| blob | 622a9242292deec405ac738a6625c8421dc29a10 |
1 ;;;; This file implements the IR1 optimization phase of the compiler.
2 ;;;; IR1 optimization is a grab-bag of optimizations that don't make
3 ;;;; major changes to the block-level control flow and don't use flow
4 ;;;; analysis. These optimizations can mostly be classified as
5 ;;;; "meta-evaluation", but there is a sizable top-down component as
6 ;;;; well.
8 ;;;; This software is part of the SBCL system. See the README file for
9 ;;;; more information.
10 ;;;;
11 ;;;; This software is derived from the CMU CL system, which was
12 ;;;; written at Carnegie Mellon University and released into the
13 ;;;; public domain. The software is in the public domain and is
14 ;;;; provided with absolutely no warranty. See the COPYING and CREDITS
15 ;;;; files for more information.
18 \f
19 ;;;; interface for obtaining results of constant folding
21 ;;; Return true for an LVAR whose sole use is a reference to a
22 ;;; constant leaf.
29 leaf)
32 ;; LEAF may be a CONSTANT behind a cast that will
33 ;; later turn out to be of the wrong type.
34 ;; And ir1-transforms suffer from this because
35 ;; they expect LVAR-VALUE to be of a restricted type.
38 ;; check for EQL types and singleton numeric types
41 ;;; Return the constant value for an LVAR whose only use is a constant
42 ;;; node.
47 leaf)
54 value))))
55 \f
56 ;;;; interface for obtaining results of type inference
58 ;;; Our best guess for the type of this lvar's value. Note that this
59 ;;; may be VALUES or FUNCTION type, which cannot be passed as an
60 ;;; argument to the normal type operations. See LVAR-TYPE.
61 ;;;
62 ;;; The result value is cached in the LVAR-%DERIVED-TYPE slot. If the
63 ;;; slot is true, just return that value, otherwise recompute and
64 ;;; stash the value there.
74 res))
77 res)))
84 ;;; Return the derived type for LVAR's first value. This is guaranteed
85 ;;; not to be a VALUES or FUNCTION type.
90 ;;; LVAR-CONSERVATIVE-TYPE
91 ;;;
92 ;;; Certain types refer to the contents of an object, which can
93 ;;; change without type derivation noticing: CONS types and ARRAY
94 ;;; types suffer from this:
95 ;;;
96 ;;; (let ((x (the (cons fixnum fixnum) (cons a b))))
97 ;;; (setf (car x) c)
98 ;;; (+ (car x) (cdr x)))
99 ;;;
100 ;;; Python doesn't realize that the SETF CAR can change the type of X -- so we
101 ;;; cannot use LVAR-TYPE which gets the derived results. Worse, still, instead
102 ;;; of (SETF CAR) we might have a call to a user-defined function FOO which
103 ;;; does the same -- so there is no way to use the derived information in
104 ;;; general.
105 ;;;
106 ;;; So, the conservative option is to use the derived type if the leaf has
107 ;;; only a single ref -- in which case there cannot be a prior call that
108 ;;; mutates it. Otherwise we use the declared type or punt to the most general
109 ;;; type we know to be correct for sure.
113 ;; Recompute using NODE-CONSERVATIVE-TYPE instead of derived type if
114 ;; necessary -- picking off some easy cases up front.
116 ;; Can't use CSUBTYPEP!
119 derived-type)
123 derived-type)
129 derived-type)
133 derived-type))))
146 derived-values-type))
147 derived-values-type)))
153 type)
159 ;; ADJUST-ARRAY may change dimensions, but rank stays same.
162 old
164 ;; Complexity cannot change.
166 ;; Element type cannot change.
169 ;; Simple arrays cannot change at all.
170 type))
172 ;; Conservative union type is an union of conservative types.
177 ;; Catch-all.
178 ;;
179 ;; If the type contains some CONS types, the conservative type contains all
180 ;; of them.
183 ;; Similarly for non-simple arrays -- it should be possible to preserve
184 ;; more information here, but really...
188 type)))
192 ;; Excluding T is necessary, because we do want type derivation to
193 ;; be able to narrow it down in case someone (most like a macro-expansion...)
194 ;; actually declares something as having type T.
195 nil)
197 ;; Covered by the next case as well, but this is a quick test.
198 t)
200 t)))
202 ;;; If LVAR is an argument of a function, return a type which the
203 ;;; function checks LVAR for.
212 ;; TODO: MV-COMBINATION
220 ;; FUN-TYPE might be (AND FUNCTION (SATISFIES ...)).
233 dest)))))
235 \f
236 ;;;; interface routines used by optimizers
238 ;;; This function is called by optimizers to indicate that something
239 ;;; interesting has happened to the value of LVAR. Optimizers must
240 ;;; make sure that they don't call for reoptimization when nothing has
241 ;;; happened, since optimization will fail to terminate.
