rustc_codegen_llvm/back/
lto.rs

1use std::collections::BTreeMap;
2use std::ffi::{CStr, CString};
3use std::fs::File;
4use std::path::{Path, PathBuf};
5use std::ptr::NonNull;
6use std::sync::Arc;
7use std::{io, iter, slice};
8
9use object::read::archive::ArchiveFile;
10use object::{Object, ObjectSection};
11use rustc_codegen_ssa::back::lto::{SerializedModule, ThinModule, ThinShared};
12use rustc_codegen_ssa::back::write::{CodegenContext, FatLtoInput};
13use rustc_codegen_ssa::traits::*;
14use rustc_codegen_ssa::{ModuleCodegen, ModuleKind, looks_like_rust_object_file};
15use rustc_data_structures::fx::FxHashMap;
16use rustc_data_structures::memmap::Mmap;
17use rustc_errors::DiagCtxtHandle;
18use rustc_hir::attrs::SanitizerSet;
19use rustc_middle::bug;
20use rustc_middle::dep_graph::WorkProduct;
21use rustc_session::config::{self, Lto};
22use tracing::{debug, info};
23
24use crate::back::write::{
25    self, CodegenDiagnosticsStage, DiagnosticHandlers, bitcode_section_name, save_temp_bitcode,
26};
27use crate::errors::{LlvmError, LtoBitcodeFromRlib};
28use crate::llvm::{self, build_string};
29use crate::{LlvmCodegenBackend, ModuleLlvm, SimpleCx};
30
31/// We keep track of the computed LTO cache keys from the previous
32/// session to determine which CGUs we can reuse.
33const THIN_LTO_KEYS_INCR_COMP_FILE_NAME: &str = "thin-lto-past-keys.bin";
34
35fn prepare_lto(
36    cgcx: &CodegenContext<LlvmCodegenBackend>,
37    exported_symbols_for_lto: &[String],
38    each_linked_rlib_for_lto: &[PathBuf],
39    dcx: DiagCtxtHandle<'_>,
40) -> (Vec<CString>, Vec<(SerializedModule<ModuleBuffer>, CString)>) {
41    let mut symbols_below_threshold = exported_symbols_for_lto
42        .iter()
43        .map(|symbol| CString::new(symbol.to_owned()).unwrap())
44        .collect::<Vec<CString>>();
45
46    if cgcx.module_config.instrument_coverage || cgcx.module_config.pgo_gen.enabled() {
47        // These are weak symbols that point to the profile version and the
48        // profile name, which need to be treated as exported so LTO doesn't nix
49        // them.
50        const PROFILER_WEAK_SYMBOLS: [&CStr; 2] =
51            [c"__llvm_profile_raw_version", c"__llvm_profile_filename"];
52
53        symbols_below_threshold.extend(PROFILER_WEAK_SYMBOLS.iter().map(|&sym| sym.to_owned()));
54    }
55
56    if cgcx.module_config.sanitizer.contains(SanitizerSet::MEMORY) {
57        let mut msan_weak_symbols = Vec::new();
58
59        // Similar to profiling, preserve weak msan symbol during LTO.
60        if cgcx.module_config.sanitizer_recover.contains(SanitizerSet::MEMORY) {
61            msan_weak_symbols.push(c"__msan_keep_going");
62        }
63
64        if cgcx.module_config.sanitizer_memory_track_origins != 0 {
65            msan_weak_symbols.push(c"__msan_track_origins");
66        }
67
68        symbols_below_threshold.extend(msan_weak_symbols.into_iter().map(|sym| sym.to_owned()));
69    }
70
71    // Preserve LLVM-injected, ASAN-related symbols.
72    // See also https://github.com/rust-lang/rust/issues/113404.
73    symbols_below_threshold.push(c"___asan_globals_registered".to_owned());
74
75    // __llvm_profile_counter_bias is pulled in at link time by an undefined reference to
76    // __llvm_profile_runtime, therefore we won't know until link time if this symbol
77    // should have default visibility.
78    symbols_below_threshold.push(c"__llvm_profile_counter_bias".to_owned());
79
80    // If we're performing LTO for the entire crate graph, then for each of our
81    // upstream dependencies, find the corresponding rlib and load the bitcode
82    // from the archive.
