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/ScalarEvolutionExpander.h"
27 #include "llvm/Support/CFG.h"
28 #include "llvm/Support/GetElementPtrTypeIterator.h"
29 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
30 #include "llvm/Transforms/Utils/Local.h"
31 #include "llvm/Target/TargetData.h"
32 #include "llvm/ADT/Statistic.h"
33 #include "llvm/Support/Debug.h"
40 Statistic<> NumReduced ("loop-reduce", "Number of GEPs strength reduced");
41 Statistic<> NumInserted("loop-reduce", "Number of PHIs inserted");
42 Statistic<> NumVariable("loop-reduce","Number of PHIs with variable strides");
44 /// IVStrideUse - Keep track of one use of a strided induction variable, where
45 /// the stride is stored externally. The Offset member keeps track of the
46 /// offset from the IV, User is the actual user of the operand, and 'Operand'
47 /// is the operand # of the User that is the use.
51 Value *OperandValToReplace;
53 // isUseOfPostIncrementedValue - True if this should use the
54 // post-incremented version of this IV, not the preincremented version.
55 // This can only be set in special cases, such as the terminating setcc
56 // instruction for a loop or uses dominated by the loop.
57 bool isUseOfPostIncrementedValue;
59 IVStrideUse(const SCEVHandle &Offs, Instruction *U, Value *O)
60 : Offset(Offs), User(U), OperandValToReplace(O),
61 isUseOfPostIncrementedValue(false) {}
64 /// IVUsersOfOneStride - This structure keeps track of all instructions that
65 /// have an operand that is based on the trip count multiplied by some stride.
66 /// The stride for all of these users is common and kept external to this
68 struct IVUsersOfOneStride {
69 /// Users - Keep track of all of the users of this stride as well as the
70 /// initial value and the operand that uses the IV.
71 std::vector<IVStrideUse> Users;
73 void addUser(const SCEVHandle &Offset,Instruction *User, Value *Operand) {
74 Users.push_back(IVStrideUse(Offset, User, Operand));
79 class LoopStrengthReduce : public FunctionPass {
84 const Type *UIntPtrTy;
87 /// MaxTargetAMSize - This is the maximum power-of-two scale value that the
88 /// target can handle for free with its addressing modes.
89 unsigned MaxTargetAMSize;
91 /// IVUsesByStride - Keep track of all uses of induction variables that we
92 /// are interested in. The key of the map is the stride of the access.
93 std::map<SCEVHandle, IVUsersOfOneStride> IVUsesByStride;
95 /// StrideOrder - An ordering of the keys in IVUsesByStride that is stable:
96 /// We use this to iterate over the IVUsesByStride collection without being
97 /// dependent on random ordering of pointers in the process.
98 std::vector<SCEVHandle> StrideOrder;
100 /// CastedValues - As we need to cast values to uintptr_t, this keeps track
101 /// of the casted version of each value. This is accessed by
102 /// getCastedVersionOf.
103 std::map<Value*, Value*> CastedPointers;
105 /// DeadInsts - Keep track of instructions we may have made dead, so that
106 /// we can remove them after we are done working.
107 std::set<Instruction*> DeadInsts;
109 LoopStrengthReduce(unsigned MTAMS = 1)
110 : MaxTargetAMSize(MTAMS) {
113 virtual bool runOnFunction(Function &) {
114 LI = &getAnalysis<LoopInfo>();
115 EF = &getAnalysis<ETForest>();
116 SE = &getAnalysis<ScalarEvolution>();
117 TD = &getAnalysis<TargetData>();
118 UIntPtrTy = TD->getIntPtrType();
121 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
127 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
128 // We split critical edges, so we change the CFG. However, we do update
129 // many analyses if they are around.
130 AU.addPreservedID(LoopSimplifyID);
131 AU.addPreserved<LoopInfo>();
132 AU.addPreserved<DominatorSet>();
133 AU.addPreserved<ETForest>();
134 AU.addPreserved<ImmediateDominators>();
135 AU.addPreserved<DominanceFrontier>();
136 AU.addPreserved<DominatorTree>();
138 AU.addRequiredID(LoopSimplifyID);
139 AU.addRequired<LoopInfo>();
140 AU.addRequired<ETForest>();
141 AU.addRequired<TargetData>();
142 AU.addRequired<ScalarEvolution>();
145 /// getCastedVersionOf - Return the specified value casted to uintptr_t.
147 Value *getCastedVersionOf(Value *V);
149 void runOnLoop(Loop *L);
150 bool AddUsersIfInteresting(Instruction *I, Loop *L,
151 std::set<Instruction*> &Processed);
152 SCEVHandle GetExpressionSCEV(Instruction *E, Loop *L);
154 void OptimizeIndvars(Loop *L);
156 void StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
157 IVUsersOfOneStride &Uses,
158 Loop *L, bool isOnlyStride);
159 void DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts);
161 RegisterOpt<LoopStrengthReduce> X("loop-reduce",
162 "Loop Strength Reduction");
165 FunctionPass *llvm::createLoopStrengthReducePass(unsigned MaxTargetAMSize) {
166 return new LoopStrengthReduce(MaxTargetAMSize);
169 /// getCastedVersionOf - Return the specified value casted to uintptr_t.
171 Value *LoopStrengthReduce::getCastedVersionOf(Value *V) {
172 if (V->getType() == UIntPtrTy) return V;
173 if (Constant *CB = dyn_cast<Constant>(V))
174 return ConstantExpr::getCast(CB, UIntPtrTy);
176 Value *&New = CastedPointers[V];
179 New = SCEVExpander::InsertCastOfTo(V, UIntPtrTy);
180 DeadInsts.insert(cast<Instruction>(New));
185 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
186 /// specified set are trivially dead, delete them and see if this makes any of
187 /// their operands subsequently dead.
