1 //===- LoopStrengthReduce.cpp - Strength Reduce IVs in Loops --------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This transformation analyzes and transforms the induction variables (and
11 // computations derived from them) into forms suitable for efficient execution
14 // This pass performs a strength reduction on array references inside loops that
15 // have as one or more of their components the loop induction variable, it
16 // rewrites expressions to take advantage of scaled-index addressing modes
17 // available on the target, and it performs a variety of other optimizations
18 // related to loop induction variables.
20 // Terminology note: this code has a lot of handling for "post-increment" or
21 // "post-inc" users. This is not talking about post-increment addressing modes;
22 // it is instead talking about code like this:
24 // %i = phi [ 0, %entry ], [ %i.next, %latch ]
26 // %i.next = add %i, 1
27 // %c = icmp eq %i.next, %n
29 // The SCEV for %i is {0,+,1}<%L>. The SCEV for %i.next is {1,+,1}<%L>, however
30 // it's useful to think about these as the same register, with some uses using
31 // the value of the register before the add and some using // it after. In this
32 // example, the icmp is a post-increment user, since it uses %i.next, which is
33 // the value of the induction variable after the increment. The other common
34 // case of post-increment users is users outside the loop.
36 // TODO: More sophistication in the way Formulae are generated and filtered.
38 // TODO: Handle multiple loops at a time.
40 // TODO: Should TargetLowering::AddrMode::BaseGV be changed to a ConstantExpr
41 // instead of a GlobalValue?
43 // TODO: When truncation is free, truncate ICmp users' operands to make it a
44 // smaller encoding (on x86 at least).
46 // TODO: When a negated register is used by an add (such as in a list of
47 // multiple base registers, or as the increment expression in an addrec),
48 // we may not actually need both reg and (-1 * reg) in registers; the
49 // negation can be implemented by using a sub instead of an add. The
50 // lack of support for taking this into consideration when making
51 // register pressure decisions is partly worked around by the "Special"
54 //===----------------------------------------------------------------------===//
56 #define DEBUG_TYPE "loop-reduce"
57 #include "llvm/Transforms/Scalar.h"
58 #include "llvm/Constants.h"
59 #include "llvm/Instructions.h"
60 #include "llvm/IntrinsicInst.h"
61 #include "llvm/DerivedTypes.h"
62 #include "llvm/Analysis/IVUsers.h"
63 #include "llvm/Analysis/Dominators.h"
64 #include "llvm/Analysis/LoopPass.h"
65 #include "llvm/Analysis/ScalarEvolutionExpander.h"
66 #include "llvm/Assembly/Writer.h"
67 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
68 #include "llvm/Transforms/Utils/Local.h"
69 #include "llvm/ADT/SmallBitVector.h"
70 #include "llvm/ADT/SetVector.h"
71 #include "llvm/ADT/DenseSet.h"
72 #include "llvm/Support/Debug.h"
73 #include "llvm/Support/CommandLine.h"
74 #include "llvm/Support/ValueHandle.h"
75 #include "llvm/Support/raw_ostream.h"
76 #include "llvm/Target/TargetLowering.h"
80 static cl::opt<bool> EnableRetry(
81 "enable-lsr-retry", cl::Hidden, cl::desc("Enable LSR retry"));
83 // Temporary flag to cleanup congruent phis after LSR phi expansion.
84 // It's currently disabled until we can determine whether it's truly useful or
85 // not. The flag should be removed after the v3.0 release.
86 // This is now needed for ivchains.
87 static cl::opt<bool> EnablePhiElim(
88 "enable-lsr-phielim", cl::Hidden, cl::init(true),
89 cl::desc("Enable LSR phi elimination"));
92 // Stress test IV chain generation.
93 static cl::opt<bool> StressIVChain(
94 "stress-ivchain", cl::Hidden, cl::init(false),
95 cl::desc("Stress test LSR IV chains"));
97 static bool StressIVChain = false;
102 /// RegSortData - This class holds data which is used to order reuse candidates.
105 /// UsedByIndices - This represents the set of LSRUse indices which reference
106 /// a particular register.
107 SmallBitVector UsedByIndices;
111 void print(raw_ostream &OS) const;
117 void RegSortData::print(raw_ostream &OS) const {
118 OS << "[NumUses=" << UsedByIndices.count() << ']';
121 void RegSortData::dump() const {
122 print(errs()); errs() << '\n';
127 /// RegUseTracker - Map register candidates to information about how they are
129 class RegUseTracker {
130 typedef DenseMap<const SCEV *, RegSortData> RegUsesTy;
132 RegUsesTy RegUsesMap;
133 SmallVector<const SCEV *, 16> RegSequence;
136 void CountRegister(const SCEV *Reg, size_t LUIdx);
137 void DropRegister(const SCEV *Reg, size_t LUIdx);
138 void SwapAndDropUse(size_t LUIdx, size_t LastLUIdx);
140 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
142 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
146 typedef SmallVectorImpl<const SCEV *>::iterator iterator;
147 typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator;
148 iterator begin() { return RegSequence.begin(); }
149 iterator end() { return RegSequence.end(); }
150 const_iterator begin() const { return RegSequence.begin(); }
151 const_iterator end() const { return RegSequence.end(); }
157 RegUseTracker::CountRegister(const SCEV *Reg, size_t LUIdx) {
158 std::pair<RegUsesTy::iterator, bool> Pair =
159 RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
160 RegSortData &RSD = Pair.first->second;
162 RegSequence.push_back(Reg);
163 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
164 RSD.UsedByIndices.set(LUIdx);
168 RegUseTracker::DropRegister(const SCEV *Reg, size_t LUIdx) {
169 RegUsesTy::iterator It = RegUsesMap.find(Reg);
170 assert(It != RegUsesMap.end());
171 RegSortData &RSD = It->second;
172 assert(RSD.UsedByIndices.size() > LUIdx);
173 RSD.UsedByIndices.reset(LUIdx);
177 RegUseTracker::SwapAndDropUse(size_t LUIdx, size_t LastLUIdx) {
178 assert(LUIdx <= LastLUIdx);
180 // Update RegUses. The data structure is not optimized for this purpose;
181 // we must iterate through it and update each of the bit vectors.
182 for (RegUsesTy::iterator I = RegUsesMap.begin(), E = RegUsesMap.end();
184 SmallBitVector &UsedByIndices = I->second.UsedByIndices;
185 if (LUIdx < UsedByIndices.size())
186 UsedByIndices[LUIdx] =
187 LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : 0;
188 UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx));
193 RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
194 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
195 if (I == RegUsesMap.end())
197 const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
198 int i = UsedByIndices.find_first();
199 if (i == -1) return false;
200 if ((size_t)i != LUIdx) return true;
201 return UsedByIndices.find_next(i) != -1;
204 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
205 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
206 assert(I != RegUsesMap.end() && "Unknown register!");
207 return I->second.UsedByIndices;
210 void RegUseTracker::clear() {
217 /// Formula - This class holds information that describes a formula for
218 /// computing satisfying a use. It may include broken-out immediates and scaled
221 /// AM - This is used to represent complex addressing, as well as other kinds
222 /// of interesting uses.
223 TargetLowering::AddrMode AM;
225 /// BaseRegs - The list of "base" registers for this use. When this is
226 /// non-empty, AM.HasBaseReg should be set to true.
227 SmallVector<const SCEV *, 2> BaseRegs;
229 /// ScaledReg - The 'scaled' register for this use. This should be non-null
230 /// when AM.Scale is not zero.
231 const SCEV *ScaledReg;
233 /// UnfoldedOffset - An additional constant offset which added near the
234 /// use. This requires a temporary register, but the offset itself can
235 /// live in an add immediate field rather than a register.
236 int64_t UnfoldedOffset;
238 Formula() : ScaledReg(0), UnfoldedOffset(0) {}
240 void InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE);
242 unsigned getNumRegs() const;
243 Type *getType() const;
245 void DeleteBaseReg(const SCEV *&S);
247 bool referencesReg(const SCEV *S) const;
248 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
249 const RegUseTracker &RegUses) const;
251 void print(raw_ostream &OS) const;
257 /// DoInitialMatch - Recursion helper for InitialMatch.
258 static void DoInitialMatch(const SCEV *S, Loop *L,
259 SmallVectorImpl<const SCEV *> &Good,
260 SmallVectorImpl<const SCEV *> &Bad,
261 ScalarEvolution &SE) {
262 // Collect expressions which properly dominate the loop header.
263 if (SE.properlyDominates(S, L->getHeader())) {
268 // Look at add operands.
269 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
270 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
272 DoInitialMatch(*I, L, Good, Bad, SE);
276 // Look at addrec operands.
277 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
278 if (!AR->getStart()->isZero()) {
279 DoInitialMatch(AR->getStart(), L, Good, Bad, SE);
280 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
281 AR->getStepRecurrence(SE),
282 // FIXME: AR->getNoWrapFlags()
283 AR->getLoop(), SCEV::FlagAnyWrap),
288 // Handle a multiplication by -1 (negation) if it didn't fold.
289 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
290 if (Mul->getOperand(0)->isAllOnesValue()) {
291 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
292 const SCEV *NewMul = SE.getMulExpr(Ops);
294 SmallVector<const SCEV *, 4> MyGood;
295 SmallVector<const SCEV *, 4> MyBad;
296 DoInitialMatch(NewMul, L, MyGood, MyBad, SE);
297 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
298 SE.getEffectiveSCEVType(NewMul->getType())));
299 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyGood.begin(),
300 E = MyGood.end(); I != E; ++I)
301 Good.push_back(SE.getMulExpr(NegOne, *I));
302 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyBad.begin(),
303 E = MyBad.end(); I != E; ++I)
304 Bad.push_back(SE.getMulExpr(NegOne, *I));
308 // Ok, we can't do anything interesting. Just stuff the whole thing into a
309 // register and hope for the best.
313 /// InitialMatch - Incorporate loop-variant parts of S into this Formula,
314 /// attempting to keep all loop-invariant and loop-computable values in a
315 /// single base register.
316 void Formula::InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) {
317 SmallVector<const SCEV *, 4> Good;
318 SmallVector<const SCEV *, 4> Bad;
319 DoInitialMatch(S, L, Good, Bad, SE);
321 const SCEV *Sum = SE.getAddExpr(Good);
323 BaseRegs.push_back(Sum);
324 AM.HasBaseReg = true;
327 const SCEV *Sum = SE.getAddExpr(Bad);
329 BaseRegs.push_back(Sum);
330 AM.HasBaseReg = true;
334 /// getNumRegs - Return the total number of register operands used by this
335 /// formula. This does not include register uses implied by non-constant
337 unsigned Formula::getNumRegs() const {
338 return !!ScaledReg + BaseRegs.size();
341 /// getType - Return the type of this formula, if it has one, or null
342 /// otherwise. This type is meaningless except for the bit size.
343 Type *Formula::getType() const {
344 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
345 ScaledReg ? ScaledReg->getType() :
346 AM.BaseGV ? AM.BaseGV->getType() :
350 /// DeleteBaseReg - Delete the given base reg from the BaseRegs list.
351 void Formula::DeleteBaseReg(const SCEV *&S) {
352 if (&S != &BaseRegs.back())
353 std::swap(S, BaseRegs.back());
357 /// referencesReg - Test if this formula references the given register.
358 bool Formula::referencesReg(const SCEV *S) const {
359 return S == ScaledReg ||
360 std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end();
363 /// hasRegsUsedByUsesOtherThan - Test whether this formula uses registers
364 /// which are used by uses other than the use with the given index.
365 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
366 const RegUseTracker &RegUses) const {
368 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
370 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
371 E = BaseRegs.end(); I != E; ++I)
372 if (RegUses.isRegUsedByUsesOtherThan(*I, LUIdx))
377 void Formula::print(raw_ostream &OS) const {
380 if (!First) OS << " + "; else First = false;
381 WriteAsOperand(OS, AM.BaseGV, /*PrintType=*/false);
383 if (AM.BaseOffs != 0) {
384 if (!First) OS << " + "; else First = false;
387 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
388 E = BaseRegs.end(); I != E; ++I) {
389 if (!First) OS << " + "; else First = false;
390 OS << "reg(" << **I << ')';
392 if (AM.HasBaseReg && BaseRegs.empty()) {
393 if (!First) OS << " + "; else First = false;
394 OS << "**error: HasBaseReg**";
395 } else if (!AM.HasBaseReg && !BaseRegs.empty()) {
396 if (!First) OS << " + "; else First = false;
397 OS << "**error: !HasBaseReg**";
400 if (!First) OS << " + "; else First = false;
401 OS << AM.Scale << "*reg(";
408 if (UnfoldedOffset != 0) {
409 if (!First) OS << " + "; else First = false;
410 OS << "imm(" << UnfoldedOffset << ')';
414 void Formula::dump() const {
415 print(errs()); errs() << '\n';
418 /// isAddRecSExtable - Return true if the given addrec can be sign-extended
419 /// without changing its value.
420 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
422 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1);
423 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
426 /// isAddSExtable - Return true if the given add can be sign-extended
427 /// without changing its value.
428 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
430 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
431 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
434 /// isMulSExtable - Return true if the given mul can be sign-extended
435 /// without changing its value.
436 static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
438 IntegerType::get(SE.getContext(),
439 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
440 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
443 /// getExactSDiv - Return an expression for LHS /s RHS, if it can be determined
444 /// and if the remainder is known to be zero, or null otherwise. If
445 /// IgnoreSignificantBits is true, expressions like (X * Y) /s Y are simplified
446 /// to Y, ignoring that the multiplication may overflow, which is useful when
447 /// the result will be used in a context where the most significant bits are
449 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
451 bool IgnoreSignificantBits = false) {
452 // Handle the trivial case, which works for any SCEV type.
454 return SE.getConstant(LHS->getType(), 1);
456 // Handle a few RHS special cases.
457 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
459 const APInt &RA = RC->getValue()->getValue();
460 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
462 if (RA.isAllOnesValue())
463 return SE.getMulExpr(LHS, RC);
464 // Handle x /s 1 as x.
469 // Check for a division of a constant by a constant.
470 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
473 const APInt &LA = C->getValue()->getValue();
474 const APInt &RA = RC->getValue()->getValue();
475 if (LA.srem(RA) != 0)
477 return SE.getConstant(LA.sdiv(RA));
480 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
481 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
482 if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) {
483 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
484 IgnoreSignificantBits);
486 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
487 IgnoreSignificantBits);
488 if (!Start) return 0;
489 // FlagNW is independent of the start value, step direction, and is
490 // preserved with smaller magnitude steps.
491 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
492 return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap);
497 // Distribute the sdiv over add operands, if the add doesn't overflow.
498 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
499 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
500 SmallVector<const SCEV *, 8> Ops;
501 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
503 const SCEV *Op = getExactSDiv(*I, RHS, SE,
504 IgnoreSignificantBits);
508 return SE.getAddExpr(Ops);
513 // Check for a multiply operand that we can pull RHS out of.
514 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
515 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
516 SmallVector<const SCEV *, 4> Ops;
518 for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end();
522 if (const SCEV *Q = getExactSDiv(S, RHS, SE,
523 IgnoreSignificantBits)) {
529 return Found ? SE.getMulExpr(Ops) : 0;
534 // Otherwise we don't know.
538 /// ExtractImmediate - If S involves the addition of a constant integer value,
539 /// return that integer value, and mutate S to point to a new SCEV with that
541 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
542 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
543 if (C->getValue()->getValue().getMinSignedBits() <= 64) {
544 S = SE.getConstant(C->getType(), 0);
545 return C->getValue()->getSExtValue();
547 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
548 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
549 int64_t Result = ExtractImmediate(NewOps.front(), SE);
551 S = SE.getAddExpr(NewOps);
553 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
554 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
555 int64_t Result = ExtractImmediate(NewOps.front(), SE);
557 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
558 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
565 /// ExtractSymbol - If S involves the addition of a GlobalValue address,
566 /// return that symbol, and mutate S to point to a new SCEV with that
568 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
569 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
570 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
571 S = SE.getConstant(GV->getType(), 0);
574 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
575 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
576 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
578 S = SE.getAddExpr(NewOps);
580 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
581 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
582 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
584 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
585 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
592 /// isAddressUse - Returns true if the specified instruction is using the
593 /// specified value as an address.
594 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
595 bool isAddress = isa<LoadInst>(Inst);
596 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
597 if (SI->getOperand(1) == OperandVal)
599 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
600 // Addressing modes can also be folded into prefetches and a variety
602 switch (II->getIntrinsicID()) {
604 case Intrinsic::prefetch:
605 case Intrinsic::x86_sse_storeu_ps:
606 case Intrinsic::x86_sse2_storeu_pd:
607 case Intrinsic::x86_sse2_storeu_dq:
608 case Intrinsic::x86_sse2_storel_dq:
609 if (II->getArgOperand(0) == OperandVal)
617 /// getAccessType - Return the type of the memory being accessed.
618 static Type *getAccessType(const Instruction *Inst) {
619 Type *AccessTy = Inst->getType();
620 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
621 AccessTy = SI->getOperand(0)->getType();
622 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
623 // Addressing modes can also be folded into prefetches and a variety
625 switch (II->getIntrinsicID()) {
627 case Intrinsic::x86_sse_storeu_ps:
628 case Intrinsic::x86_sse2_storeu_pd:
629 case Intrinsic::x86_sse2_storeu_dq:
630 case Intrinsic::x86_sse2_storel_dq:
631 AccessTy = II->getArgOperand(0)->getType();
636 // All pointers have the same requirements, so canonicalize them to an
637 // arbitrary pointer type to minimize variation.
638 if (PointerType *PTy = dyn_cast<PointerType>(AccessTy))
639 AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
640 PTy->getAddressSpace());
645 /// isExistingPhi - Return true if this AddRec is already a phi in its loop.
646 static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
647 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
648 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
649 if (SE.isSCEVable(PN->getType()) &&
650 (SE.getEffectiveSCEVType(PN->getType()) ==
651 SE.getEffectiveSCEVType(AR->getType())) &&
652 SE.getSCEV(PN) == AR)
658 /// Check if expanding this expression is likely to incur significant cost. This
659 /// is tricky because SCEV doesn't track which expressions are actually computed
660 /// by the current IR.
