Inline "noinline" function if becomes empty

[davidbolvansky]

Bugzilla Link	50011
Version	trunk
OS	Linux
CC	@aeubanks,@alinas,@dwblaikie,@mtrofin

Extended Description

static void *ptr1, *ptr2;

static void doit(void **pptr, int cond)
{
  if (cond) {
    *pptr = &ptr2;
  }
}

__attribute__ ((noinline))
static void f(int cond)
{
  doit (&ptr1, cond);
}


int main()
{
  f (1);
}



llvm -O3:
main:                                   # @​main
        push    rax
        call    f
        xor     eax, eax
        pop     rcx
        ret
f:                                      # @​f
        ret

https://godbolt.org/z/7387cer5j

[dwblaikie]

Your bug is that you want a noinline function to be inlined if the function is empty?

I don't think that's desirable - noinline functions may be needed for a variety of reasons that might still apply even if they're empty.

[aeubanks]

I think it should be fine to delete calls to empty functions.
We should be able to infer after f is empty that f is readnone, and be able to delete a call to it. This might have to do with the fact that funcattrs runs before the function simplification pipeline, not after. Will check in a bit.

[dwblaikie]

fair enough - we do do a bunch of IPC with noinline functions and probably eliminate call to them in a bunch of cases.

Feels a bit off to me, but you're probably right there.

[aeubanks]

GlobalOpt promotes ptr1 to a local alloca only if it's being used in one function. That only happens after doit is inlined into f. But GlobalOpt only runs before and after the CGSCC pipeline (which includes inlining), not during, so the FunctionAttrs pass within the CGSCC pipeline can't infer readnone.

There actually are plans to explore making GlobalOpt run interleaved with the CGSCC pipeline which might help with this case. At some point I'll have a prototype for that.

[aeubanks]

Actually it might be tricky since we'd want Inliner (inline and remove dead function) -> GlobalOpt (replace global with local alloca) -> InstCombine (remove dead stores to alloca) -> FuncAttrs (infer readnone). Interleaving GlobalOpt like that would be weird since it would run on a RefSCC, which is a larger IR type than the other passes.