1 // Copyright 2012-2015 The Rust Project Developers. See the COPYRIGHT
2 // file at the top-level directory of this distribution and at
3 // http://rust-lang.org/COPYRIGHT.
5 // Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
6 // http://www.apache.org/licenses/LICENSE-2.0> or the MIT license
7 // <LICENSE-MIT or http://opensource.org/licenses/MIT>, at your
8 // option. This file may not be copied, modified, or distributed
9 // except according to those terms.
11 //! Codegen the completed AST to the LLVM IR.
13 //! Some functions here, such as codegen_block and codegen_expr, return a value --
14 //! the result of the codegen to LLVM -- while others, such as codegen_fn
15 //! and mono_item, are called only for the side effect of adding a
16 //! particular definition to the LLVM IR output we're producing.
18 //! Hopefully useful general knowledge about codegen:
20 //! * There's no way to find out the Ty type of a Value. Doing so
21 //! would be "trying to get the eggs out of an omelette" (credit:
22 //! pcwalton). You can, instead, find out its llvm::Type by calling val_ty,
23 //! but one llvm::Type corresponds to many `Ty`s; for instance, tup(int, int,
24 //! int) and rec(x=int, y=int, z=int) will have the same llvm::Type.
26 use super::ModuleLlvm;
27 use super::ModuleCodegen;
28 use super::ModuleKind;
29 use super::CachedModuleCodegen;
35 use rustc::dep_graph::cgu_reuse_tracker::CguReuse;
36 use rustc::hir::def_id::{CrateNum, DefId, LOCAL_CRATE};
37 use rustc::middle::lang_items::StartFnLangItem;
38 use rustc::middle::weak_lang_items;
39 use rustc::mir::mono::{Linkage, Visibility, Stats, CodegenUnitNameBuilder};
40 use rustc::middle::cstore::{EncodedMetadata};
41 use rustc::ty::{self, Ty, TyCtxt};
42 use rustc::ty::layout::{self, Align, TyLayout, LayoutOf, VariantIdx, HasTyCtxt};
43 use rustc::ty::query::Providers;
44 use rustc::middle::cstore::{self, LinkagePreference};
45 use rustc::middle::exported_symbols;
46 use rustc::util::common::{time, print_time_passes_entry};
47 use rustc::util::profiling::ProfileCategory;
48 use rustc::session::config::{self, DebugInfo, EntryFnType, Lto};
49 use rustc::session::Session;
50 use rustc_incremental;
51 use mir::place::PlaceRef;
52 use builder::{Builder, MemFlags};
54 use rustc_mir::monomorphize::item::DefPathBasedNames;
55 use common::{self, IntPredicate, RealPredicate, TypeKind};
56 use context::CodegenCx;
60 use monomorphize::Instance;
61 use monomorphize::partitioning::{CodegenUnit, CodegenUnitExt};
62 use rustc_codegen_utils::symbol_names_test;
64 use mono_item::{MonoItem, MonoItemExt};
66 use rustc::util::nodemap::FxHashMap;
68 use rustc_data_structures::small_c_str::SmallCStr;
69 use rustc_data_structures::sync::Lrc;
70 use rustc_data_structures::indexed_vec::Idx;
76 use std::ffi::CString;
78 use std::ops::{Deref, DerefMut};
80 use std::time::{Instant, Duration};
82 use syntax_pos::symbol::InternedString;
84 use rustc::hir::{self, CodegenFnAttrs};
88 use mir::operand::OperandValue;
90 use rustc_codegen_utils::check_for_rustc_errors_attr;
92 pub struct StatRecorder<'a, 'tcx, Cx: 'a + CodegenMethods<'tcx>> {
96 _marker: marker::PhantomData<&'tcx ()>,
99 impl<'a, 'tcx, Cx: CodegenMethods<'tcx>> StatRecorder<'a, 'tcx, Cx> {
100 pub fn new(cx: &'a Cx, name: String) -> Self {
101 let istart = cx.stats().borrow().n_llvm_insns;
106 _marker: marker::PhantomData,
111 impl<'a, 'tcx, Cx: CodegenMethods<'tcx>> Drop for StatRecorder<'a, 'tcx, Cx> {
113 if self.cx.sess().codegen_stats() {
114 let mut stats = self.cx.stats().borrow_mut();
115 let iend = stats.n_llvm_insns;
116 stats.fn_stats.push((self.name.take().unwrap(), iend - self.istart));
118 // Reset LLVM insn count to avoid compound costs.
119 stats.n_llvm_insns = self.istart;
124 pub fn bin_op_to_icmp_predicate(op: hir::BinOpKind,
128 hir::BinOpKind::Eq => IntPredicate::IntEQ,
129 hir::BinOpKind::Ne => IntPredicate::IntNE,
130 hir::BinOpKind::Lt => if signed { IntPredicate::IntSLT } else { IntPredicate::IntULT },
131 hir::BinOpKind::Le => if signed { IntPredicate::IntSLE } else { IntPredicate::IntULE },
132 hir::BinOpKind::Gt => if signed { IntPredicate::IntSGT } else { IntPredicate::IntUGT },
133 hir::BinOpKind::Ge => if signed { IntPredicate::IntSGE } else { IntPredicate::IntUGE },
135 bug!("comparison_op_to_icmp_predicate: expected comparison operator, \
142 pub fn bin_op_to_fcmp_predicate(op: hir::BinOpKind) -> RealPredicate {
144 hir::BinOpKind::Eq => RealPredicate::RealOEQ,
145 hir::BinOpKind::Ne => RealPredicate::RealUNE,
146 hir::BinOpKind::Lt => RealPredicate::RealOLT,
147 hir::BinOpKind::Le => RealPredicate::RealOLE,
148 hir::BinOpKind::Gt => RealPredicate::RealOGT,
149 hir::BinOpKind::Ge => RealPredicate::RealOGE,
151 bug!("comparison_op_to_fcmp_predicate: expected comparison operator, \
158 pub fn compare_simd_types<'a, 'tcx: 'a, Bx: BuilderMethods<'a, 'tcx>>(
166 let signed = match t.sty {
168 let cmp = bin_op_to_fcmp_predicate(op);
169 return bx.sext(bx.fcmp(cmp, lhs, rhs), ret_ty);
171 ty::Uint(_) => false,
173 _ => bug!("compare_simd_types: invalid SIMD type"),
176 let cmp = bin_op_to_icmp_predicate(op, signed);
177 // LLVM outputs an `< size x i1 >`, so we need to perform a sign extension
178 // to get the correctly sized type. This will compile to a single instruction
179 // once the IR is converted to assembly if the SIMD instruction is supported
180 // by the target architecture.