188 void LoopStrengthReduce::
189 DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts) {
190 while (!Insts.empty()) {
191 Instruction *I = *Insts.begin();
192 Insts.erase(Insts.begin());
193 if (isInstructionTriviallyDead(I)) {
194 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
195 if (Instruction *U = dyn_cast<Instruction>(I->getOperand(i)))
197 SE->deleteInstructionFromRecords(I);
198 I->eraseFromParent();
205 /// GetExpressionSCEV - Compute and return the SCEV for the specified
207 SCEVHandle LoopStrengthReduce::GetExpressionSCEV(Instruction *Exp, Loop *L) {
208 // Scalar Evolutions doesn't know how to compute SCEV's for GEP instructions.
209 // If this is a GEP that SE doesn't know about, compute it now and insert it.
210 // If this is not a GEP, or if we have already done this computation, just let
212 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Exp);
213 if (!GEP || SE->hasSCEV(GEP))
214 return SE->getSCEV(Exp);
216 // Analyze all of the subscripts of this getelementptr instruction, looking
217 // for uses that are determined by the trip count of L. First, skip all
218 // operands the are not dependent on the IV.
220 // Build up the base expression. Insert an LLVM cast of the pointer to
222 SCEVHandle GEPVal = SCEVUnknown::get(getCastedVersionOf(GEP->getOperand(0)));
224 gep_type_iterator GTI = gep_type_begin(GEP);
226 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
227 // If this is a use of a recurrence that we can analyze, and it comes before
228 // Op does in the GEP operand list, we will handle this when we process this
230 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
231 const StructLayout *SL = TD->getStructLayout(STy);
232 unsigned Idx = cast<ConstantUInt>(GEP->getOperand(i))->getValue();
233 uint64_t Offset = SL->MemberOffsets[Idx];
234 GEPVal = SCEVAddExpr::get(GEPVal,
235 SCEVUnknown::getIntegerSCEV(Offset, UIntPtrTy));
237 Value *OpVal = getCastedVersionOf(GEP->getOperand(i));
238 SCEVHandle Idx = SE->getSCEV(OpVal);
240 uint64_t TypeSize = TD->getTypeSize(GTI.getIndexedType());
242 Idx = SCEVMulExpr::get(Idx,
243 SCEVConstant::get(ConstantUInt::get(UIntPtrTy,
245 GEPVal = SCEVAddExpr::get(GEPVal, Idx);
249 SE->setSCEV(GEP, GEPVal);
253 /// getSCEVStartAndStride - Compute the start and stride of this expression,
254 /// returning false if the expression is not a start/stride pair, or true if it
255 /// is. The stride must be a loop invariant expression, but the start may be
256 /// a mix of loop invariant and loop variant expressions.
257 static bool getSCEVStartAndStride(const SCEVHandle &SH, Loop *L,
258 SCEVHandle &Start, SCEVHandle &Stride) {
259 SCEVHandle TheAddRec = Start; // Initialize to zero.
261 // If the outer level is an AddExpr, the operands are all start values except
262 // for a nested AddRecExpr.
263 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(SH)) {
264 for (unsigned i = 0, e = AE->getNumOperands(); i != e; ++i)
265 if (SCEVAddRecExpr *AddRec =
266 dyn_cast<SCEVAddRecExpr>(AE->getOperand(i))) {
267 if (AddRec->getLoop() == L)
268 TheAddRec = SCEVAddExpr::get(AddRec, TheAddRec);
270 return false; // Nested IV of some sort?
272 Start = SCEVAddExpr::get(Start, AE->getOperand(i));
275 } else if (SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SH)) {
278 return false; // not analyzable.
281 SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(TheAddRec);
282 if (!AddRec || AddRec->getLoop() != L) return false;
284 // FIXME: Generalize to non-affine IV's.
285 if (!AddRec->isAffine()) return false;
287 Start = SCEVAddExpr::get(Start, AddRec->getOperand(0));
289 if (!isa<SCEVConstant>(AddRec->getOperand(1)))
290 DEBUG(std::cerr << "[" << L->getHeader()->getName()
291 << "] Variable stride: " << *AddRec << "\n");
293 Stride = AddRec->getOperand(1);
294 // Check that all constant strides are the unsigned type, we don't want to
295 // have two IV's one of signed stride 4 and one of unsigned stride 4 to not be
297 assert((!isa<SCEVConstant>(Stride) || Stride->getType()->isUnsigned()) &&
298 "Constants should be canonicalized to unsigned!");
303 /// IVUseShouldUsePostIncValue - We have discovered a "User" of an IV expression
304 /// and now we need to decide whether the user should use the preinc or post-inc
305 /// value. If this user should use the post-inc version of the IV, return true.
307 /// Choosing wrong here can break dominance properties (if we choose to use the
308 /// post-inc value when we cannot) or it can end up adding extra live-ranges to
309 /// the loop, resulting in reg-reg copies (if we use the pre-inc value when we
310 /// should use the post-inc value).
311 static bool IVUseShouldUsePostIncValue(Instruction *User, Instruction *IV,
312 Loop *L, ETForest *EF, Pass *P) {
313 // If the user is in the loop, use the preinc value.
