<table border="1" cellspacing="0" cellpadding="8">
    <tr>
        <th>Issue</th>
        <td>
            <a href=https://github.com/llvm/llvm-project/issues/125850>125850</a>
        </td>
    </tr>

    <tr>
        <th>Summary</th>
        <td>
            Incorrect handling of `ARM64_RELOC_PAGEOFF12` by `RuntimeDyldMachOAArch64::resolveRelocation`
        </td>
    </tr>

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

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

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

<pre>
    ## Summary
In certain situations, the `RuntimeDyldMachOAArch64::resolveRelocation` method patches an `add` instruction with an incorrect immediate operand.

## Details
The instruction being patched is an `add` with a 12-bit immediate that is paired with an `adrp` instruction to compute the address of a label.  The calculation use to compute the immediate is
```c++
(Value + RE.Addend) & 0xFFF
```
The value of `(Value + RE.Addend)` is the memory
The variable `Value` holds the base address of the memory allocated by the loader's memory manager to contain the Section, and `(Value + RE.Addend)` is the address of the label in
that memory.  The problem is that `Value` is not guaranteed to be 12-bit aligned and so the computed immediate may be incorrect.

My code generator is based on LLVM 18.1.8, but the implementation of `RuntimeDyldMachOAArch64::resolveRelocation` does not appear to have changed in 19.1.7.

## Proposed Fix

In debugging this issue, I've noticed that `RE.Addend` appears to always hold the correct patch value, so think the correct computation should either be
```c++
RE.Addend & 0xFFF
```
or
```c++
(Value - Section.getLoadAddress() + RE.Addend) & 0xFFF
```
With the caveat that in my code, I've only seen the `add` instruction patched, but I think that this also works for patching load/store instruction (the other use of `ARM64_RELOC_PAGEOFF12`).

An alternative fix might be to set the alignment requirement for the code section to $2^12$ in the call to `MemManager::reserveAllocationSpace`, but that seems more fragile.  I also noticed that the MCJIT execution engine avoids this bug because it page-aligns the memory allocated for sections.


</pre>
<img width="1" height="1" alt="" src="http://email.email.llvm.org/o/eJycVt1u2zgTfRr6ZhBBon9z4Qu1jT-kiNEiLfpdFiNyLHFLkVqScuK3X5CUnThod7sLCDBgznDOnDlnJPRetYZoy5bv2PLDDMfQWbf9aDvzSMNwmjVWnraMzxmfw5ex79GdWFnfGxDkAioDXoURg7LGM_4eQkfAVuXjaILq6cNJyz2K7lNdO9GtFmxes3ntyFt9pEfSVqRMtiqhp9BZCQMG0ZEHNPEalDKeKeODG0UMhScVuniqjLDOkQig-p6kwkBgB3JoZMHKOj4Z9AcKqLRnZf21o6urGlKmnSpKUG-K5kJQ8ZtGvS4SOgwxeEDlSF7wpEQ3vIUbLAjbD2NKJEApHXkP9gAIGhvSBUDEJVCLUSc2YPT0Nu-lvIqdsFWZH8H4u_jEbjffUI8EjL-Dx7uilpKMZPwWGF9B-bzb7V4nTnQcU4o9RPi_vCH15BOOnnqbBJCTncJGp4GnzBjYWS1zbIP-quGXfECdRk8SmlP6X1uU5Bhf-3NIjwZbcpkIk5QWA79Q4jVKDY38XdhvUCTiQRlW1mmaueQ0icHZRlOfUzFcNac8GBugHdGhCUQywmvoLBLU0UsyIfM2lZqGKF9NsMdTzLkIeJLr_gTCSoKWDDkM1sVqkUMJ1sDDw7c9VJuiKjax92YMky4GTT2ZkJWTB_nvzSct5c5wGAgT6R0eCUSHpo3gDVS3RVWsr6312dnBRoA79ZwP7g1Iasa2jc4KnfKgvB8pQr5nfH2kWEWJSNzE7cvEVuVU3sf6qJ_w5JOcJiKz25Nds27jrYlmZX5cxWTSMyW-s6OWQCp05KChn7vnguKXbrHuH3x3c9Zm0VJ4sCjrLDrGN9mFv2nL_8eFkrrBI2GY9o2BPuvjFZXW6BN4InPeuj9Zl9N2O2vm_sJWujnuPO0tPFn3w8PBuhwfhxcdyfjOB-uutybjm1jOJj7jqsqiqx_3q8X3x7uHT--_f67_d_dpt6t4suftJJraAOpAzmBQR4KDeoZetV2IZggWPGVJJxNFSYOjP0flkrwTtjxiSeDpslwZX3C2vKs44wuYdoRArdPZqtxTv8-L5KJ-ckeq9Vn9XwYUlGCeTYUhctp76GPnB4et0lQA3GeqrvQbq-3ff7z_CvRMYkygyLTKEODRqrQIo4vHFhoSGNlSUcIt3aQ2_c-XYmx26tFP5M3kdi5v57c4o221nm_Wq4pX5azbisPhQGXViMNmvtgI2VSL1XJdygWtq6pZNTO15SVflrxcVvNqPi-Lw1JseLVYN4cVlbdywRYl9ah0ofWxL6xrZ8mz24ovN8tylpalT98HnBt6Ojuax88Ft41JN83YerYotfLBv1wTVNC0vb-8qDs0Ukdt_b1i4jvhv2yx2ej0tgth8DGG7xjftSp0Y1MI2zO-i7imn5vB2T9IBMZ3qRvP-G5q97jlfwUAAP__r7P5bw">