1 use crate::back::write::{
2 self, save_temp_bitcode, to_llvm_opt_settings, with_llvm_pmb, DiagnosticHandlers,
4 use crate::llvm::{self, build_string, False, True};
5 use crate::{llvm_util, LlvmCodegenBackend, ModuleLlvm};
6 use object::read::archive::ArchiveFile;
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::{CodegenContext, FatLTOInput, TargetMachineFactoryConfig};
10 use rustc_codegen_ssa::traits::*;
11 use rustc_codegen_ssa::{looks_like_rust_object_file, ModuleCodegen, ModuleKind};
12 use rustc_data_structures::fx::FxHashMap;
13 use rustc_data_structures::memmap::Mmap;
14 use rustc_errors::{FatalError, Handler};
15 use rustc_hir::def_id::LOCAL_CRATE;
16 use rustc_middle::bug;
17 use rustc_middle::dep_graph::WorkProduct;
18 use rustc_middle::middle::exported_symbols::{SymbolExportInfo, SymbolExportLevel};
19 use rustc_session::cgu_reuse_tracker::CguReuse;
20 use rustc_session::config::{self, CrateType, Lto};
22 use std::ffi::{CStr, CString};
31 /// We keep track of the computed LTO cache keys from the previous
32 /// session to determine which CGUs we can reuse.
33 pub const THIN_LTO_KEYS_INCR_COMP_FILE_NAME: &str = "thin-lto-past-keys.bin";
35 pub fn crate_type_allows_lto(crate_type: CrateType) -> bool {
37 CrateType::Executable | CrateType::Staticlib | CrateType::Cdylib => true,
38 CrateType::Dylib | CrateType::Rlib | CrateType::ProcMacro => false,
43 cgcx: &CodegenContext<LlvmCodegenBackend>,
44 diag_handler: &Handler,
45 ) -> Result<(Vec<CString>, Vec<(SerializedModule<ModuleBuffer>, CString)>), FatalError> {
46 let export_threshold = match cgcx.lto {
47 // We're just doing LTO for our one crate
48 Lto::ThinLocal => SymbolExportLevel::Rust,
50 // We're doing LTO for the entire crate graph
51 Lto::Fat | Lto::Thin => symbol_export::crates_export_threshold(&cgcx.crate_types),
53 Lto::No => panic!("didn't request LTO but we're doing LTO"),
56 let symbol_filter = &|&(ref name, info): &(String, SymbolExportInfo)| {
57 if info.level.is_below_threshold(export_threshold) || info.used {
58 Some(CString::new(name.as_str()).unwrap())
63 let exported_symbols = cgcx.exported_symbols.as_ref().expect("needs exported symbols for LTO");
64 let mut symbols_below_threshold = {
65 let _timer = cgcx.prof.generic_activity("LLVM_lto_generate_symbols_below_threshold");
66 exported_symbols[&LOCAL_CRATE].iter().filter_map(symbol_filter).collect::<Vec<CString>>()
68 info!("{} symbols to preserve in this crate", symbols_below_threshold.len());
70 // If we're performing LTO for the entire crate graph, then for each of our
71 // upstream dependencies, find the corresponding rlib and load the bitcode
74 // We save off all the bytecode and LLVM module ids for later processing
75 // with either fat or thin LTO
76 let mut upstream_modules = Vec::new();
77 if cgcx.lto != Lto::ThinLocal {
78 if cgcx.opts.cg.prefer_dynamic {
80 .struct_err("cannot prefer dynamic linking when performing LTO")
82 "only 'staticlib', 'bin', and 'cdylib' outputs are \
86 return Err(FatalError);
89 // Make sure we actually can run LTO
90 for crate_type in cgcx.crate_types.iter() {
91 if !crate_type_allows_lto(*crate_type) {
92 let e = diag_handler.fatal(
93 "lto can only be run for executables, cdylibs and \
94 static library outputs",
100 for &(cnum, ref path) in cgcx.each_linked_rlib_for_lto.iter() {
101 let exported_symbols =
102 cgcx.exported_symbols.as_ref().expect("needs exported symbols for LTO");
105 cgcx.prof.generic_activity("LLVM_lto_generate_symbols_below_threshold");
106 symbols_below_threshold
107 .extend(exported_symbols[&cnum].iter().filter_map(symbol_filter));
110 let archive_data = unsafe {
111 Mmap::map(std::fs::File::open(&path).expect("couldn't open rlib"))
112 .expect("couldn't map rlib")
114 let archive = ArchiveFile::parse(&*archive_data).expect("wanted an rlib");
115 let obj_files = archive
117 .filter_map(|child| {
118 child.ok().and_then(|c| {
119 std::str::from_utf8(c.name()).ok().map(|name| (name.trim(), c))
122 .filter(|&(name, _)| looks_like_rust_object_file(name));
123 for (name, child) in obj_files {
124 info!("adding bitcode from {}", name);
125 match get_bitcode_slice_from_object_data(
126 child.data(&*archive_data).expect("corrupt rlib"),
129 let module = SerializedModule::FromRlib(data.to_vec());
130 upstream_modules.push((module, CString::new(name).unwrap()));
132 Err(msg) => return Err(diag_handler.fatal(&msg)),
138 Ok((symbols_below_threshold, upstream_modules))
141 fn get_bitcode_slice_from_object_data(obj: &[u8]) -> Result<&[u8], String> {
144 unsafe { llvm::LLVMRustGetBitcodeSliceFromObjectData(obj.as_ptr(), obj.len(), &mut len) };
147 let bc = unsafe { slice::from_raw_parts(data, len) };
149 // `bc` must be a sub-slice of `obj`.
