1 ;;;; This file contains the implementation-independent code for Pack
2 ;;;; phase in the compiler. Pack is responsible for assigning TNs to
3 ;;;; storage allocations or "register allocation".
5 ;;;; This software is part of the SBCL system. See the README file for
8 ;;;; This software is derived from the CMU CL system, which was
9 ;;;; written at Carnegie Mellon University and released into the
10 ;;;; public domain. The software is in the public domain and is
11 ;;;; provided with absolutely no warranty. See the COPYING and CREDITS
12 ;;;; files for more information.
14 (in-package "SB!REGALLOC")
16 ;;; for debugging: some parameters controlling which optimizations we
18 (defvar *pack-assign-costs
* t
)
19 (defvar *pack-optimize-saves
* t
)
20 ;;; FIXME: Perhaps SB-FLUID should be renamed to SB-TWEAK and these
21 ;;; should be made conditional on SB-TWEAK.
23 (declaim (ftype (function (component) index
) ir2-block-count
))
25 ;;;; conflict determination
27 ;;; Return true if the element at the specified offset, or in any of
28 ;;; the [size-1] subsequent offsets, in SB has a conflict with TN:
29 ;;; -- If a component-live TN (:COMPONENT kind), then iterate over
30 ;;; all the blocks. If the element at OFFSET is used anywhere in
31 ;;; any of the component's blocks (always-live /= 0), then there
33 ;;; -- If TN is global (Confs true), then iterate over the blocks TN
34 ;;; is live in (using TN-GLOBAL-CONFLICTS). If the TN is live
35 ;;; everywhere in the block (:LIVE), then there is a conflict
36 ;;; if the element at offset is used anywhere in the block
37 ;;; (Always-Live /= 0). Otherwise, we use the local TN number for
38 ;;; TN in block to find whether TN has a conflict at Offset in
40 ;;; -- If TN is local, then we just check for a conflict in the block
43 ;;; If there is a conflict, returns the first such conflicting offset.
44 (defun offset-conflicts-in-sb (tn sb offset
&key
(size 1))
45 (declare (type tn tn
) (type finite-sb sb
) (type index offset size
))
46 (let ((confs (tn-global-conflicts tn
))
48 (sb-conflicts (finite-sb-conflicts sb
))
49 (sb-always-live (finite-sb-always-live sb
)))
50 (macrolet ((do-offsets ((var) &body body
)
51 `(loop for
,var upfrom offset
53 thereis
(progn ,@body
))))
56 (do-offsets (offset-iter)
57 (let ((loc-live (svref sb-always-live offset-iter
)))
58 (dotimes (i (ir2-block-count *component-being-compiled
*))
59 (when (/= (sbit loc-live i
) 0)
60 (return offset-iter
))))))
62 ;; TN is global, iterate over the blocks TN is live in.
63 (do ((conf confs
(global-conflicts-next-tnwise conf
)))
66 (let* ((block (global-conflicts-block conf
))
67 (num (ir2-block-number block
)))
68 (if (eq (global-conflicts-kind conf
) :live
)
69 (do-offsets (offset-iter)
70 (let ((loc-live (svref sb-always-live offset-iter
)))
71 (when (/= (sbit loc-live num
) 0)
72 (return-from offset-conflicts-in-sb offset-iter
))))
73 (do-offsets (offset-iter)
74 (let ((loc-confs (svref sb-conflicts offset-iter
)))
75 (when (/= (sbit (svref loc-confs num
)
76 (global-conflicts-number conf
))
78 (return-from offset-conflicts-in-sb offset-iter
))))))))
80 (do-offsets (offset-iter)
81 (and (/= (sbit (svref (svref sb-conflicts offset-iter
)
82 (ir2-block-number (tn-local tn
)))
87 ;;; Return true if TN has a conflict in SC at the specified offset.
88 (declaim (ftype (function (tn sc index
) (values (or null index
) &optional
))
90 (defun conflicts-in-sc (tn sc offset
)
91 (declare (type tn tn
) (type sc sc
) (type index offset
))
92 (offset-conflicts-in-sb tn
(sc-sb sc
) offset
93 :size
(sc-element-size sc
)))
95 ;;; Add TN's conflicts into the conflicts for the location at OFFSET
96 ;;; in SC. We iterate over each location in TN, adding to the
97 ;;; conflicts for that location:
98 ;;; -- If TN is a :COMPONENT TN, then iterate over all the blocks,
99 ;;; setting all of the local conflict bits and the always-live bit.
100 ;;; This records a conflict with any TN that has a LTN number in
101 ;;; the block, as well as with :ALWAYS-LIVE and :ENVIRONMENT TNs.
102 ;;; -- If TN is global, then iterate over the blocks TN is live in. In
103 ;;; addition to setting the always-live bit to represent the conflict
104 ;;; with TNs live throughout the block, we also set bits in the
105 ;;; local conflicts. If TN is :ALWAYS-LIVE in the block, we set all
106 ;;; the bits, otherwise we OR in the local conflict bits.
107 ;;; -- If the TN is local, then we just do the block it is local to,
108 ;;; setting always-live and OR'ing in the local conflicts.
109 (defun add-location-conflicts (tn sc offset optimize
)
110 (declare (type tn tn
) (type sc sc
) (type index offset
))
111 (let ((confs (tn-global-conflicts tn
))
114 (dotimes (i (sc-element-size sc
))
115 (declare (type index i
))
116 (let* ((this-offset (+ offset i
))
117 (loc-confs (svref (finite-sb-conflicts sb
) this-offset
))
118 (loc-live (svref (finite-sb-always-live sb
) this-offset
)))
120 ((eq kind
:component
)
121 (dotimes (num (ir2-block-count *component-being-compiled
*))
122 (declare (type index num
))
123 (setf (sbit loc-live num
) 1)
124 (set-bit-vector (svref loc-confs num
))))
126 (do ((conf confs
(global-conflicts-next-tnwise conf
)))
128 (let* ((block (global-conflicts-block conf
))
129 (num (ir2-block-number block
))
130 (local-confs (svref loc-confs num
)))
131 (declare (type local-tn-bit-vector local-confs
))
132 (setf (sbit loc-live num
) 1)
133 (if (eq (global-conflicts-kind conf
) :live
)
134 (set-bit-vector local-confs
)
135 (bit-ior local-confs
(global-conflicts-conflicts conf
) t
)))))
137 (let ((num (ir2-block-number (tn-local tn
))))
138 (setf (sbit loc-live num
) 1)
139 (bit-ior (the local-tn-bit-vector
(svref loc-confs num
))
140 (tn-local-conflicts tn
) t
))))
141 ;; Calculating ALWAYS-LIVE-COUNT is moderately expensive, and
142 ;; currently the information isn't used unless (> SPEED
145 (setf (svref (finite-sb-always-live-count sb
) this-offset
)
146 (find-location-usage sb this-offset
))))))
149 ;; A rought measure of how much a given OFFSET in SB is currently
150 ;; used. Current implementation counts the amount of blocks where the
151 ;; offset has been marked as ALWAYS-LIVE.
152 (defun find-location-usage (sb offset
)
153 (declare (optimize speed
))
154 (declare (type sb sb
) (type index offset
))
155 (let* ((always-live (svref (finite-sb-always-live sb
) offset
)))
156 (declare (simple-bit-vector always-live
))
157 (count 1 always-live
)))
159 ;;; Return the total number of IR2-BLOCKs in COMPONENT.
160 (defun ir2-block-count (component)
161 (declare (type component component
))
162 (do ((2block (block-info (block-next (component-head component
)))
163 (ir2-block-next 2block
)))
165 (error "What? No ir2 blocks have a non-nil number?"))
166 (when (ir2-block-number 2block
)
167 (return (1+ (ir2-block-number 2block
))))))
169 ;;; Ensure that the conflicts vectors for each :FINITE SB are large
170 ;;; enough for the number of blocks allocated. Also clear any old
171 ;;; conflicts and reset the current size to the initial size.
172 (defun init-sb-vectors (component)
173 (let ((nblocks (ir2-block-count component
)))
174 (dolist (sb *backend-sb-list
*)
175 (unless (eq (sb-kind sb
) :non-packed
)
176 (let* ((conflicts (finite-sb-conflicts sb
))
177 (always-live (finite-sb-always-live sb
))
178 (always-live-count (finite-sb-always-live-count sb
))
179 (max-locs (length conflicts
))
180 (last-count (finite-sb-last-block-count sb
)))
181 (unless (zerop max-locs
)
182 (let ((current-size (length (the simple-vector
183 (svref conflicts
0)))))
185 ((> nblocks current-size
)
186 (let ((new-size (max nblocks
(* current-size
2))))
187 (declare (type index new-size
))
188 (dotimes (i max-locs
)
189 (declare (type index i
))
190 (let ((new-vec (make-array new-size
)))
191 (let ((old (svref conflicts i
)))
192 (declare (simple-vector old
))
193 (dotimes (j current-size
)
194 (declare (type index j
))
195 (setf (svref new-vec j
)
196 (clear-bit-vector (svref old j
)))))
198 (do ((j current-size
(1+ j
)))
200 (declare (type index j
))
201 (setf (svref new-vec j
)
202 (make-array local-tn-limit
:element-type
'bit
203 :initial-element
0)))
204 (setf (svref conflicts i
) new-vec
))
205 (setf (svref always-live i
)
206 (make-array new-size
:element-type
'bit
208 (setf (svref always-live-count i
) 0))))
210 (dotimes (i (finite-sb-current-size sb
))
211 (declare (type index i
))
212 (let ((conf (svref conflicts i
)))
213 (declare (simple-vector conf
))
214 (dotimes (j last-count
)
215 (declare (type index j
))
216 (clear-bit-vector (svref conf j
))))
217 (clear-bit-vector (svref always-live i
))
218 (setf (svref always-live-count i
) 0))))))
220 (setf (finite-sb-last-block-count sb
) nblocks
)
221 (setf (finite-sb-current-size sb
) (sb-size sb
))
222 (setf (finite-sb-last-offset sb
) 0))))))
224 ;;; Expand the :UNBOUNDED SB backing SC by either the initial size or
225 ;;; the SC element size, whichever is larger. If NEEDED-SIZE is
226 ;;; larger, then use that size.
