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 start_async_codegen, submit_codegened_module_to_llvm, submit_post_lto_module_to_llvm,
18 submit_pre_lto_module_to_llvm, 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::{DefId, LOCAL_CRATE};
34 use rustc_index::vec::Idx;
35 use rustc_middle::middle::codegen_fn_attrs::CodegenFnAttrs;
36 use rustc_middle::middle::cstore::EncodedMetadata;
37 use rustc_middle::middle::cstore::{self, LinkagePreference};
38 use rustc_middle::middle::lang_items;
39 use rustc_middle::middle::lang_items::StartFnLangItem;
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, Lto};
47 use rustc_session::Session;
49 use rustc_symbol_mangling::test as symbol_names_test;
50 use rustc_target::abi::{Abi, Align, LayoutOf, Scalar, VariantIdx};
53 use std::ops::{Deref, DerefMut};
54 use std::time::{Duration, Instant};
56 pub fn bin_op_to_icmp_predicate(op: hir::BinOpKind, signed: bool) -> IntPredicate {
58 hir::BinOpKind::Eq => IntPredicate::IntEQ,
59 hir::BinOpKind::Ne => IntPredicate::IntNE,
60 hir::BinOpKind::Lt => {
67 hir::BinOpKind::Le => {
74 hir::BinOpKind::Gt => {
81 hir::BinOpKind::Ge => {
89 "comparison_op_to_icmp_predicate: expected comparison operator, \
96 pub fn bin_op_to_fcmp_predicate(op: hir::BinOpKind) -> RealPredicate {
98 hir::BinOpKind::Eq => RealPredicate::RealOEQ,
99 hir::BinOpKind::Ne => RealPredicate::RealUNE,
100 hir::BinOpKind::Lt => RealPredicate::RealOLT,
101 hir::BinOpKind::Le => RealPredicate::RealOLE,
102 hir::BinOpKind::Gt => RealPredicate::RealOGT,
103 hir::BinOpKind::Ge => RealPredicate::RealOGE,
106 "comparison_op_to_fcmp_predicate: expected comparison operator, \
114 pub fn compare_simd_types<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
122 let signed = match t.kind {
124 let cmp = bin_op_to_fcmp_predicate(op);
125 let cmp = bx.fcmp(cmp, lhs, rhs);
126 return bx.sext(cmp, ret_ty);
128 ty::Uint(_) => false,
130 _ => bug!("compare_simd_types: invalid SIMD type"),
133 let cmp = bin_op_to_icmp_predicate(op, signed);
134 let cmp = bx.icmp(cmp, lhs, rhs);
135 // LLVM outputs an `< size x i1 >`, so we need to perform a sign extension
136 // to get the correctly sized type. This will compile to a single instruction
137 // once the IR is converted to assembly if the SIMD instruction is supported
138 // by the target architecture.
142 /// Retrieves the information we are losing (making dynamic) in an unsizing
145 /// The `old_info` argument is a bit odd. It is intended for use in an upcast,
146 /// where the new vtable for an object will be derived from the old one.
147 pub fn unsized_info<'tcx, Cx: CodegenMethods<'tcx>>(
151 old_info: Option<Cx::Value>,
153 let (source, target) =
154 cx.tcx().struct_lockstep_tails_erasing_lifetimes(source, target, cx.param_env());
155 match (&source.kind, &target.kind) {
156 (&ty::Array(_, len), &ty::Slice(_)) => {
157 cx.const_usize(len.eval_usize(cx.tcx(), ty::ParamEnv::reveal_all()))
159 (&ty::Dynamic(..), &ty::Dynamic(..)) => {
160 // For now, upcasts are limited to changes in marker
161 // traits, and hence never actually require an actual
162 // change to the vtable.
163 old_info.expect("unsized_info: missing old info for trait upcast")
165 (_, &ty::Dynamic(ref data, ..)) => {
166 let vtable_ptr = cx.layout_of(cx.tcx().mk_mut_ptr(target)).field(cx, FAT_PTR_EXTRA);
168 meth::get_vtable(cx, source, data.principal()),
169 cx.backend_type(vtable_ptr),
172 _ => bug!("unsized_info: invalid unsizing {:?} -> {:?}", source, target),
176 /// Coerces `src` to `dst_ty`. `src_ty` must be a thin pointer.
