<html>
<head>
<meta http-equiv="Content-Type" content="text/html; charset=UTF-8">
</head>
<body text="#000000" bgcolor="#FFFFFF">
<p>All of the changes in this sequence have landed, and the default
for X86 has been flipped in git 246098.</p>
<p>There will be some cleanup work and further optimization work in
a couple of weeks, but I'm going to give this a bit of burn-in
period before doing anything else to give problems time to be
reported.</p>
<p>Philip<br>
</p>
<div class="moz-cite-prefix">On 8/28/19 1:40 PM, Philip Reames via
llvm-dev wrote:<br>
</div>
<blockquote type="cite"
cite="mid:02f0536d-d7d7-5030-f5d8-caf29f279e96@philipreames.com">
<meta http-equiv="content-type" content="text/html; charset=UTF-8">
<p>I have a set of changes out for review which are possibly note
worthy, and backend contributors may wish to be aware of. <br>
</p>
<p>TLDR: atomic loads as normal LoadSDNodes w/an "isAtomic" flag.</p>
<p><br>
</p>
<p>Background</p>
<p>At the moment, we lower all atomic loads and stores as
instances of AtomicSDNode (along with cmpxchg, and atomicrmw).
This requires us to duplicate any isel rules we wish to apply
for atomic loads or stores, but does have the nice property
that it's harder to introduce a silent miscompile by adding an
transform which forgets about atomicity.</p>
<p><br>
</p>
<p>Proposed End Result<br>
</p>
<p>Represent atomic loads and stores as normal LoadSDNode or
StoreSDNodes. Analogously to volatility, provide a flag on the
node (stored in the MMO) which indicates whether the operation
is atomic. All transforms updated to check isAtomic if needed.
<br>
</p>
<p>The advantages of this representations are:<br>
1) Once the audit has been done, it makes it easier to keep
atomic and non-atomic rules in sync. <br>
2) It makes GlobalISEL easier (by eliminating the need for the
special case).<br>
3) Unify patterns flowing through other backend passes (i.e.
unordered atomics and non-atomics shouldn't generate radically
different MI structures)<br>
</p>
<p>One open question is whether we do this just for unordered
atomics, or for all atomics. I'd be open to either, but would
start with just unordered to start with either way. <br>
</p>
<p><br>
</p>
<p>Migration Plan</p>
<p>This would be done on a per-backend basis, and to start with,
I'm only proposing to port X86. <br>
</p>
<p>The basic strategy I plan on taking is:</p>
<ol class="remarkup-list">
<li class="remarkup-list-item">introduce infrastructure and a
flag for testing (<a class="moz-txt-link-freetext"
href="https://reviews.llvm.org/D66309"
moz-do-not-send="true">https://reviews.llvm.org/D66309</a>)</li>
<li class="remarkup-list-item">Audit uses of isVolatile, and
apply isAtomic conservatively*</li>
<li class="remarkup-list-item">piecemeal conservative* update
generic code and x86 backedge code in individual reviews
w/tests for cases which didn't check volatile, but can be
found with inspection</li>
<li class="remarkup-list-item">flip the (x86) flag at the end
(with minimal diffs)</li>
<li class="remarkup-list-item">Work through todo list identified
in (2) and (3) exposing performance ops</li>
</ol>
<p>(*) The "conservative" bit here is aimed at minimizing the
number of diffs involved in (4). Ideally, there'd be none. In
practice, getting it down to something reviewable by a human is
the actual goal. Note that there are (currently) no paths which
produce LoadSDNode or StoreSDNode with atomic MMOs, so we don't
need to worry about preserving any behavior there.</p>
<p>We've taken a very similar strategy twice before with success -
once at IR level, and once at the MI level (post ISEL). I'll
probably need some help with some of the ISEL patterns since
that's the part I'm not familiar with.</p>
<p><br>
</p>
<p>Thoughts?</p>
<p>Philip<br>
</p>
<br>
<fieldset class="mimeAttachmentHeader"></fieldset>
<pre class="moz-quote-pre" wrap="">_______________________________________________
LLVM Developers mailing list
<a class="moz-txt-link-abbreviated" href="mailto:llvm-dev@lists.llvm.org">llvm-dev@lists.llvm.org</a>
<a class="moz-txt-link-freetext" href="https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev">https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev</a>
</pre>
</blockquote>
</body>
</html>