227 (defun grow-sc (sc &optional
(needed-size 0))
228 (declare (type sc sc
) (type index needed-size
))
229 (let* ((sb (sc-sb sc
))
230 (size (finite-sb-current-size sb
))
231 (align-mask (1- (sc-alignment sc
)))
232 (inc (max (finite-sb-size-increment sb
)
233 (+ (sc-element-size sc
)
234 (- (logandc2 (+ size align-mask
) align-mask
)
236 (- needed-size size
)))
237 (new-size (let ((align-mask (1- (finite-sb-size-alignment sb
))))
238 (logandc2 (+ size inc align-mask
) align-mask
)))
239 (conflicts (finite-sb-conflicts sb
))
240 (block-size (if (zerop (length conflicts
))
241 (ir2-block-count *component-being-compiled
*)
242 (length (the simple-vector
(svref conflicts
0)))))
243 (padded-size (ash 1 (integer-length (1- new-size
)))))
244 (declare (type index inc new-size padded-size
))
245 (aver (eq (sb-kind sb
) :unbounded
))
247 (when (> padded-size
(length conflicts
))
248 (let ((new-conf (make-array padded-size
)))
249 (replace new-conf conflicts
)
250 (do ((i size
(1+ i
)))
252 (declare (type index i
))
253 (let ((loc-confs (make-array block-size
)))
254 (dotimes (j block-size
)
255 (setf (svref loc-confs j
)
256 (make-array local-tn-limit
258 :element-type
'bit
)))
259 (setf (svref new-conf i
) loc-confs
)))
260 (setf (finite-sb-conflicts sb
) new-conf
))
262 (let ((new-live (make-array padded-size
)))
263 (replace new-live
(finite-sb-always-live sb
))
264 (do ((i size
(1+ i
)))
266 (setf (svref new-live i
)
267 (make-array block-size
269 :element-type
'bit
)))
270 (setf (finite-sb-always-live sb
) new-live
))
272 (let ((new-live-count (make-array padded-size
)))
273 (declare (optimize speed
)) ;; FILL deftransform
274 (replace new-live-count
(finite-sb-always-live-count sb
))
275 (fill new-live-count
0 :start size
)
276 (setf (finite-sb-always-live-count sb
) new-live-count
))
278 (let ((new-tns (make-array padded-size
:initial-element nil
)))
279 (replace new-tns
(finite-sb-live-tns sb
))
280 (fill (finite-sb-live-tns sb
) nil
)
281 (setf (finite-sb-live-tns sb
) new-tns
)))
283 (setf (finite-sb-current-size sb
) new-size
))
289 ;;; Give someone a hard time because there isn't any load function
290 ;;; defined to move from SRC to DEST.
291 (defun no-load-fun-error (src dest
)
292 (let* ((src-sc (tn-sc src
))
293 (src-name (sc-name src-sc
))
294 (dest-sc (tn-sc dest
))
295 (dest-name (sc-name dest-sc
)))
296 (cond ((eq (sb-kind (sc-sb src-sc
)) :non-packed
)
297 (unless (member src-sc
(sc-constant-scs dest-sc
))
298 (error "loading from an invalid constant SC?~@
299 VM definition inconsistent, try recompiling."))
300 (error "no load function defined to load SC ~S ~
301 from its constant SC ~S"
303 ((member src-sc
(sc-alternate-scs dest-sc
))
304 (error "no load function defined to load SC ~S from its ~
307 ((member dest-sc
(sc-alternate-scs src-sc
))
308 (error "no load function defined to save SC ~S in its ~
312 ;; FIXME: "VM definition is inconsistent" shouldn't be a
313 ;; possibility in SBCL.
314 (error "loading to/from SCs that aren't alternates?~@
315 VM definition is inconsistent, try recompiling.")))))
317 ;;; Called when we failed to pack TN. If RESTRICTED is true, then we
318 ;;; are restricted to pack TN in its SC.
319 (defun failed-to-pack-error (tn restricted
)
320 (declare (type tn tn
))
321 (let* ((sc (tn-sc tn
))
322 (scs (cons sc
(sc-alternate-scs sc
))))
325 (error "failed to pack restricted TN ~S in its SC ~S"
328 (aver (not (find :unbounded scs
329 :key
(lambda (x) (sb-kind (sc-sb x
))))))
330 (let ((ptype (tn-primitive-type tn
)))
333 (aver (member (sc-number sc
) (primitive-type-scs ptype
)))
334 (error "SC ~S doesn't have any :UNBOUNDED alternate SCs, but is~@
335 a SC for primitive-type ~S."
336 (sc-name sc
) (primitive-type-name ptype
)))
338 (error "SC ~S doesn't have any :UNBOUNDED alternate SCs."
341 ;;; Return a list of format arguments describing how TN is used in
343 (defun describe-tn-use (loc tn op
)
344 (let* ((vop (tn-ref-vop op
))
345 (args (vop-args vop
))
346 (results (vop-results vop
))
347 (name (with-simple-output-to-string (stream)
348 (print-tn-guts tn stream
)))
349 (2comp (component-info *component-being-compiled
*))
352 ((setq temp
(position-in #'tn-ref-across tn args
:key
#'tn-ref-tn
))
353 `("~2D: ~A (~:R argument)" ,loc
,name
,(1+ temp
)))
354 ((setq temp
(position-in #'tn-ref-across tn results
:key
#'tn-ref-tn
))
355 `("~2D: ~A (~:R result)" ,loc
,name
,(1+ temp
)))
356 ((setq temp
(position-in #'tn-ref-across tn args
:key
#'tn-ref-load-tn
))
357 `("~2D: ~A (~:R argument load TN)" ,loc
,name
,(1+ temp
)))
358 ((setq temp
(position-in #'tn-ref-across tn results
:key
360 `("~2D: ~A (~:R result load TN)" ,loc
,name
,(1+ temp
)))
361 ((setq temp
(position-in #'tn-ref-across tn
(vop-temps vop
)
363 `("~2D: ~A (temporary ~A)" ,loc
,name
364 ,(operand-parse-name (elt (vop-parse-temps
366 (vop-info-name (vop-info vop
))))
368 ((eq (tn-kind tn
) :component
)
369 `("~2D: ~A (component live)" ,loc
,name
))
370 ((position-in #'tn-next tn
(ir2-component-wired-tns 2comp
))
371 `("~2D: ~A (wired)" ,loc
,name
))
372 ((position-in #'tn-next tn
(ir2-component-restricted-tns 2comp
))
373 `("~2D: ~A (restricted)" ,loc
,name
))
375 `("~2D: not referenced?" ,loc
)))))
377 ;;; If load TN packing fails, try to give a helpful error message. We
378 ;;; find a TN in each location that conflicts, and print it.
379 (defun failed-to-pack-load-tn-error (scs op
)
380 (declare (list scs
) (type tn-ref op
))
384 (let* ((sb (sc-sb sc
))
385 (confs (finite-sb-live-tns sb
)))
386 (aver (eq (sb-kind sb
) :finite
))
387 (dolist (el (sc-locations sc
))
388 (declare (type index el
))
389 (let ((conf (load-tn-conflicts-in-sc op sc el t
)))
391 (used (describe-tn-use el conf op
))
393 (end (+ el
(sc-element-size sc
))))
396 (declare (type index i end
))
397 (let ((victim (svref confs i
)))
399 (used (describe-tn-use el victim op
))
402 (multiple-value-bind (arg-p n more-p costs load-scs incon
)
403 (get-operand-info op
)
404 (declare (ignore costs load-scs
))
406 (error "unable to pack a Load-TN in SC ~{~A~#[~^~;, or ~:;,~]~} ~
407 for the ~:R ~:[result~;argument~] to~@
409 ~:[since all SC elements are in use:~:{~%~@?~}~%~;~
410 ~:*but these SC elements are not in use:~% ~S~%Bug?~*~]~
412 Current cost info inconsistent with that in effect at compile ~
413 time. Recompile.~%Compilation order may be incorrect.~]"
414 (mapcar #'sc-name scs
)
416 (vop-info-name (vop-info (tn-ref-vop op
)))
420 ;;; This is called when none of the SCs that we can load OP into are
421 ;;; allowed by OP's primitive-type.
