<div dir="ltr"><br><div class="gmail_extra"><br><div class="gmail_quote">On Mon, Feb 27, 2017 at 12:01 PM, Hal Finkel <span dir="ltr"><<a href="mailto:hfinkel@anl.gov" target="_blank">hfinkel@anl.gov</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">
  
    
  
  <div bgcolor="#FFFFFF"><div><div class="gmail-h5">
    <p><br>
    </p>
    <div class="gmail-m_-8252287413177567603moz-cite-prefix">On 02/27/2017 01:47 PM, Daniel Berlin
      wrote:<br>
    </div>
    <blockquote type="cite">
      
      <div dir="ltr"><br>
        <div class="gmail_extra"><br>
          <div class="gmail_quote">On Mon, Feb 27, 2017 at 11:29 AM,
            Adam Nemet <span dir="ltr"><<a href="mailto:anemet@apple.com" target="_blank">anemet@apple.com</a>></span>
            wrote:<br>
            <blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">
              <div style="word-wrap:break-word"><br>
                <div>
                  <div>
                    <div class="gmail-m_-8252287413177567603h5">
                      <blockquote type="cite">
                        <div>On Feb 27, 2017, at 10:11 AM, Hal Finkel
                          <<a href="mailto:hfinkel@anl.gov" target="_blank">hfinkel@anl.gov</a>>
                          wrote:</div>
                        <br class="gmail-m_-8252287413177567603m_-8094851141848933186Apple-interchange-newline">
                        <div>
                          <div bgcolor="#FFFFFF">
                            <p><br>
                            </p>
                            <div class="gmail-m_-8252287413177567603m_-8094851141848933186moz-cite-prefix">On
                              02/27/2017 11:47 AM, Adam Nemet wrote:<br>
                            </div>
                            <blockquote type="cite"> <br>
                              <div>
                                <blockquote type="cite">
                                  <div>On Feb 27, 2017, at 9:39 AM,
                                    Daniel Berlin <<a href="mailto:dberlin@dberlin.org" target="_blank">dberlin@dberlin.org</a>>
                                    wrote:</div>
                                  <br class="gmail-m_-8252287413177567603m_-8094851141848933186Apple-interchange-newline">
                                  <div>
                                    <div dir="ltr"><br>
                                      <div class="gmail_extra"><br>
                                        <div class="gmail_quote">On Mon,
                                          Feb 27, 2017 at 9:29 AM, Adam
                                          Nemet <span dir="ltr"><<a href="mailto:anemet@apple.com" target="_blank">anemet@apple.com</a>></span>
                                          wrote:<br>
                                          <blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">
                                            <div style="word-wrap:break-word"><br>
                                              <div><span class="gmail-m_-8252287413177567603m_-8094851141848933186gmail-">
                                                  <blockquote type="cite">
                                                    <div>On Feb 27,
                                                      2017, at 7:27 AM,
                                                      Hal Finkel <<a href="mailto:hfinkel@anl.gov" target="_blank">hfinkel@anl.gov</a>>
                                                      wrote:</div>
                                                    <br class="gmail-m_-8252287413177567603m_-8094851141848933186gmail-m_8239107560126363242Apple-interchange-newline">
                                                    <div>
                                                      <div class="gmail-m_-8252287413177567603m_-8094851141848933186gmail-m_8239107560126363242moz-cite-prefix" style="font-family:helvetica;font-size:10px;font-style:normal;font-variant-caps:normal;font-weight:normal;letter-spacing:normal;text-align:start;text-indent:0px;text-transform:none;white-space:normal;word-spacing:0px;background-color:rgb(255,255,255)"><br class="gmail-m_-8252287413177567603m_-8094851141848933186gmail-m_8239107560126363242Apple-interchange-newline">
                                                        On 02/27/2017
                                                        06:29 AM, Nema,
                                                        Ashutosh wrote:<br>
                                                      </div>
                                                      <blockquote type="cite" style="font-family:helvetica;font-size:10px;font-style:normal;font-variant-caps:normal;font-weight:normal;letter-spacing:normal;text-align:start;text-indent:0px;text-transform:none;white-space:normal;word-spacing:0px;background-color:rgb(255,255,255)">
                                                        <div class="gmail-m_-8252287413177567603m_-8094851141848933186gmail-m_8239107560126363242WordSection1">
                                                          <div>Thanks
                                                          for looking
                                                          into this.</div>
                                                          <div> </div>
                                                          <div>1) Issues
                                                          with re
                                                          running
                                                          vectorizer:</div>
                                                          <div>Vectorizer
                                                          might generate
                                                          redundant
                                                          alias checks
                                                          while
                                                          vectorizing
                                                          epilog loop.</div>
                                                          <div>Redundant
                                                          alias checks
                                                          are expensive,
                                                          we like to
                                                          reuse the
                                                          results of
                                                          already
                                                          computed alias
                                                          checks.</div>
                                                          <div>With
                                                          metadata we
                                                          can limit the
                                                          width of
                                                          epilog loop,
                                                          but not sure
                                                          about reusing
                                                          alias check
                                                          result.</div>
                                                          <div>Any
                                                          thoughts on
                                                          rerunning
                                                          vectorizer
                                                          with reusing
