1 //===- LoopStrengthReduce.cpp - Strength Reduce IVs in Loops --------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This transformation analyzes and transforms the induction variables (and
11 // computations derived from them) into forms suitable for efficient execution
14 // This pass performs a strength reduction on array references inside loops that
15 // have as one or more of their components the loop induction variable, it
16 // rewrites expressions to take advantage of scaled-index addressing modes
17 // available on the target, and it performs a variety of other optimizations
18 // related to loop induction variables.
20 // Terminology note: this code has a lot of handling for "post-increment" or
21 // "post-inc" users. This is not talking about post-increment addressing modes;
22 // it is instead talking about code like this:
24 // %i = phi [ 0, %entry ], [ %i.next, %latch ]
26 // %i.next = add %i, 1
27 // %c = icmp eq %i.next, %n
29 // The SCEV for %i is {0,+,1}<%L>. The SCEV for %i.next is {1,+,1}<%L>, however
30 // it's useful to think about these as the same register, with some uses using
31 // the value of the register before the add and some using // it after. In this
32 // example, the icmp is a post-increment user, since it uses %i.next, which is
33 // the value of the induction variable after the increment. The other common
34 // case of post-increment users is users outside the loop.
36 // TODO: More sophistication in the way Formulae are generated and filtered.
38 // TODO: Handle multiple loops at a time.
40 // TODO: Should TargetLowering::AddrMode::BaseGV be changed to a ConstantExpr
41 // instead of a GlobalValue?
43 // TODO: When truncation is free, truncate ICmp users' operands to make it a
44 // smaller encoding (on x86 at least).
46 // TODO: When a negated register is used by an add (such as in a list of
47 // multiple base registers, or as the increment expression in an addrec),
48 // we may not actually need both reg and (-1 * reg) in registers; the
49 // negation can be implemented by using a sub instead of an add. The
50 // lack of support for taking this into consideration when making
51 // register pressure decisions is partly worked around by the "Special"
54 //===----------------------------------------------------------------------===//
56 #define DEBUG_TYPE "loop-reduce"
57 #include "llvm/Transforms/Scalar.h"
58 #include "llvm/Constants.h"
59 #include "llvm/Instructions.h"
60 #include "llvm/IntrinsicInst.h"
61 #include "llvm/DerivedTypes.h"
62 #include "llvm/Analysis/IVUsers.h"
63 #include "llvm/Analysis/Dominators.h"
64 #include "llvm/Analysis/LoopPass.h"
65 #include "llvm/Analysis/ScalarEvolutionExpander.h"
66 #include "llvm/Assembly/Writer.h"
67 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
68 #include "llvm/Transforms/Utils/Local.h"
69 #include "llvm/ADT/SmallBitVector.h"
70 #include "llvm/ADT/SetVector.h"
71 #include "llvm/ADT/DenseSet.h"
72 #include "llvm/Support/Debug.h"
73 #include "llvm/Support/CommandLine.h"
74 #include "llvm/Support/ValueHandle.h"
75 #include "llvm/Support/raw_ostream.h"
76 #include "llvm/Target/TargetLowering.h"
80 /// MaxIVUsers is an arbitrary threshold that provides an early opportunitiy for
81 /// bail out. This threshold is far beyond the number of users that LSR can
82 /// conceivably solve, so it should not affect generated code, but catches the
83 /// worst cases before LSR burns too much compile time and stack space.
84 static const unsigned MaxIVUsers = 200;
86 // Temporary flag to cleanup congruent phis after LSR phi expansion.
87 // It's currently disabled until we can determine whether it's truly useful or
88 // not. The flag should be removed after the v3.0 release.
89 // This is now needed for ivchains.
90 static cl::opt<bool> EnablePhiElim(
91 "enable-lsr-phielim", cl::Hidden, cl::init(true),
92 cl::desc("Enable LSR phi elimination"));
95 // Stress test IV chain generation.
96 static cl::opt<bool> StressIVChain(
97 "stress-ivchain", cl::Hidden, cl::init(false),
98 cl::desc("Stress test LSR IV chains"));
100 static bool StressIVChain = false;
105 /// RegSortData - This class holds data which is used to order reuse candidates.
108 /// UsedByIndices - This represents the set of LSRUse indices which reference
109 /// a particular register.
110 SmallBitVector UsedByIndices;
114 void print(raw_ostream &OS) const;
120 void RegSortData::print(raw_ostream &OS) const {
121 OS << "[NumUses=" << UsedByIndices.count() << ']';
124 void RegSortData::dump() const {
125 print(errs()); errs() << '\n';
130 /// RegUseTracker - Map register candidates to information about how they are
132 class RegUseTracker {
133 typedef DenseMap<const SCEV *, RegSortData> RegUsesTy;
135 RegUsesTy RegUsesMap;
136 SmallVector<const SCEV *, 16> RegSequence;
139 void CountRegister(const SCEV *Reg, size_t LUIdx);
140 void DropRegister(const SCEV *Reg, size_t LUIdx);
141 void SwapAndDropUse(size_t LUIdx, size_t LastLUIdx);
143 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
145 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
149 typedef SmallVectorImpl<const SCEV *>::iterator iterator;
150 typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator;
151 iterator begin() { return RegSequence.begin(); }
152 iterator end() { return RegSequence.end(); }
153 const_iterator begin() const { return RegSequence.begin(); }
154 const_iterator end() const { return RegSequence.end(); }
160 RegUseTracker::CountRegister(const SCEV *Reg, size_t LUIdx) {
161 std::pair<RegUsesTy::iterator, bool> Pair =
162 RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
163 RegSortData &RSD = Pair.first->second;
165 RegSequence.push_back(Reg);
166 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
167 RSD.UsedByIndices.set(LUIdx);
171 RegUseTracker::DropRegister(const SCEV *Reg, size_t LUIdx) {
172 RegUsesTy::iterator It = RegUsesMap.find(Reg);
173 assert(It != RegUsesMap.end());
174 RegSortData &RSD = It->second;
175 assert(RSD.UsedByIndices.size() > LUIdx);
176 RSD.UsedByIndices.reset(LUIdx);
180 RegUseTracker::SwapAndDropUse(size_t LUIdx, size_t LastLUIdx) {
181 assert(LUIdx <= LastLUIdx);
183 // Update RegUses. The data structure is not optimized for this purpose;
184 // we must iterate through it and update each of the bit vectors.
185 for (RegUsesTy::iterator I = RegUsesMap.begin(), E = RegUsesMap.end();
187 SmallBitVector &UsedByIndices = I->second.UsedByIndices;
188 if (LUIdx < UsedByIndices.size())
189 UsedByIndices[LUIdx] =
190 LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : 0;
191 UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx));
196 RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
197 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
198 if (I == RegUsesMap.end())
200 const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
201 int i = UsedByIndices.find_first();
202 if (i == -1) return false;
203 if ((size_t)i != LUIdx) return true;
204 return UsedByIndices.find_next(i) != -1;
207 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
208 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
209 assert(I != RegUsesMap.end() && "Unknown register!");
210 return I->second.UsedByIndices;
213 void RegUseTracker::clear() {
220 /// Formula - This class holds information that describes a formula for
221 /// computing satisfying a use. It may include broken-out immediates and scaled
224 /// AM - This is used to represent complex addressing, as well as other kinds
225 /// of interesting uses.
226 TargetLowering::AddrMode AM;
228 /// BaseRegs - The list of "base" registers for this use. When this is
229 /// non-empty, AM.HasBaseReg should be set to true.
230 SmallVector<const SCEV *, 2> BaseRegs;
232 /// ScaledReg - The 'scaled' register for this use. This should be non-null
233 /// when AM.Scale is not zero.
234 const SCEV *ScaledReg;
236 /// UnfoldedOffset - An additional constant offset which added near the
237 /// use. This requires a temporary register, but the offset itself can
238 /// live in an add immediate field rather than a register.
239 int64_t UnfoldedOffset;
241 Formula() : ScaledReg(0), UnfoldedOffset(0) {}
243 void InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE);
245 unsigned getNumRegs() const;
246 Type *getType() const;
248 void DeleteBaseReg(const SCEV *&S);
250 bool referencesReg(const SCEV *S) const;
251 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
252 const RegUseTracker &RegUses) const;
254 void print(raw_ostream &OS) const;
260 /// DoInitialMatch - Recursion helper for InitialMatch.
261 static void DoInitialMatch(const SCEV *S, Loop *L,
262 SmallVectorImpl<const SCEV *> &Good,
263 SmallVectorImpl<const SCEV *> &Bad,
264 ScalarEvolution &SE) {
265 // Collect expressions which properly dominate the loop header.
266 if (SE.properlyDominates(S, L->getHeader())) {
271 // Look at add operands.
272 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
273 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
275 DoInitialMatch(*I, L, Good, Bad, SE);
279 // Look at addrec operands.
280 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
281 if (!AR->getStart()->isZero()) {
282 DoInitialMatch(AR->getStart(), L, Good, Bad, SE);
283 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
284 AR->getStepRecurrence(SE),
285 // FIXME: AR->getNoWrapFlags()
286 AR->getLoop(), SCEV::FlagAnyWrap),
291 // Handle a multiplication by -1 (negation) if it didn't fold.
292 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
293 if (Mul->getOperand(0)->isAllOnesValue()) {
294 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
295 const SCEV *NewMul = SE.getMulExpr(Ops);
297 SmallVector<const SCEV *, 4> MyGood;
298 SmallVector<const SCEV *, 4> MyBad;
299 DoInitialMatch(NewMul, L, MyGood, MyBad, SE);
300 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
301 SE.getEffectiveSCEVType(NewMul->getType())));
302 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyGood.begin(),
303 E = MyGood.end(); I != E; ++I)
304 Good.push_back(SE.getMulExpr(NegOne, *I));
305 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyBad.begin(),
306 E = MyBad.end(); I != E; ++I)
307 Bad.push_back(SE.getMulExpr(NegOne, *I));
311 // Ok, we can't do anything interesting. Just stuff the whole thing into a
312 // register and hope for the best.
316 /// InitialMatch - Incorporate loop-variant parts of S into this Formula,
317 /// attempting to keep all loop-invariant and loop-computable values in a
318 /// single base register.
319 void Formula::InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) {
320 SmallVector<const SCEV *, 4> Good;
321 SmallVector<const SCEV *, 4> Bad;
322 DoInitialMatch(S, L, Good, Bad, SE);
324 const SCEV *Sum = SE.getAddExpr(Good);
326 BaseRegs.push_back(Sum);
327 AM.HasBaseReg = true;
330 const SCEV *Sum = SE.getAddExpr(Bad);
332 BaseRegs.push_back(Sum);
333 AM.HasBaseReg = true;
337 /// getNumRegs - Return the total number of register operands used by this
338 /// formula. This does not include register uses implied by non-constant
340 unsigned Formula::getNumRegs() const {
341 return !!ScaledReg + BaseRegs.size();
344 /// getType - Return the type of this formula, if it has one, or null
345 /// otherwise. This type is meaningless except for the bit size.
346 Type *Formula::getType() const {
347 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
348 ScaledReg ? ScaledReg->getType() :
349 AM.BaseGV ? AM.BaseGV->getType() :
353 /// DeleteBaseReg - Delete the given base reg from the BaseRegs list.
354 void Formula::DeleteBaseReg(const SCEV *&S) {
355 if (&S != &BaseRegs.back())
356 std::swap(S, BaseRegs.back());
360 /// referencesReg - Test if this formula references the given register.
361 bool Formula::referencesReg(const SCEV *S) const {
362 return S == ScaledReg ||
363 std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end();
366 /// hasRegsUsedByUsesOtherThan - Test whether this formula uses registers
367 /// which are used by uses other than the use with the given index.
368 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
369 const RegUseTracker &RegUses) const {
371 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
373 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
374 E = BaseRegs.end(); I != E; ++I)
375 if (RegUses.isRegUsedByUsesOtherThan(*I, LUIdx))
380 void Formula::print(raw_ostream &OS) const {
383 if (!First) OS << " + "; else First = false;
384 WriteAsOperand(OS, AM.BaseGV, /*PrintType=*/false);
386 if (AM.BaseOffs != 0) {
387 if (!First) OS << " + "; else First = false;
390 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
391 E = BaseRegs.end(); I != E; ++I) {
392 if (!First) OS << " + "; else First = false;
393 OS << "reg(" << **I << ')';
395 if (AM.HasBaseReg && BaseRegs.empty()) {
396 if (!First) OS << " + "; else First = false;
397 OS << "**error: HasBaseReg**";
398 } else if (!AM.HasBaseReg && !BaseRegs.empty()) {
399 if (!First) OS << " + "; else First = false;
400 OS << "**error: !HasBaseReg**";
403 if (!First) OS << " + "; else First = false;
404 OS << AM.Scale << "*reg(";
411 if (UnfoldedOffset != 0) {
412 if (!First) OS << " + "; else First = false;
413 OS << "imm(" << UnfoldedOffset << ')';
417 void Formula::dump() const {
418 print(errs()); errs() << '\n';
421 /// isAddRecSExtable - Return true if the given addrec can be sign-extended
422 /// without changing its value.
423 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
425 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1);
426 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
429 /// isAddSExtable - Return true if the given add can be sign-extended
430 /// without changing its value.
431 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
433 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
434 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
437 /// isMulSExtable - Return true if the given mul can be sign-extended
438 /// without changing its value.
439 static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
441 IntegerType::get(SE.getContext(),
442 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
443 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
446 /// getExactSDiv - Return an expression for LHS /s RHS, if it can be determined
447 /// and if the remainder is known to be zero, or null otherwise. If
448 /// IgnoreSignificantBits is true, expressions like (X * Y) /s Y are simplified
449 /// to Y, ignoring that the multiplication may overflow, which is useful when
450 /// the result will be used in a context where the most significant bits are
452 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
454 bool IgnoreSignificantBits = false) {
455 // Handle the trivial case, which works for any SCEV type.
457 return SE.getConstant(LHS->getType(), 1);
459 // Handle a few RHS special cases.
460 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
462 const APInt &RA = RC->getValue()->getValue();
463 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
465 if (RA.isAllOnesValue())
466 return SE.getMulExpr(LHS, RC);
467 // Handle x /s 1 as x.
472 // Check for a division of a constant by a constant.
473 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
476 const APInt &LA = C->getValue()->getValue();
477 const APInt &RA = RC->getValue()->getValue();
478 if (LA.srem(RA) != 0)
480 return SE.getConstant(LA.sdiv(RA));
483 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
484 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
485 if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) {
486 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
487 IgnoreSignificantBits);
489 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
490 IgnoreSignificantBits);
491 if (!Start) return 0;
492 // FlagNW is independent of the start value, step direction, and is
493 // preserved with smaller magnitude steps.
494 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
495 return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap);
500 // Distribute the sdiv over add operands, if the add doesn't overflow.
501 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
502 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
503 SmallVector<const SCEV *, 8> Ops;
504 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
506 const SCEV *Op = getExactSDiv(*I, RHS, SE,
507 IgnoreSignificantBits);
511 return SE.getAddExpr(Ops);
516 // Check for a multiply operand that we can pull RHS out of.
517 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
518 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
519 SmallVector<const SCEV *, 4> Ops;
521 for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end();
525 if (const SCEV *Q = getExactSDiv(S, RHS, SE,
526 IgnoreSignificantBits)) {
532 return Found ? SE.getMulExpr(Ops) : 0;
537 // Otherwise we don't know.
541 /// ExtractImmediate - If S involves the addition of a constant integer value,
542 /// return that integer value, and mutate S to point to a new SCEV with that
544 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
545 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
546 if (C->getValue()->getValue().getMinSignedBits() <= 64) {
547 S = SE.getConstant(C->getType(), 0);
548 return C->getValue()->getSExtValue();
550 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
551 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
552 int64_t Result = ExtractImmediate(NewOps.front(), SE);
554 S = SE.getAddExpr(NewOps);
556 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
557 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
558 int64_t Result = ExtractImmediate(NewOps.front(), SE);
560 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
561 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
568 /// ExtractSymbol - If S involves the addition of a GlobalValue address,
569 /// return that symbol, and mutate S to point to a new SCEV with that
571 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
572 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
573 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
574 S = SE.getConstant(GV->getType(), 0);
577 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
578 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
579 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
581 S = SE.getAddExpr(NewOps);
583 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
584 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
585 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
587 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
588 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
595 /// isAddressUse - Returns true if the specified instruction is using the
596 /// specified value as an address.
597 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
598 bool isAddress = isa<LoadInst>(Inst);
599 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
600 if (SI->getOperand(1) == OperandVal)
602 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
603 // Addressing modes can also be folded into prefetches and a variety
605 switch (II->getIntrinsicID()) {
607 case Intrinsic::prefetch:
608 case Intrinsic::x86_sse_storeu_ps:
609 case Intrinsic::x86_sse2_storeu_pd:
610 case Intrinsic::x86_sse2_storeu_dq:
611 case Intrinsic::x86_sse2_storel_dq:
612 if (II->getArgOperand(0) == OperandVal)
620 /// getAccessType - Return the type of the memory being accessed.
