rust/src/librustc_trans/partitioning.rs
2016-08-17 05:50:57 +03:00

637 lines
27 KiB
Rust

// 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.
//! Partitioning Codegen Units for Incremental Compilation
//! ======================================================
//!
//! The task of this module is to take the complete set of translation items of
//! a crate and produce a set of codegen units from it, where a codegen unit
//! is a named set of (translation-item, linkage) pairs. That is, this module
//! decides which translation item appears in which codegen units with which
//! linkage. The following paragraphs describe some of the background on the
//! partitioning scheme.
//!
//! The most important opportunity for saving on compilation time with
//! incremental compilation is to avoid re-translating and re-optimizing code.
//! Since the unit of translation and optimization for LLVM is "modules" or, how
//! we call them "codegen units", the particulars of how much time can be saved
//! by incremental compilation are tightly linked to how the output program is
//! partitioned into these codegen units prior to passing it to LLVM --
//! especially because we have to treat codegen units as opaque entities once
//! they are created: There is no way for us to incrementally update an existing
//! LLVM module and so we have to build any such module from scratch if it was
//! affected by some change in the source code.
//!
//! From that point of view it would make sense to maximize the number of
//! codegen units by, for example, putting each function into its own module.
//! That way only those modules would have to be re-compiled that were actually
//! affected by some change, minimizing the number of functions that could have
//! been re-used but just happened to be located in a module that is
//! re-compiled.
//!
//! However, since LLVM optimization does not work across module boundaries,
//! using such a highly granular partitioning would lead to very slow runtime
//! code since it would effectively prohibit inlining and other inter-procedure
//! optimizations. We want to avoid that as much as possible.
//!
//! Thus we end up with a trade-off: The bigger the codegen units, the better
//! LLVM's optimizer can do its work, but also the smaller the compilation time
//! reduction we get from incremental compilation.
//!
//! Ideally, we would create a partitioning such that there are few big codegen
//! units with few interdependencies between them. For now though, we use the
//! following heuristic to determine the partitioning:
//!
//! - There are two codegen units for every source-level module:
//! - One for "stable", that is non-generic, code
//! - One for more "volatile" code, i.e. monomorphized instances of functions
//! defined in that module
//! - Code for monomorphized instances of functions from external crates gets
//! placed into every codegen unit that uses that instance.
//!
//! In order to see why this heuristic makes sense, let's take a look at when a
//! codegen unit can get invalidated:
//!
//! 1. The most straightforward case is when the BODY of a function or global
//! changes. Then any codegen unit containing the code for that item has to be
//! re-compiled. Note that this includes all codegen units where the function
//! has been inlined.
//!
//! 2. The next case is when the SIGNATURE of a function or global changes. In
//! this case, all codegen units containing a REFERENCE to that item have to be
//! re-compiled. This is a superset of case 1.
//!
//! 3. The final and most subtle case is when a REFERENCE to a generic function
//! is added or removed somewhere. Even though the definition of the function
//! might be unchanged, a new REFERENCE might introduce a new monomorphized
//! instance of this function which has to be placed and compiled somewhere.
//! Conversely, when removing a REFERENCE, it might have been the last one with
//! that particular set of generic arguments and thus we have to remove it.
//!
//! From the above we see that just using one codegen unit per source-level
//! module is not such a good idea, since just adding a REFERENCE to some
//! generic item somewhere else would invalidate everything within the module
//! containing the generic item. The heuristic above reduces this detrimental
//! side-effect of references a little by at least not touching the non-generic
//! code of the module.
//!
//! As another optimization, monomorphized functions from external crates get
//! some special handling. Since we assume that the definition of such a
//! function changes rather infrequently compared to local items, we can just
//! instantiate external functions in every codegen unit where it is referenced
//! -- without having to fear that doing this will cause a lot of unnecessary
//! re-compilations. If such a reference is added or removed, the codegen unit
//! has to be re-translated anyway.
//! (Note that this only makes sense if external crates actually don't change
//! frequently. For certain multi-crate projects this might not be a valid
//! assumption).
//!
//! A Note on Inlining
//! ------------------
//! As briefly mentioned above, in order for LLVM to be able to inline a
//! function call, the body of the function has to be available in the LLVM
//! module where the call is made. This has a few consequences for partitioning:
//!
//! - The partitioning algorithm has to take care of placing functions into all
//! codegen units where they should be available for inlining. It also has to
//! decide on the correct linkage for these functions.
//!
