<table border="1" cellspacing="0" cellpadding="8">
    <tr>
        <th>Issue</th>
        <td>
            <a href=https://github.com/llvm/llvm-project/issues/133449>133449</a>
        </td>
    </tr>

    <tr>
        <th>Summary</th>
        <td>
            [RISCV] Address instances of redundant or suboptimally canonicalised instructions in llvm-test-suite (tracking issue)
        </td>
    </tr>

    <tr>
      <th>Labels</th>
      <td>
            backend:RISC-V
      </td>
    </tr>

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

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

<pre>
    This issue tracks remaining work to remove any redundant of suboptimally canonicalised instructions as found in the output of llvm-test-suite (including SPEC). None are known to be particularly high in dynamic instruction count, but as they can be removed/improved with relatively little effort and the generated code is unambiguously better, it seems worth stepping through. This [scrappy script](https://gist.github.com/asb/c2a8af337e53f8d1b399d88f3dce3ad0) can be used for searching binaries in a directory (may be false positives). After spotting a few of these issues by eye, I thought it was probably worth being a bit more thorough (and additionally, some of them are helpful for anyone looking to minimise code size).

There are ~8k static instances in llvm-test-suite (note: hasn't been carefully checked for false positives).

## Redundant operations
* [x] No-op moves (`mv`) left after MachineCopyPropagation
  * Fixed in #129889 and #81123
* [ ] No-op moves in the form of xor/or/sub/sh*add[.uw]
  * xor/or/sub addressed in #132002, sh*add addressed in #133443 
* (TODO: link to strand of work from Mikhail, Philip and others on branches)
* [ ] Always-false or always true branches e.g. `bltu zero, t1, ...` or `bgeu zero, a2, ...`
* [ ] j to the next instruction (within the same function)
* j to the next instruction (falls through to first instruction of the next function)
  * Deleting the jump would be correct, but may be more hassle with the linker than desirable for minimal gain.
* [ ] No-op `mv` (i.e. with both operands equal)
  * Remaining cases may be due to lui/addi pairs where the addi immediate is resolved to 0 by the linker, the but the addi isn't removed.

