compiler/rustc_codegen_ssa/src/base.rs

   1 use crate::back::metadata::create_compressed_metadata_file;
   2 use crate::back::write::{
   3     compute_per_cgu_lto_type, start_async_codegen, submit_codegened_module_to_llvm,
   4     submit_post_lto_module_to_llvm, submit_pre_lto_module_to_llvm, ComputedLtoType, OngoingCodegen,
   5 };
   6 use crate::common::{IntPredicate, RealPredicate, TypeKind};
   7 use crate::meth;
   8 use crate::mir;
   9 use crate::mir::operand::OperandValue;
  10 use crate::mir::place::PlaceRef;
  11 use crate::traits::*;
  12 use crate::{CachedModuleCodegen, CompiledModule, CrateInfo, MemFlags, ModuleCodegen, ModuleKind};
  13
  14 use rustc_attr as attr;
  15 use rustc_data_structures::fx::FxHashMap;
  16 use rustc_data_structures::profiling::{get_resident_set_size, print_time_passes_entry};
  17 use rustc_data_structures::sync::{par_iter, ParallelIterator};
  18 use rustc_hir as hir;
  19 use rustc_hir::def_id::{DefId, LOCAL_CRATE};
  20 use rustc_hir::lang_items::LangItem;
  21 use rustc_index::vec::Idx;
  22 use rustc_metadata::EncodedMetadata;
  23 use rustc_middle::middle::codegen_fn_attrs::CodegenFnAttrs;
  24 use rustc_middle::middle::exported_symbols;
  25 use rustc_middle::middle::lang_items;
  26 use rustc_middle::mir::mono::{CodegenUnit, CodegenUnitNameBuilder, MonoItem};
  27 use rustc_middle::ty::layout::{HasTyCtxt, LayoutOf, TyAndLayout};
  28 use rustc_middle::ty::query::Providers;
  29 use rustc_middle::ty::{self, Instance, Ty, TyCtxt};
  30 use rustc_session::cgu_reuse_tracker::CguReuse;
  31 use rustc_session::config::{self, EntryFnType, OutputType};
  32 use rustc_session::Session;
  33 use rustc_span::symbol::sym;
  34 use rustc_target::abi::{Align, VariantIdx};
  35
  36 use std::convert::TryFrom;
  37 use std::ops::{Deref, DerefMut};
  38 use std::time::{Duration, Instant};
  39
  40 use itertools::Itertools;
  41
  42 pub fn bin_op_to_icmp_predicate(op: hir::BinOpKind, signed: bool) -> IntPredicate {
  43     match op {
  44         hir::BinOpKind::Eq => IntPredicate::IntEQ,
  45         hir::BinOpKind::Ne => IntPredicate::IntNE,
  46         hir::BinOpKind::Lt => {
  47             if signed {
  48                 IntPredicate::IntSLT
  49             } else {
  50                 IntPredicate::IntULT
  51             }
  52         }
  53         hir::BinOpKind::Le => {
  54             if signed {
  55                 IntPredicate::IntSLE
  56             } else {
  57                 IntPredicate::IntULE
  58             }
  59         }
  60         hir::BinOpKind::Gt => {
  61             if signed {
  62                 IntPredicate::IntSGT
  63             } else {
  64                 IntPredicate::IntUGT
  65             }
  66         }
  67         hir::BinOpKind::Ge => {
  68             if signed {
  69                 IntPredicate::IntSGE
  70             } else {
  71                 IntPredicate::IntUGE
  72             }
  73         }
  74         op => bug!(
  75             "comparison_op_to_icmp_predicate: expected comparison operator, \
  76              found {:?}",
  77             op
  78         ),
  79     }
  80 }
  81
  82 pub fn bin_op_to_fcmp_predicate(op: hir::BinOpKind) -> RealPredicate {
  83     match op {
  84         hir::BinOpKind::Eq => RealPredicate::RealOEQ,
  85         hir::BinOpKind::Ne => RealPredicate::RealUNE,
  86         hir::BinOpKind::Lt => RealPredicate::RealOLT,
  87         hir::BinOpKind::Le => RealPredicate::RealOLE,
  88         hir::BinOpKind::Gt => RealPredicate::RealOGT,
  89         hir::BinOpKind::Ge => RealPredicate::RealOGE,
  90         op => {
  91             bug!(
  92                 "comparison_op_to_fcmp_predicate: expected comparison operator, \
  93                  found {:?}",
  94                 op
  95             );
  96         }
  97     }
  98 }
  99
 100 pub fn compare_simd_types<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
 101     bx: &mut Bx,
 102     lhs: Bx::Value,
 103     rhs: Bx::Value,
 104     t: Ty<'tcx>,
 105     ret_ty: Bx::Type,
 106     op: hir::BinOpKind,
 107 ) -> Bx::Value {
 108     let signed = match t.kind() {
 109         ty::Float(_) => {
 110             let cmp = bin_op_to_fcmp_predicate(op);
 111             let cmp = bx.fcmp(cmp, lhs, rhs);
 112             return bx.sext(cmp, ret_ty);
 113         }
 114         ty::Uint(_) => false,
 115         ty::Int(_) => true,
 116         _ => bug!("compare_simd_types: invalid SIMD type"),
 117     };
 118
 119     let cmp = bin_op_to_icmp_predicate(op, signed);
 120     let cmp = bx.icmp(cmp, lhs, rhs);
 121     // LLVM outputs an `< size x i1 >`, so we need to perform a sign extension
 122     // to get the correctly sized type. This will compile to a single instruction
 123     // once the IR is converted to assembly if the SIMD instruction is supported
 124     // by the target architecture.
