[PATCH, fortran ieee]: PR 88678, Many gfortran.dg/ieee/ieee_X.f90 test cases fail starting with r267465

[Uros Bizjak]

On Fri, Feb 8, 2019 at 12:53 AM Steve Ellcey <sellcey@marvell.com> wrote:
>
> On Thu, 2019-01-31 at 08:46 +0100, Uros Bizjak wrote:
> > On Wed, Jan 30, 2019 at 9:51 PM Janne Blomqvist
> >
> > > This seems to change the only user of support_fpu_trap() that is
> > > different from support_fpu_flag(), so with this change one could
> > > remove support_fpu_trap() entirely and modify all callers (since
> > > it's an internal function it's not used outside libgfortran) to
> > > call support_fpu_flag() directly. Otherwise Ok.
> >
> > Some targets only support IEEE flags (so, flags in some FP status
> > register), but not exceptions (traps) based on these flags that halt
> > the program. Currently, fpu-glibc.h assumes that all flags can
> > generate exceptions (that is true for the current set of gfortran
> > targets), but some future target wants to return 0 from
> > support_fpu_trap.
> >
> > Uros.
>
> Is this actually true for all existing gfortran targets?  Specifically
> I am wondering about Arm and Aarch64.  PR 78314 says that ARM trapping
> FPU exceptions are optional.

This is assumed in the tests and in the library, please see bellow.

> I am currently seeing gfortran.dg/ieee/ieee_6.f90 fail on my Aarch64
> ThunderX box.  I wonder if we should have an Arm/Aarch64 specific
> version of the fpu header file (fpu-arm.h) that would use the previous
> version of support_fpu_trap.

The fix for PR78314 was wrong on two accounts:

a) Probing for supported traps by trying to enable and disable trap
will fire exception when the trap is disabled, but the flag in the
status register is set. You can try this on x86 by forcing it to use
unpatched previous fpu-glibc instead of fpu-387 in
libgfortran/configure.host. Please also note PR78314, comment #16,
where it is reported that the test passes if target (QEMU arm/aarch64)
supports exceptions. So, if you use the old approach in arm specific
version of the file, I suspect you will regress other tests on
arm/aarch64 targets when/if optional exceptions are supported.

b) As reported in PR78314, comment #12, libgfortran currently assumes that:

gcc/fortran/simplify.c:
gfc_expr *
simplify_ieee_support (gfc_expr *expr)
{
  /* We consider that if the IEEE modules are loaded, we have full support
     for flags, halting and rounding, which are the three functions
     (IEEE_SUPPORT_{FLAG,HALTING,ROUNDING}) allowed in constant
     expressions. One day, we will need libgfortran to detect support and
     communicate it back to us, allowing for partial support.  */

  return gfc_get_logical_expr (gfc_default_logical_kind, &expr->where,
                               true);
}

so, the reverted patch neglected this assumption. Ignoring this, we can use

--cut here--
Index: libgfortran/config/fpu-glibc.h
===================================================================
--- libgfortran/config/fpu-glibc.h      (revision 268424)
+++ libgfortran/config/fpu-glibc.h      (working copy)
@@ -129,6 +129,10 @@
 int
 support_fpu_trap (int flag)
 {
+#if defined(__arm__) || defined(__aarch64__)
+  return 0;
+#endif
+
   return support_fpu_flag (flag);
 }

--cut here--

which would in effect revert to the previous status on arm/aarch64
targets, while using correct setting for other targets, where
exception support is assumed.

Uros.