compiler/rustc_codegen_ssa/src/base.rs

   1 use crate::back::write::{
   2     compute_per_cgu_lto_type, start_async_codegen, submit_codegened_module_to_llvm,
   3     submit_post_lto_module_to_llvm, submit_pre_lto_module_to_llvm, ComputedLtoType, OngoingCodegen,
   4 };
   5 use crate::common::{IntPredicate, RealPredicate, TypeKind};
   6 use crate::meth;
   7 use crate::mir;
   8 use crate::mir::operand::OperandValue;
   9 use crate::mir::place::PlaceRef;
  10 use crate::traits::*;
  11 use crate::{CachedModuleCodegen, CrateInfo, MemFlags, ModuleCodegen, ModuleKind};
  12
  13 use rustc_attr as attr;
  14 use rustc_data_structures::fx::FxHashMap;
  15 use rustc_data_structures::profiling::{get_resident_set_size, print_time_passes_entry};
  16 use rustc_data_structures::sync::{par_iter, ParallelIterator};
  17 use rustc_hir as hir;
  18 use rustc_hir::def_id::{DefId, LOCAL_CRATE};
  19 use rustc_hir::lang_items::LangItem;
  20 use rustc_index::vec::Idx;
  21 use rustc_middle::middle::codegen_fn_attrs::CodegenFnAttrs;
  22 use rustc_middle::middle::cstore::EncodedMetadata;
  23 use rustc_middle::middle::cstore::{self, LinkagePreference};
  24 use rustc_middle::middle::lang_items;
  25 use rustc_middle::mir::mono::{CodegenUnit, CodegenUnitNameBuilder, MonoItem};
  26 use rustc_middle::ty::layout::{HasTyCtxt, TyAndLayout};
  27 use rustc_middle::ty::layout::{FAT_PTR_ADDR, FAT_PTR_EXTRA};
  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};
  32 use rustc_session::Session;
  33 use rustc_target::abi::{Align, LayoutOf, VariantIdx};
  34
  35 use std::ops::{Deref, DerefMut};
  36 use std::time::{Duration, Instant};
  37
  38 use itertools::Itertools;
  39
  40 pub fn bin_op_to_icmp_predicate(op: hir::BinOpKind, signed: bool) -> IntPredicate {
  41     match op {
  42         hir::BinOpKind::Eq => IntPredicate::IntEQ,
  43         hir::BinOpKind::Ne => IntPredicate::IntNE,
  44         hir::BinOpKind::Lt => {
  45             if signed {
  46                 IntPredicate::IntSLT
  47             } else {
  48                 IntPredicate::IntULT
  49             }
  50         }
  51         hir::BinOpKind::Le => {
  52             if signed {
  53                 IntPredicate::IntSLE
  54             } else {
  55                 IntPredicate::IntULE
  56             }
  57         }
  58         hir::BinOpKind::Gt => {
  59             if signed {
  60                 IntPredicate::IntSGT
  61             } else {
  62                 IntPredicate::IntUGT
  63             }
  64         }
  65         hir::BinOpKind::Ge => {
  66             if signed {
  67                 IntPredicate::IntSGE
  68             } else {
  69                 IntPredicate::IntUGE
  70             }
  71         }
  72         op => bug!(
  73             "comparison_op_to_icmp_predicate: expected comparison operator, \
  74              found {:?}",
  75             op
  76         ),
  77     }
  78 }
  79
  80 pub fn bin_op_to_fcmp_predicate(op: hir::BinOpKind) -> RealPredicate {
  81     match op {
  82         hir::BinOpKind::Eq => RealPredicate::RealOEQ,
  83         hir::BinOpKind::Ne => RealPredicate::RealUNE,
  84         hir::BinOpKind::Lt => RealPredicate::RealOLT,
  85         hir::BinOpKind::Le => RealPredicate::RealOLE,
  86         hir::BinOpKind::Gt => RealPredicate::RealOGT,
  87         hir::BinOpKind::Ge => RealPredicate::RealOGE,
  88         op => {
  89             bug!(
  90                 "comparison_op_to_fcmp_predicate: expected comparison operator, \
  91                  found {:?}",
  92                 op
  93             );
  94         }
  95     }
  96 }
  97
  98 pub fn compare_simd_types<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
  99     bx: &mut Bx,
 100     lhs: Bx::Value,
 101     rhs: Bx::Value,
 102     t: Ty<'tcx>,
 103     ret_ty: Bx::Type,
 104     op: hir::BinOpKind,
 105 ) -> Bx::Value {
 106     let signed = match t.kind() {
 107         ty::Float(_) => {
 108             let cmp = bin_op_to_fcmp_predicate(op);
 109             let cmp = bx.fcmp(cmp, lhs, rhs);
 110             return bx.sext(cmp, ret_ty);
 111         }
 112         ty::Uint(_) => false,
 113         ty::Int(_) => true,
 114         _ => bug!("compare_simd_types: invalid SIMD type"),
 115     };
 116
 117     let cmp = bin_op_to_icmp_predicate(op, signed);
 118     let cmp = bx.icmp(cmp, lhs, rhs);
 119     // LLVM outputs an `< size x i1 >`, so we need to perform a sign extension
 120     // to get the correctly sized type. This will compile to a single instruction
 121     // once the IR is converted to assembly if the SIMD instruction is supported
 122     // by the target architecture.
