/build/source/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp

Bug Summary

File:	build/source/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp
Warning:	line 5598, column 22 Called C++ object pointer is null
Annotated Source Code

Press '?' to see keyboard shortcuts
Show analyzer invocation
clang -cc1 -cc1 -triple x86_64-pc-linux-gnu -analyze -disable-free -clear-ast-before-backend -disable-llvm-verifier -discard-value-names -main-file-name LoopStrengthReduce.cpp -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=cplusplus -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -setup-static-analyzer -analyzer-config-compatibility-mode=true -mrelocation-model pic -pic-level 2 -mframe-pointer=none -fmath-errno -ffp-contract=on -fno-rounding-math -mconstructor-aliases -funwind-tables=2 -target-cpu x86-64 -tune-cpu generic -debugger-tuning=gdb -ffunction-sections -fdata-sections -fcoverage-compilation-dir=/build/source/build-llvm -resource-dir /usr/lib/llvm-17/lib/clang/17 -D _DEBUG -D _GLIBCXX_ASSERTIONS -D _GNU_SOURCE -D _LIBCPP_ENABLE_ASSERTIONS -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I lib/Transforms/Scalar -I /build/source/llvm/lib/Transforms/Scalar -I include -I /build/source/llvm/include -D _FORTIFY_SOURCE=2 -D NDEBUG -U NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/c++/10 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/x86_64-linux-gnu/c++/10 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/c++/10/backward -internal-isystem /usr/lib/llvm-17/lib/clang/17/include -internal-isystem /usr/local/include -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../x86_64-linux-gnu/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -fmacro-prefix-map=/build/source/build-llvm=build-llvm -fmacro-prefix-map=/build/source/= -fcoverage-prefix-map=/build/source/build-llvm=build-llvm -fcoverage-prefix-map=/build/source/= -source-date-epoch 1680607027 -O3 -Wno-unused-command-line-argument -Wno-unused-parameter -Wwrite-strings -Wno-missing-field-initializers -Wno-long-long -Wno-maybe-uninitialized -Wno-class-memaccess -Wno-redundant-move -Wno-pessimizing-move -Wno-noexcept-type -Wno-comment -Wno-misleading-indentation -std=c++17 -fdeprecated-macro -fdebug-compilation-dir=/build/source/build-llvm -fdebug-prefix-map=/build/source/build-llvm=build-llvm -fdebug-prefix-map=/build/source/= -fdebug-prefix-map=/build/source/build-llvm=build-llvm -fdebug-prefix-map=/build/source/= -ferror-limit 19 -fvisibility-inlines-hidden -stack-protector 2 -fgnuc-version=4.2.1 -fcolor-diagnostics -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -faddrsig -D__GCC_HAVE_DWARF2_CFI_ASM=1 -o /tmp/scan-build-2023-04-04-200003-16331-1 -x c++ /build/source/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp
1//===- LoopStrengthReduce.cpp - Strength Reduce IVs in Loops --------------===//
2//
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
6//
7//===----------------------------------------------------------------------===//
8//
9// This transformation analyzes and transforms the induction variables (and
10// computations derived from them) into forms suitable for efficient execution
11// on the target.
12//
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.
18//
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:
22//
23//   %i = phi [ 0, %entry ], [ %i.next, %latch ]
24//   ...
25//   %i.next = add %i, 1
26//   %c = icmp eq %i.next, %n
27//
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.
34//
35// TODO: More sophistication in the way Formulae are generated and filtered.
36//
37// TODO: Handle multiple loops at a time.
38//
39// TODO: Should the addressing mode BaseGV be changed to a ConstantExpr instead
40//       of a GlobalValue?
41//
42// TODO: When truncation is free, truncate ICmp users' operands to make it a
43//       smaller encoding (on x86 at least).
44//
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"
51//       use kind.
52//
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/Statistic.h"
68#include "llvm/ADT/iterator_range.h"
69#include "llvm/Analysis/AssumptionCache.h"
70#include "llvm/Analysis/IVUsers.h"
71#include "llvm/Analysis/LoopAnalysisManager.h"
72#include "llvm/Analysis/LoopInfo.h"
73#include "llvm/Analysis/LoopPass.h"
74#include "llvm/Analysis/MemorySSA.h"
75#include "llvm/Analysis/MemorySSAUpdater.h"
76#include "llvm/Analysis/ScalarEvolution.h"
77#include "llvm/Analysis/ScalarEvolutionExpressions.h"
78#include "llvm/Analysis/ScalarEvolutionNormalization.h"
79#include "llvm/Analysis/TargetLibraryInfo.h"
80#include "llvm/Analysis/TargetTransformInfo.h"
81#include "llvm/Analysis/ValueTracking.h"
82#include "llvm/BinaryFormat/Dwarf.h"
83#include "llvm/Config/llvm-config.h"
84#include "llvm/IR/BasicBlock.h"
85#include "llvm/IR/Constant.h"
86#include "llvm/IR/Constants.h"
87#include "llvm/IR/DebugInfoMetadata.h"
88#include "llvm/IR/DerivedTypes.h"
89#include "llvm/IR/Dominators.h"
90#include "llvm/IR/GlobalValue.h"
91#include "llvm/IR/IRBuilder.h"
92#include "llvm/IR/InstrTypes.h"
93#include "llvm/IR/Instruction.h"
94#include "llvm/IR/Instructions.h"
95#include "llvm/IR/IntrinsicInst.h"
96#include "llvm/IR/Module.h"
97#include "llvm/IR/Operator.h"
98#include "llvm/IR/PassManager.h"
99#include "llvm/IR/Type.h"
100#include "llvm/IR/Use.h"
101#include "llvm/IR/User.h"
102#include "llvm/IR/Value.h"
103#include "llvm/IR/ValueHandle.h"
104#include "llvm/InitializePasses.h"
105#include "llvm/Pass.h"
106#include "llvm/Support/Casting.h"
107#include "llvm/Support/CommandLine.h"
108#include "llvm/Support/Compiler.h"
109#include "llvm/Support/Debug.h"
110#include "llvm/Support/ErrorHandling.h"
111#include "llvm/Support/MathExtras.h"
112#include "llvm/Support/raw_ostream.h"
113#include "llvm/Transforms/Scalar.h"
114#include "llvm/Transforms/Utils.h"
115#include "llvm/Transforms/Utils/BasicBlockUtils.h"
116#include "llvm/Transforms/Utils/Local.h"
117#include "llvm/Transforms/Utils/LoopUtils.h"
118#include "llvm/Transforms/Utils/ScalarEvolutionExpander.h"
119#include <algorithm>
120#include <cassert>
121#include <cstddef>
122#include <cstdint>
123#include <iterator>
124#include <limits>
125#include <map>
126#include <numeric>
127#include <optional>
128#include <utility>

130using namespace llvm;

132#define DEBUG_TYPE"loop-reduce" "loop-reduce"

134/// MaxIVUsers is an arbitrary threshold that provides an early opportunity for
135/// bail out. This threshold is far beyond the number of users that LSR can
136/// conceivably solve, so it should not affect generated code, but catches the
137/// worst cases before LSR burns too much compile time and stack space.
138static const unsigned MaxIVUsers = 200;

140/// Limit the size of expression that SCEV-based salvaging will attempt to
141/// translate into a DIExpression.
142/// Choose a maximum size such that debuginfo is not excessively increased and
143/// the salvaging is not too expensive for the compiler.
144static const unsigned MaxSCEVSalvageExpressionSize = 64;

146// Cleanup congruent phis after LSR phi expansion.
147static cl::opt<bool> EnablePhiElim(
"enable-lsr-phielim", cl::Hidden, cl::init(true),
cl::desc("Enable LSR phi elimination"));

151// The flag adds instruction count to solutions cost comparison.
152static cl::opt<bool> InsnsCost(
"lsr-insns-cost", cl::Hidden, cl::init(true),
cl::desc("Add instruction count to a LSR cost model"));

156// Flag to choose how to narrow complex lsr solution
157static cl::opt<bool> LSRExpNarrow(
"lsr-exp-narrow", cl::Hidden, cl::init(false),
cl::desc("Narrow LSR complex solution using"
         " expectation of registers number"));

162// Flag to narrow search space by filtering non-optimal formulae with
163// the same ScaledReg and Scale.
164static cl::opt<bool> FilterSameScaledReg(
  "lsr-filter-same-scaled-reg", cl::Hidden, cl::init(true),
  cl::desc("Narrow LSR search space by filtering non-optimal formulae"
           " with the same ScaledReg and Scale"));

169static cl::opt<TTI::AddressingModeKind> PreferredAddresingMode(
"lsr-preferred-addressing-mode", cl::Hidden, cl::init(TTI::AMK_None),
 cl::desc("A flag that overrides the target's preferred addressing mode."),
 cl::values(clEnumValN(TTI::AMK_None,llvm::cl::OptionEnumValue { "none", int(TTI::AMK_None), "Don't prefer any addressing mode"
 }
                       "none",llvm::cl::OptionEnumValue { "none", int(TTI::AMK_None), "Don't prefer any addressing mode"
 }
                       "Don't prefer any addressing mode")llvm::cl::OptionEnumValue { "none", int(TTI::AMK_None), "Don't prefer any addressing mode"
 },
            clEnumValN(TTI::AMK_PreIndexed,llvm::cl::OptionEnumValue { "preindexed", int(TTI::AMK_PreIndexed
), "Prefer pre-indexed addressing mode" }
                       "preindexed",llvm::cl::OptionEnumValue { "preindexed", int(TTI::AMK_PreIndexed
), "Prefer pre-indexed addressing mode" }
                       "Prefer pre-indexed addressing mode")llvm::cl::OptionEnumValue { "preindexed", int(TTI::AMK_PreIndexed
), "Prefer pre-indexed addressing mode" },
            clEnumValN(TTI::AMK_PostIndexed,llvm::cl::OptionEnumValue { "postindexed", int(TTI::AMK_PostIndexed
), "Prefer post-indexed addressing mode" }
                       "postindexed",llvm::cl::OptionEnumValue { "postindexed", int(TTI::AMK_PostIndexed
), "Prefer post-indexed addressing mode" }
                       "Prefer post-indexed addressing mode")llvm::cl::OptionEnumValue { "postindexed", int(TTI::AMK_PostIndexed
), "Prefer post-indexed addressing mode" }));

182static cl::opt<unsigned> ComplexityLimit(
"lsr-complexity-limit", cl::Hidden,
cl::init(std::numeric_limits<uint16_t>::max()),
cl::desc("LSR search space complexity limit"));

187static cl::opt<unsigned> SetupCostDepthLimit(
  "lsr-setupcost-depth-limit", cl::Hidden, cl::init(7),
  cl::desc("The limit on recursion depth for LSRs setup cost"));

191static cl::opt<bool> AllowTerminatingConditionFoldingAfterLSR(
  "lsr-term-fold", cl::Hidden, cl::init(false),
  cl::desc("Attempt to replace primary IV with other IV."));

195static cl::opt<bool> AllowDropSolutionIfLessProfitable(
  "lsr-drop-solution", cl::Hidden, cl::init(false),
  cl::desc("Attempt to drop solution if it is less profitable"));

199STATISTIC(NumTermFold,static llvm::Statistic NumTermFold = {"loop-reduce", "NumTermFold"
, "Number of terminating condition fold recognized and performed"
}
        "Number of terminating condition fold recognized and performed")static llvm::Statistic NumTermFold = {"loop-reduce", "NumTermFold"
, "Number of terminating condition fold recognized and performed"
};

202#ifndef NDEBUG
203// Stress test IV chain generation.
204static cl::opt<bool> StressIVChain(
"stress-ivchain", cl::Hidden, cl::init(false),
cl::desc("Stress test LSR IV chains"));
207#else
208static bool StressIVChain = false;
209#endif

211namespace {

213struct MemAccessTy {
/// Used in situations where the accessed memory type is unknown.
static const unsigned UnknownAddressSpace =
    std::numeric_limits<unsigned>::max();

Type *MemTy = nullptr;
unsigned AddrSpace = UnknownAddressSpace;

MemAccessTy() = default;
MemAccessTy(Type *Ty, unsigned AS) : MemTy(Ty), AddrSpace(AS) {}

bool operator==(MemAccessTy Other) const {
  return MemTy == Other.MemTy && AddrSpace == Other.AddrSpace;
}

bool operator!=(MemAccessTy Other) const { return !(*this == Other); }

static MemAccessTy getUnknown(LLVMContext &Ctx,
                              unsigned AS = UnknownAddressSpace) {
  return MemAccessTy(Type::getVoidTy(Ctx), AS);
}

Type *getType() { return MemTy; }
236};

238/// This class holds data which is used to order reuse candidates.
239class RegSortData {
240public:
/// This represents the set of LSRUse indices which reference
/// a particular register.
SmallBitVector UsedByIndices;

void print(raw_ostream &OS) const;
void dump() const;
247};

249} // end anonymous namespace

251#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
252void RegSortData::print(raw_ostream &OS) const {
OS << "[NumUses=" << UsedByIndices.count() << ']';
254}

256LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void RegSortData::dump() const {
print(errs()); errs() << '\n';
258}
259#endif

261namespace {

263/// Map register candidates to information about how they are used.
264class RegUseTracker {
using RegUsesTy = DenseMap<const SCEV *, RegSortData>;

RegUsesTy RegUsesMap;
SmallVector<const SCEV *, 16> RegSequence;

270public:
void countRegister(const SCEV *Reg, size_t LUIdx);
void dropRegister(const SCEV *Reg, size_t LUIdx);
void swapAndDropUse(size_t LUIdx, size_t LastLUIdx);

bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;

const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;

void clear();

using iterator = SmallVectorImpl<const SCEV *>::iterator;
using const_iterator = SmallVectorImpl<const SCEV *>::const_iterator;

iterator begin() { return RegSequence.begin(); }
iterator end()   { return RegSequence.end(); }
const_iterator begin() const { return RegSequence.begin(); }
const_iterator end() const   { return RegSequence.end(); }
288};

290} // end anonymous namespace

292void
293RegUseTracker::countRegister(const SCEV *Reg, size_t LUIdx) {
std::pair<RegUsesTy::iterator, bool> Pair =
  RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
RegSortData &RSD = Pair.first->second;
if (Pair.second)
  RegSequence.push_back(Reg);
RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
RSD.UsedByIndices.set(LUIdx);
301}

303void
304RegUseTracker::dropRegister(const SCEV *Reg, size_t LUIdx) {
RegUsesTy::iterator It = RegUsesMap.find(Reg);
assert(It != RegUsesMap.end())(static_cast <bool> (It != RegUsesMap.end()) ? void (0)
 : __assert_fail ("It != RegUsesMap.end()", "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 306, __extension__ __PRETTY_FUNCTION__));
RegSortData &RSD = It->second;
assert(RSD.UsedByIndices.size() > LUIdx)(static_cast <bool> (RSD.UsedByIndices.size() > LUIdx
) ? void (0) : __assert_fail ("RSD.UsedByIndices.size() > LUIdx"
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 308, __extension__
 __PRETTY_FUNCTION__));
RSD.UsedByIndices.reset(LUIdx);
310}

312void
313RegUseTracker::swapAndDropUse(size_t LUIdx, size_t LastLUIdx) {
assert(LUIdx <= LastLUIdx)(static_cast <bool> (LUIdx <= LastLUIdx) ? void (0) :
 __assert_fail ("LUIdx <= LastLUIdx", "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 314, __extension__ __PRETTY_FUNCTION__));

// Update RegUses. The data structure is not optimized for this purpose;
// we must iterate through it and update each of the bit vectors.
for (auto &Pair : RegUsesMap) {
  SmallBitVector &UsedByIndices = Pair.second.UsedByIndices;
  if (LUIdx < UsedByIndices.size())
    UsedByIndices[LUIdx] =
      LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : false;
  UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx));
}
325}

327bool
328RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
if (I == RegUsesMap.end())
  return false;
const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
int i = UsedByIndices.find_first();
if (i == -1) return false;
if ((size_t)i != LUIdx) return true;
return UsedByIndices.find_next(i) != -1;
337}

339const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
assert(I != RegUsesMap.end() && "Unknown register!")(static_cast <bool> (I != RegUsesMap.end() && "Unknown register!"
) ? void (0) : __assert_fail ("I != RegUsesMap.end() && \"Unknown register!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 341, __extension__
 __PRETTY_FUNCTION__));
return I->second.UsedByIndices;
343}

345void RegUseTracker::clear() {
RegUsesMap.clear();
RegSequence.clear();
348}

350namespace {

352/// This class holds information that describes a formula for computing
353/// satisfying a use. It may include broken-out immediates and scaled registers.
354struct Formula {
/// Global base address used for complex addressing.
GlobalValue *BaseGV = nullptr;

/// Base offset for complex addressing.
int64_t BaseOffset = 0;

/// Whether any complex addressing has a base register.
bool HasBaseReg = false;

/// The scale of any complex addressing.
int64_t Scale = 0;

/// The list of "base" registers for this use. When this is non-empty. The
/// canonical representation of a formula is
/// 1. BaseRegs.size > 1 implies ScaledReg != NULL and
/// 2. ScaledReg != NULL implies Scale != 1 || !BaseRegs.empty().
/// 3. The reg containing recurrent expr related with currect loop in the
/// formula should be put in the ScaledReg.
/// #1 enforces that the scaled register is always used when at least two
/// registers are needed by the formula: e.g., reg1 + reg2 is reg1 + 1 * reg2.
/// #2 enforces that 1 * reg is reg.
/// #3 ensures invariant regs with respect to current loop can be combined
/// together in LSR codegen.
/// This invariant can be temporarily broken while building a formula.
/// However, every formula inserted into the LSRInstance must be in canonical
/// form.
SmallVector<const SCEV *, 4> BaseRegs;

/// The 'scaled' register for this use. This should be non-null when Scale is
/// not zero.
const SCEV *ScaledReg = nullptr;

/// An additional constant offset which added near the use. This requires a
/// temporary register, but the offset itself can live in an add immediate
/// field rather than a register.
int64_t UnfoldedOffset = 0;

Formula() = default;

void initialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE);

bool isCanonical(const Loop &L) const;

void canonicalize(const Loop &L);

bool unscale();

bool hasZeroEnd() const;

size_t getNumRegs() const;
Type *getType() const;

void deleteBaseReg(const SCEV *&S);

bool referencesReg(const SCEV *S) const;
bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
                                const RegUseTracker &RegUses) const;

void print(raw_ostream &OS) const;
void dump() const;
415};

417} // end anonymous namespace

419/// Recursion helper for initialMatch.
420static void DoInitialMatch(const SCEV *S, Loop *L,
                         SmallVectorImpl<const SCEV *> &Good,
                         SmallVectorImpl<const SCEV *> &Bad,
                         ScalarEvolution &SE) {
// Collect expressions which properly dominate the loop header.
if (SE.properlyDominates(S, L->getHeader())) {
  Good.push_back(S);
  return;
}

// Look at add operands.
if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
  for (const SCEV *S : Add->operands())
    DoInitialMatch(S, L, Good, Bad, SE);
  return;
}

// Look at addrec operands.
if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
  if (!AR->getStart()->isZero() && AR->isAffine()) {
    DoInitialMatch(AR->getStart(), L, Good, Bad, SE);
    DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
                                    AR->getStepRecurrence(SE),
                                    // FIXME: AR->getNoWrapFlags()
                                    AR->getLoop(), SCEV::FlagAnyWrap),
                   L, Good, Bad, SE);
    return;
  }

// Handle a multiplication by -1 (negation) if it didn't fold.
if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
  if (Mul->getOperand(0)->isAllOnesValue()) {
    SmallVector<const SCEV *, 4> Ops(drop_begin(Mul->operands()));
    const SCEV *NewMul = SE.getMulExpr(Ops);

    SmallVector<const SCEV *, 4> MyGood;
    SmallVector<const SCEV *, 4> MyBad;
    DoInitialMatch(NewMul, L, MyGood, MyBad, SE);
    const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
      SE.getEffectiveSCEVType(NewMul->getType())));
    for (const SCEV *S : MyGood)
      Good.push_back(SE.getMulExpr(NegOne, S));
    for (const SCEV *S : MyBad)
      Bad.push_back(SE.getMulExpr(NegOne, S));
    return;
  }

// Ok, we can't do anything interesting. Just stuff the whole thing into a
// register and hope for the best.
Bad.push_back(S);
470}

472/// Incorporate loop-variant parts of S into this Formula, attempting to keep
473/// all loop-invariant and loop-computable values in a single base register.
474void Formula::initialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) {
SmallVector<const SCEV *, 4> Good;
SmallVector<const SCEV *, 4> Bad;
DoInitialMatch(S, L, Good, Bad, SE);
if (!Good.empty()) {
  const SCEV *Sum = SE.getAddExpr(Good);
  if (!Sum->isZero())
    BaseRegs.push_back(Sum);
  HasBaseReg = true;
}
if (!Bad.empty()) {
  const SCEV *Sum = SE.getAddExpr(Bad);
  if (!Sum->isZero())
    BaseRegs.push_back(Sum);
  HasBaseReg = true;
}
canonicalize(*L);
491}

493static bool containsAddRecDependentOnLoop(const SCEV *S, const Loop &L) {
return SCEVExprContains(S, [&L](const SCEV *S) {
  return isa<SCEVAddRecExpr>(S) && (cast<SCEVAddRecExpr>(S)->getLoop() == &L);
});
497}

499/// Check whether or not this formula satisfies the canonical
500/// representation.
501/// \see Formula::BaseRegs.
502bool Formula::isCanonical(const Loop &L) const {
if (!ScaledReg)
  return BaseRegs.size() <= 1;

if (Scale != 1)
  return true;

if (Scale == 1 && BaseRegs.empty())
  return false;

if (containsAddRecDependentOnLoop(ScaledReg, L))
  return true;

// If ScaledReg is not a recurrent expr, or it is but its loop is not current
// loop, meanwhile BaseRegs contains a recurrent expr reg related with current
// loop, we want to swap the reg in BaseRegs with ScaledReg.
return none_of(BaseRegs, [&L](const SCEV *S) {
  return containsAddRecDependentOnLoop(S, L);
});
521}

523/// Helper method to morph a formula into its canonical representation.
524/// \see Formula::BaseRegs.
525/// Every formula having more than one base register, must use the ScaledReg
526/// field. Otherwise, we would have to do special cases everywhere in LSR
527/// to treat reg1 + reg2 + ... the same way as reg1 + 1*reg2 + ...
528/// On the other hand, 1*reg should be canonicalized into reg.
529void Formula::canonicalize(const Loop &L) {
if (isCanonical(L))
  return;

if (BaseRegs.empty()) {
  // No base reg? Use scale reg with scale = 1 as such.
  assert(ScaledReg && "Expected 1*reg => reg")(static_cast <bool> (ScaledReg && "Expected 1*reg => reg"
) ? void (0) : __assert_fail ("ScaledReg && \"Expected 1*reg => reg\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 535, __extension__
 __PRETTY_FUNCTION__));
  assert(Scale == 1 && "Expected 1*reg => reg")(static_cast <bool> (Scale == 1 && "Expected 1*reg => reg"
) ? void (0) : __assert_fail ("Scale == 1 && \"Expected 1*reg => reg\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 536, __extension__
 __PRETTY_FUNCTION__));
  BaseRegs.push_back(ScaledReg);
  Scale = 0;
  ScaledReg = nullptr;
  return;
}

// Keep the invariant sum in BaseRegs and one of the variant sum in ScaledReg.
if (!ScaledReg) {
  ScaledReg = BaseRegs.pop_back_val();
  Scale = 1;
}

// If ScaledReg is an invariant with respect to L, find the reg from
// BaseRegs containing the recurrent expr related with Loop L. Swap the
// reg with ScaledReg.
if (!containsAddRecDependentOnLoop(ScaledReg, L)) {
  auto I = find_if(BaseRegs, [&L](const SCEV *S) {
    return containsAddRecDependentOnLoop(S, L);
  });
  if (I != BaseRegs.end())
    std::swap(ScaledReg, *I);
}
assert(isCanonical(L) && "Failed to canonicalize?")(static_cast <bool> (isCanonical(L) && "Failed to canonicalize?"
) ? void (0) : __assert_fail ("isCanonical(L) && \"Failed to canonicalize?\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 559, __extension__
 __PRETTY_FUNCTION__));
560}

562/// Get rid of the scale in the formula.
563/// In other words, this method morphes reg1 + 1*reg2 into reg1 + reg2.
564/// \return true if it was possible to get rid of the scale, false otherwise.
565/// \note After this operation the formula may not be in the canonical form.
566bool Formula::unscale() {
if (Scale != 1)
  return false;
Scale = 0;
BaseRegs.push_back(ScaledReg);
ScaledReg = nullptr;
return true;
573}

575bool Formula::hasZeroEnd() const {
if (UnfoldedOffset || BaseOffset)
  return false;
if (BaseRegs.size() != 1 || ScaledReg)
  return false;
return true;
581}

583/// Return the total number of register operands used by this formula. This does
584/// not include register uses implied by non-constant addrec strides.
585size_t Formula::getNumRegs() const {
return !!ScaledReg + BaseRegs.size();
587}

589/// Return the type of this formula, if it has one, or null otherwise. This type
590/// is meaningless except for the bit size.
591Type *Formula::getType() const {
return !BaseRegs.empty() ? BaseRegs.front()->getType() :
       ScaledReg ? ScaledReg->getType() :
       BaseGV ? BaseGV->getType() :
       nullptr;
596}

598/// Delete the given base reg from the BaseRegs list.
599void Formula::deleteBaseReg(const SCEV *&S) {
if (&S != &BaseRegs.back())
  std::swap(S, BaseRegs.back());
BaseRegs.pop_back();
603}

605/// Test if this formula references the given register.
606bool Formula::referencesReg(const SCEV *S) const {
return S == ScaledReg || is_contained(BaseRegs, S);
608}

610/// Test whether this formula uses registers which are used by uses other than
611/// the use with the given index.
612bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
                                       const RegUseTracker &RegUses) const {
if (ScaledReg)
  if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
    return true;
for (const SCEV *BaseReg : BaseRegs)
  if (RegUses.isRegUsedByUsesOtherThan(BaseReg, LUIdx))
    return true;
return false;
621}

623#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
624void Formula::print(raw_ostream &OS) const {
bool First = true;
if (BaseGV) {
  if (!First) OS << " + "; else First = false;
  BaseGV->printAsOperand(OS, /*PrintType=*/false);
}
if (BaseOffset != 0) {
  if (!First) OS << " + "; else First = false;
  OS << BaseOffset;
}
for (const SCEV *BaseReg : BaseRegs) {
  if (!First) OS << " + "; else First = false;
  OS << "reg(" << *BaseReg << ')';
}
if (HasBaseReg && BaseRegs.empty()) {
  if (!First) OS << " + "; else First = false;
  OS << "**error: HasBaseReg**";
} else if (!HasBaseReg && !BaseRegs.empty()) {
  if (!First) OS << " + "; else First = false;
  OS << "**error: !HasBaseReg**";
}
if (Scale != 0) {
  if (!First) OS << " + "; else First = false;
  OS << Scale << "*reg(";
  if (ScaledReg)
    OS << *ScaledReg;
  else
    OS << "<unknown>";
  OS << ')';
}
if (UnfoldedOffset != 0) {
  if (!First) OS << " + ";
  OS << "imm(" << UnfoldedOffset << ')';
}
658}

660LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void Formula::dump() const {
print(errs()); errs() << '\n';
662}
663#endif

665/// Return true if the given addrec can be sign-extended without changing its
666/// value.
667static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
Type *WideTy =
  IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1);
return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
671}

673/// Return true if the given add can be sign-extended without changing its
674/// value.
675static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
Type *WideTy =
  IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
679}

681/// Return true if the given mul can be sign-extended without changing its
682/// value.
683static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
Type *WideTy =
  IntegerType::get(SE.getContext(),
                   SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
688}

690/// Return an expression for LHS /s RHS, if it can be determined and if the
691/// remainder is known to be zero, or null otherwise. If IgnoreSignificantBits
692/// is true, expressions like (X * Y) /s Y are simplified to X, ignoring that
693/// the multiplication may overflow, which is useful when the result will be
694/// used in a context where the most significant bits are ignored.
695static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
                              ScalarEvolution &SE,
                              bool IgnoreSignificantBits = false) {
// Handle the trivial case, which works for any SCEV type.
if (LHS == RHS)
  return SE.getConstant(LHS->getType(), 1);

// Handle a few RHS special cases.
const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
if (RC) {
  const APInt &RA = RC->getAPInt();
  // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
  // some folding.
  if (RA.isAllOnes()) {
    if (LHS->getType()->isPointerTy())
      return nullptr;
    return SE.getMulExpr(LHS, RC);
  }
  // Handle x /s 1 as x.
  if (RA == 1)
    return LHS;
}

// Check for a division of a constant by a constant.
if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
  if (!RC)
    return nullptr;
  const APInt &LA = C->getAPInt();
  const APInt &RA = RC->getAPInt();
  if (LA.srem(RA) != 0)
    return nullptr;
  return SE.getConstant(LA.sdiv(RA));
}

// Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
  if ((IgnoreSignificantBits || isAddRecSExtable(AR, SE)) && AR->isAffine()) {
    const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
                                    IgnoreSignificantBits);
    if (!Step) return nullptr;
    const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
                                     IgnoreSignificantBits);
    if (!Start) return nullptr;
    // FlagNW is independent of the start value, step direction, and is
    // preserved with smaller magnitude steps.
    // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
    return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap);
  }
  return nullptr;
}

