NaN with -ffast-math

[mchamberland]

NaN (still) shows up when using -ffast-math with the Fortran compiler. This happens with gcc version 4.4.7 20120313 (Red Hat 4.4.7-4) (GCC).

Sample input file for tutor7pp below.

With -ffast-math flag:

Medium 1 sige = 1.00000000362749373E-015 NaN monotone = T T

Medium 2 sige = 1.00000000362749373E-015 NaN monotone = T T

Without -ffast-math:

Medium 1 sige = 0.43671621424277174 0.43386623396694124 monotone = T T

Medium 2 sige = 0.24630246820211257 0.24325696800012028 monotone = T T

As far as I know, this does not produce any significant differences in the results of tutor7pp (or any other code), but I'm not entirely sure.

material.txt
media_bug.txt

[crcrewso]

May I ask, which input file were you using? Test1?

[mchamberland]

The input file attached to the report: "media_bug.txt"

GitHub does not allow .egsinp extensions.

[crcrewso]

Ahhh sorry it threw me not being an egsinp.

[rtownson]

Is this still a problem on the current develop branch? I could not reproduce on centos. Thanks!

[crcrewso]

Which gcc did you test with? I thought this was a 4.4 series issue.

…

On Jan 23, 2017, at 12:57 PM, Reid Townson ***@***.***> wrote: Is this still a problem on the current develop branch? I could not reproduce on centos. Thanks! — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or mute the thread.

[rtownson]

Good point - I was on gcc 4.8.5.

[Kawrakow]

Suggestion: discard the -ffast-math flag. Better (i.e., faster and less dangerous) outcomes with modern gcc versions are achieved using -march=native -mtune=native

[ftessier]

Will do, thanks for the info!