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/IntrinsicInst.h"
23 #include "llvm/Type.h"
24 #include "llvm/DerivedTypes.h"
25 #include "llvm/Analysis/Dominators.h"
26 #include "llvm/Analysis/LoopInfo.h"
27 #include "llvm/Analysis/LoopPass.h"
28 #include "llvm/Analysis/ScalarEvolutionExpander.h"
29 #include "llvm/Support/CFG.h"
30 #include "llvm/Support/GetElementPtrTypeIterator.h"
31 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
32 #include "llvm/Transforms/Utils/Local.h"
33 #include "llvm/Target/TargetData.h"
34 #include "llvm/ADT/Statistic.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/Compiler.h"
37 #include "llvm/Target/TargetLowering.h"
42 STATISTIC(NumReduced , "Number of GEPs strength reduced");
43 STATISTIC(NumInserted, "Number of PHIs inserted");
44 STATISTIC(NumVariable, "Number of PHIs with variable strides");
50 /// IVStrideUse - Keep track of one use of a strided induction variable, where
51 /// the stride is stored externally. The Offset member keeps track of the
52 /// offset from the IV, User is the actual user of the operand, and 'Operand'
53 /// is the operand # of the User that is the use.
54 struct VISIBILITY_HIDDEN IVStrideUse {
57 Value *OperandValToReplace;
59 // isUseOfPostIncrementedValue - True if this should use the
60 // post-incremented version of this IV, not the preincremented version.
61 // This can only be set in special cases, such as the terminating setcc
62 // instruction for a loop or uses dominated by the loop.
63 bool isUseOfPostIncrementedValue;
65 IVStrideUse(const SCEVHandle &Offs, Instruction *U, Value *O)
66 : Offset(Offs), User(U), OperandValToReplace(O),
67 isUseOfPostIncrementedValue(false) {}
70 /// IVUsersOfOneStride - This structure keeps track of all instructions that
71 /// have an operand that is based on the trip count multiplied by some stride.
72 /// The stride for all of these users is common and kept external to this
74 struct VISIBILITY_HIDDEN IVUsersOfOneStride {
75 /// Users - Keep track of all of the users of this stride as well as the
76 /// initial value and the operand that uses the IV.
77 std::vector<IVStrideUse> Users;
79 void addUser(const SCEVHandle &Offset,Instruction *User, Value *Operand) {
80 Users.push_back(IVStrideUse(Offset, User, Operand));
84 /// IVInfo - This structure keeps track of one IV expression inserted during
85 /// StrengthReduceStridedIVUsers. It contains the stride, the common base, as
86 /// well as the PHI node and increment value created for rewrite.
87 struct VISIBILITY_HIDDEN IVExpr {
94 : Stride(SCEVUnknown::getIntegerSCEV(0, Type::Int32Ty)),
95 Base (SCEVUnknown::getIntegerSCEV(0, Type::Int32Ty)) {}
96 IVExpr(const SCEVHandle &stride, const SCEVHandle &base, PHINode *phi,
98 : Stride(stride), Base(base), PHI(phi), IncV(incv) {}
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,
108 IVs.push_back(IVExpr(Stride, Base, PHI, IncV));
112 class VISIBILITY_HIDDEN LoopStrengthReduce : public LoopPass {
116 const TargetData *TD;
117 const Type *UIntPtrTy;
120 /// IVUsesByStride - Keep track of all uses of induction variables that we
121 /// are interested in. The key of the map is the stride of the access.
122 std::map<SCEVHandle, IVUsersOfOneStride> IVUsesByStride;
124 /// IVsByStride - Keep track of all IVs that have been inserted for a
125 /// particular stride.
126 std::map<SCEVHandle, IVsOfOneStride> IVsByStride;
128 /// StrideOrder - An ordering of the keys in IVUsesByStride that is stable:
129 /// We use this to iterate over the IVUsesByStride collection without being
130 /// dependent on random ordering of pointers in the process.
131 std::vector<SCEVHandle> StrideOrder;
133 /// CastedValues - As we need to cast values to uintptr_t, this keeps track
134 /// of the casted version of each value. This is accessed by
135 /// getCastedVersionOf.
136 std::map<Value*, Value*> CastedPointers;
138 /// DeadInsts - Keep track of instructions we may have made dead, so that
139 /// we can remove them after we are done working.
140 std::set<Instruction*> DeadInsts;
142 /// TLI - Keep a pointer of a TargetLowering to consult for determining
143 /// transformation profitability.
144 const TargetLowering *TLI;
147 static char ID; // Pass ID, replacement for typeid
148 explicit LoopStrengthReduce(const TargetLowering *tli = NULL) :
149 LoopPass((intptr_t)&ID), TLI(tli) {
152 bool runOnLoop(Loop *L, LPPassManager &LPM);
154 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
155 // We split critical edges, so we change the CFG. However, we do update
156 // many analyses if they are around.
157 AU.addPreservedID(LoopSimplifyID);
158 AU.addPreserved<LoopInfo>();
159 AU.addPreserved<DominanceFrontier>();
160 AU.addPreserved<DominatorTree>();
162 AU.addRequiredID(LoopSimplifyID);
163 AU.addRequired<LoopInfo>();
164 AU.addRequired<DominatorTree>();
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->deleteValueFromRecords(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->getABITypeSize(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, DominatorTree *DT, 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 (DT->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 (!DT->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, DT, 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);
560 Value *Base = Rewriter.expandCodeFor(NewBase, BaseInsertPt);
562 // If we are inserting the base and imm values in the same block, make sure to
563 // adjust the IP position if insertion reused a result.
