Auto merge of #53356 - michaelwoerister:itlto, r=alexcrichton

Preliminary work for incremental ThinLTO (CGU name edition)

Bring back the first half of #52266 but hopefully without the performance regression.

Author: bors

src/librustc/mir/mono.rs

@@ -8,15 +8,16 @@
 // option. This file may not be copied, modified, or distributed
 // except according to those terms.
 
-use hir::def_id::DefId;
+use hir::def_id::{DefId, CrateNum};
 use syntax::ast::NodeId;
-use syntax::symbol::InternedString;
+use syntax::symbol::{Symbol, InternedString};
 use ty::{Instance, TyCtxt};
 use util::nodemap::FxHashMap;
 use rustc_data_structures::base_n;
 use rustc_data_structures::stable_hasher::{HashStable, StableHasherResult,
                                            StableHasher};
 use ich::{Fingerprint, StableHashingContext, NodeIdHashingMode};
+use std::fmt;
 use std::hash::Hash;
 
 #[derive(PartialEq, Eq, Clone, Copy, Debug, Hash)]
@@ -239,3 +240,95 @@ impl Stats {
         self.fn_stats.extend(stats.fn_stats);
     }
 }
+
+pub struct CodegenUnitNameBuilder<'a, 'gcx: 'tcx, 'tcx: 'a> {
+    tcx: TyCtxt<'a, 'gcx, 'tcx>,
+    cache: FxHashMap<CrateNum, String>,
+}
+
+impl<'a, 'gcx: 'tcx, 'tcx: 'a> CodegenUnitNameBuilder<'a, 'gcx, 'tcx> {
+
+    pub fn new(tcx: TyCtxt<'a, 'gcx, 'tcx>) -> Self {
+        CodegenUnitNameBuilder {
+            tcx,
+            cache: FxHashMap(),
+        }
+    }
+
+    /// CGU names should fulfill the following requirements:
+    /// - They should be able to act as a file name on any kind of file system
+    /// - They should not collide with other CGU names, even for different versions
+    ///   of the same crate.
+    ///
+    /// Consequently, we don't use special characters except for '.' and '-' and we
+    /// prefix each name with the crate-name and crate-disambiguator.
+    ///
+    /// This function will build CGU names of the form:
+    ///
+    /// ```
+    /// <crate-name>.<crate-disambiguator>(-<component>)*[.<special-suffix>]
+    /// ```
+    ///
+    /// The '.' before `<special-suffix>` makes sure that names with a special
+    /// suffix can never collide with a name built out of regular Rust
+    /// identifiers (e.g. module paths).
+    pub fn build_cgu_name<I, C, S>(&mut self,
+                                   cnum: CrateNum,
+                                   components: I,
+                                   special_suffix: Option<S>)
+                                   -> InternedString
+        where I: IntoIterator<Item=C>,
+              C: fmt::Display,
+              S: fmt::Display,
+    {
+        let cgu_name = self.build_cgu_name_no_mangle(cnum,
+                                                     components,
+                                                     special_suffix);
+
+        if self.tcx.sess.opts.debugging_opts.human_readable_cgu_names {
+            cgu_name
+        } else {
+            let cgu_name = &cgu_name.as_str()[..];
+            Symbol::intern(&CodegenUnit::mangle_name(cgu_name)).as_interned_str()
+        }
+    }
+
+    /// Same as `CodegenUnit::build_cgu_name()` but will never mangle the
+    /// resulting name.
+    pub fn build_cgu_name_no_mangle<I, C, S>(&mut self,
+                                             cnum: CrateNum,
+                                             components: I,
+                                             special_suffix: Option<S>)
+                                             -> InternedString
+        where I: IntoIterator<Item=C>,
+              C: fmt::Display,
+              S: fmt::Display,
+    {
+        use std::fmt::Write;
+
+        let mut cgu_name = String::with_capacity(64);
+
+        // Start out with the crate name and disambiguator
+        let tcx = self.tcx;
+        let crate_prefix = self.cache.entry(cnum).or_insert_with(|| {
+            let crate_disambiguator = format!("{}", tcx.crate_disambiguator(cnum));
+            // Using a shortened disambiguator of about 40 bits
+            format!("{}.{}", tcx.crate_name(cnum), &crate_disambiguator[0 .. 8])
+        });
+
+        write!(cgu_name, "{}", crate_prefix).unwrap();
+
+        // Add the components
+        for component in components {
+            write!(cgu_name, "-{}", component).unwrap();
+        }
+
+        if let Some(special_suffix) = special_suffix {
+            // We add a dot in here so it cannot clash with anything in a regular
+            // Rust identifier
+            write!(cgu_name, ".{}", special_suffix).unwrap();
+        }
+
+        Symbol::intern(&cgu_name[..]).as_interned_str()
+    }
+}