150 assert!(obj.as_ptr() <= bc.as_ptr());
151 assert!(bc[bc.len()..bc.len()].as_ptr() <= obj[obj.len()..obj.len()].as_ptr());
156 let msg = llvm::last_error().unwrap_or_else(|| "unknown LLVM error".to_string());
157 Err(format!("failed to get bitcode from object file for LTO ({})", msg))
161 /// Performs fat LTO by merging all modules into a single one and returning it
162 /// for further optimization.
163 pub(crate) fn run_fat(
164 cgcx: &CodegenContext<LlvmCodegenBackend>,
165 modules: Vec<FatLTOInput<LlvmCodegenBackend>>,
166 cached_modules: Vec<(SerializedModule<ModuleBuffer>, WorkProduct)>,
167 ) -> Result<LtoModuleCodegen<LlvmCodegenBackend>, FatalError> {
168 let diag_handler = cgcx.create_diag_handler();
169 let (symbols_below_threshold, upstream_modules) = prepare_lto(cgcx, &diag_handler)?;
170 let symbols_below_threshold =
171 symbols_below_threshold.iter().map(|c| c.as_ptr()).collect::<Vec<_>>();
178 &symbols_below_threshold,
182 /// Performs thin LTO by performing necessary global analysis and returning two
183 /// lists, one of the modules that need optimization and another for modules that
184 /// can simply be copied over from the incr. comp. cache.
185 pub(crate) fn run_thin(
186 cgcx: &CodegenContext<LlvmCodegenBackend>,
187 modules: Vec<(String, ThinBuffer)>,
188 cached_modules: Vec<(SerializedModule<ModuleBuffer>, WorkProduct)>,
189 ) -> Result<(Vec<LtoModuleCodegen<LlvmCodegenBackend>>, Vec<WorkProduct>), FatalError> {
190 let diag_handler = cgcx.create_diag_handler();
191 let (symbols_below_threshold, upstream_modules) = prepare_lto(cgcx, &diag_handler)?;
192 let symbols_below_threshold =
193 symbols_below_threshold.iter().map(|c| c.as_ptr()).collect::<Vec<_>>();
194 if cgcx.opts.cg.linker_plugin_lto.enabled() {
196 "We should never reach this case if the LTO step \
197 is deferred to the linker"
206 &symbols_below_threshold,
210 pub(crate) fn prepare_thin(module: ModuleCodegen<ModuleLlvm>) -> (String, ThinBuffer) {
211 let name = module.name.clone();
212 let buffer = ThinBuffer::new(module.module_llvm.llmod(), true);
217 cgcx: &CodegenContext<LlvmCodegenBackend>,
218 diag_handler: &Handler,
219 modules: Vec<FatLTOInput<LlvmCodegenBackend>>,
220 cached_modules: Vec<(SerializedModule<ModuleBuffer>, WorkProduct)>,
221 mut serialized_modules: Vec<(SerializedModule<ModuleBuffer>, CString)>,
222 symbols_below_threshold: &[*const libc::c_char],
223 ) -> Result<LtoModuleCodegen<LlvmCodegenBackend>, FatalError> {
224 let _timer = cgcx.prof.generic_activity("LLVM_fat_lto_build_monolithic_module");
225 info!("going for a fat lto");
227 // Sort out all our lists of incoming modules into two lists.
229 // * `serialized_modules` (also and argument to this function) contains all
230 // modules that are serialized in-memory.