564 if (IP == BaseInsertPt)
565 IP = Rewriter.getInsertionPoint();
567 // Always emit the immediate (if non-zero) into the same block as the user.
568 SCEVHandle NewValSCEV = SCEVAddExpr::get(SCEVUnknown::get(Base), Imm);
569 return Rewriter.expandCodeFor(NewValSCEV, IP);
574 // Once we rewrite the code to insert the new IVs we want, update the
575 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
577 void BasedUser::RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
578 SCEVExpander &Rewriter,
580 if (!isa<PHINode>(Inst)) {
581 // By default, insert code at the user instruction.
582 BasicBlock::iterator InsertPt = Inst;
584 // However, if the Operand is itself an instruction, the (potentially
585 // complex) inserted code may be shared by many users. Because of this, we
586 // want to emit code for the computation of the operand right before its old
587 // computation. This is usually safe, because we obviously used to use the
588 // computation when it was computed in its current block. However, in some
589 // cases (e.g. use of a post-incremented induction variable) the NewBase
590 // value will be pinned to live somewhere after the original computation.
591 // In this case, we have to back off.
592 if (!isUseOfPostIncrementedValue) {
593 if (Instruction *OpInst = dyn_cast<Instruction>(OperandValToReplace)) {
595 while (isa<PHINode>(InsertPt)) ++InsertPt;
598 Value *NewVal = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
599 // Adjust the type back to match the Inst. Note that we can't use InsertPt
600 // here because the SCEVExpander may have inserted the instructions after
601 // that point, in its efforts to avoid inserting redundant expressions.
602 if (isa<PointerType>(OperandValToReplace->getType())) {
603 NewVal = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr,
605 OperandValToReplace->getType());
607 // Replace the use of the operand Value with the new Phi we just created.
608 Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
609 DOUT << " CHANGED: IMM =" << *Imm;
610 DOUT << " \tNEWBASE =" << *NewBase;
611 DOUT << " \tInst = " << *Inst;
615 // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm
616 // expression into each operand block that uses it. Note that PHI nodes can
617 // have multiple entries for the same predecessor. We use a map to make sure
618 // that a PHI node only has a single Value* for each predecessor (which also
619 // prevents us from inserting duplicate code in some blocks).
620 std::map<BasicBlock*, Value*> InsertedCode;
621 PHINode *PN = cast<PHINode>(Inst);
622 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
623 if (PN->getIncomingValue(i) == OperandValToReplace) {
624 // If this is a critical edge, split the edge so that we do not insert the
625 // code on all predecessor/successor paths. We do this unless this is the
626 // canonical backedge for this loop, as this can make some inserted code
627 // be in an illegal position.
628 BasicBlock *PHIPred = PN->getIncomingBlock(i);
629 if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 &&
630 (PN->getParent() != L->getHeader() || !L->contains(PHIPred))) {
632 // First step, split the critical edge.
633 SplitCriticalEdge(PHIPred, PN->getParent(), P, true);
635 // Next step: move the basic block. In particular, if the PHI node
636 // is outside of the loop, and PredTI is in the loop, we want to
637 // move the block to be immediately before the PHI block, not
638 // immediately after PredTI.
639 if (L->contains(PHIPred) && !L->contains(PN->getParent())) {
640 BasicBlock *NewBB = PN->getIncomingBlock(i);
641 NewBB->moveBefore(PN->getParent());
644 // Splitting the edge can reduce the number of PHI entries we have.
645 e = PN->getNumIncomingValues();
648 Value *&Code = InsertedCode[PN->getIncomingBlock(i)];
650 // Insert the code into the end of the predecessor block.
651 Instruction *InsertPt = PN->getIncomingBlock(i)->getTerminator();
652 Code = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
654 // Adjust the type back to match the PHI. Note that we can't use
655 // InsertPt here because the SCEVExpander may have inserted its
656 // instructions after that point, in its efforts to avoid inserting
657 // redundant expressions.
658 if (isa<PointerType>(PN->getType())) {
659 Code = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr,
665 // Replace the use of the operand Value with the new Phi we just created.
666 PN->setIncomingValue(i, Code);
670 DOUT << " CHANGED: IMM =" << *Imm << " Inst = " << *Inst;
674 /// isTargetConstant - Return true if the following can be referenced by the
675 /// immediate field of a target instruction.
676 static bool isTargetConstant(const SCEVHandle &V, const Type *UseTy,
677 const TargetLowering *TLI) {
678 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
679 int64_t VC = SC->getValue()->getSExtValue();
681 TargetLowering::AddrMode AM;
683 return TLI->isLegalAddressingMode(AM, UseTy);
685 // Defaults to PPC. PPC allows a sign-extended 16-bit immediate field.
686 return (VC > -(1 << 16) && VC < (1 << 16)-1);
690 if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
691 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(SU->getValue()))
692 if (TLI && CE->getOpcode() == Instruction::PtrToInt) {
693 Constant *Op0 = CE->getOperand(0);
694 if (GlobalValue *GV = dyn_cast<GlobalValue>(Op0)) {
695 TargetLowering::AddrMode AM;
697 return TLI->isLegalAddressingMode(AM, UseTy);
703 /// MoveLoopVariantsToImediateField - Move any subexpressions from Val that are
704 /// loop varying to the Imm operand.
