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, SMDiagnostic};
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};
33 use libc::{c_char, c_int, c_uint, c_void, size_t};
34 use std::ffi::CString;
36 use std::io::{self, Write};
37 use std::path::{Path, PathBuf};
42 pub fn llvm_err(handler: &rustc_errors::Handler, msg: &str) -> FatalError {
43 match llvm::last_error() {
44 Some(err) => handler.fatal(&format!("{}: {}", msg, err)),
45 None => handler.fatal(msg),
49 pub fn write_output_file<'ll>(
50 handler: &rustc_errors::Handler,
51 target: &'ll llvm::TargetMachine,
52 pm: &llvm::PassManager<'ll>,
55 dwo_output: Option<&Path>,
56 file_type: llvm::FileType,
57 self_profiler_ref: &SelfProfilerRef,
58 ) -> Result<(), FatalError> {
59 debug!("write_output_file output={:?} dwo_output={:?}", output, dwo_output);
61 let output_c = path_to_c_string(output);
63 let dwo_output_ptr = if let Some(dwo_output) = dwo_output {
64 dwo_output_c = path_to_c_string(dwo_output);
69 let result = llvm::LLVMRustWriteOutputFile(
78 // Record artifact sizes for self-profiling
79 if result == llvm::LLVMRustResult::Success {
80 let artifact_kind = match file_type {
81 llvm::FileType::ObjectFile => "object_file",
82 llvm::FileType::AssemblyFile => "assembly_file",
84 record_artifact_size(self_profiler_ref, artifact_kind, output);
85 if let Some(dwo_file) = dwo_output {
86 record_artifact_size(self_profiler_ref, "dwo_file", dwo_file);
90 result.into_result().map_err(|()| {
91 let msg = format!("could not write output to {}", output.display());
92 llvm_err(handler, &msg)
97 pub fn create_informational_target_machine(sess: &Session) -> &'static mut llvm::TargetMachine {
98 let config = TargetMachineFactoryConfig { split_dwarf_file: None };
99 // Can't use query system here quite yet because this function is invoked before the query
100 // system/tcx is set up.
101 let features = llvm_util::global_llvm_features(sess, false);
102 target_machine_factory(sess, config::OptLevel::No, &features)(config)
103 .unwrap_or_else(|err| llvm_err(sess.diagnostic(), &err).raise())
106 pub fn create_target_machine(tcx: TyCtxt<'_>, mod_name: &str) -> &'static mut llvm::TargetMachine {
107 let split_dwarf_file = if tcx.sess.target_can_use_split_dwarf() {
108 tcx.output_filenames(()).split_dwarf_path(
109 tcx.sess.split_debuginfo(),
110 tcx.sess.opts.unstable_opts.split_dwarf_kind,
116 let config = TargetMachineFactoryConfig { split_dwarf_file };
117 target_machine_factory(
119 tcx.backend_optimization_level(()),
120 tcx.global_backend_features(()),
122 .unwrap_or_else(|err| llvm_err(tcx.sess.diagnostic(), &err).raise())
125 pub fn to_llvm_opt_settings(
126 cfg: config::OptLevel,
127 ) -> (llvm::CodeGenOptLevel, llvm::CodeGenOptSize) {
128 use self::config::OptLevel::*;
130 No => (llvm::CodeGenOptLevel::None, llvm::CodeGenOptSizeNone),
131 Less => (llvm::CodeGenOptLevel::Less, llvm::CodeGenOptSizeNone),
132 Default => (llvm::CodeGenOptLevel::Default, llvm::CodeGenOptSizeNone),
133 Aggressive => (llvm::CodeGenOptLevel::Aggressive, llvm::CodeGenOptSizeNone),
134 Size => (llvm::CodeGenOptLevel::Default, llvm::CodeGenOptSizeDefault),
135 SizeMin => (llvm::CodeGenOptLevel::Default, llvm::CodeGenOptSizeAggressive),
139 fn to_pass_builder_opt_level(cfg: config::OptLevel) -> llvm::PassBuilderOptLevel {
140 use config::OptLevel::*;
142 No => llvm::PassBuilderOptLevel::O0,
143 Less => llvm::PassBuilderOptLevel::O1,
144 Default => llvm::PassBuilderOptLevel::O2,
145 Aggressive => llvm::PassBuilderOptLevel::O3,
146 Size => llvm::PassBuilderOptLevel::Os,
147 SizeMin => llvm::PassBuilderOptLevel::Oz,
151 fn to_llvm_relocation_model(relocation_model: RelocModel) -> llvm::RelocModel {
152 match relocation_model {
153 RelocModel::Static => llvm::RelocModel::Static,
154 // LLVM doesn't have a PIE relocation model, it represents PIE as PIC with an extra attribute.
155 RelocModel::Pic | RelocModel::Pie => llvm::RelocModel::PIC,
156 RelocModel::DynamicNoPic => llvm::RelocModel::DynamicNoPic,
157 RelocModel::Ropi => llvm::RelocModel::ROPI,
158 RelocModel::Rwpi => llvm::RelocModel::RWPI,
159 RelocModel::RopiRwpi => llvm::RelocModel::ROPI_RWPI,
163 pub(crate) fn to_llvm_code_model(code_model: Option<CodeModel>) -> llvm::CodeModel {
165 Some(CodeModel::Tiny) => llvm::CodeModel::Tiny,
166 Some(CodeModel::Small) => llvm::CodeModel::Small,
167 Some(CodeModel::Kernel) => llvm::CodeModel::Kernel,
168 Some(CodeModel::Medium) => llvm::CodeModel::Medium,
169 Some(CodeModel::Large) => llvm::CodeModel::Large,
170 None => llvm::CodeModel::None,
174 pub fn target_machine_factory(
176 optlvl: config::OptLevel,
177 target_features: &[String],
178 ) -> TargetMachineFactoryFn<LlvmCodegenBackend> {
179 let reloc_model = to_llvm_relocation_model(sess.relocation_model());
181 let (opt_level, _) = to_llvm_opt_settings(optlvl);
182 let use_softfp = sess.opts.cg.soft_float;
184 let ffunction_sections =
185 sess.opts.unstable_opts.function_sections.unwrap_or(sess.target.function_sections);
186 let fdata_sections = ffunction_sections;
187 let funique_section_names = !sess.opts.unstable_opts.no_unique_section_names;
189 let code_model = to_llvm_code_model(sess.code_model());
191 let mut singlethread = sess.target.singlethread;
193 // On the wasm target once the `atomics` feature is enabled that means that
194 // we're no longer single-threaded, or otherwise we don't want LLVM to
195 // lower atomic operations to single-threaded operations.
