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::middle::codegen_fn_attrs::CodegenFnAttrs;
29 use rustc::middle::cstore::EncodedMetadata;
30 use rustc::middle::cstore::{self, LinkagePreference};
31 use rustc::middle::lang_items::StartFnLangItem;
32 use rustc::middle::weak_lang_items;
33 use rustc::mir::mono::{CodegenUnit, CodegenUnitNameBuilder, MonoItem};
34 use rustc::session::config::{self, EntryFnType, Lto};
35 use rustc::session::Session;
36 use rustc::ty::layout::{self, Align, HasTyCtxt, LayoutOf, TyLayout, VariantIdx};
37 use rustc::ty::layout::{FAT_PTR_ADDR, FAT_PTR_EXTRA};
38 use rustc::ty::query::Providers;
39 use rustc::ty::{self, Instance, Ty, TyCtxt};
40 use rustc_codegen_utils::{check_for_rustc_errors_attr, symbol_names_test};
41 use rustc_data_structures::fx::FxHashMap;
42 use rustc_data_structures::profiling::print_time_passes_entry;
43 use rustc_data_structures::sync::{par_iter, Lock, ParallelIterator};
45 use rustc_hir::def_id::{DefId, LOCAL_CRATE};
46 use rustc_index::vec::Idx;
47 use rustc_session::cgu_reuse_tracker::CguReuse;
52 use std::ops::{Deref, DerefMut};
53 use std::time::{Duration, Instant};
55 pub fn bin_op_to_icmp_predicate(op: hir::BinOpKind, signed: bool) -> IntPredicate {
57 hir::BinOpKind::Eq => IntPredicate::IntEQ,
58 hir::BinOpKind::Ne => IntPredicate::IntNE,
59 hir::BinOpKind::Lt => {
66 hir::BinOpKind::Le => {
73 hir::BinOpKind::Gt => {
80 hir::BinOpKind::Ge => {
88 "comparison_op_to_icmp_predicate: expected comparison operator, \
95 pub fn bin_op_to_fcmp_predicate(op: hir::BinOpKind) -> RealPredicate {
97 hir::BinOpKind::Eq => RealPredicate::RealOEQ,
98 hir::BinOpKind::Ne => RealPredicate::RealUNE,
99 hir::BinOpKind::Lt => RealPredicate::RealOLT,
100 hir::BinOpKind::Le => RealPredicate::RealOLE,
101 hir::BinOpKind::Gt => RealPredicate::RealOGT,
102 hir::BinOpKind::Ge => RealPredicate::RealOGE,
105 "comparison_op_to_fcmp_predicate: expected comparison operator, \
113 pub fn compare_simd_types<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
121 let signed = match t.kind {
123 let cmp = bin_op_to_fcmp_predicate(op);
124 let cmp = bx.fcmp(cmp, lhs, rhs);
125 return bx.sext(cmp, ret_ty);
127 ty::Uint(_) => false,
129 _ => bug!("compare_simd_types: invalid SIMD type"),
132 let cmp = bin_op_to_icmp_predicate(op, signed);
133 let cmp = bx.icmp(cmp, lhs, rhs);
134 // LLVM outputs an `< size x i1 >`, so we need to perform a sign extension
135 // to get the correctly sized type. This will compile to a single instruction
136 // once the IR is converted to assembly if the SIMD instruction is supported
137 // by the target architecture.
141 /// Retrieves the information we are losing (making dynamic) in an unsizing
144 /// The `old_info` argument is a bit odd. It is intended for use in an upcast,
145 /// where the new vtable for an object will be derived from the old one.
146 pub fn unsized_info<'tcx, Cx: CodegenMethods<'tcx>>(
150 old_info: Option<Cx::Value>,
152 let (source, target) =
153 cx.tcx().struct_lockstep_tails_erasing_lifetimes(source, target, cx.param_env());
154 match (&source.kind, &target.kind) {
155 (&ty::Array(_, len), &ty::Slice(_)) => {
156 cx.const_usize(len.eval_usize(cx.tcx(), ty::ParamEnv::reveal_all()))
158 (&ty::Dynamic(..), &ty::Dynamic(..)) => {
159 // For now, upcasts are limited to changes in marker
160 // traits, and hence never actually require an actual
161 // change to the vtable.
