<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 --- - Bitcasts block copy forwarding"
href="https://llvm.org/bugs/show_bug.cgi?id=31697">31697</a>
</td>
</tr>
<tr>
<th>Summary</th>
<td>Bitcasts block copy forwarding
</td>
</tr>
<tr>
<th>Product</th>
<td>libraries
</td>
</tr>
<tr>
<th>Version</th>
<td>trunk
</td>
</tr>
<tr>
<th>Hardware</th>
<td>All
</td>
</tr>
<tr>
<th>OS</th>
<td>All
</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>Global Analyses
</td>
</tr>
<tr>
<th>Assignee</th>
<td>unassignedbugs@nondot.org
</td>
</tr>
<tr>
<th>Reporter</th>
<td>arcata@gmail.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>If you have a situation where memory pointed to by a readonly argument %x is
copied to a local alloca %y and then %y is projected from, then LLVM normally
eliminates the unnecessary copy. This IR:
declare void @llvm.memcpy.p0i8.p0i8.i32(i8*, i8*, i32, i32, i1)
declare void @llvm.lifetime.start(i64, i8*)
declare void @llvm.lifetime.end(i64, i8*)
define i32 @copy_then_extract({ i32,i32,i32,i32 }* noalias nocapture readonly
dereferenceable(16) %x, i32 %n) {
entry:
%y = alloca { i32, i32, i32, i32 }, align 16
%x0p = getelementptr inbounds { i32, i32, i32, i32 }, { i32, i32, i32, i32 }*
%x, i32 0, i32 0
%x0 = load i32, i32* %x0p
%x1p = getelementptr inbounds { i32, i32, i32, i32 }, { i32, i32, i32, i32 }*
%x, i32 0, i32 1
%x1 = load i32, i32* %x1p
%x2p = getelementptr inbounds { i32, i32, i32, i32 }, { i32, i32, i32, i32 }*
%x, i32 0, i32 2
%x2 = load i32, i32* %x2p
%x3p = getelementptr inbounds { i32, i32, i32, i32 }, { i32, i32, i32, i32 }*
%x, i32 0, i32 3
%x3 = load i32, i32* %x3p
%yi8 = bitcast { i32, i32, i32, i32 }* %y to i8*
call void @llvm.lifetime.start(i64 16, i8* %yi8)
%y0p = getelementptr inbounds { i32, i32, i32, i32 }, { i32, i32, i32, i32 }*
%y, i32 0, i32 0
store i32 %x0, i32* %y0p
%y1p = getelementptr inbounds { i32, i32, i32, i32 }, { i32, i32, i32, i32 }*
%y, i32 0, i32 1
store i32 %x1, i32* %y1p
%y2p = getelementptr inbounds { i32, i32, i32, i32 }, { i32, i32, i32, i32 }*
%y, i32 0, i32 2
store i32 %x2, i32* %y2p
%y3p = getelementptr inbounds { i32, i32, i32, i32 }, { i32, i32, i32, i32 }*
%y, i32 0, i32 3
store i32 %x3, i32* %y3p
%ya = bitcast i8* %yi8 to i32*
%ya2p = getelementptr inbounds i32, i32* %ya, i64 2
%y2 = load i32, i32* %ya2p
%yi8_2 = bitcast { i32, i32, i32, i32 }* %y to i8*
call void @llvm.lifetime.end(i64 16, i8* %yi8_2)
ret i32 %y2
}
reduces to:
define i32 @copy_then_extract({ i32, i32, i32, i32 }* noalias nocapture
readonly dereferenceable(16) %x, i32 %n) local_unnamed_addr #0 {
entry:
%x2p = getelementptr inbounds { i32, i32, i32, i32 }, { i32, i32, i32, i32 }*
%x, i64 0, i32 2
%x2 = load i32, i32* %x2p, align 4
ret i32 %x2
}
However, if %ya2p becomes a dynamic GEP by changing `i64 2` to `i64 %n`, then
the bitcast can no longer be reduced since that's not valid against a struct.
This ends up completely blocking copy forwarding. This IR:
declare void @llvm.memcpy.p0i8.p0i8.i32(i8*, i8*, i32, i32, i1)
declare void @llvm.lifetime.start(i64, i8*)
declare void @llvm.lifetime.end(i64, i8*)
define i32 @copy_then_extract({ i32,i32,i32,i32 }* noalias nocapture readonly
dereferenceable(16) %x, i32 %n) {
entry:
%y = alloca { i32, i32, i32, i32 }, align 16
%x0p = getelementptr inbounds { i32, i32, i32, i32 }, { i32, i32, i32, i32 }*
%x, i32 0, i32 0
%x0 = load i32, i32* %x0p
%x1p = getelementptr inbounds { i32, i32, i32, i32 }, { i32, i32, i32, i32 }*
%x, i32 0, i32 1
%x1 = load i32, i32* %x1p
%x2p = getelementptr inbounds { i32, i32, i32, i32 }, { i32, i32, i32, i32 }*
%x, i32 0, i32 2
%x2 = load i32, i32* %x2p
%x3p = getelementptr inbounds { i32, i32, i32, i32 }, { i32, i32, i32, i32 }*
%x, i32 0, i32 3
%x3 = load i32, i32* %x3p
%yi8 = bitcast { i32, i32, i32, i32 }* %y to i8*
call void @llvm.lifetime.start(i64 16, i8* %yi8)
%y0p = getelementptr inbounds { i32, i32, i32, i32 }, { i32, i32, i32, i32 }*
%y, i32 0, i32 0
store i32 %x0, i32* %y0p
%y1p = getelementptr inbounds { i32, i32, i32, i32 }, { i32, i32, i32, i32 }*
%y, i32 0, i32 1
store i32 %x1, i32* %y1p
%y2p = getelementptr inbounds { i32, i32, i32, i32 }, { i32, i32, i32, i32 }*
%y, i32 0, i32 2
store i32 %x2, i32* %y2p
%y3p = getelementptr inbounds { i32, i32, i32, i32 }, { i32, i32, i32, i32 }*
%y, i32 0, i32 3
store i32 %x3, i32* %y3p
%ya = bitcast i8* %yi8 to i32*
%ya2p = getelementptr inbounds i32, i32* %ya, i32 %n
%y2 = load i32, i32* %ya2p
%yi8_2 = bitcast { i32, i32, i32, i32 }* %y to i8*
call void @llvm.lifetime.end(i64 16, i8* %yi8_2)
ret i32 %y2
}
still keeps the unnecessary copy in the optimized IR:
define i32 @copy_then_extract({ i32, i32, i32, i32 }* noalias nocapture
readonly dereferenceable(16) %x, i32 %n) local_unnamed_addr #1 {
entry:
%y = alloca <4 x i32>, align 16
%0 = bitcast { i32, i32, i32, i32 }* %x to <4 x i32>*
%1 = load <4 x i32>, <4 x i32>* %0, align 4
%yi8 = bitcast <4 x i32>* %y to i8*
call void @llvm.lifetime.start(i64 16, i8* %yi8)
store <4 x i32> %1, <4 x i32>* %y, align 16
%2 = sext i32 %n to i64
%ya2p = getelementptr inbounds <4 x i32>, <4 x i32>* %y, i64 0, i64 %2
%y2 = load i32, i32* %ya2p, align 4
call void @llvm.lifetime.end(i64 16, i8* %yi8)
ret i32 %y2
}</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>