<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 --- - strangely pessimized code for double-word operations"
href="https://llvm.org/bugs/show_bug.cgi?id=24983">24983</a>
</td>
</tr>
<tr>
<th>Summary</th>
<td>strangely pessimized code for double-word operations
</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>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>Backend: X86
</td>
</tr>
<tr>
<th>Assignee</th>
<td>unassignedbugs@nondot.org
</td>
</tr>
<tr>
<th>Reporter</th>
<td>bonzini@gnu.org
</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>For this function:
#include <stdint.h>
void mul64(uint64_t *lo, uint64_t *hi, uint64_t a, uint64_t b)
{
typedef union {
uint64_t ll;
struct { uint32_t high, low; } l;
} LL;
LL *ra = (LL *)&a;
LL *rb = (LL *)&b;
LL r0, r32, rt;
/* Build the two products that will go in bits 32-95 */
rt.ll = (uint64_t)ra->l.high * rb->l.low;
r32.ll = (uint64_t)ra->l.low * rb->l.high;
/* Reorganize these two factors into two 33-bit numbers */
*hi = (uint64_t)r32.l.high + rt.l.high; // bits 64-127
r32.ll = (uint64_t)r32.l.low + rt.l.low; // bits 32-95
/* Now build the result 32 bits at a time. Finish bits 0-31... */
r0.ll = (uint64_t)ra->l.low * rb->l.low;
/* ... then bits 32-63... */
r32.ll += r0.l.high;
r0.l.high = r32.l.low;
*lo = r0.ll;
/* ... then bits 64-127. */
*hi += r32.l.high;
*hi += (uint64_t)ra->l.high * rb->l.high;
}
LLVM generally produces nicely optimized code, except for some strange bits:
0: 55 push %ebp
1: 53 push %ebx
2: 57 push %edi
3: 56 push %esi
4: 8b 44 24 24 mov 0x24(%esp),%eax
8: 8b 6c 24 28 mov 0x28(%esp),%ebp
c: 8b 74 24 1c mov 0x1c(%esp),%esi
10: f7 64 24 20 mull 0x20(%esp)
14: 89 d7 mov %edx,%edi
16: 89 c3 mov %eax,%ebx
18: 89 e8 mov %ebp,%eax
1a: f7 e6 mul %esi
1c: 01 fa add %edi,%edx
1e: 8b 7c 24 18 mov 0x18(%esp),%edi
22: 89 17 mov %edx,(%edi)
24: 19 d2 sbb %edx,%edx
26: 83 e2 01 and $0x1,%edx
29: 89 57 04 mov %edx,0x4(%edi)
2c: 01 d8 add %ebx,%eax
2e: 19 d2 sbb %edx,%edx
30: bb ff ff ff ff mov $0xffffffff,%ebx ; here it starts...
35: 83 c3 01 add $0x1,%ebx ; EBX = 0, CF = 1?
38: 19 c9 sbb %ecx,%ecx ; ECX = -1?
3a: 21 d1 and %edx,%ecx ; AND into -1???
3c: 21 c3 and %eax,%ebx ; AND into 0???
3e: 09 c3 or %eax,%ebx ; OR into 0???
40: 83 e1 01 and $0x1,%ecx
which could be changed into simply
mov %edx, %ecx
mov %eax, %ebx
The moves could even be coalesced with the two previous instructions, giving:
add %eax, %ebx
sbb %ecx, %ecx
because both %eax and %edx are dead at this point:
43: 8b 44 24 24 mov 0x24(%esp),%eax
47: f7 e6 mul %esi
49: 01 d3 add %edx,%ebx
4b: 8b 54 24 14 mov 0x14(%esp),%edx
...</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>