1 // Copyright 2013 The Rust Project Developers. See the COPYRIGHT
2 // file at the top-level directory of this distribution and at
3 // http://rust-lang.org/COPYRIGHT.
5 // Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
6 // http://www.apache.org/licenses/LICENSE-2.0> or the MIT license
7 // <LICENSE-MIT or http://opensource.org/licenses/MIT>, at your
8 // option. This file may not be copied, modified, or distributed
9 // except according to those terms.
11 use back::bytecode::{DecodedBytecode, RLIB_BYTECODE_EXTENSION};
12 use back::write::{ModuleConfig, with_llvm_pmb, CodegenContext};
13 use back::write::{self, DiagnosticHandlers, pre_lto_bitcode_filename};
14 use errors::{FatalError, Handler};
15 use llvm::archive_ro::ArchiveRO;
16 use llvm::{self, True, False};
18 use rustc::dep_graph::WorkProduct;
19 use rustc::dep_graph::cgu_reuse_tracker::CguReuse;
20 use rustc::hir::def_id::LOCAL_CRATE;
21 use rustc::middle::exported_symbols::SymbolExportLevel;
22 use rustc::session::config::{self, Lto};
23 use rustc::util::common::time_ext;
24 use rustc_data_structures::fx::FxHashMap;
25 use rustc_codegen_utils::symbol_export;
26 use time_graph::Timeline;
27 use {ModuleCodegen, ModuleLlvm, ModuleKind};
31 use std::ffi::{CStr, CString};
37 pub fn crate_type_allows_lto(crate_type: config::CrateType) -> bool {
39 config::CrateType::Executable |
40 config::CrateType::Staticlib |
41 config::CrateType::Cdylib => true,
43 config::CrateType::Dylib |
44 config::CrateType::Rlib |
45 config::CrateType::ProcMacro => false,
49 pub(crate) enum LtoModuleCodegen {
51 module: Option<ModuleCodegen<ModuleLlvm>>,
52 _serialized_bitcode: Vec<SerializedModule>,
58 impl LtoModuleCodegen {
59 pub fn name(&self) -> &str {
61 LtoModuleCodegen::Fat { .. } => "everything",
62 LtoModuleCodegen::Thin(ref m) => m.name(),
66 /// Optimize this module within the given codegen context.
68 /// This function is unsafe as it'll return a `ModuleCodegen` still
69 /// points to LLVM data structures owned by this `LtoModuleCodegen`.
70 /// It's intended that the module returned is immediately code generated and
71 /// dropped, and then this LTO module is dropped.
72 pub(crate) unsafe fn optimize(&mut self,
73 cgcx: &CodegenContext,
74 timeline: &mut Timeline)
75 -> Result<ModuleCodegen<ModuleLlvm>, FatalError>
78 LtoModuleCodegen::Fat { ref mut module, .. } => {
79 let module = module.take().unwrap();
81 let config = cgcx.config(module.kind);
82 let llmod = module.module_llvm.llmod();
83 let tm = &*module.module_llvm.tm;
84 run_pass_manager(cgcx, tm, llmod, config, false);
85 timeline.record("fat-done");
89 LtoModuleCodegen::Thin(ref mut thin) => thin.optimize(cgcx, timeline),
93 /// A "gauge" of how costly it is to optimize this module, used to sort
94 /// biggest modules first.
95 pub fn cost(&self) -> u64 {
97 // Only one module with fat LTO, so the cost doesn't matter.
98 LtoModuleCodegen::Fat { .. } => 0,
99 LtoModuleCodegen::Thin(ref m) => m.cost(),
104 /// Performs LTO, which in the case of full LTO means merging all modules into
105 /// a single one and returning it for further optimizing. For ThinLTO, it will
106 /// do the global analysis necessary and return two lists, one of the modules
107 /// the need optimization and another for modules that can simply be copied over
108 /// from the incr. comp. cache.
