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/signed=>signed)
639 (:args (number :scs (signed-reg) :target result
640 :load-if (not (and (sc-is number signed-stack)
641 (sc-is result signed-stack)
642 (location= number result)))))
644 (:arg-types signed-num (:constant integer))
645 (:results (result :scs (signed-reg)
646 :load-if (not (and (sc-is number signed-stack)
647 (sc-is result signed-stack)
648 (location= number result)))))
649 (:result-types signed-num)
652 (cond ((and (= amount 1) (not (location= number result)))
653 (inst lea result (make-ea :dword :index number :scale 2)))
654 ((and (= amount 2) (not (location= number result)))
655 (inst lea result (make-ea :dword :index number :scale 4)))
656 ((and (= amount 3) (not (location= number result)))
657 (inst lea result (make-ea :dword :index number :scale 8)))
660 (cond ((plusp amount) (inst shl result amount))
661 (t (inst sar result (min 31 (- amount)))))))))
663 (define-vop (fast-ash-c/unsigned=>unsigned)
666 (:args (number :scs (unsigned-reg) :target result
667 :load-if (not (and (sc-is number unsigned-stack)
668 (sc-is result unsigned-stack)
669 (location= number result)))))
671 (:arg-types unsigned-num (:constant integer))
672 (:results (result :scs (unsigned-reg)
673 :load-if (not (and (sc-is number unsigned-stack)
674 (sc-is result unsigned-stack)
675 (location= number result)))))
676 (:result-types unsigned-num)
679 (cond ((and (= amount 1) (not (location= number result)))
680 (inst lea result (make-ea :dword :index number :scale 2)))
681 ((and (= amount 2) (not (location= number result)))
682 (inst lea result (make-ea :dword :index number :scale 4)))
683 ((and (= amount 3) (not (location= number result)))
684 (inst lea result (make-ea :dword :index number :scale 8)))
687 (cond ((plusp amount) (inst shl result amount))
688 ((< amount -31) (inst xor result result))
689 (t (inst shr result (- amount))))))))
691 (define-vop (fast-ash-left/signed)
693 (:args (number :scs (signed-reg) :target result
694 :load-if (not (and (sc-is number signed-stack)
695 (sc-is result signed-stack)
696 (location= number result))))
697 (amount :scs (unsigned-reg) :target ecx))
698 (:arg-types signed-num positive-fixnum)
699 (:temporary (:sc unsigned-reg :offset ecx-offset :from (:argument 1)) ecx)
700 (:results (result :scs (signed-reg) :from (:argument 0)
701 :load-if (not (and (sc-is number signed-stack)
702 (sc-is result signed-stack)
703 (location= number result)))))
704 (:result-types signed-num)
710 (inst shl result :cl)))
712 (define-vop (fast-ash-left/unsigned)
714 (:args (number :scs (unsigned-reg) :target result
715 :load-if (not (and (sc-is number unsigned-stack)
716 (sc-is result unsigned-stack)
717 (location= number result))))
718 (amount :scs (unsigned-reg) :target ecx))
719 (:arg-types unsigned-num positive-fixnum)
720 (:temporary (:sc unsigned-reg :offset ecx-offset :from (:argument 1)) ecx)
721 (:results (result :scs (unsigned-reg) :from (:argument 0)
722 :load-if (not (and (sc-is number unsigned-stack)
723 (sc-is result unsigned-stack)
724 (location= number result)))))
725 (:result-types unsigned-num)
731 (inst shl result :cl)))
733 (define-vop (fast-ash/signed=>signed)
736 (:args (number :scs (signed-reg) :target result)
737 (amount :scs (signed-reg) :target ecx))
738 (:arg-types signed-num signed-num)
739 (:results (result :scs (signed-reg) :from (:argument 0)))
740 (:result-types signed-num)
741 (:temporary (:sc signed-reg :offset ecx-offset :from (:argument 1)) ecx)
747 (inst jmp :ns positive)
753 (inst sar result :cl)
757 ;; The result-type ensures us that this shift will not overflow.
758 (inst shl result :cl)
762 (define-vop (fast-ash/unsigned=>unsigned)
765 (:args (number :scs (unsigned-reg) :target result)
766 (amount :scs (signed-reg) :target ecx))
767 (:arg-types unsigned-num signed-num)
768 (:results (result :scs (unsigned-reg) :from (:argument 0)))
769 (:result-types unsigned-num)
770 (:temporary (:sc signed-reg :offset ecx-offset :from (:argument 1)) ecx)
776 (inst jmp :ns positive)
780 (inst xor result result)
783 (inst shr result :cl)
787 ;; The result-type ensures us that this shift will not overflow.