177 pub fn unsize_thin_ptr<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
182 ) -> (Bx::Value, Bx::Value) {
183 debug!("unsize_thin_ptr: {:?} => {:?}", src_ty, dst_ty);
184 match (&src_ty.kind, &dst_ty.kind) {
185 (&ty::Ref(_, a, _), &ty::Ref(_, b, _))
186 | (&ty::Ref(_, a, _), &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(..))
236 | (&ty::Ref(..), &ty::RawPtr(..))
237 | (&ty::RawPtr(..), &ty::RawPtr(..)) => {
238 let (base, info) = match bx.load_operand(src).val {
239 OperandValue::Pair(base, info) => {
240 // fat-ptr to fat-ptr unsize preserves the vtable
241 // i.e., &'a fmt::Debug+Send => &'a fmt::Debug
242 // So we need to pointercast the base to ensure
243 // the types match up.
244 // FIXME(eddyb) use `scalar_pair_element_backend_type` here,
245 // like `unsize_thin_ptr` does.
246 let thin_ptr = dst.layout.field(bx.cx(), FAT_PTR_ADDR);
247 (bx.pointercast(base, bx.cx().backend_type(thin_ptr)), info)
249 OperandValue::Immediate(base) => unsize_thin_ptr(bx, base, src_ty, dst_ty),
250 OperandValue::Ref(..) => bug!(),
252 OperandValue::Pair(base, info).store(bx, dst);
255 (&ty::Adt(def_a, _), &ty::Adt(def_b, _)) => {
256 assert_eq!(def_a, def_b);
258 for i in 0..def_a.variants[VariantIdx::new(0)].fields.len() {
259 let src_f = src.project_field(bx, i);
260 let dst_f = dst.project_field(bx, i);
262 if dst_f.layout.is_zst() {
266 if src_f.layout.ty == dst_f.layout.ty {
277 coerce_unsized_into(bx, src_f, dst_f);
281 _ => bug!("coerce_unsized_into: invalid coercion {:?} -> {:?}", src_ty, dst_ty,),
285 pub fn cast_shift_expr_rhs<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
291 cast_shift_rhs(bx, op, lhs, rhs)
294 fn cast_shift_rhs<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
300 // Shifts may have any size int on the rhs
302 let mut rhs_llty = bx.cx().val_ty(rhs);
303 let mut lhs_llty = bx.cx().val_ty(lhs);
304 if bx.cx().type_kind(rhs_llty) == TypeKind::Vector {
305 rhs_llty = bx.cx().element_type(rhs_llty)
307 if bx.cx().type_kind(lhs_llty) == TypeKind::Vector {
308 lhs_llty = bx.cx().element_type(lhs_llty)
310 let rhs_sz = bx.cx().int_width(rhs_llty);
311 let lhs_sz = bx.cx().int_width(lhs_llty);
313 bx.trunc(rhs, lhs_llty)
314 } else if lhs_sz > rhs_sz {
315 // FIXME (#1877: If in the future shifting by negative
316 // values is no longer undefined then this is wrong.
317 bx.zext(rhs, lhs_llty)
326 /// Returns `true` if this session's target will use SEH-based unwinding.
328 /// This is only true for MSVC targets, and even then the 64-bit MSVC target
329 /// currently uses SEH-ish unwinding with DWARF info tables to the side (same as
330 /// 64-bit MinGW) instead of "full SEH".
