1. Pypy
  2. Untitled project
  3. pypy


pypy / pypy / doc / cli-backend.rst

The CLI backend

The goal of GenCLI is to compile RPython programs to the CLI virtual machine.

Target environment and language

The target of GenCLI is the Common Language Infrastructure environment as defined by the Standard Ecma 335.

While in an ideal world we might suppose GenCLI to run fine with every implementation conforming to that standard, we know the world we live in is far from ideal, so extra efforts can be needed to maintain compatibility with more than one implementation.

At the moment of writing the two most popular implementations of the standard are supported: Microsoft Common Language Runtime (CLR) and Mono.

Then we have to choose how to generate the real executables. There are two main alternatives: generating source files in some high level language (such as C#) or generating assembly level code in Intermediate Language (IL).

The IL approach is much faster during the code generation phase, because it doesn't need to call a compiler. By contrast the high level approach has two main advantages:

  • the code generation part could be easier because the target language supports high level control structures such as structured loops;
  • the generated executables take advantage of compiler's optimizations.

In reality the first point is not an advantage in the PyPy context, because the flow graph we start from is quite low level and Python loops are already expressed in terms of branches (i.e., gotos).

About the compiler optimizations we must remember that the flow graph we receive from earlier stages is already optimized: PyPy implements a number of optimizations such a constant propagation and dead code removal, so it's not obvious if the compiler could do more.

Moreover by emitting IL instruction we are not constrained to rely on compiler choices but can directly choose how to map CLI opcodes: since the backend often know more than the compiler about the context, we might expect to produce more efficient code by selecting the most appropriate instruction; e.g., we can check for arithmetic overflow only when strictly necessary.

The last but not least reason for choosing the low level approach is flexibility in how to get an executable starting from the IL code we generate:

  • write IL code to a file, then call the ilasm assembler;
  • directly generate code on the fly by accessing the facilities exposed by the System.Reflection.Emit API.

Handling platform differences

Since our goal is to support both Microsoft CLR we have to handle the differences between the twos; in particular the main differences are in the name of the helper tools we need to call:

Tool CLR Mono
IL assembler ilasm ilasm2
C# compiler csc gmcs
Runtime ... mono

The code that handles these differences is located in the sdk.py module: it defines an abstract class which exposes some methods returning the name of the helpers and one subclass for each of the two supported platforms.

Since Microsoft ilasm is not capable of compiling the PyPy standard interpreter due to its size, on Windows machines we also look for an existing Mono installation: if present, we use CLR for everything except the assembling phase, for which we use Mono's ilasm2.

Targeting the CLI Virtual Machine

In order to write a CLI backend we have to take a number of decisions. First, we have to choose the typesystem to use: given that CLI natively supports primitives like classes and instances, ootypesystem is the most natural choice.

Once the typesystem has been chosen there is a number of steps we have to do for completing the backend:

  • map ootypesystem's types to CLI Common Type System's types;
  • map ootypesystem's low level operation to CLI instructions;
  • map Python exceptions to CLI exceptions;
  • write a code generator that translates a flow graph into a list of CLI instructions;
  • write a class generator that translates ootypesystem classes into CLI classes.

Mapping primitive types

The rtyper give us a flow graph annotated with types belonging to ootypesystem: in order to produce CLI code we need to translate these types into their Common Type System equivalents.

For numeric types the conversion is straightforward, since there is a one-to-one mapping between the two typesystems, so that e.g. Float maps to float64.

For character types the choice is more difficult: RPython has two distinct types for plain ASCII and Unicode characters (named UniChar), while .NET only supports Unicode with the char type. There are at least two ways to map plain Char to CTS:

  • map UniChar to char, thus maintaining the original distinction between the two types: this has the advantage of being a one-to-one translation, but has the disadvantage that RPython strings will not be recognized as .NET strings, since they only would be sequences of bytes;
  • map both char, so that Python strings will be treated as strings also by .NET: in this case there could be problems with existing Python modules that use strings as sequences of byte, such as the built-in struct module, so we need to pay special attention.