242 ;;;
243 ;;; We clear any cached type for the lvar and set the reoptimize flags
244 ;;; on everything in sight.
254 ;; PREV may be missing.
273 ;;; Annotate NODE to indicate that its result has been proven to be
274 ;;; TYPEP to RTYPE. After IR1 conversion has happened, this is the
275 ;;; only correct way to supply information discovered about a node's
276 ;;; type. If you screw with the NODE-DERIVED-TYPE directly, then
277 ;;; information may be lost and reoptimization may not happen.
278 ;;;
279 ;;; What we do is intersect RTYPE with NODE's DERIVED-TYPE. If the
280 ;;; intersection is different from the old type, then we do a
281 ;;; REOPTIMIZE-LVAR on the NODE-LVAR.
287 initial-type)))
298 "New inferred type ~S conflicts with old type:~
299 ~% ~S~%*** possible internal error? Please report this."
302 ;; If the new type consists of only one object, replace the
303 ;; node with a constant reference.
314 ;;; This is similar to DERIVE-NODE-TYPE, but asserts that it is an
315 ;;; error for LVAR's value not to be TYPEP to TYPE. We implement it
316 ;;; splitting off DEST a new CAST node; old LVAR will deliver values
317 ;;; to CAST. If we improve the assertion, we set TYPE-CHECK and
318 ;;; TYPE-ASSERTED to guarantee that the new assertion will be checked.
326 context)))
328 t))))
336 :value lvar
338 :arg-count arg-count
350 type-check type-asserted)
351 t)
352 cast)))
360 caller policy)))
363 \f
364 ;;;; IR1-OPTIMIZE
366 ;;; Do one forward pass over COMPONENT, deleting unreachable blocks
367 ;;; and doing IR1 optimizations. We can ignore all blocks that don't
368 ;;; have the REOPTIMIZE flag set. If COMPONENT-REOPTIMIZE is true when
369 ;;; we are done, then another iteration would be beneficial.
375 for last-block = block
378 ;; We delete blocks when there is either no predecessor or the
379 ;; block is in a lambda that has been deleted. These blocks
380 ;; would eventually be deleted by DFO recomputation, but doing
381 ;; it here immediately makes the effect available to IR1
382 ;; optimization.
388 ;; Preserve the BLOCK-SUCC invariant that almost every block has
389 ;; one successor (and a block with DELETE-P set is an acceptable
390 ;; exception).
425 ;;; Loop over the nodes in BLOCK, acting on (and clearing) REOPTIMIZE
426 ;;; flags.
427 ;;;
428 ;;; Note that although they are cleared here, REOPTIMIZE flags might
429 ;;; still be set upon return from this function, meaning that further
430 ;;; optimization is wanted (as a consequence of optimizations we did).
433 ;; We clear the node and block REOPTIMIZE flags before doing the
434 ;; optimization, not after. This ensures that the node or block will
435 ;; be reoptimized if necessary.
439 ;; As above, we clear the node REOPTIMIZE flag before optimizing.
444 ;; With a COMBINATION, we call PROPAGATE-FUN-CHANGE whenever
445 ;; the function changes, and call IR1-OPTIMIZE-COMBINATION if
446 ;; any argument changes.
451 ;; KLUDGE: We leave the NODE-OPTIMIZE flag set going into
452 ;; IR1-OPTIMIZE-RETURN, since IR1-OPTIMIZE-RETURN wants to
453 ;; clear the flag itself. -- WHN 2002-02-02, quoting original
454 ;; CMU CL comments
460 ;; With an EXIT, we derive the node's type from the VALUE's
461 ;; type.
466 ;; PROPAGATE-FROM-SETS can do a better job if NODE-REOPTIMIZE
467 ;; is accurate till the node actually has been reoptimized.
475 ;;; Try to join with a successor block. If we succeed, we return true,
476 ;;; otherwise false.
483 ;; the successor has more than one predecessor;
485 ;; the successor is the current block (infinite loop);
487 ;; the next block has a different cleanup, and thus
488 ;; we may want to insert cleanup code between the
489 ;; two blocks at some point;
492 ;; the next block has a different home lambda, and
493 ;; thus the control transfer is a non-local exit.
496 ;; Stack analysis phase wants ENTRY to start a block...
502 ;; ... and a DX-allocator to end a block.
504 ;; ... and for there to be no chance of there
505 ;; being two successive USEs of the same
506 ;; multi-valued LVAR in the same block (since
507 ;; we can only insert cleanup code at block
508 ;; boundaries, but need to discard
509 ;; multi-valued LVAR contents before they are
510 ;; overwritten).
513 nil)
516 t)))))
518 ;;; Join together two blocks. The code in BLOCK2 is moved into BLOCK1
519 ;;; and BLOCK2 is deleted from the DFO. We combine the optimize flags
520 ;;; for the two blocks so that any indicated optimization gets done.
