Merge pull request #4492 from rust-lang/rustup-2025-07-25

Automatic Rustup

9 files changed, 329 insertions, 147 deletions

diff --git a/compiler/rustc_codegen_ssa/src/back/link/raw_dylib.rs b/compiler/rustc_codegen_ssa/src/back/link/raw_dylib.rs
index 74f39022afb..b9e0c957363 100644
--- a/compiler/rustc_codegen_ssa/src/back/link/raw_dylib.rs
+++ b/compiler/rustc_codegen_ssa/src/back/link/raw_dylib.rs
@@ -4,7 +4,7 @@ use std::path::{Path, PathBuf};
 
 use rustc_abi::Endian;
 use rustc_data_structures::base_n::{CASE_INSENSITIVE, ToBaseN};
-use rustc_data_structures::fx::FxIndexMap;
+use rustc_data_structures::fx::{FxHashMap, FxIndexMap};
 use rustc_data_structures::stable_hasher::StableHasher;
 use rustc_hashes::Hash128;
 use rustc_session::Session;
@@ -214,7 +214,7 @@ pub(super) fn create_raw_dylib_elf_stub_shared_objects<'a>(
 /// It exports all the provided symbols, but is otherwise empty.
 fn create_elf_raw_dylib_stub(sess: &Session, soname: &str, symbols: &[DllImport]) -> Vec<u8> {
     use object::write::elf as write;
-    use object::{Architecture, elf};
+    use object::{AddressSize, Architecture, elf};
 
     let mut stub_buf = Vec::new();
 
@@ -226,54 +226,94 @@ fn create_elf_raw_dylib_stub(sess: &Session, soname: &str, symbols: &[DllImport]
     // It is important that the order of reservation matches the order of writing.
     // The object crate contains many debug asserts that fire if you get this wrong.
 
+    let Some((arch, sub_arch)) = sess.target.object_architecture(&sess.unstable_target_features)
+    else {
+        sess.dcx().fatal(format!(
+            "raw-dylib is not supported for the architecture `{}`",
+            sess.target.arch
+        ));
+    };
+
     let endianness = match sess.target.options.endian {
         Endian::Little => object::Endianness::Little,
         Endian::Big => object::Endianness::Big,
     };
-    let mut stub = write::Writer::new(endianness, true, &mut stub_buf);
+
+    let is_64 = match arch.address_size() {
+        Some(AddressSize::U8 | AddressSize::U16 | AddressSize::U32) => false,
+        Some(AddressSize::U64) => true,
+        _ => sess.dcx().fatal(format!(
+            "raw-dylib is not supported for the architecture `{}`",
+            sess.target.arch
+        )),
+    };
+
+    let mut stub = write::Writer::new(endianness, is_64, &mut stub_buf);
+
+    let mut vers = Vec::new();
+    let mut vers_map = FxHashMap::default();
+    let mut syms = Vec::new();
+
+    for symbol in symbols {
+        let symbol_name = symbol.name.as_str();
+        if let Some((name, version_name)) = symbol_name.split_once('@') {
+            assert!(!version_name.contains('@'));
+            let dynstr = stub.add_dynamic_string(name.as_bytes());
+            let ver = if let Some(&ver_id) = vers_map.get(version_name) {
+                ver_id
+            } else {
+                let id = vers.len();
+                vers_map.insert(version_name, id);
+                let dynstr = stub.add_dynamic_string(version_name.as_bytes());
+                vers.push((version_name, dynstr));
+                id
+            };
+            syms.push((name, dynstr, Some(ver)));
+        } else {
+            let dynstr = stub.add_dynamic_string(symbol_name.as_bytes());
+            syms.push((symbol_name, dynstr, None));
+        }
+    }
+
+    let soname = stub.add_dynamic_string(soname.as_bytes());
 
     // These initial reservations don't reserve any bytes in the binary yet,
     // they just allocate in the internal data structures.
 
-    // First, we crate the dynamic symbol table. It starts with a null symbol
+    // First, we create the dynamic symbol table. It starts with a null symbol
     // and then all the symbols and their dynamic strings.
     stub.reserve_null_dynamic_symbol_index();
 
-    let dynstrs = symbols
-        .iter()
-        .map(|sym| {
-            stub.reserve_dynamic_symbol_index();
-            (sym, stub.add_dynamic_string(sym.name.as_str().as_bytes()))
-        })
-        .collect::<Vec<_>>();
-
-    let soname = stub.add_dynamic_string(soname.as_bytes());
+    for _ in syms.iter() {
+        stub.reserve_dynamic_symbol_index();
+    }
 
     // Reserve the sections.
     // We have the minimal sections for a dynamic SO and .text where we point our dummy symbols to.
     stub.reserve_shstrtab_section_index();
     let text_section_name = stub.add_section_name(".text".as_bytes());
     let text_section = stub.reserve_section_index();
-    stub.reserve_dynstr_section_index();
     stub.reserve_dynsym_section_index();
+    stub.reserve_dynstr_section_index();
+    if !vers.is_empty() {
+        stub.reserve_gnu_versym_section_index();
+        stub.reserve_gnu_verdef_section_index();
+    }
     stub.reserve_dynamic_section_index();
 
     // These reservations now determine the actual layout order of the object file.
     stub.reserve_file_header();
     stub.reserve_shstrtab();
     stub.reserve_section_headers();
-    stub.reserve_dynstr();
     stub.reserve_dynsym();
+    stub.reserve_dynstr();
+    if !vers.is_empty() {
+        stub.reserve_gnu_versym();
+        stub.reserve_gnu_verdef(1 + vers.len(), 1 + vers.len());
+    }
     stub.reserve_dynamic(2); // DT_SONAME, DT_NULL
 
