<html>
<head>
<base href="https://bugs.llvm.org/">
</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 - Interaction between C's restrict pointer and pointer casted from integer is ambiguous in LLVM's LangRef"
href="https://bugs.llvm.org/show_bug.cgi?id=39846">39846</a>
</td>
</tr>
<tr>
<th>Summary</th>
<td>Interaction between C's restrict pointer and pointer casted from integer is ambiguous in LLVM's LangRef
</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>juneyoung.lee@sf.snu.ac.kr
</td>
</tr>
<tr>
<th>CC</th>
<td>llvm-bugs@lists.llvm.org
</td>
</tr></table>
<p>
<div>
<pre>Created <span class=""><a href="attachment.cgi?id=21180" name="attach_21180" title="The example code">attachment 21180</a> <a href="attachment.cgi?id=21180&action=edit" title="The example code">[details]</a></span>
The example code
Consider this example:
#include <stdint.h>
int foo(int *restrict x, int *restrict y) {
*y = 23;
uintptr_t xi = (uintptr_t)x;
uintptr_t yi = (uintptr_t)y;
if (xi != yi) { // xi == yi, so this is never taken
yi = xi;
}
*(int*)yi = 42; // changes *y or UB?
return *y;
}
int bar() {
int x;
return foo(&x, &x); // returns 23
}
After -O3, foo(&x, &x) is optimized into 'ret i32 23' (
<a href="https://godbolt.org/z/ty73Xk">https://godbolt.org/z/ty73Xk</a> ).
Here are several ways to explain this result:
(1) foo(&x, &x) is not allowed in LLVM IR - but I believe this makes the
semantics more undefined than C (because C allows multiple restrict pointers to
point have the same address if they are only read, but never written)
(2) *(int*)yi = 42 is not allowed - but LLVM IR LangRef's pointer aliasing
rules ( <a href="https://llvm.org/docs/LangRef.html#pointer-aliasing-rules">https://llvm.org/docs/LangRef.html#pointer-aliasing-rules</a> ) say that
(int *)yi is based on y, so I think this behavior is well-defined. Perhaps we
should update the definition of based-on relation, if we follow this option.
(3) Restrict pointer cannot be casted into integer - if we choose this option,
LangRef should be updated to deal with this.
If we're to choose (2) and (3), we should update LLVM LangRef.
If this is considered to be miscompilation, I think this is related to
<a class="bz_bug_link
bz_status_NEW "
title="NEW - InstCombine cannot blindly assume that inttoptr(ptrtoint x) -> x"
href="show_bug.cgi?id=34548">https://bugs.llvm.org/show_bug.cgi?id=34548</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>