231 // * `in_memory` contains modules which are already parsed and in-memory,
232 // such as from multi-CGU builds.
234 // All of `cached_modules` (cached from previous incremental builds) can
235 // immediately go onto the `serialized_modules` modules list and then we can
236 // split the `modules` array into these two lists.
237 let mut in_memory = Vec::new();
238 serialized_modules.extend(cached_modules.into_iter().map(|(buffer, wp)| {
239 info!("pushing cached module {:?}", wp.cgu_name);
240 (buffer, CString::new(wp.cgu_name).unwrap())
242 for module in modules {
244 FatLTOInput::InMemory(m) => in_memory.push(m),
245 FatLTOInput::Serialized { name, buffer } => {
246 info!("pushing serialized module {:?}", name);
247 let buffer = SerializedModule::Local(buffer);
248 serialized_modules.push((buffer, CString::new(name).unwrap()));
253 // Find the "costliest" module and merge everything into that codegen unit.
254 // All the other modules will be serialized and reparsed into the new
255 // context, so this hopefully avoids serializing and parsing the largest
258 // Additionally use a regular module as the base here to ensure that various
259 // file copy operations in the backend work correctly. The only other kind
260 // of module here should be an allocator one, and if your crate is smaller
261 // than the allocator module then the size doesn't really matter anyway.
262 let costliest_module = in_memory
265 .filter(|&(_, module)| module.kind == ModuleKind::Regular)
267 let cost = unsafe { llvm::LLVMRustModuleCost(module.module_llvm.llmod()) };
272 // If we found a costliest module, we're good to go. Otherwise all our
273 // inputs were serialized which could happen in the case, for example, that
274 // all our inputs were incrementally reread from the cache and we're just
275 // re-executing the LTO passes. If that's the case deserialize the first
276 // module and create a linker with it.
277 let module: ModuleCodegen<ModuleLlvm> = match costliest_module {
278 Some((_cost, i)) => in_memory.remove(i),
280 assert!(!serialized_modules.is_empty(), "must have at least one serialized module");
281 let (buffer, name) = serialized_modules.remove(0);
282 info!("no in-memory regular modules to choose from, parsing {:?}", name);
284 module_llvm: ModuleLlvm::parse(cgcx, &name, buffer.data(), diag_handler)?,
285 name: name.into_string().unwrap(),
286 kind: ModuleKind::Regular,
290 let mut serialized_bitcode = Vec::new();
292 let (llcx, llmod) = {
293 let llvm = &module.module_llvm;
294 (&llvm.llcx, llvm.llmod())
296 info!("using {:?} as a base module", module.name);
298 // The linking steps below may produce errors and diagnostics within LLVM
299 // which we'd like to handle and print, so set up our diagnostic handlers
300 // (which get unregistered when they go out of scope below).
301 let _handler = DiagnosticHandlers::new(cgcx, diag_handler, llcx);
303 // For all other modules we codegened we'll need to link them into our own
304 // bitcode. All modules were codegened in their own LLVM context, however,
305 // and we want to move everything to the same LLVM context. Currently the
306 // way we know of to do that is to serialize them to a string and them parse
307 // them later. Not great but hey, that's why it's "fat" LTO, right?
308 for module in in_memory {
309 let buffer = ModuleBuffer::new(module.module_llvm.llmod());
310 let llmod_id = CString::new(&module.name[..]).unwrap();
311 serialized_modules.push((SerializedModule::Local(buffer), llmod_id));
313 // Sort the modules to ensure we produce deterministic results.
314 serialized_modules.sort_by(|module1, module2| module1.1.cmp(&module2.1));
316 // For all serialized bitcode files we parse them and link them in as we did
317 // above, this is all mostly handled in C++. Like above, though, we don't
318 // know much about the memory management here so we err on the side of being
319 // save and persist everything with the original module.
320 let mut linker = Linker::new(llmod);
321 for (bc_decoded, name) in serialized_modules {
324 .generic_activity_with_arg_recorder("LLVM_fat_lto_link_module", |recorder| {
325 recorder.record_arg(format!("{:?}", name))
327 info!("linking {:?}", name);
328 let data = bc_decoded.data();
329 linker.add(data).map_err(|()| {
330 let msg = format!("failed to load bitcode of module {:?}", name);
331 write::llvm_err(diag_handler, &msg)
333 serialized_bitcode.push(bc_decoded);
336 save_temp_bitcode(cgcx, &module, "lto.input");
338 // Internalize everything below threshold to help strip out more modules and such.
