<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 --- - extra memory load in loop"
href="https://llvm.org/bugs/show_bug.cgi?id=27740">27740</a>
</td>
</tr>
<tr>
<th>Summary</th>
<td>extra memory load in loop
</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>carrot@google.com
</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>Compile following code with trunk llvm
$ ~/llvm/obj2/bin/clang --target=powerpc64le-grtev4-linux-gnu -O2
-fno-unroll-loops -c t17.c -save-temps
struct s{
int f0;
int f1;
float f2;
float f3;
float f4;
};
static void bar(struct s *p1, struct s *p2)
{
if (p1->f0 && p1->f3 < p2->f3)
p1->f3 = p2->f3;
else
p1->f3 = p2->f3; // p1->f4 = p2->f4;
}
void foo(struct s *p1, struct s *p2, int size)
{
for (int i=0; i<size; i++)
{
p1->f1 += p2->f1;
p1->f2 -= p2->f2;
bar(p1, p2);
p2++;
}
}
LLVM generates following code for the loop body
.LBB0_2: # %for.body
# =>This Inner Loop Header: Depth=1
lwzu 9, 20(4) // A1
lxsspx 1, 3, 7 // B1
lwz 10, 4(3) // A2
lxsspx 0, 4, 6 // B2
add 9, 10, 9 // A3
stw 9, 4(3) // A4
xssubsp 0, 1, 0 // B3
stxsspx 0, 3, 7 // B4
bc 12, 2, .LBB0_5
# BB#3: # %land.lhs.true.i
# in Loop: Header=BB0_2 Depth=1
lxsspx 0, 4, 7
lxsspx 1, 3, 8
fcmpu 1, 1, 0
bge 1, .LBB0_6
# BB#4: # %if.then.i
# in Loop: Header=BB0_2 Depth=1
stxsspx 0, 3, 8
b .LBB0_8
.p2align 4
.LBB0_5: # %entry.if.else_crit_edge.i
# in Loop: Header=BB0_2 Depth=1
lwz 9, 8(4)
b .LBB0_7
.p2align 4
.LBB0_6: # %land.lhs.true.if.else_crit_edge.i
# in Loop: Header=BB0_2 Depth=1
xscvdpspn 0, 0
xxsldwi 0, 0, 0, 3
mfvsrwz 9, 0
.LBB0_7: # %if.else.i
# in Loop: Header=BB0_2 Depth=1
stw 9, 12(3)
.LBB0_8: # %bar.exit
# in Loop: Header=BB0_2 Depth=1
addi 5, 5, -1
bdnz .LBB0_2
Statement p1->f1 += p2->f1 is translated to A1,A2,A3,A4. Statement p1->f2 -=
p2->f2 is translated to B1,B2,B3,B4. Inside the loop, all stores are through
pointer p1 only, although p1 can be aliased with p2, it is safe to keep a copy
of p1->f1 and p1->f2 inside registers, so A2 and B1 can be removed from the
loop.
In the source code, if I replace one of "p1->f3 = p2->f3" by "p1->f4 = p2->f4",
then llvm can generate expected optimized code.</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>