[llvm-commits] SSI patch

[Eli Friedman]

On Sat, Jun 27, 2009 at 9:54 PM, Nick Lewycky<nicholas at mxc.ca> wrote:
> The one thing to remember is that when placing sigma's in the first
> place you use a very simple flow of "is it a CmpInst which flows to a
> TerminatorInst". That's probably good enough for ABCD but not for a
> GCC-like VRP algorithm. You don't need to worry about that though,
> whoever tried to implement a GCC-like VRP gets to worry about it.

The idea here is that there's some sense of whether a branch actually
leads to distinct properties along the various paths, and we don't
want to bother renaming an instruction if all the interesting
properties are the same along both branches?  That somehow seems
sensitive to what exactly the pass is doing... I guess adding some
sort of callback mechanism when it becomes needed wouldn't be too
hard, though.

-Eli