MLIR

Multi-Level IR Compiler Framework

Creating a Dialect

Public dialects are typically separated into at least 3 directories:

  • mlir/include/mlir/Dialect/Foo (for public include files)
  • mlir/lib/Dialect/Foo (for sources)
  • mlir/lib/Dialect/Foo/IR (for operations)
  • mlir/lib/Dialect/Foo/Transforms (for transforms)
  • mlir/test/Dialect/Foo (for tests)

Along with other public headers, the ‘include’ directory contains a TableGen file in the ODS format, describing the operations in the dialect. This is used to generate operation declarations (FooOps.h.inc) and definitions (FooOps.cpp.inc) and operation interface declarations (FooOpsInterfaces.h.inc) and definitions (FooOpsInterfaces.cpp.inc).

The ‘IR’ directory typically contains implementations of functions for the dialect which are not automatically generated by ODS. These are typically defined in FooDialect.cpp, which includes FooOps.cpp.inc and FooOpsInterfaces.h.inc.

The ‘Transforms’ directory contains rewrite rules for the dialect, typically described in TableGen file using the DDR format.

Note that dialect names should not generally be suffixed with “Ops”, although some files pertaining only to the operations of a dialect (e.g. FooOps.cpp) might be.

CMake best practices 

TableGen Targets 

Operations in dialects are typically declared using the ODS format in tablegen in a file FooOps.td. This file forms the core of a dialect and is declared using add_mlir_dialect().

add_mlir_dialect(FooOps foo)
add_mlir_doc(FooOps FooDialect Dialects/ -gen-dialect-doc)

This generates the correct rules to run mlir-tblgen, along with a ‘MLIRFooOpsIncGen’ target which can be used to declare dependencies.

Dialect transformations are typically declared in a file FooTransforms.td. Targets for TableGen are described in typical llvm fashion.

set(LLVM_TARGET_DEFINITIONS FooTransforms.td)
mlir_tablegen(FooTransforms.h.inc -gen-rewriters)
add_public_tablegen_target(MLIRFooTransformIncGen)

The result is another ‘IncGen’ target, which runs mlir-tblgen.

Library Targets 

Dialects may have multiple libraries. Each library is typically declared with add_mlir_dialect_library(). Dialect libraries often depend on the generation of header files from TableGen (specified using the DEPENDS keyword). Dialect libraries may also depend on other dialect libraries. Typically this dependence is declared using target_link_libraries() and the PUBLIC keyword. For instance:

add_mlir_dialect_library(MLIRFoo
  DEPENDS
  MLIRFooOpsIncGen
  MLIRFooTransformsIncGen

  LINK_COMPONENTS
  Core

  LINK_LIBS PUBLIC
  MLIRBar
  <some-other-library>
  )

add_mlir_dialect_library() is a thin wrapper around add_llvm_library() which collects a list of all the dialect libraries. This list is often useful for linking tools (e.g. mlir-opt) which should have access to all dialects. This list is also linked into libMLIR.so. The list can be retrieved from the MLIR_DIALECT_LIBS global property:

get_property(dialect_libs GLOBAL PROPERTY MLIR_DIALECT_LIBS)

Note that although the Bar dialect also uses TableGen to declare its operations, it is not necessary to explicitly depend on the corresponding IncGen targets. The PUBLIC link dependency is sufficient. Also note that we avoid using add_dependencies explicitly, since the dependencies need to be available to the underlying add_llvm_library() call, allowing it to correctly create new targets with the same sources. However, dialects that depend on LLVM IR may need to depend on the LLVM ‘intrinsics_gen’ target to ensure that tablegen’d LLVM header files have been generated.

In addition, linkage to MLIR libraries is specified using the LINK_LIBS descriptor and linkage to LLVM libraries is specified using the LINK_COMPONENTS descriptor. This allows cmake infrastructure to generate new library targets with correct linkage, in particular, when BUILD_SHARED_LIBS=on or LLVM_LINK_LLVM_DYLIB=on are specified.

Dialect Conversions

Conversions from “X” to “Y” live in mlir/include/mlir/Conversion/XToY, mlir/lib/Conversion/XToY and mlir/test/Conversion/XToY, respectively.

Default file names for conversion should omit “Convert” from their name, e.g. lib/VectorToLLVM/VectorToLLVM.cpp.

Conversion passes should live separately from conversions themselves for convenience of users that only care about a pass and not about its implementation with patterns or other infrastructure. For example include/mlir/VectorToLLVM/VectorToLLVMPass.h.

Common conversion functionality from or to dialect “X” that does not belong to the dialect definition can be located in mlir/lib/Conversion/XCommon, for example mlir/lib/Conversion/GPUCommon.

CMake best practices 

Each conversion typically exists in a separate library, declared with add_mlir_conversion_library(). Conversion libraries typically depend on their source and target dialects, but may also depend on other dialects (e.g. MLIRFunc). Typically this dependence is specified using target_link_libraries() and the PUBLIC keyword. For instance:

add_mlir_conversion_library(MLIRBarToFoo
  BarToFoo.cpp

  ADDITIONAL_HEADER_DIRS
  ${MLIR_MAIN_INCLUDE_DIR}/mlir/Conversion/BarToFoo

  LINK_LIBS PUBLIC
  MLIRBar
  MLIRFoo
  )

add_mlir_conversion_library() is a thin wrapper around add_llvm_library() which collects a list of all the conversion libraries. This list is often useful for linking tools (e.g. mlir-opt) which should have access to all dialects. This list is also linked in libMLIR.so. The list can be retrieved from the MLIR_CONVERSION_LIBS global property:

get_property(dialect_libs GLOBAL PROPERTY MLIR_CONVERSION_LIBS)

Note that it is only necessary to specify a PUBLIC dependence against dialects to generate compile-time and link-time dependencies, and it is not necessary to explicitly depend on the dialects’ IncGen targets. However, conversions that directly include LLVM IR header files may need to depend on the LLVM ‘intrinsics_gen’ target to ensure that tablegen’d LLVM header files have been generated.