bpo-35193: Fix an off by one error in the RETURN_VALUE case. (GH-10418) (GH-10422)

Fix an off by one error in the peephole optimizer when checking for unreachable code beyond a return.

Do a bounds check within find_op so it can return before going past the end as a safety measure.

https://github.com/python/cpython/commit/7db3c488335168993689ddae5914a28e16188447#diff-a33329ae6ae0bb295d742f0caf93c137
introduced this off by one error while fixing another one nearby.

This bug was shipped in all Python 3.6 and 3.7 releases.

The included unittest won't fail unless you do a clang msan build.

(cherry picked from commit 49fa4a9f1ef387e16596f271414c855339eadf09)

diff --git a/Lib/test/test_compile.py b/Lib/test/test_compile.py
index 745ce029aa575cffc75549f10e42e4cd93306fd5..13cc8821b082a95e9acd712af73a1d0b20c8c0cd 100644 (file)
--- a/Lib/test/test_compile.py
+++ b/Lib/test/test_compile.py
@@ -1,3 +1,4 @@
+import dis
 import math
 import os
 import unittest
@@ -628,6 +629,24 @@ if 1:
         self.check_constant(f1, frozenset({0}))
         self.assertTrue(f1(0))
 
+    # This is a regression test for a CPython specific peephole optimizer
+    # implementation bug present in a few releases.  It's assertion verifies
+    # that peephole optimization was actually done though that isn't an
+    # indication of the bugs presence or not (crashing is).
+    @support.cpython_only
+    def test_peephole_opt_unreachable_code_array_access_in_bounds(self):
+        """Regression test for issue35193 when run under clang msan."""
+        def unused_code_at_end():
+            return 3
+            raise RuntimeError("unreachable")
+        # The above function definition will trigger the out of bounds
+        # bug in the peephole optimizer as it scans opcodes past the
+        # RETURN_VALUE opcode.  This does not always crash an interpreter.
+        # When you build with the clang memory sanitizer it reliably aborts.
+        self.assertEqual(
+            'RETURN_VALUE',
+            list(dis.get_instructions(unused_code_at_end))[-1].opname)
+
     def test_dont_merge_constants(self):
         # Issue #25843: compile() must not merge constants which are equal
         # but have a different type.

diff --git a/Misc/NEWS.d/next/Core and Builtins/2018-11-08-23-00-04.bpo-35193.WK2PDg.rst b/Misc/NEWS.d/next/Core and Builtins/2018-11-08-23-00-04.bpo-35193.WK2PDg.rst
new file mode 100644 (file)
index 0000000..a6b3c64
--- /dev/null
+++ b/Misc/NEWS.d/next/Core and Builtins/2018-11-08-23-00-04.bpo-35193.WK2PDg.rst
@@ -0,0 +1,3 @@
+Fix an off by one error in the bytecode peephole optimizer where it could
+read bytes beyond the end of bounds of an array when removing unreachable
+code. This bug was present in every release of Python 3.6 until now.

diff --git a/Python/peephole.c b/Python/peephole.c
index 31d4e92cfd33431a1a3025c5aecac5f749161621..3fa3b7fcee62843667c4d627eef02735320fae03 100644 (file)
--- a/Python/peephole.c
+++ b/Python/peephole.c
@@ -96,9 +96,9 @@ lastn_const_start(const _Py_CODEUNIT *codestr, Py_ssize_t i, Py_ssize_t n)
 
 /* Scans through EXTENDED ARGs, seeking the index of the effective opcode */
 static Py_ssize_t
-find_op(const _Py_CODEUNIT *codestr, Py_ssize_t i)
+find_op(const _Py_CODEUNIT *codestr, Py_ssize_t codelen, Py_ssize_t i)
 {
-    while (_Py_OPCODE(codestr[i]) == EXTENDED_ARG) {
+    while (i < codelen && _Py_OPCODE(codestr[i]) == EXTENDED_ARG) {
         i++;
     }
     return i;
@@ -590,7 +590,7 @@ PyCode_Optimize(PyObject *code, PyObject* consts, PyObject *names,
 
     CONST_STACK_CREATE();
 
-    for (i=find_op(codestr, 0) ; i<codelen ; i=nexti) {
+    for (i=find_op(codestr, codelen, 0) ; i<codelen ; i=nexti) {
         opcode = _Py_OPCODE(codestr[i]);
         op_start = i;
         while (op_start >= 1 && _Py_OPCODE(codestr[op_start-1]) == EXTENDED_ARG) {
@@ -755,7 +755,7 @@ PyCode_Optimize(PyObject *code, PyObject* consts, PyObject *names,
             case JUMP_IF_FALSE_OR_POP:
             case JUMP_IF_TRUE_OR_POP:
                 h = get_arg(codestr, i) / sizeof(_Py_CODEUNIT);
-                tgt = find_op(codestr, h);
+                tgt = find_op(codestr, codelen, h);
 
                 j = _Py_OPCODE(codestr[tgt]);
                 if (CONDITIONAL_JUMP(j)) {
@@ -796,7 +796,7 @@ PyCode_Optimize(PyObject *code, PyObject* consts, PyObject *names,
             case SETUP_WITH:
             case SETUP_ASYNC_WITH:
                 h = GETJUMPTGT(codestr, i);
-                tgt = find_op(codestr, h);
+                tgt = find_op(codestr, codelen, h);
                 /* Replace JUMP_* to a RETURN into just a RETURN */
                 if (UNCONDITIONAL_JUMP(opcode) &&
                     _Py_OPCODE(codestr[tgt]) == RETURN_VALUE) {
@@ -825,7 +825,7 @@ PyCode_Optimize(PyObject *code, PyObject* consts, PyObject *names,
                 }
                 if (h > i + 1) {
                     fill_nops(codestr, i + 1, h);
-                    nexti = find_op(codestr, h);
+                    nexti = find_op(codestr, codelen, h);
                 }
                 break;
         }