621 static Type *getAccessType(const Instruction *Inst) {
622 Type *AccessTy = Inst->getType();
623 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
624 AccessTy = SI->getOperand(0)->getType();
625 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
626 // Addressing modes can also be folded into prefetches and a variety
628 switch (II->getIntrinsicID()) {
630 case Intrinsic::x86_sse_storeu_ps:
631 case Intrinsic::x86_sse2_storeu_pd:
632 case Intrinsic::x86_sse2_storeu_dq:
633 case Intrinsic::x86_sse2_storel_dq:
634 AccessTy = II->getArgOperand(0)->getType();
639 // All pointers have the same requirements, so canonicalize them to an
640 // arbitrary pointer type to minimize variation.
641 if (PointerType *PTy = dyn_cast<PointerType>(AccessTy))
642 AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
643 PTy->getAddressSpace());
648 /// isExistingPhi - Return true if this AddRec is already a phi in its loop.
649 static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
650 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
651 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
652 if (SE.isSCEVable(PN->getType()) &&
653 (SE.getEffectiveSCEVType(PN->getType()) ==
654 SE.getEffectiveSCEVType(AR->getType())) &&
655 SE.getSCEV(PN) == AR)
661 /// Check if expanding this expression is likely to incur significant cost. This
662 /// is tricky because SCEV doesn't track which expressions are actually computed
663 /// by the current IR.
665 /// We currently allow expansion of IV increments that involve adds,
666 /// multiplication by constants, and AddRecs from existing phis.
668 /// TODO: Allow UDivExpr if we can find an existing IV increment that is an
669 /// obvious multiple of the UDivExpr.
670 static bool isHighCostExpansion(const SCEV *S,
671 SmallPtrSet<const SCEV*, 8> &Processed,
672 ScalarEvolution &SE) {
673 // Zero/One operand expressions
674 switch (S->getSCEVType()) {
679 return isHighCostExpansion(cast<SCEVTruncateExpr>(S)->getOperand(),
682 return isHighCostExpansion(cast<SCEVZeroExtendExpr>(S)->getOperand(),
685 return isHighCostExpansion(cast<SCEVSignExtendExpr>(S)->getOperand(),
689 if (!Processed.insert(S))
692 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
693 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
695 if (isHighCostExpansion(*I, Processed, SE))
701 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
702 if (Mul->getNumOperands() == 2) {
703 // Multiplication by a constant is ok
704 if (isa<SCEVConstant>(Mul->getOperand(0)))
705 return isHighCostExpansion(Mul->getOperand(1), Processed, SE);
707 // If we have the value of one operand, check if an existing
708 // multiplication already generates this expression.
709 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Mul->getOperand(1))) {
710 Value *UVal = U->getValue();
711 for (Value::use_iterator UI = UVal->use_begin(), UE = UVal->use_end();
713 // If U is a constant, it may be used by a ConstantExpr.
714 Instruction *User = dyn_cast<Instruction>(*UI);
715 if (User && User->getOpcode() == Instruction::Mul
716 && SE.isSCEVable(User->getType())) {
717 return SE.getSCEV(User) == Mul;
724 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
725 if (isExistingPhi(AR, SE))
729 // Fow now, consider any other type of expression (div/mul/min/max) high cost.
733 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
734 /// specified set are trivially dead, delete them and see if this makes any of
735 /// their operands subsequently dead.
737 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) {
738 bool Changed = false;
740 while (!DeadInsts.empty()) {
741 Value *V = DeadInsts.pop_back_val();
742 Instruction *I = dyn_cast_or_null<Instruction>(V);
744 if (I == 0 || !isInstructionTriviallyDead(I))
747 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
748 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
751 DeadInsts.push_back(U);
754 I->eraseFromParent();
763 /// Cost - This class is used to measure and compare candidate formulae.
765 /// TODO: Some of these could be merged. Also, a lexical ordering
766 /// isn't always optimal.
770 unsigned NumBaseAdds;
776 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0),
779 bool operator<(const Cost &Other) const;
784 // Once any of the metrics loses, they must all remain losers.
786 return ((NumRegs | AddRecCost | NumIVMuls | NumBaseAdds
787 | ImmCost | SetupCost) != ~0u)
788 || ((NumRegs & AddRecCost & NumIVMuls & NumBaseAdds
789 & ImmCost & SetupCost) == ~0u);
794 assert(isValid() && "invalid cost");
795 return NumRegs == ~0u;
798 void RateFormula(const Formula &F,
799 SmallPtrSet<const SCEV *, 16> &Regs,
800 const DenseSet<const SCEV *> &VisitedRegs,
802 const SmallVectorImpl<int64_t> &Offsets,
803 ScalarEvolution &SE, DominatorTree &DT,
804 SmallPtrSet<const SCEV *, 16> *LoserRegs = 0);
806 void print(raw_ostream &OS) const;
810 void RateRegister(const SCEV *Reg,
811 SmallPtrSet<const SCEV *, 16> &Regs,
813 ScalarEvolution &SE, DominatorTree &DT);
814 void RatePrimaryRegister(const SCEV *Reg,
815 SmallPtrSet<const SCEV *, 16> &Regs,
817 ScalarEvolution &SE, DominatorTree &DT,
818 SmallPtrSet<const SCEV *, 16> *LoserRegs);
823 /// RateRegister - Tally up interesting quantities from the given register.
824 void Cost::RateRegister(const SCEV *Reg,
825 SmallPtrSet<const SCEV *, 16> &Regs,
827 ScalarEvolution &SE, DominatorTree &DT) {
828 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
829 // If this is an addrec for another loop, don't second-guess its addrec phi
830 // nodes. LSR isn't currently smart enough to reason about more than one
831 // loop at a time. LSR has already run on inner loops, will not run on outer
832 // loops, and cannot be expected to change sibling loops.
833 if (AR->getLoop() != L) {
834 // If the AddRec exists, consider it's register free and leave it alone.
835 if (isExistingPhi(AR, SE))
838 // Otherwise, do not consider this formula at all.
842 AddRecCost += 1; /// TODO: This should be a function of the stride.
844 // Add the step value register, if it needs one.
845 // TODO: The non-affine case isn't precisely modeled here.
846 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
847 if (!Regs.count(AR->getOperand(1))) {
848 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
856 // Rough heuristic; favor registers which don't require extra setup
857 // instructions in the preheader.
858 if (!isa<SCEVUnknown>(Reg) &&
859 !isa<SCEVConstant>(Reg) &&
860 !(isa<SCEVAddRecExpr>(Reg) &&
861 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
862 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
865 NumIVMuls += isa<SCEVMulExpr>(Reg) &&
866 SE.hasComputableLoopEvolution(Reg, L);
869 /// RatePrimaryRegister - Record this register in the set. If we haven't seen it
870 /// before, rate it. Optional LoserRegs provides a way to declare any formula
871 /// that refers to one of those regs an instant loser.
872 void Cost::RatePrimaryRegister(const SCEV *Reg,
873 SmallPtrSet<const SCEV *, 16> &Regs,
875 ScalarEvolution &SE, DominatorTree &DT,
876 SmallPtrSet<const SCEV *, 16> *LoserRegs) {
877 if (LoserRegs && LoserRegs->count(Reg)) {
881 if (Regs.insert(Reg)) {
882 RateRegister(Reg, Regs, L, SE, DT);
884 LoserRegs->insert(Reg);
888 void Cost::RateFormula(const Formula &F,
889 SmallPtrSet<const SCEV *, 16> &Regs,
890 const DenseSet<const SCEV *> &VisitedRegs,
892 const SmallVectorImpl<int64_t> &Offsets,
893 ScalarEvolution &SE, DominatorTree &DT,
894 SmallPtrSet<const SCEV *, 16> *LoserRegs) {
895 // Tally up the registers.
896 if (const SCEV *ScaledReg = F.ScaledReg) {
897 if (VisitedRegs.count(ScaledReg)) {
901 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT, LoserRegs);
905 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
906 E = F.BaseRegs.end(); I != E; ++I) {
907 const SCEV *BaseReg = *I;
908 if (VisitedRegs.count(BaseReg)) {
912 RatePrimaryRegister(BaseReg, Regs, L, SE, DT, LoserRegs);
917 // Determine how many (unfolded) adds we'll need inside the loop.
918 size_t NumBaseParts = F.BaseRegs.size() + (F.UnfoldedOffset != 0);
919 if (NumBaseParts > 1)
920 NumBaseAdds += NumBaseParts - 1;
922 // Tally up the non-zero immediates.
923 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
924 E = Offsets.end(); I != E; ++I) {
925 int64_t Offset = (uint64_t)*I + F.AM.BaseOffs;
927 ImmCost += 64; // Handle symbolic values conservatively.
928 // TODO: This should probably be the pointer size.
929 else if (Offset != 0)
930 ImmCost += APInt(64, Offset, true).getMinSignedBits();
932 assert(isValid() && "invalid cost");
935 /// Loose - Set this cost to a losing value.
945 /// operator< - Choose the lower cost.
946 bool Cost::operator<(const Cost &Other) const {
947 if (NumRegs != Other.NumRegs)
948 return NumRegs < Other.NumRegs;
949 if (AddRecCost != Other.AddRecCost)
950 return AddRecCost < Other.AddRecCost;
951 if (NumIVMuls != Other.NumIVMuls)
952 return NumIVMuls < Other.NumIVMuls;
953 if (NumBaseAdds != Other.NumBaseAdds)
954 return NumBaseAdds < Other.NumBaseAdds;
955 if (ImmCost != Other.ImmCost)
956 return ImmCost < Other.ImmCost;
957 if (SetupCost != Other.SetupCost)
958 return SetupCost < Other.SetupCost;
962 void Cost::print(raw_ostream &OS) const {
963 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
965 OS << ", with addrec cost " << AddRecCost;
967 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
968 if (NumBaseAdds != 0)
969 OS << ", plus " << NumBaseAdds << " base add"
970 << (NumBaseAdds == 1 ? "" : "s");
972 OS << ", plus " << ImmCost << " imm cost";
974 OS << ", plus " << SetupCost << " setup cost";
977 void Cost::dump() const {
978 print(errs()); errs() << '\n';
983 /// LSRFixup - An operand value in an instruction which is to be replaced
984 /// with some equivalent, possibly strength-reduced, replacement.
986 /// UserInst - The instruction which will be updated.
987 Instruction *UserInst;
989 /// OperandValToReplace - The operand of the instruction which will
990 /// be replaced. The operand may be used more than once; every instance
991 /// will be replaced.
992 Value *OperandValToReplace;
994 /// PostIncLoops - If this user is to use the post-incremented value of an
995 /// induction variable, this variable is non-null and holds the loop
996 /// associated with the induction variable.
997 PostIncLoopSet PostIncLoops;
999 /// LUIdx - The index of the LSRUse describing the expression which
1000 /// this fixup needs, minus an offset (below).
1003 /// Offset - A constant offset to be added to the LSRUse expression.
1004 /// This allows multiple fixups to share the same LSRUse with different
1005 /// offsets, for example in an unrolled loop.
1008 bool isUseFullyOutsideLoop(const Loop *L) const;
1012 void print(raw_ostream &OS) const;
1018 LSRFixup::LSRFixup()
1019 : UserInst(0), OperandValToReplace(0), LUIdx(~size_t(0)), Offset(0) {}
1021 /// isUseFullyOutsideLoop - Test whether this fixup always uses its
1022 /// value outside of the given loop.
1023 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
1024 // PHI nodes use their value in their incoming blocks.
1025 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
1026 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1027 if (PN->getIncomingValue(i) == OperandValToReplace &&
1028 L->contains(PN->getIncomingBlock(i)))
1033 return !L->contains(UserInst);
1036 void LSRFixup::print(raw_ostream &OS) const {
1038 // Store is common and interesting enough to be worth special-casing.
1039 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
1041 WriteAsOperand(OS, Store->getOperand(0), /*PrintType=*/false);
1042 } else if (UserInst->getType()->isVoidTy())
1043 OS << UserInst->getOpcodeName();
1045 WriteAsOperand(OS, UserInst, /*PrintType=*/false);
1047 OS << ", OperandValToReplace=";
1048 WriteAsOperand(OS, OperandValToReplace, /*PrintType=*/false);
1050 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(),
1051 E = PostIncLoops.end(); I != E; ++I) {
1052 OS << ", PostIncLoop=";
1053 WriteAsOperand(OS, (*I)->getHeader(), /*PrintType=*/false);
1056 if (LUIdx != ~size_t(0))
1057 OS << ", LUIdx=" << LUIdx;
1060 OS << ", Offset=" << Offset;
1063 void LSRFixup::dump() const {
1064 print(errs()); errs() << '\n';
1069 /// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding
1070 /// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*.
1071 struct UniquifierDenseMapInfo {
1072 static SmallVector<const SCEV *, 2> getEmptyKey() {
1073 SmallVector<const SCEV *, 2> V;
1074 V.push_back(reinterpret_cast<const SCEV *>(-1));
1078 static SmallVector<const SCEV *, 2> getTombstoneKey() {
1079 SmallVector<const SCEV *, 2> V;
1080 V.push_back(reinterpret_cast<const SCEV *>(-2));
1084 static unsigned getHashValue(const SmallVector<const SCEV *, 2> &V) {
1085 unsigned Result = 0;
1086 for (SmallVectorImpl<const SCEV *>::const_iterator I = V.begin(),
1087 E = V.end(); I != E; ++I)
1088 Result ^= DenseMapInfo<const SCEV *>::getHashValue(*I);
1092 static bool isEqual(const SmallVector<const SCEV *, 2> &LHS,
1093 const SmallVector<const SCEV *, 2> &RHS) {
1098 /// LSRUse - This class holds the state that LSR keeps for each use in
1099 /// IVUsers, as well as uses invented by LSR itself. It includes information
1100 /// about what kinds of things can be folded into the user, information about
1101 /// the user itself, and information about how the use may be satisfied.
1102 /// TODO: Represent multiple users of the same expression in common?
1104 DenseSet<SmallVector<const SCEV *, 2>, UniquifierDenseMapInfo> Uniquifier;
1107 /// KindType - An enum for a kind of use, indicating what types of
1108 /// scaled and immediate operands it might support.
1110 Basic, ///< A normal use, with no folding.
1111 Special, ///< A special case of basic, allowing -1 scales.
1112 Address, ///< An address use; folding according to TargetLowering
1113 ICmpZero ///< An equality icmp with both operands folded into one.
1114 // TODO: Add a generic icmp too?
1120 SmallVector<int64_t, 8> Offsets;
1124 /// AllFixupsOutsideLoop - This records whether all of the fixups using this
1125 /// LSRUse are outside of the loop, in which case some special-case heuristics
1127 bool AllFixupsOutsideLoop;
1129 /// WidestFixupType - This records the widest use type for any fixup using
1130 /// this LSRUse. FindUseWithSimilarFormula can't consider uses with different
1131 /// max fixup widths to be equivalent, because the narrower one may be relying
1132 /// on the implicit truncation to truncate away bogus bits.
1133 Type *WidestFixupType;
1135 /// Formulae - A list of ways to build a value that can satisfy this user.
1136 /// After the list is populated, one of these is selected heuristically and
1137 /// used to formulate a replacement for OperandValToReplace in UserInst.
1138 SmallVector<Formula, 12> Formulae;
1140 /// Regs - The set of register candidates used by all formulae in this LSRUse.
1141 SmallPtrSet<const SCEV *, 4> Regs;
1143 LSRUse(KindType K, Type *T) : Kind(K), AccessTy(T),
1144 MinOffset(INT64_MAX),
1145 MaxOffset(INT64_MIN),
1146 AllFixupsOutsideLoop(true),
1147 WidestFixupType(0) {}
1149 bool HasFormulaWithSameRegs(const Formula &F) const;
1150 bool InsertFormula(const Formula &F);
1151 void DeleteFormula(Formula &F);
1152 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1154 void print(raw_ostream &OS) const;
1160 /// HasFormula - Test whether this use as a formula which has the same
1161 /// registers as the given formula.
1162 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1163 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1164 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1165 // Unstable sort by host order ok, because this is only used for uniquifying.
1166 std::sort(Key.begin(), Key.end());
1167 return Uniquifier.count(Key);
1170 /// InsertFormula - If the given formula has not yet been inserted, add it to
1171 /// the list, and return true. Return false otherwise.
1172 bool LSRUse::InsertFormula(const Formula &F) {
1173 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
1174 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1175 // Unstable sort by host order ok, because this is only used for uniquifying.
1176 std::sort(Key.begin(), Key.end());
1178 if (!Uniquifier.insert(Key).second)
1181 // Using a register to hold the value of 0 is not profitable.