//! - The partitioning algorithm has to know which functions are likely to get
//! inlined, so it can distribute function instantiations accordingly. Since
//! there is no way of knowing for sure which functions LLVM will decide to
//! inline in the end, we apply a heuristic here: Only functions marked with
//! #[inline] and (as stated above) functions from external crates are
//! considered for inlining by the partitioner. The current implementation
//! will not try to determine if a function is likely to be inlined by looking
//! at the functions definition.
//!
//! Note though that as a side-effect of creating a codegen units per
//! source-level module, functions from the same module will be available for
//! inlining, even when they are not marked #[inline].
use collector::InliningMap;
use llvm;
use monomorphize;
use rustc::dep_graph::{DepNode, WorkProductId};
use rustc::hir::def_id::DefId;
use rustc::hir::map::DefPathData;
use rustc::session::config::NUMBERED_CODEGEN_UNIT_MARKER;
use rustc::ty::TyCtxt;
use rustc::ty::item_path::characteristic_def_id_of_type;
use rustc::ty::subst;
use std::cmp::Ordering;
use std::hash::{Hash, Hasher, SipHasher};
use std::sync::Arc;
use symbol_map::SymbolMap;
use syntax::ast::NodeId;
use syntax::parse::token::{self, InternedString};
use trans_item::TransItem;
use util::nodemap::{FnvHashMap, FnvHashSet, NodeSet};
pub enum PartitioningStrategy {
/// Generate one codegen unit per source-level module.
PerModule,
/// Partition the whole crate into a fixed number of codegen units.
FixedUnitCount(usize)
}
pub struct CodegenUnit<'tcx> {
/// A name for this CGU. Incremental compilation requires that
/// name be unique amongst **all** crates. Therefore, it should
/// contain something unique to this crate (e.g., a module path)
/// as well as the crate name and disambiguator.
name: InternedString,
items: FnvHashMap<TransItem<'tcx>, llvm::Linkage>,
}
impl<'tcx> CodegenUnit<'tcx> {
pub fn new(name: InternedString,
items: FnvHashMap<TransItem<'tcx>, llvm::Linkage>)
-> Self {
CodegenUnit {
name: name,
items: items,
}
}
pub fn empty(name: InternedString) -> Self {
Self::new(name, FnvHashMap())
}
pub fn contains_item(&self, item: &TransItem<'tcx>) -> bool {
self.items.contains_key(item)
}
pub fn name(&self) -> &str {
&self.name
}
pub fn items(&self) -> &FnvHashMap<TransItem<'tcx>, llvm::Linkage> {
&self.items
}
pub fn work_product_id(&self) -> Arc<WorkProductId> {
Arc::new(WorkProductId(self.name().to_string()))
}
pub fn work_product_dep_node(&self) -> DepNode<DefId> {
DepNode::WorkProduct(self.work_product_id())
}
pub fn compute_symbol_name_hash(&self, tcx: TyCtxt, symbol_map: &SymbolMap) -> u64 {
let mut state = SipHasher::new();
let all_items = self.items_in_deterministic_order(tcx, symbol_map);
for (item, _) in all_items {
let symbol_name = symbol_map.get(item).unwrap();
symbol_name.hash(&mut state);
}
state.finish()
}
pub fn items_in_deterministic_order(&self,
tcx: TyCtxt,
symbol_map: &SymbolMap)
-> Vec<(TransItem<'tcx>, llvm::Linkage)> {
let mut items: Vec<(TransItem<'tcx>, llvm::Linkage)> =
self.items.iter().map(|(item, linkage)| (*item, *linkage)).collect();
// The codegen tests rely on items being process in the same order as
// they appear in the file, so for local items, we sort by node_id first
items.sort_by(|&(trans_item1, _), &(trans_item2, _)| {
let node_id1 = local_node_id(tcx, trans_item1);
let node_id2 = local_node_id(tcx, trans_item2);
match (node_id1, node_id2) {
(None, None) => {
let symbol_name1 = symbol_map.get(trans_item1).unwrap();
let symbol_name2 = symbol_map.get(trans_item2).unwrap();
symbol_name1.cmp(symbol_name2)
}
// In the following two cases we can avoid looking up the symbol
(None, Some(_)) => Ordering::Less,
(Some(_), None) => Ordering::Greater,
(Some(node_id1), Some(node_id2)) => {
let ordering = node_id1.cmp(&node_id2);
if ordering != Ordering::Equal {
return ordering;
}
let symbol_name1 = symbol_map.get(trans_item1).unwrap();
let symbol_name2 = symbol_map.get(trans_item2).unwrap();
symbol_name1.cmp(symbol_name2)
}
}
});
return items;
fn local_node_id(tcx: TyCtxt, trans_item: TransItem) -> Option<NodeId> {
match trans_item {
TransItem::Fn(instance) => {
tcx.map.as_local_node_id(instance.def)
}
TransItem::Static(node_id) => Some(node_id),
TransItem::DropGlue(_) => None,
}
}
}
}
// Anything we can't find a proper codegen unit for goes into this.
