1 use crate::back::bytecode::DecodedBytecode;
2 use crate::back::write::{self, DiagnosticHandlers, with_llvm_pmb, save_temp_bitcode,
4 use crate::llvm::archive_ro::ArchiveRO;
5 use crate::llvm::{self, True, False};
6 use crate::{ModuleLlvm, LlvmCodegenBackend};
7 use rustc_codegen_ssa::back::symbol_export;
8 use rustc_codegen_ssa::back::write::{ModuleConfig, CodegenContext, FatLTOInput};
9 use rustc_codegen_ssa::back::lto::{SerializedModule, LtoModuleCodegen, ThinShared, ThinModule};
10 use rustc_codegen_ssa::traits::*;
11 use errors::{FatalError, Handler};
12 use rustc::dep_graph::WorkProduct;
13 use rustc::dep_graph::cgu_reuse_tracker::CguReuse;
14 use rustc::hir::def_id::LOCAL_CRATE;
15 use rustc::middle::exported_symbols::SymbolExportLevel;
16 use rustc::session::config::{self, Lto};
17 use rustc::util::common::time_ext;
18 use rustc_data_structures::fx::FxHashMap;
19 use rustc_codegen_ssa::{RLIB_BYTECODE_EXTENSION, ModuleCodegen, ModuleKind};
21 use std::ffi::{CStr, CString};
26 pub fn crate_type_allows_lto(crate_type: config::CrateType) -> bool {
28 config::CrateType::Executable |
29 config::CrateType::Staticlib |
30 config::CrateType::Cdylib => true,
32 config::CrateType::Dylib |
33 config::CrateType::Rlib |
34 config::CrateType::ProcMacro => false,
38 fn prepare_lto(cgcx: &CodegenContext<LlvmCodegenBackend>,
39 diag_handler: &Handler)
40 -> Result<(Vec<CString>, Vec<(SerializedModule<ModuleBuffer>, CString)>), FatalError>
42 let export_threshold = match cgcx.lto {
43 // We're just doing LTO for our one crate
44 Lto::ThinLocal => SymbolExportLevel::Rust,
46 // We're doing LTO for the entire crate graph
47 Lto::Fat | Lto::Thin => {
48 symbol_export::crates_export_threshold(&cgcx.crate_types)
51 Lto::No => panic!("didn't request LTO but we're doing LTO"),
54 let symbol_filter = &|&(ref name, level): &(String, SymbolExportLevel)| {
55 if level.is_below_threshold(export_threshold) {
56 Some(CString::new(name.as_str()).unwrap())
61 let exported_symbols = cgcx.exported_symbols
62 .as_ref().expect("needs exported symbols for LTO");
63 let mut symbol_white_list = {
64 let _timer = cgcx.prof.generic_activity("LLVM_lto_generate_symbol_white_list");
65 exported_symbols[&LOCAL_CRATE]
67 .filter_map(symbol_filter)
68 .collect::<Vec<CString>>()
70 info!("{} symbols to preserve in this crate", symbol_white_list.len());
72 // If we're performing LTO for the entire crate graph, then for each of our
73 // upstream dependencies, find the corresponding rlib and load the bitcode
76 // We save off all the bytecode and LLVM module ids for later processing
77 // with either fat or thin LTO
78 let mut upstream_modules = Vec::new();
79 if cgcx.lto != Lto::ThinLocal {
80 if cgcx.opts.cg.prefer_dynamic {
81 diag_handler.struct_err("cannot prefer dynamic linking when performing LTO")
82 .note("only 'staticlib', 'bin', and 'cdylib' outputs are \
85 return Err(FatalError)
88 // Make sure we actually can run LTO
89 for crate_type in cgcx.crate_types.iter() {
90 if !crate_type_allows_lto(*crate_type) {
91 let e = diag_handler.fatal("lto can only be run for executables, cdylibs and \
92 static library outputs");
97 for &(cnum, ref path) in cgcx.each_linked_rlib_for_lto.iter() {
98 let exported_symbols = cgcx.exported_symbols
