1 ;;; Tree-IL partial evaluator
3 ;; Copyright (C) 2011, 2012, 2013 Free Software Foundation, Inc.
5 ;;;; This library is free software; you can redistribute it and/or
6 ;;;; modify it under the terms of the GNU Lesser General Public
7 ;;;; License as published by the Free Software Foundation; either
8 ;;;; version 3 of the License, or (at your option) any later version.
10 ;;;; This library is distributed in the hope that it will be useful,
11 ;;;; but WITHOUT ANY WARRANTY; without even the implied warranty of
12 ;;;; MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 ;;;; Lesser General Public License for more details.
15 ;;;; You should have received a copy of the GNU Lesser General Public
16 ;;;; License along with this library; if not, write to the Free Software
17 ;;;; Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
19 (define-module (language tree-il peval)
20 #:use-module (language tree-il)
21 #:use-module (language tree-il primitives)
22 #:use-module (language tree-il effects)
23 #:use-module (ice-9 vlist)
24 #:use-module (ice-9 match)
25 #:use-module (srfi srfi-1)
26 #:use-module (srfi srfi-9)
27 #:use-module (srfi srfi-11)
28 #:use-module (srfi srfi-26)
29 #:use-module (ice-9 control)
33 ;;; Partial evaluation is Guile's most important source-to-source
34 ;;; optimization pass. It performs copy propagation, dead code
35 ;;; elimination, inlining, and constant folding, all while preserving
36 ;;; the order of effects in the residual program.
38 ;;; For more on partial evaluation, see William Cook’s excellent
39 ;;; tutorial on partial evaluation at DSL 2011, called “Build your own
40 ;;; partial evaluator in 90 minutes”[0].
42 ;;; Our implementation of this algorithm was heavily influenced by
43 ;;; Waddell and Dybvig's paper, "Fast and Effective Procedure Inlining",
44 ;;; IU CS Dept. TR 484.
46 ;;; [0] http://www.cs.utexas.edu/~wcook/tutorial/.
49 ;; First, some helpers.
51 (define-syntax *logging* (identifier-syntax #f))
53 ;; For efficiency we define *logging* to inline to #f, so that the call
54 ;; to log* gets optimized out. If you want to log, uncomment these
57 ;; (define %logging #f)
58 ;; (define-syntax *logging* (identifier-syntax %logging))
60 ;; Then you can change %logging at runtime.
66 (or (eq? *logging* #t)
67 (memq 'event *logging*)))
68 (log* 'event arg ...)))))
70 (define (log* event . args)
71 (let ((pp (module-ref (resolve-interface '(ice-9 pretty-print))
73 (pp `(log ,event . ,args))
77 (define (tree-il-any proc exp)
79 (tree-il-fold (lambda (exp res)
80 (let ((res (proc exp)))
85 (define (vlist-any proc vlist)
86 (let ((len (vlist-length vlist)))
89 (or (proc (vlist-ref vlist i))
92 (define (singly-valued-expression? exp)
95 (($ <lexical-ref>) #t)
97 (($ <lexical-ref>) #t)
98 (($ <primitive-ref>) #t)
100 (($ <toplevel-ref>) #t)
101 (($ <primcall> _ (? singly-valued-primitive?)) #t)
102 (($ <primcall> _ 'values (val)) #t)
104 (($ <conditional> _ test consequent alternate)
105 (and (singly-valued-expression? consequent)
106 (singly-valued-expression? alternate)))
109 (define (truncate-values x)
110 "Discard all but the first value of X."
111 (if (singly-valued-expression? x)
113 (make-primcall (tree-il-src x) 'values (list x))))
115 ;; Peval will do a one-pass analysis on the source program to determine
116 ;; the set of assigned lexicals, and to identify unreferenced and
117 ;; singly-referenced lexicals.
119 (define-record-type <var>
120 (make-var name gensym refcount set?)
124 (refcount var-refcount set-var-refcount!)
125 (set? var-set? set-var-set?!))
127 (define* (build-var-table exp #:optional (table vlist-null))
131 (($ <lexical-ref> src name gensym)
132 (let ((var (cdr (vhash-assq gensym res))))
133 (set-var-refcount! var (1+ (var-refcount var)))
135 (($ <lambda-case> src req opt rest kw init gensyms body alt)
136 (fold (lambda (name sym res)
137 (vhash-consq sym (make-var name sym 0 #f) res))
139 (append req (or opt '()) (if rest (list rest) '())
141 ((aok? (kw name sym) ...) name)
144 (($ <let> src names gensyms vals body)
145 (fold (lambda (name sym res)
146 (vhash-consq sym (make-var name sym 0 #f) res))
148 (($ <letrec> src in-order? names gensyms vals body)
149 (fold (lambda (name sym res)
150 (vhash-consq sym (make-var name sym 0 #f) res))
152 (($ <fix> src names gensyms vals body)
153 (fold (lambda (name sym res)
154 (vhash-consq sym (make-var name sym 0 #f) res))
156 (($ <lexical-set> src name gensym exp)
157 (set-var-set?! (cdr (vhash-assq gensym res)) #t)
160 (lambda (exp res) res)
163 ;; Counters are data structures used to limit the effort that peval
164 ;; spends on particular inlining attempts. Each call site in the source
165 ;; program is allocated some amount of effort. If peval exceeds the
166 ;; effort counter while attempting to inline a call site, it aborts the
167 ;; inlining attempt and residualizes a call instead.
169 ;; As there is a fixed number of call sites, that makes `peval' O(N) in
170 ;; the number of call sites in the source program.
172 ;; Counters should limit the size of the residual program as well, but
173 ;; currently this is not implemented.
175 ;; At the top level, before seeing any peval call, there is no counter,
176 ;; because inlining will terminate as there is no recursion. When peval
177 ;; sees a call at the top level, it will make a new counter, allocating
178 ;; it some amount of effort and size.
180 ;; This top-level effort counter effectively "prints money". Within a
181 ;; toplevel counter, no more effort is printed ex nihilo; for a nested
182 ;; inlining attempt to proceed, effort must be transferred from the
183 ;; toplevel counter to the nested counter.
