1 use crate::back::lto::ThinBuffer;
2 use crate::back::profiling::{
3 selfprofile_after_pass_callback, selfprofile_before_pass_callback, LlvmSelfProfiler,
8 use crate::llvm::{self, DiagnosticInfo, PassManager};
10 use crate::type_::Type;
11 use crate::LlvmCodegenBackend;
12 use crate::ModuleLlvm;
13 use rustc_codegen_ssa::back::link::ensure_removed;
14 use rustc_codegen_ssa::back::write::{
15 BitcodeSection, CodegenContext, EmitObj, ModuleConfig, TargetMachineFactoryConfig,
16 TargetMachineFactoryFn,
18 use rustc_codegen_ssa::traits::*;
19 use rustc_codegen_ssa::{CompiledModule, ModuleCodegen};
20 use rustc_data_structures::profiling::SelfProfilerRef;
21 use rustc_data_structures::small_c_str::SmallCStr;
22 use rustc_errors::{FatalError, Handler, Level};
23 use rustc_fs_util::{link_or_copy, path_to_c_string};
24 use rustc_middle::bug;
25 use rustc_middle::ty::TyCtxt;
26 use rustc_session::config::{self, Lto, OutputType, Passes, SplitDwarfKind, SwitchWithOptPath};
27 use rustc_session::Session;
28 use rustc_span::symbol::sym;
29 use rustc_span::InnerSpan;
30 use rustc_target::spec::{CodeModel, RelocModel, SanitizerSet, SplitDebuginfo};
32 use libc::{c_char, c_int, c_uint, c_void, size_t};
33 use std::ffi::CString;
35 use std::io::{self, Write};
36 use std::path::{Path, PathBuf};
41 pub fn llvm_err(handler: &rustc_errors::Handler, msg: &str) -> FatalError {
42 match llvm::last_error() {
43 Some(err) => handler.fatal(&format!("{}: {}", msg, err)),
44 None => handler.fatal(msg),
48 pub fn write_output_file<'ll>(
49 handler: &rustc_errors::Handler,
50 target: &'ll llvm::TargetMachine,
51 pm: &llvm::PassManager<'ll>,
54 dwo_output: Option<&Path>,
55 file_type: llvm::FileType,
56 self_profiler_ref: &SelfProfilerRef,
57 ) -> Result<(), FatalError> {
58 debug!("write_output_file output={:?} dwo_output={:?}", output, dwo_output);
60 let output_c = path_to_c_string(output);
62 let dwo_output_ptr = if let Some(dwo_output) = dwo_output {
63 dwo_output_c = path_to_c_string(dwo_output);
68 let result = llvm::LLVMRustWriteOutputFile(
77 // Record artifact sizes for self-profiling
78 if result == llvm::LLVMRustResult::Success {
79 let artifact_kind = match file_type {
80 llvm::FileType::ObjectFile => "object_file",
81 llvm::FileType::AssemblyFile => "assembly_file",
83 record_artifact_size(self_profiler_ref, artifact_kind, output);
84 if let Some(dwo_file) = dwo_output {
85 record_artifact_size(self_profiler_ref, "dwo_file", dwo_file);
89 result.into_result().map_err(|()| {
90 let msg = format!("could not write output to {}", output.display());
91 llvm_err(handler, &msg)
96 pub fn create_informational_target_machine(sess: &Session) -> &'static mut llvm::TargetMachine {
97 let config = TargetMachineFactoryConfig { split_dwarf_file: None };
98 // Can't use query system here quite yet because this function is invoked before the query
99 // system/tcx is set up.
100 let features = llvm_util::global_llvm_features(sess, false);
101 target_machine_factory(sess, config::OptLevel::No, &features)(config)
102 .unwrap_or_else(|err| llvm_err(sess.diagnostic(), &err).raise())
105 pub fn create_target_machine(tcx: TyCtxt<'_>, mod_name: &str) -> &'static mut llvm::TargetMachine {
106 let split_dwarf_file = if tcx.sess.target_can_use_split_dwarf() {
107 tcx.output_filenames(()).split_dwarf_path(
108 tcx.sess.split_debuginfo(),
109 tcx.sess.opts.unstable_opts.split_dwarf_kind,
115 let config = TargetMachineFactoryConfig { split_dwarf_file };
116 target_machine_factory(
118 tcx.backend_optimization_level(()),
119 tcx.global_backend_features(()),
121 .unwrap_or_else(|err| llvm_err(tcx.sess.diagnostic(), &err).raise())
124 pub fn to_llvm_opt_settings(
125 cfg: config::OptLevel,
126 ) -> (llvm::CodeGenOptLevel, llvm::CodeGenOptSize) {
127 use self::config::OptLevel::*;
129 No => (llvm::CodeGenOptLevel::None, llvm::CodeGenOptSizeNone),
130 Less => (llvm::CodeGenOptLevel::Less, llvm::CodeGenOptSizeNone),
131 Default => (llvm::CodeGenOptLevel::Default, llvm::CodeGenOptSizeNone),
132 Aggressive => (llvm::CodeGenOptLevel::Aggressive, llvm::CodeGenOptSizeNone),
133 Size => (llvm::CodeGenOptLevel::Default, llvm::CodeGenOptSizeDefault),
134 SizeMin => (llvm::CodeGenOptLevel::Default, llvm::CodeGenOptSizeAggressive),
138 fn to_pass_builder_opt_level(cfg: config::OptLevel) -> llvm::PassBuilderOptLevel {
139 use config::OptLevel::*;
141 No => llvm::PassBuilderOptLevel::O0,
142 Less => llvm::PassBuilderOptLevel::O1,
143 Default => llvm::PassBuilderOptLevel::O2,
144 Aggressive => llvm::PassBuilderOptLevel::O3,
145 Size => llvm::PassBuilderOptLevel::Os,
146 SizeMin => llvm::PassBuilderOptLevel::Oz,
150 fn to_llvm_relocation_model(relocation_model: RelocModel) -> llvm::RelocModel {
151 match relocation_model {
152 RelocModel::Static => llvm::RelocModel::Static,
153 // LLVM doesn't have a PIE relocation model, it represents PIE as PIC with an extra attribute.
