Abseil Common Libraries (C++) (grcp 依赖) https://abseil.io/
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

125 lines
4.4 KiB

// Copyright 2020 The Abseil Authors
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// https://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "absl/hash/internal/low_level_hash.h"
#include "absl/base/internal/unaligned_access.h"
Export of internal Abseil changes -- 04cb3b22497190170aa5b774e98080c5de2ba60b by Abseil Team <absl-team@google.com>: Alternative bit mixer for LowLevelHash on ARM LowLevelHash's bit-mixer is inefficient on ARM because it calculates a 128-bit product of two 64-bit numbers. On ARM, this requires a sequence of two instructions with a high combined latency and poor throughput. This change provides alternative bit-mixing code for ARM that uses only 64-bit arithmetic (multiplication, xor, and left-shifts) and speeds things up considerably. The bit-mixing code for ARM was inspired by by Woothash[1] and xxh3[1]. Once I landed on a sequence of operations that provided good mixing, I used a test harness to search for the combination of shift / rotate factors that provided the best mixing, as indicated by SMHasher hash quality tests. The new mixing code passes 13 out of 15 of the hash quality test suites in SMHasher, with the two failures being in the noise range: e.g. 1 collision vs. zero expected in a keyset of ~8m keys. [1]: https://github.com/tommyettinger/waterhash/blob/49f5cf0b63b9/woothash.h#L16-L20 [2]: https://github.com/Cyan4973/xxHash/blob/6853ddc36e46/xxhash.h#L3240-L3265 PiperOrigin-RevId: 391833008 -- 17a4de1f9d623155c75b19285d414cd55a487cd6 by Saleem Abdulrasool <abdulras@google.com>: debugging: add support for unwinding on RISCV Linux This adds partial support for unwinding the RISCV call stack. It is largely duplicated from the AArch64 support with alterations for the ELF RISCV psABI. This covers RISCV64 and RISCV32, though not the ILP32E calling convention. PiperOrigin-RevId: 391818522 -- 32c93e449327b2cea32b32f6365e84b420fe1ed3 by Gennadiy Rozental <rogeeff@google.com>: New storage for types smaller than 8 bytes. Also adding new read interface for types smaller than or rqual to 8 bytes to avoid passing the pointer. PiperOrigin-RevId: 391726822 -- e987ac08a7787801cbfc7d7c96649e97fa8cff1a by Abseil Team <absl-team@google.com>: Extern template `find_first_non_full` to reduce linkage size for TU with single not inlined function. PiperOrigin-RevId: 391718862 -- 73af9bfcb5bf045089133e18bbd20eb5bb699172 by Gennadiy Rozental <rogeeff@google.com>: Make most non-mutable most int128 methods and friend free functions constexpr. Some functions are implemented offline (at least in some configurations) and can't be made constexpr. Mutable methods can't be made constexpr until we drop c++11 support. Fixes #978 PiperOrigin-RevId: 391706535 GitOrigin-RevId: 04cb3b22497190170aa5b774e98080c5de2ba60b Change-Id: If051fad5ff004e2e82fa53618fc04a6fe3d2d4be
3 years ago
#include "absl/numeric/bits.h"
#include "absl/numeric/int128.h"
namespace absl {
ABSL_NAMESPACE_BEGIN
namespace hash_internal {
static uint64_t Mix(uint64_t v0, uint64_t v1) {
Export of internal Abseil changes -- 04cb3b22497190170aa5b774e98080c5de2ba60b by Abseil Team <absl-team@google.com>: Alternative bit mixer for LowLevelHash on ARM LowLevelHash's bit-mixer is inefficient on ARM because it calculates a 128-bit product of two 64-bit numbers. On ARM, this requires a sequence of two instructions with a high combined latency and poor throughput. This change provides alternative bit-mixing code for ARM that uses only 64-bit arithmetic (multiplication, xor, and left-shifts) and speeds things up considerably. The bit-mixing code for ARM was inspired by by Woothash[1] and xxh3[1]. Once I landed on a sequence of operations that provided good mixing, I used a test harness to search for the combination of shift / rotate factors that provided the best mixing, as indicated by SMHasher hash quality tests. The new mixing code passes 13 out of 15 of the hash quality test suites in SMHasher, with the two failures being in the noise range: e.g. 1 collision vs. zero expected in a keyset of ~8m keys. [1]: https://github.com/tommyettinger/waterhash/blob/49f5cf0b63b9/woothash.h#L16-L20 [2]: https://github.com/Cyan4973/xxHash/blob/6853ddc36e46/xxhash.h#L3240-L3265 PiperOrigin-RevId: 391833008 -- 17a4de1f9d623155c75b19285d414cd55a487cd6 by Saleem Abdulrasool <abdulras@google.com>: debugging: add support for unwinding on RISCV Linux This adds partial support for unwinding the RISCV call stack. It is largely duplicated from the AArch64 support with alterations for the ELF RISCV psABI. This covers RISCV64 and RISCV32, though not the ILP32E calling convention. PiperOrigin-RevId: 391818522 -- 32c93e449327b2cea32b32f6365e84b420fe1ed3 by Gennadiy Rozental <rogeeff@google.com>: New storage for types smaller than 8 bytes. Also adding new read interface for types smaller than or rqual to 8 bytes to avoid passing the pointer. PiperOrigin-RevId: 391726822 -- e987ac08a7787801cbfc7d7c96649e97fa8cff1a by Abseil Team <absl-team@google.com>: Extern template `find_first_non_full` to reduce linkage size for TU with single not inlined function. PiperOrigin-RevId: 391718862 -- 73af9bfcb5bf045089133e18bbd20eb5bb699172 by Gennadiy Rozental <rogeeff@google.com>: Make most non-mutable most int128 methods and friend free functions constexpr. Some functions are implemented offline (at least in some configurations) and can't be made constexpr. Mutable methods can't be made constexpr until we drop c++11 support. Fixes #978 PiperOrigin-RevId: 391706535 GitOrigin-RevId: 04cb3b22497190170aa5b774e98080c5de2ba60b Change-Id: If051fad5ff004e2e82fa53618fc04a6fe3d2d4be
3 years ago
#if !defined(__aarch64__)
// The default bit-mixer uses 64x64->128-bit multiplication.
absl::uint128 p = v0;
p *= v1;
return absl::Uint128Low64(p) ^ absl::Uint128High64(p);
Export of internal Abseil changes -- 04cb3b22497190170aa5b774e98080c5de2ba60b by Abseil Team <absl-team@google.com>: Alternative bit mixer for LowLevelHash on ARM LowLevelHash's bit-mixer is inefficient on ARM because it calculates a 128-bit product of two 64-bit numbers. On ARM, this requires a sequence of two instructions with a high combined latency and poor throughput. This change provides alternative bit-mixing code for ARM that uses only 64-bit arithmetic (multiplication, xor, and left-shifts) and speeds things up considerably. The bit-mixing code for ARM was inspired by by Woothash[1] and xxh3[1]. Once I landed on a sequence of operations that provided good mixing, I used a test harness to search for the combination of shift / rotate factors that provided the best mixing, as indicated by SMHasher hash quality tests. The new mixing code passes 13 out of 15 of the hash quality test suites in SMHasher, with the two failures being in the noise range: e.g. 1 collision vs. zero expected in a keyset of ~8m keys. [1]: https://github.com/tommyettinger/waterhash/blob/49f5cf0b63b9/woothash.h#L16-L20 [2]: https://github.com/Cyan4973/xxHash/blob/6853ddc36e46/xxhash.h#L3240-L3265 PiperOrigin-RevId: 391833008 -- 17a4de1f9d623155c75b19285d414cd55a487cd6 by Saleem Abdulrasool <abdulras@google.com>: debugging: add support for unwinding on RISCV Linux This adds partial support for unwinding the RISCV call stack. It is largely duplicated from the AArch64 support with alterations for the ELF RISCV psABI. This covers RISCV64 and RISCV32, though not the ILP32E calling convention. PiperOrigin-RevId: 391818522 -- 32c93e449327b2cea32b32f6365e84b420fe1ed3 by Gennadiy Rozental <rogeeff@google.com>: New storage for types smaller than 8 bytes. Also adding new read interface for types smaller than or rqual to 8 bytes to avoid passing the pointer. PiperOrigin-RevId: 391726822 -- e987ac08a7787801cbfc7d7c96649e97fa8cff1a by Abseil Team <absl-team@google.com>: Extern template `find_first_non_full` to reduce linkage size for TU with single not inlined function. PiperOrigin-RevId: 391718862 -- 73af9bfcb5bf045089133e18bbd20eb5bb699172 by Gennadiy Rozental <rogeeff@google.com>: Make most non-mutable most int128 methods and friend free functions constexpr. Some functions are implemented offline (at least in some configurations) and can't be made constexpr. Mutable methods can't be made constexpr until we drop c++11 support. Fixes #978 PiperOrigin-RevId: 391706535 GitOrigin-RevId: 04cb3b22497190170aa5b774e98080c5de2ba60b Change-Id: If051fad5ff004e2e82fa53618fc04a6fe3d2d4be
3 years ago
#else
// The default bit-mixer above would perform poorly on some ARM microarchs,
// where calculating a 128-bit product requires a sequence of two
// instructions with a high combined latency and poor throughput.
