Re: [PATCH v3 1/2] rust: std_vendor: add `likely()`, `unlikely()` and `cold_path()`

["Gary Guo" <gary@garyguo.net>]

On Fri Oct 16, 2026 at 5:24 PM BST, Miguel Ojeda wrote:
> `{un,}likely()` are commonly used macros on the C side -- there are
> more than 25k occurrences in the tree. Rust is likely going to provide
> functions (not macros) for the same purpose: `core::hint::{un,}likely()`.
> Currently they are unstable under `feature(likely_unlikely)` [1].
> 
> For instance, given:
> 
>     fn f1(a: i32) -> i32 {
>         if a < 0 {
>             return 123;
>         }
> 
>         42
>     }
> 
>     fn f2(a: i32) -> i32 {
>         if likely(a < 0) {
>             return 124;
>         }
> 
>         42
>     }
> 
>     fn f3(a: i32) -> i32 {
>         if unlikely(a < 0) {
>             return 125;
>         }
> 
>         42
>     }
> 
> LLVM emits code equivalent to the one it would for similar C functions:
> 
>     f1:
>             test   %edi,%edi
>             mov    $0x7b,%ecx
>             mov    $0x2a,%eax
>             cmovs  %ecx,%eax
>             ret
> 
>     f2:
>             mov    $0x7c,%eax
>             test   %edi,%edi
>         /-- jns    <f2+0xa>
>         |   ret
>         \-> mov    $0x2a,%eax
>             ret
> 
>     f3:
>             mov    $0x2a,%eax
>             test   %edi,%edi
>         /-- js     <f3+0xa>
>         |   ret
>         \-> mov    $0x7d,%eax
>             ret
> 
> Thus vendor (essentially) the `core` library code to provide the same
> API (so that we may use `core`'s directly in the future).
> 
> We do so by implementing `{un,}likely()` like `core`'s `intrinsics` do
> (to save a layer) and based on calling our own `cold_path()` instead of
> using `core`'s (which is stable since 1.95.0 [2][3]).
> 
> It turns out that simply using `#[cold]` to emulate `cold_path()`,
> suggested by Gary [4], works in the cases I tested across all supported
> versions so far, so we can avoid the complexity of having one more Kconfig
> and conditionally enabling the feature in callers (since `cold_path()`
> was added in Rust 1.86.0 [5]).
> 
> We could optionally call `cold_path()` for Rust >= 1.86.0 within our own
> `cold_path()`, but it seems that it is not needed, and anyway since Rust
> 1.98.0 `cold_path()` is itself `#[cold]` so that it works even if the
> MIR inliner can't inline it [6].
> 
> Even if we find that it is not optimal in some cases, it can be improved,
> and meanwhile we will have the means to annotate these cases already.
> 
> Apart from that, the only differences with `core`'s version are the
> usual changes (`pr_*!` calls and their `\n`, imports, attributes, lint
> cleanup), plus a fix for a polarity typo (I sent a fix upstream [7])
> and removal of unneeded code documentation.
> 
> In addition, add them to the prelude (inlined, so that the documentation
> is rendered at least there, until we can use `core`'s).
> 
> Link: https://github.com/rust-lang/rust/issues/151619 [1]
> Link: https://github.com/rust-lang/rust/issues/136873 [2]
> Link: https://github.com/rust-lang/rust/pull/151576 [3]
> Link: https://lore.kernel.org/rust-for-linux/DGA6GUR58QJ7.1XZZ0P4VZNW86@garyguo.net/ [4]
> Link: https://github.com/rust-lang/rust/pull/133695 [5]
> Link: https://github.com/rust-lang/rust/pull/156863 [6]
> Link: https://github.com/rust-lang/rust/pull/164374 [7]
> Assisted-by: LLM
> Signed-off-by: Miguel Ojeda <ojeda@kernel.org>

Reviewed-by: Gary Guo <gary@garyguo.net>

> ---
> v3: I used Gary's `#[cold]` suggestion, and thus removed the Kconfig.
> I rebased and made sure the code is up to date with `core`, tested a
> bunch of things, compared again with C and the assembly generated, found
> a strange behavior on Clang I will report separately, a polarity typo in
> `core` (fixed here) and inlined the docs in the prelude to get it
> rendered somewhere until we bump the MSRV. I think the potential issue
> I found in v2's discussion may have been with `-Os`, but it happens the
> same way with `core`'s, so it does not matter.
> 
> Alice, Gary: I removed your Reviewed-by's just in case. It is pretty
> similar to last time, but the commit message is reworded, and the commit
> is simpler.
> 
>  rust/kernel/prelude.rs    |   7 ++
>  rust/kernel/std_vendor.rs | 136 ++++++++++++++++++++++++++++++++++++++
>  2 files changed, 143 insertions(+)