Re: [PATCH v7 0/2] x86/asm/bitops: optimize ff{s,z} functions for constant expressions

[Vincent MAILHOL <mailhol.vincent@wanadoo.fr>]

On Wed. 7 Sep. 2022 at 16:04, Nick Desaulniers <ndesaulniers@google.com> wrote:
> On Tue, Sep 6, 2022 at 11:26 AM Nick Desaulniers
> <ndesaulniers@google.com> wrote:
> >
> > On Sun, Sep 4, 2022 at 5:38 PM Vincent Mailhol
> > <mailhol.vincent@wanadoo.fr> wrote:
> > >
> > > The compilers provide some builtin expression equivalent to the ffs(),
> > > __ffs() and ffz() functions of the kernel. The kernel uses optimized
> > > assembly which produces better code than the builtin
> > > functions. However, such assembly code can not be folded when used
> > > with constant expressions.
> >
> > Another tact which may help additional sources other than just the
> > Linux kernel; it seems that compilers should be able to fold this.

Initially, I thought that you were suggesting folding the asm code
(which doesn’t seem trivial at all).

> > Vincent, if you're interested in making such an optimization in LLVM,
> > we'd welcome the contribution, and I'd be happy to show you where to
> > make such changes within LLVM; please let me know off thread.
>
> Oh right, it already does.
> https://github.com/llvm/llvm-project/blob/ea953b9d9a65c202985a79f1f95da115829baef6/llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp#L2635
> I see what's happening. Constant propagation sinks constants into a
> specialized version of ffs when there's only 1 callsite in a given
> translation unit (or multiple call sites with the same constant).
> Then dead argument elimination removes the argument, so libcall
> optimization thinks this isn't the ffs(int) you're looking for, and
> skips it.

Isn’t it a wise decision to skip it? How should the optimization be
able to decide that the redefined ffs() is equivalent to
__builtin_ffs()?

More generally, if I write my own foo() which shadows a
__builtin_foo() function, the two functions might do something totally
different and I would be pissed off if the compiler decided to
constant-fold my foo().

Dummy example:

===================
char *s;

/* ffs: fast forward string
 * @i: how many bytes to move forward
 *
 * Move forward the global s pointer by @i or strlen(s) (whoever is smaller).
 *
 * Return: how many bytes we move forward.
 */
int ffs(int i)
{
        int len = strlen(s);
        int forward = i < len ? i : len;

        s += forward;
        return forward;
}
===================

How would you instruct the compiler to constant-fold the kernel’s
ffs() but not fold above dummy ffs()?

> Nice.
> https://github.com/llvm/llvm-project/issues/57599
> I guess ffs() is usually forward declared in strings.h, so we don't
> have such a static inline definition available to constant
> prop/specialize in normal C code.