The sorted array binary search implementation no longer uses the inline
cache array, so remove all MRB_INLINE_METHOD_CACHE definitions and
related code.
By replacing the open‐addressing hash with a sorted array and binary
search, we eliminate tombstone management and improve cache locality of
method entries. This preserves the original contiguous values+keys
layout and peak memory usage, while simplifying growth logic and
delivering lookup performance gains.
Preserve the original iv_tbl heap layout and allocation pattern, but
switch iv_put to maintain sorted keys and iv_get/iv_del to perform
binary search. This eliminates extra probing overhead, improves
read-heavy lookup performance for small tables, and incurs zero
additional allocations.
My previous attempts to improve method table (mt_tbl) performance by introducing a load factor and adjusting the initial allocation size unfortunately led to undesirable increases in memory consumption.
Given that memory usage is a primary concern, I've reverted the following changes:
- Removed the load factor based rehashing logic.
- Removed the related definitions for the load factor.
- Ensured the initial allocation size is 8.
This restores the method table to its original behavior, where it rehashes only when the table is completely full or during initial allocation. This should bring memory usage back to its baseline level prior to these optimization attempts.
The method table (mt_tbl) in src/class.c previously only rehashed when it became completely full. With linear probing, this could lead to significant performance degradation for lookups and insertions as the table approached full capacity.
This change introduces a load factor (MT_LOAD_FACTOR_NUM/MT_LOAD_FACTOR_DEN, set to 3/4 or 0.75). The mt_put function now checks if adding a new element would cause the table's size to meet or exceed this load factor relative to its allocated capacity. If so, it triggers a rehash before inserting the new element.
This helps maintain more empty slots in the hash table, improving the average-case performance of linear probing and reducing the likelihood of worst-case scenarios. The existing initial allocation size and doubling strategy for rehashing are retained.
ADDI/SUBI may fall back to method call that may clear block argument
place holder, which may be a live register. So we cannot directly call
ADDI/SUBI over local variables.
Since we have introduced lrama, everyone can generate same `y.tab.c`
on any platform, without installing Bison. That was the reason we have
removed `y.tab.c` from the repository. But this change cause #6515 and
bothered out-of-tree builds. So we (reluctantly) added `y.tab.c` again.
Yukihiro "Matz" Matsumoto