rust/src/librustc_trans/back/symbol_names.rs

382 lines
15 KiB
Rust
Raw Normal View History

// Copyright 2016 The Rust Project Developers. See the COPYRIGHT
// file at the top-level directory of this distribution and at
// http://rust-lang.org/COPYRIGHT.
//
// Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
// http://www.apache.org/licenses/LICENSE-2.0> or the MIT license
// <LICENSE-MIT or http://opensource.org/licenses/MIT>, at your
// option. This file may not be copied, modified, or distributed
// except according to those terms.
//! The Rust Linkage Model and Symbol Names
//! =======================================
//!
//! The semantic model of Rust linkage is, broadly, that "there's no global
//! namespace" between crates. Our aim is to preserve the illusion of this
//! model despite the fact that it's not *quite* possible to implement on
//! modern linkers. We initially didn't use system linkers at all, but have
//! been convinced of their utility.
//!
//! There are a few issues to handle:
//!
//! - Linkers operate on a flat namespace, so we have to flatten names.
//! We do this using the C++ namespace-mangling technique. Foo::bar
//! symbols and such.
//!
//! - Symbols for distinct items with the same *name* need to get different
//! linkage-names. Examples of this are monomorphizations of functions or
//! items within anonymous scopes that end up having the same path.
//!
//! - Symbols in different crates but with same names "within" the crate need
//! to get different linkage-names.
//!
//! - Symbol names should be deterministic: Two consecutive runs of the
//! compiler over the same code base should produce the same symbol names for
//! the same items.
//!
//! - Symbol names should not depend on any global properties of the code base,
//! so that small modifications to the code base do not result in all symbols
//! changing. In previous versions of the compiler, symbol names incorporated
//! the SVH (Stable Version Hash) of the crate. This scheme turned out to be
//! infeasible when used in conjunction with incremental compilation because
//! small code changes would invalidate all symbols generated previously.
//!
//! - Even symbols from different versions of the same crate should be able to
//! live next to each other without conflict.
//!
//! In order to fulfill the above requirements the following scheme is used by
//! the compiler:
//!
//! The main tool for avoiding naming conflicts is the incorporation of a 64-bit
//! hash value into every exported symbol name. Anything that makes a difference
//! to the symbol being named, but does not show up in the regular path needs to
//! be fed into this hash:
//!
//! - Different monomorphizations of the same item have the same path but differ
//! in their concrete type parameters, so these parameters are part of the
//! data being digested for the symbol hash.
//!
//! - Rust allows items to be defined in anonymous scopes, such as in
//! `fn foo() { { fn bar() {} } { fn bar() {} } }`. Both `bar` functions have
//! the path `foo::bar`, since the anonymous scopes do not contribute to the
//! path of an item. The compiler already handles this case via so-called
//! disambiguating `DefPaths` which use indices to distinguish items with the
//! same name. The DefPaths of the functions above are thus `foo[0]::bar[0]`
//! and `foo[0]::bar[1]`. In order to incorporate this disambiguation
//! information into the symbol name too, these indices are fed into the
//! symbol hash, so that the above two symbols would end up with different
//! hash values.
//!
//! The two measures described above suffice to avoid intra-crate conflicts. In
//! order to also avoid inter-crate conflicts two more measures are taken:
//!
//! - The name of the crate containing the symbol is prepended to the symbol
//! name, i.e. symbols are "crate qualified". For example, a function `foo` in
//! module `bar` in crate `baz` would get a symbol name like
//! `baz::bar::foo::{hash}` instead of just `bar::foo::{hash}`. This avoids
//! simple conflicts between functions from different crates.
//!
//! - In order to be able to also use symbols from two versions of the same
//! crate (which naturally also have the same name), a stronger measure is
//! required: The compiler accepts an arbitrary "disambiguator" value via the
//! `-C metadata` commandline argument. This disambiguator is then fed into
//! the symbol hash of every exported item. Consequently, the symbols in two
//! identical crates but with different disambiguators are not in conflict
//! with each other. This facility is mainly intended to be used by build
//! tools like Cargo.
