Merge pull request #1264 from bjorn3/parallel_comp_refactor

Refactorings for enabling parallel compilation (part 1)

8 files changed, 482 insertions, 297 deletions

diff --git a/Readme.md b/Readme.md
index 8a2db5a43ec..1e84c7fa365 100644
--- a/Readme.md
+++ b/Readme.md
@@ -52,9 +52,7 @@ configuration options.
 ## Not yet supported
 
 * Inline assembly ([no cranelift support](https://github.com/bytecodealliance/wasmtime/issues/1041))
-    * On Linux there is support for invoking an external assembler for `global_asm!` and `asm!`.
-      `llvm_asm!` will remain unimplemented forever. `asm!` doesn't yet support reg classes. You
-      have to specify specific registers instead.
+    * On UNIX there is support for invoking an external assembler for `global_asm!` and `asm!`.
 * SIMD ([tracked here](https://github.com/bjorn3/rustc_codegen_cranelift/issues/171), some basic things work)
 
 ## License
diff --git a/example/mini_core_hello_world.rs b/example/mini_core_hello_world.rs
index 7e9cbe1bba5..e83be3a3df5 100644
--- a/example/mini_core_hello_world.rs
+++ b/example/mini_core_hello_world.rs
@@ -321,7 +321,7 @@ fn main() {
     #[cfg(not(any(jit, windows)))]
     test_tls();
 
-    #[cfg(all(not(jit), target_arch = "x86_64", target_os = "linux"))]
+    #[cfg(all(not(jit), target_arch = "x86_64", any(target_os = "linux", target_os = "darwin")))]
     unsafe {
         global_asm_test();
     }
@@ -343,7 +343,7 @@ fn main() {
 }
 }
 
-#[cfg(all(not(jit), target_arch = "x86_64", target_os = "linux"))]
+#[cfg(all(not(jit), target_arch = "x86_64", any(target_os = "linux", target_os = "darwin")))]
 extern "C" {
     fn global_asm_test();
 }
@@ -358,6 +358,16 @@ global_asm! {
     "
 }
 
+#[cfg(all(not(jit), target_arch = "x86_64", target_os = "darwin"))]
+global_asm! {
+    "
+    .global _global_asm_test
+    _global_asm_test:
+    // comment that would normally be removed by LLVM
+    ret
+    "
+}
+
 #[repr(C)]
 enum c_void {
     _1,
diff --git a/src/base.rs b/src/base.rs
index 2c04cf47268..c68d33465bc 100644
--- a/src/base.rs
+++ b/src/base.rs
@@ -5,6 +5,7 @@ use rustc_index::vec::IndexVec;
 use rustc_middle::ty::adjustment::PointerCast;
 use rustc_middle::ty::layout::FnAbiOf;
 use rustc_middle::ty::print::with_no_trimmed_paths;
+use rustc_middle::ty::SymbolName;
 
 use indexmap::IndexSet;
 
@@ -12,17 +13,42 @@ use crate::constant::ConstantCx;
 use crate::prelude::*;
 use crate::pretty_clif::CommentWriter;
 
-pub(crate) fn codegen_fn<'tcx>(
+struct CodegenedFunction<'tcx> {
+    instance: Instance<'tcx>,
+    symbol_name: SymbolName<'tcx>,
+    func_id: FuncId,
+    func: Function,
+    clif_comments: CommentWriter,
+    source_info_set: IndexSet<SourceInfo>,
+    local_map: IndexVec<mir::Local, CPlace<'tcx>>,
+}
+
+pub(crate) fn codegen_and_compile_fn<'tcx>(
     cx: &mut crate::CodegenCx<'tcx>,
+    cached_context: &mut Context,
     module: &mut dyn Module,
     instance: Instance<'tcx>,
 ) {
     let tcx = cx.tcx;
-
     let _inst_guard =
         crate::PrintOnPanic(|| format!("{:?} {}", instance, tcx.symbol_name(instance).name));
+
+    let cached_func = std::mem::replace(&mut cached_context.func, Function::new());
+    let codegened_func = codegen_fn(cx, cached_func, module, instance);
+
+    compile_fn(cx, cached_context, module, codegened_func);
+}
+
+fn codegen_fn<'tcx>(
+    cx: &mut crate::CodegenCx<'tcx>,
+    cached_func: Function,
+    module: &mut dyn Module,
+    instance: Instance<'tcx>,
+) -> CodegenedFunction<'tcx> {
     debug_assert!(!instance.substs.needs_infer());
 
+    let tcx = cx.tcx;
+
     let mir = tcx.instance_mir(instance.def);
     let _mir_guard = crate::PrintOnPanic(|| {
         let mut buf = Vec::new();
@@ -38,11 +64,10 @@ pub(crate) fn codegen_fn<'tcx>(
     let sig = get_function_sig(tcx, module.isa().triple(), instance);
     let func_id = module.declare_function(symbol_name.name, Linkage::Local, &sig).unwrap();
 
-    cx.cached_context.clear();
-
     // Make the FunctionBuilder
     let mut func_ctx = FunctionBuilderContext::new();
-    let mut func = std::mem::replace(&mut cx.cached_context.func, Function::new());
+    let mut func = cached_func;
+    func.clear();
     func.name = ExternalName::user(0, func_id.as_u32());
     func.signature = sig;
     func.collect_debug_info();
@@ -82,27 +107,7 @@ pub(crate) fn codegen_fn<'tcx>(
         next_ssa_var: 0,
     };
 
