1 //! Codegen the completed AST to the LLVM IR.
3 //! Some functions here, such as `codegen_block` and `codegen_expr`, return a value --
4 //! the result of the codegen to LLVM -- while others, such as `codegen_fn`
5 //! and `mono_item`, are called only for the side effect of adding a
6 //! particular definition to the LLVM IR output we're producing.
8 //! Hopefully useful general knowledge about codegen:
10 //! * There's no way to find out the `Ty` type of a `Value`. Doing so
11 //! would be "trying to get the eggs out of an omelette" (credit:
12 //! pcwalton). You can, instead, find out its `llvm::Type` by calling `val_ty`,
13 //! but one `llvm::Type` corresponds to many `Ty`s; for instance, `tup(int, int,
14 //! int)` and `rec(x=int, y=int, z=int)` will have the same `llvm::Type`.
16 use crate::back::write::{
17 compute_per_cgu_lto_type, start_async_codegen, submit_codegened_module_to_llvm,
18 submit_post_lto_module_to_llvm, submit_pre_lto_module_to_llvm, ComputedLtoType, OngoingCodegen,
20 use crate::common::{IntPredicate, RealPredicate, TypeKind};
23 use crate::mir::operand::OperandValue;
24 use crate::mir::place::PlaceRef;
26 use crate::{CachedModuleCodegen, CrateInfo, MemFlags, ModuleCodegen, ModuleKind};
28 use rustc_attr as attr;
29 use rustc_data_structures::fx::FxHashMap;
30 use rustc_data_structures::profiling::print_time_passes_entry;
31 use rustc_data_structures::sync::{par_iter, Lock, ParallelIterator};
33 use rustc_hir::def_id::{LocalDefId, LOCAL_CRATE};
34 use rustc_hir::lang_items::LangItem;
35 use rustc_index::vec::Idx;
36 use rustc_middle::middle::codegen_fn_attrs::CodegenFnAttrs;
37 use rustc_middle::middle::cstore::EncodedMetadata;
38 use rustc_middle::middle::cstore::{self, LinkagePreference};
39 use rustc_middle::middle::lang_items;
40 use rustc_middle::mir::mono::{CodegenUnit, CodegenUnitNameBuilder, MonoItem};
41 use rustc_middle::ty::layout::{self, HasTyCtxt, TyAndLayout};
42 use rustc_middle::ty::layout::{FAT_PTR_ADDR, FAT_PTR_EXTRA};
43 use rustc_middle::ty::query::Providers;
44 use rustc_middle::ty::{self, Instance, Ty, TyCtxt};
45 use rustc_session::cgu_reuse_tracker::CguReuse;
46 use rustc_session::config::{self, EntryFnType};
47 use rustc_session::utils::NativeLibKind;
48 use rustc_session::Session;
50 use rustc_symbol_mangling::test as symbol_names_test;
51 use rustc_target::abi::{Abi, Align, LayoutOf, Scalar, VariantIdx};
54 use std::ops::{Deref, DerefMut};
55 use std::time::{Duration, Instant};
57 pub fn bin_op_to_icmp_predicate(op: hir::BinOpKind, signed: bool) -> IntPredicate {
59 hir::BinOpKind::Eq => IntPredicate::IntEQ,
60 hir::BinOpKind::Ne => IntPredicate::IntNE,
61 hir::BinOpKind::Lt => {
68 hir::BinOpKind::Le => {
75 hir::BinOpKind::Gt => {
82 hir::BinOpKind::Ge => {
90 "comparison_op_to_icmp_predicate: expected comparison operator, \
97 pub fn bin_op_to_fcmp_predicate(op: hir::BinOpKind) -> RealPredicate {
99 hir::BinOpKind::Eq => RealPredicate::RealOEQ,
100 hir::BinOpKind::Ne => RealPredicate::RealUNE,
101 hir::BinOpKind::Lt => RealPredicate::RealOLT,
102 hir::BinOpKind::Le => RealPredicate::RealOLE,
103 hir::BinOpKind::Gt => RealPredicate::RealOGT,
104 hir::BinOpKind::Ge => RealPredicate::RealOGE,
107 "comparison_op_to_fcmp_predicate: expected comparison operator, \
115 pub fn compare_simd_types<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
123 let signed = match t.kind {
125 let cmp = bin_op_to_fcmp_predicate(op);
126 let cmp = bx.fcmp(cmp, lhs, rhs);
127 return bx.sext(cmp, ret_ty);
129 ty::Uint(_) => false,
131 _ => bug!("compare_simd_types: invalid SIMD type"),
134 let cmp = bin_op_to_icmp_predicate(op, signed);
135 let cmp = bx.icmp(cmp, lhs, rhs);
136 // LLVM outputs an `< size x i1 >`, so we need to perform a sign extension
137 // to get the correctly sized type. This will compile to a single instruction
138 // once the IR is converted to assembly if the SIMD instruction is supported
139 // by the target architecture.