     // First write the ELF header with the arch information.
-    let Some((arch, sub_arch)) = sess.target.object_architecture(&sess.unstable_target_features)
-    else {
-        sess.dcx().fatal(format!(
-            "raw-dylib is not supported for the architecture `{}`",
-            sess.target.arch
-        ));
-    };
     let e_machine = match (arch, sub_arch) {
         (Architecture::Aarch64, None) => elf::EM_AARCH64,
         (Architecture::Aarch64_Ilp32, None) => elf::EM_AARCH64,
@@ -342,18 +382,19 @@ fn create_elf_raw_dylib_stub(sess: &Session, soname: &str, symbols: &[DllImport]
         sh_addralign: 1,
         sh_entsize: 0,
     });
-    stub.write_dynstr_section_header(0);
     stub.write_dynsym_section_header(0, 1);
+    stub.write_dynstr_section_header(0);
+    if !vers.is_empty() {
+        stub.write_gnu_versym_section_header(0);
+        stub.write_gnu_verdef_section_header(0);
+    }
     stub.write_dynamic_section_header(0);
 
-    // .dynstr
-    stub.write_dynstr();
-
     // .dynsym
     stub.write_null_dynamic_symbol();
-    for (_, name) in dynstrs {
+    for (_name, dynstr, _ver) in syms.iter().copied() {
         stub.write_dynamic_symbol(&write::Sym {
-            name: Some(name),
+            name: Some(dynstr),
             st_info: (elf::STB_GLOBAL << 4) | elf::STT_NOTYPE,
             st_other: elf::STV_DEFAULT,
             section: Some(text_section),
@@ -363,10 +404,47 @@ fn create_elf_raw_dylib_stub(sess: &Session, soname: &str, symbols: &[DllImport]
         });
     }
 
+    // .dynstr
+    stub.write_dynstr();
+
+    // ld.bfd is unhappy if these sections exist without any symbols, so we only generate them when necessary.
+    if !vers.is_empty() {
+        // .gnu_version
+        stub.write_null_gnu_versym();
+        for (_name, _dynstr, ver) in syms.iter().copied() {
+            stub.write_gnu_versym(if let Some(ver) = ver {
+                assert!((2 + ver as u16) < elf::VERSYM_HIDDEN);
+                elf::VERSYM_HIDDEN | (2 + ver as u16)
+            } else {
+                1
+            });
+        }
+
+        // .gnu_version_d
+        stub.write_align_gnu_verdef();
+        stub.write_gnu_verdef(&write::Verdef {
+            version: elf::VER_DEF_CURRENT,
+            flags: elf::VER_FLG_BASE,
+            index: 1,
+            aux_count: 1,
+            name: soname,
+        });
+        for (ver, (_name, dynstr)) in vers.into_iter().enumerate() {
+            stub.write_gnu_verdef(&write::Verdef {
+                version: elf::VER_DEF_CURRENT,
+                flags: 0,
+                index: 2 + ver as u16,
+                aux_count: 1,
+                name: dynstr,
+            });
+        }
+    }
+
     // .dynamic
     // the DT_SONAME will be used by the linker to populate DT_NEEDED
     // which the loader uses to find the library.
     // DT_NULL terminates the .dynamic table.
+    stub.write_align_dynamic();
     stub.write_dynamic_string(elf::DT_SONAME, soname);
     stub.write_dynamic(elf::DT_NULL, 0);
 
diff --git a/compiler/rustc_codegen_ssa/src/back/lto.rs b/compiler/rustc_codegen_ssa/src/back/lto.rs
index b49b6783bbd..c95038375a1 100644
--- a/compiler/rustc_codegen_ssa/src/back/lto.rs
+++ b/compiler/rustc_codegen_ssa/src/back/lto.rs
@@ -2,7 +2,15 @@ use std::ffi::CString;
 use std::sync::Arc;
 
 use rustc_data_structures::memmap::Mmap;
+use rustc_hir::def_id::{CrateNum, LOCAL_CRATE};
+use rustc_middle::middle::exported_symbols::{ExportedSymbol, SymbolExportInfo, SymbolExportLevel};
+use rustc_middle::ty::TyCtxt;
+use rustc_session::config::{CrateType, Lto};
+use tracing::info;
 
+use crate::back::symbol_export::{self, symbol_name_for_instance_in_crate};
+use crate::back::write::CodegenContext;
+use crate::errors::{DynamicLinkingWithLTO, LtoDisallowed, LtoDylib, LtoProcMacro};
 use crate::traits::*;
 