162 old_info.expect("unsized_info: missing old info for trait upcast")
164 (_, &ty::Dynamic(ref data, ..)) => {
165 let vtable_ptr = cx.layout_of(cx.tcx().mk_mut_ptr(target)).field(cx, FAT_PTR_EXTRA);
167 meth::get_vtable(cx, source, data.principal()),
168 cx.backend_type(vtable_ptr),
171 _ => bug!("unsized_info: invalid unsizing {:?} -> {:?}", source, target),
175 /// Coerces `src` to `dst_ty`. `src_ty` must be a thin pointer.
176 pub fn unsize_thin_ptr<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
181 ) -> (Bx::Value, Bx::Value) {
182 debug!("unsize_thin_ptr: {:?} => {:?}", src_ty, dst_ty);
183 match (&src_ty.kind, &dst_ty.kind) {
184 (&ty::Ref(_, a, _), &ty::Ref(_, b, _))
185 | (&ty::Ref(_, a, _), &ty::RawPtr(ty::TypeAndMut { ty: b, .. }))
186 | (&ty::RawPtr(ty::TypeAndMut { ty: a, .. }), &ty::RawPtr(ty::TypeAndMut { ty: b, .. })) => {
187 assert!(bx.cx().type_is_sized(a));
188 let ptr_ty = bx.cx().type_ptr_to(bx.cx().backend_type(bx.cx().layout_of(b)));
189 (bx.pointercast(src, ptr_ty), unsized_info(bx.cx(), a, b, None))
191 (&ty::Adt(def_a, _), &ty::Adt(def_b, _)) => {
192 assert_eq!(def_a, def_b);
194 let src_layout = bx.cx().layout_of(src_ty);
195 let dst_layout = bx.cx().layout_of(dst_ty);
196 let mut result = None;
197 for i in 0..src_layout.fields.count() {
198 let src_f = src_layout.field(bx.cx(), i);
199 assert_eq!(src_layout.fields.offset(i).bytes(), 0);
200 assert_eq!(dst_layout.fields.offset(i).bytes(), 0);
204 assert_eq!(src_layout.size, src_f.size);
206 let dst_f = dst_layout.field(bx.cx(), i);
207 assert_ne!(src_f.ty, dst_f.ty);
208 assert_eq!(result, None);
209 result = Some(unsize_thin_ptr(bx, src, src_f.ty, dst_f.ty));
211 let (lldata, llextra) = result.unwrap();
212 // HACK(eddyb) have to bitcast pointers until LLVM removes pointee types.
213 // FIXME(eddyb) move these out of this `match` arm, so they're always
214 // applied, uniformly, no matter the source/destination types.
216 bx.bitcast(lldata, bx.cx().scalar_pair_element_backend_type(dst_layout, 0, true)),
217 bx.bitcast(llextra, bx.cx().scalar_pair_element_backend_type(dst_layout, 1, true)),
220 _ => bug!("unsize_thin_ptr: called on bad types"),
224 /// Coerces `src`, which is a reference to a value of type `src_ty`,
225 /// to a value of type `dst_ty`, and stores the result in `dst`.
226 pub fn coerce_unsized_into<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
228 src: PlaceRef<'tcx, Bx::Value>,
229 dst: PlaceRef<'tcx, Bx::Value>,
231 let src_ty = src.layout.ty;
232 let dst_ty = dst.layout.ty;
233 match (&src_ty.kind, &dst_ty.kind) {
234 (&ty::Ref(..), &ty::Ref(..))
235 | (&ty::Ref(..), &ty::RawPtr(..))
