Unrolled build for #145341

Rollup merge of #145341 - Kobzol:codegen-backend-gcc, r=jieyouxu

Install libgccjit into the compiler's sysroot when cg_gcc is enabled

This PR installs the `libgccjit.so` library (which is essentially GCC) into the rustc sysroot (`stageN/lib` on Linux) when the GCC codegen backend is enabled. This allows using the GCC codegen backend "out of the box" with the resulting rustc. It would be nice to get rid of the `libgccjit.so.0` alias (https://rust-lang.zulipchat.com/#narrow/channel/131828-t-compiler/topic/Forcing.20unversioned.20dylib.20dependencies.20on.20Linux/with/534180740), but it's not blocking for this change.

You can try running `x build std --set 'rust.codegen-backends=["llvm", "gcc"]'` and then compiling a hello world with `rustc +stage1 -Zcodegen-backend=gcc main.rs`.

It is now also possible to build stage2 rustc when the GCC is configured to be the default codegen backend, without any further `LD_LIBRARY_PATH` hacks: `./x build compiler --stage 2 --set 'rust.codegen-backends=["gcc"]'`.

After this change, it should be pretty simple to add a dist/rustup step for actually shipping cg_gcc to end users.

CC ```@GuillaumeGomez```

r? ```@jieyouxu```

Author: rust-timer

src/bootstrap/src/core/build_steps/compile.rs

@@ -18,7 +18,7 @@ use serde_derive::Deserialize;
 #[cfg(feature = "tracing")]
 use tracing::{instrument, span};
 
-use crate::core::build_steps::gcc::{Gcc, add_cg_gcc_cargo_flags};
+use crate::core::build_steps::gcc::{Gcc, GccOutput, add_cg_gcc_cargo_flags};
 use crate::core::build_steps::tool::{RustcPrivateCompilers, SourceType, copy_lld_artifacts};
 use crate::core::build_steps::{dist, llvm};
 use crate::core::builder;
@@ -1543,13 +1543,22 @@ impl Step for RustcLink {
     }
 }
 
+/// Output of the `compile::GccCodegenBackend` step.
+/// It includes the path to the libgccjit library on which this backend depends.
+#[derive(Clone)]
+pub struct GccCodegenBackendOutput {
+    stamp: BuildStamp,
+    gcc: GccOutput,
+}
+
 #[derive(Debug, Clone, PartialEq, Eq, Hash)]
 pub struct GccCodegenBackend {
     compilers: RustcPrivateCompilers,
 }
 
 impl Step for GccCodegenBackend {
-    type Output = BuildStamp;
+    type Output = GccCodegenBackendOutput;
+
     const ONLY_HOSTS: bool = true;
 
     fn should_run(run: ShouldRun<'_>) -> ShouldRun<'_> {
@@ -1584,6 +1593,8 @@ impl Step for GccCodegenBackend {
             &CodegenBackendKind::Gcc,
         );
 
+        let gcc = builder.ensure(Gcc { target });
+
         if builder.config.keep_stage.contains(&build_compiler.stage) {
             trace!("`keep-stage` requested");
             builder.info(
@@ -1592,7 +1603,7 @@ impl Step for GccCodegenBackend {
             );
             // Codegen backends are linked separately from this step today, so we don't do
             // anything here.
-            return stamp;
+            return GccCodegenBackendOutput { stamp, gcc };
         }
 
         let mut cargo = builder::Cargo::new(
@@ -1606,13 +1617,16 @@ impl Step for GccCodegenBackend {
         cargo.arg("--manifest-path").arg(builder.src.join("compiler/rustc_codegen_gcc/Cargo.toml"));
         rustc_cargo_env(builder, &mut cargo, target);
 
-        let gcc = builder.ensure(Gcc { target });
         add_cg_gcc_cargo_flags(&mut cargo, &gcc);
 
         let _guard =
             builder.msg(Kind::Build, "codegen backend gcc", Mode::Codegen, build_compiler, target);
         let files = run_cargo(builder, cargo, vec![], &stamp, vec![], false, false);
-        write_codegen_backend_stamp(stamp, files, builder.config.dry_run())
+
+        GccCodegenBackendOutput {
+            stamp: write_codegen_backend_stamp(stamp, files, builder.config.dry_run()),
+            gcc,
+        }
     }
 
     fn metadata(&self) -> Option<StepMetadata> {
@@ -2191,53 +2205,6 @@ impl Step for Assemble {
         );
         build_compiler.stage = actual_stage;
 
