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::definitions::DefPathData;
94 use rustc::ich::NodeIdHashingMode;
95 use rustc::ty::print::{PrettyPrinter, PrintCx, Printer};
96 use rustc::ty::query::Providers;
97 use rustc::ty::subst::SubstsRef;
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::Symbol;
108 use std::fmt::{self, Write};
110 use std::mem::{self, discriminant};
112 pub fn provide(providers: &mut Providers<'_>) {
113 *providers = Providers {
121 fn get_symbol_hash<'a, 'tcx>(
122 tcx: TyCtxt<'a, 'tcx, 'tcx>,
124 // the DefId of the item this name is for
127 // instance this name will be for
128 instance: Instance<'tcx>,
130 // type of the item, without any generic
131 // parameters substituted; this is
132 // included in the hash as a kind of
136 // values for generic type parameters,
138 substs: SubstsRef<'tcx>,
141 "get_symbol_hash(def_id={:?}, parameters={:?})",
145 let mut hasher = StableHasher::<u64>::new();
146 let mut hcx = tcx.create_stable_hashing_context();
148 record_time(&tcx.sess.perf_stats.symbol_hash_time, || {
149 // the main symbol name is not necessarily unique; hash in the
150 // compiler's internal def-path, guaranteeing each symbol has a
152 tcx.def_path_hash(def_id).hash_stable(&mut hcx, &mut hasher);
154 // Include the main item-type. Note that, in this case, the
155 // assertions about `needs_subst` may not hold, but this item-type
156 // ought to be the same for every reference anyway.
157 assert!(!item_type.has_erasable_regions());
158 hcx.while_hashing_spans(false, |hcx| {
159 hcx.with_node_id_hashing_mode(NodeIdHashingMode::HashDefPath, |hcx| {
160 item_type.hash_stable(hcx, &mut hasher);
164 // If this is a function, we hash the signature as well.
165 // This is not *strictly* needed, but it may help in some
166 // situations, see the `run-make/a-b-a-linker-guard` test.
167 if let ty::FnDef(..) = item_type.sty {
168 item_type.fn_sig(tcx).hash_stable(&mut hcx, &mut hasher);
171 // also include any type parameters (for generic items)
172 assert!(!substs.has_erasable_regions());
173 assert!(!substs.needs_subst());
174 substs.hash_stable(&mut hcx, &mut hasher);
176 let is_generic = substs.non_erasable_generics().next().is_some();
177 let avoid_cross_crate_conflicts =
178 // If this is an instance of a generic function, we also hash in
179 // the ID of the instantiating crate. This avoids symbol conflicts
180 // in case the same instances is emitted in two crates of the same
184 // If we're dealing with an instance of a function that's inlined from
185 // another crate but we're marking it as globally shared to our
186 // compliation (aka we're not making an internal copy in each of our
187 // codegen units) then this symbol may become an exported (but hidden
188 // visibility) symbol. This means that multiple crates may do the same
189 // and we want to be sure to avoid any symbol conflicts here.
190 match MonoItem::Fn(instance).instantiation_mode(tcx) {
191 InstantiationMode::GloballyShared { may_conflict: true } => true,
195 if avoid_cross_crate_conflicts {
196 let instantiating_crate = if is_generic {
197 if !def_id.is_local() && tcx.sess.opts.share_generics() {
198 // If we are re-using a monomorphization from another crate,
199 // we have to compute the symbol hash accordingly.
200 let upstream_monomorphizations = tcx.upstream_monomorphizations_for(def_id);
202 upstream_monomorphizations
203 .and_then(|monos| monos.get(&substs).cloned())
204 .unwrap_or(LOCAL_CRATE)
212 (&tcx.original_crate_name(instantiating_crate).as_str()[..])
213 .hash_stable(&mut hcx, &mut hasher);
214 (&tcx.crate_disambiguator(instantiating_crate)).hash_stable(&mut hcx, &mut hasher);
217 // We want to avoid accidental collision between different types of instances.
218 // Especially, VtableShim may overlap with its original instance without this.
219 discriminant(&instance.def).hash_stable(&mut hcx, &mut hasher);
222 // 64 bits should be enough to avoid collisions.
226 fn def_symbol_name<'a, 'tcx>(tcx: TyCtxt<'a, 'tcx, 'tcx>, def_id: DefId) -> ty::SymbolName {
227 PrintCx::with(tcx, SymbolPath::new(tcx), |cx| {
228 cx.print_def_path(def_id, None, iter::empty())
234 fn symbol_name<'a, 'tcx>(tcx: TyCtxt<'a, 'tcx, 'tcx>, instance: Instance<'tcx>) -> ty::SymbolName {
236 name: Symbol::intern(&compute_symbol_name(tcx, instance)).as_interned_str(),
240 fn compute_symbol_name<'a, 'tcx>(tcx: TyCtxt<'a, 'tcx, 'tcx>, instance: Instance<'tcx>) -> String {
241 let def_id = instance.def_id();
242 let substs = instance.substs;
244 debug!("symbol_name(def_id={:?}, substs={:?})", def_id, substs);
246 let hir_id = tcx.hir().as_local_hir_id(def_id);
248 if def_id.is_local() {
249 if tcx.plugin_registrar_fn(LOCAL_CRATE) == Some(def_id) {
250 let disambiguator = tcx.sess.local_crate_disambiguator();
251 return tcx.sess.generate_plugin_registrar_symbol(disambiguator);
253 if tcx.proc_macro_decls_static(LOCAL_CRATE) == Some(def_id) {
254 let disambiguator = tcx.sess.local_crate_disambiguator();
255 return tcx.sess.generate_proc_macro_decls_symbol(disambiguator);
259 // FIXME(eddyb) Precompute a custom symbol name based on attributes.
