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 rustc_codegen_ssa::back::symbol_export;
13 use rustc_codegen_ssa::back::write::{ModuleConfig, CodegenContext, pre_lto_bitcode_filename};
14 use rustc_codegen_ssa::back::lto::{SerializedModule, LtoModuleCodegen, ThinShared, ThinModule};
15 use rustc_codegen_ssa::traits::*;
16 use back::write::{self, DiagnosticHandlers, with_llvm_pmb, save_temp_bitcode, get_llvm_opt_level};
17 use errors::{FatalError, Handler};
18 use llvm::archive_ro::ArchiveRO;
19 use llvm::{self, True, False};
20 use rustc::dep_graph::WorkProduct;
21 use rustc::dep_graph::cgu_reuse_tracker::CguReuse;
22 use rustc::hir::def_id::LOCAL_CRATE;
23 use rustc::middle::exported_symbols::SymbolExportLevel;
24 use rustc::session::config::{self, Lto};
25 use rustc::util::common::time_ext;
26 use rustc_data_structures::fx::FxHashMap;
27 use time_graph::Timeline;
28 use {ModuleLlvm, LlvmCodegenBackend};
29 use rustc_codegen_ssa::{ModuleCodegen, ModuleKind};
33 use std::ffi::{CStr, CString};
39 pub fn crate_type_allows_lto(crate_type: config::CrateType) -> bool {
41 config::CrateType::Executable |
42 config::CrateType::Staticlib |
43 config::CrateType::Cdylib => true,
45 config::CrateType::Dylib |
46 config::CrateType::Rlib |
47 config::CrateType::ProcMacro => false,
51 fn prepare_lto(cgcx: &CodegenContext<LlvmCodegenBackend>,
52 timeline: &mut Timeline,
53 diag_handler: &Handler)
54 -> Result<(Vec<CString>, Vec<(SerializedModule<ModuleBuffer>, CString)>), FatalError>
56 let export_threshold = match cgcx.lto {
57 // We're just doing LTO for our one crate
58 Lto::ThinLocal => SymbolExportLevel::Rust,
60 // We're doing LTO for the entire crate graph
61 Lto::Fat | Lto::Thin => {
62 symbol_export::crates_export_threshold(&cgcx.crate_types)
65 Lto::No => panic!("didn't request LTO but we're doing LTO"),
68 let symbol_filter = &|&(ref name, level): &(String, SymbolExportLevel)| {
69 if level.is_below_threshold(export_threshold) {
70 let mut bytes = Vec::with_capacity(name.len() + 1);
71 bytes.extend(name.bytes());
72 Some(CString::new(bytes).unwrap())
77 let exported_symbols = cgcx.exported_symbols
78 .as_ref().expect("needs exported symbols for LTO");
79 let mut symbol_white_list = exported_symbols[&LOCAL_CRATE]
81 .filter_map(symbol_filter)
82 .collect::<Vec<CString>>();
83 timeline.record("whitelist");
84 info!("{} symbols to preserve in this crate", symbol_white_list.len());
86 // If we're performing LTO for the entire crate graph, then for each of our
87 // upstream dependencies, find the corresponding rlib and load the bitcode
90 // We save off all the bytecode and LLVM module ids for later processing
91 // with either fat or thin LTO
92 let mut upstream_modules = Vec::new();
93 if cgcx.lto != Lto::ThinLocal {
94 if cgcx.opts.cg.prefer_dynamic {
95 diag_handler.struct_err("cannot prefer dynamic linking when performing LTO")
96 .note("only 'staticlib', 'bin', and 'cdylib' outputs are \
99 return Err(FatalError)
102 // Make sure we actually can run LTO
103 for crate_type in cgcx.crate_types.iter() {
104 if !crate_type_allows_lto(*crate_type) {
105 let e = diag_handler.fatal("lto can only be run for executables, cdylibs and \
106 static library outputs");
111 for &(cnum, ref path) in cgcx.each_linked_rlib_for_lto.iter() {
112 let exported_symbols = cgcx.exported_symbols
113 .as_ref().expect("needs exported symbols for LTO");
114 symbol_white_list.extend(
115 exported_symbols[&cnum]
117 .filter_map(symbol_filter));