340 let ptr = symbols_below_threshold.as_ptr();
341 llvm::LLVMRustRunRestrictionPass(
343 ptr as *const *const libc::c_char,
344 symbols_below_threshold.len() as libc::size_t,
346 save_temp_bitcode(cgcx, &module, "lto.after-restriction");
350 Ok(LtoModuleCodegen::Fat { module, _serialized_bitcode: serialized_bitcode })
353 pub(crate) struct Linker<'a>(&'a mut llvm::Linker<'a>);
355 impl<'a> Linker<'a> {
356 pub(crate) fn new(llmod: &'a llvm::Module) -> Self {
357 unsafe { Linker(llvm::LLVMRustLinkerNew(llmod)) }
360 pub(crate) fn add(&mut self, bytecode: &[u8]) -> Result<(), ()> {
362 if llvm::LLVMRustLinkerAdd(
364 bytecode.as_ptr() as *const libc::c_char,
375 impl Drop for Linker<'_> {
378 llvm::LLVMRustLinkerFree(&mut *(self.0 as *mut _));
383 /// Prepare "thin" LTO to get run on these modules.
385 /// The general structure of ThinLTO is quite different from the structure of
386 /// "fat" LTO above. With "fat" LTO all LLVM modules in question are merged into
387 /// one giant LLVM module, and then we run more optimization passes over this
388 /// big module after internalizing most symbols. Thin LTO, on the other hand,
389 /// avoid this large bottleneck through more targeted optimization.
391 /// At a high level Thin LTO looks like:
393 /// 1. Prepare a "summary" of each LLVM module in question which describes
394 /// the values inside, cost of the values, etc.
395 /// 2. Merge the summaries of all modules in question into one "index"
396 /// 3. Perform some global analysis on this index
397 /// 4. For each module, use the index and analysis calculated previously to
398 /// perform local transformations on the module, for example inlining
399 /// small functions from other modules.
400 /// 5. Run thin-specific optimization passes over each module, and then code
401 /// generate everything at the end.
403 /// The summary for each module is intended to be quite cheap, and the global
404 /// index is relatively quite cheap to create as well. As a result, the goal of
405 /// ThinLTO is to reduce the bottleneck on LTO and enable LTO to be used in more
406 /// situations. For example one cheap optimization is that we can parallelize
407 /// all codegen modules, easily making use of all the cores on a machine.
409 /// With all that in mind, the function here is designed at specifically just
410 /// calculating the *index* for ThinLTO. This index will then be shared amongst
411 /// all of the `LtoModuleCodegen` units returned below and destroyed once
412 /// they all go out of scope.
414 cgcx: &CodegenContext<LlvmCodegenBackend>,
415 diag_handler: &Handler,
416 modules: Vec<(String, ThinBuffer)>,
417 serialized_modules: Vec<(SerializedModule<ModuleBuffer>, CString)>,
418 cached_modules: Vec<(SerializedModule<ModuleBuffer>, WorkProduct)>,
419 symbols_below_threshold: &[*const libc::c_char],
420 ) -> Result<(Vec<LtoModuleCodegen<LlvmCodegenBackend>>, Vec<WorkProduct>), FatalError> {
421 let _timer = cgcx.prof.generic_activity("LLVM_thin_lto_global_analysis");
423 info!("going for that thin, thin LTO");
425 let green_modules: FxHashMap<_, _> =
426 cached_modules.iter().map(|&(_, ref wp)| (wp.cgu_name.clone(), wp.clone())).collect();
428 let full_scope_len = modules.len() + serialized_modules.len() + cached_modules.len();
429 let mut thin_buffers = Vec::with_capacity(modules.len());
430 let mut module_names = Vec::with_capacity(full_scope_len);
431 let mut thin_modules = Vec::with_capacity(full_scope_len);
433 for (i, (name, buffer)) in modules.into_iter().enumerate() {
434 info!("local module: {} - {}", i, name);
435 let cname = CString::new(name.clone()).unwrap();
436 thin_modules.push(llvm::ThinLTOModule {
437 identifier: cname.as_ptr(),
438 data: buffer.data().as_ptr(),
439 len: buffer.data().len(),
441 thin_buffers.push(buffer);
442 module_names.push(cname);
445 // FIXME: All upstream crates are deserialized internally in the
446 // function below to extract their summary and modules. Note that
447 // unlike the loop above we *must* decode and/or read something
448 // here as these are all just serialized files on disk. An
449 // improvement, however, to make here would be to store the
450 // module summary separately from the actual module itself. Right
451 // now this is store in one large bitcode file, and the entire
452 // file is deflate-compressed. We could try to bypass some of the
453 // decompression by storing the index uncompressed and only
454 // lazily decompressing the bytecode if necessary.
