<div dir="ltr"><br><div class="gmail_extra"><br><div class="gmail_quote">On Mon, Jul 25, 2016 at 9:27 AM, 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-width:1px;border-left-style:solid;border-left-color:rgb(204,204,204);padding-left:1ex"><div><div style="font-family:arial,helvetica,sans-serif;font-size:10pt;color:rgb(0,0,0)"><br><hr><blockquote style="border-left-width:2px;border-left-style:solid;border-left-color:rgb(16,16,255);margin-left:5px;padding-left:5px;color:rgb(0,0,0);font-weight:normal;font-style:normal;text-decoration:none;font-family:helvetica,arial,sans-serif;font-size:12pt"><span class="gmail-"><b>From: </b>"Sean Silva" <<a href="mailto:chisophugis@gmail.com" target="_blank">chisophugis@gmail.com</a>><br><b>To: </b>"Chandler Carruth" <<a href="mailto:chandlerc@gmail.com" target="_blank">chandlerc@gmail.com</a>><br><b>Cc: </b>"Xinliang David Li" <<a href="mailto:davidxl@google.com" target="_blank">davidxl@google.com</a>>, "llvm-dev" <<a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a>>, "Davide Italiano" <<a href="mailto:dccitaliano@gmail.com" target="_blank">dccitaliano@gmail.com</a>>, "Tim Amini Golling" <<a href="mailto:mehdi.amini@apple.com" target="_blank">mehdi.amini@apple.com</a>>, "Hal Finkel" <<a href="mailto:hfinkel@anl.gov" target="_blank">hfinkel@anl.gov</a>>, "Sanjoy Das" <<a href="mailto:sanjoy@playingwithpointers.com" target="_blank">sanjoy@playingwithpointers.com</a>>, "Pete Cooper" <<a href="mailto:peter_cooper@apple.com" target="_blank">peter_cooper@apple.com</a>><br></span><b>Sent: </b>Friday, July 22, 2016 3:55:52 AM<span class="gmail-"><br><b>Subject: </b>Re: [PM] I think that the new PM needs to learn about inter-analysis dependencies...<br><br></span><span class="gmail-"><div dir="ltr">The more closely I look at this, the more it seems like there may be a useful incremental step in the transition to the new PM: use the new PM analysis machinery in the old PM. If this is possible, it will simplify the old PM and (hopefully) allow an incremental transition to the new PM instead of a flag day transition for the switch.<div><br></div><div>I.e., AFAICT, the new PM transition is essentially about 2 mostly orthogonal aspects of running optimization pipelines:</div><div>1. Analysis computation and analysis result lifetime management (including avoiding analysis recomputation)</div><div>2. Running transformation passes over their respective IRUnit's in some order<br><div><br></div><div>These are conflated in the old PM. In reality, the only interaction between them (with the new PM machinery for 1.) is a small number of places within 2. which need to call 'invalidate'.</div><div><br></div><div>I'm pretty sure that 2. is fairly similar in the new PM and old PM (the main difference is that the notion of "adapters" is split out in the new PM). The analysis handling seems to be what makes the old PM so difficult to understand (e.g. it is the cause of the multiple inheritance in the implementation). Trying to unify analyses and transformations (and some questionable (in hindsight) implementation decisions) seems to be the main "problem" with the design of the old PM AFAICT (there are other issues, but they are more "nice to have").</div><div><br></div><div>IMO it is an anti-pattern to think of analyses as "passes". There are just "analyses" and "transformations" and they are two separate things. In fact, the `run` method on analyses should probably be called `computeResult` or something like that to avoid confusion. </div></div></div></span></blockquote>This makes sense to me.<br><br>We do currently have some "in between" passes, like LCSSA, which are transformations, but are required by other passes, and transform the IR but whose preservation represents properties of the IR. The particulars of how we handle LCSSA aside (e.g. I think we should preserve it more, perhaps everywhere), how are we planning on handling this class of things?<br></div></div></blockquote><div><br></div><div>The new PM doesn't currently have a concept like this. As you mentioned, it is a weird cross between a transformation and an analysis: it can be "invalidated" like an analysis, but "recomputing" it actually mutates the IR like a transformation.</div><div><br></div><div>I'd like to preface the below with the following:</div><div>No matter how we ultimately address this requirement, my preference is that we do so in a way that applies to the old PM. This is a case where the old PM supports a richer set of functionality than the new PM. By incrementally refactoring the old PM away from its use of this extra capability and towards whatever "new" way there is to do it, we will understand better what it is that we actually need.