1182 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1183 "Zero allocated in a scaled register!");
1185 for (SmallVectorImpl<const SCEV *>::const_iterator I =
1186 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I)
1187 assert(!(*I)->isZero() && "Zero allocated in a base register!");
1190 // Add the formula to the list.
1191 Formulae.push_back(F);
1193 // Record registers now being used by this use.
1194 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1199 /// DeleteFormula - Remove the given formula from this use's list.
1200 void LSRUse::DeleteFormula(Formula &F) {
1201 if (&F != &Formulae.back())
1202 std::swap(F, Formulae.back());
1203 Formulae.pop_back();
1206 /// RecomputeRegs - Recompute the Regs field, and update RegUses.
1207 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1208 // Now that we've filtered out some formulae, recompute the Regs set.
1209 SmallPtrSet<const SCEV *, 4> OldRegs = Regs;
1211 for (SmallVectorImpl<Formula>::const_iterator I = Formulae.begin(),
1212 E = Formulae.end(); I != E; ++I) {
1213 const Formula &F = *I;
1214 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1215 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1218 // Update the RegTracker.
1219 for (SmallPtrSet<const SCEV *, 4>::iterator I = OldRegs.begin(),
1220 E = OldRegs.end(); I != E; ++I)
1221 if (!Regs.count(*I))
1222 RegUses.DropRegister(*I, LUIdx);
1225 void LSRUse::print(raw_ostream &OS) const {
1226 OS << "LSR Use: Kind=";
1228 case Basic: OS << "Basic"; break;
1229 case Special: OS << "Special"; break;
1230 case ICmpZero: OS << "ICmpZero"; break;
1232 OS << "Address of ";
1233 if (AccessTy->isPointerTy())
1234 OS << "pointer"; // the full pointer type could be really verbose
1239 OS << ", Offsets={";
1240 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
1241 E = Offsets.end(); I != E; ++I) {
1243 if (llvm::next(I) != E)
1248 if (AllFixupsOutsideLoop)
1249 OS << ", all-fixups-outside-loop";
1251 if (WidestFixupType)
1252 OS << ", widest fixup type: " << *WidestFixupType;
1255 void LSRUse::dump() const {
1256 print(errs()); errs() << '\n';
1259 /// isLegalUse - Test whether the use described by AM is "legal", meaning it can
1260 /// be completely folded into the user instruction at isel time. This includes
1261 /// address-mode folding and special icmp tricks.
1262 static bool isLegalUse(const TargetLowering::AddrMode &AM,
1263 LSRUse::KindType Kind, Type *AccessTy,
1264 const TargetLowering *TLI) {
1266 case LSRUse::Address:
1267 // If we have low-level target information, ask the target if it can
1268 // completely fold this address.
1269 if (TLI) return TLI->isLegalAddressingMode(AM, AccessTy);
1271 // Otherwise, just guess that reg+reg addressing is legal.
1272 return !AM.BaseGV && AM.BaseOffs == 0 && AM.Scale <= 1;
1274 case LSRUse::ICmpZero:
1275 // There's not even a target hook for querying whether it would be legal to
1276 // fold a GV into an ICmp.
1280 // ICmp only has two operands; don't allow more than two non-trivial parts.
1281 if (AM.Scale != 0 && AM.HasBaseReg && AM.BaseOffs != 0)
1284 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1285 // putting the scaled register in the other operand of the icmp.
1286 if (AM.Scale != 0 && AM.Scale != -1)
1289 // If we have low-level target information, ask the target if it can fold an
1290 // integer immediate on an icmp.
1291 if (AM.BaseOffs != 0) {
1295 // ICmpZero BaseReg + Offset => ICmp BaseReg, -Offset
1296 // ICmpZero -1*ScaleReg + Offset => ICmp ScaleReg, Offset
1297 // Offs is the ICmp immediate.
1298 int64_t Offs = AM.BaseOffs;
1300 Offs = -(uint64_t)Offs; // The cast does the right thing with INT64_MIN.
1301 return TLI->isLegalICmpImmediate(Offs);
1304 // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg
1308 // Only handle single-register values.
1309 return !AM.BaseGV && AM.Scale == 0 && AM.BaseOffs == 0;
1311 case LSRUse::Special:
1312 // Special case Basic to handle -1 scales.
1313 return !AM.BaseGV && (AM.Scale == 0 || AM.Scale == -1) && AM.BaseOffs == 0;
1316 llvm_unreachable("Invalid LSRUse Kind!");
1319 static bool isLegalUse(TargetLowering::AddrMode AM,
1320 int64_t MinOffset, int64_t MaxOffset,
1321 LSRUse::KindType Kind, Type *AccessTy,
1322 const TargetLowering *TLI) {
1323 // Check for overflow.
1324 if (((int64_t)((uint64_t)AM.BaseOffs + MinOffset) > AM.BaseOffs) !=
1327 AM.BaseOffs = (uint64_t)AM.BaseOffs + MinOffset;
1328 if (isLegalUse(AM, Kind, AccessTy, TLI)) {
1329 AM.BaseOffs = (uint64_t)AM.BaseOffs - MinOffset;
1330 // Check for overflow.
1331 if (((int64_t)((uint64_t)AM.BaseOffs + MaxOffset) > AM.BaseOffs) !=
1334 AM.BaseOffs = (uint64_t)AM.BaseOffs + MaxOffset;
1335 return isLegalUse(AM, Kind, AccessTy, TLI);
1340 static bool isAlwaysFoldable(int64_t BaseOffs,
1341 GlobalValue *BaseGV,
1343 LSRUse::KindType Kind, Type *AccessTy,
1344 const TargetLowering *TLI) {
1345 // Fast-path: zero is always foldable.
1346 if (BaseOffs == 0 && !BaseGV) return true;
1348 // Conservatively, create an address with an immediate and a
1349 // base and a scale.
1350 TargetLowering::AddrMode AM;
1351 AM.BaseOffs = BaseOffs;
1353 AM.HasBaseReg = HasBaseReg;
1354 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1356 // Canonicalize a scale of 1 to a base register if the formula doesn't
1357 // already have a base register.
1358 if (!AM.HasBaseReg && AM.Scale == 1) {
1360 AM.HasBaseReg = true;
1363 return isLegalUse(AM, Kind, AccessTy, TLI);
1366 static bool isAlwaysFoldable(const SCEV *S,
1367 int64_t MinOffset, int64_t MaxOffset,
1369 LSRUse::KindType Kind, Type *AccessTy,
1370 const TargetLowering *TLI,
1371 ScalarEvolution &SE) {
1372 // Fast-path: zero is always foldable.
1373 if (S->isZero()) return true;
1375 // Conservatively, create an address with an immediate and a
1376 // base and a scale.
1377 int64_t BaseOffs = ExtractImmediate(S, SE);
1378 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1380 // If there's anything else involved, it's not foldable.
1381 if (!S->isZero()) return false;
1383 // Fast-path: zero is always foldable.
1384 if (BaseOffs == 0 && !BaseGV) return true;
1386 // Conservatively, create an address with an immediate and a
1387 // base and a scale.
1388 TargetLowering::AddrMode AM;
1389 AM.BaseOffs = BaseOffs;
1391 AM.HasBaseReg = HasBaseReg;
1392 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1394 return isLegalUse(AM, MinOffset, MaxOffset, Kind, AccessTy, TLI);
1399 /// UseMapDenseMapInfo - A DenseMapInfo implementation for holding
1400 /// DenseMaps and DenseSets of pairs of const SCEV* and LSRUse::Kind.
1401 struct UseMapDenseMapInfo {
1402 static std::pair<const SCEV *, LSRUse::KindType> getEmptyKey() {
1403 return std::make_pair(reinterpret_cast<const SCEV *>(-1), LSRUse::Basic);
1406 static std::pair<const SCEV *, LSRUse::KindType> getTombstoneKey() {
1407 return std::make_pair(reinterpret_cast<const SCEV *>(-2), LSRUse::Basic);
1411 getHashValue(const std::pair<const SCEV *, LSRUse::KindType> &V) {
1412 unsigned Result = DenseMapInfo<const SCEV *>::getHashValue(V.first);
1413 Result ^= DenseMapInfo<unsigned>::getHashValue(unsigned(V.second));
1417 static bool isEqual(const std::pair<const SCEV *, LSRUse::KindType> &LHS,
1418 const std::pair<const SCEV *, LSRUse::KindType> &RHS) {
1423 /// IVInc - An individual increment in a Chain of IV increments.
1424 /// Relate an IV user to an expression that computes the IV it uses from the IV
1425 /// used by the previous link in the Chain.
1427 /// For the head of a chain, IncExpr holds the absolute SCEV expression for the
1428 /// original IVOperand. The head of the chain's IVOperand is only valid during
1429 /// chain collection, before LSR replaces IV users. During chain generation,
1430 /// IncExpr can be used to find the new IVOperand that computes the same
1433 Instruction *UserInst;
1435 const SCEV *IncExpr;
1437 IVInc(Instruction *U, Value *O, const SCEV *E):
1438 UserInst(U), IVOperand(O), IncExpr(E) {}
1441 // IVChain - The list of IV increments in program order.
1442 // We typically add the head of a chain without finding subsequent links.
1444 SmallVector<IVInc,1> Incs;
1445 const SCEV *ExprBase;
1447 IVChain() : ExprBase(0) {}
1449 IVChain(const IVInc &Head, const SCEV *Base)
1450 : Incs(1, Head), ExprBase(Base) {}
1452 typedef SmallVectorImpl<IVInc>::const_iterator const_iterator;
1454 // begin - return the first increment in the chain.
1455 const_iterator begin() const {
1456 assert(!Incs.empty());
1457 return llvm::next(Incs.begin());
1459 const_iterator end() const {
1463 // hasIncs - Returns true if this chain contains any increments.
1464 bool hasIncs() const { return Incs.size() >= 2; }
1466 // add - Add an IVInc to the end of this chain.
1467 void add(const IVInc &X) { Incs.push_back(X); }
1469 // tailUserInst - Returns the last UserInst in the chain.
1470 Instruction *tailUserInst() const { return Incs.back().UserInst; }
1472 // isProfitableIncrement - Returns true if IncExpr can be profitably added to
1474 bool isProfitableIncrement(const SCEV *OperExpr,
1475 const SCEV *IncExpr,
1479 /// ChainUsers - Helper for CollectChains to track multiple IV increment uses.
1480 /// Distinguish between FarUsers that definitely cross IV increments and
1481 /// NearUsers that may be used between IV increments.
1483 SmallPtrSet<Instruction*, 4> FarUsers;
1484 SmallPtrSet<Instruction*, 4> NearUsers;
1487 /// LSRInstance - This class holds state for the main loop strength reduction
1491 ScalarEvolution &SE;
1494 const TargetLowering *const TLI;
1498 /// IVIncInsertPos - This is the insert position that the current loop's
1499 /// induction variable increment should be placed. In simple loops, this is
1500 /// the latch block's terminator. But in more complicated cases, this is a
1501 /// position which will dominate all the in-loop post-increment users.
1502 Instruction *IVIncInsertPos;
1504 /// Factors - Interesting factors between use strides.
1505 SmallSetVector<int64_t, 8> Factors;
1507 /// Types - Interesting use types, to facilitate truncation reuse.
1508 SmallSetVector<Type *, 4> Types;
1510 /// Fixups - The list of operands which are to be replaced.
1511 SmallVector<LSRFixup, 16> Fixups;
1513 /// Uses - The list of interesting uses.
1514 SmallVector<LSRUse, 16> Uses;
1516 /// RegUses - Track which uses use which register candidates.
1517 RegUseTracker RegUses;
1519 // Limit the number of chains to avoid quadratic behavior. We don't expect to
1520 // have more than a few IV increment chains in a loop. Missing a Chain falls
1521 // back to normal LSR behavior for those uses.
1522 static const unsigned MaxChains = 8;
1524 /// IVChainVec - IV users can form a chain of IV increments.
1525 SmallVector<IVChain, MaxChains> IVChainVec;
1527 /// IVIncSet - IV users that belong to profitable IVChains.
1528 SmallPtrSet<Use*, MaxChains> IVIncSet;
1530 void OptimizeShadowIV();
1531 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1532 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1533 void OptimizeLoopTermCond();
1535 void ChainInstruction(Instruction *UserInst, Instruction *IVOper,
1536 SmallVectorImpl<ChainUsers> &ChainUsersVec);
1537 void FinalizeChain(IVChain &Chain);
1538 void CollectChains();
1539 void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
1540 SmallVectorImpl<WeakVH> &DeadInsts);
1542 void CollectInterestingTypesAndFactors();
1543 void CollectFixupsAndInitialFormulae();
1545 LSRFixup &getNewFixup() {
1546 Fixups.push_back(LSRFixup());
1547 return Fixups.back();
1550 // Support for sharing of LSRUses between LSRFixups.
1551 typedef DenseMap<std::pair<const SCEV *, LSRUse::KindType>,
1553 UseMapDenseMapInfo> UseMapTy;
1556 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1557 LSRUse::KindType Kind, Type *AccessTy);
1559 std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
1560 LSRUse::KindType Kind,
1563 void DeleteUse(LSRUse &LU, size_t LUIdx);
1565 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1567 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1568 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1569 void CountRegisters(const Formula &F, size_t LUIdx);
1570 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1572 void CollectLoopInvariantFixupsAndFormulae();
1574 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1575 unsigned Depth = 0);
1576 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1577 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1578 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1579 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1580 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1581 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1582 void GenerateCrossUseConstantOffsets();
1583 void GenerateAllReuseFormulae();
1585 void FilterOutUndesirableDedicatedRegisters();
1587 size_t EstimateSearchSpaceComplexity() const;
1588 void NarrowSearchSpaceByDetectingSupersets();
1589 void NarrowSearchSpaceByCollapsingUnrolledCode();
1590 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
1591 void NarrowSearchSpaceByPickingWinnerRegs();
1592 void NarrowSearchSpaceUsingHeuristics();
1594 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1596 SmallVectorImpl<const Formula *> &Workspace,
1597 const Cost &CurCost,
1598 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1599 DenseSet<const SCEV *> &VisitedRegs) const;
1600 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1602 BasicBlock::iterator
1603 HoistInsertPosition(BasicBlock::iterator IP,
1604 const SmallVectorImpl<Instruction *> &Inputs) const;
1605 BasicBlock::iterator
1606 AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1609 SCEVExpander &Rewriter) const;
1611 Value *Expand(const LSRFixup &LF,
1613 BasicBlock::iterator IP,
1614 SCEVExpander &Rewriter,
1615 SmallVectorImpl<WeakVH> &DeadInsts) const;
1616 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1618 SCEVExpander &Rewriter,
1619 SmallVectorImpl<WeakVH> &DeadInsts,
1621 void Rewrite(const LSRFixup &LF,
1623 SCEVExpander &Rewriter,
1624 SmallVectorImpl<WeakVH> &DeadInsts,
1626 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
1630 LSRInstance(const TargetLowering *tli, Loop *l, Pass *P);
1632 bool getChanged() const { return Changed; }
1634 void print_factors_and_types(raw_ostream &OS) const;
1635 void print_fixups(raw_ostream &OS) const;
1636 void print_uses(raw_ostream &OS) const;
1637 void print(raw_ostream &OS) const;
1643 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1644 /// inside the loop then try to eliminate the cast operation.
1645 void LSRInstance::OptimizeShadowIV() {
1646 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1647 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1650 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1651 UI != E; /* empty */) {
1652 IVUsers::const_iterator CandidateUI = UI;
1654 Instruction *ShadowUse = CandidateUI->getUser();
1655 Type *DestTy = NULL;
1656 bool IsSigned = false;
1658 /* If shadow use is a int->float cast then insert a second IV
1659 to eliminate this cast.
1661 for (unsigned i = 0; i < n; ++i)
1667 for (unsigned i = 0; i < n; ++i, ++d)
1670 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
1672 DestTy = UCast->getDestTy();
1674 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
1676 DestTy = SCast->getDestTy();
1678 if (!DestTy) continue;
1681 // If target does not support DestTy natively then do not apply
1682 // this transformation.
1683 EVT DVT = TLI->getValueType(DestTy);
1684 if (!TLI->isTypeLegal(DVT)) continue;
1687 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1689 if (PH->getNumIncomingValues() != 2) continue;
1691 Type *SrcTy = PH->getType();
1692 int Mantissa = DestTy->getFPMantissaWidth();
1693 if (Mantissa == -1) continue;
1694 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1697 unsigned Entry, Latch;
1698 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1706 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1707 if (!Init) continue;
1708 Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
1709 (double)Init->getSExtValue() :
1710 (double)Init->getZExtValue());
1712 BinaryOperator *Incr =
1713 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1714 if (!Incr) continue;
1715 if (Incr->getOpcode() != Instruction::Add
1716 && Incr->getOpcode() != Instruction::Sub)
1719 /* Initialize new IV, double d = 0.0 in above example. */
1720 ConstantInt *C = NULL;
1721 if (Incr->getOperand(0) == PH)
1722 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1723 else if (Incr->getOperand(1) == PH)
1724 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1730 // Ignore negative constants, as the code below doesn't handle them
1731 // correctly. TODO: Remove this restriction.