We think that mapping Python strings to .NET strings is fundamental, so we chose the second option.

Mapping built-in types

As we saw in section ootypesystem defines a set of types that take advantage of built-in types offered by the platform.

For the sake of simplicity we decided to write wrappers around .NET classes in order to match the signatures required by pypylib.dll:

ootype CLI
String System.String
StringBuilder System.Text.StringBuilder
List System.Collections.Generic.List<T>
Dict System.Collections.Generic.Dictionary<K, V>
CustomDict pypy.runtime.Dict
DictItemsIterator pypy.runtime.DictItemsIterator

Wrappers exploit inheritance for wrapping the original classes, so, for example, pypy.runtime.List<T> is a subclass of System.Collections.Generic.List<T> that provides methods whose names match those found in the _GENERIC_METHODS of ootype.List

The only exception to this rule is the String class, which is not wrapped since in .NET we can not subclass System.String. Instead, we provide a bunch of static methods in pypylib.dll that implement the methods declared by ootype.String._GENERIC_METHODS, then we call them by explicitly passing the string object in the argument list.

Mapping instructions

PyPy's low level operations are expressed in Static Single Information (SSI) form, such as this:

v2 = int_add(v0, v1)

By contrast the CLI virtual machine is stack based, which means the each operation pops its arguments from the top of the stacks and pushes its result there. The most straightforward way to translate SSI operations into stack based operations is to explicitly load the arguments and store the result into the appropriate places:


The code produced works correctly but has some inefficiency issues that can be addressed during the optimization phase.

The CLI Virtual Machine is fairly expressive, so the conversion between PyPy's low level operations and CLI instruction is relatively simple: many operations maps directly to the corresponding instruction, e.g int_add and sub.

By contrast some instructions do not have a direct correspondent and have to be rendered as a sequence of CLI instructions: this is the case of the "less-equal" and "greater-equal" family of instructions, that are rendered as "greater" or "less" followed by a boolean "not", respectively.

Finally, there are some instructions that cannot be rendered directly without increasing the complexity of the code generator, such as int_abs (which returns the absolute value of its argument). These operations are translated by calling some helper function written in C#.

The code that implements the mapping is in the modules opcodes.py.

Mapping exceptions

Both RPython and CLI have their own set of exception classes: some of these are pretty similar; e.g., we have OverflowError, ZeroDivisionError and IndexError on the first side and OverflowException, DivideByZeroException and IndexOutOfRangeException on the other side.

The first attempt was to map RPython classes to their corresponding CLI ones: this worked for simple cases, but it would have triggered subtle bugs in more complex ones, because the two exception hierarchies don't completely overlap.

At the moment we've chosen to build an RPython exception hierarchy completely independent from the CLI one, but this means that we can't rely on exceptions raised by built-in operations. The currently implemented solution is to do an exception translation on-the-fly.

As an example consider the RPython int_add_ovf operation, that sums two integers and raises an OverflowError exception in case of overflow. For implementing it we can use the built-in add.ovf CLI instruction that raises System.OverflowException when the result overflows, catch that exception and throw a new one:

    ldarg 'x_0'
    ldarg 'y_0'
    stloc 'v1'
    leave __check_block_2
catch [mscorlib]System.OverflowException
    newobj instance void class OverflowError::.ctor()

Translating flow graphs

As we saw previously in PyPy function and method bodies are represented by flow graphs that we need to translate CLI IL code. Flow graphs are expressed in a format that is very suitable for being translated to low level code, so that phase is quite straightforward, though the code is a bit involved because we need to take care of three different types of blocks.

The code doing this work is located in the Function.render method in the file function.py.

First of all it searches for variable names and types used by each block; once they are collected it emits a .local IL statement used for indicating the virtual machine the number and type of local variables used.

Then it sequentially renders all blocks in the graph, starting from the start block; special care is taken for the return block which is always rendered at last to meet CLI requirements.

