1 ;;;; the VM definition arithmetic VOPs for the x86
3 ;;;; This software is part of the SBCL system. See the README file for
6 ;;;; This software is derived from the CMU CL system, which was
7 ;;;; written at Carnegie Mellon University and released into the
8 ;;;; public domain. The software is in the public domain and is
9 ;;;; provided with absolutely no warranty. See the COPYING and CREDITS
10 ;;;; files for more information.
16 (define-vop (fast-safe-arith-op)
21 (define-vop (fixnum-unop fast-safe-arith-op)
22 (:args (x :scs (any-reg) :target res))
23 (:results (res :scs (any-reg)))
24 (:note "inline fixnum arithmetic")
25 (:arg-types tagged-num)
26 (:result-types tagged-num))
28 (define-vop (signed-unop fast-safe-arith-op)
29 (:args (x :scs (signed-reg) :target res))
30 (:results (res :scs (signed-reg)))
31 (:note "inline (signed-byte 32) arithmetic")
32 (:arg-types signed-num)
33 (:result-types signed-num))
35 (define-vop (fast-negate/fixnum fixnum-unop)
41 (define-vop (fast-negate/signed signed-unop)
47 (define-vop (fast-lognot/fixnum fixnum-unop)
51 (inst xor res (fixnumize -1))))
53 (define-vop (fast-lognot/signed signed-unop)
59 ;;;; binary fixnum operations
61 ;;; Assume that any constant operand is the second arg...
63 (define-vop (fast-fixnum-binop fast-safe-arith-op)
64 (:args (x :target r :scs (any-reg)
65 :load-if (not (and (sc-is x control-stack)
67 (sc-is r control-stack)
69 (y :scs (any-reg control-stack)))
70 (:arg-types tagged-num tagged-num)
71 (:results (r :scs (any-reg) :from (:argument 0)
72 :load-if (not (and (sc-is x control-stack)
74 (sc-is r control-stack)
76 (:result-types tagged-num)
77 (:note "inline fixnum arithmetic"))
79 (define-vop (fast-unsigned-binop fast-safe-arith-op)
80 (:args (x :target r :scs (unsigned-reg)
81 :load-if (not (and (sc-is x unsigned-stack)
82 (sc-is y unsigned-reg)
83 (sc-is r unsigned-stack)
85 (y :scs (unsigned-reg unsigned-stack)))
86 (:arg-types unsigned-num unsigned-num)
87 (:results (r :scs (unsigned-reg) :from (:argument 0)
88 :load-if (not (and (sc-is x unsigned-stack)
89 (sc-is y unsigned-reg)
90 (sc-is r unsigned-stack)
92 (:result-types unsigned-num)
93 (:note "inline (unsigned-byte 32) arithmetic"))
95 (define-vop (fast-signed-binop fast-safe-arith-op)
96 (:args (x :target r :scs (signed-reg)
97 :load-if (not (and (sc-is x signed-stack)
99 (sc-is r signed-stack)
101 (y :scs (signed-reg signed-stack)))
102 (:arg-types signed-num signed-num)
103 (:results (r :scs (signed-reg) :from (:argument 0)
104 :load-if (not (and (sc-is x signed-stack)
106 (sc-is r signed-stack)
108 (:result-types signed-num)
109 (:note "inline (signed-byte 32) arithmetic"))
111 (define-vop (fast-fixnum-binop-c fast-safe-arith-op)
112 (:args (x :target r :scs (any-reg control-stack)))
114 (:arg-types tagged-num (:constant (signed-byte 30)))
115 (:results (r :scs (any-reg)
116 :load-if (not (location= x r))))
117 (:result-types tagged-num)
118 (:note "inline fixnum arithmetic"))
120 (define-vop (fast-unsigned-binop-c fast-safe-arith-op)
121 (:args (x :target r :scs (unsigned-reg unsigned-stack)))
123 (:arg-types unsigned-num (:constant (unsigned-byte 32)))
124 (:results (r :scs (unsigned-reg)
125 :load-if (not (location= x r))))
126 (:result-types unsigned-num)
127 (:note "inline (unsigned-byte 32) arithmetic"))
129 (define-vop (fast-signed-binop-c fast-safe-arith-op)
130 (:args (x :target r :scs (signed-reg signed-stack)))
132 (:arg-types signed-num (:constant (signed-byte 32)))
133 (:results (r :scs (signed-reg)
134 :load-if (not (location= x r))))
135 (:result-types signed-num)
136 (:note "inline (signed-byte 32) arithmetic"))
138 (macrolet ((define-binop (translate untagged-penalty op)
140 (define-vop (,(symbolicate "FAST-" translate "/FIXNUM=>FIXNUM")
142 (:translate ,translate)
146 (define-vop (,(symbolicate 'fast- translate '-c/fixnum=>fixnum)
148 (:translate ,translate)
151 (inst ,op r (fixnumize y))))
152 (define-vop (,(symbolicate "FAST-" translate "/SIGNED=>SIGNED")
154 (:translate ,translate)
155 (:generator ,(1+ untagged-penalty)
158 (define-vop (,(symbolicate 'fast- translate '-c/signed=>signed)
160 (:translate ,translate)
161 (:generator ,untagged-penalty
164 (define-vop (,(symbolicate "FAST-"
166 "/UNSIGNED=>UNSIGNED")
168 (:translate ,translate)
169 (:generator ,(1+ untagged-penalty)
172 (define-vop (,(symbolicate 'fast-
174 '-c/unsigned=>unsigned)
175 fast-unsigned-binop-c)
176 (:translate ,translate)
177 (:generator ,untagged-penalty
181 ;;(define-binop + 4 add)
182 (define-binop - 4 sub)
183 (define-binop logand 2 and)
184 (define-binop logior 2 or)
185 (define-binop logxor 2 xor))
188 ;;; Special handling of add on the x86; can use lea to avoid a
189 ;;; register load, otherwise it uses add.
