When the code below is compiled with -march=pentium4 -msse2 -O3
-fomit-frame-pointer, when N == 4 f() is correct:
f:
movdqa b, %xmm0
paddd a, %xmm0
movdqa %xmm0, c
ret
However for N == 2 gcc 4.2.0 doesn't use %mm0 any more (gcc 4.0.0 worked OK).
OK, for some reason gcc decides to synthesize the operation using scalar
instructions. But why the code for f() is so scary, not even close to that of
g()? My wild guess is that gcc copies vectors onto the stack because it thinks
the vectors may overlap. But is it really possible? If so, how can I say that
they never overlap? Perhaps this is a bug too, because they are declared as
different objects, should not overlap.
#define N 2
typedef signed int data_type;
typedef data_type __attribute__((vector_size(sizeof(data_type) * N))) vector;
extern vector a, b, c;
void
f()
{
c = a + b;
}
typedef data_type vector2[N];
extern vector2 a2, b2, c2;
void
g()
{
int i;
for (i = 0; i < N; ++i)
c2[i] = a2[i] + b2[i];
}
For the code above gcc 4.2.0 produces:
f:
pushl %esi
pushl %ebx
subl $36, %esp
movl a, %ebx
movl a+4, %esi
movl b, %edx
movl b+4, %ecx
movl %edx, 24(%esp)
movl %ecx, 28(%esp)
movl 24(%esp), %eax
movl %ebx, 16(%esp)
movl %esi, 20(%esp)
addl 16(%esp), %eax
movl %eax, c
movl %edx, 8(%esp)
movl %ecx, 12(%esp)
movl 12(%esp), %eax
movl %ebx, (%esp)
movl %esi, 4(%esp)
addl 4(%esp), %eax
movl %eax, c+4
addl $36, %esp
popl %ebx
popl %esi
ret
g:
movl a2, %eax
addl b2, %eax
movl %eax, c2
movl a2+4, %eax
addl b2+4, %eax
movl %eax, c2+4
ret
--
Summary: int __attribute__((vector_size(8))) doesn't use %mm0,
produces ugly code
Product: gcc
Version: 4.2.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: middle-end
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: tomash dot brechko at gmail dot com
GCC build triplet: i686-pc-linux-gnu
GCC host triplet: i686-pc-linux-gnu
GCC target triplet: i686-pc-linux-gnu
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=32301
