Auto merge of #113923 - DianQK:restore-no-builtins-lto, r=pnkfelix

Restore `#![no_builtins]` crates participation in LTO.

After #113716, we can make `#![no_builtins]` crates participate in LTO again.

`#![no_builtins]` with LTO does not result in undefined references to the error. I believe this type of issue won't happen again.

Fixes #72140.  Fixes #112245. Fixes #110606.  Fixes #105734. Fixes #96486. Fixes #108853. Fixes #108893. Fixes #78744. Fixes #91158. Fixes https://github.com/rust-lang/cargo/issues/10118. Fixes https://github.com/rust-lang/compiler-builtins/issues/347.

 The `nightly-2023-07-20` version does not always reproduce problems due to changes in compiler-builtins, core, and user code. That's why this issue recurs and disappears.
Some issues were not tested due to the difficulty of reproducing them.

r? pnkfelix

cc `@bjorn3` `@japaric` `@alexcrichton` `@Amanieu`

5 files changed, 22 insertions, 61 deletions

diff --git a/compiler/rustc_codegen_ssa/src/back/link.rs b/compiler/rustc_codegen_ssa/src/back/link.rs
index be895417bbe..c4b052bf259 100644
--- a/compiler/rustc_codegen_ssa/src/back/link.rs
+++ b/compiler/rustc_codegen_ssa/src/back/link.rs
@@ -270,8 +270,14 @@ pub fn each_linked_rlib(
 
     for &cnum in crates {
         match fmts.get(cnum.as_usize() - 1) {
-            Some(&Linkage::NotLinked | &Linkage::Dynamic | &Linkage::IncludedFromDylib) => continue,
-            Some(_) => {}
+            Some(&Linkage::NotLinked | &Linkage::Dynamic) => continue,
+            Some(&Linkage::IncludedFromDylib) => {
+                // We always link crate `compiler_builtins` statically. When enabling LTO, we include it as well.
+                if info.compiler_builtins != Some(cnum) {
+                    continue;
+                }
+            }
+            Some(&Linkage::Static) => {}
             None => return Err(errors::LinkRlibError::MissingFormat),
         }
         let crate_name = info.crate_name[&cnum];
@@ -520,8 +526,7 @@ fn link_staticlib<'a>(
         &codegen_results.crate_info,
         Some(CrateType::Staticlib),
         &mut |cnum, path| {
-            let lto = are_upstream_rust_objects_already_included(sess)
-                && !ignored_for_lto(sess, &codegen_results.crate_info, cnum);
+            let lto = are_upstream_rust_objects_already_included(sess);
 
             let native_libs = codegen_results.crate_info.native_libraries[&cnum].iter();
             let relevant = native_libs.clone().filter(|lib| relevant_lib(sess, lib));
@@ -1256,24 +1261,6 @@ fn link_sanitizer_runtime(sess: &Session, linker: &mut dyn Linker, name: &str) {
     }
 }
 
-/// Returns a boolean indicating whether the specified crate should be ignored
-/// during LTO.
-///
-/// Crates ignored during LTO are not lumped together in the "massive object
-/// file" that we create and are linked in their normal rlib states. See
-/// comments below for what crates do not participate in LTO.
-///
-/// It's unusual for a crate to not participate in LTO. Typically only
-/// compiler-specific and unstable crates have a reason to not participate in
-/// LTO.
-pub fn ignored_for_lto(sess: &Session, info: &CrateInfo, cnum: CrateNum) -> bool {
-    // If our target enables builtin function lowering in LLVM then the
-    // crates providing these functions don't participate in LTO (e.g.
-    // no_builtins or compiler builtins crates).
-    !sess.target.no_builtins
-        && (info.compiler_builtins == Some(cnum) || info.is_no_builtins.contains(&cnum))
-}
-
 /// This functions tries to determine the appropriate linker (and corresponding LinkerFlavor) to use
 pub fn linker_and_flavor(sess: &Session) -> (PathBuf, LinkerFlavor) {
     fn infer_from(
@@ -2739,10 +2726,6 @@ fn rehome_sysroot_lib_dir<'a>(sess: &'a Session, lib_dir: &Path) -> PathBuf {
 // symbols). We must continue to include the rest of the rlib, however, as
 // it may contain static native libraries which must be linked in.
 //
-// (*) Crates marked with `#![no_builtins]` don't participate in LTO and
-// their bytecode wasn't included. The object files in those libraries must
-// still be passed to the linker.
-//
 // Note, however, that if we're not doing LTO we can just pass the rlib
 // blindly to the linker (fast) because it's fine if it's not actually
 // included as we're at the end of the dependency chain.
@@ -2768,9 +2751,7 @@ fn add_static_crate<'a>(
         cmd.link_rlib(&rlib_path);
     };
 
