1 use crate::back::write::{
2 self, save_temp_bitcode, to_llvm_opt_settings, with_llvm_pmb, DiagnosticHandlers,
4 use crate::llvm::archive_ro::ArchiveRO;
5 use crate::llvm::{self, build_string, False, True};
6 use crate::{LlvmCodegenBackend, ModuleLlvm};
7 use rustc_codegen_ssa::back::lto::{LtoModuleCodegen, SerializedModule, ThinModule, ThinShared};
8 use rustc_codegen_ssa::back::symbol_export;
9 use rustc_codegen_ssa::back::write::{
10 CodegenContext, FatLTOInput, ModuleConfig, TargetMachineFactoryConfig,
12 use rustc_codegen_ssa::traits::*;
13 use rustc_codegen_ssa::{looks_like_rust_object_file, ModuleCodegen, ModuleKind};
14 use rustc_data_structures::fx::FxHashMap;
15 use rustc_errors::{FatalError, Handler};
16 use rustc_hir::def_id::LOCAL_CRATE;
17 use rustc_middle::bug;
18 use rustc_middle::dep_graph::WorkProduct;
19 use rustc_middle::middle::exported_symbols::SymbolExportLevel;
20 use rustc_session::cgu_reuse_tracker::CguReuse;
21 use rustc_session::config::{self, CrateType, Lto};
22 use tracing::{debug, info};
24 use std::ffi::{CStr, CString};
33 /// We keep track of the computed LTO cache keys from the previous
34 /// session to determine which CGUs we can reuse.
35 pub const THIN_LTO_KEYS_INCR_COMP_FILE_NAME: &str = "thin-lto-past-keys.bin";
37 pub fn crate_type_allows_lto(crate_type: CrateType) -> bool {
39 CrateType::Executable | CrateType::Staticlib | CrateType::Cdylib => true,
40 CrateType::Dylib | CrateType::Rlib | CrateType::ProcMacro => false,
45 cgcx: &CodegenContext<LlvmCodegenBackend>,
46 diag_handler: &Handler,
47 ) -> Result<(Vec<CString>, Vec<(SerializedModule<ModuleBuffer>, CString)>), FatalError> {
48 let export_threshold = match cgcx.lto {
49 // We're just doing LTO for our one crate
50 Lto::ThinLocal => SymbolExportLevel::Rust,
52 // We're doing LTO for the entire crate graph
53 Lto::Fat | Lto::Thin => symbol_export::crates_export_threshold(&cgcx.crate_types),
55 Lto::No => panic!("didn't request LTO but we're doing LTO"),
58 let symbol_filter = &|&(ref name, level): &(String, SymbolExportLevel)| {
59 if level.is_below_threshold(export_threshold) {
60 Some(CString::new(name.as_str()).unwrap())
65 let exported_symbols = cgcx.exported_symbols.as_ref().expect("needs exported symbols for LTO");
66 let mut symbols_below_threshold = {
67 let _timer = cgcx.prof.generic_activity("LLVM_lto_generate_symbols_below_threshold");
68 exported_symbols[&LOCAL_CRATE].iter().filter_map(symbol_filter).collect::<Vec<CString>>()
70 info!("{} symbols to preserve in this crate", symbols_below_threshold.len());
72 // If we're performing LTO for the entire crate graph, then for each of our
73 // upstream dependencies, find the corresponding rlib and load the bitcode
76 // We save off all the bytecode and LLVM module ids for later processing
77 // with either fat or thin LTO
78 let mut upstream_modules = Vec::new();
79 if cgcx.lto != Lto::ThinLocal {
80 if cgcx.opts.cg.prefer_dynamic {
82 .struct_err("cannot prefer dynamic linking when performing LTO")
84 "only 'staticlib', 'bin', and 'cdylib' outputs are \
88 return Err(FatalError);
91 // Make sure we actually can run LTO
92 for crate_type in cgcx.crate_types.iter() {
93 if !crate_type_allows_lto(*crate_type) {
94 let e = diag_handler.fatal(
95 "lto can only be run for executables, cdylibs and \
96 static library outputs",
102 for &(cnum, ref path) in cgcx.each_linked_rlib_for_lto.iter() {
103 let exported_symbols =
104 cgcx.exported_symbols.as_ref().expect("needs exported symbols for LTO");
107 cgcx.prof.generic_activity("LLVM_lto_generate_symbols_below_threshold");
108 symbols_below_threshold
109 .extend(exported_symbols[&cnum].iter().filter_map(symbol_filter));
112 let archive = ArchiveRO::open(&path).expect("wanted an rlib");
113 let obj_files = archive
115 .filter_map(|child| child.ok().and_then(|c| c.name().map(|name| (name, c))))
116 .filter(|&(name, _)| looks_like_rust_object_file(name));
117 for (name, child) in obj_files {
118 info!("adding bitcode from {}", name);
119 match get_bitcode_slice_from_object_data(child.data()) {
121 let module = SerializedModule::FromRlib(data.to_vec());
122 upstream_modules.push((module, CString::new(name).unwrap()));
124 Err(msg) => return Err(diag_handler.fatal(&msg)),
130 Ok((symbols_below_threshold, upstream_modules))
133 fn get_bitcode_slice_from_object_data(obj: &[u8]) -> Result<&[u8], String> {
136 unsafe { llvm::LLVMRustGetBitcodeSliceFromObjectData(obj.as_ptr(), obj.len(), &mut len) };
139 let bc = unsafe { slice::from_raw_parts(data, len) };
141 // `bc` must be a sub-slice of `obj`.