src/librustc/session/mod.rs

@@ -24,6 +24,7 @@ use util::nodemap::{FxHashMap, FxHashSet};
 use util::common::{duration_to_secs_str, ErrorReported};
 use util::common::ProfileQueriesMsg;
 
+use rustc_data_structures::base_n;
 use rustc_data_structures::sync::{self, Lrc, Lock, LockCell, OneThread, Once, RwLock};
 
 use syntax::ast::NodeId;
@@ -48,6 +49,7 @@ use std;
 use std::cell::{self, Cell, RefCell};
 use std::collections::HashMap;
 use std::env;
+use std::fmt;
 use std::io::Write;
 use std::path::{Path, PathBuf};
 use std::time::Duration;
@@ -1221,6 +1223,14 @@ impl CrateDisambiguator {
     }
 }
 
+impl fmt::Display for CrateDisambiguator {
+    fn fmt(&self, f: &mut fmt::Formatter) -> Result<(), fmt::Error> {
+        let (a, b) = self.0.as_value();
+        let as_u128 = a as u128 | ((b as u128) << 64);
+        f.write_str(&base_n::encode(as_u128, base_n::CASE_INSENSITIVE))
+    }
+}
+
 impl From<Fingerprint> for CrateDisambiguator {
     fn from(fingerprint: Fingerprint) -> CrateDisambiguator {
         CrateDisambiguator(fingerprint)

src/librustc_codegen_llvm/back/link.rs

@@ -46,13 +46,6 @@ use std::process::{Output, Stdio};
 use std::str;
 use syntax::attr;
 
-/// The LLVM module name containing crate-metadata. This includes a `.` on
-/// purpose, so it cannot clash with the name of a user-defined module.
-pub const METADATA_MODULE_NAME: &'static str = "crate.metadata";
-
-// same as for metadata above, but for allocator shim
-pub const ALLOCATOR_MODULE_NAME: &'static str = "crate.allocator";
-
 pub use rustc_codegen_utils::link::{find_crate_name, filename_for_input, default_output_for_target,
                                   invalid_output_for_target, build_link_meta, out_filename,
                                   check_file_is_writeable};

src/librustc_codegen_llvm/back/lto.rs

@@ -242,7 +242,7 @@ fn fat_lto(cgcx: &CodegenContext,
             let llvm = module.llvm().expect("can't lto pre-codegened modules");
             (&llvm.llcx, llvm.llmod())
         };
-        info!("using {:?} as a base module", module.llmod_id);
+        info!("using {:?} as a base module", module.name);
 
         // The linking steps below may produce errors and diagnostics within LLVM
         // which we'd like to handle and print, so set up our diagnostic handlers
@@ -257,7 +257,7 @@ fn fat_lto(cgcx: &CodegenContext,
         for module in modules {
             let llvm = module.llvm().expect("can't lto pre-codegened modules");
             let buffer = ModuleBuffer::new(llvm.llmod());
-            let llmod_id = CString::new(&module.llmod_id[..]).unwrap();
+            let llmod_id = CString::new(&module.name[..]).unwrap();
             serialized_modules.push((SerializedModule::Local(buffer), llmod_id));
         }
 
@@ -384,9 +384,9 @@ fn thin_lto(diag_handler: &Handler,
         //        the most expensive portion of this small bit of global
         //        analysis!
         for (i, module) in modules.iter().enumerate() {
-            info!("local module: {} - {}", i, module.llmod_id);
+            info!("local module: {} - {}", i, module.name);
             let llvm = module.llvm().expect("can't lto precodegened module");
-            let name = CString::new(module.llmod_id.clone()).unwrap();
+            let name = CString::new(module.name.clone()).unwrap();
             let buffer = ThinBuffer::new(llvm.llmod());
             thin_modules.push(llvm::ThinLTOModule {
                 identifier: name.as_ptr(),
@@ -395,7 +395,7 @@ fn thin_lto(diag_handler: &Handler,
             });
             thin_buffers.push(buffer);
             module_names.push(name);
-            timeline.record(&module.llmod_id);
+            timeline.record(&module.name);
         }
 
