llvm-capstone/libcxx/benchmarks/string.bench.cpp
Samuel Benzaquen 27a83e99f4 Add more benchmarks for literal strings.
Summary:
Comparing against the empty string should generate much better code that
what it does today.
We can also generate better code when comparing against literals that
are larger than the SSO space.

Reviewers: EricWF

Subscribers: christof, jdoerfert, libcxx-commits

Tags: #libc

Differential Revision: https://reviews.llvm.org/D59781

llvm-svn: 357614
2019-04-03 17:40:51 +00:00

448 lines
15 KiB
C++

#include <cstdint>
#include <new>
#include <vector>
#include "CartesianBenchmarks.hpp"
#include "GenerateInput.hpp"
#include "benchmark/benchmark.h"
#include "test_macros.h"
constexpr std::size_t MAX_STRING_LEN = 8 << 14;
// Benchmark when there is no match.
static void BM_StringFindNoMatch(benchmark::State &state) {
std::string s1(state.range(0), '-');
std::string s2(8, '*');
for (auto _ : state)
benchmark::DoNotOptimize(s1.find(s2));
}
BENCHMARK(BM_StringFindNoMatch)->Range(10, MAX_STRING_LEN);
// Benchmark when the string matches first time.
static void BM_StringFindAllMatch(benchmark::State &state) {
std::string s1(MAX_STRING_LEN, '-');
std::string s2(state.range(0), '-');
for (auto _ : state)
benchmark::DoNotOptimize(s1.find(s2));
}
BENCHMARK(BM_StringFindAllMatch)->Range(1, MAX_STRING_LEN);
// Benchmark when the string matches somewhere in the end.
static void BM_StringFindMatch1(benchmark::State &state) {
std::string s1(MAX_STRING_LEN / 2, '*');
s1 += std::string(state.range(0), '-');
std::string s2(state.range(0), '-');
for (auto _ : state)
benchmark::DoNotOptimize(s1.find(s2));
}
BENCHMARK(BM_StringFindMatch1)->Range(1, MAX_STRING_LEN / 4);
// Benchmark when the string matches somewhere from middle to the end.
static void BM_StringFindMatch2(benchmark::State &state) {
std::string s1(MAX_STRING_LEN / 2, '*');
s1 += std::string(state.range(0), '-');
s1 += std::string(state.range(0), '*');
std::string s2(state.range(0), '-');
for (auto _ : state)
benchmark::DoNotOptimize(s1.find(s2));
}
BENCHMARK(BM_StringFindMatch2)->Range(1, MAX_STRING_LEN / 4);
static void BM_StringCtorDefault(benchmark::State &state) {
for (auto _ : state) {
std::string Default;
benchmark::DoNotOptimize(Default);
}
}
BENCHMARK(BM_StringCtorDefault);
enum class Length { Empty, Small, Large, Huge };
struct AllLengths : EnumValuesAsTuple<AllLengths, Length, 4> {
static constexpr const char* Names[] = {"Empty", "Small", "Large", "Huge"};
};
enum class Opacity { Opaque, Transparent };
struct AllOpacity : EnumValuesAsTuple<AllOpacity, Opacity, 2> {
static constexpr const char* Names[] = {"Opaque", "Transparent"};
};
enum class DiffType { Control, ChangeFirst, ChangeMiddle, ChangeLast };
struct AllDiffTypes : EnumValuesAsTuple<AllDiffTypes, DiffType, 4> {
static constexpr const char* Names[] = {"Control", "ChangeFirst",
"ChangeMiddle", "ChangeLast"};
};
static constexpr char SmallStringLiteral[] = "012345678";
TEST_ALWAYS_INLINE const char* getSmallString(DiffType D) {
switch (D) {
case DiffType::Control:
return SmallStringLiteral;
case DiffType::ChangeFirst:
return "-12345678";
case DiffType::ChangeMiddle:
return "0123-5678";
case DiffType::ChangeLast:
return "01234567-";
}
}
static constexpr char LargeStringLiteral[] =
"012345678901234567890123456789012345678901234567890123456789012";
TEST_ALWAYS_INLINE const char* getLargeString(DiffType D) {
#define LARGE_STRING_FIRST "123456789012345678901234567890"
#define LARGE_STRING_SECOND "234567890123456789012345678901"
