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::Namespace;
91 use rustc::hir::def_id::{CrateNum, DefId, LOCAL_CRATE};
93 use rustc::hir::CodegenFnAttrFlags;
94 use rustc::hir::map::definitions::DefPathData;
95 use rustc::ich::NodeIdHashingMode;
96 use rustc::ty::item_path::{self, ItemPathPrinter};
97 use rustc::ty::print::PrintCx;
98 use rustc::ty::query::Providers;
99 use rustc::ty::subst::SubstsRef;
100 use rustc::ty::{self, Ty, TyCtxt, TypeFoldable};
101 use rustc::util::common::record_time;
102 use rustc_data_structures::stable_hasher::{HashStable, StableHasher};
103 use rustc_mir::monomorphize::item::{InstantiationMode, MonoItem, MonoItemExt};
104 use rustc_mir::monomorphize::Instance;
106 use syntax_pos::symbol::Symbol;
111 use std::mem::discriminant;
113 pub fn provide(providers: &mut Providers<'_>) {
114 *providers = Providers {
122 fn get_symbol_hash<'a, 'tcx>(
123 tcx: TyCtxt<'a, 'tcx, 'tcx>,
125 // the DefId of the item this name is for
128 // instance this name will be for
129 instance: Instance<'tcx>,
131 // type of the item, without any generic
132 // parameters substituted; this is
133 // included in the hash as a kind of
137 // values for generic type parameters,
139 substs: SubstsRef<'tcx>,
142 "get_symbol_hash(def_id={:?}, parameters={:?})",
146 let mut hasher = StableHasher::<u64>::new();
147 let mut hcx = tcx.create_stable_hashing_context();
149 record_time(&tcx.sess.perf_stats.symbol_hash_time, || {
150 // the main symbol name is not necessarily unique; hash in the
151 // compiler's internal def-path, guaranteeing each symbol has a
153 tcx.def_path_hash(def_id).hash_stable(&mut hcx, &mut hasher);
155 // Include the main item-type. Note that, in this case, the
156 // assertions about `needs_subst` may not hold, but this item-type
157 // ought to be the same for every reference anyway.
158 assert!(!item_type.has_erasable_regions());
159 hcx.while_hashing_spans(false, |hcx| {
160 hcx.with_node_id_hashing_mode(NodeIdHashingMode::HashDefPath, |hcx| {
161 item_type.hash_stable(hcx, &mut hasher);
165 // If this is a function, we hash the signature as well.
166 // This is not *strictly* needed, but it may help in some
167 // situations, see the `run-make/a-b-a-linker-guard` test.
168 if let ty::FnDef(..) = item_type.sty {
169 item_type.fn_sig(tcx).hash_stable(&mut hcx, &mut hasher);
172 // also include any type parameters (for generic items)
173 assert!(!substs.has_erasable_regions());
174 assert!(!substs.needs_subst());
175 substs.hash_stable(&mut hcx, &mut hasher);
177 let is_generic = substs.non_erasable_generics().next().is_some();
178 let avoid_cross_crate_conflicts =
179 // If this is an instance of a generic function, we also hash in
180 // the ID of the instantiating crate. This avoids symbol conflicts
181 // in case the same instances is emitted in two crates of the same
185 // If we're dealing with an instance of a function that's inlined from
186 // another crate but we're marking it as globally shared to our
187 // compliation (aka we're not making an internal copy in each of our
188 // codegen units) then this symbol may become an exported (but hidden
189 // visibility) symbol. This means that multiple crates may do the same
190 // and we want to be sure to avoid any symbol conflicts here.
191 match MonoItem::Fn(instance).instantiation_mode(tcx) {
192 InstantiationMode::GloballyShared { may_conflict: true } => true,
196 if avoid_cross_crate_conflicts {
197 let instantiating_crate = if is_generic {
198 if !def_id.is_local() && tcx.sess.opts.share_generics() {
199 // If we are re-using a monomorphization from another crate,
200 // we have to compute the symbol hash accordingly.
