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 (let ((y (or (operand-residual-value op)
735 (visit-operand op counter 'value 10 10)
736 (operand-source op))))
738 ((and (lexical-ref? y)
739 (= (lexical-refcount (lexical-ref-gensym x)) 1))
740 ;; X is a simple alias for Y. Recurse, regardless of
741 ;; the number of aliases we were expecting.
742 (find-definition y n-aliases))
743 ((= (lexical-refcount (lexical-ref-gensym x)) n-aliases)
744 ;; We found a definition that is aliased the right
745 ;; number of times. We still recurse in case it is a
747 (values (find-definition y 1)
750 ;; We can't account for our aliases.
753 ;; A formal parameter. Can't say anything about that.
756 ;; Not a lexical: success, but only if we are looking for an
759 (else (values #f #f))))
761 (define (visit exp ctx)
762 (loop exp env counter ctx))
764 (define (for-value exp) (visit exp 'value))
765 (define (for-values exp) (visit exp 'values))
766 (define (for-test exp) (visit exp 'test))
767 (define (for-effect exp) (visit exp 'effect))
768 (define (for-call exp) (visit exp 'call))
769 (define (for-tail exp) (visit exp ctx))
772 (record-effort! counter))
774 (log 'visit ctx (and=> counter effort-counter)
775 (unparse-tree-il exp))
780 ((effect) (make-void #f))
784 ((test) (make-const #f #t))
786 (($ <lexical-ref> _ _ gensym)
787 (log 'begin-copy gensym)
788 (let ((op (lookup gensym)))
791 (log 'lexical-for-effect gensym)
793 ((operand-alias-value op)
794 ;; This is an unassigned operand that simply aliases some
795 ;; other operand. Recurse to avoid residualizing the leaf
799 ;; Don't propagate copies if we are residualizing a call.
800 (log 'residualize-lexical-call gensym op)
801 (residualize-lexical op))
802 ((var-set? (operand-var op))
803 ;; Assigned lexicals don't copy-propagate.
804 (log 'assigned-var gensym op)
805 (residualize-lexical op))
806 ((not (operand-copyable? op))
807 ;; We already know that this operand is not copyable.
808 (log 'not-copyable gensym op)
809 (residualize-lexical op))
810 ((and=> (operand-constant-value op)
811 (lambda (x) (or (const? x) (void? x) (primitive-ref? x))))
813 (let ((val (operand-constant-value op)))
814 (log 'memoized-constant gensym val)
816 ((visit-operand op counter (if (eq? ctx 'values) 'value ctx)
817 recursive-effort-limit operand-size-limit)
819 ;; If we end up deciding to residualize this value instead of
820 ;; copying it, save that residualized value.
823 ((not (constant-expression? val))
824 (log 'not-constant gensym op)
825 ;; At this point, ctx is operator, test, or value. A
826 ;; value that is non-constant in one context will be
827 ;; non-constant in the others, so it's safe to record
828 ;; that here, and avoid future visits.
829 (set-operand-copyable?! op #f)
830 (residualize-lexical op ctx val))
833 (primitive-ref? val))
834 ;; Always propagate simple values that cannot lead to
836 (log 'copy-simple gensym val)
837 ;; It could be this constant is the result of folding.
838 ;; If that is the case, cache it. This helps loop
839 ;; unrolling get farther.
840 (if (or (eq? ctx 'value) (eq? ctx 'values))
842 (log 'memoize-constant gensym val)
843 (set-operand-constant-value! op val)))
845 ((= 1 (var-refcount (operand-var op)))
846 ;; Always propagate values referenced only once.
847 (log 'copy-single gensym val)
849 ;; FIXME: do demand-driven size accounting rather than
852 ;; A pure expression in the operator position. Inline
853 ;; if it's a lambda that's small enough.
854 (if (and (lambda? val)
855 (small-expression? val operator-size-limit))
857 (log 'copy-operator gensym val)
860 (log 'too-big-for-operator gensym val)
861 (residualize-lexical op ctx val))))
863 ;; A pure expression, processed for call or for value.
