1 use crate::back::write::{self, save_temp_bitcode, DiagnosticHandlers};
3 DynamicLinkingWithLTO, LlvmError, LtoBitcodeFromRlib, LtoDisallowed, LtoDylib,
5 use crate::llvm::{self, build_string};
6 use crate::{LlvmCodegenBackend, ModuleLlvm};
7 use object::read::archive::ArchiveFile;
8 use rustc_codegen_ssa::back::lto::{LtoModuleCodegen, SerializedModule, ThinModule, ThinShared};
9 use rustc_codegen_ssa::back::symbol_export;
10 use rustc_codegen_ssa::back::write::{CodegenContext, FatLTOInput, TargetMachineFactoryConfig};
11 use rustc_codegen_ssa::traits::*;
12 use rustc_codegen_ssa::{looks_like_rust_object_file, ModuleCodegen, ModuleKind};
13 use rustc_data_structures::fx::FxHashMap;
14 use rustc_data_structures::memmap::Mmap;
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::{SymbolExportInfo, SymbolExportLevel};
20 use rustc_session::cgu_reuse_tracker::CguReuse;
21 use rustc_session::config::{self, CrateType, Lto};
23 use std::ffi::{CStr, CString};
32 /// We keep track of the computed LTO cache keys from the previous
33 /// session to determine which CGUs we can reuse.
34 pub const THIN_LTO_KEYS_INCR_COMP_FILE_NAME: &str = "thin-lto-past-keys.bin";
36 pub fn crate_type_allows_lto(crate_type: CrateType) -> bool {
38 CrateType::Executable | CrateType::Dylib | CrateType::Staticlib | CrateType::Cdylib => true,
39 CrateType::Rlib | CrateType::ProcMacro => false,
44 cgcx: &CodegenContext<LlvmCodegenBackend>,
45 diag_handler: &Handler,
46 ) -> Result<(Vec<CString>, Vec<(SerializedModule<ModuleBuffer>, CString)>), FatalError> {
47 let export_threshold = match cgcx.lto {
48 // We're just doing LTO for our one crate
49 Lto::ThinLocal => SymbolExportLevel::Rust,
51 // We're doing LTO for the entire crate graph
52 Lto::Fat | Lto::Thin => symbol_export::crates_export_threshold(&cgcx.crate_types),
54 Lto::No => panic!("didn't request LTO but we're doing LTO"),
57 let symbol_filter = &|&(ref name, info): &(String, SymbolExportInfo)| {
58 if info.level.is_below_threshold(export_threshold) || info.used {
59 Some(CString::new(name.as_str()).unwrap())
64 let exported_symbols = cgcx.exported_symbols.as_ref().expect("needs exported symbols for LTO");
65 let mut symbols_below_threshold = {
66 let _timer = cgcx.prof.generic_activity("LLVM_lto_generate_symbols_below_threshold");
67 exported_symbols[&LOCAL_CRATE].iter().filter_map(symbol_filter).collect::<Vec<CString>>()
69 info!("{} symbols to preserve in this crate", symbols_below_threshold.len());
71 // If we're performing LTO for the entire crate graph, then for each of our
72 // upstream dependencies, find the corresponding rlib and load the bitcode
75 // We save off all the bytecode and LLVM module ids for later processing
76 // with either fat or thin LTO
77 let mut upstream_modules = Vec::new();
78 if cgcx.lto != Lto::ThinLocal {
79 // Make sure we actually can run LTO
80 for crate_type in cgcx.crate_types.iter() {
81 if !crate_type_allows_lto(*crate_type) {
82 diag_handler.emit_err(LtoDisallowed);
83 return Err(FatalError);
84 } else if *crate_type == CrateType::Dylib {
85 if !cgcx.opts.unstable_opts.dylib_lto {
86 diag_handler.emit_err(LtoDylib);
87 return Err(FatalError);
92 if cgcx.opts.cg.prefer_dynamic && !cgcx.opts.unstable_opts.dylib_lto {
93 diag_handler.emit_err(DynamicLinkingWithLTO);
94 return Err(FatalError);
97 for &(cnum, ref path) in cgcx.each_linked_rlib_for_lto.iter() {
98 let exported_symbols =
99 cgcx.exported_symbols.as_ref().expect("needs exported symbols for LTO");
102 cgcx.prof.generic_activity("LLVM_lto_generate_symbols_below_threshold");
103 symbols_below_threshold
104 .extend(exported_symbols[&cnum].iter().filter_map(symbol_filter));
107 let archive_data = unsafe {
108 Mmap::map(std::fs::File::open(&path).expect("couldn't open rlib"))
109 .expect("couldn't map rlib")
111 let archive = ArchiveFile::parse(&*archive_data).expect("wanted an rlib");
112 let obj_files = archive
114 .filter_map(|child| {
115 child.ok().and_then(|c| {
116 std::str::from_utf8(c.name()).ok().map(|name| (name.trim(), c))
119 .filter(|&(name, _)| looks_like_rust_object_file(name));
120 for (name, child) in obj_files {
121 info!("adding bitcode from {}", name);
122 match get_bitcode_slice_from_object_data(
123 child.data(&*archive_data).expect("corrupt rlib"),
126 let module = SerializedModule::FromRlib(data.to_vec());
127 upstream_modules.push((module, CString::new(name).unwrap()));
130 diag_handler.emit_err(e);
131 return Err(FatalError);
138 // __llvm_profile_counter_bias is pulled in at link time by an undefined reference to
139 // __llvm_profile_runtime, therefore we won't know until link time if this symbol
140 // should have default visibility.