-    if !are_upstream_rust_objects_already_included(sess)
-        || ignored_for_lto(sess, &codegen_results.crate_info, cnum)
-    {
+    if !are_upstream_rust_objects_already_included(sess) {
         link_upstream(cratepath);
         return;
     }
@@ -2784,8 +2765,6 @@ fn add_static_crate<'a>(
         let canonical_name = name.replace('-', "_");
         let upstream_rust_objects_already_included =
             are_upstream_rust_objects_already_included(sess);
-        let is_builtins =
-            sess.target.no_builtins || !codegen_results.crate_info.is_no_builtins.contains(&cnum);
 
         let mut archive = archive_builder_builder.new_archive_builder(sess);
         if let Err(error) = archive.add_archive(
@@ -2802,9 +2781,8 @@ fn add_static_crate<'a>(
 
                 // If we're performing LTO and this is a rust-generated object
                 // file, then we don't need the object file as it's part of the
-                // LTO module. Note that `#![no_builtins]` is excluded from LTO,
-                // though, so we let that object file slide.
-                if upstream_rust_objects_already_included && is_rust_object && is_builtins {
+                // LTO module.
+                if upstream_rust_objects_already_included && is_rust_object {
                     return true;
                 }
 
diff --git a/compiler/rustc_codegen_ssa/src/back/symbol_export.rs b/compiler/rustc_codegen_ssa/src/back/symbol_export.rs
index b12ac881100..5f2fad0536b 100644
--- a/compiler/rustc_codegen_ssa/src/back/symbol_export.rs
+++ b/compiler/rustc_codegen_ssa/src/back/symbol_export.rs
@@ -54,8 +54,8 @@ fn reachable_non_generics_provider(tcx: TyCtxt<'_>, _: LocalCrate) -> DefIdMap<S
     // export level, however, as they're just implementation details.
     // Down below we'll hardwire all of the symbols to the `Rust` export
     // level instead.
-    let special_runtime_crate =
-        tcx.is_panic_runtime(LOCAL_CRATE) || tcx.is_compiler_builtins(LOCAL_CRATE);
+    let is_compiler_builtins = tcx.is_compiler_builtins(LOCAL_CRATE);
+    let special_runtime_crate = tcx.is_panic_runtime(LOCAL_CRATE) || is_compiler_builtins;
 
     let mut reachable_non_generics: DefIdMap<_> = tcx
         .reachable_set(())
@@ -107,7 +107,11 @@ fn reachable_non_generics_provider(tcx: TyCtxt<'_>, _: LocalCrate) -> DefIdMap<S
         .map(|def_id| {
             // We won't link right if this symbol is stripped during LTO.
             let name = tcx.symbol_name(Instance::mono(tcx, def_id.to_def_id())).name;
-            let used = name == "rust_eh_personality";
+            // We have to preserve the symbols of the built-in functions during LTO.
+            let is_builtin_fn = is_compiler_builtins
+                && symbol_export_level(tcx, def_id.to_def_id())
+                    .is_below_threshold(SymbolExportLevel::C);
+            let used = is_builtin_fn || name == "rust_eh_personality";
 
             let export_level = if special_runtime_crate {
                 SymbolExportLevel::Rust
diff --git a/compiler/rustc_codegen_ssa/src/back/write.rs b/compiler/rustc_codegen_ssa/src/back/write.rs
index 8f17e4ea0bd..9645a08e4a6 100644
--- a/compiler/rustc_codegen_ssa/src/back/write.rs
+++ b/compiler/rustc_codegen_ssa/src/back/write.rs
@@ -148,23 +148,12 @@ impl ModuleConfig {
 
