Use 8->7 bit encoding for ascii strings in compact_object

This commit is contained in:
Roman Gershman 2022-02-24 23:19:28 +02:00
parent 7445b20548
commit e3d7f0a34e
3 changed files with 361 additions and 28 deletions

View File

@ -16,6 +16,7 @@ extern "C" {
#include <absl/strings/str_cat.h>
#include "base/logging.h"
#include "base/pod_array.h"
namespace dfly {
using namespace std;
@ -34,18 +35,73 @@ size_t QlUsedSize(quicklist* ql) {
return res;
}
// Deniel's Lemire function validate_ascii_fast() - under Apache/MIT license.
// See https://github.com/lemire/fastvalidate-utf-8/
// The function returns true (1) if all chars passed in src are
// 7-bit values (0x00..0x7F). Otherwise, it returns false (0).
bool validate_ascii_fast(const char* src, size_t len) {
size_t i = 0;
__m128i has_error = _mm_setzero_si128();
if (len >= 16) {
for (; i <= len - 16; i += 16) {
__m128i current_bytes = _mm_loadu_si128((const __m128i*)(src + i));
has_error = _mm_or_si128(has_error, current_bytes);
}
}
int error_mask = _mm_movemask_epi8(has_error);
char tail_has_error = 0;
for (; i < len; i++) {
tail_has_error |= src[i];
}
error_mask |= (tail_has_error & 0x80);
return !error_mask;
}
// maps ascii len to 7-bit packed length. Each 8 bytes are converted to 7 bytes.
inline constexpr size_t binpacked_len(size_t ascii_len) {
return (ascii_len * 7 + 7) / 8; /* rounded up */
}
// converts 7-bit packed length back to ascii length. Note that this conversion
// is not accurate since it maps 7 bytes to 8 bytes (rounds up), while we may have
// 7 byte strings converted to 7 byte as well.
inline constexpr size_t ascii_len(size_t bin_len) {
return (bin_len * 8) / 7;
}
inline const uint8_t* to_byte(const void* s) {
return reinterpret_cast<const uint8_t*>(s);
}
static_assert(binpacked_len(7) == 7);
static_assert(binpacked_len(8) == 7);
static_assert(binpacked_len(15) == 14);
static_assert(binpacked_len(16) == 14);
static_assert(binpacked_len(17) == 15);
static_assert(binpacked_len(18) == 16);
static_assert(binpacked_len(19) == 17);
static_assert(binpacked_len(20) == 18);
static_assert(ascii_len(14) == 16);
static_assert(ascii_len(15) == 17);
static_assert(ascii_len(16) == 18);
static_assert(ascii_len(17) == 19);
struct TL {
robj tmp_robj{
.type = 0, .encoding = 0, .lru = 0, .refcount = OBJ_STATIC_REFCOUNT, .ptr = nullptr};
pmr::memory_resource* local_mr = pmr::get_default_resource();
size_t small_str_bytes;
base::PODArray<uint8_t> tmp_buf;
string tmp_str;
};
thread_local TL tl;
constexpr bool kUseSmallStrings = true;
constexpr bool kUseAsciiEncoding = true;
} // namespace
static_assert(sizeof(CompactObj) == 18);
@ -205,6 +261,84 @@ bool RobjWrapper::Equal(std::string_view sv) const {
return blob.AsView() == sv;
}
// len must be at least 16
void ascii_pack(const char* ascii, size_t len, uint8_t* bin) {
unsigned i = 0;
while (len >= 8) {
for (i = 0; i < 7; ++i) {
*bin++ = (ascii[0] >> i) | (ascii[1] << (7 - i));
++ascii;
}
++ascii;
len -= 8;
}
for (i = 0; i < len; ++i) {
*bin++ = (ascii[i] >> i) | (ascii[i + 1] << (7 - i));
}
}
// unpacks 8->7 encoded blob back to ascii.
// generally, we can not unpack inplace because ascii (dest) buffer is 8/7 bigger than
// the source buffer.
