<div dir="ltr"><br><div class="gmail_extra"><br><div class="gmail_quote">On Wed, Jan 18, 2017 at 8:05 PM, Friedman, Eli <span dir="ltr"><<a target="_blank" href="mailto:efriedma@codeaurora.org">efriedma@codeaurora.org</a>></span> wrote:<br><blockquote style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex" class="gmail_quote">
  
    
  
  <div bgcolor="#FFFFFF"><span class="gmail-">
    <div class="gmail-m_-1396804328854282964moz-cite-prefix">On 1/18/2017 2:56 AM, Thomas RUBIANO
      wrote:<br>
    </div>
    <blockquote type="cite">
      <div dir="ltr">Ty Eli for your answer.<br>
        <div>
          <div class="gmail_extra"><br>
            <div class="gmail_quote">On Tue, Jan 17, 2017 at 8:11 PM,
              Friedman, Eli <span dir="ltr"><<a target="_blank" href="mailto:efriedma@codeaurora.org">efriedma@codeaurora.org</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"><span class="gmail-m_-1396804328854282964gmail-">On 1/17/2017 7:12 AM, Thomas RUBIANO
                  via llvm-dev wrote:<br>
                  <blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">
                    Hi all!<br>
                    <br>
                    I'm new here, and would like to implement my own
                    Loop Invariant Detection adding some more
                    information on Quasi-Invariants.<br>
                    <br>
                    First, is there anything about Quasi-Invariants
                    detection in LLVM I would missed?<br>
                    <br>
                    I've seen LICM using LoopInfo::isLoopInvariant for
                    finding invariants.<br>
                    It seems that this method considers a Value
                    invariant if:<br>
                    - it's an Instruction not presents in the current
                    loop (what does it mean? There is no dependence
                    analysis on In and Out "variables" of all
                    instructions in the loop?)<br>
                  </blockquote>
                  <br>
                </span>
                isLoopInvariant just checks whether the definition of a
                value is an instruction inside the loop.<span class="gmail-m_-1396804328854282964gmail-"><br>
                </span></blockquote>
              <div><br>
              </div>
              <div>Ok, the term "definition" makes it clear. Do, if here
                it's:<br>
                <span class="gmail-m_-1396804328854282964gmail-">%1 = load i32, i32* %fact, align 4</span><br>
                <span class="gmail-m_-1396804328854282964gmail-">%mul = mul nsw i32 %1, %2<br>
                   