         let emit_obj = if !should_emit_obj {
             EmitObj::None
-        } else if sess.target.obj_is_bitcode
-            || (sess.opts.cg.linker_plugin_lto.enabled() && !no_builtins)
-        {
+        } else if sess.target.obj_is_bitcode || sess.opts.cg.linker_plugin_lto.enabled() {
             // This case is selected if the target uses objects as bitcode, or
             // if linker plugin LTO is enabled. In the linker plugin LTO case
             // the assumption is that the final link-step will read the bitcode
             // and convert it to object code. This may be done by either the
             // native linker or rustc itself.
-            //
-            // Note, however, that the linker-plugin-lto requested here is
-            // explicitly ignored for `#![no_builtins]` crates. These crates are
-            // specifically ignored by rustc's LTO passes and wouldn't work if
-            // loaded into the linker. These crates define symbols that LLVM
-            // lowers intrinsics to, and these symbol dependencies aren't known
-            // until after codegen. As a result any crate marked
-            // `#![no_builtins]` is assumed to not participate in LTO and
-            // instead goes on to generate object code.
             EmitObj::Bitcode
         } else if need_bitcode_in_object(tcx) {
             EmitObj::ObjectCode(BitcodeSection::Full)
@@ -1037,9 +1026,6 @@ fn start_executing_work<B: ExtraBackendMethods>(
 
     let mut each_linked_rlib_for_lto = Vec::new();
     drop(link::each_linked_rlib(crate_info, None, &mut |cnum, path| {
-        if link::ignored_for_lto(sess, crate_info, cnum) {
-            return;
-        }
         each_linked_rlib_for_lto.push((cnum, path.to_path_buf()));
     }));
 
diff --git a/compiler/rustc_codegen_ssa/src/base.rs b/compiler/rustc_codegen_ssa/src/base.rs
index 1d5205ac67a..b87f4b6bf89 100644
--- a/compiler/rustc_codegen_ssa/src/base.rs
+++ b/compiler/rustc_codegen_ssa/src/base.rs
@@ -858,7 +858,6 @@ impl CrateInfo {
             local_crate_name,
             compiler_builtins,
             profiler_runtime: None,
-            is_no_builtins: Default::default(),
             native_libraries: Default::default(),
             used_libraries: tcx.native_libraries(LOCAL_CRATE).iter().map(Into::into).collect(),
             crate_name: Default::default(),
@@ -885,9 +884,6 @@ impl CrateInfo {
             if tcx.is_profiler_runtime(cnum) {
                 info.profiler_runtime = Some(cnum);
             }
-            if tcx.is_no_builtins(cnum) {
-                info.is_no_builtins.insert(cnum);
-            }
         }
 
         // Handle circular dependencies in the standard library.
@@ -895,9 +891,7 @@ impl CrateInfo {
         // If global LTO is enabled then almost everything (*) is glued into a single object file,
         // so this logic is not necessary and can cause issues on some targets (due to weak lang
         // item symbols being "privatized" to that object file), so we disable it.
-        // (*) Native libs, and `#[compiler_builtins]` and `#[no_builtins]` crates are not glued,
-        // and we assume that they cannot define weak lang items. This is not currently enforced
-        // by the compiler, but that's ok because all this stuff is unstable anyway.
+        // (*) Native libs are not glued, and we assume that they cannot define weak lang items.
         let target = &tcx.sess.target;
         if !are_upstream_rust_objects_already_included(tcx.sess) {
             let missing_weak_lang_items: FxHashSet<Symbol> = info
diff --git a/compiler/rustc_codegen_ssa/src/lib.rs b/compiler/rustc_codegen_ssa/src/lib.rs
index bfd0a458884..cd5eb77e06e 100644
--- a/compiler/rustc_codegen_ssa/src/lib.rs
+++ b/compiler/rustc_codegen_ssa/src/lib.rs
@@ -25,7 +25,7 @@ extern crate tracing;
 extern crate rustc_middle;
 
 use rustc_ast as ast;
-use rustc_data_structures::fx::{FxHashMap, FxHashSet};
+use rustc_data_structures::fx::FxHashMap;
 use rustc_data_structures::sync::Lrc;
 use rustc_hir::def_id::CrateNum;
 use rustc_middle::dep_graph::WorkProduct;
@@ -157,7 +157,6 @@ pub struct CrateInfo {
     pub local_crate_name: Symbol,
     pub compiler_builtins: Option<CrateNum>,
     pub profiler_runtime: Option<CrateNum>,
-    pub is_no_builtins: FxHashSet<CrateNum>,
     pub native_libraries: FxHashMap<CrateNum, Vec<NativeLib>>,
     pub crate_name: FxHashMap<CrateNum, Symbol>,
     pub used_libraries: Vec<NativeLib>,