arl.org: Lots of thinking

arl.org

@@ -13,20 +13,19 @@ world!".  Should be relatively straightforward.
 Before we get into generating C code and then compiling it, it might
 be worth translating the parsed ARL code into a generic IR.
 
-The IR should be much more primitive in its semantics, and force clear
+The IR should be primitive in its semantics but should still
-requirements of the platform we're compiling to.  This way, at the
+encapsulate the intention behind the original ARL code.  This should
-code generator stage we can figure out:
+allow us to find a set of minimum requirements for target compilation:
 - what can we reasonably use from the target platform to satisfy
-  requirements?
+  supporting the primitive IR?
 - what do we need to hand-roll on the target in order to make this
   work?
 
 Essentially, we want to write a virtual machine, and translate ARL
 code into bytecode for that VM.  Goals:
-- Easier to optimise IR bytecode than the AST of our original program
+- Type checking
-- Easier to imagine translations from that IR bytecode into target
+- Optimiser (stretch)
-  platform code
+
-*** TODO Minimal IR representation
 We need the following clear items in our IR:
 - Static type values
 - Static type variables (possible DeBrujin numbering or other such
@@ -35,29 +34,117 @@ We need the following clear items in our IR:
 - Strongly typed primitive operators (numeric, strings, I/O) with
   packed arguments
 
-Read about [[https://en.wikipedia.org/wiki/Three-address_code][TAC]].
-*** TODO IR Compiler
 We should have a rough grouping between AST objects and this IR.  As
 ARL is Forth-like, we can use the stack semantics to generate this IR
-as we walk the AST in a linear manner.
+as we walk the AST in a linear manner.  In practice this should almost
+look like emulating a really small subset of the ARL language itself
+and executing the program in that small subset.
 
+Looking at how
+[[https://en.wikipedia.org/wiki/Three-address_code][TAC]] works, I
+think it may be a good idea to do something like that for our IR.
+Essentially we should our AST into a sequence of really simple
+bindings, with the final expression being a reference to some binding.
+
+This also simplifies type checking to just verifying each little
+binding and operation.
+*** Examples
+**** Basic example
 Consider the following ARL code:
 #+begin_src text
 34 35 +
 #+end_src
 
-When we walk through this code:
+When we walk through the above code:
 - 34 (an integer) is pushed onto the stack
 - 35 (an integer) is pushed onto the stack
-- + is encountered
+- ~+~ primitive is encountered
-  - Pop two values off the stack and verify their type against the
+  - Type check the top two values of the stack; they should be
-    contract for "+" (something like (-> i32 i32 i32))
+    integral.
-  - Generate IR, something like ~prim-add(34, 35)~
+  - ~a b +~ should correspond to ~a + b~ so the IR expression should
-*** TODO Consider optimisers
+    pack the arguments in that order: ~prim-add(34,35)~.
-Certainly we should perform optimisations on the IR itself before
+  - Bind the generated IR expression to some unique name, say ~v1~.
-passing it over to the code generator.  Currently we haven't got much
+    - Ensure this works with type checking; looking up ~v1~'s type
-in the way of optimisations to consider, but it may be worth
+      should give you the output type of the "+" operator (integer).
-considering.
+  - Push ~v1~ onto the stack.
+
+The final state of the stack should be something like ~[v1]~ where
+~v1=prim-add(34,35)~.  The final state of the stack, along with the
+bindings we form, is the IR, to pass over to the later stages of the
+compiler.
+**** Slightly more complex example
+Let's look at a slightly more complex program:
+#+begin_src text
+34 35 + 70 swap -
+#+end_src
+- 34 (integer) pushed
+- 35 (integer) pushed
+- ~+~ primitive:
+  - As stated previously, the final state of this primitive gives us
+    the name ~v1~ on the stack with the association
+    ~v1=prim-add(34,35)~.
+- 70 (integer) pushed
+- ~swap~ primitive:
+  - Requires two values on the stack, but we care little about their
+    types.  Just swaps their order on the stack.
+  - We /could/ introduce generics here to make the input/output
+    relation ship explicit (forall T, U swap:-(-> (T U) (U T))), but
+    at the same time we can just as easily get away with a type hole
+    (essentially some kind of ~any~).  Up to debate.
+  - We do not generate IR for this primitive as it simply isn't
+    necessary.  Instead we perform the swap on our IR stack and
+    continue.  The ~swap~ primitive is "transparent" in the final IR.
+  - In this situation, the stack goes from ~[v1, 70]~ to
+    ~[70, v1]~
+- ~-~ primitive:
+  - Type checks the top two values of the stack (which are both
+    integers)
+  - ~a b -~ should correspond to ~a - b~, thus the corresponding IR
+    expression should be ~prim-sub(70,v1)~
+  - Associate IR expression with name ~v2~,
+  - Push ~v2~ onto the stack.
+
+The final state of the IR should be:
+- Stack: ~[v2]~
+- Bindings:
+  - ~v1=prim-add(34,35)~
+  - ~v2=prim-sub(70,v1)~
+
+Notice how some primitives generate IR, while others manipulate IR
+themselves?  They almost seem like macros!
+
+Another thing of note is how the final state of the stack is a single
+item in this case; an IR expression representing the entire program.
+When we introduce code level bindings we won't have such nice outputs,
+but it is certainly something to consider.
+**** Hello world! example
+For our hello world:
+#+begin_src text
+"Hello, world!\n" putstr
+#+end_src
+- "Hello, world!\n" (string) pushed
+- "putstr" primitive:
+  - Type check the top of the stack (should be a string)
+  - Generate IR ~prim-putstr("Hello, world!\n")~
+  - Associate with name ~v1~ and push it onto the stack
+
+Much simpler than our
+*** TODO IR level type checking
+During IR compilation, the following should be type checked:
+- use of callables (primitives, user defined when implemented)
+- variable assignment (when implemented)
+- variable use (when implemented)
+- definition of callables (when implemented)
+
+We want to ensure no statement is unsound.
+**** TODO Primitive types
+Define the primitive types of the IR.  Remember, simplicity is key,
+but we need to mirror what we're getting on the ARL side.
+**** TODO Type contracts for callables
+Define how we can type check arguments on the stack against the types
+a callable expects for its inputs.  In the same vein, we also need to
+figure out the type of whatever is pushed onto the stack by the
+callable.
 ** TODO Code generator
 [[file:src/arl/target-c/]]