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::small_c_str::SmallCStr;
21 use rustc_errors::{FatalError, Handler, Level};
22 use rustc_fs_util::{link_or_copy, path_to_c_string};
23 use rustc_middle::bug;
24 use rustc_middle::ty::TyCtxt;
25 use rustc_session::config::{self, Lto, OutputType, Passes, SwitchWithOptPath};
26 use rustc_session::Session;
27 use rustc_span::symbol::sym;
28 use rustc_span::InnerSpan;
29 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(
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 ) -> Result<(), FatalError> {
58 let output_c = path_to_c_string(output);
59 let result = if let Some(dwo_output) = dwo_output {
60 let dwo_output_c = path_to_c_string(dwo_output);
61 llvm::LLVMRustWriteOutputFile(
66 dwo_output_c.as_ptr(),
70 llvm::LLVMRustWriteOutputFile(
79 result.into_result().map_err(|()| {
80 let msg = format!("could not write output to {}", output.display());
81 llvm_err(handler, &msg)
86 pub fn create_informational_target_machine(sess: &Session) -> &'static mut llvm::TargetMachine {
87 let config = TargetMachineFactoryConfig { split_dwarf_file: None };
88 target_machine_factory(sess, config::OptLevel::No)(config)
89 .unwrap_or_else(|err| llvm_err(sess.diagnostic(), &err).raise())
92 pub fn create_target_machine(tcx: TyCtxt<'_>, mod_name: &str) -> &'static mut llvm::TargetMachine {
93 let split_dwarf_file = if tcx.sess.target_can_use_split_dwarf() {
94 tcx.output_filenames(()).split_dwarf_path(tcx.sess.split_debuginfo(), Some(mod_name))
98 let config = TargetMachineFactoryConfig { split_dwarf_file };
99 target_machine_factory(tcx.sess, tcx.backend_optimization_level(()))(config)
100 .unwrap_or_else(|err| llvm_err(tcx.sess.diagnostic(), &err).raise())
103 pub fn to_llvm_opt_settings(
104 cfg: config::OptLevel,
105 ) -> (llvm::CodeGenOptLevel, llvm::CodeGenOptSize) {
106 use self::config::OptLevel::*;
108 No => (llvm::CodeGenOptLevel::None, llvm::CodeGenOptSizeNone),
109 Less => (llvm::CodeGenOptLevel::Less, llvm::CodeGenOptSizeNone),
110 Default => (llvm::CodeGenOptLevel::Default, llvm::CodeGenOptSizeNone),
111 Aggressive => (llvm::CodeGenOptLevel::Aggressive, llvm::CodeGenOptSizeNone),
112 Size => (llvm::CodeGenOptLevel::Default, llvm::CodeGenOptSizeDefault),
113 SizeMin => (llvm::CodeGenOptLevel::Default, llvm::CodeGenOptSizeAggressive),
117 fn to_pass_builder_opt_level(cfg: config::OptLevel) -> llvm::PassBuilderOptLevel {
118 use config::OptLevel::*;
120 No => llvm::PassBuilderOptLevel::O0,
121 Less => llvm::PassBuilderOptLevel::O1,
122 Default => llvm::PassBuilderOptLevel::O2,
123 Aggressive => llvm::PassBuilderOptLevel::O3,
124 Size => llvm::PassBuilderOptLevel::Os,
125 SizeMin => llvm::PassBuilderOptLevel::Oz,
129 fn to_llvm_relocation_model(relocation_model: RelocModel) -> llvm::RelocModel {
130 match relocation_model {
131 RelocModel::Static => llvm::RelocModel::Static,
132 // LLVM doesn't have a PIE relocation model, it represents PIE as PIC with an extra attribute.
133 RelocModel::Pic | RelocModel::Pie => llvm::RelocModel::PIC,
134 RelocModel::DynamicNoPic => llvm::RelocModel::DynamicNoPic,
135 RelocModel::Ropi => llvm::RelocModel::ROPI,
136 RelocModel::Rwpi => llvm::RelocModel::RWPI,
137 RelocModel::RopiRwpi => llvm::RelocModel::ROPI_RWPI,
141 pub(crate) fn to_llvm_code_model(code_model: Option<CodeModel>) -> llvm::CodeModel {
143 Some(CodeModel::Tiny) => llvm::CodeModel::Tiny,
144 Some(CodeModel::Small) => llvm::CodeModel::Small,
145 Some(CodeModel::Kernel) => llvm::CodeModel::Kernel,
146 Some(CodeModel::Medium) => llvm::CodeModel::Medium,
147 Some(CodeModel::Large) => llvm::CodeModel::Large,
148 None => llvm::CodeModel::None,
152 pub fn target_machine_factory(
154 optlvl: config::OptLevel,
155 ) -> TargetMachineFactoryFn<LlvmCodegenBackend> {
156 let reloc_model = to_llvm_relocation_model(sess.relocation_model());
158 let (opt_level, _) = to_llvm_opt_settings(optlvl);
159 let use_softfp = sess.opts.cg.soft_float;
161 let ffunction_sections =
162 sess.opts.debugging_opts.function_sections.unwrap_or(sess.target.function_sections);
163 let fdata_sections = ffunction_sections;
164 let funique_section_names = !sess.opts.debugging_opts.no_unique_section_names;
166 let code_model = to_llvm_code_model(sess.code_model());
168 let mut singlethread = sess.target.singlethread;
170 // On the wasm target once the `atomics` feature is enabled that means that
171 // we're no longer single-threaded, or otherwise we don't want LLVM to
172 // lower atomic operations to single-threaded operations.
