Update CI runners

[albinahlback]

In particular, remove x86 macOS runner since they are deprecated on Github.

[albinahlback]

@edgarcosta How can I find Clang systematically here? I believe Clang is not on the path, so I have to manually get the path, but I don't want to guess the version of llvm (see line 219).

[edgarcosta]

I got the runner working with the latest llvm, but then make check fails

[albinahlback]

Yeah, newer LLVM doesn't seem to like FLINT.

[albinahlback]

@fredrik-johansson I think I have fixed a lot of errors for generic macros (like add_s{n}a{2n}). We cannot have something like

# define add_ssaaaa(s1, s0, a1, a0, b1, b0) \
  do { \
    ulong __t0 = (a0); \
    (s0) = (a0) + (b0); \
    (s1) = (a1) + (b1) + ((ulong) (s0) < __t0); \
  } while (0)

# define add_sssaaaaaa(s2, s1, s0, a2, a1, a0, b2, b1, b0) \
  do { \
    ulong __t0, __t1; \
    add_ssaaaa(__t0, s0, (ulong) 0, a0, (ulong) 0, b0); \
    add_ssaaaa(__t1, s1, (ulong) 0, a1, (ulong) 0, b1); \
    add_ssaaaa(s2, s1, (a2) + (b2), s1, __t1, __t0); \
  } while (0)

since __t0 is used in both add_sssaaaaaa and add_ssaaaa. Not sure why it causes a problem, though, just that this fixes the problem. src/test, mpn_extras and ulong_extras seems to work fine now, but one matrix multiplication in nfloat seems to fail; I'm unable to track where this error stems from. Temporarily, we can check the fails on generic macros by commenting out everything between line 28 and line 58 in longlong.h, but we should integrate this into our CI.

[fredrik-johansson]

How should I configure FLINT to reproduce the nfloat failure?

[albinahlback]

How should I configure FLINT to reproduce the nfloat failure?

You can configure as normal, but just comment out every longlong_*.h inclusion inside longlong.h (so only generic macros are being used).

[fredrik-johansson]

Found it: FLINT_MPN_MUL_2X2H in nfloat/dot.c needs unique variable names too.

[edgarcosta]

Well done!

[fredrik-johansson]

Best I guess would be to choose very unique names for the temporaries in add_ssaaaa et al to minimize the risk of downstream breakage.

[albinahlback]

I will merge this when all GCC tests passes. Clang seems to be really buggy -- we have two different bugs popping up. For Ubuntu Clang, it doesn't want to compile a test file, and for Apple M1 it wrongly assumes that __builtin_clzl always returns something non-zero, which results in the failure of sdiv_qrnnd.

[fredrik-johansson]

I guess as a workaround on M1 you can cast the result of clz to a volatile variable or something like that.