Re: [patch, fortran] Fix NEAREST for float formats without infinity

["Ulrich Weigand" <uweigand at de dot ibm dot com>]

Janne Blomqvist wrote:
> On Wed, Mar 25, 2009 at 22:32, Ulrich Weigand <uweigand@de.ibm.com> wrote:
> > Hello,
> >
> > gfc_conv_intrinsic_nearest currently uses the "__builtin_inf" family
> > of intrinsics. =A0This is a problem for platforms with floating-point
> > formats that do not support infinity, like single-precision float
> > on the SPU.
> >
> > In this situation, fold_builtin_inf will issue a warning "target
> > format does not support infinity". =A0This may be appropriate for
> > explicit uses of the INFINITY macro in C code -- but in fact you
> > also get the same warning today whenever you use NEAREST on a
> > REAL (4) type in Fortran on the SPU; this seems quite confusing.
> >
> > I suggest to use __builtin_huge_val instead of __builtin_inf to
> > implement NEAREST. =A0The implementation of the two is completely
> > equivalent -- except that __builtin_huge_val does not issue the
> > warning. =A0The following patch implements this change.
> >
> > Tested on spu-elf with no regressions. =A0Fixes a number of test
> > cases that failed due to the warning on NEAREST.
> 
> This looks ok.  Thanks.

Thanks for the review!

> > OK for mainline?
> 
> Well, at the moment mainline is open only for regression and
> documentation fixes, so no. But ok for fortran-dev, and merging back
> to trunk once stage 1 opens.

I guess you could consider this to be a regression: the in-line
expansion via gfc_conv_intrinsic_nearest was newly introduced
with GCC 4.4; for GCC 4.3, you'd always get a call to a libgfortran
library routine.

This has the effect that using NEAREST on gfortran 4.3 for the SPU
does *not* result in that warning, but with gfortran 4.4 it does.

Bye,
Ulrich

-- 
  Dr. Ulrich Weigand
  GNU Toolchain for Linux on System z and Cell BE
  Ulrich.Weigand@de.ibm.com