seems more appropriate, as it is the class for a single object
Also seems to be called that on the net (don't remember where the meta came from, but it's gone)
after some serious recursive thinking it now actually makes sense.
The key was to change the actual type of the class that the meta_class manages
For objects it's (still) ok just to change the instance_type, but since the class object exists and has type, when adding instance variables, that actual type has to change
Since some weeks, Parfait uses instance variables instead of generated attribute getters (that needed type)
This makes it possible to simplify the boot process, getting rid of separate boot Space and class versions.
It is still quite order dependent, but all "normal" ruby code, (less magic) so easier to understand.
Also moved all code that can never run at runtime into the adapter. This included Space and Object new, space as the space will only ever be created at compile time and object, since that is quite different at run-time (which is where i am working towards)
as they are just the type of the meta_class, that was relatively simple.
I feel this is what oo is meant to be, instance variables and methods for the objects, and since classes are objects, for them too.
Class variables seem like a design mistake, weird scoping rules and no data hiding (left as an exercise to the reader)
args and locals got inlined into message, forgot to delete then
ripples out due to type creation
small type class api change, more ripples, but also more consistent
the "old" way of generating compilers is now obsolete
we can use ruby code with mom macros to achieve the same
Three step wip
remove old builtin
fix tests (including adding necessary methods)
fixup and inclusion of builtin code to parfait
was using exit, since raise is not implemented. This was ambiguous as all programs exit.
Using :died as special kernel code and bending it, and reporting it in interpreter.
opted to hack require to be getting on
need require for the test helper
and the files in lib/parfait
General require mechanism would still be ok, but require_ralative means implementing file source, which needs to be dragged around. I'll make an issue
Compiler now removes the module Parfait scope
and also the ::Parfait:: Scope in module names
Which means we can compile scoped code
and get unscoped code. for Parfait
Handy for tests too
I call it macro because it lets you insert basically arbitrary risc code into the ruby level. The way it works:
Reserve namespace X
map any X.some_call to a Mom instruction
by the name SomeCall
which must take the same args in constructor as given
And obviously produce whatever risc it wants
Hoping to rewrite builtin around this idea (with the existing Mom builtn instructions)
to make smaller binaries with larger integer heaps
also ran some benchmarks to see if it makes a difference
at least the binaries are smaller, calling also faster
Now we are using a statically linked list of messages. This will not work with procs, but that can be solved then.
Previous (wrong) thinking was that because of procs messages have to be allocated for every call. This was too slow, and not neccessary
Just like the args, locals are now inlined into the Message.
Message is off course bigger, but as they are created at compile time, that hardly matters
Some programs did get somewhat smaller, especially with both changes, but not super much
Super is a statement, a send really.
Not an expression (as maybe in c++)
The actual implementation will be a bit tricky, like raise, a bit of stack walking, but not impossible. Still, later
When the lambda is passed as argument, it must be moved. This triggers the generation of a corresponding parfait object (as before, and as for other constants) but now also triggers the code build. The code being the constant as it were
Also some more name fixes from renames
The parser presents the whole call that defines the block as a block. And so does the Ruby layer, as we don't want to do processing in ast.
Just making it clearer, also Vool:: block will have to be renamed
previous commit affected rather many test, as the implicit returns add extra instructions
Also added some explicit returns, so as not to test the return logic too much. return (ie return nl) is a knonwn 3 risc operation.
Parfait was depending on it, ie it created a dependency out of Parfait. But Parfiat needs to be self contained.
Moved 2 methods into parfait object, and resolved single call from text_writer to third.
Same same, just have to remembe to actually execute the condition if it is a send
Having send a possible expression, removes one tmp variable and associated move, for a little extra work.
Next return and assign (rest)
Before, when the type was determined, it was assumed that the method can be resolved. But off course tis is not true, as methods may be defined later in the file.
Two solutions for that. One could (and should) define all methods and only then start to compile. Thus having the type safety.
Or (as now) make a dynamic call and let it fail at runtime.
After having over 600 failing tests at one point, this does feel good.
Even better, most of the risc/interpreter tests where i didn't change anything came gree without changing the tests. ie we have binary compatibility.
Since Builtin generates risc, just like mom instructions, it was a design mistake to put builtin into risc in the first place. Now that borders are coming more into focus, it make much more sense to have the builtin in mom.
In fact the instructions should be moved out and a seperate invocation mechanism used , so functions can be parsed, not generated (wip)
start at Object get_interna_word
using the pattern to replace the whole risc method with a single mom instruction. Copying the original risc code into the instrucitons to_risc
also adding some very basic tests
Parfait classes must be unscoped. Now we start parsing Parfait, it must be without the module.
Luckily module_eval makes this a breeze.
Also remove string interpolation that is not yet processed
last orrurences in syscalls replaced, 2 variants
- exit does not actually need the int
- all else, preallocate the int beforehand and in syscall assume the reg name (integer_tmp)
test missing
Also pass the source into the compile method.
This way compiler can be reused for subsequent compile.
Does remove some double boots, but no major time save
Wherever space was loaded to get to the next_message
we now load the Message factory.
Otherwise much the same, only the attribute is next_object, not next_message
The binary is growing with 1k objects per factory, so i had to fix (hack) arm to handle bigger constants
close#14
slightly hacky, but in good tradition
previous implementation only worked until 16 significant bits, which is getting to little
this one just keeps adding more instructions to arrive at the constant by force
There are surely cleverer ways of doing this, ie by using the barrel shifter
A start on #15, admittedly a hack
also tried to keep the first around as was done in space
partially to not have to add the consrtants seperately
didn't work, as chain gets broken
also this has to be redone anyway
page was maybe a too low level name
pages may be the unit of the syscall, but after that objects desolve (maybe later to be added on from different pages)
Factory has the job of handing out a new instance of a type
it keeps a freelist for that and a reserve
create (load/reduce) the int once and transfer.
Save a cruicial 2 instructions
Also expanded the variable name possibilities with _self, __const , _1 and _2
now a variable has to be created before being used
thus it is save to develop contracts where a certain name
must exist in the scope
Maybe the syntax starts getting a bit weird, but at least the ! is a common symbol in ruby
these types are only needed to debug and can be gotten from the method (also in the mesage)
just saving the 6 instructions for every call
This was made possible through previous commits on fake_memory access
resolve once and reuse
also use non checking version of get/set
as index per definition must be ok
This creates some speedup, but mostly avoids some weird hang in Thread::Queue.pop for many seconds
unfortunately the writers have to have self.var =
otherwise it is just a local var
Also need to make the type explicit for all
Protocol included memory_length on the class for now
Before instance variables were used to store the actual data
While this worked it was a double mental loop and (more importantly) did not allow the same memory access
Ie, in the interpreter the types for args had to be set correctly
But they don't need to be until we walk the list, when we can get the types from the method.
In short, now the fake memory may be used as memory, by indexing into it
We now have to use attr reader/writers that map to the fake memory
And while this is possible, it means a lot more self. than i would like
Also the compiler whill have to undo those attr to instance acess at some point
as it was before blocks
(thought blocks would make reuse of messages impossible, but was wrong, this only appilies to lambdas)
(too) many tests affected
Torsten Rüger
Torsten Rüger
Torsten Ruger