diff options
author | Thiago Macieira <thiago.macieira@intel.com> | 2018-09-12 23:06:40 -0700 |
---|---|---|
committer | Thiago Macieira <thiago.macieira@intel.com> | 2018-12-08 20:28:56 +0000 |
commit | d36a4fc19709e6047fe846b36731b59909218b6d (patch) | |
tree | 1ca9b89ab023fdf97ba9f829fc9b70ce1b16603d /tests | |
parent | 2e715c31ed3a37fc196e97d4c58d0e277b1b9215 (diff) |
Optimize further the loading of 8 Latin 1 characters
This is important when AVX is enabled, which makes the VMOVQ load and
the VPMOVZXBW instruction be combined into a single VPMOVZXBW with
direct memory access. This is guaranteed to only read 8 bytes, so it's
safe even close to the end of a page. Clang and ICC do combine the
instructions like we want and I have filed a request for GCC to do so
too[1].
AVX was first introduced in 2011, so plenty of computers today would
benefit from this.
[1] https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87317
Change-Id: I8f261579aad648fdb4f0fffd1553e08e90df3171
Reviewed-by: Allan Sandfeld Jensen <allan.jensen@qt.io>
Diffstat (limited to 'tests')
0 files changed, 0 insertions, 0 deletions