Backend for generating RVSDG using the MLIR dialect

[sjalander]

Initial attempt of creating a backend for jlm-opt that outputs rvsdg as MLIR IR.

The code currently assumes that the MLIR implementation of RVSDG resides in the same root directory as the main jlm directory, i.e.,
/
jlm
mlir-rvsdg

The dialect can be cloned with
git clone [email protected]:Riphiphip/mlir_rvsdg.git mlir-rvsdg
And built by
cd mlir-rvsdg
mkdir build
cd build
cmake -G Ninja ..
-DCMAKE_C_COMPILER=clang-16
-DCMAKE_CXX_COMPILER=clang++-16
-DLLVM_DIR=/usr/lib/llvm-16/cmake/
-DMLIR_DIR=/usr/lib/llvm-16/lib/cmake/mlir
cmake --build .

At this point, it should be possible to build jlm as normal.

The backend currently only supports lambda and arithmetic constants, so the test cases are limited.
Currently, the only code that has been used to test the backend is:

test.c :
int test() {
return 1;
}

The backend can be tested by:
/usr/lib/llvm-16/bin/clang -S -emit-llvm -o test.ll test.c
jlm-opt --mlir --dne test.ll

This should print the following MLIR IR:

rvsdg.omegaNode (): {
%0 = rvsdg.lambdaNode <(!rvsdg.ioState, !rvsdg.memState, !rvsdg.loopState) -> (i32, !rvsdg.ioState, !rvsdg.memState, !rvsdg.loopState)> () : (%arg0: !rvsdg.ioState, %arg1: !rvsdg.memState, %arg2: !rvsdg.loopState): {
%c1_i32 = arith.constant 1 : i32
rvsdg.lambdaResult (%c1_i32: i32, %arg0: !rvsdg.ioState, %arg1: !rvsdg.memState, %arg2: !rvsdg.loopState)
}
rvsdg.omegaResult (%0: !rvsdg.lambdaRef<(!rvsdg.ioState, !rvsdg.memState, !rvsdg.loopState) -> (i32, !rvsdg.ioState, !rvsdg.memState, !rvsdg.loopState)>)
}

[sjalander]

When creating an MLIR backend and frontend, we need to be able to specify format for both the input and output. A question is how to handle the naming of these. The current version uses the following, non-optimal, names.

Select input format
--input-llvm - Input LLVM IR [default]
--input-mlir - Input MLIR
Select output format
--llvm - Output LLVM IR [default]
--xml - Output XML
--mlir - Output MLIR