Re: removal of static vars analysis code from mainline.

[Jan Hubicka <hubicka at ucw dot cz>]

> 
> 
> Mark Mitchell wrote:
> 
> >Kenneth Zadeck wrote:
> >
> >>Mark,
> >>
> >>Jan has explained to me that it appears to be impossible for me find 
> >>all of the variables in the mainline branch.  It appears that at the 
> >>time cgraph is run, some of the variables have disappeared from the 
> >>trees and are only in the asm file and that the front ends are still 
> >>keeping some of the variables a secret.
> >>
> >>He claims to have fixed these problems (or at least all the ones he 
> >>has found) on the tree profiling branch but believes that is not 
> >>worthwhile to port onto the mainline so the correct solution appears 
> >>to be remove my code from the mainline and work on the tree profiling 
> >>branch until things open back up on the mainline.
> >>
> >>Enclosed is a patch to remove my code. 
> >
> >
> >How sad.  Patch OK.
> >
> I would feel a lot worse if Jan had not fixed the code in tree-profile 
> branch.  I will make it work there and get it into 4.1.

Kenny,
I've been thinking about your code in the train now and I think sad
problem we run into was that I didn't realized what you really need for
your analysis (ie we always chatted only about the functions, not
variables tought it is trivial to notice that these are needed too) and
you simply assumed sane interfaces on the mostly half finished IPA
interface to be sane.

Your work is actually the second IPA pass we have (I implemented only
inlining) and it is very first that cares about the variables and thus
you sadly run into more issues than you would do if you choosed some
less pioneering work ;))

Since tree-profiling really fixes all the underlying problems I
understand to, I would suggest that I will avoid merging the revert to
the branch and apply modified version of your patch introducing master
clones (I would like to keep the files named ipa-something).
(steven, is that OK with you if I do so this week before cfg inliner or
shall I wait?)

I will try to then look into your code and hook it into current
tree-profiling interface for IPAs (ie the passmanager and the cgraph
structures) and send you the patch for review.  This way I guess I will
learn how exactly the aliasing work and hopefully will ensure if
something is still wrong and I think I can give you the more complette
testcase how things work so you can easier comment on it.
Last week I made similar experiment with the Razya's ipcp
implementation and it was ineed interesting try.

Does that sound sane to you?
Honza
> 
> Thanks for the speedy review.
> 
> kenny