update MPFR to v4.1.0

[vtjnash]

Includes build fixes to ensure it is built thread-safe. Fixes #35796 regression.

[JeffBezanson]

Backport?

[vtjnash]

Seems like AArch64 is having problems computing pi with this update. I might need someone with access to a machine to investigate.

[JeffBezanson]

Can we try 4.0.2 with threading enabled?

[vtjnash]

I assume I just need to rerun the Yggdrasil builder in a branch

[vtjnash]

Looking deeply into this, seems like something it going very badly. I see we're crashing on this copy/zero:

    if (MPFR_UNLIKELY (_destprec >= _srcprec))                              \
      {                                                                     \
        _srcs  = MPFR_PREC2LIMBS (_srcprec);                                \
        _dests = MPFR_PREC2LIMBS (_destprec) - _srcs;                       \
        MPN_COPY (_destp + _dests, srcp, _srcs);                            \
        MPN_ZERO (_destp, _dests);                                          \
        inexact = 0;                                                        \
      }                                                                     \

Where we know that _destprec == 32 and _srcprec == 32 + 63!

getting called from
MPFR_RNDRAW (inexact, a, tmp, bq + cq, rnd_mode, sign, ax2++); (mul.c:1086)
getting called from
mpfr_mul (bound_emin, bound_emin, e, MPFR_RNDD); (exp.c:119)

[vtjnash]

I'm almost certain this is a gcc (4.8.5) bug happening on some new code in mpfr-4.1.0 (prior code didn't have the cast to int)

   1084   /* b1 is 0 or 1 (most significant bit from the raw product) */
   1085   ax2 = ax + ((int) b1 - 1);
   1086   MPFR_RNDRAW (inexact, a, tmp, bq + cq, rnd_mode, sign, ax2++);

Where the round-down macro contains the if (_destprec >= _srcprec) { bad } else { good } I mention above.

The assembly though appears to have re-written that into:

if (b1 == 0) {
  ax2 = ax - 1;
  bad;
}
else {
  ax2 = ax + 0;
  good;
}

So either staying on 4.0.2 with the new flags or using a less ancient version of gcc (June 23, 2015) should resolve that.

[giordano]

Is gcc v5 less ancient enough?

[vtjnash]

I don't have gcc 4 installed to compare, but gcc 5 seems to be generating sensible code for me.