314 if (L->contains(User->getParent())) return false;
316 BasicBlock *LatchBlock = L->getLoopLatch();
318 // Ok, the user is outside of the loop. If it is dominated by the latch
319 // block, use the post-inc value.
320 if (EF->dominates(LatchBlock, User->getParent()))
323 // There is one case we have to be careful of: PHI nodes. These little guys
324 // can live in blocks that do not dominate the latch block, but (since their
325 // uses occur in the predecessor block, not the block the PHI lives in) should
326 // still use the post-inc value. Check for this case now.
327 PHINode *PN = dyn_cast<PHINode>(User);
328 if (!PN) return false; // not a phi, not dominated by latch block.
330 // Look at all of the uses of IV by the PHI node. If any use corresponds to
331 // a block that is not dominated by the latch block, give up and use the
332 // preincremented value.
333 unsigned NumUses = 0;
334 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
335 if (PN->getIncomingValue(i) == IV) {
337 if (!EF->dominates(LatchBlock, PN->getIncomingBlock(i)))
341 // Okay, all uses of IV by PN are in predecessor blocks that really are
342 // dominated by the latch block. Split the critical edges and use the
343 // post-incremented value.
344 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
345 if (PN->getIncomingValue(i) == IV) {
346 SplitCriticalEdge(PN->getIncomingBlock(i), PN->getParent(), P);
347 if (--NumUses == 0) break;
355 /// AddUsersIfInteresting - Inspect the specified instruction. If it is a
356 /// reducible SCEV, recursively add its users to the IVUsesByStride set and
357 /// return true. Otherwise, return false.
358 bool LoopStrengthReduce::AddUsersIfInteresting(Instruction *I, Loop *L,
359 std::set<Instruction*> &Processed) {
360 if (!I->getType()->isInteger() && !isa<PointerType>(I->getType()))
361 return false; // Void and FP expressions cannot be reduced.
362 if (!Processed.insert(I).second)
363 return true; // Instruction already handled.
365 // Get the symbolic expression for this instruction.
366 SCEVHandle ISE = GetExpressionSCEV(I, L);
367 if (isa<SCEVCouldNotCompute>(ISE)) return false;
369 // Get the start and stride for this expression.
370 SCEVHandle Start = SCEVUnknown::getIntegerSCEV(0, ISE->getType());
371 SCEVHandle Stride = Start;
372 if (!getSCEVStartAndStride(ISE, L, Start, Stride))
373 return false; // Non-reducible symbolic expression, bail out.
375 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;++UI){
376 Instruction *User = cast<Instruction>(*UI);
378 // Do not infinitely recurse on PHI nodes.
379 if (isa<PHINode>(User) && Processed.count(User))
382 // If this is an instruction defined in a nested loop, or outside this loop,
383 // don't recurse into it.
384 bool AddUserToIVUsers = false;
385 if (LI->getLoopFor(User->getParent()) != L) {
386 DEBUG(std::cerr << "FOUND USER in other loop: " << *User
387 << " OF SCEV: " << *ISE << "\n");
388 AddUserToIVUsers = true;
389 } else if (!AddUsersIfInteresting(User, L, Processed)) {
390 DEBUG(std::cerr << "FOUND USER: " << *User
391 << " OF SCEV: " << *ISE << "\n");
392 AddUserToIVUsers = true;
395 if (AddUserToIVUsers) {
396 IVUsersOfOneStride &StrideUses = IVUsesByStride[Stride];
397 if (StrideUses.Users.empty()) // First occurance of this stride?
398 StrideOrder.push_back(Stride);
400 // Okay, we found a user that we cannot reduce. Analyze the instruction
401 // and decide what to do with it. If we are a use inside of the loop, use
402 // the value before incrementation, otherwise use it after incrementation.
403 if (IVUseShouldUsePostIncValue(User, I, L, EF, this)) {
404 // The value used will be incremented by the stride more than we are
405 // expecting, so subtract this off.
406 SCEVHandle NewStart = SCEV::getMinusSCEV(Start, Stride);
407 StrideUses.addUser(NewStart, User, I);
408 StrideUses.Users.back().isUseOfPostIncrementedValue = true;
409 DEBUG(std::cerr << " USING POSTINC SCEV, START=" << *NewStart<< "\n");
411 StrideUses.addUser(Start, User, I);
419 /// BasedUser - For a particular base value, keep information about how we've
420 /// partitioned the expression so far.
422 /// Base - The Base value for the PHI node that needs to be inserted for
423 /// this use. As the use is processed, information gets moved from this
424 /// field to the Imm field (below). BasedUser values are sorted by this
428 /// Inst - The instruction using the induction variable.
431 /// OperandValToReplace - The operand value of Inst to replace with the
433 Value *OperandValToReplace;
435 /// Imm - The immediate value that should be added to the base immediately
436 /// before Inst, because it will be folded into the imm field of the
440 /// EmittedBase - The actual value* to use for the base value of this
441 /// operation. This is null if we should just use zero so far.
444 // isUseOfPostIncrementedValue - True if this should use the
445 // post-incremented version of this IV, not the preincremented version.
