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 the addressing mode BaseGV be changed to a ConstantExpr instead
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/ADT/DenseSet.h"
59 #include "llvm/ADT/SetVector.h"
60 #include "llvm/ADT/SmallBitVector.h"
61 #include "llvm/AddressingMode.h"
62 #include "llvm/Analysis/Dominators.h"
63 #include "llvm/Analysis/IVUsers.h"
64 #include "llvm/Analysis/LoopPass.h"
65 #include "llvm/Analysis/ScalarEvolutionExpander.h"
66 #include "llvm/Analysis/TargetTransformInfo.h"
67 #include "llvm/Assembly/Writer.h"
68 #include "llvm/IR/Constants.h"
69 #include "llvm/IR/DerivedTypes.h"
70 #include "llvm/IR/Instructions.h"
71 #include "llvm/IR/IntrinsicInst.h"
72 #include "llvm/Support/CommandLine.h"
73 #include "llvm/Support/Debug.h"
74 #include "llvm/Support/ValueHandle.h"
75 #include "llvm/Support/raw_ostream.h"
76 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
77 #include "llvm/Transforms/Utils/Local.h"
81 /// MaxIVUsers is an arbitrary threshold that provides an early opportunitiy for
82 /// bail out. This threshold is far beyond the number of users that LSR can
83 /// conceivably solve, so it should not affect generated code, but catches the
84 /// worst cases before LSR burns too much compile time and stack space.
85 static const unsigned MaxIVUsers = 200;
87 // Temporary flag to cleanup congruent phis after LSR phi expansion.
88 // It's currently disabled until we can determine whether it's truly useful or
89 // not. The flag should be removed after the v3.0 release.
90 // This is now needed for ivchains.
91 static cl::opt<bool> EnablePhiElim(
92 "enable-lsr-phielim", cl::Hidden, cl::init(true),
93 cl::desc("Enable LSR phi elimination"));
96 // Stress test IV chain generation.
97 static cl::opt<bool> StressIVChain(
98 "stress-ivchain", cl::Hidden, cl::init(false),
99 cl::desc("Stress test LSR IV chains"));
101 static bool StressIVChain = false;
106 /// RegSortData - This class holds data which is used to order reuse candidates.
109 /// UsedByIndices - This represents the set of LSRUse indices which reference
110 /// a particular register.
111 SmallBitVector UsedByIndices;
115 void print(raw_ostream &OS) const;
121 void RegSortData::print(raw_ostream &OS) const {
122 OS << "[NumUses=" << UsedByIndices.count() << ']';
125 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
126 void RegSortData::dump() const {
127 print(errs()); errs() << '\n';
133 /// RegUseTracker - Map register candidates to information about how they are
135 class RegUseTracker {
136 typedef DenseMap<const SCEV *, RegSortData> RegUsesTy;
138 RegUsesTy RegUsesMap;
139 SmallVector<const SCEV *, 16> RegSequence;
142 void CountRegister(const SCEV *Reg, size_t LUIdx);
143 void DropRegister(const SCEV *Reg, size_t LUIdx);
144 void SwapAndDropUse(size_t LUIdx, size_t LastLUIdx);
146 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
148 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
152 typedef SmallVectorImpl<const SCEV *>::iterator iterator;
153 typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator;
154 iterator begin() { return RegSequence.begin(); }
155 iterator end() { return RegSequence.end(); }
156 const_iterator begin() const { return RegSequence.begin(); }
157 const_iterator end() const { return RegSequence.end(); }
163 RegUseTracker::CountRegister(const SCEV *Reg, size_t LUIdx) {
164 std::pair<RegUsesTy::iterator, bool> Pair =
165 RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
166 RegSortData &RSD = Pair.first->second;
168 RegSequence.push_back(Reg);
169 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
170 RSD.UsedByIndices.set(LUIdx);
174 RegUseTracker::DropRegister(const SCEV *Reg, size_t LUIdx) {
175 RegUsesTy::iterator It = RegUsesMap.find(Reg);
176 assert(It != RegUsesMap.end());
177 RegSortData &RSD = It->second;
178 assert(RSD.UsedByIndices.size() > LUIdx);
179 RSD.UsedByIndices.reset(LUIdx);
183 RegUseTracker::SwapAndDropUse(size_t LUIdx, size_t LastLUIdx) {
184 assert(LUIdx <= LastLUIdx);
186 // Update RegUses. The data structure is not optimized for this purpose;
187 // we must iterate through it and update each of the bit vectors.
188 for (RegUsesTy::iterator I = RegUsesMap.begin(), E = RegUsesMap.end();
190 SmallBitVector &UsedByIndices = I->second.UsedByIndices;
191 if (LUIdx < UsedByIndices.size())
192 UsedByIndices[LUIdx] =
193 LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : 0;
194 UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx));
199 RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
200 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
201 if (I == RegUsesMap.end())
203 const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
204 int i = UsedByIndices.find_first();
205 if (i == -1) return false;
206 if ((size_t)i != LUIdx) return true;
207 return UsedByIndices.find_next(i) != -1;
210 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
211 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
212 assert(I != RegUsesMap.end() && "Unknown register!");
213 return I->second.UsedByIndices;
216 void RegUseTracker::clear() {
223 /// Formula - This class holds information that describes a formula for
224 /// computing satisfying a use. It may include broken-out immediates and scaled
227 /// AM - This is used to represent complex addressing, as well as other kinds
228 /// of interesting uses.
231 /// BaseRegs - The list of "base" registers for this use. When this is
232 /// non-empty, AM.HasBaseReg should be set to true.
233 SmallVector<const SCEV *, 2> BaseRegs;
235 /// ScaledReg - The 'scaled' register for this use. This should be non-null
236 /// when AM.Scale is not zero.
237 const SCEV *ScaledReg;
239 /// UnfoldedOffset - An additional constant offset which added near the
240 /// use. This requires a temporary register, but the offset itself can
241 /// live in an add immediate field rather than a register.
242 int64_t UnfoldedOffset;
244 Formula() : ScaledReg(0), UnfoldedOffset(0) {}
246 void InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE);
248 unsigned getNumRegs() const;
249 Type *getType() const;
251 void DeleteBaseReg(const SCEV *&S);
253 bool referencesReg(const SCEV *S) const;
254 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
255 const RegUseTracker &RegUses) const;
257 void print(raw_ostream &OS) const;
263 /// DoInitialMatch - Recursion helper for InitialMatch.
264 static void DoInitialMatch(const SCEV *S, Loop *L,
265 SmallVectorImpl<const SCEV *> &Good,
266 SmallVectorImpl<const SCEV *> &Bad,
267 ScalarEvolution &SE) {
268 // Collect expressions which properly dominate the loop header.
269 if (SE.properlyDominates(S, L->getHeader())) {
274 // Look at add operands.
275 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
276 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
278 DoInitialMatch(*I, L, Good, Bad, SE);
282 // Look at addrec operands.
283 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
284 if (!AR->getStart()->isZero()) {
285 DoInitialMatch(AR->getStart(), L, Good, Bad, SE);
286 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
287 AR->getStepRecurrence(SE),
288 // FIXME: AR->getNoWrapFlags()
289 AR->getLoop(), SCEV::FlagAnyWrap),
294 // Handle a multiplication by -1 (negation) if it didn't fold.
295 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
296 if (Mul->getOperand(0)->isAllOnesValue()) {
297 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
298 const SCEV *NewMul = SE.getMulExpr(Ops);
300 SmallVector<const SCEV *, 4> MyGood;
301 SmallVector<const SCEV *, 4> MyBad;
302 DoInitialMatch(NewMul, L, MyGood, MyBad, SE);
303 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
304 SE.getEffectiveSCEVType(NewMul->getType())));
305 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyGood.begin(),
306 E = MyGood.end(); I != E; ++I)
307 Good.push_back(SE.getMulExpr(NegOne, *I));
308 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyBad.begin(),
309 E = MyBad.end(); I != E; ++I)
310 Bad.push_back(SE.getMulExpr(NegOne, *I));
314 // Ok, we can't do anything interesting. Just stuff the whole thing into a
315 // register and hope for the best.
319 /// InitialMatch - Incorporate loop-variant parts of S into this Formula,
320 /// attempting to keep all loop-invariant and loop-computable values in a
321 /// single base register.
322 void Formula::InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) {
323 SmallVector<const SCEV *, 4> Good;
324 SmallVector<const SCEV *, 4> Bad;
325 DoInitialMatch(S, L, Good, Bad, SE);
327 const SCEV *Sum = SE.getAddExpr(Good);
329 BaseRegs.push_back(Sum);
330 AM.HasBaseReg = true;
333 const SCEV *Sum = SE.getAddExpr(Bad);
335 BaseRegs.push_back(Sum);
336 AM.HasBaseReg = true;
340 /// getNumRegs - Return the total number of register operands used by this
341 /// formula. This does not include register uses implied by non-constant
343 unsigned Formula::getNumRegs() const {
344 return !!ScaledReg + BaseRegs.size();
347 /// getType - Return the type of this formula, if it has one, or null
348 /// otherwise. This type is meaningless except for the bit size.
349 Type *Formula::getType() const {
350 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
351 ScaledReg ? ScaledReg->getType() :
352 AM.BaseGV ? AM.BaseGV->getType() :
356 /// DeleteBaseReg - Delete the given base reg from the BaseRegs list.
357 void Formula::DeleteBaseReg(const SCEV *&S) {
358 if (&S != &BaseRegs.back())
359 std::swap(S, BaseRegs.back());
363 /// referencesReg - Test if this formula references the given register.
364 bool Formula::referencesReg(const SCEV *S) const {
365 return S == ScaledReg ||
366 std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end();
369 /// hasRegsUsedByUsesOtherThan - Test whether this formula uses registers
370 /// which are used by uses other than the use with the given index.
371 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
372 const RegUseTracker &RegUses) const {
374 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
376 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
377 E = BaseRegs.end(); I != E; ++I)
378 if (RegUses.isRegUsedByUsesOtherThan(*I, LUIdx))
383 void Formula::print(raw_ostream &OS) const {
386 if (!First) OS << " + "; else First = false;
387 WriteAsOperand(OS, AM.BaseGV, /*PrintType=*/false);
389 if (AM.BaseOffs != 0) {
390 if (!First) OS << " + "; else First = false;
393 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
394 E = BaseRegs.end(); I != E; ++I) {
395 if (!First) OS << " + "; else First = false;
396 OS << "reg(" << **I << ')';
398 if (AM.HasBaseReg && BaseRegs.empty()) {
399 if (!First) OS << " + "; else First = false;
400 OS << "**error: HasBaseReg**";
401 } else if (!AM.HasBaseReg && !BaseRegs.empty()) {
402 if (!First) OS << " + "; else First = false;
403 OS << "**error: !HasBaseReg**";
406 if (!First) OS << " + "; else First = false;
407 OS << AM.Scale << "*reg(";
414 if (UnfoldedOffset != 0) {
415 if (!First) OS << " + "; else First = false;
416 OS << "imm(" << UnfoldedOffset << ')';
420 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
421 void Formula::dump() const {
422 print(errs()); errs() << '\n';
426 /// isAddRecSExtable - Return true if the given addrec can be sign-extended
427 /// without changing its value.
428 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
430 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1);
431 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
434 /// isAddSExtable - Return true if the given add can be sign-extended
435 /// without changing its value.
436 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
438 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
439 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
442 /// isMulSExtable - Return true if the given mul can be sign-extended
443 /// without changing its value.
444 static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
446 IntegerType::get(SE.getContext(),
447 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
448 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
451 /// getExactSDiv - Return an expression for LHS /s RHS, if it can be determined
452 /// and if the remainder is known to be zero, or null otherwise. If
453 /// IgnoreSignificantBits is true, expressions like (X * Y) /s Y are simplified
454 /// to Y, ignoring that the multiplication may overflow, which is useful when
455 /// the result will be used in a context where the most significant bits are
457 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
459 bool IgnoreSignificantBits = false) {
460 // Handle the trivial case, which works for any SCEV type.
462 return SE.getConstant(LHS->getType(), 1);
464 // Handle a few RHS special cases.
465 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
467 const APInt &RA = RC->getValue()->getValue();
468 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
470 if (RA.isAllOnesValue())
471 return SE.getMulExpr(LHS, RC);
472 // Handle x /s 1 as x.
477 // Check for a division of a constant by a constant.
478 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
481 const APInt &LA = C->getValue()->getValue();
482 const APInt &RA = RC->getValue()->getValue();
483 if (LA.srem(RA) != 0)
485 return SE.getConstant(LA.sdiv(RA));
488 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
489 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
490 if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) {
491 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
492 IgnoreSignificantBits);
494 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
495 IgnoreSignificantBits);
496 if (!Start) return 0;
497 // FlagNW is independent of the start value, step direction, and is
498 // preserved with smaller magnitude steps.
499 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
500 return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap);
505 // Distribute the sdiv over add operands, if the add doesn't overflow.
506 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
507 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
508 SmallVector<const SCEV *, 8> Ops;
509 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
511 const SCEV *Op = getExactSDiv(*I, RHS, SE,
512 IgnoreSignificantBits);
516 return SE.getAddExpr(Ops);
521 // Check for a multiply operand that we can pull RHS out of.
522 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
523 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
524 SmallVector<const SCEV *, 4> Ops;
526 for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end();
530 if (const SCEV *Q = getExactSDiv(S, RHS, SE,
531 IgnoreSignificantBits)) {
537 return Found ? SE.getMulExpr(Ops) : 0;
542 // Otherwise we don't know.
546 /// ExtractImmediate - If S involves the addition of a constant integer value,
547 /// return that integer value, and mutate S to point to a new SCEV with that
549 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
550 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
551 if (C->getValue()->getValue().getMinSignedBits() <= 64) {
552 S = SE.getConstant(C->getType(), 0);
553 return C->getValue()->getSExtValue();
555 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
556 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
557 int64_t Result = ExtractImmediate(NewOps.front(), SE);
559 S = SE.getAddExpr(NewOps);
561 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
562 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
563 int64_t Result = ExtractImmediate(NewOps.front(), SE);
565 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
566 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
573 /// ExtractSymbol - If S involves the addition of a GlobalValue address,
574 /// return that symbol, and mutate S to point to a new SCEV with that
576 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
577 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
578 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
579 S = SE.getConstant(GV->getType(), 0);
582 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
583 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
584 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
586 S = SE.getAddExpr(NewOps);
588 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
589 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
590 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
592 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
593 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
600 /// isAddressUse - Returns true if the specified instruction is using the
601 /// specified value as an address.
602 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
603 bool isAddress = isa<LoadInst>(Inst);
604 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
605 if (SI->getOperand(1) == OperandVal)
607 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
608 // Addressing modes can also be folded into prefetches and a variety
610 switch (II->getIntrinsicID()) {
612 case Intrinsic::prefetch:
613 case Intrinsic::x86_sse_storeu_ps:
614 case Intrinsic::x86_sse2_storeu_pd:
615 case Intrinsic::x86_sse2_storeu_dq:
616 case Intrinsic::x86_sse2_storel_dq:
617 if (II->getArgOperand(0) == OperandVal)
625 /// getAccessType - Return the type of the memory being accessed.
626 static Type *getAccessType(const Instruction *Inst) {
627 Type *AccessTy = Inst->getType();
628 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
629 AccessTy = SI->getOperand(0)->getType();
630 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
631 // Addressing modes can also be folded into prefetches and a variety
633 switch (II->getIntrinsicID()) {
635 case Intrinsic::x86_sse_storeu_ps:
636 case Intrinsic::x86_sse2_storeu_pd:
637 case Intrinsic::x86_sse2_storeu_dq:
638 case Intrinsic::x86_sse2_storel_dq:
639 AccessTy = II->getArgOperand(0)->getType();
644 // All pointers have the same requirements, so canonicalize them to an
645 // arbitrary pointer type to minimize variation.
646 if (PointerType *PTy = dyn_cast<PointerType>(AccessTy))
647 AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
648 PTy->getAddressSpace());
653 /// isExistingPhi - Return true if this AddRec is already a phi in its loop.
654 static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
655 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
656 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
657 if (SE.isSCEVable(PN->getType()) &&
658 (SE.getEffectiveSCEVType(PN->getType()) ==
659 SE.getEffectiveSCEVType(AR->getType())) &&
660 SE.getSCEV(PN) == AR)
666 /// Check if expanding this expression is likely to incur significant cost. This
667 /// is tricky because SCEV doesn't track which expressions are actually computed
668 /// by the current IR.
670 /// We currently allow expansion of IV increments that involve adds,
671 /// multiplication by constants, and AddRecs from existing phis.
673 /// TODO: Allow UDivExpr if we can find an existing IV increment that is an
674 /// obvious multiple of the UDivExpr.
675 static bool isHighCostExpansion(const SCEV *S,
676 SmallPtrSet<const SCEV*, 8> &Processed,
677 ScalarEvolution &SE) {
678 // Zero/One operand expressions
679 switch (S->getSCEVType()) {
684 return isHighCostExpansion(cast<SCEVTruncateExpr>(S)->getOperand(),
687 return isHighCostExpansion(cast<SCEVZeroExtendExpr>(S)->getOperand(),
690 return isHighCostExpansion(cast<SCEVSignExtendExpr>(S)->getOperand(),
694 if (!Processed.insert(S))
697 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
698 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
700 if (isHighCostExpansion(*I, Processed, SE))
706 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
707 if (Mul->getNumOperands() == 2) {
708 // Multiplication by a constant is ok
709 if (isa<SCEVConstant>(Mul->getOperand(0)))
710 return isHighCostExpansion(Mul->getOperand(1), Processed, SE);
712 // If we have the value of one operand, check if an existing
713 // multiplication already generates this expression.
714 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Mul->getOperand(1))) {
715 Value *UVal = U->getValue();
716 for (Value::use_iterator UI = UVal->use_begin(), UE = UVal->use_end();
718 // If U is a constant, it may be used by a ConstantExpr.
719 Instruction *User = dyn_cast<Instruction>(*UI);
720 if (User && User->getOpcode() == Instruction::Mul
721 && SE.isSCEVable(User->getType())) {
722 return SE.getSCEV(User) == Mul;
729 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
730 if (isExistingPhi(AR, SE))
734 // Fow now, consider any other type of expression (div/mul/min/max) high cost.