196 if singlethread && sess.target.is_like_wasm && sess.target_features.contains(&sym::atomics) {
197 singlethread = false;
200 let triple = SmallCStr::new(&sess.target.llvm_target);
201 let cpu = SmallCStr::new(llvm_util::target_cpu(sess));
202 let features = CString::new(target_features.join(",")).unwrap();
203 let abi = SmallCStr::new(&sess.target.llvm_abiname);
204 let trap_unreachable =
205 sess.opts.unstable_opts.trap_unreachable.unwrap_or(sess.target.trap_unreachable);
206 let emit_stack_size_section = sess.opts.unstable_opts.emit_stack_sizes;
208 let asm_comments = sess.asm_comments();
209 let relax_elf_relocations =
210 sess.opts.unstable_opts.relax_elf_relocations.unwrap_or(sess.target.relax_elf_relocations);
213 !sess.opts.unstable_opts.use_ctors_section.unwrap_or(sess.target.use_ctors_section);
215 let path_mapping = sess.source_map().path_mapping().clone();
217 Arc::new(move |config: TargetMachineFactoryConfig| {
218 let split_dwarf_file =
219 path_mapping.map_prefix(config.split_dwarf_file.unwrap_or_default()).0;
220 let split_dwarf_file = CString::new(split_dwarf_file.to_str().unwrap()).unwrap();
223 llvm::LLVMRustCreateTargetMachine(
234 funique_section_names,
238 emit_stack_size_section,
239 relax_elf_relocations,
241 split_dwarf_file.as_ptr(),
246 format!("Could not create LLVM TargetMachine for triple: {}", triple.to_str().unwrap())
251 pub(crate) fn save_temp_bitcode(
252 cgcx: &CodegenContext<LlvmCodegenBackend>,
253 module: &ModuleCodegen<ModuleLlvm>,
256 if !cgcx.save_temps {
260 let ext = format!("{}.bc", name);
261 let cgu = Some(&module.name[..]);
262 let path = cgcx.output_filenames.temp_path_ext(&ext, cgu);
263 let cstr = path_to_c_string(&path);
264 let llmod = module.module_llvm.llmod();
265 llvm::LLVMWriteBitcodeToFile(llmod, cstr.as_ptr());
269 pub struct DiagnosticHandlers<'a> {
270 data: *mut (&'a CodegenContext<LlvmCodegenBackend>, &'a Handler),
271 llcx: &'a llvm::Context,
272 old_handler: Option<&'a llvm::DiagnosticHandler>,
275 impl<'a> DiagnosticHandlers<'a> {
277 cgcx: &'a CodegenContext<LlvmCodegenBackend>,
278 handler: &'a Handler,
279 llcx: &'a llvm::Context,
281 let remark_passes_all: bool;
282 let remark_passes: Vec<CString>;
285 remark_passes_all = true;
286 remark_passes = Vec::new();
288 Passes::Some(passes) => {
289 remark_passes_all = false;
291 passes.iter().map(|name| CString::new(name.as_str()).unwrap()).collect();
294 let remark_passes: Vec<*const c_char> =
295 remark_passes.iter().map(|name: &CString| name.as_ptr()).collect();
296 let data = Box::into_raw(Box::new((cgcx, handler)));
298 let old_handler = llvm::LLVMRustContextGetDiagnosticHandler(llcx);
299 llvm::LLVMRustContextConfigureDiagnosticHandler(
304 remark_passes.as_ptr(),
307 llvm::LLVMRustSetInlineAsmDiagnosticHandler(llcx, inline_asm_handler, data.cast());
308 DiagnosticHandlers { data, llcx, old_handler }
313 impl<'a> Drop for DiagnosticHandlers<'a> {
315 use std::ptr::null_mut;
317 llvm::LLVMRustSetInlineAsmDiagnosticHandler(self.llcx, inline_asm_handler, null_mut());
318 llvm::LLVMRustContextSetDiagnosticHandler(self.llcx, self.old_handler);
319 drop(Box::from_raw(self.data));
324 fn report_inline_asm(
325 cgcx: &CodegenContext<LlvmCodegenBackend>,
327 level: llvm::DiagnosticLevel,
329 source: Option<(String, Vec<InnerSpan>)>,
331 // In LTO build we may get srcloc values from other crates which are invalid
332 // since they use a different source map. To be safe we just suppress these
334 if matches!(cgcx.lto, Lto::Fat | Lto::Thin) {
337 let level = match level {
338 llvm::DiagnosticLevel::Error => Level::Error { lint: false },
339 llvm::DiagnosticLevel::Warning => Level::Warning(None),
340 llvm::DiagnosticLevel::Note | llvm::DiagnosticLevel::Remark => Level::Note,
342 cgcx.diag_emitter.inline_asm_error(cookie as u32, msg, level, source);
345 unsafe extern "C" fn inline_asm_handler(diag: &SMDiagnostic, user: *const c_void, cookie: c_uint) {
349 let (cgcx, _) = *(user as *const (&CodegenContext<LlvmCodegenBackend>, &Handler));
351 let smdiag = llvm::diagnostic::SrcMgrDiagnostic::unpack(diag);