                  <br>
                </span><br>
                <span class="gmail-m_-1396804328854282964gmail-">load i32, i32* %fact, align 4 is
                  the def of %1 and it's inside the loop then it's not
                  invariant…<br>
                </span></div>
            </div>
          </div>
        </div>
      </div>
    </blockquote>
    <br></span>
    Exactly.  If LICM can prove the load is in fact invariant, it will
    move it out of the loop.<span class="gmail-"><br>
    <br>
    <blockquote type="cite">
      <div dir="ltr">
        <div>
          <div class="gmail_extra">
            <div class="gmail_quote">
              <div><span class="gmail-m_-1396804328854282964gmail-"><br>
                </span></div>
              <blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><span class="gmail-m_-1396804328854282964gmail-">
                  <br>
                  <blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">
                    - this Value is not an Instruction (then a Constant
                    I guess…).<br>
                  </blockquote>
                  <br>
                </span>
                Or an function argument, or a few other obscure things
                which don't really matter in this context.<span class="gmail-m_-1396804328854282964gmail-"><br>
                  <br>
                  <blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">
                    I've seen LoopAccessAnalysis using it too. What does
                    this analysis do exactly on loop invariant address?<br>
                  </blockquote>
                  <br>
                </span>
                ScalarEvolution::isLoopInvaria<wbr>nt works on SCEV
                expressions instead of Values, but it's essentially the
                same thing.<span class="gmail-m_-1396804328854282964gmail-"><br>
                </span></blockquote>
              <div><br>
              </div>
              <div>What can offer this SCEV expression more than Values?
                <br>
              </div>
            </div>
          </div>
        </div>
      </div>
    </blockquote>
    <br>
    </span><span class="gmail-m_-1396804328854282964gmail-">See the comment at the beginning of
      lib/Analysis/ScalarEvolution.<wbr>cpp for a brief description and some
      references.</span><span class="gmail-"><span class="gmail-m_-1396804328854282964gmail-"></span><br>
    <br>
    <blockquote type="cite">
      <div dir="ltr">
        <div>
          <div class="gmail_extra">
            <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"><span class="gmail-m_-1396804328854282964gmail-">
                  <br>
                  <blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">
                    <br>
                    Also DependenceAnalysis seems to give dependence
                    information on memory refs. But it seems to not be
                    used by LICM…<br>
                  </blockquote>
                  <br>
                </span>
                Yes, the current LICM uses alias analysis in a more
                direct manner (look for AliasSetTracker).<span class="gmail-m_-1396804328854282964gmail-"><br>
                  <br>
                  <blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">
                    <br>
                    Also MemoryDependenceAnalysis "determines, for a
                    given memory operation, what preceding memory
                    operations it depends on".<br>
                    <br>
                    My question is: Where is really done this dependence
                    analysis. The one which figures out which
                    Instructions depends on others?<br>
                    <br>
                    Simply if I have:<br>
                    %0 = load i32, i32* %coucou, align 4<br>
                    %1 = load i32, i32* %fact, align 4<br>
                    %2 = load i32, i32* %i, align 4<br>
                    %mul = mul nsw i32 %1, %2<br>
                    <br>
                    mul instruction will depends on the two precedents
                    loads because it uses their results %1 and %2 but
                    not the first one.<br>
                    <br>
                    I guess this is done somewhere, and there is a
                    special representation of this information but I
                    didn't find, I'm a bit lost ^^'<br>
                  </blockquote>
                  <br>
                </span>
                If you call "operands()" on the Instruction, it will
                return %1 and %2.  Please keep in mind that LLVM IR is
                SSA.<span class="gmail-m_-1396804328854282964gmail-HOEnZb"><font color="#888888"><br>
                  </font></span></blockquote>
              <div><br>
              </div>
              <div>Yes but if I understand well, the AliasSetTracker can
                tell me which load it corresponds to… <br>
                How can I use this AliasSetTracker to disambiguate the
                registers?<br>
              </div>
              <div>Here it's just having the correspondence %1 → %fact
                and %2 → %i <br>
              </div>
              <div><br>
              </div>
              <div>It would be just perfect for me to have the "real" In
                and Out of each Instruction.<br>
              </div>
              <div>Maybe I should work on another level or with another
                object representation?<br>
              </div>
              <div><br>
              </div>
              <div>Ty again :)<br>
              </div>
            </div>
          </div>
        </div>
      </div>
    </blockquote>
    <br></span>
    Can you show me the complete function you're looking at?  Have you
    run mem2reg on your IR?<span class="gmail-"><br></span></div></blockquote><div><br></div><div>I was looking the LoopInvariantCodeMotion::runOnLoop.<br><br></div><div>No I didn't run mem2reg on my IR…<br></div><div>Is it necessary?<br><br></div><div>I finally use the address of the operands and instruction to have a kind of ID of each %<num><br></div><div>It seems that it refers well what I want…<br><br></div><div>Let be the Instruction I:<br></div><div><span class="gmail-"><span class="gmail-m_-1396804328854282964gmail-">%0 = mul nsw i32 %1, %2<br><br></span></span></div><div><span class="gmail-"><span class="gmail-m_-1396804328854282964gmail-">&I = 0x3d9f850 → %0<br></span></span></div><div><span class="gmail-"><span class="gmail-m_-1396804328854282964gmail-">operands(0)→get() = 0x3d9f7bc → %1<br></span></span></div><div><span class="gmail-"><span class="gmail-m_-1396804328854282964gmail-">operands(1)→get() = 0x3d9f80c → %2<br><br></span></span></div><div><span class="gmail-"><span class="gmail-m_-1396804328854282964gmail-">then it should be ok if I take this for finding my dependencies between Instructions?<br></span></span></div><div> </div><blockquote style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex" class="gmail_quote"><div bgcolor="#FFFFFF"><span class="gmail-">
    <br>
    -Eli<br>
    <pre cols="72" class="gmail-m_-1396804328854282964moz-signature">-- 
Employee of Qualcomm Innovation Center, Inc.
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project</pre>
  </span></div>

</blockquote></div><br><br clear="all"><br>-- <br><div class="gmail_signature"> RUBIANO Thomas<br></div>
</div></div>