This is the mail archive of the
mailing list for the GCC project.
Re: [PATCH, pretty-ipa merge 0/4] IPA-SRA
On Wed, Jun 24, 2009 at 06:11:50AM -0700, H.J. Lu wrote:
> On Wed, Jun 24, 2009 at 4:20 AM, Martin Jambor<firstname.lastname@example.org> wrote:
> > Hi,
> > this patch-set is a proposed merge of IPA-SRA from the pretty-ipa
> > branch to the trunk. I have talked about IPA-SRA at the summit and it
> > is also described in the paper. Shortly, it changes the parameters of
> > local functions in the following ways:
> > - removes parameters that are not used at all,
> > - passes parameters by value instead of by reference if possible and
> > hopefully profitable
> > - passes only parts of aggregate parameters if only those are used
> > in the function (often combined with the above transformation)
> > The pass has been sitting in pretty-ipa for some time, I have
> > bootstrapped it on i386-linux and x86_64-linux where I also ran the
> > regression testsuite without any new problems popping up.
> > I'd like therefore ask for the final review and hopefully an ACK to
> > commit to trunk.
> I think you should fix all new SRA regressions first before
> another SRA related merge.
The only remaining serious SRA regression that I know of is PR 40493.
I hope I have a patch for that, in fact I am in the middle of writing
a change log right now and am about to bootstrap and test it. If
everything goes fine, I'll post it in a few hours. Don't worry, this
has been a very high priority for me this week.
There is another PR which I still should look at to determine what
effect the intraprocedural SRA has (which is however a possible
compile time regression and not an ICE or miscompilation) and that is
PR 33928. I believe IPA-SRA can be reviewed and even committed even
if it takes me a while to get to that. Moreover, IPA-SRA shares
analysis with the intraprocedural SRA but the modification almost
completely separate (except scan_function() that is).
Or are there any other SRA regressions that I do not know about?