https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102953
--- Comment #7 from Andrew Cooper <andrew.cooper3 at citrix dot com> ---
Thankyou. I've tried these two patches and they do appear to be behaving as
intended.
I've put together a slightly extended version of the original test. Compile
with gcc -Wall -fno-pic -Os -fcf-protection=branch -mmanual-endbr
-fcf-check-attribute=no
static void __attribute__((noinline)) a(void)
{
asm volatile ("sti":::"memory");
}
void __attribute__((nocf_check, noinline)) b(void)
{
asm volatile ("std":::"memory");
}
void __attribute__((cf_check, noinline)) c(void)
{
asm volatile ("cmc":::"memory");
}
void (*ptr_a)(void) = a; // Now raises a diagnostic
void (*ptr_b)(void) = b; // Did diagnose previously, still does
void (*ptr_c)(void) = c;
int test(void)
{
ptr_a();
ptr_b();
ptr_c();
a();
b();
c(); // When fully linked, does skip c's ENDBR64 instruction
return 0;
}
int main(void)
{
return 0;
}
I'll try applying this to a bigger codebase now.