154 RelocModel::Pic | RelocModel::Pie => llvm::RelocModel::PIC,
155 RelocModel::DynamicNoPic => llvm::RelocModel::DynamicNoPic,
156 RelocModel::Ropi => llvm::RelocModel::ROPI,
157 RelocModel::Rwpi => llvm::RelocModel::RWPI,
158 RelocModel::RopiRwpi => llvm::RelocModel::ROPI_RWPI,
162 pub(crate) fn to_llvm_code_model(code_model: Option<CodeModel>) -> llvm::CodeModel {
164 Some(CodeModel::Tiny) => llvm::CodeModel::Tiny,
165 Some(CodeModel::Small) => llvm::CodeModel::Small,
166 Some(CodeModel::Kernel) => llvm::CodeModel::Kernel,
167 Some(CodeModel::Medium) => llvm::CodeModel::Medium,
168 Some(CodeModel::Large) => llvm::CodeModel::Large,
169 None => llvm::CodeModel::None,
173 pub fn target_machine_factory(
175 optlvl: config::OptLevel,
176 target_features: &[String],
177 ) -> TargetMachineFactoryFn<LlvmCodegenBackend> {
178 let reloc_model = to_llvm_relocation_model(sess.relocation_model());
180 let (opt_level, _) = to_llvm_opt_settings(optlvl);
181 let use_softfp = sess.opts.cg.soft_float;
183 let ffunction_sections =
184 sess.opts.unstable_opts.function_sections.unwrap_or(sess.target.function_sections);
185 let fdata_sections = ffunction_sections;
186 let funique_section_names = !sess.opts.unstable_opts.no_unique_section_names;
188 let code_model = to_llvm_code_model(sess.code_model());
190 let mut singlethread = sess.target.singlethread;
192 // On the wasm target once the `atomics` feature is enabled that means that
193 // we're no longer single-threaded, or otherwise we don't want LLVM to
194 // lower atomic operations to single-threaded operations.
195 if singlethread && sess.target.is_like_wasm && sess.target_features.contains(&sym::atomics) {
196 singlethread = false;
199 let triple = SmallCStr::new(&sess.target.llvm_target);
200 let cpu = SmallCStr::new(llvm_util::target_cpu(sess));
201 let features = CString::new(target_features.join(",")).unwrap();
202 let abi = SmallCStr::new(&sess.target.llvm_abiname);
203 let trap_unreachable =
204 sess.opts.unstable_opts.trap_unreachable.unwrap_or(sess.target.trap_unreachable);
205 let emit_stack_size_section = sess.opts.unstable_opts.emit_stack_sizes;
207 let asm_comments = sess.asm_comments();
208 let relax_elf_relocations =
209 sess.opts.unstable_opts.relax_elf_relocations.unwrap_or(sess.target.relax_elf_relocations);
212 !sess.opts.unstable_opts.use_ctors_section.unwrap_or(sess.target.use_ctors_section);
214 let path_mapping = sess.source_map().path_mapping().clone();
216 Arc::new(move |config: TargetMachineFactoryConfig| {
217 let split_dwarf_file =
218 path_mapping.map_prefix(config.split_dwarf_file.unwrap_or_default()).0;
219 let split_dwarf_file = CString::new(split_dwarf_file.to_str().unwrap()).unwrap();
222 llvm::LLVMRustCreateTargetMachine(
233 funique_section_names,
237 emit_stack_size_section,
238 relax_elf_relocations,
240 split_dwarf_file.as_ptr(),
245 format!("Could not create LLVM TargetMachine for triple: {}", triple.to_str().unwrap())
250 pub(crate) fn save_temp_bitcode(
251 cgcx: &CodegenContext<LlvmCodegenBackend>,
252 module: &ModuleCodegen<ModuleLlvm>,
255 if !cgcx.save_temps {
259 let ext = format!("{}.bc", name);
260 let cgu = Some(&module.name[..]);
261 let path = cgcx.output_filenames.temp_path_ext(&ext, cgu);
262 let cstr = path_to_c_string(&path);
263 let llmod = module.module_llvm.llmod();
264 llvm::LLVMWriteBitcodeToFile(llmod, cstr.as_ptr());
268 pub struct DiagnosticHandlers<'a> {
269 data: *mut (&'a CodegenContext<LlvmCodegenBackend>, &'a Handler),
270 llcx: &'a llvm::Context,
271 old_handler: Option<&'a llvm::DiagnosticHandler>,
274 impl<'a> DiagnosticHandlers<'a> {
276 cgcx: &'a CodegenContext<LlvmCodegenBackend>,
277 handler: &'a Handler,
278 llcx: &'a llvm::Context,
280 let remark_passes_all: bool;
281 let remark_passes: Vec<CString>;
284 remark_passes_all = true;
285 remark_passes = Vec::new();
287 Passes::Some(passes) => {
288 remark_passes_all = false;
290 passes.iter().map(|name| CString::new(name.as_str()).unwrap()).collect();
293 let remark_passes: Vec<*const c_char> =
294 remark_passes.iter().map(|name: &CString| name.as_ptr()).collect();
295 let data = Box::into_raw(Box::new((cgcx, handler)));
297 let old_handler = llvm::LLVMRustContextGetDiagnosticHandler(llcx);
298 llvm::LLVMRustContextConfigureDiagnosticHandler(
303 remark_passes.as_ptr(),
306 DiagnosticHandlers { data, llcx, old_handler }
311 impl<'a> Drop for DiagnosticHandlers<'a> {
314 llvm::LLVMRustContextSetDiagnosticHandler(self.llcx, self.old_handler);
315 drop(Box::from_raw(self.data));
320 fn report_inline_asm(
321 cgcx: &CodegenContext<LlvmCodegenBackend>,
323 level: llvm::DiagnosticLevel,
325 source: Option<(String, Vec<InnerSpan>)>,