662 /// We currently allow expansion of IV increments that involve adds,
663 /// multiplication by constants, and AddRecs from existing phis.
665 /// TODO: Allow UDivExpr if we can find an existing IV increment that is an
666 /// obvious multiple of the UDivExpr.
667 static bool isHighCostExpansion(const SCEV *S,
668 SmallPtrSet<const SCEV*, 8> &Processed,
669 ScalarEvolution &SE) {
670 // Zero/One operand expressions
671 switch (S->getSCEVType()) {
676 return isHighCostExpansion(cast<SCEVTruncateExpr>(S)->getOperand(),
679 return isHighCostExpansion(cast<SCEVZeroExtendExpr>(S)->getOperand(),
682 return isHighCostExpansion(cast<SCEVSignExtendExpr>(S)->getOperand(),
686 if (!Processed.insert(S))
689 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
690 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
692 if (isHighCostExpansion(*I, Processed, SE))
698 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
699 if (Mul->getNumOperands() == 2) {
700 // Multiplication by a constant is ok
701 if (isa<SCEVConstant>(Mul->getOperand(0)))
702 return isHighCostExpansion(Mul->getOperand(1), Processed, SE);
704 // If we have the value of one operand, check if an existing
705 // multiplication already generates this expression.
706 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Mul->getOperand(1))) {
707 Value *UVal = U->getValue();
708 for (Value::use_iterator UI = UVal->use_begin(), UE = UVal->use_end();
710 Instruction *User = cast<Instruction>(*UI);
711 if (User->getOpcode() == Instruction::Mul
712 && SE.isSCEVable(User->getType())) {
713 return SE.getSCEV(User) == Mul;
720 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
721 if (isExistingPhi(AR, SE))
725 // Fow now, consider any other type of expression (div/mul/min/max) high cost.
729 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
730 /// specified set are trivially dead, delete them and see if this makes any of
731 /// their operands subsequently dead.
733 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) {
734 bool Changed = false;
736 while (!DeadInsts.empty()) {
737 Instruction *I = dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val());
739 if (I == 0 || !isInstructionTriviallyDead(I))
742 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
743 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
746 DeadInsts.push_back(U);
749 I->eraseFromParent();
758 /// Cost - This class is used to measure and compare candidate formulae.
760 /// TODO: Some of these could be merged. Also, a lexical ordering
761 /// isn't always optimal.
765 unsigned NumBaseAdds;
771 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0),
774 bool operator<(const Cost &Other) const;
779 // Once any of the metrics loses, they must all remain losers.
781 return ((NumRegs | AddRecCost | NumIVMuls | NumBaseAdds
782 | ImmCost | SetupCost) != ~0u)
783 || ((NumRegs & AddRecCost & NumIVMuls & NumBaseAdds
784 & ImmCost & SetupCost) == ~0u);
789 assert(isValid() && "invalid cost");
790 return NumRegs == ~0u;
793 void RateFormula(const Formula &F,
794 SmallPtrSet<const SCEV *, 16> &Regs,
795 const DenseSet<const SCEV *> &VisitedRegs,
797 const SmallVectorImpl<int64_t> &Offsets,
798 ScalarEvolution &SE, DominatorTree &DT,
799 SmallPtrSet<const SCEV *, 16> *LoserRegs = 0);
801 void print(raw_ostream &OS) const;
805 void RateRegister(const SCEV *Reg,
806 SmallPtrSet<const SCEV *, 16> &Regs,
808 ScalarEvolution &SE, DominatorTree &DT);
809 void RatePrimaryRegister(const SCEV *Reg,
810 SmallPtrSet<const SCEV *, 16> &Regs,
812 ScalarEvolution &SE, DominatorTree &DT,
813 SmallPtrSet<const SCEV *, 16> *LoserRegs);
818 /// RateRegister - Tally up interesting quantities from the given register.
819 void Cost::RateRegister(const SCEV *Reg,
820 SmallPtrSet<const SCEV *, 16> &Regs,
822 ScalarEvolution &SE, DominatorTree &DT) {
823 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
824 // If this is an addrec for another loop, don't second-guess its addrec phi
825 // nodes. LSR isn't currently smart enough to reason about more than one
826 // loop at a time. LSR has already run on inner loops, will not run on outer
827 // loops, and cannot be expected to change sibling loops.
828 if (AR->getLoop() != L) {
829 // If the AddRec exists, consider it's register free and leave it alone.
830 if (isExistingPhi(AR, SE))
833 // Otherwise, do not consider this formula at all.
837 AddRecCost += 1; /// TODO: This should be a function of the stride.
839 // Add the step value register, if it needs one.
840 // TODO: The non-affine case isn't precisely modeled here.
841 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
842 if (!Regs.count(AR->getOperand(1))) {
843 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
851 // Rough heuristic; favor registers which don't require extra setup
852 // instructions in the preheader.
853 if (!isa<SCEVUnknown>(Reg) &&
854 !isa<SCEVConstant>(Reg) &&
855 !(isa<SCEVAddRecExpr>(Reg) &&
856 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
857 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
860 NumIVMuls += isa<SCEVMulExpr>(Reg) &&
861 SE.hasComputableLoopEvolution(Reg, L);
864 /// RatePrimaryRegister - Record this register in the set. If we haven't seen it
865 /// before, rate it. Optional LoserRegs provides a way to declare any formula
866 /// that refers to one of those regs an instant loser.
867 void Cost::RatePrimaryRegister(const SCEV *Reg,
868 SmallPtrSet<const SCEV *, 16> &Regs,
870 ScalarEvolution &SE, DominatorTree &DT,
871 SmallPtrSet<const SCEV *, 16> *LoserRegs) {
872 if (LoserRegs && LoserRegs->count(Reg)) {
876 if (Regs.insert(Reg)) {
877 RateRegister(Reg, Regs, L, SE, DT);
879 LoserRegs->insert(Reg);
883 void Cost::RateFormula(const Formula &F,
884 SmallPtrSet<const SCEV *, 16> &Regs,
885 const DenseSet<const SCEV *> &VisitedRegs,
887 const SmallVectorImpl<int64_t> &Offsets,
888 ScalarEvolution &SE, DominatorTree &DT,
889 SmallPtrSet<const SCEV *, 16> *LoserRegs) {
890 // Tally up the registers.
891 if (const SCEV *ScaledReg = F.ScaledReg) {
892 if (VisitedRegs.count(ScaledReg)) {
896 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT, LoserRegs);
900 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
901 E = F.BaseRegs.end(); I != E; ++I) {
902 const SCEV *BaseReg = *I;
903 if (VisitedRegs.count(BaseReg)) {
907 RatePrimaryRegister(BaseReg, Regs, L, SE, DT, LoserRegs);
912 // Determine how many (unfolded) adds we'll need inside the loop.
913 size_t NumBaseParts = F.BaseRegs.size() + (F.UnfoldedOffset != 0);
914 if (NumBaseParts > 1)
915 NumBaseAdds += NumBaseParts - 1;
917 // Tally up the non-zero immediates.
918 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
919 E = Offsets.end(); I != E; ++I) {
920 int64_t Offset = (uint64_t)*I + F.AM.BaseOffs;
922 ImmCost += 64; // Handle symbolic values conservatively.
923 // TODO: This should probably be the pointer size.
924 else if (Offset != 0)
925 ImmCost += APInt(64, Offset, true).getMinSignedBits();
927 assert(isValid() && "invalid cost");
930 /// Loose - Set this cost to a losing value.
940 /// operator< - Choose the lower cost.
941 bool Cost::operator<(const Cost &Other) const {
942 if (NumRegs != Other.NumRegs)
943 return NumRegs < Other.NumRegs;
944 if (AddRecCost != Other.AddRecCost)
945 return AddRecCost < Other.AddRecCost;
946 if (NumIVMuls != Other.NumIVMuls)
947 return NumIVMuls < Other.NumIVMuls;
948 if (NumBaseAdds != Other.NumBaseAdds)
949 return NumBaseAdds < Other.NumBaseAdds;
950 if (ImmCost != Other.ImmCost)
951 return ImmCost < Other.ImmCost;
952 if (SetupCost != Other.SetupCost)
953 return SetupCost < Other.SetupCost;
957 void Cost::print(raw_ostream &OS) const {
958 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
960 OS << ", with addrec cost " << AddRecCost;
962 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
963 if (NumBaseAdds != 0)
964 OS << ", plus " << NumBaseAdds << " base add"
965 << (NumBaseAdds == 1 ? "" : "s");
967 OS << ", plus " << ImmCost << " imm cost";
969 OS << ", plus " << SetupCost << " setup cost";
972 void Cost::dump() const {
973 print(errs()); errs() << '\n';
978 /// LSRFixup - An operand value in an instruction which is to be replaced
979 /// with some equivalent, possibly strength-reduced, replacement.
981 /// UserInst - The instruction which will be updated.
982 Instruction *UserInst;
984 /// OperandValToReplace - The operand of the instruction which will
985 /// be replaced. The operand may be used more than once; every instance
986 /// will be replaced.
987 Value *OperandValToReplace;
989 /// PostIncLoops - If this user is to use the post-incremented value of an
990 /// induction variable, this variable is non-null and holds the loop
991 /// associated with the induction variable.
992 PostIncLoopSet PostIncLoops;
994 /// LUIdx - The index of the LSRUse describing the expression which
995 /// this fixup needs, minus an offset (below).
998 /// Offset - A constant offset to be added to the LSRUse expression.
999 /// This allows multiple fixups to share the same LSRUse with different
1000 /// offsets, for example in an unrolled loop.
1003 bool isUseFullyOutsideLoop(const Loop *L) const;
1007 void print(raw_ostream &OS) const;
1013 LSRFixup::LSRFixup()
1014 : UserInst(0), OperandValToReplace(0), LUIdx(~size_t(0)), Offset(0) {}
1016 /// isUseFullyOutsideLoop - Test whether this fixup always uses its
1017 /// value outside of the given loop.
1018 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
1019 // PHI nodes use their value in their incoming blocks.
1020 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
1021 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1022 if (PN->getIncomingValue(i) == OperandValToReplace &&
1023 L->contains(PN->getIncomingBlock(i)))
1028 return !L->contains(UserInst);
1031 void LSRFixup::print(raw_ostream &OS) const {
1033 // Store is common and interesting enough to be worth special-casing.
1034 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
1036 WriteAsOperand(OS, Store->getOperand(0), /*PrintType=*/false);
1037 } else if (UserInst->getType()->isVoidTy())
1038 OS << UserInst->getOpcodeName();
1040 WriteAsOperand(OS, UserInst, /*PrintType=*/false);
1042 OS << ", OperandValToReplace=";
1043 WriteAsOperand(OS, OperandValToReplace, /*PrintType=*/false);
1045 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(),
1046 E = PostIncLoops.end(); I != E; ++I) {
1047 OS << ", PostIncLoop=";
1048 WriteAsOperand(OS, (*I)->getHeader(), /*PrintType=*/false);
1051 if (LUIdx != ~size_t(0))
1052 OS << ", LUIdx=" << LUIdx;
1055 OS << ", Offset=" << Offset;
1058 void LSRFixup::dump() const {
1059 print(errs()); errs() << '\n';
1064 /// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding
1065 /// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*.
1066 struct UniquifierDenseMapInfo {
1067 static SmallVector<const SCEV *, 2> getEmptyKey() {
1068 SmallVector<const SCEV *, 2> V;
1069 V.push_back(reinterpret_cast<const SCEV *>(-1));
1073 static SmallVector<const SCEV *, 2> getTombstoneKey() {
1074 SmallVector<const SCEV *, 2> V;
1075 V.push_back(reinterpret_cast<const SCEV *>(-2));
1079 static unsigned getHashValue(const SmallVector<const SCEV *, 2> &V) {
1080 unsigned Result = 0;
1081 for (SmallVectorImpl<const SCEV *>::const_iterator I = V.begin(),
1082 E = V.end(); I != E; ++I)
1083 Result ^= DenseMapInfo<const SCEV *>::getHashValue(*I);
1087 static bool isEqual(const SmallVector<const SCEV *, 2> &LHS,
1088 const SmallVector<const SCEV *, 2> &RHS) {
1093 /// LSRUse - This class holds the state that LSR keeps for each use in
1094 /// IVUsers, as well as uses invented by LSR itself. It includes information
1095 /// about what kinds of things can be folded into the user, information about
1096 /// the user itself, and information about how the use may be satisfied.
1097 /// TODO: Represent multiple users of the same expression in common?
1099 DenseSet<SmallVector<const SCEV *, 2>, UniquifierDenseMapInfo> Uniquifier;
1102 /// KindType - An enum for a kind of use, indicating what types of
1103 /// scaled and immediate operands it might support.
1105 Basic, ///< A normal use, with no folding.
1106 Special, ///< A special case of basic, allowing -1 scales.
1107 Address, ///< An address use; folding according to TargetLowering
1108 ICmpZero ///< An equality icmp with both operands folded into one.
1109 // TODO: Add a generic icmp too?
1115 SmallVector<int64_t, 8> Offsets;
1119 /// AllFixupsOutsideLoop - This records whether all of the fixups using this
1120 /// LSRUse are outside of the loop, in which case some special-case heuristics
1122 bool AllFixupsOutsideLoop;
1124 /// WidestFixupType - This records the widest use type for any fixup using
1125 /// this LSRUse. FindUseWithSimilarFormula can't consider uses with different
1126 /// max fixup widths to be equivalent, because the narrower one may be relying
1127 /// on the implicit truncation to truncate away bogus bits.
1128 Type *WidestFixupType;
1130 /// Formulae - A list of ways to build a value that can satisfy this user.
1131 /// After the list is populated, one of these is selected heuristically and
1132 /// used to formulate a replacement for OperandValToReplace in UserInst.
1133 SmallVector<Formula, 12> Formulae;
1135 /// Regs - The set of register candidates used by all formulae in this LSRUse.
1136 SmallPtrSet<const SCEV *, 4> Regs;
1138 LSRUse(KindType K, Type *T) : Kind(K), AccessTy(T),
1139 MinOffset(INT64_MAX),
1140 MaxOffset(INT64_MIN),
1141 AllFixupsOutsideLoop(true),
1142 WidestFixupType(0) {}
1144 bool HasFormulaWithSameRegs(const Formula &F) const;
1145 bool InsertFormula(const Formula &F);
1146 void DeleteFormula(Formula &F);
1147 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1149 void print(raw_ostream &OS) const;
1155 /// HasFormula - Test whether this use as a formula which has the same
1156 /// registers as the given formula.
1157 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1158 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1159 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1160 // Unstable sort by host order ok, because this is only used for uniquifying.
1161 std::sort(Key.begin(), Key.end());
1162 return Uniquifier.count(Key);
1165 /// InsertFormula - If the given formula has not yet been inserted, add it to
1166 /// the list, and return true. Return false otherwise.
1167 bool LSRUse::InsertFormula(const Formula &F) {
1168 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1169 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1170 // Unstable sort by host order ok, because this is only used for uniquifying.
1171 std::sort(Key.begin(), Key.end());
1173 if (!Uniquifier.insert(Key).second)
1176 // Using a register to hold the value of 0 is not profitable.
1177 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1178 "Zero allocated in a scaled register!");
1180 for (SmallVectorImpl<const SCEV *>::const_iterator I =
1181 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I)
1182 assert(!(*I)->isZero() && "Zero allocated in a base register!");
1185 // Add the formula to the list.
1186 Formulae.push_back(F);
1188 // Record registers now being used by this use.
1189 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1194 /// DeleteFormula - Remove the given formula from this use's list.
1195 void LSRUse::DeleteFormula(Formula &F) {
1196 if (&F != &Formulae.back())
1197 std::swap(F, Formulae.back());
1198 Formulae.pop_back();
1201 /// RecomputeRegs - Recompute the Regs field, and update RegUses.
1202 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1203 // Now that we've filtered out some formulae, recompute the Regs set.
1204 SmallPtrSet<const SCEV *, 4> OldRegs = Regs;
1206 for (SmallVectorImpl<Formula>::const_iterator I = Formulae.begin(),
1207 E = Formulae.end(); I != E; ++I) {
1208 const Formula &F = *I;
1209 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1210 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1213 // Update the RegTracker.
1214 for (SmallPtrSet<const SCEV *, 4>::iterator I = OldRegs.begin(),
1215 E = OldRegs.end(); I != E; ++I)
1216 if (!Regs.count(*I))
1217 RegUses.DropRegister(*I, LUIdx);
1220 void LSRUse::print(raw_ostream &OS) const {
1221 OS << "LSR Use: Kind=";
1223 case Basic: OS << "Basic"; break;
1224 case Special: OS << "Special"; break;
1225 case ICmpZero: OS << "ICmpZero"; break;
1227 OS << "Address of ";
1228 if (AccessTy->isPointerTy())
1229 OS << "pointer"; // the full pointer type could be really verbose
1234 OS << ", Offsets={";
1235 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
1236 E = Offsets.end(); I != E; ++I) {
1238 if (llvm::next(I) != E)
1243 if (AllFixupsOutsideLoop)
1244 OS << ", all-fixups-outside-loop";
1246 if (WidestFixupType)
1247 OS << ", widest fixup type: " << *WidestFixupType;
1250 void LSRUse::dump() const {
1251 print(errs()); errs() << '\n';
1254 /// isLegalUse - Test whether the use described by AM is "legal", meaning it can
1255 /// be completely folded into the user instruction at isel time. This includes
1256 /// address-mode folding and special icmp tricks.
1257 static bool isLegalUse(const TargetLowering::AddrMode &AM,
1258 LSRUse::KindType Kind, Type *AccessTy,
1259 const TargetLowering *TLI) {
1261 case LSRUse::Address:
1262 // If we have low-level target information, ask the target if it can
1263 // completely fold this address.
1264 if (TLI) return TLI->isLegalAddressingMode(AM, AccessTy);
1266 // Otherwise, just guess that reg+reg addressing is legal.
1267 return !AM.BaseGV && AM.BaseOffs == 0 && AM.Scale <= 1;
1269 case LSRUse::ICmpZero:
1270 // There's not even a target hook for querying whether it would be legal to
1271 // fold a GV into an ICmp.
1275 // ICmp only has two operands; don't allow more than two non-trivial parts.
