<html>
<head>
<meta http-equiv="Content-Type" content="text/html; charset=utf-8">
</head>
<body text="#000000" bgcolor="#FFFFFF">
Adrerrazek,<br>
<br>
Thank for you for the fairly lengthy follow up. <br>
<br>
I've replied to few specific points inline, but the macro concern I
have is the same one I expressed in my previous email. *Why* is
shifting the extend to the load operand profitable? Why do we
believe this is a generally profitable transformation? Your answer
to date has been specific to one particular example. That does not
address the core concern which is about generality. <br>
<br>
Fair warning: I will revert this patch in a day or so unless we've
come to a satisfactory conclusion. That doesn't prevent us from
continuing the conversation and then reintroducing it later. It's
just to avoid having trunk potentially (performance wise) broken in
the mean time. <br>
<br>
Philip<br>
<br>
<div class="moz-cite-prefix">On 09/12/2018 03:56 PM, Abderrazek
Zaafrani wrote:<br>
</div>
<blockquote type="cite"
cite="mid:4066194787384d109b34900d805b651b@samsung.com">
<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:0in;
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:0in;
mso-margin-bottom-alt:auto;
margin-left:0in;
font-size:12.0pt;
font-family:"Times New Roman","serif";
color:black;}
pre
{mso-style-priority:99;
mso-style-link:"HTML Preformatted Char";
margin:0in;
margin-bottom:.0001pt;
font-size:10.0pt;
font-family:"Courier New";
color:black;}
span.phui-handle
{mso-style-name:phui-handle;}
span.HTMLPreformattedChar
{mso-style-name:"HTML Preformatted Char";
mso-style-priority:99;
mso-style-link:"HTML Preformatted";
font-family:Consolas;
color:black;}
span.EmailStyle21
{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: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]-->
<div class="WordSection1">
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D">Philip,<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">This
is a continuation of my previous reply and the main intent
is to show some generated code.<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">The
context of this work is to try to improve llvm to be on par
with gcc. A well-known benchmark used for phones show a gap
between gcc and llvm. I cannot share the proprietary code
but I tried to reconstruct my own test (see test.cpp
attached). The other two files attached are llvm IR dump
without the patch and with the patch. I am only including
the dump before and after Induction Variable Simplification
pass as it is the only pass I am modifying. </span></p>
</div>
</blockquote>
Am I correct in interpreting dump.txt to be the before IR and
dump,.txt.patch to the after IR? I'd probably have named these
before-indvars.ll and after-indvars.ll for clarity. <br>
<blockquote type="cite"
cite="mid:4066194787384d109b34900d805b651b@samsung.com">
<div class="WordSection1">
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D">If
you compare these two files, you can see that there are less
truncate instructions with my patch. Current Induction
simplification pass , among other things, tries to avoid
creating truncate instructions while widening the induction
variables. My patch does not bring any new idea or
optimization. It just tries harder to avoid creating
truncate instructions.</span></p>
</div>
</blockquote>
I disagree here. You may be extending an existing approach, but the
specifics of how to "avoid a truncate" in this case is new and needs
to be justified. Why should we believe that the widened-bin-op +
*ext form is profitable over the trunc form *in general*? (Not just
this one example.)<br>
<blockquote type="cite"
cite="mid:4066194787384d109b34900d805b651b@samsung.com">
<div class="WordSection1">
<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"><o:p> </o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D">Let’s
examine the effect of having clean code without unnecessary
truncate instructions on the final generated code.