738 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
739 /// specified set are trivially dead, delete them and see if this makes any of
740 /// their operands subsequently dead.
742 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) {
743 bool Changed = false;
745 while (!DeadInsts.empty()) {
746 Value *V = DeadInsts.pop_back_val();
747 Instruction *I = dyn_cast_or_null<Instruction>(V);
749 if (I == 0 || !isInstructionTriviallyDead(I))
752 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
753 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
756 DeadInsts.push_back(U);
759 I->eraseFromParent();
768 /// Cost - This class is used to measure and compare candidate formulae.
770 /// TODO: Some of these could be merged. Also, a lexical ordering
771 /// isn't always optimal.
775 unsigned NumBaseAdds;
781 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0),
784 bool operator<(const Cost &Other) const;
789 // Once any of the metrics loses, they must all remain losers.
791 return ((NumRegs | AddRecCost | NumIVMuls | NumBaseAdds
792 | ImmCost | SetupCost) != ~0u)
793 || ((NumRegs & AddRecCost & NumIVMuls & NumBaseAdds
794 & ImmCost & SetupCost) == ~0u);
799 assert(isValid() && "invalid cost");
800 return NumRegs == ~0u;
803 void RateFormula(const Formula &F,
804 SmallPtrSet<const SCEV *, 16> &Regs,
805 const DenseSet<const SCEV *> &VisitedRegs,
807 const SmallVectorImpl<int64_t> &Offsets,
808 ScalarEvolution &SE, DominatorTree &DT,
809 SmallPtrSet<const SCEV *, 16> *LoserRegs = 0);
811 void print(raw_ostream &OS) const;
815 void RateRegister(const SCEV *Reg,
816 SmallPtrSet<const SCEV *, 16> &Regs,
818 ScalarEvolution &SE, DominatorTree &DT);
819 void RatePrimaryRegister(const SCEV *Reg,
820 SmallPtrSet<const SCEV *, 16> &Regs,
822 ScalarEvolution &SE, DominatorTree &DT,
823 SmallPtrSet<const SCEV *, 16> *LoserRegs);
828 /// RateRegister - Tally up interesting quantities from the given register.
829 void Cost::RateRegister(const SCEV *Reg,
830 SmallPtrSet<const SCEV *, 16> &Regs,
832 ScalarEvolution &SE, DominatorTree &DT) {
833 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
834 // If this is an addrec for another loop, don't second-guess its addrec phi
835 // nodes. LSR isn't currently smart enough to reason about more than one
836 // loop at a time. LSR has already run on inner loops, will not run on outer
837 // loops, and cannot be expected to change sibling loops.
838 if (AR->getLoop() != L) {
839 // If the AddRec exists, consider it's register free and leave it alone.
840 if (isExistingPhi(AR, SE))
843 // Otherwise, do not consider this formula at all.
847 AddRecCost += 1; /// TODO: This should be a function of the stride.
849 // Add the step value register, if it needs one.
850 // TODO: The non-affine case isn't precisely modeled here.
851 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
852 if (!Regs.count(AR->getOperand(1))) {
853 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
861 // Rough heuristic; favor registers which don't require extra setup
862 // instructions in the preheader.
863 if (!isa<SCEVUnknown>(Reg) &&
864 !isa<SCEVConstant>(Reg) &&
865 !(isa<SCEVAddRecExpr>(Reg) &&
866 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
867 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
870 NumIVMuls += isa<SCEVMulExpr>(Reg) &&
871 SE.hasComputableLoopEvolution(Reg, L);
874 /// RatePrimaryRegister - Record this register in the set. If we haven't seen it
875 /// before, rate it. Optional LoserRegs provides a way to declare any formula
876 /// that refers to one of those regs an instant loser.
877 void Cost::RatePrimaryRegister(const SCEV *Reg,
878 SmallPtrSet<const SCEV *, 16> &Regs,
880 ScalarEvolution &SE, DominatorTree &DT,
881 SmallPtrSet<const SCEV *, 16> *LoserRegs) {
882 if (LoserRegs && LoserRegs->count(Reg)) {
886 if (Regs.insert(Reg)) {
887 RateRegister(Reg, Regs, L, SE, DT);
889 LoserRegs->insert(Reg);
893 void Cost::RateFormula(const Formula &F,
894 SmallPtrSet<const SCEV *, 16> &Regs,
895 const DenseSet<const SCEV *> &VisitedRegs,
897 const SmallVectorImpl<int64_t> &Offsets,
898 ScalarEvolution &SE, DominatorTree &DT,
899 SmallPtrSet<const SCEV *, 16> *LoserRegs) {
900 // Tally up the registers.
901 if (const SCEV *ScaledReg = F.ScaledReg) {
902 if (VisitedRegs.count(ScaledReg)) {
906 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT, LoserRegs);
910 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
911 E = F.BaseRegs.end(); I != E; ++I) {
912 const SCEV *BaseReg = *I;
913 if (VisitedRegs.count(BaseReg)) {
917 RatePrimaryRegister(BaseReg, Regs, L, SE, DT, LoserRegs);
922 // Determine how many (unfolded) adds we'll need inside the loop.
923 size_t NumBaseParts = F.BaseRegs.size() + (F.UnfoldedOffset != 0);
924 if (NumBaseParts > 1)
925 NumBaseAdds += NumBaseParts - 1;
927 // Tally up the non-zero immediates.
928 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
929 E = Offsets.end(); I != E; ++I) {
930 int64_t Offset = (uint64_t)*I + F.AM.BaseOffs;
932 ImmCost += 64; // Handle symbolic values conservatively.
933 // TODO: This should probably be the pointer size.
934 else if (Offset != 0)
935 ImmCost += APInt(64, Offset, true).getMinSignedBits();
937 assert(isValid() && "invalid cost");
940 /// Loose - Set this cost to a losing value.
950 /// operator< - Choose the lower cost.
951 bool Cost::operator<(const Cost &Other) const {
952 if (NumRegs != Other.NumRegs)
953 return NumRegs < Other.NumRegs;
954 if (AddRecCost != Other.AddRecCost)
955 return AddRecCost < Other.AddRecCost;
956 if (NumIVMuls != Other.NumIVMuls)
957 return NumIVMuls < Other.NumIVMuls;
958 if (NumBaseAdds != Other.NumBaseAdds)
959 return NumBaseAdds < Other.NumBaseAdds;
960 if (ImmCost != Other.ImmCost)
961 return ImmCost < Other.ImmCost;
962 if (SetupCost != Other.SetupCost)
963 return SetupCost < Other.SetupCost;
967 void Cost::print(raw_ostream &OS) const {
968 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
970 OS << ", with addrec cost " << AddRecCost;
972 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
973 if (NumBaseAdds != 0)
974 OS << ", plus " << NumBaseAdds << " base add"
975 << (NumBaseAdds == 1 ? "" : "s");
977 OS << ", plus " << ImmCost << " imm cost";
979 OS << ", plus " << SetupCost << " setup cost";
982 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
983 void Cost::dump() const {
984 print(errs()); errs() << '\n';
990 /// LSRFixup - An operand value in an instruction which is to be replaced
991 /// with some equivalent, possibly strength-reduced, replacement.
993 /// UserInst - The instruction which will be updated.
994 Instruction *UserInst;
996 /// OperandValToReplace - The operand of the instruction which will
997 /// be replaced. The operand may be used more than once; every instance
998 /// will be replaced.
999 Value *OperandValToReplace;
1001 /// PostIncLoops - If this user is to use the post-incremented value of an
1002 /// induction variable, this variable is non-null and holds the loop
1003 /// associated with the induction variable.
1004 PostIncLoopSet PostIncLoops;
1006 /// LUIdx - The index of the LSRUse describing the expression which
1007 /// this fixup needs, minus an offset (below).
1010 /// Offset - A constant offset to be added to the LSRUse expression.
1011 /// This allows multiple fixups to share the same LSRUse with different
1012 /// offsets, for example in an unrolled loop.
1015 bool isUseFullyOutsideLoop(const Loop *L) const;
1019 void print(raw_ostream &OS) const;
1025 LSRFixup::LSRFixup()
1026 : UserInst(0), OperandValToReplace(0), LUIdx(~size_t(0)), Offset(0) {}
1028 /// isUseFullyOutsideLoop - Test whether this fixup always uses its
1029 /// value outside of the given loop.
1030 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
1031 // PHI nodes use their value in their incoming blocks.
1032 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
1033 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1034 if (PN->getIncomingValue(i) == OperandValToReplace &&
1035 L->contains(PN->getIncomingBlock(i)))
1040 return !L->contains(UserInst);
1043 void LSRFixup::print(raw_ostream &OS) const {
1045 // Store is common and interesting enough to be worth special-casing.
1046 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
1048 WriteAsOperand(OS, Store->getOperand(0), /*PrintType=*/false);
1049 } else if (UserInst->getType()->isVoidTy())
1050 OS << UserInst->getOpcodeName();
1052 WriteAsOperand(OS, UserInst, /*PrintType=*/false);
1054 OS << ", OperandValToReplace=";
1055 WriteAsOperand(OS, OperandValToReplace, /*PrintType=*/false);
1057 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(),
1058 E = PostIncLoops.end(); I != E; ++I) {
1059 OS << ", PostIncLoop=";
1060 WriteAsOperand(OS, (*I)->getHeader(), /*PrintType=*/false);
1063 if (LUIdx != ~size_t(0))
1064 OS << ", LUIdx=" << LUIdx;
1067 OS << ", Offset=" << Offset;
1070 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1071 void LSRFixup::dump() const {
1072 print(errs()); errs() << '\n';
1078 /// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding
1079 /// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*.
1080 struct UniquifierDenseMapInfo {
1081 static SmallVector<const SCEV *, 2> getEmptyKey() {
1082 SmallVector<const SCEV *, 2> V;
1083 V.push_back(reinterpret_cast<const SCEV *>(-1));
1087 static SmallVector<const SCEV *, 2> getTombstoneKey() {
1088 SmallVector<const SCEV *, 2> V;
1089 V.push_back(reinterpret_cast<const SCEV *>(-2));
1093 static unsigned getHashValue(const SmallVector<const SCEV *, 2> &V) {
1094 unsigned Result = 0;
1095 for (SmallVectorImpl<const SCEV *>::const_iterator I = V.begin(),
1096 E = V.end(); I != E; ++I)
1097 Result ^= DenseMapInfo<const SCEV *>::getHashValue(*I);
1101 static bool isEqual(const SmallVector<const SCEV *, 2> &LHS,
1102 const SmallVector<const SCEV *, 2> &RHS) {
1107 /// LSRUse - This class holds the state that LSR keeps for each use in
1108 /// IVUsers, as well as uses invented by LSR itself. It includes information
1109 /// about what kinds of things can be folded into the user, information about
1110 /// the user itself, and information about how the use may be satisfied.
1111 /// TODO: Represent multiple users of the same expression in common?
1113 DenseSet<SmallVector<const SCEV *, 2>, UniquifierDenseMapInfo> Uniquifier;
1116 /// KindType - An enum for a kind of use, indicating what types of
1117 /// scaled and immediate operands it might support.
1119 Basic, ///< A normal use, with no folding.
1120 Special, ///< A special case of basic, allowing -1 scales.
1121 Address, ///< An address use; folding according to TargetLowering
1122 ICmpZero ///< An equality icmp with both operands folded into one.
1123 // TODO: Add a generic icmp too?
1129 SmallVector<int64_t, 8> Offsets;
1133 /// AllFixupsOutsideLoop - This records whether all of the fixups using this
1134 /// LSRUse are outside of the loop, in which case some special-case heuristics
1136 bool AllFixupsOutsideLoop;
1138 /// WidestFixupType - This records the widest use type for any fixup using
1139 /// this LSRUse. FindUseWithSimilarFormula can't consider uses with different
1140 /// max fixup widths to be equivalent, because the narrower one may be relying
1141 /// on the implicit truncation to truncate away bogus bits.
1142 Type *WidestFixupType;
1144 /// Formulae - A list of ways to build a value that can satisfy this user.
1145 /// After the list is populated, one of these is selected heuristically and
1146 /// used to formulate a replacement for OperandValToReplace in UserInst.
1147 SmallVector<Formula, 12> Formulae;
1149 /// Regs - The set of register candidates used by all formulae in this LSRUse.
1150 SmallPtrSet<const SCEV *, 4> Regs;
1152 LSRUse(KindType K, Type *T) : Kind(K), AccessTy(T),
1153 MinOffset(INT64_MAX),
1154 MaxOffset(INT64_MIN),
1155 AllFixupsOutsideLoop(true),
1156 WidestFixupType(0) {}
1158 bool HasFormulaWithSameRegs(const Formula &F) const;
1159 bool InsertFormula(const Formula &F);
1160 void DeleteFormula(Formula &F);
1161 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1163 void print(raw_ostream &OS) const;
1169 /// HasFormula - Test whether this use as a formula which has the same
1170 /// registers as the given formula.
1171 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1172 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1173 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1174 // Unstable sort by host order ok, because this is only used for uniquifying.
1175 std::sort(Key.begin(), Key.end());
1176 return Uniquifier.count(Key);
1179 /// InsertFormula - If the given formula has not yet been inserted, add it to
1180 /// the list, and return true. Return false otherwise.
1181 bool LSRUse::InsertFormula(const Formula &F) {
1182 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1183 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1184 // Unstable sort by host order ok, because this is only used for uniquifying.
1185 std::sort(Key.begin(), Key.end());
1187 if (!Uniquifier.insert(Key).second)
1190 // Using a register to hold the value of 0 is not profitable.
1191 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1192 "Zero allocated in a scaled register!");
1194 for (SmallVectorImpl<const SCEV *>::const_iterator I =
1195 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I)
1196 assert(!(*I)->isZero() && "Zero allocated in a base register!");
1199 // Add the formula to the list.
1200 Formulae.push_back(F);
1202 // Record registers now being used by this use.
1203 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1208 /// DeleteFormula - Remove the given formula from this use's list.
1209 void LSRUse::DeleteFormula(Formula &F) {
1210 if (&F != &Formulae.back())
1211 std::swap(F, Formulae.back());
1212 Formulae.pop_back();
1215 /// RecomputeRegs - Recompute the Regs field, and update RegUses.
1216 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1217 // Now that we've filtered out some formulae, recompute the Regs set.
1218 SmallPtrSet<const SCEV *, 4> OldRegs = Regs;
1220 for (SmallVectorImpl<Formula>::const_iterator I = Formulae.begin(),
1221 E = Formulae.end(); I != E; ++I) {
1222 const Formula &F = *I;
1223 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1224 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1227 // Update the RegTracker.
1228 for (SmallPtrSet<const SCEV *, 4>::iterator I = OldRegs.begin(),
1229 E = OldRegs.end(); I != E; ++I)
1230 if (!Regs.count(*I))
1231 RegUses.DropRegister(*I, LUIdx);
1234 void LSRUse::print(raw_ostream &OS) const {
1235 OS << "LSR Use: Kind=";
1237 case Basic: OS << "Basic"; break;
1238 case Special: OS << "Special"; break;
1239 case ICmpZero: OS << "ICmpZero"; break;
1241 OS << "Address of ";
1242 if (AccessTy->isPointerTy())
1243 OS << "pointer"; // the full pointer type could be really verbose
1248 OS << ", Offsets={";
1249 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
1250 E = Offsets.end(); I != E; ++I) {
1252 if (llvm::next(I) != E)
1257 if (AllFixupsOutsideLoop)
1258 OS << ", all-fixups-outside-loop";
1260 if (WidestFixupType)
1261 OS << ", widest fixup type: " << *WidestFixupType;
1264 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1265 void LSRUse::dump() const {
1266 print(errs()); errs() << '\n';
1270 /// isLegalUse - Test whether the use described by AM is "legal", meaning it can
1271 /// be completely folded into the user instruction at isel time. This includes
1272 /// address-mode folding and special icmp tricks.
1273 static bool isLegalUse(const TargetTransformInfo &TTI, LSRUse::KindType Kind,
1274 Type *AccessTy, GlobalValue *BaseGV, int64_t BaseOffset,
1275 bool HasBaseReg, int64_t Scale) {
1277 case LSRUse::Address:
1278 return TTI.isLegalAddressingMode(AccessTy, BaseGV, BaseOffset, HasBaseReg, Scale);
1280 // Otherwise, just guess that reg+reg addressing is legal.
1283 case LSRUse::ICmpZero:
1284 // There's not even a target hook for querying whether it would be legal to
1285 // fold a GV into an ICmp.
1289 // ICmp only has two operands; don't allow more than two non-trivial parts.
1290 if (Scale != 0 && HasBaseReg && BaseOffset != 0)
1293 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1294 // putting the scaled register in the other operand of the icmp.
1295 if (Scale != 0 && Scale != -1)
1298 // If we have low-level target information, ask the target if it can fold an
1299 // integer immediate on an icmp.
1300 if (BaseOffset != 0) {
1302 // ICmpZero BaseReg + BaseOffset => ICmp BaseReg, -BaseOffset
1303 // ICmpZero -1*ScaleReg + BaseOffset => ICmp ScaleReg, BaseOffset
1304 // Offs is the ICmp immediate.
1306 // The cast does the right thing with INT64_MIN.
1307 BaseOffset = -(uint64_t)BaseOffset;
1308 return TTI.isLegalICmpImmediate(BaseOffset);
1311 // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg
1315 // Only handle single-register values.
1316 return !BaseGV && Scale == 0 && BaseOffset == 0;
1318 case LSRUse::Special:
1319 // Special case Basic to handle -1 scales.
1320 return !BaseGV && (Scale == 0 || Scale == -1) && BaseOffset == 0;
1323 llvm_unreachable("Invalid LSRUse Kind!");
1326 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1327 int64_t MaxOffset, LSRUse::KindType Kind, Type *AccessTy,
1328 GlobalValue *BaseGV, int64_t BaseOffset, bool HasBaseReg,
1330 // Check for overflow.