705 static void MoveLoopVariantsToImediateField(SCEVHandle &Val, SCEVHandle &Imm,
707 if (Val->isLoopInvariant(L)) return; // Nothing to do.
709 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
710 std::vector<SCEVHandle> NewOps;
711 NewOps.reserve(SAE->getNumOperands());
713 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
714 if (!SAE->getOperand(i)->isLoopInvariant(L)) {
715 // If this is a loop-variant expression, it must stay in the immediate
716 // field of the expression.
717 Imm = SCEVAddExpr::get(Imm, SAE->getOperand(i));
719 NewOps.push_back(SAE->getOperand(i));
723 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
725 Val = SCEVAddExpr::get(NewOps);
726 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
727 // Try to pull immediates out of the start value of nested addrec's.
728 SCEVHandle Start = SARE->getStart();
729 MoveLoopVariantsToImediateField(Start, Imm, L);
731 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
733 Val = SCEVAddRecExpr::get(Ops, SARE->getLoop());
735 // Otherwise, all of Val is variant, move the whole thing over.
736 Imm = SCEVAddExpr::get(Imm, Val);
737 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
742 /// MoveImmediateValues - Look at Val, and pull out any additions of constants
743 /// that can fit into the immediate field of instructions in the target.
744 /// Accumulate these immediate values into the Imm value.
745 static void MoveImmediateValues(const TargetLowering *TLI,
747 SCEVHandle &Val, SCEVHandle &Imm,
748 bool isAddress, Loop *L) {
749 const Type *UseTy = User->getType();
750 if (StoreInst *SI = dyn_cast<StoreInst>(User))
751 UseTy = SI->getOperand(0)->getType();
753 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
754 std::vector<SCEVHandle> NewOps;
755 NewOps.reserve(SAE->getNumOperands());
757 for (unsigned i = 0; i != SAE->getNumOperands(); ++i) {
758 SCEVHandle NewOp = SAE->getOperand(i);
759 MoveImmediateValues(TLI, User, NewOp, Imm, isAddress, L);
761 if (!NewOp->isLoopInvariant(L)) {
762 // If this is a loop-variant expression, it must stay in the immediate
763 // field of the expression.
764 Imm = SCEVAddExpr::get(Imm, NewOp);
766 NewOps.push_back(NewOp);
771 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
773 Val = SCEVAddExpr::get(NewOps);
775 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
776 // Try to pull immediates out of the start value of nested addrec's.
777 SCEVHandle Start = SARE->getStart();
778 MoveImmediateValues(TLI, User, Start, Imm, isAddress, L);
780 if (Start != SARE->getStart()) {
781 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
783 Val = SCEVAddRecExpr::get(Ops, SARE->getLoop());
786 } else if (SCEVMulExpr *SME = dyn_cast<SCEVMulExpr>(Val)) {
787 // Transform "8 * (4 + v)" -> "32 + 8*V" if "32" fits in the immed field.
788 if (isAddress && isTargetConstant(SME->getOperand(0), UseTy, TLI) &&
789 SME->getNumOperands() == 2 && SME->isLoopInvariant(L)) {
791 SCEVHandle SubImm = SCEVUnknown::getIntegerSCEV(0, Val->getType());
792 SCEVHandle NewOp = SME->getOperand(1);
793 MoveImmediateValues(TLI, User, NewOp, SubImm, isAddress, L);
795 // If we extracted something out of the subexpressions, see if we can
797 if (NewOp != SME->getOperand(1)) {
798 // Scale SubImm up by "8". If the result is a target constant, we are
800 SubImm = SCEVMulExpr::get(SubImm, SME->getOperand(0));
801 if (isTargetConstant(SubImm, UseTy, TLI)) {
802 // Accumulate the immediate.
803 Imm = SCEVAddExpr::get(Imm, SubImm);
805 // Update what is left of 'Val'.
806 Val = SCEVMulExpr::get(SME->getOperand(0), NewOp);
813 // Loop-variant expressions must stay in the immediate field of the
815 if ((isAddress && isTargetConstant(Val, UseTy, TLI)) ||
816 !Val->isLoopInvariant(L)) {
817 Imm = SCEVAddExpr::get(Imm, Val);
818 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
822 // Otherwise, no immediates to move.
826 /// SeparateSubExprs - Decompose Expr into all of the subexpressions that are
827 /// added together. This is used to reassociate common addition subexprs
828 /// together for maximal sharing when rewriting bases.
829 static void SeparateSubExprs(std::vector<SCEVHandle> &SubExprs,
831 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) {
832 for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
833 SeparateSubExprs(SubExprs, AE->getOperand(j));
834 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) {
835 SCEVHandle Zero = SCEVUnknown::getIntegerSCEV(0, Expr->getType());
836 if (SARE->getOperand(0) == Zero) {
837 SubExprs.push_back(Expr);
839 // Compute the addrec with zero as its base.
840 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
841 Ops[0] = Zero; // Start with zero base.
842 SubExprs.push_back(SCEVAddRecExpr::get(Ops, SARE->getLoop()));
845 SeparateSubExprs(SubExprs, SARE->getOperand(0));
847 } else if (!isa<SCEVConstant>(Expr) ||
848 !cast<SCEVConstant>(Expr)->getValue()->isZero()) {
850 SubExprs.push_back(Expr);
855 /// RemoveCommonExpressionsFromUseBases - Look through all of the uses in Bases,
856 /// removing any common subexpressions from it. Anything truly common is
857 /// removed, accumulated, and returned. This looks for things like (a+b+c) and
858 /// (a+c+d) -> (a+c). The common expression is *removed* from the Bases.
