1 //! The Rust Linkage Model and Symbol Names
2 //! =======================================
4 //! The semantic model of Rust linkage is, broadly, that "there's no global
5 //! namespace" between crates. Our aim is to preserve the illusion of this
6 //! model despite the fact that it's not *quite* possible to implement on
7 //! modern linkers. We initially didn't use system linkers at all, but have
8 //! been convinced of their utility.
10 //! There are a few issues to handle:
12 //! - Linkers operate on a flat namespace, so we have to flatten names.
13 //! We do this using the C++ namespace-mangling technique. Foo::bar
16 //! - Symbols for distinct items with the same *name* need to get different
17 //! linkage-names. Examples of this are monomorphizations of functions or
18 //! items within anonymous scopes that end up having the same path.
20 //! - Symbols in different crates but with same names "within" the crate need
21 //! to get different linkage-names.
23 //! - Symbol names should be deterministic: Two consecutive runs of the
24 //! compiler over the same code base should produce the same symbol names for
27 //! - Symbol names should not depend on any global properties of the code base,
28 //! so that small modifications to the code base do not result in all symbols
29 //! changing. In previous versions of the compiler, symbol names incorporated
30 //! the SVH (Stable Version Hash) of the crate. This scheme turned out to be
31 //! infeasible when used in conjunction with incremental compilation because
32 //! small code changes would invalidate all symbols generated previously.
34 //! - Even symbols from different versions of the same crate should be able to
35 //! live next to each other without conflict.
37 //! In order to fulfill the above requirements the following scheme is used by
40 //! The main tool for avoiding naming conflicts is the incorporation of a 64-bit
41 //! hash value into every exported symbol name. Anything that makes a difference
42 //! to the symbol being named, but does not show up in the regular path needs to
43 //! be fed into this hash:
45 //! - Different monomorphizations of the same item have the same path but differ
46 //! in their concrete type parameters, so these parameters are part of the
47 //! data being digested for the symbol hash.
49 //! - Rust allows items to be defined in anonymous scopes, such as in
50 //! `fn foo() { { fn bar() {} } { fn bar() {} } }`. Both `bar` functions have
51 //! the path `foo::bar`, since the anonymous scopes do not contribute to the
52 //! path of an item. The compiler already handles this case via so-called
53 //! disambiguating `DefPaths` which use indices to distinguish items with the
54 //! same name. The DefPaths of the functions above are thus `foo[0]::bar[0]`
55 //! and `foo[0]::bar[1]`. In order to incorporate this disambiguation
56 //! information into the symbol name too, these indices are fed into the
57 //! symbol hash, so that the above two symbols would end up with different
60 //! The two measures described above suffice to avoid intra-crate conflicts. In
61 //! order to also avoid inter-crate conflicts two more measures are taken:
63 //! - The name of the crate containing the symbol is prepended to the symbol
64 //! name, i.e., symbols are "crate qualified". For example, a function `foo` in
65 //! module `bar` in crate `baz` would get a symbol name like
66 //! `baz::bar::foo::{hash}` instead of just `bar::foo::{hash}`. This avoids
67 //! simple conflicts between functions from different crates.
69 //! - In order to be able to also use symbols from two versions of the same
70 //! crate (which naturally also have the same name), a stronger measure is
71 //! required: The compiler accepts an arbitrary "disambiguator" value via the
72 //! `-C metadata` command-line argument. This disambiguator is then fed into
73 //! the symbol hash of every exported item. Consequently, the symbols in two
74 //! identical crates but with different disambiguators are not in conflict
75 //! with each other. This facility is mainly intended to be used by build
78 //! A note on symbol name stability
79 //! -------------------------------
80 //! Previous versions of the compiler resorted to feeding NodeIds into the
81 //! symbol hash in order to disambiguate between items with the same path. The
82 //! current version of the name generation algorithm takes great care not to do
83 //! that, since NodeIds are notoriously unstable: A small change to the
84 //! code base will offset all NodeIds after the change and thus, much as using
85 //! the SVH in the hash, invalidate an unbounded number of symbol names. This
86 //! makes re-using previously compiled code for incremental compilation
87 //! virtually impossible. Thus, symbol hash generation exclusively relies on
88 //! DefPaths which are much more robust in the face of changes to the code base.
