1 use crate::back::link::are_upstream_rust_objects_already_included;
2 use crate::back::metadata::create_compressed_metadata_file;
3 use crate::back::write::{
4 compute_per_cgu_lto_type, start_async_codegen, submit_codegened_module_to_llvm,
5 submit_post_lto_module_to_llvm, submit_pre_lto_module_to_llvm, ComputedLtoType, OngoingCodegen,
7 use crate::common::{IntPredicate, RealPredicate, TypeKind};
10 use crate::mir::operand::OperandValue;
11 use crate::mir::place::PlaceRef;
13 use crate::{CachedModuleCodegen, CompiledModule, CrateInfo, MemFlags, ModuleCodegen, ModuleKind};
15 use rustc_attr as attr;
16 use rustc_data_structures::fx::{FxHashMap, FxHashSet};
17 use rustc_data_structures::profiling::{get_resident_set_size, print_time_passes_entry};
19 use rustc_data_structures::sync::par_iter;
20 #[cfg(parallel_compiler)]
21 use rustc_data_structures::sync::ParallelIterator;
23 use rustc_hir::def_id::{DefId, LOCAL_CRATE};
24 use rustc_hir::lang_items::LangItem;
25 use rustc_hir::weak_lang_items::WEAK_ITEMS_SYMBOLS;
26 use rustc_index::vec::Idx;
27 use rustc_metadata::EncodedMetadata;
28 use rustc_middle::middle::codegen_fn_attrs::CodegenFnAttrs;
29 use rustc_middle::middle::exported_symbols;
30 use rustc_middle::middle::exported_symbols::SymbolExportKind;
31 use rustc_middle::middle::lang_items;
32 use rustc_middle::mir::mono::{CodegenUnit, CodegenUnitNameBuilder, MonoItem};
33 use rustc_middle::ty::layout::{HasTyCtxt, LayoutOf, TyAndLayout};
34 use rustc_middle::ty::query::Providers;
35 use rustc_middle::ty::{self, Instance, Ty, TyCtxt};
36 use rustc_session::cgu_reuse_tracker::CguReuse;
37 use rustc_session::config::{self, CrateType, EntryFnType, OutputType};
38 use rustc_session::Session;
39 use rustc_span::symbol::sym;
40 use rustc_span::Symbol;
41 use rustc_span::{DebuggerVisualizerFile, DebuggerVisualizerType};
42 use rustc_target::abi::{Align, Size, VariantIdx};
44 use std::collections::BTreeSet;
45 use std::convert::TryFrom;
46 use std::time::{Duration, Instant};
48 use itertools::Itertools;
50 pub fn bin_op_to_icmp_predicate(op: hir::BinOpKind, signed: bool) -> IntPredicate {
52 hir::BinOpKind::Eq => IntPredicate::IntEQ,
53 hir::BinOpKind::Ne => IntPredicate::IntNE,
54 hir::BinOpKind::Lt => {
61 hir::BinOpKind::Le => {
68 hir::BinOpKind::Gt => {
75 hir::BinOpKind::Ge => {
83 "comparison_op_to_icmp_predicate: expected comparison operator, \
90 pub fn bin_op_to_fcmp_predicate(op: hir::BinOpKind) -> RealPredicate {
92 hir::BinOpKind::Eq => RealPredicate::RealOEQ,
93 hir::BinOpKind::Ne => RealPredicate::RealUNE,
94 hir::BinOpKind::Lt => RealPredicate::RealOLT,
95 hir::BinOpKind::Le => RealPredicate::RealOLE,
96 hir::BinOpKind::Gt => RealPredicate::RealOGT,
97 hir::BinOpKind::Ge => RealPredicate::RealOGE,
100 "comparison_op_to_fcmp_predicate: expected comparison operator, \
108 pub fn compare_simd_types<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
116 let signed = match t.kind() {
118 let cmp = bin_op_to_fcmp_predicate(op);
119 let cmp = bx.fcmp(cmp, lhs, rhs);
120 return bx.sext(cmp, ret_ty);
122 ty::Uint(_) => false,
124 _ => bug!("compare_simd_types: invalid SIMD type"),
127 let cmp = bin_op_to_icmp_predicate(op, signed);
128 let cmp = bx.icmp(cmp, lhs, rhs);
129 // LLVM outputs an `< size x i1 >`, so we need to perform a sign extension
130 // to get the correctly sized type. This will compile to a single instruction
131 // once the IR is converted to assembly if the SIMD instruction is supported
132 // by the target architecture.
