AARCH64 support for ninja

[andriish]

Hi all,

this issue is inspired by gudrunhe/gosam#24.

Basically, I can repeat everything from that issue here.

The "problem" with that architecture is that the quadruple precision is implemented on AARCH64 natively.

That means that the precision of "long double" type is 128 bit.
That is a difference with x86_64 architecture where one needs special libraries ( e.g libquadmath ) to reach 128bit precision.
Obviously, this leads to a curious fact that e.g. libquadmath and corresponding headers are not needed on aarch64 architecture and they are absent from the GCC installation on aarch64.

That means any code that relies on quadmath.h headers will fail to compile on aarch64.

We have found it out in ATLAS during the compilation of POWHEG packages that were using some GoSaM+ninja generated code.

As an example of fixes we applied: https://gitlab.cern.ch/compilebox/compilebox/-/tree/master/COMPILEBOX/Custom/User-Processes-V2/ggHH/GoSamlib/quadninja

Best regards,

Andrii

[peraro]

Hi, thank you for the feedback. Can you please check that 93fa03e fixes this?

I have checked that it works on Linux and not on MacOS ARM64 systems, as in the latter it seems that long double is an alias for double (although it does work on Linux VMs or containers in ARM-based Macs). Either way, now the configure script on any AARCH64 system checks that long double has enough digits and in that case uses that as the quadruple precision type.

[andriish]

Hi @peraro,
many thanks for the quick reply.

I will start some checks now. Will tell about the outcome later.

Meanwhile:

• Will post some pull requests.
• yes, ARM64 on Mac is different and I haven't looked there :(

Best regards,

Andrii