Polyhedron b'marks | Linux vs Windows

[Tim Prince]

On 6/2/2014 10:32 AM, Richard Biener wrote:
> On Mon, Jun 2, 2014 at 4:28 PM, Evan Cooch <evan.cooch@gmail.com> wrote:
>>>> So (the question) -- why? I have a couple of ideas, but I was wondering
>>>> if there is a short-form 'correct answer' (correct, as in, better than
>>>> my 'couple of ideas').
>>>>
>>>> Thanks in advance...
>>> Differences that large would make one suspect you didn't clean out
>>> competing Windows processes,
>>
>> Meaning, I suspect, competing OS process, since the Polyhedron benchmark was
>> the only user-initiated task. My next experiment will be to run the PB after
>> booting Windows straight to the command line (i.e., absolute bare minimum of
>> Windows functioning -- more or less 'straight to DOS', in the classical
>> sense of a bygone era).
>>
>>> or you have ridiculously small RAM,
>>
>> 64 GB.  So, probably not this.  ;-)
>>
>>> or something is wrong with your Windows gfortran.
>>
>> Hmmm --- interesting.  I'm not sure what would be wrong with gfortran under
>> Windows that wouldn't also be wrong under GNU/Linux, but, stranger things
>> have happened.
>>
>>> To a smaller extent, your excessive unrolling (as you didn't set e.g.
>>> --param max-unroll-times=...) could hurt more on Windows.
>>
>> Easy enough to try -- thanks.
>>
>>> Failing to set -march=native ought not to incur many differences if both
>>> are set the same.  If  -march=native incurs failures, gfortran 4.9 might be
>>> a fairer test (assuming it has corrected those failures).
>>>
>> I suppose worth a try.
>>
>> What is interesting (and largely what motivated my original post) is that I
>> have a number of colleagues who report more or less the same thing (and that
>> it has been the case since at least gfortran 4.4).
> I suspect the main difference is the system supplied math library,
> glibc libm in the Linux case and whatever Windows provides.  This
> can make a tremendous difference for the benchmarks that call
> transcendental functions.
>
>
There are various library choices among Windows ports (mingw, cygwin 
(newlib), 32-bit, 64-bit, .....) but I didn't expect them to have much 
effect on Polyhedron.  Anyway, to imply that Windows chooses the library 
seems misleading.
glibc doesn't do much in the area of math function optimizations either.

-- 
Tim Prince