181 bx.sext(bx.icmp(cmp, lhs, rhs), ret_ty)
184 /// Retrieve the information we are losing (making dynamic) in an unsizing
187 /// The `old_info` argument is a bit funny. It is intended for use
188 /// in an upcast, where the new vtable for an object will be derived
189 /// from the old one.
190 pub fn unsized_info<'tcx, Cx: CodegenMethods<'tcx>>(
194 old_info: Option<Cx::Value>,
196 let (source, target) = cx.tcx().struct_lockstep_tails(source, target);
197 match (&source.sty, &target.sty) {
198 (&ty::Array(_, len), &ty::Slice(_)) => {
199 cx.const_usize(len.unwrap_usize(cx.tcx()))
201 (&ty::Dynamic(..), &ty::Dynamic(..)) => {
202 // For now, upcasts are limited to changes in marker
203 // traits, and hence never actually require an actual
204 // change to the vtable.
205 old_info.expect("unsized_info: missing old info for trait upcast")
207 (_, &ty::Dynamic(ref data, ..)) => {
208 let vtable_ptr = cx.layout_of(cx.tcx().mk_mut_ptr(target))
209 .field(cx, abi::FAT_PTR_EXTRA);
210 cx.static_ptrcast(meth::get_vtable(cx, source, data.principal()),
211 cx.backend_type(vtable_ptr))
213 _ => bug!("unsized_info: invalid unsizing {:?} -> {:?}",
219 /// Coerce `src` to `dst_ty`. `src_ty` must be a thin pointer.
220 pub fn unsize_thin_ptr<'a, 'tcx: 'a, Bx: BuilderMethods<'a, 'tcx>>(
225 ) -> (Bx::Value, Bx::Value) {
226 debug!("unsize_thin_ptr: {:?} => {:?}", src_ty, dst_ty);
227 match (&src_ty.sty, &dst_ty.sty) {
231 &ty::RawPtr(ty::TypeAndMut { ty: b, .. })) |
232 (&ty::RawPtr(ty::TypeAndMut { ty: a, .. }),
233 &ty::RawPtr(ty::TypeAndMut { ty: b, .. })) => {
234 assert!(bx.cx().type_is_sized(a));
235 let ptr_ty = bx.cx().type_ptr_to(bx.cx().backend_type(bx.cx().layout_of(b)));
236 (bx.pointercast(src, ptr_ty), unsized_info(bx.cx(), a, b, None))
238 (&ty::Adt(def_a, _), &ty::Adt(def_b, _)) if def_a.is_box() && def_b.is_box() => {
239 let (a, b) = (src_ty.boxed_ty(), dst_ty.boxed_ty());
240 assert!(bx.cx().type_is_sized(a));
241 let ptr_ty = bx.cx().type_ptr_to(bx.cx().backend_type(bx.cx().layout_of(b)));
242 (bx.pointercast(src, ptr_ty), unsized_info(bx.cx(), a, b, None))
244 (&ty::Adt(def_a, _), &ty::Adt(def_b, _)) => {
245 assert_eq!(def_a, def_b);
247 let src_layout = bx.cx().layout_of(src_ty);
248 let dst_layout = bx.cx().layout_of(dst_ty);
249 let mut result = None;
250 for i in 0..src_layout.fields.count() {
251 let src_f = src_layout.field(bx.cx(), i);
252 assert_eq!(src_layout.fields.offset(i).bytes(), 0);
253 assert_eq!(dst_layout.fields.offset(i).bytes(), 0);
257 assert_eq!(src_layout.size, src_f.size);
259 let dst_f = dst_layout.field(bx.cx(), i);
260 assert_ne!(src_f.ty, dst_f.ty);
261 assert_eq!(result, None);
262 result = Some(unsize_thin_ptr(bx, src, src_f.ty, dst_f.ty));
264 let (lldata, llextra) = result.unwrap();
265 // HACK(eddyb) have to bitcast pointers until LLVM removes pointee types.
266 (bx.bitcast(lldata, bx.cx().scalar_pair_element_backend_type(dst_layout, 0, true)),
267 bx.bitcast(llextra, bx.cx().scalar_pair_element_backend_type(dst_layout, 1, true)))
269 _ => bug!("unsize_thin_ptr: called on bad types"),
273 /// Coerce `src`, which is a reference to a value of type `src_ty`,
274 /// to a value of type `dst_ty` and store the result in `dst`
275 pub fn coerce_unsized_into<'a, 'tcx: 'a, Bx: BuilderMethods<'a, 'tcx>>(
277 src: PlaceRef<'tcx, Bx::Value>,
278 dst: PlaceRef<'tcx, Bx::Value>
280 let src_ty = src.layout.ty;
281 let dst_ty = dst.layout.ty;
282 let coerce_ptr = || {
283 let (base, info) = match bx.load_operand(src).val {
284 OperandValue::Pair(base, info) => {
285 // fat-ptr to fat-ptr unsize preserves the vtable
286 // i.e. &'a fmt::Debug+Send => &'a fmt::Debug
287 // So we need to pointercast the base to ensure
288 // the types match up.