 pub struct ThinModule<B: WriteBackendMethods> {
@@ -52,3 +60,86 @@ impl<M: ModuleBufferMethods> SerializedModule<M> {
         }
     }
 }
+
+fn crate_type_allows_lto(crate_type: CrateType) -> bool {
+    match crate_type {
+        CrateType::Executable
+        | CrateType::Dylib
+        | CrateType::Staticlib
+        | CrateType::Cdylib
+        | CrateType::ProcMacro
+        | CrateType::Sdylib => true,
+        CrateType::Rlib => false,
+    }
+}
+
+pub(super) fn exported_symbols_for_lto(
+    tcx: TyCtxt<'_>,
+    each_linked_rlib_for_lto: &[CrateNum],
+) -> Vec<String> {
+    let export_threshold = match tcx.sess.lto() {
+        // We're just doing LTO for our one crate
+        Lto::ThinLocal => SymbolExportLevel::Rust,
+
+        // We're doing LTO for the entire crate graph
+        Lto::Fat | Lto::Thin => symbol_export::crates_export_threshold(&tcx.crate_types()),
+
+        Lto::No => return vec![],
+    };
+
+    let copy_symbols = |cnum| {
+        tcx.exported_non_generic_symbols(cnum)
+            .iter()
+            .chain(tcx.exported_generic_symbols(cnum))
+            .filter_map(|&(s, info): &(ExportedSymbol<'_>, SymbolExportInfo)| {
+                if info.level.is_below_threshold(export_threshold) || info.used {
+                    Some(symbol_name_for_instance_in_crate(tcx, s, cnum))
+                } else {
+                    None
+                }
+            })
+            .collect::<Vec<_>>()
+    };
+    let mut symbols_below_threshold = {
+        let _timer = tcx.prof.generic_activity("lto_generate_symbols_below_threshold");
+        copy_symbols(LOCAL_CRATE)
+    };
+    info!("{} symbols to preserve in this crate", symbols_below_threshold.len());
+
+    // If we're performing LTO for the entire crate graph, then for each of our
+    // upstream dependencies, include their exported symbols.
+    if tcx.sess.lto() != Lto::ThinLocal {
+        for &cnum in each_linked_rlib_for_lto {
+            let _timer = tcx.prof.generic_activity("lto_generate_symbols_below_threshold");
+            symbols_below_threshold.extend(copy_symbols(cnum));
+        }
+    }
+
+    symbols_below_threshold
+}
+
+pub(super) fn check_lto_allowed<B: WriteBackendMethods>(cgcx: &CodegenContext<B>) {
+    if cgcx.lto == Lto::ThinLocal {
+        // Crate local LTO is always allowed
+        return;
+    }
+
+    let dcx = cgcx.create_dcx();
+
+    // Make sure we actually can run LTO
+    for crate_type in cgcx.crate_types.iter() {
+        if !crate_type_allows_lto(*crate_type) {
+            dcx.handle().emit_fatal(LtoDisallowed);
+        } else if *crate_type == CrateType::Dylib {
+            if !cgcx.opts.unstable_opts.dylib_lto {
+                dcx.handle().emit_fatal(LtoDylib);
+            }
+        } else if *crate_type == CrateType::ProcMacro && !cgcx.opts.unstable_opts.dylib_lto {
+            dcx.handle().emit_fatal(LtoProcMacro);
+        }
+    }
+
+    if cgcx.opts.cg.prefer_dynamic && !cgcx.opts.unstable_opts.dylib_lto {
+        dcx.handle().emit_fatal(DynamicLinkingWithLTO);
+    }
+}
diff --git a/compiler/rustc_codegen_ssa/src/back/write.rs b/compiler/rustc_codegen_ssa/src/back/write.rs
index 24e0a4eb533..7be274df1d4 100644
--- a/compiler/rustc_codegen_ssa/src/back/write.rs
+++ b/compiler/rustc_codegen_ssa/src/back/write.rs
@@ -9,7 +9,7 @@ use std::{fs, io, mem, str, thread};
 use rustc_abi::Size;
 use rustc_ast::attr;
 use rustc_ast::expand::autodiff_attrs::AutoDiffItem;
-use rustc_data_structures::fx::{FxHashMap, FxIndexMap};
+use rustc_data_structures::fx::FxIndexMap;
 use rustc_data_structures::jobserver::{self, Acquired};
 use rustc_data_structures::memmap::Mmap;
 use rustc_data_structures::profiling::{SelfProfilerRef, VerboseTimingGuard};
@@ -20,14 +20,12 @@ use rustc_errors::{
     Suggestions,
 };
 use rustc_fs_util::link_or_copy;
-use rustc_hir::def_id::{CrateNum, LOCAL_CRATE};
 use rustc_incremental::{
     copy_cgu_workproduct_to_incr_comp_cache_dir, in_incr_comp_dir, in_incr_comp_dir_sess,
 };
 use rustc_metadata::fs::copy_to_stdout;
 use rustc_middle::bug;
 use rustc_middle::dep_graph::{WorkProduct, WorkProductId};
-use rustc_middle::middle::exported_symbols::SymbolExportInfo;
 use rustc_middle::ty::TyCtxt;
 use rustc_session::Session;
 use rustc_session::config::{
@@ -40,7 +38,7 @@ use tracing::debug;
 
 use super::link::{self, ensure_removed};
 use super::lto::{self, SerializedModule};
-use super::symbol_export::symbol_name_for_instance_in_crate;
+use crate::back::lto::check_lto_allowed;
 use crate::errors::{AutodiffWithoutLto, ErrorCreatingRemarkDir};
 use crate::traits::*;
 use crate::{
@@ -332,8 +330,6 @@ pub type TargetMachineFactoryFn<B> = Arc<
         + Sync,
 >;
 
