[mlir] SEGV when converting arith op to llvm

[YuanchengJiang]

#CSR = #sparse_tensor.encoding<{map = (d0, d1) -> (d0 : dense, d1 : compressed)}>

module {
  func.func @crash_target(%arg0: tensor<10x20xf16, #CSR>) {
    // The pass attempts to convert this arith op and potentially reconciles 
    // the function signature, crashing on the unhandled sparse tensor type.
    %0 = arith.constant 0.0 : f64
    return
  }
}

reproduce: mlir-opt -pass-pipeline="builtin.module(func.func(convert-arith-to-llvm))" ./test.mlir

Segmentation fault (core dumped)

[llvmbot]

@llvm/issue-subscribers-mlir

Author: Yuancheng Jiang (YuanchengJiang)

Details

```mlir #CSR = #sparse_tensor.encoding<{map = (d0, d1) -> (d0 : dense, d1 : compressed)}>

module {
func.func @crash_target(%arg0: tensor<10x20xf16, #CSR>) {
// The pass attempts to convert this arith op and potentially reconciles
// the function signature, crashing on the unhandled sparse tensor type.
%0 = arith.constant 0.0 : f64
return
}
}

reproduce: `mlir-opt -pass-pipeline="builtin.module(func.func(convert-arith-to-llvm))" ./test.mlir`

Segmentation fault (core dumped)

</details>

[linuxlonelyeagle]

The latest MLIR has fixed it, and I haven't encountered this bug on my machine.

[CoTinker]

Fixed in latest version.