331 pub fn wants_msvc_seh(sess: &Session) -> bool {
332 sess.target.target.options.is_like_msvc
335 pub fn from_immediate<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
339 if bx.cx().val_ty(val) == bx.cx().type_i1() { bx.zext(val, bx.cx().type_i8()) } else { val }
342 pub fn to_immediate<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
345 layout: layout::TyAndLayout<'_>,
347 if let Abi::Scalar(ref scalar) = layout.abi {
348 return to_immediate_scalar(bx, val, scalar);
353 pub fn to_immediate_scalar<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
358 if scalar.is_bool() {
359 return bx.trunc(val, bx.cx().type_i1());
364 pub fn memcpy_ty<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
370 layout: TyAndLayout<'tcx>,
373 let size = layout.size.bytes();
378 bx.memcpy(dst, dst_align, src, src_align, bx.cx().const_usize(size), flags);
381 pub fn codegen_instance<'a, 'tcx: 'a, Bx: BuilderMethods<'a, 'tcx>>(
382 cx: &'a Bx::CodegenCx,
383 instance: Instance<'tcx>,
385 // this is an info! to allow collecting monomorphization statistics
386 // and to allow finding the last function before LLVM aborts from
388 info!("codegen_instance({})", instance);
390 mir::codegen_mir::<Bx>(cx, instance);
393 /// Creates the `main` function which will initialize the rust runtime and call
394 /// users main function.
395 pub fn maybe_create_entry_wrapper<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
396 cx: &'a Bx::CodegenCx,
397 ) -> Option<Bx::Function> {
398 let (main_def_id, span) = match cx.tcx().entry_fn(LOCAL_CRATE) {
399 Some((def_id, _)) => (def_id, cx.tcx().def_span(def_id)),
403 let instance = Instance::mono(cx.tcx(), main_def_id);
405 if !cx.codegen_unit().contains_item(&MonoItem::Fn(instance)) {
406 // We want to create the wrapper in the same codegen unit as Rust's main
411 let main_llfn = cx.get_fn_addr(instance);
413 return cx.tcx().entry_fn(LOCAL_CRATE).map(|(_, et)| {
414 let use_start_lang_item = EntryFnType::Start != et;
415 create_entry_fn::<Bx>(cx, span, main_llfn, main_def_id, use_start_lang_item)
418 fn create_entry_fn<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
419 cx: &'a Bx::CodegenCx,
421 rust_main: Bx::Value,
422 rust_main_def_id: DefId,
423 use_start_lang_item: bool,
425 // The entry function is either `int main(void)` or `int main(int argc, char **argv)`,
426 // depending on whether the target needs `argc` and `argv` to be passed in.
427 let llfty = if cx.sess().target.target.options.main_needs_argc_argv {
428 cx.type_func(&[cx.type_int(), cx.type_ptr_to(cx.type_i8p())], cx.type_int())
430 cx.type_func(&[], cx.type_int())
433 let main_ret_ty = cx.tcx().fn_sig(rust_main_def_id).output();
434 // Given that `main()` has no arguments,
435 // then its return type cannot have
436 // late-bound regions, since late-bound
437 // regions must appear in the argument
439 let main_ret_ty = cx.tcx().erase_regions(&main_ret_ty.no_bound_vars().unwrap());
441 if cx.get_declared_value("main").is_some() {
442 // FIXME: We should be smart and show a better diagnostic here.
444 .struct_span_err(sp, "entry symbol `main` declared multiple times")
445 .help("did you use `#[no_mangle]` on `fn main`? Use `#[start]` instead")
447 cx.sess().abort_if_errors();
450 let llfn = cx.declare_cfn("main", llfty);
452 // `main` should respect same config for frame pointer elimination as rest of code
453 cx.set_frame_pointer_elimination(llfn);
454 cx.apply_target_cpu_attr(llfn);
456 let mut bx = Bx::new_block(&cx, llfn, "top");
458 bx.insert_reference_to_gdb_debug_scripts_section_global();
460 let (arg_argc, arg_argv) = get_argc_argv(cx, &mut bx);
462 let (start_fn, args) = if use_start_lang_item {
463 let start_def_id = cx.tcx().require_lang_item(StartFnLangItem, None);
464 let start_fn = cx.get_fn_addr(
465 ty::Instance::resolve(
467 ty::ParamEnv::reveal_all(),
469 cx.tcx().intern_substs(&[main_ret_ty.into()]),
475 vec![bx.pointercast(rust_main, cx.type_ptr_to(cx.type_i8p())), arg_argc, arg_argv],
478 debug!("using user-defined start fn");
479 (rust_main, vec![arg_argc, arg_argv])
482 let result = bx.call(start_fn, &args, None);
483 let cast = bx.intcast(result, cx.type_int(), true);
490 /// Obtain the `argc` and `argv` values to pass to the rust start function.
