This document describes the toolchain that we have developed to analyze and “compile” RPython programs (like PyPy itself) to various target platforms.
It consists of three broad sections: a slightly simplified overview, a brief introduction to each of the major components of our toolchain and then a more comprehensive section describing how the pieces fit together. If you are reading this document for the first time, the Overview is likely to be most useful, if you are trying to refresh your PyPy memory then the How It Fits Together is probably what you want.
The job of the translation toolchain is to translate RPython programs into an efficient version of that program for one of various target platforms, generally one that is considerably lower-level than Python. It divides this task into several steps, and the purpose of this document is to introduce them.
As of the 1.2 release, RPython programs can be translated into the following languages/platforms: C/POSIX, CLI/.NET and Java/JVM.
The choice of the target platform affects the process somewhat, but to start with we describe the process of translating an RPython program into C (which is the default and original target).
The RPython translation toolchain never sees Python source code or syntax trees, but rather starts with the code objects that define the behaviour of the function objects one gives it as input. The bytecode evaluator and the Flow Object Space work through these code objects using abstract interpretation to produce a control flow graph (one per function): yet another representation of the source program, but one which is suitable for applying type inference and translation techniques and which is the fundamental data structure most of the translation steps operate on.
It is helpful to consider translation as being made up of the following steps (see also the figure below):
(although these steps are not quite as distinct as you might think from this presentation).
The following figure gives a simplified overview (PDF color version):
All these types are defined in rpython/flowspace/model.py (which is a rather important module in the PyPy source base, to reinforce the point).
The flow graph of a function is represented by the class FunctionGraph. It contains a reference to a collection of Blocks connected by Links.
A Block contains a list of SpaceOperations. Each SpaceOperation has an opname and a list of args and result, which are either Variables or Constants.
We have an extremely useful PyGame viewer, which allows you to visually inspect the graphs at various stages of the translation process (very useful to try to work out why things are breaking). It looks like this:
It is recommended to play with python bin/translatorshell.py on a few examples to get an idea of the structure of flow graphs. The following describes the types and their attributes in some detail:
A container for one graph (corresponding to one function).
|startblock:||the first block. It is where the control goes when the function is called. The input arguments of the startblock are the function’s arguments. If the function takes a *args argument, the args tuple is given as the last input argument of the startblock.|
|returnblock:||the (unique) block that performs a function return. It is empty, not actually containing any return operation; the return is implicit. The returned value is the unique input variable of the returnblock.|
|exceptblock:||the (unique) block that raises an exception out of the function. The two input variables are the exception class and the exception value, respectively. (No other block will actually link to the exceptblock if the function does not explicitly raise exceptions.)|
A basic block, containing a list of operations and ending in jumps to other basic blocks. All the values that are “live” during the execution of the block are stored in Variables. Each basic block uses its own distinct Variables.
|inputargs:||list of fresh, distinct Variables that represent all the values that can enter this block from any of the previous blocks.|
|operations:||list of SpaceOperations.|
|exits:||list of Links representing possible jumps from the end of this basic block to the beginning of other basic blocks.|
Each Block ends in one of the following ways:
A link from one basic block to another.
|prevblock:||the Block that this Link is an exit of.|
|target:||the target Block to which this Link points to.|
|args:||a list of Variables and Constants, of the same size as the target Block’s inputargs, which gives all the values passed into the next block. (Note that each Variable used in the prevblock may appear zero, one or more times in the args list.)|
|last_exception:||None or a Variable; see below.|
|last_exc_value:||None or a Variable; see below.|
Note that args uses Variables from the prevblock, which are matched to the target block’s inputargs by position, as in a tuple assignment or function call would do.
If the link is an exception-catching one, the last_exception and last_exc_value are set to two fresh Variables that are considered to be created when the link is entered; at run-time, they will hold the exception class and value, respectively. These two new variables can only be used in the same link’s args list, to be passed to the next block (as usual, they may actually not appear at all, or appear several times in args).