201 let upstream_monomorphizations = tcx.upstream_monomorphizations_for(def_id);
203 upstream_monomorphizations
204 .and_then(|monos| monos.get(&substs).cloned())
205 .unwrap_or(LOCAL_CRATE)
213 (&tcx.original_crate_name(instantiating_crate).as_str()[..])
214 .hash_stable(&mut hcx, &mut hasher);
215 (&tcx.crate_disambiguator(instantiating_crate)).hash_stable(&mut hcx, &mut hasher);
218 // We want to avoid accidental collision between different types of instances.
219 // Especially, VtableShim may overlap with its original instance without this.
220 discriminant(&instance.def).hash_stable(&mut hcx, &mut hasher);
223 // 64 bits should be enough to avoid collisions.
227 fn def_symbol_name<'a, 'tcx>(tcx: TyCtxt<'a, 'tcx, 'tcx>, def_id: DefId) -> ty::SymbolName {
228 item_path::with_forced_absolute_paths(|| {
229 PrintCx::new(tcx, SymbolPathPrinter)
230 .print_item_path(def_id, None, Namespace::ValueNS)
235 fn symbol_name<'a, 'tcx>(tcx: TyCtxt<'a, 'tcx, 'tcx>, instance: Instance<'tcx>) -> ty::SymbolName {
237 name: Symbol::intern(&compute_symbol_name(tcx, instance)).as_interned_str(),
241 fn compute_symbol_name<'a, 'tcx>(tcx: TyCtxt<'a, 'tcx, 'tcx>, instance: Instance<'tcx>) -> String {
242 let def_id = instance.def_id();
243 let substs = instance.substs;
245 debug!("symbol_name(def_id={:?}, substs={:?})", def_id, substs);
247 let hir_id = tcx.hir().as_local_hir_id(def_id);
249 if def_id.is_local() {
250 if tcx.plugin_registrar_fn(LOCAL_CRATE) == Some(def_id) {
251 let disambiguator = tcx.sess.local_crate_disambiguator();
252 return tcx.sess.generate_plugin_registrar_symbol(disambiguator);
254 if tcx.proc_macro_decls_static(LOCAL_CRATE) == Some(def_id) {
255 let disambiguator = tcx.sess.local_crate_disambiguator();
256 return tcx.sess.generate_proc_macro_decls_symbol(disambiguator);
260 // FIXME(eddyb) Precompute a custom symbol name based on attributes.
261 let is_foreign = if let Some(id) = hir_id {
262 match tcx.hir().get_by_hir_id(id) {
263 Node::ForeignItem(_) => true,
267 tcx.is_foreign_item(def_id)
270 let attrs = tcx.codegen_fn_attrs(def_id);
272 if let Some(name) = attrs.link_name {
273 return name.to_string();
275 // Don't mangle foreign items.
276 return tcx.item_name(def_id).to_string();
279 if let Some(name) = &attrs.export_name {
281 return name.to_string();
284 if attrs.flags.contains(CodegenFnAttrFlags::NO_MANGLE) {
286 return tcx.item_name(def_id).to_string();
289 // We want to compute the "type" of this item. Unfortunately, some
290 // kinds of items (e.g., closures) don't have an entry in the
291 // item-type array. So walk back up the find the closest parent
292 // that DOES have an entry.
293 let mut ty_def_id = def_id;
296 let key = tcx.def_key(ty_def_id);
297 match key.disambiguated_data.data {
298 DefPathData::TypeNs(_) | DefPathData::ValueNs(_) => {
299 instance_ty = tcx.type_of(ty_def_id);
303 // if we're making a symbol for something, there ought
304 // to be a value or type-def or something in there
306 ty_def_id.index = key.parent.unwrap_or_else(|| {
308 "finding type for {:?}, encountered def-id {:?} with no \
318 // Erase regions because they may not be deterministic when hashed
319 // and should not matter anyhow.
320 let instance_ty = tcx.erase_regions(&instance_ty);
322 let hash = get_symbol_hash(tcx, def_id, instance, instance_ty, substs);
324 let mut buf = SymbolPath::from_interned(tcx.def_symbol_name(def_id), tcx);
326 if instance.is_vtable_shim() {
327 buf.push("{{vtable-shim}}");
333 // Follow C++ namespace-mangling style, see
334 // http://en.wikipedia.org/wiki/Name_mangling for more info.