422 (defun no-load-scs-allowed-by-primitive-type-error (ref)
423 (declare (type tn-ref ref
))
424 (let* ((tn (tn-ref-tn ref
))
425 (ptype (tn-primitive-type tn
)))
426 (multiple-value-bind (arg-p pos more-p costs load-scs incon
)
427 (get-operand-info ref
)
428 (declare (ignore costs
))
430 (error "~S is not valid as the ~:R ~:[result~;argument~] to VOP:~
432 since the TN's primitive type ~S doesn't allow any of the SCs~@
433 allowed by the operand restriction:~% ~S~
435 Current cost info inconsistent with that in effect at compile ~
436 time. Recompile.~%Compilation order may be incorrect.~]"
438 (template-name (vop-info (tn-ref-vop ref
)))
439 (primitive-type-name ptype
)
440 (mapcar #'sc-name
(listify-restrictions load-scs
))
445 ;;; Do stuff to note that TN is spilled at VOP for the debugger's benefit.
446 (defun note-spilled-tn (tn vop
)
447 (when (and (tn-leaf tn
) (vop-save-set vop
))
448 (let ((2comp (component-info *component-being-compiled
*)))
449 (setf (gethash tn
(ir2-component-spilled-tns 2comp
)) t
)
450 (pushnew tn
(gethash vop
(ir2-component-spilled-vops 2comp
)))))
453 ;;; Make a save TN for TN, pack it, and return it. We copy various
454 ;;; conflict information from the TN so that pack does the right
456 (defun pack-save-tn (tn)
457 (declare (type tn tn
))
458 (let ((res (make-tn 0 :save nil nil
)))
459 (dolist (alt (sc-alternate-scs (tn-sc tn
))
460 (error "no unbounded alternate for SC ~S"
461 (sc-name (tn-sc tn
))))
462 (when (eq (sb-kind (sc-sb alt
)) :unbounded
)
463 (setf (tn-save-tn tn
) res
)
464 (setf (tn-save-tn res
) tn
)
465 (setf (tn-sc res
) alt
)
469 ;;; Find the load function for moving from SRC to DEST and emit a
470 ;;; MOVE-OPERAND VOP with that function as its info arg.
471 (defun emit-operand-load (node block src dest before
)
472 (declare (type node node
) (type ir2-block block
)
473 (type tn src dest
) (type (or vop null
) before
))
474 (emit-load-template node block
475 (template-or-lose 'move-operand
)
477 (list (or (svref (sc-move-funs (tn-sc dest
))
478 (sc-number (tn-sc src
)))
479 (no-load-fun-error src dest
)))
483 ;;; Find the preceding use of the VOP NAME in the emit order, starting
484 ;;; with VOP. We must find the VOP in the same IR1 block.
485 (defun reverse-find-vop (name vop
)
486 (do* ((block (vop-block vop
) (ir2-block-prev block
))
487 (last vop
(ir2-block-last-vop block
)))
489 (aver (eq (ir2-block-block block
) (ir2-block-block (vop-block vop
))))
490 (do ((current last
(vop-prev current
)))
492 (when (eq (vop-info-name (vop-info current
)) name
)
493 (return-from reverse-find-vop current
)))))
495 ;;; For TNs that have other than one writer, we save the TN before
496 ;;; each call. If a local call (MOVE-ARGS is :LOCAL-CALL), then we
497 ;;; scan back for the ALLOCATE-FRAME VOP, and emit the save there.
498 ;;; This is necessary because in a self-recursive local call, the
499 ;;; registers holding the current arguments may get trashed by setting
500 ;;; up the call arguments. The ALLOCATE-FRAME VOP marks a place at
501 ;;; which the values are known to be good.
502 (defun save-complex-writer-tn (tn vop
)
503 (let ((save (or (tn-save-tn tn
)
505 (node (vop-node vop
))
506 (block (vop-block vop
))
507 (next (vop-next vop
)))
508 (when (eq (tn-kind save
) :specified-save
)
509 (setf (tn-kind save
) :save
))
510 (aver (eq (tn-kind save
) :save
))
511 (emit-operand-load node block tn save
512 (if (eq (vop-info-move-args (vop-info vop
))
514 (reverse-find-vop 'allocate-frame vop
)
516 (emit-operand-load node block save tn next
)))
518 ;;; Return a VOP after which is an OK place to save the value of TN.
519 ;;; For correctness, it is only required that this location be after
520 ;;; any possible write and before any possible restore location.
522 ;;; In practice, we return the unique writer VOP, but give up if the
523 ;;; TN is ever read by a VOP with MOVE-ARGS :LOCAL-CALL. This prevents
524 ;;; us from being confused by non-tail local calls.
526 ;;; When looking for writes, we have to ignore uses of MOVE-OPERAND,
527 ;;; since they will correspond to restores that we have already done.
528 (defun find-single-writer (tn)
529 (declare (type tn tn
))
530 (do ((write (tn-writes tn
) (tn-ref-next write
))
534 (do ((read (tn-reads tn
) (tn-ref-next read
)))
536 (when (eq (vop-info-move-args
543 (unless (eq (vop-info-name (vop-info (tn-ref-vop write
)))
545 (when res
(return nil
))
548 ;;; Try to save TN at a single location. If we succeed, return T,
550 (defun save-single-writer-tn (tn)
551 (declare (type tn tn
))
552 (let* ((old-save (tn-save-tn tn
))
553 (save (or old-save
(pack-save-tn tn
)))
554 (writer (find-single-writer tn
)))
557 (eq (tn-kind old-save
) :specified-save
)))
558 (emit-operand-load (vop-node writer
) (vop-block writer
)
559 tn save
(vop-next writer
))
560 (setf (tn-kind save
) :save-once
)
563 ;;; Restore a TN with a :SAVE-ONCE save TN.
564 (defun restore-single-writer-tn (tn vop
)
565 (declare (type tn
) (type vop vop
))
566 (let ((save (tn-save-tn tn
)))
567 (aver (eq (tn-kind save
) :save-once
))
568 (emit-operand-load (vop-node vop
) (vop-block vop
) save tn
(vop-next vop
)))
571 ;;; Save a single TN that needs to be saved, choosing save-once if
572 ;;; appropriate. This is also called by SPILL-AND-PACK-LOAD-TN.
573 (defun basic-save-tn (tn vop
)
574 (declare (type tn tn
) (type vop vop
))
575 (let ((save (tn-save-tn tn
)))
576 (cond ((and save
(eq (tn-kind save
) :save-once
))
577 (restore-single-writer-tn tn vop
))
578 ((save-single-writer-tn tn
)
579 (restore-single-writer-tn tn vop
))
581 (save-complex-writer-tn tn vop
))))
584 ;;; Scan over the VOPs in BLOCK, emiting saving code for TNs noted in
585 ;;; the codegen info that are packed into saved SCs.
586 (defun emit-saves (block)
587 (declare (type ir2-block block
))
588 (do ((vop (ir2-block-start-vop block
) (vop-next vop
)))
590 (when (eq (vop-info-save-p (vop-info vop
)) t
)
591 (do-live-tns (tn (vop-save-set vop
) block
)
592 (when (and (sc-save-p (tn-sc tn
))
593 (not (eq (tn-kind tn
) :component
)))
594 (basic-save-tn tn vop
)))))
598 ;;;; optimized saving
600 ;;; Save TN if it isn't a single-writer TN that has already been
601 ;;; saved. If multi-write, we insert the save BEFORE the specified
602 ;;; VOP. CONTEXT is a VOP used to tell which node/block to use for the
604 (defun save-if-necessary (tn before context
)
605 (declare (type tn tn
) (type (or vop null
) before
) (type vop context
))
606 (let ((save (tn-save-tn tn
)))
607 (when (eq (tn-kind save
) :specified-save
)
608 (setf (tn-kind save
) :save
))
609 (aver (member (tn-kind save
) '(:save
:save-once
)))
610 (unless (eq (tn-kind save
) :save-once
)
611 (or (save-single-writer-tn tn
)
612 (emit-operand-load (vop-node context
) (vop-block context
)
616 ;;; Load the TN from its save location, allocating one if necessary.
617 ;;; The load is inserted BEFORE the specified VOP. CONTEXT is a VOP
618 ;;; used to tell which node/block to use for the new VOP.
619 (defun restore-tn (tn before context
)
620 (declare (type tn tn
) (type (or vop null
) before
) (type vop context
))
621 (let ((save (or (tn-save-tn tn
) (pack-save-tn tn
))))
622 (emit-operand-load (vop-node context
) (vop-block context
)
626 ;;; Start scanning backward at the end of BLOCK, looking which TNs are
627 ;;; live and looking for places where we have to save. We manipulate
628 ;;; two sets: SAVES and RESTORES.
630 ;;; SAVES is a set of all the TNs that have to be saved because they
631 ;;; are restored after some call. We normally delay saving until the
632 ;;; beginning of the block, but we must save immediately if we see a
633 ;;; write of the saved TN. We also immediately save all TNs and exit
634 ;;; when we see a NOTE-ENVIRONMENT-START VOP, since saves can't be
635 ;;; done before the environment is properly initialized.
637 ;;; RESTORES is a set of all the TNs read (and not written) between
638 ;;; here and the next call, i.e. the set of TNs that must be restored
639 ;;; when we reach the next (earlier) call VOP. Unlike SAVES, this set
640 ;;; is cleared when we do the restoring after a call. Any TNs that
641 ;;; were in RESTORES are moved into SAVES to ensure that they are
642 ;;; saved at some point.
