http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60549
Bug ID: 60549
Summary: [4.9 Regression] Run time doubled for fatigue.f90 due
to SAVE changes
Product: gcc
Version: 4.9.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: fortran
Assignee: unassigned at gcc dot gnu.org
Reporter: burnus at gcc dot gnu.org
Follow up to PR 55207 and to r208590; see
http://gcc.gnu.org/ml/fortran/2014-03/msg00108.html
Since r208590, the execution time of Polyhedron's fatigue.f90 doubled from
~1.4s to ~2.9s.
I have not closely looked at fatigue.f90, however, I believe that this is due
to the following: Fatigue's main program has internal procedures, which access
variables defined in the main program.
As static variables usually global variables, GCC does not a good job at
optimizing static variables - even if it could in case they are defined within
one function only.
Possible solution: Small variables [of the MAIN program only!] should be
(again) created on the stack - possibly even (new!) when they do have an
explicitly set SAVE attribute. I think that's especially relevant for integer
and real variables.
I think it is impossible to see the difference from within the program; the
only exception are __attribute__((destructor)) - thus, if one wants to play
safe is to exclude those with TARGET attribute.
