1 //===- LoopStrengthReduce.cpp - Strength Reduce IVs in Loops --------------===//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
9 // This transformation analyzes and transforms the induction variables (and
10 // computations derived from them) into forms suitable for efficient execution
13 // This pass performs a strength reduction on array references inside loops that
14 // have as one or more of their components the loop induction variable, it
15 // rewrites expressions to take advantage of scaled-index addressing modes
16 // available on the target, and it performs a variety of other optimizations
17 // related to loop induction variables.
19 // Terminology note: this code has a lot of handling for "post-increment" or
20 // "post-inc" users. This is not talking about post-increment addressing modes;
21 // it is instead talking about code like this:
23 // %i = phi [ 0, %entry ], [ %i.next, %latch ]
25 // %i.next = add %i, 1
26 // %c = icmp eq %i.next, %n
28 // The SCEV for %i is {0,+,1}<%L>. The SCEV for %i.next is {1,+,1}<%L>, however
29 // it's useful to think about these as the same register, with some uses using
30 // the value of the register before the add and some using it after. In this
31 // example, the icmp is a post-increment user, since it uses %i.next, which is
32 // the value of the induction variable after the increment. The other common
33 // case of post-increment users is users outside the loop.
35 // TODO: More sophistication in the way Formulae are generated and filtered.
37 // TODO: Handle multiple loops at a time.
39 // TODO: Should the addressing mode BaseGV be changed to a ConstantExpr instead
42 // TODO: When truncation is free, truncate ICmp users' operands to make it a
43 // smaller encoding (on x86 at least).
45 // TODO: When a negated register is used by an add (such as in a list of
46 // multiple base registers, or as the increment expression in an addrec),
47 // we may not actually need both reg and (-1 * reg) in registers; the
48 // negation can be implemented by using a sub instead of an add. The
49 // lack of support for taking this into consideration when making
50 // register pressure decisions is partly worked around by the "Special"
53 //===----------------------------------------------------------------------===//
55 #include "llvm/Transforms/Scalar/LoopStrengthReduce.h"
56 #include "llvm/ADT/APInt.h"
57 #include "llvm/ADT/DenseMap.h"
58 #include "llvm/ADT/DenseSet.h"
59 #include "llvm/ADT/Hashing.h"
60 #include "llvm/ADT/PointerIntPair.h"
61 #include "llvm/ADT/STLExtras.h"
62 #include "llvm/ADT/SetVector.h"
63 #include "llvm/ADT/SmallBitVector.h"
64 #include "llvm/ADT/SmallPtrSet.h"
65 #include "llvm/ADT/SmallSet.h"
66 #include "llvm/ADT/SmallVector.h"
67 #include "llvm/ADT/iterator_range.h"
68 #include "llvm/Analysis/IVUsers.h"
69 #include "llvm/Analysis/LoopAnalysisManager.h"
70 #include "llvm/Analysis/LoopInfo.h"
71 #include "llvm/Analysis/LoopPass.h"
72 #include "llvm/Analysis/ScalarEvolution.h"
73 #include "llvm/Analysis/ScalarEvolutionExpander.h"
74 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
75 #include "llvm/Analysis/ScalarEvolutionNormalization.h"
76 #include "llvm/Analysis/TargetTransformInfo.h"
77 #include "llvm/Config/llvm-config.h"
78 #include "llvm/IR/BasicBlock.h"
79 #include "llvm/IR/Constant.h"
80 #include "llvm/IR/Constants.h"
81 #include "llvm/IR/DerivedTypes.h"
82 #include "llvm/IR/Dominators.h"
83 #include "llvm/IR/GlobalValue.h"
84 #include "llvm/IR/IRBuilder.h"
85 #include "llvm/IR/InstrTypes.h"
86 #include "llvm/IR/Instruction.h"
87 #include "llvm/IR/Instructions.h"
88 #include "llvm/IR/IntrinsicInst.h"
89 #include "llvm/IR/Intrinsics.h"
90 #include "llvm/IR/Module.h"
91 #include "llvm/IR/OperandTraits.h"
92 #include "llvm/IR/Operator.h"
93 #include "llvm/IR/PassManager.h"
94 #include "llvm/IR/Type.h"
95 #include "llvm/IR/Use.h"
96 #include "llvm/IR/User.h"
97 #include "llvm/IR/Value.h"
98 #include "llvm/IR/ValueHandle.h"
99 #include "llvm/InitializePasses.h"
100 #include "llvm/Pass.h"
101 #include "llvm/Support/Casting.h"
102 #include "llvm/Support/CommandLine.h"
103 #include "llvm/Support/Compiler.h"
104 #include "llvm/Support/Debug.h"
105 #include "llvm/Support/ErrorHandling.h"
106 #include "llvm/Support/MathExtras.h"
107 #include "llvm/Support/raw_ostream.h"
108 #include "llvm/Transforms/Scalar.h"
109 #include "llvm/Transforms/Utils.h"
110 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
111 #include "llvm/Transforms/Utils/Local.h"
123 using namespace llvm;
125 #define DEBUG_TYPE "loop-reduce"
127 /// MaxIVUsers is an arbitrary threshold that provides an early opportunity for
128 /// bail out. This threshold is far beyond the number of users that LSR can
129 /// conceivably solve, so it should not affect generated code, but catches the
130 /// worst cases before LSR burns too much compile time and stack space.
131 static const unsigned MaxIVUsers = 200;
133 // Temporary flag to cleanup congruent phis after LSR phi expansion.
134 // It's currently disabled until we can determine whether it's truly useful or
135 // not. The flag should be removed after the v3.0 release.
136 // This is now needed for ivchains.
137 static cl::opt<bool> EnablePhiElim(
138 "enable-lsr-phielim", cl::Hidden, cl::init(true),
139 cl::desc("Enable LSR phi elimination"));
141 // The flag adds instruction count to solutions cost comparision.
142 static cl::opt<bool> InsnsCost(
143 "lsr-insns-cost", cl::Hidden, cl::init(true),
144 cl::desc("Add instruction count to a LSR cost model"));
146 // Flag to choose how to narrow complex lsr solution
147 static cl::opt<bool> LSRExpNarrow(
148 "lsr-exp-narrow", cl::Hidden, cl::init(false),
149 cl::desc("Narrow LSR complex solution using"
150 " expectation of registers number"));
152 // Flag to narrow search space by filtering non-optimal formulae with
153 // the same ScaledReg and Scale.
154 static cl::opt<bool> FilterSameScaledReg(
155 "lsr-filter-same-scaled-reg", cl::Hidden, cl::init(true),
156 cl::desc("Narrow LSR search space by filtering non-optimal formulae"
157 " with the same ScaledReg and Scale"));
159 static cl::opt<bool> EnableBackedgeIndexing(
160 "lsr-backedge-indexing", cl::Hidden, cl::init(true),
161 cl::desc("Enable the generation of cross iteration indexed memops"));
163 static cl::opt<unsigned> ComplexityLimit(
164 "lsr-complexity-limit", cl::Hidden,
165 cl::init(std::numeric_limits<uint16_t>::max()),
166 cl::desc("LSR search space complexity limit"));
168 static cl::opt<unsigned> SetupCostDepthLimit(
169 "lsr-setupcost-depth-limit", cl::Hidden, cl::init(7),
170 cl::desc("The limit on recursion depth for LSRs setup cost"));
173 // Stress test IV chain generation.
174 static cl::opt<bool> StressIVChain(
175 "stress-ivchain", cl::Hidden, cl::init(false),
176 cl::desc("Stress test LSR IV chains"));
178 static bool StressIVChain = false;
184 /// Used in situations where the accessed memory type is unknown.
185 static const unsigned UnknownAddressSpace =
186 std::numeric_limits<unsigned>::max();
188 Type *MemTy = nullptr;
189 unsigned AddrSpace = UnknownAddressSpace;
191 MemAccessTy() = default;
192 MemAccessTy(Type *Ty, unsigned AS) : MemTy(Ty), AddrSpace(AS) {}
194 bool operator==(MemAccessTy Other) const {
195 return MemTy == Other.MemTy && AddrSpace == Other.AddrSpace;
198 bool operator!=(MemAccessTy Other) const { return !(*this == Other); }
200 static MemAccessTy getUnknown(LLVMContext &Ctx,
201 unsigned AS = UnknownAddressSpace) {
202 return MemAccessTy(Type::getVoidTy(Ctx), AS);
205 Type *getType() { return MemTy; }
208 /// This class holds data which is used to order reuse candidates.
211 /// This represents the set of LSRUse indices which reference
212 /// a particular register.
213 SmallBitVector UsedByIndices;
215 void print(raw_ostream &OS) const;
219 } // end anonymous namespace
221 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
222 void RegSortData::print(raw_ostream &OS) const {
223 OS << "[NumUses=" << UsedByIndices.count() << ']';
226 LLVM_DUMP_METHOD void RegSortData::dump() const {
227 print(errs()); errs() << '\n';
233 /// Map register candidates to information about how they are used.
234 class RegUseTracker {
235 using RegUsesTy = DenseMap<const SCEV *, RegSortData>;
237 RegUsesTy RegUsesMap;
238 SmallVector<const SCEV *, 16> RegSequence;
241 void countRegister(const SCEV *Reg, size_t LUIdx);
242 void dropRegister(const SCEV *Reg, size_t LUIdx);
243 void swapAndDropUse(size_t LUIdx, size_t LastLUIdx);
245 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
247 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
251 using iterator = SmallVectorImpl<const SCEV *>::iterator;
252 using const_iterator = SmallVectorImpl<const SCEV *>::const_iterator;
254 iterator begin() { return RegSequence.begin(); }
255 iterator end() { return RegSequence.end(); }
256 const_iterator begin() const { return RegSequence.begin(); }
257 const_iterator end() const { return RegSequence.end(); }
260 } // end anonymous namespace
263 RegUseTracker::countRegister(const SCEV *Reg, size_t LUIdx) {
264 std::pair<RegUsesTy::iterator, bool> Pair =
265 RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
266 RegSortData &RSD = Pair.first->second;
268 RegSequence.push_back(Reg);
269 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
270 RSD.UsedByIndices.set(LUIdx);
274 RegUseTracker::dropRegister(const SCEV *Reg, size_t LUIdx) {
275 RegUsesTy::iterator It = RegUsesMap.find(Reg);
276 assert(It != RegUsesMap.end());
277 RegSortData &RSD = It->second;
278 assert(RSD.UsedByIndices.size() > LUIdx);
279 RSD.UsedByIndices.reset(LUIdx);
283 RegUseTracker::swapAndDropUse(size_t LUIdx, size_t LastLUIdx) {
284 assert(LUIdx <= LastLUIdx);
286 // Update RegUses. The data structure is not optimized for this purpose;
287 // we must iterate through it and update each of the bit vectors.
288 for (auto &Pair : RegUsesMap) {
289 SmallBitVector &UsedByIndices = Pair.second.UsedByIndices;
290 if (LUIdx < UsedByIndices.size())
291 UsedByIndices[LUIdx] =
292 LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : false;
293 UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx));
298 RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
299 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
300 if (I == RegUsesMap.end())
302 const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
303 int i = UsedByIndices.find_first();
304 if (i == -1) return false;
305 if ((size_t)i != LUIdx) return true;
306 return UsedByIndices.find_next(i) != -1;
309 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
310 RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
311 assert(I != RegUsesMap.end() && "Unknown register!");
312 return I->second.UsedByIndices;
315 void RegUseTracker::clear() {
322 /// This class holds information that describes a formula for computing
323 /// satisfying a use. It may include broken-out immediates and scaled registers.
325 /// Global base address used for complex addressing.
326 GlobalValue *BaseGV = nullptr;
328 /// Base offset for complex addressing.
329 int64_t BaseOffset = 0;
331 /// Whether any complex addressing has a base register.
332 bool HasBaseReg = false;
334 /// The scale of any complex addressing.
337 /// The list of "base" registers for this use. When this is non-empty. The
338 /// canonical representation of a formula is
339 /// 1. BaseRegs.size > 1 implies ScaledReg != NULL and
340 /// 2. ScaledReg != NULL implies Scale != 1 || !BaseRegs.empty().
341 /// 3. The reg containing recurrent expr related with currect loop in the
342 /// formula should be put in the ScaledReg.
343 /// #1 enforces that the scaled register is always used when at least two
344 /// registers are needed by the formula: e.g., reg1 + reg2 is reg1 + 1 * reg2.
345 /// #2 enforces that 1 * reg is reg.
346 /// #3 ensures invariant regs with respect to current loop can be combined
347 /// together in LSR codegen.
348 /// This invariant can be temporarily broken while building a formula.
349 /// However, every formula inserted into the LSRInstance must be in canonical
351 SmallVector<const SCEV *, 4> BaseRegs;
353 /// The 'scaled' register for this use. This should be non-null when Scale is
355 const SCEV *ScaledReg = nullptr;
357 /// An additional constant offset which added near the use. This requires a
358 /// temporary register, but the offset itself can live in an add immediate
359 /// field rather than a register.
360 int64_t UnfoldedOffset = 0;
364 void initialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE);
366 bool isCanonical(const Loop &L) const;
368 void canonicalize(const Loop &L);
372 bool hasZeroEnd() const;
374 size_t getNumRegs() const;
375 Type *getType() const;
377 void deleteBaseReg(const SCEV *&S);
379 bool referencesReg(const SCEV *S) const;
380 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
381 const RegUseTracker &RegUses) const;
383 void print(raw_ostream &OS) const;
387 } // end anonymous namespace
389 /// Recursion helper for initialMatch.
390 static void DoInitialMatch(const SCEV *S, Loop *L,
391 SmallVectorImpl<const SCEV *> &Good,
392 SmallVectorImpl<const SCEV *> &Bad,
393 ScalarEvolution &SE) {
394 // Collect expressions which properly dominate the loop header.
395 if (SE.properlyDominates(S, L->getHeader())) {
400 // Look at add operands.
401 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
402 for (const SCEV *S : Add->operands())
403 DoInitialMatch(S, L, Good, Bad, SE);
407 // Look at addrec operands.
408 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
409 if (!AR->getStart()->isZero() && AR->isAffine()) {
410 DoInitialMatch(AR->getStart(), L, Good, Bad, SE);
411 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
412 AR->getStepRecurrence(SE),
413 // FIXME: AR->getNoWrapFlags()
414 AR->getLoop(), SCEV::FlagAnyWrap),
419 // Handle a multiplication by -1 (negation) if it didn't fold.
420 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
421 if (Mul->getOperand(0)->isAllOnesValue()) {
422 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
423 const SCEV *NewMul = SE.getMulExpr(Ops);
425 SmallVector<const SCEV *, 4> MyGood;
426 SmallVector<const SCEV *, 4> MyBad;
427 DoInitialMatch(NewMul, L, MyGood, MyBad, SE);
428 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
429 SE.getEffectiveSCEVType(NewMul->getType())));
430 for (const SCEV *S : MyGood)
431 Good.push_back(SE.getMulExpr(NegOne, S));
432 for (const SCEV *S : MyBad)
433 Bad.push_back(SE.getMulExpr(NegOne, S));
437 // Ok, we can't do anything interesting. Just stuff the whole thing into a
438 // register and hope for the best.
442 /// Incorporate loop-variant parts of S into this Formula, attempting to keep
443 /// all loop-invariant and loop-computable values in a single base register.
444 void Formula::initialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) {
445 SmallVector<const SCEV *, 4> Good;
446 SmallVector<const SCEV *, 4> Bad;
447 DoInitialMatch(S, L, Good, Bad, SE);
449 const SCEV *Sum = SE.getAddExpr(Good);
451 BaseRegs.push_back(Sum);
455 const SCEV *Sum = SE.getAddExpr(Bad);
457 BaseRegs.push_back(Sum);
463 /// Check whether or not this formula satisfies the canonical
465 /// \see Formula::BaseRegs.
466 bool Formula::isCanonical(const Loop &L) const {
468 return BaseRegs.size() <= 1;
473 if (Scale == 1 && BaseRegs.empty())
476 const SCEVAddRecExpr *SAR = dyn_cast<const SCEVAddRecExpr>(ScaledReg);
477 if (SAR && SAR->getLoop() == &L)
480 // If ScaledReg is not a recurrent expr, or it is but its loop is not current
481 // loop, meanwhile BaseRegs contains a recurrent expr reg related with current
482 // loop, we want to swap the reg in BaseRegs with ScaledReg.
484 find_if(make_range(BaseRegs.begin(), BaseRegs.end()), [&](const SCEV *S) {
485 return isa<const SCEVAddRecExpr>(S) &&
486 (cast<SCEVAddRecExpr>(S)->getLoop() == &L);
488 return I == BaseRegs.end();
491 /// Helper method to morph a formula into its canonical representation.
492 /// \see Formula::BaseRegs.
493 /// Every formula having more than one base register, must use the ScaledReg
494 /// field. Otherwise, we would have to do special cases everywhere in LSR
495 /// to treat reg1 + reg2 + ... the same way as reg1 + 1*reg2 + ...
496 /// On the other hand, 1*reg should be canonicalized into reg.
497 void Formula::canonicalize(const Loop &L) {
500 // So far we did not need this case. This is easy to implement but it is
501 // useless to maintain dead code. Beside it could hurt compile time.
502 assert(!BaseRegs.empty() && "1*reg => reg, should not be needed.");
504 // Keep the invariant sum in BaseRegs and one of the variant sum in ScaledReg.
506 ScaledReg = BaseRegs.back();
511 // If ScaledReg is an invariant with respect to L, find the reg from
512 // BaseRegs containing the recurrent expr related with Loop L. Swap the
513 // reg with ScaledReg.
514 const SCEVAddRecExpr *SAR = dyn_cast<const SCEVAddRecExpr>(ScaledReg);
515 if (!SAR || SAR->getLoop() != &L) {
516 auto I = find_if(make_range(BaseRegs.begin(), BaseRegs.end()),
518 return isa<const SCEVAddRecExpr>(S) &&
519 (cast<SCEVAddRecExpr>(S)->getLoop() == &L);
521 if (I != BaseRegs.end())
522 std::swap(ScaledReg, *I);
526 /// Get rid of the scale in the formula.
527 /// In other words, this method morphes reg1 + 1*reg2 into reg1 + reg2.
528 /// \return true if it was possible to get rid of the scale, false otherwise.
529 /// \note After this operation the formula may not be in the canonical form.
530 bool Formula::unscale() {
534 BaseRegs.push_back(ScaledReg);
539 bool Formula::hasZeroEnd() const {
540 if (UnfoldedOffset || BaseOffset)
542 if (BaseRegs.size() != 1 || ScaledReg)
547 /// Return the total number of register operands used by this formula. This does
548 /// not include register uses implied by non-constant addrec strides.
549 size_t Formula::getNumRegs() const {
550 return !!ScaledReg + BaseRegs.size();
553 /// Return the type of this formula, if it has one, or null otherwise. This type
554 /// is meaningless except for the bit size.
555 Type *Formula::getType() const {
556 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
557 ScaledReg ? ScaledReg->getType() :
558 BaseGV ? BaseGV->getType() :
562 /// Delete the given base reg from the BaseRegs list.
563 void Formula::deleteBaseReg(const SCEV *&S) {
564 if (&S != &BaseRegs.back())
565 std::swap(S, BaseRegs.back());
569 /// Test if this formula references the given register.
570 bool Formula::referencesReg(const SCEV *S) const {
571 return S == ScaledReg || is_contained(BaseRegs, S);
574 /// Test whether this formula uses registers which are used by uses other than
575 /// the use with the given index.
576 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
577 const RegUseTracker &RegUses) const {
579 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
581 for (const SCEV *BaseReg : BaseRegs)
582 if (RegUses.isRegUsedByUsesOtherThan(BaseReg, LUIdx))
587 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
588 void Formula::print(raw_ostream &OS) const {
591 if (!First) OS << " + "; else First = false;
592 BaseGV->printAsOperand(OS, /*PrintType=*/false);
594 if (BaseOffset != 0) {
595 if (!First) OS << " + "; else First = false;
598 for (const SCEV *BaseReg : BaseRegs) {
599 if (!First) OS << " + "; else First = false;
600 OS << "reg(" << *BaseReg << ')';
602 if (HasBaseReg && BaseRegs.empty()) {
603 if (!First) OS << " + "; else First = false;
604 OS << "**error: HasBaseReg**";
605 } else if (!HasBaseReg && !BaseRegs.empty()) {
606 if (!First) OS << " + "; else First = false;
607 OS << "**error: !HasBaseReg**";
610 if (!First) OS << " + "; else First = false;
611 OS << Scale << "*reg(";
618 if (UnfoldedOffset != 0) {
619 if (!First) OS << " + ";
620 OS << "imm(" << UnfoldedOffset << ')';
624 LLVM_DUMP_METHOD void Formula::dump() const {
625 print(errs()); errs() << '\n';
629 /// Return true if the given addrec can be sign-extended without changing its
631 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
633 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1);
634 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
637 /// Return true if the given add can be sign-extended without changing its
639 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
641 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
642 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
645 /// Return true if the given mul can be sign-extended without changing its
647 static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
649 IntegerType::get(SE.getContext(),
650 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
651 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
654 /// Return an expression for LHS /s RHS, if it can be determined and if the
655 /// remainder is known to be zero, or null otherwise. If IgnoreSignificantBits
656 /// is true, expressions like (X * Y) /s Y are simplified to Y, ignoring that
657 /// the multiplication may overflow, which is useful when the result will be
658 /// used in a context where the most significant bits are ignored.
659 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
661 bool IgnoreSignificantBits = false) {
662 // Handle the trivial case, which works for any SCEV type.
664 return SE.getConstant(LHS->getType(), 1);
666 // Handle a few RHS special cases.
667 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
669 const APInt &RA = RC->getAPInt();
670 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
672 if (RA.isAllOnesValue())
673 return SE.getMulExpr(LHS, RC);
674 // Handle x /s 1 as x.
679 // Check for a division of a constant by a constant.
680 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
683 const APInt &LA = C->getAPInt();
684 const APInt &RA = RC->getAPInt();
685 if (LA.srem(RA) != 0)
687 return SE.getConstant(LA.sdiv(RA));
690 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
691 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
692 if ((IgnoreSignificantBits || isAddRecSExtable(AR, SE)) && AR->isAffine()) {
693 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
694 IgnoreSignificantBits);
695 if (!Step) return nullptr;
696 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
697 IgnoreSignificantBits);
698 if (!Start) return nullptr;
699 // FlagNW is independent of the start value, step direction, and is
700 // preserved with smaller magnitude steps.
701 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
702 return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap);
707 // Distribute the sdiv over add operands, if the add doesn't overflow.
708 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
709 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
710 SmallVector<const SCEV *, 8> Ops;
711 for (const SCEV *S : Add->operands()) {
712 const SCEV *Op = getExactSDiv(S, RHS, SE, IgnoreSignificantBits);
713 if (!Op) return nullptr;
716 return SE.getAddExpr(Ops);
721 // Check for a multiply operand that we can pull RHS out of.
722 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
723 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
724 SmallVector<const SCEV *, 4> Ops;
726 for (const SCEV *S : Mul->operands()) {
728 if (const SCEV *Q = getExactSDiv(S, RHS, SE,
729 IgnoreSignificantBits)) {
735 return Found ? SE.getMulExpr(Ops) : nullptr;
740 // Otherwise we don't know.
744 /// If S involves the addition of a constant integer value, return that integer
745 /// value, and mutate S to point to a new SCEV with that value excluded.
746 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
747 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
748 if (C->getAPInt().getMinSignedBits() <= 64) {
749 S = SE.getConstant(C->getType(), 0);
750 return C->getValue()->getSExtValue();
752 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
753 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
754 int64_t Result = ExtractImmediate(NewOps.front(), SE);
756 S = SE.getAddExpr(NewOps);
758 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
759 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
760 int64_t Result = ExtractImmediate(NewOps.front(), SE);
762 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
763 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
770 /// If S involves the addition of a GlobalValue address, return that symbol, and
771 /// mutate S to point to a new SCEV with that value excluded.