90 use rustc::hir::def_id::{CrateNum, DefId, LOCAL_CRATE};
92 use rustc::hir::CodegenFnAttrFlags;
93 use rustc::hir::map::{DefPathData, DisambiguatedDefPathData};
94 use rustc::ich::NodeIdHashingMode;
95 use rustc::ty::print::{PrettyPrinter, Printer, Print};
96 use rustc::ty::query::Providers;
97 use rustc::ty::subst::{Kind, SubstsRef, UnpackedKind};
98 use rustc::ty::{self, Ty, TyCtxt, TypeFoldable};
99 use rustc::util::common::record_time;
100 use rustc_data_structures::stable_hasher::{HashStable, StableHasher};
101 use rustc_mir::monomorphize::item::{InstantiationMode, MonoItem, MonoItemExt};
102 use rustc_mir::monomorphize::Instance;
104 use syntax_pos::symbol::InternedString;
108 use std::fmt::{self, Write};
109 use std::mem::{self, discriminant};
111 pub fn provide(providers: &mut Providers<'_>) {
112 *providers = Providers {
119 fn get_symbol_hash<'a, 'tcx>(
120 tcx: TyCtxt<'a, 'tcx, 'tcx>,
122 // the DefId of the item this name is for
125 // instance this name will be for
126 instance: Instance<'tcx>,
128 // type of the item, without any generic
129 // parameters substituted; this is
130 // included in the hash as a kind of
134 // values for generic type parameters,
136 substs: SubstsRef<'tcx>,
139 "get_symbol_hash(def_id={:?}, parameters={:?})",
143 let mut hasher = StableHasher::<u64>::new();
144 let mut hcx = tcx.create_stable_hashing_context();
146 record_time(&tcx.sess.perf_stats.symbol_hash_time, || {
147 // the main symbol name is not necessarily unique; hash in the
148 // compiler's internal def-path, guaranteeing each symbol has a
150 tcx.def_path_hash(def_id).hash_stable(&mut hcx, &mut hasher);
152 // Include the main item-type. Note that, in this case, the
153 // assertions about `needs_subst` may not hold, but this item-type
154 // ought to be the same for every reference anyway.
155 assert!(!item_type.has_erasable_regions());
156 hcx.while_hashing_spans(false, |hcx| {
157 hcx.with_node_id_hashing_mode(NodeIdHashingMode::HashDefPath, |hcx| {
158 item_type.hash_stable(hcx, &mut hasher);
162 // If this is a function, we hash the signature as well.
163 // This is not *strictly* needed, but it may help in some
164 // situations, see the `run-make/a-b-a-linker-guard` test.
165 if let ty::FnDef(..) = item_type.sty {
166 item_type.fn_sig(tcx).hash_stable(&mut hcx, &mut hasher);
169 // also include any type parameters (for generic items)
170 assert!(!substs.has_erasable_regions());
171 assert!(!substs.needs_subst());
172 substs.hash_stable(&mut hcx, &mut hasher);
174 let is_generic = substs.non_erasable_generics().next().is_some();
175 let avoid_cross_crate_conflicts =
176 // If this is an instance of a generic function, we also hash in
177 // the ID of the instantiating crate. This avoids symbol conflicts
178 // in case the same instances is emitted in two crates of the same
182 // If we're dealing with an instance of a function that's inlined from
183 // another crate but we're marking it as globally shared to our
184 // compliation (aka we're not making an internal copy in each of our
185 // codegen units) then this symbol may become an exported (but hidden
186 // visibility) symbol. This means that multiple crates may do the same
187 // and we want to be sure to avoid any symbol conflicts here.
188 match MonoItem::Fn(instance).instantiation_mode(tcx) {
189 InstantiationMode::GloballyShared { may_conflict: true } => true,
193 if avoid_cross_crate_conflicts {
194 let instantiating_crate = if is_generic {
195 if !def_id.is_local() && tcx.sess.opts.share_generics() {
196 // If we are re-using a monomorphization from another crate,
197 // we have to compute the symbol hash accordingly.
