<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 --- - Different index types in GEPs -> non-aliasing"
href="https://llvm.org/bugs/show_bug.cgi?id=27418">27418</a>
</td>
</tr>
<tr>
<th>Summary</th>
<td>Different index types in GEPs -> non-aliasing
</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>Global Analyses
</td>
</tr>
<tr>
<th>Assignee</th>
<td>unassignedbugs@nondot.org
</td>
</tr>
<tr>
<th>Reporter</th>
<td>mikael.holmen@ericsson.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>Created <span class=""><a href="attachment.cgi?id=16233" name="attach_16233" title="Reproducer ll file">attachment 16233</a> <a href="attachment.cgi?id=16233&action=edit" title="Reproducer ll file">[details]</a></span>
Reproducer ll file
It seems that opt thinks that the two pointers
%_tmp2 = getelementptr [3 x i16], [3 x i16]* %a, i16 0, i64 1
and
%_tmp4 = getelementptr [3 x i16], [3 x i16]* %a, i16 0, i16 1
does not alias? Is this intended or a bug?
Details below:
--------------
I found this when I ran opt on:
define i16 @f () {
%a = alloca [3 x i16]
; Write 98 at index 1 in the array.
; NB: using i64 as type of the index argument!
%_tmp2 = getelementptr [3 x i16], [3 x i16]* %a, i16 0, i64 1
store i16 98, i16* %_tmp2
; Read at index 1 in the array
; NB: using i16 as type of the index argument!
%_tmp4 = getelementptr [3 x i16], [3 x i16]* %a, i16 0, i16 1
%_tmp5 = load i16, i16* %_tmp4
; Check read value. If not 98, jump to bb1
%_tmp6 = icmp ne i16 %_tmp5, 98
br i1 %_tmp6, label %bb1, label %bb2
bb1:
; Error
ret i16 1;
bb2:
; Ok
ret i16 0;
}
opt -S -gvn foo.ll
and I got the result
; ModuleID = 'foo.ll'
define i16 @f() {
%a = alloca [3 x i16]
%_tmp2 = getelementptr [3 x i16], [3 x i16]* %a, i16 0, i64 1
store i16 98, i16* %_tmp2
%_tmp4 = getelementptr [3 x i16], [3 x i16]* %a, i16 0, i16 1
br i1 undef, label %bb1, label %bb2
bb1: ; preds = %0
ret i16 1
bb2: ; preds = %0
ret i16 0
}
Note the "undef" in the branch which GVN put there:
GVN iteration: 0
GVN removed: %_tmp5 = load i16, i16* %_tmp4
GVN removed: %_tmp6 = icmp ne i16 undef, 98
When digging into why GVN does this, I found this little piece of code in
GVN::AnalyzeLoadAvailability that triggers
// Loading the allocation -> undef.
if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI) ||
// Loading immediately after lifetime begin -> undef.
isLifetimeStart(DepInst)) {
Res = AvailableValue::get(UndefValue::get(LI->getType()));
return true;
}
And the DepInst here is
%a = alloca [3 x i16]
and not the store
store i16 98, i16* %_tmp2
Digging further it seems the reason for this is that the alias analysis
considers the two pointers used to store and load the value 98 are
non-aliasing, even if both the pointer and the indices has the same values. The
only difference is the type of the last value.
In e.g. BasicAliasAnalysis.cpp we have
bool BasicAAResult::isValueEqualInPotentialCycles(const Value *V,
const Value *V2) {
if (V != V2)
return false;
and in
aliasSameBasePointerGEPs:
// If the last (struct) indices are constants and are equal, the other
indices
// might be also be dynamically equal, so the GEPs can alias.
if (C1 && C2 && C1 == C2)
return MayAlias;
so here we consider two pointers to be completely different just because the
types of the last indices differ.</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>