864 ;; Don't inline lambdas, because they will probably won't
865 ;; fold because we don't know the operator.
866 (if (and (small-expression? val value-size-limit)
867 (not (tree-il-any lambda? val)))
869 (log 'copy-value gensym val)
872 (log 'too-big-or-has-lambda gensym val)
873 (residualize-lexical op ctx val)))))))
875 ;; Visit failed. Either the operand isn't bound, as in
876 ;; lambda formal parameters, or the copy was aborted.
877 (log 'unbound-or-aborted gensym op)
878 (residualize-lexical op)))))
879 (($ <lexical-set> src name gensym exp)
880 (let ((op (lookup gensym)))
881 (if (zero? (var-refcount (operand-var op)))
882 (let ((exp (for-effect exp)))
885 (make-seq src exp (make-void #f))))
887 (record-operand-use op)
888 (make-lexical-set src name (operand-sym op) (for-value exp))))))
891 (gensyms ... rest-sym)
892 (vals ... ($ <primcall> _ 'list rest-args))
893 ($ <primcall> asrc 'apply
896 (? (cut eq? <> rest))
898 (and (eq? sym rest-sym)
899 (= (lexical-refcount sym) 1))))))))
900 (let* ((tmps (make-list (length rest-args) 'tmp))
901 (tmp-syms (fresh-temporaries tmps)))
905 (append gensyms tmp-syms)
906 (append vals rest-args)
911 (map (cut make-lexical-ref #f <> <>)
913 (($ <let> src names gensyms vals body)
914 (define (compute-alias exp)
915 ;; It's very common for macros to introduce something like:
917 ;; ((lambda (x y) ...) x-exp y-exp)
919 ;; In that case you might end up trying to inline something like:
921 ;; (let ((x x-exp) (y y-exp)) ...)
923 ;; But if x-exp is itself a lexical-ref that aliases some much
924 ;; larger expression, perhaps it will fail to inline due to
925 ;; size. However we don't want to introduce a useless alias
926 ;; (in this case, x). So if the RHS of a let expression is a
927 ;; lexical-ref, we record that expression. If we end up having
928 ;; to residualize X, then instead we residualize X-EXP, as long
929 ;; as it isn't assigned.
932 (($ <lexical-ref> _ _ sym)
933 (let ((op (lookup sym)))
934 (and (not (var-set? (operand-var op)))
935 (or (operand-alias-value op)
939 (let* ((vars (map lookup-var gensyms))
940 (new (fresh-gensyms vars))
941 (ops (make-bound-operands vars new vals
942 (lambda (exp counter ctx)
943 (loop exp env counter ctx))
944 (map compute-alias vals)))
945 (env (fold extend-env env gensyms ops))
946 (body (loop body env counter ctx)))
949 (for-tail (list->seq src (append vals (list body)))))
950 ((and (lexical-ref? body)
951 (memq (lexical-ref-gensym body) new))
952 (let ((sym (lexical-ref-gensym body))
953 (pairs (map cons new vals)))
954 ;; (let ((x foo) (y bar) ...) x) => (begin bar ... foo)
958 (append (map cdr (alist-delete sym pairs eq?))
959 (list (assq-ref pairs sym)))))))
961 ;; Only include bindings for which lexical references
962 ;; have been residualized.
963 (prune-bindings ops #f body counter ctx
964 (lambda (names gensyms vals body)
965 (if (null? names) (error "what!" names))
966 (make-let src names gensyms vals body)))))))
967 (($ <letrec> src in-order? names gensyms vals body)
968 ;; Note the difference from the `let' case: here we use letrec*
969 ;; so that the `visit' procedure for the new operands closes over
970 ;; an environment that includes the operands. Also we don't try
971 ;; to elide aliases, because we can't sensibly reduce something
972 ;; like (letrec ((a b) (b a)) a).