143 /// Retrieves the information we are losing (making dynamic) in an unsizing
146 /// The `old_info` argument is a bit odd. It is intended for use in an upcast,
147 /// where the new vtable for an object will be derived from the old one.
148 pub fn unsized_info<'tcx, Cx: CodegenMethods<'tcx>>(
152 old_info: Option<Cx::Value>,
154 let (source, target) =
155 cx.tcx().struct_lockstep_tails_erasing_lifetimes(source, target, cx.param_env());
156 match (&source.kind, &target.kind) {
157 (&ty::Array(_, len), &ty::Slice(_)) => {
158 cx.const_usize(len.eval_usize(cx.tcx(), ty::ParamEnv::reveal_all()))
160 (&ty::Dynamic(..), &ty::Dynamic(..)) => {
161 // For now, upcasts are limited to changes in marker
162 // traits, and hence never actually require an actual
163 // change to the vtable.
164 old_info.expect("unsized_info: missing old info for trait upcast")
166 (_, &ty::Dynamic(ref data, ..)) => {
167 let vtable_ptr = cx.layout_of(cx.tcx().mk_mut_ptr(target)).field(cx, FAT_PTR_EXTRA);
169 meth::get_vtable(cx, source, data.principal()),
170 cx.backend_type(vtable_ptr),
173 _ => bug!("unsized_info: invalid unsizing {:?} -> {:?}", source, target),
177 /// Coerces `src` to `dst_ty`. `src_ty` must be a thin pointer.
178 pub fn unsize_thin_ptr<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
183 ) -> (Bx::Value, Bx::Value) {
184 debug!("unsize_thin_ptr: {:?} => {:?}", src_ty, dst_ty);
185 match (&src_ty.kind, &dst_ty.kind) {
186 (&ty::Ref(_, a, _), &ty::Ref(_, b, _) | &ty::RawPtr(ty::TypeAndMut { ty: b, .. }))
187 | (&ty::RawPtr(ty::TypeAndMut { ty: a, .. }), &ty::RawPtr(ty::TypeAndMut { ty: b, .. })) => {
188 assert!(bx.cx().type_is_sized(a));
189 let ptr_ty = bx.cx().type_ptr_to(bx.cx().backend_type(bx.cx().layout_of(b)));
190 (bx.pointercast(src, ptr_ty), unsized_info(bx.cx(), a, b, None))
192 (&ty::Adt(def_a, _), &ty::Adt(def_b, _)) => {
193 assert_eq!(def_a, def_b);
195 let src_layout = bx.cx().layout_of(src_ty);
196 let dst_layout = bx.cx().layout_of(dst_ty);
197 let mut result = None;
198 for i in 0..src_layout.fields.count() {
199 let src_f = src_layout.field(bx.cx(), i);
200 assert_eq!(src_layout.fields.offset(i).bytes(), 0);
201 assert_eq!(dst_layout.fields.offset(i).bytes(), 0);
205 assert_eq!(src_layout.size, src_f.size);
207 let dst_f = dst_layout.field(bx.cx(), i);
208 assert_ne!(src_f.ty, dst_f.ty);
209 assert_eq!(result, None);
210 result = Some(unsize_thin_ptr(bx, src, src_f.ty, dst_f.ty));
212 let (lldata, llextra) = result.unwrap();
213 // HACK(eddyb) have to bitcast pointers until LLVM removes pointee types.
214 // FIXME(eddyb) move these out of this `match` arm, so they're always
215 // applied, uniformly, no matter the source/destination types.
