add asm_cfg: #[cfg(...)] within asm!

[folkertdev]

tracking issue: #140364

This feature was discussed in #140279. It allows configuring templates and operands in the assembly macros, for example:

asm!( // or global_asm! or naked_asm!
    "nop",
    #[cfg(target_feature = "sse2")]
    "nop",
    // ...
    #[cfg(target_feature = "sse2")]
    a = const 123, // only used on sse2
);

r? @tgross35

cc @traviscross @Amanieu

This needs some squashing, but I suspect we'll iterate a bunch here. I'll leave some inline comments too.

[folkertdev]

The fact that we use parse_expr here mean that this does not work, and does not produce a great error message. This complains about an unexpected , after the 6, but the real problem is that this cannot possibly be or expand to a string literal.

asm!(
    #[cfg(false)]
    a = const 6,
    "nop",
);

Maybe we should specialize the parser to specifically only parse string literals and item macros?

[folkertdev]

attributes are now parsed before the parse_expr below, and never attached to the expression. In theory that could interact with #[feature(stmt_expr_attributes)], but:

• that feature is unstable
• any attributes it parsed would have no effect

[folkertdev]

non-cfg attributes are now accepted by the parser, but emit this error. Previously attributes were only parsed (and later rejected, unless #[feature(stmt_expr_attributes)] was enabled) on expressions. We now also always parse them on operands.

[rust-log-analyzer]

The job mingw-check failed! Check out the build log: (web) (plain)

Click to see the possible cause of the failure (guessed by this bot)

[RUSTC-TIMING] toml test:false 0.195
error[E0061]: this function takes 4 arguments but 3 arguments were supplied
   --> src/tools/rustfmt/src/parse/macros/asm.rs:10:5
    |
10  |     parse_asm_args(&mut parser, mac.span(), ast::AsmMacro::Asm).ok()
    |     ^^^^^^^^^^^^^^ ----------- argument #1 of type `&ExtCtxt<'_>` is missing
    |
note: function defined here
   --> /checkout/compiler/rustc_builtin_macros/src/asm.rs:134:8
    |
134 | pub fn parse_asm_args<'a>(
    |        ^^^^^^^^^^^^^^
help: provide the argument
    |
10  |     parse_asm_args(/* &ExtCtxt<'_> */, &mut parser, mac.span(), ast::AsmMacro::Asm).ok()
    |                    +++++++++++++++++++

For more information about this error, try `rustc --explain E0061`.
[RUSTC-TIMING] rustfmt_nightly test:false 2.371
error: could not compile `rustfmt-nightly` (lib) due to 1 previous error

[folkertdev]

Hmm, rust analyzer also parses the assembly, but from what I can see it has no way of evaluating the cfgs while parsing. Maybe that can change, but if not we'll have to fundamentally change the parser to just parse, and then later have a validation pass. I tried that before and it's possible but messy, and you need to store a lot more information in AsmArgs (that rust-analyzer then would also have to process).

@rust-lang/rust-analyzer how can we make this work?

[ChayimFriedman2]

(I haven't looked at the PR, but) parsing does not evaluated cfgs. Hir lowering does.

[folkertdev]

Well in any case the parser doesn't. I'm not exactly sure where to draw the line for when HIR lowering starts, but certainly macro expansion of builtin macros is able to do it, e.g. here for cfg!:

rust/compiler/rustc_builtin_macros/src/cfg.rs

Lines 24 to 30 in 267cae5

	let matches_cfg = attr::cfg_matches(
	&cfg,
	&cx.sess,
	cx.current_expansion.lint_node_id,
	Some(cx.ecfg.features),
	);
	MacEager::expr(cx.expr_bool(sp, matches_cfg))

Anyway, are you saying that rust-analyzer just cannot evaluate cfg's while parsing the asm! macro? We can delay it, but I think that just means that rust-analyzer would have to duplicate the logic.

[Veykril]

rust-analyzer has to duplicate these changes either way. We don't use any of the touched rustc crates here.