|Title:||Optimizing Global Variable/Attribute Access|
|Author:||skip at pobox.com (Skip Montanaro)|
The bindings for most global variables and attributes of other modules typically never change during the execution of a Python program, but because of Python's dynamic nature, code which accesses such global objects must run through a full lookup each time the object is needed. This PEP proposes a mechanism that allows code that accesses most global objects to treat them as local objects and places the burden of updating references on the code that changes the name bindings of such objects.
Consider the workhorse function sre_compile._compile. It is the internal compilation function for the sre module. It consists almost entirely of a loop over the elements of the pattern being compiled, comparing opcodes with known constant values and appending tokens to an output list. Most of the comparisons are with constants imported from the sre_constants module. This means there are lots of LOAD_GLOBAL bytecodes in the compiled output of this module. Just by reading the code it's apparent that the author intended LITERAL, NOT_LITERAL, OPCODES and many other symbols to be constants. Still, each time they are involved in an expression, they must be looked up anew.
Most global accesses are actually to objects that are "almost constants". This includes global variables in the current module as well as the attributes of other imported modules. Since they rarely change, it seems reasonable to place the burden of updating references to such objects on the code that changes the name bindings. If sre_constants.LITERAL is changed to refer to another object, perhaps it would be worthwhile for the code that modifies the sre_constants module dict to correct any active references to that object. By doing so, in many cases global variables and the attributes of many objects could be cached as local variables. If the bindings between the names given to the objects and the objects themselves changes rarely, the cost of keeping track of such objects should be low and the potential payoff fairly large.
In an attempt to gauge the effect of this proposal, I modified the Pystone benchmark program included in the Python distribution to cache global functions. Its main function, Proc0, makes calls to ten different functions inside its for loop. In addition, Func2 calls Func1 repeatedly inside a loop. If local copies of these 11 global idenfiers are made before the functions' loops are entered, performance on this particular benchmark improves by about two percent (from 5561 pystones to 5685 on my laptop). It gives some indication that performance would be improved by caching most global variable access. Note also that the pystone benchmark makes essentially no accesses of global module attributes, an anticipated area of improvement for this PEP.
I propose that the Python virtual machine be modified to include TRACK_OBJECT and UNTRACK_OBJECT opcodes. TRACK_OBJECT would associate a global name or attribute of a global name with a slot in the local variable array and perform an initial lookup of the associated object to fill in the slot with a valid value. The association it creates would be noted by the code responsible for changing the name-to-object binding to cause the associated local variable to be updated. The UNTRACK_OBJECT opcode would delete any association between the name and the local variable slot.
Operation of this code in threaded programs will be no different than in unthreaded programs. If you need to lock an object to access it, you would have had to do that before TRACK_OBJECT would have been executed and retain that lock until after you stop using it.
FIXME: I suspect I need more here.
Global variables and attributes rarely change. For example, once a function imports the math module, the binding between the name math and the module it refers to aren't likely to change. Similarly, if the function that uses the math module refers to its sin attribute, it's unlikely to change. Still, every time the module wants to call the math.sin function, it must first execute a pair of instructions:
LOAD_GLOBAL math LOAD_ATTR sin
If the client module always assumed that math.sin was a local constant and it was the responsibility of "external forces" outside the function to keep the reference correct, we might have code like this:
TRACK_OBJECT math.sin ... LOAD_FAST math.sin ... UNTRACK_OBJECT math.sin
If the LOAD_FAST was in a loop the payoff in reduced global loads and attribute lookups could be significant.
This technique could, in theory, be applied to any global variable access or attribute lookup. Consider this code:
l =  for i in range(10): l.append(math.sin(i)) return l
Even though l is a local variable, you still pay the cost of loading l.append ten times in the loop. The compiler (or an optimizer) could recognize that both math.sin and l.append are being called in the loop and decide to generate the tracked local code, avoiding it for the builtin range() function because it's only called once during loop setup. Performance issues related to accessing local variables make tracking l.append less attractive than tracking globals such as math.sin.
According to a post to python-dev by Marc-Andre Lemburg , LOAD_GLOBAL opcodes account for over 7% of all instructions executed by the Python virtual machine. This can be a very expensive instruction, at least relative to a LOAD_FAST instruction, which is a simple array index and requires no extra function calls by the virtual machine. I believe many LOAD_GLOBAL instructions and LOAD_GLOBAL/LOAD_ATTR pairs could be converted to LOAD_FAST instructions.
