Re: [tree-ssa] RFC: Making control flow more explicit

[Diego Novillo <dnovillo at redhat dot com>]

On Thu, 2003-08-07 at 17:06, Jason Merrill wrote:

> Is that a problem?  What would be the benefit of delaying the move to SSA
> form?
> 
So far the problems have been in stability more than anything.  Since
some forms of control are still hidden by the IL, doing some things like
sticking a label at an arbitrary spot in the code to make other parts
unreachable, leads to problems.  This was something that Jeff Law
recently ran into when doing some of the control transformations he's
been implementing.

The one I originally started playing with was code insertion.  We do
quite a bit of work for all the various control flow structures we need
to update.

Now, it may be that we do this lowering while in SSA form when we're
ready to do control flow manipulations of some sort.  What initially led
me down this path was the thinking that this would simplify the CFG and
IR manipulation code.


> > We probably do want to define a lower GIMPLE form.  Syntactically it
> > would be no different.  It would only accept fewer tree nodes.  Perhaps
> > we could also code the IL level in the trees to know what level we're
> > at?
> 
> I don't think that makes sense; replacing a leaf node could invalidate such
> flag on a containing node.  Rather, I think we should have global state
> indicating what level we're at, and enforce the relevant restrictions.
> 
OK.  That would work too.


Diego.