Annotation of OpenXM_contrib/gmp/mpn/x86/k7/mmx/divrem_1.asm, Revision 1.1.1.2
1.1.1.2 ! ohara 1: dnl AMD K7 mpn_divrem_1, mpn_divrem_1c, mpn_preinv_divrem_1 -- mpn by limb
! 2: dnl division.
1.1 maekawa 3:
1.1.1.2 ! ohara 4: dnl Copyright 1999, 2000, 2001, 2002 Free Software Foundation, Inc.
! 5: dnl
1.1 maekawa 6: dnl This file is part of the GNU MP Library.
1.1.1.2 ! ohara 7: dnl
1.1 maekawa 8: dnl The GNU MP Library is free software; you can redistribute it and/or
9: dnl modify it under the terms of the GNU Lesser General Public License as
10: dnl published by the Free Software Foundation; either version 2.1 of the
11: dnl License, or (at your option) any later version.
1.1.1.2 ! ohara 12: dnl
1.1 maekawa 13: dnl The GNU MP Library is distributed in the hope that it will be useful,
14: dnl but WITHOUT ANY WARRANTY; without even the implied warranty of
15: dnl MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16: dnl Lesser General Public License for more details.
1.1.1.2 ! ohara 17: dnl
1.1 maekawa 18: dnl You should have received a copy of the GNU Lesser General Public
19: dnl License along with the GNU MP Library; see the file COPYING.LIB. If
20: dnl not, write to the Free Software Foundation, Inc., 59 Temple Place -
21: dnl Suite 330, Boston, MA 02111-1307, USA.
22:
23: include(`../config.m4')
24:
25:
1.1.1.2 ! ohara 26: C K7: 17.0 cycles/limb integer part, 15.0 cycles/limb fraction part.
! 27:
! 28:
1.1 maekawa 29: C mp_limb_t mpn_divrem_1 (mp_ptr dst, mp_size_t xsize,
30: C mp_srcptr src, mp_size_t size,
31: C mp_limb_t divisor);
32: C mp_limb_t mpn_divrem_1c (mp_ptr dst, mp_size_t xsize,
33: C mp_srcptr src, mp_size_t size,
34: C mp_limb_t divisor, mp_limb_t carry);
1.1.1.2 ! ohara 35: C mp_limb_t mpn_preinv_divrem_1 (mp_ptr dst, mp_size_t xsize,
! 36: C mp_srcptr src, mp_size_t size,
! 37: C mp_limb_t divisor, mp_limb_t inverse,
! 38: C unsigned shift);
1.1 maekawa 39: C
40: C The method and nomenclature follow part 8 of "Division by Invariant
41: C Integers using Multiplication" by Granlund and Montgomery, reference in
42: C gmp.texi.
43: C
44: C The "and"s shown in the paper are done here with "cmov"s. "m" is written
45: C for m', and "d" for d_norm, which won't cause any confusion since it's
46: C only the normalized divisor that's of any use in the code. "b" is written
47: C for 2^N, the size of a limb, N being 32 here.
48: C
1.1.1.2 ! ohara 49: C mpn_divrem_1 and mpn_preinv_divrem_1 avoid one division if the src high
! 50: C limb is less than the divisor. mpn_divrem_1c doesn't check for a zero
! 51: C carry, since in normal circumstances that will be a very rare event.
! 52: C
! 53: C The test for skipping a division is branch free (once size>=1 is tested).
! 54: C The store to the destination high limb is 0 when a divide is skipped, or
! 55: C if it's not skipped then a copy of the src high limb is used. The latter
! 56: C is in case src==dst.
1.1 maekawa 57: C
58: C There's a small bias towards expecting xsize==0, by having code for
59: C xsize==0 in a straight line and xsize!=0 under forward jumps.
1.1.1.2 ! ohara 60: C
! 61: C Alternatives:
! 62: C
! 63: C If the divisor is normalized (high bit set) then a division step can
! 64: C always be skipped, since the high destination limb is always 0 or 1 in
! 65: C that case. It doesn't seem worth checking for this though, since it
! 66: C probably occurs infrequently, in particular note that big_base for a
! 67: C decimal mpn_get_str is not normalized in a 32-bit limb.
1.1 maekawa 68:
69:
70: dnl MUL_THRESHOLD is the value of xsize+size at which the multiply by
71: dnl inverse method is used, rather than plain "divl"s. Minimum value 1.