-type ExportedSymbols = FxHashMap<CrateNum, Arc<Vec<(String, SymbolExportInfo)>>>;
-
 /// Additional resources used by optimize_and_codegen (not module specific)
 #[derive(Clone)]
 pub struct CodegenContext<B: WriteBackendMethods> {
@@ -343,10 +339,8 @@ pub struct CodegenContext<B: WriteBackendMethods> {
     pub save_temps: bool,
     pub fewer_names: bool,
     pub time_trace: bool,
-    pub exported_symbols: Option<Arc<ExportedSymbols>>,
     pub opts: Arc<config::Options>,
     pub crate_types: Vec<CrateType>,
-    pub each_linked_rlib_for_lto: Vec<(CrateNum, PathBuf)>,
     pub output_filenames: Arc<OutputFilenames>,
     pub invocation_temp: Option<String>,
     pub regular_module_config: Arc<ModuleConfig>,
@@ -401,13 +395,21 @@ impl<B: WriteBackendMethods> CodegenContext<B> {
 
 fn generate_thin_lto_work<B: ExtraBackendMethods>(
     cgcx: &CodegenContext<B>,
+    exported_symbols_for_lto: &[String],
+    each_linked_rlib_for_lto: &[PathBuf],
     needs_thin_lto: Vec<(String, B::ThinBuffer)>,
     import_only_modules: Vec<(SerializedModule<B::ModuleBuffer>, WorkProduct)>,
 ) -> Vec<(WorkItem<B>, u64)> {
     let _prof_timer = cgcx.prof.generic_activity("codegen_thin_generate_lto_work");
 
-    let (lto_modules, copy_jobs) =
-        B::run_thin_lto(cgcx, needs_thin_lto, import_only_modules).unwrap_or_else(|e| e.raise());
+    let (lto_modules, copy_jobs) = B::run_thin_lto(
+        cgcx,
+        exported_symbols_for_lto,
+        each_linked_rlib_for_lto,
+        needs_thin_lto,
+        import_only_modules,
+    )
+    .unwrap_or_else(|e| e.raise());
     lto_modules
         .into_iter()
         .map(|module| {
@@ -723,6 +725,8 @@ pub(crate) enum WorkItem<B: WriteBackendMethods> {
     CopyPostLtoArtifacts(CachedModuleCodegen),
     /// Performs fat LTO on the given module.
     FatLto {
+        exported_symbols_for_lto: Arc<Vec<String>>,
+        each_linked_rlib_for_lto: Vec<PathBuf>,
         needs_fat_lto: Vec<FatLtoInput<B>>,
         import_only_modules: Vec<(SerializedModule<B::ModuleBuffer>, WorkProduct)>,
         autodiff: Vec<AutoDiffItem>,
@@ -810,7 +814,7 @@ pub(crate) enum WorkItemResult<B: WriteBackendMethods> {
 }
 
 pub enum FatLtoInput<B: WriteBackendMethods> {
-    Serialized { name: String, buffer: B::ModuleBuffer },
+    Serialized { name: String, buffer: SerializedModule<B::ModuleBuffer> },
     InMemory(ModuleCodegen<B::Module>),
 }
 
@@ -899,7 +903,10 @@ fn execute_optimize_work_item<B: ExtraBackendMethods>(
                 fs::write(&path, buffer.data()).unwrap_or_else(|e| {
                     panic!("Error writing pre-lto-bitcode file `{}`: {}", path.display(), e);
                 });
-                Ok(WorkItemResult::NeedsFatLto(FatLtoInput::Serialized { name, buffer }))
+                Ok(WorkItemResult::NeedsFatLto(FatLtoInput::Serialized {
+                    name,
+                    buffer: SerializedModule::Local(buffer),
+                }))
             }
             None => Ok(WorkItemResult::NeedsFatLto(FatLtoInput::InMemory(module))),
         },
@@ -992,12 +999,24 @@ fn execute_copy_from_cache_work_item<B: ExtraBackendMethods>(
 
 fn execute_fat_lto_work_item<B: ExtraBackendMethods>(
     cgcx: &CodegenContext<B>,
-    needs_fat_lto: Vec<FatLtoInput<B>>,
+    exported_symbols_for_lto: &[String],
+    each_linked_rlib_for_lto: &[PathBuf],
+    mut needs_fat_lto: Vec<FatLtoInput<B>>,
     import_only_modules: Vec<(SerializedModule<B::ModuleBuffer>, WorkProduct)>,
     autodiff: Vec<AutoDiffItem>,
     module_config: &ModuleConfig,
 ) -> Result<WorkItemResult<B>, FatalError> {
-    let module = B::run_and_optimize_fat_lto(cgcx, needs_fat_lto, import_only_modules, autodiff)?;
+    for (module, wp) in import_only_modules {
+        needs_fat_lto.push(FatLtoInput::Serialized { name: wp.cgu_name, buffer: module })
+    }
+
+    let module = B::run_and_optimize_fat_lto(
+        cgcx,
+        exported_symbols_for_lto,
+        each_linked_rlib_for_lto,
+        needs_fat_lto,
+        autodiff,
+    )?;
     let module = B::codegen(cgcx, module, module_config)?;
     Ok(WorkItemResult::Finished(module))
 }
@@ -1032,7 +1051,7 @@ pub(crate) enum Message<B: WriteBackendMethods> {
 
     /// The backend has finished processing a work item for a codegen unit.
     /// Sent from a backend worker thread.
-    WorkItem { result: Result<WorkItemResult<B>, Option<WorkerFatalError>>, worker_id: usize },
+    WorkItem { result: Result<WorkItemResult<B>, Option<WorkerFatalError>> },
 
