it's not clear to me why there is an ifdef __x86_64__ around the definitions of
_mm_cvtsi64_si128/_mm_cvtsi64x_si128 in emmintrin.h. the movq instruction
required to implement those has always been present in sse2 -- even on 32-bit
hosts.
gcc seems to do the right thing already if i just remove the #ifdef
__x86_64__... for example:
% cat movq.c
#include <inttypes.h>
#include <emmintrin.h>
static __inline __m128i __attribute__((__always_inline__))
_mm_cvtsi64x_si128 (long long __A)
{
return _mm_set_epi64x (0, __A);
}
void foo(__m128i *x, const uint64_t *y)
{
*x = _mm_cvtsi64x_si128(*y);
}
generates:
00000000 <foo>:
0: 55 push %ebp
1: 89 e5 mov %esp,%ebp
3: 8b 45 0c mov 0xc(%ebp),%eax
6: f3 0f 7e 00 movq (%eax),%xmm0
a: 8b 45 08 mov 0x8(%ebp),%eax
d: 66 0f 7f 00 movdqa %xmm0,(%eax)
11: 5d pop %ebp
12: c3 ret
(mind you i used a 4.1.2 gcc there, not 4.3... 'cause i neglected to build the
32-bit 4.3.x gcc. but the ifdefs are still present in the header file.)
-dean
--
Summary: _mm_cvtsi64x_si128/_mm_cvtsi64_si128 are valid in 32-bit
x86 code as well
Product: gcc
Version: 4.3.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: dean at arctic dot org
GCC build triplet: i686-unknown-linux-gnu
GCC host triplet: i686-unknown-linux-gnu
GCC target triplet: i686-unknown-linux-gnu
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=32869
