588 if (!isUseOfPostIncrementedValue) {
589 if (Instruction *OpInst = dyn_cast<Instruction>(OperandValToReplace)) {
591 while (isa<PHINode>(InsertPt)) ++InsertPt;
595 Value *NewVal = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
596 // Replace the use of the operand Value with the new Phi we just created.
597 Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
598 DOUT << " CHANGED: IMM =" << *Imm << " Inst = " << *Inst;
602 // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm
603 // expression into each operand block that uses it. Note that PHI nodes can
604 // have multiple entries for the same predecessor. We use a map to make sure
605 // that a PHI node only has a single Value* for each predecessor (which also
606 // prevents us from inserting duplicate code in some blocks).
607 std::map<BasicBlock*, Value*> InsertedCode;
608 PHINode *PN = cast<PHINode>(Inst);
609 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
610 if (PN->getIncomingValue(i) == OperandValToReplace) {
611 // If this is a critical edge, split the edge so that we do not insert the
612 // code on all predecessor/successor paths. We do this unless this is the
613 // canonical backedge for this loop, as this can make some inserted code
614 // be in an illegal position.
615 BasicBlock *PHIPred = PN->getIncomingBlock(i);
616 if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 &&
617 (PN->getParent() != L->getHeader() || !L->contains(PHIPred))) {
619 // First step, split the critical edge.
620 SplitCriticalEdge(PHIPred, PN->getParent(), P, true);
622 // Next step: move the basic block. In particular, if the PHI node
623 // is outside of the loop, and PredTI is in the loop, we want to
624 // move the block to be immediately before the PHI block, not
625 // immediately after PredTI.
626 if (L->contains(PHIPred) && !L->contains(PN->getParent())) {
627 BasicBlock *NewBB = PN->getIncomingBlock(i);
628 NewBB->moveBefore(PN->getParent());
631 // Splitting the edge can reduce the number of PHI entries we have.
632 e = PN->getNumIncomingValues();
635 Value *&Code = InsertedCode[PN->getIncomingBlock(i)];
637 // Insert the code into the end of the predecessor block.
638 Instruction *InsertPt = PN->getIncomingBlock(i)->getTerminator();
639 Code = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
642 // Replace the use of the operand Value with the new Phi we just created.
643 PN->setIncomingValue(i, Code);
647 DOUT << " CHANGED: IMM =" << *Imm << " Inst = " << *Inst;
651 /// isTargetConstant - Return true if the following can be referenced by the
652 /// immediate field of a target instruction.
653 static bool isTargetConstant(const SCEVHandle &V, const Type *UseTy,
654 const TargetLowering *TLI) {
655 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
656 int64_t VC = SC->getValue()->getSExtValue();
658 TargetLowering::AddrMode AM;
660 return TLI->isLegalAddressingMode(AM, UseTy);
662 // Defaults to PPC. PPC allows a sign-extended 16-bit immediate field.
663 return (VC > -(1 << 16) && VC < (1 << 16)-1);
667 if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
668 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(SU->getValue()))
669 if (TLI && CE->getOpcode() == Instruction::PtrToInt) {
670 Constant *Op0 = CE->getOperand(0);
671 if (GlobalValue *GV = dyn_cast<GlobalValue>(Op0)) {
672 TargetLowering::AddrMode AM;
674 return TLI->isLegalAddressingMode(AM, UseTy);
680 /// MoveLoopVariantsToImediateField - Move any subexpressions from Val that are
681 /// loop varying to the Imm operand.
682 static void MoveLoopVariantsToImediateField(SCEVHandle &Val, SCEVHandle &Imm,
684 if (Val->isLoopInvariant(L)) return; // Nothing to do.
686 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
687 std::vector<SCEVHandle> NewOps;
688 NewOps.reserve(SAE->getNumOperands());
690 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
691 if (!SAE->getOperand(i)->isLoopInvariant(L)) {
692 // If this is a loop-variant expression, it must stay in the immediate
693 // field of the expression.
694 Imm = SCEVAddExpr::get(Imm, SAE->getOperand(i));
696 NewOps.push_back(SAE->getOperand(i));
700 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
702 Val = SCEVAddExpr::get(NewOps);
703 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
704 // Try to pull immediates out of the start value of nested addrec's.
