<html xmlns:v="urn:schemas-microsoft-com:vml" xmlns:o="urn:schemas-microsoft-com:office:office" xmlns:w="urn:schemas-microsoft-com:office:word" xmlns:m="http://schemas.microsoft.com/office/2004/12/omml" xmlns="http://www.w3.org/TR/REC-html40">
<head>
<meta http-equiv="Content-Type" content="text/html; charset=utf-8">
<meta name="Generator" content="Microsoft Word 14 (filtered medium)">
<style><!--
/* Font Definitions */
@font-face
{font-family:Calibri;
panose-1:2 15 5 2 2 2 4 3 2 4;}
@font-face
{font-family:Tahoma;
panose-1:2 11 6 4 3 5 4 4 2 4;}
/* Style Definitions */
p.MsoNormal, li.MsoNormal, div.MsoNormal
{margin:0cm;
margin-bottom:.0001pt;
font-size:12.0pt;
font-family:"Times New Roman","serif";}
a:link, span.MsoHyperlink
{mso-style-priority:99;
color:blue;
text-decoration:underline;}
a:visited, span.MsoHyperlinkFollowed
{mso-style-priority:99;
color:purple;
text-decoration:underline;}
p
{mso-style-priority:99;
mso-margin-top-alt:auto;
margin-right:0cm;
mso-margin-bottom-alt:auto;
margin-left:0cm;
font-size:12.0pt;
font-family:"Times New Roman","serif";}
span.EmailStyle18
{mso-style-type:personal-reply;
font-family:"Calibri","sans-serif";
color:#1F497D;}
.MsoChpDefault
{mso-style-type:export-only;
font-family:"Calibri","sans-serif";
mso-fareast-language:EN-US;}
@page WordSection1
{size:612.0pt 792.0pt;
margin:72.0pt 72.0pt 72.0pt 72.0pt;}
div.WordSection1
{page:WordSection1;}
/* List Definitions */
@list l0
{mso-list-id:615599737;
mso-list-template-ids:-1886478414;}
ol
{margin-bottom:0cm;}
ul
{margin-bottom:0cm;}
--></style><!--[if gte mso 9]><xml>
<o:shapedefaults v:ext="edit" spidmax="1026" />
</xml><![endif]--><!--[if gte mso 9]><xml>
<o:shapelayout v:ext="edit">
<o:idmap v:ext="edit" data="1" />
</o:shapelayout></xml><![endif]-->
</head>
<body lang="EN-GB" link="blue" vlink="purple">
<div class="WordSection1">
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D">For doing PRE on the load, it looks like there’s only two things stopping GVN PRE from doing it:<o:p></o:p></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D">* GVN::PerformLoadPRE doesn’t hoist loads that are conditional. Probably this can be overcome with some kind of<o:p></o:p></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D"> heuristic that allows it to happen in loops when the blocks where a load would have to be inserted are outside<o:p></o:p></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D"> the loop.<o:p></o:p></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D">* IsFullyAvailableInBlock goes around the loop and double-counts the entry-edge into the loop as unavailable. I’m<o:p></o:p></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D"> pretty sure this can be easily fixed by marking the block that PerformLoadPRE calls IsFullyAvailableInBlock on as<o:p></o:p></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D"> FullyAvailable, so it stops there when following the back-edge.<o:p></o:p></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D">I did a quick experiment (solving the first problem by just commenting out that check) and this worked for a simple<o:p></o:p></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D">test case of<o:p></o:p></span></p>
<p class="MsoNormal" style="text-autospace:none"><span style="font-size:11.0pt;font-family:"Courier New";mso-fareast-language:EN-US">int x;<o:p></o:p></span></p>
<p class="MsoNormal" style="text-autospace:none"><span style="font-size:11.0pt;font-family:"Courier New";mso-fareast-language:EN-US">int arr[32];<o:p></o:p></span></p>
<p class="MsoNormal" style="text-autospace:none"><span style="font-size:11.0pt;font-family:"Courier New";mso-fareast-language:EN-US">void fn(int n) {<o:p></o:p></span></p>
<p class="MsoNormal" style="text-autospace:none"><span style="font-size:11.0pt;font-family:"Courier New";mso-fareast-language:EN-US"> for (int i = 0; i < n; i++) {<o:p></o:p></span></p>
<p class="MsoNormal" style="text-autospace:none"><span style="font-size:11.0pt;font-family:"Courier New";mso-fareast-language:EN-US"> if (arr[i]) {<o:p></o:p></span></p>
<p class="MsoNormal" style="text-autospace:none"><span style="font-size:11.0pt;font-family:"Courier New";mso-fareast-language:EN-US"> x++;<o:p></o:p></span></p>
<p class="MsoNormal" style="text-autospace:none"><span style="font-size:11.0pt;font-family:"Courier New";mso-fareast-language:EN-US"> }<o:p></o:p></span></p>
