1 use crate::back::metadata::create_compressed_metadata_file;
2 use crate::back::write::{
3 compute_per_cgu_lto_type, start_async_codegen, submit_codegened_module_to_llvm,
4 submit_post_lto_module_to_llvm, submit_pre_lto_module_to_llvm, ComputedLtoType, OngoingCodegen,
6 use crate::common::{IntPredicate, RealPredicate, TypeKind};
9 use crate::mir::operand::OperandValue;
10 use crate::mir::place::PlaceRef;
12 use crate::{CachedModuleCodegen, CompiledModule, CrateInfo, MemFlags, ModuleCodegen, ModuleKind};
14 use rustc_attr as attr;
15 use rustc_data_structures::fx::FxHashMap;
16 use rustc_data_structures::profiling::{get_resident_set_size, print_time_passes_entry};
18 #[cfg(parallel_compiler)]
19 use rustc_data_structures::sync::{par_iter, ParallelIterator};
21 use rustc_hir::def_id::{DefId, LOCAL_CRATE};
22 use rustc_hir::lang_items::LangItem;
23 use rustc_index::vec::Idx;
24 use rustc_metadata::EncodedMetadata;
25 use rustc_middle::middle::codegen_fn_attrs::CodegenFnAttrs;
26 use rustc_middle::middle::exported_symbols;
27 use rustc_middle::middle::lang_items;
28 use rustc_middle::mir::mono::{CodegenUnit, CodegenUnitNameBuilder, MonoItem};
29 use rustc_middle::ty::layout::{HasTyCtxt, LayoutOf, TyAndLayout};
30 use rustc_middle::ty::query::Providers;
31 use rustc_middle::ty::{self, Instance, Ty, TyCtxt};
32 use rustc_session::cgu_reuse_tracker::CguReuse;
33 use rustc_session::config::{self, EntryFnType, OutputType};
34 use rustc_session::Session;
35 use rustc_span::symbol::sym;
36 use rustc_target::abi::{Align, VariantIdx};
38 use std::convert::TryFrom;
39 use std::ops::{Deref, DerefMut};
40 use std::time::{Duration, Instant};
42 use itertools::Itertools;
44 pub fn bin_op_to_icmp_predicate(op: hir::BinOpKind, signed: bool) -> IntPredicate {
46 hir::BinOpKind::Eq => IntPredicate::IntEQ,
47 hir::BinOpKind::Ne => IntPredicate::IntNE,
48 hir::BinOpKind::Lt => {
55 hir::BinOpKind::Le => {
62 hir::BinOpKind::Gt => {
69 hir::BinOpKind::Ge => {
77 "comparison_op_to_icmp_predicate: expected comparison operator, \
84 pub fn bin_op_to_fcmp_predicate(op: hir::BinOpKind) -> RealPredicate {
86 hir::BinOpKind::Eq => RealPredicate::RealOEQ,
87 hir::BinOpKind::Ne => RealPredicate::RealUNE,
88 hir::BinOpKind::Lt => RealPredicate::RealOLT,
89 hir::BinOpKind::Le => RealPredicate::RealOLE,
90 hir::BinOpKind::Gt => RealPredicate::RealOGT,
91 hir::BinOpKind::Ge => RealPredicate::RealOGE,
94 "comparison_op_to_fcmp_predicate: expected comparison operator, \
102 pub fn compare_simd_types<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
110 let signed = match t.kind() {
112 let cmp = bin_op_to_fcmp_predicate(op);
113 let cmp = bx.fcmp(cmp, lhs, rhs);
114 return bx.sext(cmp, ret_ty);
116 ty::Uint(_) => false,
118 _ => bug!("compare_simd_types: invalid SIMD type"),
121 let cmp = bin_op_to_icmp_predicate(op, signed);
122 let cmp = bx.icmp(cmp, lhs, rhs);
123 // LLVM outputs an `< size x i1 >`, so we need to perform a sign extension
124 // to get the correctly sized type. This will compile to a single instruction
125 // once the IR is converted to assembly if the SIMD instruction is supported
126 // by the target architecture.
130 /// Retrieves the information we are losing (making dynamic) in an unsizing
133 /// The `old_info` argument is a bit odd. It is intended for use in an upcast,
134 /// where the new vtable for an object will be derived from the old one.