 125     bx.sext(cmp, ret_ty)
 126 }
 127
 128 /// Retrieves the information we are losing (making dynamic) in an unsizing
 129 /// adjustment.
 130 ///
 131 /// The `old_info` argument is a bit odd. It is intended for use in an upcast,
 132 /// where the new vtable for an object will be derived from the old one.
 133 pub fn unsized_info<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
 134     bx: &mut Bx,
 135     source: Ty<'tcx>,
 136     target: Ty<'tcx>,
 137     old_info: Option<Bx::Value>,
 138 ) -> Bx::Value {
 139     let cx = bx.cx();
 140     let (source, target) =
 141         cx.tcx().struct_lockstep_tails_erasing_lifetimes(source, target, bx.param_env());
 142     match (source.kind(), target.kind()) {
 143         (&ty::Array(_, len), &ty::Slice(_)) => {
 144             cx.const_usize(len.eval_usize(cx.tcx(), ty::ParamEnv::reveal_all()))
 145         }
 146         (&ty::Dynamic(ref data_a, ..), &ty::Dynamic(ref data_b, ..)) => {
 147             let old_info =
 148                 old_info.expect("unsized_info: missing old info for trait upcasting coercion");
 149             if data_a.principal_def_id() == data_b.principal_def_id() {
 150                 return old_info;
 151             }
 152
 153             // trait upcasting coercion
 154
 155             let vptr_entry_idx =
 156                 cx.tcx().vtable_trait_upcasting_coercion_new_vptr_slot((source, target));
 157
 158             if let Some(entry_idx) = vptr_entry_idx {
 159                 let ptr_ty = cx.type_i8p();
 160                 let ptr_align = cx.tcx().data_layout.pointer_align.abi;
 161                 let llvtable = bx.pointercast(old_info, bx.type_ptr_to(ptr_ty));
 162                 let gep = bx.inbounds_gep(
 163                     ptr_ty,
 164                     llvtable,
 165                     &[bx.const_usize(u64::try_from(entry_idx).unwrap())],
 166                 );
 167                 let new_vptr = bx.load(ptr_ty, gep, ptr_align);
 168                 bx.nonnull_metadata(new_vptr);
 169                 // Vtable loads are invariant.
 170                 bx.set_invariant_load(new_vptr);
 171                 new_vptr
 172             } else {
 173                 old_info
 174             }
 175         }
 176         (_, &ty::Dynamic(ref data, ..)) => {
 177             let vtable_ptr_ty = cx.scalar_pair_element_backend_type(
 178                 cx.layout_of(cx.tcx().mk_mut_ptr(target)),
 179                 1,
 180                 true,
 181             );
 182             cx.const_ptrcast(meth::get_vtable(cx, source, data.principal()), vtable_ptr_ty)
 183         }
 184         _ => bug!("unsized_info: invalid unsizing {:?} -> {:?}", source, target),
 185     }
 186 }
 187
 188 /// Coerces `src` to `dst_ty`. `src_ty` must be a pointer.
 189 pub fn unsize_ptr<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
 190     bx: &mut Bx,
 191     src: Bx::Value,
 192     src_ty: Ty<'tcx>,
 193     dst_ty: Ty<'tcx>,
 194     old_info: Option<Bx::Value>,
 195 ) -> (Bx::Value, Bx::Value) {
 196     debug!("unsize_ptr: {:?} => {:?}", src_ty, dst_ty);
 197     match (src_ty.kind(), dst_ty.kind()) {
 198         (&ty::Ref(_, a, _), &ty::Ref(_, b, _) | &ty::RawPtr(ty::TypeAndMut { ty: b, .. }))
 199         | (&ty::RawPtr(ty::TypeAndMut { ty: a, .. }), &ty::RawPtr(ty::TypeAndMut { ty: b, .. })) => {
 200             assert_eq!(bx.cx().type_is_sized(a), old_info.is_none());
 201             let ptr_ty = bx.cx().type_ptr_to(bx.cx().backend_type(bx.cx().layout_of(b)));
 202             (bx.pointercast(src, ptr_ty), unsized_info(bx, a, b, old_info))
 203         }
 204         (&ty::Adt(def_a, _), &ty::Adt(def_b, _)) => {
 205             assert_eq!(def_a, def_b);
 206             let src_layout = bx.cx().layout_of(src_ty);
 207             let dst_layout = bx.cx().layout_of(dst_ty);
 208             if src_ty == dst_ty {
 209                 return (src, old_info.unwrap());
 210             }
 211             let mut result = None;
 212             for i in 0..src_layout.fields.count() {
 213                 let src_f = src_layout.field(bx.cx(), i);
 214                 assert_eq!(src_layout.fields.offset(i).bytes(), 0);
 215                 assert_eq!(dst_layout.fields.offset(i).bytes(), 0);
 216                 if src_f.is_zst() {
 217                     continue;
 218                 }
 219                 assert_eq!(src_layout.size, src_f.size);
 220
 221                 let dst_f = dst_layout.field(bx.cx(), i);
 222                 assert_ne!(src_f.ty, dst_f.ty);
 223                 assert_eq!(result, None);
 224                 result = Some(unsize_ptr(bx, src, src_f.ty, dst_f.ty, old_info));
 225             }
 226             let (lldata, llextra) = result.unwrap();
 227             let lldata_ty = bx.cx().scalar_pair_element_backend_type(dst_layout, 0, true);
 228             let llextra_ty = bx.cx().scalar_pair_element_backend_type(dst_layout, 1, true);
 229             // HACK(eddyb) have to bitcast pointers until LLVM removes pointee types.
