ascii_only changes semantics of redundantly escaped regexes

[DavidBuchanan314]

$ npx terser -V
terser 5.29.2

Steps to repro:

$ cat testcase.js
console.log("hello\\world❗".split(/[\❗]/));
$ node testcase.js 
[ 'hello\\world', '' ]
$ npx terser testcase.js -f ascii_only=true | tee min.js
console.log("hello\\world\u2757".split(/[\\u2757]/));
$ node min.js
[ 'hello', 'world❗' ]

When it occurs inside a regex literal, \❗ is equivalent to just ❗ - the redundant escape is ignored by JS runtimes (as far as I can tell?).

But, once the emoji has been replaced with a unicode escape sequence (by terser), the \\ substring parses as an escaped backslash, changing the meaning of the regex.

Maybe there needs to be an initial pass to strip redundant backslashes?

(By the way, when \❗ occurs inside a string literal, it appears to be handled correctly)

[fabiosantoscode]

Cool bug!

The behavior with emojis that need two \uXXXX is also fun. The backslash of the first gets escaped, not the second.

"".split(/[\😁]/);

becomes

"".split(/\\ud83d\ude01/);