109 pub(crate) fn run(cgcx: &CodegenContext,
110 modules: Vec<ModuleCodegen<ModuleLlvm>>,
111 cached_modules: Vec<(SerializedModule, WorkProduct)>,
112 timeline: &mut Timeline)
113 -> Result<(Vec<LtoModuleCodegen>, Vec<WorkProduct>), FatalError>
115 let diag_handler = cgcx.create_diag_handler();
116 let export_threshold = match cgcx.lto {
117 // We're just doing LTO for our one crate
118 Lto::ThinLocal => SymbolExportLevel::Rust,
120 // We're doing LTO for the entire crate graph
121 Lto::Fat | Lto::Thin => {
122 symbol_export::crates_export_threshold(&cgcx.crate_types)
125 Lto::No => panic!("didn't request LTO but we're doing LTO"),
128 let symbol_filter = &|&(ref name, level): &(String, SymbolExportLevel)| {
129 if level.is_below_threshold(export_threshold) {
130 let mut bytes = Vec::with_capacity(name.len() + 1);
131 bytes.extend(name.bytes());
132 Some(CString::new(bytes).unwrap())
137 let exported_symbols = cgcx.exported_symbols
138 .as_ref().expect("needs exported symbols for LTO");
139 let mut symbol_white_list = exported_symbols[&LOCAL_CRATE]
141 .filter_map(symbol_filter)
142 .collect::<Vec<CString>>();
143 timeline.record("whitelist");
144 info!("{} symbols to preserve in this crate", symbol_white_list.len());
146 // If we're performing LTO for the entire crate graph, then for each of our
147 // upstream dependencies, find the corresponding rlib and load the bitcode
150 // We save off all the bytecode and LLVM module ids for later processing
151 // with either fat or thin LTO
152 let mut upstream_modules = Vec::new();
153 if cgcx.lto != Lto::ThinLocal {
154 if cgcx.opts.cg.prefer_dynamic {
155 diag_handler.struct_err("cannot prefer dynamic linking when performing LTO")
156 .note("only 'staticlib', 'bin', and 'cdylib' outputs are \
159 return Err(FatalError)
162 // Make sure we actually can run LTO
163 for crate_type in cgcx.crate_types.iter() {
164 if !crate_type_allows_lto(*crate_type) {
165 let e = diag_handler.fatal("lto can only be run for executables, cdylibs and \
166 static library outputs");
171 for &(cnum, ref path) in cgcx.each_linked_rlib_for_lto.iter() {
172 let exported_symbols = cgcx.exported_symbols
173 .as_ref().expect("needs exported symbols for LTO");
174 symbol_white_list.extend(
175 exported_symbols[&cnum]
177 .filter_map(symbol_filter));
179 let archive = ArchiveRO::open(&path).expect("wanted an rlib");
180 let bytecodes = archive.iter().filter_map(|child| {
181 child.ok().and_then(|c| c.name().map(|name| (name, c)))
182 }).filter(|&(name, _)| name.ends_with(RLIB_BYTECODE_EXTENSION));
183 for (name, data) in bytecodes {
184 info!("adding bytecode {}", name);
185 let bc_encoded = data.data();
187 let (bc, id) = time_ext(cgcx.time_passes, None, &format!("decode {}", name), || {
188 match DecodedBytecode::new(bc_encoded) {
189 Ok(b) => Ok((b.bytecode(), b.identifier().to_string())),
190 Err(e) => Err(diag_handler.fatal(&e)),
193 let bc = SerializedModule::FromRlib(bc);
194 upstream_modules.push((bc, CString::new(id).unwrap()));
196 timeline.record(&format!("load: {}", path.display()));
200 let symbol_white_list = symbol_white_list.iter()
202 .collect::<Vec<_>>();
205 assert!(cached_modules.is_empty());
206 let opt_jobs = fat_lto(cgcx,
212 opt_jobs.map(|opt_jobs| (opt_jobs, vec![]))
216 if cgcx.opts.debugging_opts.cross_lang_lto.enabled() {
217 unreachable!("We should never reach this case if the LTO step \
218 is deferred to the linker");
228 Lto::No => unreachable!(),
232 fn fat_lto(cgcx: &CodegenContext,
233 diag_handler: &Handler,
234 mut modules: Vec<ModuleCodegen<ModuleLlvm>>,
235 mut serialized_modules: Vec<(SerializedModule, CString)>,
236 symbol_white_list: &[*const libc::c_char],
237 timeline: &mut Timeline)
238 -> Result<Vec<LtoModuleCodegen>, FatalError>
240 info!("going for a fat lto");
242 // Find the "costliest" module and merge everything into that codegen unit.