1276 if (AM.Scale != 0 && AM.HasBaseReg && AM.BaseOffs != 0)
1279 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1280 // putting the scaled register in the other operand of the icmp.
1281 if (AM.Scale != 0 && AM.Scale != -1)
1284 // If we have low-level target information, ask the target if it can fold an
1285 // integer immediate on an icmp.
1286 if (AM.BaseOffs != 0) {
1287 if (TLI) return TLI->isLegalICmpImmediate(-(uint64_t)AM.BaseOffs);
1294 // Only handle single-register values.
1295 return !AM.BaseGV && AM.Scale == 0 && AM.BaseOffs == 0;
1297 case LSRUse::Special:
1298 // Only handle -1 scales, or no scale.
1299 return AM.Scale == 0 || AM.Scale == -1;
1302 llvm_unreachable("Invalid LSRUse Kind!");
1305 static bool isLegalUse(TargetLowering::AddrMode AM,
1306 int64_t MinOffset, int64_t MaxOffset,
1307 LSRUse::KindType Kind, Type *AccessTy,
1308 const TargetLowering *TLI) {
1309 // Check for overflow.
1310 if (((int64_t)((uint64_t)AM.BaseOffs + MinOffset) > AM.BaseOffs) !=
1313 AM.BaseOffs = (uint64_t)AM.BaseOffs + MinOffset;
1314 if (isLegalUse(AM, Kind, AccessTy, TLI)) {
1315 AM.BaseOffs = (uint64_t)AM.BaseOffs - MinOffset;
1316 // Check for overflow.
1317 if (((int64_t)((uint64_t)AM.BaseOffs + MaxOffset) > AM.BaseOffs) !=
1320 AM.BaseOffs = (uint64_t)AM.BaseOffs + MaxOffset;
1321 return isLegalUse(AM, Kind, AccessTy, TLI);
1326 static bool isAlwaysFoldable(int64_t BaseOffs,
1327 GlobalValue *BaseGV,
1329 LSRUse::KindType Kind, Type *AccessTy,
1330 const TargetLowering *TLI) {
1331 // Fast-path: zero is always foldable.
1332 if (BaseOffs == 0 && !BaseGV) return true;
1334 // Conservatively, create an address with an immediate and a
1335 // base and a scale.
1336 TargetLowering::AddrMode AM;
1337 AM.BaseOffs = BaseOffs;
1339 AM.HasBaseReg = HasBaseReg;
1340 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1342 // Canonicalize a scale of 1 to a base register if the formula doesn't
1343 // already have a base register.
1344 if (!AM.HasBaseReg && AM.Scale == 1) {
1346 AM.HasBaseReg = true;
1349 return isLegalUse(AM, Kind, AccessTy, TLI);
1352 static bool isAlwaysFoldable(const SCEV *S,
1353 int64_t MinOffset, int64_t MaxOffset,
1355 LSRUse::KindType Kind, Type *AccessTy,
1356 const TargetLowering *TLI,
1357 ScalarEvolution &SE) {
1358 // Fast-path: zero is always foldable.
1359 if (S->isZero()) return true;
1361 // Conservatively, create an address with an immediate and a
1362 // base and a scale.
1363 int64_t BaseOffs = ExtractImmediate(S, SE);
1364 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1366 // If there's anything else involved, it's not foldable.
1367 if (!S->isZero()) return false;
1369 // Fast-path: zero is always foldable.
1370 if (BaseOffs == 0 && !BaseGV) return true;
1372 // Conservatively, create an address with an immediate and a
1373 // base and a scale.
1374 TargetLowering::AddrMode AM;
1375 AM.BaseOffs = BaseOffs;
1377 AM.HasBaseReg = HasBaseReg;
1378 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1380 return isLegalUse(AM, MinOffset, MaxOffset, Kind, AccessTy, TLI);
1385 /// UseMapDenseMapInfo - A DenseMapInfo implementation for holding
1386 /// DenseMaps and DenseSets of pairs of const SCEV* and LSRUse::Kind.
1387 struct UseMapDenseMapInfo {
1388 static std::pair<const SCEV *, LSRUse::KindType> getEmptyKey() {
1389 return std::make_pair(reinterpret_cast<const SCEV *>(-1), LSRUse::Basic);
1392 static std::pair<const SCEV *, LSRUse::KindType> getTombstoneKey() {
1393 return std::make_pair(reinterpret_cast<const SCEV *>(-2), LSRUse::Basic);
1397 getHashValue(const std::pair<const SCEV *, LSRUse::KindType> &V) {
1398 unsigned Result = DenseMapInfo<const SCEV *>::getHashValue(V.first);
1399 Result ^= DenseMapInfo<unsigned>::getHashValue(unsigned(V.second));
1403 static bool isEqual(const std::pair<const SCEV *, LSRUse::KindType> &LHS,
1404 const std::pair<const SCEV *, LSRUse::KindType> &RHS) {
1409 /// IVInc - An individual increment in a Chain of IV increments.
1410 /// Relate an IV user to an expression that computes the IV it uses from the IV
1411 /// used by the previous link in the Chain.
1413 /// For the head of a chain, IncExpr holds the absolute SCEV expression for the
1414 /// original IVOperand. The head of the chain's IVOperand is only valid during
1415 /// chain collection, before LSR replaces IV users. During chain generation,
1416 /// IncExpr can be used to find the new IVOperand that computes the same
1419 Instruction *UserInst;
1421 const SCEV *IncExpr;
1423 IVInc(Instruction *U, Value *O, const SCEV *E):
1424 UserInst(U), IVOperand(O), IncExpr(E) {}
1427 // IVChain - The list of IV increments in program order.
1428 // We typically add the head of a chain without finding subsequent links.
1429 typedef SmallVector<IVInc,1> IVChain;
1431 /// ChainUsers - Helper for CollectChains to track multiple IV increment uses.
1432 /// Distinguish between FarUsers that definitely cross IV increments and
1433 /// NearUsers that may be used between IV increments.
1435 SmallPtrSet<Instruction*, 4> FarUsers;
1436 SmallPtrSet<Instruction*, 4> NearUsers;
1439 /// LSRInstance - This class holds state for the main loop strength reduction
1443 ScalarEvolution &SE;
1446 const TargetLowering *const TLI;
1450 /// IVIncInsertPos - This is the insert position that the current loop's
1451 /// induction variable increment should be placed. In simple loops, this is
1452 /// the latch block's terminator. But in more complicated cases, this is a
1453 /// position which will dominate all the in-loop post-increment users.
1454 Instruction *IVIncInsertPos;
1456 /// Factors - Interesting factors between use strides.
1457 SmallSetVector<int64_t, 8> Factors;
1459 /// Types - Interesting use types, to facilitate truncation reuse.
1460 SmallSetVector<Type *, 4> Types;
1462 /// Fixups - The list of operands which are to be replaced.
1463 SmallVector<LSRFixup, 16> Fixups;
1465 /// Uses - The list of interesting uses.
1466 SmallVector<LSRUse, 16> Uses;
1468 /// RegUses - Track which uses use which register candidates.
1469 RegUseTracker RegUses;
1471 // Limit the number of chains to avoid quadratic behavior. We don't expect to
1472 // have more than a few IV increment chains in a loop. Missing a Chain falls
1473 // back to normal LSR behavior for those uses.
1474 static const unsigned MaxChains = 8;
1476 /// IVChainVec - IV users can form a chain of IV increments.
1477 SmallVector<IVChain, MaxChains> IVChainVec;
1479 /// IVIncSet - IV users that belong to profitable IVChains.
1480 SmallPtrSet<Use*, MaxChains> IVIncSet;
1482 void OptimizeShadowIV();
1483 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1484 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1485 void OptimizeLoopTermCond();
1487 void ChainInstruction(Instruction *UserInst, Instruction *IVOper,
1488 SmallVectorImpl<ChainUsers> &ChainUsersVec);
1489 void FinalizeChain(IVChain &Chain);
1490 void CollectChains();
1491 void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
1492 SmallVectorImpl<WeakVH> &DeadInsts);
1494 void CollectInterestingTypesAndFactors();
1495 void CollectFixupsAndInitialFormulae();
1497 LSRFixup &getNewFixup() {
1498 Fixups.push_back(LSRFixup());
1499 return Fixups.back();
1502 // Support for sharing of LSRUses between LSRFixups.
1503 typedef DenseMap<std::pair<const SCEV *, LSRUse::KindType>,
1505 UseMapDenseMapInfo> UseMapTy;
1508 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1509 LSRUse::KindType Kind, Type *AccessTy);
1511 std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
1512 LSRUse::KindType Kind,
1515 void DeleteUse(LSRUse &LU, size_t LUIdx);
1517 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1519 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1520 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1521 void CountRegisters(const Formula &F, size_t LUIdx);
1522 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1524 void CollectLoopInvariantFixupsAndFormulae();
1526 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1527 unsigned Depth = 0);
1528 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1529 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1530 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1531 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1532 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1533 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1534 void GenerateCrossUseConstantOffsets();
1535 void GenerateAllReuseFormulae();
1537 void FilterOutUndesirableDedicatedRegisters();
1539 size_t EstimateSearchSpaceComplexity() const;
1540 void NarrowSearchSpaceByDetectingSupersets();
1541 void NarrowSearchSpaceByCollapsingUnrolledCode();
1542 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
1543 void NarrowSearchSpaceByPickingWinnerRegs();
1544 void NarrowSearchSpaceUsingHeuristics();
1546 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1548 SmallVectorImpl<const Formula *> &Workspace,
1549 const Cost &CurCost,
1550 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1551 DenseSet<const SCEV *> &VisitedRegs) const;
1552 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1554 BasicBlock::iterator
1555 HoistInsertPosition(BasicBlock::iterator IP,
1556 const SmallVectorImpl<Instruction *> &Inputs) const;
1557 BasicBlock::iterator
1558 AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1561 SCEVExpander &Rewriter) const;
1563 Value *Expand(const LSRFixup &LF,
1565 BasicBlock::iterator IP,
1566 SCEVExpander &Rewriter,
1567 SmallVectorImpl<WeakVH> &DeadInsts) const;
1568 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1570 SCEVExpander &Rewriter,
1571 SmallVectorImpl<WeakVH> &DeadInsts,
1573 void Rewrite(const LSRFixup &LF,
1575 SCEVExpander &Rewriter,
1576 SmallVectorImpl<WeakVH> &DeadInsts,
1578 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
1582 LSRInstance(const TargetLowering *tli, Loop *l, Pass *P);
1584 bool getChanged() const { return Changed; }
1586 void print_factors_and_types(raw_ostream &OS) const;
1587 void print_fixups(raw_ostream &OS) const;
1588 void print_uses(raw_ostream &OS) const;
1589 void print(raw_ostream &OS) const;
1595 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1596 /// inside the loop then try to eliminate the cast operation.
1597 void LSRInstance::OptimizeShadowIV() {
1598 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1599 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1602 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1603 UI != E; /* empty */) {
1604 IVUsers::const_iterator CandidateUI = UI;
1606 Instruction *ShadowUse = CandidateUI->getUser();
1607 Type *DestTy = NULL;
1608 bool IsSigned = false;
1610 /* If shadow use is a int->float cast then insert a second IV
1611 to eliminate this cast.
1613 for (unsigned i = 0; i < n; ++i)
1619 for (unsigned i = 0; i < n; ++i, ++d)
1622 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
1624 DestTy = UCast->getDestTy();
1626 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
1628 DestTy = SCast->getDestTy();
1630 if (!DestTy) continue;
1633 // If target does not support DestTy natively then do not apply
1634 // this transformation.
1635 EVT DVT = TLI->getValueType(DestTy);
1636 if (!TLI->isTypeLegal(DVT)) continue;
1639 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1641 if (PH->getNumIncomingValues() != 2) continue;
1643 Type *SrcTy = PH->getType();
1644 int Mantissa = DestTy->getFPMantissaWidth();
1645 if (Mantissa == -1) continue;
1646 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1649 unsigned Entry, Latch;
1650 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1658 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1659 if (!Init) continue;
1660 Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
1661 (double)Init->getSExtValue() :
1662 (double)Init->getZExtValue());
1664 BinaryOperator *Incr =
1665 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1666 if (!Incr) continue;
1667 if (Incr->getOpcode() != Instruction::Add
1668 && Incr->getOpcode() != Instruction::Sub)
1671 /* Initialize new IV, double d = 0.0 in above example. */
1672 ConstantInt *C = NULL;
1673 if (Incr->getOperand(0) == PH)
1674 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1675 else if (Incr->getOperand(1) == PH)
1676 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1682 // Ignore negative constants, as the code below doesn't handle them
1683 // correctly. TODO: Remove this restriction.
1684 if (!C->getValue().isStrictlyPositive()) continue;
1686 /* Add new PHINode. */
1687 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
1689 /* create new increment. '++d' in above example. */
1690 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1691 BinaryOperator *NewIncr =
1692 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1693 Instruction::FAdd : Instruction::FSub,
1694 NewPH, CFP, "IV.S.next.", Incr);
1696 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1697 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1699 /* Remove cast operation */
1700 ShadowUse->replaceAllUsesWith(NewPH);
1701 ShadowUse->eraseFromParent();
1707 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1708 /// set the IV user and stride information and return true, otherwise return
1710 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
1711 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1712 if (UI->getUser() == Cond) {
1713 // NOTE: we could handle setcc instructions with multiple uses here, but
1714 // InstCombine does it as well for simple uses, it's not clear that it
1715 // occurs enough in real life to handle.
1722 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
1723 /// a max computation.
1725 /// This is a narrow solution to a specific, but acute, problem. For loops
1731 /// } while (++i < n);
1733 /// the trip count isn't just 'n', because 'n' might not be positive. And
1734 /// unfortunately this can come up even for loops where the user didn't use
1735 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1736 /// will commonly be lowered like this:
1742 /// } while (++i < n);
1745 /// and then it's possible for subsequent optimization to obscure the if
1746 /// test in such a way that indvars can't find it.
1748 /// When indvars can't find the if test in loops like this, it creates a
1749 /// max expression, which allows it to give the loop a canonical
1750 /// induction variable:
1753 /// max = n < 1 ? 1 : n;
1756 /// } while (++i != max);
1758 /// Canonical induction variables are necessary because the loop passes
1759 /// are designed around them. The most obvious example of this is the
1760 /// LoopInfo analysis, which doesn't remember trip count values. It
1761 /// expects to be able to rediscover the trip count each time it is
1762 /// needed, and it does this using a simple analysis that only succeeds if
1763 /// the loop has a canonical induction variable.
1765 /// However, when it comes time to generate code, the maximum operation
1766 /// can be quite costly, especially if it's inside of an outer loop.
1768 /// This function solves this problem by detecting this type of loop and
1769 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1770 /// the instructions for the maximum computation.
1772 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1773 // Check that the loop matches the pattern we're looking for.
1774 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1775 Cond->getPredicate() != CmpInst::ICMP_NE)
1778 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1779 if (!Sel || !Sel->hasOneUse()) return Cond;
1781 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1782 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1784 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
1786 // Add one to the backedge-taken count to get the trip count.
1787 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
1788 if (IterationCount != SE.getSCEV(Sel)) return Cond;
1790 // Check for a max calculation that matches the pattern. There's no check
1791 // for ICMP_ULE here because the comparison would be with zero, which
1792 // isn't interesting.
1793 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
1794 const SCEVNAryExpr *Max = 0;
1795 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
1796 Pred = ICmpInst::ICMP_SLE;
1798 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
1799 Pred = ICmpInst::ICMP_SLT;
1801 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
1802 Pred = ICmpInst::ICMP_ULT;
1809 // To handle a max with more than two operands, this optimization would
1810 // require additional checking and setup.
1811 if (Max->getNumOperands() != 2)
1814 const SCEV *MaxLHS = Max->getOperand(0);
1815 const SCEV *MaxRHS = Max->getOperand(1);
1817 // ScalarEvolution canonicalizes constants to the left. For < and >, look
1818 // for a comparison with 1. For <= and >=, a comparison with zero.
1820 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
1823 // Check the relevant induction variable for conformance to
1825 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
1826 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1827 if (!AR || !AR->isAffine() ||
1828 AR->getStart() != One ||
1829 AR->getStepRecurrence(SE) != One)
1832 assert(AR->getLoop() == L &&
1833 "Loop condition operand is an addrec in a different loop!");
1835 // Check the right operand of the select, and remember it, as it will
1836 // be used in the new comparison instruction.
1838 if (ICmpInst::isTrueWhenEqual(Pred)) {
1839 // Look for n+1, and grab n.
1840 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
1841 if (isa<ConstantInt>(BO->getOperand(1)) &&
1842 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1843 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1844 NewRHS = BO->getOperand(0);
1845 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
1846 if (isa<ConstantInt>(BO->getOperand(1)) &&
1847 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1848 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1849 NewRHS = BO->getOperand(0);
1852 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
1853 NewRHS = Sel->getOperand(1);
1854 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
1855 NewRHS = Sel->getOperand(2);
1856 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
1857 NewRHS = SU->getValue();
1859 // Max doesn't match expected pattern.
1862 // Determine the new comparison opcode. It may be signed or unsigned,
1863 // and the original comparison may be either equality or inequality.
1864 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
1865 Pred = CmpInst::getInversePredicate(Pred);
1867 // Ok, everything looks ok to change the condition into an SLT or SGE and
1868 // delete the max calculation.
1870 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
1872 // Delete the max calculation instructions.
1873 Cond->replaceAllUsesWith(NewCond);
1874 CondUse->setUser(NewCond);
1875 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
1876 Cond->eraseFromParent();
1877 Sel->eraseFromParent();
1878 if (Cmp->use_empty())
1879 Cmp->eraseFromParent();
1883 /// OptimizeLoopTermCond - Change loop terminating condition to use the
1884 /// postinc iv when possible.
1886 LSRInstance::OptimizeLoopTermCond() {
1887 SmallPtrSet<Instruction *, 4> PostIncs;
1889 BasicBlock *LatchBlock = L->getLoopLatch();
1890 SmallVector<BasicBlock*, 8> ExitingBlocks;
1891 L->getExitingBlocks(ExitingBlocks);
1893 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
1894 BasicBlock *ExitingBlock = ExitingBlocks[i];
1896 // Get the terminating condition for the loop if possible. If we
1897 // can, we want to change it to use a post-incremented version of its
1898 // induction variable, to allow coalescing the live ranges for the IV into
1899 // one register value.
