On Fri, May 11, 2018 at 7:20 PM Hal Finkel <hfinkel at anl.gov> wrote:> > On 05/11/2018 08:40 PM, Daniel Berlin via llvm-dev wrote: > > > > On Fri, May 11, 2018 at 2:37 PM, Hiroshi Yamauchi <yamauchi at google.com> > wrote: > >> >> >> On Thu, May 10, 2018 at 12:49 PM Daniel Berlin <dberlin at dberlin.org> >> wrote: >> >>> >>> >>> On Thu, May 10, 2018 at 12:05 PM, Hiroshi Yamauchi <yamauchi at google.com> >>> wrote: >>> >>>> >>>> >>>> On Wed, May 9, 2018 at 8:24 PM Daniel Berlin <dberlin at dberlin.org> >>>> wrote: >>>> >>>>> >>>>> >>>>> On Wed, May 9, 2018 at 10:39 AM, Hiroshi Yamauchi <yamauchi at google.com >>>>> > wrote: >>>>> >>>>>> >>>>>> >>>>>> On Tue, May 8, 2018 at 11:15 AM Daniel Berlin <dberlin at dberlin.org> >>>>>> wrote: >>>>>> >>>>>>> >>>>>>> >>>>>>> On Tue, May 8, 2018 at 10:38 AM, Hiroshi Yamauchi via llvm-dev < >>>>>>> llvm-dev at lists.llvm.org> wrote: >>>>>>> >>>>>>>> ( >>>>>>>> I came across this issue in the context of >>>>>>>> D46336 <https://reviews.llvm.org/D46336>. >>>>>>>> >>>>>>>> Thanks, Sanjay, for starting this discussion.) >>>>>>>> >>>>>>>> If >>>>>>>> we will >>>>>>>> move >>>>>>>> reassociation, >>>>>>>> or keep additional ones >>>>>>>> , >>>>>>>> out of instcombine, >>>>>>>> open questions for me would be >>>>>>>> : >>>>>>>> >>>>>>>> >>>>>>>> 1. Since -reassociate isn't a fixed point pass, >>>>>>>> >>>>>>> >>>>>>> This is fixable, fwiw, without fixpointing it. >>>>>>> >>>>>> >>>>>> How? >>>>>> >>>>> >>>>> Depends on specifically which part you would like to know about ;) >>>>> >>>> >>>> Maybe I misunderstood what you meant by "This is fixable". Did you >>>> mean that we won't somehow need to fixpoint between instcombine and >>>> reassociate, or that the specific motivating examples from the above >>>> differentials are foldable without fixpointing? >>>> >>> >>> If by fixpointing you mean "fixpointing reassociate and instcombine", >>> then yes, that is fixable without fixpointing reassociate and instcombine, >>> but would require rewriting instcombine :) >>> >>> >>>> >>>> If the latter, that may be the case. The concern was that we may >>>> encounter examples that may need many more iterations, if not fixpointing. >>>> As long as it's feasible to fixpoint between instcombine and >>>> reassociate, it seems to work, but I guess that would probably need >>>> some pass management change. >>>> >>>> >>>>> >>>>>> >>>>>> >>>>>>> >>>>>>>> we might need to repeat "-instcombine -reassociate" multiple times >>>>>>>> to >>>>>>>> fold >>>>>>>> down to what we want (relating to my comment here >>>>>>>> <https://reviews.llvm.org/D46336#1087082>). I assumed this isn't >>>>>>>> not what we want to do >>>>>>>> ? My impression is we don't do a fixed-point with passes? >>>>>>>> >>>>>>> >>>>>>> Well, i mean there is no practical difference between passes that we >>>>>>> fixpoint externally and fixpoint internally. >>>>>>> >>>>>> >>>>>> I had the following in mind: Does the pass manager support >>>>>> fixpointing externally? Is there any performance difference? Are people >>>>>> okay with that in general? >>>>>> >>>>>> But if there is no practical difference, I don't see any problem >>>>>> with that :) >>>>>> >>>>>> >>>>>>> >>>>>>>> >>>>>>>> >>>>>>> 2. >>>>>>>> Since -reassociate needs to come up with one operand order (at >>>>>>>> least currently as the only reassociate pass), would there exist a >>>>>>>> single, unique operand order that would enable all >>>>>>>> reassociative/commutative foldings that we want? >>>>>>>> >>>>>>> >>>>>>> In what way? >>>>>>> Are you asking whether there is a single reassociation order that >>>>>>> makes all foldings occur in the same operation or something? >>>>>>> I don't feel like i understand what you are asking. >>>>>>> >>>>>> >>>>>> Does this rephrase help: with the motivating examples (like >>>>>> and-of-shifts or bit check patterns) from the above differentials in mind, >>>>>> can we come up with a single reassociation order that solves all >>>>>> those and all the others that may come up in the future? Would we need >>>>>> different reassociation orders to fold different patterns? >>>>>> >>>>> >>>>> It doesn't quite help. >>>>> When stated that generally, there can be no such ordering at all, >>>>> that's easy to prove. It is a statically undecidable problem. >>>>> >>>>> There is however, a different question and answer to a few related >>>>> problems that maybe you are really asking? >>>>> 1. Is there a way to determine and apply the a maximal or >>>>> nearly-maximal set of folds/graph transforms that could be applied to a >>>>> given set of code in a sane and principled way -> yes >>>>> >>>>> (see, e.g., http://www.cs.cornell.edu/~ross/publications/eqsat/) >>>>> >>>>> 2. Is there a way to determine all expressions in the program as it >>>>> exists that are equivalent or equivalent under constant time constant >>>>> folding/reassociation, in a reasonable time bound -> yes >>>>> >>>>> (not a single easy link, happy to talk about it) >>>>> >>>>> Your original question is basically equivalent to >>>>> Is there a way to determine all expressions in the program as it >>>>> exists that are equivalent or could be made equivalent through any type of >>>>> folding that one can think up? >>>>> The answer to that is "no", it's provable that this is not statically >>>>> decidable, so the time bound doesn't matter :) >>>>> >>>>> You have to limit the possible folding/evaluation you apply in various >>>>> ways to make this decidable, and then further limit it to make the time >>>>> bound reasonable. >>>>> >>>>> This all quickly devolves into herbrand equivalence and it's >>>>> variations. >>>>> >>>>> >>>>> >>>> Let me try one more time :) May we need multiple reassociate passes to >>>> fold different reassociative patterns? >>>> >>>> A longer version: If Sanjay wants a particular reassociative pattern to >>>> be folded (D45842), Omer wants another particular reassociative pattern >>>> to be folded (D41574), and I want yet another particular reassociative >>>> pattern to be folded (D46336), would we potentially need three >>>> different reassociate passes with each combined with instcombine, rather >>>> than just one that may be able to somehow handle those cases in one >>>> shot, (assuming we don't want to put those in instcombine)? >>>> >>>> And it sounds like the answer is yes? >>>> >>> >>> If you take the current instcombine as a base, then yes, that is correct. >>> >> >>> If you are willing to rearchitect instcombine, the answer is no, it's >>> possible to do this all in a single pass in a relatively sane way. >>> >> >> I assume by rearchitect, you mean a major rewrite as per this comment: "Is >> there a way to determine all expressions in the program as it exists that >> are equivalent or equivalent under constant time constant >> folding/reassociation, in a reasonable