243 // All the other modules will be serialized and reparsed into the new
244 // context, so this hopefully avoids serializing and parsing the largest
247 // Additionally use a regular module as the base here to ensure that various
248 // file copy operations in the backend work correctly. The only other kind
249 // of module here should be an allocator one, and if your crate is smaller
250 // than the allocator module then the size doesn't really matter anyway.
251 let (_, costliest_module) = modules.iter()
253 .filter(|&(_, module)| module.kind == ModuleKind::Regular)
256 llvm::LLVMRustModuleCost(module.module_llvm.llmod())
261 .expect("must be codegen'ing at least one module");
262 let module = modules.remove(costliest_module);
263 let mut serialized_bitcode = Vec::new();
265 let (llcx, llmod) = {
266 let llvm = &module.module_llvm;
267 (&llvm.llcx, llvm.llmod())
269 info!("using {:?} as a base module", module.name);
271 // The linking steps below may produce errors and diagnostics within LLVM
272 // which we'd like to handle and print, so set up our diagnostic handlers
273 // (which get unregistered when they go out of scope below).
274 let _handler = DiagnosticHandlers::new(cgcx, diag_handler, llcx);
276 // For all other modules we codegened we'll need to link them into our own
277 // bitcode. All modules were codegened in their own LLVM context, however,
278 // and we want to move everything to the same LLVM context. Currently the
279 // way we know of to do that is to serialize them to a string and them parse
280 // them later. Not great but hey, that's why it's "fat" LTO, right?
281 for module in modules {
282 let buffer = ModuleBuffer::new(module.module_llvm.llmod());
283 let llmod_id = CString::new(&module.name[..]).unwrap();
284 serialized_modules.push((SerializedModule::Local(buffer), llmod_id));
287 // For all serialized bitcode files we parse them and link them in as we did
288 // above, this is all mostly handled in C++. Like above, though, we don't
289 // know much about the memory management here so we err on the side of being
290 // save and persist everything with the original module.
291 let mut linker = Linker::new(llmod);
292 for (bc_decoded, name) in serialized_modules {
293 info!("linking {:?}", name);
294 time_ext(cgcx.time_passes, None, &format!("ll link {:?}", name), || {
295 let data = bc_decoded.data();
296 linker.add(&data).map_err(|()| {
297 let msg = format!("failed to load bc of {:?}", name);
298 write::llvm_err(&diag_handler, &msg)
301 timeline.record(&format!("link {:?}", name));
302 serialized_bitcode.push(bc_decoded);
305 cgcx.save_temp_bitcode(&module, "lto.input");
307 // Internalize everything that *isn't* in our whitelist to help strip out
308 // more modules and such
310 let ptr = symbol_white_list.as_ptr();
311 llvm::LLVMRustRunRestrictionPass(llmod,
312 ptr as *const *const libc::c_char,
313 symbol_white_list.len() as libc::size_t);
314 cgcx.save_temp_bitcode(&module, "lto.after-restriction");
317 if cgcx.no_landing_pads {
319 llvm::LLVMRustMarkAllFunctionsNounwind(llmod);
321 cgcx.save_temp_bitcode(&module, "lto.after-nounwind");
323 timeline.record("passes");
326 Ok(vec![LtoModuleCodegen::Fat {
327 module: Some(module),
328 _serialized_bitcode: serialized_bitcode,
332 struct Linker<'a>(&'a mut llvm::Linker<'a>);
335 fn new(llmod: &'a llvm::Module) -> Self {
336 unsafe { Linker(llvm::LLVMRustLinkerNew(llmod)) }
339 fn add(&mut self, bytecode: &[u8]) -> Result<(), ()> {
341 if llvm::LLVMRustLinkerAdd(self.0,
342 bytecode.as_ptr() as *const libc::c_char,
352 impl Drop for Linker<'a> {
354 unsafe { llvm::LLVMRustLinkerFree(&mut *(self.0 as *mut _)); }
358 /// Prepare "thin" LTO to get run on these modules.