327 // In LTO build we may get srcloc values from other crates which are invalid
328 // since they use a different source map. To be safe we just suppress these
330 if matches!(cgcx.lto, Lto::Fat | Lto::Thin) {
333 let level = match level {
334 llvm::DiagnosticLevel::Error => Level::Error { lint: false },
335 llvm::DiagnosticLevel::Warning => Level::Warning(None),
336 llvm::DiagnosticLevel::Note | llvm::DiagnosticLevel::Remark => Level::Note,
338 cgcx.diag_emitter.inline_asm_error(cookie as u32, msg, level, source);
341 unsafe extern "C" fn diagnostic_handler(info: &DiagnosticInfo, user: *mut c_void) {
345 let (cgcx, diag_handler) = *(user as *const (&CodegenContext<LlvmCodegenBackend>, &Handler));
347 match llvm::diagnostic::Diagnostic::unpack(info) {
348 llvm::diagnostic::InlineAsm(inline) => {
349 report_inline_asm(cgcx, inline.message, inline.level, inline.cookie, inline.source);
352 llvm::diagnostic::Optimization(opt) => {
353 let enabled = match cgcx.remark {
355 Passes::Some(ref v) => v.iter().any(|s| *s == opt.pass_name),
359 diag_handler.note_without_error(&format!(
361 opt.filename, opt.line, opt.column, opt.pass_name, opt.message,
365 llvm::diagnostic::PGO(diagnostic_ref) | llvm::diagnostic::Linker(diagnostic_ref) => {
366 let msg = llvm::build_string(|s| {
367 llvm::LLVMRustWriteDiagnosticInfoToString(diagnostic_ref, s)
369 .expect("non-UTF8 diagnostic");
370 diag_handler.warn(&msg);
372 llvm::diagnostic::Unsupported(diagnostic_ref) => {
373 let msg = llvm::build_string(|s| {
374 llvm::LLVMRustWriteDiagnosticInfoToString(diagnostic_ref, s)
376 .expect("non-UTF8 diagnostic");
377 diag_handler.err(&msg);
379 llvm::diagnostic::UnknownDiagnostic(..) => {}
383 fn get_pgo_gen_path(config: &ModuleConfig) -> Option<CString> {
384 match config.pgo_gen {
385 SwitchWithOptPath::Enabled(ref opt_dir_path) => {
386 let path = if let Some(dir_path) = opt_dir_path {
387 dir_path.join("default_%m.profraw")
389 PathBuf::from("default_%m.profraw")
392 Some(CString::new(format!("{}", path.display())).unwrap())
394 SwitchWithOptPath::Disabled => None,
398 fn get_pgo_use_path(config: &ModuleConfig) -> Option<CString> {
402 .map(|path_buf| CString::new(path_buf.to_string_lossy().as_bytes()).unwrap())
405 fn get_pgo_sample_use_path(config: &ModuleConfig) -> Option<CString> {
409 .map(|path_buf| CString::new(path_buf.to_string_lossy().as_bytes()).unwrap())
412 fn get_instr_profile_output_path(config: &ModuleConfig) -> Option<CString> {
413 if config.instrument_coverage {
414 Some(CString::new("default_%m_%p.profraw").unwrap())
420 pub(crate) unsafe fn optimize_with_new_llvm_pass_manager(
421 cgcx: &CodegenContext<LlvmCodegenBackend>,
422 diag_handler: &Handler,
423 module: &ModuleCodegen<ModuleLlvm>,
424 config: &ModuleConfig,
425 opt_level: config::OptLevel,
426 opt_stage: llvm::OptStage,
427 ) -> Result<(), FatalError> {
429 opt_level != config::OptLevel::Size && opt_level != config::OptLevel::SizeMin;
430 let using_thin_buffers = opt_stage == llvm::OptStage::PreLinkThinLTO || config.bitcode_needed();
431 let pgo_gen_path = get_pgo_gen_path(config);
432 let pgo_use_path = get_pgo_use_path(config);
433 let pgo_sample_use_path = get_pgo_sample_use_path(config);
434 let is_lto = opt_stage == llvm::OptStage::ThinLTO || opt_stage == llvm::OptStage::FatLTO;
435 let instr_profile_output_path = get_instr_profile_output_path(config);
436 // Sanitizer instrumentation is only inserted during the pre-link optimization stage.
437 let sanitizer_options = if !is_lto {
438 Some(llvm::SanitizerOptions {
439 sanitize_address: config.sanitizer.contains(SanitizerSet::ADDRESS),
440 sanitize_address_recover: config.sanitizer_recover.contains(SanitizerSet::ADDRESS),
441 sanitize_memory: config.sanitizer.contains(SanitizerSet::MEMORY),
442 sanitize_memory_recover: config.sanitizer_recover.contains(SanitizerSet::MEMORY),
443 sanitize_memory_track_origins: config.sanitizer_memory_track_origins as c_int,
444 sanitize_thread: config.sanitizer.contains(SanitizerSet::THREAD),
445 sanitize_hwaddress: config.sanitizer.contains(SanitizerSet::HWADDRESS),
446 sanitize_hwaddress_recover: config.sanitizer_recover.contains(SanitizerSet::HWADDRESS),
452 let mut llvm_profiler = if cgcx.prof.llvm_recording_enabled() {
453 Some(LlvmSelfProfiler::new(cgcx.prof.get_self_profiler().unwrap()))
458 let llvm_selfprofiler =
459 llvm_profiler.as_mut().map(|s| s as *mut _ as *mut c_void).unwrap_or(std::ptr::null_mut());
461 let extra_passes = if !is_lto { config.passes.join(",") } else { "".to_string() };
463 let llvm_plugins = config.llvm_plugins.join(",");
465 // FIXME: NewPM doesn't provide a facility to pass custom InlineParams.