491 fn get_argc_argv<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
492 cx: &'a Bx::CodegenCx,
494 ) -> (Bx::Value, Bx::Value) {
495 if cx.sess().target.target.options.main_needs_argc_argv {
496 // Params from native `main()` used as args for rust start function
497 let param_argc = bx.get_param(0);
498 let param_argv = bx.get_param(1);
499 let arg_argc = bx.intcast(param_argc, cx.type_isize(), true);
500 let arg_argv = param_argv;
503 // The Rust start function doesn't need `argc` and `argv`, so just pass zeros.
504 let arg_argc = bx.const_int(cx.type_int(), 0);
505 let arg_argv = bx.const_null(cx.type_ptr_to(cx.type_i8p()));
510 pub const CODEGEN_WORKER_ID: usize = ::std::usize::MAX;
512 pub fn codegen_crate<B: ExtraBackendMethods>(
515 metadata: EncodedMetadata,
516 need_metadata_module: bool,
517 ) -> OngoingCodegen<B> {
518 // Skip crate items and just output metadata in -Z no-codegen mode.
519 if tcx.sess.opts.debugging_opts.no_codegen || !tcx.sess.opts.output_types.should_codegen() {
520 let ongoing_codegen = start_async_codegen(backend, tcx, metadata, 1);
522 ongoing_codegen.codegen_finished(tcx);
526 ongoing_codegen.check_for_errors(tcx.sess);
528 return ongoing_codegen;
531 let cgu_name_builder = &mut CodegenUnitNameBuilder::new(tcx);
533 // Run the monomorphization collector and partition the collected items into
535 let codegen_units = tcx.collect_and_partition_mono_items(LOCAL_CRATE).1;
536 let codegen_units = (*codegen_units).clone();
538 // Force all codegen_unit queries so they are already either red or green
539 // when compile_codegen_unit accesses them. We are not able to re-execute
540 // the codegen_unit query from just the DepNode, so an unknown color would
541 // lead to having to re-execute compile_codegen_unit, possibly
543 if tcx.dep_graph.is_fully_enabled() {
544 for cgu in &codegen_units {
545 tcx.codegen_unit(cgu.name());
549 let ongoing_codegen = start_async_codegen(backend.clone(), tcx, metadata, codegen_units.len());
550 let ongoing_codegen = AbortCodegenOnDrop::<B>(Some(ongoing_codegen));
552 // Codegen an allocator shim, if necessary.
554 // If the crate doesn't have an `allocator_kind` set then there's definitely
555 // no shim to generate. Otherwise we also check our dependency graph for all
556 // our output crate types. If anything there looks like its a `Dynamic`
557 // linkage, then it's already got an allocator shim and we'll be using that
558 // one instead. If nothing exists then it's our job to generate the
560 let any_dynamic_crate = tcx.dependency_formats(LOCAL_CRATE).iter().any(|(_, list)| {
561 use rustc_middle::middle::dependency_format::Linkage;
562 list.iter().any(|&linkage| linkage == Linkage::Dynamic)
564 let allocator_module = if any_dynamic_crate {
566 } else if let Some(kind) = tcx.allocator_kind() {
568 cgu_name_builder.build_cgu_name(LOCAL_CRATE, &["crate"], Some("allocator")).to_string();
569 let mut modules = backend.new_metadata(tcx, &llmod_id);
571 .time("write_allocator_module", || backend.codegen_allocator(tcx, &mut modules, kind));
573 Some(ModuleCodegen { name: llmod_id, module_llvm: modules, kind: ModuleKind::Allocator })
578 if let Some(allocator_module) = allocator_module {
579 ongoing_codegen.submit_pre_codegened_module_to_llvm(tcx, allocator_module);
582 if need_metadata_module {
583 // Codegen the encoded metadata.
