On Thu, Nov 3, 2011 at 6:54 PM, Jakub Jelinek <ja...@redhat.com> wrote:
> As mentioned in the last mail, the testcase in that patch > shows a bug in the unsigned int -> float vectorization > on i?86/x86_64. E.g. 0x8000008fU converted to float > by scalar code is 0x1.000002p+31, but by vector code > is 0x1p+31, i.e. one ulp smaller. < stuff > > So, what do you prefer and do you want the expander to be moved into > i386.c or kept like this? Do we perhaps want for -ffast-math > keep the slightly faster, but imprecise version (I'd prefer not to)? IMO, let's go with the fastest solution (X313c), but please move expander to i386.c. Also, -ffast-math should use the same version. At the end of the day, real-world applications do not comprise only unsigned conversions ;) Uros.
