1 //! Partitioning Codegen Units for Incremental Compilation
2 //! ======================================================
4 //! The task of this module is to take the complete set of monomorphizations of
5 //! a crate and produce a set of codegen units from it, where a codegen unit
6 //! is a named set of (mono-item, linkage) pairs. That is, this module
7 //! decides which monomorphization appears in which codegen units with which
8 //! linkage. The following paragraphs describe some of the background on the
9 //! partitioning scheme.
11 //! The most important opportunity for saving on compilation time with
12 //! incremental compilation is to avoid re-codegenning and re-optimizing code.
13 //! Since the unit of codegen and optimization for LLVM is "modules" or, how
14 //! we call them "codegen units", the particulars of how much time can be saved
15 //! by incremental compilation are tightly linked to how the output program is
16 //! partitioned into these codegen units prior to passing it to LLVM --
17 //! especially because we have to treat codegen units as opaque entities once
18 //! they are created: There is no way for us to incrementally update an existing
19 //! LLVM module and so we have to build any such module from scratch if it was
20 //! affected by some change in the source code.
22 //! From that point of view it would make sense to maximize the number of
23 //! codegen units by, for example, putting each function into its own module.
24 //! That way only those modules would have to be re-compiled that were actually
25 //! affected by some change, minimizing the number of functions that could have
26 //! been re-used but just happened to be located in a module that is
29 //! However, since LLVM optimization does not work across module boundaries,
30 //! using such a highly granular partitioning would lead to very slow runtime
31 //! code since it would effectively prohibit inlining and other inter-procedure
32 //! optimizations. We want to avoid that as much as possible.
34 //! Thus we end up with a trade-off: The bigger the codegen units, the better
35 //! LLVM's optimizer can do its work, but also the smaller the compilation time
36 //! reduction we get from incremental compilation.
38 //! Ideally, we would create a partitioning such that there are few big codegen
39 //! units with few interdependencies between them. For now though, we use the
40 //! following heuristic to determine the partitioning:
42 //! - There are two codegen units for every source-level module:
43 //! - One for "stable", that is non-generic, code
44 //! - One for more "volatile" code, i.e., monomorphized instances of functions
45 //! defined in that module
47 //! In order to see why this heuristic makes sense, let's take a look at when a
48 //! codegen unit can get invalidated:
50 //! 1. The most straightforward case is when the BODY of a function or global
51 //! changes. Then any codegen unit containing the code for that item has to be
52 //! re-compiled. Note that this includes all codegen units where the function
55 //! 2. The next case is when the SIGNATURE of a function or global changes. In
56 //! this case, all codegen units containing a REFERENCE to that item have to be
57 //! re-compiled. This is a superset of case 1.
59 //! 3. The final and most subtle case is when a REFERENCE to a generic function
60 //! is added or removed somewhere. Even though the definition of the function
61 //! might be unchanged, a new REFERENCE might introduce a new monomorphized
62 //! instance of this function which has to be placed and compiled somewhere.
63 //! Conversely, when removing a REFERENCE, it might have been the last one with
64 //! that particular set of generic arguments and thus we have to remove it.
66 //! From the above we see that just using one codegen unit per source-level
67 //! module is not such a good idea, since just adding a REFERENCE to some
68 //! generic item somewhere else would invalidate everything within the module
69 //! containing the generic item. The heuristic above reduces this detrimental
70 //! side-effect of references a little by at least not touching the non-generic
71 //! code of the module.
73 //! A Note on Inlining
74 //! ------------------
75 //! As briefly mentioned above, in order for LLVM to be able to inline a
76 //! function call, the body of the function has to be available in the LLVM
77 //! module where the call is made. This has a few consequences for partitioning:
79 //! - The partitioning algorithm has to take care of placing functions into all
80 //! codegen units where they should be available for inlining. It also has to
81 //! decide on the correct linkage for these functions.
83 //! - The partitioning algorithm has to know which functions are likely to get
84 //! inlined, so it can distribute function instantiations accordingly. Since
85 //! there is no way of knowing for sure which functions LLVM will decide to
86 //! inline in the end, we apply a heuristic here: Only functions marked with
87 //! `#[inline]` are considered for inlining by the partitioner. The current
88 //! implementation will not try to determine if a function is likely to be
89 //! inlined by looking at the functions definition.