72: dnl
73: dnl The inverse takes about 50 cycles to calculate, but after that the
74: dnl multiply is 17 c/l versus division at 42 c/l.
75: dnl
76: dnl At 3 limbs the mul is a touch faster than div on the integer part, and
77: dnl even more so on the fractional part.
78:
79: deflit(MUL_THRESHOLD, 3)
80:
81:
1.1.1.2 ! ohara 82: defframe(PARAM_PREINV_SHIFT, 28) dnl mpn_preinv_divrem_1
! 83: defframe(PARAM_PREINV_INVERSE, 24) dnl mpn_preinv_divrem_1
! 84: defframe(PARAM_CARRY, 24) dnl mpn_divrem_1c
1.1 maekawa 85: defframe(PARAM_DIVISOR,20)
86: defframe(PARAM_SIZE, 16)
87: defframe(PARAM_SRC, 12)
88: defframe(PARAM_XSIZE, 8)
89: defframe(PARAM_DST, 4)
90:
91: defframe(SAVE_EBX, -4)
92: defframe(SAVE_ESI, -8)
93: defframe(SAVE_EDI, -12)
94: defframe(SAVE_EBP, -16)
95:
96: defframe(VAR_NORM, -20)
97: defframe(VAR_INVERSE, -24)
98: defframe(VAR_SRC, -28)
99: defframe(VAR_DST, -32)
100: defframe(VAR_DST_STOP,-36)
101:
102: deflit(STACK_SPACE, 36)
103:
1.1.1.2 ! ohara 104: TEXT
1.1 maekawa 105: ALIGN(32)
106:
1.1.1.2 ! ohara 107: PROLOGUE(mpn_preinv_divrem_1)
! 108: deflit(`FRAME',0)
! 109: movl PARAM_XSIZE, %ecx
! 110: movl PARAM_DST, %edx
! 111: subl $STACK_SPACE, %esp FRAME_subl_esp(STACK_SPACE)
! 112:
! 113: movl %esi, SAVE_ESI
! 114: movl PARAM_SRC, %esi
! 115:
! 116: movl %ebx, SAVE_EBX
! 117: movl PARAM_SIZE, %ebx
! 118:
! 119: leal 8(%edx,%ecx,4), %edx C &dst[xsize+2]
! 120: movl %ebp, SAVE_EBP
! 121: movl PARAM_DIVISOR, %ebp
! 122:
! 123: movl %edx, VAR_DST_STOP C &dst[xsize+2]
! 124: movl %edi, SAVE_EDI
! 125: xorl %edi, %edi C carry
! 126:
! 127: movl -4(%esi,%ebx,4), %eax C src high limb
! 128: xor %ecx, %ecx
! 129:
! 130: C
! 131:
! 132: C
! 133:
! 134: cmpl %ebp, %eax C high cmp divisor
! 135:
! 136: cmovc( %eax, %edi) C high is carry if high<divisor
! 137: cmovnc( %eax, %ecx) C 0 if skip div, src high if not
! 138: C (the latter in case src==dst)
! 139:
! 140: movl %ecx, -12(%edx,%ebx,4) C dst high limb
! 141: sbbl $0, %ebx C skip one division if high<divisor
! 142: movl PARAM_PREINV_SHIFT, %ecx
! 143:
! 144: leal -8(%edx,%ebx,4), %edx C &dst[xsize+size]
! 145: movl $32, %eax
! 146:
! 147: movl %edx, VAR_DST C &dst[xsize+size]
! 148:
! 149: shll %cl, %ebp C d normalized
! 150: subl %ecx, %eax
! 151: movl %ecx, VAR_NORM
! 152:
! 153: movd %eax, %mm7 C rshift
! 154: movl PARAM_PREINV_INVERSE, %eax
! 155: jmp L(start_preinv)
! 156:
! 157: EPILOGUE()
! 158:
! 159:
! 160: ALIGN(16)
