On Wed, Jun 06, 2018 at 02:52:27PM -0500, Segher Boessenkool wrote:
> On Tue, Jun 05, 2018 at 07:11:41PM -0400, Michael Meissner wrote:
> > This patch makes the long double pack/unpack built-in functions only
> > available
> > if the long double format is IBM extended double.
> >
> > I added new pack and unpack functions that use __ibm128 instead of long
> > double
> > in case somebody needs to build IBM extended double after we have changed
> > the
> > default to IEEE 128-bit.
>
> > @@ -15339,10 +15340,9 @@ rs6000_invalid_builtin (enum rs6000_buil
> > else if ((fnmask & RS6000_BTM_P9_MISC) == RS6000_BTM_P9_MISC)
> > error ("builtin function %qs requires the %qs option", name,
> > "-mcpu=power9");
> > - else if ((fnmask & (RS6000_BTM_HARD_FLOAT | RS6000_BTM_LDBL128))
> > - == (RS6000_BTM_HARD_FLOAT | RS6000_BTM_LDBL128))
> > - error ("builtin function %qs requires the %qs and %qs options",
> > - name, "-mhard-float", "-mlong-double-128");
> > + else if ((fnmask & RS6000_BTM_LDBL128) == RS6000_BTM_LDBL128)
> > + error ("builtin function %qs requires the %qs option", name,
> > + TARGET_IEEEQUAD ? "-mabi=ibmlongdouble" : "-mlong-double-128");
>
> This no longer tests for hard float. Is that correct, i.e. is hard
> float always enabled if we get here?
I originally thought that, but perhaps it is better to add an explict check
once again.
> > + /* For the pack and unpack int128 routines, fix up the builtin so it
> > + uses the correct IBM128 type. */
> > + case MISC_BUILTIN_PACK_IF:
> > + if (TARGET_LONG_DOUBLE_128 && !TARGET_IEEEQUAD)
> > + {
> > + icode = CODE_FOR_packtf;
> > + fcode = MISC_BUILTIN_PACK_TF;
> > + uns_fcode = (size_t)fcode;
> > + }
> > + break;
> > +
> > + case MISC_BUILTIN_UNPACK_IF:
> > + if (TARGET_LONG_DOUBLE_128 && !TARGET_IEEEQUAD)
> > + {
> > + icode = CODE_FOR_unpacktf;
> > + fcode = MISC_BUILTIN_UNPACK_TF;
> > + uns_fcode = (size_t)fcode;
> > + }
> > + break;
>
> (Space after cast, but nothing here does that I see).
>
> Is this correct? Why should it not use the IFmode version?
Because in the patch that I just commited (patch #2), we use the TF type for
__ibm128 when long double is IBM 128-bit. This makes templates work.
Otherwise if you try a template of __ibm128 and long double, you get an
internal compiler error.
> > +is returned. The @code{__builtin_unpack_longdouble} function is only
> > +availble if @code{long double} uses the IBM extended double
> > +representation.
>
> Typo ("available").
Ok.
> > +The @code{__builtin_pack_longdouble} function takes two @code{double}
> > +arguments and returns a @code{long double} value that combines the two
> > +arguments. The @code{__builtin_pack_longdouble} function is only
> > +availble if @code{long double} uses the IBM extended double
> > +representation.
>
> And here.
Ok.
> Okay for trunk (and backports) if those are not mistakes or you fix them,
> and maybe add some comment.
I will add a check for hard float back again.
--
Michael Meissner, IBM
IBM, M/S 2506R, 550 King Street, Littleton, MA 01460-6245, USA
email: [email protected], phone: +1 (978) 899-4797
