1 // Copyright 2013-2015 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.
12 use back
::bytecode
::{self, RLIB_BYTECODE_EXTENSION}
;
13 use back
::lto
::{self, ModuleBuffer, ThinBuffer}
;
14 use back
::link
::{self, get_linker, remove}
;
15 use back
::command
::Command
;
16 use back
::linker
::LinkerInfo
;
17 use back
::symbol_export
::ExportedSymbols
;
20 use rustc_incremental
::{save_trans_partition, in_incr_comp_dir}
;
21 use rustc
::dep_graph
::{DepGraph, WorkProductFileKind}
;
22 use rustc
::middle
::cstore
::{LinkMeta, EncodedMetadata}
;
23 use rustc
::session
::config
::{self, OutputFilenames
, OutputType
, Passes
, SomePasses
,
24 AllPasses
, Sanitizer
, Lto
};
25 use rustc
::session
::Session
;
26 use rustc
::util
::nodemap
::FxHashMap
;
27 use time_graph
::{self, TimeGraph, Timeline}
;
29 use llvm
::{ModuleRef, TargetMachineRef, PassManagerRef, DiagnosticInfoRef}
;
30 use llvm
::{SMDiagnosticRef, ContextRef}
;
31 use {CrateTranslation, ModuleSource, ModuleTranslation, CompiledModule, ModuleKind}
;
33 use rustc
::hir
::def_id
::{CrateNum, LOCAL_CRATE}
;
34 use rustc
::ty
::TyCtxt
;
35 use rustc
::util
::common
::{time_ext, time_depth, set_time_depth, print_time_passes_entry}
;
36 use rustc
::util
::common
::path2cstr
;
37 use rustc
::util
::fs
::{link_or_copy}
;
38 use errors
::{self, Handler, Level, DiagnosticBuilder, FatalError, DiagnosticId}
;
39 use errors
::emitter
::{Emitter}
;
41 use syntax
::ext
::hygiene
::Mark
;
42 use syntax_pos
::MultiSpan
;
43 use syntax_pos
::symbol
::Symbol
;
45 use context
::{is_pie_binary, get_reloc_model}
;
46 use common
::{C_bytes_in_context, val_ty}
;
47 use jobserver
::{Client, Acquired}
;
51 use std
::ffi
::{CString, CStr}
;
53 use std
::io
::{self, Write}
;
55 use std
::path
::{Path, PathBuf}
;
58 use std
::sync
::mpsc
::{channel, Sender, Receiver}
;
60 use std
::time
::Instant
;
62 use libc
::{c_uint, c_void, c_char, size_t}
;
64 pub const RELOC_MODEL_ARGS
: [(&'
static str, llvm
::RelocMode
); 7] = [
65 ("pic", llvm
::RelocMode
::PIC
),
66 ("static", llvm
::RelocMode
::Static
),
67 ("default", llvm
::RelocMode
::Default
),
68 ("dynamic-no-pic", llvm
::RelocMode
::DynamicNoPic
),
69 ("ropi", llvm
::RelocMode
::ROPI
),
70 ("rwpi", llvm
::RelocMode
::RWPI
),
71 ("ropi-rwpi", llvm
::RelocMode
::ROPI_RWPI
),
74 pub const CODE_GEN_MODEL_ARGS
: &[(&str, llvm
::CodeModel
)] = &[
75 ("small", llvm
::CodeModel
::Small
),
76 ("kernel", llvm
::CodeModel
::Kernel
),
77 ("medium", llvm
::CodeModel
::Medium
),
78 ("large", llvm
::CodeModel
::Large
),
81 pub const TLS_MODEL_ARGS
: [(&'
static str, llvm
::ThreadLocalMode
); 4] = [
82 ("global-dynamic", llvm
::ThreadLocalMode
::GeneralDynamic
),
83 ("local-dynamic", llvm
::ThreadLocalMode
::LocalDynamic
),
84 ("initial-exec", llvm
::ThreadLocalMode
::InitialExec
),
85 ("local-exec", llvm
::ThreadLocalMode
::LocalExec
),
88 pub fn llvm_err(handler
: &errors
::Handler
, msg
: String
) -> FatalError
{
89 match llvm
::last_error() {
90 Some(err
) => handler
.fatal(&format
!("{}: {}", msg
, err
)),
91 None
=> handler
.fatal(&msg
),
95 pub fn write_output_file(
96 handler
: &errors
::Handler
,
97 target
: llvm
::TargetMachineRef
,
98 pm
: llvm
::PassManagerRef
,
101 file_type
: llvm
::FileType
) -> Result
<(), FatalError
> {
103 let output_c
= path2cstr(output
);
104 let result
= llvm
::LLVMRustWriteOutputFile(
105 target
, pm
, m
, output_c
.as_ptr(), file_type
);
106 if result
.into_result().is_err() {
107 let msg
= format
!("could not write output to {}", output
.display());
108 Err(llvm_err(handler
, msg
))
115 fn get_llvm_opt_level(optimize
: config
::OptLevel
) -> llvm
::CodeGenOptLevel
{
117 config
::OptLevel
::No
=> llvm
::CodeGenOptLevel
::None
,
118 config
::OptLevel
::Less
=> llvm
::CodeGenOptLevel
::Less
,
119 config
::OptLevel
::Default
=> llvm
::CodeGenOptLevel
::Default
,
120 config
::OptLevel
::Aggressive
=> llvm
::CodeGenOptLevel
::Aggressive
,
121 _
=> llvm
::CodeGenOptLevel
::Default
,
125 fn get_llvm_opt_size(optimize
: config
::OptLevel
) -> llvm
::CodeGenOptSize
{
127 config
::OptLevel
::Size
=> llvm
::CodeGenOptSizeDefault
,
128 config
::OptLevel
::SizeMin
=> llvm
::CodeGenOptSizeAggressive
,
129 _
=> llvm
::CodeGenOptSizeNone
,
133 pub fn create_target_machine(sess
: &Session
, find_features
: bool
) -> TargetMachineRef
{
134 target_machine_factory(sess
, find_features
)().unwrap_or_else(|err
| {
135 llvm_err(sess
.diagnostic(), err
).raise()
139 // If find_features is true this won't access `sess.crate_types` by assuming
140 // that `is_pie_binary` is false. When we discover LLVM target features
141 // `sess.crate_types` is uninitialized so we cannot access it.
142 pub fn target_machine_factory(sess
: &Session
, find_features
: bool
)
143 -> Arc
<Fn() -> Result
<TargetMachineRef
, String
> + Send
+ Sync
>
145 let reloc_model
= get_reloc_model(sess
);
147 let opt_level
= get_llvm_opt_level(sess
.opts
.optimize
);
148 let use_softfp
= sess
.opts
.cg
.soft_float
;
150 let ffunction_sections
= sess
.target
.target
.options
.function_sections
;
151 let fdata_sections
= ffunction_sections
;
153 let code_model_arg
= sess
.opts
.cg
.code_model
.as_ref().or(
154 sess
.target
.target
.options
.code_model
.as_ref(),
157 let code_model
= match code_model_arg
{
159 match CODE_GEN_MODEL_ARGS
.iter().find(|arg
| arg
.0 == s
) {
162 sess
.err(&format
!("{:?} is not a valid code model",
164 sess
.abort_if_errors();
169 None
=> llvm
::CodeModel
::None
,
172 let singlethread
= sess
.target
.target
.options
.singlethread
;
174 let triple
= &sess
.target
.target
.llvm_target
;
176 let triple
= CString
::new(triple
.as_bytes()).unwrap();
177 let cpu
= match sess
.opts
.cg
.target_cpu
{
179 None
=> &*sess
.target
.target
.options
.cpu
181 let cpu
= CString
::new(cpu
.as_bytes()).unwrap();
182 let features
= attributes
::llvm_target_features(sess
)
185 let features
= CString
::new(features
).unwrap();
186 let is_pie_binary
= !find_features
&& is_pie_binary(sess
);
187 let trap_unreachable
= sess
.target
.target
.options
.trap_unreachable
;
191 llvm
::LLVMRustCreateTargetMachine(
192 triple
.as_ptr(), cpu
.as_ptr(), features
.as_ptr(),
206 Err(format
!("Could not create LLVM TargetMachine for triple: {}",
207 triple
.to_str().unwrap()))
214 /// Module-specific configuration for `optimize_and_codegen`.
215 pub struct ModuleConfig
{
216 /// Names of additional optimization passes to run.
218 /// Some(level) to optimize at a certain level, or None to run
219 /// absolutely no optimizations (used for the metadata module).
220 pub opt_level
: Option
<llvm
::CodeGenOptLevel
>,
222 /// Some(level) to optimize binary size, or None to not affect program size.
223 opt_size
: Option
<llvm
::CodeGenOptSize
>,
225 pgo_gen
: Option
<String
>,
228 // Flags indicating which outputs to produce.
229 emit_no_opt_bc
: bool
,
231 emit_bc_compressed
: bool
,
236 // Miscellaneous flags. These are mostly copied from command-line
239 no_prepopulate_passes
: bool
,
242 vectorize_loop
: bool
,
244 merge_functions
: bool
,
245 inline_threshold
: Option
<usize>,
246 // Instead of creating an object file by doing LLVM codegen, just
247 // make the object file bitcode. Provides easy compatibility with
248 // emscripten's ecc compiler, when used as the linker.
249 obj_is_bitcode
: bool
,
250 no_integrated_as
: bool
,
252 embed_bitcode_marker
: bool
,
256 fn new(passes
: Vec
<String
>) -> ModuleConfig
{
263 pgo_use
: String
::new(),
265 emit_no_opt_bc
: false,
267 emit_bc_compressed
: false,
272 obj_is_bitcode
: false,
273 embed_bitcode
: false,
274 embed_bitcode_marker
: false,
275 no_integrated_as
: false,
278 no_prepopulate_passes
: false,
281 vectorize_loop
: false,
282 vectorize_slp
: false,
283 merge_functions
: false,
284 inline_threshold
: None
288 fn set_flags(&mut self, sess
: &Session
, no_builtins
: bool
) {
289 self.no_verify
= sess
.no_verify();
290 self.no_prepopulate_passes
= sess
.opts
.cg
.no_prepopulate_passes
;
291 self.no_builtins
= no_builtins
|| sess
.target
.target
.options
.no_builtins
;
292 self.time_passes
= sess
.time_passes();
293 self.inline_threshold
= sess
.opts
.cg
.inline_threshold
;
294 self.obj_is_bitcode
= sess
.target
.target
.options
.obj_is_bitcode
;
295 let embed_bitcode
= sess
.target
.target
.options
.embed_bitcode
||
296 sess
.opts
.debugging_opts
.embed_bitcode
||
297 sess
.opts
.debugging_opts
.cross_lang_lto
;
299 match sess
.opts
.optimize
{
300 config
::OptLevel
::No
|
301 config
::OptLevel
::Less
=> {
302 self.embed_bitcode_marker
= embed_bitcode
;
304 _
=> self.embed_bitcode
= embed_bitcode
,
308 // Copy what clang does by turning on loop vectorization at O2 and
309 // slp vectorization at O3. Otherwise configure other optimization aspects
310 // of this pass manager builder.
