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;
32 use back::write::{self, OngoingCodegen};
33 use llvm::{self, TypeKind, get_param};
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};
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;
52 use mir::place::PlaceRef;
54 use builder::{Builder, MemFlags};
56 use common::{C_bool, C_bytes_in_context, C_usize};
57 use rustc_mir::monomorphize::item::DefPathBasedNames;
58 use common::{C_struct_in_context, C_array, val_ty};
60 use context::CodegenCx;
65 use monomorphize::Instance;
66 use monomorphize::partitioning::{CodegenUnit, CodegenUnitExt};
67 use rustc_codegen_utils::symbol_names_test;
69 use mono_item::{MonoItem, MonoItemExt};
71 use type_of::LayoutLlvmExt;
72 use rustc::util::nodemap::FxHashMap;
74 use rustc_data_structures::small_c_str::SmallCStr;
75 use rustc_data_structures::sync::Lrc;
76 use rustc_data_structures::indexed_vec::Idx;
78 use traits::BuilderMethods;
83 use std::ffi::CString;
84 use std::ops::{Deref, DerefMut};
86 use std::time::{Instant, Duration};
88 use syntax_pos::symbol::InternedString;
90 use rustc::hir::{self, CodegenFnAttrs};
92 use value::{Value, ValueTrait};
94 use mir::operand::OperandValue;
96 use rustc_codegen_utils::check_for_rustc_errors_attr;
98 pub struct StatRecorder<'a, 'll: 'a, 'tcx: 'll> {
99 cx: &'a CodegenCx<'ll, 'tcx>,
100 name: Option<String>,
104 impl StatRecorder<'a, 'll, 'tcx> {
105 pub fn new(cx: &'a CodegenCx<'ll, 'tcx>, name: String) -> Self {
106 let istart = cx.stats.borrow().n_llvm_insns;
115 impl Drop for StatRecorder<'a, 'll, 'tcx> {
117 if self.cx.sess().codegen_stats() {
118 let mut stats = self.cx.stats.borrow_mut();
119 let iend = stats.n_llvm_insns;
120 stats.fn_stats.push((self.name.take().unwrap(), iend - self.istart));
122 // Reset LLVM insn count to avoid compound costs.
123 stats.n_llvm_insns = self.istart;
128 pub fn bin_op_to_icmp_predicate(op: hir::BinOpKind,
130 -> llvm::IntPredicate {
132 hir::BinOpKind::Eq => llvm::IntEQ,
133 hir::BinOpKind::Ne => llvm::IntNE,
134 hir::BinOpKind::Lt => if signed { llvm::IntSLT } else { llvm::IntULT },
135 hir::BinOpKind::Le => if signed { llvm::IntSLE } else { llvm::IntULE },
136 hir::BinOpKind::Gt => if signed { llvm::IntSGT } else { llvm::IntUGT },
137 hir::BinOpKind::Ge => if signed { llvm::IntSGE } else { llvm::IntUGE },
139 bug!("comparison_op_to_icmp_predicate: expected comparison operator, \
146 pub fn bin_op_to_fcmp_predicate(op: hir::BinOpKind) -> llvm::RealPredicate {
148 hir::BinOpKind::Eq => llvm::RealOEQ,
149 hir::BinOpKind::Ne => llvm::RealUNE,
150 hir::BinOpKind::Lt => llvm::RealOLT,
151 hir::BinOpKind::Le => llvm::RealOLE,
152 hir::BinOpKind::Gt => llvm::RealOGT,
153 hir::BinOpKind::Ge => llvm::RealOGE,
155 bug!("comparison_op_to_fcmp_predicate: expected comparison operator, \
162 pub fn compare_simd_types(
163 bx: &Builder<'a, 'll, 'tcx>,
170 let signed = match t.sty {
172 let cmp = bin_op_to_fcmp_predicate(op);
173 return bx.sext(bx.fcmp(cmp, lhs, rhs), ret_ty);
175 ty::Uint(_) => false,
177 _ => bug!("compare_simd_types: invalid SIMD type"),
180 let cmp = bin_op_to_icmp_predicate(op, signed);
181 // LLVM outputs an `< size x i1 >`, so we need to perform a sign extension
182 // to get the correctly sized type. This will compile to a single instruction
183 // once the IR is converted to assembly if the SIMD instruction is supported
184 // by the target architecture.
185 bx.sext(bx.icmp(cmp, lhs, rhs), ret_ty)
188 /// Retrieve the information we are losing (making dynamic) in an unsizing
191 /// The `old_info` argument is a bit funny. It is intended for use
192 /// in an upcast, where the new vtable for an object will be derived
193 /// from the old one.
195 cx: &CodegenCx<'ll, 'tcx>,
198 old_info: Option<&'ll Value>,
200 let (source, target) = cx.tcx.struct_lockstep_tails(source, target);
201 match (&source.sty, &target.sty) {
202 (&ty::Array(_, len), &ty::Slice(_)) => {
203 C_usize(cx, len.unwrap_usize(cx.tcx))
205 (&ty::Dynamic(..), &ty::Dynamic(..)) => {
206 // For now, upcasts are limited to changes in marker
207 // traits, and hence never actually require an actual
208 // change to the vtable.