217 bx.bitcast(lldata, bx.cx().scalar_pair_element_backend_type(dst_layout, 0, true)),
218 bx.bitcast(llextra, bx.cx().scalar_pair_element_backend_type(dst_layout, 1, true)),
221 _ => bug!("unsize_thin_ptr: called on bad types"),
225 /// Coerces `src`, which is a reference to a value of type `src_ty`,
226 /// to a value of type `dst_ty`, and stores the result in `dst`.
227 pub fn coerce_unsized_into<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
229 src: PlaceRef<'tcx, Bx::Value>,
230 dst: PlaceRef<'tcx, Bx::Value>,
232 let src_ty = src.layout.ty;
233 let dst_ty = dst.layout.ty;
234 match (&src_ty.kind, &dst_ty.kind) {
235 (&ty::Ref(..), &ty::Ref(..) | &ty::RawPtr(..)) | (&ty::RawPtr(..), &ty::RawPtr(..)) => {
236 let (base, info) = match bx.load_operand(src).val {
237 OperandValue::Pair(base, info) => {
238 // fat-ptr to fat-ptr unsize preserves the vtable
239 // i.e., &'a fmt::Debug+Send => &'a fmt::Debug
240 // So we need to pointercast the base to ensure
241 // the types match up.
242 // FIXME(eddyb) use `scalar_pair_element_backend_type` here,
243 // like `unsize_thin_ptr` does.
244 let thin_ptr = dst.layout.field(bx.cx(), FAT_PTR_ADDR);
245 (bx.pointercast(base, bx.cx().backend_type(thin_ptr)), info)
247 OperandValue::Immediate(base) => unsize_thin_ptr(bx, base, src_ty, dst_ty),
248 OperandValue::Ref(..) => bug!(),
250 OperandValue::Pair(base, info).store(bx, dst);
253 (&ty::Adt(def_a, _), &ty::Adt(def_b, _)) => {
254 assert_eq!(def_a, def_b);
256 for i in 0..def_a.variants[VariantIdx::new(0)].fields.len() {
257 let src_f = src.project_field(bx, i);
258 let dst_f = dst.project_field(bx, i);
260 if dst_f.layout.is_zst() {
264 if src_f.layout.ty == dst_f.layout.ty {
275 coerce_unsized_into(bx, src_f, dst_f);
279 _ => bug!("coerce_unsized_into: invalid coercion {:?} -> {:?}", src_ty, dst_ty,),
283 pub fn cast_shift_expr_rhs<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
289 cast_shift_rhs(bx, op, lhs, rhs)
292 fn cast_shift_rhs<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
298 // Shifts may have any size int on the rhs
300 let mut rhs_llty = bx.cx().val_ty(rhs);
301 let mut lhs_llty = bx.cx().val_ty(lhs);
302 if bx.cx().type_kind(rhs_llty) == TypeKind::Vector {
303 rhs_llty = bx.cx().element_type(rhs_llty)
305 if bx.cx().type_kind(lhs_llty) == TypeKind::Vector {
306 lhs_llty = bx.cx().element_type(lhs_llty)
308 let rhs_sz = bx.cx().int_width(rhs_llty);
309 let lhs_sz = bx.cx().int_width(lhs_llty);
311 bx.trunc(rhs, lhs_llty)
312 } else if lhs_sz > rhs_sz {
313 // FIXME (#1877: If in the future shifting by negative
314 // values is no longer undefined then this is wrong.
315 bx.zext(rhs, lhs_llty)
324 /// Returns `true` if this session's target will use SEH-based unwinding.
326 /// This is only true for MSVC targets, and even then the 64-bit MSVC target
327 /// currently uses SEH-ish unwinding with DWARF info tables to the side (same as
328 /// 64-bit MinGW) instead of "full SEH".