141 symbols_below_threshold.push(CString::new("__llvm_profile_counter_bias").unwrap());
142 Ok((symbols_below_threshold, upstream_modules))
145 fn get_bitcode_slice_from_object_data(obj: &[u8]) -> Result<&[u8], LtoBitcodeFromRlib> {
148 unsafe { llvm::LLVMRustGetBitcodeSliceFromObjectData(obj.as_ptr(), obj.len(), &mut len) };
151 let bc = unsafe { slice::from_raw_parts(data, len) };
153 // `bc` must be a sub-slice of `obj`.
154 assert!(obj.as_ptr() <= bc.as_ptr());
155 assert!(bc[bc.len()..bc.len()].as_ptr() <= obj[obj.len()..obj.len()].as_ptr());
160 Err(LtoBitcodeFromRlib {
161 llvm_err: llvm::last_error().unwrap_or_else(|| "unknown LLVM error".to_string()),
166 /// Performs fat LTO by merging all modules into a single one and returning it
167 /// for further optimization.
168 pub(crate) fn run_fat(
169 cgcx: &CodegenContext<LlvmCodegenBackend>,
170 modules: Vec<FatLTOInput<LlvmCodegenBackend>>,
171 cached_modules: Vec<(SerializedModule<ModuleBuffer>, WorkProduct)>,
172 ) -> Result<LtoModuleCodegen<LlvmCodegenBackend>, FatalError> {
173 let diag_handler = cgcx.create_diag_handler();
174 let (symbols_below_threshold, upstream_modules) = prepare_lto(cgcx, &diag_handler)?;
175 let symbols_below_threshold =
176 symbols_below_threshold.iter().map(|c| c.as_ptr()).collect::<Vec<_>>();
183 &symbols_below_threshold,
187 /// Performs thin LTO by performing necessary global analysis and returning two
188 /// lists, one of the modules that need optimization and another for modules that
189 /// can simply be copied over from the incr. comp. cache.
190 pub(crate) fn run_thin(
191 cgcx: &CodegenContext<LlvmCodegenBackend>,
192 modules: Vec<(String, ThinBuffer)>,
193 cached_modules: Vec<(SerializedModule<ModuleBuffer>, WorkProduct)>,
194 ) -> Result<(Vec<LtoModuleCodegen<LlvmCodegenBackend>>, Vec<WorkProduct>), FatalError> {
195 let diag_handler = cgcx.create_diag_handler();
196 let (symbols_below_threshold, upstream_modules) = prepare_lto(cgcx, &diag_handler)?;
197 let symbols_below_threshold =
198 symbols_below_threshold.iter().map(|c| c.as_ptr()).collect::<Vec<_>>();
199 if cgcx.opts.cg.linker_plugin_lto.enabled() {
201 "We should never reach this case if the LTO step \
202 is deferred to the linker"
211 &symbols_below_threshold,
215 pub(crate) fn prepare_thin(module: ModuleCodegen<ModuleLlvm>) -> (String, ThinBuffer) {
216 let name = module.name;
217 let buffer = ThinBuffer::new(module.module_llvm.llmod(), true);
222 cgcx: &CodegenContext<LlvmCodegenBackend>,
223 diag_handler: &Handler,
224 modules: Vec<FatLTOInput<LlvmCodegenBackend>>,
225 cached_modules: Vec<(SerializedModule<ModuleBuffer>, WorkProduct)>,
226 mut serialized_modules: Vec<(SerializedModule<ModuleBuffer>, CString)>,
227 symbols_below_threshold: &[*const libc::c_char],
228 ) -> Result<LtoModuleCodegen<LlvmCodegenBackend>, FatalError> {
229 let _timer = cgcx.prof.generic_activity("LLVM_fat_lto_build_monolithic_module");
230 info!("going for a fat lto");
232 // Sort out all our lists of incoming modules into two lists.