1732 if (!C->getValue().isStrictlyPositive()) continue;
1734 /* Add new PHINode. */
1735 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
1737 /* create new increment. '++d' in above example. */
1738 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1739 BinaryOperator *NewIncr =
1740 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1741 Instruction::FAdd : Instruction::FSub,
1742 NewPH, CFP, "IV.S.next.", Incr);
1744 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1745 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1747 /* Remove cast operation */
1748 ShadowUse->replaceAllUsesWith(NewPH);
1749 ShadowUse->eraseFromParent();
1755 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1756 /// set the IV user and stride information and return true, otherwise return
1758 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
1759 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1760 if (UI->getUser() == Cond) {
1761 // NOTE: we could handle setcc instructions with multiple uses here, but
1762 // InstCombine does it as well for simple uses, it's not clear that it
1763 // occurs enough in real life to handle.
1770 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
1771 /// a max computation.
1773 /// This is a narrow solution to a specific, but acute, problem. For loops
1779 /// } while (++i < n);
1781 /// the trip count isn't just 'n', because 'n' might not be positive. And
1782 /// unfortunately this can come up even for loops where the user didn't use
1783 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1784 /// will commonly be lowered like this:
1790 /// } while (++i < n);
1793 /// and then it's possible for subsequent optimization to obscure the if
1794 /// test in such a way that indvars can't find it.
1796 /// When indvars can't find the if test in loops like this, it creates a
1797 /// max expression, which allows it to give the loop a canonical
1798 /// induction variable:
1801 /// max = n < 1 ? 1 : n;
1804 /// } while (++i != max);
1806 /// Canonical induction variables are necessary because the loop passes
1807 /// are designed around them. The most obvious example of this is the
1808 /// LoopInfo analysis, which doesn't remember trip count values. It
1809 /// expects to be able to rediscover the trip count each time it is
1810 /// needed, and it does this using a simple analysis that only succeeds if
1811 /// the loop has a canonical induction variable.
1813 /// However, when it comes time to generate code, the maximum operation
1814 /// can be quite costly, especially if it's inside of an outer loop.
1816 /// This function solves this problem by detecting this type of loop and
1817 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1818 /// the instructions for the maximum computation.
1820 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1821 // Check that the loop matches the pattern we're looking for.
1822 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1823 Cond->getPredicate() != CmpInst::ICMP_NE)
1826 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1827 if (!Sel || !Sel->hasOneUse()) return Cond;
1829 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1830 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1832 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
1834 // Add one to the backedge-taken count to get the trip count.
1835 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
1836 if (IterationCount != SE.getSCEV(Sel)) return Cond;
1838 // Check for a max calculation that matches the pattern. There's no check
1839 // for ICMP_ULE here because the comparison would be with zero, which
1840 // isn't interesting.
1841 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
1842 const SCEVNAryExpr *Max = 0;
1843 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
1844 Pred = ICmpInst::ICMP_SLE;
1846 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
1847 Pred = ICmpInst::ICMP_SLT;
1849 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
1850 Pred = ICmpInst::ICMP_ULT;
1857 // To handle a max with more than two operands, this optimization would
1858 // require additional checking and setup.
1859 if (Max->getNumOperands() != 2)
1862 const SCEV *MaxLHS = Max->getOperand(0);
1863 const SCEV *MaxRHS = Max->getOperand(1);
1865 // ScalarEvolution canonicalizes constants to the left. For < and >, look
1866 // for a comparison with 1. For <= and >=, a comparison with zero.
1868 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
1871 // Check the relevant induction variable for conformance to
1873 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
1874 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1875 if (!AR || !AR->isAffine() ||
1876 AR->getStart() != One ||
1877 AR->getStepRecurrence(SE) != One)
1880 assert(AR->getLoop() == L &&
1881 "Loop condition operand is an addrec in a different loop!");
1883 // Check the right operand of the select, and remember it, as it will
1884 // be used in the new comparison instruction.
1886 if (ICmpInst::isTrueWhenEqual(Pred)) {
1887 // Look for n+1, and grab n.
1888 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
1889 if (isa<ConstantInt>(BO->getOperand(1)) &&
1890 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1891 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1892 NewRHS = BO->getOperand(0);
1893 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
1894 if (isa<ConstantInt>(BO->getOperand(1)) &&
1895 cast<ConstantInt>(BO->getOperand(1))->isOne() &&
1896 SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1897 NewRHS = BO->getOperand(0);
1900 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
1901 NewRHS = Sel->getOperand(1);
1902 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
1903 NewRHS = Sel->getOperand(2);
1904 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
1905 NewRHS = SU->getValue();
1907 // Max doesn't match expected pattern.
1910 // Determine the new comparison opcode. It may be signed or unsigned,
1911 // and the original comparison may be either equality or inequality.
1912 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
1913 Pred = CmpInst::getInversePredicate(Pred);
1915 // Ok, everything looks ok to change the condition into an SLT or SGE and
1916 // delete the max calculation.
1918 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
1920 // Delete the max calculation instructions.
1921 Cond->replaceAllUsesWith(NewCond);
1922 CondUse->setUser(NewCond);
1923 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
1924 Cond->eraseFromParent();
1925 Sel->eraseFromParent();
1926 if (Cmp->use_empty())
1927 Cmp->eraseFromParent();
1931 /// OptimizeLoopTermCond - Change loop terminating condition to use the
1932 /// postinc iv when possible.
1934 LSRInstance::OptimizeLoopTermCond() {
1935 SmallPtrSet<Instruction *, 4> PostIncs;
1937 BasicBlock *LatchBlock = L->getLoopLatch();
1938 SmallVector<BasicBlock*, 8> ExitingBlocks;
1939 L->getExitingBlocks(ExitingBlocks);
1941 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
1942 BasicBlock *ExitingBlock = ExitingBlocks[i];
1944 // Get the terminating condition for the loop if possible. If we
1945 // can, we want to change it to use a post-incremented version of its
1946 // induction variable, to allow coalescing the live ranges for the IV into
1947 // one register value.
1949 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
1952 // FIXME: Overly conservative, termination condition could be an 'or' etc..
1953 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
1956 // Search IVUsesByStride to find Cond's IVUse if there is one.
1957 IVStrideUse *CondUse = 0;
1958 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1959 if (!FindIVUserForCond(Cond, CondUse))
1962 // If the trip count is computed in terms of a max (due to ScalarEvolution
1963 // being unable to find a sufficient guard, for example), change the loop
1964 // comparison to use SLT or ULT instead of NE.
1965 // One consequence of doing this now is that it disrupts the count-down
1966 // optimization. That's not always a bad thing though, because in such
1967 // cases it may still be worthwhile to avoid a max.
1968 Cond = OptimizeMax(Cond, CondUse);
1970 // If this exiting block dominates the latch block, it may also use
1971 // the post-inc value if it won't be shared with other uses.
1972 // Check for dominance.
1973 if (!DT.dominates(ExitingBlock, LatchBlock))
1976 // Conservatively avoid trying to use the post-inc value in non-latch
1977 // exits if there may be pre-inc users in intervening blocks.
1978 if (LatchBlock != ExitingBlock)
1979 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1980 // Test if the use is reachable from the exiting block. This dominator
1981 // query is a conservative approximation of reachability.
1982 if (&*UI != CondUse &&
1983 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
1984 // Conservatively assume there may be reuse if the quotient of their
1985 // strides could be a legal scale.
1986 const SCEV *A = IU.getStride(*CondUse, L);
1987 const SCEV *B = IU.getStride(*UI, L);
1988 if (!A || !B) continue;
1989 if (SE.getTypeSizeInBits(A->getType()) !=
1990 SE.getTypeSizeInBits(B->getType())) {
1991 if (SE.getTypeSizeInBits(A->getType()) >
1992 SE.getTypeSizeInBits(B->getType()))
1993 B = SE.getSignExtendExpr(B, A->getType());
1995 A = SE.getSignExtendExpr(A, B->getType());
1997 if (const SCEVConstant *D =
1998 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
1999 const ConstantInt *C = D->getValue();
2000 // Stride of one or negative one can have reuse with non-addresses.
2001 if (C->isOne() || C->isAllOnesValue())
2002 goto decline_post_inc;
2003 // Avoid weird situations.
2004 if (C->getValue().getMinSignedBits() >= 64 ||
2005 C->getValue().isMinSignedValue())
2006 goto decline_post_inc;
2007 // Without TLI, assume that any stride might be valid, and so any
2008 // use might be shared.
2010 goto decline_post_inc;
2011 // Check for possible scaled-address reuse.
2012 Type *AccessTy = getAccessType(UI->getUser());
2013 TargetLowering::AddrMode AM;
2014 AM.Scale = C->getSExtValue();
2015 if (TLI->isLegalAddressingMode(AM, AccessTy))
2016 goto decline_post_inc;
2017 AM.Scale = -AM.Scale;
2018 if (TLI->isLegalAddressingMode(AM, AccessTy))
2019 goto decline_post_inc;
2023 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
2026 // It's possible for the setcc instruction to be anywhere in the loop, and
2027 // possible for it to have multiple users. If it is not immediately before
2028 // the exiting block branch, move it.
2029 if (&*++BasicBlock::iterator(Cond) != TermBr) {
2030 if (Cond->hasOneUse()) {
2031 Cond->moveBefore(TermBr);
2033 // Clone the terminating condition and insert into the loopend.
2034 ICmpInst *OldCond = Cond;
2035 Cond = cast<ICmpInst>(Cond->clone());
2036 Cond->setName(L->getHeader()->getName() + ".termcond");
2037 ExitingBlock->getInstList().insert(TermBr, Cond);
2039 // Clone the IVUse, as the old use still exists!
2040 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
2041 TermBr->replaceUsesOfWith(OldCond, Cond);
2045 // If we get to here, we know that we can transform the setcc instruction to
2046 // use the post-incremented version of the IV, allowing us to coalesce the
2047 // live ranges for the IV correctly.
2048 CondUse->transformToPostInc(L);
2051 PostIncs.insert(Cond);
2055 // Determine an insertion point for the loop induction variable increment. It
2056 // must dominate all the post-inc comparisons we just set up, and it must
2057 // dominate the loop latch edge.
2058 IVIncInsertPos = L->getLoopLatch()->getTerminator();
2059 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(),
2060 E = PostIncs.end(); I != E; ++I) {
2062 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
2064 if (BB == (*I)->getParent())
2065 IVIncInsertPos = *I;
2066 else if (BB != IVIncInsertPos->getParent())
2067 IVIncInsertPos = BB->getTerminator();
2071 /// reconcileNewOffset - Determine if the given use can accommodate a fixup
2072 /// at the given offset and other details. If so, update the use and
2075 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
2076 LSRUse::KindType Kind, Type *AccessTy) {
2077 int64_t NewMinOffset = LU.MinOffset;
2078 int64_t NewMaxOffset = LU.MaxOffset;
2079 Type *NewAccessTy = AccessTy;
2081 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
2082 // something conservative, however this can pessimize in the case that one of
2083 // the uses will have all its uses outside the loop, for example.
2084 if (LU.Kind != Kind)
2086 // Conservatively assume HasBaseReg is true for now.
2087 if (NewOffset < LU.MinOffset) {
2088 if (!isAlwaysFoldable(LU.MaxOffset - NewOffset, 0, HasBaseReg,
2089 Kind, AccessTy, TLI))
2091 NewMinOffset = NewOffset;
2092 } else if (NewOffset > LU.MaxOffset) {
2093 if (!isAlwaysFoldable(NewOffset - LU.MinOffset, 0, HasBaseReg,
2094 Kind, AccessTy, TLI))
2096 NewMaxOffset = NewOffset;
2098 // Check for a mismatched access type, and fall back conservatively as needed.
2099 // TODO: Be less conservative when the type is similar and can use the same
2100 // addressing modes.
2101 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
2102 NewAccessTy = Type::getVoidTy(AccessTy->getContext());
2105 LU.MinOffset = NewMinOffset;
2106 LU.MaxOffset = NewMaxOffset;
2107 LU.AccessTy = NewAccessTy;
2108 if (NewOffset != LU.Offsets.back())
2109 LU.Offsets.push_back(NewOffset);
2113 /// getUse - Return an LSRUse index and an offset value for a fixup which
2114 /// needs the given expression, with the given kind and optional access type.
2115 /// Either reuse an existing use or create a new one, as needed.
2116 std::pair<size_t, int64_t>
2117 LSRInstance::getUse(const SCEV *&Expr,
2118 LSRUse::KindType Kind, Type *AccessTy) {
2119 const SCEV *Copy = Expr;
2120 int64_t Offset = ExtractImmediate(Expr, SE);
2122 // Basic uses can't accept any offset, for example.
2123 if (!isAlwaysFoldable(Offset, 0, /*HasBaseReg=*/true, Kind, AccessTy, TLI)) {
2128 std::pair<UseMapTy::iterator, bool> P =
2129 UseMap.insert(std::make_pair(std::make_pair(Expr, Kind), 0));
2131 // A use already existed with this base.
2132 size_t LUIdx = P.first->second;
2133 LSRUse &LU = Uses[LUIdx];
2134 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
2136 return std::make_pair(LUIdx, Offset);
2139 // Create a new use.
2140 size_t LUIdx = Uses.size();
2141 P.first->second = LUIdx;
2142 Uses.push_back(LSRUse(Kind, AccessTy));
2143 LSRUse &LU = Uses[LUIdx];
2145 // We don't need to track redundant offsets, but we don't need to go out
2146 // of our way here to avoid them.
2147 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
2148 LU.Offsets.push_back(Offset);
2150 LU.MinOffset = Offset;
2151 LU.MaxOffset = Offset;
2152 return std::make_pair(LUIdx, Offset);
2155 /// DeleteUse - Delete the given use from the Uses list.
2156 void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
2157 if (&LU != &Uses.back())
2158 std::swap(LU, Uses.back());
2162 RegUses.SwapAndDropUse(LUIdx, Uses.size());
2165 /// FindUseWithFormula - Look for a use distinct from OrigLU which is has
2166 /// a formula that has the same registers as the given formula.
2168 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
2169 const LSRUse &OrigLU) {
2170 // Search all uses for the formula. This could be more clever.
2171 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2172 LSRUse &LU = Uses[LUIdx];
2173 // Check whether this use is close enough to OrigLU, to see whether it's
2174 // worthwhile looking through its formulae.
2175 // Ignore ICmpZero uses because they may contain formulae generated by
2176 // GenerateICmpZeroScales, in which case adding fixup offsets may
2178 if (&LU != &OrigLU &&
2179 LU.Kind != LSRUse::ICmpZero &&
2180 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
2181 LU.WidestFixupType == OrigLU.WidestFixupType &&
2182 LU.HasFormulaWithSameRegs(OrigF)) {
2183 // Scan through this use's formulae.
2184 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
2185 E = LU.Formulae.end(); I != E; ++I) {
2186 const Formula &F = *I;
2187 // Check to see if this formula has the same registers and symbols
2189 if (F.BaseRegs == OrigF.BaseRegs &&
2190 F.ScaledReg == OrigF.ScaledReg &&
2191 F.AM.BaseGV == OrigF.AM.BaseGV &&
2192 F.AM.Scale == OrigF.AM.Scale &&
2193 F.UnfoldedOffset == OrigF.UnfoldedOffset) {
2194 if (F.AM.BaseOffs == 0)
2196 // This is the formula where all the registers and symbols matched;
2197 // there aren't going to be any others. Since we declined it, we
2198 // can skip the rest of the formulae and proceed to the next LSRUse.
2205 // Nothing looked good.
2209 void LSRInstance::CollectInterestingTypesAndFactors() {
2210 SmallSetVector<const SCEV *, 4> Strides;
2212 // Collect interesting types and strides.
2213 SmallVector<const SCEV *, 4> Worklist;
2214 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2215 const SCEV *Expr = IU.getExpr(*UI);
2217 // Collect interesting types.
2218 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2220 // Add strides for mentioned loops.
2221 Worklist.push_back(Expr);
2223 const SCEV *S = Worklist.pop_back_val();
2224 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2225 if (AR->getLoop() == L)
2226 Strides.insert(AR->getStepRecurrence(SE));
2227 Worklist.push_back(AR->getStart());
2228 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2229 Worklist.append(Add->op_begin(), Add->op_end());
2231 } while (!Worklist.empty());
2234 // Compute interesting factors from the set of interesting strides.