// Distribute the sdiv over add operands, if the add doesn't overflow.
if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
  if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
    SmallVector<const SCEV *, 8> Ops;
    for (const SCEV *S : Add->operands()) {
      const SCEV *Op = getExactSDiv(S, RHS, SE, IgnoreSignificantBits);
      if (!Op) return nullptr;
      Ops.push_back(Op);
    }
    return SE.getAddExpr(Ops);
  }
  return nullptr;
}

// Check for a multiply operand that we can pull RHS out of.
if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
  if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
    // Handle special case C1*X*Y /s C2*X*Y.
    if (const SCEVMulExpr *MulRHS = dyn_cast<SCEVMulExpr>(RHS)) {
      if (IgnoreSignificantBits || isMulSExtable(MulRHS, SE)) {
        const SCEVConstant *LC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
        const SCEVConstant *RC =
            dyn_cast<SCEVConstant>(MulRHS->getOperand(0));
        if (LC && RC) {
          SmallVector<const SCEV *, 4> LOps(drop_begin(Mul->operands()));
          SmallVector<const SCEV *, 4> ROps(drop_begin(MulRHS->operands()));
          if (LOps == ROps)
            return getExactSDiv(LC, RC, SE, IgnoreSignificantBits);
        }
      }
    }

    SmallVector<const SCEV *, 4> Ops;
    bool Found = false;
    for (const SCEV *S : Mul->operands()) {
      if (!Found)
        if (const SCEV *Q = getExactSDiv(S, RHS, SE,
                                         IgnoreSignificantBits)) {
          S = Q;
          Found = true;
        }
      Ops.push_back(S);
    }
    return Found ? SE.getMulExpr(Ops) : nullptr;
  }
  return nullptr;
}

// Otherwise we don't know.
return nullptr;
796}

798/// If S involves the addition of a constant integer value, return that integer
799/// value, and mutate S to point to a new SCEV with that value excluded.
800static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
  if (C->getAPInt().getSignificantBits() <= 64) {
    S = SE.getConstant(C->getType(), 0);
    return C->getValue()->getSExtValue();
  }
} else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
  SmallVector<const SCEV *, 8> NewOps(Add->operands());
  int64_t Result = ExtractImmediate(NewOps.front(), SE);
  if (Result != 0)
    S = SE.getAddExpr(NewOps);
  return Result;
} else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
  SmallVector<const SCEV *, 8> NewOps(AR->operands());
  int64_t Result = ExtractImmediate(NewOps.front(), SE);
  if (Result != 0)
    S = SE.getAddRecExpr(NewOps, AR->getLoop(),
                         // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
                         SCEV::FlagAnyWrap);
  return Result;
}
return 0;
822}

824/// If S involves the addition of a GlobalValue address, return that symbol, and
825/// mutate S to point to a new SCEV with that value excluded.
826static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
  if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
    S = SE.getConstant(GV->getType(), 0);
    return GV;
  }
} else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
  SmallVector<const SCEV *, 8> NewOps(Add->operands());
  GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
  if (Result)
    S = SE.getAddExpr(NewOps);
  return Result;
} else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
  SmallVector<const SCEV *, 8> NewOps(AR->operands());
  GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
  if (Result)
    S = SE.getAddRecExpr(NewOps, AR->getLoop(),
                         // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
                         SCEV::FlagAnyWrap);
  return Result;
}
return nullptr;
848}

850/// Returns true if the specified instruction is using the specified value as an
851/// address.
852static bool isAddressUse(const TargetTransformInfo &TTI,
                       Instruction *Inst, Value *OperandVal) {
bool isAddress = isa<LoadInst>(Inst);
if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
  if (SI->getPointerOperand() == OperandVal)
    isAddress = true;
} else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
  // Addressing modes can also be folded into prefetches and a variety
  // of intrinsics.
  switch (II->getIntrinsicID()) {
  case Intrinsic::memset:
  case Intrinsic::prefetch:
  case Intrinsic::masked_load:
    if (II->getArgOperand(0) == OperandVal)
      isAddress = true;
    break;
  case Intrinsic::masked_store:
    if (II->getArgOperand(1) == OperandVal)
      isAddress = true;
    break;
  case Intrinsic::memmove:
  case Intrinsic::memcpy:
    if (II->getArgOperand(0) == OperandVal ||
        II->getArgOperand(1) == OperandVal)
      isAddress = true;
    break;
  default: {
    MemIntrinsicInfo IntrInfo;
    if (TTI.getTgtMemIntrinsic(II, IntrInfo)) {
      if (IntrInfo.PtrVal == OperandVal)
        isAddress = true;
    }
  }
  }
} else if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Inst)) {
  if (RMW->getPointerOperand() == OperandVal)
    isAddress = true;
} else if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(Inst)) {
  if (CmpX->getPointerOperand() == OperandVal)
    isAddress = true;
}
return isAddress;
894}

896/// Return the type of the memory being accessed.
897static MemAccessTy getAccessType(const TargetTransformInfo &TTI,
                               Instruction *Inst, Value *OperandVal) {
MemAccessTy AccessTy(Inst->getType(), MemAccessTy::UnknownAddressSpace);
if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
  AccessTy.MemTy = SI->getOperand(0)->getType();
  AccessTy.AddrSpace = SI->getPointerAddressSpace();
} else if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
  AccessTy.AddrSpace = LI->getPointerAddressSpace();
} else if (const AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Inst)) {
  AccessTy.AddrSpace = RMW->getPointerAddressSpace();
} else if (const AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(Inst)) {
  AccessTy.AddrSpace = CmpX->getPointerAddressSpace();
} else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
  switch (II->getIntrinsicID()) {
  case Intrinsic::prefetch:
  case Intrinsic::memset:
    AccessTy.AddrSpace = II->getArgOperand(0)->getType()->getPointerAddressSpace();
    AccessTy.MemTy = OperandVal->getType();
    break;
  case Intrinsic::memmove:
  case Intrinsic::memcpy:
    AccessTy.AddrSpace = OperandVal->getType()->getPointerAddressSpace();
    AccessTy.MemTy = OperandVal->getType();
    break;
  case Intrinsic::masked_load:
    AccessTy.AddrSpace =
        II->getArgOperand(0)->getType()->getPointerAddressSpace();
    break;
  case Intrinsic::masked_store:
    AccessTy.MemTy = II->getOperand(0)->getType();
    AccessTy.AddrSpace =
        II->getArgOperand(1)->getType()->getPointerAddressSpace();
    break;
  default: {
    MemIntrinsicInfo IntrInfo;
    if (TTI.getTgtMemIntrinsic(II, IntrInfo) && IntrInfo.PtrVal) {
      AccessTy.AddrSpace
        = IntrInfo.PtrVal->getType()->getPointerAddressSpace();
    }

    break;
  }
  }
}

// All pointers have the same requirements, so canonicalize them to an
// arbitrary pointer type to minimize variation.
if (PointerType *PTy = dyn_cast<PointerType>(AccessTy.MemTy))
  AccessTy.MemTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
                                    PTy->getAddressSpace());

return AccessTy;
949}

951/// Return true if this AddRec is already a phi in its loop.
952static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
for (PHINode &PN : AR->getLoop()->getHeader()->phis()) {
  if (SE.isSCEVable(PN.getType()) &&
      (SE.getEffectiveSCEVType(PN.getType()) ==
       SE.getEffectiveSCEVType(AR->getType())) &&
      SE.getSCEV(&PN) == AR)
    return true;
}
return false;
961}

963/// Check if expanding this expression is likely to incur significant cost. This
964/// is tricky because SCEV doesn't track which expressions are actually computed
965/// by the current IR.
966///
967/// We currently allow expansion of IV increments that involve adds,
968/// multiplication by constants, and AddRecs from existing phis.
969///
970/// TODO: Allow UDivExpr if we can find an existing IV increment that is an
971/// obvious multiple of the UDivExpr.
972static bool isHighCostExpansion(const SCEV *S,
                              SmallPtrSetImpl<const SCEV*> &Processed,
                              ScalarEvolution &SE) {
// Zero/One operand expressions
switch (S->getSCEVType()) {
case scUnknown:
case scConstant:
case scVScale:
  return false;
case scTruncate:
  return isHighCostExpansion(cast<SCEVTruncateExpr>(S)->getOperand(),
                             Processed, SE);
case scZeroExtend:
  return isHighCostExpansion(cast<SCEVZeroExtendExpr>(S)->getOperand(),
                             Processed, SE);
case scSignExtend:
  return isHighCostExpansion(cast<SCEVSignExtendExpr>(S)->getOperand(),
                             Processed, SE);
default:
  break;
}

if (!Processed.insert(S).second)
  return false;

if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
  for (const SCEV *S : Add->operands()) {
    if (isHighCostExpansion(S, Processed, SE))
      return true;
  }
  return false;
}

if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
  if (Mul->getNumOperands() == 2) {
    // Multiplication by a constant is ok
    if (isa<SCEVConstant>(Mul->getOperand(0)))
      return isHighCostExpansion(Mul->getOperand(1), Processed, SE);

    // If we have the value of one operand, check if an existing
    // multiplication already generates this expression.
    if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Mul->getOperand(1))) {
      Value *UVal = U->getValue();
      for (User *UR : UVal->users()) {
        // If U is a constant, it may be used by a ConstantExpr.
        Instruction *UI = dyn_cast<Instruction>(UR);
        if (UI && UI->getOpcode() == Instruction::Mul &&
            SE.isSCEVable(UI->getType())) {
          return SE.getSCEV(UI) == Mul;
        }
      }
    }
  }
}

if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
  if (isExistingPhi(AR, SE))
    return false;
}

// Fow now, consider any other type of expression (div/mul/min/max) high cost.
return true;
1034}

1036namespace {

1038class LSRUse;

1040} // end anonymous namespace

1042/// Check if the addressing mode defined by \p F is completely
1043/// folded in \p LU at isel time.
1044/// This includes address-mode folding and special icmp tricks.
1045/// This function returns true if \p LU can accommodate what \p F
1046/// defines and up to 1 base + 1 scaled + offset.
1047/// In other words, if \p F has several base registers, this function may
1048/// still return true. Therefore, users still need to account for
1049/// additional base registers and/or unfolded offsets to derive an
1050/// accurate cost model.
1051static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
                               const LSRUse &LU, const Formula &F);

1054// Get the cost of the scaling factor used in F for LU.
1055static InstructionCost getScalingFactorCost(const TargetTransformInfo &TTI,
                                          const LSRUse &LU, const Formula &F,
                                          const Loop &L);

1059namespace {

1061/// This class is used to measure and compare candidate formulae.
1062class Cost {
const Loop *L = nullptr;
ScalarEvolution *SE = nullptr;
const TargetTransformInfo *TTI = nullptr;
TargetTransformInfo::LSRCost C;
TTI::AddressingModeKind AMK = TTI::AMK_None;

1069public:
Cost() = delete;
Cost(const Loop *L, ScalarEvolution &SE, const TargetTransformInfo &TTI,
     TTI::AddressingModeKind AMK) :
  L(L), SE(&SE), TTI(&TTI), AMK(AMK) {
  C.Insns = 0;
  C.NumRegs = 0;
  C.AddRecCost = 0;
  C.NumIVMuls = 0;
  C.NumBaseAdds = 0;
  C.ImmCost = 0;
  C.SetupCost = 0;
  C.ScaleCost = 0;
}

bool isLess(const Cost &Other) const;

void Lose();

1088#ifndef NDEBUG
// Once any of the metrics loses, they must all remain losers.
bool isValid() {
  return ((C.Insns | C.NumRegs | C.AddRecCost | C.NumIVMuls | C.NumBaseAdds
           | C.ImmCost | C.SetupCost | C.ScaleCost) != ~0u)
    || ((C.Insns & C.NumRegs & C.AddRecCost & C.NumIVMuls & C.NumBaseAdds
         & C.ImmCost & C.SetupCost & C.ScaleCost) == ~0u);
}
1096#endif

bool isLoser() {
  assert(isValid() && "invalid cost")(static_cast <bool> (isValid() && "invalid cost"
) ? void (0) : __assert_fail ("isValid() && \"invalid cost\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 1099, __extension__
 __PRETTY_FUNCTION__));
  return C.NumRegs == ~0u;
}

void RateFormula(const Formula &F,
                 SmallPtrSetImpl<const SCEV *> &Regs,
                 const DenseSet<const SCEV *> &VisitedRegs,
                 const LSRUse &LU,
                 SmallPtrSetImpl<const SCEV *> *LoserRegs = nullptr);

void print(raw_ostream &OS) const;
void dump() const;

1112private:
void RateRegister(const Formula &F, const SCEV *Reg,
                  SmallPtrSetImpl<const SCEV *> &Regs);
void RatePrimaryRegister(const Formula &F, const SCEV *Reg,
                         SmallPtrSetImpl<const SCEV *> &Regs,
                         SmallPtrSetImpl<const SCEV *> *LoserRegs);
1118};

1120/// An operand value in an instruction which is to be replaced with some
1121/// equivalent, possibly strength-reduced, replacement.
1122struct LSRFixup {
/// The instruction which will be updated.
Instruction *UserInst = nullptr;

/// The operand of the instruction which will be replaced. The operand may be
/// used more than once; every instance will be replaced.
Value *OperandValToReplace = nullptr;

/// If this user is to use the post-incremented value of an induction
/// variable, this set is non-empty and holds the loops associated with the
/// induction variable.
PostIncLoopSet PostIncLoops;

/// A constant offset to be added to the LSRUse expression.  This allows
/// multiple fixups to share the same LSRUse with different offsets, for
/// example in an unrolled loop.
int64_t Offset = 0;

LSRFixup() = default;

bool isUseFullyOutsideLoop(const Loop *L) const;

void print(raw_ostream &OS) const;
void dump() const;
1146};

1148/// A DenseMapInfo implementation for holding DenseMaps and DenseSets of sorted
1149/// SmallVectors of const SCEV*.
1150struct UniquifierDenseMapInfo {
static SmallVector<const SCEV *, 4> getEmptyKey() {
  SmallVector<const SCEV *, 4>  V;
  V.push_back(reinterpret_cast<const SCEV *>(-1));
  return V;
}

static SmallVector<const SCEV *, 4> getTombstoneKey() {
  SmallVector<const SCEV *, 4> V;
  V.push_back(reinterpret_cast<const SCEV *>(-2));
  return V;
}

static unsigned getHashValue(const SmallVector<const SCEV *, 4> &V) {
  return static_cast<unsigned>(hash_combine_range(V.begin(), V.end()));
}

static bool isEqual(const SmallVector<const SCEV *, 4> &LHS,
                    const SmallVector<const SCEV *, 4> &RHS) {
  return LHS == RHS;
}
1171};

1173/// This class holds the state that LSR keeps for each use in IVUsers, as well
1174/// as uses invented by LSR itself. It includes information about what kinds of
1175/// things can be folded into the user, information about the user itself, and
1176/// information about how the use may be satisfied.  TODO: Represent multiple
1177/// users of the same expression in common?
1178class LSRUse {
DenseSet<SmallVector<const SCEV *, 4>, UniquifierDenseMapInfo> Uniquifier;

1181public:
/// An enum for a kind of use, indicating what types of scaled and immediate
/// operands it might support.
enum KindType {
  Basic,   ///< A normal use, with no folding.
  Special, ///< A special case of basic, allowing -1 scales.
  Address, ///< An address use; folding according to TargetLowering
  ICmpZero ///< An equality icmp with both operands folded into one.
  // TODO: Add a generic icmp too?
};

using SCEVUseKindPair = PointerIntPair<const SCEV *, 2, KindType>;

KindType Kind;
MemAccessTy AccessTy;

/// The list of operands which are to be replaced.
SmallVector<LSRFixup, 8> Fixups;

/// Keep track of the min and max offsets of the fixups.
int64_t MinOffset = std::numeric_limits<int64_t>::max();
int64_t MaxOffset = std::numeric_limits<int64_t>::min();

/// This records whether all of the fixups using this LSRUse are outside of
/// the loop, in which case some special-case heuristics may be used.
bool AllFixupsOutsideLoop = true;

/// RigidFormula is set to true to guarantee that this use will be associated
/// with a single formula--the one that initially matched. Some SCEV
/// expressions cannot be expanded. This allows LSR to consider the registers
/// used by those expressions without the need to expand them later after
/// changing the formula.
bool RigidFormula = false;

/// This records the widest use type for any fixup using this
/// LSRUse. FindUseWithSimilarFormula can't consider uses with different max
/// fixup widths to be equivalent, because the narrower one may be relying on
/// the implicit truncation to truncate away bogus bits.
Type *WidestFixupType = nullptr;

/// A list of ways to build a value that can satisfy this user.  After the
/// list is populated, one of these is selected heuristically and used to
/// formulate a replacement for OperandValToReplace in UserInst.
SmallVector<Formula, 12> Formulae;

/// The set of register candidates used by all formulae in this LSRUse.
SmallPtrSet<const SCEV *, 4> Regs;

LSRUse(KindType K, MemAccessTy AT) : Kind(K), AccessTy(AT) {}

LSRFixup &getNewFixup() {
  Fixups.push_back(LSRFixup());
  return Fixups.back();
}

void pushFixup(LSRFixup &f) {
  Fixups.push_back(f);
  if (f.Offset > MaxOffset)
    MaxOffset = f.Offset;
  if (f.Offset < MinOffset)
    MinOffset = f.Offset;
}

bool HasFormulaWithSameRegs(const Formula &F) const;
float getNotSelectedProbability(const SCEV *Reg) const;
bool InsertFormula(const Formula &F, const Loop &L);
void DeleteFormula(Formula &F);
void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);

void print(raw_ostream &OS) const;
void dump() const;
1252};

1254} // end anonymous namespace

1256static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
                               LSRUse::KindType Kind, MemAccessTy AccessTy,
                               GlobalValue *BaseGV, int64_t BaseOffset,
                               bool HasBaseReg, int64_t Scale,
                               Instruction *Fixup = nullptr);

1262static unsigned getSetupCost(const SCEV *Reg, unsigned Depth) {
if (isa<SCEVUnknown>(Reg) || isa<SCEVConstant>(Reg))
  return 1;
if (Depth == 0)
  return 0;
if (const auto *S = dyn_cast<SCEVAddRecExpr>(Reg))
  return getSetupCost(S->getStart(), Depth - 1);
if (auto S = dyn_cast<SCEVIntegralCastExpr>(Reg))
  return getSetupCost(S->getOperand(), Depth - 1);
if (auto S = dyn_cast<SCEVNAryExpr>(Reg))
  return std::accumulate(S->operands().begin(), S->operands().end(), 0,
                         [&](unsigned i, const SCEV *Reg) {
                           return i + getSetupCost(Reg, Depth - 1);
                         });
if (auto S = dyn_cast<SCEVUDivExpr>(Reg))
  return getSetupCost(S->getLHS(), Depth - 1) +
         getSetupCost(S->getRHS(), Depth - 1);
return 0;
1280}

1282/// Tally up interesting quantities from the given register.
1283void Cost::RateRegister(const Formula &F, const SCEV *Reg,
                      SmallPtrSetImpl<const SCEV *> &Regs) {
if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
  // If this is an addrec for another loop, it should be an invariant
  // with respect to L since L is the innermost loop (at least
  // for now LSR only handles innermost loops).
  if (AR->getLoop() != L) {
    // If the AddRec exists, consider it's register free and leave it alone.
    if (isExistingPhi(AR, *SE) && AMK != TTI::AMK_PostIndexed)
      return;

    // It is bad to allow LSR for current loop to add induction variables
    // for its sibling loops.
    if (!AR->getLoop()->contains(L)) {
      Lose();
      return;
    }

    // Otherwise, it will be an invariant with respect to Loop L.
    ++C.NumRegs;
    return;
  }

  unsigned LoopCost = 1;
  if (TTI->isIndexedLoadLegal(TTI->MIM_PostInc, AR->getType()) ||
      TTI->isIndexedStoreLegal(TTI->MIM_PostInc, AR->getType())) {

    // If the step size matches the base offset, we could use pre-indexed
    // addressing.
    if (AMK == TTI::AMK_PreIndexed) {
      if (auto *Step = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE)))
        if (Step->getAPInt() == F.BaseOffset)
          LoopCost = 0;
    } else if (AMK == TTI::AMK_PostIndexed) {
      const SCEV *LoopStep = AR->getStepRecurrence(*SE);
      if (isa<SCEVConstant>(LoopStep)) {
        const SCEV *LoopStart = AR->getStart();
        if (!isa<SCEVConstant>(LoopStart) &&
            SE->isLoopInvariant(LoopStart, L))
          LoopCost = 0;
      }
    }
  }
  C.AddRecCost += LoopCost;

  // Add the step value register, if it needs one.
  // TODO: The non-affine case isn't precisely modeled here.
  if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
    if (!Regs.count(AR->getOperand(1))) {
      RateRegister(F, AR->getOperand(1), Regs);
      if (isLoser())
        return;
    }
  }
}
++C.NumRegs;

// Rough heuristic; favor registers which don't require extra setup
// instructions in the preheader.
C.SetupCost += getSetupCost(Reg, SetupCostDepthLimit);
// Ensure we don't, even with the recusion limit, produce invalid costs.
C.SetupCost = std::min<unsigned>(C.SetupCost, 1 << 16);

C.NumIVMuls += isa<SCEVMulExpr>(Reg) &&
             SE->hasComputableLoopEvolution(Reg, L);
1348}

1350/// Record this register in the set. If we haven't seen it before, rate
1351/// it. Optional LoserRegs provides a way to declare any formula that refers to
1352/// one of those regs an instant loser.
1353void Cost::RatePrimaryRegister(const Formula &F, const SCEV *Reg,
                             SmallPtrSetImpl<const SCEV *> &Regs,
                             SmallPtrSetImpl<const SCEV *> *LoserRegs) {
if (LoserRegs && LoserRegs->count(Reg)) {
  Lose();
  return;
}
if (Regs.insert(Reg).second) {
  RateRegister(F, Reg, Regs);
  if (LoserRegs && isLoser())
    LoserRegs->insert(Reg);
}
1365}

1367void Cost::RateFormula(const Formula &F,
                     SmallPtrSetImpl<const SCEV *> &Regs,
                     const DenseSet<const SCEV *> &VisitedRegs,
                     const LSRUse &LU,
                     SmallPtrSetImpl<const SCEV *> *LoserRegs) {
if (isLoser())
  return;
assert(F.isCanonical(*L) && "Cost is accurate only for canonical formula")(static_cast <bool> (F.isCanonical(*L) && "Cost is accurate only for canonical formula"
) ? void (0) : __assert_fail ("F.isCanonical(*L) && \"Cost is accurate only for canonical formula\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 1374, __extension__
 __PRETTY_FUNCTION__));
// Tally up the registers.
unsigned PrevAddRecCost = C.AddRecCost;
unsigned PrevNumRegs = C.NumRegs;
unsigned PrevNumBaseAdds = C.NumBaseAdds;
if (const SCEV *ScaledReg = F.ScaledReg) {
  if (VisitedRegs.count(ScaledReg)) {
    Lose();
    return;
  }
  RatePrimaryRegister(F, ScaledReg, Regs, LoserRegs);
  if (isLoser())
    return;
}
for (const SCEV *BaseReg : F.BaseRegs) {
  if (VisitedRegs.count(BaseReg)) {
    Lose();
    return;
  }
  RatePrimaryRegister(F, BaseReg, Regs, LoserRegs);
  if (isLoser())
    return;
}

// Determine how many (unfolded) adds we'll need inside the loop.
size_t NumBaseParts = F.getNumRegs();
if (NumBaseParts > 1)
  // Do not count the base and a possible second register if the target
  // allows to fold 2 registers.
  C.NumBaseAdds +=
      NumBaseParts - (1 + (F.Scale && isAMCompletelyFolded(*TTI, LU, F)));
C.NumBaseAdds += (F.UnfoldedOffset != 0);

// Accumulate non-free scaling amounts.
C.ScaleCost += *getScalingFactorCost(*TTI, LU, F, *L).getValue();

// Tally up the non-zero immediates.
for (const LSRFixup &Fixup : LU.Fixups) {
  int64_t O = Fixup.Offset;
  int64_t Offset = (uint64_t)O + F.BaseOffset;
  if (F.BaseGV)
    C.ImmCost += 64; // Handle symbolic values conservatively.
                   // TODO: This should probably be the pointer size.
  else if (Offset != 0)
    C.ImmCost += APInt(64, Offset, true).getSignificantBits();

  // Check with target if this offset with this instruction is
  // specifically not supported.
  if (LU.Kind == LSRUse::Address && Offset != 0 &&
      !isAMCompletelyFolded(*TTI, LSRUse::Address, LU.AccessTy, F.BaseGV,
                            Offset, F.HasBaseReg, F.Scale, Fixup.UserInst))
    C.NumBaseAdds++;
}

// If we don't count instruction cost exit here.
if (!InsnsCost) {
  assert(isValid() && "invalid cost")(static_cast <bool> (isValid() && "invalid cost"
) ? void (0) : __assert_fail ("isValid() && \"invalid cost\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 1430, __extension__
 __PRETTY_FUNCTION__));
  return;
}

// Treat every new register that exceeds TTI.getNumberOfRegisters() - 1 as
// additional instruction (at least fill).
// TODO: Need distinguish register class?
unsigned TTIRegNum = TTI->getNumberOfRegisters(
                     TTI->getRegisterClassForType(false, F.getType())) - 1;
if (C.NumRegs > TTIRegNum) {
  // Cost already exceeded TTIRegNum, then only newly added register can add
  // new instructions.
  if (PrevNumRegs > TTIRegNum)
    C.Insns += (C.NumRegs - PrevNumRegs);
  else
    C.Insns += (C.NumRegs - TTIRegNum);
}

// If ICmpZero formula ends with not 0, it could not be replaced by
// just add or sub. We'll need to compare final result of AddRec.
// That means we'll need an additional instruction. But if the target can
// macro-fuse a compare with a branch, don't count this extra instruction.
// For -10 + {0, +, 1}:
// i = i + 1;
// cmp i, 10
//
// For {-10, +, 1}:
// i = i + 1;
if (LU.Kind == LSRUse::ICmpZero && !F.hasZeroEnd() &&
    !TTI->canMacroFuseCmp())
  C.Insns++;
// Each new AddRec adds 1 instruction to calculation.
C.Insns += (C.AddRecCost - PrevAddRecCost);

// BaseAdds adds instructions for unfolded registers.
if (LU.Kind != LSRUse::ICmpZero)
  C.Insns += C.NumBaseAdds - PrevNumBaseAdds;
assert(isValid() && "invalid cost")(static_cast <bool> (isValid() && "invalid cost"
) ? void (0) : __assert_fail ("isValid() && \"invalid cost\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 1467, __extension__
 __PRETTY_FUNCTION__));
1468}

1470/// Set this cost to a losing value.
1471void Cost::Lose() {
C.Insns = std::numeric_limits<unsigned>::max();
C.NumRegs = std::numeric_limits<unsigned>::max();
C.AddRecCost = std::numeric_limits<unsigned>::max();
C.NumIVMuls = std::numeric_limits<unsigned>::max();
C.NumBaseAdds = std::numeric_limits<unsigned>::max();
C.ImmCost = std::numeric_limits<unsigned>::max();
C.SetupCost = std::numeric_limits<unsigned>::max();
C.ScaleCost = std::numeric_limits<unsigned>::max();
1480}

1482/// Choose the lower cost.
1483bool Cost::isLess(const Cost &Other) const {
if (InsnsCost.getNumOccurrences() > 0 && InsnsCost &&
    C.Insns != Other.C.Insns)
  return C.Insns < Other.C.Insns;
return TTI->isLSRCostLess(C, Other.C);
1488}

1490#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1491void Cost::print(raw_ostream &OS) const {
if (InsnsCost)
  OS << C.Insns << " instruction" << (C.Insns == 1 ? " " : "s ");
OS << C.NumRegs << " reg" << (C.NumRegs == 1 ? "" : "s");
if (C.AddRecCost != 0)
  OS << ", with addrec cost " << C.AddRecCost;
if (C.NumIVMuls != 0)
  OS << ", plus " << C.NumIVMuls << " IV mul"
     << (C.NumIVMuls == 1 ? "" : "s");
if (C.NumBaseAdds != 0)
  OS << ", plus " << C.NumBaseAdds << " base add"
     << (C.NumBaseAdds == 1 ? "" : "s");
if (C.ScaleCost != 0)
  OS << ", plus " << C.ScaleCost << " scale cost";
if (C.ImmCost != 0)
  OS << ", plus " << C.ImmCost << " imm cost";
if (C.SetupCost != 0)
  OS << ", plus " << C.SetupCost << " setup cost";
1509}

1511LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void Cost::dump() const {
print(errs()); errs() << '\n';
1513}
1514#endif

1516/// Test whether this fixup always uses its value outside of the given loop.
1517bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
// PHI nodes use their value in their incoming blocks.
if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
    if (PN->getIncomingValue(i) == OperandValToReplace &&
        L->contains(PN->getIncomingBlock(i)))
      return false;
  return true;
}

return !L->contains(UserInst);
1528}

1530#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1531void LSRFixup::print(raw_ostream &OS) const {
OS << "UserInst=";
// Store is common and interesting enough to be worth special-casing.
if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
  OS << "store ";
  Store->getOperand(0)->printAsOperand(OS, /*PrintType=*/false);
} else if (UserInst->getType()->isVoidTy())
  OS << UserInst->getOpcodeName();
else
  UserInst->printAsOperand(OS, /*PrintType=*/false);

