1 use crate::back::write::{
2 compute_per_cgu_lto_type, start_async_codegen, submit_codegened_module_to_llvm,
3 submit_post_lto_module_to_llvm, submit_pre_lto_module_to_llvm, ComputedLtoType, OngoingCodegen,
5 use crate::common::{IntPredicate, RealPredicate, TypeKind};
8 use crate::mir::operand::OperandValue;
9 use crate::mir::place::PlaceRef;
11 use crate::{CachedModuleCodegen, CrateInfo, MemFlags, ModuleCodegen, ModuleKind};
13 use rustc_attr as attr;
14 use rustc_data_structures::fx::FxHashMap;
15 use rustc_data_structures::profiling::{get_resident_set_size, print_time_passes_entry};
16 use rustc_data_structures::sync::{par_iter, ParallelIterator};
18 use rustc_hir::def_id::{DefId, LOCAL_CRATE};
19 use rustc_hir::lang_items::LangItem;
20 use rustc_index::vec::Idx;
21 use rustc_middle::middle::codegen_fn_attrs::CodegenFnAttrs;
22 use rustc_middle::middle::cstore::EncodedMetadata;
23 use rustc_middle::middle::cstore::{self, LinkagePreference};
24 use rustc_middle::middle::lang_items;
25 use rustc_middle::mir::mono::{CodegenUnit, CodegenUnitNameBuilder, MonoItem};
26 use rustc_middle::ty::layout::{HasTyCtxt, TyAndLayout};
27 use rustc_middle::ty::layout::{FAT_PTR_ADDR, FAT_PTR_EXTRA};
28 use rustc_middle::ty::query::Providers;
29 use rustc_middle::ty::{self, Instance, Ty, TyCtxt};
30 use rustc_session::cgu_reuse_tracker::CguReuse;
31 use rustc_session::config::{self, EntryFnType};
32 use rustc_session::Session;
33 use rustc_span::symbol::sym;
34 use rustc_target::abi::{Align, LayoutOf, VariantIdx};
36 use std::ops::{Deref, DerefMut};
37 use std::time::{Duration, Instant};
39 use itertools::Itertools;
41 pub fn bin_op_to_icmp_predicate(op: hir::BinOpKind, signed: bool) -> IntPredicate {
43 hir::BinOpKind::Eq => IntPredicate::IntEQ,
44 hir::BinOpKind::Ne => IntPredicate::IntNE,
45 hir::BinOpKind::Lt => {
52 hir::BinOpKind::Le => {
59 hir::BinOpKind::Gt => {
66 hir::BinOpKind::Ge => {
74 "comparison_op_to_icmp_predicate: expected comparison operator, \
81 pub fn bin_op_to_fcmp_predicate(op: hir::BinOpKind) -> RealPredicate {
83 hir::BinOpKind::Eq => RealPredicate::RealOEQ,
84 hir::BinOpKind::Ne => RealPredicate::RealUNE,
85 hir::BinOpKind::Lt => RealPredicate::RealOLT,
86 hir::BinOpKind::Le => RealPredicate::RealOLE,
87 hir::BinOpKind::Gt => RealPredicate::RealOGT,
88 hir::BinOpKind::Ge => RealPredicate::RealOGE,
91 "comparison_op_to_fcmp_predicate: expected comparison operator, \
99 pub fn compare_simd_types<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
107 let signed = match t.kind() {
109 let cmp = bin_op_to_fcmp_predicate(op);
110 let cmp = bx.fcmp(cmp, lhs, rhs);
111 return bx.sext(cmp, ret_ty);
113 ty::Uint(_) => false,
115 _ => bug!("compare_simd_types: invalid SIMD type"),
118 let cmp = bin_op_to_icmp_predicate(op, signed);
119 let cmp = bx.icmp(cmp, lhs, rhs);
120 // LLVM outputs an `< size x i1 >`, so we need to perform a sign extension
121 // to get the correctly sized type. This will compile to a single instruction
122 // once the IR is converted to assembly if the SIMD instruction is supported
123 // by the target architecture.