209 old_info.expect("unsized_info: missing old info for trait upcast")
211 (_, &ty::Dynamic(ref data, ..)) => {
212 let vtable_ptr = cx.layout_of(cx.tcx.mk_mut_ptr(target))
213 .field(cx, abi::FAT_PTR_EXTRA);
214 consts::ptrcast(meth::get_vtable(cx, source, data.principal()),
215 vtable_ptr.llvm_type(cx))
217 _ => bug!("unsized_info: invalid unsizing {:?} -> {:?}",
223 /// Coerce `src` to `dst_ty`. `src_ty` must be a thin pointer.
224 pub fn unsize_thin_ptr(
225 bx: &Builder<'a, 'll, 'tcx>,
229 ) -> (&'ll Value, &'ll Value) {
230 debug!("unsize_thin_ptr: {:?} => {:?}", src_ty, dst_ty);
231 match (&src_ty.sty, &dst_ty.sty) {
235 &ty::RawPtr(ty::TypeAndMut { ty: b, .. })) |
236 (&ty::RawPtr(ty::TypeAndMut { ty: a, .. }),
237 &ty::RawPtr(ty::TypeAndMut { ty: b, .. })) => {
238 assert!(bx.cx.type_is_sized(a));
239 let ptr_ty = bx.cx.layout_of(b).llvm_type(bx.cx).ptr_to();
240 (bx.pointercast(src, ptr_ty), unsized_info(bx.cx, a, b, None))
242 (&ty::Adt(def_a, _), &ty::Adt(def_b, _)) if def_a.is_box() && def_b.is_box() => {
243 let (a, b) = (src_ty.boxed_ty(), dst_ty.boxed_ty());
244 assert!(bx.cx.type_is_sized(a));
245 let ptr_ty = bx.cx.layout_of(b).llvm_type(bx.cx).ptr_to();
246 (bx.pointercast(src, ptr_ty), unsized_info(bx.cx, a, b, None))
248 (&ty::Adt(def_a, _), &ty::Adt(def_b, _)) => {
249 assert_eq!(def_a, def_b);
251 let src_layout = bx.cx.layout_of(src_ty);
252 let dst_layout = bx.cx.layout_of(dst_ty);
253 let mut result = None;
254 for i in 0..src_layout.fields.count() {
255 let src_f = src_layout.field(bx.cx, i);
256 assert_eq!(src_layout.fields.offset(i).bytes(), 0);
257 assert_eq!(dst_layout.fields.offset(i).bytes(), 0);
261 assert_eq!(src_layout.size, src_f.size);
263 let dst_f = dst_layout.field(bx.cx, i);
264 assert_ne!(src_f.ty, dst_f.ty);
265 assert_eq!(result, None);
266 result = Some(unsize_thin_ptr(bx, src, src_f.ty, dst_f.ty));
268 let (lldata, llextra) = result.unwrap();
269 // HACK(eddyb) have to bitcast pointers until LLVM removes pointee types.
270 (bx.bitcast(lldata, dst_layout.scalar_pair_element_llvm_type(bx.cx, 0, true)),
271 bx.bitcast(llextra, dst_layout.scalar_pair_element_llvm_type(bx.cx, 1, true)))
273 _ => bug!("unsize_thin_ptr: called on bad types"),
277 /// Coerce `src`, which is a reference to a value of type `src_ty`,
278 /// to a value of type `dst_ty` and store the result in `dst`
279 pub fn coerce_unsized_into(
280 bx: &Builder<'a, 'll, 'tcx>,
281 src: PlaceRef<'tcx, &'ll Value>,
282 dst: PlaceRef<'tcx, &'ll Value>
284 let src_ty = src.layout.ty;
285 let dst_ty = dst.layout.ty;
286 let coerce_ptr = || {
287 let (base, info) = match src.load(bx).val {
288 OperandValue::Pair(base, info) => {
289 // fat-ptr to fat-ptr unsize preserves the vtable
290 // i.e. &'a fmt::Debug+Send => &'a fmt::Debug
291 // So we need to pointercast the base to ensure
292 // the types match up.
293 let thin_ptr = dst.layout.field(bx.cx, abi::FAT_PTR_ADDR);
294 (bx.pointercast(base, thin_ptr.llvm_type(bx.cx)), info)
296 OperandValue::Immediate(base) => {
297 unsize_thin_ptr(bx, base, src_ty, dst_ty)
299 OperandValue::Ref(..) => bug!()
301 OperandValue::Pair(base, info).store(bx, dst);
303 match (&src_ty.sty, &dst_ty.sty) {
304 (&ty::Ref(..), &ty::Ref(..)) |
305 (&ty::Ref(..), &ty::RawPtr(..)) |
306 (&ty::RawPtr(..), &ty::RawPtr(..)) => {
309 (&ty::Adt(def_a, _), &ty::Adt(def_b, _)) if def_a.is_box() && def_b.is_box() => {
313 (&ty::Adt(def_a, _), &ty::Adt(def_b, _)) => {
314 assert_eq!(def_a, def_b);
316 for i in 0..def_a.variants[VariantIdx::new(0)].fields.len() {
317 let src_f = src.project_field(bx, i);
318 let dst_f = dst.project_field(bx, i);
320 if dst_f.layout.is_zst() {
324 if src_f.layout.ty == dst_f.layout.ty {
325 memcpy_ty(bx, dst_f.llval, dst_f.align, src_f.llval, src_f.align,
326 src_f.layout, MemFlags::empty());
328 coerce_unsized_into(bx, src_f, dst_f);
332 _ => bug!("coerce_unsized_into: invalid coercion {:?} -> {:?}",
338 pub fn cast_shift_expr_rhs(
339 cx: &Builder<'_, 'll, '_>, op: hir::BinOpKind, lhs: &'ll Value, rhs: &'ll Value
341 cast_shift_rhs(op, lhs, rhs, |a, b| cx.trunc(a, b), |a, b| cx.zext(a, b))
344 fn cast_shift_rhs<'ll, F, G>(op: hir::BinOpKind,
350 where F: FnOnce(&'ll Value, &'ll Type) -> &'ll Value,
351 G: FnOnce(&'ll Value, &'ll Type) -> &'ll Value
353 // Shifts may have any size int on the rhs
355 let mut rhs_llty = val_ty(rhs);
356 let mut lhs_llty = val_ty(lhs);
357 if rhs_llty.kind() == TypeKind::Vector {
358 rhs_llty = rhs_llty.element_type()
360 if lhs_llty.kind() == TypeKind::Vector {
361 lhs_llty = lhs_llty.element_type()
363 let rhs_sz = rhs_llty.int_width();
364 let lhs_sz = lhs_llty.int_width();
367 } else if lhs_sz > rhs_sz {
368 // FIXME (#1877: If in the future shifting by negative
369 // values is no longer undefined then this is wrong.
