| author |
Steve Losh <steve@stevelosh.com> |
| date |
Wed, 24 Aug 2016 15:43:20 +0000 |
| parents |
6ed3375e2921 |
| children |
fd638982ad83 |
(in-package #:temperance)
;;;; .-,--. . .
;;;; '|__/ ,-. ,-. ,-. ,-. ,-,-. ,-. . | ,-. |- . ,-. ,-.
;;;; ,| | |-' | | | | | | | | | | ,-| | | | | | |
;;;; `' ' `-' `-' `-' ' ' ' |-' ' `' `-^ `' ' `-' ' '
;;;; |
;;;; '
;;; Once we have a tokenized stream we can generate the machine instructions
;;; from it.
;;;
;;; We don't generate the ACTUAL bytecode immediately, because we want to run
;;; a few optimization passes on it first, and it's easier to work with if we
;;; have a friendlier format.
;;;
;;; So we turn a stream of tokens:
;;;
;;; (X2 = q/2), X1, X3, (X0 = p/2), X1, X2
;;;
;;; into a list of instructions, each of which is a list:
;;;
;;; (:put-structure X2 q 2)
;;; (:subterm-variable X1)
;;; (:subterm-variable X3)
;;; (:put-structure X0 p 2)
;;; (:subterm-value X1)
;;; (:subterm-value X2)
;;;
;;; The opcodes are keywords and the register arguments remain register objects.
;;; They get converted down to the raw bytes in the final "rendering" step.
;;;
;;; # Cut
;;;
;;; A quick note on cut (!): the book and original WAM do some nutty things to
;;; save one stack word per frame. They store the cut register for non-neck
;;; cuts in a "pseudovariable" on the stack, so they only have to allocate that
;;; extra stack word for things that actually USE non-neck cuts.
;;;
;;; We're going to just eat the extra stack word and store the cut register in
;;; every frame instead. This massively simplifies the implementation and lets
;;; me keep my sanity, and it MIGHT even end up being faster because there's
;;; one fewer opcode, less fucking around in the compiler, etc. But regardless:
;;; I don't want to go insane, and my laptop has sixteen gigabytes of RAM, so
;;; let's just store the damn word.
;;;
;;; # "Seen" Registers
;;;
;;; The book neglects to mention some REALLY important information about how you
;;; have to handle registers when compiling a stream of tokens. But if you've
;;; made it this far, you should be pretty used to the book omitting vital
;;; information. So hop in the clown car and take a ride with me.
;;;
;;; From the very beginning,the book mentions that certain instructions come in
;;; pairs, the first of which is used the first time the register is "seen" or
;;; "encountered", and the second used of which is used subsequent times.
;;;
;;; For example, a simple query like `p(A, A, A)` would result in:
;;;
;;; put-variable A0 X3
;;; put-value A1 X3
;;; put-value A2 X3
;;; call p/3
;;;
;;; This is all fine and dandy and works for single goals, but if you have
;;; a clause with MULTIPLE body goals you need to "reset" the list of
;;; already-seen registers after each goal. For example, consider:
;;;
;;; p() :-
;;; f(X, X),
;;; g(Y, Y).
;;;
;;; If you just apply what the book says without resetting the already-seen
;;; register list, you get:
;;;
;;; put-variable A0 X2
;;; put-value A1 X2
;;; call f/2
;;; put-value A0 X2 <--- wrong!
;;; put-value A1 X2
;;; call g/2
;;;
;;; But the variable in `g/2` is DIFFERENT than the one used in `f/2`, so that
;;; second `put-value` instruction is wrong! What we need instead is this:
;;;
;;; put-variable A0 X2
;;; put-value A1 X2
;;; call f/2
;;; put-variable A0 X2 <--- right!
;;; put-value A1 X2
;;; call g/2
;;;
;;; So the list of seen registers needs to get cleared after each body goal.
;;;
;;; But be careful: it's only TEMPORARY registers that need to get cleared! If
;;; the variables in our example WEREN'T different (`p() :- f(X, X), g(X, X)`)
;;; the instructions would be assigning to stack registers, and we WANT to do
;;; one `put-variable` and have the rest be `put-value`s.
;;;
;;; And there's one more edge case you're probably wondering about: what happens
;;; after the HEAD of a clause? Do we need to reset? The answer is: no,
;;; because the head and first body goal share registers, which is what performs
;;; the "substitution" for the first body goal (see the comment earlier for more
;;; on that rabbit hole).
