Re: [tree-ssa] replace sibcall.c take 3

[Jan Hubicka <hubicka at ucw dot cz>]

> On Tue, 2003-11-18 at 16:05, Jan Hubicka wrote:
> 
> > How should I do it?  For mainline we do have 24 hours of time for fixing
> > the failure.
> >
> On the branch we have until one of the testers dies, I find about it and
> shoot down the patch.  It can take more than 24 hours, if you're lucky
> :)
> 
> >   I am really not aware of any procedural error I made with
> > these patches and I am still having troubles reproducing problems you
> > are seeing.
> > 
> This one seems sneaky.  On one of my boxes, a --disable-checking
> bootstrap seems to have succeeded.

I do bootstrap with default flags, so this is not the case.
Do you use some special flags?  THat may be explanation.

> 
> Once I commit the reverted patch, try the combined patch in a variety of
> systems with and without checking disabled.  If that works, submit it
> again.
OK, I already made jumbo patch and will run it on x86/x86-64/alpha.
Sounds fine?

The problem with tailcall is that it really needs TREE_ADDRESSABLE flag
to be right.  I've added more sanity checks and found another latent bug
in tree-mustalias.  THe problem is that PHI nodes are not considred at
all (because they are not staements).  WHen the variable is addressed
only via constant in PHI node, we lose.

I am not sure how to fix that, as the operands are managed in separate
lists.  Since you are author of that code, could you please fix this
bug?  I will re-do testing afterwards.

Real problem seems to be that the branch is very well tunned to
bootstrap in precisely these environment it is tested so we do get a lot
of breakage when the environment change.  I hit that problem with the
cgraph changes for first time (it took me two week to get across latent
bugs).  We do regular bootstraps on x86-64 and they seem to be broken
most of time, so I will try to figure out if we can do more testing on
SuSE side.

Honza
> 
> 
> Diego.