Rollup merge of #136840 - Flakebi:linker-plugin-lto-fat, r=dianqk

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.

We solve this by not creating function summaries when fat lto is
enabled. The bitcode for the module is just directly written out.

An alternative solution would be to set the `ThinLTO=0` module flag to
signal lld to do fat lto.
The 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

try-job: x86_64-gnu-debug
try-job: aarch64-gnu-debug

Author: samueltardieu

compiler/rustc_codegen_ssa/src/back/write.rs

@@ -214,7 +214,9 @@ impl ModuleConfig {
                 false
             ),
             emit_obj,
-            emit_thin_lto: sess.opts.unstable_opts.emit_thin_lto,
+            // thin lto summaries prevent fat lto, so do not emit them if fat
+            // lto is requested. See PR #136840 for background information.
+            emit_thin_lto: sess.opts.unstable_opts.emit_thin_lto && sess.lto() != Lto::Fat,
             emit_thin_lto_summary: if_regular!(
                 sess.opts.output_types.contains_key(&OutputType::ThinLinkBitcode),
                 false

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/external_deps/rustc.rs

@@ -173,6 +173,12 @@ impl Rustc {
         self
     }
 
+    /// This flag enables LTO in the specified form.
+    pub fn lto(&mut self, option: &str) -> &mut Self {
+        self.cmd.arg(format!("-Clto={option}"));
+        self
+    }
+
     /// This flag defers LTO optimizations to the linker.
     pub fn linker_plugin_lto(&mut self, option: &str) -> &mut Self {
         self.cmd.arg(format!("-Clinker-plugin-lto={option}"));

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

@@ -63,8 +63,9 @@ pub use crate::external_deps::clang::{Clang, clang};
 pub use crate::external_deps::htmldocck::htmldocck;
 pub use crate::external_deps::llvm::{
     self, LlvmAr, LlvmBcanalyzer, LlvmDis, LlvmDwarfdump, LlvmFilecheck, LlvmNm, LlvmObjcopy,
-    LlvmObjdump, LlvmProfdata, LlvmReadobj, llvm_ar, llvm_bcanalyzer, llvm_dis, llvm_dwarfdump,
-    llvm_filecheck, llvm_nm, llvm_objcopy, llvm_objdump, llvm_profdata, llvm_readobj,
+    LlvmObjdump, LlvmProfdata, LlvmReadobj, llvm_ar, llvm_as, llvm_bcanalyzer, llvm_dis,
+    llvm_dwarfdump, llvm_filecheck, llvm_nm, llvm_objcopy, llvm_objdump, llvm_profdata,
+    llvm_readobj,
 };
 pub use crate::external_deps::python::python_command;
 pub use crate::external_deps::rustc::{self, Rustc, bare_rustc, rustc, rustc_path};

tests/run-make/cross-lang-lto-clang/rmake.rs

@@ -28,53 +28,68 @@ static C_NEVER_INLINED_PATTERN: &'static str = "bl.*<c_never_inlined>";
 static C_NEVER_INLINED_PATTERN: &'static str = "call.*c_never_inlined";
 
 fn main() {
+    test_lto(false);
+    test_lto(true);
+}
+
+fn test_lto(fat_lto: bool) {
+    let lto = if fat_lto { "fat" } else { "thin" };
+    let clang_lto = if fat_lto { "full" } else { "thin" };
+    println!("Running {lto} lto");
+
     rustc()
+        .lto(lto)
         .linker_plugin_lto("on")
         .output(static_lib_name("rustlib-xlto"))
         .opt_level("2")
         .codegen_units(1)
         .input("rustlib.rs")
         .run();
     clang()
-        .lto("thin")
+        .lto(clang_lto)
         .use_ld("lld")
         .arg("-lrustlib-xlto")
         .out_exe("cmain")
         .input("cmain.c")
         .arg("-O3")
         .run();
+
+    let dump = llvm_objdump().disassemble().input("cmain").run();
     // Make sure we don't find a call instruction to the function we expect to
     // always be inlined.
-    llvm_objdump()
-        .disassemble()
-        .input("cmain")
-        .run()
-        .assert_stdout_not_contains_regex(RUST_ALWAYS_INLINED_PATTERN);
+    dump.assert_stdout_not_contains_regex(RUST_ALWAYS_INLINED_PATTERN);
     // As a sanity check, make sure we do find a call instruction to a
     // non-inlined function
-    llvm_objdump()
-        .disassemble()
-        .input("cmain")
-        .run()
-        .assert_stdout_contains_regex(RUST_NEVER_INLINED_PATTERN);
-    clang().input("clib.c").lto("thin").arg("-c").out_exe("clib.o").arg("-O2").run();
+    #[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
+    dump.assert_stdout_contains_regex(RUST_NEVER_INLINED_PATTERN);
+    #[cfg(any(target_arch = "aarch64", target_arch = "arm"))]
+    {
+        if fat_lto {
+            // fat lto inlines this anyway
+            dump.assert_stdout_not_contains_regex(RUST_NEVER_INLINED_PATTERN);
+        } else {
+            dump.assert_stdout_contains_regex(RUST_NEVER_INLINED_PATTERN);
+        }
+    }
+
+    clang().input("clib.c").lto(clang_lto).arg("-c").out_exe("clib.o").arg("-O2").run();
     llvm_ar().obj_to_ar().output_input(static_lib_name("xyz"), "clib.o").run();
     rustc()
+        .lto(lto)
         .linker_plugin_lto("on")
         .opt_level("2")
         .linker(&env_var("CLANG"))
         .link_arg("-fuse-ld=lld")
         .input("main.rs")
         .output("rsmain")
         .run();
-    llvm_objdump()
-        .disassemble()
-        .input("rsmain")
-        .run()
-        .assert_stdout_not_contains_regex(C_ALWAYS_INLINED_PATTERN);
-    llvm_objdump()
-        .disassemble()
-        .input("rsmain")
-        .run()
-        .assert_stdout_contains_regex(C_NEVER_INLINED_PATTERN);
+
+    let dump = llvm_objdump().disassemble().input("rsmain").run();
+    dump.assert_stdout_not_contains_regex(C_ALWAYS_INLINED_PATTERN);
+    if fat_lto {
+        // fat lto inlines this anyway
+        dump.assert_stdout_not_contains_regex(C_NEVER_INLINED_PATTERN);
+    } else {
+        dump.assert_stdout_contains_regex(C_NEVER_INLINED_PATTERN);
+    }
 }

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

@@ -0,0 +1,13 @@
+#![allow(internal_features)]
+#![feature(no_core, lang_items)]
+#![no_core]
+#![crate_type = "rlib"]
+
+#[lang = "pointee_sized"]
+trait PointeeSized {}
+#[lang = "meta_sized"]
+trait MetaSized: PointeeSized {}
+#[lang = "sized"]
+trait Sized: MetaSized {}
+
+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").lto("fat").run();
+    rustc().input("main.rs").panic("abort").opt_level("3").lto("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 +1,22 @@
+#![allow(internal_features)]
+#![feature(no_core, lang_items)]
+#![no_core]
+#![crate_type = "cdylib"]
+
+#[lang = "pointee_sized"]
+trait PointeeSized {}
+#[lang = "meta_sized"]
+trait MetaSized: PointeeSized {}
+#[lang = "sized"]
+trait Sized: MetaSized {}
+
+extern "C" {
+    fn ir_callee();
+}
+
+#[no_mangle]
+extern "C" fn rs_foo() {
+    unsafe {
+        ir_callee();
+    }
+}