! 161:
1.1 maekawa 162: PROLOGUE(mpn_divrem_1c)
163: deflit(`FRAME',0)
164: movl PARAM_CARRY, %edx
165: movl PARAM_SIZE, %ecx
166: subl $STACK_SPACE, %esp
167: deflit(`FRAME',STACK_SPACE)
168:
169: movl %ebx, SAVE_EBX
170: movl PARAM_XSIZE, %ebx
171:
172: movl %edi, SAVE_EDI
173: movl PARAM_DST, %edi
174:
175: movl %ebp, SAVE_EBP
176: movl PARAM_DIVISOR, %ebp
177:
178: movl %esi, SAVE_ESI
179: movl PARAM_SRC, %esi
180:
1.1.1.2 ! ohara 181: leal -4(%edi,%ebx,4), %edi C &dst[xsize-1]
! 182: jmp L(start_1c)
1.1 maekawa 183:
184: EPILOGUE()
185:
186:
1.1.1.2 ! ohara 187: C offset 0xa1, close enough to aligned
1.1 maekawa 188: PROLOGUE(mpn_divrem_1)
189: deflit(`FRAME',0)
190:
191: movl PARAM_SIZE, %ecx
192: movl $0, %edx C initial carry (if can't skip a div)
193: subl $STACK_SPACE, %esp
194: deflit(`FRAME',STACK_SPACE)
195:
1.1.1.2 ! ohara 196: movl %esi, SAVE_ESI
! 197: movl PARAM_SRC, %esi
1.1 maekawa 198:
199: movl %ebx, SAVE_EBX
200: movl PARAM_XSIZE, %ebx
201:
1.1.1.2 ! ohara 202: movl %ebp, SAVE_EBP
! 203: movl PARAM_DIVISOR, %ebp
! 204: orl %ecx, %ecx C size
1.1 maekawa 205:
206: movl %edi, SAVE_EDI
207: movl PARAM_DST, %edi
208: leal -4(%edi,%ebx,4), %edi C &dst[xsize-1]
209:
1.1.1.2 ! ohara 210: jz L(no_skip_div) C if size==0
1.1 maekawa 211: movl -4(%esi,%ecx,4), %eax C src high limb
1.1.1.2 ! ohara 212: xorl %esi, %esi
1.1 maekawa 213:
1.1.1.2 ! ohara 214: cmpl %ebp, %eax C high cmp divisor
1.1 maekawa 215:
1.1.1.2 ! ohara 216: cmovc( %eax, %edx) C high is carry if high<divisor
! 217: cmovnc( %eax, %esi) C 0 if skip div, src high if not
! 218:
! 219: movl %esi, (%edi,%ecx,4) C dst high limb
! 220: sbbl $0, %ecx C size-1 if high<divisor
! 221: movl PARAM_SRC, %esi C reload
1.1 maekawa 222: L(no_skip_div):
223:
224:
225: L(start_1c):
1.1.1.2 ! ohara 226: C eax
1.1 maekawa 227: C ebx xsize
228: C ecx size
229: C edx carry
230: C esi src
231: C edi &dst[xsize-1]
232: C ebp divisor
233:
234: leal (%ebx,%ecx), %eax C size+xsize
235: cmpl $MUL_THRESHOLD, %eax
236: jae L(mul_by_inverse)
237:
238:
239: C With MUL_THRESHOLD set to 3, the simple loops here only do 0 to 2 limbs.
240: C It'd be possible to write them out without the looping, but no speedup
241: C would be expected.
242: C
243: C Using PARAM_DIVISOR instead of %ebp measures 1 cycle/loop faster on the
244: C integer part, but curiously not on the fractional part, where %ebp is a
245: C (fixed) couple of cycles faster.