<br></div><div><br></div><div>(and sorry for the brain dump in the rest of this post)</div><div><br></div><div><div><br></div></div><div><br></div><div>I have not seen any mention of a solution to this problem besides "we shouldn't do that", which is sort of a cop-out. Here is a strawman proposal:</div><div><br></div><div><div>If it isn't too expensive, one simple alternative is to have passes just make a call to a utility function to put things in LCSSA if they need it (this utility would also have to invalidate analyses).</div><div>If that ends up being expensive, we can have a dummy "indicator" analysis IRIsInLCSSAForm which, if cached, means "don't bother to call the utility function". We could maybe just use the LCSSA pass directly to do the transformation. LCSSA could have IRIsInLCSSAForm as an member typedef `IndicatorT` so it can be accessed generically. We could then support an API like:</div></div><div><br></div><div>```</div><div>FooTransformation.cpp:</div><div><br></div><div>PreservedAnalyses FooTransformation::run(Function &F, AnalysisManager AM) {</div><div> // Must be called before getting analyses, as it might invalidate some.<br></div><div> canonicalizeIR<LCSSA>(F, AM);</div><div><br></div><div> ...<br>}</div><div><br></div><div><br></div><div>include/IR/Canonicalization.h:</div><div><br></div><div>template <typename CanonicalizationT, typename IRUnitT></div><div>void canonicalizeIR(IRUnitT &IR, AnalysisManager &AM) {</div><div> using IndicatorT = typename CanonicalizationT::IndicatorAnalysis;<br></div><div> if (AM.getCachedResult<IndicatorT>(IR))</div><div> return;</div><div> CanonicalizationT C;</div><div> PreservedAnalysis PA = C.run(IR, AM);</div><div> AM.invalidate(IR, PA);</div><div> (void)AM.getResult<IndicatorT>(IR);<br>}</div><div><br></div><div>```</div><div><br></div><div><br></div><div>One place that talks about this problem of "requiring a transformation" is <a href="http://llvm.org/devmtg/2014-04/PDFs/Talks/Passes.pdf">http://llvm.org/devmtg/2014-04/PDFs/Talks/Passes.pdf</a> on slide 17.</div><div><br></div><div>One reason it provides for "we shouldn't do that" is that if you think about these things as "canonicalize the IR into a specific form", then when you have N>1 such dependencies (like some passes do on LoopSimplify and LCSSA), one must have a subset of the requirements of the other. I.e. you can't have two canonicalizations that "fight" each other. Using an explicit mutation API like the strawman above is a bit less bulletproof than scheduling based on statically known interferences between canonicalizations (e.g. CanonicalizationA may invalidate CanonicalizationB, but not the reverse, so it would automatically know to run CanonicalizationA before CanonicalizationB), but given that we have relatively few "canonicalizations" (to give them a name) that use this feature of the old PM, it may be livable (at least in the middle-end, it seems like there is just LCSSA, LoopSimplify, BreakCriticalEdges, and LowerSwitch in calls to addPreservedID/addRequiredID).</div><div><br></div><div>I don't find the "Causes rampant re-running of invalidated analyses" argument in that slide convincing. If a pass needs the IR in LCSSA then it needs it. There isn't much we can do about that.</div><div><br></div><div><br></div><div><br></div><div><br></div><div>One invariant I'd like to preserve in the new pass manager is that whatever pipeline is provided on the opt command line, we end up running something "valid"; so a cop-out like "if a pass needs LCSSA, you need to make sure to add LCSSA at an appropriate place before it in the pipeline" is not something I think is reasonable (way too error-prone).</div><div><br></div><div>Small rant:</div><div><br></div><div>We already are in this error-prone situation in the new PM with the need to call `getCachedResult` to access analyses from a larger IRUnitT (e.g. the situation I explained in the post-commit thread of r274712); this means that when constructing a pass pipeline to pass to opt, you need to do various things to ensure your pipeline even makes sense:</div><div>1. You need to make sure to always add the `requires<foo>` at a higher level for any pass that needs it (how do you find out what each pass needs besides grepping the source?), otherwise you get an assertion failure / report_fatal_error</div><div>2. You need to make sure that no intervening transformation at the lower level invalidates the outer analysis before getting to your pass that needs it. E.g. `module(require<foo>,function(pass-that-invalidates-foo,pass-that-requires-foo)`.