236 | (&ty::RawPtr(..), &ty::RawPtr(..)) => {
237 let (base, info) = match bx.load_operand(src).val {
238 OperandValue::Pair(base, info) => {
239 // fat-ptr to fat-ptr unsize preserves the vtable
240 // i.e., &'a fmt::Debug+Send => &'a fmt::Debug
241 // So we need to pointercast the base to ensure
242 // the types match up.
243 // FIXME(eddyb) use `scalar_pair_element_backend_type` here,
244 // like `unsize_thin_ptr` does.
245 let thin_ptr = dst.layout.field(bx.cx(), FAT_PTR_ADDR);
246 (bx.pointercast(base, bx.cx().backend_type(thin_ptr)), info)
248 OperandValue::Immediate(base) => unsize_thin_ptr(bx, base, src_ty, dst_ty),
249 OperandValue::Ref(..) => bug!(),
251 OperandValue::Pair(base, info).store(bx, dst);
254 (&ty::Adt(def_a, _), &ty::Adt(def_b, _)) => {
255 assert_eq!(def_a, def_b);
257 for i in 0..def_a.variants[VariantIdx::new(0)].fields.len() {
258 let src_f = src.project_field(bx, i);
259 let dst_f = dst.project_field(bx, i);
261 if dst_f.layout.is_zst() {
265 if src_f.layout.ty == dst_f.layout.ty {
276 coerce_unsized_into(bx, src_f, dst_f);
280 _ => bug!("coerce_unsized_into: invalid coercion {:?} -> {:?}", src_ty, dst_ty,),
284 pub fn cast_shift_expr_rhs<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
290 cast_shift_rhs(bx, op, lhs, rhs)
293 fn cast_shift_rhs<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
299 // Shifts may have any size int on the rhs
301 let mut rhs_llty = bx.cx().val_ty(rhs);
302 let mut lhs_llty = bx.cx().val_ty(lhs);
303 if bx.cx().type_kind(rhs_llty) == TypeKind::Vector {
304 rhs_llty = bx.cx().element_type(rhs_llty)
306 if bx.cx().type_kind(lhs_llty) == TypeKind::Vector {
307 lhs_llty = bx.cx().element_type(lhs_llty)
309 let rhs_sz = bx.cx().int_width(rhs_llty);
310 let lhs_sz = bx.cx().int_width(lhs_llty);
312 bx.trunc(rhs, lhs_llty)
313 } else if lhs_sz > rhs_sz {
314 // FIXME (#1877: If in the future shifting by negative
315 // values is no longer undefined then this is wrong.
316 bx.zext(rhs, lhs_llty)
325 /// Returns `true` if this session's target will use SEH-based unwinding.
327 /// This is only true for MSVC targets, and even then the 64-bit MSVC target
328 /// currently uses SEH-ish unwinding with DWARF info tables to the side (same as
329 /// 64-bit MinGW) instead of "full SEH".
330 pub fn wants_msvc_seh(sess: &Session) -> bool {
331 sess.target.target.options.is_like_msvc
334 pub fn from_immediate<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
338 if bx.cx().val_ty(val) == bx.cx().type_i1() { bx.zext(val, bx.cx().type_i8()) } else { val }
341 pub fn to_immediate<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
344 layout: layout::TyLayout<'_>,
346 if let layout::Abi::Scalar(ref scalar) = layout.abi {
347 return to_immediate_scalar(bx, val, scalar);
352 pub fn to_immediate_scalar<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
355 scalar: &layout::Scalar,
357 if scalar.is_bool() {
358 return bx.trunc(val, bx.cx().type_i1());
363 pub fn memcpy_ty<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
369 layout: TyLayout<'tcx>,
372 let size = layout.size.bytes();
377 bx.memcpy(dst, dst_align, src, src_align, bx.cx().const_usize(size), flags);
380 pub fn codegen_instance<'a, 'tcx: 'a, Bx: BuilderMethods<'a, 'tcx>>(
381 cx: &'a Bx::CodegenCx,
382 instance: Instance<'tcx>,
384 // this is an info! to allow collecting monomorphization statistics
385 // and to allow finding the last function before LLVM aborts from
387 info!("codegen_instance({})", instance);
389 mir::codegen_mir::<Bx>(cx, instance);
392 /// Creates the `main` function which will initialize the rust runtime and call
393 /// users main function.