OS << ", OperandValToReplace=";
OperandValToReplace->printAsOperand(OS, /*PrintType=*/false);

for (const Loop *PIL : PostIncLoops) {
  OS << ", PostIncLoop=";
  PIL->getHeader()->printAsOperand(OS, /*PrintType=*/false);
}

if (Offset != 0)
  OS << ", Offset=" << Offset;
1552}

1554LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void LSRFixup::dump() const {
print(errs()); errs() << '\n';
1556}
1557#endif

1559/// Test whether this use as a formula which has the same registers as the given
1560/// formula.
1561bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
SmallVector<const SCEV *, 4> Key = F.BaseRegs;
if (F.ScaledReg) Key.push_back(F.ScaledReg);
// Unstable sort by host order ok, because this is only used for uniquifying.
llvm::sort(Key);
return Uniquifier.count(Key);
1567}

1569/// The function returns a probability of selecting formula without Reg.
1570float LSRUse::getNotSelectedProbability(const SCEV *Reg) const {
unsigned FNum = 0;
for (const Formula &F : Formulae)
  if (F.referencesReg(Reg))
    FNum++;
return ((float)(Formulae.size() - FNum)) / Formulae.size();
1576}

1578/// If the given formula has not yet been inserted, add it to the list, and
1579/// return true. Return false otherwise.  The formula must be in canonical form.
1580bool LSRUse::InsertFormula(const Formula &F, const Loop &L) {
assert(F.isCanonical(L) && "Invalid canonical representation")(static_cast <bool> (F.isCanonical(L) && "Invalid canonical representation"
) ? void (0) : __assert_fail ("F.isCanonical(L) && \"Invalid canonical representation\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 1581, __extension__
 __PRETTY_FUNCTION__));

if (!Formulae.empty() && RigidFormula)
  return false;

SmallVector<const SCEV *, 4> Key = F.BaseRegs;
if (F.ScaledReg) Key.push_back(F.ScaledReg);
// Unstable sort by host order ok, because this is only used for uniquifying.
llvm::sort(Key);

if (!Uniquifier.insert(Key).second)
  return false;

// Using a register to hold the value of 0 is not profitable.
assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&(static_cast <bool> ((!F.ScaledReg || !F.ScaledReg->
isZero()) && "Zero allocated in a scaled register!") ?
 void (0) : __assert_fail ("(!F.ScaledReg || !F.ScaledReg->isZero()) && \"Zero allocated in a scaled register!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 1596, __extension__
 __PRETTY_FUNCTION__))
       "Zero allocated in a scaled register!")(static_cast <bool> ((!F.ScaledReg || !F.ScaledReg->
isZero()) && "Zero allocated in a scaled register!") ?
 void (0) : __assert_fail ("(!F.ScaledReg || !F.ScaledReg->isZero()) && \"Zero allocated in a scaled register!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 1596, __extension__
 __PRETTY_FUNCTION__));
1597#ifndef NDEBUG
for (const SCEV *BaseReg : F.BaseRegs)
  assert(!BaseReg->isZero() && "Zero allocated in a base register!")(static_cast <bool> (!BaseReg->isZero() && "Zero allocated in a base register!"
) ? void (0) : __assert_fail ("!BaseReg->isZero() && \"Zero allocated in a base register!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 1599, __extension__
 __PRETTY_FUNCTION__));
1600#endif

// Add the formula to the list.
Formulae.push_back(F);

// Record registers now being used by this use.
Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
if (F.ScaledReg)
  Regs.insert(F.ScaledReg);

return true;
1611}

1613/// Remove the given formula from this use's list.
1614void LSRUse::DeleteFormula(Formula &F) {
if (&F != &Formulae.back())
  std::swap(F, Formulae.back());
Formulae.pop_back();
1618}

1620/// Recompute the Regs field, and update RegUses.
1621void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
// Now that we've filtered out some formulae, recompute the Regs set.
SmallPtrSet<const SCEV *, 4> OldRegs = std::move(Regs);
Regs.clear();
for (const Formula &F : Formulae) {
  if (F.ScaledReg) Regs.insert(F.ScaledReg);
  Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
}

// Update the RegTracker.
for (const SCEV *S : OldRegs)
  if (!Regs.count(S))
    RegUses.dropRegister(S, LUIdx);
1634}

1636#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1637void LSRUse::print(raw_ostream &OS) const {
OS << "LSR Use: Kind=";
switch (Kind) {
case Basic:    OS << "Basic"; break;
case Special:  OS << "Special"; break;
case ICmpZero: OS << "ICmpZero"; break;
case Address:
  OS << "Address of ";
  if (AccessTy.MemTy->isPointerTy())
    OS << "pointer"; // the full pointer type could be really verbose
  else {
    OS << *AccessTy.MemTy;
  }

  OS << " in addrspace(" << AccessTy.AddrSpace << ')';
}

OS << ", Offsets={";
bool NeedComma = false;
for (const LSRFixup &Fixup : Fixups) {
  if (NeedComma) OS << ',';
  OS << Fixup.Offset;
  NeedComma = true;
}
OS << '}';

if (AllFixupsOutsideLoop)
  OS << ", all-fixups-outside-loop";

if (WidestFixupType)
  OS << ", widest fixup type: " << *WidestFixupType;
1668}

1670LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void LSRUse::dump() const {
print(errs()); errs() << '\n';
1672}
1673#endif

1675static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
                               LSRUse::KindType Kind, MemAccessTy AccessTy,
                               GlobalValue *BaseGV, int64_t BaseOffset,
                               bool HasBaseReg, int64_t Scale,
                               Instruction *Fixup/*= nullptr*/) {
switch (Kind) {
case LSRUse::Address:
  return TTI.isLegalAddressingMode(AccessTy.MemTy, BaseGV, BaseOffset,
                                   HasBaseReg, Scale, AccessTy.AddrSpace, Fixup);

case LSRUse::ICmpZero:
  // There's not even a target hook for querying whether it would be legal to
  // fold a GV into an ICmp.
  if (BaseGV)
    return false;

  // ICmp only has two operands; don't allow more than two non-trivial parts.
  if (Scale != 0 && HasBaseReg && BaseOffset != 0)
    return false;

  // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
  // putting the scaled register in the other operand of the icmp.
  if (Scale != 0 && Scale != -1)
    return false;

  // If we have low-level target information, ask the target if it can fold an
  // integer immediate on an icmp.
  if (BaseOffset != 0) {
    // We have one of:
    // ICmpZero     BaseReg + BaseOffset => ICmp BaseReg, -BaseOffset
    // ICmpZero -1*ScaleReg + BaseOffset => ICmp ScaleReg, BaseOffset
    // Offs is the ICmp immediate.
    if (Scale == 0)
      // The cast does the right thing with
      // std::numeric_limits<int64_t>::min().
      BaseOffset = -(uint64_t)BaseOffset;
    return TTI.isLegalICmpImmediate(BaseOffset);
  }

  // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg
  return true;

case LSRUse::Basic:
  // Only handle single-register values.
  return !BaseGV && Scale == 0 && BaseOffset == 0;

case LSRUse::Special:
  // Special case Basic to handle -1 scales.
  return !BaseGV && (Scale == 0 || Scale == -1) && BaseOffset == 0;
}

llvm_unreachable("Invalid LSRUse Kind!")::llvm::llvm_unreachable_internal("Invalid LSRUse Kind!", "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1726);
1727}

1729static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
                               int64_t MinOffset, int64_t MaxOffset,
                               LSRUse::KindType Kind, MemAccessTy AccessTy,
                               GlobalValue *BaseGV, int64_t BaseOffset,
                               bool HasBaseReg, int64_t Scale) {
// Check for overflow.
if (((int64_t)((uint64_t)BaseOffset + MinOffset) > BaseOffset) !=
    (MinOffset > 0))
  return false;
MinOffset = (uint64_t)BaseOffset + MinOffset;
if (((int64_t)((uint64_t)BaseOffset + MaxOffset) > BaseOffset) !=
    (MaxOffset > 0))
  return false;
MaxOffset = (uint64_t)BaseOffset + MaxOffset;

return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MinOffset,
                            HasBaseReg, Scale) &&
       isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MaxOffset,
                            HasBaseReg, Scale);
1748}

1750static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
                               int64_t MinOffset, int64_t MaxOffset,
                               LSRUse::KindType Kind, MemAccessTy AccessTy,
                               const Formula &F, const Loop &L) {
// For the purpose of isAMCompletelyFolded either having a canonical formula
// or a scale not equal to zero is correct.
// Problems may arise from non canonical formulae having a scale == 0.
// Strictly speaking it would best to just rely on canonical formulae.
// However, when we generate the scaled formulae, we first check that the
// scaling factor is profitable before computing the actual ScaledReg for
// compile time sake.
assert((F.isCanonical(L) || F.Scale != 0))(static_cast <bool> ((F.isCanonical(L) || F.Scale != 0)
) ? void (0) : __assert_fail ("(F.isCanonical(L) || F.Scale != 0)"
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 1761, __extension__
 __PRETTY_FUNCTION__));
return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
                            F.BaseGV, F.BaseOffset, F.HasBaseReg, F.Scale);
1764}

1766/// Test whether we know how to expand the current formula.
1767static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
                     int64_t MaxOffset, LSRUse::KindType Kind,
                     MemAccessTy AccessTy, GlobalValue *BaseGV,
                     int64_t BaseOffset, bool HasBaseReg, int64_t Scale) {
// We know how to expand completely foldable formulae.
return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
                            BaseOffset, HasBaseReg, Scale) ||
       // Or formulae that use a base register produced by a sum of base
       // registers.
       (Scale == 1 &&
        isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
                             BaseGV, BaseOffset, true, 0));
1779}

1781static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
                     int64_t MaxOffset, LSRUse::KindType Kind,
                     MemAccessTy AccessTy, const Formula &F) {
return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, F.BaseGV,
                  F.BaseOffset, F.HasBaseReg, F.Scale);
1786}

1788static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
                               const LSRUse &LU, const Formula &F) {
// Target may want to look at the user instructions.
if (LU.Kind == LSRUse::Address && TTI.LSRWithInstrQueries()) {
  for (const LSRFixup &Fixup : LU.Fixups)
    if (!isAMCompletelyFolded(TTI, LSRUse::Address, LU.AccessTy, F.BaseGV,
                              (F.BaseOffset + Fixup.Offset), F.HasBaseReg,
                              F.Scale, Fixup.UserInst))
      return false;
  return true;
}

return isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
                            LU.AccessTy, F.BaseGV, F.BaseOffset, F.HasBaseReg,
                            F.Scale);
1803}

1805static InstructionCost getScalingFactorCost(const TargetTransformInfo &TTI,
                                          const LSRUse &LU, const Formula &F,
                                          const Loop &L) {
if (!F.Scale)
  return 0;

// If the use is not completely folded in that instruction, we will have to
// pay an extra cost only for scale != 1.
if (!isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
                          LU.AccessTy, F, L))
  return F.Scale != 1;

switch (LU.Kind) {
case LSRUse::Address: {
  // Check the scaling factor cost with both the min and max offsets.
  InstructionCost ScaleCostMinOffset = TTI.getScalingFactorCost(
      LU.AccessTy.MemTy, F.BaseGV, F.BaseOffset + LU.MinOffset, F.HasBaseReg,
      F.Scale, LU.AccessTy.AddrSpace);
  InstructionCost ScaleCostMaxOffset = TTI.getScalingFactorCost(
      LU.AccessTy.MemTy, F.BaseGV, F.BaseOffset + LU.MaxOffset, F.HasBaseReg,
      F.Scale, LU.AccessTy.AddrSpace);

  assert(ScaleCostMinOffset.isValid() && ScaleCostMaxOffset.isValid() &&(static_cast <bool> (ScaleCostMinOffset.isValid() &&
 ScaleCostMaxOffset.isValid() && "Legal addressing mode has an illegal cost!"
) ? void (0) : __assert_fail ("ScaleCostMinOffset.isValid() && ScaleCostMaxOffset.isValid() && \"Legal addressing mode has an illegal cost!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 1828, __extension__
 __PRETTY_FUNCTION__))
         "Legal addressing mode has an illegal cost!")(static_cast <bool> (ScaleCostMinOffset.isValid() &&
 ScaleCostMaxOffset.isValid() && "Legal addressing mode has an illegal cost!"
) ? void (0) : __assert_fail ("ScaleCostMinOffset.isValid() && ScaleCostMaxOffset.isValid() && \"Legal addressing mode has an illegal cost!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 1828, __extension__
 __PRETTY_FUNCTION__));
  return std::max(ScaleCostMinOffset, ScaleCostMaxOffset);
}
case LSRUse::ICmpZero:
case LSRUse::Basic:
case LSRUse::Special:
  // The use is completely folded, i.e., everything is folded into the
  // instruction.
  return 0;
}

llvm_unreachable("Invalid LSRUse Kind!")::llvm::llvm_unreachable_internal("Invalid LSRUse Kind!", "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1839);
1840}

1842static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
                           LSRUse::KindType Kind, MemAccessTy AccessTy,
                           GlobalValue *BaseGV, int64_t BaseOffset,
                           bool HasBaseReg) {
// Fast-path: zero is always foldable.
if (BaseOffset == 0 && !BaseGV) return true;

// Conservatively, create an address with an immediate and a
// base and a scale.
int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;

// Canonicalize a scale of 1 to a base register if the formula doesn't
// already have a base register.
if (!HasBaseReg && Scale == 1) {
  Scale = 0;
  HasBaseReg = true;
}

return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, BaseOffset,
                            HasBaseReg, Scale);
1862}

1864static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
                           ScalarEvolution &SE, int64_t MinOffset,
                           int64_t MaxOffset, LSRUse::KindType Kind,
                           MemAccessTy AccessTy, const SCEV *S,
                           bool HasBaseReg) {
// Fast-path: zero is always foldable.
if (S->isZero()) return true;

// Conservatively, create an address with an immediate and a
// base and a scale.
int64_t BaseOffset = ExtractImmediate(S, SE);
GlobalValue *BaseGV = ExtractSymbol(S, SE);

// If there's anything else involved, it's not foldable.
if (!S->isZero()) return false;

// Fast-path: zero is always foldable.
if (BaseOffset == 0 && !BaseGV) return true;

// Conservatively, create an address with an immediate and a
// base and a scale.
int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;

return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
                            BaseOffset, HasBaseReg, Scale);
1889}

1891namespace {

1893/// An individual increment in a Chain of IV increments.  Relate an IV user to
1894/// an expression that computes the IV it uses from the IV used by the previous
1895/// link in the Chain.
1896///
1897/// For the head of a chain, IncExpr holds the absolute SCEV expression for the
1898/// original IVOperand. The head of the chain's IVOperand is only valid during
1899/// chain collection, before LSR replaces IV users. During chain generation,
1900/// IncExpr can be used to find the new IVOperand that computes the same
1901/// expression.
1902struct IVInc {
Instruction *UserInst;
Value* IVOperand;
const SCEV *IncExpr;

IVInc(Instruction *U, Value *O, const SCEV *E)
    : UserInst(U), IVOperand(O), IncExpr(E) {}
1909};

1911// The list of IV increments in program order.  We typically add the head of a
1912// chain without finding subsequent links.
1913struct IVChain {
SmallVector<IVInc, 1> Incs;
const SCEV *ExprBase = nullptr;

IVChain() = default;
IVChain(const IVInc &Head, const SCEV *Base)
    : Incs(1, Head), ExprBase(Base) {}

using const_iterator = SmallVectorImpl<IVInc>::const_iterator;

// Return the first increment in the chain.
const_iterator begin() const {
  assert(!Incs.empty())(static_cast <bool> (!Incs.empty()) ? void (0) : __assert_fail
 ("!Incs.empty()", "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 1925, __extension__ __PRETTY_FUNCTION__));
  return std::next(Incs.begin());
}
const_iterator end() const {
  return Incs.end();
}

// Returns true if this chain contains any increments.
bool hasIncs() const { return Incs.size() >= 2; }

// Add an IVInc to the end of this chain.
void add(const IVInc &X) { Incs.push_back(X); }

// Returns the last UserInst in the chain.
Instruction *tailUserInst() const { return Incs.back().UserInst; }

// Returns true if IncExpr can be profitably added to this chain.
bool isProfitableIncrement(const SCEV *OperExpr,
                           const SCEV *IncExpr,
                           ScalarEvolution&);
1945};

1947/// Helper for CollectChains to track multiple IV increment uses.  Distinguish
1948/// between FarUsers that definitely cross IV increments and NearUsers that may
1949/// be used between IV increments.
1950struct ChainUsers {
SmallPtrSet<Instruction*, 4> FarUsers;
SmallPtrSet<Instruction*, 4> NearUsers;
1953};

1955/// This class holds state for the main loop strength reduction logic.
1956class LSRInstance {
IVUsers &IU;
ScalarEvolution &SE;
DominatorTree &DT;
LoopInfo &LI;
AssumptionCache &AC;
TargetLibraryInfo &TLI;
const TargetTransformInfo &TTI;
Loop *const L;
MemorySSAUpdater *MSSAU;
TTI::AddressingModeKind AMK;
mutable SCEVExpander Rewriter;
bool Changed = false;

/// This is the insert position that the current loop's induction variable
/// increment should be placed. In simple loops, this is the latch block's
/// terminator. But in more complicated cases, this is a position which will
/// dominate all the in-loop post-increment users.
Instruction *IVIncInsertPos = nullptr;

/// Interesting factors between use strides.
///
/// We explicitly use a SetVector which contains a SmallSet, instead of the
/// default, a SmallDenseSet, because we need to use the full range of
/// int64_ts, and there's currently no good way of doing that with
/// SmallDenseSet.
SetVector<int64_t, SmallVector<int64_t, 8>, SmallSet<int64_t, 8>> Factors;

/// The cost of the current SCEV, the best solution by LSR will be dropped if
/// the solution is not profitable.
Cost BaselineCost;

/// Interesting use types, to facilitate truncation reuse.
SmallSetVector<Type *, 4> Types;

/// The list of interesting uses.
mutable SmallVector<LSRUse, 16> Uses;

/// Track which uses use which register candidates.
RegUseTracker RegUses;

// Limit the number of chains to avoid quadratic behavior. We don't expect to
// have more than a few IV increment chains in a loop. Missing a Chain falls
// back to normal LSR behavior for those uses.
static const unsigned MaxChains = 8;

/// IV users can form a chain of IV increments.
SmallVector<IVChain, MaxChains> IVChainVec;

/// IV users that belong to profitable IVChains.
SmallPtrSet<Use*, MaxChains> IVIncSet;

/// Induction variables that were generated and inserted by the SCEV Expander.
SmallVector<llvm::WeakVH, 2> ScalarEvolutionIVs;

void OptimizeShadowIV();
bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
void OptimizeLoopTermCond();

void ChainInstruction(Instruction *UserInst, Instruction *IVOper,
                      SmallVectorImpl<ChainUsers> &ChainUsersVec);
void FinalizeChain(IVChain &Chain);
void CollectChains();
void GenerateIVChain(const IVChain &Chain,
                     SmallVectorImpl<WeakTrackingVH> &DeadInsts);

void CollectInterestingTypesAndFactors();
void CollectFixupsAndInitialFormulae();

// Support for sharing of LSRUses between LSRFixups.
using UseMapTy = DenseMap<LSRUse::SCEVUseKindPair, size_t>;
UseMapTy UseMap;

bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
                        LSRUse::KindType Kind, MemAccessTy AccessTy);

std::pair<size_t, int64_t> getUse(const SCEV *&Expr, LSRUse::KindType Kind,
                                  MemAccessTy AccessTy);

void DeleteUse(LSRUse &LU, size_t LUIdx);

LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);

void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
void CountRegisters(const Formula &F, size_t LUIdx);
bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);

void CollectLoopInvariantFixupsAndFormulae();

void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
                            unsigned Depth = 0);

void GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
                                const Formula &Base, unsigned Depth,
                                size_t Idx, bool IsScaledReg = false);
void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
void GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
                                 const Formula &Base, size_t Idx,
                                 bool IsScaledReg = false);
void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
void GenerateConstantOffsetsImpl(LSRUse &LU, unsigned LUIdx,
                                 const Formula &Base,
                                 const SmallVectorImpl<int64_t> &Worklist,
                                 size_t Idx, bool IsScaledReg = false);
void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
void GenerateCrossUseConstantOffsets();
void GenerateAllReuseFormulae();

void FilterOutUndesirableDedicatedRegisters();

size_t EstimateSearchSpaceComplexity() const;
void NarrowSearchSpaceByDetectingSupersets();
void NarrowSearchSpaceByCollapsingUnrolledCode();
void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
void NarrowSearchSpaceByFilterFormulaWithSameScaledReg();
void NarrowSearchSpaceByFilterPostInc();
void NarrowSearchSpaceByDeletingCostlyFormulas();
void NarrowSearchSpaceByPickingWinnerRegs();
void NarrowSearchSpaceUsingHeuristics();

void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
                  Cost &SolutionCost,
                  SmallVectorImpl<const Formula *> &Workspace,
                  const Cost &CurCost,
                  const SmallPtrSet<const SCEV *, 16> &CurRegs,
                  DenseSet<const SCEV *> &VisitedRegs) const;
void Solve(SmallVectorImpl<const Formula *> &Solution) const;

BasicBlock::iterator
HoistInsertPosition(BasicBlock::iterator IP,
                    const SmallVectorImpl<Instruction *> &Inputs) const;
BasicBlock::iterator AdjustInsertPositionForExpand(BasicBlock::iterator IP,
                                                   const LSRFixup &LF,
                                                   const LSRUse &LU) const;

Value *Expand(const LSRUse &LU, const LSRFixup &LF, const Formula &F,
              BasicBlock::iterator IP,
              SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
void RewriteForPHI(PHINode *PN, const LSRUse &LU, const LSRFixup &LF,
                   const Formula &F,
                   SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
void Rewrite(const LSRUse &LU, const LSRFixup &LF, const Formula &F,
             SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution);

2106public:
LSRInstance(Loop *L, IVUsers &IU, ScalarEvolution &SE, DominatorTree &DT,
            LoopInfo &LI, const TargetTransformInfo &TTI, AssumptionCache &AC,
            TargetLibraryInfo &TLI, MemorySSAUpdater *MSSAU);

bool getChanged() const { return Changed; }
const SmallVectorImpl<WeakVH> &getScalarEvolutionIVs() const {
  return ScalarEvolutionIVs;
}

void print_factors_and_types(raw_ostream &OS) const;
void print_fixups(raw_ostream &OS) const;
void print_uses(raw_ostream &OS) const;
void print(raw_ostream &OS) const;
void dump() const;
2121};

2123} // end anonymous namespace

2125/// If IV is used in a int-to-float cast inside the loop then try to eliminate
2126/// the cast operation.
2127void LSRInstance::OptimizeShadowIV() {
const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
  return;

for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
     UI != E; /* empty */) {
  IVUsers::const_iterator CandidateUI = UI;
  ++UI;
  Instruction *ShadowUse = CandidateUI->getUser();
  Type *DestTy = nullptr;
  bool IsSigned = false;

  /* If shadow use is a int->float cast then insert a second IV
     to eliminate this cast.

       for (unsigned i = 0; i < n; ++i)
         foo((double)i);

     is transformed into

       double d = 0.0;
       for (unsigned i = 0; i < n; ++i, ++d)
         foo(d);
  */
  if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
    IsSigned = false;
    DestTy = UCast->getDestTy();
  }
  else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
    IsSigned = true;
    DestTy = SCast->getDestTy();
  }
  if (!DestTy) continue;

  // If target does not support DestTy natively then do not apply
  // this transformation.
  if (!TTI.isTypeLegal(DestTy)) continue;

  PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
  if (!PH) continue;
  if (PH->getNumIncomingValues() != 2) continue;

  // If the calculation in integers overflows, the result in FP type will
  // differ. So we only can do this transformation if we are guaranteed to not
  // deal with overflowing values
  const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE.getSCEV(PH));
  if (!AR) continue;
  if (IsSigned && !AR->hasNoSignedWrap()) continue;
  if (!IsSigned && !AR->hasNoUnsignedWrap()) continue;

  Type *SrcTy = PH->getType();
  int Mantissa = DestTy->getFPMantissaWidth();
  if (Mantissa == -1) continue;
  if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
    continue;

  unsigned Entry, Latch;
  if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
    Entry = 0;
    Latch = 1;
  } else {
    Entry = 1;
    Latch = 0;
  }

  ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
  if (!Init) continue;
  Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
                                      (double)Init->getSExtValue() :
                                      (double)Init->getZExtValue());

  BinaryOperator *Incr =
    dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
  if (!Incr) continue;
  if (Incr->getOpcode() != Instruction::Add
      && Incr->getOpcode() != Instruction::Sub)
    continue;

  /* Initialize new IV, double d = 0.0 in above example. */
  ConstantInt *C = nullptr;
  if (Incr->getOperand(0) == PH)
    C = dyn_cast<ConstantInt>(Incr->getOperand(1));
  else if (Incr->getOperand(1) == PH)
    C = dyn_cast<ConstantInt>(Incr->getOperand(0));
  else
    continue;

  if (!C) continue;

  // Ignore negative constants, as the code below doesn't handle them
  // correctly. TODO: Remove this restriction.
  if (!C->getValue().isStrictlyPositive()) continue;

  /* Add new PHINode. */
  PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);

  /* create new increment. '++d' in above example. */
  Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
  BinaryOperator *NewIncr =
    BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
                             Instruction::FAdd : Instruction::FSub,
                           NewPH, CFP, "IV.S.next.", Incr);

  NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
  NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));

  /* Remove cast operation */
  ShadowUse->replaceAllUsesWith(NewPH);
  ShadowUse->eraseFromParent();
  Changed = true;
  break;
}
2240}

2242/// If Cond has an operand that is an expression of an IV, set the IV user and
2243/// stride information and return true, otherwise return false.
2244bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
for (IVStrideUse &U : IU)
  if (U.getUser() == Cond) {
    // NOTE: we could handle setcc instructions with multiple uses here, but
    // InstCombine does it as well for simple uses, it's not clear that it
    // occurs enough in real life to handle.
    CondUse = &U;
    return true;
  }
return false;
2254}

2256/// Rewrite the loop's terminating condition if it uses a max computation.
2257///
2258/// This is a narrow solution to a specific, but acute, problem. For loops
2259/// like this:
2260///
2261///   i = 0;
2262///   do {
2263///     p[i] = 0.0;
2264///   } while (++i < n);
2265///
2266/// the trip count isn't just 'n', because 'n' might not be positive. And
2267/// unfortunately this can come up even for loops where the user didn't use
2268/// a C do-while loop. For example, seemingly well-behaved top-test loops
2269/// will commonly be lowered like this:
2270///
2271///   if (n > 0) {
2272///     i = 0;
2273///     do {
2274///       p[i] = 0.0;
2275///     } while (++i < n);
2276///   }
2277///
2278/// and then it's possible for subsequent optimization to obscure the if
2279/// test in such a way that indvars can't find it.
2280///
2281/// When indvars can't find the if test in loops like this, it creates a
2282/// max expression, which allows it to give the loop a canonical
2283/// induction variable:
2284///
2285///   i = 0;
2286///   max = n < 1 ? 1 : n;
2287///   do {
2288///     p[i] = 0.0;
2289///   } while (++i != max);
2290///
2291/// Canonical induction variables are necessary because the loop passes
2292/// are designed around them. The most obvious example of this is the
2293/// LoopInfo analysis, which doesn't remember trip count values. It
2294/// expects to be able to rediscover the trip count each time it is
2295/// needed, and it does this using a simple analysis that only succeeds if
2296/// the loop has a canonical induction variable.
2297///
2298/// However, when it comes time to generate code, the maximum operation
2299/// can be quite costly, especially if it's inside of an outer loop.
2300///
2301/// This function solves this problem by detecting this type of loop and
2302/// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
2303/// the instructions for the maximum computation.
2304ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
// Check that the loop matches the pattern we're looking for.
if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
    Cond->getPredicate() != CmpInst::ICMP_NE)
  return Cond;

SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
if (!Sel || !Sel->hasOneUse()) return Cond;

const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
  return Cond;
const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);

// Add one to the backedge-taken count to get the trip count.
const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
if (IterationCount != SE.getSCEV(Sel)) return Cond;

// Check for a max calculation that matches the pattern. There's no check
// for ICMP_ULE here because the comparison would be with zero, which
// isn't interesting.
CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
const SCEVNAryExpr *Max = nullptr;
if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
  Pred = ICmpInst::ICMP_SLE;
  Max = S;
} else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
  Pred = ICmpInst::ICMP_SLT;
  Max = S;
} else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
  Pred = ICmpInst::ICMP_ULT;
  Max = U;
} else {
  // No match; bail.
  return Cond;
}

// To handle a max with more than two operands, this optimization would
// require additional checking and setup.
if (Max->getNumOperands() != 2)
  return Cond;

const SCEV *MaxLHS = Max->getOperand(0);
const SCEV *MaxRHS = Max->getOperand(1);

// ScalarEvolution canonicalizes constants to the left. For < and >, look
// for a comparison with 1. For <= and >=, a comparison with zero.
if (!MaxLHS ||
    (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
  return Cond;

// Check the relevant induction variable for conformance to
// the pattern.
const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
if (!AR || !AR->isAffine() ||
    AR->getStart() != One ||
    AR->getStepRecurrence(SE) != One)
  return Cond;

assert(AR->getLoop() == L &&(static_cast <bool> (AR->getLoop() == L && "Loop condition operand is an addrec in a different loop!"
) ? void (0) : __assert_fail ("AR->getLoop() == L && \"Loop condition operand is an addrec in a different loop!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 2365, __extension__
 __PRETTY_FUNCTION__))
       "Loop condition operand is an addrec in a different loop!")(static_cast <bool> (AR->getLoop() == L && "Loop condition operand is an addrec in a different loop!"
) ? void (0) : __assert_fail ("AR->getLoop() == L && \"Loop condition operand is an addrec in a different loop!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 2365, __extension__
 __PRETTY_FUNCTION__));

// Check the right operand of the select, and remember it, as it will
// be used in the new comparison instruction.
Value *NewRHS = nullptr;
if (ICmpInst::isTrueWhenEqual(Pred)) {
  // Look for n+1, and grab n.
  if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
    if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
       if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
         NewRHS = BO->getOperand(0);
  if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
    if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
      if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
        NewRHS = BO->getOperand(0);
  if (!NewRHS)
    return Cond;
} else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
  NewRHS = Sel->getOperand(1);
else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
  NewRHS = Sel->getOperand(2);
else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
  NewRHS = SU->getValue();
else
  // Max doesn't match expected pattern.
  return Cond;

// Determine the new comparison opcode. It may be signed or unsigned,
// and the original comparison may be either equality or inequality.
if (Cond->getPredicate() == CmpInst::ICMP_EQ)
  Pred = CmpInst::getInversePredicate(Pred);

// Ok, everything looks ok to change the condition into an SLT or SGE and
// delete the max calculation.
ICmpInst *NewCond =
  new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");

// Delete the max calculation instructions.
NewCond->setDebugLoc(Cond->getDebugLoc());
Cond->replaceAllUsesWith(NewCond);
CondUse->setUser(NewCond);
Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
Cond->eraseFromParent();
Sel->eraseFromParent();
if (Cmp->use_empty())
  Cmp->eraseFromParent();
return NewCond;
2412}

2414/// Change loop terminating condition to use the postinc iv when possible.
2415void
2416LSRInstance::OptimizeLoopTermCond() {
SmallPtrSet<Instruction *, 4> PostIncs;

// We need a different set of heuristics for rotated and non-rotated loops.
// If a loop is rotated then the latch is also the backedge, so inserting
// post-inc expressions just before the latch is ideal. To reduce live ranges
// it also makes sense to rewrite terminating conditions to use post-inc
// expressions.
//
// If the loop is not rotated then the latch is not a backedge; the latch
// check is done in the loop head. Adding post-inc expressions before the
// latch will cause overlapping live-ranges of pre-inc and post-inc expressions
// in the loop body. In this case we do *not* want to use post-inc expressions
// in the latch check, and we want to insert post-inc expressions before
// the backedge.
BasicBlock *LatchBlock = L->getLoopLatch();
SmallVector<BasicBlock*, 8> ExitingBlocks;
L->getExitingBlocks(ExitingBlocks);
if (!llvm::is_contained(ExitingBlocks, LatchBlock)) {
  // The backedge doesn't exit the loop; treat this as a head-tested loop.
  IVIncInsertPos = LatchBlock->getTerminator();
  return;
}

// Otherwise treat this as a rotated loop.
for (BasicBlock *ExitingBlock : ExitingBlocks) {
  // Get the terminating condition for the loop if possible.  If we
  // can, we want to change it to use a post-incremented version of its
  // induction variable, to allow coalescing the live ranges for the IV into
  // one register value.

  BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
  if (!TermBr)
    continue;
  // FIXME: Overly conservative, termination condition could be an 'or' etc..
  if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
    continue;

  // Search IVUsesByStride to find Cond's IVUse if there is one.
  IVStrideUse *CondUse = nullptr;
  ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
  if (!FindIVUserForCond(Cond, CondUse))
    continue;

  // If the trip count is computed in terms of a max (due to ScalarEvolution
  // being unable to find a sufficient guard, for example), change the loop
  // comparison to use SLT or ULT instead of NE.
  // One consequence of doing this now is that it disrupts the count-down
  // optimization. That's not always a bad thing though, because in such
  // cases it may still be worthwhile to avoid a max.
  Cond = OptimizeMax(Cond, CondUse);

  // If this exiting block dominates the latch block, it may also use
  // the post-inc value if it won't be shared with other uses.
  // Check for dominance.
  if (!DT.dominates(ExitingBlock, LatchBlock))
    continue;

  // Conservatively avoid trying to use the post-inc value in non-latch
  // exits if there may be pre-inc users in intervening blocks.
  if (LatchBlock != ExitingBlock)
    for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
      // Test if the use is reachable from the exiting block. This dominator
      // query is a conservative approximation of reachability.
      if (&*UI != CondUse &&
          !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
        // Conservatively assume there may be reuse if the quotient of their
        // strides could be a legal scale.
        const SCEV *A = IU.getStride(*CondUse, L);
        const SCEV *B = IU.getStride(*UI, L);
        if (!A || !B) continue;
        if (SE.getTypeSizeInBits(A->getType()) !=
            SE.getTypeSizeInBits(B->getType())) {
          if (SE.getTypeSizeInBits(A->getType()) >
              SE.getTypeSizeInBits(B->getType()))
            B = SE.getSignExtendExpr(B, A->getType());
          else
            A = SE.getSignExtendExpr(A, B->getType());
        }
        if (const SCEVConstant *D =
              dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
          const ConstantInt *C = D->getValue();
          // Stride of one or negative one can have reuse with non-addresses.
          if (C->isOne() || C->isMinusOne())
            goto decline_post_inc;
          // Avoid weird situations.
          if (C->getValue().getSignificantBits() >= 64 ||
              C->getValue().isMinSignedValue())
            goto decline_post_inc;
          // Check for possible scaled-address reuse.
          if (isAddressUse(TTI, UI->getUser(), UI->getOperandValToReplace())) {
            MemAccessTy AccessTy = getAccessType(
                TTI, UI->getUser(), UI->getOperandValToReplace());
            int64_t Scale = C->getSExtValue();
            if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
                                          /*BaseOffset=*/0,
                                          /*HasBaseReg=*/false, Scale,
                                          AccessTy.AddrSpace))
              goto decline_post_inc;
            Scale = -Scale;
            if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
                                          /*BaseOffset=*/0,
                                          /*HasBaseReg=*/false, Scale,
                                          AccessTy.AddrSpace))
              goto decline_post_inc;
          }
        }
      }

  LLVM_DEBUG(dbgs() << "  Change loop exiting icmp to use postinc iv: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "  Change loop exiting icmp to use postinc iv: "
 << *Cond << '\n'; } } while (false)
                    << *Cond << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "  Change loop exiting icmp to use postinc iv: "
 << *Cond << '\n'; } } while (false);

  // It's possible for the setcc instruction to be anywhere in the loop, and
  // possible for it to have multiple users.  If it is not immediately before
  // the exiting block branch, move it.
  if (Cond->getNextNonDebugInstruction() != TermBr) {
    if (Cond->hasOneUse()) {
      Cond->moveBefore(TermBr);
    } else {
      // Clone the terminating condition and insert into the loopend.
      ICmpInst *OldCond = Cond;
      Cond = cast<ICmpInst>(Cond->clone());
      Cond->setName(L->getHeader()->getName() + ".termcond");
      Cond->insertInto(ExitingBlock, TermBr->getIterator());

      // Clone the IVUse, as the old use still exists!
      CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
      TermBr->replaceUsesOfWith(OldCond, Cond);
    }
  }

  // If we get to here, we know that we can transform the setcc instruction to
  // use the post-incremented version of the IV, allowing us to coalesce the
  // live ranges for the IV correctly.
  CondUse->transformToPostInc(L);
  Changed = true;

  PostIncs.insert(Cond);
decline_post_inc:;
}

// Determine an insertion point for the loop induction variable increment. It
// must dominate all the post-inc comparisons we just set up, and it must
// dominate the loop latch edge.
IVIncInsertPos = L->getLoopLatch()->getTerminator();
for (Instruction *Inst : PostIncs)
  IVIncInsertPos = DT.findNearestCommonDominator(IVIncInsertPos, Inst);
2563}

2565/// Determine if the given use can accommodate a fixup at the given offset and
2566/// other details. If so, update the use and return true.
2567bool LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset,
                                   bool HasBaseReg, LSRUse::KindType Kind,
                                   MemAccessTy AccessTy) {
int64_t NewMinOffset = LU.MinOffset;
int64_t NewMaxOffset = LU.MaxOffset;
MemAccessTy NewAccessTy = AccessTy;

// Check for a mismatched kind. It's tempting to collapse mismatched kinds to
// something conservative, however this can pessimize in the case that one of
// the uses will have all its uses outside the loop, for example.
if (LU.Kind != Kind)
  return false;

// Check for a mismatched access type, and fall back conservatively as needed.
// TODO: Be less conservative when the type is similar and can use the same
// addressing modes.
if (Kind == LSRUse::Address) {
  if (AccessTy.MemTy != LU.AccessTy.MemTy) {
    NewAccessTy = MemAccessTy::getUnknown(AccessTy.MemTy->getContext(),
                                          AccessTy.AddrSpace);
  }
}

// Conservatively assume HasBaseReg is true for now.
if (NewOffset < LU.MinOffset) {
  if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
                        LU.MaxOffset - NewOffset, HasBaseReg))
    return false;
  NewMinOffset = NewOffset;
} else if (NewOffset > LU.MaxOffset) {
  if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
                        NewOffset - LU.MinOffset, HasBaseReg))
    return false;
  NewMaxOffset = NewOffset;
}

// Update the use.
LU.MinOffset = NewMinOffset;
LU.MaxOffset = NewMaxOffset;
LU.AccessTy = NewAccessTy;
return true;
2608}

2610/// Return an LSRUse index and an offset value for a fixup which needs the given
2611/// expression, with the given kind and optional access type.  Either reuse an
2612/// existing use or create a new one, as needed.
2613std::pair<size_t, int64_t> LSRInstance::getUse(const SCEV *&Expr,
                                             LSRUse::KindType Kind,
                                             MemAccessTy AccessTy) {
const SCEV *Copy = Expr;
int64_t Offset = ExtractImmediate(Expr, SE);

// Basic uses can't accept any offset, for example.
if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ nullptr,
                      Offset, /*HasBaseReg=*/ true)) {
  Expr = Copy;
  Offset = 0;
}

std::pair<UseMapTy::iterator, bool> P =
  UseMap.insert(std::make_pair(LSRUse::SCEVUseKindPair(Expr, Kind), 0));
if (!P.second) {
  // A use already existed with this base.
  size_t LUIdx = P.first->second;
  LSRUse &LU = Uses[LUIdx];
  if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
    // Reuse this use.
    return std::make_pair(LUIdx, Offset);
}

// Create a new use.
size_t LUIdx = Uses.size();
P.first->second = LUIdx;
Uses.push_back(LSRUse(Kind, AccessTy));
LSRUse &LU = Uses[LUIdx];

LU.MinOffset = Offset;
LU.MaxOffset = Offset;
return std::make_pair(LUIdx, Offset);
2646}

2648/// Delete the given use from the Uses list.
2649void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
if (&LU != &Uses.back())
  std::swap(LU, Uses.back());
Uses.pop_back();

// Update RegUses.
RegUses.swapAndDropUse(LUIdx, Uses.size());
2656}

2658/// Look for a use distinct from OrigLU which is has a formula that has the same
2659/// registers as the given formula.
2660LSRUse *
2661LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
                                     const LSRUse &OrigLU) {
// Search all uses for the formula. This could be more clever.
for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
  LSRUse &LU = Uses[LUIdx];
  // Check whether this use is close enough to OrigLU, to see whether it's
  // worthwhile looking through its formulae.
  // Ignore ICmpZero uses because they may contain formulae generated by
  // GenerateICmpZeroScales, in which case adding fixup offsets may
  // be invalid.
  if (&LU != &OrigLU &&
      LU.Kind != LSRUse::ICmpZero &&
      LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
      LU.WidestFixupType == OrigLU.WidestFixupType &&
      LU.HasFormulaWithSameRegs(OrigF)) {
    // Scan through this use's formulae.
    for (const Formula &F : LU.Formulae) {
      // Check to see if this formula has the same registers and symbols
      // as OrigF.
      if (F.BaseRegs == OrigF.BaseRegs &&
          F.ScaledReg == OrigF.ScaledReg &&
          F.BaseGV == OrigF.BaseGV &&
          F.Scale == OrigF.Scale &&
          F.UnfoldedOffset == OrigF.UnfoldedOffset) {
        if (F.BaseOffset == 0)
          return &LU;
        // This is the formula where all the registers and symbols matched;
        // there aren't going to be any others. Since we declined it, we
        // can skip the rest of the formulae and proceed to the next LSRUse.
        break;
      }
    }
  }
}

// Nothing looked good.
return nullptr;
2698}

2700void LSRInstance::CollectInterestingTypesAndFactors() {
SmallSetVector<const SCEV *, 4> Strides;

// Collect interesting types and strides.
SmallVector<const SCEV *, 4> Worklist;
for (const IVStrideUse &U : IU) {
  const SCEV *Expr = IU.getExpr(U);

  // Collect interesting types.
  Types.insert(SE.getEffectiveSCEVType(Expr->getType()));

  // Add strides for mentioned loops.
  Worklist.push_back(Expr);
  do {
    const SCEV *S = Worklist.pop_back_val();
    if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
      if (AR->getLoop() == L)
        Strides.insert(AR->getStepRecurrence(SE));
      Worklist.push_back(AR->getStart());
    } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
      append_range(Worklist, Add->operands());
    }
  } while (!Worklist.empty());
}

// Compute interesting factors from the set of interesting strides.
for (SmallSetVector<const SCEV *, 4>::const_iterator
     I = Strides.begin(), E = Strides.end(); I != E; ++I)
  for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
       std::next(I); NewStrideIter != E; ++NewStrideIter) {
    const SCEV *OldStride = *I;
    const SCEV *NewStride = *NewStrideIter;

    if (SE.getTypeSizeInBits(OldStride->getType()) !=
        SE.getTypeSizeInBits(NewStride->getType())) {
      if (SE.getTypeSizeInBits(OldStride->getType()) >
          SE.getTypeSizeInBits(NewStride->getType()))
        NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
      else
        OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
    }
    if (const SCEVConstant *Factor =
          dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
                                                      SE, true))) {
      if (Factor->getAPInt().getSignificantBits() <= 64 && !Factor->isZero())
        Factors.insert(Factor->getAPInt().getSExtValue());
    } else if (const SCEVConstant *Factor =
                 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
                                                             NewStride,
                                                             SE, true))) {
      if (Factor->getAPInt().getSignificantBits() <= 64 && !Factor->isZero())
        Factors.insert(Factor->getAPInt().getSExtValue());
    }
  }

// If all uses use the same type, don't bother looking for truncation-based
// reuse.
if (Types.size() == 1)
  Types.clear();

LLVM_DEBUG(print_factors_and_types(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { print_factors_and_types(dbgs()); } } while
 (false);
2761}

2763/// Helper for CollectChains that finds an IV operand (computed by an AddRec in
2764/// this loop) within [OI,OE) or returns OE. If IVUsers mapped Instructions to
2765/// IVStrideUses, we could partially skip this.
2766static User::op_iterator
2767findIVOperand(User::op_iterator OI, User::op_iterator OE,
            Loop *L, ScalarEvolution &SE) {
for(; OI != OE; ++OI) {
  if (Instruction *Oper = dyn_cast<Instruction>(*OI)) {
    if (!SE.isSCEVable(Oper->getType()))
      continue;

    if (const SCEVAddRecExpr *AR =
        dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) {
      if (AR->getLoop() == L)
        break;
    }
  }
}
return OI;
2782}

2784/// IVChain logic must consistently peek base TruncInst operands, so wrap it in
2785/// a convenient helper.
2786static Value *getWideOperand(Value *Oper) {
if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper))
  return Trunc->getOperand(0);
return Oper;
2790}

2792/// Return true if we allow an IV chain to include both types.
2793static bool isCompatibleIVType(Value *LVal, Value *RVal) {
Type *LType = LVal->getType();
Type *RType = RVal->getType();
return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy() &&
                            // Different address spaces means (possibly)
                            // different types of the pointer implementation,
                            // e.g. i16 vs i32 so disallow that.
                            (LType->getPointerAddressSpace() ==
                             RType->getPointerAddressSpace()));
2802}

2804/// Return an approximation of this SCEV expression's "base", or NULL for any
2805/// constant. Returning the expression itself is conservative. Returning a
2806/// deeper subexpression is more precise and valid as long as it isn't less
2807/// complex than another subexpression. For expressions involving multiple
2808/// unscaled values, we need to return the pointer-type SCEVUnknown. This avoids
2809/// forming chains across objects, such as: PrevOper==a[i], IVOper==b[i],
2810/// IVInc==b-a.
2811///
2812/// Since SCEVUnknown is the rightmost type, and pointers are the rightmost
2813/// SCEVUnknown, we simply return the rightmost SCEV operand.
2814static const SCEV *getExprBase(const SCEV *S) {
switch (S->getSCEVType()) {
default: // including scUnknown.
  return S;
case scConstant:
case scVScale:
  return nullptr;
case scTruncate:
  return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand());
case scZeroExtend:
  return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand());
case scSignExtend:
  return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand());
case scAddExpr: {
  // Skip over scaled operands (scMulExpr) to follow add operands as long as
  // there's nothing more complex.
  // FIXME: not sure if we want to recognize negation.
  const SCEVAddExpr *Add = cast<SCEVAddExpr>(S);
  for (const SCEV *SubExpr : reverse(Add->operands())) {
    if (SubExpr->getSCEVType() == scAddExpr)
      return getExprBase(SubExpr);

    if (SubExpr->getSCEVType() != scMulExpr)
      return SubExpr;
  }
  return S; // all operands are scaled, be conservative.
}
case scAddRecExpr:
  return getExprBase(cast<SCEVAddRecExpr>(S)->getStart());
}
llvm_unreachable("Unknown SCEV kind!")::llvm::llvm_unreachable_internal("Unknown SCEV kind!", "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp"
, 2844);
2845}

2847/// Return true if the chain increment is profitable to expand into a loop
2848/// invariant value, which may require its own register. A profitable chain
2849/// increment will be an offset relative to the same base. We allow such offsets
2850/// to potentially be used as chain increment as long as it's not obviously
2851/// expensive to expand using real instructions.
2852bool IVChain::isProfitableIncrement(const SCEV *OperExpr,
                                  const SCEV *IncExpr,
                                  ScalarEvolution &SE) {
// Aggressively form chains when -stress-ivchain.
if (StressIVChain)
  return true;

// Do not replace a constant offset from IV head with a nonconstant IV
// increment.
if (!isa<SCEVConstant>(IncExpr)) {
  const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Incs[0].IVOperand));
  if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr)))
    return false;
}

SmallPtrSet<const SCEV*, 8> Processed;
return !isHighCostExpansion(IncExpr, Processed, SE);
2869}

2871/// Return true if the number of registers needed for the chain is estimated to
2872/// be less than the number required for the individual IV users. First prohibit
2873/// any IV users that keep the IV live across increments (the Users set should
2874/// be empty). Next count the number and type of increments in the chain.
2875///
2876/// Chaining IVs can lead to considerable code bloat if ISEL doesn't
2877/// effectively use postinc addressing modes. Only consider it profitable it the
2878/// increments can be computed in fewer registers when chained.
2879///
2880/// TODO: Consider IVInc free if it's already used in another chains.
2881static bool isProfitableChain(IVChain &Chain,
                            SmallPtrSetImpl<Instruction *> &Users,
                            ScalarEvolution &SE,
                            const TargetTransformInfo &TTI) {
if (StressIVChain)
  return true;

if (!Chain.hasIncs())
  return false;

if (!Users.empty()) {
  LLVM_DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " users:\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Chain: " << *Chain.
Incs[0].UserInst << " users:\n"; for (Instruction *Inst
 : Users) { dbgs() << "  " << *Inst << "\n"
; }; } } while (false)
             for (Instruction *Instdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Chain: " << *Chain.
Incs[0].UserInst << " users:\n"; for (Instruction *Inst
 : Users) { dbgs() << "  " << *Inst << "\n"
; }; } } while (false)
                  : Users) { dbgs() << "  " << *Inst << "\n"; })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Chain: " << *Chain.
Incs[0].UserInst << " users:\n"; for (Instruction *Inst
 : Users) { dbgs() << "  " << *Inst << "\n"
; }; } } while (false);
  return false;
}
assert(!Chain.Incs.empty() && "empty IV chains are not allowed")(static_cast <bool> (!Chain.Incs.empty() && "empty IV chains are not allowed"
) ? void (0) : __assert_fail ("!Chain.Incs.empty() && \"empty IV chains are not allowed\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 2897, __extension__
 __PRETTY_FUNCTION__));

// The chain itself may require a register, so intialize cost to 1.
int cost = 1;

// A complete chain likely eliminates the need for keeping the original IV in
// a register. LSR does not currently know how to form a complete chain unless
// the header phi already exists.
if (isa<PHINode>(Chain.tailUserInst())
    && SE.getSCEV(Chain.tailUserInst()) == Chain.Incs[0].IncExpr) {
  --cost;
}
const SCEV *LastIncExpr = nullptr;
unsigned NumConstIncrements = 0;
unsigned NumVarIncrements = 0;
unsigned NumReusedIncrements = 0;

if (TTI.isProfitableLSRChainElement(Chain.Incs[0].UserInst))
  return true;

for (const IVInc &Inc : Chain) {
  if (TTI.isProfitableLSRChainElement(Inc.UserInst))
    return true;
  if (Inc.IncExpr->isZero())
    continue;

  // Incrementing by zero or some constant is neutral. We assume constants can
  // be folded into an addressing mode or an add's immediate operand.
  if (isa<SCEVConstant>(Inc.IncExpr)) {
    ++NumConstIncrements;
    continue;
  }

  if (Inc.IncExpr == LastIncExpr)
    ++NumReusedIncrements;
  else
    ++NumVarIncrements;

  LastIncExpr = Inc.IncExpr;
}
// An IV chain with a single increment is handled by LSR's postinc
// uses. However, a chain with multiple increments requires keeping the IV's
// value live longer than it needs to be if chained.
if (NumConstIncrements > 1)
  --cost;

// Materializing increment expressions in the preheader that didn't exist in
// the original code may cost a register. For example, sign-extended array
// indices can produce ridiculous increments like this:
// IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64)))
cost += NumVarIncrements;

// Reusing variable increments likely saves a register to hold the multiple of
// the stride.
cost -= NumReusedIncrements;

LLVM_DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " Cost: " << costdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Chain: " << *Chain.
Incs[0].UserInst << " Cost: " << cost << "\n"
; } } while (false)
                  << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Chain: " << *Chain.
Incs[0].UserInst << " Cost: " << cost << "\n"
; } } while (false);

return cost < 0;
2957}

2959/// Add this IV user to an existing chain or make it the head of a new chain.
2960void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper,
                                 SmallVectorImpl<ChainUsers> &ChainUsersVec) {
// When IVs are used as types of varying widths, they are generally converted
// to a wider type with some uses remaining narrow under a (free) trunc.
Value *const NextIV = getWideOperand(IVOper);
const SCEV *const OperExpr = SE.getSCEV(NextIV);
const SCEV *const OperExprBase = getExprBase(OperExpr);

// Visit all existing chains. Check if its IVOper can be computed as a
// profitable loop invariant increment from the last link in the Chain.
unsigned ChainIdx = 0, NChains = IVChainVec.size();
const SCEV *LastIncExpr = nullptr;
for (; ChainIdx < NChains; ++ChainIdx) {
  IVChain &Chain = IVChainVec[ChainIdx];

  // Prune the solution space aggressively by checking that both IV operands
  // are expressions that operate on the same unscaled SCEVUnknown. This
  // "base" will be canceled by the subsequent getMinusSCEV call. Checking
  // first avoids creating extra SCEV expressions.
  if (!StressIVChain && Chain.ExprBase != OperExprBase)
    continue;

  Value *PrevIV = getWideOperand(Chain.Incs.back().IVOperand);
  if (!isCompatibleIVType(PrevIV, NextIV))
    continue;

  // A phi node terminates a chain.
  if (isa<PHINode>(UserInst) && isa<PHINode>(Chain.tailUserInst()))
    continue;

  // The increment must be loop-invariant so it can be kept in a register.
  const SCEV *PrevExpr = SE.getSCEV(PrevIV);
  const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr);
  if (isa<SCEVCouldNotCompute>(IncExpr) || !SE.isLoopInvariant(IncExpr, L))
    continue;

  if (Chain.isProfitableIncrement(OperExpr, IncExpr, SE)) {
    LastIncExpr = IncExpr;
    break;
  }
}
// If we haven't found a chain, create a new one, unless we hit the max. Don't
// bother for phi nodes, because they must be last in the chain.
if (ChainIdx == NChains) {
  if (isa<PHINode>(UserInst))
    return;
  if (NChains >= MaxChains && !StressIVChain) {
    LLVM_DEBUG(dbgs() << "IV Chain Limit\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "IV Chain Limit\n"; } } while
 (false);
    return;
  }
  LastIncExpr = OperExpr;
  // IVUsers may have skipped over sign/zero extensions. We don't currently
  // attempt to form chains involving extensions unless they can be hoisted
  // into this loop's AddRec.
  if (!isa<SCEVAddRecExpr>(LastIncExpr))
    return;
  ++NChains;
  IVChainVec.push_back(IVChain(IVInc(UserInst, IVOper, LastIncExpr),
                               OperExprBase));
  ChainUsersVec.resize(NChains);
  LLVM_DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Head: (" << *UserInstdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "IV Chain#" << ChainIdx
 << " Head: (" << *UserInst << ") IV=" <<
 *LastIncExpr << "\n"; } } while (false)
                    << ") IV=" << *LastIncExpr << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "IV Chain#" << ChainIdx
 << " Head: (" << *UserInst << ") IV=" <<
 *LastIncExpr << "\n"; } } while (false);
} else {
  LLVM_DEBUG(dbgs() << "IV Chain#" << ChainIdx << "  Inc: (" << *UserInstdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "IV Chain#" << ChainIdx
 << "  Inc: (" << *UserInst << ") IV+" <<
 *LastIncExpr << "\n"; } } while (false)
                    << ") IV+" << *LastIncExpr << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "IV Chain#" << ChainIdx
 << "  Inc: (" << *UserInst << ") IV+" <<
 *LastIncExpr << "\n"; } } while (false);
  // Add this IV user to the end of the chain.
  IVChainVec[ChainIdx].add(IVInc(UserInst, IVOper, LastIncExpr));
}
IVChain &Chain = IVChainVec[ChainIdx];

SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers;
// This chain's NearUsers become FarUsers.
if (!LastIncExpr->isZero()) {
  ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(),
                                          NearUsers.end());
  NearUsers.clear();
}

// All other uses of IVOperand become near uses of the chain.
// We currently ignore intermediate values within SCEV expressions, assuming
// they will eventually be used be the current chain, or can be computed
// from one of the chain increments. To be more precise we could
// transitively follow its user and only add leaf IV users to the set.
for (User *U : IVOper->users()) {
  Instruction *OtherUse = dyn_cast<Instruction>(U);
  if (!OtherUse)
    continue;
  // Uses in the chain will no longer be uses if the chain is formed.
  // Include the head of the chain in this iteration (not Chain.begin()).
  IVChain::const_iterator IncIter = Chain.Incs.begin();
  IVChain::const_iterator IncEnd = Chain.Incs.end();
  for( ; IncIter != IncEnd; ++IncIter) {
    if (IncIter->UserInst == OtherUse)
      break;
  }
  if (IncIter != IncEnd)
    continue;

  if (SE.isSCEVable(OtherUse->getType())
      && !isa<SCEVUnknown>(SE.getSCEV(OtherUse))
      && IU.isIVUserOrOperand(OtherUse)) {
    continue;
  }
  NearUsers.insert(OtherUse);
}

// Since this user is part of the chain, it's no longer considered a use
// of the chain.
ChainUsersVec[ChainIdx].FarUsers.erase(UserInst);
3069}

3071/// Populate the vector of Chains.
3072///
3073/// This decreases ILP at the architecture level. Targets with ample registers,
3074/// multiple memory ports, and no register renaming probably don't want
3075/// this. However, such targets should probably disable LSR altogether.
3076///
3077/// The job of LSR is to make a reasonable choice of induction variables across
3078/// the loop. Subsequent passes can easily "unchain" computation exposing more
3079/// ILP *within the loop* if the target wants it.
3080///
3081/// Finding the best IV chain is potentially a scheduling problem. Since LSR
3082/// will not reorder memory operations, it will recognize this as a chain, but
3083/// will generate redundant IV increments. Ideally this would be corrected later
3084/// by a smart scheduler:
3085///        = A[i]
3086///        = A[i+x]
3087/// A[i]   =
3088/// A[i+x] =
3089///
3090/// TODO: Walk the entire domtree within this loop, not just the path to the
3091/// loop latch. This will discover chains on side paths, but requires
3092/// maintaining multiple copies of the Chains state.
3093void LSRInstance::CollectChains() {
LLVM_DEBUG(dbgs() << "Collecting IV Chains.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Collecting IV Chains.\n";
 } } while (false);
SmallVector<ChainUsers, 8> ChainUsersVec;

SmallVector<BasicBlock *,8> LatchPath;
BasicBlock *LoopHeader = L->getHeader();
for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch());
     Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) {
  LatchPath.push_back(Rung->getBlock());
}
LatchPath.push_back(LoopHeader);

