1 //===- LoopStrengthReduce.cpp - Strength Reduce IVs in Loops --------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This transformation analyzes and transforms the induction variables (and
11 // computations derived from them) into forms suitable for efficient execution
14 // This pass performs a strength reduction on array references inside loops that
15 // have as one or more of their components the loop induction variable, it
16 // rewrites expressions to take advantage of scaled-index addressing modes
17 // available on the target, and it performs a variety of other optimizations
18 // related to loop induction variables.
20 // Terminology note: this code has a lot of handling for "post-increment" or
21 // "post-inc" users. This is not talking about post-increment addressing modes;
22 // it is instead talking about code like this:
24 // %i = phi [ 0, %entry ], [ %i.next, %latch ]
26 // %i.next = add %i, 1
27 // %c = icmp eq %i.next, %n
29 // The SCEV for %i is {0,+,1}<%L>. The SCEV for %i.next is {1,+,1}<%L>, however
30 // it's useful to think about these as the same register, with some uses using
31 // the value of the register before the add and some using // it after. In this
32 // example, the icmp is a post-increment user, since it uses %i.next, which is
33 // the value of the induction variable after the increment. The other common
34 // case of post-increment users is users outside the loop.
36 // TODO: More sophistication in the way Formulae are generated and filtered.
38 // TODO: Handle multiple loops at a time.
40 // TODO: Should the addressing mode BaseGV be changed to a ConstantExpr instead
43 // TODO: When truncation is free, truncate ICmp users' operands to make it a
44 // smaller encoding (on x86 at least).
46 // TODO: When a negated register is used by an add (such as in a list of
47 // multiple base registers, or as the increment expression in an addrec),
48 // we may not actually need both reg and (-1 * reg) in registers; the
49 // negation can be implemented by using a sub instead of an add. The
50 // lack of support for taking this into consideration when making
51 // register pressure decisions is partly worked around by the "Special"
54 //===----------------------------------------------------------------------===//
56 #define DEBUG_TYPE "loop-reduce"
57 #include "llvm/Transforms/Scalar.h"
58 #include "llvm/ADT/DenseSet.h"
59 #include "llvm/ADT/SetVector.h"
60 #include "llvm/ADT/SmallBitVector.h"
61 #include "llvm/ADT/STLExtras.h"
62 #include "llvm/Analysis/Dominators.h"
63 #include "llvm/Analysis/IVUsers.h"
64 #include "llvm/Analysis/LoopPass.h"
65 #include "llvm/Analysis/ScalarEvolutionExpander.h"
66 #include "llvm/Analysis/TargetTransformInfo.h"
67 #include "llvm/Assembly/Writer.h"
68 #include "llvm/IR/Constants.h"
69 #include "llvm/IR/DerivedTypes.h"
70 #include "llvm/IR/Instructions.h"
71 #include "llvm/IR/IntrinsicInst.h"
72 #include "llvm/Support/CommandLine.h"
73 #include "llvm/Support/Debug.h"
74 #include "llvm/Support/ValueHandle.h"
75 #include "llvm/Support/raw_ostream.h"
76 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
77 #include "llvm/Transforms/Utils/Local.h"
81 /// MaxIVUsers is an arbitrary threshold that provides an early opportunitiy for
82 /// bail out. This threshold is far beyond the number of users that LSR can
83 /// conceivably solve, so it should not affect generated code, but catches the
84 /// worst cases before LSR burns too much compile time and stack space.
85 static const unsigned MaxIVUsers = 200;
87 // Temporary flag to cleanup congruent phis after LSR phi expansion.
88 // It's currently disabled until we can determine whether it's truly useful or
89 // not. The flag should be removed after the v3.0 release.
90 // This is now needed for ivchains.
91 static cl::opt<bool> EnablePhiElim(
92 "enable-lsr-phielim", cl::Hidden, cl::init(true),
93 cl::desc("Enable LSR phi elimination"));
96 // Stress test IV chain generation.
97 static cl::opt<bool> StressIVChain(
98 "stress-ivchain", cl::Hidden, cl::init(false),
99 cl::desc("Stress test LSR IV chains"));
101 static bool StressIVChain = false;
106 /// RegSortData - This class holds data which is used to order reuse candidates.
109 /// UsedByIndices - This represents the set of LSRUse indices which reference
110 /// a particular register.
111 SmallBitVector UsedByIndices;
115 void print(raw_ostream &OS) const;
121 void RegSortData::print(raw_ostream &OS) const {
122 OS << "[NumUses=" << UsedByIndices.count() << ']';
125 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
126 void RegSortData::dump() const {
127 print(errs()); errs() << '\n';
133 /// RegUseTracker - Map register candidates to information about how they are
135 class RegUseTracker {
136 typedef DenseMap<const SCEV *, RegSortData> RegUsesTy;
138 RegUsesTy RegUsesMap;
139 SmallVector<const SCEV *, 16> RegSequence;
142 void CountRegister(const SCEV *Reg, size_t LUIdx);
143 void DropRegister(const SCEV *Reg, size_t LUIdx);
144 void SwapAndDropUse(size_t LUIdx, size_t LastLUIdx);
146 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
148 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
152 typedef SmallVectorImpl<const SCEV *>::iterator iterator;
153 typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator;
154 iterator begin() { return RegSequence.begin(); }
155 iterator end() { return RegSequence.end(); }
156 const_iterator begin() const { return RegSequence.begin(); }
157 const_iterator end() const { return RegSequence.end(); }
163 RegUseTracker::CountRegister(const SCEV *Reg, size_t LUIdx) {
164 std::pair<RegUsesTy::iterator, bool> Pair =
165 RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
166 RegSortData &RSD = Pair.first->second;
168 RegSequence.push_back(Reg);
169 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
170 RSD.UsedByIndices.set(LUIdx);
174 RegUseTracker::DropRegister(const SCEV *Reg, size_t LUIdx) {
175 RegUsesTy::iterator It = RegUsesMap.find(Reg);
176 assert(It != RegUsesMap.end());
177 RegSortData &RSD = It->second;
178 assert(RSD.UsedByIndices.size() > LUIdx);
179 RSD.UsedByIndices.reset(LUIdx);
183 RegUseTracker::SwapAndDropUse(size_t LUIdx, size_t LastLUIdx) {
184 assert(LUIdx <= LastLUIdx);
186 // Update RegUses. The data structure is not optimized for this purpose;
187 // we must iterate through it and update each of the bit vectors.
188 for (RegUsesTy::iterator I = RegUsesMap.begin(), E = RegUsesMap.end();
190 SmallBitVector &UsedByIndices = I->second.UsedByIndices;
191 if (LUIdx < UsedByIndices.size())
192 UsedByIndices[LUIdx] =
193 LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : 0;
194 UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx));
199 RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
200 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
201 if (I == RegUsesMap.end())
203 const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
204 int i = UsedByIndices.find_first();
205 if (i == -1) return false;
206 if ((size_t)i != LUIdx) return true;
207 return UsedByIndices.find_next(i) != -1;
210 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
211 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
212 assert(I != RegUsesMap.end() && "Unknown register!");
213 return I->second.UsedByIndices;
216 void RegUseTracker::clear() {
223 /// Formula - This class holds information that describes a formula for
224 /// computing satisfying a use. It may include broken-out immediates and scaled
227 /// Global base address used for complex addressing.
230 /// Base offset for complex addressing.
233 /// Whether any complex addressing has a base register.
236 /// The scale of any complex addressing.
239 /// BaseRegs - The list of "base" registers for this use. When this is
241 SmallVector<const SCEV *, 4> BaseRegs;
243 /// ScaledReg - The 'scaled' register for this use. This should be non-null
244 /// when Scale is not zero.
245 const SCEV *ScaledReg;
247 /// UnfoldedOffset - An additional constant offset which added near the
248 /// use. This requires a temporary register, but the offset itself can
249 /// live in an add immediate field rather than a register.
250 int64_t UnfoldedOffset;
253 : BaseGV(0), BaseOffset(0), HasBaseReg(false), Scale(0), ScaledReg(0),
256 void InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE);
258 unsigned getNumRegs() const;
259 Type *getType() const;
261 void DeleteBaseReg(const SCEV *&S);
263 bool referencesReg(const SCEV *S) const;
264 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
265 const RegUseTracker &RegUses) const;
267 void print(raw_ostream &OS) const;
273 /// DoInitialMatch - Recursion helper for InitialMatch.
274 static void DoInitialMatch(const SCEV *S, Loop *L,
275 SmallVectorImpl<const SCEV *> &Good,
276 SmallVectorImpl<const SCEV *> &Bad,
277 ScalarEvolution &SE) {
278 // Collect expressions which properly dominate the loop header.
279 if (SE.properlyDominates(S, L->getHeader())) {
284 // Look at add operands.
285 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
286 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
288 DoInitialMatch(*I, L, Good, Bad, SE);
292 // Look at addrec operands.
293 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
294 if (!AR->getStart()->isZero()) {
295 DoInitialMatch(AR->getStart(), L, Good, Bad, SE);
296 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
297 AR->getStepRecurrence(SE),
298 // FIXME: AR->getNoWrapFlags()
299 AR->getLoop(), SCEV::FlagAnyWrap),
304 // Handle a multiplication by -1 (negation) if it didn't fold.
305 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
306 if (Mul->getOperand(0)->isAllOnesValue()) {
307 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
308 const SCEV *NewMul = SE.getMulExpr(Ops);
310 SmallVector<const SCEV *, 4> MyGood;
311 SmallVector<const SCEV *, 4> MyBad;
312 DoInitialMatch(NewMul, L, MyGood, MyBad, SE);
313 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
314 SE.getEffectiveSCEVType(NewMul->getType())));
315 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyGood.begin(),
316 E = MyGood.end(); I != E; ++I)
317 Good.push_back(SE.getMulExpr(NegOne, *I));
318 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyBad.begin(),
319 E = MyBad.end(); I != E; ++I)
320 Bad.push_back(SE.getMulExpr(NegOne, *I));
324 // Ok, we can't do anything interesting. Just stuff the whole thing into a
325 // register and hope for the best.
329 /// InitialMatch - Incorporate loop-variant parts of S into this Formula,
330 /// attempting to keep all loop-invariant and loop-computable values in a
331 /// single base register.
332 void Formula::InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) {
333 SmallVector<const SCEV *, 4> Good;
334 SmallVector<const SCEV *, 4> Bad;
335 DoInitialMatch(S, L, Good, Bad, SE);
337 const SCEV *Sum = SE.getAddExpr(Good);
339 BaseRegs.push_back(Sum);
343 const SCEV *Sum = SE.getAddExpr(Bad);
345 BaseRegs.push_back(Sum);
350 /// getNumRegs - Return the total number of register operands used by this
351 /// formula. This does not include register uses implied by non-constant
353 unsigned Formula::getNumRegs() const {
354 return !!ScaledReg + BaseRegs.size();
357 /// getType - Return the type of this formula, if it has one, or null
358 /// otherwise. This type is meaningless except for the bit size.
359 Type *Formula::getType() const {
360 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
361 ScaledReg ? ScaledReg->getType() :
362 BaseGV ? BaseGV->getType() :
366 /// DeleteBaseReg - Delete the given base reg from the BaseRegs list.
367 void Formula::DeleteBaseReg(const SCEV *&S) {
368 if (&S != &BaseRegs.back())
369 std::swap(S, BaseRegs.back());
373 /// referencesReg - Test if this formula references the given register.
374 bool Formula::referencesReg(const SCEV *S) const {
375 return S == ScaledReg ||
376 std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end();
379 /// hasRegsUsedByUsesOtherThan - Test whether this formula uses registers
380 /// which are used by uses other than the use with the given index.
381 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
382 const RegUseTracker &RegUses) const {
384 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
386 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
387 E = BaseRegs.end(); I != E; ++I)
388 if (RegUses.isRegUsedByUsesOtherThan(*I, LUIdx))
393 void Formula::print(raw_ostream &OS) const {
396 if (!First) OS << " + "; else First = false;
397 WriteAsOperand(OS, BaseGV, /*PrintType=*/false);
399 if (BaseOffset != 0) {
400 if (!First) OS << " + "; else First = false;
403 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
404 E = BaseRegs.end(); I != E; ++I) {
405 if (!First) OS << " + "; else First = false;
406 OS << "reg(" << **I << ')';
408 if (HasBaseReg && BaseRegs.empty()) {
409 if (!First) OS << " + "; else First = false;
410 OS << "**error: HasBaseReg**";
411 } else if (!HasBaseReg && !BaseRegs.empty()) {
412 if (!First) OS << " + "; else First = false;
413 OS << "**error: !HasBaseReg**";
416 if (!First) OS << " + "; else First = false;
417 OS << Scale << "*reg(";
424 if (UnfoldedOffset != 0) {
425 if (!First) OS << " + "; else First = false;
426 OS << "imm(" << UnfoldedOffset << ')';
430 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
431 void Formula::dump() const {
432 print(errs()); errs() << '\n';
436 /// isAddRecSExtable - Return true if the given addrec can be sign-extended
437 /// without changing its value.
438 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
440 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1);
441 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
444 /// isAddSExtable - Return true if the given add can be sign-extended
445 /// without changing its value.
446 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
448 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
449 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
452 /// isMulSExtable - Return true if the given mul can be sign-extended
453 /// without changing its value.
454 static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
456 IntegerType::get(SE.getContext(),
457 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
458 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
461 /// getExactSDiv - Return an expression for LHS /s RHS, if it can be determined
462 /// and if the remainder is known to be zero, or null otherwise. If
463 /// IgnoreSignificantBits is true, expressions like (X * Y) /s Y are simplified
464 /// to Y, ignoring that the multiplication may overflow, which is useful when
465 /// the result will be used in a context where the most significant bits are
467 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
469 bool IgnoreSignificantBits = false) {
470 // Handle the trivial case, which works for any SCEV type.
472 return SE.getConstant(LHS->getType(), 1);
474 // Handle a few RHS special cases.
475 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
477 const APInt &RA = RC->getValue()->getValue();
478 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
480 if (RA.isAllOnesValue())
481 return SE.getMulExpr(LHS, RC);
482 // Handle x /s 1 as x.
487 // Check for a division of a constant by a constant.
488 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
491 const APInt &LA = C->getValue()->getValue();
492 const APInt &RA = RC->getValue()->getValue();
493 if (LA.srem(RA) != 0)
495 return SE.getConstant(LA.sdiv(RA));
498 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
499 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
500 if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) {
501 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
502 IgnoreSignificantBits);
504 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
505 IgnoreSignificantBits);
506 if (!Start) return 0;
507 // FlagNW is independent of the start value, step direction, and is
508 // preserved with smaller magnitude steps.
509 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
510 return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap);
515 // Distribute the sdiv over add operands, if the add doesn't overflow.
516 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
517 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
518 SmallVector<const SCEV *, 8> Ops;
519 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
521 const SCEV *Op = getExactSDiv(*I, RHS, SE,
522 IgnoreSignificantBits);
526 return SE.getAddExpr(Ops);
531 // Check for a multiply operand that we can pull RHS out of.
532 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
533 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
534 SmallVector<const SCEV *, 4> Ops;
536 for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end();
540 if (const SCEV *Q = getExactSDiv(S, RHS, SE,
541 IgnoreSignificantBits)) {
547 return Found ? SE.getMulExpr(Ops) : 0;
552 // Otherwise we don't know.
556 /// ExtractImmediate - If S involves the addition of a constant integer value,
557 /// return that integer value, and mutate S to point to a new SCEV with that
559 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
560 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
561 if (C->getValue()->getValue().getMinSignedBits() <= 64) {
562 S = SE.getConstant(C->getType(), 0);
563 return C->getValue()->getSExtValue();
565 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
566 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
567 int64_t Result = ExtractImmediate(NewOps.front(), SE);
569 S = SE.getAddExpr(NewOps);
571 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
572 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
573 int64_t Result = ExtractImmediate(NewOps.front(), SE);
575 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
576 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
583 /// ExtractSymbol - If S involves the addition of a GlobalValue address,
584 /// return that symbol, and mutate S to point to a new SCEV with that
586 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
587 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
588 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
589 S = SE.getConstant(GV->getType(), 0);
592 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
593 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
594 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
596 S = SE.getAddExpr(NewOps);
598 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
599 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
600 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
602 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
603 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
610 /// isAddressUse - Returns true if the specified instruction is using the
611 /// specified value as an address.
612 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
613 bool isAddress = isa<LoadInst>(Inst);
614 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
615 if (SI->getOperand(1) == OperandVal)
617 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
618 // Addressing modes can also be folded into prefetches and a variety
620 switch (II->getIntrinsicID()) {
622 case Intrinsic::prefetch:
623 case Intrinsic::x86_sse_storeu_ps:
624 case Intrinsic::x86_sse2_storeu_pd:
625 case Intrinsic::x86_sse2_storeu_dq:
626 case Intrinsic::x86_sse2_storel_dq:
627 if (II->getArgOperand(0) == OperandVal)
635 /// getAccessType - Return the type of the memory being accessed.
636 static Type *getAccessType(const Instruction *Inst) {
637 Type *AccessTy = Inst->getType();
638 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
639 AccessTy = SI->getOperand(0)->getType();
640 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
641 // Addressing modes can also be folded into prefetches and a variety
643 switch (II->getIntrinsicID()) {
645 case Intrinsic::x86_sse_storeu_ps:
646 case Intrinsic::x86_sse2_storeu_pd:
647 case Intrinsic::x86_sse2_storeu_dq:
648 case Intrinsic::x86_sse2_storel_dq:
649 AccessTy = II->getArgOperand(0)->getType();
654 // All pointers have the same requirements, so canonicalize them to an
655 // arbitrary pointer type to minimize variation.
656 if (PointerType *PTy = dyn_cast<PointerType>(AccessTy))
657 AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
658 PTy->getAddressSpace());
663 /// isExistingPhi - Return true if this AddRec is already a phi in its loop.
664 static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
665 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
666 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
667 if (SE.isSCEVable(PN->getType()) &&
668 (SE.getEffectiveSCEVType(PN->getType()) ==
669 SE.getEffectiveSCEVType(AR->getType())) &&
670 SE.getSCEV(PN) == AR)
676 /// Check if expanding this expression is likely to incur significant cost. This
677 /// is tricky because SCEV doesn't track which expressions are actually computed
678 /// by the current IR.
680 /// We currently allow expansion of IV increments that involve adds,
681 /// multiplication by constants, and AddRecs from existing phis.
683 /// TODO: Allow UDivExpr if we can find an existing IV increment that is an
684 /// obvious multiple of the UDivExpr.