 123     bx.sext(cmp, ret_ty)
 124 }
 125
 126 /// Retrieves the information we are losing (making dynamic) in an unsizing
 127 /// adjustment.
 128 ///
 129 /// The `old_info` argument is a bit odd. It is intended for use in an upcast,
 130 /// where the new vtable for an object will be derived from the old one.
 131 pub fn unsized_info<'tcx, Cx: CodegenMethods<'tcx>>(
 132     cx: &Cx,
 133     source: Ty<'tcx>,
 134     target: Ty<'tcx>,
 135     old_info: Option<Cx::Value>,
 136 ) -> Cx::Value {
 137     let (source, target) =
 138         cx.tcx().struct_lockstep_tails_erasing_lifetimes(source, target, cx.param_env());
 139     match (source.kind(), target.kind()) {
 140         (&ty::Array(_, len), &ty::Slice(_)) => {
 141             cx.const_usize(len.eval_usize(cx.tcx(), ty::ParamEnv::reveal_all()))
 142         }
 143         (&ty::Dynamic(..), &ty::Dynamic(..)) => {
 144             // For now, upcasts are limited to changes in marker
 145             // traits, and hence never actually require an actual
 146             // change to the vtable.
 147             old_info.expect("unsized_info: missing old info for trait upcast")
 148         }
 149         (_, &ty::Dynamic(ref data, ..)) => {
 150             let vtable_ptr = cx.layout_of(cx.tcx().mk_mut_ptr(target)).field(cx, FAT_PTR_EXTRA);
 151             cx.const_ptrcast(
 152                 meth::get_vtable(cx, source, data.principal()),
 153                 cx.backend_type(vtable_ptr),
 154             )
 155         }
 156         _ => bug!("unsized_info: invalid unsizing {:?} -> {:?}", source, target),
 157     }
 158 }
 159
 160 /// Coerces `src` to `dst_ty`. `src_ty` must be a thin pointer.
 161 pub fn unsize_thin_ptr<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
 162     bx: &mut Bx,
 163     src: Bx::Value,
 164     src_ty: Ty<'tcx>,
 165     dst_ty: Ty<'tcx>,
 166 ) -> (Bx::Value, Bx::Value) {
 167     debug!("unsize_thin_ptr: {:?} => {:?}", src_ty, dst_ty);
 168     match (src_ty.kind(), dst_ty.kind()) {
 169         (&ty::Ref(_, a, _), &ty::Ref(_, b, _) | &ty::RawPtr(ty::TypeAndMut { ty: b, .. }))
 170         | (&ty::RawPtr(ty::TypeAndMut { ty: a, .. }), &ty::RawPtr(ty::TypeAndMut { ty: b, .. })) => {
 171             assert!(bx.cx().type_is_sized(a));
 172             let ptr_ty = bx.cx().type_ptr_to(bx.cx().backend_type(bx.cx().layout_of(b)));
 173             (bx.pointercast(src, ptr_ty), unsized_info(bx.cx(), a, b, None))
 174         }
 175         (&ty::Adt(def_a, _), &ty::Adt(def_b, _)) => {
 176             assert_eq!(def_a, def_b);
 177
 178             let src_layout = bx.cx().layout_of(src_ty);
 179             let dst_layout = bx.cx().layout_of(dst_ty);
 180             let mut result = None;
 181             for i in 0..src_layout.fields.count() {
 182                 let src_f = src_layout.field(bx.cx(), i);
 183                 assert_eq!(src_layout.fields.offset(i).bytes(), 0);
 184                 assert_eq!(dst_layout.fields.offset(i).bytes(), 0);
 185                 if src_f.is_zst() {
 186                     continue;
 187                 }
 188                 assert_eq!(src_layout.size, src_f.size);
 189
 190                 let dst_f = dst_layout.field(bx.cx(), i);
 191                 assert_ne!(src_f.ty, dst_f.ty);
 192                 assert_eq!(result, None);
 193                 result = Some(unsize_thin_ptr(bx, src, src_f.ty, dst_f.ty));
 194             }
 195             let (lldata, llextra) = result.unwrap();
 196             // HACK(eddyb) have to bitcast pointers until LLVM removes pointee types.
 197             // FIXME(eddyb) move these out of this `match` arm, so they're always
 198             // applied, uniformly, no matter the source/destination types.
 199             (
 200                 bx.bitcast(lldata, bx.cx().scalar_pair_element_backend_type(dst_layout, 0, true)),
 201                 bx.bitcast(llextra, bx.cx().scalar_pair_element_backend_type(dst_layout, 1, true)),
 202             )
 203         }
 204         _ => bug!("unsize_thin_ptr: called on bad types"),
 205     }
 206 }
 207
 208 /// Coerces `src`, which is a reference to a value of type `src_ty`,
 209 /// to a value of type `dst_ty`, and stores the result in `dst`.
