https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121274
Bug ID: 121274
Summary: xmm extraction from zmm vector emits unnecessary
vpextrq/vpinsrq sequence
Product: gcc
Version: 16.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: mkretz at gcc dot gnu.org
Target Milestone: ---
Target: x86_64-*-*, i?86-*-*
Test case (https://compiler-explorer.com/z/PP36hnv55):
using V [[gnu::vector_size(64)]] = int;
auto f(V x)
{
return __builtin_shufflevector(x, x, 0, 1, 2, 3);
}
With -march=skylake-avx512 this compiles to
vpextrq rdx, xmm0, 1
vpinsrq xmm0, xmm0, rdx, 1
ret
which should instead be a simple 'ret'.
The same issue happens on extraction of the other 128-bit parts
(https://compiler-explorer.com/z/YYv8a1WTb):
__builtin_shufflevector(x, x, 4, 5, 6, 7);
is compiled to:
vextracti32x4 xmm2, zmm0, 1
vpextrq rdx, xmm2, 1
vpinsrq xmm0, xmm2, rdx, 1
ret
The expected result is:
vextracti32x4 xmm2, zmm0, 1
ret
>From my limited understanding of GCC tree dumps, I believe the "optimized
(tree)" pass is fine and the issue happens in the x86-target specific code?
