Re: bug in dominance algorithm

[Jan Hubicka <jh at suse dot cz>]

As I've mentioned in other email, I see why current behaviour is correct.

I am using postdominators to detect conditional branches branching to noreturn
calls.
Noreturn call is seen as block with no succesors and the conditional branch I
can find by edge between block that is not postdominated by call and block
that is.

That way I don't get confused by extra branching before reaching the call.
> But there is a middle way, which is slightly incorrect but IMHO gives more
> interesting information.  Suppose the following graph:
> 
> A-->C
> |   |\
> v   v \
> B   D-->E-->F
> |
> v
> EXIT
> 
> In some way, E could be seen as the immediate postdom of C, and also F
> postdominating C, D and E.  The "interesting information" I talked about
> is here, that "no matter which path is taken from C, nodes E and F are
> reached".  These kind of questions are the ones which usually are answered

Exactly this kind of behaviour I was somehow implicitly assuming.

> by dominance, so I think this is valuable information, even if strictly
> incorrect.  I.e. I would like to mark nodes without successors as being
> invalid for dominance questions, and otherwise provide the "dominators"
> and, may be, a way to distinguish reverse-unreachable node from normal
> ones.

Would be possible to modify algorithm to get this behaviour somehow
easilly?

Honza