311 // Turn off vectorization for emscripten, as it's not very well supported.
312 self.vectorize_loop
= !sess
.opts
.cg
.no_vectorize_loops
&&
313 (sess
.opts
.optimize
== config
::OptLevel
::Default
||
314 sess
.opts
.optimize
== config
::OptLevel
::Aggressive
) &&
315 !sess
.target
.target
.options
.is_like_emscripten
;
317 self.vectorize_slp
= !sess
.opts
.cg
.no_vectorize_slp
&&
318 sess
.opts
.optimize
== config
::OptLevel
::Aggressive
&&
319 !sess
.target
.target
.options
.is_like_emscripten
;
321 self.merge_functions
= sess
.opts
.optimize
== config
::OptLevel
::Default
||
322 sess
.opts
.optimize
== config
::OptLevel
::Aggressive
;
326 /// Assembler name and command used by codegen when no_integrated_as is enabled
327 struct AssemblerCommand
{
332 /// Additional resources used by optimize_and_codegen (not module specific)
334 pub struct CodegenContext
{
335 // Resouces needed when running LTO
336 pub time_passes
: bool
,
338 pub no_landing_pads
: bool
,
339 pub save_temps
: bool
,
340 pub fewer_names
: bool
,
341 pub exported_symbols
: Option
<Arc
<ExportedSymbols
>>,
342 pub opts
: Arc
<config
::Options
>,
343 pub crate_types
: Vec
<config
::CrateType
>,
344 pub each_linked_rlib_for_lto
: Vec
<(CrateNum
, PathBuf
)>,
345 output_filenames
: Arc
<OutputFilenames
>,
346 regular_module_config
: Arc
<ModuleConfig
>,
347 metadata_module_config
: Arc
<ModuleConfig
>,
348 allocator_module_config
: Arc
<ModuleConfig
>,
349 pub tm_factory
: Arc
<Fn() -> Result
<TargetMachineRef
, String
> + Send
+ Sync
>,
350 pub msvc_imps_needed
: bool
,
351 pub target_pointer_width
: String
,
352 debuginfo
: config
::DebugInfoLevel
,
354 // Number of cgus excluding the allocator/metadata modules
355 pub total_cgus
: usize,
356 // Handler to use for diagnostics produced during codegen.
357 pub diag_emitter
: SharedEmitter
,
358 // LLVM passes added by plugins.
359 pub plugin_passes
: Vec
<String
>,
360 // LLVM optimizations for which we want to print remarks.
362 // Worker thread number
364 // The incremental compilation session directory, or None if we are not
365 // compiling incrementally
366 pub incr_comp_session_dir
: Option
<PathBuf
>,
367 // Channel back to the main control thread to send messages to
368 coordinator_send
: Sender
<Box
<Any
+ Send
>>,
369 // A reference to the TimeGraph so we can register timings. None means that
370 // measuring is disabled.
371 time_graph
: Option
<TimeGraph
>,
372 // The assembler command if no_integrated_as option is enabled, None otherwise
373 assembler_cmd
: Option
<Arc
<AssemblerCommand
>>,
376 impl CodegenContext
{
377 pub fn create_diag_handler(&self) -> Handler
{
378 Handler
::with_emitter(true, false, Box
::new(self.diag_emitter
.clone()))
381 pub(crate) fn config(&self, kind
: ModuleKind
) -> &ModuleConfig
{
383 ModuleKind
::Regular
=> &self.regular_module_config
,
384 ModuleKind
::Metadata
=> &self.metadata_module_config
,
385 ModuleKind
::Allocator
=> &self.allocator_module_config
,
389 pub(crate) fn save_temp_bitcode(&self, trans
: &ModuleTranslation
, name
: &str) {
390 if !self.save_temps
{
394 let ext
= format
!("{}.bc", name
);
395 let cgu
= Some(&trans
.name
[..]);
396 let path
= self.output_filenames
.temp_path_ext(&ext
, cgu
);
397 let cstr
= path2cstr(&path
);
398 let llmod
= trans
.llvm().unwrap().llmod
;
399 llvm
::LLVMWriteBitcodeToFile(llmod
, cstr
.as_ptr());
404 struct DiagnosticHandlers
<'a
> {
405 inner
: Box
<(&'a CodegenContext
, &'a Handler
)>,
409 impl<'a
> DiagnosticHandlers
<'a
> {
410 fn new(cgcx
: &'a CodegenContext
,
411 handler
: &'a Handler
,
412 llcx
: ContextRef
) -> DiagnosticHandlers
<'a
> {
413 let data
= Box
::new((cgcx
, handler
));
415 let arg
= &*data
as &(_
, _
) as *const _
as *mut _
;
416 llvm
::LLVMRustSetInlineAsmDiagnosticHandler(llcx
, inline_asm_handler
, arg
);
417 llvm
::LLVMContextSetDiagnosticHandler(llcx
, diagnostic_handler
, arg
);
426 impl<'a
> Drop
for DiagnosticHandlers
<'a
> {
429 llvm
::LLVMRustSetInlineAsmDiagnosticHandler(self.llcx
, inline_asm_handler
, 0 as *mut _
);
430 llvm
::LLVMContextSetDiagnosticHandler(self.llcx
, diagnostic_handler
, 0 as *mut _
);
435 unsafe extern "C" fn report_inline_asm
<'a
, 'b
>(cgcx
: &'a CodegenContext
,
438 cgcx
.diag_emitter
.inline_asm_error(cookie
as u32, msg
.to_string());
441 unsafe extern "C" fn inline_asm_handler(diag
: SMDiagnosticRef
,
447 let (cgcx
, _
) = *(user
as *const (&CodegenContext
, &Handler
));
449 let msg
= llvm
::build_string(|s
| llvm
::LLVMRustWriteSMDiagnosticToString(diag
, s
))
450 .expect("non-UTF8 SMDiagnostic");
452 report_inline_asm(cgcx
, &msg
, cookie
);
455 unsafe extern "C" fn diagnostic_handler(info
: DiagnosticInfoRef
, user
: *mut c_void
) {
459 let (cgcx
, diag_handler
) = *(user
as *const (&CodegenContext
, &Handler
));
461 match llvm
::diagnostic
::Diagnostic
::unpack(info
) {
462 llvm
::diagnostic
::InlineAsm(inline
) => {
463 report_inline_asm(cgcx
,
464 &llvm
::twine_to_string(inline
.message
),
468 llvm
::diagnostic
::Optimization(opt
) => {
469 let enabled
= match cgcx
.remark
{
471 SomePasses(ref v
) => v
.iter().any(|s
| *s
== opt
.pass_name
),
475 diag_handler
.note_without_error(&format
!("optimization {} for {} at {}:{}:{}: {}",
484 llvm
::diagnostic
::PGO(diagnostic_ref
) => {
485 let msg
= llvm
::build_string(|s
| {
486 llvm
::LLVMRustWriteDiagnosticInfoToString(diagnostic_ref
, s
)
487 }).expect("non-UTF8 PGO diagnostic");
488 diag_handler
.warn(&msg
);
490 llvm
::diagnostic
::UnknownDiagnostic(..) => {}
,
494 // Unsafe due to LLVM calls.
495 unsafe fn optimize(cgcx
: &CodegenContext
,
496 diag_handler
: &Handler
,
497 mtrans
: &ModuleTranslation
,
498 config
: &ModuleConfig
,
499 timeline
: &mut Timeline
)
500 -> Result
<(), FatalError
>
502 let (llmod
, llcx
, tm
) = match mtrans
.source
{
503 ModuleSource
::Translated(ref llvm
) => (llvm
.llmod
, llvm
.llcx
, llvm
.tm
),
504 ModuleSource
::Preexisting(_
) => {
505 bug
!("optimize_and_codegen: called with ModuleSource::Preexisting")
509 let _handlers
= DiagnosticHandlers
::new(cgcx
, diag_handler
, llcx
);
511 let module_name
= mtrans
.name
.clone();
512 let module_name
= Some(&module_name
[..]);
514 if config
.emit_no_opt_bc
{
515 let out
= cgcx
.output_filenames
.temp_path_ext("no-opt.bc", module_name
);
516 let out
= path2cstr(&out
);
517 llvm
::LLVMWriteBitcodeToFile(llmod
, out
.as_ptr());
520 if config
.opt_level
.is_some() {
521 // Create the two optimizing pass managers. These mirror what clang
522 // does, and are by populated by LLVM's default PassManagerBuilder.
523 // Each manager has a different set of passes, but they also share
524 // some common passes.
525 let fpm
= llvm
::LLVMCreateFunctionPassManagerForModule(llmod
);
526 let mpm
= llvm
::LLVMCreatePassManager();
528 // If we're verifying or linting, add them to the function pass
530 let addpass
= |pass_name
: &str| {
531 let pass_name
= CString
::new(pass_name
).unwrap();
532 let pass
= llvm
::LLVMRustFindAndCreatePass(pass_name
.as_ptr());
536 let pass_manager
= match llvm
::LLVMRustPassKind(pass
) {
537 llvm
::PassKind
::Function
=> fpm
,
538 llvm
::PassKind
::Module
=> mpm
,
539 llvm
::PassKind
::Other
=> {
540 diag_handler
.err("Encountered LLVM pass kind we can't handle");
544 llvm
::LLVMRustAddPass(pass_manager
, pass
);
548 if !config
.no_verify { assert!(addpass("verify")); }
549 if !config
.no_prepopulate_passes
{
550 llvm
::LLVMRustAddAnalysisPasses(tm
, fpm
, llmod
);
551 llvm
::LLVMRustAddAnalysisPasses(tm
, mpm
, llmod
);
552 let opt_level
= config
.opt_level
.unwrap_or(llvm
::CodeGenOptLevel
::None
);
553 with_llvm_pmb(llmod
, &config
, opt_level
, &mut |b
| {
554 llvm
::LLVMPassManagerBuilderPopulateFunctionPassManager(b
, fpm
);
555 llvm
::LLVMPassManagerBuilderPopulateModulePassManager(b
, mpm
);
559 for pass
in &config
.passes
{
561 diag_handler
.warn(&format
!("unknown pass `{}`, ignoring",
566 for pass
in &cgcx
.plugin_passes
{
568 diag_handler
.err(&format
!("a plugin asked for LLVM pass \
569 `{}` but LLVM does not \
570 recognize it", pass
));
574 diag_handler
.abort_if_errors();
576 // Finally, run the actual optimization passes
577 time_ext(config
.time_passes
,
579 &format
!("llvm function passes [{}]", module_name
.unwrap()),
581 llvm
::LLVMRustRunFunctionPassManager(fpm
, llmod
)
583 timeline
.record("fpm");
584 time_ext(config
.time_passes
,
586 &format
!("llvm module passes [{}]", module_name
.unwrap()),
588 llvm
::LLVMRunPassManager(mpm
, llmod
)
591 // Deallocate managers that we're now done with
592 llvm
::LLVMDisposePassManager(fpm
);
593 llvm
::LLVMDisposePassManager(mpm
);
598 fn generate_lto_work(cgcx
: &CodegenContext
,
599 modules
: Vec
<ModuleTranslation
>)
600 -> Vec
<(WorkItem
, u64)>
602 let mut timeline
= cgcx
.time_graph
.as_ref().map(|tg
| {
603 tg
.start(TRANS_WORKER_TIMELINE
,
604 TRANS_WORK_PACKAGE_KIND
,
606 }).unwrap_or(Timeline
::noop());
607 let lto_modules
= lto
::run(cgcx
, modules
, &mut timeline
)
608 .unwrap_or_else(|e
| e
.raise());
610 lto_modules
.into_iter().map(|module
| {
611 let cost
= module
.cost();
612 (WorkItem
::LTO(module
), cost
)
616 unsafe fn codegen(cgcx
: &CodegenContext
,
617 diag_handler
: &Handler
,
618 mtrans
: ModuleTranslation
,
619 config
: &ModuleConfig
,
620 timeline
: &mut Timeline
)
621 -> Result
<CompiledModule
, FatalError
>
623 timeline
.record("codegen");
624 let (llmod
, llcx
, tm
) = match mtrans
.source
{
625 ModuleSource
::Translated(ref llvm
) => (llvm
.llmod
, llvm
.llcx
, llvm
.tm
),
626 ModuleSource
::Preexisting(_
) => {
627 bug
!("codegen: called with ModuleSource::Preexisting")
630 let module_name
= mtrans
.name
.clone();
631 let module_name
= Some(&module_name
[..]);
632 let handlers
= DiagnosticHandlers
::new(cgcx
, diag_handler
, llcx
);
634 if cgcx
.msvc_imps_needed
{
635 create_msvc_imps(cgcx
, llcx
, llmod
);
638 // A codegen-specific pass manager is used to generate object
639 // files for an LLVM module.
