This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC 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: Enable SSA at -O0


On Wed, 23 Jul 2008, Mark Mitchell wrote:

> Richard Guenther wrote:
> 
> > > I agree it's the right long-term strategy, but how does this make 4.4
> > > better
> > > for users?  Jan mentioned eliminating -O0 code-quality regressions, but I
> > > didn't see any numbers showing that benefit in this thread.
> > 
> > We have the possibility to enable warnings that require flow-sensitive
> > analysis.  I think this is quite important for users.
> 
> Right, so there are other not-yet-implemented improvements built on this
> infrastructure that might be beneficial.  But, we don't have those yet, as
> part of this patch, and even if we did, it's a relatively minor benefit.
> 
> > I don't think this is a likely cenario.  But of course the same thing
> > might happen with the tuples changes, with the aliasing changes and
> > with the new register allocator
> 
> Yes.  And part of managing risk is not to have too many risky things all at
> the same time.  The other things you mention are all things that have been in
> progress for a long time, with lots of visibility.  In fact, you, Joseph,
> Jakub, and I have been discussing off-list the set of things we want to get
> into 4.4 -- and turning on SSA for -O0 isn't on that list.
> 
> > ... will be different in 4.5? 
> 
> > IMHO maintainability should win here if you can not come up with
> > a concrete problem (not just fear of potential problems).
> 
> Note that I said above that I agree this is the right long-term strategy.  If
> it were Day 1 of GCC 4.5, I'd have no concern.  It's a good move, and we'll
> work the problems out.
> 
> But in 4.4, we -- as you point out -- have a pretty long list of pretty major
> changes that we want to get in.  Past experience teaches that they'll all have
> a few problems.  This has been a long cycle, and it's likely to take a while
> to get all the bugs out of the new features, and get them tuned right for lots
> of architectures.
> 
> So, I just don't see what makes this a short-term priority.

As it is a basic requirement for the things that are not ready yet it
makes it more difficult to get them developed for the stage1 of 4.5.
Basically everyone relying on SSA at -O0 has at the same time the
obligation of making SSA at -O0 happen.  Honza spent quite some
amount of time fixing all the bugs SSA at -O0 uncovered, so it would
be unfair to not include this change both for him and for the people
that pushed him to do this as they need this change for future
development (ok, that's me mostly, and eventually Diego or Andrew
for RTL expansion work).

Richard.


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