142 assert!(obj.as_ptr() <= bc.as_ptr());
143 assert!(bc[bc.len()..bc.len()].as_ptr() <= obj[obj.len()..obj.len()].as_ptr());
148 let msg = llvm::last_error().unwrap_or_else(|| "unknown LLVM error".to_string());
149 Err(format!("failed to get bitcode from object file for LTO ({})", msg))
153 /// Performs fat LTO by merging all modules into a single one and returning it
154 /// for further optimization.
155 pub(crate) fn run_fat(
156 cgcx: &CodegenContext<LlvmCodegenBackend>,
157 modules: Vec<FatLTOInput<LlvmCodegenBackend>>,
158 cached_modules: Vec<(SerializedModule<ModuleBuffer>, WorkProduct)>,
159 ) -> Result<LtoModuleCodegen<LlvmCodegenBackend>, FatalError> {
160 let diag_handler = cgcx.create_diag_handler();
161 let (symbols_below_threshold, upstream_modules) = prepare_lto(cgcx, &diag_handler)?;
162 let symbols_below_threshold =
163 symbols_below_threshold.iter().map(|c| c.as_ptr()).collect::<Vec<_>>();
170 &symbols_below_threshold,
174 /// Performs thin LTO by performing necessary global analysis and returning two
175 /// lists, one of the modules that need optimization and another for modules that
176 /// can simply be copied over from the incr. comp. cache.
177 pub(crate) fn run_thin(
178 cgcx: &CodegenContext<LlvmCodegenBackend>,
179 modules: Vec<(String, ThinBuffer)>,
180 cached_modules: Vec<(SerializedModule<ModuleBuffer>, WorkProduct)>,
181 ) -> Result<(Vec<LtoModuleCodegen<LlvmCodegenBackend>>, Vec<WorkProduct>), FatalError> {
182 let diag_handler = cgcx.create_diag_handler();
183 let (symbols_below_threshold, upstream_modules) = prepare_lto(cgcx, &diag_handler)?;
184 let symbols_below_threshold =
185 symbols_below_threshold.iter().map(|c| c.as_ptr()).collect::<Vec<_>>();
186 if cgcx.opts.cg.linker_plugin_lto.enabled() {
188 "We should never reach this case if the LTO step \
189 is deferred to the linker"
198 &symbols_below_threshold,
202 pub(crate) fn prepare_thin(module: ModuleCodegen<ModuleLlvm>) -> (String, ThinBuffer) {
203 let name = module.name.clone();
204 let buffer = ThinBuffer::new(module.module_llvm.llmod());
209 cgcx: &CodegenContext<LlvmCodegenBackend>,
210 diag_handler: &Handler,
211 modules: Vec<FatLTOInput<LlvmCodegenBackend>>,
212 cached_modules: Vec<(SerializedModule<ModuleBuffer>, WorkProduct)>,
213 mut serialized_modules: Vec<(SerializedModule<ModuleBuffer>, CString)>,
214 symbols_below_threshold: &[*const libc::c_char],
215 ) -> Result<LtoModuleCodegen<LlvmCodegenBackend>, FatalError> {
216 let _timer = cgcx.prof.generic_activity("LLVM_fat_lto_build_monolithic_module");
217 info!("going for a fat lto");
219 // Sort out all our lists of incoming modules into two lists.
221 // * `serialized_modules` (also and argument to this function) contains all
222 // modules that are serialized in-memory.
223 // * `in_memory` contains modules which are already parsed and in-memory,
224 // such as from multi-CGU builds.
226 // All of `cached_modules` (cached from previous incremental builds) can
227 // immediately go onto the `serialized_modules` modules list and then we can
228 // split the `modules` array into these two lists.