99 .as_ref().expect("needs exported symbols for LTO");
101 let _timer = cgcx.prof.generic_activity("LLVM_lto_generate_symbol_white_list");
102 symbol_white_list.extend(
103 exported_symbols[&cnum]
105 .filter_map(symbol_filter));
108 let _timer = cgcx.prof.generic_activity("LLVM_lto_load_upstream_bitcode");
109 let archive = ArchiveRO::open(&path).expect("wanted an rlib");
110 let bytecodes = archive.iter().filter_map(|child| {
111 child.ok().and_then(|c| c.name().map(|name| (name, c)))
112 }).filter(|&(name, _)| name.ends_with(RLIB_BYTECODE_EXTENSION));
113 for (name, data) in bytecodes {
114 info!("adding bytecode {}", name);
115 let bc_encoded = data.data();
117 let (bc, id) = time_ext(cgcx.time_passes, &format!("decode {}", name), || {
118 match DecodedBytecode::new(bc_encoded) {
119 Ok(b) => Ok((b.bytecode(), b.identifier().to_string())),
120 Err(e) => Err(diag_handler.fatal(&e)),
123 let bc = SerializedModule::FromRlib(bc);
124 upstream_modules.push((bc, CString::new(id).unwrap()));
129 Ok((symbol_white_list, upstream_modules))
132 /// Performs fat LTO by merging all modules into a single one and returning it
133 /// for further optimization.
134 pub(crate) fn run_fat(cgcx: &CodegenContext<LlvmCodegenBackend>,
135 modules: Vec<FatLTOInput<LlvmCodegenBackend>>,
136 cached_modules: Vec<(SerializedModule<ModuleBuffer>, WorkProduct)>)
137 -> Result<LtoModuleCodegen<LlvmCodegenBackend>, FatalError>
139 let diag_handler = cgcx.create_diag_handler();
140 let (symbol_white_list, upstream_modules) = prepare_lto(cgcx, &diag_handler)?;
141 let symbol_white_list = symbol_white_list.iter()
143 .collect::<Vec<_>>();
154 /// Performs thin LTO by performing necessary global analysis and returning two
155 /// lists, one of the modules that need optimization and another for modules that
156 /// can simply be copied over from the incr. comp. cache.
157 pub(crate) fn run_thin(cgcx: &CodegenContext<LlvmCodegenBackend>,
158 modules: Vec<(String, ThinBuffer)>,
159 cached_modules: Vec<(SerializedModule<ModuleBuffer>, WorkProduct)>)
160 -> Result<(Vec<LtoModuleCodegen<LlvmCodegenBackend>>, Vec<WorkProduct>), FatalError>
162 let diag_handler = cgcx.create_diag_handler();
163 let (symbol_white_list, upstream_modules) = prepare_lto(cgcx, &diag_handler)?;
164 let symbol_white_list = symbol_white_list.iter()
166 .collect::<Vec<_>>();
167 if cgcx.opts.cg.linker_plugin_lto.enabled() {
168 unreachable!("We should never reach this case if the LTO step \
169 is deferred to the linker");
179 pub(crate) fn prepare_thin(
180 module: ModuleCodegen<ModuleLlvm>
181 ) -> (String, ThinBuffer) {
182 let name = module.name.clone();
183 let buffer = ThinBuffer::new(module.module_llvm.llmod());
187 fn fat_lto(cgcx: &CodegenContext<LlvmCodegenBackend>,
188 diag_handler: &Handler,
189 modules: Vec<FatLTOInput<LlvmCodegenBackend>>,
190 cached_modules: Vec<(SerializedModule<ModuleBuffer>, WorkProduct)>,
191 mut serialized_modules: Vec<(SerializedModule<ModuleBuffer>, CString)>,
192 symbol_white_list: &[*const libc::c_char])
193 -> Result<LtoModuleCodegen<LlvmCodegenBackend>, FatalError>
195 let _timer = cgcx.prof.generic_activity("LLVM_fat_lto_build_monolithic_module");
196 info!("going for a fat lto");
198 // Sort out all our lists of incoming modules into two lists.
