Clearing errors at the beginning of `mrb_vm_exec()` essentially keeps the mruby VM in a non-error state.
For consistency, functions such as `mrb_funcall()` check for errors when control returns from a C function as a method.
In the case of a tail call, it should return to `mrb_vm_exec()` afterwards, so error checking is performed there.
Instructions issued while `mrb->exc` is non-null should be limited to `OP_EXCEPT`, the jump target of the catch handler table.
`string[]=(idx, replace)` should return `replace`.
## Actual (wrong)
```
string.[]=(idx, replace) → string
string.[]=(idx, len, replace) → string
```
## Expected
```
string.[]=(idx, replace) → replace
string.[]=(idx, len, replace) → replace
```
## Sidenote
As of the current mruby-compiler, `(string[idx] = 'X')` creates not only "CALL_NODE" but also "ASGN_NODE" and "OP_MOVE", overriding the wrong return value.
On the other hand, `string.[]=(idx, 'X')` creates only "CALL_NODE", exposing the wrong return value.
If my new mruby-compiler2, leveraging Prism, took the place of official compiler, `(string[idx] = 'X')` and `string.[]=(idx, 'X')` would be going to generate the same VM code without "OP_MOVE".
So I paranoidly added tests.
FYI: You can find how the new mruby-compiler2's AST and VM code look like in mruby/c's issue (mruby/c had the same bug): https://github.com/mrubyc/mrubyc/pull/210
This reverts commit ad2e626e7a.
Because of the changes made by #6282, the following code caused a problem.
```ruby
b = proc { break "BAD!" }
p self.tap { b.call }
# (expected) => break from proc-closure (LocalJumpError)
# (after #6282) => "BAD!"
```
I revived the `mrb_callinfo::blk` field to fix this, but it did not overcome the following problem.
```ruby
def m(&b); b = b.clone; GC.start; b.call; end
p m { break "OK!" }
# (expected) => "OK!"
# (revived blk) => break from proc-closure (LocalJumpError)
```
By adding a fast-path where we ignore boxed types we can gain a pretty substantial speedup of mrb_iv_get, making it about 25% faster during a standard optcarrot benchmark run.
NOTE: It is just mrb_iv_get that is that much faster, the whole benchmark seems to be about 3-5% faster with word boxing.
This allows the compiler to optimise the case in obj_iv_p into a range check. There does not seem to be any other very hot uses of this index and it grants a pretty big gain on optcarrot.
- use heap sort (O(1)) instead of merge sort (O(n)) for better space
complexity.
- method implemented in C for better performance
As a result, simple sorting now consumes far less memory and is faster.
Since it's implemented in C, fiber context switching is not allowed from
comparison, but we consider the risk is minimal (no one switches context
in the comparison, right?)
dearblue