     /// The frontend has finished generating something (backend IR or a
     /// post-LTO artifact) for a codegen unit, and it should be passed to the
@@ -1113,42 +1132,18 @@ fn start_executing_work<B: ExtraBackendMethods>(
     let autodiff_items = autodiff_items.to_vec();
 
     let mut each_linked_rlib_for_lto = Vec::new();
+    let mut each_linked_rlib_file_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()));
+        each_linked_rlib_for_lto.push(cnum);
+        each_linked_rlib_file_for_lto.push(path.to_path_buf());
     }));
 
     // Compute the set of symbols we need to retain when doing LTO (if we need to)
-    let exported_symbols = {
-        let mut exported_symbols = FxHashMap::default();
-
-        let copy_symbols = |cnum| {
-            let symbols = tcx
-                .exported_non_generic_symbols(cnum)
-                .iter()
-                .chain(tcx.exported_generic_symbols(cnum))
-                .map(|&(s, lvl)| (symbol_name_for_instance_in_crate(tcx, s, cnum), lvl))
-                .collect();
-            Arc::new(symbols)
-        };
-
-        match sess.lto() {
-            Lto::No => None,
-            Lto::ThinLocal => {
-                exported_symbols.insert(LOCAL_CRATE, copy_symbols(LOCAL_CRATE));
-                Some(Arc::new(exported_symbols))
-            }
-            Lto::Fat | Lto::Thin => {
-                exported_symbols.insert(LOCAL_CRATE, copy_symbols(LOCAL_CRATE));
-                for &(cnum, ref _path) in &each_linked_rlib_for_lto {
-                    exported_symbols.insert(cnum, copy_symbols(cnum));
-                }
-                Some(Arc::new(exported_symbols))
-            }
-        }
-    };
+    let exported_symbols_for_lto =
+        Arc::new(lto::exported_symbols_for_lto(tcx, &each_linked_rlib_for_lto));
 
     // First up, convert our jobserver into a helper thread so we can use normal
     // mpsc channels to manage our messages and such.
@@ -1183,14 +1178,12 @@ fn start_executing_work<B: ExtraBackendMethods>(
 
     let cgcx = CodegenContext::<B> {
         crate_types: tcx.crate_types().to_vec(),
-        each_linked_rlib_for_lto,
         lto: sess.lto(),
         fewer_names: sess.fewer_names(),
         save_temps: sess.opts.cg.save_temps,
         time_trace: sess.opts.unstable_opts.llvm_time_trace,
         opts: Arc::new(sess.opts.clone()),
         prof: sess.prof.clone(),
-        exported_symbols,
         remark: sess.opts.cg.remark.clone(),
         remark_dir,
         incr_comp_session_dir: sess.incr_comp_session_dir_opt().map(|r| r.clone()),
@@ -1350,18 +1343,6 @@ fn start_executing_work<B: ExtraBackendMethods>(
     // necessary. There's already optimizations in place to avoid sending work
     // back to the coordinator if LTO isn't requested.
     return B::spawn_named_thread(cgcx.time_trace, "coordinator".to_string(), move || {
-        let mut worker_id_counter = 0;
-        let mut free_worker_ids = Vec::new();
-        let mut get_worker_id = |free_worker_ids: &mut Vec<usize>| {
-            if let Some(id) = free_worker_ids.pop() {
-                id
-            } else {
-                let id = worker_id_counter;
-                worker_id_counter += 1;
-                id
-            }
-        };
-
         // This is where we collect codegen units that have gone all the way
         // through codegen and LLVM.
         let mut compiled_modules = vec![];
@@ -1442,12 +1423,7 @@ fn start_executing_work<B: ExtraBackendMethods>(
                         let (item, _) =
                             work_items.pop().expect("queue empty - queue_full_enough() broken?");
                         main_thread_state = MainThreadState::Lending;
-                        spawn_work(
-                            &cgcx,
-                            &mut llvm_start_time,
-                            get_worker_id(&mut free_worker_ids),
-                            item,
-                        );
+                        spawn_work(&cgcx, &mut llvm_start_time, item);
                     }
                 }
             } else if codegen_state == Completed {
@@ -1474,12 +1450,18 @@ fn start_executing_work<B: ExtraBackendMethods>(
                     let needs_fat_lto = mem::take(&mut needs_fat_lto);
                     let needs_thin_lto = mem::take(&mut needs_thin_lto);
                     let import_only_modules = mem::take(&mut lto_import_only_modules);
+                    let each_linked_rlib_file_for_lto =
+                        mem::take(&mut each_linked_rlib_file_for_lto);
+
+                    check_lto_allowed(&cgcx);
 
                     if !needs_fat_lto.is_empty() {
                         assert!(needs_thin_lto.is_empty());
 
                         work_items.push((
                             WorkItem::FatLto {
+                                exported_symbols_for_lto: Arc::clone(&exported_symbols_for_lto),
+                                each_linked_rlib_for_lto: each_linked_rlib_file_for_lto,
                                 needs_fat_lto,
                                 import_only_modules,
                                 autodiff: autodiff_items.clone(),
@@ -1495,9 +1477,13 @@ fn start_executing_work<B: ExtraBackendMethods>(
                             dcx.handle().emit_fatal(AutodiffWithoutLto {});
                         }
 