641 // Apparently each of these pass managers is a one-shot kind of
642 // thing, so we create a new one for each type of output. The
643 // pass manager passed to the closure should be ensured to not
644 // escape the closure itself, and the manager should only be
646 unsafe fn with_codegen
<F
, R
>(tm
: TargetMachineRef
,
650 where F
: FnOnce(PassManagerRef
) -> R
,
652 let cpm
= llvm
::LLVMCreatePassManager();
653 llvm
::LLVMRustAddAnalysisPasses(tm
, cpm
, llmod
);
654 llvm
::LLVMRustAddLibraryInfo(cpm
, llmod
, no_builtins
);
658 // If we don't have the integrated assembler, then we need to emit asm
659 // from LLVM and use `gcc` to create the object file.
660 let asm_to_obj
= config
.emit_obj
&& config
.no_integrated_as
;
662 // Change what we write and cleanup based on whether obj files are
663 // just llvm bitcode. In that case write bitcode, and possibly
664 // delete the bitcode if it wasn't requested. Don't generate the
665 // machine code, instead copy the .o file from the .bc
666 let write_bc
= config
.emit_bc
|| config
.obj_is_bitcode
;
667 let rm_bc
= !config
.emit_bc
&& config
.obj_is_bitcode
;
668 let write_obj
= config
.emit_obj
&& !config
.obj_is_bitcode
&& !asm_to_obj
;
669 let copy_bc_to_obj
= config
.emit_obj
&& config
.obj_is_bitcode
;
671 let bc_out
= cgcx
.output_filenames
.temp_path(OutputType
::Bitcode
, module_name
);
672 let obj_out
= cgcx
.output_filenames
.temp_path(OutputType
::Object
, module_name
);
675 if write_bc
|| config
.emit_bc_compressed
|| config
.embed_bitcode
{
678 let data
= if llvm
::LLVMRustThinLTOAvailable() {
679 thin
= ThinBuffer
::new(llmod
);
682 old
= ModuleBuffer
::new(llmod
);
685 timeline
.record("make-bc");
688 if let Err(e
) = fs
::write(&bc_out
, data
) {
689 diag_handler
.err(&format
!("failed to write bytecode: {}", e
));
691 timeline
.record("write-bc");
694 if config
.embed_bitcode
{
695 embed_bitcode(cgcx
, llcx
, llmod
, Some(data
));
696 timeline
.record("embed-bc");
699 if config
.emit_bc_compressed
{
700 let dst
= bc_out
.with_extension(RLIB_BYTECODE_EXTENSION
);
701 let data
= bytecode
::encode(&mtrans
.llmod_id
, data
);
702 if let Err(e
) = fs
::write(&dst
, data
) {
703 diag_handler
.err(&format
!("failed to write bytecode: {}", e
));
705 timeline
.record("compress-bc");
707 } else if config
.embed_bitcode_marker
{
708 embed_bitcode(cgcx
, llcx
, llmod
, None
);
711 time_ext(config
.time_passes
, None
, &format
!("codegen passes [{}]", module_name
.unwrap()),
712 || -> Result
<(), FatalError
> {
714 let out
= cgcx
.output_filenames
.temp_path(OutputType
::LlvmAssembly
, module_name
);
715 let out
= path2cstr(&out
);
717 extern "C" fn demangle_callback(input_ptr
: *const c_char
,
719 output_ptr
: *mut c_char
,
720 output_len
: size_t
) -> size_t
{
722 slice
::from_raw_parts(input_ptr
as *const u8, input_len
as usize)
725 let input
= match str::from_utf8(input
) {
730 let output
= unsafe {
731 slice
::from_raw_parts_mut(output_ptr
as *mut u8, output_len
as usize)
733 let mut cursor
= io
::Cursor
::new(output
);
735 let demangled
= match rustc_demangle
::try_demangle(input
) {
740 if let Err(_
) = write
!(cursor
, "{:#}", demangled
) {
741 // Possible only if provided buffer is not big enough
745 cursor
.position() as size_t
748 with_codegen(tm
, llmod
, config
.no_builtins
, |cpm
| {
749 llvm
::LLVMRustPrintModule(cpm
, llmod
, out
.as_ptr(), demangle_callback
);
750 llvm
::LLVMDisposePassManager(cpm
);
752 timeline
.record("ir");
755 if config
.emit_asm
|| asm_to_obj
{
756 let path
= cgcx
.output_filenames
.temp_path(OutputType
::Assembly
, module_name
);
758 // We can't use the same module for asm and binary output, because that triggers
759 // various errors like invalid IR or broken binaries, so we might have to clone the
760 // module to produce the asm output
761 let llmod
= if config
.emit_obj
{
762 llvm
::LLVMCloneModule(llmod
)
766 with_codegen(tm
, llmod
, config
.no_builtins
, |cpm
| {
767 write_output_file(diag_handler
, tm
, cpm
, llmod
, &path
,
768 llvm
::FileType
::AssemblyFile
)
771 llvm
::LLVMDisposeModule(llmod
);
773 timeline
.record("asm");
777 with_codegen(tm
, llmod
, config
.no_builtins
, |cpm
| {
778 write_output_file(diag_handler
, tm
, cpm
, llmod
, &obj_out
,
779 llvm
::FileType
::ObjectFile
)
781 timeline
.record("obj");
782 } else if asm_to_obj
{
783 let assembly
= cgcx
.output_filenames
.temp_path(OutputType
::Assembly
, module_name
);
784 run_assembler(cgcx
, diag_handler
, &assembly
, &obj_out
);
785 timeline
.record("asm_to_obj");
787 if !config
.emit_asm
&& !cgcx
.save_temps
{
788 drop(fs
::remove_file(&assembly
));
796 debug
!("copying bitcode {:?} to obj {:?}", bc_out
, obj_out
);
797 if let Err(e
) = link_or_copy(&bc_out
, &obj_out
) {
798 diag_handler
.err(&format
!("failed to copy bitcode to object file: {}", e
));
803 debug
!("removing_bitcode {:?}", bc_out
);
804 if let Err(e
) = fs
::remove_file(&bc_out
) {
805 diag_handler
.err(&format
!("failed to remove bitcode: {}", e
));
810 Ok(mtrans
.into_compiled_module(config
.emit_obj
,
812 config
.emit_bc_compressed
,
813 &cgcx
.output_filenames
))
816 /// Embed the bitcode of an LLVM module in the LLVM module itself.
818 /// This is done primarily for iOS where it appears to be standard to compile C
819 /// code at least with `-fembed-bitcode` which creates two sections in the
822 /// * __LLVM,__bitcode
823 /// * __LLVM,__cmdline
825 /// It appears *both* of these sections are necessary to get the linker to
826 /// recognize what's going on. For us though we just always throw in an empty
829 /// Furthermore debug/O1 builds don't actually embed bitcode but rather just
830 /// embed an empty section.
832 /// Basically all of this is us attempting to follow in the footsteps of clang
833 /// on iOS. See #35968 for lots more info.
834 unsafe fn embed_bitcode(cgcx
: &CodegenContext
,
837 bitcode
: Option
<&[u8]>) {
838 let llconst
= C_bytes_in_context(llcx
, bitcode
.unwrap_or(&[]));
839 let llglobal
= llvm
::LLVMAddGlobal(
841 val_ty(llconst
).to_ref(),
842 "rustc.embedded.module\0".as_ptr() as *const _
,
844 llvm
::LLVMSetInitializer(llglobal
, llconst
);
846 let is_apple
= cgcx
.opts
.target_triple
.triple().contains("-ios") ||
847 cgcx
.opts
.target_triple
.triple().contains("-darwin");
849 let section
= if is_apple
{
854 llvm
::LLVMSetSection(llglobal
, section
.as_ptr() as *const _
);
855 llvm
::LLVMRustSetLinkage(llglobal
, llvm
::Linkage
::PrivateLinkage
);
856 llvm
::LLVMSetGlobalConstant(llglobal
, llvm
::True
);
858 let llconst
= C_bytes_in_context(llcx
, &[]);
859 let llglobal
= llvm
::LLVMAddGlobal(
861 val_ty(llconst
).to_ref(),
862 "rustc.embedded.cmdline\0".as_ptr() as *const _
,
864 llvm
::LLVMSetInitializer(llglobal
, llconst
);
865 let section
= if is_apple
{
870 llvm
::LLVMSetSection(llglobal
, section
.as_ptr() as *const _
);
871 llvm
::LLVMRustSetLinkage(llglobal
, llvm
::Linkage
::PrivateLinkage
);
874 pub(crate) struct CompiledModules
{
875 pub modules
: Vec
<CompiledModule
>,
876 pub metadata_module
: CompiledModule
,
877 pub allocator_module
: Option
<CompiledModule
>,
880 fn need_crate_bitcode_for_rlib(sess
: &Session
) -> bool
{
881 sess
.crate_types
.borrow().contains(&config
::CrateTypeRlib
) &&
882 sess
.opts
.output_types
.contains_key(&OutputType
::Exe
)
885 pub fn start_async_translation(tcx
: TyCtxt
,
886 time_graph
: Option
<TimeGraph
>,
888 metadata
: EncodedMetadata
,
889 coordinator_receive
: Receiver
<Box
<Any
+ Send
>>,
891 -> OngoingCrateTranslation
{
893 let crate_name
= tcx
.crate_name(LOCAL_CRATE
);
894 let no_builtins
= attr
::contains_name(&tcx
.hir
.krate().attrs
, "no_builtins");
895 let subsystem
= attr
::first_attr_value_str_by_name(&tcx
.hir
.krate().attrs
,
896 "windows_subsystem");
897 let windows_subsystem
= subsystem
.map(|subsystem
| {
898 if subsystem
!= "windows" && subsystem
!= "console" {
899 tcx
.sess
.fatal(&format
!("invalid windows subsystem `{}`, only \
900 `windows` and `console` are allowed",
903 subsystem
.to_string()
906 let linker_info
= LinkerInfo
::new(tcx
);
907 let crate_info
= CrateInfo
::new(tcx
);
909 // Figure out what we actually need to build.