173 if singlethread && sess.target.is_like_wasm && sess.target_features.contains(&sym::atomics) {
174 singlethread = false;
177 let triple = SmallCStr::new(&sess.target.llvm_target);
178 let cpu = SmallCStr::new(llvm_util::target_cpu(sess));
179 let features = llvm_util::llvm_global_features(sess).join(",");
180 let features = CString::new(features).unwrap();
181 let abi = SmallCStr::new(&sess.target.llvm_abiname);
182 let trap_unreachable =
183 sess.opts.debugging_opts.trap_unreachable.unwrap_or(sess.target.trap_unreachable);
184 let emit_stack_size_section = sess.opts.debugging_opts.emit_stack_sizes;
186 let asm_comments = sess.asm_comments();
187 let relax_elf_relocations =
188 sess.opts.debugging_opts.relax_elf_relocations.unwrap_or(sess.target.relax_elf_relocations);
191 !sess.opts.debugging_opts.use_ctors_section.unwrap_or(sess.target.use_ctors_section);
193 Arc::new(move |config: TargetMachineFactoryConfig| {
194 let split_dwarf_file = config.split_dwarf_file.unwrap_or_default();
195 let split_dwarf_file = CString::new(split_dwarf_file.to_str().unwrap()).unwrap();
198 llvm::LLVMRustCreateTargetMachine(
209 funique_section_names,
213 emit_stack_size_section,
214 relax_elf_relocations,
216 split_dwarf_file.as_ptr(),
221 format!("Could not create LLVM TargetMachine for triple: {}", triple.to_str().unwrap())
226 pub(crate) fn save_temp_bitcode(
227 cgcx: &CodegenContext<LlvmCodegenBackend>,
228 module: &ModuleCodegen<ModuleLlvm>,
231 if !cgcx.save_temps {
235 let ext = format!("{}.bc", name);
236 let cgu = Some(&module.name[..]);
237 let path = cgcx.output_filenames.temp_path_ext(&ext, cgu);
238 let cstr = path_to_c_string(&path);
239 let llmod = module.module_llvm.llmod();
240 llvm::LLVMWriteBitcodeToFile(llmod, cstr.as_ptr());
244 pub struct DiagnosticHandlers<'a> {
245 data: *mut (&'a CodegenContext<LlvmCodegenBackend>, &'a Handler),
246 llcx: &'a llvm::Context,
249 impl<'a> DiagnosticHandlers<'a> {
251 cgcx: &'a CodegenContext<LlvmCodegenBackend>,
252 handler: &'a Handler,
253 llcx: &'a llvm::Context,
255 let data = Box::into_raw(Box::new((cgcx, handler)));
257 llvm::LLVMRustSetInlineAsmDiagnosticHandler(llcx, inline_asm_handler, data.cast());
258 llvm::LLVMContextSetDiagnosticHandler(llcx, diagnostic_handler, data.cast());
260 DiagnosticHandlers { data, llcx }
264 impl<'a> Drop for DiagnosticHandlers<'a> {
266 use std::ptr::null_mut;
268 llvm::LLVMRustSetInlineAsmDiagnosticHandler(self.llcx, inline_asm_handler, null_mut());
269 llvm::LLVMContextSetDiagnosticHandler(self.llcx, diagnostic_handler, null_mut());
270 drop(Box::from_raw(self.data));
275 fn report_inline_asm(
276 cgcx: &CodegenContext<LlvmCodegenBackend>,
278 level: llvm::DiagnosticLevel,
280 source: Option<(String, Vec<InnerSpan>)>,
282 // In LTO build we may get srcloc values from other crates which are invalid
283 // since they use a different source map. To be safe we just suppress these
285 if matches!(cgcx.lto, Lto::Fat | Lto::Thin) {
288 let level = match level {
289 llvm::DiagnosticLevel::Error => Level::Error,
290 llvm::DiagnosticLevel::Warning => Level::Warning,
291 llvm::DiagnosticLevel::Note | llvm::DiagnosticLevel::Remark => Level::Note,
293 cgcx.diag_emitter.inline_asm_error(cookie as u32, msg, level, source);
296 unsafe extern "C" fn inline_asm_handler(diag: &SMDiagnostic, user: *const c_void, cookie: c_uint) {
300 let (cgcx, _) = *(user as *const (&CodegenContext<LlvmCodegenBackend>, &Handler));
302 let smdiag = llvm::diagnostic::SrcMgrDiagnostic::unpack(diag);
303 report_inline_asm(cgcx, smdiag.message, smdiag.level, cookie, smdiag.source);
306 unsafe extern "C" fn diagnostic_handler(info: &DiagnosticInfo, user: *mut c_void) {
310 let (cgcx, diag_handler) = *(user as *const (&CodegenContext<LlvmCodegenBackend>, &Handler));
312 match llvm::diagnostic::Diagnostic::unpack(info) {
313 llvm::diagnostic::InlineAsm(inline) => {
314 report_inline_asm(cgcx, inline.message, inline.level, inline.cookie, inline.source);
317 llvm::diagnostic::Optimization(opt) => {
318 let enabled = match cgcx.remark {