190 (define-vop (fast-+/fixnum=>fixnum fast-safe-arith-op)
192 (:args (x :scs (any-reg) :target r
193 :load-if (not (and (sc-is x control-stack)
195 (sc-is r control-stack)
197 (y :scs (any-reg control-stack)))
198 (:arg-types tagged-num tagged-num)
199 (:results (r :scs (any-reg) :from (:argument 0)
200 :load-if (not (and (sc-is x control-stack)
202 (sc-is r control-stack)
204 (:result-types tagged-num)
205 (:note "inline fixnum arithmetic")
207 (cond ((and (sc-is x any-reg) (sc-is y any-reg) (sc-is r any-reg)
208 (not (location= x r)))
209 (inst lea r (make-ea :dword :base x :index y :scale 1)))
214 (define-vop (fast-+-c/fixnum=>fixnum fast-safe-arith-op)
216 (:args (x :target r :scs (any-reg control-stack)))
218 (:arg-types tagged-num (:constant (signed-byte 30)))
219 (:results (r :scs (any-reg)
220 :load-if (not (location= x r))))
221 (:result-types tagged-num)
222 (:note "inline fixnum arithmetic")
224 (cond ((and (sc-is x any-reg) (sc-is r any-reg) (not (location= x r)))
225 (inst lea r (make-ea :dword :base x :disp (fixnumize y))))
228 (inst add r (fixnumize y))))))
230 (define-vop (fast-+/signed=>signed fast-safe-arith-op)
232 (:args (x :scs (signed-reg) :target r
233 :load-if (not (and (sc-is x signed-stack)
235 (sc-is r signed-stack)
237 (y :scs (signed-reg signed-stack)))
238 (:arg-types signed-num signed-num)
239 (:results (r :scs (signed-reg) :from (:argument 0)
240 :load-if (not (and (sc-is x signed-stack)
243 (:result-types signed-num)
244 (:note "inline (signed-byte 32) arithmetic")
246 (cond ((and (sc-is x signed-reg) (sc-is y signed-reg) (sc-is r signed-reg)
247 (not (location= x r)))
248 (inst lea r (make-ea :dword :base x :index y :scale 1)))
254 ;;;; Special logand cases: (logand signed unsigned) => unsigned
256 (define-vop (fast-logand/signed-unsigned=>unsigned
257 fast-logand/unsigned=>unsigned)
258 (:args (x :target r :scs (signed-reg)
259 :load-if (not (and (sc-is x signed-stack)
260 (sc-is y unsigned-reg)
261 (sc-is r unsigned-stack)
263 (y :scs (unsigned-reg unsigned-stack)))
264 (:arg-types signed-num unsigned-num))
266 (define-vop (fast-logand-c/signed-unsigned=>unsigned
267 fast-logand-c/unsigned=>unsigned)
268 (:args (x :target r :scs (signed-reg signed-stack)))
269 (:arg-types signed-num (:constant (unsigned-byte 32))))
271 (define-vop (fast-logand/unsigned-signed=>unsigned
272 fast-logand/unsigned=>unsigned)
273 (:args (x :target r :scs (unsigned-reg)
274 :load-if (not (and (sc-is x unsigned-stack)
276 (sc-is r unsigned-stack)
278 (y :scs (signed-reg signed-stack)))
279 (:arg-types unsigned-num signed-num))
282 (define-vop (fast-+-c/signed=>signed fast-safe-arith-op)
284 (:args (x :target r :scs (signed-reg signed-stack)))
286 (:arg-types signed-num (:constant (signed-byte 32)))
287 (:results (r :scs (signed-reg)
288 :load-if (not (location= x r))))
289 (:result-types signed-num)
290 (:note "inline (signed-byte 32) arithmetic")
292 (cond ((and (sc-is x signed-reg) (sc-is r signed-reg)
293 (not (location= x r)))
294 (inst lea r (make-ea :dword :base x :disp y)))
301 (define-vop (fast-+/unsigned=>unsigned fast-safe-arith-op)
303 (:args (x :scs (unsigned-reg) :target r
304 :load-if (not (and (sc-is x unsigned-stack)
305 (sc-is y unsigned-reg)
306 (sc-is r unsigned-stack)
308 (y :scs (unsigned-reg unsigned-stack)))
309 (:arg-types unsigned-num unsigned-num)
310 (:results (r :scs (unsigned-reg) :from (:argument 0)
311 :load-if (not (and (sc-is x unsigned-stack)
312 (sc-is y unsigned-reg)
313 (sc-is r unsigned-stack)
315 (:result-types unsigned-num)
316 (:note "inline (unsigned-byte 32) arithmetic")
318 (cond ((and (sc-is x unsigned-reg) (sc-is y unsigned-reg)
319 (sc-is r unsigned-reg) (not (location= x r)))
320 (inst lea r (make-ea :dword :base x :index y :scale 1)))
325 (define-vop (fast-+-c/unsigned=>unsigned fast-safe-arith-op)
327 (:args (x :target r :scs (unsigned-reg unsigned-stack)))
329 (:arg-types unsigned-num (:constant (unsigned-byte 32)))
330 (:results (r :scs (unsigned-reg)
331 :load-if (not (location= x r))))
332 (:result-types unsigned-num)
333 (:note "inline (unsigned-byte 32) arithmetic")
335 (cond ((and (sc-is x unsigned-reg) (sc-is r unsigned-reg)
336 (not (location= x r)))
337 (inst lea r (make-ea :dword :base x :disp y)))
344 ;;;; multiplication and division
346 (define-vop (fast-*/fixnum=>fixnum fast-safe-arith-op)
348 ;; We need different loading characteristics.
349 (:args (x :scs (any-reg) :target r)
350 (y :scs (any-reg control-stack)))
351 (:arg-types tagged-num tagged-num)
352 (:results (r :scs (any-reg) :from (:argument 0)))
353 (:result-types tagged-num)
354 (:note "inline fixnum arithmetic")
360 (define-vop (fast-*-c/fixnum=>fixnum fast-safe-arith-op)
362 ;; We need different loading characteristics.
363 (:args (x :scs (any-reg control-stack)))
365 (:arg-types tagged-num (:constant (signed-byte 30)))
366 (:results (r :scs (any-reg)))
367 (:result-types tagged-num)
368 (:note "inline fixnum arithmetic")
372 (define-vop (fast-*/signed=>signed fast-safe-arith-op)
374 ;; We need different loading characteristics.
375 (:args (x :scs (signed-reg) :target r)
376 (y :scs (signed-reg signed-stack)))
377 (:arg-types signed-num signed-num)
378 (:results (r :scs (signed-reg) :from (:argument 0)))
379 (:result-types signed-num)
380 (:note "inline (signed-byte 32) arithmetic")
385 (define-vop (fast-*-c/signed=>signed fast-safe-arith-op)
387 ;; We need different loading characteristics.
