]> git.proxmox.com Git - rustc.git/blob - compiler/rustc_codegen_ssa/src/base.rs
Update upstream source from tag 'upstream/1.49.0_beta.4+dfsg1'
[rustc.git] / compiler / rustc_codegen_ssa / src / base.rs
1 //! Codegen the completed AST to the LLVM IR.
2 //!
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.
7 //!
8 //! Hopefully useful general knowledge about codegen:
9 //!
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`.
15
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,
19 };
20 use crate::common::{IntPredicate, RealPredicate, TypeKind};
21 use crate::meth;
22 use crate::mir;
23 use crate::mir::operand::OperandValue;
24 use crate::mir::place::PlaceRef;
25 use crate::traits::*;
26 use crate::{CachedModuleCodegen, CrateInfo, MemFlags, ModuleCodegen, ModuleKind};
27
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};
32 use rustc_hir as hir;
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::{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;
49 use rustc_symbol_mangling::test as symbol_names_test;
50 use rustc_target::abi::{Align, LayoutOf, VariantIdx};
51
52 use std::cmp;
53 use std::ops::{Deref, DerefMut};
54 use std::time::{Duration, Instant};
55
56 pub fn bin_op_to_icmp_predicate(op: hir::BinOpKind, signed: bool) -> IntPredicate {
57 match op {
58 hir::BinOpKind::Eq => IntPredicate::IntEQ,
59 hir::BinOpKind::Ne => IntPredicate::IntNE,
60 hir::BinOpKind::Lt => {
61 if signed {
62 IntPredicate::IntSLT
63 } else {
64 IntPredicate::IntULT
65 }
66 }
67 hir::BinOpKind::Le => {
68 if signed {
69 IntPredicate::IntSLE
70 } else {
71 IntPredicate::IntULE
72 }
73 }
74 hir::BinOpKind::Gt => {
75 if signed {
76 IntPredicate::IntSGT
77 } else {
78 IntPredicate::IntUGT
79 }
80 }
81 hir::BinOpKind::Ge => {
82 if signed {
83 IntPredicate::IntSGE
84 } else {
85 IntPredicate::IntUGE
86 }
87 }
88 op => bug!(
89 "comparison_op_to_icmp_predicate: expected comparison operator, \
90 found {:?}",
91 op
92 ),
93 }
94 }
95
96 pub fn bin_op_to_fcmp_predicate(op: hir::BinOpKind) -> RealPredicate {
97 match op {
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,
104 op => {
105 bug!(
106 "comparison_op_to_fcmp_predicate: expected comparison operator, \
107 found {:?}",
108 op
109 );
110 }
111 }
112 }
113
114 pub fn compare_simd_types<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
115 bx: &mut Bx,
116 lhs: Bx::Value,
117 rhs: Bx::Value,
118 t: Ty<'tcx>,
119 ret_ty: Bx::Type,
120 op: hir::BinOpKind,
121 ) -> Bx::Value {
122 let signed = match t.kind() {
123 ty::Float(_) => {
124 let cmp = bin_op_to_fcmp_predicate(op);
125 let cmp = bx.fcmp(cmp, lhs, rhs);
126 return bx.sext(cmp, ret_ty);
127 }
128 ty::Uint(_) => false,
129 ty::Int(_) => true,
130 _ => bug!("compare_simd_types: invalid SIMD type"),
131 };
132
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.
139 bx.sext(cmp, ret_ty)
140 }
141
142 /// Retrieves the information we are losing (making dynamic) in an unsizing
143 /// adjustment.