200 // * `serialized_modules` (also and argument to this function) contains all
201 // modules that are serialized in-memory.
202 // * `in_memory` contains modules which are already parsed and in-memory,
203 // such as from multi-CGU builds.
205 // All of `cached_modules` (cached from previous incremental builds) can
206 // immediately go onto the `serialized_modules` modules list and then we can
207 // split the `modules` array into these two lists.
208 let mut in_memory = Vec::new();
209 serialized_modules.extend(cached_modules.into_iter().map(|(buffer, wp)| {
210 info!("pushing cached module {:?}", wp.cgu_name);
211 (buffer, CString::new(wp.cgu_name).unwrap())
213 for module in modules {
215 FatLTOInput::InMemory(m) => in_memory.push(m),
216 FatLTOInput::Serialized { name, buffer } => {
217 info!("pushing serialized module {:?}", name);
218 let buffer = SerializedModule::Local(buffer);
219 serialized_modules.push((buffer, CString::new(name).unwrap()));
224 // Find the "costliest" module and merge everything into that codegen unit.
225 // All the other modules will be serialized and reparsed into the new
226 // context, so this hopefully avoids serializing and parsing the largest
229 // Additionally use a regular module as the base here to ensure that various
230 // file copy operations in the backend work correctly. The only other kind
231 // of module here should be an allocator one, and if your crate is smaller
232 // than the allocator module then the size doesn't really matter anyway.
233 let costliest_module = in_memory.iter()
235 .filter(|&(_, module)| module.kind == ModuleKind::Regular)
238 llvm::LLVMRustModuleCost(module.module_llvm.llmod())
244 // If we found a costliest module, we're good to go. Otherwise all our
245 // inputs were serialized which could happen in the case, for example, that
246 // all our inputs were incrementally reread from the cache and we're just
247 // re-executing the LTO passes. If that's the case deserialize the first
248 // module and create a linker with it.
249 let module: ModuleCodegen<ModuleLlvm> = match costliest_module {
250 Some((_cost, i)) => in_memory.remove(i),
252 assert!(serialized_modules.len() > 0, "must have at least one serialized module");
253 let (buffer, name) = serialized_modules.remove(0);
254 info!("no in-memory regular modules to choose from, parsing {:?}", name);
256 module_llvm: ModuleLlvm::parse(cgcx, &name, buffer.data(), diag_handler)?,
257 name: name.into_string().unwrap(),
258 kind: ModuleKind::Regular,
262 let mut serialized_bitcode = Vec::new();
264 let (llcx, llmod) = {
265 let llvm = &module.module_llvm;
266 (&llvm.llcx, llvm.llmod())
268 info!("using {:?} as a base module", module.name);
270 // The linking steps below may produce errors and diagnostics within LLVM
271 // which we'd like to handle and print, so set up our diagnostic handlers
272 // (which get unregistered when they go out of scope below).
273 let _handler = DiagnosticHandlers::new(cgcx, diag_handler, llcx);
275 // For all other modules we codegened we'll need to link them into our own
276 // bitcode. All modules were codegened in their own LLVM context, however,
277 // and we want to move everything to the same LLVM context. Currently the
278 // way we know of to do that is to serialize them to a string and them parse
279 // them later. Not great but hey, that's why it's "fat" LTO, right?
280 for module in in_memory {
281 let buffer = ModuleBuffer::new(module.module_llvm.llmod());
282 let llmod_id = CString::new(&module.name[..]).unwrap();
283 serialized_modules.push((SerializedModule::Local(buffer), llmod_id));
285 // Sort the modules to ensure we produce deterministic results.
