Merge pull request #2347 from rust-lang/rustc-pull

Rustc pull update

13 files changed, 525 insertions, 334 deletions

diff --git a/compiler/rustc_codegen_llvm/src/back/write.rs b/compiler/rustc_codegen_llvm/src/back/write.rs
index bf6138142b6..18d221d232e 100644
--- a/compiler/rustc_codegen_llvm/src/back/write.rs
+++ b/compiler/rustc_codegen_llvm/src/back/write.rs
@@ -119,14 +119,18 @@ pub(crate) fn create_target_machine(tcx: TyCtxt<'_>, mod_name: &str) -> OwnedTar
         tcx.output_filenames(()).split_dwarf_path(
             tcx.sess.split_debuginfo(),
             tcx.sess.opts.unstable_opts.split_dwarf_kind,
-            Some(mod_name),
+            mod_name,
+            tcx.sess.invocation_temp.as_deref(),
         )
     } else {
         None
     };
 
-    let output_obj_file =
-        Some(tcx.output_filenames(()).temp_path(OutputType::Object, Some(mod_name)));
+    let output_obj_file = Some(tcx.output_filenames(()).temp_path_for_cgu(
+        OutputType::Object,
+        mod_name,
+        tcx.sess.invocation_temp.as_deref(),
+    ));
     let config = TargetMachineFactoryConfig { split_dwarf_file, output_obj_file };
 
     target_machine_factory(
@@ -330,8 +334,11 @@ pub(crate) fn save_temp_bitcode(
         return;
     }
     let ext = format!("{name}.bc");
-    let cgu = Some(&module.name[..]);
-    let path = cgcx.output_filenames.temp_path_ext(&ext, cgu);
+    let path = cgcx.output_filenames.temp_path_ext_for_cgu(
+        &ext,
+        &module.name,
+        cgcx.invocation_temp.as_deref(),
+    );
     write_bitcode_to_file(module, &path)
 }
 
@@ -439,12 +446,9 @@ fn report_inline_asm(
     let span = if cookie == 0 || matches!(cgcx.lto, Lto::Fat | Lto::Thin) {
         SpanData::default()
     } else {
-        let lo = BytePos::from_u32(cookie as u32);
-        let hi = BytePos::from_u32((cookie >> 32) as u32);
         SpanData {
-            lo,
-            // LLVM version < 19 silently truncates the cookie to 32 bits in some situations.
-            hi: if hi.to_u32() != 0 { hi } else { lo },
+            lo: BytePos::from_u32(cookie as u32),
+            hi: BytePos::from_u32((cookie >> 32) as u32),
             ctxt: SyntaxContext::root(),
             parent: None,
         }
@@ -697,11 +701,12 @@ pub(crate) unsafe fn optimize(
     let llcx = &*module.module_llvm.llcx;
     let _handlers = DiagnosticHandlers::new(cgcx, dcx, llcx, module, CodegenDiagnosticsStage::Opt);
 
-    let module_name = module.name.clone();
-    let module_name = Some(&module_name[..]);
-
     if config.emit_no_opt_bc {
-        let out = cgcx.output_filenames.temp_path_ext("no-opt.bc", module_name);
+        let out = cgcx.output_filenames.temp_path_ext_for_cgu(
+            "no-opt.bc",
+            &module.name,
+            cgcx.invocation_temp.as_deref(),
+        );
         write_bitcode_to_file(module, &out)
     }
 
@@ -746,8 +751,11 @@ pub(crate) unsafe fn optimize(
         if let Some(thin_lto_buffer) = thin_lto_buffer {
             let thin_lto_buffer = unsafe { ThinBuffer::from_raw_ptr(thin_lto_buffer) };
             module.thin_lto_buffer = Some(thin_lto_buffer.data().to_vec());
-            let bc_summary_out =
-                cgcx.output_filenames.temp_path(OutputType::ThinLinkBitcode, module_name);
+            let bc_summary_out = cgcx.output_filenames.temp_path_for_cgu(
+                OutputType::ThinLinkBitcode,
+                &module.name,
+                cgcx.invocation_temp.as_deref(),
+            );
             if config.emit_thin_lto_summary
                 && let Some(thin_link_bitcode_filename) = bc_summary_out.file_name()
             {
@@ -804,8 +812,6 @@ pub(crate) unsafe fn codegen(
         let llmod = module.module_llvm.llmod();
         let llcx = &*module.module_llvm.llcx;
         let tm = &*module.module_llvm.tm;
-        let module_name = module.name.clone();
-        let module_name = Some(&module_name[..]);
         let _handlers =
             DiagnosticHandlers::new(cgcx, dcx, llcx, &module, CodegenDiagnosticsStage::Codegen);
 
@@ -817,8 +823,16 @@ pub(crate) unsafe fn codegen(
         // copy it to the .o file, and delete the bitcode if it wasn't
         // otherwise requested.
 
-        let bc_out = cgcx.output_filenames.temp_path(OutputType::Bitcode, module_name);
-        let obj_out = cgcx.output_filenames.temp_path(OutputType::Object, module_name);
+        let bc_out = cgcx.output_filenames.temp_path_for_cgu(
+            OutputType::Bitcode,
+            &module.name,
+            cgcx.invocation_temp.as_deref(),
+        );
+        let obj_out = cgcx.output_filenames.temp_path_for_cgu(
+            OutputType::Object,
+            &module.name,
+            cgcx.invocation_temp.as_deref(),
+        );
 
         if config.bitcode_needed() {
             if config.emit_bc || config.emit_obj == EmitObj::Bitcode {
@@ -860,7 +874,11 @@ pub(crate) unsafe fn codegen(
         if config.emit_ir {
             let _timer =
                 cgcx.prof.generic_activity_with_arg("LLVM_module_codegen_emit_ir", &*module.name);
-            let out = cgcx.output_filenames.temp_path(OutputType::LlvmAssembly, module_name);
+            let out = cgcx.output_filenames.temp_path_for_cgu(
+                OutputType::LlvmAssembly,
+                &module.name,
+                cgcx.invocation_temp.as_deref(),
+            );
             let out_c = path_to_c_string(&out);
 
             extern "C" fn demangle_callback(
@@ -902,7 +920,11 @@ pub(crate) unsafe fn codegen(
         if config.emit_asm {
             let _timer =
                 cgcx.prof.generic_activity_with_arg("LLVM_module_codegen_emit_asm", &*module.name);
-            let path = cgcx.output_filenames.temp_path(OutputType::Assembly, module_name);
+            let path = cgcx.output_filenames.temp_path_for_cgu(
+                OutputType::Assembly,
+                &module.name,
+                cgcx.invocation_temp.as_deref(),
+            );
 
             // We can't use the same module for asm and object code output,
             // because that triggers various errors like invalid IR or broken
@@ -932,7 +954,9 @@ pub(crate) unsafe fn codegen(
                     .prof
                     .generic_activity_with_arg("LLVM_module_codegen_emit_obj", &*module.name);
 
-                let dwo_out = cgcx.output_filenames.temp_path_dwo(module_name);
+                let dwo_out = cgcx
+                    .output_filenames
+                    .temp_path_dwo_for_cgu(&module.name, cgcx.invocation_temp.as_deref());
                 let dwo_out = match (cgcx.split_debuginfo, cgcx.split_dwarf_kind) {
                     // Don't change how DWARF is emitted when disabled.
                     (SplitDebuginfo::Off, _) => None,
@@ -997,6 +1021,7 @@ pub(crate) unsafe fn codegen(
         config.emit_asm,
         config.emit_ir,
         &cgcx.output_filenames,
+        cgcx.invocation_temp.as_deref(),
     ))
 }
 
diff --git a/compiler/rustc_codegen_llvm/src/builder.rs b/compiler/rustc_codegen_llvm/src/builder.rs
index 35134e9f5a0..27f7f95f100 100644
--- a/compiler/rustc_codegen_llvm/src/builder.rs
+++ b/compiler/rustc_codegen_llvm/src/builder.rs
@@ -123,7 +123,7 @@ impl<'a, 'll, CX: Borrow<SCx<'ll>>> GenericBuilder<'a, 'll, CX> {
 /// Empty string, to be used where LLVM expects an instruction name, indicating
 /// that the instruction is to be left unnamed (i.e. numbered, in textual IR).
 // FIXME(eddyb) pass `&CStr` directly to FFI once it's a thin pointer.
-const UNNAMED: *const c_char = c"".as_ptr();
+pub(crate) const UNNAMED: *const c_char = c"".as_ptr();
 
