Merge pull request #278 from rust-lang/feature/lto_2023-05-12

Add support for Link-Time Optimization

7 files changed, 545 insertions, 37 deletions

diff --git a/src/back/lto.rs b/src/back/lto.rs
new file mode 100644
index 00000000000..529454b119e
--- /dev/null
+++ b/src/back/lto.rs
@@ -0,0 +1,341 @@
+/// GCC requires to use the same toolchain for the whole compilation when doing LTO.
+/// So, we need the same version/commit of the linker (gcc) and lto front-end binaries (lto1,
+/// lto-wrapper, liblto_plugin.so).
+
+// FIXME(antoyo): the executables compiled with LTO are bigger than those compiled without LTO.
+// Since it is the opposite for cg_llvm, check if this is normal.
+//
+// Maybe we embed the bitcode in the final binary?
+// It doesn't look like we try to generate fat objects for the final binary.
+// Check if the way we combine the object files make it keep the LTO sections on the final link.
+// Maybe that's because the combined object files contain the IR (true) and the final link
+// does not remove it?
+//
+// TODO(antoyo): for performance, check which optimizations the C++ frontend enables.
+//
+// Fix these warnings:
+// /usr/bin/ld: warning: type of symbol `_RNvNvNvNtCs5JWOrf9uCus_5rayon11thread_pool19WORKER_THREAD_STATE7___getit5___KEY' changed from 1 to 6 in /tmp/ccKeUSiR.ltrans0.ltrans.o
+// /usr/bin/ld: warning: type of symbol `_RNvNvNvNvNtNtNtCsAj5i4SGTR7_3std4sync4mpmc5waker17current_thread_id5DUMMY7___getit5___KEY' changed from 1 to 6 in /tmp/ccKeUSiR.ltrans0.ltrans.o
+// /usr/bin/ld: warning: incremental linking of LTO and non-LTO objects; using -flinker-output=nolto-rel which will bypass whole program optimization
+
+use std::ffi::CString;
+use std::fs::{self, File};
+use std::path::{Path, PathBuf};
+
+use gccjit::OutputKind;
+use object::read::archive::ArchiveFile;
+use rustc_codegen_ssa::back::lto::{LtoModuleCodegen, SerializedModule};
+use rustc_codegen_ssa::back::symbol_export;
+use rustc_codegen_ssa::back::write::{CodegenContext, FatLtoInput};
+use rustc_codegen_ssa::traits::*;
+use rustc_codegen_ssa::{looks_like_rust_object_file, ModuleCodegen, ModuleKind};
+use rustc_data_structures::memmap::Mmap;
+use rustc_errors::{FatalError, Handler};
+use rustc_hir::def_id::LOCAL_CRATE;
+use rustc_middle::dep_graph::WorkProduct;
+use rustc_middle::middle::exported_symbols::{SymbolExportInfo, SymbolExportLevel};
+use rustc_session::config::{CrateType, Lto};
+use tempfile::{TempDir, tempdir};
+
+use crate::back::write::save_temp_bitcode;
+use crate::errors::{
+    DynamicLinkingWithLTO, LtoBitcodeFromRlib, LtoDisallowed, LtoDylib,
+};
+use crate::{GccCodegenBackend, GccContext, to_gcc_opt_level};
+
+/// We keep track of the computed LTO cache keys from the previous
+/// session to determine which CGUs we can reuse.
+//pub const THIN_LTO_KEYS_INCR_COMP_FILE_NAME: &str = "thin-lto-past-keys.bin";
+
+pub fn crate_type_allows_lto(crate_type: CrateType) -> bool {
+    match crate_type {
+        CrateType::Executable | CrateType::Dylib | CrateType::Staticlib | CrateType::Cdylib => true,
+        CrateType::Rlib | CrateType::ProcMacro => false,
+    }
+}
+
+struct LtoData {
+    // TODO(antoyo): use symbols_below_threshold.
+    //symbols_below_threshold: Vec<CString>,
+    upstream_modules: Vec<(SerializedModule<ModuleBuffer>, CString)>,
+    tmp_path: TempDir,
+}
+
+fn prepare_lto(cgcx: &CodegenContext<GccCodegenBackend>, diag_handler: &Handler) -> Result<LtoData, FatalError> {
+    let export_threshold = match cgcx.lto {
+        // We're just doing LTO for our one crate
+        Lto::ThinLocal => SymbolExportLevel::Rust,
+
+        // We're doing LTO for the entire crate graph
+        Lto::Fat | Lto::Thin => symbol_export::crates_export_threshold(&cgcx.crate_types),
+
+        Lto::No => panic!("didn't request LTO but we're doing LTO"),
+    };
+
+    let tmp_path =
+        match tempdir() {
+            Ok(tmp_path) => tmp_path,
+            Err(error) => {
+                eprintln!("Cannot create temporary directory: {}", error);
+                return Err(FatalError);
+            },
+        };
+
+    let symbol_filter = &|&(ref name, info): &(String, SymbolExportInfo)| {
+        if info.level.is_below_threshold(export_threshold) || info.used {
+            Some(CString::new(name.as_str()).unwrap())
+        } else {
+            None
+        }
+    };
+    let exported_symbols = cgcx.exported_symbols.as_ref().expect("needs exported symbols for LTO");
+    let mut symbols_below_threshold = {
+        let _timer = cgcx.prof.generic_activity("GCC_lto_generate_symbols_below_threshold");
+        exported_symbols[&LOCAL_CRATE].iter().filter_map(symbol_filter).collect::<Vec<CString>>()
+    };
+    info!("{} symbols to preserve in this crate", symbols_below_threshold.len());
+
+    // If we're performing LTO for the entire crate graph, then for each of our
+    // upstream dependencies, find the corresponding rlib and load the bitcode
+    // from the archive.
+    //
+    // We save off all the bytecode and GCC module file path for later processing
+    // with either fat or thin LTO
+    let mut upstream_modules = Vec::new();
+    if cgcx.lto != Lto::ThinLocal {
+        // Make sure we actually can run LTO
+        for crate_type in cgcx.crate_types.iter() {
+            if !crate_type_allows_lto(*crate_type) {
+                diag_handler.emit_err(LtoDisallowed);
+                return Err(FatalError);
+            } else if *crate_type == CrateType::Dylib {
