https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121176
Bug ID: 121176
Summary: Missed optimization: mixed alignas + vectorization +
inline
Product: gcc
Version: unknown
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c++
Assignee: unassigned at gcc dot gnu.org
Reporter: rockeet at gmail dot com
Target Milestone: ---
#include <utility>
#define extent(a) sizeof(a)/sizeof(a[0])
struct alignas( 4) A { int a[7]; };
struct alignas(16) B { int a[8]; };
struct alignas(16) C { A a; int f; };
static_assert(sizeof(C) == sizeof(B));
static_assert(alignof(C) == alignof(B));
void A_swap(A* x, A* y) {
for (int i = 0; i < extent(x->a); i++)
std::swap(x->a[i], y->a[i]);
}
void B_swap(B* x, B* y) {
for (int i = 0; i < extent(x->a); i++)
std::swap(x->a[i], y->a[i]);
}
void C_swap(C* x, C* y) {
A_swap(&x->a, &y->a);
std::swap(x->f, y->f);
}
----------------------------------
After A_swap was inlined in C_swap, the generated code of C_swap should be same
with B_swap, but it is not(https://godbolt.org/z/KK7frznfs):
----------------------------------
A_swap(A*, A*):
movdqu xmm0, XMMWORD PTR [rdi]
movdqu xmm1, XMMWORD PTR [rsi]
movups XMMWORD PTR [rdi], xmm1
mov rdx, QWORD PTR [rsi+16]
movups XMMWORD PTR [rsi], xmm0
mov rax, QWORD PTR [rdi+16]
mov QWORD PTR [rdi+16], rdx
mov edx, DWORD PTR [rsi+24]
mov QWORD PTR [rsi+16], rax
mov eax, DWORD PTR [rdi+24]
mov DWORD PTR [rdi+24], edx
mov DWORD PTR [rsi+24], eax
ret
B_swap(B*, B*):
movdqa xmm0, XMMWORD PTR [rdi]
movdqa xmm1, XMMWORD PTR [rsi]
movaps XMMWORD PTR [rdi], xmm1
movdqa xmm1, XMMWORD PTR [rsi+16]
movaps XMMWORD PTR [rsi], xmm0
movdqa xmm0, XMMWORD PTR [rdi+16]
movaps XMMWORD PTR [rdi+16], xmm1
movaps XMMWORD PTR [rsi+16], xmm0
ret
C_swap(C*, C*):
movdqu xmm0, XMMWORD PTR [rdi]
movdqu xmm1, XMMWORD PTR [rsi]
movups XMMWORD PTR [rdi], xmm1
mov rdx, QWORD PTR [rsi+16]
movups XMMWORD PTR [rsi], xmm0
mov rax, QWORD PTR [rdi+16]
mov QWORD PTR [rdi+16], rdx
mov edx, DWORD PTR [rsi+24]
mov QWORD PTR [rsi+16], rax
mov eax, DWORD PTR [rdi+24]
mov DWORD PTR [rdi+24], edx
mov edx, DWORD PTR [rsi+28]
mov DWORD PTR [rsi+24], eax
mov eax, DWORD PTR [rdi+28]
mov DWORD PTR [rdi+28], edx
mov DWORD PTR [rsi+28], eax
ret