584 let metadata_cgu_name =
585 cgu_name_builder.build_cgu_name(LOCAL_CRATE, &["crate"], Some("metadata")).to_string();
586 let mut metadata_llvm_module = backend.new_metadata(tcx, &metadata_cgu_name);
587 tcx.sess.time("write_compressed_metadata", || {
588 backend.write_compressed_metadata(
590 &ongoing_codegen.metadata,
591 &mut metadata_llvm_module,
595 let metadata_module = ModuleCodegen {
596 name: metadata_cgu_name,
597 module_llvm: metadata_llvm_module,
598 kind: ModuleKind::Metadata,
600 ongoing_codegen.submit_pre_codegened_module_to_llvm(tcx, metadata_module);
603 // We sort the codegen units by size. This way we can schedule work for LLVM
604 // a bit more efficiently.
605 let codegen_units = {
606 let mut codegen_units = codegen_units;
607 codegen_units.sort_by_cached_key(|cgu| cmp::Reverse(cgu.size_estimate()));
611 let total_codegen_time = Lock::new(Duration::new(0, 0));
613 // The non-parallel compiler can only translate codegen units to LLVM IR
614 // on a single thread, leading to a staircase effect where the N LLVM
615 // threads have to wait on the single codegen threads to generate work
616 // for them. The parallel compiler does not have this restriction, so
617 // we can pre-load the LLVM queue in parallel before handing off
618 // coordination to the OnGoingCodegen scheduler.
620 // This likely is a temporary measure. Once we don't have to support the
621 // non-parallel compiler anymore, we can compile CGUs end-to-end in
622 // parallel and get rid of the complicated scheduling logic.
623 let pre_compile_cgus = |cgu_reuse: &[CguReuse]| {
624 if cfg!(parallel_compiler) {
625 tcx.sess.time("compile_first_CGU_batch", || {
626 // Try to find one CGU to compile per thread.
627 let cgus: Vec<_> = cgu_reuse
630 .filter(|&(_, reuse)| reuse == &CguReuse::No)
631 .take(tcx.sess.threads())
634 // Compile the found CGUs in parallel.
637 let start_time = Instant::now();
638 let module = backend.compile_codegen_unit(tcx, codegen_units[i].name());
639 let mut time = total_codegen_time.lock();
640 *time += start_time.elapsed();
650 let mut cgu_reuse = Vec::new();
651 let mut pre_compiled_cgus: Option<FxHashMap<usize, _>> = None;
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 pre_compiled_cgus = Some(pre_compile_cgus(&cgu_reuse));
667 let cgu_reuse = cgu_reuse[i];
668 tcx.sess.cgu_reuse_tracker.set_actual_reuse(&cgu.name().as_str(), cgu_reuse);
673 if let Some(cgu) = pre_compiled_cgus.as_mut().unwrap().remove(&i) {
676 let start_time = Instant::now();
677 let module = backend.compile_codegen_unit(tcx, cgu.name());
678 let mut time = total_codegen_time.lock();
679 *time += start_time.elapsed();
682 submit_codegened_module_to_llvm(
684 &ongoing_codegen.coordinator_send,
690 CguReuse::PreLto => {
691 submit_pre_lto_module_to_llvm(
694 &ongoing_codegen.coordinator_send,
695 CachedModuleCodegen {
696 name: cgu.name().to_string(),
697 source: cgu.work_product(tcx),
702 CguReuse::PostLto => {
703 submit_post_lto_module_to_llvm(
705 &ongoing_codegen.coordinator_send,
706 CachedModuleCodegen {
707 name: cgu.name().to_string(),
708 source: cgu.work_product(tcx),
716 ongoing_codegen.codegen_finished(tcx);
718 // Since the main thread is sometimes blocked during codegen, we keep track
719 // -Ztime-passes output manually.
720 print_time_passes_entry(
721 tcx.sess.time_passes(),
722 "codegen_to_LLVM_IR",
723 total_codegen_time.into_inner(),
726 ::rustc_incremental::assert_module_sources::assert_module_sources(tcx);
728 symbol_names_test::report_symbol_names(tcx);
730 ongoing_codegen.check_for_errors(tcx.sess);
734 ongoing_codegen.into_inner()
737 /// A curious wrapper structure whose only purpose is to call `codegen_aborted`
738 /// when it's dropped abnormally.