246:
247: orl %ecx, %ecx
248: jz L(divide_no_integer)
249:
250: L(divide_integer):
251: C eax scratch (quotient)
252: C ebx xsize
253: C ecx counter
254: C edx scratch (remainder)
255: C esi src
256: C edi &dst[xsize-1]
257: C ebp divisor
258:
259: movl -4(%esi,%ecx,4), %eax
260:
261: divl PARAM_DIVISOR
262:
263: movl %eax, (%edi,%ecx,4)
264: decl %ecx
265: jnz L(divide_integer)
266:
267:
268: L(divide_no_integer):
269: movl PARAM_DST, %edi
270: orl %ebx, %ebx
271: jnz L(divide_fraction)
272:
273: L(divide_done):
274: movl SAVE_ESI, %esi
275: movl SAVE_EDI, %edi
276: movl %edx, %eax
277:
278: movl SAVE_EBX, %ebx
279: movl SAVE_EBP, %ebp
280: addl $STACK_SPACE, %esp
281:
282: ret
283:
284:
285: L(divide_fraction):
286: C eax scratch (quotient)
287: C ebx counter
288: C ecx
289: C edx scratch (remainder)
290: C esi
291: C edi dst
292: C ebp divisor
293:
294: movl $0, %eax
295:
296: divl %ebp
297:
298: movl %eax, -4(%edi,%ebx,4)
299: decl %ebx
300: jnz L(divide_fraction)
301:
302: jmp L(divide_done)
303:
304:
305:
306: C -----------------------------------------------------------------------------
307:
308: L(mul_by_inverse):
309: C eax
310: C ebx xsize
311: C ecx size
312: C edx carry
313: C esi src
314: C edi &dst[xsize-1]
315: C ebp divisor
316:
317: bsrl %ebp, %eax C 31-l
318:
1.1.1.2 ! ohara 319: leal 12(%edi), %ebx C &dst[xsize+2], loop dst stop
1.1 maekawa 320: leal 4(%edi,%ecx,4), %edi C &dst[xsize+size]
321:
322: movl %edi, VAR_DST
323: movl %ebx, VAR_DST_STOP
324:
325: movl %ecx, %ebx C size
326: movl $31, %ecx
327:
328: movl %edx, %edi C carry
329: movl $-1, %edx
330:
331: C
332:
333: xorl %eax, %ecx C l
334: incl %eax C 32-l
335:
336: shll %cl, %ebp C d normalized
337: movl %ecx, VAR_NORM
338:
339: movd %eax, %mm7
340:
341: movl $-1, %eax
342: subl %ebp, %edx C (b-d)-1 giving edx:eax = b*(b-d)-1
343:
344: divl %ebp C floor (b*(b-d)-1) / d
345:
1.1.1.2 ! ohara 346: L(start_preinv):
! 347: C eax inverse
! 348: C ebx size
! 349: C ecx shift
! 350: C edx
! 351: C esi src
! 352: C edi carry
! 353: C ebp divisor
! 354: C
! 355: C mm7 rshift
! 356:
1.1 maekawa 357: orl %ebx, %ebx C size
358: movl %eax, VAR_INVERSE
359: leal -12(%esi,%ebx,4), %eax C &src[size-3]
360:
361: jz L(start_zero)
362: movl %eax, VAR_SRC
363: cmpl $1, %ebx
364:
365: movl 8(%eax), %esi C src high limb
366: jz L(start_one)
367:
368: L(start_two_or_more):
369: movl 4(%eax), %edx C src second highest limb
370:
371: shldl( %cl, %esi, %edi) C n2 = carry,high << l
372:
373: shldl( %cl, %edx, %esi) C n10 = high,second << l
374:
375: cmpl $2, %ebx
376: je L(integer_two_left)
377: jmp L(integer_top)
378:
379:
380: L(start_one):
381: shldl( %cl, %esi, %edi) C n2 = carry,high << l
382:
383: shll %cl, %esi C n10 = high << l
384: movl %eax, VAR_SRC
385: jmp L(integer_one_left)
386:
387:
388: L(start_zero):
1.1.1.2 ! ohara 389: C Can be here with xsize==0 if mpn_preinv_divrem_1 had size==1 and
! 390: C skipped a division.
! 391:
1.1 maekawa 392: shll %cl, %edi C n2 = carry << l
1.1.1.2 ! ohara 393: movl %edi, %eax C return value for zero_done
! 394: cmpl $0, PARAM_XSIZE
1.1 maekawa 395:
1.1.1.2 ! ohara 396: je L(zero_done)
1.1 maekawa 397: jmp L(fraction_some)
398:
399:
400:
401: C -----------------------------------------------------------------------------
402: C
403: C The multiply by inverse loop is 17 cycles, and relies on some out-of-order
404: C execution. The instruction scheduling is important, with various
405: C apparently equivalent forms running 1 to 5 cycles slower.
406: C
407: C A lower bound for the time would seem to be 16 cycles, based on the
408: C following successive dependencies.
409: C
410: C cycles
411: C n2+n1 1
412: C mul 6
413: C q1+1 1
414: C mul 6
415: C sub 1
416: C addback 1
417: C ---
418: C 16
419: C
420: C This chain is what the loop has already, but 16 cycles isn't achieved.
421: C K7 has enough decode, and probably enough execute (depending maybe on what
422: C a mul actually consumes), but nothing running under 17 has been found.