553 ;;; Delete any nodes in BLOCK whose value is unused and which have no
554 ;;; side effects. We can delete sets of lexical variables when the set
555 ;;; variable has no references.
601 victim)
602 \f
603 ;;;; local call return type propagation
605 ;;; This function is called on RETURN nodes that have their REOPTIMIZE
606 ;;; flag set. It iterates over the uses of the RESULT, looking for
607 ;;; interesting stuff to update the TAIL-SET. If a use isn't a local
608 ;;; call, then we union its type together with the types of other such
609 ;;; uses. We assign to the RETURN-RESULT-TYPE the intersection of this
610 ;;; type with the RESULT's asserted type. We can make this
611 ;;; intersection now (potentially before type checking) because this
612 ;;; assertion on the result will eventually be checked (if
613 ;;; appropriate.)
614 ;;;
615 ;;; We call MAYBE-CONVERT-TAIL-LOCAL-CALL on each local non-MV
616 ;;; combination, which may change the successor of the call to be the
617 ;;; called function, and if so, checks if the call can become an
618 ;;; assignment. If we convert to an assignment, we abort, since the
619 ;;; RETURN has been deleted.
638 ;; (values-type-intersection
639 ;; (continuation-asserted-type result) ; FIXME -- APD, 2002-01-26
641 ;; )
642 ))
644 nil)
646 ;;; Do stuff to realize that something has changed about the value
647 ;;; delivered to a return node. Since we consider the return values of
648 ;;; all functions in the tail set to be equivalent, this amounts to
649 ;;; bringing the entire tail set up to date. We iterate over the
650 ;;; returns for all the functions in the tail set, reanalyzing them
651 ;;; all (not treating NODE specially.)
652 ;;;
653 ;;; When we are done, we check whether the new type is different from
654 ;;; the old TAIL-SET-TYPE. If so, we set the type and also reoptimize
655 ;;; all the lvars for references to functions in the tail set. This
656 ;;; will cause IR1-OPTIMIZE-COMBINATION to derive the new type as the
657 ;;; results of the calls.
661 :restart
681 \f
682 ;;;; IF optimization
684 ;;; Utility: return T if both argument cblocks are equivalent. For now,
685 ;;; detect only blocks that read the same leaf into the same lvar, and
686 ;;; continue to the same block.
699 ;;; Check whether the predicate is known to be true or false,
700 ;;; deleting the IF node in favor of the appropriate branch when this
701 ;;; is the case.
702 ;;; Similarly, when both branches are equivalent, branch directly to either
703 ;;; of them.
704 ;;; Also, if the test has multiple uses, replicate the node when possible...
705 ;;; in fact, splice in direct jumps to the right branch if possible.
717 alternative)
719 consequent)
722 alternative))))
730 ;; When we know that we only have a single successor, kill the victim
731 ;; ... unless the victim and the remaining successor are the same.
740 ;; When if/if conversion would leave (if ... (if nil ...)) or
741 ;; (if ... (if not-nil ...)), splice the correct successor right
742 ;; in.
761 ;; only one use left. Just kill the now-useless
762 ;; branch to avoid spurious code deletion notes.
765 node test block
771 ;; Finally, duplicate EQ-nil tests
778 ;; leave the last use as is, instead of replacing
779 ;; the (singly-referenced) CIF node with a duplicate.
782 ;;; Create a new copy of an IF node that tests the value of the node
783 ;;; USE. The test must have >1 use, and must be immediately used by
784 ;;; USE. NODE must be the only node in its block (implying that
785 ;;; block-start = if-test).
786 ;;;
787 ;;; This optimization has an effect semantically similar to the
788 ;;; source-to-source transformation:
789 ;;; (IF (IF A B C) D E) ==>
790 ;;; (IF A (IF B D E) (IF C D E))
791 ;;;
792 ;;; We clobber the NODE-SOURCE-PATH of both the original and the new
793 ;;; node so that dead code deletion notes will definitely not consider
794 ;;; either node to be part of the original source. One node might
795 ;;; become unreachable, resulting in a spurious note.
807 :consequent cblock
829 \f
830 ;;;; exit IR1 optimization
832 ;;; This function attempts to delete an exit node, returning true if
833 ;;; it deletes the block as a consequence:
834 ;;; -- If the exit is degenerate (has no ENTRY), then we don't do
835 ;;; anything, since there is nothing to be done.
836 ;;; -- If the exit node and its ENTRY have the same home lambda then
837 ;;; we know the exit is local, and can delete the exit. We change
838 ;;; uses of the Exit-Value to be uses of the original lvar,
839 ;;; then unlink the node. If the exit is to a TR context, then we
840 ;;; must do MERGE-TAIL-SETS on any local calls which delivered
841 ;;; their value to this exit.
