Add `-Z combine_cgu` flag

Introduce a compiler option to let rustc combines all regular CGUs into
a single one at the end of compilation.

Part of Issue #64191

compiler/rustc_codegen_llvm/src/back/lto.rs

@@ -346,14 +346,14 @@ fn fat_lto(
     Ok(LtoModuleCodegen::Fat { module: Some(module), _serialized_bitcode: serialized_bitcode })
 }
 
-struct Linker<'a>(&'a mut llvm::Linker<'a>);
+crate struct Linker<'a>(&'a mut llvm::Linker<'a>);
 
 impl Linker<'a> {
-    fn new(llmod: &'a llvm::Module) -> Self {
+    crate fn new(llmod: &'a llvm::Module) -> Self {
         unsafe { Linker(llvm::LLVMRustLinkerNew(llmod)) }
     }
 
-    fn add(&mut self, bytecode: &[u8]) -> Result<(), ()> {
+    crate fn add(&mut self, bytecode: &[u8]) -> Result<(), ()> {
         unsafe {
             if llvm::LLVMRustLinkerAdd(
                 self.0,

compiler/rustc_codegen_llvm/src/back/write.rs

@@ -617,6 +617,31 @@ unsafe fn add_sanitizer_passes(config: &ModuleConfig, passes: &mut Vec<&'static
     }
 }
 
+pub(crate) fn link(
+    cgcx: &CodegenContext<LlvmCodegenBackend>,
+    diag_handler: &Handler,
+    mut modules: Vec<ModuleCodegen<ModuleLlvm>>,
+) -> Result<ModuleCodegen<ModuleLlvm>, FatalError> {
+    use super::lto::{Linker, ModuleBuffer};
+    // Sort the modules by name to ensure to ensure deterministic behavior.
+    modules.sort_by(|a, b| a.name.cmp(&b.name));
+    let (first, elements) =
+        modules.split_first().expect("Bug! modules must contain at least one module.");
+
+    let mut linker = Linker::new(first.module_llvm.llmod());
+    for module in elements {
+        let _timer =
+            cgcx.prof.generic_activity_with_arg("LLVM_link_module", format!("{:?}", module.name));
+        let buffer = ModuleBuffer::new(module.module_llvm.llmod());
+        linker.add(&buffer.data()).map_err(|()| {
+            let msg = format!("failed to serialize module {:?}", module.name);
+            llvm_err(&diag_handler, &msg)
+        })?;
+    }
+    drop(linker);
+    Ok(modules.remove(0))
+}
+
 pub(crate) unsafe fn codegen(
     cgcx: &CodegenContext<LlvmCodegenBackend>,
     diag_handler: &Handler,

compiler/rustc_codegen_llvm/src/lib.rs

@@ -130,6 +130,13 @@ impl WriteBackendMethods for LlvmCodegenBackend {
             llvm::LLVMRustPrintPassTimings();
         }
     }
+    fn run_link(
+        cgcx: &CodegenContext<Self>,
+        diag_handler: &Handler,
+        modules: Vec<ModuleCodegen<Self::Module>>,
+    ) -> Result<ModuleCodegen<Self::Module>, FatalError> {
+        back::write::link(cgcx, diag_handler, modules)
+    }
     fn run_fat_lto(
         cgcx: &CodegenContext<Self>,
         modules: Vec<FatLTOInput<Self>>,

compiler/rustc_codegen_ssa/src/back/write.rs

@@ -702,6 +702,7 @@ impl<B: WriteBackendMethods> WorkItem<B> {
 
 enum WorkItemResult<B: WriteBackendMethods> {
     Compiled(CompiledModule),
+    NeedsLink(ModuleCodegen<B::Module>),
     NeedsFatLTO(FatLTOInput<B>),
     NeedsThinLTO(String, B::ThinBuffer),
 }
@@ -801,11 +802,8 @@ fn execute_optimize_work_item<B: ExtraBackendMethods>(
         None
     };
 
-    Ok(match lto_type {
-        ComputedLtoType::No => {
-            let module = unsafe { B::codegen(cgcx, &diag_handler, module, module_config)? };
-            WorkItemResult::Compiled(module)
-        }
+    match lto_type {
+        ComputedLtoType::No => finish_intra_module_work(cgcx, module, module_config),
         ComputedLtoType::Thin => {
             let (name, thin_buffer) = B::prepare_thin(module);
             if let Some(path) = bitcode {
@@ -813,7 +811,7 @@ fn execute_optimize_work_item<B: ExtraBackendMethods>(
                     panic!("Error writing pre-lto-bitcode file `{}`: {}", path.display(), e);
                 });
             }
-            WorkItemResult::NeedsThinLTO(name, thin_buffer)
+            Ok(WorkItemResult::NeedsThinLTO(name, thin_buffer))
         }
         ComputedLtoType::Fat => match bitcode {
             Some(path) => {
@@ -821,11 +819,11 @@ fn execute_optimize_work_item<B: ExtraBackendMethods>(
                 fs::write(&path, buffer.data()).unwrap_or_else(|e| {
                     panic!("Error writing pre-lto-bitcode file `{}`: {}", path.display(), e);
                 });
-                WorkItemResult::NeedsFatLTO(FatLTOInput::Serialized { name, buffer })
+                Ok(WorkItemResult::NeedsFatLTO(FatLTOInput::Serialized { name, buffer }))
             }
-            None => WorkItemResult::NeedsFatLTO(FatLTOInput::InMemory(module)),
+            None => Ok(WorkItemResult::NeedsFatLTO(FatLTOInput::InMemory(module))),
         },
-    })
+    }
 }
 
