<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 - Poor lowering of SSE4.2 string intrinsics"
href="https://bugs.llvm.org/show_bug.cgi?id=37246">37246</a>
</td>
</tr>
<tr>
<th>Summary</th>
<td>Poor lowering of SSE4.2 string intrinsics
</td>
</tr>
<tr>
<th>Product</th>
<td>new-bugs
</td>
</tr>
<tr>
<th>Version</th>
<td>unspecified
</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>enhancement
</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>chandlerc@gmail.com
</td>
</tr>
<tr>
<th>CC</th>
<td>llvm-bugs@lists.llvm.org
</td>
</tr></table>
<p>
<div>
<pre>For hypothetical memcmp code:
----
#include <cstdlib>
#include <x86intrin.h>
extern "C"
int my_memcmp(const void* lhs, const void* rhs, size_t count ) {
const unsigned char* s1 = reinterpret_cast<const unsigned char*>(lhs);
const unsigned char* s2 = reinterpret_cast<const unsigned char*>(rhs);
const __m128i s1chunk =
_mm_loadu_si128(reinterpret_cast<const __m128i*>(lhs));
const __m128i s2chunk =
_mm_loadu_si128(reinterpret_cast<const __m128i*>(rhs));
constexpr unsigned char mode =
_SIDD_UBYTE_OPS | // compare unsigned 8-bit characters
_SIDD_CMP_EQUAL_EACH | // for equality
_SIDD_NEGATIVE_POLARITY | // negate results
_SIDD_LEAST_SIGNIFICANT;
unsigned mask = _mm_cmpistri(s1chunk, s2chunk, mode);
if (mask != 0) {
return s1[mask] - s2[mask];
}
return 0;
}
----
LLVM generates annoyingly suboptimal code:
----
my_memcmp: # @my_memcmp
movdqu (%rdi), %xmm0
movdqu (%rsi), %xmm1
pcmpistri $24, %xmm1, %xmm0
testl %ecx, %ecx
je .LBB0_1
movl %ecx, %ecx
movzbl (%rdi,%rcx), %eax
movzbl (%rsi,%rcx), %ecx
subl %ecx, %eax
retq
.LBB0_1:
xorl %eax, %eax
retq
----
For comparison:
<a href="https://godbolt.org/g/LiHXxk">https://godbolt.org/g/LiHXxk</a>
Changing the code to use the flag-setting spelling doesn't really help:
----
if (_mm_cmpistrc(s1chunk, s2chunk, mode)) {
unsigned mask = _mm_cmpistri(s1chunk, s2chunk, mode);
return s1[mask] - s2[mask];
}
----
LLVM now generates:
----
my_memcmp: # @my_memcmp
movdqu (%rdi), %xmm0
movdqu (%rsi), %xmm1
pcmpistri $24, %xmm1, %xmm0
jae .LBB0_1
pcmpistri $24, %xmm1, %xmm0
movzbl (%rdi,%rcx), %eax
movzbl (%rsi,%rcx), %ecx
subl %ecx, %eax
retq
.LBB0_1:
xorl %eax, %eax
retq
----
And for comparison:
<a href="https://godbolt.org/g/UtYqu2">https://godbolt.org/g/UtYqu2</a>
What we really should be generating:
----
my_memcmp: # @my_memcmp
movdqu (%rdi), %xmm0
pcmpistri $24, (%rsi), %xmm0
jae .LBB0_1
movzbl (%rdi,%rcx), %eax
movzbl (%rsi,%rcx), %ecx
subl %ecx, %eax
retq
.LBB0_1:
xorl %eax, %eax
retq
----
ICC generates a weird self-move for %ecx? Surely that's not needed.</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>