]>
Commit | Line | Data |
---|---|---|
a2a8927a | 1 | use crate::back::metadata::create_compressed_metadata_file; |
60c5eb7d | 2 | use crate::back::write::{ |
f9f354fc XL |
3 | compute_per_cgu_lto_type, start_async_codegen, submit_codegened_module_to_llvm, |
4 | submit_post_lto_module_to_llvm, submit_pre_lto_module_to_llvm, ComputedLtoType, OngoingCodegen, | |
60c5eb7d | 5 | }; |
dfeec247 | 6 | use crate::common::{IntPredicate, RealPredicate, TypeKind}; |
60c5eb7d XL |
7 | use crate::meth; |
8 | use crate::mir; | |
9 | use crate::mir::operand::OperandValue; | |
10 | use crate::mir::place::PlaceRef; | |
11 | use crate::traits::*; | |
a2a8927a | 12 | use crate::{CachedModuleCodegen, CompiledModule, CrateInfo, MemFlags, ModuleCodegen, ModuleKind}; |
a1dfa0c6 | 13 | |
74b04a01 | 14 | use rustc_attr as attr; |
dfeec247 | 15 | use rustc_data_structures::fx::FxHashMap; |
5869c6ff | 16 | use rustc_data_structures::profiling::{get_resident_set_size, print_time_passes_entry}; |
5e7ed085 FG |
17 | |
18 | #[cfg(parallel_compiler)] | |
5869c6ff | 19 | use rustc_data_structures::sync::{par_iter, ParallelIterator}; |
dfeec247 | 20 | use rustc_hir as hir; |
cdc7bbd5 | 21 | use rustc_hir::def_id::{DefId, LOCAL_CRATE}; |
3dfed10e | 22 | use rustc_hir::lang_items::LangItem; |
e74abb32 | 23 | use rustc_index::vec::Idx; |
c295e0f8 | 24 | use rustc_metadata::EncodedMetadata; |
ba9703b0 | 25 | use rustc_middle::middle::codegen_fn_attrs::CodegenFnAttrs; |
a2a8927a | 26 | use rustc_middle::middle::exported_symbols; |
ba9703b0 XL |
27 | use rustc_middle::middle::lang_items; |
28 | use rustc_middle::mir::mono::{CodegenUnit, CodegenUnitNameBuilder, MonoItem}; | |
c295e0f8 | 29 | use rustc_middle::ty::layout::{HasTyCtxt, LayoutOf, TyAndLayout}; |
ba9703b0 XL |
30 | use rustc_middle::ty::query::Providers; |
31 | use rustc_middle::ty::{self, Instance, Ty, TyCtxt}; | |
dfeec247 | 32 | use rustc_session::cgu_reuse_tracker::CguReuse; |
a2a8927a | 33 | use rustc_session::config::{self, EntryFnType, OutputType}; |
ba9703b0 | 34 | use rustc_session::Session; |
17df50a5 | 35 | use rustc_span::symbol::sym; |
c295e0f8 | 36 | use rustc_target::abi::{Align, VariantIdx}; |
a1dfa0c6 | 37 | |
94222f64 | 38 | use std::convert::TryFrom; |
a1dfa0c6 | 39 | use std::ops::{Deref, DerefMut}; |
dfeec247 | 40 | use std::time::{Duration, Instant}; |
a1dfa0c6 | 41 | |
5869c6ff XL |
42 | use itertools::Itertools; |
43 | ||
dfeec247 | 44 | pub fn bin_op_to_icmp_predicate(op: hir::BinOpKind, signed: bool) -> IntPredicate { |
a1dfa0c6 XL |
45 | match op { |
46 | hir::BinOpKind::Eq => IntPredicate::IntEQ, | |
47 | hir::BinOpKind::Ne => IntPredicate::IntNE, | |
dfeec247 XL |
48 | hir::BinOpKind::Lt => { |
49 | if signed { | |
50 | IntPredicate::IntSLT | |
51 | } else { | |
52 | IntPredicate::IntULT | |
53 | } | |
a1dfa0c6 | 54 | } |
dfeec247 XL |
55 | hir::BinOpKind::Le => { |
56 | if signed { | |
57 | IntPredicate::IntSLE | |
58 | } else { | |
59 | IntPredicate::IntULE | |
60 | } | |
61 | } | |
62 | hir::BinOpKind::Gt => { | |
63 | if signed { | |
64 | IntPredicate::IntSGT | |
65 | } else { | |
66 | IntPredicate::IntUGT | |
67 | } | |
68 | } | |
69 | hir::BinOpKind::Ge => { | |
70 | if signed { | |
71 | IntPredicate::IntSGE | |
72 | } else { | |
73 | IntPredicate::IntUGE | |
74 | } | |
75 | } | |
76 | op => bug!( | |
77 | "comparison_op_to_icmp_predicate: expected comparison operator, \ | |
78 | found {:?}", | |
79 | op | |
80 | ), | |
a1dfa0c6 XL |
81 | } |
82 | } | |
83 | ||
84 | pub fn bin_op_to_fcmp_predicate(op: hir::BinOpKind) -> RealPredicate { | |
85 | match op { | |
86 | hir::BinOpKind::Eq => RealPredicate::RealOEQ, | |
87 | hir::BinOpKind::Ne => RealPredicate::RealUNE, | |
88 | hir::BinOpKind::Lt => RealPredicate::RealOLT, | |
89 | hir::BinOpKind::Le => RealPredicate::RealOLE, | |
90 | hir::BinOpKind::Gt => RealPredicate::RealOGT, | |
91 | hir::BinOpKind::Ge => RealPredicate::RealOGE, | |
92 | op => { | |
dfeec247 XL |
93 | bug!( |
94 | "comparison_op_to_fcmp_predicate: expected comparison operator, \ | |
95 | found {:?}", | |
96 | op | |
97 | ); | |
a1dfa0c6 XL |
98 | } |
99 | } | |
100 | } | |
101 | ||
dc9dc135 | 102 | pub fn compare_simd_types<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>( |
a1dfa0c6 XL |
103 | bx: &mut Bx, |
104 | lhs: Bx::Value, | |
105 | rhs: Bx::Value, | |
106 | t: Ty<'tcx>, | |
107 | ret_ty: Bx::Type, | |
dc9dc135 | 108 | op: hir::BinOpKind, |
a1dfa0c6 | 109 | ) -> Bx::Value { |
1b1a35ee | 110 | let signed = match t.kind() { |
a1dfa0c6 XL |
111 | ty::Float(_) => { |
112 | let cmp = bin_op_to_fcmp_predicate(op); | |
113 | let cmp = bx.fcmp(cmp, lhs, rhs); | |
114 | return bx.sext(cmp, ret_ty); | |
dfeec247 | 115 | } |
a1dfa0c6 XL |
116 | ty::Uint(_) => false, |
117 | ty::Int(_) => true, | |
118 | _ => bug!("compare_simd_types: invalid SIMD type"), | |
119 | }; | |
120 | ||
121 | let cmp = bin_op_to_icmp_predicate(op, signed); | |
122 | let cmp = bx.icmp(cmp, lhs, rhs); | |
123 | // LLVM outputs an `< size x i1 >`, so we need to perform a sign extension | |
124 | // to get the correctly sized type. This will compile to a single instruction | |
125 | // once the IR is converted to assembly if the SIMD instruction is supported | |
126 | // by the target architecture. | |
127 | bx.sext(cmp, ret_ty) | |
128 | } | |
129 | ||