 fn execute_copy_from_cache_work_item<B: ExtraBackendMethods>(
@@ -870,13 +868,26 @@ fn execute_lto_work_item<B: ExtraBackendMethods>(
     cgcx: &CodegenContext<B>,
     mut module: lto::LtoModuleCodegen<B>,
     module_config: &ModuleConfig,
+) -> Result<WorkItemResult<B>, FatalError> {
+    let module = unsafe { module.optimize(cgcx)? };
+    finish_intra_module_work(cgcx, module, module_config)
+}
+
+fn finish_intra_module_work<B: ExtraBackendMethods>(
+    cgcx: &CodegenContext<B>,
+    module: ModuleCodegen<B::Module>,
+    module_config: &ModuleConfig,
 ) -> Result<WorkItemResult<B>, FatalError> {
     let diag_handler = cgcx.create_diag_handler();
 
-    unsafe {
-        let module = module.optimize(cgcx)?;
-        let module = B::codegen(cgcx, &diag_handler, module, module_config)?;
+    if !cgcx.opts.debugging_opts.combine_cgu
+        || module.kind == ModuleKind::Metadata
+        || module.kind == ModuleKind::Allocator
+    {
+        let module = unsafe { B::codegen(cgcx, &diag_handler, module, module_config)? };
         Ok(WorkItemResult::Compiled(module))
+    } else {
+        Ok(WorkItemResult::NeedsLink(module))
     }
 }
 
@@ -891,6 +902,10 @@ pub enum Message<B: WriteBackendMethods> {
         thin_buffer: B::ThinBuffer,
         worker_id: usize,
     },
+    NeedsLink {
+        module: ModuleCodegen<B::Module>,
+        worker_id: usize,
+    },
     Done {
         result: Result<CompiledModule, Option<WorkerFatalError>>,
         worker_id: usize,
@@ -1178,6 +1193,7 @@ fn start_executing_work<B: ExtraBackendMethods>(
         let mut compiled_modules = vec![];
         let mut compiled_metadata_module = None;
         let mut compiled_allocator_module = None;
+        let mut needs_link = Vec::new();
         let mut needs_fat_lto = Vec::new();
         let mut needs_thin_lto = Vec::new();
         let mut lto_import_only_modules = Vec::new();
@@ -1434,6 +1450,10 @@ fn start_executing_work<B: ExtraBackendMethods>(
                         }
                     }
                 }
+                Message::NeedsLink { module, worker_id } => {
+                    free_worker(worker_id);
+                    needs_link.push(module);
+                }
                 Message::NeedsFatLTO { result, worker_id } => {
                     assert!(!started_lto);
                     free_worker(worker_id);
@@ -1462,6 +1482,18 @@ fn start_executing_work<B: ExtraBackendMethods>(
             }
         }
 
+        let needs_link = mem::take(&mut needs_link);
+        if !needs_link.is_empty() {
+            assert!(compiled_modules.is_empty());
+            let diag_handler = cgcx.create_diag_handler();
+            let module = B::run_link(&cgcx, &diag_handler, needs_link).map_err(|_| ())?;
+            let module = unsafe {
+                B::codegen(&cgcx, &diag_handler, module, cgcx.config(ModuleKind::Regular))
+                    .map_err(|_| ())?
+            };
+            compiled_modules.push(module);
+        }
+
         // Drop to print timings
         drop(llvm_start_time);
 
@@ -1521,6 +1553,9 @@ fn spawn_work<B: ExtraBackendMethods>(cgcx: CodegenContext<B>, work: WorkItem<B>
                     Some(Ok(WorkItemResult::Compiled(m))) => {
                         Message::Done::<B> { result: Ok(m), worker_id }
                     }
+                    Some(Ok(WorkItemResult::NeedsLink(m))) => {
+                        Message::NeedsLink::<B> { module: m, worker_id }
+                    }
                     Some(Ok(WorkItemResult::NeedsFatLTO(m))) => {
                         Message::NeedsFatLTO::<B> { result: m, worker_id }
                     }

compiler/rustc_codegen_ssa/src/traits/write.rs

@@ -13,6 +13,12 @@ pub trait WriteBackendMethods: 'static + Sized + Clone {
     type ThinData: Send + Sync;
     type ThinBuffer: ThinBufferMethods;
 
+    /// Merge all modules into main_module and returning it
+    fn run_link(
+        cgcx: &CodegenContext<Self>,
+        diag_handler: &Handler,
+        modules: Vec<ModuleCodegen<Self::Module>>,
+    ) -> Result<ModuleCodegen<Self::Module>, FatalError>;
     /// Performs fat LTO by merging all modules into a single one and returning it
     /// for further optimization.
     fn run_fat_lto(

compiler/rustc_session/src/options.rs

@@ -850,6 +850,8 @@ options! {DebuggingOptions, DebuggingSetter, basic_debugging_options,
         "enable the experimental Chalk-based trait solving engine"),
     codegen_backend: Option<String> = (None, parse_opt_string, [TRACKED],
         "the backend to use"),
+    combine_cgu: bool = (false, parse_bool, [TRACKED],
+        "combine CGUs into a single one"),
     crate_attr: Vec<String> = (Vec::new(), parse_string_push, [TRACKED],
         "inject the given attribute in the crate"),
     debug_macros: bool = (false, parse_bool, [TRACKED],