-                        for (work, cost) in
-                            generate_thin_lto_work(&cgcx, needs_thin_lto, import_only_modules)
-                        {
+                        for (work, cost) in generate_thin_lto_work(
+                            &cgcx,
+                            &exported_symbols_for_lto,
+                            &each_linked_rlib_file_for_lto,
+                            needs_thin_lto,
+                            import_only_modules,
+                        ) {
                             let insertion_index = work_items
                                 .binary_search_by_key(&cost, |&(_, cost)| cost)
                                 .unwrap_or_else(|e| e);
@@ -1516,12 +1502,7 @@ fn start_executing_work<B: ExtraBackendMethods>(
                     MainThreadState::Idle => {
                         if let Some((item, _)) = work_items.pop() {
                             main_thread_state = MainThreadState::Lending;
-                            spawn_work(
-                                &cgcx,
-                                &mut llvm_start_time,
-                                get_worker_id(&mut free_worker_ids),
-                                item,
-                            );
+                            spawn_work(&cgcx, &mut llvm_start_time, item);
                         } else {
                             // There is no unstarted work, so let the main thread
                             // take over for a running worker. Otherwise the
@@ -1557,12 +1538,7 @@ fn start_executing_work<B: ExtraBackendMethods>(
                 while running_with_own_token < tokens.len()
                     && let Some((item, _)) = work_items.pop()
                 {
-                    spawn_work(
-                        &cgcx,
-                        &mut llvm_start_time,
-                        get_worker_id(&mut free_worker_ids),
-                        item,
-                    );
+                    spawn_work(&cgcx, &mut llvm_start_time, item);
                     running_with_own_token += 1;
                 }
             }
@@ -1570,21 +1546,6 @@ fn start_executing_work<B: ExtraBackendMethods>(
             // Relinquish accidentally acquired extra tokens.
             tokens.truncate(running_with_own_token);
 
-            // If a thread exits successfully then we drop a token associated
-            // with that worker and update our `running_with_own_token` count.
-            // We may later re-acquire a token to continue running more work.
-            // We may also not actually drop a token here if the worker was
-            // running with an "ephemeral token".
-            let mut free_worker = |worker_id| {
-                if main_thread_state == MainThreadState::Lending {
-                    main_thread_state = MainThreadState::Idle;
-                } else {
-                    running_with_own_token -= 1;
-                }
-
-                free_worker_ids.push(worker_id);
-            };
-
             let msg = coordinator_receive.recv().unwrap();
             match *msg.downcast::<Message<B>>().ok().unwrap() {
                 // Save the token locally and the next turn of the loop will use
@@ -1653,8 +1614,17 @@ fn start_executing_work<B: ExtraBackendMethods>(
                     codegen_state = Aborted;
                 }
 
-                Message::WorkItem { result, worker_id } => {
-                    free_worker(worker_id);
+                Message::WorkItem { result } => {
+                    // If a thread exits successfully then we drop a token associated
+                    // with that worker and update our `running_with_own_token` count.
+                    // We may later re-acquire a token to continue running more work.
+                    // We may also not actually drop a token here if the worker was
+                    // running with an "ephemeral token".
+                    if main_thread_state == MainThreadState::Lending {
+                        main_thread_state = MainThreadState::Idle;
+                    } else {
+                        running_with_own_token -= 1;
+                    }
 
                     match result {
                         Ok(WorkItemResult::Finished(compiled_module)) => {
@@ -1800,7 +1770,6 @@ pub(crate) struct WorkerFatalError;
 fn spawn_work<'a, B: ExtraBackendMethods>(
     cgcx: &'a CodegenContext<B>,
     llvm_start_time: &mut Option<VerboseTimingGuard<'a>>,
-    worker_id: usize,
     work: WorkItem<B>,
 ) {
     if cgcx.config(work.module_kind()).time_module && llvm_start_time.is_none() {
@@ -1815,24 +1784,21 @@ fn spawn_work<'a, B: ExtraBackendMethods>(
         struct Bomb<B: ExtraBackendMethods> {
             coordinator_send: Sender<Box<dyn Any + Send>>,
             result: Option<Result<WorkItemResult<B>, FatalError>>,
-            worker_id: usize,
         }
         impl<B: ExtraBackendMethods> Drop for Bomb<B> {
             fn drop(&mut self) {
-                let worker_id = self.worker_id;
                 let msg = match self.result.take() {
-                    Some(Ok(result)) => Message::WorkItem::<B> { result: Ok(result), worker_id },
+                    Some(Ok(result)) => Message::WorkItem::<B> { result: Ok(result) },
                     Some(Err(FatalError)) => {
-                        Message::WorkItem::<B> { result: Err(Some(WorkerFatalError)), worker_id }
+                        Message::WorkItem::<B> { result: Err(Some(WorkerFatalError)) }
                     }
-                    None => Message::WorkItem::<B> { result: Err(None), worker_id },
+                    None => Message::WorkItem::<B> { result: Err(None) },
                 };
                 drop(self.coordinator_send.send(Box::new(msg)));
             }
         }
 
-        let mut bomb =
-            Bomb::<B> { coordinator_send: cgcx.coordinator_send.clone(), result: None, worker_id };
+        let mut bomb = Bomb::<B> { coordinator_send: cgcx.coordinator_send.clone(), result: None };
 