329 pub fn wants_msvc_seh(sess: &Session) -> bool {
330 sess.target.target.options.is_like_msvc
333 pub fn from_immediate<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
337 if bx.cx().val_ty(val) == bx.cx().type_i1() { bx.zext(val, bx.cx().type_i8()) } else { val }
340 pub fn to_immediate<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
343 layout: layout::TyAndLayout<'_>,
345 if let Abi::Scalar(ref scalar) = layout.abi {
346 return to_immediate_scalar(bx, val, scalar);
351 pub fn to_immediate_scalar<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
356 if scalar.is_bool() {
357 return bx.trunc(val, bx.cx().type_i1());
362 pub fn memcpy_ty<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
368 layout: TyAndLayout<'tcx>,
371 let size = layout.size.bytes();
376 bx.memcpy(dst, dst_align, src, src_align, bx.cx().const_usize(size), flags);
379 pub fn codegen_instance<'a, 'tcx: 'a, Bx: BuilderMethods<'a, 'tcx>>(
380 cx: &'a Bx::CodegenCx,
381 instance: Instance<'tcx>,
383 // this is an info! to allow collecting monomorphization statistics
384 // and to allow finding the last function before LLVM aborts from
386 info!("codegen_instance({})", instance);
388 mir::codegen_mir::<Bx>(cx, instance);
391 /// Creates the `main` function which will initialize the rust runtime and call
392 /// users main function.
393 pub fn maybe_create_entry_wrapper<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
394 cx: &'a Bx::CodegenCx,
395 ) -> Option<Bx::Function> {
396 let (main_def_id, span) = match cx.tcx().entry_fn(LOCAL_CRATE) {
397 Some((def_id, _)) => (def_id, cx.tcx().def_span(def_id)),
401 let instance = Instance::mono(cx.tcx(), main_def_id.to_def_id());
403 if !cx.codegen_unit().contains_item(&MonoItem::Fn(instance)) {
404 // We want to create the wrapper in the same codegen unit as Rust's main
409 let main_llfn = cx.get_fn_addr(instance);
411 return cx.tcx().entry_fn(LOCAL_CRATE).map(|(_, et)| {
412 let use_start_lang_item = EntryFnType::Start != et;
413 create_entry_fn::<Bx>(cx, span, main_llfn, main_def_id, use_start_lang_item)
416 fn create_entry_fn<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
417 cx: &'a Bx::CodegenCx,
419 rust_main: Bx::Value,
420 rust_main_def_id: LocalDefId,
421 use_start_lang_item: bool,
423 // The entry function is either `int main(void)` or `int main(int argc, char **argv)`,
424 // depending on whether the target needs `argc` and `argv` to be passed in.
425 let llfty = if cx.sess().target.target.options.main_needs_argc_argv {
426 cx.type_func(&[cx.type_int(), cx.type_ptr_to(cx.type_i8p())], cx.type_int())
428 cx.type_func(&[], cx.type_int())
431 let main_ret_ty = cx.tcx().fn_sig(rust_main_def_id).output();
432 // Given that `main()` has no arguments,
433 // then its return type cannot have
434 // late-bound regions, since late-bound
435 // regions must appear in the argument
437 let main_ret_ty = cx.tcx().erase_regions(&main_ret_ty.no_bound_vars().unwrap());
439 if cx.get_declared_value("main").is_some() {
440 // FIXME: We should be smart and show a better diagnostic here.
442 .struct_span_err(sp, "entry symbol `main` declared multiple times")
443 .help("did you use `#[no_mangle]` on `fn main`? Use `#[start]` instead")
445 cx.sess().abort_if_errors();
448 let llfn = cx.declare_cfn("main", llfty);
450 // `main` should respect same config for frame pointer elimination as rest of code
451 cx.set_frame_pointer_elimination(llfn);
452 cx.apply_target_cpu_attr(llfn);
454 let mut bx = Bx::new_block(&cx, llfn, "top");
456 bx.insert_reference_to_gdb_debug_scripts_section_global();
458 let (arg_argc, arg_argv) = get_argc_argv(cx, &mut bx);
460 let (start_fn, args) = if use_start_lang_item {
461 let start_def_id = cx.tcx().require_lang_item(LangItem::Start, None);
462 let start_fn = cx.get_fn_addr(
463 ty::Instance::resolve(
465 ty::ParamEnv::reveal_all(),
467 cx.tcx().intern_substs(&[main_ret_ty.into()]),
474 vec![bx.pointercast(rust_main, cx.type_ptr_to(cx.type_i8p())), arg_argc, arg_argv],
477 debug!("using user-defined start fn");
478 (rust_main, vec![arg_argc, arg_argv])
481 let result = bx.call(start_fn, &args, None);
482 let cast = bx.intcast(result, cx.type_int(), true);
489 /// Obtain the `argc` and `argv` values to pass to the rust start function.