127 /// Retrieves the information we are losing (making dynamic) in an unsizing
130 /// The `old_info` argument is a bit odd. It is intended for use in an upcast,
131 /// where the new vtable for an object will be derived from the old one.
132 pub fn unsized_info<'tcx, Cx: CodegenMethods<'tcx>>(
136 old_info: Option<Cx::Value>,
138 let (source, target) =
139 cx.tcx().struct_lockstep_tails_erasing_lifetimes(source, target, cx.param_env());
140 match (source.kind(), target.kind()) {
141 (&ty::Array(_, len), &ty::Slice(_)) => {
142 cx.const_usize(len.eval_usize(cx.tcx(), ty::ParamEnv::reveal_all()))
144 (&ty::Dynamic(..), &ty::Dynamic(..)) => {
145 // For now, upcasts are limited to changes in marker
146 // traits, and hence never actually require an actual
147 // change to the vtable.
148 old_info.expect("unsized_info: missing old info for trait upcast")
150 (_, &ty::Dynamic(ref data, ..)) => {
151 let vtable_ptr = cx.layout_of(cx.tcx().mk_mut_ptr(target)).field(cx, FAT_PTR_EXTRA);
153 meth::get_vtable(cx, source, data.principal()),
154 cx.backend_type(vtable_ptr),
157 _ => bug!("unsized_info: invalid unsizing {:?} -> {:?}", source, target),
161 /// Coerces `src` to `dst_ty`. `src_ty` must be a thin pointer.
162 pub fn unsize_thin_ptr<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
167 ) -> (Bx::Value, Bx::Value) {
168 debug!("unsize_thin_ptr: {:?} => {:?}", src_ty, dst_ty);
169 match (src_ty.kind(), dst_ty.kind()) {
170 (&ty::Ref(_, a, _), &ty::Ref(_, b, _) | &ty::RawPtr(ty::TypeAndMut { ty: b, .. }))
171 | (&ty::RawPtr(ty::TypeAndMut { ty: a, .. }), &ty::RawPtr(ty::TypeAndMut { ty: b, .. })) => {
172 assert!(bx.cx().type_is_sized(a));
173 let ptr_ty = bx.cx().type_ptr_to(bx.cx().backend_type(bx.cx().layout_of(b)));
174 (bx.pointercast(src, ptr_ty), unsized_info(bx.cx(), a, b, None))
176 (&ty::Adt(def_a, _), &ty::Adt(def_b, _)) => {
177 assert_eq!(def_a, def_b);
179 let src_layout = bx.cx().layout_of(src_ty);
180 let dst_layout = bx.cx().layout_of(dst_ty);
181 let mut result = None;
182 for i in 0..src_layout.fields.count() {
183 let src_f = src_layout.field(bx.cx(), i);
184 assert_eq!(src_layout.fields.offset(i).bytes(), 0);
185 assert_eq!(dst_layout.fields.offset(i).bytes(), 0);
189 assert_eq!(src_layout.size, src_f.size);
191 let dst_f = dst_layout.field(bx.cx(), i);
192 assert_ne!(src_f.ty, dst_f.ty);
193 assert_eq!(result, None);
194 result = Some(unsize_thin_ptr(bx, src, src_f.ty, dst_f.ty));
196 let (lldata, llextra) = result.unwrap();
197 // HACK(eddyb) have to bitcast pointers until LLVM removes pointee types.
198 // FIXME(eddyb) move these out of this `match` arm, so they're always
199 // applied, uniformly, no matter the source/destination types.
201 bx.bitcast(lldata, bx.cx().scalar_pair_element_backend_type(dst_layout, 0, true)),
202 bx.bitcast(llextra, bx.cx().scalar_pair_element_backend_type(dst_layout, 1, true)),
205 _ => bug!("unsize_thin_ptr: called on bad types"),
209 /// Coerces `src`, which is a reference to a value of type `src_ty`,
210 /// to a value of type `dst_ty`, and stores the result in `dst`.
