Help with testing 4.1.3

[Kevin Ryde]

Laurent Fousse <laurent at fousse.info> writes:
>
> | mpfr_sqrt_ui differs from sqrt for a=891885598, rnd_mode=GMP_RNDZ
> | sqrt gives 0x1.d2a1d0a4c79b2p+14, mpfr_sqrt_ui gives 
> | 0x1.d2a1d0a4c79b1p+14 (-1 ulp)

Looks like it's sqrt() which is wrong.  By my reckoning the root is
1D2A1D0A4C79B1FFD1... which rounded to zero should be ...b1 as per
mpf_sqrt_ui rather than ...b2 per sqrt().

Assuming fesetround has done it's job then it looks like sqrt is
either broken or not meant to obey the hardware rounding mode.  Such
is the folly of relying on exact answers from the math library :-(.
I'll disable that test.