289 let thin_ptr = dst.layout.field(bx.cx(), abi::FAT_PTR_ADDR);
290 (bx.pointercast(base, bx.cx().backend_type(thin_ptr)), info)
292 OperandValue::Immediate(base) => {
293 unsize_thin_ptr(bx, base, src_ty, dst_ty)
295 OperandValue::Ref(..) => bug!()
297 OperandValue::Pair(base, info).store(bx, dst);
299 match (&src_ty.sty, &dst_ty.sty) {
300 (&ty::Ref(..), &ty::Ref(..)) |
301 (&ty::Ref(..), &ty::RawPtr(..)) |
302 (&ty::RawPtr(..), &ty::RawPtr(..)) => {
305 (&ty::Adt(def_a, _), &ty::Adt(def_b, _)) if def_a.is_box() && def_b.is_box() => {
309 (&ty::Adt(def_a, _), &ty::Adt(def_b, _)) => {
310 assert_eq!(def_a, def_b);
312 for i in 0..def_a.variants[VariantIdx::new(0)].fields.len() {
313 let src_f = src.project_field(bx, i);
314 let dst_f = dst.project_field(bx, i);
316 if dst_f.layout.is_zst() {
320 if src_f.layout.ty == dst_f.layout.ty {
321 memcpy_ty(bx, dst_f.llval, dst_f.align, src_f.llval, src_f.align,
322 src_f.layout, MemFlags::empty());
324 coerce_unsized_into(bx, src_f, dst_f);
328 _ => bug!("coerce_unsized_into: invalid coercion {:?} -> {:?}",
334 pub fn cast_shift_expr_rhs<'a, 'tcx: 'a, Bx: BuilderMethods<'a, 'tcx>>(
340 cast_shift_rhs(bx, op, lhs, rhs, |a, b| bx.trunc(a, b), |a, b| bx.zext(a, b))
343 fn cast_shift_rhs<'a, 'tcx: 'a, F, G, Bx: BuilderMethods<'a, 'tcx>>(
360 // Shifts may have any size int on the rhs
362 let mut rhs_llty = bx.cx().val_ty(rhs);
363 let mut lhs_llty = bx.cx().val_ty(lhs);
364 if bx.cx().type_kind(rhs_llty) == TypeKind::Vector {
365 rhs_llty = bx.cx().element_type(rhs_llty)
367 if bx.cx().type_kind(lhs_llty) == TypeKind::Vector {
368 lhs_llty = bx.cx().element_type(lhs_llty)
370 let rhs_sz = bx.cx().int_width(rhs_llty);
371 let lhs_sz = bx.cx().int_width(lhs_llty);
374 } else if lhs_sz > rhs_sz {
375 // FIXME (#1877: If in the future shifting by negative
376 // values is no longer undefined then this is wrong.
386 /// Returns whether this session's target will use SEH-based unwinding.
388 /// This is only true for MSVC targets, and even then the 64-bit MSVC target
389 /// currently uses SEH-ish unwinding with DWARF info tables to the side (same as
390 /// 64-bit MinGW) instead of "full SEH".
391 pub fn wants_msvc_seh(sess: &Session) -> bool {
392 sess.target.target.options.is_like_msvc
395 pub fn call_assume<'a, 'tcx: 'a, Bx: BuilderMethods<'a, 'tcx>>(
399 let assume_intrinsic = bx.cx().get_intrinsic("llvm.assume");
400 bx.call(assume_intrinsic, &[val], None);
403 pub fn from_immediate<'a, 'tcx: 'a, Bx: BuilderMethods<'a, 'tcx>>(
407 if bx.cx().val_ty(val) == bx.cx().type_i1() {
408 bx.zext(val, bx.cx().type_i8())
414 pub fn to_immediate<'a, 'tcx: 'a, Bx: BuilderMethods<'a, 'tcx>>(
417 layout: layout::TyLayout,
419 if let layout::Abi::Scalar(ref scalar) = layout.abi {
420 return to_immediate_scalar(bx, val, scalar);
425 pub fn to_immediate_scalar<'a, 'tcx: 'a, Bx: BuilderMethods<'a, 'tcx>>(
428 scalar: &layout::Scalar,
430 if scalar.is_bool() {
431 return bx.trunc(val, bx.cx().type_i1());
436 pub fn memcpy_ty<'a, 'tcx: 'a, Bx: BuilderMethods<'a, 'tcx>>(
442 layout: TyLayout<'tcx>,
445 let size = layout.size.bytes();
450 bx.memcpy(dst, dst_align, src, src_align, bx.cx().const_usize(size), flags);
453 pub fn codegen_instance<'a, 'tcx: 'a, Bx: BuilderMethods<'a, 'tcx>>(
454 cx: &'a Bx::CodegenCx,
455 instance: Instance<'tcx>,
457 let _s = if cx.sess().codegen_stats() {
458 let mut instance_name = String::new();
459 DefPathBasedNames::new(cx.tcx(), true, true)
460 .push_def_path(instance.def_id(), &mut instance_name);
461 Some(StatRecorder::new(cx, instance_name))
466 // this is an info! to allow collecting monomorphization statistics
467 // and to allow finding the last function before LLVM aborts from
469 info!("codegen_instance({})", instance);
471 let sig = instance.fn_sig(cx.tcx());
472 let sig = cx.tcx().normalize_erasing_late_bound_regions(ty::ParamEnv::reveal_all(), &sig);
474 let lldecl = cx.instances().borrow().get(&instance).cloned().unwrap_or_else(||
475 bug!("Instance `{:?}` not already declared", instance));
477 cx.stats().borrow_mut().n_closures += 1;
479 let mir = cx.tcx().instance_mir(instance.def);
480 mir::codegen_mir::<Bx>(cx, lldecl, &mir, instance, sig);
483 pub fn set_link_section(llval: &Value, attrs: &CodegenFnAttrs) {
484 let sect = match attrs.link_section {
489 let buf = SmallCStr::new(§.as_str());
490 llvm::LLVMSetSection(llval, buf.as_ptr());
494 /// Create the `main` function which will initialize the rust runtime and call
495 /// users main function.
