Mercurial > crc32
diff crc32.c @ 3:6483683ac857 default tip
*: add profiling code too; expand x86 to use all eight XMM registers
basically ported verbatim from the assembly
| author | Paper <paper@tflc.us> |
|---|---|
| date | Mon, 09 Feb 2026 21:30:30 -0500 |
| parents | 422835bc1aca |
| children |
line wrap: on
line diff
--- a/crc32.c Mon Feb 09 01:21:00 2026 -0500 +++ b/crc32.c Mon Feb 09 21:30:30 2026 -0500 @@ -17,14 +17,12 @@ return; /* Calculate size needed to align */ - sz8 = align - ((uintptr_t)message % align); + sz8 = align - ((uintptr_t)*message % align); szs = MIN(*sz, sz8); *crc = crcfunc(*crc, *message, sz8); *message += sz8; *sz -= sz8; - - if (szs == sz8) assert(ALIGNED(*message, align)); } CRC32_API @@ -39,12 +37,17 @@ crc = 0xFFFFFFFF; crc32_align(&crc, crc32c_r, ALIGNOF(uint32_t), &message, &sz); if (!sz) return ~crc; -#ifdef __x86_64__ - crc32_align(&crc, crc32qw_r, 16, &message, &sz); - if (!sz) return ~crc; - return ~crc32x86_vpclmulqdq_r(crc, message, sz); -#else +#if defined(__x86_64__) && defined(__GNUC__) + /* Check at runtime if we can use vpclmulqdq */ + if (__builtin_cpu_supports("vpclmulqdq")) { + /* Align and do the rest with vpclmulqdq */ + crc32_align(&crc, crc32qw_r, 16, &message, &sz); + if (!sz) return ~crc; + + return ~crc32x86_vpclmulqdq_r(crc, message, sz); + } /* Otherwise just use 32-bit impl */ +#endif + return ~crc32qw_r(crc, message, sz); -#endif }