<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=us-ascii">
<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;}
/* Style Definitions */
p.MsoNormal, li.MsoNormal, div.MsoNormal
{margin:0in;
margin-bottom:.0001pt;
font-size:11.0pt;
font-family:"Calibri","sans-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;}
span.EmailStyle17
{mso-style-type:personal-compose;
font-family:"Calibri","sans-serif";
color:windowtext;}
.MsoChpDefault
{mso-style-type:export-only;
font-family:"Calibri","sans-serif";}
@page WordSection1
{size:8.5in 11.0in;
margin:1.0in 1.0in 1.0in 1.0in;}
div.WordSection1
{page:WordSection1;}
--></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-US" link="blue" vlink="purple">
<div class="WordSection1">
<p class="MsoNormal">Hi,<o:p></o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal">We are proposing a loop fusion pass that tries to proactive fuse loops across function call boundaries and arbitrary control flow.<o:p></o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal"><a href="http://reviews.llvm.org/D7008">http://reviews.llvm.org/D7008</a><o:p></o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal">With this pass, we get 103 loop fusions in SPECCPU INT 2006 462.libquantum with rate performance improving close to 2.5X in x86 (results from AMD A10-6700).<o:p></o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal">I took some liberties in patching up some of the code in ScalarEvolution/DependenceAnalysis/GlobalsModRef/Clang options/ and also adjusted the IPO/LTO pass managers. I would need to do a better job there but I would like to put this out
as WIP for high level reviews. At this time standard loop fusion test cases may not work (cases where control dep are same or loops are truly adjacent etc). I would be doing that as well.<o:p></o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal">Appreciate suggestions and other help. <o:p></o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal">The pass initially attempts to inline calls which have loops in them. Following inlining, a separate scalar "main" pass tries to fuse loops. It is not necessary for loops to be adjacent or have the same control dependence. Specifically
a control dependence closure is computed and loops that share a control dep closure prefix are taken as candidates for fusion, in case there are no other loops in a certain path between them. A loop graph is built with edges between loops which share the
control dependence closure prefix. A recursive application of fusion follows on the graph from "bottom-up".<o:p></o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal">During fusion, program slices are taken based on the control flow closures of the two loops being fused. Example: Suppose two loops A and B are going to be fused. They share the same control dependence prefix, but their suffices vary.
The suffices are used to duplicate Control flow paths that pass through both loops. Specifically paths from the first block in the control-dep closure suffix of the first loop, through the first loop's exit and following into the suffix of the second loop
through the second loop's exit on to the common post-dominator of either loop's exit. The number of paths grows exponentially. At this time some heuristics are used when exploring for paths. Using profile based heuristics is a better idea.<o:p></o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal">Also function unswitching helps by cleaning up control flow.<o:p></o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal">Example:<o:p></o:p></p>
<p class="MsoNormal">if (a & b) {<o:p></o:p></p>
<p class="MsoNormal"> if (a & c) {<o:p></o:p></p>
<p class="MsoNormal"> if (t & 0xF) {<o:p></o:p></p>
<p class="MsoNormal"> for (i=0; i < size; i++) {<o:p></o:p></p>
<p class="MsoNormal"> if (A[i] & d) {<o:p></o:p></p>
<p class="MsoNormal"> A[i] |= (1 << t);<o:p></o:p></p>
<p class="MsoNormal"> }<o:p></o:p></p>
<p class="MsoNormal"> }<o:p></o:p></p>
<p class="MsoNormal"> }<o:p></o:p></p>
<p class="MsoNormal"> }<o:p></o:p></p>
<p class="MsoNormal">}<o:p></o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal">if (b & d) {<o:p></o:p></p>
<p class="MsoNormal"> if (a & d) {<o:p></o:p></p>
<p class="MsoNormal"> if (t & 0xA) {<o:p></o:p></p>
<p class="MsoNormal"> for (i=0; i < size; i++) {<o:p></o:p></p>
<p class="MsoNormal"> if (A[i] & d) {<o:p></o:p></p>
<p class="MsoNormal"> A[i] |= (1 << t);<o:p></o:p></p>
<p class="MsoNormal"> }<o:p></o:p></p>
<p class="MsoNormal"> }<o:p></o:p></p>
<p class="MsoNormal"> }<o:p></o:p></p>
<p class="MsoNormal"> }<o:p></o:p></p>
<p class="MsoNormal">}<o:p></o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal">After fusion we will have:<o:p></o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal">if (a&b && a&c && t&0xF && b&d && a&d t&0xA) {<o:p></o:p></p>
<p class="MsoNormal"> for (i=0; i < size; i++) {<o:p></o:p></p>
<p class="MsoNormal"> if (A[i] & d) {<o:p></o:p></p>
<p class="MsoNormal"> A[i] |= (1 << t);<o:p></o:p></p>
<p class="MsoNormal"> }<o:p></o:p></p>
<p class="MsoNormal"> if (A[i] & d) {<o:p></o:p></p>
<p class="MsoNormal"> A[i] |= (1 << t);<o:p></o:p></p>
<p class="MsoNormal"> }<o:p></o:p></p>
<p class="MsoNormal"> }<o:p></o:p></p>
<p class="MsoNormal">} else {<o:p></o:p></p>
<p class="MsoNormal">// original code<o:p></o:p></p>
<p class="MsoNormal"> if (a & b) {<o:p></o:p></p>
<p class="MsoNormal"> if (a & c) {<o:p></o:p></p>
<p class="MsoNormal"> if (t & 0xF) {<o:p></o:p></p>
<p class="MsoNormal"> for (i=0; i < size; i++) {<o:p></o:p></p>
<p class="MsoNormal"> if (A[i] & d) {<o:p></o:p></p>
<p class="MsoNormal"> A[i] |= (1 << t);<o:p></o:p></p>
<p class="MsoNormal"> }<o:p></o:p></p>
<p class="MsoNormal"> }<o:p></o:p></p>
<p class="MsoNormal"> }<o:p></o:p></p>
<p class="MsoNormal"> }<o:p></o:p></p>
<p class="MsoNormal"> }<o:p></o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal"> if (b & d) {<o:p></o:p></p>
<p class="MsoNormal"> if (a & d) {<o:p></o:p></p>
<p class="MsoNormal"> if (t & 0xA) {<o:p></o:p></p>
<p class="MsoNormal"> for (i=0; i < size; i++) {<o:p></o:p></p>
<p class="MsoNormal"> if (A[i] & d) {<o:p></o:p></p>
<p class="MsoNormal"> A[i] |= (1 << t);<o:p></o:p></p>
<p class="MsoNormal"> }<o:p></o:p></p>
<p class="MsoNormal"> }<o:p></o:p></p>
<p class="MsoNormal"> }<o:p></o:p></p>
<p class="MsoNormal"> }<o:p></o:p></p>
<p class="MsoNormal"> }<o:p></o:p></p>
<p class="MsoNormal">}<o:p></o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal">Thanks,<o:p></o:p></p>
<p class="MsoNormal">Ramshankar<o:p></o:p></p>
</div>
</body>
</html>