496 fn maybe_create_entry_wrapper<'a, 'tcx: 'a, Bx: BuilderMethods<'a, 'tcx>>(
497 cx: &'a Bx::CodegenCx
499 let (main_def_id, span) = match *cx.sess().entry_fn.borrow() {
500 Some((id, span, _)) => {
501 (cx.tcx().hir.local_def_id(id), span)
506 let instance = Instance::mono(cx.tcx(), main_def_id);
508 if !cx.codegen_unit().contains_item(&MonoItem::Fn(instance)) {
509 // We want to create the wrapper in the same codegen unit as Rust's main
514 let main_llfn = cx.get_fn(instance);
516 let et = cx.sess().entry_fn.get().map(|e| e.2);
518 Some(EntryFnType::Main) => create_entry_fn::<Bx>(cx, span, main_llfn, main_def_id, true),
519 Some(EntryFnType::Start) => create_entry_fn::<Bx>(cx, span, main_llfn, main_def_id, false),
520 None => {} // Do nothing.
523 fn create_entry_fn<'a, 'tcx: 'a, Bx: BuilderMethods<'a, 'tcx>>(
524 cx: &'a Bx::CodegenCx,
526 rust_main: Bx::Value,
527 rust_main_def_id: DefId,
528 use_start_lang_item: bool,
531 cx.type_func(&[cx.type_int(), cx.type_ptr_to(cx.type_i8p())], cx.type_int());
533 let main_ret_ty = cx.tcx().fn_sig(rust_main_def_id).output();
534 // Given that `main()` has no arguments,
535 // then its return type cannot have
536 // late-bound regions, since late-bound
537 // regions must appear in the argument
539 let main_ret_ty = cx.tcx().erase_regions(
540 &main_ret_ty.no_bound_vars().unwrap(),
543 if cx.get_defined_value("main").is_some() {
544 // FIXME: We should be smart and show a better diagnostic here.
545 cx.sess().struct_span_err(sp, "entry symbol `main` defined multiple times")
546 .help("did you use #[no_mangle] on `fn main`? Use #[start] instead")
548 cx.sess().abort_if_errors();
551 let llfn = cx.declare_cfn("main", llfty);
553 // `main` should respect same config for frame pointer elimination as rest of code
554 cx.set_frame_pointer_elimination(llfn);
555 cx.apply_target_cpu_attr(llfn);
557 let bx = Bx::new_block(&cx, llfn, "top");
559 bx.insert_reference_to_gdb_debug_scripts_section_global();
561 // Params from native main() used as args for rust start function
562 let param_argc = cx.get_param(llfn, 0);
563 let param_argv = cx.get_param(llfn, 1);
564 let arg_argc = bx.intcast(param_argc, cx.type_isize(), true);
565 let arg_argv = param_argv;
567 let (start_fn, args) = if use_start_lang_item {
568 let start_def_id = cx.tcx().require_lang_item(StartFnLangItem);
569 let start_fn = callee::resolve_and_get_fn(
572 cx.tcx().intern_substs(&[main_ret_ty.into()]),
574 (start_fn, vec![bx.pointercast(rust_main, cx.type_ptr_to(cx.type_i8p())),
577 debug!("using user-defined start fn");
578 (rust_main, vec![arg_argc, arg_argv])
581 let result = bx.call(start_fn, &args, None);
582 bx.ret(bx.intcast(result, cx.type_int(), true));
586 pub(crate) fn write_metadata<'a, 'gcx>(
587 tcx: TyCtxt<'a, 'gcx, 'gcx>,
588 llvm_module: &ModuleLlvm
589 ) -> EncodedMetadata {
591 use flate2::Compression;
592 use flate2::write::DeflateEncoder;
594 let (metadata_llcx, metadata_llmod) = (&*llvm_module.llcx, llvm_module.llmod());
596 #[derive(PartialEq, Eq, PartialOrd, Ord)]
603 let kind = tcx.sess.crate_types.borrow().iter().map(|ty| {
605 config::CrateType::Executable |
606 config::CrateType::Staticlib |
607 config::CrateType::Cdylib => MetadataKind::None,
609 config::CrateType::Rlib => MetadataKind::Uncompressed,
611 config::CrateType::Dylib |
612 config::CrateType::ProcMacro => MetadataKind::Compressed,
614 }).max().unwrap_or(MetadataKind::None);
616 if kind == MetadataKind::None {
617 return EncodedMetadata::new();
620 let metadata = tcx.encode_metadata();
621 if kind == MetadataKind::Uncompressed {
625 assert!(kind == MetadataKind::Compressed);
626 let mut compressed = tcx.metadata_encoding_version();
627 DeflateEncoder::new(&mut compressed, Compression::fast())
628 .write_all(&metadata.raw_data).unwrap();
630 let llmeta = common::bytes_in_context(metadata_llcx, &compressed);
631 let llconst = common::struct_in_context(metadata_llcx, &[llmeta], false);
632 let name = exported_symbols::metadata_symbol_name(tcx);
633 let buf = CString::new(name).unwrap();
634 let llglobal = unsafe {
635 llvm::LLVMAddGlobal(metadata_llmod, common::val_ty(llconst), buf.as_ptr())
638 llvm::LLVMSetInitializer(llglobal, llconst);
639 let section_name = metadata::metadata_section_name(&tcx.sess.target.target);
640 let name = SmallCStr::new(section_name);
641 llvm::LLVMSetSection(llglobal, name.as_ptr());
643 // Also generate a .section directive to force no
644 // flags, at least for ELF outputs, so that the
645 // metadata doesn't get loaded into memory.
