| blob | 620d28540e42361de6b3b7f79f28b3811fb5dc97 |
1 ;;;; This file implements local call analysis. A local call is a
2 ;;;; function call between functions being compiled at the same time.
3 ;;;; If we can tell at compile time that such a call is legal, then we
4 ;;;; change the combination to call the correct lambda, mark it as
5 ;;;; local, and add this link to our call graph. Once a call is local,
6 ;;;; it is then eligible for let conversion, which places the body of
7 ;;;; the function inline.
8 ;;;;
9 ;;;; We cannot always do a local call even when we do have the
10 ;;;; function being called. Calls that cannot be shown to have legal
11 ;;;; arg counts are not converted.
13 ;;;; This software is part of the SBCL system. See the README file for
14 ;;;; more information.
15 ;;;;
16 ;;;; This software is derived from the CMU CL system, which was
17 ;;;; written at Carnegie Mellon University and released into the
18 ;;;; public domain. The software is in the public domain and is
19 ;;;; provided with absolutely no warranty. See the COPYING and CREDITS
20 ;;;; files for more information.
24 ;;; This function propagates information from the variables in the
25 ;;; function FUN to the actual arguments in CALL. This is also called
26 ;;; by the VALUES IR1 optimizer when it sleazily converts MV-BINDs to
27 ;;; LETs.
28 ;;;
29 ;;; We flush all arguments to CALL that correspond to unreferenced
30 ;;; variables in FUN. We leave NILs in the COMBINATION-ARGS so that
31 ;;; the remaining args still match up with their vars.
32 ;;;
33 ;;; We also apply the declared variable type assertion to the argument
34 ;;; continuations.
51 ;;; This function handles merging the tail sets if CALL is potentially
52 ;;; tail-recursive, and is a call to a function with a different
53 ;;; TAIL-SET than CALL's FUN. This must be called whenever we alter
54 ;;; IR1 so as to place a local call in what might be a tail-recursive
55 ;;; context. Note that any call which returns its value to a RETURN is
56 ;;; considered potentially tail-recursive, since any implicit MV-PROG1
57 ;;; might be optimized away.
58 ;;;
59 ;;; We destructively modify the set for the calling function to
60 ;;; represent both, and then change all the functions in callee's set
61 ;;; to reference the first. If we do merge, we reoptimize the
62 ;;; RETURN-RESULT continuation to cause IR1-OPTIMIZE-RETURN to
63 ;;; recompute the tail set type.
77 t)))))
79 ;;; Convert a combination into a local call. We PROPAGATE-TO-ARGS, set
80 ;;; the combination kind to :LOCAL, add FUN to the CALLS of the
81 ;;; function that the call is in, call MERGE-TAIL-SETS, then replace
82 ;;; the function in the REF node with the new function.
83 ;;;
84 ;;; We change the REF last, since changing the reference can trigger
85 ;;; LET conversion of the new function, but will only do so if the
86 ;;; call is local. Note that the replacement may trigger LET
87 ;;; conversion or other changes in IR1. We must call MERGE-TAIL-SETS
88 ;;; with NEW-FUN before the substitution, since after the substitution
89 ;;; (and LET conversion), the call may no longer be recognizable as
90 ;;; tail-recursive.
99 \f
100 ;;;; external entry point creation
102 ;;; Return a LAMBDA form that can be used as the definition of the XEP
103 ;;; for FUN.
104 ;;;
105 ;;; If FUN is a LAMBDA, then we check the number of arguments
106 ;;; (conditional on policy) and call FUN with all the arguments.
107 ;;;
108 ;;; If FUN is an OPTIONAL-DISPATCH, then we dispatch off of the number
109 ;;; of supplied arguments by doing do an = test for each entry-point,
110 ;;; calling the entry with the appropriate prefix of the passed
111 ;;; arguments.
112 ;;;
113 ;;; If there is a &MORE arg, then there are a couple of optimizations
114 ;;; that we make (more for space than anything else):
115 ;;; -- If MIN-ARGS is 0, then we make the more entry a T clause, since
116 ;;; no argument count error is possible.