144 ///
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>>(
148 cx: &Cx,
149 source: Ty<'tcx>,
150 target: Ty<'tcx>,
151 old_info: Option<Cx::Value>,
152 ) -> 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()))
158 }
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")
164 }
165 (_, &ty::Dynamic(ref data, ..)) => {
166 let vtable_ptr = cx.layout_of(cx.tcx().mk_mut_ptr(target)).field(cx, FAT_PTR_EXTRA);
167 cx.const_ptrcast(
168 meth::get_vtable(cx, source, data.principal()),
169 cx.backend_type(vtable_ptr),
170 )
171 }
172 _ => bug!("unsized_info: invalid unsizing {:?} -> {:?}", source, target),
173 }
174 }
175
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>>(
178 bx: &mut Bx,
179 src: Bx::Value,
180 src_ty: Ty<'tcx>,
181 dst_ty: Ty<'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, _) | &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))
190 }
191 (&ty::Adt(def_a, _), &ty::Adt(def_b, _)) => {
192 assert_eq!(def_a, def_b);
193
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);
201 if src_f.is_zst() {
202 continue;
203 }
204 assert_eq!(src_layout.size, src_f.size);
205
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));
210 }
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.
215 (
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)),
218 )
219 }
220 _ => bug!("unsize_thin_ptr: called on bad types"),
221 }
222 }
223
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>>(
227 bx: &mut Bx,
228 src: PlaceRef<'tcx, Bx::Value>,
229 dst: PlaceRef<'tcx, Bx::Value>,
230 ) {
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(..) | &ty::RawPtr(..)) | (&ty::RawPtr(..), &ty::RawPtr(..)) => {
235 let (base, info) = match bx.load_operand(src).val {
236 OperandValue::Pair(base, info) => {
237 // fat-ptr to fat-ptr unsize preserves the vtable
238 // i.e., &'a fmt::Debug+Send => &'a fmt::Debug
239 // So we need to pointercast the base to ensure
240 // the types match up.
241 // FIXME(eddyb) use `scalar_pair_element_backend_type` here,
242 // like `unsize_thin_ptr` does.
243 let thin_ptr = dst.layout.field(bx.cx(), FAT_PTR_ADDR);
244 (bx.pointercast(base, bx.cx().backend_type(thin_ptr)), info)
245 }
246 OperandValue::Immediate(base) => unsize_thin_ptr(bx, base, src_ty, dst_ty),
247 OperandValue::Ref(..) => bug!(),
248 };
249 OperandValue::Pair(base, info).store(bx, dst);
250 }
251
252 (&ty::Adt(def_a, _), &ty::Adt(def_b, _)) => {
253 assert_eq!(def_a, def_b);
254
255 for i in 0..def_a.variants[VariantIdx::new(0)].fields.len() {
256 let src_f = src.project_field(bx, i);
257 let dst_f = dst.project_field(bx, i);
258
259 if dst_f.layout.is_zst() {
260 continue;
261 }
262
263 if src_f.layout.ty == dst_f.layout.ty {
264 memcpy_ty(
265 bx,
266 dst_f.llval,
267 dst_f.align,
268 src_f.llval,
269 src_f.align,
270 src_f.layout,
271 MemFlags::empty(),
272 );
273 } else {
274 coerce_unsized_into(bx, src_f, dst_f);
275 }
276 }
277 }
278 _ => bug!("coerce_unsized_into: invalid coercion {:?} -> {:?}", src_ty, dst_ty,),
279 }
280 }
281
282 pub fn cast_shift_expr_rhs<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
283 bx: &mut Bx,
284 op: hir::BinOpKind,
285 lhs: Bx::Value,
286 rhs: Bx::Value,
287 ) -> Bx::Value {
288 cast_shift_rhs(bx, op, lhs, rhs)
289 }
290
291 fn cast_shift_rhs<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
292 bx: &mut Bx,
293 op: hir::BinOpKind,
294 lhs: Bx::Value,
295 rhs: Bx::Value,
296 ) -> Bx::Value {
297 // Shifts may have any size int on the rhs
298 if op.is_shift() {
299 let mut rhs_llty = bx.cx().val_ty(rhs);
300 let mut lhs_llty = bx.cx().val_ty(lhs);
301 if bx.cx().type_kind(rhs_llty) == TypeKind::Vector {
302 rhs_llty = bx.cx().element_type(rhs_llty)
303 }
304 if bx.cx().type_kind(lhs_llty) == TypeKind::Vector {
305 lhs_llty = bx.cx().element_type(lhs_llty)
306 }
307 let rhs_sz = bx.cx().int_width(rhs_llty);
308 let lhs_sz = bx.cx().int_width(lhs_llty);
309 if lhs_sz < rhs_sz {
310 bx.trunc(rhs, lhs_llty)
311 } else if lhs_sz > rhs_sz {
312 // FIXME (#1877: If in the future shifting by negative
313 // values is no longer undefined then this is wrong.
