Various disagreements between boost::regex and the JS spec

[Alcaro]

I implemented a test suite for JS v3 regex syntax as applied to 8bit strings, applied it to the three std:: implementations, and found a lot of bugs. The same tests can also be run against Boost.

https://godbolt.org/z/Pq76Pe1cz (some of the comments only apply to the regex engine I wrote alongside the tests, feel free to ignore them)

List of behavioral differences

FAIL should compile [a-b-c-d]
FAIL should compile [a-b-c-d]
FAIL should compile []
FAIL should compile [^]
FAIL expected error a[A-]]c
FAIL expected error ]
FAIL expected error \c+
FAIL expected error [a-\w]
FAIL \Z :: Z :: ret=(fail) :: exp=Z
FAIL expected error \x0
FAIL [\W\D] :: a :: ret=(fail) :: exp=a
FAIL expected error \00
FAIL expected error \01
FAIL expected error [\01]
FAIL expected error [\00]
FAIL expected error [\0-\00]
FAIL expected error a++
FAIL expected error a{,5}
FAIL expected error {
FAIL expected error }
FAIL expected error a{
FAIL expected error a{1
FAIL expected error a{1,
FAIL (abc)?\1 ::  :: ret=(fail) :: exp=/(null)
FAIL (abc)?\1 :: abc :: ret=(fail) :: exp=/(null)
FAIL (?:(a)|b)\1c :: bc :: ret=(fail) :: exp=bc/(null)
FAIL (?:a|(b))\1c :: ac :: ret=(fail) :: exp=ac/(null)
FAIL (a?){0,2}\1b :: b :: ret=b/ :: exp=b/(null)
FAIL (a?){0,2}\1b :: ab :: ret=ab/ :: exp=(fail)
FAIL (a?){0,5} :: aa :: ret=aa/ :: exp=aa/a
FAIL (?:(a)|b)\1 :: b :: ret=(fail) :: exp=b/(null)
FAIL (?:(a)b|aa)\1 :: aaa :: ret=(fail) :: exp=aa/(null)
FAIL expected error (a)b\01c
FAIL ((a)|(b))+ :: ab :: ret=ab/b/a/b :: exp=ab/b/(null)/b
FAIL ((a)|(b))+ :: ba :: ret=ba/a/a/b :: exp=ba/a/a/(null)
FAIL ((a)|(b)){2} :: ab :: ret=ab/b/a/b :: exp=ab/b/(null)/b
FAIL ((a)|(b)){2} :: ba :: ret=ba/a/a/b :: exp=ba/a/a/(null)
FAIL ((a)\2|(b)\3){2} :: aabb :: ret=aabb/bb/a/b :: exp=aabb/bb/(null)/b
FAIL ((a)\2|(b)\3){2} :: bbaa :: ret=bbaa/aa/a/b :: exp=bbaa/aa/a/(null)
FAIL (?:(a)|(b)|(c)|(d))+ :: abcd :: ret=abcd/a/b/c/d :: exp=abcd/(null)/(null)/(null)/d
FAIL (?:(a)|(b)|(c)|(d))+c :: abc :: ret=abc/a/b/(null)/(null) :: exp=abc/(null)/b/(null)/(null)
FAIL (a?)* ::  :: ret=/ :: exp=/(null)
FAIL (a?){2,4}b\1c :: aabc :: ret=aabc/ :: exp=(fail)
FAIL ((b)?(be)?){0,5}bbee :: bbebbee :: ret=bbebbee//b/be :: exp=bbebbee/bbe/b/be
FAIL ^(ab|()|a|bc)*$ :: abc :: ret=abc// :: exp=abc/bc/(null)
FAIL (?!(?!(a))) :: a :: ret=/a :: exp=/(null)
FAIL expected error (?=a)+
FAIL (?!(a))\1 :: b :: ret=(fail) :: exp=/(null)
FAIL (?:(a)|\1b)+ :: aabbaab :: ret=aa/a :: exp=aabbaab/(null)
FAIL (?:(a)|\1b)+ :: baabbaab :: ret=(fail) :: exp=baabbaab/(null)
FAIL should compile (?=)a
FAIL should compile (?![])
FAIL should compile (?![])
FAIL should compile (?![^])
FAIL should compile (?![^])

I can't find a clear specification for what regex syntax Boost tries to implement (https://www.boost.org/doc/libs/1_87_0/libs/regex/doc/html/boost_regex/ref/syntax_option_type/syntax_option_type_perl.html says it's supposed to match JS, but I can't find which version, and https://www.boost.org/doc/libs/1_87_0/libs/regex/doc/html/boost_regex/syntax/perl_syntax.html lists several Perl-only features).

As such, I don't know which of the above divergences are bugs, and which are expected; feel free to consider this a meta-bug, and close it once the above have been filed as separate bugs as appropriate. (For example, ] is explicitly called out as legal in docs, the two [a-b-c-d] are the same bug, and [\W\D] is #241.)

[jzmaddock]

The intention for Boost is to follow Perl, but of course that is an ever randomly changing target, so I suspect we're some way behind, and in general I only tend to change things when bug reports come in :(