                                                          the alias
                                                          check result ?</div>
                                                        </div>
                                                      </blockquote>
                                                      <br style="font-family:helvetica;font-size:10px;font-style:normal;font-variant-caps:normal;font-weight:normal;letter-spacing:normal;text-align:start;text-indent:0px;text-transform:none;white-space:normal;word-spacing:0px;background-color:rgb(255,255,255)">
                                                      <span style="font-family:helvetica;font-size:10px;font-style:normal;font-variant-caps:normal;font-weight:normal;letter-spacing:normal;text-align:start;text-indent:0px;text-transform:none;white-space:normal;word-spacing:0px;background-color:rgb(255,255,255);float:none;display:inline">One
                                                        way of looking
                                                        at this is:
                                                        Reusing the
                                                        alias-check
                                                        result is really
                                                        just a
                                                        conditional
                                                        propagation
                                                        problem; if we
                                                        don't already
                                                        have an
                                                        optimization
                                                        that can combine
                                                        these after the
                                                        fact, then we
                                                        should.</span><br style="font-family:helvetica;font-size:10px;font-style:normal;font-variant-caps:normal;font-weight:normal;letter-spacing:normal;text-align:start;text-indent:0px;text-transform:none;white-space:normal;word-spacing:0px;background-color:rgb(255,255,255)">
                                                    </div>
                                                  </blockquote>
                                                  <div><br>
                                                  </div>
                                                </span>
                                                <div>+Danny</div>
                                                <div><br>
                                                </div>
                                                <div>Isn’t Extended SSA
                                                  supposed to help with
                                                  this?</div>
                                              </div>
                                            </div>
                                          </blockquote>
                                          <div><br>
                                          </div>
                                          <div>Yes, it will solve this
                                            with no issue already.  GVN
                                            probably does already too.</div>
                                          <div><br>
                                          </div>
                                          <div>even if if you have</div>
                                          <div><br>
                                          </div>
                                          <div>if (a == b)</div>
                                          <div>
                                            <div>if (a == c)</div>
                                          </div>
                                          <div>
                                            <div> if (a == d)</div>
                                          </div>
                                          <div>
                                            <div> if (a == e)</div>
                                          </div>
                                          <div>
                                            <div> if (a == g)</div>
                                          </div>
                                          <div><br>
                                          </div>
                                          <div><br>
                                          </div>
                                          <div>and  we can prove a ... g
                                            equivalent, newgvn will
                                            eliminate them all and set
                                            all the branches true.</div>
                                          <div><br>
                                          </div>
                                          <div>If you need a simpler
                                            clean up pass, we could run
                                            it on sub-graphs.</div>
                                        </div>
                                      </div>
                                    </div>
                                  </div>
                                </blockquote>
                                <div><br>
                                </div>
                                <div>Yes we probably don’t want to run a
                                  full GVN after the “loop-scheduling”
                                  passes.</div>
                              </div>
                            </blockquote>
                            <br>
                            FWIW, we could, just without the
                            memory-dependence analysis enabled (i.e. set
                            the NoLoads constructor parameter to true).
                            GVN is pretty fast in that mode.<br>
                          </div>
                        </div>
                      </blockquote>
                      <div><br>
                      </div>
                    </div>
                  </div>
                  <div>OK.  Another data point is that I’ve seen cases
                    in the past where the alias checks required for the
                    loop passes could enable GVN to remove redundant
                    loads/stores.  Currently we can only pick these up
                    with LTO when GVN is rerun.</div>
                </div>
              </div>
            </blockquote>
            <div><br>
            </div>
            <div>This is just GVN brokenness, newgvn should not have
              this problem.</div>
            <div>If it does, i'd love to see it.</div>
          </div>
        </div>
      </div>
    </blockquote>
    <br></div></div>
    I thought that the problem is that we just don't run GVN after that
    point in the pipeline.<span class="gmail-HOEnZb"><font color="#888888"><br>
    <br></font></span></div></blockquote><div>Then i misunderstand.</div><div><br></div><div>"Another data point is that I’ve seen cases in the past where the alias checks required for the loop passes could enable GVN to remove redundant loads/stores."</div><div><br></div><div>Why would GVN suddenly be able to remove redundant loads and stores due to insertion of alias checks?</div><div><br></div><div>It should have been able to remove them regardless.</div><div><br></div></div></div></div>