Each block starts with an unique label that is used for jumping across, followed by the low level instructions the block is composed of; finally there is some code that jumps to the appropriate next block.

Conditional and unconditional jumps are rendered with their corresponding IL instructions: brtrue, brfalse.

Blocks that needs to catch exceptions use the native facilities offered by the CLI virtual machine: the entire block is surrounded by a .try statement followed by as many catch as needed: each catching sub-block then branches to the appropriate block:

# RPython
    # block0
except ValueError:
    # block1
except TypeError:
    # block2

// IL
  .try {
      leave block3
   catch ValueError {
      leave block1
    catch TypeError {
      leave block2
    br block3
    br block3

There is also an experimental feature that makes GenCLI to use its own exception handling mechanism instead of relying on the .NET one. Surprisingly enough, benchmarks are about 40% faster with our own exception handling machinery.

Translating classes

As we saw previously, the semantic of ootypesystem classes is very similar to the .NET one, so the translation is mostly straightforward.

The related code is located in the module class_.py. Rendered classes are composed of four parts:

  • fields;
  • user defined methods;
  • default constructor;
  • the ToString method, mainly for testing purposes

Since ootype implicitly assumes all method calls to be late bound, as an optimization before rendering the classes we search for methods that are not overridden in subclasses, and declare as "virtual" only the one that needs to.

The constructor does nothing more than calling the base class constructor and initializing class fields to their default value.

Inheritance is straightforward too, as it is natively supported by CLI. The only noticeable thing is that we map ootypesystem's ROOT class to the CLI equivalent System.Object.

The Runtime Environment

The runtime environment is a collection of helper classes and functions used and referenced by many of the GenCLI submodules. It is written in C#, compiled to a DLL (Dynamic Link Library), then linked to generated code at compile-time.

The DLL is called pypylib and is composed of three parts:

  • a set of helper functions used to implements complex RPython low-level instructions such as runtimenew and ooparse_int;
  • a set of helper classes wrapping built-in types
  • a set of helpers used by the test framework

The first two parts are contained in the pypy.runtime namespace, while the third is in the pypy.test one.

Testing GenCLI

As the rest of PyPy, GenCLI is a test-driven project: there is at least one unit test for almost each single feature of the backend. This development methodology allowed us to early discover many subtle bugs and to do some big refactoring of the code with the confidence not to break anything.

The core of the testing framework is in the module rpython.translator.cli.test.runtest; one of the most important function of this module is compile_function(): it takes a Python function, compiles it to CLI and returns a Python object that runs the just created executable when called.

This way we can test GenCLI generated code just as if it were a simple Python function; we can also directly run the generated executable, whose default name is main.exe, from a shell: the function parameters are passed as command line arguments, and the return value is printed on the standard output:

# Python source: foo.py
from rpython.translator.cli.test.runtest import compile_function

def foo(x, y):
    return x+y, x*y

f = compile_function(foo, [int, int])
assert f(3, 4) == (7, 12)

# shell
$ mono main.exe 3 4
(7, 12)

GenCLI supports only few RPython types as parameters: int, r_uint, r_longlong, r_ulonglong, bool, float and one-length strings (i.e., chars). By contrast, most types are fine for being returned: these include all primitive types, list, tuples and instances.

Installing Python for .NET on Linux

With the CLI backend, you can access .NET libraries from RPython; programs using .NET libraries will always run when translated, but you might also want to test them on top of CPython.

To do so, you can install Python for .NET. Unfortunately, it does not work out of the box under Linux.

To make it work, download and unpack the source package of Python for .NET; the only version tested with PyPy is the 1.0-rc2, but it might work also with others. Then, you need to create a file named Python.Runtime.dll.config at the root of the unpacked archive; put the following lines inside the file (assuming you are using Python 2.7):

  <dllmap dll="python27" target="libpython2.7.so.1.0" os="!windows"/>

The installation should be complete now. To run Python for .NET, simply type mono python.exe.