117 ;;; -- We can omit the = clause for the last entry-point, allowing the
118 ;;; case of 0 more args to fall through to the more entry.
119 ;;;
120 ;;; We don't bother to policy conditionalize wrong arg errors in
121 ;;; optional dispatches, since the additional overhead is negligible
122 ;;; compared to the cost of everything else going on.
123 ;;;
124 ;;; Note that if policy indicates it, argument type declarations in
125 ;;; FUN will be verified. Since nothing is known about the type of the
126 ;;; XEP arg vars, type checks will be emitted when the XEP's arg vars
127 ;;; are passed to the actual function.
141 ;; KLUDGE: The intent here is to enable tail recursion
142 ;; optimization, since leaving frames for wrapper
143 ;; functions like this on the stack is actually more
144 ;; annoying than helpful for debugging. Unfortunately
145 ;; trying to express this by messing with the
146 ;; ANSI-standard declarations is a little awkward, since
147 ;; no matter how we do it we'll tend to have side-effects
148 ;; on things like SPEED-vs.-SAFETY comparisons. Perhaps
149 ;; it'd be better to define a new SB-EXT:TAIL-RECURSIVELY
150 ;; declaration and use that? -- WHN 2002-07-08
166 ;; FIXME: Make sure that INDEX type distinguishes between
167 ;; target and host. (Probably just make the SB!XC:DEFTYPE
168 ;; different from CL:DEFTYPE.)
179 ;; KLUDGE: As above, we're trying to
180 ;; enable tail recursion optimization and
181 ;; any other effects of this declaration
182 ;; are accidental. -- WHN 2002-07-08
188 ;;; Make an external entry point (XEP) for FUN and return it. We
189 ;;; convert the result of MAKE-XEP-LAMBDA in the correct environment,
190 ;;; then associate this lambda with FUN as its XEP. After the
191 ;;; conversion, we iterate over the function's associated lambdas,
192 ;;; redoing local call analysis so that the XEP calls will get
193 ;;; converted.
194 ;;;
195 ;;; We set REANALYZE and REOPTIMIZE in the component, just in case we
196 ;;; discover an XEP after the initial local call analyze pass.
203 "XEP for ~A"
219 res)))
221 ;;; Notice a REF that is not in a local-call context. If the REF is
222 ;;; already to an XEP, then do nothing, otherwise change it to the
223 ;;; XEP, making an XEP if necessary.
224 ;;;
225 ;;; If REF is to a special :CLEANUP or :ESCAPE function, then we treat
226 ;;; it as though it was not an XEP reference (i.e. leave it alone).
234 \f
235 ;;; Attempt to convert all references to FUN to local calls. The
236 ;;; reference must be the function for a call, and the function
237 ;;; continuation must be used only once, since otherwise we cannot be
238 ;;; sure what function is to be called. The call continuation would be
239 ;;; multiply used if there is hairy stuff such as conditionals in the
240 ;;; expression that computes the function.
241 ;;;
242 ;;; If we cannot convert a reference, then we mark the referenced
243 ;;; function as an entry-point, creating a new XEP if necessary. We
244 ;;; don't try to convert calls that are in error (:ERROR kind.)
245 ;;;
246 ;;; This is broken off from LOCALL-ANALYZE-COMPONENT so that people
247 ;;; can force analysis of newly introduced calls. Note that we don't
248 ;;; do LET conversion here.
270 ;;; We examine all NEW-FUNCTIONALS in COMPONENT, attempting to convert
271 ;;; calls into local calls when it is legal. We also attempt to
272 ;;; convert each LAMBDA to a LET. LET conversion is also triggered by
273 ;;; deletion of a function reference, but functions that start out
274 ;;; eligible for conversion must be noticed sometime.
275 ;;;
276 ;;; Note that there is a lot of action going on behind the scenes
277 ;;; here, triggered by reference deletion. In particular, the
278 ;;; COMPONENT-LAMBDAS are being hacked to remove newly deleted and LET
279 ;;; converted LAMBDAs, so it is important that the LAMBDA is added to
280 ;;; the COMPONENT-LAMBDAS when it is. Also, the
281 ;;; COMPONENT-NEW-FUNCTIONALS may contain all sorts of drivel, since
282 ;;; it is not updated when we delete functions, etc. Only
283 ;;; COMPONENT-LAMBDAS is updated.
