6a9be62270
It wasn't getting inlined in aeshash256_lt32_avx256() (used by VAES +
AVX512VL variant) due to a GCC __attribute__((target())) mismatch,
causing a major loss of performance compared to the VAES + AVX2 variant.
Comparing the throughput after this fix on an Intel Core i7-1165G7
(Tiger Lake) laptop, with qHashBits modified to statically select either
[A] aeshash256() or [B] aeshash256_avx256(), out of 5 runs:
dictionary numbers paths-small uuids longstrings
A/B (avg) 103.7% 101.1% 103.5% 104.5% 100.3%
A/B (best) 103.4% 100.9% 103.2% 103.6% 100.8%
Considering that a string representation of a UUID is 37 characters (74
bytes), neither "uuids" nor "longstrings" are directly affected by this
change. However, the overhead does change, with the aeshash256_avx256()
needing slightly fewer instructions to reach aeshash256_ge32().
Benchmarking on an Intel Xeon Scalable 4th Generation (Sapphire Rapids),
the "uuids" data set has a 10% performance loss for some reason.
Pick-to: 6.6 6.5
Change-Id: I50e2158aeade4256ad1dfffd17b1b105d3cab482
Reviewed-by: Allan Sandfeld Jensen <allan.jensen@qt.io>
(cherry picked from commit 6ab4623cad39bec935f76e366f3f262922bde94a)
Reviewed-by: Qt Cherry-pick Bot <cherrypick_bot@qt-project.org>