         // Execute the work itself, and if it finishes successfully then flag
         // ourselves as a success as well.
@@ -1856,12 +1822,20 @@ fn spawn_work<'a, B: ExtraBackendMethods>(
                     );
                     Ok(execute_copy_from_cache_work_item(&cgcx, m, module_config))
                 }
-                WorkItem::FatLto { needs_fat_lto, import_only_modules, autodiff } => {
+                WorkItem::FatLto {
+                    exported_symbols_for_lto,
+                    each_linked_rlib_for_lto,
+                    needs_fat_lto,
+                    import_only_modules,
+                    autodiff,
+                } => {
                     let _timer = cgcx
                         .prof
                         .generic_activity_with_arg("codegen_module_perform_lto", "everything");
                     execute_fat_lto_work_item(
                         &cgcx,
+                        &exported_symbols_for_lto,
+                        &each_linked_rlib_for_lto,
                         needs_fat_lto,
                         import_only_modules,
                         autodiff,
diff --git a/compiler/rustc_codegen_ssa/src/common.rs b/compiler/rustc_codegen_ssa/src/common.rs
index 48565e0b4de..a6fd6c763ed 100644
--- a/compiler/rustc_codegen_ssa/src/common.rs
+++ b/compiler/rustc_codegen_ssa/src/common.rs
@@ -148,7 +148,7 @@ pub(crate) fn shift_mask_val<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
 pub fn asm_const_to_str<'tcx>(
     tcx: TyCtxt<'tcx>,
     sp: Span,
-    const_value: mir::ConstValue<'tcx>,
+    const_value: mir::ConstValue,
     ty_and_layout: TyAndLayout<'tcx>,
 ) -> String {
     let mir::ConstValue::Scalar(scalar) = const_value else {
diff --git a/compiler/rustc_codegen_ssa/src/errors.rs b/compiler/rustc_codegen_ssa/src/errors.rs
index 9040915b6af..3d787d8bdbd 100644
--- a/compiler/rustc_codegen_ssa/src/errors.rs
+++ b/compiler/rustc_codegen_ssa/src/errors.rs
@@ -1294,3 +1294,20 @@ pub(crate) struct FeatureNotValid<'a> {
     #[help]
     pub plus_hint: bool,
 }
+
+#[derive(Diagnostic)]
+#[diag(codegen_ssa_lto_disallowed)]
+pub(crate) struct LtoDisallowed;
+
+#[derive(Diagnostic)]
+#[diag(codegen_ssa_lto_dylib)]
+pub(crate) struct LtoDylib;
+
+#[derive(Diagnostic)]
+#[diag(codegen_ssa_lto_proc_macro)]
+pub(crate) struct LtoProcMacro;
+
+#[derive(Diagnostic)]
+#[diag(codegen_ssa_dynamic_linking_with_lto)]
+#[note]
+pub(crate) struct DynamicLinkingWithLTO;
diff --git a/compiler/rustc_codegen_ssa/src/mir/constant.rs b/compiler/rustc_codegen_ssa/src/mir/constant.rs
index 68a56044a07..11b6ab3cdf1 100644
--- a/compiler/rustc_codegen_ssa/src/mir/constant.rs
+++ b/compiler/rustc_codegen_ssa/src/mir/constant.rs
@@ -20,7 +20,7 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
         OperandRef::from_const(bx, val, ty)
     }
 
-    pub fn eval_mir_constant(&self, constant: &mir::ConstOperand<'tcx>) -> mir::ConstValue<'tcx> {
+    pub fn eval_mir_constant(&self, constant: &mir::ConstOperand<'tcx>) -> mir::ConstValue {
         // `MirUsedCollector` visited all required_consts before codegen began, so if we got here
         // there can be no more constants that fail to evaluate.
         self.monomorphize(constant.const_)
diff --git a/compiler/rustc_codegen_ssa/src/mir/operand.rs b/compiler/rustc_codegen_ssa/src/mir/operand.rs
index 06bedaaa4a2..8e308aac769 100644
--- a/compiler/rustc_codegen_ssa/src/mir/operand.rs
+++ b/compiler/rustc_codegen_ssa/src/mir/operand.rs
@@ -140,7 +140,7 @@ impl<'a, 'tcx, V: CodegenObject> OperandRef<'tcx, V> {
 
     pub(crate) fn from_const<Bx: BuilderMethods<'a, 'tcx, Value = V>>(
         bx: &mut Bx,
-        val: mir::ConstValue<'tcx>,
+        val: mir::ConstValue,
         ty: Ty<'tcx>,
     ) -> Self {
         let layout = bx.layout_of(ty);
@@ -154,14 +154,11 @@ impl<'a, 'tcx, V: CodegenObject> OperandRef<'tcx, V> {
                 OperandValue::Immediate(llval)
             }
             ConstValue::ZeroSized => return OperandRef::zero_sized(layout),
-            ConstValue::Slice { data, meta } => {
+            ConstValue::Slice { alloc_id, meta } => {
                 let BackendRepr::ScalarPair(a_scalar, _) = layout.backend_repr else {
                     bug!("from_const: invalid ScalarPair layout: {:#?}", layout);
                 };
-                let a = Scalar::from_pointer(
-                    Pointer::new(bx.tcx().reserve_and_set_memory_alloc(data).into(), Size::ZERO),
-                    &bx.tcx(),
-                );
+                let a = Scalar::from_pointer(Pointer::new(alloc_id.into(), Size::ZERO), &bx.tcx());
                 let a_llval = bx.scalar_to_backend(
                     a,
                     a_scalar,
diff --git a/compiler/rustc_codegen_ssa/src/mir/rvalue.rs b/compiler/rustc_codegen_ssa/src/mir/rvalue.rs
index e872f8434e5..a5759b79be4 100644
--- a/compiler/rustc_codegen_ssa/src/mir/rvalue.rs
+++ b/compiler/rustc_codegen_ssa/src/mir/rvalue.rs
@@ -288,7 +288,7 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
         // valid ranges. For example, `char`s are passed as just `i32`, with no
         // way for LLVM to know that they're 0x10FFFF at most. Thus we assume
         // the range of the input value too, not just the output range.
-        assume_scalar_range(bx, imm, from_scalar, from_backend_ty);
+        assume_scalar_range(bx, imm, from_scalar, from_backend_ty, None);
 
