ubsan complains on rpcgen-generated code, gcc's one silent

[stsp]

It appears, clang's ubsan complains on
an rpcgen-generated code, while gcc's
ubsan is silent.
The code roughly looks like this (reduced test-case):

#include <stddef.h>

struct svc_req;

static int
_a_1 (void *argp, void *result, struct svc_req *rqstp)
{
        return 0;
}

int main()
{
        int (*local)(char *, void *, struct svc_req *);
        local = (int (*) (char *, void *,  struct svc_req *))_a_1;
        return local(NULL, NULL, NULL);
}

Now do this:

$ gcc -Wall -fsanitize=undefined tst.c 
$ ./a.out
$ clang -Wall -fsanitize=undefined tst.c 
$ ./a.out
tst.c:15:9: runtime error: call to function _a_1 through pointer to incorrect function type 'int (*)(char *, void *, struct svc_req *)'
tst.c: note: _a_1 defined here
SUMMARY: UndefinedBehaviorSanitizer: undefined-behavior tst.c:15:9

As can be seen, only clang complained.
I don't know if it is a real UB or a false-positive.
But as gcc keeps silence, I assume this is
a false-positive.
There is no way to "just fix" the code, as it
is generated by rpcgen.

[DimitryAndric]

I think UBSan is correct: _a_1 is defined with a char ** as its first argument, but the call is done with a char * as its first argument. Does the runtime error go away if you fix that?

[stsp]

It does, but if, for example, I change
char ** to void * in _a_1, then error
does NOT disappear.
So its not like clang's ubsan wants the
proper pointer indirection, no, it wants
the precise type match. Even void *
doesn't match to char *.

[stsp]

I edited the original example to use
void * instead of char **.

[pinskia]

GCC does not currently implements the function mismatch specific UBSAN case yet.

[stsp]

So the UB is real and the ticket should
be closed?