185 ;; Via `data' and `prev', counters form a linked list, terminating in a
186 ;; toplevel counter. In practice `data' will be the a pointer to the
187 ;; source expression of the procedure being inlined.
189 ;; In this way peval can detect a recursive inlining attempt, by walking
190 ;; back on the `prev' links looking for matching `data'. Recursive
191 ;; counters receive a more limited effort allocation, as we don't want
192 ;; to spend all of the effort for a toplevel inlining site on loops.
193 ;; Also, recursive counters don't need a prompt at each inlining site:
194 ;; either the call chain folds entirely, or it will be residualized at
195 ;; its original call.
197 (define-record-type <counter>
198 (%make-counter effort size continuation recursive? data prev)
200 (effort effort-counter)
202 (continuation counter-continuation)
203 (recursive? counter-recursive? set-counter-recursive?!)
207 (define (abort-counter c)
208 ((counter-continuation c)))
210 (define (record-effort! c)
211 (let ((e (effort-counter c)))
212 (if (zero? (variable-ref e))
214 (variable-set! e (1- (variable-ref e))))))
216 (define (record-size! c)
217 (let ((s (size-counter c)))
218 (if (zero? (variable-ref s))
220 (variable-set! s (1- (variable-ref s))))))
222 (define (find-counter data counter)
224 (if (eq? data (counter-data counter))
226 (find-counter data (counter-prev counter)))))
228 (define* (transfer! from to #:optional
229 (effort (variable-ref (effort-counter from)))
230 (size (variable-ref (size-counter from))))
231 (define (transfer-counter! from-v to-v amount)
232 (let* ((from-balance (variable-ref from-v))
233 (to-balance (variable-ref to-v))
234 (amount (min amount from-balance)))
235 (variable-set! from-v (- from-balance amount))
236 (variable-set! to-v (+ to-balance amount))))
238 (transfer-counter! (effort-counter from) (effort-counter to) effort)
239 (transfer-counter! (size-counter from) (size-counter to) size))
241 (define (make-top-counter effort-limit size-limit continuation data)
242 (%make-counter (make-variable effort-limit)
243 (make-variable size-limit)
249 (define (make-nested-counter continuation data current)
250 (let ((c (%make-counter (make-variable 0)
256 (transfer! current c)
259 (define (make-recursive-counter effort-limit size-limit orig current)
260 (let ((c (%make-counter (make-variable 0)
262 (counter-continuation orig)
266 (transfer! current c effort-limit size-limit)
269 ;; Operand structures allow bindings to be processed lazily instead of
270 ;; eagerly. By doing so, hopefully we can get process them in a way
271 ;; appropriate to their use contexts. Operands also prevent values from
272 ;; being visited multiple times, wasting effort.
274 ;; TODO: Record value size in operand structure?
276 (define-record-type <operand>
277 (%make-operand var sym visit source visit-count use-count
278 copyable? residual-value constant-value alias-value)
282 (visit %operand-visit)
283 (source operand-source)
284 (visit-count operand-visit-count set-operand-visit-count!)
285 (use-count operand-use-count set-operand-use-count!)
286 (copyable? operand-copyable? set-operand-copyable?!)
287 (residual-value operand-residual-value %set-operand-residual-value!)
288 (constant-value operand-constant-value set-operand-constant-value!)
289 (alias-value operand-alias-value set-operand-alias-value!))
291 (define* (make-operand var sym #:optional source visit alias)
292 ;; Bind SYM to VAR, with value SOURCE. Unassigned bound operands are
293 ;; considered copyable until we prove otherwise. If we have a source
294 ;; expression, truncate it to one value. Copy propagation does not
295 ;; work on multiply-valued expressions.
296 (let ((source (and=> source truncate-values)))
297 (%make-operand var sym visit source 0 0
298 (and source (not (var-set? var))) #f #f
299 (and (not (var-set? var)) alias))))
301 (define* (make-bound-operands vars syms sources visit #:optional aliases)
303 (map (lambda (name sym source alias)
304 (make-operand name sym source visit alias))
305 vars syms sources aliases)
306 (map (lambda (name sym source)
307 (make-operand name sym source visit #f))
310 (define (make-unbound-operands vars syms)
311 (map make-operand vars syms))
313 (define (set-operand-residual-value! op val)
314 (%set-operand-residual-value!
317 (($ <primcall> src 'values (first))
318 ;; The continuation of a residualized binding does not need the
319 ;; introduced `values' node, so undo the effects of truncation.
324 (define* (visit-operand op counter ctx #:optional effort-limit size-limit)
325 ;; Peval is O(N) in call sites of the source program. However,
326 ;; visiting an operand can introduce new call sites. If we visit an
327 ;; operand outside a counter -- i.e., outside an inlining attempt --
328 ;; this can lead to divergence. So, if we are visiting an operand to
329 ;; try to copy it, and there is no counter, make a new one.
331 ;; This will only happen at most as many times as there are lexical
332 ;; references in the source program.
333 (and (zero? (operand-visit-count op))
336 (set-operand-visit-count! op (1+ (operand-visit-count op))))
338 (and (operand-source op)
339 (if (or counter (and (not effort-limit) (not size-limit)))
340 ((%operand-visit op) (operand-source op) counter ctx)
343 ;; If we abort when visiting the value in a
344 ;; fresh context, we won't succeed in any future
345 ;; attempt, so don't try to copy it again.
346 (set-operand-copyable?! op #f)
350 (make-top-counter effort-limit size-limit abort op)
353 (set-operand-visit-count! op (1- (operand-visit-count op)))))))
355 ;; A helper for constant folding.
357 (define (types-check? primitive-name args)
360 ((not pair? null? list? symbol? vector? struct?)
364 ;; FIXME: add more cases?
367 (define* (peval exp #:optional (cenv (current-module)) (env vlist-null)
369 (operator-size-limit 40)
370 (operand-size-limit 20)
371 (value-size-limit 10)
373 (recursive-effort-limit 100))
374 "Partially evaluate EXP in compilation environment CENV, with
375 top-level bindings from ENV and return the resulting expression."
377 ;; This is a simple partial evaluator. It effectively performs
378 ;; constant folding, copy propagation, dead code elimination, and
383 ;; Propagate copies across toplevel bindings, if we can prove the
384 ;; bindings to be immutable.