135 pub fn unsized_info<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
139 old_info: Option<Bx::Value>,
142 let (source, target) =
143 cx.tcx().struct_lockstep_tails_erasing_lifetimes(source, target, bx.param_env());
144 match (source.kind(), target.kind()) {
145 (&ty::Array(_, len), &ty::Slice(_)) => {
146 cx.const_usize(len.eval_usize(cx.tcx(), ty::ParamEnv::reveal_all()))
148 (&ty::Dynamic(ref data_a, ..), &ty::Dynamic(ref data_b, ..)) => {
150 old_info.expect("unsized_info: missing old info for trait upcasting coercion");
151 if data_a.principal_def_id() == data_b.principal_def_id() {
155 // trait upcasting coercion
158 cx.tcx().vtable_trait_upcasting_coercion_new_vptr_slot((source, target));
160 if let Some(entry_idx) = vptr_entry_idx {
161 let ptr_ty = cx.type_i8p();
162 let ptr_align = cx.tcx().data_layout.pointer_align.abi;
163 let llvtable = bx.pointercast(old_info, bx.type_ptr_to(ptr_ty));
164 let gep = bx.inbounds_gep(
167 &[bx.const_usize(u64::try_from(entry_idx).unwrap())],
169 let new_vptr = bx.load(ptr_ty, gep, ptr_align);
170 bx.nonnull_metadata(new_vptr);
171 // Vtable loads are invariant.
172 bx.set_invariant_load(new_vptr);
178 (_, &ty::Dynamic(ref data, ..)) => {
179 let vtable_ptr_ty = cx.scalar_pair_element_backend_type(
180 cx.layout_of(cx.tcx().mk_mut_ptr(target)),
184 cx.const_ptrcast(meth::get_vtable(cx, source, data.principal()), vtable_ptr_ty)
186 _ => bug!("unsized_info: invalid unsizing {:?} -> {:?}", source, target),
190 /// Coerces `src` to `dst_ty`. `src_ty` must be a pointer.
191 pub fn unsize_ptr<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
196 old_info: Option<Bx::Value>,
197 ) -> (Bx::Value, Bx::Value) {
198 debug!("unsize_ptr: {:?} => {:?}", src_ty, dst_ty);
199 match (src_ty.kind(), dst_ty.kind()) {
200 (&ty::Ref(_, a, _), &ty::Ref(_, b, _) | &ty::RawPtr(ty::TypeAndMut { ty: b, .. }))
201 | (&ty::RawPtr(ty::TypeAndMut { ty: a, .. }), &ty::RawPtr(ty::TypeAndMut { ty: b, .. })) => {
202 assert_eq!(bx.cx().type_is_sized(a), old_info.is_none());
203 let ptr_ty = bx.cx().type_ptr_to(bx.cx().backend_type(bx.cx().layout_of(b)));
204 (bx.pointercast(src, ptr_ty), unsized_info(bx, a, b, old_info))
206 (&ty::Adt(def_a, _), &ty::Adt(def_b, _)) => {
207 assert_eq!(def_a, def_b);
208 let src_layout = bx.cx().layout_of(src_ty);
209 let dst_layout = bx.cx().layout_of(dst_ty);
210 if src_ty == dst_ty {
211 return (src, old_info.unwrap());
213 let mut result = None;
214 for i in 0..src_layout.fields.count() {
215 let src_f = src_layout.field(bx.cx(), i);
216 assert_eq!(src_layout.fields.offset(i).bytes(), 0);
217 assert_eq!(dst_layout.fields.offset(i).bytes(), 0);
221 assert_eq!(src_layout.size, src_f.size);
223 let dst_f = dst_layout.field(bx.cx(), i);
224 assert_ne!(src_f.ty, dst_f.ty);
225 assert_eq!(result, None);
226 result = Some(unsize_ptr(bx, src, src_f.ty, dst_f.ty, old_info));
228 let (lldata, llextra) = result.unwrap();
229 let lldata_ty = bx.cx().scalar_pair_element_backend_type(dst_layout, 0, true);
230 let llextra_ty = bx.cx().scalar_pair_element_backend_type(dst_layout, 1, true);
231 // HACK(eddyb) have to bitcast pointers until LLVM removes pointee types.
232 (bx.bitcast(lldata, lldata_ty), bx.bitcast(llextra, llextra_ty))
234 _ => bug!("unsize_ptr: called on bad types"),
238 /// Coerces `src`, which is a reference to a value of type `src_ty`,
239 /// to a value of type `dst_ty`, and stores the result in `dst`.