 210 pub fn coerce_unsized_into<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
 211     bx: &mut Bx,
 212     src: PlaceRef<'tcx, Bx::Value>,
 213     dst: PlaceRef<'tcx, Bx::Value>,
 214 ) {
 215     let src_ty = src.layout.ty;
 216     let dst_ty = dst.layout.ty;
 217     match (src_ty.kind(), dst_ty.kind()) {
 218         (&ty::Ref(..), &ty::Ref(..) | &ty::RawPtr(..)) | (&ty::RawPtr(..), &ty::RawPtr(..)) => {
 219             let (base, info) = match bx.load_operand(src).val {
 220                 OperandValue::Pair(base, info) => {
 221                     // fat-ptr to fat-ptr unsize preserves the vtable
 222                     // i.e., &'a fmt::Debug+Send => &'a fmt::Debug
 223                     // So we need to pointercast the base to ensure
 224                     // the types match up.
 225                     // FIXME(eddyb) use `scalar_pair_element_backend_type` here,
 226                     // like `unsize_thin_ptr` does.
 227                     let thin_ptr = dst.layout.field(bx.cx(), FAT_PTR_ADDR);
 228                     (bx.pointercast(base, bx.cx().backend_type(thin_ptr)), info)
 229                 }
 230                 OperandValue::Immediate(base) => unsize_thin_ptr(bx, base, src_ty, dst_ty),
 231                 OperandValue::Ref(..) => bug!(),
 232             };
 233             OperandValue::Pair(base, info).store(bx, dst);
 234         }
 235
 236         (&ty::Adt(def_a, _), &ty::Adt(def_b, _)) => {
 237             assert_eq!(def_a, def_b);
 238
 239             for i in 0..def_a.variants[VariantIdx::new(0)].fields.len() {
 240                 let src_f = src.project_field(bx, i);
 241                 let dst_f = dst.project_field(bx, i);
 242
 243                 if dst_f.layout.is_zst() {
 244                     continue;
 245                 }
 246
 247                 if src_f.layout.ty == dst_f.layout.ty {
 248                     memcpy_ty(
 249                         bx,
 250                         dst_f.llval,
 251                         dst_f.align,
 252                         src_f.llval,
 253                         src_f.align,
 254                         src_f.layout,
 255                         MemFlags::empty(),
 256                     );
 257                 } else {
 258                     coerce_unsized_into(bx, src_f, dst_f);
 259                 }
 260             }
 261         }
 262         _ => bug!("coerce_unsized_into: invalid coercion {:?} -> {:?}", src_ty, dst_ty,),
 263     }
 264 }
 265
 266 pub fn cast_shift_expr_rhs<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
 267     bx: &mut Bx,
 268     op: hir::BinOpKind,
 269     lhs: Bx::Value,
 270     rhs: Bx::Value,
 271 ) -> Bx::Value {
 272     cast_shift_rhs(bx, op, lhs, rhs)
 273 }
 274
 275 fn cast_shift_rhs<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
 276     bx: &mut Bx,
 277     op: hir::BinOpKind,
 278     lhs: Bx::Value,
 279     rhs: Bx::Value,
 280 ) -> Bx::Value {
 281     // Shifts may have any size int on the rhs
 282     if op.is_shift() {
 283         let mut rhs_llty = bx.cx().val_ty(rhs);
 284         let mut lhs_llty = bx.cx().val_ty(lhs);
 285         if bx.cx().type_kind(rhs_llty) == TypeKind::Vector {
 286             rhs_llty = bx.cx().element_type(rhs_llty)
 287         }
 288         if bx.cx().type_kind(lhs_llty) == TypeKind::Vector {
 289             lhs_llty = bx.cx().element_type(lhs_llty)
 290         }
 291         let rhs_sz = bx.cx().int_width(rhs_llty);
 292         let lhs_sz = bx.cx().int_width(lhs_llty);
 293         if lhs_sz < rhs_sz {
 294             bx.trunc(rhs, lhs_llty)
 295         } else if lhs_sz > rhs_sz {
 296             // FIXME (#1877: If in the future shifting by negative
 297             // values is no longer undefined then this is wrong.
 298             bx.zext(rhs, lhs_llty)
 299         } else {
 300             rhs
 301         }
 302     } else {
 303         rhs
 304     }
 305 }
 306
 307 /// Returns `true` if this session's target will use SEH-based unwinding.
 308 ///
 309 /// This is only true for MSVC targets, and even then the 64-bit MSVC target
 310 /// currently uses SEH-ish unwinding with DWARF info tables to the side (same as
 311 /// 64-bit MinGW) instead of "full SEH".
 312 pub fn wants_msvc_seh(sess: &Session) -> bool {
 313     sess.target.is_like_msvc
 314 }
 315
 316 pub fn memcpy_ty<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
 317     bx: &mut Bx,
 318     dst: Bx::Value,
 319     dst_align: Align,
 320     src: Bx::Value,
 321     src_align: Align,
 322     layout: TyAndLayout<'tcx>,
 323     flags: MemFlags,
 324 ) {
 325     let size = layout.size.bytes();
 326     if size == 0 {
 327         return;
 328     }
 329
 330     bx.memcpy(dst, dst_align, src, src_align, bx.cx().const_usize(size), flags);
 331 }
 332
 333 pub fn codegen_instance<'a, 'tcx: 'a, Bx: BuilderMethods<'a, 'tcx>>(
 334     cx: &'a Bx::CodegenCx,
 335     instance: Instance<'tcx>,
 336 ) {
 337     // this is an info! to allow collecting monomorphization statistics
 338     // and to allow finding the last function before LLVM aborts from
 339     // release builds.
