calling functions with implicit-enabled target features is unsafe

[usamoi]

rust-analyzer version: 0.4.2302-standalone
rustc version: rustc 1.86.0-nightly (ef148cd7e 2025-02-12)
code snippet to reproduce:

#![feature(target_feature_11)]

#[target_feature(enable = "avx")]
fn f() {}

#[target_feature(enable = "avx2")]
fn g() {
    f();
    // it's safe, because enabling `avx2` enables `avx` implicitly, but ra reports that it's unsafe
    // https://github.com/rust-lang/reference/pull/1720 describe it, and nightly rustc agrees with it
}

This should be caused by

rust-analyzer/crates/hir-ty/src/utils.rs

Lines 316 to 320 in 51389bd

	// RFC 2396 <https://rust-lang.github.io/rfcs/2396-target-feature-1.1.html>.
	let callee_target_features = TargetFeatures::from_attrs(&db.attrs(func.into()));
	if !caller_target_features.enabled.is_superset(&callee_target_features.enabled) {
	return Unsafety::Unsafe;
	}

It should calculate the set of all explicit-enabled and implicit-enabled target features of caller and calee, and then check if's a subset.

The "target feature to implicit-enabled target features" relationship is recorded in https://github.com/rust-lang/rust/blob/master/compiler/rustc_target/src/target_features.rs. (How to get it in ra?)

[Veykril]

cc @ChayimFriedman2

The "target feature to implicit-enabled target features" relationship is recorded in https://github.com/rust-lang/rust/blob/master/compiler/rustc_target/src/target_features.rs. (How to get it in ra?)

We can't depend on rustc_target unfortunately (due to requiring nightly and also depending on rustc_span, both of which I tried to gate away rust-lang/rust@master...Veykril:rust:veykril/rustc_target-rust-analyzer but gave up at some point)

[ChayimFriedman2]

I could swear I checked that. Weird...

Anyway, I'm not sure it will be a good idea to embed a list of target feature implications into rust-analyzer. I will try again later to make it compiler on stable (what will be the process for that? Create a compiler team MCP?) and if I'll fail, we'll see what to do.

[Veykril]

(what will be the process for that? Create a compiler team MCP?)

Probably ask t-compiler / MCP yes.

The main reason I dropped it was that I only wanted to look briefly into it (because of the assembly things) but due to the pervasive use of symbols I kind of gave up as making all of those generic (or outright replace the symbol use with string literals) seemed like too much to propose.

[ChayimFriedman2]

What I imagine is a sym!(symbol) macro that expands to either sym::symbol or to "symbol".

[Veykril]

That might be a way forward. We should open a zulip topic for a vibe check on this ideally though, you wanna do that or should I?

[ChayimFriedman2]

Isn't that the job of the MCP?

[Veykril]

Well I guess? If you already have a concrete design in mind an MCP is probably the way to go then

[ChayimFriedman2]

MCP submitted: rust-lang/compiler-team#839.