+                if !cgcx.opts.unstable_opts.dylib_lto {
+                    diag_handler.emit_err(LtoDylib);
+                    return Err(FatalError);
+                }
+            }
+        }
+
+        if cgcx.opts.cg.prefer_dynamic && !cgcx.opts.unstable_opts.dylib_lto {
+            diag_handler.emit_err(DynamicLinkingWithLTO);
+            return Err(FatalError);
+        }
+
+        for &(cnum, ref path) in cgcx.each_linked_rlib_for_lto.iter() {
+            let exported_symbols =
+                cgcx.exported_symbols.as_ref().expect("needs exported symbols for LTO");
+            {
+                let _timer =
+                    cgcx.prof.generic_activity("GCC_lto_generate_symbols_below_threshold");
+                symbols_below_threshold
+                    .extend(exported_symbols[&cnum].iter().filter_map(symbol_filter));
+            }
+
+            let archive_data = unsafe {
+                Mmap::map(File::open(&path).expect("couldn't open rlib"))
+                    .expect("couldn't map rlib")
+            };
+            let archive = ArchiveFile::parse(&*archive_data).expect("wanted an rlib");
+            let obj_files = archive
+                .members()
+                .filter_map(|child| {
+                    child.ok().and_then(|c| {
+                        std::str::from_utf8(c.name()).ok().map(|name| (name.trim(), c))
+                    })
+                })
+                .filter(|&(name, _)| looks_like_rust_object_file(name));
+            for (name, child) in obj_files {
+                info!("adding bitcode from {}", name);
+                let path = tmp_path.path().join(name);
+                match save_as_file(child.data(&*archive_data).expect("corrupt rlib"), &path) {
+                    Ok(()) => {
+                        let buffer = ModuleBuffer::new(path);
+                        let module = SerializedModule::Local(buffer);
+                        upstream_modules.push((module, CString::new(name).unwrap()));
+                    }
+                    Err(e) => {
+                        diag_handler.emit_err(e);
+                        return Err(FatalError);
+                    }
+                }
+            }
+        }
+    }
+
+    Ok(LtoData {
+        //symbols_below_threshold,
+        upstream_modules,
+        tmp_path,
+    })
+}
+
+fn save_as_file(obj: &[u8], path: &Path) -> Result<(), LtoBitcodeFromRlib> {
+    fs::write(path, obj)
+        .map_err(|error| LtoBitcodeFromRlib {
+            gcc_err: format!("write object file to temp dir: {}", error)
+        })
+}
+
+/// Performs fat LTO by merging all modules into a single one and returning it
+/// for further optimization.
+pub(crate) fn run_fat(
+    cgcx: &CodegenContext<GccCodegenBackend>,
+    modules: Vec<FatLtoInput<GccCodegenBackend>>,
+    cached_modules: Vec<(SerializedModule<ModuleBuffer>, WorkProduct)>,
+) -> Result<LtoModuleCodegen<GccCodegenBackend>, FatalError> {
+    let diag_handler = cgcx.create_diag_handler();
+    let lto_data = prepare_lto(cgcx, &diag_handler)?;
+    /*let symbols_below_threshold =
+        lto_data.symbols_below_threshold.iter().map(|c| c.as_ptr()).collect::<Vec<_>>();*/
+    fat_lto(cgcx, &diag_handler, modules, cached_modules, lto_data.upstream_modules, lto_data.tmp_path,
+        //&symbols_below_threshold,
+    )
+}
+
+fn fat_lto(cgcx: &CodegenContext<GccCodegenBackend>, _diag_handler: &Handler, modules: Vec<FatLtoInput<GccCodegenBackend>>, cached_modules: Vec<(SerializedModule<ModuleBuffer>, WorkProduct)>, mut serialized_modules: Vec<(SerializedModule<ModuleBuffer>, CString)>, tmp_path: TempDir,
+    //symbols_below_threshold: &[*const libc::c_char],
+) -> Result<LtoModuleCodegen<GccCodegenBackend>, FatalError> {
+    let _timer = cgcx.prof.generic_activity("GCC_fat_lto_build_monolithic_module");
+    info!("going for a fat lto");
+
+    // Sort out all our lists of incoming modules into two lists.
+    //
+    // * `serialized_modules` (also and argument to this function) contains all
+    //   modules that are serialized in-memory.
+    // * `in_memory` contains modules which are already parsed and in-memory,
+    //   such as from multi-CGU builds.
+    //
+    // All of `cached_modules` (cached from previous incremental builds) can
+    // immediately go onto the `serialized_modules` modules list and then we can
+    // split the `modules` array into these two lists.
+    let mut in_memory = Vec::new();
+    serialized_modules.extend(cached_modules.into_iter().map(|(buffer, wp)| {
+        info!("pushing cached module {:?}", wp.cgu_name);
+        (buffer, CString::new(wp.cgu_name).unwrap())
+    }));
+    for module in modules {
+        match module {
+            FatLtoInput::InMemory(m) => in_memory.push(m),
+            FatLtoInput::Serialized { name, buffer } => {
+                info!("pushing serialized module {:?}", name);
+                let buffer = SerializedModule::Local(buffer);
+                serialized_modules.push((buffer, CString::new(name).unwrap()));
+            }
+        }
+    }
+
+    // Find the "costliest" module and merge everything into that codegen unit.
+    // All the other modules will be serialized and reparsed into the new
+    // context, so this hopefully avoids serializing and parsing the largest
+    // codegen unit.
+    //
+    // Additionally use a regular module as the base here to ensure that various