 impl<'ll, CX: Borrow<SCx<'ll>>> BackendTypes for GenericBuilder<'_, 'll, CX> {
     type Value = <GenericCx<'ll, CX> as BackendTypes>::Value;
diff --git a/compiler/rustc_codegen_llvm/src/builder/autodiff.rs b/compiler/rustc_codegen_llvm/src/builder/autodiff.rs
index 7d264ba4d00..e7c071d05aa 100644
--- a/compiler/rustc_codegen_llvm/src/builder/autodiff.rs
+++ b/compiler/rustc_codegen_llvm/src/builder/autodiff.rs
@@ -10,7 +10,7 @@ use rustc_middle::bug;
 use tracing::{debug, trace};
 
 use crate::back::write::llvm_err;
-use crate::builder::SBuilder;
+use crate::builder::{SBuilder, UNNAMED};
 use crate::context::SimpleCx;
 use crate::declare::declare_simple_fn;
 use crate::errors::{AutoDiffWithoutEnable, LlvmError};
@@ -51,6 +51,7 @@ fn has_sret(fnc: &Value) -> bool {
 // using iterators and peek()?
 fn match_args_from_caller_to_enzyme<'ll>(
     cx: &SimpleCx<'ll>,
+    builder: &SBuilder<'ll, 'll>,
     width: u32,
     args: &mut Vec<&'ll llvm::Value>,
     inputs: &[DiffActivity],
@@ -78,7 +79,9 @@ fn match_args_from_caller_to_enzyme<'ll>(
     let enzyme_const = cx.create_metadata("enzyme_const".to_string()).unwrap();
     let enzyme_out = cx.create_metadata("enzyme_out".to_string()).unwrap();
     let enzyme_dup = cx.create_metadata("enzyme_dup".to_string()).unwrap();
+    let enzyme_dupv = cx.create_metadata("enzyme_dupv".to_string()).unwrap();
     let enzyme_dupnoneed = cx.create_metadata("enzyme_dupnoneed".to_string()).unwrap();
+    let enzyme_dupnoneedv = cx.create_metadata("enzyme_dupnoneedv".to_string()).unwrap();
 
     while activity_pos < inputs.len() {
         let diff_activity = inputs[activity_pos as usize];
@@ -90,13 +93,34 @@ fn match_args_from_caller_to_enzyme<'ll>(
             DiffActivity::Active => (enzyme_out, false),
             DiffActivity::ActiveOnly => (enzyme_out, false),
             DiffActivity::Dual => (enzyme_dup, true),
+            DiffActivity::Dualv => (enzyme_dupv, true),
             DiffActivity::DualOnly => (enzyme_dupnoneed, true),
+            DiffActivity::DualvOnly => (enzyme_dupnoneedv, true),
             DiffActivity::Duplicated => (enzyme_dup, true),
             DiffActivity::DuplicatedOnly => (enzyme_dupnoneed, true),
-            DiffActivity::FakeActivitySize => (enzyme_const, false),
+            DiffActivity::FakeActivitySize(_) => (enzyme_const, false),
         };
         let outer_arg = outer_args[outer_pos];
         args.push(cx.get_metadata_value(activity));
+        if matches!(diff_activity, DiffActivity::Dualv) {
+            let next_outer_arg = outer_args[outer_pos + 1];
+            let elem_bytes_size: u64 = match inputs[activity_pos + 1] {
+                DiffActivity::FakeActivitySize(Some(s)) => s.into(),
+                _ => bug!("incorrect Dualv handling recognized."),
+            };
+            // stride: sizeof(T) * n_elems.
+            // n_elems is the next integer.
+            // Now we multiply `4 * next_outer_arg` to get the stride.
+            let mul = unsafe {
+                llvm::LLVMBuildMul(
+                    builder.llbuilder,
+                    cx.get_const_i64(elem_bytes_size),
+                    next_outer_arg,
+                    UNNAMED,
+                )
+            };
+            args.push(mul);
+        }
         args.push(outer_arg);
         if duplicated {
             // We know that duplicated args by construction have a following argument,
@@ -114,7 +138,7 @@ fn match_args_from_caller_to_enzyme<'ll>(
                 } else {
                     let next_activity = inputs[activity_pos + 1];
                     // We analyze the MIR types and add this dummy activity if we visit a slice.
-                    next_activity == DiffActivity::FakeActivitySize
+                    matches!(next_activity, DiffActivity::FakeActivitySize(_))
                 }
             };
             if slice {
@@ -125,7 +149,10 @@ fn match_args_from_caller_to_enzyme<'ll>(
                 // int2 >= int1, which means the shadow vector is large enough to store the gradient.
                 assert_eq!(cx.type_kind(next_outer_ty), TypeKind::Integer);
 
-                for i in 0..(width as usize) {
+                let iterations =
+                    if matches!(diff_activity, DiffActivity::Dualv) { 1 } else { width as usize };
+
+                for i in 0..iterations {
                     let next_outer_arg2 = outer_args[outer_pos + 2 * (i + 1)];
                     let next_outer_ty2 = cx.val_ty(next_outer_arg2);
                     assert_eq!(cx.type_kind(next_outer_ty2), TypeKind::Pointer);
@@ -136,7 +163,7 @@ fn match_args_from_caller_to_enzyme<'ll>(
                 }
                 args.push(cx.get_metadata_value(enzyme_const));
                 args.push(next_outer_arg);
-                outer_pos += 2 + 2 * width as usize;
+                outer_pos += 2 + 2 * iterations;
                 activity_pos += 2;
             } else {
                 // A duplicated pointer will have the following two outer_fn arguments:
@@ -201,7 +228,23 @@ fn compute_enzyme_fn_ty<'ll>(
         }
 
         if attrs.width == 1 {
-            todo!("Handle sret for scalar ad");
+            // Enzyme returns a struct of style:
+            // `{ original_ret(if requested), float, float, ... }`
+            let mut struct_elements = vec![];
+            if attrs.has_primal_ret() {
+                struct_elements.push(inner_ret_ty);
+            }
+            // Next, we push the list of active floats, since they will be lowered to `enzyme_out`,
+            // and therefore part of the return struct.
+            let param_tys = cx.func_params_types(fn_ty);
+            for (act, param_ty) in attrs.input_activity.iter().zip(param_tys) {
+                if matches!(act, DiffActivity::Active) {
+                    // Now find the float type at position i based on the fn_ty,
+                    // to know what (f16/f32/f64/...) to add to the struct.
+                    struct_elements.push(param_ty);
+                }
+            }
+            ret_ty = cx.type_struct(&struct_elements, false);
         } else {
             // First we check if we also have to deal with the primal return.
             match attrs.mode {
@@ -344,6 +387,7 @@ fn generate_enzyme_call<'ll>(
         let outer_args: Vec<&llvm::Value> = get_params(outer_fn);
         match_args_from_caller_to_enzyme(
             &cx,
+            &builder,
             attrs.width,
             &mut args,
             &attrs.input_activity,
@@ -388,7 +432,11 @@ fn generate_enzyme_call<'ll>(
                 // now store the result of the enzyme call into the sret pointer.
                 let sret_ptr = outer_args[0];
                 let call_ty = cx.val_ty(call);
-                assert_eq!(cx.type_kind(call_ty), TypeKind::Array);
+                if attrs.width == 1 {
+                    assert_eq!(cx.type_kind(call_ty), TypeKind::Struct);
+                } else {
+                    assert_eq!(cx.type_kind(call_ty), TypeKind::Array);
+                }
                 llvm::LLVMBuildStore(&builder.llbuilder, call, sret_ptr);
             }
             builder.ret_void();
diff --git a/compiler/rustc_codegen_llvm/src/common.rs b/compiler/rustc_codegen_llvm/src/common.rs
index 457e5452ce9..a6f277e4455 100644
--- a/compiler/rustc_codegen_llvm/src/common.rs
+++ b/compiler/rustc_codegen_llvm/src/common.rs
@@ -4,8 +4,8 @@ use std::borrow::Borrow;
 
