gh-93065: Fix HAMT to iterate correctly over 7-level deep trees (GH-93066)

Also while there, clarify a few things about why we reduce the hash to 32 bits.

Co-authored-by: Eli Libman <eli@hyro.ai>
Co-authored-by: Yury Selivanov <yury@edgedb.com>
Co-authored-by: Łukasz Langa <lukasz@langa.pl>

Include/internal/pycore_hamt.h

@@ -5,7 +5,19 @@
 #  error "this header requires Py_BUILD_CORE define"
 #endif
 
-#define _Py_HAMT_MAX_TREE_DEPTH 7
+
+/*
+HAMT tree is shaped by hashes of keys. Every group of 5 bits of a hash denotes
+the exact position of the key in one level of the tree. Since we're using
+32 bit hashes, we can have at most 7 such levels. Although if there are
+two distinct keys with equal hashes, they will have to occupy the same
+cell in the 7th level of the tree -- so we'd put them in a "collision" node.
+Which brings the total possible tree depth to 8. Read more about the actual
+layout of the HAMT tree in `hamt.c`.
+
+This constant is used to define a datastucture for storing iteration state.
+*/
+#define _Py_HAMT_MAX_TREE_DEPTH 8
 
 
 extern PyTypeObject _PyHamt_Type;

Lib/test/test_context.py

@@ -535,6 +535,41 @@ class HamtTest(unittest.TestCase):
         self.assertEqual(len(h4), 2)
         self.assertEqual(len(h5), 3)
 
+    def test_hamt_collision_3(self):
+        # Test that iteration works with the deepest tree possible.
+        # https://github.com/python/cpython/issues/93065
+
+        C = HashKey(0b10000000_00000000_00000000_00000000, 'C')
+        D = HashKey(0b10000000_00000000_00000000_00000000, 'D')
+
+        E = HashKey(0b00000000_00000000_00000000_00000000, 'E')
+
+        h = hamt()
+        h = h.set(C, 'C')
+        h = h.set(D, 'D')
+        h = h.set(E, 'E')
+
+        # BitmapNode(size=2 count=1 bitmap=0b1):
+        #   NULL:
+        #     BitmapNode(size=2 count=1 bitmap=0b1):
+        #       NULL:
+        #         BitmapNode(size=2 count=1 bitmap=0b1):
+        #           NULL:
+        #             BitmapNode(size=2 count=1 bitmap=0b1):
+        #               NULL:
+        #                 BitmapNode(size=2 count=1 bitmap=0b1):
+        #                   NULL:
+        #                     BitmapNode(size=2 count=1 bitmap=0b1):
+        #                       NULL:
+        #                         BitmapNode(size=4 count=2 bitmap=0b101):
+        #                           <Key name:E hash:0>: 'E'
+        #                           NULL:
+        #                             CollisionNode(size=4 id=0x107a24520):
+        #                               <Key name:C hash:2147483648>: 'C'
+        #                               <Key name:D hash:2147483648>: 'D'
+
+        self.assertEqual({k.name for k in h.keys()}, {'C', 'D', 'E'})
+
     def test_hamt_stress(self):
         COLLECTION_SIZE = 7000
         TEST_ITERS_EVERY = 647

Misc/ACKS

@@ -1062,6 +1062,7 @@ Robert Li
 Xuanji Li
 Zekun Li
 Zheao Li
+Eli Libman
 Dan Lidral-Porter
 Robert van Liere
 Ross Light

Misc/NEWS.d/next/Core and Builtins/2022-05-21-23-21-37.gh-issue-93065.5I18WC.rst

@@ -0,0 +1,5 @@
+Fix contextvars HAMT implementation to handle iteration over deep trees.
+
+The bug was discovered and fixed by Eli Libman. See
+`MagicStack/immutables#84 <https://github.com/MagicStack/immutables/issues/84>`_
+for more details.

Python/hamt.c

@@ -409,14 +409,22 @@ hamt_hash(PyObject *o)
         return -1;
     }
 
-    /* While it's suboptimal to reduce Python's 64 bit hash to
+    /* While it's somewhat suboptimal to reduce Python's 64 bit hash to
        32 bits via XOR, it seems that the resulting hash function
        is good enough (this is also how Long type is hashed in Java.)
        Storing 10, 100, 1000 Python strings results in a relatively
        shallow and uniform tree structure.
 
-       Please don't change this hashing algorithm, as there are many
-       tests that test some exact tree shape to cover all code paths.
+       Also it's worth noting that it would be possible to adapt the tree
+       structure to 64 bit hashes, but that would increase memory pressure
+       and provide little to no performance benefits for collections with
+       fewer than billions of key/value pairs.
+
+       Important: do not change this hash reducing function. There are many
+       tests that need an exact tree shape to cover all code paths and
+       we do that by specifying concrete values for test data's `__hash__`.
+       If this function is changed most of the regression tests would
+       become useless.
     */
     int32_t xored = (int32_t)(hash & 0xffffffffl) ^ (int32_t)(hash >> 32);
     return xored == -1 ? -2 : xored;