// Walk the instruction stream from the loop header to the loop latch.
for (BasicBlock *BB : reverse(LatchPath)) {
  for (Instruction &I : *BB) {
    // Skip instructions that weren't seen by IVUsers analysis.
    if (isa<PHINode>(I) || !IU.isIVUserOrOperand(&I))
      continue;

    // Ignore users that are part of a SCEV expression. This way we only
    // consider leaf IV Users. This effectively rediscovers a portion of
    // IVUsers analysis but in program order this time.
    if (SE.isSCEVable(I.getType()) && !isa<SCEVUnknown>(SE.getSCEV(&I)))
        continue;

    // Remove this instruction from any NearUsers set it may be in.
    for (unsigned ChainIdx = 0, NChains = IVChainVec.size();
         ChainIdx < NChains; ++ChainIdx) {
      ChainUsersVec[ChainIdx].NearUsers.erase(&I);
    }
    // Search for operands that can be chained.
    SmallPtrSet<Instruction*, 4> UniqueOperands;
    User::op_iterator IVOpEnd = I.op_end();
    User::op_iterator IVOpIter = findIVOperand(I.op_begin(), IVOpEnd, L, SE);
    while (IVOpIter != IVOpEnd) {
      Instruction *IVOpInst = cast<Instruction>(*IVOpIter);
      if (UniqueOperands.insert(IVOpInst).second)
        ChainInstruction(&I, IVOpInst, ChainUsersVec);
      IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
    }
  } // Continue walking down the instructions.
} // Continue walking down the domtree.
// Visit phi backedges to determine if the chain can generate the IV postinc.
for (PHINode &PN : L->getHeader()->phis()) {
  if (!SE.isSCEVable(PN.getType()))
    continue;

  Instruction *IncV =
      dyn_cast<Instruction>(PN.getIncomingValueForBlock(L->getLoopLatch()));
  if (IncV)
    ChainInstruction(&PN, IncV, ChainUsersVec);
}
// Remove any unprofitable chains.
unsigned ChainIdx = 0;
for (unsigned UsersIdx = 0, NChains = IVChainVec.size();
     UsersIdx < NChains; ++UsersIdx) {
  if (!isProfitableChain(IVChainVec[UsersIdx],
                         ChainUsersVec[UsersIdx].FarUsers, SE, TTI))
    continue;
  // Preserve the chain at UsesIdx.
  if (ChainIdx != UsersIdx)
    IVChainVec[ChainIdx] = IVChainVec[UsersIdx];
  FinalizeChain(IVChainVec[ChainIdx]);
  ++ChainIdx;
}
IVChainVec.resize(ChainIdx);
3159}

3161void LSRInstance::FinalizeChain(IVChain &Chain) {
assert(!Chain.Incs.empty() && "empty IV chains are not allowed")(static_cast <bool> (!Chain.Incs.empty() && "empty IV chains are not allowed"
) ? void (0) : __assert_fail ("!Chain.Incs.empty() && \"empty IV chains are not allowed\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 3162, __extension__
 __PRETTY_FUNCTION__));
LLVM_DEBUG(dbgs() << "Final Chain: " << *Chain.Incs[0].UserInst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Final Chain: " << *
Chain.Incs[0].UserInst << "\n"; } } while (false);

for (const IVInc &Inc : Chain) {
  LLVM_DEBUG(dbgs() << "        Inc: " << *Inc.UserInst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "        Inc: " << *
Inc.UserInst << "\n"; } } while (false);
  auto UseI = find(Inc.UserInst->operands(), Inc.IVOperand);
  assert(UseI != Inc.UserInst->op_end() && "cannot find IV operand")(static_cast <bool> (UseI != Inc.UserInst->op_end() &&
 "cannot find IV operand") ? void (0) : __assert_fail ("UseI != Inc.UserInst->op_end() && \"cannot find IV operand\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 3168, __extension__
 __PRETTY_FUNCTION__));
  IVIncSet.insert(UseI);
}
3171}

3173/// Return true if the IVInc can be folded into an addressing mode.
3174static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst,
                           Value *Operand, const TargetTransformInfo &TTI) {
const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr);
if (!IncConst || !isAddressUse(TTI, UserInst, Operand))
  return false;

if (IncConst->getAPInt().getSignificantBits() > 64)
  return false;

MemAccessTy AccessTy = getAccessType(TTI, UserInst, Operand);
int64_t IncOffset = IncConst->getValue()->getSExtValue();
if (!isAlwaysFoldable(TTI, LSRUse::Address, AccessTy, /*BaseGV=*/nullptr,
                      IncOffset, /*HasBaseReg=*/false))
  return false;

return true;
3190}

3192/// Generate an add or subtract for each IVInc in a chain to materialize the IV
3193/// user's operand from the previous IV user's operand.
3194void LSRInstance::GenerateIVChain(const IVChain &Chain,
                                SmallVectorImpl<WeakTrackingVH> &DeadInsts) {
// Find the new IVOperand for the head of the chain. It may have been replaced
// by LSR.
const IVInc &Head = Chain.Incs[0];
User::op_iterator IVOpEnd = Head.UserInst->op_end();
// findIVOperand returns IVOpEnd if it can no longer find a valid IV user.
User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(),
                                           IVOpEnd, L, SE);
Value *IVSrc = nullptr;
while (IVOpIter != IVOpEnd) {
  IVSrc = getWideOperand(*IVOpIter);

  // If this operand computes the expression that the chain needs, we may use
  // it. (Check this after setting IVSrc which is used below.)
  //
  // Note that if Head.IncExpr is wider than IVSrc, then this phi is too
  // narrow for the chain, so we can no longer use it. We do allow using a
  // wider phi, assuming the LSR checked for free truncation. In that case we
  // should already have a truncate on this operand such that
  // getSCEV(IVSrc) == IncExpr.
  if (SE.getSCEV(*IVOpIter) == Head.IncExpr
      || SE.getSCEV(IVSrc) == Head.IncExpr) {
    break;
  }
  IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
}
if (IVOpIter == IVOpEnd) {
  // Gracefully give up on this chain.
  LLVM_DEBUG(dbgs() << "Concealed chain head: " << *Head.UserInst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Concealed chain head: " <<
 *Head.UserInst << "\n"; } } while (false);
  return;
}
assert(IVSrc && "Failed to find IV chain source")(static_cast <bool> (IVSrc && "Failed to find IV chain source"
) ? void (0) : __assert_fail ("IVSrc && \"Failed to find IV chain source\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 3226, __extension__
 __PRETTY_FUNCTION__));

LLVM_DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Generate chain at: " <<
 *IVSrc << "\n"; } } while (false);
Type *IVTy = IVSrc->getType();
Type *IntTy = SE.getEffectiveSCEVType(IVTy);
const SCEV *LeftOverExpr = nullptr;
for (const IVInc &Inc : Chain) {
  Instruction *InsertPt = Inc.UserInst;
  if (isa<PHINode>(InsertPt))
    InsertPt = L->getLoopLatch()->getTerminator();

  // IVOper will replace the current IV User's operand. IVSrc is the IV
  // value currently held in a register.
  Value *IVOper = IVSrc;
  if (!Inc.IncExpr->isZero()) {
    // IncExpr was the result of subtraction of two narrow values, so must
    // be signed.
    const SCEV *IncExpr = SE.getNoopOrSignExtend(Inc.IncExpr, IntTy);
    LeftOverExpr = LeftOverExpr ?
      SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr;
  }
  if (LeftOverExpr && !LeftOverExpr->isZero()) {
    // Expand the IV increment.
    Rewriter.clearPostInc();
    Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt);
    const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc),
                                           SE.getUnknown(IncV));
    IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt);

    // If an IV increment can't be folded, use it as the next IV value.
    if (!canFoldIVIncExpr(LeftOverExpr, Inc.UserInst, Inc.IVOperand, TTI)) {
      assert(IVTy == IVOper->getType() && "inconsistent IV increment type")(static_cast <bool> (IVTy == IVOper->getType() &&
 "inconsistent IV increment type") ? void (0) : __assert_fail
 ("IVTy == IVOper->getType() && \"inconsistent IV increment type\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 3257, __extension__
 __PRETTY_FUNCTION__));
      IVSrc = IVOper;
      LeftOverExpr = nullptr;
    }
  }
  Type *OperTy = Inc.IVOperand->getType();
  if (IVTy != OperTy) {
    assert(SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) &&(static_cast <bool> (SE.getTypeSizeInBits(IVTy) >= SE
.getTypeSizeInBits(OperTy) && "cannot extend a chained IV"
) ? void (0) : __assert_fail ("SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) && \"cannot extend a chained IV\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 3265, __extension__
 __PRETTY_FUNCTION__))
           "cannot extend a chained IV")(static_cast <bool> (SE.getTypeSizeInBits(IVTy) >= SE
.getTypeSizeInBits(OperTy) && "cannot extend a chained IV"
) ? void (0) : __assert_fail ("SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) && \"cannot extend a chained IV\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 3265, __extension__
 __PRETTY_FUNCTION__));
    IRBuilder<> Builder(InsertPt);
    IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain");
  }
  Inc.UserInst->replaceUsesOfWith(Inc.IVOperand, IVOper);
  if (auto *OperandIsInstr = dyn_cast<Instruction>(Inc.IVOperand))
    DeadInsts.emplace_back(OperandIsInstr);
}
// If LSR created a new, wider phi, we may also replace its postinc. We only
// do this if we also found a wide value for the head of the chain.
if (isa<PHINode>(Chain.tailUserInst())) {
  for (PHINode &Phi : L->getHeader()->phis()) {
    if (!isCompatibleIVType(&Phi, IVSrc))
      continue;
    Instruction *PostIncV = dyn_cast<Instruction>(
        Phi.getIncomingValueForBlock(L->getLoopLatch()));
    if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc)))
      continue;
    Value *IVOper = IVSrc;
    Type *PostIncTy = PostIncV->getType();
    if (IVTy != PostIncTy) {
      assert(PostIncTy->isPointerTy() && "mixing int/ptr IV types")(static_cast <bool> (PostIncTy->isPointerTy() &&
 "mixing int/ptr IV types") ? void (0) : __assert_fail ("PostIncTy->isPointerTy() && \"mixing int/ptr IV types\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 3286, __extension__
 __PRETTY_FUNCTION__));
      IRBuilder<> Builder(L->getLoopLatch()->getTerminator());
      Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc());
      IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain");
    }
    Phi.replaceUsesOfWith(PostIncV, IVOper);
    DeadInsts.emplace_back(PostIncV);
  }
}
3295}

3297void LSRInstance::CollectFixupsAndInitialFormulae() {
BranchInst *ExitBranch = nullptr;
bool SaveCmp = TTI.canSaveCmp(L, &ExitBranch, &SE, &LI, &DT, &AC, &TLI);

// For calculating baseline cost
SmallPtrSet<const SCEV *, 16> Regs;
DenseSet<const SCEV *> VisitedRegs;
DenseSet<size_t> VisitedLSRUse;

for (const IVStrideUse &U : IU) {
  Instruction *UserInst = U.getUser();
  // Skip IV users that are part of profitable IV Chains.
  User::op_iterator UseI =
      find(UserInst->operands(), U.getOperandValToReplace());
  assert(UseI != UserInst->op_end() && "cannot find IV operand")(static_cast <bool> (UseI != UserInst->op_end() &&
 "cannot find IV operand") ? void (0) : __assert_fail ("UseI != UserInst->op_end() && \"cannot find IV operand\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 3311, __extension__
 __PRETTY_FUNCTION__));
  if (IVIncSet.count(UseI)) {
    LLVM_DEBUG(dbgs() << "Use is in profitable chain: " << **UseI << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Use is in profitable chain: "
 << **UseI << '\n'; } } while (false);
    continue;
  }

  LSRUse::KindType Kind = LSRUse::Basic;
  MemAccessTy AccessTy;
  if (isAddressUse(TTI, UserInst, U.getOperandValToReplace())) {
    Kind = LSRUse::Address;
    AccessTy = getAccessType(TTI, UserInst, U.getOperandValToReplace());
  }

  const SCEV *S = IU.getExpr(U);
  PostIncLoopSet TmpPostIncLoops = U.getPostIncLoops();

  // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
  // (N - i == 0), and this allows (N - i) to be the expression that we work
  // with rather than just N or i, so we can consider the register
  // requirements for both N and i at the same time. Limiting this code to
  // equality icmps is not a problem because all interesting loops use
  // equality icmps, thanks to IndVarSimplify.
  if (ICmpInst *CI = dyn_cast<ICmpInst>(UserInst)) {
    // If CI can be saved in some target, like replaced inside hardware loop
    // in PowerPC, no need to generate initial formulae for it.
    if (SaveCmp && CI == dyn_cast<ICmpInst>(ExitBranch->getCondition()))
      continue;
    if (CI->isEquality()) {
      // Swap the operands if needed to put the OperandValToReplace on the
      // left, for consistency.
      Value *NV = CI->getOperand(1);
      if (NV == U.getOperandValToReplace()) {
        CI->setOperand(1, CI->getOperand(0));
        CI->setOperand(0, NV);
        NV = CI->getOperand(1);
        Changed = true;
      }

      // x == y  -->  x - y == 0
      const SCEV *N = SE.getSCEV(NV);
      if (SE.isLoopInvariant(N, L) && Rewriter.isSafeToExpand(N) &&
          (!NV->getType()->isPointerTy() ||
           SE.getPointerBase(N) == SE.getPointerBase(S))) {
        // S is normalized, so normalize N before folding it into S
        // to keep the result normalized.
        N = normalizeForPostIncUse(N, TmpPostIncLoops, SE);
        Kind = LSRUse::ICmpZero;
        S = SE.getMinusSCEV(N, S);
      } else if (L->isLoopInvariant(NV) &&
                 (!isa<Instruction>(NV) ||
                  DT.dominates(cast<Instruction>(NV), L->getHeader())) &&
                 !NV->getType()->isPointerTy()) {
        // If we can't generally expand the expression (e.g. it contains
        // a divide), but it is already at a loop invariant point before the
        // loop, wrap it in an unknown (to prevent the expander from trying
        // to re-expand in a potentially unsafe way.)  The restriction to
        // integer types is required because the unknown hides the base, and
        // SCEV can't compute the difference of two unknown pointers.
        N = SE.getUnknown(NV);
        N = normalizeForPostIncUse(N, TmpPostIncLoops, SE);
        Kind = LSRUse::ICmpZero;
        S = SE.getMinusSCEV(N, S);
        assert(!isa<SCEVCouldNotCompute>(S))(static_cast <bool> (!isa<SCEVCouldNotCompute>(S)
) ? void (0) : __assert_fail ("!isa<SCEVCouldNotCompute>(S)"
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 3373, __extension__
 __PRETTY_FUNCTION__));
      }

      // -1 and the negations of all interesting strides (except the negation
      // of -1) are now also interesting.
      for (size_t i = 0, e = Factors.size(); i != e; ++i)
        if (Factors[i] != -1)
          Factors.insert(-(uint64_t)Factors[i]);
      Factors.insert(-1);
    }
  }

  // Get or create an LSRUse.
  std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
  size_t LUIdx = P.first;
  int64_t Offset = P.second;
  LSRUse &LU = Uses[LUIdx];

  // Record the fixup.
  LSRFixup &LF = LU.getNewFixup();
  LF.UserInst = UserInst;
  LF.OperandValToReplace = U.getOperandValToReplace();
  LF.PostIncLoops = TmpPostIncLoops;
  LF.Offset = Offset;
  LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);

  // Create SCEV as Formula for calculating baseline cost
  if (!VisitedLSRUse.count(LUIdx) && !LF.isUseFullyOutsideLoop(L)) {
    Formula F;
    F.initialMatch(S, L, SE);
    BaselineCost.RateFormula(F, Regs, VisitedRegs, LU);
    VisitedLSRUse.insert(LUIdx);
  }

  if (!LU.WidestFixupType ||
      SE.getTypeSizeInBits(LU.WidestFixupType) <
      SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
    LU.WidestFixupType = LF.OperandValToReplace->getType();

  // If this is the first use of this LSRUse, give it a formula.
  if (LU.Formulae.empty()) {
    InsertInitialFormula(S, LU, LUIdx);
    CountRegisters(LU.Formulae.back(), LUIdx);
  }
}

LLVM_DEBUG(print_fixups(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { print_fixups(dbgs()); } } while (false);
3420}

3422/// Insert a formula for the given expression into the given use, separating out
3423/// loop-variant portions from loop-invariant and loop-computable portions.
3424void LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU,
                                     size_t LUIdx) {
// Mark uses whose expressions cannot be expanded.
if (!Rewriter.isSafeToExpand(S))
  LU.RigidFormula = true;

Formula F;
F.initialMatch(S, L, SE);
bool Inserted = InsertFormula(LU, LUIdx, F);
assert(Inserted && "Initial formula already exists!")(static_cast <bool> (Inserted && "Initial formula already exists!"
) ? void (0) : __assert_fail ("Inserted && \"Initial formula already exists!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 3433, __extension__
 __PRETTY_FUNCTION__)); (void)Inserted;
3434}

3436/// Insert a simple single-register formula for the given expression into the
3437/// given use.
3438void
3439LSRInstance::InsertSupplementalFormula(const SCEV *S,
                                     LSRUse &LU, size_t LUIdx) {
Formula F;
F.BaseRegs.push_back(S);
F.HasBaseReg = true;
bool Inserted = InsertFormula(LU, LUIdx, F);
assert(Inserted && "Supplemental formula already exists!")(static_cast <bool> (Inserted && "Supplemental formula already exists!"
) ? void (0) : __assert_fail ("Inserted && \"Supplemental formula already exists!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 3445, __extension__
 __PRETTY_FUNCTION__)); (void)Inserted;
3446}

3448/// Note which registers are used by the given formula, updating RegUses.
3449void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
if (F.ScaledReg)
  RegUses.countRegister(F.ScaledReg, LUIdx);
for (const SCEV *BaseReg : F.BaseRegs)
  RegUses.countRegister(BaseReg, LUIdx);
3454}

3456/// If the given formula has not yet been inserted, add it to the list, and
3457/// return true. Return false otherwise.
3458bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
// Do not insert formula that we will not be able to expand.
assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F) &&(static_cast <bool> (isLegalUse(TTI, LU.MinOffset, LU.MaxOffset
, LU.Kind, LU.AccessTy, F) && "Formula is illegal") ?
 void (0) : __assert_fail ("isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F) && \"Formula is illegal\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 3461, __extension__
 __PRETTY_FUNCTION__))
       "Formula is illegal")(static_cast <bool> (isLegalUse(TTI, LU.MinOffset, LU.MaxOffset
, LU.Kind, LU.AccessTy, F) && "Formula is illegal") ?
 void (0) : __assert_fail ("isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F) && \"Formula is illegal\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 3461, __extension__
 __PRETTY_FUNCTION__));

if (!LU.InsertFormula(F, *L))
  return false;

CountRegisters(F, LUIdx);
return true;
3468}

3470/// Check for other uses of loop-invariant values which we're tracking. These
3471/// other uses will pin these values in registers, making them less profitable
3472/// for elimination.
3473/// TODO: This currently misses non-constant addrec step registers.
3474/// TODO: Should this give more weight to users inside the loop?
3475void
3476LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
SmallPtrSet<const SCEV *, 32> Visited;

while (!Worklist.empty()) {
  const SCEV *S = Worklist.pop_back_val();

  // Don't process the same SCEV twice
  if (!Visited.insert(S).second)
    continue;

  if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
    append_range(Worklist, N->operands());
  else if (const SCEVIntegralCastExpr *C = dyn_cast<SCEVIntegralCastExpr>(S))
    Worklist.push_back(C->getOperand());
  else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
    Worklist.push_back(D->getLHS());
    Worklist.push_back(D->getRHS());
  } else if (const SCEVUnknown *US = dyn_cast<SCEVUnknown>(S)) {
    const Value *V = US->getValue();
    if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
      // Look for instructions defined outside the loop.
      if (L->contains(Inst)) continue;
    } else if (isa<UndefValue>(V))
      // Undef doesn't have a live range, so it doesn't matter.
      continue;
    for (const Use &U : V->uses()) {
      const Instruction *UserInst = dyn_cast<Instruction>(U.getUser());
      // Ignore non-instructions.
      if (!UserInst)
        continue;
      // Don't bother if the instruction is an EHPad.
      if (UserInst->isEHPad())
        continue;
      // Ignore instructions in other functions (as can happen with
      // Constants).
      if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
        continue;
      // Ignore instructions not dominated by the loop.
      const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
        UserInst->getParent() :
        cast<PHINode>(UserInst)->getIncomingBlock(
          PHINode::getIncomingValueNumForOperand(U.getOperandNo()));
      if (!DT.dominates(L->getHeader(), UseBB))
        continue;
      // Don't bother if the instruction is in a BB which ends in an EHPad.
      if (UseBB->getTerminator()->isEHPad())
        continue;

      // Ignore cases in which the currently-examined value could come from
      // a basic block terminated with an EHPad. This checks all incoming
      // blocks of the phi node since it is possible that the same incoming
      // value comes from multiple basic blocks, only some of which may end
      // in an EHPad. If any of them do, a subsequent rewrite attempt by this
      // pass would try to insert instructions into an EHPad, hitting an
      // assertion.
      if (isa<PHINode>(UserInst)) {
        const auto *PhiNode = cast<PHINode>(UserInst);
        bool HasIncompatibleEHPTerminatedBlock = false;
        llvm::Value *ExpectedValue = U;
        for (unsigned int I = 0; I < PhiNode->getNumIncomingValues(); I++) {
          if (PhiNode->getIncomingValue(I) == ExpectedValue) {
            if (PhiNode->getIncomingBlock(I)->getTerminator()->isEHPad()) {
              HasIncompatibleEHPTerminatedBlock = true;
              break;
            }
          }
        }
        if (HasIncompatibleEHPTerminatedBlock) {
          continue;
        }
      }

      // Don't bother rewriting PHIs in catchswitch blocks.
      if (isa<CatchSwitchInst>(UserInst->getParent()->getTerminator()))
        continue;
      // Ignore uses which are part of other SCEV expressions, to avoid
      // analyzing them multiple times.
      if (SE.isSCEVable(UserInst->getType())) {
        const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
        // If the user is a no-op, look through to its uses.
        if (!isa<SCEVUnknown>(UserS))
          continue;
        if (UserS == US) {
          Worklist.push_back(
            SE.getUnknown(const_cast<Instruction *>(UserInst)));
          continue;
        }
      }
      // Ignore icmp instructions which are already being analyzed.
      if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
        unsigned OtherIdx = !U.getOperandNo();
        Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
        if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
          continue;
      }

      std::pair<size_t, int64_t> P = getUse(
          S, LSRUse::Basic, MemAccessTy());
      size_t LUIdx = P.first;
      int64_t Offset = P.second;
      LSRUse &LU = Uses[LUIdx];
      LSRFixup &LF = LU.getNewFixup();
      LF.UserInst = const_cast<Instruction *>(UserInst);
      LF.OperandValToReplace = U;
      LF.Offset = Offset;
      LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
      if (!LU.WidestFixupType ||
          SE.getTypeSizeInBits(LU.WidestFixupType) <
          SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
        LU.WidestFixupType = LF.OperandValToReplace->getType();
      InsertSupplementalFormula(US, LU, LUIdx);
      CountRegisters(LU.Formulae.back(), Uses.size() - 1);
      break;
    }
  }
}
3593}

3595/// Split S into subexpressions which can be pulled out into separate
3596/// registers. If C is non-null, multiply each subexpression by C.
3597///
3598/// Return remainder expression after factoring the subexpressions captured by
3599/// Ops. If Ops is complete, return NULL.
3600static const SCEV *CollectSubexprs(const SCEV *S, const SCEVConstant *C,
                                 SmallVectorImpl<const SCEV *> &Ops,
                                 const Loop *L,
                                 ScalarEvolution &SE,
                                 unsigned Depth = 0) {
// Arbitrarily cap recursion to protect compile time.
if (Depth >= 3)
  return S;

if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
  // Break out add operands.
  for (const SCEV *S : Add->operands()) {
    const SCEV *Remainder = CollectSubexprs(S, C, Ops, L, SE, Depth+1);
    if (Remainder)
      Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
  }
  return nullptr;
} else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
  // Split a non-zero base out of an addrec.
  if (AR->getStart()->isZero() || !AR->isAffine())
    return S;

  const SCEV *Remainder = CollectSubexprs(AR->getStart(),
                                          C, Ops, L, SE, Depth+1);
  // Split the non-zero AddRec unless it is part of a nested recurrence that
  // does not pertain to this loop.
  if (Remainder && (AR->getLoop() == L || !isa<SCEVAddRecExpr>(Remainder))) {
    Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
    Remainder = nullptr;
  }
  if (Remainder != AR->getStart()) {
    if (!Remainder)
      Remainder = SE.getConstant(AR->getType(), 0);
    return SE.getAddRecExpr(Remainder,
                            AR->getStepRecurrence(SE),
                            AR->getLoop(),
                            //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
                            SCEV::FlagAnyWrap);
  }
} else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
  // Break (C * (a + b + c)) into C*a + C*b + C*c.
  if (Mul->getNumOperands() != 2)
    return S;
  if (const SCEVConstant *Op0 =
      dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
    C = C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0;
    const SCEV *Remainder =
      CollectSubexprs(Mul->getOperand(1), C, Ops, L, SE, Depth+1);
    if (Remainder)
      Ops.push_back(SE.getMulExpr(C, Remainder));
    return nullptr;
  }
}
return S;
3654}

3656/// Return true if the SCEV represents a value that may end up as a
3657/// post-increment operation.
3658static bool mayUsePostIncMode(const TargetTransformInfo &TTI,
                            LSRUse &LU, const SCEV *S, const Loop *L,
                            ScalarEvolution &SE) {
if (LU.Kind != LSRUse::Address ||
    !LU.AccessTy.getType()->isIntOrIntVectorTy())
  return false;
const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S);
if (!AR)
  return false;
const SCEV *LoopStep = AR->getStepRecurrence(SE);
if (!isa<SCEVConstant>(LoopStep))
  return false;
// Check if a post-indexed load/store can be used.
if (TTI.isIndexedLoadLegal(TTI.MIM_PostInc, AR->getType()) ||
    TTI.isIndexedStoreLegal(TTI.MIM_PostInc, AR->getType())) {
  const SCEV *LoopStart = AR->getStart();
  if (!isa<SCEVConstant>(LoopStart) && SE.isLoopInvariant(LoopStart, L))
    return true;
}
return false;
3678}

3680/// Helper function for LSRInstance::GenerateReassociations.
3681void LSRInstance::GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
                                           const Formula &Base,
                                           unsigned Depth, size_t Idx,
                                           bool IsScaledReg) {
const SCEV *BaseReg = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
// Don't generate reassociations for the base register of a value that
// may generate a post-increment operator. The reason is that the
// reassociations cause extra base+register formula to be created,
// and possibly chosen, but the post-increment is more efficient.
if (AMK == TTI::AMK_PostIndexed && mayUsePostIncMode(TTI, LU, BaseReg, L, SE))
  return;
SmallVector<const SCEV *, 8> AddOps;
const SCEV *Remainder = CollectSubexprs(BaseReg, nullptr, AddOps, L, SE);
if (Remainder)
  AddOps.push_back(Remainder);

if (AddOps.size() == 1)
  return;

for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
                                                   JE = AddOps.end();
     J != JE; ++J) {
  // Loop-variant "unknown" values are uninteresting; we won't be able to
  // do anything meaningful with them.
  if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
    continue;

  // Don't pull a constant into a register if the constant could be folded
  // into an immediate field.
  if (isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
                       LU.AccessTy, *J, Base.getNumRegs() > 1))
    continue;

  // Collect all operands except *J.
  SmallVector<const SCEV *, 8> InnerAddOps(
      ((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
  InnerAddOps.append(std::next(J),
                     ((const SmallVector<const SCEV *, 8> &)AddOps).end());

  // Don't leave just a constant behind in a register if the constant could
  // be folded into an immediate field.
  if (InnerAddOps.size() == 1 &&
      isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
                       LU.AccessTy, InnerAddOps[0], Base.getNumRegs() > 1))
    continue;

  const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
  if (InnerSum->isZero())
    continue;
  Formula F = Base;

  // Add the remaining pieces of the add back into the new formula.
  const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
  if (InnerSumSC && SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
      TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
                              InnerSumSC->getValue()->getZExtValue())) {
    F.UnfoldedOffset =
        (uint64_t)F.UnfoldedOffset + InnerSumSC->getValue()->getZExtValue();
    if (IsScaledReg)
      F.ScaledReg = nullptr;
    else
      F.BaseRegs.erase(F.BaseRegs.begin() + Idx);
  } else if (IsScaledReg)
    F.ScaledReg = InnerSum;
  else
    F.BaseRegs[Idx] = InnerSum;

  // Add J as its own register, or an unfolded immediate.
  const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
  if (SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
      TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
                              SC->getValue()->getZExtValue()))
    F.UnfoldedOffset =
        (uint64_t)F.UnfoldedOffset + SC->getValue()->getZExtValue();
  else
    F.BaseRegs.push_back(*J);
  // We may have changed the number of register in base regs, adjust the
  // formula accordingly.
  F.canonicalize(*L);

  if (InsertFormula(LU, LUIdx, F))
    // If that formula hadn't been seen before, recurse to find more like
    // it.
    // Add check on Log16(AddOps.size()) - same as Log2_32(AddOps.size()) >> 2)
    // Because just Depth is not enough to bound compile time.
    // This means that every time AddOps.size() is greater 16^x we will add
    // x to Depth.
    GenerateReassociations(LU, LUIdx, LU.Formulae.back(),
                           Depth + 1 + (Log2_32(AddOps.size()) >> 2));
}
3771}