</div><div><br></div><div>I don't see the point of this restriction. In a hypothetical future where we run over inner IRUnit's in parallel we would need some awareness to avoid races when getting analyses for outer IRUnit's. But doing that serialization dynamically as the analysis is accessed (and perhaps even centralized in the analysis manager and pass management layers) seems better than just always computing it at a higher level and crossing your fingers that you got your pass pipeline right and the outer analysis doesn't get invalidated before it is needed.</div><div><br></div><div>-- Sean Silva</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 style="font-family:arial,helvetica,sans-serif;font-size:10pt;color:rgb(0,0,0)"><br>Thanks again,<br>Hal<div><div class="gmail-h5"><br><blockquote style="border-left-width:2px;border-left-style:solid;border-left-color:rgb(16,16,255);margin-left:5px;padding-left:5px;color:rgb(0,0,0);font-weight:normal;font-style:normal;text-decoration:none;font-family:helvetica,arial,sans-serif;font-size:12pt"><div dir="ltr"><div><div>And the `run` method on transformations could just as easily be called `performTransformation`.</div><div><br></div><div><br></div><div>I remember asking and getting and answer from Chandler (<a href="http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20150907/299083.html" target="_blank">http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20150907/299083.html</a>) about how to coexist the old and new PM compatibly so individual passes would be able to work for both and we wouldn't need to have a flag day. I wasn't able to find the discussions that Chandler cited, but I suspect that the answer is that we didn't have enough analyses ported at that point to consider sharing the analysis management between the old and new PM.</div><div><br></div><div><br></div><div>If this works out it may be the evolutionary path we have all been wanting for this pass manager transition. Fingers crossed. Hopefully I'm not overlooking some major issue.</div><div><br></div><div>Anyway... back to working on the analysis manager dependency tracking.</div><div><div><br></div><div>-- Sean Silva</div></div></div></div><div class="gmail_extra"><br><div class="gmail_quote">On Thu, Jul 21, 2016 at 1:59 AM, Sean Silva <span dir="ltr"><<a href="mailto:chisophugis@gmail.com" target="_blank">chisophugis@gmail.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0pt 0pt 0pt 0.8ex;border-left-width:1px;border-left-style:solid;border-left-color:rgb(204,204,204);padding-left:1ex"><div dir="ltr">We did some basic sanity checking that memory usage didn't go out of control (it doesn't; at least with with a simple preserves-all/preserves-none invalidation scheme and the current LTO pipeline). Also, I did some basic sanity checking for compile time. The simple preserves-all/preserves-none invalidation scheme seems marginally slower, but no real conclusions (besides a simple sanity check) can be drawn without the real analysis preservation semantics in place.<div><br></div><div>I'll start working on fixing the analysis managers. There seem to basically be two parts (although they may need to be done simultaneously to make sure all the pieces fit together):</div><div>- unify all the analysis managers into a single analysis manager for all IRUnitT's (requires type-erasing the IRUnit)</div><div>- introduce the dependency tracking machinery</div><div><br></div><div>I think I gave a reasonable outline in the two posts above:</div><div>- the one starting with "<span style="font-size:12.8px">To clarify, it seems like the current new PM is essentially trying to solve the problem of maintaining/updating a mapping"</span></div><div><span style="font-size:12.8px">- the one starting with "</span><span style="font-size:12.8px">Yeah, the mechanics of maintaining this fully general mapping are straightforward in the abstract"</span><br></div><div><span style="font-size:12.8px"><br></span></div><div><span style="font-size:12.8px">I'm happy to do a centralized writeup if anybody wants. Just let me know.</span></div><div><span style="font-size:12.8px"><br></span></div><div><span style="font-size:12.8px">As far as changes to the code, t</span><span style="font-size:12.8px">he updates to the new PM passes should hopefully be mechanical (despite there being many of them). However, from what I can tell, fixing this problem will require touching most lines of the analysis manager machinery so the diff will probably be a bit scary, even though I think we can keep the same basic structure (i.e. a per-IRUnit std::list holding one analysis result (at a stable address) per element, combined with a DenseMap from (Analysis, IRUnit) to an element of the per-IRUnit storage list (see AnalysisResultListMapT and AnalysisResultMapT in include/llvm/IR/PassManager.h)).