<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 15 (filtered medium)">
<style><!--
/* Font Definitions */
@font-face
{font-family:"Cambria Math";
panose-1:2 4 5 3 5 4 6 3 2 4;}
@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;
font-size:11.0pt;
font-family:"Calibri",sans-serif;}
a:link, span.MsoHyperlink
{mso-style-priority:99;
color:blue;
text-decoration:underline;}
span.EmailStyle20
{mso-style-type:personal-reply;
font-family:"Calibri",sans-serif;
color:windowtext;}
.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]-->
</head>
<body lang="EN-US" link="blue" vlink="purple" style="word-wrap:break-word">
<div class="WordSection1">
<p class="MsoNormal">> It is a common way to explain problems with -ffinite-math-only by user ignorance. However user misunderstandings and complaints may indicate a flaw in compiler implementation, which I believe we have in this case.<o:p></o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal">The *<b>intent</b>* of -ffast-math is to do absolutely everything possible to get the fastest floating point math possible. In other words, to pretend that floating point values are actually real numbers. In real math, there is no negative
0, or NaN, and infinity is not something you do algebra with. If you try to divide by 0, you just get points off your midterm; there’s no “behavior” that needs to be defined.<o:p></o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal">> Using NaN as sentinels is a natural way when you cannot spend extra memory for keeping flags for each item, spend extra cycles to read that flag and do not want to pollute cache. It does not depend on reading documentation or writing
the code from scratch. It is simply the best solution for storing data. If performance of the data processing is critical, -ffast-math is a good solution. This is a fairly legitimate use case. The fact that the compiler does not allow it is a compiler drawback.<o:p></o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal">The use case is legitimate, I agree, and the compiler does allow it. In fancyAlgo.cpp, have your input validation code that checks for NaN. in fancyAlgoFast.cpp, have your ML kernel, compiled with fast-math. If you really *<b>must</b>*
have them all be in the same TU, then compile with -fhonor-nans.<o:p></o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal">> It is a workaround, it works in some cases but does not in others. ML kernel often is a single translation unit, there may be no such thing as linker for that processor. At the same time it is computation intensive and using fast-math
in it may be very profitable.<o:p></o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal">It's a workaround for a well defined behavior that you opted into. If having multiple TU’s is not an option, then you can do some of the other solutions presented or compile with -fhonor-nans.
<o:p></o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal">> ICC and MSVC do not remove `isnan` in fast math mode. If `isnan` is implemented in libc, it is also a real check. Actually removing `isnan` creates inconsistency.<o:p></o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal">I don’t know about ICC, but this is a quote from the docs for MSVC: “Special values (NaN, +infinity, -infinity, -0.0) may not be propagated or behave strictly according to the IEEE-754 standard”. My interpretation of this is the same as
that for clang: no guarantees are provided. Maybe it works today, but that doesn’t mean it will work tomorrow.<o:p></o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal">> Removing `isnan` is only an optimization, it does not intend to change semantics. So it cannot create portability issues. Quite the contrary, it helps portability by making behavior consistent between compilers and libc implementations.
The only possible issue is performance loss, this is discussed above, it is an unlikely case. Anyway, if such loss exists and it is absolutely intolerable for a user, a hack with redefinition of `isnan` restores the previous code generation.<o:p></o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal">It creates a portability issue if we change the compiler to guarantee that isnan(x) works, and then users rely on it working. The resulting code is non portable to other compilers, and to old versions of clang.<o:p></o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal">> The documentation says about -ffinite-math-only:<o:p></o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal"> > "Allow optimizations for floating-point arithmetic that assume that arguments and results are not NaNs or +-Infs."<o:p></o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal">> Is it clear whether `isnan` is arithmetic or not?<o:p></o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal">The docs for -ffinite-math-only on <a href="https://clang.llvm.org/docs/UsersManual.html">
https://clang.llvm.org/docs/UsersManual.html</a> actually say “Allow floating-point optimizations that assume arguments and results are not NaNs or +-Inf”. Notice that it says nothing about arithmetic. I’m not sure where you got your quote, but it’s not from
the docs for the currently available release of upstream clang.<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"> Chris Tetreault<o:p></o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<div>
