<html>
<head>
<meta http-equiv="Content-Type" content="text/html; charset=UTF-8">
</head>
<body>
<p>Aren't disappearing bugs fun? :)</p>
<p>Skimming the thread, one thing we might want to consider is
teaching the vectorizer how to widen an assume. This would attack
the problem from another angle. <br>
</p>
<p>For any assume(X), if we unroll that as:<br>
assume(X1)<br>
assume(X2)<br>
....<br>
assume(XN)</p>
<p>We could also widen that into a single assume as:<br>
%vec = {X1, X2, ..., XN)<br>
%c = horizontal_and(%vec)<br>
assume(%c)</p>
<p>The catch is that I don't think we've considered whether we want
this to be a canonical form or not. For the vectorizer with %vec
already existing, I think it makes good sense though.</p>
<p>Philip<br>
</p>
<div class="moz-cite-prefix">On 4/29/21 12:42 PM, Sanjay Patel
wrote:<br>
</div>
<blockquote type="cite"
cite="mid:CA+wODivNjncGTG5dx7+rM9Dk=8eTm0JCBti1F2oqB3z6HWaL+A@mail.gmail.com">
<meta http-equiv="content-type" content="text/html; charset=UTF-8">
<div dir="ltr">
<div>Ah - the degenerate case was based on:</div>
<div><a href="https://bugs.llvm.org/show_bug.cgi?id=49785#c2"
moz-do-not-send="true">https://bugs.llvm.org/show_bug.cgi?id=49785#c2</a></div>
<div><br>
</div>
<div>But I just reverted this locally:<br>
</div>
<div><a
href="https://reviews.llvm.org/rGbb907b26e2bf2a0d5ea3fcb13ba0ccef3420a38c"
moz-do-not-send="true">https://reviews.llvm.org/rGbb907b26e2bf2a0d5ea3fcb13ba0ccef3420a38c</a></div>
<div><br>
</div>
<div>And neither the original C example from PR49785 nor the IR
reduction are taking forever to compile for me. Something else
must have changed in the past 2 weeks to mask the problem. So
I'm not sure how to trigger this again.<br>
</div>
</div>
<br>
<div class="gmail_quote">
<div dir="ltr" class="gmail_attr">On Wed, Apr 28, 2021 at 5:19
PM Philip Reames <<a
href="mailto:listmail@philipreames.com"
moz-do-not-send="true">listmail@philipreames.com</a>>
wrote:<br>
</div>
<blockquote class="gmail_quote" style="margin:0px 0px 0px
0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">
<div>
<p>Sanjay,</p>
<p>Neither of these seem to be the degenerate test case you
saw the compile time impact on. Could you point me to
that? I'm looking for something to run to understand the
compile time opportunities in this code.</p>
<p>Philip<br>
</p>
<div>On 4/15/21 6:48 AM, Sanjay Patel wrote:<br>
</div>
<blockquote type="cite">
<div dir="ltr">
<div>I'm not understanding the question. 🤔<br>
</div>
<div><br>
</div>
<div>I suspect this doesn't answer it, but more of the
trail that led here:<br>
</div>
<div>There was a gap where isKnownNonEqual was
succeeding where simplifyICmpWithZero() did not - it
was plugged with:</div>
<div><a
href="https://reviews.llvm.org/rG5ae5d25e38efad1d59ed97d969a5e930b58a5e16"
target="_blank" moz-do-not-send="true">https://reviews.llvm.org/rG5ae5d25e38efad1d59ed97d969a5e930b58a5e16</a></div>
<div><br>
</div>
<div>That was exposed by this test (and that's what the
unittest in the above commit is based on):</div>
<div><a
href="https://github.com/llvm/llvm-project/blob/9345f9fa5d6401bf9b98ce915bf4fbde9d183ce7/llvm/test/Analysis/ValueTracking/monotonic-phi.ll#L185"
target="_blank" moz-do-not-send="true">https://github.com/llvm/llvm-project/blob/9345f9fa5d6401bf9b98ce915bf4fbde9d183ce7/llvm/test/Analysis/ValueTracking/monotonic-phi.ll#L185</a></div>
<div><br>
</div>
</div>
<br>
<div class="gmail_quote">
<div dir="ltr" class="gmail_attr">On Wed, Apr 14, 2021
at 6:04 PM Philip Reames <<a
href="mailto:listmail@philipreames.com"
target="_blank" moz-do-not-send="true">listmail@philipreames.com</a>>
wrote:<br>
</div>
<blockquote class="gmail_quote" style="margin:0px 0px
0px 0.8ex;border-left:1px solid
rgb(204,204,204);padding-left:1ex">Hm, this seems
interesting.<br>
<br>
I had been thinking about inverting this API by having
the shared code <br>
return the two pointers which imply
equality/inequality and having <br>
instcombine rewrite the condition. However, I