</span></div><div><span style="font-size:12.8px">The main changes to the analysis manager will be:</span></div><div><span style="font-size:12.8px">- type-erasing the IRUnit</span></div><div><span style="font-size:12.8px">- the elements of the </span><span style="font-size:12.8px">AnalysisResultListMapT will need to keep track of any dependents</span></div><div><span style="font-size:12.8px">- the analysis manager will need to update those dependencies as it is re-entered by analyses getting results of other analyses</span></div><div><span style="font-size:12.8px">- the analysis manager will need to walk the dependencies to do transitive invalidation</span></div><div><br></div><div>I think the most robust approach is for analysis dependencies to be implicitly constructed by the analysis manager via tracking entry/exit from get{,Cached}Result.</div><div>One alternative is for analyses to explicitly pass in their ID to getResult to indicate the dependency, but that seems error-prone (and also verbose). The issue is that we will need a getResult API that does not track dependencies for use by transformation passes (since there is no dependency to track in that case); so an innocuous copy-paste from a transform pass to an analysis can result in a failure to track dependencies and risk of use-after-free (we could fight this with the type system, but I think that would get a bit verbose (I'm willing to try it though if people would prefer))</div><div>One restriction of the implicit tracking approach is that it requires all calls into the analysis manager to occur in the `run` method of the analysis (so that the dependencies are implicitly tracked via re-entrance to get{,Cached}Result); is this a reasonable restriction?</div><div><br></div><div><br></div><div>One annoying problem is that I think that the dependency links will need to be bidirectional. To use the example analysis cache from my other post:</div><span><div><div style="font-size:12.8px">(AssumptionAnalysis, function @bar) -> (AssumptionCache for @bar, [(SomeModuleAnalysis, module TheModule)])</div><div style="font-size:12.8px">(AssumptionAnalysis, function @baz) -> (AssumptionCache for @baz, [(SomeModuleAnalysis, module TheModule)])</div><div style="font-size:12.8px">(SomeModuleAnalysis, module TheModule) -> (SomeModuleAnalysisResult for TheModule, [(SomeFunctionAnalysis, function @baz)])</div><div style="font-size:12.8px">(SomeFunctionAnalysis, function @baz) -> (SomeFunctionAnalysisResult for @baz, [])</div></div><div style="font-size:12.8px"><br></div></span><div style="font-size:12.8px">if we delete function @baz, then the dependent list <span style="font-size:12.8px"> </span><span style="font-size:12.8px">[(SomeFunctionAnalysis, function @baz)] for `</span><span style="font-size:12.8px">(SomeModuleAnalysis, module TheModule)` will now have a stale pointer to function @baz. I think that in practice we can check to see if `</span><span style="font-size:12.8px">(SomeFunctionAnalysis, function @baz)` is in our hash table and if it isn't then just ignore the dependency as "this dependent analysis result has already been freed". In the worst case (memory allocator reuses the memory for another function) we may spuriously free an analysis result for a different function. However it is still unsettling (and may actually be UB in C++?).</span></div><div style="font-size:12.8px"><span style="font-size:12.8px">Ideally we would track bidirectional links; that way when we remove an analysis result we also have it remove itself from dependence lists of all of its dependencies.</span></div><span><font color="#888888"><div><br></div><div>-- Sean Silva</div></font></span></div><div><div><div class="gmail_extra"><br><div class="gmail_quote">On Fri, Jul 15, 2016 at 8:40 PM, Sean Silva <span dir="ltr"><<a href="mailto:chisophugis@gmail.com" target="_blank">chisophugis@gmail.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0pt 0pt 0pt 0.8ex;border-left-width:1px;border-left-style:solid;border-left-color:rgb(204,204,204);padding-left:1ex"><div dir="ltr"><br><div class="gmail_extra"><br><div class="gmail_quote"><span>On Fri, Jul 15, 2016 at 8:39 PM, Sean Silva <span dir="ltr"><<a href="mailto:chisophugis@gmail.com" target="_blank">chisophugis@gmail.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0pt 0pt 0pt 0.8ex;border-left-width:1px;border-left-style:solid;border-left-color:rgb(204,204,204);padding-left:1ex"><div dir="ltr">It looks like there is really no sane fix within the current infrastructure. I've had to essentially trigger invalidation (except in the PreservedAnalyses::all() case) in the function pass manager and function to loop adapters.