1331 if (((int64_t)((uint64_t)BaseOffset + MinOffset) > BaseOffset) !=
1334 MinOffset = (uint64_t)BaseOffset + MinOffset;
1335 if (((int64_t)((uint64_t)BaseOffset + MaxOffset) > BaseOffset) !=
1338 MaxOffset = (uint64_t)BaseOffset + MaxOffset;
1340 return isLegalUse(TTI, Kind, AccessTy, BaseGV, MinOffset, HasBaseReg,
1342 isLegalUse(TTI, Kind, AccessTy, BaseGV, MaxOffset, HasBaseReg, Scale);
1345 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1346 int64_t MaxOffset, LSRUse::KindType Kind, Type *AccessTy,
1348 return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, F.AM.BaseGV,
1349 F.AM.BaseOffs, F.AM.HasBaseReg, F.AM.Scale);
1352 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1353 LSRUse::KindType Kind, Type *AccessTy,
1354 GlobalValue *BaseGV, int64_t BaseOffset,
1356 // Fast-path: zero is always foldable.
1357 if (BaseOffset == 0 && !BaseGV) return true;
1359 // Conservatively, create an address with an immediate and a
1360 // base and a scale.
1361 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1363 // Canonicalize a scale of 1 to a base register if the formula doesn't
1364 // already have a base register.
1365 if (!HasBaseReg && Scale == 1) {
1370 return isLegalUse(TTI, Kind, AccessTy, BaseGV, BaseOffset, HasBaseReg, Scale);
1373 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1374 ScalarEvolution &SE, int64_t MinOffset,
1375 int64_t MaxOffset, LSRUse::KindType Kind,
1376 Type *AccessTy, const SCEV *S, bool HasBaseReg) {
1377 // Fast-path: zero is always foldable.
1378 if (S->isZero()) return true;
1380 // Conservatively, create an address with an immediate and a
1381 // base and a scale.
1382 int64_t BaseOffset = ExtractImmediate(S, SE);
1383 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1385 // If there's anything else involved, it's not foldable.
1386 if (!S->isZero()) return false;
1388 // Fast-path: zero is always foldable.
1389 if (BaseOffset == 0 && !BaseGV) return true;
1391 // Conservatively, create an address with an immediate and a
1392 // base and a scale.
1393 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1395 return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1396 BaseOffset, HasBaseReg, Scale);
1401 /// UseMapDenseMapInfo - A DenseMapInfo implementation for holding
1402 /// DenseMaps and DenseSets of pairs of const SCEV* and LSRUse::Kind.
1403 struct UseMapDenseMapInfo {
1404 static std::pair<const SCEV *, LSRUse::KindType> getEmptyKey() {
1405 return std::make_pair(reinterpret_cast<const SCEV *>(-1), LSRUse::Basic);
1408 static std::pair<const SCEV *, LSRUse::KindType> getTombstoneKey() {
1409 return std::make_pair(reinterpret_cast<const SCEV *>(-2), LSRUse::Basic);
1413 getHashValue(const std::pair<const SCEV *, LSRUse::KindType> &V) {
1414 unsigned Result = DenseMapInfo<const SCEV *>::getHashValue(V.first);
1415 Result ^= DenseMapInfo<unsigned>::getHashValue(unsigned(V.second));
1419 static bool isEqual(const std::pair<const SCEV *, LSRUse::KindType> &LHS,
1420 const std::pair<const SCEV *, LSRUse::KindType> &RHS) {
1425 /// IVInc - An individual increment in a Chain of IV increments.
1426 /// Relate an IV user to an expression that computes the IV it uses from the IV
1427 /// used by the previous link in the Chain.
1429 /// For the head of a chain, IncExpr holds the absolute SCEV expression for the
1430 /// original IVOperand. The head of the chain's IVOperand is only valid during
1431 /// chain collection, before LSR replaces IV users. During chain generation,
1432 /// IncExpr can be used to find the new IVOperand that computes the same
1435 Instruction *UserInst;
1437 const SCEV *IncExpr;
1439 IVInc(Instruction *U, Value *O, const SCEV *E):
1440 UserInst(U), IVOperand(O), IncExpr(E) {}
1443 // IVChain - The list of IV increments in program order.
1444 // We typically add the head of a chain without finding subsequent links.
1446 SmallVector<IVInc,1> Incs;
1447 const SCEV *ExprBase;
1449 IVChain() : ExprBase(0) {}
1451 IVChain(const IVInc &Head, const SCEV *Base)
1452 : Incs(1, Head), ExprBase(Base) {}
1454 typedef SmallVectorImpl<IVInc>::const_iterator const_iterator;
1456 // begin - return the first increment in the chain.
1457 const_iterator begin() const {
1458 assert(!Incs.empty());
1459 return llvm::next(Incs.begin());
1461 const_iterator end() const {
1465 // hasIncs - Returns true if this chain contains any increments.
1466 bool hasIncs() const { return Incs.size() >= 2; }
1468 // add - Add an IVInc to the end of this chain.
1469 void add(const IVInc &X) { Incs.push_back(X); }
1471 // tailUserInst - Returns the last UserInst in the chain.
1472 Instruction *tailUserInst() const { return Incs.back().UserInst; }
1474 // isProfitableIncrement - Returns true if IncExpr can be profitably added to
1476 bool isProfitableIncrement(const SCEV *OperExpr,
1477 const SCEV *IncExpr,
1481 /// ChainUsers - Helper for CollectChains to track multiple IV increment uses.
1482 /// Distinguish between FarUsers that definitely cross IV increments and
1483 /// NearUsers that may be used between IV increments.
1485 SmallPtrSet<Instruction*, 4> FarUsers;
1486 SmallPtrSet<Instruction*, 4> NearUsers;
1489 /// LSRInstance - This class holds state for the main loop strength reduction
1493 ScalarEvolution &SE;
1496 const TargetTransformInfo &TTI;
1500 /// IVIncInsertPos - This is the insert position that the current loop's
1501 /// induction variable increment should be placed. In simple loops, this is
1502 /// the latch block's terminator. But in more complicated cases, this is a
1503 /// position which will dominate all the in-loop post-increment users.
1504 Instruction *IVIncInsertPos;
1506 /// Factors - Interesting factors between use strides.
1507 SmallSetVector<int64_t, 8> Factors;
1509 /// Types - Interesting use types, to facilitate truncation reuse.
1510 SmallSetVector<Type *, 4> Types;
1512 /// Fixups - The list of operands which are to be replaced.
1513 SmallVector<LSRFixup, 16> Fixups;
1515 /// Uses - The list of interesting uses.
1516 SmallVector<LSRUse, 16> Uses;
1518 /// RegUses - Track which uses use which register candidates.
1519 RegUseTracker RegUses;
1521 // Limit the number of chains to avoid quadratic behavior. We don't expect to
1522 // have more than a few IV increment chains in a loop. Missing a Chain falls
1523 // back to normal LSR behavior for those uses.
1524 static const unsigned MaxChains = 8;
1526 /// IVChainVec - IV users can form a chain of IV increments.
1527 SmallVector<IVChain, MaxChains> IVChainVec;
1529 /// IVIncSet - IV users that belong to profitable IVChains.
1530 SmallPtrSet<Use*, MaxChains> IVIncSet;
1532 void OptimizeShadowIV();
1533 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1534 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1535 void OptimizeLoopTermCond();
1537 void ChainInstruction(Instruction *UserInst, Instruction *IVOper,
1538 SmallVectorImpl<ChainUsers> &ChainUsersVec);
1539 void FinalizeChain(IVChain &Chain);
1540 void CollectChains();
1541 void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
1542 SmallVectorImpl<WeakVH> &DeadInsts);
1544 void CollectInterestingTypesAndFactors();
1545 void CollectFixupsAndInitialFormulae();
1547 LSRFixup &getNewFixup() {
1548 Fixups.push_back(LSRFixup());
1549 return Fixups.back();
1552 // Support for sharing of LSRUses between LSRFixups.
1553 typedef DenseMap<std::pair<const SCEV *, LSRUse::KindType>,
1555 UseMapDenseMapInfo> UseMapTy;
1558 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1559 LSRUse::KindType Kind, Type *AccessTy);
1561 std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
1562 LSRUse::KindType Kind,
1565 void DeleteUse(LSRUse &LU, size_t LUIdx);
1567 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1569 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1570 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1571 void CountRegisters(const Formula &F, size_t LUIdx);
1572 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1574 void CollectLoopInvariantFixupsAndFormulae();
1576 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1577 unsigned Depth = 0);
1578 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1579 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1580 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1581 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1582 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1583 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1584 void GenerateCrossUseConstantOffsets();
1585 void GenerateAllReuseFormulae();
1587 void FilterOutUndesirableDedicatedRegisters();
1589 size_t EstimateSearchSpaceComplexity() const;
1590 void NarrowSearchSpaceByDetectingSupersets();
1591 void NarrowSearchSpaceByCollapsingUnrolledCode();
1592 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
1593 void NarrowSearchSpaceByPickingWinnerRegs();
1594 void NarrowSearchSpaceUsingHeuristics();
1596 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1598 SmallVectorImpl<const Formula *> &Workspace,
1599 const Cost &CurCost,
1600 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1601 DenseSet<const SCEV *> &VisitedRegs) const;
1602 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1604 BasicBlock::iterator
1605 HoistInsertPosition(BasicBlock::iterator IP,
1606 const SmallVectorImpl<Instruction *> &Inputs) const;
1607 BasicBlock::iterator
1608 AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1611 SCEVExpander &Rewriter) const;
1613 Value *Expand(const LSRFixup &LF,
1615 BasicBlock::iterator IP,
1616 SCEVExpander &Rewriter,
1617 SmallVectorImpl<WeakVH> &DeadInsts) const;
1618 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1620 SCEVExpander &Rewriter,
1621 SmallVectorImpl<WeakVH> &DeadInsts,
1623 void Rewrite(const LSRFixup &LF,
1625 SCEVExpander &Rewriter,
1626 SmallVectorImpl<WeakVH> &DeadInsts,
1628 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
1632 LSRInstance(Loop *L, Pass *P);
1634 bool getChanged() const { return Changed; }
1636 void print_factors_and_types(raw_ostream &OS) const;
1637 void print_fixups(raw_ostream &OS) const;
1638 void print_uses(raw_ostream &OS) const;
1639 void print(raw_ostream &OS) const;
1645 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1646 /// inside the loop then try to eliminate the cast operation.
1647 void LSRInstance::OptimizeShadowIV() {
1648 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1649 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1652 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1653 UI != E; /* empty */) {
1654 IVUsers::const_iterator CandidateUI = UI;
1656 Instruction *ShadowUse = CandidateUI->getUser();
1657 Type *DestTy = NULL;
1658 bool IsSigned = false;
1660 /* If shadow use is a int->float cast then insert a second IV
1661 to eliminate this cast.
1663 for (unsigned i = 0; i < n; ++i)
1669 for (unsigned i = 0; i < n; ++i, ++d)
1672 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
1674 DestTy = UCast->getDestTy();
1676 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
1678 DestTy = SCast->getDestTy();
1680 if (!DestTy) continue;
1682 // If target does not support DestTy natively then do not apply
1683 // this transformation.
1684 if (!TTI.isTypeLegal(DestTy)) continue;
1686 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1688 if (PH->getNumIncomingValues() != 2) continue;
1690 Type *SrcTy = PH->getType();
1691 int Mantissa = DestTy->getFPMantissaWidth();
1692 if (Mantissa == -1) continue;
1693 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1696 unsigned Entry, Latch;
1697 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1705 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1706 if (!Init) continue;
1707 Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
1708 (double)Init->getSExtValue() :
1709 (double)Init->getZExtValue());
1711 BinaryOperator *Incr =
1712 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1713 if (!Incr) continue;
1714 if (Incr->getOpcode() != Instruction::Add
1715 && Incr->getOpcode() != Instruction::Sub)
1718 /* Initialize new IV, double d = 0.0 in above example. */
1719 ConstantInt *C = NULL;
1720 if (Incr->getOperand(0) == PH)
1721 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1722 else if (Incr->getOperand(1) == PH)
1723 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1729 // Ignore negative constants, as the code below doesn't handle them
1730 // correctly. TODO: Remove this restriction.
1731 if (!C->getValue().isStrictlyPositive()) continue;
1733 /* Add new PHINode. */
1734 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
1736 /* create new increment. '++d' in above example. */
1737 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1738 BinaryOperator *NewIncr =
1739 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1740 Instruction::FAdd : Instruction::FSub,
1741 NewPH, CFP, "IV.S.next.", Incr);
1743 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1744 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1746 /* Remove cast operation */
1747 ShadowUse->replaceAllUsesWith(NewPH);
1748 ShadowUse->eraseFromParent();
1754 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1755 /// set the IV user and stride information and return true, otherwise return
1757 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
1758 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1759 if (UI->getUser() == Cond) {
1760 // NOTE: we could handle setcc instructions with multiple uses here, but
1761 // InstCombine does it as well for simple uses, it's not clear that it
1762 // occurs enough in real life to handle.
1769 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
1770 /// a max computation.
1772 /// This is a narrow solution to a specific, but acute, problem. For loops
1778 /// } while (++i < n);
1780 /// the trip count isn't just 'n', because 'n' might not be positive. And
1781 /// unfortunately this can come up even for loops where the user didn't use
1782 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1783 /// will commonly be lowered like this:
1789 /// } while (++i < n);
1792 /// and then it's possible for subsequent optimization to obscure the if
1793 /// test in such a way that indvars can't find it.
1795 /// When indvars can't find the if test in loops like this, it creates a
1796 /// max expression, which allows it to give the loop a canonical
1797 /// induction variable:
1800 /// max = n < 1 ? 1 : n;
1803 /// } while (++i != max);
1805 /// Canonical induction variables are necessary because the loop passes
1806 /// are designed around them. The most obvious example of this is the
1807 /// LoopInfo analysis, which doesn't remember trip count values. It
1808 /// expects to be able to rediscover the trip count each time it is
1809 /// needed, and it does this using a simple analysis that only succeeds if
1810 /// the loop has a canonical induction variable.
1812 /// However, when it comes time to generate code, the maximum operation
1813 /// can be quite costly, especially if it's inside of an outer loop.
1815 /// This function solves this problem by detecting this type of loop and
1816 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1817 /// the instructions for the maximum computation.
1819 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1820 // Check that the loop matches the pattern we're looking for.
1821 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1822 Cond->getPredicate() != CmpInst::ICMP_NE)
1825 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1826 if (!Sel || !Sel->hasOneUse()) return Cond;
1828 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1829 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1831 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
1833 // Add one to the backedge-taken count to get the trip count.
1834 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
1835 if (IterationCount != SE.getSCEV(Sel)) return Cond;
1837 // Check for a max calculation that matches the pattern. There's no check
1838 // for ICMP_ULE here because the comparison would be with zero, which
1839 // isn't interesting.
1840 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
1841 const SCEVNAryExpr *Max = 0;
1842 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
1843 Pred = ICmpInst::ICMP_SLE;
1845 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
1846 Pred = ICmpInst::ICMP_SLT;
1848 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
1849 Pred = ICmpInst::ICMP_ULT;
1856 // To handle a max with more than two operands, this optimization would
1857 // require additional checking and setup.
1858 if (Max->getNumOperands() != 2)
1861 const SCEV *MaxLHS = Max->getOperand(0);
1862 const SCEV *MaxRHS = Max->getOperand(1);
1864 // ScalarEvolution canonicalizes constants to the left. For < and >, look
1865 // for a comparison with 1. For <= and >=, a comparison with zero.
1867 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
1870 // Check the relevant induction variable for conformance to
1872 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
1873 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1874 if (!AR || !AR->isAffine() ||
1875 AR->getStart() != One ||
1876 AR->getStepRecurrence(SE) != One)
1879 assert(AR->getLoop() == L &&
1880 "Loop condition operand is an addrec in a different loop!");
1882 // Check the right operand of the select, and remember it, as it will
1883 // be used in the new comparison instruction.
1885 if (ICmpInst::isTrueWhenEqual(Pred)) {
1886 // Look for n+1, and grab n.
1887 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
1888 if (isa<ConstantInt>(BO->getOperand(1)) &&
1889 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1890 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1891 NewRHS = BO->getOperand(0);
1892 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
1893 if (isa<ConstantInt>(BO->getOperand(1)) &&
1894 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1895 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1896 NewRHS = BO->getOperand(0);
1899 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
1900 NewRHS = Sel->getOperand(1);
1901 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
1902 NewRHS = Sel->getOperand(2);
1903 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
1904 NewRHS = SU->getValue();
1906 // Max doesn't match expected pattern.
1909 // Determine the new comparison opcode. It may be signed or unsigned,
1910 // and the original comparison may be either equality or inequality.
1911 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
1912 Pred = CmpInst::getInversePredicate(Pred);
1914 // Ok, everything looks ok to change the condition into an SLT or SGE and
1915 // delete the max calculation.
1917 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
1919 // Delete the max calculation instructions.
1920 Cond->replaceAllUsesWith(NewCond);
1921 CondUse->setUser(NewCond);
1922 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
1923 Cond->eraseFromParent();
1924 Sel->eraseFromParent();
1925 if (Cmp->use_empty())
1926 Cmp->eraseFromParent();
1930 /// OptimizeLoopTermCond - Change loop terminating condition to use the
1931 /// postinc iv when possible.
1933 LSRInstance::OptimizeLoopTermCond() {
1934 SmallPtrSet<Instruction *, 4> PostIncs;
1936 BasicBlock *LatchBlock = L->getLoopLatch();
1937 SmallVector<BasicBlock*, 8> ExitingBlocks;
1938 L->getExitingBlocks(ExitingBlocks);
1940 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
1941 BasicBlock *ExitingBlock = ExitingBlocks[i];
1943 // Get the terminating condition for the loop if possible. If we
1944 // can, we want to change it to use a post-incremented version of its
1945 // induction variable, to allow coalescing the live ranges for the IV into
1946 // one register value.
1948 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
1951 // FIXME: Overly conservative, termination condition could be an 'or' etc..
1952 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
1955 // Search IVUsesByStride to find Cond's IVUse if there is one.
1956 IVStrideUse *CondUse = 0;
1957 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1958 if (!FindIVUserForCond(Cond, CondUse))
1961 // If the trip count is computed in terms of a max (due to ScalarEvolution
1962 // being unable to find a sufficient guard, for example), change the loop
1963 // comparison to use SLT or ULT instead of NE.
