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.

Interesting. I do wonder if/how it would adjust the flame graphs in general and if this is only clang/mold related...

I think that sample could be added to the tests as well, no?

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

With this change: f does not appear.

Ideally this could get turned into an auto test, i.e. compile this reproducer into a static binary then record it with perf and then write a test that tries to analyze the now-portable results. we have some test code that stringifies the results and that should nicely show the before/after too

that said, I'm fine with merging this as-is and trusting your judgement on this front, so thanks a lot for your contribution! considering my lack of time, I don't want to block you on writing the tests, but it would be great to have. for now I'll create a task out of this.

PS: the test would live in hotspot btw