320 Passes::Some(ref v) => v.iter().any(|s| *s == opt.pass_name),
324 diag_handler.note_without_error(&format!(
325 "optimization {} for {} at {}:{}:{}: {}",
335 llvm::diagnostic::PGO(diagnostic_ref) | llvm::diagnostic::Linker(diagnostic_ref) => {
336 let msg = llvm::build_string(|s| {
337 llvm::LLVMRustWriteDiagnosticInfoToString(diagnostic_ref, s)
339 .expect("non-UTF8 diagnostic");
340 diag_handler.warn(&msg);
342 llvm::diagnostic::Unsupported(diagnostic_ref) => {
343 let msg = llvm::build_string(|s| {
344 llvm::LLVMRustWriteDiagnosticInfoToString(diagnostic_ref, s)
346 .expect("non-UTF8 diagnostic");
347 diag_handler.err(&msg);
349 llvm::diagnostic::UnknownDiagnostic(..) => {}
353 fn get_pgo_gen_path(config: &ModuleConfig) -> Option<CString> {
354 match config.pgo_gen {
355 SwitchWithOptPath::Enabled(ref opt_dir_path) => {
356 let path = if let Some(dir_path) = opt_dir_path {
357 dir_path.join("default_%m.profraw")
359 PathBuf::from("default_%m.profraw")
362 Some(CString::new(format!("{}", path.display())).unwrap())
364 SwitchWithOptPath::Disabled => None,
368 fn get_pgo_use_path(config: &ModuleConfig) -> Option<CString> {
372 .map(|path_buf| CString::new(path_buf.to_string_lossy().as_bytes()).unwrap())
375 fn get_pgo_sample_use_path(config: &ModuleConfig) -> Option<CString> {
379 .map(|path_buf| CString::new(path_buf.to_string_lossy().as_bytes()).unwrap())
382 pub(crate) fn should_use_new_llvm_pass_manager(
383 cgcx: &CodegenContext<LlvmCodegenBackend>,
384 config: &ModuleConfig,
386 // The new pass manager is enabled by default for LLVM >= 13.
387 // This matches Clang, which also enables it since Clang 13.
389 // FIXME: There are some perf issues with the new pass manager
390 // when targeting s390x, so it is temporarily disabled for that
391 // arch, see https://github.com/rust-lang/rust/issues/89609
393 .new_llvm_pass_manager
394 .unwrap_or_else(|| cgcx.target_arch != "s390x" && llvm_util::get_version() >= (13, 0, 0))
397 pub(crate) unsafe fn optimize_with_new_llvm_pass_manager(
398 cgcx: &CodegenContext<LlvmCodegenBackend>,
399 diag_handler: &Handler,
400 module: &ModuleCodegen<ModuleLlvm>,
401 config: &ModuleConfig,
402 opt_level: config::OptLevel,
403 opt_stage: llvm::OptStage,
404 ) -> Result<(), FatalError> {
406 opt_level != config::OptLevel::Size && opt_level != config::OptLevel::SizeMin;
407 let using_thin_buffers = opt_stage == llvm::OptStage::PreLinkThinLTO || config.bitcode_needed();
408 let pgo_gen_path = get_pgo_gen_path(config);
409 let pgo_use_path = get_pgo_use_path(config);
410 let pgo_sample_use_path = get_pgo_sample_use_path(config);
411 let is_lto = opt_stage == llvm::OptStage::ThinLTO || opt_stage == llvm::OptStage::FatLTO;
412 // Sanitizer instrumentation is only inserted during the pre-link optimization stage.
413 let sanitizer_options = if !is_lto {
414 Some(llvm::SanitizerOptions {
415 sanitize_address: config.sanitizer.contains(SanitizerSet::ADDRESS),
416 sanitize_address_recover: config.sanitizer_recover.contains(SanitizerSet::ADDRESS),
417 sanitize_memory: config.sanitizer.contains(SanitizerSet::MEMORY),
418 sanitize_memory_recover: config.sanitizer_recover.contains(SanitizerSet::MEMORY),
419 sanitize_memory_track_origins: config.sanitizer_memory_track_origins as c_int,
420 sanitize_thread: config.sanitizer.contains(SanitizerSet::THREAD),
421 sanitize_hwaddress: config.sanitizer.contains(SanitizerSet::HWADDRESS),
422 sanitize_hwaddress_recover: config.sanitizer_recover.contains(SanitizerSet::HWADDRESS),
428 let mut llvm_profiler = if cgcx.prof.llvm_recording_enabled() {
429 Some(LlvmSelfProfiler::new(cgcx.prof.get_self_profiler().unwrap()))
434 let llvm_selfprofiler =
435 llvm_profiler.as_mut().map(|s| s as *mut _ as *mut c_void).unwrap_or(std::ptr::null_mut());
437 let extra_passes = config.passes.join(",");
439 // FIXME: NewPM doesn't provide a facility to pass custom InlineParams.
440 // We would have to add upstream support for this first, before we can support
441 // config.inline_threshold and our more aggressive default thresholds.