772 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
773 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
774 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
775 S = SE.getConstant(GV->getType(), 0);
778 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
779 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
780 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
782 S = SE.getAddExpr(NewOps);
784 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
785 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
786 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
788 S = SE.getAddRecExpr(NewOps, AR->getLoop(),
789 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
796 /// Returns true if the specified instruction is using the specified value as an
798 static bool isAddressUse(const TargetTransformInfo &TTI,
799 Instruction *Inst, Value *OperandVal) {
800 bool isAddress = isa<LoadInst>(Inst);
801 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
802 if (SI->getPointerOperand() == OperandVal)
804 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
805 // Addressing modes can also be folded into prefetches and a variety
807 switch (II->getIntrinsicID()) {
808 case Intrinsic::memset:
809 case Intrinsic::prefetch:
810 if (II->getArgOperand(0) == OperandVal)
813 case Intrinsic::memmove:
814 case Intrinsic::memcpy:
815 if (II->getArgOperand(0) == OperandVal ||
816 II->getArgOperand(1) == OperandVal)
820 MemIntrinsicInfo IntrInfo;
821 if (TTI.getTgtMemIntrinsic(II, IntrInfo)) {
822 if (IntrInfo.PtrVal == OperandVal)
827 } else if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Inst)) {
828 if (RMW->getPointerOperand() == OperandVal)
830 } else if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(Inst)) {
831 if (CmpX->getPointerOperand() == OperandVal)
837 /// Return the type of the memory being accessed.
838 static MemAccessTy getAccessType(const TargetTransformInfo &TTI,
839 Instruction *Inst, Value *OperandVal) {
840 MemAccessTy AccessTy(Inst->getType(), MemAccessTy::UnknownAddressSpace);
841 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
842 AccessTy.MemTy = SI->getOperand(0)->getType();
843 AccessTy.AddrSpace = SI->getPointerAddressSpace();
844 } else if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
845 AccessTy.AddrSpace = LI->getPointerAddressSpace();
846 } else if (const AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Inst)) {
847 AccessTy.AddrSpace = RMW->getPointerAddressSpace();
848 } else if (const AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(Inst)) {
849 AccessTy.AddrSpace = CmpX->getPointerAddressSpace();
850 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
851 switch (II->getIntrinsicID()) {
852 case Intrinsic::prefetch:
853 case Intrinsic::memset:
854 AccessTy.AddrSpace = II->getArgOperand(0)->getType()->getPointerAddressSpace();
855 AccessTy.MemTy = OperandVal->getType();
857 case Intrinsic::memmove:
858 case Intrinsic::memcpy:
859 AccessTy.AddrSpace = OperandVal->getType()->getPointerAddressSpace();
860 AccessTy.MemTy = OperandVal->getType();
863 MemIntrinsicInfo IntrInfo;
864 if (TTI.getTgtMemIntrinsic(II, IntrInfo) && IntrInfo.PtrVal) {
866 = IntrInfo.PtrVal->getType()->getPointerAddressSpace();
874 // All pointers have the same requirements, so canonicalize them to an
875 // arbitrary pointer type to minimize variation.
876 if (PointerType *PTy = dyn_cast<PointerType>(AccessTy.MemTy))
877 AccessTy.MemTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
878 PTy->getAddressSpace());
883 /// Return true if this AddRec is already a phi in its loop.
884 static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
885 for (PHINode &PN : AR->getLoop()->getHeader()->phis()) {
886 if (SE.isSCEVable(PN.getType()) &&
887 (SE.getEffectiveSCEVType(PN.getType()) ==
888 SE.getEffectiveSCEVType(AR->getType())) &&
889 SE.getSCEV(&PN) == AR)
895 /// Check if expanding this expression is likely to incur significant cost. This
896 /// is tricky because SCEV doesn't track which expressions are actually computed
897 /// by the current IR.
899 /// We currently allow expansion of IV increments that involve adds,
900 /// multiplication by constants, and AddRecs from existing phis.
902 /// TODO: Allow UDivExpr if we can find an existing IV increment that is an
903 /// obvious multiple of the UDivExpr.
904 static bool isHighCostExpansion(const SCEV *S,
905 SmallPtrSetImpl<const SCEV*> &Processed,
906 ScalarEvolution &SE) {
907 // Zero/One operand expressions
908 switch (S->getSCEVType()) {
913 return isHighCostExpansion(cast<SCEVTruncateExpr>(S)->getOperand(),
916 return isHighCostExpansion(cast<SCEVZeroExtendExpr>(S)->getOperand(),
919 return isHighCostExpansion(cast<SCEVSignExtendExpr>(S)->getOperand(),
923 if (!Processed.insert(S).second)
926 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
927 for (const SCEV *S : Add->operands()) {
928 if (isHighCostExpansion(S, Processed, SE))
934 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
935 if (Mul->getNumOperands() == 2) {
936 // Multiplication by a constant is ok
937 if (isa<SCEVConstant>(Mul->getOperand(0)))
938 return isHighCostExpansion(Mul->getOperand(1), Processed, SE);
940 // If we have the value of one operand, check if an existing
941 // multiplication already generates this expression.
942 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Mul->getOperand(1))) {
943 Value *UVal = U->getValue();
944 for (User *UR : UVal->users()) {
945 // If U is a constant, it may be used by a ConstantExpr.
946 Instruction *UI = dyn_cast<Instruction>(UR);
947 if (UI && UI->getOpcode() == Instruction::Mul &&
948 SE.isSCEVable(UI->getType())) {
949 return SE.getSCEV(UI) == Mul;
956 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
957 if (isExistingPhi(AR, SE))
961 // Fow now, consider any other type of expression (div/mul/min/max) high cost.
965 /// If any of the instructions in the specified set are trivially dead, delete
966 /// them and see if this makes any of their operands subsequently dead.
968 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakTrackingVH> &DeadInsts) {
969 bool Changed = false;
971 while (!DeadInsts.empty()) {
972 Value *V = DeadInsts.pop_back_val();
973 Instruction *I = dyn_cast_or_null<Instruction>(V);
975 if (!I || !isInstructionTriviallyDead(I))
978 for (Use &O : I->operands())
979 if (Instruction *U = dyn_cast<Instruction>(O)) {
982 DeadInsts.emplace_back(U);
985 I->eraseFromParent();
996 } // end anonymous namespace
998 /// Check if the addressing mode defined by \p F is completely
999 /// folded in \p LU at isel time.
1000 /// This includes address-mode folding and special icmp tricks.
1001 /// This function returns true if \p LU can accommodate what \p F
1002 /// defines and up to 1 base + 1 scaled + offset.
1003 /// In other words, if \p F has several base registers, this function may
1004 /// still return true. Therefore, users still need to account for
1005 /// additional base registers and/or unfolded offsets to derive an
1006 /// accurate cost model.
1007 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1008 const LSRUse &LU, const Formula &F);
1010 // Get the cost of the scaling factor used in F for LU.
1011 static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
1012 const LSRUse &LU, const Formula &F,
1017 /// This class is used to measure and compare candidate formulae.
1019 const Loop *L = nullptr;
1020 ScalarEvolution *SE = nullptr;
1021 const TargetTransformInfo *TTI = nullptr;
1022 TargetTransformInfo::LSRCost C;
1026 Cost(const Loop *L, ScalarEvolution &SE, const TargetTransformInfo &TTI) :
1027 L(L), SE(&SE), TTI(&TTI) {
1038 bool isLess(Cost &Other);
1043 // Once any of the metrics loses, they must all remain losers.
1045 return ((C.Insns | C.NumRegs | C.AddRecCost | C.NumIVMuls | C.NumBaseAdds
1046 | C.ImmCost | C.SetupCost | C.ScaleCost) != ~0u)
1047 || ((C.Insns & C.NumRegs & C.AddRecCost & C.NumIVMuls & C.NumBaseAdds
1048 & C.ImmCost & C.SetupCost & C.ScaleCost) == ~0u);
1053 assert(isValid() && "invalid cost");
1054 return C.NumRegs == ~0u;
1057 void RateFormula(const Formula &F,
1058 SmallPtrSetImpl<const SCEV *> &Regs,
1059 const DenseSet<const SCEV *> &VisitedRegs,
1061 SmallPtrSetImpl<const SCEV *> *LoserRegs = nullptr);
1063 void print(raw_ostream &OS) const;
1067 void RateRegister(const Formula &F, const SCEV *Reg,
1068 SmallPtrSetImpl<const SCEV *> &Regs);
1069 void RatePrimaryRegister(const Formula &F, const SCEV *Reg,
1070 SmallPtrSetImpl<const SCEV *> &Regs,
1071 SmallPtrSetImpl<const SCEV *> *LoserRegs);
1074 /// An operand value in an instruction which is to be replaced with some
1075 /// equivalent, possibly strength-reduced, replacement.
1077 /// The instruction which will be updated.
1078 Instruction *UserInst = nullptr;
1080 /// The operand of the instruction which will be replaced. The operand may be
1081 /// used more than once; every instance will be replaced.
1082 Value *OperandValToReplace = nullptr;
1084 /// If this user is to use the post-incremented value of an induction
1085 /// variable, this set is non-empty and holds the loops associated with the
1086 /// induction variable.
1087 PostIncLoopSet PostIncLoops;
1089 /// A constant offset to be added to the LSRUse expression. This allows
1090 /// multiple fixups to share the same LSRUse with different offsets, for
1091 /// example in an unrolled loop.
1094 LSRFixup() = default;
1096 bool isUseFullyOutsideLoop(const Loop *L) const;
1098 void print(raw_ostream &OS) const;
1102 /// A DenseMapInfo implementation for holding DenseMaps and DenseSets of sorted
1103 /// SmallVectors of const SCEV*.
1104 struct UniquifierDenseMapInfo {
1105 static SmallVector<const SCEV *, 4> getEmptyKey() {
1106 SmallVector<const SCEV *, 4> V;
1107 V.push_back(reinterpret_cast<const SCEV *>(-1));
1111 static SmallVector<const SCEV *, 4> getTombstoneKey() {
1112 SmallVector<const SCEV *, 4> V;
1113 V.push_back(reinterpret_cast<const SCEV *>(-2));
1117 static unsigned getHashValue(const SmallVector<const SCEV *, 4> &V) {
1118 return static_cast<unsigned>(hash_combine_range(V.begin(), V.end()));
1121 static bool isEqual(const SmallVector<const SCEV *, 4> &LHS,
1122 const SmallVector<const SCEV *, 4> &RHS) {
1127 /// This class holds the state that LSR keeps for each use in IVUsers, as well
1128 /// as uses invented by LSR itself. It includes information about what kinds of
1129 /// things can be folded into the user, information about the user itself, and
1130 /// information about how the use may be satisfied. TODO: Represent multiple
1131 /// users of the same expression in common?
1133 DenseSet<SmallVector<const SCEV *, 4>, UniquifierDenseMapInfo> Uniquifier;
1136 /// An enum for a kind of use, indicating what types of scaled and immediate
1137 /// operands it might support.
1139 Basic, ///< A normal use, with no folding.
1140 Special, ///< A special case of basic, allowing -1 scales.
1141 Address, ///< An address use; folding according to TargetLowering
1142 ICmpZero ///< An equality icmp with both operands folded into one.
1143 // TODO: Add a generic icmp too?
1146 using SCEVUseKindPair = PointerIntPair<const SCEV *, 2, KindType>;
1149 MemAccessTy AccessTy;
1151 /// The list of operands which are to be replaced.
1152 SmallVector<LSRFixup, 8> Fixups;
1154 /// Keep track of the min and max offsets of the fixups.
1155 int64_t MinOffset = std::numeric_limits<int64_t>::max();
1156 int64_t MaxOffset = std::numeric_limits<int64_t>::min();
1158 /// This records whether all of the fixups using this LSRUse are outside of
1159 /// the loop, in which case some special-case heuristics may be used.
1160 bool AllFixupsOutsideLoop = true;
1162 /// RigidFormula is set to true to guarantee that this use will be associated
1163 /// with a single formula--the one that initially matched. Some SCEV
1164 /// expressions cannot be expanded. This allows LSR to consider the registers
1165 /// used by those expressions without the need to expand them later after
1166 /// changing the formula.
1167 bool RigidFormula = false;
1169 /// This records the widest use type for any fixup using this
1170 /// LSRUse. FindUseWithSimilarFormula can't consider uses with different max
1171 /// fixup widths to be equivalent, because the narrower one may be relying on
1172 /// the implicit truncation to truncate away bogus bits.
1173 Type *WidestFixupType = nullptr;
1175 /// A list of ways to build a value that can satisfy this user. After the
1176 /// list is populated, one of these is selected heuristically and used to
1177 /// formulate a replacement for OperandValToReplace in UserInst.
1178 SmallVector<Formula, 12> Formulae;
1180 /// The set of register candidates used by all formulae in this LSRUse.
1181 SmallPtrSet<const SCEV *, 4> Regs;
1183 LSRUse(KindType K, MemAccessTy AT) : Kind(K), AccessTy(AT) {}
1185 LSRFixup &getNewFixup() {
1186 Fixups.push_back(LSRFixup());
1187 return Fixups.back();
1190 void pushFixup(LSRFixup &f) {
1191 Fixups.push_back(f);
1192 if (f.Offset > MaxOffset)
1193 MaxOffset = f.Offset;
1194 if (f.Offset < MinOffset)
1195 MinOffset = f.Offset;
1198 bool HasFormulaWithSameRegs(const Formula &F) const;
1199 float getNotSelectedProbability(const SCEV *Reg) const;
1200 bool InsertFormula(const Formula &F, const Loop &L);
1201 void DeleteFormula(Formula &F);
1202 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1204 void print(raw_ostream &OS) const;
1208 } // end anonymous namespace
1210 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1211 LSRUse::KindType Kind, MemAccessTy AccessTy,
1212 GlobalValue *BaseGV, int64_t BaseOffset,
1213 bool HasBaseReg, int64_t Scale,
1214 Instruction *Fixup = nullptr);
1216 static unsigned getSetupCost(const SCEV *Reg, unsigned Depth) {
1217 if (isa<SCEVUnknown>(Reg) || isa<SCEVConstant>(Reg))
1221 if (const auto *S = dyn_cast<SCEVAddRecExpr>(Reg))
1222 return getSetupCost(S->getStart(), Depth - 1);
1223 if (auto S = dyn_cast<SCEVCastExpr>(Reg))
1224 return getSetupCost(S->getOperand(), Depth - 1);
1225 if (auto S = dyn_cast<SCEVNAryExpr>(Reg))
1226 return std::accumulate(S->op_begin(), S->op_end(), 0,
1227 [&](unsigned i, const SCEV *Reg) {
1228 return i + getSetupCost(Reg, Depth - 1);
1230 if (auto S = dyn_cast<SCEVUDivExpr>(Reg))
1231 return getSetupCost(S->getLHS(), Depth - 1) +
1232 getSetupCost(S->getRHS(), Depth - 1);
1236 /// Tally up interesting quantities from the given register.
1237 void Cost::RateRegister(const Formula &F, const SCEV *Reg,
1238 SmallPtrSetImpl<const SCEV *> &Regs) {
1239 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
1240 // If this is an addrec for another loop, it should be an invariant
1241 // with respect to L since L is the innermost loop (at least
1242 // for now LSR only handles innermost loops).
1243 if (AR->getLoop() != L) {
1244 // If the AddRec exists, consider it's register free and leave it alone.
1245 if (isExistingPhi(AR, *SE))
1248 // It is bad to allow LSR for current loop to add induction variables
1249 // for its sibling loops.
1250 if (!AR->getLoop()->contains(L)) {
1255 // Otherwise, it will be an invariant with respect to Loop L.
1260 unsigned LoopCost = 1;
1261 if (TTI->isIndexedLoadLegal(TTI->MIM_PostInc, AR->getType()) ||
1262 TTI->isIndexedStoreLegal(TTI->MIM_PostInc, AR->getType())) {
1264 // If the step size matches the base offset, we could use pre-indexed
1266 if (TTI->shouldFavorBackedgeIndex(L)) {
1267 if (auto *Step = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE)))
1268 if (Step->getAPInt() == F.BaseOffset)
1272 if (TTI->shouldFavorPostInc()) {
1273 const SCEV *LoopStep = AR->getStepRecurrence(*SE);
1274 if (isa<SCEVConstant>(LoopStep)) {
1275 const SCEV *LoopStart = AR->getStart();
1276 if (!isa<SCEVConstant>(LoopStart) &&
1277 SE->isLoopInvariant(LoopStart, L))
1282 C.AddRecCost += LoopCost;
1284 // Add the step value register, if it needs one.
1285 // TODO: The non-affine case isn't precisely modeled here.
1286 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
1287 if (!Regs.count(AR->getOperand(1))) {
1288 RateRegister(F, AR->getOperand(1), Regs);
1296 // Rough heuristic; favor registers which don't require extra setup
1297 // instructions in the preheader.
1298 C.SetupCost += getSetupCost(Reg, SetupCostDepthLimit);
1299 // Ensure we don't, even with the recusion limit, produce invalid costs.
1300 C.SetupCost = std::min<unsigned>(C.SetupCost, 1 << 16);
1302 C.NumIVMuls += isa<SCEVMulExpr>(Reg) &&
1303 SE->hasComputableLoopEvolution(Reg, L);
1306 /// Record this register in the set. If we haven't seen it before, rate
1307 /// it. Optional LoserRegs provides a way to declare any formula that refers to
1308 /// one of those regs an instant loser.
1309 void Cost::RatePrimaryRegister(const Formula &F, const SCEV *Reg,
1310 SmallPtrSetImpl<const SCEV *> &Regs,
1311 SmallPtrSetImpl<const SCEV *> *LoserRegs) {
1312 if (LoserRegs && LoserRegs->count(Reg)) {
1316 if (Regs.insert(Reg).second) {
1317 RateRegister(F, Reg, Regs);
1318 if (LoserRegs && isLoser())
1319 LoserRegs->insert(Reg);
1323 void Cost::RateFormula(const Formula &F,
1324 SmallPtrSetImpl<const SCEV *> &Regs,
1325 const DenseSet<const SCEV *> &VisitedRegs,
1327 SmallPtrSetImpl<const SCEV *> *LoserRegs) {
1328 assert(F.isCanonical(*L) && "Cost is accurate only for canonical formula");
1329 // Tally up the registers.
1330 unsigned PrevAddRecCost = C.AddRecCost;
1331 unsigned PrevNumRegs = C.NumRegs;
1332 unsigned PrevNumBaseAdds = C.NumBaseAdds;
1333 if (const SCEV *ScaledReg = F.ScaledReg) {
1334 if (VisitedRegs.count(ScaledReg)) {
1338 RatePrimaryRegister(F, ScaledReg, Regs, LoserRegs);
1342 for (const SCEV *BaseReg : F.BaseRegs) {
1343 if (VisitedRegs.count(BaseReg)) {
1347 RatePrimaryRegister(F, BaseReg, Regs, LoserRegs);
1352 // Determine how many (unfolded) adds we'll need inside the loop.
1353 size_t NumBaseParts = F.getNumRegs();
1354 if (NumBaseParts > 1)
1355 // Do not count the base and a possible second register if the target
1356 // allows to fold 2 registers.
1358 NumBaseParts - (1 + (F.Scale && isAMCompletelyFolded(*TTI, LU, F)));
1359 C.NumBaseAdds += (F.UnfoldedOffset != 0);
1361 // Accumulate non-free scaling amounts.
1362 C.ScaleCost += getScalingFactorCost(*TTI, LU, F, *L);
1364 // Tally up the non-zero immediates.
1365 for (const LSRFixup &Fixup : LU.Fixups) {
1366 int64_t O = Fixup.Offset;
1367 int64_t Offset = (uint64_t)O + F.BaseOffset;
1369 C.ImmCost += 64; // Handle symbolic values conservatively.
1370 // TODO: This should probably be the pointer size.
1371 else if (Offset != 0)
1372 C.ImmCost += APInt(64, Offset, true).getMinSignedBits();
1374 // Check with target if this offset with this instruction is
1375 // specifically not supported.
1376 if (LU.Kind == LSRUse::Address && Offset != 0 &&
1377 !isAMCompletelyFolded(*TTI, LSRUse::Address, LU.AccessTy, F.BaseGV,
1378 Offset, F.HasBaseReg, F.Scale, Fixup.UserInst))
1382 // If we don't count instruction cost exit here.
1384 assert(isValid() && "invalid cost");
1388 // Treat every new register that exceeds TTI.getNumberOfRegisters() - 1 as
1389 // additional instruction (at least fill).
1390 // TODO: Need distinguish register class?
1391 unsigned TTIRegNum = TTI->getNumberOfRegisters(
1392 TTI->getRegisterClassForType(false, F.getType())) - 1;
1393 if (C.NumRegs > TTIRegNum) {
1394 // Cost already exceeded TTIRegNum, then only newly added register can add
1395 // new instructions.
1396 if (PrevNumRegs > TTIRegNum)
1397 C.Insns += (C.NumRegs - PrevNumRegs);
1399 C.Insns += (C.NumRegs - TTIRegNum);
1402 // If ICmpZero formula ends with not 0, it could not be replaced by
1403 // just add or sub. We'll need to compare final result of AddRec.
1404 // That means we'll need an additional instruction. But if the target can
1405 // macro-fuse a compare with a branch, don't count this extra instruction.
1406 // For -10 + {0, +, 1}:
1412 if (LU.Kind == LSRUse::ICmpZero && !F.hasZeroEnd() &&
1413 !TTI->canMacroFuseCmp())
1415 // Each new AddRec adds 1 instruction to calculation.
1416 C.Insns += (C.AddRecCost - PrevAddRecCost);
1418 // BaseAdds adds instructions for unfolded registers.
1419 if (LU.Kind != LSRUse::ICmpZero)
1420 C.Insns += C.NumBaseAdds - PrevNumBaseAdds;
1421 assert(isValid() && "invalid cost");
1424 /// Set this cost to a losing value.
1426 C.Insns = std::numeric_limits<unsigned>::max();
1427 C.NumRegs = std::numeric_limits<unsigned>::max();
1428 C.AddRecCost = std::numeric_limits<unsigned>::max();
1429 C.NumIVMuls = std::numeric_limits<unsigned>::max();
1430 C.NumBaseAdds = std::numeric_limits<unsigned>::max();
1431 C.ImmCost = std::numeric_limits<unsigned>::max();
1432 C.SetupCost = std::numeric_limits<unsigned>::max();
1433 C.ScaleCost = std::numeric_limits<unsigned>::max();
1436 /// Choose the lower cost.
1437 bool Cost::isLess(Cost &Other) {
1438 if (InsnsCost.getNumOccurrences() > 0 && InsnsCost &&
1439 C.Insns != Other.C.Insns)
1440 return C.Insns < Other.C.Insns;
1441 return TTI->isLSRCostLess(C, Other.C);
1444 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1445 void Cost::print(raw_ostream &OS) const {
1447 OS << C.Insns << " instruction" << (C.Insns == 1 ? " " : "s ");
1448 OS << C.NumRegs << " reg" << (C.NumRegs == 1 ? "" : "s");
1449 if (C.AddRecCost != 0)
1450 OS << ", with addrec cost " << C.AddRecCost;
1451 if (C.NumIVMuls != 0)
1452 OS << ", plus " << C.NumIVMuls << " IV mul"
1453 << (C.NumIVMuls == 1 ? "" : "s");
1454 if (C.NumBaseAdds != 0)
1455 OS << ", plus " << C.NumBaseAdds << " base add"
1456 << (C.NumBaseAdds == 1 ? "" : "s");
1457 if (C.ScaleCost != 0)
1458 OS << ", plus " << C.ScaleCost << " scale cost";
1460 OS << ", plus " << C.ImmCost << " imm cost";
1461 if (C.SetupCost != 0)
1462 OS << ", plus " << C.SetupCost << " setup cost";
1465 LLVM_DUMP_METHOD void Cost::dump() const {
1466 print(errs()); errs() << '\n';
1470 /// Test whether this fixup always uses its value outside of the given loop.
