Summary

Modules/_sre/sre_lib.h is included three times, once per character width, so
charset_loc_ignore is compiled as three functions. Only the bytes one can ever
run: the IN_LOC_IGNORE opcode that calls it is emitted only for LOCALE
patterns, and LOCALE is rejected for str patterns.

Evidence

Coverage from a --coverage build of main running the full test suite (552 test
files, 52,070 tests, ./python -m test -j4):

Function 'sre_ucs4_charset_loc_ignore'   Lines executed:0.00% of 39
Function 'sre_ucs2_charset_loc_ignore'   Lines executed:0.00% of 39
Function 'sre_ucs1_charset_loc_ignore'   Lines executed:87.18% of 39

78 lines that no input reaches.

Why they cannot be reached

IN_LOC_IGNORE is emitted in one place, Lib/re/_compiler.py:89:

if flags & SRE_FLAG_IGNORECASE and flags & SRE_FLAG_LOCALE:
    emit(IN_LOC_IGNORE)

and the parser refuses LOCALE for str patterns, Lib/re/_parser.py:1106:

raise ValueError("cannot use LOCALE flag with a str pattern")

The opcode is fixed when the pattern is compiled, so adding the flag later does
not help. Constructing a pattern directly through _sre.compile() with the
LOCALE flag set keeps whatever opcode the code array already carries, which for
a str pattern is IN_UNI_IGNORE, not IN_LOC_IGNORE:

import re, _sre
from re import _parser, _compiler
p = _parser.parse('[a-z]', re.IGNORECASE)
code = _compiler._code(p, re.IGNORECASE)
groupindex = p.state.groupdict
indexgroup = [None] * p.state.groups
for k, i in groupindex.items():
    indexgroup[i] = k
pat = _sre.compile('[a-z]', re.IGNORECASE | re.LOCALE, code,
                   p.state.groups - 1, groupindex, tuple(indexgroup))
pat.search('\U0001d11eA')   # matches, and still does not call the UCS4 variant

Matching a str against a bytes pattern is rejected, so there is no route from
any public or private API to the UCS2/UCS4 copies.

Question

Is the triplication intentional (the template is included and every function is
emitted regardless of reachability, and the duplicates are simply dropped by the
linker or the optimiser), or would it be worth guarding the emission so only the
bytes variant exists?

I have not prepared a patch: this is a question about intent first. If the answer
is that they should not exist, the fix is to wrap the function definition in
sre_lib.h so it is only generated for the bytes build, and the coverage build
then confirms it.

CPython versions tested on

CPython main branch (3.16.0a1+)

Operating systems tested on

Linux