446 // This can only be set in special cases, such as the terminating setcc
447 // instruction for a loop and uses outside the loop that are dominated by
449 bool isUseOfPostIncrementedValue;
451 BasedUser(IVStrideUse &IVSU)
452 : Base(IVSU.Offset), Inst(IVSU.User),
453 OperandValToReplace(IVSU.OperandValToReplace),
454 Imm(SCEVUnknown::getIntegerSCEV(0, Base->getType())), EmittedBase(0),
455 isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue) {}
457 // Once we rewrite the code to insert the new IVs we want, update the
458 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
460 void RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
461 SCEVExpander &Rewriter, Loop *L,
464 Value *InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
465 SCEVExpander &Rewriter,
466 Instruction *IP, Loop *L);
471 void BasedUser::dump() const {
472 std::cerr << " Base=" << *Base;
473 std::cerr << " Imm=" << *Imm;
475 std::cerr << " EB=" << *EmittedBase;
477 std::cerr << " Inst: " << *Inst;
480 Value *BasedUser::InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
481 SCEVExpander &Rewriter,
482 Instruction *IP, Loop *L) {
483 // Figure out where we *really* want to insert this code. In particular, if
484 // the user is inside of a loop that is nested inside of L, we really don't
485 // want to insert this expression before the user, we'd rather pull it out as
486 // many loops as possible.
487 LoopInfo &LI = Rewriter.getLoopInfo();
488 Instruction *BaseInsertPt = IP;
490 // Figure out the most-nested loop that IP is in.
491 Loop *InsertLoop = LI.getLoopFor(IP->getParent());
493 // If InsertLoop is not L, and InsertLoop is nested inside of L, figure out
494 // the preheader of the outer-most loop where NewBase is not loop invariant.
495 while (InsertLoop && NewBase->isLoopInvariant(InsertLoop)) {
496 BaseInsertPt = InsertLoop->getLoopPreheader()->getTerminator();
497 InsertLoop = InsertLoop->getParentLoop();
500 // If there is no immediate value, skip the next part.
501 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Imm))
502 if (SC->getValue()->isNullValue())
503 return Rewriter.expandCodeFor(NewBase, BaseInsertPt,
504 OperandValToReplace->getType());
506 Value *Base = Rewriter.expandCodeFor(NewBase, BaseInsertPt);
508 // Always emit the immediate (if non-zero) into the same block as the user.
509 SCEVHandle NewValSCEV = SCEVAddExpr::get(SCEVUnknown::get(Base), Imm);
510 return Rewriter.expandCodeFor(NewValSCEV, IP,
511 OperandValToReplace->getType());
515 // Once we rewrite the code to insert the new IVs we want, update the
516 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
518 void BasedUser::RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
519 SCEVExpander &Rewriter,
521 if (!isa<PHINode>(Inst)) {
522 Value *NewVal = InsertCodeForBaseAtPosition(NewBase, Rewriter, Inst, L);
523 // Replace the use of the operand Value with the new Phi we just created.
524 Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
525 DEBUG(std::cerr << " CHANGED: IMM =" << *Imm << " Inst = " << *Inst);
529 // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm
530 // expression into each operand block that uses it. Note that PHI nodes can
531 // have multiple entries for the same predecessor. We use a map to make sure
532 // that a PHI node only has a single Value* for each predecessor (which also
533 // prevents us from inserting duplicate code in some blocks).
534 std::map<BasicBlock*, Value*> InsertedCode;
535 PHINode *PN = cast<PHINode>(Inst);
536 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
537 if (PN->getIncomingValue(i) == OperandValToReplace) {
538 // If this is a critical edge, split the edge so that we do not insert the
539 // code on all predecessor/successor paths. We do this unless this is the
540 // canonical backedge for this loop, as this can make some inserted code
541 // be in an illegal position.
542 BasicBlock *PHIPred = PN->getIncomingBlock(i);
543 if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 &&
544 (PN->getParent() != L->getHeader() || !L->contains(PHIPred))) {
546 // First step, split the critical edge.
547 SplitCriticalEdge(PHIPred, PN->getParent(), P);
549 // Next step: move the basic block. In particular, if the PHI node
550 // is outside of the loop, and PredTI is in the loop, we want to
551 // move the block to be immediately before the PHI block, not
552 // immediately after PredTI.
553 if (L->contains(PHIPred) && !L->contains(PN->getParent())) {
554 BasicBlock *NewBB = PN->getIncomingBlock(i);
555 NewBB->moveBefore(PN->getParent());
559 Value *&Code = InsertedCode[PN->getIncomingBlock(i)];
561 // Insert the code into the end of the predecessor block.
562 Instruction *InsertPt = PN->getIncomingBlock(i)->getTerminator();
563 Code = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
566 // Replace the use of the operand Value with the new Phi we just created.
567 PN->setIncomingValue(i, Code);
571 DEBUG(std::cerr << " CHANGED: IMM =" << *Imm << " Inst = " << *Inst);
575 /// isTargetConstant - Return true if the following can be referenced by the
576 /// immediate field of a target instruction.
577 static bool isTargetConstant(const SCEVHandle &V) {
579 // FIXME: Look at the target to decide if &GV is a legal constant immediate.
580 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
581 // PPC allows a sign-extended 16-bit immediate field.
582 int64_t V = SC->getValue()->getSExtValue();
583 if (V > -(1 << 16) && V < (1 << 16)-1)
588 return false; // ENABLE this for x86
590 if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
591 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(SU->getValue()))
592 if (CE->getOpcode() == Instruction::Cast)
593 if (isa<GlobalValue>(CE->getOperand(0)))
594 // FIXME: should check to see that the dest is uintptr_t!
599 /// MoveLoopVariantsToImediateField - Move any subexpressions from Val that are
600 /// loop varying to the Imm operand.
601 static void MoveLoopVariantsToImediateField(SCEVHandle &Val, SCEVHandle &Imm,
603 if (Val->isLoopInvariant(L)) return; // Nothing to do.