<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">Here
is the ARM assembly for test.cpp without the patch:</span></p>
</div>
</blockquote>
Unfortunately, I'm not fluent in ARM assembly, so this doesn't help
me much. Comments on key pieces and changes would have gone a long
way. Alternatively, I am much more familiar with X86 if you'd
prefer.<br>
<blockquote type="cite"
cite="mid:4066194787384d109b34900d805b651b@samsung.com">
<div class="WordSection1">
<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"><o:p> </o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D">
.cfi_startproc<o:p></o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D">//
%bb.0: // %entry<o:p></o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D">
ldr w9, [x0]<o:p></o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D"> add
x11, x1, w3, sxtw #2<o:p></o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D"> mov
w8, wzr<o:p></o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D"> neg
w10, w9<o:p></o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D"> lsl
w12, w9, #1<o:p></o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D"> orr
w13, wzr, #0x4<o:p></o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D">.LBB0_1:
// %for.body<o:p></o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D"> //
=>This Inner Loop Header: Depth=1<o:p></o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D"> add
w14, w12, w8<o:p></o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D"> ldrb
w14, [x2, w14, sxtw]<o:p></o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D"> ldrb
w15, [x2, w8, sxtw]<o:p></o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D"> add
w16, w10, w8<o:p></o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D"> ldrb
w16, [x2, w16, sxtw]<o:p></o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D"> add
w8, w8, w9<o:p></o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D"> sub
w14, w14, w15<o:p></o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D"> add
w14, w14, w16<o:p></o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D"> str
w14, [x11, x13]<o:p></o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D"> add
x13, x13, #4 // =4<o:p></o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D"> cmp
x13, #400 // =400<o:p></o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D"> b.ne
.LBB0_1<o:p></o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D">//
%bb.2:
<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">And
with the patch:<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">
.cfi_startproc<o:p></o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D">//
%bb.0: // %entry<o:p></o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D"> ldrsw
x8, [x0]<o:p></o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D"> add
x10, x1, w3, sxtw #2<o:p></o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D"> neg
x9, x8<o:p></o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D"> lsl
x11, x8, #1<o:p></o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D"> orr
w12, wzr, #0x4<o:p></o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D">.LBB0_1:
// %for.body<o:p></o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D"> //
=>This Inner Loop Header: Depth=1<o:p></o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D"> ldrb
w13, [x2, x11]<o:p></o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D"> ldrb
w14, [x2]<o:p></o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D"> ldrb
w15, [x2, x9]<o:p></o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D">
add x2, x2, x8<o:p></o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D"> sub
w13, w13, w14<o:p></o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D"> add
w13, w13, w15<o:p></o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D"> str
w13, [x10, x12]<o:p></o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D"> add
x12, x12, #4 // =4<o:p></o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D"> cmp
x12, #400 // =400<o:p></o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D">b.ne
.LBB0_1<o:p></o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D">//
%bb.2: // %for.cond.cleanup<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">If
you look at the load instructions ldrb inside the loop for
both versions, you can see that the offsets are computed
more efficiently with the patched version than non-patched
version (offset is computed inside the loop for non-patch
version and outside the loop for patched version). </span></p>
</div>
</blockquote>
Your framing here really confuses me. What does extending the load
have to do with the addressing form used? *Why* is there a
connection here? <br>
<blockquote type="cite"
cite="mid:4066194787384d109b34900d805b651b@samsung.com">
<div class="WordSection1">
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D">Experiment
on ARM A72 shows an improvement of close to 10% for one
application in the benchmark.
</span></p>
</div>
</blockquote>
This is a single benchmark. I'm asking about the set of all
possible benchmarks. <br>
<blockquote type="cite"
cite="mid:4066194787384d109b34900d805b651b@samsung.com">
<div class="WordSection1">
<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"><o:p> </o:p></span></p>
<p class="MsoNormal"><span
style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1F497D">Abderrazek<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>
<div style="border:none;border-top:solid #B5C4DF
1.0pt;padding:3.0pt 0in 0in 0in">
<p class="MsoNormal"><b><span
style="font-size:10.0pt;font-family:"Tahoma","sans-serif";color:windowtext">From:</span></b><span
style="font-size:10.0pt;font-family:"Tahoma","sans-serif";color:windowtext">
Philip Reames [<a class="moz-txt-link-freetext" href="mailto:listmail@philipreames.com">mailto:listmail@philipreames.com</a>]
<br>
<b>Sent:</b> Monday, September 10, 2018 4:41 PM<br>
<b>To:</b> Abderrazek Zaafrani;
<a class="moz-txt-link-abbreviated" href="mailto:llvm-commits@lists.llvm.org">llvm-commits@lists.llvm.org</a>; Sebastian Pop<br>
<b>Subject:</b> Re: [llvm] r341726 - [SimplifyIndVar]
Avoid generating truncate instructions with non-hoisted
Laod operand.<o:p></o:p></span></p>
</div>
</div>
<p class="MsoNormal"><o:p> </o:p></p>
<p>I have serious concerns with this patch. I don't believe
this patch meets our usual standards. I believe this patch
warrants being reverted, and then re-reviewed from scratch.