91 //! Note though that as a side-effect of creating a codegen units per
92 //! source-level module, functions from the same module will be available for
93 //! inlining, even when they are not marked `#[inline]`.
99 use std::fs::{self, File};
100 use std::io::{BufWriter, Write};
101 use std::path::{Path, PathBuf};
103 use rustc_data_structures::fx::{FxHashMap, FxHashSet};
104 use rustc_data_structures::sync;
105 use rustc_hir::def_id::{DefIdSet, LOCAL_CRATE};
106 use rustc_middle::mir;
107 use rustc_middle::mir::mono::MonoItem;
108 use rustc_middle::mir::mono::{CodegenUnit, Linkage};
109 use rustc_middle::ty::print::with_no_trimmed_paths;
110 use rustc_middle::ty::query::Providers;
111 use rustc_middle::ty::TyCtxt;
112 use rustc_session::config::{DumpMonoStatsFormat, SwitchWithOptPath};
113 use rustc_span::symbol::Symbol;
115 use crate::collector::InliningMap;
116 use crate::collector::{self, MonoItemCollectionMode};
118 CouldntDumpMonoStats, SymbolAlreadyDefined, UnknownCguCollectionMode, UnknownPartitionStrategy,
121 pub struct PartitioningCx<'a, 'tcx> {
123 target_cgu_count: usize,
124 inlining_map: &'a InliningMap<'tcx>,
127 trait Partitioner<'tcx> {
128 fn place_root_mono_items(
130 cx: &PartitioningCx<'_, 'tcx>,
131 mono_items: &mut dyn Iterator<Item = MonoItem<'tcx>>,
132 ) -> PreInliningPartitioning<'tcx>;
134 fn merge_codegen_units(
136 cx: &PartitioningCx<'_, 'tcx>,
137 initial_partitioning: &mut PreInliningPartitioning<'tcx>,
140 fn place_inlined_mono_items(
142 cx: &PartitioningCx<'_, 'tcx>,
143 initial_partitioning: PreInliningPartitioning<'tcx>,
144 ) -> PostInliningPartitioning<'tcx>;
146 fn internalize_symbols(
148 cx: &PartitioningCx<'_, 'tcx>,
149 partitioning: &mut PostInliningPartitioning<'tcx>,
153 fn get_partitioner<'tcx>(tcx: TyCtxt<'tcx>) -> Box<dyn Partitioner<'tcx>> {
154 let strategy = match &tcx.sess.opts.unstable_opts.cgu_partitioning_strategy {
160 "default" => Box::new(default::DefaultPartitioning),
162 tcx.sess.emit_fatal(UnknownPartitionStrategy);
167 pub fn partition<'tcx>(
169 mono_items: &mut dyn Iterator<Item = MonoItem<'tcx>>,
170 max_cgu_count: usize,
171 inlining_map: &InliningMap<'tcx>,
172 ) -> Vec<CodegenUnit<'tcx>> {
173 let _prof_timer = tcx.prof.generic_activity("cgu_partitioning");
175 let mut partitioner = get_partitioner(tcx);
176 let cx = &PartitioningCx { tcx, target_cgu_count: max_cgu_count, inlining_map };
177 // In the first step, we place all regular monomorphizations into their
178 // respective 'home' codegen unit. Regular monomorphizations are all
179 // functions and statics defined in the local crate.
180 let mut initial_partitioning = {
181 let _prof_timer = tcx.prof.generic_activity("cgu_partitioning_place_roots");
182 partitioner.place_root_mono_items(cx, mono_items)
185 initial_partitioning.codegen_units.iter_mut().for_each(|cgu| cgu.create_size_estimate(tcx));
187 debug_dump(tcx, "INITIAL PARTITIONING:", initial_partitioning.codegen_units.iter());
189 // Merge until we have at most `max_cgu_count` codegen units.
191 let _prof_timer = tcx.prof.generic_activity("cgu_partitioning_merge_cgus");
192 partitioner.merge_codegen_units(cx, &mut initial_partitioning);
193 debug_dump(tcx, "POST MERGING:", initial_partitioning.codegen_units.iter());
196 // In the next step, we use the inlining map to determine which additional
197 // monomorphizations have to go into each codegen unit. These additional
198 // monomorphizations can be drop-glue, functions from external crates, and
199 // local functions the definition of which is marked with `#[inline]`.