211 pub fn coerce_unsized_into<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
213 src: PlaceRef<'tcx, Bx::Value>,
214 dst: PlaceRef<'tcx, Bx::Value>,
216 let src_ty = src.layout.ty;
217 let dst_ty = dst.layout.ty;
218 match (src_ty.kind(), dst_ty.kind()) {
219 (&ty::Ref(..), &ty::Ref(..) | &ty::RawPtr(..)) | (&ty::RawPtr(..), &ty::RawPtr(..)) => {
220 let (base, info) = match bx.load_operand(src).val {
221 OperandValue::Pair(base, info) => {
222 // fat-ptr to fat-ptr unsize preserves the vtable
223 // i.e., &'a fmt::Debug+Send => &'a fmt::Debug
224 // So we need to pointercast the base to ensure
225 // the types match up.
226 // FIXME(eddyb) use `scalar_pair_element_backend_type` here,
227 // like `unsize_thin_ptr` does.
228 let thin_ptr = dst.layout.field(bx.cx(), FAT_PTR_ADDR);
229 (bx.pointercast(base, bx.cx().backend_type(thin_ptr)), info)
231 OperandValue::Immediate(base) => unsize_thin_ptr(bx, base, src_ty, dst_ty),
232 OperandValue::Ref(..) => bug!(),
234 OperandValue::Pair(base, info).store(bx, dst);
237 (&ty::Adt(def_a, _), &ty::Adt(def_b, _)) => {
238 assert_eq!(def_a, def_b);
240 for i in 0..def_a.variants[VariantIdx::new(0)].fields.len() {
241 let src_f = src.project_field(bx, i);
242 let dst_f = dst.project_field(bx, i);
244 if dst_f.layout.is_zst() {
248 if src_f.layout.ty == dst_f.layout.ty {
259 coerce_unsized_into(bx, src_f, dst_f);
263 _ => bug!("coerce_unsized_into: invalid coercion {:?} -> {:?}", src_ty, dst_ty,),
267 pub fn cast_shift_expr_rhs<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
273 cast_shift_rhs(bx, op, lhs, rhs)
276 fn cast_shift_rhs<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
282 // Shifts may have any size int on the rhs
284 let mut rhs_llty = bx.cx().val_ty(rhs);
285 let mut lhs_llty = bx.cx().val_ty(lhs);
286 if bx.cx().type_kind(rhs_llty) == TypeKind::Vector {
287 rhs_llty = bx.cx().element_type(rhs_llty)
289 if bx.cx().type_kind(lhs_llty) == TypeKind::Vector {
290 lhs_llty = bx.cx().element_type(lhs_llty)
292 let rhs_sz = bx.cx().int_width(rhs_llty);
293 let lhs_sz = bx.cx().int_width(lhs_llty);
295 bx.trunc(rhs, lhs_llty)
296 } else if lhs_sz > rhs_sz {
297 // FIXME (#1877: If in the future shifting by negative
298 // values is no longer undefined then this is wrong.
299 bx.zext(rhs, lhs_llty)
308 /// Returns `true` if this session's target will use SEH-based unwinding.
310 /// This is only true for MSVC targets, and even then the 64-bit MSVC target
311 /// currently uses SEH-ish unwinding with DWARF info tables to the side (same as
312 /// 64-bit MinGW) instead of "full SEH".
313 pub fn wants_msvc_seh(sess: &Session) -> bool {
314 sess.target.is_like_msvc
317 pub fn memcpy_ty<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
323 layout: TyAndLayout<'tcx>,
326 let size = layout.size.bytes();
331 bx.memcpy(dst, dst_align, src, src_align, bx.cx().const_usize(size), flags);
334 pub fn codegen_instance<'a, 'tcx: 'a, Bx: BuilderMethods<'a, 'tcx>>(
335 cx: &'a Bx::CodegenCx,
336 instance: Instance<'tcx>,
338 // this is an info! to allow collecting monomorphization statistics
339 // and to allow finding the last function before LLVM aborts from
341 info!("codegen_instance({})", instance);
343 mir::codegen_mir::<Bx>(cx, instance);
346 /// Creates the `main` function which will initialize the rust runtime and call
347 /// users main function.
