PEP: 280 Title: Optimizing access to globals Author: Guido van Rossum
<guido@python.org> Status: Deferred Type: Standards Track Content-Type:
text/x-rst Created: 10-Feb-2002 Python-Version: 2.3 Post-History:

Deferral

While this PEP is a nice idea, no-one has yet emerged to do the work of
hashing out the differences between this PEP, PEP 266 and PEP 267.
Hence, it is being deferred.

Abstract

This PEP describes yet another approach to optimizing access to module
globals, providing an alternative to PEP 266 (Optimizing Global
Variable/Attribute Access by Skip Montanaro) and PEP 267 (Optimized
Access to Module Namespaces by Jeremy Hylton).

The expectation is that eventually one approach will be picked and
implemented; possibly multiple approaches will be prototyped first.

Description

(Note: Jason Orendorff writes: """I implemented this once, long ago, for
Python 1.5-ish, I believe. I got it to the point where it was only 15%
slower than ordinary Python, then abandoned it. ;) In my implementation,
"cells" were real first-class objects, and "celldict" was a
copy-and-hack version of dictionary. I forget how the rest worked."""
Reference:
https://mail.python.org/pipermail/python-dev/2002-February/019876.html)

Let a cell be a really simple Python object, containing a pointer to a
Python object and a pointer to a cell. Both pointers may be NULL. A
Python implementation could be:

    class cell(object):

        def __init__(self):
            self.objptr = NULL
            self.cellptr = NULL

The cellptr attribute is used for chaining cells together for searching
built-ins; this will be explained later.

Let a celldict be a mapping from strings (the names of a module's
globals) to objects (the values of those globals), implemented using a
dict of cells. A Python implementation could be:

    class celldict(object):

        def __init__(self):
            self.__dict = {} # dict of cells

        def getcell(self, key):
            c = self.__dict.get(key)
            if c is None:
                c = cell()
                self.__dict[key] = c
            return c

        def cellkeys(self):
            return self.__dict.keys()

        def __getitem__(self, key):
            c = self.__dict.get(key)
            if c is None:
                raise KeyError, key
            value = c.objptr
            if value is NULL:
                raise KeyError, key
            else:
                return value

        def __setitem__(self, key, value):
            c = self.__dict.get(key)
            if c is None:
                c = cell()
                self.__dict[key] = c
            c.objptr = value

        def __delitem__(self, key):
            c = self.__dict.get(key)
            if c is None or c.objptr is NULL:
                raise KeyError, key
            c.objptr = NULL

        def keys(self):
            return [k for k, c in self.__dict.iteritems()
                    if c.objptr is not NULL]

        def items(self):
            return [k, c.objptr for k, c in self.__dict.iteritems()
                    if c.objptr is not NULL]

        def values(self):
            preturn [c.objptr for c in self.__dict.itervalues()
                    if c.objptr is not NULL]

        def clear(self):
            for c in self.__dict.values():
                c.objptr = NULL

        # Etc.

It is possible that a cell exists corresponding to a given key, but the
cell's objptr is NULL; let's call such a cell empty. When the celldict
is used as a mapping, it is as if empty cells don't exist. However, once
added, a cell is never deleted from a celldict, and it is possible to
get at empty cells using the getcell() method.

The celldict implementation never uses the cellptr attribute of cells.

We change the module implementation to use a celldict for its __dict__.
The module's getattr, setattr and delattr operations now map to getitem,
setitem and delitem on the celldict. The type of <module>.__dict__ and
globals() is probably the only backwards incompatibility.

