MLIR

Multi-Level IR Compiler Framework

MLIR C API

Current status: Under development, API unstable, built by default.

Design 

Many languages can interoperate with C but have a harder time with C++ due to name mangling and memory model differences. Although the C API for MLIR can be used directly from C, it is primarily intended to be wrapped in higher-level language- or library-specific constructs. Therefore the API tends towards simplicity and feature minimalism.

Note: while the C API is expected to be more stable than C++ API, it currently offers no stability guarantees.

Scope 

The API is provided for core IR components (attributes, blocks, operations, regions, types, values), Passes and some fundamental type and attribute kinds. The core IR API is intentionally low-level, e.g. exposes a plain list of operation’s operands and attributes without attempting to assign “semantic” names to them. Users of specific dialects are expected to wrap the core API in a dialect-specific way, for example, by implementing an ODS backend.

Object Model 

Core IR components are exposed as opaque handles to an IR object existing in C++. They are not intended to be inspected by the API users (and, in many cases, cannot be meaningfully inspected). Instead the users are expected to pass handles to the appropriate manipulation functions.

The handle may or may not own the underlying object.

Naming Convention and Ownership Model 

All objects are prefixed with Mlir. They are typedefs and should be used without struct.

All functions are prefixed with mlir.

Functions primarily operating on an instance of MlirX are prefixed with mlirX. They take the instance being acted upon as their first argument (except for creation functions). For example, mlirOperationGetNumOperands inspects an MlirOperation, which it takes as its first operand.

The ownership model is encoded in the naming convention as follows.

  • By default, the ownership is not transferred.
  • Functions that transfer the ownership of the result to the caller can be in one of two forms:
    • functions that create a new object have the name mlirXCreate<...>, for example, mlirOperationCreate;
    • functions that detach an object from a parent object have the name mlirYTake<...>, for example mlirOperationStateTakeRegion.
  • Functions that take ownership of some of their arguments have the form mlirY<...>OwnedX<...> where X can refer to the type or any other sufficiently unique description of the argument, the ownership of which will be taken by the callee, for example mlirRegionAppendOwnedBlock.
  • Functions that create an object by default do not transfer its ownership to the caller, i.e. one of other objects passed in as an argument retains the ownership, they have the form mlirX<...>Get. For example, mlirTypeParseGet.
  • Functions that destroy an object owned by the caller are of the form mlirXDestroy.

If the code owns an object, it is responsible for destroying the object when it is no longer necessary. If an object that owns other objects is destroyed, any handles to those objects become invalid. Note that types and attributes are owned by the MlirContext in which they were created.

Nullity 

A handle may refer to a null object. It is the responsibility of the caller to check if an object is null by using mlirXIsNull(MlirX). API functions do not expect null objects as arguments unless explicitly stated otherwise. API functions may return null objects.

Type Hierarchies 

MLIR objects can form type hierarchies in C++. For example, all IR classes representing types are derived from mlir::Type, some of them may also be also derived from common base classes such as mlir::ShapedType or dialect-specific base classes. Type hierarchies are exposed to C API through naming conventions as follows.

  • Only the top-level class of each hierarchy is exposed, e.g. MlirType is defined as a type but MlirShapedType is not. This avoids the need for explicit upcasting when passing an object of a derived type to a function that expects a base type (this happens more often in core/standard APIs, while downcasting usually involves further checks anyway).
  • A type Y that derives from X provides a function int mlirXIsAY(MlirX) that returns a non-zero value if the given dynamic instance of X is also an instance of Y. For example, int MlirTypeIsAInteger(MlirType).
  • A function that expects a derived type as its first argument takes the base type instead and documents the expectation by using Y in its name MlirY<...>(MlirX, ...). This function asserts that the dynamic instance of its first argument is Y, and it is the responsibility of the caller to ensure it is indeed the case.

Auxiliary Types 

StringRef 

Numerous MLIR functions return instances of StringRef to refer to a non-owning segment of a string. This segment may or may not be null-terminated. In C API, these are represented as instances of MlirStringRef structure that contains a pointer to the first character of the string fragment (str) and the fragment length (length). Note that the fragment is not necessarily null-terminated, the length field must be used to identify the last character. MlirStringRef is a non-owning pointer, the caller is in charge of performing the copy or ensuring that the pointee outlives all uses of MlirStringRef.

