<html><head><meta http-equiv="Content-Type" content="text/html charset=windows-1252"></head><body style="word-wrap: break-word; -webkit-nbsp-mode: space; -webkit-line-break: after-white-space;" class="">Replying to new LLVM-commits..<div class=""><br class=""><div><blockquote type="cite" class=""><div class="">On Aug 5, 2015, at 2:32 PM, Mikhail Zolotukhin <<a href="mailto:mzolotukhin@apple.com" class="">mzolotukhin@apple.com</a>> wrote:</div><br class="Apple-interchange-newline"><div class=""><meta http-equiv="Content-Type" content="text/html charset=windows-1252" class=""><div style="word-wrap: break-word; -webkit-nbsp-mode: space; -webkit-line-break: after-white-space;" class="">Hi James,<div class=""><br class=""></div><div class="">Thanks for the good explanation, it illustrates the issues I spotted earlier very well.<br class=""><div class=""><blockquote type="cite" class=""><div class="">On Aug 4, 2015, at 2:16 AM, James Molloy <<a href="mailto:James.Molloy@arm.com" class="">James.Molloy@arm.com</a>> wrote:</div><br class="Apple-interchange-newline"><div class="">
<meta http-equiv="Content-Type" content="text/html; charset=Windows-1252" class="">
<div style="word-wrap: break-word; -webkit-nbsp-mode: space; -webkit-line-break: after-white-space;" class="">
Hi,
<div class=""><br class="">
</div>
<div class="">This all depends on how you define your intrinsic. It’s currently ambiguous - we have to choose what uabsdiff should do.</div>
<div class=""><br class="">
</div>
<div class="">Given this expression:</div>
<div class=""><br class="">
</div>
<div class="">call i8 @llvm.uabsdiff(i8 1, i8 130)</div>
<div class=""><br class="">
</div>
<div class="">What should be the result?</div>
<div class=""><br class="">
</div>
<div class="">If we expand, we get:</div>
<div class=""><br class="">
</div>
<div class="">%1 = sub i8 1, i8 130 ; pure result is -129, but has to be truncated to fit into i8.</div>
<div class=""><br class="">
</div>
<div class="">Neither unsigned nor signed comparisons will get the right result here. The result of the subtract after truncation is 127 in both signed and unsigned representations, which is greater than zero and is definitely not 129!</div>
<div class=""><br class="">
</div>
<div class="">Instead, if we expand with range expansion:</div>
<div class=""><br class="">
</div>
<div class="">%1 = zext i8 1 to i16 ; zext because we’re treating the inputs as unsigned</div>
<div class="">%2 = zext i8 130 to i16</div>
<div class="">%3 = sub i16 %1, %1 ; 0xff7f = -129</div>
<div class="">%4 = icmp sgt i16 %3, 0 ; false</div>
<div class="">%5 = sub i16 0, %3 ; 129</div>
<div class="">%6 = select i1 %4, i16 %3, %5 ; %5 = 129</div>
<div class="">%7 = trunc i16 %6 to i8 ; 129</div>
<div class=""><br class="">
</div>
<div class="">So the end result is correct and in range, if we promote the intermediate calculations to a larger bit width.</div>
<div class=""><br class="">
</div>
<div class="">So we need to decide: what should the behaviour of uabsdiff be? Should it behave as if there is larger intermediate precision or not?</div></div></div></blockquote>Would it be possible to consistently recognize such patterns in backends?</div><div class=""><br class=""></div><div class="">Michael</div><div class=""><br class=""><blockquote type="cite" class=""><div class=""><div style="word-wrap: break-word; -webkit-nbsp-mode: space; -webkit-line-break: after-white-space;" class="">
<div class=""><br class="">
</div>
<div class="">I vote yes, because all the actual absdiff implementations I know of implement this behaviour.</div>
<div class=""><br class="">
</div>
<div class="">Cheers,</div>
<div class=""><br class="">
</div>
<div class="">James</div>
<div class=""><br class="">
</div>
<div class="">
<div class="">
<div class="">On 4 Aug 2015, at 09:56, Shahid, Asghar-ahmad <<a href="mailto:Asghar-ahmad.Shahid@amd.com" class="">Asghar-ahmad.Shahid@amd.com</a>> wrote:</div>
<br class="Apple-interchange-newline">
<blockquote type="cite" class="">
<div lang="EN-US" link="blue" vlink="purple" style="font-family: Helvetica; font-size: 12px; font-style: normal; font-variant: normal; font-weight: normal; letter-spacing: normal; line-height: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px;" class="">