860 RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses) {
861 unsigned NumUses = Uses.size();
863 // Only one use? Use its base, regardless of what it is!
864 SCEVHandle Zero = SCEVUnknown::getIntegerSCEV(0, Uses[0].Base->getType());
865 SCEVHandle Result = Zero;
867 std::swap(Result, Uses[0].Base);
871 // To find common subexpressions, count how many of Uses use each expression.
872 // If any subexpressions are used Uses.size() times, they are common.
873 std::map<SCEVHandle, unsigned> SubExpressionUseCounts;
875 // UniqueSubExprs - Keep track of all of the subexpressions we see in the
876 // order we see them.
877 std::vector<SCEVHandle> UniqueSubExprs;
879 std::vector<SCEVHandle> SubExprs;
880 for (unsigned i = 0; i != NumUses; ++i) {
881 // If the base is zero (which is common), return zero now, there are no
883 if (Uses[i].Base == Zero) return Zero;
885 // Split the expression into subexprs.
886 SeparateSubExprs(SubExprs, Uses[i].Base);
887 // Add one to SubExpressionUseCounts for each subexpr present.
888 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
889 if (++SubExpressionUseCounts[SubExprs[j]] == 1)
890 UniqueSubExprs.push_back(SubExprs[j]);
894 // Now that we know how many times each is used, build Result. Iterate over
895 // UniqueSubexprs so that we have a stable ordering.
896 for (unsigned i = 0, e = UniqueSubExprs.size(); i != e; ++i) {
897 std::map<SCEVHandle, unsigned>::iterator I =
898 SubExpressionUseCounts.find(UniqueSubExprs[i]);
899 assert(I != SubExpressionUseCounts.end() && "Entry not found?");
900 if (I->second == NumUses) { // Found CSE!
901 Result = SCEVAddExpr::get(Result, I->first);
903 // Remove non-cse's from SubExpressionUseCounts.
904 SubExpressionUseCounts.erase(I);
908 // If we found no CSE's, return now.
909 if (Result == Zero) return Result;
911 // Otherwise, remove all of the CSE's we found from each of the base values.
912 for (unsigned i = 0; i != NumUses; ++i) {
913 // Split the expression into subexprs.
914 SeparateSubExprs(SubExprs, Uses[i].Base);
916 // Remove any common subexpressions.
917 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
918 if (SubExpressionUseCounts.count(SubExprs[j])) {
919 SubExprs.erase(SubExprs.begin()+j);
923 // Finally, the non-shared expressions together.
924 if (SubExprs.empty())
927 Uses[i].Base = SCEVAddExpr::get(SubExprs);
934 /// isZero - returns true if the scalar evolution expression is zero.
936 static bool isZero(SCEVHandle &V) {
937 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(V))
938 return SC->getValue()->isZero();
942 /// ValidStride - Check whether the given Scale is valid for all loads and
943 /// stores in UsersToProcess.
945 bool LoopStrengthReduce::ValidStride(int64_t Scale,
946 const std::vector<BasedUser>& UsersToProcess) {
947 for (unsigned i=0, e = UsersToProcess.size(); i!=e; ++i) {
948 // If this is a load or other access, pass the type of the access in.
949 const Type *AccessTy = Type::VoidTy;
950 if (StoreInst *SI = dyn_cast<StoreInst>(UsersToProcess[i].Inst))
951 AccessTy = SI->getOperand(0)->getType();
952 else if (LoadInst *LI = dyn_cast<LoadInst>(UsersToProcess[i].Inst))
953 AccessTy = LI->getType();
955 TargetLowering::AddrMode AM;
956 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
957 AM.BaseOffs = SC->getValue()->getSExtValue();
960 // If load[imm+r*scale] is illegal, bail out.
961 if (!TLI->isLegalAddressingMode(AM, AccessTy))
967 /// CheckForIVReuse - Returns the multiple if the stride is the multiple
968 /// of a previous stride and it is a legal value for the target addressing
969 /// mode scale component. This allows the users of this stride to be rewritten
970 /// as prev iv * factor. It returns 0 if no reuse is possible.
971 unsigned LoopStrengthReduce::CheckForIVReuse(const SCEVHandle &Stride,
972 IVExpr &IV, const Type *Ty,
973 const std::vector<BasedUser>& UsersToProcess) {
976 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Stride)) {
977 int64_t SInt = SC->getValue()->getSExtValue();
978 if (SInt == 1) return 0;
980 for (std::map<SCEVHandle, IVsOfOneStride>::iterator SI= IVsByStride.begin(),
981 SE = IVsByStride.end(); SI != SE; ++SI) {
982 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
983 if (SInt != -SSInt &&
984 (unsigned(abs(SInt)) < SSInt || (SInt % SSInt) != 0))
986 int64_t Scale = SInt / SSInt;
987 // Check that this stride is valid for all the types used for loads and
988 // stores; if it can be used for some and not others, we might as well use
989 // the original stride everywhere, since we have to create the IV for it
991 if (ValidStride(Scale, UsersToProcess))
992 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
993 IE = SI->second.IVs.end(); II != IE; ++II)
994 // FIXME: Only handle base == 0 for now.
995 // Only reuse previous IV if it would not require a type conversion.