-    let arg_uninhabited = fx
-        .mir
-        .args_iter()
-        .any(|arg| fx.layout_of(fx.monomorphize(fx.mir.local_decls[arg].ty)).abi.is_uninhabited());
-
-    if !crate::constant::check_constants(&mut fx) {
-        fx.bcx.append_block_params_for_function_params(fx.block_map[START_BLOCK]);
-        fx.bcx.switch_to_block(fx.block_map[START_BLOCK]);
-        // compilation should have been aborted
-        fx.bcx.ins().trap(TrapCode::UnreachableCodeReached);
-    } else if arg_uninhabited {
-        fx.bcx.append_block_params_for_function_params(fx.block_map[START_BLOCK]);
-        fx.bcx.switch_to_block(fx.block_map[START_BLOCK]);
-        fx.bcx.ins().trap(TrapCode::UnreachableCodeReached);
-    } else {
-        tcx.sess.time("codegen clif ir", || {
-            tcx.sess
-                .time("codegen prelude", || crate::abi::codegen_fn_prelude(&mut fx, start_block));
-            codegen_fn_content(&mut fx);
-        });
-    }
+    tcx.sess.time("codegen clif ir", || codegen_fn_body(&mut fx, start_block));
 
     // Recover all necessary data from fx, before accessing func will prevent future access to it.
     let instance = fx.instance;
@@ -124,36 +129,31 @@ pub(crate) fn codegen_fn<'tcx>(
     // Verify function
     verify_func(tcx, &clif_comments, &func);
 
-    compile_fn(
-        cx,
-        module,
+    CodegenedFunction {
         instance,
-        symbol_name.name,
+        symbol_name,
         func_id,
         func,
         clif_comments,
         source_info_set,
         local_map,
-    );
+    }
 }
 
 fn compile_fn<'tcx>(
     cx: &mut crate::CodegenCx<'tcx>,
+    cached_context: &mut Context,
     module: &mut dyn Module,
-    instance: Instance<'tcx>,
-    symbol_name: &str,
-    func_id: FuncId,
-    func: Function,
-    mut clif_comments: CommentWriter,
-    source_info_set: IndexSet<SourceInfo>,
-    local_map: IndexVec<mir::Local, CPlace<'tcx>>,
+    codegened_func: CodegenedFunction<'tcx>,
 ) {
     let tcx = cx.tcx;
 
+    let mut clif_comments = codegened_func.clif_comments;
+
     // Store function in context
-    let context = &mut cx.cached_context;
+    let context = cached_context;
     context.clear();
-    context.func = func;
+    context.func = codegened_func.func;
 
     // If the return block is not reachable, then the SSA builder may have inserted an `iconst.i128`
     // instruction, which doesn't have an encoding.
@@ -170,7 +170,7 @@ fn compile_fn<'tcx>(
         crate::optimize::optimize_function(
             tcx,
             module.isa(),
-            instance,
+            codegened_func.instance,
             context,
             &mut clif_comments,
         );
@@ -206,7 +206,7 @@ fn compile_fn<'tcx>(
     // Define function
     tcx.sess.time("define function", || {
         context.want_disasm = crate::pretty_clif::should_write_ir(tcx);
-        module.define_function(func_id, context).unwrap();
+        module.define_function(codegened_func.func_id, context).unwrap();
     });
 
     // Write optimized function to file for debugging
@@ -214,7 +214,7 @@ fn compile_fn<'tcx>(
         tcx,
         "opt",
         module.isa(),
-        instance,
+        codegened_func.instance,
         &context.func,
         &clif_comments,
     );
@@ -222,7 +222,7 @@ fn compile_fn<'tcx>(
     if let Some(disasm) = &context.mach_compile_result.as_ref().unwrap().disasm {
         crate::pretty_clif::write_ir_file(
             tcx,
-            || format!("{}.vcode", tcx.symbol_name(instance).name),
+            || format!("{}.vcode", tcx.symbol_name(codegened_func.instance).name),
             |file| file.write_all(disasm.as_bytes()),
         )
     }
@@ -234,16 +234,16 @@ fn compile_fn<'tcx>(
     tcx.sess.time("generate debug info", || {
         if let Some(debug_context) = debug_context {
             debug_context.define_function(
-                instance,
-                func_id,
-                symbol_name,
+                codegened_func.instance,
+                codegened_func.func_id,
+                codegened_func.symbol_name.name,
                 isa,
                 context,
-                &source_info_set,
-                local_map,
+                &codegened_func.source_info_set,
+                codegened_func.local_map,
             );
         }
-        unwind_context.add_function(func_id, &context, isa);
+        unwind_context.add_function(codegened_func.func_id, &context, isa);
     });
 }
 
@@ -269,7 +269,27 @@ pub(crate) fn verify_func(
     });
 }
 
-fn codegen_fn_content(fx: &mut FunctionCx<'_, '_, '_>) {
+fn codegen_fn_body(fx: &mut FunctionCx<'_, '_, '_>, start_block: Block) {
+    if !crate::constant::check_constants(fx) {
+        fx.bcx.append_block_params_for_function_params(fx.block_map[START_BLOCK]);
+        fx.bcx.switch_to_block(fx.block_map[START_BLOCK]);
+        // compilation should have been aborted
+        fx.bcx.ins().trap(TrapCode::UnreachableCodeReached);
+        return;
+    }
+
+    let arg_uninhabited = fx
+        .mir
+        .args_iter()
+        .any(|arg| fx.layout_of(fx.monomorphize(fx.mir.local_decls[arg].ty)).abi.is_uninhabited());
+    if arg_uninhabited {
+        fx.bcx.append_block_params_for_function_params(fx.block_map[START_BLOCK]);
+        fx.bcx.switch_to_block(fx.block_map[START_BLOCK]);
+        fx.bcx.ins().trap(TrapCode::UnreachableCodeReached);
+        return;
+    }
+    fx.tcx.sess.time("codegen prelude", || crate::abi::codegen_fn_prelude(fx, start_block));
+
     for (bb, bb_data) in fx.mir.basic_blocks().iter_enumerated() {
         let block = fx.get_block(bb);
         fx.bcx.switch_to_block(block);
diff --git a/src/driver/aot.rs b/src/driver/aot.rs
index 3cd1ef5639e..9d819e3995b 100644
--- a/src/driver/aot.rs
+++ b/src/driver/aot.rs
@@ -1,25 +1,31 @@
 //! The AOT driver uses [`cranelift_object`] to write object files suitable for linking into a
 //! standalone executable.
 