Printing 

IR objects can be printed using mlirXPrint(MlirX, MlirStringCallback, void *) functions. These functions accept take arguments a callback with signature void (*)(const char *, intptr_t, void *) and a pointer to user-defined data. They call the callback and supply it with chunks of the string representation, provided as a pointer to the first character and a length, and forward the user-defined data unmodified. It is up to the caller to allocate memory if the string representation must be stored and perform the copy. There is no guarantee that the pointer supplied to the callback points to a null-terminated string, the size argument should be used to find the end of the string. The callback may be called multiple times with consecutive chunks of the string representation (the printing itself is buffered).

Rationale: this approach allows the caller to have full control of the allocation and avoid unnecessary allocation and copying inside the printer.

For convenience, mlirXDump(MlirX) functions are provided to print the given object to the standard error stream.

Common Patterns 

The API adopts the following patterns for recurrent functionality in MLIR.

Indexed Components 

An object has an indexed component if it has fields accessible using a zero-based contiguous integer index, typically arrays. For example, an MlirBlock has its arguments as an indexed component. An object may have several such components. For example, an MlirOperation has attributes, operands, regions, results and successors.

For indexed components, the following pair of functions is provided.

  • intptr_t mlirXGetNum<Y>s(MlirX) returns the upper bound on the index.
  • MlirY mlirXGet<Y>(MlirX, intptr_t pos) returns ‘pos’-th subobject.

The sizes are accepted and returned as signed pointer-sized integers, i.e. intptr_t. This typedef is available in C99.

Note that the name of subobject in the function does not necessarily match the type of the subobject. For example, mlirOperationGetOperand returns an MlirValue.

Iterable Components 

An object has an iterable component if it has iterators accessing its fields in some order other than integer indexing, typically linked lists. For example, an MlirBlock has an iterable list of operations it contains. An object may have several iterable components.

For iterable components, the following triple of functions is provided.

  • MlirY mlirXGetFirst<Y>(MlirX) returns the first subobject in the list.
  • MlirY mlirYGetNextIn<X>(MlirY) returns the next subobject in the list that contains the given object, or a null object if the given object is the last in this list.
  • int mlirYIsNull(MlirY) returns 1 if the given object is null.

Note that the name of subobject in the function may or may not match its type.

This approach enables one to iterate as follows.

MlirY iter;
for (iter = mlirXGetFirst<Y>(x); !mlirYIsNull(iter);
     iter = mlirYGetNextIn<X>(iter)) {
  /* User 'iter'. */
}

Extending the API 

Extensions for Dialect Attributes and Types 

Dialect attributes and types can follow the example of builtin attributes and types, provided that implementations live in separate directories, i.e. include/mlir-c/<...>Dialect/ and lib/CAPI/<...>Dialect/. The core APIs provide implementation-private headers in include/mlir/CAPI/IR that allow one to convert between opaque C structures for core IR components and their C++ counterparts. wrap converts a C++ class into a C structure and unwrap does the inverse conversion. Once the C++ object is available, the API implementation should rely on isa to implement mlirXIsAY and is expected to use cast inside other API calls.

Extensions for Interfaces 

Interfaces can follow the example of IR interfaces and should be placed in the appropriate library (e.g., common interfaces in mlir-c/Interfaces and dialect-specific interfaces in their dialect library). Similarly to other type hierarchies, interfaces are not expected to have objects of their own type and instead operate on top-level objects: MlirAttribute, MlirOperation and MlirType. Static interface methods are expected to take as leading argument a canonical identifier of the class, MlirStringRef with the name for operations and MlirTypeID for attributes and types, followed by MlirContext in which the interfaces are registered.

Individual interfaces are expected provide a mlir<InterfaceName>TypeID() function that can be used to check whether an object or a class implements this interface using mlir<Attribute/Operation/Type>ImplementsInterface or mlir<Attribute/Operation?Type>ImplementsInterfaceStatic functions, respectively. Rationale: C++ isa only works when an object exists, static methods are usually dispatched to using templates; lookup by TypeID in MLIRContext works even without an object.