+    // file copy operations in the backend work correctly. The only other kind
+    // of module here should be an allocator one, and if your crate is smaller
+    // than the allocator module then the size doesn't really matter anyway.
+    let costliest_module = in_memory
+        .iter()
+        .enumerate()
+        .filter(|&(_, module)| module.kind == ModuleKind::Regular)
+        .map(|(i, _module)| {
+            //let cost = unsafe { llvm::LLVMRustModuleCost(module.module_llvm.llmod()) };
+            // TODO(antoyo): compute the cost of a module if GCC allows this.
+            (0, i)
+        })
+        .max();
+
+    // If we found a costliest module, we're good to go. Otherwise all our
+    // inputs were serialized which could happen in the case, for example, that
+    // all our inputs were incrementally reread from the cache and we're just
+    // re-executing the LTO passes. If that's the case deserialize the first
+    // module and create a linker with it.
+    let mut module: ModuleCodegen<GccContext> = match costliest_module {
+        Some((_cost, i)) => in_memory.remove(i),
+        None => {
+            unimplemented!("Incremental");
+            /*assert!(!serialized_modules.is_empty(), "must have at least one serialized module");
+            let (buffer, name) = serialized_modules.remove(0);
+            info!("no in-memory regular modules to choose from, parsing {:?}", name);
+            ModuleCodegen {
+                module_llvm: GccContext::parse(cgcx, &name, buffer.data(), diag_handler)?,
+                name: name.into_string().unwrap(),
+                kind: ModuleKind::Regular,
+            }*/
+        }
+    };
+    let mut serialized_bitcode = Vec::new();
+    {
+        info!("using {:?} as a base module", module.name);
+
+        // We cannot load and merge GCC contexts in memory like cg_llvm is doing.
+        // Instead, we combine the object files into a single object file.
+        for module in in_memory {
+            let path = tmp_path.path().to_path_buf().join(&module.name);
+            let path = path.to_str().expect("path");
+            let context = &module.module_llvm.context;
+            let config = cgcx.config(module.kind);
+            // NOTE: we need to set the optimization level here in order for LTO to do its job.
+            context.set_optimization_level(to_gcc_opt_level(config.opt_level));
+            context.add_command_line_option("-flto=auto");
+            context.add_command_line_option("-flto-partition=one");
+            context.compile_to_file(OutputKind::ObjectFile, path);
+            let buffer = ModuleBuffer::new(PathBuf::from(path));
+            let llmod_id = CString::new(&module.name[..]).unwrap();
+            serialized_modules.push((SerializedModule::Local(buffer), llmod_id));
+        }
+        // Sort the modules to ensure we produce deterministic results.
+        serialized_modules.sort_by(|module1, module2| module1.1.cmp(&module2.1));
+
+        // We add the object files and save in should_combine_object_files that we should combine
+        // them into a single object file when compiling later.
+        for (bc_decoded, name) in serialized_modules {
+            let _timer = cgcx
+                .prof
+                .generic_activity_with_arg_recorder("GCC_fat_lto_link_module", |recorder| {
+                    recorder.record_arg(format!("{:?}", name))
+                });
+            info!("linking {:?}", name);
+            match bc_decoded {
+                SerializedModule::Local(ref module_buffer) => {
+                    module.module_llvm.should_combine_object_files = true;
+                    module.module_llvm.context.add_driver_option(module_buffer.0.to_str().expect("path"));
+                },
+                SerializedModule::FromRlib(_) => unimplemented!("from rlib"),
+                SerializedModule::FromUncompressedFile(_) => unimplemented!("from uncompressed file"),
+            }
+            serialized_bitcode.push(bc_decoded);
+        }
+        save_temp_bitcode(cgcx, &module, "lto.input");
+
+        // Internalize everything below threshold to help strip out more modules and such.
+        /*unsafe {
+            let ptr = symbols_below_threshold.as_ptr();
+            llvm::LLVMRustRunRestrictionPass(
+                llmod,
+                ptr as *const *const libc::c_char,
+                symbols_below_threshold.len() as libc::size_t,
+            );*/
+            save_temp_bitcode(cgcx, &module, "lto.after-restriction");
+        //}
+    }
+
+    // NOTE: save the temporary directory used by LTO so that it gets deleted after linking instead
+    // of now.
+    module.module_llvm.temp_dir = Some(tmp_path);
+
+    Ok(LtoModuleCodegen::Fat { module, _serialized_bitcode: serialized_bitcode })
+}
+
+pub struct ModuleBuffer(PathBuf);
+
+impl ModuleBuffer {
+    pub fn new(path: PathBuf) -> ModuleBuffer {
+        ModuleBuffer(path)
+    }
+}
+
+impl ModuleBufferMethods for ModuleBuffer {
+    fn data(&self) -> &[u8] {
+        unimplemented!("data not needed for GCC codegen");
+    }
+}
diff --git a/src/back/mod.rs b/src/back/mod.rs
index d692799d764..10187eab0d7 100644
--- a/src/back/mod.rs
+++ b/src/back/mod.rs
@@ -1 +1,2 @@
+pub mod lto;
 pub mod write;
diff --git a/src/back/write.rs b/src/back/write.rs
index 5f54ac4ebc6..04772d7707a 100644
--- a/src/back/write.rs
+++ b/src/back/write.rs
@@ -2,27 +2,71 @@ use std::{env, fs};
 