352 report_inline_asm(cgcx, smdiag.message, smdiag.level, cookie, smdiag.source);
355 unsafe extern "C" fn diagnostic_handler(info: &DiagnosticInfo, user: *mut c_void) {
359 let (cgcx, diag_handler) = *(user as *const (&CodegenContext<LlvmCodegenBackend>, &Handler));
361 match llvm::diagnostic::Diagnostic::unpack(info) {
362 llvm::diagnostic::InlineAsm(inline) => {
363 report_inline_asm(cgcx, inline.message, inline.level, inline.cookie, inline.source);
366 llvm::diagnostic::Optimization(opt) => {
367 let enabled = match cgcx.remark {
369 Passes::Some(ref v) => v.iter().any(|s| *s == opt.pass_name),
373 diag_handler.note_without_error(&format!(
375 opt.filename, opt.line, opt.column, opt.pass_name, opt.message,
379 llvm::diagnostic::PGO(diagnostic_ref) | llvm::diagnostic::Linker(diagnostic_ref) => {
380 let msg = llvm::build_string(|s| {
381 llvm::LLVMRustWriteDiagnosticInfoToString(diagnostic_ref, s)
383 .expect("non-UTF8 diagnostic");
384 diag_handler.warn(&msg);
386 llvm::diagnostic::Unsupported(diagnostic_ref) => {
387 let msg = llvm::build_string(|s| {
388 llvm::LLVMRustWriteDiagnosticInfoToString(diagnostic_ref, s)
390 .expect("non-UTF8 diagnostic");
391 diag_handler.err(&msg);
393 llvm::diagnostic::UnknownDiagnostic(..) => {}
397 fn get_pgo_gen_path(config: &ModuleConfig) -> Option<CString> {
398 match config.pgo_gen {
399 SwitchWithOptPath::Enabled(ref opt_dir_path) => {
400 let path = if let Some(dir_path) = opt_dir_path {
401 dir_path.join("default_%m.profraw")
403 PathBuf::from("default_%m.profraw")
406 Some(CString::new(format!("{}", path.display())).unwrap())
408 SwitchWithOptPath::Disabled => None,
412 fn get_pgo_use_path(config: &ModuleConfig) -> Option<CString> {
416 .map(|path_buf| CString::new(path_buf.to_string_lossy().as_bytes()).unwrap())
419 fn get_pgo_sample_use_path(config: &ModuleConfig) -> Option<CString> {
423 .map(|path_buf| CString::new(path_buf.to_string_lossy().as_bytes()).unwrap())
426 pub(crate) unsafe fn optimize_with_new_llvm_pass_manager(
427 cgcx: &CodegenContext<LlvmCodegenBackend>,
428 diag_handler: &Handler,
429 module: &ModuleCodegen<ModuleLlvm>,
430 config: &ModuleConfig,
431 opt_level: config::OptLevel,
432 opt_stage: llvm::OptStage,
433 ) -> Result<(), FatalError> {
435 opt_level != config::OptLevel::Size && opt_level != config::OptLevel::SizeMin;
436 let using_thin_buffers = opt_stage == llvm::OptStage::PreLinkThinLTO || config.bitcode_needed();
437 let pgo_gen_path = get_pgo_gen_path(config);
438 let pgo_use_path = get_pgo_use_path(config);
439 let pgo_sample_use_path = get_pgo_sample_use_path(config);
440 let is_lto = opt_stage == llvm::OptStage::ThinLTO || opt_stage == llvm::OptStage::FatLTO;
441 // Sanitizer instrumentation is only inserted during the pre-link optimization stage.
442 let sanitizer_options = if !is_lto {
443 Some(llvm::SanitizerOptions {
444 sanitize_address: config.sanitizer.contains(SanitizerSet::ADDRESS),
445 sanitize_address_recover: config.sanitizer_recover.contains(SanitizerSet::ADDRESS),
446 sanitize_memory: config.sanitizer.contains(SanitizerSet::MEMORY),
447 sanitize_memory_recover: config.sanitizer_recover.contains(SanitizerSet::MEMORY),
448 sanitize_memory_track_origins: config.sanitizer_memory_track_origins as c_int,
449 sanitize_thread: config.sanitizer.contains(SanitizerSet::THREAD),
450 sanitize_hwaddress: config.sanitizer.contains(SanitizerSet::HWADDRESS),
451 sanitize_hwaddress_recover: config.sanitizer_recover.contains(SanitizerSet::HWADDRESS),
457 let mut llvm_profiler = if cgcx.prof.llvm_recording_enabled() {
458 Some(LlvmSelfProfiler::new(cgcx.prof.get_self_profiler().unwrap()))
463 let llvm_selfprofiler =
464 llvm_profiler.as_mut().map(|s| s as *mut _ as *mut c_void).unwrap_or(std::ptr::null_mut());
466 let extra_passes = if !is_lto { config.passes.join(",") } else { "".to_string() };
468 let llvm_plugins = config.llvm_plugins.join(",");
470 // FIXME: NewPM doesn't provide a facility to pass custom InlineParams.
471 // We would have to add upstream support for this first, before we can support
472 // config.inline_threshold and our more aggressive default thresholds.