9fa01778 | 130 | /// Retrieves the information we are losing (making dynamic) in an unsizing |
a1dfa0c6 XL |
131 | /// adjustment. |
132 | /// | |
60c5eb7d XL |
133 | /// The `old_info` argument is a bit odd. It is intended for use in an upcast, |
134 | /// where the new vtable for an object will be derived from the old one. | |
94222f64 XL |
135 | pub fn unsized_info<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>( |
136 | bx: &mut Bx, | |
a1dfa0c6 XL |
137 | source: Ty<'tcx>, |
138 | target: Ty<'tcx>, | |
94222f64 XL |
139 | old_info: Option<Bx::Value>, |
140 | ) -> Bx::Value { | |
141 | let cx = bx.cx(); | |
416331ca | 142 | let (source, target) = |
94222f64 | 143 | cx.tcx().struct_lockstep_tails_erasing_lifetimes(source, target, bx.param_env()); |
1b1a35ee | 144 | match (source.kind(), target.kind()) { |
a1dfa0c6 | 145 | (&ty::Array(_, len), &ty::Slice(_)) => { |
416331ca | 146 | cx.const_usize(len.eval_usize(cx.tcx(), ty::ParamEnv::reveal_all())) |
a1dfa0c6 | 147 | } |
94222f64 XL |
148 | (&ty::Dynamic(ref data_a, ..), &ty::Dynamic(ref data_b, ..)) => { |
149 | let old_info = | |
150 | old_info.expect("unsized_info: missing old info for trait upcasting coercion"); | |
151 | if data_a.principal_def_id() == data_b.principal_def_id() { | |
152 | return old_info; | |
153 | } | |
154 | ||
155 | // trait upcasting coercion | |
156 | ||
157 | let vptr_entry_idx = | |
158 | cx.tcx().vtable_trait_upcasting_coercion_new_vptr_slot((source, target)); | |
159 | ||
160 | if let Some(entry_idx) = vptr_entry_idx { | |
161 | let ptr_ty = cx.type_i8p(); | |
162 | let ptr_align = cx.tcx().data_layout.pointer_align.abi; | |
163 | let llvtable = bx.pointercast(old_info, bx.type_ptr_to(ptr_ty)); | |
164 | let gep = bx.inbounds_gep( | |
165 | ptr_ty, | |
166 | llvtable, | |
167 | &[bx.const_usize(u64::try_from(entry_idx).unwrap())], | |
168 | ); | |
169 | let new_vptr = bx.load(ptr_ty, gep, ptr_align); | |
170 | bx.nonnull_metadata(new_vptr); | |
171 | // Vtable loads are invariant. | |
172 | bx.set_invariant_load(new_vptr); | |
173 | new_vptr | |
174 | } else { | |
175 | old_info | |
176 | } | |
a1dfa0c6 XL |
177 | } |
178 | (_, &ty::Dynamic(ref data, ..)) => { | |
94222f64 XL |
179 | let vtable_ptr_ty = cx.scalar_pair_element_backend_type( |
180 | cx.layout_of(cx.tcx().mk_mut_ptr(target)), | |
181 | 1, | |
182 | true, | |
183 | ); | |
184 | cx.const_ptrcast(meth::get_vtable(cx, source, data.principal()), vtable_ptr_ty) | |
a1dfa0c6 | 185 | } |
dfeec247 | 186 | _ => bug!("unsized_info: invalid unsizing {:?} -> {:?}", source, target), |
a1dfa0c6 XL |
187 | } |
188 | } | |
189 | ||
94222f64 XL |
190 | /// Coerces `src` to `dst_ty`. `src_ty` must be a pointer. |
191 | pub fn unsize_ptr<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>( | |
a1dfa0c6 XL |
192 | bx: &mut Bx, |
193 | src: Bx::Value, | |
194 | src_ty: Ty<'tcx>, | |
dc9dc135 | 195 | dst_ty: Ty<'tcx>, |
94222f64 | 196 | old_info: Option<Bx::Value>, |
a1dfa0c6 | 197 | ) -> (Bx::Value, Bx::Value) { |
94222f64 | 198 | debug!("unsize_ptr: {:?} => {:?}", src_ty, dst_ty); |
1b1a35ee | 199 | match (src_ty.kind(), dst_ty.kind()) { |
ba9703b0 | 200 | (&ty::Ref(_, a, _), &ty::Ref(_, b, _) | &ty::RawPtr(ty::TypeAndMut { ty: b, .. })) |
dfeec247 | 201 | | (&ty::RawPtr(ty::TypeAndMut { ty: a, .. }), &ty::RawPtr(ty::TypeAndMut { ty: b, .. })) => { |
94222f64 | 202 | assert_eq!(bx.cx().type_is_sized(a), old_info.is_none()); |
a1dfa0c6 | 203 | let ptr_ty = bx.cx().type_ptr_to(bx.cx().backend_type(bx.cx().layout_of(b))); |
94222f64 | 204 | (bx.pointercast(src, ptr_ty), unsized_info(bx, a, b, old_info)) |
a1dfa0c6 | 205 | } |
a1dfa0c6 XL |
206 | (&ty::Adt(def_a, _), &ty::Adt(def_b, _)) => { |
207 | assert_eq!(def_a, def_b); | |
a1dfa0c6 XL |
208 | let src_layout = bx.cx().layout_of(src_ty); |
209 | let dst_layout = bx.cx().layout_of(dst_ty); | |
94222f64 XL |
210 | if src_ty == dst_ty { |
211 | return (src, old_info.unwrap()); | |
212 | } | |
a1dfa0c6 XL |
213 | let mut result = None; |
214 | for i in 0..src_layout.fields.count() { | |
215 | let src_f = src_layout.field(bx.cx(), i); | |
216 | assert_eq!(src_layout.fields.offset(i).bytes(), 0); | |
217 | assert_eq!(dst_layout.fields.offset(i).bytes(), 0); | |
218 | if src_f.is_zst() { | |
219 | continue; | |
220 | } | |
221 | assert_eq!(src_layout.size, src_f.size); | |
222 | ||
223 | let dst_f = dst_layout.field(bx.cx(), i); | |
224 | assert_ne!(src_f.ty, dst_f.ty); | |
225 | assert_eq!(result, None); | |
94222f64 | 226 | result = Some(unsize_ptr(bx, src, src_f.ty, dst_f.ty, old_info)); |
a1dfa0c6 XL |
227 | } |
228 | let (lldata, llextra) = result.unwrap(); | |
94222f64 XL |
229 | let lldata_ty = bx.cx().scalar_pair_element_backend_type(dst_layout, 0, true); |
230 | let llextra_ty = bx.cx().scalar_pair_element_backend_type(dst_layout, 1, true); | |
a1dfa0c6 | 231 | // HACK(eddyb) have to bitcast pointers until LLVM removes pointee types. |
94222f64 | 232 | (bx.bitcast(lldata, lldata_ty), bx.bitcast(llextra, llextra_ty)) |
a1dfa0c6 | 233 | } |
94222f64 | 234 | _ => bug!("unsize_ptr: called on bad types"), |
a1dfa0c6 XL |
235 | } |
236 | } | |
237 | ||
60c5eb7d XL |
238 | /// Coerces `src`, which is a reference to a value of type `src_ty`, |
239 | /// to a value of type `dst_ty`, and stores the result in `dst`. | |