229 let mut in_memory = Vec::new();
230 serialized_modules.extend(cached_modules.into_iter().map(|(buffer, wp)| {
231 info!("pushing cached module {:?}", wp.cgu_name);
232 (buffer, CString::new(wp.cgu_name).unwrap())
234 for module in modules {
236 FatLTOInput::InMemory(m) => in_memory.push(m),
237 FatLTOInput::Serialized { name, buffer } => {
238 info!("pushing serialized module {:?}", name);
239 let buffer = SerializedModule::Local(buffer);
240 serialized_modules.push((buffer, CString::new(name).unwrap()));
245 // Find the "costliest" module and merge everything into that codegen unit.
246 // All the other modules will be serialized and reparsed into the new
247 // context, so this hopefully avoids serializing and parsing the largest
250 // Additionally use a regular module as the base here to ensure that various
251 // file copy operations in the backend work correctly. The only other kind
252 // of module here should be an allocator one, and if your crate is smaller
253 // than the allocator module then the size doesn't really matter anyway.
254 let costliest_module = in_memory
257 .filter(|&(_, module)| module.kind == ModuleKind::Regular)
259 let cost = unsafe { llvm::LLVMRustModuleCost(module.module_llvm.llmod()) };
264 // If we found a costliest module, we're good to go. Otherwise all our
265 // inputs were serialized which could happen in the case, for example, that
266 // all our inputs were incrementally reread from the cache and we're just
267 // re-executing the LTO passes. If that's the case deserialize the first
268 // module and create a linker with it.
269 let module: ModuleCodegen<ModuleLlvm> = match costliest_module {
270 Some((_cost, i)) => in_memory.remove(i),
272 assert!(!serialized_modules.is_empty(), "must have at least one serialized module");
273 let (buffer, name) = serialized_modules.remove(0);
274 info!("no in-memory regular modules to choose from, parsing {:?}", name);
276 module_llvm: ModuleLlvm::parse(cgcx, &name, buffer.data(), diag_handler)?,
277 name: name.into_string().unwrap(),
278 kind: ModuleKind::Regular,
282 let mut serialized_bitcode = Vec::new();
284 let (llcx, llmod) = {
285 let llvm = &module.module_llvm;
286 (&llvm.llcx, llvm.llmod())
288 info!("using {:?} as a base module", module.name);
290 // The linking steps below may produce errors and diagnostics within LLVM
291 // which we'd like to handle and print, so set up our diagnostic handlers
292 // (which get unregistered when they go out of scope below).
293 let _handler = DiagnosticHandlers::new(cgcx, diag_handler, llcx);
295 // For all other modules we codegened we'll need to link them into our own
296 // bitcode. All modules were codegened in their own LLVM context, however,
297 // and we want to move everything to the same LLVM context. Currently the
298 // way we know of to do that is to serialize them to a string and them parse
299 // them later. Not great but hey, that's why it's "fat" LTO, right?
300 for module in in_memory {
301 let buffer = ModuleBuffer::new(module.module_llvm.llmod());
302 let llmod_id = CString::new(&module.name[..]).unwrap();
303 serialized_modules.push((SerializedModule::Local(buffer), llmod_id));
305 // Sort the modules to ensure we produce deterministic results.
306 serialized_modules.sort_by(|module1, module2| module1.1.cmp(&module2.1));
308 // For all serialized bitcode files we parse them and link them in as we did
309 // above, this is all mostly handled in C++. Like above, though, we don't
310 // know much about the memory management here so we err on the side of being
311 // save and persist everything with the original module.
312 let mut linker = Linker::new(llmod);
313 for (bc_decoded, name) in serialized_modules {
316 .generic_activity_with_arg("LLVM_fat_lto_link_module", format!("{:?}", name));
317 info!("linking {:?}", name);
318 let data = bc_decoded.data();
319 linker.add(&data).map_err(|()| {
320 let msg = format!("failed to load bc of {:?}", name);
321 write::llvm_err(&diag_handler, &msg)
323 serialized_bitcode.push(bc_decoded);
326 save_temp_bitcode(&cgcx, &module, "lto.input");
328 // Internalize everything below threshold to help strip out more modules and such.