646 let directive = format!(".section {}", section_name);
647 let directive = CString::new(directive).unwrap();
648 llvm::LLVMSetModuleInlineAsm(metadata_llmod, directive.as_ptr())
653 pub struct ValueIter<'ll> {
654 cur: Option<&'ll Value>,
655 step: unsafe extern "C" fn(&'ll Value) -> Option<&'ll Value>,
658 impl Iterator for ValueIter<'ll> {
659 type Item = &'ll Value;
661 fn next(&mut self) -> Option<&'ll Value> {
663 if let Some(old) = old {
664 self.cur = unsafe { (self.step)(old) };
670 pub fn iter_globals(llmod: &'ll llvm::Module) -> ValueIter<'ll> {
673 cur: llvm::LLVMGetFirstGlobal(llmod),
674 step: llvm::LLVMGetNextGlobal,
679 fn determine_cgu_reuse<'a, 'tcx>(tcx: TyCtxt<'a, 'tcx, 'tcx>,
680 cgu: &CodegenUnit<'tcx>)
682 if !tcx.dep_graph.is_fully_enabled() {
686 let work_product_id = &cgu.work_product_id();
687 if tcx.dep_graph.previous_work_product(work_product_id).is_none() {
688 // We don't have anything cached for this CGU. This can happen
689 // if the CGU did not exist in the previous session.
693 // Try to mark the CGU as green. If it we can do so, it means that nothing
694 // affecting the LLVM module has changed and we can re-use a cached version.
695 // If we compile with any kind of LTO, this means we can re-use the bitcode
696 // of the Pre-LTO stage (possibly also the Post-LTO version but we'll only
697 // know that later). If we are not doing LTO, there is only one optimized
698 // version of each module, so we re-use that.
699 let dep_node = cgu.codegen_dep_node(tcx);
700 assert!(!tcx.dep_graph.dep_node_exists(&dep_node),
701 "CompileCodegenUnit dep-node for CGU `{}` already exists before marking.",
704 if tcx.dep_graph.try_mark_green(tcx, &dep_node).is_some() {
705 // We can re-use either the pre- or the post-thinlto state
706 if tcx.sess.lto() != Lto::No {
716 pub fn codegen_crate<'a, 'tcx, B: BackendMethods>(
718 tcx: TyCtxt<'a, 'tcx, 'tcx>,
719 rx: mpsc::Receiver<Box<dyn Any + Send>>
720 ) -> B::OngoingCodegen {
722 check_for_rustc_errors_attr(tcx);
724 let cgu_name_builder = &mut CodegenUnitNameBuilder::new(tcx);
726 // Codegen the metadata.
727 tcx.sess.profiler(|p| p.start_activity(ProfileCategory::Codegen));
729 let metadata_cgu_name = cgu_name_builder.build_cgu_name(LOCAL_CRATE,
731 Some("metadata")).as_str()
733 let metadata_llvm_module = backend.new_metadata(tcx.sess, &metadata_cgu_name);
734 let metadata = time(tcx.sess, "write metadata", || {
735 backend.write_metadata(tcx, &metadata_llvm_module)
737 tcx.sess.profiler(|p| p.end_activity(ProfileCategory::Codegen));
739 let metadata_module = ModuleCodegen {
740 name: metadata_cgu_name,
741 module_llvm: metadata_llvm_module,
742 kind: ModuleKind::Metadata,
745 let time_graph = if tcx.sess.opts.debugging_opts.codegen_time_graph {
746 Some(time_graph::TimeGraph::new())
751 // Skip crate items and just output metadata in -Z no-codegen mode.
752 if tcx.sess.opts.debugging_opts.no_codegen ||
753 !tcx.sess.opts.output_types.should_codegen() {
754 let ongoing_codegen = backend.start_async_codegen(
761 backend.submit_pre_codegened_module_to_llvm(&ongoing_codegen, tcx, metadata_module);
762 backend.codegen_finished(&ongoing_codegen, tcx);
764 assert_and_save_dep_graph(tcx);
766 backend.check_for_errors(&ongoing_codegen, tcx.sess);
768 return ongoing_codegen;
771 // Run the monomorphization collector and partition the collected items into
773 let codegen_units = tcx.collect_and_partition_mono_items(LOCAL_CRATE).1;
774 let codegen_units = (*codegen_units).clone();
776 // Force all codegen_unit queries so they are already either red or green
777 // when compile_codegen_unit accesses them. We are not able to re-execute
778 // the codegen_unit query from just the DepNode, so an unknown color would
779 // lead to having to re-execute compile_codegen_unit, possibly
781 if tcx.dep_graph.is_fully_enabled() {
782 for cgu in &codegen_units {
783 tcx.codegen_unit(cgu.name().clone());
787 let ongoing_codegen = backend.start_async_codegen(
792 codegen_units.len());
793 let ongoing_codegen = AbortCodegenOnDrop::<B>(Some(ongoing_codegen));
795 // Codegen an allocator shim, if necessary.