3773/// Split out subexpressions from adds and the bases of addrecs.
3774void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
                                       Formula Base, unsigned Depth) {
assert(Base.isCanonical(*L) && "Input must be in the canonical form")(static_cast <bool> (Base.isCanonical(*L) && "Input must be in the canonical form"
) ? void (0) : __assert_fail ("Base.isCanonical(*L) && \"Input must be in the canonical form\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 3776, __extension__
 __PRETTY_FUNCTION__));
// Arbitrarily cap recursion to protect compile time.
if (Depth >= 3)
  return;

for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
  GenerateReassociationsImpl(LU, LUIdx, Base, Depth, i);

if (Base.Scale == 1)
  GenerateReassociationsImpl(LU, LUIdx, Base, Depth,
                             /* Idx */ -1, /* IsScaledReg */ true);
3787}

3789///  Generate a formula consisting of all of the loop-dominating registers added
3790/// into a single register.
3791void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
                                     Formula Base) {
// This method is only interesting on a plurality of registers.
if (Base.BaseRegs.size() + (Base.Scale == 1) +
    (Base.UnfoldedOffset != 0) <= 1)
  return;

// Flatten the representation, i.e., reg1 + 1*reg2 => reg1 + reg2, before
// processing the formula.
Base.unscale();
SmallVector<const SCEV *, 4> Ops;
Formula NewBase = Base;
NewBase.BaseRegs.clear();
Type *CombinedIntegerType = nullptr;
for (const SCEV *BaseReg : Base.BaseRegs) {
  if (SE.properlyDominates(BaseReg, L->getHeader()) &&
      !SE.hasComputableLoopEvolution(BaseReg, L)) {
    if (!CombinedIntegerType)
      CombinedIntegerType = SE.getEffectiveSCEVType(BaseReg->getType());
    Ops.push_back(BaseReg);
  }
  else
    NewBase.BaseRegs.push_back(BaseReg);
}

// If no register is relevant, we're done.
if (Ops.size() == 0)
  return;

// Utility function for generating the required variants of the combined
// registers.
auto GenerateFormula = [&](const SCEV *Sum) {
  Formula F = NewBase;

  // TODO: If Sum is zero, it probably means ScalarEvolution missed an
  // opportunity to fold something. For now, just ignore such cases
  // rather than proceed with zero in a register.
  if (Sum->isZero())
    return;

  F.BaseRegs.push_back(Sum);
  F.canonicalize(*L);
  (void)InsertFormula(LU, LUIdx, F);
};

// If we collected at least two registers, generate a formula combining them.
if (Ops.size() > 1) {
  SmallVector<const SCEV *, 4> OpsCopy(Ops); // Don't let SE modify Ops.
  GenerateFormula(SE.getAddExpr(OpsCopy));
}

// If we have an unfolded offset, generate a formula combining it with the
// registers collected.
if (NewBase.UnfoldedOffset) {
  assert(CombinedIntegerType && "Missing a type for the unfolded offset")(static_cast <bool> (CombinedIntegerType && "Missing a type for the unfolded offset"
) ? void (0) : __assert_fail ("CombinedIntegerType && \"Missing a type for the unfolded offset\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 3845, __extension__
 __PRETTY_FUNCTION__));
  Ops.push_back(SE.getConstant(CombinedIntegerType, NewBase.UnfoldedOffset,
                               true));
  NewBase.UnfoldedOffset = 0;
  GenerateFormula(SE.getAddExpr(Ops));
}
3851}

3853/// Helper function for LSRInstance::GenerateSymbolicOffsets.
3854void LSRInstance::GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
                                            const Formula &Base, size_t Idx,
                                            bool IsScaledReg) {
const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
GlobalValue *GV = ExtractSymbol(G, SE);
if (G->isZero() || !GV)
  return;
Formula F = Base;
F.BaseGV = GV;
if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
  return;
if (IsScaledReg)
  F.ScaledReg = G;
else
  F.BaseRegs[Idx] = G;
(void)InsertFormula(LU, LUIdx, F);
3870}

3872/// Generate reuse formulae using symbolic offsets.
3873void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
                                        Formula Base) {
// We can't add a symbolic offset if the address already contains one.
if (Base.BaseGV) return;

for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
  GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, i);
if (Base.Scale == 1)
  GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, /* Idx */ -1,
                              /* IsScaledReg */ true);
3883}

3885/// Helper function for LSRInstance::GenerateConstantOffsets.
3886void LSRInstance::GenerateConstantOffsetsImpl(
  LSRUse &LU, unsigned LUIdx, const Formula &Base,
  const SmallVectorImpl<int64_t> &Worklist, size_t Idx, bool IsScaledReg) {

auto GenerateOffset = [&](const SCEV *G, int64_t Offset) {
  Formula F = Base;
  F.BaseOffset = (uint64_t)Base.BaseOffset - Offset;

  if (isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F)) {
    // Add the offset to the base register.
    const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), Offset), G);
    // If it cancelled out, drop the base register, otherwise update it.
    if (NewG->isZero()) {
      if (IsScaledReg) {
        F.Scale = 0;
        F.ScaledReg = nullptr;
      } else
        F.deleteBaseReg(F.BaseRegs[Idx]);
      F.canonicalize(*L);
    } else if (IsScaledReg)
      F.ScaledReg = NewG;
    else
      F.BaseRegs[Idx] = NewG;

    (void)InsertFormula(LU, LUIdx, F);
  }
};

const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];

// With constant offsets and constant steps, we can generate pre-inc
// accesses by having the offset equal the step. So, for access #0 with a
// step of 8, we generate a G - 8 base which would require the first access
// to be ((G - 8) + 8),+,8. The pre-indexed access then updates the pointer
// for itself and hopefully becomes the base for other accesses. This means
// means that a single pre-indexed access can be generated to become the new
// base pointer for each iteration of the loop, resulting in no extra add/sub
// instructions for pointer updating.
if (AMK == TTI::AMK_PreIndexed && LU.Kind == LSRUse::Address) {
  if (auto *GAR = dyn_cast<SCEVAddRecExpr>(G)) {
    if (auto *StepRec =
        dyn_cast<SCEVConstant>(GAR->getStepRecurrence(SE))) {
      const APInt &StepInt = StepRec->getAPInt();
      int64_t Step = StepInt.isNegative() ?
        StepInt.getSExtValue() : StepInt.getZExtValue();

      for (int64_t Offset : Worklist) {
        Offset -= Step;
        GenerateOffset(G, Offset);
      }
    }
  }
}
for (int64_t Offset : Worklist)
  GenerateOffset(G, Offset);

int64_t Imm = ExtractImmediate(G, SE);
if (G->isZero() || Imm == 0)
  return;
Formula F = Base;
F.BaseOffset = (uint64_t)F.BaseOffset + Imm;
if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
  return;
if (IsScaledReg) {
  F.ScaledReg = G;
} else {
  F.BaseRegs[Idx] = G;
  // We may generate non canonical Formula if G is a recurrent expr reg
  // related with current loop while F.ScaledReg is not.
  F.canonicalize(*L);
}
(void)InsertFormula(LU, LUIdx, F);
3958}

3960/// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
3961void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
                                        Formula Base) {
// TODO: For now, just add the min and max offset, because it usually isn't
// worthwhile looking at everything inbetween.
SmallVector<int64_t, 2> Worklist;
Worklist.push_back(LU.MinOffset);
if (LU.MaxOffset != LU.MinOffset)
  Worklist.push_back(LU.MaxOffset);

for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
  GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, i);
if (Base.Scale == 1)
  GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, /* Idx */ -1,
                              /* IsScaledReg */ true);
3975}

3977/// For ICmpZero, check to see if we can scale up the comparison. For example, x
3978/// == y -> x*c == y*c.
3979void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
                                       Formula Base) {
if (LU.Kind != LSRUse::ICmpZero) return;

// Determine the integer type for the base formula.
Type *IntTy = Base.getType();
if (!IntTy) return;
if (SE.getTypeSizeInBits(IntTy) > 64) return;

// Don't do this if there is more than one offset.
if (LU.MinOffset != LU.MaxOffset) return;

// Check if transformation is valid. It is illegal to multiply pointer.
if (Base.ScaledReg && Base.ScaledReg->getType()->isPointerTy())
  return;
for (const SCEV *BaseReg : Base.BaseRegs)
  if (BaseReg->getType()->isPointerTy())
    return;
assert(!Base.BaseGV && "ICmpZero use is not legal!")(static_cast <bool> (!Base.BaseGV && "ICmpZero use is not legal!"
) ? void (0) : __assert_fail ("!Base.BaseGV && \"ICmpZero use is not legal!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 3997, __extension__
 __PRETTY_FUNCTION__));

// Check each interesting stride.
for (int64_t Factor : Factors) {
  // Check that Factor can be represented by IntTy
  if (!ConstantInt::isValueValidForType(IntTy, Factor))
    continue;
  // Check that the multiplication doesn't overflow.
  if (Base.BaseOffset == std::numeric_limits<int64_t>::min() && Factor == -1)
    continue;
  int64_t NewBaseOffset = (uint64_t)Base.BaseOffset * Factor;
  assert(Factor != 0 && "Zero factor not expected!")(static_cast <bool> (Factor != 0 && "Zero factor not expected!"
) ? void (0) : __assert_fail ("Factor != 0 && \"Zero factor not expected!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 4008, __extension__
 __PRETTY_FUNCTION__));
  if (NewBaseOffset / Factor != Base.BaseOffset)
    continue;
  // If the offset will be truncated at this use, check that it is in bounds.
  if (!IntTy->isPointerTy() &&
      !ConstantInt::isValueValidForType(IntTy, NewBaseOffset))
    continue;

  // Check that multiplying with the use offset doesn't overflow.
  int64_t Offset = LU.MinOffset;
  if (Offset == std::numeric_limits<int64_t>::min() && Factor == -1)
    continue;
  Offset = (uint64_t)Offset * Factor;
  if (Offset / Factor != LU.MinOffset)
    continue;
  // If the offset will be truncated at this use, check that it is in bounds.
  if (!IntTy->isPointerTy() &&
      !ConstantInt::isValueValidForType(IntTy, Offset))
    continue;

  Formula F = Base;
  F.BaseOffset = NewBaseOffset;

  // Check that this scale is legal.
  if (!isLegalUse(TTI, Offset, Offset, LU.Kind, LU.AccessTy, F))
    continue;

  // Compensate for the use having MinOffset built into it.
  F.BaseOffset = (uint64_t)F.BaseOffset + Offset - LU.MinOffset;

  const SCEV *FactorS = SE.getConstant(IntTy, Factor);

  // Check that multiplying with each base register doesn't overflow.
  for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
    F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
    if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
      goto next;
  }

  // Check that multiplying with the scaled register doesn't overflow.
  if (F.ScaledReg) {
    F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
    if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
      continue;
  }

  // Check that multiplying with the unfolded offset doesn't overflow.
  if (F.UnfoldedOffset != 0) {
    if (F.UnfoldedOffset == std::numeric_limits<int64_t>::min() &&
        Factor == -1)
      continue;
    F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
    if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
      continue;
    // If the offset will be truncated, check that it is in bounds.
    if (!IntTy->isPointerTy() &&
        !ConstantInt::isValueValidForType(IntTy, F.UnfoldedOffset))
      continue;
  }

  // If we make it here and it's legal, add it.
  (void)InsertFormula(LU, LUIdx, F);
next:;
}
4072}

4074/// Generate stride factor reuse formulae by making use of scaled-offset address
4075/// modes, for example.
4076void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
// Determine the integer type for the base formula.
Type *IntTy = Base.getType();
if (!IntTy) return;

// If this Formula already has a scaled register, we can't add another one.
// Try to unscale the formula to generate a better scale.
if (Base.Scale != 0 && !Base.unscale())
  return;

assert(Base.Scale == 0 && "unscale did not did its job!")(static_cast <bool> (Base.Scale == 0 && "unscale did not did its job!"
) ? void (0) : __assert_fail ("Base.Scale == 0 && \"unscale did not did its job!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 4086, __extension__
 __PRETTY_FUNCTION__));

// Check each interesting stride.
for (int64_t Factor : Factors) {
  Base.Scale = Factor;
  Base.HasBaseReg = Base.BaseRegs.size() > 1;
  // Check whether this scale is going to be legal.
  if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
                  Base)) {
    // As a special-case, handle special out-of-loop Basic users specially.
    // TODO: Reconsider this special case.
    if (LU.Kind == LSRUse::Basic &&
        isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LSRUse::Special,
                   LU.AccessTy, Base) &&
        LU.AllFixupsOutsideLoop)
      LU.Kind = LSRUse::Special;
    else
      continue;
  }
  // For an ICmpZero, negating a solitary base register won't lead to
  // new solutions.
  if (LU.Kind == LSRUse::ICmpZero &&
      !Base.HasBaseReg && Base.BaseOffset == 0 && !Base.BaseGV)
    continue;
  // For each addrec base reg, if its loop is current loop, apply the scale.
  for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
    const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i]);
    if (AR && (AR->getLoop() == L || LU.AllFixupsOutsideLoop)) {
      const SCEV *FactorS = SE.getConstant(IntTy, Factor);
      if (FactorS->isZero())
        continue;
      // Divide out the factor, ignoring high bits, since we'll be
      // scaling the value back up in the end.
      if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true))
        if (!Quotient->isZero()) {
          // TODO: This could be optimized to avoid all the copying.
          Formula F = Base;
          F.ScaledReg = Quotient;
          F.deleteBaseReg(F.BaseRegs[i]);
          // The canonical representation of 1*reg is reg, which is already in
          // Base. In that case, do not try to insert the formula, it will be
          // rejected anyway.
          if (F.Scale == 1 && (F.BaseRegs.empty() ||
                               (AR->getLoop() != L && LU.AllFixupsOutsideLoop)))
            continue;
          // If AllFixupsOutsideLoop is true and F.Scale is 1, we may generate
          // non canonical Formula with ScaledReg's loop not being L.
          if (F.Scale == 1 && LU.AllFixupsOutsideLoop)
            F.canonicalize(*L);
          (void)InsertFormula(LU, LUIdx, F);
        }
    }
  }
}
4140}

4142/// Generate reuse formulae from different IV types.
4143void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
// Don't bother truncating symbolic values.
if (Base.BaseGV) return;

// Determine the integer type for the base formula.
Type *DstTy = Base.getType();
if (!DstTy) return;
if (DstTy->isPointerTy())
  return;

// It is invalid to extend a pointer type so exit early if ScaledReg or
// any of the BaseRegs are pointers.
if (Base.ScaledReg && Base.ScaledReg->getType()->isPointerTy())
  return;
if (any_of(Base.BaseRegs,
           [](const SCEV *S) { return S->getType()->isPointerTy(); }))
  return;

for (Type *SrcTy : Types) {
  if (SrcTy != DstTy && TTI.isTruncateFree(SrcTy, DstTy)) {
    Formula F = Base;

    // Sometimes SCEV is able to prove zero during ext transform. It may
    // happen if SCEV did not do all possible transforms while creating the
    // initial node (maybe due to depth limitations), but it can do them while
    // taking ext.
    if (F.ScaledReg) {
      const SCEV *NewScaledReg = SE.getAnyExtendExpr(F.ScaledReg, SrcTy);
      if (NewScaledReg->isZero())
       continue;
      F.ScaledReg = NewScaledReg;
    }
    bool HasZeroBaseReg = false;
    for (const SCEV *&BaseReg : F.BaseRegs) {
      const SCEV *NewBaseReg = SE.getAnyExtendExpr(BaseReg, SrcTy);
      if (NewBaseReg->isZero()) {
        HasZeroBaseReg = true;
        break;
      }
      BaseReg = NewBaseReg;
    }
    if (HasZeroBaseReg)
      continue;

    // TODO: This assumes we've done basic processing on all uses and
    // have an idea what the register usage is.
    if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
      continue;

    F.canonicalize(*L);
    (void)InsertFormula(LU, LUIdx, F);
  }
}
4196}

4198namespace {

4200/// Helper class for GenerateCrossUseConstantOffsets. It's used to defer
4201/// modifications so that the search phase doesn't have to worry about the data
4202/// structures moving underneath it.
4203struct WorkItem {
size_t LUIdx;
int64_t Imm;
const SCEV *OrigReg;

WorkItem(size_t LI, int64_t I, const SCEV *R)
    : LUIdx(LI), Imm(I), OrigReg(R) {}

void print(raw_ostream &OS) const;
void dump() const;
4213};

4215} // end anonymous namespace

4217#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
4218void WorkItem::print(raw_ostream &OS) const {
OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
   << " , add offset " << Imm;
4221}

4223LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void WorkItem::dump() const {
print(errs()); errs() << '\n';
4225}
4226#endif

4228/// Look for registers which are a constant distance apart and try to form reuse
4229/// opportunities between them.
4230void LSRInstance::GenerateCrossUseConstantOffsets() {
// Group the registers by their value without any added constant offset.
using ImmMapTy = std::map<int64_t, const SCEV *>;

DenseMap<const SCEV *, ImmMapTy> Map;
DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
SmallVector<const SCEV *, 8> Sequence;
for (const SCEV *Use : RegUses) {
  const SCEV *Reg = Use; // Make a copy for ExtractImmediate to modify.
  int64_t Imm = ExtractImmediate(Reg, SE);
  auto Pair = Map.insert(std::make_pair(Reg, ImmMapTy()));
  if (Pair.second)
    Sequence.push_back(Reg);
  Pair.first->second.insert(std::make_pair(Imm, Use));
  UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(Use);
}

// Now examine each set of registers with the same base value. Build up
// a list of work to do and do the work in a separate step so that we're
// not adding formulae and register counts while we're searching.
SmallVector<WorkItem, 32> WorkItems;
SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
for (const SCEV *Reg : Sequence) {
  const ImmMapTy &Imms = Map.find(Reg)->second;

  // It's not worthwhile looking for reuse if there's only one offset.
  if (Imms.size() == 1)
    continue;

  LLVM_DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Generating cross-use offsets for "
 << *Reg << ':'; for (const auto &Entry : Imms
) dbgs() << ' ' << Entry.first; dbgs() << '\n'
; } } while (false)
             for (const auto &Entrydo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Generating cross-use offsets for "
 << *Reg << ':'; for (const auto &Entry : Imms
) dbgs() << ' ' << Entry.first; dbgs() << '\n'
; } } while (false)
                  : Imms) dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Generating cross-use offsets for "
 << *Reg << ':'; for (const auto &Entry : Imms
) dbgs() << ' ' << Entry.first; dbgs() << '\n'
; } } while (false)
             << ' ' << Entry.first;do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Generating cross-use offsets for "
 << *Reg << ':'; for (const auto &Entry : Imms
) dbgs() << ' ' << Entry.first; dbgs() << '\n'
; } } while (false)
             dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Generating cross-use offsets for "
 << *Reg << ':'; for (const auto &Entry : Imms
) dbgs() << ' ' << Entry.first; dbgs() << '\n'
; } } while (false);

  // Examine each offset.
  for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
       J != JE; ++J) {
    const SCEV *OrigReg = J->second;

    int64_t JImm = J->first;
    const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);

    if (!isa<SCEVConstant>(OrigReg) &&
        UsedByIndicesMap[Reg].count() == 1) {
      LLVM_DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigRegdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Skipping cross-use reuse for "
 << *OrigReg << '\n'; } } while (false)
                        << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Skipping cross-use reuse for "
 << *OrigReg << '\n'; } } while (false);
      continue;
    }

    // Conservatively examine offsets between this orig reg a few selected
    // other orig regs.
    int64_t First = Imms.begin()->first;
    int64_t Last = std::prev(Imms.end())->first;
    // Compute (First + Last)  / 2 without overflow using the fact that
    // First + Last = 2 * (First + Last) + (First ^ Last).
    int64_t Avg = (First & Last) + ((First ^ Last) >> 1);
    // If the result is negative and First is odd and Last even (or vice versa),
    // we rounded towards -inf. Add 1 in that case, to round towards 0.
    Avg = Avg + ((First ^ Last) & ((uint64_t)Avg >> 63));
    ImmMapTy::const_iterator OtherImms[] = {
        Imms.begin(), std::prev(Imms.end()),
       Imms.lower_bound(Avg)};
    for (const auto &M : OtherImms) {
      if (M == J || M == JE) continue;

      // Compute the difference between the two.
      int64_t Imm = (uint64_t)JImm - M->first;
      for (unsigned LUIdx : UsedByIndices.set_bits())
        // Make a memo of this use, offset, and register tuple.
        if (UniqueItems.insert(std::make_pair(LUIdx, Imm)).second)
          WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
    }
  }
}

Map.clear();
Sequence.clear();
UsedByIndicesMap.clear();
UniqueItems.clear();

// Now iterate through the worklist and add new formulae.
for (const WorkItem &WI : WorkItems) {
  size_t LUIdx = WI.LUIdx;
  LSRUse &LU = Uses[LUIdx];
  int64_t Imm = WI.Imm;
  const SCEV *OrigReg = WI.OrigReg;

  Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
  const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
  unsigned BitWidth = SE.getTypeSizeInBits(IntTy);

  // TODO: Use a more targeted data structure.
  for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
    Formula F = LU.Formulae[L];
    // FIXME: The code for the scaled and unscaled registers looks
    // very similar but slightly different. Investigate if they
    // could be merged. That way, we would not have to unscale the
    // Formula.
    F.unscale();
    // Use the immediate in the scaled register.
    if (F.ScaledReg == OrigReg) {
      int64_t Offset = (uint64_t)F.BaseOffset + Imm * (uint64_t)F.Scale;
      // Don't create 50 + reg(-50).
      if (F.referencesReg(SE.getSCEV(
                 ConstantInt::get(IntTy, -(uint64_t)Offset))))
        continue;
      Formula NewF = F;
      NewF.BaseOffset = Offset;
      if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
                      NewF))
        continue;
      NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);

      // If the new scale is a constant in a register, and adding the constant
      // value to the immediate would produce a value closer to zero than the
      // immediate itself, then the formula isn't worthwhile.
      if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
        if (C->getValue()->isNegative() != (NewF.BaseOffset < 0) &&
            (C->getAPInt().abs() * APInt(BitWidth, F.Scale))
                .ule(std::abs(NewF.BaseOffset)))
          continue;

      // OK, looks good.
      NewF.canonicalize(*this->L);
      (void)InsertFormula(LU, LUIdx, NewF);
    } else {
      // Use the immediate in a base register.
      for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
        const SCEV *BaseReg = F.BaseRegs[N];
        if (BaseReg != OrigReg)
          continue;
        Formula NewF = F;
        NewF.BaseOffset = (uint64_t)NewF.BaseOffset + Imm;
        if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset,
                        LU.Kind, LU.AccessTy, NewF)) {
          if (AMK == TTI::AMK_PostIndexed &&
              mayUsePostIncMode(TTI, LU, OrigReg, this->L, SE))
            continue;
          if (!TTI.isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
            continue;
          NewF = F;
          NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
        }
        NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);

        // If the new formula has a constant in a register, and adding the
        // constant value to the immediate would produce a value closer to
        // zero than the immediate itself, then the formula isn't worthwhile.
        for (const SCEV *NewReg : NewF.BaseRegs)
          if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewReg))
            if ((C->getAPInt() + NewF.BaseOffset)
                    .abs()
                    .slt(std::abs(NewF.BaseOffset)) &&
                (C->getAPInt() + NewF.BaseOffset).countr_zero() >=
                    (unsigned)llvm::countr_zero<uint64_t>(NewF.BaseOffset))
              goto skip_formula;

        // Ok, looks good.
        NewF.canonicalize(*this->L);
        (void)InsertFormula(LU, LUIdx, NewF);
        break;
      skip_formula:;
      }
    }
  }
}
4397}

4399/// Generate formulae for each use.
4400void
4401LSRInstance::GenerateAllReuseFormulae() {
// This is split into multiple loops so that hasRegsUsedByUsesOtherThan
// queries are more precise.
for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
  LSRUse &LU = Uses[LUIdx];
  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
    GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
    GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
}
for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
  LSRUse &LU = Uses[LUIdx];
  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
    GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
    GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
    GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
    GenerateScales(LU, LUIdx, LU.Formulae[i]);
}
for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
  LSRUse &LU = Uses[LUIdx];
  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
    GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
}

GenerateCrossUseConstantOffsets();

LLVM_DEBUG(dbgs() << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\n" "After generating reuse formulae:\n"
; print_uses(dbgs()); } } while (false)
                     "After generating reuse formulae:\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\n" "After generating reuse formulae:\n"
; print_uses(dbgs()); } } while (false)
           print_uses(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\n" "After generating reuse formulae:\n"
; print_uses(dbgs()); } } while (false);
4433}

4435/// If there are multiple formulae with the same set of registers used
4436/// by other uses, pick the best one and delete the others.
4437void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
DenseSet<const SCEV *> VisitedRegs;
SmallPtrSet<const SCEV *, 16> Regs;
SmallPtrSet<const SCEV *, 16> LoserRegs;
4441#ifndef NDEBUG
bool ChangedFormulae = false;
4443#endif

// Collect the best formula for each unique set of shared registers. This
// is reset for each use.
using BestFormulaeTy =
    DenseMap<SmallVector<const SCEV *, 4>, size_t, UniquifierDenseMapInfo>;

BestFormulaeTy BestFormulae;

for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
  LSRUse &LU = Uses[LUIdx];
  LLVM_DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Filtering for use "; LU.print
(dbgs()); dbgs() << '\n'; } } while (false)
             dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Filtering for use "; LU.print
(dbgs()); dbgs() << '\n'; } } while (false);

  bool Any = false;
  for (size_t FIdx = 0, NumForms = LU.Formulae.size();
       FIdx != NumForms; ++FIdx) {
    Formula &F = LU.Formulae[FIdx];

    // Some formulas are instant losers. For example, they may depend on
    // nonexistent AddRecs from other loops. These need to be filtered
    // immediately, otherwise heuristics could choose them over others leading
    // to an unsatisfactory solution. Passing LoserRegs into RateFormula here
    // avoids the need to recompute this information across formulae using the
    // same bad AddRec. Passing LoserRegs is also essential unless we remove
    // the corresponding bad register from the Regs set.
    Cost CostF(L, SE, TTI, AMK);
    Regs.clear();
    CostF.RateFormula(F, Regs, VisitedRegs, LU, &LoserRegs);
    if (CostF.isLoser()) {
      // During initial formula generation, undesirable formulae are generated
      // by uses within other loops that have some non-trivial address mode or
      // use the postinc form of the IV. LSR needs to provide these formulae
      // as the basis of rediscovering the desired formula that uses an AddRec
      // corresponding to the existing phi. Once all formulae have been
      // generated, these initial losers may be pruned.
      LLVM_DEBUG(dbgs() << "  Filtering loser "; F.print(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "  Filtering loser "; F.print
(dbgs()); dbgs() << "\n"; } } while (false)
                 dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "  Filtering loser "; F.print
(dbgs()); dbgs() << "\n"; } } while (false);
    }
    else {
      SmallVector<const SCEV *, 4> Key;
      for (const SCEV *Reg : F.BaseRegs) {
        if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
          Key.push_back(Reg);
      }
      if (F.ScaledReg &&
          RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
        Key.push_back(F.ScaledReg);
      // Unstable sort by host order ok, because this is only used for
      // uniquifying.
      llvm::sort(Key);

      std::pair<BestFormulaeTy::const_iterator, bool> P =
        BestFormulae.insert(std::make_pair(Key, FIdx));
      if (P.second)
        continue;

      Formula &Best = LU.Formulae[P.first->second];

      Cost CostBest(L, SE, TTI, AMK);
      Regs.clear();
      CostBest.RateFormula(Best, Regs, VisitedRegs, LU);
      if (CostF.isLess(CostBest))
        std::swap(F, Best);
      LLVM_DEBUG(dbgs() << "  Filtering out formula "; F.print(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "  Filtering out formula "
; F.print(dbgs()); dbgs() << "\n" "    in favor of formula "
; Best.print(dbgs()); dbgs() << '\n'; } } while (false)
                 dbgs() << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "  Filtering out formula "
; F.print(dbgs()); dbgs() << "\n" "    in favor of formula "
; Best.print(dbgs()); dbgs() << '\n'; } } while (false)
                           "    in favor of formula ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "  Filtering out formula "
; F.print(dbgs()); dbgs() << "\n" "    in favor of formula "
; Best.print(dbgs()); dbgs() << '\n'; } } while (false)
                 Best.print(dbgs()); dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "  Filtering out formula "
; F.print(dbgs()); dbgs() << "\n" "    in favor of formula "
; Best.print(dbgs()); dbgs() << '\n'; } } while (false);
    }
4512#ifndef NDEBUG
    ChangedFormulae = true;
4514#endif
    LU.DeleteFormula(F);
    --FIdx;
    --NumForms;
    Any = true;
  }

  // Now that we've filtered out some formulae, recompute the Regs set.
  if (Any)
    LU.RecomputeRegs(LUIdx, RegUses);