A recorded (or otherwise generated) basic operation.
|opname:||the name of the operation. The Flow Space produces only operations from the list in pypy.interpreter.baseobjspace, but later the names can be changed arbitrarily.|
|args:||list of arguments. Each one is a Constant or a Variable seen previously in the basic block.|
|result:||a new Variable into which the result is to be stored.|
Note that operations usually cannot implicitly raise exceptions at run-time; so for example, code generators can assume that a getitem operation on a list is safe and can be performed without bound checking. The exceptions to this rule are: (1) if the operation is the last in the block, which ends with exitswitch == Constant(last_exception), then the implicit exceptions must be checked for, generated, and caught appropriately; (2) calls to other functions, as per simple_call or call_args, can always raise whatever the called function can raise — and such exceptions must be passed through to the parent unless they are caught as above.
A placeholder for a run-time value. There is mostly debugging stuff here.
|name:||it is good style to use the Variable object itself instead of its name attribute to reference a value, although the name is guaranteed unique.|
A constant value used as argument to a SpaceOperation, or as value to pass across a Link to initialize an input Variable in the target Block.
|value:||the concrete value represented by this Constant.|
|key:||a hashable object representing the value.|
A Constant can occasionally store a mutable Python object. It represents a static, pre-initialized, read-only version of that object. The flow graph should not attempt to actually mutate such Constants.
We describe briefly below how a control flow graph can be “annotated” to discover the types of the objects. This annotation pass is a form of type inference. It operates on the control flow graphs built by the Flow Object Space.
For a more comprehensive description of the annotation process, see the corresponding section of our EU report about translation.
The major goal of the annotator is to “annotate” each variable that appears in a flow graph. An “annotation” describes all the possible Python objects that this variable could contain at run-time, based on a whole-program analysis of all the flow graphs – one per function.
An “annotation” is an instance of a subclass of SomeObject. Each subclass that represents a specific family of objects.
Here is an overview (see pypy/annotation/model/):
The result of the annotation pass is essentially a large dictionary mapping Variables to annotations.
All the SomeXxx instances are immutable. If the annotator needs to revise its belief about what a Variable can contain, it does so creating a new annotation, not mutating the existing one.
Mutable objects need special treatment during annotation, because the annotation of contained values needs to be possibly updated to account for mutation operations, and consequently the annotation information reflown through the relevant parts of the flow graphs.
SomeInstance stands for an instance of the given class or any subclass of it. For each user-defined class seen by the annotator, we maintain a ClassDef (pypy.annotation.classdef) describing the attributes of the instances of the class; essentially, a ClassDef gives the set of all class-level and instance-level attributes, and for each one, a corresponding SomeXxx annotation.
Instance-level attributes are discovered progressively as the annotation progresses. Assignments like:
inst.attr = value
update the ClassDef of the given instance to record that the given attribute exists and can be as general as the given value.
For every attribute, the ClassDef also records all the positions where the attribute is read. If, at some later time, we discover an assignment that forces the annotation about the attribute to be generalized, then all the places that read the attribute so far are marked as invalid and the annotator will restart its analysis from there.
The distinction between instance-level and class-level attributes is thin; class-level attributes are essentially considered as initial values for instance-level attributes. Methods are not special in this respect, except that they are bound to the instance (i.e. self = SomeInstance(cls)) when considered as the initial value for the instance.
The inheritance rules are as follows: the union of two SomeInstance annotations is the SomeInstance of the most precise common base class. If an attribute is considered (i.e. read or written) through a SomeInstance of a parent class, then we assume that all subclasses also have the same attribute, and that the same annotation applies to them all (so code like return self.x in a method of a parent class forces the parent class and all its subclasses to have an attribute x, whose annotation is general enough to contain all the values that all the subclasses might want to store in x). However, distinct subclasses can have attributes of the same names with different, unrelated annotations if they are not used in a general way through the parent class.
The RTyper is the first place where the choice of backend makes a difference; as outlined above we are assuming that ANSI C is the target.
The RPython Typer is the bridge between the Annotator and the code generator. The information computed by the annotator is high-level, in the sense that it describe RPython types like lists or instances of user-defined classes.
To emit code we need to represent these high-level annotations in the low-level model of the target language; for C, this means structures and pointers and arrays. The Typer both determines the appropriate low-level type for each annotation and replaces each high-level operation in the control flow graphs with one or a few low-level operations. Just like low-level types, there is only a fairly restricted set of low-level operations, along the lines of reading or writing from or to a field of a structure.
