https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63791
Bug ID: 63791
Summary: use 32-byte version of vpbroadcastb on AVX2 platform
Product: gcc
Version: 4.9.2
Status: UNCONFIRMED
Severity: minor
Priority: P3
Component: c
Assignee: unassigned at gcc dot gnu.org
Reporter: marcus.kool at urlfilterdb dot com
Created attachment 33926
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=33926&action=edit
code with _mm256_set1_epi8, _mm256_loadu_si256, _mm256_cmpeq_epi8,
_mm256_movemask_epi8
With gcc 4.9.2 and compile options
-std=c99 -mavx2 -mbmi -mbmi2 -O3 -fno-tree-vectorize
on an Intel Haswell CPU
the intrinsic function _mm256_set1_epi8() generates 3 instructions while it
could do better with only 2 instructions.
Generated code is either
vmovd reg, xmmreg
vpbroadcastb xmmreg, xmmreg
vinserti128 $1, xmmreg, ymmreg, ymmreg
or
vmovd reg, xmmreg
vpbroadcastb xmmreg, xmmreg
vperm2i128 $0, ymmreg, ymmreg, ymmreg
But it could generate faster code instead:
vmovd reg, xmmreg
vpbroadcastb xmmreg, ymmreg
Example C source is in the attachment.
