[llvm-commits] [PATCH] Promote all possible memory to registers

[Kenneth Uildriks]

On Sun, Oct 24, 2010 at 9:51 AM, Chris Lattner <clattner at apple.com> wrote:
> Splitting scalar and memory GVN apart from each other isn't practical: each unblocks important cse's in the other.
>
> -Chris
>

I'm thinking that if multiple conceptually different optimizations are
living in one pass so that they can be repeated until neither helps
the other, PassManager really ought to handle that sort of thing
instead somehow.  Maybe have a pass request to be repeated after some
other pass (if it is turned on) changes the code.  Then each can point
to another that way and be alternated until they stop finding
improvements.  We'd have to make sure that they actually stop at some
point, but the same is true for the present GVN's loop inside of
runOnFunction.