1471 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
1472 // PHI nodes use their value in their incoming blocks.
1473 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
1474 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1475 if (PN->getIncomingValue(i) == OperandValToReplace &&
1476 L->contains(PN->getIncomingBlock(i)))
1481 return !L->contains(UserInst);
1484 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1485 void LSRFixup::print(raw_ostream &OS) const {
1487 // Store is common and interesting enough to be worth special-casing.
1488 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
1490 Store->getOperand(0)->printAsOperand(OS, /*PrintType=*/false);
1491 } else if (UserInst->getType()->isVoidTy())
1492 OS << UserInst->getOpcodeName();
1494 UserInst->printAsOperand(OS, /*PrintType=*/false);
1496 OS << ", OperandValToReplace=";
1497 OperandValToReplace->printAsOperand(OS, /*PrintType=*/false);
1499 for (const Loop *PIL : PostIncLoops) {
1500 OS << ", PostIncLoop=";
1501 PIL->getHeader()->printAsOperand(OS, /*PrintType=*/false);
1505 OS << ", Offset=" << Offset;
1508 LLVM_DUMP_METHOD void LSRFixup::dump() const {
1509 print(errs()); errs() << '\n';
1513 /// Test whether this use as a formula which has the same registers as the given
1515 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1516 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1517 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1518 // Unstable sort by host order ok, because this is only used for uniquifying.
1520 return Uniquifier.count(Key);
1523 /// The function returns a probability of selecting formula without Reg.
1524 float LSRUse::getNotSelectedProbability(const SCEV *Reg) const {
1526 for (const Formula &F : Formulae)
1527 if (F.referencesReg(Reg))
1529 return ((float)(Formulae.size() - FNum)) / Formulae.size();
1532 /// If the given formula has not yet been inserted, add it to the list, and
1533 /// return true. Return false otherwise. The formula must be in canonical form.
1534 bool LSRUse::InsertFormula(const Formula &F, const Loop &L) {
1535 assert(F.isCanonical(L) && "Invalid canonical representation");
1537 if (!Formulae.empty() && RigidFormula)
1540 SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1541 if (F.ScaledReg) Key.push_back(F.ScaledReg);
1542 // Unstable sort by host order ok, because this is only used for uniquifying.
1545 if (!Uniquifier.insert(Key).second)
1548 // Using a register to hold the value of 0 is not profitable.
1549 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1550 "Zero allocated in a scaled register!");
1552 for (const SCEV *BaseReg : F.BaseRegs)
1553 assert(!BaseReg->isZero() && "Zero allocated in a base register!");
1556 // Add the formula to the list.
1557 Formulae.push_back(F);
1559 // Record registers now being used by this use.
1560 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1562 Regs.insert(F.ScaledReg);
1567 /// Remove the given formula from this use's list.
1568 void LSRUse::DeleteFormula(Formula &F) {
1569 if (&F != &Formulae.back())
1570 std::swap(F, Formulae.back());
1571 Formulae.pop_back();
1574 /// Recompute the Regs field, and update RegUses.
1575 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1576 // Now that we've filtered out some formulae, recompute the Regs set.
1577 SmallPtrSet<const SCEV *, 4> OldRegs = std::move(Regs);
1579 for (const Formula &F : Formulae) {
1580 if (F.ScaledReg) Regs.insert(F.ScaledReg);
1581 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1584 // Update the RegTracker.
1585 for (const SCEV *S : OldRegs)
1587 RegUses.dropRegister(S, LUIdx);
1590 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1591 void LSRUse::print(raw_ostream &OS) const {
1592 OS << "LSR Use: Kind=";
1594 case Basic: OS << "Basic"; break;
1595 case Special: OS << "Special"; break;
1596 case ICmpZero: OS << "ICmpZero"; break;
1598 OS << "Address of ";
1599 if (AccessTy.MemTy->isPointerTy())
1600 OS << "pointer"; // the full pointer type could be really verbose
1602 OS << *AccessTy.MemTy;
1605 OS << " in addrspace(" << AccessTy.AddrSpace << ')';
1608 OS << ", Offsets={";
1609 bool NeedComma = false;
1610 for (const LSRFixup &Fixup : Fixups) {
1611 if (NeedComma) OS << ',';
1617 if (AllFixupsOutsideLoop)
1618 OS << ", all-fixups-outside-loop";
1620 if (WidestFixupType)
1621 OS << ", widest fixup type: " << *WidestFixupType;
1624 LLVM_DUMP_METHOD void LSRUse::dump() const {
1625 print(errs()); errs() << '\n';
1629 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1630 LSRUse::KindType Kind, MemAccessTy AccessTy,
1631 GlobalValue *BaseGV, int64_t BaseOffset,
1632 bool HasBaseReg, int64_t Scale,
1633 Instruction *Fixup/*= nullptr*/) {
1635 case LSRUse::Address:
1636 return TTI.isLegalAddressingMode(AccessTy.MemTy, BaseGV, BaseOffset,
1637 HasBaseReg, Scale, AccessTy.AddrSpace, Fixup);
1639 case LSRUse::ICmpZero:
1640 // There's not even a target hook for querying whether it would be legal to
1641 // fold a GV into an ICmp.
1645 // ICmp only has two operands; don't allow more than two non-trivial parts.
1646 if (Scale != 0 && HasBaseReg && BaseOffset != 0)
1649 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1650 // putting the scaled register in the other operand of the icmp.
1651 if (Scale != 0 && Scale != -1)
1654 // If we have low-level target information, ask the target if it can fold an
1655 // integer immediate on an icmp.
1656 if (BaseOffset != 0) {
1658 // ICmpZero BaseReg + BaseOffset => ICmp BaseReg, -BaseOffset
1659 // ICmpZero -1*ScaleReg + BaseOffset => ICmp ScaleReg, BaseOffset
1660 // Offs is the ICmp immediate.
1662 // The cast does the right thing with
1663 // std::numeric_limits<int64_t>::min().
1664 BaseOffset = -(uint64_t)BaseOffset;
1665 return TTI.isLegalICmpImmediate(BaseOffset);
1668 // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg
1672 // Only handle single-register values.
1673 return !BaseGV && Scale == 0 && BaseOffset == 0;
1675 case LSRUse::Special:
1676 // Special case Basic to handle -1 scales.
1677 return !BaseGV && (Scale == 0 || Scale == -1) && BaseOffset == 0;
1680 llvm_unreachable("Invalid LSRUse Kind!");
1683 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1684 int64_t MinOffset, int64_t MaxOffset,
1685 LSRUse::KindType Kind, MemAccessTy AccessTy,
1686 GlobalValue *BaseGV, int64_t BaseOffset,
1687 bool HasBaseReg, int64_t Scale) {
1688 // Check for overflow.
1689 if (((int64_t)((uint64_t)BaseOffset + MinOffset) > BaseOffset) !=
1692 MinOffset = (uint64_t)BaseOffset + MinOffset;
1693 if (((int64_t)((uint64_t)BaseOffset + MaxOffset) > BaseOffset) !=
1696 MaxOffset = (uint64_t)BaseOffset + MaxOffset;
1698 return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MinOffset,
1699 HasBaseReg, Scale) &&
1700 isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MaxOffset,
1704 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1705 int64_t MinOffset, int64_t MaxOffset,
1706 LSRUse::KindType Kind, MemAccessTy AccessTy,
1707 const Formula &F, const Loop &L) {
1708 // For the purpose of isAMCompletelyFolded either having a canonical formula
1709 // or a scale not equal to zero is correct.
1710 // Problems may arise from non canonical formulae having a scale == 0.
1711 // Strictly speaking it would best to just rely on canonical formulae.
1712 // However, when we generate the scaled formulae, we first check that the
1713 // scaling factor is profitable before computing the actual ScaledReg for
1714 // compile time sake.
1715 assert((F.isCanonical(L) || F.Scale != 0));
1716 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1717 F.BaseGV, F.BaseOffset, F.HasBaseReg, F.Scale);
1720 /// Test whether we know how to expand the current formula.
1721 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1722 int64_t MaxOffset, LSRUse::KindType Kind,
1723 MemAccessTy AccessTy, GlobalValue *BaseGV,
1724 int64_t BaseOffset, bool HasBaseReg, int64_t Scale) {
1725 // We know how to expand completely foldable formulae.
1726 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1727 BaseOffset, HasBaseReg, Scale) ||
1728 // Or formulae that use a base register produced by a sum of base
1731 isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1732 BaseGV, BaseOffset, true, 0));
1735 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1736 int64_t MaxOffset, LSRUse::KindType Kind,
1737 MemAccessTy AccessTy, const Formula &F) {
1738 return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, F.BaseGV,
1739 F.BaseOffset, F.HasBaseReg, F.Scale);
1742 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1743 const LSRUse &LU, const Formula &F) {
1744 // Target may want to look at the user instructions.
1745 if (LU.Kind == LSRUse::Address && TTI.LSRWithInstrQueries()) {
1746 for (const LSRFixup &Fixup : LU.Fixups)
1747 if (!isAMCompletelyFolded(TTI, LSRUse::Address, LU.AccessTy, F.BaseGV,
1748 (F.BaseOffset + Fixup.Offset), F.HasBaseReg,
1749 F.Scale, Fixup.UserInst))
1754 return isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1755 LU.AccessTy, F.BaseGV, F.BaseOffset, F.HasBaseReg,
1759 static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
1760 const LSRUse &LU, const Formula &F,
1765 // If the use is not completely folded in that instruction, we will have to
1766 // pay an extra cost only for scale != 1.
1767 if (!isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1769 return F.Scale != 1;
1772 case LSRUse::Address: {
1773 // Check the scaling factor cost with both the min and max offsets.
1774 int ScaleCostMinOffset = TTI.getScalingFactorCost(
1775 LU.AccessTy.MemTy, F.BaseGV, F.BaseOffset + LU.MinOffset, F.HasBaseReg,
1776 F.Scale, LU.AccessTy.AddrSpace);
1777 int ScaleCostMaxOffset = TTI.getScalingFactorCost(
1778 LU.AccessTy.MemTy, F.BaseGV, F.BaseOffset + LU.MaxOffset, F.HasBaseReg,
1779 F.Scale, LU.AccessTy.AddrSpace);
1781 assert(ScaleCostMinOffset >= 0 && ScaleCostMaxOffset >= 0 &&
1782 "Legal addressing mode has an illegal cost!");
1783 return std::max(ScaleCostMinOffset, ScaleCostMaxOffset);
1785 case LSRUse::ICmpZero:
1787 case LSRUse::Special:
1788 // The use is completely folded, i.e., everything is folded into the
1793 llvm_unreachable("Invalid LSRUse Kind!");
1796 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1797 LSRUse::KindType Kind, MemAccessTy AccessTy,
1798 GlobalValue *BaseGV, int64_t BaseOffset,
1800 // Fast-path: zero is always foldable.
1801 if (BaseOffset == 0 && !BaseGV) return true;
1803 // Conservatively, create an address with an immediate and a
1804 // base and a scale.
1805 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1807 // Canonicalize a scale of 1 to a base register if the formula doesn't
1808 // already have a base register.
1809 if (!HasBaseReg && Scale == 1) {
1814 return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, BaseOffset,
1818 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1819 ScalarEvolution &SE, int64_t MinOffset,
1820 int64_t MaxOffset, LSRUse::KindType Kind,
1821 MemAccessTy AccessTy, const SCEV *S,
1823 // Fast-path: zero is always foldable.
1824 if (S->isZero()) return true;
1826 // Conservatively, create an address with an immediate and a
1827 // base and a scale.
1828 int64_t BaseOffset = ExtractImmediate(S, SE);
1829 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1831 // If there's anything else involved, it's not foldable.
1832 if (!S->isZero()) return false;
1834 // Fast-path: zero is always foldable.
1835 if (BaseOffset == 0 && !BaseGV) return true;
1837 // Conservatively, create an address with an immediate and a
1838 // base and a scale.
1839 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1841 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1842 BaseOffset, HasBaseReg, Scale);
1847 /// An individual increment in a Chain of IV increments. Relate an IV user to
1848 /// an expression that computes the IV it uses from the IV used by the previous
1849 /// link in the Chain.
1851 /// For the head of a chain, IncExpr holds the absolute SCEV expression for the
1852 /// original IVOperand. The head of the chain's IVOperand is only valid during
1853 /// chain collection, before LSR replaces IV users. During chain generation,
1854 /// IncExpr can be used to find the new IVOperand that computes the same
1857 Instruction *UserInst;
1859 const SCEV *IncExpr;
1861 IVInc(Instruction *U, Value *O, const SCEV *E)
1862 : UserInst(U), IVOperand(O), IncExpr(E) {}
1865 // The list of IV increments in program order. We typically add the head of a
1866 // chain without finding subsequent links.
1868 SmallVector<IVInc, 1> Incs;
1869 const SCEV *ExprBase = nullptr;
1871 IVChain() = default;
1872 IVChain(const IVInc &Head, const SCEV *Base)
1873 : Incs(1, Head), ExprBase(Base) {}
1875 using const_iterator = SmallVectorImpl<IVInc>::const_iterator;
1877 // Return the first increment in the chain.
1878 const_iterator begin() const {
1879 assert(!Incs.empty());
1880 return std::next(Incs.begin());
1882 const_iterator end() const {
1886 // Returns true if this chain contains any increments.
1887 bool hasIncs() const { return Incs.size() >= 2; }
1889 // Add an IVInc to the end of this chain.
1890 void add(const IVInc &X) { Incs.push_back(X); }
1892 // Returns the last UserInst in the chain.
1893 Instruction *tailUserInst() const { return Incs.back().UserInst; }
1895 // Returns true if IncExpr can be profitably added to this chain.
1896 bool isProfitableIncrement(const SCEV *OperExpr,
1897 const SCEV *IncExpr,
1901 /// Helper for CollectChains to track multiple IV increment uses. Distinguish
1902 /// between FarUsers that definitely cross IV increments and NearUsers that may
1903 /// be used between IV increments.
1905 SmallPtrSet<Instruction*, 4> FarUsers;
1906 SmallPtrSet<Instruction*, 4> NearUsers;
1909 /// This class holds state for the main loop strength reduction logic.
1912 ScalarEvolution &SE;
1915 AssumptionCache &AC;
1916 TargetLibraryInfo &LibInfo;
1917 const TargetTransformInfo &TTI;
1919 bool FavorBackedgeIndex = false;
1920 bool Changed = false;
1922 /// This is the insert position that the current loop's induction variable
1923 /// increment should be placed. In simple loops, this is the latch block's
1924 /// terminator. But in more complicated cases, this is a position which will
1925 /// dominate all the in-loop post-increment users.
1926 Instruction *IVIncInsertPos = nullptr;
1928 /// Interesting factors between use strides.
1930 /// We explicitly use a SetVector which contains a SmallSet, instead of the
1931 /// default, a SmallDenseSet, because we need to use the full range of
1932 /// int64_ts, and there's currently no good way of doing that with
1934 SetVector<int64_t, SmallVector<int64_t, 8>, SmallSet<int64_t, 8>> Factors;
1936 /// Interesting use types, to facilitate truncation reuse.
1937 SmallSetVector<Type *, 4> Types;
1939 /// The list of interesting uses.
1940 mutable SmallVector<LSRUse, 16> Uses;
1942 /// Track which uses use which register candidates.
1943 RegUseTracker RegUses;
1945 // Limit the number of chains to avoid quadratic behavior. We don't expect to
1946 // have more than a few IV increment chains in a loop. Missing a Chain falls
1947 // back to normal LSR behavior for those uses.
1948 static const unsigned MaxChains = 8;
1950 /// IV users can form a chain of IV increments.
1951 SmallVector<IVChain, MaxChains> IVChainVec;
1953 /// IV users that belong to profitable IVChains.
1954 SmallPtrSet<Use*, MaxChains> IVIncSet;
1956 void OptimizeShadowIV();
1957 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1958 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1959 void OptimizeLoopTermCond();
1961 void ChainInstruction(Instruction *UserInst, Instruction *IVOper,
1962 SmallVectorImpl<ChainUsers> &ChainUsersVec);
1963 void FinalizeChain(IVChain &Chain);
1964 void CollectChains();
1965 void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
1966 SmallVectorImpl<WeakTrackingVH> &DeadInsts);
1968 void CollectInterestingTypesAndFactors();
1969 void CollectFixupsAndInitialFormulae();
1971 // Support for sharing of LSRUses between LSRFixups.
1972 using UseMapTy = DenseMap<LSRUse::SCEVUseKindPair, size_t>;
1975 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1976 LSRUse::KindType Kind, MemAccessTy AccessTy);
1978 std::pair<size_t, int64_t> getUse(const SCEV *&Expr, LSRUse::KindType Kind,
1979 MemAccessTy AccessTy);
1981 void DeleteUse(LSRUse &LU, size_t LUIdx);
1983 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1985 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1986 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1987 void CountRegisters(const Formula &F, size_t LUIdx);
1988 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1990 void CollectLoopInvariantFixupsAndFormulae();
1992 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1993 unsigned Depth = 0);
1995 void GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
1996 const Formula &Base, unsigned Depth,
1997 size_t Idx, bool IsScaledReg = false);
1998 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1999 void GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
2000 const Formula &Base, size_t Idx,
2001 bool IsScaledReg = false);
2002 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
2003 void GenerateConstantOffsetsImpl(LSRUse &LU, unsigned LUIdx,
2004 const Formula &Base,
2005 const SmallVectorImpl<int64_t> &Worklist,
2006 size_t Idx, bool IsScaledReg = false);
2007 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
2008 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
2009 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
2010 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
2011 void GenerateCrossUseConstantOffsets();
2012 void GenerateAllReuseFormulae();
2014 void FilterOutUndesirableDedicatedRegisters();
2016 size_t EstimateSearchSpaceComplexity() const;
2017 void NarrowSearchSpaceByDetectingSupersets();
2018 void NarrowSearchSpaceByCollapsingUnrolledCode();
2019 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
2020 void NarrowSearchSpaceByFilterFormulaWithSameScaledReg();
2021 void NarrowSearchSpaceByDeletingCostlyFormulas();
2022 void NarrowSearchSpaceByPickingWinnerRegs();
2023 void NarrowSearchSpaceUsingHeuristics();
2025 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
2027 SmallVectorImpl<const Formula *> &Workspace,
2028 const Cost &CurCost,
2029 const SmallPtrSet<const SCEV *, 16> &CurRegs,
2030 DenseSet<const SCEV *> &VisitedRegs) const;
2031 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
2033 BasicBlock::iterator
2034 HoistInsertPosition(BasicBlock::iterator IP,
2035 const SmallVectorImpl<Instruction *> &Inputs) const;
2036 BasicBlock::iterator
2037 AdjustInsertPositionForExpand(BasicBlock::iterator IP,
2040 SCEVExpander &Rewriter) const;
2042 Value *Expand(const LSRUse &LU, const LSRFixup &LF, const Formula &F,
2043 BasicBlock::iterator IP, SCEVExpander &Rewriter,
2044 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
2045 void RewriteForPHI(PHINode *PN, const LSRUse &LU, const LSRFixup &LF,
2046 const Formula &F, SCEVExpander &Rewriter,
2047 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
2048 void Rewrite(const LSRUse &LU, const LSRFixup &LF, const Formula &F,
2049 SCEVExpander &Rewriter,
2050 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
2051 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution);
2054 LSRInstance(Loop *L, IVUsers &IU, ScalarEvolution &SE, DominatorTree &DT,
2055 LoopInfo &LI, const TargetTransformInfo &TTI, AssumptionCache &AC,
2056 TargetLibraryInfo &LibInfo);
2058 bool getChanged() const { return Changed; }
2060 void print_factors_and_types(raw_ostream &OS) const;
2061 void print_fixups(raw_ostream &OS) const;
2062 void print_uses(raw_ostream &OS) const;
2063 void print(raw_ostream &OS) const;
2067 } // end anonymous namespace
2069 /// If IV is used in a int-to-float cast inside the loop then try to eliminate
2070 /// the cast operation.
2071 void LSRInstance::OptimizeShadowIV() {
2072 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
2073 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2076 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
2077 UI != E; /* empty */) {
2078 IVUsers::const_iterator CandidateUI = UI;
2080 Instruction *ShadowUse = CandidateUI->getUser();
2081 Type *DestTy = nullptr;
2082 bool IsSigned = false;
2084 /* If shadow use is a int->float cast then insert a second IV
2085 to eliminate this cast.
2087 for (unsigned i = 0; i < n; ++i)
2093 for (unsigned i = 0; i < n; ++i, ++d)
2096 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
2098 DestTy = UCast->getDestTy();
2100 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
2102 DestTy = SCast->getDestTy();
2104 if (!DestTy) continue;
2106 // If target does not support DestTy natively then do not apply
2107 // this transformation.
2108 if (!TTI.isTypeLegal(DestTy)) continue;
2110 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
2112 if (PH->getNumIncomingValues() != 2) continue;
2114 // If the calculation in integers overflows, the result in FP type will
2115 // differ. So we only can do this transformation if we are guaranteed to not
2116 // deal with overflowing values
2117 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE.getSCEV(PH));
2119 if (IsSigned && !AR->hasNoSignedWrap()) continue;
2120 if (!IsSigned && !AR->hasNoUnsignedWrap()) continue;
2122 Type *SrcTy = PH->getType();
2123 int Mantissa = DestTy->getFPMantissaWidth();
2124 if (Mantissa == -1) continue;
2125 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
2128 unsigned Entry, Latch;
2129 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
2137 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
2138 if (!Init) continue;
2139 Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
2140 (double)Init->getSExtValue() :
2141 (double)Init->getZExtValue());
2143 BinaryOperator *Incr =
2144 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
2145 if (!Incr) continue;
2146 if (Incr->getOpcode() != Instruction::Add
2147 && Incr->getOpcode() != Instruction::Sub)
2150 /* Initialize new IV, double d = 0.0 in above example. */
2151 ConstantInt *C = nullptr;
2152 if (Incr->getOperand(0) == PH)
2153 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
2154 else if (Incr->getOperand(1) == PH)
2155 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
2161 // Ignore negative constants, as the code below doesn't handle them
2162 // correctly. TODO: Remove this restriction.
2163 if (!C->getValue().isStrictlyPositive()) continue;
2165 /* Add new PHINode. */
2166 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
2168 /* create new increment. '++d' in above example. */
2169 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
2170 BinaryOperator *NewIncr =
2171 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
2172 Instruction::FAdd : Instruction::FSub,
2173 NewPH, CFP, "IV.S.next.", Incr);
2175 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
2176 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
2178 /* Remove cast operation */
2179 ShadowUse->replaceAllUsesWith(NewPH);
2180 ShadowUse->eraseFromParent();
2186 /// If Cond has an operand that is an expression of an IV, set the IV user and
2187 /// stride information and return true, otherwise return false.
2188 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
2189 for (IVStrideUse &U : IU)
2190 if (U.getUser() == Cond) {
2191 // NOTE: we could handle setcc instructions with multiple uses here, but
2192 // InstCombine does it as well for simple uses, it's not clear that it
2193 // occurs enough in real life to handle.
2200 /// Rewrite the loop's terminating condition if it uses a max computation.
2202 /// This is a narrow solution to a specific, but acute, problem. For loops
2208 /// } while (++i < n);
2210 /// the trip count isn't just 'n', because 'n' might not be positive. And
2211 /// unfortunately this can come up even for loops where the user didn't use
2212 /// a C do-while loop. For example, seemingly well-behaved top-test loops
2213 /// will commonly be lowered like this:
2219 /// } while (++i < n);
2222 /// and then it's possible for subsequent optimization to obscure the if
2223 /// test in such a way that indvars can't find it.
2225 /// When indvars can't find the if test in loops like this, it creates a
2226 /// max expression, which allows it to give the loop a canonical
2227 /// induction variable:
2230 /// max = n < 1 ? 1 : n;
2233 /// } while (++i != max);
2235 /// Canonical induction variables are necessary because the loop passes
2236 /// are designed around them. The most obvious example of this is the
2237 /// LoopInfo analysis, which doesn't remember trip count values. It
2238 /// expects to be able to rediscover the trip count each time it is
2239 /// needed, and it does this using a simple analysis that only succeeds if
2240 /// the loop has a canonical induction variable.