 230             (bx.bitcast(lldata, lldata_ty), bx.bitcast(llextra, llextra_ty))
 231         }
 232         _ => bug!("unsize_ptr: called on bad types"),
 233     }
 234 }
 235
 236 /// Coerces `src`, which is a reference to a value of type `src_ty`,
 237 /// to a value of type `dst_ty`, and stores the result in `dst`.
 238 pub fn coerce_unsized_into<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
 239     bx: &mut Bx,
 240     src: PlaceRef<'tcx, Bx::Value>,
 241     dst: PlaceRef<'tcx, Bx::Value>,
 242 ) {
 243     let src_ty = src.layout.ty;
 244     let dst_ty = dst.layout.ty;
 245     match (src_ty.kind(), dst_ty.kind()) {
 246         (&ty::Ref(..), &ty::Ref(..) | &ty::RawPtr(..)) | (&ty::RawPtr(..), &ty::RawPtr(..)) => {
 247             let (base, info) = match bx.load_operand(src).val {
 248                 OperandValue::Pair(base, info) => unsize_ptr(bx, base, src_ty, dst_ty, Some(info)),
 249                 OperandValue::Immediate(base) => unsize_ptr(bx, base, src_ty, dst_ty, None),
 250                 OperandValue::Ref(..) => bug!(),
 251             };
 252             OperandValue::Pair(base, info).store(bx, dst);
 253         }
 254
 255         (&ty::Adt(def_a, _), &ty::Adt(def_b, _)) => {
 256             assert_eq!(def_a, def_b);
 257
 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);
 261
 262                 if dst_f.layout.is_zst() {
 263                     continue;
 264                 }
 265
 266                 if src_f.layout.ty == dst_f.layout.ty {
 267                     memcpy_ty(
 268                         bx,
 269                         dst_f.llval,
 270                         dst_f.align,
 271                         src_f.llval,
 272                         src_f.align,
 273                         src_f.layout,
 274                         MemFlags::empty(),
 275                     );
 276                 } else {
 277                     coerce_unsized_into(bx, src_f, dst_f);
 278                 }
 279             }
 280         }
 281         _ => bug!("coerce_unsized_into: invalid coercion {:?} -> {:?}", src_ty, dst_ty,),
 282     }
 283 }
 284
 285 pub fn cast_shift_expr_rhs<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
 286     bx: &mut Bx,
 287     op: hir::BinOpKind,
 288     lhs: Bx::Value,
 289     rhs: Bx::Value,
 290 ) -> Bx::Value {
 291     cast_shift_rhs(bx, op, lhs, rhs)
 292 }
 293
 294 fn cast_shift_rhs<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
 295     bx: &mut Bx,
 296     op: hir::BinOpKind,
 297     lhs: Bx::Value,
 298     rhs: Bx::Value,
 299 ) -> Bx::Value {
 300     // Shifts may have any size int on the rhs
 301     if op.is_shift() {
 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)
 306         }
 307         if bx.cx().type_kind(lhs_llty) == TypeKind::Vector {
 308             lhs_llty = bx.cx().element_type(lhs_llty)
 309         }
 310         let rhs_sz = bx.cx().int_width(rhs_llty);
 311         let lhs_sz = bx.cx().int_width(lhs_llty);
 312         if lhs_sz < rhs_sz {
 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)
 318         } else {
 319             rhs
 320         }
 321     } else {
 322         rhs
 323     }
 324 }
 325
 326 /// Returns `true` if this session's target will use SEH-based unwinding.
 327 ///
 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.is_like_msvc
 333 }
 334
 335 pub fn memcpy_ty<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
 336     bx: &mut Bx,
 337     dst: Bx::Value,
 338     dst_align: Align,
 339     src: Bx::Value,
 340     src_align: Align,
 341     layout: TyAndLayout<'tcx>,
 342     flags: MemFlags,
 343 ) {
 344     let size = layout.size.bytes();
 345     if size == 0 {
 346         return;
 347     }
 348
 349     bx.memcpy(dst, dst_align, src, src_align, bx.cx().const_usize(size), flags);
 350 }
 351
 352 pub fn codegen_instance<'a, 'tcx: 'a, Bx: BuilderMethods<'a, 'tcx>>(
 353     cx: &'a Bx::CodegenCx,
 354     instance: Instance<'tcx>,
 355 ) {
 356     // this is an info! to allow collecting monomorphization statistics
 357     // and to allow finding the last function before LLVM aborts from
 358     // release builds.
 359     info!("codegen_instance({})", instance);
 360
 361     mir::codegen_mir::<Bx>(cx, instance);
 362 }
 363
 364 /// Creates the `main` function which will initialize the rust runtime and call
 365 /// users main function.