234 // * `serialized_modules` (also and argument to this function) contains all
235 // modules that are serialized in-memory.
236 // * `in_memory` contains modules which are already parsed and in-memory,
237 // such as from multi-CGU builds.
239 // All of `cached_modules` (cached from previous incremental builds) can
240 // immediately go onto the `serialized_modules` modules list and then we can
241 // split the `modules` array into these two lists.
242 let mut in_memory = Vec::new();
243 serialized_modules.extend(cached_modules.into_iter().map(|(buffer, wp)| {
244 info!("pushing cached module {:?}", wp.cgu_name);
245 (buffer, CString::new(wp.cgu_name).unwrap())
247 for module in modules {
249 FatLTOInput::InMemory(m) => in_memory.push(m),
250 FatLTOInput::Serialized { name, buffer } => {
251 info!("pushing serialized module {:?}", name);
252 let buffer = SerializedModule::Local(buffer);
253 serialized_modules.push((buffer, CString::new(name).unwrap()));
258 // Find the "costliest" module and merge everything into that codegen unit.
259 // All the other modules will be serialized and reparsed into the new
260 // context, so this hopefully avoids serializing and parsing the largest
263 // Additionally use a regular module as the base here to ensure that various
264 // file copy operations in the backend work correctly. The only other kind
265 // of module here should be an allocator one, and if your crate is smaller
266 // than the allocator module then the size doesn't really matter anyway.
267 let costliest_module = in_memory
270 .filter(|&(_, module)| module.kind == ModuleKind::Regular)
272 let cost = unsafe { llvm::LLVMRustModuleCost(module.module_llvm.llmod()) };
277 // If we found a costliest module, we're good to go. Otherwise all our
278 // inputs were serialized which could happen in the case, for example, that
279 // all our inputs were incrementally reread from the cache and we're just
280 // re-executing the LTO passes. If that's the case deserialize the first
281 // module and create a linker with it.
282 let module: ModuleCodegen<ModuleLlvm> = match costliest_module {
283 Some((_cost, i)) => in_memory.remove(i),
285 assert!(!serialized_modules.is_empty(), "must have at least one serialized module");
286 let (buffer, name) = serialized_modules.remove(0);
287 info!("no in-memory regular modules to choose from, parsing {:?}", name);
289 module_llvm: ModuleLlvm::parse(cgcx, &name, buffer.data(), diag_handler)?,
290 name: name.into_string().unwrap(),
291 kind: ModuleKind::Regular,
295 let mut serialized_bitcode = Vec::new();
297 let (llcx, llmod) = {
298 let llvm = &module.module_llvm;
299 (&llvm.llcx, llvm.llmod())
301 info!("using {:?} as a base module", module.name);
303 // The linking steps below may produce errors and diagnostics within LLVM
304 // which we'd like to handle and print, so set up our diagnostic handlers
305 // (which get unregistered when they go out of scope below).
306 let _handler = DiagnosticHandlers::new(cgcx, diag_handler, llcx);
308 // For all other modules we codegened we'll need to link them into our own
309 // bitcode. All modules were codegened in their own LLVM context, however,
310 // and we want to move everything to the same LLVM context. Currently the
311 // way we know of to do that is to serialize them to a string and them parse
312 // them later. Not great but hey, that's why it's "fat" LTO, right?
313 for module in in_memory {
314 let buffer = ModuleBuffer::new(module.module_llvm.llmod());
315 let llmod_id = CString::new(&module.name[..]).unwrap();
316 serialized_modules.push((SerializedModule::Local(buffer), llmod_id));
318 // Sort the modules to ensure we produce deterministic results.