const FALLBACK_CODEGEN_UNIT: &'static str = "__rustc_fallback_codegen_unit";
pub fn partition<'a, 'tcx, I>(tcx: TyCtxt<'a, 'tcx, 'tcx>,
trans_items: I,
strategy: PartitioningStrategy,
inlining_map: &InliningMap<'tcx>,
reachable: &NodeSet)
-> Vec<CodegenUnit<'tcx>>
where I: Iterator<Item = TransItem<'tcx>>
{
if let PartitioningStrategy::FixedUnitCount(1) = strategy {
// If there is only a single codegen-unit, we can use a very simple
// scheme and don't have to bother with doing much analysis.
return vec![single_codegen_unit(tcx, trans_items, reachable)];
}
// In the first step, we place all regular translation items into their
// respective 'home' codegen unit. Regular translation items are all
// functions and statics defined in the local crate.
let mut initial_partitioning = place_root_translation_items(tcx,
trans_items,
reachable);
debug_dump(tcx, "INITIAL PARTITONING:", initial_partitioning.codegen_units.iter());
// If the partitioning should produce a fixed count of codegen units, merge
// until that count is reached.
if let PartitioningStrategy::FixedUnitCount(count) = strategy {
merge_codegen_units(&mut initial_partitioning, count, &tcx.crate_name[..]);
debug_dump(tcx, "POST MERGING:", initial_partitioning.codegen_units.iter());
}
// In the next step, we use the inlining map to determine which addtional
// translation items have to go into each codegen unit. These additional
// translation items can be drop-glue, functions from external crates, and
// local functions the definition of which is marked with #[inline].
let post_inlining = place_inlined_translation_items(initial_partitioning,
inlining_map);
debug_dump(tcx, "POST INLINING:", post_inlining.0.iter());
// Finally, sort by codegen unit name, so that we get deterministic results
let mut result = post_inlining.0;
result.sort_by(|cgu1, cgu2| {
(&cgu1.name[..]).cmp(&cgu2.name[..])
});
result
}
struct PreInliningPartitioning<'tcx> {
codegen_units: Vec<CodegenUnit<'tcx>>,
roots: FnvHashSet<TransItem<'tcx>>,
}
struct PostInliningPartitioning<'tcx>(Vec<CodegenUnit<'tcx>>);
fn place_root_translation_items<'a, 'tcx, I>(tcx: TyCtxt<'a, 'tcx, 'tcx>,
trans_items: I,
_reachable: &NodeSet)
-> PreInliningPartitioning<'tcx>
where I: Iterator<Item = TransItem<'tcx>>
{
let mut roots = FnvHashSet();
let mut codegen_units = FnvHashMap();
for trans_item in trans_items {
let is_root = !trans_item.is_instantiated_only_on_demand();
if is_root {
let characteristic_def_id = characteristic_def_id_of_trans_item(tcx, trans_item);
let is_volatile = trans_item.is_generic_fn();
let codegen_unit_name = match characteristic_def_id {
Some(def_id) => compute_codegen_unit_name(tcx, def_id, is_volatile),
None => InternedString::new(FALLBACK_CODEGEN_UNIT),
};
let make_codegen_unit = || {
CodegenUnit::empty(codegen_unit_name.clone())
};
let mut codegen_unit = codegen_units.entry(codegen_unit_name.clone())
.or_insert_with(make_codegen_unit);
let linkage = match trans_item.explicit_linkage(tcx) {
Some(explicit_linkage) => explicit_linkage,
None => {
match trans_item {
TransItem::Static(..) => llvm::ExternalLinkage,
TransItem::DropGlue(..) => unreachable!(),
// Is there any benefit to using ExternalLinkage?:
TransItem::Fn(ref instance) => {
if instance.substs.types.is_empty() {
// This is a non-generic functions, we always
// make it visible externally on the chance that
// it might be used in another codegen unit.
llvm::ExternalLinkage
} else {
// In the current setup, generic functions cannot
// be roots.
