[llvm] [llvm-dwp] Fix infinite loop on DWARFv5 DW_FORM_implicit_const (PR #205567)

[David Blaikie via llvm-commits]

================
@@ -39,6 +39,23 @@ static uint64_t debugStrOffsetsHeaderSize(DataExtractor StrOffsetsData,
   return 8;    // unit length: 4 bytes, version: 2 bytes, padding: 2 bytes.
 }
 
+// Read the next (attribute, form) pair from an abbreviation declaration.
+// DW_FORM_implicit_const is the only form that stores an extra value -- an
+// SLEB128 -- in the abbreviation declaration itself; consume it here so the
+// (attribute, form) walk stays aligned with the rest of the table.
+// return True if can keep going- Name or Form is not 0
+static bool readAbbrevAttribute(const DataExtractor &AbbrevData,
+                                uint64_t *Offset, uint64_t &Name,
+                                dwarf::Form &Form) {
+  Name = AbbrevData.getULEB128(Offset);
+  Form = static_cast<dwarf::Form>(AbbrevData.getULEB128(Offset));
+  if (Form == dwarf::DW_FORM_implicit_const)
+    AbbrevData.getSLEB128(Offset);
+  if (Name != 0 || Form != 0)
----------------
dwblaikie wrote:

I'd prefer a switch - the loop should be exiting gracefully under the DWARF spec'd condition, of a 0 byte for the abbrev code. And then should have some handling for failures to read/corrupted input.

(& probably best not to mark conversations as "resolved" when they have active questions in them like this - it hides them (it's not "resolved" as in "I've done my part, I'm waiting for someone else to do their part" - github's "resolved" is more "everything's done here, no need to look" (which also means you have to have high confidence marking /someone else's/ concerns as "resolved" - you should probably leave it open for them to decide if you've adequately addressed their concern/resolved the issue))

https://github.com/llvm/llvm-project/pull/205567