314 bx.zext(rhs, lhs_llty)
315 } else {
316 rhs
317 }
318 } else {
319 rhs
320 }
321 }
322
323 /// Returns `true` if this session's target will use SEH-based unwinding.
324 ///
325 /// This is only true for MSVC targets, and even then the 64-bit MSVC target
326 /// currently uses SEH-ish unwinding with DWARF info tables to the side (same as
327 /// 64-bit MinGW) instead of "full SEH".
328 pub fn wants_msvc_seh(sess: &Session) -> bool {
329 sess.target.is_like_msvc
330 }
331
332 pub fn memcpy_ty<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
333 bx: &mut Bx,
334 dst: Bx::Value,
335 dst_align: Align,
336 src: Bx::Value,
337 src_align: Align,
338 layout: TyAndLayout<'tcx>,
339 flags: MemFlags,
340 ) {
341 let size = layout.size.bytes();
342 if size == 0 {
343 return;
344 }
345
346 bx.memcpy(dst, dst_align, src, src_align, bx.cx().const_usize(size), flags);
347 }
348
349 pub fn codegen_instance<'a, 'tcx: 'a, Bx: BuilderMethods<'a, 'tcx>>(
350 cx: &'a Bx::CodegenCx,
351 instance: Instance<'tcx>,
352 ) {
353 // this is an info! to allow collecting monomorphization statistics
354 // and to allow finding the last function before LLVM aborts from
355 // release builds.
356 info!("codegen_instance({})", instance);
357
358 mir::codegen_mir::<Bx>(cx, instance);
359 }
360
361 /// Creates the `main` function which will initialize the rust runtime and call
362 /// users main function.
363 pub fn maybe_create_entry_wrapper<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
364 cx: &'a Bx::CodegenCx,
365 ) -> Option<Bx::Function> {
366 let main_def_id = cx.tcx().entry_fn(LOCAL_CRATE).map(|(def_id, _)| def_id)?;
367 let instance = Instance::mono(cx.tcx(), main_def_id.to_def_id());
368
369 if !cx.codegen_unit().contains_item(&MonoItem::Fn(instance)) {
370 // We want to create the wrapper in the same codegen unit as Rust's main
371 // function.
372 return None;
373 }
374
375 let main_llfn = cx.get_fn_addr(instance);
376
377 return cx.tcx().entry_fn(LOCAL_CRATE).map(|(_, et)| {
378 let use_start_lang_item = EntryFnType::Start != et;
379 create_entry_fn::<Bx>(cx, main_llfn, main_def_id, use_start_lang_item)
380 });
381
382 fn create_entry_fn<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
383 cx: &'a Bx::CodegenCx,
384 rust_main: Bx::Value,
385 rust_main_def_id: LocalDefId,
386 use_start_lang_item: bool,
387 ) -> Bx::Function {
388 // The entry function is either `int main(void)` or `int main(int argc, char **argv)`,
389 // depending on whether the target needs `argc` and `argv` to be passed in.
390 let llfty = if cx.sess().target.main_needs_argc_argv {
391 cx.type_func(&[cx.type_int(), cx.type_ptr_to(cx.type_i8p())], cx.type_int())
392 } else {
393 cx.type_func(&[], cx.type_int())
394 };
395
396 let main_ret_ty = cx.tcx().fn_sig(rust_main_def_id).output();
397 // Given that `main()` has no arguments,
398 // then its return type cannot have
399 // late-bound regions, since late-bound
400 // regions must appear in the argument
401 // listing.
402 let main_ret_ty = cx.tcx().erase_regions(&main_ret_ty.no_bound_vars().unwrap());
403
404 let llfn = match cx.declare_c_main(llfty) {
405 Some(llfn) => llfn,
406 None => {
407 // FIXME: We should be smart and show a better diagnostic here.
