| Re: [cdt-dev] Future of Managed Build |
Alexander,
FYI it is me Jantje who posted the mail you replied on and I did not consult Jonah.
Honestly I do not understand most of what you write. I have used many modeller tools but never a modeller modellers tool. I have no clue what EMF, GMF, "direct model implementation" nor "interpretation of modeled structures" stand for. That is clearly your expertise so it is great to have you on board.
Secondly, We are very early in the process. I'm shooting ideas
and requesting ideas so this idea can become more tangible. I do
not expect anything I'm shooting here to become the final
solution. I do hope some of my ideas become part of the solution.
When I wrote my mail I wanted to show something that resembled
how I see a complex toolchain modelled. For that I took the most
complex one I know; which is arduino one. Because I do not have
any modelling tools on my system and I had a column based model in
my head I used excel to draw something. This is a quick and good
way to validate the model. Basically you can think of the metadata
that assigned to the components and see whether you can get the
desired outcome.
It's that process that lead me to the notes of which one that said
"I understand that note 4 and 5 are a serious constraint for a
more generic modelling tool.". What I'm trying to say with that
note is: "I've been writing this mail now for over an hour and now
I see this model is not good enough; so I'll add this note and
send the mail any-ways hoping we can improve as a group" ;-)
I do not know what to expect from the modelling work (which I
won't be doing/driving). Based on your remarks it looks to me as
if you want to generate make cmake ... files directly from the
model. It is pretty hard for me to get my head around that. For
sure because each tool has it's limitations which restrict the
model.
Can you elaborate on how this would work?
For the routers I was assuming that one could have options like
"1 on 1", "All on one" or "custom". Each with their own metadata.
The metadata of the custom option could point to a class that
implements a interface and other needed stuff. The "1 on 1" and
"All on one" would be implementations of the same interface.
Do you agree that in order to work as a team these interfaces need
to be specified and be stable before we start implementing?
Best regards
Jantje
Jonah,
After re-reading your vision I have a perception that we have different understanding how the modeling should be applied here. Please don't take it as a criticism, the subject is to big and complex and I may be wrong with my interpretation of your ideas, but let me share my thoughts.
I assume we will use EMF - best in class for today and native to Eclipse ecosystem. And EMF is the most effective for declarative approach. It supports the "methods" for objects but this way will quickly direct you to a swamp of "@generated not".
The good example of the right EMF usage is Sirius: after years of "GMF Gen"-based exercises they found a way to switch from "direct model implementation" to a kind of "interpretation of modeled structures".
May be I'm not so clear in one sentence - this is a big topic more suitable for a book - but it is very important to understand the limits of effective applicability for a particular tool, EMF in our case.
I'll try to describe it on other words: we should not model "how". but "what". I know that you are familiar with Maven it asks us "what" do we have and then applies a set of "how" plugins according to more or less fixed algo. Gradle goes further and provides much more customization, but still with a good focus on declarative approach. But if you remember Ant - it is completely different, it requires to specify each and every step that is really not scalable and hardly maintainable approach.
So, the most important part is to describe carefully and structure effectively the "metadata" that "router" will operate on, not to model a "router" algo. Why? Because "router" who operates on metadata will be most effectively implemented either with OSGi services or by an existing tools, or by a chain of tools, but not with EMF. For example "make" works just perfect and we are not going to re-implement make, right? But we would like to avoid creating make files manually, yes. Well, not a problem, we can generate them if we have sufficient metadata, or generate cmake files, or generate manifests for other build systems. It will be relatively easy if we have good metadata that includes every aspect we may need to care about: host capabilities, target hardware, tool options, project structure, whatever else we need.
Having EMF-based metadata available via OSGi services - we already successfully tested this architecture with Eclipse Passage - "routers" can be developed in parallel according to your pictures. What is important - the same metadata will be available during all the project lyfecycle: create/export/import/refactor/build/run/debug/etc.
Nope, it is not planned to be multi years project before we can _start to_ get it into CDT, it may be done faster. However, it depends on resources, as you can guess this work is not funded ATM.
Also from organizational perspective it is not clear how it can get into CDT as it will have some "volume" with EMF-generated sources and everything:
- it may be a separate proposal - in this case it will need to go through incubation,
- or it may be a separate repository (GitHub, please, to get more people involved easily) inside the CDT project to be utilized part-by-part,
- or you may have other suggestions.
