coverage: Deal with unused functions and their names in one place

Author: Zalathar

compiler/rustc_codegen_llvm/src/coverageinfo/mapgen.rs

@@ -73,12 +73,11 @@ pub(crate) fn finalize(cx: &CodegenCx<'_, '_>) {
     // 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 = unused::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 global_file_table,
+            &mut covfun_records,
+        );
     }
 
     // If there are no covfun records for this CGU, don't generate a covmap record.
@@ -97,33 +96,10 @@ pub(crate) fn finalize(cx: &CodegenCx<'_, '_>) {
     // contain multiple covmap records from different compilation units.
     let filenames_hash = llvm_cov::hash_bytes(&filenames_buffer);
 
-    let mut unused_function_names = vec![];
-
     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);
-    }
-
     // 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.)

compiler/rustc_codegen_llvm/src/coverageinfo/mapgen/covfun.rs

@@ -37,14 +37,6 @@ pub(crate) struct CovfunRecord<'tcx> {
     regions: ffi::Regions,
 }
 
-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)
-    }
-}
-
 pub(crate) fn prepare_covfun_record<'tcx>(
     tcx: TyCtxt<'tcx>,
     global_file_table: &mut GlobalFileTable,

compiler/rustc_codegen_llvm/src/coverageinfo/mapgen/unused.rs

@@ -1,3 +1,4 @@
+use rustc_codegen_ssa::traits::{BaseTypeCodegenMethods, ConstCodegenMethods};
 use rustc_data_structures::fx::FxHashSet;
 use rustc_hir::def_id::{DefId, LocalDefId};
 use rustc_middle::mir;
@@ -6,6 +7,9 @@ use rustc_middle::ty::{self, TyCtxt};
 use rustc_span::def_id::DefIdSet;
 
 use crate::common::CodegenCx;
+use crate::coverageinfo::mapgen::GlobalFileTable;
+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
@@ -15,9 +19,48 @@ use crate::common::CodegenCx;
 /// 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 gather_unused_function_instances<'tcx>(
+pub(crate) fn prepare_covfun_records_for_unused_functions<'tcx>(
     cx: &CodegenCx<'_, 'tcx>,
-) -> Vec<ty::Instance<'tcx>> {
+    global_file_table: &mut GlobalFileTable,
+    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, global_file_table, 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;
@@ -45,6 +88,7 @@ pub(crate) fn gather_unused_function_instances<'tcx>(
         .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<_>>()
 }