+use std::fs::File;
 use std::path::PathBuf;
+use std::sync::Arc;
 
-use rustc_ast::{InlineAsmOptions, InlineAsmTemplatePiece};
 use rustc_codegen_ssa::back::metadata::create_compressed_metadata_file;
 use rustc_codegen_ssa::{CodegenResults, CompiledModule, CrateInfo, ModuleKind};
+use rustc_data_structures::profiling::SelfProfilerRef;
 use rustc_data_structures::stable_hasher::{HashStable, StableHasher};
 use rustc_metadata::EncodedMetadata;
 use rustc_middle::dep_graph::{WorkProduct, WorkProductId};
 use rustc_middle::mir::mono::{CodegenUnit, MonoItem};
 use rustc_session::cgu_reuse_tracker::CguReuse;
-use rustc_session::config::{DebugInfo, OutputType};
+use rustc_session::config::{DebugInfo, OutputFilenames, OutputType};
 use rustc_session::Session;
 
-use cranelift_codegen::isa::TargetIsa;
 use cranelift_object::{ObjectBuilder, ObjectModule};
 
+use crate::global_asm::GlobalAsmConfig;
 use crate::{prelude::*, BackendConfig};
 
-struct ModuleCodegenResult(CompiledModule, Option<(WorkProductId, WorkProduct)>);
+struct ModuleCodegenResult {
+    module_regular: CompiledModule,
+    module_global_asm: Option<CompiledModule>,
+    existing_work_product: Option<(WorkProductId, WorkProduct)>,
+}
 
 impl<HCX> HashStable<HCX> for ModuleCodegenResult {
     fn hash_stable(&self, _: &mut HCX, _: &mut StableHasher) {
@@ -27,7 +33,79 @@ impl<HCX> HashStable<HCX> for ModuleCodegenResult {
     }
 }
 
-fn make_module(sess: &Session, isa: Box<dyn TargetIsa>, name: String) -> ObjectModule {
+pub(crate) struct OngoingCodegen {
+    modules: Vec<Result<ModuleCodegenResult, String>>,
+    allocator_module: Option<CompiledModule>,
+    metadata_module: Option<CompiledModule>,
+    metadata: EncodedMetadata,
+    crate_info: CrateInfo,
+}
+
+impl OngoingCodegen {
+    pub(crate) fn join(
+        self,
+        sess: &Session,
+        backend_config: &BackendConfig,
+    ) -> (CodegenResults, FxHashMap<WorkProductId, WorkProduct>) {
+        let mut work_products = FxHashMap::default();
+        let mut modules = vec![];
+
+        for module_codegen_result in self.modules {
+            let module_codegen_result = match module_codegen_result {
+                Ok(module_codegen_result) => module_codegen_result,
+                Err(err) => sess.fatal(&err),
+            };
+            let ModuleCodegenResult { module_regular, module_global_asm, existing_work_product } =
+                module_codegen_result;
+
+            if let Some((work_product_id, work_product)) = existing_work_product {
+                work_products.insert(work_product_id, work_product);
+            } else {
+                let work_product = if backend_config.disable_incr_cache {
+                    None
+                } else if let Some(module_global_asm) = &module_global_asm {
+                    rustc_incremental::copy_cgu_workproduct_to_incr_comp_cache_dir(
+                        sess,
+                        &module_regular.name,
+                        &[
+                            ("o", &module_regular.object.as_ref().unwrap()),
+                            ("asm.o", &module_global_asm.object.as_ref().unwrap()),
+                        ],
+                    )
+                } else {
+                    rustc_incremental::copy_cgu_workproduct_to_incr_comp_cache_dir(
+                        sess,
+                        &module_regular.name,
+                        &[("o", &module_regular.object.as_ref().unwrap())],
+                    )
+                };
+                if let Some((work_product_id, work_product)) = work_product {
+                    work_products.insert(work_product_id, work_product);
+                }
+            }
+
+            modules.push(module_regular);
+            if let Some(module_global_asm) = module_global_asm {
+                modules.push(module_global_asm);
+            }
+        }
+
+        (
+            CodegenResults {
+                modules,
+                allocator_module: self.allocator_module,
+                metadata_module: self.metadata_module,
+                metadata: self.metadata,
+                crate_info: self.crate_info,
+            },
+            work_products,
+        )
+    }
+}
+
+fn make_module(sess: &Session, backend_config: &BackendConfig, name: String) -> ObjectModule {
+    let isa = crate::build_isa(sess, backend_config);
+
     let mut builder =
         ObjectBuilder::new(isa, name + ".o", cranelift_module::default_libcall_names()).unwrap();
     // Unlike cg_llvm, cg_clif defaults to disabling -Zfunction-sections. For cg_llvm binary size
@@ -37,15 +115,15 @@ fn make_module(sess: &Session, isa: Box<dyn TargetIsa>, name: String) -> ObjectM
     ObjectModule::new(builder)
 }
 
-fn emit_module(
-    tcx: TyCtxt<'_>,
-    backend_config: &BackendConfig,
+fn emit_cgu(
+    output_filenames: &OutputFilenames,
+    prof: &SelfProfilerRef,
     name: String,
-    kind: ModuleKind,
     module: ObjectModule,
     debug: Option<DebugContext<'_>>,
     unwind_context: UnwindContext,
-) -> ModuleCodegenResult {
+    global_asm_object_file: Option<PathBuf>,
+) -> Result<ModuleCodegenResult, String> {
     let mut product = module.finish();
 
     if let Some(mut debug) = debug {
@@ -54,71 +132,117 @@ fn emit_module(
 
     unwind_context.emit(&mut product);
 
-    let tmp_file = tcx.output_filenames(()).temp_path(OutputType::Object, Some(&name));
-    let obj = product.object.write().unwrap();
+    let module_regular =
+        emit_module(output_filenames, prof, product.object, ModuleKind::Regular, name.clone())?;
+
+    Ok(ModuleCodegenResult {
+        module_regular,
+        module_global_asm: global_asm_object_file.map(|global_asm_object_file| CompiledModule {
+            name: format!("{name}.asm"),
+            kind: ModuleKind::Regular,
+            object: Some(global_asm_object_file),
+            dwarf_object: None,
+            bytecode: None,
+        }),
+        existing_work_product: None,
+    })
+}
 
