While looking at get_all_args, I noted that it doesn't check for too many arguments. In a few cases, like mktime, that's exploited to take care of some optional arguments afterwards. In the majority of cases however it just makes the functions too forgiving for argument errors (c.f. code generated from cmods that normally doesn't allow too many arguments).
I'm thinking of adding support for optional arguments to get_all_args and to let it check for too many arguments. That means that it would complain by default, which is incompatible (notwithstanding a provision for 7.4 compat mode, which of course should be added). Opinions?
/ Martin Stjernholm, Roxen IS
Previous text:
2004-02-28 21:18: Subject: Re: Problem with function prototypes
This is quite easy, why it doesn't - since before this check there is another (line 373):
if (ret <= args)
and, of course, value of "ret" is 0, value of "args" is 0 too, so... :)
Yes, that looks bogus. It should be < there.
But I wonder why this bug doesn't happen in 7.4 (same check order)...
The bug was that a freed svalue on the stack was copied into the exception struct if the function was called with no arguments. That only led to a failure if the stack svalue pointed to a refcounted value of some sort. So whether or not you get the bug depended on the code further up. It took me a while to figure out a test that consistently trigs it:
void trig_bug() { lambda (mixed a, mixed b, mixed c) {} (({time()}), ({time()}), ({time()})); // set_weak_flag uses get_all_args and requires two arguments. ([function] set_weak_flag)(); }
I guess there's some kind of subtle difference in the stack handling between 7.4 and 7.5 that caused it to happen only in 7.5 in your case.
Also, compiling pike with --with-rtldebug and without --without-cdebug (or without stripping) helps quite a lot in tracking down bugs.
Perhaps, you mean --with-cdebug?
That should work too, but since it's the default I did actually mean without --without-cdebug.
No idea why gdb shows no args in backtrace...
The binary doesn't get stripped somewhere? When I hack Pike on the C level, I usually compile it with a plain "make" and run it directly in the build tree through the bin/pike script. It's not stripped then, at least.
Anyway, the backtrace is strange (perhaps because of longjmp() - there is no trace of get_all_args()).
That was because the use of the bogus svalue caused an error first when the exception is handled.
/ Martin Stjernholm, Roxen IS