330 let ptr = symbols_below_threshold.as_ptr();
331 llvm::LLVMRustRunRestrictionPass(
333 ptr as *const *const libc::c_char,
334 symbols_below_threshold.len() as libc::size_t,
336 save_temp_bitcode(&cgcx, &module, "lto.after-restriction");
339 if cgcx.no_landing_pads {
341 llvm::LLVMRustMarkAllFunctionsNounwind(llmod);
343 save_temp_bitcode(&cgcx, &module, "lto.after-nounwind");
347 Ok(LtoModuleCodegen::Fat { module: Some(module), _serialized_bitcode: serialized_bitcode })
350 crate struct Linker<'a>(&'a mut llvm::Linker<'a>);
353 crate fn new(llmod: &'a llvm::Module) -> Self {
354 unsafe { Linker(llvm::LLVMRustLinkerNew(llmod)) }
357 crate fn add(&mut self, bytecode: &[u8]) -> Result<(), ()> {
359 if llvm::LLVMRustLinkerAdd(
361 bytecode.as_ptr() as *const libc::c_char,
372 impl Drop for Linker<'a> {
375 llvm::LLVMRustLinkerFree(&mut *(self.0 as *mut _));
380 /// Prepare "thin" LTO to get run on these modules.
382 /// The general structure of ThinLTO is quite different from the structure of
383 /// "fat" LTO above. With "fat" LTO all LLVM modules in question are merged into
384 /// one giant LLVM module, and then we run more optimization passes over this
385 /// big module after internalizing most symbols. Thin LTO, on the other hand,
386 /// avoid this large bottleneck through more targeted optimization.
388 /// At a high level Thin LTO looks like:
390 /// 1. Prepare a "summary" of each LLVM module in question which describes
391 /// the values inside, cost of the values, etc.
392 /// 2. Merge the summaries of all modules in question into one "index"
393 /// 3. Perform some global analysis on this index
394 /// 4. For each module, use the index and analysis calculated previously to
395 /// perform local transformations on the module, for example inlining
396 /// small functions from other modules.
397 /// 5. Run thin-specific optimization passes over each module, and then code
398 /// generate everything at the end.
400 /// The summary for each module is intended to be quite cheap, and the global
401 /// index is relatively quite cheap to create as well. As a result, the goal of
402 /// ThinLTO is to reduce the bottleneck on LTO and enable LTO to be used in more
403 /// situations. For example one cheap optimization is that we can parallelize
404 /// all codegen modules, easily making use of all the cores on a machine.
406 /// With all that in mind, the function here is designed at specifically just
407 /// calculating the *index* for ThinLTO. This index will then be shared amongst
408 /// all of the `LtoModuleCodegen` units returned below and destroyed once
409 /// they all go out of scope.
411 cgcx: &CodegenContext<LlvmCodegenBackend>,
412 diag_handler: &Handler,
413 modules: Vec<(String, ThinBuffer)>,
414 serialized_modules: Vec<(SerializedModule<ModuleBuffer>, CString)>,
415 cached_modules: Vec<(SerializedModule<ModuleBuffer>, WorkProduct)>,
416 symbols_below_threshold: &[*const libc::c_char],
417 ) -> Result<(Vec<LtoModuleCodegen<LlvmCodegenBackend>>, Vec<WorkProduct>), FatalError> {
418 let _timer = cgcx.prof.generic_activity("LLVM_thin_lto_global_analysis");
420 info!("going for that thin, thin LTO");
422 let green_modules: FxHashMap<_, _> =
423 cached_modules.iter().map(|&(_, ref wp)| (wp.cgu_name.clone(), wp.clone())).collect();
425 let full_scope_len = modules.len() + serialized_modules.len() + cached_modules.len();
426 let mut thin_buffers = Vec::with_capacity(modules.len());
427 let mut module_names = Vec::with_capacity(full_scope_len);
428 let mut thin_modules = Vec::with_capacity(full_scope_len);
430 for (i, (name, buffer)) in modules.into_iter().enumerate() {
431 info!("local module: {} - {}", i, name);
432 let cname = CString::new(name.clone()).unwrap();
433 thin_modules.push(llvm::ThinLTOModule {
434 identifier: cname.as_ptr(),
435 data: buffer.data().as_ptr(),
436 len: buffer.data().len(),
438 thin_buffers.push(buffer);
439 module_names.push(cname);
442 // FIXME: All upstream crates are deserialized internally in the
443 // function below to extract their summary and modules. Note that
444 // unlike the loop above we *must* decode and/or read something
445 // here as these are all just serialized files on disk. An
446 // improvement, however, to make here would be to store the
447 // module summary separately from the actual module itself. Right
448 // now this is store in one large bitcode file, and the entire
449 // file is deflate-compressed. We could try to bypass some of the
450 // decompression by storing the index uncompressed and only
451 // lazily decompressing the bytecode if necessary.
453 // Note that truly taking advantage of this optimization will
454 // likely be further down the road. We'd have to implement
455 // incremental ThinLTO first where we could actually avoid
456 // looking at upstream modules entirely sometimes (the contents,
457 // we must always unconditionally look at the index).