 use gccjit::OutputKind;
 use rustc_codegen_ssa::{CompiledModule, ModuleCodegen};
-use rustc_codegen_ssa::back::write::{CodegenContext, EmitObj, ModuleConfig};
+use rustc_codegen_ssa::back::link::ensure_removed;
+use rustc_codegen_ssa::back::write::{BitcodeSection, CodegenContext, EmitObj, ModuleConfig};
 use rustc_errors::Handler;
+use rustc_fs_util::link_or_copy;
 use rustc_session::config::OutputType;
 use rustc_span::fatal_error::FatalError;
 use rustc_target::spec::SplitDebuginfo;
 
 use crate::{GccCodegenBackend, GccContext};
+use crate::errors::CopyBitcode;
 
-pub(crate) unsafe fn codegen(cgcx: &CodegenContext<GccCodegenBackend>, _diag_handler: &Handler, module: ModuleCodegen<GccContext>, config: &ModuleConfig) -> Result<CompiledModule, FatalError> {
-    let _timer = cgcx.prof.generic_activity_with_arg("LLVM_module_codegen", &*module.name);
+pub(crate) unsafe fn codegen(cgcx: &CodegenContext<GccCodegenBackend>, diag_handler: &Handler, module: ModuleCodegen<GccContext>, config: &ModuleConfig) -> Result<CompiledModule, FatalError> {
+    let _timer = cgcx.prof.generic_activity_with_arg("GCC_module_codegen", &*module.name);
     {
         let context = &module.module_llvm.context;
 
         let module_name = module.name.clone();
+
+        let should_combine_object_files = module.module_llvm.should_combine_object_files;
+
         let module_name = Some(&module_name[..]);
 
-        let _bc_out = cgcx.output_filenames.temp_path(OutputType::Bitcode, module_name);
+        // NOTE: Only generate object files with GIMPLE when this environment variable is set for
+        // now because this requires a particular setup (same gcc/lto1/lto-wrapper commit as libgccjit).
+        let fat_lto = env::var("EMBED_LTO_BITCODE").as_deref() == Ok("1");
+
+        let bc_out = cgcx.output_filenames.temp_path(OutputType::Bitcode, module_name);
         let obj_out = cgcx.output_filenames.temp_path(OutputType::Object, module_name);
 