284 ;;;
285 ;;; COMPONENT-REANALYZE-FUNCTIONALS is treated similarly to
286 ;;; COMPONENT-NEW-FUNCTIONALS, but we don't add lambdas to the
287 ;;; LAMBDAS.
305 ;; Fix/check FUNCTIONAL's relationship to COMPONENT-LAMDBAS.
307 ;; Since FUNCTIONAL isn't a LAMBDA, this doesn't
308 ;; apply: no-op.
311 ; NEW-FUNCTIONALS, hence is new.
312 ;; FUNCTIONAL becomes part of COMPONENT-LAMBDAS now.
317 ;; FUNCTIONAL should be in COMPONENT-LAMBDAS already.
319 component)))))
331 ;; The original CMU CL code seemed to implicitly assume that
332 ;; COMPONENT is the only one here. Let's make that explicit.
342 ;;; If policy is auspicious and CALL is not in an XEP and we don't seem
343 ;;; to be in an infinite recursive loop, then change the reference to
344 ;;; reference a fresh copy. We return whichever function we decide to
345 ;;; reference.
360 "local inline ~A"
362 original-functional)))))))
366 calls this LET-converted local function:~
367 ~% ~S"
369 original-functional)
372 converted-lambda)))
373 original-functional))
375 ;;; Dispatch to the appropriate function to attempt to convert a call.
376 ;;; REF must be a reference to a FUNCTIONAL. This is called in IR1
377 ;;; optimization as well as in local call analysis. If the call is is
378 ;;; already :LOCAL, we do nothing. If the call is already scheduled
379 ;;; for deletion, also do nothing (in addition to saving time, this
380 ;;; also avoids some problems with optimizing collections of functions
381 ;;; that are partially deleted.)
382 ;;;
383 ;;; This is called both before and after FIND-INITIAL-DFO runs. When
384 ;;; called on a :INITIAL component, we don't care whether the caller
385 ;;; and callee are in the same component. Afterward, we must stick
386 ;;; with whatever component division we have chosen.
387 ;;;
388 ;;; Before attempting to convert a call, we see whether the function
389 ;;; is supposed to be inline expanded. Call conversion proceeds as
390 ;;; before after any expansion.
391 ;;;
392 ;;; We bind *COMPILER-ERROR-CONTEXT* to the node for the call so that
393 ;;; warnings will get the right context.
409 component))))
412 original-fun))
430 ;;; Attempt to convert a multiple-value call. The only interesting
431 ;;; case is a call to a function that LOOKS-LIKE-AN-MV-BIND, has
432 ;;; exactly one reference and no XEP, and is called with one values
433 ;;; continuation.
434 ;;;
435 ;;; We change the call to be to the last optional entry point and
436 ;;; change the call to be local. Due to our preconditions, the call
437 ;;; should eventually be converted to a let, but we can't do that now,
438 ;;; since there may be stray references to the e-p lambda due to
439 ;;; optional defaulting code.
440 ;;;
441 ;;; We also use variable types for the called function to construct an
442 ;;; assertion for the values continuation.
443 ;;;
444 ;;; See CONVERT-CALL for additional notes on MERGE-TAIL-SETS, etc.
464 ;;; Attempt to convert a call to a lambda. If the number of args is
465 ;;; wrong, we give a warning and mark the call as :ERROR to remove it
466 ;;; from future consideration. If the argcount is O.K. then we just
467 ;;; convert it.
475 ;; FIXME: ANSI requires in "3.2.5 Exceptional Situations in the
476 ;; Compiler" that calling a function with "the wrong number of
477 ;; arguments" be only a STYLE-ERROR. I think, though, that this
478 ;; should only apply when the number of arguments is inferred
479 ;; from a previous definition. If the number of arguments
480 ;; is DECLAIMed, surely calling with the wrong number is a
481 ;; real WARNING. As long as SBCL continues to use CMU CL's
482 ;; non-ANSI DEFUN-is-a-DECLAIM policy, we're in violation here,
483 ;; but as long as we continue to use that policy, that's the
484 ;; not our biggest problem.:-| When we fix that policy, this
485 ;; should come back into compliance. (So fix that policy!)
