fsmonitor: layer on another hack in bser.c for os.stat() compat (issue5811)

It's unclear to me how these `bserobj_tuple` objects are used, other
than as stat objects. This should fix fsmonitor in the wake of
ffa3026d4196 and similar changes. I regret the hack here, but the code
already has plenty of hg-specific hacks. :(

It feels like we should be able to use int(result.st_mtime) globally,
but that doesn't work. See issue4836 for a bug that was hard to track
down relating to rounding behavior causing very subtle dirstate
problems.

Differential Revision: https://phab.mercurial-scm.org/D2939

hgext/fsmonitor/pywatchman/bser.c

@@ -128,17 +128,27 b' static PyObject* bserobj_getattrro(PyObj'
   Py_ssize_t i, n;
   PyObject* name_bytes = NULL;
   PyObject* ret = NULL;
-  const char* namestr;
+  const char* namestr = NULL;
 
   if (PyIndex_Check(name)) {
     i = PyNumber_AsSsize_t(name, PyExc_IndexError);
     if (i == -1 && PyErr_Occurred()) {
       goto bail;
     }
+
+    if (i == 8 && PySequence_Size(obj->values) < 9) {
+      // Hack alert: Python 3 removed support for os.stat().st_mtime
+      // being an integer.Instead, if you need an integer, you have to
+      // use os.stat()[stat.ST_MTIME] instead. stat.ST_MTIME is 8, and
+      // our stat tuples are shorter than that, so we can detect
+      // requests for index 8 on tuples shorter than that and return
+      // st_mtime instead.
+      namestr = "st_mtime";
+    } else {
     ret = PySequence_GetItem(obj->values, i);
     goto bail;
   }
-
+  } else {
   // We can be passed in Unicode objects here -- we don't support anything other
   // than UTF-8 for keys.
   if (PyUnicode_Check(name)) {
@@ -150,6 +160,7 b' static PyObject* bserobj_getattrro(PyObj'
   } else {
     namestr = PyBytes_AsString(name);
   }
+  }
 
   if (namestr == NULL) {
     goto bail;