-        if config.bitcode_needed() {
+        if config.bitcode_needed() && fat_lto {
+            let _timer = cgcx
+                .prof
+                .generic_activity_with_arg("GCC_module_codegen_make_bitcode", &*module.name);
+
             // TODO(antoyo)
+            /*if let Some(bitcode_filename) = bc_out.file_name() {
+                cgcx.prof.artifact_size(
+                    "llvm_bitcode",
+                    bitcode_filename.to_string_lossy(),
+                    data.len() as u64,
+                );
+            }*/
+
+            if config.emit_bc || config.emit_obj == EmitObj::Bitcode {
+                let _timer = cgcx
+                    .prof
+                    .generic_activity_with_arg("GCC_module_codegen_emit_bitcode", &*module.name);
+                context.add_command_line_option("-flto=auto");
+                context.add_command_line_option("-flto-partition=one");
+                context.compile_to_file(OutputKind::ObjectFile, bc_out.to_str().expect("path to str"));
+            }
+
+            if config.emit_obj == EmitObj::ObjectCode(BitcodeSection::Full) {
+                let _timer = cgcx
+                    .prof
+                    .generic_activity_with_arg("GCC_module_codegen_embed_bitcode", &*module.name);
+                // TODO(antoyo): maybe we should call embed_bitcode to have the proper iOS fixes?
+                //embed_bitcode(cgcx, llcx, llmod, &config.bc_cmdline, data);
+
+                context.add_command_line_option("-flto=auto");
+                context.add_command_line_option("-flto-partition=one");
+                context.add_command_line_option("-ffat-lto-objects");
+                // TODO(antoyo): Send -plugin/usr/lib/gcc/x86_64-pc-linux-gnu/11.1.0/liblto_plugin.so to linker (this should be done when specifying the appropriate rustc cli argument).
+                context.compile_to_file(OutputKind::ObjectFile, bc_out.to_str().expect("path to str"));
+            }
         }
 
         if config.emit_ir {
@@ -32,7 +76,7 @@ pub(crate) unsafe fn codegen(cgcx: &CodegenContext<GccCodegenBackend>, _diag_han
         if config.emit_asm {
             let _timer = cgcx
                 .prof
-                .generic_activity_with_arg("LLVM_module_codegen_emit_asm", &*module.name);
+                .generic_activity_with_arg("GCC_module_codegen_emit_asm", &*module.name);
             let path = cgcx.output_filenames.temp_path(OutputType::Assembly, module_name);
             context.compile_to_file(OutputKind::Assembler, path.to_str().expect("path to str"));
         }
@@ -41,7 +85,7 @@ pub(crate) unsafe fn codegen(cgcx: &CodegenContext<GccCodegenBackend>, _diag_han
             EmitObj::ObjectCode(_) => {
                 let _timer = cgcx
                     .prof
-                    .generic_activity_with_arg("LLVM_module_codegen_emit_obj", &*module.name);
+                    .generic_activity_with_arg("GCC_module_codegen_emit_obj", &*module.name);
                 if env::var("CG_GCCJIT_DUMP_MODULE_NAMES").as_deref() == Ok("1") {
                     println!("Module {}", module.name);
                 }
@@ -60,11 +104,36 @@ pub(crate) unsafe fn codegen(cgcx: &CodegenContext<GccCodegenBackend>, _diag_han
                     context.set_debug_info(true);
                     context.dump_to_file(path, true);
                 }
-                context.compile_to_file(OutputKind::ObjectFile, obj_out.to_str().expect("path to str"));
+                if should_combine_object_files && fat_lto {
+                    context.add_command_line_option("-flto=auto");
+                    context.add_command_line_option("-flto-partition=one");
+
+                    context.add_driver_option("-Wl,-r");
+                    // NOTE: we need -nostdlib, otherwise, we get the following error:
+                    // /usr/bin/ld: cannot find -lgcc_s: No such file or directory
+                    context.add_driver_option("-nostdlib");
+                    // NOTE: without -fuse-linker-plugin, we get the following error:
+                    // lto1: internal compiler error: decompressed stream: Destination buffer is too small
+                    context.add_driver_option("-fuse-linker-plugin");
+
+                    // NOTE: this doesn't actually generate an executable. With the above flags, it combines the .o files together in another .o.
+                    context.compile_to_file(OutputKind::Executable, obj_out.to_str().expect("path to str"));
+                }
+                else {
+                    context.compile_to_file(OutputKind::ObjectFile, obj_out.to_str().expect("path to str"));
+                }
             }
 
             EmitObj::Bitcode => {
-                // TODO(antoyo)
+                debug!("copying bitcode {:?} to obj {:?}", bc_out, obj_out);
+                if let Err(err) = link_or_copy(&bc_out, &obj_out) {
+                    diag_handler.emit_err(CopyBitcode { err });
+                }
+
+                if !config.emit_bc {
+                    debug!("removing_bitcode {:?}", bc_out);
+                    ensure_removed(diag_handler, &bc_out);
+                }
             }
 