//!
//! A note on symbol name stability
//! -------------------------------
//! Previous versions of the compiler resorted to feeding NodeIds into the
//! symbol hash in order to disambiguate between items with the same path. The
//! current version of the name generation algorithm takes great care not to do
//! that, since NodeIds are notoriously unstable: A small change to the
//! code base will offset all NodeIds after the change and thus, much as using
//! the SVH in the hash, invalidate an unbounded number of symbol names. This
//! makes re-using previously compiled code for incremental compilation
//! virtually impossible. Thus, symbol hash generation exclusively relies on
//! DefPaths which are much more robust in the face of changes to the code base.
use common::SharedCrateContext;
use monomorphize::Instance;
use rustc::middle::weak_lang_items;
use rustc::hir::def_id::LOCAL_CRATE;
use rustc::hir::map as hir_map;
use rustc::ty::{self, Ty, TypeFoldable};
use rustc::ty::fold::TypeVisitor;
use rustc::ty::item_path::{self, ItemPathBuffer, RootMode};
use rustc::ty::subst::Substs;
2016-03-29 00:50:44 -05:00
use rustc::hir::map::definitions::{DefPath, DefPathData};
use rustc::util::common::record_time;
use syntax::attr;
use syntax::symbol::{Symbol, InternedString};
fn get_symbol_hash<'a, 'tcx>(scx: &SharedCrateContext<'a, 'tcx>,
// path to the item this name is for
def_path: &DefPath,
// type of the item, without any generic
// parameters substituted; this is
// included in the hash as a kind of
// safeguard.
item_type: Ty<'tcx>,
// values for generic type parameters,
// if any.
substs: Option<&'tcx Substs<'tcx>>)
-> String {
debug!("get_symbol_hash(def_path={:?}, parameters={:?})",
def_path, substs);
let tcx = scx.tcx();
let mut hasher = ty::util::TypeIdHasher::<u64>::new(tcx);
record_time(&tcx.sess.perf_stats.symbol_hash_time, || {
// the main symbol name is not necessarily unique; hash in the
// compiler's internal def-path, guaranteeing each symbol has a
// truly unique path
hasher.def_path(def_path);
// Include the main item-type. Note that, in this case, the
// assertions about `needs_subst` may not hold, but this item-type
// ought to be the same for every reference anyway.
assert!(!item_type.has_erasable_regions());
hasher.visit_ty(item_type);
// also include any type parameters (for generic items)
if let Some(substs) = substs {
assert!(!substs.has_erasable_regions());
assert!(!substs.needs_subst());
substs.visit_with(&mut hasher);
// If this is an instance of a generic function, we also hash in
// the ID of the instantiating crate. This avoids symbol conflicts
// in case the same instances is emitted in two crates of the same
// project.
if substs.types().next().is_some() {
hasher.hash(scx.tcx().crate_name.as_str());
hasher.hash(scx.sess().local_crate_disambiguator().as_str());
}
}
});
// 64 bits should be enough to avoid collisions.