198 let upstream_monomorphizations = tcx.upstream_monomorphizations_for(def_id);
200 upstream_monomorphizations
201 .and_then(|monos| monos.get(&substs).cloned())
202 .unwrap_or(LOCAL_CRATE)
210 (&tcx.original_crate_name(instantiating_crate).as_str()[..])
211 .hash_stable(&mut hcx, &mut hasher);
212 (&tcx.crate_disambiguator(instantiating_crate)).hash_stable(&mut hcx, &mut hasher);
215 // We want to avoid accidental collision between different types of instances.
216 // Especially, VtableShim may overlap with its original instance without this.
217 discriminant(&instance.def).hash_stable(&mut hcx, &mut hasher);
220 // 64 bits should be enough to avoid collisions.
224 fn symbol_name(tcx: TyCtxt<'_, 'tcx, 'tcx>, instance: Instance<'tcx>) -> ty::SymbolName {
226 name: compute_symbol_name(tcx, instance),
230 fn compute_symbol_name(tcx: TyCtxt<'_, 'tcx, 'tcx>, instance: Instance<'tcx>) -> InternedString {
231 let def_id = instance.def_id();
232 let substs = instance.substs;
234 debug!("symbol_name(def_id={:?}, substs={:?})", def_id, substs);
236 let hir_id = tcx.hir().as_local_hir_id(def_id);
238 if def_id.is_local() {
239 if tcx.plugin_registrar_fn(LOCAL_CRATE) == Some(def_id) {
240 let disambiguator = tcx.sess.local_crate_disambiguator();
242 InternedString::intern(&tcx.sess.generate_plugin_registrar_symbol(disambiguator));
244 if tcx.proc_macro_decls_static(LOCAL_CRATE) == Some(def_id) {
245 let disambiguator = tcx.sess.local_crate_disambiguator();
247 InternedString::intern(&tcx.sess.generate_proc_macro_decls_symbol(disambiguator));
251 // FIXME(eddyb) Precompute a custom symbol name based on attributes.
252 let is_foreign = if let Some(id) = hir_id {
253 match tcx.hir().get_by_hir_id(id) {
254 Node::ForeignItem(_) => true,
258 tcx.is_foreign_item(def_id)
261 let attrs = tcx.codegen_fn_attrs(def_id);
263 if let Some(name) = attrs.link_name {
264 return name.as_interned_str();
266 // Don't mangle foreign items.
267 return tcx.item_name(def_id);
270 if let Some(name) = &attrs.export_name {
272 return name.as_interned_str();
275 if attrs.flags.contains(CodegenFnAttrFlags::NO_MANGLE) {
277 return tcx.item_name(def_id);
280 // We want to compute the "type" of this item. Unfortunately, some
281 // kinds of items (e.g., closures) don't have an entry in the
282 // item-type array. So walk back up the find the closest parent
283 // that DOES have an entry.
284 let mut ty_def_id = def_id;
287 let key = tcx.def_key(ty_def_id);
288 match key.disambiguated_data.data {
289 DefPathData::TypeNs(_) | DefPathData::ValueNs(_) => {
290 instance_ty = tcx.type_of(ty_def_id);
294 // if we're making a symbol for something, there ought
295 // to be a value or type-def or something in there
297 ty_def_id.index = key.parent.unwrap_or_else(|| {
299 "finding type for {:?}, encountered def-id {:?} with no \
309 // Erase regions because they may not be deterministic when hashed
310 // and should not matter anyhow.
311 let instance_ty = tcx.erase_regions(&instance_ty);
313 let hash = get_symbol_hash(tcx, def_id, instance, instance_ty, substs);
315 let mut printer = SymbolPrinter {
317 path: SymbolPath::new(),
318 keep_within_component: false,
319 }.print_def_path(def_id, &[]).unwrap();
321 if instance.is_vtable_shim() {
322 let _ = printer.write_str("{{vtable-shim}}");
325 InternedString::intern(&printer.path.finish(hash))
328 // Follow C++ namespace-mangling style, see
329 // http://en.wikipedia.org/wiki/Name_mangling for more info.