705 SCEVHandle Start = SARE->getStart();
706 MoveLoopVariantsToImediateField(Start, Imm, L);
708 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
710 Val = SCEVAddRecExpr::get(Ops, SARE->getLoop());
712 // Otherwise, all of Val is variant, move the whole thing over.
713 Imm = SCEVAddExpr::get(Imm, Val);
714 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
719 /// MoveImmediateValues - Look at Val, and pull out any additions of constants
720 /// that can fit into the immediate field of instructions in the target.
721 /// Accumulate these immediate values into the Imm value.
722 static void MoveImmediateValues(const TargetLowering *TLI,
724 SCEVHandle &Val, SCEVHandle &Imm,
725 bool isAddress, Loop *L) {
726 const Type *UseTy = User->getType();
727 if (StoreInst *SI = dyn_cast<StoreInst>(User))
728 UseTy = SI->getOperand(0)->getType();
730 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
731 std::vector<SCEVHandle> NewOps;
732 NewOps.reserve(SAE->getNumOperands());
734 for (unsigned i = 0; i != SAE->getNumOperands(); ++i) {
735 SCEVHandle NewOp = SAE->getOperand(i);
736 MoveImmediateValues(TLI, User, NewOp, Imm, isAddress, L);
738 if (!NewOp->isLoopInvariant(L)) {
739 // If this is a loop-variant expression, it must stay in the immediate
740 // field of the expression.
741 Imm = SCEVAddExpr::get(Imm, NewOp);
743 NewOps.push_back(NewOp);
748 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
750 Val = SCEVAddExpr::get(NewOps);
752 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
753 // Try to pull immediates out of the start value of nested addrec's.
754 SCEVHandle Start = SARE->getStart();
755 MoveImmediateValues(TLI, User, Start, Imm, isAddress, L);
757 if (Start != SARE->getStart()) {
758 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
760 Val = SCEVAddRecExpr::get(Ops, SARE->getLoop());
763 } else if (SCEVMulExpr *SME = dyn_cast<SCEVMulExpr>(Val)) {
764 // Transform "8 * (4 + v)" -> "32 + 8*V" if "32" fits in the immed field.
765 if (isAddress && isTargetConstant(SME->getOperand(0), UseTy, TLI) &&
766 SME->getNumOperands() == 2 && SME->isLoopInvariant(L)) {
768 SCEVHandle SubImm = SCEVUnknown::getIntegerSCEV(0, Val->getType());
769 SCEVHandle NewOp = SME->getOperand(1);
770 MoveImmediateValues(TLI, User, NewOp, SubImm, isAddress, L);
772 // If we extracted something out of the subexpressions, see if we can
774 if (NewOp != SME->getOperand(1)) {
775 // Scale SubImm up by "8". If the result is a target constant, we are
777 SubImm = SCEVMulExpr::get(SubImm, SME->getOperand(0));
778 if (isTargetConstant(SubImm, UseTy, TLI)) {
779 // Accumulate the immediate.
780 Imm = SCEVAddExpr::get(Imm, SubImm);
782 // Update what is left of 'Val'.
783 Val = SCEVMulExpr::get(SME->getOperand(0), NewOp);
790 // Loop-variant expressions must stay in the immediate field of the
792 if ((isAddress && isTargetConstant(Val, UseTy, TLI)) ||
793 !Val->isLoopInvariant(L)) {
794 Imm = SCEVAddExpr::get(Imm, Val);
795 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
799 // Otherwise, no immediates to move.
803 /// SeparateSubExprs - Decompose Expr into all of the subexpressions that are
804 /// added together. This is used to reassociate common addition subexprs
805 /// together for maximal sharing when rewriting bases.
806 static void SeparateSubExprs(std::vector<SCEVHandle> &SubExprs,
808 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) {
809 for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
810 SeparateSubExprs(SubExprs, AE->getOperand(j));
811 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) {
812 SCEVHandle Zero = SCEVUnknown::getIntegerSCEV(0, Expr->getType());
813 if (SARE->getOperand(0) == Zero) {
814 SubExprs.push_back(Expr);
816 // Compute the addrec with zero as its base.