740 /// In the process of working on rust-lang/rust#55238 a mysterious segfault was
741 /// stumbled upon. The segfault was never reproduced locally, but it was
742 /// suspected to be related to the fact that codegen worker threads were
743 /// sticking around by the time the main thread was exiting, causing issues.
745 /// This structure is an attempt to fix that issue where the `codegen_aborted`
746 /// message will block until all workers have finished. This should ensure that
747 /// even if the main codegen thread panics we'll wait for pending work to
748 /// complete before returning from the main thread, hopefully avoiding
751 /// If you see this comment in the code, then it means that this workaround
752 /// worked! We may yet one day track down the mysterious cause of that
754 struct AbortCodegenOnDrop<B: ExtraBackendMethods>(Option<OngoingCodegen<B>>);
756 impl<B: ExtraBackendMethods> AbortCodegenOnDrop<B> {
757 fn into_inner(mut self) -> OngoingCodegen<B> {
758 self.0.take().unwrap()
762 impl<B: ExtraBackendMethods> Deref for AbortCodegenOnDrop<B> {
763 type Target = OngoingCodegen<B>;
765 fn deref(&self) -> &OngoingCodegen<B> {
766 self.0.as_ref().unwrap()
770 impl<B: ExtraBackendMethods> DerefMut for AbortCodegenOnDrop<B> {
771 fn deref_mut(&mut self) -> &mut OngoingCodegen<B> {
772 self.0.as_mut().unwrap()
776 impl<B: ExtraBackendMethods> Drop for AbortCodegenOnDrop<B> {
778 if let Some(codegen) = self.0.take() {
779 codegen.codegen_aborted();
784 fn finalize_tcx(tcx: TyCtxt<'_>) {
785 tcx.sess.time("assert_dep_graph", || ::rustc_incremental::assert_dep_graph(tcx));
786 tcx.sess.time("serialize_dep_graph", || ::rustc_incremental::save_dep_graph(tcx));
788 // We assume that no queries are run past here. If there are new queries
789 // after this point, they'll show up as "<unknown>" in self-profiling data.
791 let _prof_timer = tcx.prof.generic_activity("self_profile_alloc_query_strings");
792 tcx.alloc_self_profile_query_strings();
797 pub fn new(tcx: TyCtxt<'_>) -> CrateInfo {
798 let mut info = CrateInfo {
800 compiler_builtins: None,
801 profiler_runtime: None,
802 is_no_builtins: Default::default(),
803 native_libraries: Default::default(),
804 used_libraries: tcx.native_libraries(LOCAL_CRATE),
805 link_args: tcx.link_args(LOCAL_CRATE),
806 crate_name: Default::default(),
807 used_crates_dynamic: cstore::used_crates(tcx, LinkagePreference::RequireDynamic),
808 used_crates_static: cstore::used_crates(tcx, LinkagePreference::RequireStatic),
809 used_crate_source: Default::default(),
810 lang_item_to_crate: Default::default(),
811 missing_lang_items: Default::default(),
812 dependency_formats: tcx.dependency_formats(LOCAL_CRATE),
814 let lang_items = tcx.lang_items();
816 let crates = tcx.crates();
818 let n_crates = crates.len();
819 info.native_libraries.reserve(n_crates);
820 info.crate_name.reserve(n_crates);
821 info.used_crate_source.reserve(n_crates);
822 info.missing_lang_items.reserve(n_crates);
824 for &cnum in crates.iter() {
825 info.native_libraries.insert(cnum, tcx.native_libraries(cnum));
826 info.crate_name.insert(cnum, tcx.crate_name(cnum).to_string());
827 info.used_crate_source.insert(cnum, tcx.used_crate_source(cnum));
828 if tcx.is_panic_runtime(cnum) {
829 info.panic_runtime = Some(cnum);
831 if tcx.is_compiler_builtins(cnum) {
832 info.compiler_builtins = Some(cnum);
834 if tcx.is_profiler_runtime(cnum) {
835 info.profiler_runtime = Some(cnum);
837 if tcx.is_no_builtins(cnum) {
838 info.is_no_builtins.insert(cnum);
840 let missing = tcx.missing_lang_items(cnum);
841 for &item in missing.iter() {
842 if let Ok(id) = lang_items.require(item) {
843 info.lang_item_to_crate.insert(item, id.krate);
847 // No need to look for lang items that are whitelisted and don't
848 // actually need to exist.