  // Reset this to prepare for the next use.
  BestFormulae.clear();
}

LLVM_DEBUG(if (ChangedFormulae) {do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { if (ChangedFormulae) { dbgs() << "\n"
 "After filtering out undesirable candidates:\n"; print_uses(
dbgs()); }; } } while (false)
  dbgs() << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { if (ChangedFormulae) { dbgs() << "\n"
 "After filtering out undesirable candidates:\n"; print_uses(
dbgs()); }; } } while (false)
            "After filtering out undesirable candidates:\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { if (ChangedFormulae) { dbgs() << "\n"
 "After filtering out undesirable candidates:\n"; print_uses(
dbgs()); }; } } while (false)
  print_uses(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { if (ChangedFormulae) { dbgs() << "\n"
 "After filtering out undesirable candidates:\n"; print_uses(
dbgs()); }; } } while (false)
})do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { if (ChangedFormulae) { dbgs() << "\n"
 "After filtering out undesirable candidates:\n"; print_uses(
dbgs()); }; } } while (false);
4534}

4536/// Estimate the worst-case number of solutions the solver might have to
4537/// consider. It almost never considers this many solutions because it prune the
4538/// search space, but the pruning isn't always sufficient.
4539size_t LSRInstance::EstimateSearchSpaceComplexity() const {
size_t Power = 1;
for (const LSRUse &LU : Uses) {
  size_t FSize = LU.Formulae.size();
  if (FSize >= ComplexityLimit) {
    Power = ComplexityLimit;
    break;
  }
  Power *= FSize;
  if (Power >= ComplexityLimit)
    break;
}
return Power;
4552}

4554/// When one formula uses a superset of the registers of another formula, it
4555/// won't help reduce register pressure (though it may not necessarily hurt
4556/// register pressure); remove it to simplify the system.
4557void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
  LLVM_DEBUG(dbgs() << "The search space is too complex.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "The search space is too complex.\n"
; } } while (false);

  LLVM_DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Narrowing the search space by eliminating formulae "
 "which use a superset of registers used by other " "formulae.\n"
; } } while (false)
                       "which use a superset of registers used by other "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Narrowing the search space by eliminating formulae "
 "which use a superset of registers used by other " "formulae.\n"
; } } while (false)
                       "formulae.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Narrowing the search space by eliminating formulae "
 "which use a superset of registers used by other " "formulae.\n"
; } } while (false);

  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
    LSRUse &LU = Uses[LUIdx];
    bool Any = false;
    for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
      Formula &F = LU.Formulae[i];
      // Look for a formula with a constant or GV in a register. If the use
      // also has a formula with that same value in an immediate field,
      // delete the one that uses a register.
      for (SmallVectorImpl<const SCEV *>::const_iterator
           I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
        if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
          Formula NewF = F;
          //FIXME: Formulas should store bitwidth to do wrapping properly.
          //       See PR41034.
          NewF.BaseOffset += (uint64_t)C->getValue()->getSExtValue();
          NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
                              (I - F.BaseRegs.begin()));
          if (LU.HasFormulaWithSameRegs(NewF)) {
            LLVM_DEBUG(dbgs() << "  Deleting "; F.print(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "  Deleting "; F.print(dbgs
()); dbgs() << '\n'; } } while (false)
                       dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "  Deleting "; F.print(dbgs
()); dbgs() << '\n'; } } while (false);
            LU.DeleteFormula(F);
            --i;
            --e;
            Any = true;
            break;
          }
        } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
          if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
            if (!F.BaseGV) {
              Formula NewF = F;
              NewF.BaseGV = GV;
              NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
                                  (I - F.BaseRegs.begin()));
              if (LU.HasFormulaWithSameRegs(NewF)) {
                LLVM_DEBUG(dbgs() << "  Deleting "; F.print(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "  Deleting "; F.print(dbgs
()); dbgs() << '\n'; } } while (false)
                           dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "  Deleting "; F.print(dbgs
()); dbgs() << '\n'; } } while (false);
                LU.DeleteFormula(F);
                --i;
                --e;
                Any = true;
                break;
              }
            }
        }
      }
    }
    if (Any)
      LU.RecomputeRegs(LUIdx, RegUses);
  }

  LLVM_DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "After pre-selection:\n"; print_uses
(dbgs()); } } while (false);
}
4617}

4619/// When there are many registers for expressions like A, A+1, A+2, etc.,
4620/// allocate a single register for them.
4621void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
if (EstimateSearchSpaceComplexity() < ComplexityLimit)
  return;

LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "The search space is too complex.\n"
 "Narrowing the search space by assuming that uses separated "
 "by a constant offset will use the same registers.\n"; } } while
 (false)
    dbgs() << "The search space is too complex.\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "The search space is too complex.\n"
 "Narrowing the search space by assuming that uses separated "
 "by a constant offset will use the same registers.\n"; } } while
 (false)
              "Narrowing the search space by assuming that uses separated "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "The search space is too complex.\n"
 "Narrowing the search space by assuming that uses separated "
 "by a constant offset will use the same registers.\n"; } } while
 (false)
              "by a constant offset will use the same registers.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "The search space is too complex.\n"
 "Narrowing the search space by assuming that uses separated "
 "by a constant offset will use the same registers.\n"; } } while
 (false);

// This is especially useful for unrolled loops.

for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
  LSRUse &LU = Uses[LUIdx];
  for (const Formula &F : LU.Formulae) {
    if (F.BaseOffset == 0 || (F.Scale != 0 && F.Scale != 1))
      continue;

    LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU);
    if (!LUThatHas)
      continue;

    if (!reconcileNewOffset(*LUThatHas, F.BaseOffset, /*HasBaseReg=*/ false,
                            LU.Kind, LU.AccessTy))
      continue;

    LLVM_DEBUG(dbgs() << "  Deleting use "; LU.print(dbgs()); dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "  Deleting use "; LU.print
(dbgs()); dbgs() << '\n'; } } while (false);

    LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;

    // Transfer the fixups of LU to LUThatHas.
    for (LSRFixup &Fixup : LU.Fixups) {
      Fixup.Offset += F.BaseOffset;
      LUThatHas->pushFixup(Fixup);
      LLVM_DEBUG(dbgs() << "New fixup has offset " << Fixup.Offset << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "New fixup has offset " <<
 Fixup.Offset << '\n'; } } while (false);
    }

    // Delete formulae from the new use which are no longer legal.
    bool Any = false;
    for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
      Formula &F = LUThatHas->Formulae[i];
      if (!isLegalUse(TTI, LUThatHas->MinOffset, LUThatHas->MaxOffset,
                      LUThatHas->Kind, LUThatHas->AccessTy, F)) {
        LLVM_DEBUG(dbgs() << "  Deleting "; F.print(dbgs()); dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "  Deleting "; F.print(dbgs
()); dbgs() << '\n'; } } while (false);
        LUThatHas->DeleteFormula(F);
        --i;
        --e;
        Any = true;
      }
    }

    if (Any)
      LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);

    // Delete the old use.
    DeleteUse(LU, LUIdx);
    --LUIdx;
    --NumUses;
    break;
  }
}

LLVM_DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "After pre-selection:\n"; print_uses
(dbgs()); } } while (false);
4683}

4685/// Call FilterOutUndesirableDedicatedRegisters again, if necessary, now that
4686/// we've done more filtering, as it may be able to find more formulae to
4687/// eliminate.
4688void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
  LLVM_DEBUG(dbgs() << "The search space is too complex.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "The search space is too complex.\n"
; } } while (false);

  LLVM_DEBUG(dbgs() << "Narrowing the search space by re-filtering out "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Narrowing the search space by re-filtering out "
 "undesirable dedicated registers.\n"; } } while (false)
                       "undesirable dedicated registers.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Narrowing the search space by re-filtering out "
 "undesirable dedicated registers.\n"; } } while (false);

  FilterOutUndesirableDedicatedRegisters();

  LLVM_DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "After pre-selection:\n"; print_uses
(dbgs()); } } while (false);
}
4699}

4701/// If a LSRUse has multiple formulae with the same ScaledReg and Scale.
4702/// Pick the best one and delete the others.
4703/// This narrowing heuristic is to keep as many formulae with different
4704/// Scale and ScaledReg pair as possible while narrowing the search space.
4705/// The benefit is that it is more likely to find out a better solution
4706/// from a formulae set with more Scale and ScaledReg variations than
4707/// a formulae set with the same Scale and ScaledReg. The picking winner
4708/// reg heuristic will often keep the formulae with the same Scale and
4709/// ScaledReg and filter others, and we want to avoid that if possible.
4710void LSRInstance::NarrowSearchSpaceByFilterFormulaWithSameScaledReg() {
if (EstimateSearchSpaceComplexity() < ComplexityLimit)
  return;

LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "The search space is too complex.\n"
 "Narrowing the search space by choosing the best Formula " "from the Formulae with the same Scale and ScaledReg.\n"
; } } while (false)
    dbgs() << "The search space is too complex.\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "The search space is too complex.\n"
 "Narrowing the search space by choosing the best Formula " "from the Formulae with the same Scale and ScaledReg.\n"
; } } while (false)
              "Narrowing the search space by choosing the best Formula "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "The search space is too complex.\n"
 "Narrowing the search space by choosing the best Formula " "from the Formulae with the same Scale and ScaledReg.\n"
; } } while (false)
              "from the Formulae with the same Scale and ScaledReg.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "The search space is too complex.\n"
 "Narrowing the search space by choosing the best Formula " "from the Formulae with the same Scale and ScaledReg.\n"
; } } while (false);

// Map the "Scale * ScaledReg" pair to the best formula of current LSRUse.
using BestFormulaeTy = DenseMap<std::pair<const SCEV *, int64_t>, size_t>;

BestFormulaeTy BestFormulae;
4723#ifndef NDEBUG
bool ChangedFormulae = false;
4725#endif
DenseSet<const SCEV *> VisitedRegs;
SmallPtrSet<const SCEV *, 16> Regs;

for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
  LSRUse &LU = Uses[LUIdx];
  LLVM_DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Filtering for use "; LU.print
(dbgs()); dbgs() << '\n'; } } while (false)
             dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Filtering for use "; LU.print
(dbgs()); dbgs() << '\n'; } } while (false);

  // Return true if Formula FA is better than Formula FB.
  auto IsBetterThan = [&](Formula &FA, Formula &FB) {
    // First we will try to choose the Formula with fewer new registers.
    // For a register used by current Formula, the more the register is
    // shared among LSRUses, the less we increase the register number
    // counter of the formula.
    size_t FARegNum = 0;
    for (const SCEV *Reg : FA.BaseRegs) {
      const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(Reg);
      FARegNum += (NumUses - UsedByIndices.count() + 1);
    }
    size_t FBRegNum = 0;
    for (const SCEV *Reg : FB.BaseRegs) {
      const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(Reg);
      FBRegNum += (NumUses - UsedByIndices.count() + 1);
    }
    if (FARegNum != FBRegNum)
      return FARegNum < FBRegNum;

    // If the new register numbers are the same, choose the Formula with
    // less Cost.
    Cost CostFA(L, SE, TTI, AMK);
    Cost CostFB(L, SE, TTI, AMK);
    Regs.clear();
    CostFA.RateFormula(FA, Regs, VisitedRegs, LU);
    Regs.clear();
    CostFB.RateFormula(FB, Regs, VisitedRegs, LU);
    return CostFA.isLess(CostFB);
  };

  bool Any = false;
  for (size_t FIdx = 0, NumForms = LU.Formulae.size(); FIdx != NumForms;
       ++FIdx) {
    Formula &F = LU.Formulae[FIdx];
    if (!F.ScaledReg)
      continue;
    auto P = BestFormulae.insert({{F.ScaledReg, F.Scale}, FIdx});
    if (P.second)
      continue;

    Formula &Best = LU.Formulae[P.first->second];
    if (IsBetterThan(F, Best))
      std::swap(F, Best);
    LLVM_DEBUG(dbgs() << "  Filtering out formula "; F.print(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "  Filtering out formula "
; F.print(dbgs()); dbgs() << "\n" "    in favor of formula "
; Best.print(dbgs()); dbgs() << '\n'; } } while (false)
               dbgs() << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "  Filtering out formula "
; F.print(dbgs()); dbgs() << "\n" "    in favor of formula "
; Best.print(dbgs()); dbgs() << '\n'; } } while (false)
                         "    in favor of formula ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "  Filtering out formula "
; F.print(dbgs()); dbgs() << "\n" "    in favor of formula "
; Best.print(dbgs()); dbgs() << '\n'; } } while (false)
               Best.print(dbgs()); dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "  Filtering out formula "
; F.print(dbgs()); dbgs() << "\n" "    in favor of formula "
; Best.print(dbgs()); dbgs() << '\n'; } } while (false);
4781#ifndef NDEBUG
    ChangedFormulae = true;
4783#endif
    LU.DeleteFormula(F);
    --FIdx;
    --NumForms;
    Any = true;
  }
  if (Any)
    LU.RecomputeRegs(LUIdx, RegUses);

  // Reset this to prepare for the next use.
  BestFormulae.clear();
}

LLVM_DEBUG(if (ChangedFormulae) {do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { if (ChangedFormulae) { dbgs() << "\n"
 "After filtering out undesirable candidates:\n"; print_uses(
dbgs()); }; } } while (false)
  dbgs() << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { if (ChangedFormulae) { dbgs() << "\n"
 "After filtering out undesirable candidates:\n"; print_uses(
dbgs()); }; } } while (false)
            "After filtering out undesirable candidates:\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { if (ChangedFormulae) { dbgs() << "\n"
 "After filtering out undesirable candidates:\n"; print_uses(
dbgs()); }; } } while (false)
  print_uses(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { if (ChangedFormulae) { dbgs() << "\n"
 "After filtering out undesirable candidates:\n"; print_uses(
dbgs()); }; } } while (false)
})do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { if (ChangedFormulae) { dbgs() << "\n"
 "After filtering out undesirable candidates:\n"; print_uses(
dbgs()); }; } } while (false);
4801}

4803/// If we are over the complexity limit, filter out any post-inc prefering
4804/// variables to only post-inc values.
4805void LSRInstance::NarrowSearchSpaceByFilterPostInc() {
if (AMK != TTI::AMK_PostIndexed)
  return;
if (EstimateSearchSpaceComplexity() < ComplexityLimit)
  return;

LLVM_DEBUG(dbgs() << "The search space is too complex.\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "The search space is too complex.\n"
 "Narrowing the search space by choosing the lowest " "register Formula for PostInc Uses.\n"
; } } while (false)
                     "Narrowing the search space by choosing the lowest "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "The search space is too complex.\n"
 "Narrowing the search space by choosing the lowest " "register Formula for PostInc Uses.\n"
; } } while (false)
                     "register Formula for PostInc Uses.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "The search space is too complex.\n"
 "Narrowing the search space by choosing the lowest " "register Formula for PostInc Uses.\n"
; } } while (false);

for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
  LSRUse &LU = Uses[LUIdx];

  if (LU.Kind != LSRUse::Address)
    continue;
  if (!TTI.isIndexedLoadLegal(TTI.MIM_PostInc, LU.AccessTy.getType()) &&
      !TTI.isIndexedStoreLegal(TTI.MIM_PostInc, LU.AccessTy.getType()))
    continue;

  size_t MinRegs = std::numeric_limits<size_t>::max();
  for (const Formula &F : LU.Formulae)
    MinRegs = std::min(F.getNumRegs(), MinRegs);

  bool Any = false;
  for (size_t FIdx = 0, NumForms = LU.Formulae.size(); FIdx != NumForms;
       ++FIdx) {
    Formula &F = LU.Formulae[FIdx];
    if (F.getNumRegs() > MinRegs) {
      LLVM_DEBUG(dbgs() << "  Filtering out formula "; F.print(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "  Filtering out formula "
; F.print(dbgs()); dbgs() << "\n"; } } while (false)
                 dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "  Filtering out formula "
; F.print(dbgs()); dbgs() << "\n"; } } while (false);
      LU.DeleteFormula(F);
      --FIdx;
      --NumForms;
      Any = true;
    }
  }
  if (Any)
    LU.RecomputeRegs(LUIdx, RegUses);

  if (EstimateSearchSpaceComplexity() < ComplexityLimit)
    break;
}

LLVM_DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "After pre-selection:\n"; print_uses
(dbgs()); } } while (false);
4849}

4851/// The function delete formulas with high registers number expectation.
4852/// Assuming we don't know the value of each formula (already delete
4853/// all inefficient), generate probability of not selecting for each
4854/// register.
4855/// For example,
4856/// Use1:
4857///  reg(a) + reg({0,+,1})
4858///  reg(a) + reg({-1,+,1}) + 1
4859///  reg({a,+,1})
4860/// Use2:
4861///  reg(b) + reg({0,+,1})
4862///  reg(b) + reg({-1,+,1}) + 1
4863///  reg({b,+,1})
4864/// Use3:
4865///  reg(c) + reg(b) + reg({0,+,1})
4866///  reg(c) + reg({b,+,1})
4867///
4868/// Probability of not selecting
4869///                 Use1   Use2    Use3
4870/// reg(a)         (1/3) *   1   *   1
4871/// reg(b)           1   * (1/3) * (1/2)
4872/// reg({0,+,1})   (2/3) * (2/3) * (1/2)
4873/// reg({-1,+,1})  (2/3) * (2/3) *   1
4874/// reg({a,+,1})   (2/3) *   1   *   1
4875/// reg({b,+,1})     1   * (2/3) * (2/3)
4876/// reg(c)           1   *   1   *   0
4877///
4878/// Now count registers number mathematical expectation for each formula:
4879/// Note that for each use we exclude probability if not selecting for the use.
4880/// For example for Use1 probability for reg(a) would be just 1 * 1 (excluding
4881/// probabilty 1/3 of not selecting for Use1).
4882/// Use1:
4883///  reg(a) + reg({0,+,1})          1 + 1/3       -- to be deleted
4884///  reg(a) + reg({-1,+,1}) + 1     1 + 4/9       -- to be deleted
4885///  reg({a,+,1})                   1
4886/// Use2:
4887///  reg(b) + reg({0,+,1})          1/2 + 1/3     -- to be deleted
4888///  reg(b) + reg({-1,+,1}) + 1     1/2 + 2/3     -- to be deleted
4889///  reg({b,+,1})                   2/3
4890/// Use3:
4891///  reg(c) + reg(b) + reg({0,+,1}) 1 + 1/3 + 4/9 -- to be deleted
4892///  reg(c) + reg({b,+,1})          1 + 2/3
4893void LSRInstance::NarrowSearchSpaceByDeletingCostlyFormulas() {
if (EstimateSearchSpaceComplexity() < ComplexityLimit)
  return;
// Ok, we have too many of formulae on our hands to conveniently handle.
// Use a rough heuristic to thin out the list.

// Set of Regs wich will be 100% used in final solution.
// Used in each formula of a solution (in example above this is reg(c)).
// We can skip them in calculations.
SmallPtrSet<const SCEV *, 4> UniqRegs;
LLVM_DEBUG(dbgs() << "The search space is too complex.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "The search space is too complex.\n"
; } } while (false);

// Map each register to probability of not selecting
DenseMap <const SCEV *, float> RegNumMap;
for (const SCEV *Reg : RegUses) {
  if (UniqRegs.count(Reg))
    continue;
  float PNotSel = 1;
  for (const LSRUse &LU : Uses) {
    if (!LU.Regs.count(Reg))
      continue;
    float P = LU.getNotSelectedProbability(Reg);
    if (P != 0.0)
      PNotSel *= P;
    else
      UniqRegs.insert(Reg);
  }
  RegNumMap.insert(std::make_pair(Reg, PNotSel));
}

LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Narrowing the search space by deleting costly formulas\n"
; } } while (false)
    dbgs() << "Narrowing the search space by deleting costly formulas\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Narrowing the search space by deleting costly formulas\n"
; } } while (false);

// Delete formulas where registers number expectation is high.
for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
  LSRUse &LU = Uses[LUIdx];
  // If nothing to delete - continue.
  if (LU.Formulae.size() < 2)
    continue;
  // This is temporary solution to test performance. Float should be
  // replaced with round independent type (based on integers) to avoid
  // different results for different target builds.
  float FMinRegNum = LU.Formulae[0].getNumRegs();
  float FMinARegNum = LU.Formulae[0].getNumRegs();
  size_t MinIdx = 0;
  for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
    Formula &F = LU.Formulae[i];
    float FRegNum = 0;
    float FARegNum = 0;
    for (const SCEV *BaseReg : F.BaseRegs) {
      if (UniqRegs.count(BaseReg))
        continue;
      FRegNum += RegNumMap[BaseReg] / LU.getNotSelectedProbability(BaseReg);
      if (isa<SCEVAddRecExpr>(BaseReg))
        FARegNum +=
            RegNumMap[BaseReg] / LU.getNotSelectedProbability(BaseReg);
    }
    if (const SCEV *ScaledReg = F.ScaledReg) {
      if (!UniqRegs.count(ScaledReg)) {
        FRegNum +=
            RegNumMap[ScaledReg] / LU.getNotSelectedProbability(ScaledReg);
        if (isa<SCEVAddRecExpr>(ScaledReg))
          FARegNum +=
              RegNumMap[ScaledReg] / LU.getNotSelectedProbability(ScaledReg);
      }
    }
    if (FMinRegNum > FRegNum ||
        (FMinRegNum == FRegNum && FMinARegNum > FARegNum)) {
      FMinRegNum = FRegNum;
      FMinARegNum = FARegNum;
      MinIdx = i;
    }
  }
  LLVM_DEBUG(dbgs() << "  The formula "; LU.Formulae[MinIdx].print(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "  The formula "; LU.Formulae
[MinIdx].print(dbgs()); dbgs() << " with min reg num " <<
 FMinRegNum << '\n'; } } while (false)
             dbgs() << " with min reg num " << FMinRegNum << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "  The formula "; LU.Formulae
[MinIdx].print(dbgs()); dbgs() << " with min reg num " <<
 FMinRegNum << '\n'; } } while (false);
  if (MinIdx != 0)
    std::swap(LU.Formulae[MinIdx], LU.Formulae[0]);
  while (LU.Formulae.size() != 1) {
    LLVM_DEBUG(dbgs() << "  Deleting "; LU.Formulae.back().print(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "  Deleting "; LU.Formulae
.back().print(dbgs()); dbgs() << '\n'; } } while (false
)
               dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "  Deleting "; LU.Formulae
.back().print(dbgs()); dbgs() << '\n'; } } while (false
);
    LU.Formulae.pop_back();
  }
  LU.RecomputeRegs(LUIdx, RegUses);
  assert(LU.Formulae.size() == 1 && "Should be exactly 1 min regs formula")(static_cast <bool> (LU.Formulae.size() == 1 &&
 "Should be exactly 1 min regs formula") ? void (0) : __assert_fail
 ("LU.Formulae.size() == 1 && \"Should be exactly 1 min regs formula\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 4976, __extension__
 __PRETTY_FUNCTION__));
  Formula &F = LU.Formulae[0];
  LLVM_DEBUG(dbgs() << "  Leaving only "; F.print(dbgs()); dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "  Leaving only "; F.print
(dbgs()); dbgs() << '\n'; } } while (false);
  // When we choose the formula, the regs become unique.
  UniqRegs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
  if (F.ScaledReg)
    UniqRegs.insert(F.ScaledReg);
}
LLVM_DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "After pre-selection:\n"; print_uses
(dbgs()); } } while (false);
4985}

4987// Check if Best and Reg are SCEVs separated by a constant amount C, and if so
4988// would the addressing offset +C would be legal where the negative offset -C is
4989// not.
4990static bool IsSimplerBaseSCEVForTarget(const TargetTransformInfo &TTI,
                                     ScalarEvolution &SE, const SCEV *Best,
                                     const SCEV *Reg,
                                     MemAccessTy AccessType) {
if (Best->getType() != Reg->getType() ||
    (isa<SCEVAddRecExpr>(Best) && isa<SCEVAddRecExpr>(Reg) &&
     cast<SCEVAddRecExpr>(Best)->getLoop() !=
         cast<SCEVAddRecExpr>(Reg)->getLoop()))
  return false;
const auto *Diff = dyn_cast<SCEVConstant>(SE.getMinusSCEV(Best, Reg));
if (!Diff)
  return false;

return TTI.isLegalAddressingMode(
           AccessType.MemTy, /*BaseGV=*/nullptr,
           /*BaseOffset=*/Diff->getAPInt().getSExtValue(),
           /*HasBaseReg=*/false, /*Scale=*/0, AccessType.AddrSpace) &&
       !TTI.isLegalAddressingMode(
           AccessType.MemTy, /*BaseGV=*/nullptr,
           /*BaseOffset=*/-Diff->getAPInt().getSExtValue(),
           /*HasBaseReg=*/false, /*Scale=*/0, AccessType.AddrSpace);
5011}

5013/// Pick a register which seems likely to be profitable, and then in any use
5014/// which has any reference to that register, delete all formulae which do not
5015/// reference that register.
5016void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
// With all other options exhausted, loop until the system is simple
// enough to handle.
SmallPtrSet<const SCEV *, 4> Taken;
while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
  // Ok, we have too many of formulae on our hands to conveniently handle.
  // Use a rough heuristic to thin out the list.
  LLVM_DEBUG(dbgs() << "The search space is too complex.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "The search space is too complex.\n"
; } } while (false);

  // Pick the register which is used by the most LSRUses, which is likely
  // to be a good reuse register candidate.
  const SCEV *Best = nullptr;
  unsigned BestNum = 0;
  for (const SCEV *Reg : RegUses) {
    if (Taken.count(Reg))
      continue;
    if (!Best) {
      Best = Reg;
      BestNum = RegUses.getUsedByIndices(Reg).count();
    } else {
      unsigned Count = RegUses.getUsedByIndices(Reg).count();
      if (Count > BestNum) {
        Best = Reg;
        BestNum = Count;
      }

      // If the scores are the same, but the Reg is simpler for the target
      // (for example {x,+,1} as opposed to {x+C,+,1}, where the target can
      // handle +C but not -C), opt for the simpler formula.
      if (Count == BestNum) {
        int LUIdx = RegUses.getUsedByIndices(Reg).find_first();
        if (LUIdx >= 0 && Uses[LUIdx].Kind == LSRUse::Address &&
            IsSimplerBaseSCEVForTarget(TTI, SE, Best, Reg,
                                       Uses[LUIdx].AccessTy)) {
          Best = Reg;
          BestNum = Count;
        }
      }
    }
  }
  assert(Best && "Failed to find best LSRUse candidate")(static_cast <bool> (Best && "Failed to find best LSRUse candidate"
) ? void (0) : __assert_fail ("Best && \"Failed to find best LSRUse candidate\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 5056, __extension__
 __PRETTY_FUNCTION__));

  LLVM_DEBUG(dbgs() << "Narrowing the search space by assuming " << *Bestdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Narrowing the search space by assuming "
 << *Best << " will yield profitable reuse.\n"; }
 } while (false)
                    << " will yield profitable reuse.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Narrowing the search space by assuming "
 << *Best << " will yield profitable reuse.\n"; }
 } while (false);
  Taken.insert(Best);

  // In any use with formulae which references this register, delete formulae
  // which don't reference it.
  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
    LSRUse &LU = Uses[LUIdx];
    if (!LU.Regs.count(Best)) continue;

    bool Any = false;
    for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
      Formula &F = LU.Formulae[i];
      if (!F.referencesReg(Best)) {
        LLVM_DEBUG(dbgs() << "  Deleting "; F.print(dbgs()); dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "  Deleting "; F.print(dbgs
()); dbgs() << '\n'; } } while (false);
        LU.DeleteFormula(F);
        --e;
        --i;
        Any = true;
        assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?")(static_cast <bool> (e != 0 && "Use has no formulae left! Is Regs inconsistent?"
) ? void (0) : __assert_fail ("e != 0 && \"Use has no formulae left! Is Regs inconsistent?\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 5077, __extension__
 __PRETTY_FUNCTION__));
        continue;
      }
    }

    if (Any)
      LU.RecomputeRegs(LUIdx, RegUses);
  }

  LLVM_DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "After pre-selection:\n"; print_uses
(dbgs()); } } while (false);
}
5088}

5090/// If there are an extraordinary number of formulae to choose from, use some
5091/// rough heuristics to prune down the number of formulae. This keeps the main
5092/// solver from taking an extraordinary amount of time in some worst-case
5093/// scenarios.
5094void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
NarrowSearchSpaceByDetectingSupersets();
NarrowSearchSpaceByCollapsingUnrolledCode();
NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
if (FilterSameScaledReg)
  NarrowSearchSpaceByFilterFormulaWithSameScaledReg();
NarrowSearchSpaceByFilterPostInc();
if (LSRExpNarrow)
  NarrowSearchSpaceByDeletingCostlyFormulas();
else
  NarrowSearchSpaceByPickingWinnerRegs();
5105}

5107/// This is the recursive solver.
5108void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
                             Cost &SolutionCost,
                             SmallVectorImpl<const Formula *> &Workspace,
                             const Cost &CurCost,
                             const SmallPtrSet<const SCEV *, 16> &CurRegs,
                             DenseSet<const SCEV *> &VisitedRegs) const {
// Some ideas:
//  - prune more:
//    - use more aggressive filtering
//    - sort the formula so that the most profitable solutions are found first
//    - sort the uses too
//  - search faster:
//    - don't compute a cost, and then compare. compare while computing a cost
//      and bail early.
//    - track register sets with SmallBitVector

const LSRUse &LU = Uses[Workspace.size()];

// If this use references any register that's already a part of the
// in-progress solution, consider it a requirement that a formula must
// reference that register in order to be considered. This prunes out
// unprofitable searching.
SmallSetVector<const SCEV *, 4> ReqRegs;
for (const SCEV *S : CurRegs)
  if (LU.Regs.count(S))
    ReqRegs.insert(S);

