<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;}
@font-face
{font-family:Consolas;
panose-1:2 11 6 9 2 2 4 3 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";
color:black;}
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";
color:black;}
pre
{mso-style-priority:99;
mso-style-link:"HTML Preformatted Char";
margin:0cm;
margin-bottom:.0001pt;
font-size:10.0pt;
font-family:"Courier New";
color:black;}
span.HTMLPreformattedChar
{mso-style-name:"HTML Preformatted Char";
mso-style-priority:99;
mso-style-link:"HTML Preformatted";
font-family:Consolas;
color:black;}
span.EmailStyle20
{mso-style-type:personal-reply;
font-family:"Calibri","sans-serif";
color:#1F497D;}
.MsoChpDefault
{mso-style-type:export-only;
font-size:10.0pt;}
@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:939336499;
mso-list-template-ids:845055970;}
@list l0:level1
{mso-level-start-at:2;
mso-level-tab-stop:36.0pt;
mso-level-number-position:left;
text-indent:-18.0pt;}
@list l1
{mso-list-id:1636136645;
mso-list-template-ids:845055970;}
@list l1:level1
{mso-level-start-at:2;
mso-level-tab-stop:36.0pt;
mso-level-number-position:left;
text-indent:-18.0pt;}
@list l2
{mso-list-id:1949267029;
mso-list-template-ids:845055970;}
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 bgcolor="white" lang="EN-GB" link="blue" vlink="purple">
<div class="WordSection1">
<p class="MsoNormal">> Here's another motivating bug where early transform to select prevents optimization:<o:p></o:p></p>
<p class="MsoNormal">> <a href="https://bugs.llvm.org/show_bug.cgi?id=36760">https://bugs.llvm.org/show_bug.cgi?id=36760</a><o:p></o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal">> Is that case affected by this patch?<o:p></o:p></p>
<p class="MsoNormal" style="mso-margin-top-alt:auto;mso-margin-bottom-alt:auto">From a quick reading of that bug it looks like CVP needs to be enhanced to do a certain transformation, but the presence<br>
of a select means that even with that enhancement it wouldn’t do anything. With this patch we have a phi and so I think<br>
that if the CVP enhancement were done it should work.<o:p></o:p></p>
<p class="MsoNormal" style="mso-margin-top-alt:auto;mso-margin-bottom-alt:auto">>>> 1. LICM does not know how to hoist or sink regions. It does know how to hoist and sink selects.<br>
>>I think we should teach LICM to do this eventually.<br>
>Agreed. I just don't currently see patches to do so. Once LICM supports region hoisting, much of my concern disappears.<o:p></o:p></p>
<p class="MsoNormal" style="mso-margin-top-alt:auto;mso-margin-bottom-alt:auto">I’ve looked into what happens with LICM with this change, and it looks like everything except the phi gets hoisted, then<br>
later the phi is turned into a select, then a later LICM pass hoists the select. For the very simple tests I tried we end up with<br>
the same result in the end, but I wouldn’t be surprised if indeed LICM not doing this makes things worse in more complex<br>
cases.<o:p></o:p></p>
<p class="MsoNormal" style="mso-margin-top-alt:auto;mso-margin-bottom-alt:auto">I’ve hacked together a modification to LICM to make it able to hoist phis (by essentially hoisting entire blocks instead of<br>
moving everything into a single preheader block) and it seemed to work OK, so I’ll get to work on getting that working<br>
properly and committed before I touch phi-to-select transformation.<o:p></o:p></p>
<p class="MsoNormal" style="mso-margin-top-alt:auto;mso-margin-bottom-alt:auto">> InstCombine has limited support for triangles/diamonds, but fairly extensive support for selects.<o:p></o:p></p>
<p class="MsoNormal" style="mso-margin-top-alt:auto;mso-margin-bottom-alt:auto">Currently I have InstCombine running directly after phis have been turned into selects, so this should only be a problem if<br>
something before that relies on the select instcombine having already happened. I’ll look into that to see if it is indeed a<br>
problem.<o:p></o:p></p>
<p class="MsoNormal" style="mso-margin-top-alt:auto;mso-margin-bottom-alt:auto">>>>EarlyCSE and GVN do not know how to eliminate fully redundant triangles/diamonds. PRE *may* get some of these cases, but that is by no means guaranteed or likely to be robust.<br>
>>Agreed, we’ll need a plan to deal with these issues.<br>