1901 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
1904 // FIXME: Overly conservative, termination condition could be an 'or' etc..
1905 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
1908 // Search IVUsesByStride to find Cond's IVUse if there is one.
1909 IVStrideUse *CondUse = 0;
1910 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1911 if (!FindIVUserForCond(Cond, CondUse))
1914 // If the trip count is computed in terms of a max (due to ScalarEvolution
1915 // being unable to find a sufficient guard, for example), change the loop
1916 // comparison to use SLT or ULT instead of NE.
1917 // One consequence of doing this now is that it disrupts the count-down
1918 // optimization. That's not always a bad thing though, because in such
1919 // cases it may still be worthwhile to avoid a max.
1920 Cond = OptimizeMax(Cond, CondUse);
1922 // If this exiting block dominates the latch block, it may also use
1923 // the post-inc value if it won't be shared with other uses.
1924 // Check for dominance.
1925 if (!DT.dominates(ExitingBlock, LatchBlock))
1928 // Conservatively avoid trying to use the post-inc value in non-latch
1929 // exits if there may be pre-inc users in intervening blocks.
1930 if (LatchBlock != ExitingBlock)
1931 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1932 // Test if the use is reachable from the exiting block. This dominator
1933 // query is a conservative approximation of reachability.
1934 if (&*UI != CondUse &&
1935 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
1936 // Conservatively assume there may be reuse if the quotient of their
1937 // strides could be a legal scale.
1938 const SCEV *A = IU.getStride(*CondUse, L);
1939 const SCEV *B = IU.getStride(*UI, L);
1940 if (!A || !B) continue;
1941 if (SE.getTypeSizeInBits(A->getType()) !=
1942 SE.getTypeSizeInBits(B->getType())) {
1943 if (SE.getTypeSizeInBits(A->getType()) >
1944 SE.getTypeSizeInBits(B->getType()))
1945 B = SE.getSignExtendExpr(B, A->getType());
1947 A = SE.getSignExtendExpr(A, B->getType());
1949 if (const SCEVConstant *D =
1950 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
1951 const ConstantInt *C = D->getValue();
1952 // Stride of one or negative one can have reuse with non-addresses.
1953 if (C->isOne() || C->isAllOnesValue())
1954 goto decline_post_inc;
1955 // Avoid weird situations.
1956 if (C->getValue().getMinSignedBits() >= 64 ||
1957 C->getValue().isMinSignedValue())
1958 goto decline_post_inc;
1959 // Without TLI, assume that any stride might be valid, and so any
1960 // use might be shared.
1962 goto decline_post_inc;
1963 // Check for possible scaled-address reuse.
1964 Type *AccessTy = getAccessType(UI->getUser());
1965 TargetLowering::AddrMode AM;
1966 AM.Scale = C->getSExtValue();
1967 if (TLI->isLegalAddressingMode(AM, AccessTy))
1968 goto decline_post_inc;
1969 AM.Scale = -AM.Scale;
1970 if (TLI->isLegalAddressingMode(AM, AccessTy))
1971 goto decline_post_inc;
1975 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
1978 // It's possible for the setcc instruction to be anywhere in the loop, and
1979 // possible for it to have multiple users. If it is not immediately before
1980 // the exiting block branch, move it.
1981 if (&*++BasicBlock::iterator(Cond) != TermBr) {
1982 if (Cond->hasOneUse()) {
1983 Cond->moveBefore(TermBr);
1985 // Clone the terminating condition and insert into the loopend.
1986 ICmpInst *OldCond = Cond;
1987 Cond = cast<ICmpInst>(Cond->clone());
1988 Cond->setName(L->getHeader()->getName() + ".termcond");
1989 ExitingBlock->getInstList().insert(TermBr, Cond);
1991 // Clone the IVUse, as the old use still exists!
1992 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
1993 TermBr->replaceUsesOfWith(OldCond, Cond);
1997 // If we get to here, we know that we can transform the setcc instruction to
1998 // use the post-incremented version of the IV, allowing us to coalesce the
1999 // live ranges for the IV correctly.
2000 CondUse->transformToPostInc(L);
2003 PostIncs.insert(Cond);
2007 // Determine an insertion point for the loop induction variable increment. It
2008 // must dominate all the post-inc comparisons we just set up, and it must
2009 // dominate the loop latch edge.
2010 IVIncInsertPos = L->getLoopLatch()->getTerminator();
2011 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(),
2012 E = PostIncs.end(); I != E; ++I) {
2014 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
2016 if (BB == (*I)->getParent())
2017 IVIncInsertPos = *I;
2018 else if (BB != IVIncInsertPos->getParent())
2019 IVIncInsertPos = BB->getTerminator();
2023 /// reconcileNewOffset - Determine if the given use can accommodate a fixup
2024 /// at the given offset and other details. If so, update the use and
2027 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
2028 LSRUse::KindType Kind, Type *AccessTy) {
2029 int64_t NewMinOffset = LU.MinOffset;
2030 int64_t NewMaxOffset = LU.MaxOffset;
2031 Type *NewAccessTy = AccessTy;
2033 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
2034 // something conservative, however this can pessimize in the case that one of
2035 // the uses will have all its uses outside the loop, for example.
2036 if (LU.Kind != Kind)
2038 // Conservatively assume HasBaseReg is true for now.
2039 if (NewOffset < LU.MinOffset) {
2040 if (!isAlwaysFoldable(LU.MaxOffset - NewOffset, 0, HasBaseReg,
2041 Kind, AccessTy, TLI))
2043 NewMinOffset = NewOffset;
2044 } else if (NewOffset > LU.MaxOffset) {
2045 if (!isAlwaysFoldable(NewOffset - LU.MinOffset, 0, HasBaseReg,
2046 Kind, AccessTy, TLI))
2048 NewMaxOffset = NewOffset;
2050 // Check for a mismatched access type, and fall back conservatively as needed.
2051 // TODO: Be less conservative when the type is similar and can use the same
2052 // addressing modes.
2053 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
2054 NewAccessTy = Type::getVoidTy(AccessTy->getContext());
2057 LU.MinOffset = NewMinOffset;
2058 LU.MaxOffset = NewMaxOffset;
2059 LU.AccessTy = NewAccessTy;
2060 if (NewOffset != LU.Offsets.back())
2061 LU.Offsets.push_back(NewOffset);
2065 /// getUse - Return an LSRUse index and an offset value for a fixup which
2066 /// needs the given expression, with the given kind and optional access type.
2067 /// Either reuse an existing use or create a new one, as needed.
2068 std::pair<size_t, int64_t>
2069 LSRInstance::getUse(const SCEV *&Expr,
2070 LSRUse::KindType Kind, Type *AccessTy) {
2071 const SCEV *Copy = Expr;
2072 int64_t Offset = ExtractImmediate(Expr, SE);
2074 // Basic uses can't accept any offset, for example.
2075 if (!isAlwaysFoldable(Offset, 0, /*HasBaseReg=*/true, Kind, AccessTy, TLI)) {
2080 std::pair<UseMapTy::iterator, bool> P =
2081 UseMap.insert(std::make_pair(std::make_pair(Expr, Kind), 0));
2083 // A use already existed with this base.
2084 size_t LUIdx = P.first->second;
2085 LSRUse &LU = Uses[LUIdx];
2086 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
2088 return std::make_pair(LUIdx, Offset);
2091 // Create a new use.
2092 size_t LUIdx = Uses.size();
2093 P.first->second = LUIdx;
2094 Uses.push_back(LSRUse(Kind, AccessTy));
2095 LSRUse &LU = Uses[LUIdx];
2097 // We don't need to track redundant offsets, but we don't need to go out
2098 // of our way here to avoid them.
2099 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
2100 LU.Offsets.push_back(Offset);
2102 LU.MinOffset = Offset;
2103 LU.MaxOffset = Offset;
2104 return std::make_pair(LUIdx, Offset);
2107 /// DeleteUse - Delete the given use from the Uses list.
2108 void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
2109 if (&LU != &Uses.back())
2110 std::swap(LU, Uses.back());
2114 RegUses.SwapAndDropUse(LUIdx, Uses.size());
2117 /// FindUseWithFormula - Look for a use distinct from OrigLU which is has
2118 /// a formula that has the same registers as the given formula.
2120 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
2121 const LSRUse &OrigLU) {
2122 // Search all uses for the formula. This could be more clever.
2123 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2124 LSRUse &LU = Uses[LUIdx];
2125 // Check whether this use is close enough to OrigLU, to see whether it's
2126 // worthwhile looking through its formulae.
2127 // Ignore ICmpZero uses because they may contain formulae generated by
2128 // GenerateICmpZeroScales, in which case adding fixup offsets may
2130 if (&LU != &OrigLU &&
2131 LU.Kind != LSRUse::ICmpZero &&
2132 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
2133 LU.WidestFixupType == OrigLU.WidestFixupType &&
2134 LU.HasFormulaWithSameRegs(OrigF)) {
2135 // Scan through this use's formulae.
2136 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
2137 E = LU.Formulae.end(); I != E; ++I) {
2138 const Formula &F = *I;
2139 // Check to see if this formula has the same registers and symbols
2141 if (F.BaseRegs == OrigF.BaseRegs &&
2142 F.ScaledReg == OrigF.ScaledReg &&
2143 F.AM.BaseGV == OrigF.AM.BaseGV &&
2144 F.AM.Scale == OrigF.AM.Scale &&
2145 F.UnfoldedOffset == OrigF.UnfoldedOffset) {
2146 if (F.AM.BaseOffs == 0)
2148 // This is the formula where all the registers and symbols matched;
2149 // there aren't going to be any others. Since we declined it, we
2150 // can skip the rest of the formulae and procede to the next LSRUse.
2157 // Nothing looked good.
2161 void LSRInstance::CollectInterestingTypesAndFactors() {
2162 SmallSetVector<const SCEV *, 4> Strides;
2164 // Collect interesting types and strides.
2165 SmallVector<const SCEV *, 4> Worklist;
2166 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2167 const SCEV *Expr = IU.getExpr(*UI);
2169 // Collect interesting types.
2170 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2172 // Add strides for mentioned loops.
2173 Worklist.push_back(Expr);
2175 const SCEV *S = Worklist.pop_back_val();
2176 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2177 if (AR->getLoop() == L)
2178 Strides.insert(AR->getStepRecurrence(SE));
2179 Worklist.push_back(AR->getStart());
2180 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2181 Worklist.append(Add->op_begin(), Add->op_end());
2183 } while (!Worklist.empty());
2186 // Compute interesting factors from the set of interesting strides.
2187 for (SmallSetVector<const SCEV *, 4>::const_iterator
2188 I = Strides.begin(), E = Strides.end(); I != E; ++I)
2189 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
2190 llvm::next(I); NewStrideIter != E; ++NewStrideIter) {
2191 const SCEV *OldStride = *I;
2192 const SCEV *NewStride = *NewStrideIter;
2194 if (SE.getTypeSizeInBits(OldStride->getType()) !=
2195 SE.getTypeSizeInBits(NewStride->getType())) {
2196 if (SE.getTypeSizeInBits(OldStride->getType()) >
2197 SE.getTypeSizeInBits(NewStride->getType()))
2198 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2200 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2202 if (const SCEVConstant *Factor =
2203 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2205 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2206 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2207 } else if (const SCEVConstant *Factor =
2208 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2211 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2212 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2216 // If all uses use the same type, don't bother looking for truncation-based
2218 if (Types.size() == 1)
2221 DEBUG(print_factors_and_types(dbgs()));
2224 /// findIVOperand - Helper for CollectChains that finds an IV operand (computed
2225 /// by an AddRec in this loop) within [OI,OE) or returns OE. If IVUsers mapped
2226 /// Instructions to IVStrideUses, we could partially skip this.
2227 static User::op_iterator
2228 findIVOperand(User::op_iterator OI, User::op_iterator OE,
2229 Loop *L, ScalarEvolution &SE) {
2230 for(; OI != OE; ++OI) {
2231 if (Instruction *Oper = dyn_cast<Instruction>(*OI)) {
2232 if (!SE.isSCEVable(Oper->getType()))
2235 if (const SCEVAddRecExpr *AR =
2236 dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) {
2237 if (AR->getLoop() == L)
2245 /// getWideOperand - IVChain logic must consistenctly peek base TruncInst
2246 /// operands, so wrap it in a convenient helper.
2247 static Value *getWideOperand(Value *Oper) {
2248 if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper))
2249 return Trunc->getOperand(0);
2253 /// isCompatibleIVType - Return true if we allow an IV chain to include both
2255 static bool isCompatibleIVType(Value *LVal, Value *RVal) {
2256 Type *LType = LVal->getType();
2257 Type *RType = RVal->getType();
2258 return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy());
2261 /// getExprBase - Return an approximation of this SCEV expression's "base", or
2262 /// NULL for any constant. Returning the expression itself is
2263 /// conservative. Returning a deeper subexpression is more precise and valid as
2264 /// long as it isn't less complex than another subexpression. For expressions
2265 /// involving multiple unscaled values, we need to return the pointer-type
2266 /// SCEVUnknown. This avoids forming chains across objects, such as:
2267 /// PrevOper==a[i], IVOper==b[i], IVInc==b-a.
2269 /// Since SCEVUnknown is the rightmost type, and pointers are the rightmost
2270 /// SCEVUnknown, we simply return the rightmost SCEV operand.
2271 static const SCEV *getExprBase(const SCEV *S) {
2272 switch (S->getSCEVType()) {
2273 default: // uncluding scUnknown.
2278 return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand());
2280 return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand());
2282 return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand());
2284 // Skip over scaled operands (scMulExpr) to follow add operands as long as
2285 // there's nothing more complex.
2286 // FIXME: not sure if we want to recognize negation.
2287 const SCEVAddExpr *Add = cast<SCEVAddExpr>(S);
2288 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(Add->op_end()),
2289 E(Add->op_begin()); I != E; ++I) {
2290 const SCEV *SubExpr = *I;
2291 if (SubExpr->getSCEVType() == scAddExpr)
2292 return getExprBase(SubExpr);
2294 if (SubExpr->getSCEVType() != scMulExpr)
2297 return S; // all operands are scaled, be conservative.
2300 return getExprBase(cast<SCEVAddRecExpr>(S)->getStart());
2304 /// Return true if the chain increment is profitable to expand into a loop
2305 /// invariant value, which may require its own register. A profitable chain
2306 /// increment will be an offset relative to the same base. We allow such offsets
2307 /// to potentially be used as chain increment as long as it's not obviously
2308 /// expensive to expand using real instructions.
2310 getProfitableChainIncrement(Value *NextIV, Value *PrevIV,
2311 const IVChain &Chain, Loop *L,
2312 ScalarEvolution &SE, const TargetLowering *TLI) {
2313 // Prune the solution space aggressively by checking that both IV operands
2314 // are expressions that operate on the same unscaled SCEVUnknown. This
2315 // "base" will be canceled by the subsequent getMinusSCEV call. Checking first
2316 // avoids creating extra SCEV expressions.
2317 const SCEV *OperExpr = SE.getSCEV(NextIV);
2318 const SCEV *PrevExpr = SE.getSCEV(PrevIV);
2319 if (getExprBase(OperExpr) != getExprBase(PrevExpr) && !StressIVChain)
2322 const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr);
2323 if (!SE.isLoopInvariant(IncExpr, L))
2326 // We are not able to expand an increment unless it is loop invariant,
2327 // however, the following checks are purely for profitability.
2331 // Do not replace a constant offset from IV head with a nonconstant IV
2333 if (!isa<SCEVConstant>(IncExpr)) {
2334 const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Chain[0].IVOperand));
2335 if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr)))
2339 SmallPtrSet<const SCEV*, 8> Processed;
2340 if (isHighCostExpansion(IncExpr, Processed, SE))
2346 /// Return true if the number of registers needed for the chain is estimated to
2347 /// be less than the number required for the individual IV users. First prohibit
2348 /// any IV users that keep the IV live across increments (the Users set should
2349 /// be empty). Next count the number and type of increments in the chain.
2351 /// Chaining IVs can lead to considerable code bloat if ISEL doesn't
2352 /// effectively use postinc addressing modes. Only consider it profitable it the
2353 /// increments can be computed in fewer registers when chained.
2355 /// TODO: Consider IVInc free if it's already used in another chains.
2357 isProfitableChain(IVChain &Chain, SmallPtrSet<Instruction*, 4> &Users,
2358 ScalarEvolution &SE, const TargetLowering *TLI) {
2362 if (Chain.size() <= 2)
2365 if (!Users.empty()) {
2366 DEBUG(dbgs() << "Chain: " << *Chain[0].UserInst << " users:\n";
2367 for (SmallPtrSet<Instruction*, 4>::const_iterator I = Users.begin(),
2368 E = Users.end(); I != E; ++I) {
2369 dbgs() << " " << **I << "\n";
2373 assert(!Chain.empty() && "empty IV chains are not allowed");
2375 // The chain itself may require a register, so intialize cost to 1.
2378 // A complete chain likely eliminates the need for keeping the original IV in
2379 // a register. LSR does not currently know how to form a complete chain unless
2380 // the header phi already exists.
2381 if (isa<PHINode>(Chain.back().UserInst)
2382 && SE.getSCEV(Chain.back().UserInst) == Chain[0].IncExpr) {
2385 const SCEV *LastIncExpr = 0;
2386 unsigned NumConstIncrements = 0;
2387 unsigned NumVarIncrements = 0;
2388 unsigned NumReusedIncrements = 0;
2389 for (IVChain::const_iterator I = llvm::next(Chain.begin()), E = Chain.end();
2392 if (I->IncExpr->isZero())
2395 // Incrementing by zero or some constant is neutral. We assume constants can
2396 // be folded into an addressing mode or an add's immediate operand.
2397 if (isa<SCEVConstant>(I->IncExpr)) {
2398 ++NumConstIncrements;
2402 if (I->IncExpr == LastIncExpr)
2403 ++NumReusedIncrements;
2407 LastIncExpr = I->IncExpr;
2409 // An IV chain with a single increment is handled by LSR's postinc
2410 // uses. However, a chain with multiple increments requires keeping the IV's
2411 // value live longer than it needs to be if chained.