dc9dc135 | 240 | pub fn coerce_unsized_into<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>( |
a1dfa0c6 XL |
241 | bx: &mut Bx, |
242 | src: PlaceRef<'tcx, Bx::Value>, | |
dc9dc135 XL |
243 | dst: PlaceRef<'tcx, Bx::Value>, |
244 | ) { | |
a1dfa0c6 XL |
245 | let src_ty = src.layout.ty; |
246 | let dst_ty = dst.layout.ty; | |
1b1a35ee | 247 | match (src_ty.kind(), dst_ty.kind()) { |
ba9703b0 | 248 | (&ty::Ref(..), &ty::Ref(..) | &ty::RawPtr(..)) | (&ty::RawPtr(..), &ty::RawPtr(..)) => { |
e74abb32 | 249 | let (base, info) = match bx.load_operand(src).val { |
94222f64 XL |
250 | OperandValue::Pair(base, info) => unsize_ptr(bx, base, src_ty, dst_ty, Some(info)), |
251 | OperandValue::Immediate(base) => unsize_ptr(bx, base, src_ty, dst_ty, None), | |
dfeec247 | 252 | OperandValue::Ref(..) => bug!(), |
e74abb32 XL |
253 | }; |
254 | OperandValue::Pair(base, info).store(bx, dst); | |
a1dfa0c6 XL |
255 | } |
256 | ||
257 | (&ty::Adt(def_a, _), &ty::Adt(def_b, _)) => { | |
258 | assert_eq!(def_a, def_b); | |
259 | ||
5e7ed085 | 260 | for i in 0..def_a.variant(VariantIdx::new(0)).fields.len() { |
a1dfa0c6 XL |
261 | let src_f = src.project_field(bx, i); |
262 | let dst_f = dst.project_field(bx, i); | |
263 | ||
264 | if dst_f.layout.is_zst() { | |
265 | continue; | |
266 | } | |
267 | ||
268 | if src_f.layout.ty == dst_f.layout.ty { | |
dfeec247 XL |
269 | memcpy_ty( |
270 | bx, | |
271 | dst_f.llval, | |
272 | dst_f.align, | |
273 | src_f.llval, | |
274 | src_f.align, | |
275 | src_f.layout, | |
276 | MemFlags::empty(), | |
277 | ); | |
a1dfa0c6 XL |
278 | } else { |
279 | coerce_unsized_into(bx, src_f, dst_f); | |
280 | } | |
281 | } | |
282 | } | |
dfeec247 | 283 | _ => bug!("coerce_unsized_into: invalid coercion {:?} -> {:?}", src_ty, dst_ty,), |
a1dfa0c6 XL |
284 | } |
285 | } | |
286 | ||
dc9dc135 | 287 | pub fn cast_shift_expr_rhs<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>( |
a1dfa0c6 XL |
288 | bx: &mut Bx, |
289 | op: hir::BinOpKind, | |
290 | lhs: Bx::Value, | |
dc9dc135 | 291 | rhs: Bx::Value, |
a1dfa0c6 XL |
292 | ) -> Bx::Value { |
293 | cast_shift_rhs(bx, op, lhs, rhs) | |
294 | } | |
295 | ||
dc9dc135 | 296 | fn cast_shift_rhs<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>( |
a1dfa0c6 XL |
297 | bx: &mut Bx, |
298 | op: hir::BinOpKind, | |
299 | lhs: Bx::Value, | |
300 | rhs: Bx::Value, | |
301 | ) -> Bx::Value { | |
302 | // Shifts may have any size int on the rhs | |
303 | if op.is_shift() { | |
304 | let mut rhs_llty = bx.cx().val_ty(rhs); | |
305 | let mut lhs_llty = bx.cx().val_ty(lhs); | |
306 | if bx.cx().type_kind(rhs_llty) == TypeKind::Vector { | |
307 | rhs_llty = bx.cx().element_type(rhs_llty) | |
308 | } | |
309 | if bx.cx().type_kind(lhs_llty) == TypeKind::Vector { | |
310 | lhs_llty = bx.cx().element_type(lhs_llty) | |
311 | } | |
312 | let rhs_sz = bx.cx().int_width(rhs_llty); | |
313 | let lhs_sz = bx.cx().int_width(lhs_llty); | |
314 | if lhs_sz < rhs_sz { | |
315 | bx.trunc(rhs, lhs_llty) | |
316 | } else if lhs_sz > rhs_sz { | |
317 | // FIXME (#1877: If in the future shifting by negative | |
318 | // values is no longer undefined then this is wrong. | |
319 | bx.zext(rhs, lhs_llty) | |
320 | } else { | |
321 | rhs | |
322 | } | |
323 | } else { | |
324 | rhs | |
325 | } | |
326 | } | |
327 | ||
9fa01778 | 328 | /// Returns `true` if this session's target will use SEH-based unwinding. |
a1dfa0c6 XL |
329 | /// |
330 | /// This is only true for MSVC targets, and even then the 64-bit MSVC target | |
331 | /// currently uses SEH-ish unwinding with DWARF info tables to the side (same as | |
332 | /// 64-bit MinGW) instead of "full SEH". | |
333 | pub fn wants_msvc_seh(sess: &Session) -> bool { | |
29967ef6 | 334 | sess.target.is_like_msvc |
a1dfa0c6 XL |
335 | } |
336 | ||
dc9dc135 | 337 | pub fn memcpy_ty<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>( |
a1dfa0c6 XL |
338 | bx: &mut Bx, |
339 | dst: Bx::Value, | |
340 | dst_align: Align, | |
341 | src: Bx::Value, | |
342 | src_align: Align, | |
ba9703b0 | 343 | layout: TyAndLayout<'tcx>, |
a1dfa0c6 XL |
344 | flags: MemFlags, |
345 | ) { | |
346 | let size = layout.size.bytes(); | |
347 | if size == 0 { | |
348 | return; | |
349 | } | |
350 | ||
351 | bx.memcpy(dst, dst_align, src, src_align, bx.cx().const_usize(size), flags); | |
352 | } | |
353 | ||
354 | pub fn codegen_instance<'a, 'tcx: 'a, Bx: BuilderMethods<'a, 'tcx>>( | |
355 | cx: &'a Bx::CodegenCx, | |
356 | instance: Instance<'tcx>, | |
357 | ) { | |
a1dfa0c6 XL |
358 | // this is an info! to allow collecting monomorphization statistics |
359 | // and to allow finding the last function before LLVM aborts from | |
360 | // release builds. | |
361 | info!("codegen_instance({})", instance); | |
362 | ||
60c5eb7d | 363 | mir::codegen_mir::<Bx>(cx, instance); |
a1dfa0c6 XL |
364 | } |
365 | ||
9fa01778 | 366 | /// Creates the `main` function which will initialize the rust runtime and call |
a1dfa0c6 | 367 | /// users main function. |
74b04a01 XL |
368 | pub fn maybe_create_entry_wrapper<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>( |
369 | cx: &'a Bx::CodegenCx, | |
370 | ) -> Option<Bx::Function> { | |
17df50a5 | 371 | let (main_def_id, entry_type) = cx.tcx().entry_fn(())?; |
cdc7bbd5 XL |
372 | let main_is_local = main_def_id.is_local(); |
373 | let instance = Instance::mono(cx.tcx(), main_def_id); | |