973 (letrec* ((visit (lambda (exp counter ctx)
974 (loop exp env* counter ctx)))
975 (vars (map lookup-var gensyms))
976 (new (fresh-gensyms vars))
977 (ops (make-bound-operands vars new vals visit))
978 (env* (fold extend-env env gensyms ops))
979 (body* (visit body counter ctx)))
980 (if (and (const? body*) (every constant-expression? vals))
981 ;; We may have folded a loop completely, even though there
982 ;; might be cyclical references between the bound values.
983 ;; Handle this degenerate case specially.
985 (prune-bindings ops in-order? body* counter ctx
986 (lambda (names gensyms vals body)
987 (make-letrec src in-order?
988 names gensyms vals body))))))
989 (($ <fix> src names gensyms vals body)
990 (letrec* ((visit (lambda (exp counter ctx)
991 (loop exp env* counter ctx)))
992 (vars (map lookup-var gensyms))
993 (new (fresh-gensyms vars))
994 (ops (make-bound-operands vars new vals visit))
995 (env* (fold extend-env env gensyms ops))
996 (body* (visit body counter ctx)))
999 (prune-bindings ops #f body* counter ctx
1000 (lambda (names gensyms vals body)
1001 (make-fix src names gensyms vals body))))))
1002 (($ <let-values> lv-src producer consumer)
1003 ;; Peval the producer, then try to inline the consumer into
1004 ;; the producer. If that succeeds, peval again. Otherwise
1005 ;; reconstruct the let-values, pevaling the consumer.
1006 (let ((producer (for-values producer)))
1008 (($ <lambda-case> src (req-name) #f #f #f () (req-sym) body #f)
1010 (make-let src (list req-name) (list req-sym) (list producer)
1012 ((and ($ <lambda-case> src () #f rest #f () (rest-sym) body #f)
1013 (? (lambda _ (singly-valued-expression? producer))))
1014 (let ((tmp (gensym "tmp ")))
1015 (record-new-temporary! 'tmp tmp 1)
1018 src (list 'tmp) (list tmp) (list producer)
1020 src (list rest) (list rest-sym)
1022 (make-primcall #f 'list
1023 (list (make-lexical-ref #f 'tmp tmp))))
1025 (($ <lambda-case> src req opt rest #f inits gensyms body #f)
1026 (let* ((nmin (length req))
1027 (nmax (and (not rest) (+ nmin (if opt (length opt) 0)))))
1029 ((inline-values lv-src producer nmin nmax consumer)
1033 (make-let-values lv-src producer (for-tail consumer)))))
1034 (($ <toplevel-ref> src (? effect-free-primitive? name))
1037 ;; todo: open private local bindings.
1039 (($ <module-ref> src module (? effect-free-primitive? name) #f)
1040 (let ((module (false-if-exception
1041 (resolve-module module #:ensure #f))))
1042 (if (module? module)
1043 (let ((var (module-variable module name)))
1044 (if (eq? var (module-variable the-scm-module name))
1045 (make-primitive-ref src name)
1050 (($ <module-set> src mod name public? exp)
1051 (make-module-set src mod name public? (for-value exp)))
1052 (($ <toplevel-define> src name exp)
1053 (make-toplevel-define src name (for-value exp)))
1054 (($ <toplevel-set> src name exp)
1055 (make-toplevel-set src name (for-value exp)))
1056 (($ <primitive-ref>)
1058 ((effect) (make-void #f))
1059 ((test) (make-const #f #t))
1061 (($ <conditional> src condition subsequent alternate)
1062 (define (call-with-failure-thunk exp proc)
1064 (($ <call> _ _ ()) (proc exp))
1065 (($ <primcall> _ _ ()) (proc exp))
1066 (($ <const>) (proc exp))
1067 (($ <void>) (proc exp))
1068 (($ <lexical-ref>) (proc exp))
1070 (let ((t (gensym "failure-")))
1071 (record-new-temporary! 'failure t 2)
1073 src (list 'failure) (list t)
1077 (make-lambda-case #f '() #f #f #f '() '() exp #f)))
1078 (proc (make-call #f (make-lexical-ref #f 'failure t)
1080 (define (simplify-conditional c)
1082 ;; Swap the arms of (if (not FOO) A B), to simplify.