394 pub fn maybe_create_entry_wrapper<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(cx: &'a Bx::CodegenCx) {
395 let (main_def_id, span) = match cx.tcx().entry_fn(LOCAL_CRATE) {
396 Some((def_id, _)) => (def_id, cx.tcx().def_span(def_id)),
400 let instance = Instance::mono(cx.tcx(), main_def_id);
402 if !cx.codegen_unit().contains_item(&MonoItem::Fn(instance)) {
403 // We want to create the wrapper in the same codegen unit as Rust's main
408 let main_llfn = cx.get_fn_addr(instance);
410 let et = cx.tcx().entry_fn(LOCAL_CRATE).map(|e| e.1);
412 Some(EntryFnType::Main) => create_entry_fn::<Bx>(cx, span, main_llfn, main_def_id, true),
413 Some(EntryFnType::Start) => create_entry_fn::<Bx>(cx, span, main_llfn, main_def_id, false),
414 None => {} // Do nothing.
417 fn create_entry_fn<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
418 cx: &'a Bx::CodegenCx,
420 rust_main: Bx::Value,
421 rust_main_def_id: DefId,
422 use_start_lang_item: bool,
424 // The entry function is either `int main(void)` or `int main(int argc, char **argv)`,
425 // depending on whether the target needs `argc` and `argv` to be passed in.
426 let llfty = if cx.sess().target.target.options.main_needs_argc_argv {
427 cx.type_func(&[cx.type_int(), cx.type_ptr_to(cx.type_i8p())], cx.type_int())
429 cx.type_func(&[], cx.type_int())
432 let main_ret_ty = cx.tcx().fn_sig(rust_main_def_id).output();
433 // Given that `main()` has no arguments,
434 // then its return type cannot have
435 // late-bound regions, since late-bound
436 // regions must appear in the argument
438 let main_ret_ty = cx.tcx().erase_regions(&main_ret_ty.no_bound_vars().unwrap());
440 if cx.get_defined_value("main").is_some() {
441 // FIXME: We should be smart and show a better diagnostic here.
443 .struct_span_err(sp, "entry symbol `main` defined multiple times")
444 .help("did you use `#[no_mangle]` on `fn main`? Use `#[start]` instead")
446 cx.sess().abort_if_errors();
449 let llfn = cx.declare_cfn("main", llfty);
451 // `main` should respect same config for frame pointer elimination as rest of code
452 cx.set_frame_pointer_elimination(llfn);
453 cx.apply_target_cpu_attr(llfn);
455 let mut bx = Bx::new_block(&cx, llfn, "top");
457 bx.insert_reference_to_gdb_debug_scripts_section_global();
459 let (arg_argc, arg_argv) = get_argc_argv(cx, &mut bx);
461 let (start_fn, args) = if use_start_lang_item {
462 let start_def_id = cx.tcx().require_lang_item(StartFnLangItem, None);
463 let start_fn = cx.get_fn_addr(
464 ty::Instance::resolve(
466 ty::ParamEnv::reveal_all(),
468 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);
487 /// Obtain the `argc` and `argv` values to pass to the rust start function.
488 fn get_argc_argv<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
489 cx: &'a Bx::CodegenCx,
491 ) -> (Bx::Value, Bx::Value) {
492 if cx.sess().target.target.options.main_needs_argc_argv {
493 // Params from native `main()` used as args for rust start function
494 let param_argc = bx.get_param(0);
495 let param_argv = bx.get_param(1);
496 let arg_argc = bx.intcast(param_argc, cx.type_isize(), true);
497 let arg_argv = param_argv;
500 // The Rust start function doesn't need `argc` and `argv`, so just pass zeros.