490 fn get_argc_argv<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
491 cx: &'a Bx::CodegenCx,
493 ) -> (Bx::Value, Bx::Value) {
494 if cx.sess().target.target.options.main_needs_argc_argv {
495 // Params from native `main()` used as args for rust start function
496 let param_argc = bx.get_param(0);
497 let param_argv = bx.get_param(1);
498 let arg_argc = bx.intcast(param_argc, cx.type_isize(), true);
499 let arg_argv = param_argv;
502 // The Rust start function doesn't need `argc` and `argv`, so just pass zeros.
503 let arg_argc = bx.const_int(cx.type_int(), 0);
504 let arg_argv = bx.const_null(cx.type_ptr_to(cx.type_i8p()));
509 pub const CODEGEN_WORKER_ID: usize = usize::MAX;
511 pub fn codegen_crate<B: ExtraBackendMethods>(
514 metadata: EncodedMetadata,
515 need_metadata_module: bool,
516 ) -> OngoingCodegen<B> {
517 // Skip crate items and just output metadata in -Z no-codegen mode.
518 if tcx.sess.opts.debugging_opts.no_codegen || !tcx.sess.opts.output_types.should_codegen() {
519 let ongoing_codegen = start_async_codegen(backend, tcx, metadata, 1);
521 ongoing_codegen.codegen_finished(tcx);
525 ongoing_codegen.check_for_errors(tcx.sess);
527 return ongoing_codegen;
530 let cgu_name_builder = &mut CodegenUnitNameBuilder::new(tcx);
532 // Run the monomorphization collector and partition the collected items into
534 let codegen_units = tcx.collect_and_partition_mono_items(LOCAL_CRATE).1;
536 // Force all codegen_unit queries so they are already either red or green
537 // when compile_codegen_unit accesses them. We are not able to re-execute
538 // the codegen_unit query from just the DepNode, so an unknown color would
539 // lead to having to re-execute compile_codegen_unit, possibly
541 if tcx.dep_graph.is_fully_enabled() {
542 for cgu in codegen_units {
543 tcx.ensure().codegen_unit(cgu.name());
547 let ongoing_codegen = start_async_codegen(backend.clone(), tcx, metadata, codegen_units.len());
548 let ongoing_codegen = AbortCodegenOnDrop::<B>(Some(ongoing_codegen));
550 // Codegen an allocator shim, if necessary.
552 // If the crate doesn't have an `allocator_kind` set then there's definitely
553 // no shim to generate. Otherwise we also check our dependency graph for all
554 // our output crate types. If anything there looks like its a `Dynamic`
555 // linkage, then it's already got an allocator shim and we'll be using that
556 // one instead. If nothing exists then it's our job to generate the
558 let any_dynamic_crate = tcx.dependency_formats(LOCAL_CRATE).iter().any(|(_, list)| {
559 use rustc_middle::middle::dependency_format::Linkage;
560 list.iter().any(|&linkage| linkage == Linkage::Dynamic)
562 let allocator_module = if any_dynamic_crate {
564 } else if let Some(kind) = tcx.allocator_kind() {
566 cgu_name_builder.build_cgu_name(LOCAL_CRATE, &["crate"], Some("allocator")).to_string();
567 let mut modules = backend.new_metadata(tcx, &llmod_id);
569 .time("write_allocator_module", || backend.codegen_allocator(tcx, &mut modules, kind));
571 Some(ModuleCodegen { name: llmod_id, module_llvm: modules, kind: ModuleKind::Allocator })
576 if let Some(allocator_module) = allocator_module {
577 ongoing_codegen.submit_pre_codegened_module_to_llvm(tcx, allocator_module);
580 if need_metadata_module {
581 // Codegen the encoded metadata.
582 let metadata_cgu_name =
583 cgu_name_builder.build_cgu_name(LOCAL_CRATE, &["crate"], Some("metadata")).to_string();
584 let mut metadata_llvm_module = backend.new_metadata(tcx, &metadata_cgu_name);
585 tcx.sess.time("write_compressed_metadata", || {
586 backend.write_compressed_metadata(
588 &ongoing_codegen.metadata,
589 &mut metadata_llvm_module,
593 let metadata_module = ModuleCodegen {
594 name: metadata_cgu_name,
595 module_llvm: metadata_llvm_module,
596 kind: ModuleKind::Metadata,
598 ongoing_codegen.submit_pre_codegened_module_to_llvm(tcx, metadata_module);
601 // We sort the codegen units by size. This way we can schedule work for LLVM
602 // a bit more efficiently.