685 static bool isHighCostExpansion(const SCEV *S,
686 SmallPtrSet<const SCEV*, 8> &Processed,
687 ScalarEvolution &SE) {
688 // Zero/One operand expressions
689 switch (S->getSCEVType()) {
694 return isHighCostExpansion(cast<SCEVTruncateExpr>(S)->getOperand(),
697 return isHighCostExpansion(cast<SCEVZeroExtendExpr>(S)->getOperand(),
700 return isHighCostExpansion(cast<SCEVSignExtendExpr>(S)->getOperand(),
704 if (!Processed.insert(S))
707 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
708 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
710 if (isHighCostExpansion(*I, Processed, SE))
716 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
717 if (Mul->getNumOperands() == 2) {
718 // Multiplication by a constant is ok
719 if (isa<SCEVConstant>(Mul->getOperand(0)))
720 return isHighCostExpansion(Mul->getOperand(1), Processed, SE);
722 // If we have the value of one operand, check if an existing
723 // multiplication already generates this expression.
724 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Mul->getOperand(1))) {
725 Value *UVal = U->getValue();
726 for (Value::use_iterator UI = UVal->use_begin(), UE = UVal->use_end();
728 // If U is a constant, it may be used by a ConstantExpr.
729 Instruction *User = dyn_cast<Instruction>(*UI);
730 if (User && User->getOpcode() == Instruction::Mul
731 && SE.isSCEVable(User->getType())) {
732 return SE.getSCEV(User) == Mul;
739 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
740 if (isExistingPhi(AR, SE))
744 // Fow now, consider any other type of expression (div/mul/min/max) high cost.
748 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
749 /// specified set are trivially dead, delete them and see if this makes any of
750 /// their operands subsequently dead.
752 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) {
753 bool Changed = false;
755 while (!DeadInsts.empty()) {
756 Value *V = DeadInsts.pop_back_val();
757 Instruction *I = dyn_cast_or_null<Instruction>(V);
759 if (I == 0 || !isInstructionTriviallyDead(I))
762 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
763 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
766 DeadInsts.push_back(U);
769 I->eraseFromParent();
779 // Check if it is legal to fold 2 base registers.
780 static bool isLegal2RegAMUse(const TargetTransformInfo &TTI, const LSRUse &LU,
782 // Get the cost of the scaling factor used in F for LU.
783 static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
784 const LSRUse &LU, const Formula &F);
788 /// Cost - This class is used to measure and compare candidate formulae.
790 /// TODO: Some of these could be merged. Also, a lexical ordering
791 /// isn't always optimal.
795 unsigned NumBaseAdds;
802 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0),
803 SetupCost(0), ScaleCost(0) {}
805 bool operator<(const Cost &Other) const;
810 // Once any of the metrics loses, they must all remain losers.
812 return ((NumRegs | AddRecCost | NumIVMuls | NumBaseAdds
813 | ImmCost | SetupCost | ScaleCost) != ~0u)
814 || ((NumRegs & AddRecCost & NumIVMuls & NumBaseAdds
815 & ImmCost & SetupCost & ScaleCost) == ~0u);
820 assert(isValid() && "invalid cost");
821 return NumRegs == ~0u;
824 void RateFormula(const TargetTransformInfo &TTI,
826 SmallPtrSet<const SCEV *, 16> &Regs,
827 const DenseSet<const SCEV *> &VisitedRegs,
829 const SmallVectorImpl<int64_t> &Offsets,
830 ScalarEvolution &SE, DominatorTree &DT,
832 SmallPtrSet<const SCEV *, 16> *LoserRegs = 0);
834 void print(raw_ostream &OS) const;
838 void RateRegister(const SCEV *Reg,
839 SmallPtrSet<const SCEV *, 16> &Regs,
841 ScalarEvolution &SE, DominatorTree &DT);
842 void RatePrimaryRegister(const SCEV *Reg,
843 SmallPtrSet<const SCEV *, 16> &Regs,
845 ScalarEvolution &SE, DominatorTree &DT,
846 SmallPtrSet<const SCEV *, 16> *LoserRegs);
851 /// RateRegister - Tally up interesting quantities from the given register.
852 void Cost::RateRegister(const SCEV *Reg,
853 SmallPtrSet<const SCEV *, 16> &Regs,
855 ScalarEvolution &SE, DominatorTree &DT) {
856 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
857 // If this is an addrec for another loop, don't second-guess its addrec phi
858 // nodes. LSR isn't currently smart enough to reason about more than one
859 // loop at a time. LSR has already run on inner loops, will not run on outer
860 // loops, and cannot be expected to change sibling loops.
861 if (AR->getLoop() != L) {
862 // If the AddRec exists, consider it's register free and leave it alone.
863 if (isExistingPhi(AR, SE))
866 // Otherwise, do not consider this formula at all.
870 AddRecCost += 1; /// TODO: This should be a function of the stride.
872 // Add the step value register, if it needs one.
873 // TODO: The non-affine case isn't precisely modeled here.
874 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
875 if (!Regs.count(AR->getOperand(1))) {
876 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
884 // Rough heuristic; favor registers which don't require extra setup
885 // instructions in the preheader.
886 if (!isa<SCEVUnknown>(Reg) &&
887 !isa<SCEVConstant>(Reg) &&
888 !(isa<SCEVAddRecExpr>(Reg) &&
889 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
890 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
893 NumIVMuls += isa<SCEVMulExpr>(Reg) &&
894 SE.hasComputableLoopEvolution(Reg, L);
897 /// RatePrimaryRegister - Record this register in the set. If we haven't seen it
898 /// before, rate it. Optional LoserRegs provides a way to declare any formula
899 /// that refers to one of those regs an instant loser.
900 void Cost::RatePrimaryRegister(const SCEV *Reg,
901 SmallPtrSet<const SCEV *, 16> &Regs,
903 ScalarEvolution &SE, DominatorTree &DT,
904 SmallPtrSet<const SCEV *, 16> *LoserRegs) {
905 if (LoserRegs && LoserRegs->count(Reg)) {
909 if (Regs.insert(Reg)) {
910 RateRegister(Reg, Regs, L, SE, DT);
911 if (LoserRegs && isLoser())
912 LoserRegs->insert(Reg);
916 void Cost::RateFormula(const TargetTransformInfo &TTI,
918 SmallPtrSet<const SCEV *, 16> &Regs,
919 const DenseSet<const SCEV *> &VisitedRegs,
921 const SmallVectorImpl<int64_t> &Offsets,
922 ScalarEvolution &SE, DominatorTree &DT,
924 SmallPtrSet<const SCEV *, 16> *LoserRegs) {
925 // Tally up the registers.
926 if (const SCEV *ScaledReg = F.ScaledReg) {
927 if (VisitedRegs.count(ScaledReg)) {
931 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT, LoserRegs);
935 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
936 E = F.BaseRegs.end(); I != E; ++I) {
937 const SCEV *BaseReg = *I;
938 if (VisitedRegs.count(BaseReg)) {
942 RatePrimaryRegister(BaseReg, Regs, L, SE, DT, LoserRegs);
947 // Determine how many (unfolded) adds we'll need inside the loop.
948 size_t NumBaseParts = F.BaseRegs.size() + (F.UnfoldedOffset != 0);
949 if (NumBaseParts > 1)
950 // Do not count the base and a possible second register if the target
951 // allows to fold 2 registers.
952 NumBaseAdds += NumBaseParts - (1 + isLegal2RegAMUse(TTI, LU, F));
954 // Accumulate non-free scaling amounts.
955 ScaleCost += getScalingFactorCost(TTI, LU, F);
957 // Tally up the non-zero immediates.
958 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
959 E = Offsets.end(); I != E; ++I) {
960 int64_t Offset = (uint64_t)*I + F.BaseOffset;
962 ImmCost += 64; // Handle symbolic values conservatively.
963 // TODO: This should probably be the pointer size.
964 else if (Offset != 0)
965 ImmCost += APInt(64, Offset, true).getMinSignedBits();
967 assert(isValid() && "invalid cost");
970 /// Loose - Set this cost to a losing value.
981 /// operator< - Choose the lower cost.
982 bool Cost::operator<(const Cost &Other) const {
983 if (NumRegs != Other.NumRegs)
984 return NumRegs < Other.NumRegs;
985 if (AddRecCost != Other.AddRecCost)
986 return AddRecCost < Other.AddRecCost;
987 if (NumIVMuls != Other.NumIVMuls)
988 return NumIVMuls < Other.NumIVMuls;
989 if (NumBaseAdds != Other.NumBaseAdds)
990 return NumBaseAdds < Other.NumBaseAdds;
991 if (ScaleCost != Other.ScaleCost)
992 return ScaleCost < Other.ScaleCost;
993 if (ImmCost != Other.ImmCost)
994 return ImmCost < Other.ImmCost;
995 if (SetupCost != Other.SetupCost)
996 return SetupCost < Other.SetupCost;
1000 void Cost::print(raw_ostream &OS) const {
1001 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
1002 if (AddRecCost != 0)
1003 OS << ", with addrec cost " << AddRecCost;
1005 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
1006 if (NumBaseAdds != 0)
1007 OS << ", plus " << NumBaseAdds << " base add"
1008 << (NumBaseAdds == 1 ? "" : "s");
1010 OS << ", plus " << ScaleCost << " scale cost";
1012 OS << ", plus " << ImmCost << " imm cost";
1014 OS << ", plus " << SetupCost << " setup cost";
1017 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1018 void Cost::dump() const {
1019 print(errs()); errs() << '\n';
1025 /// LSRFixup - An operand value in an instruction which is to be replaced
1026 /// with some equivalent, possibly strength-reduced, replacement.
1028 /// UserInst - The instruction which will be updated.
1029 Instruction *UserInst;
1031 /// OperandValToReplace - The operand of the instruction which will
1032 /// be replaced. The operand may be used more than once; every instance
1033 /// will be replaced.
1034 Value *OperandValToReplace;
1036 /// PostIncLoops - If this user is to use the post-incremented value of an
1037 /// induction variable, this variable is non-null and holds the loop
1038 /// associated with the induction variable.
1039 PostIncLoopSet PostIncLoops;
1041 /// LUIdx - The index of the LSRUse describing the expression which
1042 /// this fixup needs, minus an offset (below).
1045 /// Offset - A constant offset to be added to the LSRUse expression.
1046 /// This allows multiple fixups to share the same LSRUse with different
1047 /// offsets, for example in an unrolled loop.
1050 bool isUseFullyOutsideLoop(const Loop *L) const;
1054 void print(raw_ostream &OS) const;
1060 LSRFixup::LSRFixup()
1061 : UserInst(0), OperandValToReplace(0), LUIdx(~size_t(0)), Offset(0) {}
1063 /// isUseFullyOutsideLoop - Test whether this fixup always uses its
1064 /// value outside of the given loop.
1065 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
1066 // PHI nodes use their value in their incoming blocks.
1067 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
1068 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1069 if (PN->getIncomingValue(i) == OperandValToReplace &&
1070 L->contains(PN->getIncomingBlock(i)))
1075 return !L->contains(UserInst);
1078 void LSRFixup::print(raw_ostream &OS) const {
1080 // Store is common and interesting enough to be worth special-casing.
1081 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
1083 WriteAsOperand(OS, Store->getOperand(0), /*PrintType=*/false);
1084 } else if (UserInst->getType()->isVoidTy())
1085 OS << UserInst->getOpcodeName();
1087 WriteAsOperand(OS, UserInst, /*PrintType=*/false);
1089 OS << ", OperandValToReplace=";
1090 WriteAsOperand(OS, OperandValToReplace, /*PrintType=*/false);
1092 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(),
1093 E = PostIncLoops.end(); I != E; ++I) {
1094 OS << ", PostIncLoop=";
1095 WriteAsOperand(OS, (*I)->getHeader(), /*PrintType=*/false);
1098 if (LUIdx != ~size_t(0))
1099 OS << ", LUIdx=" << LUIdx;
1102 OS << ", Offset=" << Offset;
1105 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1106 void LSRFixup::dump() const {
1107 print(errs()); errs() << '\n';
1113 /// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding
1114 /// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*.
1115 struct UniquifierDenseMapInfo {
1116 static SmallVector<const SCEV *, 4> getEmptyKey() {
1117 SmallVector<const SCEV *, 4> V;
1118 V.push_back(reinterpret_cast<const SCEV *>(-1));
1122 static SmallVector<const SCEV *, 4> getTombstoneKey() {
1123 SmallVector<const SCEV *, 4> V;
1124 V.push_back(reinterpret_cast<const SCEV *>(-2));
1128 static unsigned getHashValue(const SmallVector<const SCEV *, 4> &V) {
1129 unsigned Result = 0;
1130 for (SmallVectorImpl<const SCEV *>::const_iterator I = V.begin(),
1131 E = V.end(); I != E; ++I)
1132 Result ^= DenseMapInfo<const SCEV *>::getHashValue(*I);
1136 static bool isEqual(const SmallVector<const SCEV *, 4> &LHS,
1137 const SmallVector<const SCEV *, 4> &RHS) {
1142 /// LSRUse - This class holds the state that LSR keeps for each use in
1143 /// IVUsers, as well as uses invented by LSR itself. It includes information
1144 /// about what kinds of things can be folded into the user, information about
1145 /// the user itself, and information about how the use may be satisfied.
1146 /// TODO: Represent multiple users of the same expression in common?
1148 DenseSet<SmallVector<const SCEV *, 4>, UniquifierDenseMapInfo> Uniquifier;
1151 /// KindType - An enum for a kind of use, indicating what types of
1152 /// scaled and immediate operands it might support.
1154 Basic, ///< A normal use, with no folding.
1155 Special, ///< A special case of basic, allowing -1 scales.
1156 Address, ///< An address use; folding according to TargetLowering
1157 ICmpZero ///< An equality icmp with both operands folded into one.
1158 // TODO: Add a generic icmp too?
1164 SmallVector<int64_t, 8> Offsets;
1168 /// AllFixupsOutsideLoop - This records whether all of the fixups using this
1169 /// LSRUse are outside of the loop, in which case some special-case heuristics
1171 bool AllFixupsOutsideLoop;
1173 /// RigidFormula is set to true to guarantee that this use will be associated
1174 /// with a single formula--the one that initially matched. Some SCEV
1175 /// expressions cannot be expanded. This allows LSR to consider the registers
1176 /// used by those expressions without the need to expand them later after
1177 /// changing the formula.
1180 /// WidestFixupType - This records the widest use type for any fixup using
1181 /// this LSRUse. FindUseWithSimilarFormula can't consider uses with different
1182 /// max fixup widths to be equivalent, because the narrower one may be relying
1183 /// on the implicit truncation to truncate away bogus bits.
1184 Type *WidestFixupType;
1186 /// Formulae - A list of ways to build a value that can satisfy this user.
1187 /// After the list is populated, one of these is selected heuristically and
1188 /// used to formulate a replacement for OperandValToReplace in UserInst.
1189 SmallVector<Formula, 12> Formulae;
1191 /// Regs - The set of register candidates used by all formulae in this LSRUse.
1192 SmallPtrSet<const SCEV *, 4> Regs;
1194 LSRUse(KindType K, Type *T) : Kind(K), AccessTy(T),
1195 MinOffset(INT64_MAX),
1196 MaxOffset(INT64_MIN),
1197 AllFixupsOutsideLoop(true),
1198 RigidFormula(false),
1199 WidestFixupType(0) {}
1201 bool HasFormulaWithSameRegs(const Formula &F) const;
1202 bool InsertFormula(const Formula &F);
1203 void DeleteFormula(Formula &F);
1204 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1206 void print(raw_ostream &OS) const;
1212 /// HasFormula - Test whether this use as a formula which has the same
1213 /// registers as the given formula.
1214 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1215 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1216 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1217 // Unstable sort by host order ok, because this is only used for uniquifying.
1218 std::sort(Key.begin(), Key.end());
1219 return Uniquifier.count(Key);
1222 /// InsertFormula - If the given formula has not yet been inserted, add it to
1223 /// the list, and return true. Return false otherwise.
1224 bool LSRUse::InsertFormula(const Formula &F) {
1225 if (!Formulae.empty() && RigidFormula)
1228 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1229 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1230 // Unstable sort by host order ok, because this is only used for uniquifying.
1231 std::sort(Key.begin(), Key.end());
1233 if (!Uniquifier.insert(Key).second)
1236 // Using a register to hold the value of 0 is not profitable.
1237 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1238 "Zero allocated in a scaled register!");
1240 for (SmallVectorImpl<const SCEV *>::const_iterator I =
1241 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I)
1242 assert(!(*I)->isZero() && "Zero allocated in a base register!");
1245 // Add the formula to the list.
1246 Formulae.push_back(F);
1248 // Record registers now being used by this use.
1249 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1254 /// DeleteFormula - Remove the given formula from this use's list.
1255 void LSRUse::DeleteFormula(Formula &F) {
1256 if (&F != &Formulae.back())
1257 std::swap(F, Formulae.back());
1258 Formulae.pop_back();
1261 /// RecomputeRegs - Recompute the Regs field, and update RegUses.
1262 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1263 // Now that we've filtered out some formulae, recompute the Regs set.
1264 SmallPtrSet<const SCEV *, 4> OldRegs = Regs;
1266 for (SmallVectorImpl<Formula>::const_iterator I = Formulae.begin(),
1267 E = Formulae.end(); I != E; ++I) {
1268 const Formula &F = *I;
1269 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1270 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1273 // Update the RegTracker.
1274 for (SmallPtrSet<const SCEV *, 4>::iterator I = OldRegs.begin(),
1275 E = OldRegs.end(); I != E; ++I)
1276 if (!Regs.count(*I))
1277 RegUses.DropRegister(*I, LUIdx);
1280 void LSRUse::print(raw_ostream &OS) const {
1281 OS << "LSR Use: Kind=";
1283 case Basic: OS << "Basic"; break;
1284 case Special: OS << "Special"; break;
1285 case ICmpZero: OS << "ICmpZero"; break;
1287 OS << "Address of ";
1288 if (AccessTy->isPointerTy())
1289 OS << "pointer"; // the full pointer type could be really verbose
1294 OS << ", Offsets={";
1295 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
1296 E = Offsets.end(); I != E; ++I) {
1298 if (llvm::next(I) != E)
1303 if (AllFixupsOutsideLoop)
1304 OS << ", all-fixups-outside-loop";
1306 if (WidestFixupType)
1307 OS << ", widest fixup type: " << *WidestFixupType;
1310 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1311 void LSRUse::dump() const {
1312 print(errs()); errs() << '\n';
1316 /// isLegalUse - Test whether the use described by AM is "legal", meaning it can
1317 /// be completely folded into the user instruction at isel time. This includes
1318 /// address-mode folding and special icmp tricks.
1319 static bool isLegalUse(const TargetTransformInfo &TTI, LSRUse::KindType Kind,
1320 Type *AccessTy, GlobalValue *BaseGV, int64_t BaseOffset,
1321 bool HasBaseReg, int64_t Scale) {
1323 case LSRUse::Address:
1324 return TTI.isLegalAddressingMode(AccessTy, BaseGV, BaseOffset, HasBaseReg, Scale);
1326 // Otherwise, just guess that reg+reg addressing is legal.