119 let archive = ArchiveRO::open(&path).expect("wanted an rlib");
120 let bytecodes = archive.iter().filter_map(|child| {
121 child.ok().and_then(|c| c.name().map(|name| (name, c)))
122 }).filter(|&(name, _)| name.ends_with(RLIB_BYTECODE_EXTENSION));
123 for (name, data) in bytecodes {
124 info!("adding bytecode {}", name);
125 let bc_encoded = data.data();
127 let (bc, id) = time_ext(cgcx.time_passes, None, &format!("decode {}", name), || {
128 match DecodedBytecode::new(bc_encoded) {
129 Ok(b) => Ok((b.bytecode(), b.identifier().to_string())),
130 Err(e) => Err(diag_handler.fatal(&e)),
133 let bc = SerializedModule::FromRlib(bc);
134 upstream_modules.push((bc, CString::new(id).unwrap()));
136 timeline.record(&format!("load: {}", path.display()));
140 Ok((symbol_white_list, upstream_modules))
143 /// Performs fat LTO by merging all modules into a single one and returning it
144 /// for further optimization.
145 pub(crate) fn run_fat(cgcx: &CodegenContext<LlvmCodegenBackend>,
146 modules: Vec<ModuleCodegen<ModuleLlvm>>,
147 timeline: &mut Timeline)
148 -> Result<LtoModuleCodegen<LlvmCodegenBackend>, FatalError>
150 let diag_handler = cgcx.create_diag_handler();
151 let (symbol_white_list, upstream_modules) = prepare_lto(cgcx, timeline, &diag_handler)?;
152 let symbol_white_list = symbol_white_list.iter()
154 .collect::<Vec<_>>();
155 fat_lto(cgcx, &diag_handler, modules, upstream_modules, &symbol_white_list, timeline)
158 /// Performs thin LTO by performing necessary global analysis and returning two
159 /// lists, one of the modules that need optimization and another for modules that
160 /// can simply be copied over from the incr. comp. cache.
161 pub(crate) fn run_thin(cgcx: &CodegenContext<LlvmCodegenBackend>,
162 modules: Vec<(String, ThinBuffer)>,
163 cached_modules: Vec<(SerializedModule<ModuleBuffer>, WorkProduct)>,
164 timeline: &mut Timeline)
165 -> Result<(Vec<LtoModuleCodegen<LlvmCodegenBackend>>, Vec<WorkProduct>), FatalError>
167 let diag_handler = cgcx.create_diag_handler();
168 let (symbol_white_list, upstream_modules) = prepare_lto(cgcx, timeline, &diag_handler)?;
169 let symbol_white_list = symbol_white_list.iter()
171 .collect::<Vec<_>>();
172 if cgcx.opts.debugging_opts.cross_lang_lto.enabled() {
173 unreachable!("We should never reach this case if the LTO step \
174 is deferred to the linker");
185 pub(crate) fn prepare_thin(
186 cgcx: &CodegenContext<LlvmCodegenBackend>,
187 module: ModuleCodegen<ModuleLlvm>
188 ) -> (String, ThinBuffer) {
189 let name = module.name.clone();
190 let buffer = ThinBuffer::new(module.module_llvm.llmod());
192 // We emit the module after having serialized it into a ThinBuffer
193 // because only then it will contain the ThinLTO module summary.
194 if let Some(ref incr_comp_session_dir) = cgcx.incr_comp_session_dir {
195 if cgcx.config(module.kind).emit_pre_thin_lto_bc {
196 let path = incr_comp_session_dir
197 .join(pre_lto_bitcode_filename(&name));
199 fs::write(&path, buffer.data()).unwrap_or_else(|e| {
200 panic!("Error writing pre-lto-bitcode file `{}`: {}",
210 fn fat_lto(cgcx: &CodegenContext<LlvmCodegenBackend>,
211 diag_handler: &Handler,
212 mut modules: Vec<ModuleCodegen<ModuleLlvm>>,
213 mut serialized_modules: Vec<(SerializedModule<ModuleBuffer>, CString)>,
214 symbol_white_list: &[*const libc::c_char],
215 timeline: &mut Timeline)
216 -> Result<LtoModuleCodegen<LlvmCodegenBackend>, FatalError>
218 info!("going for a fat lto");
220 // Find the "costliest" module and merge everything into that codegen unit.