473 let result = llvm::LLVMRustOptimizeWithNewPassManager(
474 module.module_llvm.llmod(),
475 &*module.module_llvm.tm,
476 to_pass_builder_opt_level(opt_level),
478 config.no_prepopulate_passes,
479 config.verify_llvm_ir,
481 config.merge_functions,
483 config.vectorize_slp,
484 config.vectorize_loop,
486 config.emit_lifetime_markers,
487 sanitizer_options.as_ref(),
488 pgo_gen_path.as_ref().map_or(std::ptr::null(), |s| s.as_ptr()),
489 pgo_use_path.as_ref().map_or(std::ptr::null(), |s| s.as_ptr()),
490 config.instrument_coverage,
491 config.instrument_gcov,
492 pgo_sample_use_path.as_ref().map_or(std::ptr::null(), |s| s.as_ptr()),
493 config.debug_info_for_profiling,
495 selfprofile_before_pass_callback,
496 selfprofile_after_pass_callback,
497 extra_passes.as_ptr().cast(),
499 llvm_plugins.as_ptr().cast(),
502 result.into_result().map_err(|()| llvm_err(diag_handler, "failed to run LLVM passes"))
505 // Unsafe due to LLVM calls.
506 pub(crate) unsafe fn optimize(
507 cgcx: &CodegenContext<LlvmCodegenBackend>,
508 diag_handler: &Handler,
509 module: &ModuleCodegen<ModuleLlvm>,
510 config: &ModuleConfig,
511 ) -> Result<(), FatalError> {
512 let _timer = cgcx.prof.generic_activity_with_arg("LLVM_module_optimize", &*module.name);
514 let llmod = module.module_llvm.llmod();
515 let llcx = &*module.module_llvm.llcx;
516 let tm = &*module.module_llvm.tm;
517 let _handlers = DiagnosticHandlers::new(cgcx, diag_handler, llcx);
519 let module_name = module.name.clone();
520 let module_name = Some(&module_name[..]);
522 if let Some(false) = config.new_llvm_pass_manager && llvm_util::get_version() >= (15, 0, 0) {
524 "ignoring `-Z new-llvm-pass-manager=no`, which is no longer supported with LLVM 15",
528 if config.emit_no_opt_bc {
529 let out = cgcx.output_filenames.temp_path_ext("no-opt.bc", module_name);
530 let out = path_to_c_string(&out);
531 llvm::LLVMWriteBitcodeToFile(llmod, out.as_ptr());
534 if let Some(opt_level) = config.opt_level {
535 if llvm_util::should_use_new_llvm_pass_manager(
536 &config.new_llvm_pass_manager,
539 let opt_stage = match cgcx.lto {
540 Lto::Fat => llvm::OptStage::PreLinkFatLTO,
541 Lto::Thin | Lto::ThinLocal => llvm::OptStage::PreLinkThinLTO,
542 _ if cgcx.opts.cg.linker_plugin_lto.enabled() => llvm::OptStage::PreLinkThinLTO,
543 _ => llvm::OptStage::PreLinkNoLTO,
545 return optimize_with_new_llvm_pass_manager(
555 if cgcx.prof.llvm_recording_enabled() {
557 .warn("`-Z self-profile-events = llvm` requires `-Z new-llvm-pass-manager`");
560 // Create the two optimizing pass managers. These mirror what clang
561 // does, and are by populated by LLVM's default PassManagerBuilder.
562 // Each manager has a different set of passes, but they also share
563 // some common passes.
564 let fpm = llvm::LLVMCreateFunctionPassManagerForModule(llmod);
565 let mpm = llvm::LLVMCreatePassManager();
568 let find_pass = |pass_name: &str| {
569 let pass_name = SmallCStr::new(pass_name);
570 llvm::LLVMRustFindAndCreatePass(pass_name.as_ptr())
573 if config.verify_llvm_ir {
574 // Verification should run as the very first pass.
575 llvm::LLVMRustAddPass(fpm, find_pass("verify").unwrap());
578 let mut extra_passes = Vec::new();
579 let mut have_name_anon_globals_pass = false;
581 for pass_name in &config.passes {
582 if pass_name == "lint" {
583 // Linting should also be performed early, directly on the generated IR.
584 llvm::LLVMRustAddPass(fpm, find_pass("lint").unwrap());
588 if let Some(pass) = find_pass(pass_name) {
589 extra_passes.push(pass);
591 diag_handler.warn(&format!("unknown pass `{}`, ignoring", pass_name));
594 if pass_name == "name-anon-globals" {
595 have_name_anon_globals_pass = true;
599 // Instrumentation must be inserted before optimization,
600 // otherwise LLVM may optimize some functions away which
603 // This mirrors what Clang does in lib/CodeGen/BackendUtil.cpp.
604 if config.instrument_gcov {
605 llvm::LLVMRustAddPass(mpm, find_pass("insert-gcov-profiling").unwrap());
607 if config.instrument_coverage {
608 llvm::LLVMRustAddPass(mpm, find_pass("instrprof").unwrap());
610 if config.debug_info_for_profiling {
611 llvm::LLVMRustAddPass(mpm, find_pass("add-discriminators").unwrap());
614 add_sanitizer_passes(config, &mut extra_passes);
616 // Some options cause LLVM bitcode to be emitted, which uses ThinLTOBuffers, so we need
617 // to make sure we run LLVM's NameAnonGlobals pass when emitting bitcode; otherwise
618 // we'll get errors in LLVM.