386 ;; Specialize lambda expressions with invariant arguments.
388 (define local-toplevel-env
389 ;; The top-level environment of the module being compiled.
391 (define (env-folder x env)
393 (($ <toplevel-define> _ name)
394 (vhash-consq name #t env))
395 (($ <seq> _ head tail)
396 (env-folder tail (env-folder head env)))
398 (env-folder exp vlist-null)))
400 (define (local-toplevel? name)
401 (vhash-assq name local-toplevel-env))
404 ;; renamed-term -> original-term
406 (define store (build-var-table exp))
408 (define (record-new-temporary! name sym refcount)
409 (set! store (vhash-consq sym (make-var name sym refcount #f) store)))
411 (define (lookup-var sym)
412 (let ((v (vhash-assq sym store)))
413 (if v (cdr v) (error "unbound var" sym (vlist->list store)))))
415 (define (fresh-gensyms vars)
417 (let ((new (gensym (string-append (symbol->string (var-name var))
419 (set! store (vhash-consq new var store))
423 (define (fresh-temporaries ls)
425 (let ((new (gensym "tmp ")))
426 (record-new-temporary! 'tmp new 1)
430 (define (assigned-lexical? sym)
431 (var-set? (lookup-var sym)))
433 (define (lexical-refcount sym)
434 (var-refcount (lookup-var sym)))
436 (define (with-temporaries src exps refcount can-copy? k)
437 (let* ((pairs (map (match-lambda
438 ((and exp (? can-copy?))
441 (let ((sym (gensym "tmp ")))
442 (record-new-temporary! 'tmp sym refcount)
445 (tmps (filter car pairs)))
450 (make-list (length tmps) 'tmp)
453 (k (map (match-lambda
456 (make-lexical-ref #f 'tmp sym)))
459 (define (make-begin0 src first second)
463 (let ((vals (gensym "vals ")))
464 (record-new-temporary! 'vals vals 1)
467 '() #f 'vals #f '() (list vals)
471 (make-primcall #f 'apply
473 (make-primitive-ref #f 'values)
474 (make-lexical-ref #f 'vals vals))))
477 ;; ORIG has been alpha-renamed to NEW. Analyze NEW and record a link
480 (define (record-source-expression! orig new)
481 (set! store (vhash-consq new (source-expression orig) store))
484 ;; Find the source expression corresponding to NEW. Used to detect
485 ;; recursive inlining attempts.
487 (define (source-expression new)
488 (let ((x (vhash-assq new store)))
491 (define (record-operand-use op)
492 (set-operand-use-count! op (1+ (operand-use-count op))))
494 (define (unrecord-operand-uses op n)
495 (let ((count (- (operand-use-count op) n)))
497 (set-operand-residual-value! op #f))
498 (set-operand-use-count! op count)))
500 (define* (residualize-lexical op #:optional ctx val)
501 (log 'residualize op)
502 (record-operand-use op)
503 (if (memq ctx '(value values))
504 (set-operand-residual-value! op val))
505 (make-lexical-ref #f (var-name (operand-var op)) (operand-sym op)))
507 (define (fold-constants src name args ctx)
508 (define (apply-primitive name args)
509 ;; todo: further optimize commutative primitives
514 (apply (module-ref the-scm-module name) args))
516 (values #t results))))
519 (define (make-values src values)
521 ((single) single) ; 1 value
522 ((_ ...) ; 0, or 2 or more values
523 (make-primcall src 'values values))))
524 (define (residualize-call)
525 (make-primcall src name args))
528 (let-values (((success? values)
529 (apply-primitive name (map const-exp args))))
530 (log 'fold success? values name args)
533 ((effect) (make-void src))
535 ;; Values truncation: only take the first
538 (make-const src (car values))
539 (make-values src '())))
541 (make-values src (map (cut make-const src <>) values))))
542 (residualize-call))))
543 ((and (eq? ctx 'effect) (types-check? name args))
546 (residualize-call))))
548 (define (inline-values src exp nmin nmax consumer)
549 (let loop ((exp exp))
551 ;; Some expression types are always singly-valued.
559 ($ <lexical-set>) ; FIXME: these set! expressions
560 ($ <toplevel-set>) ; could return zero values in
561 ($ <toplevel-define>) ; the future
563 ($ <primcall> src (? singly-valued-primitive?)))
564 (and (<= nmin 1) (or (not nmax) (>= nmax 1))
565 (make-call src (make-lambda #f '() consumer) (list exp))))
567 ;; Statically-known number of values.
568 (($ <primcall> src 'values vals)
569 (and (<= nmin (length vals)) (or (not nmax) (>= nmax (length vals)))
570 (make-call src (make-lambda #f '() consumer) vals)))
572 ;; Not going to copy code into both branches.
573 (($ <conditional>) #f)
575 ;; Bail on other applications.
579 ;; Bail on prompt and abort.
583 ;; Propagate to tail positions.
584 (($ <let> src names gensyms vals body)
585 (let ((body (loop body)))
587 (make-let src names gensyms vals body))))
588 (($ <letrec> src in-order? names gensyms vals body)
589 (let ((body (loop body)))
591 (make-letrec src in-order? names gensyms vals body))))
592 (($ <fix> src names gensyms vals body)
593 (let ((body (loop body)))
595 (make-fix src names gensyms vals body))))
596 (($ <let-values> src exp
597 ($ <lambda-case> src2 req opt rest kw inits gensyms body #f))
598 (let ((body (loop body)))
600 (make-let-values src exp
601 (make-lambda-case src2 req opt rest kw
602 inits gensyms body #f)))))
603 (($ <seq> src head tail)
604 (let ((tail (loop tail)))
605 (and tail (make-seq src head tail)))))))
607 (define compute-effects
608 (make-effects-analyzer assigned-lexical?))
610 (define (constant-expression? x)
611 ;; Return true if X is constant, for the purposes of copying or
612 ;; elision---i.e., if it is known to have no effects, does not
613 ;; allocate storage for a mutable object, and does not access
614 ;; mutable data (like `car' or toplevel references).