360 /// The general structure of ThinLTO is quite different from the structure of
361 /// "fat" LTO above. With "fat" LTO all LLVM modules in question are merged into
362 /// one giant LLVM module, and then we run more optimization passes over this
363 /// big module after internalizing most symbols. Thin LTO, on the other hand,
364 /// avoid this large bottleneck through more targeted optimization.
366 /// At a high level Thin LTO looks like:
368 /// 1. Prepare a "summary" of each LLVM module in question which describes
369 /// the values inside, cost of the values, etc.
370 /// 2. Merge the summaries of all modules in question into one "index"
371 /// 3. Perform some global analysis on this index
372 /// 4. For each module, use the index and analysis calculated previously to
373 /// perform local transformations on the module, for example inlining
374 /// small functions from other modules.
375 /// 5. Run thin-specific optimization passes over each module, and then code
376 /// generate everything at the end.
378 /// The summary for each module is intended to be quite cheap, and the global
379 /// index is relatively quite cheap to create as well. As a result, the goal of
380 /// ThinLTO is to reduce the bottleneck on LTO and enable LTO to be used in more
381 /// situations. For example one cheap optimization is that we can parallelize
382 /// all codegen modules, easily making use of all the cores on a machine.
384 /// With all that in mind, the function here is designed at specifically just
385 /// calculating the *index* for ThinLTO. This index will then be shared amongst
386 /// all of the `LtoModuleCodegen` units returned below and destroyed once
387 /// they all go out of scope.
388 fn thin_lto(cgcx: &CodegenContext,
389 diag_handler: &Handler,
390 modules: Vec<ModuleCodegen<ModuleLlvm>>,
391 serialized_modules: Vec<(SerializedModule, CString)>,
392 cached_modules: Vec<(SerializedModule, WorkProduct)>,
393 symbol_white_list: &[*const libc::c_char],
394 timeline: &mut Timeline)
395 -> Result<(Vec<LtoModuleCodegen>, Vec<WorkProduct>), FatalError>
398 info!("going for that thin, thin LTO");
400 let green_modules: FxHashMap<_, _> = cached_modules
402 .map(|&(_, ref wp)| (wp.cgu_name.clone(), wp.clone()))
405 let mut thin_buffers = Vec::new();
406 let mut module_names = Vec::new();
407 let mut thin_modules = Vec::new();
409 // FIXME: right now, like with fat LTO, we serialize all in-memory
410 // modules before working with them and ThinLTO. We really
411 // shouldn't do this, however, and instead figure out how to
412 // extract a summary from an in-memory module and then merge that
413 // into the global index. It turns out that this loop is by far
414 // the most expensive portion of this small bit of global
416 for (i, module) in modules.iter().enumerate() {
417 info!("local module: {} - {}", i, module.name);
418 let name = CString::new(module.name.clone()).unwrap();
419 let buffer = ThinBuffer::new(module.module_llvm.llmod());
421 // We emit the module after having serialized it into a ThinBuffer
422 // because only then it will contain the ThinLTO module summary.