817 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
818 Ops[0] = Zero; // Start with zero base.
819 SubExprs.push_back(SCEVAddRecExpr::get(Ops, SARE->getLoop()));
822 SeparateSubExprs(SubExprs, SARE->getOperand(0));
824 } else if (!isa<SCEVConstant>(Expr) ||
825 !cast<SCEVConstant>(Expr)->getValue()->isZero()) {
827 SubExprs.push_back(Expr);
832 /// RemoveCommonExpressionsFromUseBases - Look through all of the uses in Bases,
833 /// removing any common subexpressions from it. Anything truly common is
834 /// removed, accumulated, and returned. This looks for things like (a+b+c) and
835 /// (a+c+d) -> (a+c). The common expression is *removed* from the Bases.
837 RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses) {
838 unsigned NumUses = Uses.size();
840 // Only one use? Use its base, regardless of what it is!
841 SCEVHandle Zero = SCEVUnknown::getIntegerSCEV(0, Uses[0].Base->getType());
842 SCEVHandle Result = Zero;
844 std::swap(Result, Uses[0].Base);
848 // To find common subexpressions, count how many of Uses use each expression.
849 // If any subexpressions are used Uses.size() times, they are common.
850 std::map<SCEVHandle, unsigned> SubExpressionUseCounts;
852 // UniqueSubExprs - Keep track of all of the subexpressions we see in the
853 // order we see them.
854 std::vector<SCEVHandle> UniqueSubExprs;
856 std::vector<SCEVHandle> SubExprs;
857 for (unsigned i = 0; i != NumUses; ++i) {
858 // If the base is zero (which is common), return zero now, there are no
860 if (Uses[i].Base == Zero) return Zero;
862 // Split the expression into subexprs.
863 SeparateSubExprs(SubExprs, Uses[i].Base);
864 // Add one to SubExpressionUseCounts for each subexpr present.
865 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
866 if (++SubExpressionUseCounts[SubExprs[j]] == 1)
867 UniqueSubExprs.push_back(SubExprs[j]);
871 // Now that we know how many times each is used, build Result. Iterate over
872 // UniqueSubexprs so that we have a stable ordering.
873 for (unsigned i = 0, e = UniqueSubExprs.size(); i != e; ++i) {
874 std::map<SCEVHandle, unsigned>::iterator I =
875 SubExpressionUseCounts.find(UniqueSubExprs[i]);
876 assert(I != SubExpressionUseCounts.end() && "Entry not found?");
877 if (I->second == NumUses) { // Found CSE!
878 Result = SCEVAddExpr::get(Result, I->first);
880 // Remove non-cse's from SubExpressionUseCounts.
881 SubExpressionUseCounts.erase(I);
885 // If we found no CSE's, return now.
886 if (Result == Zero) return Result;
888 // Otherwise, remove all of the CSE's we found from each of the base values.
889 for (unsigned i = 0; i != NumUses; ++i) {
890 // Split the expression into subexprs.
891 SeparateSubExprs(SubExprs, Uses[i].Base);
893 // Remove any common subexpressions.
894 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
895 if (SubExpressionUseCounts.count(SubExprs[j])) {
896 SubExprs.erase(SubExprs.begin()+j);
900 // Finally, the non-shared expressions together.
901 if (SubExprs.empty())
904 Uses[i].Base = SCEVAddExpr::get(SubExprs);
911 /// isZero - returns true if the scalar evolution expression is zero.
913 static bool isZero(SCEVHandle &V) {
914 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(V))
915 return SC->getValue()->isZero();
919 /// ValidStride - Check whether the given Scale is valid for all loads and
920 /// stores in UsersToProcess.
922 bool LoopStrengthReduce::ValidStride(int64_t Scale,
923 const std::vector<BasedUser>& UsersToProcess) {
924 for (unsigned i=0, e = UsersToProcess.size(); i!=e; ++i) {
925 // If this is a load or other access, pass the type of the access in.
926 const Type *AccessTy = Type::VoidTy;
927 if (StoreInst *SI = dyn_cast<StoreInst>(UsersToProcess[i].Inst))
928 AccessTy = SI->getOperand(0)->getType();
929 else if (LoadInst *LI = dyn_cast<LoadInst>(UsersToProcess[i].Inst))
930 AccessTy = LI->getType();
932 TargetLowering::AddrMode AM;
933 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
934 AM.BaseOffs = SC->getValue()->getSExtValue();
937 // If load[imm+r*scale] is illegal, bail out.
938 if (!TLI->isLegalAddressingMode(AM, AccessTy))
944 /// CheckForIVReuse - Returns the multiple if the stride is the multiple
945 /// of a previous stride and it is a legal value for the target addressing
946 /// mode scale component. This allows the users of this stride to be rewritten
947 /// as prev iv * factor. It returns 0 if no reuse is possible.