442 let result = llvm::LLVMRustOptimizeWithNewPassManager(
443 module.module_llvm.llmod(),
444 &*module.module_llvm.tm,
445 to_pass_builder_opt_level(opt_level),
447 config.no_prepopulate_passes,
448 config.verify_llvm_ir,
450 config.merge_functions,
452 config.vectorize_slp,
453 config.vectorize_loop,
455 config.emit_lifetime_markers,
456 sanitizer_options.as_ref(),
457 pgo_gen_path.as_ref().map_or(std::ptr::null(), |s| s.as_ptr()),
458 pgo_use_path.as_ref().map_or(std::ptr::null(), |s| s.as_ptr()),
459 config.instrument_coverage,
460 config.instrument_gcov,
461 pgo_sample_use_path.as_ref().map_or(std::ptr::null(), |s| s.as_ptr()),
462 config.debug_info_for_profiling,
464 selfprofile_before_pass_callback,
465 selfprofile_after_pass_callback,
466 extra_passes.as_ptr().cast(),
469 result.into_result().map_err(|()| llvm_err(diag_handler, "failed to run LLVM passes"))
472 // Unsafe due to LLVM calls.
473 pub(crate) unsafe fn optimize(
474 cgcx: &CodegenContext<LlvmCodegenBackend>,
475 diag_handler: &Handler,
476 module: &ModuleCodegen<ModuleLlvm>,
477 config: &ModuleConfig,
478 ) -> Result<(), FatalError> {
479 let _timer = cgcx.prof.generic_activity_with_arg("LLVM_module_optimize", &module.name[..]);
481 let llmod = module.module_llvm.llmod();
482 let llcx = &*module.module_llvm.llcx;
483 let tm = &*module.module_llvm.tm;
484 let _handlers = DiagnosticHandlers::new(cgcx, diag_handler, llcx);
486 let module_name = module.name.clone();
487 let module_name = Some(&module_name[..]);
489 if config.emit_no_opt_bc {
490 let out = cgcx.output_filenames.temp_path_ext("no-opt.bc", module_name);
491 let out = path_to_c_string(&out);
492 llvm::LLVMWriteBitcodeToFile(llmod, out.as_ptr());
495 if let Some(opt_level) = config.opt_level {
496 if should_use_new_llvm_pass_manager(cgcx, config) {
497 let opt_stage = match cgcx.lto {
498 Lto::Fat => llvm::OptStage::PreLinkFatLTO,
499 Lto::Thin | Lto::ThinLocal => llvm::OptStage::PreLinkThinLTO,
500 _ if cgcx.opts.cg.linker_plugin_lto.enabled() => llvm::OptStage::PreLinkThinLTO,
501 _ => llvm::OptStage::PreLinkNoLTO,
503 return optimize_with_new_llvm_pass_manager(
513 if cgcx.prof.llvm_recording_enabled() {
515 .warn("`-Z self-profile-events = llvm` requires `-Z new-llvm-pass-manager`");
518 // Create the two optimizing pass managers. These mirror what clang
519 // does, and are by populated by LLVM's default PassManagerBuilder.
520 // Each manager has a different set of passes, but they also share
521 // some common passes.
522 let fpm = llvm::LLVMCreateFunctionPassManagerForModule(llmod);
523 let mpm = llvm::LLVMCreatePassManager();
526 let find_pass = |pass_name: &str| {
527 let pass_name = SmallCStr::new(pass_name);
528 llvm::LLVMRustFindAndCreatePass(pass_name.as_ptr())
531 if config.verify_llvm_ir {
532 // Verification should run as the very first pass.
533 llvm::LLVMRustAddPass(fpm, find_pass("verify").unwrap());
536 let mut extra_passes = Vec::new();
537 let mut have_name_anon_globals_pass = false;
539 for pass_name in &config.passes {
540 if pass_name == "lint" {
541 // Linting should also be performed early, directly on the generated IR.
542 llvm::LLVMRustAddPass(fpm, find_pass("lint").unwrap());
546 if let Some(pass) = find_pass(pass_name) {
547 extra_passes.push(pass);
549 diag_handler.warn(&format!("unknown pass `{}`, ignoring", pass_name));
552 if pass_name == "name-anon-globals" {
553 have_name_anon_globals_pass = true;
557 // Instrumentation must be inserted before optimization,
558 // otherwise LLVM may optimize some functions away which
561 // This mirrors what Clang does in lib/CodeGen/BackendUtil.cpp.
562 if config.instrument_gcov {
563 llvm::LLVMRustAddPass(mpm, find_pass("insert-gcov-profiling").unwrap());
565 if config.instrument_coverage {
566 llvm::LLVMRustAddPass(mpm, find_pass("instrprof").unwrap());
568 if config.debug_info_for_profiling {
569 llvm::LLVMRustAddPass(mpm, find_pass("add-discriminators").unwrap());
572 add_sanitizer_passes(config, &mut extra_passes);
574 // Some options cause LLVM bitcode to be emitted, which uses ThinLTOBuffers, so we need
575 // to make sure we run LLVM's NameAnonGlobals pass when emitting bitcode; otherwise
576 // we'll get errors in LLVM.