200 let mut post_inlining = {
201 let _prof_timer = tcx.prof.generic_activity("cgu_partitioning_place_inline_items");
202 partitioner.place_inlined_mono_items(cx, initial_partitioning)
205 post_inlining.codegen_units.iter_mut().for_each(|cgu| cgu.create_size_estimate(tcx));
207 debug_dump(tcx, "POST INLINING:", post_inlining.codegen_units.iter());
209 // Next we try to make as many symbols "internal" as possible, so LLVM has
210 // more freedom to optimize.
211 if !tcx.sess.link_dead_code() {
212 let _prof_timer = tcx.prof.generic_activity("cgu_partitioning_internalize_symbols");
213 partitioner.internalize_symbols(cx, &mut post_inlining);
216 let instrument_dead_code =
217 tcx.sess.instrument_coverage() && !tcx.sess.instrument_coverage_except_unused_functions();
219 if instrument_dead_code {
221 post_inlining.codegen_units.len() > 0,
222 "There must be at least one CGU that code coverage data can be generated in."
225 // Find the smallest CGU that has exported symbols and put the dead
226 // function stubs in that CGU. We look for exported symbols to increase
227 // the likelihood the linker won't throw away the dead functions.
228 // FIXME(#92165): In order to truly resolve this, we need to make sure
229 // the object file (CGU) containing the dead function stubs is included
230 // in the final binary. This will probably require forcing these
231 // function symbols to be included via `-u` or `/include` linker args.
232 let mut cgus: Vec<_> = post_inlining.codegen_units.iter_mut().collect();
233 cgus.sort_by_key(|cgu| cgu.size_estimate());
236 if let Some(cgu) = cgus.into_iter().rev().find(|cgu| {
237 cgu.items().iter().any(|(_, (linkage, _))| *linkage == Linkage::External)
241 // If there are no CGUs that have externally linked items,
242 // then we just pick the first CGU as a fallback.
243 &mut post_inlining.codegen_units[0]
245 dead_code_cgu.make_code_coverage_dead_code_cgu();
248 // Finally, sort by codegen unit name, so that we get deterministic results.
249 let PostInliningPartitioning {
250 codegen_units: mut result,
251 mono_item_placements: _,
252 internalization_candidates: _,
255 result.sort_by(|a, b| a.name().as_str().partial_cmp(b.name().as_str()).unwrap());
260 pub struct PreInliningPartitioning<'tcx> {
261 codegen_units: Vec<CodegenUnit<'tcx>>,
262 roots: FxHashSet<MonoItem<'tcx>>,
263 internalization_candidates: FxHashSet<MonoItem<'tcx>>,
266 /// For symbol internalization, we need to know whether a symbol/mono-item is
267 /// accessed from outside the codegen unit it is defined in. This type is used
268 /// to keep track of that.
269 #[derive(Clone, PartialEq, Eq, Debug)]
270 enum MonoItemPlacement {
271 SingleCgu { cgu_name: Symbol },
275 struct PostInliningPartitioning<'tcx> {
276 codegen_units: Vec<CodegenUnit<'tcx>>,
277 mono_item_placements: FxHashMap<MonoItem<'tcx>, MonoItemPlacement>,
278 internalization_candidates: FxHashSet<MonoItem<'tcx>>,
281 fn debug_dump<'a, 'tcx, I>(tcx: TyCtxt<'tcx>, label: &str, cgus: I)
283 I: Iterator<Item = &'a CodegenUnit<'tcx>>,
289 let s = &mut String::new();
290 let _ = writeln!(s, "{label}");
293 writeln!(s, "CodegenUnit {} estimated size {} :", cgu.name(), cgu.size_estimate());
295 for (mono_item, linkage) in cgu.items() {
296 let symbol_name = mono_item.symbol_name(tcx).name;
297 let symbol_hash_start = symbol_name.rfind('h');
298 let symbol_hash = symbol_hash_start.map_or("<no hash>", |i| &symbol_name[i..]);
302 " - {} [{:?}] [{}] estimated size {}",
306 mono_item.size_estimate(tcx)
316 debug!("{}", dump());
319 #[inline(never)] // give this a place in the profiler
320 fn assert_symbols_are_distinct<'a, 'tcx, I>(tcx: TyCtxt<'tcx>, mono_items: I)
322 I: Iterator<Item = &'a MonoItem<'tcx>>,