388 (:args (x :scs (signed-reg signed-stack)))
390 (:arg-types signed-num (:constant (signed-byte 32)))
391 (:results (r :scs (signed-reg)))
392 (:result-types signed-num)
393 (:note "inline (signed-byte 32) arithmetic")
397 (define-vop (fast-*/unsigned=>unsigned fast-safe-arith-op)
399 (:args (x :scs (unsigned-reg) :target eax)
400 (y :scs (unsigned-reg unsigned-stack)))
401 (:arg-types unsigned-num unsigned-num)
402 (:temporary (:sc unsigned-reg :offset eax-offset :target result
403 :from (:argument 0) :to :result) eax)
404 (:temporary (:sc unsigned-reg :offset edx-offset
405 :from :eval :to :result) edx)
407 (:results (result :scs (unsigned-reg)))
408 (:result-types unsigned-num)
409 (:note "inline (unsigned-byte 32) arithmetic")
411 (:save-p :compute-only)
418 (define-vop (fast-truncate/fixnum=>fixnum fast-safe-arith-op)
419 (:translate truncate)
420 (:args (x :scs (any-reg) :target eax)
421 (y :scs (any-reg control-stack)))
422 (:arg-types tagged-num tagged-num)
423 (:temporary (:sc signed-reg :offset eax-offset :target quo
424 :from (:argument 0) :to (:result 0)) eax)
425 (:temporary (:sc unsigned-reg :offset edx-offset :target rem
426 :from (:argument 0) :to (:result 1)) edx)
427 (:results (quo :scs (any-reg))
428 (rem :scs (any-reg)))
429 (:result-types tagged-num tagged-num)
430 (:note "inline fixnum arithmetic")
432 (:save-p :compute-only)
434 (let ((zero (generate-error-code vop division-by-zero-error x y)))
435 (if (sc-is y any-reg)
436 (inst test y y) ; smaller instruction
442 (if (location= quo eax)
444 (inst lea quo (make-ea :dword :index eax :scale 4)))
447 (define-vop (fast-truncate-c/fixnum=>fixnum fast-safe-arith-op)
448 (:translate truncate)
449 (:args (x :scs (any-reg) :target eax))
451 (:arg-types tagged-num (:constant (signed-byte 30)))
452 (:temporary (:sc signed-reg :offset eax-offset :target quo
453 :from :argument :to (:result 0)) eax)
454 (:temporary (:sc any-reg :offset edx-offset :target rem
455 :from :eval :to (:result 1)) edx)
456 (:temporary (:sc any-reg :from :eval :to :result) y-arg)
457 (:results (quo :scs (any-reg))
458 (rem :scs (any-reg)))
459 (:result-types tagged-num tagged-num)
460 (:note "inline fixnum arithmetic")
462 (:save-p :compute-only)
466 (inst mov y-arg (fixnumize y))
467 (inst idiv eax y-arg)
468 (if (location= quo eax)
470 (inst lea quo (make-ea :dword :index eax :scale 4)))
473 (define-vop (fast-truncate/unsigned=>unsigned fast-safe-arith-op)
474 (:translate truncate)
475 (:args (x :scs (unsigned-reg) :target eax)
476 (y :scs (unsigned-reg signed-stack)))
477 (:arg-types unsigned-num unsigned-num)
478 (:temporary (:sc unsigned-reg :offset eax-offset :target quo
479 :from (:argument 0) :to (:result 0)) eax)
480 (:temporary (:sc unsigned-reg :offset edx-offset :target rem
481 :from (:argument 0) :to (:result 1)) edx)
482 (:results (quo :scs (unsigned-reg))
483 (rem :scs (unsigned-reg)))
484 (:result-types unsigned-num unsigned-num)
485 (:note "inline (unsigned-byte 32) arithmetic")
487 (:save-p :compute-only)
489 (let ((zero (generate-error-code vop division-by-zero-error x y)))
490 (if (sc-is y unsigned-reg)
491 (inst test y y) ; smaller instruction
500 (define-vop (fast-truncate-c/unsigned=>unsigned fast-safe-arith-op)
501 (:translate truncate)
502 (:args (x :scs (unsigned-reg) :target eax))
504 (:arg-types unsigned-num (:constant (unsigned-byte 32)))
505 (:temporary (:sc unsigned-reg :offset eax-offset :target quo
506 :from :argument :to (:result 0)) eax)
507 (:temporary (:sc unsigned-reg :offset edx-offset :target rem
508 :from :eval :to (:result 1)) edx)
509 (:temporary (:sc unsigned-reg :from :eval :to :result) y-arg)
510 (:results (quo :scs (unsigned-reg))
511 (rem :scs (unsigned-reg)))
512 (:result-types unsigned-num unsigned-num)
513 (:note "inline (unsigned-byte 32) arithmetic")
515 (:save-p :compute-only)
524 (define-vop (fast-truncate/signed=>signed fast-safe-arith-op)
525 (:translate truncate)
526 (:args (x :scs (signed-reg) :target eax)
527 (y :scs (signed-reg signed-stack)))
528 (:arg-types signed-num signed-num)
529 (:temporary (:sc signed-reg :offset eax-offset :target quo
530 :from (:argument 0) :to (:result 0)) eax)
531 (:temporary (:sc signed-reg :offset edx-offset :target rem
532 :from (:argument 0) :to (:result 1)) edx)
533 (:results (quo :scs (signed-reg))
534 (rem :scs (signed-reg)))
535 (:result-types signed-num signed-num)
536 (:note "inline (signed-byte 32) arithmetic")
538 (:save-p :compute-only)
540 (let ((zero (generate-error-code vop division-by-zero-error x y)))
541 (if (sc-is y signed-reg)
542 (inst test y y) ; smaller instruction
551 (define-vop (fast-truncate-c/signed=>signed fast-safe-arith-op)
552 (:translate truncate)
553 (:args (x :scs (signed-reg) :target eax))
555 (:arg-types signed-num (:constant (signed-byte 32)))
556 (:temporary (:sc signed-reg :offset eax-offset :target quo
557 :from :argument :to (:result 0)) eax)
558 (:temporary (:sc signed-reg :offset edx-offset :target rem
559 :from :eval :to (:result 1)) edx)
560 (:temporary (:sc signed-reg :from :eval :to :result) y-arg)
561 (:results (quo :scs (signed-reg))
562 (rem :scs (signed-reg)))
563 (:result-types signed-num signed-num)
564 (:note "inline (signed-byte 32) arithmetic")
566 (:save-p :compute-only)
571 (inst idiv eax y-arg)
578 (define-vop (fast-ash-c/fixnum=>fixnum)
581 (:args (number :scs (any-reg) :target result
582 :load-if (not (and (sc-is number any-reg control-stack)
583 (sc-is result any-reg control-stack)
584 (location= number result)))))
586 (:arg-types tagged-num (:constant integer))
587 (:results (result :scs (any-reg)
588 :load-if (not (and (sc-is number control-stack)
589 (sc-is result control-stack)
590 (location= number result)))))
591 (:result-types tagged-num)
594 (cond ((and (= amount 1) (not (location= number result)))
595 (inst lea result (make-ea :dword :index number :scale 2)))
596 ((and (= amount 2) (not (location= number result)))
597 (inst lea result (make-ea :dword :index number :scale 4)))
598 ((and (= amount 3) (not (location= number result)))
599 (inst lea result (make-ea :dword :index number :scale 8)))
602 (cond ((plusp amount)
603 ;; We don't have to worry about overflow because of the
604 ;; result type restriction.
