<table border="1" cellspacing="0" cellpadding="8">
    <tr>
        <th>Issue</th>
        <td>
            <a href=https://github.com/llvm/llvm-project/issues/143335>143335</a>
        </td>
    </tr>

    <tr>
        <th>Summary</th>
        <td>
            [MLIR] Inconsistent output when executing MLIR program with and without `-test-math-polynomial-approximation="enable-avx2"`
        </td>
    </tr>

    <tr>
      <th>Labels</th>
      <td>
            mlir
      </td>
    </tr>

    <tr>
      <th>Assignees</th>
      <td>
      </td>
    </tr>

    <tr>
      <th>Reporter</th>
      <td>
          Lambor24
      </td>
    </tr>
</table>

<pre>
    My git version is [90beda2](https://github.com/llvm/llvm-project/commit/90beda2aba3cac34052827c560449fcb184c7313).

## Description:
I am experiencing an inconsistent result when executing the same MLIR program with and without the `-test-math-polynomial-approximation="enable-avx2"`.

## Steps to Reproduce:

### 1. **MLIR Program (test.mlir)**:

test.mlir:

```
module attributes {llvm.data_layout = ""} {
  func.func private @printMemrefI64(tensor<*xi64>)
  func.func private @printMemrefF32(tensor<*xf32>)
  func.func @main() {
    %0 = "tosa.const"() <{values = dense<0.000000e+00> : tensor<1xf32>}> : () -> tensor<1xf32>
    %1 = "tosa.const"() <{values = dense<-1.023950e+03> : tensor<1x1x6x3xf32>}> : () -> tensor<1x1x6x3xf32>
    %2 = "tosa.const"() <{values = dense<3.671420e+03> : tensor<4x4x3x1xf32>}> : () -> tensor<4x4x3x1xf32>
    %3 = "tosa.const"() <{values = dense<4.235330e+03> : tensor<3xf32>}> : () -> tensor<3xf32>
    %4 = tosa.depthwise_conv2d %1, %2, %3, %0, %0 {acc_type = f32, dilation = array<i64: 1, 2>, pad = array<i64: 1, 2, 2, 2>, stride = array<i64: 2, 1>} : (tensor<1x1x6x3xf32>, tensor<4x4x3x1xf32>, tensor<3xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<1x1x4x3xf32>
    %5 = tosa.cos %4 : (tensor<1x1x4x3xf32>) -> tensor<1x1x4x3xf32>
    %cast = tensor.cast %5 : tensor<1x1x4x3xf32> to tensor<*xf32>
    call @printMemrefF32(%cast) : (tensor<*xf32>) -> ()
 return
  }
}
```

### 2. **Command to Run Without `-test-math-polynomial-approximation="enable-avx2"`:**

```
/path/llvm-project/build/bin/mlir-opt test.mlir -pass-pipeline='builtin.module(func.func(tosa-to-linalg-named,tosa-to-linalg))' | \
/path/llvm-project/build/bin/mlir-opt -tosa-to-arith -one-shot-bufferize="bufferize-function-boundaries" -convert-linalg-to-affine-loops -test-lower-to-llvm | \
/path/llvm-project/build/bin/mlir-runner -e main -entry-point-result=void \
-shared-libs=/path/llvm-project/build/lib/libmlir_runner_utils.so \
-shared-libs=/path/llvm-project/build/lib/libmlir_c_runner_utils.so \
-shared-libs=/path/llvm-project/build/lib/libmlir_async_runtime.so
```

### 3. **Output Without `-test-math-polynomial-approximation="enable-avx2"`:**

```
[[[[-0.290655, -0.290655,     -0.290655], 
   [-0.290655,     -0.290655, -0.290655], 
   [-0.290655,     -0.290655,     -0.290655], 
   [-0.290655, -0.290655,     -0.290655]]]]
```

### 4. **Command to Run With `-test-math-polynomial-approximation="enable-avx2"`:**

```
/path/llvm-project/build/bin/mlir-opt test.mlir -pass-pipeline='builtin.module(func.func(tosa-to-linalg-named,tosa-to-linalg))' | \
/path/llvm-project/build/bin/mlir-opt -tosa-to-arith -one-shot-bufferize="bufferize-function-boundaries" -convert-linalg-to-affine-loops -test-math-polynomial-approximation="enable-avx2" -test-lower-to-llvm | \
/path/llvm-project/build/bin/mlir-runner -e main -entry-point-result=void \
-shared-libs=/path/llvm-project/build/lib/libmlir_runner_utils.so \
-shared-libs=/path/llvm-project/build/lib/libmlir_c_runner_utils.so \
-shared-libs=/path/llvm-project/build/lib/libmlir_async_runtime.so
```

