<div dir="ltr"><div dir="ltr"><br></div><br><div class="gmail_quote"><div dir="ltr" class="gmail_attr">On Sat, Feb 29, 2020 at 4:00 PM Nicholas Krause <<a href="mailto:xerofoify@gmail.com">xerofoify@gmail.com</a>> wrote:<br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">
  
    
  
  <div>
    <br>
    <br>
    <div>On 2/29/20 6:17 PM, River Riddle via
      llvm-dev wrote:<br>
    </div>
    <blockquote type="cite">
      
      <div dir="ltr">
        <div dir="ltr"><br>
        </div>
        <br>
        <div class="gmail_quote">
          <div dir="ltr" class="gmail_attr">On Sat, Feb 29, 2020 at 2:25
            PM David Blaikie via llvm-dev <<a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a>>
            wrote:<br>
          </div>
          <blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">
            <div dir="ltr">
              <div dir="ltr"><br>
              </div>
              <br>
              <div class="gmail_quote">
                <div dir="ltr" class="gmail_attr">On Sat, Feb 29, 2020
                  at 2:19 PM Chris Lattner <<a href="mailto:clattner@nondot.org" target="_blank">clattner@nondot.org</a>>
                  wrote:<br>
                </div>
                <blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">
                  <div>On Feb 29, 2020, at 2:08 PM, David Blaikie <<a href="mailto:dblaikie@gmail.com" target="_blank">dblaikie@gmail.com</a>>
                    wrote:
                    <div>
                      <blockquote type="cite">
                        <div>
                          <blockquote class="gmail_quote" style="font-family:Helvetica;font-size:12px;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;text-decoration:none;margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">I've<span> </span><br>
                            curious as<br>
                            to how MLIR deals with IPO as that's the
                            problem I was running into.<span> </span><br>
                          </blockquote>
                          <div style="font-family:Helvetica;font-size:12px;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;text-decoration:none"><br>
                            FWIW I believe LLVM's new pass manager (NPM)
                            was designed with parallelism and the
                            ability to support this situation (that MLIR
                            doesn't? Or doesn't to the degree/way in
                            which the NPM does). I'll leave it to folks
                            (Chandler probably has the most context
                            here) to provide some more detail there if
                            they can/have time.<br>
                          </div>
                        </div>
                      </blockquote>
                    </div>
                    <br>
                    <div>Historically speaking, all of the LLVM pass
                      managers have been designed to support
                      multithreaded compilation (check out the ancient
                      history of the <a href="http://llvm.org/docs/WritingAnLLVMPass.html" target="_blank">WritingAnLLVMPass</a> doc
                      if curious).</div>
                  </div>
                </blockquote>
                <div><br>
                </div>
                <div>I think the specific thing that might'v been a bit
                  different in the NPM was to do with analysis
                  invalidation in a way that's more parallelism friendly
                  than the previous one - but I may be
                  misrepresenting/misundrstanding some of it.</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">
                  <div>
                    <div>The problem is that LLVM has global use-def
                      chains on constants, functions and globals, etc,
                      so it is impractical to do this.  Every
                      “inst->setOperand” would have to be able to
                      take locks or use something like software
                      transactional memory techniques in their
                      implementation.  This would be very complicated
                      and very slow.<br>
                    </div>
                  </div>
                </blockquote>
                <div><br>
                  Oh, yeah - I recall that particular limitation being
                  discussed/not addressed as yet.<br>
                   </div>
                <blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">
                  <div>
                    <div>MLIR defines this away from the beginning. 
                      This is a result of the core IR design, not the
                      pass manager design itself.<br>
                    </div>
                  </div>
                </blockquote>
                <div><br>
                  What does MLIR do differently here/how does it define
                  that issue away? (doesn't have use-lists built-in?)<br>
                </div>
              </div>
            </div>
          </blockquote>
          <div><br>
          </div>
          <div>The major thing is that constants and global-like objects
            don't produce SSA values and thus don't have use-lists. <a href="https://mlir.llvm.org/docs/Rationale/#multithreading-the-compiler" target="_blank">https://mlir.llvm.org/docs/Rationale/#multithreading-the-compiler</a> discusses