423 if let Some(ref incr_comp_session_dir) = cgcx.incr_comp_session_dir {
424 if cgcx.config(module.kind).emit_pre_thin_lto_bc {
425 let path = incr_comp_session_dir
426 .join(pre_lto_bitcode_filename(&module.name));
428 fs::write(&path, buffer.data()).unwrap_or_else(|e| {
429 panic!("Error writing pre-lto-bitcode file `{}`: {}",
436 thin_modules.push(llvm::ThinLTOModule {
437 identifier: name.as_ptr(),
438 data: buffer.data().as_ptr(),
439 len: buffer.data().len(),
441 thin_buffers.push(buffer);
442 module_names.push(name);
443 timeline.record(&module.name);
446 // FIXME: All upstream crates are deserialized internally in the
447 // function below to extract their summary and modules. Note that
448 // unlike the loop above we *must* decode and/or read something
449 // here as these are all just serialized files on disk. An
450 // improvement, however, to make here would be to store the
451 // module summary separately from the actual module itself. Right
452 // now this is store in one large bitcode file, and the entire
453 // file is deflate-compressed. We could try to bypass some of the
454 // decompression by storing the index uncompressed and only
455 // lazily decompressing the bytecode if necessary.
457 // Note that truly taking advantage of this optimization will
458 // likely be further down the road. We'd have to implement
459 // incremental ThinLTO first where we could actually avoid
460 // looking at upstream modules entirely sometimes (the contents,
461 // we must always unconditionally look at the index).
462 let mut serialized = Vec::new();
464 let cached_modules = cached_modules.into_iter().map(|(sm, wp)| {
465 (sm, CString::new(wp.cgu_name).unwrap())
468 for (module, name) in serialized_modules.into_iter().chain(cached_modules) {
469 info!("upstream or cached module {:?}", name);
470 thin_modules.push(llvm::ThinLTOModule {
471 identifier: name.as_ptr(),
472 data: module.data().as_ptr(),
473 len: module.data().len(),
475 serialized.push(module);
476 module_names.push(name);
480 assert_eq!(thin_modules.len(), module_names.len());
482 // Delegate to the C++ bindings to create some data here. Once this is a
483 // tried-and-true interface we may wish to try to upstream some of this
484 // to LLVM itself, right now we reimplement a lot of what they do
486 let data = llvm::LLVMRustCreateThinLTOData(
487 thin_modules.as_ptr(),
488 thin_modules.len() as u32,
489 symbol_white_list.as_ptr(),
490 symbol_white_list.len() as u32,
492 write::llvm_err(&diag_handler, "failed to prepare thin LTO context")
495 info!("thin LTO data created");
496 timeline.record("data");
498 let import_map = if cgcx.incr_comp_session_dir.is_some() {
499 ThinLTOImports::from_thin_lto_data(data)
501 // If we don't compile incrementally, we don't need to load the
502 // import data from LLVM.
503 assert!(green_modules.is_empty());
504 ThinLTOImports::default()
506 info!("thin LTO import map loaded");
507 timeline.record("import-map-loaded");
509 let data = ThinData(data);
511 // Throw our data in an `Arc` as we'll be sharing it across threads. We
512 // also put all memory referenced by the C++ data (buffers, ids, etc)
513 // into the arc as well. After this we'll create a thin module
514 // codegen per module in this data.
515 let shared = Arc::new(ThinShared {
518 serialized_modules: serialized,
522 let mut copy_jobs = vec![];
523 let mut opt_jobs = vec![];
525 info!("checking which modules can be-reused and which have to be re-optimized.");
526 for (module_index, module_name) in shared.module_names.iter().enumerate() {
527 let module_name = module_name_to_str(module_name);
529 // If the module hasn't changed and none of the modules it imports
530 // from has changed, we can re-use the post-ThinLTO version of the
532 if green_modules.contains_key(module_name) {
533 let imports_all_green = import_map.modules_imported_by(module_name)
535 .all(|imported_module| green_modules.contains_key(imported_module));
537 if imports_all_green {
538 let work_product = green_modules[module_name].clone();
539 copy_jobs.push(work_product);
540 info!(" - {}: re-used", module_name);
541 cgcx.cgu_reuse_tracker.set_actual_reuse(module_name,
547 info!(" - {}: re-compiled", module_name);
548 opt_jobs.push(LtoModuleCodegen::Thin(ThinModule {
549 shared: shared.clone(),
554 Ok((opt_jobs, copy_jobs))
558 fn run_pass_manager(cgcx: &CodegenContext,
559 tm: &llvm::TargetMachine,
560 llmod: &llvm::Module,
561 config: &ModuleConfig,
563 // Now we have one massive module inside of llmod. Time to run the
564 // LTO-specific optimization passes that LLVM provides.