2412 if (NumConstIncrements > 1)
2415 // Materializing increment expressions in the preheader that didn't exist in
2416 // the original code may cost a register. For example, sign-extended array
2417 // indices can produce ridiculous increments like this:
2418 // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64)))
2419 cost += NumVarIncrements;
2421 // Reusing variable increments likely saves a register to hold the multiple of
2423 cost -= NumReusedIncrements;
2425 DEBUG(dbgs() << "Chain: " << *Chain[0].UserInst << " Cost: " << cost << "\n");
2430 /// ChainInstruction - Add this IV user to an existing chain or make it the head
2432 void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper,
2433 SmallVectorImpl<ChainUsers> &ChainUsersVec) {
2434 // When IVs are used as types of varying widths, they are generally converted
2435 // to a wider type with some uses remaining narrow under a (free) trunc.
2436 Value *NextIV = getWideOperand(IVOper);
2438 // Visit all existing chains. Check if its IVOper can be computed as a
2439 // profitable loop invariant increment from the last link in the Chain.
2440 unsigned ChainIdx = 0, NChains = IVChainVec.size();
2441 const SCEV *LastIncExpr = 0;
2442 for (; ChainIdx < NChains; ++ChainIdx) {
2443 Value *PrevIV = getWideOperand(IVChainVec[ChainIdx].back().IVOperand);
2444 if (!isCompatibleIVType(PrevIV, NextIV))
2447 // A phi nodes terminates a chain.
2448 if (isa<PHINode>(UserInst)
2449 && isa<PHINode>(IVChainVec[ChainIdx].back().UserInst))
2452 if (const SCEV *IncExpr =
2453 getProfitableChainIncrement(NextIV, PrevIV, IVChainVec[ChainIdx],
2455 LastIncExpr = IncExpr;
2459 // If we haven't found a chain, create a new one, unless we hit the max. Don't
2460 // bother for phi nodes, because they must be last in the chain.
2461 if (ChainIdx == NChains) {
2462 if (isa<PHINode>(UserInst))
2464 if (NChains >= MaxChains && !StressIVChain) {
2465 DEBUG(dbgs() << "IV Chain Limit\n");
2468 LastIncExpr = SE.getSCEV(NextIV);
2469 // IVUsers may have skipped over sign/zero extensions. We don't currently
2470 // attempt to form chains involving extensions unless they can be hoisted
2471 // into this loop's AddRec.
2472 if (!isa<SCEVAddRecExpr>(LastIncExpr))
2475 IVChainVec.resize(NChains);
2476 ChainUsersVec.resize(NChains);
2477 DEBUG(dbgs() << "IV Head: (" << *UserInst << ") IV=" << *LastIncExpr
2481 DEBUG(dbgs() << "IV Inc: (" << *UserInst << ") IV+" << *LastIncExpr
2484 // Add this IV user to the end of the chain.
2485 IVChainVec[ChainIdx].push_back(IVInc(UserInst, IVOper, LastIncExpr));
2487 SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers;
2488 // This chain's NearUsers become FarUsers.
2489 if (!LastIncExpr->isZero()) {
2490 ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(),
2495 // All other uses of IVOperand become near uses of the chain.
2496 // We currently ignore intermediate values within SCEV expressions, assuming
2497 // they will eventually be used be the current chain, or can be computed
2498 // from one of the chain increments. To be more precise we could
2499 // transitively follow its user and only add leaf IV users to the set.
2500 for (Value::use_iterator UseIter = IVOper->use_begin(),
2501 UseEnd = IVOper->use_end(); UseIter != UseEnd; ++UseIter) {
2502 Instruction *OtherUse = dyn_cast<Instruction>(*UseIter);
2503 if (SE.isSCEVable(OtherUse->getType())
2504 && !isa<SCEVUnknown>(SE.getSCEV(OtherUse))
2505 && IU.isIVUserOrOperand(OtherUse)) {
2508 if (OtherUse && OtherUse != UserInst)
2509 NearUsers.insert(OtherUse);
2512 // Since this user is part of the chain, it's no longer considered a use
2514 ChainUsersVec[ChainIdx].FarUsers.erase(UserInst);
2517 /// CollectChains - Populate the vector of Chains.
2519 /// This decreases ILP at the architecture level. Targets with ample registers,
2520 /// multiple memory ports, and no register renaming probably don't want
2521 /// this. However, such targets should probably disable LSR altogether.
2523 /// The job of LSR is to make a reasonable choice of induction variables across
2524 /// the loop. Subsequent passes can easily "unchain" computation exposing more
2525 /// ILP *within the loop* if the target wants it.
2527 /// Finding the best IV chain is potentially a scheduling problem. Since LSR
2528 /// will not reorder memory operations, it will recognize this as a chain, but
2529 /// will generate redundant IV increments. Ideally this would be corrected later
2530 /// by a smart scheduler:
2536 /// TODO: Walk the entire domtree within this loop, not just the path to the
2537 /// loop latch. This will discover chains on side paths, but requires
2538 /// maintaining multiple copies of the Chains state.
2539 void LSRInstance::CollectChains() {
2540 SmallVector<ChainUsers, 8> ChainUsersVec;
2542 SmallVector<BasicBlock *,8> LatchPath;
2543 BasicBlock *LoopHeader = L->getHeader();
2544 for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch());
2545 Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) {
2546 LatchPath.push_back(Rung->getBlock());
2548 LatchPath.push_back(LoopHeader);
2550 // Walk the instruction stream from the loop header to the loop latch.
2551 for (SmallVectorImpl<BasicBlock *>::reverse_iterator
2552 BBIter = LatchPath.rbegin(), BBEnd = LatchPath.rend();
2553 BBIter != BBEnd; ++BBIter) {
2554 for (BasicBlock::iterator I = (*BBIter)->begin(), E = (*BBIter)->end();
2556 // Skip instructions that weren't seen by IVUsers analysis.
2557 if (isa<PHINode>(I) || !IU.isIVUserOrOperand(I))
2560 // Ignore users that are part of a SCEV expression. This way we only
2561 // consider leaf IV Users. This effectively rediscovers a portion of
2562 // IVUsers analysis but in program order this time.
2563 if (SE.isSCEVable(I->getType()) && !isa<SCEVUnknown>(SE.getSCEV(I)))
2566 // Remove this instruction from any NearUsers set it may be in.
2567 for (unsigned ChainIdx = 0, NChains = IVChainVec.size();
2568 ChainIdx < NChains; ++ChainIdx) {
2569 ChainUsersVec[ChainIdx].NearUsers.erase(I);
2571 // Search for operands that can be chained.
2572 SmallPtrSet<Instruction*, 4> UniqueOperands;
2573 User::op_iterator IVOpEnd = I->op_end();
2574 User::op_iterator IVOpIter = findIVOperand(I->op_begin(), IVOpEnd, L, SE);
2575 while (IVOpIter != IVOpEnd) {
2576 Instruction *IVOpInst = cast<Instruction>(*IVOpIter);
2577 if (UniqueOperands.insert(IVOpInst))
2578 ChainInstruction(I, IVOpInst, ChainUsersVec);
2579 IVOpIter = findIVOperand(llvm::next(IVOpIter), IVOpEnd, L, SE);
2581 } // Continue walking down the instructions.
2582 } // Continue walking down the domtree.
2583 // Visit phi backedges to determine if the chain can generate the IV postinc.
2584 for (BasicBlock::iterator I = L->getHeader()->begin();
2585 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2586 if (!SE.isSCEVable(PN->getType()))
2590 dyn_cast<Instruction>(PN->getIncomingValueForBlock(L->getLoopLatch()));
2592 ChainInstruction(PN, IncV, ChainUsersVec);
2594 // Remove any unprofitable chains.
2595 unsigned ChainIdx = 0;
2596 for (unsigned UsersIdx = 0, NChains = IVChainVec.size();
2597 UsersIdx < NChains; ++UsersIdx) {
2598 if (!isProfitableChain(IVChainVec[UsersIdx],
2599 ChainUsersVec[UsersIdx].FarUsers, SE, TLI))
2601 // Preserve the chain at UsesIdx.
2602 if (ChainIdx != UsersIdx)
2603 IVChainVec[ChainIdx] = IVChainVec[UsersIdx];
2604 FinalizeChain(IVChainVec[ChainIdx]);
2607 IVChainVec.resize(ChainIdx);
2610 void LSRInstance::FinalizeChain(IVChain &Chain) {
2611 assert(!Chain.empty() && "empty IV chains are not allowed");
2612 DEBUG(dbgs() << "Final Chain: " << *Chain[0].UserInst << "\n");
2614 for (IVChain::const_iterator I = llvm::next(Chain.begin()), E = Chain.end();
2616 DEBUG(dbgs() << " Inc: " << *I->UserInst << "\n");
2617 User::op_iterator UseI =
2618 std::find(I->UserInst->op_begin(), I->UserInst->op_end(), I->IVOperand);
2619 assert(UseI != I->UserInst->op_end() && "cannot find IV operand");
2620 IVIncSet.insert(UseI);
2624 /// Return true if the IVInc can be folded into an addressing mode.
2625 static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst,
2626 Value *Operand, const TargetLowering *TLI) {
2627 const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr);
2628 if (!IncConst || !isAddressUse(UserInst, Operand))
2631 if (IncConst->getValue()->getValue().getMinSignedBits() > 64)
2634 int64_t IncOffset = IncConst->getValue()->getSExtValue();
2635 if (!isAlwaysFoldable(IncOffset, /*BaseGV=*/0, /*HaseBaseReg=*/false,
2636 LSRUse::Address, getAccessType(UserInst), TLI))
2642 /// GenerateIVChains - Generate an add or subtract for each IVInc in a chain to
2643 /// materialize the IV user's operand from the previous IV user's operand.
2644 void LSRInstance::GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
2645 SmallVectorImpl<WeakVH> &DeadInsts) {
2646 // Find the new IVOperand for the head of the chain. It may have been replaced
2648 const IVInc &Head = Chain[0];
2649 User::op_iterator IVOpEnd = Head.UserInst->op_end();
2650 User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(),
2653 while (IVOpIter != IVOpEnd) {
2654 IVSrc = getWideOperand(*IVOpIter);
2656 // If this operand computes the expression that the chain needs, we may use
2657 // it. (Check this after setting IVSrc which is used below.)
2659 // Note that if Head.IncExpr is wider than IVSrc, then this phi is too
2660 // narrow for the chain, so we can no longer use it. We do allow using a
2661 // wider phi, assuming the LSR checked for free truncation. In that case we
2662 // should already have a truncate on this operand such that
2663 // getSCEV(IVSrc) == IncExpr.
2664 if (SE.getSCEV(*IVOpIter) == Head.IncExpr
2665 || SE.getSCEV(IVSrc) == Head.IncExpr) {
2668 IVOpIter = findIVOperand(llvm::next(IVOpIter), IVOpEnd, L, SE);
2670 if (IVOpIter == IVOpEnd) {
2671 // Gracefully give up on this chain.
2672 DEBUG(dbgs() << "Concealed chain head: " << *Head.UserInst << "\n");
2676 DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n");
2677 Type *IVTy = IVSrc->getType();
2678 Type *IntTy = SE.getEffectiveSCEVType(IVTy);
2679 const SCEV *LeftOverExpr = 0;
2680 for (IVChain::const_iterator IncI = llvm::next(Chain.begin()),
2681 IncE = Chain.end(); IncI != IncE; ++IncI) {
2683 Instruction *InsertPt = IncI->UserInst;
2684 if (isa<PHINode>(InsertPt))
2685 InsertPt = L->getLoopLatch()->getTerminator();
2687 // IVOper will replace the current IV User's operand. IVSrc is the IV
2688 // value currently held in a register.
2689 Value *IVOper = IVSrc;
2690 if (!IncI->IncExpr->isZero()) {
2691 // IncExpr was the result of subtraction of two narrow values, so must
2693 const SCEV *IncExpr = SE.getNoopOrSignExtend(IncI->IncExpr, IntTy);
2694 LeftOverExpr = LeftOverExpr ?
2695 SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr;
2697 if (LeftOverExpr && !LeftOverExpr->isZero()) {
2698 // Expand the IV increment.
2699 Rewriter.clearPostInc();
2700 Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt);
2701 const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc),
2702 SE.getUnknown(IncV));
2703 IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt);
2705 // If an IV increment can't be folded, use it as the next IV value.
2706 if (!canFoldIVIncExpr(LeftOverExpr, IncI->UserInst, IncI->IVOperand,
2708 assert(IVTy == IVOper->getType() && "inconsistent IV increment type");
2713 Type *OperTy = IncI->IVOperand->getType();
2714 if (IVTy != OperTy) {
2715 assert(SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) &&
2716 "cannot extend a chained IV");
2717 IRBuilder<> Builder(InsertPt);
2718 IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain");
2720 IncI->UserInst->replaceUsesOfWith(IncI->IVOperand, IVOper);
2721 DeadInsts.push_back(IncI->IVOperand);
2723 // If LSR created a new, wider phi, we may also replace its postinc. We only
2724 // do this if we also found a wide value for the head of the chain.
2725 if (isa<PHINode>(Chain.back().UserInst)) {
2726 for (BasicBlock::iterator I = L->getHeader()->begin();
2727 PHINode *Phi = dyn_cast<PHINode>(I); ++I) {
2728 if (!isCompatibleIVType(Phi, IVSrc))
2730 Instruction *PostIncV = dyn_cast<Instruction>(
2731 Phi->getIncomingValueForBlock(L->getLoopLatch()));
2732 if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc)))
2734 Value *IVOper = IVSrc;
2735 Type *PostIncTy = PostIncV->getType();
2736 if (IVTy != PostIncTy) {
2737 assert(PostIncTy->isPointerTy() && "mixing int/ptr IV types");
2738 IRBuilder<> Builder(L->getLoopLatch()->getTerminator());
2739 Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc());
2740 IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain");
2742 Phi->replaceUsesOfWith(PostIncV, IVOper);
2743 DeadInsts.push_back(PostIncV);
2748 void LSRInstance::CollectFixupsAndInitialFormulae() {
2749 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2750 Instruction *UserInst = UI->getUser();
2751 // Skip IV users that are part of profitable IV Chains.
2752 User::op_iterator UseI = std::find(UserInst->op_begin(), UserInst->op_end(),
2753 UI->getOperandValToReplace());
2754 assert(UseI != UserInst->op_end() && "cannot find IV operand");
2755 if (IVIncSet.count(UseI))
2759 LSRFixup &LF = getNewFixup();
2760 LF.UserInst = UserInst;
2761 LF.OperandValToReplace = UI->getOperandValToReplace();
2762 LF.PostIncLoops = UI->getPostIncLoops();
2764 LSRUse::KindType Kind = LSRUse::Basic;
2766 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
2767 Kind = LSRUse::Address;
2768 AccessTy = getAccessType(LF.UserInst);
2771 const SCEV *S = IU.getExpr(*UI);
2773 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
2774 // (N - i == 0), and this allows (N - i) to be the expression that we work
2775 // with rather than just N or i, so we can consider the register
2776 // requirements for both N and i at the same time. Limiting this code to
2777 // equality icmps is not a problem because all interesting loops use
2778 // equality icmps, thanks to IndVarSimplify.
2779 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
2780 if (CI->isEquality()) {
2781 // Swap the operands if needed to put the OperandValToReplace on the
2782 // left, for consistency.
2783 Value *NV = CI->getOperand(1);
2784 if (NV == LF.OperandValToReplace) {
2785 CI->setOperand(1, CI->getOperand(0));
2786 CI->setOperand(0, NV);
2787 NV = CI->getOperand(1);
2791 // x == y --> x - y == 0
2792 const SCEV *N = SE.getSCEV(NV);
2793 if (SE.isLoopInvariant(N, L)) {
2794 // S is normalized, so normalize N before folding it into S
2795 // to keep the result normalized.
2796 N = TransformForPostIncUse(Normalize, N, CI, 0,
2797 LF.PostIncLoops, SE, DT);
2798 Kind = LSRUse::ICmpZero;
2799 S = SE.getMinusSCEV(N, S);
2802 // -1 and the negations of all interesting strides (except the negation
2803 // of -1) are now also interesting.
2804 for (size_t i = 0, e = Factors.size(); i != e; ++i)
2805 if (Factors[i] != -1)
2806 Factors.insert(-(uint64_t)Factors[i]);
2810 // Set up the initial formula for this use.
2811 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
2813 LF.Offset = P.second;
2814 LSRUse &LU = Uses[LF.LUIdx];
2815 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2816 if (!LU.WidestFixupType ||
2817 SE.getTypeSizeInBits(LU.WidestFixupType) <
2818 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2819 LU.WidestFixupType = LF.OperandValToReplace->getType();
2821 // If this is the first use of this LSRUse, give it a formula.
2822 if (LU.Formulae.empty()) {
2823 InsertInitialFormula(S, LU, LF.LUIdx);
2824 CountRegisters(LU.Formulae.back(), LF.LUIdx);
2828 DEBUG(print_fixups(dbgs()));
2831 /// InsertInitialFormula - Insert a formula for the given expression into
2832 /// the given use, separating out loop-variant portions from loop-invariant
2833 /// and loop-computable portions.
2835 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
2837 F.InitialMatch(S, L, SE);
2838 bool Inserted = InsertFormula(LU, LUIdx, F);
2839 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
2842 /// InsertSupplementalFormula - Insert a simple single-register formula for
2843 /// the given expression into the given use.
2845 LSRInstance::InsertSupplementalFormula(const SCEV *S,
2846 LSRUse &LU, size_t LUIdx) {
2848 F.BaseRegs.push_back(S);
2849 F.AM.HasBaseReg = true;
2850 bool Inserted = InsertFormula(LU, LUIdx, F);
2851 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
2854 /// CountRegisters - Note which registers are used by the given formula,
2855 /// updating RegUses.
2856 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
2858 RegUses.CountRegister(F.ScaledReg, LUIdx);
2859 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
2860 E = F.BaseRegs.end(); I != E; ++I)
2861 RegUses.CountRegister(*I, LUIdx);
2864 /// InsertFormula - If the given formula has not yet been inserted, add it to
2865 /// the list, and return true. Return false otherwise.
2866 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
2867 if (!LU.InsertFormula(F))
2870 CountRegisters(F, LUIdx);
2874 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of
2875 /// loop-invariant values which we're tracking. These other uses will pin these
2876 /// values in registers, making them less profitable for elimination.