136 /// Retrieves the information we are losing (making dynamic) in an unsizing
139 /// The `old_info` argument is a bit odd. It is intended for use in an upcast,
140 /// where the new vtable for an object will be derived from the old one.
141 pub fn unsized_info<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
145 old_info: Option<Bx::Value>,
148 let (source, target) =
149 cx.tcx().struct_lockstep_tails_erasing_lifetimes(source, target, bx.param_env());
150 match (source.kind(), target.kind()) {
151 (&ty::Array(_, len), &ty::Slice(_)) => {
152 cx.const_usize(len.eval_usize(cx.tcx(), ty::ParamEnv::reveal_all()))
155 &ty::Dynamic(ref data_a, _, src_dyn_kind),
156 &ty::Dynamic(ref data_b, _, target_dyn_kind),
158 assert_eq!(src_dyn_kind, target_dyn_kind);
161 old_info.expect("unsized_info: missing old info for trait upcasting coercion");
162 if data_a.principal_def_id() == data_b.principal_def_id() {
163 // A NOP cast that doesn't actually change anything, should be allowed even with invalid vtables.
167 // trait upcasting coercion
170 cx.tcx().vtable_trait_upcasting_coercion_new_vptr_slot((source, target));
172 if let Some(entry_idx) = vptr_entry_idx {
173 let ptr_ty = cx.type_i8p();
174 let ptr_align = cx.tcx().data_layout.pointer_align.abi;
175 let vtable_ptr_ty = vtable_ptr_ty(cx, target, target_dyn_kind);
176 let llvtable = bx.pointercast(old_info, bx.type_ptr_to(ptr_ty));
177 let gep = bx.inbounds_gep(
180 &[bx.const_usize(u64::try_from(entry_idx).unwrap())],
182 let new_vptr = bx.load(ptr_ty, gep, ptr_align);
183 bx.nonnull_metadata(new_vptr);
184 // VTable loads are invariant.
185 bx.set_invariant_load(new_vptr);
186 bx.pointercast(new_vptr, vtable_ptr_ty)
191 (_, &ty::Dynamic(ref data, _, target_dyn_kind)) => {
192 let vtable_ptr_ty = vtable_ptr_ty(cx, target, target_dyn_kind);
193 cx.const_ptrcast(meth::get_vtable(cx, source, data.principal()), vtable_ptr_ty)
195 _ => bug!("unsized_info: invalid unsizing {:?} -> {:?}", source, target),
199 // Returns the vtable pointer type of a `dyn` or `dyn*` type
200 fn vtable_ptr_ty<'tcx, Cx: CodegenMethods<'tcx>>(
204 ) -> <Cx as BackendTypes>::Type {
205 cx.scalar_pair_element_backend_type(
206 cx.layout_of(match kind {
207 // vtable is the second field of `*mut dyn Trait`
208 ty::Dyn => cx.tcx().mk_mut_ptr(target),
209 // vtable is the second field of `dyn* Trait`
210 ty::DynStar => target,
217 /// Coerces `src` to `dst_ty`. `src_ty` must be a pointer.
218 pub fn unsize_ptr<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
223 old_info: Option<Bx::Value>,
224 ) -> (Bx::Value, Bx::Value) {
225 debug!("unsize_ptr: {:?} => {:?}", src_ty, dst_ty);
226 match (src_ty.kind(), dst_ty.kind()) {
227 (&ty::Ref(_, a, _), &ty::Ref(_, b, _) | &ty::RawPtr(ty::TypeAndMut { ty: b, .. }))
228 | (&ty::RawPtr(ty::TypeAndMut { ty: a, .. }), &ty::RawPtr(ty::TypeAndMut { ty: b, .. })) => {
229 assert_eq!(bx.cx().type_is_sized(a), old_info.is_none());
230 let ptr_ty = bx.cx().type_ptr_to(bx.cx().backend_type(bx.cx().layout_of(b)));
231 (bx.pointercast(src, ptr_ty), unsized_info(bx, a, b, old_info))
233 (&ty::Adt(def_a, _), &ty::Adt(def_b, _)) => {
234 assert_eq!(def_a, def_b);
235 let src_layout = bx.cx().layout_of(src_ty);
236 let dst_layout = bx.cx().layout_of(dst_ty);
237 if src_ty == dst_ty {
238 return (src, old_info.unwrap());
240 let mut result = None;
241 for i in 0..src_layout.fields.count() {
242 let src_f = src_layout.field(bx.cx(), i);
247 assert_eq!(src_layout.fields.offset(i).bytes(), 0);
248 assert_eq!(dst_layout.fields.offset(i).bytes(), 0);
249 assert_eq!(src_layout.size, src_f.size);
251 let dst_f = dst_layout.field(bx.cx(), i);
252 assert_ne!(src_f.ty, dst_f.ty);
253 assert_eq!(result, None);
254 result = Some(unsize_ptr(bx, src, src_f.ty, dst_f.ty, old_info));
256 let (lldata, llextra) = result.unwrap();
257 let lldata_ty = bx.cx().scalar_pair_element_backend_type(dst_layout, 0, true);
258 let llextra_ty = bx.cx().scalar_pair_element_backend_type(dst_layout, 1, true);
259 // HACK(eddyb) have to bitcast pointers until LLVM removes pointee types.