348 pub fn maybe_create_entry_wrapper<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
349 cx: &'a Bx::CodegenCx,
350 ) -> Option<Bx::Function> {
351 let (main_def_id, entry_type) = cx.tcx().entry_fn(())?;
352 let main_is_local = main_def_id.is_local();
353 let instance = Instance::mono(cx.tcx(), main_def_id);
356 // We want to create the wrapper in the same codegen unit as Rust's main
358 if !cx.codegen_unit().contains_item(&MonoItem::Fn(instance)) {
361 } else if !cx.codegen_unit().is_primary() {
362 // We want to create the wrapper only when the codegen unit is the primary one
366 let main_llfn = cx.get_fn_addr(instance);
368 let use_start_lang_item = EntryFnType::Start != entry_type;
369 let entry_fn = create_entry_fn::<Bx>(cx, main_llfn, main_def_id, use_start_lang_item);
370 return Some(entry_fn);
372 fn create_entry_fn<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
373 cx: &'a Bx::CodegenCx,
374 rust_main: Bx::Value,
375 rust_main_def_id: DefId,
376 use_start_lang_item: bool,
378 // The entry function is either `int main(void)` or `int main(int argc, char **argv)`,
379 // depending on whether the target needs `argc` and `argv` to be passed in.
380 let llfty = if cx.sess().target.main_needs_argc_argv {
381 cx.type_func(&[cx.type_int(), cx.type_ptr_to(cx.type_i8p())], cx.type_int())
383 cx.type_func(&[], cx.type_int())
386 let main_ret_ty = cx.tcx().fn_sig(rust_main_def_id).output();
387 // Given that `main()` has no arguments,
388 // then its return type cannot have
389 // late-bound regions, since late-bound
390 // regions must appear in the argument
392 let main_ret_ty = cx.tcx().erase_regions(main_ret_ty.no_bound_vars().unwrap());
394 let llfn = match cx.declare_c_main(llfty) {
397 // FIXME: We should be smart and show a better diagnostic here.
398 let span = cx.tcx().def_span(rust_main_def_id);
400 .struct_span_err(span, "entry symbol `main` declared multiple times")
401 .help("did you use `#[no_mangle]` on `fn main`? Use `#[start]` instead")
403 cx.sess().abort_if_errors();
408 // `main` should respect same config for frame pointer elimination as rest of code
409 cx.set_frame_pointer_elimination(llfn);
410 cx.apply_target_cpu_attr(llfn);
412 let llbb = Bx::append_block(&cx, llfn, "top");
413 let mut bx = Bx::build(&cx, llbb);
415 bx.insert_reference_to_gdb_debug_scripts_section_global();
417 let (arg_argc, arg_argv) = get_argc_argv(cx, &mut bx);
419 let (start_fn, args) = if use_start_lang_item {
420 let start_def_id = cx.tcx().require_lang_item(LangItem::Start, None);
421 let start_fn = cx.get_fn_addr(
422 ty::Instance::resolve(
424 ty::ParamEnv::reveal_all(),
426 cx.tcx().intern_substs(&[main_ret_ty.into()]),
433 vec![bx.pointercast(rust_main, cx.type_ptr_to(cx.type_i8p())), arg_argc, arg_argv],
436 debug!("using user-defined start fn");
437 (rust_main, vec![arg_argc, arg_argv])
440 let result = bx.call(start_fn, &args, None);
441 let cast = bx.intcast(result, cx.type_int(), true);
448 /// Obtain the `argc` and `argv` values to pass to the rust start function.
449 fn get_argc_argv<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
450 cx: &'a Bx::CodegenCx,
452 ) -> (Bx::Value, Bx::Value) {
453 if cx.sess().target.main_needs_argc_argv {
454 // Params from native `main()` used as args for rust start function
455 let param_argc = bx.get_param(0);
456 let param_argv = bx.get_param(1);
457 let arg_argc = bx.intcast(param_argc, cx.type_isize(), true);
458 let arg_argv = param_argv;
461 // The Rust start function doesn't need `argc` and `argv`, so just pass zeros.