<o:p></o:p></p>
<p>@<span class="phui-handle">sebpop</span>, I don't think it
was appropriate to approve this patch given the public
history. I wouldn't normally call that out, but this seems
like a really clear cut case.<o:p></o:p></p>
<p>@Abderrazek, please don't let me scare you off here. While I
think this patch does need reverted, I think this can be
adjusted into something clearly worthwhile, and I encourage
you to do so. I'm happy to help as well. (Fair warning, my
time is very limited, so my help will only go so far.)<o:p></o:p></p>
<p>Philip<o:p></o:p></p>
<div>
<p class="MsoNormal">On 09/07/2018 03:41 PM, Abderrazek
Zaafrani via llvm-commits wrote:<o:p></o:p></p>
</div>
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<pre>Author: az<o:p></o:p></pre>
<pre>Date: Fri Sep 7 15:41:57 2018<o:p></o:p></pre>
<pre>New Revision: 341726<o:p></o:p></pre>
<pre><o:p> </o:p></pre>
<pre>URL: <a href="http://llvm.org/viewvc/llvm-project?rev=341726&view=rev" moz-do-not-send="true">http://llvm.org/viewvc/llvm-project?rev=341726&view=rev</a><o:p></o:p></pre>
<pre>Log:<o:p></o:p></pre>
<pre>[SimplifyIndVar] Avoid generating truncate instructions with non-hoisted Laod operand.<o:p></o:p></pre>
</blockquote>
<p class="MsoNormal">(For the future, you had a nice comment in
the review about the reasoning. The submit comment should
include that too.)<br>
<br>
Your reasoning here is explained purely in terms of what we do
for loop-invariant loads. You don't state what this is, or
provide a clear description of what the code generation should
be. I've read over the review and the code, and can reverse
engineer what you're going for, but it is your responsibility
to make this clear from the submitted code. This patch does
not achieve that objective.<br>
<br>
Reverse engineering the code, here's what I *think* you're
trying to accomplish:<br>
If we find a binary operator where we need to extend one side,
check to see if we can cheaply extend the other operand and
widen the load instead. Many architectures can efficiently
generate an extending load, so we consider loads to be one
such cheaply extensible case.<br>
<br>
(I can't connect this understanding of the code to your
comments in the review though, so maybe I'm missing something
still?)<br>
<br>
In terms of the review discussion, I am concerned that you
appear to be mixing TBAA and AliasSet terminology. You appear
to have rejected an approach based on a possibly flawed
understanding. Your reviewers should have asked for a
reference w.r.t. the rejected approach (e.g. email discussion,
review, or rejected patch). <br>
<br>
<br>
<br>
<o:p></o:p></p>
<pre><o:p> </o:p></pre>
<pre><o:p> </o:p></pre>
<pre>Differential Revision: <a href="https://reviews.llvm.org/D49151" moz-do-not-send="true">https://reviews.llvm.org/D49151</a><o:p></o:p></pre>
<pre><o:p> </o:p></pre>
<pre>Modified:<o:p></o:p></pre>
<pre> llvm/trunk/lib/Transforms/Scalar/IndVarSimplify.cpp<o:p></o:p></pre>
<pre> llvm/trunk/test/Transforms/IndVarSimplify/iv-widen-elim-ext.ll<o:p></o:p></pre>
<pre><o:p> </o:p></pre>
<pre>Modified: llvm/trunk/lib/Transforms/Scalar/IndVarSimplify.cpp<o:p></o:p></pre>
<pre>URL: <a href="http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/Scalar/IndVarSimplify.cpp?rev=341726&r1=341725&r2=341726&view=diff" moz-do-not-send="true">http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/Scalar/IndVarSimplify.cpp?rev=341726&r1=341725&r2=341726&view=diff</a><o:p></o:p></pre>
<pre>==============================================================================<o:p></o:p></pre>
<pre>--- llvm/trunk/lib/Transforms/Scalar/IndVarSimplify.cpp (original)<o:p></o:p></pre>
<pre>+++ llvm/trunk/lib/Transforms/Scalar/IndVarSimplify.cpp Fri Sep 7 15:41:57 2018<o:p></o:p></pre>
<pre>@@ -1017,6 +1017,8 @@ protected:<o:p></o:p></pre>
<pre> Instruction *widenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter);<o:p></o:p></pre>
<pre> <o:p></o:p></pre>
<pre> bool widenLoopCompare(NarrowIVDefUse DU);<o:p></o:p></pre>
<pre>+ bool widenWithVariantLoadUse(NarrowIVDefUse DU);<o:p></o:p></pre>
<pre>+ void widenWithVariantLoadUseCodegen(NarrowIVDefUse DU);<o:p></o:p></pre>
<pre> <o:p></o:p></pre>
<pre> void pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef);<o:p></o:p></pre>
<pre> };<o:p></o:p></pre>
<pre>@@ -1361,6 +1363,146 @@ bool WidenIV::widenLoopCompare(NarrowIVD<o:p></o:p></pre>
<pre> return true;<o:p></o:p></pre>
<pre> }<o:p></o:p></pre>
<pre> <o:p></o:p></pre>
<pre>+/// If the narrow use is an instruction whose two operands are the defining<o:p></o:p></pre>
<pre>+/// instruction of DU and a load instruction, then we have the following:<o:p></o:p></pre>
<pre>+/// if the load is hoisted outside the loop, then we do not reach this function<o:p></o:p></pre>
<pre>+/// as scalar evolution analysis works fine in widenIVUse with variables<o:p></o:p></pre>
<pre>+/// hoisted outside the loop and efficient code is subsequently generated by<o:p></o:p></pre>
<pre>+/// not emitting truncate instructions. But when the load is not hoisted<o:p></o:p></pre>
<pre>+/// (whether due to limitation in alias analysis or due to a true legality),<o:p></o:p></pre>
<pre>+/// then scalar evolution can not proceed with loop variant values and<o:p></o:p></pre>
<pre>+/// inefficient code is generated. This function handles the non-hoisted load<o:p></o:p></pre>
<pre>+/// special case by making the optimization generate the same type of code for<o:p></o:p></pre>
<pre>+/// hoisted and non-hoisted load (widen use and eliminate sign extend<o:p></o:p></pre>
<pre>+/// instruction). This special case is important especially when the induction<o:p></o:p></pre>
<pre>+/// variables are affecting addressing mode in code generation.<o:p></o:p></pre>
<p class="MsoNormal">This comment really tells me nothing about
what this function does, or what it's goal is.
