This is the mail archive of the fortran@gcc.gnu.org mailing list for the GNU Fortran project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [Patch,Fortran] PR 45756 Fix muliple-decl issue with PARAMETER

Paul,

On 09/28/2010 04:19 PM, Paul Richard Thomas wrote: 
[PR 44945]
Ahhh yes, I looked at the PR and am now reminded - you are right.  I
even made a relatively relevant comment :-)

Are you going to work on that PR?

(However, the example also shows that having the correct decl does not
necessarily make the program faster - actually, fatigue [cf. PR] is 40%
slower with LTO.)
I didn't know that it is that bad. Why is that?

-fno-function-inline fixes it. The problem is seemingly that functions get inlined which shouldn't while the really hot function does not get inlined. Thus, the problem are the parameter settings for inlining which give up too soon (some count exceeds the limit) - and seemingly also some deeply nested loops are not recognized as being hot although they are. Like all inlining decisions, its highly heuristic and probably needs to have some extra tuning for Fortran, which tends to have deeply nested loops and does not have the "inline" attribute C has.

Tobias

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]