260 (bx.bitcast(lldata, lldata_ty), bx.bitcast(llextra, llextra_ty))
262 _ => bug!("unsize_ptr: called on bad types"),
266 /// Coerces `src` to `dst_ty` which is guaranteed to be a `dyn*` type.
267 pub fn cast_to_dyn_star<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
270 src_ty_and_layout: TyAndLayout<'tcx>,
272 old_info: Option<Bx::Value>,
273 ) -> (Bx::Value, Bx::Value) {
274 debug!("cast_to_dyn_star: {:?} => {:?}", src_ty_and_layout.ty, dst_ty);
276 matches!(dst_ty.kind(), ty::Dynamic(_, _, ty::DynStar)),
277 "destination type must be a dyn*"
279 // FIXME(dyn-star): this is probably not the best way to check if this is
280 // a pointer, and really we should ensure that the value is a suitable
281 // pointer earlier in the compilation process.
282 let src = match src_ty_and_layout.pointee_info_at(bx.cx(), Size::ZERO) {
283 Some(_) => bx.ptrtoint(src, bx.cx().type_isize()),
284 None => bx.bitcast(src, bx.type_isize()),
286 (src, unsized_info(bx, src_ty_and_layout.ty, dst_ty, old_info))
289 /// Coerces `src`, which is a reference to a value of type `src_ty`,
290 /// to a value of type `dst_ty`, and stores the result in `dst`.
291 pub fn coerce_unsized_into<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
293 src: PlaceRef<'tcx, Bx::Value>,
294 dst: PlaceRef<'tcx, Bx::Value>,
296 let src_ty = src.layout.ty;
297 let dst_ty = dst.layout.ty;
298 match (src_ty.kind(), dst_ty.kind()) {
299 (&ty::Ref(..), &ty::Ref(..) | &ty::RawPtr(..)) | (&ty::RawPtr(..), &ty::RawPtr(..)) => {
300 let (base, info) = match bx.load_operand(src).val {
301 OperandValue::Pair(base, info) => unsize_ptr(bx, base, src_ty, dst_ty, Some(info)),
302 OperandValue::Immediate(base) => unsize_ptr(bx, base, src_ty, dst_ty, None),
303 OperandValue::Ref(..) => bug!(),
305 OperandValue::Pair(base, info).store(bx, dst);
308 (&ty::Adt(def_a, _), &ty::Adt(def_b, _)) => {
309 assert_eq!(def_a, def_b);
311 for i in 0..def_a.variant(VariantIdx::new(0)).fields.len() {
312 let src_f = src.project_field(bx, i);
313 let dst_f = dst.project_field(bx, i);
315 if dst_f.layout.is_zst() {
319 if src_f.layout.ty == dst_f.layout.ty {
330 coerce_unsized_into(bx, src_f, dst_f);
334 _ => bug!("coerce_unsized_into: invalid coercion {:?} -> {:?}", src_ty, dst_ty,),
338 pub fn cast_shift_expr_rhs<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
343 // Shifts may have any size int on the rhs
344 let mut rhs_llty = bx.cx().val_ty(rhs);
345 let mut lhs_llty = bx.cx().val_ty(lhs);
346 if bx.cx().type_kind(rhs_llty) == TypeKind::Vector {
347 rhs_llty = bx.cx().element_type(rhs_llty)
349 if bx.cx().type_kind(lhs_llty) == TypeKind::Vector {
350 lhs_llty = bx.cx().element_type(lhs_llty)
352 let rhs_sz = bx.cx().int_width(rhs_llty);
353 let lhs_sz = bx.cx().int_width(lhs_llty);
355 bx.trunc(rhs, lhs_llty)
356 } else if lhs_sz > rhs_sz {
357 // FIXME (#1877: If in the future shifting by negative
358 // values is no longer undefined then this is wrong.