When a module is initialized, its __builtins__ is initialized from the
__builtin__ module's __dict__, which is itself a celldict. For each cell
in __builtins__, the new module's __dict__ adds a cell with a NULL
objptr, whose cellptr points to the corresponding cell of __builtins__.
Python pseudo-code (ignoring rexec):

    import __builtin__

    class module(object):

        def __init__(self):
            self.__dict__ = d = celldict()
            d['__builtins__'] = bd = __builtin__.__dict__
            for k in bd.cellkeys():
                c = self.__dict__.getcell(k)
                c.cellptr = bd.getcell(k)

        def __getattr__(self, k):
            try:
                return self.__dict__[k]
            except KeyError:
                raise IndexError, k

        def __setattr__(self, k, v):
            self.__dict__[k] = v

        def __delattr__(self, k):
            del self.__dict__[k]

The compiler generates LOAD_GLOBAL_CELL <i> (and STORE_GLOBAL_CELL <i>
etc.) opcodes for references to globals, where <i> is a small index with
meaning only within one code object like the const index in LOAD_CONST.
The code object has a new tuple, co_globals, giving the names of the
globals referenced by the code indexed by <i>. No new analysis is
required to be able to do this.

When a function object is created from a code object and a celldict, the
function object creates an array of cell pointers by asking the celldict
for cells corresponding to the names in the code object's co_globals. If
the celldict doesn't already have a cell for a particular name, it
creates and an empty one. This array of cell pointers is stored on the
function object as func_cells. When a function object is created from a
regular dict instead of a celldict, func_cells is a NULL pointer.

When the VM executes a LOAD_GLOBAL_CELL <i> instruction, it gets cell
number <i> from func_cells. It then looks in the cell's PyObject
pointer, and if not NULL, that's the global value. If it is NULL, it
follows the cell's cell pointer to the next cell, if it is not NULL, and
looks in the PyObject pointer in that cell. If that's also NULL, or if
there is no second cell, NameError is raised. (It could follow the chain
of cell pointers until a NULL cell pointer is found; but I have no use
for this.) Similar for STORE_GLOBAL_CELL <i>, except it doesn't follow
the cell pointer chain -- it always stores in the first cell.

There are fallbacks in the VM for the case where the function's globals
aren't a celldict, and hence func_cells is NULL. In that case, the code
object's co_globals is indexed with <i> to find the name of the
corresponding global and this name is used to index the function's
globals dict.

Additional Ideas

-   Never make func_cell a NULL pointer; instead, make up an array of
    empty cells, so that LOAD_GLOBAL_CELL can index func_cells without a
    NULL check.

-   Make c.cellptr equal to c when a cell is created, so that
    LOAD_GLOBAL_CELL can always dereference c.cellptr without a NULL
    check.

    With these two additional ideas added, here's Python pseudo-code for
    LOAD_GLOBAL_CELL:

        def LOAD_GLOBAL_CELL(self, i):
            # self is the frame
            c = self.func_cells[i]
            obj = c.objptr
            if obj is not NULL:
                return obj # Existing global
            return c.cellptr.objptr # Built-in or NULL

-   Be more aggressive: put the actual values of builtins into module
    dicts, not just pointers to cells containing the actual values.

    There are two points to this: (1) Simplify and speed access, which
    is the most common operation. (2) Support faithful emulation of
    extreme existing corner cases.

    WRT #2, the set of builtins in the scheme above is captured at the
    time a module dict is first created. Mutations to the set of builtin
    names following that don't get reflected in the module dicts.
    Example: consider files main.py and cheater.py:

        [main.py]
        import cheater
        def f():
            cheater.cheat()
            return pachinko()
        print f()

        [cheater.py]
        def cheat():
            import __builtin__
            __builtin__.pachinko = lambda: 666

    If main.py is run under Python 2.2 (or before), 666 is printed. But
    under the proposal, __builtin__.pachinko doesn't exist at the time
    main's __dict__ is initialized. When the function object for f is
    created, main.__dict__ grows a pachinko cell mapping to two NULLs.
    When cheat() is called, __builtin__.__dict__ grows a pachinko cell
    too, but main.__dict__ doesn't know-- and will never know --about
    that. When f's return stmt references pachinko, in will still find
    the double-NULLs in main.__dict__'s pachinko cell, and so raise
    NameError.

    A similar (in cause) break in compatibility can occur if a module
    global foo is del'ed, but a builtin foo was created prior to that
    but after the module dict was first created. Then the builtin foo
    becomes visible in the module under 2.2 and before, but remains
    invisible under the proposal.