<p class="MsoNormal" style="text-autospace:none"><span style="font-size:11.0pt;font-family:"Courier New";mso-fareast-language:EN-US"> }<o:p></o:p></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Courier New";mso-fareast-language:EN-US">}</span><span style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D"><o:p></o:p></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D">So perhaps the load PRE half can be done without a separate pass.<o:p></o:p></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D"><o:p> </o:p></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D">John<o:p></o:p></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D"><o:p> </o:p></span></p>
<div style="border:none;border-left:solid blue 1.5pt;padding:0cm 0cm 0cm 4.0pt">
<div>
<div style="border:none;border-top:solid #B5C4DF 1.0pt;padding:3.0pt 0cm 0cm 0cm">
<p class="MsoNormal"><b><span lang="EN-US" style="font-size:10.0pt;font-family:"Tahoma","sans-serif"">From:</span></b><span lang="EN-US" style="font-size:10.0pt;font-family:"Tahoma","sans-serif""> llvm-dev [mailto:llvm-dev-bounces@lists.llvm.org]
<b>On Behalf Of </b>Alina Sbirlea via llvm-dev<br>
<b>Sent:</b> 13 September 2018 22:01<br>
<b>To:</b> listmail@philipreames.com<br>
<b>Cc:</b> llvm-dev@lists.llvm.org<br>
<b>Subject:</b> Re: [llvm-dev] Generalizing load/store promotion in LICM<o:p></o:p></span></p>
</div>
</div>
<p class="MsoNormal"><o:p> </o:p></p>
<div>
<p class="MsoNormal">For context, I've been looking into replacing the use of AST (AliasSetTracker) with MemorySSA in LICM. This works great for sinking/hoisting but does not apply well for promotion, and one of the solutions I considered is precisely ripping
out the promotion part and replacing its functionality with a separate PRE pass + possibly store sinking. FWIW I think that's the right way to go.<o:p></o:p></p>
<div>
<p class="MsoNormal">I did not get deep enough into working on the solution but I would gladly have a detailed discussion to move this forward.<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">Reading into detail now.<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">Thanks,<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal">Alina<o:p></o:p></p>
</div>
</div>
<p class="MsoNormal"><o:p> </o:p></p>
<div>
<div>
<p class="MsoNormal">On Thu, Sep 13, 2018 at 1:43 PM Philip Reames <<a href="mailto:listmail@philipreames.com">listmail@philipreames.com</a>> wrote:<o:p></o:p></p>
</div>
<blockquote style="border:none;border-left:solid #CCCCCC 1.0pt;padding:0cm 0cm 0cm 6.0pt;margin-left:4.8pt;margin-right:0cm">
<div>
<p class="MsoNormal">(minor inline additions)<br>
<br>
On 09/13/2018 01:51 AM, Chandler Carruth wrote:<br>
<br>
<o:p></o:p></p>
<div>
<p class="MsoNormal">Haven't had time to dig into this, but wanted to add <a href="mailto:asbirlea@google.com" target="_blank" id="m_-2310710337215009520IloFPc-1">+Alina Sbirlea</a> to the thread as she has been working on promotion and other aspects of LICM
for a long time here.<o:p></o:p></p>
</div>
<p class="MsoNormal">Thanks!<br>
<br>
<o:p></o:p></p>
<div>
<div>
<p class="MsoNormal">On Wed, Sep 12, 2018 at 11:41 PM Philip Reames <<a href="mailto:listmail@philipreames.com" target="_blank">listmail@philipreames.com</a>> wrote:<o:p></o:p></p>
</div>
<blockquote style="border:none;border-left:solid #CCCCCC 1.0pt;padding:0cm 0cm 0cm 6.0pt;margin-left:4.8pt;margin-right:0cm">
<div>
<p>I'm thinking about making some semi radical changes to load store promotion works in LICM, and figured it would be a good idea to get buy in before I actually started writing code. :)<o:p></o:p></p>
<p>TLDR: legality of sinking stores to exits is hard, can we separate load handling into a super aggressive form of PRE, and use predicated stores to avoid solving legality question?<o:p></o:p></p>
<p><o:p> </o:p></p>
<p>Background<o:p></o:p></p>
<p>We've been seeing an interesting class of problems recently that looks roughly like this:<o:p></o:p></p>
<p>for (int = 0; i < N; i++)<br>
if (a[i] == 0) // some data dependent check<br>
g_count++; // conditional load and store to shared location<o:p></o:p></p>
<p>The critical detail in this example is that g_count is a global location which may be accessed concurrently* by multiple threads. The challenge here is that we don't know whether the store ever executes, and we're not allowed to insert a store along any
path that didn't originally contain them. Said differently, if all elements in "a" are non-zero, we're not allowed to store to g_count. We do know that the g_count location is dereferenceable though.