         // FIXME: All upstream crates are deserialized internally in the
@@ -668,7 +668,6 @@ impl ThinModule {
                 llcx,
                 tm,
             }),
-            llmod_id: self.name().to_string(),
             name: self.name().to_string(),
             kind: ModuleKind::Regular,
         };

src/librustc_codegen_llvm/back/write.rs

@@ -728,7 +728,7 @@ unsafe fn codegen(cgcx: &CodegenContext,
 
             if config.emit_bc_compressed {
                 let dst = bc_out.with_extension(RLIB_BYTECODE_EXTENSION);
-                let data = bytecode::encode(&module.llmod_id, data);
+                let data = bytecode::encode(&module.name, data);
                 if let Err(e) = fs::write(&dst, data) {
                     diag_handler.err(&format!("failed to write bytecode: {}", e));
                 }
@@ -1338,7 +1338,6 @@ fn execute_work_item(cgcx: &CodegenContext,
         assert_eq!(bytecode_compressed.is_some(), config.emit_bc_compressed);
 
         Ok(WorkItemResult::Compiled(CompiledModule {
-            llmod_id: module.llmod_id.clone(),
             name: module_name,
             kind: ModuleKind::Regular,
             pre_existing: true,

src/librustc_codegen_llvm/base.rs

@@ -36,7 +36,7 @@ use metadata;
 use rustc::hir::def_id::{CrateNum, DefId, LOCAL_CRATE};
 use rustc::middle::lang_items::StartFnLangItem;
 use rustc::middle::weak_lang_items;
-use rustc::mir::mono::{Linkage, Visibility, Stats};
+use rustc::mir::mono::{Linkage, Visibility, Stats, CodegenUnitNameBuilder};
 use rustc::middle::cstore::{EncodedMetadata};
 use rustc::ty::{self, Ty, TyCtxt};
 use rustc::ty::layout::{self, Align, TyLayout, LayoutOf};
@@ -742,19 +742,23 @@ pub fn codegen_crate<'a, 'tcx>(tcx: TyCtxt<'a, 'tcx, 'tcx>,
 
     let crate_hash = tcx.crate_hash(LOCAL_CRATE);
     let link_meta = link::build_link_meta(crate_hash);
+    let cgu_name_builder = &mut CodegenUnitNameBuilder::new(tcx);
 
     // Codegen the metadata.
     tcx.sess.profiler(|p| p.start_activity(ProfileCategory::Codegen));
-    let llmod_id = "metadata";
-    let metadata_llvm_module = ModuleLlvm::new(tcx.sess, llmod_id);
+
+    let metadata_cgu_name = cgu_name_builder.build_cgu_name(LOCAL_CRATE,
+                                                            &["crate"],
+                                                            Some("metadata")).as_str()
+                                                                             .to_string();
+    let metadata_llvm_module = ModuleLlvm::new(tcx.sess, &metadata_cgu_name);
     let metadata = time(tcx.sess, "write metadata", || {
         write_metadata(tcx, &metadata_llvm_module, &link_meta)
     });
     tcx.sess.profiler(|p| p.end_activity(ProfileCategory::Codegen));
 
     let metadata_module = ModuleCodegen {
-        name: link::METADATA_MODULE_NAME.to_string(),
-        llmod_id: llmod_id.to_string(),
+        name: metadata_cgu_name,
         source: ModuleSource::Codegened(metadata_llvm_module),
         kind: ModuleKind::Metadata,
     };
@@ -833,20 +837,22 @@ pub fn codegen_crate<'a, 'tcx>(tcx: TyCtxt<'a, 'tcx, 'tcx>,
     let allocator_module = if any_dynamic_crate {
         None
     } else if let Some(kind) = *tcx.sess.allocator_kind.get() {
-        unsafe {
-            let llmod_id = "allocator";
-            let modules = ModuleLlvm::new(tcx.sess, llmod_id);
-            time(tcx.sess, "write allocator module", || {
+        let llmod_id = cgu_name_builder.build_cgu_name(LOCAL_CRATE,
+                                                       &["crate"],
+                                                       Some("allocator")).as_str()
+                                                                         .to_string();
+        let modules = ModuleLlvm::new(tcx.sess, &llmod_id);
+        time(tcx.sess, "write allocator module", || {
+            unsafe {
                 allocator::codegen(tcx, &modules, kind)
-            });
+            }
+        });
 
