This is lighter-weight than building a full index->cursor table for
the string, adding a constant two words to the memory required to
store a string, as opposed to one word for every n characters. The
cached cursor is used for any string-ref operation requesting an index
after the most-recently-requested index, making potentially quadratic
repeated string-ref procedures run in linear time. In theory, it could
also use a heuristic to speed up moving backwards through the string
when it thinks that moving the old cursor backwards would be faster
than starting again at the start of the string. In practice, my
logging of when the cached cursor is actually reused during the Chibi
compilation and startup process shows that the most common case of
moving backwards is going back to the start of the string anyway.
Benchmarks to follow.
Without this chibi can crash after 129 open file descriptors.
Note the bug referenced would also indirectly be fixed if
process->string-list properly closed its ports, but we
shouldn't rely on that.
This fix is rather dumb, but it prevents things from crashing when
forking a lot and creating file handles. I assume that this is where
the filehandles go, but I don't have a good guess.
Replace explicit recursion by heap allocations
in sexp_mark_one code.
This prevents crashes caused by stack overflow.
In particular, fixes issue #601.
As an optimization, allocate a fixed sized stack buffer first,
which should be enough for "normal" uses.
When that stack overflows, switch to heap.
Also, store "ranges" on the stack, instead of the actual sexp's,
using the fact that sexp's of a single parent are continous in memory.
This patch doesn't remove recursion on the context saves
because it didn't seem like they overflow in practice.
But changing that is simple having the stack interface.