619 let using_thin_buffers = config.bitcode_needed();
620 if !config.no_prepopulate_passes {
621 llvm::LLVMAddAnalysisPasses(tm, fpm);
622 llvm::LLVMAddAnalysisPasses(tm, mpm);
623 let opt_level = to_llvm_opt_settings(opt_level).0;
624 let prepare_for_thin_lto = cgcx.lto == Lto::Thin
625 || cgcx.lto == Lto::ThinLocal
626 || (cgcx.lto != Lto::Fat && cgcx.opts.cg.linker_plugin_lto.enabled());
627 with_llvm_pmb(llmod, config, opt_level, prepare_for_thin_lto, &mut |b| {
628 llvm::LLVMRustAddLastExtensionPasses(
630 extra_passes.as_ptr(),
631 extra_passes.len() as size_t,
633 llvm::LLVMRustPassManagerBuilderPopulateFunctionPassManager(b, fpm);
634 llvm::LLVMRustPassManagerBuilderPopulateModulePassManager(b, mpm);
637 have_name_anon_globals_pass = have_name_anon_globals_pass || prepare_for_thin_lto;
638 if using_thin_buffers && !prepare_for_thin_lto {
639 llvm::LLVMRustAddPass(mpm, find_pass("name-anon-globals").unwrap());
640 have_name_anon_globals_pass = true;
643 // If we don't use the standard pipeline, directly populate the MPM
644 // with the extra passes.
645 for pass in extra_passes {
646 llvm::LLVMRustAddPass(mpm, pass);
650 if using_thin_buffers && !have_name_anon_globals_pass {
651 // As described above, this will probably cause an error in LLVM
652 if config.no_prepopulate_passes {
654 "The current compilation is going to use thin LTO buffers \
655 without running LLVM's NameAnonGlobals pass. \
656 This will likely cause errors in LLVM. Consider adding \
657 -C passes=name-anon-globals to the compiler command line.",
661 "We are using thin LTO buffers without running the NameAnonGlobals pass. \
662 This will likely cause errors in LLVM and should never happen."
668 diag_handler.abort_if_errors();
670 // Finally, run the actual optimization passes
672 let _timer = cgcx.prof.extra_verbose_generic_activity(
673 "LLVM_module_optimize_function_passes",
676 llvm::LLVMRustRunFunctionPassManager(fpm, llmod);
679 let _timer = cgcx.prof.extra_verbose_generic_activity(
680 "LLVM_module_optimize_module_passes",
683 llvm::LLVMRunPassManager(mpm, llmod);
686 // Deallocate managers that we're now done with
687 llvm::LLVMDisposePassManager(fpm);
688 llvm::LLVMDisposePassManager(mpm);
693 unsafe fn add_sanitizer_passes(config: &ModuleConfig, passes: &mut Vec<&'static mut llvm::Pass>) {
694 if config.sanitizer.contains(SanitizerSet::ADDRESS) {
695 let recover = config.sanitizer_recover.contains(SanitizerSet::ADDRESS);
696 passes.push(llvm::LLVMRustCreateAddressSanitizerFunctionPass(recover));
697 passes.push(llvm::LLVMRustCreateModuleAddressSanitizerPass(recover));
699 if config.sanitizer.contains(SanitizerSet::MEMORY) {
700 let track_origins = config.sanitizer_memory_track_origins as c_int;
701 let recover = config.sanitizer_recover.contains(SanitizerSet::MEMORY);
702 passes.push(llvm::LLVMRustCreateMemorySanitizerPass(track_origins, recover));
704 if config.sanitizer.contains(SanitizerSet::THREAD) {
705 passes.push(llvm::LLVMRustCreateThreadSanitizerPass());
707 if config.sanitizer.contains(SanitizerSet::HWADDRESS) {
708 let recover = config.sanitizer_recover.contains(SanitizerSet::HWADDRESS);
709 passes.push(llvm::LLVMRustCreateHWAddressSanitizerPass(recover));
714 cgcx: &CodegenContext<LlvmCodegenBackend>,
715 diag_handler: &Handler,
716 mut modules: Vec<ModuleCodegen<ModuleLlvm>>,
717 ) -> Result<ModuleCodegen<ModuleLlvm>, FatalError> {
718 use super::lto::{Linker, ModuleBuffer};
719 // Sort the modules by name to ensure to ensure deterministic behavior.
720 modules.sort_by(|a, b| a.name.cmp(&b.name));
721 let (first, elements) =
722 modules.split_first().expect("Bug! modules must contain at least one module.");
724 let mut linker = Linker::new(first.module_llvm.llmod());
725 for module in elements {
726 let _timer = cgcx.prof.generic_activity_with_arg("LLVM_link_module", &*module.name);
727 let buffer = ModuleBuffer::new(module.module_llvm.llmod());
728 linker.add(buffer.data()).map_err(|()| {
729 let msg = format!("failed to serialize module {:?}", module.name);
730 llvm_err(diag_handler, &msg)
734 Ok(modules.remove(0))
737 pub(crate) unsafe fn codegen(
738 cgcx: &CodegenContext<LlvmCodegenBackend>,
739 diag_handler: &Handler,
740 module: ModuleCodegen<ModuleLlvm>,
741 config: &ModuleConfig,
742 ) -> Result<CompiledModule, FatalError> {
743 let _timer = cgcx.prof.generic_activity_with_arg("LLVM_module_codegen", &*module.name);
745 let llmod = module.module_llvm.llmod();
746 let llcx = &*module.module_llvm.llcx;
747 let tm = &*module.module_llvm.tm;
748 let module_name = module.name.clone();
749 let module_name = Some(&module_name[..]);
750 let handlers = DiagnosticHandlers::new(cgcx, diag_handler, llcx);
752 if cgcx.msvc_imps_needed {
753 create_msvc_imps(cgcx, llcx, llmod);
756 // A codegen-specific pass manager is used to generate object
757 // files for an LLVM module.