221 // All the other modules will be serialized and reparsed into the new
222 // context, so this hopefully avoids serializing and parsing the largest
225 // Additionally use a regular module as the base here to ensure that various
226 // file copy operations in the backend work correctly. The only other kind
227 // of module here should be an allocator one, and if your crate is smaller
228 // than the allocator module then the size doesn't really matter anyway.
229 let (_, costliest_module) = modules.iter()
231 .filter(|&(_, module)| module.kind == ModuleKind::Regular)
234 llvm::LLVMRustModuleCost(module.module_llvm.llmod())
239 .expect("must be codegen'ing at least one module");
240 let module = modules.remove(costliest_module);
241 let mut serialized_bitcode = Vec::new();
243 let (llcx, llmod) = {
244 let llvm = &module.module_llvm;
245 (&llvm.llcx, llvm.llmod())
247 info!("using {:?} as a base module", module.name);
249 // The linking steps below may produce errors and diagnostics within LLVM
250 // which we'd like to handle and print, so set up our diagnostic handlers
251 // (which get unregistered when they go out of scope below).
252 let _handler = DiagnosticHandlers::new(cgcx, diag_handler, llcx);
254 // For all other modules we codegened we'll need to link them into our own
255 // bitcode. All modules were codegened in their own LLVM context, however,
256 // and we want to move everything to the same LLVM context. Currently the
257 // way we know of to do that is to serialize them to a string and them parse
258 // them later. Not great but hey, that's why it's "fat" LTO, right?
259 serialized_modules.extend(modules.into_iter().map(|module| {
260 let buffer = ModuleBuffer::new(module.module_llvm.llmod());
261 let llmod_id = CString::new(&module.name[..]).unwrap();
263 (SerializedModule::Local(buffer), llmod_id)
266 // For all serialized bitcode files we parse them and link them in as we did
267 // above, this is all mostly handled in C++. Like above, though, we don't
268 // know much about the memory management here so we err on the side of being
269 // save and persist everything with the original module.
270 let mut linker = Linker::new(llmod);
271 for (bc_decoded, name) in serialized_modules {
272 info!("linking {:?}", name);
273 time_ext(cgcx.time_passes, None, &format!("ll link {:?}", name), || {
274 let data = bc_decoded.data();
275 linker.add(&data).map_err(|()| {
276 let msg = format!("failed to load bc of {:?}", name);
277 write::llvm_err(&diag_handler, &msg)
280 timeline.record(&format!("link {:?}", name));
281 serialized_bitcode.push(bc_decoded);
284 save_temp_bitcode(&cgcx, &module, "lto.input");
286 // Internalize everything that *isn't* in our whitelist to help strip out
287 // more modules and such
289 let ptr = symbol_white_list.as_ptr();
290 llvm::LLVMRustRunRestrictionPass(llmod,
291 ptr as *const *const libc::c_char,
292 symbol_white_list.len() as libc::size_t);
293 save_temp_bitcode(&cgcx, &module, "lto.after-restriction");
296 if cgcx.no_landing_pads {
298 llvm::LLVMRustMarkAllFunctionsNounwind(llmod);
300 save_temp_bitcode(&cgcx, &module, "lto.after-nounwind");
302 timeline.record("passes");
305 Ok(LtoModuleCodegen::Fat {
306 module: Some(module),
307 _serialized_bitcode: serialized_bitcode,
311 struct Linker<'a>(&'a mut llvm::Linker<'a>);
314 fn new(llmod: &'a llvm::Module) -> Self {
315 unsafe { Linker(llvm::LLVMRustLinkerNew(llmod)) }
318 fn add(&mut self, bytecode: &[u8]) -> Result<(), ()> {
320 if llvm::LLVMRustLinkerAdd(self.0,
321 bytecode.as_ptr() as *const libc::c_char,
331 impl Drop for Linker<'a> {
333 unsafe { llvm::LLVMRustLinkerFree(&mut *(self.0 as *mut _)); }
337 /// Prepare "thin" LTO to get run on these modules.