<br>
<br>
<o:p></o:p></p>
<pre><o:p> </o:p></pre>
<pre>+bool WidenIV::widenWithVariantLoadUse(NarrowIVDefUse DU) {<o:p></o:p></pre>
<pre>+ Instruction *NarrowUse = DU.NarrowUse;<o:p></o:p></pre>
<pre>+ Instruction *NarrowDef = DU.NarrowDef;<o:p></o:p></pre>
<pre>+ Instruction *WideDef = DU.WideDef;<o:p></o:p></pre>
<pre>+<o:p></o:p></pre>
<pre>+ // Handle the common case of add<nsw/nuw><o:p></o:p></pre>
<pre>+ const unsigned OpCode = NarrowUse->getOpcode();<o:p></o:p></pre>
<pre>+ // Only Add/Sub/Mul instructions are supported.<o:p></o:p></pre>
<p class="MsoNormal">Why?<br>
<br>
<o:p></o:p></p>
<pre><o:p> </o:p></pre>
<pre>+ if (OpCode != Instruction::Add && OpCode != Instruction::Sub &&<o:p></o:p></pre>
<pre>+ OpCode != Instruction::Mul)<o:p></o:p></pre>
<pre>+ return false;<o:p></o:p></pre>
<pre>+<o:p></o:p></pre>
<pre>+ // The operand that is not defined by NarrowDef of DU. Let's call it the<o:p></o:p></pre>
<pre>+ // other operand.<o:p></o:p></pre>
<pre>+ unsigned ExtendOperIdx = DU.NarrowUse->getOperand(0) == NarrowDef ? 1 : 0;<o:p></o:p></pre>
<pre>+ assert(DU.NarrowUse->getOperand(1 - ExtendOperIdx) == DU.NarrowDef &&<o:p></o:p></pre>
<pre>+ "bad DU");<o:p></o:p></pre>
<pre>+<o:p></o:p></pre>
<pre>+ const SCEV *ExtendOperExpr = nullptr;<o:p></o:p></pre>
<pre>+ const OverflowingBinaryOperator *OBO =<o:p></o:p></pre>
<pre>+ cast<OverflowingBinaryOperator>(NarrowUse);<o:p></o:p></pre>
<pre>+ ExtendKind ExtKind = getExtendKind(NarrowDef);<o:p></o:p></pre>
<pre>+ if (ExtKind == SignExtended && OBO->hasNoSignedWrap())<o:p></o:p></pre>
<pre>+ ExtendOperExpr = SE->getSignExtendExpr(<o:p></o:p></pre>
<pre>+ SE->getSCEV(NarrowUse->getOperand(ExtendOperIdx)), WideType);<o:p></o:p></pre>
<pre>+ else if (ExtKind == ZeroExtended && OBO->hasNoUnsignedWrap())<o:p></o:p></pre>
<pre>+ ExtendOperExpr = SE->getZeroExtendExpr(<o:p></o:p></pre>
<pre>+ SE->getSCEV(NarrowUse->getOperand(ExtendOperIdx)), WideType);<o:p></o:p></pre>
<pre>+ else<o:p></o:p></pre>
<pre>+ return false;<o:p></o:p></pre>
<pre>+<o:p></o:p></pre>
<pre>+ // We are interested in the other operand being a load instruction.<o:p></o:p></pre>
<pre>+ // But, we should look into relaxing this restriction later on.<o:p></o:p></pre>
<pre>+ auto *I = dyn_cast<Instruction>(NarrowUse->getOperand(ExtendOperIdx));<o:p></o:p></pre>
<pre>+ if (I && I->getOpcode() != Instruction::Load)<o:p></o:p></pre>
<pre>+ return false;<o:p></o:p></pre>
<pre>+<o:p></o:p></pre>
<pre>+ // Verifying that Defining operand is an AddRec<o:p></o:p></pre>
<pre>+ const SCEV *Op1 = SE->getSCEV(WideDef);<o:p></o:p></pre>
<pre>+ const SCEVAddRecExpr *AddRecOp1 = dyn_cast<SCEVAddRecExpr>(Op1);<o:p></o:p></pre>
<pre>+ if (!AddRecOp1 || AddRecOp1->getLoop() != L)<o:p></o:p></pre>
<pre>+ return false;<o:p></o:p></pre>
<pre>+ // Verifying that other operand is an Extend.<o:p></o:p></pre>
<pre>+ if (ExtKind == SignExtended) {<o:p></o:p></pre>
<pre>+ if (!isa<SCEVSignExtendExpr>(ExtendOperExpr))<o:p></o:p></pre>
<pre>+ return false;<o:p></o:p></pre>
<pre>+ } else {<o:p></o:p></pre>
<pre>+ if (!isa<SCEVZeroExtendExpr>(ExtendOperExpr))<o:p></o:p></pre>
<pre>+ return false;<o:p></o:p></pre>
<pre>+ }<o:p></o:p></pre>
<pre>+<o:p></o:p></pre>
<pre>+ if (ExtKind == SignExtended) {<o:p></o:p></pre>
<pre>+ for (Use &U : NarrowUse->uses()) {<o:p></o:p></pre>