-    tcx.sess.prof.artifact_size("object_file", name.clone(), obj.len().try_into().unwrap());
+fn emit_module(
+    output_filenames: &OutputFilenames,
+    prof: &SelfProfilerRef,
+    object: cranelift_object::object::write::Object<'_>,
+    kind: ModuleKind,
+    name: String,
+) -> Result<CompiledModule, String> {
+    let tmp_file = output_filenames.temp_path(OutputType::Object, Some(&name));
+    let mut file = match File::create(&tmp_file) {
+        Ok(file) => file,
+        Err(err) => return Err(format!("error creating object file: {}", err)),
+    };
 
-    if let Err(err) = std::fs::write(&tmp_file, obj) {
-        tcx.sess.fatal(&format!("error writing object file: {}", err));
+    if let Err(err) = object.write_stream(&mut file) {
+        return Err(format!("error writing object file: {}", err));
     }
 
-    let work_product = if backend_config.disable_incr_cache {
-        None
-    } else {
-        rustc_incremental::copy_cgu_workproduct_to_incr_comp_cache_dir(
-            tcx.sess,
-            &name,
-            &[("o", &tmp_file)],
-        )
-    };
+    prof.artifact_size("object_file", &*name, file.metadata().unwrap().len());
 
-    ModuleCodegenResult(
-        CompiledModule { name, kind, object: Some(tmp_file), dwarf_object: None, bytecode: None },
-        work_product,
-    )
+    Ok(CompiledModule { name, kind, object: Some(tmp_file), dwarf_object: None, bytecode: None })
 }
 
 fn reuse_workproduct_for_cgu(
     tcx: TyCtxt<'_>,
     cgu: &CodegenUnit<'_>,
-    work_products: &mut FxHashMap<WorkProductId, WorkProduct>,
-) -> CompiledModule {
+) -> Result<ModuleCodegenResult, String> {
     let work_product = cgu.previous_work_product(tcx);
-    let obj_out = tcx.output_filenames(()).temp_path(OutputType::Object, Some(cgu.name().as_str()));
-    let source_file = rustc_incremental::in_incr_comp_dir_sess(
+    let obj_out_regular =
+        tcx.output_filenames(()).temp_path(OutputType::Object, Some(cgu.name().as_str()));
+    let source_file_regular = rustc_incremental::in_incr_comp_dir_sess(
         &tcx.sess,
         &work_product.saved_files.get("o").expect("no saved object file in work product"),
     );
-    if let Err(err) = rustc_fs_util::link_or_copy(&source_file, &obj_out) {
-        tcx.sess.err(&format!(
+
+    if let Err(err) = rustc_fs_util::link_or_copy(&source_file_regular, &obj_out_regular) {
+        return Err(format!(
             "unable to copy {} to {}: {}",
-            source_file.display(),
-            obj_out.display(),
+            source_file_regular.display(),
+            obj_out_regular.display(),
             err
         ));
     }
+    let obj_out_global_asm =
+        crate::global_asm::add_file_stem_postfix(obj_out_regular.clone(), ".asm");
+    let has_global_asm = if let Some(asm_o) = work_product.saved_files.get("asm.o") {
+        let source_file_global_asm = rustc_incremental::in_incr_comp_dir_sess(&tcx.sess, asm_o);
+        if let Err(err) = rustc_fs_util::link_or_copy(&source_file_global_asm, &obj_out_global_asm)
+        {
+            return Err(format!(
+                "unable to copy {} to {}: {}",
+                source_file_regular.display(),
+                obj_out_regular.display(),
+                err
+            ));
+        }
+        true
+    } else {
+        false
+    };
 
-    work_products.insert(cgu.work_product_id(), work_product);
-
-    CompiledModule {
-        name: cgu.name().to_string(),
-        kind: ModuleKind::Regular,
-        object: Some(obj_out),
-        dwarf_object: None,
-        bytecode: None,
-    }
+    Ok(ModuleCodegenResult {
+        module_regular: CompiledModule {
+            name: cgu.name().to_string(),
+            kind: ModuleKind::Regular,
+            object: Some(obj_out_regular),
+            dwarf_object: None,
+            bytecode: None,
+        },
+        module_global_asm: if has_global_asm {
+            Some(CompiledModule {
+                name: cgu.name().to_string(),
+                kind: ModuleKind::Regular,
+                object: Some(obj_out_global_asm),
+                dwarf_object: None,
+                bytecode: None,
+            })
+        } else {
+            None
+        },
+        existing_work_product: Some((cgu.work_product_id(), work_product)),
+    })
 }
 
 fn module_codegen(
     tcx: TyCtxt<'_>,
-    (backend_config, cgu_name): (BackendConfig, rustc_span::Symbol),
-) -> ModuleCodegenResult {
+    (backend_config, global_asm_config, cgu_name): (
+        BackendConfig,
+        Arc<GlobalAsmConfig>,
+        rustc_span::Symbol,
+    ),
+) -> Result<ModuleCodegenResult, String> {
     let cgu = tcx.codegen_unit(cgu_name);
     let mono_items = cgu.items_in_deterministic_order(tcx);
 