605 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
606 std::vector<SCEVHandle> NewOps;
607 NewOps.reserve(SAE->getNumOperands());
609 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
610 if (!SAE->getOperand(i)->isLoopInvariant(L)) {
611 // If this is a loop-variant expression, it must stay in the immediate
612 // field of the expression.
613 Imm = SCEVAddExpr::get(Imm, SAE->getOperand(i));
615 NewOps.push_back(SAE->getOperand(i));
619 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
621 Val = SCEVAddExpr::get(NewOps);
622 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
623 // Try to pull immediates out of the start value of nested addrec's.
624 SCEVHandle Start = SARE->getStart();
625 MoveLoopVariantsToImediateField(Start, Imm, L);
627 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
629 Val = SCEVAddRecExpr::get(Ops, SARE->getLoop());
631 // Otherwise, all of Val is variant, move the whole thing over.
632 Imm = SCEVAddExpr::get(Imm, Val);
633 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
638 /// MoveImmediateValues - Look at Val, and pull out any additions of constants
639 /// that can fit into the immediate field of instructions in the target.
640 /// Accumulate these immediate values into the Imm value.
641 static void MoveImmediateValues(SCEVHandle &Val, SCEVHandle &Imm,
642 bool isAddress, Loop *L) {
643 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
644 std::vector<SCEVHandle> NewOps;
645 NewOps.reserve(SAE->getNumOperands());
647 for (unsigned i = 0; i != SAE->getNumOperands(); ++i) {
648 SCEVHandle NewOp = SAE->getOperand(i);
649 MoveImmediateValues(NewOp, Imm, isAddress, L);
651 if (!NewOp->isLoopInvariant(L)) {
652 // If this is a loop-variant expression, it must stay in the immediate
653 // field of the expression.
654 Imm = SCEVAddExpr::get(Imm, NewOp);
656 NewOps.push_back(NewOp);
661 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
663 Val = SCEVAddExpr::get(NewOps);
665 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
666 // Try to pull immediates out of the start value of nested addrec's.
667 SCEVHandle Start = SARE->getStart();
668 MoveImmediateValues(Start, Imm, isAddress, L);
670 if (Start != SARE->getStart()) {
671 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
673 Val = SCEVAddRecExpr::get(Ops, SARE->getLoop());
676 } else if (SCEVMulExpr *SME = dyn_cast<SCEVMulExpr>(Val)) {
677 // Transform "8 * (4 + v)" -> "32 + 8*V" if "32" fits in the immed field.
678 if (isAddress && isTargetConstant(SME->getOperand(0)) &&
679 SME->getNumOperands() == 2 && SME->isLoopInvariant(L)) {
681 SCEVHandle SubImm = SCEVUnknown::getIntegerSCEV(0, Val->getType());
682 SCEVHandle NewOp = SME->getOperand(1);
683 MoveImmediateValues(NewOp, SubImm, isAddress, L);
685 // If we extracted something out of the subexpressions, see if we can
687 if (NewOp != SME->getOperand(1)) {
688 // Scale SubImm up by "8". If the result is a target constant, we are
690 SubImm = SCEVMulExpr::get(SubImm, SME->getOperand(0));
691 if (isTargetConstant(SubImm)) {
692 // Accumulate the immediate.
693 Imm = SCEVAddExpr::get(Imm, SubImm);
695 // Update what is left of 'Val'.
696 Val = SCEVMulExpr::get(SME->getOperand(0), NewOp);
703 // Loop-variant expressions must stay in the immediate field of the
705 if ((isAddress && isTargetConstant(Val)) ||
706 !Val->isLoopInvariant(L)) {
707 Imm = SCEVAddExpr::get(Imm, Val);
708 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
712 // Otherwise, no immediates to move.
716 /// IncrementAddExprUses - Decompose the specified expression into its added
717 /// subexpressions, and increment SubExpressionUseCounts for each of these
718 /// decomposed parts.
719 static void SeparateSubExprs(std::vector<SCEVHandle> &SubExprs,
721 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) {
722 for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
723 SeparateSubExprs(SubExprs, AE->getOperand(j));
724 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) {
725 SCEVHandle Zero = SCEVUnknown::getIntegerSCEV(0, Expr->getType());
726 if (SARE->getOperand(0) == Zero) {
727 SubExprs.push_back(Expr);
729 // Compute the addrec with zero as its base.
730 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
731 Ops[0] = Zero; // Start with zero base.
732 SubExprs.push_back(SCEVAddRecExpr::get(Ops, SARE->getLoop()));
735 SeparateSubExprs(SubExprs, SARE->getOperand(0));
737 } else if (!isa<SCEVConstant>(Expr) ||
738 !cast<SCEVConstant>(Expr)->getValue()->isNullValue()) {
740 SubExprs.push_back(Expr);
745 /// RemoveCommonExpressionsFromUseBases - Look through all of the uses in Bases,
746 /// removing any common subexpressions from it. Anything truly common is
747 /// removed, accumulated, and returned. This looks for things like (a+b+c) and
748 /// (a+c+d) -> (a+c). The common expression is *removed* from the Bases.
750 RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses) {
751 unsigned NumUses = Uses.size();
753 // Only one use? Use its base, regardless of what it is!
754 SCEVHandle Zero = SCEVUnknown::getIntegerSCEV(0, Uses[0].Base->getType());
755 SCEVHandle Result = Zero;
757 std::swap(Result, Uses[0].Base);
761 // To find common subexpressions, count how many of Uses use each expression.
762 // If any subexpressions are used Uses.size() times, they are common.