615 (constant? (compute-effects x)))
617 (define (prune-bindings ops in-order? body counter ctx build-result)
618 ;; This helper handles both `let' and `letrec'/`fix'. In the latter
619 ;; cases we need to make sure that if referenced binding A needs
620 ;; as-yet-unreferenced binding B, that B is processed for value.
621 ;; Likewise if C, when processed for effect, needs otherwise
622 ;; unreferenced D, then D needs to be processed for value too.
624 (define (referenced? op)
625 ;; When we visit lambdas in operator context, we just copy them,
626 ;; as we will process their body later. However this does have
627 ;; the problem that any free var referenced by the lambda is not
628 ;; marked as needing residualization. Here we hack around this
629 ;; and treat all bindings as referenced if we are in operator
631 (or (eq? ctx 'operator)
632 (not (zero? (operand-use-count op)))))
634 ;; values := (op ...)
635 ;; effects := (op ...)
636 (define (residualize values effects)
637 ;; Note, values and effects are reversed.
640 (let ((values (filter operand-residual-value ops)))
643 (build-result (map (compose var-name operand-var) values)
644 (map operand-sym values)
645 (map operand-residual-value values)
651 (let ((effect-vals (map operand-residual-value effects)))
652 (list->seq #f (reverse (cons body effect-vals)))))))
655 (let ((values (reverse values)))
656 (build-result (map (compose var-name operand-var) values)
657 (map operand-sym values)
658 (map operand-residual-value values)
662 ;; values := (op ...)
663 ;; effects := ((op . value) ...)
664 (let prune ((old (map referenced? ops)) (values '()) (effects '()))
665 (let lp ((ops* ops) (values values) (effects effects))
668 (let ((new (map referenced? ops)))
669 (if (not (equal? new old))
670 (prune new values '())
672 (map (lambda (op val)
673 (set-operand-residual-value! op val)
675 (map car effects) (map cdr effects))))))
677 (let ((op (car ops*)))
680 (lp (cdr ops*) values effects))
681 ((operand-residual-value op)
682 (lp (cdr ops*) (cons op values) effects))
684 (set-operand-residual-value! op (visit-operand op counter 'value))
685 (lp (cdr ops*) (cons op values) effects))
689 (let ((effect (visit-operand op counter 'effect)))
692 (acons op effect effects))))))))))))
694 (define (small-expression? x limit)
697 (lambda (x res) ; down
706 (define (extend-env sym op env)
707 (vhash-consq (operand-sym op) op (vhash-consq sym op env)))
710 (env vlist-null) ; vhash of gensym -> <operand>
711 (counter #f) ; inlined call stack
712 (ctx 'values)) ; effect, value, values, test, operator, or call
715 ((vhash-assq var env) => cdr)
716 (else (error "unbound var" var))))
718 ;; Find a value referenced a specific number of times. This is a hack
719 ;; that's used for propagating fresh data structures like rest lists and
720 ;; prompt tags. Usually we wouldn't copy consed data, but we can do so in
721 ;; some special cases like `apply' or prompts if we can account
722 ;; for all of its uses.
724 ;; You don't want to use this in general because it introduces a slight
725 ;; nonlinearity by running peval again (though with a small effort and size
728 (define (find-definition x n-aliases)
732 ((lookup (lexical-ref-gensym x))
734 (if (var-set? (operand-var op))
736 (let ((y (or (operand-residual-value op)
737 (visit-operand op counter 'value 10 10)
738 (operand-source op))))
740 ((and (lexical-ref? y)
741 (= (lexical-refcount (lexical-ref-gensym x)) 1))
742 ;; X is a simple alias for Y. Recurse, regardless of
743 ;; the number of aliases we were expecting.
744 (find-definition y n-aliases))
745 ((= (lexical-refcount (lexical-ref-gensym x)) n-aliases)
746 ;; We found a definition that is aliased the right
747 ;; number of times. We still recurse in case it is a
749 (values (find-definition y 1)
752 ;; We can't account for our aliases.
755 ;; A formal parameter. Can't say anything about that.
758 ;; Not a lexical: success, but only if we are looking for an
761 (else (values #f #f))))
763 (define (visit exp ctx)
764 (loop exp env counter ctx))
766 (define (for-value exp) (visit exp 'value))
767 (define (for-values exp) (visit exp 'values))
768 (define (for-test exp) (visit exp 'test))
769 (define (for-effect exp) (visit exp 'effect))
770 (define (for-call exp) (visit exp 'call))
771 (define (for-tail exp) (visit exp ctx))
774 (record-effort! counter))
776 (log 'visit ctx (and=> counter effort-counter)
777 (unparse-tree-il exp))
782 ((effect) (make-void #f))
786 ((test) (make-const #f #t))
788 (($ <lexical-ref> _ _ gensym)
789 (log 'begin-copy gensym)
790 (let ((op (lookup gensym)))
793 (log 'lexical-for-effect gensym)
795 ((operand-alias-value op)
796 ;; This is an unassigned operand that simply aliases some
797 ;; other operand. Recurse to avoid residualizing the leaf
801 ;; Don't propagate copies if we are residualizing a call.
802 (log 'residualize-lexical-call gensym op)
803 (residualize-lexical op))
804 ((var-set? (operand-var op))
805 ;; Assigned lexicals don't copy-propagate.
806 (log 'assigned-var gensym op)
807 (residualize-lexical op))
808 ((not (operand-copyable? op))
809 ;; We already know that this operand is not copyable.
810 (log 'not-copyable gensym op)
811 (residualize-lexical op))
812 ((and=> (operand-constant-value op)
813 (lambda (x) (or (const? x) (void? x) (primitive-ref? x))))
815 (let ((val (operand-constant-value op)))
816 (log 'memoized-constant gensym val)
818 ((visit-operand op counter (if (eq? ctx 'values) 'value ctx)
819 recursive-effort-limit operand-size-limit)
821 ;; If we end up deciding to residualize this value instead of
822 ;; copying it, save that residualized value.
825 ((not (constant-expression? val))
826 (log 'not-constant gensym op)
827 ;; At this point, ctx is operator, test, or value. A
828 ;; value that is non-constant in one context will be
829 ;; non-constant in the others, so it's safe to record
830 ;; that here, and avoid future visits.
