Reference: https://mlir.llvm.org/docs/Tutorials/Toy/Ch-4/

Note: Check Setup the Environment of MLIR for the environment setup.

1. Introduction

Problem: Naively implementing each transformation for each dialect leads to large amounts of code duplication, as the internal algorithms are generally very similar.

Solution: To provide the ability for transformations to opaquely hook into dialects like Toy to get the information they need.

2. Add OpPrintInterface

Define a new env var:

export TOY_CH4_HOME="$MLIR_HOME/examples/toy/Ch4"

2.1. Define OpPrintInterface

First, create a new file $TOY_CH4_HOME/include/toy/OpPrintInterface.td, define OpPrintOpInterface with method opPrint which returns a std::string:

#ifndef PRINT_INTERFACE
#define PRINT_INTERFACE

include "mlir/IR/OpBase.td"

def OpPrintOpInterface : OpInterface<"OpPrint">
{
    let description = [{
        Interface to print something in an operator.
    }];

    let methods = [
        InterfaceMethod< "Print some information in the current operation", "std::string", "opPrint" >
    ];
}


#endif // PRINT_INTERFACE

Second, create a file $TOY_CH4_HOME/include/toy/OpPrintInterface.hpp:

#ifndef OPPRINTINTERFACE_HPP_
#define OPPRINTINTERFACE_HPP_

#include "mlir/IR/OpDefinition.h"

namespace mlir::toy {

/// Include the auto-generated declarations.
#include "toy/OpPrintOpInterface.h.inc"

} // namespace mlir::toy


#endif

Third, in $TOY_CH4_HOME/include/toy/Dialect.h, include the new interface:

#include "toy/OpPrintInterface.hpp"

Fourth, make some modifications in $TOY_CH4_HOME/include/toy/Ops.td. Include the interface’s td at the beginning of the file:

include "toy/OpPrintInterface.td"

Then, for example, change the AddOp to declare that it implements the OpPrint interface:

def AddOp : Toy_Op<"add", [
    Pure, 
    DeclareOpInterfaceMethods<ShapeInferenceOpInterface>,
    DeclareOpInterfaceMethods<OpPrintOpInterface>
  ]> {
  // ...
}

You can also do the similar declaration for other operations.

Finally, some CMakeLists need to be modified.

Add following lines in $TOY_CH4_HOME/include/toy/CMakeLists.txt:

set(LLVM_TARGET_DEFINITIONS OpPrintInterface.td)
mlir_tablegen(OpPrintOpInterface.h.inc -gen-op-interface-decls)
mlir_tablegen(OpPrintOpInterface.cpp.inc -gen-op-interface-defs)
add_public_tablegen_target(ToyCh4OpPrintInterfaceIncGen)

Change the add_toy_chapter in $TOY_CH4_HOME/CMakelists.txt to:

add_toy_chapter(toyc-ch4
  toyc.cpp
  parser/AST.cpp
  mlir/MLIRGen.cpp
  mlir/Dialect.cpp
  mlir/ShapeInferencePass.cpp
  mlir/OpPrintInterfacePass.cpp
  mlir/ToyCombine.cpp

  DEPENDS
  ToyCh4OpsIncGen
  ToyCh4ShapeInferenceInterfaceIncGen
  ToyCh4CombineIncGen
  ToyCh4OpPrintInterfaceIncGen
  )

To match the listed source files, create a blank file $TOY_CH4_HOME/mlir/OpPrintInterfacePass.cpp. We will implement the pass later.

Now build the MLIR:

bash $LLVM_PROJ_HOME/scripts/build-mlir.sh

Errors pop out because we haven’t implemented the OpPrintInterface which is declared in AddOp. Don’t worry, it will be implemented in the next section.

Now you can check the generated C++ class declarations in, for example, $LLVM_PROJ_HOME/build/tools/mlir/examples/toy/Ch4/include/toy/OpPrintOpInterface.h.inc.

2.3. Implement OpPrintInterface

Since the interface is declared in AddOp (which is actually implemented by inheriting OpPrintOpInterface which provides a pure virtual function opPrint), we need to implement the function.

In $TOY_CH4_HOME/mlir/Dialect.cpp, add the following code:

std::string AddOp::opPrint() { return "I am AddOp"; }

2.4. Implement OpPrintPass

In $TOY_CH4_HOME/include/toy/Passes.h, add the following line under mlir/examples/toy/Ch4/include/toy/Passes.h (inside namespace mlir::toy):

std::unique_ptr<Pass> createOpPrintPass();

In $TOY_CH4_HOME/mlir/OpPrintInterfacePass.cpp, add the following code:

#include "mlir/IR/Operation.h"
#include "mlir/Pass/Pass.h"
#include "mlir/Support/LLVM.h"
#include "mlir/Support/TypeID.h"
#include "toy/Dialect.h"
#include "toy/OpPrintInterface.hpp"
#include "toy/Passes.h"
#include <iostream>
#include <memory>

using namespace mlir;
using namespace toy;

#include "toy/OpPrintOpInterface.cpp.inc"

namespace
{
struct OpPrintIntercfacePass
    : public mlir::PassWrapper<OpPrintIntercfacePass, OperationPass<toy::FuncOp>>
{
    MLIR_DEFINE_EXPLICIT_INTERNAL_INLINE_TYPE_ID(OpPrintIntercfacePass)

    void runOnOperation() override
    {
        auto f = getOperation();
        f.walk([&](mlir::Operation* op) {
            if (auto shapeOp = dyn_cast<OpPrint>(op)) {
                std::cout << shapeOp.opPrint() << std::endl;
            }
        });
    }
};

}  // namespace

std::unique_ptr<Pass> createOpPrintPass()
{
    return std::make_unique<OpPrintIntercfacePass>();
}

2.5. Add Pass to Pass Manager

In $TOY_CH4_HOME/toyc.cpp, add the following line after mlir::OpPassManager &optPM = pm.nest<mlir::toy::FuncOp>();:

optPM.addPass(mlir::toy::createOpPrintPass());

Implement OpPrintPass in $TOY_CH4_HOME/mlir/OpPrintInterfacePass.cpp:

#include "mlir/IR/Operation.h"
#include "mlir/Pass/Pass.h"
#include "mlir/Support/LLVM.h"
#include "mlir/Support/TypeID.h"
#include "toy/Dialect.h"
#include "toy/OpPrintInterface.hpp"
#include "toy/Passes.h"
#include <iostream>
#include <memory>

using namespace mlir;
using namespace toy;

#include "toy/OpPrintOpInterface.cpp.inc"

namespace
{
struct OpPrintIntercfacePass
    : public mlir::PassWrapper<OpPrintIntercfacePass, OperationPass<toy::FuncOp>>
{
    MLIR_DEFINE_EXPLICIT_INTERNAL_INLINE_TYPE_ID(OpPrintIntercfacePass)

    void runOnOperation() override
    {
        auto f = getOperation();
        f.walk([&](mlir::Operation* op) {
            if (auto shapeOp = dyn_cast<OpPrint>(op)) {
                std::cout << shapeOp.opPrint() << std::endl;
            }
        });
    }
};

}  // namespace

std::unique_ptr<Pass> mlir::toy::createOpPrintPass()
{
    return std::make_unique<OpPrintIntercfacePass>();
}

2.6. Test the Pass

Now, rebuild the MLIR:

bash $LLVM_PROJ_HOME/scripts/build-mlir.sh

Run the Test:

toyc-ch4 $MLIR_HOME/test/Examples/Toy/Ch4/codegen.toy -emit=mlir -opt