260 let is_foreign = if let Some(id) = hir_id {
261 match tcx.hir().get_by_hir_id(id) {
262 Node::ForeignItem(_) => true,
266 tcx.is_foreign_item(def_id)
269 let attrs = tcx.codegen_fn_attrs(def_id);
271 if let Some(name) = attrs.link_name {
272 return name.to_string();
274 // Don't mangle foreign items.
275 return tcx.item_name(def_id).to_string();
278 if let Some(name) = &attrs.export_name {
280 return name.to_string();
283 if attrs.flags.contains(CodegenFnAttrFlags::NO_MANGLE) {
285 return tcx.item_name(def_id).to_string();
288 // We want to compute the "type" of this item. Unfortunately, some
289 // kinds of items (e.g., closures) don't have an entry in the
290 // item-type array. So walk back up the find the closest parent
291 // that DOES have an entry.
292 let mut ty_def_id = def_id;
295 let key = tcx.def_key(ty_def_id);
296 match key.disambiguated_data.data {
297 DefPathData::TypeNs(_) | DefPathData::ValueNs(_) => {
298 instance_ty = tcx.type_of(ty_def_id);
302 // if we're making a symbol for something, there ought
303 // to be a value or type-def or something in there
305 ty_def_id.index = key.parent.unwrap_or_else(|| {
307 "finding type for {:?}, encountered def-id {:?} with no \
317 // Erase regions because they may not be deterministic when hashed
318 // and should not matter anyhow.
319 let instance_ty = tcx.erase_regions(&instance_ty);
321 let hash = get_symbol_hash(tcx, def_id, instance, instance_ty, substs);
323 let mut buf = SymbolPath::from_interned(tcx.def_symbol_name(def_id), tcx);
325 if instance.is_vtable_shim() {
326 let _ = buf.write_str("{{vtable-shim}}");
332 // Follow C++ namespace-mangling style, see
333 // http://en.wikipedia.org/wiki/Name_mangling for more info.
335 // It turns out that on macOS you can actually have arbitrary symbols in
336 // function names (at least when given to LLVM), but this is not possible
337 // when using unix's linker. Perhaps one day when we just use a linker from LLVM
338 // we won't need to do this name mangling. The problem with name mangling is
339 // that it seriously limits the available characters. For example we can't
340 // have things like &T in symbol names when one would theoretically
341 // want them for things like impls of traits on that type.
343 // To be able to work on all platforms and get *some* reasonable output, we
344 // use C++ name-mangling.
351 // When `true`, `finalize_pending_component` isn't used.
352 // This is needed when recursing into `path_qualified`,
353 // or `path_generic_args`, as any nested paths are
354 // logically within one component.
355 keep_within_component: bool,
359 fn new(tcx: TyCtxt<'_, '_, '_>) -> Self {
360 let mut result = SymbolPath {
361 result: String::with_capacity(64),
362 temp_buf: String::with_capacity(16),
363 strict_naming: tcx.has_strict_asm_symbol_naming(),
364 keep_within_component: false,
366 result.result.push_str("_ZN"); // _Z == Begin name-sequence, N == nested
370 fn from_interned(symbol: ty::SymbolName, tcx: TyCtxt<'_, '_, '_>) -> Self {
371 let mut result = SymbolPath {
372 result: String::with_capacity(64),
373 temp_buf: String::with_capacity(16),
374 strict_naming: tcx.has_strict_asm_symbol_naming(),
375 keep_within_component: false,
377 result.result.push_str(&symbol.as_str());
381 fn into_interned(mut self) -> ty::SymbolName {
382 self.finalize_pending_component();
384 name: Symbol::intern(&self.result).as_interned_str(),
388 fn finalize_pending_component(&mut self) {
389 if !self.temp_buf.is_empty() {
390 let _ = write!(self.result, "{}{}", self.temp_buf.len(), self.temp_buf);
391 self.temp_buf.clear();
395 fn finish(mut self, hash: u64) -> String {
396 self.finalize_pending_component();
397 // E = end name-sequence
398 let _ = write!(self.result, "17h{:016x}E", hash);
403 // HACK(eddyb) this relies on using the `fmt` interface to get
404 // `PrettyPrinter` aka pretty printing of e.g. types in paths,
405 // symbol names should have their own printing machinery.