 366 pub fn maybe_create_entry_wrapper<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
 367     cx: &'a Bx::CodegenCx,
 368 ) -> Option<Bx::Function> {
 369     let (main_def_id, entry_type) = cx.tcx().entry_fn(())?;
 370     let main_is_local = main_def_id.is_local();
 371     let instance = Instance::mono(cx.tcx(), main_def_id);
 372
 373     if main_is_local {
 374         // We want to create the wrapper in the same codegen unit as Rust's main
 375         // function.
 376         if !cx.codegen_unit().contains_item(&MonoItem::Fn(instance)) {
 377             return None;
 378         }
 379     } else if !cx.codegen_unit().is_primary() {
 380         // We want to create the wrapper only when the codegen unit is the primary one
 381         return None;
 382     }
 383
 384     let main_llfn = cx.get_fn_addr(instance);
 385
 386     let use_start_lang_item = EntryFnType::Start != entry_type;
 387     let entry_fn = create_entry_fn::<Bx>(cx, main_llfn, main_def_id, use_start_lang_item);
 388     return Some(entry_fn);
 389
 390     fn create_entry_fn<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
 391         cx: &'a Bx::CodegenCx,
 392         rust_main: Bx::Value,
 393         rust_main_def_id: DefId,
 394         use_start_lang_item: bool,
 395     ) -> Bx::Function {
 396         // The entry function is either `int main(void)` or `int main(int argc, char **argv)`,
 397         // depending on whether the target needs `argc` and `argv` to be passed in.
 398         let llfty = if cx.sess().target.main_needs_argc_argv {
 399             cx.type_func(&[cx.type_int(), cx.type_ptr_to(cx.type_i8p())], cx.type_int())
 400         } else {
 401             cx.type_func(&[], cx.type_int())
 402         };
 403
 404         let main_ret_ty = cx.tcx().fn_sig(rust_main_def_id).output();
 405         // Given that `main()` has no arguments,
 406         // then its return type cannot have
 407         // late-bound regions, since late-bound
 408         // regions must appear in the argument
 409         // listing.
 410         let main_ret_ty = cx.tcx().erase_regions(main_ret_ty.no_bound_vars().unwrap());
 411
 412         let llfn = match cx.declare_c_main(llfty) {
 413             Some(llfn) => llfn,
 414             None => {
 415                 // FIXME: We should be smart and show a better diagnostic here.
 416                 let span = cx.tcx().def_span(rust_main_def_id);
 417                 cx.sess()
 418                     .struct_span_err(span, "entry symbol `main` declared multiple times")
 419                     .help("did you use `#[no_mangle]` on `fn main`? Use `#[start]` instead")
 420                     .emit();
 421                 cx.sess().abort_if_errors();
 422                 bug!();
 423             }
 424         };
 425
 426         // `main` should respect same config for frame pointer elimination as rest of code
 427         cx.set_frame_pointer_type(llfn);
 428         cx.apply_target_cpu_attr(llfn);
 429
 430         let llbb = Bx::append_block(&cx, llfn, "top");
 431         let mut bx = Bx::build(&cx, llbb);
 432
 433         bx.insert_reference_to_gdb_debug_scripts_section_global();
 434
 435         let isize_ty = cx.type_isize();
 436         let i8pp_ty = cx.type_ptr_to(cx.type_i8p());
 437         let (arg_argc, arg_argv) = get_argc_argv(cx, &mut bx);
 438
 439         let (start_fn, start_ty, args) = if use_start_lang_item {
 440             let start_def_id = cx.tcx().require_lang_item(LangItem::Start, None);
 441             let start_fn = cx.get_fn_addr(
 442                 ty::Instance::resolve(
 443                     cx.tcx(),
 444                     ty::ParamEnv::reveal_all(),
 445                     start_def_id,
 446                     cx.tcx().intern_substs(&[main_ret_ty.into()]),
 447                 )
 448                 .unwrap()
 449                 .unwrap(),
 450             );
 451             let start_ty = cx.type_func(&[cx.val_ty(rust_main), isize_ty, i8pp_ty], isize_ty);
 452             (start_fn, start_ty, vec![rust_main, arg_argc, arg_argv])
 453         } else {
 454             debug!("using user-defined start fn");
 455             let start_ty = cx.type_func(&[isize_ty, i8pp_ty], isize_ty);
 456             (rust_main, start_ty, vec![arg_argc, arg_argv])
 457         };
 458
 459         let result = bx.call(start_ty, start_fn, &args, None);
 460         let cast = bx.intcast(result, cx.type_int(), true);
 461         bx.ret(cast);
 462
 463         llfn
 464     }
 465 }
 466
 467 /// Obtain the `argc` and `argv` values to pass to the rust start function.
 468 fn get_argc_argv<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
 469     cx: &'a Bx::CodegenCx,
 470     bx: &mut Bx,
 471 ) -> (Bx::Value, Bx::Value) {
 472     if cx.sess().target.main_needs_argc_argv {
 473         // Params from native `main()` used as args for rust start function
 474         let param_argc = bx.get_param(0);
 475         let param_argv = bx.get_param(1);
 476         let arg_argc = bx.intcast(param_argc, cx.type_isize(), true);
 477         let arg_argv = param_argv;
 478         (arg_argc, arg_argv)
 479     } else {
 480         // The Rust start function doesn't need `argc` and `argv`, so just pass zeros.