a1dfa0c6 | 374 | |
cdc7bbd5 | 375 | if main_is_local { |
a1dfa0c6 XL |
376 | // We want to create the wrapper in the same codegen unit as Rust's main |
377 | // function. | |
cdc7bbd5 XL |
378 | if !cx.codegen_unit().contains_item(&MonoItem::Fn(instance)) { |
379 | return None; | |
380 | } | |
17df50a5 XL |
381 | } else if !cx.codegen_unit().is_primary() { |
382 | // We want to create the wrapper only when the codegen unit is the primary one | |
383 | return None; | |
a1dfa0c6 XL |
384 | } |
385 | ||
e74abb32 | 386 | let main_llfn = cx.get_fn_addr(instance); |
a1dfa0c6 | 387 | |
17df50a5 XL |
388 | let use_start_lang_item = EntryFnType::Start != entry_type; |
389 | let entry_fn = create_entry_fn::<Bx>(cx, main_llfn, main_def_id, use_start_lang_item); | |
390 | return Some(entry_fn); | |
a1dfa0c6 | 391 | |
dc9dc135 | 392 | fn create_entry_fn<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>( |
a1dfa0c6 | 393 | cx: &'a Bx::CodegenCx, |
a1dfa0c6 | 394 | rust_main: Bx::Value, |
cdc7bbd5 | 395 | rust_main_def_id: DefId, |
a1dfa0c6 | 396 | use_start_lang_item: bool, |
74b04a01 | 397 | ) -> Bx::Function { |
e74abb32 XL |
398 | // The entry function is either `int main(void)` or `int main(int argc, char **argv)`, |
399 | // depending on whether the target needs `argc` and `argv` to be passed in. | |
29967ef6 | 400 | let llfty = if cx.sess().target.main_needs_argc_argv { |
e74abb32 XL |
401 | cx.type_func(&[cx.type_int(), cx.type_ptr_to(cx.type_i8p())], cx.type_int()) |
402 | } else { | |
403 | cx.type_func(&[], cx.type_int()) | |
404 | }; | |
a1dfa0c6 XL |
405 | |
406 | let main_ret_ty = cx.tcx().fn_sig(rust_main_def_id).output(); | |
407 | // Given that `main()` has no arguments, | |
408 | // then its return type cannot have | |
409 | // late-bound regions, since late-bound | |
410 | // regions must appear in the argument | |
411 | // listing. | |
5e7ed085 FG |
412 | let main_ret_ty = cx.tcx().normalize_erasing_regions( |
413 | ty::ParamEnv::reveal_all(), | |
414 | main_ret_ty.no_bound_vars().unwrap(), | |
415 | ); | |
416 | ||
417 | let Some(llfn) = cx.declare_c_main(llfty) else { | |
418 | // FIXME: We should be smart and show a better diagnostic here. | |
419 | let span = cx.tcx().def_span(rust_main_def_id); | |
420 | cx.sess() | |
421 | .struct_span_err(span, "entry symbol `main` declared multiple times") | |
422 | .help("did you use `#[no_mangle]` on `fn main`? Use `#[start]` instead") | |
423 | .emit(); | |
424 | cx.sess().abort_if_errors(); | |
425 | bug!(); | |
1b1a35ee | 426 | }; |
a1dfa0c6 XL |
427 | |
428 | // `main` should respect same config for frame pointer elimination as rest of code | |
136023e0 | 429 | cx.set_frame_pointer_type(llfn); |
a1dfa0c6 XL |
430 | cx.apply_target_cpu_attr(llfn); |
431 | ||
17df50a5 XL |
432 | let llbb = Bx::append_block(&cx, llfn, "top"); |
433 | let mut bx = Bx::build(&cx, llbb); | |
a1dfa0c6 XL |
434 | |
435 | bx.insert_reference_to_gdb_debug_scripts_section_global(); | |
436 | ||
94222f64 XL |
437 | let isize_ty = cx.type_isize(); |
438 | let i8pp_ty = cx.type_ptr_to(cx.type_i8p()); | |
e74abb32 | 439 | let (arg_argc, arg_argv) = get_argc_argv(cx, &mut bx); |
a1dfa0c6 | 440 | |
94222f64 | 441 | let (start_fn, start_ty, args) = if use_start_lang_item { |
3dfed10e | 442 | let start_def_id = cx.tcx().require_lang_item(LangItem::Start, None); |
e74abb32 XL |
443 | let start_fn = cx.get_fn_addr( |
444 | ty::Instance::resolve( | |
445 | cx.tcx(), | |
446 | ty::ParamEnv::reveal_all(), | |
447 | start_def_id, | |
448 | cx.tcx().intern_substs(&[main_ret_ty.into()]), | |
dfeec247 | 449 | ) |
f9f354fc | 450 | .unwrap() |
dfeec247 | 451 | .unwrap(), |
a1dfa0c6 | 452 | ); |
94222f64 XL |
453 | let start_ty = cx.type_func(&[cx.val_ty(rust_main), isize_ty, i8pp_ty], isize_ty); |
454 | (start_fn, start_ty, vec![rust_main, arg_argc, arg_argv]) | |
a1dfa0c6 XL |
455 | } else { |
456 | debug!("using user-defined start fn"); | |
94222f64 XL |
457 | let start_ty = cx.type_func(&[isize_ty, i8pp_ty], isize_ty); |
458 | (rust_main, start_ty, vec![arg_argc, arg_argv]) | |
a1dfa0c6 XL |
459 | }; |
460 | ||
94222f64 | 461 | let result = bx.call(start_ty, start_fn, &args, None); |
a1dfa0c6 XL |
462 | let cast = bx.intcast(result, cx.type_int(), true); |
463 | bx.ret(cast); | |
74b04a01 XL |
464 | |
465 | llfn | |
a1dfa0c6 XL |
466 | } |
467 | } | |
468 | ||
e74abb32 XL |
469 | /// Obtain the `argc` and `argv` values to pass to the rust start function. |
470 | fn get_argc_argv<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>( | |
471 | cx: &'a Bx::CodegenCx, | |
dfeec247 XL |
472 | bx: &mut Bx, |
473 | ) -> (Bx::Value, Bx::Value) { | |
29967ef6 | 474 | if cx.sess().target.main_needs_argc_argv { |
e74abb32 XL |
475 | // Params from native `main()` used as args for rust start function |
476 | let param_argc = bx.get_param(0); | |
477 | let param_argv = bx.get_param(1); | |
478 | let arg_argc = bx.intcast(param_argc, cx.type_isize(), true); | |
479 | let arg_argv = param_argv; | |
480 | (arg_argc, arg_argv) | |
481 | } else { | |
482 | // The Rust start function doesn't need `argc` and `argv`, so just pass zeros. | |
483 | let arg_argc = bx.const_int(cx.type_int(), 0); | |
484 | let arg_argv = bx.const_null(cx.type_ptr_to(cx.type_i8p())); | |
485 | (arg_argc, arg_argv) | |
486 | } | |
487 | } | |