2242 /// However, when it comes time to generate code, the maximum operation
2243 /// can be quite costly, especially if it's inside of an outer loop.
2245 /// This function solves this problem by detecting this type of loop and
2246 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
2247 /// the instructions for the maximum computation.
2248 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
2249 // Check that the loop matches the pattern we're looking for.
2250 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
2251 Cond->getPredicate() != CmpInst::ICMP_NE)
2254 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
2255 if (!Sel || !Sel->hasOneUse()) return Cond;
2257 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
2258 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2260 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
2262 // Add one to the backedge-taken count to get the trip count.
2263 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
2264 if (IterationCount != SE.getSCEV(Sel)) return Cond;
2266 // Check for a max calculation that matches the pattern. There's no check
2267 // for ICMP_ULE here because the comparison would be with zero, which
2268 // isn't interesting.
2269 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
2270 const SCEVNAryExpr *Max = nullptr;
2271 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
2272 Pred = ICmpInst::ICMP_SLE;
2274 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
2275 Pred = ICmpInst::ICMP_SLT;
2277 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
2278 Pred = ICmpInst::ICMP_ULT;
2285 // To handle a max with more than two operands, this optimization would
2286 // require additional checking and setup.
2287 if (Max->getNumOperands() != 2)
2290 const SCEV *MaxLHS = Max->getOperand(0);
2291 const SCEV *MaxRHS = Max->getOperand(1);
2293 // ScalarEvolution canonicalizes constants to the left. For < and >, look
2294 // for a comparison with 1. For <= and >=, a comparison with zero.
2296 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
2299 // Check the relevant induction variable for conformance to
2301 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
2302 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2303 if (!AR || !AR->isAffine() ||
2304 AR->getStart() != One ||
2305 AR->getStepRecurrence(SE) != One)
2308 assert(AR->getLoop() == L &&
2309 "Loop condition operand is an addrec in a different loop!");
2311 // Check the right operand of the select, and remember it, as it will
2312 // be used in the new comparison instruction.
2313 Value *NewRHS = nullptr;
2314 if (ICmpInst::isTrueWhenEqual(Pred)) {
2315 // Look for n+1, and grab n.
2316 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
2317 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2318 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2319 NewRHS = BO->getOperand(0);
2320 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
2321 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2322 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2323 NewRHS = BO->getOperand(0);
2326 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
2327 NewRHS = Sel->getOperand(1);
2328 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
2329 NewRHS = Sel->getOperand(2);
2330 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
2331 NewRHS = SU->getValue();
2333 // Max doesn't match expected pattern.
2336 // Determine the new comparison opcode. It may be signed or unsigned,
2337 // and the original comparison may be either equality or inequality.
2338 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
2339 Pred = CmpInst::getInversePredicate(Pred);
2341 // Ok, everything looks ok to change the condition into an SLT or SGE and
2342 // delete the max calculation.
2344 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
2346 // Delete the max calculation instructions.
2347 Cond->replaceAllUsesWith(NewCond);
2348 CondUse->setUser(NewCond);
2349 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2350 Cond->eraseFromParent();
2351 Sel->eraseFromParent();
2352 if (Cmp->use_empty())
2353 Cmp->eraseFromParent();
2357 /// Change loop terminating condition to use the postinc iv when possible.
2359 LSRInstance::OptimizeLoopTermCond() {
2360 SmallPtrSet<Instruction *, 4> PostIncs;
2362 // We need a different set of heuristics for rotated and non-rotated loops.
2363 // If a loop is rotated then the latch is also the backedge, so inserting
2364 // post-inc expressions just before the latch is ideal. To reduce live ranges
2365 // it also makes sense to rewrite terminating conditions to use post-inc
2368 // If the loop is not rotated then the latch is not a backedge; the latch
2369 // check is done in the loop head. Adding post-inc expressions before the
2370 // latch will cause overlapping live-ranges of pre-inc and post-inc expressions
2371 // in the loop body. In this case we do *not* want to use post-inc expressions
2372 // in the latch check, and we want to insert post-inc expressions before
2374 BasicBlock *LatchBlock = L->getLoopLatch();
2375 SmallVector<BasicBlock*, 8> ExitingBlocks;
2376 L->getExitingBlocks(ExitingBlocks);
2377 if (llvm::all_of(ExitingBlocks, [&LatchBlock](const BasicBlock *BB) {
2378 return LatchBlock != BB;
2380 // The backedge doesn't exit the loop; treat this as a head-tested loop.
2381 IVIncInsertPos = LatchBlock->getTerminator();
2385 // Otherwise treat this as a rotated loop.
2386 for (BasicBlock *ExitingBlock : ExitingBlocks) {
2387 // Get the terminating condition for the loop if possible. If we
2388 // can, we want to change it to use a post-incremented version of its
2389 // induction variable, to allow coalescing the live ranges for the IV into
2390 // one register value.
2392 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2395 // FIXME: Overly conservative, termination condition could be an 'or' etc..
2396 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
2399 // Search IVUsesByStride to find Cond's IVUse if there is one.
2400 IVStrideUse *CondUse = nullptr;
2401 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2402 if (!FindIVUserForCond(Cond, CondUse))
2405 // If the trip count is computed in terms of a max (due to ScalarEvolution
2406 // being unable to find a sufficient guard, for example), change the loop
2407 // comparison to use SLT or ULT instead of NE.
2408 // One consequence of doing this now is that it disrupts the count-down
2409 // optimization. That's not always a bad thing though, because in such
2410 // cases it may still be worthwhile to avoid a max.
2411 Cond = OptimizeMax(Cond, CondUse);
2413 // If this exiting block dominates the latch block, it may also use
2414 // the post-inc value if it won't be shared with other uses.
2415 // Check for dominance.
2416 if (!DT.dominates(ExitingBlock, LatchBlock))
2419 // Conservatively avoid trying to use the post-inc value in non-latch
2420 // exits if there may be pre-inc users in intervening blocks.
2421 if (LatchBlock != ExitingBlock)
2422 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
2423 // Test if the use is reachable from the exiting block. This dominator
2424 // query is a conservative approximation of reachability.
2425 if (&*UI != CondUse &&
2426 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
2427 // Conservatively assume there may be reuse if the quotient of their
2428 // strides could be a legal scale.
2429 const SCEV *A = IU.getStride(*CondUse, L);
2430 const SCEV *B = IU.getStride(*UI, L);
2431 if (!A || !B) continue;
2432 if (SE.getTypeSizeInBits(A->getType()) !=
2433 SE.getTypeSizeInBits(B->getType())) {
2434 if (SE.getTypeSizeInBits(A->getType()) >
2435 SE.getTypeSizeInBits(B->getType()))
2436 B = SE.getSignExtendExpr(B, A->getType());
2438 A = SE.getSignExtendExpr(A, B->getType());
2440 if (const SCEVConstant *D =
2441 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
2442 const ConstantInt *C = D->getValue();
2443 // Stride of one or negative one can have reuse with non-addresses.
2444 if (C->isOne() || C->isMinusOne())
2445 goto decline_post_inc;
2446 // Avoid weird situations.
2447 if (C->getValue().getMinSignedBits() >= 64 ||
2448 C->getValue().isMinSignedValue())
2449 goto decline_post_inc;
2450 // Check for possible scaled-address reuse.
2451 if (isAddressUse(TTI, UI->getUser(), UI->getOperandValToReplace())) {
2452 MemAccessTy AccessTy = getAccessType(
2453 TTI, UI->getUser(), UI->getOperandValToReplace());
2454 int64_t Scale = C->getSExtValue();
2455 if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
2457 /*HasBaseReg=*/false, Scale,
2458 AccessTy.AddrSpace))
2459 goto decline_post_inc;
2461 if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
2463 /*HasBaseReg=*/false, Scale,
2464 AccessTy.AddrSpace))
2465 goto decline_post_inc;
2470 LLVM_DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
2473 // It's possible for the setcc instruction to be anywhere in the loop, and
2474 // possible for it to have multiple users. If it is not immediately before
2475 // the exiting block branch, move it.
2476 if (&*++BasicBlock::iterator(Cond) != TermBr) {
2477 if (Cond->hasOneUse()) {
2478 Cond->moveBefore(TermBr);
2480 // Clone the terminating condition and insert into the loopend.
2481 ICmpInst *OldCond = Cond;
2482 Cond = cast<ICmpInst>(Cond->clone());
2483 Cond->setName(L->getHeader()->getName() + ".termcond");
2484 ExitingBlock->getInstList().insert(TermBr->getIterator(), Cond);
2486 // Clone the IVUse, as the old use still exists!
2487 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
2488 TermBr->replaceUsesOfWith(OldCond, Cond);
2492 // If we get to here, we know that we can transform the setcc instruction to
2493 // use the post-incremented version of the IV, allowing us to coalesce the
2494 // live ranges for the IV correctly.
2495 CondUse->transformToPostInc(L);
2498 PostIncs.insert(Cond);
2502 // Determine an insertion point for the loop induction variable increment. It
2503 // must dominate all the post-inc comparisons we just set up, and it must
2504 // dominate the loop latch edge.
2505 IVIncInsertPos = L->getLoopLatch()->getTerminator();
2506 for (Instruction *Inst : PostIncs) {
2508 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
2510 if (BB == Inst->getParent())
2511 IVIncInsertPos = Inst;
2512 else if (BB != IVIncInsertPos->getParent())
2513 IVIncInsertPos = BB->getTerminator();
2517 /// Determine if the given use can accommodate a fixup at the given offset and
2518 /// other details. If so, update the use and return true.
2519 bool LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset,
2520 bool HasBaseReg, LSRUse::KindType Kind,
2521 MemAccessTy AccessTy) {
2522 int64_t NewMinOffset = LU.MinOffset;
2523 int64_t NewMaxOffset = LU.MaxOffset;
2524 MemAccessTy NewAccessTy = AccessTy;
2526 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
2527 // something conservative, however this can pessimize in the case that one of
2528 // the uses will have all its uses outside the loop, for example.
2529 if (LU.Kind != Kind)
2532 // Check for a mismatched access type, and fall back conservatively as needed.
2533 // TODO: Be less conservative when the type is similar and can use the same
2534 // addressing modes.
2535 if (Kind == LSRUse::Address) {
2536 if (AccessTy.MemTy != LU.AccessTy.MemTy) {
2537 NewAccessTy = MemAccessTy::getUnknown(AccessTy.MemTy->getContext(),
2538 AccessTy.AddrSpace);
2542 // Conservatively assume HasBaseReg is true for now.
2543 if (NewOffset < LU.MinOffset) {
2544 if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2545 LU.MaxOffset - NewOffset, HasBaseReg))
2547 NewMinOffset = NewOffset;
2548 } else if (NewOffset > LU.MaxOffset) {
2549 if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2550 NewOffset - LU.MinOffset, HasBaseReg))
2552 NewMaxOffset = NewOffset;
2556 LU.MinOffset = NewMinOffset;
2557 LU.MaxOffset = NewMaxOffset;
2558 LU.AccessTy = NewAccessTy;
2562 /// Return an LSRUse index and an offset value for a fixup which needs the given
2563 /// expression, with the given kind and optional access type. Either reuse an
2564 /// existing use or create a new one, as needed.
2565 std::pair<size_t, int64_t> LSRInstance::getUse(const SCEV *&Expr,
2566 LSRUse::KindType Kind,
2567 MemAccessTy AccessTy) {
2568 const SCEV *Copy = Expr;
2569 int64_t Offset = ExtractImmediate(Expr, SE);
2571 // Basic uses can't accept any offset, for example.
2572 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ nullptr,
2573 Offset, /*HasBaseReg=*/ true)) {
2578 std::pair<UseMapTy::iterator, bool> P =
2579 UseMap.insert(std::make_pair(LSRUse::SCEVUseKindPair(Expr, Kind), 0));
2581 // A use already existed with this base.
2582 size_t LUIdx = P.first->second;
2583 LSRUse &LU = Uses[LUIdx];
2584 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
2586 return std::make_pair(LUIdx, Offset);
2589 // Create a new use.
2590 size_t LUIdx = Uses.size();
2591 P.first->second = LUIdx;
2592 Uses.push_back(LSRUse(Kind, AccessTy));
2593 LSRUse &LU = Uses[LUIdx];
2595 LU.MinOffset = Offset;
2596 LU.MaxOffset = Offset;
2597 return std::make_pair(LUIdx, Offset);
2600 /// Delete the given use from the Uses list.
2601 void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
2602 if (&LU != &Uses.back())
2603 std::swap(LU, Uses.back());
2607 RegUses.swapAndDropUse(LUIdx, Uses.size());
2610 /// Look for a use distinct from OrigLU which is has a formula that has the same
2611 /// registers as the given formula.
2613 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
2614 const LSRUse &OrigLU) {
2615 // Search all uses for the formula. This could be more clever.
2616 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2617 LSRUse &LU = Uses[LUIdx];
2618 // Check whether this use is close enough to OrigLU, to see whether it's
2619 // worthwhile looking through its formulae.
2620 // Ignore ICmpZero uses because they may contain formulae generated by
2621 // GenerateICmpZeroScales, in which case adding fixup offsets may
2623 if (&LU != &OrigLU &&
2624 LU.Kind != LSRUse::ICmpZero &&
2625 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
2626 LU.WidestFixupType == OrigLU.WidestFixupType &&
2627 LU.HasFormulaWithSameRegs(OrigF)) {
2628 // Scan through this use's formulae.
2629 for (const Formula &F : LU.Formulae) {
2630 // Check to see if this formula has the same registers and symbols
2632 if (F.BaseRegs == OrigF.BaseRegs &&
2633 F.ScaledReg == OrigF.ScaledReg &&
2634 F.BaseGV == OrigF.BaseGV &&
2635 F.Scale == OrigF.Scale &&
2636 F.UnfoldedOffset == OrigF.UnfoldedOffset) {
2637 if (F.BaseOffset == 0)
2639 // This is the formula where all the registers and symbols matched;
2640 // there aren't going to be any others. Since we declined it, we
2641 // can skip the rest of the formulae and proceed to the next LSRUse.
2648 // Nothing looked good.
2652 void LSRInstance::CollectInterestingTypesAndFactors() {
2653 SmallSetVector<const SCEV *, 4> Strides;
2655 // Collect interesting types and strides.
2656 SmallVector<const SCEV *, 4> Worklist;
2657 for (const IVStrideUse &U : IU) {
2658 const SCEV *Expr = IU.getExpr(U);
2660 // Collect interesting types.
2661 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2663 // Add strides for mentioned loops.
2664 Worklist.push_back(Expr);
2666 const SCEV *S = Worklist.pop_back_val();
2667 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2668 if (AR->getLoop() == L)
2669 Strides.insert(AR->getStepRecurrence(SE));
2670 Worklist.push_back(AR->getStart());
2671 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2672 Worklist.append(Add->op_begin(), Add->op_end());
2674 } while (!Worklist.empty());
2677 // Compute interesting factors from the set of interesting strides.
2678 for (SmallSetVector<const SCEV *, 4>::const_iterator
2679 I = Strides.begin(), E = Strides.end(); I != E; ++I)
2680 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
2681 std::next(I); NewStrideIter != E; ++NewStrideIter) {
2682 const SCEV *OldStride = *I;
2683 const SCEV *NewStride = *NewStrideIter;
2685 if (SE.getTypeSizeInBits(OldStride->getType()) !=
2686 SE.getTypeSizeInBits(NewStride->getType())) {
2687 if (SE.getTypeSizeInBits(OldStride->getType()) >
2688 SE.getTypeSizeInBits(NewStride->getType()))
2689 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2691 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2693 if (const SCEVConstant *Factor =
2694 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2696 if (Factor->getAPInt().getMinSignedBits() <= 64)
2697 Factors.insert(Factor->getAPInt().getSExtValue());
2698 } else if (const SCEVConstant *Factor =
2699 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2702 if (Factor->getAPInt().getMinSignedBits() <= 64)
2703 Factors.insert(Factor->getAPInt().getSExtValue());
2707 // If all uses use the same type, don't bother looking for truncation-based
2709 if (Types.size() == 1)
2712 LLVM_DEBUG(print_factors_and_types(dbgs()));
2715 /// Helper for CollectChains that finds an IV operand (computed by an AddRec in
2716 /// this loop) within [OI,OE) or returns OE. If IVUsers mapped Instructions to
2717 /// IVStrideUses, we could partially skip this.
2718 static User::op_iterator
2719 findIVOperand(User::op_iterator OI, User::op_iterator OE,
2720 Loop *L, ScalarEvolution &SE) {
2721 for(; OI != OE; ++OI) {
2722 if (Instruction *Oper = dyn_cast<Instruction>(*OI)) {
2723 if (!SE.isSCEVable(Oper->getType()))
2726 if (const SCEVAddRecExpr *AR =
2727 dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) {
2728 if (AR->getLoop() == L)
2736 /// IVChain logic must consistently peek base TruncInst operands, so wrap it in
2737 /// a convenient helper.
2738 static Value *getWideOperand(Value *Oper) {
2739 if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper))
2740 return Trunc->getOperand(0);
2744 /// Return true if we allow an IV chain to include both types.
2745 static bool isCompatibleIVType(Value *LVal, Value *RVal) {
2746 Type *LType = LVal->getType();
2747 Type *RType = RVal->getType();
2748 return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy() &&
2749 // Different address spaces means (possibly)
2750 // different types of the pointer implementation,
2751 // e.g. i16 vs i32 so disallow that.
2752 (LType->getPointerAddressSpace() ==
2753 RType->getPointerAddressSpace()));
2756 /// Return an approximation of this SCEV expression's "base", or NULL for any
2757 /// constant. Returning the expression itself is conservative. Returning a
2758 /// deeper subexpression is more precise and valid as long as it isn't less
2759 /// complex than another subexpression. For expressions involving multiple
2760 /// unscaled values, we need to return the pointer-type SCEVUnknown. This avoids
2761 /// forming chains across objects, such as: PrevOper==a[i], IVOper==b[i],
2764 /// Since SCEVUnknown is the rightmost type, and pointers are the rightmost
2765 /// SCEVUnknown, we simply return the rightmost SCEV operand.
2766 static const SCEV *getExprBase(const SCEV *S) {
2767 switch (S->getSCEVType()) {
2768 default: // uncluding scUnknown.
2773 return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand());
2775 return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand());
2777 return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand());
2779 // Skip over scaled operands (scMulExpr) to follow add operands as long as
2780 // there's nothing more complex.
2781 // FIXME: not sure if we want to recognize negation.
2782 const SCEVAddExpr *Add = cast<SCEVAddExpr>(S);
2783 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(Add->op_end()),
2784 E(Add->op_begin()); I != E; ++I) {
2785 const SCEV *SubExpr = *I;
2786 if (SubExpr->getSCEVType() == scAddExpr)
2787 return getExprBase(SubExpr);
2789 if (SubExpr->getSCEVType() != scMulExpr)
2792 return S; // all operands are scaled, be conservative.
2795 return getExprBase(cast<SCEVAddRecExpr>(S)->getStart());
2799 /// Return true if the chain increment is profitable to expand into a loop
2800 /// invariant value, which may require its own register. A profitable chain
2801 /// increment will be an offset relative to the same base. We allow such offsets
2802 /// to potentially be used as chain increment as long as it's not obviously
2803 /// expensive to expand using real instructions.
2804 bool IVChain::isProfitableIncrement(const SCEV *OperExpr,
2805 const SCEV *IncExpr,
2806 ScalarEvolution &SE) {
2807 // Aggressively form chains when -stress-ivchain.
2811 // Do not replace a constant offset from IV head with a nonconstant IV
2813 if (!isa<SCEVConstant>(IncExpr)) {
2814 const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Incs[0].IVOperand));
2815 if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr)))
2819 SmallPtrSet<const SCEV*, 8> Processed;
2820 return !isHighCostExpansion(IncExpr, Processed, SE);
2823 /// Return true if the number of registers needed for the chain is estimated to
2824 /// be less than the number required for the individual IV users. First prohibit
2825 /// any IV users that keep the IV live across increments (the Users set should
2826 /// be empty). Next count the number and type of increments in the chain.
2828 /// Chaining IVs can lead to considerable code bloat if ISEL doesn't
2829 /// effectively use postinc addressing modes. Only consider it profitable it the
2830 /// increments can be computed in fewer registers when chained.
2832 /// TODO: Consider IVInc free if it's already used in another chains.
2834 isProfitableChain(IVChain &Chain, SmallPtrSetImpl<Instruction*> &Users,
2835 ScalarEvolution &SE) {
2839 if (!Chain.hasIncs())
2842 if (!Users.empty()) {
2843 LLVM_DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " users:\n";
2844 for (Instruction *Inst
2845 : Users) { dbgs() << " " << *Inst << "\n"; });
2848 assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
2850 // The chain itself may require a register, so intialize cost to 1.
2853 // A complete chain likely eliminates the need for keeping the original IV in
2854 // a register. LSR does not currently know how to form a complete chain unless
2855 // the header phi already exists.
2856 if (isa<PHINode>(Chain.tailUserInst())
2857 && SE.getSCEV(Chain.tailUserInst()) == Chain.Incs[0].IncExpr) {
2860 const SCEV *LastIncExpr = nullptr;
2861 unsigned NumConstIncrements = 0;
2862 unsigned NumVarIncrements = 0;
2863 unsigned NumReusedIncrements = 0;
2864 for (const IVInc &Inc : Chain) {
2865 if (Inc.IncExpr->isZero())
2868 // Incrementing by zero or some constant is neutral. We assume constants can
2869 // be folded into an addressing mode or an add's immediate operand.
2870 if (isa<SCEVConstant>(Inc.IncExpr)) {
2871 ++NumConstIncrements;
2875 if (Inc.IncExpr == LastIncExpr)
2876 ++NumReusedIncrements;
2880 LastIncExpr = Inc.IncExpr;
2882 // An IV chain with a single increment is handled by LSR's postinc
2883 // uses. However, a chain with multiple increments requires keeping the IV's
2884 // value live longer than it needs to be if chained.
2885 if (NumConstIncrements > 1)
2888 // Materializing increment expressions in the preheader that didn't exist in
2889 // the original code may cost a register. For example, sign-extended array
2890 // indices can produce ridiculous increments like this:
2891 // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64)))
2892 cost += NumVarIncrements;
2894 // Reusing variable increments likely saves a register to hold the multiple of
2896 cost -= NumReusedIncrements;
2898 LLVM_DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " Cost: " << cost
2904 /// Add this IV user to an existing chain or make it the head of a new chain.
2905 void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper,
2906 SmallVectorImpl<ChainUsers> &ChainUsersVec) {
2907 // When IVs are used as types of varying widths, they are generally converted
2908 // to a wider type with some uses remaining narrow under a (free) trunc.
2909 Value *const NextIV = getWideOperand(IVOper);
2910 const SCEV *const OperExpr = SE.getSCEV(NextIV);
2911 const SCEV *const OperExprBase = getExprBase(OperExpr);
2913 // Visit all existing chains. Check if its IVOper can be computed as a
2914 // profitable loop invariant increment from the last link in the Chain.
2915 unsigned ChainIdx = 0, NChains = IVChainVec.size();
2916 const SCEV *LastIncExpr = nullptr;
2917 for (; ChainIdx < NChains; ++ChainIdx) {
2918 IVChain &Chain = IVChainVec[ChainIdx];
2920 // Prune the solution space aggressively by checking that both IV operands
2921 // are expressions that operate on the same unscaled SCEVUnknown. This
2922 // "base" will be canceled by the subsequent getMinusSCEV call. Checking
2923 // first avoids creating extra SCEV expressions.
2924 if (!StressIVChain && Chain.ExprBase != OperExprBase)
2927 Value *PrevIV = getWideOperand(Chain.Incs.back().IVOperand);
2928 if (!isCompatibleIVType(PrevIV, NextIV))
2931 // A phi node terminates a chain.