797 // If the crate doesn't have an `allocator_kind` set then there's definitely
798 // no shim to generate. Otherwise we also check our dependency graph for all
799 // our output crate types. If anything there looks like its a `Dynamic`
800 // linkage, then it's already got an allocator shim and we'll be using that
801 // one instead. If nothing exists then it's our job to generate the
803 let any_dynamic_crate = tcx.sess.dependency_formats.borrow()
806 use rustc::middle::dependency_format::Linkage;
807 list.iter().any(|&linkage| linkage == Linkage::Dynamic)
809 let allocator_module = if any_dynamic_crate {
811 } else if let Some(kind) = *tcx.sess.allocator_kind.get() {
812 let llmod_id = cgu_name_builder.build_cgu_name(LOCAL_CRATE,
814 Some("allocator")).as_str()
816 let modules = backend.new_metadata(tcx.sess, &llmod_id);
817 time(tcx.sess, "write allocator module", || {
818 backend.codegen_allocator(tcx, &modules, kind)
823 module_llvm: modules,
824 kind: ModuleKind::Allocator,
830 if let Some(allocator_module) = allocator_module {
831 backend.submit_pre_codegened_module_to_llvm(&ongoing_codegen, tcx, allocator_module);
834 backend.submit_pre_codegened_module_to_llvm(&ongoing_codegen, tcx, metadata_module);
836 // We sort the codegen units by size. This way we can schedule work for LLVM
837 // a bit more efficiently.
838 let codegen_units = {
839 let mut codegen_units = codegen_units;
840 codegen_units.sort_by_cached_key(|cgu| cmp::Reverse(cgu.size_estimate()));
844 let mut total_codegen_time = Duration::new(0, 0);
845 let mut all_stats = Stats::default();
847 for cgu in codegen_units.into_iter() {
848 backend.wait_for_signal_to_codegen_item(&ongoing_codegen);
849 backend.check_for_errors(&ongoing_codegen, tcx.sess);
851 let cgu_reuse = determine_cgu_reuse(tcx, &cgu);
852 tcx.sess.cgu_reuse_tracker.set_actual_reuse(&cgu.name().as_str(), cgu_reuse);
856 let _timing_guard = time_graph.as_ref().map(|time_graph| {
857 time_graph.start(write::CODEGEN_WORKER_TIMELINE,
858 write::CODEGEN_WORK_PACKAGE_KIND,
859 &format!("codegen {}", cgu.name()))
861 let start_time = Instant::now();
862 let stats = compile_codegen_unit(tcx, *cgu.name());
863 all_stats.extend(stats);
864 total_codegen_time += start_time.elapsed();
867 CguReuse::PreLto => {
868 write::submit_pre_lto_module_to_llvm(tcx, CachedModuleCodegen {
869 name: cgu.name().to_string(),
870 source: cgu.work_product(tcx),
874 CguReuse::PostLto => {
875 write::submit_post_lto_module_to_llvm(tcx, CachedModuleCodegen {
876 name: cgu.name().to_string(),
877 source: cgu.work_product(tcx),
884 backend.codegen_finished(&ongoing_codegen, tcx);
886 // Since the main thread is sometimes blocked during codegen, we keep track
887 // -Ztime-passes output manually.
888 print_time_passes_entry(tcx.sess.time_passes(),
889 "codegen to LLVM IR",
892 rustc_incremental::assert_module_sources::assert_module_sources(tcx);
894 symbol_names_test::report_symbol_names(tcx);
896 if tcx.sess.codegen_stats() {
897 println!("--- codegen stats ---");
898 println!("n_glues_created: {}", all_stats.n_glues_created);
899 println!("n_null_glues: {}", all_stats.n_null_glues);
900 println!("n_real_glues: {}", all_stats.n_real_glues);
902 println!("n_fns: {}", all_stats.n_fns);
903 println!("n_inlines: {}", all_stats.n_inlines);
904 println!("n_closures: {}", all_stats.n_closures);
905 println!("fn stats:");
906 all_stats.fn_stats.sort_by_key(|&(_, insns)| insns);
907 for &(ref name, insns) in all_stats.fn_stats.iter() {
908 println!("{} insns, {}", insns, *name);
912 if tcx.sess.count_llvm_insns() {
913 for (k, v) in all_stats.llvm_insns.iter() {
914 println!("{:7} {}", *v, *k);
918 backend.check_for_errors(&ongoing_codegen, tcx.sess);
920 assert_and_save_dep_graph(tcx);
921 ongoing_codegen.into_inner()
924 /// A curious wrapper structure whose only purpose is to call `codegen_aborted`
925 /// when it's dropped abnormally.
927 /// In the process of working on rust-lang/rust#55238 a mysterious segfault was
928 /// stumbled upon. The segfault was never reproduced locally, but it was
929 /// suspected to be related to the fact that codegen worker threads were
930 /// sticking around by the time the main thread was exiting, causing issues.
