Re: [PATCH v2 -tip] x86/percpu: Use C for arch_raw_cpu_ptr()

[Uros Bizjak]

On Thu, Oct 12, 2023 at 7:10 PM Linus Torvalds
<torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
>
> On Thu, 12 Oct 2023 at 09:55, Uros Bizjak <ubizjak@xxxxxxxxx> wrote:
> >
> > An example:
>
> Oh, I'm convinced.
>
> The fix seems to be a simple one-liner, ie just
>
> -       asm(__pcpu_op2_##size(op, __percpu_arg(P[var]), "%[val]")       \
> +       asm(__pcpu_op2_##size(op, __percpu_arg(a[var]), "%[val]")       \

The effect of the change:

25542442        4387686  808452 30738580        1d50894 vmlinux-new.o
25546484        4387686  808452 30742622        1d5185e vmlinux-old.o

> and it turns out that we have other places where I think we could use that '%a',
>
> For example, we have things like this:
>
>         asm ("lea sme_cmdline_arg(%%rip), %0"
>              : "=r" (cmdline_arg)
>              : "p" (sme_cmdline_arg));
>
> and I think the only reason we do that ridiculous asm is that the code
> in question really does want that (%rip) encoding. It sounds like this
> could just do
>
>         asm ("lea %a1, %0"
>              : "=r" (cmdline_arg)
>              : "p" (sme_cmdline_arg));
>
> instead. Once again, I claim ignorance of the operand modifiers as the
> reason for these kinds of things.
>
> But coming back to the stable op thing, I do wonder if there is some
> way we could avoid the unnecessary reload.
>
> I don't hate Nadav's patch, so that part is fine, but I'd like to
> understand what it is that makes gcc think it needs to reload. We have
> other cases (like the ALTERNATIVE() uses) where we *have* to use
> inline asm, so it would be good to know...
>
> Is it just that "p" (in the constraint, not "P" in the modifier) ends
> up always being seen as a memory access, even when we only use the
> address?
>
> That part has never really been something we've been entirely clear
> on. We *are* passing in just the address, so the hope in *that* place
> is that it's only an address dependency, not a memory one.

Let's see the difference of:

--cut here--
int m;

void foo (void)
{
  asm ("# %a0" :: "p" (&m));
}

void bar (void)
{
  asm ("# %0" :: "m" (m));
}
--cut here--

The internal dump shows:

(insn:TI 5 2 15 2 (parallel [
            (asm_operands/v ("# %a0") ("") 0 [
                    (symbol_ref:DI ("m") [flags 0x2]  <var_decl
0x7f3175011bd0 m>)
                ]
                 [
                    (asm_input:DI ("p") rip.c:5)
                ]
                 [] rip.c:5)
            (clobber (reg:CC 17 flags))
        ]) "rip.c":5:3 -1
     (expr_list:REG_UNUSED (reg:CC 17 flags)
        (nil)))

vs:

(insn:TI 5 2 13 2 (parallel [
            (asm_operands/v ("# %0") ("") 0 [
                    (mem/c:SI (symbol_ref:DI ("m") [flags 0x2]
<var_decl 0x7f3175011bd0 m>) [1 m+0 S4 A32])
                ]
                 [
                    (asm_input:SI ("m") rip.c:10)
                ]
                 [] rip.c:10)
            (clobber (reg:CC 17 flags))
        ]) "rip.c":10:3 -1
     (expr_list:REG_UNUSED (reg:CC 17 flags)
        (nil)))

The first argument is internally regarded as "constant":

-- Macro: CONSTANT_P (X)
    'CONSTANT_P', which is defined by target-independent code, accepts
    integer-values expressions whose values are not explicitly known,
    such as 'symbol_ref', 'label_ref', and 'high' expressions and
    'const' arithmetic expressions, in addition to 'const_int' and
    'const_double' expressions.

So, it should not have any dependency.

Perhaps a testcase should be created and posted to gcc-bugs for
further analysis.

Uros.