577 let using_thin_buffers = config.bitcode_needed();
578 if !config.no_prepopulate_passes {
579 llvm::LLVMAddAnalysisPasses(tm, fpm);
580 llvm::LLVMAddAnalysisPasses(tm, mpm);
581 let opt_level = to_llvm_opt_settings(opt_level).0;
582 let prepare_for_thin_lto = cgcx.lto == Lto::Thin
583 || cgcx.lto == Lto::ThinLocal
584 || (cgcx.lto != Lto::Fat && cgcx.opts.cg.linker_plugin_lto.enabled());
585 with_llvm_pmb(llmod, config, opt_level, prepare_for_thin_lto, &mut |b| {
586 llvm::LLVMRustAddLastExtensionPasses(
588 extra_passes.as_ptr(),
589 extra_passes.len() as size_t,
591 llvm::LLVMPassManagerBuilderPopulateFunctionPassManager(b, fpm);
592 llvm::LLVMPassManagerBuilderPopulateModulePassManager(b, mpm);
595 have_name_anon_globals_pass = have_name_anon_globals_pass || prepare_for_thin_lto;
596 if using_thin_buffers && !prepare_for_thin_lto {
597 llvm::LLVMRustAddPass(mpm, find_pass("name-anon-globals").unwrap());
598 have_name_anon_globals_pass = true;
601 // If we don't use the standard pipeline, directly populate the MPM
602 // with the extra passes.
603 for pass in extra_passes {
604 llvm::LLVMRustAddPass(mpm, pass);
608 if using_thin_buffers && !have_name_anon_globals_pass {
609 // As described above, this will probably cause an error in LLVM
610 if config.no_prepopulate_passes {
612 "The current compilation is going to use thin LTO buffers \
613 without running LLVM's NameAnonGlobals pass. \
614 This will likely cause errors in LLVM. Consider adding \
615 -C passes=name-anon-globals to the compiler command line.",
619 "We are using thin LTO buffers without running the NameAnonGlobals pass. \
620 This will likely cause errors in LLVM and should never happen."
626 diag_handler.abort_if_errors();
628 // Finally, run the actual optimization passes
630 let _timer = cgcx.prof.extra_verbose_generic_activity(
631 "LLVM_module_optimize_function_passes",
634 llvm::LLVMRustRunFunctionPassManager(fpm, llmod);
637 let _timer = cgcx.prof.extra_verbose_generic_activity(
638 "LLVM_module_optimize_module_passes",
641 llvm::LLVMRunPassManager(mpm, llmod);
644 // Deallocate managers that we're now done with
645 llvm::LLVMDisposePassManager(fpm);
646 llvm::LLVMDisposePassManager(mpm);
651 unsafe fn add_sanitizer_passes(config: &ModuleConfig, passes: &mut Vec<&'static mut llvm::Pass>) {
652 if config.sanitizer.contains(SanitizerSet::ADDRESS) {
653 let recover = config.sanitizer_recover.contains(SanitizerSet::ADDRESS);
654 passes.push(llvm::LLVMRustCreateAddressSanitizerFunctionPass(recover));
655 passes.push(llvm::LLVMRustCreateModuleAddressSanitizerPass(recover));
657 if config.sanitizer.contains(SanitizerSet::MEMORY) {
658 let track_origins = config.sanitizer_memory_track_origins as c_int;
659 let recover = config.sanitizer_recover.contains(SanitizerSet::MEMORY);
660 passes.push(llvm::LLVMRustCreateMemorySanitizerPass(track_origins, recover));
662 if config.sanitizer.contains(SanitizerSet::THREAD) {
663 passes.push(llvm::LLVMRustCreateThreadSanitizerPass());
665 if config.sanitizer.contains(SanitizerSet::HWADDRESS) {
666 let recover = config.sanitizer_recover.contains(SanitizerSet::HWADDRESS);
667 passes.push(llvm::LLVMRustCreateHWAddressSanitizerPass(recover));
672 cgcx: &CodegenContext<LlvmCodegenBackend>,
673 diag_handler: &Handler,
674 mut modules: Vec<ModuleCodegen<ModuleLlvm>>,
675 ) -> Result<ModuleCodegen<ModuleLlvm>, FatalError> {
676 use super::lto::{Linker, ModuleBuffer};
677 // Sort the modules by name to ensure to ensure deterministic behavior.
678 modules.sort_by(|a, b| a.name.cmp(&b.name));
679 let (first, elements) =
680 modules.split_first().expect("Bug! modules must contain at least one module.");
682 let mut linker = Linker::new(first.module_llvm.llmod());
683 for module in elements {
685 cgcx.prof.generic_activity_with_arg("LLVM_link_module", format!("{:?}", module.name));
686 let buffer = ModuleBuffer::new(module.module_llvm.llmod());
687 linker.add(buffer.data()).map_err(|()| {
688 let msg = format!("failed to serialize module {:?}", module.name);
689 llvm_err(diag_handler, &msg)
693 Ok(modules.remove(0))
696 pub(crate) unsafe fn codegen(
697 cgcx: &CodegenContext<LlvmCodegenBackend>,
698 diag_handler: &Handler,
699 module: ModuleCodegen<ModuleLlvm>,
700 config: &ModuleConfig,
701 ) -> Result<CompiledModule, FatalError> {
702 let _timer = cgcx.prof.generic_activity_with_arg("LLVM_module_codegen", &module.name[..]);
704 let llmod = module.module_llvm.llmod();
705 let llcx = &*module.module_llvm.llcx;
706 let tm = &*module.module_llvm.tm;
707 let module_name = module.name.clone();
708 let module_name = Some(&module_name[..]);
709 let handlers = DiagnosticHandlers::new(cgcx, diag_handler, llcx);
711 if cgcx.msvc_imps_needed {
712 create_msvc_imps(cgcx, llcx, llmod);
715 // A codegen-specific pass manager is used to generate object
716 // files for an LLVM module.