486 ;; ..but..
487 ;; FIXME, continued: Except that section "3.2.2.3 Semantic
488 ;; Constraints" says that if it's within the same file, it's
489 ;; wrong. And we're in locall.lisp here, so it's probably
490 ;; (haven't checked this..) a call to something in the same
491 ;; file. So maybe it deserves a full warning anyway.
493 "function called with ~R argument~:P, but wants exactly ~R"
494 call-args nargs)
496 \f
497 ;;;; &OPTIONAL, &MORE and &KEYWORD calls
499 ;;; This is similar to CONVERT-LAMBDA-CALL, but deals with
500 ;;; OPTIONAL-DISPATCHes. If only fixed args are supplied, then convert
501 ;;; a call to the correct entry point. If &KEY args are supplied, then
502 ;;; dispatch to a subfunction. We don't convert calls to functions
503 ;;; that have a &MORE (or &REST) arg.
511 ;; FIXME: See FIXME note at the previous
512 ;; wrong-number-of-arguments warnings in this file.
514 "function called with ~R argument~:P, but wants at least ~R"
515 call-args min-args)
524 ;; FIXME: See FIXME note at the previous
525 ;; wrong-number-of-arguments warnings in this file.
527 "function called with ~R argument~:P, but wants at most ~R"
528 call-args max-args)
532 ;;; This function is used to convert a call to an entry point when
533 ;;; complex transformations need to be done on the original arguments.
534 ;;; ENTRY is the entry point function that we are calling. VARS is a
535 ;;; list of variable names which are bound to the original call
536 ;;; arguments. IGNORES is the subset of VARS which are ignored. ARGS
537 ;;; is the list of arguments to the entry point function.
538 ;;;
539 ;;; In order to avoid gruesome graph grovelling, we introduce a new
540 ;;; function that rearranges the arguments and calls the entry point.
541 ;;; We analyze the new function and the entry point immediately so
542 ;;; that everything gets converted during the single pass.
559 ;;; Use CONVERT-HAIRY-FUN-ENTRY to convert a &MORE-arg call to a known
560 ;;; function into a local call to the MAIN-ENTRY.
561 ;;;
562 ;;; First we verify that all keywords are constant and legal. If there
563 ;;; aren't, then we warn the user and don't attempt to convert the call.
564 ;;;
565 ;;; We massage the supplied &KEY arguments into the order expected
566 ;;; by the main entry. This is done by binding all the arguments to
567 ;;; the keyword call to variables in the introduced lambda, then
568 ;;; passing these values variables in the correct order when calling
569 ;;; the main entry. Unused arguments (such as the keywords themselves)
570 ;;; are discarded simply by not passing them along.
571 ;;;
572 ;;; If there is a &REST arg, then we bundle up the args and pass them
573 ;;; to LIST.
595 (:keyword
624 ;; FIXME: check whether KEY was supplied earlier
647 loser)
654 (())
658 (:optional
662 (:rest
665 (:keyword
683 \f
684 ;;;; LET conversion
685 ;;;;
686 ;;;; Converting to a LET has differing significance to various parts
687 ;;;; of the compiler:
688 ;;;; -- The body of a LET is spliced in immediately after the
689 ;;;; corresponding combination node, making the control transfer
690 ;;;; explicit and allowing LETs to be mashed together into a single
691 ;;;; block. The value of the LET is delivered directly to the
692 ;;;; original continuation for the call, eliminating the need to
693 ;;;; propagate information from the dummy result continuation.
694 ;;;; -- As far as IR1 optimization is concerned, it is interesting in
695 ;;;; that there is only one expression that the variable can be bound
696 ;;;; to, and this is easily substituted for.
697 ;;;; -- LETs are interesting to environment analysis and to the back
698 ;;;; end because in most ways a LET can be considered to be "the
699 ;;;; same function" as its home function.