462 let arg_argc = bx.const_int(cx.type_int(), 0);
463 let arg_argv = bx.const_null(cx.type_ptr_to(cx.type_i8p()));
468 pub fn codegen_crate<B: ExtraBackendMethods>(
472 metadata: EncodedMetadata,
473 need_metadata_module: bool,
474 ) -> OngoingCodegen<B> {
475 // Skip crate items and just output metadata in -Z no-codegen mode.
476 if tcx.sess.opts.debugging_opts.no_codegen || !tcx.sess.opts.output_types.should_codegen() {
477 let ongoing_codegen = start_async_codegen(backend, tcx, target_cpu, metadata, 1);
479 ongoing_codegen.codegen_finished(tcx);
481 ongoing_codegen.check_for_errors(tcx.sess);
483 return ongoing_codegen;
486 let cgu_name_builder = &mut CodegenUnitNameBuilder::new(tcx);
488 // Run the monomorphization collector and partition the collected items into
490 let codegen_units = tcx.collect_and_partition_mono_items(()).1;
492 // Force all codegen_unit queries so they are already either red or green
493 // when compile_codegen_unit accesses them. We are not able to re-execute
494 // the codegen_unit query from just the DepNode, so an unknown color would
495 // lead to having to re-execute compile_codegen_unit, possibly
497 if tcx.dep_graph.is_fully_enabled() {
498 for cgu in codegen_units {
499 tcx.ensure().codegen_unit(cgu.name());
503 let ongoing_codegen =
504 start_async_codegen(backend.clone(), tcx, target_cpu, metadata, codegen_units.len());
505 let ongoing_codegen = AbortCodegenOnDrop::<B>(Some(ongoing_codegen));
507 // Codegen an allocator shim, if necessary.
509 // If the crate doesn't have an `allocator_kind` set then there's definitely
510 // no shim to generate. Otherwise we also check our dependency graph for all
511 // our output crate types. If anything there looks like its a `Dynamic`
512 // linkage, then it's already got an allocator shim and we'll be using that
513 // one instead. If nothing exists then it's our job to generate the
515 let any_dynamic_crate = tcx.dependency_formats(()).iter().any(|(_, list)| {
516 use rustc_middle::middle::dependency_format::Linkage;
517 list.iter().any(|&linkage| linkage == Linkage::Dynamic)
519 let allocator_module = if any_dynamic_crate {
521 } else if let Some(kind) = tcx.allocator_kind() {
523 cgu_name_builder.build_cgu_name(LOCAL_CRATE, &["crate"], Some("allocator")).to_string();
524 let mut modules = backend.new_metadata(tcx, &llmod_id);
525 tcx.sess.time("write_allocator_module", || {
526 backend.codegen_allocator(tcx, &mut modules, kind, tcx.lang_items().oom().is_some())
529 Some(ModuleCodegen { name: llmod_id, module_llvm: modules, kind: ModuleKind::Allocator })
534 if let Some(allocator_module) = allocator_module {
535 ongoing_codegen.submit_pre_codegened_module_to_llvm(tcx, allocator_module);
538 if need_metadata_module {
539 // Codegen the encoded metadata.
540 let metadata_cgu_name =
541 cgu_name_builder.build_cgu_name(LOCAL_CRATE, &["crate"], Some("metadata")).to_string();
542 let mut metadata_llvm_module = backend.new_metadata(tcx, &metadata_cgu_name);
543 tcx.sess.time("write_compressed_metadata", || {
544 backend.write_compressed_metadata(
546 &ongoing_codegen.metadata,
547 &mut metadata_llvm_module,
551 let metadata_module = ModuleCodegen {
552 name: metadata_cgu_name,
553 module_llvm: metadata_llvm_module,
554 kind: ModuleKind::Metadata,
556 ongoing_codegen.submit_pre_codegened_module_to_llvm(tcx, metadata_module);
559 // For better throughput during parallel processing by LLVM, we used to sort
560 // CGUs largest to smallest. This would lead to better thread utilization
561 // by, for example, preventing a large CGU from being processed last and
562 // having only one LLVM thread working while the rest remained idle.
564 // However, this strategy would lead to high memory usage, as it meant the
565 // LLVM-IR for all of the largest CGUs would be resident in memory at once.
