The way how _bytesify_input handles errors and warnings is too inflexible:

• For bytes_mode=None (default), the function emits a bytes warning when the input is text, then encodes it as UTF-8 bytes. Warnings are slow and can cause performance regressions (see Possible performance regression with LDAPBytesWarning #133). The warnings are always emitted, even when warnings are disabled and not shown to the user. The warnings module does not allow early opt-out.
• For bytes_mode=False (strictly future-compatible) there is no warnings mode at all. When the function is called with bytes as input, it always fails with TypeError. TypeErrors make it really, really hard to port an existing application to future compatible API. There should be a mode that emits just warnings.

Proposal

I suggest that we introduce a bytes_warning flag:

• bytes_warning=BYTES_WARNINGS_SILENT emits no warnings and doesn't raise TypeError for bytes_mode None, True, and False.
• bytes_warning=BYTES_WARNINGS_WARN emits a warning when input is text for bytes_mode=None / bytes_mode=True or input is bytes for bytes_mode=False.
• bytes_warning=BYTES_WARNINGS_ERROR raises a TypeError when input is text for bytes_mode=None / bytes_mode=True or input is bytes for bytes_mode=False.
• bytes_warning=None (default) depends on bytes_mode:
  • bytes_mode=None -> BYTES_WARNINGS_WARN
  • bytes_mode=True -> BYTES_WARNINGS_ERROR
  • bytes_mode=False -> BYTES_WARNINGS_ERROR

The flags are defined as:

BYTES_WARNINGS_SILENT = 0
BYTES_WARNINGS_WARN = 1
BYTES_WARNINGS_ERROR = 2