[R] make check failed on linux-amd64 using PGI compilers

[Prof Brian Ripley]

On Thu, 10 Nov 2005, Liaw, Andy wrote:

>> From: pd at pubhealth.ku.dk
>>
>> "Liaw, Andy" <andy_liaw at merck.com> writes:
>>
>>> Dear R-help,
>>>
>>> I am trying to build R-2.2.0-patched (2005-11-07 r36217) on
>> the head node of
>>> a Scyld cluster (dual Opteron 250s) using PGI compilers
>> (v6.0).  I used the
>>> flags suggested by Jennifer Lai on R-devel (taken from
>> R-admin, except that
>>> I had to add -L/usr/X11R6/lib64 to LDFLAGS).  The build
>> went fine, but make
>>> check-all failed when running tests/Examples/graphics-Ex.R, at:
>>>
>>>> plot(1:2, xaxs = "i") # 'inner-axis' w/o extra space
>>>> stopifnot(par("xaxp")[1:2] == 1:2 &&
>>> +           par("usr")[1:2] == 1:2)
>>> Error: par("xaxp")]1:2] == 1:2 && par("usr")[1:2] == 1:2 is not TRUE
>>>
>>> The above looks a bit strange to me, as running the R built
>> as above,
>>> par("xaxp")[1:2] - 1:2 gives
>>>
>>> [1] -1.110223e-16 -2.220446e-16
>>>
>>> Is my R build faulty?  I'd very much appreciate any advise.
>>
>> Relative errors of 1 ULP would not ruin my sleep.  It's not obvious to
>> me which is the floating point operation responsible for the effect,
>> though, and it is a bit suspicious given that it should just be the
>> x-range of the data.
>
> It's the par() output that has storage mode "double", as that stores the
> coordinates.  1:2 has storage mode "integer".
>
> Those comparisons looked a bit suspicious to me.  Isn't it recommended time
> and again on this list not to test for exact equality of numerics?
> all.equal() of the above does return TRUE.

Yes.  It is also not the intention to have examples be regression tests, 
something we have moved away from in general.

But I think Peter's point was that par("usr") should be exactly the range 
of the data (it is), and par("xaxp") should be par("usr") with xaxs="i" 
when the end points are round numbers. Now setting par("xaxp") is a 
tangled tale and involves a call to GPretty, and the latter is what will 
be doing floating-point computations.  It is a little strange that they 
are not exact, but although all.equal with the default tolerance would be 
too loose, a tolerance of say 100 ulp would suffice.


-- 
Brian D. Ripley,                  ripley at stats.ox.ac.uk
Professor of Applied Statistics,  http://www.stats.ox.ac.uk/~ripley/
University of Oxford,             Tel:  +44 1865 272861 (self)
1 South Parks Road,                     +44 1865 272866 (PA)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595