<div class="WordSection1" style="page: WordSection1;">
<div style="margin: 0in 0in 0.0001pt; font-size: 12pt; font-family: 'Times New Roman', serif;" class="">
<span style="font-size: 11pt; font-family: Calibri, sans-serif; color: rgb(31, 73, 125);" class="">Hi Mikhail,<o:p class=""></o:p></span></div>
<div style="margin: 0in 0in 0.0001pt; font-size: 12pt; font-family: 'Times New Roman', serif;" class="">
<span style="font-size: 11pt; font-family: Calibri, sans-serif; color: rgb(31, 73, 125);" class=""> </span></div>
<div style="margin: 0in 0in 0.0001pt; font-size: 12pt; font-family: 'Times New Roman', serif;" class="">
<span style="font-size: 11pt; font-family: Calibri, sans-serif; color: rgb(31, 73, 125);" class="">AFAIU, I think we need to use “ugt” as the “modulo 2^n” result of the difference for unsigned overflow<o:p class=""></o:p></span></div>
<div style="margin: 0in 0in 0.0001pt; font-size: 12pt; font-family: 'Times New Roman', serif;" class="">
<span style="font-size: 11pt; font-family: Calibri, sans-serif; color: rgb(31, 73, 125);" class="">will be wrapped to zero.<o:p class=""></o:p></span></div>
<div style="margin: 0in 0in 0.0001pt; font-size: 12pt; font-family: 'Times New Roman', serif;" class="">
<span style="font-size: 11pt; font-family: Calibri, sans-serif; color: rgb(31, 73, 125);" class=""> </span></div>
<div style="margin: 0in 0in 0.0001pt; font-size: 12pt; font-family: 'Times New Roman', serif;" class="">
<span style="font-size: 11pt; font-family: Calibri, sans-serif; color: rgb(31, 73, 125);" class="">But I too would be happy to be corrected if I'm wrong:)<o:p class=""></o:p></span></div>
<div style="margin: 0in 0in 0.0001pt; font-size: 12pt; font-family: 'Times New Roman', serif;" class="">
<span style="font-size: 11pt; font-family: Calibri, sans-serif; color: rgb(31, 73, 125);" class=""> </span></div>
<div style="margin: 0in 0in 0.0001pt; font-size: 12pt; font-family: 'Times New Roman', serif;" class="">
<span style="font-size: 11pt; font-family: Calibri, sans-serif; color: rgb(31, 73, 125);" class="">Regards,<o:p class=""></o:p></span></div>
<div style="margin: 0in 0in 0.0001pt; font-size: 12pt; font-family: 'Times New Roman', serif;" class="">
<span style="font-size: 11pt; font-family: Calibri, sans-serif; color: rgb(31, 73, 125);" class="">Shahid<o:p class=""></o:p></span></div>
<div style="margin: 0in 0in 0.0001pt; font-size: 12pt; font-family: 'Times New Roman', serif;" class="">
<span style="font-size: 11pt; font-family: Calibri, sans-serif; color: rgb(31, 73, 125);" class=""> </span></div>
<div style="border-style: none none none solid; border-left-color: blue; border-left-width: 1.5pt; padding: 0in 0in 0in 4pt;" class="">
<div class="">
<div style="border-style: solid none none; border-top-color: rgb(181, 196, 223); border-top-width: 1pt; padding: 3pt 0in 0in;" class="">
<div style="margin: 0in 0in 0.0001pt; font-size: 12pt; font-family: 'Times New Roman', serif;" class="">
<b class=""><span style="font-size: 10pt; font-family: Tahoma, sans-serif;" class="">From:</span></b><span style="font-size: 10pt; font-family: Tahoma, sans-serif;" class=""><span class="Apple-converted-space"> </span>Mikhail Zolotukhin [<a href="mailto:mzolotukhin@apple.com" style="color: purple; text-decoration: underline;" class="">mailto:mzolotukhin@apple.com</a>]<span class="Apple-converted-space"> </span><br class="">
<b class="">Sent:</b><span class="Apple-converted-space"> </span>Tuesday, August 04, 2015 3:02 AM<br class="">
<b class="">To:</b><span class="Apple-converted-space"> </span>James Molloy<br class="">
<b class="">Cc:</b><span class="Apple-converted-space"> </span><a href="mailto:reviews+D11678+public+e92bec0f352bb617@reviews.llvm.org" style="color: purple; text-decoration: underline;" class="">reviews+D11678+public+e92bec0f352bb617@reviews.llvm.org</a>; Shahid, Asghar-ahmad;<span class="Apple-converted-space"> </span><a href="mailto:james.molloy@arm.com" style="color: purple; text-decoration: underline;" class="">james.molloy@arm.com</a>;<span class="Apple-converted-space"> </span><a href="mailto:hfinkel@anl.gov" style="color: purple; text-decoration: underline;" class="">hfinkel@anl.gov</a>;<span class="Apple-converted-space"> </span><a href="mailto:llvm-commits@cs.uiuc.edu" style="color: purple; text-decoration: underline;" class="">llvm-commits@cs.uiuc.edu</a><br class="">
