[llvm-commits] [patch] Add dominance computation for edges

[Bill Wendling]

On Aug 6, 2012, at 12:15 PM, Rafael Espíndola <rafael.espindola at gmail.com> wrote:

>> I don't understand this at all. I don't see an 'invoke' in the example. And the '%x' in '%foo' cannot be replaced by '0' because it may not be '0' (i.e., it took the 'bb1' route to 'bb2).
> 
> Exactly and exactly :-)
> 
(I need to remember to drink caffeine before responding to emails.)

> Sorry if it wasn't clear. The idea of the refactoring is to allow the
> new methods to be used in cases where we *don't* have an invoke, but
> we have a case where some information is available about an area
> dominated by an edge.
> 
> I want GVN to know that it can replace the 'x' in 'cond', but not the
> 'x' in 'foo', which is exactly what this analysis tells it (once GVN
> is modified to use it).
> 
Okay. That sounds better. :) It looks good to me if no one else objects.

-bw