<table border="1" cellspacing="0" cellpadding="8">
    <tr>
        <th>Issue</th>
        <td>
            <a href=https://github.com/llvm/llvm-project/issues/94025>94025</a>
        </td>
    </tr>

    <tr>
        <th>Summary</th>
        <td>
            AVX-512 mask registers spill to stack when GPRs are available
        </td>
    </tr>

    <tr>
      <th>Labels</th>
      <td>
            new issue
      </td>
    </tr>

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

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

<pre>
    I hit some cases where LLVM spills mask registers to the stack in memory-bound code, where extra loads and stores to the stack are quite expensive. Some of these spills may be avoidable through better scheduling, but I think improving the performance of the spills (when unavoidable) would also be useful.

I noticed that GCC spills these registers to GPRs -- could LLVM do the same?

Here is a minimal example:
* clang spills mask registers to the stack: https://godbolt.org/z/e9nGb6z8s
* gcc spills mask registers to GPRs: https://godbolt.org/z/h46cdje7a

Here is a more realistic scenario, where better scheduling could theoretically eliminate the spills (but doesn't): https://godbolt.org/z/Pz1dsrh53

For the realistic scenario, I also observed that GCC spills to GPRs, while LLVM spills to the stack.

I chatted offline with @MatzeB, who indicated that this would probably require a lot of work in the register allocator. I understand if the gains aren't large enough to justify this work. But I still thought it might be useful to share this data point with the community.
</pre>
<img width="1px" height="1px" alt="" src="http://email.email.llvm.org/o/eJyMVMFu2zgQ_Rr6MrAhU1ZsHXRIUjgboAWKXaDY64gai5NQpEuO7Dpfv6DsNKm3BXKxDZjz3syb9wZT4t4TNaq6U9WnGY5iQ2xoaPtZG7pT8wiWBVIYCAwmSnC0FAk-f_72BdKenUswYHqGSD0noZhAAoglSILmGdjDQEOIp3kbRt-BCR0pfX9BoR8SEVzALgH6DpKESFcIGAm-jyz59Z584gMt4J_cT9jlZ4ne-jhBS4CHwB22jkBsDGNvoSURipCMpW507PvcQTsKPIJY9s_Awz6GA_t-4t1T3IU4oDevHK8MSm-OljyM_ieJ0jUcw-g6QJdC5h8T7Ua3UMUnVdyePx_BB2FDHYhFgYf7-1fE8wC_iPfw9e8E8zmYCXUSursIggOpcvse-a8sIydAGNjzgA7oBw57R6q8vFD6FoxD339gXaq8BSuyT7lab5Xe9qFrg5NFiL3S2xelt1T7h_bmZZPe4Htj_gyex_kIrl3dmO6J1viH8ULMOqHjJGwgGfIYObx56X9LvggolkIkYYPOnYAcD-xR6Gqt2Q1doOSVXovS9Uc6_vqy7FK0Vfm-422IE_TvW308uyS0ieLhd3a46DVNxe7XnL3f1JW9jEUR6iDsdo49wZHFgloVX1Be6O4MF4B9xwbllVcsp4t39zG02LoTRPo-ciTIqZTs_mOIU4rPM533CuhcMCghLuARRt9RTJLzy-e09Mg-5eBOaoLD2BOQn7IoAZ7GJLw7vfLH5wXcTWFMws6B2PxQgAUGzj9-ZioXJ5sPwlTaoSDsA3s5z5upTRiG0bOcFrOuKbu6rHFGzXK9XC83q2W5nNmmLvWmbeva6PXu5qYq2lWHptBmc7Oq1_WaZtzoQq-Kqlwu67Iql4tdhbgpN2W12lXt2hRqVdCA7BbOHYZsiBmnNFJTrwpdzRy25NJ0ULX2dITpT6V1vq-xyTXzduyTWhXZIekNRVgcNbff_p1XS30dpckEkwLTWZzu0HQqsh54QHb5HM3G6Jor47LYsV2YMCi9zVyXr_k-hicyovR26jApvT1PcGj0fwEAAP__HzkRfQ">