 use libc::{c_char, c_uint};
 use rustc_abi as abi;
+use rustc_abi::HasDataLayout;
 use rustc_abi::Primitive::Pointer;
-use rustc_abi::{AddressSpace, HasDataLayout};
 use rustc_ast::Mutability;
 use rustc_codegen_ssa::common::TypeKind;
 use rustc_codegen_ssa::traits::*;
@@ -269,7 +269,8 @@ impl<'ll, 'tcx> ConstCodegenMethods for CodegenCx<'ll, 'tcx> {
             }
             Scalar::Ptr(ptr, _size) => {
                 let (prov, offset) = ptr.into_parts();
-                let (base_addr, base_addr_space) = match self.tcx.global_alloc(prov.alloc_id()) {
+                let global_alloc = self.tcx.global_alloc(prov.alloc_id());
+                let base_addr = match global_alloc {
                     GlobalAlloc::Memory(alloc) => {
                         // For ZSTs directly codegen an aligned pointer.
                         // This avoids generating a zero-sized constant value and actually needing a
@@ -301,12 +302,10 @@ impl<'ll, 'tcx> ConstCodegenMethods for CodegenCx<'ll, 'tcx> {
                                     format!("alloc_{hash:032x}").as_bytes(),
                                 );
                             }
-                            (value, AddressSpace::DATA)
+                            value
                         }
                     }
-                    GlobalAlloc::Function { instance, .. } => {
-                        (self.get_fn_addr(instance), self.data_layout().instruction_address_space)
-                    }
+                    GlobalAlloc::Function { instance, .. } => self.get_fn_addr(instance),
                     GlobalAlloc::VTable(ty, dyn_ty) => {
                         let alloc = self
                             .tcx
@@ -319,14 +318,15 @@ impl<'ll, 'tcx> ConstCodegenMethods for CodegenCx<'ll, 'tcx> {
                             .unwrap_memory();
                         let init = const_alloc_to_llvm(self, alloc, /*static*/ false);
                         let value = self.static_addr_of_impl(init, alloc.inner().align, None);
-                        (value, AddressSpace::DATA)
+                        value
                     }
                     GlobalAlloc::Static(def_id) => {
                         assert!(self.tcx.is_static(def_id));
                         assert!(!self.tcx.is_thread_local_static(def_id));
-                        (self.get_static(def_id), AddressSpace::DATA)
+                        self.get_static(def_id)
                     }
                 };
+                let base_addr_space = global_alloc.address_space(self);
                 let llval = unsafe {
                     llvm::LLVMConstInBoundsGEP2(
                         self.type_i8(),
diff --git a/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen.rs b/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen.rs
index 9a2473d6cf2..55b1e728b70 100644
--- a/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen.rs
+++ b/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen.rs
@@ -5,15 +5,11 @@ use rustc_abi::Align;
 use rustc_codegen_ssa::traits::{
     BaseTypeCodegenMethods, ConstCodegenMethods, StaticCodegenMethods,
 };
-use rustc_data_structures::fx::{FxHashSet, FxIndexMap};
-use rustc_hir::def_id::{DefId, LocalDefId};
+use rustc_data_structures::fx::FxIndexMap;
 use rustc_index::IndexVec;
-use rustc_middle::mir;
-use rustc_middle::mir::mono::MonoItemPartitions;
-use rustc_middle::ty::{self, TyCtxt};
+use rustc_middle::ty::TyCtxt;
 use rustc_session::RemapFileNameExt;
 use rustc_session::config::RemapPathScopeComponents;
-use rustc_span::def_id::DefIdSet;
 use rustc_span::{SourceFile, StableSourceFileId};
 use tracing::debug;
 
@@ -24,6 +20,7 @@ use crate::llvm;
 
 mod covfun;
 mod spans;
+mod unused;
 
 /// Generates and exports the coverage map, which is embedded in special
 /// linker sections in the final binary.
@@ -56,13 +53,6 @@ pub(crate) fn finalize(cx: &CodegenCx<'_, '_>) {
         None => return,
     };
 
-    // The order of entries in this global file table needs to be deterministic,
-    // and ideally should also be independent of the details of stable-hashing,
-    // because coverage tests snapshots (`.cov-map`) can observe the order and
-    // would need to be re-blessed if it changes. As long as those requirements
-    // are satisfied, the order can be arbitrary.
-    let mut global_file_table = GlobalFileTable::new();
-
     let mut covfun_records = instances_used
         .iter()
         .copied()
@@ -70,18 +60,13 @@ pub(crate) fn finalize(cx: &CodegenCx<'_, '_>) {
         // order that doesn't depend on the stable-hash-based order in which
         // instances were visited during codegen.
         .sorted_by_cached_key(|&instance| tcx.symbol_name(instance).name)
-        .filter_map(|instance| prepare_covfun_record(tcx, &mut global_file_table, instance, true))
+        .filter_map(|instance| prepare_covfun_record(tcx, instance, true))
         .collect::<Vec<_>>();
 
     // In a single designated CGU, also prepare covfun records for functions
     // in this crate that were instrumented for coverage, but are unused.
     if cx.codegen_unit.is_code_coverage_dead_code_cgu() {
-        let mut unused_instances = gather_unused_function_instances(cx);
-        // Sort the unused instances by symbol name, for the same reason as the used ones.
-        unused_instances.sort_by_cached_key(|&instance| tcx.symbol_name(instance).name);
-        covfun_records.extend(unused_instances.into_iter().filter_map(|instance| {
-            prepare_covfun_record(tcx, &mut global_file_table, instance, false)
-        }));
+        unused::prepare_covfun_records_for_unused_functions(cx, &mut covfun_records);
     }
 
     // If there are no covfun records for this CGU, don't generate a covmap record.
@@ -93,91 +78,88 @@ pub(crate) fn finalize(cx: &CodegenCx<'_, '_>) {
         return;
     }
 
-    // Encode all filenames referenced by coverage mappings in this CGU.
-    let filenames_buffer = global_file_table.make_filenames_buffer(tcx);
-    // The `llvm-cov` tool uses this hash to associate each covfun record with
-    // its corresponding filenames table, since the final binary will typically
-    // contain multiple covmap records from different compilation units.
-    let filenames_hash = llvm_cov::hash_bytes(&filenames_buffer);
-
-    let mut unused_function_names = vec![];
+    // Prepare the global file table for this CGU, containing all paths needed
+    // by one or more covfun records.
+    let global_file_table =
+        GlobalFileTable::build(tcx, covfun_records.iter().flat_map(|c| c.all_source_files()));
 
     for covfun in &covfun_records {
-        unused_function_names.extend(covfun.mangled_function_name_if_unused());
-
-        covfun::generate_covfun_record(cx, filenames_hash, covfun)
-    }
-
-    // For unused functions, we need to take their mangled names and store them
-    // in a specially-named global array. LLVM's `InstrProfiling` pass will
-    // detect this global and include those names in its `__llvm_prf_names`
-    // section. (See `llvm/lib/Transforms/Instrumentation/InstrProfiling.cpp`.)
-    if !unused_function_names.is_empty() {
-        assert!(cx.codegen_unit.is_code_coverage_dead_code_cgu());
-
-        let name_globals = unused_function_names
-            .into_iter()
-            .map(|mangled_function_name| cx.const_str(mangled_function_name).0)
-            .collect::<Vec<_>>();
-        let initializer = cx.const_array(cx.type_ptr(), &name_globals);
-
-        let array = llvm::add_global(cx.llmod, cx.val_ty(initializer), c"__llvm_coverage_names");
-        llvm::set_global_constant(array, true);
-        llvm::set_linkage(array, llvm::Linkage::InternalLinkage);
-        llvm::set_initializer(array, initializer);
+        covfun::generate_covfun_record(cx, &global_file_table, covfun)
     }
 
     // Generate the coverage map header, which contains the filenames used by
     // this CGU's coverage mappings, and store it in a well-known global.
     // (This is skipped if we returned early due to having no covfun records.)
-    generate_covmap_record(cx, covmap_version, &filenames_buffer);
+    generate_covmap_record(cx, covmap_version, &global_file_table.filenames_buffer);
 }
 