456 // Note that truly taking advantage of this optimization will
457 // likely be further down the road. We'd have to implement
458 // incremental ThinLTO first where we could actually avoid
459 // looking at upstream modules entirely sometimes (the contents,
460 // we must always unconditionally look at the index).
461 let mut serialized = Vec::with_capacity(serialized_modules.len() + cached_modules.len());
464 cached_modules.into_iter().map(|(sm, wp)| (sm, CString::new(wp.cgu_name).unwrap()));
466 for (module, name) in serialized_modules.into_iter().chain(cached_modules) {
467 info!("upstream or cached module {:?}", name);
468 thin_modules.push(llvm::ThinLTOModule {
469 identifier: name.as_ptr(),
470 data: module.data().as_ptr(),
471 len: module.data().len(),
473 serialized.push(module);
474 module_names.push(name);
478 assert_eq!(thin_modules.len(), module_names.len());
480 // Delegate to the C++ bindings to create some data here. Once this is a
481 // tried-and-true interface we may wish to try to upstream some of this
482 // to LLVM itself, right now we reimplement a lot of what they do
484 let data = llvm::LLVMRustCreateThinLTOData(
485 thin_modules.as_ptr(),
486 thin_modules.len() as u32,
487 symbols_below_threshold.as_ptr(),
488 symbols_below_threshold.len() as u32,
490 .ok_or_else(|| write::llvm_err(diag_handler, "failed to prepare thin LTO context"))?;
492 let data = ThinData(data);
494 info!("thin LTO data created");
496 let (key_map_path, prev_key_map, curr_key_map) = if let Some(ref incr_comp_session_dir) =
497 cgcx.incr_comp_session_dir
499 let path = incr_comp_session_dir.join(THIN_LTO_KEYS_INCR_COMP_FILE_NAME);
500 // If the previous file was deleted, or we get an IO error
501 // reading the file, then we'll just use `None` as the
502 // prev_key_map, which will force the code to be recompiled.
504 if path.exists() { ThinLTOKeysMap::load_from_file(&path).ok() } else { None };
505 let curr = ThinLTOKeysMap::from_thin_lto_modules(&data, &thin_modules, &module_names);
506 (Some(path), prev, curr)
508 // If we don't compile incrementally, we don't need to load the
509 // import data from LLVM.
510 assert!(green_modules.is_empty());
511 let curr = ThinLTOKeysMap::default();
514 info!("thin LTO cache key map loaded");
515 info!("prev_key_map: {:#?}", prev_key_map);
516 info!("curr_key_map: {:#?}", curr_key_map);
518 // Throw our data in an `Arc` as we'll be sharing it across threads. We
519 // also put all memory referenced by the C++ data (buffers, ids, etc)
520 // into the arc as well. After this we'll create a thin module
521 // codegen per module in this data.
522 let shared = Arc::new(ThinShared {
525 serialized_modules: serialized,
529 let mut copy_jobs = vec![];
530 let mut opt_jobs = vec![];
532 info!("checking which modules can be-reused and which have to be re-optimized.");
533 for (module_index, module_name) in shared.module_names.iter().enumerate() {
534 let module_name = module_name_to_str(module_name);
535 if let (Some(prev_key_map), true) =
536 (prev_key_map.as_ref(), green_modules.contains_key(module_name))
538 assert!(cgcx.incr_comp_session_dir.is_some());
540 // If a module exists in both the current and the previous session,
541 // and has the same LTO cache key in both sessions, then we can re-use it
542 if prev_key_map.keys.get(module_name) == curr_key_map.keys.get(module_name) {
543 let work_product = green_modules[module_name].clone();
544 copy_jobs.push(work_product);
545 info!(" - {}: re-used", module_name);
546 assert!(cgcx.incr_comp_session_dir.is_some());
547 cgcx.cgu_reuse_tracker.set_actual_reuse(module_name, CguReuse::PostLto);
552 info!(" - {}: re-compiled", module_name);
553 opt_jobs.push(LtoModuleCodegen::Thin(ThinModule {
554 shared: shared.clone(),
559 // Save the current ThinLTO import information for the next compilation
560 // session, overwriting the previous serialized data (if any).