788 (inst shl result :cl)
792 ;;; FIXME: before making knowledge of this too public, it needs to be
793 ;;; fixed so that it's actually _faster_ than the non-CMOV version; at
794 ;;; least on my Celeron-XXX laptop, this version is marginally slower
795 ;;; than the above version with branches. -- CSR, 2003-09-04
796 (define-vop (fast-cmov-ash/unsigned=>unsigned)
799 (:args (number :scs (unsigned-reg) :target result)
800 (amount :scs (signed-reg) :target ecx))
801 (:arg-types unsigned-num signed-num)
802 (:results (result :scs (unsigned-reg) :from (:argument 0)))
803 (:result-types unsigned-num)
804 (:temporary (:sc signed-reg :offset ecx-offset :from (:argument 1)) ecx)
805 (:temporary (:sc any-reg :from (:eval 0) :to (:eval 1)) zero)
807 (:guard (member :cmov *backend-subfeatures*))
812 (inst jmp :ns positive)
815 (inst shr result :cl)
817 (inst cmov :nbe result zero)
821 ;; The result-type ensures us that this shift will not overflow.
822 (inst shl result :cl)
826 ;;; Note: documentation for this function is wrong - rtfm
827 (define-vop (signed-byte-32-len)
828 (:translate integer-length)
829 (:note "inline (signed-byte 32) integer-length")
831 (:args (arg :scs (signed-reg) :target res))
832 (:arg-types signed-num)
833 (:results (res :scs (unsigned-reg)))
834 (:result-types unsigned-num)
849 (define-vop (unsigned-byte-32-len)
850 (:translate integer-length)
851 (:note "inline (unsigned-byte 32) integer-length")
853 (:args (arg :scs (unsigned-reg)))
854 (:arg-types unsigned-num)
855 (:results (res :scs (unsigned-reg)))
856 (:result-types unsigned-num)
866 (define-vop (unsigned-byte-32-count)
867 (:translate logcount)
868 (:note "inline (unsigned-byte 32) logcount")
870 (:args (arg :scs (unsigned-reg)))
871 (:arg-types unsigned-num)
872 (:results (result :scs (unsigned-reg)))
873 (:result-types positive-fixnum)
874 (:temporary (:sc unsigned-reg :from (:argument 0)) temp)
878 (inst mov temp result)
880 (inst and result #x55555555)
881 (inst and temp #x55555555)
882 (inst add result temp)
884 (inst mov temp result)
886 (inst and result #x33333333)
887 (inst and temp #x33333333)
888 (inst add result temp)
890 (inst mov temp result)
892 (inst and result #x0f0f0f0f)
893 (inst and temp #x0f0f0f0f)
894 (inst add result temp)
896 (inst mov temp result)
898 (inst and result #x00ff00ff)
899 (inst and temp #x00ff00ff)
900 (inst add result temp)
902 (inst mov temp result)
904 (inst and result #x0000ffff)
905 (inst and temp #x0000ffff)
906 (inst add result temp)))
908 ;;;; binary conditional VOPs
910 (define-vop (fast-conditional)
915 (:policy :fast-safe))
917 (define-vop (fast-conditional/fixnum fast-conditional)
918 (:args (x :scs (any-reg)
919 :load-if (not (and (sc-is x control-stack)
921 (y :scs (any-reg control-stack)))
922 (:arg-types tagged-num tagged-num)
923 (:note "inline fixnum comparison"))
925 (define-vop (fast-conditional-c/fixnum fast-conditional/fixnum)
926 (:args (x :scs (any-reg control-stack)))
927 (:arg-types tagged-num (:constant (signed-byte 30)))
928 (:info target not-p y))
930 (define-vop (fast-conditional/signed fast-conditional)
931 (:args (x :scs (signed-reg)
932 :load-if (not (and (sc-is x signed-stack)
933 (sc-is y signed-reg))))
934 (y :scs (signed-reg signed-stack)))
935 (:arg-types signed-num signed-num)
936 (:note "inline (signed-byte 32) comparison"))
938 (define-vop (fast-conditional-c/signed fast-conditional/signed)
939 (:args (x :scs (signed-reg signed-stack)))
940 (:arg-types signed-num (:constant (signed-byte 32)))
941 (:info target not-p y))
943 (define-vop (fast-conditional/unsigned fast-conditional)
944 (:args (x :scs (unsigned-reg)
945 :load-if (not (and (sc-is x unsigned-stack)
946 (sc-is y unsigned-reg))))
947 (y :scs (unsigned-reg unsigned-stack)))
948 (:arg-types unsigned-num unsigned-num)
949 (:note "inline (unsigned-byte 32) comparison"))
951 (define-vop (fast-conditional-c/unsigned fast-conditional/unsigned)
952 (:args (x :scs (unsigned-reg unsigned-stack)))
953 (:arg-types unsigned-num (:constant (unsigned-byte 32)))
954 (:info target not-p y))
957 (macrolet ((define-conditional-vop (tran cond unsigned not-cond not-unsigned)
960 (lambda (suffix cost signed)
961 `(define-vop (;; FIXME: These could be done more
962 ;; cleanly with SYMBOLICATE.