### 5. **Output With `-test-math-polynomial-approximation="enable-avx2"`:**

```
[[[[-0.841471, -0.841471,     -0.841471], 
   [-0.841471,     -0.841471, -0.841471], 
   [-0.841471,     -0.841471,     -0.841471], 
   [-0.841471, -0.841471,     -0.841471]]]]
```

I'm not sure if there is any bug in my program or if the wrong usage of the above passes caused this result.
</pre>
<img width="1" height="1" alt="" src="http://email.email.llvm.org/o/eJzsWFFvozgQ_jXOywhkbAjJQx7SZCNV2upOew_3WBmYBJ_ARrZJk_v1JxuSNm3aa7u3p33YCEEw428-j4dv0Ahr5U4hLkh2Q7L1RPSu1mbxVbSFNiydFLo6Lu6OsJMO9mis1AqkBZLdzGmBlWAkWxM2q53rLOFLwjaEbXbS1X0Rl7olbNM0-9Ml6oz-C0tH2KbUbSv9nxFGFIKXouQpzdiM5WU2pWk635ZFMkvLnCecsHlM6NIfjBPGYY22NLJzUivvmC5vQbSAhw6NRFVKtQOhQKpSKyutQ-XAoO0bBw81KsADlr3zVq5GsKJFuPt6-w06o3dGtPAgXQ1CVeGP7l0wI1MaObQuaoWro043R6VbKZpIdJ3RB9mKgc6aMIZKFA1GYn9ghDEypZf0_3DYWXAavmFndNWXOKzibOGNkhgIWxK2DNR-H6kRNvMk4raRhrD5YHGe_fjojDel40GXra76BkE4Z2TRO7RA8hu_N3ElnLhvxNGvlfA1eNKMkXztLQhdAmx7Vcb-BJ2Re-EQSEo7I5W7w9bg9naaBmrKakP4irDlQU5Twr94ku8B2HD2HGDL2VUAktJWSEXYjLD5mSEAYRk90XfaithvvwtrGSz5iuQ3e9H0ful8DRUqi4SvaEzDDwm7oZTwL0D4Es5ckhOTfH16NiJG_v6l3ZlO8ik6URJTxufZwIdf4ZMcpgf-blYX1mdu7FPceDzNk5S9Ri09pAd-eG_AnlmfqfFPUUtjxjPOX6P23ni9jFQa3AQuFXaufpAW70ut9qwKm0zYKgR0vPLxSk9Xn6KiLO_dscMA5R2wFVSyCZoRxoQx4kj4Krw1SwigQ_qvoBPVGzaPp8HaOiMrvDYhmCVDDE4BeC1P2OrVfXr66Kr9VcvHwWsZml7J0Owx7qW2p514yfrJ3Hdjl8IOUjdYxsP94HP52nQv2dcUakAtRdNcFbXR3ZC8l_QvZG6gPiSlxzToeqMCuk9auhzPTyT9omKwU8VY6bb15ctXmF7Bn2MV-64KFgp8OF7UFcI2nXD1y0Jf9LKp_NWL9cbXpUh3Ds5VCqJOWBt1ssNGKgyecz_JSRUP1Yqw2Vn4feC0FZHTUSOVaHaREi1WhK0uh0NdnBOWA8lXQLLVJzhGJ0hh_MdApBVGttYuKvrtFo38e2DLzreRZ-jjFxW6V5UwEi1hDCIvFGjcibKH3G6lwqjRurMw7EijH9CEFTT79pO0Ta8UGogQfHWECJUzx6jTUrlo-PohfL3XshqxI1sLg1XUyMKGxfyLq0YWw9l7ux-83fdONja2-j_BLH8IqrBHFZCdbDG2-q1XiJ9eod961_Xux786_st7PCIaszmdZpmXzYsb_3sc8B_dKxhE5_m0S8sLnA9Ne7_Ht6edjzdinr4hW7806-fVrA_vyC-t-6m0Lruidf-X0M3SJM2TUT8eb0b9GAdeys5rlhc4H5r2fo9vT3td6G4Jy1tQ2oHtDYLcgqvR_7Eg1BGKfgdSQXs8dx-0GY3gwWi1g96KHYIehkSh9whegtBCKXqLFbha2rG_EU-qBa_mfC4muEjydJ7y2Ywnk3pBeZJvk6yiWGGSI1ZVkWO6TbKi3ApapRO5YJRldErnlCc8S-Ii47NtJvJ5luYpTWYkpdgK2cShZaDNbiKt7XGRpJzzbNKIAhsbekmMDe0JRrL1xCxC5hb9zpKUNtI6-4jgpGtC_-nu6-03kq3h9mnPRg-J-axn83ar5rsSeNKbZvHhflYIgyVsM0Ziv2D_BAAA__9-_7TY">