603 let codegen_units = {
604 let mut codegen_units = codegen_units.iter().collect::<Vec<_>>();
605 codegen_units.sort_by_cached_key(|cgu| cmp::Reverse(cgu.size_estimate()));
609 let total_codegen_time = Lock::new(Duration::new(0, 0));
611 // The non-parallel compiler can only translate codegen units to LLVM IR
612 // on a single thread, leading to a staircase effect where the N LLVM
613 // threads have to wait on the single codegen threads to generate work
614 // for them. The parallel compiler does not have this restriction, so
615 // we can pre-load the LLVM queue in parallel before handing off
616 // coordination to the OnGoingCodegen scheduler.
618 // This likely is a temporary measure. Once we don't have to support the
619 // non-parallel compiler anymore, we can compile CGUs end-to-end in
620 // parallel and get rid of the complicated scheduling logic.
621 let pre_compile_cgus = |cgu_reuse: &[CguReuse]| {
622 if cfg!(parallel_compiler) {
623 tcx.sess.time("compile_first_CGU_batch", || {
624 // Try to find one CGU to compile per thread.
625 let cgus: Vec<_> = cgu_reuse
628 .filter(|&(_, reuse)| reuse == &CguReuse::No)
629 .take(tcx.sess.threads())
632 // Compile the found CGUs in parallel.
635 let start_time = Instant::now();
636 let module = backend.compile_codegen_unit(tcx, codegen_units[i].name());
637 let mut time = total_codegen_time.lock();
638 *time += start_time.elapsed();
648 let mut cgu_reuse = Vec::new();
649 let mut pre_compiled_cgus: Option<FxHashMap<usize, _>> = None;
651 for (i, cgu) in codegen_units.iter().enumerate() {
652 ongoing_codegen.wait_for_signal_to_codegen_item();
653 ongoing_codegen.check_for_errors(tcx.sess);
655 // Do some setup work in the first iteration
656 if pre_compiled_cgus.is_none() {
657 // Calculate the CGU reuse
658 cgu_reuse = tcx.sess.time("find_cgu_reuse", || {
659 codegen_units.iter().map(|cgu| determine_cgu_reuse(tcx, &cgu)).collect()
661 // Pre compile some CGUs
662 pre_compiled_cgus = Some(pre_compile_cgus(&cgu_reuse));
665 let cgu_reuse = cgu_reuse[i];
666 tcx.sess.cgu_reuse_tracker.set_actual_reuse(&cgu.name().as_str(), cgu_reuse);
671 if let Some(cgu) = pre_compiled_cgus.as_mut().unwrap().remove(&i) {
674 let start_time = Instant::now();
675 let module = backend.compile_codegen_unit(tcx, cgu.name());
676 let mut time = total_codegen_time.lock();
677 *time += start_time.elapsed();
680 submit_codegened_module_to_llvm(
682 &ongoing_codegen.coordinator_send,
688 CguReuse::PreLto => {
689 submit_pre_lto_module_to_llvm(
692 &ongoing_codegen.coordinator_send,
693 CachedModuleCodegen {
694 name: cgu.name().to_string(),
695 source: cgu.work_product(tcx),
700 CguReuse::PostLto => {
701 submit_post_lto_module_to_llvm(
703 &ongoing_codegen.coordinator_send,
704 CachedModuleCodegen {
705 name: cgu.name().to_string(),
706 source: cgu.work_product(tcx),
714 ongoing_codegen.codegen_finished(tcx);
716 // Since the main thread is sometimes blocked during codegen, we keep track
717 // -Ztime-passes output manually.
718 print_time_passes_entry(
719 tcx.sess.time_passes(),
720 "codegen_to_LLVM_IR",
721 total_codegen_time.into_inner(),
724 ::rustc_incremental::assert_module_sources::assert_module_sources(tcx);
726 symbol_names_test::report_symbol_names(tcx);
728 ongoing_codegen.check_for_errors(tcx.sess);
732 ongoing_codegen.into_inner()
735 /// A curious wrapper structure whose only purpose is to call `codegen_aborted`
736 /// when it's dropped abnormally.