759 // Apparently each of these pass managers is a one-shot kind of
760 // thing, so we create a new one for each type of output. The
761 // pass manager passed to the closure should be ensured to not
762 // escape the closure itself, and the manager should only be
764 unsafe fn with_codegen<'ll, F, R>(
765 tm: &'ll llvm::TargetMachine,
766 llmod: &'ll llvm::Module,
771 F: FnOnce(&'ll mut PassManager<'ll>) -> R,
773 let cpm = llvm::LLVMCreatePassManager();
774 llvm::LLVMAddAnalysisPasses(tm, cpm);
775 llvm::LLVMRustAddLibraryInfo(cpm, llmod, no_builtins);
779 // Two things to note:
780 // - If object files are just LLVM bitcode we write bitcode, copy it to
781 // the .o file, and delete the bitcode if it wasn't otherwise
783 // - If we don't have the integrated assembler then we need to emit
784 // asm from LLVM and use `gcc` to create the object file.
786 let bc_out = cgcx.output_filenames.temp_path(OutputType::Bitcode, module_name);
787 let obj_out = cgcx.output_filenames.temp_path(OutputType::Object, module_name);
789 if config.bitcode_needed() {
792 .generic_activity_with_arg("LLVM_module_codegen_make_bitcode", &*module.name);
793 let thin = ThinBuffer::new(llmod, config.emit_thin_lto);
794 let data = thin.data();
796 if let Some(bitcode_filename) = bc_out.file_name() {
797 cgcx.prof.artifact_size(
799 bitcode_filename.to_string_lossy(),
804 if config.emit_bc || config.emit_obj == EmitObj::Bitcode {
807 .generic_activity_with_arg("LLVM_module_codegen_emit_bitcode", &*module.name);
808 if let Err(e) = fs::write(&bc_out, data) {
809 let msg = format!("failed to write bytecode to {}: {}", bc_out.display(), e);
810 diag_handler.err(&msg);
814 if config.emit_obj == EmitObj::ObjectCode(BitcodeSection::Full) {
817 .generic_activity_with_arg("LLVM_module_codegen_embed_bitcode", &*module.name);
818 embed_bitcode(cgcx, llcx, llmod, &config.bc_cmdline, data);
824 cgcx.prof.generic_activity_with_arg("LLVM_module_codegen_emit_ir", &*module.name);
825 let out = cgcx.output_filenames.temp_path(OutputType::LlvmAssembly, module_name);
826 let out_c = path_to_c_string(&out);
828 extern "C" fn demangle_callback(
829 input_ptr: *const c_char,
831 output_ptr: *mut c_char,
835 unsafe { slice::from_raw_parts(input_ptr as *const u8, input_len as usize) };
837 let Ok(input) = str::from_utf8(input) else { return 0 };
839 let output = unsafe {
840 slice::from_raw_parts_mut(output_ptr as *mut u8, output_len as usize)
842 let mut cursor = io::Cursor::new(output);
844 let Ok(demangled) = rustc_demangle::try_demangle(input) else { return 0 };
846 if write!(cursor, "{:#}", demangled).is_err() {
847 // Possible only if provided buffer is not big enough
851 cursor.position() as size_t
854 let result = llvm::LLVMRustPrintModule(llmod, out_c.as_ptr(), demangle_callback);
856 if result == llvm::LLVMRustResult::Success {
857 record_artifact_size(&cgcx.prof, "llvm_ir", &out);
860 result.into_result().map_err(|()| {
861 let msg = format!("failed to write LLVM IR to {}", out.display());
862 llvm_err(diag_handler, &msg)
868 cgcx.prof.generic_activity_with_arg("LLVM_module_codegen_emit_asm", &*module.name);
869 let path = cgcx.output_filenames.temp_path(OutputType::Assembly, module_name);
871 // We can't use the same module for asm and object code output,
872 // because that triggers various errors like invalid IR or broken
873 // binaries. So we must clone the module to produce the asm output
874 // if we are also producing object code.
875 let llmod = if let EmitObj::ObjectCode(_) = config.emit_obj {
876 llvm::LLVMCloneModule(llmod)
880 with_codegen(tm, llmod, config.no_builtins, |cpm| {
888 llvm::FileType::AssemblyFile,
894 match config.emit_obj {
895 EmitObj::ObjectCode(_) => {
898 .generic_activity_with_arg("LLVM_module_codegen_emit_obj", &*module.name);
900 let dwo_out = cgcx.output_filenames.temp_path_dwo(module_name);
901 let dwo_out = match (cgcx.split_debuginfo, cgcx.split_dwarf_kind) {
902 // Don't change how DWARF is emitted when disabled.
903 (SplitDebuginfo::Off, _) => None,
904 // Don't provide a DWARF object path if split debuginfo is enabled but this is
905 // a platform that doesn't support Split DWARF.
906 _ if !cgcx.target_can_use_split_dwarf => None,
907 // Don't provide a DWARF object path in single mode, sections will be written
908 // into the object as normal but ignored by linker.
909 (_, SplitDwarfKind::Single) => None,
910 // Emit (a subset of the) DWARF into a separate dwarf object file in split
912 (_, SplitDwarfKind::Split) => Some(dwo_out.as_path()),
915 with_codegen(tm, llmod, config.no_builtins, |cpm| {
923 llvm::FileType::ObjectFile,
929 EmitObj::Bitcode => {
930 debug!("copying bitcode {:?} to obj {:?}", bc_out, obj_out);
931 if let Err(e) = link_or_copy(&bc_out, &obj_out) {
932 diag_handler.err(&format!("failed to copy bitcode to object file: {}", e));
936 debug!("removing_bitcode {:?}", bc_out);
937 ensure_removed(diag_handler, &bc_out);
947 Ok(module.into_compiled_module(
948 config.emit_obj != EmitObj::None,
949 cgcx.target_can_use_split_dwarf
950 && cgcx.split_debuginfo != SplitDebuginfo::Off
951 && cgcx.split_dwarf_kind == SplitDwarfKind::Split,
953 &cgcx.output_filenames,
957 fn create_section_with_flags_asm(section_name: &str, section_flags: &str, data: &[u8]) -> Vec<u8> {
958 let mut asm = format!(".section {},\"{}\"\n", section_name, section_flags).into_bytes();
959 asm.extend_from_slice(b".ascii \"");
960 asm.reserve(data.len());
962 if byte == b'\\' || byte == b'"' {
965 } else if byte < 0x20 || byte >= 0x80 {
966 // Avoid non UTF-8 inline assembly. Use octal escape sequence, because it is fixed
967 // width, while hex escapes will consume following characters.