1329 case LSRUse::ICmpZero:
1330 // There's not even a target hook for querying whether it would be legal to
1331 // fold a GV into an ICmp.
1335 // ICmp only has two operands; don't allow more than two non-trivial parts.
1336 if (Scale != 0 && HasBaseReg && BaseOffset != 0)
1339 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1340 // putting the scaled register in the other operand of the icmp.
1341 if (Scale != 0 && Scale != -1)
1344 // If we have low-level target information, ask the target if it can fold an
1345 // integer immediate on an icmp.
1346 if (BaseOffset != 0) {
1348 // ICmpZero BaseReg + BaseOffset => ICmp BaseReg, -BaseOffset
1349 // ICmpZero -1*ScaleReg + BaseOffset => ICmp ScaleReg, BaseOffset
1350 // Offs is the ICmp immediate.
1352 // The cast does the right thing with INT64_MIN.
1353 BaseOffset = -(uint64_t)BaseOffset;
1354 return TTI.isLegalICmpImmediate(BaseOffset);
1357 // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg
1361 // Only handle single-register values.
1362 return !BaseGV && Scale == 0 && BaseOffset == 0;
1364 case LSRUse::Special:
1365 // Special case Basic to handle -1 scales.
1366 return !BaseGV && (Scale == 0 || Scale == -1) && BaseOffset == 0;
1369 llvm_unreachable("Invalid LSRUse Kind!");
1372 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1373 int64_t MaxOffset, LSRUse::KindType Kind, Type *AccessTy,
1374 GlobalValue *BaseGV, int64_t BaseOffset, bool HasBaseReg,
1376 // Check for overflow.
1377 if (((int64_t)((uint64_t)BaseOffset + MinOffset) > BaseOffset) !=
1380 MinOffset = (uint64_t)BaseOffset + MinOffset;
1381 if (((int64_t)((uint64_t)BaseOffset + MaxOffset) > BaseOffset) !=
1384 MaxOffset = (uint64_t)BaseOffset + MaxOffset;
1386 return isLegalUse(TTI, Kind, AccessTy, BaseGV, MinOffset, HasBaseReg,
1388 isLegalUse(TTI, Kind, AccessTy, BaseGV, MaxOffset, HasBaseReg, Scale);
1391 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1392 int64_t MaxOffset, LSRUse::KindType Kind, Type *AccessTy,
1394 return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, F.BaseGV,
1395 F.BaseOffset, F.HasBaseReg, F.Scale);
1398 static bool isLegal2RegAMUse(const TargetTransformInfo &TTI, const LSRUse &LU,
1400 // If F is used as an Addressing Mode, it may fold one Base plus one
1401 // scaled register. If the scaled register is nil, do as if another
1402 // element of the base regs is a 1-scaled register.
1403 // This is possible if BaseRegs has at least 2 registers.
1405 // If this is not an address calculation, this is not an addressing mode
1407 if (LU.Kind != LSRUse::Address)
1410 // F is already scaled.
1414 // We need to keep one register for the base and one to scale.
1415 if (F.BaseRegs.size() < 2)
1418 return isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
1419 F.BaseGV, F.BaseOffset, F.HasBaseReg, 1);
1422 static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
1423 const LSRUse &LU, const Formula &F) {
1426 assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1427 LU.AccessTy, F) && "Illegal formula in use.");
1430 case LSRUse::Address: {
1431 // Check the scaling factor cost with both the min and max offsets.
1432 int ScaleCostMinOffset =
1433 TTI.getScalingFactorCost(LU.AccessTy, F.BaseGV,
1434 F.BaseOffset + LU.MinOffset,
1435 F.HasBaseReg, F.Scale);
1436 int ScaleCostMaxOffset =
1437 TTI.getScalingFactorCost(LU.AccessTy, F.BaseGV,
1438 F.BaseOffset + LU.MaxOffset,
1439 F.HasBaseReg, F.Scale);
1441 assert(ScaleCostMinOffset >= 0 && ScaleCostMaxOffset >= 0 &&
1442 "Legal addressing mode has an illegal cost!");
1443 return std::max(ScaleCostMinOffset, ScaleCostMaxOffset);
1445 case LSRUse::ICmpZero:
1446 // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg.
1447 // Therefore, return 0 in case F.Scale == -1.
1448 return F.Scale != -1;
1451 case LSRUse::Special:
1455 llvm_unreachable("Invalid LSRUse Kind!");
1458 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1459 LSRUse::KindType Kind, Type *AccessTy,
1460 GlobalValue *BaseGV, int64_t BaseOffset,
1462 // Fast-path: zero is always foldable.
1463 if (BaseOffset == 0 && !BaseGV) return true;
1465 // Conservatively, create an address with an immediate and a
1466 // base and a scale.
1467 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1469 // Canonicalize a scale of 1 to a base register if the formula doesn't
1470 // already have a base register.
1471 if (!HasBaseReg && Scale == 1) {
1476 return isLegalUse(TTI, Kind, AccessTy, BaseGV, BaseOffset, HasBaseReg, Scale);
1479 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1480 ScalarEvolution &SE, int64_t MinOffset,
1481 int64_t MaxOffset, LSRUse::KindType Kind,
1482 Type *AccessTy, const SCEV *S, bool HasBaseReg) {
1483 // Fast-path: zero is always foldable.
1484 if (S->isZero()) return true;
1486 // Conservatively, create an address with an immediate and a
1487 // base and a scale.
1488 int64_t BaseOffset = ExtractImmediate(S, SE);
1489 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1491 // If there's anything else involved, it's not foldable.
1492 if (!S->isZero()) return false;
1494 // Fast-path: zero is always foldable.
1495 if (BaseOffset == 0 && !BaseGV) return true;
1497 // Conservatively, create an address with an immediate and a
1498 // base and a scale.
1499 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1501 return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1502 BaseOffset, HasBaseReg, Scale);
1507 /// UseMapDenseMapInfo - A DenseMapInfo implementation for holding
1508 /// DenseMaps and DenseSets of pairs of const SCEV* and LSRUse::Kind.
1509 struct UseMapDenseMapInfo {
1510 static std::pair<const SCEV *, LSRUse::KindType> getEmptyKey() {
1511 return std::make_pair(reinterpret_cast<const SCEV *>(-1), LSRUse::Basic);
1514 static std::pair<const SCEV *, LSRUse::KindType> getTombstoneKey() {
1515 return std::make_pair(reinterpret_cast<const SCEV *>(-2), LSRUse::Basic);
1519 getHashValue(const std::pair<const SCEV *, LSRUse::KindType> &V) {
1520 unsigned Result = DenseMapInfo<const SCEV *>::getHashValue(V.first);
1521 Result ^= DenseMapInfo<unsigned>::getHashValue(unsigned(V.second));
1525 static bool isEqual(const std::pair<const SCEV *, LSRUse::KindType> &LHS,
1526 const std::pair<const SCEV *, LSRUse::KindType> &RHS) {
1531 /// IVInc - An individual increment in a Chain of IV increments.
1532 /// Relate an IV user to an expression that computes the IV it uses from the IV
1533 /// used by the previous link in the Chain.
1535 /// For the head of a chain, IncExpr holds the absolute SCEV expression for the
1536 /// original IVOperand. The head of the chain's IVOperand is only valid during
1537 /// chain collection, before LSR replaces IV users. During chain generation,
1538 /// IncExpr can be used to find the new IVOperand that computes the same
1541 Instruction *UserInst;
1543 const SCEV *IncExpr;
1545 IVInc(Instruction *U, Value *O, const SCEV *E):
1546 UserInst(U), IVOperand(O), IncExpr(E) {}
1549 // IVChain - The list of IV increments in program order.
1550 // We typically add the head of a chain without finding subsequent links.
1552 SmallVector<IVInc,1> Incs;
1553 const SCEV *ExprBase;
1555 IVChain() : ExprBase(0) {}
1557 IVChain(const IVInc &Head, const SCEV *Base)
1558 : Incs(1, Head), ExprBase(Base) {}
1560 typedef SmallVectorImpl<IVInc>::const_iterator const_iterator;
1562 // begin - return the first increment in the chain.
1563 const_iterator begin() const {
1564 assert(!Incs.empty());
1565 return llvm::next(Incs.begin());
1567 const_iterator end() const {
1571 // hasIncs - Returns true if this chain contains any increments.
1572 bool hasIncs() const { return Incs.size() >= 2; }
1574 // add - Add an IVInc to the end of this chain.
1575 void add(const IVInc &X) { Incs.push_back(X); }
1577 // tailUserInst - Returns the last UserInst in the chain.
1578 Instruction *tailUserInst() const { return Incs.back().UserInst; }
1580 // isProfitableIncrement - Returns true if IncExpr can be profitably added to
1582 bool isProfitableIncrement(const SCEV *OperExpr,
1583 const SCEV *IncExpr,
1587 /// ChainUsers - Helper for CollectChains to track multiple IV increment uses.
1588 /// Distinguish between FarUsers that definitely cross IV increments and
1589 /// NearUsers that may be used between IV increments.
1591 SmallPtrSet<Instruction*, 4> FarUsers;
1592 SmallPtrSet<Instruction*, 4> NearUsers;
1595 /// LSRInstance - This class holds state for the main loop strength reduction
1599 ScalarEvolution &SE;
1602 const TargetTransformInfo &TTI;
1606 /// IVIncInsertPos - This is the insert position that the current loop's
1607 /// induction variable increment should be placed. In simple loops, this is
1608 /// the latch block's terminator. But in more complicated cases, this is a
1609 /// position which will dominate all the in-loop post-increment users.
1610 Instruction *IVIncInsertPos;
1612 /// Factors - Interesting factors between use strides.
1613 SmallSetVector<int64_t, 8> Factors;
1615 /// Types - Interesting use types, to facilitate truncation reuse.
1616 SmallSetVector<Type *, 4> Types;
1618 /// Fixups - The list of operands which are to be replaced.
1619 SmallVector<LSRFixup, 16> Fixups;
1621 /// Uses - The list of interesting uses.
1622 SmallVector<LSRUse, 16> Uses;
1624 /// RegUses - Track which uses use which register candidates.
1625 RegUseTracker RegUses;
1627 // Limit the number of chains to avoid quadratic behavior. We don't expect to
1628 // have more than a few IV increment chains in a loop. Missing a Chain falls
1629 // back to normal LSR behavior for those uses.
1630 static const unsigned MaxChains = 8;
1632 /// IVChainVec - IV users can form a chain of IV increments.
1633 SmallVector<IVChain, MaxChains> IVChainVec;
1635 /// IVIncSet - IV users that belong to profitable IVChains.
1636 SmallPtrSet<Use*, MaxChains> IVIncSet;
1638 void OptimizeShadowIV();
1639 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1640 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1641 void OptimizeLoopTermCond();
1643 void ChainInstruction(Instruction *UserInst, Instruction *IVOper,
1644 SmallVectorImpl<ChainUsers> &ChainUsersVec);
1645 void FinalizeChain(IVChain &Chain);
1646 void CollectChains();
1647 void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
1648 SmallVectorImpl<WeakVH> &DeadInsts);
1650 void CollectInterestingTypesAndFactors();
1651 void CollectFixupsAndInitialFormulae();
1653 LSRFixup &getNewFixup() {
1654 Fixups.push_back(LSRFixup());
1655 return Fixups.back();
1658 // Support for sharing of LSRUses between LSRFixups.
1659 typedef DenseMap<std::pair<const SCEV *, LSRUse::KindType>,
1661 UseMapDenseMapInfo> UseMapTy;
1664 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1665 LSRUse::KindType Kind, Type *AccessTy);
1667 std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
1668 LSRUse::KindType Kind,
1671 void DeleteUse(LSRUse &LU, size_t LUIdx);
1673 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1675 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1676 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1677 void CountRegisters(const Formula &F, size_t LUIdx);
1678 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1680 void CollectLoopInvariantFixupsAndFormulae();
1682 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1683 unsigned Depth = 0);
1684 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1685 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1686 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1687 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1688 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1689 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1690 void GenerateCrossUseConstantOffsets();
1691 void GenerateAllReuseFormulae();
1693 void FilterOutUndesirableDedicatedRegisters();
1695 size_t EstimateSearchSpaceComplexity() const;
1696 void NarrowSearchSpaceByDetectingSupersets();
1697 void NarrowSearchSpaceByCollapsingUnrolledCode();
1698 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
1699 void NarrowSearchSpaceByPickingWinnerRegs();
1700 void NarrowSearchSpaceUsingHeuristics();
1702 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1704 SmallVectorImpl<const Formula *> &Workspace,
1705 const Cost &CurCost,
1706 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1707 DenseSet<const SCEV *> &VisitedRegs) const;
1708 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1710 BasicBlock::iterator
1711 HoistInsertPosition(BasicBlock::iterator IP,
1712 const SmallVectorImpl<Instruction *> &Inputs) const;
1713 BasicBlock::iterator
1714 AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1717 SCEVExpander &Rewriter) const;
1719 Value *Expand(const LSRFixup &LF,
1721 BasicBlock::iterator IP,
1722 SCEVExpander &Rewriter,
1723 SmallVectorImpl<WeakVH> &DeadInsts) const;
1724 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1726 SCEVExpander &Rewriter,
1727 SmallVectorImpl<WeakVH> &DeadInsts,
1729 void Rewrite(const LSRFixup &LF,
1731 SCEVExpander &Rewriter,
1732 SmallVectorImpl<WeakVH> &DeadInsts,
1734 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
1738 LSRInstance(Loop *L, Pass *P);
1740 bool getChanged() const { return Changed; }
1742 void print_factors_and_types(raw_ostream &OS) const;
1743 void print_fixups(raw_ostream &OS) const;
1744 void print_uses(raw_ostream &OS) const;
1745 void print(raw_ostream &OS) const;
1751 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1752 /// inside the loop then try to eliminate the cast operation.
1753 void LSRInstance::OptimizeShadowIV() {
1754 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1755 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1758 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1759 UI != E; /* empty */) {
1760 IVUsers::const_iterator CandidateUI = UI;
1762 Instruction *ShadowUse = CandidateUI->getUser();
1764 bool IsSigned = false;
1766 /* If shadow use is a int->float cast then insert a second IV
1767 to eliminate this cast.
1769 for (unsigned i = 0; i < n; ++i)
1775 for (unsigned i = 0; i < n; ++i, ++d)
1778 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
1780 DestTy = UCast->getDestTy();
1782 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
1784 DestTy = SCast->getDestTy();
1786 if (!DestTy) continue;
1788 // If target does not support DestTy natively then do not apply
1789 // this transformation.
1790 if (!TTI.isTypeLegal(DestTy)) continue;
1792 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1794 if (PH->getNumIncomingValues() != 2) continue;
1796 Type *SrcTy = PH->getType();
1797 int Mantissa = DestTy->getFPMantissaWidth();
1798 if (Mantissa == -1) continue;
1799 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1802 unsigned Entry, Latch;
1803 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1811 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1812 if (!Init) continue;
1813 Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
1814 (double)Init->getSExtValue() :
1815 (double)Init->getZExtValue());
1817 BinaryOperator *Incr =
1818 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1819 if (!Incr) continue;
1820 if (Incr->getOpcode() != Instruction::Add
1821 && Incr->getOpcode() != Instruction::Sub)
1824 /* Initialize new IV, double d = 0.0 in above example. */
1826 if (Incr->getOperand(0) == PH)
1827 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1828 else if (Incr->getOperand(1) == PH)
1829 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1835 // Ignore negative constants, as the code below doesn't handle them
1836 // correctly. TODO: Remove this restriction.
1837 if (!C->getValue().isStrictlyPositive()) continue;
1839 /* Add new PHINode. */
1840 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
1842 /* create new increment. '++d' in above example. */
1843 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1844 BinaryOperator *NewIncr =
1845 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1846 Instruction::FAdd : Instruction::FSub,
1847 NewPH, CFP, "IV.S.next.", Incr);
1849 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1850 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1852 /* Remove cast operation */
1853 ShadowUse->replaceAllUsesWith(NewPH);
1854 ShadowUse->eraseFromParent();
1860 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1861 /// set the IV user and stride information and return true, otherwise return
1863 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
1864 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1865 if (UI->getUser() == Cond) {
1866 // NOTE: we could handle setcc instructions with multiple uses here, but
1867 // InstCombine does it as well for simple uses, it's not clear that it
1868 // occurs enough in real life to handle.
1875 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
1876 /// a max computation.
1878 /// This is a narrow solution to a specific, but acute, problem. For loops
1884 /// } while (++i < n);
1886 /// the trip count isn't just 'n', because 'n' might not be positive. And
1887 /// unfortunately this can come up even for loops where the user didn't use
1888 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1889 /// will commonly be lowered like this:
1895 /// } while (++i < n);
1898 /// and then it's possible for subsequent optimization to obscure the if
1899 /// test in such a way that indvars can't find it.
1901 /// When indvars can't find the if test in loops like this, it creates a
1902 /// max expression, which allows it to give the loop a canonical
1903 /// induction variable:
1906 /// max = n < 1 ? 1 : n;
1909 /// } while (++i != max);
1911 /// Canonical induction variables are necessary because the loop passes
1912 /// are designed around them. The most obvious example of this is the
1913 /// LoopInfo analysis, which doesn't remember trip count values. It
1914 /// expects to be able to rediscover the trip count each time it is
1915 /// needed, and it does this using a simple analysis that only succeeds if
1916 /// the loop has a canonical induction variable.
1918 /// However, when it comes time to generate code, the maximum operation
1919 /// can be quite costly, especially if it's inside of an outer loop.
1921 /// This function solves this problem by detecting this type of loop and
1922 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1923 /// the instructions for the maximum computation.
1925 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1926 // Check that the loop matches the pattern we're looking for.
1927 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1928 Cond->getPredicate() != CmpInst::ICMP_NE)
1931 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1932 if (!Sel || !Sel->hasOneUse()) return Cond;
1934 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1935 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1937 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
1939 // Add one to the backedge-taken count to get the trip count.
1940 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
1941 if (IterationCount != SE.getSCEV(Sel)) return Cond;
1943 // Check for a max calculation that matches the pattern. There's no check
1944 // for ICMP_ULE here because the comparison would be with zero, which
1945 // isn't interesting.
1946 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
1947 const SCEVNAryExpr *Max = 0;
1948 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
1949 Pred = ICmpInst::ICMP_SLE;
1951 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
1952 Pred = ICmpInst::ICMP_SLT;
1954 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
1955 Pred = ICmpInst::ICMP_ULT;
1962 // To handle a max with more than two operands, this optimization would
1963 // require additional checking and setup.