-        let mut codegen_backend_stamps = vec![];
-        {
-            #[cfg(feature = "tracing")]
-            let _codegen_backend_span =
-                span!(tracing::Level::DEBUG, "building requested codegen backends").entered();
-
-            for backend in builder.config.enabled_codegen_backends(target_compiler.host) {
-                // FIXME: this is a horrible hack used to make `x check` work when other codegen
-                // backends are enabled.
-                // `x check` will check stage 1 rustc, which copies its rmetas to the stage0 sysroot.
-                // Then it checks codegen backends, which correctly use these rmetas.
-                // Then it needs to check std, but for that it needs to build stage 1 rustc.
-                // This copies the build rmetas into the stage0 sysroot, effectively poisoning it,
-                // because we then have both check and build rmetas in the same sysroot.
-                // That would be fine on its own. However, when another codegen backend is enabled,
-                // then building stage 1 rustc implies also building stage 1 codegen backend (even if
-                // it isn't used for anything). And since that tries to use the poisoned
-                // rmetas, it fails to build.
-                // We don't actually need to build rustc-private codegen backends for checking std,
-                // so instead we skip that.
-                // Note: this would be also an issue for other rustc-private tools, but that is "solved"
-                // by check::Std being last in the list of checked things (see
-                // `Builder::get_step_descriptions`).
-                if builder.kind == Kind::Check && builder.top_stage == 1 {
-                    continue;
-                }
-
-                let prepare_compilers = || {
-                    RustcPrivateCompilers::from_build_and_target_compiler(
-                        build_compiler,
-                        target_compiler,
-                    )
-                };
-
-                let stamp = match backend {
-                    CodegenBackendKind::Cranelift => {
-                        builder.ensure(CraneliftCodegenBackend { compilers: prepare_compilers() })
-                    }
-                    CodegenBackendKind::Gcc => {
-                        builder.ensure(GccCodegenBackend { compilers: prepare_compilers() })
-                    }
-                    CodegenBackendKind::Llvm | CodegenBackendKind::Custom(_) => continue,
-                };
-                codegen_backend_stamps.push(stamp);
-            }
-        }
-
         let stage = target_compiler.stage;
         let host = target_compiler.host;
         let (host_info, dir_name) = if build_compiler.host == host {
@@ -2296,9 +2263,56 @@ impl Step for Assemble {
             }
         }
 
-        debug!("copying codegen backends to sysroot");
-        for stamp in codegen_backend_stamps {
-            copy_codegen_backends_to_sysroot(builder, stamp, target_compiler);
+        {
+            #[cfg(feature = "tracing")]
+            let _codegen_backend_span =
+                span!(tracing::Level::DEBUG, "building requested codegen backends").entered();
+
+            for backend in builder.config.enabled_codegen_backends(target_compiler.host) {
+                // FIXME: this is a horrible hack used to make `x check` work when other codegen
+                // backends are enabled.
+                // `x check` will check stage 1 rustc, which copies its rmetas to the stage0 sysroot.
+                // Then it checks codegen backends, which correctly use these rmetas.
+                // Then it needs to check std, but for that it needs to build stage 1 rustc.
+                // This copies the build rmetas into the stage0 sysroot, effectively poisoning it,
+                // because we then have both check and build rmetas in the same sysroot.
+                // That would be fine on its own. However, when another codegen backend is enabled,
+                // then building stage 1 rustc implies also building stage 1 codegen backend (even if
+                // it isn't used for anything). And since that tries to use the poisoned
+                // rmetas, it fails to build.
+                // We don't actually need to build rustc-private codegen backends for checking std,
+                // so instead we skip that.
+                // Note: this would be also an issue for other rustc-private tools, but that is "solved"
+                // by check::Std being last in the list of checked things (see
+                // `Builder::get_step_descriptions`).
+                if builder.kind == Kind::Check && builder.top_stage == 1 {
+                    continue;
+                }
+
+                let prepare_compilers = || {
+                    RustcPrivateCompilers::from_build_and_target_compiler(
+                        build_compiler,
+                        target_compiler,
+                    )
+                };
+
+                match backend {
+                    CodegenBackendKind::Cranelift => {
+                        let stamp = builder
+                            .ensure(CraneliftCodegenBackend { compilers: prepare_compilers() });
+                        copy_codegen_backends_to_sysroot(builder, stamp, target_compiler);
+                    }
+                    CodegenBackendKind::Gcc => {
+                        let output =
+                            builder.ensure(GccCodegenBackend { compilers: prepare_compilers() });
+                        copy_codegen_backends_to_sysroot(builder, output.stamp, target_compiler);
+                        // Also copy libgccjit to the library sysroot, so that it is available for
+                        // the codegen backend.
+                        output.gcc.install_to(builder, &rustc_libdir);
+                    }
+                    CodegenBackendKind::Llvm | CodegenBackendKind::Custom(_) => continue,
+                }
+            }
         }
 
