icpc 11.1 fails to compile has_float and related cmath headers

[Joel Falcou]

The problem is strkingly similar to : ​http://groups.google.com/group/boost-list/browse_thread/thread/99ee648b03bcbf88 and Intel say that using mathimf.h insetad of cmath seems to fix the problem as per ​http://software.intel.com/en-us/forums/showthread.php?t=64188).

Seems fixing boost/config/no_tr1/cmath.hpp is enough to fix this bug.

Fidn attached modified files

[Joel Falcou]

Fixed cmath.hpp

[anonymous]

That doesn't actually fix things - in fact it causes large numbers of failures in our regression tests. Note that Intel on Linux normally handles Boost.Math just fine (and passes all our tests), it's only when you use icpc with an unsupported (ie too recent!) gcc version that the problems occur.

I'd much rather we found some other workaround, or else documented a workaround, or better got Intel to fix this - after all we really should be able to #include <cmath> which is all we're doing here!

[Joel Falcou]

Oh my bad :/ Now the think is that the only gcc i have is 4.4.3 ... i don't find that too recent :s

[anonymous]

"Now the think is that the only gcc i have is 4.4.3 ... i don't find that too recent"

No me neither :-(

I've been bitten by this myself from time to time BTW, and never found a good workaround, seems like the gcc guys keep changing their intrinsics and Intel are forever playing catchup :-(

John.

[Joel Falcou]

well :-( it renders nt2 unsupported by 11.1 on random machine it seems. I guess I'll work around that locally and wait for some intel solution then.

[anonymous]

I've been able to reproduce this locally, and will be committing a workaround shortly. In the mean time I've filed a bug report with Intel - it's issue #600230.

[John Maddock]

(In [65015]) Added workaround for Intel compiler, which has sporadic issues with fpclassify on Linux. Fixes #4592.