<div dir="ltr">I filed a couple more, in case they're actually different issues:<div>- <a href="http://llvm.org/bugs/show_bug.cgi?id=22412" target="_blank">http://llvm.org/bugs/show_bug.cgi?id=22412</a></div><div>- <a href="http://llvm.org/bugs/show_bug.cgi?id=22413" target="_blank">http://llvm.org/bugs/show_bug.cgi?id=22413</a></div><div><br></div><div>And that's pretty much it for internal changes. I'm fine with flipping the switch; Quentin, are you?</div><div>Also, just to have an idea, do you (or someone else!) plan to tackle these in the near future?<br></div><div><br></div><div class="gmail_extra"><div><div>-Ahmed</div></div>
<br><div class="gmail_quote">On Thu, Jan 29, 2015 at 11:50 AM, Ahmed Bougacha <span dir="ltr"><<a href="mailto:ahmed.bougacha@gmail.com" target="_blank">ahmed.bougacha@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"><div class="gmail_extra"><div><div><br></div></div><div class="gmail_quote"><div><div>On Wed, Jan 28, 2015 at 4:47 PM, Chandler Carruth <span dir="ltr"><<a href="mailto:chandlerc@gmail.com" target="_blank">chandlerc@gmail.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div dir="ltr"><div class="gmail_extra"><br><div class="gmail_quote"><span>On Wed, Jan 28, 2015 at 4:05 PM, Ahmed Bougacha <span dir="ltr"><<a href="mailto:ahmed.bougacha@gmail.com" target="_blank">ahmed.bougacha@gmail.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div dir="ltr">Hi Chandler,<br><br>I've been looking at the regressions Quentin mentioned, and filed a PR<br>for the most egregious one: <a href="http://llvm.org/bugs/show_bug.cgi?id=22377" target="_blank">http://llvm.org/bugs/show_bug.cgi?id=22377</a><br><br>As for the others, I'm working on reducing them, but for now, here are<br>some raw observations, in case any of it rings a bell:<br></div></blockquote><div><br></div></span><div>Very cool, and thanks for the analysis!</div><div><div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div dir="ltr"><br><br>Another problem I'm seeing is that in some cases we can't fold memory anymore:<br> vpermilps $-0x6d, -0xXX(%rdx), %xmm2 ## xmm2 = mem[3,0,1,2]<br> vblendps $0x1, %xmm2, %xmm0, %xmm0<br>becomes:<br> vmovaps -0xXX(%rdx), %xmm2<br> vshufps $0x3, %xmm0, %xmm2, %xmm3 ## xmm3 = xmm2[3,0],xmm0[0,0]<br> vshufps $-0x68, %xmm0, %xmm3, %xmm0 ## xmm0 = xmm3[0,2],xmm0[1,2]<br><br><br>Also, I see differences when some loads are shuffled, that I'm a bit<br>conflicted about:<br> vmovaps -0xXX(%rbp), %xmm3<br> ...<br> vinsertps $0xc0, %xmm4, %xmm3, %xmm5 ## xmm5 = xmm4[3],xmm3[1,2,3]<br>becomes:<br> vpermilps $-0x6d, -0xXX(%rbp), %xmm2 ## xmm2 = mem[3,0,1,2]<br> ...<br> vinsertps $0xc0, %xmm4, %xmm2, %xmm2 ## xmm2 = xmm4[3],xmm2[1,2,3]<br><br>Note that the second version does the shuffle in-place, in xmm2.<br><br><br>Some are blends (har har) of those two:<br> vpermilps $-0x6d, %xmm_mem_1, %xmm6 ## xmm6 = xmm_mem_1[3,0,1,2]<br> vpermilps $-0x6d, -0xXX(%rax), %xmm1 ## xmm1 = mem_2[3,0,1,2]<br> vblendps $0x1, %xmm1, %xmm6, %xmm0 ## xmm0 = xmm1[0],xmm6[1,2,3]<br>becomes:<br> vmovaps -0xXX(%rax), %xmm0 ## %xmm0 = mem_2[0,1,2,3]<br> vpermilps $-0x6d, %xmm0, %xmm1 ## xmm1 = xmm0[3,0,1,2]<br> vshufps $0x3, %xmm_mem_1, %xmm0, %xmm0 ## xmm0 = xmm0[3,0],xmm_mem_1[0,0]<br> vshufps $-0x68, %xmm_mem_1, %xmm0, %xmm0 ## xmm0 = xmm0[0,2],xmm_mem_1[1,2]<br><br><br>I also see a lot of somewhat neutral (focusing on Haswell for now)<br>domain changes such as (xmm5 and 0 are initially integers, and are<br>dead after the store):<br> vpshufd $-0x5c, %xmm0, %xmm0 ## xmm0 = xmm0[0,1,2,2]<br> vpalignr $0xc, %xmm0, %xmm5, %xmm0 ## xmm0 = xmm0[12,13,14,15],xmm5[0,1,2,3,4,5,6,7,8,9,10,11]<br> vmovdqu %xmm0, 0x20(%rax)<br>turning into:<br> vshufps $0x2, %xmm5, %xmm0, %xmm0 ## xmm0 = xmm0[2,0],xmm5[0,0]<br> vshufps $-0x68, %xmm5, %xmm0, %xmm0 ## xmm0 = xmm0[0,2],xmm5[1,2]<br> vmovups %xmm0, 0x20(%rax)</div></blockquote></div></div></div><br>All of these stem from what I think is the same core weakness of the current algorithm: we prefer the fully general shufps+shufps 4-way shuffle/blend far too often. Here is how I would more precisely classify the two things missing here:</div><div class="gmail_extra"><br></div><div class="gmail_extra">- Check if either inputs are "in place" and we can do a fast single-input shuffle with a fixed blend.</div></div></blockquote><div><br></div></div></div><div>I believe this would be <a href="http://llvm.org/bugs/show_bug.cgi?id=22390" target="_blank">http://llvm.org/bugs/show_bug.cgi?id=22390</a></div><span><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div dir="ltr"><div class="gmail_extra">- Check if we can form a rotation and use palignr to finish a shuffle/blend</div></div></blockquote><div><br></div></span><div>.. and this would be <a href="http://llvm.org/bugs/show_bug.cgi?id=22391" target="_blank">http://llvm.org/bugs/show_bug.cgi?id=22391</a></div><div><br></div><div>I think this about covers the Haswell regressions I'm seeing. Now for some pre-AVX fun!</div><span><font color="#888888"><div><br></div><div>-Ahmed</div></font></span><span><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div dir="ltr"><div class="gmail_extra">There may be other patterns we're missing, but these two seem to jump out based on your analysis, and may be fairly easy to tackle.</div></div>
</blockquote></span></div><br></div></div>
</blockquote></div><br></div></div>