700 ;;;; -- LET conversion has dynamic scope implications, since control
701 ;;;; transfers within the same environment are local. In a local
702 ;;;; control transfer, cleanup code must be emitted to remove
703 ;;;; dynamic bindings that are no longer in effect.
705 ;;; Set up the control transfer to the called CLAMBDA. We split the
706 ;;; call block immediately after the call, and link the head of
707 ;;; CLAMBDA to the call block. The successor block after splitting
708 ;;; (where we return to) is returned.
709 ;;;
710 ;;; If the lambda is is a different component than the call, then we
711 ;;; call JOIN-COMPONENTS. This only happens in block compilation
712 ;;; before FIND-INITIAL-DFO.
725 ;; FIXME: Use DESTRUCTURING-BIND here, and grep for other
726 ;; uses of '=.*length' which could also be converted to use
727 ;; DESTRUCTURING-BIND or PROPER-LIST-OF-LENGTH-P.
732 next-block)))
734 ;;; Remove CLAMBDA from the tail set of anything it used to be in the
735 ;;; same set as; but leave CLAMBDA with a valid tail set value of
736 ;;; its own, for the benefit of code which might try to pull
737 ;;; something out of it (e.g. return type).
739 ;; Until sbcl-0.pre7.37.flaky5.2, we did
740 ;; (LET ((TAILS (LAMBDA-TAIL-SET CLAMBDA)))
741 ;; (SETF (TAIL-SET-FUNS TAILS)
742 ;; (DELETE CLAMBDA (TAIL-SET-FUNS TAILS))))
743 ;; (SETF (LAMBDA-TAIL-SET CLAMBDA) NIL)
744 ;; here. Apparently the idea behind the (SETF .. NIL) was that since
745 ;; TAIL-SET-FUNS no longer thinks we're in the tail set, it's
746 ;; inconsistent, and perhaps unsafe, for us to think we're in the
747 ;; tail set. Unfortunately..
748 ;;
749 ;; The (SETF .. NIL) caused problems in sbcl-0.pre7.37.flaky5.2 when
750 ;; I was trying to get Python to emit :EXTERNAL LAMBDAs directly
751 ;; (instead of only being able to emit funny little :TOPLEVEL stubs
752 ;; which you called in order to get the address of an external LAMBDA):
753 ;; the external function was defined in terms of internal function,
754 ;; which was LET-converted, and then things blew up downstream when
755 ;; FINALIZE-XEP-DEFINITION tried to find out its DEFINED-TYPE from
756 ;; the now-NILed-out TAIL-SET. So..
757 ;;
758 ;; To deal with this problem, we no longer NIL out
759 ;; (LAMBDA-TAIL-SET CLAMBDA) here. Instead:
760 ;; * If we're the only function in TAIL-SET-FUNS, it should
761 ;; be safe to leave ourself linked to it, and it to you.
762 ;; * If there are other functions in TAIL-SET-FUNS, then we're
763 ;; afraid of future optimizations on those functions causing
764 ;; the TAIL-SET object no longer to be valid to describe our
765 ;; return value. Thus, we delete ourselves from that object;
766 ;; but we save a newly-allocated tail-set, derived from the old
767 ;; one, for ourselves, for the use of later code (e.g.
768 ;; FINALIZE-XEP-DEFINITION) which might want to
769 ;; know about our return type.
778 ;; The documentation on TAIL-SET-INFO doesn't tell whether it could
779 ;; remain valid in this case, so we nuke it on the theory that
780 ;; missing information tends to be less dangerous than incorrect
781 ;; information.
784 ;;; Handle the PHYSENV semantics of LET conversion. We add CLAMBDA and
785 ;;; its LETs to LETs for the CALL's home function. We merge the calls
786 ;;; for CLAMBDA with the calls for the home function, removing CLAMBDA
787 ;;; in the process. We also merge the ENTRIES.
788 ;;;
789 ;;; We also unlink the function head from the component head and set
790 ;;; COMPONENT-REANALYZE to true to indicate that the DFO should be
791 ;;; recomputed.
810 ;; CLAMBDA belongs to HOME now.
