-march=native failing on Windows laptop

[Tim Prince]

On 11/28/2013 2:32 PM, Toon Moene wrote:
> On 11/28/2013 05:13 PM, Tim Prince wrote:
>
>> Under gdb, it throws a segfault in the loop which is likely hanging
>> outside of gdb.  There are avx2 instructions in that loop.
>
> For comparison, I have *two* completely different weather forecasting 
> systems running on my home computer:
>
> proce
>
> If the example code is short, we could try to turn it into a 
> Bugzillable report ...
>
>
The bug shows up in a somewhat reduced case with the following 
combination (all conditions must exist together)
-march=native  (on a Haswell/corei7-4)
-fopenmp
-ftree-vectorize  (explicitly or by -O3)
!$omp parallel if(.false.)
so the case actually runs well with the OpenMP if() clauses removed. In 
the double precision version, it doesn't hang at the trouble point, it 
simply terminates silently.
  I run the test on various platforms from 4 to 118 hardware-supported 
threads, so I didn't want to go parallel on the smaller cases.

on a slightly divergent topic, I'm reluctant to attempt adding linux to 
my new Ultrabook on account of documented incompatibilities with Windows 
8.1 (where it took me hours to set up compilers) and unfamiliarity with 
linux installation without a DVD drive. (not to mention misgivings about 
linux distros which may support touch screen).  I do have an external 
hard drive available.

-- 
Tim Prince