842 ;;; -- If there is no value (as in a GO), then we skip the value
843 ;;; semantics.
844 ;;;
845 ;;; This function is also called by environment analysis, since it
846 ;;; wants all exits to be optimized even if normal optimization was
847 ;;; omitted.
857 ;; We can't simply use DELETE-FILTER to unlink the node
858 ;; and substitute some LVAR magic, as this can confuse the
859 ;; stack analysis if there's another EXIT to the same
860 ;; continuation. Instead, we fabricate a new block (in
861 ;; the same lexenv as the ENTRY, so it can't be merged
862 ;; backwards), insert a gimmicked CAST node to link up the
863 ;; LVAR holding the value being returned to the LVAR which
864 ;; is expecting to accept the value, thus placing the
865 ;; return value where it needs to be while still providing
866 ;; the hook required for stack analysis.
867 ;; -- AJB, 2014-Mar-03
872 (%make-cast
875 :value value
879 ;; We only expect a single successor to EXIT-BLOCK,
880 ;; because it contains an EXIT node (which must end its
881 ;; block) and the only blocks that have more than once
882 ;; successor are those with IF nodes (which also must
883 ;; end their blocks). Still, just to be sure, we use a
884 ;; construct that guarantees an error if this
885 ;; expectation is violated.
890 ;; Finish creating the new block.
894 ;; Link the new block into the control sequence.
899 ;; Finish re-pointing the value-holding LVAR to the
900 ;; CAST node.
905 ;; Register the CAST node as providing a value to the
906 ;; LVAR for the continuation.
910 ;; Remove the EXIT node.
913 ;; And, because we created a new block, we need to
914 ;; force component reanalysis (to assign a DFO number
915 ;; to the block if nothing else).
919 \f
920 ;;;; combination IR1 optimization
922 ;;; Report as we try each transform?
923 #!+sb-show
931 "The return value of ~A should not be discarded."
934 ;;; Do IR1 optimizations on a COMBINATION node.
944 (:local
949 (:error
953 (:full
974 ;; Check against the DEFINED-TYPE unless TYPE is already good.
984 (:known
992 ;; If somebody is really sure that they want to modify
993 ;; constants, let them.
1022 ;; First give the VM a peek at the call
1026 (:direct
1027 ;; The VM knows how to handle this.
1028 )
1029 (:transform
1030 ;; The VM mostly knows how to handle this. We need
1031 ;; to massage the call slightly, though.
1034 ;; Let transforms have a crack at it.
1036 #!+sb-show
1042 #!+sb-show
1057 ;;; If NODE doesn't return (i.e. return type is NIL), then terminate
1058 ;;; the block there, and link it to the component tail.
1059 ;;;
1060 ;;; Except when called during IR1 convertion, we delete the
1061 ;;; continuation if it has no other uses. (If it does have other uses,
1062 ;;; we reoptimize.)
1063 ;;;
1064 ;;; Termination on the basis of a continuation type is
1065 ;;; inhibited when:
1066 ;;; -- The continuation is deleted (hence the assertion is spurious), or
1067 ;;; -- We are in IR1 conversion (where THE assertions are subject to
1068 ;;; weakening.) FIXME: Now THE assertions are not weakened, but new
1069 ;;; uses can(?) be added later. -- APD, 2003-07-17
1070 ;;;
1071 ;;; Why do we need to consider LVAR type? -- APD, 2003-07-30
1082 ;; Even if the combination will never return, don't terminate if this
1083 ;; is the tail call of a XEP: doing that would inhibit TCO.
1110 t))))
1112 ;;; This is called both by IR1 conversion and IR1 optimization when
1113 ;;; they have verified the type signature for the call, and are
1114 ;;; wondering if something should be done to special-case the call. If
1115 ;;; CALL is a call to a global function, then see whether it defined
1116 ;;; or known:
1117 ;;; -- If a DEFINED-FUN should be inline expanded, then convert
1118 ;;; the expansion and change the call to call it. Expansion is
1119 ;;; enabled if :INLINE or if SPACE=0. If the FUNCTIONAL slot is
1120 ;;; true, we never expand, since this function has already been
1121 ;;; converted. Local call analysis will duplicate the definition
1122 ;;; if necessary. We claim that the parent form is LABELS for
1123 ;;; context declarations, since we don't want it to be considered
1124 ;;; a real global function.
1125 ;;; -- If it is a known function, mark it as such by setting the KIND.
1126 ;;;
1127 ;;; We return the leaf referenced (NIL if not a leaf) and the
1128 ;;; FUN-INFO assigned.
1152 ;; Inline: if the function has already been converted at another call
1153 ;; site in this component, we point this REF to the functional. If not,
1154 ;; we convert the expansion.