1964 if (Max->getNumOperands() != 2)
1967 const SCEV *MaxLHS = Max->getOperand(0);
1968 const SCEV *MaxRHS = Max->getOperand(1);
1970 // ScalarEvolution canonicalizes constants to the left. For < and >, look
1971 // for a comparison with 1. For <= and >=, a comparison with zero.
1973 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
1976 // Check the relevant induction variable for conformance to
1978 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
1979 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1980 if (!AR || !AR->isAffine() ||
1981 AR->getStart() != One ||
1982 AR->getStepRecurrence(SE) != One)
1985 assert(AR->getLoop() == L &&
1986 "Loop condition operand is an addrec in a different loop!");
1988 // Check the right operand of the select, and remember it, as it will
1989 // be used in the new comparison instruction.
1991 if (ICmpInst::isTrueWhenEqual(Pred)) {
1992 // Look for n+1, and grab n.
1993 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
1994 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
1995 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
1996 NewRHS = BO->getOperand(0);
1997 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
1998 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
1999 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2000 NewRHS = BO->getOperand(0);
2003 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
2004 NewRHS = Sel->getOperand(1);
2005 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
2006 NewRHS = Sel->getOperand(2);
2007 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
2008 NewRHS = SU->getValue();
2010 // Max doesn't match expected pattern.
2013 // Determine the new comparison opcode. It may be signed or unsigned,
2014 // and the original comparison may be either equality or inequality.
2015 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
2016 Pred = CmpInst::getInversePredicate(Pred);
2018 // Ok, everything looks ok to change the condition into an SLT or SGE and
2019 // delete the max calculation.
2021 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
2023 // Delete the max calculation instructions.
2024 Cond->replaceAllUsesWith(NewCond);
2025 CondUse->setUser(NewCond);
2026 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2027 Cond->eraseFromParent();
2028 Sel->eraseFromParent();
2029 if (Cmp->use_empty())
2030 Cmp->eraseFromParent();
2034 /// OptimizeLoopTermCond - Change loop terminating condition to use the
2035 /// postinc iv when possible.
2037 LSRInstance::OptimizeLoopTermCond() {
2038 SmallPtrSet<Instruction *, 4> PostIncs;
2040 BasicBlock *LatchBlock = L->getLoopLatch();
2041 SmallVector<BasicBlock*, 8> ExitingBlocks;
2042 L->getExitingBlocks(ExitingBlocks);
2044 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
2045 BasicBlock *ExitingBlock = ExitingBlocks[i];
2047 // Get the terminating condition for the loop if possible. If we
2048 // can, we want to change it to use a post-incremented version of its
2049 // induction variable, to allow coalescing the live ranges for the IV into
2050 // one register value.
2052 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2055 // FIXME: Overly conservative, termination condition could be an 'or' etc..
2056 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
2059 // Search IVUsesByStride to find Cond's IVUse if there is one.
2060 IVStrideUse *CondUse = 0;
2061 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2062 if (!FindIVUserForCond(Cond, CondUse))
2065 // If the trip count is computed in terms of a max (due to ScalarEvolution
2066 // being unable to find a sufficient guard, for example), change the loop
2067 // comparison to use SLT or ULT instead of NE.
2068 // One consequence of doing this now is that it disrupts the count-down
2069 // optimization. That's not always a bad thing though, because in such
2070 // cases it may still be worthwhile to avoid a max.
2071 Cond = OptimizeMax(Cond, CondUse);
2073 // If this exiting block dominates the latch block, it may also use
2074 // the post-inc value if it won't be shared with other uses.
2075 // Check for dominance.
2076 if (!DT.dominates(ExitingBlock, LatchBlock))
2079 // Conservatively avoid trying to use the post-inc value in non-latch
2080 // exits if there may be pre-inc users in intervening blocks.
2081 if (LatchBlock != ExitingBlock)
2082 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
2083 // Test if the use is reachable from the exiting block. This dominator
2084 // query is a conservative approximation of reachability.
2085 if (&*UI != CondUse &&
2086 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
2087 // Conservatively assume there may be reuse if the quotient of their
2088 // strides could be a legal scale.
2089 const SCEV *A = IU.getStride(*CondUse, L);
2090 const SCEV *B = IU.getStride(*UI, L);
2091 if (!A || !B) continue;
2092 if (SE.getTypeSizeInBits(A->getType()) !=
2093 SE.getTypeSizeInBits(B->getType())) {
2094 if (SE.getTypeSizeInBits(A->getType()) >
2095 SE.getTypeSizeInBits(B->getType()))
2096 B = SE.getSignExtendExpr(B, A->getType());
2098 A = SE.getSignExtendExpr(A, B->getType());
2100 if (const SCEVConstant *D =
2101 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
2102 const ConstantInt *C = D->getValue();
2103 // Stride of one or negative one can have reuse with non-addresses.
2104 if (C->isOne() || C->isAllOnesValue())
2105 goto decline_post_inc;
2106 // Avoid weird situations.
2107 if (C->getValue().getMinSignedBits() >= 64 ||
2108 C->getValue().isMinSignedValue())
2109 goto decline_post_inc;
2110 // Check for possible scaled-address reuse.
2111 Type *AccessTy = getAccessType(UI->getUser());
2112 int64_t Scale = C->getSExtValue();
2113 if (TTI.isLegalAddressingMode(AccessTy, /*BaseGV=*/ 0,
2115 /*HasBaseReg=*/ false, Scale))
2116 goto decline_post_inc;
2118 if (TTI.isLegalAddressingMode(AccessTy, /*BaseGV=*/ 0,
2120 /*HasBaseReg=*/ false, Scale))
2121 goto decline_post_inc;
2125 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
2128 // It's possible for the setcc instruction to be anywhere in the loop, and
2129 // possible for it to have multiple users. If it is not immediately before
2130 // the exiting block branch, move it.
2131 if (&*++BasicBlock::iterator(Cond) != TermBr) {
2132 if (Cond->hasOneUse()) {
2133 Cond->moveBefore(TermBr);
2135 // Clone the terminating condition and insert into the loopend.
2136 ICmpInst *OldCond = Cond;
2137 Cond = cast<ICmpInst>(Cond->clone());
2138 Cond->setName(L->getHeader()->getName() + ".termcond");
2139 ExitingBlock->getInstList().insert(TermBr, Cond);
2141 // Clone the IVUse, as the old use still exists!
2142 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
2143 TermBr->replaceUsesOfWith(OldCond, Cond);
2147 // If we get to here, we know that we can transform the setcc instruction to
2148 // use the post-incremented version of the IV, allowing us to coalesce the
2149 // live ranges for the IV correctly.
2150 CondUse->transformToPostInc(L);
2153 PostIncs.insert(Cond);
2157 // Determine an insertion point for the loop induction variable increment. It
2158 // must dominate all the post-inc comparisons we just set up, and it must
2159 // dominate the loop latch edge.
2160 IVIncInsertPos = L->getLoopLatch()->getTerminator();
2161 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(),
2162 E = PostIncs.end(); I != E; ++I) {
2164 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
2166 if (BB == (*I)->getParent())
2167 IVIncInsertPos = *I;
2168 else if (BB != IVIncInsertPos->getParent())
2169 IVIncInsertPos = BB->getTerminator();
2173 /// reconcileNewOffset - Determine if the given use can accommodate a fixup
2174 /// at the given offset and other details. If so, update the use and
2177 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
2178 LSRUse::KindType Kind, Type *AccessTy) {
2179 int64_t NewMinOffset = LU.MinOffset;
2180 int64_t NewMaxOffset = LU.MaxOffset;
2181 Type *NewAccessTy = AccessTy;
2183 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
2184 // something conservative, however this can pessimize in the case that one of
2185 // the uses will have all its uses outside the loop, for example.
2186 if (LU.Kind != Kind)
2188 // Conservatively assume HasBaseReg is true for now.
2189 if (NewOffset < LU.MinOffset) {
2190 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ 0,
2191 LU.MaxOffset - NewOffset, HasBaseReg))
2193 NewMinOffset = NewOffset;
2194 } else if (NewOffset > LU.MaxOffset) {
2195 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ 0,
2196 NewOffset - LU.MinOffset, HasBaseReg))
2198 NewMaxOffset = NewOffset;
2200 // Check for a mismatched access type, and fall back conservatively as needed.
2201 // TODO: Be less conservative when the type is similar and can use the same
2202 // addressing modes.
2203 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
2204 NewAccessTy = Type::getVoidTy(AccessTy->getContext());
2207 LU.MinOffset = NewMinOffset;
2208 LU.MaxOffset = NewMaxOffset;
2209 LU.AccessTy = NewAccessTy;
2210 if (NewOffset != LU.Offsets.back())
2211 LU.Offsets.push_back(NewOffset);
2215 /// getUse - Return an LSRUse index and an offset value for a fixup which
2216 /// needs the given expression, with the given kind and optional access type.
2217 /// Either reuse an existing use or create a new one, as needed.
2218 std::pair<size_t, int64_t>
2219 LSRInstance::getUse(const SCEV *&Expr,
2220 LSRUse::KindType Kind, Type *AccessTy) {
2221 const SCEV *Copy = Expr;
2222 int64_t Offset = ExtractImmediate(Expr, SE);
2224 // Basic uses can't accept any offset, for example.
2225 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ 0,
2226 Offset, /*HasBaseReg=*/ true)) {
2231 std::pair<UseMapTy::iterator, bool> P =
2232 UseMap.insert(std::make_pair(std::make_pair(Expr, Kind), 0));
2234 // A use already existed with this base.
2235 size_t LUIdx = P.first->second;
2236 LSRUse &LU = Uses[LUIdx];
2237 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
2239 return std::make_pair(LUIdx, Offset);
2242 // Create a new use.
2243 size_t LUIdx = Uses.size();
2244 P.first->second = LUIdx;
2245 Uses.push_back(LSRUse(Kind, AccessTy));
2246 LSRUse &LU = Uses[LUIdx];
2248 // We don't need to track redundant offsets, but we don't need to go out
2249 // of our way here to avoid them.
2250 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
2251 LU.Offsets.push_back(Offset);
2253 LU.MinOffset = Offset;
2254 LU.MaxOffset = Offset;
2255 return std::make_pair(LUIdx, Offset);
2258 /// DeleteUse - Delete the given use from the Uses list.
2259 void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
2260 if (&LU != &Uses.back())
2261 std::swap(LU, Uses.back());
2265 RegUses.SwapAndDropUse(LUIdx, Uses.size());
2268 /// FindUseWithFormula - Look for a use distinct from OrigLU which is has
2269 /// a formula that has the same registers as the given formula.
2271 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
2272 const LSRUse &OrigLU) {
2273 // Search all uses for the formula. This could be more clever.
2274 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2275 LSRUse &LU = Uses[LUIdx];
2276 // Check whether this use is close enough to OrigLU, to see whether it's
2277 // worthwhile looking through its formulae.
2278 // Ignore ICmpZero uses because they may contain formulae generated by
2279 // GenerateICmpZeroScales, in which case adding fixup offsets may
2281 if (&LU != &OrigLU &&
2282 LU.Kind != LSRUse::ICmpZero &&
2283 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
2284 LU.WidestFixupType == OrigLU.WidestFixupType &&
2285 LU.HasFormulaWithSameRegs(OrigF)) {
2286 // Scan through this use's formulae.
2287 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
2288 E = LU.Formulae.end(); I != E; ++I) {
2289 const Formula &F = *I;
2290 // Check to see if this formula has the same registers and symbols
2292 if (F.BaseRegs == OrigF.BaseRegs &&
2293 F.ScaledReg == OrigF.ScaledReg &&
2294 F.BaseGV == OrigF.BaseGV &&
2295 F.Scale == OrigF.Scale &&
2296 F.UnfoldedOffset == OrigF.UnfoldedOffset) {
2297 if (F.BaseOffset == 0)
2299 // This is the formula where all the registers and symbols matched;
2300 // there aren't going to be any others. Since we declined it, we
2301 // can skip the rest of the formulae and proceed to the next LSRUse.
2308 // Nothing looked good.
2312 void LSRInstance::CollectInterestingTypesAndFactors() {
2313 SmallSetVector<const SCEV *, 4> Strides;
2315 // Collect interesting types and strides.
2316 SmallVector<const SCEV *, 4> Worklist;
2317 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2318 const SCEV *Expr = IU.getExpr(*UI);
2320 // Collect interesting types.
2321 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2323 // Add strides for mentioned loops.
2324 Worklist.push_back(Expr);
2326 const SCEV *S = Worklist.pop_back_val();
2327 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2328 if (AR->getLoop() == L)
2329 Strides.insert(AR->getStepRecurrence(SE));
2330 Worklist.push_back(AR->getStart());
2331 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2332 Worklist.append(Add->op_begin(), Add->op_end());
2334 } while (!Worklist.empty());
2337 // Compute interesting factors from the set of interesting strides.
2338 for (SmallSetVector<const SCEV *, 4>::const_iterator
2339 I = Strides.begin(), E = Strides.end(); I != E; ++I)
2340 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
2341 llvm::next(I); NewStrideIter != E; ++NewStrideIter) {
2342 const SCEV *OldStride = *I;
2343 const SCEV *NewStride = *NewStrideIter;
2345 if (SE.getTypeSizeInBits(OldStride->getType()) !=
2346 SE.getTypeSizeInBits(NewStride->getType())) {
2347 if (SE.getTypeSizeInBits(OldStride->getType()) >
2348 SE.getTypeSizeInBits(NewStride->getType()))
2349 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2351 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2353 if (const SCEVConstant *Factor =
2354 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2356 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2357 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2358 } else if (const SCEVConstant *Factor =
2359 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2362 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
2363 Factors.insert(Factor->getValue()->getValue().getSExtValue());
2367 // If all uses use the same type, don't bother looking for truncation-based
2369 if (Types.size() == 1)
2372 DEBUG(print_factors_and_types(dbgs()));
2375 /// findIVOperand - Helper for CollectChains that finds an IV operand (computed
2376 /// by an AddRec in this loop) within [OI,OE) or returns OE. If IVUsers mapped
2377 /// Instructions to IVStrideUses, we could partially skip this.
2378 static User::op_iterator
2379 findIVOperand(User::op_iterator OI, User::op_iterator OE,
2380 Loop *L, ScalarEvolution &SE) {
2381 for(; OI != OE; ++OI) {
2382 if (Instruction *Oper = dyn_cast<Instruction>(*OI)) {
2383 if (!SE.isSCEVable(Oper->getType()))
2386 if (const SCEVAddRecExpr *AR =
2387 dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) {
2388 if (AR->getLoop() == L)
2396 /// getWideOperand - IVChain logic must consistenctly peek base TruncInst
2397 /// operands, so wrap it in a convenient helper.
2398 static Value *getWideOperand(Value *Oper) {
2399 if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper))
2400 return Trunc->getOperand(0);
2404 /// isCompatibleIVType - Return true if we allow an IV chain to include both
2406 static bool isCompatibleIVType(Value *LVal, Value *RVal) {
2407 Type *LType = LVal->getType();
2408 Type *RType = RVal->getType();
2409 return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy());
2412 /// getExprBase - Return an approximation of this SCEV expression's "base", or
2413 /// NULL for any constant. Returning the expression itself is
2414 /// conservative. Returning a deeper subexpression is more precise and valid as
2415 /// long as it isn't less complex than another subexpression. For expressions
2416 /// involving multiple unscaled values, we need to return the pointer-type
2417 /// SCEVUnknown. This avoids forming chains across objects, such as:
2418 /// PrevOper==a[i], IVOper==b[i], IVInc==b-a.
2420 /// Since SCEVUnknown is the rightmost type, and pointers are the rightmost
2421 /// SCEVUnknown, we simply return the rightmost SCEV operand.
2422 static const SCEV *getExprBase(const SCEV *S) {
2423 switch (S->getSCEVType()) {
2424 default: // uncluding scUnknown.
2429 return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand());
2431 return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand());
2433 return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand());
2435 // Skip over scaled operands (scMulExpr) to follow add operands as long as
2436 // there's nothing more complex.
2437 // FIXME: not sure if we want to recognize negation.
2438 const SCEVAddExpr *Add = cast<SCEVAddExpr>(S);
2439 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(Add->op_end()),
2440 E(Add->op_begin()); I != E; ++I) {
2441 const SCEV *SubExpr = *I;
2442 if (SubExpr->getSCEVType() == scAddExpr)
2443 return getExprBase(SubExpr);
2445 if (SubExpr->getSCEVType() != scMulExpr)
2448 return S; // all operands are scaled, be conservative.
2451 return getExprBase(cast<SCEVAddRecExpr>(S)->getStart());
2455 /// Return true if the chain increment is profitable to expand into a loop
2456 /// invariant value, which may require its own register. A profitable chain
2457 /// increment will be an offset relative to the same base. We allow such offsets
2458 /// to potentially be used as chain increment as long as it's not obviously
2459 /// expensive to expand using real instructions.
2460 bool IVChain::isProfitableIncrement(const SCEV *OperExpr,
2461 const SCEV *IncExpr,
2462 ScalarEvolution &SE) {
2463 // Aggressively form chains when -stress-ivchain.
2467 // Do not replace a constant offset from IV head with a nonconstant IV
2469 if (!isa<SCEVConstant>(IncExpr)) {
2470 const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Incs[0].IVOperand));
2471 if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr)))
2475 SmallPtrSet<const SCEV*, 8> Processed;
2476 return !isHighCostExpansion(IncExpr, Processed, SE);
2479 /// Return true if the number of registers needed for the chain is estimated to
2480 /// be less than the number required for the individual IV users. First prohibit
2481 /// any IV users that keep the IV live across increments (the Users set should
2482 /// be empty). Next count the number and type of increments in the chain.
2484 /// Chaining IVs can lead to considerable code bloat if ISEL doesn't
2485 /// effectively use postinc addressing modes. Only consider it profitable it the
2486 /// increments can be computed in fewer registers when chained.
2488 /// TODO: Consider IVInc free if it's already used in another chains.
2490 isProfitableChain(IVChain &Chain, SmallPtrSet<Instruction*, 4> &Users,
2491 ScalarEvolution &SE, const TargetTransformInfo &TTI) {
2495 if (!Chain.hasIncs())
2498 if (!Users.empty()) {
2499 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " users:\n";
2500 for (SmallPtrSet<Instruction*, 4>::const_iterator I = Users.begin(),
2501 E = Users.end(); I != E; ++I) {
2502 dbgs() << " " << **I << "\n";
2506 assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
2508 // The chain itself may require a register, so intialize cost to 1.
2511 // A complete chain likely eliminates the need for keeping the original IV in
2512 // a register. LSR does not currently know how to form a complete chain unless
2513 // the header phi already exists.
2514 if (isa<PHINode>(Chain.tailUserInst())
2515 && SE.getSCEV(Chain.tailUserInst()) == Chain.Incs[0].IncExpr) {
2518 const SCEV *LastIncExpr = 0;
2519 unsigned NumConstIncrements = 0;
2520 unsigned NumVarIncrements = 0;
2521 unsigned NumReusedIncrements = 0;
2522 for (IVChain::const_iterator I = Chain.begin(), E = Chain.end();
2525 if (I->IncExpr->isZero())
2528 // Incrementing by zero or some constant is neutral. We assume constants can
2529 // be folded into an addressing mode or an add's immediate operand.