567 // Instead, we can compromise by ordering CGUs such that the largest and
568 // smallest are first, second largest and smallest are next, etc. If there
569 // are large size variations, this can reduce memory usage significantly.
570 let codegen_units: Vec<_> = {
571 let mut sorted_cgus = codegen_units.iter().collect::<Vec<_>>();
572 sorted_cgus.sort_by_cached_key(|cgu| cgu.size_estimate());
574 let (first_half, second_half) = sorted_cgus.split_at(sorted_cgus.len() / 2);
575 second_half.iter().rev().interleave(first_half).copied().collect()
578 // The non-parallel compiler can only translate codegen units to LLVM IR
579 // on a single thread, leading to a staircase effect where the N LLVM
580 // threads have to wait on the single codegen threads to generate work
581 // for them. The parallel compiler does not have this restriction, so
582 // we can pre-load the LLVM queue in parallel before handing off
583 // coordination to the OnGoingCodegen scheduler.
585 // This likely is a temporary measure. Once we don't have to support the
586 // non-parallel compiler anymore, we can compile CGUs end-to-end in
587 // parallel and get rid of the complicated scheduling logic.
588 let pre_compile_cgus = |cgu_reuse: &[CguReuse]| {
589 if cfg!(parallel_compiler) {
590 tcx.sess.time("compile_first_CGU_batch", || {
591 // Try to find one CGU to compile per thread.
592 let cgus: Vec<_> = cgu_reuse
595 .filter(|&(_, reuse)| reuse == &CguReuse::No)
596 .take(tcx.sess.threads())
599 // Compile the found CGUs in parallel.
600 let start_time = Instant::now();
602 let pre_compiled_cgus = par_iter(cgus)
604 let module = backend.compile_codegen_unit(tcx, codegen_units[i].name());
609 (pre_compiled_cgus, start_time.elapsed())
612 (FxHashMap::default(), Duration::new(0, 0))
616 let mut cgu_reuse = Vec::new();
617 let mut pre_compiled_cgus: Option<FxHashMap<usize, _>> = None;
618 let mut total_codegen_time = Duration::new(0, 0);
619 let start_rss = tcx.sess.time_passes().then(|| get_resident_set_size());
621 for (i, cgu) in codegen_units.iter().enumerate() {
622 ongoing_codegen.wait_for_signal_to_codegen_item();
623 ongoing_codegen.check_for_errors(tcx.sess);
625 // Do some setup work in the first iteration
626 if pre_compiled_cgus.is_none() {
627 // Calculate the CGU reuse
628 cgu_reuse = tcx.sess.time("find_cgu_reuse", || {
629 codegen_units.iter().map(|cgu| determine_cgu_reuse(tcx, &cgu)).collect()
631 // Pre compile some CGUs
632 let (compiled_cgus, codegen_time) = pre_compile_cgus(&cgu_reuse);
633 pre_compiled_cgus = Some(compiled_cgus);
634 total_codegen_time += codegen_time;
637 let cgu_reuse = cgu_reuse[i];
638 tcx.sess.cgu_reuse_tracker.set_actual_reuse(&cgu.name().as_str(), cgu_reuse);
643 if let Some(cgu) = pre_compiled_cgus.as_mut().unwrap().remove(&i) {
646 let start_time = Instant::now();
647 let module = backend.compile_codegen_unit(tcx, cgu.name());
648 total_codegen_time += start_time.elapsed();
651 // This will unwind if there are errors, which triggers our `AbortCodegenOnDrop`
652 // guard. Unfortunately, just skipping the `submit_codegened_module_to_llvm` makes
653 // compilation hang on post-monomorphization errors.
654 tcx.sess.abort_if_errors();
656 submit_codegened_module_to_llvm(
658 &ongoing_codegen.coordinator_send,
664 CguReuse::PreLto => {
665 submit_pre_lto_module_to_llvm(
668 &ongoing_codegen.coordinator_send,
669 CachedModuleCodegen {
670 name: cgu.name().to_string(),
671 source: cgu.work_product(tcx),
676 CguReuse::PostLto => {
677 submit_post_lto_module_to_llvm(
679 &ongoing_codegen.coordinator_send,
680 CachedModuleCodegen {
681 name: cgu.name().to_string(),
682 source: cgu.work_product(tcx),
690 ongoing_codegen.codegen_finished(tcx);
692 // Since the main thread is sometimes blocked during codegen, we keep track
693 // -Ztime-passes output manually.