963 ,(intern (format nil "~:@(FAST-IF-~A~A~)"
966 (format nil "~:@(FAST-CONDITIONAL~A~)"
971 ,(if (eq suffix '-c/fixnum)
982 '(/fixnum -c/fixnum /signed -c/signed /unsigned -c/unsigned)
984 '(t t t t nil nil)))))
986 (define-conditional-vop < :l :b :ge :ae)
987 (define-conditional-vop > :g :a :le :be))
989 (define-vop (fast-if-eql/signed fast-conditional/signed)
993 (inst jmp (if not-p :ne :e) target)))
995 (define-vop (fast-if-eql-c/signed fast-conditional-c/signed)
998 (cond ((and (sc-is x signed-reg) (zerop y))
999 (inst test x x)) ; smaller instruction
1002 (inst jmp (if not-p :ne :e) target)))
1004 (define-vop (fast-if-eql/unsigned fast-conditional/unsigned)
1008 (inst jmp (if not-p :ne :e) target)))
1010 (define-vop (fast-if-eql-c/unsigned fast-conditional-c/unsigned)
1013 (cond ((and (sc-is x unsigned-reg) (zerop y))
1014 (inst test x x)) ; smaller instruction
1017 (inst jmp (if not-p :ne :e) target)))
1019 ;;; EQL/FIXNUM is funny because the first arg can be of any type, not just a
1022 ;;; These versions specify a fixnum restriction on their first arg. We have
1023 ;;; also generic-eql/fixnum VOPs which are the same, but have no restriction on
1024 ;;; the first arg and a higher cost. The reason for doing this is to prevent
1025 ;;; fixnum specific operations from being used on word integers, spuriously
1026 ;;; consing the argument.
1028 (define-vop (fast-eql/fixnum fast-conditional)
1029 (:args (x :scs (any-reg)
1030 :load-if (not (and (sc-is x control-stack)
1031 (sc-is y any-reg))))
1032 (y :scs (any-reg control-stack)))
1033 (:arg-types tagged-num tagged-num)
1034 (:note "inline fixnum comparison")
1038 (inst jmp (if not-p :ne :e) target)))
1039 (define-vop (generic-eql/fixnum fast-eql/fixnum)
1040 (:args (x :scs (any-reg descriptor-reg)
1041 :load-if (not (and (sc-is x control-stack)
1042 (sc-is y any-reg))))
1043 (y :scs (any-reg control-stack)))
1044 (:arg-types * tagged-num)
1047 (define-vop (fast-eql-c/fixnum fast-conditional/fixnum)
1048 (:args (x :scs (any-reg control-stack)))
1049 (:arg-types tagged-num (:constant (signed-byte 30)))
1050 (:info target not-p y)
1053 (cond ((and (sc-is x any-reg) (zerop y))
1054 (inst test x x)) ; smaller instruction
1056 (inst cmp x (fixnumize y))))
1057 (inst jmp (if not-p :ne :e) target)))
1058 (define-vop (generic-eql-c/fixnum fast-eql-c/fixnum)
1059 (:args (x :scs (any-reg descriptor-reg control-stack)))
1060 (:arg-types * (:constant (signed-byte 30)))
1063 ;;;; 32-bit logical operations
1065 (define-vop (merge-bits)
1066 (:translate merge-bits)
1067 (:args (shift :scs (signed-reg unsigned-reg) :target ecx)
1068 (prev :scs (unsigned-reg) :target result)
1069 (next :scs (unsigned-reg)))
1070 (:arg-types tagged-num unsigned-num unsigned-num)
1071 (:temporary (:sc signed-reg :offset ecx-offset :from (:argument 0)) ecx)
1072 (:results (result :scs (unsigned-reg) :from (:argument 1)))
1073 (:result-types unsigned-num)
1074 (:policy :fast-safe)
1078 (inst shrd result next :cl)))
1080 (define-source-transform 32bit-logical-not (x)
1081 `(logand (lognot (the (unsigned-byte 32) ,x)) #.(1- (ash 1 32))))
1083 (deftransform 32bit-logical-and ((x y))
1086 (define-source-transform 32bit-logical-nand (x y)
1087 `(32bit-logical-not (32bit-logical-and ,x ,y)))
1089 (deftransform 32bit-logical-or ((x y))
1092 (define-source-transform 32bit-logical-nor (x y)
