<table border="1" cellspacing="0" cellpadding="8">
    <tr>
        <th>Issue</th>
        <td>
            <a href=https://github.com/llvm/llvm-project/issues/133394>133394</a>
        </td>
    </tr>

    <tr>
        <th>Summary</th>
        <td>
            [RISCV] How to model chaining in RVV scheduler?
        </td>
    </tr>

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

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

    <tr>
      <th>Reporter</th>
      <td>
          wangpc-pp
      </td>
    </tr>
</table>

<pre>
    IIUC, we have already considered LMUL/SEW in our RVV scheduler, but not chaining. So, we may overestimate the latency of RVV instructions. For example:
```
vsetvli a0, zero, e8, m8, ta, ma
vadd.vv v3, v2, v1
vmul.vv v5, v4, v3
```
The part of v3 can be read by `vmul.vv` before the commit of `vadd.vv`, so the whole latency is lower than the sum of latencies of these two instructions.

We have `ReadAdvance` to bypass some cycles, but now we don't apply this mechanism to existing models. Instead, we set the `Latency` to `LMUL_1`'s cycles and occupy the vector unit during the whole execution via setting `AcquireAtCycles` and `ReleaseAtCycles` (correct me if I understand it wrong). I don't know if this way  is OK or good enough.

So some questions here:
1.  Is current approach good enough?
2. Any another thoughts about modeling chaining?

@michaelmaitland @mshockwave 
</pre>
<img width="1" height="1" alt="" src="http://email.email.llvm.org/o/eJxcVF1v6zYP_jXKDVHDkZyvC1_kTd9gwXowoF17LgdZYmKt-vDRh13v1w9SvNN1gKHEIvmQz0OaPAR1s4gt2fyPbB5XPMXe-Xbi9jaIh2FYdU7O7eXyeiL0BBNCz0cErj1yOYNwNiiJHiU8fXt9IvT88v_voCy45OH57Q2C6FEmjT5HdymCdRFEz5VV9lbBi1tQDZ_BjegxRGV4RIg9guYRrZjBXQuWsiH6JKJyNlRwdh7wg5tBI2FHUh_Jtl6e-jgGjKNWwOuM_xf6kgf3-TTljLz859mZS1mNI4wsX420nOtsMEkXw6ZcNeVk_0n1e48wcB9zlSMDwS10CFke6GYg23pBIdsaOrw6f-cmnDGqBGWXewUZkJ4guOIx9U5_aqACaDehh9hzW-whmRx-d1AY8kvsMSDEyX0VK5dcH78vzSPb-hm5PMqRW4G5ruigmwceAgRnEMQsNIbPjk25Q9JZQncR-DDoGWKvAhgUPbcqmAyAHypEZW9gnEQdKrjYEJHLpcEBY6mabOunO6Ulcb749vr0x7qw34UlO3ArwQmRhrnEjSii85CsiiCTz4k-RcIPFClThVHxnKoUQrb1UfxIyuMxnu6UtnXBLQpo5OGLhdC9cN6jiGAQ1BUukKxEH2KOUREm7-yN0EMFl59yvGd51PUuyMRnyJ367VdwHm7OSUDr0q1fOvDi7gL_SHnOnQ3Qo1_Gd10BXAKI5D3aIrN3XPT_RiHsTOojreBoZ-DWxb7MQzbFALxzKd7Vz-z_-cjuQflpaqNEz1EbrqIuOjS1Cb0T71Oei5VsmTywA19hu941bL2p2Zqu-nYv6f7a7a-MNbTB7XXdbemuEVd6YHy3P7CVamlNNzWj-3pPN2tacSYZE7jtxE52dN2QpkbDla60Hk3l_G2lQkjYrhljh2aleYc6lBVEacfFO1pJ2PH58nJ6eCM0f5GEUqMeyjrJF5vHlW8z2EOXboE0tVYhhk_4qKIuOy1jvJHNI_zipjxuRZ-f4uRV9XVNsfMqed32MQ4hN4aeCT3fVOxTVwlnCD3nHMvPw-DdnygioefCJxB6XiiNLf07AAD__yKrwEU">