<pre>+ SExtInst *User = dyn_cast<SExtInst>(U.getUser());<o:p></o:p></pre>
<pre>+ if (!User || User->getType() != WideType)<o:p></o:p></pre>
<pre>+ return false;<o:p></o:p></pre>
<pre>+ }<o:p></o:p></pre>
<pre>+ } else { // ExtKind == ZeroExtended<o:p></o:p></pre>
<pre>+ for (Use &U : NarrowUse->uses()) {<o:p></o:p></pre>
<pre>+ ZExtInst *User = dyn_cast<ZExtInst>(U.getUser());<o:p></o:p></pre>
<pre>+ if (!User || User->getType() != WideType)<o:p></o:p></pre>
<pre>+ return false;<o:p></o:p></pre>
<pre>+ }<o:p></o:p></pre>
<pre>+ }<o:p></o:p></pre>
<p class="MsoNormal">This and the previous block of restrictions
seem awfully narrow with no clear reasoning of why they need
to be.<br>
<br>
<o:p></o:p></p>
<pre><o:p> </o:p></pre>
<pre>+<o:p></o:p></pre>
<pre>+ return true;<o:p></o:p></pre>
<pre>+}<o:p></o:p></pre>
<pre>+<o:p></o:p></pre>
<pre>+/// Special Case for widening with variant Loads (see<o:p></o:p></pre>
<pre>+/// WidenIV::widenWithVariantLoadUse). This is the code generation part.<o:p></o:p></pre>
<pre>+void WidenIV::widenWithVariantLoadUseCodegen(NarrowIVDefUse DU) {<o:p></o:p></pre>
<pre>+ Instruction *NarrowUse = DU.NarrowUse;<o:p></o:p></pre>
<pre>+ Instruction *NarrowDef = DU.NarrowDef;<o:p></o:p></pre>
<pre>+ Instruction *WideDef = DU.WideDef;<o:p></o:p></pre>
<pre>+<o:p></o:p></pre>
<pre>+ ExtendKind ExtKind = getExtendKind(NarrowDef);<o:p></o:p></pre>
<pre>+<o:p></o:p></pre>
<pre>+ LLVM_DEBUG(dbgs() << "Cloning arithmetic IVUser: " << *NarrowUse << "\n");<o:p></o:p></pre>
<pre>+<o:p></o:p></pre>
<pre>+ // Generating a widening use instruction.<o:p></o:p></pre>
<pre>+ Value *LHS = (NarrowUse->getOperand(0) == NarrowDef)<o:p></o:p></pre>
<pre>+ ? WideDef<o:p></o:p></pre>
<pre>+ : createExtendInst(NarrowUse->getOperand(0), WideType,<o:p></o:p></pre>
<pre>+ ExtKind, NarrowUse);<o:p></o:p></pre>
<pre>+ Value *RHS = (NarrowUse->getOperand(1) == NarrowDef)<o:p></o:p></pre>
<pre>+ ? WideDef<o:p></o:p></pre>
<pre>+ : createExtendInst(NarrowUse->getOperand(1), WideType,<o:p></o:p></pre>
<pre>+ ExtKind, NarrowUse);<o:p></o:p></pre>
<pre>+<o:p></o:p></pre>
<pre>+ auto *NarrowBO = cast<BinaryOperator>(NarrowUse);<o:p></o:p></pre>
<pre>+ auto *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(), LHS, RHS,<o:p></o:p></pre>
<pre>+ NarrowBO->getName());<o:p></o:p></pre>
<pre>+ IRBuilder<> Builder(NarrowUse);<o:p></o:p></pre>
<pre>+ Builder.Insert(WideBO);<o:p></o:p></pre>
<pre>+ WideBO->copyIRFlags(NarrowBO);<o:p></o:p></pre>
<pre>+<o:p></o:p></pre>
<pre>+ if (ExtKind == SignExtended)<o:p></o:p></pre>
<pre>+ ExtendKindMap[NarrowUse] = SignExtended;<o:p></o:p></pre>
<pre>+ else<o:p></o:p></pre>
<pre>+ ExtendKindMap[NarrowUse] = ZeroExtended;<o:p></o:p></pre>
<pre>+<o:p></o:p></pre>
<pre>+ // Update the Use.<o:p></o:p></pre>
<pre>+ if (ExtKind == SignExtended) {<o:p></o:p></pre>
<pre>+ for (Use &U : NarrowUse->uses()) {<o:p></o:p></pre>
<pre>+ SExtInst *User = dyn_cast<SExtInst>(U.getUser());<o:p></o:p></pre>
<pre>+ if (User && User->getType() == WideType) {<o:p></o:p></pre>
<pre>+ LLVM_DEBUG(dbgs() << "INDVARS: eliminating " << *User << " replaced by "<o:p></o:p></pre>
<pre>+ << *WideBO << "\n");<o:p></o:p></pre>
<pre>+ ++NumElimExt;<o:p></o:p></pre>
<pre>+ User->replaceAllUsesWith(WideBO);<o:p></o:p></pre>
<pre>+ DeadInsts.emplace_back(User);<o:p></o:p></pre>