286 serialized_modules.sort_by(|module1, module2| module1.1.cmp(&module2.1));
288 // For all serialized bitcode files we parse them and link them in as we did
289 // above, this is all mostly handled in C++. Like above, though, we don't
290 // know much about the memory management here so we err on the side of being
291 // save and persist everything with the original module.
292 let mut linker = Linker::new(llmod);
293 for (bc_decoded, name) in serialized_modules {
294 let _timer = cgcx.prof.generic_activity("LLVM_fat_lto_link_module");
295 info!("linking {:?}", name);
296 time_ext(cgcx.time_passes, &format!("ll link {:?}", name), || {
297 let data = bc_decoded.data();
298 linker.add(&data).map_err(|()| {
299 let msg = format!("failed to load bc of {:?}", name);
300 write::llvm_err(&diag_handler, &msg)
303 serialized_bitcode.push(bc_decoded);
306 save_temp_bitcode(&cgcx, &module, "lto.input");
308 // Internalize everything that *isn't* in our whitelist to help strip out
309 // more modules and such
311 let ptr = symbol_white_list.as_ptr();
312 llvm::LLVMRustRunRestrictionPass(llmod,
313 ptr as *const *const libc::c_char,
314 symbol_white_list.len() as libc::size_t);
315 save_temp_bitcode(&cgcx, &module, "lto.after-restriction");
318 if cgcx.no_landing_pads {
320 llvm::LLVMRustMarkAllFunctionsNounwind(llmod);
322 save_temp_bitcode(&cgcx, &module, "lto.after-nounwind");
326 Ok(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<LlvmCodegenBackend>,
389 diag_handler: &Handler,
390 modules: Vec<(String, ThinBuffer)>,
391 serialized_modules: Vec<(SerializedModule<ModuleBuffer>, CString)>,
392 cached_modules: Vec<(SerializedModule<ModuleBuffer>, WorkProduct)>,
393 symbol_white_list: &[*const libc::c_char])
394 -> Result<(Vec<LtoModuleCodegen<LlvmCodegenBackend>>, Vec<WorkProduct>), FatalError>
396 let _timer = cgcx.prof.generic_activity("LLVM_thin_lto_global_analysis");
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 full_scope_len = modules.len() + serialized_modules.len() + cached_modules.len();
406 let mut thin_buffers = Vec::with_capacity(modules.len());
407 let mut module_names = Vec::with_capacity(full_scope_len);
408 let mut thin_modules = Vec::with_capacity(full_scope_len);
410 for (i, (name, buffer)) in modules.into_iter().enumerate() {
411 info!("local module: {} - {}", i, name);
412 let cname = CString::new(name.clone()).unwrap();
413 thin_modules.push(llvm::ThinLTOModule {
414 identifier: cname.as_ptr(),
415 data: buffer.data().as_ptr(),
416 len: buffer.data().len(),
418 thin_buffers.push(buffer);
419 module_names.push(cname);
422 // FIXME: All upstream crates are deserialized internally in the
423 // function below to extract their summary and modules. Note that
424 // unlike the loop above we *must* decode and/or read something
425 // here as these are all just serialized files on disk. An
426 // improvement, however, to make here would be to store the
427 // module summary separately from the actual module itself. Right
428 // now this is store in one large bitcode file, and the entire
429 // file is deflate-compressed. We could try to bypass some of the
430 // decompression by storing the index uncompressed and only
431 // lazily decompressing the bytecode if necessary.
433 // Note that truly taking advantage of this optimization will
434 // likely be further down the road. We'd have to implement
435 // incremental ThinLTO first where we could actually avoid
436 // looking at upstream modules entirely sometimes (the contents,
437 // we must always unconditionally look at the index).
