Fix the metadata deduplication

The llvmlite debug metadata cache does not work as expected due to
the encoding value using address of DIToken instance and serves as
part of the key.

Solution is to extract the string value from the encoding object.

Since the DIToken class is designed to be a wrapper of enumeration
value, e.g. the DW_* enumerations, that should appear bare in the
emitted metadata, in this change, stringify DIToken is made to fix
the problem.

This solved llvmlite issue #1164

Author: jiel-nv

llvmlite/ir/module.py

@@ -42,6 +42,15 @@ def _fix_di_operands(self, operands):
             fixed_ops.append((name, op))
         return fixed_ops
 
+    def str_ditok_operands(self, operands):
+        str_ops = []
+        for name, op in operands:
+            if name == 'encoding' and isinstance(op, values.DIToken):
+                # use string value instead of address of object
+                op = op.value
+            str_ops.append((name, op))
+        return str_ops
+
     def add_metadata(self, operands):
         """
         Add an unnamed metadata to the module with the given *operands*

llvmlite/ir/values.py

@@ -705,13 +705,6 @@ class DIToken:
 
     Use this to wrap known constants, e.g. the DW_* enumerations.
     """
-    _instance = None
-
-    def __new__(cls, value):
-        if not cls._instance or cls._instance.value != value:
-            cls._instance = super().__new__(cls)
-            cls._instance.value = value
-        return cls._instance
 
     def __init__(self, value):
         self.value = value