718 // Apparently each of these pass managers is a one-shot kind of
719 // thing, so we create a new one for each type of output. The
720 // pass manager passed to the closure should be ensured to not
721 // escape the closure itself, and the manager should only be
723 unsafe fn with_codegen<'ll, F, R>(
724 tm: &'ll llvm::TargetMachine,
725 llmod: &'ll llvm::Module,
730 F: FnOnce(&'ll mut PassManager<'ll>) -> R,
732 let cpm = llvm::LLVMCreatePassManager();
733 llvm::LLVMAddAnalysisPasses(tm, cpm);
734 llvm::LLVMRustAddLibraryInfo(cpm, llmod, no_builtins);
738 // Two things to note:
739 // - If object files are just LLVM bitcode we write bitcode, copy it to
740 // the .o file, and delete the bitcode if it wasn't otherwise
742 // - If we don't have the integrated assembler then we need to emit
743 // asm from LLVM and use `gcc` to create the object file.
745 let bc_out = cgcx.output_filenames.temp_path(OutputType::Bitcode, module_name);
746 let obj_out = cgcx.output_filenames.temp_path(OutputType::Object, module_name);
748 if config.bitcode_needed() {
751 .generic_activity_with_arg("LLVM_module_codegen_make_bitcode", &module.name[..]);
752 let thin = ThinBuffer::new(llmod);
753 let data = thin.data();
755 if config.emit_bc || config.emit_obj == EmitObj::Bitcode {
756 let _timer = cgcx.prof.generic_activity_with_arg(
757 "LLVM_module_codegen_emit_bitcode",
760 if let Err(e) = fs::write(&bc_out, data) {
761 let msg = format!("failed to write bytecode to {}: {}", bc_out.display(), e);
762 diag_handler.err(&msg);
766 if config.emit_obj == EmitObj::ObjectCode(BitcodeSection::Full) {
767 let _timer = cgcx.prof.generic_activity_with_arg(
768 "LLVM_module_codegen_embed_bitcode",
771 embed_bitcode(cgcx, llcx, llmod, &config.bc_cmdline, data);
778 .generic_activity_with_arg("LLVM_module_codegen_emit_ir", &module.name[..]);
779 let out = cgcx.output_filenames.temp_path(OutputType::LlvmAssembly, module_name);
780 let out_c = path_to_c_string(&out);
782 extern "C" fn demangle_callback(
783 input_ptr: *const c_char,
785 output_ptr: *mut c_char,
789 unsafe { slice::from_raw_parts(input_ptr as *const u8, input_len as usize) };
791 let input = match str::from_utf8(input) {
796 let output = unsafe {
797 slice::from_raw_parts_mut(output_ptr as *mut u8, output_len as usize)
799 let mut cursor = io::Cursor::new(output);
801 let demangled = match rustc_demangle::try_demangle(input) {
806 if write!(cursor, "{:#}", demangled).is_err() {
807 // Possible only if provided buffer is not big enough
811 cursor.position() as size_t
814 let result = llvm::LLVMRustPrintModule(llmod, out_c.as_ptr(), demangle_callback);
815 result.into_result().map_err(|()| {
816 let msg = format!("failed to write LLVM IR to {}", out.display());
817 llvm_err(diag_handler, &msg)
824 .generic_activity_with_arg("LLVM_module_codegen_emit_asm", &module.name[..]);
825 let path = cgcx.output_filenames.temp_path(OutputType::Assembly, module_name);
827 // We can't use the same module for asm and object code output,
828 // because that triggers various errors like invalid IR or broken
829 // binaries. So we must clone the module to produce the asm output
830 // if we are also producing object code.
831 let llmod = if let EmitObj::ObjectCode(_) = config.emit_obj {
832 llvm::LLVMCloneModule(llmod)
836 with_codegen(tm, llmod, config.no_builtins, |cpm| {
844 llvm::FileType::AssemblyFile,
849 match config.emit_obj {
850 EmitObj::ObjectCode(_) => {
853 .generic_activity_with_arg("LLVM_module_codegen_emit_obj", &module.name[..]);
855 let dwo_out = cgcx.output_filenames.temp_path_dwo(module_name);
856 let dwo_out = match cgcx.split_debuginfo {
857 // Don't change how DWARF is emitted in single mode (or when disabled).
858 SplitDebuginfo::Off | SplitDebuginfo::Packed => None,
859 // Emit (a subset of the) DWARF into a separate file in split mode.
