<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 - [x86] keep math/logic on vector unit to avoid insert/extract"
href="https://bugs.llvm.org/show_bug.cgi?id=40135">40135</a>
</td>
</tr>
<tr>
<th>Summary</th>
<td>[x86] keep math/logic on vector unit to avoid insert/extract
</td>
</tr>
<tr>
<th>Product</th>
<td>libraries
</td>
</tr>
<tr>
<th>Version</th>
<td>trunk
</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>Backend: X86
</td>
</tr>
<tr>
<th>Assignee</th>
<td>unassignedbugs@nondot.org
</td>
</tr>
<tr>
<th>Reporter</th>
<td>spatel+llvm@rotateright.com
</td>
</tr>
<tr>
<th>CC</th>
<td>craig.topper@gmail.com, llvm-bugs@lists.llvm.org, llvm-dev@redking.me.uk, spatel+llvm@rotateright.com
</td>
</tr></table>
<p>
<div>
<pre>We have a potential regression with shift lowering via rotates while trying to
do scalarization in:
<a href="https://reviews.llvm.org/D55722">https://reviews.llvm.org/D55722</a>
I don't think we're allowed to do anything with this in IR even if the pattern
was appearing there:
define <4 x i32> @shiftamount_math(<4 x i32> %x, <4 x i32> %y) {
%e = extractelement <4 x i32> %x, i32 0
%s = sub i32 32, %e
%i = insertelement <4 x i32> undef, i32 %s, i32 0
%splat = shufflevector <4 x i32> %i, <4 x i32> undef, <4 x i32>
zeroinitializer
%shift = shl <4 x i32> %y, %splat
ret <4 x i32> %shift
}
$ llc -o - vecshift.ll
...
movd %xmm0, %eax
movl $32, %ecx
subl %eax, %ecx
movd %ecx, %xmm0
pslld %xmm0, %xmm1
movdqa %xmm1, %xmm0
retq
We would probably be better off loading the '32' as a splat constant and doing
the 'sub' as a vector op:
movdqa LCPI1_0(%rip), %xmm2 ## xmm2 = [32,32,32,32]
psubd %xmm0, %xmm2
pxor %xmm0, %xmm0
movss %xmm2, %xmm0 ## xmm0 = xmm2[0],xmm0[1,2,3]
pslld %xmm0, %xmm1
movdqa %xmm1, %xmm0
It's ugly because of the x86
vector-shift-by-variable-but-not-really-a-vector-shift instruction that takes a
64-bit shift amount, but that beats transferring to scalar and back?</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>