423: C
424: C In theory n2+n1 could be done in the sub and addback stages (by
425: C calculating both n2 and n2+n1 there), but lack of registers makes this an
426: C unlikely proposition.
427: C
428: C The jz in the loop keeps the q1+1 stage to 1 cycle. Handling an overflow
429: C from q1+1 with an "sbbl $0, %ebx" would add a cycle to the dependent
430: C chain, and nothing better than 18 cycles has been found when using it.
431: C The jump is taken only when q1 is 0xFFFFFFFF, and on random data this will
432: C be an extremely rare event.
433: C
434: C Branch mispredictions will hit random occurrances of q1==0xFFFFFFFF, but
435: C if some special data is coming out with this always, the q1_ff special
436: C case actually runs at 15 c/l. 0x2FFF...FFFD divided by 3 is a good way to
437: C induce the q1_ff case, for speed measurements or testing. Note that
438: C 0xFFF...FFF divided by 1 or 2 doesn't induce it.
439: C
440: C The instruction groupings and empty comments show the cycles for a naive
441: C in-order view of the code (conveniently ignoring the load latency on
442: C VAR_INVERSE). This shows some of where the time is going, but is nonsense
443: C to the extent that out-of-order execution rearranges it. In this case
444: C there's 19 cycles shown, but it executes at 17.
445:
446: ALIGN(16)
447: L(integer_top):
448: C eax scratch
449: C ebx scratch (nadj, q1)
450: C ecx scratch (src, dst)
451: C edx scratch
452: C esi n10
453: C edi n2
454: C ebp divisor
455: C
456: C mm0 scratch (src qword)
457: C mm7 rshift for normalization
458:
459: cmpl $0x80000000, %esi C n1 as 0=c, 1=nc
460: movl %edi, %eax C n2
461: movl VAR_SRC, %ecx
462:
463: leal (%ebp,%esi), %ebx
464: cmovc( %esi, %ebx) C nadj = n10 + (-n1 & d), ignoring overflow
465: sbbl $-1, %eax C n2+n1
466:
467: mull VAR_INVERSE C m*(n2+n1)
468:
469: movq (%ecx), %mm0 C next limb and the one below it
470: subl $4, %ecx
471:
472: movl %ecx, VAR_SRC
473:
474: C
475:
476: addl %ebx, %eax C m*(n2+n1) + nadj, low giving carry flag
1.1.1.2 ! ohara 477: leal 1(%edi), %ebx C n2+1
1.1 maekawa 478: movl %ebp, %eax C d
479:
480: C
481:
482: adcl %edx, %ebx C 1 + high(n2<<32 + m*(n2+n1) + nadj) = q1+1
483: jz L(q1_ff)
484: movl VAR_DST, %ecx
485:
486: mull %ebx C (q1+1)*d
487:
488: psrlq %mm7, %mm0
489:
490: leal -4(%ecx), %ecx
491:
492: C
493:
494: subl %eax, %esi
495: movl VAR_DST_STOP, %eax
496:
497: C
498:
499: sbbl %edx, %edi C n - (q1+1)*d
500: movl %esi, %edi C remainder -> n2
501: leal (%ebp,%esi), %edx
502:
503: movd %mm0, %esi
504:
505: cmovc( %edx, %edi) C n - q1*d if underflow from using q1+1
506: sbbl $0, %ebx C q
507: cmpl %eax, %ecx
508:
509: movl %ebx, (%ecx)
510: movl %ecx, VAR_DST
511: jne L(integer_top)
512:
513:
514: L(integer_loop_done):
515:
516:
517: C -----------------------------------------------------------------------------
518: C
519: C Here, and in integer_one_left below, an sbbl $0 is used rather than a jz
520: C q1_ff special case. This make the code a bit smaller and simpler, and
521: C costs only 1 cycle (each).