488 | ||
a1dfa0c6 XL |
489 | pub fn codegen_crate<B: ExtraBackendMethods>( |
490 | backend: B, | |
a2a8927a | 491 | tcx: TyCtxt<'_>, |
17df50a5 | 492 | target_cpu: String, |
48663c56 XL |
493 | metadata: EncodedMetadata, |
494 | need_metadata_module: bool, | |
a1dfa0c6 | 495 | ) -> OngoingCodegen<B> { |
a1dfa0c6 | 496 | // Skip crate items and just output metadata in -Z no-codegen mode. |
dfeec247 | 497 | if tcx.sess.opts.debugging_opts.no_codegen || !tcx.sess.opts.output_types.should_codegen() { |
a2a8927a | 498 | let ongoing_codegen = start_async_codegen(backend, tcx, target_cpu, metadata, None, 1); |
a1dfa0c6 | 499 | |
a1dfa0c6 XL |
500 | ongoing_codegen.codegen_finished(tcx); |
501 | ||
a1dfa0c6 XL |
502 | ongoing_codegen.check_for_errors(tcx.sess); |
503 | ||
504 | return ongoing_codegen; | |
505 | } | |
506 | ||
48663c56 XL |
507 | let cgu_name_builder = &mut CodegenUnitNameBuilder::new(tcx); |
508 | ||
a1dfa0c6 XL |
509 | // Run the monomorphization collector and partition the collected items into |
510 | // codegen units. | |
17df50a5 | 511 | let codegen_units = tcx.collect_and_partition_mono_items(()).1; |
a1dfa0c6 XL |
512 | |
513 | // Force all codegen_unit queries so they are already either red or green | |
514 | // when compile_codegen_unit accesses them. We are not able to re-execute | |
515 | // the codegen_unit query from just the DepNode, so an unknown color would | |
516 | // lead to having to re-execute compile_codegen_unit, possibly | |
517 | // unnecessarily. | |
518 | if tcx.dep_graph.is_fully_enabled() { | |
ba9703b0 | 519 | for cgu in codegen_units { |
f9f354fc | 520 | tcx.ensure().codegen_unit(cgu.name()); |
a1dfa0c6 XL |
521 | } |
522 | } | |
523 | ||
a2a8927a XL |
524 | let metadata_module = if need_metadata_module { |
525 | // Emit compressed metadata object. | |
526 | let metadata_cgu_name = | |
527 | cgu_name_builder.build_cgu_name(LOCAL_CRATE, &["crate"], Some("metadata")).to_string(); | |
528 | tcx.sess.time("write_compressed_metadata", || { | |
529 | let file_name = | |
530 | tcx.output_filenames(()).temp_path(OutputType::Metadata, Some(&metadata_cgu_name)); | |
531 | let data = create_compressed_metadata_file( | |
532 | tcx.sess, | |
533 | &metadata, | |
534 | &exported_symbols::metadata_symbol_name(tcx), | |
535 | ); | |
536 | if let Err(err) = std::fs::write(&file_name, data) { | |
537 | tcx.sess.fatal(&format!("error writing metadata object file: {}", err)); | |
538 | } | |
539 | Some(CompiledModule { | |
540 | name: metadata_cgu_name, | |
541 | kind: ModuleKind::Metadata, | |
542 | object: Some(file_name), | |
543 | dwarf_object: None, | |
544 | bytecode: None, | |
545 | }) | |
546 | }) | |
547 | } else { | |
548 | None | |
549 | }; | |
550 | ||
551 | let ongoing_codegen = start_async_codegen( | |
552 | backend.clone(), | |
553 | tcx, | |
554 | target_cpu, | |
555 | metadata, | |
556 | metadata_module, | |
557 | codegen_units.len(), | |
558 | ); | |
a1dfa0c6 XL |
559 | let ongoing_codegen = AbortCodegenOnDrop::<B>(Some(ongoing_codegen)); |
560 | ||
561 | // Codegen an allocator shim, if necessary. | |
562 | // | |
563 | // If the crate doesn't have an `allocator_kind` set then there's definitely | |
564 | // no shim to generate. Otherwise we also check our dependency graph for all | |
565 | // our output crate types. If anything there looks like its a `Dynamic` | |
566 | // linkage, then it's already got an allocator shim and we'll be using that | |
567 | // one instead. If nothing exists then it's our job to generate the | |
568 | // allocator! | |
17df50a5 | 569 | let any_dynamic_crate = tcx.dependency_formats(()).iter().any(|(_, list)| { |
ba9703b0 | 570 | use rustc_middle::middle::dependency_format::Linkage; |
dfeec247 XL |
571 | list.iter().any(|&linkage| linkage == Linkage::Dynamic) |
572 | }); | |
a1dfa0c6 XL |
573 | let allocator_module = if any_dynamic_crate { |
574 | None | |
136023e0 | 575 | } else if let Some(kind) = tcx.allocator_kind(()) { |
dfeec247 XL |
576 | let llmod_id = |
577 | cgu_name_builder.build_cgu_name(LOCAL_CRATE, &["crate"], Some("allocator")).to_string(); | |
c295e0f8 | 578 | let mut module_llvm = backend.new_metadata(tcx, &llmod_id); |
29967ef6 | 579 | tcx.sess.time("write_allocator_module", || { |
c295e0f8 XL |
580 | backend.codegen_allocator( |
581 | tcx, | |
582 | &mut module_llvm, | |
583 | &llmod_id, | |
584 | kind, | |
585 | tcx.lang_items().oom().is_some(), | |
586 | ) | |
29967ef6 | 587 | }); |
a1dfa0c6 | 588 | |
c295e0f8 | 589 | Some(ModuleCodegen { name: llmod_id, module_llvm, kind: ModuleKind::Allocator }) |
a1dfa0c6 XL |
590 | } else { |
591 | None | |
592 | }; | |
593 | ||
594 | if let Some(allocator_module) = allocator_module { | |
595 | ongoing_codegen.submit_pre_codegened_module_to_llvm(tcx, allocator_module); | |
596 | } | |
597 | ||
5869c6ff XL |
598 | // For better throughput during parallel processing by LLVM, we used to sort |
599 | // CGUs largest to smallest. This would lead to better thread utilization | |
600 | // by, for example, preventing a large CGU from being processed last and | |
601 | // having only one LLVM thread working while the rest remained idle. | |
602 | // | |
603 | // However, this strategy would lead to high memory usage, as it meant the | |
604 | // LLVM-IR for all of the largest CGUs would be resident in memory at once. | |