408 let span = cx.tcx().def_span(rust_main_def_id);
409 cx.sess()
410 .struct_span_err(span, "entry symbol `main` declared multiple times")
411 .help("did you use `#[no_mangle]` on `fn main`? Use `#[start]` instead")
412 .emit();
413 cx.sess().abort_if_errors();
414 bug!();
415 }
416 };
417
418 // `main` should respect same config for frame pointer elimination as rest of code
419 cx.set_frame_pointer_elimination(llfn);
420 cx.apply_target_cpu_attr(llfn);
421
422 let mut bx = Bx::new_block(&cx, llfn, "top");
423
424 bx.insert_reference_to_gdb_debug_scripts_section_global();
425
426 let (arg_argc, arg_argv) = get_argc_argv(cx, &mut bx);
427
428 let (start_fn, args) = if use_start_lang_item {
429 let start_def_id = cx.tcx().require_lang_item(LangItem::Start, None);
430 let start_fn = cx.get_fn_addr(
431 ty::Instance::resolve(
432 cx.tcx(),
433 ty::ParamEnv::reveal_all(),
434 start_def_id,
435 cx.tcx().intern_substs(&[main_ret_ty.into()]),
436 )
437 .unwrap()
438 .unwrap(),
439 );
440 (
441 start_fn,
442 vec![bx.pointercast(rust_main, cx.type_ptr_to(cx.type_i8p())), arg_argc, arg_argv],
443 )
444 } else {
445 debug!("using user-defined start fn");
446 (rust_main, vec![arg_argc, arg_argv])
447 };
448
449 let result = bx.call(start_fn, &args, None);
450 let cast = bx.intcast(result, cx.type_int(), true);
451 bx.ret(cast);
452
453 llfn
454 }
455 }
456
457 /// Obtain the `argc` and `argv` values to pass to the rust start function.
458 fn get_argc_argv<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
459 cx: &'a Bx::CodegenCx,
460 bx: &mut Bx,
461 ) -> (Bx::Value, Bx::Value) {
462 if cx.sess().target.main_needs_argc_argv {
463 // Params from native `main()` used as args for rust start function
464 let param_argc = bx.get_param(0);
465 let param_argv = bx.get_param(1);
466 let arg_argc = bx.intcast(param_argc, cx.type_isize(), true);
467 let arg_argv = param_argv;
468 (arg_argc, arg_argv)
469 } else {
470 // The Rust start function doesn't need `argc` and `argv`, so just pass zeros.
471 let arg_argc = bx.const_int(cx.type_int(), 0);
472 let arg_argv = bx.const_null(cx.type_ptr_to(cx.type_i8p()));
473 (arg_argc, arg_argv)
474 }
475 }
476
477 pub fn codegen_crate<B: ExtraBackendMethods>(
478 backend: B,
479 tcx: TyCtxt<'tcx>,
480 metadata: EncodedMetadata,
481 need_metadata_module: bool,
482 ) -> OngoingCodegen<B> {
483 // Skip crate items and just output metadata in -Z no-codegen mode.
484 if tcx.sess.opts.debugging_opts.no_codegen || !tcx.sess.opts.output_types.should_codegen() {
485 let ongoing_codegen = start_async_codegen(backend, tcx, metadata, 1);
486
487 ongoing_codegen.codegen_finished(tcx);
488
489 finalize_tcx(tcx);
490
491 ongoing_codegen.check_for_errors(tcx.sess);
492
493 return ongoing_codegen;
494 }
495
496 let cgu_name_builder = &mut CodegenUnitNameBuilder::new(tcx);
497
498 // Run the monomorphization collector and partition the collected items into
499 // codegen units.
500 let codegen_units = tcx.collect_and_partition_mono_items(LOCAL_CRATE).1;
501
502 // Force all codegen_unit queries so they are already either red or green
503 // when compile_codegen_unit accesses them. We are not able to re-execute
504 // the codegen_unit query from just the DepNode, so an unknown color would
505 // lead to having to re-execute compile_codegen_unit, possibly
506 // unnecessarily.
