Re: GCC3 to GCC4 performance regression. Bug?

[James E Wilson <wilson at specifixinc dot com>]

On Thu, 2005-03-24 at 15:52, Steven Bosscher wrote:
> I'd suggest trying -fmove-loop-invariants, and report a bug about
> that instead if it does not move those loop invariants.  We really
> should move away from loop.c anyway.

In general, yes, but we will probably always need some RTL loop
optimizations.  Lowering gimple to RTL may expose target dependent loop
invariants that were not present in the gimple.  Hence, we still need
the RTL loop pass to work.

There is also the issue of the special looping branches, which is a low
level target dependent issue that is probably better handled in RTL than
gimple.  The -fmove-loop-invariants option can't fix this problem, and
this results in a few percentage points of performance loss for this
testcase.  We need the RTL loop pass for this.

There is also the more pragmatic problem that we are suffering user
visible performance losses right now, and we shouldn't force users to
wait for future tree-ssa enhancements to fix them when an apparently
relatively simple RTL optimizer change can fix them.
-- 
Jim Wilson, GNU Tools Support, http://www.SpecifixInc.com