319 serialized_modules.sort_by(|module1, module2| module1.1.cmp(&module2.1));
321 // For all serialized bitcode files we parse them and link them in as we did
322 // above, this is all mostly handled in C++. Like above, though, we don't
323 // know much about the memory management here so we err on the side of being
324 // save and persist everything with the original module.
325 let mut linker = Linker::new(llmod);
326 for (bc_decoded, name) in serialized_modules {
329 .generic_activity_with_arg_recorder("LLVM_fat_lto_link_module", |recorder| {
330 recorder.record_arg(format!("{:?}", name))
332 info!("linking {:?}", name);
333 let data = bc_decoded.data();
336 .map_err(|()| write::llvm_err(diag_handler, LlvmError::LoadBitcode { name }))?;
337 serialized_bitcode.push(bc_decoded);
340 save_temp_bitcode(cgcx, &module, "lto.input");
342 // Internalize everything below threshold to help strip out more modules and such.
344 let ptr = symbols_below_threshold.as_ptr();
345 llvm::LLVMRustRunRestrictionPass(
347 ptr as *const *const libc::c_char,
348 symbols_below_threshold.len() as libc::size_t,
350 save_temp_bitcode(cgcx, &module, "lto.after-restriction");
354 Ok(LtoModuleCodegen::Fat { module, _serialized_bitcode: serialized_bitcode })
357 pub(crate) struct Linker<'a>(&'a mut llvm::Linker<'a>);
359 impl<'a> Linker<'a> {
360 pub(crate) fn new(llmod: &'a llvm::Module) -> Self {
361 unsafe { Linker(llvm::LLVMRustLinkerNew(llmod)) }
364 pub(crate) fn add(&mut self, bytecode: &[u8]) -> Result<(), ()> {
366 if llvm::LLVMRustLinkerAdd(
368 bytecode.as_ptr() as *const libc::c_char,
379 impl Drop for Linker<'_> {
382 llvm::LLVMRustLinkerFree(&mut *(self.0 as *mut _));
387 /// Prepare "thin" LTO to get run on these modules.
389 /// The general structure of ThinLTO is quite different from the structure of
390 /// "fat" LTO above. With "fat" LTO all LLVM modules in question are merged into
391 /// one giant LLVM module, and then we run more optimization passes over this
392 /// big module after internalizing most symbols. Thin LTO, on the other hand,
393 /// avoid this large bottleneck through more targeted optimization.
395 /// At a high level Thin LTO looks like:
397 /// 1. Prepare a "summary" of each LLVM module in question which describes
398 /// the values inside, cost of the values, etc.
399 /// 2. Merge the summaries of all modules in question into one "index"
400 /// 3. Perform some global analysis on this index
401 /// 4. For each module, use the index and analysis calculated previously to
402 /// perform local transformations on the module, for example inlining
403 /// small functions from other modules.
404 /// 5. Run thin-specific optimization passes over each module, and then code
405 /// generate everything at the end.
407 /// The summary for each module is intended to be quite cheap, and the global
408 /// index is relatively quite cheap to create as well. As a result, the goal of
409 /// ThinLTO is to reduce the bottleneck on LTO and enable LTO to be used in more
410 /// situations. For example one cheap optimization is that we can parallelize
411 /// all codegen modules, easily making use of all the cores on a machine.
413 /// With all that in mind, the function here is designed at specifically just
414 /// calculating the *index* for ThinLTO. This index will then be shared amongst
415 /// all of the `LtoModuleCodegen` units returned below and destroyed once
416 /// they all go out of scope.