>Again, I'd love to see these issues fixed. Once that's done, the convert about phi vs select as canonical form is much less relevant. <o:p></o:p></p>
<p class="MsoNormal" style="mso-margin-top-alt:auto;mso-margin-bottom-alt:auto">I’ll look into this.<o:p></o:p></p>
<p class="MsoNormal" style="mso-margin-top-alt:auto;mso-margin-bottom-alt:auto">> For the one of the specific cases mentioned, teaching GVN to do FRE and PRE for loads from selects of pointers just doesn't seem that painful. I would be really tempted to do
that<br>
> instead. Similarly, walking facts back from select uses in CVP seems generally useful since we have use dependent facts in a bunch of contexts, not just selects. (Call arguments for
<br>
> instance, non-null from unconditional deref, etc..)<o:p></o:p></p>
<p class="MsoNormal" style="mso-margin-top-alt:auto;mso-margin-bottom-alt:auto">Actually making GVN able to handle selects being painful is exactly the reason I’m proposing doing this. GVN has a hard<br>
constraint that a single block can only have a single available value, and attempting to change that (which you have to do<br>
when selecting between values defined in the same block) causes a ton of problems that have to be resolved.<o:p></o:p></p>
<p class="MsoNormal" style="mso-margin-top-alt:auto;mso-margin-bottom-alt:auto">John<o:p></o:p></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";color:windowtext">From:</span></b><span lang="EN-US" style="font-size:10.0pt;font-family:"Tahoma","sans-serif";color:windowtext"> Philip Reames [mailto:listmail@philipreames.com]
<br>
<b>Sent:</b> 17 August 2018 21:17<br>
<b>To:</b> Amara Emerson; Hal Finkel; John Brawn<br>
<b>Cc:</b> Sanjay Patel; llvm-dev@lists.llvm.org; nd<br>
<b>Subject:</b> Re: [llvm-dev] [RFC] Delaying phi-to-select transformation until later in the pass pipeline<o:p></o:p></span></p>
</div>
</div>
<p class="MsoNormal"><o:p> </o:p></p>
<p><o:p> </o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<div>
<p class="MsoNormal">On 08/17/2018 09:02 AM, Amara Emerson wrote:<o:p></o:p></p>
</div>
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<p class="MsoNormal"><o:p> </o:p></p>
<div>
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<div>
<p class="MsoNormal">On Aug 15, 2018, at 10:57 PM, Hal Finkel via llvm-dev <<a href="mailto:llvm-dev@lists.llvm.org">llvm-dev@lists.llvm.org</a>> wrote:<o:p></o:p></p>
</div>
<p class="MsoNormal"><o:p> </o:p></p>
<div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
<div>
<p class="MsoNormal">On 08/15/2018 02:38 PM, Philip Reames via llvm-dev wrote:<o:p></o:p></p>
</div>
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<p class="MsoNormal" style="mso-margin-top-alt:auto;mso-margin-bottom-alt:auto">I'm concerned that we're focusing on one side of this. Let me point out a few concerns w/changing the canonical form 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:l2 level1 lfo1">
LICM does not know how to hoist or sink regions. It does know how to hoist and sink selects.<o:p></o:p></li></ol>
</blockquote>
</div>
</div>
</blockquote>
<p class="MsoNormal">I think we should teach LICM to do this eventually.<o:p></o:p></p>
</div>
</blockquote>
<p class="MsoNormal">Agreed. I just don't currently see patches to do so. Once LICM supports region hoisting, much of my concern disappears.<br>
<br>
<o:p></o:p></p>
<div>
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<div>
<div>
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<ol start="2" type="1">
<li class="MsoNormal" style="mso-margin-top-alt:auto;mso-margin-bottom-alt:auto;mso-list:l0 level1 lfo2">
InstCombine has limited support for triangles/diamonds, but fairly extensive support for selects.<o:p></o:p></li><li class="MsoNormal" style="mso-margin-top-alt:auto;mso-margin-bottom-alt:auto;mso-list:l0 level1 lfo2">
EarlyCSE and GVN do not know how to eliminate fully redundant triangles/diamonds. PRE *may* get some of these cases, but that is by no means guaranteed or likely to be robust.<o:p></o:p></li></ol>
</blockquote>
</div>
</div>
</blockquote>
<p class="MsoNormal">Agreed, we’ll need a plan to deal with these issues.<o:p></o:p></p>
</div>
<p class="MsoNormal">Again, I'd love to see these issues fixed. Once that's done, the convert about phi vs select as canonical form is much less relevant.