## Suboptimally canonicalised operations
* [ ] Reg-reg moves encoded as e.g. `sh1add` with `rs1=zer0`.
  * Produced after MachineCopyPropagation. Not compressible, while the plain `mv` is.
  * Likely best solved by late stage canonicalisation. This _could_ only be done when compression is enabled, but given it's never worse than neutral, prefer to keep codegen the same for compressed vs non-compressed.
* [ ]  `or rd, zero zero`
    * Not compressible, want to use `c.li`. Also, a number of instances of this in 502.gcc_r immediately followed by bnez on the loaded value..
* [ ]  `andi` with `rs1==zero`, `zext.w` of `zero`
* [ ] `beq zero, rs2, ...` and `bne zero, rs2, ..`
   * Should compress to `c.bnez`/`c.beqz` and print an appropriate alias
* [ ] `seqz` /`snez` with `zero` oeprand
* [ ] `sll/srl/...` with `zero` operand for `rs`
* [ ] `addw rd, zero, rt`
   * Should be `addiw` for better compressibility
* Other minor variants of the above
</pre>
<img width="1" height="1" alt="" src="http://email.email.llvm.org/o/eJyMV01v4zgS_TXMpRBBpuLYPvjgTtbAADvTje7GXBslsSSxTZFqkrKjHPa3L4qSYyftHezFiSSyPl69ekViCLqxRFux_CSWz3c4xNb5LYbyrnRq3H5vdQAdwkAQPVaHAJ461FbbBk7OHyA6fuOOBGhH8KQGq9BGcDWEoXR91B0aM0KF1lldodGBFGgboh-qqJ0NgAFqN1h-C7ElcEPsh2TBmGN3HynE-zDoSCDkWtvKDIrdf_vyrychNxn85SwBeoKDdSfLEZUEPfqoq8GgNyO0umnZuhotdrq6dg-VG2wU8gnKIXIosaUULRuZMlNC7nXXe_4XTjq24Mlg1EcyIxgdoyGgunY-AlqVUmjIksdICiqnCHSAwWJX6mZwQzAjlBQjefaqIwSiLjCasYUQqe85u9h6NzRtBqkCYvkpVB77foRQed1HsXwWct3G2AdR7ITcC7lvdIhZo2M7lFnlOiH3GEoh95XENdZFsaJlUa_Voiw2G7Ve14WqqECVC7k5JzxwcWrnIRD6quVASm3RawqMH4LSnqro_Mi16JAzgRpNIOhd0IxJSDXZ1ZE8hN7FyEYQajpxRWNLgSZCBShHoJEYhT8gtpxuZDxOGKD3rsTSjDMsJU1WSh2hc554ecKHw2DQUSnN5WSuscHgOpr9dYkbLZm-HkxKDu3IlDHOHRLUDjptdacDTeUK-pU4C5HvRL773pKf-PWf9QFCxDgzCG01wXKDptZFEsUOWgxWyFWEkshChZ7qIbVDS9VhxvoGfpNnIQshC_h6aaqeWcVNk77umBcvYvkMf7l71wOTNbB38Zh3R_GYKmuojoCpHH8il5SeXD9-8a7HJtkS-Q6Aje31S2pNELJYyM16vUl8FrJYLxayuLiEjy7nxq2d7xj0F-eF3KefMDADQyvkDpUSy0_ZcGLunn2-X8pl9BTCJYxC5rlMBZ1N3FhSPDwUcI5Orr9_fv7M0BttkzyF6DkNV0-CVXvXwZ_60KI2bPhLq43uU6YutuQDOAulR1u1qRYf0t6ZE47hfqoZcyk9Q_QDve0CypoMxGNemjjAK3nHjuKCf7MsE4857-TvDV2-o7x8_-D0J-fBCFt6ie_US8g1K9JcgIAdQT3Y9O0S-z9ur9GYcJYbXlhrH96vmvpo2v3e-lTEZzIUJ9Ei-Dl0PZzcYBRrQ-U8C8ZZX2fFSC3cYgiGJkHljVwv8hBbtKAoaI-lSZyauhMNNKhtdpOGZ8KnCZFRNlktXWynnrEqAP0a0FyH_fVtklUYKJyDUzzrHJhBs4IqpaFH7QOckg5wpOml7jpSGmNSd0_BGR4P0UHOunZJKJW-pZT-ZfOsCvOAed_v3_733LwlAAmGr9Tce2rmfiTLOqZ4nJ2pGNoFN-BjPkEjHnMfFqJ4fiWfi8c8e0Pli3dqqHjzP2gGD90Ilet67kVdmqTip1abCaHeIHfnuSo6XOz_Wx8oTcAQYQatHMEwkCFiQ9cZz77SCPxRMad-gLNmKhNL-KllVT2H4SzXgiwTR50p1-gjWdBRyFUAS0fyrAOBJqJZGqLHJAS9p5r55-BA1KdJ0NB1Xzn_5ooUHANYZ-8vbz4yk9N3HnyKhJt86vTU3AATGDdRZKWPjmcxm6gyo7lAsDNh0gmwQ1eS57a8jKHUozpJ8TKXWVNVP_yFo2aE2hnjThPapaVXlrnEUodMlSOagbJbOaBV-nfeTNRJ6cgnfv1KLzE7JW2rp-dzstcGWfTo15vm-XAletO0ecxLS78veMONrX1rk76ckWO0ElKcV4poPz3Sr9ez3d5ry-czwL73rvepc9Fo_NhJ3CvzvslOmIy-5T8nBo56FpYb243hceb5d87st72TKCVOJUhvIoVKna74k_CIt4AoaV6uUwHY6nTCvCKXNjqOs4_PPOhYVp2HI3qNNoazymPpjnSntoXaFBu8o-1i9VAsVsWDLO7abbWpVrQuiofVgqqK6hVt6mJZ4KqWS7WQ8k5vZS6XeSHXi4dlLotssSge6sd1UeerIl-vFuIhZ901GZ-aMuebu3Qa3KY5vrkzWJIJ6ToiZYnVgawSxe7rH9-e7v8WUvIlxW_TiascmiAecqNDDBdrUUeTrjO85e80sqcTw_tmubqp-P_7pnL7rJcuRjxGUiJCbu4Gb7Yfz-ZXx3K2Mf-57737mQbkfjoUC7mfkThu5X8DAAD__2HReOQ">