This is the mail archive of the 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: Fixes to ipa branch

Razya Ladelsky wrote:
> Jan Hubicka <> wrote on 24/08/2006 00:48:07:
>>> Hi,
>>> This patch includes several fixes to ipa branch.
>>> The targhooks and rs6000 fixes are necessary for compilation on 
> powerpc 
>>> and 
>>> also for testing the powerpc suite.
>> What are precisely those changes shooting for?  Is it something that can
>> be brought in by another merge?  (I am leaving for two weeks of vacation
>> at friday, so I will unlikely to be able to do it before that, but they
>> looks like something that is quite independent on IPA branch).
> I think so, yes.
> These are parts that I took from mainline.
> We had troubles bootstraping the branch on powerpc without them.
>>> The rest of the changes refer to the dominance info (they were 
> suggested 
>>> by Daniel Berlin a while ago)
>> My approach to dominance info so far has been to always calculate just
>> one (ie not have it privatized).  I am not oposed to privatizing it at
>> all (just we probably should be sure to not leave it hanging around all
>> CFGs for no good reason), but I wonder what do you need it for?
> Well, when I ran some tests (for instance art benchmark) with matrix 
> flattening enabled,
> I got segmentation fault during verify_flow that referred to dominance 
> information.
> Perhaps there is another way to handle it.
> Paulo wrote that he had another way of working around it:
> "I was working around 
> this by calling free_dominance_info before every pop_cfun."
> Your thoughts?

Jan, if you don't do this, it means you can't have dominance info in an
IPA pass, unless you very explicitly blow away the dominance info before
looking at another function.
This seems, uh, silly.

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