On Apr 14, 2022, Segher Boessenkool <[email protected]> wrote:
> Lol, the dates line up very well, I didn't realise it was from 2021 :-)
Heh, indeed. Same testsuite results cleanup season, too ;-)
>> The relevant fact, described in yesterday's message, is that -mfloat128
>> is not enabled by default, even with -mcpu=power9, except on target
>> variants that define TARGET_FLOAT128_ENABLE_TYPE to nonzero. As you
>> stated, its overall default is zero (though GNU/Linux overrides it to
>> nonzero), so the existing tests do not conform with the machine's
>> defaults in assuming -mfloat128 is enabled by -mcpu=power9.
> First off, vxworks.h should not disable it again.
Erhm... I'm not sure what the 'it' is.
For abundance of clarity, we do *not* disable vsx when -mcpu=power9 is
given. vsx is enabled for these tests. But neither -mcpu=power9 nor
having vsx enabled are enough for the _Float128/_ieee128 type to be
defined.
The target-specific option that controls whether _Float128/_ieee128 is
defined when VSX is enabled is TARGET_FLOAT128_ENABLE_TYPE. The only
file that defines it as nonzero is rs6000/linux64.h, which backs up the
comment in rs6000.cc before the statement that carries out this choice:
/* Enable the default support for IEEE 128-bit floating point on
[GNU/]Linux VSX sytems. [...] */
TARGET_FLOAT128_TYPE = TARGET_FLOAT128_ENABLE_TYPE && TARGET_VSX;
So, if the 'it' refers to VSX, I reaffirm it's enabled as it should.
But if 'it' refers to TARGET_FLOAT128_ENABLE_TYPE, then it would seem
that you're saying that this is no longer a choice available to targets,
and that _Float128/_ieee128 are now mandatory when VSX is available.
That would be quite a departure from the current state.
Now, we are looking into the possibility of enabling _Float128/_ieee128
on ppc64-vx7r2, but keep in mind it's a nonfree system, so if system
libraries (or kernel) aren't up to it, that would be a blocker. So I'd
prefer if both choices for TARGET_FLOAT128_ENABLE_TYPE remained
available.
> Then, this needs to be fixed, indeed. But that would be a code fix, not
> a testsuite workaround. If you use -mcpu=power9 it should support QP
> float.
I guess there's room for improvement indeed, especially in light of the
second patch for pr79004.c sent out ealier today, but I don't think I'd
risk such changes at this stage of development of gcc-12, let alone when
maintainer and implementation seem to me to disagree as to what the
expected behavior is :-(
--
Alexandre Oliva, happy hacker https://FSFLA.org/blogs/lxo/
Free Software Activist GNU Toolchain Engineer
Disinformation flourishes because many people care deeply about injustice
but very few check the facts. Ask me about <https://stallmansupport.org>