format!("h{:016x}", hasher.finish())
}
impl<'a, 'tcx> Instance<'tcx> {
pub fn symbol_name(self, scx: &SharedCrateContext<'a, 'tcx>) -> String {
let Instance { def: def_id, substs } = self;
debug!("symbol_name(def_id={:?}, substs={:?})",
def_id, substs);
let node_id = scx.tcx().hir.as_local_node_id(def_id);
if let Some(id) = node_id {
if scx.sess().plugin_registrar_fn.get() == Some(id) {
let svh = &scx.link_meta().crate_hash;
let idx = def_id.index;
return scx.sess().generate_plugin_registrar_symbol(svh, idx);
}
rustc: Implement custom derive (macros 1.1) This commit is an implementation of [RFC 1681] which adds support to the compiler for first-class user-define custom `#[derive]` modes with a far more stable API than plugins have today. [RFC 1681]: https://github.com/rust-lang/rfcs/blob/master/text/1681-macros-1.1.md The main features added by this commit are: * A new `rustc-macro` crate-type. This crate type represents one which will provide custom `derive` implementations and perhaps eventually flower into the implementation of macros 2.0 as well. * A new `rustc_macro` crate in the standard distribution. This crate will provide the runtime interface between macro crates and the compiler. The API here is particularly conservative right now but has quite a bit of room to expand into any manner of APIs required by macro authors. * The ability to load new derive modes through the `#[macro_use]` annotations on other crates. All support added here is gated behind the `rustc_macro` feature gate, both for the library support (the `rustc_macro` crate) as well as the language features. There are a few minor differences from the implementation outlined in the RFC, such as the `rustc_macro` crate being available as a dylib and all symbols are `dlsym`'d directly instead of having a shim compiled. These should only affect the implementation, however, not the public interface. This commit also ended up touching a lot of code related to `#[derive]`, making a few notable changes: * Recognized derive attributes are no longer desugared to `derive_Foo`. Wasn't sure how to keep this behavior and *not* expose it to custom derive. * Derive attributes no longer have access to unstable features by default, they have to opt in on a granular level. * The `derive(Copy,Clone)` optimization is now done through another "obscure attribute" which is just intended to ferry along in the compiler that such an optimization is possible. The `derive(PartialEq,Eq)` optimization was also updated to do something similar. --- One part of this PR which needs to be improved before stabilizing are the errors and exact interfaces here. The error messages are relatively poor quality and there are surprising spects of this such as `#[derive(PartialEq, Eq, MyTrait)]` not working by default. The custom attributes added by the compiler end up becoming unstable again when going through a custom impl. Hopefully though this is enough to start allowing experimentation on crates.io! syntax-[breaking-change]
2016-08-22 19:07:11 -05:00
if scx.sess().derive_registrar_fn.get() == Some(id) {
let svh = &scx.link_meta().crate_hash;
let idx = def_id.index;
return scx.sess().generate_derive_registrar_symbol(svh, idx);
}
}
// FIXME(eddyb) Precompute a custom symbol name based on attributes.
let attrs = scx.tcx().get_attrs(def_id);
let is_foreign = if let Some(id) = node_id {
match scx.tcx().hir.get(id) {
hir_map::NodeForeignItem(_) => true,
_ => false
}
} else {
scx.sess().cstore.is_foreign_item(def_id)
};
if let Some(name) = weak_lang_items::link_name(&attrs) {
return name.to_string();
}
if is_foreign {
if let Some(name) = attr::first_attr_value_str_by_name(&attrs, "link_name") {
return name.to_string();
}
// Don't mangle foreign items.
return scx.tcx().item_name(def_id).as_str().to_string();
}
if let Some(name) = attr::find_export_name_attr(scx.sess().diagnostic(), &attrs) {
// Use provided name
return name.to_string();
}
if attr::contains_name(&attrs, "no_mangle") {
// Don't mangle
return scx.tcx().item_name(def_id).as_str().to_string();
}
let def_path = scx.tcx().def_path(def_id);
// We want to compute the "type" of this item. Unfortunately, some
// kinds of items (e.g., closures) don't have an entry in the
// item-type array. So walk back up the find the closest parent
// that DOES have an entry.
let mut ty_def_id = def_id;
let instance_ty;
loop {
let key = scx.tcx().def_key(ty_def_id);
match key.disambiguated_data.data {
DefPathData::TypeNs(_) |
DefPathData::ValueNs(_) => {
instance_ty = scx.tcx().item_type(ty_def_id);
break;
}
_ => {
// if we're making a symbol for something, there ought
// to be a value or type-def or something in there
// *somewhere*
ty_def_id.index = key.parent.unwrap_or_else(|| {
bug!("finding type for {:?}, encountered def-id {:?} with no \
parent", def_id, ty_def_id);
});
}
}
}
// Erase regions because they may not be deterministic when hashed
// and should not matter anyhow.