501 let arg_argc = bx.const_int(cx.type_int(), 0);
502 let arg_argv = bx.const_null(cx.type_ptr_to(cx.type_i8p()));
507 pub const CODEGEN_WORKER_ID: usize = ::std::usize::MAX;
509 pub fn codegen_crate<B: ExtraBackendMethods>(
512 metadata: EncodedMetadata,
513 need_metadata_module: bool,
514 ) -> OngoingCodegen<B> {
515 check_for_rustc_errors_attr(tcx);
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;
535 let codegen_units = (*codegen_units).clone();
537 // Force all codegen_unit queries so they are already either red or green
538 // when compile_codegen_unit accesses them. We are not able to re-execute
539 // the codegen_unit query from just the DepNode, so an unknown color would
540 // lead to having to re-execute compile_codegen_unit, possibly
542 if tcx.dep_graph.is_fully_enabled() {
543 for cgu in &codegen_units {
544 tcx.codegen_unit(cgu.name());
548 let ongoing_codegen = start_async_codegen(backend.clone(), tcx, metadata, codegen_units.len());
549 let ongoing_codegen = AbortCodegenOnDrop::<B>(Some(ongoing_codegen));
551 // Codegen an allocator shim, if necessary.
553 // If the crate doesn't have an `allocator_kind` set then there's definitely
554 // no shim to generate. Otherwise we also check our dependency graph for all
555 // our output crate types. If anything there looks like its a `Dynamic`
556 // linkage, then it's already got an allocator shim and we'll be using that
557 // one instead. If nothing exists then it's our job to generate the
559 let any_dynamic_crate = tcx.dependency_formats(LOCAL_CRATE).iter().any(|(_, list)| {
560 use rustc::middle::dependency_format::Linkage;
561 list.iter().any(|&linkage| linkage == Linkage::Dynamic)
563 let allocator_module = if any_dynamic_crate {
565 } else if let Some(kind) = tcx.allocator_kind() {
567 cgu_name_builder.build_cgu_name(LOCAL_CRATE, &["crate"], Some("allocator")).to_string();
568 let mut modules = backend.new_metadata(tcx, &llmod_id);
570 .time("write_allocator_module", || backend.codegen_allocator(tcx, &mut modules, kind));
572 Some(ModuleCodegen { name: llmod_id, module_llvm: modules, kind: ModuleKind::Allocator })
577 if let Some(allocator_module) = allocator_module {
578 ongoing_codegen.submit_pre_codegened_module_to_llvm(tcx, allocator_module);
581 if need_metadata_module {
582 // Codegen the encoded metadata.
583 let metadata_cgu_name =
584 cgu_name_builder.build_cgu_name(LOCAL_CRATE, &["crate"], Some("metadata")).to_string();
585 let mut metadata_llvm_module = backend.new_metadata(tcx, &metadata_cgu_name);
586 tcx.sess.time("write_compressed_metadata", || {
587 backend.write_compressed_metadata(
589 &ongoing_codegen.metadata,
590 &mut metadata_llvm_module,
594 let metadata_module = ModuleCodegen {
595 name: metadata_cgu_name,
596 module_llvm: metadata_llvm_module,
597 kind: ModuleKind::Metadata,
599 ongoing_codegen.submit_pre_codegened_module_to_llvm(tcx, metadata_module);
602 // We sort the codegen units by size. This way we can schedule work for LLVM
603 // a bit more efficiently.
604 let codegen_units = {
605 let mut codegen_units = codegen_units;
606 codegen_units.sort_by_cached_key(|cgu| cmp::Reverse(cgu.size_estimate()));
610 let total_codegen_time = Lock::new(Duration::new(0, 0));
612 // The non-parallel compiler can only translate codegen units to LLVM IR
613 // on a single thread, leading to a staircase effect where the N LLVM
614 // threads have to wait on the single codegen threads to generate work
615 // for them. The parallel compiler does not have this restriction, so
616 // we can pre-load the LLVM queue in parallel before handing off
617 // coordination to the OnGoingCodegen scheduler.
