[clang] [llvm] [LTO] Reduce memory usage for import lists (PR #106772)

[Kazu Hirata via cfe-commits]

================
@@ -215,13 +215,49 @@ class FunctionImporter {
     SmallVector<StringRef, 0> getSourceModules() const;
 
     std::optional<GlobalValueSummary::ImportKind>
-    getImportType(const FunctionsToImportTy &GUIDToImportType,
-                  GlobalValue::GUID GUID) const;
+    getImportType(StringRef FromModule, GlobalValue::GUID GUID) const;
+
+    // Iterate over the import list.  The caller gets tuples of FromModule,
+    // GUID, and ImportKind instead of import IDs.
+    auto begin() const { return map_iterator(Imports.begin(), IDs); }
+    auto end() const { return map_iterator(Imports.end(), IDs); }
+
+    friend class SortedImportList;
+
+  private:
+    ImportIDTable &IDs;
+    DenseSet<ImportIDTable::ImportIDTy> Imports;
+  };
+
+  // A read-only copy of ImportMapTy with its contents sorted according to the
+  // given comparison function.
+  class SortedImportList {
+  public:
+    SortedImportList(const ImportMapTy &ImportMap,
+                     llvm::function_ref<
+                         bool(const std::pair<StringRef, GlobalValue::GUID> &,
+                              const std::pair<StringRef, GlobalValue::GUID> &)>
+                         Comp)
+        : IDs(ImportMap.IDs), Imports(iterator_range(ImportMap.Imports)) {
+      llvm::sort(Imports, [&](ImportIDTable::ImportIDTy L,
----------------
kazutakahirata wrote:

May I ask why?  `llvm::stable_sort` implies that the original order matters in some way, but the original order here is the iteration order of `DenseSet`.  There is nothing worth preserving there.

https://github.com/llvm/llvm-project/pull/106772