let instance_ty = scx.tcx().erase_regions(&instance_ty);
let hash = get_symbol_hash(scx, &def_path, instance_ty, Some(substs));
let mut buffer = SymbolPathBuffer {
names: Vec::with_capacity(def_path.data.len())
};
item_path::with_forced_absolute_paths(|| {
scx.tcx().push_item_path(&mut buffer, def_id);
});
mangle(buffer.names.into_iter(), &hash)
}
}
struct SymbolPathBuffer {
names: Vec<InternedString>,
}
impl ItemPathBuffer for SymbolPathBuffer {
fn root_mode(&self) -> &RootMode {
const ABSOLUTE: &'static RootMode = &RootMode::Absolute;
ABSOLUTE
}
fn push(&mut self, text: &str) {
self.names.push(Symbol::intern(text).as_str());
}
}
pub fn exported_name_from_type_and_prefix<'a, 'tcx>(scx: &SharedCrateContext<'a, 'tcx>,
t: Ty<'tcx>,
prefix: &str)
-> String {
let empty_def_path = DefPath {
data: vec![],
krate: LOCAL_CRATE,
};
let hash = get_symbol_hash(scx, &empty_def_path, t, None);
let path = [Symbol::intern(prefix).as_str()];
mangle(path.iter().cloned(), &hash)
2016-03-01 07:18:21 -06:00
}
// Name sanitation. LLVM will happily accept identifiers with weird names, but
// gas doesn't!
// gas accepts the following characters in symbols: a-z, A-Z, 0-9, ., _, $
pub fn sanitize(s: &str) -> String {
let mut result = String::new();
for c in s.chars() {
match c {
// Escape these with $ sequences
'@' => result.push_str("$SP$"),
'*' => result.push_str("$BP$"),
'&' => result.push_str("$RF$"),
'<' => result.push_str("$LT$"),
'>' => result.push_str("$GT$"),
'(' => result.push_str("$LP$"),
')' => result.push_str("$RP$"),
',' => result.push_str("$C$"),
// '.' doesn't occur in types and functions, so reuse it
// for ':' and '-'
'-' | ':' => result.push('.'),
// These are legal symbols
'a' ... 'z'
| 'A' ... 'Z'
| '0' ... '9'
| '_' | '.' | '$' => result.push(c),
_ => {
result.push('$');
for c in c.escape_unicode().skip(1) {
match c {
'{' => {},
'}' => result.push('$'),
c => result.push(c),
}
}
}
}
}
// Underscore-qualify anything that didn't start as an ident.
if !result.is_empty() &&
result.as_bytes()[0] != '_' as u8 &&
! (result.as_bytes()[0] as char).is_xid_start() {
return format!("_{}", &result[..]);
}
return result;
}
fn mangle<PI: Iterator<Item=InternedString>>(path: PI, hash: &str) -> String {
2016-03-01 07:18:21 -06:00
// Follow C++ namespace-mangling style, see
// http://en.wikipedia.org/wiki/Name_mangling for more info.
//
// It turns out that on macOS you can actually have arbitrary symbols in
2016-03-01 07:18:21 -06:00
// function names (at least when given to LLVM), but this is not possible
// when using unix's linker. Perhaps one day when we just use a linker from LLVM
// we won't need to do this name mangling. The problem with name mangling is
// that it seriously limits the available characters. For example we can't
// have things like &T in symbol names when one would theoretically
// want them for things like impls of traits on that type.
//
// To be able to work on all platforms and get *some* reasonable output, we
// use C++ name-mangling.
let mut n = String::from("_ZN"); // _Z == Begin name-sequence, N == nested
fn push(n: &mut String, s: &str) {
let sani = sanitize(s);
n.push_str(&format!("{}{}", sani.len(), sani));
}
// First, connect each component with <len, name> pairs.
for data in path {
push(&mut n, &data);
}
push(&mut n, hash);
2016-03-01 07:18:21 -06:00
n.push('E'); // End name-sequence.
n
}