Hi all, I think Pablo answered most of the questions. But there is another important concern: Hermes is not designed to be a general purpose serializer. So it not in it goals to be extensible or fast. The main goals behind hermes are: - to be small (actually smaller than the compiler) - to have as few dependencies as possible
This is because we want to have it in the first bootstrapped image to have a small kernel with faster build time, and be able to load the compiler afterwards using it. Concretely, after bootstrap we use Hermes to load only four packages (Compiler + 3 dependencies). Once that is done, we can go back into using the compiler to load the rest. We did not think about putting it at the level of monticello, that has to be thought for the future. In any case, moving to git/filetree makes sharing binary mcz more difficult. Another two points that hermes is not managing so far to be a more general code loading solution is that it does not manage source files (it installs methods without source) nor the reloading/update of code. >From an implementation/format point of view, Hermes writes something somehow equivalent to a java .class file but with a package granularity. I think it may be valuable to nicely document the format. That could allow other dialects to have their implementation and share hermes files (delta some runtime compatibility :) ). Guille On Tue, Aug 1, 2017 at 11:14 PM, teso...@gmail.com <teso...@gmail.com> wrote: > Hi Mariano, > The definition is not serialized. What is serialized is all the > information to regenerate the class (eg: superclass name, instance variable > names, class variables names, share pool names, category, layout class, and > so on). This information is later used to regenerate the class as it is > created initially by the class builder and installer. The same happens with > the methods, they are built up from scratch without using the compiler. The > bytecode is stored, the literals are serialized and the information in the > header is serialized, as it has to be rebuilt when install. > > The same is applicable to the traits and extension methods. > > Basically it can be seen as a class builder / installer, that get all the > information from the serialized version and not from the text > representation of the class or the methods. > > That's why I said Hermes is so couple with Pharo. > > Thanks for clarifying, I am starting to understand where yours doubts came > from. > > Cheers, > Pablo > > On Tue, Aug 1, 2017 at 10:52 PM, Mariano Martinez Peck < > marianop...@gmail.com> wrote: > >> >> >> On Tue, Aug 1, 2017 at 5:31 PM, teso...@gmail.com <teso...@gmail.com> >> wrote: >> >>> Hi, >>> Yes for sure there are objects, and that is very powerful, but >>> comparing the algorithm of serialization of Hermes with Fuel (I don't >>> really know Parcels) is not possible. It is like comparing Filetree with >>> Fuel, The problems they have to take care are completely different. >>> >>> Maybe I am not seeing the point. >>> >>> >> Let me add one sentence that may explain Eliot questions... if you >> serialize the class definition, then it means you would need the compiler. >> I guess Eliot may be interested in NOT having the compiler at all and >> instead bring classes directly from binary format (like Fuel / Tanker ) >> >> >> >> >>> Cheers. >>> >>> On Tue, Aug 1, 2017 at 10:25 PM, Eliot Miranda <eliot.mira...@gmail.com> >>> wrote: >>> >>>> Hi Pablo, >>>> >>>> On Tue, Aug 1, 2017 at 12:36 PM, teso...@gmail.com <teso...