339 /// The general structure of ThinLTO is quite different from the structure of
340 /// "fat" LTO above. With "fat" LTO all LLVM modules in question are merged into
341 /// one giant LLVM module, and then we run more optimization passes over this
342 /// big module after internalizing most symbols. Thin LTO, on the other hand,
343 /// avoid this large bottleneck through more targeted optimization.
345 /// At a high level Thin LTO looks like:
347 /// 1. Prepare a "summary" of each LLVM module in question which describes
348 /// the values inside, cost of the values, etc.
349 /// 2. Merge the summaries of all modules in question into one "index"
350 /// 3. Perform some global analysis on this index
351 /// 4. For each module, use the index and analysis calculated previously to
352 /// perform local transformations on the module, for example inlining
353 /// small functions from other modules.
354 /// 5. Run thin-specific optimization passes over each module, and then code
355 /// generate everything at the end.
357 /// The summary for each module is intended to be quite cheap, and the global
358 /// index is relatively quite cheap to create as well. As a result, the goal of
359 /// ThinLTO is to reduce the bottleneck on LTO and enable LTO to be used in more
360 /// situations. For example one cheap optimization is that we can parallelize
361 /// all codegen modules, easily making use of all the cores on a machine.
363 /// With all that in mind, the function here is designed at specifically just
364 /// calculating the *index* for ThinLTO. This index will then be shared amongst
365 /// all of the `LtoModuleCodegen` units returned below and destroyed once
366 /// they all go out of scope.
367 fn thin_lto(cgcx: &CodegenContext<LlvmCodegenBackend>,
368 diag_handler: &Handler,
369 modules: Vec<(String, ThinBuffer)>,
370 serialized_modules: Vec<(SerializedModule<ModuleBuffer>, CString)>,
371 cached_modules: Vec<(SerializedModule<ModuleBuffer>, WorkProduct)>,
372 symbol_white_list: &[*const libc::c_char],
373 timeline: &mut Timeline)
374 -> Result<(Vec<LtoModuleCodegen<LlvmCodegenBackend>>, Vec<WorkProduct>), FatalError>
377 info!("going for that thin, thin LTO");
379 let green_modules: FxHashMap<_, _> = cached_modules
381 .map(|&(_, ref wp)| (wp.cgu_name.clone(), wp.clone()))
384 let full_scope_len = modules.len() + serialized_modules.len() + cached_modules.len();
385 let mut thin_buffers = Vec::with_capacity(modules.len());
386 let mut module_names = Vec::with_capacity(full_scope_len);
387 let mut thin_modules = Vec::with_capacity(full_scope_len);
389 for (i, (name, buffer)) in modules.into_iter().enumerate() {
390 info!("local module: {} - {}", i, name);
391 let cname = CString::new(name.clone()).unwrap();
392 thin_modules.push(llvm::ThinLTOModule {
393 identifier: cname.as_ptr(),
394 data: buffer.data().as_ptr(),
395 len: buffer.data().len(),
397 thin_buffers.push(buffer);
398 module_names.push(cname);
399 timeline.record(&name);
402 // FIXME: All upstream crates are deserialized internally in the
403 // function below to extract their summary and modules. Note that
404 // unlike the loop above we *must* decode and/or read something
405 // here as these are all just serialized files on disk. An
406 // improvement, however, to make here would be to store the
407 // module summary separately from the actual module itself. Right
408 // now this is store in one large bitcode file, and the entire
409 // file is deflate-compressed. We could try to bypass some of the
410 // decompression by storing the index uncompressed and only
411 // lazily decompressing the bytecode if necessary.