240 pub fn coerce_unsized_into<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
242 src: PlaceRef<'tcx, Bx::Value>,
243 dst: PlaceRef<'tcx, Bx::Value>,
245 let src_ty = src.layout.ty;
246 let dst_ty = dst.layout.ty;
247 match (src_ty.kind(), dst_ty.kind()) {
248 (&ty::Ref(..), &ty::Ref(..) | &ty::RawPtr(..)) | (&ty::RawPtr(..), &ty::RawPtr(..)) => {
249 let (base, info) = match bx.load_operand(src).val {
250 OperandValue::Pair(base, info) => unsize_ptr(bx, base, src_ty, dst_ty, Some(info)),
251 OperandValue::Immediate(base) => unsize_ptr(bx, base, src_ty, dst_ty, None),
252 OperandValue::Ref(..) => bug!(),
254 OperandValue::Pair(base, info).store(bx, dst);
257 (&ty::Adt(def_a, _), &ty::Adt(def_b, _)) => {
258 assert_eq!(def_a, def_b);
260 for i in 0..def_a.variant(VariantIdx::new(0)).fields.len() {
261 let src_f = src.project_field(bx, i);
262 let dst_f = dst.project_field(bx, i);
264 if dst_f.layout.is_zst() {
268 if src_f.layout.ty == dst_f.layout.ty {
279 coerce_unsized_into(bx, src_f, dst_f);
283 _ => bug!("coerce_unsized_into: invalid coercion {:?} -> {:?}", src_ty, dst_ty,),
287 pub fn cast_shift_expr_rhs<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
293 cast_shift_rhs(bx, op, lhs, rhs)
296 fn cast_shift_rhs<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
302 // Shifts may have any size int on the rhs
304 let mut rhs_llty = bx.cx().val_ty(rhs);
305 let mut lhs_llty = bx.cx().val_ty(lhs);
306 if bx.cx().type_kind(rhs_llty) == TypeKind::Vector {
307 rhs_llty = bx.cx().element_type(rhs_llty)
309 if bx.cx().type_kind(lhs_llty) == TypeKind::Vector {
310 lhs_llty = bx.cx().element_type(lhs_llty)
312 let rhs_sz = bx.cx().int_width(rhs_llty);
313 let lhs_sz = bx.cx().int_width(lhs_llty);
315 bx.trunc(rhs, lhs_llty)
316 } else if lhs_sz > rhs_sz {
317 // FIXME (#1877: If in the future shifting by negative
318 // values is no longer undefined then this is wrong.
319 bx.zext(rhs, lhs_llty)
328 /// Returns `true` if this session's target will use SEH-based unwinding.
330 /// This is only true for MSVC targets, and even then the 64-bit MSVC target
331 /// currently uses SEH-ish unwinding with DWARF info tables to the side (same as
332 /// 64-bit MinGW) instead of "full SEH".
333 pub fn wants_msvc_seh(sess: &Session) -> bool {
334 sess.target.is_like_msvc
337 pub fn memcpy_ty<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
343 layout: TyAndLayout<'tcx>,
346 let size = layout.size.bytes();
351 bx.memcpy(dst, dst_align, src, src_align, bx.cx().const_usize(size), flags);
354 pub fn codegen_instance<'a, 'tcx: 'a, Bx: BuilderMethods<'a, 'tcx>>(
355 cx: &'a Bx::CodegenCx,
356 instance: Instance<'tcx>,
358 // this is an info! to allow collecting monomorphization statistics
359 // and to allow finding the last function before LLVM aborts from
361 info!("codegen_instance({})", instance);
363 mir::codegen_mir::<Bx>(cx, instance);
366 /// Creates the `main` function which will initialize the rust runtime and call
367 /// users main function.