1083 (($ <conditional> src ($ <primcall> _ 'not (pred))
1084 subsequent alternate)
1085 (simplify-conditional
1086 (make-conditional src pred alternate subsequent)))
1087 ;; Special cases for common tests in the predicates of chains
1088 ;; of if expressions.
1089 (($ <conditional> src
1090 ($ <conditional> src* outer-test inner-test ($ <const> _ #f))
1093 (let lp ((alternate alternate))
1095 ;; Lift a common repeated test out of a chain of if
1097 (($ <conditional> _ (? (cut tree-il=? outer-test <>))
1098 other-subsequent alternate)
1101 (simplify-conditional
1102 (make-conditional src* inner-test inner-subsequent
1105 ;; Likewise, but punching through any surrounding
1106 ;; failure continuations.
1107 (($ <let> let-src (name) (sym) ((and thunk ($ <lambda>))) body)
1109 let-src (list name) (list sym) (list thunk)
1111 ;; Otherwise, rotate AND tests to expose a simple
1112 ;; condition in the front. Although this may result in
1113 ;; lexically binding failure thunks, the thunks will be
1114 ;; compiled to labels allocation, so there's no actual
1117 (call-with-failure-thunk
1122 (simplify-conditional
1123 (make-conditional src* inner-test inner-subsequent failure))
1126 (match (for-test condition)
1129 (for-tail subsequent)
1130 (for-tail alternate)))
1132 (simplify-conditional
1133 (make-conditional src c (for-tail subsequent)
1134 (for-tail alternate))))))
1135 (($ <primcall> src 'call-with-values
1139 ;; No optional or kwargs.
1141 _ req #f rest #f () gensyms body #f)))))
1142 (for-tail (make-let-values src (make-call src producer '())
1144 (($ <primcall> src 'dynamic-wind (w thunk u))
1147 src (list w u) 2 constant-expression?
1156 ;; fixme: introduce logic to fold thunk?
1157 (make-primcall src 'thunk? (list u))
1158 (make-call src w '())
1162 (make-const #f 'wrong-type-arg)
1163 (make-const #f "dynamic-wind")
1164 (make-const #f "Wrong type (expecting thunk): ~S")
1165 (make-primcall #f 'list (list u))
1166 (make-primcall #f 'list (list u)))))
1167 (make-primcall src 'wind (list w u)))
1169 (make-call src thunk '())
1171 (make-primcall src 'unwind '())
1172 (make-call src u '())))))))))
1174 (($ <primcall> src 'with-fluid* (f v thunk))
1177 src (list f v thunk) 1 constant-expression?
1181 (make-primcall src 'push-fluid (list f v))
1183 (make-call src thunk '())
1184 (make-primcall src 'pop-fluid '()))))))))
1186 (($ <primcall> src 'values exps)
1189 (if (eq? ctx 'effect)
1193 (let ((vals (map for-value exps)))
1195 ((value test effect) #t)
1196 (else (null? (cdr vals))))
1197 (every singly-valued-expression? vals))
1198 (for-tail (list->seq src (append (cdr vals) (list (car vals)))))
1199 (make-primcall src 'values vals))))))
1201 (($ <primcall> src 'apply (proc args ... tail))
1202 (let lp ((tail* (find-definition tail 1)) (speculative? #t))
1203 (define (copyable? x)
1204 ;; Inlining a result from find-definition effectively copies it,
1205 ;; relying on the let-pruning to remove its original binding. We
1206 ;; shouldn't copy non-constant expressions.
1207 (or (not speculative?) (constant-expression? x)))
1209 (($ <const> _ (args* ...))
1210 (let ((args* (map (cut make-const #f <>) args*)))
1211 (for-tail (make-call src proc (append args args*)))))
1212 (($ <primcall> _ 'cons
1213 ((and head (? copyable?)) (and tail (? copyable?))))
1214 (for-tail (make-primcall src 'apply
1216 (append args (list head tail))))))
1217 (($ <primcall> _ 'list
1218 (and args* ((? copyable?) ...)))
