[llvm] Add support for verifying .debug_names in split DWARF for CUs and TUs. (PR #101775)

[David Blaikie via llvm-commits]

================
@@ -1626,8 +1626,56 @@ unsigned DWARFVerifier::verifyNameIndexEntries(
       UnitOffset = NI.getCUOffset(*CUIndex);
     if (!UnitOffset)
       continue;
-    uint64_t DIEOffset = *UnitOffset + *EntryOr->getDIEUnitOffset();
-    DWARFDie DIE = DCtx.getDIEForOffset(DIEOffset);
+    // For split DWARF entries we need to make sure we find the non skeleton
+    // DWARF unit that is needed and use that's DWARF unit offset as the
+    // DIE offset to add the DW_IDX_die_offset to.
+    DWARFUnit *DU = DCtx.getUnitForOffset(*UnitOffset);
+    if (DU == nullptr || DU->getOffset() != *UnitOffset) {
+      // If we didn't find a DWARF Unit from the UnitOffset, or if the offset
+      // of the unit doesn't match exactly, report an error.
+      ErrorCategory.Report(
+          "Name Index entry contains invalid CU or TU offset", [&]() {
+            error() << formatv("Name Index @ {0:x}: Entry @ {1:x} contains an "
+                               "invalid CU or TU offset {1:x}.\n",
+                               NI.getUnitOffset(), EntryID, *UnitOffset);
+          });
+      ++NumErrors;
+      continue;
+    }
+    // This function will try to get the non skeleton unit DIE, but if it is
+    // unable to load the .dwo file from the .dwo or .dwp, it will return the
+    // unit DIE of the DWARFUnit in "DU". So we need to check if the DWARFUnit
+    // has a .dwo file, but we couldn't load it.
+    DWARFDie NonSkeletonUnitDie = DU->getNonSkeletonUnitDIE();
+    if (DU->getDWOId() && DU->getUnitDIE() == NonSkeletonUnitDie) {
----------------
dwblaikie wrote:

> I believe it is intentional and expected behavior because as it if want all DIEs from a skeleton unit or a normal CU or TU, you call this API to try to get the DIEs regardless of if it is a skeleton unit or not.

The behavior of returning the non-split unit DIE makes sense given the name, but the behavior of returning the skeleton unit DIE if the split unit can't be found seems surprising/unintentional.

Returning the non-split unit DIE allows the "calling this function regardless of if it is a skeleton unit or not" work out, which seems good - but returning the skeleton unit DIE when the split unit isn't found seems like it'd be bad.

https://github.com/llvm/llvm-project/pull/101775