Garbage collector reimplementation #25
Conversation
basically base frame just needs to exist as a regular gc value with no special caveats.
| /// iterate the generation, collect within each page | ||
| fn collect(gc: *Gc) void { | ||
| gc.generation +%= 1; | ||
| gc.markVal(gc.base_frame); |
There was a problem hiding this comment.
The previous root marking system also detected temporary values on the program's stack. Only marking values referenced by the virtual machine's stack leaves those values susceptible to being collected while in use.
There was a problem hiding this comment.
Aha, ok. How can I access those values during marking?
There was a problem hiding this comment.
That's what the @frameAddress stuff was for. It wasn't perfect by any means and seemingly didn't work on windows but I never got around to debugging it further.
There was a problem hiding this comment.
I have two ideas for a solution to this:
- Reference frames on the caller's stack during the callee's runtime
- Tracking the current call stack in the VM
I think I like the first better since it requires very few any additional mechanics to be added. These also would be reusable for creating error traces, or do error traces already exist? What do you think of these solutions?
| @@ -280,25 +274,23 @@ pub fn compileAndRun(vm: *Vm, file_path: []const u8) !*Value { | |||
| else => |e| return e, | |||
| }; | |||
|
|
|||
| var frame = Frame{ | |||
| const frame = try vm.gc.gpa.create(Frame); | |||
| frame.* = Frame{ | |||
There was a problem hiding this comment.
Removing the defer here leaks the stack and err_handlers and now that the frame is allocated it also needs to be freed.
There was a problem hiding this comment.
This was the source of a segfault that I spent an hour or so debugging. Since my implementation tracks the base frame like any other value, it gets collected after clearBaseFrame() is called and this shouldn't leak anything
There was a problem hiding this comment.
If it isn't stored on the heap, there has to be a special case added to either Value.deinit or gc.collect or both and that felt like a bit of a dirty solution
There was a problem hiding this comment.
Freeing the garbage collector pages doesn't free the array lists allocated with the gpa. I previously reset the frame value to an int to avoid the segfault:
defer frame_val.* = .{ .int = 0 }; // clear frameThere was a problem hiding this comment.
Which arraylists are you referring to? I feel that it would generally be preferable to be able to treat frames the same as any other value, or is there something I'm missing?
A pretty simple mark-sweep garbage collector. In most cases this should be similarly performant to a bump allocator.
The most interesting thing about it is the used/free object tracking, which uses a bit set in with a binary-heap indexing strategy to keep track of objects that are either in use or not collected yet, the idea being that looking for a free object in a relatively full list should remain fast. I'm not sure this is actually more useful than keeping a simple list of freed indices, but I thought the idea was worth taking a stab at.