time bound -> yes". Any pointer >> or time to chat? >> > > I'm happy to do both. > > >> >> I think that an approach like >> D >> >> >> 4 >> >> >> 6 >> >> >> 3 >> >> >> 3 >> >> >> 6 >> >> >> / >> >> D >> >> >> 4 >> >> >> 6 >> >> >> 5 >> >> >> 9 >> >> >> 5 >> >> has a merit: it would adds a bit of complexity, but would not require: >> >> 1. a major rewrite of instcombine, >> 2. writing multiple (potentially many) reassociate passes and figuring >> out how to fixpoint them with instcombine, or >> 3. writing a self-contained folding pass for a specific pattern >> >> If you look at the diffs in the existing .ll files in >> D46336 >> , it helps fold some previously-unfolded reassociation patterns beyond >> the bit check patterns that it originally targeted. >> >> Sure, and it does so by adding another O(N) cost to evaluation in each > case. Instcombine doesn't even do lazy reevaluation through tracking > dependencies, so it'll do so a lot of times as well. > > To me, that's not a good tradeoff, especially given how slow instcombine > is *already*. The code it produces is "good enough" to stop for a while > and do something else and not suffer horribly in performance.[1] > > Let me ask a different question: > > At what point would anyone here be willing to stop adding things to > instcombine and start doing something else instead, instead of waiting for > someone else to do it? > As far as i can tell, the answer is: "never", which makes most of these > discussions just pointless rehashes as we slowly repeat the same disaster > that became gcc's instruction combiner :) > > If the answer is "something", great, i'll set a mail filter and ignore > these threads until that something happens :) > > Personally, in my experience people will never do more here unless pushed > somewhat, or the thing becomes such a complete disaster no one wants to > touch it. > > > I've said this before, but I think a major impediment to forward progress > here is coming up with an agreement on what the "something else" should be. > Some of us have talked for years about having some TableGen-driven > replacement, or maybe we want something with a syntax more like what is > used by the Alive tool, but regardless, in order to gain in efficiency I > suspect we need a model that is more restrictive than more-or-less > arbitrary C++ code, and so we should pick a model and figure out how things > might work. > > > [1] Last year i computed the "improvement in performance on applications" > due to instcombine for a bunch of google apps and open source apps that had > easy to use benchmarks (IE I isolated about two years of instcombine > changes and made them to a current compiler piece by piece while measuring > performance). > I also computed the compile time increase in single instcombine passes > over the same time period. > > On x86, but the numbers basically said we were basically gaining nearly > nothing for high cost. IE our drive for better looking output does not > appear to translate into any real gains that i can find. Either > improvements to other opts hid them, or they simply didn't matter on the > processors i tested on. > > Certainly, apps/workloads/architectures may vary here, and my goal is not > to claim it's all worthless. > My actual goal in all of this was to get a sense of whether my perspective > on instcombine was still "reasonable", not to do a true scientific > exploration :) > I didn't have time/energy/etc to run it elsewhere, and again, my goal was > not to give certainty/try to give exact percentages. > > > This also matches my experience, but I draw a somewhat different lesson. I > often tell application developers that *this* is why they must file > compiler bug reports. Waiting and assuming that someone else will hit the > same problem, and file the report, is a bad strategy. I think that this is > due to two things: > > 1. As far as things go, the tail of the distribution is often really > long, and probability that the particular thing hampering one piece of hot > code is the same thing hampering another piece of hot code is often small. > > 2. We tend to add special cases instead of adding more-general > algorithms. The more-general work is often hard because figuring out the > cost modeling is often highly non-trivial. Also, when it's finally done, > the chances that the old special cases are removed is also small (so we'll > still accumulate cruft without specific effort). > >I don't have better or larger study results or data, but for fixing some smaller but important enough performance degradations (often in microbenchmarks, but occasionally in larger settings) smaller compiler improvements (not necessarily in instcombine but around that level) do seem to matter at least in individual contexts. It's not clear how much those actually mattered in a grander scheme of things, though. It sounds like we don't want to add to instcombine due to added cost/complexity, rearchitecting would be hard, it's not clear if incremental changes in instcombine did much per cost, which would make it harder to justify rearchitecting it... What do people think of approaches like D41574 and D45842?> -Hal > > > --Dan > > > > > > > > _______________________________________________ > LLVM Developers mailing listllvm-dev at lists.llvm.orghttp://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev > > > -- > Hal Finkel > Lead, Compiler Technology and Programming Languages > Leadership Computing Facility > Argonne National Laboratory > >-------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20180514/fd3d6a32/attachment-0001.html>
I mentioned this earlier in the thread - I would like to see something like D41574 in the optimizer. It's optimizing code that no other pass does currently, and I don't see any other near-term proposal that gets us those optimizations. Omer, can you rebase that to trunk? I think a header has moved, so it doesn't build as-is. I'd like to know if it can catch the cases in D45842. If not, why not? If it can handle those easily, I'll abandon D45842. Also, I don't know if it's better to include that functionality as another iteration of the existing -reassociate or split it off as its own pass. But I think it should do the distributive simplifications that are currently in -instcombine (InstCombiner::SimplifyUsingDistributiveLaws). Using that instsimplify logic for analysis lets us decide if the reassociation is worthwhile in the 1st place, it removes the risk that some other pass would somehow mess up the pattern before instcombine could zap it, and it reduces the burden on instcombine to be the entire optimizer. :) On Mon, May 14, 2018 at 1:34 PM, Hiroshi Yamauchi via llvm-dev < llvm-dev at lists.llvm.org> wrote:> > > On Fri, May 11, 2018 at 7:20 PM Hal Finkel <hfinkel at anl.gov> wrote: > >> >> On 05/11/2018 