-    let isa = crate::build_isa(tcx.sess, &backend_config);
-    let mut module = make_module(tcx.sess, isa, cgu_name.as_str().to_string());
+    let mut module = make_module(tcx.sess, &backend_config, cgu_name.as_str().to_string());
 
     let mut cx = crate::CodegenCx::new(
         tcx,
@@ -128,32 +252,22 @@ fn module_codegen(
         cgu_name,
     );
     super::predefine_mono_items(tcx, &mut module, &mono_items);
+    let mut cached_context = Context::new();
     for (mono_item, _) in mono_items {
         match mono_item {
             MonoItem::Fn(inst) => {
-                cx.tcx
-                    .sess
-                    .time("codegen fn", || crate::base::codegen_fn(&mut cx, &mut module, inst));
+                cx.tcx.sess.time("codegen fn", || {
+                    crate::base::codegen_and_compile_fn(
+                        &mut cx,
+                        &mut cached_context,
+                        &mut module,
+                        inst,
+                    )
+                });
             }
             MonoItem::Static(def_id) => crate::constant::codegen_static(tcx, &mut module, def_id),
             MonoItem::GlobalAsm(item_id) => {
-                let item = cx.tcx.hir().item(item_id);
-                if let rustc_hir::ItemKind::GlobalAsm(asm) = item.kind {
-                    if !asm.options.contains(InlineAsmOptions::ATT_SYNTAX) {
-                        cx.global_asm.push_str("\n.intel_syntax noprefix\n");
-                    } else {
-                        cx.global_asm.push_str("\n.att_syntax\n");
-                    }
-                    for piece in asm.template {
-                        match *piece {
-                            InlineAsmTemplatePiece::String(ref s) => cx.global_asm.push_str(s),
-                            InlineAsmTemplatePiece::Placeholder { .. } => todo!(),
-                        }
-                    }
-                    cx.global_asm.push_str("\n.att_syntax\n\n");
-                } else {
-                    bug!("Expected GlobalAsm found {:?}", item);
-                }
+                crate::global_asm::codegen_global_asm_item(tcx, &mut cx.global_asm, item_id);
             }
         }
     }
@@ -165,23 +279,28 @@ fn module_codegen(
         cgu.is_primary(),
     );
 
+    let global_asm_object_file = match crate::global_asm::compile_global_asm(
+        &global_asm_config,
+        cgu.name().as_str(),
+        &cx.global_asm,
+    ) {
+        Ok(global_asm_object_file) => global_asm_object_file,
+        Err(err) => tcx.sess.fatal(&err),
+    };
+
     let debug_context = cx.debug_context;
     let unwind_context = cx.unwind_context;
-    let codegen_result = tcx.sess.time("write object file", || {
-        emit_module(
-            tcx,
-            &backend_config,
+    tcx.sess.time("write object file", || {
+        emit_cgu(
+            &global_asm_config.output_filenames,
+            &tcx.sess.prof,
             cgu.name().as_str().to_string(),
-            ModuleKind::Regular,
             module,
             debug_context,
             unwind_context,
+            global_asm_object_file,
         )
-    });
-
-    codegen_global_asm(tcx, cgu.name().as_str(), &cx.global_asm);
-
-    codegen_result
+    })
 }
 
 pub(crate) fn run_aot(
@@ -189,9 +308,7 @@ pub(crate) fn run_aot(
     backend_config: BackendConfig,
     metadata: EncodedMetadata,
     need_metadata_module: bool,
-) -> Box<(CodegenResults, FxHashMap<WorkProductId, WorkProduct>)> {
-    let mut work_products = FxHashMap::default();
-
+) -> Box<OngoingCodegen> {
     let cgus = if tcx.sess.opts.output_types.should_codegen() {
         tcx.collect_and_partition_mono_items(()).1
     } else {
@@ -206,62 +323,59 @@ pub(crate) fn run_aot(
         }
     }
 
+    let global_asm_config = Arc::new(crate::global_asm::GlobalAsmConfig::new(tcx));
+
     let modules = super::time(tcx, backend_config.display_cg_time, "codegen mono items", || {
         cgus.iter()
             .map(|cgu| {
-                let cgu_reuse = determine_cgu_reuse(tcx, cgu);
+                let cgu_reuse = if backend_config.disable_incr_cache {
+                    CguReuse::No
+                } else {
+                    determine_cgu_reuse(tcx, cgu)
+                };
                 tcx.sess.cgu_reuse_tracker.set_actual_reuse(cgu.name().as_str(), cgu_reuse);
 
                 match cgu_reuse {
-                    _ if backend_config.disable_incr_cache => {}
-                    CguReuse::No => {}
-                    CguReuse::PreLto => {
-                        return reuse_workproduct_for_cgu(tcx, &*cgu, &mut work_products);
+                    CguReuse::No => {
+                        let dep_node = cgu.codegen_dep_node(tcx);
+                        tcx.dep_graph
+                            .with_task(
+                                dep_node,
+                                tcx,
+                                (backend_config.clone(), global_asm_config.clone(), cgu.name()),
+                                module_codegen,
+                                Some(rustc_middle::dep_graph::hash_result),
+                            )
+                            .0
                     }
+                    CguReuse::PreLto => reuse_workproduct_for_cgu(tcx, &*cgu),
                     CguReuse::PostLto => unreachable!(),
                 }
-
-                let dep_node = cgu.codegen_dep_node(tcx);
-                let (ModuleCodegenResult(module, work_product), _) = tcx.dep_graph.with_task(
-                    dep_node,
-                    tcx,
-                    (backend_config.clone(), cgu.name()),
-                    module_codegen,
-                    Some(rustc_middle::dep_graph::hash_result),
-                );
-
-                if let Some((id, product)) = work_product {
-                    work_products.insert(id, product);
-                }
-
-                module
             })
             .collect::<Vec<_>>()
     });
 
     tcx.sess.abort_if_errors();
 
-    let isa = crate::build_isa(tcx.sess, &backend_config);
-    let mut allocator_module = make_module(tcx.sess, isa, "allocator_shim".to_string());
-    assert_eq!(pointer_ty(tcx), allocator_module.target_config().pointer_type());
+    let mut allocator_module = make_module(tcx.sess, &backend_config, "allocator_shim".to_string());
     let mut allocator_unwind_context = UnwindContext::new(allocator_module.isa(), true);
     let created_alloc_shim =
         crate::allocator::codegen(tcx, &mut allocator_module, &mut allocator_unwind_context);
 
     let allocator_module = if created_alloc_shim {
-        let ModuleCodegenResult(module, work_product) = emit_module(
-            tcx,
-            &backend_config,
-            "allocator_shim".to_string(),
+        let mut product = allocator_module.finish();
+        allocator_unwind_context.emit(&mut product);
+
+        match emit_module(
+            tcx.output_filenames(()),
+            &tcx.sess.prof,
+            product.object,
             ModuleKind::Allocator,
-            allocator_module,
-            None,
-            allocator_unwind_context,
-        );
-        if let Some((id, product)) = work_product {
-            work_products.insert(id, product);
+            "allocator_shim".to_owned(),
+        ) {
+            Ok(allocator_module) => Some(allocator_module),
+            Err(err) => tcx.sess.fatal(err),
         }
-        Some(module)
     } else {
         None
     };
@@ -308,102 +422,13 @@ pub(crate) fn run_aot(
     }
     .to_owned();
 