1964 // One consequence of doing this now is that it disrupts the count-down
1965 // optimization. That's not always a bad thing though, because in such
1966 // cases it may still be worthwhile to avoid a max.
1967 Cond = OptimizeMax(Cond, CondUse);
1969 // If this exiting block dominates the latch block, it may also use
1970 // the post-inc value if it won't be shared with other uses.
1971 // Check for dominance.
1972 if (!DT.dominates(ExitingBlock, LatchBlock))
1975 // Conservatively avoid trying to use the post-inc value in non-latch
1976 // exits if there may be pre-inc users in intervening blocks.
1977 if (LatchBlock != ExitingBlock)
1978 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1979 // Test if the use is reachable from the exiting block. This dominator
1980 // query is a conservative approximation of reachability.
1981 if (&*UI != CondUse &&
1982 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
1983 // Conservatively assume there may be reuse if the quotient of their
1984 // strides could be a legal scale.
1985 const SCEV *A = IU.getStride(*CondUse, L);
1986 const SCEV *B = IU.getStride(*UI, L);
1987 if (!A || !B) continue;
1988 if (SE.getTypeSizeInBits(A->getType()) !=
1989 SE.getTypeSizeInBits(B->getType())) {
1990 if (SE.getTypeSizeInBits(A->getType()) >
1991 SE.getTypeSizeInBits(B->getType()))
1992 B = SE.getSignExtendExpr(B, A->getType());
1994 A = SE.getSignExtendExpr(A, B->getType());
1996 if (const SCEVConstant *D =
1997 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
1998 const ConstantInt *C = D->getValue();
1999 // Stride of one or negative one can have reuse with non-addresses.
2000 if (C->isOne() || C->isAllOnesValue())
2001 goto decline_post_inc;
2002 // Avoid weird situations.
2003 if (C->getValue().getMinSignedBits() >= 64 ||
2004 C->getValue().isMinSignedValue())
2005 goto decline_post_inc;
2006 // Check for possible scaled-address reuse.
2007 Type *AccessTy = getAccessType(UI->getUser());
2008 int64_t Scale = C->getSExtValue();
2009 if (TTI.isLegalAddressingMode(AccessTy, /*BaseGV=*/ 0,
2011 /*HasBaseReg=*/ false, Scale))
2012 goto decline_post_inc;
2014 if (TTI.isLegalAddressingMode(AccessTy, /*BaseGV=*/ 0,
2016 /*HasBaseReg=*/ false, Scale))
2017 goto decline_post_inc;
2021 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
2024 // It's possible for the setcc instruction to be anywhere in the loop, and
2025 // possible for it to have multiple users. If it is not immediately before
2026 // the exiting block branch, move it.
2027 if (&*++BasicBlock::iterator(Cond) != TermBr) {
2028 if (Cond->hasOneUse()) {
2029 Cond->moveBefore(TermBr);
2031 // Clone the terminating condition and insert into the loopend.
2032 ICmpInst *OldCond = Cond;
2033 Cond = cast<ICmpInst>(Cond->clone());
2034 Cond->setName(L->getHeader()->getName() + ".termcond");
2035 ExitingBlock->getInstList().insert(TermBr, Cond);
2037 // Clone the IVUse, as the old use still exists!
2038 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
2039 TermBr->replaceUsesOfWith(OldCond, Cond);
2043 // If we get to here, we know that we can transform the setcc instruction to
2044 // use the post-incremented version of the IV, allowing us to coalesce the
2045 // live ranges for the IV correctly.
2046 CondUse->transformToPostInc(L);
2049 PostIncs.insert(Cond);
2053 // Determine an insertion point for the loop induction variable increment. It
2054 // must dominate all the post-inc comparisons we just set up, and it must
2055 // dominate the loop latch edge.
2056 IVIncInsertPos = L->getLoopLatch()->getTerminator();
2057 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(),
2058 E = PostIncs.end(); I != E; ++I) {
2060 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
2062 if (BB == (*I)->getParent())
2063 IVIncInsertPos = *I;
2064 else if (BB != IVIncInsertPos->getParent())
2065 IVIncInsertPos = BB->getTerminator();
2069 /// reconcileNewOffset - Determine if the given use can accommodate a fixup
2070 /// at the given offset and other details. If so, update the use and
2073 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
2074 LSRUse::KindType Kind, Type *AccessTy) {
2075 int64_t NewMinOffset = LU.MinOffset;
2076 int64_t NewMaxOffset = LU.MaxOffset;
2077 Type *NewAccessTy = AccessTy;
2079 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
2080 // something conservative, however this can pessimize in the case that one of
2081 // the uses will have all its uses outside the loop, for example.
2082 if (LU.Kind != Kind)
2084 // Conservatively assume HasBaseReg is true for now.
2085 if (NewOffset < LU.MinOffset) {
2086 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ 0,
2087 LU.MaxOffset - NewOffset, HasBaseReg))
2089 NewMinOffset = NewOffset;
2090 } else if (NewOffset > LU.MaxOffset) {
2091 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ 0,
2092 NewOffset - LU.MinOffset, HasBaseReg))
2094 NewMaxOffset = NewOffset;
2096 // Check for a mismatched access type, and fall back conservatively as needed.
2097 // TODO: Be less conservative when the type is similar and can use the same
2098 // addressing modes.
2099 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
2100 NewAccessTy = Type::getVoidTy(AccessTy->getContext());
2103 LU.MinOffset = NewMinOffset;
2104 LU.MaxOffset = NewMaxOffset;
2105 LU.AccessTy = NewAccessTy;
2106 if (NewOffset != LU.Offsets.back())
2107 LU.Offsets.push_back(NewOffset);
2111 /// getUse - Return an LSRUse index and an offset value for a fixup which
2112 /// needs the given expression, with the given kind and optional access type.
2113 /// Either reuse an existing use or create a new one, as needed.
2114 std::pair<size_t, int64_t>
2115 LSRInstance::getUse(const SCEV *&Expr,
2116 LSRUse::KindType Kind, Type *AccessTy) {
2117 const SCEV *Copy = Expr;
2118 int64_t Offset = ExtractImmediate(Expr, SE);
2120 // Basic uses can't accept any offset, for example.
2121 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ 0,
2122 Offset, /*HasBaseReg=*/ true)) {
2127 std::pair<UseMapTy::iterator, bool> P =
2128 UseMap.insert(std::make_pair(std::make_pair(Expr, Kind), 0));
2130 // A use already existed with this base.
2131 size_t LUIdx = P.first->second;
2132 LSRUse &LU = Uses[LUIdx];
2133 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
2135 return std::make_pair(LUIdx, Offset);
2138 // Create a new use.
2139 size_t LUIdx = Uses.size();
2140 P.first->second = LUIdx;
2141 Uses.push_back(LSRUse(Kind, AccessTy));
2142 LSRUse &LU = Uses[LUIdx];
2144 // We don't need to track redundant offsets, but we don't need to go out
2145 // of our way here to avoid them.
2146 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
2147 LU.Offsets.push_back(Offset);
2149 LU.MinOffset = Offset;
2150 LU.MaxOffset = Offset;
2151 return std::make_pair(LUIdx, Offset);
2154 /// DeleteUse - Delete the given use from the Uses list.
2155 void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
2156 if (&LU != &Uses.back())
2157 std::swap(LU, Uses.back());
2161 RegUses.SwapAndDropUse(LUIdx, Uses.size());
2164 /// FindUseWithFormula - Look for a use distinct from OrigLU which is has
2165 /// a formula that has the same registers as the given formula.
2167 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
2168 const LSRUse &OrigLU) {
2169 // Search all uses for the formula. This could be more clever.
2170 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2171 LSRUse &LU = Uses[LUIdx];
2172 // Check whether this use is close enough to OrigLU, to see whether it's
2173 // worthwhile looking through its formulae.
2174 // Ignore ICmpZero uses because they may contain formulae generated by
2175 // GenerateICmpZeroScales, in which case adding fixup offsets may
2177 if (&LU != &OrigLU &&
2178 LU.Kind != LSRUse::ICmpZero &&
2179 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
2180 LU.WidestFixupType == OrigLU.WidestFixupType &&
2181 LU.HasFormulaWithSameRegs(OrigF)) {
2182 // Scan through this use's formulae.
2183 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
2184 E = LU.Formulae.end(); I != E; ++I) {
2185 const Formula &F = *I;
2186 // Check to see if this formula has the same registers and symbols
2188 if (F.BaseRegs == OrigF.BaseRegs &&
2189 F.ScaledReg == OrigF.ScaledReg &&
2190 F.AM.BaseGV == OrigF.AM.BaseGV &&
2191 F.AM.Scale == OrigF.AM.Scale &&
2192 F.UnfoldedOffset == OrigF.UnfoldedOffset) {
2193 if (F.AM.BaseOffs == 0)
2195 // This is the formula where all the registers and symbols matched;
2196 // there aren't going to be any others. Since we declined it, we
2197 // can skip the rest of the formulae and proceed to the next LSRUse.
2204 // Nothing looked good.
2208 void LSRInstance::CollectInterestingTypesAndFactors() {
2209 SmallSetVector<const SCEV *, 4> Strides;
2211 // Collect interesting types and strides.
2212 SmallVector<const SCEV *, 4> Worklist;
2213 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2214 const SCEV *Expr = IU.getExpr(*UI);
2216 // Collect interesting types.
2217 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2219 // Add strides for mentioned loops.
2220 Worklist.push_back(Expr);
2222 const SCEV *S = Worklist.pop_back_val();
2223 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2224 if (AR->getLoop() == L)
2225 Strides.insert(AR->getStepRecurrence(SE));
2226 Worklist.push_back(AR->getStart());
2227 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2228 Worklist.append(Add->op_begin(), Add->op_end());
2230 } while (!Worklist.empty());
2233 // Compute interesting factors from the set of interesting strides.
2234 for (SmallSetVector<const SCEV *, 4>::const_iterator
2235 I = Strides.begin(), E = Strides.end(); I != E; ++I)
2236 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
2237 llvm::next(I); NewStrideIter != E; ++NewStrideIter) {
2238 const SCEV *OldStride = *I;
2239 const SCEV *NewStride = *NewStrideIter;
2241 if (SE.getTypeSizeInBits(OldStride->getType()) !=
2242 SE.getTypeSizeInBits(NewStride->getType())) {
2243 if (SE.getTypeSizeInBits(OldStride->getType()) >
2244 SE.getTypeSizeInBits(NewStride->getType()))
2245 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2247 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2249 if (const SCEVConstant *Factor =
2250 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2252 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2253 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2254 } else if (const SCEVConstant *Factor =
2255 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2258 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2259 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2263 // If all uses use the same type, don't bother looking for truncation-based
2265 if (Types.size() == 1)
2268 DEBUG(print_factors_and_types(dbgs()));
2271 /// findIVOperand - Helper for CollectChains that finds an IV operand (computed
2272 /// by an AddRec in this loop) within [OI,OE) or returns OE. If IVUsers mapped
2273 /// Instructions to IVStrideUses, we could partially skip this.
2274 static User::op_iterator
2275 findIVOperand(User::op_iterator OI, User::op_iterator OE,
2276 Loop *L, ScalarEvolution &SE) {
2277 for(; OI != OE; ++OI) {
2278 if (Instruction *Oper = dyn_cast<Instruction>(*OI)) {
2279 if (!SE.isSCEVable(Oper->getType()))
2282 if (const SCEVAddRecExpr *AR =
2283 dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) {
2284 if (AR->getLoop() == L)
2292 /// getWideOperand - IVChain logic must consistenctly peek base TruncInst
2293 /// operands, so wrap it in a convenient helper.
2294 static Value *getWideOperand(Value *Oper) {
2295 if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper))
2296 return Trunc->getOperand(0);
2300 /// isCompatibleIVType - Return true if we allow an IV chain to include both
2302 static bool isCompatibleIVType(Value *LVal, Value *RVal) {
2303 Type *LType = LVal->getType();
2304 Type *RType = RVal->getType();
2305 return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy());
2308 /// getExprBase - Return an approximation of this SCEV expression's "base", or
2309 /// NULL for any constant. Returning the expression itself is
2310 /// conservative. Returning a deeper subexpression is more precise and valid as
2311 /// long as it isn't less complex than another subexpression. For expressions
2312 /// involving multiple unscaled values, we need to return the pointer-type
2313 /// SCEVUnknown. This avoids forming chains across objects, such as:
2314 /// PrevOper==a[i], IVOper==b[i], IVInc==b-a.
2316 /// Since SCEVUnknown is the rightmost type, and pointers are the rightmost
2317 /// SCEVUnknown, we simply return the rightmost SCEV operand.
2318 static const SCEV *getExprBase(const SCEV *S) {
2319 switch (S->getSCEVType()) {
2320 default: // uncluding scUnknown.
2325 return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand());
2327 return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand());
2329 return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand());
2331 // Skip over scaled operands (scMulExpr) to follow add operands as long as
2332 // there's nothing more complex.
2333 // FIXME: not sure if we want to recognize negation.
2334 const SCEVAddExpr *Add = cast<SCEVAddExpr>(S);
2335 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(Add->op_end()),
2336 E(Add->op_begin()); I != E; ++I) {
2337 const SCEV *SubExpr = *I;
2338 if (SubExpr->getSCEVType() == scAddExpr)
2339 return getExprBase(SubExpr);
2341 if (SubExpr->getSCEVType() != scMulExpr)
2344 return S; // all operands are scaled, be conservative.
2347 return getExprBase(cast<SCEVAddRecExpr>(S)->getStart());
2351 /// Return true if the chain increment is profitable to expand into a loop
2352 /// invariant value, which may require its own register. A profitable chain
2353 /// increment will be an offset relative to the same base. We allow such offsets
2354 /// to potentially be used as chain increment as long as it's not obviously
2355 /// expensive to expand using real instructions.
2356 bool IVChain::isProfitableIncrement(const SCEV *OperExpr,
2357 const SCEV *IncExpr,
2358 ScalarEvolution &SE) {
2359 // Aggressively form chains when -stress-ivchain.
2363 // Do not replace a constant offset from IV head with a nonconstant IV
2365 if (!isa<SCEVConstant>(IncExpr)) {
2366 const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Incs[0].IVOperand));
2367 if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr)))
2371 SmallPtrSet<const SCEV*, 8> Processed;
2372 return !isHighCostExpansion(IncExpr, Processed, SE);
2375 /// Return true if the number of registers needed for the chain is estimated to
2376 /// be less than the number required for the individual IV users. First prohibit
2377 /// any IV users that keep the IV live across increments (the Users set should
2378 /// be empty). Next count the number and type of increments in the chain.
2380 /// Chaining IVs can lead to considerable code bloat if ISEL doesn't
2381 /// effectively use postinc addressing modes. Only consider it profitable it the
2382 /// increments can be computed in fewer registers when chained.
2384 /// TODO: Consider IVInc free if it's already used in another chains.
2386 isProfitableChain(IVChain &Chain, SmallPtrSet<Instruction*, 4> &Users,
2387 ScalarEvolution &SE, const TargetTransformInfo &TTI) {
2391 if (!Chain.hasIncs())
2394 if (!Users.empty()) {
2395 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " users:\n";
2396 for (SmallPtrSet<Instruction*, 4>::const_iterator I = Users.begin(),
2397 E = Users.end(); I != E; ++I) {
2398 dbgs() << " " << **I << "\n";
2402 assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
2404 // The chain itself may require a register, so intialize cost to 1.
2407 // A complete chain likely eliminates the need for keeping the original IV in
2408 // a register. LSR does not currently know how to form a complete chain unless
2409 // the header phi already exists.
2410 if (isa<PHINode>(Chain.tailUserInst())
2411 && SE.getSCEV(Chain.tailUserInst()) == Chain.Incs[0].IncExpr) {
2414 const SCEV *LastIncExpr = 0;
2415 unsigned NumConstIncrements = 0;
2416 unsigned NumVarIncrements = 0;
2417 unsigned NumReusedIncrements = 0;
2418 for (IVChain::const_iterator I = Chain.begin(), E = Chain.end();
2421 if (I->IncExpr->isZero())
2424 // Incrementing by zero or some constant is neutral. We assume constants can
2425 // be folded into an addressing mode or an add's immediate operand.
2426 if (isa<SCEVConstant>(I->IncExpr)) {
2427 ++NumConstIncrements;
2431 if (I->IncExpr == LastIncExpr)
2432 ++NumReusedIncrements;
2436 LastIncExpr = I->IncExpr;
2438 // An IV chain with a single increment is handled by LSR's postinc
2439 // uses. However, a chain with multiple increments requires keeping the IV's
2440 // value live longer than it needs to be if chained.
2441 if (NumConstIncrements > 1)
2444 // Materializing increment expressions in the preheader that didn't exist in
2445 // the original code may cost a register. For example, sign-extended array
2446 // indices can produce ridiculous increments like this:
2447 // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64)))
2448 cost += NumVarIncrements;
2450 // Reusing variable increments likely saves a register to hold the multiple of
2452 cost -= NumReusedIncrements;
2454 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " Cost: " << cost
2460 /// ChainInstruction - Add this IV user to an existing chain or make it the head
2462 void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper,
2463 SmallVectorImpl<ChainUsers> &ChainUsersVec) {
2464 // When IVs are used as types of varying widths, they are generally converted
2465 // to a wider type with some uses remaining narrow under a (free) trunc.
2466 Value *const NextIV = getWideOperand(IVOper);
2467 const SCEV *const OperExpr = SE.getSCEV(NextIV);
2468 const SCEV *const OperExprBase = getExprBase(OperExpr);
2470 // Visit all existing chains. Check if its IVOper can be computed as a
2471 // profitable loop invariant increment from the last link in the Chain.
2472 unsigned ChainIdx = 0, NChains = IVChainVec.size();
2473 const SCEV *LastIncExpr = 0;
2474 for (; ChainIdx < NChains; ++ChainIdx) {
2475 IVChain &Chain = IVChainVec[ChainIdx];
2477 // Prune the solution space aggressively by checking that both IV operands
2478 // are expressions that operate on the same unscaled SCEVUnknown. This
2479 // "base" will be canceled by the subsequent getMinusSCEV call. Checking
2480 // first avoids creating extra SCEV expressions.
