http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56253
--- Comment #7 from Richard Biener <rguenth at gcc dot gnu.org> 2013-02-11
11:36:29 UTC ---
So to re-cap, for vector intrinsics we can use the vector GCC extensions
while for the scalar intrinsics we would have to use target builtin folding.
Not sure if it's worth the asymmetry (thus instead do everything with
target builtin folding).
Testcase that still needs to "work" (emit addsd) with -m32 -msse2
and without -mfpmath=sse:
#include <emmintrin.h>
double foo (double x)
{
return _mm_cvtsd_f64 (_mm_add_sd (_mm_set_sd (x), _mm_set_sd (1.0)));
}
Exposing the intrinsics internals to the compiler also would cause us to
not "literally" emitting the code the user may have asked for (though
RTL optimization already might do that - but RTL for example never
re-associates FP, even with -ffast-math).