368 pub fn maybe_create_entry_wrapper<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
369 cx: &'a Bx::CodegenCx,
370 ) -> Option<Bx::Function> {
371 let (main_def_id, entry_type) = cx.tcx().entry_fn(())?;
372 let main_is_local = main_def_id.is_local();
373 let instance = Instance::mono(cx.tcx(), main_def_id);
376 // We want to create the wrapper in the same codegen unit as Rust's main
378 if !cx.codegen_unit().contains_item(&MonoItem::Fn(instance)) {
381 } else if !cx.codegen_unit().is_primary() {
382 // We want to create the wrapper only when the codegen unit is the primary one
386 let main_llfn = cx.get_fn_addr(instance);
388 let use_start_lang_item = EntryFnType::Start != entry_type;
389 let entry_fn = create_entry_fn::<Bx>(cx, main_llfn, main_def_id, use_start_lang_item);
390 return Some(entry_fn);
392 fn create_entry_fn<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
393 cx: &'a Bx::CodegenCx,
394 rust_main: Bx::Value,
395 rust_main_def_id: DefId,
396 use_start_lang_item: bool,
398 // The entry function is either `int main(void)` or `int main(int argc, char **argv)`,
399 // depending on whether the target needs `argc` and `argv` to be passed in.
400 let llfty = if cx.sess().target.main_needs_argc_argv {
401 cx.type_func(&[cx.type_int(), cx.type_ptr_to(cx.type_i8p())], cx.type_int())
403 cx.type_func(&[], cx.type_int())
406 let main_ret_ty = cx.tcx().fn_sig(rust_main_def_id).output();
407 // Given that `main()` has no arguments,
408 // then its return type cannot have
409 // late-bound regions, since late-bound
410 // regions must appear in the argument
412 let main_ret_ty = cx.tcx().normalize_erasing_regions(
413 ty::ParamEnv::reveal_all(),
414 main_ret_ty.no_bound_vars().unwrap(),
417 let Some(llfn) = cx.declare_c_main(llfty) else {
418 // FIXME: We should be smart and show a better diagnostic here.
419 let span = cx.tcx().def_span(rust_main_def_id);
421 .struct_span_err(span, "entry symbol `main` declared multiple times")
422 .help("did you use `#[no_mangle]` on `fn main`? Use `#[start]` instead")
424 cx.sess().abort_if_errors();
428 // `main` should respect same config for frame pointer elimination as rest of code
429 cx.set_frame_pointer_type(llfn);
430 cx.apply_target_cpu_attr(llfn);
432 let llbb = Bx::append_block(&cx, llfn, "top");
433 let mut bx = Bx::build(&cx, llbb);
435 bx.insert_reference_to_gdb_debug_scripts_section_global();
437 let isize_ty = cx.type_isize();
438 let i8pp_ty = cx.type_ptr_to(cx.type_i8p());
439 let (arg_argc, arg_argv) = get_argc_argv(cx, &mut bx);
441 let (start_fn, start_ty, args) = if use_start_lang_item {
442 let start_def_id = cx.tcx().require_lang_item(LangItem::Start, None);
443 let start_fn = cx.get_fn_addr(
444 ty::Instance::resolve(
446 ty::ParamEnv::reveal_all(),
448 cx.tcx().intern_substs(&[main_ret_ty.into()]),
453 let start_ty = cx.type_func(&[cx.val_ty(rust_main), isize_ty, i8pp_ty], isize_ty);
454 (start_fn, start_ty, vec![rust_main, arg_argc, arg_argv])
456 debug!("using user-defined start fn");
457 let start_ty = cx.type_func(&[isize_ty, i8pp_ty], isize_ty);
458 (rust_main, start_ty, vec![arg_argc, arg_argv])
461 let result = bx.call(start_ty, start_fn, &args, None);
462 let cast = bx.intcast(result, cx.type_int(), true);
469 /// Obtain the `argc` and `argv` values to pass to the rust start function.
470 fn get_argc_argv<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
471 cx: &'a Bx::CodegenCx,
473 ) -> (Bx::Value, Bx::Value) {
474 if cx.sess().target.main_needs_argc_argv {
475 // Params from native `main()` used as args for rust start function
476 let param_argc = bx.get_param(0);
477 let param_argv = bx.get_param(1);
478 let arg_argc = bx.intcast(param_argc, cx.type_isize(), true);
479 let arg_argv = param_argv;
482 // The Rust start function doesn't need `argc` and `argv`, so just pass zeros.
483 let arg_argc = bx.const_int(cx.type_int(), 0);
484 let arg_argv = bx.const_null(cx.type_ptr_to(cx.type_i8p()));
489 pub fn codegen_crate<B: ExtraBackendMethods>(
493 metadata: EncodedMetadata,
494 need_metadata_module: bool,
495 ) -> OngoingCodegen<B> {
496 // Skip crate items and just output metadata in -Z no-codegen mode.