SmallPtrSet<const SCEV *, 16> NewRegs;
Cost NewCost(L, SE, TTI, AMK);
for (const Formula &F : LU.Formulae) {
  // Ignore formulae which may not be ideal in terms of register reuse of
  // ReqRegs.  The formula should use all required registers before
  // introducing new ones.
  // This can sometimes (notably when trying to favour postinc) lead to
  // sub-optimial decisions. There it is best left to the cost modelling to
  // get correct.
  if (AMK != TTI::AMK_PostIndexed || LU.Kind != LSRUse::Address) {
    int NumReqRegsToFind = std::min(F.getNumRegs(), ReqRegs.size());
    for (const SCEV *Reg : ReqRegs) {
      if ((F.ScaledReg && F.ScaledReg == Reg) ||
          is_contained(F.BaseRegs, Reg)) {
        --NumReqRegsToFind;
        if (NumReqRegsToFind == 0)
          break;
      }
    }
    if (NumReqRegsToFind != 0) {
      // If none of the formulae satisfied the required registers, then we could
      // clear ReqRegs and try again. Currently, we simply give up in this case.
      continue;
    }
  }

  // Evaluate the cost of the current formula. If it's already worse than
  // the current best, prune the search at that point.
  NewCost = CurCost;
  NewRegs = CurRegs;
  NewCost.RateFormula(F, NewRegs, VisitedRegs, LU);
  if (NewCost.isLess(SolutionCost)) {
    Workspace.push_back(&F);
    if (Workspace.size() != Uses.size()) {
      SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
                   NewRegs, VisitedRegs);
      if (F.getNumRegs() == 1 && Workspace.size() == 1)
        VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
    } else {
      LLVM_DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "New best at "; NewCost.print
(dbgs()); dbgs() << ".\nRegs:\n"; for (const SCEV *S : NewRegs
) dbgs() << "- " << *S << "\n"; dbgs() <<
 '\n'; } } while (false)
                 dbgs() << ".\nRegs:\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "New best at "; NewCost.print
(dbgs()); dbgs() << ".\nRegs:\n"; for (const SCEV *S : NewRegs
) dbgs() << "- " << *S << "\n"; dbgs() <<
 '\n'; } } while (false)
                 for (const SCEV *S : NewRegs) dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "New best at "; NewCost.print
(dbgs()); dbgs() << ".\nRegs:\n"; for (const SCEV *S : NewRegs
) dbgs() << "- " << *S << "\n"; dbgs() <<
 '\n'; } } while (false)
                    << "- " << *S << "\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "New best at "; NewCost.print
(dbgs()); dbgs() << ".\nRegs:\n"; for (const SCEV *S : NewRegs
) dbgs() << "- " << *S << "\n"; dbgs() <<
 '\n'; } } while (false)
                 dbgs() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "New best at "; NewCost.print
(dbgs()); dbgs() << ".\nRegs:\n"; for (const SCEV *S : NewRegs
) dbgs() << "- " << *S << "\n"; dbgs() <<
 '\n'; } } while (false);

      SolutionCost = NewCost;
      Solution = Workspace;
    }
    Workspace.pop_back();
  }
}
5186}

5188/// Choose one formula from each use. Return the results in the given Solution
5189/// vector.
5190void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
SmallVector<const Formula *, 8> Workspace;
Cost SolutionCost(L, SE, TTI, AMK);
SolutionCost.Lose();
Cost CurCost(L, SE, TTI, AMK);
SmallPtrSet<const SCEV *, 16> CurRegs;
DenseSet<const SCEV *> VisitedRegs;
Workspace.reserve(Uses.size());

// SolveRecurse does all the work.
SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
             CurRegs, VisitedRegs);
if (Solution.empty()) {
  LLVM_DEBUG(dbgs() << "\nNo Satisfactory Solution\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\nNo Satisfactory Solution\n"
; } } while (false);
  return;
}

// Ok, we've now made all our decisions.
LLVM_DEBUG(dbgs() << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\n" "The chosen solution requires "
; SolutionCost.print(dbgs()); dbgs() << ":\n"; for (size_t
 i = 0, e = Uses.size(); i != e; ++i) { dbgs() << "  ";
 Uses[i].print(dbgs()); dbgs() << "\n" "    "; Solution
[i]->print(dbgs()); dbgs() << '\n'; }; } } while (false
)
                     "The chosen solution requires ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\n" "The chosen solution requires "
; SolutionCost.print(dbgs()); dbgs() << ":\n"; for (size_t
 i = 0, e = Uses.size(); i != e; ++i) { dbgs() << "  ";
 Uses[i].print(dbgs()); dbgs() << "\n" "    "; Solution
[i]->print(dbgs()); dbgs() << '\n'; }; } } while (false
)
           SolutionCost.print(dbgs()); dbgs() << ":\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\n" "The chosen solution requires "
; SolutionCost.print(dbgs()); dbgs() << ":\n"; for (size_t
 i = 0, e = Uses.size(); i != e; ++i) { dbgs() << "  ";
 Uses[i].print(dbgs()); dbgs() << "\n" "    "; Solution
[i]->print(dbgs()); dbgs() << '\n'; }; } } while (false
)
           for (size_t i = 0, e = Uses.size(); i != e; ++i) {do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\n" "The chosen solution requires "
; SolutionCost.print(dbgs()); dbgs() << ":\n"; for (size_t
 i = 0, e = Uses.size(); i != e; ++i) { dbgs() << "  ";
 Uses[i].print(dbgs()); dbgs() << "\n" "    "; Solution
[i]->print(dbgs()); dbgs() << '\n'; }; } } while (false
)
             dbgs() << "  ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\n" "The chosen solution requires "
; SolutionCost.print(dbgs()); dbgs() << ":\n"; for (size_t
 i = 0, e = Uses.size(); i != e; ++i) { dbgs() << "  ";
 Uses[i].print(dbgs()); dbgs() << "\n" "    "; Solution
[i]->print(dbgs()); dbgs() << '\n'; }; } } while (false
)
             Uses[i].print(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\n" "The chosen solution requires "
; SolutionCost.print(dbgs()); dbgs() << ":\n"; for (size_t
 i = 0, e = Uses.size(); i != e; ++i) { dbgs() << "  ";
 Uses[i].print(dbgs()); dbgs() << "\n" "    "; Solution
[i]->print(dbgs()); dbgs() << '\n'; }; } } while (false
)
             dbgs() << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\n" "The chosen solution requires "
; SolutionCost.print(dbgs()); dbgs() << ":\n"; for (size_t
 i = 0, e = Uses.size(); i != e; ++i) { dbgs() << "  ";
 Uses[i].print(dbgs()); dbgs() << "\n" "    "; Solution
[i]->print(dbgs()); dbgs() << '\n'; }; } } while (false
)
                       "    ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\n" "The chosen solution requires "
; SolutionCost.print(dbgs()); dbgs() << ":\n"; for (size_t
 i = 0, e = Uses.size(); i != e; ++i) { dbgs() << "  ";
 Uses[i].print(dbgs()); dbgs() << "\n" "    "; Solution
[i]->print(dbgs()); dbgs() << '\n'; }; } } while (false
)
             Solution[i]->print(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\n" "The chosen solution requires "
; SolutionCost.print(dbgs()); dbgs() << ":\n"; for (size_t
 i = 0, e = Uses.size(); i != e; ++i) { dbgs() << "  ";
 Uses[i].print(dbgs()); dbgs() << "\n" "    "; Solution
[i]->print(dbgs()); dbgs() << '\n'; }; } } while (false
)
             dbgs() << '\n';do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\n" "The chosen solution requires "
; SolutionCost.print(dbgs()); dbgs() << ":\n"; for (size_t
 i = 0, e = Uses.size(); i != e; ++i) { dbgs() << "  ";
 Uses[i].print(dbgs()); dbgs() << "\n" "    "; Solution
[i]->print(dbgs()); dbgs() << '\n'; }; } } while (false
)
           })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "\n" "The chosen solution requires "
; SolutionCost.print(dbgs()); dbgs() << ":\n"; for (size_t
 i = 0, e = Uses.size(); i != e; ++i) { dbgs() << "  ";
 Uses[i].print(dbgs()); dbgs() << "\n" "    "; Solution
[i]->print(dbgs()); dbgs() << '\n'; }; } } while (false
);

assert(Solution.size() == Uses.size() && "Malformed solution!")(static_cast <bool> (Solution.size() == Uses.size() &&
 "Malformed solution!") ? void (0) : __assert_fail ("Solution.size() == Uses.size() && \"Malformed solution!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 5220, __extension__
 __PRETTY_FUNCTION__));

if (BaselineCost.isLess(SolutionCost)) {
  LLVM_DEBUG(dbgs() << "The baseline solution requires ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "The baseline solution requires "
; BaselineCost.print(dbgs()); dbgs() << "\n"; } } while
 (false)
             BaselineCost.print(dbgs()); dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "The baseline solution requires "
; BaselineCost.print(dbgs()); dbgs() << "\n"; } } while
 (false);
  if (!AllowDropSolutionIfLessProfitable)
    LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Baseline is more profitable than chosen solution, "
 "add option 'lsr-drop-solution' to drop LSR solution.\n"; } }
 while (false)
        dbgs() << "Baseline is more profitable than chosen solution, "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Baseline is more profitable than chosen solution, "
 "add option 'lsr-drop-solution' to drop LSR solution.\n"; } }
 while (false)
                  "add option 'lsr-drop-solution' to drop LSR solution.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Baseline is more profitable than chosen solution, "
 "add option 'lsr-drop-solution' to drop LSR solution.\n"; } }
 while (false);
  else {
    LLVM_DEBUG(dbgs() << "Baseline is more profitable than chosen "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Baseline is more profitable than chosen "
 "solution, dropping LSR solution.\n";; } } while (false)
                         "solution, dropping LSR solution.\n";)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("loop-reduce")) { dbgs() << "Baseline is more profitable than chosen "
 "solution, dropping LSR solution.\n";; } } while (false);
    Solution.clear();
  }
}
5235}

5237/// Helper for AdjustInsertPositionForExpand. Climb up the dominator tree far as
5238/// we can go while still being dominated by the input positions. This helps
5239/// canonicalize the insert position, which encourages sharing.
5240BasicBlock::iterator
5241LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
                               const SmallVectorImpl<Instruction *> &Inputs)
                                                                       const {
Instruction *Tentative = &*IP;
while (true) {
  bool AllDominate = true;
  Instruction *BetterPos = nullptr;
  // Don't bother attempting to insert before a catchswitch, their basic block
  // cannot have other non-PHI instructions.
  if (isa<CatchSwitchInst>(Tentative))
    return IP;

  for (Instruction *Inst : Inputs) {
    if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
      AllDominate = false;
      break;
    }
    // Attempt to find an insert position in the middle of the block,
    // instead of at the end, so that it can be used for other expansions.
    if (Tentative->getParent() == Inst->getParent() &&
        (!BetterPos || !DT.dominates(Inst, BetterPos)))
      BetterPos = &*std::next(BasicBlock::iterator(Inst));
  }
  if (!AllDominate)
    break;
  if (BetterPos)
    IP = BetterPos->getIterator();
  else
    IP = Tentative->getIterator();

  const Loop *IPLoop = LI.getLoopFor(IP->getParent());
  unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;

  BasicBlock *IDom;
  for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
    if (!Rung) return IP;
    Rung = Rung->getIDom();
    if (!Rung) return IP;
    IDom = Rung->getBlock();

    // Don't climb into a loop though.
    const Loop *IDomLoop = LI.getLoopFor(IDom);
    unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
    if (IDomDepth <= IPLoopDepth &&
        (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
      break;
  }

  Tentative = IDom->getTerminator();
}

return IP;
5293}

5295/// Determine an input position which will be dominated by the operands and
5296/// which will dominate the result.
5297BasicBlock::iterator LSRInstance::AdjustInsertPositionForExpand(
  BasicBlock::iterator LowestIP, const LSRFixup &LF, const LSRUse &LU) const {
// Collect some instructions which must be dominated by the
// expanding replacement. These must be dominated by any operands that
// will be required in the expansion.
SmallVector<Instruction *, 4> Inputs;
if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
  Inputs.push_back(I);
if (LU.Kind == LSRUse::ICmpZero)
  if (Instruction *I =
        dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
    Inputs.push_back(I);
if (LF.PostIncLoops.count(L)) {
  if (LF.isUseFullyOutsideLoop(L))
    Inputs.push_back(L->getLoopLatch()->getTerminator());
  else
    Inputs.push_back(IVIncInsertPos);
}
// The expansion must also be dominated by the increment positions of any
// loops it for which it is using post-inc mode.
for (const Loop *PIL : LF.PostIncLoops) {
  if (PIL == L) continue;

  // Be dominated by the loop exit.
  SmallVector<BasicBlock *, 4> ExitingBlocks;
  PIL->getExitingBlocks(ExitingBlocks);
  if (!ExitingBlocks.empty()) {
    BasicBlock *BB = ExitingBlocks[0];
    for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
      BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
    Inputs.push_back(BB->getTerminator());
  }
}

assert(!isa<PHINode>(LowestIP) && !LowestIP->isEHPad()(static_cast <bool> (!isa<PHINode>(LowestIP) &&
 !LowestIP->isEHPad() && !isa<DbgInfoIntrinsic>
(LowestIP) && "Insertion point must be a normal instruction"
) ? void (0) : __assert_fail ("!isa<PHINode>(LowestIP) && !LowestIP->isEHPad() && !isa<DbgInfoIntrinsic>(LowestIP) && \"Insertion point must be a normal instruction\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 5333, __extension__
 __PRETTY_FUNCTION__))
       && !isa<DbgInfoIntrinsic>(LowestIP) &&(static_cast <bool> (!isa<PHINode>(LowestIP) &&
 !LowestIP->isEHPad() && !isa<DbgInfoIntrinsic>
(LowestIP) && "Insertion point must be a normal instruction"
) ? void (0) : __assert_fail ("!isa<PHINode>(LowestIP) && !LowestIP->isEHPad() && !isa<DbgInfoIntrinsic>(LowestIP) && \"Insertion point must be a normal instruction\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 5333, __extension__
 __PRETTY_FUNCTION__))
       "Insertion point must be a normal instruction")(static_cast <bool> (!isa<PHINode>(LowestIP) &&
 !LowestIP->isEHPad() && !isa<DbgInfoIntrinsic>
(LowestIP) && "Insertion point must be a normal instruction"
) ? void (0) : __assert_fail ("!isa<PHINode>(LowestIP) && !LowestIP->isEHPad() && !isa<DbgInfoIntrinsic>(LowestIP) && \"Insertion point must be a normal instruction\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 5333, __extension__
 __PRETTY_FUNCTION__));

// Then, climb up the immediate dominator tree as far as we can go while
// still being dominated by the input positions.
BasicBlock::iterator IP = HoistInsertPosition(LowestIP, Inputs);

// Don't insert instructions before PHI nodes.
while (isa<PHINode>(IP)) ++IP;

// Ignore landingpad instructions.
while (IP->isEHPad()) ++IP;

// Ignore debug intrinsics.
while (isa<DbgInfoIntrinsic>(IP)) ++IP;

// Set IP below instructions recently inserted by SCEVExpander. This keeps the
// IP consistent across expansions and allows the previously inserted
// instructions to be reused by subsequent expansion.
while (Rewriter.isInsertedInstruction(&*IP) && IP != LowestIP)
  ++IP;

return IP;
5355}

5357/// Emit instructions for the leading candidate expression for this LSRUse (this
5358/// is called "expanding").
5359Value *LSRInstance::Expand(const LSRUse &LU, const LSRFixup &LF,
                         const Formula &F, BasicBlock::iterator IP,
                         SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
if (LU.RigidFormula)
  return LF.OperandValToReplace;

// Determine an input position which will be dominated by the operands and
// which will dominate the result.
IP = AdjustInsertPositionForExpand(IP, LF, LU);
Rewriter.setInsertPoint(&*IP);

// Inform the Rewriter if we have a post-increment use, so that it can
// perform an advantageous expansion.
Rewriter.setPostInc(LF.PostIncLoops);

// This is the type that the user actually needs.
Type *OpTy = LF.OperandValToReplace->getType();
// This will be the type that we'll initially expand to.
Type *Ty = F.getType();
if (!Ty)
  // No type known; just expand directly to the ultimate type.
  Ty = OpTy;
else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
  // Expand directly to the ultimate type if it's the right size.
  Ty = OpTy;
// This is the type to do integer arithmetic in.
Type *IntTy = SE.getEffectiveSCEVType(Ty);

// Build up a list of operands to add together to form the full base.
SmallVector<const SCEV *, 8> Ops;

// Expand the BaseRegs portion.
for (const SCEV *Reg : F.BaseRegs) {
  assert(!Reg->isZero() && "Zero allocated in a base register!")(static_cast <bool> (!Reg->isZero() && "Zero allocated in a base register!"
) ? void (0) : __assert_fail ("!Reg->isZero() && \"Zero allocated in a base register!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 5392, __extension__
 __PRETTY_FUNCTION__));

  // If we're expanding for a post-inc user, make the post-inc adjustment.
  Reg = denormalizeForPostIncUse(Reg, LF.PostIncLoops, SE);
  Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, nullptr)));
}

// Expand the ScaledReg portion.
Value *ICmpScaledV = nullptr;
if (F.Scale != 0) {
  const SCEV *ScaledS = F.ScaledReg;

  // If we're expanding for a post-inc user, make the post-inc adjustment.
  PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
  ScaledS = denormalizeForPostIncUse(ScaledS, Loops, SE);

  if (LU.Kind == LSRUse::ICmpZero) {
    // Expand ScaleReg as if it was part of the base regs.
    if (F.Scale == 1)
      Ops.push_back(
          SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr)));
    else {
      // An interesting way of "folding" with an icmp is to use a negated
      // scale, which we'll implement by inserting it into the other operand
      // of the icmp.
      assert(F.Scale == -1 &&(static_cast <bool> (F.Scale == -1 && "The only scale supported by ICmpZero uses is -1!"
) ? void (0) : __assert_fail ("F.Scale == -1 && \"The only scale supported by ICmpZero uses is -1!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 5418, __extension__
 __PRETTY_FUNCTION__))
             "The only scale supported by ICmpZero uses is -1!")(static_cast <bool> (F.Scale == -1 && "The only scale supported by ICmpZero uses is -1!"
) ? void (0) : __assert_fail ("F.Scale == -1 && \"The only scale supported by ICmpZero uses is -1!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 5418, __extension__
 __PRETTY_FUNCTION__));
      ICmpScaledV = Rewriter.expandCodeFor(ScaledS, nullptr);
    }
  } else {
    // Otherwise just expand the scaled register and an explicit scale,
    // which is expected to be matched as part of the address.

    // Flush the operand list to suppress SCEVExpander hoisting address modes.
    // Unless the addressing mode will not be folded.
    if (!Ops.empty() && LU.Kind == LSRUse::Address &&
        isAMCompletelyFolded(TTI, LU, F)) {
      Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), nullptr);
      Ops.clear();
      Ops.push_back(SE.getUnknown(FullV));
    }
    ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr));
    if (F.Scale != 1)
      ScaledS =
          SE.getMulExpr(ScaledS, SE.getConstant(ScaledS->getType(), F.Scale));
    Ops.push_back(ScaledS);
  }
}

// Expand the GV portion.
if (F.BaseGV) {
  // Flush the operand list to suppress SCEVExpander hoisting.
  if (!Ops.empty()) {
    Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), IntTy);
    Ops.clear();
    Ops.push_back(SE.getUnknown(FullV));
  }
  Ops.push_back(SE.getUnknown(F.BaseGV));
}

// Flush the operand list to suppress SCEVExpander hoisting of both folded and
// unfolded offsets. LSR assumes they both live next to their uses.
if (!Ops.empty()) {
  Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty);
  Ops.clear();
  Ops.push_back(SE.getUnknown(FullV));
}

// Expand the immediate portion.
int64_t Offset = (uint64_t)F.BaseOffset + LF.Offset;
if (Offset != 0) {
  if (LU.Kind == LSRUse::ICmpZero) {
    // The other interesting way of "folding" with an ICmpZero is to use a
    // negated immediate.
    if (!ICmpScaledV)
      ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset);
    else {
      Ops.push_back(SE.getUnknown(ICmpScaledV));
      ICmpScaledV = ConstantInt::get(IntTy, Offset);
    }
  } else {
    // Just add the immediate values. These again are expected to be matched
    // as part of the address.
    Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
  }
}

// Expand the unfolded offset portion.
int64_t UnfoldedOffset = F.UnfoldedOffset;
if (UnfoldedOffset != 0) {
  // Just add the immediate values.
  Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy,
                                                     UnfoldedOffset)));
}

// Emit instructions summing all the operands.
const SCEV *FullS = Ops.empty() ?
                    SE.getConstant(IntTy, 0) :
                    SE.getAddExpr(Ops);
Value *FullV = Rewriter.expandCodeFor(FullS, Ty);

// We're done expanding now, so reset the rewriter.
Rewriter.clearPostInc();

// An ICmpZero Formula represents an ICmp which we're handling as a
// comparison against zero. Now that we've expanded an expression for that
// form, update the ICmp's other operand.
if (LU.Kind == LSRUse::ICmpZero) {
  ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
  if (auto *OperandIsInstr = dyn_cast<Instruction>(CI->getOperand(1)))
    DeadInsts.emplace_back(OperandIsInstr);
  assert(!F.BaseGV && "ICmp does not support folding a global value and "(static_cast <bool> (!F.BaseGV && "ICmp does not support folding a global value and "
 "a scale at the same time!") ? void (0) : __assert_fail ("!F.BaseGV && \"ICmp does not support folding a global value and \" \"a scale at the same time!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 5504, __extension__
 __PRETTY_FUNCTION__))
                         "a scale at the same time!")(static_cast <bool> (!F.BaseGV && "ICmp does not support folding a global value and "
 "a scale at the same time!") ? void (0) : __assert_fail ("!F.BaseGV && \"ICmp does not support folding a global value and \" \"a scale at the same time!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 5504, __extension__
 __PRETTY_FUNCTION__));
  if (F.Scale == -1) {
    if (ICmpScaledV->getType() != OpTy) {
      Instruction *Cast =
        CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
                                                 OpTy, false),
                         ICmpScaledV, OpTy, "tmp", CI);
      ICmpScaledV = Cast;
    }
    CI->setOperand(1, ICmpScaledV);
  } else {
    // A scale of 1 means that the scale has been expanded as part of the
    // base regs.
    assert((F.Scale == 0 || F.Scale == 1) &&(static_cast <bool> ((F.Scale == 0 || F.Scale == 1) &&
 "ICmp does not support folding a global value and " "a scale at the same time!"
) ? void (0) : __assert_fail ("(F.Scale == 0 || F.Scale == 1) && \"ICmp does not support folding a global value and \" \"a scale at the same time!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 5519, __extension__
 __PRETTY_FUNCTION__))
           "ICmp does not support folding a global value and "(static_cast <bool> ((F.Scale == 0 || F.Scale == 1) &&
 "ICmp does not support folding a global value and " "a scale at the same time!"
) ? void (0) : __assert_fail ("(F.Scale == 0 || F.Scale == 1) && \"ICmp does not support folding a global value and \" \"a scale at the same time!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 5519, __extension__
 __PRETTY_FUNCTION__))
           "a scale at the same time!")(static_cast <bool> ((F.Scale == 0 || F.Scale == 1) &&
 "ICmp does not support folding a global value and " "a scale at the same time!"
) ? void (0) : __assert_fail ("(F.Scale == 0 || F.Scale == 1) && \"ICmp does not support folding a global value and \" \"a scale at the same time!\""
, "llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp", 5519, __extension__
 __PRETTY_FUNCTION__));
    Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
                                         -(uint64_t)Offset);
    if (C->getType() != OpTy)
      C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
                                                        OpTy, false),
                                C, OpTy);

    CI->setOperand(1, C);
  }
}

return FullV;
5532}

5534/// Helper for Rewrite. PHI nodes are special because the use of their operands
5535/// effectively happens in their predecessor blocks, so the expression may need
5536/// to be expanded in multiple places.
5537void LSRInstance::RewriteForPHI(
  PHINode *PN, const LSRUse &LU, const LSRFixup &LF, const Formula &F,
  SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
DenseMap<BasicBlock *, Value *> Inserted;
for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1
Assuming 'i' is not equal to 'e'→
2
←
Loop condition is true.  Entering loop body→
  if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
3
←
Assuming pointer value is null→
4
←
Taking true branch→
    bool needUpdateFixups = false;
    BasicBlock *BB = PN->getIncomingBlock(i);

    // If this is a critical edge, split the edge so that we do not insert
    // the code on all predecessor/successor paths.  We do this unless this
    // is the canonical backedge for this loop, which complicates post-inc
    // users.
    if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
5
←
Assuming 'e' is not equal to 1→
6
←
Assuming the condition is false→
7
←
Taking false branch→
        !isa<IndirectBrInst>(BB->getTerminator()) &&
        !isa<CatchSwitchInst>(BB->getTerminator())) {
      BasicBlock *Parent = PN->getParent();
      Loop *PNLoop = LI.getLoopFor(Parent);
      if (!PNLoop || Parent != PNLoop->getHeader()) {
        // Split the critical edge.
        BasicBlock *NewBB = nullptr;
        if (!Parent->isLandingPad()) {
          NewBB =
              SplitCriticalEdge(BB, Parent,
                                CriticalEdgeSplittingOptions(&DT, &LI, MSSAU)
                                    .setMergeIdenticalEdges()
                                    .setKeepOneInputPHIs());
        } else {
          SmallVector<BasicBlock*, 2> NewBBs;
          SplitLandingPadPredecessors(Parent, BB, "", "", NewBBs, &DT, &LI);
          NewBB = NewBBs[0];
        }
        // If NewBB==NULL, then SplitCriticalEdge refused to split because all
        // phi predecessors are identical. The simple thing to do is skip
        // splitting in this case rather than complicate the API.
        if (NewBB) {
          // If PN is outside of the loop and BB is in the loop, we want to
          // move the block to be immediately before the PHI block, not
          // immediately after BB.
          if (L->contains(BB) && !L->contains(PN))
            NewBB->moveBefore(PN->getParent());

          // Splitting the edge can reduce the number of PHI entries we have.
          e = PN->getNumIncomingValues();
          BB = NewBB;
          i = PN->getBasicBlockIndex(BB);

          needUpdateFixups = true;
        }
      }
    }

    std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
      Inserted.insert(std::make_pair(BB, static_cast<Value *>(nullptr)));
    if (!Pair.second)
8
←
Assuming field 'second' is true→
9
←
Taking false branch→
      PN->setIncomingValue(i, Pair.first->second);
    else {
      Value *FullV =
          Expand(LU, LF, F, BB->getTerminator()->getIterator(), DeadInsts);

      // If this is reuse-by-noop-cast, insert the noop cast.
      Type *OpTy = LF.OperandValToReplace->getType();
10
←
Called C++ object pointer is null
      if (FullV->getType() != OpTy)
        FullV =
          CastInst::Create(CastInst::getCastOpcode(FullV, false,
                                                   OpTy, false),
                           FullV, LF.OperandValToReplace->getType(),
                           "tmp", BB->getTerminator());

      PN->setIncomingValue(i, FullV);
      Pair.first->second = FullV;
    }

    // If LSR splits critical edge and phi node has other pending
    // fixup operands, we need to update those pending fixups. Otherwise
    // formulae will not be implemented completely and some instructions
    // will not be eliminated.
    if (needUpdateFixups) {
      for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx)
        for (LSRFixup &Fixup : Uses[LUIdx].Fixups)
          // If fixup is supposed to rewrite some operand in the phi
          // that was just updated, it may be already moved to
          // another phi node. Such fixup requires update.
          if (Fixup.UserInst == PN) {
            // Check if the operand we try to replace still exists in the
            // original phi.
            bool foundInOriginalPHI = false;
            for (const auto &val : PN->incoming_values())
              if (val == Fixup.OperandValToReplace) {
                foundInOriginalPHI = true;
                break;
              }

            // If fixup operand found in original PHI - nothing to do.
            if (foundInOriginalPHI)
              continue;

            // Otherwise it might be moved to another PHI and requires update.
            // If fixup operand not found in any of the incoming blocks that
            // means we have already rewritten it - nothing to do.
            for (const auto &Block : PN->blocks())
              for (BasicBlock::iterator I = Block->begin(); isa<PHINode>(I);
                   ++I) {
                PHINode *NewPN = cast<PHINode>(I);
                for (const auto &val : NewPN->incoming_values())
                  if (val == Fixup.OperandValToReplace)
                    Fixup.UserInst = NewPN;
              }
          }
    }
  }
5648}

5650/// Emit instructions for the leading candidate expression for this LSRUse (this
5651/// is called "expanding"), and update the UserInst to reference the newly
5652/// expanded value.
5653void LSRInstance::Rewrite(const LSRUse &LU, const LSRFixup &LF,
                        const Formula &F,
                        SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
// First, find an insertion point that dominates UserInst. For PHI nodes,
// find the nearest block which dominates all the relevant uses.
if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
  RewriteForPHI(PN, LU, LF, F, DeadInsts);
} else {
  Value *FullV = Expand(LU, LF, F, LF.UserInst->getIterator(), DeadInsts);

  // If this is reuse-by-noop-cast, insert the noop cast.
  Type *OpTy = LF.OperandValToReplace->getType();
  if (FullV->getType() != OpTy) {
    Instruction *Cast =
      CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
                       FullV, OpTy, "tmp", LF.UserInst);