379 /// Returns whether this session's target will use SEH-based unwinding.
381 /// This is only true for MSVC targets, and even then the 64-bit MSVC target
382 /// currently uses SEH-ish unwinding with DWARF info tables to the side (same as
383 /// 64-bit MinGW) instead of "full SEH".
384 pub fn wants_msvc_seh(sess: &Session) -> bool {
385 sess.target.target.options.is_like_msvc
388 pub fn call_assume(bx: &Builder<'_, 'll, '_>, val: &'ll Value) {
389 let assume_intrinsic = bx.cx.get_intrinsic("llvm.assume");
390 bx.call(assume_intrinsic, &[val], None);
393 pub fn from_immediate<'a, 'll: 'a, 'tcx: 'll,
395 Builder: BuilderMethods<'a, 'll, 'tcx, Value, BasicBlock>>(
398 ) -> &'ll Value where Value : ValueTrait {
399 if val_ty(val) == Type::i1(bx.cx()) {
400 bx.zext(val, Type::i8(bx.cx()))
407 bx: &Builder<'_, 'll, '_>,
409 layout: layout::TyLayout,
411 if let layout::Abi::Scalar(ref scalar) = layout.abi {
412 return to_immediate_scalar(bx, val, scalar);
417 pub fn to_immediate_scalar(
418 bx: &Builder<'_, 'll, '_>,
420 scalar: &layout::Scalar,
422 if scalar.is_bool() {
423 return bx.trunc(val, Type::i1(bx.cx));
428 pub fn call_memcpy<'a, 'll: 'a, 'tcx: 'll,
430 Builder: BuilderMethods<'a, 'll, 'tcx, Value, BasicBlock>>(
438 ) where Value : ValueTrait {
439 if flags.contains(MemFlags::NONTEMPORAL) {
440 // HACK(nox): This is inefficient but there is no nontemporal memcpy.
441 let val = bx.load(src, src_align);
442 let ptr = bx.pointercast(dst, val_ty(val).ptr_to());
443 bx.store_with_flags(val, ptr, dst_align, flags);
447 let src_ptr = bx.pointercast(src, Type::i8p(cx));
448 let dst_ptr = bx.pointercast(dst, Type::i8p(cx));
449 let size = bx.intcast(n_bytes, cx.isize_ty, false);
450 let volatile = flags.contains(MemFlags::VOLATILE);
451 bx.memcpy(dst_ptr, dst_align.abi(), src_ptr, src_align.abi(), size, volatile);
454 pub fn memcpy_ty<'a, 'll: 'a, 'tcx: 'll,
456 Builder: BuilderMethods<'a, 'll, 'tcx, Value, BasicBlock>>(
462 layout: TyLayout<'tcx>,
464 ) where Value : ValueTrait {
465 let size = layout.size.bytes();
470 call_memcpy(bx, dst, dst_align, src, src_align, C_usize(bx.cx(), size), flags);
474 bx: &Builder<'_, 'll, '_>,
476 fill_byte: &'ll Value,
481 let ptr_width = &bx.cx.sess().target.target.target_pointer_width;
482 let intrinsic_key = format!("llvm.memset.p0i8.i{}", ptr_width);
483 let llintrinsicfn = bx.cx.get_intrinsic(&intrinsic_key);
484 let volatile = C_bool(bx.cx, volatile);
485 bx.call(llintrinsicfn, &[ptr, fill_byte, size, align, volatile], None)
488 pub fn codegen_instance<'a, 'tcx>(cx: &CodegenCx<'a, 'tcx>, instance: Instance<'tcx>) {
489 let _s = if cx.sess().codegen_stats() {
490 let mut instance_name = String::new();
491 DefPathBasedNames::new(cx.tcx, true, true)
492 .push_def_path(instance.def_id(), &mut instance_name);
493 Some(StatRecorder::new(cx, instance_name))
498 // this is an info! to allow collecting monomorphization statistics
499 // and to allow finding the last function before LLVM aborts from
501 info!("codegen_instance({})", instance);
503 let sig = instance.fn_sig(cx.tcx);
504 let sig = cx.tcx.normalize_erasing_late_bound_regions(ty::ParamEnv::reveal_all(), &sig);
506 let lldecl = cx.instances.borrow().get(&instance).cloned().unwrap_or_else(||
507 bug!("Instance `{:?}` not already declared", instance));
509 cx.stats.borrow_mut().n_closures += 1;
511 let mir = cx.tcx.instance_mir(instance.def);
512 mir::codegen_mir(cx, lldecl, &mir, instance, sig);
515 pub fn set_link_section(llval: &Value, attrs: &CodegenFnAttrs) {
516 let sect = match attrs.link_section {
521 let buf = SmallCStr::new(§.as_str());
522 llvm::LLVMSetSection(llval, buf.as_ptr());
526 /// Create the `main` function which will initialize the rust runtime and call
527 /// users main function.