             EmitObj::None => {}
@@ -82,3 +151,18 @@ pub(crate) unsafe fn codegen(cgcx: &CodegenContext<GccCodegenBackend>, _diag_han
 pub(crate) fn link(_cgcx: &CodegenContext<GccCodegenBackend>, _diag_handler: &Handler, mut _modules: Vec<ModuleCodegen<GccContext>>) -> Result<ModuleCodegen<GccContext>, FatalError> {
     unimplemented!();
 }
+
+pub(crate) fn save_temp_bitcode(cgcx: &CodegenContext<GccCodegenBackend>, _module: &ModuleCodegen<GccContext>, _name: &str) {
+    if !cgcx.save_temps {
+        return;
+    }
+    unimplemented!();
+    /*unsafe {
+        let ext = format!("{}.bc", name);
+        let cgu = Some(&module.name[..]);
+        let path = cgcx.output_filenames.temp_path_ext(&ext, cgu);
+        let cstr = path_to_c_string(&path);
+        let llmod = module.module_llvm.llmod();
+        llvm::LLVMWriteBitcodeToFile(llmod, cstr.as_ptr());
+    }*/
+}
diff --git a/src/base.rs b/src/base.rs
index bf0309fea14..266d60da10c 100644
--- a/src/base.rs
+++ b/src/base.rs
@@ -56,6 +56,7 @@ pub fn global_linkage_to_gcc(linkage: Linkage) -> GlobalKind {
 pub fn linkage_to_gcc(linkage: Linkage) -> FunctionType {
     match linkage {
         Linkage::External => FunctionType::Exported,
+        // TODO(antoyo): set the attribute externally_visible.
         Linkage::AvailableExternally => FunctionType::Extern,
         Linkage::LinkOnceAny => unimplemented!(),
         Linkage::LinkOnceODR => unimplemented!(),
@@ -91,7 +92,6 @@ pub fn compile_codegen_unit(tcx: TyCtxt<'_>, cgu_name: Symbol, target_info: Arc<
     fn module_codegen(tcx: TyCtxt<'_>, (cgu_name, target_info): (Symbol, Arc<TargetInfo>)) -> ModuleCodegen<GccContext> {
         let cgu = tcx.codegen_unit(cgu_name);
         // Instantiate monomorphizations without filling out definitions yet...
-        //let llvm_module = ModuleLlvm::new(tcx, &cgu_name.as_str());
         let context = Context::default();
 
         context.add_command_line_option("-fexceptions");
@@ -152,7 +152,10 @@ pub fn compile_codegen_unit(tcx: TyCtxt<'_>, cgu_name: Symbol, target_info: Arc<
             context.add_command_line_option("-fdump-rtl-all");
         }
         if env::var("CG_GCCJIT_DUMP_TREE_ALL").as_deref() == Ok("1") {
-            context.add_command_line_option("-fdump-tree-all");
+            context.add_command_line_option("-fdump-tree-all-eh");
+        }
+        if env::var("CG_GCCJIT_DUMP_IPA_ALL").as_deref() == Ok("1") {
+            context.add_command_line_option("-fdump-ipa-all-eh");
         }
         if env::var("CG_GCCJIT_DUMP_CODE").as_deref() == Ok("1") {
             context.set_dump_code_on_compile(true);
@@ -168,6 +171,10 @@ pub fn compile_codegen_unit(tcx: TyCtxt<'_>, cgu_name: Symbol, target_info: Arc<
             context.set_keep_intermediates(true);
         }
 
+        if env::var("CG_GCCJIT_VERBOSE").as_deref() == Ok("1") {
+            context.add_driver_option("-v");
+        }
+
         // NOTE: The codegen generates unrechable blocks.
         context.set_allow_unreachable_blocks(true);
 
@@ -197,7 +204,9 @@ pub fn compile_codegen_unit(tcx: TyCtxt<'_>, cgu_name: Symbol, target_info: Arc<
         ModuleCodegen {
             name: cgu_name.to_string(),
             module_llvm: GccContext {
-                context
+                context,
+                should_combine_object_files: false,
+                temp_dir: None,
             },
             kind: ModuleKind::Regular,
         }
diff --git a/src/declare.rs b/src/declare.rs
index 493626c3cf5..e673d0af4c7 100644
--- a/src/declare.rs
+++ b/src/declare.rs
@@ -1,4 +1,6 @@
 use gccjit::{Function, FunctionType, GlobalKind, LValue, RValue, Type};
+#[cfg(feature="master")]
+use gccjit::{FnAttribute, ToRValue};
 use rustc_codegen_ssa::traits::BaseTypeMethods;
 use rustc_middle::ty::Ty;
 use rustc_span::Symbol;
@@ -114,6 +116,44 @@ fn declare_raw_fn<'gcc>(cx: &CodegenCx<'gcc, '_>, name: &str, _callconv: () /*ll
                 .collect();
             let func = cx.context.new_function(None, cx.linkage.get(), return_type, &params, mangle_name(name), variadic);
             cx.functions.borrow_mut().insert(name.to_string(), func);
+
+            #[cfg(feature="master")]
+            if name == "rust_eh_personality" {
+                // NOTE: GCC will sometimes change the personality function set on a function from
+                // rust_eh_personality to __gcc_personality_v0 as an optimization.
+                // As such, we need to create a weak alias from __gcc_personality_v0 to
+                // rust_eh_personality in order to avoid a linker error.
+                // This needs to be weak in order to still allow using the standard
+                // __gcc_personality_v0 when the linking to it.
+                // Since aliases don't work (maybe because of a bug in LTO partitioning?), we
+                // create a wrapper function that calls rust_eh_personality.
+
+                let params: Vec<_> = param_types.into_iter().enumerate()
+                    .map(|(index, param)| cx.context.new_parameter(None, *param, &format!("param{}", index))) // TODO(antoyo): set name.
+                    .collect();
+                let gcc_func = cx.context.new_function(None, FunctionType::Exported, return_type, &params, "__gcc_personality_v0", variadic);
+
+                // We need a normal extern function for the crates that access rust_eh_personality
+                // without defining it, otherwise we'll get a compiler error.
+                //
+                // For the crate defining it, that needs to be a weak alias instead.
+                gcc_func.add_attribute(FnAttribute::Weak);
+
+                let block = gcc_func.new_block("start");
+                let mut args = vec![];
+                for param in &params {
+                    args.push(param.to_rvalue());
+                }
+                let call = cx.context.new_call(None, func, &args);
+                if return_type == cx.type_void() {
+                    block.add_eval(None, call);
+                    block.end_with_void_return(None);
+                }
+                else {
+                    block.end_with_return(None, call);
+                }
+            }
+
             func
         };
 