2235 for (SmallSetVector<const SCEV *, 4>::const_iterator
2236 I = Strides.begin(), E = Strides.end(); I != E; ++I)
2237 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
2238 llvm::next(I); NewStrideIter != E; ++NewStrideIter) {
2239 const SCEV *OldStride = *I;
2240 const SCEV *NewStride = *NewStrideIter;
2242 if (SE.getTypeSizeInBits(OldStride->getType()) !=
2243 SE.getTypeSizeInBits(NewStride->getType())) {
2244 if (SE.getTypeSizeInBits(OldStride->getType()) >
2245 SE.getTypeSizeInBits(NewStride->getType()))
2246 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2248 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2250 if (const SCEVConstant *Factor =
2251 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2253 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2254 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2255 } else if (const SCEVConstant *Factor =
2256 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2259 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2260 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2264 // If all uses use the same type, don't bother looking for truncation-based
2266 if (Types.size() == 1)
2269 DEBUG(print_factors_and_types(dbgs()));
2272 /// findIVOperand - Helper for CollectChains that finds an IV operand (computed
2273 /// by an AddRec in this loop) within [OI,OE) or returns OE. If IVUsers mapped
2274 /// Instructions to IVStrideUses, we could partially skip this.
2275 static User::op_iterator
2276 findIVOperand(User::op_iterator OI, User::op_iterator OE,
2277 Loop *L, ScalarEvolution &SE) {
2278 for(; OI != OE; ++OI) {
2279 if (Instruction *Oper = dyn_cast<Instruction>(*OI)) {
2280 if (!SE.isSCEVable(Oper->getType()))
2283 if (const SCEVAddRecExpr *AR =
2284 dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) {
2285 if (AR->getLoop() == L)
2293 /// getWideOperand - IVChain logic must consistenctly peek base TruncInst
2294 /// operands, so wrap it in a convenient helper.
2295 static Value *getWideOperand(Value *Oper) {
2296 if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper))
2297 return Trunc->getOperand(0);
2301 /// isCompatibleIVType - Return true if we allow an IV chain to include both
2303 static bool isCompatibleIVType(Value *LVal, Value *RVal) {
2304 Type *LType = LVal->getType();
2305 Type *RType = RVal->getType();
2306 return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy());
2309 /// getExprBase - Return an approximation of this SCEV expression's "base", or
2310 /// NULL for any constant. Returning the expression itself is
2311 /// conservative. Returning a deeper subexpression is more precise and valid as
2312 /// long as it isn't less complex than another subexpression. For expressions
2313 /// involving multiple unscaled values, we need to return the pointer-type
2314 /// SCEVUnknown. This avoids forming chains across objects, such as:
2315 /// PrevOper==a[i], IVOper==b[i], IVInc==b-a.
2317 /// Since SCEVUnknown is the rightmost type, and pointers are the rightmost
2318 /// SCEVUnknown, we simply return the rightmost SCEV operand.
2319 static const SCEV *getExprBase(const SCEV *S) {
2320 switch (S->getSCEVType()) {
2321 default: // uncluding scUnknown.
2326 return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand());
2328 return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand());
2330 return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand());
2332 // Skip over scaled operands (scMulExpr) to follow add operands as long as
2333 // there's nothing more complex.
2334 // FIXME: not sure if we want to recognize negation.
2335 const SCEVAddExpr *Add = cast<SCEVAddExpr>(S);
2336 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(Add->op_end()),
2337 E(Add->op_begin()); I != E; ++I) {
2338 const SCEV *SubExpr = *I;
2339 if (SubExpr->getSCEVType() == scAddExpr)
2340 return getExprBase(SubExpr);
2342 if (SubExpr->getSCEVType() != scMulExpr)
2345 return S; // all operands are scaled, be conservative.
2348 return getExprBase(cast<SCEVAddRecExpr>(S)->getStart());
2352 /// Return true if the chain increment is profitable to expand into a loop
2353 /// invariant value, which may require its own register. A profitable chain
2354 /// increment will be an offset relative to the same base. We allow such offsets
2355 /// to potentially be used as chain increment as long as it's not obviously
2356 /// expensive to expand using real instructions.
2357 bool IVChain::isProfitableIncrement(const SCEV *OperExpr,
2358 const SCEV *IncExpr,
2359 ScalarEvolution &SE) {
2360 // Aggressively form chains when -stress-ivchain.
2364 // Do not replace a constant offset from IV head with a nonconstant IV
2366 if (!isa<SCEVConstant>(IncExpr)) {
2367 const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Incs[0].IVOperand));
2368 if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr)))
2372 SmallPtrSet<const SCEV*, 8> Processed;
2373 return !isHighCostExpansion(IncExpr, Processed, SE);
2376 /// Return true if the number of registers needed for the chain is estimated to
2377 /// be less than the number required for the individual IV users. First prohibit
2378 /// any IV users that keep the IV live across increments (the Users set should
2379 /// be empty). Next count the number and type of increments in the chain.
2381 /// Chaining IVs can lead to considerable code bloat if ISEL doesn't
2382 /// effectively use postinc addressing modes. Only consider it profitable it the
2383 /// increments can be computed in fewer registers when chained.
2385 /// TODO: Consider IVInc free if it's already used in another chains.
2387 isProfitableChain(IVChain &Chain, SmallPtrSet<Instruction*, 4> &Users,
2388 ScalarEvolution &SE, const TargetLowering *TLI) {
2392 if (!Chain.hasIncs())
2395 if (!Users.empty()) {
2396 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " users:\n";
2397 for (SmallPtrSet<Instruction*, 4>::const_iterator I = Users.begin(),
2398 E = Users.end(); I != E; ++I) {
2399 dbgs() << " " << **I << "\n";
2403 assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
2405 // The chain itself may require a register, so intialize cost to 1.
2408 // A complete chain likely eliminates the need for keeping the original IV in
2409 // a register. LSR does not currently know how to form a complete chain unless
2410 // the header phi already exists.
2411 if (isa<PHINode>(Chain.tailUserInst())
2412 && SE.getSCEV(Chain.tailUserInst()) == Chain.Incs[0].IncExpr) {
2415 const SCEV *LastIncExpr = 0;
2416 unsigned NumConstIncrements = 0;
2417 unsigned NumVarIncrements = 0;
2418 unsigned NumReusedIncrements = 0;
2419 for (IVChain::const_iterator I = Chain.begin(), E = Chain.end();
2422 if (I->IncExpr->isZero())
2425 // Incrementing by zero or some constant is neutral. We assume constants can
2426 // be folded into an addressing mode or an add's immediate operand.
2427 if (isa<SCEVConstant>(I->IncExpr)) {
2428 ++NumConstIncrements;
2432 if (I->IncExpr == LastIncExpr)
2433 ++NumReusedIncrements;
2437 LastIncExpr = I->IncExpr;
2439 // An IV chain with a single increment is handled by LSR's postinc
2440 // uses. However, a chain with multiple increments requires keeping the IV's
2441 // value live longer than it needs to be if chained.
2442 if (NumConstIncrements > 1)
2445 // Materializing increment expressions in the preheader that didn't exist in
2446 // the original code may cost a register. For example, sign-extended array
2447 // indices can produce ridiculous increments like this:
2448 // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64)))
2449 cost += NumVarIncrements;
2451 // Reusing variable increments likely saves a register to hold the multiple of
2453 cost -= NumReusedIncrements;
2455 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " Cost: " << cost
2461 /// ChainInstruction - Add this IV user to an existing chain or make it the head
2463 void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper,
2464 SmallVectorImpl<ChainUsers> &ChainUsersVec) {
2465 // When IVs are used as types of varying widths, they are generally converted
2466 // to a wider type with some uses remaining narrow under a (free) trunc.
2467 Value *const NextIV = getWideOperand(IVOper);
2468 const SCEV *const OperExpr = SE.getSCEV(NextIV);
2469 const SCEV *const OperExprBase = getExprBase(OperExpr);
2471 // Visit all existing chains. Check if its IVOper can be computed as a
2472 // profitable loop invariant increment from the last link in the Chain.
2473 unsigned ChainIdx = 0, NChains = IVChainVec.size();
2474 const SCEV *LastIncExpr = 0;
2475 for (; ChainIdx < NChains; ++ChainIdx) {
2476 IVChain &Chain = IVChainVec[ChainIdx];
2478 // Prune the solution space aggressively by checking that both IV operands
2479 // are expressions that operate on the same unscaled SCEVUnknown. This
2480 // "base" will be canceled by the subsequent getMinusSCEV call. Checking
2481 // first avoids creating extra SCEV expressions.
2482 if (!StressIVChain && Chain.ExprBase != OperExprBase)
2485 Value *PrevIV = getWideOperand(Chain.Incs.back().IVOperand);
2486 if (!isCompatibleIVType(PrevIV, NextIV))
2489 // A phi node terminates a chain.
2490 if (isa<PHINode>(UserInst) && isa<PHINode>(Chain.tailUserInst()))
2493 // The increment must be loop-invariant so it can be kept in a register.
2494 const SCEV *PrevExpr = SE.getSCEV(PrevIV);
2495 const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr);
2496 if (!SE.isLoopInvariant(IncExpr, L))
2499 if (Chain.isProfitableIncrement(OperExpr, IncExpr, SE)) {
2500 LastIncExpr = IncExpr;
2504 // If we haven't found a chain, create a new one, unless we hit the max. Don't
2505 // bother for phi nodes, because they must be last in the chain.
2506 if (ChainIdx == NChains) {
2507 if (isa<PHINode>(UserInst))
2509 if (NChains >= MaxChains && !StressIVChain) {
2510 DEBUG(dbgs() << "IV Chain Limit\n");
2513 LastIncExpr = OperExpr;
2514 // IVUsers may have skipped over sign/zero extensions. We don't currently
2515 // attempt to form chains involving extensions unless they can be hoisted
2516 // into this loop's AddRec.
2517 if (!isa<SCEVAddRecExpr>(LastIncExpr))
2520 IVChainVec.push_back(IVChain(IVInc(UserInst, IVOper, LastIncExpr),
2522 ChainUsersVec.resize(NChains);
2523 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Head: (" << *UserInst
2524 << ") IV=" << *LastIncExpr << "\n");
2526 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Inc: (" << *UserInst
2527 << ") IV+" << *LastIncExpr << "\n");
2528 // Add this IV user to the end of the chain.
2529 IVChainVec[ChainIdx].add(IVInc(UserInst, IVOper, LastIncExpr));
2532 SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers;
2533 // This chain's NearUsers become FarUsers.
2534 if (!LastIncExpr->isZero()) {
2535 ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(),
2540 // All other uses of IVOperand become near uses of the chain.
2541 // We currently ignore intermediate values within SCEV expressions, assuming
2542 // they will eventually be used be the current chain, or can be computed
2543 // from one of the chain increments. To be more precise we could
2544 // transitively follow its user and only add leaf IV users to the set.
2545 for (Value::use_iterator UseIter = IVOper->use_begin(),
2546 UseEnd = IVOper->use_end(); UseIter != UseEnd; ++UseIter) {
2547 Instruction *OtherUse = dyn_cast<Instruction>(*UseIter);
2548 if (!OtherUse || OtherUse == UserInst)
2550 if (SE.isSCEVable(OtherUse->getType())
2551 && !isa<SCEVUnknown>(SE.getSCEV(OtherUse))
2552 && IU.isIVUserOrOperand(OtherUse)) {
2555 NearUsers.insert(OtherUse);
2558 // Since this user is part of the chain, it's no longer considered a use
2560 ChainUsersVec[ChainIdx].FarUsers.erase(UserInst);
2563 /// CollectChains - Populate the vector of Chains.
2565 /// This decreases ILP at the architecture level. Targets with ample registers,
2566 /// multiple memory ports, and no register renaming probably don't want
2567 /// this. However, such targets should probably disable LSR altogether.
2569 /// The job of LSR is to make a reasonable choice of induction variables across
2570 /// the loop. Subsequent passes can easily "unchain" computation exposing more
2571 /// ILP *within the loop* if the target wants it.
2573 /// Finding the best IV chain is potentially a scheduling problem. Since LSR
2574 /// will not reorder memory operations, it will recognize this as a chain, but
2575 /// will generate redundant IV increments. Ideally this would be corrected later
2576 /// by a smart scheduler:
2582 /// TODO: Walk the entire domtree within this loop, not just the path to the
2583 /// loop latch. This will discover chains on side paths, but requires
2584 /// maintaining multiple copies of the Chains state.
2585 void LSRInstance::CollectChains() {
2586 DEBUG(dbgs() << "Collecting IV Chains.\n");
2587 SmallVector<ChainUsers, 8> ChainUsersVec;
2589 SmallVector<BasicBlock *,8> LatchPath;
2590 BasicBlock *LoopHeader = L->getHeader();
2591 for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch());
2592 Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) {
2593 LatchPath.push_back(Rung->getBlock());
2595 LatchPath.push_back(LoopHeader);
2597 // Walk the instruction stream from the loop header to the loop latch.
2598 for (SmallVectorImpl<BasicBlock *>::reverse_iterator
2599 BBIter = LatchPath.rbegin(), BBEnd = LatchPath.rend();
2600 BBIter != BBEnd; ++BBIter) {
2601 for (BasicBlock::iterator I = (*BBIter)->begin(), E = (*BBIter)->end();
2603 // Skip instructions that weren't seen by IVUsers analysis.
2604 if (isa<PHINode>(I) || !IU.isIVUserOrOperand(I))
2607 // Ignore users that are part of a SCEV expression. This way we only
2608 // consider leaf IV Users. This effectively rediscovers a portion of
2609 // IVUsers analysis but in program order this time.
2610 if (SE.isSCEVable(I->getType()) && !isa<SCEVUnknown>(SE.getSCEV(I)))
2613 // Remove this instruction from any NearUsers set it may be in.
2614 for (unsigned ChainIdx = 0, NChains = IVChainVec.size();
2615 ChainIdx < NChains; ++ChainIdx) {
2616 ChainUsersVec[ChainIdx].NearUsers.erase(I);
2618 // Search for operands that can be chained.
2619 SmallPtrSet<Instruction*, 4> UniqueOperands;
2620 User::op_iterator IVOpEnd = I->op_end();
2621 User::op_iterator IVOpIter = findIVOperand(I->op_begin(), IVOpEnd, L, SE);
2622 while (IVOpIter != IVOpEnd) {
2623 Instruction *IVOpInst = cast<Instruction>(*IVOpIter);
2624 if (UniqueOperands.insert(IVOpInst))
2625 ChainInstruction(I, IVOpInst, ChainUsersVec);
2626 IVOpIter = findIVOperand(llvm::next(IVOpIter), IVOpEnd, L, SE);
2628 } // Continue walking down the instructions.
2629 } // Continue walking down the domtree.
2630 // Visit phi backedges to determine if the chain can generate the IV postinc.
2631 for (BasicBlock::iterator I = L->getHeader()->begin();
2632 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2633 if (!SE.isSCEVable(PN->getType()))
2637 dyn_cast<Instruction>(PN->getIncomingValueForBlock(L->getLoopLatch()));
2639 ChainInstruction(PN, IncV, ChainUsersVec);
2641 // Remove any unprofitable chains.
2642 unsigned ChainIdx = 0;
2643 for (unsigned UsersIdx = 0, NChains = IVChainVec.size();
2644 UsersIdx < NChains; ++UsersIdx) {
2645 if (!isProfitableChain(IVChainVec[UsersIdx],
2646 ChainUsersVec[UsersIdx].FarUsers, SE, TLI))
2648 // Preserve the chain at UsesIdx.
2649 if (ChainIdx != UsersIdx)
2650 IVChainVec[ChainIdx] = IVChainVec[UsersIdx];
2651 FinalizeChain(IVChainVec[ChainIdx]);
2654 IVChainVec.resize(ChainIdx);
2657 void LSRInstance::FinalizeChain(IVChain &Chain) {
2658 assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
2659 DEBUG(dbgs() << "Final Chain: " << *Chain.Incs[0].UserInst << "\n");
2661 for (IVChain::const_iterator I = Chain.begin(), E = Chain.end();
2663 DEBUG(dbgs() << " Inc: " << *I->UserInst << "\n");
2664 User::op_iterator UseI =
2665 std::find(I->UserInst->op_begin(), I->UserInst->op_end(), I->IVOperand);
2666 assert(UseI != I->UserInst->op_end() && "cannot find IV operand");
2667 IVIncSet.insert(UseI);
2671 /// Return true if the IVInc can be folded into an addressing mode.
2672 static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst,
2673 Value *Operand, const TargetLowering *TLI) {
2674 const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr);
2675 if (!IncConst || !isAddressUse(UserInst, Operand))
2678 if (IncConst->getValue()->getValue().getMinSignedBits() > 64)
2681 int64_t IncOffset = IncConst->getValue()->getSExtValue();
2682 if (!isAlwaysFoldable(IncOffset, /*BaseGV=*/0, /*HaseBaseReg=*/false,
2683 LSRUse::Address, getAccessType(UserInst), TLI))
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(*J, LU.MinOffset, LU.MaxOffset,
3099 Base.getNumRegs() > 1,
3100 LU.Kind, LU.AccessTy, TLI, SE))
3103 // Collect all operands except *J.
