[PATCH] D19821: [EarlyCSE] Port to use MemorySSA (disabled by default). NFC.

[Philip Reames via llvm-commits]

On 05/17/2016 11:40 AM, George Burgess IV wrote:
> > There are a couple of FIXME comments in this change that identify 
> cases where MemorySSA is maybe being too conservative (e.g. when 
> dealing with fence release instructions and load atomic instructions).
>
> Yup! D16875 that should make MemSSA more willing to handle 
> atomics/volatile ops. Philip gave it a verbal LGTM*, but IIRC he 
> thought it may be better to wait until we get MSSA integrated into a 
> few passes (to get any big bugs worked out) before we start making it 
> super aggressive with memory ordering. I have no opinion on when it 
> should land; if we want it in now, I'm happy to commit it today.
>
> > Similarly, what do you think of Phillip's suggestion to look at 
> using ValueHandles in MemorySSA to make removal invalidating more 
> automated?
>
> I was planning for that to come with the overhauled cache (most recent 
> attempt is abandoned at D19503). I should have more time to take 
> another stab at it this weekend. If not, I'll at least end up tweaking 
> the current implementation to use VHes.
>
>
> * - (To be as clear as I can be, it was an LGTM with comments. 
> Addressing the comments required refactoring the code a bit. I'm 
> assuming the refactoring doesn't matter, and the LGTM still applies.)
It does.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20160820/224c1cf0/attachment.html>