Lower op with tensor type to llvm ir

Beginners
, ,
1

Hi all, I’m new here, studying how to convert mlir to llvm-ir(and finally, to machine code); Soon i met the problem:

A.mlir: 
module {
  func.func public @main(%arg0: tensor<f32> ) -> (tensor<f32>) {
    %0 = arith.mulf %arg0, %arg0 : tensor<f32>
    return %0 : tensor<f32>
  }
}
B.mlir:
module {
  func.func public @main(%arg0: f32 ) -> (f32) {
    %0 = arith.mulf %arg0, %arg0 : f32
    return %0 : f32
  }
}

When i execute: mlir-opt A.mlir --convert-to-llvm -o llvm.ir, nothing changed in the output file, while convertting B.mlir successfully emits llvm-ir:

module {
  llvm.func @main(%arg0: f32) -> f32 {
    %0 = llvm.fmul %arg0, %arg0  : f32
    llvm.return %0 : f32
  }
}

I wonder what makes ‘tensor’ type special ? Because it has variadic size so we should be aware of tensor alloc/dealloc/ownership?
If so, how can I generate llvm-ir by specifying the mlir-opt command line flags?
And what is the function signature if I want to call this function from “C” ?

Thanks a lot!

2

i’m new too, tensor dialect should be lower to memref by -one-shot-bufferize pass, i think.
some lowering pass don’t named as “convert-tensor-to-llvm”, which is sometimes hard to discover.