763 std::map<SCEVHandle, unsigned> SubExpressionUseCounts;
765 // UniqueSubExprs - Keep track of all of the subexpressions we see in the
766 // order we see them.
767 std::vector<SCEVHandle> UniqueSubExprs;
769 std::vector<SCEVHandle> SubExprs;
770 for (unsigned i = 0; i != NumUses; ++i) {
771 // If the base is zero (which is common), return zero now, there are no
773 if (Uses[i].Base == Zero) return Zero;
775 // Split the expression into subexprs.
776 SeparateSubExprs(SubExprs, Uses[i].Base);
777 // Add one to SubExpressionUseCounts for each subexpr present.
778 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
779 if (++SubExpressionUseCounts[SubExprs[j]] == 1)
780 UniqueSubExprs.push_back(SubExprs[j]);
784 // Now that we know how many times each is used, build Result. Iterate over
785 // UniqueSubexprs so that we have a stable ordering.
786 for (unsigned i = 0, e = UniqueSubExprs.size(); i != e; ++i) {
787 std::map<SCEVHandle, unsigned>::iterator I =
788 SubExpressionUseCounts.find(UniqueSubExprs[i]);
789 assert(I != SubExpressionUseCounts.end() && "Entry not found?");
790 if (I->second == NumUses) { // Found CSE!
791 Result = SCEVAddExpr::get(Result, I->first);
793 // Remove non-cse's from SubExpressionUseCounts.
794 SubExpressionUseCounts.erase(I);
798 // If we found no CSE's, return now.
799 if (Result == Zero) return Result;
801 // Otherwise, remove all of the CSE's we found from each of the base values.
802 for (unsigned i = 0; i != NumUses; ++i) {
803 // Split the expression into subexprs.
804 SeparateSubExprs(SubExprs, Uses[i].Base);
806 // Remove any common subexpressions.
807 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
808 if (SubExpressionUseCounts.count(SubExprs[j])) {
809 SubExprs.erase(SubExprs.begin()+j);
813 // Finally, the non-shared expressions together.
814 if (SubExprs.empty())
817 Uses[i].Base = SCEVAddExpr::get(SubExprs);
825 /// StrengthReduceStridedIVUsers - Strength reduce all of the users of a single
826 /// stride of IV. All of the users may have different starting values, and this
827 /// may not be the only stride (we know it is if isOnlyStride is true).
828 void LoopStrengthReduce::StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
829 IVUsersOfOneStride &Uses,
832 // Transform our list of users and offsets to a bit more complex table. In
833 // this new vector, each 'BasedUser' contains 'Base' the base of the
834 // strided accessas well as the old information from Uses. We progressively
835 // move information from the Base field to the Imm field, until we eventually
836 // have the full access expression to rewrite the use.
837 std::vector<BasedUser> UsersToProcess;
838 UsersToProcess.reserve(Uses.Users.size());
839 for (unsigned i = 0, e = Uses.Users.size(); i != e; ++i) {
840 UsersToProcess.push_back(Uses.Users[i]);
842 // Move any loop invariant operands from the offset field to the immediate
843 // field of the use, so that we don't try to use something before it is
845 MoveLoopVariantsToImediateField(UsersToProcess.back().Base,
846 UsersToProcess.back().Imm, L);
847 assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
848 "Base value is not loop invariant!");
851 // We now have a whole bunch of uses of like-strided induction variables, but
852 // they might all have different bases. We want to emit one PHI node for this
853 // stride which we fold as many common expressions (between the IVs) into as
854 // possible. Start by identifying the common expressions in the base values
855 // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
856 // "A+B"), emit it to the preheader, then remove the expression from the
857 // UsersToProcess base values.
858 SCEVHandle CommonExprs = RemoveCommonExpressionsFromUseBases(UsersToProcess);
860 // Next, figure out what we can represent in the immediate fields of
861 // instructions. If we can represent anything there, move it to the imm
862 // fields of the BasedUsers. We do this so that it increases the commonality
863 // of the remaining uses.
864 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
865 // If the user is not in the current loop, this means it is using the exit
866 // value of the IV. Do not put anything in the base, make sure it's all in
867 // the immediate field to allow as much factoring as possible.
868 if (!L->contains(UsersToProcess[i].Inst->getParent())) {
869 UsersToProcess[i].Imm = SCEVAddExpr::get(UsersToProcess[i].Imm,
870 UsersToProcess[i].Base);
871 UsersToProcess[i].Base =
872 SCEVUnknown::getIntegerSCEV(0, UsersToProcess[i].Base->getType());
875 // Addressing modes can be folded into loads and stores. Be careful that
876 // the store is through the expression, not of the expression though.
877 bool isAddress = isa<LoadInst>(UsersToProcess[i].Inst);
878 if (StoreInst *SI = dyn_cast<StoreInst>(UsersToProcess[i].Inst))
879 if (SI->getOperand(1) == UsersToProcess[i].OperandValToReplace)
882 MoveImmediateValues(UsersToProcess[i].Base, UsersToProcess[i].Imm,
887 // Now that we know what we need to do, insert the PHI node itself.
889 DEBUG(std::cerr << "INSERTING IV of STRIDE " << *Stride << " and BASE "
890 << *CommonExprs << " :\n");
892 SCEVExpander Rewriter(*SE, *LI);
893 SCEVExpander PreheaderRewriter(*SE, *LI);
895 BasicBlock *Preheader = L->getLoopPreheader();
896 Instruction *PreInsertPt = Preheader->getTerminator();
897 Instruction *PhiInsertBefore = L->getHeader()->begin();
899 BasicBlock *LatchBlock = L->getLoopLatch();
901 // Create a new Phi for this base, and stick it in the loop header.
