_mm_cvtpi32x2_ps in xmmintrin.h could be made more efficient. the existing
version:
static __inline __m128 __attribute__((__always_inline__))
_mm_cvtpi32x2_ps(__m64 __A, __m64 __B)
{
__v4sf __zero = (__v4sf) _mm_setzero_ps ();
__v4sf __sfa = __builtin_ia32_cvtpi2ps (__zero, (__v2si)__A);
__v4sf __sfb = __builtin_ia32_cvtpi2ps (__zero, (__v2si)__B);
return (__m128) __builtin_ia32_movlhps (__sfa, __sfb);
}
generates unneccesary zeroes and copies of xmm registers. as cvtpi2ps and
movlhps both overwrite the target portion of the target register, it is not
neccesary to create and copy a zero-filled register in this function. as
written, it also saves the value of __sfa before the movlhps instruction. i've
modified it to:
static __inline __m128 __attribute__((__always_inline__))
_mm_cvtpi32x2_ps(__m64 __A, __m64 __B)
{
__v4sf __sfa = __builtin_ia32_cvtpi2ps (__sfa, (__v2si)__A);
__v4sf __sfb = __builtin_ia32_cvtpi2ps (__sfa, (__v2si)__B);
return (__m128) (__sfa = __builtin_ia32_movlhps (__sfa, __sfb));
}
and gotten about 25% shorter run-time in a trivial test app that converts an
array of long ints to floats.
--
Summary: faster _mm_cvtpi32x2_ps for xmmintrin.h
Product: gcc
Version: 4.1.1
Status: UNCONFIRMED
Severity: enhancement
Priority: P3
Component: other
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: andrew dot mahone 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=29096