Code that uses global variables heavily often resorts to various tricks to avoid global variable and attribute lookup. The aforementioned sre_compile._compile function caches the append method of the growing output list. Many people commonly abuse functions' default argument feature to cache global variable lookups. Both of these schemes are hackish and rarely address all the available opportunities for optimization. (For example, sre_compile._compile does not cache the two globals that it uses most frequently: the builtin len function and the global OPCODES array that it imports from sre_constants.py.
I believe the global interpreter lock will protect values from being corrupted. In any case, the situation would be no worse than it is today. If one thread modified math.sin after another thread had already executed LOAD_GLOBAL math, but before it executed LOAD_ATTR sin, the client thread would see the old value of math.sin.
The idea is this. I use a multi-attribute load below as an example, not because it would happen very often, but because by demonstrating the recursive nature with an extra call hopefully it will become clearer what I have in mind. Suppose a function defined in module foo wants to access spam.eggs.ham and that spam is a module imported at the module level in foo:
import spam ... def somefunc(): ... x = spam.eggs.ham
Upon entry to somefunc, a TRACK_GLOBAL instruction will be executed:
TRACK_GLOBAL spam.eggs.ham n
spam.eggs.ham is a string literal stored in the function's constants array. n is a fastlocals index. &fastlocals[n] is a reference to slot n in the executing frame's fastlocals array, the location in which the spam.eggs.ham reference will be stored. Here's what I envision happening:
The TRACK_GLOBAL instruction locates the object referred to by the name spam and finds it in its module scope. It then executes a C function like:
_PyObject_TrackName(m, "spam.eggs.ham", &fastlocals[n])
where m is the module object with an attribute spam.
The module object strips the leading spam. and stores the necessary information (eggs.ham and &fastlocals[n]) in case its binding for the name eggs changes. It then locates the object referred to by the key eggs in its dict and recursively calls:
_PyObject_TrackName(eggs, "eggs.ham", &fastlocals[n])
The eggs object strips the leading eggs., stores the (ham, &fastlocals[n]) info, locates the object in its namespace called ham and calls _PyObject_TrackName once again:
_PyObject_TrackName(ham, "ham", &fastlocals[n])
The ham object strips the leading string (no "." this time, but that's a minor point), sees that the result is empty, then uses its own value (self, probably) to update the location it was handed:
Py_XDECREF(&fastlocals[n]); &fastlocals[n] = self; Py_INCREF(&fastlocals[n]);
At this point, each object involved in resolving spam.eggs.ham knows which entry in its namespace needs to be tracked and what location to update if that name changes. Furthermore, if the one name it is tracking in its local storage changes, it can call _PyObject_TrackName using the new object once the change has been made. At the bottom end of the food chain, the last object will always strip a name, see the empty string and know that its value should be stuffed into the location it's been passed.
When the object referred to by the dotted expression spam.eggs.ham is going to go out of scope, an UNTRACK_GLOBAL spam.eggs.ham n instruction is executed. It has the effect of deleting all the tracking information that TRACK_GLOBAL established.
The tracking operation may seem expensive, but recall that the objects being tracked are assumed to be "almost constant", so the setup cost will be traded off against hopefully multiple local instead of global loads. For globals with attributes the tracking setup cost grows but is offset by avoiding the extra LOAD_ATTR cost. The TRACK_GLOBAL instruction needs to perform a PyDict_GetItemString for the first name in the chain to determine where the top-level object resides. Each object in the chain has to store a string and an address somewhere, probably in a dict that uses storage locations as keys (e.g. the &fastlocals[n]) and strings as values. (This dict could possibly be a central dict of dicts whose keys are object addresses instead of a per-object dict.) It shouldn't be the other way around because multiple active frames may want to track spam.eggs.ham, but only one frame will want to associate that name with one of its fast locals slots.
What about this (dumb) code?:
l =  lock = threading.Lock() ... def fill_l():: for i in range(1000):: lock.acquire() l.append(math.sin(i)) lock.release() ... def consume_l():: while 1:: lock.acquire() if l:: elt = l.pop() lock.release() fiddle(elt)
It's not clear from a static analysis of the code what the lock is protecting. (You can't tell at compile-time that threads are even involved can you?) Would or should it affect attempts to track l.append or math.sin in the fill_l function?