522:
523: L(integer_two_left):
524: C eax scratch
525: C ebx scratch (nadj, q1)
526: C ecx scratch (src, dst)
527: C edx scratch
528: C esi n10
529: C edi n2
530: C ebp divisor
531: C
532: C mm7 rshift
533:
534: cmpl $0x80000000, %esi C n1 as 0=c, 1=nc
535: movl %edi, %eax C n2
536: movl PARAM_SRC, %ecx
537:
538: leal (%ebp,%esi), %ebx
539: cmovc( %esi, %ebx) C nadj = n10 + (-n1 & d), ignoring overflow
540: sbbl $-1, %eax C n2+n1
541:
542: mull VAR_INVERSE C m*(n2+n1)
543:
544: movd (%ecx), %mm0 C src low limb
545:
546: movl VAR_DST_STOP, %ecx
547:
548: C
549:
550: addl %ebx, %eax C m*(n2+n1) + nadj, low giving carry flag
1.1.1.2 ! ohara 551: leal 1(%edi), %ebx C n2+1
1.1 maekawa 552: movl %ebp, %eax C d
553:
554: adcl %edx, %ebx C 1 + high(n2<<32 + m*(n2+n1) + nadj) = q1+1
555:
556: sbbl $0, %ebx
557:
558: mull %ebx C (q1+1)*d
559:
560: psllq $32, %mm0
561:
562: psrlq %mm7, %mm0
563:
564: C
565:
566: subl %eax, %esi
567:
568: C
569:
570: sbbl %edx, %edi C n - (q1+1)*d
571: movl %esi, %edi C remainder -> n2
572: leal (%ebp,%esi), %edx
573:
574: movd %mm0, %esi
575:
576: cmovc( %edx, %edi) C n - q1*d if underflow from using q1+1
577: sbbl $0, %ebx C q
578:
579: movl %ebx, -4(%ecx)
580:
581:
582: C -----------------------------------------------------------------------------
583: L(integer_one_left):
584: C eax scratch
585: C ebx scratch (nadj, q1)
586: C ecx dst
587: C edx scratch
588: C esi n10
589: C edi n2
590: C ebp divisor
591: C
592: C mm7 rshift
593:
594: movl VAR_DST_STOP, %ecx
595: cmpl $0x80000000, %esi C n1 as 0=c, 1=nc
596: movl %edi, %eax C n2
597:
598: leal (%ebp,%esi), %ebx
599: cmovc( %esi, %ebx) C nadj = n10 + (-n1 & d), ignoring overflow
600: sbbl $-1, %eax C n2+n1
601:
602: mull VAR_INVERSE C m*(n2+n1)
603:
604: C
605:
606: C
607:
608: C
609:
610: addl %ebx, %eax C m*(n2+n1) + nadj, low giving carry flag
1.1.1.2 ! ohara 611: leal 1(%edi), %ebx C n2+1
1.1 maekawa 612: movl %ebp, %eax C d
613:
614: C
615:
616: adcl %edx, %ebx C 1 + high(n2<<32 + m*(n2+n1) + nadj) = q1+1
617:
618: sbbl $0, %ebx C q1 if q1+1 overflowed
619:
620: mull %ebx
621:
622: C
623:
624: C
625:
626: C
627:
628: subl %eax, %esi
629:
630: C
631:
632: sbbl %edx, %edi C n - (q1+1)*d
633: movl %esi, %edi C remainder -> n2
634: leal (%ebp,%esi), %edx
635:
636: cmovc( %edx, %edi) C n - q1*d if underflow from using q1+1
637: sbbl $0, %ebx C q
638:
639: movl %ebx, -8(%ecx)
640: subl $8, %ecx
641:
642:
643:
644: L(integer_none):
645: cmpl $0, PARAM_XSIZE
646: jne L(fraction_some)
647:
648: movl %edi, %eax
649: L(fraction_done):
650: movl VAR_NORM, %ecx
1.1.1.2 ! ohara 651: L(zero_done):
1.1 maekawa 652: movl SAVE_EBP, %ebp
653:
654: movl SAVE_EDI, %edi
655: movl SAVE_ESI, %esi
656:
657: movl SAVE_EBX, %ebx
658: addl $STACK_SPACE, %esp
659:
660: shrl %cl, %eax
661: emms
662:
663: ret
664:
665:
666: C -----------------------------------------------------------------------------
667: C
668: C Special case for q1=0xFFFFFFFF, giving q=0xFFFFFFFF meaning the low dword
669: C of q*d is simply -d and the remainder n-q*d = n10+d
670:
671: L(q1_ff):
672: C eax (divisor)
673: C ebx (q1+1 == 0)
674: C ecx
675: C edx
676: C esi n10
677: C edi n2
678: C ebp divisor
679:
680: movl VAR_DST, %ecx
681: movl VAR_DST_STOP, %edx
682: subl $4, %ecx
683:
684: psrlq %mm7, %mm0
685: leal (%ebp,%esi), %edi C n-q*d remainder -> next n2
686: movl %ecx, VAR_DST
687:
688: movd %mm0, %esi C next n10
689:
690: movl $-1, (%ecx)
691: cmpl %ecx, %edx
692: jne L(integer_top)
693:
694: jmp L(integer_loop_done)
695:
696:
697:
698: C -----------------------------------------------------------------------------
699: C
700: C Being the fractional part, the "source" limbs are all zero, meaning
701: C n10=0, n1=0, and hence nadj=0, leading to many instructions eliminated.