 481         let arg_argc = bx.const_int(cx.type_int(), 0);
 482         let arg_argv = bx.const_null(cx.type_ptr_to(cx.type_i8p()));
 483         (arg_argc, arg_argv)
 484     }
 485 }
 486
 487 pub fn codegen_crate<B: ExtraBackendMethods>(
 488     backend: B,
 489     tcx: TyCtxt<'_>,
 490     target_cpu: String,
 491     metadata: EncodedMetadata,
 492     need_metadata_module: bool,
 493 ) -> OngoingCodegen<B> {
 494     // Skip crate items and just output metadata in -Z no-codegen mode.
 495     if tcx.sess.opts.debugging_opts.no_codegen || !tcx.sess.opts.output_types.should_codegen() {
 496         let ongoing_codegen = start_async_codegen(backend, tcx, target_cpu, metadata, None, 1);
 497
 498         ongoing_codegen.codegen_finished(tcx);
 499
 500         ongoing_codegen.check_for_errors(tcx.sess);
 501
 502         return ongoing_codegen;
 503     }
 504
 505     let cgu_name_builder = &mut CodegenUnitNameBuilder::new(tcx);
 506
 507     // Run the monomorphization collector and partition the collected items into
 508     // codegen units.
 509     let codegen_units = tcx.collect_and_partition_mono_items(()).1;
 510
 511     // Force all codegen_unit queries so they are already either red or green
 512     // when compile_codegen_unit accesses them. We are not able to re-execute
 513     // the codegen_unit query from just the DepNode, so an unknown color would
 514     // lead to having to re-execute compile_codegen_unit, possibly
 515     // unnecessarily.
 516     if tcx.dep_graph.is_fully_enabled() {
 517         for cgu in codegen_units {
 518             tcx.ensure().codegen_unit(cgu.name());
 519         }
 520     }
 521
 522     let metadata_module = if need_metadata_module {
 523         // Emit compressed metadata object.
 524         let metadata_cgu_name =
 525             cgu_name_builder.build_cgu_name(LOCAL_CRATE, &["crate"], Some("metadata")).to_string();
 526         tcx.sess.time("write_compressed_metadata", || {
 527             let file_name =
 528                 tcx.output_filenames(()).temp_path(OutputType::Metadata, Some(&metadata_cgu_name));
 529             let data = create_compressed_metadata_file(
 530                 tcx.sess,
 531                 &metadata,
 532                 &exported_symbols::metadata_symbol_name(tcx),
 533             );
 534             if let Err(err) = std::fs::write(&file_name, data) {
 535                 tcx.sess.fatal(&format!("error writing metadata object file: {}", err));
 536             }
 537             Some(CompiledModule {
 538                 name: metadata_cgu_name,
 539                 kind: ModuleKind::Metadata,
 540                 object: Some(file_name),
 541                 dwarf_object: None,
 542                 bytecode: None,
 543             })
 544         })
 545     } else {
 546         None
 547     };
 548
 549     let ongoing_codegen = start_async_codegen(
 550         backend.clone(),
 551         tcx,
 552         target_cpu,
 553         metadata,
 554         metadata_module,
 555         codegen_units.len(),
 556     );
 557     let ongoing_codegen = AbortCodegenOnDrop::<B>(Some(ongoing_codegen));
 558
 559     // Codegen an allocator shim, if necessary.
 560     //
 561     // If the crate doesn't have an `allocator_kind` set then there's definitely
 562     // no shim to generate. Otherwise we also check our dependency graph for all
 563     // our output crate types. If anything there looks like its a `Dynamic`
 564     // linkage, then it's already got an allocator shim and we'll be using that
 565     // one instead. If nothing exists then it's our job to generate the
 566     // allocator!
 567     let any_dynamic_crate = tcx.dependency_formats(()).iter().any(|(_, list)| {
 568         use rustc_middle::middle::dependency_format::Linkage;
 569         list.iter().any(|&linkage| linkage == Linkage::Dynamic)
 570     });
 571     let allocator_module = if any_dynamic_crate {
 572         None
 573     } else if let Some(kind) = tcx.allocator_kind(()) {
 574         let llmod_id =
 575             cgu_name_builder.build_cgu_name(LOCAL_CRATE, &["crate"], Some("allocator")).to_string();
 576         let mut module_llvm = backend.new_metadata(tcx, &llmod_id);
 577         tcx.sess.time("write_allocator_module", || {
 578             backend.codegen_allocator(
 579                 tcx,
 580                 &mut module_llvm,
 581                 &llmod_id,
 582                 kind,
 583                 tcx.lang_items().oom().is_some(),
 584             )
 585         });
 586
 587         Some(ModuleCodegen { name: llmod_id, module_llvm, kind: ModuleKind::Allocator })
 588     } else {
 589         None
 590     };
 591
 592     if let Some(allocator_module) = allocator_module {
 593         ongoing_codegen.submit_pre_codegened_module_to_llvm(tcx, allocator_module);
 594     }
 595
 596     // For better throughput during parallel processing by LLVM, we used to sort
 597     // CGUs largest to smallest. This would lead to better thread utilization
 598     // by, for example, preventing a large CGU from being processed last and
 599     // having only one LLVM thread working while the rest remained idle.
