Re: CVS commit: src/tests/lib/libc/locale

[Robert Elz <kre%munnari.OZ.AU@localhost>]

    Date:        Thu, 30 Nov 2017 23:09:07 +0100
    From:        Manuel Bouyer <bouyer%antioche.eu.org@localhost>
    Message-ID:  <20171130220907.GA2318%antioche.eu.org@localhost>

  | Shouldn't it be made Xfail on i386 in this case ?

I don't think so, especially not now the problem is understood - it is
trivial to make it work on i386, the only question is which is the best
way to do that.    Right now I have -ffloat-store added to i386 builds
of this one file (not committed) which works, but I am not sure that
just reverting to the test for (close enough) is not better.  In fact
while composing this e-mail I have convinced myself that it is, so that
is what I am proposing to do.

An xfail of this test would give the impression that there's something
broken about locales on i386, which isn't the case at all (or not in any
way relevant to this test.)

A new test of simple floating point arith / functions / whatever, could
possibly xfail on i386 if it is actually appropriate - however it seems
that the C standards actually say that giving more accurate results than
the (floating) data type can hold is actually perfectly OK, so there is
actually nothing broken except the test, which is assuming that does not
happen, and should not be - either by forcing it using a compiler flag,
or by simply coding for it properly.

Another way to handle this, would be to change the value being tested
to one that can be represented perfectly in less than 56 mantissa bits.
That is, if more bits are provided, all of them will be 0.   I don't
think that's really the right way forward though.

kre