831 (set-operand-copyable?! op #f)
832 (residualize-lexical op ctx val))
835 (primitive-ref? val))
836 ;; Always propagate simple values that cannot lead to
838 (log 'copy-simple gensym val)
839 ;; It could be this constant is the result of folding.
840 ;; If that is the case, cache it. This helps loop
841 ;; unrolling get farther.
842 (if (or (eq? ctx 'value) (eq? ctx 'values))
844 (log 'memoize-constant gensym val)
845 (set-operand-constant-value! op val)))
847 ((= 1 (var-refcount (operand-var op)))
848 ;; Always propagate values referenced only once.
849 (log 'copy-single gensym val)
851 ;; FIXME: do demand-driven size accounting rather than
854 ;; A pure expression in the operator position. Inline
855 ;; if it's a lambda that's small enough.
856 (if (and (lambda? val)
857 (small-expression? val operator-size-limit))
859 (log 'copy-operator gensym val)
862 (log 'too-big-for-operator gensym val)
863 (residualize-lexical op ctx val))))
865 ;; A pure expression, processed for call or for value.
866 ;; Don't inline lambdas, because they will probably won't
867 ;; fold because we don't know the operator.
868 (if (and (small-expression? val value-size-limit)
869 (not (tree-il-any lambda? val)))
871 (log 'copy-value gensym val)
874 (log 'too-big-or-has-lambda gensym val)
875 (residualize-lexical op ctx val)))))))
877 ;; Visit failed. Either the operand isn't bound, as in
878 ;; lambda formal parameters, or the copy was aborted.
879 (log 'unbound-or-aborted gensym op)
880 (residualize-lexical op)))))
881 (($ <lexical-set> src name gensym exp)
882 (let ((op (lookup gensym)))
883 (if (zero? (var-refcount (operand-var op)))
884 (let ((exp (for-effect exp)))
887 (make-seq src exp (make-void #f))))
889 (record-operand-use op)
890 (make-lexical-set src name (operand-sym op) (for-value exp))))))
893 (gensyms ... rest-sym)
894 (vals ... ($ <primcall> _ 'list rest-args))
895 ($ <primcall> asrc 'apply
898 (? (cut eq? <> rest))
900 (and (eq? sym rest-sym)
901 (= (lexical-refcount sym) 1))))))))
902 (let* ((tmps (make-list (length rest-args) 'tmp))
903 (tmp-syms (fresh-temporaries tmps)))
907 (append gensyms tmp-syms)
908 (append vals rest-args)
913 (map (cut make-lexical-ref #f <> <>)
915 (($ <let> src names gensyms vals body)
916 (define (compute-alias exp)
917 ;; It's very common for macros to introduce something like:
919 ;; ((lambda (x y) ...) x-exp y-exp)
921 ;; In that case you might end up trying to inline something like:
923 ;; (let ((x x-exp) (y y-exp)) ...)
925 ;; But if x-exp is itself a lexical-ref that aliases some much
926 ;; larger expression, perhaps it will fail to inline due to
927 ;; size. However we don't want to introduce a useless alias
928 ;; (in this case, x). So if the RHS of a let expression is a
929 ;; lexical-ref, we record that expression. If we end up having
930 ;; to residualize X, then instead we residualize X-EXP, as long
931 ;; as it isn't assigned.
934 (($ <lexical-ref> _ _ sym)
935 (let ((op (lookup sym)))
936 (and (not (var-set? (operand-var op)))
937 (or (operand-alias-value op)
941 (let* ((vars (map lookup-var gensyms))
942 (new (fresh-gensyms vars))
943 (ops (make-bound-operands vars new vals
944 (lambda (exp counter ctx)
945 (loop exp env counter ctx))
946 (map compute-alias vals)))
947 (env (fold extend-env env gensyms ops))
948 (body (loop body env counter ctx)))
951 (for-tail (list->seq src (append vals (list body)))))
952 ((and (lexical-ref? body)
953 (memq (lexical-ref-gensym body) new))
954 (let ((sym (lexical-ref-gensym body))
955 (pairs (map cons new vals)))
956 ;; (let ((x foo) (y bar) ...) x) => (begin bar ... foo)
960 (append (map cdr (alist-delete sym pairs eq?))
961 (list (assq-ref pairs sym)))))))
963 ;; Only include bindings for which lexical references
964 ;; have been residualized.
965 (prune-bindings ops #f body counter ctx
966 (lambda (names gensyms vals body)
967 (if (null? names) (error "what!" names))
968 (make-let src names gensyms vals body)))))))
969 (($ <letrec> src in-order? names gensyms vals body)
970 ;; Note the difference from the `let' case: here we use letrec*
971 ;; so that the `visit' procedure for the new operands closes over
972 ;; an environment that includes the operands. Also we don't try
973 ;; to elide aliases, because we can't sensibly reduce something
974 ;; like (letrec ((a b) (b a)) a).
975 (letrec* ((visit (lambda (exp counter ctx)
976 (loop exp env* counter ctx)))
977 (vars (map lookup-var gensyms))
978 (new (fresh-gensyms vars))
979 (ops (make-bound-operands vars new vals visit))
980 (env* (fold extend-env env gensyms ops))
981 (body* (visit body counter ctx)))
982 (if (and (const? body*) (every constant-expression? vals))
983 ;; We may have folded a loop completely, even though there
984 ;; might be cyclical references between the bound values.
985 ;; Handle this degenerate case specially.
987 (prune-bindings ops in-order? body* counter ctx
988 (lambda (names gensyms vals body)
989 (make-letrec src in-order?
990 names gensyms vals body))))))
991 (($ <fix> src names gensyms vals body)
992 (letrec* ((visit (lambda (exp counter ctx)
993 (loop exp env* counter ctx)))
994 (vars (map lookup-var gensyms))
995 (new (fresh-gensyms vars))
996 (ops (make-bound-operands vars new vals visit))
997 (env* (fold extend-env env gensyms ops))
998 (body* (visit body counter ctx)))
1001 (prune-bindings ops #f body* counter ctx
1002 (lambda (names gensyms vals body)
1003 (make-fix src names gensyms vals body))))))
1004 (($ <let-values> lv-src producer consumer)
1005 ;; Peval the producer, then try to inline the consumer into
1006 ;; the producer. If that succeeds, peval again. Otherwise
1007 ;; reconstruct the let-values, pevaling the consumer.