1155 ;;
1156 ;; For :INLINE case local call analysis will copy the expansion later,
1157 ;; but for :MAYBE-INLINE and NIL cases we only get one copy of the
1158 ;; expansion per component.
1159 ;;
1160 ;; FIXME: We also convert in :INLINE & FUNCTIONAL-KIND case below. What
1161 ;; is it for?
1166 ;; Convert.
1169 name
1171 leaf
1172 inlinep
1174 ;; Allow backward references to this function from following
1175 ;; forms. (Reused only if policy matches.)
1180 ;; If we've already converted, change ref to the converted
1181 ;; functional.
1193 ;;; Check whether CALL satisfies TYPE. If so, apply the type to the
1194 ;;; call, and do MAYBE-TERMINATE-BLOCK and return the values of
1195 ;;; RECOGNIZE-KNOWN-CALL. If an error, set the combination kind and
1196 ;;; return NIL, NIL. If the type is just FUNCTION, then skip the
1197 ;;; syntax check, arg/result type processing, but still call
1198 ;;; RECOGNIZE-KNOWN-CALL, since the call might be to a known lambda,
1199 ;;; and that checking is done by local call analysis.
1208 ;; Using the defined-type too early is a bit of a waste: during
1209 ;; conversion we cannot use the untrusted ASSERT-CALL-TYPE, etc.
1214 ;; Don't validate multiple times against the same leaf --
1215 ;; it doesn't add any information, but may generate the same warning
1216 ;; multiple times.
1220 :result-test nil
1222 #'compiler-warn
1225 same-file-p)
1232 :result-test nil
1242 ;;; This is called by IR1-OPTIMIZE when the function for a call has
1243 ;;; changed. If the call is local, we try to LET-convert it, and
1244 ;;; derive the result type. If it is a :FULL call, we validate it
1245 ;;; against the type, which recognizes known calls, does inline
1246 ;;; expansion, etc. If a call to a predicate in a non-conditional
1247 ;;; position or to a function with a source transform, then we
1248 ;;; reconvert the form to give IR1 another chance.
1255 (:local
1260 (:full
1268 call))
1276 predicate)
1290 \f
1291 ;;;; known function optimization
1293 ;;; Add a failed optimization note to FAILED-OPTIMZATIONS for NODE,
1294 ;;; FUN and ARGS. If there is already a note for NODE and TRANSFORM,
1295 ;;; replace it, otherwise add a new one.
1306 ;;; Attempt to transform NODE using TRANSFORM-FUNCTION, subject to the
1307 ;;; call type constraint TRANSFORM-TYPE. If we are inhibited from
1308 ;;; doing the transform for some reason and FLAME is true, then we
1309 ;;; make a note of the message in FAILED-OPTIMIZATIONS for IR1
1310 ;;; finalize to pick up. We return true if the transform failed, and
1311 ;;; thus further transformation should be attempted. We return false
1312 ;;; if either the transform succeeded or was aborted.
1336 (:none
1338 nil)
1339 (:aborted
1344 nil)
1345 (:failure
1351 t)
1352 (:delayed
1354 nil))))
1357 type
1361 t)
1363 t))))
1365 ;;; When we don't like an IR1 transform, we throw the severity/reason
1366 ;;; and args.
1367 ;;;
1368 ;;; GIVE-UP-IR1-TRANSFORM is used to throw out of an IR1 transform,
1369 ;;; aborting this attempt to transform the call, but admitting the
1370 ;;; possibility that this or some other transform will later succeed.
1371 ;;; If arguments are supplied, they are format arguments for an
1372 ;;; efficiency note.
1373 ;;;
1374 ;;; ABORT-IR1-TRANSFORM is used to throw out of an IR1 transform and
1375 ;;; force a normal call to the function at run time. No further
1376 ;;; optimizations will be attempted.
1377 ;;;
1378 ;;; DELAY-IR1-TRANSFORM is used to throw out of an IR1 transform, and
1379 ;;; delay the transform on the node until later. REASONS specifies
1380 ;;; when the transform will be later retried. The :OPTIMIZE reason
1381 ;;; causes the transform to be delayed until after the current IR1
1382 ;;; optimization pass. The :CONSTRAINT reason causes the transform to
1383 ;;; be delayed until after constraint propagation.
1384 ;;;
1385 ;;; FIXME: Now (0.6.11.44) that there are 4 variants of this (GIVE-UP,
1386 ;;; ABORT, DELAY/:OPTIMIZE, DELAY/:CONSTRAINT) and we're starting to
1387 ;;; do CASE operations on the various REASON values, it might be a
1388 ;;; good idea to go OO, representing the reasons by objects, using
1389 ;;; CLOS methods on the objects instead of CASE, and (possibly) using
1390 ;;; SIGNAL instead of THROW.
