[PATCH] x86-64: Use TI->SF and TI->DF conversions in soft-fp

[Terry Guo]

On Tue, Jan 22, 2019 at 7:48 AM Joseph Myers <joseph@codesourcery.com> wrote:
>
> On Mon, 21 Jan 2019, H.J. Lu wrote:
>
> > TI->SF and TI->DF conversions in libgcc2.c:
> >
> > FSTYPE
> > FUNC (DWtype u)
> > {
> >   ...
> > }
> >
> > have no rounding mode support.  We should replace __floattisf, __floattidf,
> > __floatuntisf and __floatuntidf in libgcc2.c with these from soft-fp.
>
> Please explain what you mean by "have no rounding mode support" (i.e., the
> exact flow through a function that is incorrect in a non-default rounding
> mode).  This patch is missing testcases - which of course should be
> architecture-independent.  (Any bug in libgcc2.c should first have an
> architecture-independent fix - it can't be considered fixed based on a fix
> for one architecture.  Then, if some other approach is optimal on
> particular architectures, they can get optimized variants.)
>
> I believe all those function implementations are designed so that only a
> single rounding occurs, which is for the final result, so no explicit
> handling of rounding modes is ever needed (the integer code before then
> may set up sticky bits appropriately to ensure the floating-point parts of
> the code only need a single rounding, which works in all modes), but maybe
> there are bugs in certain cases.  To identify the correct fix, we need
> details of the exact code path being used (the exact values of the various
> macros, choices for the various conditional parts of the function, values
> each variable has at each point) and where the existing,
> rounding-mode-independent logic goes wrong.
>
> --
> Joseph S. Myers
> joseph@codesourcery.com

Hi Joseph,

I believe HJ is proposing patch to fix bug
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88931. In the test case
of the bug, the "#pragma STDC FENV_ACCESS ON" is used and there are
four rounding modes:

  {
    ROUNDING (FE_DOWNWARD),
    ROUNDING (FE_UPWARD),
    ROUNDING (FE_TOWARDZERO),
    ROUNDING (FE_TONEAREST)
  }

The current _floattisf from libgcc2 doesn't support those four rounding modes.

BR,
Terry