Re: condition DCE of partially dead calls

[Xinliang David Li <davidxl at google dot com>]

Andrew Pinski wrote:
> On Fri, May 30, 2008 at 10:21 AM, Xinliang David Li <davidxl@google.com> wrote:
> > Andrew Pinski wrote:
> > > On Fri, May 30, 2008 at 12:11 AM, Xinliang David Li <davidxl@google.com>
> > > wrote:
> > > > 2) Enabled only when target float format is IEEE format.
> > > +  /* The limit constants used in the implementation
> > > +     assume IEEE floating point format.  Other formats
> > > +     can be supported in the future if needed.  */
> > > +  return (flag_tree_builtin_call_dce != 0
> > > +          && TARGET_FLOAT_FORMAT == IEEE_FLOAT_FORMAT);
> > >
> > > This check is incorrect really as the float format might be different
> > > from the double format which might be different from the long double
> > > format.
> > You mean format for each mode may be overridden to be non-ieee format
> > when the TARGET_FLOAT_FORMAT is set to IEEE_FLOAT_FORMAT?
> >
> > > Plus REAL_MODE_FORMAT is what you should be checking as
> > > TARGET_FLOAT_FORMAT just supplies the default but REAL_MODE_FORMAT can
> > > be changed:
> > >  if (TARGET_LONG_DOUBLE_128 && !TARGET_IEEEQUAD)
> > >    REAL_MODE_FORMAT (TFmode) = &ibm_extended_format;
> > So after this override of quad mode format, the TARGET_FLOAT_FORMAT can
> >  still remain to be IEEE_FLOAT_FORMAT instead of some other value?
>
> Yes, TARGET_FLOAT_FORMAT is only setting the default format and nothing more.

Ok.


> > Basically, you are suggesting check format for individual mode. How
> > about long double? It is mapped to double-extended mode (XF) on some
> > targets (x86, ia64), quadMode on some target, while double mode on
> > others. XF mode may not exist on some targets, how do I check it safely
> > for all platforms?
>
> On PowerPC, the format for long double can either be IEEE 128bit,
> 64bit IEEE or the IBM 128bit format (which is a double-double format
> and still is related to IEEE format as it has two IEEE 64bit floats).
> So you really should not be checking IEEE format.  Also why are you
> checking IEEE format rather than if the format supports infinites and
> NaNs?

What is the macro to use for that check?

> So you need check the mode when you decide if you are going to do the
> DCE rather than just limiting the pass to where IEEE formats are used.
>  Also doing this optimization on more than IEEE formats seems correct
> anyways.

IEEE formats are majorities, and technically it is the only one I can
test and trust (thus the limitation for now) :(.


> > > Once the spu float format are submitted the spu port will be changed
> > > such TARGET_FLOAT_FORMAT is defined to SPU_FLOAT_FORMAT but the double
> > > format will be ieee.
> > I don't see problem with this case, as the TARGET_FLOAT_FORMAT value
> > does get changed to non IEEE.
>
> Because the format for double is IEEE, just float is non IEEE.  That
> is my problem with it, the check is too restrictive and not correct.

Ok.  I would love to take your suggestions for testing individual modes, 
I will see if there is a good way for long doubles.


Thanks,

David



> Thanks,
> Andrew Pinski