458 let mut serialized = Vec::with_capacity(serialized_modules.len() + cached_modules.len());
461 cached_modules.into_iter().map(|(sm, wp)| (sm, CString::new(wp.cgu_name).unwrap()));
463 for (module, name) in serialized_modules.into_iter().chain(cached_modules) {
464 info!("upstream or cached module {:?}", name);
465 thin_modules.push(llvm::ThinLTOModule {
466 identifier: name.as_ptr(),
467 data: module.data().as_ptr(),
468 len: module.data().len(),
470 serialized.push(module);
471 module_names.push(name);
475 assert_eq!(thin_modules.len(), module_names.len());
477 // Delegate to the C++ bindings to create some data here. Once this is a
478 // tried-and-true interface we may wish to try to upstream some of this
479 // to LLVM itself, right now we reimplement a lot of what they do
481 let data = llvm::LLVMRustCreateThinLTOData(
482 thin_modules.as_ptr(),
483 thin_modules.len() as u32,
484 symbols_below_threshold.as_ptr(),
485 symbols_below_threshold.len() as u32,
487 .ok_or_else(|| write::llvm_err(&diag_handler, "failed to prepare thin LTO context"))?;
489 let data = ThinData(data);
491 info!("thin LTO data created");
493 let (key_map_path, prev_key_map, curr_key_map) = if let Some(ref incr_comp_session_dir) =
494 cgcx.incr_comp_session_dir
496 let path = incr_comp_session_dir.join(THIN_LTO_KEYS_INCR_COMP_FILE_NAME);
497 // If the previous file was deleted, or we get an IO error
498 // reading the file, then we'll just use `None` as the
499 // prev_key_map, which will force the code to be recompiled.
501 if path.exists() { ThinLTOKeysMap::load_from_file(&path).ok() } else { None };
502 let curr = ThinLTOKeysMap::from_thin_lto_modules(&data, &thin_modules, &module_names);
503 (Some(path), prev, curr)
505 // If we don't compile incrementally, we don't need to load the
506 // import data from LLVM.
507 assert!(green_modules.is_empty());
508 let curr = ThinLTOKeysMap::default();
511 info!("thin LTO cache key map loaded");
512 info!("prev_key_map: {:#?}", prev_key_map);
513 info!("curr_key_map: {:#?}", curr_key_map);
515 // Throw our data in an `Arc` as we'll be sharing it across threads. We
516 // also put all memory referenced by the C++ data (buffers, ids, etc)
517 // into the arc as well. After this we'll create a thin module
518 // codegen per module in this data.
519 let shared = Arc::new(ThinShared {
522 serialized_modules: serialized,
526 let mut copy_jobs = vec![];
527 let mut opt_jobs = vec![];
529 info!("checking which modules can be-reused and which have to be re-optimized.");
530 for (module_index, module_name) in shared.module_names.iter().enumerate() {
531 let module_name = module_name_to_str(module_name);
532 if let (Some(prev_key_map), true) =
533 (prev_key_map.as_ref(), green_modules.contains_key(module_name))
535 assert!(cgcx.incr_comp_session_dir.is_some());
537 // If a module exists in both the current and the previous session,
538 // and has the same LTO cache key in both sessions, then we can re-use it
539 if prev_key_map.keys.get(module_name) == curr_key_map.keys.get(module_name) {
540 let work_product = green_modules[module_name].clone();
541 copy_jobs.push(work_product);
542 info!(" - {}: re-used", module_name);
543 assert!(cgcx.incr_comp_session_dir.is_some());
544 cgcx.cgu_reuse_tracker.set_actual_reuse(module_name, CguReuse::PostLto);
549 info!(" - {}: re-compiled", module_name);
550 opt_jobs.push(LtoModuleCodegen::Thin(ThinModule {
551 shared: shared.clone(),
556 // Save the current ThinLTO import information for the next compilation
557 // session, overwriting the previous serialized data (if any).
558 if let Some(path) = key_map_path {
559 if let Err(err) = curr_key_map.save_to_file(&path) {
560 let msg = format!("Error while writing ThinLTO key data: {}", err);
561 return Err(write::llvm_err(&diag_handler, &msg));
565 Ok((opt_jobs, copy_jobs))
569 pub(crate) fn run_pass_manager(
570 cgcx: &CodegenContext<LlvmCodegenBackend>,
571 module: &ModuleCodegen<ModuleLlvm>,
572 config: &ModuleConfig,
575 let _timer = cgcx.prof.extra_verbose_generic_activity("LLVM_lto_optimize", &module.name[..]);
577 // Now we have one massive module inside of llmod. Time to run the
578 // LTO-specific optimization passes that LLVM provides.