331 // It turns out that on macOS you can actually have arbitrary symbols in
332 // function names (at least when given to LLVM), but this is not possible
333 // when using unix's linker. Perhaps one day when we just use a linker from LLVM
334 // we won't need to do this name mangling. The problem with name mangling is
335 // that it seriously limits the available characters. For example we can't
336 // have things like &T in symbol names when one would theoretically
337 // want them for things like impls of traits on that type.
339 // To be able to work on all platforms and get *some* reasonable output, we
340 // use C++ name-mangling.
349 let mut result = SymbolPath {
350 result: String::with_capacity(64),
351 temp_buf: String::with_capacity(16),
353 result.result.push_str("_ZN"); // _Z == Begin name-sequence, N == nested
357 fn finalize_pending_component(&mut self) {
358 if !self.temp_buf.is_empty() {
359 let _ = write!(self.result, "{}{}", self.temp_buf.len(), self.temp_buf);
360 self.temp_buf.clear();
364 fn finish(mut self, hash: u64) -> String {
365 self.finalize_pending_component();
366 // E = end name-sequence
367 let _ = write!(self.result, "17h{:016x}E", hash);
372 struct SymbolPrinter<'a, 'tcx> {
373 tcx: TyCtxt<'a, 'tcx, 'tcx>,
376 // When `true`, `finalize_pending_component` isn't used.
377 // This is needed when recursing into `path_qualified`,
378 // or `path_generic_args`, as any nested paths are
379 // logically within one component.
380 keep_within_component: bool,
383 // HACK(eddyb) this relies on using the `fmt` interface to get
384 // `PrettyPrinter` aka pretty printing of e.g. types in paths,
385 // symbol names should have their own printing machinery.
387 impl Printer<'tcx, 'tcx> for SymbolPrinter<'_, 'tcx> {
388 type Error = fmt::Error;
393 type DynExistential = Self;
395 fn tcx(&'a self) -> TyCtxt<'a, 'tcx, 'tcx> {
401 _region: ty::Region<'_>,
402 ) -> Result<Self::Region, Self::Error> {
409 ) -> Result<Self::Type, Self::Error> {
411 // Print all nominal types as paths (unlike `pretty_print_type`).
412 ty::FnDef(def_id, substs) |
413 ty::Opaque(def_id, substs) |
414 ty::Projection(ty::ProjectionTy { item_def_id: def_id, substs }) |
415 ty::UnnormalizedProjection(ty::ProjectionTy { item_def_id: def_id, substs }) |
416 ty::Closure(def_id, ty::ClosureSubsts { substs }) |
417 ty::Generator(def_id, ty::GeneratorSubsts { substs }, _) => {
418 self.print_def_path(def_id, substs)
420 _ => self.pretty_print_type(ty),
424 fn print_dyn_existential(
426 predicates: &'tcx ty::List<ty::ExistentialPredicate<'tcx>>,
427 ) -> Result<Self::DynExistential, Self::Error> {
428 let mut first = false;
429 for p in predicates {
434 self = p.print(self)?;
442 ) -> Result<Self::Path, Self::Error> {
443 self.write_str(&self.tcx.original_crate_name(cnum).as_str())?;
449 trait_ref: Option<ty::TraitRef<'tcx>>,
450 ) -> Result<Self::Path, Self::Error> {
451 // Similar to `pretty_path_qualified`, but for the other
452 // types that are printed as paths (see `print_type` above).
457 ty::UnnormalizedProjection(_) |
460 if trait_ref.is_none() =>
462 self.print_type(self_ty)
465 _ => self.pretty_path_qualified(self_ty, trait_ref)
471 print_prefix: impl FnOnce(Self) -> Result<Self::Path, Self::Error>,
472 _disambiguated_data: &DisambiguatedDefPathData,
474 trait_ref: Option<ty::TraitRef<'tcx>>,
475 ) -> Result<Self::Path, Self::Error> {
476 self.pretty_path_append_impl(
478 cx = print_prefix(cx)?;
480 if cx.keep_within_component {
481 // HACK(eddyb) print the path similarly to how `FmtPrinter` prints it.
