Fail to legalize operation on math dialect

[WangJialei-A]

Please see the correctness check result at here:
https://github.com/intel/graph-compiler/actions/runs/10447798151/job/28927236931?pr=161

A reproducible module

module {
  func.func @entry(%arg0: tensor<5x6xf32>) -> tensor<5x6xf32> attributes {llvm.emit_c_interface} {
    %cst = arith.constant 0.000000e+00 : f32
    %0 = tensor.empty() : tensor<5x6xf32>
    %1 = linalg.fill ins(%cst : f32) outs(%0 : tensor<5x6xf32>) -> tensor<5x6xf32>
    %2 = linalg.abs ins(%arg0 : tensor<5x6xf32>) outs(%1 : tensor<5x6xf32>) -> tensor<5x6xf32>
    return %2 : tensor<5x6xf32>
  }
}

[xurui1995]

IR Dump After ConvertMemRefToCPURuntime Failed , @crazydemo could you take a look at this?

<unknown>:0: error: failed to legalize operation 'math.absf'
<unknown>:0: note: see current operation: %6 = "math.absf"(%5) <{fastmath = #arith.fastmath<none>}> : (f32) -> f32
// -----// IR Dump After ConvertMemRefToCPURuntime Failed (convert-memref-to-cpuruntime) //----- //
func.func @entry(%arg0: memref<5x6xf32>, %arg1: memref<5x6xf32>) attributes {llvm.emit_c_interface} {
  %cst = arith.constant 0.000000e+00 : f32
  %c0 = arith.constant 0 : index
  %c5 = arith.constant 5 : index
  %c1 = arith.constant 1 : index
  %c6 = arith.constant 6 : index
  scf.parallel (%arg2, %arg3) = (%c0, %c0) to (%c5, %c6) step (%c1, %c1) {
    memref.store %cst, %arg1[%arg2, %arg3] : memref<5x6xf32>
    scf.reduce 
  }
  scf.parallel (%arg2, %arg3) = (%c0, %c0) to (%c5, %c6) step (%c1, %c1) {
    %0 = memref.load %arg0[%arg2, %arg3] : memref<5x6xf32>
    %1 = math.absf %0 : f32
    memref.store %1, %arg1[%arg2, %arg3] : memref<5x6xf32>
    scf.reduce 
  }
  return
}

[WangJialei-A]

@crazydemo rsqrt is not working after your fix.
See the result here
https://github.com/intel/graph-compiler/actions/runs/10448587776/job/28929309658?pr=161

[crazydemo]

@crazydemo rsqrt is not working after your fix. See the result here https://github.com/intel/graph-compiler/actions/runs/10448587776/job/28929309658?pr=161

This issue is related to the upstream llvm. math::RsqrtOp is labled as illegal in mathtolibm pass. Will submit PR to LLVM community. Suggest to disable this test case for now.