| Age | Commit message (Collapse) | Author |
|---|
|
#### 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.
|
|
* The allocated memory is guaranteed to be aligned for any data type (it was
not guaranteed when string type is embed).
* Make allocation size exactly specified size (does not allocate space for a
null byte).
|
|
|
|
For efficiency with `MRB_WORD_BOXING` (implement type predicate macros for
all `enum mrb_vtype`).
|
|
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.
|
|
Because immediate values are not registered.
|
|
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.
|
|
|
|
|
|
Fix leaked function symbols
|
|
- `free_heap()` in src/gc.c
- `symhash()` in src/symbol.c
- `no_optimize()` in mrbgems/mruby-compiler/core/codegen.c
|
|
|
|
[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
|
|
By dffa203 that reclaim `env` objects from heaps, there's more chance
for `env` objects referenced from fibers may be freed from heap pages.
|
|
The old condition marks the top-level callinfo even after the
fiber is terminated.
|
|
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`
}
```
|
|
|
|
|
|
This means reducing one word per a call frame.
|
|
Renamed flag macro names as well:
`MRB_FLAG_IS_FROZEN` -> `MRB_FL_OBJ_FROZEN`
`MRB_FLAG_IS_PREPENDED` -> `MRB_FL_CLASS_IS_PREPENDED`
`MRB_FLAG_IS_ORIGIN` -> `MRB_FL_CLASS_IS_ORIGIN`
`MRB_FLAG_IS_INHERITED` -> `MRB_FL_CLASS_IS_INHERITED`
|
|
If MRB_TT_ENV has stack shared flag or bidx flag, flags is too large
than the real stack size.
|
|
|
|
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).
|
|
Because 200% is consumes too much memory for some cases.
|
|
Because unlike other GC ratio constants, it's not a default value.
|
|
|
|
Memory leak fixed; ref #3711
|
|
|
|
|
|
|
|
The issue #3837 is a regression of #3819. I was mistakenly removed
a important code to mark fibers from `env`.
|
|
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`.
|
|
|
|
|
|
`RSTR_FSHARED` use frozen strings as shared body instead of
`struct mrb_shared_string`. This reduces allocation from
literal strings.
|
|
|
|
possible loss of data
|
|
possible loss of data
|
|
possible loss of data
|
|
|
|
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`.
|
|
|
|
This reverts commit 15d48efa4bf6a6b999dab99720fb9fe3c4a06ef0.
|
|
|
|
Otherwise dead object may be seen from `#each_object`.
|
|
|
|
|
|
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.
|
