[PATCH] Fix information loss in branch probability computation.

[Duncan P. N. Exon Smith]

> On 2015 May 7, at 12:06, Duncan P. N. Exon Smith <dexonsmith at apple.com> wrote:
> 
>> 
>> On 2015 May 7, at 11:57, Hans Wennborg <hans at chromium.org> wrote:
>> 
>> On Thu, May 7, 2015 at 8:51 AM, Diego Novillo <dnovillo at google.com> wrote:
>>> On Thu, May 7, 2015 at 8:44 AM, Duncan P. N. Exon Smith
>>> <dexonsmith at apple.com> wrote:
>>> 
>>>> This assertion will fire under circumstances such as:
>>>> 
>>>> - `1` successor has a weight of `3`.
>>>> - `UINT32_MAX - 2` successors have a weight of `0` each.
>>>> 
>>>> This won't get rescaled by the new BPI, but it violates the local
>>>> assumptions (with the floor, the sum will be `UINT32_MAX + 1`).
>>> 
>>> Oh, absolutely. But that's not a problem now. Hans is going to rewrite
>>> the balancing algorithm to tolerate 0 weights. With that, I will be
>>> able to simplify the patch. I'll wait for that fix before committing
>>> (thanks Hans!)
>> 
>> I've updated the switch lowering code to handle 0 weights without
>> tampering with them in r236739.
> 
> Ah!  I didn't understand this was going on in parallel.  Sorry for being
> out of it.
> 
> I'll go back and review the patch.

So after r236739 it looks like you can drop your change to
`SelectionDAG`, in which case the patch LGTM!  Thanks Diego.