Hi Alexandre,
on 2023/4/6 12:43, Alexandre Oliva wrote:
> Hello, Kewen,
>
> Thanks for the feedback.
>
> On Mar 27, 2023, "Kewen.Lin" <li...@linux.ibm.com> wrote:
>
>> on 2023/3/25 16:37, Alexandre Oliva via Gcc-patches wrote:
>>>
>>> When long double is 64-bit wide, as on vxworks, the rs6000 backend
>>> defines neither the __ibm128 type nor the __SIZEOF_IBM128__ macro, but
>>> pr99708.c expected both to be always defined. Adjust the test to
>>> match the implementation.
>
>> There is one patch from Mike to define type __ibm128 even without
>> IEEE 128-bit floating point support, it's at the link:
>
>> https://gcc.gnu.org/pipermail/gcc-patches/2022-August/599984.html
>
>> I would expect this issue would be gone if the adjustment on the
>> support of type __ibm128 gets landed in future.
>
> Yeah, the issue would then be gone, but the patch is compatible with
> that proposed change: if __ibm128 and the corresponding SIZEOF macro are
> defined, the proposed change is a no-op.
Yeah, I agree.
>
>> So maybe we can just xfail this for longdouble64? What do you
>> think?
>
> I wouldn't object to that, and I could even write and test the alternate
> patch for that, but I fail to understand why that would be more
> desirable. Would you be so kind as to enlighten me?
The reason why personally I preferred to fix it with xfail is that:
1) if the mentioned patch changing the condition of defining __ibm128 type
gets re-tested, this case would change from XFAIL to XPASS, then it gets
our attentions and shows some benefit of that patch (it can be also
mentioned in that commit log).
2) once the mentioned patch gets landed, the below hunk in the proposed patch
becomes unreachable:
+#else
+ || __SIZEOF_LONG_DOUBLE__ * __CHAR_BIT__ != 64
+#endif
And it's very likely we won't remember to remove it. At that time when
someone reads the code, he/she can probably get the impression that type
__ibm128 is not always defined even under effective target ppc_float128_sw,
it could cause some misunderstanding.
Does it make sense to you?
BR,
Kewen