910 let mut modules_config
= ModuleConfig
::new(sess
.opts
.cg
.passes
.clone());
911 let mut metadata_config
= ModuleConfig
::new(vec
![]);
912 let mut allocator_config
= ModuleConfig
::new(vec
![]);
914 if let Some(ref sanitizer
) = sess
.opts
.debugging_opts
.sanitizer
{
916 Sanitizer
::Address
=> {
917 modules_config
.passes
.push("asan".to_owned());
918 modules_config
.passes
.push("asan-module".to_owned());
920 Sanitizer
::Memory
=> {
921 modules_config
.passes
.push("msan".to_owned())
923 Sanitizer
::Thread
=> {
924 modules_config
.passes
.push("tsan".to_owned())
930 if sess
.opts
.debugging_opts
.profile
{
931 modules_config
.passes
.push("insert-gcov-profiling".to_owned())
934 modules_config
.pgo_gen
= sess
.opts
.debugging_opts
.pgo_gen
.clone();
935 modules_config
.pgo_use
= sess
.opts
.debugging_opts
.pgo_use
.clone();
937 modules_config
.opt_level
= Some(get_llvm_opt_level(sess
.opts
.optimize
));
938 modules_config
.opt_size
= Some(get_llvm_opt_size(sess
.opts
.optimize
));
940 // Save all versions of the bytecode if we're saving our temporaries.
941 if sess
.opts
.cg
.save_temps
{
942 modules_config
.emit_no_opt_bc
= true;
943 modules_config
.emit_bc
= true;
944 modules_config
.emit_lto_bc
= true;
945 metadata_config
.emit_bc
= true;
946 allocator_config
.emit_bc
= true;
949 // Emit compressed bitcode files for the crate if we're emitting an rlib.
950 // Whenever an rlib is created, the bitcode is inserted into the archive in
951 // order to allow LTO against it.
952 if need_crate_bitcode_for_rlib(sess
) {
953 modules_config
.emit_bc_compressed
= true;
954 allocator_config
.emit_bc_compressed
= true;
957 modules_config
.no_integrated_as
= tcx
.sess
.opts
.cg
.no_integrated_as
||
958 tcx
.sess
.target
.target
.options
.no_integrated_as
;
960 for output_type
in sess
.opts
.output_types
.keys() {
962 OutputType
::Bitcode
=> { modules_config.emit_bc = true; }
963 OutputType
::LlvmAssembly
=> { modules_config.emit_ir = true; }
964 OutputType
::Assembly
=> {
965 modules_config
.emit_asm
= true;
966 // If we're not using the LLVM assembler, this function
967 // could be invoked specially with output_type_assembly, so
968 // in this case we still want the metadata object file.
969 if !sess
.opts
.output_types
.contains_key(&OutputType
::Assembly
) {
970 metadata_config
.emit_obj
= true;
971 allocator_config
.emit_obj
= true;
974 OutputType
::Object
=> { modules_config.emit_obj = true; }
975 OutputType
::Metadata
=> { metadata_config.emit_obj = true; }
977 modules_config
.emit_obj
= true;
978 metadata_config
.emit_obj
= true;
979 allocator_config
.emit_obj
= true;
981 OutputType
::Mir
=> {}
982 OutputType
::DepInfo
=> {}
986 modules_config
.set_flags(sess
, no_builtins
);
987 metadata_config
.set_flags(sess
, no_builtins
);
988 allocator_config
.set_flags(sess
, no_builtins
);
990 // Exclude metadata and allocator modules from time_passes output, since
991 // they throw off the "LLVM passes" measurement.
992 metadata_config
.time_passes
= false;
993 allocator_config
.time_passes
= false;
995 let (shared_emitter
, shared_emitter_main
) = SharedEmitter
::new();
996 let (trans_worker_send
, trans_worker_receive
) = channel();
998 let coordinator_thread
= start_executing_work(tcx
,
1002 coordinator_receive
,
1004 sess
.jobserver
.clone(),
1006 Arc
::new(modules_config
),
1007 Arc
::new(metadata_config
),
1008 Arc
::new(allocator_config
));
1010 OngoingCrateTranslation
{
1019 coordinator_send
: tcx
.tx_to_llvm_workers
.lock().clone(),
1020 trans_worker_receive
,
1021 shared_emitter_main
,
1022 future
: coordinator_thread
,
1023 output_filenames
: tcx
.output_filenames(LOCAL_CRATE
),
1027 fn copy_module_artifacts_into_incr_comp_cache(sess
: &Session
,
1028 dep_graph
: &DepGraph
,
1029 compiled_modules
: &CompiledModules
) {
1030 if sess
.opts
.incremental
.is_none() {
1034 for module
in compiled_modules
.modules
.iter() {
1035 let mut files
= vec
![];
1037 if let Some(ref path
) = module
.object
{
1038 files
.push((WorkProductFileKind
::Object
, path
.clone()));
1040 if let Some(ref path
) = module
.bytecode
{
1041 files
.push((WorkProductFileKind
::Bytecode
, path
.clone()));
1043 if let Some(ref path
) = module
.bytecode_compressed
{
1044 files
.push((WorkProductFileKind
::BytecodeCompressed
, path
.clone()));
1047 save_trans_partition(sess
, dep_graph
, &module
.name
, &files
);
1051 fn produce_final_output_artifacts(sess
: &Session
,
1052 compiled_modules
: &CompiledModules
,
1053 crate_output
: &OutputFilenames
) {
1054 let mut user_wants_bitcode
= false;
1055 let mut user_wants_objects
= false;
1057 // Produce final compile outputs.
1058 let copy_gracefully
= |from
: &Path
, to
: &Path
| {
1059 if let Err(e
) = fs
::copy(from
, to
) {
1060 sess
.err(&format
!("could not copy {:?} to {:?}: {}", from
, to
, e
));
1064 let copy_if_one_unit
= |output_type
: OutputType
,
1065 keep_numbered
: bool
| {
1066 if compiled_modules
.modules
.len() == 1 {
1067 // 1) Only one codegen unit. In this case it's no difficulty
1068 // to copy `foo.0.x` to `foo.x`.
1069 let module_name
= Some(&compiled_modules
.modules
[0].name
[..]);
1070 let path
= crate_output
.temp_path(output_type
, module_name
);
1071 copy_gracefully(&path
,
1072 &crate_output
.path(output_type
));
1073 if !sess
.opts
.cg
.save_temps
&& !keep_numbered
{
1074 // The user just wants `foo.x`, not `foo.#module-name#.x`.
1075 remove(sess
, &path
);
1078 let ext
= crate_output
.temp_path(output_type
, None
)
1085 if crate_output
.outputs
.contains_key(&output_type
) {
1086 // 2) Multiple codegen units, with `--emit foo=some_name`. We have
1087 // no good solution for this case, so warn the user.
1088 sess
.warn(&format
!("ignoring emit path because multiple .{} files \
1089 were produced", ext
));
1090 } else if crate_output
.single_output_file
.is_some() {
1091 // 3) Multiple codegen units, with `-o some_name`. We have
1092 // no good solution for this case, so warn the user.
1093 sess
.warn(&format
!("ignoring -o because multiple .{} files \
1094 were produced", ext
));
1096 // 4) Multiple codegen units, but no explicit name. We
1097 // just leave the `foo.0.x` files in place.
1098 // (We don't have to do any work in this case.)
1103 // Flag to indicate whether the user explicitly requested bitcode.
1104 // Otherwise, we produced it only as a temporary output, and will need
1105 // to get rid of it.
1106 for output_type
in crate_output
.outputs
.keys() {
1107 match *output_type
{
1108 OutputType
::Bitcode
=> {
1109 user_wants_bitcode
= true;
1110 // Copy to .bc, but always keep the .0.bc. There is a later
1111 // check to figure out if we should delete .0.bc files, or keep
1112 // them for making an rlib.
1113 copy_if_one_unit(OutputType
::Bitcode
, true);
1115 OutputType
::LlvmAssembly
=> {
1116 copy_if_one_unit(OutputType
::LlvmAssembly
, false);
1118 OutputType
::Assembly
=> {
1119 copy_if_one_unit(OutputType
::Assembly
, false);
1121 OutputType
::Object
=> {
1122 user_wants_objects
= true;
1123 copy_if_one_unit(OutputType
::Object
, true);
1126 OutputType
::Metadata
|
1128 OutputType
::DepInfo
=> {}
1132 // Clean up unwanted temporary files.
1134 // We create the following files by default:
1135 // - #crate#.#module-name#.bc
1136 // - #crate#.#module-name#.o
1137 // - #crate#.crate.metadata.bc
1138 // - #crate#.crate.metadata.o
1139 // - #crate#.o (linked from crate.##.o)
1140 // - #crate#.bc (copied from crate.##.bc)
1141 // We may create additional files if requested by the user (through
1142 // `-C save-temps` or `--emit=` flags).
1144 if !sess
.opts
.cg
.save_temps
{
1145 // Remove the temporary .#module-name#.o objects. If the user didn't
1146 // explicitly request bitcode (with --emit=bc), and the bitcode is not
1147 // needed for building an rlib, then we must remove .#module-name#.bc as
1150 // Specific rules for keeping .#module-name#.bc:
1151 // - If the user requested bitcode (`user_wants_bitcode`), and
1152 // codegen_units > 1, then keep it.
1153 // - If the user requested bitcode but codegen_units == 1, then we
1154 // can toss .#module-name#.bc because we copied it to .bc earlier.
1155 // - If we're not building an rlib and the user didn't request
1156 // bitcode, then delete .#module-name#.bc.
1157 // If you change how this works, also update back::link::link_rlib,
1158 // where .#module-name#.bc files are (maybe) deleted after making an
1160 let needs_crate_object
= crate_output
.outputs
.contains_key(&OutputType
::Exe
);
1162 let keep_numbered_bitcode
= user_wants_bitcode
&& sess
.codegen_units() > 1;
1164 let keep_numbered_objects
= needs_crate_object
||
1165 (user_wants_objects
&& sess
.codegen_units() > 1);
1167 for module
in compiled_modules
.modules
.iter() {
1168 if let Some(ref path
) = module
.object
{
1169 if !keep_numbered_objects
{
1174 if let Some(ref path
) = module
.bytecode
{
1175 if !keep_numbered_bitcode
{
1181 if !user_wants_bitcode
{
1182 if let Some(ref path
) = compiled_modules
.metadata_module
.bytecode
{
1183 remove(sess
, &path
);
1186 if let Some(ref allocator_module
) = compiled_modules
.allocator_module
{
1187 if let Some(ref path
) = allocator_module
.bytecode
{
1194 // We leave the following files around by default:
1196 // - #crate#.crate.metadata.o
1198 // These are used in linking steps and will be cleaned up afterward.