<b class="">Subject:</b><span class="Apple-converted-space"> </span>Re: [PATCH] D11678: [CodeGen] Fixes *absdiff* intrinsic: LangRef doc/test case improvement and corresponding code change<o:p class=""></o:p></span></div>
</div>
</div>
<div style="margin: 0in 0in 0.0001pt; font-size: 12pt; font-family: 'Times New Roman', serif;" class="">
<o:p class=""> </o:p></div>
<div style="margin: 0in 0in 0.0001pt; font-size: 12pt; font-family: 'Times New Roman', serif;" class="">
Hi James,<o:p class=""></o:p></div>
<div class="">
<div style="margin: 0in 0in 0.0001pt; font-size: 12pt; font-family: 'Times New Roman', serif;" class="">
<o:p class=""> </o:p></div>
</div>
<div class="">
<div style="margin: 0in 0in 0.0001pt; font-size: 12pt; font-family: 'Times New Roman', serif;" class="">
According to LLVM Language reference manual, arithmetic operations do perform truncation:<o:p class=""></o:p></div>
</div>
<div class="">
<div style="margin: 0in 0in 0.0001pt; font-size: 12pt; font-family: 'Times New Roman', serif;" class="">
<span style="font-size: 10.5pt; font-family: 'Lucida Sans Unicode', sans-serif; background-color: white; background-position: initial initial; background-repeat: initial initial;" class="">If the difference has unsigned overflow, the result returned is the mathematical
result modulo 2</span><sup class=""><span style="font-family: 'Lucida Sans Unicode', sans-serif;" class="">n</span></sup><span style="font-size: 10.5pt; font-family: 'Lucida Sans Unicode', sans-serif; background-color: white; background-position: initial initial; background-repeat: initial initial;" class="">,
where n is the bit width of the result.</span><o:p class=""></o:p></div>
</div>
<div class="">
<div style="margin: 0in 0in 0.0001pt; font-size: 12pt; font-family: 'Times New Roman', serif;" class="">
<o:p class=""> </o:p></div>
</div>
<div class="">
<div style="margin: 0in 0in 0.0001pt; font-size: 12pt; font-family: 'Times New Roman', serif;" class="">
So, I think that if we use "uge" for unsigned, we're actually fine here. But I'd be happy to be corrected if I'm wrong here:)<o:p class=""></o:p></div>
</div>
<div class="">
<div style="margin: 0in 0in 0.0001pt; font-size: 12pt; font-family: 'Times New Roman', serif;" class="">
<o:p class=""> </o:p></div>
</div>
<div class="">
<div style="margin: 0in 0in 0.0001pt; font-size: 12pt; font-family: 'Times New Roman', serif;" class="">
Michael<o:p class=""></o:p></div>
</div>
<div class="">
<div style="margin: 0in 0in 0.0001pt; font-size: 12pt; font-family: 'Times New Roman', serif;" class="">
<o:p class=""> </o:p></div>
<div class="">
<blockquote style="margin-top: 5pt; margin-bottom: 5pt;" class="">
<div class="">
<div style="margin: 0in 0in 0.0001pt; font-size: 12pt; font-family: 'Times New Roman', serif;" class="">
On Aug 3, 2015, at 1:09 PM, James Molloy <<a href="mailto:james@jamesmolloy.co.uk" style="color: purple; text-decoration: underline;" class="">james@jamesmolloy.co.uk</a>> wrote:<o:p class=""></o:p></div>
</div>
<div style="margin: 0in 0in 0.0001pt; font-size: 12pt; font-family: 'Times New Roman', serif;" class="">
<o:p class=""> </o:p></div>
<div class="">
<div style="margin: 0in 0in 0.0001pt; font-size: 12pt; font-family: 'Times New Roman', serif;" class="">
Hi Michael,<br class="">
<br class="">
I think the complexity comes from the subtraction having as its domain two unsigned integers- so it's range must be a larger signed integer.<br class="">
<br class="">
Signed comparison for unsigned values is clearly wrong as you say, but I could contruct a testcase that shows incorrect behaviour with an unsigned comparison too. I think the only correct behaviour is to extend the inputs first and truncate the result.<span class="Apple-converted-space"> </span><br class="">
<br class="">
But I've been wrong before :)<br class="">