359 bx.zext(rhs, lhs_llty)
365 /// Returns `true` if this session's target will use SEH-based unwinding.
367 /// This is only true for MSVC targets, and even then the 64-bit MSVC target
368 /// currently uses SEH-ish unwinding with DWARF info tables to the side (same as
369 /// 64-bit MinGW) instead of "full SEH".
370 pub fn wants_msvc_seh(sess: &Session) -> bool {
371 sess.target.is_like_msvc
374 pub fn memcpy_ty<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
380 layout: TyAndLayout<'tcx>,
383 let size = layout.size.bytes();
388 bx.memcpy(dst, dst_align, src, src_align, bx.cx().const_usize(size), flags);
391 pub fn codegen_instance<'a, 'tcx: 'a, Bx: BuilderMethods<'a, 'tcx>>(
392 cx: &'a Bx::CodegenCx,
393 instance: Instance<'tcx>,
395 // this is an info! to allow collecting monomorphization statistics
396 // and to allow finding the last function before LLVM aborts from
398 info!("codegen_instance({})", instance);
400 mir::codegen_mir::<Bx>(cx, instance);
403 /// Creates the `main` function which will initialize the rust runtime and call
404 /// users main function.
405 pub fn maybe_create_entry_wrapper<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
406 cx: &'a Bx::CodegenCx,
407 ) -> Option<Bx::Function> {
408 let (main_def_id, entry_type) = cx.tcx().entry_fn(())?;
409 let main_is_local = main_def_id.is_local();
410 let instance = Instance::mono(cx.tcx(), main_def_id);
413 // We want to create the wrapper in the same codegen unit as Rust's main
415 if !cx.codegen_unit().contains_item(&MonoItem::Fn(instance)) {
418 } else if !cx.codegen_unit().is_primary() {
419 // We want to create the wrapper only when the codegen unit is the primary one
423 let main_llfn = cx.get_fn_addr(instance);
425 let entry_fn = create_entry_fn::<Bx>(cx, main_llfn, main_def_id, entry_type);
426 return Some(entry_fn);
428 fn create_entry_fn<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
429 cx: &'a Bx::CodegenCx,
430 rust_main: Bx::Value,
431 rust_main_def_id: DefId,
432 entry_type: EntryFnType,
434 // The entry function is either `int main(void)` or `int main(int argc, char **argv)`,
435 // depending on whether the target needs `argc` and `argv` to be passed in.
436 let llfty = if cx.sess().target.main_needs_argc_argv {
437 cx.type_func(&[cx.type_int(), cx.type_ptr_to(cx.type_i8p())], cx.type_int())
439 cx.type_func(&[], cx.type_int())
442 let main_ret_ty = cx.tcx().fn_sig(rust_main_def_id).output();
443 // Given that `main()` has no arguments,
444 // then its return type cannot have
445 // late-bound regions, since late-bound
446 // regions must appear in the argument
448 let main_ret_ty = cx.tcx().normalize_erasing_regions(
449 ty::ParamEnv::reveal_all(),
450 main_ret_ty.no_bound_vars().unwrap(),
453 let Some(llfn) = cx.declare_c_main(llfty) else {
454 // FIXME: We should be smart and show a better diagnostic here.
455 let span = cx.tcx().def_span(rust_main_def_id);
457 .struct_span_err(span, "entry symbol `main` declared multiple times")
458 .help("did you use `#[no_mangle]` on `fn main`? Use `#[start]` instead")
460 cx.sess().abort_if_errors();
464 // `main` should respect same config for frame pointer elimination as rest of code
465 cx.set_frame_pointer_type(llfn);
466 cx.apply_target_cpu_attr(llfn);
468 let llbb = Bx::append_block(&cx, llfn, "top");
469 let mut bx = Bx::build(&cx, llbb);
471 bx.insert_reference_to_gdb_debug_scripts_section_global();
473 let isize_ty = cx.type_isize();
474 let i8pp_ty = cx.type_ptr_to(cx.type_i8p());
475 let (arg_argc, arg_argv) = get_argc_argv(cx, &mut bx);
477 let (start_fn, start_ty, args) = if let EntryFnType::Main { sigpipe } = entry_type {
478 let start_def_id = cx.tcx().require_lang_item(LangItem::Start, None);
479 let start_fn = cx.get_fn_addr(
480 ty::Instance::resolve(
482 ty::ParamEnv::reveal_all(),
484 cx.tcx().intern_substs(&[main_ret_ty.into()]),
490 let i8_ty = cx.type_i8();
491 let arg_sigpipe = bx.const_u8(sigpipe);
494 cx.type_func(&[cx.val_ty(rust_main), isize_ty, i8pp_ty, i8_ty], isize_ty);
495 (start_fn, start_ty, vec![rust_main, arg_argc, arg_argv, arg_sigpipe])
497 debug!("using user-defined start fn");
498 let start_ty = cx.type_func(&[isize_ty, i8pp_ty], isize_ty);
499 (rust_main, start_ty, vec![arg_argc, arg_argv])
502 let result = bx.call(start_ty, None, start_fn, &args, None);
503 let cast = bx.intcast(result, cx.type_int(), true);
510 /// Obtain the `argc` and `argv` values to pass to the rust start function.
