[Tom Lane <tgl@sss.pgh.pa.us>]

I wrote:
>>> A small problem with the "rejiggering" is that it now makes the wrong
>>> choice for K&R-style function definitions, causing them to be weirdly
>>> indented.  For our purposes, that's a non-problem so I'm not excited
>>> about trying to make it smart enough to recognize those.  We do have
>>> a couple of amazingly old and crufty K&R-style functions in src/port/,
>>> though, so probably we'd wish to fix those.

To be concrete about that: the existing behavior when trying to decide
whether "foo(" starts a function declaration or a function definition
is to scan to the matching right paren[1] and then, if the immediately
next character is ';' or ',', it's a declaration; else it's a definition.
This fails on decls with attributes, such as

extern void pg_re_throw(void) __attribute__((noreturn));

My patch changes the rule to be "scan until we see a ';' ',' or '{'
that's not within parens; then it's a definition if that character
is '{', otherwise a declaration".  So it gets the right answer
for the above, but the wrong answer for

int
isinf(x)
double		x;
{

It doesn't really seem practical to me to make the lookahead function
smart enough to tell the difference between attributes and K&R-style
parameter declarations.  What I'm thinking of doing to have an
upstreamable patch is to invent a new switch, perhaps '-kr'/'-nkr',
to indicate whether the user is more worried about K&R function
declarations than she is about function attributes.  The default
for this switch could be debated perhaps, but I'd just stick an
explicit -nkr into pgindent's invocation, so we wouldn't care too
much.  (This would also ensure that pgindent would fail if you
try to use a too-old copy of pg_bsd_indent...)

			regards, tom lane

[1] Um, well, actually the existing behavior is to scan to the
*first* right paren.  But upgrading it to count parens correctly
seems like an unobjectionable improvement in any case.