[PATCH] Fortran/PR31593 Speed up some loops

[Tobias Schlüter]

Hi,

we have a special case to deal with loops with step +-1.  The code 
generated in this case uses the do variable also as the loop counter. 
This saves some arithmetic, but has the disadvantage that if the do 
variable is used as an argument to a procedure, the compiler will have 
to assume that it changed all over the place, and therefore this can 
produce slower code.

In the PR, comment #23, is a testcase where the attached simple patch, 
which reintroduces the separation between loop count and do variable, 
while still saving some arithmetic, creates a 10% benefit.  I ran the 
polyhedron testsuite, but with good faith I can only say that the patch 
is performance neutral.  This may be because I used -funroll-loops where
the loop count is treated separately anyway.

Tested on i386-darwin, and I also checked that I can run the polyhedron 
benchmarks on i686-linux successfully with it.

Ok for the trunk?

Cheers,
- Tobi
-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: better_loops.diff.txt
URL: <http://gcc.gnu.org/pipermail/fortran/attachments/20090815/97f408fb/attachment.txt>