<table border="1" cellspacing="0" cellpadding="8">
    <tr>
        <th>Issue</th>
        <td>
            <a href=https://github.com/llvm/llvm-project/issues/154116>154116</a>
        </td>
    </tr>

    <tr>
        <th>Summary</th>
        <td>
            separate-const-offset-from-gep miscompile related to "trunc nuw"
        </td>
    </tr>

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

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

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

<pre>
    llvm commit: 673750feea

Reproduce with:
```
opt -passes=separate-const-offset-from-gep bbi-109669_sep.ll -S -o - -mtriple=hexagon
```
The input program does
```
%0 = phi i32 [ -20, %entry ]
%1 = add i32 %0, 20
%2 = trunc nuw i32 %1 to i16
```
where the "trunc nuw" is ok since %1 is 0. No truncated bits are set.

However, in the result we get:
```
%0 = phi i32 [ -20, %entry ]
[...]
%1 = trunc nuw i32 %0 to i16
```
so here the input to "trunc nuw" is negative and the result is poison.
Langref https://llvm.org/docs/LangRef.html#trunc-to-instruction says
"If the nuw keyword is present, and any of the truncated bits are non-zero, the result is a [poison value](https://llvm.org/docs/LangRef.html#poisonvalues)."

Then %1 is further used to compute the return value from the function so the whole thing results in poison which it did not in the input.

(I stumbled upon this as a miscompile after
 https://github.com/llvm/llvm-project/pull/152990
which I first incorrectly blamed.)

[bbi-109669_sep.ll.gz](https://github.com/user-attachments/files/21836248/bbi-109669_sep.ll.gz)
</pre>
<img width="1" height="1" alt="" src="http://email.email.llvm.org/o/eJyUVF1vqzgT_jXOzQhkhkDggov2jaK30movzp77lQMDeGtsZA_N5vz6lYF2z2krrSpFioDHM8_HeFQIerBEjSgeRXE-qIVH55tJPysyozMT2cPVdffGmJcJWjdNmkX-AOUpPxWyJ1JCPgj58I1m77qlJbhpHkUe34lS7j_54GaGZFYhUBD5OdCsvGJKWmcDJ67vA3HSezclA81wveokk3VZ1n8GmlNjIPkDEgcJJBN7PRsS-Xmkv9Xg7Ls-30cCbeeFYfZu8GqCzlF4BxJYSBD5GeZRg84RRPEICUqB_wOBBVn2dxDFeYNmK1R13QbFYsXhXgjXr-wX24Jdbq-YDNiBzsp3nW8jeQIeCQTi2yGBCDqAe4agbUvbeR1ApvC722orpg6umgMoTxCI0833_7sbvZCPjLRdC3sKi2G4EQzEH4P4ivbiMU3Tdz58UCo_VxocvGndAmH3mWhLg2L9QqBs9zN_HWB2Ojgbhf6m7OCph5F5DlETXgRe4kimzg8CL51rg8BLxH2jPh15MgLztVfCLtE2sF9a1s5CUPd1HhCf-rVh1PJM95vz3drVUyDL0Y9ISdk7uA34SRDW2eQHeRfRv5JX0dhNAbwos1D0EasvKtgKrOeDwDoViFvu30eyb3PSL55H8rAE6qLNrZvmhWlnxIvfKUC8YevbfrG7G259vo3ORLy2w64hxHna-d9G3Y6gGTrdgXX8OmprrvsgCqyeIPAyXQ11sMwuQqIP0YpJh8hJGwLVM3khH95lOWgel2vaumm3Zf9LZu_-opYFXubFGIGXrMC63u5SZPUEvfYhUmqd99SyucPVqIm6VGC9cyseP6yUdPjxWSK_8FgC-UQxq3acyHLMp9cmJnHBrMpLPFYCL5-WxvrQNXlX57U6UJOdiqLAvKiKw9jUp1aWJ9kX1Payq0o69t0J8-tRFrKvsvygG5RYyCqrshxLzFMls6pq-7xEWdV9ReIoaVLapK_zc9AhLNRkxTHLyoNRVzJh3eiIlm6wfo2jU5wPft3kyXUZgjhKowOHf8uwZkPNf2znn8L0ZNb78PFqHxZvmi9HvBINa8irkpcG_wkAAP__hEcWRQ">