<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 --- - LICM::sink has exponential complexity on certain inputs"
href="http://llvm.org/bugs/show_bug.cgi?id=19835">19835</a>
</td>
</tr>
<tr>
<th>Summary</th>
<td>LICM::sink has exponential complexity on certain inputs
</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>Scalar Optimizations
</td>
</tr>
<tr>
<th>Assignee</th>
<td>unassignedbugs@nondot.org
</td>
</tr>
<tr>
<th>Reporter</th>
<td>eugeni.stepanov@gmail.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>$ cat 1.cc
#define Q(i, j) b##j = a##j + x; a##i = a##j * b##j;
#define D(i) int a##i; int b##i;
int f(int x) {
D(0)
D(1)
D(2)
D(3)
D(4)
D(5)
D(6)
D(7)
D(8)
D(9)
for (int i = 0; i < 100; ++i) {
int volatile z = x;
int a0 = z;
Q(1, 0)
Q(2, 1)
Q(3, 2)
Q(4, 3)
Q(5, 4)
Q(6, 5)
Q(7, 6)
Q(8, 7)
Q(9, 8)
}
return a9;
}
$ ./bin/clang++ -S -emit-llvm -O1 1.cc -o - |& grep '%.*=' | wc -l
1027
Playing with the number of macro calls in the above source, I see that the
number of instructions in the IR is 2**(N+1) + 3.
IR itself contains lots of duplicate instructions:
define i32 @_Z1fi(i32 %x) #0 {
entry:
%z = alloca i32, align 4
br label %for.body
for.body: ; preds = %for.body, %entry
%i.033 = phi i32 [ 0, %entry ], [ %inc, %for.body ]
store volatile i32 %x, i32* %z, align 4
%z.0.z.0. = load volatile i32* %z, align 4
%inc = add nsw i32 %i.033, 1
%exitcond = icmp eq i32 %inc, 100
br i1 %exitcond, label %for.end, label %for.body
for.end: ; preds = %for.body
%add.le1275 = add nsw i32 %z.0.z.0., %x
%add.le1273 = add nsw i32 %z.0.z.0., %x
%add.le1271 = add nsw i32 %z.0.z.0., %x
%add.le1269 = add nsw i32 %z.0.z.0., %x
%add.le1267 = add nsw i32 %z.0.z.0., %x
%add.le1265 = add nsw i32 %z.0.z.0., %x
%add.le1263 = add nsw i32 %z.0.z.0., %x
%add.le1261 = add nsw i32 %z.0.z.0., %x
%add.le1259 = add nsw i32 %z.0.z.0., %x
%add.le1257 = add nsw i32 %z.0.z.0., %x
%add.le1255 = add nsw i32 %z.0.z.0., %x
%add.le1253 = add nsw i32 %z.0.z.0., %x
%add.le1251 = add nsw i32 %z.0.z.0., %x
%add.le1249 = add nsw i32 %z.0.z.0., %x
%add.le1247 = add nsw i32 %z.0.z.0., %x
%add.le1245 = add nsw i32 %z.0.z.0., %x
%add.le1243 = add nsw i32 %z.0.z.0., %x
%add.le1241 = add nsw i32 %z.0.z.0., %x
%add.le1239 = add nsw i32 %z.0.z.0., %x
%add.le1237 = add nsw i32 %z.0.z.0., %x
%add.le1235 = add nsw i32 %z.0.z.0., %x
%add.le1233 = add nsw i32 %z.0.z.0., %x
...
%add.le = add nsw i32 %z.0.z.0., %x
%mul.le1019 = mul nsw i32 %add.le, %z.0.z.0.
%mul.le1016 = mul nsw i32 %add.le1023, %z.0.z.0.
%mul.le1013 = mul nsw i32 %add.le1025, %z.0.z.0.
%mul.le1010 = mul nsw i32 %add.le1027, %z.0.z.0.
%mul.le1007 = mul nsw i32 %add.le1029, %z.0.z.0.
%mul.le1004 = mul nsw i32 %add.le1031, %z.0.z.0.
%mul.le1001 = mul nsw i32 %add.le1033, %z.0.z.0.
%mul.le998 = mul nsw i32 %add.le1035, %z.0.z.0.
%mul.le995 = mul nsw i32 %add.le1037, %z.0.z.0.
%mul.le992 = mul nsw i32 %add.le1039, %z.0.z.0.
%mul.le989 = mul nsw i32 %add.le1041, %z.0.z.0.
%mul.le986 = mul nsw i32 %add.le1043, %z.0.z.0.
%mul.le983 = mul nsw i32 %add.le1045, %z.0.z.0.
%mul.le980 = mul nsw i32 %add.le1047, %z.0.z.0.
%mul.le977 = mul nsw i32 %add.le1049, %z.0.z.0.
%mul.le974 = mul nsw i32 %add.le1051, %z.0.z.0.
%mul.le971 = mul nsw i32 %add.le1053, %z.0.z.0.
...
The problem is in LICM::sink, which does not handle the following pattern well:
f = op g
e = op f, g
d = op e
c = op d, e
b = op c
a = op b, c
When an instruction with N uses is sunk, each of its operands get N new uses
(all of them - phi nodes). In the example above, if a had 1 use, c would have
2, e would have 4, and g would have 8.
This has regressed in r201148:
<a href="http://llvm.org/viewvc/llvm-project?rev=201148&view=rev">http://llvm.org/viewvc/llvm-project?rev=201148&view=rev</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>