1219 (for-tail (make-call src proc (append args args*))))
1222 (lp (for-value tail) #f)
1223 (let ((args (append (map for-value args) (list tail*))))
1224 (make-primcall src 'apply
1225 (cons (for-value proc) args))))))))
1227 (($ <primcall> src (? constructor-primitive? name) args)
1229 ((and (memq ctx '(effect test))
1230 (match (cons name args)
1235 ('make-prompt-tag ($ <const> _ (? string?))))
1238 ;; Some expressions can be folded without visiting the
1239 ;; arguments for value.
1240 (let ((res (if (eq? ctx 'effect)
1242 (make-const #f #t))))
1243 (for-tail (list->seq src (append args (list res))))))
1245 (match (cons name (map for-value args))
1246 (('cons x ($ <const> _ (? (cut eq? <> '()))))
1247 (make-primcall src 'list (list x)))
1248 (('cons x ($ <primcall> _ 'list elts))
1249 (make-primcall src 'list (cons x elts)))
1251 (make-primcall src name args))))))
1253 (($ <primcall> src 'thunk? (proc))
1254 (match (for-value proc)
1255 (($ <lambda> _ _ ($ <lambda-case> _ req))
1256 (for-tail (make-const src (null? req))))
1259 ((effect) (make-void src))
1260 (else (make-primcall src 'thunk? (list proc)))))))
1262 (($ <primcall> src (? accessor-primitive? name) args)
1263 (match (cons name (map for-value args))
1264 ;; FIXME: these for-tail recursions could take place outside
1265 ;; an effort counter.
1266 (('car ($ <primcall> src 'cons (head tail)))
1267 (for-tail (make-seq src tail head)))
1268 (('cdr ($ <primcall> src 'cons (head tail)))
1269 (for-tail (make-seq src head tail)))
1270 (('car ($ <primcall> src 'list (head . tail)))
1271 (for-tail (list->seq src (append tail (list head)))))
1272 (('cdr ($ <primcall> src 'list (head . tail)))
1273 (for-tail (make-seq src head (make-primcall #f 'list tail))))
1275 (('car ($ <const> src (head . tail)))
1276 (for-tail (make-const src head)))
1277 (('cdr ($ <const> src (head . tail)))
1278 (for-tail (make-const src tail)))
1279 (((or 'memq 'memv) k ($ <const> _ (elts ...)))
1284 (make-seq src k (make-void #f))))
1288 ;; A shortcut. The `else' case would handle it, but
1289 ;; this way is faster.
1290 (let ((member (case name ((memq) memq) ((memv) memv))))
1291 (make-const #f (and (member (const-exp k) elts) #t))))
1294 (make-seq src k (make-const #f #f))))
1296 (let ((t (gensym "t "))
1297 (eq (if (eq? name 'memq) 'eq? 'eqv?)))
1298 (record-new-temporary! 't t (length elts))
1301 src (list 't) (list t) (list k)
1302 (let lp ((elts elts))
1304 (make-primcall #f eq
1305 (list (make-lexical-ref #f 't t)
1306 (make-const #f (car elts)))))
1307 (if (null? (cdr elts))
1309 (make-conditional src test
1311 (lp (cdr elts)))))))))))
1315 (let ((member (case name ((memq) memq) ((memv) memv))))
1316 (make-const #f (member (const-exp k) elts))))
1318 (for-tail (make-seq src k (make-const #f #f))))
1320 (make-primcall src name (list k (make-const #f elts))))))))
1322 (fold-constants src name args ctx))))
1324 (($ <primcall> src (? equality-primitive? name) (a b))
1325 (let ((val-a (for-value a))
1326 (val-b (for-value b)))
1327 (log 'equality-primitive name val-a val-b)
1328 (cond ((and (lexical-ref? val-a) (lexical-ref? val-b)
1329 (eq? (lexical-ref-gensym val-a)
1330 (lexical-ref-gensym val-b)))
1331 (for-tail (make-const #f #t)))
1333 (fold-constants src name (list val-a val-b) ctx)))))
1335 (($ <primcall> src (? effect-free-primitive? name) args)
1336 (fold-constants src name (map for-value args) ctx))
1338 (($ <primcall> src name args)
1339 (make-primcall src name (map for-value args)))
1341 (($ <call> src orig-proc orig-args)
1342 ;; todo: augment the global env with specialized functions
1343 (let revisit-proc ((proc (visit orig-proc 'operator)))
1345 (($ <primitive-ref> _ name)
1346 (for-tail (make-primcall src name orig-args)))
1348 ($ <lambda-case> _ req opt rest #f inits gensyms body #f))
1349 ;; Simple case: no keyword arguments.