diff --git a/src/errors.rs b/src/errors.rs
index 693367192b1..19a967cb489 100644
--- a/src/errors.rs
+++ b/src/errors.rs
@@ -40,3 +40,34 @@ pub(crate) struct TiedTargetFeatures {
     pub span: Span,
     pub features: String,
 }
+
+#[derive(Diagnostic)]
+#[diag(codegen_gcc_copy_bitcode)]
+pub(crate) struct CopyBitcode {
+    pub err: std::io::Error,
+}
+
+#[derive(Diagnostic)]
+#[diag(codegen_gcc_dynamic_linking_with_lto)]
+#[note]
+pub(crate) struct DynamicLinkingWithLTO;
+
+#[derive(Diagnostic)]
+#[diag(codegen_gcc_load_bitcode)]
+pub(crate) struct LoadBitcode {
+    name: String,
+}
+
+#[derive(Diagnostic)]
+#[diag(codegen_gcc_lto_disallowed)]
+pub(crate) struct LtoDisallowed;
+
+#[derive(Diagnostic)]
+#[diag(codegen_gcc_lto_dylib)]
+pub(crate) struct LtoDylib;
+
+#[derive(Diagnostic)]
+#[diag(codegen_gcc_lto_bitcode_from_rlib)]
+pub(crate) struct LtoBitcodeFromRlib {
+    pub gcc_err: String,
+}
diff --git a/src/lib.rs b/src/lib.rs
index 7b55a4e4082..2de8fb3fc70 100644
--- a/src/lib.rs
+++ b/src/lib.rs
@@ -2,6 +2,12 @@
  * TODO(antoyo): implement equality in libgccjit based on https://zpz.github.io/blog/overloading-equality-operator-in-cpp-class-hierarchy/ (for type equality?)
  * TODO(antoyo): support #[inline] attributes.
  * TODO(antoyo): support LTO (gcc's equivalent to Full LTO is -flto -flto-partition=one — https://documentation.suse.com/sbp/all/html/SBP-GCC-10/index.html).
+ * For Thin LTO, this might be helpful:
+ * In gcc 4.6 -fwhopr was removed and became default with -flto. The non-whopr path can still be executed via -flto-partition=none.
+ *
+ * Maybe some missing optizations enabled by rustc's LTO is in there: https://gcc.gnu.org/onlinedocs/gcc/Optimize-Options.html
+ * Like -fipa-icf (should be already enabled) and maybe -fdevirtualize-at-ltrans.
+ * TODO: disable debug info always being emitted. Perhaps this slows down things?
  *
  * TODO(antoyo): remove the patches.
  */
@@ -28,6 +34,7 @@ extern crate rustc_codegen_ssa;
 extern crate rustc_data_structures;
 extern crate rustc_errors;
 extern crate rustc_fluent_macro;
+extern crate rustc_fs_util;
 extern crate rustc_hir;
 extern crate rustc_macros;
 extern crate rustc_metadata;
@@ -35,6 +42,8 @@ extern crate rustc_middle;
 extern crate rustc_session;
 extern crate rustc_span;
 extern crate rustc_target;
+#[macro_use]
+extern crate tracing;
 
 // This prevents duplicating functions and statics that are already part of the host rustc process.
 #[allow(unused_extern_crates)]
@@ -65,22 +74,24 @@ mod type_of;
 use std::any::Any;
 use std::sync::Arc;
 #[cfg(not(feature="master"))]
-use std::sync::atomic::{AtomicBool, Ordering};
+use std::sync::atomic::AtomicBool;
+#[cfg(not(feature="master"))]
+use std::sync::atomic::Ordering;
 
