------- Comment #2 from chrbr at gcc dot gnu dot org 2009-03-11 08:46 -------
I observed some large performance regressions in 4.3 and upwards for many
benchmarks for the superh targets, There are many causes but the main one is
reduced to the indirect+offset access :
int
foo (int tab[], int index)
{
return tab [index+1];
}
compiles (-O2 -fomit-frame-pointer) into
mov r5,r0
add #1,r0
shll2 r0
rts
mov.l @(r0,r4),r0
instead of
shll2 r5
add r4,r5
rts
mov.l @(4,r5),r0
Note that in more complex code the problem is emphasized because only r0
register class can be used as indirect register index, putting extra pressure
on reload.
It seems to be that the problem is in the way that the constant index is now
hidden by gimple, so we now have
return *(tab + ((unsigned int) index + 1) * 4)
instead of
return *(tab + 4B + (int *) ((unsigned int) index * 4))
It seems more easy to change gimple, but this is a target dependant
transformation. On the other hand the RTL code gen should be able to
redistribute the factorization, but that seems extra work to undo what was done
previously.
--
chrbr at gcc dot gnu dot org changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|RESOLVED |UNCONFIRMED
Resolution|INVALID |
Summary|[ |[SH] performance
| |regression: lost mov
| |@(disp,Rn)
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=39423