605 (inst shl result amount))
607 ;; If the amount is greater than 31, only shift by 31. We
608 ;; have to do this because the shift instructions only look
609 ;; at the low five bits of the result.
610 (inst sar result (min 31 (- amount)))
611 ;; Fixnum correction.
612 (inst and result #xfffffffc)))))))
614 (define-vop (fast-ash-left/fixnum=>fixnum)
616 (:args (number :scs (any-reg) :target result
617 :load-if (not (and (sc-is number control-stack)
618 (sc-is result control-stack)
619 (location= number result))))
620 (amount :scs (unsigned-reg) :target ecx))
621 (:arg-types tagged-num positive-fixnum)
622 (:temporary (:sc unsigned-reg :offset ecx-offset :from (:argument 1)) ecx)
623 (:results (result :scs (any-reg) :from (:argument 0)
624 :load-if (not (and (sc-is number control-stack)
625 (sc-is result control-stack)
626 (location= number result)))))
627 (:result-types tagged-num)
633 ;; The result-type ensures us that this shift will not overflow.
634 (inst shl result :cl)))
636 (define-vop (fast-ash-c)
639 (:args (number :scs (signed-reg unsigned-reg) :target result
640 :load-if (not (and (sc-is number signed-stack unsigned-stack)
641 (sc-is result signed-stack unsigned-stack)
642 (location= number result)))))
644 (:arg-types (:or signed-num unsigned-num) (:constant integer))
645 (:results (result :scs (signed-reg unsigned-reg)
647 (and (sc-is number signed-stack unsigned-stack)
648 (sc-is result signed-stack unsigned-stack)
649 (location= number result)))))
650 (:result-types (:or signed-num unsigned-num))
653 (cond ((and (= amount 1) (not (location= number result)))
654 (inst lea result (make-ea :dword :index number :scale 2)))
655 ((and (= amount 2) (not (location= number result)))
656 (inst lea result (make-ea :dword :index number :scale 4)))
657 ((and (= amount 3) (not (location= number result)))
658 (inst lea result (make-ea :dword :index number :scale 8)))
661 (cond ((plusp amount)
662 ;; We don't have to worry about overflow because of the
663 ;; result type restriction.
664 (inst shl result amount))
665 ((sc-is number signed-reg signed-stack)
666 ;; If the amount is greater than 31, only shift by 31. We
667 ;; have to do this because the shift instructions only look
668 ;; at the low five bits of the result.
669 (inst sar result (min 31 (- amount))))
671 (inst shr result (min 31 (- amount)))))))))
673 (define-vop (fast-ash-left)
675 (:args (number :scs (signed-reg unsigned-reg) :target result
676 :load-if (not (and (sc-is number signed-stack unsigned-stack)
677 (sc-is result signed-stack unsigned-stack)
678 (location= number result))))
679 (amount :scs (unsigned-reg) :target ecx))
680 (:arg-types (:or signed-num unsigned-num) positive-fixnum)
681 (:temporary (:sc unsigned-reg :offset ecx-offset :from (:argument 1)) ecx)
682 (:results (result :scs (signed-reg unsigned-reg) :from (:argument 0)
684 (and (sc-is number signed-stack unsigned-stack)
685 (sc-is result signed-stack unsigned-stack)
686 (location= number result)))))
687 (:result-types (:or signed-num unsigned-num))
693 ;; The result-type ensures us that this shift will not overflow.
694 (inst shl result :cl)))
696 (define-vop (fast-ash)
699 (:args (number :scs (signed-reg unsigned-reg) :target result)
700 (amount :scs (signed-reg) :target ecx))
701 (:arg-types (:or signed-num unsigned-num) signed-num)
702 (:results (result :scs (signed-reg unsigned-reg) :from (:argument 0)))
703 (:result-types (:or signed-num unsigned-num))
704 (:temporary (:sc signed-reg :offset ecx-offset :from (:argument 1)) ecx)
710 (inst jmp :ns positive)
717 (signed-reg (inst sar result :cl))
718 (unsigned-reg (inst shr result :cl)))
722 ;; The result-type ensures us that this shift will not overflow.
723 (inst shl result :cl)
727 ;;; Note: documentation for this function is wrong - rtfm
728 (define-vop (signed-byte-32-len)
729 (:translate integer-length)
730 (:note "inline (signed-byte 32) integer-length")
732 (:args (arg :scs (signed-reg) :target res))
733 (:arg-types signed-num)
734 (:results (res :scs (unsigned-reg)))
735 (:result-types unsigned-num)
750 (define-vop (unsigned-byte-32-len)
751 (:translate integer-length)
752 (:note "inline (unsigned-byte 32) integer-length")
754 (:args (arg :scs (unsigned-reg)))
755 (:arg-types unsigned-num)
756 (:results (res :scs (unsigned-reg)))
757 (:result-types unsigned-num)
767 (define-vop (unsigned-byte-32-count)
768 (:translate logcount)
769 (:note "inline (unsigned-byte 32) logcount")
771 (:args (arg :scs (unsigned-reg)))
772 (:arg-types unsigned-num)
773 (:results (result :scs (unsigned-reg)))
774 (:result-types positive-fixnum)
775 (:temporary (:sc unsigned-reg :from (:argument 0)) temp)
779 (inst mov temp result)
781 (inst and result #x55555555)
782 (inst and temp #x55555555)
783 (inst add result temp)
785 (inst mov temp result)
787 (inst and result #x33333333)
788 (inst and temp #x33333333)
789 (inst add result temp)
791 (inst mov temp result)
793 (inst and result #x0f0f0f0f)
794 (inst and temp #x0f0f0f0f)
795 (inst add result temp)
797 (inst mov temp result)
799 (inst and result #x00ff00ff)
800 (inst and temp #x00ff00ff)
801 (inst add result temp)
803 (inst mov temp result)
805 (inst and result #x0000ffff)
806 (inst and temp #x0000ffff)
807 (inst add result temp)))
809 ;;;; binary conditional VOPs
811 (define-vop (fast-conditional)
816 (:policy :fast-safe))
818 (define-vop (fast-conditional/fixnum fast-conditional)
819 (:args (x :scs (any-reg)
820 :load-if (not (and (sc-is x control-stack)
822 (y :scs (any-reg control-stack)))
823 (:arg-types tagged-num tagged-num)
824 (:note "inline fixnum comparison"))
826 (define-vop (fast-conditional-c/fixnum fast-conditional/fixnum)
827 (:args (x :scs (any-reg control-stack)))
828 (:arg-types tagged-num (:constant (signed-byte 30)))
829 (:info target not-p y))
831 (define-vop (fast-conditional/signed fast-conditional)
832 (:args (x :scs (signed-reg)
833 :load-if (not (and (sc-is x signed-stack)
834 (sc-is y signed-reg))))
835 (y :scs (signed-reg signed-stack)))
836 (:arg-types signed-num signed-num)
837 (:note "inline (signed-byte 32) comparison"))
839 (define-vop (fast-conditional-c/signed fast-conditional/signed)
840 (:args (x :scs (signed-reg signed-stack)))
841 (:arg-types signed-num (:constant (signed-byte 32)))
842 (:info target not-p y))
844 (define-vop (fast-conditional/unsigned fast-conditional)
845 (:args (x :scs (unsigned-reg)
846 :load-if (not (and (sc-is x unsigned-stack)
847 (sc-is y unsigned-reg))))
848 (y :scs (unsigned-reg unsigned-stack)))
849 (:arg-types unsigned-num unsigned-num)
850 (:note "inline (unsigned-byte 32) comparison"))
852 (define-vop (fast-conditional-c/unsigned fast-conditional/unsigned)
853 (:args (x :scs (unsigned-reg unsigned-stack)))
854 (:arg-types unsigned-num (:constant (unsigned-byte 32)))
855 (:info target not-p y))
858 (macrolet ((define-conditional-vop (tran cond unsigned not-cond not-unsigned)
861 (lambda (suffix cost signed)
862 `(define-vop (;; FIXME: These could be done more
863 ;; cleanly with SYMBOLICATE.
