Instead we added `%define parse.error verbose`.

The newer `clang` warns implicit float conversions.

Get an irep symbol if it's `OP_GETSV` or `OP_SETSV`

Update `doc/opcode.md` [ci skip]

The difference of `include/mruby/ops.h` is applied.

- OP_NOP - update semantics
- OP_GETSV - update semantics
- OP_SETSV - update semantics
- OP_GETUPVAR - update prefix
- OP_SETUPVAR - update prefix
- OP_JMPIF - update operands and semantics
- OP_JMPNOT - update operands and semantics
- OP_JMPNIL - add entry
- OP_ONERR - update semantics
- OP_POPERR - update prefix
- OP_EPOP - update prefix
- OP_SENDB - update semantics
- OP_ADD - update prefix and operands
- OP_ADDI - update operands and semantics
- OP_SUB - update prefix and operands
- OP_SUBI - update semantics
- OP_MUL - update prefix and operands
- OP_DIV - update prefix and operands
- OP_EQ - update prefix and operands
- OP_LT - update prefix and operands
- OP_LE - update prefix and operands
- OP_GT - update prefix and operands
- OP_GE - update prefix and operands
- OP_ARYDUP - add entry
- OP_INTERN - add entry
- OP_HASHCAT - add entry
- OP_ERR - update semantics

Remove byteorder constants; ref 87576b81

Fix boundary check for `OP_LOADI16`; ref fa8668c7

The `FLAG_BYTEORDER_NATIVE` and `FLAG_BYTEORDER_NONATIVE` are no longer
needed.

It was making a negative integer if the highest-order bit of a 16-bit
integer was 1.

no patched:

```ruby
p 0x7fff  # => 32767
p 0x8000  # => -32768
p 0xffff  # => -1
p 0x10000 # => 65536
```

Fix some `Hash` methods are inconsistent with `values`

Inconsistent when hash has duplicate key.

### Example

  ```ruby
  # example.rb
  keys = (1..3).map{[_1]}
  h = keys.to_h{[_1, _1[0]]}
  keys[0][0] = 2
  p h.values
  p h.each_value.to_a
  p h
  ```

#### Before this patch:

  ```console
  $ bin/mruby example.rb
  [1, 2, 3]
  [1, 1, 3]
  {[2]=>1, [2]=>1, [3]=>3}
  ```

#### After this patch (same as Ruby)

  ```console
  $ bin/mruby example.rb
  [1, 2, 3]
  [1, 2, 3]
  {[2]=>1, [2]=>2, [3]=>3}
  ```

`cmpnum` function may return `nil` on error.

Since `mruby 2.0`, compiled bytecode no longer depends on the
endian of the machine.

You don't have to specify endian since `mruby 2.0`.

- Add `OP_LOADI16`
- Update keyword argument related ops (e.g. OP_KEY_P, etc.)
- Fix some wrong descriptions

Which loads 16bit integer to the register. The instruction number should
be reorder on massive instruction refactoring. The instruction is added
for `mruby/c` which had performance issue with `OP_EXT`. With this
instruction, `mruby/c` VM can just raise errors on `OP_EXT` extension
instructions.

`mrb_run` requires to push callinfo stack before calling, which is very
hard from outside of `vm.c`. So there should be virtually no correct
usage of the function, hence the cause of #4488. We removed it.

You can use `mrb_top_run(mrb, proc, self, 0)` instead of
`mrb_run(mrb, proc self)`.

Enable MRB_METHOD_T_STRUCT by default on 32bit GUN/Linux

Because we can't use the highest 2 bits of function pointers.

Avoid changing directory in `mruby-io` test

Refine `.travis.yml`

* Disable automatic update and clean up on `brew install` (install time
  160 sec -> 5 sec).
* Avoid using deprecated keys.

Sorting the list of gems at build time by name

Avoid comparing pointers and integers

Makes `mrb_any_to_s()` accept an object whose class is `NULL`

When using `mrb_any_to_s()` for debugging purposes, giving an object
whose class is `NULL` no longer causes a SIGSEGV and no crash.
This is achieved by making `mrb_class_name()` and `mrb_str_cat_cstr()`
null safe.

Printing them in sorted order makes it easier to find the desired gem.
But it has come to completely ignore the dependency.

`assert()` can be completely removed when `NDEBUG` is set.