[PATCH] Fix PR 23525 - Separate header mass propagation in irregular loops.

[Duncan P. N. Exon Smith]

> On 2015-Jun-10, at 08:43, Diego Novillo <dnovillo at google.com> wrote:
> 
> > I think you have two options:
> >
> >   - Force the size of the `BackedgeMass` vector to be exactly
> >     `NumHeaders`.  In this code, do a binary search through the range
> >     between `NodeList.begin()` and `NodeList.begin() + NumHeaders` to
> >     find the index of the header in `NodeList`, and use the same index
> >     for `BackedgeMass`.  This will have logarithmic runtime cost in the
> >     number of loop headers (locally), and linear memory requirements
> >     overall.
> >   - Use a `SmallDenseMap<BlockNode, BlockMass, 2>`.
> >
> > I prefer the former approach, unless you have some reason against it?
> 
> No particular preference. The map approach seems a tad more convenient to implement, though. Seems to me that it would also make the code more readable?
> 

Seems like adding a map unfairly punishes the compile-time of the normal
case, where there's exactly one loop header.  But if you don't have any
compile-time regressions in LNT or in a clang LTO bootstrap, it doesn't
really matter though.