 340     info!("codegen_instance({})", instance);
 341
 342     mir::codegen_mir::<Bx>(cx, instance);
 343 }
 344
 345 /// Creates the `main` function which will initialize the rust runtime and call
 346 /// users main function.
 347 pub fn maybe_create_entry_wrapper<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
 348     cx: &'a Bx::CodegenCx,
 349 ) -> Option<Bx::Function> {
 350     let main_def_id = cx.tcx().entry_fn(LOCAL_CRATE).map(|(def_id, _)| def_id)?;
 351     let main_is_local = main_def_id.is_local();
 352     let instance = Instance::mono(cx.tcx(), main_def_id);
 353
 354     if main_is_local {
 355         // We want to create the wrapper in the same codegen unit as Rust's main
 356         // function.
 357         if !cx.codegen_unit().contains_item(&MonoItem::Fn(instance)) {
 358             return None;
 359         }
 360     } else {
 361         // FIXME: Add support for non-local main fn codegen
 362         let span = cx.tcx().main_def.unwrap().span;
 363         let n = 28937;
 364         cx.sess()
 365             .struct_span_err(span, "entry symbol `main` from foreign crate is not yet supported.")
 366             .note(&format!(
 367                 "see issue #{} <https://github.com/rust-lang/rust/issues/{}> \
 368                  for more information",
 369                 n, n,
 370             ))
 371             .emit();
 372         cx.sess().abort_if_errors();
 373         bug!();
 374     }
 375
 376     let main_llfn = cx.get_fn_addr(instance);
 377
 378     return cx.tcx().entry_fn(LOCAL_CRATE).map(|(_, et)| {
 379         let use_start_lang_item = EntryFnType::Start != et;
 380         create_entry_fn::<Bx>(cx, main_llfn, main_def_id, use_start_lang_item)
 381     });
 382
 383     fn create_entry_fn<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
 384         cx: &'a Bx::CodegenCx,
 385         rust_main: Bx::Value,
 386         rust_main_def_id: DefId,
 387         use_start_lang_item: bool,
 388     ) -> Bx::Function {
 389         // The entry function is either `int main(void)` or `int main(int argc, char **argv)`,
 390         // depending on whether the target needs `argc` and `argv` to be passed in.
 391         let llfty = if cx.sess().target.main_needs_argc_argv {
 392             cx.type_func(&[cx.type_int(), cx.type_ptr_to(cx.type_i8p())], cx.type_int())
 393         } else {
 394             cx.type_func(&[], cx.type_int())
 395         };
 396
 397         let main_ret_ty = cx.tcx().fn_sig(rust_main_def_id).output();
 398         // Given that `main()` has no arguments,
 399         // then its return type cannot have
 400         // late-bound regions, since late-bound
 401         // regions must appear in the argument
 402         // listing.
 403         let main_ret_ty = cx.tcx().erase_regions(main_ret_ty.no_bound_vars().unwrap());
 404
 405         let llfn = match cx.declare_c_main(llfty) {
 406             Some(llfn) => llfn,
 407             None => {
 408                 // FIXME: We should be smart and show a better diagnostic here.
 409                 let span = cx.tcx().def_span(rust_main_def_id);
 410                 cx.sess()
 411                     .struct_span_err(span, "entry symbol `main` declared multiple times")
 412                     .help("did you use `#[no_mangle]` on `fn main`? Use `#[start]` instead")
 413                     .emit();
 414                 cx.sess().abort_if_errors();
 415                 bug!();
 416             }
 417         };
 418
 419         // `main` should respect same config for frame pointer elimination as rest of code
 420         cx.set_frame_pointer_elimination(llfn);
 421         cx.apply_target_cpu_attr(llfn);
 422
 423         let mut bx = Bx::new_block(&cx, llfn, "top");
 424
 425         bx.insert_reference_to_gdb_debug_scripts_section_global();
 426
 427         let (arg_argc, arg_argv) = get_argc_argv(cx, &mut bx);
 428
 429         let (start_fn, args) = if use_start_lang_item {
 430             let start_def_id = cx.tcx().require_lang_item(LangItem::Start, None);
 431             let start_fn = cx.get_fn_addr(
 432                 ty::Instance::resolve(
 433                     cx.tcx(),
 434                     ty::ParamEnv::reveal_all(),
 435                     start_def_id,
 436                     cx.tcx().intern_substs(&[main_ret_ty.into()]),
 437                 )
 438                 .unwrap()
 439                 .unwrap(),
 440             );
 441             (
 442                 start_fn,
 443                 vec![bx.pointercast(rust_main, cx.type_ptr_to(cx.type_i8p())), arg_argc, arg_argv],
 444             )
 445         } else {
 446             debug!("using user-defined start fn");
 447             (rust_main, vec![arg_argc, arg_argv])
 448         };
 449
 450         let result = bx.call(start_fn, &args, None);
 451         let cast = bx.intcast(result, cx.type_int(), true);
 452         bx.ret(cast);
 453
 454         llfn
 455     }
 456 }
 457
 458 /// Obtain the `argc` and `argv` values to pass to the rust start function.
