https://gcc.gnu.org/bugzilla/show_bug.cgi?id=117267
Bug ID: 117267
Summary: 'setjmp' can never be copied because it receives a
non-local goto
Product: gcc
Version: 14.2.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c
Assignee: unassigned at gcc dot gnu.org
Reporter: stsp at users dot sourceforge.net
Target Milestone: ---
The test code below works with
clang but is rejected by gcc:
---
#include <stdio.h>
#if 0
#include <setjmp.h>
#else
typedef struct __jmp_buf {
void *buf[5];
volatile int rc;
} jmp_buf[1];
__attribute__((noreturn))
static inline void longjmp(jmp_buf env, int val)
{
env[0].rc = val;
__builtin_longjmp(env[0].buf, 1);
}
__attribute__((returns_twice,always_inline))
static inline int setjmp(jmp_buf env)
{
int rc = __builtin_setjmp(env[0].buf);
return (rc ? env[0].rc : 0);
}
#endif
static jmp_buf jb;
static void foo(void)
{
longjmp(jb, 1);
}
int main(void)
{
if (setjmp(jb))
printf("Jumped\n");
else
foo();
return 0;
}
---
If I remove attribute always_inline,
which is obviously a wrong thing to do,
then the compilation succeeds, but gcc
generates the nonsentical code. clang
in that case still generates correct code,
but of course you grab the invalid stack
frame, so its not working as expected.
But gcc for some reason generates garbage.