couldn't find a case <br>
where the existing logic didn't already do that
inversion through some <br>
other (more general) path. Do you happen to have an
IR fragment where <br>
we trip this example? I'd like to take a look at it
with that lens.<br>
<br>
Philip<br>
<br>
On 4/14/21 6:12 AM, Sanjay Patel via llvm-commits
wrote:<br>
> Author: Sanjay Patel<br>
> Date: 2021-04-14T09:04:15-04:00<br>
> New Revision:
7ef2c68a3d24af0b0d540e748e8b564180f4e18a<br>
><br>
> URL: <a
href="https://github.com/llvm/llvm-project/commit/7ef2c68a3d24af0b0d540e748e8b564180f4e18a"
rel="noreferrer" target="_blank"
moz-do-not-send="true">https://github.com/llvm/llvm-project/commit/7ef2c68a3d24af0b0d540e748e8b564180f4e18a</a><br>
> DIFF: <a
href="https://github.com/llvm/llvm-project/commit/7ef2c68a3d24af0b0d540e748e8b564180f4e18a.diff"
rel="noreferrer" target="_blank"
moz-do-not-send="true">https://github.com/llvm/llvm-project/commit/7ef2c68a3d24af0b0d540e748e8b564180f4e18a.diff</a><br>
><br>
> LOG: [InstSimplify] improve efficiency for
detecting non-zero value<br>
><br>
> Stepping through callstacks in the example from
D99759 reveals<br>
> this potential compile-time improvement.<br>
><br>
> The savings come from avoiding ValueTracking's
computing known<br>
> bits if we have already dealt with special-case
patterns.<br>
><br>
> Further improvements in this direction seem
possible.<br>
><br>
> This makes a degenerate test based on PR49785
about 40x faster<br>
> (25 sec -> 0.6 sec), but it does not address
the larger question<br>
> of how to limit computeKnownBitsFromAssume(). Ie,
the original<br>
> test there is still infinite-time for all
practical purposes.<br>
><br>
> Differential Revision: <a
href="https://reviews.llvm.org/D100408"
rel="noreferrer" target="_blank"
moz-do-not-send="true">https://reviews.llvm.org/D100408</a><br>
><br>
> Added:<br>
> <br>
><br>
> Modified:<br>
> llvm/lib/Analysis/InstructionSimplify.cpp<br>
><br>
> Removed:<br>
> <br>
><br>
><br>
>
################################################################################<br>
> diff --git
a/llvm/lib/Analysis/InstructionSimplify.cpp
b/llvm/lib/Analysis/InstructionSimplify.cpp<br>
> index b233a0f3eb2d..08f504a0ce37 100644<br>
> --- a/llvm/lib/Analysis/InstructionSimplify.cpp<br>
> +++ b/llvm/lib/Analysis/InstructionSimplify.cpp<br>
> @@ -3432,6 +3432,8 @@ static Value
*SimplifyICmpInst(unsigned Predicate, Value *LHS,
Value *RHS,<br>
> if (Value *V = simplifyICmpOfBools(Pred, LHS,
RHS, Q))<br>
> return V;<br>
> <br>
> + // TODO: Sink/common this with other
potentially expensive calls that use<br>
> + // ValueTracking? See comment below for
isKnownNonEqual().<br>
> if (Value *V = simplifyICmpWithZero(Pred,
LHS, RHS, Q))<br>
> return V;<br>
> <br>
> @@ -3637,7 +3639,9 @@ static Value
*SimplifyICmpInst(unsigned Predicate, Value *LHS,
Value *RHS,<br>
> }<br>
> <br>
> // icmp eq|ne X, Y -> false|true if X != Y<br>
> - if (ICmpInst::isEquality(Pred) &&<br>
> + // This is potentially expensive, and we have
already computedKnownBits for<br>
> + // compares with 0 above here, so only try
this for a non-zero compare.<br>
> + if (ICmpInst::isEquality(Pred) &&
!match(RHS, m_Zero()) &&<br>
> isKnownNonEqual(LHS, RHS, Q.DL, <a
href="http://Q.AC" rel="noreferrer" target="_blank"
moz-do-not-send="true">Q.AC</a>, Q.CxtI, Q.DT,
Q.IIQ.UseInstrInfo)) {<br>
> return Pred == ICmpInst::ICMP_NE ?
getTrue(ITy) : getFalse(ITy);<br>
> }<br>
><br>
><br>
> <br>
> _______________________________________________<br>
> llvm-commits mailing list<br>
> <a href="mailto:llvm-commits@lists.llvm.org"
target="_blank" moz-do-not-send="true">llvm-commits@lists.llvm.org</a><br>
> <a
href="https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits"
rel="noreferrer" target="_blank"
moz-do-not-send="true">https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits</a><br>
</blockquote>
</div>
</blockquote>
</div>
</blockquote>
</div>
</blockquote>
</body>
</html>