If we annotate the code with mythical track_object and untrack_object builtins (I'm not proposing this, just illustrating where stuff would go!), we get:
l =  lock = threading.Lock() ... def fill_l():: track_object("l.append", append) track_object("math.sin", sin) for i in range(1000):: lock.acquire() append(sin(i)) lock.release() untrack_object("math.sin", sin) untrack_object("l.append", append) ... def consume_l():: while 1:: lock.acquire() if l:: elt = l.pop() lock.release() fiddle(elt)
Is that correct both with and without threads (or at least equally incorrect with and without threads)?
The presence of nested scopes will affect where TRACK_GLOBAL finds a global variable, but shouldn't affect anything after that. (I think.)
Suppose I am tracking the object referred to by spam.eggs.ham and spam.eggs is rebound to an object that does not have a ham attribute. It's clear this will be an AttributeError if the programmer attempts to resolve spam.eggs.ham in the current Python virtual machine, but suppose the programmer has anticipated this case:
if hasattr(spam.eggs, "ham"): print spam.eggs.ham elif hasattr(spam.eggs, "bacon"): print spam.eggs.bacon else: print "what? no meat?"
You can't raise an AttributeError when the tracking information is recalculated. If it does not raise AttributeError and instead lets the tracking stand, it may be setting the programmer up for a very subtle error.
One solution to this problem would be to track the shortest possible root of each dotted expression the function refers to directly. In the above example, spam.eggs would be tracked, but spam.eggs.ham and spam.eggs.bacon would not.
In the Questions section I postulated the existence of a _PyObject_TrackName function. While the API is fairly easy to specify, the implementation behind-the-scenes is not so obvious. A central dictionary could be used to track the name/location mappings, but it appears that all setattr functions might need to be modified to accommodate this new functionality.
If all types used the PyObject_GenericSetAttr function to set attributes that would localize the update code somewhat. They don't however (which is not too surprising), so it seems that all getattrfunc and getattrofunc functions will have to be updated. In addition, this would place an absolute requirement on C extension module authors to call some function when an attribute changes value (PyObject_TrackUpdate?).
Finally, it's quite possible that some attributes will be set by side effect and not by any direct call to a setattr method of some sort. Consider a device interface module that has an interrupt routine that copies the contents of a device register into a slot in the object's struct whenever it changes. In these situations, more extensive modifications would have to be made by the module author. To identify such situations at compile time would be impossible. I think an extra slot could be added to PyTypeObjects to indicate if an object's code is safe for global tracking. It would have a default value of 0 (Py_TRACKING_NOT_SAFE). If an extension module author has implemented the necessary tracking support, that field could be initialized to 1 (Py_TRACKING_SAFE). _PyObject_TrackName could check that field and issue a warning if it is asked to track an object that the author has not explicitly said was safe for tracking.
Jeremy Hylton has an alternate proposal on the table . His proposal seeks to create a hybrid dictionary/list object for use in global name lookups that would make global variable access look more like local variable access. While there is no C code available to examine, the Python implementation given in his proposal still appears to require dictionary key lookup. It doesn't appear that his proposal could speed local variable attribute lookup, which might be worthwhile in some situations if potential performance burdens could be addressed.
I don't believe there will be any serious issues of backward compatibility. Obviously, Python bytecode that contains TRACK_OBJECT opcodes could not be executed by earlier versions of the interpreter, but breakage at the bytecode level is often assumed between versions.
TBD. This is where I need help. I believe there should be either a central name/location registry or the code that modifies object attributes should be modified, but I'm not sure the best way to go about this. If you look at the code that implements the STORE_GLOBAL and STORE_ATTR opcodes, it seems likely that some changes will be required to PyDict_SetItem and PyObject_SetAttr or their String variants. Ideally, there'd be a fairly central place to localize these changes. If you begin considering tracking attributes of local variables you get into issues of modifying STORE_FAST as well, which could be a problem, since the name bindings for local variables are changed much more frequently. (I think an optimizer could avoid inserting the tracking code for the attributes for any local variables where the variable's name binding changes.)
I believe (though I have no code to prove it at this point), that implementing TRACK_OBJECT will generally not be much more expensive than a single LOAD_GLOBAL instruction or a LOAD_GLOBAL/LOAD_ATTR pair. An optimizer should be able to avoid converting LOAD_GLOBAL and LOAD_GLOBAL/LOAD_ATTR to the new scheme unless the object access occurred within a loop. Further down the line, a register-oriented replacement for the current Python virtual machine  could conceivably eliminate most of the LOAD_FAST instructions as well.
The number of tracked objects should be relatively small. All active frames of all active threads could conceivably be tracking objects, but this seems small compared to the number of functions defined in a given application.
This document has been placed in the public domain.