528 fn maybe_create_entry_wrapper(cx: &CodegenCx) {
529 let (main_def_id, span) = match *cx.sess().entry_fn.borrow() {
530 Some((id, span, _)) => {
531 (cx.tcx.hir.local_def_id(id), span)
536 let instance = Instance::mono(cx.tcx, main_def_id);
538 if !cx.codegen_unit.contains_item(&MonoItem::Fn(instance)) {
539 // We want to create the wrapper in the same codegen unit as Rust's main
544 let main_llfn = callee::get_fn(cx, instance);
546 let et = cx.sess().entry_fn.get().map(|e| e.2);
548 Some(EntryFnType::Main) => create_entry_fn(cx, span, main_llfn, main_def_id, true),
549 Some(EntryFnType::Start) => create_entry_fn(cx, span, main_llfn, main_def_id, false),
550 None => {} // Do nothing.
554 cx: &CodegenCx<'ll, '_>,
556 rust_main: &'ll Value,
557 rust_main_def_id: DefId,
558 use_start_lang_item: bool,
561 Type::func::<Value>(&[Type::c_int(cx), Type::i8p(cx).ptr_to()], Type::c_int(cx));
563 let main_ret_ty = cx.tcx.fn_sig(rust_main_def_id).output();
564 // Given that `main()` has no arguments,
565 // then its return type cannot have
566 // late-bound regions, since late-bound
567 // regions must appear in the argument
569 let main_ret_ty = cx.tcx.erase_regions(
570 &main_ret_ty.no_bound_vars().unwrap(),
573 if declare::get_defined_value(cx, "main").is_some() {
574 // FIXME: We should be smart and show a better diagnostic here.
575 cx.sess().struct_span_err(sp, "entry symbol `main` defined multiple times")
576 .help("did you use #[no_mangle] on `fn main`? Use #[start] instead")
578 cx.sess().abort_if_errors();
581 let llfn = declare::declare_cfn(cx, "main", llfty);
583 // `main` should respect same config for frame pointer elimination as rest of code
584 attributes::set_frame_pointer_elimination(cx, llfn);
585 attributes::apply_target_cpu_attr(cx, llfn);
587 let bx = Builder::new_block(cx, llfn, "top");
589 debuginfo::gdb::insert_reference_to_gdb_debug_scripts_section_global(&bx);
591 // Params from native main() used as args for rust start function
592 let param_argc = get_param(llfn, 0);
593 let param_argv = get_param(llfn, 1);
594 let arg_argc = bx.intcast(param_argc, cx.isize_ty, true);
595 let arg_argv = param_argv;
597 let (start_fn, args) = if use_start_lang_item {
598 let start_def_id = cx.tcx.require_lang_item(StartFnLangItem);
599 let start_fn = callee::resolve_and_get_fn(
602 cx.tcx.intern_substs(&[main_ret_ty.into()]),
604 (start_fn, vec![bx.pointercast(rust_main, Type::i8p(cx).ptr_to()),
607 debug!("using user-defined start fn");
608 (rust_main, vec![arg_argc, arg_argv])
611 let result = bx.call(start_fn, &args, None);
612 bx.ret(bx.intcast(result, Type::c_int(cx), true));
616 fn write_metadata<'a, 'gcx>(tcx: TyCtxt<'a, 'gcx, 'gcx>,
617 llvm_module: &ModuleLlvm)
620 use flate2::Compression;
621 use flate2::write::DeflateEncoder;
623 let (metadata_llcx, metadata_llmod) = (&*llvm_module.llcx, llvm_module.llmod());
625 #[derive(PartialEq, Eq, PartialOrd, Ord)]
632 let kind = tcx.sess.crate_types.borrow().iter().map(|ty| {
634 config::CrateType::Executable |
635 config::CrateType::Staticlib |
636 config::CrateType::Cdylib => MetadataKind::None,
638 config::CrateType::Rlib => MetadataKind::Uncompressed,
640 config::CrateType::Dylib |
641 config::CrateType::ProcMacro => MetadataKind::Compressed,
643 }).max().unwrap_or(MetadataKind::None);
645 if kind == MetadataKind::None {
646 return EncodedMetadata::new();
649 let metadata = tcx.encode_metadata();
650 if kind == MetadataKind::Uncompressed {
654 assert!(kind == MetadataKind::Compressed);
655 let mut compressed = tcx.metadata_encoding_version();
656 DeflateEncoder::new(&mut compressed, Compression::fast())
657 .write_all(&metadata.raw_data).unwrap();
659 let llmeta = C_bytes_in_context(metadata_llcx, &compressed);
660 let llconst = C_struct_in_context(metadata_llcx, &[llmeta], false);
661 let name = exported_symbols::metadata_symbol_name(tcx);
662 let buf = CString::new(name).unwrap();
663 let llglobal = unsafe {
664 llvm::LLVMAddGlobal(metadata_llmod, val_ty(llconst), buf.as_ptr())
667 llvm::LLVMSetInitializer(llglobal, llconst);
668 let section_name = metadata::metadata_section_name(&tcx.sess.target.target);
669 let name = SmallCStr::new(section_name);
670 llvm::LLVMSetSection(llglobal, name.as_ptr());
672 // Also generate a .section directive to force no
673 // flags, at least for ELF outputs, so that the
674 // metadata doesn't get loaded into memory.