644 ;;; SAVES and RESTORES are represented using both a list and a
645 ;;; bit-vector so that we can quickly iterate and test for membership.
646 ;;; The incoming SAVES and RESTORES args are used for computing these
647 ;;; sets (the initial contents are ignored.)
649 ;;; When we hit a VOP with :COMPUTE-ONLY SAVE-P (an internal error
650 ;;; location), we pretend that all live TNs were read, unless (= speed
651 ;;; 3), in which case we mark all the TNs that are live but not
652 ;;; restored as spilled.
653 (defun optimized-emit-saves-block (block saves restores
)
654 (declare (type ir2-block block
) (type simple-bit-vector saves restores
))
655 (let ((1block (ir2-block-block block
))
659 (declare (list saves-list restores-list
))
660 (clear-bit-vector saves
)
661 (clear-bit-vector restores
)
662 (do-live-tns (tn (ir2-block-live-in block
) block
)
663 (when (and (sc-save-p (tn-sc tn
))
664 (not (eq (tn-kind tn
) :component
)))
665 (let ((num (tn-number tn
)))
666 (setf (sbit restores num
) 1)
667 (push tn restores-list
))))
669 (do ((block block
(ir2-block-prev block
))
671 ((not (eq (ir2-block-block block
) 1block
))
672 (aver (not skipping
))
673 (dolist (save saves-list
)
674 (let ((start (ir2-block-start-vop prev
)))
675 (save-if-necessary save start start
)))
677 (do ((vop (ir2-block-last-vop block
) (vop-prev vop
)))
679 (let ((info (vop-info vop
)))
680 (case (vop-info-name info
)
684 (note-environment-start
685 (aver (not skipping
))
686 (dolist (save saves-list
)
687 (save-if-necessary save
(vop-next vop
) vop
))
688 (return-from optimized-emit-saves-block block
)))
691 (do ((write (vop-results vop
) (tn-ref-across write
)))
693 (let* ((tn (tn-ref-tn write
))
694 (num (tn-number tn
)))
695 (unless (zerop (sbit restores num
))
696 (setf (sbit restores num
) 0)
698 (delete tn restores-list
:test
#'eq
)))
699 (unless (zerop (sbit saves num
))
700 (setf (sbit saves num
) 0)
701 (save-if-necessary tn
(vop-next vop
) vop
)
703 (delete tn saves-list
:test
#'eq
))))))
705 (macrolet ((save-note-read (tn)
707 (num (tn-number tn
)))
708 (when (and (sc-save-p (tn-sc tn
))
709 (zerop (sbit restores num
))
710 (not (eq (tn-kind tn
) :component
)))
711 (setf (sbit restores num
) 1)
712 (push tn restores-list
)))))
714 (case (vop-info-save-p info
)
716 (dolist (tn restores-list
)
717 (restore-tn tn
(vop-next vop
) vop
)
718 (let ((num (tn-number tn
)))
719 (when (zerop (sbit saves num
))
721 (setf (sbit saves num
) 1))))
722 (setq restores-list nil
)
723 (clear-bit-vector restores
))
725 (cond ((policy (vop-node vop
) (= speed
3))
726 (do-live-tns (tn (vop-save-set vop
) block
)
727 (when (zerop (sbit restores
(tn-number tn
)))
728 (note-spilled-tn tn vop
))))
730 (do-live-tns (tn (vop-save-set vop
) block
)
731 (save-note-read tn
))))))
733 (if (eq (vop-info-move-args info
) :local-call
)
735 (do ((read (vop-args vop
) (tn-ref-across read
)))
737 (save-note-read (tn-ref-tn read
))))))))))
739 ;;; This is like EMIT-SAVES, only different. We avoid redundant saving
740 ;;; within the block, and don't restore values that aren't used before
741 ;;; the next call. This function is just the top level loop over the
742 ;;; blocks in the component, which locates blocks that need saving
744 (defun optimized-emit-saves (component)
745 (declare (type component component
))
746 (let* ((gtn-count (1+ (ir2-component-global-tn-counter
747 (component-info component
))))
748 (saves (make-array gtn-count
:element-type
'bit
))
749 (restores (make-array gtn-count
:element-type
'bit
))
750 (block (ir2-block-prev (block-info (component-tail component
))))
751 (head (block-info (component-head component
))))
753 (when (eq block head
) (return))
754 (when (do ((vop (ir2-block-start-vop block
) (vop-next vop
)))
756 (when (eq (vop-info-save-p (vop-info vop
)) t
)
758 (setq block
(optimized-emit-saves-block block saves restores
)))
759 (setq block
(ir2-block-prev block
)))))
761 ;;; Iterate over the normal TNs, finding the cost of packing on the
762 ;;; stack in units of the number of references. We count all read
763 ;;; references as +1, write references as + *tn-write-cost*, and
764 ;;; subtract out REGISTER-SAVE-PENALTY for each place where we would
765 ;;; have to save a register.
766 ;;; The subtraction reflects the fact that having a value in a
767 ;;; register around a call means that code to spill and unspill must
770 ;;; The costs also take into account the loop depth at which each
771 ;;; reference occurs: the penalty or cost is incremented by the depth
772 ;;; scaled by *tn-loop-depth-multiplier*. The default (NIL) is to let
773 ;;; this be one more than the max of the cost for reads (1), for write
774 ;;; references and for being live across a call.
775 (defvar *tn-write-cost
* 2)
776 (defvar *tn-loop-depth-multiplier
* nil
)
778 (defun assign-tn-costs (component)
779 (let* ((save-penalty *backend-register-save-penalty
*)
780 (write-cost *tn-write-cost
*)
781 (depth-scale (or *tn-loop-depth-multiplier
*
782 (1+ (max 1 write-cost save-penalty
)))))
783 (flet ((vop-depth-cost (vop)
784 (let ((loop (block-loop
788 (* depth-scale
(loop-depth loop
))
790 (do-ir2-blocks (block component
)
791 (do ((vop (ir2-block-start-vop block
) (vop-next vop
)))
793 (when (eq (vop-info-save-p (vop-info vop
)) t
)
794 (let ((penalty (+ save-penalty
(vop-depth-cost vop
))))
795 (do-live-tns (tn (vop-save-set vop
) block
)
796 (decf (tn-cost tn
) penalty
))))))
798 (do ((tn (ir2-component-normal-tns (component-info component
))
801 (let ((cost (tn-cost tn
)))
802 (declare (fixnum cost
))
803 (do ((ref (tn-reads tn
) (tn-ref-next ref
)))
805 (incf cost
(1+ (vop-depth-cost (tn-ref-vop ref
)))))
806 (do ((ref (tn-writes tn
) (tn-ref-next ref
)))
808 (incf cost
(+ write-cost
(vop-depth-cost (tn-ref-vop ref
)))))
809 (setf (tn-cost tn
) cost
))))))
811 ;;; Iterate over the normal TNs, folding over the depth of the looops
812 ;;; that the TN is used in and storing the result in TN-LOOP-DEPTH.
813 ;;: reducer is the function used to join depth values together. #'max
814 ;;; gives the maximum depth, #'+ the sum.
815 (defun assign-tn-depths (component &key
(reducer #'max
))
816 (declare (type function reducer
))
818 ;; We only use tn depth for normal TNs
819 (do ((tn (ir2-component-normal-tns (component-info component
))
823 (declare (type fixnum depth
))
825 (declare (type (or null tn-ref
) ref
))
826 (do ((ref ref
(tn-ref-next ref
)))
828 (let* ((vop (tn-ref-vop ref
))
829 (block (ir2-block-block (vop-block vop
)))
830 (loop (block-loop block
)))
831 (setf depth
(funcall reducer
837 (frob (tn-writes tn
))
838 (setf (tn-loop-depth tn
) depth
))))))
842 ;;; These variables indicate the last location at which we computed
843 ;;; the Live-TNs. They hold the BLOCK and VOP values that were passed
844 ;;; to COMPUTE-LIVE-TNS.
845 (defvar *live-block
*)
848 ;;; If we unpack some TNs, then we mark all affected blocks by
849 ;;; sticking them in this hash-table. This is initially null. We
850 ;;; create the hashtable if we do any unpacking.
851 (defvar *repack-blocks
*)
852 (declaim (type list
*repack-blocks
*))
854 ;;; Set the LIVE-TNS vectors in all :FINITE SBs to represent the TNs
855 ;;; live at the end of BLOCK.
856 (defun init-live-tns (block)
857 (dolist (sb *backend-sb-list
*)
858 (when (eq (sb-kind sb
) :finite
)
859 (fill (finite-sb-live-tns sb
) nil
)))
861 (do-live-tns (tn (ir2-block-live-in block
) block
)
862 (let* ((sc (tn-sc tn
))
864 (when (eq (sb-kind sb
) :finite
)
865 ;; KLUDGE: we can have "live" TNs that are neither read
866 ;; to nor written from, due to more aggressive (type-
867 ;; directed) constant propagation. Such TNs will never
868 ;; be assigned an offset nor be in conflict with anything.
