This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: Fix recent SPEC2000/x86-64 ICEs


> 
> Hi Jan,
> > Roger's patches to add math builtins are not working very well on x86-64
> > where we use TARGET_128BIT_LONG_DOUBLE so we must use TFmode instead.
> 
> One quick question.  On x86-64 in 128-bit long double mode does the x87
> instruction "fldpi" load an 80-bit or 128-bit approximation to pi?  I
> know nothing about AMD64, but it looks like your patch just extends the
> truncated 80-bit value.  Its just that this seems a bit odd.

The 128-bit long doubles on x86-64 are just i386 80bit doubles with 48
bits of padding to get loads/stores faster, so everything works as
usual, just you need to use TFmode instead of XFmode everywhere as GCC
does not like idea of mixing multiple extended FP modes.

Honza
> 
> Roger
> --


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]