 600     //
 601     // However, this strategy would lead to high memory usage, as it meant the
 602     // LLVM-IR for all of the largest CGUs would be resident in memory at once.
 603     //
 604     // Instead, we can compromise by ordering CGUs such that the largest and
 605     // smallest are first, second largest and smallest are next, etc. If there
 606     // are large size variations, this can reduce memory usage significantly.
 607     let codegen_units: Vec<_> = {
 608         let mut sorted_cgus = codegen_units.iter().collect::<Vec<_>>();
 609         sorted_cgus.sort_by_cached_key(|cgu| cgu.size_estimate());
 610
 611         let (first_half, second_half) = sorted_cgus.split_at(sorted_cgus.len() / 2);
 612         second_half.iter().rev().interleave(first_half).copied().collect()
 613     };
 614
 615     // The non-parallel compiler can only translate codegen units to LLVM IR
 616     // on a single thread, leading to a staircase effect where the N LLVM
 617     // threads have to wait on the single codegen threads to generate work
 618     // for them. The parallel compiler does not have this restriction, so
 619     // we can pre-load the LLVM queue in parallel before handing off
 620     // coordination to the OnGoingCodegen scheduler.
 621     //
 622     // This likely is a temporary measure. Once we don't have to support the
 623     // non-parallel compiler anymore, we can compile CGUs end-to-end in
 624     // parallel and get rid of the complicated scheduling logic.
 625     let pre_compile_cgus = |cgu_reuse: &[CguReuse]| {
 626         if cfg!(parallel_compiler) {
 627             tcx.sess.time("compile_first_CGU_batch", || {
 628                 // Try to find one CGU to compile per thread.
 629                 let cgus: Vec<_> = cgu_reuse
 630                     .iter()
 631                     .enumerate()
 632                     .filter(|&(_, reuse)| reuse == &CguReuse::No)
 633                     .take(tcx.sess.threads())
 634                     .collect();
 635
 636                 // Compile the found CGUs in parallel.
 637                 let start_time = Instant::now();
 638
 639                 let pre_compiled_cgus = par_iter(cgus)
 640                     .map(|(i, _)| {
 641                         let module = backend.compile_codegen_unit(tcx, codegen_units[i].name());
 642                         (i, module)
 643                     })
 644                     .collect();
 645
 646                 (pre_compiled_cgus, start_time.elapsed())
 647             })
 648         } else {
 649             (FxHashMap::default(), Duration::new(0, 0))
 650         }
 651     };
 652
 653     let mut cgu_reuse = Vec::new();
 654     let mut pre_compiled_cgus: Option<FxHashMap<usize, _>> = None;
 655     let mut total_codegen_time = Duration::new(0, 0);
 656     let start_rss = tcx.sess.time_passes().then(|| get_resident_set_size());
 657
 658     for (i, cgu) in codegen_units.iter().enumerate() {
 659         ongoing_codegen.wait_for_signal_to_codegen_item();
 660         ongoing_codegen.check_for_errors(tcx.sess);
 661
 662         // Do some setup work in the first iteration
 663         if pre_compiled_cgus.is_none() {
 664             // Calculate the CGU reuse
 665             cgu_reuse = tcx.sess.time("find_cgu_reuse", || {
 666                 codegen_units.iter().map(|cgu| determine_cgu_reuse(tcx, &cgu)).collect()
 667             });
 668             // Pre compile some CGUs
 669             let (compiled_cgus, codegen_time) = pre_compile_cgus(&cgu_reuse);
 670             pre_compiled_cgus = Some(compiled_cgus);
 671             total_codegen_time += codegen_time;
 672         }
 673
 674         let cgu_reuse = cgu_reuse[i];
 675         tcx.sess.cgu_reuse_tracker.set_actual_reuse(cgu.name().as_str(), cgu_reuse);
 676
 677         match cgu_reuse {
 678             CguReuse::No => {
 679                 let (module, cost) =
 680                     if let Some(cgu) = pre_compiled_cgus.as_mut().unwrap().remove(&i) {
 681                         cgu
 682                     } else {
 683                         let start_time = Instant::now();
 684                         let module = backend.compile_codegen_unit(tcx, cgu.name());
 685                         total_codegen_time += start_time.elapsed();
 686                         module
 687                     };
 688                 // This will unwind if there are errors, which triggers our `AbortCodegenOnDrop`
 689                 // guard. Unfortunately, just skipping the `submit_codegened_module_to_llvm` makes
 690                 // compilation hang on post-monomorphization errors.
 691                 tcx.sess.abort_if_errors();
 692
 693                 submit_codegened_module_to_llvm(
 694                     &backend,
 695                     &ongoing_codegen.coordinator_send,
 696                     module,
 697                     cost,
 698                 );
 699                 false
 700             }
 701             CguReuse::PreLto => {
 702                 submit_pre_lto_module_to_llvm(
 703                     &backend,
 704                     tcx,
 705                     &ongoing_codegen.coordinator_send,
 706                     CachedModuleCodegen {
 707                         name: cgu.name().to_string(),
 708                         source: cgu.work_product(tcx),
 709                     },
 710                 );
 711                 true
 712             }
 713             CguReuse::PostLto => {
 714                 submit_post_lto_module_to_llvm(
 715                     &backend,
 716                     &ongoing_codegen.coordinator_send,
 717                     CachedModuleCodegen {
 718                         name: cgu.name().to_string(),
 719                         source: cgu.work_product(tcx),
 720                     },
 721                 );
 722                 true
 723             }
 724         };
 725     }
 726
 727     ongoing_codegen.codegen_finished(tcx);
 728
 729     // Since the main thread is sometimes blocked during codegen, we keep track
 730     // -Ztime-passes output manually.