948 unsigned LoopStrengthReduce::CheckForIVReuse(const SCEVHandle &Stride,
949 IVExpr &IV, const Type *Ty,
950 const std::vector<BasedUser>& UsersToProcess) {
953 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Stride)) {
954 int64_t SInt = SC->getValue()->getSExtValue();
955 if (SInt == 1) return 0;
957 for (std::map<SCEVHandle, IVsOfOneStride>::iterator SI= IVsByStride.begin(),
958 SE = IVsByStride.end(); SI != SE; ++SI) {
959 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
960 if (SInt != -SSInt &&
961 (unsigned(abs(SInt)) < SSInt || (SInt % SSInt) != 0))
963 int64_t Scale = SInt / SSInt;
964 // Check that this stride is valid for all the types used for loads and
965 // stores; if it can be used for some and not others, we might as well use
966 // the original stride everywhere, since we have to create the IV for it
968 if (ValidStride(Scale, UsersToProcess))
969 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
970 IE = SI->second.IVs.end(); II != IE; ++II)
971 // FIXME: Only handle base == 0 for now.
972 // Only reuse previous IV if it would not require a type conversion.
973 if (isZero(II->Base) && II->Base->getType() == Ty) {
982 /// PartitionByIsUseOfPostIncrementedValue - Simple boolean predicate that
983 /// returns true if Val's isUseOfPostIncrementedValue is true.
984 static bool PartitionByIsUseOfPostIncrementedValue(const BasedUser &Val) {
985 return Val.isUseOfPostIncrementedValue;
988 /// StrengthReduceStridedIVUsers - Strength reduce all of the users of a single
989 /// stride of IV. All of the users may have different starting values, and this
990 /// may not be the only stride (we know it is if isOnlyStride is true).
991 void LoopStrengthReduce::StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
992 IVUsersOfOneStride &Uses,
995 // Transform our list of users and offsets to a bit more complex table. In
996 // this new vector, each 'BasedUser' contains 'Base' the base of the
997 // strided accessas well as the old information from Uses. We progressively
998 // move information from the Base field to the Imm field, until we eventually
999 // have the full access expression to rewrite the use.
1000 std::vector<BasedUser> UsersToProcess;
1001 UsersToProcess.reserve(Uses.Users.size());
1002 for (unsigned i = 0, e = Uses.Users.size(); i != e; ++i) {
1003 UsersToProcess.push_back(Uses.Users[i]);
1005 // Move any loop invariant operands from the offset field to the immediate
1006 // field of the use, so that we don't try to use something before it is
1008 MoveLoopVariantsToImediateField(UsersToProcess.back().Base,
1009 UsersToProcess.back().Imm, L);
1010 assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
1011 "Base value is not loop invariant!");
1014 // We now have a whole bunch of uses of like-strided induction variables, but
1015 // they might all have different bases. We want to emit one PHI node for this
1016 // stride which we fold as many common expressions (between the IVs) into as
1017 // possible. Start by identifying the common expressions in the base values
1018 // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
1019 // "A+B"), emit it to the preheader, then remove the expression from the
1020 // UsersToProcess base values.
1021 SCEVHandle CommonExprs =
1022 RemoveCommonExpressionsFromUseBases(UsersToProcess);
1024 // Next, figure out what we can represent in the immediate fields of
1025 // instructions. If we can represent anything there, move it to the imm
1026 // fields of the BasedUsers. We do this so that it increases the commonality
1027 // of the remaining uses.
1028 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1029 // If the user is not in the current loop, this means it is using the exit
1030 // value of the IV. Do not put anything in the base, make sure it's all in
1031 // the immediate field to allow as much factoring as possible.
1032 if (!L->contains(UsersToProcess[i].Inst->getParent())) {
1033 UsersToProcess[i].Imm = SCEVAddExpr::get(UsersToProcess[i].Imm,
1034 UsersToProcess[i].Base);
1035 UsersToProcess[i].Base =
1036 SCEVUnknown::getIntegerSCEV(0, UsersToProcess[i].Base->getType());
1039 // Addressing modes can be folded into loads and stores. Be careful that
1040 // the store is through the expression, not of the expression though.