497 if tcx.sess.opts.debugging_opts.no_codegen || !tcx.sess.opts.output_types.should_codegen() {
498 let ongoing_codegen = start_async_codegen(backend, tcx, target_cpu, metadata, None, 1);
500 ongoing_codegen.codegen_finished(tcx);
502 ongoing_codegen.check_for_errors(tcx.sess);
504 return ongoing_codegen;
507 let cgu_name_builder = &mut CodegenUnitNameBuilder::new(tcx);
509 // Run the monomorphization collector and partition the collected items into
511 let codegen_units = tcx.collect_and_partition_mono_items(()).1;
513 // Force all codegen_unit queries so they are already either red or green
514 // when compile_codegen_unit accesses them. We are not able to re-execute
515 // the codegen_unit query from just the DepNode, so an unknown color would
516 // lead to having to re-execute compile_codegen_unit, possibly
518 if tcx.dep_graph.is_fully_enabled() {
519 for cgu in codegen_units {
520 tcx.ensure().codegen_unit(cgu.name());
524 let metadata_module = if need_metadata_module {
525 // Emit compressed metadata object.
526 let metadata_cgu_name =
527 cgu_name_builder.build_cgu_name(LOCAL_CRATE, &["crate"], Some("metadata")).to_string();
528 tcx.sess.time("write_compressed_metadata", || {
530 tcx.output_filenames(()).temp_path(OutputType::Metadata, Some(&metadata_cgu_name));
531 let data = create_compressed_metadata_file(
534 &exported_symbols::metadata_symbol_name(tcx),
536 if let Err(err) = std::fs::write(&file_name, data) {
537 tcx.sess.fatal(&format!("error writing metadata object file: {}", err));
539 Some(CompiledModule {
540 name: metadata_cgu_name,
541 kind: ModuleKind::Metadata,
542 object: Some(file_name),
551 let ongoing_codegen = start_async_codegen(
559 let ongoing_codegen = AbortCodegenOnDrop::<B>(Some(ongoing_codegen));
561 // Codegen an allocator shim, if necessary.
563 // If the crate doesn't have an `allocator_kind` set then there's definitely
564 // no shim to generate. Otherwise we also check our dependency graph for all
565 // our output crate types. If anything there looks like its a `Dynamic`
566 // linkage, then it's already got an allocator shim and we'll be using that
567 // one instead. If nothing exists then it's our job to generate the
569 let any_dynamic_crate = tcx.dependency_formats(()).iter().any(|(_, list)| {
570 use rustc_middle::middle::dependency_format::Linkage;
571 list.iter().any(|&linkage| linkage == Linkage::Dynamic)
573 let allocator_module = if any_dynamic_crate {
575 } else if let Some(kind) = tcx.allocator_kind(()) {
577 cgu_name_builder.build_cgu_name(LOCAL_CRATE, &["crate"], Some("allocator")).to_string();
578 let module_llvm = tcx.sess.time("write_allocator_module", || {
579 backend.codegen_allocator(tcx, &llmod_id, kind, tcx.lang_items().oom().is_some())
582 Some(ModuleCodegen { name: llmod_id, module_llvm, kind: ModuleKind::Allocator })
587 if let Some(allocator_module) = allocator_module {
588 ongoing_codegen.submit_pre_codegened_module_to_llvm(tcx, allocator_module);
591 // For better throughput during parallel processing by LLVM, we used to sort
592 // CGUs largest to smallest. This would lead to better thread utilization
593 // by, for example, preventing a large CGU from being processed last and
594 // having only one LLVM thread working while the rest remained idle.
596 // However, this strategy would lead to high memory usage, as it meant the
597 // LLVM-IR for all of the largest CGUs would be resident in memory at once.
599 // Instead, we can compromise by ordering CGUs such that the largest and
600 // smallest are first, second largest and smallest are next, etc. If there
601 // are large size variations, this can reduce memory usage significantly.