566 // This code is based off the code found in llvm's LTO code generator:
567 // tools/lto/LTOCodeGenerator.cpp
568 debug!("running the pass manager");
570 let pm = llvm::LLVMCreatePassManager();
571 llvm::LLVMRustAddAnalysisPasses(tm, pm, llmod);
573 if config.verify_llvm_ir {
574 let pass = llvm::LLVMRustFindAndCreatePass("verify\0".as_ptr() as *const _);
575 llvm::LLVMRustAddPass(pm, pass.unwrap());
578 // When optimizing for LTO we don't actually pass in `-O0`, but we force
579 // it to always happen at least with `-O1`.
581 // With ThinLTO we mess around a lot with symbol visibility in a way
582 // that will actually cause linking failures if we optimize at O0 which
583 // notable is lacking in dead code elimination. To ensure we at least
584 // get some optimizations and correctly link we forcibly switch to `-O1`
585 // to get dead code elimination.
587 // Note that in general this shouldn't matter too much as you typically
588 // only turn on ThinLTO when you're compiling with optimizations
590 let opt_level = config.opt_level.unwrap_or(llvm::CodeGenOptLevel::None);
591 let opt_level = match opt_level {
592 llvm::CodeGenOptLevel::None => llvm::CodeGenOptLevel::Less,
595 with_llvm_pmb(llmod, config, opt_level, false, &mut |b| {
597 llvm::LLVMRustPassManagerBuilderPopulateThinLTOPassManager(b, pm);
599 llvm::LLVMPassManagerBuilderPopulateLTOPassManager(b, pm,
600 /* Internalize = */ False,
601 /* RunInliner = */ True);
605 // We always generate bitcode through ThinLTOBuffers,
606 // which do not support anonymous globals
607 if config.bitcode_needed() {
608 let pass = llvm::LLVMRustFindAndCreatePass("name-anon-globals\0".as_ptr() as *const _);
609 llvm::LLVMRustAddPass(pm, pass.unwrap());
612 if config.verify_llvm_ir {
613 let pass = llvm::LLVMRustFindAndCreatePass("verify\0".as_ptr() as *const _);
614 llvm::LLVMRustAddPass(pm, pass.unwrap());
617 time_ext(cgcx.time_passes, None, "LTO passes", || llvm::LLVMRunPassManager(pm, llmod));
619 llvm::LLVMDisposePassManager(pm);
624 pub enum SerializedModule {
627 FromUncompressedFile(memmap::Mmap),
630 impl SerializedModule {
631 fn data(&self) -> &[u8] {
633 SerializedModule::Local(ref m) => m.data(),
634 SerializedModule::FromRlib(ref m) => m,
635 SerializedModule::FromUncompressedFile(ref m) => m,
640 pub struct ModuleBuffer(&'static mut llvm::ModuleBuffer);
642 unsafe impl Send for ModuleBuffer {}
643 unsafe impl Sync for ModuleBuffer {}
646 pub fn new(m: &llvm::Module) -> ModuleBuffer {
647 ModuleBuffer(unsafe {
648 llvm::LLVMRustModuleBufferCreate(m)
652 pub fn data(&self) -> &[u8] {
654 let ptr = llvm::LLVMRustModuleBufferPtr(self.0);
655 let len = llvm::LLVMRustModuleBufferLen(self.0);
656 slice::from_raw_parts(ptr, len)
661 impl Drop for ModuleBuffer {
663 unsafe { llvm::LLVMRustModuleBufferFree(&mut *(self.0 as *mut _)); }
667 pub struct ThinModule {
668 shared: Arc<ThinShared>,
674 thin_buffers: Vec<ThinBuffer>,
675 serialized_modules: Vec<SerializedModule>,
676 module_names: Vec<CString>,
679 struct ThinData(&'static mut llvm::ThinLTOData);
681 unsafe impl Send for ThinData {}
682 unsafe impl Sync for ThinData {}
684 impl Drop for ThinData {
687 llvm::LLVMRustFreeThinLTOData(&mut *(self.0 as *mut _));