-    Box::new((
-        CodegenResults {
-            modules,
-            allocator_module,
-            metadata_module,
-            metadata,
-            crate_info: CrateInfo::new(tcx, target_cpu),
-        },
-        work_products,
-    ))
-}
-
-fn codegen_global_asm(tcx: TyCtxt<'_>, cgu_name: &str, global_asm: &str) {
-    use std::io::Write;
-    use std::process::{Command, Stdio};
-
-    if global_asm.is_empty() {
-        return;
-    }
-
-    if cfg!(not(feature = "inline_asm"))
-        || tcx.sess.target.is_like_osx
-        || tcx.sess.target.is_like_windows
-    {
-        if global_asm.contains("__rust_probestack") {
-            return;
-        }
-
-        // FIXME fix linker error on macOS
-        if cfg!(not(feature = "inline_asm")) {
-            tcx.sess.fatal(
-                "asm! and global_asm! support is disabled while compiling rustc_codegen_cranelift",
-            );
-        } else {
-            tcx.sess.fatal("asm! and global_asm! are not yet supported on macOS and Windows");
-        }
-    }
-
-    let assembler = crate::toolchain::get_toolchain_binary(tcx.sess, "as");
-    let linker = crate::toolchain::get_toolchain_binary(tcx.sess, "ld");
-
-    // Remove all LLVM style comments
-    let global_asm = global_asm
-        .lines()
-        .map(|line| if let Some(index) = line.find("//") { &line[0..index] } else { line })
-        .collect::<Vec<_>>()
-        .join("\n");
-
-    let output_object_file = tcx.output_filenames(()).temp_path(OutputType::Object, Some(cgu_name));
-
-    // Assemble `global_asm`
-    let global_asm_object_file = add_file_stem_postfix(output_object_file.clone(), ".asm");
-    let mut child = Command::new(assembler)
-        .arg("-o")
-        .arg(&global_asm_object_file)
-        .stdin(Stdio::piped())
-        .spawn()
-        .expect("Failed to spawn `as`.");
-    child.stdin.take().unwrap().write_all(global_asm.as_bytes()).unwrap();
-    let status = child.wait().expect("Failed to wait for `as`.");
-    if !status.success() {
-        tcx.sess.fatal(&format!("Failed to assemble `{}`", global_asm));
-    }
-
-    // Link the global asm and main object file together
-    let main_object_file = add_file_stem_postfix(output_object_file.clone(), ".main");
-    std::fs::rename(&output_object_file, &main_object_file).unwrap();
-    let status = Command::new(linker)
-        .arg("-r") // Create a new object file
-        .arg("-o")
-        .arg(output_object_file)
-        .arg(&main_object_file)
-        .arg(&global_asm_object_file)
-        .status()
-        .unwrap();
-    if !status.success() {
-        tcx.sess.fatal(&format!(
-            "Failed to link `{}` and `{}` together",
-            main_object_file.display(),
-            global_asm_object_file.display(),
-        ));
-    }
-
-    std::fs::remove_file(global_asm_object_file).unwrap();
-    std::fs::remove_file(main_object_file).unwrap();
-}
-
-fn add_file_stem_postfix(mut path: PathBuf, postfix: &str) -> PathBuf {
-    let mut new_filename = path.file_stem().unwrap().to_owned();
-    new_filename.push(postfix);
-    if let Some(extension) = path.extension() {
-        new_filename.push(".");
-        new_filename.push(extension);
-    }
-    path.set_file_name(new_filename);
-    path
+    Box::new(OngoingCodegen {
+        modules,
+        allocator_module,
+        metadata_module,
+        metadata,
+        crate_info: CrateInfo::new(tcx, target_cpu),
+    })
 }
 
 // Adapted from https://github.com/rust-lang/rust/blob/303d8aff6092709edd4dbd35b1c88e9aa40bf6d8/src/librustc_codegen_ssa/base.rs#L922-L953
diff --git a/src/driver/jit.rs b/src/driver/jit.rs
index a56a9100059..1b046d7ec6e 100644
--- a/src/driver/jit.rs
+++ b/src/driver/jit.rs
@@ -111,6 +111,7 @@ pub(crate) fn run_jit(tcx: TyCtxt<'_>, backend_config: BackendConfig) -> ! {
         &backend_config,
         matches!(backend_config.codegen_mode, CodegenMode::JitLazy),
     );
+    let mut cached_context = Context::new();
 
     let (_, cgus) = tcx.collect_and_partition_mono_items(());
     let mono_items = cgus
@@ -129,10 +130,17 @@ pub(crate) fn run_jit(tcx: TyCtxt<'_>, backend_config: BackendConfig) -> ! {
                     CodegenMode::Aot => unreachable!(),
                     CodegenMode::Jit => {
                         cx.tcx.sess.time("codegen fn", || {
-                            crate::base::codegen_fn(&mut cx, &mut jit_module, inst)
+                            crate::base::codegen_and_compile_fn(
+                                &mut cx,
+                                &mut cached_context,
+                                &mut jit_module,
+                                inst,
+                            )
                         });
                     }
-                    CodegenMode::JitLazy => codegen_shim(&mut cx, &mut jit_module, inst),
+                    CodegenMode::JitLazy => {
+                        codegen_shim(&mut cx, &mut cached_context, &mut jit_module, inst)
+                    }
                 },
                 MonoItem::Static(def_id) => {
                     crate::constant::codegen_static(tcx, &mut jit_module, def_id);
@@ -259,7 +267,14 @@ fn jit_fn(instance_ptr: *const Instance<'static>, trampoline_ptr: *const u8) ->
                 false,
                 Symbol::intern("dummy_cgu_name"),
             );
-            tcx.sess.time("codegen fn", || crate::base::codegen_fn(&mut cx, jit_module, instance));
+            tcx.sess.time("codegen fn", || {
+                crate::base::codegen_and_compile_fn(
+                    &mut cx,
+                    &mut Context::new(),
+                    jit_module,
+                    instance,
+                )
+            });
 
             assert!(cx.global_asm.is_empty());
             jit_module.finalize_definitions();
@@ -334,7 +349,12 @@ fn load_imported_symbols_for_jit(
     imported_symbols
 }
 