</div></blockquote><div><br></div></span><div>invalidation of *everything* I mean.</div><span><font color="#888888"><div><br></div><div>-- Sean Silva</div></font></span><div><div><div> </div><blockquote class="gmail_quote" style="margin:0pt 0pt 0pt 0.8ex;border-left-width:1px;border-left-style:solid;border-left-color:rgb(204,204,204);padding-left:1ex"><div dir="ltr"><div><br></div><div>So we basically need to get the analysis manager dependency tracking fixed.</div><div><br></div><div>Davide and I will get measurements on the resident set impact of all this caching (even with the overconservative invalidation for now) to see the impact. If there is a big rss impact then we probably want to consider that problem at the same time as the rewrite of the analysis manager.</div><span><font color="#888888"><div><br></div><div>-- Sean Silva</div></font></span></div><div><div><div class="gmail_extra"><br><div class="gmail_quote">On Thu, Jul 14, 2016 at 12:51 AM, Sean Silva <span dir="ltr"><<a href="mailto:chisophugis@gmail.com" target="_blank">chisophugis@gmail.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0pt 0pt 0pt 0.8ex;border-left-width:1px;border-left-style:solid;border-left-color:rgb(204,204,204);padding-left:1ex"><div dir="ltr"><br><div class="gmail_extra"><br><div class="gmail_quote"><div><div>On Wed, Jul 13, 2016 at 1:48 AM, Sean Silva <span dir="ltr"><<a href="mailto:chisophugis@gmail.com" target="_blank">chisophugis@gmail.com</a>></span> wrote:<br><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"><br><div class="gmail_extra"><br><div class="gmail_quote"><div><div>On Wed, Jul 13, 2016 at 12:34 AM, Chandler Carruth <span dir="ltr"><<a href="mailto:chandlerc@gmail.com" target="_blank">chandlerc@gmail.com</a>></span> wrote:<br><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><div><div dir="ltr">On Wed, Jul 13, 2016 at 12:25 AM Sean Silva <<a href="mailto:chisophugis@gmail.com" target="_blank">chisophugis@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 dir="ltr"><div class="gmail_extra"><div class="gmail_quote">On Tue, Jul 12, 2016 at 11:39 PM, Chandler Carruth <span dir="ltr"><<a href="mailto:chandlerc@gmail.com" target="_blank">chandlerc@gmail.com</a>></span> wrote:<br><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"><span><div dir="ltr">On Tue, Jul 12, 2016 at 11:34 PM Sean Silva <<a href="mailto:chisophugis@gmail.com" target="_blank">chisophugis@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 dir="ltr"><div class="gmail_extra"><div class="gmail_quote">On Tue, Jul 12, 2016 at 11:32 PM, Xinliang David Li <span dir="ltr"><<a href="mailto:davidxl@google.com" target="_blank">davidxl@google.com</a>></span> wrote:<br><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"><br><div class="gmail_extra"><br><div class="gmail_quote"><span>On Tue, Jul 12, 2016 at 10:57 PM, Chandler Carruth <span dir="ltr"><<a href="mailto:chandlerc@gmail.com" target="_blank">chandlerc@gmail.com</a>></span> wrote:<br><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>Yea, this is a nasty problem.</div><div><br></div><div>One important thing to understand is that this is specific to analyses which hold references to other analyses. While this isn't unheard of, it isn't as common as it could be. Still, definitely something we need to address.</div></div></div></blockquote><div><br></div></span><div>We can call this type of dependencies (holding references) hard-dependency. The soft dependency refers to the case where analysis 'A' depends on 'B' during computation, but does not need 'B' once it is computed.</div><div><br></div><div>There are actually quite a few examples of hard-dependency case. For instance LoopAccessInfo, LazyValueInfo etc which hold references to other analyses.</div><div><br></div><div>Problem involving hard-dependency is actually easier to detect, as it is usually a compile time problem. Issues involving soft dependencies are more subtle and can lead to wrong code gen.</div></div></div></div></blockquote><div><br></div></div></div></div><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote"><div>Did you mean to say that soft-dependency problems are easier to detect? At least my intuition is that soft-dependency is easier because there is no risk of dangling pointers to other analyses.<br></div></div></div></div></blockquote><div><br></div></span><div>The issue is that the fact that there is *any* dependency isn't clear.</div><div><br></div><div>However, I think the only real problem here are these "hard dependencies" (I don't really like that term though). For others, only an analysis that is *explicitly* preserved survives. So I'm not worried about the fact that people have to remember this.