511 fn get_argc_argv<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
512 cx: &'a Bx::CodegenCx,
514 ) -> (Bx::Value, Bx::Value) {
515 if cx.sess().target.main_needs_argc_argv {
516 // Params from native `main()` used as args for rust start function
517 let param_argc = bx.get_param(0);
518 let param_argv = bx.get_param(1);
519 let arg_argc = bx.intcast(param_argc, cx.type_isize(), true);
520 let arg_argv = param_argv;
523 // The Rust start function doesn't need `argc` and `argv`, so just pass zeros.
524 let arg_argc = bx.const_int(cx.type_int(), 0);
525 let arg_argv = bx.const_null(cx.type_ptr_to(cx.type_i8p()));
530 /// This function returns all of the debugger visualizers specified for the
531 /// current crate as well as all upstream crates transitively that match the
532 /// `visualizer_type` specified.
533 pub fn collect_debugger_visualizers_transitive(
535 visualizer_type: DebuggerVisualizerType,
536 ) -> BTreeSet<DebuggerVisualizerFile> {
537 tcx.debugger_visualizers(LOCAL_CRATE)
543 let used_crate_source = tcx.used_crate_source(*cnum);
544 used_crate_source.rlib.is_some() || used_crate_source.rmeta.is_some()
546 .flat_map(|&cnum| tcx.debugger_visualizers(cnum)),
548 .filter(|visualizer| visualizer.visualizer_type == visualizer_type)
550 .collect::<BTreeSet<_>>()
553 pub fn codegen_crate<B: ExtraBackendMethods>(
557 metadata: EncodedMetadata,
558 need_metadata_module: bool,
559 ) -> OngoingCodegen<B> {
560 // Skip crate items and just output metadata in -Z no-codegen mode.
561 if tcx.sess.opts.unstable_opts.no_codegen || !tcx.sess.opts.output_types.should_codegen() {
562 let ongoing_codegen = start_async_codegen(backend, tcx, target_cpu, metadata, None, 1);
564 ongoing_codegen.codegen_finished(tcx);
566 ongoing_codegen.check_for_errors(tcx.sess);
568 return ongoing_codegen;
571 let cgu_name_builder = &mut CodegenUnitNameBuilder::new(tcx);
573 // Run the monomorphization collector and partition the collected items into
575 let codegen_units = tcx.collect_and_partition_mono_items(()).1;
577 // Force all codegen_unit queries so they are already either red or green
578 // when compile_codegen_unit accesses them. We are not able to re-execute
579 // the codegen_unit query from just the DepNode, so an unknown color would
580 // lead to having to re-execute compile_codegen_unit, possibly
582 if tcx.dep_graph.is_fully_enabled() {
583 for cgu in codegen_units {
584 tcx.ensure().codegen_unit(cgu.name());
588 let metadata_module = if need_metadata_module {
589 // Emit compressed metadata object.
590 let metadata_cgu_name =
591 cgu_name_builder.build_cgu_name(LOCAL_CRATE, &["crate"], Some("metadata")).to_string();
592 tcx.sess.time("write_compressed_metadata", || {
594 tcx.output_filenames(()).temp_path(OutputType::Metadata, Some(&metadata_cgu_name));
595 let data = create_compressed_metadata_file(
598 &exported_symbols::metadata_symbol_name(tcx),
600 if let Err(err) = std::fs::write(&file_name, data) {
601 tcx.sess.fatal(&format!("error writing metadata object file: {}", err));
603 Some(CompiledModule {
604 name: metadata_cgu_name,
605 kind: ModuleKind::Metadata,
606 object: Some(file_name),
615 let ongoing_codegen = start_async_codegen(
624 // Codegen an allocator shim, if necessary.