// however, if binary data is positioned on the right of the ascii buffer with empty space on the
// left than we can unpack inplace.
void ascii_unpack(const uint8_t* bin, size_t ascii_len, char* ascii) {
constexpr uint8_t kM = 0x7F;
uint8_t p = 0;
unsigned i = 0;
auto step = [&] {
uint8_t src = *bin; // keep on stack in case we unpack inplace.
*ascii++ = (p >> (8 - i)) | ((src << i) & kM);
p = src;
++bin;
};
while (ascii_len >= 8) {
for (i = 0; i < 7; ++i) {
step();
}
ascii_len -= 8;
*ascii++ = p >> 1;
}
for (i = 0; i < ascii_len; ++i) {
uint8_t src = *bin;
*ascii++ = (p >> (8 - i)) | ((src << i) & kM);
p = src;
++bin;
}
}
// compares packed and unpacked strings. packed must be of length = binpacked_len(ascii_len).
bool compare_packed(const uint8_t* packed, const char* ascii, size_t ascii_len) {
unsigned i = 0;
bool res = true;
while (ascii_len >= 8) {
for (i = 0; i < 7; ++i) {
uint8_t conv = (ascii[0] >> i) | (ascii[1] << (7 - i));
res &= (conv == *packed);
++ascii;
++packed;
}
if (!res)
return false;
++ascii;
ascii_len -= 8;
}
for (i = 0; i < ascii_len; ++i) {
uint8_t b = (ascii[i] >> i) | (ascii[i + 1] << (7 - i));
res &= (b == *packed);
++packed;
}
return res;
}
} // namespace detail
using namespace std;
@ -218,6 +352,7 @@ auto CompactObj::GetStats() -> Stats {
void CompactObj::InitThreadLocal(pmr::memory_resource* mr) {
tl.local_mr = mr;
tl.tmp_buf = base::PODArray<uint8_t>{mr};
SmallString::InitThreadLocal();
}
@ -256,10 +391,22 @@ size_t CompactObj::StrSize() const {
}
uint64_t CompactObj::HashCode() const {
uint8_t encoded = (mask_ & kEncMask);
if (IsInline()) {
if (encoded) {
char buf[kInlineLen * 2];
detail::ascii_unpack(to_byte(u_.inline_str), taglen_, buf);
return XXH3_64bits_withSeed(buf, DecodedLen(taglen_), kHashSeed);
}
return XXH3_64bits_withSeed(u_.inline_str, taglen_, kHashSeed);
}
if (encoded) {
GetString(&tl.tmp_str);
return XXH3_64bits_withSeed(tl.tmp_str.data(), tl.tmp_str.size(), kHashSeed);
}
switch (taglen_) {
case SMALL_TAG:
return u_.small_str.HashCode();
@ -285,7 +432,7 @@ unsigned CompactObj::ObjType() const {
if (taglen_ == ROBJ_TAG)
return u_.r_obj.type;
LOG(FATAL) << "TBD " << taglen_;
LOG(FATAL) << "TBD " << int(taglen_);
return 0;
}
@ -390,37 +537,108 @@ void CompactObj::SetString(std::string_view str) {
}
}
std::string_view input = str;
if (str.size() <= kInlineLen) {
SetMeta(str.size(), 0);
return;
}
if (kUseSmallStrings && taglen_ == 0 && str.size() < (1 << 15)) {
u_.small_str.Reset();
SetMeta(SMALL_TAG, 0);
u_.small_str.Assign(str);
string_view encoded = str;
uint8_t mask = 0;
bool is_ascii = kUseAsciiEncoding && validate_ascii_fast(str.data(), str.size());
if (is_ascii) {
size_t encode_len = binpacked_len(str.size());