870 ;; Ideally, it seems to me we could make sure these TNs
871 ;; are never allocated in the first place in
872 ;; ASSIGN-LAMBDA-VAR-TNS.
874 (do ((offset (tn-offset tn
) (1+ offset
))
875 (end (+ (tn-offset tn
) (sc-element-size sc
))))
877 (declare (type index offset end
))
878 (setf (svref (finite-sb-live-tns sb
) offset
) tn
))
879 (aver (and (null (tn-reads tn
)) (null (tn-writes tn
))))))))
881 (setq *live-block
* block
)
882 (setq *live-vop
* (ir2-block-last-vop block
))
886 ;;; Set the LIVE-TNs in :FINITE SBs to represent the TNs live
887 ;;; immediately after the evaluation of VOP in BLOCK, excluding
888 ;;; results of the VOP. If VOP is null, then compute the live TNs at
889 ;;; the beginning of the block. Sequential calls on the same block
890 ;;; must be in reverse VOP order.
891 (defun compute-live-tns (block vop
)
892 (declare (type ir2-block block
) (type vop vop
))
893 (unless (eq block
*live-block
*)
894 (init-live-tns block
))
896 (do ((current *live-vop
* (vop-prev current
)))
898 (do ((res (vop-results vop
) (tn-ref-across res
)))
900 (let* ((tn (tn-ref-tn res
))
903 (when (eq (sb-kind sb
) :finite
)
904 (do ((offset (tn-offset tn
) (1+ offset
))
905 (end (+ (tn-offset tn
) (sc-element-size sc
))))
907 (declare (type index offset end
))
908 (setf (svref (finite-sb-live-tns sb
) offset
) nil
))))))
909 (do ((ref (vop-refs current
) (tn-ref-next-ref ref
)))
911 (let ((ltn (tn-ref-load-tn ref
)))
913 (let* ((sc (tn-sc ltn
))
915 (when (eq (sb-kind sb
) :finite
)
916 (let ((tns (finite-sb-live-tns sb
)))
917 (do ((offset (tn-offset ltn
) (1+ offset
))
918 (end (+ (tn-offset ltn
) (sc-element-size sc
))))
920 (declare (type index offset end
))
921 (aver (null (svref tns offset
)))))))))
923 (let* ((tn (tn-ref-tn ref
))
926 (when (eq (sb-kind sb
) :finite
)
927 (let ((tns (finite-sb-live-tns sb
)))
928 (do ((offset (tn-offset tn
) (1+ offset
))
929 (end (+ (tn-offset tn
) (sc-element-size sc
))))
931 (declare (type index offset end
))
932 (if (tn-ref-write-p ref
)
933 (setf (svref tns offset
) nil
)
934 (let ((old (svref tns offset
)))
935 (aver (or (null old
) (eq old tn
)))
936 (setf (svref tns offset
) tn
)))))))))
938 (setq *live-vop
* vop
)
941 ;;; This is kind of like OFFSET-CONFLICTS-IN-SB, except that it uses
942 ;;; the VOP refs to determine whether a Load-TN for OP could be packed
943 ;;; in the specified location, disregarding conflicts with TNs not
944 ;;; referenced by this VOP. There is a conflict if either:
945 ;;; 1. The reference is a result, and the same location is either:
946 ;;; -- Used by some other result.
947 ;;; -- Used in any way after the reference (exclusive).
948 ;;; 2. The reference is an argument, and the same location is either:
949 ;;; -- Used by some other argument.
950 ;;; -- Used in any way before the reference (exclusive).
952 ;;; In 1 (and 2) above, the first bullet corresponds to result-result
953 ;;; (and argument-argument) conflicts. We need this case because there
954 ;;; aren't any TN-REFs to represent the implicit reading of results or
955 ;;; writing of arguments.
957 ;;; The second bullet corresponds to conflicts with temporaries or
958 ;;; between arguments and results.
960 ;;; We consider both the TN-REF-TN and the TN-REF-LOAD-TN (if any) to
961 ;;; be referenced simultaneously and in the same way. This causes
962 ;;; load-TNs to appear live to the beginning (or end) of the VOP, as
965 ;;; We return a conflicting TN if there is a conflict.
966 (defun load-tn-offset-conflicts-in-sb (op sb offset
)
967 (declare (type tn-ref op
) (type finite-sb sb
) (type index offset
))
968 (aver (eq (sb-kind sb
) :finite
))
969 (let ((vop (tn-ref-vop op
)))
970 (labels ((tn-overlaps (tn)
971 (let ((sc (tn-sc tn
))
972 (tn-offset (tn-offset tn
)))
973 (when (and (eq (sc-sb sc
) sb
)
974 (<= tn-offset offset
)
977 (+ tn-offset
(sc-element-size sc
)))))
980 (let ((tn (tn-ref-tn ref
))
981 (ltn (tn-ref-load-tn ref
)))
983 (and ltn
(tn-overlaps ltn
)))))
985 (do ((ops ops
(tn-ref-across ops
)))
987 (let ((found (same ops
)))
988 (when (and found
(not (eq ops op
)))
991 (do ((refs refs
(tn-ref-next-ref refs
)))
993 (let ((found (same refs
)))
994 (when found
(return found
))))))
995 (declare (inline is-op is-ref tn-overlaps
))
996 (if (tn-ref-write-p op
)
997 (or (is-op (vop-results vop
))
998 (is-ref (vop-refs vop
) op
))
999 (or (is-op (vop-args vop
))
1000 (is-ref (tn-ref-next-ref op
) nil
))))))
1002 ;;; Iterate over all the elements in the SB that would be allocated by
1003 ;;; allocating a TN in SC at Offset, checking for conflict with
1004 ;;; load-TNs or other TNs (live in the LIVE-TNS, which must be set
1005 ;;; up.) We also return true if there aren't enough locations after
1006 ;;; Offset to hold a TN in SC. If Ignore-Live is true, then we ignore
1007 ;;; the live-TNs, considering only references within Op's VOP.
1009 ;;; We return a conflicting TN, or :OVERFLOW if the TN won't fit.
1010 (defun load-tn-conflicts-in-sc (op sc offset ignore-live
)
1011 (let* ((sb (sc-sb sc
))
1012 (size (finite-sb-current-size sb
)))
1013 (do ((i offset
(1+ i
))
1014 (end (+ offset
(sc-element-size sc
))))
1016 (declare (type index i end
))
1017 (let ((res (or (when (>= i size
) :overflow
)
1018 (and (not ignore-live
)
1019 (svref (finite-sb-live-tns sb
) i
))
1020 (load-tn-offset-conflicts-in-sb op sb i
))))
1021 (when res
(return res
))))))
1023 ;;; If a load-TN for OP is targeted to a legal location in SC, then
1024 ;;; return the offset, otherwise return NIL. We see whether the target
1025 ;;; of the operand is packed, and try that location. There isn't any
1026 ;;; need to chain down the target path, since everything is packed
1029 ;;; We require the target to be in SC (and not merely to overlap with
1030 ;;; SC). This prevents SC information from being lost in load TNs (we
1031 ;;; won't pack a load TN in ANY-REG when it is targeted to a
1032 ;;; DESCRIPTOR-REG.) This shouldn't hurt the code as long as all
1033 ;;; relevant overlapping SCs are allowed in the operand SC
1035 (defun find-load-tn-target (op sc
)
1036 (declare (inline member
))
1037 (let ((target (tn-ref-target op
)))
1039 (let* ((tn (tn-ref-tn target
))
1040 (loc (tn-offset tn
)))
1041 (if (and (eq (tn-sc tn
) sc
)
1042 (member (the index loc
) (sc-locations sc
))
1043 (not (load-tn-conflicts-in-sc op sc loc nil
)))
1047 ;;; Select a legal location for a load TN for Op in SC. We just
1048 ;;; iterate over the SC's locations. If we can't find a legal
1049 ;;; location, return NIL.
1050 (defun select-load-tn-location (op sc
)
1051 (declare (type tn-ref op
) (type sc sc
))
1053 ;; Check any target location first.
1054 (let ((target (tn-ref-target op
)))
1056 (let* ((tn (tn-ref-tn target
))
1057 (loc (tn-offset tn
)))
1058 (when (and (eq (sc-sb sc
) (sc-sb (tn-sc tn
)))
1059 (member (the index loc
) (sc-locations sc
))
1060 (not (load-tn-conflicts-in-sc op sc loc nil
)))
1061 (return-from select-load-tn-location loc
)))))
1063 (dolist (loc (sc-locations sc
) nil
)
1064 (unless (load-tn-conflicts-in-sc op sc loc nil
)
1067 (defevent unpack-tn
"Unpacked a TN to satisfy operand SC restriction.")
1069 ;;; Make TN's location the same as for its save TN (allocating a save
1070 ;;; TN if necessary.) Delete any save/restore code that has been
1071 ;;; emitted thus far. Mark all blocks containing references as needing
1073 (defun unpack-tn (tn)
1075 (let ((stn (or (tn-save-tn tn
)
1076 (pack-save-tn tn
))))
1077 (setf (tn-sc tn
) (tn-sc stn
))
1078 (setf (tn-offset tn
) (tn-offset stn
))
1080 (do ((ref refs
(tn-ref-next ref
)))
1082 (let ((vop (tn-ref-vop ref
)))
1083 (if (eq (vop-info-name (vop-info vop
)) 'move-operand
)
1085 (pushnew (vop-block vop
) *repack-blocks
*))))))
1087 (zot (tn-writes tn
))))
1091 (defevent unpack-fallback
"Unpacked some operand TN.")