605 | // | |
606 | // Instead, we can compromise by ordering CGUs such that the largest and | |
607 | // smallest are first, second largest and smallest are next, etc. If there | |
608 | // are large size variations, this can reduce memory usage significantly. | |
609 | let codegen_units: Vec<_> = { | |
610 | let mut sorted_cgus = codegen_units.iter().collect::<Vec<_>>(); | |
611 | sorted_cgus.sort_by_cached_key(|cgu| cgu.size_estimate()); | |
612 | ||
613 | let (first_half, second_half) = sorted_cgus.split_at(sorted_cgus.len() / 2); | |
614 | second_half.iter().rev().interleave(first_half).copied().collect() | |
a1dfa0c6 XL |
615 | }; |
616 | ||
dfeec247 XL |
617 | // The non-parallel compiler can only translate codegen units to LLVM IR |
618 | // on a single thread, leading to a staircase effect where the N LLVM | |
619 | // threads have to wait on the single codegen threads to generate work | |
620 | // for them. The parallel compiler does not have this restriction, so | |
621 | // we can pre-load the LLVM queue in parallel before handing off | |
622 | // coordination to the OnGoingCodegen scheduler. | |
623 | // | |
624 | // This likely is a temporary measure. Once we don't have to support the | |
625 | // non-parallel compiler anymore, we can compile CGUs end-to-end in | |
626 | // parallel and get rid of the complicated scheduling logic. | |
5e7ed085 | 627 | #[cfg(parallel_compiler)] |
dfeec247 | 628 | let pre_compile_cgus = |cgu_reuse: &[CguReuse]| { |
5e7ed085 FG |
629 | tcx.sess.time("compile_first_CGU_batch", || { |
630 | // Try to find one CGU to compile per thread. | |
631 | let cgus: Vec<_> = cgu_reuse | |
632 | .iter() | |
633 | .enumerate() | |
634 | .filter(|&(_, reuse)| reuse == &CguReuse::No) | |
635 | .take(tcx.sess.threads()) | |
636 | .collect(); | |
637 | ||
638 | // Compile the found CGUs in parallel. | |
639 | let start_time = Instant::now(); | |
640 | ||
641 | let pre_compiled_cgus = par_iter(cgus) | |
642 | .map(|(i, _)| { | |
643 | let module = backend.compile_codegen_unit(tcx, codegen_units[i].name()); | |
644 | (i, module) | |
645 | }) | |
646 | .collect(); | |
647 | ||
648 | (pre_compiled_cgus, start_time.elapsed()) | |
649 | }) | |
dfeec247 XL |
650 | }; |
651 | ||
5e7ed085 FG |
652 | #[cfg(not(parallel_compiler))] |
653 | let pre_compile_cgus = |_: &[CguReuse]| (FxHashMap::default(), Duration::new(0, 0)); | |
654 | ||
dfeec247 XL |
655 | let mut cgu_reuse = Vec::new(); |
656 | let mut pre_compiled_cgus: Option<FxHashMap<usize, _>> = None; | |
5869c6ff XL |
657 | let mut total_codegen_time = Duration::new(0, 0); |
658 | let start_rss = tcx.sess.time_passes().then(|| get_resident_set_size()); | |
a1dfa0c6 | 659 | |
dfeec247 | 660 | for (i, cgu) in codegen_units.iter().enumerate() { |
a1dfa0c6 XL |
661 | ongoing_codegen.wait_for_signal_to_codegen_item(); |
662 | ongoing_codegen.check_for_errors(tcx.sess); | |
663 | ||
dfeec247 XL |
664 | // Do some setup work in the first iteration |
665 | if pre_compiled_cgus.is_none() { | |
666 | // Calculate the CGU reuse | |
667 | cgu_reuse = tcx.sess.time("find_cgu_reuse", || { | |
668 | codegen_units.iter().map(|cgu| determine_cgu_reuse(tcx, &cgu)).collect() | |
669 | }); | |
670 | // Pre compile some CGUs | |
5869c6ff XL |
671 | let (compiled_cgus, codegen_time) = pre_compile_cgus(&cgu_reuse); |
672 | pre_compiled_cgus = Some(compiled_cgus); | |
673 | total_codegen_time += codegen_time; | |
dfeec247 XL |
674 | } |
675 | ||
676 | let cgu_reuse = cgu_reuse[i]; | |
a2a8927a | 677 | tcx.sess.cgu_reuse_tracker.set_actual_reuse(cgu.name().as_str(), cgu_reuse); |
a1dfa0c6 XL |
678 | |
679 | match cgu_reuse { | |
680 | CguReuse::No => { | |
dfeec247 XL |
681 | let (module, cost) = |
682 | if let Some(cgu) = pre_compiled_cgus.as_mut().unwrap().remove(&i) { | |
683 | cgu | |
684 | } else { | |
685 | let start_time = Instant::now(); | |
686 | let module = backend.compile_codegen_unit(tcx, cgu.name()); | |
5869c6ff | 687 | total_codegen_time += start_time.elapsed(); |
dfeec247 XL |
688 | module |
689 | }; | |
5869c6ff XL |
690 | // This will unwind if there are errors, which triggers our `AbortCodegenOnDrop` |
691 | // guard. Unfortunately, just skipping the `submit_codegened_module_to_llvm` makes | |
692 | // compilation hang on post-monomorphization errors. | |
693 | tcx.sess.abort_if_errors(); | |
694 | ||
dfeec247 XL |
695 | submit_codegened_module_to_llvm( |
696 | &backend, | |
697 | &ongoing_codegen.coordinator_send, | |
698 | module, | |
699 | cost, | |
700 | ); | |
a1dfa0c6 XL |
701 | false |
702 | } | |
703 | CguReuse::PreLto => { | |
dfeec247 XL |
704 | submit_pre_lto_module_to_llvm( |
705 | &backend, | |
706 | tcx, | |
707 | &ongoing_codegen.coordinator_send, | |
708 | CachedModuleCodegen { | |
709 | name: cgu.name().to_string(), | |
710 | source: cgu.work_product(tcx), | |
711 | }, | |
712 | ); | |
a1dfa0c6 XL |
713 | true |
714 | } | |
715 | CguReuse::PostLto => { | |
dfeec247 XL |
716 | submit_post_lto_module_to_llvm( |
717 | &backend, | |
718 | &ongoing_codegen.coordinator_send, | |
719 | CachedModuleCodegen { | |
720 | name: cgu.name().to_string(), | |
721 | source: cgu.work_product(tcx), | |
722 | }, | |
723 | ); | |
a1dfa0c6 XL |
724 | true |
725 | } | |
726 | }; | |
727 | } | |
728 | ||
729 | ongoing_codegen.codegen_finished(tcx); | |