2877 /// TODO: This currently misses non-constant addrec step registers.
2878 /// TODO: Should this give more weight to users inside the loop?
2880 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
2881 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
2882 SmallPtrSet<const SCEV *, 8> Inserted;
2884 while (!Worklist.empty()) {
2885 const SCEV *S = Worklist.pop_back_val();
2887 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
2888 Worklist.append(N->op_begin(), N->op_end());
2889 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
2890 Worklist.push_back(C->getOperand());
2891 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
2892 Worklist.push_back(D->getLHS());
2893 Worklist.push_back(D->getRHS());
2894 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2895 if (!Inserted.insert(U)) continue;
2896 const Value *V = U->getValue();
2897 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
2898 // Look for instructions defined outside the loop.
2899 if (L->contains(Inst)) continue;
2900 } else if (isa<UndefValue>(V))
2901 // Undef doesn't have a live range, so it doesn't matter.
2903 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
2905 const Instruction *UserInst = dyn_cast<Instruction>(*UI);
2906 // Ignore non-instructions.
2909 // Ignore instructions in other functions (as can happen with
2911 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
2913 // Ignore instructions not dominated by the loop.
2914 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
2915 UserInst->getParent() :
2916 cast<PHINode>(UserInst)->getIncomingBlock(
2917 PHINode::getIncomingValueNumForOperand(UI.getOperandNo()));
2918 if (!DT.dominates(L->getHeader(), UseBB))
2920 // Ignore uses which are part of other SCEV expressions, to avoid
2921 // analyzing them multiple times.
2922 if (SE.isSCEVable(UserInst->getType())) {
2923 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
2924 // If the user is a no-op, look through to its uses.
2925 if (!isa<SCEVUnknown>(UserS))
2929 SE.getUnknown(const_cast<Instruction *>(UserInst)));
2933 // Ignore icmp instructions which are already being analyzed.
2934 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
2935 unsigned OtherIdx = !UI.getOperandNo();
2936 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
2937 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
2941 LSRFixup &LF = getNewFixup();
2942 LF.UserInst = const_cast<Instruction *>(UserInst);
2943 LF.OperandValToReplace = UI.getUse();
2944 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0);
2946 LF.Offset = P.second;
2947 LSRUse &LU = Uses[LF.LUIdx];
2948 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2949 if (!LU.WidestFixupType ||
2950 SE.getTypeSizeInBits(LU.WidestFixupType) <
2951 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2952 LU.WidestFixupType = LF.OperandValToReplace->getType();
2953 InsertSupplementalFormula(U, LU, LF.LUIdx);
2954 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
2961 /// CollectSubexprs - Split S into subexpressions which can be pulled out into
2962 /// separate registers. If C is non-null, multiply each subexpression by C.
2963 static void CollectSubexprs(const SCEV *S, const SCEVConstant *C,
2964 SmallVectorImpl<const SCEV *> &Ops,
2966 ScalarEvolution &SE) {
2967 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2968 // Break out add operands.
2969 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
2971 CollectSubexprs(*I, C, Ops, L, SE);
2973 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2974 // Split a non-zero base out of an addrec.
2975 if (!AR->getStart()->isZero()) {
2976 CollectSubexprs(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
2977 AR->getStepRecurrence(SE),
2979 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
2982 CollectSubexprs(AR->getStart(), C, Ops, L, SE);
2985 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
2986 // Break (C * (a + b + c)) into C*a + C*b + C*c.
2987 if (Mul->getNumOperands() == 2)
2988 if (const SCEVConstant *Op0 =
2989 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
2990 CollectSubexprs(Mul->getOperand(1),
2991 C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0,
2997 // Otherwise use the value itself, optionally with a scale applied.
2998 Ops.push_back(C ? SE.getMulExpr(C, S) : S);
3001 /// GenerateReassociations - Split out subexpressions from adds and the bases of
3003 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
3006 // Arbitrarily cap recursion to protect compile time.
3007 if (Depth >= 3) return;
3009 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3010 const SCEV *BaseReg = Base.BaseRegs[i];
3012 SmallVector<const SCEV *, 8> AddOps;
3013 CollectSubexprs(BaseReg, 0, AddOps, L, SE);
3015 if (AddOps.size() == 1) continue;
3017 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
3018 JE = AddOps.end(); J != JE; ++J) {
3020 // Loop-variant "unknown" values are uninteresting; we won't be able to
3021 // do anything meaningful with them.
3022 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
3025 // Don't pull a constant into a register if the constant could be folded
3026 // into an immediate field.
3027 if (isAlwaysFoldable(*J, LU.MinOffset, LU.MaxOffset,
3028 Base.getNumRegs() > 1,
3029 LU.Kind, LU.AccessTy, TLI, SE))
3032 // Collect all operands except *J.
3033 SmallVector<const SCEV *, 8> InnerAddOps
3034 (((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
3036 (llvm::next(J), ((const SmallVector<const SCEV *, 8> &)AddOps).end());
3038 // Don't leave just a constant behind in a register if the constant could
3039 // be folded into an immediate field.
3040 if (InnerAddOps.size() == 1 &&
3041 isAlwaysFoldable(InnerAddOps[0], LU.MinOffset, LU.MaxOffset,
3042 Base.getNumRegs() > 1,
3043 LU.Kind, LU.AccessTy, TLI, SE))
3046 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
3047 if (InnerSum->isZero())
3051 // Add the remaining pieces of the add back into the new formula.
3052 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
3053 if (TLI && InnerSumSC &&
3054 SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
3055 TLI->isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3056 InnerSumSC->getValue()->getZExtValue())) {
3057 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
3058 InnerSumSC->getValue()->getZExtValue();
3059 F.BaseRegs.erase(F.BaseRegs.begin() + i);
3061 F.BaseRegs[i] = InnerSum;
3063 // Add J as its own register, or an unfolded immediate.
3064 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
3065 if (TLI && SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
3066 TLI->isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3067 SC->getValue()->getZExtValue()))
3068 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
3069 SC->getValue()->getZExtValue();
3071 F.BaseRegs.push_back(*J);
3073 if (InsertFormula(LU, LUIdx, F))
3074 // If that formula hadn't been seen before, recurse to find more like
3076 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1);
3081 /// GenerateCombinations - Generate a formula consisting of all of the
3082 /// loop-dominating registers added into a single register.
3083 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
3085 // This method is only interesting on a plurality of registers.
3086 if (Base.BaseRegs.size() <= 1) return;
3090 SmallVector<const SCEV *, 4> Ops;
3091 for (SmallVectorImpl<const SCEV *>::const_iterator
3092 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
3093 const SCEV *BaseReg = *I;
3094 if (SE.properlyDominates(BaseReg, L->getHeader()) &&
3095 !SE.hasComputableLoopEvolution(BaseReg, L))
3096 Ops.push_back(BaseReg);
3098 F.BaseRegs.push_back(BaseReg);
3100 if (Ops.size() > 1) {
3101 const SCEV *Sum = SE.getAddExpr(Ops);
3102 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
3103 // opportunity to fold something. For now, just ignore such cases
3104 // rather than proceed with zero in a register.
3105 if (!Sum->isZero()) {
3106 F.BaseRegs.push_back(Sum);
3107 (void)InsertFormula(LU, LUIdx, F);
3112 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
3113 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
3115 // We can't add a symbolic offset if the address already contains one.
3116 if (Base.AM.BaseGV) return;
3118 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3119 const SCEV *G = Base.BaseRegs[i];
3120 GlobalValue *GV = ExtractSymbol(G, SE);
3121 if (G->isZero() || !GV)
3125 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
3126 LU.Kind, LU.AccessTy, TLI))
3129 (void)InsertFormula(LU, LUIdx, F);
3133 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
3134 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
3136 // TODO: For now, just add the min and max offset, because it usually isn't
3137 // worthwhile looking at everything inbetween.
3138 SmallVector<int64_t, 2> Worklist;
3139 Worklist.push_back(LU.MinOffset);
3140 if (LU.MaxOffset != LU.MinOffset)
3141 Worklist.push_back(LU.MaxOffset);
3143 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3144 const SCEV *G = Base.BaseRegs[i];
3146 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
3147 E = Worklist.end(); I != E; ++I) {
3149 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs - *I;
3150 if (isLegalUse(F.AM, LU.MinOffset - *I, LU.MaxOffset - *I,
3151 LU.Kind, LU.AccessTy, TLI)) {
3152 // Add the offset to the base register.
3153 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), *I), G);
3154 // If it cancelled out, drop the base register, otherwise update it.
3155 if (NewG->isZero()) {
3156 std::swap(F.BaseRegs[i], F.BaseRegs.back());
3157 F.BaseRegs.pop_back();
3159 F.BaseRegs[i] = NewG;
3161 (void)InsertFormula(LU, LUIdx, F);
3165 int64_t Imm = ExtractImmediate(G, SE);
3166 if (G->isZero() || Imm == 0)
3169 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Imm;
3170 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
3171 LU.Kind, LU.AccessTy, TLI))
3174 (void)InsertFormula(LU, LUIdx, F);
3178 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
3179 /// the comparison. For example, x == y -> x*c == y*c.
3180 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
3182 if (LU.Kind != LSRUse::ICmpZero) return;
3184 // Determine the integer type for the base formula.
3185 Type *IntTy = Base.getType();
3187 if (SE.getTypeSizeInBits(IntTy) > 64) return;
3189 // Don't do this if there is more than one offset.
3190 if (LU.MinOffset != LU.MaxOffset) return;
3192 assert(!Base.AM.BaseGV && "ICmpZero use is not legal!");
3194 // Check each interesting stride.
3195 for (SmallSetVector<int64_t, 8>::const_iterator
3196 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3197 int64_t Factor = *I;
3199 // Check that the multiplication doesn't overflow.
3200 if (Base.AM.BaseOffs == INT64_MIN && Factor == -1)
3202 int64_t NewBaseOffs = (uint64_t)Base.AM.BaseOffs * Factor;
3203 if (NewBaseOffs / Factor != Base.AM.BaseOffs)
3206 // Check that multiplying with the use offset doesn't overflow.
3207 int64_t Offset = LU.MinOffset;
3208 if (Offset == INT64_MIN && Factor == -1)
3210 Offset = (uint64_t)Offset * Factor;
3211 if (Offset / Factor != LU.MinOffset)
3215 F.AM.BaseOffs = NewBaseOffs;
3217 // Check that this scale is legal.
3218 if (!isLegalUse(F.AM, Offset, Offset, LU.Kind, LU.AccessTy, TLI))
3221 // Compensate for the use having MinOffset built into it.
3222 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Offset - LU.MinOffset;
3224 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3226 // Check that multiplying with each base register doesn't overflow.
3227 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
3228 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
3229 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
3233 // Check that multiplying with the scaled register doesn't overflow.
3235 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
3236 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
3240 // Check that multiplying with the unfolded offset doesn't overflow.
3241 if (F.UnfoldedOffset != 0) {
3242 if (F.UnfoldedOffset == INT64_MIN && Factor == -1)
3244 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
3245 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
3249 // If we make it here and it's legal, add it.
3250 (void)InsertFormula(LU, LUIdx, F);
3255 /// GenerateScales - Generate stride factor reuse formulae by making use of
3256 /// scaled-offset address modes, for example.
3257 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
3258 // Determine the integer type for the base formula.
3259 Type *IntTy = Base.getType();
3262 // If this Formula already has a scaled register, we can't add another one.
3263 if (Base.AM.Scale != 0) return;
3265 // Check each interesting stride.
3266 for (SmallSetVector<int64_t, 8>::const_iterator
3267 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3268 int64_t Factor = *I;
3270 Base.AM.Scale = Factor;
3271 Base.AM.HasBaseReg = Base.BaseRegs.size() > 1;
3272 // Check whether this scale is going to be legal.
3273 if (!isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
3274 LU.Kind, LU.AccessTy, TLI)) {
3275 // As a special-case, handle special out-of-loop Basic users specially.
3276 // TODO: Reconsider this special case.
3277 if (LU.Kind == LSRUse::Basic &&
3278 isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
3279 LSRUse::Special, LU.AccessTy, TLI) &&
3280 LU.AllFixupsOutsideLoop)
3281 LU.Kind = LSRUse::Special;
3285 // For an ICmpZero, negating a solitary base register won't lead to
3287 if (LU.Kind == LSRUse::ICmpZero &&
3288 !Base.AM.HasBaseReg && Base.AM.BaseOffs == 0 && !Base.AM.BaseGV)
3290 // For each addrec base reg, apply the scale, if possible.
3291 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3292 if (const SCEVAddRecExpr *AR =
3293 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
3294 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3295 if (FactorS->isZero())
3297 // Divide out the factor, ignoring high bits, since we'll be
3298 // scaling the value back up in the end.
3299 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
3300 // TODO: This could be optimized to avoid all the copying.
3302 F.ScaledReg = Quotient;
3303 F.DeleteBaseReg(F.BaseRegs[i]);
3304 (void)InsertFormula(LU, LUIdx, F);
3310 /// GenerateTruncates - Generate reuse formulae from different IV types.
3311 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
3312 // This requires TargetLowering to tell us which truncates are free.
3315 // Don't bother truncating symbolic values.
3316 if (Base.AM.BaseGV) return;
3318 // Determine the integer type for the base formula.
3319 Type *DstTy = Base.getType();
3321 DstTy = SE.getEffectiveSCEVType(DstTy);
3323 for (SmallSetVector<Type *, 4>::const_iterator
3324 I = Types.begin(), E = Types.end(); I != E; ++I) {
3326 if (SrcTy != DstTy && TLI->isTruncateFree(SrcTy, DstTy)) {
3329 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
3330 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
3331 JE = F.BaseRegs.end(); J != JE; ++J)
3332 *J = SE.getAnyExtendExpr(*J, SrcTy);
3334 // TODO: This assumes we've done basic processing on all uses and
3335 // have an idea what the register usage is.
3336 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
3339 (void)InsertFormula(LU, LUIdx, F);
3346 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
3347 /// defer modifications so that the search phase doesn't have to worry about
3348 /// the data structures moving underneath it.
3352 const SCEV *OrigReg;
3354 WorkItem(size_t LI, int64_t I, const SCEV *R)
3355 : LUIdx(LI), Imm(I), OrigReg(R) {}
3357 void print(raw_ostream &OS) const;
3363 void WorkItem::print(raw_ostream &OS) const {
3364 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
3365 << " , add offset " << Imm;
3368 void WorkItem::dump() const {
3369 print(errs()); errs() << '\n';
3372 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
3373 /// distance apart and try to form reuse opportunities between them.
3374 void LSRInstance::GenerateCrossUseConstantOffsets() {
3375 // Group the registers by their value without any added constant offset.
3376 typedef std::map<int64_t, const SCEV *> ImmMapTy;
3377 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
3379 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
3380 SmallVector<const SCEV *, 8> Sequence;
3381 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
3383 const SCEV *Reg = *I;
3384 int64_t Imm = ExtractImmediate(Reg, SE);
3385 std::pair<RegMapTy::iterator, bool> Pair =
3386 Map.insert(std::make_pair(Reg, ImmMapTy()));
3388 Sequence.push_back(Reg);
3389 Pair.first->second.insert(std::make_pair(Imm, *I));
3390 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
3393 // Now examine each set of registers with the same base value. Build up
3394 // a list of work to do and do the work in a separate step so that we're
3395 // not adding formulae and register counts while we're searching.
3396 SmallVector<WorkItem, 32> WorkItems;
3397 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
3398 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
3399 E = Sequence.end(); I != E; ++I) {
3400 const SCEV *Reg = *I;
3401 const ImmMapTy &Imms = Map.find(Reg)->second;
3403 // It's not worthwhile looking for reuse if there's only one offset.
3404 if (Imms.size() == 1)
3407 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
3408 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3410 dbgs() << ' ' << J->first;
3413 // Examine each offset.
3414 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3416 const SCEV *OrigReg = J->second;
3418 int64_t JImm = J->first;
3419 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
3421 if (!isa<SCEVConstant>(OrigReg) &&
3422 UsedByIndicesMap[Reg].count() == 1) {
3423 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
3427 // Conservatively examine offsets between this orig reg a few selected
3429 ImmMapTy::const_iterator OtherImms[] = {
3430 Imms.begin(), prior(Imms.end()),
3431 Imms.lower_bound((Imms.begin()->first + prior(Imms.end())->first) / 2)
3433 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
3434 ImmMapTy::const_iterator M = OtherImms[i];
3435 if (M == J || M == JE) continue;
3437 // Compute the difference between the two.
3438 int64_t Imm = (uint64_t)JImm - M->first;
3439 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
3440 LUIdx = UsedByIndices.find_next(LUIdx))
3441 // Make a memo of this use, offset, and register tuple.
3442 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
3443 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
3450 UsedByIndicesMap.clear();
3451 UniqueItems.clear();
3453 // Now iterate through the worklist and add new formulae.
3454 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
3455 E = WorkItems.end(); I != E; ++I) {
3456 const WorkItem &WI = *I;
3457 size_t LUIdx = WI.LUIdx;
3458 LSRUse &LU = Uses[LUIdx];
3459 int64_t Imm = WI.Imm;
3460 const SCEV *OrigReg = WI.OrigReg;
3462 Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
3463 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
3464 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
3466 // TODO: Use a more targeted data structure.
3467 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
3468 const Formula &F = LU.Formulae[L];
3469 // Use the immediate in the scaled register.
3470 if (F.ScaledReg == OrigReg) {
3471 int64_t Offs = (uint64_t)F.AM.BaseOffs +
3472 Imm * (uint64_t)F.AM.Scale;
3473 // Don't create 50 + reg(-50).
3474 if (F.referencesReg(SE.getSCEV(
3475 ConstantInt::get(IntTy, -(uint64_t)Offs))))
3478 NewF.AM.BaseOffs = Offs;
3479 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
3480 LU.Kind, LU.AccessTy, TLI))
3482 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
3484 // If the new scale is a constant in a register, and adding the constant
3485 // value to the immediate would produce a value closer to zero than the
3486 // immediate itself, then the formula isn't worthwhile.
