SubstitutionString can't contain escape sequences \n, \t, etc. during replace

[digital-carver]

While trying to create documentation for SubstitutionString (#26497), I found that replace using SubstitutionStrings doesn't work if the substition string has a newline or other escape sequence in it.

julia> msg = "#Hello# from Julia";

julia> re = r"#(.+)# from (?<name>\w+)";

julia> subst = s"FROM: \g<name>\n MESSAGE: \1"
s"FROM: \\g<name>\\n MESSAGE: \\1"

julia> replace(msg, re => subst)
ERROR: Bad replacement string: FROM: \g<name>\n MESSAGE: \1
Stacktrace:
 [1] error at .\error.jl:33 [inlined]
 [2] replace_err at .\regex.jl:241 [inlined]
 [3] _replace(::Base.GenericIOBuffer{Array{UInt8,1}}, ::Base.SubstitutionString{String}, ::String, ::UnitRange{Int64}, ::Regex) at .\regex.jl:302
 [4] #replace#350(::Int64, ::Function, ::String, ::Pair{Regex,Base.SubstitutionString{String}}) at .\strings\util.jl:427
 [5] replace(::String, ::Pair{Regex,Base.SubstitutionString{String}}) at .\strings\util.jl:415
 [6] top-level scope

julia> subst = s"FROM: \g<name>; MESSAGE: \1"
s"FROM: \\g<name>; MESSAGE: \\1"

julia> replace(msg, re => subst) #works fine when \n is removed
"FROM: Julia; MESSAGE: Hello"

Looking at the code, when the _replace function (in regex.jl) sees a backslash, it expects another backslash, a digit or a 'g', and throws an error with replace_err if it doesn't see one of those. Another consequence of this is that the replacement string can't contain things like Unicode codepoint sequences (\u0B85).

[nalimilan]

The problem is that \n isn't interpreted as in other strings, it's just stored as is:

julia> s"\n".string == "\\n"
true

The same thing applies to regexes, but it's less visible since PCRE interprets escapes like \n. Not sure what we should do.

[digital-carver]

How about calling unescape_string on the string first, with a special additional argument to keep capture-related escape sequences as it is (i.e. the \N and \g<> sequences that _replace handles itself). Something like:

--- a/base/strings/io.jl
+++ b/base/strings/io.jl
@@ -316,19 +316,21 @@ end
 General unescaping of traditional C and Unicode escape sequences. Reverse of
 [`escape_string`](@ref).
 """
-unescape_string(s::AbstractString) = sprint(unescape_string, s, sizehint=lastindex(s))
+unescape_string(s::AbstractString, keep_esc::AbstractArray{<:AbstractChar}=Char[]) = sprint(unescape_string, s, keep_esc; sizehint=lastindex(s))
 
 """
     unescape_string(io, str::AbstractString) -> Nothing
 
 Unescapes sequences and prints result to `io`. See also [`escape_string`](@ref).
 """
-function unescape_string(io, s::AbstractString)
+function unescape_string(io, s::AbstractString, keep_esc::AbstractArray{<:AbstractChar}=Char[])
     a = Iterators.Stateful(s)
     for c in a
         if !isempty(a) && c == '\\'
             c = popfirst!(a)
-            if c == 'x' || c == 'u' || c == 'U'
+            if c in keep_esc
+                print(io, '\\', c)
+            elseif c == 'x' || c == 'u' || c == 'U'
                 n = k = 0
                 m = c == 'x' ? 2 :
                     c == 'u' ? 4 : 8
--- a/base/regex.jl
+++ b/base/regex.jl
@@ -254,7 +254,8 @@ function _replace(io, repl_s::SubstitutionString, str, r, re)
     GROUP_CHAR = 'g'
     LBRACKET = '<'
     RBRACKET = '>'
-    repl = repl_s.string
+    keep_esc = [SUB_CHAR, GROUP_CHAR, collect('0':'9')...]
+    repl = unescape_string(repl_s.string, keep_esc)
     i = firstindex(repl)
     e = lastindex(repl)
     while i <= e

There's a change in behaviour due to this though: invalid single letter sequences like \m, \j, etc. are silently replaced by their corresponding character (m, j, ...) by unescape_string, whereas currently _replace emits the "Bad replacement string" error for those. Perhaps unescape_string needs another optional argument to enable a strict mode, or... is there a reason unescape_string is currently not strict about invalid C sequences, while it errors out on invalid Unicode sequences?

[StefanKarpinski]

Spot on: a custom unescaping pass is what is needed to make this work correctly.

[digital-carver]

Note for future: even after other escape sequences are allowed in a SubstitutionString, octal sequences still can't be used since \73 refers to the 73rd capture group and not to ';'. Should be documented.

Related, if the user wants to use the second capture group and then have "45" in the output string, there doesn't currently seem to be a way to do that: \245 will try to get the 245th capture group. Perl solves this by allowing \g{2}45 in the substitution string - allowing the \g syntax to be used with capture group numbers also, in addition to with names. Similarly, Julia should probably allow a \g<2>45 syntax for referring to unnamed capture groups by ordinal number.

To avoid ambiguity then, named capture groups should disallow numerical names, like r"(?g<2>\W+)" (which Perl also does).

[pocc]

Noting that a way to get around this for \n is to manually entering a return like so:

julia> print(replace("beautiful_string", r"_"  => s"
       "))
beautiful
string