694 if tcx.sess.time_passes() {
695 let end_rss = get_resident_set_size();
697 print_time_passes_entry(
698 "codegen_to_LLVM_IR",
705 ongoing_codegen.check_for_errors(tcx.sess);
707 ongoing_codegen.into_inner()
710 /// A curious wrapper structure whose only purpose is to call `codegen_aborted`
711 /// when it's dropped abnormally.
713 /// In the process of working on rust-lang/rust#55238 a mysterious segfault was
714 /// stumbled upon. The segfault was never reproduced locally, but it was
715 /// suspected to be related to the fact that codegen worker threads were
716 /// sticking around by the time the main thread was exiting, causing issues.
718 /// This structure is an attempt to fix that issue where the `codegen_aborted`
719 /// message will block until all workers have finished. This should ensure that
720 /// even if the main codegen thread panics we'll wait for pending work to
721 /// complete before returning from the main thread, hopefully avoiding
724 /// If you see this comment in the code, then it means that this workaround
725 /// worked! We may yet one day track down the mysterious cause of that
727 struct AbortCodegenOnDrop<B: ExtraBackendMethods>(Option<OngoingCodegen<B>>);
729 impl<B: ExtraBackendMethods> AbortCodegenOnDrop<B> {
730 fn into_inner(mut self) -> OngoingCodegen<B> {
731 self.0.take().unwrap()
735 impl<B: ExtraBackendMethods> Deref for AbortCodegenOnDrop<B> {
736 type Target = OngoingCodegen<B>;
738 fn deref(&self) -> &OngoingCodegen<B> {
739 self.0.as_ref().unwrap()
743 impl<B: ExtraBackendMethods> DerefMut for AbortCodegenOnDrop<B> {
744 fn deref_mut(&mut self) -> &mut OngoingCodegen<B> {
745 self.0.as_mut().unwrap()
749 impl<B: ExtraBackendMethods> Drop for AbortCodegenOnDrop<B> {
751 if let Some(codegen) = self.0.take() {
752 codegen.codegen_aborted();
758 pub fn new(tcx: TyCtxt<'_>) -> CrateInfo {
759 let local_crate_name = tcx.crate_name(LOCAL_CRATE);
760 let crate_attrs = tcx.hir().attrs(rustc_hir::CRATE_HIR_ID);
761 let subsystem = tcx.sess.first_attr_value_str_by_name(crate_attrs, sym::windows_subsystem);
762 let windows_subsystem = subsystem.map(|subsystem| {
763 if subsystem != sym::windows && subsystem != sym::console {
764 tcx.sess.fatal(&format!(
765 "invalid windows subsystem `{}`, only \
766 `windows` and `console` are allowed",
770 subsystem.to_string()
773 let mut info = CrateInfo {
776 compiler_builtins: None,
777 profiler_runtime: None,
778 is_no_builtins: Default::default(),
779 native_libraries: Default::default(),
780 used_libraries: tcx.native_libraries(LOCAL_CRATE).iter().map(Into::into).collect(),
781 crate_name: Default::default(),
782 used_crates_dynamic: cstore::used_crates(tcx, LinkagePreference::RequireDynamic),
783 used_crates_static: cstore::used_crates(tcx, LinkagePreference::RequireStatic),
784 used_crate_source: Default::default(),
785 lang_item_to_crate: Default::default(),
786 missing_lang_items: Default::default(),
787 dependency_formats: tcx.dependency_formats(()),
790 let lang_items = tcx.lang_items();
792 let crates = tcx.crates();
794 let n_crates = crates.len();
795 info.native_libraries.reserve(n_crates);
796 info.crate_name.reserve(n_crates);
797 info.used_crate_source.reserve(n_crates);
798 info.missing_lang_items.reserve(n_crates);
800 for &cnum in crates.iter() {
801 info.native_libraries
802 .insert(cnum, tcx.native_libraries(cnum).iter().map(Into::into).collect());
803 info.crate_name.insert(cnum, tcx.crate_name(cnum).to_string());
804 info.used_crate_source.insert(cnum, tcx.used_crate_source(cnum));
805 if tcx.is_panic_runtime(cnum) {
806 info.panic_runtime = Some(cnum);
808 if tcx.is_compiler_builtins(cnum) {
809 info.compiler_builtins = Some(cnum);
811 if tcx.is_profiler_runtime(cnum) {
812 info.profiler_runtime = Some(cnum);
814 if tcx.is_no_builtins(cnum) {
815 info.is_no_builtins.insert(cnum);
817 let missing = tcx.missing_lang_items(cnum);
818 for &item in missing.iter() {
819 if let Ok(id) = lang_items.require(item) {
820 info.lang_item_to_crate.insert(item, id.krate);
824 // No need to look for lang items that don't actually need to exist.