3104 SmallVector<const SCEV *, 8> InnerAddOps
3105 (((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
3107 (llvm::next(J), ((const SmallVector<const SCEV *, 8> &)AddOps).end());
3109 // Don't leave just a constant behind in a register if the constant could
3110 // be folded into an immediate field.
3111 if (InnerAddOps.size() == 1 &&
3112 isAlwaysFoldable(InnerAddOps[0], LU.MinOffset, LU.MaxOffset,
3113 Base.getNumRegs() > 1,
3114 LU.Kind, LU.AccessTy, TLI, SE))
3117 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
3118 if (InnerSum->isZero())
3122 // Add the remaining pieces of the add back into the new formula.
3123 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
3124 if (TLI && InnerSumSC &&
3125 SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
3126 TLI->isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3127 InnerSumSC->getValue()->getZExtValue())) {
3128 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
3129 InnerSumSC->getValue()->getZExtValue();
3130 F.BaseRegs.erase(F.BaseRegs.begin() + i);
3132 F.BaseRegs[i] = InnerSum;
3134 // Add J as its own register, or an unfolded immediate.
3135 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
3136 if (TLI && SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
3137 TLI->isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3138 SC->getValue()->getZExtValue()))
3139 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
3140 SC->getValue()->getZExtValue();
3142 F.BaseRegs.push_back(*J);
3144 if (InsertFormula(LU, LUIdx, F))
3145 // If that formula hadn't been seen before, recurse to find more like
3147 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1);
3152 /// GenerateCombinations - Generate a formula consisting of all of the
3153 /// loop-dominating registers added into a single register.
3154 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
3156 // This method is only interesting on a plurality of registers.
3157 if (Base.BaseRegs.size() <= 1) return;
3161 SmallVector<const SCEV *, 4> Ops;
3162 for (SmallVectorImpl<const SCEV *>::const_iterator
3163 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
3164 const SCEV *BaseReg = *I;
3165 if (SE.properlyDominates(BaseReg, L->getHeader()) &&
3166 !SE.hasComputableLoopEvolution(BaseReg, L))
3167 Ops.push_back(BaseReg);
3169 F.BaseRegs.push_back(BaseReg);
3171 if (Ops.size() > 1) {
3172 const SCEV *Sum = SE.getAddExpr(Ops);
3173 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
3174 // opportunity to fold something. For now, just ignore such cases
3175 // rather than proceed with zero in a register.
3176 if (!Sum->isZero()) {
3177 F.BaseRegs.push_back(Sum);
3178 (void)InsertFormula(LU, LUIdx, F);
3183 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
3184 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
3186 // We can't add a symbolic offset if the address already contains one.
3187 if (Base.AM.BaseGV) return;
3189 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3190 const SCEV *G = Base.BaseRegs[i];
3191 GlobalValue *GV = ExtractSymbol(G, SE);
3192 if (G->isZero() || !GV)
3196 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
3197 LU.Kind, LU.AccessTy, TLI))
3200 (void)InsertFormula(LU, LUIdx, F);
3204 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
3205 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
3207 // TODO: For now, just add the min and max offset, because it usually isn't
3208 // worthwhile looking at everything inbetween.
3209 SmallVector<int64_t, 2> Worklist;
3210 Worklist.push_back(LU.MinOffset);
3211 if (LU.MaxOffset != LU.MinOffset)
3212 Worklist.push_back(LU.MaxOffset);
3214 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3215 const SCEV *G = Base.BaseRegs[i];
3217 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
3218 E = Worklist.end(); I != E; ++I) {
3220 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs - *I;
3221 if (isLegalUse(F.AM, LU.MinOffset - *I, LU.MaxOffset - *I,
3222 LU.Kind, LU.AccessTy, TLI)) {
3223 // Add the offset to the base register.
3224 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), *I), G);
3225 // If it cancelled out, drop the base register, otherwise update it.
3226 if (NewG->isZero()) {
3227 std::swap(F.BaseRegs[i], F.BaseRegs.back());
3228 F.BaseRegs.pop_back();
3230 F.BaseRegs[i] = NewG;
3232 (void)InsertFormula(LU, LUIdx, F);
3236 int64_t Imm = ExtractImmediate(G, SE);
3237 if (G->isZero() || Imm == 0)
3240 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Imm;
3241 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
3242 LU.Kind, LU.AccessTy, TLI))
3245 (void)InsertFormula(LU, LUIdx, F);
3249 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
3250 /// the comparison. For example, x == y -> x*c == y*c.
3251 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
3253 if (LU.Kind != LSRUse::ICmpZero) return;
3255 // Determine the integer type for the base formula.
3256 Type *IntTy = Base.getType();
3258 if (SE.getTypeSizeInBits(IntTy) > 64) return;
3260 // Don't do this if there is more than one offset.
3261 if (LU.MinOffset != LU.MaxOffset) return;
3263 assert(!Base.AM.BaseGV && "ICmpZero use is not legal!");
3265 // Check each interesting stride.
3266 for (SmallSetVector<int64_t, 8>::const_iterator
3267 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3268 int64_t Factor = *I;
3270 // Check that the multiplication doesn't overflow.
3271 if (Base.AM.BaseOffs == INT64_MIN && Factor == -1)
3273 int64_t NewBaseOffs = (uint64_t)Base.AM.BaseOffs * Factor;
3274 if (NewBaseOffs / Factor != Base.AM.BaseOffs)
3277 // Check that multiplying with the use offset doesn't overflow.
3278 int64_t Offset = LU.MinOffset;
3279 if (Offset == INT64_MIN && Factor == -1)
3281 Offset = (uint64_t)Offset * Factor;
3282 if (Offset / Factor != LU.MinOffset)
3286 F.AM.BaseOffs = NewBaseOffs;
3288 // Check that this scale is legal.
3289 if (!isLegalUse(F.AM, Offset, Offset, LU.Kind, LU.AccessTy, TLI))
3292 // Compensate for the use having MinOffset built into it.
3293 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Offset - LU.MinOffset;
3295 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3297 // Check that multiplying with each base register doesn't overflow.
3298 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
3299 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
3300 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
3304 // Check that multiplying with the scaled register doesn't overflow.
3306 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
3307 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
3311 // Check that multiplying with the unfolded offset doesn't overflow.
3312 if (F.UnfoldedOffset != 0) {
3313 if (F.UnfoldedOffset == INT64_MIN && Factor == -1)
3315 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
3316 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
3320 // If we make it here and it's legal, add it.
3321 (void)InsertFormula(LU, LUIdx, F);
3326 /// GenerateScales - Generate stride factor reuse formulae by making use of
3327 /// scaled-offset address modes, for example.
3328 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
3329 // Determine the integer type for the base formula.
3330 Type *IntTy = Base.getType();
3333 // If this Formula already has a scaled register, we can't add another one.
3334 if (Base.AM.Scale != 0) return;
3336 // Check each interesting stride.
3337 for (SmallSetVector<int64_t, 8>::const_iterator
3338 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3339 int64_t Factor = *I;
3341 Base.AM.Scale = Factor;
3342 Base.AM.HasBaseReg = Base.BaseRegs.size() > 1;
3343 // Check whether this scale is going to be legal.
3344 if (!isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
3345 LU.Kind, LU.AccessTy, TLI)) {
3346 // As a special-case, handle special out-of-loop Basic users specially.
3347 // TODO: Reconsider this special case.
3348 if (LU.Kind == LSRUse::Basic &&
3349 isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
3350 LSRUse::Special, LU.AccessTy, TLI) &&
3351 LU.AllFixupsOutsideLoop)
3352 LU.Kind = LSRUse::Special;
3356 // For an ICmpZero, negating a solitary base register won't lead to
3358 if (LU.Kind == LSRUse::ICmpZero &&
3359 !Base.AM.HasBaseReg && Base.AM.BaseOffs == 0 && !Base.AM.BaseGV)
3361 // For each addrec base reg, apply the scale, if possible.
3362 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3363 if (const SCEVAddRecExpr *AR =
3364 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
3365 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3366 if (FactorS->isZero())
3368 // Divide out the factor, ignoring high bits, since we'll be
3369 // scaling the value back up in the end.
3370 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
3371 // TODO: This could be optimized to avoid all the copying.
3373 F.ScaledReg = Quotient;
3374 F.DeleteBaseReg(F.BaseRegs[i]);
3375 (void)InsertFormula(LU, LUIdx, F);
3381 /// GenerateTruncates - Generate reuse formulae from different IV types.
3382 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
3383 // This requires TargetLowering to tell us which truncates are free.
3386 // Don't bother truncating symbolic values.
3387 if (Base.AM.BaseGV) return;
3389 // Determine the integer type for the base formula.
3390 Type *DstTy = Base.getType();
3392 DstTy = SE.getEffectiveSCEVType(DstTy);
3394 for (SmallSetVector<Type *, 4>::const_iterator
3395 I = Types.begin(), E = Types.end(); I != E; ++I) {
3397 if (SrcTy != DstTy && TLI->isTruncateFree(SrcTy, DstTy)) {
3400 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
3401 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
3402 JE = F.BaseRegs.end(); J != JE; ++J)
3403 *J = SE.getAnyExtendExpr(*J, SrcTy);
3405 // TODO: This assumes we've done basic processing on all uses and
3406 // have an idea what the register usage is.
3407 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
3410 (void)InsertFormula(LU, LUIdx, F);
3417 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
3418 /// defer modifications so that the search phase doesn't have to worry about
3419 /// the data structures moving underneath it.
3423 const SCEV *OrigReg;
3425 WorkItem(size_t LI, int64_t I, const SCEV *R)
3426 : LUIdx(LI), Imm(I), OrigReg(R) {}
3428 void print(raw_ostream &OS) const;
3434 void WorkItem::print(raw_ostream &OS) const {
3435 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
3436 << " , add offset " << Imm;
3439 void WorkItem::dump() const {
3440 print(errs()); errs() << '\n';
3443 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
3444 /// distance apart and try to form reuse opportunities between them.
3445 void LSRInstance::GenerateCrossUseConstantOffsets() {
3446 // Group the registers by their value without any added constant offset.
3447 typedef std::map<int64_t, const SCEV *> ImmMapTy;
3448 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
3450 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
3451 SmallVector<const SCEV *, 8> Sequence;
3452 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
3454 const SCEV *Reg = *I;
3455 int64_t Imm = ExtractImmediate(Reg, SE);
3456 std::pair<RegMapTy::iterator, bool> Pair =
3457 Map.insert(std::make_pair(Reg, ImmMapTy()));
3459 Sequence.push_back(Reg);
3460 Pair.first->second.insert(std::make_pair(Imm, *I));
3461 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
3464 // Now examine each set of registers with the same base value. Build up
3465 // a list of work to do and do the work in a separate step so that we're
3466 // not adding formulae and register counts while we're searching.
3467 SmallVector<WorkItem, 32> WorkItems;
3468 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
3469 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
3470 E = Sequence.end(); I != E; ++I) {
3471 const SCEV *Reg = *I;
3472 const ImmMapTy &Imms = Map.find(Reg)->second;
3474 // It's not worthwhile looking for reuse if there's only one offset.
3475 if (Imms.size() == 1)
3478 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
3479 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3481 dbgs() << ' ' << J->first;
3484 // Examine each offset.
3485 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3487 const SCEV *OrigReg = J->second;
3489 int64_t JImm = J->first;
3490 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
3492 if (!isa<SCEVConstant>(OrigReg) &&
3493 UsedByIndicesMap[Reg].count() == 1) {
3494 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
3498 // Conservatively examine offsets between this orig reg a few selected
3500 ImmMapTy::const_iterator OtherImms[] = {
3501 Imms.begin(), prior(Imms.end()),
3502 Imms.lower_bound((Imms.begin()->first + prior(Imms.end())->first) / 2)
3504 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
3505 ImmMapTy::const_iterator M = OtherImms[i];
3506 if (M == J || M == JE) continue;
3508 // Compute the difference between the two.
3509 int64_t Imm = (uint64_t)JImm - M->first;
3510 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
3511 LUIdx = UsedByIndices.find_next(LUIdx))
3512 // Make a memo of this use, offset, and register tuple.
3513 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
3514 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
3521 UsedByIndicesMap.clear();
3522 UniqueItems.clear();
3524 // Now iterate through the worklist and add new formulae.
3525 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
3526 E = WorkItems.end(); I != E; ++I) {
3527 const WorkItem &WI = *I;
3528 size_t LUIdx = WI.LUIdx;
3529 LSRUse &LU = Uses[LUIdx];
3530 int64_t Imm = WI.Imm;
3531 const SCEV *OrigReg = WI.OrigReg;
3533 Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
3534 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
3535 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
3537 // TODO: Use a more targeted data structure.
3538 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
3539 const Formula &F = LU.Formulae[L];
3540 // Use the immediate in the scaled register.
3541 if (F.ScaledReg == OrigReg) {
3542 int64_t Offs = (uint64_t)F.AM.BaseOffs +
3543 Imm * (uint64_t)F.AM.Scale;
3544 // Don't create 50 + reg(-50).
3545 if (F.referencesReg(SE.getSCEV(
3546 ConstantInt::get(IntTy, -(uint64_t)Offs))))
3549 NewF.AM.BaseOffs = Offs;
3550 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
3551 LU.Kind, LU.AccessTy, TLI))
3553 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
3555 // If the new scale is a constant in a register, and adding the constant
3556 // value to the immediate would produce a value closer to zero than the
3557 // immediate itself, then the formula isn't worthwhile.
3558 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
3559 if (C->getValue()->isNegative() !=
3560 (NewF.AM.BaseOffs < 0) &&
3561 (C->getValue()->getValue().abs() * APInt(BitWidth, F.AM.Scale))
3562 .ule(abs64(NewF.AM.BaseOffs)))
3566 (void)InsertFormula(LU, LUIdx, NewF);
3568 // Use the immediate in a base register.
3569 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
3570 const SCEV *BaseReg = F.BaseRegs[N];
3571 if (BaseReg != OrigReg)
3574 NewF.AM.BaseOffs = (uint64_t)NewF.AM.BaseOffs + Imm;
3575 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
3576 LU.Kind, LU.AccessTy, TLI)) {
3578 !TLI->isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
3581 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
3583 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
3585 // If the new formula has a constant in a register, and adding the
3586 // constant value to the immediate would produce a value closer to
3587 // zero than the immediate itself, then the formula isn't worthwhile.
3588 for (SmallVectorImpl<const SCEV *>::const_iterator
3589 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
3591 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
3592 if ((C->getValue()->getValue() + NewF.AM.BaseOffs).abs().slt(
3593 abs64(NewF.AM.BaseOffs)) &&
3594 (C->getValue()->getValue() +
3595 NewF.AM.BaseOffs).countTrailingZeros() >=
3596 CountTrailingZeros_64(NewF.AM.BaseOffs))
3600 (void)InsertFormula(LU, LUIdx, NewF);
3609 /// GenerateAllReuseFormulae - Generate formulae for each use.
3611 LSRInstance::GenerateAllReuseFormulae() {
3612 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
3613 // queries are more precise.
3614 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3615 LSRUse &LU = Uses[LUIdx];
3616 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3617 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
3618 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3619 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
3621 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3622 LSRUse &LU = Uses[LUIdx];
3623 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3624 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
3625 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3626 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
3627 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3628 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
3629 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3630 GenerateScales(LU, LUIdx, LU.Formulae[i]);
3632 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3633 LSRUse &LU = Uses[LUIdx];
3634 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3635 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
3638 GenerateCrossUseConstantOffsets();
3640 DEBUG(dbgs() << "\n"
3641 "After generating reuse formulae:\n";
3642 print_uses(dbgs()));
3645 /// If there are multiple formulae with the same set of registers used
3646 /// by other uses, pick the best one and delete the others.
3647 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
3648 DenseSet<const SCEV *> VisitedRegs;
3649 SmallPtrSet<const SCEV *, 16> Regs;
3650 SmallPtrSet<const SCEV *, 16> LoserRegs;
3652 bool ChangedFormulae = false;
3655 // Collect the best formula for each unique set of shared registers. This
3656 // is reset for each use.
3657 typedef DenseMap<SmallVector<const SCEV *, 2>, size_t, UniquifierDenseMapInfo>
3659 BestFormulaeTy BestFormulae;
3661 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3662 LSRUse &LU = Uses[LUIdx];
3663 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
3666 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
3667 FIdx != NumForms; ++FIdx) {
3668 Formula &F = LU.Formulae[FIdx];
3670 // Some formulas are instant losers. For example, they may depend on
3671 // nonexistent AddRecs from other loops. These need to be filtered
3672 // immediately, otherwise heuristics could choose them over others leading
3673 // to an unsatisfactory solution. Passing LoserRegs into RateFormula here
3674 // avoids the need to recompute this information across formulae using the
3675 // same bad AddRec. Passing LoserRegs is also essential unless we remove
3676 // the corresponding bad register from the Regs set.