932 /// This structure is an attempt to fix that issue where the `codegen_aborted`
933 /// message will block until all workers have finished. This should ensure that
934 /// even if the main codegen thread panics we'll wait for pending work to
935 /// complete before returning from the main thread, hopefully avoiding
938 /// If you see this comment in the code, then it means that this workaround
939 /// worked! We may yet one day track down the mysterious cause of that
941 struct AbortCodegenOnDrop<B: BackendMethods>(Option<B::OngoingCodegen>);
943 impl<B: BackendMethods> AbortCodegenOnDrop<B> {
944 fn into_inner(mut self) -> B::OngoingCodegen {
945 self.0.take().unwrap()
949 impl<B: BackendMethods> Deref for AbortCodegenOnDrop<B> {
950 type Target = B::OngoingCodegen;
952 fn deref(&self) -> &B::OngoingCodegen {
953 self.0.as_ref().unwrap()
957 impl<B: BackendMethods> DerefMut for AbortCodegenOnDrop<B> {
958 fn deref_mut(&mut self) -> &mut B::OngoingCodegen {
959 self.0.as_mut().unwrap()
963 impl<B: BackendMethods> Drop for AbortCodegenOnDrop<B> {
965 if let Some(codegen) = self.0.take() {
966 B::codegen_aborted(codegen);
971 fn assert_and_save_dep_graph<'a, 'tcx>(tcx: TyCtxt<'a, 'tcx, 'tcx>) {
974 || rustc_incremental::assert_dep_graph(tcx));
977 "serialize dep graph",
978 || rustc_incremental::save_dep_graph(tcx));
982 pub fn new(tcx: TyCtxt) -> CrateInfo {
983 let mut info = CrateInfo {
985 compiler_builtins: None,
986 profiler_runtime: None,
987 sanitizer_runtime: None,
988 is_no_builtins: Default::default(),
989 native_libraries: Default::default(),
990 used_libraries: tcx.native_libraries(LOCAL_CRATE),
991 link_args: tcx.link_args(LOCAL_CRATE),
992 crate_name: Default::default(),
993 used_crates_dynamic: cstore::used_crates(tcx, LinkagePreference::RequireDynamic),
994 used_crates_static: cstore::used_crates(tcx, LinkagePreference::RequireStatic),
995 used_crate_source: Default::default(),
996 wasm_imports: Default::default(),
997 lang_item_to_crate: Default::default(),
998 missing_lang_items: Default::default(),
1000 let lang_items = tcx.lang_items();
1002 let load_wasm_items = tcx.sess.crate_types.borrow()
1004 .any(|c| *c != config::CrateType::Rlib) &&
1005 tcx.sess.opts.target_triple.triple() == "wasm32-unknown-unknown";
1007 if load_wasm_items {
1008 info.load_wasm_imports(tcx, LOCAL_CRATE);
1011 let crates = tcx.crates();
1013 let n_crates = crates.len();
1014 info.native_libraries.reserve(n_crates);
1015 info.crate_name.reserve(n_crates);
1016 info.used_crate_source.reserve(n_crates);
1017 info.missing_lang_items.reserve(n_crates);
1019 for &cnum in crates.iter() {
1020 info.native_libraries.insert(cnum, tcx.native_libraries(cnum));
1021 info.crate_name.insert(cnum, tcx.crate_name(cnum).to_string());
1022 info.used_crate_source.insert(cnum, tcx.used_crate_source(cnum));
1023 if tcx.is_panic_runtime(cnum) {
1024 info.panic_runtime = Some(cnum);
1026 if tcx.is_compiler_builtins(cnum) {
1027 info.compiler_builtins = Some(cnum);
1029 if tcx.is_profiler_runtime(cnum) {
1030 info.profiler_runtime = Some(cnum);
1032 if tcx.is_sanitizer_runtime(cnum) {
1033 info.sanitizer_runtime = Some(cnum);
1035 if tcx.is_no_builtins(cnum) {
1036 info.is_no_builtins.insert(cnum);
1038 if load_wasm_items {
1039 info.load_wasm_imports(tcx, cnum);
1041 let missing = tcx.missing_lang_items(cnum);
1042 for &item in missing.iter() {
1043 if let Ok(id) = lang_items.require(item) {
1044 info.lang_item_to_crate.insert(item, id.krate);
1048 // No need to look for lang items that are whitelisted and don't
1049 // actually need to exist.
1050 let missing = missing.iter()
1052 .filter(|&l| !weak_lang_items::whitelisted(tcx, l))
1054 info.missing_lang_items.insert(cnum, missing);
1060 fn load_wasm_imports(&mut self, tcx: TyCtxt, cnum: CrateNum) {
1061 self.wasm_imports.extend(tcx.wasm_import_module_map(cnum).iter().map(|(&id, module)| {
1062 let instance = Instance::mono(tcx, id);
1063 let import_name = tcx.symbol_name(instance);
1065 (import_name.to_string(), module.clone())
1070 fn compile_codegen_unit<'a, 'tcx>(tcx: TyCtxt<'a, 'tcx, 'tcx>,
1071 cgu_name: InternedString)
1073 let start_time = Instant::now();
1075 let dep_node = tcx.codegen_unit(cgu_name).codegen_dep_node(tcx);
1076 let ((stats, module), _) = tcx.dep_graph.with_task(dep_node,
1080 let time_to_codegen = start_time.elapsed();
1082 // We assume that the cost to run LLVM on a CGU is proportional to
1083 // the time we needed for codegenning it.
1084 let cost = time_to_codegen.as_secs() * 1_000_000_000 +
1085 time_to_codegen.subsec_nanos() as u64;
1087 write::submit_codegened_module_to_llvm(tcx,
1092 fn module_codegen<'a, 'tcx>(
1093 tcx: TyCtxt<'a, 'tcx, 'tcx>,
1094 cgu_name: InternedString)
1095 -> (Stats, ModuleCodegen<ModuleLlvm>)
1097 let cgu = tcx.codegen_unit(cgu_name);
1099 // Instantiate monomorphizations without filling out definitions yet...