580 // This code is based off the code found in llvm's LTO code generator:
581 // tools/lto/LTOCodeGenerator.cpp
582 debug!("running the pass manager");
584 if write::should_use_new_llvm_pass_manager(config) {
585 let opt_stage = if thin { llvm::OptStage::ThinLTO } else { llvm::OptStage::FatLTO };
586 let opt_level = config.opt_level.unwrap_or(config::OptLevel::No);
587 // See comment below for why this is necessary.
588 let opt_level = if let config::OptLevel::No = opt_level {
589 config::OptLevel::Less
593 write::optimize_with_new_llvm_pass_manager(cgcx, module, config, opt_level, opt_stage);
598 let pm = llvm::LLVMCreatePassManager();
599 llvm::LLVMAddAnalysisPasses(module.module_llvm.tm, pm);
601 if config.verify_llvm_ir {
602 let pass = llvm::LLVMRustFindAndCreatePass("verify\0".as_ptr().cast());
603 llvm::LLVMRustAddPass(pm, pass.unwrap());
606 // When optimizing for LTO we don't actually pass in `-O0`, but we force
607 // it to always happen at least with `-O1`.
609 // With ThinLTO we mess around a lot with symbol visibility in a way
610 // that will actually cause linking failures if we optimize at O0 which
611 // notable is lacking in dead code elimination. To ensure we at least
612 // get some optimizations and correctly link we forcibly switch to `-O1`
613 // to get dead code elimination.
615 // Note that in general this shouldn't matter too much as you typically
616 // only turn on ThinLTO when you're compiling with optimizations
618 let opt_level = config
620 .map(|x| to_llvm_opt_settings(x).0)
621 .unwrap_or(llvm::CodeGenOptLevel::None);
622 let opt_level = match opt_level {
623 llvm::CodeGenOptLevel::None => llvm::CodeGenOptLevel::Less,
626 with_llvm_pmb(module.module_llvm.llmod(), config, opt_level, false, &mut |b| {
628 llvm::LLVMRustPassManagerBuilderPopulateThinLTOPassManager(b, pm);
630 llvm::LLVMPassManagerBuilderPopulateLTOPassManager(
631 b, pm, /* Internalize = */ False, /* RunInliner = */ True,
636 // We always generate bitcode through ThinLTOBuffers,
637 // which do not support anonymous globals
638 if config.bitcode_needed() {
639 let pass = llvm::LLVMRustFindAndCreatePass("name-anon-globals\0".as_ptr().cast());
640 llvm::LLVMRustAddPass(pm, pass.unwrap());
643 if config.verify_llvm_ir {
644 let pass = llvm::LLVMRustFindAndCreatePass("verify\0".as_ptr().cast());
645 llvm::LLVMRustAddPass(pm, pass.unwrap());
648 llvm::LLVMRunPassManager(pm, module.module_llvm.llmod());
650 llvm::LLVMDisposePassManager(pm);
655 pub struct ModuleBuffer(&'static mut llvm::ModuleBuffer);
657 unsafe impl Send for ModuleBuffer {}
658 unsafe impl Sync for ModuleBuffer {}
661 pub fn new(m: &llvm::Module) -> ModuleBuffer {
662 ModuleBuffer(unsafe { llvm::LLVMRustModuleBufferCreate(m) })
666 impl ModuleBufferMethods for ModuleBuffer {
667 fn data(&self) -> &[u8] {
669 let ptr = llvm::LLVMRustModuleBufferPtr(self.0);
670 let len = llvm::LLVMRustModuleBufferLen(self.0);
671 slice::from_raw_parts(ptr, len)
676 impl Drop for ModuleBuffer {
679 llvm::LLVMRustModuleBufferFree(&mut *(self.0 as *mut _));
684 pub struct ThinData(&'static mut llvm::ThinLTOData);
686 unsafe impl Send for ThinData {}
687 unsafe impl Sync for ThinData {}
689 impl Drop for ThinData {
692 llvm::LLVMRustFreeThinLTOData(&mut *(self.0 as *mut _));
697 pub struct ThinBuffer(&'static mut llvm::ThinLTOBuffer);
699 unsafe impl Send for ThinBuffer {}
700 unsafe impl Sync for ThinBuffer {}
703 pub fn new(m: &llvm::Module) -> ThinBuffer {
705 let buffer = llvm::LLVMRustThinLTOBufferCreate(m);
711 impl ThinBufferMethods for ThinBuffer {
712 fn data(&self) -> &[u8] {
714 let ptr = llvm::LLVMRustThinLTOBufferPtr(self.0) as *const _;
715 let len = llvm::LLVMRustThinLTOBufferLen(self.0);
716 slice::from_raw_parts(ptr, len)
721 impl Drop for ThinBuffer {
724 llvm::LLVMRustThinLTOBufferFree(&mut *(self.0 as *mut _));
729 pub unsafe fn optimize_thin_module(
730 thin_module: &mut ThinModule<LlvmCodegenBackend>,
731 cgcx: &CodegenContext<LlvmCodegenBackend>,
732 ) -> Result<ModuleCodegen<ModuleLlvm>, FatalError> {
733 let diag_handler = cgcx.create_diag_handler();
735 let module_name = &thin_module.shared.module_names[thin_module.idx];
736 let tm_factory_config = TargetMachineFactoryConfig::new(cgcx, module_name.to_str().unwrap());
738 (cgcx.tm_factory)(tm_factory_config).map_err(|e| write::llvm_err(&diag_handler, &e))?;
740 // Right now the implementation we've got only works over serialized
741 // modules, so we create a fresh new LLVM context and parse the module
742 // into that context. One day, however, we may do this for upstream
743 // crates but for locally codegened modules we may be able to reuse
744 // that LLVM Context and Module.