413 // Note that truly taking advantage of this optimization will
414 // likely be further down the road. We'd have to implement
415 // incremental ThinLTO first where we could actually avoid
416 // looking at upstream modules entirely sometimes (the contents,
417 // we must always unconditionally look at the index).
418 let mut serialized = Vec::with_capacity(serialized_modules.len() + cached_modules.len());
420 let cached_modules = cached_modules.into_iter().map(|(sm, wp)| {
421 (sm, CString::new(wp.cgu_name).unwrap())
424 for (module, name) in serialized_modules.into_iter().chain(cached_modules) {
425 info!("upstream or cached module {:?}", name);
426 thin_modules.push(llvm::ThinLTOModule {
427 identifier: name.as_ptr(),
428 data: module.data().as_ptr(),
429 len: module.data().len(),
431 serialized.push(module);
432 module_names.push(name);
436 assert_eq!(thin_modules.len(), module_names.len());
438 // Delegate to the C++ bindings to create some data here. Once this is a
439 // tried-and-true interface we may wish to try to upstream some of this
440 // to LLVM itself, right now we reimplement a lot of what they do
442 let data = llvm::LLVMRustCreateThinLTOData(
443 thin_modules.as_ptr(),
444 thin_modules.len() as u32,
445 symbol_white_list.as_ptr(),
446 symbol_white_list.len() as u32,
448 write::llvm_err(&diag_handler, "failed to prepare thin LTO context")
451 info!("thin LTO data created");
452 timeline.record("data");
454 let import_map = if cgcx.incr_comp_session_dir.is_some() {
455 ThinLTOImports::from_thin_lto_data(data)
457 // If we don't compile incrementally, we don't need to load the
458 // import data from LLVM.
459 assert!(green_modules.is_empty());
460 ThinLTOImports::default()
462 info!("thin LTO import map loaded");
463 timeline.record("import-map-loaded");
465 let data = ThinData(data);
467 // Throw our data in an `Arc` as we'll be sharing it across threads. We
468 // also put all memory referenced by the C++ data (buffers, ids, etc)
469 // into the arc as well. After this we'll create a thin module
470 // codegen per module in this data.
471 let shared = Arc::new(ThinShared {
474 serialized_modules: serialized,
478 let mut copy_jobs = vec![];
479 let mut opt_jobs = vec![];
481 info!("checking which modules can be-reused and which have to be re-optimized.");
482 for (module_index, module_name) in shared.module_names.iter().enumerate() {
483 let module_name = module_name_to_str(module_name);
485 // If the module hasn't changed and none of the modules it imports
486 // from has changed, we can re-use the post-ThinLTO version of the
488 if green_modules.contains_key(module_name) {
489 let imports_all_green = import_map.modules_imported_by(module_name)
491 .all(|imported_module| green_modules.contains_key(imported_module));
493 if imports_all_green {
494 let work_product = green_modules[module_name].clone();
495 copy_jobs.push(work_product);
496 info!(" - {}: re-used", module_name);
497 cgcx.cgu_reuse_tracker.set_actual_reuse(module_name,
503 info!(" - {}: re-compiled", module_name);
504 opt_jobs.push(LtoModuleCodegen::Thin(ThinModule {
505 shared: shared.clone(),
510 Ok((opt_jobs, copy_jobs))
514 pub(crate) fn run_pass_manager(cgcx: &CodegenContext<LlvmCodegenBackend>,
515 module: &ModuleCodegen<ModuleLlvm>,
516 config: &ModuleConfig,
518 // Now we have one massive module inside of llmod. Time to run the
519 // LTO-specific optimization passes that LLVM provides.
521 // This code is based off the code found in llvm's LTO code generator:
522 // tools/lto/LTOCodeGenerator.cpp
523 debug!("running the pass manager");
525 let pm = llvm::LLVMCreatePassManager();
526 llvm::LLVMRustAddAnalysisPasses(module.module_llvm.tm, pm, module.module_llvm.llmod());
528 if config.verify_llvm_ir {
529 let pass = llvm::LLVMRustFindAndCreatePass("verify\0".as_ptr() as *const _);
530 llvm::LLVMRustAddPass(pm, pass.unwrap());
533 // When optimizing for LTO we don't actually pass in `-O0`, but we force
534 // it to always happen at least with `-O1`.
