Re: [PATCH] x86: Optimize __get_user_asm() assembly

[Uros Bizjak]

On Wed, Nov 26, 2025 at 4:18 PM david laight <david.laight@xxxxxxxxxx> wrote:
>
> On Wed, 26 Nov 2025 14:23:32 +0100
> Uros Bizjak <ubizjak@xxxxxxxxx> wrote:
>
> > On x86, byte and word moves (MOVB, MOVW) write only to the low
> > portion of a 32-bit register, leaving the upper bits unchanged.
> > Modern compilers therefore prefer using MOVZBL and MOVZWL to load
> > 8-bit and 16-bit values with zero-extension to the full register
> > width.
> >
> > Update __get_user_asm() to follow this convention by explicitly
> > zero-extending 8-bit and 16-bit loads from memory.
> >
> > An additional benefit of this change is that it enables the full
> > integer register set to be used for 8-bit loads. Also, it
> > eliminates the need for manual zero-extension of 8-bit values.
> >
> > There is only a minimal increase in code size:
>
> Interesting, where does that come from.
> I'd have thought it should shrink it.
> The mov[sz]x is one byte longer.
> Perhaps a lot of 8-bit values are never zero-extended?

bloat-o-meter says:

add/remove: 0/0 grow/shrink: 4/0 up/down: 16/0 (16)
Function                                     old     new   delta
strncpy_from_kernel_nofault                  151     158      +7
copy_from_kernel_nofault                     207     214      +7
strncpy_from_user                            218     219      +1
fault_in_readable                            150     151      +1
Total: Before=23919018, After=23919034, chg +0.00%

where the difference in copy_from_kernel_nofault() is expected:

-       66 8b 45 00             mov    0x0(%rbp),%ax
+       0f b7 45 00             movzwl 0x0(%rbp),%eax
...
-       8a 45 00                mov    0x0(%rbp),%al
+       0f b6 45 00             movzbl 0x0(%rbp),%eax

and then some NOPs at the end due to function entry alignment.

> I'm trying to remember what the magic 'k' is for.
> Does gas treat %krxx as %dxx ?
> Which would matter if 'x' is 64bit when using "=r" for 32bit reads.
> But, in that case, I think you need it on the 'movl' as well.

%k is gcc operand modifier:

‘k’        Print the SImode name of the register.

that overrides operand width and prints a 32-bit register name.

> There is the slight problem (which affects for of the asm) is that
> gcc (I think) can't assume that the high 32bits of a register result
> from an inline asm function are zero.

No, the compiler can't assume that. It doesn't know that highpart is
zero and ...

> But if you return 'u64' then everything at the call site has to
> be extended as well.

... has to zero-extend from u32 to u64 even if we know highpart is zero.

> I wonder (not looked at the generated code yet) whether casting
> the u64 result from the function to u32 has the effect of avoiding
> the zero extend if the value is needed as a 64bit one.

No, it still has to follow C promotion rules, so e.g.:

unsigned int foo (unsigned int a)
{
    return (unsigned char)a;
}

always extends function arguments:

foo:
       movzbl  %dil, %eax
       ret

Uros.