<div style="border:none;border-top:solid #E1E1E1 1.0pt;padding:3.0pt 0in 0in 0in">
<p class="MsoNormal"><b>From:</b> Serge Pavlov <sepavloff@gmail.com> <br>
<b>Sent:</b> Thursday, September 16, 2021 8:23 PM<br>
<b>To:</b> Chris Tetreault <ctetreau@quicinc.com><br>
<b>Cc:</b> Michael Kruse <cfe-dev@meinersbur.de>; llvm-dev@lists.llvm.org; cfe-dev@lists.llvm.org<br>
<b>Subject:</b> Re: [cfe-dev] [llvm-dev] Should isnan be optimized out in fast-math mode?<o:p></o:p></p>
</div>
</div>
<p class="MsoNormal"><o:p> </o:p></p>
<p align="center" style="text-align:center"><strong><span style="font-size:10.5pt;font-family:"Arial",sans-serif;color:black;background:yellow">WARNING:</span></strong><span style="font-size:10.5pt;font-family:"Arial",sans-serif;color:black;background:yellow">
This email originated from outside of Qualcomm. Please be wary of any links or attachments, and do not enable macros.</span><o:p></o:p></p>
<div>
<div>
<div>
<p class="MsoNormal">On Fri, Sep 17, 2021 at 3:11 AM Chris Tetreault <<a href="mailto:ctetreau@quicinc.com">ctetreau@quicinc.com</a>> wrote:<o:p></o:p></p>
</div>
<div>
<blockquote style="border:none;border-left:solid #CCCCCC 1.0pt;padding:0in 0in 0in 6.0pt;margin-left:4.8pt;margin-right:0in">
<p class="MsoNormal">The difference there is that doing pointer arithmetic on null pointers doesn't *usually* work, unless you turn on -ffast-pointers.<br>
<br>
It seems to me that most confusion related to -ffast-math is likely caused by people who are transitioning to using it. I have some codebase, and I turn on fast math, and then a few months down the road I notice a strangeness that I did not catch during the
initial transition period. If you're writing new code with fast-math, you don't do things like try to use NaN as a sentinel value in a TU with fast math turned on. This is the sort of thing you catch when you try to transition an existing codebase. Forgive
me for the uncharitable interpretation, but it's much easier to ask the compiler to change to accommodate your use case than it is to refactor your code.<o:p></o:p></p>
</blockquote>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">It is a common way to explain problems with -ffinite-math-only by user ignorance. However user misunderstandings and complaints may indicate a flaw in compiler implementation, which I believe we have in this case.<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">Using NaN as sentinels is a natural way when you cannot spend extra memory for keeping flags for each item, spend extra cycles to read that flag and do not want to pollute cache. It does not depend on reading documentation or writing the
code from scratch. It is simply the best solution for storing data. If performance of the data processing is critical, -ffast-math is a good solution. This is a fairly legitimate use case. The fact that the compiler does not allow it is a compiler drawback.<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<blockquote style="border:none;border-left:solid #CCCCCC 1.0pt;padding:0in 0in 0in 6.0pt;margin-left:4.8pt;margin-right:0in">
<p class="MsoNormal"><br>
To me, I think Mehdi had the best solution: The algorithm that is the bottleneck, and experiences the huge speedup using fast-math should be separated into its own source file. This source file, and only this source file should be compiled with fast-math. The
outer driver loop should not be compiled with fast math. This solution is clean, (probably) easy, and doesn't require a change in the compiler.<o:p></o:p></p>
</blockquote>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">It is a workaround, it works in some cases but does not in others. ML kernel often is a single translation unit, there may be no such thing as linker for that processor. At the same time it is computation intensive and using fast-math in
it may be very profitable.<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"> <o:p></o:p></p>
</div>
<blockquote style="border:none;border-left:solid #CCCCCC 1.0pt;padding:0in 0in 0in 6.0pt;margin-left:4.8pt;margin-right:0in">
<p class="MsoNormal">Changing the compiler is hard, affects everybody who uses the compiler, and creates inconsistency in behavior between clang and gcc (and msvc with /fp:fast), and clang and old versions of clang.<o:p></o:p></p>
</blockquote>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">ICC and MSVC do not remove `isnan` in fast math mode. If `isnan` is implemented in libc, it is also a real check. Actually removing `isnan` creates inconsistency.<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"> <o:p></o:p></p>
</div>
<blockquote style="border:none;border-left:solid #CCCCCC 1.0pt;padding:0in 0in 0in 6.0pt;margin-left:4.8pt;margin-right:0in">
<p class="MsoNormal">The behavior of fast-math with respect to NaN is consistent across the mainstream c/c++ compilers: no promises are made, users should not assume that they can use it for anything. Changing it now would create a major portability issue for
user codebases, which in and of itself is a very strong reason to not make this change.<o:p></o:p></p>
</blockquote>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">Removing `isnan` is only an optimization, it does not intend to change semantics. So it cannot create portability issues. Quite the contrary, it helps portability by making behavior consistent between compilers and libc implementations.