08:40 PM, Daniel Berlin via llvm-dev wrote: >> >> >> >> On Fri, May 11, 2018 at 2:37 PM, Hiroshi Yamauchi <yamauchi at google.com> >> wrote: >> >>> >>> >>> On Thu, May 10, 2018 at 12:49 PM Daniel Berlin <dberlin at dberlin.org> >>> wrote: >>> >>>> >>>> >>>> On Thu, May 10, 2018 at 12:05 PM, Hiroshi Yamauchi <yamauchi at google.com >>>> > wrote: >>>> >>>>> >>>>> >>>>> On Wed, May 9, 2018 at 8:24 PM Daniel Berlin <dberlin at dberlin.org> >>>>> wrote: >>>>> >>>>>> >>>>>> >>>>>> On Wed, May 9, 2018 at 10:39 AM, Hiroshi Yamauchi < >>>>>> yamauchi at google.com> wrote: >>>>>> >>>>>>> >>>>>>> >>>>>>> On Tue, May 8, 2018 at 11:15 AM Daniel Berlin <dberlin at dberlin.org> >>>>>>> wrote: >>>>>>> >>>>>>>> >>>>>>>> >>>>>>>> On Tue, May 8, 2018 at 10:38 AM, Hiroshi Yamauchi via llvm-dev < >>>>>>>> llvm-dev at lists.llvm.org> wrote: >>>>>>>> >>>>>>>>> ( >>>>>>>>> I came across this issue in the context of >>>>>>>>> D46336 <https://reviews.llvm.org/D46336>. >>>>>>>>> >>>>>>>>> Thanks, Sanjay, for starting this discussion.) >>>>>>>>> >>>>>>>>> If >>>>>>>>> we will >>>>>>>>> move >>>>>>>>> reassociation, >>>>>>>>> or keep additional ones >>>>>>>>> , >>>>>>>>> out of instcombine, >>>>>>>>> open questions for me would be >>>>>>>>> : >>>>>>>>> >>>>>>>>> >>>>>>>>> 1. Since -reassociate isn't a fixed point pass, >>>>>>>>> >>>>>>>> >>>>>>>> This is fixable, fwiw, without fixpointing it. >>>>>>>> >>>>>>> >>>>>>> How? >>>>>>> >>>>>> >>>>>> Depends on specifically which part you would like to know about ;) >>>>>> >>>>> >>>>> Maybe I misunderstood what you meant by "This is fixable". Did you >>>>> mean that we won't somehow need to fixpoint between instcombine and >>>>> reassociate, or that the specific motivating examples from the above >>>>> differentials are foldable without fixpointing? >>>>> >>>> >>>> If by fixpointing you mean "fixpointing reassociate and instcombine", >>>> then yes, that is fixable without fixpointing reassociate and instcombine, >>>> but would require rewriting instcombine :) >>>> >>>> >>>>> >>>>> If the latter, that may be the case. The concern was that we may >>>>> encounter examples that may need many more iterations, if not fixpointing. >>>>> As long as it's feasible to fixpoint between instcombine and >>>>> reassociate, it seems to work, but I guess that would probably need >>>>> some pass management change. >>>>> >>>>> >>>>>> >>>>>>> >>>>>>> >>>>>>>> >>>>>>>>> we might need to repeat "-instcombine -reassociate" multiple times >>>>>>>>> to >>>>>>>>> fold >>>>>>>>> down to what we want (relating to my comment here >>>>>>>>> <https://reviews.llvm.org/D46336#1087082>). I assumed this isn't >>>>>>>>> not what we want to do >>>>>>>>> ? My impression is we don't do a fixed-point with passes? >>>>>>>>> >>>>>>>> >>>>>>>> Well, i mean there is no practical difference between passes that >>>>>>>> we fixpoint externally and fixpoint internally. >>>>>>>> >>>>>>> >>>>>>> I had the following in mind: Does the pass manager support >>>>>>> fixpointing externally? Is there any performance difference? Are people >>>>>>> okay with that in general? >>>>>>> >>>>>>> But if there is no practical difference, I don't see any problem >>>>>>> with that :) >>>>>>> >>>>>>> >>>>>>>> >>>>>>>>> >>>>>>>>> >>>>>>>> 2. >>>>>>>>> Since -reassociate needs to come up with one operand order (at >>>>>>>>> least currently as the only reassociate pass), would there exist >>>>>>>>> a single, unique operand order that would enable all >>>>>>>>> reassociative/commutative foldings that we want? >>>>>>>>> >>>>>>>> >>>>>>>> In what way? >>>>>>>> Are you asking whether there is a single reassociation order that >>>>>>>> makes all foldings occur in the same operation or something? >>>>>>>> I don't feel like i understand what you are asking. >>>>>>>> >>>>>>> >>>>>>> Does this rephrase help: with the motivating examples (like >>>>>>> and-of-shifts or bit check patterns) from the above differentials in mind, >>>>>>> can we come up with a single reassociation order that solves all >>>>>>> those and all the others that may come up in the future? Would we need >>>>>>> different reassociation orders to fold different patterns? >>>>>>> >>>>>> >>>>>> It doesn't quite help. >>>>>> When stated that generally, there can be no such ordering at all, >>>>>> that's easy to prove. It is a statically undecidable problem. >>>>>> >>>>>> There is however, a different question and answer to a few related >>>>>> problems that maybe you are really asking? >>>>>> 1. Is there a way to determine and apply the a maximal or >>>>>> nearly-maximal set of folds/graph transforms that could be applied to a >>>>>> given set of code in a sane and principled way -> yes >>>>>> >>>>>> (see, e.g., http://www.cs.cornell.edu/~ross/publications/eqsat/) >>>>>> >>>>>> 2. Is there a way to determine all expressions in the program as it >>>>>> exists that are equivalent or equivalent under constant time constant >>>>>> folding/reassociation, in a reasonable time bound -> yes >>>>>> >>>>>> (not a single easy link, happy to talk about it) >>>>>> >>>>>> Your original question is basically equivalent to >>>>>> Is there a way to determine all expressions in the program as it >>>>>> exists that are equivalent or could be made equivalent through any type of >>>>>> folding that one can think up? >>>>>> The answer to that is "no", it's provable that this is not statically >>>>>> decidable, so the time bound doesn't matter :) >>>>>> >>>>>> You have to limit the possible folding/evaluation you apply in >>>>>> various ways to make this decidable, and then further limit it to make the >>>>>> time bound reasonable. >>>>>> >>>>>> This all quickly devolves into herbrand equivalence and it's >>>>>> variations. >>>>>> >>>>>> >>>>>> >>>>> Let me try one more time :) May we need multiple reassociate passes >>>>> to fold different reassociative patterns? >>>>> >>>>> A longer version: If Sanjay wants a particular reassociative pattern >>>>> to be folded (D45842), Omer wants another particular reassociative >>>>> pattern to be folded (D41574), and I want yet another particular >>>>> reassociative pattern to be folded (D46336), would we potentially >>>>> need three different reassociate passes with each combined with >>>>> instcombine, rather than just one that may be able to somehow handle >>>>> those cases in one shot, (assuming we don't want to put those in >>>>> instcombine)? >>>>> >>>>> And it sounds like the answer is yes? >>>>> >>>> >>>> If you take the current instcombine as a base, then yes, that is >>>> correct. >>>> >>> >>>> If you are willing to rearchitect instcombine, the answer is no, it's >>>> possible to do this all in a single pass in a relatively sane way. >>>> >>> >>> I assume by rearchitect, you mean a major rewrite as per this comment: "Is >>> there a way to determine all expressions in the program