619 // This likely is a temporary measure. Once we don't have to support the
620 // non-parallel compiler anymore, we can compile CGUs end-to-end in
621 // parallel and get rid of the complicated scheduling logic.
622 let pre_compile_cgus = |cgu_reuse: &[CguReuse]| {
623 if cfg!(parallel_compiler) {
624 tcx.sess.time("compile_first_CGU_batch", || {
625 // Try to find one CGU to compile per thread.
626 let cgus: Vec<_> = cgu_reuse
629 .filter(|&(_, reuse)| reuse == &CguReuse::No)
630 .take(tcx.sess.threads())
633 // Compile the found CGUs in parallel.
636 let start_time = Instant::now();
637 let module = backend.compile_codegen_unit(tcx, codegen_units[i].name());
638 let mut time = total_codegen_time.lock();
639 *time += start_time.elapsed();
649 let mut cgu_reuse = Vec::new();
650 let mut pre_compiled_cgus: Option<FxHashMap<usize, _>> = None;
652 for (i, cgu) in codegen_units.iter().enumerate() {
653 ongoing_codegen.wait_for_signal_to_codegen_item();
654 ongoing_codegen.check_for_errors(tcx.sess);
656 // Do some setup work in the first iteration
657 if pre_compiled_cgus.is_none() {
658 // Calculate the CGU reuse
659 cgu_reuse = tcx.sess.time("find_cgu_reuse", || {
660 codegen_units.iter().map(|cgu| determine_cgu_reuse(tcx, &cgu)).collect()
662 // Pre compile some CGUs
663 pre_compiled_cgus = Some(pre_compile_cgus(&cgu_reuse));
666 let cgu_reuse = cgu_reuse[i];
667 tcx.sess.cgu_reuse_tracker.set_actual_reuse(&cgu.name().as_str(), cgu_reuse);
672 if let Some(cgu) = pre_compiled_cgus.as_mut().unwrap().remove(&i) {
675 let start_time = Instant::now();
676 let module = backend.compile_codegen_unit(tcx, cgu.name());
677 let mut time = total_codegen_time.lock();
678 *time += start_time.elapsed();
681 submit_codegened_module_to_llvm(
683 &ongoing_codegen.coordinator_send,
689 CguReuse::PreLto => {
690 submit_pre_lto_module_to_llvm(
693 &ongoing_codegen.coordinator_send,
694 CachedModuleCodegen {
695 name: cgu.name().to_string(),
696 source: cgu.work_product(tcx),
701 CguReuse::PostLto => {
702 submit_post_lto_module_to_llvm(
704 &ongoing_codegen.coordinator_send,
705 CachedModuleCodegen {
706 name: cgu.name().to_string(),
707 source: cgu.work_product(tcx),
715 ongoing_codegen.codegen_finished(tcx);
717 // Since the main thread is sometimes blocked during codegen, we keep track
718 // -Ztime-passes output manually.
719 print_time_passes_entry(
720 tcx.sess.time_passes(),
721 "codegen_to_LLVM_IR",
722 total_codegen_time.into_inner(),
725 ::rustc_incremental::assert_module_sources::assert_module_sources(tcx);
727 symbol_names_test::report_symbol_names(tcx);
729 ongoing_codegen.check_for_errors(tcx.sess);
733 ongoing_codegen.into_inner()
736 /// A curious wrapper structure whose only purpose is to call `codegen_aborted`
737 /// when it's dropped abnormally.
739 /// In the process of working on rust-lang/rust#55238 a mysterious segfault was
740 /// stumbled upon. The segfault was never reproduced locally, but it was
741 /// suspected to be related to the fact that codegen worker threads were
742 /// sticking around by the time the main thread was exiting, causing issues.
