https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116512
Bug ID: 116512
Summary: Wrong vzeroupper at the function epilogue
Product: gcc
Version: 14.2.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c++
Assignee: unassigned at gcc dot gnu.org
Reporter: stlim0727 at gmail dot com
Target Milestone: ---
Code:
#include <immintrin.h>
struct B {
union {
__m512 f;
__m512i s;
};
};
auto foo(int n) {
B res;
res.s = _mm512_set1_epi32(n);
return res;
}
auto bar(int n) {
B res;
res.s = _mm512_set1_epi32(n);
return res.s;
}
Compiled with:
g++ -O2 -std=c++2a -march=skylake-avx512 test.cpp
Result:
foo(int):
vpbroadcastd zmm0, edi
vzeroupper
ret
bar(int):
vpbroadcastd zmm0, edi
ret
Unlike bar(), foo() has wrong vzeroupper which clears upper bits of zmm0 except
the lower 128-bits.