850 missing.iter().cloned().filter(|&l| !lang_items::whitelisted(tcx, l)).collect();
851 info.missing_lang_items.insert(cnum, missing);
858 pub fn provide_both(providers: &mut Providers<'_>) {
859 providers.backend_optimization_level = |tcx, cratenum| {
860 let for_speed = match tcx.sess.opts.optimize {
861 // If globally no optimisation is done, #[optimize] has no effect.
863 // This is done because if we ended up "upgrading" to `-O2` here, we’d populate the
864 // pass manager and it is likely that some module-wide passes (such as inliner or
865 // cross-function constant propagation) would ignore the `optnone` annotation we put
866 // on the functions, thus necessarily involving these functions into optimisations.
867 config::OptLevel::No => return config::OptLevel::No,
868 // If globally optimise-speed is already specified, just use that level.
869 config::OptLevel::Less => return config::OptLevel::Less,
870 config::OptLevel::Default => return config::OptLevel::Default,
871 config::OptLevel::Aggressive => return config::OptLevel::Aggressive,
872 // If globally optimize-for-size has been requested, use -O2 instead (if optimize(size)
874 config::OptLevel::Size => config::OptLevel::Default,
875 config::OptLevel::SizeMin => config::OptLevel::Default,
878 let (defids, _) = tcx.collect_and_partition_mono_items(cratenum);
880 let CodegenFnAttrs { optimize, .. } = tcx.codegen_fn_attrs(*id);
882 attr::OptimizeAttr::None => continue,
883 attr::OptimizeAttr::Size => continue,
884 attr::OptimizeAttr::Speed => {
889 tcx.sess.opts.optimize
892 providers.dllimport_foreign_items = |tcx, krate| {
893 let module_map = tcx.foreign_modules(krate);
895 module_map.iter().map(|lib| (lib.def_id, lib)).collect::<FxHashMap<_, _>>();
898 .native_libraries(krate)
901 if lib.kind != cstore::NativeLibraryKind::NativeUnknown {
904 let cfg = match lib.cfg {
905 Some(ref cfg) => cfg,
908 attr::cfg_matches(cfg, &tcx.sess.parse_sess, None)
910 .filter_map(|lib| lib.foreign_module)
911 .map(|id| &module_map[&id])
912 .flat_map(|module| module.foreign_items.iter().cloned())
914 tcx.arena.alloc(dllimports)
917 providers.is_dllimport_foreign_item =
918 |tcx, def_id| tcx.dllimport_foreign_items(def_id.krate).contains(&def_id);
921 fn determine_cgu_reuse<'tcx>(tcx: TyCtxt<'tcx>, cgu: &CodegenUnit<'tcx>) -> CguReuse {
922 if !tcx.dep_graph.is_fully_enabled() {
926 let work_product_id = &cgu.work_product_id();
927 if tcx.dep_graph.previous_work_product(work_product_id).is_none() {
928 // We don't have anything cached for this CGU. This can happen
929 // if the CGU did not exist in the previous session.
933 // Try to mark the CGU as green. If it we can do so, it means that nothing
934 // affecting the LLVM module has changed and we can re-use a cached version.
935 // If we compile with any kind of LTO, this means we can re-use the bitcode
936 // of the Pre-LTO stage (possibly also the Post-LTO version but we'll only
937 // know that later). If we are not doing LTO, there is only one optimized
938 // version of each module, so we re-use that.
939 let dep_node = cgu.codegen_dep_node(tcx);
941 !tcx.dep_graph.dep_node_exists(&dep_node),
942 "CompileCodegenUnit dep-node for CGU `{}` already exists before marking.",
946 if tcx.dep_graph.try_mark_green(tcx, &dep_node).is_some() {
947 // We can re-use either the pre- or the post-thinlto state
948 if tcx.sess.lto() != Lto::No { CguReuse::PreLto } else { CguReuse::PostLto }