864 ,(intern (format nil "~:@(FAST-IF-~A~A~)"
867 (format nil "~:@(FAST-CONDITIONAL~A~)"
872 ,(if (eq suffix '-c/fixnum)
883 '(/fixnum -c/fixnum /signed -c/signed /unsigned -c/unsigned)
885 '(t t t t nil nil)))))
887 (define-conditional-vop < :l :b :ge :ae)
888 (define-conditional-vop > :g :a :le :be))
890 (define-vop (fast-if-eql/signed fast-conditional/signed)
894 (inst jmp (if not-p :ne :e) target)))
896 (define-vop (fast-if-eql-c/signed fast-conditional-c/signed)
899 (cond ((and (sc-is x signed-reg) (zerop y))
900 (inst test x x)) ; smaller instruction
903 (inst jmp (if not-p :ne :e) target)))
905 (define-vop (fast-if-eql/unsigned fast-conditional/unsigned)
909 (inst jmp (if not-p :ne :e) target)))
911 (define-vop (fast-if-eql-c/unsigned fast-conditional-c/unsigned)
914 (cond ((and (sc-is x unsigned-reg) (zerop y))
915 (inst test x x)) ; smaller instruction
918 (inst jmp (if not-p :ne :e) target)))
920 ;;; EQL/FIXNUM is funny because the first arg can be of any type, not just a
923 ;;; These versions specify a fixnum restriction on their first arg. We have
924 ;;; also generic-eql/fixnum VOPs which are the same, but have no restriction on
925 ;;; the first arg and a higher cost. The reason for doing this is to prevent
926 ;;; fixnum specific operations from being used on word integers, spuriously
927 ;;; consing the argument.
929 (define-vop (fast-eql/fixnum fast-conditional)
930 (:args (x :scs (any-reg)
931 :load-if (not (and (sc-is x control-stack)
933 (y :scs (any-reg control-stack)))
934 (:arg-types tagged-num tagged-num)
935 (:note "inline fixnum comparison")
939 (inst jmp (if not-p :ne :e) target)))
940 (define-vop (generic-eql/fixnum fast-eql/fixnum)
941 (:args (x :scs (any-reg descriptor-reg)
942 :load-if (not (and (sc-is x control-stack)
944 (y :scs (any-reg control-stack)))
945 (:arg-types * tagged-num)
948 (define-vop (fast-eql-c/fixnum fast-conditional/fixnum)
949 (:args (x :scs (any-reg control-stack)))
950 (:arg-types tagged-num (:constant (signed-byte 30)))
951 (:info target not-p y)
954 (cond ((and (sc-is x any-reg) (zerop y))
955 (inst test x x)) ; smaller instruction
957 (inst cmp x (fixnumize y))))
958 (inst jmp (if not-p :ne :e) target)))
959 (define-vop (generic-eql-c/fixnum fast-eql-c/fixnum)
960 (:args (x :scs (any-reg descriptor-reg control-stack)))
961 (:arg-types * (:constant (signed-byte 30)))
964 ;;;; 32-bit logical operations
966 (define-vop (merge-bits)
967 (:translate merge-bits)
968 (:args (shift :scs (signed-reg unsigned-reg) :target ecx)
969 (prev :scs (unsigned-reg) :target result)
970 (next :scs (unsigned-reg)))
971 (:arg-types tagged-num unsigned-num unsigned-num)
972 (:temporary (:sc signed-reg :offset ecx-offset :from (:argument 0)) ecx)
973 (:results (result :scs (unsigned-reg) :from (:argument 1)))
974 (:result-types unsigned-num)
979 (inst shrd result next :cl)))
981 (define-source-transform 32bit-logical-not (x)
982 `(logand (lognot (the (unsigned-byte 32) ,x)) #.(1- (ash 1 32))))
984 (deftransform 32bit-logical-and ((x y))
987 (define-source-transform 32bit-logical-nand (x y)
988 `(32bit-logical-not (32bit-logical-and ,x ,y)))
990 (deftransform 32bit-logical-or ((x y))
993 (define-source-transform 32bit-logical-nor (x y)
994 `(32bit-logical-not (32bit-logical-or ,x ,y)))
996 (deftransform 32bit-logical-xor ((x y))
999 (define-source-transform 32bit-logical-eqv (x y)
1000 `(32bit-logical-not (32bit-logical-xor ,x ,y)))
1002 (define-source-transform 32bit-logical-orc1 (x y)
1003 `(32bit-logical-or (32bit-logical-not ,x) ,y))
1005 (define-source-transform 32bit-logical-orc2 (x y)
1006 `(32bit-logical-or ,x (32bit-logical-not ,y)))
1008 (define-source-transform 32bit-logical-andc1 (x y)
1009 `(32bit-logical-and (32bit-logical-not ,x) ,y))
1011 (define-source-transform 32bit-logical-andc2 (x y)
1012 `(32bit-logical-and ,x (32bit-logical-not ,y)))
1014 ;;; Only the lower 5 bits of the shift amount are significant.