83    //
84    // We save off all the bytecode and LLVM module ids for later processing
85    // with either fat or thin LTO
86    let mut upstream_modules = Vec::new();
87    if cgcx.lto != Lto::ThinLocal {
88        for path in each_linked_rlib_for_lto {
89            let archive_data = unsafe {
90                Mmap::map(std::fs::File::open(&path).expect("couldn't open rlib"))
91                    .expect("couldn't map rlib")
92            };
93            let archive = ArchiveFile::parse(&*archive_data).expect("wanted an rlib");
94            let obj_files = archive
95                .members()
96                .filter_map(|child| {
97                    child.ok().and_then(|c| {
98                        std::str::from_utf8(c.name()).ok().map(|name| (name.trim(), c))
99                    })
100                })
101                .filter(|&(name, _)| looks_like_rust_object_file(name));
102            for (name, child) in obj_files {
103                info!("adding bitcode from {}", name);
104                match get_bitcode_slice_from_object_data(
105                    child.data(&*archive_data).expect("corrupt rlib"),
106                    cgcx,
107                ) {
108                    Ok(data) => {
109                        let module = SerializedModule::FromRlib(data.to_vec());
110                        upstream_modules.push((module, CString::new(name).unwrap()));
111                    }
112                    Err(e) => dcx.emit_fatal(e),
113                }
114            }
115        }
116    }
117
118    (symbols_below_threshold, upstream_modules)
119}
120
121fn get_bitcode_slice_from_object_data<'a>(
122    obj: &'a [u8],
123    cgcx: &CodegenContext<LlvmCodegenBackend>,
124) -> Result<&'a [u8], LtoBitcodeFromRlib> {
125    // We're about to assume the data here is an object file with sections, but if it's raw LLVM IR
126    // that won't work. Fortunately, if that's what we have we can just return the object directly,
127    // so we sniff the relevant magic strings here and return.
128    if obj.starts_with(b"\xDE\xC0\x17\x0B") || obj.starts_with(b"BC\xC0\xDE") {
129        return Ok(obj);
130    }
131    // We drop the "__LLVM," prefix here because on Apple platforms there's a notion of "segment
132    // name" which in the public API for sections gets treated as part of the section name, but
133    // internally in MachOObjectFile.cpp gets treated separately.
134    let section_name = bitcode_section_name(cgcx).to_str().unwrap().trim_start_matches("__LLVM,");
135
136    let obj =
137        object::File::parse(obj).map_err(|err| LtoBitcodeFromRlib { err: err.to_string() })?;
138
139    let section = obj
140        .section_by_name(section_name)
141        .ok_or_else(|| LtoBitcodeFromRlib { err: format!("Can't find section {section_name}") })?;
142
143    section.data().map_err(|err| LtoBitcodeFromRlib { err: err.to_string() })
144}
145
146/// Performs fat LTO by merging all modules into a single one and returning it
147/// for further optimization.
148pub(crate) fn run_fat(
149    cgcx: &CodegenContext<LlvmCodegenBackend>,
150    exported_symbols_for_lto: &[String],
151    each_linked_rlib_for_lto: &[PathBuf],
152    modules: Vec<FatLtoInput<LlvmCodegenBackend>>,
153) -> ModuleCodegen<ModuleLlvm> {
154    let dcx = cgcx.create_dcx();
155    let dcx = dcx.handle();
156    let (symbols_below_threshold, upstream_modules) =
157        prepare_lto(cgcx, exported_symbols_for_lto, each_linked_rlib_for_lto, dcx);
158    let symbols_below_threshold =
159        symbols_below_threshold.iter().map(|c| c.as_ptr()).collect::<Vec<_>>();
160    fat_lto(cgcx, dcx, modules, upstream_modules, &symbols_below_threshold)
161}
162
163/// Performs thin LTO by performing necessary global analysis and returning two
164/// lists, one of the modules that need optimization and another for modules that
165/// can simply be copied over from the incr. comp. cache.
166pub(crate) fn run_thin(
167    cgcx: &CodegenContext<LlvmCodegenBackend>,
168    exported_symbols_for_lto: &[String],
169    each_linked_rlib_for_lto: &[PathBuf],
170    modules: Vec<(String, ThinBuffer)>,
171    cached_modules: Vec<(SerializedModule<ModuleBuffer>, WorkProduct)>,
172) -> (Vec<ThinModule<LlvmCodegenBackend>>, Vec<WorkProduct>) {
173    let dcx = cgcx.create_dcx();
174    let dcx = dcx.handle();
175    let (symbols_below_threshold, upstream_modules) =
176        prepare_lto(cgcx, exported_symbols_for_lto, each_linked_rlib_for_lto, dcx);
177    let symbols_below_threshold =
178        symbols_below_threshold.iter().map(|c| c.as_ptr()).collect::<Vec<_>>();
179    if cgcx.opts.cg.linker_plugin_lto.enabled() {
180        unreachable!(
181            "We should never reach this case if the LTO step \
182                      is deferred to the linker"
183        );
184    }
185    thin_lto(cgcx, dcx, modules, upstream_modules, cached_modules, &symbols_below_threshold)
186}
187
188pub(crate) fn prepare_thin(
189    module: ModuleCodegen<ModuleLlvm>,
190    emit_summary: bool,
191) -> (String, ThinBuffer) {
192    let name = module.name;
193    let buffer = ThinBuffer::new(module.module_llvm.llmod(), true, emit_summary);
194    (name, buffer)
195}
196
197fn fat_lto(
198    cgcx: &CodegenContext<LlvmCodegenBackend>,
199    dcx: DiagCtxtHandle<'_>,
200    modules: Vec<FatLtoInput<LlvmCodegenBackend>>,
201    mut serialized_modules: Vec<(SerializedModule<ModuleBuffer>, CString)>,
202    symbols_below_threshold: &[*const libc::c_char],
203) -> ModuleCodegen<ModuleLlvm> {
204    let _timer = cgcx.prof.generic_activity("LLVM_fat_lto_build_monolithic_module");
205    info!("going for a fat lto");
206
207    // Sort out all our lists of incoming modules into two lists.