602 let codegen_units: Vec<_> = {
603 let mut sorted_cgus = codegen_units.iter().collect::<Vec<_>>();
604 sorted_cgus.sort_by_cached_key(|cgu| cgu.size_estimate());
606 let (first_half, second_half) = sorted_cgus.split_at(sorted_cgus.len() / 2);
607 second_half.iter().rev().interleave(first_half).copied().collect()
610 // The non-parallel compiler can only translate codegen units to LLVM IR
611 // on a single thread, leading to a staircase effect where the N LLVM
612 // threads have to wait on the single codegen threads to generate work
613 // for them. The parallel compiler does not have this restriction, so
614 // we can pre-load the LLVM queue in parallel before handing off
615 // coordination to the OnGoingCodegen scheduler.
617 // This likely is a temporary measure. Once we don't have to support the
618 // non-parallel compiler anymore, we can compile CGUs end-to-end in
619 // parallel and get rid of the complicated scheduling logic.
620 #[cfg(parallel_compiler)]
621 let pre_compile_cgus = |cgu_reuse: &[CguReuse]| {
622 tcx.sess.time("compile_first_CGU_batch", || {
623 // Try to find one CGU to compile per thread.
624 let cgus: Vec<_> = cgu_reuse
627 .filter(|&(_, reuse)| reuse == &CguReuse::No)
628 .take(tcx.sess.threads())
631 // Compile the found CGUs in parallel.
632 let start_time = Instant::now();
634 let pre_compiled_cgus = par_iter(cgus)
636 let module = backend.compile_codegen_unit(tcx, codegen_units[i].name());
641 (pre_compiled_cgus, start_time.elapsed())
645 #[cfg(not(parallel_compiler))]
646 let pre_compile_cgus = |_: &[CguReuse]| (FxHashMap::default(), Duration::new(0, 0));
648 let mut cgu_reuse = Vec::new();
649 let mut pre_compiled_cgus: Option<FxHashMap<usize, _>> = None;
650 let mut total_codegen_time = Duration::new(0, 0);
651 let start_rss = tcx.sess.time_passes().then(|| get_resident_set_size());
653 for (i, cgu) in codegen_units.iter().enumerate() {
654 ongoing_codegen.wait_for_signal_to_codegen_item();
655 ongoing_codegen.check_for_errors(tcx.sess);
657 // Do some setup work in the first iteration
658 if pre_compiled_cgus.is_none() {
659 // Calculate the CGU reuse
660 cgu_reuse = tcx.sess.time("find_cgu_reuse", || {
661 codegen_units.iter().map(|cgu| determine_cgu_reuse(tcx, &cgu)).collect()
663 // Pre compile some CGUs
664 let (compiled_cgus, codegen_time) = pre_compile_cgus(&cgu_reuse);
665 pre_compiled_cgus = Some(compiled_cgus);
666 total_codegen_time += codegen_time;
669 let cgu_reuse = cgu_reuse[i];
670 tcx.sess.cgu_reuse_tracker.set_actual_reuse(cgu.name().as_str(), cgu_reuse);
675 if let Some(cgu) = pre_compiled_cgus.as_mut().unwrap().remove(&i) {
678 let start_time = Instant::now();
679 let module = backend.compile_codegen_unit(tcx, cgu.name());
680 total_codegen_time += start_time.elapsed();
683 // This will unwind if there are errors, which triggers our `AbortCodegenOnDrop`
684 // guard. Unfortunately, just skipping the `submit_codegened_module_to_llvm` makes
685 // compilation hang on post-monomorphization errors.
686 tcx.sess.abort_if_errors();
688 submit_codegened_module_to_llvm(
690 &ongoing_codegen.coordinator_send,
696 CguReuse::PreLto => {
697 submit_pre_lto_module_to_llvm(
700 &ongoing_codegen.coordinator_send,
701 CachedModuleCodegen {
702 name: cgu.name().to_string(),
703 source: cgu.work_product(tcx),
708 CguReuse::PostLto => {
709 submit_post_lto_module_to_llvm(
711 &ongoing_codegen.coordinator_send,
712 CachedModuleCodegen {
713 name: cgu.name().to_string(),
714 source: cgu.work_product(tcx),
722 ongoing_codegen.codegen_finished(tcx);
724 // Since the main thread is sometimes blocked during codegen, we keep track
725 // -Ztime-passes output manually.
726 if tcx.sess.time_passes() {
727 let end_rss = get_resident_set_size();
729 print_time_passes_entry(
730 "codegen_to_LLVM_IR",
737 ongoing_codegen.check_for_errors(tcx.sess);
739 ongoing_codegen.into_inner()
742 /// A curious wrapper structure whose only purpose is to call `codegen_aborted`
743 /// when it's dropped abnormally.