(defun find-opcode-register (first-seen register)
(let ((register-variant (when register
(ecase (register-type register)
((:local :argument) :local)
((:permanent) :stack)
((:anonymous) :void)))))
(if first-seen
(ecase register-variant
(:local :subterm-variable-local)
(:stack :subterm-variable-stack)
(:void :subterm-void))
(ecase register-variant
(:local :subterm-value-local)
(:stack :subterm-value-stack)
(:void :subterm-void)))))
(defun find-opcode-list (mode)
(ecase mode
(:program :get-list)
(:query :put-list)))
(defun find-opcode-lisp-object (mode)
(ecase mode
(:program :get-lisp-object)
(:query :put-lisp-object)))
(defun find-opcode-structure (mode)
(ecase mode
(:program :get-structure)
(:query :put-structure)))
(defun find-opcode-argument (first-seen mode register)
(let ((register-variant (ecase (register-type register)
((:local :argument) :local)
((:permanent) :stack))))
(if first-seen
(ecase mode
(:program (ecase register-variant
(:local :get-variable-local)
(:stack :get-variable-stack)))
(:query (ecase register-variant
(:local :put-variable-local)
(:stack :put-variable-stack))))
(ecase mode
(:program (ecase register-variant
(:local :get-value-local)
(:stack :get-value-stack)))
(:query (ecase register-variant
(:local :put-value-local)
(:stack :put-value-stack)))))))
(defun precompile-tokens (head-tokens body-tokens)
"Generate a series of machine instructions from a stream of head and body
tokens.
The `head-tokens` should be program-style tokens, and are compiled in program
mode. The `body-tokens` should be query-style tokens, and are compiled in
query mode.
Actual queries are a special case where the `head-tokens` stream is `nil`
The compiled instructions will be returned as a circle.
"
(let ((seen (list))
(mode nil)
(instructions (make-empty-circle)))
(labels
((push-instruction (&rest instruction)
(circle-insert-end instructions instruction))
(reset-seen ()
;; Reset the list of seen registers (grep for "clown car" above)
(setf seen (remove-if #'register-temporary-p seen)))
(handle-argument (argument-register source-register)
(if (register-anonymous-p source-register)
(ecase mode
;; Query terms need to put an unbound var into their argument
;; register for each anonymous variable.
(:query (push-instruction :put-void argument-register))
;; Crazy, but for program terms we can just drop
;; argument-position anonymous variables on the floor.
(:program nil))
;; OP X_n A_i
(let ((first-seen (push-if-new source-register seen :test #'register=)))
(push-instruction
(find-opcode-argument first-seen mode source-register)
source-register
argument-register))))
(handle-structure (destination-register functor arity)
;; OP functor reg
(push destination-register seen)
(push-instruction (find-opcode-structure mode)
functor
arity
destination-register))
(handle-list (register)
(push register seen)
(push-instruction (find-opcode-list mode)
register))
(handle-lisp-object (register object)
;; OP object register
(push register seen)
(push-instruction (find-opcode-lisp-object mode) object register))
(handle-cut ()
(push-instruction :cut))
(handle-procedure-call (functor arity is-jump)
(if (and (eq functor 'call)
(= arity 1))
;; DYNAMIC-[CALL/JUMP]
(push-instruction (if is-jump :dynamic-jump :dynamic-call))
;; [CALL/JUMP] functor
(push-instruction (if is-jump :jump :call) functor arity))
;; This is a little janky, but at this point the body goals have been
;; turned into one single stream of tokens, so we don't have a nice
;; clean way to tell when one ends. But in practice, a body goal is
;; going to end with a CALL instruction, so we can use this as
;; a kludge to know when to reset.
;;
;; TODO: We should probably dekludge this by emitting an extra "end
;; body goal" token, especially once we add some special forms that
;; might need to do some resetting but not end in a CALL.
(reset-seen))
(handle-register (register)
(if (register-anonymous-p register)
;; VOID 1
(push-instruction (find-opcode-register nil register) 1)
;; OP reg
(let ((first-seen (push-if-new register seen :test #'register=)))
(push-instruction
(find-opcode-register first-seen register)
register))))
(handle-token (token)
(etypecase token
(argument-variable-token
(handle-argument (token-register token)
(token-target token)))
(structure-token
(handle-structure (token-register token)
(token-functor token)
(token-arity token)))
(list-token
(handle-list (token-register token)))
(lisp-object-token
(handle-lisp-object (token-register token)
(token-object token)))
(cut-token
(handle-cut))
(jump-token
(handle-procedure-call (token-functor token)
(token-arity token)
t))
(call-token
(handle-procedure-call (token-functor token)
(token-arity token)
nil))
(register-token
(handle-register (token-register token)))))
(handle-stream (tokens)
(map nil #'handle-token tokens)))
(when head-tokens
(setf mode :program)
(handle-stream head-tokens))
(setf mode :query)
(handle-stream body-tokens)
instructions)))
(defun precompile-clause (head body)
"Precompile the clause.