1041 bool isAddress = isa<LoadInst>(UsersToProcess[i].Inst);
1042 if (StoreInst *SI = dyn_cast<StoreInst>(UsersToProcess[i].Inst))
1043 if (SI->getOperand(1) == UsersToProcess[i].OperandValToReplace)
1046 MoveImmediateValues(TLI, UsersToProcess[i].Inst, UsersToProcess[i].Base,
1047 UsersToProcess[i].Imm, isAddress, L);
1051 // Check if it is possible to reuse a IV with stride that is factor of this
1052 // stride. And the multiple is a number that can be encoded in the scale
1053 // field of the target addressing mode. And we will have a valid
1054 // instruction after this substition, including the immediate field, if any.
1055 PHINode *NewPHI = NULL;
1058 unsigned RewriteFactor = CheckForIVReuse(Stride, ReuseIV,
1059 CommonExprs->getType(),
1061 if (RewriteFactor != 0) {
1062 DOUT << "BASED ON IV of STRIDE " << *ReuseIV.Stride
1063 << " and BASE " << *ReuseIV.Base << " :\n";
1064 NewPHI = ReuseIV.PHI;
1065 IncV = ReuseIV.IncV;
1068 const Type *ReplacedTy = CommonExprs->getType();
1070 // Now that we know what we need to do, insert the PHI node itself.
1072 DOUT << "INSERTING IV of TYPE " << *ReplacedTy << " of STRIDE "
1073 << *Stride << " and BASE " << *CommonExprs << " :\n";
1075 SCEVExpander Rewriter(*SE, *LI);
1076 SCEVExpander PreheaderRewriter(*SE, *LI);
1078 BasicBlock *Preheader = L->getLoopPreheader();
1079 Instruction *PreInsertPt = Preheader->getTerminator();
1080 Instruction *PhiInsertBefore = L->getHeader()->begin();
1082 BasicBlock *LatchBlock = L->getLoopLatch();
1085 // Emit the initial base value into the loop preheader.
1087 = PreheaderRewriter.expandCodeFor(CommonExprs, PreInsertPt,
1090 if (RewriteFactor == 0) {
1091 // Create a new Phi for this base, and stick it in the loop header.
1092 NewPHI = new PHINode(ReplacedTy, "iv.", PhiInsertBefore);
1095 // Add common base to the new Phi node.
1096 NewPHI->addIncoming(CommonBaseV, Preheader);
1098 // Insert the stride into the preheader.
1099 Value *StrideV = PreheaderRewriter.expandCodeFor(Stride, PreInsertPt,
1101 if (!isa<ConstantInt>(StrideV)) ++NumVariable;
1103 // Emit the increment of the base value before the terminator of the loop
1104 // latch block, and add it to the Phi node.
1105 SCEVHandle IncExp = SCEVAddExpr::get(SCEVUnknown::get(NewPHI),
1106 SCEVUnknown::get(StrideV));
1108 IncV = Rewriter.expandCodeFor(IncExp, LatchBlock->getTerminator(),
1110 IncV->setName(NewPHI->getName()+".inc");
1111 NewPHI->addIncoming(IncV, LatchBlock);
1113 // Remember this in case a later stride is multiple of this.
1114 IVsByStride[Stride].addIV(Stride, CommonExprs, NewPHI, IncV);
1116 Constant *C = dyn_cast<Constant>(CommonBaseV);
1118 (!C->isNullValue() &&
1119 !isTargetConstant(SCEVUnknown::get(CommonBaseV), ReplacedTy, TLI)))
1120 // We want the common base emitted into the preheader! This is just
1121 // using cast as a copy so BitCast (no-op cast) is appropriate
1122 CommonBaseV = new BitCastInst(CommonBaseV, CommonBaseV->getType(),
1123 "commonbase", PreInsertPt);
1126 // We want to emit code for users inside the loop first. To do this, we
1127 // rearrange BasedUser so that the entries at the end have
1128 // isUseOfPostIncrementedValue = false, because we pop off the end of the
1129 // vector (so we handle them first).
1130 std::partition(UsersToProcess.begin(), UsersToProcess.end(),
1131 PartitionByIsUseOfPostIncrementedValue);
1133 // Sort this by base, so that things with the same base are handled
1134 // together. By partitioning first and stable-sorting later, we are
1135 // guaranteed that within each base we will pop off users from within the
1136 // loop before users outside of the loop with a particular base.
