fix: subtract 1 from call chain addresses other than the first one

[pcc]

For stack frames other than the first one, ip is a return address, so it points to the instruction after the call instruction. By using the return address as is, we were incorrectly attributing these call chain samples to the instruction following the call. Therefore, subtract 1 from the return address so that it gets correctly attributed to the call instruction.

[pcc]

Reproducer:

$ cat 1.c
int x;

void f() {
  x++;
}

__attribute__((weak))
void force_frame_pointer() {}

__attribute__((noinline))
void infloop() {
#ifdef FP
  force_frame_pointer();
#endif
  // Use asm to prevent the compiler from optimizing away the rest of main.
#ifdef __aarch64__
  __asm__ __volatile__("b .");
#elif defined(__x86_64__)
  __asm__ __volatile__("jmp .");
#endif
}

int main() {
  infloop();
  f();
}
$ clang -fuse-ld=mold 1.c -gdwarf-4 -DFP -fno-omit-frame-pointer -O
$ perf record -g ./a.out
$ hotspot perf.data

Without this change: f appears in flame graph even though it is never executed.

With this change: f does not appear.