In theory, this step is optional; a code generator might be able to read directly the high-level types. Our experience, however, suggests that this is very unlikely to be practical. “Compiling” high-level types into low-level ones is rather more messy than one would expect and this was the motivation for making this step explicit and isolated in a single place. After RTyping, the graphs only contain operations that already live on the level of the target language, which makes the job of the code generators much simpler.
For more detailed information, see the documentation for the RTyper.
Integer operations are make an easy example. Assume a graph containing the following operation:
v3 = add(v1, v2)
v1 -> SomeInteger() v2 -> SomeInteger() v3 -> SomeInteger()
then obviously we want to type it and replace it with:
v3 = int_add(v1, v2)
where – in C notation – all three variables v1, v2 and v3 are typed int. This is done by attaching an attribute concretetype to v1, v2 and v3 (which might be instances of Variable or possibly Constant). In our model, this concretetype is pypy.rpython.lltypesystem.lltype.Signed. Of course, the purpose of replacing the operation called add with int_add is that code generators no longer have to worry about what kind of addition (or concatenation maybe?) it means.
Between RTyping and C source generation there are two optional transforms: the “backend optimizations” and the “stackless transform”. See also D07.1 Massive Parallelism and Translation Aspects for further details.
The point of the backend optimizations are to make the compiled program run faster. Compared to many parts of the PyPy translator, which are very unlike a traditional compiler, most of these will be fairly familiar to people who know how compilers work.
To reduce the overhead of the many function calls that occur when running the PyPy interpreter we implemented function inlining. This is an optimization which takes a flow graph and a callsite and inserts a copy of the flow graph into the graph of the calling function, renaming occurring variables as appropriate. This leads to problems if the original function was surrounded by a try: ... except: ... guard. In this case inlining is not always possible. If the called function is not directly raising an exception (but an exception is potentially raised by further called functions) inlining is safe, though.
In addition we also implemented heuristics which function to inline where. For this purpose we assign every function a “size”. This size should roughly correspond to the increase in code-size which is to be expected should the function be inlined somewhere. This estimate is the sum of two numbers: for one every operations is assigned a specific weight, the default being a weight of one. Some operations are considered to be more effort than others, e.g. memory allocation and calls; others are considered to be no effort at all (casts...). The size estimate is for one the sum of the weights of all operations occurring in the graph. This is called the “static instruction count”. The other part of the size estimate of a graph is the “median execution cost”. This is again the sum of the weight of all operations in the graph, but this time weighted with a guess how often the operation is executed. To arrive at this guess we assume that at every branch we take both paths equally often, except for branches that are the end of loops, where the jump back to the end of the loop is considered more likely. This leads to a system of equations which can be solved to get approximate weights for all operations.
After the size estimate for all function has been determined, functions are being inlined into their callsites, starting from the smallest functions. Every time a function is being inlined into another function, the size of the outer function is recalculated. This is done until the remaining functions all have a size greater than a predefined limit.
Since RPython is a garbage collected language there is a lot of heap memory allocation going on all the time, which would either not occur at all in a more traditional explicitly managed language or results in an object which dies at a time known in advance and can thus be explicitly deallocated. For example a loop of the following form:
for i in range(n): ...
which simply iterates over all numbers from 0 to n - 1 is equivalent to the following in Python:
l = range(n) iterator = iter(l) try: while 1: i = iterator.next() ... except StopIteration: pass
Which means that three memory allocations are executed: The range object, the iterator for the range object and the StopIteration instance, which ends the loop.
After a small bit of inlining all these three objects are never even passed as arguments to another function and are also not stored into a globally reachable position. In such a situation the object can be removed (since it would die anyway after the function returns) and can be replaced by its contained values.
This pattern (an allocated object never leaves the current function and thus dies after the function returns) occurs quite often, especially after some inlining has happened. Therefore we implemented an optimization which “explodes” objects and thus saves one allocation in this simple (but quite common) situation.
Another technique to reduce the memory allocation penalty is to use stack allocation for objects that can be proved not to life longer than the stack frame they have been allocated in. This proved not to really gain us any speed, so over time it was removed again.
