Open Table Of Contents


This module provides a variety of transforms that transform the AST into a final form ready for code generation.

Below follows an explanation and justification of the design of the main compilation stages in numba.

We start with a Python AST, compiled from source code or decompiled from bytecode using meta. We run the following transformations:

  1. Type inference

    Infer types of all expressions, and fix the types of all local variables. Local variable types are promoted (for instance float to double), but cannot change (e.g. string cannot be assigned to float).

    When the type inferencer cannot determine a type, such as when it calls a Python function or method that is not a Numba function, it assumes type object. Object variables may be coerced to and from most native types.

    The type inferencer inserts CoercionNode nodes that perform such coercions, as well as coercions between promotable native types. It also resolves the return type of many math functions called in the numpy, math and cmath modules.

    Each AST expression node has a Variable that holds the type of the expression, as well as any meta-data such as constant values that have been determined.

  2. Transform for loops

    Provides some transformations of for loops over arrays to loops over a range. Iteration over range/xrange is resolved at compilation time.

    What I would like to see is the generation of a RangeNode holding a ast.Compare and an iteration variable incrementing ast.BinOp.

  3. Low level specializations (LateSpecializer)

    This stage performs low-level specializations. For instance it resolves coercions to and from object into calls such as PyFloat_FromDouble, with a fallback to Py_BuildValue/PyArg_ParseTuple.

    This specializer also has the responsibility to ensure that new references are accounted for by refcounting ObjectTempNode nodes. This node absorbs the references and lets parent nodes borrow the reference. At function cleanup, it decrefs its value. In loops, it also decrefs any previous value, if set. Hence, these temporaries must be initialized to NULL.

    An object temporary is specific to one specific sub-expression, and they are not reused (like in Cython).

    It also rewrites object attribute access and method calls into PyObject_GetAttrString etc.

  4. Code generation

    Generate LLVM code from the transformed AST.

    This should be as minimal as possible, and should not contain blobs of code performing complex operations. Instead, complex operations should be broken down by AST transformers into fundamental operations that are already supported by the AST.

    This way we maximize code reuse, and make potential future additions of different code generation backends easier. This can be taken only so far, since low-level transformations must also tailor to limitations of the code generation backend, such as intrinsic LLVM calls or calls into libc. However, code reuse is especially convenient in the face of type coercions, which LLVM does not provide any leniency for.

class numba.transforms.BuiltinResolverMixinBase

Base class for mixins resolving calls to built-in functions.

Methods called _resolve_<built-in name> are called to handle calls to the built-in of that name.

class numba.transforms.LateBuiltinResolverMixin

Perform final low-level transformations such as abs(value) -> fabs(value)