- There was some unnecessary complexity in `OP_BREAK` introduced in commit ad2e626 (#6282).
- Since `mrb->c` is never NULL, there is no need to check it with `MRB_ENV_ONSTACK_P()` beforehand.
The part removed in this patch was introduced by commit c7c9543bed.
The current mechanism should be able to trace from block objects created by `eval` to higher level blocks without any problems.
There are two issues to be fixed:
- `mrb_irep` could leak if `mrb_calloc()` encountered an out-of-memory exception
- `mrb_proc_merge_lvar()` allocated one extra variable name.
`irep->lv` can always refer to only one less range than `irep->nlocals`.
Also, when `mrb_proc_merge_lvar()` extends `irep->lv`, `mrb_realloc()` with `NULL` has the same behavior as `mrb_malloc()`.
Set `env->cxt` to `NULL` when it is detached from the call frame.
In other words, we can determine if `env->cxt` is `NULL` or not.
Also, `mruby-binding` had been setting `env->cxt` unnecessarily, so this has been fixed.
We need to include stdlib.h and malloc.h to use malloc()/free() but
they aren't included in src/string.c with WIN32_LEAN_AND_MEAN. It
generates build time warnings.
We can solve this by including stdlib.h and malloc.h explicitly.
In the following example, the fiber context and call stack may be in an incomplete state.
- In case another thread running mruby is terminated abnormally
- In case of a global jump that is out of management by mruby
Only go to exception handling if `mrb->exc` is non-null.
This may cause some compatibility problems, but I doubt that it is necessary to maintain that compatibility.
Here is how I see the incompatibility with the change at this time:
- If `mrb->exc` is non-null and `mrb_vm_exec()` is called, an exception will be thrown immediately.
- If `MRB_THROW()` is used while `mrb->exc` is `NULL`, it will not go to exception handling.
Previous SWAR version assumes valid UTF-8 to count number of code points
in the string, but we need to handle invalid sequence as well. We now
use `search_nonascii` to skip counting single byte characters for
performance. The new version is even faster than SWAR version (probably
because `search_nonascii` uses SSE2 on Intel compatible CPU (which I use).
The patch assumes that `struct REnv::cxt` only performs checks with the `OP_BREAK` and `OP_RETURN_BLK` instructions, and does not reference the entity.
Therefore, by changing to a weak reference, it is possible to collect fibers that are no longer directly referenced while in the suspended state.
However, we need to detach the living env objects that remain in the call stack of the fiber.
So, in effect, it involves a revert of following commits.
- commit a3365d8b3f
- commit 57ffa1c150
Examples of the effects of change are shown below.
Note that it was built with `rake MRUBY_CONFIG=host-debug`.
```ruby
f = Fiber.new { (x, y, z) = "X", "Y", "Z"; Fiber.yield -> { [x, y, z] } }
g = f.resume
GC.start
p ObjectSpace.memsize_of_all
# => 59532
g.call
# => ["X", "Y", "Z"]
f = nil
GC.start
ObjectSpace.memsize_of_all
# BEFORE => 59532
# AFTER => 58044
g.call
# => ["X", "Y", "Z"]
```
`mrb_env_unshare()` calls `mrb_realloc_simple()` and follows `mrb_full_gc()` to avoid an infinite loop where `mrb_env_unshare()` is called again.
This does not occur at this time, but may occur in subsequent patches.
`mrb_vm_run()` is,
- It does not change the fiber context.
- When control is returned, only one ci prepared by the caller is popped.
If the ci equals cibase when called, the ci position does not change.
related commits:
- commit 4e84bdb507
- commit 34dd258c63
- commit ebd6636a1e
- commit c6736357a7
- commit 23a4e7149d
- commit 31a961acf1
- Don't create multiple envs on one ci.
- Don't share a env to different ci.
- Don't attach a closed env to any ci.
Changes in `envadjust()` can be simplified with those guarantees.
dearblue