@gmail.com> >>>> wrote: >>>> >>>>> Hi again, >>>>> thanks for the clarification, I haven't understood the question >>>>> initially, but now I think I can answer it. >>>>> >>>>> Hermes is only exporting the classes, not objects. so it does not have >>>>> to handle complex graphs of objects. Basically it serializes the >>>>> definition >>>>> of the classes, traits and methods. Then they are loaded in sequence. The >>>>> only caring during the generation is to serialize first the superclasses, >>>>> and then the subclasses. There is no way of serializing objects outside >>>>> the >>>>> classes, methods, traits and literals in a method. >>>>> >>>> >>>> Classes, traits, methods, literals *are* a graph of objects :-). The >>>> Parcel architecture, from which Fuel derived its architecture, was designed >>>> for loading code in VisualWorks. In fact, last time I looked Parcels were >>>> used only to store code and not as a general purpose (de)serializer. >>>> >>>> So the answer is that is not using Parcels or Fuel architecture because >>>>> they are intended for different uses. >>>>> >>>> >>>> Well, I don't think that's the reason ;-), but fine. I understand that >>>> it doesn't use this architecture. Thanks. >>>> >>>> >>>>> Also the format itself has no special design or considerations to be >>>>> fast to generate or fast to read (as It happens with Fuel) >>>>> >>>>> Cheers, >>>>> Pablo >>>>> >>>>> On Tue, Aug 1, 2017 at 9:16 PM, Eliot Miranda <eliot.mira...@gmail.com >>>>> > wrote: >>>>> >>>>>> Hi Pablo, >>>>>> >>>>>> I understand that Hermes has its own format. My question was >>>>>> about architecture. The Parcels/Fuel architecture is, AFAIA, the best >>>>>> architecture for binary storage because it is virtual, is fast to load >>>>>> and >>>>>> cleanly separates loading from initialization. What do I mean? >>>>>> >>>>>> - concrete formats like ImageSegment are tied to internal >>>>>> representations in the VM and do are difficult to read in forever FB >>>>>> systems and very hard to write in foreign systems. So a virtual format >>>>>> (like BOSS, Fuel, Parcels, etc) is better but unless done well can be >>>>>> slow >>>>>> >>>>>> - a first generation system like BOSS is a flattening of a graph >>>>>> traversal using a simple grammar. Each phrase is either an object >>>>>> reference (object id) or an object definition (object is followed by type >>>>>> info and contents, contents being a sequence of phrases. This is slow to >>>>>> parse (because each phrase must be decoded), and has invalid intermediate >>>>>> states (e.g. a Set whose contents are not all present, hence whose >>>>>> external >>>>>> hash may change during loading, etc). A second generation system like >>>>>> Parcels and Fuel separates objects from references (nodes from arcs) and >>>>>> batches object creation, grouping all instances of a given class for >>>>>> bulk, >>>>>> and hence fast, instantiation. Following my the objects are the >>>>>> references. So loading has three phases: >>>>>> - instantiate the objects >>>>>> - connect the graph by filling in object references >>>>>> - initialize the objects that require it (e.g. rehash sets) >>>>>> >>>>>> Consequently the writer must be two pass, the first pass to collect >>>>>> the objects in the graph and sort them by class, the second to write the >>>>>> nodes followed by the references >>>>>> >>>>>> So let me ask again, does Hermes use the Parcels/Fuel architecture? >>>>>> >>>>>> _,,,^..^,,,_ (phone) >>>>>> >>>>>> On Aug 1, 2017, at 11:52 AM, "teso...@gmail.com" <teso...