1015 (define-vop (shift-towards-someplace)
1016 (:policy :fast-safe)
1017 (:args (num :scs (unsigned-reg) :target r)
1018 (amount :scs (signed-reg) :target ecx))
1019 (:arg-types unsigned-num tagged-num)
1020 (:temporary (:sc signed-reg :offset ecx-offset :from (:argument 1)) ecx)
1021 (:results (r :scs (unsigned-reg) :from (:argument 0)))
1022 (:result-types unsigned-num))
1024 (define-vop (shift-towards-start shift-towards-someplace)
1025 (:translate shift-towards-start)
1026 (:note "SHIFT-TOWARDS-START")
1032 (define-vop (shift-towards-end shift-towards-someplace)
1033 (:translate shift-towards-end)
1034 (:note "SHIFT-TOWARDS-END")
1040 ;;;; Modular functions
1042 (define-modular-fun +-mod32 (x y) + 32)
1043 (define-vop (fast-+-mod32/unsigned=>unsigned fast-+/unsigned=>unsigned)
1044 (:translate +-mod32))
1045 (define-vop (fast-+-mod32-c/unsigned=>unsigned fast-+-c/unsigned=>unsigned)
1046 (:translate +-mod32))
1048 ;;; logical operations
1049 (define-modular-fun lognot-mod32 (x) lognot 32)
1050 (define-vop (lognot-mod32/unsigned=>unsigned)
1051 (:translate lognot-mod32)
1052 (:args (x :scs (unsigned-reg unsigned-stack) :target r
1053 :load-if (not (and (sc-is x unsigned-stack)
1054 (sc-is r unsigned-stack)
1056 (:arg-types unsigned-num)
1057 (:results (r :scs (unsigned-reg)
1058 :load-if (not (and (sc-is x unsigned-stack)
1059 (sc-is r unsigned-stack)
1061 (:result-types unsigned-num)
1062 (:policy :fast-safe)
1067 (define-modular-fun logxor-mod32 (x y) logxor 32)
1068 (define-vop (fast-logxor-mod32/unsigned=>unsigned
1069 fast-logxor/unsigned=>unsigned)
1070 (:translate logxor-mod32))
1071 (define-vop (fast-logxor-mod32-c/unsigned=>unsigned
1072 fast-logxor-c/unsigned=>unsigned)
1073 (:translate logxor-mod32))
1077 (define-vop (bignum-length get-header-data)
1078 (:translate sb!bignum::%bignum-length)
1079 (:policy :fast-safe))
1081 (define-vop (bignum-set-length set-header-data)
1082 (:translate sb!bignum::%bignum-set-length)
1083 (:policy :fast-safe))
1085 (define-full-reffer bignum-ref * bignum-digits-offset other-pointer-lowtag
1086 (unsigned-reg) unsigned-num sb!bignum::%bignum-ref)
1088 (define-full-setter bignum-set * bignum-digits-offset other-pointer-lowtag
1089 (unsigned-reg) unsigned-num sb!bignum::%bignum-set)
1091 (define-vop (digit-0-or-plus)
1092 (:translate sb!bignum::%digit-0-or-plusp)
1093 (:policy :fast-safe)
1094 (:args (digit :scs (unsigned-reg)))
1095 (:arg-types unsigned-num)
1097 (:info target not-p)
1099 (inst or digit digit)
1100 (inst jmp (if not-p :s :ns) target)))
1103 ;;; For add and sub with carry the sc of carry argument is any-reg so
1104 ;;; the it may be passed as a fixnum or word and thus may be 0, 1, or
1105 ;;; 4. This is easy to deal with and may save a fixnum-word
1107 (define-vop (add-w/carry)
1108 (:translate sb!bignum::%add-with-carry)
1109 (:policy :fast-safe)
1110 (:args (a :scs (unsigned-reg) :target result)
1111 (b :scs (unsigned-reg unsigned-stack) :to :eval)
1112 (c :scs (any-reg) :target temp))
1113 (:arg-types unsigned-num unsigned-num positive-fixnum)
1114 (:temporary (:sc any-reg :from (:argument 2) :to :eval) temp)
1115 (:results (result :scs (unsigned-reg) :from (:argument 0))
1116 (carry :scs (unsigned-reg)))
1117 (:result-types unsigned-num positive-fixnum)
1121 (inst neg temp) ; Set the carry flag to 0 if c=0 else to 1
1124 (inst adc carry carry)))
1126 ;;; Note: the borrow is the oppostite of the x86 convention - 1 for no
1127 ;;; borrow and 0 for a borrow.
1128 (define-vop (sub-w/borrow)
1129 (:translate sb!bignum::%subtract-with-borrow)
1130 (:policy :fast-safe)
1131 (:args (a :scs (unsigned-reg) :to :eval :target result)
1132 (b :scs (unsigned-reg unsigned-stack) :to :result)
1133 (c :scs (any-reg control-stack)))
1134 (:arg-types unsigned-num unsigned-num positive-fixnum)
1135 (:results (result :scs (unsigned-reg) :from :eval)
1136 (borrow :scs (unsigned-reg)))
1137 (:result-types unsigned-num positive-fixnum)
1139 (inst cmp c 1) ; Set the carry flag to 1 if c=0 else to 0
1143 (inst adc borrow borrow)
1144 (inst xor borrow 1)))
1147 (define-vop (bignum-mult-and-add-3-arg)
1148 (:translate sb!bignum::%multiply-and-add)
1149 (:policy :fast-safe)
1150 (:args (x :scs (unsigned-reg) :target eax)
1151 (y :scs (unsigned-reg unsigned-stack))
1152 (carry-in :scs (unsigned-reg unsigned-stack)))
1153 (:arg-types unsigned-num unsigned-num unsigned-num)
1154 (:temporary (:sc unsigned-reg :offset eax-offset :from (:argument 0)
1155 :to (:result 1) :target lo) eax)
1156 (:temporary (:sc unsigned-reg :offset edx-offset :from (:argument 1)
1157 :to (:result 0) :target hi) edx)
1158 (:results (hi :scs (unsigned-reg))
1159 (lo :scs (unsigned-reg)))
1160 (:result-types unsigned-num unsigned-num)
1164 (inst add eax carry-in)
1169 (define-vop (bignum-mult-and-add-4-arg)
1170 (:translate sb!bignum::%multiply-and-add)
1171 (:policy :fast-safe)
1172 (:args (x :scs (unsigned-reg) :target eax)
1173 (y :scs (unsigned-reg unsigned-stack))
1174 (prev :scs (unsigned-reg unsigned-stack))
1175 (carry-in :scs (unsigned-reg unsigned-stack)))
1176 (:arg-types unsigned-num unsigned-num unsigned-num unsigned-num)