675 let directive = format!(".section {}", section_name);
676 let directive = CString::new(directive).unwrap();
677 llvm::LLVMSetModuleInlineAsm(metadata_llmod, directive.as_ptr())
682 pub struct ValueIter<'ll> {
683 cur: Option<&'ll Value>,
684 step: unsafe extern "C" fn(&'ll Value) -> Option<&'ll Value>,
687 impl Iterator for ValueIter<'ll> {
688 type Item = &'ll Value;
690 fn next(&mut self) -> Option<&'ll Value> {
692 if let Some(old) = old {
693 self.cur = unsafe { (self.step)(old) };
699 pub fn iter_globals(llmod: &'ll llvm::Module) -> ValueIter<'ll> {
702 cur: llvm::LLVMGetFirstGlobal(llmod),
703 step: llvm::LLVMGetNextGlobal,
708 fn determine_cgu_reuse<'a, 'tcx>(tcx: TyCtxt<'a, 'tcx, 'tcx>,
709 cgu: &CodegenUnit<'tcx>)
711 if !tcx.dep_graph.is_fully_enabled() {
715 let work_product_id = &cgu.work_product_id();
716 if tcx.dep_graph.previous_work_product(work_product_id).is_none() {
717 // We don't have anything cached for this CGU. This can happen
718 // if the CGU did not exist in the previous session.
722 // Try to mark the CGU as green. If it we can do so, it means that nothing
723 // affecting the LLVM module has changed and we can re-use a cached version.
724 // If we compile with any kind of LTO, this means we can re-use the bitcode
725 // of the Pre-LTO stage (possibly also the Post-LTO version but we'll only
726 // know that later). If we are not doing LTO, there is only one optimized
727 // version of each module, so we re-use that.
728 let dep_node = cgu.codegen_dep_node(tcx);
729 assert!(!tcx.dep_graph.dep_node_exists(&dep_node),
730 "CompileCodegenUnit dep-node for CGU `{}` already exists before marking.",
733 if tcx.dep_graph.try_mark_green(tcx, &dep_node).is_some() {
734 // We can re-use either the pre- or the post-thinlto state
735 if tcx.sess.lto() != Lto::No {
745 pub fn codegen_crate<'a, 'tcx>(tcx: TyCtxt<'a, 'tcx, 'tcx>,
746 rx: mpsc::Receiver<Box<dyn Any + Send>>)
749 check_for_rustc_errors_attr(tcx);
751 let cgu_name_builder = &mut CodegenUnitNameBuilder::new(tcx);
753 // Codegen the metadata.
754 tcx.sess.profiler(|p| p.start_activity(ProfileCategory::Codegen));
756 let metadata_cgu_name = cgu_name_builder.build_cgu_name(LOCAL_CRATE,
758 Some("metadata")).as_str()
760 let metadata_llvm_module = ModuleLlvm::new(tcx.sess, &metadata_cgu_name);
761 let metadata = time(tcx.sess, "write metadata", || {
762 write_metadata(tcx, &metadata_llvm_module)
764 tcx.sess.profiler(|p| p.end_activity(ProfileCategory::Codegen));
766 let metadata_module = ModuleCodegen {
767 name: metadata_cgu_name,
768 module_llvm: metadata_llvm_module,
769 kind: ModuleKind::Metadata,
772 let time_graph = if tcx.sess.opts.debugging_opts.codegen_time_graph {
773 Some(time_graph::TimeGraph::new())
778 // Skip crate items and just output metadata in -Z no-codegen mode.
779 if tcx.sess.opts.debugging_opts.no_codegen ||
780 !tcx.sess.opts.output_types.should_codegen() {
781 let ongoing_codegen = write::start_async_codegen(
788 ongoing_codegen.submit_pre_codegened_module_to_llvm(tcx, metadata_module);
789 ongoing_codegen.codegen_finished(tcx);
791 assert_and_save_dep_graph(tcx);
793 ongoing_codegen.check_for_errors(tcx.sess);
795 return ongoing_codegen;
798 // Run the monomorphization collector and partition the collected items into
800 let codegen_units = tcx.collect_and_partition_mono_items(LOCAL_CRATE).1;
801 let codegen_units = (*codegen_units).clone();
803 // Force all codegen_unit queries so they are already either red or green
804 // when compile_codegen_unit accesses them. We are not able to re-execute
805 // the codegen_unit query from just the DepNode, so an unknown color would
806 // lead to having to re-execute compile_codegen_unit, possibly
808 if tcx.dep_graph.is_fully_enabled() {
809 for cgu in &codegen_units {
810 tcx.codegen_unit(cgu.name().clone());
814 let ongoing_codegen = write::start_async_codegen(
819 codegen_units.len());
820 let ongoing_codegen = AbortCodegenOnDrop(Some(ongoing_codegen));
822 // Codegen an allocator shim, if necessary.