            this a bit. </div>
          <div><br>
          </div>
          <div>For constants, the data is stored as an Attribute(context
            uniqued metadata, have no use-list, not SSA). This attribute
            can either placed in the attribute list(if the operand is
            always constant, like for the value of a switch case),
            otherwise it must be explicitly materialized via some
            operation. For example, the `<a href="https://mlir.llvm.org/docs/Dialects/Standard/#constant-operation" target="_blank">std.constant</a>` operation will
            materialize an SSA value from some attribute data.</div>
          <div><br>
          </div>
          <div>For references to functions and other global-like
            objects, we have a non-SSA mechanism built around `symbols`.
            This is essentially using a special attribute to reference
            the function by-name, instead of by ssa value. You can find
            more information on <a href="https://mlir.llvm.org/docs/SymbolsAndSymbolTables/" target="_blank">MLIR symbols here</a>. </div>
          <div><br>
          </div>
          <div>Along with the above, there is a trait that can be
            attached to operations called `<a href="https://mlir.llvm.org/docs/Traits/#isolatedfromabove" target="_blank">IsolatedFromAbove</a>`. This
            essentially means that no SSA values defined above a region
            can be referenced from within that region. The pass manager
            only allows schedule passes on operations that have this
            property, meaning that all pipelines are implicitly
            multi-threaded.</div>
          <div><br>
          </div>
          <div>The pass manager in MLIR was heavily inspired by the work
            on the new pass manager in LLVM, but with specific
            constraints/requirements that are unique to the design of
            MLIR. That being said, there are some usability features
            added that would also make great additions to LLVM: instance
            specific pass options and statistics, pipeline crash
            reproducer generation, etc.</div>
          <div><br>
          </div>
          <div>Not sure if any of the above helps clarify, but happy to
            chat more if you are interested.</div>
          <div><br>
          </div>
          <div>-- River</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">
            <div dir="ltr">
              <div class="gmail_quote">
                <div>- Dave<br>
                </div>
              </div>
            </div>
          </blockquote>
        </div>
      </div>
    </blockquote>
    River,<br>
    The big thing from my reading of the Pass Manager in MLIR is that it
    allows us to iterate through<br>
    a pass per function or module as a group allowing it to run in
    async. I've proposed this <br>
    on the GCC side:<br>
    <a href="https://gcc.gnu.org/ml/gcc/2020-02/msg00247.html" target="_blank">https://gcc.gnu.org/ml/gcc/2020-02/msg00247.html</a><br>
    <br>
    Its to walk through the IPA passes which are similar to analyze
    passes on the LLVM side.<br></div></blockquote><div><br></div><div>Hi Nicholas,</div><div><br></div><div>I can't say anything about the GCC side, but this isn't a particularly novel aspect of the MLIR pass manager. In many ways, the pass manager is the easiest/simplest part of the multi-threading problem. The bigger problem is making sure that the rest of the compiler infrastructure is structured in a way that is thread-safe, or can be made thread-safe. This is why most of the discussion is based around how to model things like constants, global values, etc. When I made MLIR multi-threaded a year ago, a large majority of my time was spent outside of the pass manager. For a real example, I spent much more time just on <a href="https://mlir.llvm.org/docs/WritingAPass/#multi-threaded-pass-timing">multi-threaded pass timing</a> than making the pass manager itself multi-threaded.</div><div><br></div><div>-- River</div><div><br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><div>
    <br>
    Nick <br>
    <br>
    <blockquote type="cite">
      <div dir="ltr">
        <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">
            <div dir="ltr">
              <div class="gmail_quote">
                <div> </div>
                <blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">
                  <div>
                    <div><br>
                    </div>
                    <div>-Chris</div>
                  </div>
                </blockquote>
              </div>
            </div>
            _______________________________________________<br>
            LLVM Developers mailing list<br>
            <a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a><br>
            <a href="https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev" rel="noreferrer" target="_blank">https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev</a><br>
          </blockquote>
        </div>
      </div>
      <br>
      <fieldset></fieldset>
      <pre>_______________________________________________
LLVM Developers mailing list
<a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a>
<a href="https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev" target="_blank">https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev</a>
</pre>
    </blockquote>
    <br>
  </div>

</blockquote></div></div>