<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 - ARC optimizations interact badly with catchswitch"
href="https://bugs.llvm.org/show_bug.cgi?id=37332">37332</a>
</td>
</tr>
<tr>
<th>Summary</th>
<td>ARC optimizations interact badly with catchswitch
</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>All
</td>
</tr>
<tr>
<th>Status</th>
<td>NEW
</td>
</tr>
<tr>
<th>Severity</th>
<td>enhancement
</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>smeenai@fb.com
</td>
</tr>
<tr>
<th>CC</th>
<td>ahatanak@gmail.com, compnerd@compnerd.org, llvm-bugs@lists.llvm.org, rjmccall@apple.com, rnk@google.com
</td>
</tr></table>
<p>
<div>
<pre>Created <span class=""><a href="attachment.cgi?id=20260" name="attach_20260" title="Reduced IR">attachment 20260</a> <a href="attachment.cgi?id=20260&action=edit" title="Reduced IR">[details]</a></span>
Reduced IR
If you run the attached IR with opt -objc-arc, the verifier will complain
because ARC optimizations attempted to insert an objc_release before a
catchswitch, but a catchswitch must be the only non-PHI instruction in its
basic block. Specifically, the ARC optimizations were attempting to duplicate
an objc_release across both edges of an invoke (presumably to shorten a
lifetime), so they wanted an insertion point that was guaranteed to be reached
from the unwind edge of the invoke; unfortunately, with the funclet-style
exception instructions, there's no such guaranteed insertion point in the
existing CFG in the general case.
One way I see of fixing this is splitting the invoke edge if it goes to a
catchswitch and inserting a cleanuppad to hold the release. E.g. in my
particular example the transformed IR would be:
%call = invoke i8* @f(i8* %p, i8* %q)
to label %invoke.cont unwind label %ehcleanup
ehcleanup:
%tmp = cleanuppad within none []
tail call void @objc_release(i8* %p) [ "funclet"(token %tmp) ]
cleanupret from %tmp to label %catch.dispatch
%catch.dispatch:
...
I think this is the "right" thing to do in some sense, but it also involves
changing the CFG, which is probably going to be fraught with pain. I admittedly
haven't looked too much into how much work it would be to keep the analysis
correct in the face of a CFG change, but I'm guessing there's a reason the
existing transform explicitly avoids splitting edges :) (See
<a href="https://reviews.llvm.org/diffusion/L/browse/llvm/trunk/lib/Transforms/ObjCARC/PtrState.cpp;331493$262-272">https://reviews.llvm.org/diffusion/L/browse/llvm/trunk/lib/Transforms/ObjCARC/PtrState.cpp;331493$262-272</a>.)
The other option is to just bail out of the objc_release movement if there's a
catchswitch in the way. Unfortunately, I can't figure out where to do that
check. Marking SetCFGHazardAfflicted inside
BottomUpPtrState::HandlePotentialUse doesn't seem to do the trick; I think the
CFGHazardAfflicted state somehow gets lost at some point during merging. I can
check for the presence of a catchswitch inside ObjCARCOpt::CheckForCFGHazards,
but that seems overly conservative, since we only care if the catchswitch is a
potential insertion point. Any help on that front would be very appreciated.
(I'm also not sure why the retain/release for %p isn't just elided entirely in
this particular example. I have a separate reduction for that, and I'll file a
different bug.)</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>