-/// Maps "global" (per-CGU) file ID numbers to their underlying source files.
+/// Maps "global" (per-CGU) file ID numbers to their underlying source file paths.
+#[derive(Debug)]
 struct GlobalFileTable {
     /// This "raw" table doesn't include the working dir, so a file's
     /// global ID is its index in this set **plus one**.
-    raw_file_table: FxIndexMap<StableSourceFileId, Arc<SourceFile>>,
+    raw_file_table: FxIndexMap<StableSourceFileId, String>,
+
+    /// The file table in encoded form (possibly compressed), which can be
+    /// included directly in this CGU's `__llvm_covmap` record.
+    filenames_buffer: Vec<u8>,
+
+    /// Truncated hash of the bytes in `filenames_buffer`.
+    ///
+    /// The `llvm-cov` tool uses this hash to associate each covfun record with
+    /// its corresponding filenames table, since the final binary will typically
+    /// contain multiple covmap records from different compilation units.
+    filenames_hash: u64,
 }
 
 impl GlobalFileTable {
-    fn new() -> Self {
-        Self { raw_file_table: FxIndexMap::default() }
-    }
+    /// Builds a "global file table" for this CGU, mapping numeric IDs to
+    /// path strings.
+    fn build<'a>(tcx: TyCtxt<'_>, all_files: impl Iterator<Item = &'a SourceFile>) -> Self {
+        let mut raw_file_table = FxIndexMap::default();
+
+        for file in all_files {
+            raw_file_table.entry(file.stable_id).or_insert_with(|| {
+                file.name
+                    .for_scope(tcx.sess, RemapPathScopeComponents::MACRO)
+                    .to_string_lossy()
+                    .into_owned()
+            });
+        }
 
-    fn global_file_id_for_file(&mut self, file: &Arc<SourceFile>) -> GlobalFileId {
-        // Ensure the given file has a table entry, and get its index.
-        let entry = self.raw_file_table.entry(file.stable_id);
-        let raw_id = entry.index();
-        entry.or_insert_with(|| Arc::clone(file));
+        // FIXME(Zalathar): Consider sorting the file table here, but maybe
+        // only after adding filename support to coverage-dump, so that the
+        // table order isn't directly visible in `.coverage-map` snapshots.
 
-        // The raw file table doesn't include an entry for the working dir
-        // (which has ID 0), so add 1 to get the correct ID.
-        GlobalFileId::from_usize(raw_id + 1)
-    }
+        let mut table = Vec::with_capacity(raw_file_table.len() + 1);
 
-    fn make_filenames_buffer(&self, tcx: TyCtxt<'_>) -> Vec<u8> {
-        let mut table = Vec::with_capacity(self.raw_file_table.len() + 1);
-
-        // LLVM Coverage Mapping Format version 6 (zero-based encoded as 5)
-        // requires setting the first filename to the compilation directory.
-        // Since rustc generates coverage maps with relative paths, the
-        // compilation directory can be combined with the relative paths
-        // to get absolute paths, if needed.
-        table.push(
-            tcx.sess
-                .opts
-                .working_dir
-                .for_scope(tcx.sess, RemapPathScopeComponents::MACRO)
-                .to_string_lossy(),
-        );
+        // Since version 6 of the LLVM coverage mapping format, the first entry
+        // in the global file table is treated as a base directory, used to
+        // resolve any other entries that are stored as relative paths.
+        let base_dir = tcx
+            .sess
+            .opts
+            .working_dir
+            .for_scope(tcx.sess, RemapPathScopeComponents::MACRO)
+            .to_string_lossy();
+        table.push(base_dir.as_ref());
 
         // Add the regular entries after the base directory.
-        table.extend(self.raw_file_table.values().map(|file| {
-            file.name.for_scope(tcx.sess, RemapPathScopeComponents::MACRO).to_string_lossy()
-        }));
+        table.extend(raw_file_table.values().map(|name| name.as_str()));
 
-        llvm_cov::write_filenames_to_buffer(&table)
+        // Encode the file table into a buffer, and get the hash of its encoded
+        // bytes, so that we can embed that hash in `__llvm_covfun` records.
+        let filenames_buffer = llvm_cov::write_filenames_to_buffer(&table);
+        let filenames_hash = llvm_cov::hash_bytes(&filenames_buffer);
+
+        Self { raw_file_table, filenames_buffer, filenames_hash }
+    }
+
+    fn get_existing_id(&self, file: &SourceFile) -> Option<GlobalFileId> {
+        let raw_id = self.raw_file_table.get_index_of(&file.stable_id)?;
+        // The raw file table doesn't include an entry for the base dir
+        // (which has ID 0), so add 1 to get the correct ID.
+        Some(GlobalFileId::from_usize(raw_id + 1))
     }
 }
 
@@ -193,26 +175,31 @@ rustc_index::newtype_index! {
     struct LocalFileId {}
 }
 
-/// Holds a mapping from "local" (per-function) file IDs to "global" (per-CGU)
-/// file IDs.
+/// Holds a mapping from "local" (per-function) file IDs to their corresponding
+/// source files.
 #[derive(Debug, Default)]
 struct VirtualFileMapping {
-    local_to_global: IndexVec<LocalFileId, GlobalFileId>,
-    global_to_local: FxIndexMap<GlobalFileId, LocalFileId>,
+    local_file_table: IndexVec<LocalFileId, Arc<SourceFile>>,
 }
 
 impl VirtualFileMapping {
-    fn local_id_for_global(&mut self, global_file_id: GlobalFileId) -> LocalFileId {
-        *self
-            .global_to_local
-            .entry(global_file_id)
-            .or_insert_with(|| self.local_to_global.push(global_file_id))
+    fn push_file(&mut self, source_file: &Arc<SourceFile>) -> LocalFileId {
+        self.local_file_table.push(Arc::clone(source_file))
     }
 
-    fn to_vec(&self) -> Vec<u32> {
-        // This clone could be avoided by transmuting `&[GlobalFileId]` to `&[u32]`,
-        // but it isn't hot or expensive enough to justify the extra unsafety.
-        self.local_to_global.iter().map(|&global| GlobalFileId::as_u32(global)).collect()
+    /// Resolves all of the filenames in this local file mapping to a list of
+    /// global file IDs in its CGU, for inclusion in this function's
+    /// `__llvm_covfun` record.
+    ///
+    /// The global file IDs are returned as `u32` to make FFI easier.
+    fn resolve_all(&self, global_file_table: &GlobalFileTable) -> Option<Vec<u32>> {
+        self.local_file_table
+            .iter()
+            .map(|file| try {
+                let id = global_file_table.get_existing_id(file)?;
+                GlobalFileId::as_u32(id)
+            })
+            .collect::<Option<Vec<_>>>()
     }
 }
 
@@ -249,121 +236,3 @@ fn generate_covmap_record<'ll>(cx: &CodegenCx<'ll, '_>, version: u32, filenames_
 