815 ;; All of CLAMBDA's LETs belong to HOME now.
821 ;; CLAMBDA no longer has an independent existence as an entity
822 ;; which has LETs.
825 ;; HOME no longer calls CLAMBDA, and owns all of CLAMBDA's old
826 ;; DFO dependencies.
831 ;; CLAMBDA no longer has an independent existence as an entity
832 ;; which calls things or has DFO dependencies.
835 ;; All of CLAMBDA's ENTRIES belong to HOME now.
839 ;; CLAMBDA no longer has an independent existence as an entity
840 ;; with ENTRIES.
845 ;;; Handle the value semantics of LET conversion. Delete FUN's return
846 ;;; node, and change the control flow to transfer to NEXT-BLOCK
847 ;;; instead. Move all the uses of the result continuation to CALL's
848 ;;; CONT.
849 ;;;
850 ;;; If the actual continuation is only used by the LET call, then we
851 ;;; intersect the type assertion on the dummy continuation with the
852 ;;; assertion for the actual continuation; in all other cases
853 ;;; assertions on the dummy continuation are lost.
854 ;;;
855 ;;; We also intersect the derived type of the CALL with the derived
856 ;;; type of all the dummy continuation's uses. This serves mainly to
857 ;;; propagate TRULY-THE through LETs.
873 cont
882 ;;; Change all CONT for all the calls to FUN to be the start
883 ;;; continuation for the bind node. This allows the blocks to be
884 ;;; joined if the caller count ever goes to one.
894 ;;; We are converting FUN to be a LET when the call is in a non-tail
895 ;;; position. Any previously tail calls in FUN are no longer tail
896 ;;; calls, and must be restored to normal calls which transfer to
897 ;;; NEXT-BLOCK (FUN's return point.) We can't do this by DO-USES on
898 ;;; the RETURN-RESULT, because the return might have been deleted (if
899 ;;; all calls were TR.)
919 ;; The called function might be an assignment in the
920 ;; case where we are currently converting that function.
921 ;; In steady-state, assignments never appear as a called
922 ;; function.
923 (:assignment
927 ;;; Deal with returning from a LET or assignment that we are
928 ;;; converting. FUN is the function we are calling, CALL is a call to
929 ;;; FUN, and NEXT-BLOCK is the return point for a non-tail call, or
930 ;;; NULL if call is a tail call.
931 ;;;
932 ;;; If the call is not a tail call, then we must do
933 ;;; UNCONVERT-TAIL-CALLS, since a tail call is a call which returns
934 ;;; its value out of the enclosing non-let function. When call is
935 ;;; non-TR, we must convert it back to an ordinary local call, since
936 ;;; the value must be delivered to the receiver of CALL's value.
937 ;;;
938 ;;; We do different things depending on whether the caller and callee
939 ;;; have returns left:
941 ;;; -- If the callee has no return we just do MOVE-LET-CALL-CONT.
942 ;;; Either the function doesn't return, or all returns are via
943 ;;; tail-recursive local calls.
944 ;;; -- If CALL is a non-tail call, or if both have returns, then
945 ;;; we delete the callee's return, move its uses to the call's
946 ;;; result continuation, and transfer control to the appropriate
947 ;;; return point.
948 ;;; -- If the callee has a return, but the caller doesn't, then we
949 ;;; move the return to the caller.
970 ;;; Actually do LET conversion. We call subfunctions to do most of the
971 ;;; work. We change the CALL's CONT to be the continuation heading the
972 ;;; BIND block, and also do REOPTIMIZE-CONTINUATION on the args and
973 ;;; CONT so that LET-specific IR1 optimizations get a chance. We blow
974 ;;; away any entry for the function in *FREE-FUNS* so that nobody
975 ;;; will create new references to it.
979 nil
984 ;;; Reoptimize all of CALL's args and its result.
993 ;;; Are there any declarations in force to say CLAMBDA shouldn't be
994 ;;; LET converted?
996 ;; From the user's point of view, LET-converting something that
997 ;; has a name is inlining it. (The user can't see what we're doing
998 ;; with anonymous things, and suppressing inlining
999 ;; for such things can easily give Python acute indigestion, so
1000 ;; we don't.)
