50da6a40f2
also needs a branch in register. This way the register level is self sufficient (noticed while writing debugger) |
||
---|---|---|
lib | ||
stash | ||
test | ||
.gitignore | ||
CNAME | ||
CodeStyle.md | ||
FEATURES.md | ||
Gemfile | ||
Gemfile.lock | ||
LICENSE.txt | ||
Rakefile | ||
README.md | ||
salama.gemspec | ||
ToDo.md |
Salama
Salama is about native code generation in and of ruby.
It is probably best to read the The Book first.
Currently the work is to get the system to bootstrap, ie produce and executable that does what the ruby code says.
Later that executable needs to extend itself, but to do so it will be able to use the same code.
Done
Some things that are finished (for a definition of finished, ie started)
Assembly
Produce binary that represents code. Traditionally called assembling, but there is no need for an external file representation.
Most instructions are in fact assembling correctly. Meaning i have tests, and i can use objdump to verify the correct assembler code is disassembled
I even polished the dsl and so (from the tests), this is a valid hello world:
hello = "Hello World\n"
@program.main do
mov r7, 4 # 4 == write
mov r0 , 1 # stdout
add r1 , pc , hello # address of "hello World"
mov r2 , hello.length
swi 0 #software interupt, ie kernel syscall
mov r7, 1 # 1 == exit
swi 0
end
write(7 + hello.length/4 + 1 , 'hello')
Linking
Package the code into an executable, currently elf, and very simple at that.
Above Hello World can be linked and run. And will say its thing.
There is no way to link c code currently and not planned either, for some time (see next)
Syscalls
Some small portion of what libc usually provides is needed even right at the beginning. Mainly file open and read, exit, that kind of thing. Looking at libc implementations and kernel "api" docs, this is quite simple to do.
As said, "Hello world" comes out and does use syscall 4. Also the program stops by syscall exit. The full list is on the net and involves mostly just work.
Parse ruby
Parse simple code, using Parslet. This has been separated out as it's own gem, salama-reader.
Parsing is a surprisingly fiddly process, very space and order sensitive. But Parslet is great and simple expressions (including function definitions and calls) are starting to work.
I spent some time on the parse testing framework, so it is safe to fiddle and add. In fact it is very modular and so ot is easy to add.
Virtual: Compile the Ast
Since we now have an Abstact syntax tree, it needs to be compiled to a virtual machine Instruction format. For the parsed subset that's almost done.
It took me a while to come up with a decent but simple machine model. I had tried to map straight to hardware but failed. The current Virtual directory represent a machine with basic oo features.
Instead of having more Layers to go from virtual to arm, i opted to have passes that go over the data structure and modify it.
This allows optimization after every pass as we have a data structure at every point in time.
Parfait - the runtime
After an initial phase where i aimed for a really really small run-time, i have now started to implement a more decent set classes and functions. This is a process off course.
I was encouraged by the thought that a large amount of the run-time code can actually be reused at compile time, by using inlining. That off course assumes that i figure out how to do inlining, but i have at least an idea.
Sof
Salama Object File format is a yaml like format to look at code dumps and help testing. The dumper is ok and does produce (as intended) considerably denser dumps than yaml
When a reader is done (not started) the idea is to use sof as pre-compiled, language independent exchange format, have the core read that, and use the mechanism to achieve language independence.
Status - Dynmaic function lookup
It proved to be quite a big step to go from static function calling to oo method lookup. Also ruby is very introspective and that means much of the compiled code needs to be accessible in the runtime (not just present, accessible).
This has taken me the better part of a year, but is starting to come around.
So the current status is that i can
- parse a usable subset of ruby
- compile that to my vm model
- generate assembler for all higher level constructs in the vm model
- assemble and link the code and objects (strings/arrays/hashes) into an executable
- run the executable and debug :-(
PS: the current current status (05/15) is that even that is broken as i am reworking the run-time. But i am hopefull.
Future
Blocks
Implement ruby Blocks, and make new vm classes to deal with that. This is in fact a little open, but i have a general notion that blocks are "just" methods with even more implicit arguments.
Exceptions
Implement Exceptions. Conceptually this is not so difficult in an oo machine as it would be in c.
I have a post about it
which boils down to the fact that we can treat the address to return to in an exception quite like a return address from a function. Ie just another implicit parameter (as return is really an implicit parameter, a little like self for oo)
C linking
Implement a way to call libc and other c libraries. I am not placing a large emphasis on this personally, but expect somebody will come along and have library they want to use so much they can't stop themselves. Personally i think a fresh start is what we need much more. I once counted the call chain from a simple printf to the actual kernel invocation in some libc, and it was getting to 10! I hope with dynamic (re)compiling and intelligent inlining, we can do better than that.
Stary sky
Iterate:
- more cpus (ie intel)
- more systems (ie mac)
- more syscalls, there are after all some hundreds
- Ruby is full of niceties that are not done, also negative tests are non existant
- A lot of modern cpu's functionality has to be mapped to ruby and implemented in assembler to be useful
- Different sized machines, with different register types ?
- on 64bit, there would be 8 bits for types and thus allow for rational, complex, and whatnot
- Housekeeping (the superset of gc) is abundant
- Any amount of time could be spent on a decent digital tree (see judy). Or possibly Dr.Cliffs hash.
- Also better string/arrays would be good.
- The minor point of threads and hopefully lock free primitives to deal with that.
- Other languages, python at least, maybe others
- translation of the vm instructions to another vm, say js
And generally optimize and work towards that perfect world (we never seem to be able to attain).
Contributing to salama
Probably best to talk to me, if it's not a typo or so.
I do have a todo, for the adventurous.
Fork and create a branch before sending pulls.
== Copyright
Copyright (c) 2014 Torsten Ruger. See LICENSE.txt for further details.