626 // If the crate doesn't have an `allocator_kind` set then there's definitely
627 // no shim to generate. Otherwise we also check our dependency graph for all
628 // our output crate types. If anything there looks like its a `Dynamic`
629 // linkage, then it's already got an allocator shim and we'll be using that
630 // one instead. If nothing exists then it's our job to generate the
632 let any_dynamic_crate = tcx.dependency_formats(()).iter().any(|(_, list)| {
633 use rustc_middle::middle::dependency_format::Linkage;
634 list.iter().any(|&linkage| linkage == Linkage::Dynamic)
636 let allocator_module = if any_dynamic_crate {
638 } else if let Some(kind) = tcx.allocator_kind(()) {
640 cgu_name_builder.build_cgu_name(LOCAL_CRATE, &["crate"], Some("allocator")).to_string();
641 let module_llvm = tcx.sess.time("write_allocator_module", || {
642 backend.codegen_allocator(tcx, &llmod_id, kind, tcx.lang_items().oom().is_some())
645 Some(ModuleCodegen { name: llmod_id, module_llvm, kind: ModuleKind::Allocator })
650 if let Some(allocator_module) = allocator_module {
651 ongoing_codegen.submit_pre_codegened_module_to_llvm(tcx, allocator_module);
654 // For better throughput during parallel processing by LLVM, we used to sort
655 // CGUs largest to smallest. This would lead to better thread utilization
656 // by, for example, preventing a large CGU from being processed last and
657 // having only one LLVM thread working while the rest remained idle.
659 // However, this strategy would lead to high memory usage, as it meant the
660 // LLVM-IR for all of the largest CGUs would be resident in memory at once.
662 // Instead, we can compromise by ordering CGUs such that the largest and
663 // smallest are first, second largest and smallest are next, etc. If there
664 // are large size variations, this can reduce memory usage significantly.
665 let codegen_units: Vec<_> = {
666 let mut sorted_cgus = codegen_units.iter().collect::<Vec<_>>();
667 sorted_cgus.sort_by_cached_key(|cgu| cgu.size_estimate());
669 let (first_half, second_half) = sorted_cgus.split_at(sorted_cgus.len() / 2);
670 second_half.iter().rev().interleave(first_half).copied().collect()
673 // Calculate the CGU reuse
674 let cgu_reuse = tcx.sess.time("find_cgu_reuse", || {
675 codegen_units.iter().map(|cgu| determine_cgu_reuse(tcx, &cgu)).collect::<Vec<_>>()
678 let mut total_codegen_time = Duration::new(0, 0);
679 let start_rss = tcx.sess.time_passes().then(|| get_resident_set_size());
681 // The non-parallel compiler can only translate codegen units to LLVM IR
682 // on a single thread, leading to a staircase effect where the N LLVM
683 // threads have to wait on the single codegen threads to generate work
684 // for them. The parallel compiler does not have this restriction, so
685 // we can pre-load the LLVM queue in parallel before handing off
686 // coordination to the OnGoingCodegen scheduler.
688 // This likely is a temporary measure. Once we don't have to support the
689 // non-parallel compiler anymore, we can compile CGUs end-to-end in
690 // parallel and get rid of the complicated scheduling logic.
691 let mut pre_compiled_cgus = if cfg!(parallel_compiler) {
692 tcx.sess.time("compile_first_CGU_batch", || {
693 // Try to find one CGU to compile per thread.
694 let cgus: Vec<_> = cgu_reuse
697 .filter(|&(_, reuse)| reuse == &CguReuse::No)
698 .take(tcx.sess.threads())
701 // Compile the found CGUs in parallel.
702 let start_time = Instant::now();
704 let pre_compiled_cgus = par_iter(cgus)
706 let module = backend.compile_codegen_unit(tcx, codegen_units[i].name());
711 total_codegen_time += start_time.elapsed();
719 for (i, cgu) in codegen_units.iter().enumerate() {
720 ongoing_codegen.wait_for_signal_to_codegen_item();
721 ongoing_codegen.check_for_errors(tcx.sess);
723 let cgu_reuse = cgu_reuse[i];
724 tcx.sess.cgu_reuse_tracker.set_actual_reuse(cgu.name().as_str(), cgu_reuse);
728 let (module, cost) = if let Some(cgu) = pre_compiled_cgus.remove(&i) {
731 let start_time = Instant::now();
732 let module = backend.compile_codegen_unit(tcx, cgu.name());
733 total_codegen_time += start_time.elapsed();
736 // This will unwind if there are errors, which triggers our `AbortCodegenOnDrop`
737 // guard. Unfortunately, just skipping the `submit_codegened_module_to_llvm` makes
738 // compilation hang on post-monomorphization errors.