438 let mut serialized = Vec::with_capacity(serialized_modules.len() + cached_modules.len());
440 let cached_modules = cached_modules.into_iter().map(|(sm, wp)| {
441 (sm, CString::new(wp.cgu_name).unwrap())
444 for (module, name) in serialized_modules.into_iter().chain(cached_modules) {
445 info!("upstream or cached module {:?}", name);
446 thin_modules.push(llvm::ThinLTOModule {
447 identifier: name.as_ptr(),
448 data: module.data().as_ptr(),
449 len: module.data().len(),
451 serialized.push(module);
452 module_names.push(name);
456 assert_eq!(thin_modules.len(), module_names.len());
458 // Delegate to the C++ bindings to create some data here. Once this is a
459 // tried-and-true interface we may wish to try to upstream some of this
460 // to LLVM itself, right now we reimplement a lot of what they do
462 let data = llvm::LLVMRustCreateThinLTOData(
463 thin_modules.as_ptr(),
464 thin_modules.len() as u32,
465 symbol_white_list.as_ptr(),
466 symbol_white_list.len() as u32,
468 write::llvm_err(&diag_handler, "failed to prepare thin LTO context")
471 info!("thin LTO data created");
473 let import_map = if cgcx.incr_comp_session_dir.is_some() {
474 ThinLTOImports::from_thin_lto_data(data)
476 // If we don't compile incrementally, we don't need to load the
477 // import data from LLVM.
478 assert!(green_modules.is_empty());
479 ThinLTOImports::default()
481 info!("thin LTO import map loaded");
483 let data = ThinData(data);
485 // Throw our data in an `Arc` as we'll be sharing it across threads. We
486 // also put all memory referenced by the C++ data (buffers, ids, etc)
487 // into the arc as well. After this we'll create a thin module
488 // codegen per module in this data.
489 let shared = Arc::new(ThinShared {
492 serialized_modules: serialized,
496 let mut copy_jobs = vec![];
497 let mut opt_jobs = vec![];
499 info!("checking which modules can be-reused and which have to be re-optimized.");
500 for (module_index, module_name) in shared.module_names.iter().enumerate() {
501 let module_name = module_name_to_str(module_name);
503 // If the module hasn't changed and none of the modules it imports
504 // from has changed, we can re-use the post-ThinLTO version of the
506 if green_modules.contains_key(module_name) {
507 let imports_all_green = import_map.modules_imported_by(module_name)
509 .all(|imported_module| green_modules.contains_key(imported_module));
511 if imports_all_green {
512 let work_product = green_modules[module_name].clone();
513 copy_jobs.push(work_product);
514 info!(" - {}: re-used", module_name);
515 cgcx.cgu_reuse_tracker.set_actual_reuse(module_name,
521 info!(" - {}: re-compiled", module_name);
522 opt_jobs.push(LtoModuleCodegen::Thin(ThinModule {
523 shared: shared.clone(),
528 Ok((opt_jobs, copy_jobs))
532 pub(crate) fn run_pass_manager(cgcx: &CodegenContext<LlvmCodegenBackend>,
533 module: &ModuleCodegen<ModuleLlvm>,
534 config: &ModuleConfig,
536 // Now we have one massive module inside of llmod. Time to run the
537 // LTO-specific optimization passes that LLVM provides.
539 // This code is based off the code found in llvm's LTO code generator:
540 // tools/lto/LTOCodeGenerator.cpp
541 debug!("running the pass manager");
543 let pm = llvm::LLVMCreatePassManager();
544 llvm::LLVMRustAddAnalysisPasses(module.module_llvm.tm, pm, module.module_llvm.llmod());
546 if config.verify_llvm_ir {
547 let pass = llvm::LLVMRustFindAndCreatePass("verify\0".as_ptr().cast());
548 llvm::LLVMRustAddPass(pm, pass.unwrap());
551 // When optimizing for LTO we don't actually pass in `-O0`, but we force
552 // it to always happen at least with `-O1`.
554 // With ThinLTO we mess around a lot with symbol visibility in a way
555 // that will actually cause linking failures if we optimize at O0 which
556 // notable is lacking in dead code elimination. To ensure we at least
557 // get some optimizations and correctly link we forcibly switch to `-O1`
558 // to get dead code elimination.
