Re: [tree-ssa] Latent edge insertion bug

[Dale Johannesen <dalej at apple dot com>]

On Dec 23, 2003, at 1:57 PM, law@redhat.com wrote:
> Note carefully that the initialization of iftmp.4_7 occurs _after_ the 
> PHI
> node which references iftmp.4_7.   Not good.
>
> I haven't seen this trigger with the branch sources, but I'm pretty 
> sure
> it is a latent problem.  I can easily trigger it with the code which 
> threads
> through blocks with real statements.

I have seen it in one of the Specmarks.  I was thinking about how to 
fix it
when I updated and it went away...

> Anyway, this patch avoids using the destination block for the edge
> insertion when the destination block has a PHI node.  We could possibly
> actually look at the instructions we want to insert and the PHI nodes
> and see if there is a conflict, but it doesn't seem to be worth the
> headache.

Doing that was mostly what I was thinking about.  If you don't bother, 
that
leads to splitting a fair number of edges that don't really need to be 
split;
just how bad is that?