[ruby-core:47396] [ruby-trunk - Feature #6962][Open] Use lighter hash structure for methods table, instance variable positions, constants

Issue #6962 has been reported by funny_falcon (Yura Sokolov).

----------------------------------------
Feature #6962: Use lighter hash structure for methods table, instance variable positions, constants
https://bugs.ruby-lang.org/issues/6962

Author: funny_falcon (Yura Sokolov)
Status: Open
Priority: Normal
Assignee: matz (Yukihiro Matsumoto)
Category: core
Target version: 


I've experimented with replacing `struct st_table` with lighter hash structure.

I were inspired by Lua tables, so that hash were created similar:
- open addressing but with elements chaining instead of any kind of probing
- there is sibling probing on insert, to ensure cache locality
- on insert, already inserted collision item which is not in main position is moved to free place,
  so that chains started from different main positions are not joined.
  This allows to keep chains short.
- struct header is 24 bytes on 64bit platform (compared with 40 bytes for `st_table`), so that i've
  placed them directly into `rb_classext_struct` instead of separate allocation.

Since keys are integers (ID), there is no separate field for hash. And I used `unsigned int` instead of
`unsigned long` for key, cause ID almost never reach 2**32. It allows to pack whole entry into 16 bytes
on 64bit platform (compared to 48 bytes for `st_table_entry`).

Such structure gives about 6% improvement without GC tuning, and 10% improvement when GC is tuned
(I've used `RUBY_GC_MALLOC_LIMIT=60000000 RUBY_FREE_MIN=200000`)

Also I add per class method cache instead of single global hash array. It has advantage over global
cache cause there is no collision over different classes and methods. But it seems that Rails applications
too often creates singleton classes, so that improvement is not very big (but no slowdown too).

Currently I have patch for Ruby 1.9.3 https://github.com/funny-falcon/ruby/compare/ruby_1_9_3...sparse_array/ruby_1_9_3

Single incompatible external api change is almost neglible, cause `rb_generic_ivar_table` methos is not very usefull
https://github.com/funny-falcon/ruby/compare/ruby_1_9_3...sparse_array/ruby_1_9_3#L5L852

Silly, but there is too many changes in trunk, so that, I could not simply rebase it, but ought to make whole work again.
I'll try to do it shortly.

With regards, Sokolov Yura aka funny-falcon.


-- 
http://bugs.ruby-lang.org/