325 let _prof_timer = tcx.prof.generic_activity("assert_symbols_are_distinct");
327 let mut symbols: Vec<_> =
328 mono_items.map(|mono_item| (mono_item, mono_item.symbol_name(tcx))).collect();
330 symbols.sort_by_key(|sym| sym.1);
332 for &[(mono_item1, ref sym1), (mono_item2, ref sym2)] in symbols.array_windows() {
334 let span1 = mono_item1.local_span(tcx);
335 let span2 = mono_item2.local_span(tcx);
337 // Deterministically select one of the spans for error reporting
338 let span = match (span1, span2) {
339 (Some(span1), Some(span2)) => {
340 Some(if span1.lo().0 > span2.lo().0 { span1 } else { span2 })
342 (span1, span2) => span1.or(span2),
345 tcx.sess.emit_fatal(SymbolAlreadyDefined { span, symbol: sym1.to_string() });
350 fn collect_and_partition_mono_items(tcx: TyCtxt<'_>, (): ()) -> (&DefIdSet, &[CodegenUnit<'_>]) {
351 let collection_mode = match tcx.sess.opts.unstable_opts.print_mono_items {
353 let mode = s.to_lowercase();
354 let mode = mode.trim();
356 MonoItemCollectionMode::Eager
359 tcx.sess.emit_warning(UnknownCguCollectionMode { mode });
362 MonoItemCollectionMode::Lazy
366 if tcx.sess.link_dead_code() {
367 MonoItemCollectionMode::Eager
369 MonoItemCollectionMode::Lazy
374 let (items, inlining_map) = collector::collect_crate_mono_items(tcx, collection_mode);
376 tcx.sess.abort_if_errors();
378 let (codegen_units, _) = tcx.sess.time("partition_and_assert_distinct_symbols", || {
381 let mut codegen_units = partition(
383 &mut items.iter().cloned(),
384 tcx.sess.codegen_units(),
387 codegen_units[0].make_primary();
388 &*tcx.arena.alloc_from_iter(codegen_units)
390 || assert_symbols_are_distinct(tcx, items.iter()),
394 if tcx.prof.enabled() {
395 // Record CGU size estimates for self-profiling.
396 for cgu in codegen_units {
397 tcx.prof.artifact_size(
398 "codegen_unit_size_estimate",
400 cgu.size_estimate() as u64,
405 let mono_items: DefIdSet = items
407 .filter_map(|mono_item| match *mono_item {
408 MonoItem::Fn(ref instance) => Some(instance.def_id()),
409 MonoItem::Static(def_id) => Some(def_id),
414 // Output monomorphization stats per def_id
415 if let SwitchWithOptPath::Enabled(ref path) = tcx.sess.opts.unstable_opts.dump_mono_stats {
417 dump_mono_items_stats(tcx, &codegen_units, path, tcx.crate_name(LOCAL_CRATE))
419 tcx.sess.emit_fatal(CouldntDumpMonoStats { error: err.to_string() });
423 if tcx.sess.opts.unstable_opts.print_mono_items.is_some() {
424 let mut item_to_cgus: FxHashMap<_, Vec<_>> = Default::default();
426 for cgu in codegen_units {
427 for (&mono_item, &linkage) in cgu.items() {
428 item_to_cgus.entry(mono_item).or_default().push((cgu.name(), linkage));
432 let mut item_keys: Vec<_> = items
435 let mut output = with_no_trimmed_paths!(i.to_string());
436 output.push_str(" @@");
437 let mut empty = Vec::new();
438 let cgus = item_to_cgus.get_mut(i).unwrap_or(&mut empty);
439 cgus.sort_by_key(|(name, _)| *name);
441 for &(ref cgu_name, (linkage, _)) in cgus.iter() {
443 output.push_str(cgu_name.as_str());
445 let linkage_abbrev = match linkage {
446 Linkage::External => "External",
447 Linkage::AvailableExternally => "Available",
448 Linkage::LinkOnceAny => "OnceAny",
449 Linkage::LinkOnceODR => "OnceODR",
450 Linkage::WeakAny => "WeakAny",
451 Linkage::WeakODR => "WeakODR",
452 Linkage::Appending => "Appending",
453 Linkage::Internal => "Internal",
454 Linkage::Private => "Private",
455 Linkage::ExternalWeak => "ExternalWeak",
456 Linkage::Common => "Common",
460 output.push_str(linkage_abbrev);
469 for item in item_keys {
470 println!("MONO_ITEM {item}");
474 (tcx.arena.alloc(mono_items), codegen_units)
477 /// Outputs stats about instantation counts and estimated size, per `MonoItem`'s
478 /// def, to a file in the given output directory.