2481 if (!StressIVChain && Chain.ExprBase != OperExprBase)
2484 Value *PrevIV = getWideOperand(Chain.Incs.back().IVOperand);
2485 if (!isCompatibleIVType(PrevIV, NextIV))
2488 // A phi node terminates a chain.
2489 if (isa<PHINode>(UserInst) && isa<PHINode>(Chain.tailUserInst()))
2492 // The increment must be loop-invariant so it can be kept in a register.
2493 const SCEV *PrevExpr = SE.getSCEV(PrevIV);
2494 const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr);
2495 if (!SE.isLoopInvariant(IncExpr, L))
2498 if (Chain.isProfitableIncrement(OperExpr, IncExpr, SE)) {
2499 LastIncExpr = IncExpr;
2503 // If we haven't found a chain, create a new one, unless we hit the max. Don't
2504 // bother for phi nodes, because they must be last in the chain.
2505 if (ChainIdx == NChains) {
2506 if (isa<PHINode>(UserInst))
2508 if (NChains >= MaxChains && !StressIVChain) {
2509 DEBUG(dbgs() << "IV Chain Limit\n");
2512 LastIncExpr = OperExpr;
2513 // IVUsers may have skipped over sign/zero extensions. We don't currently
2514 // attempt to form chains involving extensions unless they can be hoisted
2515 // into this loop's AddRec.
2516 if (!isa<SCEVAddRecExpr>(LastIncExpr))
2519 IVChainVec.push_back(IVChain(IVInc(UserInst, IVOper, LastIncExpr),
2521 ChainUsersVec.resize(NChains);
2522 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Head: (" << *UserInst
2523 << ") IV=" << *LastIncExpr << "\n");
2525 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Inc: (" << *UserInst
2526 << ") IV+" << *LastIncExpr << "\n");
2527 // Add this IV user to the end of the chain.
2528 IVChainVec[ChainIdx].add(IVInc(UserInst, IVOper, LastIncExpr));
2531 SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers;
2532 // This chain's NearUsers become FarUsers.
2533 if (!LastIncExpr->isZero()) {
2534 ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(),
2539 // All other uses of IVOperand become near uses of the chain.
2540 // We currently ignore intermediate values within SCEV expressions, assuming
2541 // they will eventually be used be the current chain, or can be computed
2542 // from one of the chain increments. To be more precise we could
2543 // transitively follow its user and only add leaf IV users to the set.
2544 for (Value::use_iterator UseIter = IVOper->use_begin(),
2545 UseEnd = IVOper->use_end(); UseIter != UseEnd; ++UseIter) {
2546 Instruction *OtherUse = dyn_cast<Instruction>(*UseIter);
2547 if (!OtherUse || OtherUse == UserInst)
2549 if (SE.isSCEVable(OtherUse->getType())
2550 && !isa<SCEVUnknown>(SE.getSCEV(OtherUse))
2551 && IU.isIVUserOrOperand(OtherUse)) {
2554 NearUsers.insert(OtherUse);
2557 // Since this user is part of the chain, it's no longer considered a use
2559 ChainUsersVec[ChainIdx].FarUsers.erase(UserInst);
2562 /// CollectChains - Populate the vector of Chains.
2564 /// This decreases ILP at the architecture level. Targets with ample registers,
2565 /// multiple memory ports, and no register renaming probably don't want
2566 /// this. However, such targets should probably disable LSR altogether.
2568 /// The job of LSR is to make a reasonable choice of induction variables across
2569 /// the loop. Subsequent passes can easily "unchain" computation exposing more
2570 /// ILP *within the loop* if the target wants it.
2572 /// Finding the best IV chain is potentially a scheduling problem. Since LSR
2573 /// will not reorder memory operations, it will recognize this as a chain, but
2574 /// will generate redundant IV increments. Ideally this would be corrected later
2575 /// by a smart scheduler:
2581 /// TODO: Walk the entire domtree within this loop, not just the path to the
2582 /// loop latch. This will discover chains on side paths, but requires
2583 /// maintaining multiple copies of the Chains state.
2584 void LSRInstance::CollectChains() {
2585 DEBUG(dbgs() << "Collecting IV Chains.\n");
2586 SmallVector<ChainUsers, 8> ChainUsersVec;
2588 SmallVector<BasicBlock *,8> LatchPath;
2589 BasicBlock *LoopHeader = L->getHeader();
2590 for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch());
2591 Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) {
2592 LatchPath.push_back(Rung->getBlock());
2594 LatchPath.push_back(LoopHeader);
2596 // Walk the instruction stream from the loop header to the loop latch.
2597 for (SmallVectorImpl<BasicBlock *>::reverse_iterator
2598 BBIter = LatchPath.rbegin(), BBEnd = LatchPath.rend();
2599 BBIter != BBEnd; ++BBIter) {
2600 for (BasicBlock::iterator I = (*BBIter)->begin(), E = (*BBIter)->end();
2602 // Skip instructions that weren't seen by IVUsers analysis.
2603 if (isa<PHINode>(I) || !IU.isIVUserOrOperand(I))
2606 // Ignore users that are part of a SCEV expression. This way we only
2607 // consider leaf IV Users. This effectively rediscovers a portion of
2608 // IVUsers analysis but in program order this time.
2609 if (SE.isSCEVable(I->getType()) && !isa<SCEVUnknown>(SE.getSCEV(I)))
2612 // Remove this instruction from any NearUsers set it may be in.
2613 for (unsigned ChainIdx = 0, NChains = IVChainVec.size();
2614 ChainIdx < NChains; ++ChainIdx) {
2615 ChainUsersVec[ChainIdx].NearUsers.erase(I);
2617 // Search for operands that can be chained.
2618 SmallPtrSet<Instruction*, 4> UniqueOperands;
2619 User::op_iterator IVOpEnd = I->op_end();
2620 User::op_iterator IVOpIter = findIVOperand(I->op_begin(), IVOpEnd, L, SE);
2621 while (IVOpIter != IVOpEnd) {
2622 Instruction *IVOpInst = cast<Instruction>(*IVOpIter);
2623 if (UniqueOperands.insert(IVOpInst))
2624 ChainInstruction(I, IVOpInst, ChainUsersVec);
2625 IVOpIter = findIVOperand(llvm::next(IVOpIter), IVOpEnd, L, SE);
2627 } // Continue walking down the instructions.
2628 } // Continue walking down the domtree.
2629 // Visit phi backedges to determine if the chain can generate the IV postinc.
2630 for (BasicBlock::iterator I = L->getHeader()->begin();
2631 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2632 if (!SE.isSCEVable(PN->getType()))
2636 dyn_cast<Instruction>(PN->getIncomingValueForBlock(L->getLoopLatch()));
2638 ChainInstruction(PN, IncV, ChainUsersVec);
2640 // Remove any unprofitable chains.
2641 unsigned ChainIdx = 0;
2642 for (unsigned UsersIdx = 0, NChains = IVChainVec.size();
2643 UsersIdx < NChains; ++UsersIdx) {
2644 if (!isProfitableChain(IVChainVec[UsersIdx],
2645 ChainUsersVec[UsersIdx].FarUsers, SE, TTI))
2647 // Preserve the chain at UsesIdx.
2648 if (ChainIdx != UsersIdx)
2649 IVChainVec[ChainIdx] = IVChainVec[UsersIdx];
2650 FinalizeChain(IVChainVec[ChainIdx]);
2653 IVChainVec.resize(ChainIdx);
2656 void LSRInstance::FinalizeChain(IVChain &Chain) {
2657 assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
2658 DEBUG(dbgs() << "Final Chain: " << *Chain.Incs[0].UserInst << "\n");
2660 for (IVChain::const_iterator I = Chain.begin(), E = Chain.end();
2662 DEBUG(dbgs() << " Inc: " << *I->UserInst << "\n");
2663 User::op_iterator UseI =
2664 std::find(I->UserInst->op_begin(), I->UserInst->op_end(), I->IVOperand);
2665 assert(UseI != I->UserInst->op_end() && "cannot find IV operand");
2666 IVIncSet.insert(UseI);
2670 /// Return true if the IVInc can be folded into an addressing mode.
2671 static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst,
2672 Value *Operand, const TargetTransformInfo &TTI) {
2673 const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr);
2674 if (!IncConst || !isAddressUse(UserInst, Operand))
2677 if (IncConst->getValue()->getValue().getMinSignedBits() > 64)
2680 int64_t IncOffset = IncConst->getValue()->getSExtValue();
2681 if (!isAlwaysFoldable(TTI, LSRUse::Address,
2682 getAccessType(UserInst), /*BaseGV=*/ 0,
2683 IncOffset, /*HaseBaseReg=*/ false))
2689 /// GenerateIVChains - Generate an add or subtract for each IVInc in a chain to
2690 /// materialize the IV user's operand from the previous IV user's operand.
2691 void LSRInstance::GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
2692 SmallVectorImpl<WeakVH> &DeadInsts) {
2693 // Find the new IVOperand for the head of the chain. It may have been replaced
2695 const IVInc &Head = Chain.Incs[0];
2696 User::op_iterator IVOpEnd = Head.UserInst->op_end();
2697 User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(),
2700 while (IVOpIter != IVOpEnd) {
2701 IVSrc = getWideOperand(*IVOpIter);
2703 // If this operand computes the expression that the chain needs, we may use
2704 // it. (Check this after setting IVSrc which is used below.)
2706 // Note that if Head.IncExpr is wider than IVSrc, then this phi is too
2707 // narrow for the chain, so we can no longer use it. We do allow using a
2708 // wider phi, assuming the LSR checked for free truncation. In that case we
2709 // should already have a truncate on this operand such that
2710 // getSCEV(IVSrc) == IncExpr.
2711 if (SE.getSCEV(*IVOpIter) == Head.IncExpr
2712 || SE.getSCEV(IVSrc) == Head.IncExpr) {
2715 IVOpIter = findIVOperand(llvm::next(IVOpIter), IVOpEnd, L, SE);
2717 if (IVOpIter == IVOpEnd) {
2718 // Gracefully give up on this chain.
2719 DEBUG(dbgs() << "Concealed chain head: " << *Head.UserInst << "\n");
2723 DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n");
2724 Type *IVTy = IVSrc->getType();
2725 Type *IntTy = SE.getEffectiveSCEVType(IVTy);
2726 const SCEV *LeftOverExpr = 0;
2727 for (IVChain::const_iterator IncI = Chain.begin(),
2728 IncE = Chain.end(); IncI != IncE; ++IncI) {
2730 Instruction *InsertPt = IncI->UserInst;
2731 if (isa<PHINode>(InsertPt))
2732 InsertPt = L->getLoopLatch()->getTerminator();
2734 // IVOper will replace the current IV User's operand. IVSrc is the IV
2735 // value currently held in a register.
2736 Value *IVOper = IVSrc;
2737 if (!IncI->IncExpr->isZero()) {
2738 // IncExpr was the result of subtraction of two narrow values, so must
2740 const SCEV *IncExpr = SE.getNoopOrSignExtend(IncI->IncExpr, IntTy);
2741 LeftOverExpr = LeftOverExpr ?
2742 SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr;
2744 if (LeftOverExpr && !LeftOverExpr->isZero()) {
2745 // Expand the IV increment.
2746 Rewriter.clearPostInc();
2747 Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt);
2748 const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc),
2749 SE.getUnknown(IncV));
2750 IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt);
2752 // If an IV increment can't be folded, use it as the next IV value.
2753 if (!canFoldIVIncExpr(LeftOverExpr, IncI->UserInst, IncI->IVOperand,
2755 assert(IVTy == IVOper->getType() && "inconsistent IV increment type");
2760 Type *OperTy = IncI->IVOperand->getType();
2761 if (IVTy != OperTy) {
2762 assert(SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) &&
2763 "cannot extend a chained IV");
2764 IRBuilder<> Builder(InsertPt);
2765 IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain");
2767 IncI->UserInst->replaceUsesOfWith(IncI->IVOperand, IVOper);
2768 DeadInsts.push_back(IncI->IVOperand);
2770 // If LSR created a new, wider phi, we may also replace its postinc. We only
2771 // do this if we also found a wide value for the head of the chain.
2772 if (isa<PHINode>(Chain.tailUserInst())) {
2773 for (BasicBlock::iterator I = L->getHeader()->begin();
2774 PHINode *Phi = dyn_cast<PHINode>(I); ++I) {
2775 if (!isCompatibleIVType(Phi, IVSrc))
2777 Instruction *PostIncV = dyn_cast<Instruction>(
2778 Phi->getIncomingValueForBlock(L->getLoopLatch()));
2779 if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc)))
2781 Value *IVOper = IVSrc;
2782 Type *PostIncTy = PostIncV->getType();
2783 if (IVTy != PostIncTy) {
2784 assert(PostIncTy->isPointerTy() && "mixing int/ptr IV types");
2785 IRBuilder<> Builder(L->getLoopLatch()->getTerminator());
2786 Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc());
2787 IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain");
2789 Phi->replaceUsesOfWith(PostIncV, IVOper);
2790 DeadInsts.push_back(PostIncV);
2795 void LSRInstance::CollectFixupsAndInitialFormulae() {
2796 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2797 Instruction *UserInst = UI->getUser();
2798 // Skip IV users that are part of profitable IV Chains.
2799 User::op_iterator UseI = std::find(UserInst->op_begin(), UserInst->op_end(),
2800 UI->getOperandValToReplace());
2801 assert(UseI != UserInst->op_end() && "cannot find IV operand");
2802 if (IVIncSet.count(UseI))
2806 LSRFixup &LF = getNewFixup();
2807 LF.UserInst = UserInst;
2808 LF.OperandValToReplace = UI->getOperandValToReplace();
2809 LF.PostIncLoops = UI->getPostIncLoops();
2811 LSRUse::KindType Kind = LSRUse::Basic;
2813 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
2814 Kind = LSRUse::Address;
2815 AccessTy = getAccessType(LF.UserInst);
2818 const SCEV *S = IU.getExpr(*UI);
2820 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
2821 // (N - i == 0), and this allows (N - i) to be the expression that we work
2822 // with rather than just N or i, so we can consider the register
2823 // requirements for both N and i at the same time. Limiting this code to
2824 // equality icmps is not a problem because all interesting loops use
2825 // equality icmps, thanks to IndVarSimplify.
2826 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
2827 if (CI->isEquality()) {
2828 // Swap the operands if needed to put the OperandValToReplace on the
2829 // left, for consistency.
2830 Value *NV = CI->getOperand(1);
2831 if (NV == LF.OperandValToReplace) {
2832 CI->setOperand(1, CI->getOperand(0));
2833 CI->setOperand(0, NV);
2834 NV = CI->getOperand(1);
2838 // x == y --> x - y == 0
2839 const SCEV *N = SE.getSCEV(NV);
2840 if (SE.isLoopInvariant(N, L) && isSafeToExpand(N)) {
2841 // S is normalized, so normalize N before folding it into S
2842 // to keep the result normalized.
2843 N = TransformForPostIncUse(Normalize, N, CI, 0,
2844 LF.PostIncLoops, SE, DT);
2845 Kind = LSRUse::ICmpZero;
2846 S = SE.getMinusSCEV(N, S);
2849 // -1 and the negations of all interesting strides (except the negation
2850 // of -1) are now also interesting.
2851 for (size_t i = 0, e = Factors.size(); i != e; ++i)
2852 if (Factors[i] != -1)
2853 Factors.insert(-(uint64_t)Factors[i]);
2857 // Set up the initial formula for this use.
2858 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
2860 LF.Offset = P.second;
2861 LSRUse &LU = Uses[LF.LUIdx];
2862 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2863 if (!LU.WidestFixupType ||
2864 SE.getTypeSizeInBits(LU.WidestFixupType) <
2865 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2866 LU.WidestFixupType = LF.OperandValToReplace->getType();
2868 // If this is the first use of this LSRUse, give it a formula.
2869 if (LU.Formulae.empty()) {
2870 InsertInitialFormula(S, LU, LF.LUIdx);
2871 CountRegisters(LU.Formulae.back(), LF.LUIdx);
2875 DEBUG(print_fixups(dbgs()));
2878 /// InsertInitialFormula - Insert a formula for the given expression into
2879 /// the given use, separating out loop-variant portions from loop-invariant
2880 /// and loop-computable portions.
2882 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
2884 F.InitialMatch(S, L, SE);
2885 bool Inserted = InsertFormula(LU, LUIdx, F);
2886 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
2889 /// InsertSupplementalFormula - Insert a simple single-register formula for
2890 /// the given expression into the given use.
2892 LSRInstance::InsertSupplementalFormula(const SCEV *S,
2893 LSRUse &LU, size_t LUIdx) {
2895 F.BaseRegs.push_back(S);
2896 F.AM.HasBaseReg = true;
2897 bool Inserted = InsertFormula(LU, LUIdx, F);
2898 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
2901 /// CountRegisters - Note which registers are used by the given formula,
2902 /// updating RegUses.
2903 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
2905 RegUses.CountRegister(F.ScaledReg, LUIdx);
2906 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
2907 E = F.BaseRegs.end(); I != E; ++I)
2908 RegUses.CountRegister(*I, LUIdx);
2911 /// InsertFormula - If the given formula has not yet been inserted, add it to
2912 /// the list, and return true. Return false otherwise.
2913 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
2914 if (!LU.InsertFormula(F))
2917 CountRegisters(F, LUIdx);
2921 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of
2922 /// loop-invariant values which we're tracking. These other uses will pin these
2923 /// values in registers, making them less profitable for elimination.
2924 /// TODO: This currently misses non-constant addrec step registers.
2925 /// TODO: Should this give more weight to users inside the loop?
2927 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
2928 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
2929 SmallPtrSet<const SCEV *, 8> Inserted;
2931 while (!Worklist.empty()) {
2932 const SCEV *S = Worklist.pop_back_val();
2934 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
2935 Worklist.append(N->op_begin(), N->op_end());
2936 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
2937 Worklist.push_back(C->getOperand());
2938 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
2939 Worklist.push_back(D->getLHS());
2940 Worklist.push_back(D->getRHS());
2941 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
2942 if (!Inserted.insert(U)) continue;
2943 const Value *V = U->getValue();
2944 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
2945 // Look for instructions defined outside the loop.