<pre>+ }<o:p></o:p></pre>
<pre>+ }<o:p></o:p></pre>
<pre>+ } else { // ExtKind == ZeroExtended<o:p></o:p></pre>
<pre>+ for (Use &U : NarrowUse->uses()) {<o:p></o:p></pre>
<pre>+ ZExtInst *User = dyn_cast<ZExtInst>(U.getUser());<o:p></o:p></pre>
<pre>+ if (User && User->getType() == WideType) {<o:p></o:p></pre>
<pre>+ LLVM_DEBUG(dbgs() << "INDVARS: eliminating " << *User << " replaced by "<o:p></o:p></pre>
<pre>+ << *WideBO << "\n");<o:p></o:p></pre>
<pre>+ ++NumElimExt;<o:p></o:p></pre>
<pre>+ User->replaceAllUsesWith(WideBO);<o:p></o:p></pre>
<pre>+ DeadInsts.emplace_back(User);<o:p></o:p></pre>
<pre>+ }<o:p></o:p></pre>
<pre>+ }<o:p></o:p></pre>
<pre>+ }<o:p></o:p></pre>
<pre>+}<o:p></o:p></pre>
<pre>+<o:p></o:p></pre>
<pre> /// Determine whether an individual user of the narrow IV can be widened. If so,<o:p></o:p></pre>
<pre> /// return the wide clone of the user.<o:p></o:p></pre>
<pre> Instruction *WidenIV::widenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter) {<o:p></o:p></pre>
<pre>@@ -1458,6 +1600,16 @@ Instruction *WidenIV::widenIVUse(NarrowI<o:p></o:p></pre>
<pre> if (widenLoopCompare(DU))<o:p></o:p></pre>
<pre> return nullptr;<o:p></o:p></pre>
<pre> <o:p></o:p></pre>
<pre>+ // We are here about to generate a truncate instruction that may hurt<o:p></o:p></pre>
<pre>+ // performance because the scalar evolution expression computed earlier<o:p></o:p></pre>
<pre>+ // in WideAddRec.first does not indicate a polynomial induction expression.<o:p></o:p></pre>
<pre>+ // In that case, look at the operands of the use instruction to determine<o:p></o:p></pre>
<pre>+ // if we can still widen the use instead of truncating its operand.<o:p></o:p></pre>
<pre>+ if (widenWithVariantLoadUse(DU)) {<o:p></o:p></pre>
<pre>+ widenWithVariantLoadUseCodegen(DU);<o:p></o:p></pre>
<pre>+ return nullptr;<o:p></o:p></pre>
<pre>+ }<o:p></o:p></pre>
<pre>+<o:p></o:p></pre>
<pre> // This user does not evaluate to a recurrence after widening, so don't<o:p></o:p></pre>
<pre> // follow it. Instead insert a Trunc to kill off the original use,<o:p></o:p></pre>
<pre> // eventually isolating the original narrow IV so it can be removed.<o:p></o:p></pre>
<pre><o:p> </o:p></pre>
<pre>Modified: llvm/trunk/test/Transforms/IndVarSimplify/iv-widen-elim-ext.ll<o:p></o:p></pre>
<pre>URL: <a href="http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/IndVarSimplify/iv-widen-elim-ext.ll?rev=341726&r1=341725&r2=341726&view=diff" moz-do-not-send="true">http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/IndVarSimplify/iv-widen-elim-ext.ll?rev=341726&r1=341725&r2=341726&view=diff</a><o:p></o:p></pre>
<pre>==============================================================================<o:p></o:p></pre>
<pre>--- llvm/trunk/test/Transforms/IndVarSimplify/iv-widen-elim-ext.ll (original)<o:p></o:p></pre>
<pre>+++ llvm/trunk/test/Transforms/IndVarSimplify/iv-widen-elim-ext.ll Fri Sep 7 15:41:57 2018<o:p></o:p></pre>
<pre>@@ -273,3 +273,87 @@ for.end:<o:p></o:p></pre>
<pre> %call = call i32 @dummy(i32* getelementptr inbounds ([100 x i32], [100 x i32]* @a, i32 0, i32 0), i32* getelementptr inbounds ([100 x i32], [100 x i32]* @b, i32 0, i32 0))<o:p></o:p></pre>
<pre> ret i32 0<o:p></o:p></pre>
<pre> }<o:p></o:p></pre>
<p class="MsoNormal">The testing included here is insufficient.