744 /// This structure is an attempt to fix that issue where the `codegen_aborted`
745 /// message will block until all workers have finished. This should ensure that
746 /// even if the main codegen thread panics we'll wait for pending work to
747 /// complete before returning from the main thread, hopefully avoiding
750 /// If you see this comment in the code, then it means that this workaround
751 /// worked! We may yet one day track down the mysterious cause of that
753 struct AbortCodegenOnDrop<B: ExtraBackendMethods>(Option<OngoingCodegen<B>>);
755 impl<B: ExtraBackendMethods> AbortCodegenOnDrop<B> {
756 fn into_inner(mut self) -> OngoingCodegen<B> {
757 self.0.take().unwrap()
761 impl<B: ExtraBackendMethods> Deref for AbortCodegenOnDrop<B> {
762 type Target = OngoingCodegen<B>;
764 fn deref(&self) -> &OngoingCodegen<B> {
765 self.0.as_ref().unwrap()
769 impl<B: ExtraBackendMethods> DerefMut for AbortCodegenOnDrop<B> {
770 fn deref_mut(&mut self) -> &mut OngoingCodegen<B> {
771 self.0.as_mut().unwrap()
775 impl<B: ExtraBackendMethods> Drop for AbortCodegenOnDrop<B> {
777 if let Some(codegen) = self.0.take() {
778 codegen.codegen_aborted();
783 fn finalize_tcx(tcx: TyCtxt<'_>) {
784 tcx.sess.time("assert_dep_graph", || ::rustc_incremental::assert_dep_graph(tcx));
785 tcx.sess.time("serialize_dep_graph", || ::rustc_incremental::save_dep_graph(tcx));
787 // We assume that no queries are run past here. If there are new queries
788 // after this point, they'll show up as "<unknown>" in self-profiling data.
790 let _prof_timer = tcx.prof.generic_activity("self_profile_alloc_query_strings");
791 tcx.alloc_self_profile_query_strings();
796 pub fn new(tcx: TyCtxt<'_>) -> CrateInfo {
797 let mut info = CrateInfo {
799 compiler_builtins: None,
800 profiler_runtime: None,
801 is_no_builtins: Default::default(),
802 native_libraries: Default::default(),
803 used_libraries: tcx.native_libraries(LOCAL_CRATE),
804 link_args: tcx.link_args(LOCAL_CRATE),
805 crate_name: Default::default(),
806 used_crates_dynamic: cstore::used_crates(tcx, LinkagePreference::RequireDynamic),
807 used_crates_static: cstore::used_crates(tcx, LinkagePreference::RequireStatic),
808 used_crate_source: Default::default(),
809 lang_item_to_crate: Default::default(),
810 missing_lang_items: Default::default(),
811 dependency_formats: tcx.dependency_formats(LOCAL_CRATE),
813 let lang_items = tcx.lang_items();
815 let crates = tcx.crates();
817 let n_crates = crates.len();
818 info.native_libraries.reserve(n_crates);
819 info.crate_name.reserve(n_crates);
820 info.used_crate_source.reserve(n_crates);
821 info.missing_lang_items.reserve(n_crates);
823 for &cnum in crates.iter() {
824 info.native_libraries.insert(cnum, tcx.native_libraries(cnum));
825 info.crate_name.insert(cnum, tcx.crate_name(cnum).to_string());
826 info.used_crate_source.insert(cnum, tcx.used_crate_source(cnum));
827 if tcx.is_panic_runtime(cnum) {
828 info.panic_runtime = Some(cnum);
830 if tcx.is_compiler_builtins(cnum) {
831 info.compiler_builtins = Some(cnum);
833 if tcx.is_profiler_runtime(cnum) {
834 info.profiler_runtime = Some(cnum);
836 if tcx.is_no_builtins(cnum) {
837 info.is_no_builtins.insert(cnum);
839 let missing = tcx.missing_lang_items(cnum);
840 for &item in missing.iter() {
841 if let Ok(id) = lang_items.require(item) {
842 info.lang_item_to_crate.insert(item, id.krate);
846 // No need to look for lang items that are whitelisted and don't
847 // actually need to exist.