2530 if (isa<SCEVConstant>(I->IncExpr)) {
2531 ++NumConstIncrements;
2535 if (I->IncExpr == LastIncExpr)
2536 ++NumReusedIncrements;
2540 LastIncExpr = I->IncExpr;
2542 // An IV chain with a single increment is handled by LSR's postinc
2543 // uses. However, a chain with multiple increments requires keeping the IV's
2544 // value live longer than it needs to be if chained.
2545 if (NumConstIncrements > 1)
2548 // Materializing increment expressions in the preheader that didn't exist in
2549 // the original code may cost a register. For example, sign-extended array
2550 // indices can produce ridiculous increments like this:
2551 // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64)))
2552 cost += NumVarIncrements;
2554 // Reusing variable increments likely saves a register to hold the multiple of
2556 cost -= NumReusedIncrements;
2558 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " Cost: " << cost
2564 /// ChainInstruction - Add this IV user to an existing chain or make it the head
2566 void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper,
2567 SmallVectorImpl<ChainUsers> &ChainUsersVec) {
2568 // When IVs are used as types of varying widths, they are generally converted
2569 // to a wider type with some uses remaining narrow under a (free) trunc.
2570 Value *const NextIV = getWideOperand(IVOper);
2571 const SCEV *const OperExpr = SE.getSCEV(NextIV);
2572 const SCEV *const OperExprBase = getExprBase(OperExpr);
2574 // Visit all existing chains. Check if its IVOper can be computed as a
2575 // profitable loop invariant increment from the last link in the Chain.
2576 unsigned ChainIdx = 0, NChains = IVChainVec.size();
2577 const SCEV *LastIncExpr = 0;
2578 for (; ChainIdx < NChains; ++ChainIdx) {
2579 IVChain &Chain = IVChainVec[ChainIdx];
2581 // Prune the solution space aggressively by checking that both IV operands
2582 // are expressions that operate on the same unscaled SCEVUnknown. This
2583 // "base" will be canceled by the subsequent getMinusSCEV call. Checking
2584 // first avoids creating extra SCEV expressions.
2585 if (!StressIVChain && Chain.ExprBase != OperExprBase)
2588 Value *PrevIV = getWideOperand(Chain.Incs.back().IVOperand);
2589 if (!isCompatibleIVType(PrevIV, NextIV))
2592 // A phi node terminates a chain.
2593 if (isa<PHINode>(UserInst) && isa<PHINode>(Chain.tailUserInst()))
2596 // The increment must be loop-invariant so it can be kept in a register.
2597 const SCEV *PrevExpr = SE.getSCEV(PrevIV);
2598 const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr);
2599 if (!SE.isLoopInvariant(IncExpr, L))
2602 if (Chain.isProfitableIncrement(OperExpr, IncExpr, SE)) {
2603 LastIncExpr = IncExpr;
2607 // If we haven't found a chain, create a new one, unless we hit the max. Don't
2608 // bother for phi nodes, because they must be last in the chain.
2609 if (ChainIdx == NChains) {
2610 if (isa<PHINode>(UserInst))
2612 if (NChains >= MaxChains && !StressIVChain) {
2613 DEBUG(dbgs() << "IV Chain Limit\n");
2616 LastIncExpr = OperExpr;
2617 // IVUsers may have skipped over sign/zero extensions. We don't currently
2618 // attempt to form chains involving extensions unless they can be hoisted
2619 // into this loop's AddRec.
2620 if (!isa<SCEVAddRecExpr>(LastIncExpr))
2623 IVChainVec.push_back(IVChain(IVInc(UserInst, IVOper, LastIncExpr),
2625 ChainUsersVec.resize(NChains);
2626 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Head: (" << *UserInst
2627 << ") IV=" << *LastIncExpr << "\n");
2629 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Inc: (" << *UserInst
2630 << ") IV+" << *LastIncExpr << "\n");
2631 // Add this IV user to the end of the chain.
2632 IVChainVec[ChainIdx].add(IVInc(UserInst, IVOper, LastIncExpr));
2634 IVChain &Chain = IVChainVec[ChainIdx];
2636 SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers;
2637 // This chain's NearUsers become FarUsers.
2638 if (!LastIncExpr->isZero()) {
2639 ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(),
2644 // All other uses of IVOperand become near uses of the chain.
2645 // We currently ignore intermediate values within SCEV expressions, assuming
2646 // they will eventually be used be the current chain, or can be computed
2647 // from one of the chain increments. To be more precise we could
2648 // transitively follow its user and only add leaf IV users to the set.
2649 for (Value::use_iterator UseIter = IVOper->use_begin(),
2650 UseEnd = IVOper->use_end(); UseIter != UseEnd; ++UseIter) {
2651 Instruction *OtherUse = dyn_cast<Instruction>(*UseIter);
2654 // Uses in the chain will no longer be uses if the chain is formed.
2655 // Include the head of the chain in this iteration (not Chain.begin()).
2656 IVChain::const_iterator IncIter = Chain.Incs.begin();
2657 IVChain::const_iterator IncEnd = Chain.Incs.end();
2658 for( ; IncIter != IncEnd; ++IncIter) {
2659 if (IncIter->UserInst == OtherUse)
2662 if (IncIter != IncEnd)
2665 if (SE.isSCEVable(OtherUse->getType())
2666 && !isa<SCEVUnknown>(SE.getSCEV(OtherUse))
2667 && IU.isIVUserOrOperand(OtherUse)) {
2670 NearUsers.insert(OtherUse);
2673 // Since this user is part of the chain, it's no longer considered a use
2675 ChainUsersVec[ChainIdx].FarUsers.erase(UserInst);
2678 /// CollectChains - Populate the vector of Chains.
2680 /// This decreases ILP at the architecture level. Targets with ample registers,
2681 /// multiple memory ports, and no register renaming probably don't want
2682 /// this. However, such targets should probably disable LSR altogether.
2684 /// The job of LSR is to make a reasonable choice of induction variables across
2685 /// the loop. Subsequent passes can easily "unchain" computation exposing more
2686 /// ILP *within the loop* if the target wants it.
2688 /// Finding the best IV chain is potentially a scheduling problem. Since LSR
2689 /// will not reorder memory operations, it will recognize this as a chain, but
2690 /// will generate redundant IV increments. Ideally this would be corrected later
2691 /// by a smart scheduler:
2697 /// TODO: Walk the entire domtree within this loop, not just the path to the
2698 /// loop latch. This will discover chains on side paths, but requires
2699 /// maintaining multiple copies of the Chains state.
2700 void LSRInstance::CollectChains() {
2701 DEBUG(dbgs() << "Collecting IV Chains.\n");
2702 SmallVector<ChainUsers, 8> ChainUsersVec;
2704 SmallVector<BasicBlock *,8> LatchPath;
2705 BasicBlock *LoopHeader = L->getHeader();
2706 for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch());
2707 Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) {
2708 LatchPath.push_back(Rung->getBlock());
2710 LatchPath.push_back(LoopHeader);
2712 // Walk the instruction stream from the loop header to the loop latch.
2713 for (SmallVectorImpl<BasicBlock *>::reverse_iterator
2714 BBIter = LatchPath.rbegin(), BBEnd = LatchPath.rend();
2715 BBIter != BBEnd; ++BBIter) {
2716 for (BasicBlock::iterator I = (*BBIter)->begin(), E = (*BBIter)->end();
2718 // Skip instructions that weren't seen by IVUsers analysis.
2719 if (isa<PHINode>(I) || !IU.isIVUserOrOperand(I))
2722 // Ignore users that are part of a SCEV expression. This way we only
2723 // consider leaf IV Users. This effectively rediscovers a portion of
2724 // IVUsers analysis but in program order this time.
2725 if (SE.isSCEVable(I->getType()) && !isa<SCEVUnknown>(SE.getSCEV(I)))
2728 // Remove this instruction from any NearUsers set it may be in.
2729 for (unsigned ChainIdx = 0, NChains = IVChainVec.size();
2730 ChainIdx < NChains; ++ChainIdx) {
2731 ChainUsersVec[ChainIdx].NearUsers.erase(I);
2733 // Search for operands that can be chained.
2734 SmallPtrSet<Instruction*, 4> UniqueOperands;
2735 User::op_iterator IVOpEnd = I->op_end();
2736 User::op_iterator IVOpIter = findIVOperand(I->op_begin(), IVOpEnd, L, SE);
2737 while (IVOpIter != IVOpEnd) {
2738 Instruction *IVOpInst = cast<Instruction>(*IVOpIter);
2739 if (UniqueOperands.insert(IVOpInst))
2740 ChainInstruction(I, IVOpInst, ChainUsersVec);
2741 IVOpIter = findIVOperand(llvm::next(IVOpIter), IVOpEnd, L, SE);
2743 } // Continue walking down the instructions.
2744 } // Continue walking down the domtree.
2745 // Visit phi backedges to determine if the chain can generate the IV postinc.
2746 for (BasicBlock::iterator I = L->getHeader()->begin();
2747 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2748 if (!SE.isSCEVable(PN->getType()))
2752 dyn_cast<Instruction>(PN->getIncomingValueForBlock(L->getLoopLatch()));
2754 ChainInstruction(PN, IncV, ChainUsersVec);
2756 // Remove any unprofitable chains.
2757 unsigned ChainIdx = 0;
2758 for (unsigned UsersIdx = 0, NChains = IVChainVec.size();
2759 UsersIdx < NChains; ++UsersIdx) {
2760 if (!isProfitableChain(IVChainVec[UsersIdx],
2761 ChainUsersVec[UsersIdx].FarUsers, SE, TTI))
2763 // Preserve the chain at UsesIdx.
2764 if (ChainIdx != UsersIdx)
2765 IVChainVec[ChainIdx] = IVChainVec[UsersIdx];
2766 FinalizeChain(IVChainVec[ChainIdx]);
2769 IVChainVec.resize(ChainIdx);
2772 void LSRInstance::FinalizeChain(IVChain &Chain) {
2773 assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
2774 DEBUG(dbgs() << "Final Chain: " << *Chain.Incs[0].UserInst << "\n");
2776 for (IVChain::const_iterator I = Chain.begin(), E = Chain.end();
2778 DEBUG(dbgs() << " Inc: " << *I->UserInst << "\n");
2779 User::op_iterator UseI =
2780 std::find(I->UserInst->op_begin(), I->UserInst->op_end(), I->IVOperand);
2781 assert(UseI != I->UserInst->op_end() && "cannot find IV operand");
2782 IVIncSet.insert(UseI);
2786 /// Return true if the IVInc can be folded into an addressing mode.
2787 static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst,
2788 Value *Operand, const TargetTransformInfo &TTI) {
2789 const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr);
2790 if (!IncConst || !isAddressUse(UserInst, Operand))
2793 if (IncConst->getValue()->getValue().getMinSignedBits() > 64)
2796 int64_t IncOffset = IncConst->getValue()->getSExtValue();
2797 if (!isAlwaysFoldable(TTI, LSRUse::Address,
2798 getAccessType(UserInst), /*BaseGV=*/ 0,
2799 IncOffset, /*HaseBaseReg=*/ false))
2805 /// GenerateIVChains - Generate an add or subtract for each IVInc in a chain to
2806 /// materialize the IV user's operand from the previous IV user's operand.
2807 void LSRInstance::GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
2808 SmallVectorImpl<WeakVH> &DeadInsts) {
2809 // Find the new IVOperand for the head of the chain. It may have been replaced
2811 const IVInc &Head = Chain.Incs[0];
2812 User::op_iterator IVOpEnd = Head.UserInst->op_end();
2813 // findIVOperand returns IVOpEnd if it can no longer find a valid IV user.
2814 User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(),
2817 while (IVOpIter != IVOpEnd) {
2818 IVSrc = getWideOperand(*IVOpIter);
2820 // If this operand computes the expression that the chain needs, we may use
2821 // it. (Check this after setting IVSrc which is used below.)
2823 // Note that if Head.IncExpr is wider than IVSrc, then this phi is too
2824 // narrow for the chain, so we can no longer use it. We do allow using a
2825 // wider phi, assuming the LSR checked for free truncation. In that case we
2826 // should already have a truncate on this operand such that
2827 // getSCEV(IVSrc) == IncExpr.
2828 if (SE.getSCEV(*IVOpIter) == Head.IncExpr
2829 || SE.getSCEV(IVSrc) == Head.IncExpr) {
2832 IVOpIter = findIVOperand(llvm::next(IVOpIter), IVOpEnd, L, SE);
2834 if (IVOpIter == IVOpEnd) {
2835 // Gracefully give up on this chain.
2836 DEBUG(dbgs() << "Concealed chain head: " << *Head.UserInst << "\n");
2840 DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n");
2841 Type *IVTy = IVSrc->getType();
2842 Type *IntTy = SE.getEffectiveSCEVType(IVTy);
2843 const SCEV *LeftOverExpr = 0;
2844 for (IVChain::const_iterator IncI = Chain.begin(),
2845 IncE = Chain.end(); IncI != IncE; ++IncI) {
2847 Instruction *InsertPt = IncI->UserInst;
2848 if (isa<PHINode>(InsertPt))
2849 InsertPt = L->getLoopLatch()->getTerminator();
2851 // IVOper will replace the current IV User's operand. IVSrc is the IV
2852 // value currently held in a register.
2853 Value *IVOper = IVSrc;
2854 if (!IncI->IncExpr->isZero()) {
2855 // IncExpr was the result of subtraction of two narrow values, so must
2857 const SCEV *IncExpr = SE.getNoopOrSignExtend(IncI->IncExpr, IntTy);
2858 LeftOverExpr = LeftOverExpr ?
2859 SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr;
2861 if (LeftOverExpr && !LeftOverExpr->isZero()) {
2862 // Expand the IV increment.
2863 Rewriter.clearPostInc();
2864 Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt);
2865 const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc),
2866 SE.getUnknown(IncV));
2867 IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt);
2869 // If an IV increment can't be folded, use it as the next IV value.
2870 if (!canFoldIVIncExpr(LeftOverExpr, IncI->UserInst, IncI->IVOperand,
2872 assert(IVTy == IVOper->getType() && "inconsistent IV increment type");
2877 Type *OperTy = IncI->IVOperand->getType();
2878 if (IVTy != OperTy) {
2879 assert(SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) &&
2880 "cannot extend a chained IV");
2881 IRBuilder<> Builder(InsertPt);
2882 IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain");
2884 IncI->UserInst->replaceUsesOfWith(IncI->IVOperand, IVOper);
2885 DeadInsts.push_back(IncI->IVOperand);
2887 // If LSR created a new, wider phi, we may also replace its postinc. We only
2888 // do this if we also found a wide value for the head of the chain.
2889 if (isa<PHINode>(Chain.tailUserInst())) {
2890 for (BasicBlock::iterator I = L->getHeader()->begin();
2891 PHINode *Phi = dyn_cast<PHINode>(I); ++I) {
2892 if (!isCompatibleIVType(Phi, IVSrc))
2894 Instruction *PostIncV = dyn_cast<Instruction>(
2895 Phi->getIncomingValueForBlock(L->getLoopLatch()));
2896 if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc)))
2898 Value *IVOper = IVSrc;
2899 Type *PostIncTy = PostIncV->getType();
2900 if (IVTy != PostIncTy) {
2901 assert(PostIncTy->isPointerTy() && "mixing int/ptr IV types");
2902 IRBuilder<> Builder(L->getLoopLatch()->getTerminator());
2903 Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc());
2904 IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain");
2906 Phi->replaceUsesOfWith(PostIncV, IVOper);
2907 DeadInsts.push_back(PostIncV);
2912 void LSRInstance::CollectFixupsAndInitialFormulae() {
2913 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
2914 Instruction *UserInst = UI->getUser();
2915 // Skip IV users that are part of profitable IV Chains.
2916 User::op_iterator UseI = std::find(UserInst->op_begin(), UserInst->op_end(),
2917 UI->getOperandValToReplace());
2918 assert(UseI != UserInst->op_end() && "cannot find IV operand");
2919 if (IVIncSet.count(UseI))
2923 LSRFixup &LF = getNewFixup();
2924 LF.UserInst = UserInst;
2925 LF.OperandValToReplace = UI->getOperandValToReplace();
2926 LF.PostIncLoops = UI->getPostIncLoops();
2928 LSRUse::KindType Kind = LSRUse::Basic;
2930 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
2931 Kind = LSRUse::Address;
2932 AccessTy = getAccessType(LF.UserInst);
2935 const SCEV *S = IU.getExpr(*UI);
2937 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
2938 // (N - i == 0), and this allows (N - i) to be the expression that we work
2939 // with rather than just N or i, so we can consider the register
2940 // requirements for both N and i at the same time. Limiting this code to
2941 // equality icmps is not a problem because all interesting loops use
2942 // equality icmps, thanks to IndVarSimplify.
2943 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
2944 if (CI->isEquality()) {
2945 // Swap the operands if needed to put the OperandValToReplace on the
2946 // left, for consistency.
2947 Value *NV = CI->getOperand(1);
2948 if (NV == LF.OperandValToReplace) {
2949 CI->setOperand(1, CI->getOperand(0));
2950 CI->setOperand(0, NV);
2951 NV = CI->getOperand(1);
2955 // x == y --> x - y == 0
2956 const SCEV *N = SE.getSCEV(NV);
2957 if (SE.isLoopInvariant(N, L) && isSafeToExpand(N, SE)) {
2958 // S is normalized, so normalize N before folding it into S
2959 // to keep the result normalized.
2960 N = TransformForPostIncUse(Normalize, N, CI, 0,
2961 LF.PostIncLoops, SE, DT);
2962 Kind = LSRUse::ICmpZero;
2963 S = SE.getMinusSCEV(N, S);
2966 // -1 and the negations of all interesting strides (except the negation
2967 // of -1) are now also interesting.
2968 for (size_t i = 0, e = Factors.size(); i != e; ++i)
2969 if (Factors[i] != -1)
2970 Factors.insert(-(uint64_t)Factors[i]);
2974 // Set up the initial formula for this use.
2975 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
2977 LF.Offset = P.second;
2978 LSRUse &LU = Uses[LF.LUIdx];
2979 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
2980 if (!LU.WidestFixupType ||
2981 SE.getTypeSizeInBits(LU.WidestFixupType) <
2982 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
2983 LU.WidestFixupType = LF.OperandValToReplace->getType();
2985 // If this is the first use of this LSRUse, give it a formula.
