Re: real.h: REAL_VALUE_TO_TARGET_DOUBLE

[Richard Henderson <rth at redhat dot com>]

Omar Torres wrote:
> On Thu, Oct 2, 2008 at 11:59 PM, Ian Lance Taylor <iant@google.com> wrote:
> > "Omar Torres" <gcc.omar@gmail.com> writes:
> >
> > >  Shouldn't this macro:
> > > #define REAL_VALUE_TO_TARGET_DOUBLE(IN, OUT) \
> > >   real_to_target (OUT, &(IN), mode_for_size (64, MODE_FLOAT, 0))
> > >
> > >  be using DOUBLE_TYPE_SIZE instead of the hard coded '64'? Am I missing
> > > something here?
> > That would certainly be more logical.  In order to make that change,
> > you would have to consider each use of the macro to confirm that the
> > change would be valid.
> >
> >
> > > In the target I am currently working, DOUBLE_TYPE_SIZE is defined as
> > > 32-bit, so the hard-coded '64' is causing trouble. I am planning on
> > > doing this change on my local source tree, but wanted to discuss with
> > > the community to see if there is some implementations details somewhere
> > > that I need to consider as well.
> > As far as I can see, all but one use of this macro is in target
> > dependent code.  Are you having trouble from the single target
> > independent use?  Or perhaps your target dependent code should just
> > use REAL_VALUE_TO_TARGET_SINGLE instead?
> >
> > Ian
>  Both.
>  Yes, I can use REAL_VALUE_TO_TARGET_SINGLE in the target dependent
> files to work around this.
>  I do not see a similar solution for the target independent code,
> since the size mismatch cause the incorrect format (encode/decode
> functions) to be use.

The size of the C type "double" is controlled by DOUBLE_TYPE_SIZE,
not the size of the compiler mode "DFmode".  This macro is referring
to the latter -- a double-precision floating point mode.


r~