incr.comp.: Turn translation-related attributes into a query.

[michaelwoerister]

There are a few attributes, like #[inline], #[cold], and #[target_feature], that would profit from being turned into a query:

• directly accessing the ast::Attribute will lead to a dependency edge to HIR which is brittle because of spans,
• at least #[inline] is searched for a few times, so caching might be good.

The attributes in question are listed in trans::attributes::from_fn_attrs():

rust/src/librustc_trans/attributes.rs

Lines 97 to 132 in 27ede55

	pub fn from_fn_attrs(ccx: &CrateContext, attrs: &[ast::Attribute], llfn: ValueRef) {
	use syntax::attr::*;
	inline(llfn, find_inline_attr(Some(ccx.sess().diagnostic()), attrs));
	set_frame_pointer_elimination(ccx, llfn);
	set_probestack(ccx, llfn);
	let mut target_features = vec![];
	for attr in attrs {
	if attr.check_name("target_feature") {
	if let Some(val) = attr.value_str() {
	for feat in val.as_str().split(",").map(|f| f.trim()) {
	if !feat.is_empty() && !feat.contains('\0') {
	target_features.push(feat.to_string());
	}
	}
	}
	} else if attr.check_name("cold") {
	Attribute::Cold.apply_llfn(Function, llfn);
	} else if attr.check_name("naked") {
	naked(llfn, true);
	} else if attr.check_name("allocator") {
	Attribute::NoAlias.apply_llfn(
	llvm::AttributePlace::ReturnValue, llfn);
	} else if attr.check_name("unwind") {
	unwind(llfn, true);
	} else if attr.check_name("rustc_allocator_nounwind") {
	unwind(llfn, false);
	}
	}
	if !target_features.is_empty() {
	let val = CString::new(target_features.join(",")).unwrap();
	llvm::AddFunctionAttrStringValue(
	llfn, llvm::AttributePlace::Function,
	cstr("target-features\0"), &val);
	}
	}

A query could return a TransFnAttrs value that looks like:

struct TransFnAttrs {
    inline: InlineAttr, // from syntax::attr
    flags: TransFnAttrFlags,
    target_features: Symbol,
}

bitflags! {
    pub struct TransFnAttrFlags: u8 {
        const COLD                      = 0b0000_0001;
        const ALLOCATOR                 = 0b0000_0010;
        const UNWIND                    = 0b0000_0100;
        const RUSTC_ALLOCATOR_NOUNWIND  = 0b0000_1000;
        const NAKED                     = 0b0001_0000;
    }
}

The syntax::attr::{find_inline_attr, requests_inline} should be replaced by this new query then.

The ultimate goal of this refactoring is to reduce the number of false positives during incr. comp. cache invalidation by providing a "firewall" between HIR and the compile_codegen_unit query.

[michaelwoerister]

TransFnAttrFlags should also contain the #[rustc_std_internal_symbol] attribute used here:

rust/src/librustc_trans/back/symbol_export.rs

Lines 216 to 217 in 27ede55

	let std_internal = attr::contains_name(&tcx.get_attrs(sym_def_id),
	"rustc_std_internal_symbol");

It should also subsume the contains_extern_indicator and export_name queries.

[michaelwoerister]

#[linkage] should also be included as an Option<rustc::mir::mono::Linkage> field.

[alexcrichton]

FWIW I started doing this for target_feature in #47223, although that'd probably just be used for the final implementation of this!

[michaelwoerister]

#47223 looks like a step in the right direction. The important thing is to extract things from attributes into queries so they can be checked for changes separately. However, since many of these things occur together, it would be a nice optimization to coalesce them.

[alexcrichton]

Indeed!

[wesleywiser]

I'm starting to work on this.

[michaelwoerister]

Awesome, thank you @wesleywiser! Let me know if you need any further directions.

[wesleywiser]

This comment is just for me to organize a TODO list of what needs to be done:

• inline attribute
• Various flags attributes
• target_features attribute
• Remove syntax::attr::find_inline_attr
• Remove syntax::attr:requests_inline
• rustc_std_internal_symbol attribute
• Remove contains_extern_indicator query
• Remove export_name query
• linkage attribute

[michaelwoerister]

Thanks for keeping at it, @wesleywiser!

[wesleywiser]

Sorry it's taking so long @michaelwoerister. My changes for the inline attribute broke the stage1 compiler and it took awhile to find the issue.

[michaelwoerister]

@wesleywiser, btw, I sent you a pm on gitter. I didn't find another way of contacting you.