<div dir="ltr"><div class="gmail_extra"><div><div class="gmail_signature">On Thu, Feb 8, 2018 at 4:04 PM, Rafael Avila de Espindola <span dir="ltr"><<a href="mailto:rafael.espindola@gmail.com" target="_blank">rafael.espindola@gmail.com</a>></span> wrote:<br></div></div><div class="gmail_quote"><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">Looking a bit more on why I might not be measuring a performance<br>
improvement I noticed that the callgraph I was using was missing<br>
conditional calls. The attached script fixes that.<br>
<br>
I have uploaded a new version of the test with the complete call graph<br>
to <a href="https://s3-us-west-2.amazonaws.com/linker-tests/t2.tar.xz" rel="noreferrer" target="_blank">https://s3-us-west-2.<wbr>amazonaws.com/linker-tests/t2.<wbr>tar.xz</a>.<br>
<br>
I also noticed that we were not considering the case of multiple symbols<br>
in the same section. The attached patch fixes that.<br></blockquote><div><br></div><div>This is already handled in `CallGraphSort::CallGraphSort()` here:</div><div><br></div><div><div> NodeIndex From = GetOrCreateNode(FromSB);</div><div> NodeIndex To = GetOrCreateNode(ToSB);</div><div><br></div><div> Nodes[To].Weight = SaturatingAdd(Nodes[To].Weight, Weight);</div><div><br></div><div> if (>From == To)</div><div> continue;</div></div><div><br></div><div>It's specifically done here after the node weight adjustment so that density calculation later on takes this into account.</div><div><br></div><div>
<span style="color:rgb(34,34,34);font-family:arial,sans-serif;font-size:small;font-style:normal;font-variant-ligatures:normal;font-variant-caps:normal;font-weight:400;letter-spacing:normal;text-align:start;text-indent:0px;text-transform:none;white-space:normal;word-spacing:0px;background-color:rgb(255,255,255);text-decoration-style:initial;text-decoration-color:initial;float:none;display:inline">- Michael Spencer</span>
<br></div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">
<br>
Even with these changes I still get a iTLB regression.<br>
<br>
I am now going to try building hfsort and compare its results.<br>
<br>
Please upload a new patch on top of tree and include the attached fixes.<br>
<br>
<br>diff --git a/ELF/Driver.cpp b/ELF/Driver.cpp<br>
index 1fc9a0ad5..a8d28de4c 100644<br>
--- a/ELF/Driver.cpp<br>
+++ b/ELF/Driver.cpp<br>
@@ -590,8 +590,12 @@ static void readCallGraph(MemoryBufferRef MB) {<br>
fatal("parse error");<br>
InputSectionBase *FromSec = SymbolSection.lookup(Fields[0]<wbr>);<br>
InputSectionBase *ToSec = SymbolSection.lookup(Fields[1]<wbr>);<br>
- if (FromSec && ToSec)<br>
- Config->CallGraphProfile[std::<wbr>make_pair(FromSec, ToSec)] = Count;<br>
+ if (FromSec == ToSec)<br>
+ continue;<br>
+ if (FromSec && ToSec) {<br>
+ uint64_t &V = Config->CallGraphProfile[std::<wbr>make_pair(FromSec, ToSec)];<br>
+ V += Count;<br>
+ }<br>
}<br>
}<br>
<br>
<br><br>
Thanks,<br>
Rafael<br>
<br>
Michael Spencer <<a href="mailto:bigcheesegs@gmail.com">bigcheesegs@gmail.com</a>> writes:<br>
<br>
> On Thu, Feb 8, 2018 at 10:41 AM, Rafael Avila de Espindola <<br>
> <a href="mailto:rafael.espindola@gmail.com">rafael.espindola@gmail.com</a>> wrote:<br>
><br>
>> Michael Spencer <<a href="mailto:bigcheesegs@gmail.com">bigcheesegs@gmail.com</a>> writes:<br>
>><br>
>> > On Tue, Feb 6, 2018 at 6:53 PM, Rafael Avila de Espindola <<br>
>> > <a href="mailto:rafael.espindola@gmail.com">rafael.espindola@gmail.com</a>> wrote:<br>
>> ><br>
>> >> I have benchmarked this by timing lld ltoing FileCheck. The working set<br>
>> >> is much larger this time. The old callgraph had 4079 calls, this one has<br>
>> >> 30616.<br>
>> >><br>
>> >> The results are somewhat similar:<br>
>> >><br>
>> >> Performance counter stats for '../default-ld.lld @response.txt' (10<br>
>> runs):<br>
>> >><br>
>> >> 498,771 iTLB-load-misses<br>
>> >> ( +- 0.10% )<br>
>> >> 224,751,360 L1-icache-load-misses<br>
>> >> ( +- 0.00% )<br>
>> >><br>
>> >> 2.339864606 seconds time elapsed<br>
>> >> ( +- 0.06% )<br>
>> >><br>
>> >> Performance counter stats for '../sorted-ld.lld @response.txt' (10<br>
>> runs):<br>
>> >><br>
>> >> 556,999 iTLB-load-misses<br>
>> >> ( +- 0.17% )<br>
>> >> 216,788,838 L1-icache-load-misses<br>
>> >> ( +- 0.01% )<br>
>> >><br>
>> >> 2.326596163 seconds time elapsed<br>
>> >> ( +- 0.04% )<br>
>> >><br>
>> >> As with the previous test iTLB gets worse and L1 gets better. The net<br>
>> >> result is a very small speedup.<br>
>> >><br>
>> >> Do you know how big the chromium call graph is?<br>
>> >><br>
>> ><br>
>> > Not sure, but the call graph for a high profile internal game I tested is<br>
>> > about 10k functions and 17 MiB of .text, and I got a %2-%4 speedup.<br>
>> Given<br>
>> > that it's a game it runs a decent portion of that 17MiB 60 times a<br>
>> second,<br>
>> > while llvm is heavily pass based, so I don't expect the instruction<br>
>> working<br>
>> > set over a small period of time to be that high.<br>
>><br>
>> One difference from the paper and the script I am using to create the<br>
>> call graph is that the script I have records every call the exact number<br>
>> of times. The script is attached.<br>
>><br>
>> With sampling, a call foo->long_running_bar would be recorded multiple<br>
>> times and show up as multiple calls.<br>
>><br>
>> The first seems better, but I wonder if sampling somehow produces a<br>
>> better result.<br>
>><br>
>> With instrumentation (which I assume is what you used in the game), you<br>
>> also get an exact callgraph, no?<br>
>><br>
><br>
> You get an exact callgraph minus indirect calls as those currently aren't<br>
> captured.<br>
><br>
> - Michael Spencer<br>
><br>
><br>
>><br>
>> ><br>
>> > I am however surprised by the 10% increase in iTLB misses.<br>
>><br>
>><br>
>><br>
>><br>
>> Cheers,<br>
>> Rafael<br>
>><br>
>><br>
<br></blockquote></div><br></div></div>