[llvm] [DebugInfo] Make DISubprogram's hashing always produce the same result (PR #90770)

[Adrian Prantl via llvm-commits]

================
@@ -825,19 +825,26 @@ template <> struct MDNodeKeyImpl<DISubprogram> {
   bool isDefinition() const { return SPFlags & DISubprogram::SPFlagDefinition; }
 
   unsigned getHashValue() const {
+    // Use the Scope's linkage name instead of using the scope directly, as the
+    // scope may be a temporary one which can replaced, which would produce a
+    // different hash for the same DISubprogram.
+    llvm::StringRef ScopeLinkageName;
+    if (auto *CT = dyn_cast_or_null<DICompositeType>(Scope))
+      if (auto *ID = CT->getRawIdentifier())
+        ScopeLinkageName = ID->getString();
+
     // If this is a declaration inside an ODR type, only hash the type and the
     // name.  Otherwise the hash will be stronger than
     // MDNodeSubsetEqualImpl::isDeclarationOfODRMember().
-    if (!isDefinition() && LinkageName)
-      if (auto *CT = dyn_cast_or_null<DICompositeType>(Scope))
-        if (CT->getRawIdentifier())
-          return hash_combine(LinkageName, Scope);
+    if (!isDefinition() && LinkageName &&
+        isa_and_nonnull<DICompositeType>(Scope))
+      return hash_combine(LinkageName, ScopeLinkageName);
 
     // Intentionally computes the hash on a subset of the operands for
     // performance reason. The subset has to be significant enough to avoid
     // collision "most of the time". There is no correctness issue in case of
     // collision because of the full check above.
-    return hash_combine(Name, Scope, File, Type, Line);
+    return hash_combine(Name, ScopeLinkageName, File, Type, Line);
----------------
adrian-prantl wrote:

Also the File/Line should provide enough uniqueness.

https://github.com/llvm/llvm-project/pull/90770