208    //
209    // * `serialized_modules` (also and argument to this function) contains all
210    //   modules that are serialized in-memory.
211    // * `in_memory` contains modules which are already parsed and in-memory,
212    //   such as from multi-CGU builds.
213    let mut in_memory = Vec::new();
214    for module in modules {
215        match module {
216            FatLtoInput::InMemory(m) => in_memory.push(m),
217            FatLtoInput::Serialized { name, buffer } => {
218                info!("pushing serialized module {:?}", name);
219                serialized_modules.push((buffer, CString::new(name).unwrap()));
220            }
221        }
222    }
223
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
227    // codegen unit.
228    //
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
234        .iter()
235        .enumerate()
236        .filter(|&(_, module)| module.kind == ModuleKind::Regular)
237        .map(|(i, module)| {
238            let cost = unsafe { llvm::LLVMRustModuleCost(module.module_llvm.llmod()) };
239            (cost, i)
240        })
241        .max();
242
243    // If we found a costliest module, we're good to go. Otherwise all our
244    // inputs were serialized which could happen in the case, for example, that
245    // all our inputs were incrementally reread from the cache and we're just
246    // re-executing the LTO passes. If that's the case deserialize the first
247    // module and create a linker with it.
248    let module: ModuleCodegen<ModuleLlvm> = match costliest_module {
249        Some((_cost, i)) => in_memory.remove(i),
250        None => {
251            assert!(!serialized_modules.is_empty(), "must have at least one serialized module");
252            let (buffer, name) = serialized_modules.remove(0);
253            info!("no in-memory regular modules to choose from, parsing {:?}", name);
254            let llvm_module = ModuleLlvm::parse(cgcx, &name, buffer.data(), dcx);
255            ModuleCodegen::new_regular(name.into_string().unwrap(), llvm_module)
256        }
257    };
258    {
259        let (llcx, llmod) = {
260            let llvm = &module.module_llvm;
261            (&llvm.llcx, llvm.llmod())
262        };
263        info!("using {:?} as a base module", module.name);
264
265        // The linking steps below may produce errors and diagnostics within LLVM
266        // which we'd like to handle and print, so set up our diagnostic handlers
267        // (which get unregistered when they go out of scope below).
268        let _handler =
269            DiagnosticHandlers::new(cgcx, dcx, llcx, &module, CodegenDiagnosticsStage::LTO);
270
271        // For all other modules we codegened we'll need to link them into our own
272        // bitcode. All modules were codegened in their own LLVM context, however,
273        // and we want to move everything to the same LLVM context. Currently the
274        // way we know of to do that is to serialize them to a string and them parse
275        // them later. Not great but hey, that's why it's "fat" LTO, right?
276        for module in in_memory {
277            let buffer = ModuleBuffer::new(module.module_llvm.llmod());
278            let llmod_id = CString::new(&module.name[..]).unwrap();
279            serialized_modules.push((SerializedModule::Local(buffer), llmod_id));
280        }
281        // Sort the modules to ensure we produce deterministic results.
282        serialized_modules.sort_by(|module1, module2| module1.1.cmp(&module2.1));
283
284        // For all serialized bitcode files we parse them and link them in as we did
285        // above, this is all mostly handled in C++.
286        let mut linker = Linker::new(llmod);
287        for (bc_decoded, name) in serialized_modules {
288            let _timer = cgcx
289                .prof
290                .generic_activity_with_arg_recorder("LLVM_fat_lto_link_module", |recorder| {
291                    recorder.record_arg(format!("{name:?}"))
292                });
293            info!("linking {:?}", name);
294            let data = bc_decoded.data();
295            linker
296                .add(data)
297                .unwrap_or_else(|()| write::llvm_err(dcx, LlvmError::LoadBitcode { name }));
298        }
299        drop(linker);
300        save_temp_bitcode(cgcx, &module, "lto.input");
301
302        // Internalize everything below threshold to help strip out more modules and such.