466 // We would have to add upstream support for this first, before we can support
467 // config.inline_threshold and our more aggressive default thresholds.
468 let result = llvm::LLVMRustOptimizeWithNewPassManager(
469 module.module_llvm.llmod(),
470 &*module.module_llvm.tm,
471 to_pass_builder_opt_level(opt_level),
473 config.no_prepopulate_passes,
474 config.verify_llvm_ir,
476 config.merge_functions,
478 config.vectorize_slp,
479 config.vectorize_loop,
481 config.emit_lifetime_markers,
482 sanitizer_options.as_ref(),
483 pgo_gen_path.as_ref().map_or(std::ptr::null(), |s| s.as_ptr()),
484 pgo_use_path.as_ref().map_or(std::ptr::null(), |s| s.as_ptr()),
485 config.instrument_coverage,
486 instr_profile_output_path.as_ref().map_or(std::ptr::null(), |s| s.as_ptr()),
487 config.instrument_gcov,
488 pgo_sample_use_path.as_ref().map_or(std::ptr::null(), |s| s.as_ptr()),
489 config.debug_info_for_profiling,
491 selfprofile_before_pass_callback,
492 selfprofile_after_pass_callback,
493 extra_passes.as_ptr().cast(),
495 llvm_plugins.as_ptr().cast(),
498 result.into_result().map_err(|()| llvm_err(diag_handler, "failed to run LLVM passes"))
501 // Unsafe due to LLVM calls.
502 pub(crate) unsafe fn optimize(
503 cgcx: &CodegenContext<LlvmCodegenBackend>,
504 diag_handler: &Handler,
505 module: &ModuleCodegen<ModuleLlvm>,
506 config: &ModuleConfig,
507 ) -> Result<(), FatalError> {
508 let _timer = cgcx.prof.generic_activity_with_arg("LLVM_module_optimize", &*module.name);
510 let llmod = module.module_llvm.llmod();
511 let llcx = &*module.module_llvm.llcx;
512 let tm = &*module.module_llvm.tm;
513 let _handlers = DiagnosticHandlers::new(cgcx, diag_handler, llcx);
515 let module_name = module.name.clone();
516 let module_name = Some(&module_name[..]);
518 if let Some(false) = config.new_llvm_pass_manager && llvm_util::get_version() >= (15, 0, 0) {
520 "ignoring `-Z new-llvm-pass-manager=no`, which is no longer supported with LLVM 15",
524 if config.emit_no_opt_bc {
525 let out = cgcx.output_filenames.temp_path_ext("no-opt.bc", module_name);
526 let out = path_to_c_string(&out);
527 llvm::LLVMWriteBitcodeToFile(llmod, out.as_ptr());
530 if let Some(opt_level) = config.opt_level {
531 if llvm_util::should_use_new_llvm_pass_manager(
532 &config.new_llvm_pass_manager,
535 let opt_stage = match cgcx.lto {
536 Lto::Fat => llvm::OptStage::PreLinkFatLTO,
537 Lto::Thin | Lto::ThinLocal => llvm::OptStage::PreLinkThinLTO,
538 _ if cgcx.opts.cg.linker_plugin_lto.enabled() => llvm::OptStage::PreLinkThinLTO,
539 _ => llvm::OptStage::PreLinkNoLTO,
541 return optimize_with_new_llvm_pass_manager(
551 if cgcx.prof.llvm_recording_enabled() {
553 .warn("`-Z self-profile-events = llvm` requires `-Z new-llvm-pass-manager`");
556 // Create the two optimizing pass managers. These mirror what clang
557 // does, and are by populated by LLVM's default PassManagerBuilder.
558 // Each manager has a different set of passes, but they also share
559 // some common passes.
560 let fpm = llvm::LLVMCreateFunctionPassManagerForModule(llmod);
561 let mpm = llvm::LLVMCreatePassManager();
564 let find_pass = |pass_name: &str| {
565 let pass_name = SmallCStr::new(pass_name);
566 llvm::LLVMRustFindAndCreatePass(pass_name.as_ptr())
569 if config.verify_llvm_ir {
570 // Verification should run as the very first pass.
571 llvm::LLVMRustAddPass(fpm, find_pass("verify").unwrap());
574 let mut extra_passes = Vec::new();
575 let mut have_name_anon_globals_pass = false;
577 for pass_name in &config.passes {
578 if pass_name == "lint" {
579 // Linting should also be performed early, directly on the generated IR.
580 llvm::LLVMRustAddPass(fpm, find_pass("lint").unwrap());
584 if let Some(pass) = find_pass(pass_name) {
585 extra_passes.push(pass);
587 diag_handler.warn(&format!("unknown pass `{}`, ignoring", pass_name));
590 if pass_name == "name-anon-globals" {
591 have_name_anon_globals_pass = true;
595 // Instrumentation must be inserted before optimization,
596 // otherwise LLVM may optimize some functions away which
599 // This mirrors what Clang does in lib/CodeGen/BackendUtil.cpp.
600 if config.instrument_gcov {
601 llvm::LLVMRustAddPass(mpm, find_pass("insert-gcov-profiling").unwrap());
603 if config.instrument_coverage {
604 llvm::LLVMRustAddPass(mpm, find_pass("instrprof").unwrap());
606 if config.debug_info_for_profiling {
607 llvm::LLVMRustAddPass(mpm, find_pass("add-discriminators").unwrap());
610 add_sanitizer_passes(config, &mut extra_passes);
612 // Some options cause LLVM bitcode to be emitted, which uses ThinLTOBuffers, so we need
613 // to make sure we run LLVM's NameAnonGlobals pass when emitting bitcode; otherwise
614 // we'll get errors in LLVM.