860 SplitDebuginfo::Unpacked => {
861 if cgcx.target_can_use_split_dwarf {
862 Some(dwo_out.as_path())
869 with_codegen(tm, llmod, config.no_builtins, |cpm| {
877 llvm::FileType::ObjectFile,
882 EmitObj::Bitcode => {
883 debug!("copying bitcode {:?} to obj {:?}", bc_out, obj_out);
884 if let Err(e) = link_or_copy(&bc_out, &obj_out) {
885 diag_handler.err(&format!("failed to copy bitcode to object file: {}", e));
889 debug!("removing_bitcode {:?}", bc_out);
890 ensure_removed(diag_handler, &bc_out);
900 Ok(module.into_compiled_module(
901 config.emit_obj != EmitObj::None,
902 cgcx.target_can_use_split_dwarf && cgcx.split_debuginfo == SplitDebuginfo::Unpacked,
904 &cgcx.output_filenames,
908 /// Embed the bitcode of an LLVM module in the LLVM module itself.
910 /// This is done primarily for iOS where it appears to be standard to compile C
911 /// code at least with `-fembed-bitcode` which creates two sections in the
914 /// * __LLVM,__bitcode
915 /// * __LLVM,__cmdline
917 /// It appears *both* of these sections are necessary to get the linker to
918 /// recognize what's going on. A suitable cmdline value is taken from the
921 /// Furthermore debug/O1 builds don't actually embed bitcode but rather just
922 /// embed an empty section.
924 /// Basically all of this is us attempting to follow in the footsteps of clang
925 /// on iOS. See #35968 for lots more info.
926 unsafe fn embed_bitcode(
927 cgcx: &CodegenContext<LlvmCodegenBackend>,
928 llcx: &llvm::Context,
929 llmod: &llvm::Module,
933 let llconst = common::bytes_in_context(llcx, bitcode);
934 let llglobal = llvm::LLVMAddGlobal(
936 common::val_ty(llconst),
937 "rustc.embedded.module\0".as_ptr().cast(),
939 llvm::LLVMSetInitializer(llglobal, llconst);
941 let is_apple = cgcx.opts.target_triple.triple().contains("-ios")
942 || cgcx.opts.target_triple.triple().contains("-darwin")
943 || cgcx.opts.target_triple.triple().contains("-tvos");
945 let section = if is_apple { "__LLVM,__bitcode\0" } else { ".llvmbc\0" };
946 llvm::LLVMSetSection(llglobal, section.as_ptr().cast());
947 llvm::LLVMRustSetLinkage(llglobal, llvm::Linkage::PrivateLinkage);
948 llvm::LLVMSetGlobalConstant(llglobal, llvm::True);
950 let llconst = common::bytes_in_context(llcx, cmdline.as_bytes());
951 let llglobal = llvm::LLVMAddGlobal(
953 common::val_ty(llconst),
954 "rustc.embedded.cmdline\0".as_ptr().cast(),
956 llvm::LLVMSetInitializer(llglobal, llconst);
957 let section = if is_apple { "__LLVM,__cmdline\0" } else { ".llvmcmd\0" };
958 llvm::LLVMSetSection(llglobal, section.as_ptr().cast());
959 llvm::LLVMRustSetLinkage(llglobal, llvm::Linkage::PrivateLinkage);
961 // We're adding custom sections to the output object file, but we definitely
962 // do not want these custom sections to make their way into the final linked
963 // executable. The purpose of these custom sections is for tooling
964 // surrounding object files to work with the LLVM IR, if necessary. For
965 // example rustc's own LTO will look for LLVM IR inside of the object file
966 // in these sections by default.
968 // To handle this is a bit different depending on the object file format
969 // used by the backend, broken down into a few different categories:
971 // * Mach-O - this is for macOS. Inspecting the source code for the native
972 // linker here shows that the `.llvmbc` and `.llvmcmd` sections are
973 // automatically skipped by the linker. In that case there's nothing extra
974 // that we need to do here.
976 // * Wasm - the native LLD linker is hard-coded to skip `.llvmbc` and
977 // `.llvmcmd` sections, so there's nothing extra we need to do.
979 // * COFF - if we don't do anything the linker will by default copy all
980 // these sections to the output artifact, not what we want! To subvert
981 // this we want to flag the sections we inserted here as
982 // `IMAGE_SCN_LNK_REMOVE`. Unfortunately though LLVM has no native way to
983 // do this. Thankfully though we can do this with some inline assembly,
984 // which is easy enough to add via module-level global inline asm.
986 // * ELF - this is very similar to COFF above. One difference is that these
987 // sections are removed from the output linked artifact when
988 // `--gc-sections` is passed, which we pass by default. If that flag isn't
989 // passed though then these sections will show up in the final output.
990 // Additionally the flag that we need to set here is `SHF_EXCLUDE`.