561 if let Some(path) = key_map_path {
562 if let Err(err) = curr_key_map.save_to_file(&path) {
563 let msg = format!("Error while writing ThinLTO key data: {}", err);
564 return Err(write::llvm_err(diag_handler, &msg));
568 Ok((opt_jobs, copy_jobs))
572 pub(crate) fn run_pass_manager(
573 cgcx: &CodegenContext<LlvmCodegenBackend>,
574 diag_handler: &Handler,
575 module: &mut ModuleCodegen<ModuleLlvm>,
577 ) -> Result<(), FatalError> {
578 let _timer = cgcx.prof.extra_verbose_generic_activity("LLVM_lto_optimize", &*module.name);
579 let config = cgcx.config(module.kind);
581 // Now we have one massive module inside of llmod. Time to run the
582 // LTO-specific optimization passes that LLVM provides.
584 // This code is based off the code found in llvm's LTO code generator:
585 // llvm/lib/LTO/LTOCodeGenerator.cpp
586 debug!("running the pass manager");
588 if !llvm::LLVMRustHasModuleFlag(
589 module.module_llvm.llmod(),
590 "LTOPostLink".as_ptr().cast(),
593 llvm::LLVMRustAddModuleFlag(
594 module.module_llvm.llmod(),
595 llvm::LLVMModFlagBehavior::Error,
596 "LTOPostLink\0".as_ptr().cast(),
600 if llvm_util::should_use_new_llvm_pass_manager(
601 &config.new_llvm_pass_manager,
604 let opt_stage = if thin { llvm::OptStage::ThinLTO } else { llvm::OptStage::FatLTO };
605 let opt_level = config.opt_level.unwrap_or(config::OptLevel::No);
606 write::optimize_with_new_llvm_pass_manager(
618 let pm = llvm::LLVMCreatePassManager();
619 llvm::LLVMAddAnalysisPasses(module.module_llvm.tm, pm);
621 if config.verify_llvm_ir {
622 let pass = llvm::LLVMRustFindAndCreatePass("verify\0".as_ptr().cast());
623 llvm::LLVMRustAddPass(pm, pass.unwrap());
626 let opt_level = config
628 .map(|x| to_llvm_opt_settings(x).0)
629 .unwrap_or(llvm::CodeGenOptLevel::None);
630 with_llvm_pmb(module.module_llvm.llmod(), config, opt_level, false, &mut |b| {
632 llvm::LLVMRustPassManagerBuilderPopulateThinLTOPassManager(b, pm);
634 llvm::LLVMRustPassManagerBuilderPopulateLTOPassManager(
635 b, pm, /* Internalize = */ False, /* RunInliner = */ True,
640 // We always generate bitcode through ThinLTOBuffers,
641 // which do not support anonymous globals
642 if config.bitcode_needed() {
643 let pass = llvm::LLVMRustFindAndCreatePass("name-anon-globals\0".as_ptr().cast());
644 llvm::LLVMRustAddPass(pm, pass.unwrap());
647 if config.verify_llvm_ir {
648 let pass = llvm::LLVMRustFindAndCreatePass("verify\0".as_ptr().cast());
649 llvm::LLVMRustAddPass(pm, pass.unwrap());
652 llvm::LLVMRunPassManager(pm, module.module_llvm.llmod());
654 llvm::LLVMDisposePassManager(pm);
660 pub struct ModuleBuffer(&'static mut llvm::ModuleBuffer);
662 unsafe impl Send for ModuleBuffer {}
663 unsafe impl Sync for ModuleBuffer {}
666 pub fn new(m: &llvm::Module) -> ModuleBuffer {
667 ModuleBuffer(unsafe { llvm::LLVMRustModuleBufferCreate(m) })
671 impl ModuleBufferMethods for ModuleBuffer {
672 fn data(&self) -> &[u8] {
674 let ptr = llvm::LLVMRustModuleBufferPtr(self.0);
675 let len = llvm::LLVMRustModuleBufferLen(self.0);
676 slice::from_raw_parts(ptr, len)
681 impl Drop for ModuleBuffer {
684 llvm::LLVMRustModuleBufferFree(&mut *(self.0 as *mut _));
689 pub struct ThinData(&'static mut llvm::ThinLTOData);
691 unsafe impl Send for ThinData {}
692 unsafe impl Sync for ThinData {}
694 impl Drop for ThinData {
697 llvm::LLVMRustFreeThinLTOData(&mut *(self.0 as *mut _));
702 pub struct ThinBuffer(&'static mut llvm::ThinLTOBuffer);
704 unsafe impl Send for ThinBuffer {}
705 unsafe impl Sync for ThinBuffer {}
708 pub fn new(m: &llvm::Module, is_thin: bool) -> ThinBuffer {
710 let buffer = llvm::LLVMRustThinLTOBufferCreate(m, is_thin);
716 impl ThinBufferMethods for ThinBuffer {
717 fn data(&self) -> &[u8] {
719 let ptr = llvm::LLVMRustThinLTOBufferPtr(self.0) as *const _;
720 let len = llvm::LLVMRustThinLTOBufferLen(self.0);
721 slice::from_raw_parts(ptr, len)