615 let using_thin_buffers = config.bitcode_needed();
616 if !config.no_prepopulate_passes {
617 llvm::LLVMAddAnalysisPasses(tm, fpm);
618 llvm::LLVMAddAnalysisPasses(tm, mpm);
619 let opt_level = to_llvm_opt_settings(opt_level).0;
620 let prepare_for_thin_lto = cgcx.lto == Lto::Thin
621 || cgcx.lto == Lto::ThinLocal
622 || (cgcx.lto != Lto::Fat && cgcx.opts.cg.linker_plugin_lto.enabled());
623 with_llvm_pmb(llmod, config, opt_level, prepare_for_thin_lto, &mut |b| {
624 llvm::LLVMRustAddLastExtensionPasses(
626 extra_passes.as_ptr(),
627 extra_passes.len() as size_t,
629 llvm::LLVMRustPassManagerBuilderPopulateFunctionPassManager(b, fpm);
630 llvm::LLVMRustPassManagerBuilderPopulateModulePassManager(b, mpm);
633 have_name_anon_globals_pass = have_name_anon_globals_pass || prepare_for_thin_lto;
634 if using_thin_buffers && !prepare_for_thin_lto {
635 llvm::LLVMRustAddPass(mpm, find_pass("name-anon-globals").unwrap());
636 have_name_anon_globals_pass = true;
639 // If we don't use the standard pipeline, directly populate the MPM
640 // with the extra passes.
641 for pass in extra_passes {
642 llvm::LLVMRustAddPass(mpm, pass);
646 if using_thin_buffers && !have_name_anon_globals_pass {
647 // As described above, this will probably cause an error in LLVM
648 if config.no_prepopulate_passes {
650 "The current compilation is going to use thin LTO buffers \
651 without running LLVM's NameAnonGlobals pass. \
652 This will likely cause errors in LLVM. Consider adding \
653 -C passes=name-anon-globals to the compiler command line.",
657 "We are using thin LTO buffers without running the NameAnonGlobals pass. \
658 This will likely cause errors in LLVM and should never happen."
664 diag_handler.abort_if_errors();
666 // Finally, run the actual optimization passes
668 let _timer = cgcx.prof.extra_verbose_generic_activity(
669 "LLVM_module_optimize_function_passes",
672 llvm::LLVMRustRunFunctionPassManager(fpm, llmod);
675 let _timer = cgcx.prof.extra_verbose_generic_activity(
676 "LLVM_module_optimize_module_passes",
679 llvm::LLVMRunPassManager(mpm, llmod);
682 // Deallocate managers that we're now done with
683 llvm::LLVMDisposePassManager(fpm);
684 llvm::LLVMDisposePassManager(mpm);
689 unsafe fn add_sanitizer_passes(config: &ModuleConfig, passes: &mut Vec<&'static mut llvm::Pass>) {
690 if config.sanitizer.contains(SanitizerSet::ADDRESS) {
691 let recover = config.sanitizer_recover.contains(SanitizerSet::ADDRESS);
692 passes.push(llvm::LLVMRustCreateAddressSanitizerFunctionPass(recover));
693 passes.push(llvm::LLVMRustCreateModuleAddressSanitizerPass(recover));
695 if config.sanitizer.contains(SanitizerSet::MEMORY) {
696 let track_origins = config.sanitizer_memory_track_origins as c_int;
697 let recover = config.sanitizer_recover.contains(SanitizerSet::MEMORY);
698 passes.push(llvm::LLVMRustCreateMemorySanitizerPass(track_origins, recover));
700 if config.sanitizer.contains(SanitizerSet::THREAD) {
701 passes.push(llvm::LLVMRustCreateThreadSanitizerPass());
703 if config.sanitizer.contains(SanitizerSet::HWADDRESS) {
704 let recover = config.sanitizer_recover.contains(SanitizerSet::HWADDRESS);
705 passes.push(llvm::LLVMRustCreateHWAddressSanitizerPass(recover));
710 cgcx: &CodegenContext<LlvmCodegenBackend>,
711 diag_handler: &Handler,
712 mut modules: Vec<ModuleCodegen<ModuleLlvm>>,
713 ) -> Result<ModuleCodegen<ModuleLlvm>, FatalError> {
714 use super::lto::{Linker, ModuleBuffer};
715 // Sort the modules by name to ensure to ensure deterministic behavior.
716 modules.sort_by(|a, b| a.name.cmp(&b.name));
717 let (first, elements) =
718 modules.split_first().expect("Bug! modules must contain at least one module.");
720 let mut linker = Linker::new(first.module_llvm.llmod());
721 for module in elements {
722 let _timer = cgcx.prof.generic_activity_with_arg("LLVM_link_module", &*module.name);
723 let buffer = ModuleBuffer::new(module.module_llvm.llmod());
724 linker.add(buffer.data()).map_err(|()| {
725 let msg = format!("failed to serialize module {:?}", module.name);
726 llvm_err(diag_handler, &msg)
730 Ok(modules.remove(0))
733 pub(crate) unsafe fn codegen(
734 cgcx: &CodegenContext<LlvmCodegenBackend>,
735 diag_handler: &Handler,
736 module: ModuleCodegen<ModuleLlvm>,
737 config: &ModuleConfig,
738 ) -> Result<CompiledModule, FatalError> {
739 let _timer = cgcx.prof.generic_activity_with_arg("LLVM_module_codegen", &*module.name);
741 let llmod = module.module_llvm.llmod();
742 let llcx = &*module.module_llvm.llcx;
743 let tm = &*module.module_llvm.tm;
744 let module_name = module.name.clone();
745 let module_name = Some(&module_name[..]);
746 let handlers = DiagnosticHandlers::new(cgcx, diag_handler, llcx);
748 if cgcx.msvc_imps_needed {
749 create_msvc_imps(cgcx, llcx, llmod);
752 // A codegen-specific pass manager is used to generate object
753 // files for an LLVM module.