2932 if (isa<PHINode>(UserInst) && isa<PHINode>(Chain.tailUserInst()))
2935 // The increment must be loop-invariant so it can be kept in a register.
2936 const SCEV *PrevExpr = SE.getSCEV(PrevIV);
2937 const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr);
2938 if (!SE.isLoopInvariant(IncExpr, L))
2941 if (Chain.isProfitableIncrement(OperExpr, IncExpr, SE)) {
2942 LastIncExpr = IncExpr;
2946 // If we haven't found a chain, create a new one, unless we hit the max. Don't
2947 // bother for phi nodes, because they must be last in the chain.
2948 if (ChainIdx == NChains) {
2949 if (isa<PHINode>(UserInst))
2951 if (NChains >= MaxChains && !StressIVChain) {
2952 LLVM_DEBUG(dbgs() << "IV Chain Limit\n");
2955 LastIncExpr = OperExpr;
2956 // IVUsers may have skipped over sign/zero extensions. We don't currently
2957 // attempt to form chains involving extensions unless they can be hoisted
2958 // into this loop's AddRec.
2959 if (!isa<SCEVAddRecExpr>(LastIncExpr))
2962 IVChainVec.push_back(IVChain(IVInc(UserInst, IVOper, LastIncExpr),
2964 ChainUsersVec.resize(NChains);
2965 LLVM_DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Head: (" << *UserInst
2966 << ") IV=" << *LastIncExpr << "\n");
2968 LLVM_DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Inc: (" << *UserInst
2969 << ") IV+" << *LastIncExpr << "\n");
2970 // Add this IV user to the end of the chain.
2971 IVChainVec[ChainIdx].add(IVInc(UserInst, IVOper, LastIncExpr));
2973 IVChain &Chain = IVChainVec[ChainIdx];
2975 SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers;
2976 // This chain's NearUsers become FarUsers.
2977 if (!LastIncExpr->isZero()) {
2978 ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(),
2983 // All other uses of IVOperand become near uses of the chain.
2984 // We currently ignore intermediate values within SCEV expressions, assuming
2985 // they will eventually be used be the current chain, or can be computed
2986 // from one of the chain increments. To be more precise we could
2987 // transitively follow its user and only add leaf IV users to the set.
2988 for (User *U : IVOper->users()) {
2989 Instruction *OtherUse = dyn_cast<Instruction>(U);
2992 // Uses in the chain will no longer be uses if the chain is formed.
2993 // Include the head of the chain in this iteration (not Chain.begin()).
2994 IVChain::const_iterator IncIter = Chain.Incs.begin();
2995 IVChain::const_iterator IncEnd = Chain.Incs.end();
2996 for( ; IncIter != IncEnd; ++IncIter) {
2997 if (IncIter->UserInst == OtherUse)
3000 if (IncIter != IncEnd)
3003 if (SE.isSCEVable(OtherUse->getType())
3004 && !isa<SCEVUnknown>(SE.getSCEV(OtherUse))
3005 && IU.isIVUserOrOperand(OtherUse)) {
3008 NearUsers.insert(OtherUse);
3011 // Since this user is part of the chain, it's no longer considered a use
3013 ChainUsersVec[ChainIdx].FarUsers.erase(UserInst);
3016 /// Populate the vector of Chains.
3018 /// This decreases ILP at the architecture level. Targets with ample registers,
3019 /// multiple memory ports, and no register renaming probably don't want
3020 /// this. However, such targets should probably disable LSR altogether.
3022 /// The job of LSR is to make a reasonable choice of induction variables across
3023 /// the loop. Subsequent passes can easily "unchain" computation exposing more
3024 /// ILP *within the loop* if the target wants it.
3026 /// Finding the best IV chain is potentially a scheduling problem. Since LSR
3027 /// will not reorder memory operations, it will recognize this as a chain, but
3028 /// will generate redundant IV increments. Ideally this would be corrected later
3029 /// by a smart scheduler:
3035 /// TODO: Walk the entire domtree within this loop, not just the path to the
3036 /// loop latch. This will discover chains on side paths, but requires
3037 /// maintaining multiple copies of the Chains state.
3038 void LSRInstance::CollectChains() {
3039 LLVM_DEBUG(dbgs() << "Collecting IV Chains.\n");
3040 SmallVector<ChainUsers, 8> ChainUsersVec;
3042 SmallVector<BasicBlock *,8> LatchPath;
3043 BasicBlock *LoopHeader = L->getHeader();
3044 for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch());
3045 Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) {
3046 LatchPath.push_back(Rung->getBlock());
3048 LatchPath.push_back(LoopHeader);
3050 // Walk the instruction stream from the loop header to the loop latch.
3051 for (BasicBlock *BB : reverse(LatchPath)) {
3052 for (Instruction &I : *BB) {
3053 // Skip instructions that weren't seen by IVUsers analysis.
3054 if (isa<PHINode>(I) || !IU.isIVUserOrOperand(&I))
3057 // Ignore users that are part of a SCEV expression. This way we only
3058 // consider leaf IV Users. This effectively rediscovers a portion of
3059 // IVUsers analysis but in program order this time.
3060 if (SE.isSCEVable(I.getType()) && !isa<SCEVUnknown>(SE.getSCEV(&I)))
3063 // Remove this instruction from any NearUsers set it may be in.
3064 for (unsigned ChainIdx = 0, NChains = IVChainVec.size();
3065 ChainIdx < NChains; ++ChainIdx) {
3066 ChainUsersVec[ChainIdx].NearUsers.erase(&I);
3068 // Search for operands that can be chained.
3069 SmallPtrSet<Instruction*, 4> UniqueOperands;
3070 User::op_iterator IVOpEnd = I.op_end();
3071 User::op_iterator IVOpIter = findIVOperand(I.op_begin(), IVOpEnd, L, SE);
3072 while (IVOpIter != IVOpEnd) {
3073 Instruction *IVOpInst = cast<Instruction>(*IVOpIter);
3074 if (UniqueOperands.insert(IVOpInst).second)
3075 ChainInstruction(&I, IVOpInst, ChainUsersVec);
3076 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
3078 } // Continue walking down the instructions.
3079 } // Continue walking down the domtree.
3080 // Visit phi backedges to determine if the chain can generate the IV postinc.
3081 for (PHINode &PN : L->getHeader()->phis()) {
3082 if (!SE.isSCEVable(PN.getType()))
3086 dyn_cast<Instruction>(PN.getIncomingValueForBlock(L->getLoopLatch()));
3088 ChainInstruction(&PN, IncV, ChainUsersVec);
3090 // Remove any unprofitable chains.
3091 unsigned ChainIdx = 0;
3092 for (unsigned UsersIdx = 0, NChains = IVChainVec.size();
3093 UsersIdx < NChains; ++UsersIdx) {
3094 if (!isProfitableChain(IVChainVec[UsersIdx],
3095 ChainUsersVec[UsersIdx].FarUsers, SE))
3097 // Preserve the chain at UsesIdx.
3098 if (ChainIdx != UsersIdx)
3099 IVChainVec[ChainIdx] = IVChainVec[UsersIdx];
3100 FinalizeChain(IVChainVec[ChainIdx]);
3103 IVChainVec.resize(ChainIdx);
3106 void LSRInstance::FinalizeChain(IVChain &Chain) {
3107 assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
3108 LLVM_DEBUG(dbgs() << "Final Chain: " << *Chain.Incs[0].UserInst << "\n");
3110 for (const IVInc &Inc : Chain) {
3111 LLVM_DEBUG(dbgs() << " Inc: " << *Inc.UserInst << "\n");
3112 auto UseI = find(Inc.UserInst->operands(), Inc.IVOperand);
3113 assert(UseI != Inc.UserInst->op_end() && "cannot find IV operand");
3114 IVIncSet.insert(UseI);
3118 /// Return true if the IVInc can be folded into an addressing mode.
3119 static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst,
3120 Value *Operand, const TargetTransformInfo &TTI) {
3121 const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr);
3122 if (!IncConst || !isAddressUse(TTI, UserInst, Operand))
3125 if (IncConst->getAPInt().getMinSignedBits() > 64)
3128 MemAccessTy AccessTy = getAccessType(TTI, UserInst, Operand);
3129 int64_t IncOffset = IncConst->getValue()->getSExtValue();
3130 if (!isAlwaysFoldable(TTI, LSRUse::Address, AccessTy, /*BaseGV=*/nullptr,
3131 IncOffset, /*HasBaseReg=*/false))
3137 /// Generate an add or subtract for each IVInc in a chain to materialize the IV
3138 /// user's operand from the previous IV user's operand.
3139 void LSRInstance::GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
3140 SmallVectorImpl<WeakTrackingVH> &DeadInsts) {
3141 // Find the new IVOperand for the head of the chain. It may have been replaced
3143 const IVInc &Head = Chain.Incs[0];
3144 User::op_iterator IVOpEnd = Head.UserInst->op_end();
3145 // findIVOperand returns IVOpEnd if it can no longer find a valid IV user.
3146 User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(),
3148 Value *IVSrc = nullptr;
3149 while (IVOpIter != IVOpEnd) {
3150 IVSrc = getWideOperand(*IVOpIter);
3152 // If this operand computes the expression that the chain needs, we may use
3153 // it. (Check this after setting IVSrc which is used below.)
3155 // Note that if Head.IncExpr is wider than IVSrc, then this phi is too
3156 // narrow for the chain, so we can no longer use it. We do allow using a
3157 // wider phi, assuming the LSR checked for free truncation. In that case we
3158 // should already have a truncate on this operand such that
3159 // getSCEV(IVSrc) == IncExpr.
3160 if (SE.getSCEV(*IVOpIter) == Head.IncExpr
3161 || SE.getSCEV(IVSrc) == Head.IncExpr) {
3164 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
3166 if (IVOpIter == IVOpEnd) {
3167 // Gracefully give up on this chain.
3168 LLVM_DEBUG(dbgs() << "Concealed chain head: " << *Head.UserInst << "\n");
3171 assert(IVSrc && "Failed to find IV chain source");
3173 LLVM_DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n");
3174 Type *IVTy = IVSrc->getType();
3175 Type *IntTy = SE.getEffectiveSCEVType(IVTy);
3176 const SCEV *LeftOverExpr = nullptr;
3177 for (const IVInc &Inc : Chain) {
3178 Instruction *InsertPt = Inc.UserInst;
3179 if (isa<PHINode>(InsertPt))
3180 InsertPt = L->getLoopLatch()->getTerminator();
3182 // IVOper will replace the current IV User's operand. IVSrc is the IV
3183 // value currently held in a register.
3184 Value *IVOper = IVSrc;
3185 if (!Inc.IncExpr->isZero()) {
3186 // IncExpr was the result of subtraction of two narrow values, so must
3188 const SCEV *IncExpr = SE.getNoopOrSignExtend(Inc.IncExpr, IntTy);
3189 LeftOverExpr = LeftOverExpr ?
3190 SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr;
3192 if (LeftOverExpr && !LeftOverExpr->isZero()) {
3193 // Expand the IV increment.
3194 Rewriter.clearPostInc();
3195 Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt);
3196 const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc),
3197 SE.getUnknown(IncV));
3198 IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt);
3200 // If an IV increment can't be folded, use it as the next IV value.
3201 if (!canFoldIVIncExpr(LeftOverExpr, Inc.UserInst, Inc.IVOperand, TTI)) {
3202 assert(IVTy == IVOper->getType() && "inconsistent IV increment type");
3204 LeftOverExpr = nullptr;
3207 Type *OperTy = Inc.IVOperand->getType();
3208 if (IVTy != OperTy) {
3209 assert(SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) &&
3210 "cannot extend a chained IV");
3211 IRBuilder<> Builder(InsertPt);
3212 IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain");
3214 Inc.UserInst->replaceUsesOfWith(Inc.IVOperand, IVOper);
3215 DeadInsts.emplace_back(Inc.IVOperand);
3217 // If LSR created a new, wider phi, we may also replace its postinc. We only
3218 // do this if we also found a wide value for the head of the chain.
3219 if (isa<PHINode>(Chain.tailUserInst())) {
3220 for (PHINode &Phi : L->getHeader()->phis()) {
3221 if (!isCompatibleIVType(&Phi, IVSrc))
3223 Instruction *PostIncV = dyn_cast<Instruction>(
3224 Phi.getIncomingValueForBlock(L->getLoopLatch()));
3225 if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc)))
3227 Value *IVOper = IVSrc;
3228 Type *PostIncTy = PostIncV->getType();
3229 if (IVTy != PostIncTy) {
3230 assert(PostIncTy->isPointerTy() && "mixing int/ptr IV types");
3231 IRBuilder<> Builder(L->getLoopLatch()->getTerminator());
3232 Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc());
3233 IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain");
3235 Phi.replaceUsesOfWith(PostIncV, IVOper);
3236 DeadInsts.emplace_back(PostIncV);
3241 void LSRInstance::CollectFixupsAndInitialFormulae() {
3242 BranchInst *ExitBranch = nullptr;
3243 bool SaveCmp = TTI.canSaveCmp(L, &ExitBranch, &SE, &LI, &DT, &AC, &LibInfo);
3245 for (const IVStrideUse &U : IU) {
3246 Instruction *UserInst = U.getUser();
3247 // Skip IV users that are part of profitable IV Chains.
3248 User::op_iterator UseI =
3249 find(UserInst->operands(), U.getOperandValToReplace());
3250 assert(UseI != UserInst->op_end() && "cannot find IV operand");
3251 if (IVIncSet.count(UseI)) {
3252 LLVM_DEBUG(dbgs() << "Use is in profitable chain: " << **UseI << '\n');
3256 LSRUse::KindType Kind = LSRUse::Basic;
3257 MemAccessTy AccessTy;
3258 if (isAddressUse(TTI, UserInst, U.getOperandValToReplace())) {
3259 Kind = LSRUse::Address;
3260 AccessTy = getAccessType(TTI, UserInst, U.getOperandValToReplace());
3263 const SCEV *S = IU.getExpr(U);
3264 PostIncLoopSet TmpPostIncLoops = U.getPostIncLoops();
3266 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
3267 // (N - i == 0), and this allows (N - i) to be the expression that we work
3268 // with rather than just N or i, so we can consider the register
3269 // requirements for both N and i at the same time. Limiting this code to
3270 // equality icmps is not a problem because all interesting loops use
3271 // equality icmps, thanks to IndVarSimplify.
3272 if (ICmpInst *CI = dyn_cast<ICmpInst>(UserInst)) {
3273 // If CI can be saved in some target, like replaced inside hardware loop
3274 // in PowerPC, no need to generate initial formulae for it.
3275 if (SaveCmp && CI == dyn_cast<ICmpInst>(ExitBranch->getCondition()))
3277 if (CI->isEquality()) {
3278 // Swap the operands if needed to put the OperandValToReplace on the
3279 // left, for consistency.
3280 Value *NV = CI->getOperand(1);
3281 if (NV == U.getOperandValToReplace()) {
3282 CI->setOperand(1, CI->getOperand(0));
3283 CI->setOperand(0, NV);
3284 NV = CI->getOperand(1);
3288 // x == y --> x - y == 0
3289 const SCEV *N = SE.getSCEV(NV);
3290 if (SE.isLoopInvariant(N, L) && isSafeToExpand(N, SE)) {
3291 // S is normalized, so normalize N before folding it into S
3292 // to keep the result normalized.
3293 N = normalizeForPostIncUse(N, TmpPostIncLoops, SE);
3294 Kind = LSRUse::ICmpZero;
3295 S = SE.getMinusSCEV(N, S);
3298 // -1 and the negations of all interesting strides (except the negation
3299 // of -1) are now also interesting.
3300 for (size_t i = 0, e = Factors.size(); i != e; ++i)
3301 if (Factors[i] != -1)
3302 Factors.insert(-(uint64_t)Factors[i]);
3307 // Get or create an LSRUse.
3308 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
3309 size_t LUIdx = P.first;
3310 int64_t Offset = P.second;
3311 LSRUse &LU = Uses[LUIdx];
3313 // Record the fixup.
3314 LSRFixup &LF = LU.getNewFixup();
3315 LF.UserInst = UserInst;
3316 LF.OperandValToReplace = U.getOperandValToReplace();
3317 LF.PostIncLoops = TmpPostIncLoops;
3319 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3321 if (!LU.WidestFixupType ||
3322 SE.getTypeSizeInBits(LU.WidestFixupType) <
3323 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3324 LU.WidestFixupType = LF.OperandValToReplace->getType();
3326 // If this is the first use of this LSRUse, give it a formula.
3327 if (LU.Formulae.empty()) {
3328 InsertInitialFormula(S, LU, LUIdx);
3329 CountRegisters(LU.Formulae.back(), LUIdx);
3333 LLVM_DEBUG(print_fixups(dbgs()));
3336 /// Insert a formula for the given expression into the given use, separating out
3337 /// loop-variant portions from loop-invariant and loop-computable portions.
3339 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
3340 // Mark uses whose expressions cannot be expanded.
3341 if (!isSafeToExpand(S, SE))
3342 LU.RigidFormula = true;
3345 F.initialMatch(S, L, SE);
3346 bool Inserted = InsertFormula(LU, LUIdx, F);
3347 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
3350 /// Insert a simple single-register formula for the given expression into the
3353 LSRInstance::InsertSupplementalFormula(const SCEV *S,
3354 LSRUse &LU, size_t LUIdx) {
3356 F.BaseRegs.push_back(S);
3357 F.HasBaseReg = true;
3358 bool Inserted = InsertFormula(LU, LUIdx, F);
3359 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
3362 /// Note which registers are used by the given formula, updating RegUses.
3363 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
3365 RegUses.countRegister(F.ScaledReg, LUIdx);
3366 for (const SCEV *BaseReg : F.BaseRegs)
3367 RegUses.countRegister(BaseReg, LUIdx);
3370 /// If the given formula has not yet been inserted, add it to the list, and
3371 /// return true. Return false otherwise.
3372 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
3373 // Do not insert formula that we will not be able to expand.
3374 assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F) &&
3375 "Formula is illegal");
3377 if (!LU.InsertFormula(F, *L))
3380 CountRegisters(F, LUIdx);
3384 /// Check for other uses of loop-invariant values which we're tracking. These
3385 /// other uses will pin these values in registers, making them less profitable
3386 /// for elimination.
3387 /// TODO: This currently misses non-constant addrec step registers.
3388 /// TODO: Should this give more weight to users inside the loop?
3390 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
3391 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
3392 SmallPtrSet<const SCEV *, 32> Visited;
3394 while (!Worklist.empty()) {
3395 const SCEV *S = Worklist.pop_back_val();
3397 // Don't process the same SCEV twice
3398 if (!Visited.insert(S).second)
3401 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
3402 Worklist.append(N->op_begin(), N->op_end());
3403 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
3404 Worklist.push_back(C->getOperand());
3405 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
3406 Worklist.push_back(D->getLHS());
3407 Worklist.push_back(D->getRHS());
3408 } else if (const SCEVUnknown *US = dyn_cast<SCEVUnknown>(S)) {
3409 const Value *V = US->getValue();
3410 if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
3411 // Look for instructions defined outside the loop.
3412 if (L->contains(Inst)) continue;
3413 } else if (isa<UndefValue>(V))
3414 // Undef doesn't have a live range, so it doesn't matter.
3416 for (const Use &U : V->uses()) {
3417 const Instruction *UserInst = dyn_cast<Instruction>(U.getUser());
3418 // Ignore non-instructions.
3421 // Ignore instructions in other functions (as can happen with
3423 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
3425 // Ignore instructions not dominated by the loop.
3426 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
3427 UserInst->getParent() :
3428 cast<PHINode>(UserInst)->getIncomingBlock(
3429 PHINode::getIncomingValueNumForOperand(U.getOperandNo()));
3430 if (!DT.dominates(L->getHeader(), UseBB))
3432 // Don't bother if the instruction is in a BB which ends in an EHPad.
3433 if (UseBB->getTerminator()->isEHPad())
3435 // Don't bother rewriting PHIs in catchswitch blocks.
3436 if (isa<CatchSwitchInst>(UserInst->getParent()->getTerminator()))
3438 // Ignore uses which are part of other SCEV expressions, to avoid
3439 // analyzing them multiple times.
3440 if (SE.isSCEVable(UserInst->getType())) {
3441 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
3442 // If the user is a no-op, look through to its uses.
3443 if (!isa<SCEVUnknown>(UserS))
3447 SE.getUnknown(const_cast<Instruction *>(UserInst)));
3451 // Ignore icmp instructions which are already being analyzed.
3452 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
3453 unsigned OtherIdx = !U.getOperandNo();
3454 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
3455 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
3459 std::pair<size_t, int64_t> P = getUse(
3460 S, LSRUse::Basic, MemAccessTy());
3461 size_t LUIdx = P.first;
3462 int64_t Offset = P.second;
3463 LSRUse &LU = Uses[LUIdx];
3464 LSRFixup &LF = LU.getNewFixup();
3465 LF.UserInst = const_cast<Instruction *>(UserInst);
3466 LF.OperandValToReplace = U;
3468 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3469 if (!LU.WidestFixupType ||
3470 SE.getTypeSizeInBits(LU.WidestFixupType) <
3471 SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3472 LU.WidestFixupType = LF.OperandValToReplace->getType();
3473 InsertSupplementalFormula(US, LU, LUIdx);
3474 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
3481 /// Split S into subexpressions which can be pulled out into separate
3482 /// registers. If C is non-null, multiply each subexpression by C.
3484 /// Return remainder expression after factoring the subexpressions captured by
3485 /// Ops. If Ops is complete, return NULL.
3486 static const SCEV *CollectSubexprs(const SCEV *S, const SCEVConstant *C,
3487 SmallVectorImpl<const SCEV *> &Ops,
3489 ScalarEvolution &SE,
3490 unsigned Depth = 0) {
3491 // Arbitrarily cap recursion to protect compile time.
3495 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3496 // Break out add operands.
3497 for (const SCEV *S : Add->operands()) {
3498 const SCEV *Remainder = CollectSubexprs(S, C, Ops, L, SE, Depth+1);
3500 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3503 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
3504 // Split a non-zero base out of an addrec.
3505 if (AR->getStart()->isZero() || !AR->isAffine())
3508 const SCEV *Remainder = CollectSubexprs(AR->getStart(),
3509 C, Ops, L, SE, Depth+1);
3510 // Split the non-zero AddRec unless it is part of a nested recurrence that
3511 // does not pertain to this loop.
3512 if (Remainder && (AR->getLoop() == L || !isa<SCEVAddRecExpr>(Remainder))) {
3513 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3514 Remainder = nullptr;
3516 if (Remainder != AR->getStart()) {
3518 Remainder = SE.getConstant(AR->getType(), 0);
3519 return SE.getAddRecExpr(Remainder,
3520 AR->getStepRecurrence(SE),
3522 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
3525 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3526 // Break (C * (a + b + c)) into C*a + C*b + C*c.
3527 if (Mul->getNumOperands() != 2)
3529 if (const SCEVConstant *Op0 =
3530 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
3531 C = C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0;
3532 const SCEV *Remainder =
3533 CollectSubexprs(Mul->getOperand(1), C, Ops, L, SE, Depth+1);
3535 Ops.push_back(SE.getMulExpr(C, Remainder));
3542 /// Return true if the SCEV represents a value that may end up as a
3543 /// post-increment operation.
3544 static bool mayUsePostIncMode(const TargetTransformInfo &TTI,
3545 LSRUse &LU, const SCEV *S, const Loop *L,
3546 ScalarEvolution &SE) {
3547 if (LU.Kind != LSRUse::Address ||
3548 !LU.AccessTy.getType()->isIntOrIntVectorTy())
3550 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S);
3553 const SCEV *LoopStep = AR->getStepRecurrence(SE);
3554 if (!isa<SCEVConstant>(LoopStep))
3556 if (LU.AccessTy.getType()->getScalarSizeInBits() !=
3557 LoopStep->getType()->getScalarSizeInBits())
3559 // Check if a post-indexed load/store can be used.