The stackless transform converts functions into a form that knows how to save the execution point and active variables into a heap structure and resume execution at that point. This was used to implement coroutines as an RPython-level feature, which in turn are used to implement coroutines, greenlets and tasklets as an application level feature for the Standard Interpreter.
The stackless transformation has been deprecated and is no longer available in trunk. It has been replaced with continulets.
This, perhaps slightly vaguely named, stage is the most recent to appear as a separate step. Its job is to make the final implementation decisions before source generation – experience has shown that you really don’t want to be doing any thinking at the same time as actually generating source code. For the C backend, this step does three things:
- inserts explicit exception handling,
- inserts explicit memory management operations,
- decides on the names functions and types will have in the final source (this mapping of objects to names is sometimes referred to as the “low-level database”).
RPython code is free to use exceptions in much the same way as unrestricted Python, but the final result is a C program, and C has no concept of exceptions. The exception transformer implements exception handling in a similar way to CPython: exceptions are indicated by special return values and the current exception is stored in a global data structure.
In a sense the input to the exception transformer is a program in terms of the lltypesystem with exceptions and the output is a program in terms of the bare lltypesystem.
As well as featuring exceptions, RPython is a garbage collected language; again, C is not. To square this circle, decisions about memory management must be made. In keeping with PyPy’s approach to flexibility, there is freedom to change how to do it. There are three approaches implemented today:
- reference counting (deprecated, too slow)
- using the Boehm-Demers-Weiser conservative garbage collector
- using one of our custom exact GCs implemented in RPython
Almost all application-level Python code allocates objects at a very fast rate; this means that the memory management implementation is critical to the performance of the PyPy interpreter.
You can choose which garbage collection strategy to use with –gc.
GenC is usually the most actively maintained backend – everyone working on PyPy has a C compiler, for one thing – and is usually where new features are implemented first.
As this document has shown, the translation step is divided into more steps than one might at first expect. It is certainly divided into more steps than we expected when the project started; the very first version of GenC operated on the high-level flow graphs and the output of the annotator, and even the concept of the RTyper didn’t exist yet. More recently, the fact that preparing the graphs for source generation (“databasing”) and actually generating the source are best considered separately has become clear.
Use the –backend option to choose which backend to use.
The Object-Oriented backends target platforms that are less C-like and support classes, instance etc. If such a platform is targeted, the OO type system is used while rtyping. Of the OO backends, both gencli and genjava can translate the full Python interpreter.
It is the most advanced of the object oriented backends – it can compile the PyPy interpreter as well as our two standard benchmarks, RPyStone (CPython’s PyStone benchmark modified slightly to be RPython) and a RPython version of the common Richards benchmark.
It is almost entirely the work of Antonio Cuni, who started this backend as part of his Master’s thesis, the Google’s Summer of Code 2006 program and the Summer of PyPy program.
GenJVM targets the Java Virtual Machine: it translates RPython programs directly into Java bytecode, similarly to what GenCLI does.
So far it is the second most mature high level backend after GenCLI: it still can’t translate the full Standard Interpreter, but after the Leysin sprint we were able to compile and run the rpystone and richards benchmarks.
GenJVM is almost entirely the work of Niko Matsakis, who worked on it also as part of the Summer of PyPy program.
As should be clear by now, the translation toolchain of PyPy is a flexible and complicated beast, formed from many separate components.
The following image summarizes the various parts of the toolchain as of the 0.9 release, with the default translation to C highlighted:
A detail that has not yet been emphasized is the interaction of the various components. It makes for a nice presentation to say that after the annotator has finished the RTyper processes the graphs and then the exception handling is made explicit and so on, but it’s not entirely true. For example, the RTyper inserts calls to many low-level helpers which must first be annotated, and the GC transformer can use inlining (one of the backend optimizations) of some of its small helper functions to improve performance. The following picture attempts to summarize the components involved in performing each step of the default translation process:
A component not mentioned before is the “MixLevelAnnotator”; it provides a convenient interface for a “late” (after RTyping) translation step to declare that it needs to be able to call each of a collection of functions (which may refer to each other in a mutually recursive fashion) and annotate and rtype them all at once.