726 impl Drop for ThinBuffer {
729 llvm::LLVMRustThinLTOBufferFree(&mut *(self.0 as *mut _));
734 pub unsafe fn optimize_thin_module(
735 thin_module: ThinModule<LlvmCodegenBackend>,
736 cgcx: &CodegenContext<LlvmCodegenBackend>,
737 ) -> Result<ModuleCodegen<ModuleLlvm>, FatalError> {
738 let diag_handler = cgcx.create_diag_handler();
740 let module_name = &thin_module.shared.module_names[thin_module.idx];
741 let tm_factory_config = TargetMachineFactoryConfig::new(cgcx, module_name.to_str().unwrap());
743 (cgcx.tm_factory)(tm_factory_config).map_err(|e| write::llvm_err(&diag_handler, &e))?;
745 // Right now the implementation we've got only works over serialized
746 // modules, so we create a fresh new LLVM context and parse the module
747 // into that context. One day, however, we may do this for upstream
748 // crates but for locally codegened modules we may be able to reuse
749 // that LLVM Context and Module.
750 let llcx = llvm::LLVMRustContextCreate(cgcx.fewer_names);
751 let llmod_raw = parse_module(llcx, module_name, thin_module.data(), &diag_handler)? as *const _;
752 let mut module = ModuleCodegen {
753 module_llvm: ModuleLlvm { llmod_raw, llcx, tm },
754 name: thin_module.name().to_string(),
755 kind: ModuleKind::Regular,
758 let target = &*module.module_llvm.tm;
759 let llmod = module.module_llvm.llmod();
760 save_temp_bitcode(cgcx, &module, "thin-lto-input");
762 // Before we do much else find the "main" `DICompileUnit` that we'll be
763 // using below. If we find more than one though then rustc has changed
764 // in a way we're not ready for, so generate an ICE by returning
766 let mut cu1 = ptr::null_mut();
767 let mut cu2 = ptr::null_mut();
768 llvm::LLVMRustThinLTOGetDICompileUnit(llmod, &mut cu1, &mut cu2);
770 let msg = "multiple source DICompileUnits found";
771 return Err(write::llvm_err(&diag_handler, msg));
774 // Up next comes the per-module local analyses that we do for Thin LTO.
775 // Each of these functions is basically copied from the LLVM
776 // implementation and then tailored to suit this implementation. Ideally
777 // each of these would be supported by upstream LLVM but that's perhaps
778 // a patch for another day!
780 // You can find some more comments about these functions in the LLVM
781 // bindings we've got (currently `PassWrapper.cpp`)
784 cgcx.prof.generic_activity_with_arg("LLVM_thin_lto_rename", thin_module.name());
785 if !llvm::LLVMRustPrepareThinLTORename(thin_module.shared.data.0, llmod, target) {
786 let msg = "failed to prepare thin LTO module";
787 return Err(write::llvm_err(&diag_handler, msg));
789 save_temp_bitcode(cgcx, &module, "thin-lto-after-rename");
795 .generic_activity_with_arg("LLVM_thin_lto_resolve_weak", thin_module.name());
796 if !llvm::LLVMRustPrepareThinLTOResolveWeak(thin_module.shared.data.0, llmod) {
797 let msg = "failed to prepare thin LTO module";
798 return Err(write::llvm_err(&diag_handler, msg));
800 save_temp_bitcode(cgcx, &module, "thin-lto-after-resolve");
806 .generic_activity_with_arg("LLVM_thin_lto_internalize", thin_module.name());
807 if !llvm::LLVMRustPrepareThinLTOInternalize(thin_module.shared.data.0, llmod) {
808 let msg = "failed to prepare thin LTO module";
809 return Err(write::llvm_err(&diag_handler, msg));
811 save_temp_bitcode(cgcx, &module, "thin-lto-after-internalize");
816 cgcx.prof.generic_activity_with_arg("LLVM_thin_lto_import", thin_module.name());
817 if !llvm::LLVMRustPrepareThinLTOImport(thin_module.shared.data.0, llmod, target) {
818 let msg = "failed to prepare thin LTO module";
819 return Err(write::llvm_err(&diag_handler, msg));
821 save_temp_bitcode(cgcx, &module, "thin-lto-after-import");
824 // Ok now this is a bit unfortunate. This is also something you won't
825 // find upstream in LLVM's ThinLTO passes! This is a hack for now to
826 // work around bugs in LLVM.
