Re: [tree-ssa] Switch stmts and inserting on edges

[Andrew MacLeod <amacleod at redhat dot com>]

On Fri, 2003-05-30 at 14:25, Michael Matz wrote:
> Hi,
> 
> On 30 May 2003, Andrew MacLeod wrote:
> 
> > The problem is that the preceeding case stmt drops into the default
> > case. Arg. That means the default block has 2 predecessors, and the
> > switch stmt has multiple successors, so the only way to insert is to
> > split the edge.
> >
> > Splitting that edge is not trivial...
> 
> It shouldn't be any more difficult, than splitting _any_
> (non-abnormal) edge.
> 
> > Since it can't be properly split,
> 
> Why not?

Well, not without mucking around in the switch stmt expression list
would be the qualifier.
> 
> > Id have to go to all the other predecessors, change them to goto a label
> > that would be inserted after the stmt Im trying to insert.
> 
> No, the way this would work is to preserve the old default label, to
> which the other blocks still have the edge, and instead redirect the
> target of the split edge (i.e. the switch->default) to be the new block,
> i.e.:

And I beleive thats the trick, I think its easier to change where the
fallthru edges go than it is change the switch->default..
> 
> > case blah:
> >   <...>
> >   goto new_label;
> >
> > default:
> >   <inserted_stmt>
> > new_label:
> >   default code.
> 
> 
> case blah:
>    <....>
>    goto default;
> new_label:
>    <inserted_stmt>
> default:
>    default code.
> 
> 
> Then only the edge from the switch needs to be redirected to "new_label".
> Hmm, or is it not possible to replace the default label in this way due to
> some inhomogenity in tree-ssa's representation of switches?
> 
I beleive the switch stmt has a list of case elements, and in order for
things to work, I'd have to arrange for the switch stmt to know that the
edge that use to go to a certain block now goes to a different block,
and I dont know how easy that is to map back to the switch case
expression's knowledge of its case targets.   Thats the main reason I
asked, I was hoping someone who knew cases better in tree form would be
able to tell me I dont have to hack around like Im afraid Im going to
:-). I suspect in order to do that Id have to go in and start looking at
the case expressions... At least the one for thatr edge. Im hoping I m
wrong...



Andrew