1408 ;;; Poor man's catching and resignalling
1409 ;;; Implicit %GIVE-UP macrolet will resignal the give-up "condition"
1419 (:delayed
1427 ;;; Clear any delayed transform with no reasons - these should have
1428 ;;; been tried in the last pass. Then remove the reason from the
1429 ;;; delayed transform reasons, and if any become empty then set
1430 ;;; reoptimize flags for the node. Return true if any transforms are
1431 ;;; to be retried.
1447 reoptimize))
1449 ;;; Take the lambda-expression RES, IR1 convert it in the proper
1450 ;;; environment, and then install it as the function for the call
1451 ;;; NODE. We do local call analysis so that the new function is
1452 ;;; integrated into the control flow.
1453 ;;;
1454 ;;; We require the original function source name in order to generate
1455 ;;; a meaningful debug name for the lambda we set up. (It'd be
1456 ;;; possible to do this starting from debug names as well as source
1457 ;;; names, but as of sbcl-0.7.1.5, there was no need for this
1458 ;;; generality, since source names are always known to our callers.)
1466 :policy
1467 ;; The internal variables of a transform are not going to be
1468 ;; interesting to the debugger, so there's no sense in
1469 ;; suppressing the substitution of variables with only one use
1470 ;; (the extra variables can slow down constraint propagation).
1472 preserve-single-use-debug-variables
1473 0
1480 res
1492 t)
1501 ;;; Return T if the function is foldable and if it's marked as CALL
1502 ;;; all function arguments are FOLDABLE too.
1508 nil)
1513 args)
1518 combination)))
1522 ;;; Replace a call to a foldable function of constant arguments with
1523 ;;; the result of evaluating the form. If there is an error during the
1524 ;;; evaluation, we give a warning and leave the call alone, making the
1525 ;;; call a :ERROR call.
1526 ;;;
1527 ;;; If there is more than one value, then we transform the call into a
1528 ;;; VALUES form.
1539 args
1540 call
1541 ;; Note: CMU CL had COMPILER-WARN here, and that
1542 ;; seems more natural, but it's probably not.
1543 ;;
1544 ;; It's especially not while bug 173 exists:
1545 ;; Expressions like
1546 ;; (COND (END
1547 ;; (UNLESS (OR UNSAFE? (<= END SIZE)))
1548 ;; ...))
1549 ;; can cause constant-folding TYPE-ERRORs (in
1550 ;; #'<=) when END can be proved to be NIL, even
1551 ;; though the code is perfectly legal and safe
1552 ;; because a NIL value of END means that the
1553 ;; #'<= will never be executed.
1554 ;;
1555 ;; Moreover, even without bug 173,
1556 ;; quite-possibly-valid code like
1557 ;; (COND ((NONINLINED-PREDICATE END)
1558 ;; (UNLESS (<= END SIZE))
1559 ;; ...))
1560 ;; (where NONINLINED-PREDICATE is something the
1561 ;; compiler can't do at compile time, but which
1562 ;; turns out to make the #'<= expression
1563 ;; unreachable when END=NIL) could cause errors
1564 ;; when the compiler tries to constant-fold (<=
1565 ;; END SIZE).
1566 ;;
1567 ;; So, with or without bug 173, it'd be
1568 ;; unnecessarily evil to do a full
1569 ;; COMPILER-WARNING (and thus return FAILURE-P=T
1570 ;; from COMPILE-FILE) for legal code, so we we
1571 ;; use a wimpier COMPILE-STYLE-WARNING instead.
1573 ;; On the other hand, for code we control, we
1574 ;; should be able to work around any bug
1575 ;; 173-related problems, and in particular we
1576 ;; want to be alerted to calls to our own
1577 ;; functions which aren't being folded away; a
1578 ;; COMPILER-WARNING is butch enough to stop the
1579 ;; SBCL build itself in its tracks.
1599 call
1603 fun-name))))))
1605 \f
1606 ;;;; local call optimization
1608 ;;; Propagate TYPE to LEAF and its REFS, marking things changed.
1609 ;;;
1610 ;;; If the leaf type is a function type, then just leave it alone, since TYPE
1611 ;;; is never going to be more specific than that (and TYPE-INTERSECTION would
1612 ;;; choke.)
1613 ;;;
1614 ;;; Also, if the type is one requiring special care don't touch it if the leaf
1615 ;;; has multiple references -- otherwise LVAR-CONSERVATIVE-TYPE is screwed.
1630 ;; KLUDGE: LET var substitution
1636 ;;; Iteration variable: exactly one SETQ of the form:
1637 ;;;
1638 ;;; (let ((var initial))
1639 ;;; ...
1640 ;;; (setq var (+ var step))
1641 ;;; ...)
1667 ;; Detect cases like (LOOP FOR 1.0 to 5.0 ...), where
1668 ;; the initial and the step are of different types,
1669 ;; and the step is less contagious.