         if builder.config.lld_enabled {

src/bootstrap/src/core/build_steps/dist.rs

@@ -478,7 +478,19 @@ impl Step for Rustc {
             if libdir_relative.to_str() != Some("bin") {
                 let libdir = builder.rustc_libdir(compiler);
                 for entry in builder.read_dir(&libdir) {
-                    if is_dylib(&entry.path()) {
+                    // A safeguard that we will not ship libgccjit.so from the libdir, in case the
+                    // GCC codegen backend is enabled by default.
+                    // Long-term we should probably split the config options for:
+                    // - Include cg_gcc in the rustc sysroot by default
+                    // - Run dist of a specific codegen backend in `x dist` by default
+                    if is_dylib(&entry.path())
+                        && !entry
+                            .path()
+                            .file_name()
+                            .and_then(|n| n.to_str())
+                            .map(|n| n.contains("libgccjit"))
+                            .unwrap_or(false)
+                    {
                         // Don't use custom libdir here because ^lib/ will be resolved again
                         // with installer
                         builder.install(&entry.path(), &image.join("lib"), FileType::NativeLibrary);

src/bootstrap/src/core/build_steps/gcc.rs

@@ -12,6 +12,7 @@ use std::fs;
 use std::path::{Path, PathBuf};
 use std::sync::OnceLock;
 
+use crate::FileType;
 use crate::core::builder::{Builder, Cargo, Kind, RunConfig, ShouldRun, Step};
 use crate::core::config::TargetSelection;
 use crate::utils::build_stamp::{BuildStamp, generate_smart_stamp_hash};
@@ -28,6 +29,21 @@ pub struct GccOutput {
     pub libgccjit: PathBuf,
 }
 
+impl GccOutput {
+    /// Install the required libgccjit library file(s) to the specified `path`.
+    pub fn install_to(&self, builder: &Builder<'_>, directory: &Path) {
+        // At build time, cg_gcc has to link to libgccjit.so (the unversioned symbol).
+        // However, at runtime, it will by default look for libgccjit.so.0.
+        // So when we install the built libgccjit.so file to the target `directory`, we add it there
+        // with the `.0` suffix.
+        let mut target_filename = self.libgccjit.file_name().unwrap().to_str().unwrap().to_string();
+        target_filename.push_str(".0");
+
+        let dst = directory.join(target_filename);
+        builder.copy_link(&self.libgccjit, &dst, FileType::NativeLibrary);
+    }
+}
+
 impl Step for Gcc {
     type Output = GccOutput;
 
@@ -61,23 +77,13 @@ impl Step for Gcc {
         }
 
         build_gcc(&metadata, builder, target);
-        create_lib_alias(builder, &libgccjit_path);
 
         t!(metadata.stamp.write());
 
         GccOutput { libgccjit: libgccjit_path }
     }
 }
 
-/// Creates a libgccjit.so.0 alias next to libgccjit.so if it does not
-/// already exist
-fn create_lib_alias(builder: &Builder<'_>, libgccjit: &PathBuf) {
-    let lib_alias = libgccjit.parent().unwrap().join("libgccjit.so.0");
-    if !lib_alias.exists() {
-        t!(builder.symlink_file(libgccjit, lib_alias));
-    }
-}
-
 pub struct Meta {
     stamp: BuildStamp,
     out_dir: PathBuf,
@@ -124,7 +130,6 @@ fn try_download_gcc(builder: &Builder<'_>, target: TargetSelection) -> Option<Pa
             }
 
             let libgccjit = root.join("lib").join("libgccjit.so");
-            create_lib_alias(builder, &libgccjit);
             Some(libgccjit)
         }
         PathFreshness::HasLocalModifications { .. } => {