     cx.add_used_global(covmap_global);
 }
-
-/// Each CGU will normally only emit coverage metadata for the functions that it actually generates.
-/// But since we don't want unused functions to disappear from coverage reports, we also scan for
-/// functions that were instrumented but are not participating in codegen.
-///
-/// These unused functions don't need to be codegenned, but we do need to add them to the function
-/// coverage map (in a single designated CGU) so that we still emit coverage mappings for them.
-/// We also end up adding their symbol names to a special global array that LLVM will include in
-/// its embedded coverage data.
-fn gather_unused_function_instances<'tcx>(cx: &CodegenCx<'_, 'tcx>) -> Vec<ty::Instance<'tcx>> {
-    assert!(cx.codegen_unit.is_code_coverage_dead_code_cgu());
-
-    let tcx = cx.tcx;
-    let usage = prepare_usage_sets(tcx);
-
-    let is_unused_fn = |def_id: LocalDefId| -> bool {
-        // Usage sets expect `DefId`, so convert from `LocalDefId`.
-        let d: DefId = LocalDefId::to_def_id(def_id);
-        // To be potentially eligible for "unused function" mappings, a definition must:
-        // - Be eligible for coverage instrumentation
-        // - Not participate directly in codegen (or have lost all its coverage statements)
-        // - Not have any coverage statements inlined into codegenned functions
-        tcx.is_eligible_for_coverage(def_id)
-            && (!usage.all_mono_items.contains(&d) || usage.missing_own_coverage.contains(&d))
-            && !usage.used_via_inlining.contains(&d)
-    };
-
-    // FIXME(#79651): Consider trying to filter out dummy instantiations of
-    // unused generic functions from library crates, because they can produce
-    // "unused instantiation" in coverage reports even when they are actually
-    // used by some downstream crate in the same binary.
-
-    tcx.mir_keys(())
-        .iter()
-        .copied()
-        .filter(|&def_id| is_unused_fn(def_id))
-        .map(|def_id| make_dummy_instance(tcx, def_id))
-        .collect::<Vec<_>>()
-}
-
-struct UsageSets<'tcx> {
-    all_mono_items: &'tcx DefIdSet,
-    used_via_inlining: FxHashSet<DefId>,
-    missing_own_coverage: FxHashSet<DefId>,
-}
-
-/// Prepare sets of definitions that are relevant to deciding whether something
-/// is an "unused function" for coverage purposes.
-fn prepare_usage_sets<'tcx>(tcx: TyCtxt<'tcx>) -> UsageSets<'tcx> {
-    let MonoItemPartitions { all_mono_items, codegen_units, .. } =
-        tcx.collect_and_partition_mono_items(());
-
-    // Obtain a MIR body for each function participating in codegen, via an
-    // arbitrary instance.
-    let mut def_ids_seen = FxHashSet::default();
-    let def_and_mir_for_all_mono_fns = codegen_units
-        .iter()
-        .flat_map(|cgu| cgu.items().keys())
-        .filter_map(|item| match item {
-            mir::mono::MonoItem::Fn(instance) => Some(instance),
-            mir::mono::MonoItem::Static(_) | mir::mono::MonoItem::GlobalAsm(_) => None,
-        })
-        // We only need one arbitrary instance per definition.
-        .filter(move |instance| def_ids_seen.insert(instance.def_id()))
-        .map(|instance| {
-            // We don't care about the instance, just its underlying MIR.
-            let body = tcx.instance_mir(instance.def);
-            (instance.def_id(), body)
-        });
-
-    // Functions whose coverage statements were found inlined into other functions.
-    let mut used_via_inlining = FxHashSet::default();
-    // Functions that were instrumented, but had all of their coverage statements
-    // removed by later MIR transforms (e.g. UnreachablePropagation).
-    let mut missing_own_coverage = FxHashSet::default();
-
-    for (def_id, body) in def_and_mir_for_all_mono_fns {
-        let mut saw_own_coverage = false;
-
-        // Inspect every coverage statement in the function's MIR.
-        for stmt in body
-            .basic_blocks
-            .iter()
-            .flat_map(|block| &block.statements)
-            .filter(|stmt| matches!(stmt.kind, mir::StatementKind::Coverage(_)))
-        {
-            if let Some(inlined) = stmt.source_info.scope.inlined_instance(&body.source_scopes) {
-                // This coverage statement was inlined from another function.
-                used_via_inlining.insert(inlined.def_id());
-            } else {
-                // Non-inlined coverage statements belong to the enclosing function.
-                saw_own_coverage = true;
-            }
-        }
-
-        if !saw_own_coverage && body.function_coverage_info.is_some() {
-            missing_own_coverage.insert(def_id);
-        }
-    }
-
-    UsageSets { all_mono_items, used_via_inlining, missing_own_coverage }
-}
-
-fn make_dummy_instance<'tcx>(tcx: TyCtxt<'tcx>, local_def_id: LocalDefId) -> ty::Instance<'tcx> {
-    let def_id = local_def_id.to_def_id();
-
-    // Make a dummy instance that fills in all generics with placeholders.
-    ty::Instance::new(
-        def_id,
-        ty::GenericArgs::for_item(tcx, def_id, |param, _| {
-            if let ty::GenericParamDefKind::Lifetime = param.kind {
-                tcx.lifetimes.re_erased.into()
-            } else {
-                tcx.mk_param_from_def(param)
-            }
-        }),
-    )
-}
diff --git a/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen/covfun.rs b/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen/covfun.rs
index 048e1988c32..7bdbc685952 100644
--- a/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen/covfun.rs
+++ b/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen/covfun.rs
@@ -5,6 +5,7 @@
 //! [^win]: On Windows the section name is `.lcovfun`.
 
 use std::ffi::CString;
+use std::sync::Arc;
 
 use rustc_abi::Align;
 use rustc_codegen_ssa::traits::{
@@ -15,7 +16,7 @@ use rustc_middle::mir::coverage::{
     MappingKind, Op,
 };
 use rustc_middle::ty::{Instance, TyCtxt};
-use rustc_span::Span;
+use rustc_span::{SourceFile, Span};
 use rustc_target::spec::HasTargetSpec;
 use tracing::debug;
 
@@ -38,16 +39,15 @@ pub(crate) struct CovfunRecord<'tcx> {
 }
 
 impl<'tcx> CovfunRecord<'tcx> {
-    /// FIXME(Zalathar): Make this the responsibility of the code that determines
-    /// which functions are unused.
-    pub(crate) fn mangled_function_name_if_unused(&self) -> Option<&'tcx str> {
-        (!self.is_used).then_some(self.mangled_function_name)
+    /// Iterator that yields all source files referred to by this function's
+    /// coverage mappings. Used to build the global file table for the CGU.
+    pub(crate) fn all_source_files(&self) -> impl Iterator<Item = &SourceFile> {
+        self.virtual_file_mapping.local_file_table.iter().map(Arc::as_ref)
     }
 }
 
 pub(crate) fn prepare_covfun_record<'tcx>(
     tcx: TyCtxt<'tcx>,
-    global_file_table: &mut GlobalFileTable,
     instance: Instance<'tcx>,
     is_used: bool,
 ) -> Option<CovfunRecord<'tcx>> {
@@ -65,7 +65,7 @@ pub(crate) fn prepare_covfun_record<'tcx>(
         regions: ffi::Regions::default(),
     };
 
-    fill_region_tables(tcx, global_file_table, fn_cov_info, ids_info, &mut covfun);
+    fill_region_tables(tcx, fn_cov_info, ids_info, &mut covfun);
 
     if covfun.regions.has_no_regions() {
         debug!(?covfun, "function has no mappings to embed; skipping");
@@ -100,7 +100,6 @@ fn prepare_expressions(ids_info: &CoverageIdsInfo) -> Vec<ffi::CounterExpression
 /// Populates the mapping region tables in the current function's covfun record.
 fn fill_region_tables<'tcx>(
     tcx: TyCtxt<'tcx>,
-    global_file_table: &mut GlobalFileTable,
     fn_cov_info: &'tcx FunctionCoverageInfo,
     ids_info: &'tcx CoverageIdsInfo,
     covfun: &mut CovfunRecord<'tcx>,
@@ -114,11 +113,7 @@ fn fill_region_tables<'tcx>(
     };
     let source_file = source_map.lookup_source_file(first_span.lo());
 
-    // Look up the global file ID for that file.
-    let global_file_id = global_file_table.global_file_id_for_file(&source_file);
-
-    // Associate that global file ID with a local file ID for this function.
-    let local_file_id = covfun.virtual_file_mapping.local_id_for_global(global_file_id);
+    let local_file_id = covfun.virtual_file_mapping.push_file(&source_file);
 
     // In rare cases, _all_ of a function's spans are discarded, and coverage
     // codegen needs to handle that gracefully to avoid #133606.
@@ -187,7 +182,7 @@ fn fill_region_tables<'tcx>(
 /// as a global variable in the `__llvm_covfun` section.
 pub(crate) fn generate_covfun_record<'tcx>(
     cx: &CodegenCx<'_, 'tcx>,
-    filenames_hash: u64,
+    global_file_table: &GlobalFileTable,
     covfun: &CovfunRecord<'tcx>,
 ) {
     let &CovfunRecord {
@@ -199,12 +194,19 @@ pub(crate) fn generate_covfun_record<'tcx>(
         ref regions,
     } = covfun;
 