-fn codegen_shim<'tcx>(cx: &mut CodegenCx<'tcx>, module: &mut JITModule, inst: Instance<'tcx>) {
+fn codegen_shim<'tcx>(
+    cx: &mut CodegenCx<'tcx>,
+    cached_context: &mut Context,
+    module: &mut JITModule,
+    inst: Instance<'tcx>,
+) {
     let tcx = cx.tcx;
 
     let pointer_type = module.target_config().pointer_type();
@@ -357,8 +377,9 @@ fn codegen_shim<'tcx>(cx: &mut CodegenCx<'tcx>, module: &mut JITModule, inst: In
         )
         .unwrap();
 
-    cx.cached_context.clear();
-    let trampoline = &mut cx.cached_context.func;
+    let context = cached_context;
+    context.clear();
+    let trampoline = &mut context.func;
     trampoline.signature = sig.clone();
 
     let mut builder_ctx = FunctionBuilderContext::new();
@@ -381,5 +402,5 @@ fn codegen_shim<'tcx>(cx: &mut CodegenCx<'tcx>, module: &mut JITModule, inst: In
     let ret_vals = trampoline_builder.func.dfg.inst_results(call_inst).to_vec();
     trampoline_builder.ins().return_(&ret_vals);
 
-    module.define_function(func_id, &mut cx.cached_context).unwrap();
+    module.define_function(func_id, context).unwrap();
 }
diff --git a/src/global_asm.rs b/src/global_asm.rs
new file mode 100644
index 00000000000..dcbcaba30fe
--- /dev/null
+++ b/src/global_asm.rs
@@ -0,0 +1,114 @@
+//! The AOT driver uses [`cranelift_object`] to write object files suitable for linking into a
+//! standalone executable.
+
+use std::io::Write;
+use std::path::PathBuf;
+use std::process::{Command, Stdio};
+use std::sync::Arc;
+
+use rustc_ast::{InlineAsmOptions, InlineAsmTemplatePiece};
+use rustc_hir::ItemId;
+use rustc_session::config::{OutputFilenames, OutputType};
+
+use crate::prelude::*;
+
+pub(crate) fn codegen_global_asm_item(tcx: TyCtxt<'_>, global_asm: &mut String, item_id: ItemId) {
+    let item = tcx.hir().item(item_id);
+    if let rustc_hir::ItemKind::GlobalAsm(asm) = item.kind {
+        if !asm.options.contains(InlineAsmOptions::ATT_SYNTAX) {
+            global_asm.push_str("\n.intel_syntax noprefix\n");
+        } else {
+            global_asm.push_str("\n.att_syntax\n");
+        }
+        for piece in asm.template {
+            match *piece {
+                InlineAsmTemplatePiece::String(ref s) => global_asm.push_str(s),
+                InlineAsmTemplatePiece::Placeholder { .. } => todo!(),
+            }
+        }
+        global_asm.push_str("\n.att_syntax\n\n");
+    } else {
+        bug!("Expected GlobalAsm found {:?}", item);
+    }
+}
+
+#[derive(Debug)]
+pub(crate) struct GlobalAsmConfig {
+    asm_enabled: bool,
+    assembler: PathBuf,
+    pub(crate) output_filenames: Arc<OutputFilenames>,
+}
+
+impl GlobalAsmConfig {
+    pub(crate) fn new(tcx: TyCtxt<'_>) -> Self {
+        let asm_enabled = cfg!(feature = "inline_asm") && !tcx.sess.target.is_like_windows;
+
+        GlobalAsmConfig {
+            asm_enabled,
+            assembler: crate::toolchain::get_toolchain_binary(tcx.sess, "as"),
+            output_filenames: tcx.output_filenames(()).clone(),
+        }
+    }
+}
+
+pub(crate) fn compile_global_asm(
+    config: &GlobalAsmConfig,
+    cgu_name: &str,
+    global_asm: &str,
+) -> Result<Option<PathBuf>, String> {
+    if global_asm.is_empty() {
+        return Ok(None);
+    }
+
+    if !config.asm_enabled {
+        if global_asm.contains("__rust_probestack") {
+            return Ok(None);
+        }
+
+        // FIXME fix linker error on macOS
+        if cfg!(not(feature = "inline_asm")) {
+            return Err(
+                "asm! and global_asm! support is disabled while compiling rustc_codegen_cranelift"
+                    .to_owned(),
+            );
+        } else {
+            return Err("asm! and global_asm! are not yet supported on Windows".to_owned());
+        }
+    }
+
+    // Remove all LLVM style comments
+    let global_asm = global_asm
+        .lines()
+        .map(|line| if let Some(index) = line.find("//") { &line[0..index] } else { line })
+        .collect::<Vec<_>>()
+        .join("\n");
+
+    let output_object_file = config.output_filenames.temp_path(OutputType::Object, Some(cgu_name));
+
+    // Assemble `global_asm`
+    let global_asm_object_file = add_file_stem_postfix(output_object_file.clone(), ".asm");
+    let mut child = Command::new(&config.assembler)
+        .arg("-o")
+        .arg(&global_asm_object_file)
+        .stdin(Stdio::piped())
+        .spawn()
+        .expect("Failed to spawn `as`.");
+    child.stdin.take().unwrap().write_all(global_asm.as_bytes()).unwrap();
+    let status = child.wait().expect("Failed to wait for `as`.");
+    if !status.success() {
+        return Err(format!("Failed to assemble `{}`", global_asm));
+    }
+
+    Ok(Some(global_asm_object_file))
+}
+
+pub(crate) fn add_file_stem_postfix(mut path: PathBuf, postfix: &str) -> PathBuf {
+    let mut new_filename = path.file_stem().unwrap().to_owned();
+    new_filename.push(postfix);
+    if let Some(extension) = path.extension() {
+        new_filename.push(".");
+        new_filename.push(extension);
+    }
+    path.set_file_name(new_filename);
+    path
+}
diff --git a/src/lib.rs b/src/lib.rs
index bb0793b1deb..909f4f00f1e 100644
--- a/src/lib.rs
+++ b/src/lib.rs
@@ -25,7 +25,7 @@ extern crate rustc_target;
 extern crate rustc_driver;
 
