Re: [PATCH v7 16/31] rust: ptr: add const_align_up()

[Alice Ryhl]

On Fri, Mar 20, 2026 at 11:27 AM David Rheinsberg <david@xxxxxxxxxxxx> wrote:
>
> Hi Alice!
>
> On Fri, Mar 20, 2026, at 10:47 AM, Alice Ryhl wrote:
> > On Fri, Mar 20, 2026 at 10:27 AM David Rheinsberg <david@xxxxxxxxxxxx> wrote:
> >> Kinda off-topic: why doesn't `Alignment` store a u8 that represents the exponent, rather than the power? The left-shift when needing the power should be effectively free, shouldn't it? It would avoid all the unsafety in the impl.
> >
> > For one, it mirrors the design of the unstable stdlib Alignment type.
> > For another, it'd make Alignment::of and similar a pain to implement.
>
> Fair enough!
>
> >> Do you need this helper then at all? I assume it is added because `Alignable` cannot be used in const. But it hard-codes `usize` as type, yet does not reflect that in the name. It comes down to which one is more readable, I guess:
> >>
> >>     const_align_up(value, align)
> >> vs
> >>     value.checked_next_multiple_of(align.as_usize())
> >>
> >> Meh, I don't mind too much. Just wanted to point out that the standard library provides this exactly.
> >
> > The stdlib implementation probably invokes an expensive division
> > operation in place of our bitwise-and operation when the argument is
> > not a compile-time constant value.
>
> Fortunately, it does not! Yay! `Alignment::as_nonzero()` has enough annotations to produce competitive assembly. If you care, see my example on x86_64 below, which uses:
>
> align_up1: v.checked_next_multiple_of(align.as_nonzero())
> align_up2: const_align_up(v, align)
>
> Thanks
> David
>
>
> align_up1:
>         lea     rax, [rsi - 1]
>         and     rax, rdi
>         sub     rsi, rax
>         lea     rdx, [rsi + rdi]
>         cmp     rdx, rdi
>         setae   cl
>         test    rax, rax
>         sete    al
>         cmove   rdx, rdi
>         or      al, cl
>         movzx   eax, al
>         ret
>
> align_up2:
>         lea     rdx, [rdi + rsi]
>         dec     rdx
>         xor     eax, eax
>         cmp     rdx, rdi
>         setae   al
>         neg     rsi
>         and     rdx, rsi
>         ret

The conditional move instruction still seems worse as it breaks data
dependencies in the cpu pipeline, right?

Alice