[WIP] Implement LTO support

Serializing and deserializing Cranelift IR is supported, but due to the
lack of interprocedural optimizations in Cranelift there is no runtime
perf benefit. Only a compile time hit. It may still be useful for things
like the JIT mode where invoking a regular linker is not possible.

Cargo.toml

@@ -16,11 +16,13 @@ cranelift-jit = { version = "0.114.0", optional = true }
 cranelift-object = { version = "0.114.0" }
 target-lexicon = "0.12.0"
 gimli = { version = "0.31", default-features = false, features = ["write"] }
-object = { version = "0.36", default-features = false, features = ["std", "read_core", "write", "archive", "coff", "elf", "macho", "pe"] }
+object = { version = "0.36", default-features = false, features = ["std", "read_core", "write", "archive", "coff", "elf", "macho", "pe", "unaligned"] }
 
 indexmap = "2.0.0"
 libloading = { version = "0.8.0", optional = true }
 smallvec = "1.8.1"
+serde = { version = "1.0.203", features = ["derive"], optional = true }
+postcard = { version = "1.0.8", default-features = false, features = ["use-std"], optional = true }
 
 [patch.crates-io]
 # Uncomment to use local checkout of cranelift
@@ -35,8 +37,9 @@ smallvec = "1.8.1"
 
 [features]
 # Enable features not ready to be enabled when compiling as part of rustc
-unstable-features = ["jit", "inline_asm_sym"]
+unstable-features = ["jit", "inline_asm_sym", "lto"]
 jit = ["cranelift-jit", "libloading"]
+lto = ["serde", "postcard", "cranelift-codegen/enable-serde", "cranelift-module/enable-serde"]
 inline_asm_sym = []
 
 [package.metadata.rust-analyzer]

build_system/build_sysroot.rs

@@ -243,6 +243,9 @@ fn build_clif_sysroot_for_triple(
             prefix.to_str().unwrap()
         ));
     }
+    rustflags.push("-Clto=thin".to_owned());
+    rustflags.push("-Zdylib-lto".to_owned());
+    rustflags.push("-Cembed-bitcode=yes".to_owned());
     compiler.rustflags.extend(rustflags);
     let mut build_cmd = STANDARD_LIBRARY.build(&compiler, dirs);
     build_cmd.arg("--release");
@@ -253,6 +256,7 @@ fn build_clif_sysroot_for_triple(
     if compiler.triple.contains("apple") {
         build_cmd.env("CARGO_PROFILE_RELEASE_SPLIT_DEBUGINFO", "packed");
     }
+    build_cmd.env("CARGO_PROFILE_RELEASE_LTO", "thin");
     spawn_and_wait(build_cmd);
 
     for entry in fs::read_dir(build_dir.join("deps")).unwrap() {

build_system/tests.rs

@@ -57,7 +57,7 @@ impl TestCase {
 }
 
 const NO_SYSROOT_SUITE: &[TestCase] = &[
-    TestCase::build_lib("build.mini_core", "example/mini_core.rs", "lib,dylib"),
+    TestCase::build_lib("build.mini_core", "example/mini_core.rs", "lib"),
     TestCase::build_lib("build.example", "example/example.rs", "lib"),
     TestCase::jit_bin("jit.mini_core_hello_world", "example/mini_core_hello_world.rs", "abc bcd"),
     TestCase::build_bin_and_run(
@@ -400,6 +400,7 @@ impl<'a> TestRunner<'a> {
                     }
                     spawn_and_wait(jit_cmd);
 
+                    /*
                     eprintln!("[JIT-lazy] {testname}");
                     let mut jit_cmd = self.rustc_command([
                         "-Zunstable-options",
@@ -413,6 +414,7 @@ impl<'a> TestRunner<'a> {
                         jit_cmd.env("CG_CLIF_JIT_ARGS", args);
                     }
                     spawn_and_wait(jit_cmd);
+                    */
                 }
             }
         }
@@ -431,6 +433,7 @@ impl<'a> TestRunner<'a> {
         cmd.arg("--out-dir");
         cmd.arg(BUILD_EXAMPLE_OUT_DIR.to_path(&self.dirs));
         cmd.arg("-Cdebuginfo=2");
+        cmd.arg("-Clto=thin");
         cmd.arg("--target");
         cmd.arg(&self.target_compiler.triple);
         cmd.arg("-Cpanic=abort");

example/mini_core.rs

@@ -786,6 +786,7 @@ struct PanicLocation {
     column: u32,
 }
 
+/*
 #[no_mangle]
 #[cfg(not(all(windows, target_env = "gnu")))]
 pub fn get_tls() -> u8 {
@@ -794,3 +795,4 @@ pub fn get_tls() -> u8 {
 
     A
 }
+*/

example/mini_core_hello_world.rs

@@ -332,11 +332,13 @@ fn main() {
     #[cfg(all(not(jit), not(all(windows, target_env = "gnu"))))]
     test_tls();
 
+    /*
     #[cfg(all(not(jit), target_arch = "x86_64", any(target_os = "linux", target_os = "macos")))]
     unsafe {
         global_asm_test();
         naked_test();
     }
+    */
 
     // Both statics have a reference that points to the same anonymous allocation.
     static REF1: &u8 = &42;
@@ -361,6 +363,7 @@ fn stack_val_align() {
     assert_eq!(&a as *const Foo as usize % 8192, 0);
 }
 
+/*
 #[cfg(all(not(jit), target_arch = "x86_64", any(target_os = "linux", target_os = "macos")))]
 extern "C" {
     fn global_asm_test();
@@ -385,6 +388,7 @@ global_asm! {
     ret
     "
 }
+*/
 