 use std::any::Any;
-use std::cell::Cell;
+use std::cell::{Cell, RefCell};
 
 use rustc_codegen_ssa::traits::CodegenBackend;
 use rustc_codegen_ssa::CodegenResults;
@@ -56,6 +56,7 @@ mod constant;
 mod debuginfo;
 mod discriminant;
 mod driver;
+mod global_asm;
 mod inline_asm;
 mod intrinsics;
 mod linkage;
@@ -123,7 +124,6 @@ struct CodegenCx<'tcx> {
     tcx: TyCtxt<'tcx>,
     global_asm: String,
     inline_asm_index: Cell<usize>,
-    cached_context: Context,
     debug_context: Option<DebugContext<'tcx>>,
     unwind_context: UnwindContext,
     cgu_name: Symbol,
@@ -150,7 +150,6 @@ impl<'tcx> CodegenCx<'tcx> {
             tcx,
             global_asm: String::new(),
             inline_asm_index: Cell::new(0),
-            cached_context: Context::new(),
             debug_context,
             unwind_context,
             cgu_name,
@@ -159,7 +158,7 @@ impl<'tcx> CodegenCx<'tcx> {
 }
 
 pub struct CraneliftCodegenBackend {
-    pub config: Option<BackendConfig>,
+    pub config: RefCell<Option<BackendConfig>>,
 }
 
 impl CodegenBackend for CraneliftCodegenBackend {
@@ -169,6 +168,13 @@ impl CodegenBackend for CraneliftCodegenBackend {
             Lto::No | Lto::ThinLocal => {}
             Lto::Thin | Lto::Fat => sess.warn("LTO is not supported. You may get a linker error."),
         }
+
+        let mut config = self.config.borrow_mut();
+        if config.is_none() {
+            let new_config = BackendConfig::from_opts(&sess.opts.cg.llvm_args)
+                .unwrap_or_else(|err| sess.fatal(&err));
+            *config = Some(new_config);
+        }
     }
 
     fn target_features(&self, _sess: &Session, _allow_unstable: bool) -> Vec<rustc_span::Symbol> {
@@ -186,15 +192,7 @@ impl CodegenBackend for CraneliftCodegenBackend {
         need_metadata_module: bool,
     ) -> Box<dyn Any> {
         tcx.sess.abort_if_errors();
-        let config = if let Some(config) = self.config.clone() {
-            config
-        } else {
-            if !tcx.sess.unstable_options() && !tcx.sess.opts.cg.llvm_args.is_empty() {
-                tcx.sess.fatal("`-Z unstable-options` must be passed to allow configuring cg_clif");
-            }
-            BackendConfig::from_opts(&tcx.sess.opts.cg.llvm_args)
-                .unwrap_or_else(|err| tcx.sess.fatal(&err))
-        };
+        let config = self.config.borrow().clone().unwrap();
         match config.codegen_mode {
             CodegenMode::Aot => driver::aot::run_aot(tcx, config, metadata, need_metadata_module),
             CodegenMode::Jit | CodegenMode::JitLazy => {
@@ -210,12 +208,13 @@ impl CodegenBackend for CraneliftCodegenBackend {
     fn join_codegen(
         &self,
         ongoing_codegen: Box<dyn Any>,
-        _sess: &Session,
+        sess: &Session,
         _outputs: &OutputFilenames,
     ) -> Result<(CodegenResults, FxHashMap<WorkProductId, WorkProduct>), ErrorGuaranteed> {
-        Ok(*ongoing_codegen
-            .downcast::<(CodegenResults, FxHashMap<WorkProductId, WorkProduct>)>()
-            .unwrap())
+        Ok(ongoing_codegen
+            .downcast::<driver::aot::OngoingCodegen>()
+            .unwrap()
+            .join(sess, self.config.borrow().as_ref().unwrap()))
     }
 
     fn link(
@@ -312,5 +311,5 @@ fn build_isa(sess: &Session, backend_config: &BackendConfig) -> Box<dyn isa::Tar
 /// This is the entrypoint for a hot plugged rustc_codegen_cranelift
 #[no_mangle]
 pub fn __rustc_codegen_backend() -> Box<dyn CodegenBackend> {
-    Box::new(CraneliftCodegenBackend { config: None })
+    Box::new(CraneliftCodegenBackend { config: RefCell::new(None) })
 }
diff --git a/src/toolchain.rs b/src/toolchain.rs
index f86236ef3ea..b6b465e1f4e 100644
--- a/src/toolchain.rs
+++ b/src/toolchain.rs
@@ -8,10 +8,8 @@ use rustc_session::Session;
 /// Tries to infer the path of a binary for the target toolchain from the linker name.
 pub(crate) fn get_toolchain_binary(sess: &Session, tool: &str) -> PathBuf {
     let (mut linker, _linker_flavor) = linker_and_flavor(sess);
-    let linker_file_name = linker
-        .file_name()
-        .and_then(|name| name.to_str())
-        .unwrap_or_else(|| sess.fatal("couldn't extract file name from specified linker"));
+    let linker_file_name =
+        linker.file_name().unwrap().to_str().expect("linker filename should be valid UTF-8");
 
     if linker_file_name == "ld.lld" {
         if tool != "ld" {