Verified Commit 347ade20 authored by
Fix various GC bugs and improve GC performance
This fixes various Immix related bugs and makes some improvements to the collector's performance. Bug fixes --------- The following bugs have been fixed: 1. Evacuating required one or more fragmented blocks, which were set based on histogram statistics. These statistics were only populated when there were one or more fragmented blocks. This resulted in evacuating never happening. Evacuating now happens if the previous collection increased the allocation threshold. 2. The fragmentation status of a block was never cleared after a collection. This could result in an increase of fragmentation if the collector statistics deemed evacuating to not be needed. 3. The remembered set using a HashSet meant that updating forward pointers in the remembered set would result in duplicate entries, as the hash codes would be different. 4. Objects promoted to the mature generation would not be remembered if they contained any young pointers not promoted. We now trace mature objects to check if they need to be remembered, instead of always remembering them; drastically reducing the size of the remembered set. Improvements ------------ 1. The index to the eden space has been changed to a u8, allowing us to add some additional fields without growing the size of a Bucket. 2. We no longer update line maps of mature blocks when performing a young collection. This is not needed in a young collection. 3. We now use a chunked list for the remembered set and an additional object bit to prevent duplicate entries. This makes the remembered set more memory efficient, and removes the need for hashing. This does require the use of a 64-bits architecture, but Inko never really officially supported 32-bits anyway. 4. Objects with a single hole are not taken into account when calculating fragmentation statuses. This way we don't overestimate the number of free lines as much, reducing the amount of evacuating objects. 5. Messages are now copied into a process' local heap directly, instead of being copied into a separate mailbox heap. This reduces memory needed when sending messages, overall memory usage, and quite drastically simplifies the mailbox and garbage collection implementation. Access to the mailbox is synchronised using a spinlock, which is fast enough to be used by a mutator. 6. Various garbage collection timers (e.g. the time spent preparing a collection) that weren't very useful have been removed. 7. Parallel tracing performance has been improved by tracing objects in parallel, instead of only processing stack frames in parallel. When moving of objects is not needed this can significantly reduce garbage collection timings. 8. The INKO_CONCURRENCY variable has been replaced with separate variables, as greater control is needed over some of the thread pool settings. 9. GC timings can now be printed when setting INKO_PRINT_GC_TIMINGS to "true". This is an internal option and will only stick around until we have a better way of aggregating and presenting VM/GC statistics. 10. Updating of garbage collection statistics at the end of a collection has been improved a bit, removing the need for counting the total number of blocks separately. The performance improvement of this is minor, but it does clean up the code quite a bit.
Showing with 2980 additions and 3240 deletions
This diff is collapsed.