unreachable!()
}
}
}
}
};
codegen_unit.items.insert(trans_item, linkage);
roots.insert(trans_item);
}
}
// always ensure we have at least one CGU; otherwise, if we have a
// crate with just types (for example), we could wind up with no CGU
if codegen_units.is_empty() {
let codegen_unit_name = InternedString::new(FALLBACK_CODEGEN_UNIT);
codegen_units.entry(codegen_unit_name.clone())
.or_insert_with(|| CodegenUnit::empty(codegen_unit_name.clone()));
}
PreInliningPartitioning {
codegen_units: codegen_units.into_iter()
.map(|(_, codegen_unit)| codegen_unit)
.collect(),
roots: roots,
}
}
fn merge_codegen_units<'tcx>(initial_partitioning: &mut PreInliningPartitioning<'tcx>,
target_cgu_count: usize,
crate_name: &str) {
assert!(target_cgu_count >= 1);
let codegen_units = &mut initial_partitioning.codegen_units;
// Merge the two smallest codegen units until the target size is reached.
// Note that "size" is estimated here rather inaccurately as the number of
// translation items in a given unit. This could be improved on.
while codegen_units.len() > target_cgu_count {
// Sort small cgus to the back
codegen_units.sort_by_key(|cgu| -(cgu.items.len() as i64));
let smallest = codegen_units.pop().unwrap();
let second_smallest = codegen_units.last_mut().unwrap();
for (k, v) in smallest.items.into_iter() {
second_smallest.items.insert(k, v);
}
}
for (index, cgu) in codegen_units.iter_mut().enumerate() {
cgu.name = numbered_codegen_unit_name(crate_name, index);
}
// If the initial partitioning contained less than target_cgu_count to begin
// with, we won't have enough codegen units here, so add a empty units until
// we reach the target count
while codegen_units.len() < target_cgu_count {
let index = codegen_units.len();
codegen_units.push(
CodegenUnit::empty(numbered_codegen_unit_name(crate_name, index)));
}
}
fn place_inlined_translation_items<'tcx>(initial_partitioning: PreInliningPartitioning<'tcx>,
inlining_map: &InliningMap<'tcx>)
-> PostInliningPartitioning<'tcx> {
let mut new_partitioning = Vec::new();
for codegen_unit in &initial_partitioning.codegen_units[..] {
// Collect all items that need to be available in this codegen unit
let mut reachable = FnvHashSet();
for root in codegen_unit.items.keys() {
follow_inlining(*root, inlining_map, &mut reachable);
}
let mut new_codegen_unit =
CodegenUnit::empty(codegen_unit.name.clone());
// Add all translation items that are not already there
for trans_item in reachable {
if let Some(linkage) = codegen_unit.items.get(&trans_item) {
// This is a root, just copy it over
new_codegen_unit.items.insert(trans_item, *linkage);
} else if initial_partitioning.roots.contains(&trans_item) {
// This item will be instantiated in some other codegen unit,
// so we just add it here with AvailableExternallyLinkage
// FIXME(mw): I have not seen it happening yet but having
// available_externally here could potentially lead
// to the same problem with exception handling tables
// as in the case below.
new_codegen_unit.items.insert(trans_item,
llvm::AvailableExternallyLinkage);
} else if trans_item.is_from_extern_crate() && !trans_item.is_generic_fn() {
// FIXME(mw): It would be nice if we could mark these as
// `AvailableExternallyLinkage`, since they should have
// been instantiated in the extern crate. But this
// sometimes leads to crashes on Windows because LLVM
// does not handle exception handling table instantiation
// reliably in that case.
new_codegen_unit.items.insert(trans_item, llvm::InternalLinkage);
} else {
assert!(trans_item.is_instantiated_only_on_demand());
// We can't be sure if this will also be instantiated
// somewhere else, so we add an instance here with
// InternalLinkage so we don't get any conflicts.
new_codegen_unit.items.insert(trans_item,
llvm::InternalLinkage);
}
}
new_partitioning.push(new_codegen_unit);
}
return PostInliningPartitioning(new_partitioning);
fn follow_inlining<'tcx>(trans_item: TransItem<'tcx>,
inlining_map: &InliningMap<'tcx>,
visited: &mut FnvHashSet<TransItem<'tcx>>) {
if !visited.insert(trans_item) {
return;
}
inlining_map.with_inlining_candidates(trans_item, |target| {
follow_inlining(target, inlining_map, visited);
});
}
}
fn characteristic_def_id_of_trans_item<'a, 'tcx>(tcx: TyCtxt<'a, 'tcx, 'tcx>,
trans_item: TransItem<'tcx>)
-> Option<DefId> {
match trans_item {
TransItem::Fn(instance) => {
// If this is a method, we want to put it into the same module as
// its self-type. If the self-type does not provide a characteristic
// DefId, we use the location of the impl after all.