1093 `(32bit-logical-not (32bit-logical-or ,x ,y)))
1095 (deftransform 32bit-logical-xor ((x y))
1098 (define-source-transform 32bit-logical-eqv (x y)
1099 `(32bit-logical-not (32bit-logical-xor ,x ,y)))
1101 (define-source-transform 32bit-logical-orc1 (x y)
1102 `(32bit-logical-or (32bit-logical-not ,x) ,y))
1104 (define-source-transform 32bit-logical-orc2 (x y)
1105 `(32bit-logical-or ,x (32bit-logical-not ,y)))
1107 (define-source-transform 32bit-logical-andc1 (x y)
1108 `(32bit-logical-and (32bit-logical-not ,x) ,y))
1110 (define-source-transform 32bit-logical-andc2 (x y)
1111 `(32bit-logical-and ,x (32bit-logical-not ,y)))
1113 ;;; Only the lower 5 bits of the shift amount are significant.
1114 (define-vop (shift-towards-someplace)
1115 (:policy :fast-safe)
1116 (:args (num :scs (unsigned-reg) :target r)
1117 (amount :scs (signed-reg) :target ecx))
1118 (:arg-types unsigned-num tagged-num)
1119 (:temporary (:sc signed-reg :offset ecx-offset :from (:argument 1)) ecx)
1120 (:results (r :scs (unsigned-reg) :from (:argument 0)))
1121 (:result-types unsigned-num))
1123 (define-vop (shift-towards-start shift-towards-someplace)
1124 (:translate shift-towards-start)
1125 (:note "SHIFT-TOWARDS-START")
1131 (define-vop (shift-towards-end shift-towards-someplace)
1132 (:translate shift-towards-end)
1133 (:note "SHIFT-TOWARDS-END")
1139 ;;;; Modular functions
1141 (define-modular-fun +-mod32 (x y) + 32)
1142 (define-vop (fast-+-mod32/unsigned=>unsigned fast-+/unsigned=>unsigned)
1143 (:translate +-mod32))
1144 (define-vop (fast-+-mod32-c/unsigned=>unsigned fast-+-c/unsigned=>unsigned)
1145 (:translate +-mod32))
1147 ;;; logical operations
1148 (define-modular-fun lognot-mod32 (x) lognot 32)
1149 (define-vop (lognot-mod32/unsigned=>unsigned)
1150 (:translate lognot-mod32)
1151 (:args (x :scs (unsigned-reg unsigned-stack) :target r
1152 :load-if (not (and (sc-is x unsigned-stack)
1153 (sc-is r unsigned-stack)
1155 (:arg-types unsigned-num)
1156 (:results (r :scs (unsigned-reg)
1157 :load-if (not (and (sc-is x unsigned-stack)
1158 (sc-is r unsigned-stack)
1160 (:result-types unsigned-num)
1161 (:policy :fast-safe)
1166 (define-modular-fun logxor-mod32 (x y) logxor 32)
1167 (define-vop (fast-logxor-mod32/unsigned=>unsigned
1168 fast-logxor/unsigned=>unsigned)
1169 (:translate logxor-mod32))
1170 (define-vop (fast-logxor-mod32-c/unsigned=>unsigned
1171 fast-logxor-c/unsigned=>unsigned)
1172 (:translate logxor-mod32))
1176 (define-vop (bignum-length get-header-data)
1177 (:translate sb!bignum::%bignum-length)
1178 (:policy :fast-safe))
1180 (define-vop (bignum-set-length set-header-data)
1181 (:translate sb!bignum::%bignum-set-length)
1182 (:policy :fast-safe))
1184 (define-full-reffer bignum-ref * bignum-digits-offset other-pointer-lowtag
1185 (unsigned-reg) unsigned-num sb!bignum::%bignum-ref)
1187 (define-full-setter bignum-set * bignum-digits-offset other-pointer-lowtag
1188 (unsigned-reg) unsigned-num sb!bignum::%bignum-set)
1190 (define-vop (digit-0-or-plus)
1191 (:translate sb!bignum::%digit-0-or-plusp)
1192 (:policy :fast-safe)
1193 (:args (digit :scs (unsigned-reg)))
1194 (:arg-types unsigned-num)
1196 (:info target not-p)
1198 (inst or digit digit)
1199 (inst jmp (if not-p :s :ns) target)))
1202 ;;; For add and sub with carry the sc of carry argument is any-reg so