+    let Some(local_file_table) = virtual_file_mapping.resolve_all(global_file_table) else {
+        debug_assert!(
+            false,
+            "all local files should be present in the global file table: \
+                global_file_table = {global_file_table:?}, \
+                virtual_file_mapping = {virtual_file_mapping:?}"
+        );
+        return;
+    };
+
     // Encode the function's coverage mappings into a buffer.
-    let coverage_mapping_buffer = llvm_cov::write_function_mappings_to_buffer(
-        &virtual_file_mapping.to_vec(),
-        expressions,
-        regions,
-    );
+    let coverage_mapping_buffer =
+        llvm_cov::write_function_mappings_to_buffer(&local_file_table, expressions, regions);
 
     // A covfun record consists of four target-endian integers, followed by the
     // encoded mapping data in bytes. Note that the length field is 32 bits.
@@ -217,7 +219,7 @@ pub(crate) fn generate_covfun_record<'tcx>(
             cx.const_u64(func_name_hash),
             cx.const_u32(coverage_mapping_buffer.len() as u32),
             cx.const_u64(source_hash),
-            cx.const_u64(filenames_hash),
+            cx.const_u64(global_file_table.filenames_hash),
             cx.const_bytes(&coverage_mapping_buffer),
         ],
         // This struct needs to be packed, so that the 32-bit length field
diff --git a/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen/unused.rs b/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen/unused.rs
new file mode 100644
index 00000000000..68f60f169b5
--- /dev/null
+++ b/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen/unused.rs
@@ -0,0 +1,170 @@
+use rustc_codegen_ssa::traits::{BaseTypeCodegenMethods, ConstCodegenMethods};
+use rustc_data_structures::fx::FxHashSet;
+use rustc_hir::def_id::{DefId, LocalDefId};
+use rustc_middle::mir;
+use rustc_middle::mir::mono::MonoItemPartitions;
+use rustc_middle::ty::{self, TyCtxt};
+use rustc_span::def_id::DefIdSet;
+
+use crate::common::CodegenCx;
+use crate::coverageinfo::mapgen::covfun::{CovfunRecord, prepare_covfun_record};
+use crate::llvm;
+
+/// Each CGU will normally only emit coverage metadata for the functions that it actually generates.
+/// But since we don't want unused functions to disappear from coverage reports, we also scan for
+/// functions that were instrumented but are not participating in codegen.
+///
+/// These unused functions don't need to be codegenned, but we do need to add them to the function
+/// coverage map (in a single designated CGU) so that we still emit coverage mappings for them.
+/// We also end up adding their symbol names to a special global array that LLVM will include in
+/// its embedded coverage data.
+pub(crate) fn prepare_covfun_records_for_unused_functions<'tcx>(
+    cx: &CodegenCx<'_, 'tcx>,
+    covfun_records: &mut Vec<CovfunRecord<'tcx>>,
+) {
+    assert!(cx.codegen_unit.is_code_coverage_dead_code_cgu());
+
+    let mut unused_instances = gather_unused_function_instances(cx);
+    // Sort the unused instances by symbol name, so that their order isn't hash-sensitive.
+    unused_instances.sort_by_key(|instance| instance.symbol_name);
+
+    // Try to create a covfun record for each unused function.
+    let mut name_globals = Vec::with_capacity(unused_instances.len());
+    covfun_records.extend(unused_instances.into_iter().filter_map(|unused| try {
+        let record = prepare_covfun_record(cx.tcx, unused.instance, false)?;
+        // If successful, also store its symbol name in a global constant.
+        name_globals.push(cx.const_str(unused.symbol_name.name).0);
+        record
+    }));
+
+    // Store the names of unused functions in a specially-named global array.
+    // LLVM's `InstrProfilling` pass will detect this array, and include the
+    // referenced names in its `__llvm_prf_names` section.
+    // (See `llvm/lib/Transforms/Instrumentation/InstrProfiling.cpp`.)
+    if !name_globals.is_empty() {
+        let initializer = cx.const_array(cx.type_ptr(), &name_globals);
+
+        let array = llvm::add_global(cx.llmod, cx.val_ty(initializer), c"__llvm_coverage_names");
+        llvm::set_global_constant(array, true);
+        llvm::set_linkage(array, llvm::Linkage::InternalLinkage);
+        llvm::set_initializer(array, initializer);
+    }
+}
+
+/// Holds a dummy function instance along with its symbol name, to avoid having
+/// to repeatedly query for the name.
+struct UnusedInstance<'tcx> {
+    instance: ty::Instance<'tcx>,
+    symbol_name: ty::SymbolName<'tcx>,
+}
+
+fn gather_unused_function_instances<'tcx>(cx: &CodegenCx<'_, 'tcx>) -> Vec<UnusedInstance<'tcx>> {
+    assert!(cx.codegen_unit.is_code_coverage_dead_code_cgu());
+
+    let tcx = cx.tcx;
+    let usage = prepare_usage_sets(tcx);
+
+    let is_unused_fn = |def_id: LocalDefId| -> bool {
+        // Usage sets expect `DefId`, so convert from `LocalDefId`.
+        let d: DefId = LocalDefId::to_def_id(def_id);
+        // To be potentially eligible for "unused function" mappings, a definition must:
+        // - Be eligible for coverage instrumentation
+        // - Not participate directly in codegen (or have lost all its coverage statements)
+        // - Not have any coverage statements inlined into codegenned functions
+        tcx.is_eligible_for_coverage(def_id)
+            && (!usage.all_mono_items.contains(&d) || usage.missing_own_coverage.contains(&d))
+            && !usage.used_via_inlining.contains(&d)
+    };
+
+    // FIXME(#79651): Consider trying to filter out dummy instantiations of
+    // unused generic functions from library crates, because they can produce
+    // "unused instantiation" in coverage reports even when they are actually
+    // used by some downstream crate in the same binary.
+
+    tcx.mir_keys(())
+        .iter()
+        .copied()
+        .filter(|&def_id| is_unused_fn(def_id))
+        .map(|def_id| make_dummy_instance(tcx, def_id))
+        .map(|instance| UnusedInstance { instance, symbol_name: tcx.symbol_name(instance) })
+        .collect::<Vec<_>>()
+}
+
+struct UsageSets<'tcx> {
+    all_mono_items: &'tcx DefIdSet,
+    used_via_inlining: FxHashSet<DefId>,
+    missing_own_coverage: FxHashSet<DefId>,
+}
+
+/// Prepare sets of definitions that are relevant to deciding whether something
+/// is an "unused function" for coverage purposes.
+fn prepare_usage_sets<'tcx>(tcx: TyCtxt<'tcx>) -> UsageSets<'tcx> {
+    let MonoItemPartitions { all_mono_items, codegen_units, .. } =
+        tcx.collect_and_partition_mono_items(());
+
+    // Obtain a MIR body for each function participating in codegen, via an
+    // arbitrary instance.
+    let mut def_ids_seen = FxHashSet::default();
+    let def_and_mir_for_all_mono_fns = codegen_units
+        .iter()
+        .flat_map(|cgu| cgu.items().keys())
+        .filter_map(|item| match item {
+            mir::mono::MonoItem::Fn(instance) => Some(instance),
+            mir::mono::MonoItem::Static(_) | mir::mono::MonoItem::GlobalAsm(_) => None,
+        })
+        // We only need one arbitrary instance per definition.
+        .filter(move |instance| def_ids_seen.insert(instance.def_id()))
+        .map(|instance| {
+            // We don't care about the instance, just its underlying MIR.
+            let body = tcx.instance_mir(instance.def);
+            (instance.def_id(), body)
+        });
+
+    // Functions whose coverage statements were found inlined into other functions.
+    let mut used_via_inlining = FxHashSet::default();
+    // Functions that were instrumented, but had all of their coverage statements
+    // removed by later MIR transforms (e.g. UnreachablePropagation).
+    let mut missing_own_coverage = FxHashSet::default();
+
+    for (def_id, body) in def_and_mir_for_all_mono_fns {
+        let mut saw_own_coverage = false;
+
+        // Inspect every coverage statement in the function's MIR.
+        for stmt in body
+            .basic_blocks
+            .iter()
+            .flat_map(|block| &block.statements)
+            .filter(|stmt| matches!(stmt.kind, mir::StatementKind::Coverage(_)))
+        {
+            if let Some(inlined) = stmt.source_info.scope.inlined_instance(&body.source_scopes) {
+                // This coverage statement was inlined from another function.
+                used_via_inlining.insert(inlined.def_id());
+            } else {
+                // Non-inlined coverage statements belong to the enclosing function.
+                saw_own_coverage = true;
+            }
+        }
+
+        if !saw_own_coverage && body.function_coverage_info.is_some() {
+            missing_own_coverage.insert(def_id);
+        }
+    }
+
+    UsageSets { all_mono_items, used_via_inlining, missing_own_coverage }
+}
+
+fn make_dummy_instance<'tcx>(tcx: TyCtxt<'tcx>, local_def_id: LocalDefId) -> ty::Instance<'tcx> {
+    let def_id = local_def_id.to_def_id();
+
+    // Make a dummy instance that fills in all generics with placeholders.
+    ty::Instance::new(
+        def_id,
+        ty::GenericArgs::for_item(tcx, def_id, |param, _| {
+            if let ty::GenericParamDefKind::Lifetime = param.kind {
+                tcx.lifetimes.re_erased.into()
+            } else {
+                tcx.mk_param_from_def(param)
+            }
+        }),
+    )
+}
diff --git a/compiler/rustc_codegen_llvm/src/debuginfo/create_scope_map.rs b/compiler/rustc_codegen_llvm/src/debuginfo/create_scope_map.rs
index f52991b3697..d2591139d6e 100644
--- a/compiler/rustc_codegen_llvm/src/debuginfo/create_scope_map.rs
+++ b/compiler/rustc_codegen_llvm/src/debuginfo/create_scope_map.rs
@@ -3,7 +3,6 @@ use std::collections::hash_map::Entry;
 use rustc_codegen_ssa::mir::debuginfo::{DebugScope, FunctionDebugContext};
 use rustc_codegen_ssa::traits::*;
 use rustc_data_structures::fx::FxHashMap;
-use rustc_index::Idx;
 use rustc_index::bit_set::DenseBitSet;
 use rustc_middle::mir::{Body, SourceScope};
 use rustc_middle::ty::layout::{FnAbiOf, HasTypingEnv};
@@ -43,8 +42,7 @@ pub(crate) fn compute_mir_scopes<'ll, 'tcx>(
     let mut instantiated = DenseBitSet::new_empty(mir.source_scopes.len());
     let mut discriminators = FxHashMap::default();
     // Instantiate all scopes.
-    for idx in 0..mir.source_scopes.len() {
-        let scope = SourceScope::new(idx);
+    for scope in mir.source_scopes.indices() {
         make_mir_scope(
             cx,
             instance,
diff --git a/compiler/rustc_codegen_llvm/src/debuginfo/metadata.rs b/compiler/rustc_codegen_llvm/src/debuginfo/metadata.rs
index 2eaaf127e41..7f3e486ca31 100644
--- a/compiler/rustc_codegen_llvm/src/debuginfo/metadata.rs
+++ b/compiler/rustc_codegen_llvm/src/debuginfo/metadata.rs
@@ -910,7 +910,8 @@ pub(crate) fn build_compile_unit_di_node<'ll, 'tcx>(
         && let Some(f) = output_filenames.split_dwarf_path(
             tcx.sess.split_debuginfo(),
             tcx.sess.opts.unstable_opts.split_dwarf_kind,
-            Some(codegen_unit_name),
+            codegen_unit_name,
+            tcx.sess.invocation_temp.as_deref(),
         ) {
         // We get a path relative to the working directory from split_dwarf_path
         Some(tcx.sess.source_map().path_mapping().to_real_filename(f))
@@ -1314,31 +1315,21 @@ fn build_generic_type_param_di_nodes<'ll, 'tcx>(
     ty: Ty<'tcx>,
 ) -> SmallVec<Option<&'ll DIType>> {
     if let ty::Adt(def, args) = *ty.kind() {
-        let generics = cx.tcx.generics_of(def.did());
-        return get_template_parameters(cx, generics, args);
-    }
-
-    return smallvec![];
-}
-
-pub(super) fn get_template_parameters<'ll, 'tcx>(
-    cx: &CodegenCx<'ll, 'tcx>,
-    generics: &ty::Generics,
-    args: ty::GenericArgsRef<'tcx>,
-) -> SmallVec<Option<&'ll DIType>> {
-    if args.types().next().is_some() {
-        let names = get_parameter_names(cx, generics);
-        let template_params: SmallVec<_> = iter::zip(args, names)
-            .filter_map(|(kind, name)| {
-                kind.as_type().map(|ty| {
-                    let actual_type = cx.tcx.normalize_erasing_regions(cx.typing_env(), ty);
-                    let actual_type_di_node = type_di_node(cx, actual_type);
-                    Some(cx.create_template_type_parameter(name.as_str(), actual_type_di_node))
+        if args.types().next().is_some() {
+            let generics = cx.tcx.generics_of(def.did());
+            let names = get_parameter_names(cx, generics);
+            let template_params: SmallVec<_> = iter::zip(args, names)
+                .filter_map(|(kind, name)| {
+                    kind.as_type().map(|ty| {
+                        let actual_type = cx.tcx.normalize_erasing_regions(cx.typing_env(), ty);
+                        let actual_type_di_node = type_di_node(cx, actual_type);
+                        Some(cx.create_template_type_parameter(name.as_str(), actual_type_di_node))
+                    })
                 })
-            })
-            .collect();
+                .collect();
 