Specifically:<br>
1) The tests do not appear to be maximally reduced.<br>
2) The tests do not clearly show what the expected output is.
(i.e. positive check)<br>
3) It includes only minimal negative tests (i.e. exercising
cases which shouldn't trigger). Clearly missing cases
include: non-load RHS, other binary op types, missing nsw/nuw,
etc...<br>
<br>
<br>
<o:p></o:p></p>
<pre><o:p> </o:p></pre>
<pre>+<o:p></o:p></pre>
<pre>+%struct.image = type {i32, i32}<o:p></o:p></pre>
<pre>+define i32 @foo4(%struct.image* %input, i32 %length, i32* %in) {<o:p></o:p></pre>
<pre>+entry:<o:p></o:p></pre>
<pre>+ %stride = getelementptr inbounds %struct.image, %struct.image* %input, i64 0, i32 1<o:p></o:p></pre>
<pre>+ %0 = load i32, i32* %stride, align 4<o:p></o:p></pre>
<pre>+ %cmp17 = icmp sgt i32 %length, 1<o:p></o:p></pre>
<pre>+ br i1 %cmp17, label %for.body.lr.ph, label %for.cond.cleanup<o:p></o:p></pre>
<pre>+<o:p></o:p></pre>
<pre>+for.body.lr.ph: ; preds = %entry<o:p></o:p></pre>
<pre>+ %channel = getelementptr inbounds %struct.image, %struct.image* %input, i64 0, i32 0<o:p></o:p></pre>
<pre>+ br label %for.body<o:p></o:p></pre>
<pre>+<o:p></o:p></pre>
<pre>+for.cond.cleanup.loopexit: ; preds = %for.body<o:p></o:p></pre>
<pre>+ %1 = phi i32 [ %6, %for.body ]<o:p></o:p></pre>
<pre>+ br label %for.cond.cleanup<o:p></o:p></pre>
<pre>+<o:p></o:p></pre>
<pre>+for.cond.cleanup: ; preds = %for.cond.cleanup.loopexit, %entry<o:p></o:p></pre>
<pre>+ %2 = phi i32 [ 0, %entry ], [ %1, %for.cond.cleanup.loopexit ]<o:p></o:p></pre>
<pre>+ ret i32 %2<o:p></o:p></pre>
<pre>+<o:p></o:p></pre>
<pre>+; mul instruction below is widened instead of generating a truncate instruction for it<o:p></o:p></pre>
<pre>+; regardless if Load operand of mul is inside or outside the loop (we have both cases).<o:p></o:p></pre>
<pre>+; CHECK: for.body:<o:p></o:p></pre>
<pre>+; CHECK-NOT: trunc<o:p></o:p></pre>
<pre>+for.body: ; preds = %for.body.lr.ph, %for.body<o:p></o:p></pre>
<pre>+ %x.018 = phi i32 [ 1, %for.body.lr.ph ], [ %add, %for.body ]<o:p></o:p></pre>
<pre>+ %add = add nuw nsw i32 %x.018, 1<o:p></o:p></pre>
<pre>+ %3 = load i32, i32* %channel, align 8<o:p></o:p></pre>
<pre>+ %mul = mul nsw i32 %3, %add<o:p></o:p></pre>
<pre>+ %idx.ext = sext i32 %mul to i64<o:p></o:p></pre>
<pre>+ %add.ptr = getelementptr inbounds i32, i32* %in, i64 %idx.ext<o:p></o:p></pre>
<pre>+ %4 = load i32, i32* %add.ptr, align 4<o:p></o:p></pre>
<pre>+ %mul1 = mul nsw i32 %0, %add<o:p></o:p></pre>
<pre>+ %idx.ext1 = sext i32 %mul1 to i64<o:p></o:p></pre>
<pre>+ %add.ptr1 = getelementptr inbounds i32, i32* %in, i64 %idx.ext1<o:p></o:p></pre>
<pre>+ %5 = load i32, i32* %add.ptr1, align 4<o:p></o:p></pre>
<pre>+ %6 = add i32 %4, %5<o:p></o:p></pre>
<pre>+ %cmp = icmp slt i32 %add, %length<o:p></o:p></pre>
<pre>+ br i1 %cmp, label %for.body, label %for.cond.cleanup.loopexit<o:p></o:p></pre>
<pre>+}<o:p></o:p></pre>
<pre>+<o:p></o:p></pre>
<pre>+<o:p></o:p></pre>
<pre>+define i32 @foo5(%struct.image* %input, i32 %length, i32* %in) {<o:p></o:p></pre>