as it exists that >>> are equivalent or equivalent under constant time constant >>> folding/reassociation, in a reasonable time bound -> yes". Any pointer >>> or time to chat? >>> >> >> I'm happy to do both. >> >> >>> >>> I think that an approach like >>> D >>> >>> >>> 4 >>> >>> >>> 6 >>> >>> >>> 3 >>> >>> >>> 3 >>> >>> >>> 6 >>> >>> >>> / >>> >>> D >>> >>> >>> 4 >>> >>> >>> 6 >>> >>> >>> 5 >>> >>> >>> 9 >>> >>> >>> 5 >>> >>> has a merit: it would adds a bit of complexity, but would not >>> require: >>> >>> 1. a major rewrite of instcombine, >>> 2. writing multiple (potentially many) reassociate passes and figuring >>> out how to fixpoint them with instcombine, or >>> 3. writing a self-contained folding pass for a specific pattern >>> >>> If you look at the diffs in the existing .ll files in >>> D46336 >>> , it helps fold some previously-unfolded reassociation patterns beyond >>> the bit check patterns that it originally targeted. >>> >>> Sure, and it does so by adding another O(N) cost to evaluation in each >> case. Instcombine doesn't even do lazy reevaluation through tracking >> dependencies, so it'll do so a lot of times as well. >> >> To me, that's not a good tradeoff, especially given how slow instcombine >> is *already*. The code it produces is "good enough" to stop for a while >> and do something else and not suffer horribly in performance.[1] >> >> Let me ask a different question: >> >> At what point would anyone here be willing to stop adding things to >> instcombine and start doing something else instead, instead of waiting for >> someone else to do it? >> As far as i can tell, the answer is: "never", which makes most of these >> discussions just pointless rehashes as we slowly repeat the same disaster >> that became gcc's instruction combiner :) >> >> If the answer is "something", great, i'll set a mail filter and ignore >> these threads until that something happens :) >> >> Personally, in my experience people will never do more here unless pushed >> somewhat, or the thing becomes such a complete disaster no one wants to >> touch it. >> >> >> I've said this before, but I think a major impediment to forward progress >> here is coming up with an agreement on what the "something else" should be. >> Some of us have talked for years about having some TableGen-driven >> replacement, or maybe we want something with a syntax more like what is >> used by the Alive tool, but regardless, in order to gain in efficiency I >> suspect we need a model that is more restrictive than more-or-less >> arbitrary C++ code, and so we should pick a model and figure out how things >> might work. >> >> >> [1] Last year i computed the "improvement in performance on >> applications" due to instcombine for a bunch of google apps and open source >> apps that had easy to use benchmarks (IE I isolated about two years of >> instcombine changes and made them to a current compiler piece by piece >> while measuring performance). >> I also computed the compile time increase in single instcombine passes >> over the same time period. >> >> On x86, but the numbers basically said we were basically gaining nearly >> nothing for high cost. IE our drive for better looking output does not >> appear to translate into any real gains that i can find. Either >> improvements to other opts hid them, or they simply didn't matter on the >> processors i tested on. >> >> Certainly, apps/workloads/architectures may vary here, and my goal is not >> to claim it's all worthless. >> My actual goal in all of this was to get a sense of whether my >> perspective on instcombine was still "reasonable", not to do a true >> scientific exploration :) >> I didn't have time/energy/etc to run it elsewhere, and again, my goal was >> not to give certainty/try to give exact percentages. >> >> >> This also matches my experience, but I draw a somewhat different lesson. >> I often tell application developers that *this* is why they must file >> compiler bug reports. Waiting and assuming that someone else will hit the >> same problem, and file the report, is a bad strategy. I think that this is >> due to two things: >> >> 1. As far as things go, the tail of the distribution is often really >> long, and probability that the particular thing hampering one piece of hot >> code is the same thing hampering another piece of hot code is often small. >> >> 2. We tend to add special cases instead of adding more-general >> algorithms. The more-general work is often hard because figuring out the >> cost modeling is often highly non-trivial. Also, when it's finally done, >> the chances that the old special cases are removed is also small (so we'll >> still accumulate cruft without specific effort). >> >> > I don't have better or larger study results or data, but for fixing > some smaller but important enough performance degradations (often in > microbenchmarks, but occasionally in larger settings) smaller compiler > improvements (not necessarily in instcombine but around that level) do > seem to matter at least in individual contexts. It's not clear how much > those actually mattered in a grander scheme of things, though. > > It sounds like we don't want to add to instcombine due to added > cost/complexity, rearchitecting would be hard, it's not clear if > incremental changes in instcombine did much per cost, which would make it > harder to justify rearchitecting it... > > What do people think of approaches like D41574 and D45842? > > > >> -Hal >> >> >> --Dan >> >> >> >> >> >> >> >> _______________________________________________ >> LLVM Developers mailing listllvm-dev at lists.llvm.orghttp://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev >> >> >> -- >> Hal Finkel >> Lead, Compiler Technology and Programming Languages >> Leadership Computing Facility >> Argonne National Laboratory >> >> > _______________________________________________ > LLVM Developers mailing list > llvm-dev at lists.llvm.org > http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev > >-------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20180518/d082c260/attachment.html>