1203 ;;; the it may be passed as a fixnum or word and thus may be 0, 1, or
1204 ;;; 4. This is easy to deal with and may save a fixnum-word
1206 (define-vop (add-w/carry)
1207 (:translate sb!bignum::%add-with-carry)
1208 (:policy :fast-safe)
1209 (:args (a :scs (unsigned-reg) :target result)
1210 (b :scs (unsigned-reg unsigned-stack) :to :eval)
1211 (c :scs (any-reg) :target temp))
1212 (:arg-types unsigned-num unsigned-num positive-fixnum)
1213 (:temporary (:sc any-reg :from (:argument 2) :to :eval) temp)
1214 (:results (result :scs (unsigned-reg) :from (:argument 0))
1215 (carry :scs (unsigned-reg)))
1216 (:result-types unsigned-num positive-fixnum)
1220 (inst neg temp) ; Set the carry flag to 0 if c=0 else to 1
1223 (inst adc carry carry)))
1225 ;;; Note: the borrow is the oppostite of the x86 convention - 1 for no
1226 ;;; borrow and 0 for a borrow.
1227 (define-vop (sub-w/borrow)
1228 (:translate sb!bignum::%subtract-with-borrow)
1229 (:policy :fast-safe)
1230 (:args (a :scs (unsigned-reg) :to :eval :target result)
1231 (b :scs (unsigned-reg unsigned-stack) :to :result)
1232 (c :scs (any-reg control-stack)))
1233 (:arg-types unsigned-num unsigned-num positive-fixnum)
1234 (:results (result :scs (unsigned-reg) :from :eval)
1235 (borrow :scs (unsigned-reg)))
1236 (:result-types unsigned-num positive-fixnum)
1238 (inst cmp c 1) ; Set the carry flag to 1 if c=0 else to 0
1242 (inst adc borrow borrow)
1243 (inst xor borrow 1)))
1246 (define-vop (bignum-mult-and-add-3-arg)
1247 (:translate sb!bignum::%multiply-and-add)
1248 (:policy :fast-safe)
1249 (:args (x :scs (unsigned-reg) :target eax)
1250 (y :scs (unsigned-reg unsigned-stack))
1251 (carry-in :scs (unsigned-reg unsigned-stack)))
1252 (:arg-types unsigned-num unsigned-num unsigned-num)
1253 (:temporary (:sc unsigned-reg :offset eax-offset :from (:argument 0)
1254 :to (:result 1) :target lo) eax)
1255 (:temporary (:sc unsigned-reg :offset edx-offset :from (:argument 1)
1256 :to (:result 0) :target hi) edx)
1257 (:results (hi :scs (unsigned-reg))
1258 (lo :scs (unsigned-reg)))
1259 (:result-types unsigned-num unsigned-num)
1263 (inst add eax carry-in)
1268 (define-vop (bignum-mult-and-add-4-arg)
1269 (:translate sb!bignum::%multiply-and-add)
1270 (:policy :fast-safe)
1271 (:args (x :scs (unsigned-reg) :target eax)
1272 (y :scs (unsigned-reg unsigned-stack))
1273 (prev :scs (unsigned-reg unsigned-stack))
1274 (carry-in :scs (unsigned-reg unsigned-stack)))
1275 (:arg-types unsigned-num unsigned-num unsigned-num unsigned-num)
1276 (:temporary (:sc unsigned-reg :offset eax-offset :from (:argument 0)
1277 :to (:result 1) :target lo) eax)
1278 (:temporary (:sc unsigned-reg :offset edx-offset :from (:argument 1)
1279 :to (:result 0) :target hi) edx)
1280 (:results (hi :scs (unsigned-reg))
1281 (lo :scs (unsigned-reg)))
1282 (:result-types unsigned-num unsigned-num)
1288 (inst add eax carry-in)
1294 (define-vop (bignum-mult)
1295 (:translate sb!bignum::%multiply)
1296 (:policy :fast-safe)
1297 (:args (x :scs (unsigned-reg) :target eax)
1298 (y :scs (unsigned-reg unsigned-stack)))
1299 (:arg-types unsigned-num unsigned-num)
1300 (:temporary (:sc unsigned-reg :offset eax-offset :from (:argument 0)
1301 :to (:result 1) :target lo) eax)
1302 (:temporary (:sc unsigned-reg :offset edx-offset :from (:argument 1)