1201 pub(crate) fn dump_incremental_data(trans
: &CrateTranslation
) {
1202 println
!("[incremental] Re-using {} out of {} modules",
1203 trans
.modules
.iter().filter(|m
| m
.pre_existing
).count(),
1204 trans
.modules
.len());
1208 Optimize(ModuleTranslation
),
1209 LTO(lto
::LtoModuleTranslation
),
1213 fn kind(&self) -> ModuleKind
{
1215 WorkItem
::Optimize(ref m
) => m
.kind
,
1216 WorkItem
::LTO(_
) => ModuleKind
::Regular
,
1220 fn name(&self) -> String
{
1222 WorkItem
::Optimize(ref m
) => format
!("optimize: {}", m
.name
),
1223 WorkItem
::LTO(ref m
) => format
!("lto: {}", m
.name()),
1228 enum WorkItemResult
{
1229 Compiled(CompiledModule
),
1230 NeedsLTO(ModuleTranslation
),
1233 fn execute_work_item(cgcx
: &CodegenContext
,
1234 work_item
: WorkItem
,
1235 timeline
: &mut Timeline
)
1236 -> Result
<WorkItemResult
, FatalError
>
1238 let diag_handler
= cgcx
.create_diag_handler();
1239 let config
= cgcx
.config(work_item
.kind());
1240 let mtrans
= match work_item
{
1241 WorkItem
::Optimize(mtrans
) => mtrans
,
1242 WorkItem
::LTO(mut lto
) => {
1244 let module
= lto
.optimize(cgcx
, timeline
)?
;
1245 let module
= codegen(cgcx
, &diag_handler
, module
, config
, timeline
)?
;
1246 return Ok(WorkItemResult
::Compiled(module
))
1250 let module_name
= mtrans
.name
.clone();
1252 let pre_existing
= match mtrans
.source
{
1253 ModuleSource
::Translated(_
) => None
,
1254 ModuleSource
::Preexisting(ref wp
) => Some(wp
.clone()),
1257 if let Some(wp
) = pre_existing
{
1258 let incr_comp_session_dir
= cgcx
.incr_comp_session_dir
1261 let name
= &mtrans
.name
;
1262 let mut object
= None
;
1263 let mut bytecode
= None
;
1264 let mut bytecode_compressed
= None
;
1265 for (kind
, saved_file
) in wp
.saved_files
{
1266 let obj_out
= match kind
{
1267 WorkProductFileKind
::Object
=> {
1268 let path
= cgcx
.output_filenames
.temp_path(OutputType
::Object
, Some(name
));
1269 object
= Some(path
.clone());
1272 WorkProductFileKind
::Bytecode
=> {
1273 let path
= cgcx
.output_filenames
.temp_path(OutputType
::Bitcode
, Some(name
));
1274 bytecode
= Some(path
.clone());
1277 WorkProductFileKind
::BytecodeCompressed
=> {
1278 let path
= cgcx
.output_filenames
.temp_path(OutputType
::Bitcode
, Some(name
))
1279 .with_extension(RLIB_BYTECODE_EXTENSION
);
1280 bytecode_compressed
= Some(path
.clone());
1284 let source_file
= in_incr_comp_dir(&incr_comp_session_dir
,
1286 debug
!("copying pre-existing module `{}` from {:?} to {}",
1290 match link_or_copy(&source_file
, &obj_out
) {
1293 diag_handler
.err(&format
!("unable to copy {} to {}: {}",
1294 source_file
.display(),
1300 assert_eq
!(object
.is_some(), config
.emit_obj
);
1301 assert_eq
!(bytecode
.is_some(), config
.emit_bc
);
1302 assert_eq
!(bytecode_compressed
.is_some(), config
.emit_bc_compressed
);
1304 Ok(WorkItemResult
::Compiled(CompiledModule
{
1305 llmod_id
: mtrans
.llmod_id
.clone(),
1307 kind
: ModuleKind
::Regular
,
1311 bytecode_compressed
,
1314 debug
!("llvm-optimizing {:?}", module_name
);
1317 optimize(cgcx
, &diag_handler
, &mtrans
, config
, timeline
)?
;
1319 // After we've done the initial round of optimizations we need to
1320 // decide whether to synchronously codegen this module or ship it
1321 // back to the coordinator thread for further LTO processing (which
1322 // has to wait for all the initial modules to be optimized).
1324 // Here we dispatch based on the `cgcx.lto` and kind of module we're
1326 let needs_lto
= match cgcx
.lto
{
1329 // Here we've got a full crate graph LTO requested. We ignore
1330 // this, however, if the crate type is only an rlib as there's
1331 // no full crate graph to process, that'll happen later.
1333 // This use case currently comes up primarily for targets that
1334 // require LTO so the request for LTO is always unconditionally
1335 // passed down to the backend, but we don't actually want to do
1336 // anything about it yet until we've got a final product.
1337 Lto
::Yes
| Lto
::Fat
| Lto
::Thin
=> {
1338 cgcx
.crate_types
.len() != 1 ||
1339 cgcx
.crate_types
[0] != config
::CrateTypeRlib
1342 // When we're automatically doing ThinLTO for multi-codegen-unit
1343 // builds we don't actually want to LTO the allocator modules if
1344 // it shows up. This is due to various linker shenanigans that
1345 // we'll encounter later.
1347 // Additionally here's where we also factor in the current LLVM
1348 // version. If it doesn't support ThinLTO we skip this.
1350 mtrans
.kind
!= ModuleKind
::Allocator
&&
1351 llvm
::LLVMRustThinLTOAvailable()
1355 // Metadata modules never participate in LTO regardless of the lto
1357 let needs_lto
= needs_lto
&& mtrans
.kind
!= ModuleKind
::Metadata
;
1359 // Don't run LTO passes when cross-lang LTO is enabled. The linker
1360 // will do that for us in this case.
1361 let needs_lto
= needs_lto
&& !cgcx
.opts
.debugging_opts
.cross_lang_lto
;
1364 Ok(WorkItemResult
::NeedsLTO(mtrans
))
1366 let module
= codegen(cgcx
, &diag_handler
, mtrans
, config
, timeline
)?
;
1367 Ok(WorkItemResult
::Compiled(module
))
1374 Token(io
::Result
<Acquired
>),
1376 result
: ModuleTranslation
,
1380 result
: Result
<CompiledModule
, ()>,
1384 llvm_work_item
: WorkItem
,
1387 TranslationComplete
,
1393 code
: Option
<DiagnosticId
>,
1397 #[derive(PartialEq, Clone, Copy, Debug)]
1398 enum MainThreadWorkerState
{
1404 fn start_executing_work(tcx
: TyCtxt
,
1405 crate_info
: &CrateInfo
,
1406 shared_emitter
: SharedEmitter
,
1407 trans_worker_send
: Sender
<Message
>,
1408 coordinator_receive
: Receiver
<Box
<Any
+ Send
>>,
1411 time_graph
: Option
<TimeGraph
>,
1412 modules_config
: Arc
<ModuleConfig
>,
1413 metadata_config
: Arc
<ModuleConfig
>,
1414 allocator_config
: Arc
<ModuleConfig
>)
1415 -> thread
::JoinHandle
<Result
<CompiledModules
, ()>> {
1416 let coordinator_send
= tcx
.tx_to_llvm_workers
.lock().clone();
1417 let sess
= tcx
.sess
;
1419 // Compute the set of symbols we need to retain when doing LTO (if we need to)
1420 let exported_symbols
= {
1421 let mut exported_symbols
= FxHashMap();
1423 let copy_symbols
= |cnum
| {
1424 let symbols
= tcx
.exported_symbols(cnum
)
1426 .map(|&(s
, lvl
)| (s
.symbol_name(tcx
).to_string(), lvl
))
1434 exported_symbols
.insert(LOCAL_CRATE
, copy_symbols(LOCAL_CRATE
));
1435 Some(Arc
::new(exported_symbols
))
1437 Lto
::Yes
| Lto
::Fat
| Lto
::Thin
=> {
1438 exported_symbols
.insert(LOCAL_CRATE
, copy_symbols(LOCAL_CRATE
));
1439 for &cnum
in tcx
.crates().iter() {
1440 exported_symbols
.insert(cnum
, copy_symbols(cnum
));
1442 Some(Arc
::new(exported_symbols
))
1447 // First up, convert our jobserver into a helper thread so we can use normal
1448 // mpsc channels to manage our messages and such.
1449 // After we've requested tokens then we'll, when we can,
1450 // get tokens on `coordinator_receive` which will
1451 // get managed in the main loop below.
1452 let coordinator_send2
= coordinator_send
.clone();
1453 let helper
= jobserver
.into_helper_thread(move |token
| {
1454 drop(coordinator_send2
.send(Box
::new(Message
::Token(token
))));
1455 }).expect("failed to spawn helper thread");
1457 let mut each_linked_rlib_for_lto
= Vec
::new();
1458 drop(link
::each_linked_rlib(sess
, crate_info
, &mut |cnum
, path
| {
1459 if link
::ignored_for_lto(sess
, crate_info
, cnum
) {
1462 each_linked_rlib_for_lto
.push((cnum
, path
.to_path_buf()));
1465 let assembler_cmd
= if modules_config
.no_integrated_as
{
1466 // HACK: currently we use linker (gcc) as our assembler
1467 let (name
, mut cmd
) = get_linker(sess
);
1468 cmd
.args(&sess
.target
.target
.options
.asm_args
);
1469 Some(Arc
::new(AssemblerCommand
{
1477 let cgcx
= CodegenContext
{
1478 crate_types
: sess
.crate_types
.borrow().clone(),
1479 each_linked_rlib_for_lto
,
1481 no_landing_pads
: sess
.no_landing_pads(),
1482 fewer_names
: sess
.fewer_names(),
1483 save_temps
: sess
.opts
.cg
.save_temps
,
1484 opts
: Arc
::new(sess
.opts
.clone()),
1485 time_passes
: sess
.time_passes(),
1487 plugin_passes
: sess
.plugin_llvm_passes
.borrow().clone(),
1488 remark
: sess
.opts
.cg
.remark
.clone(),
1490 incr_comp_session_dir
: sess
.incr_comp_session_dir_opt().map(|r
| r
.clone()),
1492 diag_emitter
: shared_emitter
.clone(),
1494 output_filenames
: tcx
.output_filenames(LOCAL_CRATE
),
1495 regular_module_config
: modules_config
,
1496 metadata_module_config
: metadata_config
,
1497 allocator_module_config
: allocator_config
,
1498 tm_factory
: target_machine_factory(tcx
.sess
, false),
1500 msvc_imps_needed
: msvc_imps_needed(tcx
),
1501 target_pointer_width
: tcx
.sess
.target
.target
.target_pointer_width
.clone(),
1502 debuginfo
: tcx
.sess
.opts
.debuginfo
,
1506 // This is the "main loop" of parallel work happening for parallel codegen.
