Are new sequences really O(1)?
Marcin 'Qrczak' Kowalczyk
qrczak at knm.org.pl
Tue May 31 15:15:55 EDT 2005
"Simon Marlow" <simonmar at microsoft.com> writes:
> Nice trick. Unfortunately the same assumptions don't hold for GHC's
> garbage collector - objects are aged in the youngest generation, so
> it is usually at least two GCs before an object is promoted. We
> could still do the same trick, but instead we'd have to find maximum
> stack depth at which all objects below are in an older generation.
I don't know how to do it under these assumptions, i.e. how to find
> Incedentally, how do you mark the stack depth? Overwrite a return
> address with a special one, and keep the real one around in a known
This is the first thing I tried :-) and even implemented it before I
realized that it doesn't work under my assumptions. The problem is
that I pass the return address in a virtual register, and it's saved
on the stack by the callee only if it performs some non-tail calls.
The return address is saved at the end of the stack frame.
This means that a tail call followed by a non-tail call might read the
special return address from the stack, but without jumping through it
immediately. This return address is put again on the stack by a
different function, with a possibly different stack frame size, so it
lands in a different position on the stack, and thus it can't be found
when GC wants to restore it.
While changing the stack frame layout could perhaps make this
workable, I found a much simpler solution. Since forever each stack
frame contained a pointer used only by the GC and by stack trace
printing. It points to a static structure containing the frame size
and return address -> source location mapping. This pointer is now
marked in the lowest bit by the GC. Pushing a fresh stack frame always
puts an even pointer.
BTW, a few days earlier I fixed a similar behavior for extensible
arrays. Most mutable objects use a write barrier which stores changed
locations, but arrays store references to changed objects because
their payload is malloced and may be reallocated without a GC. This
meant that code which repeatedly appends to a given array has O(N^2)
complexity for huge arrays, as each GC scans the whole array. So I now
store the size of the initial part of the array which is known to be
unchanged since last GC. This is updated manually by particular
__("< Marcin Kowalczyk
\__/ qrczak at knm.org.pl
More information about the Glasgow-haskell-users