Free too soon in mark_roots() (suspected bug)

[clinuxrulz]

I do not have a failing test case for this. However I believe this is not a safe place to free a node:

bacon-rajan-cc/src/collect.rs

Line 223 in 9d59ed4

free(s);

I believe the free-ing should be defered to the end of collect_roots, just like collect_white does.

The reason is that the early drop() it could be potentially be changing the shape of the memory graph that has already been traversed by mark_grey(). Potentially causing scan_black() not to restore the reference counts properly.

[clinuxrulz]

Also scan_black which is called in scan will occasionally need to cycle blacks back over the visited white nodes (in the depth direction if that makes sense). So trace needs the capability to loop around and revisit the same nodes. However the trace implementation for RefCell does not revisit the same nodes in a recursive loop-like style when they are already locked for reading.

That means that this library needs its own RefCell separate from the one supplied by the std.

[jrmuizel]

I believe I've fixed the issue that you pointed out in the first comment in 7cd96af.

I don't really follow your second comment. Perhaps you could create a test case that shows the problem?

[clinuxrulz]

Its been a while since I've looked at this code.
If I remember correctly:
In second comment just means that a user using RefCell will not allow you to visit the same gc node twice in the gc algorithm. Which can break the algorithm with resetting colors etc.
To avoid the problem trace can be done iteratively instead of recursively (add a stack vec from thread local and use it to convert recursive trace into an iterative trace)
If I find some time, I'll wip up a test case to demonstrate the problem.