 731     if tcx.sess.time_passes() {
 732         let end_rss = get_resident_set_size();
 733
 734         print_time_passes_entry(
 735             "codegen_to_LLVM_IR",
 736             total_codegen_time,
 737             start_rss.unwrap(),
 738             end_rss,
 739         );
 740     }
 741
 742     ongoing_codegen.check_for_errors(tcx.sess);
 743
 744     ongoing_codegen.into_inner()
 745 }
 746
 747 /// A curious wrapper structure whose only purpose is to call `codegen_aborted`
 748 /// when it's dropped abnormally.
 749 ///
 750 /// In the process of working on rust-lang/rust#55238 a mysterious segfault was
 751 /// stumbled upon. The segfault was never reproduced locally, but it was
 752 /// suspected to be related to the fact that codegen worker threads were
 753 /// sticking around by the time the main thread was exiting, causing issues.
 754 ///
 755 /// This structure is an attempt to fix that issue where the `codegen_aborted`
 756 /// message will block until all workers have finished. This should ensure that
 757 /// even if the main codegen thread panics we'll wait for pending work to
 758 /// complete before returning from the main thread, hopefully avoiding
 759 /// segfaults.
 760 ///
 761 /// If you see this comment in the code, then it means that this workaround
 762 /// worked! We may yet one day track down the mysterious cause of that
 763 /// segfault...
 764 struct AbortCodegenOnDrop<B: ExtraBackendMethods>(Option<OngoingCodegen<B>>);
 765
 766 impl<B: ExtraBackendMethods> AbortCodegenOnDrop<B> {
 767     fn into_inner(mut self) -> OngoingCodegen<B> {
 768         self.0.take().unwrap()
 769     }
 770 }
 771
 772 impl<B: ExtraBackendMethods> Deref for AbortCodegenOnDrop<B> {
 773     type Target = OngoingCodegen<B>;
 774
 775     fn deref(&self) -> &OngoingCodegen<B> {
 776         self.0.as_ref().unwrap()
 777     }
 778 }
 779
 780 impl<B: ExtraBackendMethods> DerefMut for AbortCodegenOnDrop<B> {
 781     fn deref_mut(&mut self) -> &mut OngoingCodegen<B> {
 782         self.0.as_mut().unwrap()
 783     }
 784 }
 785
 786 impl<B: ExtraBackendMethods> Drop for AbortCodegenOnDrop<B> {
 787     fn drop(&mut self) {
 788         if let Some(codegen) = self.0.take() {
 789             codegen.codegen_aborted();
 790         }
 791     }
 792 }
 793
 794 impl CrateInfo {
 795     pub fn new(tcx: TyCtxt<'_>, target_cpu: String) -> CrateInfo {
 796         let exported_symbols = tcx
 797             .sess
 798             .crate_types()
 799             .iter()
 800             .map(|&c| (c, crate::back::linker::exported_symbols(tcx, c)))
 801             .collect();
 802         let local_crate_name = tcx.crate_name(LOCAL_CRATE);
 803         let crate_attrs = tcx.hir().attrs(rustc_hir::CRATE_HIR_ID);
 804         let subsystem = tcx.sess.first_attr_value_str_by_name(crate_attrs, sym::windows_subsystem);
 805         let windows_subsystem = subsystem.map(|subsystem| {
 806             if subsystem != sym::windows && subsystem != sym::console {
 807                 tcx.sess.fatal(&format!(
 808                     "invalid windows subsystem `{}`, only \
 809                                      `windows` and `console` are allowed",
 810                     subsystem
 811                 ));
 812             }
 813             subsystem.to_string()
 814         });
 815
 816         // This list is used when generating the command line to pass through to
 817         // system linker. The linker expects undefined symbols on the left of the
 818         // command line to be defined in libraries on the right, not the other way
 819         // around. For more info, see some comments in the add_used_library function
 820         // below.
 821         //
 822         // In order to get this left-to-right dependency ordering, we use the reverse
 823         // postorder of all crates putting the leaves at the right-most positions.