// Instead, we mix bits using only 64-bit arithmetic, which is faster.
uint64_t p = v0 ^ absl::rotl(v1, 40);
p *= v1 ^ absl::rotl(v0, 39);
return p ^ (p >> 11);
#endif
}
uint64_t LowLevelHash(const void* data, size_t len, uint64_t seed,
const uint64_t salt[5]) {
const uint8_t* ptr = static_cast<const uint8_t*>(data);
uint64_t starting_length = static_cast<uint64_t>(len);
uint64_t current_state = seed ^ salt[0];
if (len > 64) {
// If we have more than 64 bytes, we're going to handle chunks of 64
// bytes at a time. We're going to build up two separate hash states
// which we will then hash together.
uint64_t duplicated_state = current_state;
do {
uint64_t a = absl::base_internal::UnalignedLoad64(ptr);
uint64_t b = absl::base_internal::UnalignedLoad64(ptr + 8);
uint64_t c = absl::base_internal::UnalignedLoad64(ptr + 16);
uint64_t d = absl::base_internal::UnalignedLoad64(ptr + 24);
uint64_t e = absl::base_internal::UnalignedLoad64(ptr + 32);
uint64_t f = absl::base_internal::UnalignedLoad64(ptr + 40);
uint64_t g = absl::base_internal::UnalignedLoad64(ptr + 48);
uint64_t h = absl::base_internal::UnalignedLoad64(ptr + 56);
uint64_t cs0 = Mix(a ^ salt[1], b ^ current_state);
uint64_t cs1 = Mix(c ^ salt[2], d ^ current_state);
current_state = (cs0 ^ cs1);
uint64_t ds0 = Mix(e ^ salt[3], f ^ duplicated_state);
uint64_t ds1 = Mix(g ^ salt[4], h ^ duplicated_state);
duplicated_state = (ds0 ^ ds1);
ptr += 64;
len -= 64;
} while (len > 64);
current_state = current_state ^ duplicated_state;
}
// We now have a data `ptr` with at most 64 bytes and the current state
// of the hashing state machine stored in current_state.
while (len > 16) {
uint64_t a = absl::base_internal::UnalignedLoad64(ptr);
uint64_t b = absl::base_internal::UnalignedLoad64(ptr + 8);
current_state = Mix(a ^ salt[1], b ^ current_state);
ptr += 16;
len -= 16;
}
// We now have a data `ptr` with at most 16 bytes.
uint64_t a = 0;
uint64_t b = 0;
if (len > 8) {
// When we have at least 9 and at most 16 bytes, set A to the first 64
// bits of the input and B to the last 64 bits of the input. Yes, they will
// overlap in the middle if we are working with less than the full 16
// bytes.
a = absl::base_internal::UnalignedLoad64(ptr);
b = absl::base_internal::UnalignedLoad64(ptr + len - 8);
} else if (len > 3) {
// If we have at least 4 and at most 8 bytes, set A to the first 32
// bits and B to the last 32 bits.
a = absl::base_internal::UnalignedLoad32(ptr);
b = absl::base_internal::UnalignedLoad32(ptr + len - 4);
} else if (len > 0) {
// If we have at least 1 and at most 3 bytes, read all of the provided
// bits into A, with some adjustments.
a = static_cast<uint64_t>((ptr[0] << 16) | (ptr[len >> 1] << 8) |
ptr[len - 1]);
b = 0;
} else {
a = 0;
b = 0;
}
uint64_t w = Mix(a ^ salt[1], b ^ current_state);
uint64_t z = salt[1] ^ starting_length;
return Mix(w, z);
}
} // namespace hash_internal
ABSL_NAMESPACE_END
} // namespace absl