Bug #15497
Updated by nobu (Nobuyoshi Nakada) over 5 years ago
This seems to happen mostly for internal errors, as `raise` in Ruby code of course just uses the passed String's encoding for the message.
Example:
```ruby
name = "été"
p name.encoding
begin
Module.new.const_set(name, 1)
rescue => e
p e
p e.message.encoding
end
```
When run, it gives:
```
$ LANG=en_US.UTF-8 ruby c.rb
#<Encoding:UTF-8>
#<NameError: wrong constant name été>
#<Encoding:UTF-8>
$ LANG=C ruby c.rb
#<Encoding:UTF-8>
#<NameError: wrong constant name "\u00E9t\u00E9">
#<Encoding:US-ASCII>
```
Depending on the locale encoding, the encoding of the message changes!
This seems very unexpected, is inconvenient for testing (e.g., https://github.com/ruby/spec/commit/a6101a6e and any test checking exception messages with non-US-ASCII characters),
and does not represent what is in the source code (here it's clearly a valid UTF-8 String).
I think for such a case, the encoding of the constant name should be used, i.e., UTF-8.
Another way to see it is the message should be built like `"wrong constant name ".force_encoding('us-ascii') + constant_name`.
Indeed, if we do build the message manually like that it works as expected:
```ruby ```
name = "été"
begin
raise "wrong constant name ".force_encoding('US-ASCII') + name
rescue => e
p e
p e.message.encoding
end
```
gives
```
$ LANG=en_US.UTF-8 ruby c.rb
#<Encoding:UTF-8>
#<RuntimeError: wrong constant name été>
#<Encoding:UTF-8>
$ LANG=C ruby c.rb
#<Encoding:UTF-8>
#<RuntimeError: wrong constant name \u00E9t\u00E9>
#<Encoding:UTF-8>
```
Note that the message still looks different, but that's the effect of `Kernel#p`, because it does not know how to display UTF-8 characters in a US-ASCII terminal.
Nevertheless, both messages have the same bytes and encoding, which fixes all 3 problems mentioned above.
Setting `Encoding.default_internal` can workaround this but it's a bad workaround as this cannot work reliably in a multithreaded Ruby application,
affects many more things than just error messages, and the default behavior should be error messages with a deterministic encoding, just like `raise` in Ruby code.