745 /// In the process of working on rust-lang/rust#55238 a mysterious segfault was
746 /// stumbled upon. The segfault was never reproduced locally, but it was
747 /// suspected to be related to the fact that codegen worker threads were
748 /// sticking around by the time the main thread was exiting, causing issues.
750 /// This structure is an attempt to fix that issue where the `codegen_aborted`
751 /// message will block until all workers have finished. This should ensure that
752 /// even if the main codegen thread panics we'll wait for pending work to
753 /// complete before returning from the main thread, hopefully avoiding
756 /// If you see this comment in the code, then it means that this workaround
757 /// worked! We may yet one day track down the mysterious cause of that
759 struct AbortCodegenOnDrop<B: ExtraBackendMethods>(Option<OngoingCodegen<B>>);
761 impl<B: ExtraBackendMethods> AbortCodegenOnDrop<B> {
762 fn into_inner(mut self) -> OngoingCodegen<B> {
763 self.0.take().unwrap()
767 impl<B: ExtraBackendMethods> Deref for AbortCodegenOnDrop<B> {
768 type Target = OngoingCodegen<B>;
770 fn deref(&self) -> &OngoingCodegen<B> {
771 self.0.as_ref().unwrap()
775 impl<B: ExtraBackendMethods> DerefMut for AbortCodegenOnDrop<B> {
776 fn deref_mut(&mut self) -> &mut OngoingCodegen<B> {
777 self.0.as_mut().unwrap()
781 impl<B: ExtraBackendMethods> Drop for AbortCodegenOnDrop<B> {
783 if let Some(codegen) = self.0.take() {
784 codegen.codegen_aborted();
790 pub fn new(tcx: TyCtxt<'_>, target_cpu: String) -> CrateInfo {
791 let exported_symbols = tcx
795 .map(|&c| (c, crate::back::linker::exported_symbols(tcx, c)))
797 let linked_symbols = tcx
801 .map(|&c| (c, crate::back::linker::linked_symbols(tcx, c)))
803 let local_crate_name = tcx.crate_name(LOCAL_CRATE);
804 let crate_attrs = tcx.hir().attrs(rustc_hir::CRATE_HIR_ID);
805 let subsystem = tcx.sess.first_attr_value_str_by_name(crate_attrs, sym::windows_subsystem);
806 let windows_subsystem = subsystem.map(|subsystem| {
807 if subsystem != sym::windows && subsystem != sym::console {
808 tcx.sess.fatal(&format!(
809 "invalid windows subsystem `{}`, only \
810 `windows` and `console` are allowed",
814 subsystem.to_string()
817 // This list is used when generating the command line to pass through to
818 // system linker. The linker expects undefined symbols on the left of the
819 // command line to be defined in libraries on the right, not the other way
820 // around. For more info, see some comments in the add_used_library function
823 // In order to get this left-to-right dependency ordering, we use the reverse
824 // postorder of all crates putting the leaves at the right-most positions.
825 let used_crates = tcx
830 .filter(|&cnum| !tcx.dep_kind(cnum).macros_only())
833 let mut info = CrateInfo {
838 compiler_builtins: None,
839 profiler_runtime: None,
840 is_no_builtins: Default::default(),
841 native_libraries: Default::default(),
842 used_libraries: tcx.native_libraries(LOCAL_CRATE).iter().map(Into::into).collect(),
843 crate_name: Default::default(),
845 used_crate_source: Default::default(),
846 lang_item_to_crate: Default::default(),
847 missing_lang_items: Default::default(),
848 dependency_formats: tcx.dependency_formats(()).clone(),
850 debugger_visualizers: Default::default(),
852 let debugger_visualizers = tcx.debugger_visualizers(LOCAL_CRATE).clone();
853 if !debugger_visualizers.is_empty() {
854 info.debugger_visualizers.insert(LOCAL_CRATE, debugger_visualizers);
857 let lang_items = tcx.lang_items();
859 let crates = tcx.crates(());
861 let n_crates = crates.len();
862 info.native_libraries.reserve(n_crates);
863 info.crate_name.reserve(n_crates);
864 info.used_crate_source.reserve(n_crates);
865 info.missing_lang_items.reserve(n_crates);
867 for &cnum in crates.iter() {
868 info.native_libraries
869 .insert(cnum, tcx.native_libraries(cnum).iter().map(Into::into).collect());
870 info.crate_name.insert(cnum, tcx.crate_name(cnum));
872 let used_crate_source = tcx.used_crate_source(cnum);
873 info.used_crate_source.insert(cnum, used_crate_source.clone());
874 if tcx.is_compiler_builtins(cnum) {
875 info.compiler_builtins = Some(cnum);
877 if tcx.is_profiler_runtime(cnum) {
878 info.profiler_runtime = Some(cnum);
880 if tcx.is_no_builtins(cnum) {
881 info.is_no_builtins.insert(cnum);
883 let missing = tcx.missing_lang_items(cnum);
884 for &item in missing.iter() {
885 if let Ok(id) = lang_items.require(item) {
886 info.lang_item_to_crate.insert(item, id.krate);
890 // No need to look for lang items that don't actually need to exist.