303        unsafe {
304            let ptr = symbols_below_threshold.as_ptr();
305            llvm::LLVMRustRunRestrictionPass(
306                llmod,
307                ptr as *const *const libc::c_char,
308                symbols_below_threshold.len() as libc::size_t,
309            );
310        }
311        save_temp_bitcode(cgcx, &module, "lto.after-restriction");
312    }
313
314    module
315}
316
317pub(crate) struct Linker<'a>(&'a mut llvm::Linker<'a>);
318
319impl<'a> Linker<'a> {
320    pub(crate) fn new(llmod: &'a llvm::Module) -> Self {
321        unsafe { Linker(llvm::LLVMRustLinkerNew(llmod)) }
322    }
323
324    pub(crate) fn add(&mut self, bytecode: &[u8]) -> Result<(), ()> {
325        unsafe {
326            if llvm::LLVMRustLinkerAdd(
327                self.0,
328                bytecode.as_ptr() as *const libc::c_char,
329                bytecode.len(),
330            ) {
331                Ok(())
332            } else {
333                Err(())
334            }
335        }
336    }
337}
338
339impl Drop for Linker<'_> {
340    fn drop(&mut self) {
341        unsafe {
342            llvm::LLVMRustLinkerFree(&mut *(self.0 as *mut _));
343        }
344    }
345}
346
347/// Prepare "thin" LTO to get run on these modules.
348///
349/// The general structure of ThinLTO is quite different from the structure of
350/// "fat" LTO above. With "fat" LTO all LLVM modules in question are merged into
351/// one giant LLVM module, and then we run more optimization passes over this
352/// big module after internalizing most symbols. Thin LTO, on the other hand,
353/// avoid this large bottleneck through more targeted optimization.
354///
355/// At a high level Thin LTO looks like:
356///
357///    1. Prepare a "summary" of each LLVM module in question which describes
358///       the values inside, cost of the values, etc.
359///    2. Merge the summaries of all modules in question into one "index"
360///    3. Perform some global analysis on this index
361///    4. For each module, use the index and analysis calculated previously to
362///       perform local transformations on the module, for example inlining
363///       small functions from other modules.
364///    5. Run thin-specific optimization passes over each module, and then code
365///       generate everything at the end.
366///
367/// The summary for each module is intended to be quite cheap, and the global
368/// index is relatively quite cheap to create as well. As a result, the goal of
369/// ThinLTO is to reduce the bottleneck on LTO and enable LTO to be used in more
370/// situations. For example one cheap optimization is that we can parallelize
371/// all codegen modules, easily making use of all the cores on a machine.
372///
373/// With all that in mind, the function here is designed at specifically just
374/// calculating the *index* for ThinLTO. This index will then be shared amongst
375/// all of the `LtoModuleCodegen` units returned below and destroyed once
376/// they all go out of scope.
377fn thin_lto(
378    cgcx: &CodegenContext<LlvmCodegenBackend>,
379    dcx: DiagCtxtHandle<'_>,
380    modules: Vec<(String, ThinBuffer)>,
381    serialized_modules: Vec<(SerializedModule<ModuleBuffer>, CString)>,
382    cached_modules: Vec<(SerializedModule<ModuleBuffer>, WorkProduct)>,
383    symbols_below_threshold: &[*const libc::c_char],
384) -> (Vec<ThinModule<LlvmCodegenBackend>>, Vec<WorkProduct>) {
385    let _timer = cgcx.prof.generic_activity("LLVM_thin_lto_global_analysis");
386    unsafe {
387        info!("going for that thin, thin LTO");
388
389        let green_modules: FxHashMap<_, _> =
390            cached_modules.iter().map(|(_, wp)| (wp.cgu_name.clone(), wp.clone())).collect();
391
392        let full_scope_len = modules.len() + serialized_modules.len() + cached_modules.len();
393        let mut thin_buffers = Vec::with_capacity(modules.len());
394        let mut module_names = Vec::with_capacity(full_scope_len);
395        let mut thin_modules = Vec::with_capacity(full_scope_len);
396
397        for (i, (name, buffer)) in modules.into_iter().enumerate() {
398            info!("local module: {} - {}", i, name);
399            let cname = CString::new(name.as_bytes()).unwrap();
400            thin_modules.push(llvm::ThinLTOModule {
401                identifier: cname.as_ptr(),
402                data: buffer.data().as_ptr(),
403                len: buffer.data().len(),
404            });
405            thin_buffers.push(buffer);
406            module_names.push(cname);
407        }
408
409        // FIXME: All upstream crates are deserialized internally in the
410        //        function below to extract their summary and modules. Note that
411        //        unlike the loop above we *must* decode and/or read something
412        //        here as these are all just serialized files on disk. An
413        //        improvement, however, to make here would be to store the
414        //        module summary separately from the actual module itself. Right
415        //        now this is store in one large bitcode file, and the entire
416        //        file is deflate-compressed. We could try to bypass some of the
417        //        decompression by storing the index uncompressed and only
418        //        lazily decompressing the bytecode if necessary.