996 if (isZero(II->Base) && II->Base->getType() == Ty) {
1005 /// PartitionByIsUseOfPostIncrementedValue - Simple boolean predicate that
1006 /// returns true if Val's isUseOfPostIncrementedValue is true.
1007 static bool PartitionByIsUseOfPostIncrementedValue(const BasedUser &Val) {
1008 return Val.isUseOfPostIncrementedValue;
1011 /// isNonConstantNegative - REturn true if the specified scev is negated, but
1013 static bool isNonConstantNegative(const SCEVHandle &Expr) {
1014 SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Expr);
1015 if (!Mul) return false;
1017 // If there is a constant factor, it will be first.
1018 SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
1019 if (!SC) return false;
1021 // Return true if the value is negative, this matches things like (-42 * V).
1022 return SC->getValue()->getValue().isNegative();
1025 /// StrengthReduceStridedIVUsers - Strength reduce all of the users of a single
1026 /// stride of IV. All of the users may have different starting values, and this
1027 /// may not be the only stride (we know it is if isOnlyStride is true).
1028 void LoopStrengthReduce::StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
1029 IVUsersOfOneStride &Uses,
1031 bool isOnlyStride) {
1032 // Transform our list of users and offsets to a bit more complex table. In
1033 // this new vector, each 'BasedUser' contains 'Base' the base of the
1034 // strided accessas well as the old information from Uses. We progressively
1035 // move information from the Base field to the Imm field, until we eventually
1036 // have the full access expression to rewrite the use.
1037 std::vector<BasedUser> UsersToProcess;
1038 UsersToProcess.reserve(Uses.Users.size());
1039 for (unsigned i = 0, e = Uses.Users.size(); i != e; ++i) {
1040 UsersToProcess.push_back(Uses.Users[i]);
1042 // Move any loop invariant operands from the offset field to the immediate
1043 // field of the use, so that we don't try to use something before it is
1045 MoveLoopVariantsToImediateField(UsersToProcess.back().Base,
1046 UsersToProcess.back().Imm, L);
1047 assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
1048 "Base value is not loop invariant!");
1051 // We now have a whole bunch of uses of like-strided induction variables, but
1052 // they might all have different bases. We want to emit one PHI node for this
1053 // stride which we fold as many common expressions (between the IVs) into as
1054 // possible. Start by identifying the common expressions in the base values
1055 // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
1056 // "A+B"), emit it to the preheader, then remove the expression from the
1057 // UsersToProcess base values.
1058 SCEVHandle CommonExprs =
1059 RemoveCommonExpressionsFromUseBases(UsersToProcess);
1061 // Next, figure out what we can represent in the immediate fields of
1062 // instructions. If we can represent anything there, move it to the imm
1063 // fields of the BasedUsers. We do this so that it increases the commonality
1064 // of the remaining uses.
1065 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1066 // If the user is not in the current loop, this means it is using the exit
1067 // value of the IV. Do not put anything in the base, make sure it's all in
1068 // the immediate field to allow as much factoring as possible.
1069 if (!L->contains(UsersToProcess[i].Inst->getParent())) {
1070 UsersToProcess[i].Imm = SCEVAddExpr::get(UsersToProcess[i].Imm,
1071 UsersToProcess[i].Base);
1072 UsersToProcess[i].Base =
1073 SCEVUnknown::getIntegerSCEV(0, UsersToProcess[i].Base->getType());
1076 // Addressing modes can be folded into loads and stores. Be careful that
1077 // the store is through the expression, not of the expression though.
1078 bool isAddress = isa<LoadInst>(UsersToProcess[i].Inst);
1079 if (StoreInst *SI = dyn_cast<StoreInst>(UsersToProcess[i].Inst)) {
1080 if (SI->getOperand(1) == UsersToProcess[i].OperandValToReplace)
1082 } else if (IntrinsicInst *II =
1083 dyn_cast<IntrinsicInst>(UsersToProcess[i].Inst)) {
1084 // Addressing modes can also be folded into prefetches.
1085 if (II->getIntrinsicID() == Intrinsic::prefetch &&
1086 II->getOperand(1) == UsersToProcess[i].OperandValToReplace)
1090 MoveImmediateValues(TLI, UsersToProcess[i].Inst, UsersToProcess[i].Base,
1091 UsersToProcess[i].Imm, isAddress, L);
1095 // Check if it is possible to reuse a IV with stride that is factor of this
1096 // stride. And the multiple is a number that can be encoded in the scale
1097 // field of the target addressing mode. And we will have a valid
1098 // instruction after this substition, including the immediate field, if any.
1099 PHINode *NewPHI = NULL;
1102 unsigned RewriteFactor = CheckForIVReuse(Stride, ReuseIV,
1103 CommonExprs->getType(),
1105 if (RewriteFactor != 0) {
1106 DOUT << "BASED ON IV of STRIDE " << *ReuseIV.Stride
1107 << " and BASE " << *ReuseIV.Base << " :\n";
1108 NewPHI = ReuseIV.PHI;
1109 IncV = ReuseIV.IncV;
1112 const Type *ReplacedTy = CommonExprs->getType();
1114 // Now that we know what we need to do, insert the PHI node itself.