730 | ||
731 | // Since the main thread is sometimes blocked during codegen, we keep track | |
732 | // -Ztime-passes output manually. | |
5869c6ff XL |
733 | if tcx.sess.time_passes() { |
734 | let end_rss = get_resident_set_size(); | |
735 | ||
736 | print_time_passes_entry( | |
737 | "codegen_to_LLVM_IR", | |
738 | total_codegen_time, | |
739 | start_rss.unwrap(), | |
740 | end_rss, | |
741 | ); | |
742 | } | |
a1dfa0c6 | 743 | |
a1dfa0c6 XL |
744 | ongoing_codegen.check_for_errors(tcx.sess); |
745 | ||
a1dfa0c6 XL |
746 | ongoing_codegen.into_inner() |
747 | } | |
748 | ||
749 | /// A curious wrapper structure whose only purpose is to call `codegen_aborted` | |
750 | /// when it's dropped abnormally. | |
751 | /// | |
752 | /// In the process of working on rust-lang/rust#55238 a mysterious segfault was | |
753 | /// stumbled upon. The segfault was never reproduced locally, but it was | |
754 | /// suspected to be related to the fact that codegen worker threads were | |
755 | /// sticking around by the time the main thread was exiting, causing issues. | |
756 | /// | |
757 | /// This structure is an attempt to fix that issue where the `codegen_aborted` | |
758 | /// message will block until all workers have finished. This should ensure that | |
759 | /// even if the main codegen thread panics we'll wait for pending work to | |
760 | /// complete before returning from the main thread, hopefully avoiding | |
761 | /// segfaults. | |
762 | /// | |
763 | /// If you see this comment in the code, then it means that this workaround | |
764 | /// worked! We may yet one day track down the mysterious cause of that | |
765 | /// segfault... | |
766 | struct AbortCodegenOnDrop<B: ExtraBackendMethods>(Option<OngoingCodegen<B>>); | |
767 | ||
768 | impl<B: ExtraBackendMethods> AbortCodegenOnDrop<B> { | |
769 | fn into_inner(mut self) -> OngoingCodegen<B> { | |
770 | self.0.take().unwrap() | |
771 | } | |
772 | } | |
773 | ||
774 | impl<B: ExtraBackendMethods> Deref for AbortCodegenOnDrop<B> { | |
775 | type Target = OngoingCodegen<B>; | |
776 | ||
777 | fn deref(&self) -> &OngoingCodegen<B> { | |
778 | self.0.as_ref().unwrap() | |
779 | } | |
780 | } | |
781 | ||
782 | impl<B: ExtraBackendMethods> DerefMut for AbortCodegenOnDrop<B> { | |
783 | fn deref_mut(&mut self) -> &mut OngoingCodegen<B> { | |
784 | self.0.as_mut().unwrap() | |
785 | } | |
786 | } | |
787 | ||
788 | impl<B: ExtraBackendMethods> Drop for AbortCodegenOnDrop<B> { | |
789 | fn drop(&mut self) { | |
790 | if let Some(codegen) = self.0.take() { | |
791 | codegen.codegen_aborted(); | |
792 | } | |
793 | } | |
794 | } | |
795 | ||
a1dfa0c6 | 796 | impl CrateInfo { |
136023e0 XL |
797 | pub fn new(tcx: TyCtxt<'_>, target_cpu: String) -> CrateInfo { |
798 | let exported_symbols = tcx | |
799 | .sess | |
800 | .crate_types() | |
801 | .iter() | |
802 | .map(|&c| (c, crate::back::linker::exported_symbols(tcx, c))) | |
803 | .collect(); | |
17df50a5 XL |
804 | let local_crate_name = tcx.crate_name(LOCAL_CRATE); |
805 | let crate_attrs = tcx.hir().attrs(rustc_hir::CRATE_HIR_ID); | |
806 | let subsystem = tcx.sess.first_attr_value_str_by_name(crate_attrs, sym::windows_subsystem); | |
807 | let windows_subsystem = subsystem.map(|subsystem| { | |
808 | if subsystem != sym::windows && subsystem != sym::console { | |
809 | tcx.sess.fatal(&format!( | |
810 | "invalid windows subsystem `{}`, only \ | |
811 | `windows` and `console` are allowed", | |
812 | subsystem | |
813 | )); | |
814 | } | |
815 | subsystem.to_string() | |
816 | }); | |
817 | ||
136023e0 XL |
818 | // This list is used when generating the command line to pass through to |
819 | // system linker. The linker expects undefined symbols on the left of the | |
820 | // command line to be defined in libraries on the right, not the other way | |
821 | // around. For more info, see some comments in the add_used_library function | |
822 | // below. | |
823 | // | |
824 | // In order to get this left-to-right dependency ordering, we use the reverse | |
825 | // postorder of all crates putting the leaves at the right-most positions. | |
826 | let used_crates = tcx | |
827 | .postorder_cnums(()) | |
828 | .iter() | |
829 | .rev() | |
830 | .copied() | |
831 | .filter(|&cnum| !tcx.dep_kind(cnum).macros_only()) | |
832 | .collect(); | |
833 | ||
a1dfa0c6 | 834 | let mut info = CrateInfo { |
136023e0 XL |
835 | target_cpu, |
836 | exported_symbols, | |
17df50a5 | 837 | local_crate_name, |
a1dfa0c6 XL |
838 | compiler_builtins: None, |
839 | profiler_runtime: None, | |
a1dfa0c6 XL |
840 | is_no_builtins: Default::default(), |
841 | native_libraries: Default::default(), | |
fc512014 | 842 | used_libraries: tcx.native_libraries(LOCAL_CRATE).iter().map(Into::into).collect(), |
a1dfa0c6 | 843 | crate_name: Default::default(), |
136023e0 | 844 | used_crates, |
a1dfa0c6 | 845 | used_crate_source: Default::default(), |
a1dfa0c6 XL |
846 | lang_item_to_crate: Default::default(), |
847 | missing_lang_items: Default::default(), | |
5099ac24 | 848 | dependency_formats: tcx.dependency_formats(()).clone(), |
17df50a5 | 849 | windows_subsystem, |
a1dfa0c6 XL |
850 | }; |
851 | let lang_items = tcx.lang_items(); | |
852 | ||
136023e0 | 853 | let crates = tcx.crates(()); |
a1dfa0c6 XL |
854 | |
855 | let n_crates = crates.len(); | |