-            Some(ModuleCodegen {
-                name: link::ALLOCATOR_MODULE_NAME.to_string(),
-                llmod_id: llmod_id.to_string(),
-                source: ModuleSource::Codegened(modules),
-                kind: ModuleKind::Allocator,
-            })
-        }
+        Some(ModuleCodegen {
+            name: llmod_id,
+            source: ModuleSource::Codegened(modules),
+            kind: ModuleKind::Allocator,
+        })
     } else {
         None
     };
@@ -889,21 +895,10 @@ pub fn codegen_crate<'a, 'tcx>(tcx: TyCtxt<'a, 'tcx, 'tcx>,
                 // succeed it means that none of the dependencies has changed
                 // and we can safely re-use.
                 if let Some(dep_node_index) = tcx.dep_graph.try_mark_green(tcx, dep_node) {
-                    // Append ".rs" to LLVM module identifier.
-                    //
-                    // LLVM code generator emits a ".file filename" directive
-                    // for ELF backends. Value of the "filename" is set as the
-                    // LLVM module identifier.  Due to a LLVM MC bug[1], LLVM
-                    // crashes if the module identifier is same as other symbols
-                    // such as a function name in the module.
-                    // 1. http://llvm.org/bugs/show_bug.cgi?id=11479
-                    let llmod_id = format!("{}.rs", cgu.name());
-
                     let module = ModuleCodegen {
                         name: cgu.name().to_string(),
                         source: ModuleSource::Preexisting(buf),
                         kind: ModuleKind::Regular,
-                        llmod_id,
                     };
                     tcx.dep_graph.mark_loaded_from_cache(dep_node_index, true);
                     write::submit_codegened_module_to_llvm(tcx, module, 0);
@@ -1212,21 +1207,8 @@ fn compile_codegen_unit<'a, 'tcx>(tcx: TyCtxt<'a, 'tcx, 'tcx>,
     {
         let cgu_name = cgu.name().to_string();
 
-        // Append ".rs" to LLVM module identifier.
-        //
-        // LLVM code generator emits a ".file filename" directive
-        // for ELF backends. Value of the "filename" is set as the
-        // LLVM module identifier.  Due to a LLVM MC bug[1], LLVM
-        // crashes if the module identifier is same as other symbols
-        // such as a function name in the module.
-        // 1. http://llvm.org/bugs/show_bug.cgi?id=11479
-        let llmod_id = format!("{}-{}.rs",
-                               cgu.name(),
-                               tcx.crate_disambiguator(LOCAL_CRATE)
-                                   .to_fingerprint().to_hex());
-
         // Instantiate monomorphizations without filling out definitions yet...
-        let llvm_module = ModuleLlvm::new(tcx.sess, &llmod_id);
+        let llvm_module = ModuleLlvm::new(tcx.sess, &cgu_name);
         let stats = {
             let cx = CodegenCx::new(tcx, cgu, &llvm_module);
             let mono_items = cx.codegen_unit
@@ -1282,7 +1264,6 @@ fn compile_codegen_unit<'a, 'tcx>(tcx: TyCtxt<'a, 'tcx, 'tcx>,
             name: cgu_name,
             source: ModuleSource::Codegened(llvm_module),
             kind: ModuleKind::Regular,
-            llmod_id,
         })
     }
 }

src/librustc_codegen_llvm/lib.rs

@@ -269,8 +269,8 @@ struct ModuleCodegen {
     /// unique amongst **all** crates.  Therefore, it should contain
     /// something unique to this crate (e.g., a module path) as well
     /// as the crate name and disambiguator.
+    /// We currently generate these names via CodegenUnit::build_cgu_name().
     name: String,
-    llmod_id: String,
     source: ModuleSource,
     kind: ModuleKind,
 }
@@ -317,7 +317,6 @@ impl ModuleCodegen {
         };
 
         CompiledModule {
-            llmod_id: self.llmod_id,
             name: self.name.clone(),
             kind: self.kind,
             pre_existing,
@@ -331,7 +330,6 @@ impl ModuleCodegen {
 #[derive(Debug)]
 struct CompiledModule {
     name: String,
-    llmod_id: String,
     kind: ModuleKind,
     pre_existing: bool,
     object: Option<PathBuf>,