560 // Note that in general this shouldn't matter too much as you typically
561 // only turn on ThinLTO when you're compiling with optimizations
563 let opt_level = config.opt_level.map(|x| to_llvm_opt_settings(x).0)
564 .unwrap_or(llvm::CodeGenOptLevel::None);
565 let opt_level = match opt_level {
566 llvm::CodeGenOptLevel::None => llvm::CodeGenOptLevel::Less,
569 with_llvm_pmb(module.module_llvm.llmod(), config, opt_level, false, &mut |b| {
571 llvm::LLVMRustPassManagerBuilderPopulateThinLTOPassManager(b, pm);
573 llvm::LLVMPassManagerBuilderPopulateLTOPassManager(b, pm,
574 /* Internalize = */ False,
575 /* RunInliner = */ True);
579 // We always generate bitcode through ThinLTOBuffers,
580 // which do not support anonymous globals
581 if config.bitcode_needed() {
582 let pass = llvm::LLVMRustFindAndCreatePass("name-anon-globals\0".as_ptr().cast());
583 llvm::LLVMRustAddPass(pm, pass.unwrap());
586 if config.verify_llvm_ir {
587 let pass = llvm::LLVMRustFindAndCreatePass("verify\0".as_ptr().cast());
588 llvm::LLVMRustAddPass(pm, pass.unwrap());
591 time_ext(cgcx.time_passes, "LTO passes", ||
592 llvm::LLVMRunPassManager(pm, module.module_llvm.llmod()));
594 llvm::LLVMDisposePassManager(pm);
599 pub struct ModuleBuffer(&'static mut llvm::ModuleBuffer);
601 unsafe impl Send for ModuleBuffer {}
602 unsafe impl Sync for ModuleBuffer {}
605 pub fn new(m: &llvm::Module) -> ModuleBuffer {
606 ModuleBuffer(unsafe {
607 llvm::LLVMRustModuleBufferCreate(m)
612 impl ModuleBufferMethods for ModuleBuffer {
613 fn data(&self) -> &[u8] {
615 let ptr = llvm::LLVMRustModuleBufferPtr(self.0);
616 let len = llvm::LLVMRustModuleBufferLen(self.0);
617 slice::from_raw_parts(ptr, len)
622 impl Drop for ModuleBuffer {
624 unsafe { llvm::LLVMRustModuleBufferFree(&mut *(self.0 as *mut _)); }
628 pub struct ThinData(&'static mut llvm::ThinLTOData);
630 unsafe impl Send for ThinData {}
631 unsafe impl Sync for ThinData {}
633 impl Drop for ThinData {
636 llvm::LLVMRustFreeThinLTOData(&mut *(self.0 as *mut _));
641 pub struct ThinBuffer(&'static mut llvm::ThinLTOBuffer);
643 unsafe impl Send for ThinBuffer {}
644 unsafe impl Sync for ThinBuffer {}
647 pub fn new(m: &llvm::Module) -> ThinBuffer {
649 let buffer = llvm::LLVMRustThinLTOBufferCreate(m);
655 impl ThinBufferMethods for ThinBuffer {
656 fn data(&self) -> &[u8] {
658 let ptr = llvm::LLVMRustThinLTOBufferPtr(self.0) as *const _;
659 let len = llvm::LLVMRustThinLTOBufferLen(self.0);
660 slice::from_raw_parts(ptr, len)
665 impl Drop for ThinBuffer {
668 llvm::LLVMRustThinLTOBufferFree(&mut *(self.0 as *mut _));
673 pub unsafe fn optimize_thin_module(
674 thin_module: &mut ThinModule<LlvmCodegenBackend>,
675 cgcx: &CodegenContext<LlvmCodegenBackend>,
676 ) -> Result<ModuleCodegen<ModuleLlvm>, FatalError> {
677 let diag_handler = cgcx.create_diag_handler();
678 let tm = (cgcx.tm_factory.0)().map_err(|e| {
679 write::llvm_err(&diag_handler, &e)
682 // Right now the implementation we've got only works over serialized
683 // modules, so we create a fresh new LLVM context and parse the module
684 // into that context. One day, however, we may do this for upstream
685 // crates but for locally codegened modules we may be able to reuse
686 // that LLVM Context and Module.