1350 ;; todo: handle the more complex cases
1351 (let* ((nargs (length orig-args))
1353 (nopt (if opt (length opt) 0))
1354 (key (source-expression proc)))
1355 (define (inlined-call)
1359 (if rest (list rest) '()))
1361 (if (> nargs (+ nreq nopt))
1362 (append (list-head orig-args (+ nreq nopt))
1366 (drop orig-args (+ nreq nopt)))))
1368 (drop inits (- nargs nreq))
1370 (list (make-const #f '()))
1375 ((or (< nargs nreq) (and (not rest) (> nargs (+ nreq nopt))))
1376 ;; An error, or effecting arguments.
1377 (make-call src (for-call orig-proc) (map for-value orig-args)))
1378 ((or (and=> (find-counter key counter) counter-recursive?)
1379 (lambda? orig-proc))
1380 ;; A recursive call, or a lambda in the operator
1381 ;; position of the source expression. Process again in
1384 ;; In the recursive case, mark intervening counters as
1385 ;; recursive, so we can handle a toplevel counter that
1386 ;; recurses mutually with some other procedure.
1387 ;; Otherwise, the next time we see the other procedure,
1388 ;; the effort limit would be clamped to 100.
1390 (let ((found (find-counter key counter)))
1391 (if (and found (counter-recursive? found))
1392 (let lp ((counter counter))
1393 (if (not (eq? counter found))
1395 (set-counter-recursive?! counter #t)
1396 (lp (counter-prev counter)))))))
1398 (log 'inline-recurse key)
1399 (loop (inlined-call) env counter ctx))
1401 ;; An integration at the top-level, the first
1402 ;; recursion of a recursive procedure, or a nested
1403 ;; integration of a procedure that hasn't been seen
1405 (log 'inline-begin exp)
1408 (log 'inline-abort exp)
1409 (k (make-call src (for-call orig-proc)
1410 (map for-value orig-args))))
1413 ;; These first two cases will transfer effort
1414 ;; from the current counter into the new
1416 ((find-counter key counter)
1418 (make-recursive-counter recursive-effort-limit
1422 (make-nested-counter abort key counter))
1423 ;; This case opens a new account, effectively
1424 ;; printing money. It should only do so once
1425 ;; for each call site in the source program.
1427 (make-top-counter effort-limit operand-size-limit
1430 (loop (inlined-call) env new-counter ctx))
1433 ;; The nested inlining attempt succeeded.
1434 ;; Deposit the unspent effort and size back
1435 ;; into the current counter.
1436 (transfer! new-counter counter))
1438 (log 'inline-end result exp)
1440 (($ <let> _ _ _ vals _)
1441 ;; Attempt to inline `let' in the operator position.
1443 ;; We have to re-visit the proc in value mode, since the
1444 ;; `let' bindings might have been introduced or renamed,
1445 ;; whereas the lambda (if any) in operator position has not
1447 (if (or (and-map constant-expression? vals)
1448 (and-map constant-expression? orig-args))
1449 ;; The arguments and the let-bound values commute.