738 /// In the process of working on rust-lang/rust#55238 a mysterious segfault was
739 /// stumbled upon. The segfault was never reproduced locally, but it was
740 /// suspected to be related to the fact that codegen worker threads were
741 /// sticking around by the time the main thread was exiting, causing issues.
743 /// This structure is an attempt to fix that issue where the `codegen_aborted`
744 /// message will block until all workers have finished. This should ensure that
745 /// even if the main codegen thread panics we'll wait for pending work to
746 /// complete before returning from the main thread, hopefully avoiding
749 /// If you see this comment in the code, then it means that this workaround
750 /// worked! We may yet one day track down the mysterious cause of that
752 struct AbortCodegenOnDrop<B: ExtraBackendMethods>(Option<OngoingCodegen<B>>);
754 impl<B: ExtraBackendMethods> AbortCodegenOnDrop<B> {
755 fn into_inner(mut self) -> OngoingCodegen<B> {
756 self.0.take().unwrap()
760 impl<B: ExtraBackendMethods> Deref for AbortCodegenOnDrop<B> {
761 type Target = OngoingCodegen<B>;
763 fn deref(&self) -> &OngoingCodegen<B> {
764 self.0.as_ref().unwrap()
768 impl<B: ExtraBackendMethods> DerefMut for AbortCodegenOnDrop<B> {
769 fn deref_mut(&mut self) -> &mut OngoingCodegen<B> {
770 self.0.as_mut().unwrap()
774 impl<B: ExtraBackendMethods> Drop for AbortCodegenOnDrop<B> {
776 if let Some(codegen) = self.0.take() {
777 codegen.codegen_aborted();
782 fn finalize_tcx(tcx: TyCtxt<'_>) {
783 tcx.sess.time("assert_dep_graph", || ::rustc_incremental::assert_dep_graph(tcx));
784 tcx.sess.time("serialize_dep_graph", || ::rustc_incremental::save_dep_graph(tcx));
786 // We assume that no queries are run past here. If there are new queries
787 // after this point, they'll show up as "<unknown>" in self-profiling data.
789 let _prof_timer = tcx.prof.generic_activity("self_profile_alloc_query_strings");
790 tcx.alloc_self_profile_query_strings();
795 pub fn new(tcx: TyCtxt<'_>) -> CrateInfo {
796 let mut info = CrateInfo {
798 compiler_builtins: None,
799 profiler_runtime: None,
800 is_no_builtins: Default::default(),
801 native_libraries: Default::default(),
802 used_libraries: tcx.native_libraries(LOCAL_CRATE),
803 link_args: tcx.link_args(LOCAL_CRATE),
804 crate_name: Default::default(),
805 used_crates_dynamic: cstore::used_crates(tcx, LinkagePreference::RequireDynamic),
806 used_crates_static: cstore::used_crates(tcx, LinkagePreference::RequireStatic),
807 used_crate_source: Default::default(),
808 lang_item_to_crate: Default::default(),
809 missing_lang_items: Default::default(),
810 dependency_formats: tcx.dependency_formats(LOCAL_CRATE),
812 let lang_items = tcx.lang_items();
814 let crates = tcx.crates();
816 let n_crates = crates.len();
817 info.native_libraries.reserve(n_crates);
818 info.crate_name.reserve(n_crates);
819 info.used_crate_source.reserve(n_crates);
820 info.missing_lang_items.reserve(n_crates);
822 for &cnum in crates.iter() {
823 info.native_libraries.insert(cnum, tcx.native_libraries(cnum));
824 info.crate_name.insert(cnum, tcx.crate_name(cnum).to_string());
825 info.used_crate_source.insert(cnum, tcx.used_crate_source(cnum));
826 if tcx.is_panic_runtime(cnum) {
827 info.panic_runtime = Some(cnum);
829 if tcx.is_compiler_builtins(cnum) {
830 info.compiler_builtins = Some(cnum);
832 if tcx.is_profiler_runtime(cnum) {
833 info.profiler_runtime = Some(cnum);
835 if tcx.is_no_builtins(cnum) {
836 info.is_no_builtins.insert(cnum);
838 let missing = tcx.missing_lang_items(cnum);
839 for &item in missing.iter() {
840 if let Ok(id) = lang_items.require(item) {
841 info.lang_item_to_crate.insert(item, id.krate);
845 // No need to look for lang items that don't actually need to exist.