-        return template_params;
+            return template_params;
+        }
     }
 
     return smallvec![];
diff --git a/compiler/rustc_codegen_llvm/src/debuginfo/metadata/enums/mod.rs b/compiler/rustc_codegen_llvm/src/debuginfo/metadata/enums/mod.rs
index 6792c307fdc..7c701926d2c 100644
--- a/compiler/rustc_codegen_llvm/src/debuginfo/metadata/enums/mod.rs
+++ b/compiler/rustc_codegen_llvm/src/debuginfo/metadata/enums/mod.rs
@@ -363,7 +363,6 @@ fn build_coroutine_variant_struct_type_di_node<'ll, 'tcx>(
 
             state_specific_fields.into_iter().chain(common_fields).collect()
         },
-        // FIXME: this is a no-op. `build_generic_type_param_di_nodes` only works for Adts.
         |cx| build_generic_type_param_di_nodes(cx, coroutine_type_and_layout.ty),
     )
     .di_node
diff --git a/compiler/rustc_codegen_llvm/src/debuginfo/metadata/type_map.rs b/compiler/rustc_codegen_llvm/src/debuginfo/metadata/type_map.rs
index ae2ab32ef53..56fb12d3c22 100644
--- a/compiler/rustc_codegen_llvm/src/debuginfo/metadata/type_map.rs
+++ b/compiler/rustc_codegen_llvm/src/debuginfo/metadata/type_map.rs
@@ -247,6 +247,16 @@ pub(super) fn stub<'ll, 'tcx>(
     StubInfo { metadata, unique_type_id }
 }
 
+struct AdtStackPopGuard<'ll, 'tcx, 'a> {
+    cx: &'a CodegenCx<'ll, 'tcx>,
+}
+
+impl<'ll, 'tcx, 'a> Drop for AdtStackPopGuard<'ll, 'tcx, 'a> {
+    fn drop(&mut self) {
+        debug_context(self.cx).adt_stack.borrow_mut().pop();
+    }
+}
+
 /// This function enables creating debuginfo nodes that can recursively refer to themselves.
 /// It will first insert the given stub into the type map and only then execute the `members`
 /// and `generics` closures passed in. These closures have access to the stub so they can
@@ -261,6 +271,44 @@ pub(super) fn build_type_with_children<'ll, 'tcx>(
 ) -> DINodeCreationResult<'ll> {
     assert_eq!(debug_context(cx).type_map.di_node_for_unique_id(stub_info.unique_type_id), None);
 