<br>
<br>
<o:p></o:p></p>
<div>
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<div>
<div>
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<p class="MsoNormal" style="mso-margin-top-alt:auto;mso-margin-bottom-alt:auto">My personal opinion is that selects are the appropriate canonical form. <o:p></o:p></p>
<p class="MsoNormal" style="mso-margin-top-alt:auto;mso-margin-bottom-alt:auto">For the one of the specific cases mentioned, teaching GVN to do FRE and PRE for loads from selects of pointers just doesn't seem that painful. I would be really tempted to do that
instead. Similarly, walking facts back from select uses in CVP seems generally useful since we have use dependent facts in a bunch of contexts, not just selects. (Call arguments for instance, non-null from unconditional deref, etc..)<o:p></o:p></p>
<p class="MsoNormal" style="mso-margin-top-alt:auto;mso-margin-bottom-alt:auto">To be clear, I am raising concerns, not actively objecting to this. If you want to move forward and commit to work through all of the issues identified I wont actively stand in
the way.<o:p></o:p></p>
</blockquote>
<p class="MsoNormal"><br>
As I've expressed in the past, I think that not using select as part of our canonical form is potentially a superior design. However, it would be a major change. In addition to the issues that Philip mentions, there's also the fact that we'll just have more
basic blocks and that, in itself, could lead to an increase in compile time. However, working through these issues will likely leave us with a more-robust optimizer.<br>
<br>
See also: <a href="https://bugs.llvm.org/show_bug.cgi?id=34603#c19">https://bugs.llvm.org/show_bug.cgi?id=34603#c19</a><o:p></o:p></p>
</div>
</div>
</blockquote>
<p class="MsoNormal">Canonicalizing to phis is my preference too.<br>
<br>
<o:p></o:p></p>
<div>
<div>
<p class="MsoNormal"><br>
-Hal<br>
<br>
<br>
<o:p></o:p></p>
<p class="MsoNormal" style="mso-margin-top-alt:auto;mso-margin-bottom-alt:auto">Philip<o:p></o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<div>
<p class="MsoNormal">On 08/14/2018 12:39 PM, Sanjay Patel via llvm-dev wrote:<o:p></o:p></p>
</div>
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<div>
<div>
<p class="MsoNormal">I didn't look closely at the new patch, but this appears to be a small extension to:<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><a href="https://reviews.llvm.org/D38566">https://reviews.llvm.org/D38566</a><o:p></o:p></p>
</div>
<div>
<p class="MsoNormal">...and the GVN-based reasons for delaying transformation to 'select' are discussed in detail in the motivating bug for that patch:<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><a href="https://bugs.llvm.org/show_bug.cgi?id=34603">https://bugs.llvm.org/show_bug.cgi?id=34603</a><o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">So this sounds like the right direction to me. Note that there was objection to the implementation (a pile of pass options vs. uniquely named passes).<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">Here's another motivating bug where early transform to select prevents optimization:<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><a href="https://bugs.llvm.org/show_bug.cgi?id=36760">https://bugs.llvm.org/show_bug.cgi?id=36760</a><o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">Is that case affected by this patch?<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
<div>