2946 if (L->contains(Inst)) continue;
2947 } else if (isa<UndefValue>(V))
2948 // Undef doesn't have a live range, so it doesn't matter.
2950 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
2952 const Instruction *UserInst = dyn_cast<Instruction>(*UI);
2953 // Ignore non-instructions.
2956 // Ignore instructions in other functions (as can happen with
2958 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
2960 // Ignore instructions not dominated by the loop.
2961 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
2962 UserInst->getParent() :
2963 cast<PHINode>(UserInst)->getIncomingBlock(
2964 PHINode::getIncomingValueNumForOperand(UI.getOperandNo()));
2965 if (!DT.dominates(L->getHeader(), UseBB))
2967 // Ignore uses which are part of other SCEV expressions, to avoid
2968 // analyzing them multiple times.
2969 if (SE.isSCEVable(UserInst->getType())) {
2970 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
2971 // If the user is a no-op, look through to its uses.
2972 if (!isa<SCEVUnknown>(UserS))
2976 SE.getUnknown(const_cast<Instruction *>(UserInst)));
2980 // Ignore icmp instructions which are already being analyzed.
2981 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
2982 unsigned OtherIdx = !UI.getOperandNo();
2983 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
2984 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
2988 LSRFixup &LF = getNewFixup();
2989 LF.UserInst = const_cast<Instruction *>(UserInst);
2990 LF.OperandValToReplace = UI.getUse();
2991 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0);
2993 LF.Offset = P.second;
2994 LSRUse &LU = Uses[LF.LUIdx];
2995 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2996 if (!LU.WidestFixupType ||
2997 SE.getTypeSizeInBits(LU.WidestFixupType) <
2998 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2999 LU.WidestFixupType = LF.OperandValToReplace->getType();
3000 InsertSupplementalFormula(U, LU, LF.LUIdx);
3001 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
3008 /// CollectSubexprs - Split S into subexpressions which can be pulled out into
3009 /// separate registers. If C is non-null, multiply each subexpression by C.
3011 /// Return remainder expression after factoring the subexpressions captured by
3012 /// Ops. If Ops is complete, return NULL.
3013 static const SCEV *CollectSubexprs(const SCEV *S, const SCEVConstant *C,
3014 SmallVectorImpl<const SCEV *> &Ops,
3016 ScalarEvolution &SE,
3017 unsigned Depth = 0) {
3018 // Arbitrarily cap recursion to protect compile time.
3022 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3023 // Break out add operands.
3024 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
3026 const SCEV *Remainder = CollectSubexprs(*I, C, Ops, L, SE, Depth+1);
3028 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3031 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
3032 // Split a non-zero base out of an addrec.
3033 if (AR->getStart()->isZero())
3036 const SCEV *Remainder = CollectSubexprs(AR->getStart(),
3037 C, Ops, L, SE, Depth+1);
3038 // Split the non-zero AddRec unless it is part of a nested recurrence that
3039 // does not pertain to this loop.
3040 if (Remainder && (AR->getLoop() == L || !isa<SCEVAddRecExpr>(Remainder))) {
3041 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3044 if (Remainder != AR->getStart()) {
3046 Remainder = SE.getConstant(AR->getType(), 0);
3047 return SE.getAddRecExpr(Remainder,
3048 AR->getStepRecurrence(SE),
3050 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
3053 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3054 // Break (C * (a + b + c)) into C*a + C*b + C*c.
3055 if (Mul->getNumOperands() != 2)
3057 if (const SCEVConstant *Op0 =
3058 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
3059 C = C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0;
3060 const SCEV *Remainder =
3061 CollectSubexprs(Mul->getOperand(1), C, Ops, L, SE, Depth+1);
3063 Ops.push_back(SE.getMulExpr(C, Remainder));
3070 /// GenerateReassociations - Split out subexpressions from adds and the bases of
3072 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
3075 // Arbitrarily cap recursion to protect compile time.
3076 if (Depth >= 3) return;
3078 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3079 const SCEV *BaseReg = Base.BaseRegs[i];
3081 SmallVector<const SCEV *, 8> AddOps;
3082 const SCEV *Remainder = CollectSubexprs(BaseReg, 0, AddOps, L, SE);
3084 AddOps.push_back(Remainder);
3086 if (AddOps.size() == 1) continue;
3088 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
3089 JE = AddOps.end(); J != JE; ++J) {
3091 // Loop-variant "unknown" values are uninteresting; we won't be able to
3092 // do anything meaningful with them.
3093 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
3096 // Don't pull a constant into a register if the constant could be folded
3097 // into an immediate field.
3098 if (isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3099 LU.AccessTy, *J, Base.getNumRegs() > 1))
3102 // Collect all operands except *J.
3103 SmallVector<const SCEV *, 8> InnerAddOps
3104 (((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
3106 (llvm::next(J), ((const SmallVector<const SCEV *, 8> &)AddOps).end());
3108 // Don't leave just a constant behind in a register if the constant could
3109 // be folded into an immediate field.
3110 if (InnerAddOps.size() == 1 &&
3111 isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3112 LU.AccessTy, InnerAddOps[0], Base.getNumRegs() > 1))
3115 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
3116 if (InnerSum->isZero())
3120 // Add the remaining pieces of the add back into the new formula.
3121 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
3123 SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
3124 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3125 InnerSumSC->getValue()->getZExtValue())) {
3126 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
3127 InnerSumSC->getValue()->getZExtValue();
3128 F.BaseRegs.erase(F.BaseRegs.begin() + i);
3130 F.BaseRegs[i] = InnerSum;
3132 // Add J as its own register, or an unfolded immediate.
3133 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
3134 if (SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
3135 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3136 SC->getValue()->getZExtValue()))
3137 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
3138 SC->getValue()->getZExtValue();
3140 F.BaseRegs.push_back(*J);
3142 if (InsertFormula(LU, LUIdx, F))
3143 // If that formula hadn't been seen before, recurse to find more like
3145 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1);
3150 /// GenerateCombinations - Generate a formula consisting of all of the
3151 /// loop-dominating registers added into a single register.
3152 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
3154 // This method is only interesting on a plurality of registers.
3155 if (Base.BaseRegs.size() <= 1) return;
3159 SmallVector<const SCEV *, 4> Ops;
3160 for (SmallVectorImpl<const SCEV *>::const_iterator
3161 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
3162 const SCEV *BaseReg = *I;
3163 if (SE.properlyDominates(BaseReg, L->getHeader()) &&
3164 !SE.hasComputableLoopEvolution(BaseReg, L))
3165 Ops.push_back(BaseReg);
3167 F.BaseRegs.push_back(BaseReg);
3169 if (Ops.size() > 1) {
3170 const SCEV *Sum = SE.getAddExpr(Ops);
3171 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
3172 // opportunity to fold something. For now, just ignore such cases
3173 // rather than proceed with zero in a register.
3174 if (!Sum->isZero()) {
3175 F.BaseRegs.push_back(Sum);
3176 (void)InsertFormula(LU, LUIdx, F);
3181 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
3182 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
3184 // We can't add a symbolic offset if the address already contains one.
3185 if (Base.AM.BaseGV) return;
3187 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3188 const SCEV *G = Base.BaseRegs[i];
3189 GlobalValue *GV = ExtractSymbol(G, SE);
3190 if (G->isZero() || !GV)
3194 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3197 (void)InsertFormula(LU, LUIdx, F);
3201 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
3202 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
3204 // TODO: For now, just add the min and max offset, because it usually isn't
3205 // worthwhile looking at everything inbetween.
3206 SmallVector<int64_t, 2> Worklist;
3207 Worklist.push_back(LU.MinOffset);
3208 if (LU.MaxOffset != LU.MinOffset)
3209 Worklist.push_back(LU.MaxOffset);
3211 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3212 const SCEV *G = Base.BaseRegs[i];
3214 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
3215 E = Worklist.end(); I != E; ++I) {
3217 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs - *I;
3218 if (isLegalUse(TTI, LU.MinOffset - *I, LU.MaxOffset - *I, LU.Kind,
3220 // Add the offset to the base register.
3221 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), *I), G);
3222 // If it cancelled out, drop the base register, otherwise update it.
3223 if (NewG->isZero()) {
3224 std::swap(F.BaseRegs[i], F.BaseRegs.back());
3225 F.BaseRegs.pop_back();
3227 F.BaseRegs[i] = NewG;
3229 (void)InsertFormula(LU, LUIdx, F);
3233 int64_t Imm = ExtractImmediate(G, SE);
3234 if (G->isZero() || Imm == 0)
3237 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Imm;
3238 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3241 (void)InsertFormula(LU, LUIdx, F);
3245 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
3246 /// the comparison. For example, x == y -> x*c == y*c.
3247 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
3249 if (LU.Kind != LSRUse::ICmpZero) return;
3251 // Determine the integer type for the base formula.
3252 Type *IntTy = Base.getType();
3254 if (SE.getTypeSizeInBits(IntTy) > 64) return;
3256 // Don't do this if there is more than one offset.
3257 if (LU.MinOffset != LU.MaxOffset) return;
3259 assert(!Base.AM.BaseGV && "ICmpZero use is not legal!");
3261 // Check each interesting stride.
3262 for (SmallSetVector<int64_t, 8>::const_iterator
3263 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3264 int64_t Factor = *I;
3266 // Check that the multiplication doesn't overflow.
3267 if (Base.AM.BaseOffs == INT64_MIN && Factor == -1)
3269 int64_t NewBaseOffs = (uint64_t)Base.AM.BaseOffs * Factor;
3270 if (NewBaseOffs / Factor != Base.AM.BaseOffs)
3273 // Check that multiplying with the use offset doesn't overflow.
3274 int64_t Offset = LU.MinOffset;
3275 if (Offset == INT64_MIN && Factor == -1)
3277 Offset = (uint64_t)Offset * Factor;
3278 if (Offset / Factor != LU.MinOffset)
3282 F.AM.BaseOffs = NewBaseOffs;
3284 // Check that this scale is legal.
3285 if (!isLegalUse(TTI, Offset, Offset, LU.Kind, LU.AccessTy, F))
3288 // Compensate for the use having MinOffset built into it.
3289 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Offset - LU.MinOffset;
3291 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3293 // Check that multiplying with each base register doesn't overflow.
3294 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
3295 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
3296 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
3300 // Check that multiplying with the scaled register doesn't overflow.
3302 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
3303 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
3307 // Check that multiplying with the unfolded offset doesn't overflow.
3308 if (F.UnfoldedOffset != 0) {
3309 if (F.UnfoldedOffset == INT64_MIN && Factor == -1)
3311 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
3312 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
3316 // If we make it here and it's legal, add it.
3317 (void)InsertFormula(LU, LUIdx, F);
3322 /// GenerateScales - Generate stride factor reuse formulae by making use of
3323 /// scaled-offset address modes, for example.
3324 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
3325 // Determine the integer type for the base formula.
3326 Type *IntTy = Base.getType();
3329 // If this Formula already has a scaled register, we can't add another one.
3330 if (Base.AM.Scale != 0) return;
3332 // Check each interesting stride.
3333 for (SmallSetVector<int64_t, 8>::const_iterator
3334 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3335 int64_t Factor = *I;
3337 Base.AM.Scale = Factor;
3338 Base.AM.HasBaseReg = Base.BaseRegs.size() > 1;
3339 // Check whether this scale is going to be legal.
3340 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3342 // As a special-case, handle special out-of-loop Basic users specially.
3343 // TODO: Reconsider this special case.
3344 if (LU.Kind == LSRUse::Basic &&
3345 isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LSRUse::Special,
3346 LU.AccessTy, Base) &&
3347 LU.AllFixupsOutsideLoop)
3348 LU.Kind = LSRUse::Special;
3352 // For an ICmpZero, negating a solitary base register won't lead to
3354 if (LU.Kind == LSRUse::ICmpZero &&
3355 !Base.AM.HasBaseReg && Base.AM.BaseOffs == 0 && !Base.AM.BaseGV)
3357 // For each addrec base reg, apply the scale, if possible.
3358 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3359 if (const SCEVAddRecExpr *AR =
3360 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
3361 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3362 if (FactorS->isZero())
3364 // Divide out the factor, ignoring high bits, since we'll be
3365 // scaling the value back up in the end.
3366 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
3367 // TODO: This could be optimized to avoid all the copying.
3369 F.ScaledReg = Quotient;
3370 F.DeleteBaseReg(F.BaseRegs[i]);
3371 (void)InsertFormula(LU, LUIdx, F);
3377 /// GenerateTruncates - Generate reuse formulae from different IV types.
3378 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
3379 // Don't bother truncating symbolic values.
3380 if (Base.AM.BaseGV) return;
3382 // Determine the integer type for the base formula.
3383 Type *DstTy = Base.getType();
3385 DstTy = SE.getEffectiveSCEVType(DstTy);
3387 for (SmallSetVector<Type *, 4>::const_iterator
3388 I = Types.begin(), E = Types.end(); I != E; ++I) {
3390 if (SrcTy != DstTy && TTI.isTruncateFree(SrcTy, DstTy)) {
3393 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
3394 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
3395 JE = F.BaseRegs.end(); J != JE; ++J)
3396 *J = SE.getAnyExtendExpr(*J, SrcTy);
3398 // TODO: This assumes we've done basic processing on all uses and
3399 // have an idea what the register usage is.
3400 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
3403 (void)InsertFormula(LU, LUIdx, F);
3410 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
3411 /// defer modifications so that the search phase doesn't have to worry about
3412 /// the data structures moving underneath it.
3416 const SCEV *OrigReg;
3418 WorkItem(size_t LI, int64_t I, const SCEV *R)
3419 : LUIdx(LI), Imm(I), OrigReg(R) {}
3421 void print(raw_ostream &OS) const;
3427 void WorkItem::print(raw_ostream &OS) const {
3428 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
3429 << " , add offset " << Imm;
3432 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
3433 void WorkItem::dump() const {
3434 print(errs()); errs() << '\n';
3438 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
3439 /// distance apart and try to form reuse opportunities between them.
3440 void LSRInstance::GenerateCrossUseConstantOffsets() {
3441 // Group the registers by their value without any added constant offset.
3442 typedef std::map<int64_t, const SCEV *> ImmMapTy;
3443 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
3445 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
3446 SmallVector<const SCEV *, 8> Sequence;
3447 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
3449 const SCEV *Reg = *I;
3450 int64_t Imm = ExtractImmediate(Reg, SE);
3451 std::pair<RegMapTy::iterator, bool> Pair =
3452 Map.insert(std::make_pair(Reg, ImmMapTy()));
3454 Sequence.push_back(Reg);
3455 Pair.first->second.insert(std::make_pair(Imm, *I));
3456 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
3459 // Now examine each set of registers with the same base value. Build up
3460 // a list of work to do and do the work in a separate step so that we're
3461 // not adding formulae and register counts while we're searching.
3462 SmallVector<WorkItem, 32> WorkItems;
3463 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
3464 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
3465 E = Sequence.end(); I != E; ++I) {
3466 const SCEV *Reg = *I;
3467 const ImmMapTy &Imms = Map.find(Reg)->second;
3469 // It's not worthwhile looking for reuse if there's only one offset.
3470 if (Imms.size() == 1)
3473 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
3474 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3476 dbgs() << ' ' << J->first;
3479 // Examine each offset.
3480 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3482 const SCEV *OrigReg = J->second;
3484 int64_t JImm = J->first;
3485 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
3487 if (!isa<SCEVConstant>(OrigReg) &&
3488 UsedByIndicesMap[Reg].count() == 1) {
3489 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
3493 // Conservatively examine offsets between this orig reg a few selected
3495 ImmMapTy::const_iterator OtherImms[] = {
3496 Imms.begin(), prior(Imms.end()),
3497 Imms.lower_bound((Imms.begin()->first + prior(Imms.end())->first) / 2)
3499 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
3500 ImmMapTy::const_iterator M = OtherImms[i];
3501 if (M == J || M == JE) continue;
3503 // Compute the difference between the two.
3504 int64_t Imm = (uint64_t)JImm - M->first;
3505 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
3506 LUIdx = UsedByIndices.find_next(LUIdx))
3507 // Make a memo of this use, offset, and register tuple.
3508 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
3509 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
3516 UsedByIndicesMap.clear();
3517 UniqueItems.clear();
3519 // Now iterate through the worklist and add new formulae.
3520 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
3521 E = WorkItems.end(); I != E; ++I) {
3522 const WorkItem &WI = *I;
3523 size_t LUIdx = WI.LUIdx;
3524 LSRUse &LU = Uses[LUIdx];
3525 int64_t Imm = WI.Imm;
3526 const SCEV *OrigReg = WI.OrigReg;
3528 Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
3529 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
3530 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
3532 // TODO: Use a more targeted data structure.
3533 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
3534 const Formula &F = LU.Formulae[L];
3535 // Use the immediate in the scaled register.
3536 if (F.ScaledReg == OrigReg) {
3537 int64_t Offs = (uint64_t)F.AM.BaseOffs +
3538 Imm * (uint64_t)F.AM.Scale;
3539 // Don't create 50 + reg(-50).
3540 if (F.referencesReg(SE.getSCEV(
3541 ConstantInt::get(IntTy, -(uint64_t)Offs))))
3544 NewF.AM.BaseOffs = Offs;
3545 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3548 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
3550 // If the new scale is a constant in a register, and adding the constant
3551 // value to the immediate would produce a value closer to zero than the
3552 // immediate itself, then the formula isn't worthwhile.
3553 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
3554 if (C->getValue()->isNegative() !=
3555 (NewF.AM.BaseOffs < 0) &&
3556 (C->getValue()->getValue().abs() * APInt(BitWidth, F.AM.Scale))
3557 .ule(abs64(NewF.AM.BaseOffs)))
3561 (void)InsertFormula(LU, LUIdx, NewF);
3563 // Use the immediate in a base register.