892 missing.iter().cloned().filter(|&l| lang_items::required(tcx, l)).collect();
893 info.missing_lang_items.insert(cnum, missing);
895 // Only include debugger visualizer files from crates that will be statically linked.
896 if used_crate_source.rlib.is_some() || used_crate_source.rmeta.is_some() {
897 let debugger_visualizers = tcx.debugger_visualizers(cnum).clone();
898 if !debugger_visualizers.is_empty() {
899 info.debugger_visualizers.insert(cnum, debugger_visualizers);
908 pub fn provide(providers: &mut Providers) {
909 providers.backend_optimization_level = |tcx, cratenum| {
910 let for_speed = match tcx.sess.opts.optimize {
911 // If globally no optimisation is done, #[optimize] has no effect.
913 // This is done because if we ended up "upgrading" to `-O2` here, we’d populate the
914 // pass manager and it is likely that some module-wide passes (such as inliner or
915 // cross-function constant propagation) would ignore the `optnone` annotation we put
916 // on the functions, thus necessarily involving these functions into optimisations.
917 config::OptLevel::No => return config::OptLevel::No,
918 // If globally optimise-speed is already specified, just use that level.
919 config::OptLevel::Less => return config::OptLevel::Less,
920 config::OptLevel::Default => return config::OptLevel::Default,
921 config::OptLevel::Aggressive => return config::OptLevel::Aggressive,
922 // If globally optimize-for-size has been requested, use -O2 instead (if optimize(size)
924 config::OptLevel::Size => config::OptLevel::Default,
925 config::OptLevel::SizeMin => config::OptLevel::Default,
928 let (defids, _) = tcx.collect_and_partition_mono_items(cratenum);
930 let CodegenFnAttrs { optimize, .. } = tcx.codegen_fn_attrs(*id);
932 attr::OptimizeAttr::None => continue,
933 attr::OptimizeAttr::Size => continue,
934 attr::OptimizeAttr::Speed => {
939 tcx.sess.opts.optimize
943 fn determine_cgu_reuse<'tcx>(tcx: TyCtxt<'tcx>, cgu: &CodegenUnit<'tcx>) -> CguReuse {
944 if !tcx.dep_graph.is_fully_enabled() {
948 let work_product_id = &cgu.work_product_id();
949 if tcx.dep_graph.previous_work_product(work_product_id).is_none() {
950 // We don't have anything cached for this CGU. This can happen
951 // if the CGU did not exist in the previous session.
955 // Try to mark the CGU as green. If it we can do so, it means that nothing
956 // affecting the LLVM module has changed and we can re-use a cached version.
957 // If we compile with any kind of LTO, this means we can re-use the bitcode
958 // of the Pre-LTO stage (possibly also the Post-LTO version but we'll only
959 // know that later). If we are not doing LTO, there is only one optimized
960 // version of each module, so we re-use that.
961 let dep_node = cgu.codegen_dep_node(tcx);
963 !tcx.dep_graph.dep_node_exists(&dep_node),
964 "CompileCodegenUnit dep-node for CGU `{}` already exists before marking.",
968 if tcx.try_mark_green(&dep_node) {
969 // We can re-use either the pre- or the post-thinlto state. If no LTO is
970 // being performed then we can use post-LTO artifacts, otherwise we must
971 // reuse pre-LTO artifacts
972 match compute_per_cgu_lto_type(
975 &tcx.sess.crate_types(),
978 ComputedLtoType::No => CguReuse::PostLto,
979 _ => CguReuse::PreLto,