1138 // We would like to use stable_sort here, but we can't. The problem is that
1139 // SCEVHandle's don't have a deterministic ordering w.r.t to each other, so
1140 // we don't have anything to do a '<' comparison on. Because we think the
1141 // number of uses is small, do a horrible bubble sort which just relies on
1143 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1144 // Get a base value.
1145 SCEVHandle Base = UsersToProcess[i].Base;
1147 // Compact everything with this base to be consequetive with this one.
1148 for (unsigned j = i+1; j != e; ++j) {
1149 if (UsersToProcess[j].Base == Base) {
1150 std::swap(UsersToProcess[i+1], UsersToProcess[j]);
1156 // Process all the users now. This outer loop handles all bases, the inner
1157 // loop handles all users of a particular base.
1158 while (!UsersToProcess.empty()) {
1159 SCEVHandle Base = UsersToProcess.back().Base;
1161 DOUT << " INSERTING code for BASE = " << *Base << ":\n";
1163 // Emit the code for Base into the preheader.
1164 Value *BaseV = PreheaderRewriter.expandCodeFor(Base, PreInsertPt,
1167 // If BaseV is a constant other than 0, make sure that it gets inserted into
1168 // the preheader, instead of being forward substituted into the uses. We do
1169 // this by forcing a BitCast (noop cast) to be inserted into the preheader
1171 if (Constant *C = dyn_cast<Constant>(BaseV)) {
1172 if (!C->isNullValue() && !isTargetConstant(Base, ReplacedTy, TLI)) {
1173 // We want this constant emitted into the preheader! This is just
1174 // using cast as a copy so BitCast (no-op cast) is appropriate
1175 BaseV = new BitCastInst(BaseV, BaseV->getType(), "preheaderinsert",
1180 // Emit the code to add the immediate offset to the Phi value, just before
1181 // the instructions that we identified as using this stride and base.
1183 // FIXME: Use emitted users to emit other users.
1184 BasedUser &User = UsersToProcess.back();
1186 // If this instruction wants to use the post-incremented value, move it
1187 // after the post-inc and use its value instead of the PHI.
1188 Value *RewriteOp = NewPHI;
1189 if (User.isUseOfPostIncrementedValue) {
1192 // If this user is in the loop, make sure it is the last thing in the
1193 // loop to ensure it is dominated by the increment.
1194 if (L->contains(User.Inst->getParent()))
1195 User.Inst->moveBefore(LatchBlock->getTerminator());
1197 if (RewriteOp->getType() != ReplacedTy) {
1198 Instruction::CastOps opcode = Instruction::Trunc;
1199 if (ReplacedTy->getPrimitiveSizeInBits() ==
1200 RewriteOp->getType()->getPrimitiveSizeInBits())
1201 opcode = Instruction::BitCast;
1202 RewriteOp = SCEVExpander::InsertCastOfTo(opcode, RewriteOp, ReplacedTy);
1205 SCEVHandle RewriteExpr = SCEVUnknown::get(RewriteOp);
1207 // Clear the SCEVExpander's expression map so that we are guaranteed
1208 // to have the code emitted where we expect it.
1211 // If we are reusing the iv, then it must be multiplied by a constant
1212 // factor take advantage of addressing mode scale component.
1213 if (RewriteFactor != 0) {
1215 SCEVMulExpr::get(SCEVUnknown::getIntegerSCEV(RewriteFactor,
1216 RewriteExpr->getType()),
1219 // The common base is emitted in the loop preheader. But since we
1220 // are reusing an IV, it has not been used to initialize the PHI node.
1221 // Add it to the expression used to rewrite the uses.
1222 if (!isa<ConstantInt>(CommonBaseV) ||
1223 !cast<ConstantInt>(CommonBaseV)->isZero())
1224 RewriteExpr = SCEVAddExpr::get(RewriteExpr,
1225 SCEVUnknown::get(CommonBaseV));
1228 // Now that we know what we need to do, insert code before User for the
1229 // immediate and any loop-variant expressions.
1230 if (!isa<ConstantInt>(BaseV) || !cast<ConstantInt>(BaseV)->isZero())
1231 // Add BaseV to the PHI value if needed.