size_t rev_len = ascii_len(encode_len);
CHECK_GE(rev_len, str.size() - 1) << "Bad ascii encoding for len " << str.size();
if (rev_len == str.size() - 1) {
mask |= ASCII1_ENC_BIT;
} else {
mask |= ASCII2_ENC_BIT;
}
tl.tmp_buf.resize(encode_len);
detail::ascii_pack(str.data(), str.size(), tl.tmp_buf.data());
encoded = string_view{reinterpret_cast<char*>(tl.tmp_buf.data()), encode_len};
if (encoded.size() <= kInlineLen) {
SetMeta(encoded.size(), mask);
detail::ascii_pack(str.data(), str.size(), reinterpret_cast<uint8_t*>(u_.inline_str));
return;
}
}
if (kUseSmallStrings && taglen_ == 0 && encoded.size() < (1 << 15)) {
SetMeta(SMALL_TAG, mask);
u_.small_str.Assign(encoded);
tl.small_str_bytes += u_.small_str.MallocUsed();
return;
}
if (taglen_ != ROBJ_TAG || u_.r_obj.type != OBJ_STRING) {
SetMeta(ROBJ_TAG);
SetMeta(ROBJ_TAG, mask);
u_.r_obj.type = OBJ_STRING;
u_.r_obj.encoding = OBJ_ENCODING_RAW;
}
DCHECK(taglen_ == ROBJ_TAG && u_.r_obj.type == OBJ_STRING);
CHECK_EQ(OBJ_ENCODING_RAW, u_.r_obj.encoding);
u_.r_obj.blob.Assign(input, tl.local_mr);
u_.r_obj.blob.Assign(encoded, tl.local_mr);
}
std::string_view CompactObj::GetSlice(std::string* scratch) const {
string_view CompactObj::GetSlice(string* scratch) const {
uint8_t is_encoded = mask_ & kEncMask;
if (IsInline()) {
return std::string_view{u_.inline_str, taglen_};
if (is_encoded) {
size_t decoded_len = taglen_ + 2;
// must be this because we either shortened 17 or 18.
DCHECK_EQ(is_encoded, ASCII2_ENC_BIT);
DCHECK_EQ(decoded_len, ascii_len(taglen_));
scratch->resize(decoded_len);
detail::ascii_unpack(to_byte(u_.inline_str), decoded_len, scratch->data());
return *scratch;
}
return string_view{u_.inline_str, taglen_};
}
if (taglen_ == INT_TAG) {
absl::AlphaNum an(u_.ival);
scratch->assign(an.Piece());
return *scratch;
}
if (is_encoded) {
if (taglen_ == ROBJ_TAG) {
CHECK_EQ(OBJ_STRING, u_.r_obj.type);
DCHECK_EQ(OBJ_ENCODING_RAW, u_.r_obj.encoding);
size_t decoded_len = DecodedLen(u_.r_obj.blob.size());
scratch->resize(decoded_len);
detail::ascii_unpack(to_byte(u_.r_obj.blob.ptr()), decoded_len, scratch->data());
} else if (taglen_ == SMALL_TAG) {
size_t decoded_len = DecodedLen(u_.small_str.size());
size_t pref_len = decoded_len - u_.small_str.size();
scratch->resize(decoded_len);
string_view slices[2];
unsigned num = u_.small_str.GetV(slices);
DCHECK_EQ(2u, num);
char* next = scratch->data() + pref_len;
memcpy(next, slices[0].data(), slices[0].size());
next += slices[0].size();
memcpy(next, slices[1].data(), slices[1].size());
detail::ascii_unpack(reinterpret_cast<uint8_t*>(scratch->data() + pref_len), decoded_len,
scratch->data());
} else {
LOG(FATAL) << "Unsupported tag " << int(taglen_);
}
return *scratch;
}
// no encoding.