The only possible issue is performance loss, this is discussed above, it is an unlikely case. Anyway, if such loss exists and it is absolutely intolerable for a user, a hack with redefinition of `isnan` restores the previous code generation.<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<blockquote style="border:none;border-left:solid #CCCCCC 1.0pt;padding:0in 0in 0in 6.0pt;margin-left:4.8pt;margin-right:0in">
<p class="MsoNormal"><br>
If the behavior is confusing to users, that's because it's poorly explained. <o:p>
</o:p></p>
</blockquote>
<div>
<p class="MsoNormal"> <o:p></o:p></p>
</div>
<div>
<div>
<p class="MsoNormal">What is confusing? That the explicitly written call to a function is not removed? According to user feedback it is the silent removal of `isnan` that confuses users.<o:p></o:p></p>
</div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<blockquote style="border:none;border-left:solid #CCCCCC 1.0pt;padding:0in 0in 0in 6.0pt;margin-left:4.8pt;margin-right:0in">
<p class="MsoNormal">Honestly, I think the docs are pretty clear, but "It's clear, you just need to learn to read" is never an acceptable answer so it could certainly be improved. This is the only thing that needs to be fixed in my opinion.<o:p></o:p></p>
</blockquote>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">The documentation says about -ffinite-math-only:<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal"> "Allow optimizations for floating-point arithmetic that assume that arguments and results are not NaNs or +-Infs."<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">Is it clear whether `isnan` is arithmetic or not?<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<blockquote style="border:none;border-left:solid #CCCCCC 1.0pt;padding:0in 0in 0in 6.0pt;margin-left:4.8pt;margin-right:0in">
<p class="MsoNormal"><br>
Thanks,<br>
Chris Tetreault<br>
<br>
-----Original Message-----<br>
From: Michael Kruse <<a href="mailto:cfe-dev@meinersbur.de" target="_blank">cfe-dev@meinersbur.de</a>>
<br>
Sent: Thursday, September 16, 2021 12:29 PM<br>
To: Chris Tetreault <<a href="mailto:ctetreau@quicinc.com" target="_blank">ctetreau@quicinc.com</a>><br>
Cc: Serge Pavlov <<a href="mailto:sepavloff@gmail.com" target="_blank">sepavloff@gmail.com</a>>;
<a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a>;
<a href="mailto:cfe-dev@lists.llvm.org" target="_blank">cfe-dev@lists.llvm.org</a><br>
Subject: Re: [cfe-dev] [llvm-dev] Should isnan be optimized out in fast-math mode?<br>
<br>
WARNING: This email originated from outside of Qualcomm. Please be wary of any links or attachments, and do not enable macros.<br>
<br>
Am Mo., 13. Sept. 2021 um 11:46 Uhr schrieb Chris Tetreault via cfe-dev <<a href="mailto:cfe-dev@lists.llvm.org" target="_blank">cfe-dev@lists.llvm.org</a>>:<br>
> As a user, if I read that:<br>
><br>
><br>
><br>
> ```<br>
><br>
> if (isnan(x)) {<br>
><br>
> ```<br>
><br>
><br>
><br>
> … is guaranteed to work, and I read that fast-math enables the compiler to reason about constructs like `x + 0` being equal to `x`, then I’m going to be very confused when:<br>
><br>
><br>
><br>
> ```<br>
><br>
> if (isnan(x + 0)) {<br>
><br>
> ```<br>
><br>
><br>
><br>
> … does not also work. I’m going to open a bug and complain, and the slide down the slippery slope will continue. You and I understand the difference, and the technical reason why `isnan(x)` is supported but `isnan(x + 0)` isn’t, but Joe Coder just trying
to figure out why he’s got NaN in his matrices despite his careful NaN handling code. Joe is not a compiler expert, and on the face of it, it seems like a silly limitation. This will never end until fast-math is gutted.<br>
<br>
C/C++ already has cases like this. Pointer arithmetic on null pointers is undefined behaviour, even if adding[1,2]/subtracting[3] zero. I don't think it is too far fetched to expect from users to know that an operation is undefined behaviour even if one of
the operands is zero.<br>
<br>
Michael<br>
<br>
[1] <a href="https://github.com/llvm/llvm-project/blob/main/clang/test/Sema/pointer-addition.c" target="_blank">
https://github.com/llvm/llvm-project/blob/main/clang/test/Sema/pointer-addition.c</a><br>
[2] <a href="https://github.com/llvm/llvm-project/blob/main/compiler-rt/test/ubsan/TestCases/Pointer/nullptr-and-nonzero-offset-constants.cpp" target="_blank">
https://github.com/llvm/llvm-project/blob/main/compiler-rt/test/ubsan/TestCases/Pointer/nullptr-and-nonzero-offset-constants.cpp</a><br>
[3] <a href="https://github.com/llvm/llvm-project/blob/main/clang/test/Sema/pointer-subtraction.c" target="_blank">
https://github.com/llvm/llvm-project/blob/main/clang/test/Sema/pointer-subtraction.c</a><br>
<br>
<br>
Michael<o:p></o:p></p>
</blockquote>
</div>
</div>
</div>
</div>
</body>
</html>