<html>
<head>
<base href="http://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 --- - Pathological register allocation within gigantic single basic block function (x86-64)"
href="http://llvm.org/bugs/show_bug.cgi?id=20014">20014</a>
</td>
</tr>
<tr>
<th>Summary</th>
<td>Pathological register allocation within gigantic single basic block function (x86-64)
</td>
</tr>
<tr>
<th>Product</th>
<td>new-bugs
</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>new bugs
</td>
</tr>
<tr>
<th>Assignee</th>
<td>unassignedbugs@nondot.org
</td>
</tr>
<tr>
<th>Reporter</th>
<td>vlad@drpetric.com
</td>
</tr>
<tr>
<th>CC</th>
<td>llvmbugs@cs.uiuc.edu
</td>
</tr>
<tr>
<th>Classification</th>
<td>Unclassified
</td>
</tr></table>
<p>
<div>
<pre>The code in question is from an Othello minimax engine.
<a href="https://github.com/vladpetric/ntest/branches">https://github.com/vladpetric/ntest/branches</a> (jmobs2 branch)
The static evaluation function of ntest is a huge basic block, of around 550
instructions. It's a complex basic block, with a lot of live values.
Unfortunately, clang produces no fewer than 75 spills! I count a spill as a
store/load pair to the stack, which includes pushes, pops, moves and other
instructions with an operand in memory and using %rsp (callee-managed registers
are included here as well). At the same time gcc 4.8.2 generates 9, and icc
only 4. Given that this is a single linear basic block, all instructions are
executed exactly the same number of times.
Instruction statistics for the function CEvaluator::EvalMobs :
gcc 4.8.2:
Total st stores loads st loads moves alu mult popcnt
547 9 122 9 117 261 28 1
icc:
Total st stores loads st loads moves alu mult popcnt
541 4 122 4 122 260 28 1
clang (3.5 trunk):
Total st stores loads st loads moves alu mult popcnt
704 76 122 75 81 321 28 1
A few notes:
* there are no branches or non-stack stores. st loads mean stack loads, st
stores mean stack stores.
* non-stack loads are loads from evaluation tables. The mults come primarily
from the "magic multiplier" bit gather trick. popcnt - well, that's explicitly
called. There's a constant number of them.
* an x86_64 ALU instruction that has a memory operand is counted twice - both
ALU and load. Such operations are likely broken into two micro-ops, generally
speaking (one load and one alu).
The code can be compiled by fetching the sourcecode, then a straightfoward
cmake project, but I've also attached the .ll file which includes the
pathological function (CEvaluator::EvalMobs).</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>