1008 (let ((producer (for-values producer)))
1010 (($ <lambda-case> src (req-name) #f #f #f () (req-sym) body #f)
1012 (make-let src (list req-name) (list req-sym) (list producer)
1014 ((and ($ <lambda-case> src () #f rest #f () (rest-sym) body #f)
1015 (? (lambda _ (singly-valued-expression? producer))))
1016 (let ((tmp (gensym "tmp ")))
1017 (record-new-temporary! 'tmp tmp 1)
1020 src (list 'tmp) (list tmp) (list producer)
1022 src (list rest) (list rest-sym)
1024 (make-primcall #f 'list
1025 (list (make-lexical-ref #f 'tmp tmp))))
1027 (($ <lambda-case> src req opt rest #f inits gensyms body #f)
1028 (let* ((nmin (length req))
1029 (nmax (and (not rest) (+ nmin (if opt (length opt) 0)))))
1031 ((inline-values lv-src producer nmin nmax consumer)
1035 (make-let-values lv-src producer (for-tail consumer)))))
1036 (($ <toplevel-ref> src (? effect-free-primitive? name))
1039 ;; todo: open private local bindings.
1041 (($ <module-ref> src module (? effect-free-primitive? name) #f)
1042 (let ((module (false-if-exception
1043 (resolve-module module #:ensure #f))))
1044 (if (module? module)
1045 (let ((var (module-variable module name)))
1046 (if (eq? var (module-variable the-scm-module name))
1047 (make-primitive-ref src name)
1052 (($ <module-set> src mod name public? exp)
1053 (make-module-set src mod name public? (for-value exp)))
1054 (($ <toplevel-define> src name exp)
1055 (make-toplevel-define src name (for-value exp)))
1056 (($ <toplevel-set> src name exp)
1057 (make-toplevel-set src name (for-value exp)))
1058 (($ <primitive-ref>)
1060 ((effect) (make-void #f))
1061 ((test) (make-const #f #t))
1063 (($ <conditional> src condition subsequent alternate)
1064 (define (call-with-failure-thunk exp proc)
1066 (($ <call> _ _ ()) (proc exp))
1067 (($ <primcall> _ _ ()) (proc exp))
1068 (($ <const>) (proc exp))
1069 (($ <void>) (proc exp))
1070 (($ <lexical-ref>) (proc exp))
1072 (let ((t (gensym "failure-")))
1073 (record-new-temporary! 'failure t 2)
1075 src (list 'failure) (list t)
1079 (make-lambda-case #f '() #f #f #f '() '() exp #f)))
1080 (proc (make-call #f (make-lexical-ref #f 'failure t)
1082 (define (simplify-conditional c)
1084 ;; Swap the arms of (if (not FOO) A B), to simplify.
1085 (($ <conditional> src ($ <primcall> _ 'not (pred))
1086 subsequent alternate)
1087 (simplify-conditional
1088 (make-conditional src pred alternate subsequent)))
1089 ;; Special cases for common tests in the predicates of chains
1090 ;; of if expressions.
1091 (($ <conditional> src
1092 ($ <conditional> src* outer-test inner-test ($ <const> _ #f))
1095 (let lp ((alternate alternate))
1097 ;; Lift a common repeated test out of a chain of if
1099 (($ <conditional> _ (? (cut tree-il=? outer-test <>))
1100 other-subsequent alternate)
1103 (simplify-conditional
1104 (make-conditional src* inner-test inner-subsequent
1107 ;; Likewise, but punching through any surrounding
1108 ;; failure continuations.
1109 (($ <let> let-src (name) (sym) ((and thunk ($ <lambda>))) body)
1111 let-src (list name) (list sym) (list thunk)
1113 ;; Otherwise, rotate AND tests to expose a simple
1114 ;; condition in the front. Although this may result in
1115 ;; lexically binding failure thunks, the thunks will be
1116 ;; compiled to labels allocation, so there's no actual
1119 (call-with-failure-thunk
1124 (simplify-conditional
1125 (make-conditional src* inner-test inner-subsequent failure))
1128 (match (for-test condition)
1131 (for-tail subsequent)
1132 (for-tail alternate)))
1134 (simplify-conditional
1135 (make-conditional src c (for-tail subsequent)
1136 (for-tail alternate))))))
1137 (($ <primcall> src 'call-with-values
1141 ;; No optional or kwargs.
1143 _ req #f rest #f () gensyms body #f)))))
1144 (for-tail (make-let-values src (make-call src producer '())
1146 (($ <primcall> src 'dynamic-wind (w thunk u))
1149 src (list w u) 2 constant-expression?
1158 ;; fixme: introduce logic to fold thunk?
1159 (make-primcall src 'thunk? (list u))
1160 (make-call src w '())
1164 (make-const #f 'wrong-type-arg)
1165 (make-const #f "dynamic-wind")
1166 (make-const #f "Wrong type (expecting thunk): ~S")
1167 (make-primcall #f 'list (list u))
1168 (make-primcall #f 'list (list u)))))
1169 (make-primcall src 'wind (list w u)))
1171 (make-call src thunk '())
1173 (make-primcall src 'unwind '())
1174 (make-call src u '())))))))))
1176 (($ <primcall> src 'with-fluid* (f v thunk))
1179 src (list f v thunk) 1 constant-expression?
1183 (make-primcall src 'push-fluid (list f v))
1185 (make-call src thunk '())
1186 (make-primcall src 'pop-fluid '()))))))))
1188 (($ <primcall> src 'values exps)
1191 (if (eq? ctx 'effect)
1195 (let ((vals (map for-value exps)))
1197 ((value test effect) #t)
1198 (else (null? (cdr vals))))
1199 (every singly-valued-expression? vals))
1200 (for-tail (list->seq src (append (cdr vals) (list (car vals)))))
1201 (make-primcall src 'values vals))))))
1203 (($ <primcall> src 'apply (proc args ... tail))
1204 (let lp ((tail* (find-definition tail 1)) (speculative? #t))
1205 (define (copyable? x)
1206 ;; Inlining a result from find-definition effectively copies it,
1207 ;; relying on the let-pruning to remove its original binding. We
1208 ;; shouldn't copy non-constant expressions.