<o:p></o:p></p>
<p>(*Please, let's avoid the memory model derailment here. I'm simplifying and C++ language rules aren't real useful for my Java language frontend anyways. In practice, all the access are atomic, but unordered, but we'll leave that out of discussion otherwise.)<o:p></o:p></p>
<p>I have two approaches I'm considering here. These are orthogonal, but I suspect we'll want to implement both.<o:p></o:p></p>
<p><o:p> </o:p></p>
<p>Proposal A - Use predicated stores in loop exits<o:p></o:p></p>
<p>The basic idea is that we don't worry about solving the legality question above, and just insert a store which is predicated on a condition which is true exactly when the original store ran. In pseudo code, this looks something like:<o:p></o:p></p>
<p>bool StoreLegal = false;<br>
int LocalCount = g_count;<br>
for (int = 0; i < N; i++)<br>
if (a[i] == 0) {<br>
LocalCount++;<br>
StoreLegal = true;<br>
}<br>
if (StoreLegal) g_count = LocalCount; <o:p></o:p></p>
<p>There are two obvious concerns here:<o:p></o:p></p>
<ol start="1" type="1">
<li class="MsoNormal" style="mso-margin-top-alt:auto;mso-margin-bottom-alt:auto;mso-list:l0 level1 lfo1">
The predicated store might be expensive in practice - true for most current architectures.<o:p></o:p></li><li class="MsoNormal" style="mso-margin-top-alt:auto;mso-margin-bottom-alt:auto;mso-list:l0 level1 lfo1">
We''re introducing an extra boolean phi cycle around the loop. <o:p></o:p></li></ol>
<p>Talking to a couple of folks offline at the socials over the last few months, the former seems to be the main objection. I think we can control this by restricting this transform to when the loop is believed to have a high trip count and the conditional
path is taken with some frequency. Getting feedback on this particular point is one of my main reasons for writing this mail.
<o:p></o:p></p>
<p>The second objection can frequently be resolved by finding other expressions which evaluate to the same boolean. (In this case, if LocalCount != LocalCountOrig assuming i doesn't overflow.) We already have a framework with SCEV to do these replacements.
Though, from some quick testing, it would definitely need strengthening. However, SCEV can't remove the extra phi in all cases, so we have to be okay with the extra phi cycle in the general case. This seems worthwhile to me, but thoughts?<o:p></o:p></p>
<p><o:p> </o:p></p>
<p>Proposal B - Separate load and store handling into distinct transforms<o:p></o:p></p>
<p>(For folks I've discussed this with before, this part is all new.)<o:p></o:p></p>
<p>Thinking about the problem, it finally occurred to me that we can decompose the original example into two steps: getting the loads out of the loop, and sinking the stores out of the loop. If we can accomplish the former, but not the later, we've still made
meaningful progress. <o:p></o:p></p>
<p>So, what'd we'd essentially have is a load only transformation which produces this:<br>
int LocalCount = g_count;<br>
for (int = 0; i < N; i++)<br>
if (a[i] == 0) {<br>
LocalCount++;<br>
g_count = LocalCount;<br>
}<o:p></o:p></p>
<p>At this point, we've reduced memory traffic by half, and opened up the possibility that other parts of the optimizer can exploit the simplified form. The last point is particular interesting since we generally try to canonicalize loads out of loops, and
much of the optimizer is tuned for a form with as much as possible being loop invariant. As one example, completely by accident, there's some work going on in the LoopVectorizer right now to handle such stores to loop invariant addresses during vectorization.
Putting the two pieces together would let us fully vectorize this loop without needing to sink the stores at all.
<o:p></o:p></p>
<p>In practice, the existing implementation in LICM would cleanly split along these lines with little problem.
<o:p></o:p></p>
<p>One way of looking at this load specific transform is as an extreme form of PRE (partial redundancy elimination). Our current PRE implementation doesn't handle cases this complicated.
<o:p></o:p></p>
</div>
</blockquote>
</div>
<p class="MsoNormal">It occurred to my later that simply framing the new transform as a separate pass (LoopPRE) and using the same AST + SSA construction approach would be straight forward. So, if folks think that having an aggressive form of load PRE in LICM
is going a bit too far, it'd be easy to represent as an optional separate pass. I'd still prefer having LICM contain the logic though.
<br>
<br>
<o:p></o:p></p>
<div>
<blockquote style="border:none;border-left:solid #CCCCCC 1.0pt;padding:0cm 0cm 0cm 6.0pt;margin-left:4.8pt;margin-right:0cm">
<div>
<p>Thoughts?<o:p></o:p></p>
</div>
<div>
<p>Philip<o:p></o:p></p>
</div>
</blockquote>
</div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
</blockquote>
</div>
</div>
</div>
</body>
</html>