<pre>+entry:<o:p></o:p></pre>
<pre>+ %stride = getelementptr inbounds %struct.image, %struct.image* %input, i64 0, i32 1<o:p></o:p></pre>
<pre>+ %0 = load i32, i32* %stride, align 4<o:p></o:p></pre>
<pre>+ %cmp17 = icmp sgt i32 %length, 1<o:p></o:p></pre>
<pre>+ br i1 %cmp17, label %for.body.lr.ph, label %for.cond.cleanup<o:p></o:p></pre>
<pre>+<o:p></o:p></pre>
<pre>+for.body.lr.ph: ; preds = %entry<o:p></o:p></pre>
<pre>+ %channel = getelementptr inbounds %struct.image, %struct.image* %input, i64 0, i32 0<o:p></o:p></pre>
<pre>+ br label %for.body<o:p></o:p></pre>
<pre>+<o:p></o:p></pre>
<pre>+for.cond.cleanup.loopexit: ; preds = %for.body<o:p></o:p></pre>
<pre>+ %1 = phi i32 [ %7, %for.body ]<o:p></o:p></pre>
<pre>+ br label %for.cond.cleanup<o:p></o:p></pre>
<pre>+<o:p></o:p></pre>
<pre>+for.cond.cleanup: ; preds = %for.cond.cleanup.loopexit, %entry<o:p></o:p></pre>
<pre>+ %2 = phi i32 [ 0, %entry ], [ %1, %for.cond.cleanup.loopexit ]<o:p></o:p></pre>
<pre>+ ret i32 %2<o:p></o:p></pre>
<pre>+<o:p></o:p></pre>
<pre>+; This example is the same as above except that the first mul is used in two places<o:p></o:p></pre>
<pre>+; and this may result in having two versions of the multiply: an i32 and i64 version.<o:p></o:p></pre>
<pre>+; In this case, keep the trucate instructions to avoid this redundancy.<o:p></o:p></pre>
<pre>+; CHECK: for.body:<o:p></o:p></pre>
<pre>+; CHECK: trunc<o:p></o:p></pre>
<pre>+for.body: ; preds = %for.body.lr.ph, %for.body<o:p></o:p></pre>
<pre>+ %x.018 = phi i32 [ 1, %for.body.lr.ph ], [ %add, %for.body ]<o:p></o:p></pre>
<pre>+ %add = add nuw nsw i32 %x.018, 1<o:p></o:p></pre>
<pre>+ %3 = load i32, i32* %channel, align 8<o:p></o:p></pre>
<pre>+ %mul = mul nsw i32 %3, %add<o:p></o:p></pre>
<pre>+ %idx.ext = sext i32 %mul to i64<o:p></o:p></pre>
<pre>+ %add.ptr = getelementptr inbounds i32, i32* %in, i64 %idx.ext<o:p></o:p></pre>
<pre>+ %4 = load i32, i32* %add.ptr, align 4<o:p></o:p></pre>
<pre>+ %mul1 = mul nsw i32 %0, %add<o:p></o:p></pre>
<pre>+ %idx.ext1 = sext i32 %mul1 to i64<o:p></o:p></pre>
<pre>+ %add.ptr1 = getelementptr inbounds i32, i32* %in, i64 %idx.ext1<o:p></o:p></pre>
<pre>+ %5 = load i32, i32* %add.ptr1, align 4<o:p></o:p></pre>
<pre>+ %6 = add i32 %4, %5<o:p></o:p></pre>
<pre>+ %7 = add i32 %6, %mul<o:p></o:p></pre>
<pre>+ %cmp = icmp slt i32 %add, %length<o:p></o:p></pre>
<pre>+ br i1 %cmp, label %for.body, label %for.cond.cleanup.loopexit<o:p></o:p></pre>
<pre>+}<o:p></o:p></pre>
<pre><o:p> </o:p></pre>
<pre><o:p> </o:p></pre>
<pre>_______________________________________________<o:p></o:p></pre>
<pre>llvm-commits mailing list<o:p></o:p></pre>
<pre><a href="mailto:llvm-commits@lists.llvm.org" moz-do-not-send="true">llvm-commits@lists.llvm.org</a><o:p></o:p></pre>
<pre><a href="http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits" moz-do-not-send="true">http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits</a><o:p></o:p></pre>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
</blockquote>
<br>
</body>
</html>