755 // Apparently each of these pass managers is a one-shot kind of
756 // thing, so we create a new one for each type of output. The
757 // pass manager passed to the closure should be ensured to not
758 // escape the closure itself, and the manager should only be
760 unsafe fn with_codegen<'ll, F, R>(
761 tm: &'ll llvm::TargetMachine,
762 llmod: &'ll llvm::Module,
767 F: FnOnce(&'ll mut PassManager<'ll>) -> R,
769 let cpm = llvm::LLVMCreatePassManager();
770 llvm::LLVMAddAnalysisPasses(tm, cpm);
771 llvm::LLVMRustAddLibraryInfo(cpm, llmod, no_builtins);
775 // Two things to note:
776 // - If object files are just LLVM bitcode we write bitcode, copy it to
777 // the .o file, and delete the bitcode if it wasn't otherwise
779 // - If we don't have the integrated assembler then we need to emit
780 // asm from LLVM and use `gcc` to create the object file.
782 let bc_out = cgcx.output_filenames.temp_path(OutputType::Bitcode, module_name);
783 let obj_out = cgcx.output_filenames.temp_path(OutputType::Object, module_name);
785 if config.bitcode_needed() {
788 .generic_activity_with_arg("LLVM_module_codegen_make_bitcode", &*module.name);
789 let thin = ThinBuffer::new(llmod, config.emit_thin_lto);
790 let data = thin.data();
792 if let Some(bitcode_filename) = bc_out.file_name() {
793 cgcx.prof.artifact_size(
795 bitcode_filename.to_string_lossy(),
800 if config.emit_bc || config.emit_obj == EmitObj::Bitcode {
803 .generic_activity_with_arg("LLVM_module_codegen_emit_bitcode", &*module.name);
804 if let Err(e) = fs::write(&bc_out, data) {
805 let msg = format!("failed to write bytecode to {}: {}", bc_out.display(), e);
806 diag_handler.err(&msg);
810 if config.emit_obj == EmitObj::ObjectCode(BitcodeSection::Full) {
813 .generic_activity_with_arg("LLVM_module_codegen_embed_bitcode", &*module.name);
814 embed_bitcode(cgcx, llcx, llmod, &config.bc_cmdline, data);
820 cgcx.prof.generic_activity_with_arg("LLVM_module_codegen_emit_ir", &*module.name);
821 let out = cgcx.output_filenames.temp_path(OutputType::LlvmAssembly, module_name);
822 let out_c = path_to_c_string(&out);
824 extern "C" fn demangle_callback(
825 input_ptr: *const c_char,
827 output_ptr: *mut c_char,
831 unsafe { slice::from_raw_parts(input_ptr as *const u8, input_len as usize) };
833 let Ok(input) = str::from_utf8(input) else { return 0 };
835 let output = unsafe {
836 slice::from_raw_parts_mut(output_ptr as *mut u8, output_len as usize)
838 let mut cursor = io::Cursor::new(output);
840 let Ok(demangled) = rustc_demangle::try_demangle(input) else { return 0 };
842 if write!(cursor, "{:#}", demangled).is_err() {
843 // Possible only if provided buffer is not big enough
847 cursor.position() as size_t
850 let result = llvm::LLVMRustPrintModule(llmod, out_c.as_ptr(), demangle_callback);
852 if result == llvm::LLVMRustResult::Success {
853 record_artifact_size(&cgcx.prof, "llvm_ir", &out);
856 result.into_result().map_err(|()| {
857 let msg = format!("failed to write LLVM IR to {}", out.display());
858 llvm_err(diag_handler, &msg)
864 cgcx.prof.generic_activity_with_arg("LLVM_module_codegen_emit_asm", &*module.name);
865 let path = cgcx.output_filenames.temp_path(OutputType::Assembly, module_name);
867 // We can't use the same module for asm and object code output,
868 // because that triggers various errors like invalid IR or broken
869 // binaries. So we must clone the module to produce the asm output
870 // if we are also producing object code.
871 let llmod = if let EmitObj::ObjectCode(_) = config.emit_obj {
872 llvm::LLVMCloneModule(llmod)
876 with_codegen(tm, llmod, config.no_builtins, |cpm| {
884 llvm::FileType::AssemblyFile,
890 match config.emit_obj {
891 EmitObj::ObjectCode(_) => {
894 .generic_activity_with_arg("LLVM_module_codegen_emit_obj", &*module.name);
896 let dwo_out = cgcx.output_filenames.temp_path_dwo(module_name);
897 let dwo_out = match (cgcx.split_debuginfo, cgcx.split_dwarf_kind) {
898 // Don't change how DWARF is emitted when disabled.
899 (SplitDebuginfo::Off, _) => None,
900 // Don't provide a DWARF object path if split debuginfo is enabled but this is
901 // a platform that doesn't support Split DWARF.
902 _ if !cgcx.target_can_use_split_dwarf => None,
903 // Don't provide a DWARF object path in single mode, sections will be written
904 // into the object as normal but ignored by linker.
905 (_, SplitDwarfKind::Single) => None,
906 // Emit (a subset of the) DWARF into a separate dwarf object file in split
908 (_, SplitDwarfKind::Split) => Some(dwo_out.as_path()),
911 with_codegen(tm, llmod, config.no_builtins, |cpm| {
919 llvm::FileType::ObjectFile,
925 EmitObj::Bitcode => {
926 debug!("copying bitcode {:?} to obj {:?}", bc_out, obj_out);
927 if let Err(e) = link_or_copy(&bc_out, &obj_out) {
928 diag_handler.err(&format!("failed to copy bitcode to object file: {}", e));
932 debug!("removing_bitcode {:?}", bc_out);
933 ensure_removed(diag_handler, &bc_out);
943 Ok(module.into_compiled_module(
944 config.emit_obj != EmitObj::None,
945 cgcx.target_can_use_split_dwarf
946 && cgcx.split_debuginfo != SplitDebuginfo::Off
947 && cgcx.split_dwarf_kind == SplitDwarfKind::Split,
949 &cgcx.output_filenames,
953 fn create_section_with_flags_asm(section_name: &str, section_flags: &str, data: &[u8]) -> Vec<u8> {
954 let mut asm = format!(".section {},\"{}\"\n", section_name, section_flags).into_bytes();
955 asm.extend_from_slice(b".ascii \"");
956 asm.reserve(data.len());
958 if byte == b'\\' || byte == b'"' {
961 } else if byte < 0x20 || byte >= 0x80 {
962 // Avoid non UTF-8 inline assembly. Use octal escape sequence, because it is fixed
963 // width, while hex escapes will consume following characters.