824 // If the crate doesn't have an `allocator_kind` set then there's definitely
825 // no shim to generate. Otherwise we also check our dependency graph for all
826 // our output crate types. If anything there looks like its a `Dynamic`
827 // linkage, then it's already got an allocator shim and we'll be using that
828 // one instead. If nothing exists then it's our job to generate the
830 let any_dynamic_crate = tcx.sess.dependency_formats.borrow()
833 use rustc::middle::dependency_format::Linkage;
834 list.iter().any(|&linkage| linkage == Linkage::Dynamic)
836 let allocator_module = if any_dynamic_crate {
838 } else if let Some(kind) = *tcx.sess.allocator_kind.get() {
839 let llmod_id = cgu_name_builder.build_cgu_name(LOCAL_CRATE,
841 Some("allocator")).as_str()
843 let modules = ModuleLlvm::new(tcx.sess, &llmod_id);
844 time(tcx.sess, "write allocator module", || {
846 allocator::codegen(tcx, &modules, kind)
852 module_llvm: modules,
853 kind: ModuleKind::Allocator,
859 if let Some(allocator_module) = allocator_module {
860 ongoing_codegen.submit_pre_codegened_module_to_llvm(tcx, allocator_module);
863 ongoing_codegen.submit_pre_codegened_module_to_llvm(tcx, metadata_module);
865 // We sort the codegen units by size. This way we can schedule work for LLVM
866 // a bit more efficiently.
867 let codegen_units = {
868 let mut codegen_units = codegen_units;
869 codegen_units.sort_by_cached_key(|cgu| cmp::Reverse(cgu.size_estimate()));
873 let mut total_codegen_time = Duration::new(0, 0);
874 let mut all_stats = Stats::default();
876 for cgu in codegen_units.into_iter() {
877 ongoing_codegen.wait_for_signal_to_codegen_item();
878 ongoing_codegen.check_for_errors(tcx.sess);
880 let cgu_reuse = determine_cgu_reuse(tcx, &cgu);
881 tcx.sess.cgu_reuse_tracker.set_actual_reuse(&cgu.name().as_str(), cgu_reuse);
885 let _timing_guard = time_graph.as_ref().map(|time_graph| {
886 time_graph.start(write::CODEGEN_WORKER_TIMELINE,
887 write::CODEGEN_WORK_PACKAGE_KIND,
888 &format!("codegen {}", cgu.name()))
890 let start_time = Instant::now();
891 let stats = compile_codegen_unit(tcx, *cgu.name());
892 all_stats.extend(stats);
893 total_codegen_time += start_time.elapsed();
896 CguReuse::PreLto => {
897 write::submit_pre_lto_module_to_llvm(tcx, CachedModuleCodegen {
898 name: cgu.name().to_string(),
899 source: cgu.work_product(tcx),
903 CguReuse::PostLto => {
904 write::submit_post_lto_module_to_llvm(tcx, CachedModuleCodegen {
905 name: cgu.name().to_string(),
906 source: cgu.work_product(tcx),
913 ongoing_codegen.codegen_finished(tcx);
915 // Since the main thread is sometimes blocked during codegen, we keep track
916 // -Ztime-passes output manually.
917 print_time_passes_entry(tcx.sess.time_passes(),
918 "codegen to LLVM IR",
921 rustc_incremental::assert_module_sources::assert_module_sources(tcx);
923 symbol_names_test::report_symbol_names(tcx);
925 if tcx.sess.codegen_stats() {
926 println!("--- codegen stats ---");
927 println!("n_glues_created: {}", all_stats.n_glues_created);
928 println!("n_null_glues: {}", all_stats.n_null_glues);
929 println!("n_real_glues: {}", all_stats.n_real_glues);
931 println!("n_fns: {}", all_stats.n_fns);
932 println!("n_inlines: {}", all_stats.n_inlines);
933 println!("n_closures: {}", all_stats.n_closures);
934 println!("fn stats:");
935 all_stats.fn_stats.sort_by_key(|&(_, insns)| insns);
936 for &(ref name, insns) in all_stats.fn_stats.iter() {
937 println!("{} insns, {}", insns, *name);
941 if tcx.sess.count_llvm_insns() {
942 for (k, v) in all_stats.llvm_insns.iter() {
943 println!("{:7} {}", *v, *k);
947 ongoing_codegen.check_for_errors(tcx.sess);
949 assert_and_save_dep_graph(tcx);
950 ongoing_codegen.into_inner()
953 /// A curious wrapper structure whose only purpose is to call `codegen_aborted`
954 /// when it's dropped abnormally.
956 /// In the process of working on rust-lang/rust#55238 a mysterious segfault was
957 /// stumbled upon. The segfault was never reproduced locally, but it was
958 /// suspected to be related to the fact that codegen worker threads were
959 /// sticking around by the time the main thread was exiting, causing issues.