507 if tcx.dep_graph.is_fully_enabled() {
508 for cgu in codegen_units {
509 tcx.ensure().codegen_unit(cgu.name());
510 }
511 }
512
513 let ongoing_codegen = start_async_codegen(backend.clone(), tcx, metadata, codegen_units.len());
514 let ongoing_codegen = AbortCodegenOnDrop::<B>(Some(ongoing_codegen));
515
516 // Codegen an allocator shim, if necessary.
517 //
518 // If the crate doesn't have an `allocator_kind` set then there's definitely
519 // no shim to generate. Otherwise we also check our dependency graph for all
520 // our output crate types. If anything there looks like its a `Dynamic`
521 // linkage, then it's already got an allocator shim and we'll be using that
522 // one instead. If nothing exists then it's our job to generate the
523 // allocator!
524 let any_dynamic_crate = tcx.dependency_formats(LOCAL_CRATE).iter().any(|(_, list)| {
525 use rustc_middle::middle::dependency_format::Linkage;
526 list.iter().any(|&linkage| linkage == Linkage::Dynamic)
527 });
528 let allocator_module = if any_dynamic_crate {
529 None
530 } else if let Some(kind) = tcx.allocator_kind() {
531 let llmod_id =
532 cgu_name_builder.build_cgu_name(LOCAL_CRATE, &["crate"], Some("allocator")).to_string();
533 let mut modules = backend.new_metadata(tcx, &llmod_id);
534 tcx.sess.time("write_allocator_module", || {
535 backend.codegen_allocator(tcx, &mut modules, kind, tcx.lang_items().oom().is_some())
536 });
537
538 Some(ModuleCodegen { name: llmod_id, module_llvm: modules, kind: ModuleKind::Allocator })
539 } else {
540 None
541 };
542
543 if let Some(allocator_module) = allocator_module {
544 ongoing_codegen.submit_pre_codegened_module_to_llvm(tcx, allocator_module);
545 }
546
547 if need_metadata_module {
548 // Codegen the encoded metadata.
549 let metadata_cgu_name =
550 cgu_name_builder.build_cgu_name(LOCAL_CRATE, &["crate"], Some("metadata")).to_string();
551 let mut metadata_llvm_module = backend.new_metadata(tcx, &metadata_cgu_name);
552 tcx.sess.time("write_compressed_metadata", || {
553 backend.write_compressed_metadata(
554 tcx,
555 &ongoing_codegen.metadata,
556 &mut metadata_llvm_module,
557 );
558 });
559
560 let metadata_module = ModuleCodegen {
561 name: metadata_cgu_name,
562 module_llvm: metadata_llvm_module,
563 kind: ModuleKind::Metadata,
564 };
565 ongoing_codegen.submit_pre_codegened_module_to_llvm(tcx, metadata_module);
566 }
567
568 // We sort the codegen units by size. This way we can schedule work for LLVM
569 // a bit more efficiently.
570 let codegen_units = {
571 let mut codegen_units = codegen_units.iter().collect::<Vec<_>>();
572 codegen_units.sort_by_cached_key(|cgu| cmp::Reverse(cgu.size_estimate()));
573 codegen_units
574 };
575
576 let total_codegen_time = Lock::new(Duration::new(0, 0));
577
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.
584 //
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
593 .iter()
594 .enumerate()
595 .filter(|&(_, reuse)| reuse == &CguReuse::No)
596 .take(tcx.sess.threads())
597 .collect();
598
599 // Compile the found CGUs in parallel.