if (taglen_ == ROBJ_TAG) {
CHECK_EQ(OBJ_STRING, u_.r_obj.type);
DCHECK_EQ(OBJ_ENCODING_RAW, u_.r_obj.encoding);
@ -432,16 +650,9 @@ std::string_view CompactObj::GetSlice(std::string* scratch) const {
return *scratch;
}
if (taglen_ == INT_TAG) {
absl::AlphaNum an(u_.ival);
scratch->assign(an.Piece());
return *scratch;
}
LOG(FATAL) << "Bad tag " << int(taglen_);
return std::string_view{};
return string_view{};
}
bool CompactObj::HasAllocated() const {
@ -454,7 +665,7 @@ bool CompactObj::HasAllocated() const {
}
void CompactObj::GetString(string* res) const {
std::string_view slice = GetSlice(res);
string_view slice = GetSlice(res);
if (res->data() != slice.data()) {
res->assign(slice);
}
@ -478,7 +689,7 @@ void CompactObj::Free() {
tl.small_str_bytes -= u_.small_str.MallocUsed();
u_.small_str.Free();
} else {
LOG(FATAL) << "Bad compact object type " << int(taglen_);
LOG(FATAL) << "Unsupported tag " << int(taglen_);
}
memset(u_.inline_str, 0, kInlineLen);
@ -501,10 +712,14 @@ size_t CompactObj::MallocUsed() const {
}
bool CompactObj::operator==(const CompactObj& o) const {
uint8_t m1 = mask_ & kEncMask;
uint8_t m2 = mask_ & kEncMask;
if (m1 != m2)
return false;
if (taglen_ == ROBJ_TAG || o.taglen_ == ROBJ_TAG) {
if (o.taglen_ != taglen_)
return false;
return u_.r_obj.Equal(o.u_.r_obj);
}
@ -528,6 +743,7 @@ bool CompactObj::EqualNonInline(std::string_view sv) const {
absl::AlphaNum an(u_.ival);
return sv == an.Piece();
}
case ROBJ_TAG:
return u_.r_obj.Equal(sv);
case SMALL_TAG:
@ -538,4 +754,79 @@ bool CompactObj::EqualNonInline(std::string_view sv) const {
return false;
}
bool CompactObj::CmpEncoded(string_view sv) const {
size_t encode_len = binpacked_len(sv.size());
if (IsInline()) {
if (encode_len != taglen_)
return false;
char buf[kInlineLen * 2];
detail::ascii_unpack(to_byte(u_.inline_str), sv.size(), buf);
return sv == string_view(buf, sv.size());
}
if (taglen_ == ROBJ_TAG) {
if (u_.r_obj.type != OBJ_STRING)
return false;
if (u_.r_obj.blob.size() != encode_len)
return false;
if (!validate_ascii_fast(sv.data(), sv.size()))
return false;
return detail::compare_packed(to_byte(u_.r_obj.blob.ptr()), sv.data(), sv.size());
}
if (taglen_ == SMALL_TAG) {
if (u_.small_str.size() != encode_len)
return false;
if (!validate_ascii_fast(sv.data(), sv.size()))
return false;
// We need to compare an unpacked sv with 2 packed parts.
// To compare easily ascii with binary we would need to split ascii at 8 bytes boundaries
// so that we could pack it into complete binary bytes (8 ascii chars produce 7 bytes).
// I choose a minimal 16 byte prefix:
// 1. sv must be longer than 16 if we are here (at least 18 actually).
// 2. 16 chars produce 14 byte blob that should cover the first slice (10 bytes) and 4 bytes
// of the second slice.
// 3. I assume that the first slice is less than 14 bytes which is correct since small string
// has only 9-10 bytes in its inline prefix storage.
DCHECK_GT(sv.size(), 16u); // we would not be in SMALL_TAG, otherwise.
string_view slice[2];
unsigned num = u_.small_str.GetV(slice);
DCHECK_EQ(2u, num);
DCHECK_LT(slice[0].size(), 14u);
uint8_t tmpbuf[14];
detail::ascii_pack(sv.data(), 16, tmpbuf);
// Compare the first slice.
if (memcmp(slice[0].data(), tmpbuf, slice[0].size()) != 0)
return false;
// Compare the prefix of the second slice.
size_t pref_len = 14 - slice[0].size();
if (memcmp(slice[1].data(), tmpbuf + slice[0].size(), pref_len) != 0)
return false;
// We verified that the first 16 chars (or 14 bytes) are equal.