961 /// This structure is an attempt to fix that issue where the `codegen_aborted`
962 /// message will block until all workers have finished. This should ensure that
963 /// even if the main codegen thread panics we'll wait for pending work to
964 /// complete before returning from the main thread, hopefully avoiding
967 /// If you see this comment in the code, then it means that this workaround
968 /// worked! We may yet one day track down the mysterious cause of that
970 struct AbortCodegenOnDrop(Option<OngoingCodegen>);
972 impl AbortCodegenOnDrop {
973 fn into_inner(mut self) -> OngoingCodegen {
974 self.0.take().unwrap()
978 impl Deref for AbortCodegenOnDrop {
979 type Target = OngoingCodegen;
981 fn deref(&self) -> &OngoingCodegen {
982 self.0.as_ref().unwrap()
986 impl DerefMut for AbortCodegenOnDrop {
987 fn deref_mut(&mut self) -> &mut OngoingCodegen {
988 self.0.as_mut().unwrap()
992 impl Drop for AbortCodegenOnDrop {
994 if let Some(codegen) = self.0.take() {
995 codegen.codegen_aborted();
1000 fn assert_and_save_dep_graph<'a, 'tcx>(tcx: TyCtxt<'a, 'tcx, 'tcx>) {
1003 || rustc_incremental::assert_dep_graph(tcx));
1006 "serialize dep graph",
1007 || rustc_incremental::save_dep_graph(tcx));
1011 pub fn new(tcx: TyCtxt) -> CrateInfo {
1012 let mut info = CrateInfo {
1013 panic_runtime: None,
1014 compiler_builtins: None,
1015 profiler_runtime: None,
1016 sanitizer_runtime: None,
1017 is_no_builtins: Default::default(),
1018 native_libraries: Default::default(),
1019 used_libraries: tcx.native_libraries(LOCAL_CRATE),
1020 link_args: tcx.link_args(LOCAL_CRATE),
1021 crate_name: Default::default(),
1022 used_crates_dynamic: cstore::used_crates(tcx, LinkagePreference::RequireDynamic),
1023 used_crates_static: cstore::used_crates(tcx, LinkagePreference::RequireStatic),
1024 used_crate_source: Default::default(),
1025 wasm_imports: Default::default(),
1026 lang_item_to_crate: Default::default(),
1027 missing_lang_items: Default::default(),
1029 let lang_items = tcx.lang_items();
1031 let load_wasm_items = tcx.sess.crate_types.borrow()
1033 .any(|c| *c != config::CrateType::Rlib) &&
1034 tcx.sess.opts.target_triple.triple() == "wasm32-unknown-unknown";
1036 if load_wasm_items {
1037 info.load_wasm_imports(tcx, LOCAL_CRATE);
1040 let crates = tcx.crates();
1042 let n_crates = crates.len();
1043 info.native_libraries.reserve(n_crates);
1044 info.crate_name.reserve(n_crates);
1045 info.used_crate_source.reserve(n_crates);
1046 info.missing_lang_items.reserve(n_crates);
1048 for &cnum in crates.iter() {
1049 info.native_libraries.insert(cnum, tcx.native_libraries(cnum));
1050 info.crate_name.insert(cnum, tcx.crate_name(cnum).to_string());
1051 info.used_crate_source.insert(cnum, tcx.used_crate_source(cnum));
1052 if tcx.is_panic_runtime(cnum) {
1053 info.panic_runtime = Some(cnum);
1055 if tcx.is_compiler_builtins(cnum) {
1056 info.compiler_builtins = Some(cnum);
1058 if tcx.is_profiler_runtime(cnum) {
1059 info.profiler_runtime = Some(cnum);
1061 if tcx.is_sanitizer_runtime(cnum) {
1062 info.sanitizer_runtime = Some(cnum);
1064 if tcx.is_no_builtins(cnum) {
1065 info.is_no_builtins.insert(cnum);
1067 if load_wasm_items {
1068 info.load_wasm_imports(tcx, cnum);
1070 let missing = tcx.missing_lang_items(cnum);
1071 for &item in missing.iter() {
1072 if let Ok(id) = lang_items.require(item) {
1073 info.lang_item_to_crate.insert(item, id.krate);
1077 // No need to look for lang items that are whitelisted and don't
1078 // actually need to exist.
1079 let missing = missing.iter()
1081 .filter(|&l| !weak_lang_items::whitelisted(tcx, l))
1083 info.missing_lang_items.insert(cnum, missing);
1089 fn load_wasm_imports(&mut self, tcx: TyCtxt, cnum: CrateNum) {
1090 self.wasm_imports.extend(tcx.wasm_import_module_map(cnum).iter().map(|(&id, module)| {
1091 let instance = Instance::mono(tcx, id);
1092 let import_name = tcx.symbol_name(instance);
1094 (import_name.to_string(), module.clone())
1099 fn compile_codegen_unit<'a, 'tcx>(tcx: TyCtxt<'a, 'tcx, 'tcx>,
1100 cgu_name: InternedString)
1102 let start_time = Instant::now();
1104 let dep_node = tcx.codegen_unit(cgu_name).codegen_dep_node(tcx);
1105 let ((stats, module), _) = tcx.dep_graph.with_task(dep_node,
1109 let time_to_codegen = start_time.elapsed();
1111 // We assume that the cost to run LLVM on a CGU is proportional to
1112 // the time we needed for codegenning it.
1113 let cost = time_to_codegen.as_secs() * 1_000_000_000 +
1114 time_to_codegen.subsec_nanos() as u64;
1116 write::submit_codegened_module_to_llvm(tcx,
1121 fn module_codegen<'a, 'tcx>(
1122 tcx: TyCtxt<'a, 'tcx, 'tcx>,
1123 cgu_name: InternedString)
1124 -> (Stats, ModuleCodegen)
1126 let cgu = tcx.codegen_unit(cgu_name);
1128 // Instantiate monomorphizations without filling out definitions yet...
