my coalescer on ML kit: bad

[Henry Cejtin]

Well,  that was certainly not very impressive.  For the ML kit, here are some
statistics:
                                    old coalescer           new coalescer
        number of chunks                   11,425                   7,677
        number of trampolines         217,944,056             206,952,722
        number of returns             220,922,989             220,945,534
        number of inter-chunk returns 157,022,980             103,291,304
        number of limit checks        119,171,837             119,188,003
        CPU time                           164.34                  165.45

I'm surprised that  the  number  of  trampolines  (=  inter-chunk  transfers,
right?)  only went down by 5.3% when the number of chunks went down by 48.8%.
Still, that could be.  I don't understand why the number of returns went  up.
The  number  of inter-chunk returns went down by 52%, which is good.  Why did
the number of inter-chunk returns go down by  so  much  when  the  number  of
trampolines went down hardly any?  Is it really just luck, and almost all the
old inter-chunk calls were still inter-chunk?

Any idea why did the number of limit checks changed (up by epsilon)?  Why did
the CPU time go up (also only by epsilon).  coalescer trampolines