-use crate::errors::LTONotSupported;
 use gccjit::{Context, OptimizationLevel};
 #[cfg(feature="master")]
 use gccjit::TargetInfo;
 #[cfg(not(feature="master"))]
 use gccjit::CType;
+use errors::LTONotSupported;
 use rustc_ast::expand::allocator::AllocatorKind;
 use rustc_codegen_ssa::{CodegenResults, CompiledModule, ModuleCodegen};
 use rustc_codegen_ssa::base::codegen_crate;
 use rustc_codegen_ssa::back::write::{CodegenContext, FatLtoInput, ModuleConfig, TargetMachineFactoryFn};
 use rustc_codegen_ssa::back::lto::{LtoModuleCodegen, SerializedModule, ThinModule};
 use rustc_codegen_ssa::target_features::supported_target_features;
-use rustc_codegen_ssa::traits::{CodegenBackend, ExtraBackendMethods, ModuleBufferMethods, ThinBufferMethods, WriteBackendMethods};
 use rustc_data_structures::fx::FxIndexMap;
+use rustc_codegen_ssa::traits::{CodegenBackend, ExtraBackendMethods, ThinBufferMethods, WriteBackendMethods};
 use rustc_errors::{DiagnosticMessage, ErrorGuaranteed, Handler, SubdiagnosticMessage};
 use rustc_fluent_macro::fluent_messages;
 use rustc_metadata::EncodedMetadata;
@@ -91,9 +102,10 @@ use rustc_session::config::{Lto, OptLevel, OutputFilenames};
 use rustc_session::Session;
 use rustc_span::Symbol;
 use rustc_span::fatal_error::FatalError;
-#[cfg(not(feature="master"))]
 use tempfile::TempDir;
 
+use crate::back::lto::ModuleBuffer;
+
 fluent_messages! { "../messages.ftl" }
 
 pub struct PrintOnPanic<F: Fn() -> String>(pub F);
@@ -136,7 +148,7 @@ impl CodegenBackend for GccCodegenBackend {
     fn init(&self, sess: &Session) {
         #[cfg(feature="master")]
         gccjit::set_global_personality_function_name(b"rust_eh_personality\0");
-        if sess.lto() != Lto::No {
+        if sess.lto() == Lto::Thin {
             sess.emit_warning(LTONotSupported {});
         }
 
@@ -194,7 +206,12 @@ impl ExtraBackendMethods for GccCodegenBackend {
     fn codegen_allocator<'tcx>(&self, tcx: TyCtxt<'tcx>, module_name: &str, kind: AllocatorKind, alloc_error_handler_kind: AllocatorKind) -> Self::Module {
         let mut mods = GccContext {
             context: Context::default(),
+            should_combine_object_files: false,
+            temp_dir: None,
         };
+
+        // TODO(antoyo): only set for x86.
+        mods.context.add_command_line_option("-masm=intel");
         unsafe { allocator::codegen(tcx, &mut mods, module_name, kind, alloc_error_handler_kind); }
         mods
     }
@@ -211,14 +228,6 @@ impl ExtraBackendMethods for GccCodegenBackend {
     }
 }
 
-pub struct ModuleBuffer;
-
-impl ModuleBufferMethods for ModuleBuffer {
-    fn data(&self) -> &[u8] {
-        unimplemented!();
-    }
-}
-
 pub struct ThinBuffer;
 
 impl ThinBufferMethods for ThinBuffer {
@@ -229,6 +238,9 @@ impl ThinBufferMethods for ThinBuffer {
 
 pub struct GccContext {
     context: Context<'static>,
+    should_combine_object_files: bool,
+    // Temporary directory used by LTO. We keep it here so that it's not removed before linking.
+    temp_dir: Option<TempDir>,
 }
 
 unsafe impl Send for GccContext {}
@@ -243,18 +255,8 @@ impl WriteBackendMethods for GccCodegenBackend {
     type ThinData = ();
     type ThinBuffer = ThinBuffer;
 
-    fn run_fat_lto(_cgcx: &CodegenContext<Self>, mut modules: Vec<FatLtoInput<Self>>, _cached_modules: Vec<(SerializedModule<Self::ModuleBuffer>, WorkProduct)>) -> Result<LtoModuleCodegen<Self>, FatalError> {
-        // TODO(antoyo): implement LTO by sending -flto to libgccjit and adding the appropriate gcc linker plugins.
-        // NOTE: implemented elsewhere.
-        // TODO(antoyo): what is implemented elsewhere ^ ?
-        let module =
-            match modules.remove(0) {
-                FatLtoInput::InMemory(module) => module,
-                FatLtoInput::Serialized { .. } => {
-                    unimplemented!();
-                }
-            };
-        Ok(LtoModuleCodegen::Fat { module, _serialized_bitcode: vec![] })
+    fn run_fat_lto(cgcx: &CodegenContext<Self>, modules: Vec<FatLtoInput<Self>>, cached_modules: Vec<(SerializedModule<Self::ModuleBuffer>, WorkProduct)>) -> Result<LtoModuleCodegen<Self>, FatalError> {
+        back::lto::run_fat(cgcx, modules, cached_modules)
     }
 
     fn run_thin_lto(_cgcx: &CodegenContext<Self>, _modules: Vec<(String, Self::ThinBuffer)>, _cached_modules: Vec<(SerializedModule<Self::ModuleBuffer>, WorkProduct)>) -> Result<(Vec<LtoModuleCodegen<Self>>, Vec<WorkProduct>), FatalError> {