965 asm.push(b'0' + ((byte >> 6) & 0x7));
966 asm.push(b'0' + ((byte >> 3) & 0x7));
967 asm.push(b'0' + ((byte >> 0) & 0x7));
972 asm.extend_from_slice(b"\"\n");
976 /// Embed the bitcode of an LLVM module in the LLVM module itself.
978 /// This is done primarily for iOS where it appears to be standard to compile C
979 /// code at least with `-fembed-bitcode` which creates two sections in the
982 /// * __LLVM,__bitcode
983 /// * __LLVM,__cmdline
985 /// It appears *both* of these sections are necessary to get the linker to
986 /// recognize what's going on. A suitable cmdline value is taken from the
989 /// Furthermore debug/O1 builds don't actually embed bitcode but rather just
990 /// embed an empty section.
992 /// Basically all of this is us attempting to follow in the footsteps of clang
993 /// on iOS. See #35968 for lots more info.
994 unsafe fn embed_bitcode(
995 cgcx: &CodegenContext<LlvmCodegenBackend>,
996 llcx: &llvm::Context,
997 llmod: &llvm::Module,
1001 // We're adding custom sections to the output object file, but we definitely
1002 // do not want these custom sections to make their way into the final linked
1003 // executable. The purpose of these custom sections is for tooling
1004 // surrounding object files to work with the LLVM IR, if necessary. For
1005 // example rustc's own LTO will look for LLVM IR inside of the object file
1006 // in these sections by default.
1008 // To handle this is a bit different depending on the object file format
1009 // used by the backend, broken down into a few different categories:
1011 // * Mach-O - this is for macOS. Inspecting the source code for the native
1012 // linker here shows that the `.llvmbc` and `.llvmcmd` sections are
1013 // automatically skipped by the linker. In that case there's nothing extra
1014 // that we need to do here.
1016 // * Wasm - the native LLD linker is hard-coded to skip `.llvmbc` and
1017 // `.llvmcmd` sections, so there's nothing extra we need to do.
1019 // * COFF - if we don't do anything the linker will by default copy all
1020 // these sections to the output artifact, not what we want! To subvert
1021 // this we want to flag the sections we inserted here as
1022 // `IMAGE_SCN_LNK_REMOVE`.
1024 // * ELF - this is very similar to COFF above. One difference is that these
1025 // sections are removed from the output linked artifact when
1026 // `--gc-sections` is passed, which we pass by default. If that flag isn't
1027 // passed though then these sections will show up in the final output.
1028 // Additionally the flag that we need to set here is `SHF_EXCLUDE`.
1030 // Unfortunately, LLVM provides no way to set custom section flags. For ELF
1031 // and COFF we emit the sections using module level inline assembly for that
1032 // reason (see issue #90326 for historical background).
1033 let is_apple = cgcx.opts.target_triple.triple().contains("-ios")
1034 || cgcx.opts.target_triple.triple().contains("-darwin")
1035 || cgcx.opts.target_triple.triple().contains("-tvos")
1036 || cgcx.opts.target_triple.triple().contains("-watchos");
1038 || cgcx.opts.target_triple.triple().starts_with("wasm")
1039 || cgcx.opts.target_triple.triple().starts_with("asmjs")
1041 // We don't need custom section flags, create LLVM globals.
1042 let llconst = common::bytes_in_context(llcx, bitcode);
1043 let llglobal = llvm::LLVMAddGlobal(
1045 common::val_ty(llconst),
1046 "rustc.embedded.module\0".as_ptr().cast(),
1048 llvm::LLVMSetInitializer(llglobal, llconst);
1050 let section = if is_apple { "__LLVM,__bitcode\0" } else { ".llvmbc\0" };
1051 llvm::LLVMSetSection(llglobal, section.as_ptr().cast());
1052 llvm::LLVMRustSetLinkage(llglobal, llvm::Linkage::PrivateLinkage);
1053 llvm::LLVMSetGlobalConstant(llglobal, llvm::True);
1055 let llconst = common::bytes_in_context(llcx, cmdline.as_bytes());
1056 let llglobal = llvm::LLVMAddGlobal(
1058 common::val_ty(llconst),
1059 "rustc.embedded.cmdline\0".as_ptr().cast(),
1061 llvm::LLVMSetInitializer(llglobal, llconst);
1062 let section = if is_apple { "__LLVM,__cmdline\0" } else { ".llvmcmd\0" };
1063 llvm::LLVMSetSection(llglobal, section.as_ptr().cast());
1064 llvm::LLVMRustSetLinkage(llglobal, llvm::Linkage::PrivateLinkage);
1066 // We need custom section flags, so emit module-level inline assembly.