3487 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
3488 if (C->getValue()->isNegative() !=
3489 (NewF.AM.BaseOffs < 0) &&
3490 (C->getValue()->getValue().abs() * APInt(BitWidth, F.AM.Scale))
3491 .ule(abs64(NewF.AM.BaseOffs)))
3495 (void)InsertFormula(LU, LUIdx, NewF);
3497 // Use the immediate in a base register.
3498 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
3499 const SCEV *BaseReg = F.BaseRegs[N];
3500 if (BaseReg != OrigReg)
3503 NewF.AM.BaseOffs = (uint64_t)NewF.AM.BaseOffs + Imm;
3504 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
3505 LU.Kind, LU.AccessTy, TLI)) {
3507 !TLI->isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
3510 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
3512 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
3514 // If the new formula has a constant in a register, and adding the
3515 // constant value to the immediate would produce a value closer to
3516 // zero than the immediate itself, then the formula isn't worthwhile.
3517 for (SmallVectorImpl<const SCEV *>::const_iterator
3518 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
3520 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
3521 if ((C->getValue()->getValue() + NewF.AM.BaseOffs).abs().slt(
3522 abs64(NewF.AM.BaseOffs)) &&
3523 (C->getValue()->getValue() +
3524 NewF.AM.BaseOffs).countTrailingZeros() >=
3525 CountTrailingZeros_64(NewF.AM.BaseOffs))
3529 (void)InsertFormula(LU, LUIdx, NewF);
3538 /// GenerateAllReuseFormulae - Generate formulae for each use.
3540 LSRInstance::GenerateAllReuseFormulae() {
3541 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
3542 // queries are more precise.
3543 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3544 LSRUse &LU = Uses[LUIdx];
3545 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3546 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
3547 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3548 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
3550 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3551 LSRUse &LU = Uses[LUIdx];
3552 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3553 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
3554 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3555 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
3556 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3557 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
3558 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3559 GenerateScales(LU, LUIdx, LU.Formulae[i]);
3561 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3562 LSRUse &LU = Uses[LUIdx];
3563 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3564 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
3567 GenerateCrossUseConstantOffsets();
3569 DEBUG(dbgs() << "\n"
3570 "After generating reuse formulae:\n";
3571 print_uses(dbgs()));
3574 /// If there are multiple formulae with the same set of registers used
3575 /// by other uses, pick the best one and delete the others.
3576 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
3577 DenseSet<const SCEV *> VisitedRegs;
3578 SmallPtrSet<const SCEV *, 16> Regs;
3579 SmallPtrSet<const SCEV *, 16> LoserRegs;
3581 bool ChangedFormulae = false;
3584 // Collect the best formula for each unique set of shared registers. This
3585 // is reset for each use.
3586 typedef DenseMap<SmallVector<const SCEV *, 2>, size_t, UniquifierDenseMapInfo>
3588 BestFormulaeTy BestFormulae;
3590 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3591 LSRUse &LU = Uses[LUIdx];
3592 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
3595 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
3596 FIdx != NumForms; ++FIdx) {
3597 Formula &F = LU.Formulae[FIdx];
3599 // Some formulas are instant losers. For example, they may depend on
3600 // nonexistent AddRecs from other loops. These need to be filtered
3601 // immediately, otherwise heuristics could choose them over others leading
3602 // to an unsatisfactory solution. Passing LoserRegs into RateFormula here
3603 // avoids the need to recompute this information across formulae using the
3604 // same bad AddRec. Passing LoserRegs is also essential unless we remove
3605 // the corresponding bad register from the Regs set.
3608 CostF.RateFormula(F, Regs, VisitedRegs, L, LU.Offsets, SE, DT,
3610 if (CostF.isLoser()) {
3611 // During initial formula generation, undesirable formulae are generated
3612 // by uses within other loops that have some non-trivial address mode or
3613 // use the postinc form of the IV. LSR needs to provide these formulae
3614 // as the basis of rediscovering the desired formula that uses an AddRec
3615 // corresponding to the existing phi. Once all formulae have been
3616 // generated, these initial losers may be pruned.
3617 DEBUG(dbgs() << " Filtering loser "; F.print(dbgs());
3621 SmallVector<const SCEV *, 2> Key;
3622 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
3623 JE = F.BaseRegs.end(); J != JE; ++J) {
3624 const SCEV *Reg = *J;
3625 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
3629 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
3630 Key.push_back(F.ScaledReg);
3631 // Unstable sort by host order ok, because this is only used for
3633 std::sort(Key.begin(), Key.end());
3635 std::pair<BestFormulaeTy::const_iterator, bool> P =
3636 BestFormulae.insert(std::make_pair(Key, FIdx));
3640 Formula &Best = LU.Formulae[P.first->second];
3644 CostBest.RateFormula(Best, Regs, VisitedRegs, L, LU.Offsets, SE, DT);
3645 if (CostF < CostBest)
3647 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
3649 " in favor of formula "; Best.print(dbgs());
3653 ChangedFormulae = true;
3655 LU.DeleteFormula(F);
3661 // Now that we've filtered out some formulae, recompute the Regs set.
3663 LU.RecomputeRegs(LUIdx, RegUses);
3665 // Reset this to prepare for the next use.
3666 BestFormulae.clear();
3669 DEBUG(if (ChangedFormulae) {
3671 "After filtering out undesirable candidates:\n";
3676 // This is a rough guess that seems to work fairly well.
3677 static const size_t ComplexityLimit = UINT16_MAX;
3679 /// EstimateSearchSpaceComplexity - Estimate the worst-case number of
3680 /// solutions the solver might have to consider. It almost never considers
3681 /// this many solutions because it prune the search space, but the pruning
3682 /// isn't always sufficient.
3683 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
3685 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3686 E = Uses.end(); I != E; ++I) {
3687 size_t FSize = I->Formulae.size();
3688 if (FSize >= ComplexityLimit) {
3689 Power = ComplexityLimit;
3693 if (Power >= ComplexityLimit)
3699 /// NarrowSearchSpaceByDetectingSupersets - When one formula uses a superset
3700 /// of the registers of another formula, it won't help reduce register
3701 /// pressure (though it may not necessarily hurt register pressure); remove
3702 /// it to simplify the system.
3703 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
3704 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3705 DEBUG(dbgs() << "The search space is too complex.\n");
3707 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
3708 "which use a superset of registers used by other "
3711 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3712 LSRUse &LU = Uses[LUIdx];
3714 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3715 Formula &F = LU.Formulae[i];
3716 // Look for a formula with a constant or GV in a register. If the use
3717 // also has a formula with that same value in an immediate field,
3718 // delete the one that uses a register.
3719 for (SmallVectorImpl<const SCEV *>::const_iterator
3720 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
3721 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
3723 NewF.AM.BaseOffs += C->getValue()->getSExtValue();
3724 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
3725 (I - F.BaseRegs.begin()));
3726 if (LU.HasFormulaWithSameRegs(NewF)) {
3727 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3728 LU.DeleteFormula(F);
3734 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
3735 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
3738 NewF.AM.BaseGV = GV;
3739 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
3740 (I - F.BaseRegs.begin()));
3741 if (LU.HasFormulaWithSameRegs(NewF)) {
3742 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3744 LU.DeleteFormula(F);
3755 LU.RecomputeRegs(LUIdx, RegUses);
3758 DEBUG(dbgs() << "After pre-selection:\n";
3759 print_uses(dbgs()));
3763 /// NarrowSearchSpaceByCollapsingUnrolledCode - When there are many registers
3764 /// for expressions like A, A+1, A+2, etc., allocate a single register for
3766 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
3767 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3768 DEBUG(dbgs() << "The search space is too complex.\n");
3770 DEBUG(dbgs() << "Narrowing the search space by assuming that uses "
3771 "separated by a constant offset will use the same "
3774 // This is especially useful for unrolled loops.
3776 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3777 LSRUse &LU = Uses[LUIdx];
3778 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3779 E = LU.Formulae.end(); I != E; ++I) {
3780 const Formula &F = *I;
3781 if (F.AM.BaseOffs != 0 && F.AM.Scale == 0) {
3782 if (LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU)) {
3783 if (reconcileNewOffset(*LUThatHas, F.AM.BaseOffs,
3784 /*HasBaseReg=*/false,
3785 LU.Kind, LU.AccessTy)) {
3786 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs());
3789 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
3791 // Update the relocs to reference the new use.
3792 for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(),
3793 E = Fixups.end(); I != E; ++I) {
3794 LSRFixup &Fixup = *I;
3795 if (Fixup.LUIdx == LUIdx) {
3796 Fixup.LUIdx = LUThatHas - &Uses.front();
3797 Fixup.Offset += F.AM.BaseOffs;
3798 // Add the new offset to LUThatHas' offset list.
3799 if (LUThatHas->Offsets.back() != Fixup.Offset) {
3800 LUThatHas->Offsets.push_back(Fixup.Offset);
3801 if (Fixup.Offset > LUThatHas->MaxOffset)
3802 LUThatHas->MaxOffset = Fixup.Offset;
3803 if (Fixup.Offset < LUThatHas->MinOffset)
3804 LUThatHas->MinOffset = Fixup.Offset;
3806 DEBUG(dbgs() << "New fixup has offset "
3807 << Fixup.Offset << '\n');
3809 if (Fixup.LUIdx == NumUses-1)
3810 Fixup.LUIdx = LUIdx;
3813 // Delete formulae from the new use which are no longer legal.
3815 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
3816 Formula &F = LUThatHas->Formulae[i];
3817 if (!isLegalUse(F.AM,
3818 LUThatHas->MinOffset, LUThatHas->MaxOffset,
3819 LUThatHas->Kind, LUThatHas->AccessTy, TLI)) {
3820 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3822 LUThatHas->DeleteFormula(F);
3829 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
3831 // Delete the old use.
3832 DeleteUse(LU, LUIdx);
3842 DEBUG(dbgs() << "After pre-selection:\n";
3843 print_uses(dbgs()));
3847 /// NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters - Call
3848 /// FilterOutUndesirableDedicatedRegisters again, if necessary, now that
3849 /// we've done more filtering, as it may be able to find more formulae to
3851 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
3852 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3853 DEBUG(dbgs() << "The search space is too complex.\n");
3855 DEBUG(dbgs() << "Narrowing the search space by re-filtering out "
3856 "undesirable dedicated registers.\n");
3858 FilterOutUndesirableDedicatedRegisters();
3860 DEBUG(dbgs() << "After pre-selection:\n";
3861 print_uses(dbgs()));
3865 /// NarrowSearchSpaceByPickingWinnerRegs - Pick a register which seems likely
3866 /// to be profitable, and then in any use which has any reference to that
3867 /// register, delete all formulae which do not reference that register.
3868 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
3869 // With all other options exhausted, loop until the system is simple
3870 // enough to handle.
3871 SmallPtrSet<const SCEV *, 4> Taken;
3872 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3873 // Ok, we have too many of formulae on our hands to conveniently handle.
3874 // Use a rough heuristic to thin out the list.
3875 DEBUG(dbgs() << "The search space is too complex.\n");
3877 // Pick the register which is used by the most LSRUses, which is likely
3878 // to be a good reuse register candidate.
3879 const SCEV *Best = 0;
3880 unsigned BestNum = 0;
3881 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
3883 const SCEV *Reg = *I;
3884 if (Taken.count(Reg))
3889 unsigned Count = RegUses.getUsedByIndices(Reg).count();
3890 if (Count > BestNum) {
3897 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
3898 << " will yield profitable reuse.\n");
3901 // In any use with formulae which references this register, delete formulae
3902 // which don't reference it.
3903 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3904 LSRUse &LU = Uses[LUIdx];
3905 if (!LU.Regs.count(Best)) continue;
3908 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3909 Formula &F = LU.Formulae[i];
3910 if (!F.referencesReg(Best)) {
3911 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3912 LU.DeleteFormula(F);
3916 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
3922 LU.RecomputeRegs(LUIdx, RegUses);
3925 DEBUG(dbgs() << "After pre-selection:\n";
3926 print_uses(dbgs()));
3930 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
3931 /// formulae to choose from, use some rough heuristics to prune down the number
3932 /// of formulae. This keeps the main solver from taking an extraordinary amount
3933 /// of time in some worst-case scenarios.
3934 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
3935 NarrowSearchSpaceByDetectingSupersets();
3936 NarrowSearchSpaceByCollapsingUnrolledCode();
3937 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
3938 NarrowSearchSpaceByPickingWinnerRegs();
3941 /// SolveRecurse - This is the recursive solver.
3942 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
3944 SmallVectorImpl<const Formula *> &Workspace,
3945 const Cost &CurCost,
3946 const SmallPtrSet<const SCEV *, 16> &CurRegs,
3947 DenseSet<const SCEV *> &VisitedRegs) const {
3950 // - use more aggressive filtering
3951 // - sort the formula so that the most profitable solutions are found first
3952 // - sort the uses too
3954 // - don't compute a cost, and then compare. compare while computing a cost
3956 // - track register sets with SmallBitVector
3958 const LSRUse &LU = Uses[Workspace.size()];
3960 // If this use references any register that's already a part of the
3961 // in-progress solution, consider it a requirement that a formula must
3962 // reference that register in order to be considered. This prunes out
3963 // unprofitable searching.
3964 SmallSetVector<const SCEV *, 4> ReqRegs;
3965 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(),
3966 E = CurRegs.end(); I != E; ++I)
3967 if (LU.Regs.count(*I))
3970 bool AnySatisfiedReqRegs = false;
3971 SmallPtrSet<const SCEV *, 16> NewRegs;
3974 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3975 E = LU.Formulae.end(); I != E; ++I) {
3976 const Formula &F = *I;
3978 // Ignore formulae which do not use any of the required registers.
3979 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
3980 JE = ReqRegs.end(); J != JE; ++J) {
3981 const SCEV *Reg = *J;
3982 if ((!F.ScaledReg || F.ScaledReg != Reg) &&
3983 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) ==
3987 AnySatisfiedReqRegs = true;
3989 // Evaluate the cost of the current formula. If it's already worse than
3990 // the current best, prune the search at that point.
3993 NewCost.RateFormula(F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT);
3994 if (NewCost < SolutionCost) {
3995 Workspace.push_back(&F);
3996 if (Workspace.size() != Uses.size()) {
3997 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
3998 NewRegs, VisitedRegs);
3999 if (F.getNumRegs() == 1 && Workspace.size() == 1)
4000 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
4002 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
4003 dbgs() << ".\n Regs:";
4004 for (SmallPtrSet<const SCEV *, 16>::const_iterator
4005 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I)
4006 dbgs() << ' ' << **I;
4009 SolutionCost = NewCost;
4010 Solution = Workspace;
4012 Workspace.pop_back();
4017 if (!EnableRetry && !AnySatisfiedReqRegs)
4020 // If none of the formulae had all of the required registers, relax the
4021 // constraint so that we don't exclude all formulae.
4022 if (!AnySatisfiedReqRegs) {
4023 assert(!ReqRegs.empty() && "Solver failed even without required registers");
4029 /// Solve - Choose one formula from each use. Return the results in the given
4030 /// Solution vector.
4031 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
4032 SmallVector<const Formula *, 8> Workspace;
4034 SolutionCost.Loose();
4036 SmallPtrSet<const SCEV *, 16> CurRegs;
4037 DenseSet<const SCEV *> VisitedRegs;
4038 Workspace.reserve(Uses.size());
4040 // SolveRecurse does all the work.
4041 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
4042 CurRegs, VisitedRegs);
4043 if (Solution.empty()) {
4044 DEBUG(dbgs() << "\nNo Satisfactory Solution\n");
4048 // Ok, we've now made all our decisions.
4049 DEBUG(dbgs() << "\n"
4050 "The chosen solution requires "; SolutionCost.print(dbgs());
4052 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
4054 Uses[i].print(dbgs());
4057 Solution[i]->print(dbgs());
4061 assert(Solution.size() == Uses.size() && "Malformed solution!");
4064 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
4065 /// the dominator tree far as we can go while still being dominated by the
4066 /// input positions. This helps canonicalize the insert position, which
4067 /// encourages sharing.
4068 BasicBlock::iterator
4069 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
4070 const SmallVectorImpl<Instruction *> &Inputs)
4073 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
4074 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
4077 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
4078 if (!Rung) return IP;
4079 Rung = Rung->getIDom();
4080 if (!Rung) return IP;
4081 IDom = Rung->getBlock();
4083 // Don't climb into a loop though.
4084 const Loop *IDomLoop = LI.getLoopFor(IDom);
4085 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
4086 if (IDomDepth <= IPLoopDepth &&
4087 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
4091 bool AllDominate = true;
4092 Instruction *BetterPos = 0;
4093 Instruction *Tentative = IDom->getTerminator();
4094 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
4095 E = Inputs.end(); I != E; ++I) {
4096 Instruction *Inst = *I;
4097 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
4098 AllDominate = false;
4101 // Attempt to find an insert position in the middle of the block,
4102 // instead of at the end, so that it can be used for other expansions.
4103 if (IDom == Inst->getParent() &&
4104 (!BetterPos || DT.dominates(BetterPos, Inst)))
4105 BetterPos = llvm::next(BasicBlock::iterator(Inst));
4118 /// AdjustInsertPositionForExpand - Determine an input position which will be
4119 /// dominated by the operands and which will dominate the result.
4120 BasicBlock::iterator
4121 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator LowestIP,
4124 SCEVExpander &Rewriter) const {
4125 // Collect some instructions which must be dominated by the
4126 // expanding replacement. These must be dominated by any operands that
4127 // will be required in the expansion.
4128 SmallVector<Instruction *, 4> Inputs;
4129 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
4130 Inputs.push_back(I);
4131 if (LU.Kind == LSRUse::ICmpZero)
4132 if (Instruction *I =
4133 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
4134 Inputs.push_back(I);
4135 if (LF.PostIncLoops.count(L)) {
4136 if (LF.isUseFullyOutsideLoop(L))
4137 Inputs.push_back(L->getLoopLatch()->getTerminator());
4139 Inputs.push_back(IVIncInsertPos);
4141 // The expansion must also be dominated by the increment positions of any
4142 // loops it for which it is using post-inc mode.
4143 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
4144 E = LF.PostIncLoops.end(); I != E; ++I) {
4145 const Loop *PIL = *I;
4146 if (PIL == L) continue;
4148 // Be dominated by the loop exit.