687 let llcx = llvm::LLVMRustContextCreate(cgcx.fewer_names);
688 let llmod_raw = parse_module(
690 &thin_module.shared.module_names[thin_module.idx],
694 let module = ModuleCodegen {
695 module_llvm: ModuleLlvm {
700 name: thin_module.name().to_string(),
701 kind: ModuleKind::Regular,
704 let llmod = module.module_llvm.llmod();
705 save_temp_bitcode(&cgcx, &module, "thin-lto-input");
707 // Before we do much else find the "main" `DICompileUnit` that we'll be
708 // using below. If we find more than one though then rustc has changed
709 // in a way we're not ready for, so generate an ICE by returning
711 let mut cu1 = ptr::null_mut();
712 let mut cu2 = ptr::null_mut();
713 llvm::LLVMRustThinLTOGetDICompileUnit(llmod, &mut cu1, &mut cu2);
715 let msg = "multiple source DICompileUnits found";
716 return Err(write::llvm_err(&diag_handler, msg))
719 // Like with "fat" LTO, get some better optimizations if landing pads
720 // are disabled by removing all landing pads.
721 if cgcx.no_landing_pads {
722 let _timer = cgcx.prof.generic_activity("LLVM_thin_lto_remove_landing_pads");
723 llvm::LLVMRustMarkAllFunctionsNounwind(llmod);
724 save_temp_bitcode(&cgcx, &module, "thin-lto-after-nounwind");
727 // Up next comes the per-module local analyses that we do for Thin LTO.
728 // Each of these functions is basically copied from the LLVM
729 // implementation and then tailored to suit this implementation. Ideally
730 // each of these would be supported by upstream LLVM but that's perhaps
731 // a patch for another day!
733 // You can find some more comments about these functions in the LLVM
734 // bindings we've got (currently `PassWrapper.cpp`)
736 let _timer = cgcx.prof.generic_activity("LLVM_thin_lto_rename");
737 if !llvm::LLVMRustPrepareThinLTORename(thin_module.shared.data.0, llmod) {
738 let msg = "failed to prepare thin LTO module";
739 return Err(write::llvm_err(&diag_handler, msg))
741 save_temp_bitcode(cgcx, &module, "thin-lto-after-rename");
745 let _timer = cgcx.prof.generic_activity("LLVM_thin_lto_resolve_weak");
746 if !llvm::LLVMRustPrepareThinLTOResolveWeak(thin_module.shared.data.0, llmod) {
747 let msg = "failed to prepare thin LTO module";
748 return Err(write::llvm_err(&diag_handler, msg))
750 save_temp_bitcode(cgcx, &module, "thin-lto-after-resolve");
754 let _timer = cgcx.prof.generic_activity("LLVM_thin_lto_internalize");
755 if !llvm::LLVMRustPrepareThinLTOInternalize(thin_module.shared.data.0, llmod) {
756 let msg = "failed to prepare thin LTO module";
757 return Err(write::llvm_err(&diag_handler, msg))
759 save_temp_bitcode(cgcx, &module, "thin-lto-after-internalize");
763 let _timer = cgcx.prof.generic_activity("LLVM_thin_lto_import");
764 if !llvm::LLVMRustPrepareThinLTOImport(thin_module.shared.data.0, llmod) {
765 let msg = "failed to prepare thin LTO module";
766 return Err(write::llvm_err(&diag_handler, msg))
768 save_temp_bitcode(cgcx, &module, "thin-lto-after-import");
771 // Ok now this is a bit unfortunate. This is also something you won't
772 // find upstream in LLVM's ThinLTO passes! This is a hack for now to
773 // work around bugs in LLVM.
