[clang] [clang] Fix incorrect filenames for hardlinks (PR #189475)

[Adrian Prantl via cfe-commits]

================
@@ -542,14 +580,27 @@ FileID SourceManager::createFileID(FileEntryRef SourceFile,
                                    SourceLocation::UIntTy LoadedOffset) {
   SrcMgr::ContentCache &IR = getOrCreateContentCache(SourceFile,
                                                      isSystem(FileCharacter));
+  SrcMgr::ContentCache *Cache = &IR;
+  StringRef Filename = SourceFile.getName();
+
+  if (IR.OrigEntry && !IR.OrigEntry->isSameRef(SourceFile)) {
+    if (referencesDistinctFilePaths(getFileManager(), *IR.OrigEntry,
+                                    SourceFile)) {
+      Cache = cloneContentCache(ContentCacheAlloc, IR);
+      Cache->OrigEntry = SourceFile;
+      FileIDContentCaches.push_back(Cache);
+    } else {
+      Filename = IR.OrigEntry->getName();
----------------
adrian-prantl wrote:

AFAIK, we don't document LLVM's desired behavior when it comes to path normalization in debug info. The benefits of doing any kind of normalization are mostly about efficiency: it allows us to store paths only once, and stat them only once. DWARF specifically can store paths relative to the compilation directory, which saves space. Debug info clients often do simple prefix-based path remapping, which may either get easier or harder if you normalize paths depending on how you look at it. The safe option is often to leave paths alone.
`CGDebugInfo::createFile()` currently normalizes paths, which is code I added years ago while chasing after a more efficient encoding. At that time I didn't think about Windows paths nearly as much as I do now, and I am no longer convinced that normalization for saving a few bytes in the encoding is necessary the right trade off any more.

https://github.com/llvm/llvm-project/pull/189475