1232 RewriteExpr = SCEVAddExpr::get(RewriteExpr, SCEVUnknown::get(BaseV));
1234 User.RewriteInstructionToUseNewBase(RewriteExpr, Rewriter, L, this);
1236 // Mark old value we replaced as possibly dead, so that it is elminated
1237 // if we just replaced the last use of that value.
1238 DeadInsts.insert(cast<Instruction>(User.OperandValToReplace));
1240 UsersToProcess.pop_back();
1243 // If there are any more users to process with the same base, process them
1244 // now. We sorted by base above, so we just have to check the last elt.
1245 } while (!UsersToProcess.empty() && UsersToProcess.back().Base == Base);
1246 // TODO: Next, find out which base index is the most common, pull it out.
1249 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
1250 // different starting values, into different PHIs.
1253 /// FindIVForUser - If Cond has an operand that is an expression of an IV,
1254 /// set the IV user and stride information and return true, otherwise return
1256 bool LoopStrengthReduce::FindIVForUser(ICmpInst *Cond, IVStrideUse *&CondUse,
1257 const SCEVHandle *&CondStride) {
1258 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e && !CondUse;
1260 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1261 IVUsesByStride.find(StrideOrder[Stride]);
1262 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1264 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
1265 E = SI->second.Users.end(); UI != E; ++UI)
1266 if (UI->User == Cond) {
1267 // NOTE: we could handle setcc instructions with multiple uses here, but
1268 // InstCombine does it as well for simple uses, it's not clear that it
1269 // occurs enough in real life to handle.
1271 CondStride = &SI->first;
1278 // OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
1279 // uses in the loop, look to see if we can eliminate some, in favor of using
1280 // common indvars for the different uses.
1281 void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
1282 // TODO: implement optzns here.
1284 // Finally, get the terminating condition for the loop if possible. If we
1285 // can, we want to change it to use a post-incremented version of its
1286 // induction variable, to allow coalescing the live ranges for the IV into
1287 // one register value.
1288 PHINode *SomePHI = cast<PHINode>(L->getHeader()->begin());
1289 BasicBlock *Preheader = L->getLoopPreheader();
1290 BasicBlock *LatchBlock =
1291 SomePHI->getIncomingBlock(SomePHI->getIncomingBlock(0) == Preheader);
1292 BranchInst *TermBr = dyn_cast<BranchInst>(LatchBlock->getTerminator());
1293 if (!TermBr || TermBr->isUnconditional() ||
1294 !isa<ICmpInst>(TermBr->getCondition()))
1296 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1298 // Search IVUsesByStride to find Cond's IVUse if there is one.
1299 IVStrideUse *CondUse = 0;
1300 const SCEVHandle *CondStride = 0;
1302 if (!FindIVForUser(Cond, CondUse, CondStride))
1303 return; // setcc doesn't use the IV.
1306 // It's possible for the setcc instruction to be anywhere in the loop, and
1307 // possible for it to have multiple users. If it is not immediately before
1308 // the latch block branch, move it.
1309 if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
1310 if (Cond->hasOneUse()) { // Condition has a single use, just move it.
1311 Cond->moveBefore(TermBr);
1313 // Otherwise, clone the terminating condition and insert into the loopend.
1314 Cond = cast<ICmpInst>(Cond->clone());
1315 Cond->setName(L->getHeader()->getName() + ".termcond");
1316 LatchBlock->getInstList().insert(TermBr, Cond);
1318 // Clone the IVUse, as the old use still exists!
1319 IVUsesByStride[*CondStride].addUser(CondUse->Offset, Cond,
1320 CondUse->OperandValToReplace);
1321 CondUse = &IVUsesByStride[*CondStride].Users.back();
1325 // If we get to here, we know that we can transform the setcc instruction to
1326 // use the post-incremented version of the IV, allowing us to coalesce the
1327 // live ranges for the IV correctly.
1328 CondUse->Offset = SCEV::getMinusSCEV(CondUse->Offset, *CondStride);
1329 CondUse->isUseOfPostIncrementedValue = true;
1333 // Constant strides come first which in turns are sorted by their absolute
1334 // values. If absolute values are the same, then positive strides comes first.