@gmail.com> >>>>>> wrote: >>>>>> >>>>>> Hi Eliot, >>>>>> >>>>>> The last version of Hermes, that I have generated today and have to >>>>>> be tested for real yet, is able to load and save from and to 32 and 64 >>>>>> bits, handling the conversion. Because the Hermes representation (if we >>>>>> can >>>>>> call it like that, because is really simple, does not care about the >>>>>> architecture). >>>>>> >>>>>> Hermes is using a custom format completely separated from Fuel or any >>>>>> other tool. >>>>>> >>>>>> This is a consequence of the design of Hermes. Hermes has been >>>>>> designed to only work as light way of loading code during the bootstrap. >>>>>> So >>>>>> that, it is heavily coupled with Pharo (it can be changed, for example >>>>>> implementing another loader / installer) and uses only really core >>>>>> classes. >>>>>> >>>>>> Of course it can be extended and used in another Smalltalk dialects, >>>>>> but I am not sure if the tool can be useful for them, or even for the >>>>>> Pharo >>>>>> community outside the Pharo bootstrap process. >>>>>> >>>>>> Today to be useful outside the bootstrap process, some work has to be >>>>>> done (perhaps a lot), but mainly new use cases have to be thinked. Once >>>>>> the >>>>>> Compiler, Monticello or Metacello is available, there is no need for >>>>>> Hermes. It is only used to generate some binary loadable packages for >>>>>> the >>>>>> compiler from the source code. >>>>>> >>>>>> Again, it was only thinked as a tool to improve the speed during the >>>>>> bootstrap. However, if somebody has ideas to use it or want to >>>>>> collaborate >>>>>> are of course welcomed. >>>>>> >>>>>> Hermes is actually in Github, because to me is more comfortable to >>>>>> have it there, but if somebody is really wanting to use it.... we can >>>>>> manage to have an export process to Smalltalkhub. >>>>>> >>>>>> I hope I have clarified a little the idea behind Hermes. >>>>>> >>>>>> Maybe you have better ideas of what can we do with it. >>>>>> >>>>>> >>>>>> On Tue, Aug 1, 2017 at 7:59 PM, Eliot Miranda < >>>>>> eliot.mira...@gmail.com> wrote: >>>>>> >>>>>>> Hi Guille, >>>>>>> >>>>>>> On Aug 1, 2017, at 4:29 AM, Guillermo Polito < >>>>>>> guillermopol...@gmail.com> wrote: >>>>>>> >>>>>>> >>>>>>> >>>>>>> On Tue, Aug 1, 2017 at 12:57 PM, philippe.b...@highoctane.be < >>>>>>> philippe.b...@gmail.com> wrote: >>>>>>> >>>>>>>> Massive. >>>>>>>> >>>>>>>> What is in the super small image? >>>>>>>> Is Hermes going to be a generic binary content loader? >>>>>>>> >>>>>>> >>>>>>> There is still work to do in this front. >>>>>>> - Hermes only works for 32bits format. We should adapt it to 64 >>>>>>> bits (immediate floats and so on...) >>>>>>> - There is no format validation. We should add one for safety. >>>>>>> >>>>>>> >>>>>>> Will Hermes be able to save on 32-bits and load on 64-bits and vice >>>>>>> verse? >>>>>>> >>>>>>> Does Hermes use the Parcels/Fuel architecture of saving nodes, >>>>>>> grouped by class, followed by the arcs? >>>>>>> >>>>>>> If yes to both of these, are you willing to keep it in Monticello >>>>>>> and collaborate with the Squeak & Cuis communities in developing Hermes? >>>>>>> >>>>>>> >>>>>>> >>>>>>>> Phil >>>>>>>> >>>>>>>> On Aug 1, 2017 12:36, "Stephane Ducasse" <stepharo.s...@gmail.com> >>>>>>>> wrote: >>>>>>>> >>>>>>>>> Hi Pavel >>>>>>>>> >>>>>>>>> This is super excellent! IMPRESSIVE. An image without the compiler >>>>>>>>> and >>>>>>>>> a reloadable compiler. >>>>>>>>> Super cool. >>>>>>>>> >>>>>>>>> Stef >>>>>>>>> >>>>>>>>> >>>>>>>>> On Tue, Aug 1, 2017 at 11:57 AM, Pavel Krivanek >>>>>>>>> <pavel.kriva...