    Mutating builtins is extremely rare (most programs never mutate the
    builtins, and it's hard to imagine a plausible use for frequent
    mutation of the builtins -- I've never seen or heard of one), so it
    doesn't matter how expensive mutating the builtins becomes. OTOH,
    referencing globals and builtins is very common. Combining those
    observations suggests a more aggressive caching of builtins in
    module globals, speeding access at the expense of making mutations
    of the builtins (potentially much) more expensive to keep the caches
    in synch.

    Much of the scheme above remains the same, and most of the rest is
    just a little different. A cell changes to:

        class cell(object):
            def __init__(self, obj=NULL, builtin=0):
                self.objptr = obj
                self.builtinflag = builtin

    and a celldict maps strings to this version of cells. builtinflag is
    true when and only when objptr contains a value obtained from the
    builtins; in other words, it's true when and only when a cell is
    acting as a cached value. When builtinflag is false, objptr is the
    value of a module global (possibly NULL). celldict changes to:

        class celldict(object):

            def __init__(self, builtindict=()):
                self.basedict = builtindict
                self.__dict = d = {}
                for k, v in builtindict.items():
                    d[k] = cell(v, 1)

            def __getitem__(self, key):
                c = self.__dict.get(key)
                if c is None or c.objptr is NULL or c.builtinflag:
                    raise KeyError, key
                return c.objptr

            def __setitem__(self, key, value):
                c = self.__dict.get(key)
                if c is None:
                    c = cell()
                    self.__dict[key] = c
                c.objptr = value
                c.builtinflag = 0

            def __delitem__(self, key):
                c = self.__dict.get(key)
                if c is None or c.objptr is NULL or c.builtinflag:
                    raise KeyError, key
                c.objptr = NULL
                # We may have unmasked a builtin.  Note that because
                # we're checking the builtin dict for that *now*, this
                # still works if the builtin first came into existence
                # after we were constructed.  Note too that del on
                # namespace dicts is rare, so the expense of this check
                # shouldn't matter.
                if key in self.basedict:
                    c.objptr = self.basedict[key]
                    assert c.objptr is not NULL # else "in" lied
                    c.builtinflag = 1
                else:
                    # There is no builtin with the same name.
                    assert not c.builtinflag

            def keys(self):
                return [k for k, c in self.__dict.iteritems()
                        if c.objptr is not NULL and not c.builtinflag]

            def items(self):
                return [k, c.objptr for k, c in self.__dict.iteritems()
                        if c.objptr is not NULL and not c.builtinflag]

            def values(self):
                preturn [c.objptr for c in self.__dict.itervalues()
                        if c.objptr is not NULL and not c.builtinflag]

            def clear(self):
                for c in self.__dict.values():
                    if not c.builtinflag:
                        c.objptr = NULL

            # Etc.

    The speed benefit comes from simplifying LOAD_GLOBAL_CELL, which I
    expect is executed more frequently than all other namespace
    operations combined:

        def LOAD_GLOBAL_CELL(self, i):
            # self is the frame
            c = self.func_cells[i]
            return c.objptr   # may be NULL (also true before)

    That is, accessing builtins and accessing module globals are equally
    fast. For module globals, a NULL-pointer test+branch is saved. For
    builtins, an additional pointer chase is also saved.

    The other part needed to make this fly is expensive, propagating
    mutations of builtins into the module dicts that were initialized
    from the builtins. This is much like, in 2.2, propagating changes in
    new-style base classes to their descendants: the builtins need to
    maintain a list of weakrefs to the modules (or module dicts)
    initialized from the builtin's dict. Given a mutation to the builtin
    dict (adding a new key, changing the value associated with an
    existing key, or deleting a key), traverse the list of module dicts
    and make corresponding mutations to them. This is straightforward;
    for example, if a key is deleted from builtins, execute
    reflect_bltin_del in each module:

        def reflect_bltin_del(self, key):
            c = self.__dict.get(key)
            assert c is not None # else we were already out of synch
            if c.builtinflag:
                # Put us back in synch.
                c.objptr = NULL
                c.builtinflag = 0
            # Else we're shadowing the builtin, so don't care that
            # the builtin went away.