536 // With ThinLTO we mess around a lot with symbol visibility in a way
537 // that will actually cause linking failures if we optimize at O0 which
538 // notable is lacking in dead code elimination. To ensure we at least
539 // get some optimizations and correctly link we forcibly switch to `-O1`
540 // to get dead code elimination.
542 // Note that in general this shouldn't matter too much as you typically
543 // only turn on ThinLTO when you're compiling with optimizations
545 let opt_level = config.opt_level.map(get_llvm_opt_level)
546 .unwrap_or(llvm::CodeGenOptLevel::None);
547 let opt_level = match opt_level {
548 llvm::CodeGenOptLevel::None => llvm::CodeGenOptLevel::Less,
551 with_llvm_pmb(module.module_llvm.llmod(), config, opt_level, false, &mut |b| {
553 llvm::LLVMRustPassManagerBuilderPopulateThinLTOPassManager(b, pm);
555 llvm::LLVMPassManagerBuilderPopulateLTOPassManager(b, pm,
556 /* Internalize = */ False,
557 /* RunInliner = */ True);
561 // We always generate bitcode through ThinLTOBuffers,
562 // which do not support anonymous globals
563 if config.bitcode_needed() {
564 let pass = llvm::LLVMRustFindAndCreatePass("name-anon-globals\0".as_ptr() as *const _);
565 llvm::LLVMRustAddPass(pm, pass.unwrap());
568 if config.verify_llvm_ir {
569 let pass = llvm::LLVMRustFindAndCreatePass("verify\0".as_ptr() as *const _);
570 llvm::LLVMRustAddPass(pm, pass.unwrap());
573 time_ext(cgcx.time_passes, None, "LTO passes", ||
574 llvm::LLVMRunPassManager(pm, module.module_llvm.llmod()));
576 llvm::LLVMDisposePassManager(pm);
581 pub struct ModuleBuffer(&'static mut llvm::ModuleBuffer);
583 unsafe impl Send for ModuleBuffer {}
584 unsafe impl Sync for ModuleBuffer {}
587 pub fn new(m: &llvm::Module) -> ModuleBuffer {
588 ModuleBuffer(unsafe {
589 llvm::LLVMRustModuleBufferCreate(m)
594 impl ModuleBufferMethods for ModuleBuffer {
595 fn data(&self) -> &[u8] {
597 let ptr = llvm::LLVMRustModuleBufferPtr(self.0);
598 let len = llvm::LLVMRustModuleBufferLen(self.0);
599 slice::from_raw_parts(ptr, len)
604 impl Drop for ModuleBuffer {
606 unsafe { llvm::LLVMRustModuleBufferFree(&mut *(self.0 as *mut _)); }
610 pub struct ThinData(&'static mut llvm::ThinLTOData);
612 unsafe impl Send for ThinData {}
613 unsafe impl Sync for ThinData {}
615 impl Drop for ThinData {
618 llvm::LLVMRustFreeThinLTOData(&mut *(self.0 as *mut _));
623 pub struct ThinBuffer(&'static mut llvm::ThinLTOBuffer);
625 unsafe impl Send for ThinBuffer {}
626 unsafe impl Sync for ThinBuffer {}
629 pub fn new(m: &llvm::Module) -> ThinBuffer {
631 let buffer = llvm::LLVMRustThinLTOBufferCreate(m);
637 impl ThinBufferMethods for ThinBuffer {
638 fn data(&self) -> &[u8] {
640 let ptr = llvm::LLVMRustThinLTOBufferPtr(self.0) as *const _;
641 let len = llvm::LLVMRustThinLTOBufferLen(self.0);
642 slice::from_raw_parts(ptr, len)
647 impl Drop for ThinBuffer {
650 llvm::LLVMRustThinLTOBufferFree(&mut *(self.0 as *mut _));
655 pub unsafe fn optimize_thin_module(
656 thin_module: &mut ThinModule<LlvmCodegenBackend>,
657 cgcx: &CodegenContext<LlvmCodegenBackend>,
658 timeline: &mut Timeline
659 ) -> Result<ModuleCodegen<ModuleLlvm>, FatalError> {
660 let diag_handler = cgcx.create_diag_handler();
661 let tm = (cgcx.tm_factory.0)().map_err(|e| {
662 write::llvm_err(&diag_handler, &e)
665 // Right now the implementation we've got only works over serialized
666 // modules, so we create a fresh new LLVM context and parse the module
667 // into that context. One day, however, we may do this for upstream
668 // crates but for locally codegened modules we may be able to reuse
669 // that LLVM Context and Module.