484 cx.path.finalize_pending_component();
495 print_prefix: impl FnOnce(Self) -> Result<Self::Path, Self::Error>,
496 disambiguated_data: &DisambiguatedDefPathData,
497 ) -> Result<Self::Path, Self::Error> {
498 self = print_prefix(self)?;
500 // Skip `::{{constructor}}` on tuple/unit structs.
501 match disambiguated_data.data {
502 DefPathData::Ctor => return Ok(self),
506 if self.keep_within_component {
507 // HACK(eddyb) print the path similarly to how `FmtPrinter` prints it.
508 self.write_str("::")?;
510 self.path.finalize_pending_component();
513 self.write_str(&disambiguated_data.data.as_interned_str().as_str())?;
516 fn path_generic_args(
518 print_prefix: impl FnOnce(Self) -> Result<Self::Path, Self::Error>,
520 ) -> Result<Self::Path, Self::Error> {
521 self = print_prefix(self)?;
523 let args = args.iter().cloned().filter(|arg| {
525 UnpackedKind::Lifetime(_) => false,
530 if args.clone().next().is_some() {
531 self.generic_delimiters(|cx| cx.comma_sep(args))
538 impl PrettyPrinter<'tcx, 'tcx> for SymbolPrinter<'_, 'tcx> {
539 fn region_should_not_be_omitted(
541 _region: ty::Region<'_>,
547 mut elems: impl Iterator<Item = T>,
548 ) -> Result<Self, Self::Error>
549 where T: Print<'tcx, 'tcx, Self, Output = Self, Error = Self::Error>
551 if let Some(first) = elems.next() {
552 self = first.print(self)?;
554 self.write_str(",")?;
555 self = elem.print(self)?;
561 fn generic_delimiters(
563 f: impl FnOnce(Self) -> Result<Self, Self::Error>,
564 ) -> Result<Self, Self::Error> {
567 let kept_within_component =
568 mem::replace(&mut self.keep_within_component, true);
570 self.keep_within_component = kept_within_component;
578 impl fmt::Write for SymbolPrinter<'_, '_> {
579 fn write_str(&mut self, s: &str) -> fmt::Result {
580 // Name sanitation. LLVM will happily accept identifiers with weird names, but
582 // gas accepts the following characters in symbols: a-z, A-Z, 0-9, ., _, $
583 // NVPTX assembly has more strict naming rules than gas, so additionally, dots
584 // are replaced with '$' there.
587 if self.path.temp_buf.is_empty() {
589 'a'..='z' | 'A'..='Z' | '_' => {}
591 // Underscore-qualify anything that didn't start as an ident.
592 self.path.temp_buf.push('_');
597 // Escape these with $ sequences
598 '@' => self.path.temp_buf.push_str("$SP$"),
599 '*' => self.path.temp_buf.push_str("$BP$"),
600 '&' => self.path.temp_buf.push_str("$RF$"),
601 '<' => self.path.temp_buf.push_str("$LT$"),
602 '>' => self.path.temp_buf.push_str("$GT$"),
603 '(' => self.path.temp_buf.push_str("$LP$"),
604 ')' => self.path.temp_buf.push_str("$RP$"),
605 ',' => self.path.temp_buf.push_str("$C$"),
607 '-' | ':' | '.' if self.tcx.has_strict_asm_symbol_naming() => {
608 // NVPTX doesn't support these characters in symbol names.
609 self.path.temp_buf.push('$')
612 // '.' doesn't occur in types and functions, so reuse it
614 '-' | ':' => self.path.temp_buf.push('.'),
616 // These are legal symbols
617 'a'..='z' | 'A'..='Z' | '0'..='9' | '_' | '.' | '$' => self.path.temp_buf.push(c),
620 self.path.temp_buf.push('$');
621 for c in c.escape_unicode().skip(1) {
624 '}' => self.path.temp_buf.push('$'),
625 c => self.path.temp_buf.push(c),