https://gcc.gnu.org/bugzilla/show_bug.cgi?id=47781
Gabriel Dougherty <contact at gabrieldougherty dot com> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |contact at gabrieldougherty
dot co
| |m
--- Comment #32 from Gabriel Dougherty <contact at gabrieldougherty dot com> ---
Hi, I'm not sure if this is out of my league, but I am at the very least
interested in working on this.
My basic idea is to have an attribute like the following:
#include <printf.h>
struct My {};
struct MyL {};
struct MyLL {};
__attribute__((printf_format_specifier(
"M",
/* pointer_count */ 1,
/* flag_chars */ "-+0 #wp",
/* flags2 */ "iWR[]<>",
/* length token pairs */
"none", __typeof__(struct My *),
"l", __typeof__(struct MyL *),
"ll", __typeof__(struct MyLL *)
)))
static int
custom_arginfo_M (const struct printf_info *info, size_t n,
int *argtypes, int *size)
{
/* ... */
}
/* ... */
register_printf_specifier ('M', custom_printf_M, custom_arginfo_M);
This is a direct conversion from GCC internals, with the intention to support
all mainstream printf single-character specifier cases, while leaving out
things like Solaris-style multi-argument specs and the ability to extend
existing specifiers. I also do not allow per-spec STD_* setting (STD_C89,
STD_C99 etc), they stay at STD_EXT. So this proposal does not get full parity
with all that format_char_info offers; it would just cover what I imagine to be
the majority of cases based on reading uses in the wild.
pointer_count (meaning the level of pointer indirection, with 0 meaning no
indirection, 1 meaning one level, *p, and so on) applies uniformly to all
length entries (%l, %ll, ..), there is not a per-length override.
flag_chars and flags2 define what printf flags are supported by this new
specifier. Both are validated against printf’s flag_specs; any character not
defined there is rejected.
For the length token pairs, "none" is a special value meaning format_lengths
FMT_LEN_none signifying no length modifier (%M in this example). If "none" is
not provided, then bare %M in this case would be disallowed. The other options
are all matched against printf's format_kind_info.length_char_specs and
translated by printf_length_specs. If we get a truly unknown token, that
becomes a diagnostic.
The gcc format_spec attribute can be set on the custom specifier's print
function handler (custom_printf_M in this example) or on a dummy function if
desired, it just needs to be available in the Translation Unit where the format
checking is done by GCC.
Happy to contribute an implementation (in the form of a plugin to start) and
tests, just wanted to get alignment on the interface first.
Thanks
Gabriel
