This will only work if the filesystem encoding happens to be utf-8, though, right? (That's usually the case, but not always). I think instead this needs to use sys.getfilesystemencoding().

Three variables must be considered:

• Regular expressions compiler
• output of subprocess, which is processed with the reqular expressions compiler. By default returns bytes
• fontexts synonyms, which are plain strings at the moment.

Things to consider:

• The reqular expressions compiler and the output of subprocess encoding must match, which is the case, when using a bytes reqular expression or decoding the bytes object to a string. utf-8 is only the default for bytes.decode and could also be omited.
• Even if the system file encoding will be differently, subprocess returns byes, which can be decoded to whatever you like. So utf-8, what is the default for bytes.decode anyway looks fine.

The only reason for sys.getfilesystemencoding() would be, that matplotlib uses the same encoding internally than the file system. But even there, I think it would be more sane to directly use utf-8 instead and would favour that in any case...

I think you've raised a valid issue with the fix in #1767, however I think this still is not correct.

If the file system encoding were, for example latin-1, it's possible, and actually quite easy, to have filenames that are not valid utf-8. (Note: I'm not saying not valid Unicode -- utf-8 is only an encoding, and not all bytestrings are valid utf-8). I looked deeper into fc-match -- the manpage doesn't really address encodings, but it does look from the code like the output is basically utf-8, though the filename part is passed in directly through, so it can be in the filesystem encoding. So I think the correct fix is:

--- a/lib/matplotlib/font_manager.py
+++ b/lib/matplotlib/font_manager.py
@@ -1289,11 +1289,12 @@ if USE_FONTCONFIG and sys.platform != 'win32':
         if pipe.returncode == 0:
             for match in _fc_match_regex.finditer(output):
                 file = match.group(1)
+                file = file.decode(sys.getfilesystemencoding())
                 if os.path.splitext(file)[1][1:] in fontexts:
                     return file
         return None

-    _fc_match_regex = re.compile(r'\sfile:\s+"([^"]*)"')
+    _fc_match_regex = re.compile(rb'\sfile:\s+"([^"]*)"')
     _fc_match_cache = {}

     def findfont(prop, fontext='ttf'):

I've marked this a blocker for v1.3 as it sounds pretty nasty. @mdboom & @tomspur - do we have a consensus on where this PR should go?

@pelson I think @mdboom's appoach is more reasonable. Unfortunately, I currently have other problems with my former reproducer from above (unrelated to matplotlib) and cannot verify his solution currently...