[Vm-dev] Overview of the OpenSmalltalk VM and its build process

Tue Oct 17 17:03:56 UTC 2017

2017-10-17 16:00 GMT+02:00 Clément Bera <bera.clement at gmail.com>:

>
> Hi Raffaello,
>
> Answers inlined...
>
> On Tue, Oct 17, 2017 at 3:07 PM, Raffaello Giulietti <
> raffaello.giulietti at lifeware.ch> wrote:
>
>>
>> Hi,
>>
>> I'm interested in understanding the overall OpenSmalltalk VM structure
>> and design and its build process.
>>
>> I read [1] and [2] to get a high level picture of the vm and its
>> interesting strategy to development (Slang, VMMaker, etc.)
>>
>> Apart from these, are there other useful documents/sites/tutorials on the
>> vm development process in general and about the underlying ideas in the
>> build process? I'm aware that vm construction is a rather special topic but
>> any useful reference or hint is welcome.
>>
>> I'm already familiar with the interesting [3] and [4] websites and their
>> authors papers, talks and slides, but these do not seem to help much in
>> clarifying the development process itself.
>>
>> In particular, I'm targeting Pharo 64 bit/Win 64 bit (x64-86). I hope to
>> be able someday to actively participate in bringing Cog/Spur/Sista to a
>> more mature status on this platform.
>>
>> The questions I would like to find an answer to are of the following kind:
>> * Where should I start?
>>
>
> A good start would be to load the Squeak VMMaker dev image (See here:
> http://www.mirandabanda.org/cogblog/build-image/) and read the different
> opened workspaces. In addition, ...
>
> *Compilation toolchain*
>
> To check you have everything installed to compile, you can go to the one
> of the build folder (One of the windows build folder if you're on windows,
> for example build.Win64x64, in most case we don't recommend cross
> compilation), read the HowToBuild file, then go to one of the inner folder
> and use the mvm script to compile a VM from pre-generated C sources.
>

See below though, on windows, we target mingw so it's kind of
cross-compilation anyway.

> If it does not work, check on the openSmalltalk-VM repo how the VM is
> built using Travis CI and set-up your machine similarly to the Travis
> slaves.
>
> *Changes in Slang (GC/Interpreter/JIT/...)*
>
> In general when you want to change something in the GC / Interpreter the
> process is as follow:
>
> 1) Make your changes work with the StackVM (interpreter-only) simulator
> 2) Make your changes work with the compiled StackVM (interpreter-only)
> 3) Make your changes work with the CogVM simulator
> 4) Make your changes work with the compiled CogVM
>
> In general when you want to change something in the JIT the process is as
> follow:
>
> 1) Make your changes generate correct in-image machine code
> 2) Make your changes work with the CogVM simulator
> 3) Make your changes work with the compiled CogVM
>
> So a good place to start is to have the simulators working. You could use
> this to help:
> https://clementbera.wordpress.com/2016/05/30/simulating-the-cog-vm/
>
> Once the simulator works with your changes, you can generate C sources
> with something like "VMMaker generateConfiguration"
>
> *Changes in Platform code*
>
> Go to the platform folder at the root and change the C files with a text
> editor
>
>
>> * Should I use Squaek or can I use Pharo for vm development purposes?
>>
>
> Currently I still use Squeak but I plan to migrate to Pharo. I'd say it's
> still a bit easier from Squeak due to package compatibility making it hard
> to commit from Pharo without uncommitting something else, some scripts are
> also present by default in the Squeak dev image, but everything should work
> on both environments. No big deal in using one of the other.
>
>
>> * What are the restrictions of Slang?
>>
>
> ...
> As long as the VM compiles and run, this is correct Slang code, else it's
> not.
> In general you should just think in C and write Smalltalk instead.
> I don't know what to say. This is not dynamic at all. No dictionaries, no
> look-up, no nothing. Even arrays look like Smalltalk arrays but like C
> arrays you don't have the actual size.
> Blocks work as long as they're inlined and removed when performing
> inlining in the Slang to C compiler.
>
> * Where does VMMaker live?
>>
>
> Half there:
> https://github.com/OpenSmalltalk/opensmalltalk-vm
> Accessible from git
>
> Half there:
> http://source.squeak.org/VMMaker/
> Accessible from Monticello
>
> * Which C toolchain shall I use?
>
> * Which release of gcc, ld, make are known to work?
>> * What about clang?
>
>
> Depends on the platform. The mvm script normally chooses for you.
>
> All VMs are built from Travis CI so you can check how to set up your
> machine from the Travis set-up.
>
> On linux it's gcc by default. All recent gcc versions should work since 4.4
> On Mac, gcc is not supported anymore, so clang is used. I've used many
> different versions of clang and they all worked.
>
> Specific compiler versions that don't work are mentioned in the HowToBuild
> file in the platform folders (if any).
>
>
>> * Which version of cygwin64 or mingw64 are OK?
>>
>
> Arf. Both worked at some point, I don't know what the default script use
> right now I would start with the version used by the mvm script / the
> Travis build.
>
>
gcc is broken for win64 (or at least not satisfying our expectations about
stack management)
So we have to use clang.
For this reason, we prefer to compile thru cygwin which has prebuilt
(downlodable) clang modules.
It's much more difficult to get a working clang in a mingw environment, so
this brand is not maintained anymore.

But the target is a mingw target, not a cygwin target (we don't want VM to
depend on cygwin layer, we want to stay minimal).
So we use a cross compiler. See the line

TOOLPREFIX:=x86_64-w64-mingw32-

in
https://github.com/OpenSmalltalk/opensmalltalk-vm/blob/Cog/build.win64x64/common/Makefile.tools

Using cross-compiler means that we can eventually build a x86_64 .exe from
a cygwin32.
But cygwin32 does not provide a debugger (gdb) for x86_64, that's why we
prefer developing from a cygwin64 environment.

Eventually, I have a hand-assembled visual studio project for compiling the
vm with MSVC (2015 if I remember...)
Of course that kind of project should better be generated (as was possible
via cmake) rather than hand-assembled...
It's absolutely cumbersome to maintain such .vc(x)proj across several
versions of MSVC.
Anyway, it does not produce working VM yet.
I retried with 2017 but encountered a code generation bug.

We could try to adapt the gnu makefiles to use MSVC thru command line, but
IMO the main interest of MSVC is to perform complementary analysis thru IDE.

>
>> * etc.
>>
>>
>>
>> Greetings
>> Raffaello
>>
>> ---
>>
>> [1] http://www.vpri.org/pdf/tr1997001_backto.pdf
>> [2] http://design.cs.iastate.edu/vmil/2011/papers/p03-miranda.pdf
>> [3] http://www.mirandabanda.org/cogblog/
>> [4] https://clementbera.wordpress.com/
>>
>
>
>
> --
> Clément Béra
> Pharo consortium engineer
> https://clementbera.wordpress.com/
> Bâtiment B 40, avenue Halley 59650 Villeneuve d'Ascq
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.squeakfoundation.org/pipermail/vm-dev/attachments/20171017/0a152a61/attachment.html>