969 asm.push(b'0' + ((byte >> 6) & 0x7));
970 asm.push(b'0' + ((byte >> 3) & 0x7));
971 asm.push(b'0' + ((byte >> 0) & 0x7));
976 asm.extend_from_slice(b"\"\n");
980 /// Embed the bitcode of an LLVM module in the LLVM module itself.
982 /// This is done primarily for iOS where it appears to be standard to compile C
983 /// code at least with `-fembed-bitcode` which creates two sections in the
986 /// * __LLVM,__bitcode
987 /// * __LLVM,__cmdline
989 /// It appears *both* of these sections are necessary to get the linker to
990 /// recognize what's going on. A suitable cmdline value is taken from the
993 /// Furthermore debug/O1 builds don't actually embed bitcode but rather just
994 /// embed an empty section.
996 /// Basically all of this is us attempting to follow in the footsteps of clang
997 /// on iOS. See #35968 for lots more info.
998 unsafe fn embed_bitcode(
999 cgcx: &CodegenContext<LlvmCodegenBackend>,
1000 llcx: &llvm::Context,
1001 llmod: &llvm::Module,
1005 // We're adding custom sections to the output object file, but we definitely
1006 // do not want these custom sections to make their way into the final linked
1007 // executable. The purpose of these custom sections is for tooling
1008 // surrounding object files to work with the LLVM IR, if necessary. For
1009 // example rustc's own LTO will look for LLVM IR inside of the object file
1010 // in these sections by default.
1012 // To handle this is a bit different depending on the object file format
1013 // used by the backend, broken down into a few different categories:
1015 // * Mach-O - this is for macOS. Inspecting the source code for the native
1016 // linker here shows that the `.llvmbc` and `.llvmcmd` sections are
1017 // automatically skipped by the linker. In that case there's nothing extra
1018 // that we need to do here.
1020 // * Wasm - the native LLD linker is hard-coded to skip `.llvmbc` and
1021 // `.llvmcmd` sections, so there's nothing extra we need to do.
1023 // * COFF - if we don't do anything the linker will by default copy all
1024 // these sections to the output artifact, not what we want! To subvert
1025 // this we want to flag the sections we inserted here as
1026 // `IMAGE_SCN_LNK_REMOVE`.
1028 // * ELF - this is very similar to COFF above. One difference is that these
1029 // sections are removed from the output linked artifact when
1030 // `--gc-sections` is passed, which we pass by default. If that flag isn't
1031 // passed though then these sections will show up in the final output.
1032 // Additionally the flag that we need to set here is `SHF_EXCLUDE`.
1034 // Unfortunately, LLVM provides no way to set custom section flags. For ELF
1035 // and COFF we emit the sections using module level inline assembly for that
1036 // reason (see issue #90326 for historical background).
1037 let is_apple = cgcx.opts.target_triple.triple().contains("-ios")
1038 || cgcx.opts.target_triple.triple().contains("-darwin")
1039 || cgcx.opts.target_triple.triple().contains("-tvos")
1040 || cgcx.opts.target_triple.triple().contains("-watchos");
1042 || cgcx.opts.target_triple.triple().starts_with("wasm")
1043 || cgcx.opts.target_triple.triple().starts_with("asmjs")
1045 // We don't need custom section flags, create LLVM globals.
1046 let llconst = common::bytes_in_context(llcx, bitcode);
1047 let llglobal = llvm::LLVMAddGlobal(
1049 common::val_ty(llconst),
1050 "rustc.embedded.module\0".as_ptr().cast(),
1052 llvm::LLVMSetInitializer(llglobal, llconst);
1054 let section = if is_apple { "__LLVM,__bitcode\0" } else { ".llvmbc\0" };
1055 llvm::LLVMSetSection(llglobal, section.as_ptr().cast());
1056 llvm::LLVMRustSetLinkage(llglobal, llvm::Linkage::PrivateLinkage);
1057 llvm::LLVMSetGlobalConstant(llglobal, llvm::True);
1059 let llconst = common::bytes_in_context(llcx, cmdline.as_bytes());
1060 let llglobal = llvm::LLVMAddGlobal(
1062 common::val_ty(llconst),
1063 "rustc.embedded.cmdline\0".as_ptr().cast(),
1065 llvm::LLVMSetInitializer(llglobal, llconst);
1066 let section = if is_apple { "__LLVM,__cmdline\0" } else { ".llvmcmd\0" };
1067 llvm::LLVMSetSection(llglobal, section.as_ptr().cast());
1068 llvm::LLVMRustSetLinkage(llglobal, llvm::Linkage::PrivateLinkage);
1070 // We need custom section flags, so emit module-level inline assembly.