692 pub struct ThinBuffer(&'static mut llvm::ThinLTOBuffer);
694 unsafe impl Send for ThinBuffer {}
695 unsafe impl Sync for ThinBuffer {}
698 pub fn new(m: &llvm::Module) -> ThinBuffer {
700 let buffer = llvm::LLVMRustThinLTOBufferCreate(m);
705 pub fn data(&self) -> &[u8] {
707 let ptr = llvm::LLVMRustThinLTOBufferPtr(self.0) as *const _;
708 let len = llvm::LLVMRustThinLTOBufferLen(self.0);
709 slice::from_raw_parts(ptr, len)
714 impl Drop for ThinBuffer {
717 llvm::LLVMRustThinLTOBufferFree(&mut *(self.0 as *mut _));
723 fn name(&self) -> &str {
724 self.shared.module_names[self.idx].to_str().unwrap()
727 fn cost(&self) -> u64 {
728 // Yes, that's correct, we're using the size of the bytecode as an
729 // indicator for how costly this codegen unit is.
730 self.data().len() as u64
733 fn data(&self) -> &[u8] {
734 let a = self.shared.thin_buffers.get(self.idx).map(|b| b.data());
735 a.unwrap_or_else(|| {
736 let len = self.shared.thin_buffers.len();
737 self.shared.serialized_modules[self.idx - len].data()
741 unsafe fn optimize(&mut self, cgcx: &CodegenContext, timeline: &mut Timeline)
742 -> Result<ModuleCodegen<ModuleLlvm>, FatalError>
744 let diag_handler = cgcx.create_diag_handler();
745 let tm = (cgcx.tm_factory)().map_err(|e| {
746 write::llvm_err(&diag_handler, &e)
749 // Right now the implementation we've got only works over serialized
750 // modules, so we create a fresh new LLVM context and parse the module
751 // into that context. One day, however, we may do this for upstream
752 // crates but for locally codegened modules we may be able to reuse
753 // that LLVM Context and Module.
754 let llcx = llvm::LLVMRustContextCreate(cgcx.fewer_names);
755 let llmod_raw = llvm::LLVMRustParseBitcodeForThinLTO(
757 self.data().as_ptr(),
759 self.shared.module_names[self.idx].as_ptr(),
761 let msg = "failed to parse bitcode for thin LTO module";
762 write::llvm_err(&diag_handler, msg)
764 let module = ModuleCodegen {
765 module_llvm: ModuleLlvm {
770 name: self.name().to_string(),
771 kind: ModuleKind::Regular,
774 let llmod = module.module_llvm.llmod();
775 cgcx.save_temp_bitcode(&module, "thin-lto-input");
777 // Before we do much else find the "main" `DICompileUnit` that we'll be
778 // using below. If we find more than one though then rustc has changed
779 // in a way we're not ready for, so generate an ICE by returning
781 let mut cu1 = ptr::null_mut();
782 let mut cu2 = ptr::null_mut();
783 llvm::LLVMRustThinLTOGetDICompileUnit(llmod, &mut cu1, &mut cu2);
785 let msg = "multiple source DICompileUnits found";
786 return Err(write::llvm_err(&diag_handler, msg))
789 // Like with "fat" LTO, get some better optimizations if landing pads
790 // are disabled by removing all landing pads.
791 if cgcx.no_landing_pads {
792 llvm::LLVMRustMarkAllFunctionsNounwind(llmod);
793 cgcx.save_temp_bitcode(&module, "thin-lto-after-nounwind");
794 timeline.record("nounwind");
797 // Up next comes the per-module local analyses that we do for Thin LTO.
798 // Each of these functions is basically copied from the LLVM
799 // implementation and then tailored to suit this implementation. Ideally
800 // each of these would be supported by upstream LLVM but that's perhaps
801 // a patch for another day!
