On Mon, Aug 28, 2023 at 3:53 AM David Laight <David.Laight@xxxxxxxxxx> wrote:
>
> From: Linus Torvalds
> > Sent: 25 August 2023 21:43
> ....
> > Clang turns this:
> >
> > return __ffs64(val);
> >
> > into this horror:
> >
> > pushq %rax
> > movq %rdi, (%rsp)
> > #APP
> > rep
> > bsfq (%rsp), %rax
> > #NO_APP
> > popq %rcx
> >
> > which is just incredibly broken on so many levels. It *should* be a
> > single instruction, like gcc does:
> >
> > rep; bsf %rdi,%rax # tmp87, word
> >
> > but clang decides that it really wants to put the argument on the
> > stack, and apparently also wants to do that nonsensical stack
> > alignment thing to make things even worse.
> >
> > We use this:
> >
> > static __always_inline unsigned long variable__ffs(unsigned long word)
> > {
> > asm("rep; bsf %1,%0"
> > : "=r" (word)
> > : "rm" (word));
> > return word;
> > }
> >
> > for the definition, and it looks like clang royally just screws up
> > here. Yes, "m" is _allowed_ in that input set, but it damn well
> > shouldn't be used for something that is already in a register, since
> > "r" is also allowed, and is the first choice.
>
> Why don't we just remove the "m" option?
>
> Pretty much the only time it will be worse is it the value
> is in memory and loading it into a register causes a spill
> to stack.
>
> While it is possible to generate code where that happens it
> is pretty unlikely.
As Linus expressed below, register exhaustion could occur. Besides,
this is a bug in clang that we acknowledge, and should fix. I have
the general idea where things are going wrong, I just don't yet have
the muscle memory (or time) to dive into the register allocator.
>
> David
>
> -
> Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
> Registration No: 1397386 (Wales)
--
Thanks,
~Nick Desaulniers