1093 ;;; This is called by PACK-LOAD-TN where there isn't any location free
1094 ;;; that we can pack into. What we do is move some live TN in one of
1095 ;;; the specified SCs to memory, then mark all blocks that reference
1096 ;;; the TN as needing repacking. If we succeed, we throw to UNPACKED-TN.
1097 ;;; If we fail, we return NIL.
1099 ;;; We can unpack any live TN that appears in the NORMAL-TNs list
1100 ;;; (isn't wired or restricted.) We prefer to unpack TNs that are not
1101 ;;; used by the VOP. If we can't find any such TN, then we unpack some
1102 ;;; argument or result TN. The only way we can fail is if all
1103 ;;; locations in SC are used by load-TNs or temporaries in VOP.
1104 (defun unpack-for-load-tn (sc op
)
1105 (declare (type sc sc
) (type tn-ref op
))
1106 (let ((sb (sc-sb sc
))
1107 (normal-tns (ir2-component-normal-tns
1108 (component-info *component-being-compiled
*)))
1109 (node (vop-node (tn-ref-vop op
)))
1111 (flet ((unpack-em (victims)
1112 (pushnew (vop-block (tn-ref-vop op
)) *repack-blocks
*)
1113 (dolist (victim victims
)
1114 (event unpack-tn node
)
1116 (throw 'unpacked-tn nil
)))
1117 (dolist (loc (sc-locations sc
))
1118 (declare (type index loc
))
1120 (collect ((victims nil adjoin
))
1122 (end (+ loc
(sc-element-size sc
))))
1124 (declare (type index i end
))
1125 (let ((victim (svref (finite-sb-live-tns sb
) i
)))
1127 (unless (find-in #'tn-next victim normal-tns
)
1131 (let ((conf (load-tn-conflicts-in-sc op sc loc t
)))
1133 (unpack-em (victims)))
1134 ((eq conf
:overflow
))
1136 (cond ((find conf
(victims))
1137 (setq fallback
(victims)))
1138 ((find-in #'tn-next conf normal-tns
)
1139 (setq fallback
(list conf
))))))))))
1142 (event unpack-fallback node
)
1143 (unpack-em fallback
))))
1147 ;;; Try to pack a load TN in the SCs indicated by Load-SCs. If we run
1148 ;;; out of SCs, then we unpack some TN and try again. We return the
1151 ;;; Note: we allow a Load-TN to be packed in the target location even
1152 ;;; if that location is in a SC not allowed by the primitive type.
1153 ;;; (The SC must still be allowed by the operand restriction.) This
1154 ;;; makes move VOPs more efficient, since we won't do a move from the
1155 ;;; stack into a non-descriptor any-reg through a descriptor argument
1156 ;;; load-TN. This does give targeting some real semantics, making it
1157 ;;; not a pure advisory to pack. It allows pack to do some packing it
1158 ;;; wouldn't have done before.
1159 (defun pack-load-tn (load-scs op
)
1160 (declare (type sc-vector load-scs
) (type tn-ref op
))
1161 (let ((vop (tn-ref-vop op
)))
1162 (compute-live-tns (vop-block vop
) vop
))
1164 (let* ((tn (tn-ref-tn op
))
1165 (ptype (tn-primitive-type tn
))
1166 (scs (svref load-scs
(sc-number (tn-sc tn
)))))
1167 (let ((current-scs scs
)
1173 (no-load-scs-allowed-by-primitive-type-error op
))
1174 (dolist (sc allowed
)
1175 (unpack-for-load-tn sc op
))
1176 (failed-to-pack-load-tn-error allowed op
))
1178 (let* ((sc (svref *backend-sc-numbers
* (pop current-scs
)))
1179 (target (find-load-tn-target op sc
)))
1180 (when (or target
(sc-allowed-by-primitive-type sc ptype
))
1181 (let ((loc (or target
1182 (select-load-tn-location op sc
))))
1184 (let ((res (make-tn 0 :load nil sc
)))
1185 (setf (tn-offset res
) loc
)
1187 (push sc allowed
)))))))))
1189 ;;; Scan a list of load-SCs vectors and a list of TN-REFS threaded by
1190 ;;; TN-REF-ACROSS. When we find a reference whose TN doesn't satisfy
1191 ;;; the restriction, we pack a Load-TN and load the operand into it.
1192 ;;; If a load-tn has already been allocated, we can assume that the
1193 ;;; restriction is satisfied.
1194 #!-sb-fluid
(declaim (inline check-operand-restrictions
))
1195 (defun check-operand-restrictions (scs ops
)
1196 (declare (list scs
) (type (or tn-ref null
) ops
))
1198 ;; Check the targeted operands first.
1199 (do ((scs scs
(cdr scs
))
1200 (op ops
(tn-ref-across op
)))
1202 (let ((target (tn-ref-target op
)))
1204 (let* ((load-tn (tn-ref-load-tn op
))
1205 (load-scs (svref (car scs
)
1207 (tn-sc (or load-tn
(tn-ref-tn op
)))))))
1209 (aver (eq load-scs t
))
1210 (unless (eq load-scs t
)
1211 (setf (tn-ref-load-tn op
)
1212 (pack-load-tn (car scs
) op
))))))))
1214 (do ((scs scs
(cdr scs
))
1215 (op ops
(tn-ref-across op
)))
1217 (let ((target (tn-ref-target op
)))
1219 (let* ((load-tn (tn-ref-load-tn op
))
1220 (load-scs (svref (car scs
)
1222 (tn-sc (or load-tn
(tn-ref-tn op
)))))))
1224 (aver (eq load-scs t
))
1225 (unless (eq load-scs t
)
1226 (setf (tn-ref-load-tn op
)
1227 (pack-load-tn (car scs
) op
))))))))
1231 ;;; Scan the VOPs in BLOCK, looking for operands whose SC restrictions
1232 ;;; aren't satisfied. We do the results first, since they are
1233 ;;; evaluated later, and our conflict analysis is a backward scan.
1234 (defun pack-load-tns (block)
1236 (let ((*live-block
* nil
)
1238 (do ((vop (ir2-block-last-vop block
) (vop-prev vop
)))
1240 (let ((info (vop-info vop
)))
1241 (check-operand-restrictions (vop-info-result-load-scs info
)
1243 (check-operand-restrictions (vop-info-arg-load-scs info
)
1249 ;;; Link the TN-REFS READ and WRITE together using the TN-REF-TARGET
1250 ;;; when this seems like a good idea. Currently we always do, as this
1251 ;;; increases the success of load-TN targeting.
1252 (defun target-if-desirable (read write
)
1253 (declare (type tn-ref read write
))
1254 ;; As per the comments at the definition of TN-REF-TARGET, read and
1255 ;; write refs are always paired, with TARGET in the read pointing to
1256 ;; the write and vice versa.
1257 (aver (eq (tn-ref-write-p read
)
1258 (not (tn-ref-write-p write
))))
1259 (setf (tn-ref-target read
) write
)
1260 (setf (tn-ref-target write
) read
))
1262 ;;; If TN can be packed into SC so as to honor a preference to TARGET,
1263 ;;; then return the offset to pack at, otherwise return NIL. TARGET
1264 ;;; must be already packed.
1265 (defun check-ok-target (target tn sc
)
1266 (declare (type tn target tn
) (type sc sc
) (inline member
))
1267 (let* ((loc (tn-offset target
))
1268 (target-sc (tn-sc target
))
1269 (target-sb (sc-sb target-sc
)))
1270 (declare (type index loc
))
1271 ;; We can honor a preference if:
1272 ;; -- TARGET's location is in SC's locations.
1273 ;; -- The element sizes of the two SCs are the same.
1274 ;; -- TN doesn't conflict with target's location.
1275 (if (and (eq target-sb
(sc-sb sc
))
1276 (or (eq (sb-kind target-sb
) :unbounded
)
1277 (member loc
(sc-locations sc
)))
1278 (= (sc-element-size target-sc
) (sc-element-size sc
))
1279 (not (conflicts-in-sc tn sc loc
))
1280 (zerop (mod loc
(sc-alignment sc
))))
1284 ;;; Scan along the target path from TN, looking at readers or writers.
1285 ;;; When we find a TN, call CALLEE with that TN, and then resume
1286 ;;; walking down that TN's target. As soon as there is no target, or
1287 ;;; if the TN has multiple readers (writers), we stop walking the
1288 ;;; targetting chain. We also always stop after 10 iterations to get
1289 ;;; around potential circularity problems.