`head` should be the head of the clause for program clauses, or `nil` for
query clauses.
`body` is the body of the clause, or `nil` for facts.
Returns a circle of instructions and the properties of the clause.
"
(let* ((clause-props
(determine-clause-properties head body))
(head-tokens
(when head
(tokenize-program-term head clause-props)))
(clause-type
(cond ((null head) :query)
((null body) :fact)
((null (rest body)) :chain)
(t :rule)))
(body-tokens
(when body
(loop
:with first = t
:for (goal . remaining) :on body
:append
(if (eq goal '!) ; gross
;; cut just gets emitted straight, but DOESN'T flip `first`...
;; TODO: fix the cut layering violation here...
(list (make-instance 'cut-token))
(prog1
(tokenize-query-term
goal clause-props
:in-nead first
;; For actual WAM queries we're running, we don't want to
;; LCO the final CALL because we need that stack frame
;; (for storing the results).
:is-tail (and (not (eq clause-type :query))
(null remaining)))
(setf first nil)))))))
(let ((instructions (precompile-tokens head-tokens body-tokens))
(variable-count (length (clause-permanent-vars clause-props))))
;; We need to compile facts and rules differently. Facts end with
;; a PROCEED and rules are wrapped in ALOC/DEAL.
(ecase clause-type
(:chain
;; Chain rules don't need anything at all. They just unify, set up
;; the next predicate's arguments, and JUMP. By definition, in a chain
;; rule all variables must be temporary, so we don't need a stack frame
;; at all!
nil)
(:rule ; a full-ass rule
;; Non-chain rules need an ALLOC at the head and a DEALLOC right before
;; the tail call:
;;
;; ALLOC n
;; ...
;; DEAL
;; JUMP
(circle-insert-beginning instructions `(:allocate ,variable-count))
(circle-insert-before (circle-backward instructions) `(:deallocate)))
(:fact
(circle-insert-end instructions `(:proceed)))
(:query
;; The book doesn't have this ALOC here, but we do it to aid in result
;; extraction. Basically, to make extracting th results of a query
;; easier we allocate all of its variables on the stack, so we need
;; push a stack frame for them before we get started. We don't DEAL
;; because we want the frame to be left on the stack at the end so we
;; can poke at it.
(circle-insert-beginning instructions `(:allocate ,variable-count))
(circle-insert-end instructions `(:done))))
(values instructions clause-props))))
(defun precompile-query (query)
"Compile `query`, returning the instructions and permanent variables.
`query` should be a list of goal terms.
"
(multiple-value-bind (instructions clause-props)
(precompile-clause nil query)
(values instructions
(clause-permanent-vars clause-props))))
(defun find-predicate (clause)
"Return the functor and arity of the predicate of `clause`."
;; ( (f ?x ?y) | head ||| clause
;; (foo ?x) || body |||
;; (bar ?y) ) || |||
(let ((head (car clause)))
(etypecase head
(null (error "Clause ~S has a NIL head." clause))
(symbol (values head 0)) ; constants are 0-arity
(cons (values (car head) ; (f ...)
(1- (length head))))
(t (error "Clause ~S has a malformed head." clause)))))
(defun precompile-rules (rules)
"Compile a single predicate's `rules` into a list of instructions.
All the rules must for the same predicate. This is not checked, for
performance reasons. Don't fuck it up.
Each rule in `rules` should be a clause consisting of a head term and zero or
more body terms. A rule with no body is called a fact.
Returns the circle of compiled instructions, as well as the functor and arity
of the rules being compiled.
"
(assert rules () "Cannot compile an empty program.")
(multiple-value-bind (functor arity) (find-predicate (first rules))
(values
(if (= 1 (length rules))
;; Single-clause rules don't need to bother setting up a choice point.
(destructuring-bind ((head . body)) rules
(precompile-clause head body))
;; Otherwise we need to loop through each of the clauses, pushing their
;; choice point instruction first, then their actual code.
;;
;; The `nil` clause addresses will get filled in later, during rendering.
(loop :with instructions = (make-empty-circle)
:for ((head . body) . remaining) :on rules
:for first-p = t :then nil
:for last-p = (null remaining)
:for clause-instructions = (precompile-clause head body)
:do (progn
(circle-insert-end
instructions
(cond (first-p `(:try ,+choice-point-placeholder+))
(last-p `(:trust))
(t `(:retry ,+choice-point-placeholder+))))
(circle-append-circle instructions clause-instructions))
:finally (return instructions)))
functor
arity)))