856 | info.native_libraries.reserve(n_crates); | |
857 | info.crate_name.reserve(n_crates); | |
858 | info.used_crate_source.reserve(n_crates); | |
859 | info.missing_lang_items.reserve(n_crates); | |
860 | ||
861 | for &cnum in crates.iter() { | |
fc512014 XL |
862 | info.native_libraries |
863 | .insert(cnum, tcx.native_libraries(cnum).iter().map(Into::into).collect()); | |
a1dfa0c6 | 864 | info.crate_name.insert(cnum, tcx.crate_name(cnum).to_string()); |
5099ac24 | 865 | info.used_crate_source.insert(cnum, tcx.used_crate_source(cnum).clone()); |
a1dfa0c6 XL |
866 | if tcx.is_compiler_builtins(cnum) { |
867 | info.compiler_builtins = Some(cnum); | |
868 | } | |
869 | if tcx.is_profiler_runtime(cnum) { | |
870 | info.profiler_runtime = Some(cnum); | |
871 | } | |
a1dfa0c6 XL |
872 | if tcx.is_no_builtins(cnum) { |
873 | info.is_no_builtins.insert(cnum); | |
874 | } | |
a1dfa0c6 XL |
875 | let missing = tcx.missing_lang_items(cnum); |
876 | for &item in missing.iter() { | |
877 | if let Ok(id) = lang_items.require(item) { | |
878 | info.lang_item_to_crate.insert(item, id.krate); | |
879 | } | |
880 | } | |
881 | ||
f035d41b | 882 | // No need to look for lang items that don't actually need to exist. |
74b04a01 | 883 | let missing = |
f035d41b | 884 | missing.iter().cloned().filter(|&l| lang_items::required(tcx, l)).collect(); |
a1dfa0c6 XL |
885 | info.missing_lang_items.insert(cnum, missing); |
886 | } | |
887 | ||
ba9703b0 | 888 | info |
a1dfa0c6 | 889 | } |
a1dfa0c6 XL |
890 | } |
891 | ||
5869c6ff | 892 | pub fn provide(providers: &mut Providers) { |
9fa01778 XL |
893 | providers.backend_optimization_level = |tcx, cratenum| { |
894 | let for_speed = match tcx.sess.opts.optimize { | |
895 | // If globally no optimisation is done, #[optimize] has no effect. | |
896 | // | |
897 | // This is done because if we ended up "upgrading" to `-O2` here, we’d populate the | |
898 | // pass manager and it is likely that some module-wide passes (such as inliner or | |
899 | // cross-function constant propagation) would ignore the `optnone` annotation we put | |
900 | // on the functions, thus necessarily involving these functions into optimisations. | |
901 | config::OptLevel::No => return config::OptLevel::No, | |
902 | // If globally optimise-speed is already specified, just use that level. | |
903 | config::OptLevel::Less => return config::OptLevel::Less, | |
904 | config::OptLevel::Default => return config::OptLevel::Default, | |
905 | config::OptLevel::Aggressive => return config::OptLevel::Aggressive, | |
906 | // If globally optimize-for-size has been requested, use -O2 instead (if optimize(size) | |
907 | // are present). | |
908 | config::OptLevel::Size => config::OptLevel::Default, | |
909 | config::OptLevel::SizeMin => config::OptLevel::Default, | |
910 | }; | |
911 | ||
912 | let (defids, _) = tcx.collect_and_partition_mono_items(cratenum); | |
913 | for id in &*defids { | |
dfeec247 | 914 | let CodegenFnAttrs { optimize, .. } = tcx.codegen_fn_attrs(*id); |
9fa01778 XL |
915 | match optimize { |
916 | attr::OptimizeAttr::None => continue, | |
917 | attr::OptimizeAttr::Size => continue, | |
918 | attr::OptimizeAttr::Speed => { | |
919 | return for_speed; | |
920 | } | |
921 | } | |
922 | } | |
ba9703b0 | 923 | tcx.sess.opts.optimize |
9fa01778 | 924 | }; |
a1dfa0c6 XL |
925 | } |
926 | ||
dc9dc135 | 927 | fn determine_cgu_reuse<'tcx>(tcx: TyCtxt<'tcx>, cgu: &CodegenUnit<'tcx>) -> CguReuse { |
a1dfa0c6 | 928 | if !tcx.dep_graph.is_fully_enabled() { |
dfeec247 | 929 | return CguReuse::No; |
a1dfa0c6 XL |
930 | } |
931 | ||
932 | let work_product_id = &cgu.work_product_id(); | |
933 | if tcx.dep_graph.previous_work_product(work_product_id).is_none() { | |
934 | // We don't have anything cached for this CGU. This can happen | |
935 | // if the CGU did not exist in the previous session. | |
dfeec247 | 936 | return CguReuse::No; |
a1dfa0c6 XL |
937 | } |
938 | ||
939 | // Try to mark the CGU as green. If it we can do so, it means that nothing | |
940 | // affecting the LLVM module has changed and we can re-use a cached version. | |
941 | // If we compile with any kind of LTO, this means we can re-use the bitcode | |
942 | // of the Pre-LTO stage (possibly also the Post-LTO version but we'll only | |
943 | // know that later). If we are not doing LTO, there is only one optimized | |
944 | // version of each module, so we re-use that. | |
945 | let dep_node = cgu.codegen_dep_node(tcx); | |
dfeec247 XL |
946 | assert!( |
947 | !tcx.dep_graph.dep_node_exists(&dep_node), | |
a1dfa0c6 | 948 | "CompileCodegenUnit dep-node for CGU `{}` already exists before marking.", |
dfeec247 XL |
949 | cgu.name() |
950 | ); | |
a1dfa0c6 | 951 | |
6a06907d | 952 | if tcx.try_mark_green(&dep_node) { |
f9f354fc XL |
953 | // We can re-use either the pre- or the post-thinlto state. If no LTO is |
954 | // being performed then we can use post-LTO artifacts, otherwise we must | |
955 | // reuse pre-LTO artifacts | |
956 | match compute_per_cgu_lto_type( | |
957 | &tcx.sess.lto(), | |
958 | &tcx.sess.opts, | |
959 | &tcx.sess.crate_types(), | |
960 | ModuleKind::Regular, | |
961 | ) { | |
962 | ComputedLtoType::No => CguReuse::PostLto, | |
963 | _ => CguReuse::PreLto, | |
964 | } | |
a1dfa0c6 XL |
965 | } else { |
966 | CguReuse::No | |
967 | } | |
968 | } |