1672 step-type))))
1711 :low low
1712 :high high
1715 "check for iteration variable reoptimization"
1725 ;;; Figure out the type of a LET variable that has sets. We compute
1726 ;;; the union of the INITIAL-TYPE and the types of all the set
1727 ;;; values and to a PROPAGATE-TO-REFS with this type.
1747 ;;; If a LET variable, find the initial value's type and do
1748 ;;; PROPAGATE-FROM-SETS. We also derive the VALUE's type as the node's
1749 ;;; type.
1765 ;;; Return true if the value of REF will always be the same (and is
1766 ;;; thus legal to substitute.)
1778 (:global-function
1785 ;;; If we have a non-set LET var with a single use, then (if possible)
1786 ;;; replace the variable reference's LVAR with the arg lvar.
1787 ;;;
1788 ;;; We change the REF to be a reference to NIL with unused value, and
1789 ;;; let it be flushed as dead code. A side effect of this substitution
1790 ;;; is to delete the variable.
1798 ;; Think about (LET ((A ...)) (IF ... A ...)): two
1799 ;; LVAR-USEs should not be met on one path. Another problem
1800 ;; is with dynamic-extent.
1803 ;; If the destinatation is dynamic extent, don't substitute unless
1804 ;; the source is as well.
1809 ;; we should not change lifetime of unknown values lvars
1816 ;; CASTs can disappear, don't substitute if
1817 ;; DEST-LVAR has other uses (this will be
1818 ;; insufficient if we have a CAST-CAST chain, but
1819 ;; works well for a single CAST)
1827 ;; While CRETURN and EXIT nodes may be known-values,
1828 ;; they have their own complications, such as
1829 ;; substitution into CRETURN may create new tail calls.
1830 nil)
1833 t))
1839 ;; Really it is (EQ (LVAR-USES LVAR) REF):
1840 t)
1845 ;; KLUDGE: it should be (TYPE-CHECK 0)
1849 ;; FIXME
1850 t)
1858 t)))
1860 ;;; Delete a LET, removing the call and bind nodes, and warning about
1861 ;;; any unreferenced variables. Note that FLUSH-DEAD-CODE will come
1862 ;;; along right away and delete the REF and then the lambda, since we
1863 ;;; flush the FUN lvar.
1878 ;;; This function is called when one of the arguments to a LET
1879 ;;; changes. We look at each changed argument. If the corresponding
1880 ;;; variable is set, then we call PROPAGATE-FROM-SETS. Otherwise, we
1881 ;;; consider substituting for the variable, and also propagate
1882 ;;; derived-type information for the arg to all the VAR's refs.
1883 ;;;
1884 ;;; Substitution is inhibited when the arg leaf's derived type isn't a
1885 ;;; subtype of the argument's leaf type. This prevents type checking
1886 ;;; from being defeated, and also ensures that the best representation
1887 ;;; for the variable can be used.
1888 ;;;
1889 ;;; Substitution of individual references is inhibited if the
1890 ;;; reference is in a different component from the home. This can only
1891 ;;; happen with closures over top level lambda vars. In such cases,
1892 ;;; the references may have already been compiled, and thus can't be
1893 ;;; retroactively modified.
1894 ;;;
1895 ;;; If all of the variables are deleted (have no references) when we
1896 ;;; are done, then we delete the LET.
1897 ;;;
1898 ;;; Note that we are responsible for clearing the LVAR-REOPTIMIZE
1899 ;;; flags.
1914 ;; (NODE-DERIVED-TYPE USE) would
1915 ;; be better -- APD, 2003-05-15
1923 t)
1926 nil)))
1927 leaf var)))
1928 t)))))
1940 ;;; This function is called when one of the args to a non-LET local
1941 ;;; call changes. For each changed argument corresponding to an unset
1942 ;;; variable, we compute the union of the types across all calls and
1943 ;;; propagate this type information to the var's refs.
1944 ;;;
1945 ;;; If the function has an entry-fun, then we don't do anything: since
1946 ;;; it has a XEP we would not discover anything.
1947 ;;;
1948 ;;; If the function is an optional-entry-point, we will just make sure
1949 ;;; &REST lists are known to be lists. Doing the regular rigamarole
1950 ;;; can erronously propagate too strict types into refs: see
1951 ;;; BUG-655203-REGRESSION in tests/compiler.pure.lisp.
1952 ;;;
1953 ;;; We can clear the LVAR-REOPTIMIZE flags for arguments in all calls
1954 ;;; corresponding to changed arguments in CALL, since the only use in
1955 ;;; IR1 optimization of the REOPTIMIZE flag for local call args is
1956 ;;; right here.
1961 ;; We can still make sure &REST is known to be a list.
1966 ;; The normal case.
