[PATCH v2 15/16] powerpc/powernv/sriov: Make single PE mode a per-BAR setting

[Segher Boessenkool]

On Mon, Aug 03, 2020 at 03:57:11PM +1000, Michael Ellerman wrote:
> > I would assume the function should still be generated since those checks
> > are relevant, just the return value is bogus.
> 
> Yeah, just sometimes missing warnings boil down to the compiler eliding
> whole sections of code, if it can convince itself they're unreachable.

Including when the compiler considers they must be unreachable because
they would perform undefined behaviour, like, act on uninitialised
values.  Such code is removed by cddce ("control dependence dead code
elimination", enabled by -ftree-dce at -O2 or above).

> AFAICS there's nothing weird going on here that should confuse GCC, it's
> about as straight forward as it gets.

Yes.  Please open a bug?

> Actually I can reproduce it with:
> 
> $ cat > test.c <<EOF
> int foo(void *p)
> {
>         unsigned align;
> 
>         if (!p)
>                 return align;
> 
>         return 0;
> }
> EOF
> 
> $ gcc -Wall -Wno-maybe-uninitialized -c test.c
> $
> 
> No warning.

The whole if() is deleted pretty early.

===
static int foo(void *p)
{
        unsigned align;

        if (!p)
                return align;

        return 42;
}
int bork(void) { return foo(0); }
===

doesn't warn either, although that always uses the uninitialised var
(actually, that code is *removed*, and it always does that "return 42").

> But I guess that's behaving as documented. The compiler can't prove that
> foo() will be called with p == NULL, so it doesn't warn.

-Wmaybe-uninitialized doesn't warn for this either, btw.


Segher