479 fn dump_mono_items_stats<'tcx>(
481 codegen_units: &[CodegenUnit<'tcx>],
482 output_directory: &Option<PathBuf>,
484 ) -> Result<(), Box<dyn std::error::Error>> {
485 let output_directory = if let Some(ref directory) = output_directory {
486 fs::create_dir_all(directory)?;
492 let format = tcx.sess.opts.unstable_opts.dump_mono_stats_format;
493 let ext = format.extension();
494 let filename = format!("{crate_name}.mono_items.{ext}");
495 let output_path = output_directory.join(&filename);
496 let file = File::create(&output_path)?;
497 let mut file = BufWriter::new(file);
499 // Gather instantiated mono items grouped by def_id
500 let mut items_per_def_id: FxHashMap<_, Vec<_>> = Default::default();
501 for cgu in codegen_units {
502 for (&mono_item, _) in cgu.items() {
503 // Avoid variable-sized compiler-generated shims
504 if mono_item.is_user_defined() {
505 items_per_def_id.entry(mono_item.def_id()).or_default().push(mono_item);
510 #[derive(serde::Serialize)]
513 instantiation_count: usize,
514 size_estimate: usize,
515 total_estimate: usize,
518 // Output stats sorted by total instantiated size, from heaviest to lightest
519 let mut stats: Vec<_> = items_per_def_id
521 .map(|(def_id, items)| {
522 let name = with_no_trimmed_paths!(tcx.def_path_str(def_id));
523 let instantiation_count = items.len();
524 let size_estimate = items[0].size_estimate(tcx);
525 let total_estimate = instantiation_count * size_estimate;
526 MonoItem { name, instantiation_count, size_estimate, total_estimate }
529 stats.sort_unstable_by_key(|item| cmp::Reverse(item.total_estimate));
531 if !stats.is_empty() {
533 DumpMonoStatsFormat::Json => serde_json::to_writer(file, &stats)?,
534 DumpMonoStatsFormat::Markdown => {
537 "| Item | Instantiation count | Estimated Cost Per Instantiation | Total Estimated Cost |"
539 writeln!(file, "| --- | ---: | ---: | ---: |")?;
541 for MonoItem { name, instantiation_count, size_estimate, total_estimate } in stats {
544 "| `{name}` | {instantiation_count} | {size_estimate} | {total_estimate} |"
554 fn codegened_and_inlined_items(tcx: TyCtxt<'_>, (): ()) -> &DefIdSet {
555 let (items, cgus) = tcx.collect_and_partition_mono_items(());
556 let mut visited = DefIdSet::default();
557 let mut result = items.clone();
560 for (item, _) in cgu.items() {
561 if let MonoItem::Fn(ref instance) = item {
562 let did = instance.def_id();
563 if !visited.insert(did) {
566 let body = tcx.instance_mir(instance.def);
567 for block in body.basic_blocks.iter() {
568 for statement in &block.statements {
569 let mir::StatementKind::Coverage(_) = statement.kind else { continue };
570 let scope = statement.source_info.scope;
571 if let Some(inlined) = scope.inlined_instance(&body.source_scopes) {
572 result.insert(inlined.def_id());
580 tcx.arena.alloc(result)
583 pub fn provide(providers: &mut Providers) {
584 providers.collect_and_partition_mono_items = collect_and_partition_mono_items;
585 providers.codegened_and_inlined_items = codegened_and_inlined_items;
587 providers.is_codegened_item = |tcx, def_id| {
588 let (all_mono_items, _) = tcx.collect_and_partition_mono_items(());
589 all_mono_items.contains(&def_id)
592 providers.codegen_unit = |tcx, name| {
593 let (_, all) = tcx.collect_and_partition_mono_items(());
595 .find(|cgu| cgu.name() == name)
596 .unwrap_or_else(|| panic!("failed to find cgu with name {name:?}"))