3560 if (TTI.isIndexedLoadLegal(TTI.MIM_PostInc, AR->getType()) ||
3561 TTI.isIndexedStoreLegal(TTI.MIM_PostInc, AR->getType())) {
3562 const SCEV *LoopStart = AR->getStart();
3563 if (!isa<SCEVConstant>(LoopStart) && SE.isLoopInvariant(LoopStart, L))
3569 /// Helper function for LSRInstance::GenerateReassociations.
3570 void LSRInstance::GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
3571 const Formula &Base,
3572 unsigned Depth, size_t Idx,
3574 const SCEV *BaseReg = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3575 // Don't generate reassociations for the base register of a value that
3576 // may generate a post-increment operator. The reason is that the
3577 // reassociations cause extra base+register formula to be created,
3578 // and possibly chosen, but the post-increment is more efficient.
3579 if (TTI.shouldFavorPostInc() && mayUsePostIncMode(TTI, LU, BaseReg, L, SE))
3581 SmallVector<const SCEV *, 8> AddOps;
3582 const SCEV *Remainder = CollectSubexprs(BaseReg, nullptr, AddOps, L, SE);
3584 AddOps.push_back(Remainder);
3586 if (AddOps.size() == 1)
3589 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
3592 // Loop-variant "unknown" values are uninteresting; we won't be able to
3593 // do anything meaningful with them.
3594 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
3597 // Don't pull a constant into a register if the constant could be folded
3598 // into an immediate field.
3599 if (isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3600 LU.AccessTy, *J, Base.getNumRegs() > 1))
3603 // Collect all operands except *J.
3604 SmallVector<const SCEV *, 8> InnerAddOps(
3605 ((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
3606 InnerAddOps.append(std::next(J),
3607 ((const SmallVector<const SCEV *, 8> &)AddOps).end());
3609 // Don't leave just a constant behind in a register if the constant could
3610 // be folded into an immediate field.
3611 if (InnerAddOps.size() == 1 &&
3612 isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3613 LU.AccessTy, InnerAddOps[0], Base.getNumRegs() > 1))
3616 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
3617 if (InnerSum->isZero())
3621 // Add the remaining pieces of the add back into the new formula.
3622 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
3623 if (InnerSumSC && SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
3624 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3625 InnerSumSC->getValue()->getZExtValue())) {
3627 (uint64_t)F.UnfoldedOffset + InnerSumSC->getValue()->getZExtValue();
3629 F.ScaledReg = nullptr;
3631 F.BaseRegs.erase(F.BaseRegs.begin() + Idx);
3632 } else if (IsScaledReg)
3633 F.ScaledReg = InnerSum;
3635 F.BaseRegs[Idx] = InnerSum;
3637 // Add J as its own register, or an unfolded immediate.
3638 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
3639 if (SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
3640 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3641 SC->getValue()->getZExtValue()))
3643 (uint64_t)F.UnfoldedOffset + SC->getValue()->getZExtValue();
3645 F.BaseRegs.push_back(*J);
3646 // We may have changed the number of register in base regs, adjust the
3647 // formula accordingly.
3650 if (InsertFormula(LU, LUIdx, F))
3651 // If that formula hadn't been seen before, recurse to find more like
3653 // Add check on Log16(AddOps.size()) - same as Log2_32(AddOps.size()) >> 2)
3654 // Because just Depth is not enough to bound compile time.
3655 // This means that every time AddOps.size() is greater 16^x we will add
3657 GenerateReassociations(LU, LUIdx, LU.Formulae.back(),
3658 Depth + 1 + (Log2_32(AddOps.size()) >> 2));
3662 /// Split out subexpressions from adds and the bases of addrecs.
3663 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
3664 Formula Base, unsigned Depth) {
3665 assert(Base.isCanonical(*L) && "Input must be in the canonical form");
3666 // Arbitrarily cap recursion to protect compile time.
3670 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3671 GenerateReassociationsImpl(LU, LUIdx, Base, Depth, i);
3673 if (Base.Scale == 1)
3674 GenerateReassociationsImpl(LU, LUIdx, Base, Depth,
3675 /* Idx */ -1, /* IsScaledReg */ true);
3678 /// Generate a formula consisting of all of the loop-dominating registers added
3679 /// into a single register.
3680 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
3682 // This method is only interesting on a plurality of registers.
3683 if (Base.BaseRegs.size() + (Base.Scale == 1) +
3684 (Base.UnfoldedOffset != 0) <= 1)
3687 // Flatten the representation, i.e., reg1 + 1*reg2 => reg1 + reg2, before
3688 // processing the formula.
3690 SmallVector<const SCEV *, 4> Ops;
3691 Formula NewBase = Base;
3692 NewBase.BaseRegs.clear();
3693 Type *CombinedIntegerType = nullptr;
3694 for (const SCEV *BaseReg : Base.BaseRegs) {
3695 if (SE.properlyDominates(BaseReg, L->getHeader()) &&
3696 !SE.hasComputableLoopEvolution(BaseReg, L)) {
3697 if (!CombinedIntegerType)
3698 CombinedIntegerType = SE.getEffectiveSCEVType(BaseReg->getType());
3699 Ops.push_back(BaseReg);
3702 NewBase.BaseRegs.push_back(BaseReg);
3705 // If no register is relevant, we're done.
3706 if (Ops.size() == 0)
3709 // Utility function for generating the required variants of the combined
3711 auto GenerateFormula = [&](const SCEV *Sum) {
3712 Formula F = NewBase;
3714 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
3715 // opportunity to fold something. For now, just ignore such cases
3716 // rather than proceed with zero in a register.
3720 F.BaseRegs.push_back(Sum);
3722 (void)InsertFormula(LU, LUIdx, F);
3725 // If we collected at least two registers, generate a formula combining them.
3726 if (Ops.size() > 1) {
3727 SmallVector<const SCEV *, 4> OpsCopy(Ops); // Don't let SE modify Ops.
3728 GenerateFormula(SE.getAddExpr(OpsCopy));
3731 // If we have an unfolded offset, generate a formula combining it with the
3732 // registers collected.
3733 if (NewBase.UnfoldedOffset) {
3734 assert(CombinedIntegerType && "Missing a type for the unfolded offset");
3735 Ops.push_back(SE.getConstant(CombinedIntegerType, NewBase.UnfoldedOffset,
3737 NewBase.UnfoldedOffset = 0;
3738 GenerateFormula(SE.getAddExpr(Ops));
3742 /// Helper function for LSRInstance::GenerateSymbolicOffsets.
3743 void LSRInstance::GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
3744 const Formula &Base, size_t Idx,
3746 const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3747 GlobalValue *GV = ExtractSymbol(G, SE);
3748 if (G->isZero() || !GV)
3752 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3757 F.BaseRegs[Idx] = G;
3758 (void)InsertFormula(LU, LUIdx, F);
3761 /// Generate reuse formulae using symbolic offsets.
3762 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
3764 // We can't add a symbolic offset if the address already contains one.
3765 if (Base.BaseGV) return;
3767 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3768 GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, i);
3769 if (Base.Scale == 1)
3770 GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, /* Idx */ -1,
3771 /* IsScaledReg */ true);
3774 /// Helper function for LSRInstance::GenerateConstantOffsets.
3775 void LSRInstance::GenerateConstantOffsetsImpl(
3776 LSRUse &LU, unsigned LUIdx, const Formula &Base,
3777 const SmallVectorImpl<int64_t> &Worklist, size_t Idx, bool IsScaledReg) {
3779 auto GenerateOffset = [&](const SCEV *G, int64_t Offset) {
3781 F.BaseOffset = (uint64_t)Base.BaseOffset - Offset;
3783 if (isLegalUse(TTI, LU.MinOffset - Offset, LU.MaxOffset - Offset, LU.Kind,
3785 // Add the offset to the base register.
3786 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), Offset), G);
3787 // If it cancelled out, drop the base register, otherwise update it.
3788 if (NewG->isZero()) {
3791 F.ScaledReg = nullptr;
3793 F.deleteBaseReg(F.BaseRegs[Idx]);
3795 } else if (IsScaledReg)
3798 F.BaseRegs[Idx] = NewG;
3800 (void)InsertFormula(LU, LUIdx, F);
3804 const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3806 // With constant offsets and constant steps, we can generate pre-inc
3807 // accesses by having the offset equal the step. So, for access #0 with a
3808 // step of 8, we generate a G - 8 base which would require the first access
3809 // to be ((G - 8) + 8),+,8. The pre-indexed access then updates the pointer
3810 // for itself and hopefully becomes the base for other accesses. This means
3811 // means that a single pre-indexed access can be generated to become the new
3812 // base pointer for each iteration of the loop, resulting in no extra add/sub
3813 // instructions for pointer updating.
3814 if (FavorBackedgeIndex && LU.Kind == LSRUse::Address) {
3815 if (auto *GAR = dyn_cast<SCEVAddRecExpr>(G)) {
3817 dyn_cast<SCEVConstant>(GAR->getStepRecurrence(SE))) {
3818 const APInt &StepInt = StepRec->getAPInt();
3819 int64_t Step = StepInt.isNegative() ?
3820 StepInt.getSExtValue() : StepInt.getZExtValue();
3822 for (int64_t Offset : Worklist) {
3824 GenerateOffset(G, Offset);
3829 for (int64_t Offset : Worklist)
3830 GenerateOffset(G, Offset);
3832 int64_t Imm = ExtractImmediate(G, SE);
3833 if (G->isZero() || Imm == 0)
3836 F.BaseOffset = (uint64_t)F.BaseOffset + Imm;
3837 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3842 F.BaseRegs[Idx] = G;
3843 (void)InsertFormula(LU, LUIdx, F);
3846 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
3847 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
3849 // TODO: For now, just add the min and max offset, because it usually isn't
3850 // worthwhile looking at everything inbetween.
3851 SmallVector<int64_t, 2> Worklist;
3852 Worklist.push_back(LU.MinOffset);
3853 if (LU.MaxOffset != LU.MinOffset)
3854 Worklist.push_back(LU.MaxOffset);
3856 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3857 GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, i);
3858 if (Base.Scale == 1)
3859 GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, /* Idx */ -1,
3860 /* IsScaledReg */ true);
3863 /// For ICmpZero, check to see if we can scale up the comparison. For example, x
3864 /// == y -> x*c == y*c.
3865 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
3867 if (LU.Kind != LSRUse::ICmpZero) return;
3869 // Determine the integer type for the base formula.
3870 Type *IntTy = Base.getType();
3872 if (SE.getTypeSizeInBits(IntTy) > 64) return;
3874 // Don't do this if there is more than one offset.
3875 if (LU.MinOffset != LU.MaxOffset) return;
3877 // Check if transformation is valid. It is illegal to multiply pointer.
3878 if (Base.ScaledReg && Base.ScaledReg->getType()->isPointerTy())
3880 for (const SCEV *BaseReg : Base.BaseRegs)
3881 if (BaseReg->getType()->isPointerTy())
3883 assert(!Base.BaseGV && "ICmpZero use is not legal!");
3885 // Check each interesting stride.
3886 for (int64_t Factor : Factors) {
3887 // Check that the multiplication doesn't overflow.
3888 if (Base.BaseOffset == std::numeric_limits<int64_t>::min() && Factor == -1)
3890 int64_t NewBaseOffset = (uint64_t)Base.BaseOffset * Factor;
3891 if (NewBaseOffset / Factor != Base.BaseOffset)
3893 // If the offset will be truncated at this use, check that it is in bounds.
3894 if (!IntTy->isPointerTy() &&
3895 !ConstantInt::isValueValidForType(IntTy, NewBaseOffset))
3898 // Check that multiplying with the use offset doesn't overflow.
3899 int64_t Offset = LU.MinOffset;
3900 if (Offset == std::numeric_limits<int64_t>::min() && Factor == -1)
3902 Offset = (uint64_t)Offset * Factor;
3903 if (Offset / Factor != LU.MinOffset)
3905 // If the offset will be truncated at this use, check that it is in bounds.
3906 if (!IntTy->isPointerTy() &&
3907 !ConstantInt::isValueValidForType(IntTy, Offset))
3911 F.BaseOffset = NewBaseOffset;
3913 // Check that this scale is legal.
3914 if (!isLegalUse(TTI, Offset, Offset, LU.Kind, LU.AccessTy, F))
3917 // Compensate for the use having MinOffset built into it.
3918 F.BaseOffset = (uint64_t)F.BaseOffset + Offset - LU.MinOffset;
3920 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3922 // Check that multiplying with each base register doesn't overflow.
3923 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
3924 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
3925 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
3929 // Check that multiplying with the scaled register doesn't overflow.
3931 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
3932 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
3936 // Check that multiplying with the unfolded offset doesn't overflow.
3937 if (F.UnfoldedOffset != 0) {
3938 if (F.UnfoldedOffset == std::numeric_limits<int64_t>::min() &&
3941 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
3942 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
3944 // If the offset will be truncated, check that it is in bounds.
3945 if (!IntTy->isPointerTy() &&
3946 !ConstantInt::isValueValidForType(IntTy, F.UnfoldedOffset))
3950 // If we make it here and it's legal, add it.
3951 (void)InsertFormula(LU, LUIdx, F);
3956 /// Generate stride factor reuse formulae by making use of scaled-offset address
3957 /// modes, for example.
3958 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
3959 // Determine the integer type for the base formula.
3960 Type *IntTy = Base.getType();
3963 // If this Formula already has a scaled register, we can't add another one.
3964 // Try to unscale the formula to generate a better scale.
3965 if (Base.Scale != 0 && !Base.unscale())
3968 assert(Base.Scale == 0 && "unscale did not did its job!");
3970 // Check each interesting stride.
3971 for (int64_t Factor : Factors) {
3972 Base.Scale = Factor;
3973 Base.HasBaseReg = Base.BaseRegs.size() > 1;
3974 // Check whether this scale is going to be legal.
3975 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3977 // As a special-case, handle special out-of-loop Basic users specially.
3978 // TODO: Reconsider this special case.
3979 if (LU.Kind == LSRUse::Basic &&
3980 isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LSRUse::Special,
3981 LU.AccessTy, Base) &&
3982 LU.AllFixupsOutsideLoop)
3983 LU.Kind = LSRUse::Special;
3987 // For an ICmpZero, negating a solitary base register won't lead to
3989 if (LU.Kind == LSRUse::ICmpZero &&
3990 !Base.HasBaseReg && Base.BaseOffset == 0 && !Base.BaseGV)
3992 // For each addrec base reg, if its loop is current loop, apply the scale.
3993 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3994 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i]);
3995 if (AR && (AR->getLoop() == L || LU.AllFixupsOutsideLoop)) {
3996 const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3997 if (FactorS->isZero())
3999 // Divide out the factor, ignoring high bits, since we'll be
4000 // scaling the value back up in the end.
4001 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
4002 // TODO: This could be optimized to avoid all the copying.
4004 F.ScaledReg = Quotient;
4005 F.deleteBaseReg(F.BaseRegs[i]);
4006 // The canonical representation of 1*reg is reg, which is already in
4007 // Base. In that case, do not try to insert the formula, it will be
4009 if (F.Scale == 1 && (F.BaseRegs.empty() ||
4010 (AR->getLoop() != L && LU.AllFixupsOutsideLoop)))
4012 // If AllFixupsOutsideLoop is true and F.Scale is 1, we may generate
4013 // non canonical Formula with ScaledReg's loop not being L.
4014 if (F.Scale == 1 && LU.AllFixupsOutsideLoop)
4016 (void)InsertFormula(LU, LUIdx, F);
4023 /// Generate reuse formulae from different IV types.
4024 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
4025 // Don't bother truncating symbolic values.
4026 if (Base.BaseGV) return;
4028 // Determine the integer type for the base formula.
4029 Type *DstTy = Base.getType();
4031 DstTy = SE.getEffectiveSCEVType(DstTy);
4033 for (Type *SrcTy : Types) {
4034 if (SrcTy != DstTy && TTI.isTruncateFree(SrcTy, DstTy)) {
4037 // Sometimes SCEV is able to prove zero during ext transform. It may
4038 // happen if SCEV did not do all possible transforms while creating the
4039 // initial node (maybe due to depth limitations), but it can do them while
4042 const SCEV *NewScaledReg = SE.getAnyExtendExpr(F.ScaledReg, SrcTy);
4043 if (NewScaledReg->isZero())
4045 F.ScaledReg = NewScaledReg;
4047 bool HasZeroBaseReg = false;
4048 for (const SCEV *&BaseReg : F.BaseRegs) {
4049 const SCEV *NewBaseReg = SE.getAnyExtendExpr(BaseReg, SrcTy);
4050 if (NewBaseReg->isZero()) {
4051 HasZeroBaseReg = true;
4054 BaseReg = NewBaseReg;
4059 // TODO: This assumes we've done basic processing on all uses and
4060 // have an idea what the register usage is.
4061 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
4065 (void)InsertFormula(LU, LUIdx, F);
4072 /// Helper class for GenerateCrossUseConstantOffsets. It's used to defer
4073 /// modifications so that the search phase doesn't have to worry about the data
4074 /// structures moving underneath it.
4078 const SCEV *OrigReg;
4080 WorkItem(size_t LI, int64_t I, const SCEV *R)
4081 : LUIdx(LI), Imm(I), OrigReg(R) {}
4083 void print(raw_ostream &OS) const;
4087 } // end anonymous namespace
4089 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
4090 void WorkItem::print(raw_ostream &OS) const {
4091 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
4092 << " , add offset " << Imm;
4095 LLVM_DUMP_METHOD void WorkItem::dump() const {
4096 print(errs()); errs() << '\n';
4100 /// Look for registers which are a constant distance apart and try to form reuse
4101 /// opportunities between them.
4102 void LSRInstance::GenerateCrossUseConstantOffsets() {
4103 // Group the registers by their value without any added constant offset.
4104 using ImmMapTy = std::map<int64_t, const SCEV *>;
4106 DenseMap<const SCEV *, ImmMapTy> Map;
4107 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
4108 SmallVector<const SCEV *, 8> Sequence;
4109 for (const SCEV *Use : RegUses) {
4110 const SCEV *Reg = Use; // Make a copy for ExtractImmediate to modify.
4111 int64_t Imm = ExtractImmediate(Reg, SE);
4112 auto Pair = Map.insert(std::make_pair(Reg, ImmMapTy()));
4114 Sequence.push_back(Reg);
4115 Pair.first->second.insert(std::make_pair(Imm, Use));
4116 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(Use);
4119 // Now examine each set of registers with the same base value. Build up
4120 // a list of work to do and do the work in a separate step so that we're
4121 // not adding formulae and register counts while we're searching.
4122 SmallVector<WorkItem, 32> WorkItems;
4123 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
4124 for (const SCEV *Reg : Sequence) {
4125 const ImmMapTy &Imms = Map.find(Reg)->second;
4127 // It's not worthwhile looking for reuse if there's only one offset.
4128 if (Imms.size() == 1)
4131 LLVM_DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
4132 for (const auto &Entry
4134 << ' ' << Entry.first;
4137 // Examine each offset.
4138 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
4140 const SCEV *OrigReg = J->second;
4142 int64_t JImm = J->first;
4143 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
4145 if (!isa<SCEVConstant>(OrigReg) &&
4146 UsedByIndicesMap[Reg].count() == 1) {
4147 LLVM_DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg
4152 // Conservatively examine offsets between this orig reg a few selected
4154 int64_t First = Imms.begin()->first;
4155 int64_t Last = std::prev(Imms.end())->first;
4156 // Compute (First + Last) / 2 without overflow using the fact that
4157 // First + Last = 2 * (First + Last) + (First ^ Last).
4158 int64_t Avg = (First & Last) + ((First ^ Last) >> 1);
4159 // If the result is negative and First is odd and Last even (or vice versa),
4160 // we rounded towards -inf. Add 1 in that case, to round towards 0.
4161 Avg = Avg + ((First ^ Last) & ((uint64_t)Avg >> 63));
4162 ImmMapTy::const_iterator OtherImms[] = {
4163 Imms.begin(), std::prev(Imms.end()),
4164 Imms.lower_bound(Avg)};
4165 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
4166 ImmMapTy::const_iterator M = OtherImms[i];
4167 if (M == J || M == JE) continue;
4169 // Compute the difference between the two.
4170 int64_t Imm = (uint64_t)JImm - M->first;
4171 for (unsigned LUIdx : UsedByIndices.set_bits())
4172 // Make a memo of this use, offset, and register tuple.
4173 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)).second)
4174 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
4181 UsedByIndicesMap.clear();
4182 UniqueItems.clear();
4184 // Now iterate through the worklist and add new formulae.
4185 for (const WorkItem &WI : WorkItems) {
4186 size_t LUIdx = WI.LUIdx;
4187 LSRUse &LU = Uses[LUIdx];
4188 int64_t Imm = WI.Imm;
4189 const SCEV *OrigReg = WI.OrigReg;
4191 Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
4192 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
4193 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
4195 // TODO: Use a more targeted data structure.
4196 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
4197 Formula F = LU.Formulae[L];
4198 // FIXME: The code for the scaled and unscaled registers looks
4199 // very similar but slightly different. Investigate if they
4200 // could be merged. That way, we would not have to unscale the
4203 // Use the immediate in the scaled register.
4204 if (F.ScaledReg == OrigReg) {
4205 int64_t Offset = (uint64_t)F.BaseOffset + Imm * (uint64_t)F.Scale;
4206 // Don't create 50 + reg(-50).
4207 if (F.referencesReg(SE.getSCEV(
4208 ConstantInt::get(IntTy, -(uint64_t)Offset))))
4211 NewF.BaseOffset = Offset;
4212 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
4215 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
4217 // If the new scale is a constant in a register, and adding the constant
4218 // value to the immediate would produce a value closer to zero than the
4219 // immediate itself, then the formula isn't worthwhile.
4220 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
4221 if (C->getValue()->isNegative() != (NewF.BaseOffset < 0) &&
4222 (C->getAPInt().abs() * APInt(BitWidth, F.Scale))
4223 .ule(std::abs(NewF.BaseOffset)))
4227 NewF.canonicalize(*this->L);
4228 (void)InsertFormula(LU, LUIdx, NewF);
4230 // Use the immediate in a base register.
4231 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
4232 const SCEV *BaseReg = F.BaseRegs[N];
4233 if (BaseReg != OrigReg)
4236 NewF.BaseOffset = (uint64_t)NewF.BaseOffset + Imm;
4237 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset,
4238 LU.Kind, LU.AccessTy, NewF)) {
4239 if (TTI.shouldFavorPostInc() &&
4240 mayUsePostIncMode(TTI, LU, OrigReg, this->L, SE))
4242 if (!TTI.isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
4245 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
4247 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
4249 // If the new formula has a constant in a register, and adding the
4250 // constant value to the immediate would produce a value closer to
4251 // zero than the immediate itself, then the formula isn't worthwhile.
4252 for (const SCEV *NewReg : NewF.BaseRegs)
4253 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewReg))
4254 if ((C->getAPInt() + NewF.BaseOffset)
4256 .slt(std::abs(NewF.BaseOffset)) &&
4257 (C->getAPInt() + NewF.BaseOffset).countTrailingZeros() >=
4258 countTrailingZeros<uint64_t>(NewF.BaseOffset))
4262 NewF.canonicalize(*this->L);
4263 (void)InsertFormula(LU, LUIdx, NewF);
4272 /// Generate formulae for each use.
4274 LSRInstance::GenerateAllReuseFormulae() {
4275 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
4276 // queries are more precise.
4277 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4278 LSRUse &LU = Uses[LUIdx];
4279 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4280 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
4281 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4282 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
4284 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4285 LSRUse &LU = Uses[LUIdx];
4286 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4287 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
4288 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4289 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
4290 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4291 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
4292 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4293 GenerateScales(LU, LUIdx, LU.Formulae[i]);
4295 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4296 LSRUse &LU = Uses[LUIdx];
4297 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4298 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
4301 GenerateCrossUseConstantOffsets();
4303 LLVM_DEBUG(dbgs() << "\n"
4304 "After generating reuse formulae:\n";
4305 print_uses(dbgs()));
4308 /// If there are multiple formulae with the same set of registers used
4309 /// by other uses, pick the best one and delete the others.