<p class="MsoNormal">On Tue, Aug 14, 2018 at 11:17 AM, John Brawn via llvm-dev <<a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a>> wrote:<o:p></o:p></p>
<p class="MsoNormal">Summary<br>
=======<br>
<br>
I'm planning on adjusting SimplifyCFG so that it doesn't turn two-entry phi<br>
nodes into selects until later in the pass pipeline, to give passes which can<br>
understand phis but not selects more opportunity to optimize. The thing I'm<br>
trying to do which made me think of doing this is described below, but from the<br>
benchmarking I've done it looks like this is overall a good idea regardless of<br>
if I manage to get that done or not.<br>
<br>
Motivation<br>
==========<br>
<br>
My goal is to get clang to optimize some code containing a call to<br>
std::min_element which is dereferenced, so something like:<br>
<br>
float min_element_example(float *data, int size)<br>
{<br>
return *std::min_element(data, data+size);<br>
}<br>
<br>
which, after inlining a specialization, looks like:<br>
<br>
float min_element_example_inlined(float *first, float *last)<br>
{<br>
for (float *p = first; p != last; ++p)<br>
{<br>
if (*p < *first)<br>
first = p;<br>
}<br>
return *first;<br>
}<br>
<br>
There are two loads in the loop, *p and *first, but actually the load *p can be<br>
eliminated by using either the previous load *p or the previous *first,<br>
depending on if the if-condition was taken or not. However the<br>
"if (*p < *first) first = p" gets turned by simplifycfg into a select and this<br>
makes optimizing this a lot harder because you no longer have distinct paths<br>
through the CFG.<br>
<br>
I have some ideas on how to do the optimization (see my previous RFC "Making GVN<br>
able to visit the same block more than once" posted in April, though I've<br>
decided that the specific idea presented there isn't the right way to do it),<br>
but I think the first step is to make sure we don't have a select when we try<br>
to optimise this.<br>
<br>
Approach<br>
========<br>
<br>
I've posted a patch to <a href="https://reviews.llvm.org/D50723" target="_blank">
https://reviews.llvm.org/D50723</a> showing what I'm<br>
intending to do. An extra parameter is added to SimplifyCFG to control whether<br>
two-entry phi nodes are converted into select, and this is set to false in all<br>
instances before the end of module simplification. At the end of module<br>
simplification we do SimplifyCFG, then Instcombine to optimise the selects that<br>
are introduced, then EarlyCSE to eliminate common subexpressions introduced by<br>
instcombine.<o:p></o:p></p>
</div>
</div>
</blockquote>
</div>
</div>
<p class="MsoNormal">Is scheduling another simplifycfg, instcombine + earlycse pass really necessary? I’m concerned about the compile time impact.<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">Cheers,<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal">Amara<br>
<br>
<o:p></o:p></p>
<div>
<div>
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<div>
<div>
<blockquote style="border:none;border-left:solid #CCCCCC 1.0pt;padding:0cm 0cm 0cm 6.0pt;margin-left:4.8pt;margin-right:0cm">
<p class="MsoNormal"><br>
Benchmark Results<br>
=================<br>
<br>
These are performance differences reported by LNT when running llvm-test-suite,<br>
spec2000, and spec2006 at -O3 with and without the patch linked above (using<br>
trunk llvm from a week or so ago).<br>
<br>
AArch64 results on ARM Cortex-A72:<br>
<br>
Performance Regressions - execution_time Change<br>
SingleSource/Benchmarks/Shootout/Shootout-ary3 9.48%<br>