 824         let used_crates = tcx
 825             .postorder_cnums(())
 826             .iter()
 827             .rev()
 828             .copied()
 829             .filter(|&cnum| !tcx.dep_kind(cnum).macros_only())
 830             .collect();
 831
 832         let mut info = CrateInfo {
 833             target_cpu,
 834             exported_symbols,
 835             local_crate_name,
 836             compiler_builtins: None,
 837             profiler_runtime: None,
 838             is_no_builtins: Default::default(),
 839             native_libraries: Default::default(),
 840             used_libraries: tcx.native_libraries(LOCAL_CRATE).iter().map(Into::into).collect(),
 841             crate_name: Default::default(),
 842             used_crates,
 843             used_crate_source: Default::default(),
 844             lang_item_to_crate: Default::default(),
 845             missing_lang_items: Default::default(),
 846             dependency_formats: tcx.dependency_formats(()),
 847             windows_subsystem,
 848         };
 849         let lang_items = tcx.lang_items();
 850
 851         let crates = tcx.crates(());
 852
 853         let n_crates = crates.len();
 854         info.native_libraries.reserve(n_crates);
 855         info.crate_name.reserve(n_crates);
 856         info.used_crate_source.reserve(n_crates);
 857         info.missing_lang_items.reserve(n_crates);
 858
 859         for &cnum in crates.iter() {
 860             info.native_libraries
 861                 .insert(cnum, tcx.native_libraries(cnum).iter().map(Into::into).collect());
 862             info.crate_name.insert(cnum, tcx.crate_name(cnum).to_string());
 863             info.used_crate_source.insert(cnum, tcx.used_crate_source(cnum));
 864             if tcx.is_compiler_builtins(cnum) {
 865                 info.compiler_builtins = Some(cnum);
 866             }
 867             if tcx.is_profiler_runtime(cnum) {
 868                 info.profiler_runtime = Some(cnum);
 869             }
 870             if tcx.is_no_builtins(cnum) {
 871                 info.is_no_builtins.insert(cnum);
 872             }
 873             let missing = tcx.missing_lang_items(cnum);
 874             for &item in missing.iter() {
 875                 if let Ok(id) = lang_items.require(item) {
 876                     info.lang_item_to_crate.insert(item, id.krate);
 877                 }
 878             }
 879
 880             // No need to look for lang items that don't actually need to exist.
 881             let missing =
 882                 missing.iter().cloned().filter(|&l| lang_items::required(tcx, l)).collect();
 883             info.missing_lang_items.insert(cnum, missing);
 884         }
 885
 886         info
 887     }
 888 }
 889
 890 pub fn provide(providers: &mut Providers) {
 891     providers.backend_optimization_level = |tcx, cratenum| {
 892         let for_speed = match tcx.sess.opts.optimize {
 893             // If globally no optimisation is done, #[optimize] has no effect.
 894             //
 895             // This is done because if we ended up "upgrading" to `-O2` here, we’d populate the
 896             // pass manager and it is likely that some module-wide passes (such as inliner or
 897             // cross-function constant propagation) would ignore the `optnone` annotation we put
 898             // on the functions, thus necessarily involving these functions into optimisations.
 899             config::OptLevel::No => return config::OptLevel::No,
 900             // If globally optimise-speed is already specified, just use that level.
 901             config::OptLevel::Less => return config::OptLevel::Less,
 902             config::OptLevel::Default => return config::OptLevel::Default,
 903             config::OptLevel::Aggressive => return config::OptLevel::Aggressive,
 904             // If globally optimize-for-size has been requested, use -O2 instead (if optimize(size)
 905             // are present).
 906             config::OptLevel::Size => config::OptLevel::Default,
 907             config::OptLevel::SizeMin => config::OptLevel::Default,
 908         };
 909
 910         let (defids, _) = tcx.collect_and_partition_mono_items(cratenum);
 911         for id in &*defids {
 912             let CodegenFnAttrs { optimize, .. } = tcx.codegen_fn_attrs(*id);
 913             match optimize {
 914                 attr::OptimizeAttr::None => continue,
 915                 attr::OptimizeAttr::Size => continue,
 916                 attr::OptimizeAttr::Speed => {
 917                     return for_speed;
 918                 }
 919             }
 920         }
 921         tcx.sess.opts.optimize
 922     };
 923 }
 924
 925 fn determine_cgu_reuse<'tcx>(tcx: TyCtxt<'tcx>, cgu: &CodegenUnit<'tcx>) -> CguReuse {
 926     if !tcx.dep_graph.is_fully_enabled() {
 927         return CguReuse::No;
 928     }
 929
 930     let work_product_id = &cgu.work_product_id();
 931     if tcx.dep_graph.previous_work_product(work_product_id).is_none() {
 932         // We don't have anything cached for this CGU. This can happen
 933         // if the CGU did not exist in the previous session.
 934         return CguReuse::No;
 935     }
 936
 937     // Try to mark the CGU as green. If it we can do so, it means that nothing
 938     // affecting the LLVM module has changed and we can re-use a cached version.
 939     // If we compile with any kind of LTO, this means we can re-use the bitcode
 940     // of the Pre-LTO stage (possibly also the Post-LTO version but we'll only
 941     // know that later). If we are not doing LTO, there is only one optimized
 942     // version of each module, so we re-use that.
 943     let dep_node = cgu.codegen_dep_node(tcx);
 944     assert!(
 945         !tcx.dep_graph.dep_node_exists(&dep_node),
 946         "CompileCodegenUnit dep-node for CGU `{}` already exists before marking.",
 947         cgu.name()
 948     );
 949
 950     if tcx.try_mark_green(&dep_node) {
 951         // We can re-use either the pre- or the post-thinlto state. If no LTO is
 952         // being performed then we can use post-LTO artifacts, otherwise we must
 953         // reuse pre-LTO artifacts
 954         match compute_per_cgu_lto_type(
 955             &tcx.sess.lto(),
 956             &tcx.sess.opts,
 957             &tcx.sess.crate_types(),
 958             ModuleKind::Regular,
 959         ) {
 960             ComputedLtoType::No => CguReuse::PostLto,
 961             _ => CguReuse::PreLto,
 962         }
 963     } else {
 964         CguReuse::No
 965     }
 966 }