1209 (or (not speculative?) (constant-expression? x)))
1211 (($ <const> _ (args* ...))
1212 (let ((args* (map (cut make-const #f <>) args*)))
1213 (for-tail (make-call src proc (append args args*)))))
1214 (($ <primcall> _ 'cons
1215 ((and head (? copyable?)) (and tail (? copyable?))))
1216 (for-tail (make-primcall src 'apply
1218 (append args (list head tail))))))
1219 (($ <primcall> _ 'list
1220 (and args* ((? copyable?) ...)))
1221 (for-tail (make-call src proc (append args args*))))
1224 (lp (for-value tail) #f)
1225 (let ((args (append (map for-value args) (list tail*))))
1226 (make-primcall src 'apply
1227 (cons (for-value proc) args))))))))
1229 (($ <primcall> src (? constructor-primitive? name) args)
1231 ((and (memq ctx '(effect test))
1232 (match (cons name args)
1237 ('make-prompt-tag ($ <const> _ (? string?))))
1240 ;; Some expressions can be folded without visiting the
1241 ;; arguments for value.
1242 (let ((res (if (eq? ctx 'effect)
1244 (make-const #f #t))))
1245 (for-tail (list->seq src (append args (list res))))))
1247 (match (cons name (map for-value args))
1248 (('cons x ($ <const> _ (? (cut eq? <> '()))))
1249 (make-primcall src 'list (list x)))
1250 (('cons x ($ <primcall> _ 'list elts))
1251 (make-primcall src 'list (cons x elts)))
1253 (make-primcall src name args))))))
1255 (($ <primcall> src 'thunk? (proc))
1258 (for-tail (make-seq src proc (make-void src))))
1260 (match (for-value proc)
1261 (($ <lambda> _ _ ($ <lambda-case> _ req))
1262 (for-tail (make-const src (null? req))))
1264 (match (find-definition proc 2)
1265 (($ <lambda> _ _ ($ <lambda-case> _ req))
1266 (for-tail (make-const src (null? req))))
1268 (make-primcall src 'thunk? (list proc)))))))))
1270 (($ <primcall> src name args)
1271 (match (cons name (map for-value args))
1272 ;; FIXME: these for-tail recursions could take place outside
1273 ;; an effort counter.
1274 (('car ($ <primcall> src 'cons (head tail)))
1275 (for-tail (make-seq src tail head)))
1276 (('cdr ($ <primcall> src 'cons (head tail)))
1277 (for-tail (make-seq src head tail)))
1278 (('car ($ <primcall> src 'list (head . tail)))
1279 (for-tail (list->seq src (append tail (list head)))))
1280 (('cdr ($ <primcall> src 'list (head . tail)))
1281 (for-tail (make-seq src head (make-primcall #f 'list tail))))
1283 (('car ($ <const> src (head . tail)))
1284 (for-tail (make-const src head)))
1285 (('cdr ($ <const> src (head . tail)))
1286 (for-tail (make-const src tail)))
1287 (((or 'memq 'memv) k ($ <const> _ (elts ...)))
1292 (make-seq src k (make-void #f))))
1296 ;; A shortcut. The `else' case would handle it, but
1297 ;; this way is faster.
1298 (let ((member (case name ((memq) memq) ((memv) memv))))
1299 (make-const #f (and (member (const-exp k) elts) #t))))
1302 (make-seq src k (make-const #f #f))))
1304 (let ((t (gensym "t "))
1305 (eq (if (eq? name 'memq) 'eq? 'eqv?)))
1306 (record-new-temporary! 't t (length elts))
1309 src (list 't) (list t) (list k)
1310 (let lp ((elts elts))
1312 (make-primcall #f eq
1313 (list (make-lexical-ref #f 't t)
1314 (make-const #f (car elts)))))
1315 (if (null? (cdr elts))
1317 (make-conditional src test
1319 (lp (cdr elts)))))))))))
1323 (let ((member (case name ((memq) memq) ((memv) memv))))
1324 (make-const #f (member (const-exp k) elts))))
1326 (for-tail (make-seq src k (make-const #f #f))))
1328 (make-primcall src name (list k (make-const #f elts))))))))
1329 (((? equality-primitive?)
1330 ($ <lexical-ref> _ _ sym) ($ <lexical-ref> _ _ sym))
1331 (for-tail (make-const #f #t)))
1333 (((? effect-free-primitive?) . args)
1334 (fold-constants src name args ctx))
1337 (make-primcall src name args))))
1339 (($ <call> src orig-proc orig-args)
1340 ;; todo: augment the global env with specialized functions
1341 (let revisit-proc ((proc (visit orig-proc 'operator)))
1343 (($ <primitive-ref> _ name)
1344 (for-tail (make-primcall src name orig-args)))
1346 ($ <lambda-case> _ req opt rest #f inits gensyms body #f))
1347 ;; Simple case: no keyword arguments.
1348 ;; todo: handle the more complex cases
1349 (let* ((nargs (length orig-args))
1352 (rest (if rest (list rest) '()))
1354 (key (source-expression proc)))
1355 (define (inlined-call)
1356 (let ((req-vals (list-head orig-args nreq))
1357 (opt-vals (let lp ((args (drop orig-args nreq))
1365 (lp '() inits (cons init out)))
1367 (lp args inits (cons arg out))))))))
1369 ((> nargs (+ nreq nopt))
1370 (list (make-primcall
1372 (drop orig-args (+ nreq nopt)))))
1373 (rest (list (make-const #f '())))
1375 (if (>= nargs (+ nreq nopt))
1377 (append req opt rest)
1379 (append req-vals opt-vals rest-vals)
1381 ;; The required argument values are in the scope
1382 ;; of the optional argument initializers.