1450 (match (for-value orig-proc)
1451 (($ <let> lsrc names syms vals body)
1452 (log 'inline-let orig-proc)
1454 (make-let lsrc names syms vals
1455 (make-call src body orig-args))))
1456 ;; It's possible for a `let' to go away after the
1457 ;; visit due to the fact that visiting a procedure in
1458 ;; value context will prune unused bindings, whereas
1459 ;; visiting in operator mode can't because it doesn't
1460 ;; traverse through lambdas. In that case re-visit
1462 (proc (revisit-proc proc)))
1463 (make-call src (for-call orig-proc)
1464 (map for-value orig-args))))
1466 (make-call src (for-call orig-proc) (map for-value orig-args))))))
1467 (($ <lambda> src meta body)
1469 ((effect) (make-void #f))
1470 ((test) (make-const #f #t))
1472 (else (record-source-expression!
1474 (make-lambda src meta (and body (for-values body)))))))
1475 (($ <lambda-case> src req opt rest kw inits gensyms body alt)
1476 (define (lift-applied-lambda body gensyms)
1477 (and (not opt) rest (not kw)
1479 (($ <primcall> _ 'apply
1480 (($ <lambda> _ _ (and lcase ($ <lambda-case>)))
1481 ($ <lexical-ref> _ _ sym)
1483 (and (equal? sym gensyms)
1484 (not (lambda-case-alternate lcase))
1487 (let* ((vars (map lookup-var gensyms))
1488 (new (fresh-gensyms vars))
1489 (env (fold extend-env env gensyms
1490 (make-unbound-operands vars new)))
1491 (new-sym (lambda (old)
1492 (operand-sym (cdr (vhash-assq old env)))))
1493 (body (loop body env counter ctx)))
1495 ;; (lambda args (apply (lambda ...) args)) => (lambda ...)
1496 (lift-applied-lambda body new)
1497 (make-lambda-case src req opt rest
1499 ((aok? (kw name old) ...)
1500 (cons aok? (map list kw name (map new-sym old))))
1502 (map (cut loop <> env counter 'value) inits)
1505 (and alt (for-tail alt))))))
1506 (($ <seq> src head tail)
1507 (let ((head (for-effect head))
1508 (tail (for-tail tail)))
1512 (if (and (seq? head)
1513 (void? (seq-tail head)))
1517 (($ <prompt> src escape-only? tag body handler)
1518 (define (make-prompt-tag? x)
1520 (($ <primcall> _ 'make-prompt-tag (or () ((? constant-expression?))))
1524 (let ((tag (for-value tag))
1525 (body (for-value body)))
1527 ((find-definition tag 1)
1529 (make-prompt-tag? val))
1531 ;; There is no way that an <abort> could know the tag
1532 ;; for this <prompt>, so we can elide the <prompt>
1534 (unrecord-operand-uses op 1)
1535 (for-tail (make-call src body '()))))
1536 ((find-definition tag 2)
1538 (and (make-prompt-tag? val)
1541 ($ <lambda-case> _ () #f #f #f () ()
1542 ($ <abort> _ (? (cut tree-il=? <> tag)))))
1546 ;; (let ((t (make-prompt-tag)))
1547 ;; (call-with-prompt t
1548 ;; (lambda () (abort-to-prompt t val ...))
1549 ;; (lambda (k arg ...) e ...)))
1550 ;; => (call-with-values (lambda () (values values val ...))
1551 ;; (lambda (k arg ...) e ...))
1552 (unrecord-operand-uses op 2)
1555 ($ <lambda-case> _ () #f #f #f () ()
1556 ($ <abort> _ _ args tail)))
1559 src 'call-with-values
1563 #f '() #f #f #f '() '()
1564 (make-primcall #f 'apply
1565 `(,(make-primitive-ref #f 'values)
1566 ,(make-primitive-ref #f 'values)
1572 (let ((handler (for-value handler)))
1573 (define (escape-only-handler? handler)
1576 ($ <lambda-case> _ (_ . _) _ _ _ _ (k . _) body #f))
1579 (($ <lexical-ref> _ _ (? (cut eq? <> k))) #t)
1583 (make-prompt src (or escape-only? (escape-only-handler? handler))
1584 tag body (for-value handler)))))))
1585 (($ <abort> src tag args tail)
1586 (make-abort src (for-value tag) (map for-value args)
1587 (for-value tail))))))