407 impl Printer for SymbolPath {
408 type Error = fmt::Error;
415 self: PrintCx<'_, '_, '_, Self>,
416 _region: ty::Region<'_>,
417 ) -> Result<Self::Region, Self::Error> {
422 self: PrintCx<'_, '_, 'tcx, Self>,
424 ) -> Result<Self::Type, Self::Error> {
425 self.pretty_print_type(ty)
429 mut self: PrintCx<'_, '_, '_, Self>,
431 ) -> Result<Self::Path, Self::Error> {
432 self.printer.write_str(&self.tcx.original_crate_name(cnum).as_str())?;
436 self: PrintCx<'_, '_, 'tcx, Self>,
438 trait_ref: Option<ty::TraitRef<'tcx>>,
439 ) -> Result<Self::Path, Self::Error> {
440 self.pretty_path_qualified(self_ty, trait_ref)
443 fn path_append_impl<'gcx, 'tcx>(
444 self: PrintCx<'_, 'gcx, 'tcx, Self>,
445 print_prefix: impl FnOnce(
446 PrintCx<'_, 'gcx, 'tcx, Self>,
447 ) -> Result<Self::Path, Self::Error>,
449 trait_ref: Option<ty::TraitRef<'tcx>>,
450 ) -> Result<Self::Path, Self::Error> {
451 self.pretty_path_append_impl(
452 |cx| cx.path_append(print_prefix, ""),
457 fn path_append<'gcx, 'tcx>(
458 self: PrintCx<'_, 'gcx, 'tcx, Self>,
459 print_prefix: impl FnOnce(
460 PrintCx<'_, 'gcx, 'tcx, Self>,
461 ) -> Result<Self::Path, Self::Error>,
463 ) -> Result<Self::Path, Self::Error> {
464 let mut path = print_prefix(self)?;
466 if path.keep_within_component {
467 // HACK(eddyb) print the path similarly to how `FmtPrinter` prints it.
468 path.write_str("::")?;
470 path.finalize_pending_component();
473 path.write_str(text)?;
476 fn path_generic_args<'gcx, 'tcx>(
477 self: PrintCx<'_, 'gcx, 'tcx, Self>,
478 print_prefix: impl FnOnce(
479 PrintCx<'_, 'gcx, 'tcx, Self>,
480 ) -> Result<Self::Path, Self::Error>,
481 params: &[ty::GenericParamDef],
482 substs: SubstsRef<'tcx>,
483 projections: impl Iterator<Item = ty::ExistentialProjection<'tcx>>,
484 ) -> Result<Self::Path, Self::Error> {
485 self.pretty_path_generic_args(print_prefix, params, substs, projections)
489 impl PrettyPrinter for SymbolPath {
490 fn print_region_outputs_anything(
491 self: &PrintCx<'_, '_, '_, Self>,
492 _region: ty::Region<'_>,
497 fn generic_delimiters<'gcx, 'tcx>(
498 mut self: PrintCx<'_, 'gcx, 'tcx, Self>,
499 f: impl FnOnce(PrintCx<'_, 'gcx, 'tcx, Self>) -> Result<Self, Self::Error>,
500 ) -> Result<Self, Self::Error> {
501 write!(self.printer, "<")?;
503 let kept_within_component =
504 mem::replace(&mut self.printer.keep_within_component, true);
505 let mut path = f(self)?;
506 path.keep_within_component = kept_within_component;
514 impl fmt::Write for SymbolPath {
515 fn write_str(&mut self, s: &str) -> fmt::Result {
516 // Name sanitation. LLVM will happily accept identifiers with weird names, but
518 // gas accepts the following characters in symbols: a-z, A-Z, 0-9, ., _, $
519 // NVPTX assembly has more strict naming rules than gas, so additionally, dots
520 // are replaced with '$' there.
523 if self.temp_buf.is_empty() {
525 'a'..='z' | 'A'..='Z' | '_' => {}
527 // Underscore-qualify anything that didn't start as an ident.
528 self.temp_buf.push('_');
533 // Escape these with $ sequences
534 '@' => self.temp_buf.push_str("$SP$"),
535 '*' => self.temp_buf.push_str("$BP$"),
536 '&' => self.temp_buf.push_str("$RF$"),
537 '<' => self.temp_buf.push_str("$LT$"),
538 '>' => self.temp_buf.push_str("$GT$"),
539 '(' => self.temp_buf.push_str("$LP$"),
540 ')' => self.temp_buf.push_str("$RP$"),
541 ',' => self.temp_buf.push_str("$C$"),
543 '-' | ':' | '.' if self.strict_naming => {
544 // NVPTX doesn't support these characters in symbol names.
545 self.temp_buf.push('$')
548 // '.' doesn't occur in types and functions, so reuse it
550 '-' | ':' => self.temp_buf.push('.'),
552 // These are legal symbols
553 'a'..='z' | 'A'..='Z' | '0'..='9' | '_' | '.' | '$' => self.temp_buf.push(c),
556 self.temp_buf.push('$');
557 for c in c.escape_unicode().skip(1) {
560 '}' => self.temp_buf.push('$'),
561 c => self.temp_buf.push(c),