336 // It turns out that on macOS you can actually have arbitrary symbols in
337 // function names (at least when given to LLVM), but this is not possible
338 // when using unix's linker. Perhaps one day when we just use a linker from LLVM
339 // we won't need to do this name mangling. The problem with name mangling is
340 // that it seriously limits the available characters. For example we can't
341 // have things like &T in symbol names when one would theoretically
342 // want them for things like impls of traits on that type.
344 // To be able to work on all platforms and get *some* reasonable output, we
345 // use C++ name-mangling.
354 fn new(tcx: TyCtxt<'_, '_, '_>) -> Self {
355 let mut result = SymbolPath {
356 result: String::with_capacity(64),
357 temp_buf: String::with_capacity(16),
358 strict_naming: tcx.has_strict_asm_symbol_naming(),
360 result.result.push_str("_ZN"); // _Z == Begin name-sequence, N == nested
364 fn from_interned(symbol: ty::SymbolName, tcx: TyCtxt<'_, '_, '_>) -> Self {
365 let mut result = SymbolPath {
366 result: String::with_capacity(64),
367 temp_buf: String::with_capacity(16),
368 strict_naming: tcx.has_strict_asm_symbol_naming(),
370 result.result.push_str(&symbol.as_str());
374 fn into_interned(self) -> ty::SymbolName {
376 name: Symbol::intern(&self.result).as_interned_str(),
380 fn push(&mut self, text: &str) {
381 self.temp_buf.clear();
382 let need_underscore = sanitize(&mut self.temp_buf, text, self.strict_naming);
386 self.temp_buf.len() + (need_underscore as usize)
389 self.result.push('_');
391 self.result.push_str(&self.temp_buf);
394 fn finish(mut self, hash: u64) -> String {
395 // E = end name-sequence
396 let _ = write!(self.result, "17h{:016x}E", hash);
401 struct SymbolPathPrinter;
403 impl ItemPathPrinter for SymbolPathPrinter {
404 type Path = SymbolPath;
406 fn path_crate(self: &mut PrintCx<'_, '_, '_, Self>, cnum: CrateNum) -> Self::Path {
407 let mut path = SymbolPath::new(self.tcx);
408 path.push(&self.tcx.original_crate_name(cnum).as_str());
411 fn path_impl(self: &mut PrintCx<'_, '_, '_, Self>, text: &str) -> Self::Path {
412 let mut path = SymbolPath::new(self.tcx);
417 self: &mut PrintCx<'_, '_, '_, Self>,
418 mut path: Self::Path,
426 // Name sanitation. LLVM will happily accept identifiers with weird names, but
428 // gas accepts the following characters in symbols: a-z, A-Z, 0-9, ., _, $
429 // NVPTX assembly has more strict naming rules than gas, so additionally, dots
430 // are replaced with '$' there.
432 // returns true if an underscore must be added at the start
433 fn sanitize(result: &mut String, s: &str, strict_naming: bool) -> bool {
436 // Escape these with $ sequences
437 '@' => result.push_str("$SP$"),
438 '*' => result.push_str("$BP$"),
439 '&' => result.push_str("$RF$"),
440 '<' => result.push_str("$LT$"),
441 '>' => result.push_str("$GT$"),
442 '(' => result.push_str("$LP$"),
443 ')' => result.push_str("$RP$"),
444 ',' => result.push_str("$C$"),
446 '-' | ':' | '.' if strict_naming => {
447 // NVPTX doesn't support these characters in symbol names.
451 // '.' doesn't occur in types and functions, so reuse it
453 '-' | ':' => result.push('.'),
455 // These are legal symbols
456 'a'..='z' | 'A'..='Z' | '0'..='9' | '_' | '.' | '$' => result.push(c),
460 for c in c.escape_unicode().skip(1) {
463 '}' => result.push('$'),
471 // Underscore-qualify anything that didn't start as an ident.
472 !result.is_empty() && result.as_bytes()[0] != '_' as u8
473 && !(result.as_bytes()[0] as char).is_xid_start()