902 const Type *ReplacedTy = CommonExprs->getType();
903 PHINode *NewPHI = new PHINode(ReplacedTy, "iv.", PhiInsertBefore);
906 // Insert the stride into the preheader.
907 Value *StrideV = PreheaderRewriter.expandCodeFor(Stride, PreInsertPt,
909 if (!isa<ConstantInt>(StrideV)) ++NumVariable;
912 // Emit the initial base value into the loop preheader, and add it to the
914 Value *PHIBaseV = PreheaderRewriter.expandCodeFor(CommonExprs, PreInsertPt,
916 NewPHI->addIncoming(PHIBaseV, Preheader);
918 // Emit the increment of the base value before the terminator of the loop
919 // latch block, and add it to the Phi node.
920 SCEVHandle IncExp = SCEVAddExpr::get(SCEVUnknown::get(NewPHI),
921 SCEVUnknown::get(StrideV));
923 Value *IncV = Rewriter.expandCodeFor(IncExp, LatchBlock->getTerminator(),
925 IncV->setName(NewPHI->getName()+".inc");
926 NewPHI->addIncoming(IncV, LatchBlock);
928 // Sort by the base value, so that all IVs with identical bases are next to
930 while (!UsersToProcess.empty()) {
931 SCEVHandle Base = UsersToProcess.back().Base;
933 DEBUG(std::cerr << " INSERTING code for BASE = " << *Base << ":\n");
935 // Emit the code for Base into the preheader.
936 Value *BaseV = PreheaderRewriter.expandCodeFor(Base, PreInsertPt,
939 // If BaseV is a constant other than 0, make sure that it gets inserted into
940 // the preheader, instead of being forward substituted into the uses. We do
941 // this by forcing a noop cast to be inserted into the preheader in this
943 if (Constant *C = dyn_cast<Constant>(BaseV))
944 if (!C->isNullValue() && !isTargetConstant(Base)) {
945 // We want this constant emitted into the preheader!
946 BaseV = new CastInst(BaseV, BaseV->getType(), "preheaderinsert",
950 // Emit the code to add the immediate offset to the Phi value, just before
951 // the instructions that we identified as using this stride and base.
952 unsigned ScanPos = 0;
954 BasedUser &User = UsersToProcess.back();
956 // If this instruction wants to use the post-incremented value, move it
957 // after the post-inc and use its value instead of the PHI.
958 Value *RewriteOp = NewPHI;
959 if (User.isUseOfPostIncrementedValue) {
962 // If this user is in the loop, make sure it is the last thing in the
963 // loop to ensure it is dominated by the increment.
964 if (L->contains(User.Inst->getParent()))
965 User.Inst->moveBefore(LatchBlock->getTerminator());
967 SCEVHandle RewriteExpr = SCEVUnknown::get(RewriteOp);
969 // Clear the SCEVExpander's expression map so that we are guaranteed
970 // to have the code emitted where we expect it.
973 // Now that we know what we need to do, insert code before User for the
974 // immediate and any loop-variant expressions.
975 if (!isa<ConstantInt>(BaseV) || !cast<ConstantInt>(BaseV)->isNullValue())
976 // Add BaseV to the PHI value if needed.
977 RewriteExpr = SCEVAddExpr::get(RewriteExpr, SCEVUnknown::get(BaseV));
979 User.RewriteInstructionToUseNewBase(RewriteExpr, Rewriter, L, this);
981 // Mark old value we replaced as possibly dead, so that it is elminated
982 // if we just replaced the last use of that value.
983 DeadInsts.insert(cast<Instruction>(User.OperandValToReplace));
985 UsersToProcess.pop_back();
988 // If there are any more users to process with the same base, move one of
989 // them to the end of the list so that we will process it.
990 if (!UsersToProcess.empty()) {
991 for (unsigned e = UsersToProcess.size(); ScanPos != e; ++ScanPos)
992 if (UsersToProcess[ScanPos].Base == Base) {
993 std::swap(UsersToProcess[ScanPos], UsersToProcess.back());
997 } while (!UsersToProcess.empty() && UsersToProcess.back().Base == Base);
998 // TODO: Next, find out which base index is the most common, pull it out.
1001 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
1002 // different starting values, into different PHIs.
1005 // OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
1006 // uses in the loop, look to see if we can eliminate some, in favor of using
1007 // common indvars for the different uses.
1008 void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
1009 // TODO: implement optzns here.
1014 // Finally, get the terminating condition for the loop if possible. If we
1015 // can, we want to change it to use a post-incremented version of its
1016 // induction variable, to allow coallescing the live ranges for the IV into
1017 // one register value.
1018 PHINode *SomePHI = cast<PHINode>(L->getHeader()->begin());
1019 BasicBlock *Preheader = L->getLoopPreheader();
1020 BasicBlock *LatchBlock =
1021 SomePHI->getIncomingBlock(SomePHI->getIncomingBlock(0) == Preheader);
1022 BranchInst *TermBr = dyn_cast<BranchInst>(LatchBlock->getTerminator());
1023 if (!TermBr || TermBr->isUnconditional() ||
1024 !isa<SetCondInst>(TermBr->getCondition()))
1026 SetCondInst *Cond = cast<SetCondInst>(TermBr->getCondition());
1028 // Search IVUsesByStride to find Cond's IVUse if there is one.