1116 DOUT << "INSERTING IV of TYPE " << *ReplacedTy << " of STRIDE "
1117 << *Stride << " and BASE " << *CommonExprs << ": ";
1119 SCEVExpander Rewriter(*SE, *LI);
1120 SCEVExpander PreheaderRewriter(*SE, *LI);
1122 BasicBlock *Preheader = L->getLoopPreheader();
1123 Instruction *PreInsertPt = Preheader->getTerminator();
1124 Instruction *PhiInsertBefore = L->getHeader()->begin();
1126 BasicBlock *LatchBlock = L->getLoopLatch();
1129 // Emit the initial base value into the loop preheader.
1131 = PreheaderRewriter.expandCodeFor(CommonExprs, PreInsertPt);
1133 if (RewriteFactor == 0) {
1134 // Create a new Phi for this base, and stick it in the loop header.
1135 NewPHI = new PHINode(ReplacedTy, "iv.", PhiInsertBefore);
1138 // Add common base to the new Phi node.
1139 NewPHI->addIncoming(CommonBaseV, Preheader);
1141 // If the stride is negative, insert a sub instead of an add for the
1143 bool isNegative = isNonConstantNegative(Stride);
1144 SCEVHandle IncAmount = Stride;
1146 IncAmount = SCEV::getNegativeSCEV(Stride);
1148 // Insert the stride into the preheader.
1149 Value *StrideV = PreheaderRewriter.expandCodeFor(IncAmount, PreInsertPt);
1150 if (!isa<ConstantInt>(StrideV)) ++NumVariable;
1152 // Emit the increment of the base value before the terminator of the loop
1153 // latch block, and add it to the Phi node.
1154 SCEVHandle IncExp = SCEVUnknown::get(StrideV);
1156 IncExp = SCEV::getNegativeSCEV(IncExp);
1157 IncExp = SCEVAddExpr::get(SCEVUnknown::get(NewPHI), IncExp);
1159 IncV = Rewriter.expandCodeFor(IncExp, LatchBlock->getTerminator());
1160 IncV->setName(NewPHI->getName()+".inc");
1161 NewPHI->addIncoming(IncV, LatchBlock);
1163 // Remember this in case a later stride is multiple of this.
1164 IVsByStride[Stride].addIV(Stride, CommonExprs, NewPHI, IncV);
1166 DOUT << " IV=%" << NewPHI->getNameStr() << " INC=%" << IncV->getNameStr();
1168 Constant *C = dyn_cast<Constant>(CommonBaseV);
1170 (!C->isNullValue() &&
1171 !isTargetConstant(SCEVUnknown::get(CommonBaseV), ReplacedTy, TLI)))
1172 // We want the common base emitted into the preheader! This is just
1173 // using cast as a copy so BitCast (no-op cast) is appropriate
1174 CommonBaseV = new BitCastInst(CommonBaseV, CommonBaseV->getType(),
1175 "commonbase", PreInsertPt);
1179 // We want to emit code for users inside the loop first. To do this, we
1180 // rearrange BasedUser so that the entries at the end have
1181 // isUseOfPostIncrementedValue = false, because we pop off the end of the
1182 // vector (so we handle them first).
1183 std::partition(UsersToProcess.begin(), UsersToProcess.end(),
1184 PartitionByIsUseOfPostIncrementedValue);
1186 // Sort this by base, so that things with the same base are handled
1187 // together. By partitioning first and stable-sorting later, we are
1188 // guaranteed that within each base we will pop off users from within the
1189 // loop before users outside of the loop with a particular base.
1191 // We would like to use stable_sort here, but we can't. The problem is that
1192 // SCEVHandle's don't have a deterministic ordering w.r.t to each other, so
1193 // we don't have anything to do a '<' comparison on. Because we think the
1194 // number of uses is small, do a horrible bubble sort which just relies on
1196 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1197 // Get a base value.
1198 SCEVHandle Base = UsersToProcess[i].Base;
1200 // Compact everything with this base to be consequetive with this one.
1201 for (unsigned j = i+1; j != e; ++j) {
1202 if (UsersToProcess[j].Base == Base) {
1203 std::swap(UsersToProcess[i+1], UsersToProcess[j]);
1209 // Process all the users now. This outer loop handles all bases, the inner
1210 // loop handles all users of a particular base.
1211 while (!UsersToProcess.empty()) {
1212 SCEVHandle Base = UsersToProcess.back().Base;
1214 // Emit the code for Base into the preheader.
1215 Value *BaseV = PreheaderRewriter.expandCodeFor(Base, PreInsertPt);
1217 DOUT << " INSERTING code for BASE = " << *Base << ":";
1218 if (BaseV->hasName())
1219 DOUT << " Result value name = %" << BaseV->getNameStr();
1222 // If BaseV is a constant other than 0, make sure that it gets inserted into
1223 // the preheader, instead of being forward substituted into the uses. We do
1224 // this by forcing a BitCast (noop cast) to be inserted into the preheader
1226 if (Constant *C = dyn_cast<Constant>(BaseV)) {
1227 if (!C->isNullValue() && !isTargetConstant(Base, ReplacedTy, TLI)) {
1228 // We want this constant emitted into the preheader! This is just
1229 // using cast as a copy so BitCast (no-op cast) is appropriate
1230 BaseV = new BitCastInst(BaseV, BaseV->getType(), "preheaderinsert",
1235 // Emit the code to add the immediate offset to the Phi value, just before
1236 // the instructions that we identified as using this stride and base.
1238 // FIXME: Use emitted users to emit other users.
1239 BasedUser &User = UsersToProcess.back();
1241 // If this instruction wants to use the post-incremented value, move it
1242 // after the post-inc and use its value instead of the PHI.
