<table border="1" cellspacing="0" cellpadding="8">
    <tr>
        <th>Issue</th>
        <td>
            <a href=https://github.com/llvm/llvm-project/issues/134516>134516</a>
        </td>
    </tr>

    <tr>
        <th>Summary</th>
        <td>
            [mlir][vector] Error when running `--canonicalize` with out of bounds positions in `vector.extract`
        </td>
    </tr>

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

    <tr>
      <th>Assignees</th>
      <td>
            math-fehr
      </td>
    </tr>

    <tr>
      <th>Reporter</th>
      <td>
          math-fehr
      </td>
    </tr>
</table>

<pre>
    The following code
```
func.func @extract_poison_idx(%arg0: vector<16xf32>) -> f32 {
  %x = arith.constant -2 : index
  %0 = vector.extract %arg0[%x]: f32 from vector<16xf32>
  return %0 : f32
}
```
raises the following error when running `--canonicalize`
```
<source>:3:8: error: 'vector.extract' op expected position attribute #1 to be a non-negative integer smaller than the corresponding vector dimension or poison (-1)
  %0 = vector.extract %arg0[%x]: f32 from vector<16xf32>
       ^
```

This is due to `-1` being the only accepted static position that is out of bounds in a `vector.extract` or `vector.insert`.

There should be two ways to fix this:
* Either we allow out of bounds static positions that are not `-1` (which are also poison).
* Or the `vector.extract` builders should map out of bounds static positions to `-1`.
* Or it is the role of the users to make sure they do not pass non `-1` out of bounds static positions.

I feel both options come with a cost, and would personally prefer solution 1, as it follows LLVM semantics (`extractelement <4 x i32> %vec, i32 -2` is poison as well), and is roughly more modular. I feel solution 3 is exactly why we have this bug in the first place, and would be quite error-prone overall, because it is really easy to miss. Solution 2 is *okay* for me, but I feel following LLVM would make more sense here.

What do you think? I can write the patch, I just want to make sure we agree on a solution first before going into it. It also applies for `vector.insert` and `vector.shuffle` from what I see.
</pre>
<img width="1" height="1" alt="" src="http://email.email.llvm.org/o/eJysVUuP4rgW_jVmcwQKNoFiwYKuh1RSX93Fbd1ZthznhLjLsTM-NoH59aOThKarVJqZxSAEkXMe3_f5PDSRPXnEgyi_CCk7ndplg20UUoryaaFzakM8_DxeVKG-Hr61CE1wLgzWn8CEGkVxFNti_hbHJnuz4h8QmwIvKWqTvvfBUvDfbX0R8kHIUsdTIdQRzmhSiEI9rreXRkmhnoXcw1KoZ2iUBLH7IoojgJDlBYR6Ah1talcmeEraJ1hK4CjW13i5GRaj4RR4NeeHW0omWl5E-cRunKGJofsMBQeLmHL0t5ijPXPdPX1gHLUlJEjvlMEYQ4ShRQ8xe89HYlssl0b74K3Rzv6Bk_uvoYR6pJCjQcagjkqo4wOnHqPxg5C799SE3EHoAS89moQ19IFsssGDTinaKicEIdUaUoAKQYMPfunxpJM9I1if8IQRqNPOYYTUaj_yMCFGpD74moFPGaG2HXri2CHCdKMg5MNyLeT-31d__Ijy-aNCxfFbawksQZ2RabGsa7EtoEIGy_CDd1fQxmDPklDSyZq7MqnVif1DThAaqEL2NYH1oDnWB3m3BbO9n1tPGPl4dQODEYHakF3NCqchwKCvxMgae4HUWhKKLYU8wrNNLUYYEDQXygcIH4DShFRHBB_SnaeQD0NrTTu-0Y7CfBlC7ldznv_GUYdP6VTZuhoj3UB3uv9bHHeVf8lgRxU5TwwO2Z-fM3HsFKDTbwiUI_LxFeowsug1ERfhnc5f555lfoUG0UEVUguhn1CZ0CEMNrWgwQRKQj6C9jUMI60eIwWvnbtCH7HhKg8ujwWwHi2JCUwNS_D16___A4Sd9skaYonF9ja90GGHPoFQjxu4gB1rlIv6jIYjWSVhKZmKpVtfaIIBnePOmFFZghjyqXVX6EJE6EKdnY4rmKn9RKfYFC_aJHeFob1ytbT6jGMpQZVPXKrjsLGREvROG3xPvUL4PduE09hY9jF4hHDGqBnRI1RodCacLzDiKBJquo7XZolW8L8bGskmQh7Dm77yvTchQjfmq3K6Yb-PvVHHYa6rN5yoEnpC4EaZb_M3rus6wDVkZuXfhHqBVzDawxAZONPrdTItJ3qFH5kSDDzx39UVt9EpIvc76LuAky4VNpz7FBiW9SmATSt4TVPH6L53Fmmk80lzj2Lez6nNTeN4Xk8ja2D8r0CIq0V9UPVe7fUCD-vdRu2LdVGoRXt4MArXer-ulMENljtTVJvNfl9XuN5Iud0v7EEWsiw2xbYoyu1GrbApd1Wt1F42-91OK96fnbZu5dy5W4V4WliijIe12pTr7cLpCh3dlrezvB_mWTot8Hhgx2WVTyQ2hbOU6B4q2eTGzT96lk-i_DI7l0_w_M-W19R87_v3PjSs_3T-LHJ0hzalfhyL8kXIl5NNba5WJnRCvjDC-Y8r9wfyjnsZmZOQLzP580H-GQAA__8I4NmV">