## Implementation Summary
* Only keys and only values of hash table are contiguous to eliminate
structure padding.
* Change upper limit of `iv_tbl` size to `UINT16_MAX` (it seems to be
acceptable in mruby because the total number of classes/modules
immediately after starting Redmine is 20,000 or less).
* `iv_tbl*` point hash buckets directly.
## Benchmark Summary
Only the results of typical situations on 64-bit Word-boxing are present
here. For more detailed information, including consideration, see below
report (although most of the body is written in Japanese).
* https://shuujii.github.io/mruby-iv-benchmark
### Memory Usage
Lower value is better.
| iv_tbl Size | Baseline | New | Factor |
|------------:|---------------:|---------------:|-----------:|
| 4 | 88B | 52B | 0.59091x |
| 30 | 536B | 388B | 0.72388x |
| 100 | 2072B | 1540B | 0.74324x |
| 200 | 4120B | 3076B | 0.74660x |
Although not mentioned in the above report, the memory usage of `mrbtest`
(full-core gembox) is as follows in the result by Valgrind.
* Baseline: 108,086 allocs, 16,313,122 bytes allocated
* New: 94,273 allocs, 15,875,214 bytes allocated
### Performance
Higher value is better.
#### `mrb_obj_iv_set`
| iv_tbl Size | Baseline | New | Factor |
|------------:|---------------:|---------------:|-----------:|
| 4 | 88.63003M i/s | 92.60611M i/s | 1.04486x |
| 30 | 32.97066M i/s | 25.25095M i/s | 0.76586x |
| 100 | 16.33224M i/s | 22.74998M i/s | 1.39295x |
| 200 | 5.64484M i/s | 6.79949M i/s | 1.20455x |
#### `mrb_obj_iv_get`
| iv_tbl Size | Baseline | New | Factor |
|------------:|---------------:|---------------:|-----------:|
| 4 | 217.58391M i/s | 237.59912M i/s | 1.09199x |
| 30 | 139.56195M i/s | 160.49470M i/s | 1.14999x |
| 100 | 143.09716M i/s | 190.95047M i/s | 1.33441x |
| 200 | 89.75291M i/s | 134.78717M i/s | 1.50176x |
### Binary Size
Lower value is better.
| File | Baseline | New | Factor |
|:------------|---------------:|---------------:|-----------:|
| mruby | 697,520B | 697,520B | 1.00000x |
| libmruby.a | 1,046,570B | 1,046,682B | 0.99989x |
## Note
The address in `struct RObject::iv` may change after initialization because
`iv_tbl*` points directly to hash buckets. Therefore, the address cannot be
copied and shared when include/prepend. So, when sharing `iv_tbl`, refer to
it via the sharing source class. As a result, the following bug have also
been fixed.
* [An `iv_tbl` is not shared when a class includes or prepends an empty module](https://gist.github.com/shuujii/0ac23fa24b0c55b2c602b534d81e4a95)
They are basically the copy of instance variable tables. On my Linux
box, memory consumption of `mrbtest` measured by `valgrind` is:
- old: 17,683,830 bytes
- new: 14,283,749 bytes
* The `Fixnum` constant is now an alias for the `Integer` class.
* Remove `struct mrb_state::fixnum_class` member.
If necessary, use `struct mrb_state::integer_class` instead.
- `MRB_WITHOUT_FLOAT` => `MRB_NO_FLOAT`
- `MRB_USE_FLOAT` => `MRB_USE_FLOAT32`
The former is to use `USE_XXX` naming convention. The latter is to make
sure `float` is 32bit float and not floating point number in general.
Introduced `MRB_NO_METHOD_CACHE` which is inverse of `MRB_METHOD_CACHE`
that should be enabled intestinally. In addition, the default cache is
made bigger (128 -> 256).
#### Example (with `MRB_METHOD_CACHE`)
```ruby
GC.start
c = Class.new
p c #=> #<Class:0x7fd6a180e790>
c.new #=> cache `c.new`
c = nil
GC.start #=> `c` is GCed
r = Range.dup
p r #=> #<Class:0x7fd6a180e790>
# [same pointer as `c`]
r.new(2, 3) #=> ArgumentError: 'initialize':
# wrong number of arguments (2 for 0)
# [`c.new` is called instead of `r.new`]
```
#### Cause
An entry of method cache is identified by class pointer and method
id. However, reusing memory after GC may create a class with the same
pointer as the cached class.
#### Treatment
Cleared method caches of the class when the class is GCed.
Because location info (file name and line number) is kept in the backtrace,
it should not be kept in the result of `inspect` (and the exception object
itself), I think.
### Example
```ruby
# example.rb
begin
raise "err"
rescue => e
p e
end
```
#### Before this patch:
```
$ bin/mruby example.rb
example.rb:2: err (RuntimeError)
```
#### After this patch:
```
$ bin/mruby example.rb
err (RuntimeError)
```
Before this patch:
$ bin/mruby -e 'p Struct.new("A-")'
#=> Struct::"A-"
After this patch:
$ bin/mruby -e 'p Struct.new("A-")'
#=> NameError: identifier A- needs to be constant
Now the method tables (in classes/modules and caches) keeps C function
pointers without wrapping in `struct RProc` objects. For the sake of
portability, `mrb_method_t` is represented by the struct and union, but
if the most significant bit of the pointer is not used by the platform,
`mrb_method_t` should be packed in `uintptr_t` to reduce memory usage.
`MRB_METHOD_TABLE_INLINE` is turned on by default for linux.
To enable method cache, define `MRB_METHOD_CACHE` or
`MRB_METHOD_CACHE_SIZE`. The cache size must be power of 2.
The default cache size is 128.
The measurement:
I measured simple benchmarks found in benchmark/ directory. With
method cache enabled, we gained 6-8% performance improvement, with
97-99% cache hit rate.
The change removes several internal instance variables used by
class name management. The variables `__classid__` and `__classpath__`
are no longer available. `__outer__` is used only for unnamed outer
classes/modules (and will be removed after they are named).
[Important note]
Along with this change we removed several public functions.
- mrb_class_outer_module()
- mrb_class_sym()
We believe no one have used those functions, but if you do, please
ask us for the workaround.
It isn't needed as it's very similar to mrb_define_method_raw() and also
there's only one place where mrb_proc_ptr() actually has to be called.
Inspired by @furunkel's method cache patch (#2764).
KOBAYASHI Shuji