[clang] [llvm] [clang][Modules][serialization] Deduplicate source location entries across loaded modules (PR #209795)

[Jan Svoboda via cfe-commits]

================
@@ -1815,6 +1815,97 @@ llvm::Error ASTReader::ReadSourceManagerBlock(ModuleFile &F) {
   }
 }
 
+const ASTReader::PrimaryLoadedFileLoc *
+ASTReader::getPrimaryLoadedFile(const SLocFileIdentity &Id) const {
+  assert(!Id.Name.empty() && "querying a non-file entry");
+  auto It = PrimaryLoadedFiles.find(Id.Name);
+  if (It == PrimaryLoadedFiles.end())
+    return nullptr;
+  const PrimaryLoadedFileLoc &Primary = It->second;
+  // A matching name but different size or time is a different file, e.g. two
+  // modules built against different versions of the same path. Don't merge.
+  if (Primary.Size != Id.Size || Primary.Time != Id.Time)
----------------
jansvoboda11 wrote:

The size+mtime of input files is already baked into Clang's PCMs by default, but can be (mostly?) disabled through command-line options I believe.

What I'd like to know is if we're okay with not deduplicating the input files due the names not matching exactly (i.e. capitalization, path separators, absolute/relative mismatch, benign path components differences, etc.).

https://github.com/llvm/llvm-project/pull/209795