600 par_iter(cgus)
601 .map(|(i, _)| {
602 let start_time = Instant::now();
603 let module = backend.compile_codegen_unit(tcx, codegen_units[i].name());
604 let mut time = total_codegen_time.lock();
605 *time += start_time.elapsed();
606 (i, module)
607 })
608 .collect()
609 })
610 } else {
611 FxHashMap::default()
612 }
613 };
614
615 let mut cgu_reuse = Vec::new();
616 let mut pre_compiled_cgus: Option<FxHashMap<usize, _>> = None;
617
618 for (i, cgu) in codegen_units.iter().enumerate() {
619 ongoing_codegen.wait_for_signal_to_codegen_item();
620 ongoing_codegen.check_for_errors(tcx.sess);
621
622 // Do some setup work in the first iteration
623 if pre_compiled_cgus.is_none() {
624 // Calculate the CGU reuse
625 cgu_reuse = tcx.sess.time("find_cgu_reuse", || {
626 codegen_units.iter().map(|cgu| determine_cgu_reuse(tcx, &cgu)).collect()
627 });
628 // Pre compile some CGUs
629 pre_compiled_cgus = Some(pre_compile_cgus(&cgu_reuse));
630 }
631
632 let cgu_reuse = cgu_reuse[i];
633 tcx.sess.cgu_reuse_tracker.set_actual_reuse(&cgu.name().as_str(), cgu_reuse);
634
635 match cgu_reuse {
636 CguReuse::No => {
637 let (module, cost) =
638 if let Some(cgu) = pre_compiled_cgus.as_mut().unwrap().remove(&i) {
639 cgu
640 } else {
641 let start_time = Instant::now();
642 let module = backend.compile_codegen_unit(tcx, cgu.name());
643 let mut time = total_codegen_time.lock();
644 *time += start_time.elapsed();
645 module
646 };
647 submit_codegened_module_to_llvm(
648 &backend,
649 &ongoing_codegen.coordinator_send,
650 module,
651 cost,
652 );
653 false
654 }
655 CguReuse::PreLto => {
656 submit_pre_lto_module_to_llvm(
657 &backend,
658 tcx,
659 &ongoing_codegen.coordinator_send,
660 CachedModuleCodegen {
661 name: cgu.name().to_string(),
662 source: cgu.work_product(tcx),
663 },
664 );
665 true
666 }
667 CguReuse::PostLto => {
668 submit_post_lto_module_to_llvm(
669 &backend,
670 &ongoing_codegen.coordinator_send,
671 CachedModuleCodegen {
672 name: cgu.name().to_string(),
673 source: cgu.work_product(tcx),
674 },
675 );
676 true
677 }
678 };
679 }
680
681 ongoing_codegen.codegen_finished(tcx);
682
683 // Since the main thread is sometimes blocked during codegen, we keep track
684 // -Ztime-passes output manually.
685 print_time_passes_entry(
686 tcx.sess.time_passes(),
687 "codegen_to_LLVM_IR",
688 total_codegen_time.into_inner(),
689 );
690
691 rustc_incremental::assert_module_sources::assert_module_sources(tcx);
692
693 symbol_names_test::report_symbol_names(tcx);
694
695 ongoing_codegen.check_for_errors(tcx.sess);
696
697 finalize_tcx(tcx);
698
699 ongoing_codegen.into_inner()
700 }
701
702 /// A curious wrapper structure whose only purpose is to call `codegen_aborted`
703 /// when it's dropped abnormally.
704 ///
705 /// In the process of working on rust-lang/rust#55238 a mysterious segfault was
706 /// stumbled upon. The segfault was never reproduced locally, but it was
707 /// suspected to be related to the fact that codegen worker threads were
708 /// sticking around by the time the main thread was exiting, causing issues.
709 ///
710 /// This structure is an attempt to fix that issue where the `codegen_aborted`
711 /// message will block until all workers have finished. This should ensure that
712 /// even if the main codegen thread panics we'll wait for pending work to
713 /// complete before returning from the main thread, hopefully avoiding
714 /// segfaults.
715 ///
716 /// If you see this comment in the code, then it means that this workaround
717 /// worked! We may yet one day track down the mysterious cause of that
718 /// segfault...