1071 let section_flags = if cgcx.is_pe_coff { "n" } else { "e" };
1072 let asm = create_section_with_flags_asm(".llvmbc", section_flags, bitcode);
1073 llvm::LLVMRustAppendModuleInlineAsm(llmod, asm.as_ptr().cast(), asm.len());
1074 let asm = create_section_with_flags_asm(".llvmcmd", section_flags, cmdline.as_bytes());
1075 llvm::LLVMRustAppendModuleInlineAsm(llmod, asm.as_ptr().cast(), asm.len());
1079 pub unsafe fn with_llvm_pmb(
1080 llmod: &llvm::Module,
1081 config: &ModuleConfig,
1082 opt_level: llvm::CodeGenOptLevel,
1083 prepare_for_thin_lto: bool,
1084 f: &mut dyn FnMut(&llvm::PassManagerBuilder),
1088 // Create the PassManagerBuilder for LLVM. We configure it with
1089 // reasonable defaults and prepare it to actually populate the pass
1091 let builder = llvm::LLVMRustPassManagerBuilderCreate();
1092 let opt_size = config.opt_size.map_or(llvm::CodeGenOptSizeNone, |x| to_llvm_opt_settings(x).1);
1093 let inline_threshold = config.inline_threshold;
1094 let pgo_gen_path = get_pgo_gen_path(config);
1095 let pgo_use_path = get_pgo_use_path(config);
1096 let pgo_sample_use_path = get_pgo_sample_use_path(config);
1098 llvm::LLVMRustConfigurePassManagerBuilder(
1101 config.merge_functions,
1102 config.vectorize_slp,
1103 config.vectorize_loop,
1104 prepare_for_thin_lto,
1105 pgo_gen_path.as_ref().map_or(ptr::null(), |s| s.as_ptr()),
1106 pgo_use_path.as_ref().map_or(ptr::null(), |s| s.as_ptr()),
1107 pgo_sample_use_path.as_ref().map_or(ptr::null(), |s| s.as_ptr()),
1111 llvm::LLVMRustAddBuilderLibraryInfo(builder, llmod, config.no_builtins);
1113 // Here we match what clang does (kinda). For O0 we only inline
1114 // always-inline functions (but don't add lifetime intrinsics), at O1 we
1115 // inline with lifetime intrinsics, and O2+ we add an inliner with a
1116 // thresholds copied from clang.
1117 match (opt_level, opt_size, inline_threshold) {
1119 llvm::LLVMRustPassManagerBuilderUseInlinerWithThreshold(builder, t);
1121 (llvm::CodeGenOptLevel::Aggressive, ..) => {
1122 llvm::LLVMRustPassManagerBuilderUseInlinerWithThreshold(builder, 275);
1124 (_, llvm::CodeGenOptSizeDefault, _) => {
1125 llvm::LLVMRustPassManagerBuilderUseInlinerWithThreshold(builder, 75);
1127 (_, llvm::CodeGenOptSizeAggressive, _) => {
1128 llvm::LLVMRustPassManagerBuilderUseInlinerWithThreshold(builder, 25);
1130 (llvm::CodeGenOptLevel::None, ..) => {
1131 llvm::LLVMRustAddAlwaysInlinePass(builder, config.emit_lifetime_markers);
1133 (llvm::CodeGenOptLevel::Less, ..) => {
1134 llvm::LLVMRustAddAlwaysInlinePass(builder, config.emit_lifetime_markers);
1136 (llvm::CodeGenOptLevel::Default, ..) => {
1137 llvm::LLVMRustPassManagerBuilderUseInlinerWithThreshold(builder, 225);
1142 llvm::LLVMRustPassManagerBuilderDispose(builder);
1145 // Create a `__imp_<symbol> = &symbol` global for every public static `symbol`.
1146 // This is required to satisfy `dllimport` references to static data in .rlibs
1147 // when using MSVC linker. We do this only for data, as linker can fix up
1148 // code references on its own.
1149 // See #26591, #27438
1150 fn create_msvc_imps(
1151 cgcx: &CodegenContext<LlvmCodegenBackend>,
1152 llcx: &llvm::Context,
1153 llmod: &llvm::Module,
1155 if !cgcx.msvc_imps_needed {
1158 // The x86 ABI seems to require that leading underscores are added to symbol
1159 // names, so we need an extra underscore on x86. There's also a leading
1160 // '\x01' here which disables LLVM's symbol mangling (e.g., no extra
1161 // underscores added in front).
1162 let prefix = if cgcx.target_arch == "x86" { "\x01__imp__" } else { "\x01__imp_" };
1165 let i8p_ty = Type::i8p_llcx(llcx);
1166 let globals = base::iter_globals(llmod)
1168 llvm::LLVMRustGetLinkage(val) == llvm::Linkage::ExternalLinkage
1169 && llvm::LLVMIsDeclaration(val) == 0
1172 // Exclude some symbols that we know are not Rust symbols.
1173 let name = llvm::get_value_name(val);
1174 if ignored(name) { None } else { Some((val, name)) }
1176 .map(move |(val, name)| {
1177 let mut imp_name = prefix.as_bytes().to_vec();
1178 imp_name.extend(name);
1179 let imp_name = CString::new(imp_name).unwrap();
1182 .collect::<Vec<_>>();
1184 for (imp_name, val) in globals {
1185 let imp = llvm::LLVMAddGlobal(llmod, i8p_ty, imp_name.as_ptr().cast());
1186 llvm::LLVMSetInitializer(imp, consts::ptrcast(val, i8p_ty));
1187 llvm::LLVMRustSetLinkage(imp, llvm::Linkage::ExternalLinkage);
1191 // Use this function to exclude certain symbols from `__imp` generation.
1192 fn ignored(symbol_name: &[u8]) -> bool {
1193 // These are symbols generated by LLVM's profiling instrumentation
1194 symbol_name.starts_with(b"__llvm_profile_")
1198 fn record_artifact_size(
1199 self_profiler_ref: &SelfProfilerRef,
1200 artifact_kind: &'static str,
1203 // Don't stat the file if we are not going to record its size.
1204 if !self_profiler_ref.enabled() {
1208 if let Some(artifact_name) = path.file_name() {
1209 let file_size = std::fs::metadata(path).map(|m| m.len()).unwrap_or(0);
1210 self_profiler_ref.artifact_size(artifact_kind, artifact_name.to_string_lossy(), file_size);