775 // First discovered in #45511 it was found that as part of ThinLTO
776 // importing passes LLVM will import `DICompileUnit` metadata
777 // information across modules. This means that we'll be working with one
778 // LLVM module that has multiple `DICompileUnit` instances in it (a
779 // bunch of `llvm.dbg.cu` members). Unfortunately there's a number of
780 // bugs in LLVM's backend which generates invalid DWARF in a situation
783 // https://bugs.llvm.org/show_bug.cgi?id=35212
784 // https://bugs.llvm.org/show_bug.cgi?id=35562
786 // While the first bug there is fixed the second ended up causing #46346
787 // which was basically a resurgence of #45511 after LLVM's bug 35212 was
790 // This function below is a huge hack around this problem. The function
791 // below is defined in `PassWrapper.cpp` and will basically "merge"
792 // all `DICompileUnit` instances in a module. Basically it'll take all
793 // the objects, rewrite all pointers of `DISubprogram` to point to the
794 // first `DICompileUnit`, and then delete all the other units.
796 // This is probably mangling to the debug info slightly (but hopefully
797 // not too much) but for now at least gets LLVM to emit valid DWARF (or
798 // so it appears). Hopefully we can remove this once upstream bugs are
801 let _timer = cgcx.prof.generic_activity("LLVM_thin_lto_patch_debuginfo");
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 let _timer = cgcx.prof.generic_activity("LLVM_thin_lto_optimize");
813 info!("running thin lto passes over {}", module.name);
814 let config = cgcx.config(module.kind);
815 run_pass_manager(cgcx, &module, config, true);
816 save_temp_bitcode(cgcx, &module, "thin-lto-after-pm");
822 #[derive(Debug, Default)]
823 pub struct ThinLTOImports {
824 // key = llvm name of importing module, value = list of modules it imports from
825 imports: FxHashMap<String, Vec<String>>,
828 impl ThinLTOImports {
829 fn modules_imported_by(&self, llvm_module_name: &str) -> &[String] {
830 self.imports.get(llvm_module_name).map(|v| &v[..]).unwrap_or(&[])
833 /// Loads the ThinLTO import map from ThinLTOData.
834 unsafe fn from_thin_lto_data(data: *const llvm::ThinLTOData) -> ThinLTOImports {
835 unsafe extern "C" fn imported_module_callback(payload: *mut libc::c_void,
836 importing_module_name: *const libc::c_char,
837 imported_module_name: *const libc::c_char) {
838 let map = &mut* (payload as *mut ThinLTOImports);
839 let importing_module_name = CStr::from_ptr(importing_module_name);
840 let importing_module_name = module_name_to_str(&importing_module_name);
841 let imported_module_name = CStr::from_ptr(imported_module_name);
842 let imported_module_name = module_name_to_str(&imported_module_name);
844 if !map.imports.contains_key(importing_module_name) {
845 map.imports.insert(importing_module_name.to_owned(), vec![]);
849 .get_mut(importing_module_name)
851 .push(imported_module_name.to_owned());
853 let mut map = ThinLTOImports::default();
854 llvm::LLVMRustGetThinLTOModuleImports(data,
855 imported_module_callback,
856 &mut map as *mut _ as *mut libc::c_void);
861 fn module_name_to_str(c_str: &CStr) -> &str {
862 c_str.to_str().unwrap_or_else(|e|
863 bug!("Encountered non-utf8 LLVM module name `{}`: {}", c_str.to_string_lossy(), e))
866 pub fn parse_module<'a>(
867 cx: &'a llvm::Context,
870 diag_handler: &Handler,
871 ) -> Result<&'a llvm::Module, FatalError> {
873 llvm::LLVMRustParseBitcodeForLTO(
879 let msg = "failed to parse bitcode for LTO module";
880 write::llvm_err(&diag_handler, msg)