1507 // It's here that we manage parallelism, schedule work, and work with
1508 // messages coming from clients.
1510 // There are a few environmental pre-conditions that shape how the system
1513 // - Error reporting only can happen on the main thread because that's the
1514 // only place where we have access to the compiler `Session`.
1515 // - LLVM work can be done on any thread.
1516 // - Translation can only happen on the main thread.
1517 // - Each thread doing substantial work most be in possession of a `Token`
1518 // from the `Jobserver`.
1519 // - The compiler process always holds one `Token`. Any additional `Tokens`
1520 // have to be requested from the `Jobserver`.
1524 // The error reporting restriction is handled separately from the rest: We
1525 // set up a `SharedEmitter` the holds an open channel to the main thread.
1526 // When an error occurs on any thread, the shared emitter will send the
1527 // error message to the receiver main thread (`SharedEmitterMain`). The
1528 // main thread will periodically query this error message queue and emit
1529 // any error messages it has received. It might even abort compilation if
1530 // has received a fatal error. In this case we rely on all other threads
1531 // being torn down automatically with the main thread.
1532 // Since the main thread will often be busy doing translation work, error
1533 // reporting will be somewhat delayed, since the message queue can only be
1534 // checked in between to work packages.
1536 // Work Processing Infrastructure
1537 // ==============================
1538 // The work processing infrastructure knows three major actors:
1540 // - the coordinator thread,
1541 // - the main thread, and
1542 // - LLVM worker threads
1544 // The coordinator thread is running a message loop. It instructs the main
1545 // thread about what work to do when, and it will spawn off LLVM worker
1546 // threads as open LLVM WorkItems become available.
1548 // The job of the main thread is to translate CGUs into LLVM work package
1549 // (since the main thread is the only thread that can do this). The main
1550 // thread will block until it receives a message from the coordinator, upon
1551 // which it will translate one CGU, send it to the coordinator and block
1552 // again. This way the coordinator can control what the main thread is
1555 // The coordinator keeps a queue of LLVM WorkItems, and when a `Token` is
1556 // available, it will spawn off a new LLVM worker thread and let it process
1557 // that a WorkItem. When a LLVM worker thread is done with its WorkItem,
1558 // it will just shut down, which also frees all resources associated with
1559 // the given LLVM module, and sends a message to the coordinator that the
1560 // has been completed.
1564 // The scheduler's goal is to minimize the time it takes to complete all
1565 // work there is, however, we also want to keep memory consumption low
1566 // if possible. These two goals are at odds with each other: If memory
1567 // consumption were not an issue, we could just let the main thread produce
1568 // LLVM WorkItems at full speed, assuring maximal utilization of
1569 // Tokens/LLVM worker threads. However, since translation usual is faster
1570 // than LLVM processing, the queue of LLVM WorkItems would fill up and each
1571 // WorkItem potentially holds on to a substantial amount of memory.
1573 // So the actual goal is to always produce just enough LLVM WorkItems as
1574 // not to starve our LLVM worker threads. That means, once we have enough
1575 // WorkItems in our queue, we can block the main thread, so it does not
1576 // produce more until we need them.
1578 // Doing LLVM Work on the Main Thread
1579 // ----------------------------------
1580 // Since the main thread owns the compiler processes implicit `Token`, it is
1581 // wasteful to keep it blocked without doing any work. Therefore, what we do
1582 // in this case is: We spawn off an additional LLVM worker thread that helps
1583 // reduce the queue. The work it is doing corresponds to the implicit
1584 // `Token`. The coordinator will mark the main thread as being busy with
1585 // LLVM work. (The actual work happens on another OS thread but we just care
1586 // about `Tokens`, not actual threads).
1588 // When any LLVM worker thread finishes while the main thread is marked as
1589 // "busy with LLVM work", we can do a little switcheroo: We give the Token
1590 // of the just finished thread to the LLVM worker thread that is working on
1591 // behalf of the main thread's implicit Token, thus freeing up the main
1592 // thread again. The coordinator can then again decide what the main thread
1593 // should do. This allows the coordinator to make decisions at more points
1596 // Striking a Balance between Throughput and Memory Consumption
1597 // ------------------------------------------------------------
1598 // Since our two goals, (1) use as many Tokens as possible and (2) keep
1599 // memory consumption as low as possible, are in conflict with each other,
1600 // we have to find a trade off between them. Right now, the goal is to keep
1601 // all workers busy, which means that no worker should find the queue empty
1602 // when it is ready to start.
1603 // How do we do achieve this? Good question :) We actually never know how
1604 // many `Tokens` are potentially available so it's hard to say how much to
1605 // fill up the queue before switching the main thread to LLVM work. Also we
1606 // currently don't have a means to estimate how long a running LLVM worker
1607 // will still be busy with it's current WorkItem. However, we know the
1608 // maximal count of available Tokens that makes sense (=the number of CPU
1609 // cores), so we can take a conservative guess. The heuristic we use here
1610 // is implemented in the `queue_full_enough()` function.
1612 // Some Background on Jobservers
1613 // -----------------------------
1614 // It's worth also touching on the management of parallelism here. We don't
1615 // want to just spawn a thread per work item because while that's optimal
1616 // parallelism it may overload a system with too many threads or violate our
1617 // configuration for the maximum amount of cpu to use for this process. To
1618 // manage this we use the `jobserver` crate.
1620 // Job servers are an artifact of GNU make and are used to manage
1621 // parallelism between processes. A jobserver is a glorified IPC semaphore
1622 // basically. Whenever we want to run some work we acquire the semaphore,
1623 // and whenever we're done with that work we release the semaphore. In this
1624 // manner we can ensure that the maximum number of parallel workers is
1625 // capped at any one point in time.
1627 // LTO and the coordinator thread
1628 // ------------------------------
1630 // The final job the coordinator thread is responsible for is managing LTO
1631 // and how that works. When LTO is requested what we'll to is collect all
1632 // optimized LLVM modules into a local vector on the coordinator. Once all
1633 // modules have been translated and optimized we hand this to the `lto`
1634 // module for further optimization. The `lto` module will return back a list
1635 // of more modules to work on, which the coordinator will continue to spawn
1638 // Each LLVM module is automatically sent back to the coordinator for LTO if
1639 // necessary. There's already optimizations in place to avoid sending work
1640 // back to the coordinator if LTO isn't requested.
1641 return thread
::spawn(move || {
1642 // We pretend to be within the top-level LLVM time-passes task here:
1645 let max_workers
= ::num_cpus
::get();
1646 let mut worker_id_counter
= 0;
1647 let mut free_worker_ids
= Vec
::new();
1648 let mut get_worker_id
= |free_worker_ids
: &mut Vec
<usize>| {
1649 if let Some(id
) = free_worker_ids
.pop() {
1652 let id
= worker_id_counter
;
1653 worker_id_counter
+= 1;
1658 // This is where we collect codegen units that have gone all the way
1659 // through translation and LLVM.
1660 let mut compiled_modules
= vec
![];
1661 let mut compiled_metadata_module
= None
;
1662 let mut compiled_allocator_module
= None
;
1663 let mut needs_lto
= Vec
::new();
1664 let mut started_lto
= false;
1666 // This flag tracks whether all items have gone through translations
1667 let mut translation_done
= false;
1669 // This is the queue of LLVM work items that still need processing.
1670 let mut work_items
= Vec
::<(WorkItem
, u64)>::new();
1672 // This are the Jobserver Tokens we currently hold. Does not include
1673 // the implicit Token the compiler process owns no matter what.
1674 let mut tokens
= Vec
::new();
1676 let mut main_thread_worker_state
= MainThreadWorkerState
::Idle
;
1677 let mut running
= 0;
1679 let mut llvm_start_time
= None
;
1681 // Run the message loop while there's still anything that needs message
1683 while !translation_done
||
1684 work_items
.len() > 0 ||
1686 needs_lto
.len() > 0 ||
1687 main_thread_worker_state
!= MainThreadWorkerState
::Idle
{
1689 // While there are still CGUs to be translated, the coordinator has
1690 // to decide how to utilize the compiler processes implicit Token:
1691 // For translating more CGU or for running them through LLVM.
1692 if !translation_done
{
1693 if main_thread_worker_state
== MainThreadWorkerState
::Idle
{
1694 if !queue_full_enough(work_items
.len(), running
, max_workers
) {
1695 // The queue is not full enough, translate more items:
1696 if let Err(_
) = trans_worker_send
.send(Message
::TranslateItem
) {
1697 panic
!("Could not send Message::TranslateItem to main thread")
1699 main_thread_worker_state
= MainThreadWorkerState
::Translating
;
1701 // The queue is full enough to not let the worker
1702 // threads starve. Use the implicit Token to do some
1704 let (item
, _
) = work_items
.pop()
1705 .expect("queue empty - queue_full_enough() broken?");
1706 let cgcx
= CodegenContext
{
1707 worker
: get_worker_id(&mut free_worker_ids
),
1710 maybe_start_llvm_timer(cgcx
.config(item
.kind()),
1711 &mut llvm_start_time
);
1712 main_thread_worker_state
= MainThreadWorkerState
::LLVMing
;
1713 spawn_work(cgcx
, item
);
1717 // If we've finished everything related to normal translation
1718 // then it must be the case that we've got some LTO work to do.
1719 // Perform the serial work here of figuring out what we're
1720 // going to LTO and then push a bunch of work items onto our
1722 if work_items
.len() == 0 &&
1724 main_thread_worker_state
== MainThreadWorkerState
::Idle
{
1725 assert
!(!started_lto
);
1726 assert
!(needs_lto
.len() > 0);
1728 let modules
= mem
::replace(&mut needs_lto
, Vec
::new());
1729 for (work
, cost
) in generate_lto_work(&cgcx
, modules
) {
1730 let insertion_index
= work_items
1731 .binary_search_by_key(&cost
, |&(_
, cost
)| cost
)
1732 .unwrap_or_else(|e
| e
);
1733 work_items
.insert(insertion_index
, (work
, cost
));
1734 helper
.request_token();
1738 // In this branch, we know that everything has been translated,
1739 // so it's just a matter of determining whether the implicit
1740 // Token is free to use for LLVM work.
1741 match main_thread_worker_state
{
1742 MainThreadWorkerState
::Idle
=> {
1743 if let Some((item
, _
)) = work_items
.pop() {
1744 let cgcx
= CodegenContext
{
1745 worker
: get_worker_id(&mut free_worker_ids
),
1748 maybe_start_llvm_timer(cgcx
.config(item
.kind()),
1749 &mut llvm_start_time
);
1750 main_thread_worker_state
= MainThreadWorkerState
::LLVMing
;
1751 spawn_work(cgcx
, item
);
1753 // There is no unstarted work, so let the main thread
1754 // take over for a running worker. Otherwise the
1755 // implicit token would just go to waste.
