-
Eileen M. Uchitelle authored
* Avoid pinning `storage_head` in `iseq_mark_and_move` This refactor changes the behavior of `iseq_mark_and_move` to avoid pinning the `storage_head`. Previously pinning was required because they could be gc'd during `iseq_set_sequence` it would be possible to end up with a half build array of instructions. However, in order to implement a moving immix algorithm we can't pin these objects so this rafactoring changes the code to mark and move. To accomplish this, it was required to add `iseq_size`, `iseq_encoded`, and the `mark_bits` union to the `iseq_compile_data` struct. In addition `iseq_compile_data` sets a bool for whether there is a single or list of mark bits. While this change is needed for moving immix, it should be better for Ruby's GC as well. * Don't allocate mark_offset_bits for one word If only one word is needed, we don't need to allocate mark_offset_bits and can instead directly write to it. --------- Co-authored-by:
Peter Zhu <peter@peterzhu.ca>
Eileen M. Uchitelle authored* Avoid pinning `storage_head` in `iseq_mark_and_move` This refactor changes the behavior of `iseq_mark_and_move` to avoid pinning the `storage_head`. Previously pinning was required because they could be gc'd during `iseq_set_sequence` it would be possible to end up with a half build array of instructions. However, in order to implement a moving immix algorithm we can't pin these objects so this rafactoring changes the code to mark and move. To accomplish this, it was required to add `iseq_size`, `iseq_encoded`, and the `mark_bits` union to the `iseq_compile_data` struct. In addition `iseq_compile_data` sets a bool for whether there is a single or list of mark bits. While this change is needed for moving immix, it should be better for Ruby's GC as well. * Don't allocate mark_offset_bits for one word If only one word is needed, we don't need to allocate mark_offset_bits and can instead directly write to it. --------- Co-authored-by:
Peter Zhu <peter@peterzhu.ca>
Loading