828 // First discovered in #45511 it was found that as part of ThinLTO
829 // importing passes LLVM will import `DICompileUnit` metadata
830 // information across modules. This means that we'll be working with one
831 // LLVM module that has multiple `DICompileUnit` instances in it (a
832 // bunch of `llvm.dbg.cu` members). Unfortunately there's a number of
833 // bugs in LLVM's backend which generates invalid DWARF in a situation
836 // https://bugs.llvm.org/show_bug.cgi?id=35212
837 // https://bugs.llvm.org/show_bug.cgi?id=35562
839 // While the first bug there is fixed the second ended up causing #46346
840 // which was basically a resurgence of #45511 after LLVM's bug 35212 was
843 // This function below is a huge hack around this problem. The function
844 // below is defined in `PassWrapper.cpp` and will basically "merge"
845 // all `DICompileUnit` instances in a module. Basically it'll take all
846 // the objects, rewrite all pointers of `DISubprogram` to point to the
847 // first `DICompileUnit`, and then delete all the other units.
849 // This is probably mangling to the debug info slightly (but hopefully
850 // not too much) but for now at least gets LLVM to emit valid DWARF (or
851 // so it appears). Hopefully we can remove this once upstream bugs are
856 .generic_activity_with_arg("LLVM_thin_lto_patch_debuginfo", thin_module.name());
857 llvm::LLVMRustThinLTOPatchDICompileUnit(llmod, cu1);
858 save_temp_bitcode(cgcx, &module, "thin-lto-after-patch");
861 // Alright now that we've done everything related to the ThinLTO
862 // analysis it's time to run some optimizations! Here we use the same
863 // `run_pass_manager` as the "fat" LTO above except that we tell it to
864 // populate a thin-specific pass manager, which presumably LLVM treats a
865 // little differently.
867 info!("running thin lto passes over {}", module.name);
868 run_pass_manager(cgcx, &diag_handler, &mut module, true)?;
869 save_temp_bitcode(cgcx, &module, "thin-lto-after-pm");
875 /// Maps LLVM module identifiers to their corresponding LLVM LTO cache keys
876 #[derive(Debug, Default)]
877 pub struct ThinLTOKeysMap {
878 // key = llvm name of importing module, value = LLVM cache key
879 keys: FxHashMap<String, String>,
882 impl ThinLTOKeysMap {
883 fn save_to_file(&self, path: &Path) -> io::Result<()> {
885 let file = File::create(path)?;
886 let mut writer = io::BufWriter::new(file);
887 for (module, key) in &self.keys {
888 writeln!(writer, "{} {}", module, key)?;
893 fn load_from_file(path: &Path) -> io::Result<Self> {
894 use std::io::BufRead;
895 let mut keys = FxHashMap::default();
896 let file = File::open(path)?;
897 for line in io::BufReader::new(file).lines() {
899 let mut split = line.split(' ');
900 let module = split.next().unwrap();
901 let key = split.next().unwrap();
902 assert_eq!(split.next(), None, "Expected two space-separated values, found {:?}", line);
903 keys.insert(module.to_string(), key.to_string());
908 fn from_thin_lto_modules(
910 modules: &[llvm::ThinLTOModule],
913 let keys = iter::zip(modules, names)
914 .map(|(module, name)| {
915 let key = build_string(|rust_str| unsafe {
916 llvm::LLVMRustComputeLTOCacheKey(rust_str, module.identifier, data.0);
918 .expect("Invalid ThinLTO module key");
919 (name.clone().into_string().unwrap(), key)
926 fn module_name_to_str(c_str: &CStr) -> &str {
927 c_str.to_str().unwrap_or_else(|e| {
928 bug!("Encountered non-utf8 LLVM module name `{}`: {}", c_str.to_string_lossy(), e)
932 pub fn parse_module<'a>(
933 cx: &'a llvm::Context,
936 diag_handler: &Handler,
937 ) -> Result<&'a llvm::Module, FatalError> {
939 llvm::LLVMRustParseBitcodeForLTO(cx, data.as_ptr(), data.len(), name.as_ptr()).ok_or_else(
941 let msg = "failed to parse bitcode for LTO module";
942 write::llvm_err(diag_handler, msg)