1177 (:temporary (:sc unsigned-reg :offset eax-offset :from (:argument 0)
1178 :to (:result 1) :target lo) eax)
1179 (:temporary (:sc unsigned-reg :offset edx-offset :from (:argument 1)
1180 :to (:result 0) :target hi) edx)
1181 (:results (hi :scs (unsigned-reg))
1182 (lo :scs (unsigned-reg)))
1183 (:result-types unsigned-num unsigned-num)
1189 (inst add eax carry-in)
1195 (define-vop (bignum-mult)
1196 (:translate sb!bignum::%multiply)
1197 (:policy :fast-safe)
1198 (:args (x :scs (unsigned-reg) :target eax)
1199 (y :scs (unsigned-reg unsigned-stack)))
1200 (:arg-types unsigned-num unsigned-num)
1201 (:temporary (:sc unsigned-reg :offset eax-offset :from (:argument 0)
1202 :to (:result 1) :target lo) eax)
1203 (:temporary (:sc unsigned-reg :offset edx-offset :from (:argument 1)
1204 :to (:result 0) :target hi) edx)
1205 (:results (hi :scs (unsigned-reg))
1206 (lo :scs (unsigned-reg)))
1207 (:result-types unsigned-num unsigned-num)
1214 (define-vop (bignum-lognot lognot-mod32/unsigned=>unsigned)
1215 (:translate sb!bignum::%lognot))
1217 (define-vop (fixnum-to-digit)
1218 (:translate sb!bignum::%fixnum-to-digit)
1219 (:policy :fast-safe)
1220 (:args (fixnum :scs (any-reg control-stack) :target digit))
1221 (:arg-types tagged-num)
1222 (:results (digit :scs (unsigned-reg)
1223 :load-if (not (and (sc-is fixnum control-stack)
1224 (sc-is digit unsigned-stack)
1225 (location= fixnum digit)))))
1226 (:result-types unsigned-num)
1229 (inst sar digit 2)))
1231 (define-vop (bignum-floor)
1232 (:translate sb!bignum::%floor)
1233 (:policy :fast-safe)
1234 (:args (div-high :scs (unsigned-reg) :target edx)
1235 (div-low :scs (unsigned-reg) :target eax)
1236 (divisor :scs (unsigned-reg unsigned-stack)))
1237 (:arg-types unsigned-num unsigned-num unsigned-num)
1238 (:temporary (:sc unsigned-reg :offset eax-offset :from (:argument 1)
1239 :to (:result 0) :target quo) eax)
1240 (:temporary (:sc unsigned-reg :offset edx-offset :from (:argument 0)
1241 :to (:result 1) :target rem) edx)
1242 (:results (quo :scs (unsigned-reg))
1243 (rem :scs (unsigned-reg)))
1244 (:result-types unsigned-num unsigned-num)
1248 (inst div eax divisor)
1252 (define-vop (signify-digit)
1253 (:translate sb!bignum::%fixnum-digit-with-correct-sign)
1254 (:policy :fast-safe)
1255 (:args (digit :scs (unsigned-reg unsigned-stack) :target res))
1256 (:arg-types unsigned-num)
1257 (:results (res :scs (any-reg signed-reg)
1258 :load-if (not (and (sc-is digit unsigned-stack)
1259 (sc-is res control-stack signed-stack)
1260 (location= digit res)))))
1261 (:result-types signed-num)
1264 (when (sc-is res any-reg control-stack)
1267 (define-vop (digit-ashr)
1268 (:translate sb!bignum::%ashr)
1269 (:policy :fast-safe)
1270 (:args (digit :scs (unsigned-reg unsigned-stack) :target result)
1271 (count :scs (unsigned-reg) :target ecx))
1272 (:arg-types unsigned-num positive-fixnum)
1273 (:temporary (:sc unsigned-reg :offset ecx-offset :from (:argument 1)) ecx)
1274 (:results (result :scs (unsigned-reg) :from (:argument 0)
1275 :load-if (not (and (sc-is result unsigned-stack)
1276 (location= digit result)))))
1277 (:result-types unsigned-num)
1281 (inst sar result :cl)))
1283 (define-vop (digit-lshr digit-ashr)
1284 (:translate sb!bignum::%digit-logical-shift-right)
1288 (inst shr result :cl)))
1290 (define-vop (digit-ashl digit-ashr)
1291 (:translate sb!bignum::%ashl)
1295 (inst shl result :cl)))
1297 ;;;; static functions
1299 (define-static-fun two-arg-/ (x y) :translate /)
1301 (define-static-fun two-arg-gcd (x y) :translate gcd)
1302 (define-static-fun two-arg-lcm (x y) :translate lcm)
1304 (define-static-fun two-arg-and (x y) :translate logand)
1305 (define-static-fun two-arg-ior (x y) :translate logior)
1306 (define-static-fun two-arg-xor (x y) :translate logxor)
1309 ;;; Support for the Mersenne Twister, MT19937, random number generator
1310 ;;; due to Matsumoto and Nishimura.
1312 ;;; Makoto Matsumoto and T. Nishimura, "Mersenne twister: A
1313 ;;; 623-dimensionally equidistributed uniform pseudorandom number
1314 ;;; generator.", ACM Transactions on Modeling and Computer Simulation,
1315 ;;; 1997, to appear.
1318 ;;; 0-1: Constant matrix A. [0, #x9908b0df] (not used here)
1319 ;;; 2: Index; init. to 1.
1321 (defknown random-mt19937 ((simple-array (unsigned-byte 32) (*)))
1322 (unsigned-byte 32) ())
1323 (define-vop (random-mt19937)
1324 (:policy :fast-safe)
1325 (:translate random-mt19937)
1326 (:args (state :scs (descriptor-reg) :to :result))
1327 (:arg-types simple-array-unsigned-byte-32)
1328 (:temporary (:sc unsigned-reg :from (:eval 0) :to :result) k)
1329 (:temporary (:sc unsigned-reg :offset eax-offset
1330 :from (:eval 0) :to :result) tmp)
1331 (:results (y :scs (unsigned-reg) :from (:eval 0)))
1332 (:result-types unsigned-num)
1334 (inst mov k (make-ea :dword :base state
1335 :disp (- (* (+ 2 vector-data-offset)
1337 other-pointer-lowtag)))
1339 (inst jmp :ne no-update)
1340 (inst mov tmp state) ; The state is passed in EAX.
1341 (inst call (make-fixup 'random-mt19937-update :assembly-routine))
1342 ;; Restore k, and set to 0.