419        //
420        //        Note that truly taking advantage of this optimization will
421        //        likely be further down the road. We'd have to implement
422        //        incremental ThinLTO first where we could actually avoid
423        //        looking at upstream modules entirely sometimes (the contents,
424        //        we must always unconditionally look at the index).
425        let mut serialized = Vec::with_capacity(serialized_modules.len() + cached_modules.len());
426
427        let cached_modules =
428            cached_modules.into_iter().map(|(sm, wp)| (sm, CString::new(wp.cgu_name).unwrap()));
429
430        for (module, name) in serialized_modules.into_iter().chain(cached_modules) {
431            info!("upstream or cached module {:?}", name);
432            thin_modules.push(llvm::ThinLTOModule {
433                identifier: name.as_ptr(),
434                data: module.data().as_ptr(),
435                len: module.data().len(),
436            });
437            serialized.push(module);
438            module_names.push(name);
439        }
440
441        // Sanity check
442        assert_eq!(thin_modules.len(), module_names.len());
443
444        // Delegate to the C++ bindings to create some data here. Once this is a
445        // tried-and-true interface we may wish to try to upstream some of this
446        // to LLVM itself, right now we reimplement a lot of what they do
447        // upstream...
448        let data = llvm::LLVMRustCreateThinLTOData(
449            thin_modules.as_ptr(),
450            thin_modules.len(),
451            symbols_below_threshold.as_ptr(),
452            symbols_below_threshold.len(),
453        )
454        .unwrap_or_else(|| write::llvm_err(dcx, LlvmError::PrepareThinLtoContext));
455
456        let data = ThinData(data);
457
458        info!("thin LTO data created");
459
460        let (key_map_path, prev_key_map, curr_key_map) = if let Some(ref incr_comp_session_dir) =
461            cgcx.incr_comp_session_dir
462        {
463            let path = incr_comp_session_dir.join(THIN_LTO_KEYS_INCR_COMP_FILE_NAME);
464            // If the previous file was deleted, or we get an IO error
465            // reading the file, then we'll just use `None` as the
466            // prev_key_map, which will force the code to be recompiled.
467            let prev =
468                if path.exists() { ThinLTOKeysMap::load_from_file(&path).ok() } else { None };
469            let curr = ThinLTOKeysMap::from_thin_lto_modules(&data, &thin_modules, &module_names);
470            (Some(path), prev, curr)
471        } else {
472            // If we don't compile incrementally, we don't need to load the
473            // import data from LLVM.
474            assert!(green_modules.is_empty());
475            let curr = ThinLTOKeysMap::default();
476            (None, None, curr)
477        };
478        info!("thin LTO cache key map loaded");
479        info!("prev_key_map: {:#?}", prev_key_map);
480        info!("curr_key_map: {:#?}", curr_key_map);
481
482        // Throw our data in an `Arc` as we'll be sharing it across threads. We
483        // also put all memory referenced by the C++ data (buffers, ids, etc)
484        // into the arc as well. After this we'll create a thin module
485        // codegen per module in this data.
486        let shared = Arc::new(ThinShared {
487            data,
488            thin_buffers,
489            serialized_modules: serialized,
490            module_names,
491        });
492
493        let mut copy_jobs = vec![];
494        let mut opt_jobs = vec![];
495
496        info!("checking which modules can be-reused and which have to be re-optimized.");
497        for (module_index, module_name) in shared.module_names.iter().enumerate() {
498            let module_name = module_name_to_str(module_name);
499            if let (Some(prev_key_map), true) =
500                (prev_key_map.as_ref(), green_modules.contains_key(module_name))
501            {
502                assert!(cgcx.incr_comp_session_dir.is_some());
503
504                // If a module exists in both the current and the previous session,
505                // and has the same LTO cache key in both sessions, then we can re-use it
506                if prev_key_map.keys.get(module_name) == curr_key_map.keys.get(module_name) {
507                    let work_product = green_modules[module_name].clone();
508                    copy_jobs.push(work_product);
509                    info!(" - {}: re-used", module_name);
510                    assert!(cgcx.incr_comp_session_dir.is_some());
511                    continue;
512                }
513            }
514
515            info!(" - {}: re-compiled", module_name);
516            opt_jobs.push(ThinModule { shared: Arc::clone(&shared), idx: module_index });
517        }
518
519        // Save the current ThinLTO import information for the next compilation
520        // session, overwriting the previous serialized data (if any).
521        if let Some(path) = key_map_path
522            && let Err(err) = curr_key_map.save_to_file(&path)
523        {
524            write::llvm_err(dcx, LlvmError::WriteThinLtoKey { err });
525        }
526
527        (opt_jobs, copy_jobs)
528    }
529}
530
531fn enable_autodiff_settings(ad: &[config::AutoDiff]) {
532    for val in ad {
533        // We intentionally don't use a wildcard, to not forget handling anything new.