2986 if (LU.Formulae.empty()) {
2987 InsertInitialFormula(S, LU, LF.LUIdx);
2988 CountRegisters(LU.Formulae.back(), LF.LUIdx);
2992 DEBUG(print_fixups(dbgs()));
2995 /// InsertInitialFormula - Insert a formula for the given expression into
2996 /// the given use, separating out loop-variant portions from loop-invariant
2997 /// and loop-computable portions.
2999 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
3000 // Mark uses whose expressions cannot be expanded.
3001 if (!isSafeToExpand(S, SE))
3002 LU.RigidFormula = true;
3005 F.InitialMatch(S, L, SE);
3006 bool Inserted = InsertFormula(LU, LUIdx, F);
3007 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
3010 /// InsertSupplementalFormula - Insert a simple single-register formula for
3011 /// the given expression into the given use.
3013 LSRInstance::InsertSupplementalFormula(const SCEV *S,
3014 LSRUse &LU, size_t LUIdx) {
3016 F.BaseRegs.push_back(S);
3017 F.HasBaseReg = true;
3018 bool Inserted = InsertFormula(LU, LUIdx, F);
3019 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
3022 /// CountRegisters - Note which registers are used by the given formula,
3023 /// updating RegUses.
3024 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
3026 RegUses.CountRegister(F.ScaledReg, LUIdx);
3027 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
3028 E = F.BaseRegs.end(); I != E; ++I)
3029 RegUses.CountRegister(*I, LUIdx);
3032 /// InsertFormula - If the given formula has not yet been inserted, add it to
3033 /// the list, and return true. Return false otherwise.
3034 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
3035 if (!LU.InsertFormula(F))
3038 CountRegisters(F, LUIdx);
3042 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of
3043 /// loop-invariant values which we're tracking. These other uses will pin these
3044 /// values in registers, making them less profitable for elimination.
3045 /// TODO: This currently misses non-constant addrec step registers.
3046 /// TODO: Should this give more weight to users inside the loop?
3048 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
3049 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
3050 SmallPtrSet<const SCEV *, 8> Inserted;
3052 while (!Worklist.empty()) {
3053 const SCEV *S = Worklist.pop_back_val();
3055 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
3056 Worklist.append(N->op_begin(), N->op_end());
3057 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
3058 Worklist.push_back(C->getOperand());
3059 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
3060 Worklist.push_back(D->getLHS());
3061 Worklist.push_back(D->getRHS());
3062 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
3063 if (!Inserted.insert(U)) continue;
3064 const Value *V = U->getValue();
3065 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
3066 // Look for instructions defined outside the loop.
3067 if (L->contains(Inst)) continue;
3068 } else if (isa<UndefValue>(V))
3069 // Undef doesn't have a live range, so it doesn't matter.
3071 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
3073 const Instruction *UserInst = dyn_cast<Instruction>(*UI);
3074 // Ignore non-instructions.
3077 // Ignore instructions in other functions (as can happen with
3079 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
3081 // Ignore instructions not dominated by the loop.
3082 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
3083 UserInst->getParent() :
3084 cast<PHINode>(UserInst)->getIncomingBlock(
3085 PHINode::getIncomingValueNumForOperand(UI.getOperandNo()));
3086 if (!DT.dominates(L->getHeader(), UseBB))
3088 // Ignore uses which are part of other SCEV expressions, to avoid
3089 // analyzing them multiple times.
3090 if (SE.isSCEVable(UserInst->getType())) {
3091 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
3092 // If the user is a no-op, look through to its uses.
3093 if (!isa<SCEVUnknown>(UserS))
3097 SE.getUnknown(const_cast<Instruction *>(UserInst)));
3101 // Ignore icmp instructions which are already being analyzed.
3102 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
3103 unsigned OtherIdx = !UI.getOperandNo();
3104 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
3105 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
3109 LSRFixup &LF = getNewFixup();
3110 LF.UserInst = const_cast<Instruction *>(UserInst);
3111 LF.OperandValToReplace = UI.getUse();
3112 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0);
3114 LF.Offset = P.second;
3115 LSRUse &LU = Uses[LF.LUIdx];
3116 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3117 if (!LU.WidestFixupType ||
3118 SE.getTypeSizeInBits(LU.WidestFixupType) <
3119 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3120 LU.WidestFixupType = LF.OperandValToReplace->getType();
3121 InsertSupplementalFormula(U, LU, LF.LUIdx);
3122 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
3129 /// CollectSubexprs - Split S into subexpressions which can be pulled out into
3130 /// separate registers. If C is non-null, multiply each subexpression by C.
3132 /// Return remainder expression after factoring the subexpressions captured by
3133 /// Ops. If Ops is complete, return NULL.
3134 static const SCEV *CollectSubexprs(const SCEV *S, const SCEVConstant *C,
3135 SmallVectorImpl<const SCEV *> &Ops,
3137 ScalarEvolution &SE,
3138 unsigned Depth = 0) {
3139 // Arbitrarily cap recursion to protect compile time.
3143 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3144 // Break out add operands.
3145 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
3147 const SCEV *Remainder = CollectSubexprs(*I, C, Ops, L, SE, Depth+1);
3149 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3152 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
3153 // Split a non-zero base out of an addrec.
3154 if (AR->getStart()->isZero())
3157 const SCEV *Remainder = CollectSubexprs(AR->getStart(),
3158 C, Ops, L, SE, Depth+1);
3159 // Split the non-zero AddRec unless it is part of a nested recurrence that
3160 // does not pertain to this loop.
3161 if (Remainder && (AR->getLoop() == L || !isa<SCEVAddRecExpr>(Remainder))) {
3162 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3165 if (Remainder != AR->getStart()) {
3167 Remainder = SE.getConstant(AR->getType(), 0);
3168 return SE.getAddRecExpr(Remainder,
3169 AR->getStepRecurrence(SE),
3171 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
3174 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3175 // Break (C * (a + b + c)) into C*a + C*b + C*c.
3176 if (Mul->getNumOperands() != 2)
3178 if (const SCEVConstant *Op0 =
3179 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
3180 C = C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0;
3181 const SCEV *Remainder =
3182 CollectSubexprs(Mul->getOperand(1), C, Ops, L, SE, Depth+1);
3184 Ops.push_back(SE.getMulExpr(C, Remainder));
3191 /// GenerateReassociations - Split out subexpressions from adds and the bases of
3193 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
3196 // Arbitrarily cap recursion to protect compile time.
3197 if (Depth >= 3) return;
3199 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3200 const SCEV *BaseReg = Base.BaseRegs[i];
3202 SmallVector<const SCEV *, 8> AddOps;
3203 const SCEV *Remainder = CollectSubexprs(BaseReg, 0, AddOps, L, SE);
3205 AddOps.push_back(Remainder);
3207 if (AddOps.size() == 1) continue;
3209 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
3210 JE = AddOps.end(); J != JE; ++J) {
3212 // Loop-variant "unknown" values are uninteresting; we won't be able to
3213 // do anything meaningful with them.
3214 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
3217 // Don't pull a constant into a register if the constant could be folded
3218 // into an immediate field.
3219 if (isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3220 LU.AccessTy, *J, Base.getNumRegs() > 1))
3223 // Collect all operands except *J.
3224 SmallVector<const SCEV *, 8> InnerAddOps
3225 (((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
3227 (llvm::next(J), ((const SmallVector<const SCEV *, 8> &)AddOps).end());
3229 // Don't leave just a constant behind in a register if the constant could
3230 // be folded into an immediate field.
3231 if (InnerAddOps.size() == 1 &&
3232 isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3233 LU.AccessTy, InnerAddOps[0], Base.getNumRegs() > 1))
3236 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
3237 if (InnerSum->isZero())
3241 // Add the remaining pieces of the add back into the new formula.
3242 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
3244 SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
3245 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3246 InnerSumSC->getValue()->getZExtValue())) {
3247 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
3248 InnerSumSC->getValue()->getZExtValue();
3249 F.BaseRegs.erase(F.BaseRegs.begin() + i);
3251 F.BaseRegs[i] = InnerSum;
3253 // Add J as its own register, or an unfolded immediate.
3254 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
3255 if (SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
3256 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3257 SC->getValue()->getZExtValue()))
3258 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
3259 SC->getValue()->getZExtValue();
3261 F.BaseRegs.push_back(*J);
3263 if (InsertFormula(LU, LUIdx, F))
3264 // If that formula hadn't been seen before, recurse to find more like
3266 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1);
3271 /// GenerateCombinations - Generate a formula consisting of all of the
3272 /// loop-dominating registers added into a single register.
3273 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
3275 // This method is only interesting on a plurality of registers.
3276 if (Base.BaseRegs.size() <= 1) return;
3280 SmallVector<const SCEV *, 4> Ops;
3281 for (SmallVectorImpl<const SCEV *>::const_iterator
3282 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
3283 const SCEV *BaseReg = *I;
3284 if (SE.properlyDominates(BaseReg, L->getHeader()) &&
3285 !SE.hasComputableLoopEvolution(BaseReg, L))
3286 Ops.push_back(BaseReg);
3288 F.BaseRegs.push_back(BaseReg);
3290 if (Ops.size() > 1) {
3291 const SCEV *Sum = SE.getAddExpr(Ops);
3292 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
3293 // opportunity to fold something. For now, just ignore such cases
3294 // rather than proceed with zero in a register.
3295 if (!Sum->isZero()) {
3296 F.BaseRegs.push_back(Sum);
3297 (void)InsertFormula(LU, LUIdx, F);
3302 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
3303 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
3305 // We can't add a symbolic offset if the address already contains one.
3306 if (Base.BaseGV) return;
3308 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3309 const SCEV *G = Base.BaseRegs[i];
3310 GlobalValue *GV = ExtractSymbol(G, SE);
3311 if (G->isZero() || !GV)
3315 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3318 (void)InsertFormula(LU, LUIdx, F);
3322 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
3323 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
3325 // TODO: For now, just add the min and max offset, because it usually isn't
3326 // worthwhile looking at everything inbetween.
3327 SmallVector<int64_t, 2> Worklist;
3328 Worklist.push_back(LU.MinOffset);
3329 if (LU.MaxOffset != LU.MinOffset)
3330 Worklist.push_back(LU.MaxOffset);
3332 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3333 const SCEV *G = Base.BaseRegs[i];
3335 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
3336 E = Worklist.end(); I != E; ++I) {
3338 F.BaseOffset = (uint64_t)Base.BaseOffset - *I;
3339 if (isLegalUse(TTI, LU.MinOffset - *I, LU.MaxOffset - *I, LU.Kind,
3341 // Add the offset to the base register.
3342 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), *I), G);
3343 // If it cancelled out, drop the base register, otherwise update it.
3344 if (NewG->isZero()) {
3345 std::swap(F.BaseRegs[i], F.BaseRegs.back());
3346 F.BaseRegs.pop_back();
3348 F.BaseRegs[i] = NewG;
3350 (void)InsertFormula(LU, LUIdx, F);
3354 int64_t Imm = ExtractImmediate(G, SE);
3355 if (G->isZero() || Imm == 0)
3358 F.BaseOffset = (uint64_t)F.BaseOffset + Imm;
3359 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3362 (void)InsertFormula(LU, LUIdx, F);
3366 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
3367 /// the comparison. For example, x == y -> x*c == y*c.
3368 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
3370 if (LU.Kind != LSRUse::ICmpZero) return;
3372 // Determine the integer type for the base formula.
3373 Type *IntTy = Base.getType();
3375 if (SE.getTypeSizeInBits(IntTy) > 64) return;
3377 // Don't do this if there is more than one offset.
3378 if (LU.MinOffset != LU.MaxOffset) return;
3380 assert(!Base.BaseGV && "ICmpZero use is not legal!");
3382 // Check each interesting stride.
3383 for (SmallSetVector<int64_t, 8>::const_iterator
3384 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3385 int64_t Factor = *I;
3387 // Check that the multiplication doesn't overflow.
3388 if (Base.BaseOffset == INT64_MIN && Factor == -1)
3390 int64_t NewBaseOffset = (uint64_t)Base.BaseOffset * Factor;
3391 if (NewBaseOffset / Factor != Base.BaseOffset)
3394 // Check that multiplying with the use offset doesn't overflow.
3395 int64_t Offset = LU.MinOffset;
3396 if (Offset == INT64_MIN && Factor == -1)
3398 Offset = (uint64_t)Offset * Factor;
3399 if (Offset / Factor != LU.MinOffset)
3403 F.BaseOffset = NewBaseOffset;
3405 // Check that this scale is legal.
3406 if (!isLegalUse(TTI, Offset, Offset, LU.Kind, LU.AccessTy, F))
3409 // Compensate for the use having MinOffset built into it.
3410 F.BaseOffset = (uint64_t)F.BaseOffset + Offset - LU.MinOffset;
3412 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3414 // Check that multiplying with each base register doesn't overflow.
3415 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
3416 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
3417 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
3421 // Check that multiplying with the scaled register doesn't overflow.
3423 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
3424 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
3428 // Check that multiplying with the unfolded offset doesn't overflow.
3429 if (F.UnfoldedOffset != 0) {
3430 if (F.UnfoldedOffset == INT64_MIN && Factor == -1)
3432 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
3433 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
3437 // If we make it here and it's legal, add it.
3438 (void)InsertFormula(LU, LUIdx, F);
3443 /// GenerateScales - Generate stride factor reuse formulae by making use of
3444 /// scaled-offset address modes, for example.
3445 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
3446 // Determine the integer type for the base formula.
3447 Type *IntTy = Base.getType();
3450 // If this Formula already has a scaled register, we can't add another one.
3451 if (Base.Scale != 0) return;
3453 // Check each interesting stride.
3454 for (SmallSetVector<int64_t, 8>::const_iterator
3455 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3456 int64_t Factor = *I;
3458 Base.Scale = Factor;
3459 Base.HasBaseReg = Base.BaseRegs.size() > 1;
3460 // Check whether this scale is going to be legal.
3461 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3463 // As a special-case, handle special out-of-loop Basic users specially.
3464 // TODO: Reconsider this special case.
3465 if (LU.Kind == LSRUse::Basic &&
3466 isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LSRUse::Special,
3467 LU.AccessTy, Base) &&
3468 LU.AllFixupsOutsideLoop)
3469 LU.Kind = LSRUse::Special;
3473 // For an ICmpZero, negating a solitary base register won't lead to
3475 if (LU.Kind == LSRUse::ICmpZero &&
3476 !Base.HasBaseReg && Base.BaseOffset == 0 && !Base.BaseGV)
3478 // For each addrec base reg, apply the scale, if possible.
3479 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3480 if (const SCEVAddRecExpr *AR =
3481 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
3482 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3483 if (FactorS->isZero())
3485 // Divide out the factor, ignoring high bits, since we'll be
3486 // scaling the value back up in the end.
3487 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
3488 // TODO: This could be optimized to avoid all the copying.
3490 F.ScaledReg = Quotient;
3491 F.DeleteBaseReg(F.BaseRegs[i]);
3492 (void)InsertFormula(LU, LUIdx, F);
3498 /// GenerateTruncates - Generate reuse formulae from different IV types.
3499 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
3500 // Don't bother truncating symbolic values.
3501 if (Base.BaseGV) return;
3503 // Determine the integer type for the base formula.
3504 Type *DstTy = Base.getType();
3506 DstTy = SE.getEffectiveSCEVType(DstTy);
3508 for (SmallSetVector<Type *, 4>::const_iterator
3509 I = Types.begin(), E = Types.end(); I != E; ++I) {
3511 if (SrcTy != DstTy && TTI.isTruncateFree(SrcTy, DstTy)) {
3514 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
3515 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
3516 JE = F.BaseRegs.end(); J != JE; ++J)
3517 *J = SE.getAnyExtendExpr(*J, SrcTy);
3519 // TODO: This assumes we've done basic processing on all uses and
3520 // have an idea what the register usage is.
3521 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
3524 (void)InsertFormula(LU, LUIdx, F);
3531 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
3532 /// defer modifications so that the search phase doesn't have to worry about
3533 /// the data structures moving underneath it.
3537 const SCEV *OrigReg;
3539 WorkItem(size_t LI, int64_t I, const SCEV *R)
3540 : LUIdx(LI), Imm(I), OrigReg(R) {}
3542 void print(raw_ostream &OS) const;
3548 void WorkItem::print(raw_ostream &OS) const {
3549 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
3550 << " , add offset " << Imm;
3553 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
3554 void WorkItem::dump() const {
3555 print(errs()); errs() << '\n';
3559 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
3560 /// distance apart and try to form reuse opportunities between them.
3561 void LSRInstance::GenerateCrossUseConstantOffsets() {
3562 // Group the registers by their value without any added constant offset.
3563 typedef std::map<int64_t, const SCEV *> ImmMapTy;
3564 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
3566 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
3567 SmallVector<const SCEV *, 8> Sequence;
3568 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
3570 const SCEV *Reg = *I;
3571 int64_t Imm = ExtractImmediate(Reg, SE);
3572 std::pair<RegMapTy::iterator, bool> Pair =
3573 Map.insert(std::make_pair(Reg, ImmMapTy()));
3575 Sequence.push_back(Reg);
3576 Pair.first->second.insert(std::make_pair(Imm, *I));
3577 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
3580 // Now examine each set of registers with the same base value. Build up
3581 // a list of work to do and do the work in a separate step so that we're
3582 // not adding formulae and register counts while we're searching.
3583 SmallVector<WorkItem, 32> WorkItems;
3584 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
3585 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
3586 E = Sequence.end(); I != E; ++I) {
3587 const SCEV *Reg = *I;
3588 const ImmMapTy &Imms = Map.find(Reg)->second;
3590 // It's not worthwhile looking for reuse if there's only one offset.
3591 if (Imms.size() == 1)
3594 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
3595 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3597 dbgs() << ' ' << J->first;
3600 // Examine each offset.
3601 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3603 const SCEV *OrigReg = J->second;
3605 int64_t JImm = J->first;
3606 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
3608 if (!isa<SCEVConstant>(OrigReg) &&
3609 UsedByIndicesMap[Reg].count() == 1) {
3610 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
3614 // Conservatively examine offsets between this orig reg a few selected
3616 ImmMapTy::const_iterator OtherImms[] = {
3617 Imms.begin(), prior(Imms.end()),
3618 Imms.lower_bound((Imms.begin()->first + prior(Imms.end())->first) / 2)
3620 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
3621 ImmMapTy::const_iterator M = OtherImms[i];
3622 if (M == J || M == JE) continue;
3624 // Compute the difference between the two.
3625 int64_t Imm = (uint64_t)JImm - M->first;
3626 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
3627 LUIdx = UsedByIndices.find_next(LUIdx))
3628 // Make a memo of this use, offset, and register tuple.
3629 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
3630 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
3637 UsedByIndicesMap.clear();
3638 UniqueItems.clear();
3640 // Now iterate through the worklist and add new formulae.
