Auto merge of rust-lang#136840 - Flakebi:linker-plugin-lto-fat, r=<try>

Fix linker-plugin-lto only doing thin lto

When rust provides LLVM bitcode files to lld and the bitcode contains
function summaries as used for thin lto, lld defaults to using thin lto.
This prevents some optimizations that are only applied for fat lto.

Set the `ThinLTO=0` module flag to signal lld to do fat lto.

The analogous code in clang that sets this flag is here:
https://github.com/llvm/llvm-project/blob/560149b5e3c891c64899e9912e29467a69dc3a4c/clang/lib/CodeGen/BackendUtil.cpp#L1150

The code in LLVM that queries the flag and defaults to thin lto if not
set is here:
https://github.com/llvm/llvm-project/blob/e258bca9505f35e0a22cb213a305eea9b76d11ea/llvm/lib/Bitcode/Writer/BitcodeWriter.cpp#L4441-L4446

r? `@workingjubilee,` as you’ve been reviewing most other amdgpu patches, not sure if there should be other reviewers for lto.

Author: bors

compiler/rustc_codegen_llvm/src/context.rs

@@ -21,10 +21,10 @@ use rustc_middle::ty::layout::{
 };
 use rustc_middle::ty::{self, Instance, Ty, TyCtxt};
 use rustc_middle::{bug, span_bug};
-use rustc_session::Session;
 use rustc_session::config::{
     BranchProtection, CFGuard, CFProtection, CrateType, DebugInfo, FunctionReturn, PAuthKey, PacRet,
 };
+use rustc_session::{Session, config};
 use rustc_span::source_map::Spanned;
 use rustc_span::{DUMMY_SP, Span};
 use rustc_symbol_mangling::mangle_internal_symbol;
@@ -311,6 +311,11 @@ pub(crate) unsafe fn create_module<'ll>(
         );
     }
 
+    // If fat lto is requested, lld still defaults to thin lto. Set ThinLTO=0 to force fat lto in lld.
+    if sess.lto() == config::Lto::Fat {
+        llvm::add_module_flag_u32(llmod, llvm::ModuleFlagMergeBehavior::Override, "ThinLTO", 0);
+    }
+
     // Add "kcfi" module flag if KCFI is enabled. (See https://reviews.llvm.org/D119296.)
     if sess.is_sanitizer_kcfi_enabled() {
         llvm::add_module_flag_u32(llmod, llvm::ModuleFlagMergeBehavior::Override, "kcfi", 1);

src/tools/run-make-support/src/external_deps/llvm.rs

@@ -60,6 +60,12 @@ pub fn llvm_pdbutil() -> LlvmPdbutil {
     LlvmPdbutil::new()
 }
 
+/// Construct a new `llvm-as` invocation. This assumes that `llvm-as` is available
+/// at `$LLVM_BIN_DIR/llvm-as`.
+pub fn llvm_as() -> LlvmAs {
+    LlvmAs::new()
+}
+
 /// Construct a new `llvm-dis` invocation. This assumes that `llvm-dis` is available
 /// at `$LLVM_BIN_DIR/llvm-dis`.
 pub fn llvm_dis() -> LlvmDis {
@@ -135,6 +141,13 @@ pub struct LlvmPdbutil {
     cmd: Command,
 }
 
+/// A `llvm-as` invocation builder.
+#[derive(Debug)]
+#[must_use]
+pub struct LlvmAs {
+    cmd: Command,
+}
+
 /// A `llvm-dis` invocation builder.
 #[derive(Debug)]
 #[must_use]
@@ -158,6 +171,7 @@ crate::macros::impl_common_helpers!(LlvmNm);
 crate::macros::impl_common_helpers!(LlvmBcanalyzer);
 crate::macros::impl_common_helpers!(LlvmDwarfdump);
 crate::macros::impl_common_helpers!(LlvmPdbutil);
+crate::macros::impl_common_helpers!(LlvmAs);
 crate::macros::impl_common_helpers!(LlvmDis);
 crate::macros::impl_common_helpers!(LlvmObjcopy);
 
@@ -441,6 +455,22 @@ impl LlvmObjcopy {
     }
 }
 
+impl LlvmAs {
+    /// Construct a new `llvm-as` invocation. This assumes that `llvm-as` is available
+    /// at `$LLVM_BIN_DIR/llvm-as`.
+    pub fn new() -> Self {
+        let llvm_as = llvm_bin_dir().join("llvm-as");
+        let cmd = Command::new(llvm_as);
+        Self { cmd }
+    }
+
+    /// Provide an input file.
+    pub fn input<P: AsRef<Path>>(&mut self, path: P) -> &mut Self {
+        self.cmd.arg(path.as_ref());
+        self
+    }
+}
+
 impl LlvmDis {
     /// Construct a new `llvm-dis` invocation. This assumes that `llvm-dis` is available
     /// at `$LLVM_BIN_DIR/llvm-dis`.

