Hi!
On Mon, Apr 16, 2018 at 01:41:29PM -0400, Michael Meissner wrote:
> As I was working on PR target/85075 (to flesh some bugs with IEEE 128-bit
> support on the PowerPC, particularly with switching the default of long
> double), I noticed that for explicit IBM extended double, the compiler was
> converting the __ibm128 type to an IEEE 128-bit type, because those types had
> hardware support in ISA 3.0 (power9).
>
> Unfortunately, IBM extended double (a pair of doubles meant to give you more
> mantissa precision but not more expoenent range), is not completely
> representable in IEEE 128-bit floating point. There are likely some exposed
> corner cases involved, and the bottom bits might not be the same.
>
> While I would prefer IBM extended double to disappear entirely, I do think we
> need to deal with it for a few versions still to come.
There are many subtargets that still need it, and nothing to change
that has been planned.
> I tried to come up with machine dependent ways to prevent the automatic
> widening from occuring, but I couldn't get anything to work reliably. This
> patch adds a new target hook that says whether the automatic widening between
> two modes should be allowed. The default hook says to allow all of the
> default
> widenings to occur, while the PowerPC override says IBM extended double does
> not widen to IEEE 128-bit and vice versa.
>
> Given we are in stage4 right now, it is not the time to add new hooks, but
> here
> is the patch. If it doesn't go into GCC 8, is it acceptable for being put
> early into GCC 9 with a backport before 8.2 comes out?
>
> I have tested this patch using bootstrap builds on a power8 little system, a
> power7 big endian system (both 32-bit and 64-bit), and an x86_64 bit system
> with no regressions.
Can't you just set up things such that GET_MODE_WIDER_TYPE does not
return ieee128 for ibm128? With maybe a new hook yes, if that is
necessary.
> --- gcc/target.def (revision 259376)
> +++ gcc/target.def (working copy)
> @@ -3486,6 +3486,13 @@ If this hook allows @code{val} to have a
> hook_bool_mode_uhwi_false)
>
> DEFHOOK
> +(default_widening_p,
> + "Return true if GCC can automatically widen from @var{from_mode} to\n\
> +@var{to_mode}. Conversions are unsigned if @var{unsigned_p} is true.",
> + bool, (machine_mode, machine_mode, bool),
> + hook_bool_mode_mode_bool_true)
This should have those names (from_mode, to_mode, unsigned_p) in the
prototype then, for the generated documentation to make sense.
Segher