534        match val {
535            config::AutoDiff::PrintPerf => {
536                llvm::set_print_perf(true);
537            }
538            config::AutoDiff::PrintAA => {
539                llvm::set_print_activity(true);
540            }
541            config::AutoDiff::PrintTA => {
542                llvm::set_print_type(true);
543            }
544            config::AutoDiff::PrintTAFn(fun) => {
545                llvm::set_print_type(true); // Enable general type printing
546                llvm::set_print_type_fun(&fun); // Set specific function to analyze
547            }
548            config::AutoDiff::Inline => {
549                llvm::set_inline(true);
550            }
551            config::AutoDiff::LooseTypes => {
552                llvm::set_loose_types(true);
553            }
554            config::AutoDiff::PrintSteps => {
555                llvm::set_print(true);
556            }
557            // We handle this in the PassWrapper.cpp
558            config::AutoDiff::PrintPasses => {}
559            // We handle this in the PassWrapper.cpp
560            config::AutoDiff::PrintModBefore => {}
561            // We handle this in the PassWrapper.cpp
562            config::AutoDiff::PrintModAfter => {}
563            // We handle this in the PassWrapper.cpp
564            config::AutoDiff::PrintModFinal => {}
565            // This is required and already checked
566            config::AutoDiff::Enable => {}
567            // We handle this below
568            config::AutoDiff::NoPostopt => {}
569        }
570    }
571    // This helps with handling enums for now.
572    llvm::set_strict_aliasing(false);
573    // FIXME(ZuseZ4): Test this, since it was added a long time ago.
574    llvm::set_rust_rules(true);
575}
576
577pub(crate) fn run_pass_manager(
578    cgcx: &CodegenContext<LlvmCodegenBackend>,
579    dcx: DiagCtxtHandle<'_>,
580    module: &mut ModuleCodegen<ModuleLlvm>,
581    thin: bool,
582) {
583    let _timer = cgcx.prof.generic_activity_with_arg("LLVM_lto_optimize", &*module.name);
584    let config = &cgcx.module_config;
585
586    // Now we have one massive module inside of llmod. Time to run the
587    // LTO-specific optimization passes that LLVM provides.
588    //
589    // This code is based off the code found in llvm's LTO code generator:
590    //      llvm/lib/LTO/LTOCodeGenerator.cpp
591    debug!("running the pass manager");
592    let opt_stage = if thin { llvm::OptStage::ThinLTO } else { llvm::OptStage::FatLTO };
593    let opt_level = config.opt_level.unwrap_or(config::OptLevel::No);
594
595    // The PostAD behavior is the same that we would have if no autodiff was used.
596    // It will run the default optimization pipeline. If AD is enabled we select
597    // the DuringAD stage, which will disable vectorization and loop unrolling, and
598    // schedule two autodiff optimization + differentiation passes.
599    // We then run the llvm_optimize function a second time, to optimize the code which we generated
600    // in the enzyme differentiation pass.
601    let enable_ad = config.autodiff.contains(&config::AutoDiff::Enable);
602    let enable_gpu = config.offload.contains(&config::Offload::Enable);
603    let stage = if thin {
604        write::AutodiffStage::PreAD
605    } else {
606        if enable_ad { write::AutodiffStage::DuringAD } else { write::AutodiffStage::PostAD }
607    };
608
609    if enable_ad {
610        enable_autodiff_settings(&config.autodiff);
611    }
612
613    unsafe {
614        write::llvm_optimize(cgcx, dcx, module, None, config, opt_level, opt_stage, stage);
615    }
616
617    if enable_gpu && !thin {
618        let cx =
619            SimpleCx::new(module.module_llvm.llmod(), &module.module_llvm.llcx, cgcx.pointer_size);
620        crate::builder::gpu_offload::handle_gpu_code(cgcx, &cx);
621    }
622
623    if cfg!(llvm_enzyme) && enable_ad && !thin {
624        let opt_stage = llvm::OptStage::FatLTO;
625        let stage = write::AutodiffStage::PostAD;
626        if !config.autodiff.contains(&config::AutoDiff::NoPostopt) {
627            unsafe {
628                write::llvm_optimize(cgcx, dcx, module, None, config, opt_level, opt_stage, stage);
629            }
630        }
631
632        // This is the final IR, so people should be able to inspect the optimized autodiff output,
633        // for manual inspection.
