[Bug middle-end/99339] Poor codegen with simple varargs

["redbeard0531 at gmail dot com" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99339

--- Comment #6 from Mathias Stearn <redbeard0531 at gmail dot com> ---
> The question is how common in the wild it is and if it is worth the work.

I'm not sure how common it is, but this is my use case. The code in the bug
report is a slightly simplified example from some Real World Code I am working
on:
https://source.wiredtiger.com/develop/struct_w_t___c_u_r_s_o_r.html#af19f6f9d9c7fc248ab38879032620b2f.
Essentially WT_CURSOR is a structure of function pointers that all take a
WT_CURSOR pointer as the first argument, similar to C++ virtual functions. The
get_key() and get_value() "methods" both take (WT_CURSOR*, ...) arguments, and
unpack the arguments based on the format string that was set up when you opened
the cursor. The expectation is that they will be called many times for each
cursor object. Since we know the format string when creating the cursor I was
experimenting with creating specialized functions for common formats rather
than dispatching through the generic format-inspecting logic every time. I
noticed that I was able to get even more performance by declaring the functions
as taking the arguments they actually take rather than dealing with va_args,
then casting the function pointers to the generic (WT_CURSOR, ...) type when
assigning into the method slot. I assume this is UB in C (I don't know the C
standard nearly as well as C++) but all ABIs I know of are designed to make
this kind of thing work.

I'd rather not have to do that kind of questionable shenanigans in order to get
the same performance.