Rebased D41574 to trunk. From: Sanjay Patel [mailto:spatel at rotateright.com] Sent: Saturday, May 19, 2018 00:31 To: Hiroshi Yamauchi <yamauchi at google.com> Cc: Hal Finkel <hfinkel at anl.gov>; llvm-dev <llvm-dev at lists.llvm.org>; Paparo Bivas, Omer <omer.paparo.bivas at intel.com> Subject: Re: [llvm-dev] more reassociation in IR I mentioned this earlier in the thread - I would like to see something like D41574 in the optimizer. It's optimizing code that no other pass does currently, and I don't see any other near-term proposal that gets us those optimizations. Omer, can you rebase that to trunk? I think a header has moved, so it doesn't build as-is. I'd like to know if it can catch the cases in D45842. If not, why not? If it can handle those easily, I'll abandon D45842. Also, I don't know if it's better to include that functionality as another iteration of the existing -reassociate or split it off as its own pass. But I think it should do the distributive simplifications that are currently in -instcombine (InstCombiner::SimplifyUsingDistributiveLaws). Using that instsimplify logic for analysis lets us decide if the reassociation is worthwhile in the 1st place, it removes the risk that some other pass would somehow mess up the pattern before instcombine could zap it, and it reduces the burden on instcombine to be the entire optimizer. :) On Mon, May 14, 2018 at 1:34 PM, Hiroshi Yamauchi via llvm-dev <llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org>> wrote: On Fri, May 11, 2018 at 7:20 PM Hal Finkel <hfinkel at anl.gov<mailto:hfinkel at anl.gov>> wrote: On 05/11/2018 08:40 PM, Daniel Berlin via llvm-dev wrote: On Fri, May 11, 2018 at 2:37 PM, Hiroshi Yamauchi <yamauchi at google.com<mailto:yamauchi at google.com>> wrote: On Thu, May 10, 2018 at 12:49 PM Daniel Berlin <dberlin at dberlin.org<mailto:dberlin at dberlin.org>> wrote: On Thu, May 10, 2018 at 12:05 PM, Hiroshi Yamauchi <yamauchi at google.com<mailto:yamauchi at google.com>> wrote: On Wed, May 9, 2018 at 8:24 PM Daniel Berlin <dberlin at dberlin.org<mailto:dberlin at dberlin.org>> wrote: On Wed, May 9, 2018 at 10:39 AM, Hiroshi Yamauchi <yamauchi at google.com<mailto:yamauchi at google.com>> wrote: On Tue, May 8, 2018 at 11:15 AM Daniel Berlin <dberlin at dberlin.org<mailto:dberlin at dberlin.org>> wrote: On Tue, May 8, 2018 at 10:38 AM, Hiroshi Yamauchi via llvm-dev <llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org>> wrote: ( I came across this issue in the context of D46336<https://reviews.llvm.org/D46336>. Thanks, Sanjay, for starting this discussion.) If we will move reassociation, or keep additional ones , out of instcombine, open questions for me would be : 1. Since -reassociate isn't a fixed point pass, This is fixable, fwiw, without fixpointing it. How? Depends on specifically which part you would like to know about ;) Maybe I misunderstood what you meant by "This is fixable". Did you mean that we won't somehow need to fixpoint between instcombine and reassociate, or that the specific motivating examples from the above differentials are foldable without fixpointing? If by fixpointing you mean "fixpointing reassociate and instcombine", then yes, that is fixable without fixpointing reassociate and instcombine, but would require rewriting instcombine :) If the latter, that may be the case. The concern was that we may encounter examples that may need many more iterations, if not fixpointing. As long as it's feasible to fixpoint between instcombine and reassociate, it seems to work, but I guess that would probably need some pass management change. we might need to repeat "-instcombine -reassociate" multiple times to fold down to what we want (relating to my comment here<https://reviews.llvm.org/D46336#1087082>). I assumed this isn't not what we want to do ? My impression is we don't do a fixed-point with passes? Well, i mean there is no practical difference between passes that we fixpoint externally and fixpoint internally. I had the following in mind: Does the pass manager support fixpointing externally? Is there any performance difference? Are people okay with that in general? But if there is no practical difference, I don't see any problem with that :) 2. Since -reassociate needs to come up with one operand order (at least currently as the only reassociate pass), would there exist a single, unique operand order that would enable all reassociative/commutative foldings that we want? In what way? Are you asking whether there is a single reassociation order that makes all foldings occur in the same operation or something? I don't feel like i understand what you are asking. Does this rephrase help: with the motivating examples (like and-of-shifts or bit check patterns) from the above differentials in mind, can we come up with a single reassociation order that solves all those and all the others that may come up in the future? Would we need different reassociation orders to fold different patterns? It doesn't quite help. When stated that generally, there can be no such ordering at all, that's easy to prove. It is a statically undecidable problem. There is however, a different question and answer to a few related problems that maybe you are really asking? 1. Is there a way to determine and apply the a maximal or nearly-maximal set of folds/graph transforms that could be applied to a given set of code in a sane and principled way -> yes (see, e.g., http://www.cs.cornell.edu/~ross/publications/eqsat/<http://www.cs.cornell.edu/%7Eross/publications/eqsat/>) 2. Is there a way to determine all expressions in the program as it exists that are equivalent or equivalent under constant time constant folding/reassociation, in a reasonable time bound -> yes (not a single easy link, happy to talk about it) Your original question is basically equivalent to Is there a way to determine all expressions in the program as it exists that are equivalent or could be made equivalent through any type of folding that one can think up? The answer to that is "no", it's provable that this is not statically decidable, so the time bound doesn't matter :) You have to limit the possible folding/evaluation you apply in various ways to make this decidable, and then further limit it to make the time bound reasonable. This all quickly devolves into herbrand equivalence and it's variations. Let me try one more time :) May we need multiple reassociate passes to fold different reassociative patterns? A longer version: If Sanjay wants a particular reassociative pattern to be folded (D45842), Omer wants another particular reassociative pattern to be folded (D41574), and I want yet another particular reassociative pattern to be folded (D46336), would we potentially need three different reassociate passes with each combined with instcombine, rather than just one that may be able to somehow handle those cases in one shot, (assuming we don't want to put those in instcombine)? And it sounds like the answer is yes? If you take the current instcombine as a base, then yes, that is correct. If you are willing to rearchitect instcombine, the answer is no, it's possible to do this all in a single pass in a relatively sane way. I assume by rearchitect, you mean a major rewrite as per this comment: "Is there a way to determine all expressions in the program as it exists that are equivalent or equivalent under constant time constant folding/reassociation, in a reasonable time bound -> yes". Any pointer or time to chat? I'm happy to do both. I think that an approach like D 4 6 3 3 6 / D 4 6 5 9 5 has a merit: it would adds a bit of complexity, but would not require: 1. a major rewrite of instcombine, 2. writing multiple (potentially many) reassociate passes and figuring out how to fixpoint them with instcombine, or 3. writing a self-contained folding pass for a specific pattern If you look at the diffs in the existing .ll files in D46336 , it helps fold some previously-unfolded reassociation patterns beyond the bit check patterns that it originally targeted. Sure, and it does so by adding another O(N) cost to evaluation in each case. Instcombine doesn't even do lazy reevaluation through tracking dependencies, so it'll do so a lot of times as well. To me, that's not a good tradeoff, especially given how slow instcombine is *already*. The code it produces is "good enough" to stop for a while and do something else and not suffer horribly in performance.