I opened another ticket (39171) to track the -ffast-math issue. I
believe it will be a significant amount of work to get make check to
succeed with -ffast-math so I don't see this being solved any time soon
- and I'm not even sure its possible, or even desirable, to rewrite all
of the affected routines in a way which will run consistently with these
optimizations.
For the other intel specific bug (39152), I agree it would be useful to
get make check to work out of the box with icc. One problem though is I
think older versions of icc don't have the -fp-model flag...they
required turning on/off many different optimization options, so
unfortunately any autoconf foo will be ugly. Maybe we should just add
the -fp-model precise when icc > X.XX version is detected
Patrick
On 06/05/2013 04:48 AM, Rhys Ulerich wrote:
Hi Patrick,
Thank you for investigating.
To turn off the unsafe intel optimizations, use:
make CFLAGS="-fp-model precise -g -O2"
These flags pass all tests for me with icc 13.0.1. Could you check this with
your 12.1 and 13.1?
With this CFLAGS setting, my 12.1 and 13.1 builds pass make check.
If it works I think we can close this bug.
Should we add autoconf foo to prevent Intel from shooting the user in the foot?
I suppose its
another (rather complicated) question whether we want to get GSL to pass all
the tests with gcc -ffast-math.
Agreed. Think that's worth a ticket? In my work I tend to gather
fast-math safe logic into one convenience library, no-fast-math into
another convenience library, and then create the installed .so from
those two outputs. This permits using -fsomething-sane only where
necessary.
- Rhys