</div><div><br></div><div>The question is how often there are cross-data-structure references. David mentions a few examples, and I'm sure there are more, but it isn't clear to me yet whether this is pervasive or occasional.</div></div></div></blockquote><div><br></div></div></div></div><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote"><div><div><br>I just did a quick run-through of PassRegistry.def and this is what I found:</div><div><br></div><div>Module analyses: 0/5 hold pointers to other analyses</div><div>CallGraph: No pointers to other analyses.</div><div>LazyCallGraph: No pointers to other analyses.</div><div>ProfileSummaryAnalysis: No pointers to other analyses.</div><div>TargetLibraryAnalysis: No pointers to other analyses.<br></div><div>VerifierAnalysis: No pointers to other analyses.</div><div><br></div><div>Module alias analyses: 1/1 keeps pointer to other analysis.</div><div>GlobalsAA: Result keeps pointer to TLI (this is a function analysis).</div><div><br></div><div>Function analyses: 9/17 keep pointers to other analysis</div><div>AAManager: Its Result holds TLI pointer and pointers to individual AA result objects.</div><div>AssumptionAnalysis: No pointers to other analyses.<br></div><div>BlockFrequencyAnalysis: Its Result holds pointers to LoopInfo and BPI.<br></div><div>BranchProbabilityAnalysis: Stores no pointers to other analyses. (uses LoopInfo to "recalculate" though)<br></div><div>DominatorTreeAnalysis: Stores no pointers to other analyses.<br></div><div>PostDominatorTreeAnalysis: Stores no pointers to other analyses.</div><div>DemandedBitsAnalysis: Stores pointers to AssumptionCache and DominatorTree<br></div><div>DominanceFrontierAnalysis: Stores no pointers to other analyses. (uses DominatorTreeAnalysis for "recalculate" though).<br></div><div>LoopInfo: Uses DominatorTreeAnalysis for "recalculate" but stores no pointers.<br></div><div>LazyValueAnalysis: Stores pointers to AssumptionCache, TargetLibraryInfo, DominatorTree.<br></div><div>DependenceAnalysis: Stores pointers to AliasAnalysis, ScalarEvolution, LoopInfo</div><div>MemoryDependenceAnalysis: Stores pointers to AliasAnalysis, AssumptionCache, TargetLibraryInfo, DominatorTree<br></div><div>MemorySSAAnalysis: Stores pointers to AliasAnalysis, DominatorTree<br></div><div>RegionInfoAnalysis: Stores pointers to DomTree, PostDomTree, DomFrontier<br></div><div>ScalarEvolutionAnalysis: Stores pointers to TargetLibraryInfo, AssumptionCache, DominatorTree, LoopInfo</div><div>TargetLibraryAnalysis: Has no dependencies<br></div><div>TargetIRAnalysis: Has no dependencies.<br></div><div><br></div><div>Function alias analyses: 3/5 keep pointers to other analyses</div><div>BasicAA: Keeps pointers to TargetLibraryInfo, AssumptionCache, DominatorTree, LoopInfo</div><div>CFLAA: Keeps pointer to TargetLibraryInfo</div><div>SCEVAA: Keeps pointer to ScalarEvolution</div><div>ScopedNoAliasAA: No dependencies<br></div><div>TypeBasedAA: No dependencies<br></div><div><br></div><div><br></div><div>Total: 13/28 analyses (~50%) hold pointers to other analyses.</div><div>Of the 15/28 analyses that don't hold pointers, 12/15 simply have no dependencies. Only 3/15 (BPI, LoopInfo, DominanceFrontier) have dependencies that are used just for a "recalculate" step that retains no pointers.</div><div>So I think it is fair to say that analyses which hold pointers to other analyses is not an exceptional case. In fact, analyses that use other analyses just for a "recalculate" step seems to be the exceptional case (only 3/28 or about 10%)</div></div></div></div></div></blockquote><div><br></div></div></div><div>Interesting!</div><div><br></div><div>Most of these look like they hold a pointer to the root analysis as opposed to detailed objects *inside* the analysis?</div><div><br></div><div>It might make sense to try to handle this very specific pattern in a special way of overriding the invalidate routines is too error prone.... <span style="line-height:1.5">We could try to make this work "automatically" but I'm worried this would be challenging to get right. Open to suggestions of course.</span></div><div><br></div><div>Any other ideas about what would make sense to handle this?</div><div><br></div><div>Does it make sense to override the invalidate routines now and iterate from there? I feel like you've done a lot of the research necessary for this already...</div></div></div></blockquote><div><br></div></div></div><div>I'll keep pushing forward tomorrow with building test-suite successfully using the new PM for the LTO pipeline (I was doing some unrelated LLD stuff for most of today). It will be interesting to see how many `invalidate` overrides will be needed to avoid these issues for just the LTO pipeline on test-suite.