 459 fn get_argc_argv<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
 460     cx: &'a Bx::CodegenCx,
 461     bx: &mut Bx,
 462 ) -> (Bx::Value, Bx::Value) {
 463     if cx.sess().target.main_needs_argc_argv {
 464         // Params from native `main()` used as args for rust start function
 465         let param_argc = bx.get_param(0);
 466         let param_argv = bx.get_param(1);
 467         let arg_argc = bx.intcast(param_argc, cx.type_isize(), true);
 468         let arg_argv = param_argv;
 469         (arg_argc, arg_argv)
 470     } else {
 471         // The Rust start function doesn't need `argc` and `argv`, so just pass zeros.
 472         let arg_argc = bx.const_int(cx.type_int(), 0);
 473         let arg_argv = bx.const_null(cx.type_ptr_to(cx.type_i8p()));
 474         (arg_argc, arg_argv)
 475     }
 476 }
 477
 478 pub fn codegen_crate<B: ExtraBackendMethods>(
 479     backend: B,
 480     tcx: TyCtxt<'tcx>,
 481     metadata: EncodedMetadata,
 482     need_metadata_module: bool,
 483 ) -> OngoingCodegen<B> {
 484     // Skip crate items and just output metadata in -Z no-codegen mode.
 485     if tcx.sess.opts.debugging_opts.no_codegen || !tcx.sess.opts.output_types.should_codegen() {
 486         let ongoing_codegen = start_async_codegen(backend, tcx, metadata, 1);
 487
 488         ongoing_codegen.codegen_finished(tcx);
 489
 490         ongoing_codegen.check_for_errors(tcx.sess);
 491
 492         return ongoing_codegen;
 493     }
 494
 495     let cgu_name_builder = &mut CodegenUnitNameBuilder::new(tcx);
 496
 497     // Run the monomorphization collector and partition the collected items into
 498     // codegen units.
 499     let codegen_units = tcx.collect_and_partition_mono_items(LOCAL_CRATE).1;
 500
 501     // Force all codegen_unit queries so they are already either red or green
 502     // when compile_codegen_unit accesses them. We are not able to re-execute
 503     // the codegen_unit query from just the DepNode, so an unknown color would
 504     // lead to having to re-execute compile_codegen_unit, possibly
 505     // unnecessarily.
 506     if tcx.dep_graph.is_fully_enabled() {
 507         for cgu in codegen_units {
 508             tcx.ensure().codegen_unit(cgu.name());
 509         }
 510     }
 511
 512     let ongoing_codegen = start_async_codegen(backend.clone(), tcx, metadata, codegen_units.len());
 513     let ongoing_codegen = AbortCodegenOnDrop::<B>(Some(ongoing_codegen));
 514
 515     // Codegen an allocator shim, if necessary.
 516     //
 517     // If the crate doesn't have an `allocator_kind` set then there's definitely
 518     // no shim to generate. Otherwise we also check our dependency graph for all
 519     // our output crate types. If anything there looks like its a `Dynamic`
 520     // linkage, then it's already got an allocator shim and we'll be using that
 521     // one instead. If nothing exists then it's our job to generate the
 522     // allocator!
 523     let any_dynamic_crate = tcx.dependency_formats(LOCAL_CRATE).iter().any(|(_, list)| {
 524         use rustc_middle::middle::dependency_format::Linkage;
 525         list.iter().any(|&linkage| linkage == Linkage::Dynamic)
 526     });
 527     let allocator_module = if any_dynamic_crate {
 528         None
 529     } else if let Some(kind) = tcx.allocator_kind() {
 530         let llmod_id =
 531             cgu_name_builder.build_cgu_name(LOCAL_CRATE, &["crate"], Some("allocator")).to_string();
 532         let mut modules = backend.new_metadata(tcx, &llmod_id);
 533         tcx.sess.time("write_allocator_module", || {
 534             backend.codegen_allocator(tcx, &mut modules, kind, tcx.lang_items().oom().is_some())
 535         });
 536
 537         Some(ModuleCodegen { name: llmod_id, module_llvm: modules, kind: ModuleKind::Allocator })
 538     } else {
 539         None
 540     };
 541
 542     if let Some(allocator_module) = allocator_module {
 543         ongoing_codegen.submit_pre_codegened_module_to_llvm(tcx, allocator_module);
 544     }
 545
 546     if need_metadata_module {
 547         // Codegen the encoded metadata.
