Re: [PATCH] Fix -Wundef warnins for __FP_FAST_FMA*

[Roland McGrath <roland at hack dot frob dot com>]

Indeed installed headers and compiler predefines are importantly different
cases than most of what we want to clean up in this regard.  Still, I think
we should be aiming for the same goal: minimize the possibility of typo
bugs not detected by -Wundef.

The context of installed headers just means that we have to take care with
the name space issues and the wide variety of older (and non-GNU) compilers
and their option combinations that might be in use, and beware of how users
might mistakenly misuse any internal macros we expose.

The context of predefines usually means that the base thing we're looking
at provided by the compiler has to be tested with #ifdef.  (For any future
predefines, we should work with compiler folks to suggest that new things
follow the #if-friendly pattern of macros always defined to either 0 or 1
rather than the typo-prone pattern of sometimes-defined macros.)  But that
doesn't mean there isn't cleanup and typo-proofing we should be doing.  We
should make sure that each predefine is tested with #ifdef in exactly one
place.  That place can then define internal macros to 0 or 1 so they can be
used elsewhere in our headers with #if and -Wundef.  The one place might be
simply at the top of an installed header (if the need to test it is solely
within that file or others that already must #include it).  It might be in
a new topical helper header (e.g. bits/foobar-options.h) if that seems to
make the most sense.  It might be in some new consolidated place for
handling many such macros (e.g. bits/predefines.h), or perhaps directly in
features.h).  I think we'll want to see all the variety of uses of such
macros in our installed headers before we decide.  (That is, audit the
macro use more generally, not just the things popping out with -Wundef.)

In short, these are the difficult cases.  I'd kind of figured we'd knock
out all the cases in private code before tackling the nontrivial installed
header cases.


Thanks,
Roland