[1] Let me ask a different question: At what point would anyone here be willing to stop adding things to instcombine and start doing something else instead, instead of waiting for someone else to do it? As far as i can tell, the answer is: "never", which makes most of these discussions just pointless rehashes as we slowly repeat the same disaster that became gcc's instruction combiner :) If the answer is "something", great, i'll set a mail filter and ignore these threads until that something happens :) Personally, in my experience people will never do more here unless pushed somewhat, or the thing becomes such a complete disaster no one wants to touch it. I've said this before, but I think a major impediment to forward progress here is coming up with an agreement on what the "something else" should be. Some of us have talked for years about having some TableGen-driven replacement, or maybe we want something with a syntax more like what is used by the Alive tool, but regardless, in order to gain in efficiency I suspect we need a model that is more restrictive than more-or-less arbitrary C++ code, and so we should pick a model and figure out how things might work. [1] Last year i computed the "improvement in performance on applications" due to instcombine for a bunch of google apps and open source apps that had easy to use benchmarks (IE I isolated about two years of instcombine changes and made them to a current compiler piece by piece while measuring performance). I also computed the compile time increase in single instcombine passes over the same time period. On x86, but the numbers basically said we were basically gaining nearly nothing for high cost. IE our drive for better looking output does not appear to translate into any real gains that i can find. Either improvements to other opts hid them, or they simply didn't matter on the processors i tested on. Certainly, apps/workloads/architectures may vary here, and my goal is not to claim it's all worthless. My actual goal in all of this was to get a sense of whether my perspective on instcombine was still "reasonable", not to do a true scientific exploration :) I didn't have time/energy/etc to run it elsewhere, and again, my goal was not to give certainty/try to give exact percentages. This also matches my experience, but I draw a somewhat different lesson. I often tell application developers that *this* is why they must file compiler bug reports. Waiting and assuming that someone else will hit the same problem, and file the report, is a bad strategy. I think that this is due to two things: 1. As far as things go, the tail of the distribution is often really long, and probability that the particular thing hampering one piece of hot code is the same thing hampering another piece of hot code is often small. 2. We tend to add special cases instead of adding more-general algorithms. The more-general work is often hard because figuring out the cost modeling is often highly non-trivial. Also, when it's finally done, the chances that the old special cases are removed is also small (so we'll still accumulate cruft without specific effort). I don't have better or larger study results or data, but for fixing some smaller but important enough performance degradations (often in microbenchmarks, but occasionally in larger settings) smaller compiler improvements (not necessarily in instcombine but around that level) do seem to matter at least in individual contexts. It's not clear how much those actually mattered in a grander scheme of things, though. It sounds like we don't want to add to instcombine due to added cost/complexity, rearchitecting would be hard, it's not clear if incremental changes in instcombine did much per cost, which would make it harder to justify rearchitecting it... What do people think of approaches like D41574 and D45842? -Hal --Dan _______________________________________________ LLVM Developers mailing list llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev -- Hal Finkel Lead, Compiler Technology and Programming Languages Leadership Computing Facility Argonne National Laboratory _______________________________________________ LLVM Developers mailing list llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev --------------------------------------------------------------------- Intel Israel (74) Limited This e-mail and any attachments may contain confidential material for the sole use of the intended recipient(s). Any review or distribution by others is strictly prohibited. If you are not the intended recipient, please contact the sender and delete all copies. -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20180522/95a2aad5/attachment-0001.html>
Taking an approach like D41574 and/or D45842 would mean modifying the existing reassociate pass, adding some new reassociate-like passes, etc. I think if there's any concern/objection to it as a general direction in the short term, it'd be good to discuss here. But there doesn't seem to be? On Fri, May 18, 2018 at 2:31 PM Sanjay Patel <spatel at rotateright.com> wrote:> I mentioned this earlier in the thread - I would like to see something > like D41574 in the optimizer. It's optimizing code that no other pass does > currently, and I don't see any other near-term proposal that gets us those > optimizations. >> Omer, can you rebase that to trunk? I think a header has moved, so it > doesn't build as-is. I'd like to know if it can catch the cases in D45842. > If not, why not? If it can handle those easily, I'll abandon D45842. > > Also, I don't know if it's better to include that functionality as another > iteration of the existing -reassociate or split it off as its own pass. But > I think it should do the distributive simplifications that are currently in > -instcombine (InstCombiner::SimplifyUsingDistributiveLaws). Using that > instsimplify logic for analysis lets us decide if the reassociation is > worthwhile in the 1st place, it removes the risk