mirror of
https://github.com/MariaDB/server.git
synced 2025-10-27 17:08:14 +01:00
305576e9c6
For the algorithms in crc32_arm64.c to work consistenly, all multibyte integers loaded from the input buffer have to be loaded in little-endian. Tested on an aarch64_be-unknown-linux-musl system, with: ninja crc32-t && unittest/mysys/crc32-t Test results: unpatched: +------------------------------------+-------------------+--------------------+ | Code path | GCC 15.2.0 | Clang 20.1.8 | +------------------------------------+-------------------+--------------------+ |`crc32c_aarch64` | 25 failed tests | 25 failed tests | +------------------------------------+-------------------+--------------------+ |`crc32c_aarch64_pmull` (assembly) | 25 failed tests | doesn't compile | +------------------------------------+-------------------+--------------------+ |`crc32c_aarch64_pmull` (intrinsics) | 25 failed tests | 26 failed tests | +------------------------------------+-------------------+--------------------+ patched: +------------------------------------+-------------------+--------------------+ | Code path | GCC 15.2.0 | Clang 20.1.8 | +------------------------------------+-------------------+--------------------+ |`crc32c_aarch64` | success | success | +------------------------------------+-------------------+--------------------+ |`crc32c_aarch64_pmull` (assembly) | success | doesn't compile | +------------------------------------+-------------------+--------------------+ |`crc32c_aarch64_pmull` (intrinsics) | success | 1 failed test | +------------------------------------+-------------------+--------------------+ Implementation notes: - uintNkorr uses the byte-wise implementation of a fixed-endian load, which can't always be turned back into a single load instruction. On aarch64-unknown-linux-musl (little endian) with GCC 15.2, this resulted in a code pessimization at -O0 and -O1, but -Os and -O2 are fine, so this is not a problem. On MSVC/Windows there does appear to be a pessimization[1]. - Using le16toh etc. are a potential alternative, as these macros are no-ops on little-endian platforms that support them. The endian(3) manpage suggests that NetBSD, FreeBSD, and glibc (in addition to musl), but not OpenBSD, support them[2]. Windows support is unclear. leNNtoh has the added disadvantage of being unusual in the MariaDB codebase, so it's better to stick with uintNkorr. [1]: https://godbolt.org/z/oncv9Gj19 [2]: https://man.he.net/man3/htole64 |
||
|---|---|---|
| .. | ||
| clang_workaround.h | ||
| crc32_arm64.c | ||
| crc32_ppc64.c | ||
| crc32_x86.c | ||
| crc32c.cc | ||
| crc32c_amd64.cc | ||
| crc32c_ppc.c | ||
| crc32c_ppc.h | ||
| crc32c_x86.cc | ||
| crc_ppc64.h | ||
| pcc_crc32_constants.h | ||
| pcc_crc32c_constants.h | ||