634        if config.autodiff.contains(&config::AutoDiff::PrintModFinal) {
635            unsafe { llvm::LLVMDumpModule(module.module_llvm.llmod()) };
636        }
637    }
638
639    debug!("lto done");
640}
641
642pub struct ModuleBuffer(&'static mut llvm::ModuleBuffer);
643
644unsafe impl Send for ModuleBuffer {}
645unsafe impl Sync for ModuleBuffer {}
646
647impl ModuleBuffer {
648    pub(crate) fn new(m: &llvm::Module) -> ModuleBuffer {
649        ModuleBuffer(unsafe { llvm::LLVMRustModuleBufferCreate(m) })
650    }
651}
652
653impl ModuleBufferMethods for ModuleBuffer {
654    fn data(&self) -> &[u8] {
655        unsafe {
656            let ptr = llvm::LLVMRustModuleBufferPtr(self.0);
657            let len = llvm::LLVMRustModuleBufferLen(self.0);
658            slice::from_raw_parts(ptr, len)
659        }
660    }
661}
662
663impl Drop for ModuleBuffer {
664    fn drop(&mut self) {
665        unsafe {
666            llvm::LLVMRustModuleBufferFree(&mut *(self.0 as *mut _));
667        }
668    }
669}
670
671pub struct ThinData(&'static mut llvm::ThinLTOData);
672
673unsafe impl Send for ThinData {}
674unsafe impl Sync for ThinData {}
675
676impl Drop for ThinData {
677    fn drop(&mut self) {
678        unsafe {
679            llvm::LLVMRustFreeThinLTOData(&mut *(self.0 as *mut _));
680        }
681    }
682}
683
684pub struct ThinBuffer(&'static mut llvm::ThinLTOBuffer);
685
686unsafe impl Send for ThinBuffer {}
687unsafe impl Sync for ThinBuffer {}
688
689impl ThinBuffer {
690    pub(crate) fn new(m: &llvm::Module, is_thin: bool, emit_summary: bool) -> ThinBuffer {
691        unsafe {
692            let buffer = llvm::LLVMRustThinLTOBufferCreate(m, is_thin, emit_summary);
693            ThinBuffer(buffer)
694        }
695    }
696
697    pub(crate) unsafe fn from_raw_ptr(ptr: *mut llvm::ThinLTOBuffer) -> ThinBuffer {
698        let mut ptr = NonNull::new(ptr).unwrap();
699        ThinBuffer(unsafe { ptr.as_mut() })
700    }
701}
702
703impl ThinBufferMethods for ThinBuffer {
704    fn data(&self) -> &[u8] {
705        unsafe {
706            let ptr = llvm::LLVMRustThinLTOBufferPtr(self.0) as *const _;
707            let len = llvm::LLVMRustThinLTOBufferLen(self.0);
708            slice::from_raw_parts(ptr, len)
709        }
710    }
711
712    fn thin_link_data(&self) -> &[u8] {
713        unsafe {
714            let ptr = llvm::LLVMRustThinLTOBufferThinLinkDataPtr(self.0) as *const _;
715            let len = llvm::LLVMRustThinLTOBufferThinLinkDataLen(self.0);
716            slice::from_raw_parts(ptr, len)
717        }
718    }
719}
720
721impl Drop for ThinBuffer {
722    fn drop(&mut self) {
723        unsafe {
724            llvm::LLVMRustThinLTOBufferFree(&mut *(self.0 as *mut _));
725        }
726    }
727}
728
729pub(crate) fn optimize_thin_module(
730    thin_module: ThinModule<LlvmCodegenBackend>,
731    cgcx: &CodegenContext<LlvmCodegenBackend>,
732) -> ModuleCodegen<ModuleLlvm> {
733    let dcx = cgcx.create_dcx();
734    let dcx = dcx.handle();
735
736    let module_name = &thin_module.shared.module_names[thin_module.idx];
737
738    // Right now the implementation we've got only works over serialized
739    // modules, so we create a fresh new LLVM context and parse the module
740    // into that context. One day, however, we may do this for upstream
741    // crates but for locally codegened modules we may be able to reuse
742    // that LLVM Context and Module.
743    let module_llvm = ModuleLlvm::parse(cgcx, module_name, thin_module.data(), dcx);
744    let mut module = ModuleCodegen::new_regular(thin_module.name(), module_llvm);
745    // Given that the newly created module lacks a thinlto buffer for embedding, we need to re-add it here.
746    if cgcx.module_config.embed_bitcode() {
747        module.thin_lto_buffer = Some(thin_module.data().to_vec());
748    }
749    {
750        let target = &*module.module_llvm.tm;
751        let llmod = module.module_llvm.llmod();
752        save_temp_bitcode(cgcx, &module, "thin-lto-input");
753
754        // Up next comes the per-module local analyses that we do for Thin LTO.
755        // Each of these functions is basically copied from the LLVM
756        // implementation and then tailored to suit this implementation. Ideally
757        // each of these would be supported by upstream LLVM but that's perhaps
758        // a patch for another day!
