[cfe-dev] proposal: every warning should have a -W flag

[Ted Kremenek]

On Aug 9, 2011, at 10:44 PM, David Blaikie wrote:

>> All of our diagnostics have unique tablegen entries, so getting unique names
>> for -W flags might not be that hard.  When one looks at the .td files, many
>> of the diagnostic names directly correspond to the -W flag.
> 
> Oh, of course. So would it be worth just making every diagnostic its
> own flag by definition, using the tablegen name for the -W flag?

I don't think so, or at least I wouldn't make it automatic.  Some diagnostics are just variants of basically the same thing, or logically part of a single warning.  For example, consider the different format string diagnostics:

def warn_printf_insufficient_data_args : Warning<
  "more '%%' conversions than data arguments">, InGroup<Format>;
def warn_format_invalid_conversion : Warning<
  "invalid conversion specifier '%0'">, InGroup<Format>;

and so on.

Is there real value in placing each of these under a separate flag?  In the general case, I don't think so, but where it makes sense it seems reasonable to break them down into separate -W flags (which we do in some cases for format string warnings).

> 
>> I was specifically proposing that…
>> (a) every warning is part of a group
>> and
>> (b) every group that contains a warning that can only be activated by that group maps to a -W flag
> 
> Does LLVM have the concept of groups that don't have an associated -W
> flag? what are they for?

That's a good point.  I don't think one can define a warning group without an associated flag.  Groups, however, can contain other groups, so when we report warnings we include the flag with the most precise coverage.