4149 SmallVector<BasicBlock *, 4> ExitingBlocks;
4150 PIL->getExitingBlocks(ExitingBlocks);
4151 if (!ExitingBlocks.empty()) {
4152 BasicBlock *BB = ExitingBlocks[0];
4153 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
4154 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
4155 Inputs.push_back(BB->getTerminator());
4159 assert(!isa<PHINode>(LowestIP) && !isa<LandingPadInst>(LowestIP)
4160 && !isa<DbgInfoIntrinsic>(LowestIP) &&
4161 "Insertion point must be a normal instruction");
4163 // Then, climb up the immediate dominator tree as far as we can go while
4164 // still being dominated by the input positions.
4165 BasicBlock::iterator IP = HoistInsertPosition(LowestIP, Inputs);
4167 // Don't insert instructions before PHI nodes.
4168 while (isa<PHINode>(IP)) ++IP;
4170 // Ignore landingpad instructions.
4171 while (isa<LandingPadInst>(IP)) ++IP;
4173 // Ignore debug intrinsics.
4174 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
4176 // Set IP below instructions recently inserted by SCEVExpander. This keeps the
4177 // IP consistent across expansions and allows the previously inserted
4178 // instructions to be reused by subsequent expansion.
4179 while (Rewriter.isInsertedInstruction(IP) && IP != LowestIP) ++IP;
4184 /// Expand - Emit instructions for the leading candidate expression for this
4185 /// LSRUse (this is called "expanding").
4186 Value *LSRInstance::Expand(const LSRFixup &LF,
4188 BasicBlock::iterator IP,
4189 SCEVExpander &Rewriter,
4190 SmallVectorImpl<WeakVH> &DeadInsts) const {
4191 const LSRUse &LU = Uses[LF.LUIdx];
4193 // Determine an input position which will be dominated by the operands and
4194 // which will dominate the result.
4195 IP = AdjustInsertPositionForExpand(IP, LF, LU, Rewriter);
4197 // Inform the Rewriter if we have a post-increment use, so that it can
4198 // perform an advantageous expansion.
4199 Rewriter.setPostInc(LF.PostIncLoops);
4201 // This is the type that the user actually needs.
4202 Type *OpTy = LF.OperandValToReplace->getType();
4203 // This will be the type that we'll initially expand to.
4204 Type *Ty = F.getType();
4206 // No type known; just expand directly to the ultimate type.
4208 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
4209 // Expand directly to the ultimate type if it's the right size.
4211 // This is the type to do integer arithmetic in.
4212 Type *IntTy = SE.getEffectiveSCEVType(Ty);
4214 // Build up a list of operands to add together to form the full base.
4215 SmallVector<const SCEV *, 8> Ops;
4217 // Expand the BaseRegs portion.
4218 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
4219 E = F.BaseRegs.end(); I != E; ++I) {
4220 const SCEV *Reg = *I;
4221 assert(!Reg->isZero() && "Zero allocated in a base register!");
4223 // If we're expanding for a post-inc user, make the post-inc adjustment.
4224 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4225 Reg = TransformForPostIncUse(Denormalize, Reg,
4226 LF.UserInst, LF.OperandValToReplace,
4229 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP)));
4232 // Flush the operand list to suppress SCEVExpander hoisting.
4234 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4236 Ops.push_back(SE.getUnknown(FullV));
4239 // Expand the ScaledReg portion.
4240 Value *ICmpScaledV = 0;
4241 if (F.AM.Scale != 0) {
4242 const SCEV *ScaledS = F.ScaledReg;
4244 // If we're expanding for a post-inc user, make the post-inc adjustment.
4245 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4246 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
4247 LF.UserInst, LF.OperandValToReplace,
4250 if (LU.Kind == LSRUse::ICmpZero) {
4251 // An interesting way of "folding" with an icmp is to use a negated
4252 // scale, which we'll implement by inserting it into the other operand
4254 assert(F.AM.Scale == -1 &&
4255 "The only scale supported by ICmpZero uses is -1!");
4256 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP);
4258 // Otherwise just expand the scaled register and an explicit scale,
4259 // which is expected to be matched as part of the address.
4260 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP));
4261 ScaledS = SE.getMulExpr(ScaledS,
4262 SE.getConstant(ScaledS->getType(), F.AM.Scale));
4263 Ops.push_back(ScaledS);
4265 // Flush the operand list to suppress SCEVExpander hoisting.
4266 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4268 Ops.push_back(SE.getUnknown(FullV));
4272 // Expand the GV portion.
4274 Ops.push_back(SE.getUnknown(F.AM.BaseGV));
4276 // Flush the operand list to suppress SCEVExpander hoisting.
4277 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4279 Ops.push_back(SE.getUnknown(FullV));
4282 // Expand the immediate portion.
4283 int64_t Offset = (uint64_t)F.AM.BaseOffs + LF.Offset;
4285 if (LU.Kind == LSRUse::ICmpZero) {
4286 // The other interesting way of "folding" with an ICmpZero is to use a
4287 // negated immediate.
4289 ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset);
4291 Ops.push_back(SE.getUnknown(ICmpScaledV));
4292 ICmpScaledV = ConstantInt::get(IntTy, Offset);
4295 // Just add the immediate values. These again are expected to be matched
4296 // as part of the address.
4297 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
4301 // Expand the unfolded offset portion.
4302 int64_t UnfoldedOffset = F.UnfoldedOffset;
4303 if (UnfoldedOffset != 0) {
4304 // Just add the immediate values.
4305 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy,
4309 // Emit instructions summing all the operands.
4310 const SCEV *FullS = Ops.empty() ?
4311 SE.getConstant(IntTy, 0) :
4313 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
4315 // We're done expanding now, so reset the rewriter.
4316 Rewriter.clearPostInc();
4318 // An ICmpZero Formula represents an ICmp which we're handling as a
4319 // comparison against zero. Now that we've expanded an expression for that
4320 // form, update the ICmp's other operand.
4321 if (LU.Kind == LSRUse::ICmpZero) {
4322 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
4323 DeadInsts.push_back(CI->getOperand(1));
4324 assert(!F.AM.BaseGV && "ICmp does not support folding a global value and "
4325 "a scale at the same time!");
4326 if (F.AM.Scale == -1) {
4327 if (ICmpScaledV->getType() != OpTy) {
4329 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
4331 ICmpScaledV, OpTy, "tmp", CI);
4334 CI->setOperand(1, ICmpScaledV);
4336 assert(F.AM.Scale == 0 &&
4337 "ICmp does not support folding a global value and "
4338 "a scale at the same time!");
4339 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
4341 if (C->getType() != OpTy)
4342 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
4346 CI->setOperand(1, C);
4353 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use
4354 /// of their operands effectively happens in their predecessor blocks, so the
4355 /// expression may need to be expanded in multiple places.
4356 void LSRInstance::RewriteForPHI(PHINode *PN,
4359 SCEVExpander &Rewriter,
4360 SmallVectorImpl<WeakVH> &DeadInsts,
4362 DenseMap<BasicBlock *, Value *> Inserted;
4363 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
4364 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
4365 BasicBlock *BB = PN->getIncomingBlock(i);
4367 // If this is a critical edge, split the edge so that we do not insert
4368 // the code on all predecessor/successor paths. We do this unless this
4369 // is the canonical backedge for this loop, which complicates post-inc
4371 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
4372 !isa<IndirectBrInst>(BB->getTerminator())) {
4373 BasicBlock *Parent = PN->getParent();
4374 Loop *PNLoop = LI.getLoopFor(Parent);
4375 if (!PNLoop || Parent != PNLoop->getHeader()) {
4376 // Split the critical edge.
4377 BasicBlock *NewBB = 0;
4378 if (!Parent->isLandingPad()) {
4379 NewBB = SplitCriticalEdge(BB, Parent, P,
4380 /*MergeIdenticalEdges=*/true,
4381 /*DontDeleteUselessPhis=*/true);
4383 SmallVector<BasicBlock*, 2> NewBBs;
4384 SplitLandingPadPredecessors(Parent, BB, "", "", P, NewBBs);
4388 // If PN is outside of the loop and BB is in the loop, we want to
4389 // move the block to be immediately before the PHI block, not
4390 // immediately after BB.
4391 if (L->contains(BB) && !L->contains(PN))
4392 NewBB->moveBefore(PN->getParent());
4394 // Splitting the edge can reduce the number of PHI entries we have.
4395 e = PN->getNumIncomingValues();
4397 i = PN->getBasicBlockIndex(BB);
4401 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
4402 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0)));
4404 PN->setIncomingValue(i, Pair.first->second);
4406 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts);
4408 // If this is reuse-by-noop-cast, insert the noop cast.
4409 Type *OpTy = LF.OperandValToReplace->getType();
4410 if (FullV->getType() != OpTy)
4412 CastInst::Create(CastInst::getCastOpcode(FullV, false,
4414 FullV, LF.OperandValToReplace->getType(),
4415 "tmp", BB->getTerminator());
4417 PN->setIncomingValue(i, FullV);
4418 Pair.first->second = FullV;
4423 /// Rewrite - Emit instructions for the leading candidate expression for this
4424 /// LSRUse (this is called "expanding"), and update the UserInst to reference
4425 /// the newly expanded value.
4426 void LSRInstance::Rewrite(const LSRFixup &LF,
4428 SCEVExpander &Rewriter,
4429 SmallVectorImpl<WeakVH> &DeadInsts,
4431 // First, find an insertion point that dominates UserInst. For PHI nodes,
4432 // find the nearest block which dominates all the relevant uses.
4433 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
4434 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P);
4436 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts);
4438 // If this is reuse-by-noop-cast, insert the noop cast.
4439 Type *OpTy = LF.OperandValToReplace->getType();
4440 if (FullV->getType() != OpTy) {
4442 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
4443 FullV, OpTy, "tmp", LF.UserInst);
4447 // Update the user. ICmpZero is handled specially here (for now) because
4448 // Expand may have updated one of the operands of the icmp already, and
4449 // its new value may happen to be equal to LF.OperandValToReplace, in
4450 // which case doing replaceUsesOfWith leads to replacing both operands
4451 // with the same value. TODO: Reorganize this.
4452 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
4453 LF.UserInst->setOperand(0, FullV);
4455 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
4458 DeadInsts.push_back(LF.OperandValToReplace);
4461 /// ImplementSolution - Rewrite all the fixup locations with new values,
4462 /// following the chosen solution.
4464 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
4466 // Keep track of instructions we may have made dead, so that
4467 // we can remove them after we are done working.
4468 SmallVector<WeakVH, 16> DeadInsts;
4470 SCEVExpander Rewriter(SE, "lsr");
4472 Rewriter.setDebugType(DEBUG_TYPE);
4474 Rewriter.disableCanonicalMode();
4475 Rewriter.enableLSRMode();
4476 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
4478 // Mark phi nodes that terminate chains so the expander tries to reuse them.
4479 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(),
4480 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) {
4481 if (PHINode *PN = dyn_cast<PHINode>(ChainI->back().UserInst))
4482 Rewriter.setChainedPhi(PN);
4485 // Expand the new value definitions and update the users.
4486 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
4487 E = Fixups.end(); I != E; ++I) {
4488 const LSRFixup &Fixup = *I;
4490 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P);
4495 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(),
4496 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) {
4497 GenerateIVChain(*ChainI, Rewriter, DeadInsts);
4500 // Clean up after ourselves. This must be done before deleting any
4504 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
4507 LSRInstance::LSRInstance(const TargetLowering *tli, Loop *l, Pass *P)
4508 : IU(P->getAnalysis<IVUsers>()),
4509 SE(P->getAnalysis<ScalarEvolution>()),
4510 DT(P->getAnalysis<DominatorTree>()),
4511 LI(P->getAnalysis<LoopInfo>()),
4512 TLI(tli), L(l), Changed(false), IVIncInsertPos(0) {
4514 // If LoopSimplify form is not available, stay out of trouble.
4515 if (!L->isLoopSimplifyForm())
4518 // If there's no interesting work to be done, bail early.
4519 if (IU.empty()) return;
4522 // All dominating loops must have preheaders, or SCEVExpander may not be able
4523 // to materialize an AddRecExpr whose Start is an outer AddRecExpr.
4525 // IVUsers analysis should only create users that are dominated by simple loop
4526 // headers. Since this loop should dominate all of its users, its user list
4527 // should be empty if this loop itself is not within a simple loop nest.
4528 for (DomTreeNode *Rung = DT.getNode(L->getLoopPreheader());
4529 Rung; Rung = Rung->getIDom()) {
4530 BasicBlock *BB = Rung->getBlock();
4531 const Loop *DomLoop = LI.getLoopFor(BB);
4532 if (DomLoop && DomLoop->getHeader() == BB) {
4533 assert(DomLoop->getLoopPreheader() && "LSR needs a simplified loop nest");
4538 DEBUG(dbgs() << "\nLSR on loop ";
4539 WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false);
4542 // First, perform some low-level loop optimizations.
4544 OptimizeLoopTermCond();
4546 // If loop preparation eliminates all interesting IV users, bail.
4547 if (IU.empty()) return;
4549 // Skip nested loops until we can model them better with formulae.
4551 DEBUG(dbgs() << "LSR skipping outer loop " << *L << "\n");
4555 // Start collecting data and preparing for the solver.
4557 CollectInterestingTypesAndFactors();
4558 CollectFixupsAndInitialFormulae();
4559 CollectLoopInvariantFixupsAndFormulae();
4561 assert(!Uses.empty() && "IVUsers reported at least one use");
4562 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
4563 print_uses(dbgs()));
4565 // Now use the reuse data to generate a bunch of interesting ways
4566 // to formulate the values needed for the uses.
4567 GenerateAllReuseFormulae();
4569 FilterOutUndesirableDedicatedRegisters();
4570 NarrowSearchSpaceUsingHeuristics();
4572 SmallVector<const Formula *, 8> Solution;
4575 // Release memory that is no longer needed.
4580 if (Solution.empty())
4584 // Formulae should be legal.
4585 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
4586 E = Uses.end(); I != E; ++I) {
4587 const LSRUse &LU = *I;
4588 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
4589 JE = LU.Formulae.end(); J != JE; ++J)
4590 assert(isLegalUse(J->AM, LU.MinOffset, LU.MaxOffset,
4591 LU.Kind, LU.AccessTy, TLI) &&
4592 "Illegal formula generated!");
4596 // Now that we've decided what we want, make it so.
4597 ImplementSolution(Solution, P);
4600 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
4601 if (Factors.empty() && Types.empty()) return;
4603 OS << "LSR has identified the following interesting factors and types: ";
4606 for (SmallSetVector<int64_t, 8>::const_iterator
4607 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
4608 if (!First) OS << ", ";
4613 for (SmallSetVector<Type *, 4>::const_iterator
4614 I = Types.begin(), E = Types.end(); I != E; ++I) {
4615 if (!First) OS << ", ";
4617 OS << '(' << **I << ')';
4622 void LSRInstance::print_fixups(raw_ostream &OS) const {
4623 OS << "LSR is examining the following fixup sites:\n";
4624 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
4625 E = Fixups.end(); I != E; ++I) {
4632 void LSRInstance::print_uses(raw_ostream &OS) const {
4633 OS << "LSR is examining the following uses:\n";
4634 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
4635 E = Uses.end(); I != E; ++I) {
4636 const LSRUse &LU = *I;
4640 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
4641 JE = LU.Formulae.end(); J != JE; ++J) {
4649 void LSRInstance::print(raw_ostream &OS) const {
4650 print_factors_and_types(OS);
4655 void LSRInstance::dump() const {
4656 print(errs()); errs() << '\n';
4661 class LoopStrengthReduce : public LoopPass {
4662 /// TLI - Keep a pointer of a TargetLowering to consult for determining
4663 /// transformation profitability.
4664 const TargetLowering *const TLI;
4667 static char ID; // Pass ID, replacement for typeid
4668 explicit LoopStrengthReduce(const TargetLowering *tli = 0);
4671 bool runOnLoop(Loop *L, LPPassManager &LPM);
4672 void getAnalysisUsage(AnalysisUsage &AU) const;
4677 char LoopStrengthReduce::ID = 0;
4678 INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce",
4679 "Loop Strength Reduction", false, false)
4680 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
4681 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
4682 INITIALIZE_PASS_DEPENDENCY(IVUsers)
4683 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
4684 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
4685 INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce",
4686 "Loop Strength Reduction", false, false)
4689 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
4690 return new LoopStrengthReduce(TLI);
4693 LoopStrengthReduce::LoopStrengthReduce(const TargetLowering *tli)
4694 : LoopPass(ID), TLI(tli) {
4695 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry());
4698 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
4699 // We split critical edges, so we change the CFG. However, we do update
4700 // many analyses if they are around.
4701 AU.addPreservedID(LoopSimplifyID);
4703 AU.addRequired<LoopInfo>();
4704 AU.addPreserved<LoopInfo>();
4705 AU.addRequiredID(LoopSimplifyID);
4706 AU.addRequired<DominatorTree>();
4707 AU.addPreserved<DominatorTree>();
4708 AU.addRequired<ScalarEvolution>();
4709 AU.addPreserved<ScalarEvolution>();
4710 // Requiring LoopSimplify a second time here prevents IVUsers from running
4711 // twice, since LoopSimplify was invalidated by running ScalarEvolution.
4712 AU.addRequiredID(LoopSimplifyID);
4713 AU.addRequired<IVUsers>();
4714 AU.addPreserved<IVUsers>();
4717 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
4718 bool Changed = false;
4720 // Run the main LSR transformation.
4721 Changed |= LSRInstance(TLI, L, this).getChanged();
4723 // Remove any extra phis created by processing inner loops.
4724 Changed |= DeleteDeadPHIs(L->getHeader());
4725 if (EnablePhiElim) {
4726 SmallVector<WeakVH, 16> DeadInsts;
4727 SCEVExpander Rewriter(getAnalysis<ScalarEvolution>(), "lsr");
4729 Rewriter.setDebugType(DEBUG_TYPE);
4731 unsigned numFolded = Rewriter.
4732 replaceCongruentIVs(L, &getAnalysis<DominatorTree>(), DeadInsts, TLI);
4735 DeleteTriviallyDeadInstructions(DeadInsts);
4736 DeleteDeadPHIs(L->getHeader());