3641 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
3642 E = WorkItems.end(); I != E; ++I) {
3643 const WorkItem &WI = *I;
3644 size_t LUIdx = WI.LUIdx;
3645 LSRUse &LU = Uses[LUIdx];
3646 int64_t Imm = WI.Imm;
3647 const SCEV *OrigReg = WI.OrigReg;
3649 Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
3650 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
3651 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
3653 // TODO: Use a more targeted data structure.
3654 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
3655 const Formula &F = LU.Formulae[L];
3656 // Use the immediate in the scaled register.
3657 if (F.ScaledReg == OrigReg) {
3658 int64_t Offset = (uint64_t)F.BaseOffset + Imm * (uint64_t)F.Scale;
3659 // Don't create 50 + reg(-50).
3660 if (F.referencesReg(SE.getSCEV(
3661 ConstantInt::get(IntTy, -(uint64_t)Offset))))
3664 NewF.BaseOffset = Offset;
3665 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3668 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
3670 // If the new scale is a constant in a register, and adding the constant
3671 // value to the immediate would produce a value closer to zero than the
3672 // immediate itself, then the formula isn't worthwhile.
3673 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
3674 if (C->getValue()->isNegative() !=
3675 (NewF.BaseOffset < 0) &&
3676 (C->getValue()->getValue().abs() * APInt(BitWidth, F.Scale))
3677 .ule(abs64(NewF.BaseOffset)))
3681 (void)InsertFormula(LU, LUIdx, NewF);
3683 // Use the immediate in a base register.
3684 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
3685 const SCEV *BaseReg = F.BaseRegs[N];
3686 if (BaseReg != OrigReg)
3689 NewF.BaseOffset = (uint64_t)NewF.BaseOffset + Imm;
3690 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset,
3691 LU.Kind, LU.AccessTy, NewF)) {
3692 if (!TTI.isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
3695 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
3697 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
3699 // If the new formula has a constant in a register, and adding the
3700 // constant value to the immediate would produce a value closer to
3701 // zero than the immediate itself, then the formula isn't worthwhile.
3702 for (SmallVectorImpl<const SCEV *>::const_iterator
3703 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
3705 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
3706 if ((C->getValue()->getValue() + NewF.BaseOffset).abs().slt(
3707 abs64(NewF.BaseOffset)) &&
3708 (C->getValue()->getValue() +
3709 NewF.BaseOffset).countTrailingZeros() >=
3710 countTrailingZeros<uint64_t>(NewF.BaseOffset))
3714 (void)InsertFormula(LU, LUIdx, NewF);
3723 /// GenerateAllReuseFormulae - Generate formulae for each use.
3725 LSRInstance::GenerateAllReuseFormulae() {
3726 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
3727 // queries are more precise.
3728 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3729 LSRUse &LU = Uses[LUIdx];
3730 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3731 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
3732 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3733 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
3735 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3736 LSRUse &LU = Uses[LUIdx];
3737 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3738 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
3739 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3740 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
3741 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3742 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
3743 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3744 GenerateScales(LU, LUIdx, LU.Formulae[i]);
3746 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3747 LSRUse &LU = Uses[LUIdx];
3748 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
3749 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
3752 GenerateCrossUseConstantOffsets();
3754 DEBUG(dbgs() << "\n"
3755 "After generating reuse formulae:\n";
3756 print_uses(dbgs()));
3759 /// If there are multiple formulae with the same set of registers used
3760 /// by other uses, pick the best one and delete the others.
3761 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
3762 DenseSet<const SCEV *> VisitedRegs;
3763 SmallPtrSet<const SCEV *, 16> Regs;
3764 SmallPtrSet<const SCEV *, 16> LoserRegs;
3766 bool ChangedFormulae = false;
3769 // Collect the best formula for each unique set of shared registers. This
3770 // is reset for each use.
3771 typedef DenseMap<SmallVector<const SCEV *, 4>, size_t, UniquifierDenseMapInfo>
3773 BestFormulaeTy BestFormulae;
3775 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3776 LSRUse &LU = Uses[LUIdx];
3777 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
3780 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
3781 FIdx != NumForms; ++FIdx) {
3782 Formula &F = LU.Formulae[FIdx];
3784 // Some formulas are instant losers. For example, they may depend on
3785 // nonexistent AddRecs from other loops. These need to be filtered
3786 // immediately, otherwise heuristics could choose them over others leading
3787 // to an unsatisfactory solution. Passing LoserRegs into RateFormula here
3788 // avoids the need to recompute this information across formulae using the
3789 // same bad AddRec. Passing LoserRegs is also essential unless we remove
3790 // the corresponding bad register from the Regs set.
3793 CostF.RateFormula(TTI, F, Regs, VisitedRegs, L, LU.Offsets, SE, DT, LU,
3795 if (CostF.isLoser()) {
3796 // During initial formula generation, undesirable formulae are generated
3797 // by uses within other loops that have some non-trivial address mode or
3798 // use the postinc form of the IV. LSR needs to provide these formulae
3799 // as the basis of rediscovering the desired formula that uses an AddRec
3800 // corresponding to the existing phi. Once all formulae have been
3801 // generated, these initial losers may be pruned.
3802 DEBUG(dbgs() << " Filtering loser "; F.print(dbgs());
3806 SmallVector<const SCEV *, 4> Key;
3807 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
3808 JE = F.BaseRegs.end(); J != JE; ++J) {
3809 const SCEV *Reg = *J;
3810 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
3814 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
3815 Key.push_back(F.ScaledReg);
3816 // Unstable sort by host order ok, because this is only used for
3818 std::sort(Key.begin(), Key.end());
3820 std::pair<BestFormulaeTy::const_iterator, bool> P =
3821 BestFormulae.insert(std::make_pair(Key, FIdx));
3825 Formula &Best = LU.Formulae[P.first->second];
3829 CostBest.RateFormula(TTI, Best, Regs, VisitedRegs, L, LU.Offsets, SE,
3831 if (CostF < CostBest)
3833 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
3835 " in favor of formula "; Best.print(dbgs());
3839 ChangedFormulae = true;
3841 LU.DeleteFormula(F);
3847 // Now that we've filtered out some formulae, recompute the Regs set.
3849 LU.RecomputeRegs(LUIdx, RegUses);
3851 // Reset this to prepare for the next use.
3852 BestFormulae.clear();
3855 DEBUG(if (ChangedFormulae) {
3857 "After filtering out undesirable candidates:\n";
3862 // This is a rough guess that seems to work fairly well.
3863 static const size_t ComplexityLimit = UINT16_MAX;
3865 /// EstimateSearchSpaceComplexity - Estimate the worst-case number of
3866 /// solutions the solver might have to consider. It almost never considers
3867 /// this many solutions because it prune the search space, but the pruning
3868 /// isn't always sufficient.
3869 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
3871 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3872 E = Uses.end(); I != E; ++I) {
3873 size_t FSize = I->Formulae.size();
3874 if (FSize >= ComplexityLimit) {
3875 Power = ComplexityLimit;
3879 if (Power >= ComplexityLimit)
3885 /// NarrowSearchSpaceByDetectingSupersets - When one formula uses a superset
3886 /// of the registers of another formula, it won't help reduce register
3887 /// pressure (though it may not necessarily hurt register pressure); remove
3888 /// it to simplify the system.
3889 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
3890 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
3891 DEBUG(dbgs() << "The search space is too complex.\n");
3893 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
3894 "which use a superset of registers used by other "
3897 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3898 LSRUse &LU = Uses[LUIdx];
3900 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
3901 Formula &F = LU.Formulae[i];
3902 // Look for a formula with a constant or GV in a register. If the use
3903 // also has a formula with that same value in an immediate field,
3904 // delete the one that uses a register.
3905 for (SmallVectorImpl<const SCEV *>::const_iterator
3906 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
3907 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
3909 NewF.BaseOffset += C->getValue()->getSExtValue();
3910 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
3911 (I - F.BaseRegs.begin()));
3912 if (LU.HasFormulaWithSameRegs(NewF)) {
3913 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
3914 LU.DeleteFormula(F);
3920 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
3921 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
3925 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
3926 (I - F.BaseRegs.begin()));
3927 if (LU.HasFormulaWithSameRegs(NewF)) {
3928 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
3930 LU.DeleteFormula(F);
3941 LU.RecomputeRegs(LUIdx, RegUses);
3944 DEBUG(dbgs() << "After pre-selection:\n";
3945 print_uses(dbgs()));
3949 /// NarrowSearchSpaceByCollapsingUnrolledCode - When there are many registers
3950 /// for expressions like A, A+1, A+2, etc., allocate a single register for
3952 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
3953 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
3956 DEBUG(dbgs() << "The search space is too complex.\n"
3957 "Narrowing the search space by assuming that uses separated "
3958 "by a constant offset will use the same registers.\n");
3960 // This is especially useful for unrolled loops.
3962 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
3963 LSRUse &LU = Uses[LUIdx];
3964 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
3965 E = LU.Formulae.end(); I != E; ++I) {
3966 const Formula &F = *I;
3967 if (F.BaseOffset == 0 || F.Scale != 0)
3970 LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU);
3974 if (!reconcileNewOffset(*LUThatHas, F.BaseOffset, /*HasBaseReg=*/ false,
3975 LU.Kind, LU.AccessTy))
3978 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs()); dbgs() << '\n');
3980 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
3982 // Update the relocs to reference the new use.
3983 for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(),
3984 E = Fixups.end(); I != E; ++I) {
3985 LSRFixup &Fixup = *I;
3986 if (Fixup.LUIdx == LUIdx) {
3987 Fixup.LUIdx = LUThatHas - &Uses.front();
3988 Fixup.Offset += F.BaseOffset;
3989 // Add the new offset to LUThatHas' offset list.
3990 if (LUThatHas->Offsets.back() != Fixup.Offset) {
3991 LUThatHas->Offsets.push_back(Fixup.Offset);
3992 if (Fixup.Offset > LUThatHas->MaxOffset)
3993 LUThatHas->MaxOffset = Fixup.Offset;
3994 if (Fixup.Offset < LUThatHas->MinOffset)
3995 LUThatHas->MinOffset = Fixup.Offset;
3997 DEBUG(dbgs() << "New fixup has offset " << Fixup.Offset << '\n');
3999 if (Fixup.LUIdx == NumUses-1)
4000 Fixup.LUIdx = LUIdx;
4003 // Delete formulae from the new use which are no longer legal.
4005 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
4006 Formula &F = LUThatHas->Formulae[i];
4007 if (!isLegalUse(TTI, LUThatHas->MinOffset, LUThatHas->MaxOffset,
4008 LUThatHas->Kind, LUThatHas->AccessTy, F)) {
4009 DEBUG(dbgs() << " Deleting "; F.print(dbgs());
4011 LUThatHas->DeleteFormula(F);
4019 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
4021 // Delete the old use.
4022 DeleteUse(LU, LUIdx);
4029 DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
4032 /// NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters - Call
4033 /// FilterOutUndesirableDedicatedRegisters again, if necessary, now that
4034 /// we've done more filtering, as it may be able to find more formulae to
4036 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
4037 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4038 DEBUG(dbgs() << "The search space is too complex.\n");
4040 DEBUG(dbgs() << "Narrowing the search space by re-filtering out "
4041 "undesirable dedicated registers.\n");
4043 FilterOutUndesirableDedicatedRegisters();
4045 DEBUG(dbgs() << "After pre-selection:\n";
4046 print_uses(dbgs()));
4050 /// NarrowSearchSpaceByPickingWinnerRegs - Pick a register which seems likely
4051 /// to be profitable, and then in any use which has any reference to that
4052 /// register, delete all formulae which do not reference that register.
4053 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
4054 // With all other options exhausted, loop until the system is simple
4055 // enough to handle.
4056 SmallPtrSet<const SCEV *, 4> Taken;
4057 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4058 // Ok, we have too many of formulae on our hands to conveniently handle.
4059 // Use a rough heuristic to thin out the list.
4060 DEBUG(dbgs() << "The search space is too complex.\n");
4062 // Pick the register which is used by the most LSRUses, which is likely
4063 // to be a good reuse register candidate.
4064 const SCEV *Best = 0;
4065 unsigned BestNum = 0;
4066 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
4068 const SCEV *Reg = *I;
4069 if (Taken.count(Reg))
4074 unsigned Count = RegUses.getUsedByIndices(Reg).count();
4075 if (Count > BestNum) {
4082 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
4083 << " will yield profitable reuse.\n");
4086 // In any use with formulae which references this register, delete formulae
4087 // which don't reference it.
4088 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4089 LSRUse &LU = Uses[LUIdx];
4090 if (!LU.Regs.count(Best)) continue;
4093 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4094 Formula &F = LU.Formulae[i];
4095 if (!F.referencesReg(Best)) {
4096 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
4097 LU.DeleteFormula(F);
4101 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
4107 LU.RecomputeRegs(LUIdx, RegUses);
4110 DEBUG(dbgs() << "After pre-selection:\n";
4111 print_uses(dbgs()));
4115 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
4116 /// formulae to choose from, use some rough heuristics to prune down the number
4117 /// of formulae. This keeps the main solver from taking an extraordinary amount
4118 /// of time in some worst-case scenarios.
4119 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
4120 NarrowSearchSpaceByDetectingSupersets();
4121 NarrowSearchSpaceByCollapsingUnrolledCode();
4122 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
4123 NarrowSearchSpaceByPickingWinnerRegs();
4126 /// SolveRecurse - This is the recursive solver.
4127 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
4129 SmallVectorImpl<const Formula *> &Workspace,
4130 const Cost &CurCost,
4131 const SmallPtrSet<const SCEV *, 16> &CurRegs,
4132 DenseSet<const SCEV *> &VisitedRegs) const {
4135 // - use more aggressive filtering
4136 // - sort the formula so that the most profitable solutions are found first
4137 // - sort the uses too
4139 // - don't compute a cost, and then compare. compare while computing a cost
4141 // - track register sets with SmallBitVector
4143 const LSRUse &LU = Uses[Workspace.size()];
4145 // If this use references any register that's already a part of the
4146 // in-progress solution, consider it a requirement that a formula must
4147 // reference that register in order to be considered. This prunes out
4148 // unprofitable searching.
4149 SmallSetVector<const SCEV *, 4> ReqRegs;
4150 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(),
4151 E = CurRegs.end(); I != E; ++I)
4152 if (LU.Regs.count(*I))
4155 SmallPtrSet<const SCEV *, 16> NewRegs;
4157 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
4158 E = LU.Formulae.end(); I != E; ++I) {
4159 const Formula &F = *I;
4161 // Ignore formulae which do not use any of the required registers.
4162 bool SatisfiedReqReg = true;
4163 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
4164 JE = ReqRegs.end(); J != JE; ++J) {
4165 const SCEV *Reg = *J;
4166 if ((!F.ScaledReg || F.ScaledReg != Reg) &&
4167 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) ==
4169 SatisfiedReqReg = false;
4173 if (!SatisfiedReqReg) {
4174 // If none of the formulae satisfied the required registers, then we could
4175 // clear ReqRegs and try again. Currently, we simply give up in this case.
4179 // Evaluate the cost of the current formula. If it's already worse than
4180 // the current best, prune the search at that point.
4183 NewCost.RateFormula(TTI, F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT,
4185 if (NewCost < SolutionCost) {
4186 Workspace.push_back(&F);
4187 if (Workspace.size() != Uses.size()) {
4188 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
4189 NewRegs, VisitedRegs);
4190 if (F.getNumRegs() == 1 && Workspace.size() == 1)
4191 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
4193 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
4194 dbgs() << ".\n Regs:";
4195 for (SmallPtrSet<const SCEV *, 16>::const_iterator
4196 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I)
4197 dbgs() << ' ' << **I;
4200 SolutionCost = NewCost;
4201 Solution = Workspace;
4203 Workspace.pop_back();
4208 /// Solve - Choose one formula from each use. Return the results in the given
4209 /// Solution vector.
4210 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
4211 SmallVector<const Formula *, 8> Workspace;
4213 SolutionCost.Loose();
4215 SmallPtrSet<const SCEV *, 16> CurRegs;
4216 DenseSet<const SCEV *> VisitedRegs;
4217 Workspace.reserve(Uses.size());
4219 // SolveRecurse does all the work.
4220 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
4221 CurRegs, VisitedRegs);
4222 if (Solution.empty()) {
4223 DEBUG(dbgs() << "\nNo Satisfactory Solution\n");
4227 // Ok, we've now made all our decisions.
4228 DEBUG(dbgs() << "\n"
4229 "The chosen solution requires "; SolutionCost.print(dbgs());
4231 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
4233 Uses[i].print(dbgs());
4236 Solution[i]->print(dbgs());
4240 assert(Solution.size() == Uses.size() && "Malformed solution!");
4243 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
4244 /// the dominator tree far as we can go while still being dominated by the
4245 /// input positions. This helps canonicalize the insert position, which
4246 /// encourages sharing.
4247 BasicBlock::iterator
4248 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
4249 const SmallVectorImpl<Instruction *> &Inputs)
4252 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
4253 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
4256 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
4257 if (!Rung) return IP;
4258 Rung = Rung->getIDom();
4259 if (!Rung) return IP;
4260 IDom = Rung->getBlock();
4262 // Don't climb into a loop though.
4263 const Loop *IDomLoop = LI.getLoopFor(IDom);
4264 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
4265 if (IDomDepth <= IPLoopDepth &&
4266 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
4270 bool AllDominate = true;
4271 Instruction *BetterPos = 0;
4272 Instruction *Tentative = IDom->getTerminator();
4273 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
4274 E = Inputs.end(); I != E; ++I) {
4275 Instruction *Inst = *I;
4276 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
4277 AllDominate = false;
4280 // Attempt to find an insert position in the middle of the block,
4281 // instead of at the end, so that it can be used for other expansions.
4282 if (IDom == Inst->getParent() &&
4283 (!BetterPos || !DT.dominates(Inst, BetterPos)))
4284 BetterPos = llvm::next(BasicBlock::iterator(Inst));
4297 /// AdjustInsertPositionForExpand - Determine an input position which will be
4298 /// dominated by the operands and which will dominate the result.
4299 BasicBlock::iterator
4300 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator LowestIP,
4303 SCEVExpander &Rewriter) const {
4304 // Collect some instructions which must be dominated by the
4305 // expanding replacement. These must be dominated by any operands that
4306 // will be required in the expansion.
4307 SmallVector<Instruction *, 4> Inputs;
4308 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
4309 Inputs.push_back(I);
4310 if (LU.Kind == LSRUse::ICmpZero)
4311 if (Instruction *I =
4312 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
4313 Inputs.push_back(I);
4314 if (LF.PostIncLoops.count(L)) {
4315 if (LF.isUseFullyOutsideLoop(L))
4316 Inputs.push_back(L->getLoopLatch()->getTerminator());
4318 Inputs.push_back(IVIncInsertPos);
4320 // The expansion must also be dominated by the increment positions of any
4321 // loops it for which it is using post-inc mode.