1291 ;;; Why the single-reader/writer constraint? As far as I can tell,
1292 ;;; this is concerned with straight pipeline of data, e.g. CASTs. In
1293 ;;; that case, limiting to chains of length 10 seems to be more than
1295 (declaim (inline %call-with-target-tns
))
1296 (defun %call-with-target-tns
(tn callee
1297 &key
(limit 10) (reads t
) (writes t
))
1298 (declare (type tn tn
) (type function callee
) (type index limit
))
1299 (flet ((frob-slot (slot-function)
1300 (declare (type function slot-function
))
1303 (declare (type index count
))
1305 (let ((refs (funcall slot-function current
)))
1306 (unless (and (plusp count
)
1308 (not (tn-ref-next refs
)))
1310 (let ((target (tn-ref-target refs
)))
1311 (unless target
(return nil
))
1312 (setq current
(tn-ref-tn target
))
1313 (funcall callee current
)
1316 (frob-slot #'tn-reads
))
1318 (frob-slot #'tn-writes
))
1321 (defmacro do-target-tns
((target-variable source-tn
1322 &rest keys
&key limit reads writes
)
1324 (declare (ignore limit reads writes
))
1325 (let ((callback (sb!xc
:gensym
"CALLBACK")))
1326 `(flet ((,callback
(,target-variable
)
1328 (declare (dynamic-extent #',callback
))
1329 (%call-with-target-tns
,source-tn
#',callback
,@keys
))))
1331 (defun find-ok-target-offset (tn sc
)
1332 (declare (type tn tn
) (type sc sc
))
1333 (do-target-tns (target tn
)
1334 (awhen (and (tn-offset target
)
1335 (check-ok-target target tn sc
))
1336 (return-from find-ok-target-offset it
))))
1338 ;;;; location selection
1340 ;;; Select some location for TN in SC, returning the offset if we
1341 ;;; succeed, and NIL if we fail.
1343 ;;; For :UNBOUNDED SCs just find the smallest correctly aligned offset
1344 ;;; where the TN doesn't conflict with the TNs that have already been
1345 ;;; packed. For :FINITE SCs try to pack the TN into the most heavily
1346 ;;; used locations first (as estimated in FIND-LOCATION-USAGE).
1348 ;;; Historically SELECT-LOCATION tried did the opposite and tried to
1349 ;;; distribute the TNs evenly across the available locations. At least
1350 ;;; on register-starved architectures (x86) this seems to be a bad
1351 ;;; strategy. -- JES 2004-09-11
1352 (defun select-location (tn sc
&key use-reserved-locs optimize
)
1353 (declare (type tn tn
) (type sc sc
) (inline member
))
1354 (let* ((sb (sc-sb sc
))
1355 (element-size (sc-element-size sc
))
1356 (alignment (sc-alignment sc
))
1357 (align-mask (1- alignment
))
1358 (size (finite-sb-current-size sb
)))
1359 (flet ((attempt-location (start-offset)
1360 (let ((conflict (conflicts-in-sc tn sc start-offset
)))
1362 (logandc2 (+ conflict align-mask
1)
1364 (return-from select-location start-offset
)))))
1365 (if (eq (sb-kind sb
) :unbounded
)
1366 (loop with offset
= 0
1367 until
(> (+ offset element-size
) size
) do
1368 (setf offset
(attempt-location offset
)))
1369 (let ((locations (sc-locations sc
)))
1372 (schwartzian-stable-sort-list
1374 :key
(lambda (location-offset)
1375 (loop for offset from location-offset
1378 (finite-sb-always-live-count sb
)
1380 (dolist (offset locations
)
1381 (when (or use-reserved-locs
1383 (sc-reserve-locations sc
))))
1384 (attempt-location offset
))))))))
1386 ;;; If a save TN, return the saved TN, otherwise return TN. This is
1387 ;;; useful for getting the conflicts of a TN that might be a save TN.
1388 (defun original-tn (tn)
1389 (declare (type tn tn
))
1390 (if (member (tn-kind tn
) '(:save
:save-once
:specified-save
))
1397 (declaim (inline unbounded-sc-p
))
1398 (defun unbounded-sc-p (sc)
1399 (eq (sb-kind (sc-sb sc
)) :unbounded
))
1401 (defun unbounded-tn-p (tn)
1402 (unbounded-sc-p (tn-sc tn
)))
1403 (declaim (notinline unbounded-sc-p
))
1405 ;;; Attempt to pack TN in all possible SCs, first in the SC chosen by
1406 ;;; representation selection, then in the alternate SCs in the order
1407 ;;; they were specified in the SC definition. If the TN-COST is
1408 ;;; negative, then we don't attempt to pack in SCs that must be saved.
1409 ;;; If Restricted, then we can only pack in TN-SC, not in any
1412 ;;; If we are attempting to pack in the SC of the save TN for a TN
1413 ;;; with a :SPECIFIED-SAVE TN, then we pack in that location, instead
1414 ;;; of allocating a new stack location.
1415 (defun pack-tn (tn restricted optimize
&key
(allow-unbounded-sc t
))
1416 (declare (type tn tn
))
1417 (aver (not (tn-offset tn
)))
1418 (let* ((original (original-tn tn
))
1420 (alternates (unless restricted
(sc-alternate-scs fsc
)))
1421 (save (tn-save-tn tn
))
1424 (eq (tn-kind save
) :specified-save
))
1426 (do ((sc fsc
(pop alternates
)))
1428 (failed-to-pack-error tn restricted
))
1429 (unless (or allow-unbounded-sc
1430 (not (unbounded-sc-p sc
)))
1432 (when (eq sc specified-save-sc
)
1433 (unless (tn-offset save
)
1434 (pack-tn save nil optimize
))
1435 (setf (tn-offset tn
) (tn-offset save
))
1436 (setf (tn-sc tn
) (tn-sc save
))
1438 (when (or restricted
1439 (not (and (minusp (tn-cost tn
)) (sc-save-p sc
))))
1440 (let ((loc (or (find-ok-target-offset original sc
)
1441 (select-location original sc
:optimize optimize
)
1443 (select-location original sc
:use-reserved-locs t
1444 :optimize optimize
))
1445 (when (unbounded-sc-p sc
)
1447 (or (select-location original sc
)
1448 (error "failed to pack after growing SC?"))))))
1450 (add-location-conflicts original sc loc optimize
)
1451 (setf (tn-sc tn
) sc
)
1452 (setf (tn-offset tn
) loc
)
1456 ;;; Pack a wired TN, checking that the offset is in bounds for the SB,
1457 ;;; and that the TN doesn't conflict with some other TN already packed
1458 ;;; in that location. If the TN is wired to a location beyond the end
1459 ;;; of a :UNBOUNDED SB, then grow the SB enough to hold the TN.
1461 ;;; ### Checking for conflicts is disabled for :SPECIFIED-SAVE TNs.
1462 ;;; This is kind of a hack to make specifying wired stack save
1463 ;;; locations for local call arguments (such as OLD-FP) work, since
1464 ;;; the caller and callee OLD-FP save locations may conflict when the
1465 ;;; save locations don't really (due to being in different frames.)
1466 (defun pack-wired-tn (tn optimize
)
1467 (declare (type tn tn
))
1468 (let* ((sc (tn-sc tn
))
1470 (offset (tn-offset tn
))
1471 (end (+ offset
(sc-element-size sc
)))
1472 (original (original-tn tn
)))
1473 (when (> end
(finite-sb-current-size sb
))
1474 (unless (eq (sb-kind sb
) :unbounded
)
1475 (error "~S is wired to a location that is out of bounds." tn
))
1478 ;; For non-x86 ports the presence of a save-tn associated with a
1479 ;; tn is used to identify the old-fp and return-pc tns. It depends
1480 ;; on the old-fp and return-pc being passed in registers.
1481 #!-
(or x86 x86-64 arm arm64
)
1482 (when (and (not (eq (tn-kind tn
) :specified-save
))
1483 (conflicts-in-sc original sc offset
))
1484 (error "~S is wired to a location that it conflicts with." tn
))
1486 ;; Use the above check, but only print a verbose warning. This can
1487 ;; be helpful for debugging the x86 port.
1489 (when (and (not (eq (tn-kind tn
) :specified-save
))
1490 (conflicts-in-sc original sc offset
))
1491 (format t
"~&* Pack-wired-tn possible conflict:~% ~
1492 tn: ~S; tn-kind: ~S~% ~
1494 sb: ~S; sb-name: ~S; sb-kind: ~S~% ~
1495 offset: ~S; end: ~S~% ~
1497 tn-save-tn: ~S; tn-kind of tn-save-tn: ~S~%"
1499 sb
(sb-name sb
) (sb-kind sb
)
1502 (tn-save-tn tn
) (tn-kind (tn-save-tn tn
))))
1504 ;; On the x86 ports the old-fp and return-pc are often passed on
1505 ;; the stack so the above hack for the other ports does not always
1506 ;; work. Here the old-fp and return-pc tns are identified by being
1507 ;; on the stack in their standard save locations.
1508 #!+(or x86 x86-64 arm arm64
)
1509 (when (and (not (eq (tn-kind tn
) :specified-save
))
1510 (not (and (string= (sb-name sb
)
1511 #!-
(or arm arm64
) "STACK"
1512 #!+(or arm arm64
) "CONTROL-STACK")
1515 (conflicts-in-sc original sc offset
))
1516 (error "~S is wired to location ~D in SC ~A of kind ~S that it conflicts with."
1517 tn offset sc
(tn-kind tn
)))
1519 (add-location-conflicts original sc offset optimize
)))
1521 (defevent repack-block
"Repacked a block due to TN unpacking.")