// Lets verify the rest - suffix of the second slice and the suffix of sv.
return detail::compare_packed(to_byte(slice[1].data() + pref_len), sv.data() + 16,
sv.size() - 16);
}
LOG(FATAL) << "Unsupported tag " << int(taglen_);
return false;
}
size_t CompactObj::DecodedLen(size_t sz) const {
return ascii_len(sz) - ((mask_ & ASCII1_ENC_BIT) ? 1 : 0);
}
} // namespace dfly

View File

@ -76,6 +76,16 @@ struct RobjWrapper {
}
} __attribute__((packed));
// unpacks 8->7 encoded blob back to ascii.
// generally, we can not unpack inplace because ascii (dest) buffer is 8/7 bigger than
// the source buffer.
// however, if binary data is positioned on the right of the ascii buffer with empty space on the
// left than we can unpack inplace.
void ascii_unpack(const uint8_t* bin, size_t ascii_len, char* ascii);
// packs ascii string (does not verify) into binary form saving 1 bit per byte on average (12.5%).
void ascii_pack(const char* ascii, size_t len, uint8_t* bin);
} // namespace detail
class CompactObj {
@ -95,8 +105,17 @@ class CompactObj {
REF_BIT = 1,
EXPIRE_BIT = 2,
FLAG_BIT = 4,
// ascii encoding is not an injective function. it compresses 8 bytes to 7 but also 7 to 7.
// therefore, in order to know the original length we introduce 2 flags that
// correct the length upon decoding. ASCII1_ENC_BIT rounds down the decoded length,
// while ASCII2_ENC_BIT rounds it up. See DecodedLen implementation for more info.
ASCII1_ENC_BIT = 8,
ASCII2_ENC_BIT = 0x10,
};
static constexpr uint8_t kEncMask = ASCII1_ENC_BIT | ASCII2_ENC_BIT;
public:
using PrefixArray = std::vector<std::string_view>;
@ -216,7 +235,6 @@ class CompactObj {
return kInlineLen;
}
struct Stats {
size_t small_string_bytes = 0;
};
@ -226,6 +244,8 @@ class CompactObj {
static void InitThreadLocal(std::pmr::memory_resource* mr);
private:
size_t DecodedLen(size_t sz) const;
bool EqualNonInline(std::string_view sv) const;
// Requires: HasAllocated() - true.
@ -233,6 +253,8 @@ class CompactObj {
bool HasAllocated() const;
bool CmpEncoded(std::string_view sv) const;
void SetMeta(uint8_t taglen, uint8_t mask = 0) {
if (HasAllocated()) {
Free();
@ -268,6 +290,9 @@ class CompactObj {
};
inline bool CompactObj::operator==(std::string_view sv) const {
if (mask_ & kEncMask)
return CmpEncoded(sv);
if (IsInline()) {
return std::string_view{u_.inline_str, taglen_} == sv;
}

View File

@ -60,6 +60,10 @@ TEST_F(CompactObjectTest, NonInline) {
EXPECT_EQ(18261733907982517826UL, expected_val);
EXPECT_EQ(expected_val, obj.HashCode());
EXPECT_EQ(s, obj);
s.assign(25, 'b');
obj.SetString(s);
EXPECT_EQ(s, obj);
}
TEST_F(CompactObjectTest, Int) {
@ -81,4 +85,17 @@ TEST_F(CompactObjectTest, MediumString) {
obj.Reset();
}
TEST_F(CompactObjectTest, AsciiUtil) {
std::string_view data{"aaaaaabb"};
uint8_t buf[32];
char ascii2[] = "xxxxxxxxxxxxxx";
detail::ascii_pack(data.data(), 7, buf);
detail::ascii_unpack(buf, 7, ascii2);
ASSERT_EQ('x', ascii2[7]) << ascii2;
std::string_view actual{ascii2, 7};
ASSERT_EQ(data.substr(0, 7), actual);
}
} // namespace dfly