that some other pass would > somehow mess up the pattern before instcombine could zap it, and it reduces > the burden on instcombine to be the entire optimizer. :) > > > On Mon, May 14, 2018 at 1:34 PM, Hiroshi Yamauchi via llvm-dev < > llvm-dev at lists.llvm.org> wrote: > >> >> >> On Fri, May 11, 2018 at 7:20 PM Hal Finkel <hfinkel at anl.gov> wrote: >> >>> >>> On 05/11/2018 08:40 PM, Daniel Berlin via llvm-dev wrote: >>> >>> >>> >>> On Fri, May 11, 2018 at 2:37 PM, Hiroshi Yamauchi <yamauchi at google.com> >>> wrote: >>> >>>> >>>> >>>> On Thu, May 10, 2018 at 12:49 PM Daniel Berlin <dberlin at dberlin.org> >>>> wrote: >>>> >>>>> >>>>> >>>>> On Thu, May 10, 2018 at 12:05 PM, Hiroshi Yamauchi < >>>>> yamauchi at google.com> wrote: >>>>> >>>>>> >>>>>> >>>>>> On Wed, May 9, 2018 at 8:24 PM Daniel Berlin <dberlin at dberlin.org> >>>>>> wrote: >>>>>> >>>>>>> >>>>>>> >>>>>>> On Wed, May 9, 2018 at 10:39 AM, Hiroshi Yamauchi < >>>>>>> yamauchi at google.com> wrote: >>>>>>> >>>>>>>> >>>>>>>> >>>>>>>> On Tue, May 8, 2018 at 11:15 AM Daniel Berlin <dberlin at dberlin.org> >>>>>>>> wrote: >>>>>>>> >>>>>>>>> >>>>>>>>> >>>>>>>>> On Tue, May 8, 2018 at 10:38 AM, Hiroshi Yamauchi via llvm-dev < >>>>>>>>> llvm-dev at lists.llvm.org> wrote: >>>>>>>>> >>>>>>>>>> ( >>>>>>>>>> I came across this issue in the context of >>>>>>>>>> D46336 <https://reviews.llvm.org/D46336>. >>>>>>>>>> >>>>>>>>>> Thanks, Sanjay, for starting this discussion.) >>>>>>>>>> >>>>>>>>>> If >>>>>>>>>> we will >>>>>>>>>> move >>>>>>>>>> reassociation, >>>>>>>>>> or keep additional ones >>>>>>>>>> , >>>>>>>>>> out of instcombine, >>>>>>>>>> open questions for me would be >>>>>>>>>> : >>>>>>>>>> >>>>>>>>>> >>>>>>>>>> 1. Since -reassociate isn't a fixed point pass, >>>>>>>>>> >>>>>>>>> >>>>>>>>> This is fixable, fwiw, without fixpointing it. >>>>>>>>> >>>>>>>> >>>>>>>> How? >>>>>>>> >>>>>>> >>>>>>> Depends on specifically which part you would like to know about ;) >>>>>>> >>>>>> >>>>>> Maybe I misunderstood what you meant by "This is fixable". Did you >>>>>> mean that we won't somehow need to fixpoint between instcombine and >>>>>> reassociate, or that the specific motivating examples from the above >>>>>> differentials are foldable without fixpointing? >>>>>> >>>>> >>>>> If by fixpointing you mean "fixpointing reassociate and instcombine", >>>>> then yes, that is fixable without fixpointing reassociate and instcombine, >>>>> but would require rewriting instcombine :) >>>>> >>>>> >>>>>> >>>>>> If the latter, that may be the case. The concern was that we may >>>>>> encounter examples that may need many more iterations, if not fixpointing. >>>>>> As long as it's feasible to fixpoint between instcombine and >>>>>> reassociate, it seems to work, but I guess that would probably need >>>>>> some pass management change. >>>>>> >>>>>> >>>>>>> >>>>>>>> >>>>>>>> >>>>>>>>> >>>>>>>>>> we might need to repeat "-instcombine -reassociate" multiple >>>>>>>>>> times to >>>>>>>>>> fold >>>>>>>>>> down to what we want (relating to my comment here >>>>>>>>>> <https://reviews.llvm.org/D46336#1087082>). I assumed this isn't >>>>>>>>>> not what we want to do >>>>>>>>>> ? My impression is we don't do a fixed-point with passes? >>>>>>>>>> >>>>>>>>> >>>>>>>>> Well, i mean there is no practical difference between passes that >>>>>>>>> we fixpoint externally and fixpoint internally. >>>>>>>>> >>>>>>>> >>>>>>>> I had the following in mind: Does the pass manager support >>>>>>>> fixpointing externally? Is there any performance difference? Are people >>>>>>>> okay with that in general? >>>>>>>> >>>>>>>> But if there is no practical difference, I don't see any problem >>>>>>>> with that :) >>>>>>>> >>>>>>>> >>>>>>>>> >>>>>>>>>> >>>>>>>>>> >>>>>>>>> 2. >>>>>>>>>> Since -reassociate needs to come up with one operand order (at >>>>>>>>>> least currently as the only reassociate pass), would there exist >>>>>>>>>> a single, unique operand order that would enable all >>>>>>>>>> reassociative/commutative foldings that we want? >>>>>>>>>> >>>>>>>>> >>>>>>>>> In what way? >>>>>>>>> Are you asking whether there is a single reassociation order that >>>>>>>>> makes all foldings occur in the same operation or something? >>>>>>>>> I don't feel like i understand what you are asking. >>>>>>>>> >>>>>>>> >>>>>>>> Does this rephrase help: with the motivating examples (like >>>>>>>> and-of-shifts or bit check patterns) from the above differentials in mind, >>>>>>>> can we come up with a single reassociation order that solves all >>>>>>>> those and all the others that may come up in the future? Would we need >>>>>>>> different reassociation orders to fold different patterns? >>>>>>>> >>>>>>> >>>>>>> It doesn't quite help. >>>>>>> When stated that generally, there can be no such ordering at all, >>>>>>> that's easy to prove. It is a statically undecidable problem. >>>>>>> >>>>>>> There is however, a different question and answer to a few related >>>>>>> problems that maybe you are really asking? >>>>>>> 1. Is there a way to determine and apply the a maximal or >>>>>>> nearly-maximal set of folds/graph transforms that could be applied to a >>>>>>> given set of code in a sane and principled way -> yes >>>>>>> >>>>>>> (see, e.g., http://www.cs.cornell.edu/~ross/publications/eqsat/) >>>>>>> >>>>>>> 2. Is there a way to determine all expressions in the program as it >>>>>>> exists that are equivalent or equivalent under constant time constant >>>>>>> folding/reassociation, in a reasonable time bound -> yes >>>>>>> >>>>>>> (not a single easy link, happy to talk about it) >>>>>>> >>>>>>> Your original question is basically equivalent to >>>>>>> Is there a way to determine all expressions in the program as it >>>>>>> exists that are equivalent or could be made equivalent through any type of >>>>>>> folding that one can think up? >>>>>>> The answer to that is "no", it's provable that this is not >>>>>>> statically decidable, so the time bound doesn't matter :) >>>>>>> >>>>>>> You have to limit the possible folding/evaluation you apply in >>>>>>> various ways to make this decidable, and then further limit it to make the >>>>>>> time bound reasonable. >>>>>>> >>>>>>> This all quickly devolves into herbrand equivalence and it's >>>>>>> variations. >>>>>>> >>>>>>> >>>>>>> >>>>>> Let me try one more time :) May we need multiple reassociate passes >>>>>> to fold different reassociative patterns? >>>>>> >>>>>> A longer version: If Sanjay wants a particular reassociative pattern >>>>>> to be folded (D45842), Omer wants another particular reassociative >>>>>> pattern to be folded (D41574), and I want yet another particular reassociative >>>>>> pattern to be folded (D46336), would we potentially need three >>>>>> different reassociate passes with each combined with instcombine, rather >>>>>> than just one that may be able to