Re: [RFC PATCH] pre-process: add __VA_OPT__ support

[Al Viro <viro@zeniv.linux.org.uk>]

On Wed, Feb 25, 2026 at 12:14:12AM -0800, Eric Zhang wrote:

> For # __VA_OPT__(), a stringify flag on TOKEN_VA_OPT_START could
> signal substitute() to stringify or produce "" depending on whether
> varargs are empty.

IMO it's better to turn that into
	TOKEN_VA_OPT[<token-list>]
and
	TOKEN_QUOTED_VA_OPT[<token-list>]
with list hanging off the cannibalized token.  Interpreter (substitute())
can easily keep track of where it is.

FWIW, the way they patched __VA_OPT__ into 10.5.1 is unfortunate -
I can understand wanting to keep the changes localized, but it ends
up very convoluted ;-/  In part it's due to the way ## and # evaluation
order is left unspecified, but... ouch.

Basically, #__VA_OPT__(<token-list>) is treated the following way:
it's "" if va-opt is suppressed, otherwise we
	* do argument substitution in <token-list>
	* [unspecified, but everyone does that] process # and ##
in the token-list.
	* do *NOT* remove placemaker tokens
	* do *NOT* rescan
	* stringify the resulting token list, same way we would if
that token list had been passed as an argument and we were processing
#<that argument> (as per 6.10.5.2[3]).

## vs. __VA_OPT__ is similar; the tricky part is placemaker treatment.
For normal arguments it's either a non-empty list or a solitary placemaker;
here we might have placemakers with non-empty list.

#define F1(X, Y, ...) Y ## __VA_OPT__(X X) ## Y
#define F2(X, Y, ...) Y ## __VA_OPT__(X) ## Y
F1(,a,_)
F2(,a,_)

We get a ## placemaker placemaker ## a (i.e. a a) or a ## placemaker
## a (i.e. aa) respectively.  Approach without explicit ## tokens
at expansion time is easy to adapt to that - we just interpret the
translated token-list hanging off __VA_OPT__, then return to the rest
of the body; state is updated as usual.  Quoted __VA_OPT__ == run the
interpreter (starting from Normal) on the token-list, then feed that
to stringify() and use the result as if it came from quoted argument
(note that concatenation with previous token *is* possible -
L ## #__VA_OPT__(something) is not invalid).

Since __VA_OPT__ can't nest, it's easy to save the body->next into a
local variable, set body to body->va_opt_list and, after the main loop
check if that local variable is non-NULL.  In that case we just set body
to that, clear that variable and go back to the beginning of the loop.
Quoted ones get a recursive call...

NOTE: substitute() is the second hottest loop in the entire thing; only
tokenizer is hotter.  And gcc is too enthusiastic about the inlining
around that function, ending up with bad register spills, along with
a bunch of stalls.  Worse, decisions are sensitive to minor changes in
places textually far away, making it a real bitch to deal with.
Makes for fun reordering the commits in local queue... ;-/

> > Another problem is that having no __VA_ARGS__ in the body should
> > *not* be treated as "vararg is empty"
> 
> Missed in the test case, but it seems like it can work in the current
> version.

AFAICS you won't even try to expand it at expand_arguments(), so the
reference will remain NULL.