1756 // We reduce the `running` counter by one. The
1757 // `tokens.truncate()` below will take care of
1758 // giving the Token back.
1759 debug_assert
!(running
> 0);
1761 main_thread_worker_state
= MainThreadWorkerState
::LLVMing
;
1764 MainThreadWorkerState
::Translating
=> {
1765 bug
!("trans worker should not be translating after \
1766 translation was already completed")
1768 MainThreadWorkerState
::LLVMing
=> {
1769 // Already making good use of that token
1774 // Spin up what work we can, only doing this while we've got available
1775 // parallelism slots and work left to spawn.
1776 while work_items
.len() > 0 && running
< tokens
.len() {
1777 let (item
, _
) = work_items
.pop().unwrap();
1779 maybe_start_llvm_timer(cgcx
.config(item
.kind()),
1780 &mut llvm_start_time
);
1782 let cgcx
= CodegenContext
{
1783 worker
: get_worker_id(&mut free_worker_ids
),
1787 spawn_work(cgcx
, item
);
1791 // Relinquish accidentally acquired extra tokens
1792 tokens
.truncate(running
);
1794 let msg
= coordinator_receive
.recv().unwrap();
1795 match *msg
.downcast
::<Message
>().ok().unwrap() {
1796 // Save the token locally and the next turn of the loop will use
1797 // this to spawn a new unit of work, or it may get dropped
1798 // immediately if we have no more work to spawn.
1799 Message
::Token(token
) => {
1804 if main_thread_worker_state
== MainThreadWorkerState
::LLVMing
{
1805 // If the main thread token is used for LLVM work
1806 // at the moment, we turn that thread into a regular
1807 // LLVM worker thread, so the main thread is free
1808 // to react to translation demand.
1809 main_thread_worker_state
= MainThreadWorkerState
::Idle
;
1814 let msg
= &format
!("failed to acquire jobserver token: {}", e
);
1815 shared_emitter
.fatal(msg
);
1816 // Exit the coordinator thread
1822 Message
::TranslationDone { llvm_work_item, cost }
=> {
1823 // We keep the queue sorted by estimated processing cost,
1824 // so that more expensive items are processed earlier. This
1825 // is good for throughput as it gives the main thread more
1826 // time to fill up the queue and it avoids scheduling
1827 // expensive items to the end.
1828 // Note, however, that this is not ideal for memory
1829 // consumption, as LLVM module sizes are not evenly
1831 let insertion_index
=
1832 work_items
.binary_search_by_key(&cost
, |&(_
, cost
)| cost
);
1833 let insertion_index
= match insertion_index
{
1834 Ok(idx
) | Err(idx
) => idx
1836 work_items
.insert(insertion_index
, (llvm_work_item
, cost
));
1838 helper
.request_token();
1839 assert_eq
!(main_thread_worker_state
,
1840 MainThreadWorkerState
::Translating
);
1841 main_thread_worker_state
= MainThreadWorkerState
::Idle
;
1844 Message
::TranslationComplete
=> {
1845 translation_done
= true;
1846 assert_eq
!(main_thread_worker_state
,
1847 MainThreadWorkerState
::Translating
);
1848 main_thread_worker_state
= MainThreadWorkerState
::Idle
;
1851 // If a thread exits successfully then we drop a token associated
1852 // with that worker and update our `running` count. We may later
1853 // re-acquire a token to continue running more work. We may also not
1854 // actually drop a token here if the worker was running with an
1855 // "ephemeral token"
1857 // Note that if the thread failed that means it panicked, so we
1858 // abort immediately.
1859 Message
::Done { result: Ok(compiled_module), worker_id }
=> {
1860 if main_thread_worker_state
== MainThreadWorkerState
::LLVMing
{
1861 main_thread_worker_state
= MainThreadWorkerState
::Idle
;
1866 free_worker_ids
.push(worker_id
);
1868 match compiled_module
.kind
{
1869 ModuleKind
::Regular
=> {
1870 compiled_modules
.push(compiled_module
);
1872 ModuleKind
::Metadata
=> {
1873 assert
!(compiled_metadata_module
.is_none());
1874 compiled_metadata_module
= Some(compiled_module
);
1876 ModuleKind
::Allocator
=> {
1877 assert
!(compiled_allocator_module
.is_none());
1878 compiled_allocator_module
= Some(compiled_module
);
1882 Message
::NeedsLTO { result, worker_id }
=> {
1883 assert
!(!started_lto
);
1884 if main_thread_worker_state
== MainThreadWorkerState
::LLVMing
{
1885 main_thread_worker_state
= MainThreadWorkerState
::Idle
;
1890 free_worker_ids
.push(worker_id
);
1891 needs_lto
.push(result
);
1893 Message
::Done { result: Err(()), worker_id: _ }
=> {
1894 shared_emitter
.fatal("aborting due to worker thread failure");
1895 // Exit the coordinator thread
1898 Message
::TranslateItem
=> {
1899 bug
!("the coordinator should not receive translation requests")
1904 if let Some(llvm_start_time
) = llvm_start_time
{
1905 let total_llvm_time
= Instant
::now().duration_since(llvm_start_time
);
1906 // This is the top-level timing for all of LLVM, set the time-depth
1909 print_time_passes_entry(cgcx
.time_passes
,
1914 // Regardless of what order these modules completed in, report them to
1915 // the backend in the same order every time to ensure that we're handing
1916 // out deterministic results.
1917 compiled_modules
.sort_by(|a
, b
| a
.name
.cmp(&b
.name
));
1919 let compiled_metadata_module
= compiled_metadata_module
1920 .expect("Metadata module not compiled?");
1922 Ok(CompiledModules
{
1923 modules
: compiled_modules
,
1924 metadata_module
: compiled_metadata_module
,
1925 allocator_module
: compiled_allocator_module
,
1929 // A heuristic that determines if we have enough LLVM WorkItems in the
1930 // queue so that the main thread can do LLVM work instead of translation
1931 fn queue_full_enough(items_in_queue
: usize,
1932 workers_running
: usize,
1933 max_workers
: usize) -> bool
{
1935 items_in_queue
> 0 &&
1936 items_in_queue
>= max_workers
.saturating_sub(workers_running
/ 2)
1939 fn maybe_start_llvm_timer(config
: &ModuleConfig
,
1940 llvm_start_time
: &mut Option
<Instant
>) {
1941 // We keep track of the -Ztime-passes output manually,
1942 // since the closure-based interface does not fit well here.
1943 if config
.time_passes
{
1944 if llvm_start_time
.is_none() {
1945 *llvm_start_time
= Some(Instant
::now());
1951 pub const TRANS_WORKER_ID
: usize = ::std
::usize::MAX
;
1952 pub const TRANS_WORKER_TIMELINE
: time_graph
::TimelineId
=
1953 time_graph
::TimelineId(TRANS_WORKER_ID
);
1954 pub const TRANS_WORK_PACKAGE_KIND
: time_graph
::WorkPackageKind
=
1955 time_graph
::WorkPackageKind(&["#DE9597", "#FED1D3", "#FDC5C7", "#B46668", "#88494B"]);
1956 const LLVM_WORK_PACKAGE_KIND
: time_graph
::WorkPackageKind
=
1957 time_graph
::WorkPackageKind(&["#7DB67A", "#C6EEC4", "#ACDAAA", "#579354", "#3E6F3C"]);
1959 fn spawn_work(cgcx
: CodegenContext
, work
: WorkItem
) {
1960 let depth
= time_depth();
1962 thread
::spawn(move || {
1963 set_time_depth(depth
);
1965 // Set up a destructor which will fire off a message that we're done as
1968 coordinator_send
: Sender
<Box
<Any
+ Send
>>,
1969 result
: Option
<WorkItemResult
>,
1972 impl Drop
for Bomb
{
1973 fn drop(&mut self) {
1974 let worker_id
= self.worker_id
;
1975 let msg
= match self.result
.take() {
1976 Some(WorkItemResult
::Compiled(m
)) => {
1977 Message
::Done { result: Ok(m), worker_id }
1979 Some(WorkItemResult
::NeedsLTO(m
)) => {
1980 Message
::NeedsLTO { result: m, worker_id }
1982 None
=> Message
::Done { result: Err(()), worker_id }
1984 drop(self.coordinator_send
.send(Box
::new(msg
)));
1988 let mut bomb
= Bomb
{
1989 coordinator_send
: cgcx
.coordinator_send
.clone(),
1991 worker_id
: cgcx
.worker
,
1994 // Execute the work itself, and if it finishes successfully then flag
1995 // ourselves as a success as well.
1997 // Note that we ignore any `FatalError` coming out of `execute_work_item`,
1998 // as a diagnostic was already sent off to the main thread - just
1999 // surface that there was an error in this worker.
2001 let timeline
= cgcx
.time_graph
.as_ref().map(|tg
| {
2002 tg
.start(time_graph
::TimelineId(cgcx
.worker
),
2003 LLVM_WORK_PACKAGE_KIND
,
2006 let mut timeline
= timeline
.unwrap_or(Timeline
::noop());
2007 execute_work_item(&cgcx
, work
, &mut timeline
).ok()
2012 pub fn run_assembler(cgcx
: &CodegenContext
, handler
: &Handler
, assembly
: &Path
, object
: &Path
) {
2013 let assembler
= cgcx
.assembler_cmd
2015 .expect("cgcx.assembler_cmd is missing?");
2017 let pname
= &assembler
.name
;
2018 let mut cmd
= assembler
.cmd
.clone();
2019 cmd
.arg("-c").arg("-o").arg(object
).arg(assembly
);
2020 debug
!("{:?}", cmd
);
2022 match cmd
.output() {
2024 if !prog
.status
.success() {
2025 let mut note
= prog
.stderr
.clone();
2026 note
.extend_from_slice(&prog
.stdout
);
2028 handler
.struct_err(&format
!("linking with `{}` failed: {}",
2031 .note(&format
!("{:?}", &cmd
))
2032 .note(str::from_utf8(¬e
[..]).unwrap())
2034 handler
.abort_if_errors();
2038 handler
.err(&format
!("could not exec the linker `{}`: {}", pname
.display(), e
));
2039 handler
.abort_if_errors();
2044 pub unsafe fn with_llvm_pmb(llmod
: ModuleRef
,
2045 config
: &ModuleConfig
,
2046 opt_level
: llvm
::CodeGenOptLevel
,
2047 f
: &mut FnMut(llvm
::PassManagerBuilderRef
)) {
2050 // Create the PassManagerBuilder for LLVM. We configure it with
2051 // reasonable defaults and prepare it to actually populate the pass
2053 let builder
= llvm
::LLVMPassManagerBuilderCreate();
2054 let opt_size
= config
.opt_size
.unwrap_or(llvm
::CodeGenOptSizeNone
);
2055 let inline_threshold
= config
.inline_threshold
;
2057 let pgo_gen_path
= config
.pgo_gen
.as_ref().map(|s
| {
2058 let s
= if s
.is_empty() { "default_%m.profraw" }
else { s }
;
2059 CString
::new(s
.as_bytes()).unwrap()
2062 let pgo_use_path
= if config
.pgo_use
.is_empty() {
2065 Some(CString
::new(config
.pgo_use
.as_bytes()).unwrap())
2068 llvm
::LLVMRustConfigurePassManagerBuilder(
2071 config
.merge_functions
,
2072 config
.vectorize_slp
,
2073 config
.vectorize_loop
,
2074 pgo_gen_path
.as_ref().map_or(ptr
::null(), |s
| s
.as_ptr()),
2075 pgo_use_path
.as_ref().map_or(ptr
::null(), |s
| s
.as_ptr()),
2078 llvm
::LLVMPassManagerBuilderSetSizeLevel(builder
, opt_size
as u32);
2080 if opt_size
!= llvm
::CodeGenOptSizeNone
{
2081 llvm
::LLVMPassManagerBuilderSetDisableUnrollLoops(builder
, 1);
2084 llvm
::LLVMRustAddBuilderLibraryInfo(builder
, llmod
, config
.no_builtins
);
2086 // Here we match what clang does (kinda). For O0 we only inline
2087 // always-inline functions (but don't add lifetime intrinsics), at O1 we
2088 // inline with lifetime intrinsics, and O2+ we add an inliner with a
2089 // thresholds copied from clang.