somehow handle those cases in one >>>>>> shot, (assuming we don't want to put those in instcombine)? >>>>>> >>>>>> And it sounds like the answer is yes? >>>>>> >>>>> >>>>> If you take the current instcombine as a base, then yes, that is >>>>> correct. >>>>> >>>> >>>>> If you are willing to rearchitect instcombine, the answer is no, it's >>>>> possible to do this all in a single pass in a relatively sane way. >>>>> >>>> >>>> I assume by rearchitect, you mean a major rewrite as per this comment: >>>> "Is there a way to determine all expressions in the program as it >>>> exists that are equivalent or equivalent under constant time constant >>>> folding/reassociation, in a reasonable time bound -> yes". Any pointer >>>> or time to chat? >>>> >>> >>> I'm happy to do both. >>> >>> >>>> >>>> I think that an approach like >>>> D >>>> >>>> >>>> 4 >>>> >>>> >>>> 6 >>>> >>>> >>>> 3 >>>> >>>> >>>> 3 >>>> >>>> >>>> 6 >>>> >>>> >>>> / >>>> >>>> D >>>> >>>> >>>> 4 >>>> >>>> >>>> 6 >>>> >>>> >>>> 5 >>>> >>>> >>>> 9 >>>> >>>> >>>> 5 >>>> >>>> has a merit: it would adds a bit of complexity, but would not >>>> require: >>>> >>>> 1. a major rewrite of instcombine, >>>> 2. writing multiple (potentially many) reassociate passes and figuring >>>> out how to fixpoint them with instcombine, or >>>> 3. writing a self-contained folding pass for a specific pattern >>>> >>>> If you look at the diffs in the existing .ll files in >>>> D46336 >>>> , it helps fold some previously-unfolded reassociation patterns >>>> beyond the bit check patterns that it originally targeted. >>>> >>>> Sure, and it does so by adding another O(N) cost to evaluation in each >>> case. Instcombine doesn't even do lazy reevaluation through tracking >>> dependencies, so it'll do so a lot of times as well. >>> >>> To me, that's not a good tradeoff, especially given how slow instcombine >>> is *already*. The code it produces is "good enough" to stop for a while >>> and do something else and not suffer horribly in performance.[1] >>> >>> Let me ask a different question: >>> >>> At what point would anyone here be willing to stop adding things to >>> instcombine and start doing something else instead, instead of waiting for >>> someone else to do it? >>> As far as i can tell, the answer is: "never", which makes most of these >>> discussions just pointless rehashes as we slowly repeat the same disaster >>> that became gcc's instruction combiner :) >>> >>> If the answer is "something", great, i'll set a mail filter and ignore >>> these threads until that something happens :) >>> >>> Personally, in my experience people will never do more here unless >>> pushed somewhat, or the thing becomes such a complete disaster no one wants >>> to touch it. >>> >>> >>> I've said this before, but I think a major impediment to forward >>> progress here is coming up with an agreement on what the "something else" >>> should be. Some of us have talked for years about having some >>> TableGen-driven replacement, or maybe we want something with a syntax more >>> like what is used by the Alive tool, but regardless, in order to gain in >>> efficiency I suspect we need a model that is more restrictive than >>> more-or-less arbitrary C++ code, and so we should pick a model and figure >>> out how things might work. >>> >>> >>> [1] Last year i computed the "improvement in performance on >>> applications" due to instcombine for a bunch of google apps and open source >>> apps that had easy to use benchmarks (IE I isolated about two years of >>> instcombine changes and made them to a current compiler piece by piece >>> while measuring performance). >>> I also computed the compile time increase in single instcombine passes >>> over the same time period. >>> >>> On x86, but the numbers basically said we were basically gaining nearly >>> nothing for high cost. IE our drive for better looking output does not >>> appear to translate into any real gains that i can find. Either >>> improvements to other opts hid them, or they simply didn't matter on the >>> processors i tested on. >>> >>> Certainly, apps/workloads/architectures may vary here, and my goal is >>> not to claim it's all worthless. >>> My actual goal in all of this was to get a sense of whether my >>> perspective on instcombine was still "reasonable", not to do a true >>> scientific exploration :) >>> I didn't have time/energy/etc to run it elsewhere, and again, my goal >>> was not to give certainty/try to give exact percentages. >>> >>> >>> This also matches my experience, but I draw a somewhat different lesson. >>> I often tell application developers that *this* is why they must file >>> compiler bug reports. Waiting and assuming that someone else will hit the >>> same problem, and file the report, is a bad strategy. I think that this is >>> due to two things: >>> >>> 1. As far as things go, the tail of the distribution is often really >>> long, and probability that the particular thing hampering one piece of hot >>> code is the same thing hampering another piece of hot code is often small. >>> >>> 2. We tend to add special cases instead of adding more-general >>> algorithms. The more-general work is often hard because figuring out the >>> cost modeling is often highly non-trivial. Also, when it's finally done, >>> the chances that the old special cases are removed is also small (so we'll >>> still accumulate cruft without specific effort). >>> >>> >> I don't have better or larger study results or data, but for fixing >> some smaller but important enough performance degradations (often in >> microbenchmarks, but occasionally in larger settings) smaller compiler >> improvements (not necessarily in instcombine but around that level) do >> seem to matter at least in individual contexts. It's not clear how much >> those actually mattered in a grander scheme of things, though. >> >> It sounds like we don't want to add to instcombine due to added >> cost/complexity, rearchitecting would be hard, it's not clear if >> incremental changes in instcombine did much per cost, which would make it >> harder to justify rearchitecting it... >> >> What do people think of approaches like D41574 and D45842? >> >> >> >>> -Hal >>> >>> >>> --Dan >>> >>> >>> >>> >>> >>> >>> >>> _______________________________________________ >>> LLVM Developers mailing listllvm-dev at lists.llvm.orghttp://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev >>> >>> >>> -- >>> Hal Finkel >>> Lead, Compiler Technology and Programming Languages >>> Leadership Computing Facility >>> Argonne National Laboratory >>> >>> >> _______________________________________________ >> LLVM Developers mailing list >> llvm-dev at lists.llvm.org >> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev >> >> >-------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20180525/fbc567a6/attachment-0001.html>