670 let llcx = llvm::LLVMRustContextCreate(cgcx.fewer_names);
671 let llmod_raw = llvm::LLVMRustParseBitcodeForThinLTO(
673 thin_module.data().as_ptr(),
674 thin_module.data().len(),
675 thin_module.shared.module_names[thin_module.idx].as_ptr(),
677 let msg = "failed to parse bitcode for thin LTO module";
678 write::llvm_err(&diag_handler, msg)
680 let module = ModuleCodegen {
681 module_llvm: ModuleLlvm {
686 name: thin_module.name().to_string(),
687 kind: ModuleKind::Regular,
690 let llmod = module.module_llvm.llmod();
691 save_temp_bitcode(&cgcx, &module, "thin-lto-input");
693 // Before we do much else find the "main" `DICompileUnit` that we'll be
694 // using below. If we find more than one though then rustc has changed
695 // in a way we're not ready for, so generate an ICE by returning
697 let mut cu1 = ptr::null_mut();
698 let mut cu2 = ptr::null_mut();
699 llvm::LLVMRustThinLTOGetDICompileUnit(llmod, &mut cu1, &mut cu2);
701 let msg = "multiple source DICompileUnits found";
702 return Err(write::llvm_err(&diag_handler, msg))
705 // Like with "fat" LTO, get some better optimizations if landing pads
706 // are disabled by removing all landing pads.
707 if cgcx.no_landing_pads {
708 llvm::LLVMRustMarkAllFunctionsNounwind(llmod);
709 save_temp_bitcode(&cgcx, &module, "thin-lto-after-nounwind");
710 timeline.record("nounwind");
713 // Up next comes the per-module local analyses that we do for Thin LTO.
714 // Each of these functions is basically copied from the LLVM
715 // implementation and then tailored to suit this implementation. Ideally
716 // each of these would be supported by upstream LLVM but that's perhaps
717 // a patch for another day!
719 // You can find some more comments about these functions in the LLVM
720 // bindings we've got (currently `PassWrapper.cpp`)
721 if !llvm::LLVMRustPrepareThinLTORename(thin_module.shared.data.0, llmod) {
722 let msg = "failed to prepare thin LTO module";
723 return Err(write::llvm_err(&diag_handler, msg))
725 save_temp_bitcode(cgcx, &module, "thin-lto-after-rename");
726 timeline.record("rename");
727 if !llvm::LLVMRustPrepareThinLTOResolveWeak(thin_module.shared.data.0, llmod) {
728 let msg = "failed to prepare thin LTO module";
729 return Err(write::llvm_err(&diag_handler, msg))
731 save_temp_bitcode(cgcx, &module, "thin-lto-after-resolve");
732 timeline.record("resolve");
733 if !llvm::LLVMRustPrepareThinLTOInternalize(thin_module.shared.data.0, llmod) {
734 let msg = "failed to prepare thin LTO module";
735 return Err(write::llvm_err(&diag_handler, msg))
737 save_temp_bitcode(cgcx, &module, "thin-lto-after-internalize");
738 timeline.record("internalize");
739 if !llvm::LLVMRustPrepareThinLTOImport(thin_module.shared.data.0, llmod) {
740 let msg = "failed to prepare thin LTO module";
741 return Err(write::llvm_err(&diag_handler, msg))
743 save_temp_bitcode(cgcx, &module, "thin-lto-after-import");
744 timeline.record("import");
746 // Ok now this is a bit unfortunate. This is also something you won't
747 // find upstream in LLVM's ThinLTO passes! This is a hack for now to
748 // work around bugs in LLVM.