switch (D) {
case DiffType::Control:
return "0" LARGE_STRING_FIRST "1" LARGE_STRING_SECOND "2";
case DiffType::ChangeFirst:
return "-" LARGE_STRING_FIRST "1" LARGE_STRING_SECOND "2";
case DiffType::ChangeMiddle:
return "0" LARGE_STRING_FIRST "-" LARGE_STRING_SECOND "2";
case DiffType::ChangeLast:
return "0" LARGE_STRING_FIRST "1" LARGE_STRING_SECOND "-";
}
}
TEST_ALWAYS_INLINE const char* getHugeString(DiffType D) {
#define HUGE_STRING0 "0123456789"
#define HUGE_STRING1 HUGE_STRING0 HUGE_STRING0 HUGE_STRING0 HUGE_STRING0
#define HUGE_STRING2 HUGE_STRING1 HUGE_STRING1 HUGE_STRING1 HUGE_STRING1
#define HUGE_STRING3 HUGE_STRING2 HUGE_STRING2 HUGE_STRING2 HUGE_STRING2
#define HUGE_STRING4 HUGE_STRING3 HUGE_STRING3 HUGE_STRING3 HUGE_STRING3
switch (D) {
case DiffType::Control:
return "0123456789" HUGE_STRING4 "0123456789" HUGE_STRING4 "0123456789";
case DiffType::ChangeFirst:
return "-123456789" HUGE_STRING4 "0123456789" HUGE_STRING4 "0123456789";
case DiffType::ChangeMiddle:
return "0123456789" HUGE_STRING4 "01234-6789" HUGE_STRING4 "0123456789";
case DiffType::ChangeLast:
return "0123456789" HUGE_STRING4 "0123456789" HUGE_STRING4 "012345678-";
}
}
TEST_ALWAYS_INLINE std::string makeString(Length L,
DiffType D = DiffType::Control,
Opacity O = Opacity::Transparent) {
switch (L) {
case Length::Empty:
return maybeOpaque("", O == Opacity::Opaque);
case Length::Small:
return maybeOpaque(getSmallString(D), O == Opacity::Opaque);
case Length::Large:
return maybeOpaque(getLargeString(D), O == Opacity::Opaque);
case Length::Huge:
return maybeOpaque(getHugeString(D), O == Opacity::Opaque);
}
}
template <class Length, class Opaque>
struct StringConstructDestroyCStr {
static void run(benchmark::State& state) {
for (auto _ : state) {
benchmark::DoNotOptimize(
makeString(Length(), DiffType::Control, Opaque()));
}
}
static std::string name() {
return "BM_StringConstructDestroyCStr" + Length::name() + Opaque::name();
}
};
template <class Length, bool MeasureCopy, bool MeasureDestroy>
static void StringCopyAndDestroy(benchmark::State& state) {
static constexpr size_t NumStrings = 1024;
auto Orig = makeString(Length());
std::aligned_storage<sizeof(std::string)>::type Storage[NumStrings];
while (state.KeepRunningBatch(NumStrings)) {
if (!MeasureCopy)
state.PauseTiming();
for (size_t I = 0; I < NumStrings; ++I) {
::new (static_cast<void*>(Storage + I)) std::string(Orig);
}
if (!MeasureCopy)
state.ResumeTiming();
if (!MeasureDestroy)
state.PauseTiming();
for (size_t I = 0; I < NumStrings; ++I) {
using S = std::string;
reinterpret_cast<S*>(Storage + I)->~S();
}
if (!MeasureDestroy)
state.ResumeTiming();
}
}
template <class Length>
struct StringCopy {
static void run(benchmark::State& state) {
StringCopyAndDestroy<Length, true, false>(state);
}
static std::string name() { return "BM_StringCopy" + Length::name(); }
};
template <class Length>
struct StringDestroy {
static void run(benchmark::State& state) {
StringCopyAndDestroy<Length, false, true>(state);
}
static std::string name() { return "BM_StringDestroy" + Length::name(); }
};
template <class Length>
struct StringMove {
static void run(benchmark::State& state) {
// Keep two object locations and move construct back and forth.
std::aligned_storage<sizeof(std::string), alignof(std::string)>::type Storage[2];
using S = std::string;
size_t I = 0;
S *newS = new (static_cast<void*>(Storage)) std::string(makeString(Length()));
for (auto _ : state) {
// Switch locations.
I ^= 1;
benchmark::DoNotOptimize(Storage);