1303 :to (:result 0) :target hi) edx)
1304 (:results (hi :scs (unsigned-reg))
1305 (lo :scs (unsigned-reg)))
1306 (:result-types unsigned-num unsigned-num)
1313 (define-vop (bignum-lognot lognot-mod32/unsigned=>unsigned)
1314 (:translate sb!bignum::%lognot))
1316 (define-vop (fixnum-to-digit)
1317 (:translate sb!bignum::%fixnum-to-digit)
1318 (:policy :fast-safe)
1319 (:args (fixnum :scs (any-reg control-stack) :target digit))
1320 (:arg-types tagged-num)
1321 (:results (digit :scs (unsigned-reg)
1322 :load-if (not (and (sc-is fixnum control-stack)
1323 (sc-is digit unsigned-stack)
1324 (location= fixnum digit)))))
1325 (:result-types unsigned-num)
1328 (inst sar digit 2)))
1330 (define-vop (bignum-floor)
1331 (:translate sb!bignum::%floor)
1332 (:policy :fast-safe)
1333 (:args (div-high :scs (unsigned-reg) :target edx)
1334 (div-low :scs (unsigned-reg) :target eax)
1335 (divisor :scs (unsigned-reg unsigned-stack)))
1336 (:arg-types unsigned-num unsigned-num unsigned-num)
1337 (:temporary (:sc unsigned-reg :offset eax-offset :from (:argument 1)
1338 :to (:result 0) :target quo) eax)
1339 (:temporary (:sc unsigned-reg :offset edx-offset :from (:argument 0)
1340 :to (:result 1) :target rem) edx)
1341 (:results (quo :scs (unsigned-reg))
1342 (rem :scs (unsigned-reg)))
1343 (:result-types unsigned-num unsigned-num)
1347 (inst div eax divisor)
1351 (define-vop (signify-digit)
1352 (:translate sb!bignum::%fixnum-digit-with-correct-sign)
1353 (:policy :fast-safe)
1354 (:args (digit :scs (unsigned-reg unsigned-stack) :target res))
1355 (:arg-types unsigned-num)
1356 (:results (res :scs (any-reg signed-reg)
1357 :load-if (not (and (sc-is digit unsigned-stack)
1358 (sc-is res control-stack signed-stack)
1359 (location= digit res)))))
1360 (:result-types signed-num)
1363 (when (sc-is res any-reg control-stack)
1366 (define-vop (digit-ashr)
1367 (:translate sb!bignum::%ashr)
1368 (:policy :fast-safe)
1369 (:args (digit :scs (unsigned-reg unsigned-stack) :target result)
1370 (count :scs (unsigned-reg) :target ecx))
1371 (:arg-types unsigned-num positive-fixnum)
1372 (:temporary (:sc unsigned-reg :offset ecx-offset :from (:argument 1)) ecx)
1373 (:results (result :scs (unsigned-reg) :from (:argument 0)
1374 :load-if (not (and (sc-is result unsigned-stack)
1375 (location= digit result)))))
1376 (:result-types unsigned-num)
1380 (inst sar result :cl)))
1382 (define-vop (digit-lshr digit-ashr)
1383 (:translate sb!bignum::%digit-logical-shift-right)
1387 (inst shr result :cl)))
1389 (define-vop (digit-ashl digit-ashr)
1390 (:translate sb!bignum::%ashl)
1394 (inst shl result :cl)))
1396 ;;;; static functions
1398 (define-static-fun two-arg-/ (x y) :translate /)
1400 (define-static-fun two-arg-gcd (x y) :translate gcd)
1401 (define-static-fun two-arg-lcm (x y) :translate lcm)
1403 (define-static-fun two-arg-and (x y) :translate logand)
1404 (define-static-fun two-arg-ior (x y) :translate logior)
1405 (define-static-fun two-arg-xor (x y) :translate logxor)
1408 ;;; Support for the Mersenne Twister, MT19937, random number generator
1409 ;;; due to Matsumoto and Nishimura.
1411 ;;; Makoto Matsumoto and T. Nishimura, "Mersenne twister: A
1412 ;;; 623-dimensionally equidistributed uniform pseudorandom number
1413 ;;; generator.", ACM Transactions on Modeling and Computer Simulation,
1414 ;;; 1997, to appear.