1336 // 4, -1, X, 1, 2 ==> 1, -1, 2, 4, X
1337 struct StrideCompare {
1338 bool operator()(const SCEVHandle &LHS, const SCEVHandle &RHS) {
1339 SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS);
1340 SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS);
1342 int64_t LV = LHSC->getValue()->getSExtValue();
1343 int64_t RV = RHSC->getValue()->getSExtValue();
1344 uint64_t ALV = (LV < 0) ? -LV : LV;
1345 uint64_t ARV = (RV < 0) ? -RV : RV;
1351 return (LHSC && !RHSC);
1356 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager &LPM) {
1358 LI = &getAnalysis<LoopInfo>();
1359 EF = &getAnalysis<ETForest>();
1360 SE = &getAnalysis<ScalarEvolution>();
1361 TD = &getAnalysis<TargetData>();
1362 UIntPtrTy = TD->getIntPtrType();
1364 // Find all uses of induction variables in this loop, and catagorize
1365 // them by stride. Start by finding all of the PHI nodes in the header for
1366 // this loop. If they are induction variables, inspect their uses.
1367 std::set<Instruction*> Processed; // Don't reprocess instructions.
1368 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I)
1369 AddUsersIfInteresting(I, L, Processed);
1371 // If we have nothing to do, return.
1372 if (IVUsesByStride.empty()) return false;
1374 // Optimize induction variables. Some indvar uses can be transformed to use
1375 // strides that will be needed for other purposes. A common example of this
1376 // is the exit test for the loop, which can often be rewritten to use the
1377 // computation of some other indvar to decide when to terminate the loop.
1381 // FIXME: We can widen subreg IV's here for RISC targets. e.g. instead of
1382 // doing computation in byte values, promote to 32-bit values if safe.
1384 // FIXME: Attempt to reuse values across multiple IV's. In particular, we
1385 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should be
1386 // codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC. Need
1387 // to be careful that IV's are all the same type. Only works for intptr_t
1390 // If we only have one stride, we can more aggressively eliminate some things.
1391 bool HasOneStride = IVUsesByStride.size() == 1;
1394 DOUT << "\nLSR on ";
1398 // IVsByStride keeps IVs for one particular loop.
1399 IVsByStride.clear();
1401 // Sort the StrideOrder so we process larger strides first.
1402 std::stable_sort(StrideOrder.begin(), StrideOrder.end(), StrideCompare());
1404 // Note: this processes each stride/type pair individually. All users passed
1405 // into StrengthReduceStridedIVUsers have the same type AND stride. Also,
1406 // node that we iterate over IVUsesByStride indirectly by using StrideOrder.
1407 // This extra layer of indirection makes the ordering of strides deterministic
1408 // - not dependent on map order.
1409 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e; ++Stride) {
1410 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1411 IVUsesByStride.find(StrideOrder[Stride]);
1412 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1413 StrengthReduceStridedIVUsers(SI->first, SI->second, L, HasOneStride);
1416 // Clean up after ourselves
1417 if (!DeadInsts.empty()) {
1418 DeleteTriviallyDeadInstructions(DeadInsts);
1420 BasicBlock::iterator I = L->getHeader()->begin();
1422 while ((PN = dyn_cast<PHINode>(I))) {
1423 ++I; // Preincrement iterator to avoid invalidating it when deleting PN.
1425 // At this point, we know that we have killed one or more GEP
1426 // instructions. It is worth checking to see if the cann indvar is also
1427 // dead, so that we can remove it as well. The requirements for the cann
1428 // indvar to be considered dead are:
1429 // 1. the cann indvar has one use
1430 // 2. the use is an add instruction
1431 // 3. the add has one use
1432 // 4. the add is used by the cann indvar
1433 // If all four cases above are true, then we can remove both the add and
1435 // FIXME: this needs to eliminate an induction variable even if it's being
1436 // compared against some value to decide loop termination.
1437 if (PN->hasOneUse()) {
1438 Instruction *BO = dyn_cast<Instruction>(*PN->use_begin());
1439 if (BO && (isa<BinaryOperator>(BO) || isa<CmpInst>(BO))) {
1440 if (BO->hasOneUse() && PN == *(BO->use_begin())) {
1441 DeadInsts.insert(BO);
1442 // Break the cycle, then delete the PHI.
1443 PN->replaceAllUsesWith(UndefValue::get(PN->getType()));
1444 SE->deleteInstructionFromRecords(PN);
1445 PN->eraseFromParent();
1450 DeleteTriviallyDeadInstructions(DeadInsts);
1453 CastedPointers.clear();
1454 IVUsesByStride.clear();
1455 StrideOrder.clear();