// Move construct into the new location,
S *tmpS = new (static_cast<void*>(Storage + I)) S(std::move(*newS));
// then destroy the old one.
newS->~S();
newS = tmpS;
}
newS->~S();
}
static std::string name() { return "BM_StringMove" + Length::name(); }
};
enum class Relation { Eq, Less, Compare };
struct AllRelations : EnumValuesAsTuple<AllRelations, Relation, 3> {
static constexpr const char* Names[] = {"Eq", "Less", "Compare"};
};
template <class Rel, class LHLength, class RHLength, class DiffType>
struct StringRelational {
static void run(benchmark::State& state) {
auto Lhs = makeString(RHLength());
auto Rhs = makeString(LHLength(), DiffType());
for (auto _ : state) {
benchmark::DoNotOptimize(Lhs);
benchmark::DoNotOptimize(Rhs);
switch (Rel()) {
case Relation::Eq:
benchmark::DoNotOptimize(Lhs == Rhs);
break;
case Relation::Less:
benchmark::DoNotOptimize(Lhs < Rhs);
break;
case Relation::Compare:
benchmark::DoNotOptimize(Lhs.compare(Rhs));
break;
}
}
}
static bool skip() {
// Eq is commutative, so skip half the matrix.
if (Rel() == Relation::Eq && LHLength() > RHLength())
return true;
// We only care about control when the lengths differ.
if (LHLength() != RHLength() && DiffType() != ::DiffType::Control)
return true;
// For empty, only control matters.
if (LHLength() == Length::Empty && DiffType() != ::DiffType::Control)
return true;
return false;
}
static std::string name() {
return "BM_StringRelational" + Rel::name() + LHLength::name() +
RHLength::name() + DiffType::name();
}
};
template <class Rel, class LHLength, class RHLength, class DiffType>
struct StringRelationalLiteral {
static void run(benchmark::State& state) {
auto Lhs = makeString(LHLength(), DiffType());
for (auto _ : state) {
benchmark::DoNotOptimize(Lhs);
constexpr const char* Literal = RHLength::value == Length::Empty
? ""
: RHLength::value == Length::Small
? SmallStringLiteral
: LargeStringLiteral;
switch (Rel()) {
case Relation::Eq:
benchmark::DoNotOptimize(Lhs == Literal);
break;
case Relation::Less:
benchmark::DoNotOptimize(Lhs < Literal);
break;
case Relation::Compare:
benchmark::DoNotOptimize(Lhs.compare(Literal));
break;
}
}
}
static bool skip() {
// Doesn't matter how they differ if they have different size.
if (LHLength() != RHLength() && DiffType() != ::DiffType::Control)
return true;
// We don't need huge. Doensn't give anything different than Large.
if (LHLength() == Length::Huge || RHLength() == Length::Huge)
return true;
return false;
}
static std::string name() {
return "BM_StringRelationalLiteral" + Rel::name() + LHLength::name() +
RHLength::name() + DiffType::name();
}
};
enum class Depth { Shallow, Deep };
struct AllDepths : EnumValuesAsTuple<AllDepths, Depth, 2> {
static constexpr const char* Names[] = {"Shallow", "Deep"};
};
enum class Temperature { Hot, Cold };
struct AllTemperatures : EnumValuesAsTuple<AllTemperatures, Temperature, 2> {
static constexpr const char* Names[] = {"Hot", "Cold"};
};
template <class Temperature, class Depth, class Length>
struct StringRead {
void run(benchmark::State& state) const {
static constexpr size_t NumStrings =
Temperature() == ::Temperature::Hot
? 1 << 10
: /* Enough strings to overflow the cache */ 1 << 20;
static_assert((NumStrings & (NumStrings - 1)) == 0,
"NumStrings should be a power of two to reduce overhead.");
std::vector<std::string> Values(NumStrings, makeString(Length()));
size_t I = 0;
for (auto _ : state) {
// Jump long enough to defeat cache locality, and use a value that is