MultiSource/Benchmarks/TSVC/Packing-flt/Packing-flt 3.79%<br>
SingleSource/Benchmarks/CoyoteBench/huffbench 1.40%<br>
<br>
Performance Improvements - execution_time Change<br>
MultiSource/Benchmarks/TSVC/Searching-dbl/Searching-dbl -23.74%<br>
External/SPEC/CINT2000/256.bzip2/256.bzip2 -9.82%<br>
MultiSource/Benchmarks/TSVC/Searching-flt/Searching-flt -9.57%<br>
MultiSource/Benchmarks/TSVC/Equivalencing-flt/Equivalencing-flt -4.38%<br>
MultiSource/Benchmarks/TSVC/LinearDependence-flt/LinearDependence-flt -3.94%<br>
MultiSource/Benchmarks/TSVC/Packing-dbl/Packing-dbl -3.44%<br>
External/SPEC/CFP2006/453.povray/453.povray -2.50%<br>
SingleSource/Benchmarks/Adobe-C++/stepanov_vector -1.49%<br>
<br>
X86_64 results on Intel Xeon E5-2690:<br>
<br>
Performance Regressions - execution_time Change<br>
MultiSource/Benchmarks/Ptrdist/yacr2/yacr2 5.62%<br>
<br>
Performance Improvements - execution_time Change<br>
SingleSource/Benchmarks/Misc-C++/Large/sphereflake -4.43%<br>
External/SPEC/CINT2006/456.hmmer/456.hmmer -2.50%<br>
External/SPEC/CINT2006/464.h264ref/464.h264ref -1.60%<br>
MultiSource/Benchmarks/nbench/nbench -1.19%<br>
SingleSource/Benchmarks/Adobe-C++/functionobjects -1.07%<br>
<br>
I had a brief look at the regressions and they all look to be caused by<br>
getting bad luck with branch mispredictions: I looked into the Shootout-ary3 and<br>
yacr2 cases and in both the hot code path was the same with and without the<br>
patch, but with more mispredictions probably caused by changes elsewhere.<br>
<br>
John<br>
<br>
_______________________________________________<br>
LLVM Developers mailing list<br>
<a href="mailto:llvm-dev@lists.llvm.org">llvm-dev@lists.llvm.org</a><br>
<a href="http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev" target="_blank">http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev</a><o:p></o:p></p>
</blockquote>
</div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<p class="MsoNormal"><br>
<br>
<br>
<o:p></o:p></p>
<pre>_______________________________________________<o:p></o:p></pre>
<pre>LLVM Developers mailing list<o:p></o:p></pre>
<pre><a href="mailto:llvm-dev@lists.llvm.org">llvm-dev@lists.llvm.org</a><o:p></o:p></pre>
<pre><a href="http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev">http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev</a><o:p></o:p></pre>
</blockquote>
<p class="MsoNormal"><br>
<br>
<br>
<br>
<o:p></o:p></p>
<pre>_______________________________________________<o:p></o:p></pre>
<pre>LLVM Developers mailing list<o:p></o:p></pre>
<pre><a href="mailto:llvm-dev@lists.llvm.org">llvm-dev@lists.llvm.org</a><o:p></o:p></pre>
<pre><a href="http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev">http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev</a><o:p></o:p></pre>
</blockquote>
<p class="MsoNormal"><br>
<br>
<o:p></o:p></p>
<pre>-- <o:p></o:p></pre>
<pre>Hal Finkel<o:p></o:p></pre>
<pre>Lead, Compiler Technology and Programming Languages<o:p></o:p></pre>
<pre>Leadership Computing Facility<o:p></o:p></pre>
<pre>Argonne National Laboratory<o:p></o:p></pre>
</div>
<p class="MsoNormal">_______________________________________________<br>
LLVM Developers mailing list<br>
<a href="mailto:llvm-dev@lists.llvm.org">llvm-dev@lists.llvm.org</a><br>
<a href="http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev">http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev</a><o:p></o:p></p>
</div>
</div>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
</div>
</body>
</html>