 548         let metadata_cgu_name =
 549             cgu_name_builder.build_cgu_name(LOCAL_CRATE, &["crate"], Some("metadata")).to_string();
 550         let mut metadata_llvm_module = backend.new_metadata(tcx, &metadata_cgu_name);
 551         tcx.sess.time("write_compressed_metadata", || {
 552             backend.write_compressed_metadata(
 553                 tcx,
 554                 &ongoing_codegen.metadata,
 555                 &mut metadata_llvm_module,
 556             );
 557         });
 558
 559         let metadata_module = ModuleCodegen {
 560             name: metadata_cgu_name,
 561             module_llvm: metadata_llvm_module,
 562             kind: ModuleKind::Metadata,
 563         };
 564         ongoing_codegen.submit_pre_codegened_module_to_llvm(tcx, metadata_module);
 565     }
 566
 567     // For better throughput during parallel processing by LLVM, we used to sort
 568     // CGUs largest to smallest. This would lead to better thread utilization
 569     // by, for example, preventing a large CGU from being processed last and
 570     // having only one LLVM thread working while the rest remained idle.
 571     //
 572     // However, this strategy would lead to high memory usage, as it meant the
 573     // LLVM-IR for all of the largest CGUs would be resident in memory at once.
 574     //
 575     // Instead, we can compromise by ordering CGUs such that the largest and
 576     // smallest are first, second largest and smallest are next, etc. If there
 577     // are large size variations, this can reduce memory usage significantly.
 578     let codegen_units: Vec<_> = {
 579         let mut sorted_cgus = codegen_units.iter().collect::<Vec<_>>();
 580         sorted_cgus.sort_by_cached_key(|cgu| cgu.size_estimate());
 581
 582         let (first_half, second_half) = sorted_cgus.split_at(sorted_cgus.len() / 2);
 583         second_half.iter().rev().interleave(first_half).copied().collect()
 584     };
 585
 586     // The non-parallel compiler can only translate codegen units to LLVM IR
 587     // on a single thread, leading to a staircase effect where the N LLVM
 588     // threads have to wait on the single codegen threads to generate work
 589     // for them. The parallel compiler does not have this restriction, so
 590     // we can pre-load the LLVM queue in parallel before handing off
 591     // coordination to the OnGoingCodegen scheduler.
 592     //
 593     // This likely is a temporary measure. Once we don't have to support the
 594     // non-parallel compiler anymore, we can compile CGUs end-to-end in
 595     // parallel and get rid of the complicated scheduling logic.
 596     let pre_compile_cgus = |cgu_reuse: &[CguReuse]| {
 597         if cfg!(parallel_compiler) {
 598             tcx.sess.time("compile_first_CGU_batch", || {
 599                 // Try to find one CGU to compile per thread.
 600                 let cgus: Vec<_> = cgu_reuse
 601                     .iter()
 602                     .enumerate()
 603                     .filter(|&(_, reuse)| reuse == &CguReuse::No)
 604                     .take(tcx.sess.threads())
 605                     .collect();
 606
 607                 // Compile the found CGUs in parallel.
 608                 let start_time = Instant::now();
 609
 610                 let pre_compiled_cgus = par_iter(cgus)
 611                     .map(|(i, _)| {
 612                         let module = backend.compile_codegen_unit(tcx, codegen_units[i].name());
 613                         (i, module)
 614                     })
 615                     .collect();
 616
 617                 (pre_compiled_cgus, start_time.elapsed())
 618             })
 619         } else {
 620             (FxHashMap::default(), Duration::new(0, 0))
 621         }
 622     };
 623
 624     let mut cgu_reuse = Vec::new();
 625     let mut pre_compiled_cgus: Option<FxHashMap<usize, _>> = None;
 626     let mut total_codegen_time = Duration::new(0, 0);
 627     let start_rss = tcx.sess.time_passes().then(|| get_resident_set_size());
 628
 629     for (i, cgu) in codegen_units.iter().enumerate() {
 630         ongoing_codegen.wait_for_signal_to_codegen_item();
 631         ongoing_codegen.check_for_errors(tcx.sess);
 632
 633         // Do some setup work in the first iteration
 634         if pre_compiled_cgus.is_none() {
 635             // Calculate the CGU reuse
 636             cgu_reuse = tcx.sess.time("find_cgu_reuse", || {
 637                 codegen_units.iter().map(|cgu| determine_cgu_reuse(tcx, &cgu)).collect()
 638             });
 639             // Pre compile some CGUs
 640             let (compiled_cgus, codegen_time) = pre_compile_cgus(&cgu_reuse);
 641             pre_compiled_cgus = Some(compiled_cgus);
 642             total_codegen_time += codegen_time;
 643         }
 644
 645         let cgu_reuse = cgu_reuse[i];
 646         tcx.sess.cgu_reuse_tracker.set_actual_reuse(&cgu.name().as_str(), cgu_reuse);
 647
 648         match cgu_reuse {
 649             CguReuse::No => {
 650                 let (module, cost) =
 651                     if let Some(cgu) = pre_compiled_cgus.as_mut().unwrap().remove(&i) {
 652                         cgu
 653                     } else {
 654                         let start_time = Instant::now();
 655                         let module = backend.compile_codegen_unit(tcx, cgu.name());
 656                         total_codegen_time += start_time.elapsed();
 657                         module
 658                     };
 659                 // This will unwind if there are errors, which triggers our `AbortCodegenOnDrop`
 660                 // guard. Unfortunately, just skipping the `submit_codegened_module_to_llvm` makes
 661                 // compilation hang on post-monomorphization errors.