if tcx.trait_of_item(instance.def).is_some() {
let self_ty = *instance.substs.types.get(subst::TypeSpace, 0);
// This is an implementation of a trait method.
return characteristic_def_id_of_type(self_ty).or(Some(instance.def));
}
if let Some(impl_def_id) = tcx.impl_of_method(instance.def) {
// This is a method within an inherent impl, find out what the
// self-type is:
let impl_self_ty = tcx.lookup_item_type(impl_def_id).ty;
let impl_self_ty = tcx.erase_regions(&impl_self_ty);
let impl_self_ty = monomorphize::apply_param_substs(tcx,
instance.substs,
&impl_self_ty);
if let Some(def_id) = characteristic_def_id_of_type(impl_self_ty) {
return Some(def_id);
}
}
Some(instance.def)
}
TransItem::DropGlue(dg) => characteristic_def_id_of_type(dg.ty()),
TransItem::Static(node_id) => Some(tcx.map.local_def_id(node_id)),
}
}
fn compute_codegen_unit_name<'a, 'tcx>(tcx: TyCtxt<'a, 'tcx, 'tcx>,
def_id: DefId,
volatile: bool)
-> InternedString {
// Unfortunately we cannot just use the `ty::item_path` infrastructure here
// because we need paths to modules and the DefIds of those are not
// available anymore for external items.
let mut mod_path = String::with_capacity(64);
let def_path = tcx.def_path(def_id);
mod_path.push_str(&tcx.crate_name(def_path.krate));
for part in tcx.def_path(def_id)
.data
.iter()
.take_while(|part| {
match part.data {
DefPathData::Module(..) => true,
_ => false,
}
}) {
mod_path.push_str("-");
mod_path.push_str(&part.data.as_interned_str());
}
if volatile {
mod_path.push_str(".volatile");
}
return token::intern_and_get_ident(&mod_path[..]);
}
fn single_codegen_unit<'a, 'tcx, I>(tcx: TyCtxt<'a, 'tcx, 'tcx>,
trans_items: I,
reachable: &NodeSet)
-> CodegenUnit<'tcx>
where I: Iterator<Item = TransItem<'tcx>>
{
let mut items = FnvHashMap();
for trans_item in trans_items {
let linkage = trans_item.explicit_linkage(tcx).unwrap_or_else(|| {
match trans_item {
TransItem::Static(node_id) => {
if reachable.contains(&node_id) {
llvm::ExternalLinkage
} else {
llvm::PrivateLinkage
}
}
TransItem::DropGlue(_) => {
llvm::InternalLinkage
}
TransItem::Fn(instance) => {
if trans_item.is_generic_fn() {
// FIXME(mw): Assigning internal linkage to all
// monomorphizations is potentially a waste of space
// since monomorphizations could be shared between
// crates. The main reason for making them internal is
// a limitation in MingW's binutils that cannot deal
// with COFF object that have more than 2^15 sections,
// which is something that can happen for large programs
// when every function gets put into its own COMDAT
// section.
llvm::InternalLinkage
} else if trans_item.is_from_extern_crate() {
// FIXME(mw): It would be nice if we could mark these as
// `AvailableExternallyLinkage`, since they should have
// been instantiated in the extern crate. But this
// sometimes leads to crashes on Windows because LLVM
// does not handle exception handling table instantiation
// reliably in that case.
llvm::InternalLinkage
} else if reachable.contains(&tcx.map
.as_local_node_id(instance.def)
.unwrap()) {
llvm::ExternalLinkage
} else {
// Functions that are not visible outside this crate can
// be marked as internal.
llvm::InternalLinkage
}
}
}
});
items.insert(trans_item, linkage);
}
CodegenUnit::new(
numbered_codegen_unit_name(&tcx.crate_name[..], 0),
items)
}
fn numbered_codegen_unit_name(crate_name: &str, index: usize) -> InternedString {
token::intern_and_get_ident(&format!("{}{}{}",
crate_name,
NUMBERED_CODEGEN_UNIT_MARKER,
index)[..])
}
fn debug_dump<'a, 'b, 'tcx, I>(tcx: TyCtxt<'a, 'tcx, 'tcx>,
label: &str,
cgus: I)
where I: Iterator<Item=&'b CodegenUnit<'tcx>>,
'tcx: 'a + 'b
{
if cfg!(debug_assertions) {
debug!("{}", label);
for cgu in cgus {
debug!("CodegenUnit {}:", cgu.name);
for (trans_item, linkage) in &cgu.items {
debug!(" - {} [{:?}]", trans_item.to_string(tcx), linkage);
}
debug!("");
}
}
}