https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116574
Bug ID: 116574
Summary: GCC generates redundant moves in simple vector
addition
Product: gcc
Version: 14.2.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c++
Assignee: unassigned at gcc dot gnu.org
Reporter: bernardo at bernardosulzbach dot com
Target Milestone: ---
When given
#include <array>
using V = std::array<float, 2>;
V add(V v1, V v2) { return {v1[0] + v2[0], v1[1] + v2[1]}; }
GCC 14.2 with -Os produces
add(std::array<float, 2ul>, std::array<float, 2ul>):
movq rax, xmm0
movq rdx, xmm1
addss xmm0, xmm1
shr rax, 32
shr rdx, 32
movd xmm3, edx
movd xmm2, eax
addss xmm2, xmm3
movd edx, xmm0
movd eax, xmm2
sal rax, 32
or rdx, rax
movq xmm0, rdx
ret
and with -O2 or -O3 produces
add(std::array<float, 2ul>, std::array<float, 2ul>):
movq xmm1, xmm1
movq xmm0, xmm0
addps xmm0, xmm1
ret
while Clang always produces
add(std::array<float, 2ul>, std::array<float, 2ul>):
addps xmm0, xmm1
ret
as I'd expect.
Compiler Explorer (godbolt.org) link (for convenience, it doesn't have anything
I didn't write here): https://godbolt.org/z/chsqqjPPx.
The -Os version should probably not generate so much code, and I think the
`movq {x}, {x}` in the -O2/-O3 version are redundant.
I picked C++ as the component, because I don't know what component to blame for
this.