 662                 tcx.sess.abort_if_errors();
 663
 664                 submit_codegened_module_to_llvm(
 665                     &backend,
 666                     &ongoing_codegen.coordinator_send,
 667                     module,
 668                     cost,
 669                 );
 670                 false
 671             }
 672             CguReuse::PreLto => {
 673                 submit_pre_lto_module_to_llvm(
 674                     &backend,
 675                     tcx,
 676                     &ongoing_codegen.coordinator_send,
 677                     CachedModuleCodegen {
 678                         name: cgu.name().to_string(),
 679                         source: cgu.work_product(tcx),
 680                     },
 681                 );
 682                 true
 683             }
 684             CguReuse::PostLto => {
 685                 submit_post_lto_module_to_llvm(
 686                     &backend,
 687                     &ongoing_codegen.coordinator_send,
 688                     CachedModuleCodegen {
 689                         name: cgu.name().to_string(),
 690                         source: cgu.work_product(tcx),
 691                     },
 692                 );
 693                 true
 694             }
 695         };
 696     }
 697
 698     ongoing_codegen.codegen_finished(tcx);
 699
 700     // Since the main thread is sometimes blocked during codegen, we keep track
 701     // -Ztime-passes output manually.
 702     if tcx.sess.time_passes() {
 703         let end_rss = get_resident_set_size();
 704
 705         print_time_passes_entry(
 706             "codegen_to_LLVM_IR",
 707             total_codegen_time,
 708             start_rss.unwrap(),
 709             end_rss,
 710         );
 711     }
 712
 713     ongoing_codegen.check_for_errors(tcx.sess);
 714
 715     ongoing_codegen.into_inner()
 716 }
 717
 718 /// A curious wrapper structure whose only purpose is to call `codegen_aborted`
 719 /// when it's dropped abnormally.
 720 ///
 721 /// In the process of working on rust-lang/rust#55238 a mysterious segfault was
 722 /// stumbled upon. The segfault was never reproduced locally, but it was
 723 /// suspected to be related to the fact that codegen worker threads were
 724 /// sticking around by the time the main thread was exiting, causing issues.
 725 ///
 726 /// This structure is an attempt to fix that issue where the `codegen_aborted`
 727 /// message will block until all workers have finished. This should ensure that
 728 /// even if the main codegen thread panics we'll wait for pending work to
 729 /// complete before returning from the main thread, hopefully avoiding
 730 /// segfaults.
 731 ///
 732 /// If you see this comment in the code, then it means that this workaround
 733 /// worked! We may yet one day track down the mysterious cause of that
 734 /// segfault...
 735 struct AbortCodegenOnDrop<B: ExtraBackendMethods>(Option<OngoingCodegen<B>>);
 736
 737 impl<B: ExtraBackendMethods> AbortCodegenOnDrop<B> {
 738     fn into_inner(mut self) -> OngoingCodegen<B> {
 739         self.0.take().unwrap()
 740     }
 741 }
 742
 743 impl<B: ExtraBackendMethods> Deref for AbortCodegenOnDrop<B> {
 744     type Target = OngoingCodegen<B>;
 745
 746     fn deref(&self) -> &OngoingCodegen<B> {
 747         self.0.as_ref().unwrap()
 748     }
 749 }
 750
 751 impl<B: ExtraBackendMethods> DerefMut for AbortCodegenOnDrop<B> {
 752     fn deref_mut(&mut self) -> &mut OngoingCodegen<B> {
 753         self.0.as_mut().unwrap()
 754     }
 755 }
 756
 757 impl<B: ExtraBackendMethods> Drop for AbortCodegenOnDrop<B> {
 758     fn drop(&mut self) {
 759         if let Some(codegen) = self.0.take() {
 760             codegen.codegen_aborted();
 761         }
 762     }
 763 }
 764
 765 impl CrateInfo {
 766     pub fn new(tcx: TyCtxt<'_>) -> CrateInfo {
 767         let mut info = CrateInfo {
 768             panic_runtime: None,
 769             compiler_builtins: None,
 770             profiler_runtime: None,
 771             is_no_builtins: Default::default(),
 772             native_libraries: Default::default(),
 773             used_libraries: tcx.native_libraries(LOCAL_CRATE).iter().map(Into::into).collect(),
 774             crate_name: Default::default(),
 775             used_crates_dynamic: cstore::used_crates(tcx, LinkagePreference::RequireDynamic),
 776             used_crates_static: cstore::used_crates(tcx, LinkagePreference::RequireStatic),
 777             used_crate_source: Default::default(),
 778             lang_item_to_crate: Default::default(),
 779             missing_lang_items: Default::default(),
 780             dependency_formats: tcx.dependency_formats(LOCAL_CRATE),
 781         };
 782         let lang_items = tcx.lang_items();
 783
 784         let crates = tcx.crates();
 785
 786         let n_crates = crates.len();
 787         info.native_libraries.reserve(n_crates);
 788         info.crate_name.reserve(n_crates);
 789         info.used_crate_source.reserve(n_crates);
 790         info.missing_lang_items.reserve(n_crates);
 791
 792         for &cnum in crates.iter() {
 793             info.native_libraries
 794                 .insert(cnum, tcx.native_libraries(cnum).iter().map(Into::into).collect());
 795             info.crate_name.insert(cnum, tcx.crate_name(cnum).to_string());
 796             info.used_crate_source.insert(cnum, tcx.used_crate_source(cnum));
 797             if tcx.is_panic_runtime(cnum) {
 798                 info.panic_runtime = Some(cnum);
 799             }
 800             if tcx.is_compiler_builtins(cnum) {
 801                 info.compiler_builtins = Some(cnum);
 802             }
 803             if tcx.is_profiler_runtime(cnum) {
 804                 info.profiler_runtime = Some(cnum);
 805             }
 806             if tcx.is_no_builtins(cnum) {
 807                 info.is_no_builtins.insert(cnum);
 808             }
 809             let missing = tcx.missing_lang_items(cnum);
 810             for &item in missing.iter() {
 811                 if let Ok(id) = lang_items.require(item) {
 812                     info.lang_item_to_crate.insert(item, id.krate);
 813                 }
 814             }
 815
 816             // No need to look for lang items that don't actually need to exist.