418 cgcx: &CodegenContext<LlvmCodegenBackend>,
419 diag_handler: &Handler,
420 modules: Vec<(String, ThinBuffer)>,
421 serialized_modules: Vec<(SerializedModule<ModuleBuffer>, CString)>,
422 cached_modules: Vec<(SerializedModule<ModuleBuffer>, WorkProduct)>,
423 symbols_below_threshold: &[*const libc::c_char],
424 ) -> Result<(Vec<LtoModuleCodegen<LlvmCodegenBackend>>, Vec<WorkProduct>), FatalError> {
425 let _timer = cgcx.prof.generic_activity("LLVM_thin_lto_global_analysis");
427 info!("going for that thin, thin LTO");
429 let green_modules: FxHashMap<_, _> =
430 cached_modules.iter().map(|(_, wp)| (wp.cgu_name.clone(), wp.clone())).collect();
432 let full_scope_len = modules.len() + serialized_modules.len() + cached_modules.len();
433 let mut thin_buffers = Vec::with_capacity(modules.len());
434 let mut module_names = Vec::with_capacity(full_scope_len);
435 let mut thin_modules = Vec::with_capacity(full_scope_len);
437 for (i, (name, buffer)) in modules.into_iter().enumerate() {
438 info!("local module: {} - {}", i, name);
439 let cname = CString::new(name.clone()).unwrap();
440 thin_modules.push(llvm::ThinLTOModule {
441 identifier: cname.as_ptr(),
442 data: buffer.data().as_ptr(),
443 len: buffer.data().len(),
445 thin_buffers.push(buffer);
446 module_names.push(cname);
449 // FIXME: All upstream crates are deserialized internally in the
450 // function below to extract their summary and modules. Note that
451 // unlike the loop above we *must* decode and/or read something
452 // here as these are all just serialized files on disk. An
453 // improvement, however, to make here would be to store the
454 // module summary separately from the actual module itself. Right
455 // now this is store in one large bitcode file, and the entire
456 // file is deflate-compressed. We could try to bypass some of the
457 // decompression by storing the index uncompressed and only
458 // lazily decompressing the bytecode if necessary.
460 // Note that truly taking advantage of this optimization will
461 // likely be further down the road. We'd have to implement
462 // incremental ThinLTO first where we could actually avoid
463 // looking at upstream modules entirely sometimes (the contents,
464 // we must always unconditionally look at the index).
465 let mut serialized = Vec::with_capacity(serialized_modules.len() + cached_modules.len());
468 cached_modules.into_iter().map(|(sm, wp)| (sm, CString::new(wp.cgu_name).unwrap()));
470 for (module, name) in serialized_modules.into_iter().chain(cached_modules) {
471 info!("upstream or cached module {:?}", name);
472 thin_modules.push(llvm::ThinLTOModule {
473 identifier: name.as_ptr(),
474 data: module.data().as_ptr(),
475 len: module.data().len(),
477 serialized.push(module);
478 module_names.push(name);
482 assert_eq!(thin_modules.len(), module_names.len());
484 // Delegate to the C++ bindings to create some data here. Once this is a
485 // tried-and-true interface we may wish to try to upstream some of this
486 // to LLVM itself, right now we reimplement a lot of what they do
488 let data = llvm::LLVMRustCreateThinLTOData(
489 thin_modules.as_ptr(),
490 thin_modules.len() as u32,
491 symbols_below_threshold.as_ptr(),
492 symbols_below_threshold.len() as u32,
494 .ok_or_else(|| write::llvm_err(diag_handler, LlvmError::PrepareThinLtoContext))?;
496 let data = ThinData(data);
498 info!("thin LTO data created");
500 let (key_map_path, prev_key_map, curr_key_map) = if let Some(ref incr_comp_session_dir) =
501 cgcx.incr_comp_session_dir
503 let path = incr_comp_session_dir.join(THIN_LTO_KEYS_INCR_COMP_FILE_NAME);
504 // If the previous file was deleted, or we get an IO error
505 // reading the file, then we'll just use `None` as the
506 // prev_key_map, which will force the code to be recompiled.
508 if path.exists() { ThinLTOKeysMap::load_from_file(&path).ok() } else { None };
509 let curr = ThinLTOKeysMap::from_thin_lto_modules(&data, &thin_modules, &module_names);
510 (Some(path), prev, curr)
512 // If we don't compile incrementally, we don't need to load the
513 // import data from LLVM.
514 assert!(green_modules.is_empty());
515 let curr = ThinLTOKeysMap::default();
518 info!("thin LTO cache key map loaded");
519 info!("prev_key_map: {:#?}", prev_key_map);
520 info!("curr_key_map: {:#?}", curr_key_map);
522 // Throw our data in an `Arc` as we'll be sharing it across threads. We
523 // also put all memory referenced by the C++ data (buffers, ids, etc)
524 // into the arc as well. After this we'll create a thin module
525 // codegen per module in this data.
