Re: test-round fails on alpha

[Neil Jerram]

address@hidden (Ludovic Courtès) writes:

> Anyway, I think experience from Debian (Rob & Neil) on that matter could
> be helpful.

It was being seen on m68k, and apparently that is not a
release-critical architecture for Lenny (the next Debian release), so
I stopped looking at it.

I've appended relevant history below, but basically this still needs
investigation.

       Neil


Me:
> For the rest...  There is a remaining bug in test-round in m68k:
> 
> lt-test-round: test-round.c:86: test_scm_c_round: Assertion
> `scm_c_round (x) == want' failed.
> 
> Will that prevent the upload from reaching testing?  If so, I'll look
> into a fix.

Me again:
> It seems to be a fairly obscure rounding test that is failing:
> 
>      x = ldexp (1.0, DBL_MANT_DIG - 1) - 1.0 + 0.5;
>      want = ldexp (1.0, DBL_MANT_DIG - 1);
>      assert (scm_c_round (x) == want);
> 
> i.e., IIUC, testing rounding near the upper limit of what can be
> represented as a floating point number.
> 
> Looking at this from the "what changed" point of view: I don't believe
> that either the code (scm_c_round() in libguile/numbers.c) or the test
> (test-suite/standalone/test-round.c) has changed significantly since
> 2004.  As far as Debian is concerned, therefore, I would have expected
> to see exactly the same m68k build failure for 1.8.1+1-4 (the current
> stable version) and 1.8.4+1.2 (current testing).  But I haven't been
> able to find any build logs to confirm or deny that.

Rob:
> Note that if I recall correctly, m68k isn't a release critical
> architecture for lenny, so although I think it would be good to fix
> this, it may not affect guile's inclusion in lenny.

Me:
> Excellent, sounds like we're all set then.