3679 CostF.RateFormula(F, Regs, VisitedRegs, L, LU.Offsets, SE, DT,
3681 if (CostF.isLoser()) {
3682 // During initial formula generation, undesirable formulae are generated
3683 // by uses within other loops that have some non-trivial address mode or
3684 // use the postinc form of the IV. LSR needs to provide these formulae
3685 // as the basis of rediscovering the desired formula that uses an AddRec
3686 // corresponding to the existing phi. Once all formulae have been
3687 // generated, these initial losers may be pruned.
3688 DEBUG(dbgs() << " Filtering loser "; F.print(dbgs());
3692 SmallVector<const SCEV *, 2> Key;
3693 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
3694 JE = F.BaseRegs.end(); J != JE; ++J) {
3695 const SCEV *Reg = *J;
3696 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
3700 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
3701 Key.push_back(F.ScaledReg);
3702 // Unstable sort by host order ok, because this is only used for
3704 std::sort(Key.begin(), Key.end());
3706 std::pair<BestFormulaeTy::const_iterator, bool> P =
3707 BestFormulae.insert(std::make_pair(Key, FIdx));
3711 Formula &Best = LU.Formulae[P.first->second];
3715 CostBest.RateFormula(Best, Regs, VisitedRegs, L, LU.Offsets, SE, DT);
3716 if (CostF < CostBest)
3718 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
3720 " in favor of formula "; Best.print(dbgs());
3724 ChangedFormulae = true;
3726 LU.DeleteFormula(F);
3732 // Now that we've filtered out some formulae, recompute the Regs set.
3734 LU.RecomputeRegs(LUIdx, RegUses);
3736 // Reset this to prepare for the next use.
3737 BestFormulae.clear();
3740 DEBUG(if (ChangedFormulae) {
3742 "After filtering out undesirable candidates:\n";
3747 // This is a rough guess that seems to work fairly well.
3748 static const size_t ComplexityLimit = UINT16_MAX;
3750 /// EstimateSearchSpaceComplexity - Estimate the worst-case number of
3751 /// solutions the solver might have to consider. It almost never considers
3752 /// this many solutions because it prune the search space, but the pruning
3753 /// isn't always sufficient.
3754 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
3756 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3757 E = Uses.end(); I != E; ++I) {
3758 size_t FSize = I->Formulae.size();
3759 if (FSize >= ComplexityLimit) {
3760 Power = ComplexityLimit;
3764 if (Power >= ComplexityLimit)
3770 /// NarrowSearchSpaceByDetectingSupersets - When one formula uses a superset
3771 /// of the registers of another formula, it won't help reduce register
3772 /// pressure (though it may not necessarily hurt register pressure); remove
3773 /// it to simplify the system.
3774 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
3775 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3776 DEBUG(dbgs() << "The search space is too complex.\n");
3778 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
3779 "which use a superset of registers used by other "
3782 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3783 LSRUse &LU = Uses[LUIdx];
3785 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3786 Formula &F = LU.Formulae[i];
3787 // Look for a formula with a constant or GV in a register. If the use
3788 // also has a formula with that same value in an immediate field,
3789 // delete the one that uses a register.
3790 for (SmallVectorImpl<const SCEV *>::const_iterator
3791 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
3792 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
3794 NewF.AM.BaseOffs += C->getValue()->getSExtValue();
3795 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
3796 (I - F.BaseRegs.begin()));
3797 if (LU.HasFormulaWithSameRegs(NewF)) {
3798 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3799 LU.DeleteFormula(F);
3805 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
3806 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
3809 NewF.AM.BaseGV = GV;
3810 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
3811 (I - F.BaseRegs.begin()));
3812 if (LU.HasFormulaWithSameRegs(NewF)) {
3813 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3815 LU.DeleteFormula(F);
3826 LU.RecomputeRegs(LUIdx, RegUses);
3829 DEBUG(dbgs() << "After pre-selection:\n";
3830 print_uses(dbgs()));
3834 /// NarrowSearchSpaceByCollapsingUnrolledCode - When there are many registers
3835 /// for expressions like A, A+1, A+2, etc., allocate a single register for
3837 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
3838 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3839 DEBUG(dbgs() << "The search space is too complex.\n");
3841 DEBUG(dbgs() << "Narrowing the search space by assuming that uses "
3842 "separated by a constant offset will use the same "
3845 // This is especially useful for unrolled loops.
3847 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3848 LSRUse &LU = Uses[LUIdx];
3849 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3850 E = LU.Formulae.end(); I != E; ++I) {
3851 const Formula &F = *I;
3852 if (F.AM.BaseOffs != 0 && F.AM.Scale == 0) {
3853 if (LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU)) {
3854 if (reconcileNewOffset(*LUThatHas, F.AM.BaseOffs,
3855 /*HasBaseReg=*/false,
3856 LU.Kind, LU.AccessTy)) {
3857 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs());
3860 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
3862 // Update the relocs to reference the new use.
3863 for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(),
3864 E = Fixups.end(); I != E; ++I) {
3865 LSRFixup &Fixup = *I;
3866 if (Fixup.LUIdx == LUIdx) {
3867 Fixup.LUIdx = LUThatHas - &Uses.front();
3868 Fixup.Offset += F.AM.BaseOffs;
3869 // Add the new offset to LUThatHas' offset list.
3870 if (LUThatHas->Offsets.back() != Fixup.Offset) {
3871 LUThatHas->Offsets.push_back(Fixup.Offset);
3872 if (Fixup.Offset > LUThatHas->MaxOffset)
3873 LUThatHas->MaxOffset = Fixup.Offset;
3874 if (Fixup.Offset < LUThatHas->MinOffset)
3875 LUThatHas->MinOffset = Fixup.Offset;
3877 DEBUG(dbgs() << "New fixup has offset "
3878 << Fixup.Offset << '\n');
3880 if (Fixup.LUIdx == NumUses-1)
3881 Fixup.LUIdx = LUIdx;
3884 // Delete formulae from the new use which are no longer legal.
3886 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
3887 Formula &F = LUThatHas->Formulae[i];
3888 if (!isLegalUse(F.AM,
3889 LUThatHas->MinOffset, LUThatHas->MaxOffset,
3890 LUThatHas->Kind, LUThatHas->AccessTy, TLI)) {
3891 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3893 LUThatHas->DeleteFormula(F);
3900 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
3902 // Delete the old use.
3903 DeleteUse(LU, LUIdx);
3913 DEBUG(dbgs() << "After pre-selection:\n";
3914 print_uses(dbgs()));
3918 /// NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters - Call
3919 /// FilterOutUndesirableDedicatedRegisters again, if necessary, now that
3920 /// we've done more filtering, as it may be able to find more formulae to
3922 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
3923 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3924 DEBUG(dbgs() << "The search space is too complex.\n");
3926 DEBUG(dbgs() << "Narrowing the search space by re-filtering out "
3927 "undesirable dedicated registers.\n");
3929 FilterOutUndesirableDedicatedRegisters();
3931 DEBUG(dbgs() << "After pre-selection:\n";
3932 print_uses(dbgs()));
3936 /// NarrowSearchSpaceByPickingWinnerRegs - Pick a register which seems likely
3937 /// to be profitable, and then in any use which has any reference to that
3938 /// register, delete all formulae which do not reference that register.
3939 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
3940 // With all other options exhausted, loop until the system is simple
3941 // enough to handle.
3942 SmallPtrSet<const SCEV *, 4> Taken;
3943 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3944 // Ok, we have too many of formulae on our hands to conveniently handle.
3945 // Use a rough heuristic to thin out the list.
3946 DEBUG(dbgs() << "The search space is too complex.\n");
3948 // Pick the register which is used by the most LSRUses, which is likely
3949 // to be a good reuse register candidate.
3950 const SCEV *Best = 0;
3951 unsigned BestNum = 0;
3952 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
3954 const SCEV *Reg = *I;
3955 if (Taken.count(Reg))
3960 unsigned Count = RegUses.getUsedByIndices(Reg).count();
3961 if (Count > BestNum) {
3968 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
3969 << " will yield profitable reuse.\n");
3972 // In any use with formulae which references this register, delete formulae
3973 // which don't reference it.
3974 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3975 LSRUse &LU = Uses[LUIdx];
3976 if (!LU.Regs.count(Best)) continue;
3979 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3980 Formula &F = LU.Formulae[i];
3981 if (!F.referencesReg(Best)) {
3982 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3983 LU.DeleteFormula(F);
3987 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
3993 LU.RecomputeRegs(LUIdx, RegUses);
3996 DEBUG(dbgs() << "After pre-selection:\n";
3997 print_uses(dbgs()));
4001 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
4002 /// formulae to choose from, use some rough heuristics to prune down the number
4003 /// of formulae. This keeps the main solver from taking an extraordinary amount
4004 /// of time in some worst-case scenarios.
4005 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
4006 NarrowSearchSpaceByDetectingSupersets();
4007 NarrowSearchSpaceByCollapsingUnrolledCode();
4008 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
4009 NarrowSearchSpaceByPickingWinnerRegs();
4012 /// SolveRecurse - This is the recursive solver.
4013 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
4015 SmallVectorImpl<const Formula *> &Workspace,
4016 const Cost &CurCost,
4017 const SmallPtrSet<const SCEV *, 16> &CurRegs,
4018 DenseSet<const SCEV *> &VisitedRegs) const {
4021 // - use more aggressive filtering
4022 // - sort the formula so that the most profitable solutions are found first
4023 // - sort the uses too
4025 // - don't compute a cost, and then compare. compare while computing a cost
4027 // - track register sets with SmallBitVector
4029 const LSRUse &LU = Uses[Workspace.size()];
4031 // If this use references any register that's already a part of the
4032 // in-progress solution, consider it a requirement that a formula must
4033 // reference that register in order to be considered. This prunes out
4034 // unprofitable searching.
4035 SmallSetVector<const SCEV *, 4> ReqRegs;
4036 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(),
4037 E = CurRegs.end(); I != E; ++I)
4038 if (LU.Regs.count(*I))
4041 SmallPtrSet<const SCEV *, 16> NewRegs;
4043 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
4044 E = LU.Formulae.end(); I != E; ++I) {
4045 const Formula &F = *I;
4047 // Ignore formulae which do not use any of the required registers.
4048 bool SatisfiedReqReg = true;
4049 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
4050 JE = ReqRegs.end(); J != JE; ++J) {
4051 const SCEV *Reg = *J;
4052 if ((!F.ScaledReg || F.ScaledReg != Reg) &&
4053 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) ==
4055 SatisfiedReqReg = false;
4059 if (!SatisfiedReqReg) {
4060 // If none of the formulae satisfied the required registers, then we could
4061 // clear ReqRegs and try again. Currently, we simply give up in this case.
4065 // Evaluate the cost of the current formula. If it's already worse than
4066 // the current best, prune the search at that point.
4069 NewCost.RateFormula(F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT);
4070 if (NewCost < SolutionCost) {
4071 Workspace.push_back(&F);
4072 if (Workspace.size() != Uses.size()) {
4073 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
4074 NewRegs, VisitedRegs);
4075 if (F.getNumRegs() == 1 && Workspace.size() == 1)
4076 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
4078 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
4079 dbgs() << ".\n Regs:";
4080 for (SmallPtrSet<const SCEV *, 16>::const_iterator
4081 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I)
4082 dbgs() << ' ' << **I;
4085 SolutionCost = NewCost;
4086 Solution = Workspace;
4088 Workspace.pop_back();
4093 /// Solve - Choose one formula from each use. Return the results in the given
4094 /// Solution vector.
4095 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
4096 SmallVector<const Formula *, 8> Workspace;
4098 SolutionCost.Loose();
4100 SmallPtrSet<const SCEV *, 16> CurRegs;
4101 DenseSet<const SCEV *> VisitedRegs;
4102 Workspace.reserve(Uses.size());
4104 // SolveRecurse does all the work.
4105 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
4106 CurRegs, VisitedRegs);
4107 if (Solution.empty()) {
4108 DEBUG(dbgs() << "\nNo Satisfactory Solution\n");
4112 // Ok, we've now made all our decisions.
4113 DEBUG(dbgs() << "\n"
4114 "The chosen solution requires "; SolutionCost.print(dbgs());
4116 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
4118 Uses[i].print(dbgs());
4121 Solution[i]->print(dbgs());
4125 assert(Solution.size() == Uses.size() && "Malformed solution!");
4128 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
4129 /// the dominator tree far as we can go while still being dominated by the
4130 /// input positions. This helps canonicalize the insert position, which
4131 /// encourages sharing.
4132 BasicBlock::iterator
4133 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
4134 const SmallVectorImpl<Instruction *> &Inputs)
4137 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
4138 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
4141 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
4142 if (!Rung) return IP;
4143 Rung = Rung->getIDom();
4144 if (!Rung) return IP;
4145 IDom = Rung->getBlock();
4147 // Don't climb into a loop though.
4148 const Loop *IDomLoop = LI.getLoopFor(IDom);
4149 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
4150 if (IDomDepth <= IPLoopDepth &&
4151 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
4155 bool AllDominate = true;
4156 Instruction *BetterPos = 0;
4157 Instruction *Tentative = IDom->getTerminator();
4158 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
4159 E = Inputs.end(); I != E; ++I) {
4160 Instruction *Inst = *I;
4161 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
4162 AllDominate = false;
4165 // Attempt to find an insert position in the middle of the block,
4166 // instead of at the end, so that it can be used for other expansions.
4167 if (IDom == Inst->getParent() &&
4168 (!BetterPos || !DT.dominates(Inst, BetterPos)))
4169 BetterPos = llvm::next(BasicBlock::iterator(Inst));
4182 /// AdjustInsertPositionForExpand - Determine an input position which will be
4183 /// dominated by the operands and which will dominate the result.
4184 BasicBlock::iterator
4185 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator LowestIP,
4188 SCEVExpander &Rewriter) const {
4189 // Collect some instructions which must be dominated by the
4190 // expanding replacement. These must be dominated by any operands that
4191 // will be required in the expansion.
4192 SmallVector<Instruction *, 4> Inputs;
4193 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
4194 Inputs.push_back(I);
4195 if (LU.Kind == LSRUse::ICmpZero)
4196 if (Instruction *I =
4197 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
4198 Inputs.push_back(I);
4199 if (LF.PostIncLoops.count(L)) {
4200 if (LF.isUseFullyOutsideLoop(L))
4201 Inputs.push_back(L->getLoopLatch()->getTerminator());
4203 Inputs.push_back(IVIncInsertPos);
4205 // The expansion must also be dominated by the increment positions of any
4206 // loops it for which it is using post-inc mode.
4207 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
4208 E = LF.PostIncLoops.end(); I != E; ++I) {
4209 const Loop *PIL = *I;
4210 if (PIL == L) continue;
4212 // Be dominated by the loop exit.
4213 SmallVector<BasicBlock *, 4> ExitingBlocks;
4214 PIL->getExitingBlocks(ExitingBlocks);
4215 if (!ExitingBlocks.empty()) {
4216 BasicBlock *BB = ExitingBlocks[0];
4217 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
4218 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
4219 Inputs.push_back(BB->getTerminator());
4223 assert(!isa<PHINode>(LowestIP) && !isa<LandingPadInst>(LowestIP)
4224 && !isa<DbgInfoIntrinsic>(LowestIP) &&
4225 "Insertion point must be a normal instruction");
4227 // Then, climb up the immediate dominator tree as far as we can go while
4228 // still being dominated by the input positions.
4229 BasicBlock::iterator IP = HoistInsertPosition(LowestIP, Inputs);
4231 // Don't insert instructions before PHI nodes.
4232 while (isa<PHINode>(IP)) ++IP;
4234 // Ignore landingpad instructions.
4235 while (isa<LandingPadInst>(IP)) ++IP;
4237 // Ignore debug intrinsics.
4238 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
4240 // Set IP below instructions recently inserted by SCEVExpander. This keeps the
4241 // IP consistent across expansions and allows the previously inserted
4242 // instructions to be reused by subsequent expansion.
4243 while (Rewriter.isInsertedInstruction(IP) && IP != LowestIP) ++IP;
4248 /// Expand - Emit instructions for the leading candidate expression for this
4249 /// LSRUse (this is called "expanding").
4250 Value *LSRInstance::Expand(const LSRFixup &LF,
4252 BasicBlock::iterator IP,
4253 SCEVExpander &Rewriter,
4254 SmallVectorImpl<WeakVH> &DeadInsts) const {
4255 const LSRUse &LU = Uses[LF.LUIdx];
4257 // Determine an input position which will be dominated by the operands and
4258 // which will dominate the result.
