diff options
Diffstat (limited to 'libblake_internal_blake2b_output_digest.c')
| -rw-r--r-- | libblake_internal_blake2b_output_digest.c | 36 |
1 files changed, 34 insertions, 2 deletions
diff --git a/libblake_internal_blake2b_output_digest.c b/libblake_internal_blake2b_output_digest.c index bc5b407..8700c8c 100644 --- a/libblake_internal_blake2b_output_digest.c +++ b/libblake_internal_blake2b_output_digest.c @@ -4,8 +4,18 @@ static void encode_uint64_le(unsigned char *out, uint_least64_t value, size_t bytes) { + /* Adding LIKELY to indicate that the default case is the + * expected does not affact the output */ switch (bytes) { default: + /* + * The following optimisation have been tested: + * + * 1) Changing the default case, on amd64, to + * *(uint64_t *)out = (uint64_t)value; + * break; + * result: halved performance + */ out[7] = (unsigned char)((value >> 56) & 255); /* fall through */ case 7: @@ -39,6 +49,28 @@ libblake_internal_blake2b_output_digest(struct libblake_blake2b_state *state, si { size_t i, j; - for (i = 0, j = 0; i < output_len; i += 8, j += 1) - encode_uint64_le(&output[i], state->h[j], output_len - i); +#ifdef LITTLE_ENDIAN + if ((uint_least64_t)(UINT_LEAST64_C(0xFFFFffffFFFFffff) + 1) == 0) { + /* 37.5x performance improvement; + * even though the compiler is smart enough to optimise + * `encode_uint64_le(&output[i], state->h[j], 8);` to a + * movq (on amd64), it is note smart enough to optimise + * the rest */ + memcpy(output, state->h, output_len); + return; + } +#endif + + /* Estimated to have similar performance benefit as above + * on big-endian machines */ + for (i = 0, j = 0; i + 8 < output_len; i += 8, j += 1) + encode_uint64_le(&output[i], state->h[j], 8); + encode_uint64_le(&output[i], state->h[j], output_len - i); + + /* + * Unoptimised code: + * + * for (i = 0, j = 0; i < output_len; i += 8, j += 1) + * encode_uint64_le(&output[i], state->h[j], output_len - i); + */ } |