 817             let missing =
 818                 missing.iter().cloned().filter(|&l| lang_items::required(tcx, l)).collect();
 819             info.missing_lang_items.insert(cnum, missing);
 820         }
 821
 822         info
 823     }
 824 }
 825
 826 pub fn provide(providers: &mut Providers) {
 827     providers.backend_optimization_level = |tcx, cratenum| {
 828         let for_speed = match tcx.sess.opts.optimize {
 829             // If globally no optimisation is done, #[optimize] has no effect.
 830             //
 831             // This is done because if we ended up "upgrading" to `-O2` here, we’d populate the
 832             // pass manager and it is likely that some module-wide passes (such as inliner or
 833             // cross-function constant propagation) would ignore the `optnone` annotation we put
 834             // on the functions, thus necessarily involving these functions into optimisations.
 835             config::OptLevel::No => return config::OptLevel::No,
 836             // If globally optimise-speed is already specified, just use that level.
 837             config::OptLevel::Less => return config::OptLevel::Less,
 838             config::OptLevel::Default => return config::OptLevel::Default,
 839             config::OptLevel::Aggressive => return config::OptLevel::Aggressive,
 840             // If globally optimize-for-size has been requested, use -O2 instead (if optimize(size)
 841             // are present).
 842             config::OptLevel::Size => config::OptLevel::Default,
 843             config::OptLevel::SizeMin => config::OptLevel::Default,
 844         };
 845
 846         let (defids, _) = tcx.collect_and_partition_mono_items(cratenum);
 847         for id in &*defids {
 848             let CodegenFnAttrs { optimize, .. } = tcx.codegen_fn_attrs(*id);
 849             match optimize {
 850                 attr::OptimizeAttr::None => continue,
 851                 attr::OptimizeAttr::Size => continue,
 852                 attr::OptimizeAttr::Speed => {
 853                     return for_speed;
 854                 }
 855             }
 856         }
 857         tcx.sess.opts.optimize
 858     };
 859 }
 860
 861 fn determine_cgu_reuse<'tcx>(tcx: TyCtxt<'tcx>, cgu: &CodegenUnit<'tcx>) -> CguReuse {
 862     if !tcx.dep_graph.is_fully_enabled() {
 863         return CguReuse::No;
 864     }
 865
 866     let work_product_id = &cgu.work_product_id();
 867     if tcx.dep_graph.previous_work_product(work_product_id).is_none() {
 868         // We don't have anything cached for this CGU. This can happen
 869         // if the CGU did not exist in the previous session.
 870         return CguReuse::No;
 871     }
 872
 873     // Try to mark the CGU as green. If it we can do so, it means that nothing
 874     // affecting the LLVM module has changed and we can re-use a cached version.
 875     // If we compile with any kind of LTO, this means we can re-use the bitcode
 876     // of the Pre-LTO stage (possibly also the Post-LTO version but we'll only
 877     // know that later). If we are not doing LTO, there is only one optimized
 878     // version of each module, so we re-use that.
 879     let dep_node = cgu.codegen_dep_node(tcx);
 880     assert!(
 881         !tcx.dep_graph.dep_node_exists(&dep_node),
 882         "CompileCodegenUnit dep-node for CGU `{}` already exists before marking.",
 883         cgu.name()
 884     );
 885
 886     if tcx.try_mark_green(&dep_node) {
 887         // We can re-use either the pre- or the post-thinlto state. If no LTO is
 888         // being performed then we can use post-LTO artifacts, otherwise we must
 889         // reuse pre-LTO artifacts
 890         match compute_per_cgu_lto_type(
 891             &tcx.sess.lto(),
 892             &tcx.sess.opts,
 893             &tcx.sess.crate_types(),
 894             ModuleKind::Regular,
 895         ) {
 896             ComputedLtoType::No => CguReuse::PostLto,
 897             _ => CguReuse::PreLto,
 898         }
 899     } else {
 900         CguReuse::No
 901     }
 902 }