3564 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
3565 const SCEV *BaseReg = F.BaseRegs[N];
3566 if (BaseReg != OrigReg)
3569 NewF.AM.BaseOffs = (uint64_t)NewF.AM.BaseOffs + Imm;
3570 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset,
3571 LU.Kind, LU.AccessTy, NewF)) {
3572 if (!TTI.isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
3575 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
3577 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
3579 // If the new formula has a constant in a register, and adding the
3580 // constant value to the immediate would produce a value closer to
3581 // zero than the immediate itself, then the formula isn't worthwhile.
3582 for (SmallVectorImpl<const SCEV *>::const_iterator
3583 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
3585 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
3586 if ((C->getValue()->getValue() + NewF.AM.BaseOffs).abs().slt(
3587 abs64(NewF.AM.BaseOffs)) &&
3588 (C->getValue()->getValue() +
3589 NewF.AM.BaseOffs).countTrailingZeros() >=
3590 CountTrailingZeros_64(NewF.AM.BaseOffs))
3594 (void)InsertFormula(LU, LUIdx, NewF);
3603 /// GenerateAllReuseFormulae - Generate formulae for each use.
3605 LSRInstance::GenerateAllReuseFormulae() {
3606 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
3607 // queries are more precise.
3608 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3609 LSRUse &LU = Uses[LUIdx];
3610 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3611 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
3612 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3613 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
3615 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3616 LSRUse &LU = Uses[LUIdx];
3617 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3618 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
3619 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3620 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
3621 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3622 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
3623 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3624 GenerateScales(LU, LUIdx, LU.Formulae[i]);
3626 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3627 LSRUse &LU = Uses[LUIdx];
3628 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3629 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
3632 GenerateCrossUseConstantOffsets();
3634 DEBUG(dbgs() << "\n"
3635 "After generating reuse formulae:\n";
3636 print_uses(dbgs()));
3639 /// If there are multiple formulae with the same set of registers used
3640 /// by other uses, pick the best one and delete the others.
3641 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
3642 DenseSet<const SCEV *> VisitedRegs;
3643 SmallPtrSet<const SCEV *, 16> Regs;
3644 SmallPtrSet<const SCEV *, 16> LoserRegs;
3646 bool ChangedFormulae = false;
3649 // Collect the best formula for each unique set of shared registers. This
3650 // is reset for each use.
3651 typedef DenseMap<SmallVector<const SCEV *, 2>, size_t, UniquifierDenseMapInfo>
3653 BestFormulaeTy BestFormulae;
3655 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3656 LSRUse &LU = Uses[LUIdx];
3657 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
3660 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
3661 FIdx != NumForms; ++FIdx) {
3662 Formula &F = LU.Formulae[FIdx];
3664 // Some formulas are instant losers. For example, they may depend on
3665 // nonexistent AddRecs from other loops. These need to be filtered
3666 // immediately, otherwise heuristics could choose them over others leading
3667 // to an unsatisfactory solution. Passing LoserRegs into RateFormula here
3668 // avoids the need to recompute this information across formulae using the
3669 // same bad AddRec. Passing LoserRegs is also essential unless we remove
3670 // the corresponding bad register from the Regs set.
3673 CostF.RateFormula(F, Regs, VisitedRegs, L, LU.Offsets, SE, DT,
3675 if (CostF.isLoser()) {
3676 // During initial formula generation, undesirable formulae are generated
3677 // by uses within other loops that have some non-trivial address mode or
3678 // use the postinc form of the IV. LSR needs to provide these formulae
3679 // as the basis of rediscovering the desired formula that uses an AddRec
3680 // corresponding to the existing phi. Once all formulae have been
3681 // generated, these initial losers may be pruned.
3682 DEBUG(dbgs() << " Filtering loser "; F.print(dbgs());
3686 SmallVector<const SCEV *, 2> Key;
3687 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
3688 JE = F.BaseRegs.end(); J != JE; ++J) {
3689 const SCEV *Reg = *J;
3690 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
3694 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
3695 Key.push_back(F.ScaledReg);
3696 // Unstable sort by host order ok, because this is only used for
3698 std::sort(Key.begin(), Key.end());
3700 std::pair<BestFormulaeTy::const_iterator, bool> P =
3701 BestFormulae.insert(std::make_pair(Key, FIdx));
3705 Formula &Best = LU.Formulae[P.first->second];
3709 CostBest.RateFormula(Best, Regs, VisitedRegs, L, LU.Offsets, SE, DT);
3710 if (CostF < CostBest)
3712 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
3714 " in favor of formula "; Best.print(dbgs());
3718 ChangedFormulae = true;
3720 LU.DeleteFormula(F);
3726 // Now that we've filtered out some formulae, recompute the Regs set.
3728 LU.RecomputeRegs(LUIdx, RegUses);
3730 // Reset this to prepare for the next use.
3731 BestFormulae.clear();
3734 DEBUG(if (ChangedFormulae) {
3736 "After filtering out undesirable candidates:\n";
3741 // This is a rough guess that seems to work fairly well.
3742 static const size_t ComplexityLimit = UINT16_MAX;
3744 /// EstimateSearchSpaceComplexity - Estimate the worst-case number of
3745 /// solutions the solver might have to consider. It almost never considers
3746 /// this many solutions because it prune the search space, but the pruning
3747 /// isn't always sufficient.
3748 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
3750 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3751 E = Uses.end(); I != E; ++I) {
3752 size_t FSize = I->Formulae.size();
3753 if (FSize >= ComplexityLimit) {
3754 Power = ComplexityLimit;
3758 if (Power >= ComplexityLimit)
3764 /// NarrowSearchSpaceByDetectingSupersets - When one formula uses a superset
3765 /// of the registers of another formula, it won't help reduce register
3766 /// pressure (though it may not necessarily hurt register pressure); remove
3767 /// it to simplify the system.
3768 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
3769 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3770 DEBUG(dbgs() << "The search space is too complex.\n");
3772 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
3773 "which use a superset of registers used by other "
3776 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3777 LSRUse &LU = Uses[LUIdx];
3779 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3780 Formula &F = LU.Formulae[i];
3781 // Look for a formula with a constant or GV in a register. If the use
3782 // also has a formula with that same value in an immediate field,
3783 // delete the one that uses a register.
3784 for (SmallVectorImpl<const SCEV *>::const_iterator
3785 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
3786 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
3788 NewF.AM.BaseOffs += C->getValue()->getSExtValue();
3789 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
3790 (I - F.BaseRegs.begin()));
3791 if (LU.HasFormulaWithSameRegs(NewF)) {
3792 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3793 LU.DeleteFormula(F);
3799 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
3800 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
3803 NewF.AM.BaseGV = GV;
3804 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
3805 (I - F.BaseRegs.begin()));
3806 if (LU.HasFormulaWithSameRegs(NewF)) {
3807 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3809 LU.DeleteFormula(F);
3820 LU.RecomputeRegs(LUIdx, RegUses);
3823 DEBUG(dbgs() << "After pre-selection:\n";
3824 print_uses(dbgs()));
3828 /// NarrowSearchSpaceByCollapsingUnrolledCode - When there are many registers
3829 /// for expressions like A, A+1, A+2, etc., allocate a single register for
3831 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
3832 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3833 DEBUG(dbgs() << "The search space is too complex.\n");
3835 DEBUG(dbgs() << "Narrowing the search space by assuming that uses "
3836 "separated by a constant offset will use the same "
3839 // This is especially useful for unrolled loops.
3841 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3842 LSRUse &LU = Uses[LUIdx];
3843 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3844 E = LU.Formulae.end(); I != E; ++I) {
3845 const Formula &F = *I;
3846 if (F.AM.BaseOffs != 0 && F.AM.Scale == 0) {
3847 if (LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU)) {
3848 if (reconcileNewOffset(*LUThatHas, F.AM.BaseOffs,
3849 /*HasBaseReg=*/false,
3850 LU.Kind, LU.AccessTy)) {
3851 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs());
3854 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
3856 // Update the relocs to reference the new use.
3857 for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(),
3858 E = Fixups.end(); I != E; ++I) {
3859 LSRFixup &Fixup = *I;
3860 if (Fixup.LUIdx == LUIdx) {
3861 Fixup.LUIdx = LUThatHas - &Uses.front();
3862 Fixup.Offset += F.AM.BaseOffs;
3863 // Add the new offset to LUThatHas' offset list.
3864 if (LUThatHas->Offsets.back() != Fixup.Offset) {
3865 LUThatHas->Offsets.push_back(Fixup.Offset);
3866 if (Fixup.Offset > LUThatHas->MaxOffset)
3867 LUThatHas->MaxOffset = Fixup.Offset;
3868 if (Fixup.Offset < LUThatHas->MinOffset)
3869 LUThatHas->MinOffset = Fixup.Offset;
3871 DEBUG(dbgs() << "New fixup has offset "
3872 << Fixup.Offset << '\n');
3874 if (Fixup.LUIdx == NumUses-1)
3875 Fixup.LUIdx = LUIdx;
3878 // Delete formulae from the new use which are no longer legal.
3880 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
3881 Formula &F = LUThatHas->Formulae[i];
3882 if (!isLegalUse(TTI, LUThatHas->MinOffset, LUThatHas->MaxOffset,
3883 LUThatHas->Kind, LUThatHas->AccessTy, F)) {
3884 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3886 LUThatHas->DeleteFormula(F);
3893 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
3895 // Delete the old use.
3896 DeleteUse(LU, LUIdx);
3906 DEBUG(dbgs() << "After pre-selection:\n";
3907 print_uses(dbgs()));
3911 /// NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters - Call
3912 /// FilterOutUndesirableDedicatedRegisters again, if necessary, now that
3913 /// we've done more filtering, as it may be able to find more formulae to
3915 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
3916 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3917 DEBUG(dbgs() << "The search space is too complex.\n");
3919 DEBUG(dbgs() << "Narrowing the search space by re-filtering out "
3920 "undesirable dedicated registers.\n");
3922 FilterOutUndesirableDedicatedRegisters();
3924 DEBUG(dbgs() << "After pre-selection:\n";
3925 print_uses(dbgs()));
3929 /// NarrowSearchSpaceByPickingWinnerRegs - Pick a register which seems likely
3930 /// to be profitable, and then in any use which has any reference to that
3931 /// register, delete all formulae which do not reference that register.
3932 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
3933 // With all other options exhausted, loop until the system is simple
3934 // enough to handle.
3935 SmallPtrSet<const SCEV *, 4> Taken;
3936 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3937 // Ok, we have too many of formulae on our hands to conveniently handle.
3938 // Use a rough heuristic to thin out the list.
3939 DEBUG(dbgs() << "The search space is too complex.\n");
3941 // Pick the register which is used by the most LSRUses, which is likely
3942 // to be a good reuse register candidate.
3943 const SCEV *Best = 0;
3944 unsigned BestNum = 0;
3945 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
3947 const SCEV *Reg = *I;
3948 if (Taken.count(Reg))
3953 unsigned Count = RegUses.getUsedByIndices(Reg).count();
3954 if (Count > BestNum) {
3961 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
3962 << " will yield profitable reuse.\n");
3965 // In any use with formulae which references this register, delete formulae
3966 // which don't reference it.
3967 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3968 LSRUse &LU = Uses[LUIdx];
3969 if (!LU.Regs.count(Best)) continue;
3972 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3973 Formula &F = LU.Formulae[i];
3974 if (!F.referencesReg(Best)) {
3975 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3976 LU.DeleteFormula(F);
3980 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
3986 LU.RecomputeRegs(LUIdx, RegUses);
3989 DEBUG(dbgs() << "After pre-selection:\n";
3990 print_uses(dbgs()));
3994 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
3995 /// formulae to choose from, use some rough heuristics to prune down the number
3996 /// of formulae. This keeps the main solver from taking an extraordinary amount
3997 /// of time in some worst-case scenarios.
3998 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
3999 NarrowSearchSpaceByDetectingSupersets();
4000 NarrowSearchSpaceByCollapsingUnrolledCode();
4001 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
4002 NarrowSearchSpaceByPickingWinnerRegs();
4005 /// SolveRecurse - This is the recursive solver.
4006 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
4008 SmallVectorImpl<const Formula *> &Workspace,
4009 const Cost &CurCost,
4010 const SmallPtrSet<const SCEV *, 16> &CurRegs,
4011 DenseSet<const SCEV *> &VisitedRegs) const {
4014 // - use more aggressive filtering
4015 // - sort the formula so that the most profitable solutions are found first
4016 // - sort the uses too
4018 // - don't compute a cost, and then compare. compare while computing a cost
4020 // - track register sets with SmallBitVector
4022 const LSRUse &LU = Uses[Workspace.size()];
4024 // If this use references any register that's already a part of the
4025 // in-progress solution, consider it a requirement that a formula must
4026 // reference that register in order to be considered. This prunes out
4027 // unprofitable searching.
4028 SmallSetVector<const SCEV *, 4> ReqRegs;
4029 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(),
4030 E = CurRegs.end(); I != E; ++I)
4031 if (LU.Regs.count(*I))
4034 SmallPtrSet<const SCEV *, 16> NewRegs;
4036 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
4037 E = LU.Formulae.end(); I != E; ++I) {
4038 const Formula &F = *I;
4040 // Ignore formulae which do not use any of the required registers.
4041 bool SatisfiedReqReg = true;
4042 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
4043 JE = ReqRegs.end(); J != JE; ++J) {
4044 const SCEV *Reg = *J;
4045 if ((!F.ScaledReg || F.ScaledReg != Reg) &&
4046 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) ==
4048 SatisfiedReqReg = false;
4052 if (!SatisfiedReqReg) {
4053 // If none of the formulae satisfied the required registers, then we could
4054 // clear ReqRegs and try again. Currently, we simply give up in this case.
4058 // Evaluate the cost of the current formula. If it's already worse than
4059 // the current best, prune the search at that point.
4062 NewCost.RateFormula(F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT);
4063 if (NewCost < SolutionCost) {
4064 Workspace.push_back(&F);
4065 if (Workspace.size() != Uses.size()) {
4066 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
4067 NewRegs, VisitedRegs);
4068 if (F.getNumRegs() == 1 && Workspace.size() == 1)
4069 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
4071 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
4072 dbgs() << ".\n Regs:";
4073 for (SmallPtrSet<const SCEV *, 16>::const_iterator
4074 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I)
4075 dbgs() << ' ' << **I;
4078 SolutionCost = NewCost;
4079 Solution = Workspace;
4081 Workspace.pop_back();
4086 /// Solve - Choose one formula from each use. Return the results in the given
4087 /// Solution vector.
4088 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
4089 SmallVector<const Formula *, 8> Workspace;
4091 SolutionCost.Loose();
4093 SmallPtrSet<const SCEV *, 16> CurRegs;
4094 DenseSet<const SCEV *> VisitedRegs;
4095 Workspace.reserve(Uses.size());
4097 // SolveRecurse does all the work.
4098 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
4099 CurRegs, VisitedRegs);
4100 if (Solution.empty()) {
4101 DEBUG(dbgs() << "\nNo Satisfactory Solution\n");
4105 // Ok, we've now made all our decisions.
4106 DEBUG(dbgs() << "\n"
4107 "The chosen solution requires "; SolutionCost.print(dbgs());
4109 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
4111 Uses[i].print(dbgs());
4114 Solution[i]->print(dbgs());
4118 assert(Solution.size() == Uses.size() && "Malformed solution!");
4121 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
4122 /// the dominator tree far as we can go while still being dominated by the
4123 /// input positions. This helps canonicalize the insert position, which
4124 /// encourages sharing.
4125 BasicBlock::iterator
4126 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
4127 const SmallVectorImpl<Instruction *> &Inputs)
4130 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
4131 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
4134 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
4135 if (!Rung) return IP;
4136 Rung = Rung->getIDom();
4137 if (!Rung) return IP;
4138 IDom = Rung->getBlock();
4140 // Don't climb into a loop though.
4141 const Loop *IDomLoop = LI.getLoopFor(IDom);
4142 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
4143 if (IDomDepth <= IPLoopDepth &&
4144 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
4148 bool AllDominate = true;
4149 Instruction *BetterPos = 0;
4150 Instruction *Tentative = IDom->getTerminator();
4151 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
4152 E = Inputs.end(); I != E; ++I) {
4153 Instruction *Inst = *I;
4154 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
4155 AllDominate = false;
4158 // Attempt to find an insert position in the middle of the block,
4159 // instead of at the end, so that it can be used for other expansions.
4160 if (IDom == Inst->getParent() &&
4161 (!BetterPos || !DT.dominates(Inst, BetterPos)))
4162 BetterPos = llvm::next(BasicBlock::iterator(Inst));
4175 /// AdjustInsertPositionForExpand - Determine an input position which will be
4176 /// dominated by the operands and which will dominate the result.
4177 BasicBlock::iterator
4178 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator LowestIP,
4181 SCEVExpander &Rewriter) const {
4182 // Collect some instructions which must be dominated by the
4183 // expanding replacement. These must be dominated by any operands that
4184 // will be required in the expansion.
4185 SmallVector<Instruction *, 4> Inputs;
4186 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
4187 Inputs.push_back(I);
4188 if (LU.Kind == LSRUse::ICmpZero)
4189 if (Instruction *I =
4190 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
4191 Inputs.push_back(I);
4192 if (LF.PostIncLoops.count(L)) {
4193 if (LF.isUseFullyOutsideLoop(L))
4194 Inputs.push_back(L->getLoopLatch()->getTerminator());
4196 Inputs.push_back(IVIncInsertPos);
4198 // The expansion must also be dominated by the increment positions of any
4199 // loops it for which it is using post-inc mode.
4200 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
4201 E = LF.PostIncLoops.end(); I != E; ++I) {
4202 const Loop *PIL = *I;
4203 if (PIL == L) continue;
4205 // Be dominated by the loop exit.
4206 SmallVector<BasicBlock *, 4> ExitingBlocks;
4207 PIL->getExitingBlocks(ExitingBlocks);
4208 if (!ExitingBlocks.empty()) {
4209 BasicBlock *BB = ExitingBlocks[0];
4210 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
4211 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
4212 Inputs.push_back(BB->getTerminator());
4216 assert(!isa<PHINode>(LowestIP) && !isa<LandingPadInst>(LowestIP)
4217 && !isa<DbgInfoIntrinsic>(LowestIP) &&
4218 "Insertion point must be a normal instruction");
4220 // Then, climb up the immediate dominator tree as far as we can go while
4221 // still being dominated by the input positions.
