[polybench] kernel symm raising failure

[avik-pal]

module @reactant_kernel_... attributes {mhlo.num_partitions = 1 : i64, mhlo.num_replicas = 1 : i64} {
  func.func @main(%arg0: tensor<f32> {enzymexla.memory_effects = []}, %arg1: tensor<f32> {enzymexla.memory_effects = []}, %arg2: tensor<64x64xf32> {enzymexla.memory_effects = []}, %arg3: tensor<32x64xf32> {enzymexla.memory_effects = []}) -> tensor<32x64xf32> attributes {enzymexla.memory_effects = []} {
    %c = stablehlo.constant dense<0> : tensor<i32>
    %c_0 = stablehlo.constant dense<"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tensor<64x2xi32>
    %c_1 = stablehlo.constant dense<64> : tensor<i64>
    %cst = stablehlo.constant dense<0.000000e+00> : tensor<32x64xf32>
    %c_2 = stablehlo.constant dense<1> : tensor<i32>
    %c_3 = stablehlo.constant dense<0> : tensor<i64>
    %c_4 = stablehlo.constant dense<1> : tensor<i64>
    %0 = stablehlo.transpose %arg2, dims = [1, 0] : (tensor<64x64xf32>) -> tensor<64x64xf32>
    %1 = stablehlo.transpose %arg3, dims = [1, 0] : (tensor<32x64xf32>) -> tensor<64x32xf32>
    %2 = stablehlo.broadcast_in_dim %arg2, dims = [2, 0] : (tensor<64x64xf32>) -> tensor<64x32x64x1xf32>
    %3 = stablehlo.dot_general %arg2, %arg3, contracting_dims = [0] x [1] : (tensor<64x64xf32>, tensor<32x64xf32>) -> tensor<64x32xf32>
    %4 = stablehlo.broadcast_in_dim %arg0, dims = [] : (tensor<f32>) -> tensor<64x32xf32>
    %5 = stablehlo.multiply %4, %1 : tensor<64x32xf32>
    %6 = "stablehlo.gather"(%0, %c_0) <{dimension_numbers = #stablehlo.gather<collapsed_slice_dims = [0, 1], start_index_map = [0, 1], index_vector_dim = 1>, indices_are_sorted = false, slice_sizes = array<i64: 1, 1>}> : (tensor<64x64xf32>, tensor<64x2xi32>) -> tensor<64xf32>
    %7 = stablehlo.broadcast_in_dim %6, dims = [0] : (tensor<64xf32>) -> tensor<64x32xf32>
    %8 = stablehlo.multiply %5, %7 : tensor<64x32xf32>
    %9 = stablehlo.broadcast_in_dim %5, dims = [0, 1] : (tensor<64x32xf32>) -> tensor<64x32x64x1xf32>
    %10 = stablehlo.broadcast_in_dim %arg0, dims = [] : (tensor<f32>) -> tensor<64x32x1x1xf32>
    %11 = stablehlo.reshape %3 : (tensor<64x32xf32>) -> tensor<64x32x1x1xf32>
    %12 = stablehlo.multiply %10, %11 : tensor<64x32x1x1xf32>
    %13 = stablehlo.multiply %9, %2 : tensor<64x32x64x1xf32>
    %14 = stablehlo.broadcast_in_dim %arg1, dims = [] : (tensor<f32>) -> tensor<32x1x1xf32>
    %15:2 = stablehlo.while(%iterArg = %c_3, %iterArg_5 = %cst) : tensor<i64>, tensor<32x64xf32> attributes {enzyme.disable_mincut, enzymexla.symmetric_matrix = [#enzymexla<guaranteed UNKNOWN>, #enzymexla<guaranteed UNKNOWN>, #enzymexla<guaranteed UNKNOWN>, #enzymexla<guaranteed UNKNOWN>, #enzymexla<guaranteed UNKNOWN>, #enzymexla<guaranteed UNKNOWN>, #enzymexla<guaranteed NOTGUARANTEED>, #enzymexla<guaranteed UNKNOWN>, #enzymexla<guaranteed UNKNOWN>]}
    cond {
      %16 = stablehlo.compare  LT, %iterArg, %c_1 : (tensor<i64>, tensor<i64>) -> tensor<i1>
      stablehlo.return %16 : tensor<i1>
    } do {
      %16 = stablehlo.add %c_4, %iterArg {enzymexla.bounds = [[1, 64]]} : tensor<i64>
      %17 = stablehlo.convert %16 {enzymexla.bounds = [[1, 64]]} : (tensor<i64>) -> tensor<i32>
      %18 = stablehlo.subtract %17, %c_2 {enzymexla.bounds = [[0, 63]]} : tensor<i32>
      %19 = stablehlo.reshape %iterArg_5 : (tensor<32x64xf32>) -> tensor<32x64x1xf32>
      %20 = stablehlo.dynamic_slice %13, %iterArg, %c_3, %c_3, %c_3, sizes = [1, 32, 64, 1] : (tensor<64x32x64x1xf32>, tensor<i64>, tensor<i64>, tensor<i64>, tensor<i64>) -> tensor<1x32x64x1xf32>
      %21 = stablehlo.reshape %20 : (tensor<1x32x64x1xf32>) -> tensor<32x64x1xf32>
      %22 = stablehlo.add %19, %21 : tensor<32x64x1xf32>
      %23 = stablehlo.convert %18 {enzymexla.bounds = [[0, 63]]} : (tensor<i32>) -> tensor<i64>
      %24 = stablehlo.dynamic_slice %22, %c_3, %23, %c_3, sizes = [32, 1, 1] : (tensor<32x64x1xf32>, tensor<i64>, tensor<i64>, tensor<i64>) -> tensor<32x1x1xf32>
      %25 = stablehlo.multiply %14, %24 : tensor<32x1x1xf32>
      %26 = stablehlo.dynamic_slice %8, %iterArg, %c_3, sizes = [1, 32] : (tensor<64x32xf32>, tensor<i64>, tensor<i64>) -> tensor<1x32xf32>
      %27 = stablehlo.reshape %26 : (tensor<1x32xf32>) -> tensor<32x1x1xf32>
      %28 = stablehlo.add %25, %27 : tensor<32x1x1xf32>
      %29 = stablehlo.dynamic_slice %12, %iterArg, %c_3, %c_3, %c_3, sizes = [1, 32, 1, 1] : (tensor<64x32x1x1xf32>, tensor<i64>, tensor<i64>, tensor<i64>, tensor<i64>) -> tensor<1x32x1x1xf32>
      %30 = stablehlo.reshape %29 : (tensor<1x32x1x1xf32>) -> tensor<32x1x1xf32>
      %31 = stablehlo.add %28, %30 : tensor<32x1x1xf32>
      %32 = stablehlo.reshape %22 : (tensor<32x64x1xf32>) -> tensor<32x1x64xf32>
      %33 = stablehlo.dynamic_update_slice %32, %31, %c, %c, %18 : (tensor<32x1x64xf32>, tensor<32x1x1xf32>, tensor<i32>, tensor<i32>, tensor<i32>) -> tensor<32x1x64xf32>
      %34 = stablehlo.reshape %33 : (tensor<32x1x64xf32>) -> tensor<32x64xf32>
      stablehlo.return %16, %34 : tensor<i64>, tensor<32x64xf32>
    }
    return %15#1 : tensor<32x64xf32>
  }
}

we are able to raise the 2 inner loops. the outer loop remains