759        //
760        // You can find some more comments about these functions in the LLVM
761        // bindings we've got (currently `PassWrapper.cpp`)
762        {
763            let _timer =
764                cgcx.prof.generic_activity_with_arg("LLVM_thin_lto_rename", thin_module.name());
765            unsafe {
766                llvm::LLVMRustPrepareThinLTORename(thin_module.shared.data.0, llmod, target.raw())
767            };
768            save_temp_bitcode(cgcx, &module, "thin-lto-after-rename");
769        }
770
771        {
772            let _timer = cgcx
773                .prof
774                .generic_activity_with_arg("LLVM_thin_lto_resolve_weak", thin_module.name());
775            if unsafe { !llvm::LLVMRustPrepareThinLTOResolveWeak(thin_module.shared.data.0, llmod) }
776            {
777                write::llvm_err(dcx, LlvmError::PrepareThinLtoModule);
778            }
779            save_temp_bitcode(cgcx, &module, "thin-lto-after-resolve");
780        }
781
782        {
783            let _timer = cgcx
784                .prof
785                .generic_activity_with_arg("LLVM_thin_lto_internalize", thin_module.name());
786            if unsafe { !llvm::LLVMRustPrepareThinLTOInternalize(thin_module.shared.data.0, llmod) }
787            {
788                write::llvm_err(dcx, LlvmError::PrepareThinLtoModule);
789            }
790            save_temp_bitcode(cgcx, &module, "thin-lto-after-internalize");
791        }
792
793        {
794            let _timer =
795                cgcx.prof.generic_activity_with_arg("LLVM_thin_lto_import", thin_module.name());
796            if unsafe {
797                !llvm::LLVMRustPrepareThinLTOImport(thin_module.shared.data.0, llmod, target.raw())
798            } {
799                write::llvm_err(dcx, LlvmError::PrepareThinLtoModule);
800            }
801            save_temp_bitcode(cgcx, &module, "thin-lto-after-import");
802        }
803
804        // Alright now that we've done everything related to the ThinLTO
805        // analysis it's time to run some optimizations! Here we use the same
806        // `run_pass_manager` as the "fat" LTO above except that we tell it to
807        // populate a thin-specific pass manager, which presumably LLVM treats a
808        // little differently.
809        {
810            info!("running thin lto passes over {}", module.name);
811            run_pass_manager(cgcx, dcx, &mut module, true);
812            save_temp_bitcode(cgcx, &module, "thin-lto-after-pm");
813        }
814    }
815    module
816}
817
818/// Maps LLVM module identifiers to their corresponding LLVM LTO cache keys
819#[derive(Debug, Default)]
820struct ThinLTOKeysMap {
821    // key = llvm name of importing module, value = LLVM cache key
822    keys: BTreeMap<String, String>,
823}
824
825impl ThinLTOKeysMap {
826    fn save_to_file(&self, path: &Path) -> io::Result<()> {
827        use std::io::Write;
828        let mut writer = File::create_buffered(path)?;
829        // The entries are loaded back into a hash map in `load_from_file()`, so
830        // the order in which we write them to file here does not matter.
831        for (module, key) in &self.keys {
832            writeln!(writer, "{module} {key}")?;
833        }
834        Ok(())
835    }
836
837    fn load_from_file(path: &Path) -> io::Result<Self> {
838        use std::io::BufRead;
839        let mut keys = BTreeMap::default();
840        let file = File::open_buffered(path)?;
841        for line in file.lines() {
842            let line = line?;
843            let mut split = line.split(' ');
844            let module = split.next().unwrap();
845            let key = split.next().unwrap();
846            assert_eq!(split.next(), None, "Expected two space-separated values, found {line:?}");
847            keys.insert(module.to_string(), key.to_string());
848        }
849        Ok(Self { keys })
850    }
851
852    fn from_thin_lto_modules(
853        data: &ThinData,
854        modules: &[llvm::ThinLTOModule],
855        names: &[CString],
856    ) -> Self {
857        let keys = iter::zip(modules, names)
858            .map(|(module, name)| {
859                let key = build_string(|rust_str| unsafe {
860                    llvm::LLVMRustComputeLTOCacheKey(rust_str, module.identifier, data.0);
861                })
862                .expect("Invalid ThinLTO module key");
863                (module_name_to_str(name).to_string(), key)
864            })
865            .collect();
866        Self { keys }
867    }
868}
869
870fn module_name_to_str(c_str: &CStr) -> &str {
871    c_str.to_str().unwrap_or_else(|e| {
872        bug!("Encountered non-utf8 LLVM module name `{}`: {}", c_str.to_string_lossy(), e)
873    })
874}
875
876pub(crate) fn parse_module<'a>(
877    cx: &'a llvm::Context,
878    name: &CStr,
879    data: &[u8],
880    dcx: DiagCtxtHandle<'_>,
881) -> &'a llvm::Module {
882    unsafe {
883        llvm::LLVMRustParseBitcodeForLTO(cx, data.as_ptr(), data.len(), name.as_ptr())
884            .unwrap_or_else(|| write::llvm_err(dcx, LlvmError::ParseBitcode))
885    }
886}
rustc_codegen_llvm/back/lto.rs

rustc_codegen_llvm/back/
lto.rs
