<html>
<head>
<base href="https://llvm.org/bugs/" />
</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 --- - [AVX2] vpermd + vpshufb intrinsics merged into wrong code at -O1 and higher"
href="https://llvm.org/bugs/show_bug.cgi?id=27320">27320</a>
</td>
</tr>
<tr>
<th>Summary</th>
<td>[AVX2] vpermd + vpshufb intrinsics merged into wrong code at -O1 and higher
</td>
</tr>
<tr>
<th>Product</th>
<td>new-bugs
</td>
</tr>
<tr>
<th>Version</th>
<td>3.8
</td>
</tr>
<tr>
<th>Hardware</th>
<td>PC
</td>
</tr>
<tr>
<th>OS</th>
<td>Linux
</td>
</tr>
<tr>
<th>Status</th>
<td>NEW
</td>
</tr>
<tr>
<th>Severity</th>
<td>normal
</td>
</tr>
<tr>
<th>Priority</th>
<td>P
</td>
</tr>
<tr>
<th>Component</th>
<td>new bugs
</td>
</tr>
<tr>
<th>Assignee</th>
<td>unassignedbugs@nondot.org
</td>
</tr>
<tr>
<th>Reporter</th>
<td>peter@cordes.ca
</td>
</tr>
<tr>
<th>CC</th>
<td>llvm-bugs@lists.llvm.org
</td>
</tr>
<tr>
<th>Classification</th>
<td>Unclassified
</td>
</tr></table>
<p>
<div>
<pre>clang 3.8 (but not clang 3.7.1) mis-compiles a sequence of two shuffles.
(Background:
<a href="http://stackoverflow.com/questions/36327100/complex-data-reorganization-with-vector-instructions">http://stackoverflow.com/questions/36327100/complex-data-reorganization-with-vector-instructions</a>)
__m256i doubleshuff(__m256i in) {
const __m256i byteshuf = ...;
const __m256i lane_adjust_shuf = ...;
in = _mm256_permutevar8x32_epi32(in, lane_adjust_shuf);
return _mm256_shuffle_epi8(in, byteshuf);
}
clang 3.8.1-svn265380-1~exp1 (from llvm repo for Ubuntu 15.10) compiles that to
(with -O3 -march=haswell):
vmovdqa ymm0, ymmword ptr [rip + .LCPI3_0] # ymm0 = [0,1,2,2,3,4,5,5]
vpshufb ymm0, ymm0, ymmword ptr [rip + .LCPI3_1] # ymm0 =
ymm0[0,1,1,2,3,4,4,5,6,7,7,8,9,10,10,11,28,29,29,30,31,16,16,17,18,19,19,20,21,22,22,23]
ret
which is obviously wrong because it doesn't depend on the input. It's
shuffling the vpermd shuffle mask instead of shuffling the input.
gcc 5.2 emits:
vmovdqa ymm1, YMMWORD PTR .LC7[rip]
vpermd ymm0, ymm1, ymm0
vpshufb ymm0, ymm0, YMMWORD PTR .LC8[rip]
ret
I created a test-case that exits with true or false status, suitable for
testing with git bisect. (I don't have a checkout of the clang tree myself, or
a fast machine).
test-case:
#include <immintrin.h>
#include <assert.h>
#include <string.h>
__m256i doubleshuff(__m256i in) {
const __m256i byteshuf = _mm256_setr_epi8(0,1,1,2, 3,4,4,5, 6,7,7,8,
9,10,10,11,
0,1,1,2, 3,4,4,5, 6,7,7,8, 9,10,10,11);
// or broadcast128
const __m256i lane_adjust_shuf = _mm256_setr_epi32(0,1,2,2, 3,4,5,5);
// const __m256i lane_adjust_shuf =
_mm256_cvtepu8_epi32(_mm_setr_epi8(0,1,2,2, 3,4,5,5,
// /* unused padding that isn't optimized away :( */
0,0,0,0, 0,0,0,0));
in = _mm256_permutevar8x32_epi32(in, lane_adjust_shuf);
return _mm256_shuffle_epi8(in, byteshuf);
}
static unsigned char dst[33];
static const unsigned char src[33] =
"ABCDEFGHIJKL"
"abcdefghijkl"
"XXXXXXXX";
static const unsigned char good_output[33] =
"ABBCDEEFGHHIJKKL"
"abbcdeefghhijkkl";
// return false if result doesn't match
int shuffle_test(void) {
__m256i srcv = _mm256_loadu_si256((const __m256i*)src);
__m256i expanded = doubleshuff(srcv);
_mm256_storeu_si256((__m256i*)dst, expanded);
return !memcmp(dst, good_output, 32);
}
int main(int argc, char**argv) {
int test_passed = shuffle_test();
// assert(test_passed);
return !test_passed;
}
See also this code and a bigger loop using it on godbolt:
<a href="https://godbolt.org/g/nJqyJc">https://godbolt.org/g/nJqyJc</a></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>