1243 Value *RewriteOp = NewPHI;
1244 if (User.isUseOfPostIncrementedValue) {
1247 // If this user is in the loop, make sure it is the last thing in the
1248 // loop to ensure it is dominated by the increment.
1249 if (L->contains(User.Inst->getParent()))
1250 User.Inst->moveBefore(LatchBlock->getTerminator());
1252 if (RewriteOp->getType() != ReplacedTy) {
1253 Instruction::CastOps opcode = Instruction::Trunc;
1254 if (ReplacedTy->getPrimitiveSizeInBits() ==
1255 RewriteOp->getType()->getPrimitiveSizeInBits())
1256 opcode = Instruction::BitCast;
1257 RewriteOp = SCEVExpander::InsertCastOfTo(opcode, RewriteOp, ReplacedTy);
1260 SCEVHandle RewriteExpr = SCEVUnknown::get(RewriteOp);
1262 // Clear the SCEVExpander's expression map so that we are guaranteed
1263 // to have the code emitted where we expect it.
1266 // If we are reusing the iv, then it must be multiplied by a constant
1267 // factor take advantage of addressing mode scale component.
1268 if (RewriteFactor != 0) {
1270 SCEVMulExpr::get(SCEVUnknown::getIntegerSCEV(RewriteFactor,
1271 RewriteExpr->getType()),
1274 // The common base is emitted in the loop preheader. But since we
1275 // are reusing an IV, it has not been used to initialize the PHI node.
1276 // Add it to the expression used to rewrite the uses.
1277 if (!isa<ConstantInt>(CommonBaseV) ||
1278 !cast<ConstantInt>(CommonBaseV)->isZero())
1279 RewriteExpr = SCEVAddExpr::get(RewriteExpr,
1280 SCEVUnknown::get(CommonBaseV));
1283 // Now that we know what we need to do, insert code before User for the
1284 // immediate and any loop-variant expressions.
1285 if (!isa<ConstantInt>(BaseV) || !cast<ConstantInt>(BaseV)->isZero())
1286 // Add BaseV to the PHI value if needed.
1287 RewriteExpr = SCEVAddExpr::get(RewriteExpr, SCEVUnknown::get(BaseV));
1289 User.RewriteInstructionToUseNewBase(RewriteExpr, Rewriter, L, this);
1291 // Mark old value we replaced as possibly dead, so that it is elminated
1292 // if we just replaced the last use of that value.
1293 DeadInsts.insert(cast<Instruction>(User.OperandValToReplace));
1295 UsersToProcess.pop_back();
1298 // If there are any more users to process with the same base, process them
1299 // now. We sorted by base above, so we just have to check the last elt.
1300 } while (!UsersToProcess.empty() && UsersToProcess.back().Base == Base);
1301 // TODO: Next, find out which base index is the most common, pull it out.
1304 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
1305 // different starting values, into different PHIs.
1308 /// FindIVForUser - If Cond has an operand that is an expression of an IV,
1309 /// set the IV user and stride information and return true, otherwise return
1311 bool LoopStrengthReduce::FindIVForUser(ICmpInst *Cond, IVStrideUse *&CondUse,
1312 const SCEVHandle *&CondStride) {
1313 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e && !CondUse;
1315 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1316 IVUsesByStride.find(StrideOrder[Stride]);
1317 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1319 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
1320 E = SI->second.Users.end(); UI != E; ++UI)
1321 if (UI->User == Cond) {
1322 // NOTE: we could handle setcc instructions with multiple uses here, but
1323 // InstCombine does it as well for simple uses, it's not clear that it
1324 // occurs enough in real life to handle.
1326 CondStride = &SI->first;
1333 // OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
1334 // uses in the loop, look to see if we can eliminate some, in favor of using
1335 // common indvars for the different uses.
1336 void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
1337 // TODO: implement optzns here.
1339 // Finally, get the terminating condition for the loop if possible. If we
1340 // can, we want to change it to use a post-incremented version of its
1341 // induction variable, to allow coalescing the live ranges for the IV into
1342 // one register value.
1343 PHINode *SomePHI = cast<PHINode>(L->getHeader()->begin());
1344 BasicBlock *Preheader = L->getLoopPreheader();
1345 BasicBlock *LatchBlock =
1346 SomePHI->getIncomingBlock(SomePHI->getIncomingBlock(0) == Preheader);
1347 BranchInst *TermBr = dyn_cast<BranchInst>(LatchBlock->getTerminator());
1348 if (!TermBr || TermBr->isUnconditional() ||
1349 !isa<ICmpInst>(TermBr->getCondition()))
1351 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1353 // Search IVUsesByStride to find Cond's IVUse if there is one.
1354 IVStrideUse *CondUse = 0;
1355 const SCEVHandle *CondStride = 0;
1357 if (!FindIVForUser(Cond, CondUse, CondStride))
1358 return; // setcc doesn't use the IV.
1361 // It's possible for the setcc instruction to be anywhere in the loop, and
1362 // possible for it to have multiple users. If it is not immediately before
1363 // the latch block branch, move it.
1364 if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
1365 if (Cond->hasOneUse()) { // Condition has a single use, just move it.
1366 Cond->moveBefore(TermBr);
1368 // Otherwise, clone the terminating condition and insert into the loopend.
1369 Cond = cast<ICmpInst>(Cond->clone());
1370 Cond->setName(L->getHeader()->getName() + ".termcond");
1371 LatchBlock->getInstList().insert(TermBr, Cond);
1373 // Clone the IVUse, as the old use still exists!