4222 BasicBlock::iterator IP = HoistInsertPosition(LowestIP, Inputs);
4224 // Don't insert instructions before PHI nodes.
4225 while (isa<PHINode>(IP)) ++IP;
4227 // Ignore landingpad instructions.
4228 while (isa<LandingPadInst>(IP)) ++IP;
4230 // Ignore debug intrinsics.
4231 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
4233 // Set IP below instructions recently inserted by SCEVExpander. This keeps the
4234 // IP consistent across expansions and allows the previously inserted
4235 // instructions to be reused by subsequent expansion.
4236 while (Rewriter.isInsertedInstruction(IP) && IP != LowestIP) ++IP;
4241 /// Expand - Emit instructions for the leading candidate expression for this
4242 /// LSRUse (this is called "expanding").
4243 Value *LSRInstance::Expand(const LSRFixup &LF,
4245 BasicBlock::iterator IP,
4246 SCEVExpander &Rewriter,
4247 SmallVectorImpl<WeakVH> &DeadInsts) const {
4248 const LSRUse &LU = Uses[LF.LUIdx];
4250 // Determine an input position which will be dominated by the operands and
4251 // which will dominate the result.
4252 IP = AdjustInsertPositionForExpand(IP, LF, LU, Rewriter);
4254 // Inform the Rewriter if we have a post-increment use, so that it can
4255 // perform an advantageous expansion.
4256 Rewriter.setPostInc(LF.PostIncLoops);
4258 // This is the type that the user actually needs.
4259 Type *OpTy = LF.OperandValToReplace->getType();
4260 // This will be the type that we'll initially expand to.
4261 Type *Ty = F.getType();
4263 // No type known; just expand directly to the ultimate type.
4265 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
4266 // Expand directly to the ultimate type if it's the right size.
4268 // This is the type to do integer arithmetic in.
4269 Type *IntTy = SE.getEffectiveSCEVType(Ty);
4271 // Build up a list of operands to add together to form the full base.
4272 SmallVector<const SCEV *, 8> Ops;
4274 // Expand the BaseRegs portion.
4275 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
4276 E = F.BaseRegs.end(); I != E; ++I) {
4277 const SCEV *Reg = *I;
4278 assert(!Reg->isZero() && "Zero allocated in a base register!");
4280 // If we're expanding for a post-inc user, make the post-inc adjustment.
4281 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4282 Reg = TransformForPostIncUse(Denormalize, Reg,
4283 LF.UserInst, LF.OperandValToReplace,
4286 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP)));
4289 // Expand the ScaledReg portion.
4290 Value *ICmpScaledV = 0;
4291 if (F.AM.Scale != 0) {
4292 const SCEV *ScaledS = F.ScaledReg;
4294 // If we're expanding for a post-inc user, make the post-inc adjustment.
4295 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4296 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
4297 LF.UserInst, LF.OperandValToReplace,
4300 if (LU.Kind == LSRUse::ICmpZero) {
4301 // An interesting way of "folding" with an icmp is to use a negated
4302 // scale, which we'll implement by inserting it into the other operand
4304 assert(F.AM.Scale == -1 &&
4305 "The only scale supported by ICmpZero uses is -1!");
4306 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP);
4308 // Otherwise just expand the scaled register and an explicit scale,
4309 // which is expected to be matched as part of the address.
4311 // Flush the operand list to suppress SCEVExpander hoisting address modes.
4312 if (!Ops.empty() && LU.Kind == LSRUse::Address) {
4313 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4315 Ops.push_back(SE.getUnknown(FullV));
4317 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP));
4318 ScaledS = SE.getMulExpr(ScaledS,
4319 SE.getConstant(ScaledS->getType(), F.AM.Scale));
4320 Ops.push_back(ScaledS);
4324 // Expand the GV portion.
4326 // Flush the operand list to suppress SCEVExpander hoisting.
4328 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4330 Ops.push_back(SE.getUnknown(FullV));
4332 Ops.push_back(SE.getUnknown(F.AM.BaseGV));
4335 // Flush the operand list to suppress SCEVExpander hoisting of both folded and
4336 // unfolded offsets. LSR assumes they both live next to their uses.
4338 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4340 Ops.push_back(SE.getUnknown(FullV));
4343 // Expand the immediate portion.
4344 int64_t Offset = (uint64_t)F.AM.BaseOffs + LF.Offset;
4346 if (LU.Kind == LSRUse::ICmpZero) {
4347 // The other interesting way of "folding" with an ICmpZero is to use a
4348 // negated immediate.
4350 ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset);
4352 Ops.push_back(SE.getUnknown(ICmpScaledV));
4353 ICmpScaledV = ConstantInt::get(IntTy, Offset);
4356 // Just add the immediate values. These again are expected to be matched
4357 // as part of the address.
4358 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
4362 // Expand the unfolded offset portion.
4363 int64_t UnfoldedOffset = F.UnfoldedOffset;
4364 if (UnfoldedOffset != 0) {
4365 // Just add the immediate values.
4366 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy,
4370 // Emit instructions summing all the operands.
4371 const SCEV *FullS = Ops.empty() ?
4372 SE.getConstant(IntTy, 0) :
4374 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
4376 // We're done expanding now, so reset the rewriter.
4377 Rewriter.clearPostInc();
4379 // An ICmpZero Formula represents an ICmp which we're handling as a
4380 // comparison against zero. Now that we've expanded an expression for that
4381 // form, update the ICmp's other operand.
4382 if (LU.Kind == LSRUse::ICmpZero) {
4383 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
4384 DeadInsts.push_back(CI->getOperand(1));
4385 assert(!F.AM.BaseGV && "ICmp does not support folding a global value and "
4386 "a scale at the same time!");
4387 if (F.AM.Scale == -1) {
4388 if (ICmpScaledV->getType() != OpTy) {
4390 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
4392 ICmpScaledV, OpTy, "tmp", CI);
4395 CI->setOperand(1, ICmpScaledV);
4397 assert(F.AM.Scale == 0 &&
4398 "ICmp does not support folding a global value and "
4399 "a scale at the same time!");
4400 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
4402 if (C->getType() != OpTy)
4403 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
4407 CI->setOperand(1, C);
4414 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use
4415 /// of their operands effectively happens in their predecessor blocks, so the
4416 /// expression may need to be expanded in multiple places.
4417 void LSRInstance::RewriteForPHI(PHINode *PN,
4420 SCEVExpander &Rewriter,
4421 SmallVectorImpl<WeakVH> &DeadInsts,
4423 DenseMap<BasicBlock *, Value *> Inserted;
4424 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
4425 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
4426 BasicBlock *BB = PN->getIncomingBlock(i);
4428 // If this is a critical edge, split the edge so that we do not insert
4429 // the code on all predecessor/successor paths. We do this unless this
4430 // is the canonical backedge for this loop, which complicates post-inc
4432 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
4433 !isa<IndirectBrInst>(BB->getTerminator())) {
4434 BasicBlock *Parent = PN->getParent();
4435 Loop *PNLoop = LI.getLoopFor(Parent);
4436 if (!PNLoop || Parent != PNLoop->getHeader()) {
4437 // Split the critical edge.
4438 BasicBlock *NewBB = 0;
4439 if (!Parent->isLandingPad()) {
4440 NewBB = SplitCriticalEdge(BB, Parent, P,
4441 /*MergeIdenticalEdges=*/true,
4442 /*DontDeleteUselessPhis=*/true);
4444 SmallVector<BasicBlock*, 2> NewBBs;
4445 SplitLandingPadPredecessors(Parent, BB, "", "", P, NewBBs);
4448 // If NewBB==NULL, then SplitCriticalEdge refused to split because all
4449 // phi predecessors are identical. The simple thing to do is skip
4450 // splitting in this case rather than complicate the API.
4452 // If PN is outside of the loop and BB is in the loop, we want to
4453 // move the block to be immediately before the PHI block, not
4454 // immediately after BB.
4455 if (L->contains(BB) && !L->contains(PN))
4456 NewBB->moveBefore(PN->getParent());
4458 // Splitting the edge can reduce the number of PHI entries we have.
4459 e = PN->getNumIncomingValues();
4461 i = PN->getBasicBlockIndex(BB);
4466 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
4467 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0)));
4469 PN->setIncomingValue(i, Pair.first->second);
4471 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts);
4473 // If this is reuse-by-noop-cast, insert the noop cast.
4474 Type *OpTy = LF.OperandValToReplace->getType();
4475 if (FullV->getType() != OpTy)
4477 CastInst::Create(CastInst::getCastOpcode(FullV, false,
4479 FullV, LF.OperandValToReplace->getType(),
4480 "tmp", BB->getTerminator());
4482 PN->setIncomingValue(i, FullV);
4483 Pair.first->second = FullV;
4488 /// Rewrite - Emit instructions for the leading candidate expression for this
4489 /// LSRUse (this is called "expanding"), and update the UserInst to reference
4490 /// the newly expanded value.
4491 void LSRInstance::Rewrite(const LSRFixup &LF,
4493 SCEVExpander &Rewriter,
4494 SmallVectorImpl<WeakVH> &DeadInsts,
4496 // First, find an insertion point that dominates UserInst. For PHI nodes,
4497 // find the nearest block which dominates all the relevant uses.
4498 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
4499 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P);
4501 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts);
4503 // If this is reuse-by-noop-cast, insert the noop cast.
4504 Type *OpTy = LF.OperandValToReplace->getType();
4505 if (FullV->getType() != OpTy) {
4507 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
4508 FullV, OpTy, "tmp", LF.UserInst);
4512 // Update the user. ICmpZero is handled specially here (for now) because
4513 // Expand may have updated one of the operands of the icmp already, and
4514 // its new value may happen to be equal to LF.OperandValToReplace, in
4515 // which case doing replaceUsesOfWith leads to replacing both operands
4516 // with the same value. TODO: Reorganize this.
4517 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
4518 LF.UserInst->setOperand(0, FullV);
4520 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
4523 DeadInsts.push_back(LF.OperandValToReplace);
4526 /// ImplementSolution - Rewrite all the fixup locations with new values,
4527 /// following the chosen solution.
4529 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
4531 // Keep track of instructions we may have made dead, so that
4532 // we can remove them after we are done working.
4533 SmallVector<WeakVH, 16> DeadInsts;
4535 SCEVExpander Rewriter(SE, "lsr");
4537 Rewriter.setDebugType(DEBUG_TYPE);
4539 Rewriter.disableCanonicalMode();
4540 Rewriter.enableLSRMode();
4541 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
4543 // Mark phi nodes that terminate chains so the expander tries to reuse them.
4544 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(),
4545 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) {
4546 if (PHINode *PN = dyn_cast<PHINode>(ChainI->tailUserInst()))
4547 Rewriter.setChainedPhi(PN);
4550 // Expand the new value definitions and update the users.
4551 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
4552 E = Fixups.end(); I != E; ++I) {
4553 const LSRFixup &Fixup = *I;
4555 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P);
4560 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(),
4561 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) {
4562 GenerateIVChain(*ChainI, Rewriter, DeadInsts);
4565 // Clean up after ourselves. This must be done before deleting any
4569 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
4572 LSRInstance::LSRInstance(Loop *L, Pass *P)
4573 : IU(P->getAnalysis<IVUsers>()), SE(P->getAnalysis<ScalarEvolution>()),
4574 DT(P->getAnalysis<DominatorTree>()), LI(P->getAnalysis<LoopInfo>()),
4575 TTI(P->getAnalysis<TargetTransformInfo>()), L(L), Changed(false),
4577 // If LoopSimplify form is not available, stay out of trouble.
4578 if (!L->isLoopSimplifyForm())
4581 // If there's no interesting work to be done, bail early.
4582 if (IU.empty()) return;
4584 // If there's too much analysis to be done, bail early. We won't be able to
4585 // model the problem anyway.
4586 unsigned NumUsers = 0;
4587 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
4588 if (++NumUsers > MaxIVUsers) {
4589 DEBUG(dbgs() << "LSR skipping loop, too many IV Users in " << *L
4596 // All dominating loops must have preheaders, or SCEVExpander may not be able
4597 // to materialize an AddRecExpr whose Start is an outer AddRecExpr.
4599 // IVUsers analysis should only create users that are dominated by simple loop
4600 // headers. Since this loop should dominate all of its users, its user list
4601 // should be empty if this loop itself is not within a simple loop nest.
4602 for (DomTreeNode *Rung = DT.getNode(L->getLoopPreheader());
4603 Rung; Rung = Rung->getIDom()) {
4604 BasicBlock *BB = Rung->getBlock();
4605 const Loop *DomLoop = LI.getLoopFor(BB);
4606 if (DomLoop && DomLoop->getHeader() == BB) {
4607 assert(DomLoop->getLoopPreheader() && "LSR needs a simplified loop nest");
4612 DEBUG(dbgs() << "\nLSR on loop ";
4613 WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false);
4616 // First, perform some low-level loop optimizations.
4618 OptimizeLoopTermCond();
4620 // If loop preparation eliminates all interesting IV users, bail.
4621 if (IU.empty()) return;
4623 // Skip nested loops until we can model them better with formulae.
4625 DEBUG(dbgs() << "LSR skipping outer loop " << *L << "\n");
4629 // Start collecting data and preparing for the solver.
4631 CollectInterestingTypesAndFactors();
4632 CollectFixupsAndInitialFormulae();
4633 CollectLoopInvariantFixupsAndFormulae();
4635 assert(!Uses.empty() && "IVUsers reported at least one use");
4636 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
4637 print_uses(dbgs()));
4639 // Now use the reuse data to generate a bunch of interesting ways
4640 // to formulate the values needed for the uses.
4641 GenerateAllReuseFormulae();
4643 FilterOutUndesirableDedicatedRegisters();
4644 NarrowSearchSpaceUsingHeuristics();
4646 SmallVector<const Formula *, 8> Solution;
4649 // Release memory that is no longer needed.
4654 if (Solution.empty())
4658 // Formulae should be legal.
4659 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(), E = Uses.end();
4661 const LSRUse &LU = *I;
4662 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
4663 JE = LU.Formulae.end();
4665 assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
4666 *J) && "Illegal formula generated!");
4670 // Now that we've decided what we want, make it so.
4671 ImplementSolution(Solution, P);
4674 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
4675 if (Factors.empty() && Types.empty()) return;
4677 OS << "LSR has identified the following interesting factors and types: ";
4680 for (SmallSetVector<int64_t, 8>::const_iterator
4681 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
4682 if (!First) OS << ", ";
4687 for (SmallSetVector<Type *, 4>::const_iterator
4688 I = Types.begin(), E = Types.end(); I != E; ++I) {
4689 if (!First) OS << ", ";
4691 OS << '(' << **I << ')';
4696 void LSRInstance::print_fixups(raw_ostream &OS) const {
4697 OS << "LSR is examining the following fixup sites:\n";
4698 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
4699 E = Fixups.end(); I != E; ++I) {
4706 void LSRInstance::print_uses(raw_ostream &OS) const {
4707 OS << "LSR is examining the following uses:\n";
4708 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
4709 E = Uses.end(); I != E; ++I) {
4710 const LSRUse &LU = *I;
4714 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
4715 JE = LU.Formulae.end(); J != JE; ++J) {
4723 void LSRInstance::print(raw_ostream &OS) const {
4724 print_factors_and_types(OS);
4729 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
4730 void LSRInstance::dump() const {
4731 print(errs()); errs() << '\n';
4737 class LoopStrengthReduce : public LoopPass {
4739 static char ID; // Pass ID, replacement for typeid
4740 LoopStrengthReduce();
4743 bool runOnLoop(Loop *L, LPPassManager &LPM);
4744 void getAnalysisUsage(AnalysisUsage &AU) const;
4749 char LoopStrengthReduce::ID = 0;
4750 INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce",
4751 "Loop Strength Reduction", false, false)
4752 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
4753 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
4754 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
4755 INITIALIZE_PASS_DEPENDENCY(IVUsers)
4756 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
4757 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
4758 INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce",
4759 "Loop Strength Reduction", false, false)
4762 Pass *llvm::createLoopStrengthReducePass() {
4763 return new LoopStrengthReduce();
4766 LoopStrengthReduce::LoopStrengthReduce() : LoopPass(ID) {
4767 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry());
4770 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
4771 // We split critical edges, so we change the CFG. However, we do update
4772 // many analyses if they are around.
4773 AU.addPreservedID(LoopSimplifyID);
4775 AU.addRequired<LoopInfo>();
4776 AU.addPreserved<LoopInfo>();
4777 AU.addRequiredID(LoopSimplifyID);
4778 AU.addRequired<DominatorTree>();
4779 AU.addPreserved<DominatorTree>();
4780 AU.addRequired<ScalarEvolution>();
4781 AU.addPreserved<ScalarEvolution>();
4782 // Requiring LoopSimplify a second time here prevents IVUsers from running
4783 // twice, since LoopSimplify was invalidated by running ScalarEvolution.
4784 AU.addRequiredID(LoopSimplifyID);
4785 AU.addRequired<IVUsers>();
4786 AU.addPreserved<IVUsers>();
4787 AU.addRequired<TargetTransformInfo>();
4790 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
4791 bool Changed = false;
4793 // Run the main LSR transformation.
4794 Changed |= LSRInstance(L, this).getChanged();
4796 // Remove any extra phis created by processing inner loops.
4797 Changed |= DeleteDeadPHIs(L->getHeader());
4798 if (EnablePhiElim && L->isLoopSimplifyForm()) {
4799 SmallVector<WeakVH, 16> DeadInsts;
4800 SCEVExpander Rewriter(getAnalysis<ScalarEvolution>(), "lsr");
4802 Rewriter.setDebugType(DEBUG_TYPE);
4804 unsigned numFolded =
4805 Rewriter.replaceCongruentIVs(L, &getAnalysis<DominatorTree>(),
4807 &getAnalysis<TargetTransformInfo>());
4810 DeleteTriviallyDeadInstructions(DeadInsts);
4811 DeleteDeadPHIs(L->getHeader());