719 struct AbortCodegenOnDrop<B: ExtraBackendMethods>(Option<OngoingCodegen<B>>);
720
721 impl<B: ExtraBackendMethods> AbortCodegenOnDrop<B> {
722 fn into_inner(mut self) -> OngoingCodegen<B> {
723 self.0.take().unwrap()
724 }
725 }
726
727 impl<B: ExtraBackendMethods> Deref for AbortCodegenOnDrop<B> {
728 type Target = OngoingCodegen<B>;
729
730 fn deref(&self) -> &OngoingCodegen<B> {
731 self.0.as_ref().unwrap()
732 }
733 }
734
735 impl<B: ExtraBackendMethods> DerefMut for AbortCodegenOnDrop<B> {
736 fn deref_mut(&mut self) -> &mut OngoingCodegen<B> {
737 self.0.as_mut().unwrap()
738 }
739 }
740
741 impl<B: ExtraBackendMethods> Drop for AbortCodegenOnDrop<B> {
742 fn drop(&mut self) {
743 if let Some(codegen) = self.0.take() {
744 codegen.codegen_aborted();
745 }
746 }
747 }
748
749 fn finalize_tcx(tcx: TyCtxt<'_>) {
750 tcx.sess.time("assert_dep_graph", || rustc_incremental::assert_dep_graph(tcx));
751 tcx.sess.time("serialize_dep_graph", || rustc_incremental::save_dep_graph(tcx));
752
753 // We assume that no queries are run past here. If there are new queries
754 // after this point, they'll show up as "<unknown>" in self-profiling data.
755 {
756 let _prof_timer = tcx.prof.generic_activity("self_profile_alloc_query_strings");
757 tcx.alloc_self_profile_query_strings();
758 }
759 }
760
761 impl CrateInfo {
762 pub fn new(tcx: TyCtxt<'_>) -> CrateInfo {
763 let mut info = CrateInfo {
764 panic_runtime: None,
765 compiler_builtins: None,
766 profiler_runtime: None,
767 is_no_builtins: Default::default(),
768 native_libraries: Default::default(),
769 used_libraries: tcx.native_libraries(LOCAL_CRATE),
770 link_args: tcx.link_args(LOCAL_CRATE),
771 crate_name: Default::default(),
772 used_crates_dynamic: cstore::used_crates(tcx, LinkagePreference::RequireDynamic),
773 used_crates_static: cstore::used_crates(tcx, LinkagePreference::RequireStatic),
774 used_crate_source: Default::default(),
775 lang_item_to_crate: Default::default(),
776 missing_lang_items: Default::default(),
777 dependency_formats: tcx.dependency_formats(LOCAL_CRATE),
778 };
779 let lang_items = tcx.lang_items();
780
781 let crates = tcx.crates();
782
783 let n_crates = crates.len();
784 info.native_libraries.reserve(n_crates);
785 info.crate_name.reserve(n_crates);
786 info.used_crate_source.reserve(n_crates);
787 info.missing_lang_items.reserve(n_crates);
788
789 for &cnum in crates.iter() {
790 info.native_libraries.insert(cnum, tcx.native_libraries(cnum));
791 info.crate_name.insert(cnum, tcx.crate_name(cnum).to_string());
792 info.used_crate_source.insert(cnum, tcx.used_crate_source(cnum));
793 if tcx.is_panic_runtime(cnum) {
794 info.panic_runtime = Some(cnum);
795 }
796 if tcx.is_compiler_builtins(cnum) {
797 info.compiler_builtins = Some(cnum);
798 }
799 if tcx.is_profiler_runtime(cnum) {
800 info.profiler_runtime = Some(cnum);
801 }
802 if tcx.is_no_builtins(cnum) {
803 info.is_no_builtins.insert(cnum);
804 }
805 let missing = tcx.missing_lang_items(cnum);
806 for &item in missing.iter() {
807 if let Ok(id) = lang_items.require(item) {
808 info.lang_item_to_crate.insert(item, id.krate);
809 }
810 }
811
812 // No need to look for lang items that don't actually need to exist.
813 let missing =
814 missing.iter().cloned().filter(|&l| lang_items::required(tcx, l)).collect();
815 info.missing_lang_items.insert(cnum, missing);
816 }
817
818 info
819 }
820 }
821
822 pub fn provide_both(providers: &mut Providers) {
823 providers.backend_optimization_level = |tcx, cratenum| {
824 let for_speed = match tcx.sess.opts.optimize {
825 // If globally no optimisation is done, #[optimize] has no effect.