526 let shared = Arc::new(ThinShared {
529 serialized_modules: serialized,
533 let mut copy_jobs = vec![];
534 let mut opt_jobs = vec![];
536 info!("checking which modules can be-reused and which have to be re-optimized.");
537 for (module_index, module_name) in shared.module_names.iter().enumerate() {
538 let module_name = module_name_to_str(module_name);
539 if let (Some(prev_key_map), true) =
540 (prev_key_map.as_ref(), green_modules.contains_key(module_name))
542 assert!(cgcx.incr_comp_session_dir.is_some());
544 // If a module exists in both the current and the previous session,
545 // and has the same LTO cache key in both sessions, then we can re-use it
546 if prev_key_map.keys.get(module_name) == curr_key_map.keys.get(module_name) {
547 let work_product = green_modules[module_name].clone();
548 copy_jobs.push(work_product);
549 info!(" - {}: re-used", module_name);
550 assert!(cgcx.incr_comp_session_dir.is_some());
551 cgcx.cgu_reuse_tracker.set_actual_reuse(module_name, CguReuse::PostLto);
556 info!(" - {}: re-compiled", module_name);
557 opt_jobs.push(LtoModuleCodegen::Thin(ThinModule {
558 shared: shared.clone(),
563 // Save the current ThinLTO import information for the next compilation
564 // session, overwriting the previous serialized data (if any).
565 if let Some(path) = key_map_path {
566 if let Err(err) = curr_key_map.save_to_file(&path) {
567 return Err(write::llvm_err(diag_handler, LlvmError::WriteThinLtoKey { err }));
571 Ok((opt_jobs, copy_jobs))
575 pub(crate) fn run_pass_manager(
576 cgcx: &CodegenContext<LlvmCodegenBackend>,
577 diag_handler: &Handler,
578 module: &mut ModuleCodegen<ModuleLlvm>,
580 ) -> Result<(), FatalError> {
581 let _timer = cgcx.prof.verbose_generic_activity_with_arg("LLVM_lto_optimize", &*module.name);
582 let config = cgcx.config(module.kind);
584 // Now we have one massive module inside of llmod. Time to run the
585 // LTO-specific optimization passes that LLVM provides.
587 // This code is based off the code found in llvm's LTO code generator:
588 // llvm/lib/LTO/LTOCodeGenerator.cpp
589 debug!("running the pass manager");
591 if !llvm::LLVMRustHasModuleFlag(
592 module.module_llvm.llmod(),
593 "LTOPostLink".as_ptr().cast(),
596 llvm::LLVMRustAddModuleFlag(
597 module.module_llvm.llmod(),
598 llvm::LLVMModFlagBehavior::Error,
599 "LTOPostLink\0".as_ptr().cast(),
603 let opt_stage = if thin { llvm::OptStage::ThinLTO } else { llvm::OptStage::FatLTO };
604 let opt_level = config.opt_level.unwrap_or(config::OptLevel::No);
605 write::llvm_optimize(cgcx, diag_handler, module, config, opt_level, opt_stage)?;
611 pub struct ModuleBuffer(&'static mut llvm::ModuleBuffer);
613 unsafe impl Send for ModuleBuffer {}
614 unsafe impl Sync for ModuleBuffer {}
617 pub fn new(m: &llvm::Module) -> ModuleBuffer {
618 ModuleBuffer(unsafe { llvm::LLVMRustModuleBufferCreate(m) })
622 impl ModuleBufferMethods for ModuleBuffer {
623 fn data(&self) -> &[u8] {
625 let ptr = llvm::LLVMRustModuleBufferPtr(self.0);
626 let len = llvm::LLVMRustModuleBufferLen(self.0);
627 slice::from_raw_parts(ptr, len)
632 impl Drop for ModuleBuffer {
635 llvm::LLVMRustModuleBufferFree(&mut *(self.0 as *mut _));
640 pub struct ThinData(&'static mut llvm::ThinLTOData);
642 unsafe impl Send for ThinData {}
643 unsafe impl Sync for ThinData {}
645 impl Drop for ThinData {
648 llvm::LLVMRustFreeThinLTOData(&mut *(self.0 as *mut _));
653 pub struct ThinBuffer(&'static mut llvm::ThinLTOBuffer);
655 unsafe impl Send for ThinBuffer {}
656 unsafe impl Sync for ThinBuffer {}
659 pub fn new(m: &llvm::Module, is_thin: bool) -> ThinBuffer {
661 let buffer = llvm::LLVMRustThinLTOBufferCreate(m, is_thin);
667 impl ThinBufferMethods for ThinBuffer {
668 fn data(&self) -> &[u8] {
670 let ptr = llvm::LLVMRustThinLTOBufferPtr(self.0) as *const _;
671 let len = llvm::LLVMRustThinLTOBufferLen(self.0);
672 slice::from_raw_parts(ptr, len)
677 impl Drop for ThinBuffer {
680 llvm::LLVMRustThinLTOBufferFree(&mut *(self.0 as *mut _));
685 pub unsafe fn optimize_thin_module(
686 thin_module: ThinModule<LlvmCodegenBackend>,
687 cgcx: &CodegenContext<LlvmCodegenBackend>,
688 ) -> Result<ModuleCodegen<ModuleLlvm>, FatalError> {
689 let diag_handler = cgcx.create_diag_handler();
691 let module_name = &thin_module.shared.module_names[thin_module.idx];
692 let tm_factory_config = TargetMachineFactoryConfig::new(cgcx, module_name.to_str().unwrap());
693 let tm = (cgcx.tm_factory)(tm_factory_config).map_err(|e| write::llvm_err(&diag_handler, e))?;
695 // Right now the implementation we've got only works over serialized
696 // modules, so we create a fresh new LLVM context and parse the module
697 // into that context. One day, however, we may do this for upstream
698 // crates but for locally codegened modules we may be able to reuse
699 // that LLVM Context and Module.
