Misc #20434
openDeprecate encoding-related regular expression modifiers
Description
This is a follow-up to @duerst's comment here: https://bugs.ruby-lang.org/issues/20406#note-6.
As noted in the other issue, there are many encodings that factor in to how a regular expression operates. This includes:
- The encoding of the file
- The encoding of the string parts within the regular expression
- The regular expression encoding modifiers
- The encoding of the string being matched
At the time the modifiers were introduced, I believe the modifiers may have been the only (??) encoding that factored in here. At this point, however, they can lead to quite a bit of confusion, as noted in the other ticket.
I would like to propose to deprecate the regular expression encoding modifiers. Instead, we could suggest in a warning to instead create a regular expression with an encoded string. For example, when we find:
/\x81\x40/s
we would instead suggest:
::Regexp.new(::String.new("\x81\x40", encoding: "Windows-31J"))
or equivalent. As a migration path, we could do the following:
- Emit a warning to change to the suggested expression
- Change the compiler to compile to the suggested expression when those flags are found
- Remove support for the flags
Step 2 may be unnecessary depending on how long of a timeline we would like to provide. To be clear, I'm not advocating for any particular timeline, and would be fine with this being multiple years/versions to give plenty of time for people to migrate. But I do think this would be a good change to eliminate confusion about the interaction between the four different encodings at play.
Updated by shyouhei (Shyouhei Urabe) 7 months ago
- Subject changed from Deprecate regular expression modifiers to Deprecate encoding-releated regular expression modifiers
+1 for deprecating encoding modifiers, but they're not everything that a regexp can take. For instance /foo/i
is a valid regular expression literal in ruby, perl, PHP(preg), and Javascript.
I'm sure Kevin didn't intend to kill everything. Let me narrow the scope of this request; subject updated.
Updated by duerst (Martin Dürst) 7 months ago · Edited
I guess there might still be some use for the encoding-related modifiers in single-line scripts and the like. But I don't have an actual use case; I hope whoever has such an use case comes forward.
The replacement code (::Regexp.new(::String.new("\x81\x40", encoding: "Windows-31J"))
) is quite lengthy. This makes it clear that while each regular expression has an encodings in the same way as each String has an encoding, regular expressions don't really allow to manipulate the encoding. Strings have #force_encoding and #encode, so maybe adding one or both methods to Regexp would help. The example could then be written as /\x81\x40/.force_encoding("Windows-31J")
or /\3000/.encode("Windows-31J")
.
Updated by byroot (Jean Boussier) 7 months ago
/\x81\x40/.force_encoding("Windows-31J")
wouldn't work because String#force_encoding
mutates the string, and Regexp literals are immutable.
Similarly String#encode
doesn't just change the string encoding attribute, but convert the bytes to the new encoding. So I'd expect /\3000/.encode("Windows-31J")
to fail with:
\x81" on UTF-8 (Encoding::InvalidByteSequenceError)
So I think the String API to mirror would be String.new(encoding:)
Regexp.new(/\x81\x40/, encoding: Encoding::WINDOWS_31J)
Regexp.new("\x81\x40", encoding: Encoding::WINDOWS_31J)
But if we want an instance method, I think something like:
/\x81\x40/.encoded(Encoding::WINDOWS_31J)
, which by the way would also be useful on String
, e.g., this is common:
# frozen_string_literal: true
THING = "fée".dup.force_encoding(Encoding::ISO8859_1)
So it could become:
# frozen_string_literal: true
THING = "fée".encoded(Encoding::ISO8859_1)
Updated by Eregon (Benoit Daloze) 7 months ago
- Subject changed from Deprecate encoding-releated regular expression modifiers to Deprecate encoding-related regular expression modifiers
Updated by Eregon (Benoit Daloze) 7 months ago
This seems a good simplification to me, I think the semantics of these encoding modifiers are confusing to most Rubyists.
I wouldn't be worried too much about length of the replacement, because /.../s
//.../e
are likely very rare (using file encoding seems a good replacement for those).
/.../u
seems redundant with the default source encoding, so the u
can likely just be removed in most cases.
I'm not so sure /.../n
, that may be more frequent.
Methods to convert an existing Regexp from one encoding to another feel suboptiomal, because that will cause an extra Regexp instance and creating a Regexp is not cheap due to many checks, allocations, and even compilation (AFAIK eager in CRuby at least).
So I think the existing Regexp.new("a".dup.force_encoding(Encoding::WINDOWS_31J))
is good enough.
And since this would address a deprecation, it seems very important that the code also works on older Ruby versions.
(I'm all for String#{encoded,with_encoding}
but it seems best to propose that as a separate ticket)
I would be interested to have a good textual description in #20406 of how the encoding of a Regexp is computed currently, it seems quite complex, but having it in text would allow to reason more easily about it.
Maybe we could simplify it while remaining compatible (i.e. the specific value of Regexp#encoding matters not so much, what matters is a Regexp can still be matched against Strings of various encoding like it could before).
Updated by Eregon (Benoit Daloze) 7 months ago
- Related to Misc #20406: Question about Regexp encoding negotiation added
Updated by kddnewton (Kevin Newton) 7 months ago
Thanks @shyouhei (Shyouhei Urabe) — I definitely only meant encoding-relate modifiers. I really like the other ones!