<br class="">
James<o:p class=""></o:p></div>
<div class="">
<div class="">
<div style="margin: 0in 0in 0.0001pt; font-size: 12pt; font-family: 'Times New Roman', serif;" class="">
On Mon, 3 Aug 2015 at 21:05, Michael Zolotukhin <<a href="mailto:mzolotukhin@apple.com" style="color: purple; text-decoration: underline;" class="">mzolotukhin@apple.com</a>> wrote:<o:p class=""></o:p></div>
</div>
<blockquote style="border-style: none none none solid; border-left-color: rgb(204, 204, 204); border-left-width: 1pt; padding: 0in 0in 0in 6pt; margin-left: 4.8pt; margin-right: 0in;" class="">
<div style="margin: 0in 0in 0.0001pt; font-size: 12pt; font-family: 'Times New Roman', serif;" class="">
mzolotukhin added inline comments.<br class="">
<br class="">
================<br class="">
Comment at: docs/LangRef.rst:10387-10390<br class="">
@@ -10386,6 +10386,6 @@<br class="">
<br class="">
%sub = sub nsw <4 x i32> %a, %b<br class="">
- %ispos = icmp sgt <4 x i32> %sub, <i32 -1, i32 -1, i32 -1, i32 -1><br class="">
+ %ispos = icmp sge <4 x i32> %sub, zeroinitializer<br class="">
%neg = sub nsw <4 x i32> zeroinitializer, %sub<br class="">
%1 = select <4 x i1> %ispos, <4 x i32> %sub, <4 x i32> %neg<br class="">
<br class="">
----------------<br class="">
ashahid wrote:<br class="">
> mzolotukhin wrote:<br class="">
> > What's the difference between `llvm.uabsdiff` and `llvm.sabsdiff` then?<br class="">
> The difference is the presence of NSW flag in case of llvm.sabsdiff.<br class="">
I still don't think it's correct. NSW is just a hint to optimizers, but it doesn't add any additional logic. It does assert that the expression won't overflow, but the operations we execute are still the same. That is, currently the only difference between
signed and unsigned version is that for signed version we could get an undefined behavior in some cases. This is clearly incorrect, because we should get different results without undefined behavior in some cases (e.g. `<-1,-1,-1,-1>` and `<1,1,1,1>` - it
should give `<254,254,254,254>` for `uabsdiff.v4i8` and `<2,2,2,2>` for `sabsdiff.v4i8`).<br class="">
<br class="">
What really should be the difference, as far is I understand, is condition code in the comparison:<br class="">
```<br class="">
%ispos = icmp sge <4 x i32> %sub, zeroinitializer<br class="">
```<br class="">
As far as I understand, we should use `uge` for unsigned and `sge` for signed case.<br class="">
<br class="">
<br class="">
<br class="">
Repository:<br class="">
rL LLVM<br class="">
<br class="">
<a href="http://reviews.llvm.org/D11678" target="_blank" style="color: purple; text-decoration: underline;" class="">http://reviews.llvm.org/D11678</a><br class="">
<br class="">
<br class="">
<br class="">
<br class="">
_______________________________________________<br class="">
llvm-commits mailing list<br class="">
<a href="mailto:llvm-commits@cs.uiuc.edu" target="_blank" style="color: purple; text-decoration: underline;" class="">llvm-commits@cs.uiuc.edu</a><br class="">
<a href="http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits" target="_blank" style="color: purple; text-decoration: underline;" class="">http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits</a></div>
</blockquote>
</div>
</div>
</blockquote>
</div>
</div>
</div>
</div>
</div>
</blockquote>
</div>
<br class="">
</div>
<br class="">
<font face="Arial" size="2" class="">-- IMPORTANT NOTICE: The contents of this email and any attachments are confidential and may also be privileged. If you are not the intended recipient, please notify the sender immediately and do not disclose the contents
to any other person, use it for any purpose, or store or copy the information in any medium. Thank you.<br class="">
<br class="">
ARM Limited, Registered office 110 Fulbourn Road, Cambridge CB1 9NJ, Registered in England & Wales, Company No: 2557590<br class="">
ARM Holdings plc, Registered office 110 Fulbourn Road, Cambridge CB1 9NJ, Registered in England & Wales, Company No: 2548782<br class="">
</font>
</div>
</div></blockquote></div><br class=""></div></div></div></blockquote></div><br class=""></div></body></html>