700 let llcx = llvm::LLVMRustContextCreate(cgcx.fewer_names);
701 let llmod_raw = parse_module(llcx, module_name, thin_module.data(), &diag_handler)? as *const _;
702 let mut module = ModuleCodegen {
703 module_llvm: ModuleLlvm { llmod_raw, llcx, tm },
704 name: thin_module.name().to_string(),
705 kind: ModuleKind::Regular,
708 let target = &*module.module_llvm.tm;
709 let llmod = module.module_llvm.llmod();
710 save_temp_bitcode(cgcx, &module, "thin-lto-input");
712 // Before we do much else find the "main" `DICompileUnit` that we'll be
713 // using below. If we find more than one though then rustc has changed
714 // in a way we're not ready for, so generate an ICE by returning
716 let mut cu1 = ptr::null_mut();
717 let mut cu2 = ptr::null_mut();
718 llvm::LLVMRustThinLTOGetDICompileUnit(llmod, &mut cu1, &mut cu2);
720 return Err(write::llvm_err(&diag_handler, LlvmError::MultipleSourceDiCompileUnit));
723 // Up next comes the per-module local analyses that we do for Thin LTO.
724 // Each of these functions is basically copied from the LLVM
725 // implementation and then tailored to suit this implementation. Ideally
726 // each of these would be supported by upstream LLVM but that's perhaps
727 // a patch for another day!
729 // You can find some more comments about these functions in the LLVM
730 // bindings we've got (currently `PassWrapper.cpp`)
733 cgcx.prof.generic_activity_with_arg("LLVM_thin_lto_rename", thin_module.name());
734 if !llvm::LLVMRustPrepareThinLTORename(thin_module.shared.data.0, llmod, target) {
735 return Err(write::llvm_err(&diag_handler, LlvmError::PrepareThinLtoModule));
737 save_temp_bitcode(cgcx, &module, "thin-lto-after-rename");
743 .generic_activity_with_arg("LLVM_thin_lto_resolve_weak", thin_module.name());
744 if !llvm::LLVMRustPrepareThinLTOResolveWeak(thin_module.shared.data.0, llmod) {
745 return Err(write::llvm_err(&diag_handler, LlvmError::PrepareThinLtoModule));
747 save_temp_bitcode(cgcx, &module, "thin-lto-after-resolve");
753 .generic_activity_with_arg("LLVM_thin_lto_internalize", thin_module.name());
754 if !llvm::LLVMRustPrepareThinLTOInternalize(thin_module.shared.data.0, llmod) {
755 return Err(write::llvm_err(&diag_handler, LlvmError::PrepareThinLtoModule));
757 save_temp_bitcode(cgcx, &module, "thin-lto-after-internalize");
762 cgcx.prof.generic_activity_with_arg("LLVM_thin_lto_import", thin_module.name());
763 if !llvm::LLVMRustPrepareThinLTOImport(thin_module.shared.data.0, llmod, target) {
764 return Err(write::llvm_err(&diag_handler, LlvmError::PrepareThinLtoModule));
766 save_temp_bitcode(cgcx, &module, "thin-lto-after-import");
769 // Ok now this is a bit unfortunate. This is also something you won't
770 // find upstream in LLVM's ThinLTO passes! This is a hack for now to
771 // work around bugs in LLVM.