750 // First discovered in #45511 it was found that as part of ThinLTO
751 // importing passes LLVM will import `DICompileUnit` metadata
752 // information across modules. This means that we'll be working with one
753 // LLVM module that has multiple `DICompileUnit` instances in it (a
754 // bunch of `llvm.dbg.cu` members). Unfortunately there's a number of
755 // bugs in LLVM's backend which generates invalid DWARF in a situation
758 // https://bugs.llvm.org/show_bug.cgi?id=35212
759 // https://bugs.llvm.org/show_bug.cgi?id=35562
761 // While the first bug there is fixed the second ended up causing #46346
762 // which was basically a resurgence of #45511 after LLVM's bug 35212 was
765 // This function below is a huge hack around this problem. The function
766 // below is defined in `PassWrapper.cpp` and will basically "merge"
767 // all `DICompileUnit` instances in a module. Basically it'll take all
768 // the objects, rewrite all pointers of `DISubprogram` to point to the
769 // first `DICompileUnit`, and then delete all the other units.
771 // This is probably mangling to the debug info slightly (but hopefully
772 // not too much) but for now at least gets LLVM to emit valid DWARF (or
773 // so it appears). Hopefully we can remove this once upstream bugs are
775 llvm::LLVMRustThinLTOPatchDICompileUnit(llmod, cu1);
776 save_temp_bitcode(cgcx, &module, "thin-lto-after-patch");
777 timeline.record("patch");
779 // Alright now that we've done everything related to the ThinLTO
780 // analysis it's time to run some optimizations! Here we use the same
781 // `run_pass_manager` as the "fat" LTO above except that we tell it to
782 // populate a thin-specific pass manager, which presumably LLVM treats a
783 // little differently.
784 info!("running thin lto passes over {}", module.name);
785 let config = cgcx.config(module.kind);
786 run_pass_manager(cgcx, &module, config, true);
787 save_temp_bitcode(cgcx, &module, "thin-lto-after-pm");
788 timeline.record("thin-done");
793 #[derive(Debug, Default)]
794 pub struct ThinLTOImports {
795 // key = llvm name of importing module, value = list of modules it imports from
796 imports: FxHashMap<String, Vec<String>>,
799 impl ThinLTOImports {
800 fn modules_imported_by(&self, llvm_module_name: &str) -> &[String] {
801 self.imports.get(llvm_module_name).map(|v| &v[..]).unwrap_or(&[])
804 /// Load the ThinLTO import map from ThinLTOData.
805 unsafe fn from_thin_lto_data(data: *const llvm::ThinLTOData) -> ThinLTOImports {
806 unsafe extern "C" fn imported_module_callback(payload: *mut libc::c_void,
807 importing_module_name: *const libc::c_char,
808 imported_module_name: *const libc::c_char) {
809 let map = &mut* (payload as *mut ThinLTOImports);
810 let importing_module_name = CStr::from_ptr(importing_module_name);
811 let importing_module_name = module_name_to_str(&importing_module_name);
812 let imported_module_name = CStr::from_ptr(imported_module_name);
813 let imported_module_name = module_name_to_str(&imported_module_name);
815 if !map.imports.contains_key(importing_module_name) {
816 map.imports.insert(importing_module_name.to_owned(), vec![]);
820 .get_mut(importing_module_name)
822 .push(imported_module_name.to_owned());
824 let mut map = ThinLTOImports::default();
825 llvm::LLVMRustGetThinLTOModuleImports(data,
826 imported_module_callback,
827 &mut map as *mut _ as *mut libc::c_void);
832 fn module_name_to_str(c_str: &CStr) -> &str {
833 c_str.to_str().unwrap_or_else(|e|
834 bug!("Encountered non-utf8 LLVM module name `{}`: {}", c_str.to_string_lossy(), e))