src/tools/run-make-support/src/lib.rs

@@ -63,7 +63,7 @@ pub use cargo::cargo;
 pub use clang::{clang, Clang};
 pub use htmldocck::htmldocck;
 pub use llvm::{
-    llvm_ar, llvm_bcanalyzer, llvm_dis, llvm_dwarfdump, llvm_filecheck, llvm_nm, llvm_objcopy,
+    llvm_ar, llvm_bcanalyzer, llvm_as, llvm_dis, llvm_dwarfdump, llvm_filecheck, llvm_nm, llvm_objcopy,
     llvm_objdump, llvm_profdata, llvm_readobj, LlvmAr, LlvmBcanalyzer, LlvmDis, LlvmDwarfdump,
     LlvmFilecheck, LlvmNm, LlvmObjcopy, LlvmObjdump, LlvmProfdata, LlvmReadobj,
 };

tests/run-make/fat-then-thin-lto/lib.rs

@@ -0,0 +1,8 @@
+#![feature(no_core, lang_items)]
+#![no_core]
+#![crate_type = "rlib"]
+
+#[lang = "sized"]
+trait Sized {}
+
+pub fn foo() {}

tests/run-make/fat-then-thin-lto/main.rs

@@ -0,0 +1,11 @@
+#![allow(internal_features)]
+#![feature(no_core, lang_items)]
+#![no_core]
+#![crate_type = "cdylib"]
+
+extern crate lib;
+
+#[unsafe(no_mangle)]
+pub fn bar() {
+    lib::foo();
+}

tests/run-make/fat-then-thin-lto/rmake.rs

@@ -0,0 +1,25 @@
+// Compile a library with lto=fat, then compile a binary with lto=thin
+// and check that lto is applied with the library.
+// The goal is to mimic the standard library being build with lto=fat
+// and allowing users to build with lto=thin.
+
+//@ only-x86_64-unknown-linux-gnu
+
+use run_make_support::{dynamic_lib_name, llvm_objdump, rustc};
+
+fn main() {
+    rustc().input("lib.rs").opt_level("3").arg("-Clto=fat").run();
+    rustc().input("main.rs").panic("abort").opt_level("3").arg("-Clto=thin").run();
+
+    llvm_objdump()
+        .input(dynamic_lib_name("main"))
+        .arg("--disassemble-symbols=bar")
+        .run()
+        // The called function should be inlined.
+        // Check that we have a ret (to detect tail
+        // calls with a jmp) and no call.
+        .assert_stdout_contains("bar")
+        .assert_stdout_contains("ret")
+        .assert_stdout_not_contains("foo")
+        .assert_stdout_not_contains("call");
+}

tests/run-make/linker-plugin-lto-fat/ir.ll

@@ -0,0 +1,6 @@
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+define void @ir_callee() {
+  ret void
+}

tests/run-make/linker-plugin-lto-fat/main.rs

@@ -0,0 +1,17 @@
+#![feature(no_core, lang_items)]
+#![no_core]
+#![crate_type = "cdylib"]
+
+#[lang = "sized"]
+trait Sized {}
+
+extern "C" {
+    fn ir_callee();
+}
+
+#[no_mangle]
+extern "C" fn rs_foo() {
+    unsafe {
+        ir_callee();
+    }
+}

tests/run-make/linker-plugin-lto-fat/rmake.rs

@@ -0,0 +1,29 @@
+// Check that -C lto=fat with -C linker-plugin-lto actually works and can inline functions.
+// A library is created from LLVM IR, defining a single function. Then a dylib is compiled,
+// linking to the library and calling the function from the library.
+// The function from the library should end up inlined and disappear from the output.
+
+//@ only-x86_64-unknown-linux-gnu
+//@ needs-rust-lld
+
+use run_make_support::{dynamic_lib_name, llvm_as, llvm_objdump, rustc};
+
+fn main() {
+    llvm_as().input("ir.ll").run();
+    rustc()
+        .input("main.rs")
+        .opt_level("3")
+        .args(&["-Clto=fat", "-Clinker-plugin-lto", "-Zlinker-features=+lld", "-Clink-arg=ir.bc"])
+        .run();
+
+    llvm_objdump()
+        .input(dynamic_lib_name("main"))
+        .arg("--disassemble-symbols=rs_foo")
+        .run()
+        // The called function should be inlined.
+        // Check that we have a ret (to detect tail
+        // calls with a jmp) and no call.
+        .assert_stdout_contains("foo")
+        .assert_stdout_contains("ret")
+        .assert_stdout_not_contains("call");
+}