+    let mut _adt_stack_pop_guard = None;
+    if let UniqueTypeId::Ty(ty, ..) = stub_info.unique_type_id
+        && let ty::Adt(adt_def, args) = ty.kind()
+    {
+        let def_id = adt_def.did();
+        // If any sub type reference the original type definition and the sub type has a type
+        // parameter that strictly contains the original parameter, the original type is a recursive
+        // type that can expanding indefinitely. Example,
+        // ```
+        // enum Recursive<T> {
+        //     Recurse(*const Recursive<Wrap<T>>),
+        //     Item(T),
+        // }
+        // ```
+        let is_expanding_recursive =
+            debug_context(cx).adt_stack.borrow().iter().any(|(parent_def_id, parent_args)| {
+                if def_id == *parent_def_id {
+                    args.iter().zip(parent_args.iter()).any(|(arg, parent_arg)| {
+                        if let (Some(arg), Some(parent_arg)) = (arg.as_type(), parent_arg.as_type())
+                        {
+                            arg != parent_arg && arg.contains(parent_arg)
+                        } else {
+                            false
+                        }
+                    })
+                } else {
+                    false
+                }
+            });
+        if is_expanding_recursive {
+            // FIXME: indicate that this is an expanding recursive type in stub metadata?
+            return DINodeCreationResult::new(stub_info.metadata, false);
+        } else {
+            debug_context(cx).adt_stack.borrow_mut().push((def_id, args));
+            _adt_stack_pop_guard = Some(AdtStackPopGuard { cx });
+        }
+    }
+
     debug_context(cx).type_map.insert(stub_info.unique_type_id, stub_info.metadata);
 
     let members: SmallVec<_> =
diff --git a/compiler/rustc_codegen_llvm/src/debuginfo/mod.rs b/compiler/rustc_codegen_llvm/src/debuginfo/mod.rs
index ae7d080db66..c5085927923 100644
--- a/compiler/rustc_codegen_llvm/src/debuginfo/mod.rs
+++ b/compiler/rustc_codegen_llvm/src/debuginfo/mod.rs
@@ -2,8 +2,8 @@
 
 use std::cell::{OnceCell, RefCell};
 use std::ops::Range;
-use std::ptr;
 use std::sync::Arc;
+use std::{iter, ptr};
 
 use libc::c_uint;
 use metadata::create_subroutine_type;
@@ -66,6 +66,7 @@ pub(crate) struct CodegenUnitDebugContext<'ll, 'tcx> {
     created_files: RefCell<UnordMap<Option<(StableSourceFileId, SourceFileHash)>, &'ll DIFile>>,
 
     type_map: metadata::TypeMap<'ll, 'tcx>,
+    adt_stack: RefCell<Vec<(DefId, GenericArgsRef<'tcx>)>>,
     namespace_map: RefCell<DefIdMap<&'ll DIScope>>,
     recursion_marker_type: OnceCell<&'ll DIType>,
 }
@@ -80,6 +81,7 @@ impl<'ll, 'tcx> CodegenUnitDebugContext<'ll, 'tcx> {
             builder,
             created_files: Default::default(),
             type_map: Default::default(),
+            adt_stack: Default::default(),
             namespace_map: RefCell::new(Default::default()),
             recursion_marker_type: OnceCell::new(),
         }
@@ -486,10 +488,40 @@ impl<'ll, 'tcx> DebugInfoCodegenMethods<'tcx> for CodegenCx<'ll, 'tcx> {
             generics: &ty::Generics,
             args: GenericArgsRef<'tcx>,
         ) -> &'ll DIArray {
-            let template_params = metadata::get_template_parameters(cx, generics, args);
+            if args.types().next().is_none() {
+                return create_DIArray(DIB(cx), &[]);
+            }
+
+            // Again, only create type information if full debuginfo is enabled
+            let template_params: Vec<_> = if cx.sess().opts.debuginfo == DebugInfo::Full {
+                let names = get_parameter_names(cx, generics);
+                iter::zip(args, names)
+                    .filter_map(|(kind, name)| {
+                        kind.as_type().map(|ty| {
+                            let actual_type = cx.tcx.normalize_erasing_regions(cx.typing_env(), ty);
+                            let actual_type_metadata = type_di_node(cx, actual_type);
+                            Some(cx.create_template_type_parameter(
+                                name.as_str(),
+                                actual_type_metadata,
+                            ))
+                        })
+                    })
+                    .collect()
+            } else {
+                vec![]
+            };
+
             create_DIArray(DIB(cx), &template_params)
         }
 
+        fn get_parameter_names(cx: &CodegenCx<'_, '_>, generics: &ty::Generics) -> Vec<Symbol> {
+            let mut names = generics.parent.map_or_else(Vec::new, |def_id| {
+                get_parameter_names(cx, cx.tcx.generics_of(def_id))
+            });
+            names.extend(generics.own_params.iter().map(|param| param.name));
+            names
+        }
+
         /// Returns a scope, plus `true` if that's a type scope for "class" methods,
         /// otherwise `false` for plain namespace scopes.
         fn get_containing_scope<'ll, 'tcx>(
diff --git a/compiler/rustc_codegen_llvm/src/intrinsic.rs b/compiler/rustc_codegen_llvm/src/intrinsic.rs
index 67135fcc308..d1d6bcebd33 100644
--- a/compiler/rustc_codegen_llvm/src/intrinsic.rs
+++ b/compiler/rustc_codegen_llvm/src/intrinsic.rs
@@ -1421,7 +1421,7 @@ fn generic_simd_intrinsic<'ll, 'tcx>(
         return Ok(bx.shuffle_vector(args[0].immediate(), args[1].immediate(), indices));
     }
 
-    if name == sym::simd_insert {
+    if name == sym::simd_insert || name == sym::simd_insert_dyn {
         require!(
             in_elem == arg_tys[2],
             InvalidMonomorphization::InsertedType {
@@ -1432,40 +1432,49 @@ fn generic_simd_intrinsic<'ll, 'tcx>(
                 out_ty: arg_tys[2]
             }
         );
-        let idx = bx
-            .const_to_opt_u128(args[1].immediate(), false)
-            .expect("typeck should have ensure that this is a const");
-        if idx >= in_len.into() {
-            return_error!(InvalidMonomorphization::SimdIndexOutOfBounds {
-                span,
-                name,
-                arg_idx: 1,
-                total_len: in_len.into(),
-            });
-        }
-        return Ok(bx.insert_element(
-            args[0].immediate(),
-            args[2].immediate(),
-            bx.const_i32(idx as i32),
-        ));
+
+        let index_imm = if name == sym::simd_insert {
+            let idx = bx
+                .const_to_opt_u128(args[1].immediate(), false)
+                .expect("typeck should have ensure that this is a const");
+            if idx >= in_len.into() {
+                return_error!(InvalidMonomorphization::SimdIndexOutOfBounds {
+                    span,
+                    name,
+                    arg_idx: 1,
+                    total_len: in_len.into(),
+                });
+            }
+            bx.const_i32(idx as i32)
+        } else {
+            args[1].immediate()
+        };
+
+        return Ok(bx.insert_element(args[0].immediate(), args[2].immediate(), index_imm));
     }
-    if name == sym::simd_extract {
+    if name == sym::simd_extract || name == sym::simd_extract_dyn {
         require!(
             ret_ty == in_elem,
             InvalidMonomorphization::ReturnType { span, name, in_elem, in_ty, ret_ty }
         );
-        let idx = bx
-            .const_to_opt_u128(args[1].immediate(), false)
-            .expect("typeck should have ensure that this is a const");
-        if idx >= in_len.into() {
-            return_error!(InvalidMonomorphization::SimdIndexOutOfBounds {
-                span,
-                name,
-                arg_idx: 1,
-                total_len: in_len.into(),
-            });
-        }
-        return Ok(bx.extract_element(args[0].immediate(), bx.const_i32(idx as i32)));
+        let index_imm = if name == sym::simd_extract {
+            let idx = bx
+                .const_to_opt_u128(args[1].immediate(), false)
+                .expect("typeck should have ensure that this is a const");
+            if idx >= in_len.into() {
+                return_error!(InvalidMonomorphization::SimdIndexOutOfBounds {
+                    span,
+                    name,
+                    arg_idx: 1,
+                    total_len: in_len.into(),
+                });
+            }
+            bx.const_i32(idx as i32)
+        } else {
+            args[1].immediate()
+        };
+
+        return Ok(bx.extract_element(args[0].immediate(), index_imm));
     }
 
     if name == sym::simd_select {