826 //
827 // This is done because if we ended up "upgrading" to `-O2` here, we’d populate the
828 // pass manager and it is likely that some module-wide passes (such as inliner or
829 // cross-function constant propagation) would ignore the `optnone` annotation we put
830 // on the functions, thus necessarily involving these functions into optimisations.
831 config::OptLevel::No => return config::OptLevel::No,
832 // If globally optimise-speed is already specified, just use that level.
833 config::OptLevel::Less => return config::OptLevel::Less,
834 config::OptLevel::Default => return config::OptLevel::Default,
835 config::OptLevel::Aggressive => return config::OptLevel::Aggressive,
836 // If globally optimize-for-size has been requested, use -O2 instead (if optimize(size)
837 // are present).
838 config::OptLevel::Size => config::OptLevel::Default,
839 config::OptLevel::SizeMin => config::OptLevel::Default,
840 };
841
842 let (defids, _) = tcx.collect_and_partition_mono_items(cratenum);
843 for id in &*defids {
844 let CodegenFnAttrs { optimize, .. } = tcx.codegen_fn_attrs(*id);
845 match optimize {
846 attr::OptimizeAttr::None => continue,
847 attr::OptimizeAttr::Size => continue,
848 attr::OptimizeAttr::Speed => {
849 return for_speed;
850 }
851 }
852 }
853 tcx.sess.opts.optimize
854 };
855
856 providers.dllimport_foreign_items = |tcx, krate| {
857 let module_map = tcx.foreign_modules(krate);
858
859 let dllimports = tcx
860 .native_libraries(krate)
861 .iter()
862 .filter(|lib| {
863 if !matches!(lib.kind, NativeLibKind::Dylib | NativeLibKind::Unspecified) {
864 return false;
865 }
866 let cfg = match lib.cfg {
867 Some(ref cfg) => cfg,
868 None => return true,
869 };
870 attr::cfg_matches(cfg, &tcx.sess.parse_sess, None)
871 })
872 .filter_map(|lib| lib.foreign_module)
873 .map(|id| &module_map[&id])
874 .flat_map(|module| module.foreign_items.iter().cloned())
875 .collect();
876 dllimports
877 };
878
879 providers.is_dllimport_foreign_item =
880 |tcx, def_id| tcx.dllimport_foreign_items(def_id.krate).contains(&def_id);
881 }
882
883 fn determine_cgu_reuse<'tcx>(tcx: TyCtxt<'tcx>, cgu: &CodegenUnit<'tcx>) -> CguReuse {
884 if !tcx.dep_graph.is_fully_enabled() {
885 return CguReuse::No;
886 }
887
888 let work_product_id = &cgu.work_product_id();
889 if tcx.dep_graph.previous_work_product(work_product_id).is_none() {
890 // We don't have anything cached for this CGU. This can happen
891 // if the CGU did not exist in the previous session.
892 return CguReuse::No;
893 }
894
895 // Try to mark the CGU as green. If it we can do so, it means that nothing
896 // affecting the LLVM module has changed and we can re-use a cached version.
897 // If we compile with any kind of LTO, this means we can re-use the bitcode
898 // of the Pre-LTO stage (possibly also the Post-LTO version but we'll only
899 // know that later). If we are not doing LTO, there is only one optimized
900 // version of each module, so we re-use that.
901 let dep_node = cgu.codegen_dep_node(tcx);
902 assert!(
903 !tcx.dep_graph.dep_node_exists(&dep_node),
904 "CompileCodegenUnit dep-node for CGU `{}` already exists before marking.",
905 cgu.name()
906 );
907
908 if tcx.dep_graph.try_mark_green(tcx, &dep_node).is_some() {
909 // We can re-use either the pre- or the post-thinlto state. If no LTO is
910 // being performed then we can use post-LTO artifacts, otherwise we must
911 // reuse pre-LTO artifacts
912 match compute_per_cgu_lto_type(
913 &tcx.sess.lto(),
914 &tcx.sess.opts,
915 &tcx.sess.crate_types(),
916 ModuleKind::Regular,
917 ) {
918 ComputedLtoType::No => CguReuse::PostLto,
919 _ => CguReuse::PreLto,
920 }
921 } else {
922 CguReuse::No
923 }
924 }