702: C
703: C The loop runs at 15 cycles. The dependent chain is the same as the
704: C general case above, but without the n2+n1 stage (due to n1==0), so 15
705: C would seem to be the lower bound.
706: C
707: C A not entirely obvious simplification is that q1+1 never overflows a limb,
708: C and so there's no need for the sbbl $0 or jz q1_ff from the general case.
709: C q1 is the high word of m*n2+b*n2 and the following shows q1<=b-2 always.
710: C rnd() means rounding down to a multiple of d.
711: C
712: C m*n2 + b*n2 <= m*(d-1) + b*(d-1)
713: C = m*d + b*d - m - b
714: C = floor((b(b-d)-1)/d)*d + b*d - m - b
715: C = rnd(b(b-d)-1) + b*d - m - b
716: C = rnd(b(b-d)-1 + b*d) - m - b
717: C = rnd(b*b-1) - m - b
718: C <= (b-2)*b
719: C
720: C Unchanged from the general case is that the final quotient limb q can be
721: C either q1 or q1+1, and the q1+1 case occurs often. This can be seen from
722: C equation 8.4 of the paper which simplifies as follows when n1==0 and
723: C n0==0.
724: C
725: C n-q1*d = (n2*k+q0*d)/b <= d + (d*d-2d)/b
726: C
727: C As before, the instruction groupings and empty comments show a naive
728: C in-order view of the code, which is made a nonsense by out of order
729: C execution. There's 17 cycles shown, but it executes at 15.
730: C
731: C Rotating the store q and remainder->n2 instructions up to the top of the
732: C loop gets the run time down from 16 to 15.
733:
734: ALIGN(16)
735: L(fraction_some):
736: C eax
737: C ebx
738: C ecx
739: C edx
740: C esi
741: C edi carry
742: C ebp divisor
743:
744: movl PARAM_DST, %esi
1.1.1.2 ! ohara 745: movl VAR_DST_STOP, %ecx C &dst[xsize+2]
1.1 maekawa 746: movl %edi, %eax
747:
1.1.1.2 ! ohara 748: subl $8, %ecx C &dst[xsize]
1.1 maekawa 749: jmp L(fraction_entry)
750:
751:
752: ALIGN(16)
753: L(fraction_top):
754: C eax n2 carry, then scratch
755: C ebx scratch (nadj, q1)
756: C ecx dst, decrementing
757: C edx scratch
758: C esi dst stop point
759: C edi (will be n2)
760: C ebp divisor
761:
762: movl %ebx, (%ecx) C previous q
763: movl %eax, %edi C remainder->n2
764:
765: L(fraction_entry):
766: mull VAR_INVERSE C m*n2
767:
768: movl %ebp, %eax C d
769: subl $4, %ecx C dst
770: leal 1(%edi), %ebx
771:
772: C
773:
774: C
775:
776: C
777:
778: C
779:
780: addl %edx, %ebx C 1 + high(n2<<32 + m*n2) = q1+1
781:
782: mull %ebx C (q1+1)*d
783:
784: C
785:
786: C
787:
788: C
789:
790: negl %eax C low of n - (q1+1)*d
791:
792: C
793:
794: sbbl %edx, %edi C high of n - (q1+1)*d, caring only about carry
795: leal (%ebp,%eax), %edx
796:
797: cmovc( %edx, %eax) C n - q1*d if underflow from using q1+1
798: sbbl $0, %ebx C q
799: cmpl %esi, %ecx
800:
801: jne L(fraction_top)
802:
803:
804: movl %ebx, (%ecx)
805: jmp L(fraction_done)
806:
807: EPILOGUE()
FreeBSD-CVSweb <freebsd-cvsweb@FreeBSD.org>