4310 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
4311 DenseSet<const SCEV *> VisitedRegs;
4312 SmallPtrSet<const SCEV *, 16> Regs;
4313 SmallPtrSet<const SCEV *, 16> LoserRegs;
4315 bool ChangedFormulae = false;
4318 // Collect the best formula for each unique set of shared registers. This
4319 // is reset for each use.
4320 using BestFormulaeTy =
4321 DenseMap<SmallVector<const SCEV *, 4>, size_t, UniquifierDenseMapInfo>;
4323 BestFormulaeTy BestFormulae;
4325 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4326 LSRUse &LU = Uses[LUIdx];
4327 LLVM_DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs());
4331 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
4332 FIdx != NumForms; ++FIdx) {
4333 Formula &F = LU.Formulae[FIdx];
4335 // Some formulas are instant losers. For example, they may depend on
4336 // nonexistent AddRecs from other loops. These need to be filtered
4337 // immediately, otherwise heuristics could choose them over others leading
4338 // to an unsatisfactory solution. Passing LoserRegs into RateFormula here
4339 // avoids the need to recompute this information across formulae using the
4340 // same bad AddRec. Passing LoserRegs is also essential unless we remove
4341 // the corresponding bad register from the Regs set.
4342 Cost CostF(L, SE, TTI);
4344 CostF.RateFormula(F, Regs, VisitedRegs, LU, &LoserRegs);
4345 if (CostF.isLoser()) {
4346 // During initial formula generation, undesirable formulae are generated
4347 // by uses within other loops that have some non-trivial address mode or
4348 // use the postinc form of the IV. LSR needs to provide these formulae
4349 // as the basis of rediscovering the desired formula that uses an AddRec
4350 // corresponding to the existing phi. Once all formulae have been
4351 // generated, these initial losers may be pruned.
4352 LLVM_DEBUG(dbgs() << " Filtering loser "; F.print(dbgs());
4356 SmallVector<const SCEV *, 4> Key;
4357 for (const SCEV *Reg : F.BaseRegs) {
4358 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
4362 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
4363 Key.push_back(F.ScaledReg);
4364 // Unstable sort by host order ok, because this is only used for
4368 std::pair<BestFormulaeTy::const_iterator, bool> P =
4369 BestFormulae.insert(std::make_pair(Key, FIdx));
4373 Formula &Best = LU.Formulae[P.first->second];
4375 Cost CostBest(L, SE, TTI);
4377 CostBest.RateFormula(Best, Regs, VisitedRegs, LU);
4378 if (CostF.isLess(CostBest))
4380 LLVM_DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
4382 " in favor of formula ";
4383 Best.print(dbgs()); dbgs() << '\n');
4386 ChangedFormulae = true;
4388 LU.DeleteFormula(F);
4394 // Now that we've filtered out some formulae, recompute the Regs set.
4396 LU.RecomputeRegs(LUIdx, RegUses);
4398 // Reset this to prepare for the next use.
4399 BestFormulae.clear();
4402 LLVM_DEBUG(if (ChangedFormulae) {
4404 "After filtering out undesirable candidates:\n";
4409 /// Estimate the worst-case number of solutions the solver might have to
4410 /// consider. It almost never considers this many solutions because it prune the
4411 /// search space, but the pruning isn't always sufficient.
4412 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
4414 for (const LSRUse &LU : Uses) {
4415 size_t FSize = LU.Formulae.size();
4416 if (FSize >= ComplexityLimit) {
4417 Power = ComplexityLimit;
4421 if (Power >= ComplexityLimit)
4427 /// When one formula uses a superset of the registers of another formula, it
4428 /// won't help reduce register pressure (though it may not necessarily hurt
4429 /// register pressure); remove it to simplify the system.
4430 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
4431 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4432 LLVM_DEBUG(dbgs() << "The search space is too complex.\n");
4434 LLVM_DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
4435 "which use a superset of registers used by other "
4438 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4439 LSRUse &LU = Uses[LUIdx];
4441 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4442 Formula &F = LU.Formulae[i];
4443 // Look for a formula with a constant or GV in a register. If the use
4444 // also has a formula with that same value in an immediate field,
4445 // delete the one that uses a register.
4446 for (SmallVectorImpl<const SCEV *>::const_iterator
4447 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
4448 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
4450 //FIXME: Formulas should store bitwidth to do wrapping properly.
4452 NewF.BaseOffset += (uint64_t)C->getValue()->getSExtValue();
4453 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
4454 (I - F.BaseRegs.begin()));
4455 if (LU.HasFormulaWithSameRegs(NewF)) {
4456 LLVM_DEBUG(dbgs() << " Deleting "; F.print(dbgs());
4458 LU.DeleteFormula(F);
4464 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
4465 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
4469 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
4470 (I - F.BaseRegs.begin()));
4471 if (LU.HasFormulaWithSameRegs(NewF)) {
4472 LLVM_DEBUG(dbgs() << " Deleting "; F.print(dbgs());
4474 LU.DeleteFormula(F);
4485 LU.RecomputeRegs(LUIdx, RegUses);
4488 LLVM_DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
4492 /// When there are many registers for expressions like A, A+1, A+2, etc.,
4493 /// allocate a single register for them.
4494 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
4495 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4499 dbgs() << "The search space is too complex.\n"
4500 "Narrowing the search space by assuming that uses separated "
4501 "by a constant offset will use the same registers.\n");
4503 // This is especially useful for unrolled loops.
4505 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4506 LSRUse &LU = Uses[LUIdx];
4507 for (const Formula &F : LU.Formulae) {
4508 if (F.BaseOffset == 0 || (F.Scale != 0 && F.Scale != 1))
4511 LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU);
4515 if (!reconcileNewOffset(*LUThatHas, F.BaseOffset, /*HasBaseReg=*/ false,
4516 LU.Kind, LU.AccessTy))
4519 LLVM_DEBUG(dbgs() << " Deleting use "; LU.print(dbgs()); dbgs() << '\n');
4521 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
4523 // Transfer the fixups of LU to LUThatHas.
4524 for (LSRFixup &Fixup : LU.Fixups) {
4525 Fixup.Offset += F.BaseOffset;
4526 LUThatHas->pushFixup(Fixup);
4527 LLVM_DEBUG(dbgs() << "New fixup has offset " << Fixup.Offset << '\n');
4530 // Delete formulae from the new use which are no longer legal.
4532 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
4533 Formula &F = LUThatHas->Formulae[i];
4534 if (!isLegalUse(TTI, LUThatHas->MinOffset, LUThatHas->MaxOffset,
4535 LUThatHas->Kind, LUThatHas->AccessTy, F)) {
4536 LLVM_DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
4537 LUThatHas->DeleteFormula(F);
4545 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
4547 // Delete the old use.
4548 DeleteUse(LU, LUIdx);
4555 LLVM_DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
4558 /// Call FilterOutUndesirableDedicatedRegisters again, if necessary, now that
4559 /// we've done more filtering, as it may be able to find more formulae to
4561 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
4562 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4563 LLVM_DEBUG(dbgs() << "The search space is too complex.\n");
4565 LLVM_DEBUG(dbgs() << "Narrowing the search space by re-filtering out "
4566 "undesirable dedicated registers.\n");
4568 FilterOutUndesirableDedicatedRegisters();
4570 LLVM_DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
4574 /// If a LSRUse has multiple formulae with the same ScaledReg and Scale.
4575 /// Pick the best one and delete the others.
4576 /// This narrowing heuristic is to keep as many formulae with different
4577 /// Scale and ScaledReg pair as possible while narrowing the search space.
4578 /// The benefit is that it is more likely to find out a better solution
4579 /// from a formulae set with more Scale and ScaledReg variations than
4580 /// a formulae set with the same Scale and ScaledReg. The picking winner
4581 /// reg heuristic will often keep the formulae with the same Scale and
4582 /// ScaledReg and filter others, and we want to avoid that if possible.
4583 void LSRInstance::NarrowSearchSpaceByFilterFormulaWithSameScaledReg() {
4584 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4588 dbgs() << "The search space is too complex.\n"
4589 "Narrowing the search space by choosing the best Formula "
4590 "from the Formulae with the same Scale and ScaledReg.\n");
4592 // Map the "Scale * ScaledReg" pair to the best formula of current LSRUse.
4593 using BestFormulaeTy = DenseMap<std::pair<const SCEV *, int64_t>, size_t>;
4595 BestFormulaeTy BestFormulae;
4597 bool ChangedFormulae = false;
4599 DenseSet<const SCEV *> VisitedRegs;
4600 SmallPtrSet<const SCEV *, 16> Regs;
4602 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4603 LSRUse &LU = Uses[LUIdx];
4604 LLVM_DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs());
4607 // Return true if Formula FA is better than Formula FB.
4608 auto IsBetterThan = [&](Formula &FA, Formula &FB) {
4609 // First we will try to choose the Formula with fewer new registers.
4610 // For a register used by current Formula, the more the register is
4611 // shared among LSRUses, the less we increase the register number
4612 // counter of the formula.
4613 size_t FARegNum = 0;
4614 for (const SCEV *Reg : FA.BaseRegs) {
4615 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(Reg);
4616 FARegNum += (NumUses - UsedByIndices.count() + 1);
4618 size_t FBRegNum = 0;
4619 for (const SCEV *Reg : FB.BaseRegs) {
4620 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(Reg);
4621 FBRegNum += (NumUses - UsedByIndices.count() + 1);
4623 if (FARegNum != FBRegNum)
4624 return FARegNum < FBRegNum;
4626 // If the new register numbers are the same, choose the Formula with
4628 Cost CostFA(L, SE, TTI);
4629 Cost CostFB(L, SE, TTI);
4631 CostFA.RateFormula(FA, Regs, VisitedRegs, LU);
4633 CostFB.RateFormula(FB, Regs, VisitedRegs, LU);
4634 return CostFA.isLess(CostFB);
4638 for (size_t FIdx = 0, NumForms = LU.Formulae.size(); FIdx != NumForms;
4640 Formula &F = LU.Formulae[FIdx];
4643 auto P = BestFormulae.insert({{F.ScaledReg, F.Scale}, FIdx});
4647 Formula &Best = LU.Formulae[P.first->second];
4648 if (IsBetterThan(F, Best))
4650 LLVM_DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
4652 " in favor of formula ";
4653 Best.print(dbgs()); dbgs() << '\n');
4655 ChangedFormulae = true;
4657 LU.DeleteFormula(F);
4663 LU.RecomputeRegs(LUIdx, RegUses);
4665 // Reset this to prepare for the next use.
4666 BestFormulae.clear();
4669 LLVM_DEBUG(if (ChangedFormulae) {
4671 "After filtering out undesirable candidates:\n";
4676 /// The function delete formulas with high registers number expectation.
4677 /// Assuming we don't know the value of each formula (already delete
4678 /// all inefficient), generate probability of not selecting for each
4682 /// reg(a) + reg({0,+,1})
4683 /// reg(a) + reg({-1,+,1}) + 1
4686 /// reg(b) + reg({0,+,1})
4687 /// reg(b) + reg({-1,+,1}) + 1
4690 /// reg(c) + reg(b) + reg({0,+,1})
4691 /// reg(c) + reg({b,+,1})
4693 /// Probability of not selecting
4695 /// reg(a) (1/3) * 1 * 1
4696 /// reg(b) 1 * (1/3) * (1/2)
4697 /// reg({0,+,1}) (2/3) * (2/3) * (1/2)
4698 /// reg({-1,+,1}) (2/3) * (2/3) * 1
4699 /// reg({a,+,1}) (2/3) * 1 * 1
4700 /// reg({b,+,1}) 1 * (2/3) * (2/3)
4701 /// reg(c) 1 * 1 * 0
4703 /// Now count registers number mathematical expectation for each formula:
4704 /// Note that for each use we exclude probability if not selecting for the use.
4705 /// For example for Use1 probability for reg(a) would be just 1 * 1 (excluding
4706 /// probabilty 1/3 of not selecting for Use1).
4708 /// reg(a) + reg({0,+,1}) 1 + 1/3 -- to be deleted
4709 /// reg(a) + reg({-1,+,1}) + 1 1 + 4/9 -- to be deleted
4712 /// reg(b) + reg({0,+,1}) 1/2 + 1/3 -- to be deleted
4713 /// reg(b) + reg({-1,+,1}) + 1 1/2 + 2/3 -- to be deleted
4714 /// reg({b,+,1}) 2/3
4716 /// reg(c) + reg(b) + reg({0,+,1}) 1 + 1/3 + 4/9 -- to be deleted
4717 /// reg(c) + reg({b,+,1}) 1 + 2/3
4718 void LSRInstance::NarrowSearchSpaceByDeletingCostlyFormulas() {
4719 if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4721 // Ok, we have too many of formulae on our hands to conveniently handle.
4722 // Use a rough heuristic to thin out the list.
4724 // Set of Regs wich will be 100% used in final solution.
4725 // Used in each formula of a solution (in example above this is reg(c)).
4726 // We can skip them in calculations.
4727 SmallPtrSet<const SCEV *, 4> UniqRegs;
4728 LLVM_DEBUG(dbgs() << "The search space is too complex.\n");
4730 // Map each register to probability of not selecting
4731 DenseMap <const SCEV *, float> RegNumMap;
4732 for (const SCEV *Reg : RegUses) {
4733 if (UniqRegs.count(Reg))
4736 for (const LSRUse &LU : Uses) {
4737 if (!LU.Regs.count(Reg))
4739 float P = LU.getNotSelectedProbability(Reg);
4743 UniqRegs.insert(Reg);
4745 RegNumMap.insert(std::make_pair(Reg, PNotSel));
4749 dbgs() << "Narrowing the search space by deleting costly formulas\n");
4751 // Delete formulas where registers number expectation is high.
4752 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4753 LSRUse &LU = Uses[LUIdx];
4754 // If nothing to delete - continue.
4755 if (LU.Formulae.size() < 2)
4757 // This is temporary solution to test performance. Float should be
4758 // replaced with round independent type (based on integers) to avoid
4759 // different results for different target builds.
4760 float FMinRegNum = LU.Formulae[0].getNumRegs();
4761 float FMinARegNum = LU.Formulae[0].getNumRegs();
4763 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4764 Formula &F = LU.Formulae[i];
4767 for (const SCEV *BaseReg : F.BaseRegs) {
4768 if (UniqRegs.count(BaseReg))
4770 FRegNum += RegNumMap[BaseReg] / LU.getNotSelectedProbability(BaseReg);
4771 if (isa<SCEVAddRecExpr>(BaseReg))
4773 RegNumMap[BaseReg] / LU.getNotSelectedProbability(BaseReg);
4775 if (const SCEV *ScaledReg = F.ScaledReg) {
4776 if (!UniqRegs.count(ScaledReg)) {
4778 RegNumMap[ScaledReg] / LU.getNotSelectedProbability(ScaledReg);
4779 if (isa<SCEVAddRecExpr>(ScaledReg))
4781 RegNumMap[ScaledReg] / LU.getNotSelectedProbability(ScaledReg);
4784 if (FMinRegNum > FRegNum ||
4785 (FMinRegNum == FRegNum && FMinARegNum > FARegNum)) {
4786 FMinRegNum = FRegNum;
4787 FMinARegNum = FARegNum;
4791 LLVM_DEBUG(dbgs() << " The formula "; LU.Formulae[MinIdx].print(dbgs());
4792 dbgs() << " with min reg num " << FMinRegNum << '\n');
4794 std::swap(LU.Formulae[MinIdx], LU.Formulae[0]);
4795 while (LU.Formulae.size() != 1) {
4796 LLVM_DEBUG(dbgs() << " Deleting "; LU.Formulae.back().print(dbgs());
4798 LU.Formulae.pop_back();
4800 LU.RecomputeRegs(LUIdx, RegUses);
4801 assert(LU.Formulae.size() == 1 && "Should be exactly 1 min regs formula");
4802 Formula &F = LU.Formulae[0];
4803 LLVM_DEBUG(dbgs() << " Leaving only "; F.print(dbgs()); dbgs() << '\n');
4804 // When we choose the formula, the regs become unique.
4805 UniqRegs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
4807 UniqRegs.insert(F.ScaledReg);
4809 LLVM_DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
4812 /// Pick a register which seems likely to be profitable, and then in any use
4813 /// which has any reference to that register, delete all formulae which do not
4814 /// reference that register.
4815 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
4816 // With all other options exhausted, loop until the system is simple
4817 // enough to handle.
4818 SmallPtrSet<const SCEV *, 4> Taken;
4819 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4820 // Ok, we have too many of formulae on our hands to conveniently handle.
4821 // Use a rough heuristic to thin out the list.
4822 LLVM_DEBUG(dbgs() << "The search space is too complex.\n");
4824 // Pick the register which is used by the most LSRUses, which is likely
4825 // to be a good reuse register candidate.
4826 const SCEV *Best = nullptr;
4827 unsigned BestNum = 0;
4828 for (const SCEV *Reg : RegUses) {
4829 if (Taken.count(Reg))
4833 BestNum = RegUses.getUsedByIndices(Reg).count();
4835 unsigned Count = RegUses.getUsedByIndices(Reg).count();
4836 if (Count > BestNum) {
4842 assert(Best && "Failed to find best LSRUse candidate");
4844 LLVM_DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
4845 << " will yield profitable reuse.\n");
4848 // In any use with formulae which references this register, delete formulae
4849 // which don't reference it.
4850 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4851 LSRUse &LU = Uses[LUIdx];
4852 if (!LU.Regs.count(Best)) continue;
4855 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4856 Formula &F = LU.Formulae[i];
4857 if (!F.referencesReg(Best)) {
4858 LLVM_DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
4859 LU.DeleteFormula(F);
4863 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
4869 LU.RecomputeRegs(LUIdx, RegUses);
4872 LLVM_DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
4876 /// If there are an extraordinary number of formulae to choose from, use some
4877 /// rough heuristics to prune down the number of formulae. This keeps the main
4878 /// solver from taking an extraordinary amount of time in some worst-case
4880 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
4881 NarrowSearchSpaceByDetectingSupersets();
4882 NarrowSearchSpaceByCollapsingUnrolledCode();
4883 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
4884 if (FilterSameScaledReg)
4885 NarrowSearchSpaceByFilterFormulaWithSameScaledReg();
4887 NarrowSearchSpaceByDeletingCostlyFormulas();
4889 NarrowSearchSpaceByPickingWinnerRegs();
4892 /// This is the recursive solver.
4893 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
4895 SmallVectorImpl<const Formula *> &Workspace,
4896 const Cost &CurCost,
4897 const SmallPtrSet<const SCEV *, 16> &CurRegs,
4898 DenseSet<const SCEV *> &VisitedRegs) const {
4901 // - use more aggressive filtering
4902 // - sort the formula so that the most profitable solutions are found first
4903 // - sort the uses too
4905 // - don't compute a cost, and then compare. compare while computing a cost
4907 // - track register sets with SmallBitVector
4909 const LSRUse &LU = Uses[Workspace.size()];
4911 // If this use references any register that's already a part of the
4912 // in-progress solution, consider it a requirement that a formula must
4913 // reference that register in order to be considered. This prunes out
4914 // unprofitable searching.
4915 SmallSetVector<const SCEV *, 4> ReqRegs;
4916 for (const SCEV *S : CurRegs)
4917 if (LU.Regs.count(S))
4920 SmallPtrSet<const SCEV *, 16> NewRegs;
4921 Cost NewCost(L, SE, TTI);
4922 for (const Formula &F : LU.Formulae) {
4923 // Ignore formulae which may not be ideal in terms of register reuse of
4924 // ReqRegs. The formula should use all required registers before
4925 // introducing new ones.
4926 int NumReqRegsToFind = std::min(F.getNumRegs(), ReqRegs.size());
4927 for (const SCEV *Reg : ReqRegs) {
4928 if ((F.ScaledReg && F.ScaledReg == Reg) ||
4929 is_contained(F.BaseRegs, Reg)) {
4931 if (NumReqRegsToFind == 0)
4935 if (NumReqRegsToFind != 0) {
4936 // If none of the formulae satisfied the required registers, then we could
4937 // clear ReqRegs and try again. Currently, we simply give up in this case.
4941 // Evaluate the cost of the current formula. If it's already worse than
4942 // the current best, prune the search at that point.
4945 NewCost.RateFormula(F, NewRegs, VisitedRegs, LU);
4946 if (NewCost.isLess(SolutionCost)) {
4947 Workspace.push_back(&F);
4948 if (Workspace.size() != Uses.size()) {
4949 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
4950 NewRegs, VisitedRegs);
4951 if (F.getNumRegs() == 1 && Workspace.size() == 1)
4952 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
4954 LLVM_DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
4955 dbgs() << ".\nRegs:\n";
4956 for (const SCEV *S : NewRegs) dbgs()
4957 << "- " << *S << "\n";
4960 SolutionCost = NewCost;
4961 Solution = Workspace;
4963 Workspace.pop_back();
4968 /// Choose one formula from each use. Return the results in the given Solution
4970 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
4971 SmallVector<const Formula *, 8> Workspace;
4972 Cost SolutionCost(L, SE, TTI);
4973 SolutionCost.Lose();
4974 Cost CurCost(L, SE, TTI);
4975 SmallPtrSet<const SCEV *, 16> CurRegs;
4976 DenseSet<const SCEV *> VisitedRegs;
4977 Workspace.reserve(Uses.size());
4979 // SolveRecurse does all the work.
4980 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
4981 CurRegs, VisitedRegs);
4982 if (Solution.empty()) {
4983 LLVM_DEBUG(dbgs() << "\nNo Satisfactory Solution\n");
4987 // Ok, we've now made all our decisions.
4988 LLVM_DEBUG(dbgs() << "\n"
4989 "The chosen solution requires ";
4990 SolutionCost.print(dbgs()); dbgs() << ":\n";
4991 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
4993 Uses[i].print(dbgs());
4996 Solution[i]->print(dbgs());
5000 assert(Solution.size() == Uses.size() && "Malformed solution!");
5003 /// Helper for AdjustInsertPositionForExpand. Climb up the dominator tree far as
5004 /// we can go while still being dominated by the input positions. This helps
5005 /// canonicalize the insert position, which encourages sharing.
5006 BasicBlock::iterator
5007 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
5008 const SmallVectorImpl<Instruction *> &Inputs)
5010 Instruction *Tentative = &*IP;
5012 bool AllDominate = true;
5013 Instruction *BetterPos = nullptr;
5014 // Don't bother attempting to insert before a catchswitch, their basic block
5015 // cannot have other non-PHI instructions.
5016 if (isa<CatchSwitchInst>(Tentative))
5019 for (Instruction *Inst : Inputs) {
5020 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
5021 AllDominate = false;
5024 // Attempt to find an insert position in the middle of the block,
5025 // instead of at the end, so that it can be used for other expansions.
5026 if (Tentative->getParent() == Inst->getParent() &&
5027 (!BetterPos || !DT.dominates(Inst, BetterPos)))
5028 BetterPos = &*std::next(BasicBlock::iterator(Inst));
5033 IP = BetterPos->getIterator();
5035 IP = Tentative->getIterator();
5037 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
5038 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
5041 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
5042 if (!Rung) return IP;
5043 Rung = Rung->getIDom();
5044 if (!Rung) return IP;
5045 IDom = Rung->getBlock();
5047 // Don't climb into a loop though.
5048 const Loop *IDomLoop = LI.getLoopFor(IDom);
5049 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
5050 if (IDomDepth <= IPLoopDepth &&
5051 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
5055 Tentative = IDom->getTerminator();
5061 /// Determine an input position which will be dominated by the operands and
5062 /// which will dominate the result.
5063 BasicBlock::iterator
5064 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator LowestIP,
5067 SCEVExpander &Rewriter) const {
5068 // Collect some instructions which must be dominated by the
5069 // expanding replacement. These must be dominated by any operands that
5070 // will be required in the expansion.