1100 let llvm_module = ModuleLlvm::new(tcx.sess, &cgu_name.as_str());
1102 let cx = CodegenCx::new(tcx, cgu, &llvm_module);
1103 let mono_items = cx.codegen_unit
1104 .items_in_deterministic_order(cx.tcx);
1105 for &(mono_item, (linkage, visibility)) in &mono_items {
1106 mono_item.predefine(&cx, linkage, visibility);
1109 // ... and now that we have everything pre-defined, fill out those definitions.
1110 for &(mono_item, _) in &mono_items {
1111 mono_item.define(&cx);
1114 // If this codegen unit contains the main function, also create the
1116 maybe_create_entry_wrapper::<Builder<&Value>>(&cx);
1118 // Run replace-all-uses-with for statics that need it
1119 for &(old_g, new_g) in cx.statics_to_rauw.borrow().iter() {
1121 let bitcast = llvm::LLVMConstPointerCast(new_g, cx.val_ty(old_g));
1122 llvm::LLVMReplaceAllUsesWith(old_g, bitcast);
1123 llvm::LLVMDeleteGlobal(old_g);
1127 // Create the llvm.used variable
1128 // This variable has type [N x i8*] and is stored in the llvm.metadata section
1129 if !cx.used_statics.borrow().is_empty() {
1130 let name = const_cstr!("llvm.used");
1131 let section = const_cstr!("llvm.metadata");
1132 let array = cx.const_array(
1133 &cx.type_ptr_to(cx.type_i8()),
1134 &*cx.used_statics.borrow()
1138 let g = llvm::LLVMAddGlobal(cx.llmod,
1141 llvm::LLVMSetInitializer(g, array);
1142 llvm::LLVMRustSetLinkage(g, llvm::Linkage::AppendingLinkage);
1143 llvm::LLVMSetSection(g, section.as_ptr());
1147 // Finalize debuginfo
1148 if cx.sess().opts.debuginfo != DebugInfo::None {
1149 debuginfo::finalize(&cx);
1152 cx.stats.into_inner()
1155 (stats, ModuleCodegen {
1156 name: cgu_name.to_string(),
1157 module_llvm: llvm_module,
1158 kind: ModuleKind::Regular,
1163 pub fn provide_both(providers: &mut Providers) {
1164 providers.dllimport_foreign_items = |tcx, krate| {
1165 let module_map = tcx.foreign_modules(krate);
1166 let module_map = module_map.iter()
1167 .map(|lib| (lib.def_id, lib))
1168 .collect::<FxHashMap<_, _>>();
1170 let dllimports = tcx.native_libraries(krate)
1173 if lib.kind != cstore::NativeLibraryKind::NativeUnknown {
1176 let cfg = match lib.cfg {
1177 Some(ref cfg) => cfg,
1178 None => return true,
1180 attr::cfg_matches(cfg, &tcx.sess.parse_sess, None)
1182 .filter_map(|lib| lib.foreign_module)
1183 .map(|id| &module_map[&id])
1184 .flat_map(|module| module.foreign_items.iter().cloned())
1186 Lrc::new(dllimports)
1189 providers.is_dllimport_foreign_item = |tcx, def_id| {
1190 tcx.dllimport_foreign_items(def_id.krate).contains(&def_id)
1194 pub fn linkage_to_llvm(linkage: Linkage) -> llvm::Linkage {
1196 Linkage::External => llvm::Linkage::ExternalLinkage,
1197 Linkage::AvailableExternally => llvm::Linkage::AvailableExternallyLinkage,
1198 Linkage::LinkOnceAny => llvm::Linkage::LinkOnceAnyLinkage,
1199 Linkage::LinkOnceODR => llvm::Linkage::LinkOnceODRLinkage,
1200 Linkage::WeakAny => llvm::Linkage::WeakAnyLinkage,
1201 Linkage::WeakODR => llvm::Linkage::WeakODRLinkage,
1202 Linkage::Appending => llvm::Linkage::AppendingLinkage,
1203 Linkage::Internal => llvm::Linkage::InternalLinkage,
1204 Linkage::Private => llvm::Linkage::PrivateLinkage,
1205 Linkage::ExternalWeak => llvm::Linkage::ExternalWeakLinkage,
1206 Linkage::Common => llvm::Linkage::CommonLinkage,
1210 pub fn visibility_to_llvm(linkage: Visibility) -> llvm::Visibility {
1212 Visibility::Default => llvm::Visibility::Default,
1213 Visibility::Hidden => llvm::Visibility::Hidden,
1214 Visibility::Protected => llvm::Visibility::Protected,
1218 // FIXME(mw): Anything that is produced via DepGraph::with_task() must implement
1219 // the HashStable trait. Normally DepGraph::with_task() calls are
1220 // hidden behind queries, but CGU creation is a special case in two
1221 // ways: (1) it's not a query and (2) CGU are output nodes, so their
1222 // Fingerprints are not actually needed. It remains to be clarified
1223 // how exactly this case will be handled in the red/green system but
1224 // for now we content ourselves with providing a no-op HashStable
1225 // implementation for CGUs.
1226 mod temp_stable_hash_impls {
1227 use rustc_data_structures::stable_hasher::{StableHasherResult, StableHasher,
1229 use {ModuleCodegen, ModuleLlvm};
1231 impl<HCX> HashStable<HCX> for ModuleCodegen<ModuleLlvm> {
1232 fn hash_stable<W: StableHasherResult>(&self,
1234 _: &mut StableHasher<W>) {