847 missing.iter().cloned().filter(|&l| lang_items::required(tcx, l)).collect();
848 info.missing_lang_items.insert(cnum, missing);
855 pub fn provide_both(providers: &mut Providers) {
856 providers.backend_optimization_level = |tcx, cratenum| {
857 let for_speed = match tcx.sess.opts.optimize {
858 // If globally no optimisation is done, #[optimize] has no effect.
860 // This is done because if we ended up "upgrading" to `-O2` here, we’d populate the
861 // pass manager and it is likely that some module-wide passes (such as inliner or
862 // cross-function constant propagation) would ignore the `optnone` annotation we put
863 // on the functions, thus necessarily involving these functions into optimisations.
864 config::OptLevel::No => return config::OptLevel::No,
865 // If globally optimise-speed is already specified, just use that level.
866 config::OptLevel::Less => return config::OptLevel::Less,
867 config::OptLevel::Default => return config::OptLevel::Default,
868 config::OptLevel::Aggressive => return config::OptLevel::Aggressive,
869 // If globally optimize-for-size has been requested, use -O2 instead (if optimize(size)
871 config::OptLevel::Size => config::OptLevel::Default,
872 config::OptLevel::SizeMin => config::OptLevel::Default,
875 let (defids, _) = tcx.collect_and_partition_mono_items(cratenum);
877 let CodegenFnAttrs { optimize, .. } = tcx.codegen_fn_attrs(*id);
879 attr::OptimizeAttr::None => continue,
880 attr::OptimizeAttr::Size => continue,
881 attr::OptimizeAttr::Speed => {
886 tcx.sess.opts.optimize
889 providers.dllimport_foreign_items = |tcx, krate| {
890 let module_map = tcx.foreign_modules(krate);
892 module_map.iter().map(|lib| (lib.def_id, lib)).collect::<FxHashMap<_, _>>();
895 .native_libraries(krate)
898 if !matches!(lib.kind, NativeLibKind::Dylib | NativeLibKind::Unspecified) {
901 let cfg = match lib.cfg {
902 Some(ref cfg) => cfg,
905 attr::cfg_matches(cfg, &tcx.sess.parse_sess, None)
907 .filter_map(|lib| lib.foreign_module)
908 .map(|id| &module_map[&id])
909 .flat_map(|module| module.foreign_items.iter().cloned())
914 providers.is_dllimport_foreign_item =
915 |tcx, def_id| tcx.dllimport_foreign_items(def_id.krate).contains(&def_id);
918 fn determine_cgu_reuse<'tcx>(tcx: TyCtxt<'tcx>, cgu: &CodegenUnit<'tcx>) -> CguReuse {
919 if !tcx.dep_graph.is_fully_enabled() {
923 let work_product_id = &cgu.work_product_id();
924 if tcx.dep_graph.previous_work_product(work_product_id).is_none() {
925 // We don't have anything cached for this CGU. This can happen
926 // if the CGU did not exist in the previous session.
930 // Try to mark the CGU as green. If it we can do so, it means that nothing
931 // affecting the LLVM module has changed and we can re-use a cached version.
932 // If we compile with any kind of LTO, this means we can re-use the bitcode
933 // of the Pre-LTO stage (possibly also the Post-LTO version but we'll only
934 // know that later). If we are not doing LTO, there is only one optimized
935 // version of each module, so we re-use that.
936 let dep_node = cgu.codegen_dep_node(tcx);
938 !tcx.dep_graph.dep_node_exists(&dep_node),
939 "CompileCodegenUnit dep-node for CGU `{}` already exists before marking.",
943 if tcx.dep_graph.try_mark_green(tcx, &dep_node).is_some() {
944 // We can re-use either the pre- or the post-thinlto state. If no LTO is
945 // being performed then we can use post-LTO artifacts, otherwise we must
946 // reuse pre-LTO artifacts
947 match compute_per_cgu_lto_type(
950 &tcx.sess.crate_types(),
953 ComputedLtoType::No => CguReuse::PostLto,
954 _ => CguReuse::PreLto,