4322 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
4323 E = LF.PostIncLoops.end(); I != E; ++I) {
4324 const Loop *PIL = *I;
4325 if (PIL == L) continue;
4327 // Be dominated by the loop exit.
4328 SmallVector<BasicBlock *, 4> ExitingBlocks;
4329 PIL->getExitingBlocks(ExitingBlocks);
4330 if (!ExitingBlocks.empty()) {
4331 BasicBlock *BB = ExitingBlocks[0];
4332 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
4333 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
4334 Inputs.push_back(BB->getTerminator());
4338 assert(!isa<PHINode>(LowestIP) && !isa<LandingPadInst>(LowestIP)
4339 && !isa<DbgInfoIntrinsic>(LowestIP) &&
4340 "Insertion point must be a normal instruction");
4342 // Then, climb up the immediate dominator tree as far as we can go while
4343 // still being dominated by the input positions.
4344 BasicBlock::iterator IP = HoistInsertPosition(LowestIP, Inputs);
4346 // Don't insert instructions before PHI nodes.
4347 while (isa<PHINode>(IP)) ++IP;
4349 // Ignore landingpad instructions.
4350 while (isa<LandingPadInst>(IP)) ++IP;
4352 // Ignore debug intrinsics.
4353 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
4355 // Set IP below instructions recently inserted by SCEVExpander. This keeps the
4356 // IP consistent across expansions and allows the previously inserted
4357 // instructions to be reused by subsequent expansion.
4358 while (Rewriter.isInsertedInstruction(IP) && IP != LowestIP) ++IP;
4363 /// Expand - Emit instructions for the leading candidate expression for this
4364 /// LSRUse (this is called "expanding").
4365 Value *LSRInstance::Expand(const LSRFixup &LF,
4367 BasicBlock::iterator IP,
4368 SCEVExpander &Rewriter,
4369 SmallVectorImpl<WeakVH> &DeadInsts) const {
4370 const LSRUse &LU = Uses[LF.LUIdx];
4371 if (LU.RigidFormula)
4372 return LF.OperandValToReplace;
4374 // Determine an input position which will be dominated by the operands and
4375 // which will dominate the result.
4376 IP = AdjustInsertPositionForExpand(IP, LF, LU, Rewriter);
4378 // Inform the Rewriter if we have a post-increment use, so that it can
4379 // perform an advantageous expansion.
4380 Rewriter.setPostInc(LF.PostIncLoops);
4382 // This is the type that the user actually needs.
4383 Type *OpTy = LF.OperandValToReplace->getType();
4384 // This will be the type that we'll initially expand to.
4385 Type *Ty = F.getType();
4387 // No type known; just expand directly to the ultimate type.
4389 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
4390 // Expand directly to the ultimate type if it's the right size.
4392 // This is the type to do integer arithmetic in.
4393 Type *IntTy = SE.getEffectiveSCEVType(Ty);
4395 // Build up a list of operands to add together to form the full base.
4396 SmallVector<const SCEV *, 8> Ops;
4398 // Expand the BaseRegs portion.
4399 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
4400 E = F.BaseRegs.end(); I != E; ++I) {
4401 const SCEV *Reg = *I;
4402 assert(!Reg->isZero() && "Zero allocated in a base register!");
4404 // If we're expanding for a post-inc user, make the post-inc adjustment.
4405 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4406 Reg = TransformForPostIncUse(Denormalize, Reg,
4407 LF.UserInst, LF.OperandValToReplace,
4410 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP)));
4413 // Expand the ScaledReg portion.
4414 Value *ICmpScaledV = 0;
4416 const SCEV *ScaledS = F.ScaledReg;
4418 // If we're expanding for a post-inc user, make the post-inc adjustment.
4419 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4420 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
4421 LF.UserInst, LF.OperandValToReplace,
4424 if (LU.Kind == LSRUse::ICmpZero) {
4425 // An interesting way of "folding" with an icmp is to use a negated
4426 // scale, which we'll implement by inserting it into the other operand
4428 assert(F.Scale == -1 &&
4429 "The only scale supported by ICmpZero uses is -1!");
4430 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP);
4432 // Otherwise just expand the scaled register and an explicit scale,
4433 // which is expected to be matched as part of the address.
4435 // Flush the operand list to suppress SCEVExpander hoisting address modes.
4436 if (!Ops.empty() && LU.Kind == LSRUse::Address) {
4437 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4439 Ops.push_back(SE.getUnknown(FullV));
4441 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP));
4442 ScaledS = SE.getMulExpr(ScaledS,
4443 SE.getConstant(ScaledS->getType(), F.Scale));
4444 Ops.push_back(ScaledS);
4448 // Expand the GV portion.
4450 // Flush the operand list to suppress SCEVExpander hoisting.
4452 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4454 Ops.push_back(SE.getUnknown(FullV));
4456 Ops.push_back(SE.getUnknown(F.BaseGV));
4459 // Flush the operand list to suppress SCEVExpander hoisting of both folded and
4460 // unfolded offsets. LSR assumes they both live next to their uses.
4462 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
4464 Ops.push_back(SE.getUnknown(FullV));
4467 // Expand the immediate portion.
4468 int64_t Offset = (uint64_t)F.BaseOffset + LF.Offset;
4470 if (LU.Kind == LSRUse::ICmpZero) {
4471 // The other interesting way of "folding" with an ICmpZero is to use a
4472 // negated immediate.
4474 ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset);
4476 Ops.push_back(SE.getUnknown(ICmpScaledV));
4477 ICmpScaledV = ConstantInt::get(IntTy, Offset);
4480 // Just add the immediate values. These again are expected to be matched
4481 // as part of the address.
4482 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
4486 // Expand the unfolded offset portion.
4487 int64_t UnfoldedOffset = F.UnfoldedOffset;
4488 if (UnfoldedOffset != 0) {
4489 // Just add the immediate values.
4490 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy,
4494 // Emit instructions summing all the operands.
4495 const SCEV *FullS = Ops.empty() ?
4496 SE.getConstant(IntTy, 0) :
4498 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
4500 // We're done expanding now, so reset the rewriter.
4501 Rewriter.clearPostInc();
4503 // An ICmpZero Formula represents an ICmp which we're handling as a
4504 // comparison against zero. Now that we've expanded an expression for that
4505 // form, update the ICmp's other operand.
4506 if (LU.Kind == LSRUse::ICmpZero) {
4507 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
4508 DeadInsts.push_back(CI->getOperand(1));
4509 assert(!F.BaseGV && "ICmp does not support folding a global value and "
4510 "a scale at the same time!");
4511 if (F.Scale == -1) {
4512 if (ICmpScaledV->getType() != OpTy) {
4514 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
4516 ICmpScaledV, OpTy, "tmp", CI);
4519 CI->setOperand(1, ICmpScaledV);
4521 assert(F.Scale == 0 &&
4522 "ICmp does not support folding a global value and "
4523 "a scale at the same time!");
4524 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
4526 if (C->getType() != OpTy)
4527 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
4531 CI->setOperand(1, C);
4538 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use
4539 /// of their operands effectively happens in their predecessor blocks, so the
4540 /// expression may need to be expanded in multiple places.
4541 void LSRInstance::RewriteForPHI(PHINode *PN,
4544 SCEVExpander &Rewriter,
4545 SmallVectorImpl<WeakVH> &DeadInsts,
4547 DenseMap<BasicBlock *, Value *> Inserted;
4548 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
4549 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
4550 BasicBlock *BB = PN->getIncomingBlock(i);
4552 // If this is a critical edge, split the edge so that we do not insert
4553 // the code on all predecessor/successor paths. We do this unless this
4554 // is the canonical backedge for this loop, which complicates post-inc
4556 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
4557 !isa<IndirectBrInst>(BB->getTerminator())) {
4558 BasicBlock *Parent = PN->getParent();
4559 Loop *PNLoop = LI.getLoopFor(Parent);
4560 if (!PNLoop || Parent != PNLoop->getHeader()) {
4561 // Split the critical edge.
4562 BasicBlock *NewBB = 0;
4563 if (!Parent->isLandingPad()) {
4564 NewBB = SplitCriticalEdge(BB, Parent, P,
4565 /*MergeIdenticalEdges=*/true,
4566 /*DontDeleteUselessPhis=*/true);
4568 SmallVector<BasicBlock*, 2> NewBBs;
4569 SplitLandingPadPredecessors(Parent, BB, "", "", P, NewBBs);
4572 // If NewBB==NULL, then SplitCriticalEdge refused to split because all
4573 // phi predecessors are identical. The simple thing to do is skip
4574 // splitting in this case rather than complicate the API.
4576 // If PN is outside of the loop and BB is in the loop, we want to
4577 // move the block to be immediately before the PHI block, not
4578 // immediately after BB.
4579 if (L->contains(BB) && !L->contains(PN))
4580 NewBB->moveBefore(PN->getParent());
4582 // Splitting the edge can reduce the number of PHI entries we have.
4583 e = PN->getNumIncomingValues();
4585 i = PN->getBasicBlockIndex(BB);
4590 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
4591 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0)));
4593 PN->setIncomingValue(i, Pair.first->second);
4595 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts);
4597 // If this is reuse-by-noop-cast, insert the noop cast.
4598 Type *OpTy = LF.OperandValToReplace->getType();
4599 if (FullV->getType() != OpTy)
4601 CastInst::Create(CastInst::getCastOpcode(FullV, false,
4603 FullV, LF.OperandValToReplace->getType(),
4604 "tmp", BB->getTerminator());
4606 PN->setIncomingValue(i, FullV);
4607 Pair.first->second = FullV;
4612 /// Rewrite - Emit instructions for the leading candidate expression for this
4613 /// LSRUse (this is called "expanding"), and update the UserInst to reference
4614 /// the newly expanded value.
4615 void LSRInstance::Rewrite(const LSRFixup &LF,
4617 SCEVExpander &Rewriter,
4618 SmallVectorImpl<WeakVH> &DeadInsts,
4620 // First, find an insertion point that dominates UserInst. For PHI nodes,
4621 // find the nearest block which dominates all the relevant uses.
4622 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
4623 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P);
4625 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts);
4627 // If this is reuse-by-noop-cast, insert the noop cast.
4628 Type *OpTy = LF.OperandValToReplace->getType();
4629 if (FullV->getType() != OpTy) {
4631 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
4632 FullV, OpTy, "tmp", LF.UserInst);
4636 // Update the user. ICmpZero is handled specially here (for now) because
4637 // Expand may have updated one of the operands of the icmp already, and
4638 // its new value may happen to be equal to LF.OperandValToReplace, in
4639 // which case doing replaceUsesOfWith leads to replacing both operands
4640 // with the same value. TODO: Reorganize this.
4641 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
4642 LF.UserInst->setOperand(0, FullV);
4644 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
4647 DeadInsts.push_back(LF.OperandValToReplace);
4650 /// ImplementSolution - Rewrite all the fixup locations with new values,
4651 /// following the chosen solution.
4653 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
4655 // Keep track of instructions we may have made dead, so that
4656 // we can remove them after we are done working.
4657 SmallVector<WeakVH, 16> DeadInsts;
4659 SCEVExpander Rewriter(SE, "lsr");
4661 Rewriter.setDebugType(DEBUG_TYPE);
4663 Rewriter.disableCanonicalMode();
4664 Rewriter.enableLSRMode();
4665 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
4667 // Mark phi nodes that terminate chains so the expander tries to reuse them.
4668 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(),
4669 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) {
4670 if (PHINode *PN = dyn_cast<PHINode>(ChainI->tailUserInst()))
4671 Rewriter.setChainedPhi(PN);
4674 // Expand the new value definitions and update the users.
4675 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
4676 E = Fixups.end(); I != E; ++I) {
4677 const LSRFixup &Fixup = *I;
4679 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P);
4684 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(),
4685 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) {
4686 GenerateIVChain(*ChainI, Rewriter, DeadInsts);
4689 // Clean up after ourselves. This must be done before deleting any
4693 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
4696 LSRInstance::LSRInstance(Loop *L, Pass *P)
4697 : IU(P->getAnalysis<IVUsers>()), SE(P->getAnalysis<ScalarEvolution>()),
4698 DT(P->getAnalysis<DominatorTree>()), LI(P->getAnalysis<LoopInfo>()),
4699 TTI(P->getAnalysis<TargetTransformInfo>()), L(L), Changed(false),
4701 // If LoopSimplify form is not available, stay out of trouble.
4702 if (!L->isLoopSimplifyForm())
4705 // If there's no interesting work to be done, bail early.
4706 if (IU.empty()) return;
4708 // If there's too much analysis to be done, bail early. We won't be able to
4709 // model the problem anyway.
4710 unsigned NumUsers = 0;
4711 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
4712 if (++NumUsers > MaxIVUsers) {
4713 DEBUG(dbgs() << "LSR skipping loop, too many IV Users in " << *L
4720 // All dominating loops must have preheaders, or SCEVExpander may not be able
4721 // to materialize an AddRecExpr whose Start is an outer AddRecExpr.
4723 // IVUsers analysis should only create users that are dominated by simple loop
4724 // headers. Since this loop should dominate all of its users, its user list
4725 // should be empty if this loop itself is not within a simple loop nest.
4726 for (DomTreeNode *Rung = DT.getNode(L->getLoopPreheader());
4727 Rung; Rung = Rung->getIDom()) {
4728 BasicBlock *BB = Rung->getBlock();
4729 const Loop *DomLoop = LI.getLoopFor(BB);
4730 if (DomLoop && DomLoop->getHeader() == BB) {
4731 assert(DomLoop->getLoopPreheader() && "LSR needs a simplified loop nest");
4736 DEBUG(dbgs() << "\nLSR on loop ";
4737 WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false);
4740 // First, perform some low-level loop optimizations.
4742 OptimizeLoopTermCond();
4744 // If loop preparation eliminates all interesting IV users, bail.
4745 if (IU.empty()) return;
4747 // Skip nested loops until we can model them better with formulae.
4749 DEBUG(dbgs() << "LSR skipping outer loop " << *L << "\n");
4753 // Start collecting data and preparing for the solver.
4755 CollectInterestingTypesAndFactors();
4756 CollectFixupsAndInitialFormulae();
4757 CollectLoopInvariantFixupsAndFormulae();
4759 assert(!Uses.empty() && "IVUsers reported at least one use");
4760 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
4761 print_uses(dbgs()));
4763 // Now use the reuse data to generate a bunch of interesting ways
4764 // to formulate the values needed for the uses.
4765 GenerateAllReuseFormulae();
4767 FilterOutUndesirableDedicatedRegisters();
4768 NarrowSearchSpaceUsingHeuristics();
4770 SmallVector<const Formula *, 8> Solution;
4773 // Release memory that is no longer needed.
4778 if (Solution.empty())
4782 // Formulae should be legal.
4783 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(), E = Uses.end();
4785 const LSRUse &LU = *I;
4786 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
4787 JE = LU.Formulae.end();
4789 assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
4790 *J) && "Illegal formula generated!");
4794 // Now that we've decided what we want, make it so.
4795 ImplementSolution(Solution, P);
4798 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
4799 if (Factors.empty() && Types.empty()) return;
4801 OS << "LSR has identified the following interesting factors and types: ";
4804 for (SmallSetVector<int64_t, 8>::const_iterator
4805 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
4806 if (!First) OS << ", ";
4811 for (SmallSetVector<Type *, 4>::const_iterator
4812 I = Types.begin(), E = Types.end(); I != E; ++I) {
4813 if (!First) OS << ", ";
4815 OS << '(' << **I << ')';
4820 void LSRInstance::print_fixups(raw_ostream &OS) const {
4821 OS << "LSR is examining the following fixup sites:\n";
4822 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
4823 E = Fixups.end(); I != E; ++I) {
4830 void LSRInstance::print_uses(raw_ostream &OS) const {
4831 OS << "LSR is examining the following uses:\n";
4832 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
4833 E = Uses.end(); I != E; ++I) {
4834 const LSRUse &LU = *I;
4838 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
4839 JE = LU.Formulae.end(); J != JE; ++J) {
4847 void LSRInstance::print(raw_ostream &OS) const {
4848 print_factors_and_types(OS);
4853 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
4854 void LSRInstance::dump() const {
4855 print(errs()); errs() << '\n';
4861 class LoopStrengthReduce : public LoopPass {
4863 static char ID; // Pass ID, replacement for typeid
4864 LoopStrengthReduce();
4867 bool runOnLoop(Loop *L, LPPassManager &LPM);
4868 void getAnalysisUsage(AnalysisUsage &AU) const;
4873 char LoopStrengthReduce::ID = 0;
4874 INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce",
4875 "Loop Strength Reduction", false, false)
4876 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
4877 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
4878 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
4879 INITIALIZE_PASS_DEPENDENCY(IVUsers)
4880 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
4881 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
4882 INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce",
4883 "Loop Strength Reduction", false, false)
4886 Pass *llvm::createLoopStrengthReducePass() {
4887 return new LoopStrengthReduce();
4890 LoopStrengthReduce::LoopStrengthReduce() : LoopPass(ID) {
4891 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry());
4894 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
4895 // We split critical edges, so we change the CFG. However, we do update
4896 // many analyses if they are around.
4897 AU.addPreservedID(LoopSimplifyID);
4899 AU.addRequired<LoopInfo>();
4900 AU.addPreserved<LoopInfo>();
4901 AU.addRequiredID(LoopSimplifyID);
4902 AU.addRequired<DominatorTree>();
4903 AU.addPreserved<DominatorTree>();
4904 AU.addRequired<ScalarEvolution>();
4905 AU.addPreserved<ScalarEvolution>();
4906 // Requiring LoopSimplify a second time here prevents IVUsers from running
4907 // twice, since LoopSimplify was invalidated by running ScalarEvolution.
4908 AU.addRequiredID(LoopSimplifyID);
4909 AU.addRequired<IVUsers>();
4910 AU.addPreserved<IVUsers>();
4911 AU.addRequired<TargetTransformInfo>();
4914 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
4915 bool Changed = false;
4917 // Run the main LSR transformation.
4918 Changed |= LSRInstance(L, this).getChanged();
4920 // Remove any extra phis created by processing inner loops.
4921 Changed |= DeleteDeadPHIs(L->getHeader());
4922 if (EnablePhiElim && L->isLoopSimplifyForm()) {
4923 SmallVector<WeakVH, 16> DeadInsts;
4924 SCEVExpander Rewriter(getAnalysis<ScalarEvolution>(), "lsr");
4926 Rewriter.setDebugType(DEBUG_TYPE);
4928 unsigned numFolded =
4929 Rewriter.replaceCongruentIVs(L, &getAnalysis<DominatorTree>(),
4931 &getAnalysis<TargetTransformInfo>());
4934 DeleteTriviallyDeadInstructions(DeadInsts);
4935 DeleteDeadPHIs(L->getHeader());