[RSB PATCH] 6: Merge the MacOS M silicon patch from gcc-12 to gcc-13

Chris Johns chrisj at rtems.org
Mon Dec 18 22:11:00 UTC 2023


On 18/12/2023 8:50 pm, Karel Gardas wrote:
> On 12/18/23 07:05, Chris Johns wrote:
>>> Mac with M1/M2/M3 work fine with the latest tooling.
>>
>> I think there is a middle ground here and that means some investigation is
>> needed to determine what works and what is at issue then deciding how much
>> further work is done. I have done some of this. The results are based on what I
>> have working:
>>
>> Builds:
>>
>>    Sonoma
>>    Montery
>>    Big Sur
>>
>> Fails:
>>
>>    Ventura
>>
>> The failure on Ventura is in GMP called from MPFR. Running the GMP tests a
>> number fail and removing --disable-shared improves the test results but GCC does
>> not build. I have not looked any further.
> 
> Just a rant.
> 
> IIRC when Ventura cames it was a disaster for RSB since on Monterey this worked
> and on Ventura it stopped. 

I think that was the removal of Python from MacOS.

> IIRC less RAM on system shows more sensibility to
> Ventura problems. And IIRC you were the person who solved this somehow.

Yes

> Does this mean we're back to the original state with Ventura making troubles
> again?

It is the different problem. I suspect something in Xcode and GMP have issues
that are not present on the other builds of MacOS and Xcode.

> I'm not sure if this is even technically possible (partly due to how
> Apple supports OS reverts) to test every version of Ventura released by Apple to
> see reliably if whole family is sick or there are some versions which may be
> recommended for use. E.g. Ventura spans from 13.0 to 13.5 already...

The issue is a pretty simple call by the compiler to MPFR then GMP to convert an
ASCII  real number to binary. The call crashes the compiler. I suspect something
is not right as the `make test` for GMP fails. I tested older versions that use
to work and they also fail.

A solution is create a suitable bug report and to post to GMP for them to look into.

Chris


More information about the devel mailing list