1417 ;;; 0-1: Constant matrix A. [0, #x9908b0df] (not used here)
1418 ;;; 2: Index; init. to 1.
1420 (defknown random-mt19937 ((simple-array (unsigned-byte 32) (*)))
1421 (unsigned-byte 32) ())
1422 (define-vop (random-mt19937)
1423 (:policy :fast-safe)
1424 (:translate random-mt19937)
1425 (:args (state :scs (descriptor-reg) :to :result))
1426 (:arg-types simple-array-unsigned-byte-32)
1427 (:temporary (:sc unsigned-reg :from (:eval 0) :to :result) k)
1428 (:temporary (:sc unsigned-reg :offset eax-offset
1429 :from (:eval 0) :to :result) tmp)
1430 (:results (y :scs (unsigned-reg) :from (:eval 0)))
1431 (:result-types unsigned-num)
1433 (inst mov k (make-ea :dword :base state
1434 :disp (- (* (+ 2 vector-data-offset)
1436 other-pointer-lowtag)))
1438 (inst jmp :ne no-update)
1439 (inst mov tmp state) ; The state is passed in EAX.
1440 (inst call (make-fixup 'random-mt19937-update :assembly-routine))
1441 ;; Restore k, and set to 0.
1445 (inst mov y (make-ea :dword :base state :index k :scale 4
1446 :disp (- (* (+ 3 vector-data-offset)
1448 other-pointer-lowtag)))
1451 (inst xor y (make-ea :dword :base state :index k :scale 4
1452 :disp (- (* (+ 3 vector-data-offset)
1454 other-pointer-lowtag)))
1455 ;; y ^= (y << 7) & #x9d2c5680
1459 (inst mov (make-ea :dword :base state
1460 :disp (- (* (+ 2 vector-data-offset)
1462 other-pointer-lowtag))
1464 (inst and tmp #x9d2c5680)
1466 ;; y ^= (y << 15) & #xefc60000
1469 (inst and tmp #xefc60000)
1478 (defknown %lea ((or (signed-byte 32) (unsigned-byte 32))
1479 (or (signed-byte 32) (unsigned-byte 32))
1480 (member 1 2 4 8) (signed-byte 32))
1481 (or (signed-byte 32) (unsigned-byte 32))
1482 (foldable flushable))
1484 (defoptimizer (%lea derive-type) ((base index scale disp))
1485 (when (and (constant-lvar-p scale)
1486 (constant-lvar-p disp))
1487 (let ((scale (lvar-value scale))
1488 (disp (lvar-value disp))
1489 (base-type (lvar-type base))
1490 (index-type (lvar-type index)))
1491 (when (and (numeric-type-p base-type)
1492 (numeric-type-p index-type))
1493 (let ((base-lo (numeric-type-low base-type))
1494 (base-hi (numeric-type-high base-type))
1495 (index-lo (numeric-type-low index-type))
1496 (index-hi (numeric-type-high index-type)))
1497 (make-numeric-type :class 'integer
1499 :low (when (and base-lo index-lo)
1500 (+ base-lo (* index-lo scale) disp))
1501 :high (when (and base-hi index-hi)
1502 (+ base-hi (* index-hi scale) disp))))))))
1504 (defun %lea (base index scale disp)
1505 (+ base (* index scale) disp))
1507 (in-package "SB!VM")
1509 (define-vop (%lea/unsigned=>unsigned)
1511 (:policy :fast-safe)
1512 (:args (base :scs (unsigned-reg))
1513 (index :scs (unsigned-reg)))
1515 (:arg-types unsigned-num unsigned-num
1516 (:constant (member 1 2 4 8))
1517 (:constant (signed-byte 32)))
1518 (:results (r :scs (unsigned-reg)))
1519 (:result-types unsigned-num)
1521 (inst lea r (make-ea :dword :base base :index index
1522 :scale scale :disp disp))))
1524 (define-vop (%lea/signed=>signed)
1526 (:policy :fast-safe)
1527 (:args (base :scs (signed-reg))
1528 (index :scs (signed-reg)))
1530 (:arg-types signed-num signed-num
1531 (:constant (member 1 2 4 8))
1532 (:constant (signed-byte 32)))
1533 (:results (r :scs (signed-reg)))
1534 (:result-types signed-num)
1536 (inst lea r (make-ea :dword :base base :index index
1537 :scale scale :disp disp))))
1539 (define-vop (%lea/fixnum=>fixnum)
1541 (:policy :fast-safe)
1542 (:args (base :scs (any-reg))
1543 (index :scs (any-reg)))
1545 (:arg-types tagged-num tagged-num
1546 (:constant (member 1 2 4 8))
1547 (:constant (signed-byte 32)))
1548 (:results (r :scs (any-reg)))
1549 (:result-types tagged-num)
1551 (inst lea r (make-ea :dword :base base :index index
1552 :scale scale :disp disp))))
1556 ;;; This is essentially a straight implementation of the algorithm in
1557 ;;; "Strength Reduction of Multiplications by Integer Constants",
1558 ;;; Youfeng Wu, ACM SIGPLAN Notices, Vol. 30, No.2, February 1995.