848 let missing = missing
851 .filter(|&l| !weak_lang_items::whitelisted(tcx, l))
853 info.missing_lang_items.insert(cnum, missing);
860 pub fn provide_both(providers: &mut Providers<'_>) {
861 providers.backend_optimization_level = |tcx, cratenum| {
862 let for_speed = match tcx.sess.opts.optimize {
863 // If globally no optimisation is done, #[optimize] has no effect.
865 // This is done because if we ended up "upgrading" to `-O2` here, we’d populate the
866 // pass manager and it is likely that some module-wide passes (such as inliner or
867 // cross-function constant propagation) would ignore the `optnone` annotation we put
868 // on the functions, thus necessarily involving these functions into optimisations.
869 config::OptLevel::No => return config::OptLevel::No,
870 // If globally optimise-speed is already specified, just use that level.
871 config::OptLevel::Less => return config::OptLevel::Less,
872 config::OptLevel::Default => return config::OptLevel::Default,
873 config::OptLevel::Aggressive => return config::OptLevel::Aggressive,
874 // If globally optimize-for-size has been requested, use -O2 instead (if optimize(size)
876 config::OptLevel::Size => config::OptLevel::Default,
877 config::OptLevel::SizeMin => config::OptLevel::Default,
880 let (defids, _) = tcx.collect_and_partition_mono_items(cratenum);
882 let CodegenFnAttrs { optimize, .. } = tcx.codegen_fn_attrs(*id);
884 attr::OptimizeAttr::None => continue,
885 attr::OptimizeAttr::Size => continue,
886 attr::OptimizeAttr::Speed => {
891 return tcx.sess.opts.optimize;
894 providers.dllimport_foreign_items = |tcx, krate| {
895 let module_map = tcx.foreign_modules(krate);
897 module_map.iter().map(|lib| (lib.def_id, lib)).collect::<FxHashMap<_, _>>();
900 .native_libraries(krate)
903 if lib.kind != cstore::NativeLibraryKind::NativeUnknown {
906 let cfg = match lib.cfg {
907 Some(ref cfg) => cfg,
910 attr::cfg_matches(cfg, &tcx.sess.parse_sess, None)
912 .filter_map(|lib| lib.foreign_module)
913 .map(|id| &module_map[&id])
914 .flat_map(|module| module.foreign_items.iter().cloned())
916 tcx.arena.alloc(dllimports)
919 providers.is_dllimport_foreign_item =
920 |tcx, def_id| tcx.dllimport_foreign_items(def_id.krate).contains(&def_id);
923 fn determine_cgu_reuse<'tcx>(tcx: TyCtxt<'tcx>, cgu: &CodegenUnit<'tcx>) -> CguReuse {
924 if !tcx.dep_graph.is_fully_enabled() {
928 let work_product_id = &cgu.work_product_id();
929 if tcx.dep_graph.previous_work_product(work_product_id).is_none() {
930 // We don't have anything cached for this CGU. This can happen
931 // if the CGU did not exist in the previous session.
935 // Try to mark the CGU as green. If it we can do so, it means that nothing
936 // affecting the LLVM module has changed and we can re-use a cached version.
937 // If we compile with any kind of LTO, this means we can re-use the bitcode
938 // of the Pre-LTO stage (possibly also the Post-LTO version but we'll only
939 // know that later). If we are not doing LTO, there is only one optimized
940 // version of each module, so we re-use that.
941 let dep_node = cgu.codegen_dep_node(tcx);
943 !tcx.dep_graph.dep_node_exists(&dep_node),
944 "CompileCodegenUnit dep-node for CGU `{}` already exists before marking.",
948 if tcx.dep_graph.try_mark_green(tcx, &dep_node).is_some() {
949 // We can re-use either the pre- or the post-thinlto state
950 if tcx.sess.lto() != Lto::No { CguReuse::PreLto } else { CguReuse::PostLto }