@gmail.com> wrote: >>>>>>>>> > Hello, >>>>>>>>> > >>>>>>>>> > we are checking a huge pull request #177 >>>>>>>>> > (https://github.com/pharo-project/pharo/pull/177) that will >>>>>>>>> change some >>>>>>>>> > basics of the bootstrap process: >>>>>>>>> > >>>>>>>>> > Now we will bootstrap a smaller image that will not include >>>>>>>>> compiler/parser. >>>>>>>>> > Compiler and related packages are exported and loaded using a >>>>>>>>> binary >>>>>>>>> > exporter named Hermes. >>>>>>>>> > The compiler is then used to load FileSystem and Monticello. The >>>>>>>>> rest of the >>>>>>>>> > bootstrap process will be the same as before. >>>>>>>>> > As the result we will have faster bootstrap and better system >>>>>>>>> modularization >>>>>>>>> > and possibilities. >>>>>>>>> > >>>>>>>>> > It required some modularization efforts: >>>>>>>>> > >>>>>>>>> > - simplified initialization scripts >>>>>>>>> > - Use Zinc converters and encoders instead of FilePathEncoder >>>>>>>>> and old >>>>>>>>> > TextConverter >>>>>>>>> > - Use Stdio instead of FileStream >>>>>>>>> > - Using File instead of FileSystem >>>>>>>>> > - Deprecated FileStream & childs (Moved to Deprecated70) >>>>>>>>> > - Extracted Path classes to their on package: FileSystem-Path >>>>>>>>> > - Moved OpalEncoders to their own package. They are required by >>>>>>>>> the runtime >>>>>>>>> > (not only for compilation) >>>>>>>>> > - Introduced AsciiCharset in the kernel to answer to #isLetter >>>>>>>>> #isUppercase >>>>>>>>> > and so on without requiring full Unicode tables from the >>>>>>>>> beginning >>>>>>>>> > - Cleaning up a bit the full exception logging infrastructure >>>>>>>>> (streams, >>>>>>>>> > transcript, files, stack printing...) >>>>>>>>> > - Split Ring methods required for system navigation to the >>>>>>>>> Ring-Navigation >>>>>>>>> > package >>>>>>>>> > - Remove usages of #translated in the kernel >>>>>>>>> > - Refactored the bootstrapping classes to remove duplications >>>>>>>>> > - Cleaning up dependencies in CompiledMethod>>printOn: >>>>>>>>> > - fix path printing >>>>>>>>> > >>>>>>>>> > We need to merge these changes at once and of course it can >>>>>>>>> cause some >>>>>>>>> > conflicts with the existing pull requests or external code. >>>>>>>>> Anyway, we need >>>>>>>>> > to merge it as soon as possible. >>>>>>>>> > >>>>>>>>> > So please, try to look at the PR and test the resultant image >>>>>>>>> [1] to avoid >>>>>>>>> > some major problems. >>>>>>>>> > >>>>>>>>> > [1] >>>>>>>>> > https://ci.inria.fr/pharo/view/7.0/job/70-PR-Check-Load/last >>>>>>>>> SuccessfulBuild/artifact/bootstrap-cache/Pharo7.0-32bit-9c06 >>>>>>>>> 91d.zip >>>>>>>>> > >>>>>>>>> > Cheers, >>>>>>>>> > -- Pavel >>>>>>>>> > >>>>>>>>> >>>>>>>>> >>>>>>> -- >>>>>>> >>>>>>> >>>>>>> >>>>>>> Guille Polito >>>>>>> >>>>>>> >>>>>>> Research Engineer >>>>>>> >>>>>>> French National Center for Scientific Research - >>>>>>> *http://www.cnrs.fr* <http://www.cnrs.fr> >>>>>>> >>>>>>> >>>>>>> >>>>>>> *Web:* *http://guillep.github.io* <http://guillep.github.io> >>>>>>> >>>>>>> *Phone: *+33 06 52 70 66 13 <+33%206%2052%2070%2066%2013> >>>>>>> >>>>>>> >>>>>> >>>>>> >>>>>> -- >>>>>> Pablo Tesone. >>>>>> teso...@gmail.com >>>>>> >>>>>> >>>>> >>>>> >>>>> -- >>>>> Pablo Tesone. >>>>> teso...@gmail.com >>>>> >>>> >>>> >>>> >>>> -- >>>> _,,,^..^,,,_ >>>> best, Eliot >>>> >>> >>> >>> >>> -- >>> Pablo Tesone. >>> teso...@gmail.com >>> >> >> >> >> -- >> Mariano >> http://marianopeck.wordpress.com >> > > > > -- > Pablo Tesone. > teso...@gmail.com > -- Guille Polito Research Engineer French National Center for Scientific Research - *http://www.cnrs.fr* <http://www.cnrs.fr> *Web:* *http://guillep.github.io* <http://guillep.github.io> *Phone: *+33 06 52 70 66 13
