[inductor] Restore ExpandView sanity checks (pytorch#127251)

This restores the assertion removed in pytorch#124864

The handling of unbacked symints is incidental, the main purpose of this assert
was to catch bugs in lowerings.

Pull Request resolved: pytorch#127251
Approved by: https://github.com/lezcano

torch/_inductor/ir.py

@@ -1937,6 +1937,7 @@ class ExpandView(BaseView):
     @staticmethod
     def _normalize_size(x, new_size):
         """Replace `-1` with correct sizes"""
+        sizevars = V.graph.sizevars
         new_size = list(map(sympy.expand, new_size))
         old_size = x.get_size()
         old_size = [None] * (len(new_size) - len(old_size)) + list(old_size)
@@ -1948,12 +1949,14 @@ def _normalize_size(x, new_size):
             elif old_size[i] is None or old_size[i] == 1:
                 pass
             else:
-                # NB: new_size[i] == old_size[i] is known because the meta
-                # formula was expected to have taught us this equality.
-                # We can't conveniently check it right now because
-                # statically_known_equals doesn't know to consult preexisting
-                # guards
-                pass
+                # Sanity check: Expect broadcast compatibility
+                #
+                # NB: new_size[i] == old_size[i] is expected to already be
+                # guarded because the meta formula was expected to have taught
+                # us this equality.
+                assert (
+                    sizevars.size_hint(new_size[i] - old_size[i], fallback=0) == 0
+                ), "Broadcast failed in ExpandView({x.get_size()}, {new_size}) on dimension {i}"
         return new_size
 
     @classmethod