    Note that c.builtinflag protects from us erroneously deleting a
    module global of the same name. Adding a new (key, value) builtin
    pair is similar:

        def reflect_bltin_new(self, key, value):
            c = self.__dict.get(key)
            if c is None:
                # Never heard of it before:  cache the builtin value.
                self.__dict[key] = cell(value, 1)
            elif c.objptr is NULL:
                # This used to exist in the module or the builtins,
                # but doesn't anymore; rehabilitate it.
                assert not c.builtinflag
                c.objptr = value
                c.builtinflag = 1
            else:
                # We're shadowing it already.
                assert not c.builtinflag

    Changing the value of an existing builtin:

        def reflect_bltin_change(self, key, newvalue):
            c = self.__dict.get(key)
            assert c is not None # else we were already out of synch
            if c.builtinflag:
                # Put us back in synch.
                c.objptr = newvalue
            # Else we're shadowing the builtin, so don't care that
            # the builtin changed.

FAQs

-   Q: Will it still be possible to:

    a) install new builtins in the __builtin__ namespace and have them
    available in all already loaded modules right away ?

    b) override builtins (e.g. open()) with my own copies (e.g. to
    increase security) in a way that makes these new copies override the
    previous ones in all modules ?

    A: Yes, this is the whole point of this design. In the original
    approach, when LOAD_GLOBAL_CELL finds a NULL in the second cell, it
    should go back to see if the __builtins__ dict has been modified
    (the pseudo code doesn't have this yet). Tim's "more aggressive"
    alternative also takes care of this.

-   Q: How does the new scheme get along with the restricted execution
    model?

    A: It is intended to support that fully.

-   Q: What happens when a global is deleted?

    A: The module's celldict would have a cell with a NULL objptr for
    that key. This is true in both variations, but the "aggressive"
    variation goes on to see whether this unmasks a builtin of the same
    name, and if so copies its value (just a pointer-copy of the
    ultimate PyObject*) into the cell's objptr and sets the cell's
    builtinflag to true.

-   Q: What would the C code for LOAD_GLOBAL_CELL look like?

    A: The first version, with the first two bullets under "Additional
    ideas" incorporated, could look like this:

        case LOAD_GLOBAL_CELL:
            cell = func_cells[oparg];
            x = cell->objptr;
            if (x == NULL) {
                x = cell->cellptr->objptr;
                if (x == NULL) {
                    ... error recovery ...
                    break;
                }
            }
            Py_INCREF(x);
            PUSH(x);
            continue;

    We could even write it like this (idea courtesy of Ka-Ping Yee):

        case LOAD_GLOBAL_CELL:
            cell = func_cells[oparg];
            x = cell->cellptr->objptr;
            if (x != NULL) {
                Py_INCREF(x);
                PUSH(x);
                continue;
            }
            ... error recovery ...
            break;

    In modern CPU architectures, this reduces the number of branches
    taken for built-ins, which might be a really good thing, while any
    decent memory cache should realize that cell->cellptr is the same as
    cell for regular globals and hence this should be very fast in that
    case too.

    For the aggressive variant:

        case LOAD_GLOBAL_CELL:
            cell = func_cells[oparg];
            x = cell->objptr;
            if (x != NULL) {
                Py_INCREF(x);
                PUSH(x);
                continue;
            }
            ... error recovery ...
            break;

-   Q: What happens in the module's top-level code where there is
    presumably no func_cells array?

    A: We could do some code analysis and create a func_cells array, or
    we could use LOAD_NAME which should use PyMapping_GetItem on the
    globals dict.

Graphics

Ka-Ping Yee supplied a drawing of the state of things after "import
spam", where spam.py contains:

    import eggs

    i = -2
    max = 3

    def foo(n):
        y = abs(i) + max
        return eggs.ham(y + n)

The drawing is at http://web.lfw.org/repo/cells.gif; a larger version is
at http://lfw.org/repo/cells-big.gif; the source is at
http://lfw.org/repo/cells.ai.

Comparison

XXX Here, a comparison of the three approaches could be added.

Copyright

This document has been placed in the public domain.