// coprime with NumStrings to ensure we visit every element.
I = (I + 17) % NumStrings;
const auto& V = Values[I];
// Read everything first. Escaping data() through DoNotOptimize might
// cause the compiler to have to recalculate information about `V` due to
// aliasing.
const char* const Data = V.data();
const size_t Size = V.size();
benchmark::DoNotOptimize(Data);
benchmark::DoNotOptimize(Size);
if (Depth() == ::Depth::Deep) {
// Read into the payload. This mainly shows the benefit of SSO when the
// data is cold.
benchmark::DoNotOptimize(*Data);
}
}
}
static bool skip() {
// Huge does not give us anything that Large doesn't have. Skip it.
if (Length() == ::Length::Huge) {
return true;
}
return false;
}
std::string name() const {
return "BM_StringRead" + Temperature::name() + Depth::name() +
Length::name();
}
};
void sanityCheckGeneratedStrings() {
for (auto Lhs : {Length::Empty, Length::Small, Length::Large, Length::Huge}) {
const auto LhsString = makeString(Lhs);
for (auto Rhs :
{Length::Empty, Length::Small, Length::Large, Length::Huge}) {
if (Lhs > Rhs)
continue;
const auto RhsString = makeString(Rhs);
// The smaller one must be a prefix of the larger one.
if (RhsString.find(LhsString) != 0) {
fprintf(stderr, "Invalid autogenerated strings for sizes (%d,%d).\n",
static_cast<int>(Lhs), static_cast<int>(Rhs));
std::abort();
}
}
}
// Verify the autogenerated diffs
for (auto L : {Length::Small, Length::Large, Length::Huge}) {
const auto Control = makeString(L);
const auto Verify = [&](std::string Exp, size_t Pos) {
// Only change on the Pos char.
if (Control[Pos] != Exp[Pos]) {
Exp[Pos] = Control[Pos];
if (Control == Exp)
return;
}
fprintf(stderr, "Invalid autogenerated diff with size %d\n",
static_cast<int>(L));
std::abort();
};
Verify(makeString(L, DiffType::ChangeFirst), 0);
Verify(makeString(L, DiffType::ChangeMiddle), Control.size() / 2);
Verify(makeString(L, DiffType::ChangeLast), Control.size() - 1);
}
}
// Some small codegen thunks to easily see generated code.
bool StringEqString(const std::string& a, const std::string& b) {
return a == b;
}
bool StringEqCStr(const std::string& a, const char* b) { return a == b; }
bool CStrEqString(const char* a, const std::string& b) { return a == b; }
bool StringEqCStrLiteralEmpty(const std::string& a) {
return a == "";
}
bool StringEqCStrLiteralSmall(const std::string& a) {
return a == SmallStringLiteral;
}
bool StringEqCStrLiteralLarge(const std::string& a) {
return a == LargeStringLiteral;
}
int main(int argc, char** argv) {
benchmark::Initialize(&argc, argv);
if (benchmark::ReportUnrecognizedArguments(argc, argv))
return 1;
sanityCheckGeneratedStrings();
makeCartesianProductBenchmark<StringConstructDestroyCStr, AllLengths,
AllOpacity>();
makeCartesianProductBenchmark<StringCopy, AllLengths>();
makeCartesianProductBenchmark<StringMove, AllLengths>();
makeCartesianProductBenchmark<StringDestroy, AllLengths>();
makeCartesianProductBenchmark<StringRelational, AllRelations, AllLengths,
AllLengths, AllDiffTypes>();
makeCartesianProductBenchmark<StringRelationalLiteral, AllRelations,
AllLengths, AllLengths, AllDiffTypes>();
makeCartesianProductBenchmark<StringRead, AllTemperatures, AllDepths,
AllLengths>();
benchmark::RunSpecifiedBenchmarks();
if (argc < 0) {
// ODR-use the functions to force them being generated in the binary.
auto functions = std::make_tuple(
StringEqString, StringEqCStr, CStrEqString, StringEqCStrLiteralEmpty,
StringEqCStrLiteralSmall, StringEqCStrLiteralLarge);
printf("%p", &functions);
}
}