</div></div></div></div></blockquote><div><br></div></div></div><div>I spent the better part of today working on this and will continue tomorrow; this problem seems nastier than I thought. For some reason the LTO pipeline (or something about LTO) seems to hit on these issues much more (I'm talking like 40k lines of ASan error reports from building test-suite with the LTO pipeline in the new PM; per-TU steps still using the old PM). Some notes:</div><div><br></div><div>- BasicAA's dependence on domtree and loopinfo in the new PM seems to account for quite a few of the problems.</div><div>- BasicAA and other stuff are marked (by overriding `invalidate` to return false) to never be invalidated because they are "stateless". However they still hold pointers and so they do need to be invalidated.</div><div>- CallGraph uses AssertingVH (PR28400) and so I needed a workaround similar to r274656 in various passes.</div><div>- <span style="font-size:12.8px">D21921 is holding up -- I haven't hit any issues with the core logic of that patch.</span></div><div>- AAResults holds handles to various AA result objects. This means it pretty much always needs to be invalidated unless you are sure that none of the AA's will get invalidated.</div><div><br></div><div><br></div><div>The existing `invalidate` method doesn't have the right semantics for even an error-prone solution :( We are going to need to make some significant changes to even get basic sanity I think. Perhaps each analysis can expose a "preserve" static function. E.g. instead of `PA.preserve<Foo>();` you have to do `Foo::setPreserved(PA);`.</div><div>I'm actually not quite sure that that will even work. Once I have test-suite fully building successfully with the LTO pipeline in the new PM I'll be able to give a more confident answer (esp. w.r.t. the manager for different IRUnitT's).</div><div>But at this point I'm not confident running *any* pass pipeline in the new PM without at least assertions+ASan.<br></div><div><br></div><div>We may want to have a proper design discussion around this problem though.</div><div><br></div><div>Also I'd like to have test-suite working (by hook or by crook) with LTO in the new PM so we can get some numbers on the resident set impact of all this caching; if it is really problematic then we may need to start talking front-and-center about different invalidation policies for keeping this in check instead of leaving it as something that we will be able to patch later.<br></div><div><br></div><div><br></div><div><br></div><div>The more I think about it, the more I'm convinced that the real "hard" problem that the new PM is exposing us to is having the ability for any pass to ask for any analysis on any IRUnitT (and any specific IRUnit of that IRUnitT) and have the result stored somewhere and then invalidated. This means that "getAnalysisUsage" is not just a list of passes, but much more complicated and is essentially a set of arbitrary pairs "(analysis, IRUnit)" (and the associated potential tangle of dependencies between the state cached on these tuples). With the old PM, you essentially are looking at a problem of scheduling the lifetime of analyses of the same IRUnit intermingled with transformation passes on that same IRUnit, so you only have the "analysis" part of the tuple above, making things much simpler (and handling dependencies is much simpler too). We've obviously outgrown this model with examples like LAA, AssumptionCacheTracker, etc. that hack around this in the old PM. We may want to have a fresh re-examination of what problems we are exactly trying to solve.</div><div><br></div><div>For me, my main concern now is what changes need to be made in order to feel confident running a pipeline in the new PM without assertions+ASan.</div><div><br></div><div><br></div><div>Sorry for the long post, just brain-dumping before heading home.</div><div><br></div><div>-- Sean Silva</div><div><br></div><div><br></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_extra"><div class="gmail_quote"><span><font color="#888888"><div><br></div><div>-- Sean Silva</div></font></span></div><br></div></div>
</blockquote></div><br></div></div>
</blockquote></div><br></div>
</div></div></blockquote></div></div></div><br></div></div>
</blockquote></div><br></div>
</div></div></blockquote></div><br></div>
</blockquote><br><br><br></div></div><span class="gmail-">-- <br><div><span name="x"></span>Hal Finkel<br>Assistant Computational Scientist<br>Leadership Computing Facility<br>Argonne National Laboratory<span name="x"></span><br></div></span></div></div></blockquote></div><br></div></div>