Regards,
AF
30.01.2020 1:56, Jan Baeyens пишет:
Op 27/01/2020 om 14:47 schreef Alexander Fedorov:
Hello,Great
We (with my good friends @ArSysOp) are working on Toolchain metamodel prototype that should cover the needs of CDT and will be flexible enough to be applied for sibling areas.
Of course there is a number of EMF usage tricks collected through all that years that we are going to use during this work. We know the model-to-text generation technologies from JET to Acceleo, so we will be able to create whatever we need, not only the default Java classes.
This is what I would expect. Basically in toolchain land there are "files" and "tools". There is only a very limited set of combinations you can have with "files" and "tools".
Currently we are polishing things that were presented for Postgres community 1 year ago during the "IDE for stored procedures" talk and discussion. Yes, it may sound curious, but the "toolchain model" idea with all the "target hardware description" and "project structure model" appears to be very similar for a lot of ecosystems.
IMHO recursivity at the file level does not exists and tool sequence is pretty obvious at the model level. However: correct me if I'm wrong.
I have been thinking hard about how to model a gcc toolchain and below is a representation of how I see things now
I haven't yet found a good name for what I called router. It is absolutely crucial to understand router to understand the model. All other stuff is pretty obvious
In it's simplest form router will create a list of "input file; output file" (like "input.cpp" "input.cpp.o"). for each item in the list tool can add a command which results in "input file; output file; command" which -in make world- is called a "rule".
A more complicated router may have multiple input files and multiple output files which means that the router produces a list of "list of input files; list of output files".
Also the tool can produce more than one command so the output of a router tool combinations should be a list of "list of input files; list of output files; ordered list of commands".The router can also filter input files. For instance in the example above the output of the "c to object" tool is send to both the archiver and the linker. It is up to the "collection of routers" to make sure the files are processed at the right place. Given a specific file: the model supports all options (archiver; linker; archiver and linker;none) and assumes the need for coding to support this construction.
The builder can simply go from left to right processing the router/tool columns one by one. Asking the routers the "list of input files; list of output files" checking whether a command needs to be run and if so ask the command to generate the "ordered list of commands" to run. Run these and move on to the next.
Note 1 that the router explicitly names the output file(s).
Note 2 when implementing this model it seems logical to have 2 generic routers ("1 on 1" -fi compile- and "all on 1" -fi archive-) "1 on 1" would need a "name provider" (append .o) and "all on 1" will need a output file (archive.ar). The model above needs "custom build" routers next to the generic ones
Note 3 Looking an the linker input files (which can be object files and archive files) one can argue that the "list of input files' needs to be a "list of lists of input files". One can also argue it is up to the tool to "triage the files". I haven't decided yet what I think is the best.
Note 4: this model assumes there is no order of commands in the column. It behaves as if all the commands from one column can be executed at the same time.
Note 5: the model assumes that all actions in a column must have finished before the next column can be "evaluated/executed". This allows for "non file based commands" like "pre build actions"; "post build actions" to have their own column like demonstrated below
Note 6:I understand that note 4 and 5 are a serious constraint for a more generic modelling tool.
Do you think this will be multi year or multi month project before we get it into CDT?
The plan is to share it on GitHub as a set of Ecore metamodels and then go through several iteration of minor "releases" until we will find good solution. Then we can have it either as a separate Eclipse project or as a part of CDT - this may be decided later.
Regards,
AF
23.01.2020 23:35, Jonah Graham пишет:
P.S. Well, as we started the introduction session ... actually, my family name is Fedorov, from the Greek "Θεόδωρος " you may know it as the Latin "Theodor", that means "God-given" :)
I have no experience with emf modelling so I don't know it is capable enough. I also have no experience getting this in java classes. Help will be needed if we go this way. Learning EMF is now on my todo :-)
Is there some doc from the discussion from Alexander Federov and William Riley?
_______________________________________________ cdt-dev mailing list cdt-dev@xxxxxxxxxxx To change your delivery options, retrieve your password, or unsubscribe from this list, visit https://www.eclipse.org/mailman/listinfo/cdt-dev
_______________________________________________ cdt-dev mailing list cdt-dev@xxxxxxxxxxx To change your delivery options, retrieve your password, or unsubscribe from this list, visit https://www.eclipse.org/mailman/listinfo/cdt-dev
_______________________________________________ cdt-dev mailing list cdt-dev@xxxxxxxxxxx To change your delivery options, retrieve your password, or unsubscribe from this list, visit https://www.eclipse.org/mailman/listinfo/cdt-dev