1029 IVStrideUse *CondUse = 0;
1030 const SCEVHandle *CondStride = 0;
1032 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e && !CondUse;
1034 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1035 IVUsesByStride.find(StrideOrder[Stride]);
1036 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1038 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
1039 E = SI->second.Users.end(); UI != E; ++UI)
1040 if (UI->User == Cond) {
1042 CondStride = &SI->first;
1043 // NOTE: we could handle setcc instructions with multiple uses here, but
1044 // InstCombine does it as well for simple uses, it's not clear that it
1045 // occurs enough in real life to handle.
1049 if (!CondUse) return; // setcc doesn't use the IV.
1051 // setcc stride is complex, don't mess with users.
1052 // FIXME: Evaluate whether this is a good idea or not.
1053 if (!isa<SCEVConstant>(*CondStride)) return;
1055 // It's possible for the setcc instruction to be anywhere in the loop, and
1056 // possible for it to have multiple users. If it is not immediately before
1057 // the latch block branch, move it.
1058 if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
1059 if (Cond->hasOneUse()) { // Condition has a single use, just move it.
1060 Cond->moveBefore(TermBr);
1062 // Otherwise, clone the terminating condition and insert into the loopend.
1063 Cond = cast<SetCondInst>(Cond->clone());
1064 Cond->setName(L->getHeader()->getName() + ".termcond");
1065 LatchBlock->getInstList().insert(TermBr, Cond);
1067 // Clone the IVUse, as the old use still exists!
1068 IVUsesByStride[*CondStride].addUser(CondUse->Offset, Cond,
1069 CondUse->OperandValToReplace);
1070 CondUse = &IVUsesByStride[*CondStride].Users.back();
1074 // If we get to here, we know that we can transform the setcc instruction to
1075 // use the post-incremented version of the IV, allowing us to coallesce the
1076 // live ranges for the IV correctly.
1077 CondUse->Offset = SCEV::getMinusSCEV(CondUse->Offset, *CondStride);
1078 CondUse->isUseOfPostIncrementedValue = true;
1081 void LoopStrengthReduce::runOnLoop(Loop *L) {
1082 // First step, transform all loops nesting inside of this loop.
1083 for (LoopInfo::iterator I = L->begin(), E = L->end(); I != E; ++I)
1086 // Next, find all uses of induction variables in this loop, and catagorize
1087 // them by stride. Start by finding all of the PHI nodes in the header for
1088 // this loop. If they are induction variables, inspect their uses.
1089 std::set<Instruction*> Processed; // Don't reprocess instructions.
1090 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I)
1091 AddUsersIfInteresting(I, L, Processed);
1093 // If we have nothing to do, return.
1094 if (IVUsesByStride.empty()) return;
1096 // Optimize induction variables. Some indvar uses can be transformed to use
1097 // strides that will be needed for other purposes. A common example of this
1098 // is the exit test for the loop, which can often be rewritten to use the
1099 // computation of some other indvar to decide when to terminate the loop.
1103 // FIXME: We can widen subreg IV's here for RISC targets. e.g. instead of
1104 // doing computation in byte values, promote to 32-bit values if safe.
1106 // FIXME: Attempt to reuse values across multiple IV's. In particular, we
1107 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should be
1108 // codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC. Need
1109 // to be careful that IV's are all the same type. Only works for intptr_t
1112 // If we only have one stride, we can more aggressively eliminate some things.
1113 bool HasOneStride = IVUsesByStride.size() == 1;
1115 // Note: this processes each stride/type pair individually. All users passed
1116 // into StrengthReduceStridedIVUsers have the same type AND stride. Also,
1117 // node that we iterate over IVUsesByStride indirectly by using StrideOrder.
1118 // This extra layer of indirection makes the ordering of strides deterministic
1119 // - not dependent on map order.
1120 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e; ++Stride) {
1121 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1122 IVUsesByStride.find(StrideOrder[Stride]);
1123 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1124 StrengthReduceStridedIVUsers(SI->first, SI->second, L, HasOneStride);
1127 // Clean up after ourselves
1128 if (!DeadInsts.empty()) {
1129 DeleteTriviallyDeadInstructions(DeadInsts);
1131 BasicBlock::iterator I = L->getHeader()->begin();
1133 while ((PN = dyn_cast<PHINode>(I))) {
1134 ++I; // Preincrement iterator to avoid invalidating it when deleting PN.
1136 // At this point, we know that we have killed one or more GEP
1137 // instructions. It is worth checking to see if the cann indvar is also
1138 // dead, so that we can remove it as well. The requirements for the cann
1139 // indvar to be considered dead are:
1140 // 1. the cann indvar has one use
1141 // 2. the use is an add instruction
1142 // 3. the add has one use
1143 // 4. the add is used by the cann indvar
1144 // If all four cases above are true, then we can remove both the add and
1146 // FIXME: this needs to eliminate an induction variable even if it's being
1147 // compared against some value to decide loop termination.
1148 if (PN->hasOneUse()) {
1149 BinaryOperator *BO = dyn_cast<BinaryOperator>(*(PN->use_begin()));
1150 if (BO && BO->hasOneUse()) {
1151 if (PN == *(BO->use_begin())) {
1152 DeadInsts.insert(BO);
1153 // Break the cycle, then delete the PHI.
1154 PN->replaceAllUsesWith(UndefValue::get(PN->getType()));
1155 SE->deleteInstructionFromRecords(PN);
1156 PN->eraseFromParent();
1161 DeleteTriviallyDeadInstructions(DeadInsts);
1164 CastedPointers.clear();
1165 IVUsesByStride.clear();
1166 StrideOrder.clear();