https://gcc.gnu.org/bugzilla/show_bug.cgi?id=89226
Bug ID: 89226
Summary: codegen for copying a 512-bit object fails to use avx
instructions
Product: gcc
Version: 9.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c++
Assignee: unassigned at gcc dot gnu.org
Reporter: barry.revzin at gmail dot com
Target Milestone: ---
Consider the following example:
#include <x86intrin.h>
// DUMB PAIR
struct dumb_pair {
alignas(2*sizeof(__m256i)) __m256i x[2];
};
void copy1(const dumb_pair& from, dumb_pair& to) {
to = from;
}
// SMART PAIR
struct foo512 {
__m256i a;
__m256i b;
auto& operator=(const foo512& f) {
a = f.a;
b = f.b;
return *this;
}
};
struct smart_pair {
union {
foo512 y;
__m256i x[2];
};
smart_pair(const smart_pair& sp) {
y = sp.y;
}
smart_pair& operator=(const smart_pair& sp) {
y = sp.y;
return *this;
}
};
void copy2(const smart_pair& from, smart_pair& to) {
to = from;
}
when compiled with:
g++ -mavx -O3 -march=corei7-avx -mtune=corei7-avx
on latest gcc (either trunk or 8.2 or 7.4) emits
(https://godbolt.org/z/mZj4VU):
copy1(dumb_pair const&, dumb_pair&):
vmovdqa xmm0, XMMWORD PTR [rdi]
vmovaps XMMWORD PTR [rsi], xmm0
vmovdqa xmm1, XMMWORD PTR [rdi+16]
vmovaps XMMWORD PTR [rsi+16], xmm1
vmovdqa xmm2, XMMWORD PTR [rdi+32]
vmovaps XMMWORD PTR [rsi+32], xmm2
vmovdqa xmm3, XMMWORD PTR [rdi+48]
vmovaps XMMWORD PTR [rsi+48], xmm3
ret
copy2(smart_pair const&, smart_pair&):
vmovdqa ymm0, YMMWORD PTR [rdi]
vmovdqa ymm1, YMMWORD PTR [rdi+32]
vmovdqa YMMWORD PTR [rsi], ymm0
vmovdqa YMMWORD PTR [rsi+32], ymm1
vzeroupper
ret
copy2() is better than copy1(). If we remove the user-provided copy assignment
operator from foo512 (even though the user-provided implementation is the same
as the default), the smart_pair code becomes the same as dumb_pair code.
clang++ emits the same code in both cases: the same code as copy2() in this
example.