1346 (inst mov y (make-ea :dword :base state :index k :scale 4
1347 :disp (- (* (+ 3 vector-data-offset)
1349 other-pointer-lowtag)))
1352 (inst xor y (make-ea :dword :base state :index k :scale 4
1353 :disp (- (* (+ 3 vector-data-offset)
1355 other-pointer-lowtag)))
1356 ;; y ^= (y << 7) & #x9d2c5680
1360 (inst mov (make-ea :dword :base state
1361 :disp (- (* (+ 2 vector-data-offset)
1363 other-pointer-lowtag))
1365 (inst and tmp #x9d2c5680)
1367 ;; y ^= (y << 15) & #xefc60000
1370 (inst and tmp #xefc60000)
1379 (defknown %lea ((or (signed-byte 32) (unsigned-byte 32))
1380 (or (signed-byte 32) (unsigned-byte 32))
1381 (member 1 2 4 8) (signed-byte 32))
1382 (or (signed-byte 32) (unsigned-byte 32))
1383 (foldable flushable))
1385 (defoptimizer (%lea derive-type) ((base index scale disp))
1386 (when (and (constant-continuation-p scale)
1387 (constant-continuation-p disp))
1388 (let ((scale (continuation-value scale))
1389 (disp (continuation-value disp))
1390 (base-type (continuation-type base))
1391 (index-type (continuation-type index)))
1392 (when (and (numeric-type-p base-type)
1393 (numeric-type-p index-type))
1394 (let ((base-lo (numeric-type-low base-type))
1395 (base-hi (numeric-type-high base-type))
1396 (index-lo (numeric-type-low index-type))
1397 (index-hi (numeric-type-high index-type)))
1398 (make-numeric-type :class 'integer
1400 :low (when (and base-lo index-lo)
1401 (+ base-lo (* index-lo scale) disp))
1402 :high (when (and base-hi index-hi)
1403 (+ base-hi (* index-hi scale) disp))))))))
1405 (defun %lea (base index scale disp)
1406 (+ base (* index scale) disp))
1408 (in-package "SB!VM")
1410 (define-vop (%lea/unsigned=>unsigned)
1412 (:policy :fast-safe)
1413 (:args (base :scs (unsigned-reg))
1414 (index :scs (unsigned-reg)))
1416 (:arg-types unsigned-num unsigned-num
1417 (:constant (member 1 2 4 8))
1418 (:constant (signed-byte 32)))
1419 (:results (r :scs (unsigned-reg)))
1420 (:result-types unsigned-num)
1422 (inst lea r (make-ea :dword :base base :index index
1423 :scale scale :disp disp))))
1425 (define-vop (%lea/signed=>signed)
1427 (:policy :fast-safe)
1428 (:args (base :scs (signed-reg))
1429 (index :scs (signed-reg)))
1431 (:arg-types signed-num signed-num
1432 (:constant (member 1 2 4 8))
1433 (:constant (signed-byte 32)))
1434 (:results (r :scs (signed-reg)))
1435 (:result-types signed-num)
1437 (inst lea r (make-ea :dword :base base :index index
1438 :scale scale :disp disp))))
1440 (define-vop (%lea/fixnum=>fixnum)
1442 (:policy :fast-safe)
1443 (:args (base :scs (any-reg))
1444 (index :scs (any-reg)))
1446 (:arg-types tagged-num tagged-num
1447 (:constant (member 1 2 4 8))
1448 (:constant (signed-byte 32)))
1449 (:results (r :scs (any-reg)))
1450 (:result-types tagged-num)
1452 (inst lea r (make-ea :dword :base base :index index
1453 :scale scale :disp disp))))
1457 ;;; This is essentially a straight implementation of the algorithm in
1458 ;;; "Strength Reduction of Multiplications by Integer Constants",
1459 ;;; Youfeng Wu, ACM SIGPLAN Notices, Vol. 30, No.2, February 1995.
1460 (defun basic-decompose-multiplication (arg num n-bits condensed)
1461 (case (aref condensed 0)
1463 (let ((tmp (min 3 (aref condensed 1))))
1464 (decf (aref condensed 1) tmp)
1465 `(truly-the (unsigned-byte 32)
1467 ,(decompose-multiplication
1468 arg (ash (1- num) (- tmp)) (1- n-bits) (subseq condensed 1))
1471 (let ((r0 (aref condensed 0)))
1472 (incf (aref condensed 1) r0)
1473 `(truly-the (unsigned-byte 32)
1474 (%lea ,(decompose-multiplication
1475 arg (- num (ash 1 r0)) (1- n-bits) (subseq condensed 1))
1478 (t (let ((r0 (aref condensed 0)))
1479 (setf (aref condensed 0) 0)
1480 `(truly-the (unsigned-byte 32)
1481 (ash ,(decompose-multiplication
1482 arg (ash num (- r0)) n-bits condensed)
1485 (defun decompose-multiplication (arg num n-bits condensed)
1490 `(truly-the (unsigned-byte 32) (ash ,arg ,(1- (integer-length num)))))
1491 ((let ((max 0) (end 0))
1492 (loop for i from 2 to (length condensed)
1493 for j = (reduce #'+ (subseq condensed 0 i))
1494 when (and (> (- (* 2 i) 3 j) max)
1495 (< (+ (ash 1 (1+ j))
1496 (ash (ldb (byte (- 32 (1+ j)) (1+ j)) num)
1499 do (setq max (- (* 2 i) 3 j)
1502 (let ((j (reduce #'+ (subseq condensed 0 end))))
1503 (let ((n2 (+ (ash 1 (1+ j))
1504 (ash (ldb (byte (- 32 (1+ j)) (1+ j)) num) (1+ j))))
1505 (n1 (1+ (ldb (byte (1+ j) 0) (lognot num)))))
1506 `(truly-the (unsigned-byte 32)
1507 (- ,(optimize-multiply arg n2) ,(optimize-multiply arg n1))))))))
1508 ((dolist (i '(9 5 3))
1509 (when (integerp (/ num i))
1510 (when (< (logcount (/ num i)) (logcount num))
1512 (return `(let ((,x ,(optimize-multiply arg (/ num i))))
1513 (truly-the (unsigned-byte 32)
1514 (%lea ,x ,x (1- ,i) 0)))))))))
1515 (t (basic-decompose-multiplication arg num n-bits condensed))))
1517 (defun optimize-multiply (arg x)
1518 (let* ((n-bits (logcount x))
1519 (condensed (make-array n-bits)))
1520 (let ((count 0) (bit 0))
1522 (cond ((logbitp i x)
1523 (setf (aref condensed bit) count)
1527 (decompose-multiplication arg x n-bits condensed)))
1529 (deftransform * ((x y)
1530 ((unsigned-byte 32) (constant-arg (unsigned-byte 32)))
1532 "recode as leas, shifts and adds"
1533 (let ((y (continuation-value y)))
1535 ((= y (ash 1 (integer-length y)))
1536 ;; there's a generic transform for y = 2^k
1537 (give-up-ir1-transform))
1538 ((member y '(3 5 9))
1539 ;; we can do these multiplications directly using LEA
1540 `(%lea x x ,(1- y) 0))
1541 ((member :pentium4 *backend-subfeatures*)
1542 ;; the pentium4's multiply unit is reportedly very good
1543 (give-up-ir1-transform))
1544 ;; FIXME: should make this more fine-grained. If nothing else,
1545 ;; there should probably be a cutoff of about 9 instructions on
1546 ;; pentium-class machines.
1547 (t (optimize-multiply 'x y)))))
1549 ;;; FIXME: we should also be able to write an optimizer or two to
1550 ;;; convert (+ (* x 2) 17), (- (* x 9) 5) to a %LEA.