1374 IVUsesByStride[*CondStride].addUser(CondUse->Offset, Cond,
1375 CondUse->OperandValToReplace);
1376 CondUse = &IVUsesByStride[*CondStride].Users.back();
1380 // If we get to here, we know that we can transform the setcc instruction to
1381 // use the post-incremented version of the IV, allowing us to coalesce the
1382 // live ranges for the IV correctly.
1383 CondUse->Offset = SCEV::getMinusSCEV(CondUse->Offset, *CondStride);
1384 CondUse->isUseOfPostIncrementedValue = true;
1388 // Constant strides come first which in turns are sorted by their absolute
1389 // values. If absolute values are the same, then positive strides comes first.
1391 // 4, -1, X, 1, 2 ==> 1, -1, 2, 4, X
1392 struct StrideCompare {
1393 bool operator()(const SCEVHandle &LHS, const SCEVHandle &RHS) {
1394 SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS);
1395 SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS);
1397 int64_t LV = LHSC->getValue()->getSExtValue();
1398 int64_t RV = RHSC->getValue()->getSExtValue();
1399 uint64_t ALV = (LV < 0) ? -LV : LV;
1400 uint64_t ARV = (RV < 0) ? -RV : RV;
1406 return (LHSC && !RHSC);
1411 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager &LPM) {
1413 LI = &getAnalysis<LoopInfo>();
1414 DT = &getAnalysis<DominatorTree>();
1415 SE = &getAnalysis<ScalarEvolution>();
1416 TD = &getAnalysis<TargetData>();
1417 UIntPtrTy = TD->getIntPtrType();
1419 // Find all uses of induction variables in this loop, and catagorize
1420 // them by stride. Start by finding all of the PHI nodes in the header for
1421 // this loop. If they are induction variables, inspect their uses.
1422 std::set<Instruction*> Processed; // Don't reprocess instructions.
1423 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I)
1424 AddUsersIfInteresting(I, L, Processed);
1426 // If we have nothing to do, return.
1427 if (IVUsesByStride.empty()) return false;
1429 // Optimize induction variables. Some indvar uses can be transformed to use
1430 // strides that will be needed for other purposes. A common example of this
1431 // is the exit test for the loop, which can often be rewritten to use the
1432 // computation of some other indvar to decide when to terminate the loop.
1436 // FIXME: We can widen subreg IV's here for RISC targets. e.g. instead of
1437 // doing computation in byte values, promote to 32-bit values if safe.
1439 // FIXME: Attempt to reuse values across multiple IV's. In particular, we
1440 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should be
1441 // codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC. Need
1442 // to be careful that IV's are all the same type. Only works for intptr_t
1445 // If we only have one stride, we can more aggressively eliminate some things.
1446 bool HasOneStride = IVUsesByStride.size() == 1;
1449 DOUT << "\nLSR on ";
1453 // IVsByStride keeps IVs for one particular loop.
1454 IVsByStride.clear();
1456 // Sort the StrideOrder so we process larger strides first.
1457 std::stable_sort(StrideOrder.begin(), StrideOrder.end(), StrideCompare());
1459 // Note: this processes each stride/type pair individually. All users passed
1460 // into StrengthReduceStridedIVUsers have the same type AND stride. Also,
1461 // node that we iterate over IVUsesByStride indirectly by using StrideOrder.
1462 // This extra layer of indirection makes the ordering of strides deterministic
1463 // - not dependent on map order.
1464 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e; ++Stride) {
1465 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1466 IVUsesByStride.find(StrideOrder[Stride]);
1467 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1468 StrengthReduceStridedIVUsers(SI->first, SI->second, L, HasOneStride);
1471 // Clean up after ourselves
1472 if (!DeadInsts.empty()) {
1473 DeleteTriviallyDeadInstructions(DeadInsts);
1475 BasicBlock::iterator I = L->getHeader()->begin();
1477 while ((PN = dyn_cast<PHINode>(I))) {
1478 ++I; // Preincrement iterator to avoid invalidating it when deleting PN.
1480 // At this point, we know that we have killed one or more GEP
1481 // instructions. It is worth checking to see if the cann indvar is also
1482 // dead, so that we can remove it as well. The requirements for the cann
1483 // indvar to be considered dead are:
1484 // 1. the cann indvar has one use
1485 // 2. the use is an add instruction
1486 // 3. the add has one use
1487 // 4. the add is used by the cann indvar
1488 // If all four cases above are true, then we can remove both the add and
1490 // FIXME: this needs to eliminate an induction variable even if it's being
1491 // compared against some value to decide loop termination.
1492 if (PN->hasOneUse()) {
1493 Instruction *BO = dyn_cast<Instruction>(*PN->use_begin());
1494 if (BO && (isa<BinaryOperator>(BO) || isa<CmpInst>(BO))) {
1495 if (BO->hasOneUse() && PN == *(BO->use_begin())) {
1496 DeadInsts.insert(BO);
1497 // Break the cycle, then delete the PHI.
1498 PN->replaceAllUsesWith(UndefValue::get(PN->getType()));
1499 SE->deleteValueFromRecords(PN);
1500 PN->eraseFromParent();
1505 DeleteTriviallyDeadInstructions(DeadInsts);
1508 CastedPointers.clear();
1509 IVUsesByStride.clear();
1510 StrideOrder.clear();