         imm = match (from_scalar.primitive(), to_scalar.primitive()) {
             (Int(_, is_signed), Int(..)) => bx.intcast(imm, to_backend_ty, is_signed),
@@ -869,7 +869,7 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
                     let ltext = bx.zext(is_lt, bx.type_i8());
                     bx.unchecked_ssub(gtext, ltext)
                 } else {
-                    // These operations are those expected by `tests/codegen/integer-cmp.rs`,
+                    // These operations are those expected by `tests/codegen-llvm/integer-cmp.rs`,
                     // from <https://github.com/rust-lang/rust/pull/63767>.
                     let is_lt = bx.icmp(pred(mir::BinOp::Lt), lhs, rhs);
                     let is_ne = bx.icmp(pred(mir::BinOp::Ne), lhs, rhs);
@@ -1064,7 +1064,7 @@ pub(super) fn transmute_scalar<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
     // That said, last time we tried removing this, it didn't actually help
     // the rustc-perf results, so might as well keep doing it
     // <https://github.com/rust-lang/rust/pull/135610#issuecomment-2599275182>
-    assume_scalar_range(bx, imm, from_scalar, from_backend_ty);
+    assume_scalar_range(bx, imm, from_scalar, from_backend_ty, Some(&to_scalar));
 
     imm = match (from_scalar.primitive(), to_scalar.primitive()) {
         (Int(..) | Float(_), Int(..) | Float(_)) => bx.bitcast(imm, to_backend_ty),
@@ -1092,22 +1092,42 @@ pub(super) fn transmute_scalar<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
     // since it's never passed to something with parameter metadata (especially
     // after MIR inlining) so the only way to tell the backend about the
     // constraint that the `transmute` introduced is to `assume` it.
-    assume_scalar_range(bx, imm, to_scalar, to_backend_ty);
+    assume_scalar_range(bx, imm, to_scalar, to_backend_ty, Some(&from_scalar));
 
     imm = bx.to_immediate_scalar(imm, to_scalar);
     imm
 }
 
+/// Emits an `assume` call that `imm`'s value is within the known range of `scalar`.
+///
+/// If `known` is `Some`, only emits the assume if it's more specific than
+/// whatever is already known from the range of *that* scalar.
 fn assume_scalar_range<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
     bx: &mut Bx,
     imm: Bx::Value,
     scalar: abi::Scalar,
     backend_ty: Bx::Type,
+    known: Option<&abi::Scalar>,
 ) {
-    if matches!(bx.cx().sess().opts.optimize, OptLevel::No) || scalar.is_always_valid(bx.cx()) {
+    if matches!(bx.cx().sess().opts.optimize, OptLevel::No) {
         return;
     }
 
+    match (scalar, known) {
+        (abi::Scalar::Union { .. }, _) => return,
+        (_, None) => {
+            if scalar.is_always_valid(bx.cx()) {
+                return;
+            }
+        }
+        (abi::Scalar::Initialized { valid_range, .. }, Some(known)) => {
+            let known_range = known.valid_range(bx.cx());
+            if valid_range.contains_range(known_range, scalar.size(bx.cx())) {
+                return;
+            }
+        }
+    }
+
     match scalar.primitive() {
         abi::Primitive::Int(..) => {
             let range = scalar.valid_range(bx.cx());
diff --git a/compiler/rustc_codegen_ssa/src/traits/write.rs b/compiler/rustc_codegen_ssa/src/traits/write.rs
index 5e993640472..8e78cbe1963 100644
--- a/compiler/rustc_codegen_ssa/src/traits/write.rs
+++ b/compiler/rustc_codegen_ssa/src/traits/write.rs
@@ -1,3 +1,5 @@
+use std::path::PathBuf;
+
 use rustc_ast::expand::autodiff_attrs::AutoDiffItem;
 use rustc_errors::{DiagCtxtHandle, FatalError};
 use rustc_middle::dep_graph::WorkProduct;
@@ -24,8 +26,9 @@ pub trait WriteBackendMethods: Clone + 'static {
     /// if necessary and running any further optimizations
     fn run_and_optimize_fat_lto(
         cgcx: &CodegenContext<Self>,
+        exported_symbols_for_lto: &[String],
+        each_linked_rlib_for_lto: &[PathBuf],
         modules: Vec<FatLtoInput<Self>>,
-        cached_modules: Vec<(SerializedModule<Self::ModuleBuffer>, WorkProduct)>,
         diff_fncs: Vec<AutoDiffItem>,
     ) -> Result<ModuleCodegen<Self::Module>, FatalError>;
     /// Performs thin LTO by performing necessary global analysis and returning two
@@ -33,6 +36,8 @@ pub trait WriteBackendMethods: Clone + 'static {
     /// can simply be copied over from the incr. comp. cache.
     fn run_thin_lto(
         cgcx: &CodegenContext<Self>,
+        exported_symbols_for_lto: &[String],
+        each_linked_rlib_for_lto: &[PathBuf],
         modules: Vec<(String, Self::ThinBuffer)>,
         cached_modules: Vec<(SerializedModule<Self::ModuleBuffer>, WorkProduct)>,
     ) -> Result<(Vec<ThinModule<Self>>, Vec<WorkProduct>), FatalError>;