On 64-bit CPU, there is padding in `RBasic`, so reorder the fields and use
it as buffer of embedded string. This change allows 4 more bytes to be
embedded on 64-bit CPU.
However, an incompatibility will occur if `RString::as::ary` is accessed
directly because `RString` structure has changed.
I think the string buffer of NOFREE string always exists and does not need
to be released, so it can be shared as another NOFREE string.
Also changed the `mrb_shared_string` field order so that eliminate padding if
`int` and `mrb_int` sizes are less than pointer size.
The binary sizes (gems are only `mruby-bin-mruby`) are reduced slightly in
my environment than before the introduction of new specifiers/modifiers
(5116789a) with this change.
------------+-------------------+-------------------+--------
BINARY | BEFORE (5116789a) | AFTER (This PR) | RATIO
------------+-------------------+-------------------+--------
mruby | 593416 bytes | 593208 bytes | -0.04%
libmruby.a | 769048 bytes | 767264 bytes | -0.23%
------------+-------------------+-------------------+--------
BTW, I accidentally changed `tasks/toolchains/visualcpp.rake` at #4613,
so I put it back.
This time, the allocated memory comes from the string object, which is
referenced from GC arena. The memory region will be reclaimed when the C
function called from VM is terminated, or the GC arena is restored.
[Breaking changes]
Developers must use following APIs for accessing attributes of RRange
because RRange structure depends on boxing setting.
- mrb_range_beg
- mrb_range_end
- mrb_range_excl_p
The following code mistakenly exits from the function without marking
the env object as `MRB_TT_FREE`.
``` ruby
if (MRB_ENV_STACK_SHARED_P(e)) {
/* cannot be freed */
return; // <- should be `break`
}
```
When object allocation rate during incremental GC is too high, the
`gc->majorgc_old_threshold` becomes too big. This means major GC
start slower and old objects stay longer (and consume too much memory).
Instead of `irep` links, we added a `upper` link to `struct RProc`.
To make a space for the `upper` link, we moved `target_class` reference.
If a `Proc` does not have `env`, `target_class` is saved in an `union`
shared with `env` (if a `Proc` has env, you can tell it by `MRB_PROC_ENV_P()).
Otherwise `target_class` is referenced from `env->c`. We removed links
in `env` as well.
This change removes 2 members from `mrb_irep` struct, thus saving 2
words per method/proc/block. This also fixes potential memory leaks
due to the circular references caused by a link from `mrb_irep`.
Instead of `irep -> proc` chain, we use `irep -> irep` chain to
avoid GC bugs like #3804. We added `target_class` reference to
`mrb_irep` struct. That means one more word consumption per `irep`.
It reduces the memory consumption and sometimes improve the
performance as well. For example, the consumed memory size
of `bench/bm_ao_render.rb` is reduced from 1.2GB to 1GB, and
its total execution time become 18.795 sec from 22.229 sec.
KOBAYASHI Shuji