1067 let section_flags = if cgcx.is_pe_coff { "n" } else { "e" };
1068 let asm = create_section_with_flags_asm(".llvmbc", section_flags, bitcode);
1069 llvm::LLVMRustAppendModuleInlineAsm(llmod, asm.as_ptr().cast(), asm.len());
1070 let asm = create_section_with_flags_asm(".llvmcmd", section_flags, cmdline.as_bytes());
1071 llvm::LLVMRustAppendModuleInlineAsm(llmod, asm.as_ptr().cast(), asm.len());
1075 pub unsafe fn with_llvm_pmb(
1076 llmod: &llvm::Module,
1077 config: &ModuleConfig,
1078 opt_level: llvm::CodeGenOptLevel,
1079 prepare_for_thin_lto: bool,
1080 f: &mut dyn FnMut(&llvm::PassManagerBuilder),
1084 // Create the PassManagerBuilder for LLVM. We configure it with
1085 // reasonable defaults and prepare it to actually populate the pass
1087 let builder = llvm::LLVMRustPassManagerBuilderCreate();
1088 let opt_size = config.opt_size.map_or(llvm::CodeGenOptSizeNone, |x| to_llvm_opt_settings(x).1);
1089 let inline_threshold = config.inline_threshold;
1090 let pgo_gen_path = get_pgo_gen_path(config);
1091 let pgo_use_path = get_pgo_use_path(config);
1092 let pgo_sample_use_path = get_pgo_sample_use_path(config);
1094 llvm::LLVMRustConfigurePassManagerBuilder(
1097 config.merge_functions,
1098 config.vectorize_slp,
1099 config.vectorize_loop,
1100 prepare_for_thin_lto,
1101 pgo_gen_path.as_ref().map_or(ptr::null(), |s| s.as_ptr()),
1102 pgo_use_path.as_ref().map_or(ptr::null(), |s| s.as_ptr()),
1103 pgo_sample_use_path.as_ref().map_or(ptr::null(), |s| s.as_ptr()),
1107 llvm::LLVMRustAddBuilderLibraryInfo(builder, llmod, config.no_builtins);
1109 // Here we match what clang does (kinda). For O0 we only inline
1110 // always-inline functions (but don't add lifetime intrinsics), at O1 we
1111 // inline with lifetime intrinsics, and O2+ we add an inliner with a
1112 // thresholds copied from clang.
1113 match (opt_level, opt_size, inline_threshold) {
1115 llvm::LLVMRustPassManagerBuilderUseInlinerWithThreshold(builder, t);
1117 (llvm::CodeGenOptLevel::Aggressive, ..) => {
1118 llvm::LLVMRustPassManagerBuilderUseInlinerWithThreshold(builder, 275);
1120 (_, llvm::CodeGenOptSizeDefault, _) => {
1121 llvm::LLVMRustPassManagerBuilderUseInlinerWithThreshold(builder, 75);
1123 (_, llvm::CodeGenOptSizeAggressive, _) => {
1124 llvm::LLVMRustPassManagerBuilderUseInlinerWithThreshold(builder, 25);
1126 (llvm::CodeGenOptLevel::None, ..) => {
1127 llvm::LLVMRustAddAlwaysInlinePass(builder, config.emit_lifetime_markers);
1129 (llvm::CodeGenOptLevel::Less, ..) => {
1130 llvm::LLVMRustAddAlwaysInlinePass(builder, config.emit_lifetime_markers);
1132 (llvm::CodeGenOptLevel::Default, ..) => {
1133 llvm::LLVMRustPassManagerBuilderUseInlinerWithThreshold(builder, 225);
1138 llvm::LLVMRustPassManagerBuilderDispose(builder);
1141 // Create a `__imp_<symbol> = &symbol` global for every public static `symbol`.
1142 // This is required to satisfy `dllimport` references to static data in .rlibs
1143 // when using MSVC linker. We do this only for data, as linker can fix up
1144 // code references on its own.
1145 // See #26591, #27438
1146 fn create_msvc_imps(
1147 cgcx: &CodegenContext<LlvmCodegenBackend>,
1148 llcx: &llvm::Context,
1149 llmod: &llvm::Module,
1151 if !cgcx.msvc_imps_needed {
1154 // The x86 ABI seems to require that leading underscores are added to symbol
1155 // names, so we need an extra underscore on x86. There's also a leading
1156 // '\x01' here which disables LLVM's symbol mangling (e.g., no extra
1157 // underscores added in front).
1158 let prefix = if cgcx.target_arch == "x86" { "\x01__imp__" } else { "\x01__imp_" };
1161 let i8p_ty = Type::i8p_llcx(llcx);
1162 let globals = base::iter_globals(llmod)
1164 llvm::LLVMRustGetLinkage(val) == llvm::Linkage::ExternalLinkage
1165 && llvm::LLVMIsDeclaration(val) == 0
1168 // Exclude some symbols that we know are not Rust symbols.
1169 let name = llvm::get_value_name(val);
1170 if ignored(name) { None } else { Some((val, name)) }
1172 .map(move |(val, name)| {
1173 let mut imp_name = prefix.as_bytes().to_vec();
1174 imp_name.extend(name);
1175 let imp_name = CString::new(imp_name).unwrap();
1178 .collect::<Vec<_>>();
1180 for (imp_name, val) in globals {
1181 let imp = llvm::LLVMAddGlobal(llmod, i8p_ty, imp_name.as_ptr().cast());
1182 llvm::LLVMSetInitializer(imp, consts::ptrcast(val, i8p_ty));
1183 llvm::LLVMRustSetLinkage(imp, llvm::Linkage::ExternalLinkage);
1187 // Use this function to exclude certain symbols from `__imp` generation.
1188 fn ignored(symbol_name: &[u8]) -> bool {
1189 // These are symbols generated by LLVM's profiling instrumentation
1190 symbol_name.starts_with(b"__llvm_profile_")
1194 fn record_artifact_size(
1195 self_profiler_ref: &SelfProfilerRef,
1196 artifact_kind: &'static str,
1199 // Don't stat the file if we are not going to record its size.
1200 if !self_profiler_ref.enabled() {
1204 if let Some(artifact_name) = path.file_name() {
1205 let file_size = std::fs::metadata(path).map(|m| m.len()).unwrap_or(0);
1206 self_profiler_ref.artifact_size(artifact_kind, artifact_name.to_string_lossy(), file_size);