826 missing.iter().cloned().filter(|&l| lang_items::required(tcx, l)).collect();
827 info.missing_lang_items.insert(cnum, missing);
834 pub fn provide(providers: &mut Providers) {
835 providers.backend_optimization_level = |tcx, cratenum| {
836 let for_speed = match tcx.sess.opts.optimize {
837 // If globally no optimisation is done, #[optimize] has no effect.
839 // This is done because if we ended up "upgrading" to `-O2` here, we’d populate the
840 // pass manager and it is likely that some module-wide passes (such as inliner or
841 // cross-function constant propagation) would ignore the `optnone` annotation we put
842 // on the functions, thus necessarily involving these functions into optimisations.
843 config::OptLevel::No => return config::OptLevel::No,
844 // If globally optimise-speed is already specified, just use that level.
845 config::OptLevel::Less => return config::OptLevel::Less,
846 config::OptLevel::Default => return config::OptLevel::Default,
847 config::OptLevel::Aggressive => return config::OptLevel::Aggressive,
848 // If globally optimize-for-size has been requested, use -O2 instead (if optimize(size)
850 config::OptLevel::Size => config::OptLevel::Default,
851 config::OptLevel::SizeMin => config::OptLevel::Default,
854 let (defids, _) = tcx.collect_and_partition_mono_items(cratenum);
856 let CodegenFnAttrs { optimize, .. } = tcx.codegen_fn_attrs(*id);
858 attr::OptimizeAttr::None => continue,
859 attr::OptimizeAttr::Size => continue,
860 attr::OptimizeAttr::Speed => {
865 tcx.sess.opts.optimize
869 fn determine_cgu_reuse<'tcx>(tcx: TyCtxt<'tcx>, cgu: &CodegenUnit<'tcx>) -> CguReuse {
870 if !tcx.dep_graph.is_fully_enabled() {
874 let work_product_id = &cgu.work_product_id();
875 if tcx.dep_graph.previous_work_product(work_product_id).is_none() {
876 // We don't have anything cached for this CGU. This can happen
877 // if the CGU did not exist in the previous session.
881 // Try to mark the CGU as green. If it we can do so, it means that nothing
882 // affecting the LLVM module has changed and we can re-use a cached version.
883 // If we compile with any kind of LTO, this means we can re-use the bitcode
884 // of the Pre-LTO stage (possibly also the Post-LTO version but we'll only
885 // know that later). If we are not doing LTO, there is only one optimized
886 // version of each module, so we re-use that.
887 let dep_node = cgu.codegen_dep_node(tcx);
889 !tcx.dep_graph.dep_node_exists(&dep_node),
890 "CompileCodegenUnit dep-node for CGU `{}` already exists before marking.",
894 if tcx.try_mark_green(&dep_node) {
895 // We can re-use either the pre- or the post-thinlto state. If no LTO is
896 // being performed then we can use post-LTO artifacts, otherwise we must
897 // reuse pre-LTO artifacts
898 match compute_per_cgu_lto_type(
901 &tcx.sess.crate_types(),
904 ComputedLtoType::No => CguReuse::PostLto,
905 _ => CguReuse::PreLto,