4259 IP = AdjustInsertPositionForExpand(IP, LF, LU, Rewriter);
4261 // Inform the Rewriter if we have a post-increment use, so that it can
4262 // perform an advantageous expansion.
4263 Rewriter.setPostInc(LF.PostIncLoops);
4265 // This is the type that the user actually needs.
4266 Type *OpTy = LF.OperandValToReplace->getType();
4267 // This will be the type that we'll initially expand to.
4268 Type *Ty = F.getType();
4270 // No type known; just expand directly to the ultimate type.
4272 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
4273 // Expand directly to the ultimate type if it's the right size.
4275 // This is the type to do integer arithmetic in.
4276 Type *IntTy = SE.getEffectiveSCEVType(Ty);
4278 // Build up a list of operands to add together to form the full base.
4279 SmallVector<const SCEV *, 8> Ops;
4281 // Expand the BaseRegs portion.
4282 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
4283 E = F.BaseRegs.end(); I != E; ++I) {
4284 const SCEV *Reg = *I;
4285 assert(!Reg->isZero() && "Zero allocated in a base register!");
4287 // If we're expanding for a post-inc user, make the post-inc adjustment.
4288 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4289 Reg = TransformForPostIncUse(Denormalize, Reg,
4290 LF.UserInst, LF.OperandValToReplace,
4293 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP)));
4296 // Expand the ScaledReg portion.
4297 Value *ICmpScaledV = 0;
4298 if (F.AM.Scale != 0) {
4299 const SCEV *ScaledS = F.ScaledReg;
4301 // If we're expanding for a post-inc user, make the post-inc adjustment.
4302 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4303 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
4304 LF.UserInst, LF.OperandValToReplace,
4307 if (LU.Kind == LSRUse::ICmpZero) {
4308 // An interesting way of "folding" with an icmp is to use a negated
4309 // scale, which we'll implement by inserting it into the other operand
4311 assert(F.AM.Scale == -1 &&
4312 "The only scale supported by ICmpZero uses is -1!");
4313 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP);
4315 // Otherwise just expand the scaled register and an explicit scale,
4316 // which is expected to be matched as part of the address.
4318 // Flush the operand list to suppress SCEVExpander hoisting address modes.
4319 if (!Ops.empty() && LU.Kind == LSRUse::Address) {
4320 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4322 Ops.push_back(SE.getUnknown(FullV));
4324 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP));
4325 ScaledS = SE.getMulExpr(ScaledS,
4326 SE.getConstant(ScaledS->getType(), F.AM.Scale));
4327 Ops.push_back(ScaledS);
4331 // Expand the GV portion.
4333 // Flush the operand list to suppress SCEVExpander hoisting.
4335 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4337 Ops.push_back(SE.getUnknown(FullV));
4339 Ops.push_back(SE.getUnknown(F.AM.BaseGV));
4342 // Flush the operand list to suppress SCEVExpander hoisting of both folded and
4343 // unfolded offsets. LSR assumes they both live next to their uses.
4345 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4347 Ops.push_back(SE.getUnknown(FullV));
4350 // Expand the immediate portion.
4351 int64_t Offset = (uint64_t)F.AM.BaseOffs + LF.Offset;
4353 if (LU.Kind == LSRUse::ICmpZero) {
4354 // The other interesting way of "folding" with an ICmpZero is to use a
4355 // negated immediate.
4357 ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset);
4359 Ops.push_back(SE.getUnknown(ICmpScaledV));
4360 ICmpScaledV = ConstantInt::get(IntTy, Offset);
4363 // Just add the immediate values. These again are expected to be matched
4364 // as part of the address.
4365 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
4369 // Expand the unfolded offset portion.
4370 int64_t UnfoldedOffset = F.UnfoldedOffset;
4371 if (UnfoldedOffset != 0) {
4372 // Just add the immediate values.
4373 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy,
4377 // Emit instructions summing all the operands.
4378 const SCEV *FullS = Ops.empty() ?
4379 SE.getConstant(IntTy, 0) :
4381 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
4383 // We're done expanding now, so reset the rewriter.
4384 Rewriter.clearPostInc();
4386 // An ICmpZero Formula represents an ICmp which we're handling as a
4387 // comparison against zero. Now that we've expanded an expression for that
4388 // form, update the ICmp's other operand.
4389 if (LU.Kind == LSRUse::ICmpZero) {
4390 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
4391 DeadInsts.push_back(CI->getOperand(1));
4392 assert(!F.AM.BaseGV && "ICmp does not support folding a global value and "
4393 "a scale at the same time!");
4394 if (F.AM.Scale == -1) {
4395 if (ICmpScaledV->getType() != OpTy) {
4397 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
4399 ICmpScaledV, OpTy, "tmp", CI);
4402 CI->setOperand(1, ICmpScaledV);
4404 assert(F.AM.Scale == 0 &&
4405 "ICmp does not support folding a global value and "
4406 "a scale at the same time!");
4407 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
4409 if (C->getType() != OpTy)
4410 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
4414 CI->setOperand(1, C);
4421 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use
4422 /// of their operands effectively happens in their predecessor blocks, so the
4423 /// expression may need to be expanded in multiple places.
4424 void LSRInstance::RewriteForPHI(PHINode *PN,
4427 SCEVExpander &Rewriter,
4428 SmallVectorImpl<WeakVH> &DeadInsts,
4430 DenseMap<BasicBlock *, Value *> Inserted;
4431 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
4432 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
4433 BasicBlock *BB = PN->getIncomingBlock(i);
4435 // If this is a critical edge, split the edge so that we do not insert
4436 // the code on all predecessor/successor paths. We do this unless this
4437 // is the canonical backedge for this loop, which complicates post-inc
4439 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
4440 !isa<IndirectBrInst>(BB->getTerminator())) {
4441 BasicBlock *Parent = PN->getParent();
4442 Loop *PNLoop = LI.getLoopFor(Parent);
4443 if (!PNLoop || Parent != PNLoop->getHeader()) {
4444 // Split the critical edge.
4445 BasicBlock *NewBB = 0;
4446 if (!Parent->isLandingPad()) {
4447 NewBB = SplitCriticalEdge(BB, Parent, P,
4448 /*MergeIdenticalEdges=*/true,
4449 /*DontDeleteUselessPhis=*/true);
4451 SmallVector<BasicBlock*, 2> NewBBs;
4452 SplitLandingPadPredecessors(Parent, BB, "", "", P, NewBBs);
4456 // If PN is outside of the loop and BB is in the loop, we want to
4457 // move the block to be immediately before the PHI block, not
4458 // immediately after BB.
4459 if (L->contains(BB) && !L->contains(PN))
4460 NewBB->moveBefore(PN->getParent());
4462 // Splitting the edge can reduce the number of PHI entries we have.
4463 e = PN->getNumIncomingValues();
4465 i = PN->getBasicBlockIndex(BB);
4469 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
4470 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0)));
4472 PN->setIncomingValue(i, Pair.first->second);
4474 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts);
4476 // If this is reuse-by-noop-cast, insert the noop cast.
4477 Type *OpTy = LF.OperandValToReplace->getType();
4478 if (FullV->getType() != OpTy)
4480 CastInst::Create(CastInst::getCastOpcode(FullV, false,
4482 FullV, LF.OperandValToReplace->getType(),
4483 "tmp", BB->getTerminator());
4485 PN->setIncomingValue(i, FullV);
4486 Pair.first->second = FullV;
4491 /// Rewrite - Emit instructions for the leading candidate expression for this
4492 /// LSRUse (this is called "expanding"), and update the UserInst to reference
4493 /// the newly expanded value.
4494 void LSRInstance::Rewrite(const LSRFixup &LF,
4496 SCEVExpander &Rewriter,
4497 SmallVectorImpl<WeakVH> &DeadInsts,
4499 // First, find an insertion point that dominates UserInst. For PHI nodes,
4500 // find the nearest block which dominates all the relevant uses.
4501 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
4502 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P);
4504 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts);
4506 // If this is reuse-by-noop-cast, insert the noop cast.
4507 Type *OpTy = LF.OperandValToReplace->getType();
4508 if (FullV->getType() != OpTy) {
4510 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
4511 FullV, OpTy, "tmp", LF.UserInst);
4515 // Update the user. ICmpZero is handled specially here (for now) because
4516 // Expand may have updated one of the operands of the icmp already, and
4517 // its new value may happen to be equal to LF.OperandValToReplace, in
4518 // which case doing replaceUsesOfWith leads to replacing both operands
4519 // with the same value. TODO: Reorganize this.
4520 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
4521 LF.UserInst->setOperand(0, FullV);
4523 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
4526 DeadInsts.push_back(LF.OperandValToReplace);
4529 /// ImplementSolution - Rewrite all the fixup locations with new values,
4530 /// following the chosen solution.
4532 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
4534 // Keep track of instructions we may have made dead, so that
4535 // we can remove them after we are done working.
4536 SmallVector<WeakVH, 16> DeadInsts;
4538 SCEVExpander Rewriter(SE, "lsr");
4540 Rewriter.setDebugType(DEBUG_TYPE);
4542 Rewriter.disableCanonicalMode();
4543 Rewriter.enableLSRMode();
4544 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
4546 // Mark phi nodes that terminate chains so the expander tries to reuse them.
4547 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(),
4548 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) {
4549 if (PHINode *PN = dyn_cast<PHINode>(ChainI->tailUserInst()))
4550 Rewriter.setChainedPhi(PN);
4553 // Expand the new value definitions and update the users.
4554 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
4555 E = Fixups.end(); I != E; ++I) {
4556 const LSRFixup &Fixup = *I;
4558 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P);
4563 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(),
4564 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) {
4565 GenerateIVChain(*ChainI, Rewriter, DeadInsts);
4568 // Clean up after ourselves. This must be done before deleting any
4572 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
4575 LSRInstance::LSRInstance(const TargetLowering *tli, Loop *l, Pass *P)
4576 : IU(P->getAnalysis<IVUsers>()),
4577 SE(P->getAnalysis<ScalarEvolution>()),
4578 DT(P->getAnalysis<DominatorTree>()),
4579 LI(P->getAnalysis<LoopInfo>()),
4580 TLI(tli), L(l), Changed(false), IVIncInsertPos(0) {
4582 // If LoopSimplify form is not available, stay out of trouble.
4583 if (!L->isLoopSimplifyForm())
4586 // If there's no interesting work to be done, bail early.
4587 if (IU.empty()) return;
4589 // If there's too much analysis to be done, bail early. We won't be able to
4590 // model the problem anyway.
4591 unsigned NumUsers = 0;
4592 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
4593 if (++NumUsers > MaxIVUsers) {
4594 DEBUG(dbgs() << "LSR skipping loop, too many IV Users in " << *L
4601 // All dominating loops must have preheaders, or SCEVExpander may not be able
4602 // to materialize an AddRecExpr whose Start is an outer AddRecExpr.
4604 // IVUsers analysis should only create users that are dominated by simple loop
4605 // headers. Since this loop should dominate all of its users, its user list
4606 // should be empty if this loop itself is not within a simple loop nest.
4607 for (DomTreeNode *Rung = DT.getNode(L->getLoopPreheader());
4608 Rung; Rung = Rung->getIDom()) {
4609 BasicBlock *BB = Rung->getBlock();
4610 const Loop *DomLoop = LI.getLoopFor(BB);
4611 if (DomLoop && DomLoop->getHeader() == BB) {
4612 assert(DomLoop->getLoopPreheader() && "LSR needs a simplified loop nest");
4617 DEBUG(dbgs() << "\nLSR on loop ";
4618 WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false);
4621 // First, perform some low-level loop optimizations.
4623 OptimizeLoopTermCond();
4625 // If loop preparation eliminates all interesting IV users, bail.
4626 if (IU.empty()) return;
4628 // Skip nested loops until we can model them better with formulae.
4630 DEBUG(dbgs() << "LSR skipping outer loop " << *L << "\n");
4634 // Start collecting data and preparing for the solver.
4636 CollectInterestingTypesAndFactors();
4637 CollectFixupsAndInitialFormulae();
4638 CollectLoopInvariantFixupsAndFormulae();
4640 assert(!Uses.empty() && "IVUsers reported at least one use");
4641 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
4642 print_uses(dbgs()));
4644 // Now use the reuse data to generate a bunch of interesting ways
4645 // to formulate the values needed for the uses.
4646 GenerateAllReuseFormulae();
4648 FilterOutUndesirableDedicatedRegisters();
4649 NarrowSearchSpaceUsingHeuristics();
4651 SmallVector<const Formula *, 8> Solution;
4654 // Release memory that is no longer needed.
4659 if (Solution.empty())
4663 // Formulae should be legal.
4664 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
4665 E = Uses.end(); I != E; ++I) {
4666 const LSRUse &LU = *I;
4667 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
4668 JE = LU.Formulae.end(); J != JE; ++J)
4669 assert(isLegalUse(J->AM, LU.MinOffset, LU.MaxOffset,
4670 LU.Kind, LU.AccessTy, TLI) &&
4671 "Illegal formula generated!");
4675 // Now that we've decided what we want, make it so.
4676 ImplementSolution(Solution, P);
4679 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
4680 if (Factors.empty() && Types.empty()) return;
4682 OS << "LSR has identified the following interesting factors and types: ";
4685 for (SmallSetVector<int64_t, 8>::const_iterator
4686 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
4687 if (!First) OS << ", ";
4692 for (SmallSetVector<Type *, 4>::const_iterator
4693 I = Types.begin(), E = Types.end(); I != E; ++I) {
4694 if (!First) OS << ", ";
4696 OS << '(' << **I << ')';
4701 void LSRInstance::print_fixups(raw_ostream &OS) const {
4702 OS << "LSR is examining the following fixup sites:\n";
4703 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
4704 E = Fixups.end(); I != E; ++I) {
4711 void LSRInstance::print_uses(raw_ostream &OS) const {
4712 OS << "LSR is examining the following uses:\n";
4713 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
4714 E = Uses.end(); I != E; ++I) {
4715 const LSRUse &LU = *I;
4719 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
4720 JE = LU.Formulae.end(); J != JE; ++J) {
4728 void LSRInstance::print(raw_ostream &OS) const {
4729 print_factors_and_types(OS);
4734 void LSRInstance::dump() const {
4735 print(errs()); errs() << '\n';
4740 class LoopStrengthReduce : public LoopPass {
4741 /// TLI - Keep a pointer of a TargetLowering to consult for determining
4742 /// transformation profitability.
4743 const TargetLowering *const TLI;
4746 static char ID; // Pass ID, replacement for typeid
4747 explicit LoopStrengthReduce(const TargetLowering *tli = 0);
4750 bool runOnLoop(Loop *L, LPPassManager &LPM);
4751 void getAnalysisUsage(AnalysisUsage &AU) const;
4756 char LoopStrengthReduce::ID = 0;
4757 INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce",
4758 "Loop Strength Reduction", false, false)
4759 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
4760 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
4761 INITIALIZE_PASS_DEPENDENCY(IVUsers)
4762 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
4763 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
4764 INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce",
4765 "Loop Strength Reduction", false, false)
4768 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
4769 return new LoopStrengthReduce(TLI);
4772 LoopStrengthReduce::LoopStrengthReduce(const TargetLowering *tli)
4773 : LoopPass(ID), TLI(tli) {
4774 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry());
4777 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
4778 // We split critical edges, so we change the CFG. However, we do update
4779 // many analyses if they are around.
4780 AU.addPreservedID(LoopSimplifyID);
4782 AU.addRequired<LoopInfo>();
4783 AU.addPreserved<LoopInfo>();
4784 AU.addRequiredID(LoopSimplifyID);
4785 AU.addRequired<DominatorTree>();
4786 AU.addPreserved<DominatorTree>();
4787 AU.addRequired<ScalarEvolution>();
4788 AU.addPreserved<ScalarEvolution>();
4789 // Requiring LoopSimplify a second time here prevents IVUsers from running
4790 // twice, since LoopSimplify was invalidated by running ScalarEvolution.
4791 AU.addRequiredID(LoopSimplifyID);
4792 AU.addRequired<IVUsers>();
4793 AU.addPreserved<IVUsers>();
4796 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
4797 bool Changed = false;
4799 // Run the main LSR transformation.
4800 Changed |= LSRInstance(TLI, L, this).getChanged();
4802 // Remove any extra phis created by processing inner loops.
4803 Changed |= DeleteDeadPHIs(L->getHeader());
4804 if (EnablePhiElim) {
4805 SmallVector<WeakVH, 16> DeadInsts;
4806 SCEVExpander Rewriter(getAnalysis<ScalarEvolution>(), "lsr");
4808 Rewriter.setDebugType(DEBUG_TYPE);
4810 unsigned numFolded = Rewriter.
4811 replaceCongruentIVs(L, &getAnalysis<DominatorTree>(), DeadInsts, TLI);
4814 DeleteTriviallyDeadInstructions(DeadInsts);
4815 DeleteDeadPHIs(L->getHeader());