745 let llcx = llvm::LLVMRustContextCreate(cgcx.fewer_names);
747 parse_module(llcx, &module_name, thin_module.data(), &diag_handler)? as *const _;
748 let module = ModuleCodegen {
749 module_llvm: ModuleLlvm { llmod_raw, llcx, tm },
750 name: thin_module.name().to_string(),
751 kind: ModuleKind::Regular,
754 let target = &*module.module_llvm.tm;
755 let llmod = module.module_llvm.llmod();
756 save_temp_bitcode(&cgcx, &module, "thin-lto-input");
758 // Before we do much else find the "main" `DICompileUnit` that we'll be
759 // using below. If we find more than one though then rustc has changed
760 // in a way we're not ready for, so generate an ICE by returning
762 let mut cu1 = ptr::null_mut();
763 let mut cu2 = ptr::null_mut();
764 llvm::LLVMRustThinLTOGetDICompileUnit(llmod, &mut cu1, &mut cu2);
766 let msg = "multiple source DICompileUnits found";
767 return Err(write::llvm_err(&diag_handler, msg));
770 // Like with "fat" LTO, get some better optimizations if landing pads
771 // are disabled by removing all landing pads.
772 if cgcx.no_landing_pads {
775 .generic_activity_with_arg("LLVM_thin_lto_remove_landing_pads", thin_module.name());
776 llvm::LLVMRustMarkAllFunctionsNounwind(llmod);
777 save_temp_bitcode(&cgcx, &module, "thin-lto-after-nounwind");
780 // Up next comes the per-module local analyses that we do for Thin LTO.
781 // Each of these functions is basically copied from the LLVM
782 // implementation and then tailored to suit this implementation. Ideally
783 // each of these would be supported by upstream LLVM but that's perhaps
784 // a patch for another day!
786 // You can find some more comments about these functions in the LLVM
787 // bindings we've got (currently `PassWrapper.cpp`)
790 cgcx.prof.generic_activity_with_arg("LLVM_thin_lto_rename", thin_module.name());
791 if !llvm::LLVMRustPrepareThinLTORename(thin_module.shared.data.0, llmod, target) {
792 let msg = "failed to prepare thin LTO module";
793 return Err(write::llvm_err(&diag_handler, msg));
795 save_temp_bitcode(cgcx, &module, "thin-lto-after-rename");
801 .generic_activity_with_arg("LLVM_thin_lto_resolve_weak", thin_module.name());
802 if !llvm::LLVMRustPrepareThinLTOResolveWeak(thin_module.shared.data.0, llmod) {
803 let msg = "failed to prepare thin LTO module";
804 return Err(write::llvm_err(&diag_handler, msg));
806 save_temp_bitcode(cgcx, &module, "thin-lto-after-resolve");
812 .generic_activity_with_arg("LLVM_thin_lto_internalize", thin_module.name());
813 if !llvm::LLVMRustPrepareThinLTOInternalize(thin_module.shared.data.0, llmod) {
814 let msg = "failed to prepare thin LTO module";
815 return Err(write::llvm_err(&diag_handler, msg));
817 save_temp_bitcode(cgcx, &module, "thin-lto-after-internalize");
822 cgcx.prof.generic_activity_with_arg("LLVM_thin_lto_import", thin_module.name());
823 if !llvm::LLVMRustPrepareThinLTOImport(thin_module.shared.data.0, llmod, target) {
824 let msg = "failed to prepare thin LTO module";
825 return Err(write::llvm_err(&diag_handler, msg));
827 save_temp_bitcode(cgcx, &module, "thin-lto-after-import");
830 // Ok now this is a bit unfortunate. This is also something you won't
831 // find upstream in LLVM's ThinLTO passes! This is a hack for now to
832 // work around bugs in LLVM.
