[gclist] Compiler lifetime analysis

Hans Boehm boehm@hoh.mti.sgi.com
Mon, 19 Aug 1996 09:55:15 -0700


In a collector like ours, there's a different potential advantage to lifetime
prediction (in the nonguaranteed sense):  You can arrange to have the contents
of a region of memory (i.e. page) contain objects of the same age.  That means
writes to young objects are much less likely to pollute dirty bits of old
objects, giving you essentially the same advantages of a copying young
generation collector. (A lot of this seems to happen by itself anyway, somewhat
reducing the advantage.)

The trick is to implement lifetime prediction that's sufficiently accurate to
be useful, and sufficiently cheap that it still makes sense.  We haven't
succeeded at that.  But I haven't looked at this since the Barrett and Zorn
paper.

Hans

-- 
Hans-Juergen Boehm
boehm@mti.sgi.com