1974 vars))
1990 union)))))
1993 and type in union
1997 \f
1998 ;;;; multiple values optimization
2000 ;;; Do stuff to notice a change to a MV combination node. There are
2001 ;;; two main branches here:
2002 ;;; -- If the call is local, then it is already a MV let, or should
2003 ;;; become one. Note that although all :LOCAL MV calls must eventually
2004 ;;; be converted to :MV-LETs, there can be a window when the call
2005 ;;; is local, but has not been LET converted yet. This is because
2006 ;;; the entry-point lambdas may have stray references (in other
2007 ;;; entry points) that have not been deleted yet.
2008 ;;; -- The call is full. This case is somewhat similar to the non-MV
2009 ;;; combination optimization: we propagate return type information and
2010 ;;; notice non-returning calls. We also have an optimization
2011 ;;; which tries to convert MV-CALLs into MV-binds.
2017 (:local
2022 do
2027 (:full
2049 ;;; Propagate derived type info from the values lvar to the vars.
2060 and type in types
2065 do
2080 sum nvals)))
2084 "MULTIPLE-VALUE-CALL with ~R values when the function expects ~
2085 at least ~R."
2086 total-nvals min)
2091 "MULTIPLE-VALUE-CALL with ~R values when the function expects ~
2092 at most ~R."
2093 total-nvals max)
2099 min)
2117 ;;; If we see:
2118 ;;; (multiple-value-bind
2119 ;;; (x y)
2120 ;;; (values xx yy)
2121 ;;; ...)
2122 ;;; Convert to:
2123 ;;; (let ((x xx)
2124 ;;; (y yy))
2125 ;;; ...)
2126 ;;;
2127 ;;; What we actually do is convert the VALUES combination into a
2128 ;;; normal LET combination calling the original :MV-LET lambda. If
2129 ;;; there are extra args to VALUES, discard the corresponding
2130 ;;; lvars. If there are insufficient args, insert references to NIL.
2154 for prev = node-prev then ctran
2160 use ctran))
2174 ;; Propagate derived types from the VALUES call to its args:
2175 ;; transforms can leave the VALUES call with a better type
2176 ;; than its args have, so make sure not to throw that away.
2182 ;; Propagate declared types of MV-BIND variables.
2185 t)))
2187 ;;; If we see:
2188 ;;; (values-list (list x y z))
2189 ;;;
2190 ;;; Convert to:
2191 ;;; (values x y z)
2192 ;;;
2193 ;;; In implementation, this is somewhat similar to
2194 ;;; CONVERT-MV-BIND-TO-LET. We grab the args of LIST and make them
2195 ;;; args of the VALUES-LIST call, flushing the old argument lvar
2196 ;;; (allowing the LIST to be flushed.)
2197 ;;;
2198 ;;; FIXME: Thus we lose possible type assertions on (LIST ...).
2205 ;; FIXME: VALUES might not satisfy an assertion on NODE-LVAR.
2217 t)))
2219 ;;; If VALUES appears in a non-MV context, then effectively convert it
2220 ;;; to a PROG1. This allows the computation of the additional values
2221 ;;; to become dead code.
2233 val))
2234 nil))
2236 ;;; TODO:
2237 ;;; - CAST chains;
2256 ;; Vestigial exits are only introduced when eliminating a local
2257 ;; RETURN-FROM. We may delete them only when we can show that
2258 ;; there are no other code paths that use the entry LVAR that
2259 ;; are live from within the block that contained the deleted
2260 ;; EXIT and that all uses of the entry LVAR have the same
2261 ;; dynamic-extent environment. The conservative version of this
2262 ;; is that there are no EXITs for any ENTRY introduced between
2263 ;; the LEXENV of the deleted EXIT and the LEXENV of the target
2264 ;; ENTRY and that there is no :DYNAMIC-EXTENT cleanup between
2265 ;; the deleted EXIT and the target ENTRY (which the CAST shares
2266 ;; a lexenv with). See ONLY-HARMLESS-CLEANUPS in locall.lisp
2267 ;; for a similar analysis.
2281 #+sb-xc-host context
2303 lvar)
2308 :arg-count
2315 lvar)
2316 ;; Pathwise removing of CAST
2319 next-block)
2353 ;; No need to transform into an analog of
2354 ;; %COMPILE-TIME-TYPE-ERROR, %CHECK-BOUND will signal at
2355 ;; run-time and %CHECK-BOUND ir2-converter will signal at
2356 ;; compile-time if it survives further stages of ir1
2357 ;; optimization.
2360 ;; FIXME: Do it in one step.
2365 value
2371 ;; FIXME: Derived type.
2381 ',detail
2384 ;; KLUDGE: FILTER-LVAR does not work for non-returning
2385 ;; functions, so we declare the return type of
2386 ;; %COMPILE-TIME-TYPE-ERROR to be * and derive the real type
2387 ;; here.
2391 ;; FIXME: Is it necessary?