<table border="1" cellspacing="0" cellpadding="8">
    <tr>
        <th>Issue</th>
        <td>
            <a href=https://github.com/llvm/llvm-project/issues/77748>77748</a>
        </td>
    </tr>

    <tr>
        <th>Summary</th>
        <td>
            Miscompile of dav1d on ppc64le
        </td>
    </tr>

    <tr>
      <th>Labels</th>
      <td>
            backend:PowerPC,
            miscompilation
      </td>
    </tr>

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

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

<pre>
    The checkasm test of https://code.videolan.org/videolan/dav1d fails on ppc64le when built with `meson --build-type=plain`.

The issue bisects to 1a2e77cf9e11dbf56b5720c607313a566eebb16e, which presumably just exposes the issue.

Using opt-bisect-limit and debug-counter=instcombine-visit-count we can narrow this down to an InstCombine transform. However, I believe the transform itself is correct (it just converts some extractelement/insertelement into a shufflevector) and this just exposes an issue in the backend.

The before.ll and after.ll in https://gist.github.com/nikic/b4c015523bcd22ea159d183c0648deea are before and after the transform.

The llc output for these two files is before.s and after.s in https://gist.github.com/nikic/ea9909b8216726762c8f2f277abc8006. (I've run `llc` with `opt-bisect-limit=0` to mostly isolate isel only.)

The difference is in two vperms and corresponding constant pool shuffle indices. 
</pre>
<img width="1px" height="1px" alt="" src="http://email.email.llvm.org/o/eJyUlM1yozgQx59GvnSZEsKAOXDIxuXaHLZqDrsPoI_GaCIkSt3Yk7ffAseTSeY0Fyh9df9a_79aE_lLROxF_ZeoTzu98JhyH_2rtzuT3Fv_74hgR7SvmiZgJIY0wMg8k6iehDoLdbbJYXH1DlPQsUj5ItT5MRTq7PS1dDBoHwhShHm2zSEg3EaMYBYfGG6eRxCNnJBShP1-nXV7fptRVKc5aB9FIwshT0I-3b8rlSdaEIwntEzACUqtsG3t0GFZOjPUjalbJW0j26qsdN00iMaUDQr1DLfR2xHmjLRM2oQ3-L4QA_6YEyEBP8J_Svof-XiBNPP-nnQf_OQZdHTg0CyXvU1LZMyiOvlIbNNkfMT91ZPn-xrcEKyOEHXO6QY8egKXbnGl1xFeIvHz_RRw1pGGlKcC_k43vGJesV_AYPB4xQ3x5x7wTBgG8AQ25YyWQaij53tVNsUrZiagNCHgD87aMgacMLJQZx8J82MMPq4sQOMyDAGvaDllobqtyo3300Xp-C6DjxuR0fYVo_tNK4NDyliEsMXRA2NeBz5-sdLFExcXz-NiCpsmoc6bFYU6m4OVZV2rylinFOqy7lx5rKxsDkeHqEHnR5qPHJ9v6TeqECykheeFYUjbZkLgW4LBB6T1Nt-56Rds-hNq1F0nO3NUZdOqpm2UPQ5qUG2rjT1K2RSrTi9CtVeEvMT1EYRgRSN_vomvdhPVSa7rnGBKxOENPKWgeTUsBkgxvBVCdV9LdX4YMGO0675NrVuC64x5ute22YbmFN3qcZsisY4Mc0rh4QXw0XmLVMDO9ZXrqk7vsC9beeiarlLVbuxN3dalbQ7loam1VF1tbN0O0jWVtEbrZud7JdVBlmVZyqpWVXG0WFlXdqbqnOzqShwkTtqHIoTrtPaS3eavvm3bw3EXtMFAW7NS6t1ronr6lm6Yvz0LpYRav5Mnm6bZB80-xXW6Pu1yv4bcm-VC4iCDJ6aPJOw5YP_P4xyuTe7etz461m7Jof8q_C-ar8Hef_s5p-9ot9e14pNQ562C_wMAAP__jSDPbA">