739 tcx.sess.abort_if_errors();
741 submit_codegened_module_to_llvm(
743 &ongoing_codegen.coordinator.sender,
749 CguReuse::PreLto => {
750 submit_pre_lto_module_to_llvm(
753 &ongoing_codegen.coordinator.sender,
754 CachedModuleCodegen {
755 name: cgu.name().to_string(),
756 source: cgu.previous_work_product(tcx),
761 CguReuse::PostLto => {
762 submit_post_lto_module_to_llvm(
764 &ongoing_codegen.coordinator.sender,
765 CachedModuleCodegen {
766 name: cgu.name().to_string(),
767 source: cgu.previous_work_product(tcx),
775 ongoing_codegen.codegen_finished(tcx);
777 // Since the main thread is sometimes blocked during codegen, we keep track
778 // -Ztime-passes output manually.
779 if tcx.sess.time_passes() {
780 let end_rss = get_resident_set_size();
782 print_time_passes_entry(
783 "codegen_to_LLVM_IR",
790 ongoing_codegen.check_for_errors(tcx.sess);
795 pub fn new(tcx: TyCtxt<'_>, target_cpu: String) -> CrateInfo {
796 let exported_symbols = tcx
800 .map(|&c| (c, crate::back::linker::exported_symbols(tcx, c)))
802 let linked_symbols = tcx
806 .map(|&c| (c, crate::back::linker::linked_symbols(tcx, c)))
808 let local_crate_name = tcx.crate_name(LOCAL_CRATE);
809 let crate_attrs = tcx.hir().attrs(rustc_hir::CRATE_HIR_ID);
810 let subsystem = tcx.sess.first_attr_value_str_by_name(crate_attrs, sym::windows_subsystem);
811 let windows_subsystem = subsystem.map(|subsystem| {
812 if subsystem != sym::windows && subsystem != sym::console {
813 tcx.sess.fatal(&format!(
814 "invalid windows subsystem `{}`, only \
815 `windows` and `console` are allowed",
819 subsystem.to_string()
822 // This list is used when generating the command line to pass through to
823 // system linker. The linker expects undefined symbols on the left of the
824 // command line to be defined in libraries on the right, not the other way
825 // around. For more info, see some comments in the add_used_library function
828 // In order to get this left-to-right dependency ordering, we use the reverse
829 // postorder of all crates putting the leaves at the right-most positions.
830 let used_crates = tcx
835 .filter(|&cnum| !tcx.dep_kind(cnum).macros_only())
838 let mut info = CrateInfo {
843 compiler_builtins: None,
844 profiler_runtime: None,
845 is_no_builtins: Default::default(),
846 native_libraries: Default::default(),
847 used_libraries: tcx.native_libraries(LOCAL_CRATE).iter().map(Into::into).collect(),
848 crate_name: Default::default(),
850 used_crate_source: Default::default(),
851 dependency_formats: tcx.dependency_formats(()).clone(),
853 natvis_debugger_visualizers: Default::default(),
855 let crates = tcx.crates(());
857 let n_crates = crates.len();
858 info.native_libraries.reserve(n_crates);
859 info.crate_name.reserve(n_crates);
860 info.used_crate_source.reserve(n_crates);
862 for &cnum in crates.iter() {
863 info.native_libraries
864 .insert(cnum, tcx.native_libraries(cnum).iter().map(Into::into).collect());
865 info.crate_name.insert(cnum, tcx.crate_name(cnum));
867 let used_crate_source = tcx.used_crate_source(cnum);
868 info.used_crate_source.insert(cnum, used_crate_source.clone());
869 if tcx.is_compiler_builtins(cnum) {
870 info.compiler_builtins = Some(cnum);
872 if tcx.is_profiler_runtime(cnum) {
873 info.profiler_runtime = Some(cnum);
875 if tcx.is_no_builtins(cnum) {
876 info.is_no_builtins.insert(cnum);
880 // Handle circular dependencies in the standard library.
881 // See comment before `add_linked_symbol_object` function for the details.
882 // If global LTO is enabled then almost everything (*) is glued into a single object file,
883 // so this logic is not necessary and can cause issues on some targets (due to weak lang
884 // item symbols being "privatized" to that object file), so we disable it.
885 // (*) Native libs, and `#[compiler_builtins]` and `#[no_builtins]` crates are not glued,
886 // and we assume that they cannot define weak lang items. This is not currently enforced
887 // by the compiler, but that's ok because all this stuff is unstable anyway.