1523 ;;; KLUDGE: Prior to SBCL version 0.8.9.xx, this function was known as
1524 ;;; PACK-BEFORE-GC-HOOK, but was non-functional since approximately
1525 ;;; version 0.8.3.xx since the removal of GC hooks from the system.
1526 ;;; This currently (as of 2004-04-12) runs now after every call to
1527 ;;; PACK, rather than -- as was originally intended -- once per GC
1528 ;;; cycle; this is probably non-optimal, and might require tuning,
1529 ;;; maybe to be called when the data structures exceed a certain size,
1530 ;;; or maybe once every N times. The KLUDGE is that this rewrite has
1531 ;;; done nothing to improve the reentrance or threadsafety of the
1532 ;;; compiler; it still fails to be callable from several threads at
1535 ;;; Brief experiments indicate that during a compilation cycle this
1536 ;;; causes about 10% more consing, and takes about 1%-2% more time.
1538 ;;; -- CSR, 2004-04-12
1539 (defun clean-up-pack-structures ()
1540 (dolist (sb *backend-sb-list
*)
1541 (unless (eq (sb-kind sb
) :non-packed
)
1542 (let ((size (sb-size sb
)))
1543 (fill (finite-sb-always-live sb
) nil
)
1544 (setf (finite-sb-always-live sb
)
1545 (make-array size
:initial-element
#*))
1546 (setf (finite-sb-always-live-count sb
)
1547 (make-array size
:initial-element
0))
1549 (fill (finite-sb-conflicts sb
) nil
)
1550 (setf (finite-sb-conflicts sb
)
1551 (make-array size
:initial-element
'#()))
1553 (fill (finite-sb-live-tns sb
) nil
)
1554 (setf (finite-sb-live-tns sb
)
1555 (make-array size
:initial-element nil
))))))
1557 (defun tn-lexical-depth (tn)
1558 (let ((path t
)) ; dummy initial value
1559 (labels ((path (lambda)
1561 (lambda lambda
(lambda-parent lambda
)))
1564 (register-scope (lambda)
1565 (let ((new-path (path lambda
)))
1566 (setf path
(if (eql path t
)
1569 0 (mismatch path new-path
))))))
1571 (do ((ref ref
(tn-ref-next ref
)))
1574 (awhen (vop-node (tn-ref-vop ref
))
1575 (register-scope (lexenv-lambda (node-lexenv it
)))))))
1576 (walk-tn-refs (tn-reads tn
))
1577 (walk-tn-refs (tn-writes tn
))
1579 most-positive-fixnum
1582 (declaim (type (member :iterative
:greedy
:adaptive
)
1583 *register-allocation-method
*))
1584 (defvar *register-allocation-method
* :adaptive
)
1586 (declaim (ftype function pack-greedy pack-iterative
))
1588 (defun pack (component)
1590 (let ((optimize nil
)
1592 (2comp (component-info component
)))
1593 (init-sb-vectors component
)
1595 ;; Determine whether we want to do more expensive packing by
1596 ;; checking whether any blocks in the component have (> SPEED
1599 ;; Also, determine if any such block also declares (speed 3),
1600 ;; in which case :adaptive register allocation will switch to
1601 ;; the iterative Chaitin-Briggs spilling/coloring algorithm.
1603 ;; FIXME: This means that a declaration can have a minor
1604 ;; effect even outside its scope, and as the packing is done
1605 ;; component-globally it'd be tricky to use strict scoping. I
1606 ;; think this is still acceptable since it's just a tradeoff
1607 ;; between compilation speed and allocation quality and
1608 ;; doesn't affect the semantics of the generated code in any
1609 ;; way. -- JES 2004-10-06
1610 (do-ir2-blocks (block component
)
1611 (let ((block (block-last (ir2-block-block block
))))
1612 (when (policy block
(> speed compilation-speed
))
1614 (when (policy block
(= speed
3))
1618 ;; Assign costs to normal TNs so we know which ones should always
1619 ;; be packed on the stack, and which are important not to spill.
1620 (when *pack-assign-costs
*
1621 (assign-tn-costs component
))
1623 ;; Actually allocate registers for most TNs. After this, only
1624 ;; :normal tns may be left unallocated (or TNs :restricted to
1625 ;; an unbounded SC).
1626 (funcall (ecase *register-allocation-method
*
1627 (:greedy
#'pack-greedy
)
1628 (:iterative
#'pack-iterative
)
1629 (:adaptive
(if speed-3
#'pack-iterative
#'pack-greedy
)))
1630 component
2comp optimize
)
1632 ;; Pack any leftover normal/restricted TN that is not already
1633 ;; allocated to a finite SC, or TNs that do not appear in any
1634 ;; local TN map (e.g. :MORE TNs). Since we'll likely be
1635 ;; allocating on the stack, first allocate TNs that are
1636 ;; associated with code at shallow lexical depths: this will
1637 ;; allocate long live ranges (i.e. TNs with more conflicts)
1638 ;; first, and hopefully minimise stack fragmentation.
1639 ;; Component TNs are a degenerate case: they are always live.
1640 (let ((component-tns '())
1641 (contiguous-tns '())
1643 (flet ((register-tn (tn)
1644 (unless (tn-offset tn
)
1647 (push tn component-tns
))
1648 ((:environment
:debug-environment
)
1649 (push tn contiguous-tns
))
1652 (do ((tn (ir2-component-restricted-tns 2comp
) (tn-next tn
)))
1654 ;; by this time, restricted TNs must either be
1655 ;; allocated in the right SC or unbounded
1656 (aver (or (tn-offset tn
) (unbounded-tn-p tn
)))
1658 (do ((tn (ir2-component-normal-tns 2comp
) (tn-next tn
)))
1661 (flet ((pack-tns (tns &optional in-order
)
1662 (dolist (tn (if in-order
1664 (schwartzian-stable-sort-list
1665 tns
#'< :key
#'tn-lexical-depth
)))
1666 (unless (tn-offset tn
)
1667 (pack-tn tn nil optimize
)))))
1668 ;; first pack TNs that are known to have simple live
1669 ;; ranges (contiguous lexical scopes)
1670 (pack-tns component-tns t
)
1671 (pack-tns contiguous-tns
)
1674 ;; Do load TN packing and emit saves.
1675 (let ((*repack-blocks
* nil
))
1676 (cond ((and optimize
*pack-optimize-saves
*)
1677 (optimized-emit-saves component
)
1678 (do-ir2-blocks (block component
)
1679 (pack-load-tns block
)))
1681 (do-ir2-blocks (block component
)
1683 (pack-load-tns block
))))
1685 (unless *repack-blocks
* (return))
1686 (let ((orpb *repack-blocks
*))
1687 (setq *repack-blocks
* nil
)
1688 (dolist (block orpb
)
1689 (event repack-block
)
1690 (pack-load-tns block
)))))
1693 (clean-up-pack-structures)))
1695 (defun pack-greedy (component 2comp optimize
)
1696 (declare (type component component
)
1697 (type ir2-component
2comp
))
1698 ;; Pack wired TNs first.
1699 (do ((tn (ir2-component-wired-tns 2comp
) (tn-next tn
)))
1701 (pack-wired-tn tn optimize
))
1703 ;; Then, pack restricted TNs, ones that are live over the whole
1704 ;; component first (they cause no fragmentation). Sort by TN cost
1705 ;; to help important TNs get good targeting.
1706 (collect ((component)
1708 (do ((tn (ir2-component-restricted-tns 2comp
) (tn-next tn
)))
1710 (unless (or (tn-offset tn
) (unbounded-tn-p tn
))
1711 (if (eq :component
(tn-kind tn
))
1714 (flet ((pack-tns (tns)
1715 (dolist (tn (stable-sort tns
#'> :key
#'tn-cost
))
1716 (pack-tn tn t optimize
))))
1717 (pack-tns (component))
1718 (pack-tns (normal))))
1720 (cond ((and *loop-analyze
* *pack-assign-costs
*)
1721 ;; Allocate normal TNs, starting with the TNs that are
1722 ;; heavily used in deep loops (which is taken into account in
1723 ;; TN spill costs). Only allocate in finite SCs (i.e. not on
1726 (do ((tn (ir2-component-normal-tns 2comp
) (tn-next tn
)))
1728 (unless (or (tn-offset tn
)
1729 (eq (tn-kind tn
) :more
)
1731 (and (sc-save-p (tn-sc tn
)) ; SC caller-save, but TN
1732 (minusp (tn-cost tn
)))) ; lives over many calls
1734 (dolist (tn (stable-sort (tns) #'> :key
#'tn-cost
))
1735 (unless (tn-offset tn
)
1736 ;; if it can't fit in a bounded SC, the final pass will
1737 ;; take care of stack packing.
1738 (pack-tn tn nil optimize
:allow-unbounded-sc nil
)))))
1740 ;; If loop analysis has been disabled we might as well revert
1741 ;; to the old behaviour of just packing TNs linearly as they
1743 (do-ir2-blocks (block component
)
1744 (let ((ltns (ir2-block-local-tns block
)))
1745 (do ((i (1- (ir2-block-local-tn-count block
)) (1- i
)))
1747 (declare (fixnum i
))
1748 (let ((tn (svref ltns i
)))
1749 (unless (or (null tn
)
1752 (unbounded-tn-p tn
))
1753 (pack-tn tn nil optimize
:allow-unbounded-sc nil
)))))))))