803 // You can find some more comments about these functions in the LLVM
804 // bindings we've got (currently `PassWrapper.cpp`)
805 if !llvm::LLVMRustPrepareThinLTORename(self.shared.data.0, llmod) {
806 let msg = "failed to prepare thin LTO module";
807 return Err(write::llvm_err(&diag_handler, msg))
809 cgcx.save_temp_bitcode(&module, "thin-lto-after-rename");
810 timeline.record("rename");
811 if !llvm::LLVMRustPrepareThinLTOResolveWeak(self.shared.data.0, llmod) {
812 let msg = "failed to prepare thin LTO module";
813 return Err(write::llvm_err(&diag_handler, msg))
815 cgcx.save_temp_bitcode(&module, "thin-lto-after-resolve");
816 timeline.record("resolve");
817 if !llvm::LLVMRustPrepareThinLTOInternalize(self.shared.data.0, llmod) {
818 let msg = "failed to prepare thin LTO module";
819 return Err(write::llvm_err(&diag_handler, msg))
821 cgcx.save_temp_bitcode(&module, "thin-lto-after-internalize");
822 timeline.record("internalize");
823 if !llvm::LLVMRustPrepareThinLTOImport(self.shared.data.0, llmod) {
824 let msg = "failed to prepare thin LTO module";
825 return Err(write::llvm_err(&diag_handler, msg))
827 cgcx.save_temp_bitcode(&module, "thin-lto-after-import");
828 timeline.record("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
859 llvm::LLVMRustThinLTOPatchDICompileUnit(llmod, cu1);
860 cgcx.save_temp_bitcode(&module, "thin-lto-after-patch");
861 timeline.record("patch");
863 // Alright now that we've done everything related to the ThinLTO
864 // analysis it's time to run some optimizations! Here we use the same
865 // `run_pass_manager` as the "fat" LTO above except that we tell it to
866 // populate a thin-specific pass manager, which presumably LLVM treats a
867 // little differently.
868 info!("running thin lto passes over {}", module.name);
869 let config = cgcx.config(module.kind);
870 run_pass_manager(cgcx, module.module_llvm.tm, llmod, config, true);
871 cgcx.save_temp_bitcode(&module, "thin-lto-after-pm");
872 timeline.record("thin-done");
879 #[derive(Debug, Default)]
880 pub struct ThinLTOImports {
881 // key = llvm name of importing module, value = list of modules it imports from
882 imports: FxHashMap<String, Vec<String>>,
885 impl ThinLTOImports {
886 fn modules_imported_by(&self, llvm_module_name: &str) -> &[String] {
887 self.imports.get(llvm_module_name).map(|v| &v[..]).unwrap_or(&[])
890 /// Load the ThinLTO import map from ThinLTOData.
891 unsafe fn from_thin_lto_data(data: *const llvm::ThinLTOData) -> ThinLTOImports {
892 unsafe extern "C" fn imported_module_callback(payload: *mut libc::c_void,
893 importing_module_name: *const libc::c_char,
894 imported_module_name: *const libc::c_char) {
895 let map = &mut* (payload as *mut ThinLTOImports);
896 let importing_module_name = CStr::from_ptr(importing_module_name);
897 let importing_module_name = module_name_to_str(&importing_module_name);
898 let imported_module_name = CStr::from_ptr(imported_module_name);
899 let imported_module_name = module_name_to_str(&imported_module_name);
901 if !map.imports.contains_key(importing_module_name) {
902 map.imports.insert(importing_module_name.to_owned(), vec![]);
906 .get_mut(importing_module_name)
908 .push(imported_module_name.to_owned());
910 let mut map = ThinLTOImports::default();
911 llvm::LLVMRustGetThinLTOModuleImports(data,
912 imported_module_callback,
913 &mut map as *mut _ as *mut libc::c_void);
918 fn module_name_to_str(c_str: &CStr) -> &str {
919 c_str.to_str().unwrap_or_else(|e|
920 bug!("Encountered non-utf8 LLVM module name `{}`: {}", c_str.to_string_lossy(), e))