1385 (append (list-head gensyms nreq)
1386 (last-pair gensyms))
1387 (append req-vals rest-vals)
1390 (list-head (drop gensyms nreq) nopt)
1395 ((or (< nargs nreq) (and (not rest) (> nargs (+ nreq nopt))))
1396 ;; An error, or effecting arguments.
1397 (make-call src (for-call orig-proc) (map for-value orig-args)))
1398 ((or (and=> (find-counter key counter) counter-recursive?)
1399 (lambda? orig-proc))
1400 ;; A recursive call, or a lambda in the operator
1401 ;; position of the source expression. Process again in
1404 ;; In the recursive case, mark intervening counters as
1405 ;; recursive, so we can handle a toplevel counter that
1406 ;; recurses mutually with some other procedure.
1407 ;; Otherwise, the next time we see the other procedure,
1408 ;; the effort limit would be clamped to 100.
1410 (let ((found (find-counter key counter)))
1411 (if (and found (counter-recursive? found))
1412 (let lp ((counter counter))
1413 (if (not (eq? counter found))
1415 (set-counter-recursive?! counter #t)
1416 (lp (counter-prev counter)))))))
1418 (log 'inline-recurse key)
1419 (loop (inlined-call) env counter ctx))
1421 ;; An integration at the top-level, the first
1422 ;; recursion of a recursive procedure, or a nested
1423 ;; integration of a procedure that hasn't been seen
1425 (log 'inline-begin exp)
1428 (log 'inline-abort exp)
1429 (k (make-call src (for-call orig-proc)
1430 (map for-value orig-args))))
1433 ;; These first two cases will transfer effort
1434 ;; from the current counter into the new
1436 ((find-counter key counter)
1438 (make-recursive-counter recursive-effort-limit
1442 (make-nested-counter abort key counter))
1443 ;; This case opens a new account, effectively
1444 ;; printing money. It should only do so once
1445 ;; for each call site in the source program.
1447 (make-top-counter effort-limit operand-size-limit
1450 (loop (inlined-call) env new-counter ctx))
1453 ;; The nested inlining attempt succeeded.
1454 ;; Deposit the unspent effort and size back
1455 ;; into the current counter.
1456 (transfer! new-counter counter))
1458 (log 'inline-end result exp)
1460 (($ <let> _ _ _ vals _)
1461 ;; Attempt to inline `let' in the operator position.
1463 ;; We have to re-visit the proc in value mode, since the
1464 ;; `let' bindings might have been introduced or renamed,
1465 ;; whereas the lambda (if any) in operator position has not
1467 (if (or (and-map constant-expression? vals)
1468 (and-map constant-expression? orig-args))
1469 ;; The arguments and the let-bound values commute.
1470 (match (for-value orig-proc)
1471 (($ <let> lsrc names syms vals body)
1472 (log 'inline-let orig-proc)
1474 (make-let lsrc names syms vals
1475 (make-call src body orig-args))))
1476 ;; It's possible for a `let' to go away after the
1477 ;; visit due to the fact that visiting a procedure in
1478 ;; value context will prune unused bindings, whereas
1479 ;; visiting in operator mode can't because it doesn't
1480 ;; traverse through lambdas. In that case re-visit
1482 (proc (revisit-proc proc)))
1483 (make-call src (for-call orig-proc)
1484 (map for-value orig-args))))
1486 (make-call src (for-call orig-proc) (map for-value orig-args))))))
1487 (($ <lambda> src meta body)
1489 ((effect) (make-void #f))
1490 ((test) (make-const #f #t))
1492 (else (record-source-expression!
1494 (make-lambda src meta (and body (for-values body)))))))
1495 (($ <lambda-case> src req opt rest kw inits gensyms body alt)
1496 (define (lift-applied-lambda body gensyms)
1497 (and (not opt) rest (not kw)
1499 (($ <primcall> _ 'apply
1500 (($ <lambda> _ _ (and lcase ($ <lambda-case>)))
1501 ($ <lexical-ref> _ _ sym)
1503 (and (equal? sym gensyms)
1504 (not (lambda-case-alternate lcase))
1507 (let* ((vars (map lookup-var gensyms))
1508 (new (fresh-gensyms vars))
1509 (env (fold extend-env env gensyms
1510 (make-unbound-operands vars new)))
1511 (new-sym (lambda (old)
1512 (operand-sym (cdr (vhash-assq old env)))))
1513 (body (loop body env counter ctx)))
1515 ;; (lambda args (apply (lambda ...) args)) => (lambda ...)
1516 (lift-applied-lambda body new)
1517 (make-lambda-case src req opt rest
1519 ((aok? (kw name old) ...)
1520 (cons aok? (map list kw name (map new-sym old))))
1522 (map (cut loop <> env counter 'value) inits)
1525 (and alt (for-tail alt))))))
1526 (($ <seq> src head tail)
1527 (let ((head (for-effect head))
1528 (tail (for-tail tail)))
1532 (if (and (seq? head)
1533 (void? (seq-tail head)))
1537 (($ <prompt> src escape-only? tag body handler)
1538 (define (make-prompt-tag? x)
1540 (($ <primcall> _ 'make-prompt-tag (or () ((? constant-expression?))))
1544 (let ((tag (for-value tag))
1545 (body (if escape-only? (for-tail body) (for-value body))))
1547 ((find-definition tag 1)
1549 (make-prompt-tag? val))
1551 ;; There is no way that an <abort> could know the tag
1552 ;; for this <prompt>, so we can elide the <prompt>
1554 (unrecord-operand-uses op 1)
1555 (for-tail (if escape-only? body (make-call src body '())))))
1557 (let ((handler (for-value handler)))
1558 (define (escape-only-handler? handler)
1561 ($ <lambda-case> _ (_ . _) _ _ _ _ (k . _) body #f))
1564 (($ <lexical-ref> _ _ (? (cut eq? <> k))) #t)
1568 (if (and (not escape-only?) (escape-only-handler? handler))
1569 ;; Prompt transitioning to escape-only; transition body
1570 ;; to be an expression.
1572 (make-prompt src #t tag (make-call #f body '()) handler))
1573 (make-prompt src escape-only? tag body handler)))))))
1575 (($ <abort> src tag args tail)
1576 (make-abort src (for-value tag) (map for-value args)
1577 (for-value tail))))))