<div dir="ltr"><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 5:14 PM Nicholas Krause via llvm-dev <<a href="mailto:llvm-dev@lists.llvm.org">llvm-dev@lists.llvm.org</a>> wrote:<br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-style:solid;border-left-color:rgb(204,204,204);padding-left:1ex">
  
    
  
  <div>
    <br>
    <br>
    <div>On 2/29/20 7:23 PM, River Riddle 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 4:00
            PM Nicholas Krause <<a href="mailto:xerofoify@gmail.com" target="_blank">xerofoify@gmail.com</a>> wrote:<br>
          </div>
          <blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-style:solid;border-left-color: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-width:1px;border-left-style:solid;border-left-color: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-width:1px;border-left-style:solid;border-left-color: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-width:1px;border-left-style:solid;border-left-color: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-width:1px;border-left-style:solid;border-left-color: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-width:1px;border-left-style:solid;border-left-color: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-width:1px;border-left-style:solid;border-left-color: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" target="_blank">multi-threaded pass timing</a> than
            making the pass manager itself multi-threaded.</div>
          <div><br>
          </div>
          <div>-- River</div>
        </div>
      </div>
    </blockquote>
    Actually in my experience, the biggest problem is if we can detect
    IPO and run async guarantees on that. MLIR runs operations but only
    for a module or set of functions<br>
    without this. One of my dreams would be to run passes in parallel
    including IPO detection and stop if it cannot continue pass a IPO
    pass or set of passes due to changes.<br>
    <br>
    Maybe MLIR does do that but its the one bottleneck that is really
    hard to fix,<br></div></blockquote><div><br></div><div>What MLIR does (that would require quite some work in LLVM) is making sure that you can process and transform functions in isolation, allowing to run *local* optimizations in parallel. This does not solve the IPO problem you're after. As I understand it, this is a difficult thing to design, and it requires consideration about how you think the passes and the pass-pipeline entirely.</div><div></div><div><br></div><div>Running function-passes and "local" optimizations in parallel in LLVM isn't possible because the structures in the LLVMContext aren't thread-safe, and because the IR itself isn't thread-safe. Something like just DCE or CSE a function call requires to modify the callee (through its use-list).</div><div><br></div><div>-- </div><div>Mehdi</div></div></div></div>