<table border="1" cellspacing="0" cellpadding="8">
    <tr>
        <th>Issue</th>
        <td>
            <a href=https://github.com/llvm/llvm-project/issues/92906>92906</a>
        </td>
    </tr>

    <tr>
        <th>Summary</th>
        <td>
            Codegen issue with SVML vectorization on Windows. (wrong calling convention?)
        </td>
    </tr>

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

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

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

<pre>
    I am building on Clang-CL 17.0.3 via Visual Studio 2022 on Windows.

I'm using SVML to vectorise some loops (with the flags `/O2 /arch:AVX2 -openmp /clang:-fveclib=SVML /clang:-fno-math-errno`), and linking against svml_dispmd.lib as shipped with the 2024.0 Intel compiler

I have a attached a [Repro solution](https://github.com/llvm/llvm-project/files/15389811/CppApp.zip) that reproduces the problem.  You can see with a breakpoint in the main method that the array is filled with `-inf` when SVML is used instead of the correct values. 

As far as I understand it, I think the issue is that the Intel SVML functions use a vectorcall calling convention, and the arguments should be passed in ymm0 and ymm1, however the compiler seems to be calling them using cdecl with pointers in rcx and rdx:

```
00007FF67DDF1090  vmovups     ymm0,ymmword ptr [rsi+rdi] 
00007FF67DDF1095  vmovaps     ymmword ptr [rsp+80h],ymm0 
00007FF67DDF109E  vmovapd     ymm0,ymmword ptr [rsp+40h] 
00007FF67DDF10A4  vmovapd     ymmword ptr [rsp+0A0h],ymm0 
00007FF67DDF10AD  lea         rcx,[rsp+0A0h]  
00007FF67DDF10B5  mov rdx,rbx  
00007FF67DDF10B8  call        __svml_pow4 (07FF67DDF1E00h) 
```

Testing on [godbolt](https://godbolt.org/z/bM4EaP1ae), I see no issue with the generated assembly, so I suspect this is windows-specific.

```
        vmovups ymm1, ymmword ptr [rbx + r13]
        vmovupd ymm0, ymmword ptr [rsp + 48]    # 32-byte Reload
 call    __svml_pow4@PLT
```

Unfortunately as I'm running from VS I don't have bugpoint, but hopefully the above description and repro helps narrow this down.
</pre>
<img width="1px" height="1px" alt="" src="http://email.email.llvm.org/o/eJyUVk2P2zgS_TX0pWCBpiRbOvjgjwhrIEGCJNu7cwoosWRxQpECScnt_PoBKTvdSTwZjNFgwzL5WFXv1Stx5-RZI25Jvif5ccFH3xm7fa_khPY_XCneL2ojrtsT8B7qUSoh9RmMhoPi-rw8vIXVJqFJCpPk8CTdyBV88qOQBhhlLOz8n9TCXFxC6JHQ3byeCNv0MLoA9unp3VvwBiZsvLHSITjTIyhjBgeEFRfpO_AdQqv42QFZU8Kq9wwIq7htOpLudk__Z7A0A-p-CI-bEBtJd8t2wkbJmqTHeMkPP2mz7LnvlmitNhG0JOwAXAtQUn8NkfEzl9p5cFOvvgjphl4kStbAHbhODgMK-B4coyxLKJy0RwWN6Qep0P6QMnR8QuDAvedNhwI4kHz_EQdrwBk1emk0yY-EFZ33gyPpjrCKsOosfTfWSWN6wiqlpvu_5WDNn9h4wqpWKnSEVas8LcpitSKsOgzDbhiSb3IgrATfcQ82XCXGBl2MeLCmVtgnAH-YERquwSHOCXGoLfKvg5Hag9Rxe8-lhh59Z8QMFx5ya_kVpINWKnUvB1nTpdQtWVO4dKhngqWD0aGAUFDkAkwbARpjLTYeJq5GdAm8rtjOQcttqPYJRi3QOh_YkT7wdALfSf01gkjnxrC-xDXTEC9uR92E0sb7gd9k1nClICyB58boCXWs_00Bc27nsUftA9lmVAJqhIG7OQm49j2NW699vwrHOnPBCe0tq5n_UNHeBXHX-P023-Fd-o3ARs1Fi7VG6wK2bZ4jtBXPQQWvShJ0Ov_Fr5RSuqmq9eZ4rFa0pABTb6ZxcBA-IUTCDte-vxgrYPA2CM46SdjeCknyIzyEyWcY_gLz4_mBsH1Bu6jVQ6zDQ5g3dxjx22gCWhbRHsLssl9gfj1Pd_8Yzu4IoJDD_WObZ8IOPyPAw7P7HKA3U-SDHWz9_DfbCogk36_48iUax2AuWfCxl51vKO1CVz7kdF4_o_M3pyX5_mxEbZR_6A7zT4mxZ8Kqb4RV9bvsDf-w4nhztFPsa21ubfLdsc6o0XIfjMg57Gt1DbudCQdGN4Su9J10oa8us4Uvw1PZyib5jSbv2d-VeO-Pn4mrn4GwPdhVGtJ6cFTcJfOA83g0KyJjAISlkLJlffUIH1EZLm54dzZeMUEy-uHt599U_r-6NdaPmntU12g-cVbZUetASGtND0-f4AQi2MXGz7Zej-fYwSHcevTQmQHbUanrbCW1mRAEusbKIfjM3N7R-jtUgwPNrTWXueDCXHSyENtUlGnJF7hdbVZrVmYlyxbdNk0LXhe0LJoCszUr8mK9WaclzbKyaQVtF3IbZhHN2WrF0pxlSSEwayjmad4KWm7WJKPYc6mSMEWCcBZRGduSlXS9ULxG5eILAWMaL7NsCGPh_cBu4-Spx7MjGVXSefeC4qVXuD0YgWfUr9UWffg23r_xmP-r94I4460JdvirH6cVYeVitGr7r4diDCBMxZjXXwEAAP__DO3Kjw">