992 || cgcx.opts.target_triple.triple().starts_with("wasm")
993 || cgcx.opts.target_triple.triple().starts_with("asmjs")
995 // nothing to do here
996 } else if cgcx.is_pe_coff {
998 .section .llvmbc,\"n\"
999 .section .llvmcmd,\"n\"
1001 llvm::LLVMRustAppendModuleInlineAsm(llmod, asm.as_ptr().cast(), asm.len());
1004 .section .llvmbc,\"e\"
1005 .section .llvmcmd,\"e\"
1007 llvm::LLVMRustAppendModuleInlineAsm(llmod, asm.as_ptr().cast(), asm.len());
1011 pub unsafe fn with_llvm_pmb(
1012 llmod: &llvm::Module,
1013 config: &ModuleConfig,
1014 opt_level: llvm::CodeGenOptLevel,
1015 prepare_for_thin_lto: bool,
1016 f: &mut dyn FnMut(&llvm::PassManagerBuilder),
1020 // Create the PassManagerBuilder for LLVM. We configure it with
1021 // reasonable defaults and prepare it to actually populate the pass
1023 let builder = llvm::LLVMPassManagerBuilderCreate();
1024 let opt_size = config.opt_size.map_or(llvm::CodeGenOptSizeNone, |x| to_llvm_opt_settings(x).1);
1025 let inline_threshold = config.inline_threshold;
1026 let pgo_gen_path = get_pgo_gen_path(config);
1027 let pgo_use_path = get_pgo_use_path(config);
1028 let pgo_sample_use_path = get_pgo_sample_use_path(config);
1030 llvm::LLVMRustConfigurePassManagerBuilder(
1033 config.merge_functions,
1034 config.vectorize_slp,
1035 config.vectorize_loop,
1036 prepare_for_thin_lto,
1037 pgo_gen_path.as_ref().map_or(ptr::null(), |s| s.as_ptr()),
1038 pgo_use_path.as_ref().map_or(ptr::null(), |s| s.as_ptr()),
1039 pgo_sample_use_path.as_ref().map_or(ptr::null(), |s| s.as_ptr()),
1042 llvm::LLVMPassManagerBuilderSetSizeLevel(builder, opt_size as u32);
1044 if opt_size != llvm::CodeGenOptSizeNone {
1045 llvm::LLVMPassManagerBuilderSetDisableUnrollLoops(builder, 1);
1048 llvm::LLVMRustAddBuilderLibraryInfo(builder, llmod, config.no_builtins);
1050 // Here we match what clang does (kinda). For O0 we only inline
1051 // always-inline functions (but don't add lifetime intrinsics), at O1 we
1052 // inline with lifetime intrinsics, and O2+ we add an inliner with a
1053 // thresholds copied from clang.
1054 match (opt_level, opt_size, inline_threshold) {
1056 llvm::LLVMPassManagerBuilderUseInlinerWithThreshold(builder, t);
1058 (llvm::CodeGenOptLevel::Aggressive, ..) => {
1059 llvm::LLVMPassManagerBuilderUseInlinerWithThreshold(builder, 275);
1061 (_, llvm::CodeGenOptSizeDefault, _) => {
1062 llvm::LLVMPassManagerBuilderUseInlinerWithThreshold(builder, 75);
1064 (_, llvm::CodeGenOptSizeAggressive, _) => {
1065 llvm::LLVMPassManagerBuilderUseInlinerWithThreshold(builder, 25);
1067 (llvm::CodeGenOptLevel::None, ..) => {
1068 llvm::LLVMRustAddAlwaysInlinePass(builder, config.emit_lifetime_markers);
1070 (llvm::CodeGenOptLevel::Less, ..) => {
1071 llvm::LLVMRustAddAlwaysInlinePass(builder, config.emit_lifetime_markers);
1073 (llvm::CodeGenOptLevel::Default, ..) => {
1074 llvm::LLVMPassManagerBuilderUseInlinerWithThreshold(builder, 225);
1079 llvm::LLVMPassManagerBuilderDispose(builder);
1082 // Create a `__imp_<symbol> = &symbol` global for every public static `symbol`.
1083 // This is required to satisfy `dllimport` references to static data in .rlibs
1084 // when using MSVC linker. We do this only for data, as linker can fix up
1085 // code references on its own.
1086 // See #26591, #27438
1087 fn create_msvc_imps(
1088 cgcx: &CodegenContext<LlvmCodegenBackend>,
1089 llcx: &llvm::Context,
1090 llmod: &llvm::Module,
1092 if !cgcx.msvc_imps_needed {
1095 // The x86 ABI seems to require that leading underscores are added to symbol
1096 // names, so we need an extra underscore on x86. There's also a leading
1097 // '\x01' here which disables LLVM's symbol mangling (e.g., no extra
1098 // underscores added in front).
1099 let prefix = if cgcx.target_arch == "x86" { "\x01__imp__" } else { "\x01__imp_" };
1102 let i8p_ty = Type::i8p_llcx(llcx);
1103 let globals = base::iter_globals(llmod)
1105 llvm::LLVMRustGetLinkage(val) == llvm::Linkage::ExternalLinkage
1106 && llvm::LLVMIsDeclaration(val) == 0
1109 // Exclude some symbols that we know are not Rust symbols.
1110 let name = llvm::get_value_name(val);
1111 if ignored(name) { None } else { Some((val, name)) }
1113 .map(move |(val, name)| {
1114 let mut imp_name = prefix.as_bytes().to_vec();
1115 imp_name.extend(name);
1116 let imp_name = CString::new(imp_name).unwrap();
1119 .collect::<Vec<_>>();
1121 for (imp_name, val) in globals {
1122 let imp = llvm::LLVMAddGlobal(llmod, i8p_ty, imp_name.as_ptr().cast());
1123 llvm::LLVMSetInitializer(imp, consts::ptrcast(val, i8p_ty));
1124 llvm::LLVMRustSetLinkage(imp, llvm::Linkage::ExternalLinkage);
1128 // Use this function to exclude certain symbols from `__imp` generation.
1129 fn ignored(symbol_name: &[u8]) -> bool {
1130 // These are symbols generated by LLVM's profiling instrumentation
1131 symbol_name.starts_with(b"__llvm_profile_")