888 let target = &tcx.sess.target;
889 if !are_upstream_rust_objects_already_included(tcx.sess) {
890 let missing_weak_lang_items: FxHashSet<&Symbol> = info
894 tcx.missing_lang_items(*cnum)
896 .filter(|l| lang_items::required(tcx, **l))
897 .filter_map(|item| WEAK_ITEMS_SYMBOLS.get(item))
900 let prefix = if target.is_like_windows && target.arch == "x86" { "_" } else { "" };
903 .filter(|(crate_type, _)| {
904 !matches!(crate_type, CrateType::Rlib | CrateType::Staticlib)
906 .for_each(|(_, linked_symbols)| {
907 linked_symbols.extend(
908 missing_weak_lang_items
910 .map(|item| (format!("{prefix}{item}"), SymbolExportKind::Text)),
915 let embed_visualizers = tcx.sess.crate_types().iter().any(|&crate_type| match crate_type {
916 CrateType::Executable | CrateType::Dylib | CrateType::Cdylib => {
917 // These are crate types for which we invoke the linker and can embed
918 // NatVis visualizers.
921 CrateType::ProcMacro => {
922 // We could embed NatVis for proc macro crates too (to improve the debugging
923 // experience for them) but it does not seem like a good default, since
924 // this is a rare use case and we don't want to slow down the common case.
927 CrateType::Staticlib | CrateType::Rlib => {
928 // We don't invoke the linker for these, so we don't need to collect the NatVis for them.
933 if target.is_like_msvc && embed_visualizers {
934 info.natvis_debugger_visualizers =
935 collect_debugger_visualizers_transitive(tcx, DebuggerVisualizerType::Natvis);
942 pub fn provide(providers: &mut Providers) {
943 providers.backend_optimization_level = |tcx, cratenum| {
944 let for_speed = match tcx.sess.opts.optimize {
945 // If globally no optimisation is done, #[optimize] has no effect.
947 // This is done because if we ended up "upgrading" to `-O2` here, we’d populate the
948 // pass manager and it is likely that some module-wide passes (such as inliner or
949 // cross-function constant propagation) would ignore the `optnone` annotation we put
950 // on the functions, thus necessarily involving these functions into optimisations.
951 config::OptLevel::No => return config::OptLevel::No,
952 // If globally optimise-speed is already specified, just use that level.
953 config::OptLevel::Less => return config::OptLevel::Less,
954 config::OptLevel::Default => return config::OptLevel::Default,
955 config::OptLevel::Aggressive => return config::OptLevel::Aggressive,
956 // If globally optimize-for-size has been requested, use -O2 instead (if optimize(size)
958 config::OptLevel::Size => config::OptLevel::Default,
959 config::OptLevel::SizeMin => config::OptLevel::Default,
962 let (defids, _) = tcx.collect_and_partition_mono_items(cratenum);
964 let CodegenFnAttrs { optimize, .. } = tcx.codegen_fn_attrs(*id);
966 attr::OptimizeAttr::None => continue,
967 attr::OptimizeAttr::Size => continue,
968 attr::OptimizeAttr::Speed => {
973 tcx.sess.opts.optimize
977 fn determine_cgu_reuse<'tcx>(tcx: TyCtxt<'tcx>, cgu: &CodegenUnit<'tcx>) -> CguReuse {
978 if !tcx.dep_graph.is_fully_enabled() {
982 let work_product_id = &cgu.work_product_id();
983 if tcx.dep_graph.previous_work_product(work_product_id).is_none() {
984 // We don't have anything cached for this CGU. This can happen
985 // if the CGU did not exist in the previous session.
989 // Try to mark the CGU as green. If it we can do so, it means that nothing
990 // affecting the LLVM module has changed and we can re-use a cached version.
991 // If we compile with any kind of LTO, this means we can re-use the bitcode
992 // of the Pre-LTO stage (possibly also the Post-LTO version but we'll only
993 // know that later). If we are not doing LTO, there is only one optimized
994 // version of each module, so we re-use that.
995 let dep_node = cgu.codegen_dep_node(tcx);
997 !tcx.dep_graph.dep_node_exists(&dep_node),
998 "CompileCodegenUnit dep-node for CGU `{}` already exists before marking.",
1002 if tcx.try_mark_green(&dep_node) {
1003 // We can re-use either the pre- or the post-thinlto state. If no LTO is
1004 // being performed then we can use post-LTO artifacts, otherwise we must
1005 // reuse pre-LTO artifacts
1006 match compute_per_cgu_lto_type(
1009 &tcx.sess.crate_types(),
1010 ModuleKind::Regular,
1012 ComputedLtoType::No => CguReuse::PostLto,
1013 _ => CguReuse::PreLto,