834 // First discovered in #45511 it was found that as part of ThinLTO
835 // importing passes LLVM will import `DICompileUnit` metadata
836 // information across modules. This means that we'll be working with one
837 // LLVM module that has multiple `DICompileUnit` instances in it (a
838 // bunch of `llvm.dbg.cu` members). Unfortunately there's a number of
839 // bugs in LLVM's backend which generates invalid DWARF in a situation
842 // https://bugs.llvm.org/show_bug.cgi?id=35212
843 // https://bugs.llvm.org/show_bug.cgi?id=35562
845 // While the first bug there is fixed the second ended up causing #46346
846 // which was basically a resurgence of #45511 after LLVM's bug 35212 was
849 // This function below is a huge hack around this problem. The function
850 // below is defined in `PassWrapper.cpp` and will basically "merge"
851 // all `DICompileUnit` instances in a module. Basically it'll take all
852 // the objects, rewrite all pointers of `DISubprogram` to point to the
853 // first `DICompileUnit`, and then delete all the other units.
855 // This is probably mangling to the debug info slightly (but hopefully
856 // not too much) but for now at least gets LLVM to emit valid DWARF (or
857 // so it appears). Hopefully we can remove this once upstream bugs are
862 .generic_activity_with_arg("LLVM_thin_lto_patch_debuginfo", thin_module.name());
863 llvm::LLVMRustThinLTOPatchDICompileUnit(llmod, cu1);
864 save_temp_bitcode(cgcx, &module, "thin-lto-after-patch");
867 // Alright now that we've done everything related to the ThinLTO
868 // analysis it's time to run some optimizations! Here we use the same
869 // `run_pass_manager` as the "fat" LTO above except that we tell it to
870 // populate a thin-specific pass manager, which presumably LLVM treats a
871 // little differently.
873 info!("running thin lto passes over {}", module.name);
874 let config = cgcx.config(module.kind);
875 run_pass_manager(cgcx, &module, config, true);
876 save_temp_bitcode(cgcx, &module, "thin-lto-after-pm");
882 /// Maps LLVM module identifiers to their corresponding LLVM LTO cache keys
883 #[derive(Debug, Default)]
884 pub struct ThinLTOKeysMap {
885 // key = llvm name of importing module, value = LLVM cache key
886 keys: FxHashMap<String, String>,
889 impl ThinLTOKeysMap {
890 fn save_to_file(&self, path: &Path) -> io::Result<()> {
892 let file = File::create(path)?;
893 let mut writer = io::BufWriter::new(file);
894 for (module, key) in &self.keys {
895 writeln!(writer, "{} {}", module, key)?;
900 fn load_from_file(path: &Path) -> io::Result<Self> {
901 use std::io::BufRead;
902 let mut keys = FxHashMap::default();
903 let file = File::open(path)?;
904 for line in io::BufReader::new(file).lines() {
906 let mut split = line.split(' ');
907 let module = split.next().unwrap();
908 let key = split.next().unwrap();
909 assert_eq!(split.next(), None, "Expected two space-separated values, found {:?}", line);
910 keys.insert(module.to_string(), key.to_string());
915 fn from_thin_lto_modules(
917 modules: &[llvm::ThinLTOModule],
920 let keys = iter::zip(modules, names)
921 .map(|(module, name)| {
922 let key = build_string(|rust_str| unsafe {
923 llvm::LLVMRustComputeLTOCacheKey(rust_str, module.identifier, data.0);
925 .expect("Invalid ThinLTO module key");
926 (name.clone().into_string().unwrap(), key)
933 fn module_name_to_str(c_str: &CStr) -> &str {
934 c_str.to_str().unwrap_or_else(|e| {
935 bug!("Encountered non-utf8 LLVM module name `{}`: {}", c_str.to_string_lossy(), e)
939 pub fn parse_module<'a>(
940 cx: &'a llvm::Context,
943 diag_handler: &Handler,
944 ) -> Result<&'a llvm::Module, FatalError> {
946 llvm::LLVMRustParseBitcodeForLTO(cx, data.as_ptr(), data.len(), name.as_ptr()).ok_or_else(
948 let msg = "failed to parse bitcode for LTO module";
949 write::llvm_err(&diag_handler, msg)