1002 ;; ANSI requires that explicit NOTINLINE be respected.
1004 ;; If (> DEBUG SPEED) we can guess that inlining generally
1005 ;; won't be appreciated, but if the user specifically requests
1006 ;; inlining, that takes precedence over our general guess.
1010 ;;; We also don't convert calls to named functions which appear in the
1011 ;;; initial component, delaying this until optimization. This
1012 ;;; minimizes the likelihood that we will LET-convert a function which
1013 ;;; may have references added due to later local inline expansion.
1020 ;;; This function is called when there is some reason to believe that
1021 ;;; CLAMBDA might be converted into a LET. This is done after local
1022 ;;; call analysis, and also when a reference is deleted. We return
1023 ;;; true if we converted.
1027 ;; We only convert to a LET when the function is a normal local
1028 ;; function, has no XEP, and is referenced in exactly one local
1029 ;; call. Conversion is also inhibited if the only reference is in
1030 ;; a block about to be deleted.
1031 ;;
1032 ;; These rules limiting LET conversion may seem unnecessarily
1033 ;; restrictive, since there are some cases where we could do the
1034 ;; return with a jump that don't satisfy these requirements. The
1035 ;; reason for doing things this way is that it makes the concept
1036 ;; of a LET much more useful at the level of IR1 semantics. The
1037 ;; :ASSIGNMENT function kind provides another way to optimize
1038 ;; calls to single-return/multiple call functions.
1039 ;;
1040 ;; We don't attempt to convert calls to functions that have an
1041 ;; XEP, since we might be embarrassed later when we want to
1042 ;; convert a newly discovered local call. Also, see
1043 ;; OK-INITIAL-CONVERT-P.
1060 nil)))
1069 t))))
1070 \f
1071 ;;;; tail local calls and assignments
1073 ;;; Return T if there are no cleanups between BLOCK1 and BLOCK2, or if
1074 ;;; they definitely won't generate any cleanup code. Currently we
1075 ;;; recognize lexical entry points that are only used locally (if at
1076 ;;; all).
1090 ;;; If a potentially TR local call really is TR, then convert it to
1091 ;;; jump directly to the called function. We also call
1092 ;;; MAYBE-CONVERT-TO-ASSIGNMENT. The first value is true if we
1093 ;;; tail-convert. The second is the value of M-C-T-A. We can switch
1094 ;;; the succesor (potentially deleting the RETURN node) unless:
1095 ;;; -- The call has already been converted.
1096 ;;; -- The call isn't TR (some implicit MV PROG1.)
1097 ;;; -- The call is in an XEP (thus we might decide to make it non-tail
1098 ;;; so that we can use known return inside the component.)
1099 ;;; -- There is a change in the cleanup between the call in the return,
1100 ;;; so we might need to introduce cleanup code.
1119 ;;; This is called when we believe it might make sense to convert
1120 ;;; CLAMBDA to an assignment. All this function really does is
1121 ;;; determine when a function with more than one call can still be
1122 ;;; combined with the calling function's environment. We can convert
1123 ;;; when:
1124 ;;; -- The function is a normal, non-entry function, and
1125 ;;; -- Except for one call, all calls must be tail recursive calls
1126 ;;; in the called function (i.e. are self-recursive tail calls)
1127 ;;; -- OK-INITIAL-CONVERT-P is true.
1128 ;;;
1129 ;;; There may be one outside call, and it need not be tail-recursive.
1130 ;;; Since all tail local calls have already been converted to direct
1131 ;;; transfers, the only control semantics needed are to splice in the
1132 ;;; body at the non-tail call. If there is no non-tail call, then we
1133 ;;; need only merge the environments. Both cases are handled by
1134 ;;; LET-CONVERT.
1135 ;;;
1136 ;;; ### It would actually be possible to allow any number of outside
1137 ;;; calls as long as they all return to the same place (i.e. have the
1138 ;;; same conceptual continuation.) A special case of this would be
1139 ;;; when all of the outside calls are tail recursive.
1165 t)
1167 nil))))))