<div dir="ltr"><div dir="ltr">Am Fr., 7. Feb. 2020 um 13:06 Uhr schrieb Chris Lattner <<a href="mailto:clattner@nondot.org">clattner@nondot.org</a>>:<br></div><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 style="overflow-wrap: break-word;"><div><br></div><div><blockquote type="cite"><div><blockquote type="cite"><div>Furthermore, my experience is that specialized IRs never get the investment in (e.g.). testing, location propagation for debugging optimized code, textual round tripping, pass management, and the multitude of other things that are required for a world class compiler implementation.</div></blockquote><p>I agree, but I think that these are well-known requirements within this ecosystem. What this means also depends on context (it might be more helpful to think of this more like SCEV than like a custom IR).</p></div></blockquote></div><div>SCEV is a great example of the problem that I’m talking about here. SCEV cannot be easily tested, because the IR doesn’t round trip to a file-check-able textual representation, and cannot have passes applied to it. This is barely acceptable for SCEV, because there are workarounds for the simple cases, but would be extremely problematic for something as critical as the IR for a loop optimizer as a whole. SCEV also doesn’t propagate location information, and has several other problems.</div></div></blockquote><div><br></div><div>A textual representation is not the highest in my priority list, but certainly doable. I'd wait until the data structure settles to something sufficiently stable. For instance, does the textual representation encompass the entire version history or just one snapshot (i.e. just the red tree?). Until then, I think it is sufficient when there is a predictable path from IR to red/green tree.</div><div><br></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 style="overflow-wrap: break-word;"><div></div><div>The IR for a loop framework is going to require other duplications and have other problems, e.g. you’ll want a pass manager etc.</div></div></blockquote><div><br></div><div></div><div>I am not sure that a pass manager is that necessary. It adds flexibility to the optimization process, but as long we don't support dynamically adding transformations, I think organizing the transformations in code is sufficient at the beginning. Since transformations apply on subtrees, are prioritized, and can add candidates to be selected by a cost function, a pass manager has to be structured differently.</div><div><br></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 style="overflow-wrap: break-word;"><div><blockquote type="cite"><div><div><p>Michael and I have discussed this offilne, and I think that more
quantitative information here would be helpful. One phrasing might
be: For a reasonable test suite of programs, for the loops we
might optimize, how many different loop-nest variants do we need
to represent for costing purposes (also note that, depending on
profiling information and/or heuristics, we may need to cost at
multiple trip-count values)? Given that, what is the relative cost
of creating deep copies of the loop nests and transforming them
vs. the cost of using a red/green tree? Does that cost seem
significant?</p><p>We don't want to invest significant effort into a infrastructure
that we know ahead of time won't scale sufficiently.<br>
</p></div></div></blockquote><div>Again, as I mentioned above, I completely agree with your goals. That said, I have a few concerns about this:</div><div><br></div><div>1) My understanding is that this approach is more of a research project that needs to be proven, than an implemented design the entire LLVM community should back as “the plan". If you’d like to pursue this, then I think the right way to go is to do some significant implementation work to build it out, see how it works, get experience with it, then propose inclusion as “the” loop optimizer when it can be measured empirically. This is effectively what Polly did.</div></div></div></blockquote><div><br></div><div>100% agree</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 style="overflow-wrap: break-word;"><div><div><br></div><div>2) On MLIR, in addition to being a nice technical solution to the IR data structures and infrastructure backing it, a large slice of the MLIR community has exactly the same goals as you propose. They are literally building loop transformation technology driving state of the art forward, and they also have phase ordering issues :-). These folks have lots of advanced thoughts on this, and a lot of expertise in vectorization, memory hierarchy optimizations, accelerators, HPC, and low-level code generation etc. Leveraging this work makes sense to me, and duplicating it seems wasteful.</div></div></div></blockquote><div><br></div><div>Agreed, although I think there is some difference the level we operate on. My focus was on optimizing code coming from relatively low-level C/C++ ensuring that we do not loose opportunities because the high-level representation does not exactly match the low-level IR. This does not mean we cannot learn from each other's experience.</div><div><br></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 style="overflow-wrap: break-word;"><div><div>3) On the technical point design of red/green trees, my personal opinion is that this isn’t solving the important part of the problem. Red/green trees (as far <a href="https://llvm.org/devmtg/2018-10/slides/Kruse-LoopTransforms.pdf" target="_blank">as I’m aware</a>) are a memory saving optimization. </div></div></div></blockquote><div><br></div><div>Not entirely, there would also be a run-time cost if we had to copy the complete IR before for creating a new "red" representation. If this was necessary, I think it would be a barrier to speculatively apply transformations.</div><div><br></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 style="overflow-wrap: break-word;"><div><div> However, “solving" the phase ordering issues require exploring an exponential (e.g. in depth of transformations / pass pipeline) search space problem (which is expensive in cycles, not just memory), and it relies on having a very good cost model for the generated code (which is a universally hard problem). Furthermore, on the cost model point, the performance of the code depends on the later “non loop optimizations” as well, which are not going to get changed to use this new IR.</div></div></div></blockquote><div><br></div><div>Very good point. I don't claim a red/green tree is the universal solution to this problem, but I think it makes problem space exploration significantly easier, especially when combining multiple transformations before knowing the outcome.</div><div><br></div><div>As you point out, "non-loop optimizations" will be necessary after loop optimizations as well (same with LLVM's current loop optimizations: for instance, undo LCSAA) such as new InstCombine opportunities after loop fusion. This can be difficult to predict by a cost function, but they are usually less significant than what can be gained by a loop transformations. Others, such as LICM, can be significant though and should be considered by a cost function (or simply applied to the red/green tree before evaluating the cost function to be automatically be considered).</div><div><br></div><div>This is why I am enthusiastic about speculative transformations -- a generic cost function can handle many cases without requiring code for each transformation.</div><div> </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 style="overflow-wrap: break-word;"><div><div><br></div><div>4) While red/green trees are interesting, there are other approaches to tackle these problems. However, they are also researchy and need to be proven. I’m happy to explain my thoughts on this if you are interested.</div><div><br></div><div><br></div><div>Overall, I am super enthusiastic about new research directions, but I think they should be done along the lines of Polly, where the “new thing” gets built and can then be quantitatively evaluated based on its merits in practice, rather than being an up-front decision based on ideas. I am also a bit concerned that this effort is highly duplicative of work already happening in the MLIR community, but duplication isn’t necessarily bad when the path forward isn’t perfectly clear. In either case, I’d love to see more cross pollination between the efforts!</div></div></div></blockquote><div><br></div><div>One of the motivations of writing the RFC is to clarify goals and requirements. I would like to avoid the situation we had with Polly to discover 'blockers' when mainstreaming.</div><div><br></div><div>Another motivation is to find more interested parties. </div><div><br></div><div><br></div><div>Michael</div><div><br></div></div></div>