<div dir="ltr">Hi Kevin,<div><br></div><div>[re-cc'ing list]</div><div><br></div><div>I see - this makes sense. Thanks for the explanation.</div><div><br></div><div>Cheers,</div><div><br></div><div>James</div></div><div class="gmail_extra">
<br><br><div class="gmail_quote">On 2 September 2014 13:09, Kevin Qin <span dir="ltr"><<a href="mailto:kevinqindev@gmail.com" target="_blank">kevinqindev@gmail.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
<div dir="ltr">Hi James,<div><br></div><div>Yes, For the original solution, switching to use a lookup/jump table may bring better performance, but doesn't help to recude the code size. And the performance benefit should not be significant, because most benefit from runtime unrolling are coming from the unrolled loop body, not from the prolog. Andthe unrolled prolog will cost extra10% code size comparing to the rolled one , which is not a good deal from my eyes.</div>
<div><br></div><div><br></div><div>Regards,</div><div>Kevin</div></div><div class="gmail_extra"><br><br><div class="gmail_quote">2014-09-02 12:48 GMT+01:00 James Molloy <span dir="ltr"><<a href="mailto:james@jamesmolloy.co.uk" target="_blank">james@jamesmolloy.co.uk</a>></span>:<div>
<div class="h5"><br>
<blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div dir="ltr">Hi Kevin,<div><br></div><div>The "obvious" (to me at least) prologue would be to use something similar to Duff's Device:</div>
<div><div><br></div><div><font face="courier new, monospace"><span style="font-size:13px">extraiters = tripcount % loopfactor</span><br>
</font></div></div><div><span style="font-size:13px"><font face="courier new, monospace">switch (extraiters) {</font></span></div><div><span style="font-size:13px"><font face="courier new, monospace">case 0: jump loop:</font></span></div>
<div><span style="font-size:13px"><font face="courier new, monospace">case 1: jump L1</font></span></div><div><span style="font-size:13px"><font face="courier new, monospace">case 2: jump L2</font></span></div><div><span style="font-size:13px"><font face="courier new, monospace">case 3: jump L3</font></span></div>
<div><span style="font-size:13px"><font face="courier new, monospace">}</font></span></div><div><span style="font-size:13px"><font face="courier new, monospace"><br></font></span></div><div><span style="font-size:13px"><font face="courier new, monospace">Loop:</font></span></div>
<div><span style="font-size:13px"><font face="courier new, monospace"> tripcount --;</font></span></div><div><font face="courier new, monospace"> LoopBody</font></div><div><font face="courier new, monospace">L1:</font></div>
<div><font face="courier new, monospace"> tripcount --;</font></div><div><font face="courier new, monospace"> LoopBody</font></div><div><font face="courier new, monospace">L2:</font></div><div><font face="courier new, monospace"> tripcount --;</font></div>
<div><font face="courier new, monospace"> LoopBody</font></div><div><font face="courier new, monospace">L3:</font></div><div><font face="courier new, monospace"> tripcount --;</font></div><div><font face="courier new, monospace"> LoopBody</font></div>
<div><font face="courier new, monospace"> </font></div><div><font face="courier new, monospace"> if (tripcount >= 0) jump Loop else jump Out</font></div><div><font face="courier new, monospace"><br></font></div><div>
<font face="courier new, monospace">Out:</font></div><div><font face="arial, sans-serif"><br></font></div><div><font face="arial, sans-serif">The switch would be changed into a lookup/jump table. Wouldn't this produce better code too?</font></div>
<div><font face="arial, sans-serif"><br></font></div><div><font face="arial, sans-serif">Cheers,</font></div><div><font face="arial, sans-serif"><br></font></div><div><font face="arial, sans-serif">James</font></div></div>
<div class="gmail_extra"><br><br><div class="gmail_quote"><div><div>On 2 September 2014 10:56, Kevin Qin <span dir="ltr"><<a href="mailto:kevinqindev@gmail.com" target="_blank">kevinqindev@gmail.com</a>></span> wrote:<br>
</div></div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div><div>
Runtime unrolling will create a prologue to execute the extra iterations which is can't divided by the unroll factor. It generates an if-then-else sequence to jump into a factor -1 times unrolled loop body, like<br>
<br>
extraiters = tripcount % loopfactor<br>
if (extraiters == 0) jump Loop:<br>
if (extraiters == loopfactor) jump L1<br>
if (extraiters == loopfactor-1) jump L2<br>
...<br>
L1: LoopBody;<br>
L2: LoopBody;<br>
...<br>
if tripcount < loopfactor jump End<br>
Loop:<br>
...<br>
End:<br>
<br>
It means if the unroll factor is 4, the loop body will be 7 times unrolled, 3 are in loop prologue, and 4 are in the loop.<br>
This patch is to use a loop to execute the extra iterations in prologue, like<br>
<br>
extraiters = tripcount % loopfactor<br>
if (extraiters == 0) jump Loop:<br>
else jump Prol<br>
Prol: LoopBody;<br>
extraiters -= 1 // Omitted if unroll factor is 2.<br>
if (extraiters != 0) jump Prol: // Omitted if unroll factor is 2.<br>
if (tripcount < loopfactor) jump End<br>
Loop:<br>
...<br>
End:<br>
<br>
Then when unroll factor is 4, the loop body will be copied by only 5 times, 1 in the prologue loop, 4 in the original loop. And if the unroll factor is 2, new loop won't be created, just as the original solution.<br>
<br>
On AArch64 target, if runtime unrolling enabled, after applying this patch, the code size will drop by 10%.<br>
<br>
Also, the sequence of if-then-else sequence is saved, which could bring very slightly performance benefit, which is less than 0.1% on X86 and AArch64 target.<br>
<br>
So overall, this patch can bring a lot of code size improvement, and have no harm to performance.<br>
<br>
Is it OK to commit?<br>
<br>
Thanks,<br>
Kevin<br>
<br>
<a href="http://reviews.llvm.org/D5147" target="_blank">http://reviews.llvm.org/D5147</a><br>
<br>
Files:<br>
lib/Transforms/Utils/LoopUnrollRuntime.cpp<br>
test/Transforms/LoopUnroll/PowerPC/a2-unrolling.ll<br>
test/Transforms/LoopUnroll/runtime-loop.ll<br>
test/Transforms/LoopUnroll/runtime-loop1.ll<br>
test/Transforms/LoopUnroll/runtime-loop2.ll<br>
<br></div></div>_______________________________________________<br>
llvm-commits mailing list<br>
<a href="mailto:llvm-commits@cs.uiuc.edu" target="_blank">llvm-commits@cs.uiuc.edu</a><br>
<a href="http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits" target="_blank">http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits</a><br>
<br></blockquote></div><br></div>
</blockquote></div></div></div><span class="HOEnZb"><font color="#888888"><br><br clear="all"><div><br></div>-- <br><div dir="ltr">Best Regards,<div><br></div><div>Kevin Qin</div></div>
</font></span></div>
</blockquote></div><br></div>