1129 let llvm_module = ModuleLlvm::new(tcx.sess, &cgu_name.as_str());
1131 let cx = CodegenCx::new(tcx, cgu, &llvm_module);
1132 let mono_items = cx.codegen_unit
1133 .items_in_deterministic_order(cx.tcx);
1134 for &(mono_item, (linkage, visibility)) in &mono_items {
1135 mono_item.predefine(&cx, linkage, visibility);
1138 // ... and now that we have everything pre-defined, fill out those definitions.
1139 for &(mono_item, _) in &mono_items {
1140 mono_item.define(&cx);
1143 // If this codegen unit contains the main function, also create the
1145 maybe_create_entry_wrapper(&cx);
1147 // Run replace-all-uses-with for statics that need it
1148 for &(old_g, new_g) in cx.statics_to_rauw.borrow().iter() {
1150 let bitcast = llvm::LLVMConstPointerCast(new_g, val_ty(old_g));
1151 llvm::LLVMReplaceAllUsesWith(old_g, bitcast);
1152 llvm::LLVMDeleteGlobal(old_g);
1156 // Create the llvm.used variable
1157 // This variable has type [N x i8*] and is stored in the llvm.metadata section
1158 if !cx.used_statics.borrow().is_empty() {
1159 let name = const_cstr!("llvm.used");
1160 let section = const_cstr!("llvm.metadata");
1161 let array = C_array(Type::i8(&cx).ptr_to(), &*cx.used_statics.borrow());
1164 let g = llvm::LLVMAddGlobal(cx.llmod,
1167 llvm::LLVMSetInitializer(g, array);
1168 llvm::LLVMRustSetLinkage(g, llvm::Linkage::AppendingLinkage);
1169 llvm::LLVMSetSection(g, section.as_ptr());
1173 // Finalize debuginfo
1174 if cx.sess().opts.debuginfo != DebugInfo::None {
1175 debuginfo::finalize(&cx);
1178 cx.stats.into_inner()
1181 (stats, ModuleCodegen {
1182 name: cgu_name.to_string(),
1183 module_llvm: llvm_module,
1184 kind: ModuleKind::Regular,
1189 pub fn provide_both(providers: &mut Providers) {
1190 providers.dllimport_foreign_items = |tcx, krate| {
1191 let module_map = tcx.foreign_modules(krate);
1192 let module_map = module_map.iter()
1193 .map(|lib| (lib.def_id, lib))
1194 .collect::<FxHashMap<_, _>>();
1196 let dllimports = tcx.native_libraries(krate)
1199 if lib.kind != cstore::NativeLibraryKind::NativeUnknown {
1202 let cfg = match lib.cfg {
1203 Some(ref cfg) => cfg,
1204 None => return true,
1206 attr::cfg_matches(cfg, &tcx.sess.parse_sess, None)
1208 .filter_map(|lib| lib.foreign_module)
1209 .map(|id| &module_map[&id])
1210 .flat_map(|module| module.foreign_items.iter().cloned())
1212 Lrc::new(dllimports)
1215 providers.is_dllimport_foreign_item = |tcx, def_id| {
1216 tcx.dllimport_foreign_items(def_id.krate).contains(&def_id)
1220 pub fn linkage_to_llvm(linkage: Linkage) -> llvm::Linkage {
1222 Linkage::External => llvm::Linkage::ExternalLinkage,
1223 Linkage::AvailableExternally => llvm::Linkage::AvailableExternallyLinkage,
1224 Linkage::LinkOnceAny => llvm::Linkage::LinkOnceAnyLinkage,
1225 Linkage::LinkOnceODR => llvm::Linkage::LinkOnceODRLinkage,
1226 Linkage::WeakAny => llvm::Linkage::WeakAnyLinkage,
1227 Linkage::WeakODR => llvm::Linkage::WeakODRLinkage,
1228 Linkage::Appending => llvm::Linkage::AppendingLinkage,
1229 Linkage::Internal => llvm::Linkage::InternalLinkage,
1230 Linkage::Private => llvm::Linkage::PrivateLinkage,
1231 Linkage::ExternalWeak => llvm::Linkage::ExternalWeakLinkage,
1232 Linkage::Common => llvm::Linkage::CommonLinkage,
1236 pub fn visibility_to_llvm(linkage: Visibility) -> llvm::Visibility {
1238 Visibility::Default => llvm::Visibility::Default,
1239 Visibility::Hidden => llvm::Visibility::Hidden,
1240 Visibility::Protected => llvm::Visibility::Protected,
1244 // FIXME(mw): Anything that is produced via DepGraph::with_task() must implement
1245 // the HashStable trait. Normally DepGraph::with_task() calls are
1246 // hidden behind queries, but CGU creation is a special case in two
1247 // ways: (1) it's not a query and (2) CGU are output nodes, so their
1248 // Fingerprints are not actually needed. It remains to be clarified
1249 // how exactly this case will be handled in the red/green system but
1250 // for now we content ourselves with providing a no-op HashStable
1251 // implementation for CGUs.
1252 mod temp_stable_hash_impls {
1253 use rustc_data_structures::stable_hasher::{StableHasherResult, StableHasher,
1257 impl<HCX> HashStable<HCX> for ModuleCodegen {
1258 fn hash_stable<W: StableHasherResult>(&self,
1260 _: &mut StableHasher<W>) {