<html>
<head>
<base href="https://bugs.llvm.org/">
</head>
<body><table border="1" cellspacing="0" cellpadding="8">
<tr>
<th>Bug ID</th>
<td><a class="bz_bug_link
bz_status_NEW "
title="NEW - _mm256_shuffle_ps(_,_,0x44) compiles to vunpcklpd with -march= icelake-client"
href="https://bugs.llvm.org/show_bug.cgi?id=44113">44113</a>
</td>
</tr>
<tr>
<th>Summary</th>
<td>_mm256_shuffle_ps(_,_,0x44) compiles to vunpcklpd with -march= icelake-client
</td>
</tr>
<tr>
<th>Product</th>
<td>clang
</td>
</tr>
<tr>
<th>Version</th>
<td>9.0
</td>
</tr>
<tr>
<th>Hardware</th>
<td>PC
</td>
</tr>
<tr>
<th>OS</th>
<td>All
</td>
</tr>
<tr>
<th>Status</th>
<td>NEW
</td>
</tr>
<tr>
<th>Severity</th>
<td>enhancement
</td>
</tr>
<tr>
<th>Priority</th>
<td>P
</td>
</tr>
<tr>
<th>Component</th>
<td>C++
</td>
</tr>
<tr>
<th>Assignee</th>
<td>unassignedclangbugs@nondot.org
</td>
</tr>
<tr>
<th>Reporter</th>
<td>DrTroll@gmx.de
</td>
</tr>
<tr>
<th>CC</th>
<td>blitzrakete@gmail.com, dgregor@apple.com, erik.pilkington@gmail.com, llvm-bugs@lists.llvm.org, richard-llvm@metafoo.co.uk
</td>
</tr></table>
<p>
<div>
<pre>In a recent Stackoverflow discussion (link:
<a href="https://stackoverflow.com/questions/58954801/avx-equivalent-for-mm-movelh-ps">https://stackoverflow.com/questions/58954801/avx-equivalent-for-mm-movelh-ps</a>)
we found out that the instruction _mm256_shuffle_ps(_,_,0x44) is compiled to
vunpcklpd by Clang. This is a possible optimization for Skylake and other
processors which have identical throughput for shuffles and unpacks.
But as the Stackoverflow-user Peter Cordes mentioned in his answer, Ice Lake
processors have a higher throughput for shuffles than for unpacks:
unpack:
<a href="https://www.uops.info/table.html?search=vunpcklp&cb_lat=on&cb_tp=on&cb_uops=on&cb_ports=on&cb_SKL=on&cb_ICL=on&cb_measurements=on&cb_iaca30=on&cb_doc=on&cb_avx=on&cb_avx2=on">https://www.uops.info/table.html?search=vunpcklp&cb_lat=on&cb_tp=on&cb_uops=on&cb_ports=on&cb_SKL=on&cb_ICL=on&cb_measurements=on&cb_iaca30=on&cb_doc=on&cb_avx=on&cb_avx2=on</a>
shuffle:
<a href="https://www.uops.info/table.html?search=vshufp&cb_lat=on&cb_tp=on&cb_uops=on&cb_ports=on&cb_SKL=on&cb_ICL=on&cb_measurements=on&cb_iaca30=on&cb_doc=on&cb_avx=on&cb_avx2=on">https://www.uops.info/table.html?search=vshufp&cb_lat=on&cb_tp=on&cb_uops=on&cb_ports=on&cb_SKL=on&cb_ICL=on&cb_measurements=on&cb_iaca30=on&cb_doc=on&cb_avx=on&cb_avx2=on</a>
Therefore, the performed replacement of the shuffle is contra-productive on an
Ice Lake processor.
Even with -march= icelake-client, Clang replaces the shuffle with vunpcklpd:
<a href="https://godbolt.org/z/MqtJXY">https://godbolt.org/z/MqtJXY</a>
Same goes for _mm256_shuffle_ps(_,_,0xee) vunpckhpd.</pre>
</div>
</p>
<hr>
<span>You are receiving this mail because:</span>
<ul>
<li>You are on the CC list for the bug.</li>
</ul>
</body>
</html>