 #[cfg(all(not(jit), target_arch = "x86_64"))]
 #[naked]

patches/0029-stdlib-dylib-crate-type.patch

@@ -0,0 +1,25 @@
+From 0910cbe862990b0c3a17c67bca199ebb4452b0ec Mon Sep 17 00:00:00 2001
+From: bjorn3 <[email protected]>
+Date: Tue, 28 Mar 2023 17:09:01 +0000
+Subject: [PATCH] Disable dylib crate type
+
+---
+ library/std/Cargo.toml  | 2 +-
+ 1 files changed, 1 insertions(+), 1 deletions(-)
+
+diff --git a/library/std/Cargo.toml b/library/std/Cargo.toml
+index 598a4bf..3e68680 100644
+--- a/library/std/Cargo.toml
++++ b/library/std/Cargo.toml
+@@ -7,7 +7,7 @@ description = "The Rust Standard Library"
+ edition = "2021"
+
+ [lib]
+-crate-type = ["dylib", "rlib"]
++crate-type = ["rlib"]
+
+ [dependencies]
+ alloc = { path = "../alloc", public = true }
+-- 
+2.34.1
+

src/base.rs

@@ -252,6 +252,7 @@ pub(crate) fn compile_fn(
         }
     }
 
+    /*
     // Define debuginfo for function
     let debug_context = &mut cx.debug_context;
     cx.profiler.generic_activity("generate debug info").run(|| {
@@ -263,6 +264,7 @@ pub(crate) fn compile_fn(
             );
         }
     });
+    */
 }
 
 fn verify_func(tcx: TyCtxt<'_>, writer: &crate::pretty_clif::CommentWriter, func: &Function) {

src/driver/aot.rs

@@ -141,7 +141,7 @@ impl OngoingCodegen {
 }
 
 // Adapted from https://github.com/rust-lang/rust/blob/73476d49904751f8d90ce904e16dfbc278083d2c/compiler/rustc_codegen_ssa/src/back/write.rs#L547C1-L706C2
-fn produce_final_output_artifacts(
+pub(super) fn produce_final_output_artifacts(
     sess: &Session,
     codegen_results: &CodegenResults,
     crate_output: &OutputFilenames,
@@ -328,7 +328,7 @@ fn produce_final_output_artifacts(
     // These are used in linking steps and will be cleaned up afterward.
 }
 
-fn make_module(sess: &Session, name: String) -> UnwindModule<ObjectModule> {
+pub(super) fn make_module(sess: &Session, name: String) -> UnwindModule<ObjectModule> {
     let isa = crate::build_isa(sess);
 
     let mut builder =
@@ -379,7 +379,7 @@ fn emit_cgu(
     })
 }
 
-fn emit_module(
+pub(super) fn emit_module(
     output_filenames: &OutputFilenames,
     prof: &SelfProfilerRef,
     mut object: cranelift_object::object::write::Object<'_>,
@@ -488,7 +488,7 @@ fn reuse_workproduct_for_cgu(
     })
 }
 
-fn codegen_cgu_content(
+pub(super) fn codegen_cgu_content(
     tcx: TyCtxt<'_>,
     module: &mut dyn Module,
     cgu_name: rustc_span::Symbol,
@@ -680,7 +680,7 @@ pub(crate) fn run_aot(
     })
 }
 
-fn emit_allocator_module(tcx: TyCtxt<'_>) -> Option<CompiledModule> {
+pub(super) fn emit_allocator_module(tcx: TyCtxt<'_>) -> Option<CompiledModule> {
     let mut allocator_module = make_module(tcx.sess, "allocator_shim".to_string());
     let created_alloc_shim = crate::allocator::codegen(tcx, &mut allocator_module);
 
@@ -703,7 +703,7 @@ fn emit_allocator_module(tcx: TyCtxt<'_>) -> Option<CompiledModule> {
     }
 }
 
-fn emit_metadata_module(tcx: TyCtxt<'_>, metadata: &EncodedMetadata) -> CompiledModule {
+pub(super) fn emit_metadata_module(tcx: TyCtxt<'_>, metadata: &EncodedMetadata) -> CompiledModule {
     tcx.sess.time("write compressed metadata", || {
         use rustc_middle::mir::mono::CodegenUnitNameBuilder;
 

src/driver/jit.rs

@@ -66,10 +66,15 @@ fn create_jit_module(tcx: TyCtxt<'_>, hotswap: bool) -> (UnwindModule<JITModule>
     let mut jit_builder = JITBuilder::with_isa(isa, cranelift_module::default_libcall_names());
     jit_builder.hotswap(hotswap);
     crate::compiler_builtins::register_functions_for_jit(&mut jit_builder);
-    jit_builder.symbol_lookup_fn(dep_symbol_lookup_fn(tcx.sess, crate_info));
+    //jit_builder.symbol_lookup_fn(dep_symbol_lookup_fn(tcx.sess, crate_info));
     jit_builder.symbol("__clif_jit_fn", clif_jit_fn as *const u8);
     let mut jit_module = UnwindModule::new(JITModule::new(jit_builder), false);
 
+    #[cfg(feature = "lto")]
+    for (_name, module) in super::lto::load_lto_modules(tcx, &crate_info) {
+        module.apply_to(&mut jit_module);
+    }
+
     let cx = crate::CodegenCx::new(tcx, jit_module.isa(), false, sym::dummy_cgu_name);
 
     crate::allocator::codegen(tcx, &mut jit_module);