5071 SmallVector<Instruction *, 4> Inputs;
5072 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
5073 Inputs.push_back(I);
5074 if (LU.Kind == LSRUse::ICmpZero)
5075 if (Instruction *I =
5076 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
5077 Inputs.push_back(I);
5078 if (LF.PostIncLoops.count(L)) {
5079 if (LF.isUseFullyOutsideLoop(L))
5080 Inputs.push_back(L->getLoopLatch()->getTerminator());
5082 Inputs.push_back(IVIncInsertPos);
5084 // The expansion must also be dominated by the increment positions of any
5085 // loops it for which it is using post-inc mode.
5086 for (const Loop *PIL : LF.PostIncLoops) {
5087 if (PIL == L) continue;
5089 // Be dominated by the loop exit.
5090 SmallVector<BasicBlock *, 4> ExitingBlocks;
5091 PIL->getExitingBlocks(ExitingBlocks);
5092 if (!ExitingBlocks.empty()) {
5093 BasicBlock *BB = ExitingBlocks[0];
5094 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
5095 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
5096 Inputs.push_back(BB->getTerminator());
5100 assert(!isa<PHINode>(LowestIP) && !LowestIP->isEHPad()
5101 && !isa<DbgInfoIntrinsic>(LowestIP) &&
5102 "Insertion point must be a normal instruction");
5104 // Then, climb up the immediate dominator tree as far as we can go while
5105 // still being dominated by the input positions.
5106 BasicBlock::iterator IP = HoistInsertPosition(LowestIP, Inputs);
5108 // Don't insert instructions before PHI nodes.
5109 while (isa<PHINode>(IP)) ++IP;
5111 // Ignore landingpad instructions.
5112 while (IP->isEHPad()) ++IP;
5114 // Ignore debug intrinsics.
5115 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
5117 // Set IP below instructions recently inserted by SCEVExpander. This keeps the
5118 // IP consistent across expansions and allows the previously inserted
5119 // instructions to be reused by subsequent expansion.
5120 while (Rewriter.isInsertedInstruction(&*IP) && IP != LowestIP)
5126 /// Emit instructions for the leading candidate expression for this LSRUse (this
5127 /// is called "expanding").
5128 Value *LSRInstance::Expand(const LSRUse &LU, const LSRFixup &LF,
5129 const Formula &F, BasicBlock::iterator IP,
5130 SCEVExpander &Rewriter,
5131 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
5132 if (LU.RigidFormula)
5133 return LF.OperandValToReplace;
5135 // Determine an input position which will be dominated by the operands and
5136 // which will dominate the result.
5137 IP = AdjustInsertPositionForExpand(IP, LF, LU, Rewriter);
5138 Rewriter.setInsertPoint(&*IP);
5140 // Inform the Rewriter if we have a post-increment use, so that it can
5141 // perform an advantageous expansion.
5142 Rewriter.setPostInc(LF.PostIncLoops);
5144 // This is the type that the user actually needs.
5145 Type *OpTy = LF.OperandValToReplace->getType();
5146 // This will be the type that we'll initially expand to.
5147 Type *Ty = F.getType();
5149 // No type known; just expand directly to the ultimate type.
5151 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
5152 // Expand directly to the ultimate type if it's the right size.
5154 // This is the type to do integer arithmetic in.
5155 Type *IntTy = SE.getEffectiveSCEVType(Ty);
5157 // Build up a list of operands to add together to form the full base.
5158 SmallVector<const SCEV *, 8> Ops;
5160 // Expand the BaseRegs portion.
5161 for (const SCEV *Reg : F.BaseRegs) {
5162 assert(!Reg->isZero() && "Zero allocated in a base register!");
5164 // If we're expanding for a post-inc user, make the post-inc adjustment.
5165 Reg = denormalizeForPostIncUse(Reg, LF.PostIncLoops, SE);
5166 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, nullptr)));
5169 // Expand the ScaledReg portion.
5170 Value *ICmpScaledV = nullptr;
5172 const SCEV *ScaledS = F.ScaledReg;
5174 // If we're expanding for a post-inc user, make the post-inc adjustment.
5175 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
5176 ScaledS = denormalizeForPostIncUse(ScaledS, Loops, SE);
5178 if (LU.Kind == LSRUse::ICmpZero) {
5179 // Expand ScaleReg as if it was part of the base regs.
5182 SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr)));
5184 // An interesting way of "folding" with an icmp is to use a negated
5185 // scale, which we'll implement by inserting it into the other operand
5187 assert(F.Scale == -1 &&
5188 "The only scale supported by ICmpZero uses is -1!");
5189 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, nullptr);
5192 // Otherwise just expand the scaled register and an explicit scale,
5193 // which is expected to be matched as part of the address.
5195 // Flush the operand list to suppress SCEVExpander hoisting address modes.
5196 // Unless the addressing mode will not be folded.
5197 if (!Ops.empty() && LU.Kind == LSRUse::Address &&
5198 isAMCompletelyFolded(TTI, LU, F)) {
5199 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), nullptr);
5201 Ops.push_back(SE.getUnknown(FullV));
5203 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr));
5206 SE.getMulExpr(ScaledS, SE.getConstant(ScaledS->getType(), F.Scale));
5207 Ops.push_back(ScaledS);
5211 // Expand the GV portion.
5213 // Flush the operand list to suppress SCEVExpander hoisting.
5215 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty);
5217 Ops.push_back(SE.getUnknown(FullV));
5219 Ops.push_back(SE.getUnknown(F.BaseGV));
5222 // Flush the operand list to suppress SCEVExpander hoisting of both folded and
5223 // unfolded offsets. LSR assumes they both live next to their uses.
5225 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty);
5227 Ops.push_back(SE.getUnknown(FullV));
5230 // Expand the immediate portion.
5231 int64_t Offset = (uint64_t)F.BaseOffset + LF.Offset;
5233 if (LU.Kind == LSRUse::ICmpZero) {
5234 // The other interesting way of "folding" with an ICmpZero is to use a
5235 // negated immediate.
5237 ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset);
5239 Ops.push_back(SE.getUnknown(ICmpScaledV));
5240 ICmpScaledV = ConstantInt::get(IntTy, Offset);
5243 // Just add the immediate values. These again are expected to be matched
5244 // as part of the address.
5245 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
5249 // Expand the unfolded offset portion.
5250 int64_t UnfoldedOffset = F.UnfoldedOffset;
5251 if (UnfoldedOffset != 0) {
5252 // Just add the immediate values.
5253 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy,
5257 // Emit instructions summing all the operands.
5258 const SCEV *FullS = Ops.empty() ?
5259 SE.getConstant(IntTy, 0) :
5261 Value *FullV = Rewriter.expandCodeFor(FullS, Ty);
5263 // We're done expanding now, so reset the rewriter.
5264 Rewriter.clearPostInc();
5266 // An ICmpZero Formula represents an ICmp which we're handling as a
5267 // comparison against zero. Now that we've expanded an expression for that
5268 // form, update the ICmp's other operand.
5269 if (LU.Kind == LSRUse::ICmpZero) {
5270 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
5271 DeadInsts.emplace_back(CI->getOperand(1));
5272 assert(!F.BaseGV && "ICmp does not support folding a global value and "
5273 "a scale at the same time!");
5274 if (F.Scale == -1) {
5275 if (ICmpScaledV->getType() != OpTy) {
5277 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
5279 ICmpScaledV, OpTy, "tmp", CI);
5282 CI->setOperand(1, ICmpScaledV);
5284 // A scale of 1 means that the scale has been expanded as part of the
5286 assert((F.Scale == 0 || F.Scale == 1) &&
5287 "ICmp does not support folding a global value and "
5288 "a scale at the same time!");
5289 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
5291 if (C->getType() != OpTy)
5292 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
5296 CI->setOperand(1, C);
5303 /// Helper for Rewrite. PHI nodes are special because the use of their operands
5304 /// effectively happens in their predecessor blocks, so the expression may need
5305 /// to be expanded in multiple places.
5306 void LSRInstance::RewriteForPHI(
5307 PHINode *PN, const LSRUse &LU, const LSRFixup &LF, const Formula &F,
5308 SCEVExpander &Rewriter, SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
5309 DenseMap<BasicBlock *, Value *> Inserted;
5310 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
5311 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
5312 bool needUpdateFixups = false;
5313 BasicBlock *BB = PN->getIncomingBlock(i);
5315 // If this is a critical edge, split the edge so that we do not insert
5316 // the code on all predecessor/successor paths. We do this unless this
5317 // is the canonical backedge for this loop, which complicates post-inc
5319 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
5320 !isa<IndirectBrInst>(BB->getTerminator()) &&
5321 !isa<CatchSwitchInst>(BB->getTerminator())) {
5322 BasicBlock *Parent = PN->getParent();
5323 Loop *PNLoop = LI.getLoopFor(Parent);
5324 if (!PNLoop || Parent != PNLoop->getHeader()) {
5325 // Split the critical edge.
5326 BasicBlock *NewBB = nullptr;
5327 if (!Parent->isLandingPad()) {
5328 NewBB = SplitCriticalEdge(BB, Parent,
5329 CriticalEdgeSplittingOptions(&DT, &LI)
5330 .setMergeIdenticalEdges()
5331 .setKeepOneInputPHIs());
5333 SmallVector<BasicBlock*, 2> NewBBs;
5334 SplitLandingPadPredecessors(Parent, BB, "", "", NewBBs, &DT, &LI);
5337 // If NewBB==NULL, then SplitCriticalEdge refused to split because all
5338 // phi predecessors are identical. The simple thing to do is skip
5339 // splitting in this case rather than complicate the API.
5341 // If PN is outside of the loop and BB is in the loop, we want to
5342 // move the block to be immediately before the PHI block, not
5343 // immediately after BB.
5344 if (L->contains(BB) && !L->contains(PN))
5345 NewBB->moveBefore(PN->getParent());
5347 // Splitting the edge can reduce the number of PHI entries we have.
5348 e = PN->getNumIncomingValues();
5350 i = PN->getBasicBlockIndex(BB);
5352 needUpdateFixups = true;
5357 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
5358 Inserted.insert(std::make_pair(BB, static_cast<Value *>(nullptr)));
5360 PN->setIncomingValue(i, Pair.first->second);
5362 Value *FullV = Expand(LU, LF, F, BB->getTerminator()->getIterator(),
5363 Rewriter, DeadInsts);
5365 // If this is reuse-by-noop-cast, insert the noop cast.
5366 Type *OpTy = LF.OperandValToReplace->getType();
5367 if (FullV->getType() != OpTy)
5369 CastInst::Create(CastInst::getCastOpcode(FullV, false,
5371 FullV, LF.OperandValToReplace->getType(),
5372 "tmp", BB->getTerminator());
5374 PN->setIncomingValue(i, FullV);
5375 Pair.first->second = FullV;
5378 // If LSR splits critical edge and phi node has other pending
5379 // fixup operands, we need to update those pending fixups. Otherwise
5380 // formulae will not be implemented completely and some instructions
5381 // will not be eliminated.
5382 if (needUpdateFixups) {
5383 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx)
5384 for (LSRFixup &Fixup : Uses[LUIdx].Fixups)
5385 // If fixup is supposed to rewrite some operand in the phi
5386 // that was just updated, it may be already moved to
5387 // another phi node. Such fixup requires update.
5388 if (Fixup.UserInst == PN) {
5389 // Check if the operand we try to replace still exists in the
5391 bool foundInOriginalPHI = false;
5392 for (const auto &val : PN->incoming_values())
5393 if (val == Fixup.OperandValToReplace) {
5394 foundInOriginalPHI = true;
5398 // If fixup operand found in original PHI - nothing to do.
5399 if (foundInOriginalPHI)
5402 // Otherwise it might be moved to another PHI and requires update.
5403 // If fixup operand not found in any of the incoming blocks that
5404 // means we have already rewritten it - nothing to do.
5405 for (const auto &Block : PN->blocks())
5406 for (BasicBlock::iterator I = Block->begin(); isa<PHINode>(I);
5408 PHINode *NewPN = cast<PHINode>(I);
5409 for (const auto &val : NewPN->incoming_values())
5410 if (val == Fixup.OperandValToReplace)
5411 Fixup.UserInst = NewPN;
5418 /// Emit instructions for the leading candidate expression for this LSRUse (this
5419 /// is called "expanding"), and update the UserInst to reference the newly
5421 void LSRInstance::Rewrite(const LSRUse &LU, const LSRFixup &LF,
5422 const Formula &F, SCEVExpander &Rewriter,
5423 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
5424 // First, find an insertion point that dominates UserInst. For PHI nodes,
5425 // find the nearest block which dominates all the relevant uses.
5426 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
5427 RewriteForPHI(PN, LU, LF, F, Rewriter, DeadInsts);
5430 Expand(LU, LF, F, LF.UserInst->getIterator(), Rewriter, DeadInsts);
5432 // If this is reuse-by-noop-cast, insert the noop cast.
5433 Type *OpTy = LF.OperandValToReplace->getType();
5434 if (FullV->getType() != OpTy) {
5436 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
5437 FullV, OpTy, "tmp", LF.UserInst);
5441 // Update the user. ICmpZero is handled specially here (for now) because
5442 // Expand may have updated one of the operands of the icmp already, and
5443 // its new value may happen to be equal to LF.OperandValToReplace, in
5444 // which case doing replaceUsesOfWith leads to replacing both operands
5445 // with the same value. TODO: Reorganize this.
5446 if (LU.Kind == LSRUse::ICmpZero)
5447 LF.UserInst->setOperand(0, FullV);
5449 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
5452 DeadInsts.emplace_back(LF.OperandValToReplace);
5455 /// Rewrite all the fixup locations with new values, following the chosen
5457 void LSRInstance::ImplementSolution(
5458 const SmallVectorImpl<const Formula *> &Solution) {
5459 // Keep track of instructions we may have made dead, so that
5460 // we can remove them after we are done working.
5461 SmallVector<WeakTrackingVH, 16> DeadInsts;
5463 SCEVExpander Rewriter(SE, L->getHeader()->getModule()->getDataLayout(),
5466 Rewriter.setDebugType(DEBUG_TYPE);
5468 Rewriter.disableCanonicalMode();
5469 Rewriter.enableLSRMode();
5470 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
5472 // Mark phi nodes that terminate chains so the expander tries to reuse them.
5473 for (const IVChain &Chain : IVChainVec) {
5474 if (PHINode *PN = dyn_cast<PHINode>(Chain.tailUserInst()))
5475 Rewriter.setChainedPhi(PN);
5478 // Expand the new value definitions and update the users.
5479 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx)
5480 for (const LSRFixup &Fixup : Uses[LUIdx].Fixups) {
5481 Rewrite(Uses[LUIdx], Fixup, *Solution[LUIdx], Rewriter, DeadInsts);
5485 for (const IVChain &Chain : IVChainVec) {
5486 GenerateIVChain(Chain, Rewriter, DeadInsts);
5489 // Clean up after ourselves. This must be done before deleting any
5493 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
5496 LSRInstance::LSRInstance(Loop *L, IVUsers &IU, ScalarEvolution &SE,
5497 DominatorTree &DT, LoopInfo &LI,
5498 const TargetTransformInfo &TTI, AssumptionCache &AC,
5499 TargetLibraryInfo &LibInfo)
5500 : IU(IU), SE(SE), DT(DT), LI(LI), AC(AC), LibInfo(LibInfo), TTI(TTI), L(L),
5501 FavorBackedgeIndex(EnableBackedgeIndexing &&
5502 TTI.shouldFavorBackedgeIndex(L)) {
5503 // If LoopSimplify form is not available, stay out of trouble.
5504 if (!L->isLoopSimplifyForm())
5507 // If there's no interesting work to be done, bail early.
5508 if (IU.empty()) return;
5510 // If there's too much analysis to be done, bail early. We won't be able to
5511 // model the problem anyway.
5512 unsigned NumUsers = 0;
5513 for (const IVStrideUse &U : IU) {
5514 if (++NumUsers > MaxIVUsers) {
5516 LLVM_DEBUG(dbgs() << "LSR skipping loop, too many IV Users in " << U
5520 // Bail out if we have a PHI on an EHPad that gets a value from a
5521 // CatchSwitchInst. Because the CatchSwitchInst cannot be split, there is
5522 // no good place to stick any instructions.
5523 if (auto *PN = dyn_cast<PHINode>(U.getUser())) {
5524 auto *FirstNonPHI = PN->getParent()->getFirstNonPHI();
5525 if (isa<FuncletPadInst>(FirstNonPHI) ||
5526 isa<CatchSwitchInst>(FirstNonPHI))
5527 for (BasicBlock *PredBB : PN->blocks())
5528 if (isa<CatchSwitchInst>(PredBB->getFirstNonPHI()))
5534 // All dominating loops must have preheaders, or SCEVExpander may not be able
5535 // to materialize an AddRecExpr whose Start is an outer AddRecExpr.
5537 // IVUsers analysis should only create users that are dominated by simple loop
5538 // headers. Since this loop should dominate all of its users, its user list
5539 // should be empty if this loop itself is not within a simple loop nest.
5540 for (DomTreeNode *Rung = DT.getNode(L->getLoopPreheader());
5541 Rung; Rung = Rung->getIDom()) {
5542 BasicBlock *BB = Rung->getBlock();
5543 const Loop *DomLoop = LI.getLoopFor(BB);
5544 if (DomLoop && DomLoop->getHeader() == BB) {
5545 assert(DomLoop->getLoopPreheader() && "LSR needs a simplified loop nest");
5550 LLVM_DEBUG(dbgs() << "\nLSR on loop ";
5551 L->getHeader()->printAsOperand(dbgs(), /*PrintType=*/false);
5554 // First, perform some low-level loop optimizations.
5556 OptimizeLoopTermCond();
5558 // If loop preparation eliminates all interesting IV users, bail.
5559 if (IU.empty()) return;
5561 // Skip nested loops until we can model them better with formulae.
5563 LLVM_DEBUG(dbgs() << "LSR skipping outer loop " << *L << "\n");
5567 // Start collecting data and preparing for the solver.
5569 CollectInterestingTypesAndFactors();
5570 CollectFixupsAndInitialFormulae();
5571 CollectLoopInvariantFixupsAndFormulae();
5576 LLVM_DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
5577 print_uses(dbgs()));
5579 // Now use the reuse data to generate a bunch of interesting ways
5580 // to formulate the values needed for the uses.
5581 GenerateAllReuseFormulae();
5583 FilterOutUndesirableDedicatedRegisters();
5584 NarrowSearchSpaceUsingHeuristics();
5586 SmallVector<const Formula *, 8> Solution;
5589 // Release memory that is no longer needed.
5594 if (Solution.empty())
5598 // Formulae should be legal.
5599 for (const LSRUse &LU : Uses) {
5600 for (const Formula &F : LU.Formulae)
5601 assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
5602 F) && "Illegal formula generated!");
5606 // Now that we've decided what we want, make it so.
5607 ImplementSolution(Solution);
5610 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
5611 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
5612 if (Factors.empty() && Types.empty()) return;
5614 OS << "LSR has identified the following interesting factors and types: ";
5617 for (int64_t Factor : Factors) {
5618 if (!First) OS << ", ";
5620 OS << '*' << Factor;
5623 for (Type *Ty : Types) {
5624 if (!First) OS << ", ";
5626 OS << '(' << *Ty << ')';
5631 void LSRInstance::print_fixups(raw_ostream &OS) const {
5632 OS << "LSR is examining the following fixup sites:\n";
5633 for (const LSRUse &LU : Uses)
5634 for (const LSRFixup &LF : LU.Fixups) {
5641 void LSRInstance::print_uses(raw_ostream &OS) const {
5642 OS << "LSR is examining the following uses:\n";
5643 for (const LSRUse &LU : Uses) {
5647 for (const Formula &F : LU.Formulae) {
5655 void LSRInstance::print(raw_ostream &OS) const {
5656 print_factors_and_types(OS);
5661 LLVM_DUMP_METHOD void LSRInstance::dump() const {
5662 print(errs()); errs() << '\n';
5668 class LoopStrengthReduce : public LoopPass {
5670 static char ID; // Pass ID, replacement for typeid
5672 LoopStrengthReduce();
5675 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
5676 void getAnalysisUsage(AnalysisUsage &AU) const override;
5679 } // end anonymous namespace
5681 LoopStrengthReduce::LoopStrengthReduce() : LoopPass(ID) {
5682 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry());
5685 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
5686 // We split critical edges, so we change the CFG. However, we do update
5687 // many analyses if they are around.
5688 AU.addPreservedID(LoopSimplifyID);
5690 AU.addRequired<LoopInfoWrapperPass>();
5691 AU.addPreserved<LoopInfoWrapperPass>();
5692 AU.addRequiredID(LoopSimplifyID);
5693 AU.addRequired<DominatorTreeWrapperPass>();
5694 AU.addPreserved<DominatorTreeWrapperPass>();
5695 AU.addRequired<ScalarEvolutionWrapperPass>();
5696 AU.addPreserved<ScalarEvolutionWrapperPass>();
5697 AU.addRequired<AssumptionCacheTracker>();
5698 AU.addRequired<TargetLibraryInfoWrapperPass>();
5699 // Requiring LoopSimplify a second time here prevents IVUsers from running
5700 // twice, since LoopSimplify was invalidated by running ScalarEvolution.
5701 AU.addRequiredID(LoopSimplifyID);
5702 AU.addRequired<IVUsersWrapperPass>();
5703 AU.addPreserved<IVUsersWrapperPass>();
5704 AU.addRequired<TargetTransformInfoWrapperPass>();
5707 static bool ReduceLoopStrength(Loop *L, IVUsers &IU, ScalarEvolution &SE,
5708 DominatorTree &DT, LoopInfo &LI,
5709 const TargetTransformInfo &TTI,
5710 AssumptionCache &AC,
5711 TargetLibraryInfo &LibInfo) {
5713 bool Changed = false;
5715 // Run the main LSR transformation.
5716 Changed |= LSRInstance(L, IU, SE, DT, LI, TTI, AC, LibInfo).getChanged();
5718 // Remove any extra phis created by processing inner loops.
5719 Changed |= DeleteDeadPHIs(L->getHeader());
5720 if (EnablePhiElim && L->isLoopSimplifyForm()) {
5721 SmallVector<WeakTrackingVH, 16> DeadInsts;
5722 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
5723 SCEVExpander Rewriter(SE, DL, "lsr");
5725 Rewriter.setDebugType(DEBUG_TYPE);
5727 unsigned numFolded = Rewriter.replaceCongruentIVs(L, &DT, DeadInsts, &TTI);
5730 DeleteTriviallyDeadInstructions(DeadInsts);
5731 DeleteDeadPHIs(L->getHeader());
5737 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
5741 auto &IU = getAnalysis<IVUsersWrapperPass>().getIU();
5742 auto &SE = getAnalysis<ScalarEvolutionWrapperPass>().getSE();
5743 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
5744 auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
5745 const auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
5746 *L->getHeader()->getParent());
5747 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(
5748 *L->getHeader()->getParent());
5749 auto &LibInfo = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(
5750 *L->getHeader()->getParent());
5751 return ReduceLoopStrength(L, IU, SE, DT, LI, TTI, AC, LibInfo);
5754 PreservedAnalyses LoopStrengthReducePass::run(Loop &L, LoopAnalysisManager &AM,
5755 LoopStandardAnalysisResults &AR,
5757 if (!ReduceLoopStrength(&L, AM.getResult<IVUsersAnalysis>(L, AR), AR.SE,
5758 AR.DT, AR.LI, AR.TTI, AR.AC, AR.TLI))
5759 return PreservedAnalyses::all();
5761 return getLoopPassPreservedAnalyses();
5764 char LoopStrengthReduce::ID = 0;
5766 INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce",
5767 "Loop Strength Reduction", false, false)
5768 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
5769 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
5770 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
5771 INITIALIZE_PASS_DEPENDENCY(IVUsersWrapperPass)
5772 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
5773 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5774 INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce",
5775 "Loop Strength Reduction", false, false)
5777 Pass *llvm::createLoopStrengthReducePass() { return new LoopStrengthReduce(); }