773 // First discovered in #45511 it was found that as part of ThinLTO
774 // importing passes LLVM will import `DICompileUnit` metadata
775 // information across modules. This means that we'll be working with one
776 // LLVM module that has multiple `DICompileUnit` instances in it (a
777 // bunch of `llvm.dbg.cu` members). Unfortunately there's a number of
778 // bugs in LLVM's backend which generates invalid DWARF in a situation
781 // https://bugs.llvm.org/show_bug.cgi?id=35212
782 // https://bugs.llvm.org/show_bug.cgi?id=35562
784 // While the first bug there is fixed the second ended up causing #46346
785 // which was basically a resurgence of #45511 after LLVM's bug 35212 was
788 // This function below is a huge hack around this problem. The function
789 // below is defined in `PassWrapper.cpp` and will basically "merge"
790 // all `DICompileUnit` instances in a module. Basically it'll take all
791 // the objects, rewrite all pointers of `DISubprogram` to point to the
792 // first `DICompileUnit`, and then delete all the other units.
794 // This is probably mangling to the debug info slightly (but hopefully
795 // not too much) but for now at least gets LLVM to emit valid DWARF (or
796 // so it appears). Hopefully we can remove this once upstream bugs are
801 .generic_activity_with_arg("LLVM_thin_lto_patch_debuginfo", thin_module.name());
802 llvm::LLVMRustThinLTOPatchDICompileUnit(llmod, cu1);
803 save_temp_bitcode(cgcx, &module, "thin-lto-after-patch");
806 // Alright now that we've done everything related to the ThinLTO
807 // analysis it's time to run some optimizations! Here we use the same
808 // `run_pass_manager` as the "fat" LTO above except that we tell it to
809 // populate a thin-specific pass manager, which presumably LLVM treats a
810 // little differently.
812 info!("running thin lto passes over {}", module.name);
813 run_pass_manager(cgcx, &diag_handler, &mut module, true)?;
814 save_temp_bitcode(cgcx, &module, "thin-lto-after-pm");
820 /// Maps LLVM module identifiers to their corresponding LLVM LTO cache keys
821 #[derive(Debug, Default)]
822 pub struct ThinLTOKeysMap {
823 // key = llvm name of importing module, value = LLVM cache key
824 keys: FxHashMap<String, String>,
827 impl ThinLTOKeysMap {
828 fn save_to_file(&self, path: &Path) -> io::Result<()> {
830 let file = File::create(path)?;
831 let mut writer = io::BufWriter::new(file);
832 for (module, key) in &self.keys {
833 writeln!(writer, "{} {}", module, key)?;
838 fn load_from_file(path: &Path) -> io::Result<Self> {
839 use std::io::BufRead;
840 let mut keys = FxHashMap::default();
841 let file = File::open(path)?;
842 for line in io::BufReader::new(file).lines() {
844 let mut split = line.split(' ');
845 let module = split.next().unwrap();
846 let key = split.next().unwrap();
847 assert_eq!(split.next(), None, "Expected two space-separated values, found {:?}", line);
848 keys.insert(module.to_string(), key.to_string());
853 fn from_thin_lto_modules(
855 modules: &[llvm::ThinLTOModule],
858 let keys = iter::zip(modules, names)
859 .map(|(module, name)| {
860 let key = build_string(|rust_str| unsafe {
861 llvm::LLVMRustComputeLTOCacheKey(rust_str, module.identifier, data.0);
863 .expect("Invalid ThinLTO module key");
864 (name.clone().into_string().unwrap(), key)
871 fn module_name_to_str(c_str: &CStr) -> &str {
872 c_str.to_str().unwrap_or_else(|e| {
873 bug!("Encountered non-utf8 LLVM module name `{}`: {}", c_str.to_string_lossy(), e)
877 pub fn parse_module<'a>(
878 cx: &'a llvm::Context,
881 diag_handler: &Handler,
882 ) -> Result<&'a llvm::Module, FatalError> {
884 llvm::LLVMRustParseBitcodeForLTO(cx, data.as_ptr(), data.len(), name.as_ptr())
885 .ok_or_else(|| write::llvm_err(diag_handler, LlvmError::ParseBitcode))