2090 match (opt_level
, opt_size
, inline_threshold
) {
2092 llvm
::LLVMPassManagerBuilderUseInlinerWithThreshold(builder
, t
as u32);
2094 (llvm
::CodeGenOptLevel
::Aggressive
, ..) => {
2095 llvm
::LLVMPassManagerBuilderUseInlinerWithThreshold(builder
, 275);
2097 (_
, llvm
::CodeGenOptSizeDefault
, _
) => {
2098 llvm
::LLVMPassManagerBuilderUseInlinerWithThreshold(builder
, 75);
2100 (_
, llvm
::CodeGenOptSizeAggressive
, _
) => {
2101 llvm
::LLVMPassManagerBuilderUseInlinerWithThreshold(builder
, 25);
2103 (llvm
::CodeGenOptLevel
::None
, ..) => {
2104 llvm
::LLVMRustAddAlwaysInlinePass(builder
, false);
2106 (llvm
::CodeGenOptLevel
::Less
, ..) => {
2107 llvm
::LLVMRustAddAlwaysInlinePass(builder
, true);
2109 (llvm
::CodeGenOptLevel
::Default
, ..) => {
2110 llvm
::LLVMPassManagerBuilderUseInlinerWithThreshold(builder
, 225);
2112 (llvm
::CodeGenOptLevel
::Other
, ..) => {
2113 bug
!("CodeGenOptLevel::Other selected")
2118 llvm
::LLVMPassManagerBuilderDispose(builder
);
2122 enum SharedEmitterMessage
{
2123 Diagnostic(Diagnostic
),
2124 InlineAsmError(u32, String
),
2130 pub struct SharedEmitter
{
2131 sender
: Sender
<SharedEmitterMessage
>,
2134 pub struct SharedEmitterMain
{
2135 receiver
: Receiver
<SharedEmitterMessage
>,
2138 impl SharedEmitter
{
2139 pub fn new() -> (SharedEmitter
, SharedEmitterMain
) {
2140 let (sender
, receiver
) = channel();
2142 (SharedEmitter { sender }
, SharedEmitterMain { receiver }
)
2145 fn inline_asm_error(&self, cookie
: u32, msg
: String
) {
2146 drop(self.sender
.send(SharedEmitterMessage
::InlineAsmError(cookie
, msg
)));
2149 fn fatal(&self, msg
: &str) {
2150 drop(self.sender
.send(SharedEmitterMessage
::Fatal(msg
.to_string())));
2154 impl Emitter
for SharedEmitter
{
2155 fn emit(&mut self, db
: &DiagnosticBuilder
) {
2156 drop(self.sender
.send(SharedEmitterMessage
::Diagnostic(Diagnostic
{
2158 code
: db
.code
.clone(),
2161 for child
in &db
.children
{
2162 drop(self.sender
.send(SharedEmitterMessage
::Diagnostic(Diagnostic
{
2163 msg
: child
.message(),
2168 drop(self.sender
.send(SharedEmitterMessage
::AbortIfErrors
));
2172 impl SharedEmitterMain
{
2173 pub fn check(&self, sess
: &Session
, blocking
: bool
) {
2175 let message
= if blocking
{
2176 match self.receiver
.recv() {
2177 Ok(message
) => Ok(message
),
2181 match self.receiver
.try_recv() {
2182 Ok(message
) => Ok(message
),
2188 Ok(SharedEmitterMessage
::Diagnostic(diag
)) => {
2189 let handler
= sess
.diagnostic();
2192 handler
.emit_with_code(&MultiSpan
::new(),
2198 handler
.emit(&MultiSpan
::new(),
2204 Ok(SharedEmitterMessage
::InlineAsmError(cookie
, msg
)) => {
2205 match Mark
::from_u32(cookie
).expn_info() {
2206 Some(ei
) => sess
.span_err(ei
.call_site
, &msg
),
2207 None
=> sess
.err(&msg
),
2210 Ok(SharedEmitterMessage
::AbortIfErrors
) => {
2211 sess
.abort_if_errors();
2213 Ok(SharedEmitterMessage
::Fatal(msg
)) => {
2225 pub struct OngoingCrateTranslation
{
2228 metadata
: EncodedMetadata
,
2229 windows_subsystem
: Option
<String
>,
2230 linker_info
: LinkerInfo
,
2231 crate_info
: CrateInfo
,
2232 time_graph
: Option
<TimeGraph
>,
2233 coordinator_send
: Sender
<Box
<Any
+ Send
>>,
2234 trans_worker_receive
: Receiver
<Message
>,
2235 shared_emitter_main
: SharedEmitterMain
,
2236 future
: thread
::JoinHandle
<Result
<CompiledModules
, ()>>,
2237 output_filenames
: Arc
<OutputFilenames
>,
2240 impl OngoingCrateTranslation
{
2241 pub(crate) fn join(self, sess
: &Session
, dep_graph
: &DepGraph
) -> CrateTranslation
{
2242 self.shared_emitter_main
.check(sess
, true);
2243 let compiled_modules
= match self.future
.join() {
2244 Ok(Ok(compiled_modules
)) => compiled_modules
,
2246 sess
.abort_if_errors();
2247 panic
!("expected abort due to worker thread errors")
2250 sess
.fatal("Error during translation/LLVM phase.");
2254 sess
.abort_if_errors();
2256 if let Some(time_graph
) = self.time_graph
{
2257 time_graph
.dump(&format
!("{}-timings", self.crate_name
));
2260 copy_module_artifacts_into_incr_comp_cache(sess
,
2263 produce_final_output_artifacts(sess
,
2265 &self.output_filenames
);
2267 // FIXME: time_llvm_passes support - does this use a global context or
2269 if sess
.codegen_units() == 1 && sess
.time_llvm_passes() {
2270 unsafe { llvm::LLVMRustPrintPassTimings(); }
2273 let trans
= CrateTranslation
{
2274 crate_name
: self.crate_name
,
2276 metadata
: self.metadata
,
2277 windows_subsystem
: self.windows_subsystem
,
2278 linker_info
: self.linker_info
,
2279 crate_info
: self.crate_info
,
2281 modules
: compiled_modules
.modules
,
2282 allocator_module
: compiled_modules
.allocator_module
,
2283 metadata_module
: compiled_modules
.metadata_module
,
2289 pub(crate) fn submit_pre_translated_module_to_llvm(&self,
2291 mtrans
: ModuleTranslation
) {
2292 self.wait_for_signal_to_translate_item();
2293 self.check_for_errors(tcx
.sess
);
2295 // These are generally cheap and won't through off scheduling.
2297 submit_translated_module_to_llvm(tcx
, mtrans
, cost
);
2300 pub fn translation_finished(&self, tcx
: TyCtxt
) {
2301 self.wait_for_signal_to_translate_item();
2302 self.check_for_errors(tcx
.sess
);
2303 drop(self.coordinator_send
.send(Box
::new(Message
::TranslationComplete
)));
2306 pub fn check_for_errors(&self, sess
: &Session
) {
2307 self.shared_emitter_main
.check(sess
, false);
2310 pub fn wait_for_signal_to_translate_item(&self) {
2311 match self.trans_worker_receive
.recv() {
2312 Ok(Message
::TranslateItem
) => {
2315 Ok(_
) => panic
!("unexpected message"),
2317 // One of the LLVM threads must have panicked, fall through so
2318 // error handling can be reached.
2324 pub(crate) fn submit_translated_module_to_llvm(tcx
: TyCtxt
,
2325 mtrans
: ModuleTranslation
,
2327 let llvm_work_item
= WorkItem
::Optimize(mtrans
);
2328 drop(tcx
.tx_to_llvm_workers
.lock().send(Box
::new(Message
::TranslationDone
{
2334 fn msvc_imps_needed(tcx
: TyCtxt
) -> bool
{
2335 tcx
.sess
.target
.target
.options
.is_like_msvc
&&
2336 tcx
.sess
.crate_types
.borrow().iter().any(|ct
| *ct
== config
::CrateTypeRlib
)
2339 // Create a `__imp_<symbol> = &symbol` global for every public static `symbol`.
2340 // This is required to satisfy `dllimport` references to static data in .rlibs
2341 // when using MSVC linker. We do this only for data, as linker can fix up
2342 // code references on its own.
2343 // See #26591, #27438
2344 fn create_msvc_imps(cgcx
: &CodegenContext
, llcx
: ContextRef
, llmod
: ModuleRef
) {
2345 if !cgcx
.msvc_imps_needed
{
2348 // The x86 ABI seems to require that leading underscores are added to symbol
2349 // names, so we need an extra underscore on 32-bit. There's also a leading
2350 // '\x01' here which disables LLVM's symbol mangling (e.g. no extra
2351 // underscores added in front).
2352 let prefix
= if cgcx
.target_pointer_width
== "32" {
2358 let i8p_ty
= Type
::i8p_llcx(llcx
);
2359 let globals
= base
::iter_globals(llmod
)
2361 llvm
::LLVMRustGetLinkage(val
) == llvm
::Linkage
::ExternalLinkage
&&
2362 llvm
::LLVMIsDeclaration(val
) == 0
2365 let name
= CStr
::from_ptr(llvm
::LLVMGetValueName(val
));
2366 let mut imp_name
= prefix
.as_bytes().to_vec();
2367 imp_name
.extend(name
.to_bytes());
2368 let imp_name
= CString
::new(imp_name
).unwrap();
2371 .collect
::<Vec
<_
>>();
2372 for (imp_name
, val
) in globals
{
2373 let imp
= llvm
::LLVMAddGlobal(llmod
,
2375 imp_name
.as_ptr() as *const _
);
2376 llvm
::LLVMSetInitializer(imp
, consts
::ptrcast(val
, i8p_ty
));
2377 llvm
::LLVMRustSetLinkage(imp
, llvm
::Linkage
::ExternalLinkage
);