<table border="1" cellspacing="0" cellpadding="8">
    <tr>
        <th>Issue</th>
        <td>
            <a href=https://github.com/llvm/llvm-project/issues/61930>61930</a>
        </td>
    </tr>

    <tr>
        <th>Summary</th>
        <td>
            Fold adds with global addresses into load offset
        </td>
    </tr>

    <tr>
      <th>Labels</th>
      <td>
            backend:WebAssembly,
            llvm:codegen
      </td>
    </tr>

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

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

<pre>
    From https://reviews.llvm.org/D139645:

Given the following C:

```c
extern int *data;
int f(unsigned idx) {
  return data[idx];
}
```

You would think we could emit

```wasm
i32.load8_u     data($0)
```

But instead we currently emit

```wasm
h:                                      # @h
        .functype       h (i32) -> (i32)
# %bb.0: # %entry
        i64.const       data
        local.get       0
        i64.extend_i32_u
        i64.add 
        i32.load8_s     0
```

We can't naively fold the add into the load offset though, because the effective address calculation is unsigned, which can overflow with negative offsets:

> OK, so looking back at https://github.com/llvm/llvm-project/issues/29497 and [D24053](https://reviews.llvm.org/D24053) this actually does look like the problem that we fixed back then.
> IIRC the problem is that sometimes the address operand of the store (i.e. local 0, or [L1](https://reviews.llvm.org/L1) here) can have a negative value (here it's -128), but the store's effective address calculation (i.e. the operand plus the constant offset, [L1](https://reviews.llvm.org/L1) + args + 128 with this CL applied) is unsigned, so it will overflow. 
> So we have to ensure that the calculation that recombines the native local value with the compile-time constant (here 128) happens with an add instruction rather than getting folded.

For signed types, this is a problem. But there should be a way to do the folding for unsigned types.
</pre>
<img width="1px" height="1px" alt="" src="http://email.email.llvm.org/o/eJycVUuv4jgT_TVmU7qR4yRAFizg0nxqfS2NNLNozarlxEXiuY6NbAeafz8qJzz6qtVzZxAi-FGvc05VZAi6s4gbVu1YtV_IMfbOb5qxaeSiceq6OXg3QB_jKbBiy8SBiYPHs8ZLyIw5D5nzHROHfV7Uy7KiK3zP-Pz7P31GC7FHODpj3EXbDl7f3WFLPn3baY3fI3oL2kZgYqtklKzYTUe0d2RiPdqUtAKtvjNRA1vNFwA8xtFbSGbVjs6r_d2erfbvYj4n8qcb4eJGoyD22r7BBaFNSxx0_GnGFxmGObNCZMZJtf42Ml6n6GLNRMmZqH8RcjdG0DZElCqFG71HG831IyF7VmzhQx8mCmAl728O6-w42jZeT8h43QMTa10IwvGFFZ8ey_k6GYuqaTJOAecl2uivd396WWatsyHear8dGNdKk3VIB_yH68SyVd90IQixpwOpFDw27rCGJw8_Q_IrQistE6sIVuozmitJTiXtkUtto0sLcgfueAwYIfZu7HomXqHBVo4B0w08HrGN-pwMPYYArTTtaGTUzoIOcJMfGV563fYUGtwZ_dG4C1x07MFiJ5OPKVR4L_riE_z2f3IQHBjn3qgzGtm-gYzvmq3TsR-brHUDEwfquPnxcvLuL2wjEwcdwoiBiYOoy3oF0ipg1W4vSl4VpH-x_uf-nS6LmsQfQLZxlMZcQTkMKUEw-m3C5-RdY3CA2MtIqj3q76im5GOPNntU-Pnz768_mOgwWQU3YNQDhhs_CWZ3Qk-5u2PaDtF5THLMMIMkJeAEmfNU3pf8g6V9yamsHj3Sk6jqJZH74OgszZgi0SXQkYlVgJecOrhO6hjjI6N0-GuN3HImm1tRJzNO1aZGkTbOyiD__74aJnYgfRfSn1ysJ9El6l6_gDydjCZ91u_lGhzoCBdtzF2vGTwI-8MRoQme6ABtGD1OjKXMn0pMmx5bNzTazjzaCc2JqQnTOS2qejhpgy_E-gOCG-QT1NDL0wltmKyknRs3RD-2KaiXsUdPsS10GCM1DXU5quy5uQ7Ow_yCoCEXqPAEDQn7psUMdhOrHiH0ac43JIqLvFLtyt1eW2qK4u84Tk6zhdoUqi5qucBNvlzzJa_5Siz6TVupI_JizWWdL9dK8ErxdrVsywZVnUtc6I3gouAlL_O8ynOR1U0j1nlTLkVVr9umZCXHQWpz532ROnyzzOuCL4xs0IT0whaC2g6tYsX2KzbbEHBozJUJwcQrEyKNi2LbOoUdWtqu9gu_SeOjGbvASm50iA-BLaKOBjcHGp1SqZmKzrhGmpvUMUzT9GmSLkZvNv99bKW6_g4AAP__8Palpw">