RFC 2396: target-feature-1.1

lang (typesystem | target_feature | effects)

Summary

This RFC attempts to resolve some of the unresolved questions in RFC 2045 (target_feature). In particular, it allows:

It achieves this by proposing three incremental steps that we can sequentially make to improve the ergonomics and the safety of target-specific functionality without adding run-time overhead.

Motivation

This is a brief recap of RFC 2045 (target_feature).

The #[target_feature] attribute allows Rust to generate machine code for a function under the assumption that the hardware where the function will be executed on supports some specific "features".

If the hardware does not support the features, the machine code was generated under assumptions that do not hold, and the behavior of executing the function is undefined.

RFC 2045 (target_feature) guarantees safety by requiring all #[target_feature] functions to be unsafe fn, thus preventing them from being called from safe code. That is, users have to open an unsafe { } block to call these functions, and they have to manually ensure that their pre-conditions hold - for example, that they will only be executed on the appropriate hardware by doing run-time feature detection, or using conditional compilation.

And that's it. That's all RFC 2045 (target_feature) had to say about this. Back then, there were many other problems that needed to be solved for all of this to be minimally useful, and RFC 2045 (target_feature) dealt with those.

However, the consensus back then was that this is far from ideal for many reasons:

The solution proposed in this RFC solves these problems.

Guide-level explanation

Currently, we require that #[target_feature] functions be declared as unsafe fn. This RFC relaxes this restriction:

// Example 1:
#[target_feature(enable = "sse2")] unsafe fn foo() { }  // RFC2045
#[target_feature(enable = "sse2")] fn bar() { }  // NEW

// This function does not have the "sse2" target feature:
fn meow() {
    foo(); // ERROR (unsafe block required)
    unsafe { foo() }; // OK
    bar(); // ERROR (meow is not sse2)
    unsafe { bar() }; // OK
}

#[target_feature(enable = "sse2")]
fn bark() {
    foo(); // ERROR (foo is unsafe: unsafe block required)
    unsafe { foo() }; // OK
    bar(); // OK (bark is sse2 and bar is safe)
    unsafe { bar() }; // OK (as well - warning: unnecessary unsafe block)
}

#[target_feature(enable = "avx")]  // avx != sse2
fn moo() {
    foo(); // ERROR (unsafe block required)
    unsafe { foo() }; // OK
    bar(); // ERROR (moo is not sse2 but bar requires it)
    unsafe { bar() }; // OK 
}

Note: while it is safe to call an SSE2 function from some AVX functions, this would require specifying how features relate to each other in hierarchies. It is unclear whether those hierarchies actually exist, but adding them to this RFC would unnecessarily complicate it and can be done later or in parallel to this one, once we agree on the fundamentals.

First, this is still sound. The caller has a super-set of #[target_features] of the callee. That is, the #[target_feature]-related pre-conditions of the callee are uphold by the caller, therefore calling the callee is safe.

This change already solves all three issues mentioned in the motivation:

The #[target_feature] attribute continues to be allowed on inherent methods - this RFC does not change that.

The #[target_feature] attribute continues to not be allowed on safe trait method implementations because that would require an unsafe trait method declaration:

// Example 2:
trait Foo { fn foo(); }
struct Fooish();
impl Foo for Fooish { 
    #[target_feature(enable = "sse2")] fn foo() { }  
    // ^ ERROR: #[target_feature] on trait method impl requires 
    // unsafe fn but Foo::foo is safe
    // (this is already an error per RFC2045)
}

trait Bar { unsafe fn bar(); }
struct Barish();
impl Bar for Barish { 
    #[target_feature(enable = "sse2")] unsafe fn bar() { }  // OK (RFC2045)
}
// Example 3
#[target_feature(enable = "avx")] fn meow() {}

static x: fn () -> () = meow;
// ^ ERROR: meow can only be assigned to unsafe fn pointers due to 
// #[target_feature] but function pointer x with type fn()->() is safe.
static y: unsafe fn () -> () = meow as unsafe fn()->(); // OK

Reference-level explanation

This RFC proposes to changes to the language with respect to RFC 2045 (target_feature):

Drawbacks

This RFC extends the typing rules for #[target_feature], which might unnecessarily complicate future language features like an effect system.

Rationale and alternatives

Since #[target_feature] are effects or restrictions (depending on whether we enable or disable them), the alternative would be to integrate them with an effect system.

Prior art

RFC2212 target feature unsafe attempted to solve this problem. This RFC builds on the discussion that was produced by that RFC and by many discussions in the stdsimd repo.

Unresolved questions

None.

Future possibilities

Negative features

RFC 2045 (target_feature) introduced the #[target_feature(enable = "x")] syntax to allow introducing negative features in future RFCs in the form of #[target_feature(disable = "y")]. Since these have not been introduced yet we can only speculate about how they would interact with the extensions proposed in this RFC but we probably can make the following work in some form:

// #[target_feature(enable = "sse")]
fn foo() {}

#[target_feature(disable = "sse")] 
fn bar() {
    foo(); // ERROR: (bar is not sse)
    unsafe { foo() }; // OK
}

fn baz() {
  bar(); // OK 
}

Effect system

It is unclear how #[target_feature] would interact with an effect system for Rust like the one being tracked here and discussed in RFC2237.

In particular, it is unclear how the typing rules being proposed here would be covered by such an effect system, and whether such system would support attributes in effect/restriction position.

Such an effect-system might need to introduce first-class target-features into the language (beyond just a simple attribute) which could lead to the deprecation of the #[target_feature] attribute.

It is also unclear how any of this interacts with effect-polymorphism at this point, but we could maybe support something like impl const Trait and T: const Trait:

impl #[target_feature(enable = "...")] Trait for Type { ... }
fn foo<T: #[target_feature(enable = "...")] Trait>(...) { ...}

if all trait methods are unsafe; otherwise they can't have the #[target_feature] attribute.