1559 (defun basic-decompose-multiplication (arg num n-bits condensed)
1560 (case (aref condensed 0)
1562 (let ((tmp (min 3 (aref condensed 1))))
1563 (decf (aref condensed 1) tmp)
1564 `(truly-the (unsigned-byte 32)
1566 ,(decompose-multiplication
1567 arg (ash (1- num) (- tmp)) (1- n-bits) (subseq condensed 1))
1570 (let ((r0 (aref condensed 0)))
1571 (incf (aref condensed 1) r0)
1572 `(truly-the (unsigned-byte 32)
1573 (%lea ,(decompose-multiplication
1574 arg (- num (ash 1 r0)) (1- n-bits) (subseq condensed 1))
1577 (t (let ((r0 (aref condensed 0)))
1578 (setf (aref condensed 0) 0)
1579 `(truly-the (unsigned-byte 32)
1580 (ash ,(decompose-multiplication
1581 arg (ash num (- r0)) n-bits condensed)
1584 (defun decompose-multiplication (arg num n-bits condensed)
1589 `(truly-the (unsigned-byte 32) (ash ,arg ,(1- (integer-length num)))))
1590 ((let ((max 0) (end 0))
1591 (loop for i from 2 to (length condensed)
1592 for j = (reduce #'+ (subseq condensed 0 i))
1593 when (and (> (- (* 2 i) 3 j) max)
1594 (< (+ (ash 1 (1+ j))
1595 (ash (ldb (byte (- 32 (1+ j)) (1+ j)) num)
1598 do (setq max (- (* 2 i) 3 j)
1601 (let ((j (reduce #'+ (subseq condensed 0 end))))
1602 (let ((n2 (+ (ash 1 (1+ j))
1603 (ash (ldb (byte (- 32 (1+ j)) (1+ j)) num) (1+ j))))
1604 (n1 (1+ (ldb (byte (1+ j) 0) (lognot num)))))
1605 `(truly-the (unsigned-byte 32)
1606 (- ,(optimize-multiply arg n2) ,(optimize-multiply arg n1))))))))
1607 ((dolist (i '(9 5 3))
1608 (when (integerp (/ num i))
1609 (when (< (logcount (/ num i)) (logcount num))
1611 (return `(let ((,x ,(optimize-multiply arg (/ num i))))
1612 (truly-the (unsigned-byte 32)
1613 (%lea ,x ,x (1- ,i) 0)))))))))
1614 (t (basic-decompose-multiplication arg num n-bits condensed))))
1616 (defun optimize-multiply (arg x)
1617 (let* ((n-bits (logcount x))
1618 (condensed (make-array n-bits)))
1619 (let ((count 0) (bit 0))
1621 (cond ((logbitp i x)
1622 (setf (aref condensed bit) count)
1626 (decompose-multiplication arg x n-bits condensed)))
1628 (deftransform * ((x y)
1629 ((unsigned-byte 32) (constant-arg (unsigned-byte 32)))
1631 "recode as leas, shifts and adds"
1632 (let ((y (lvar-value y)))
1634 ((= y (ash 1 (integer-length y)))
1635 ;; there's a generic transform for y = 2^k
1636 (give-up-ir1-transform))
1637 ((member y '(3 5 9))
1638 ;; we can do these multiplications directly using LEA
1639 `(%lea x x ,(1- y) 0))
1640 ((member :pentium4 *backend-subfeatures*)
1641 ;; the pentium4's multiply unit is reportedly very good
1642 (give-up-ir1-transform))
1643 ;; FIXME: should make this more fine-grained. If nothing else,
1644 ;; there should probably be a cutoff of about 9 instructions on
1645 ;; pentium-class machines.
1646 (t (optimize-multiply 'x y)))))
1648 ;;; FIXME: we should also be able to write an optimizer or two to
1649 ;;; convert (+ (* x 2) 17), (- (* x 9) 5) to a %LEA.