Annotation of OpenXM_contrib/gmp/mpn/x86/p6/mod_1.asm, Revision 1.1.1.1
1.1 ohara 1: dnl Intel P6 mpn_mod_1 -- mpn by limb remainder.
2:
3: dnl Copyright 1999, 2000, 2002 Free Software Foundation, Inc.
4: dnl
5: dnl This file is part of the GNU MP Library.
6: dnl
7: dnl The GNU MP Library is free software; you can redistribute it and/or
8: dnl modify it under the terms of the GNU Lesser General Public License as
9: dnl published by the Free Software Foundation; either version 2.1 of the
10: dnl License, or (at your option) any later version.
11: dnl
12: dnl The GNU MP Library is distributed in the hope that it will be useful,
13: dnl but WITHOUT ANY WARRANTY; without even the implied warranty of
14: dnl MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15: dnl Lesser General Public License for more details.
16: dnl
17: dnl You should have received a copy of the GNU Lesser General Public
18: dnl License along with the GNU MP Library; see the file COPYING.LIB. If
19: dnl not, write to the Free Software Foundation, Inc., 59 Temple Place -
20: dnl Suite 330, Boston, MA 02111-1307, USA.
21:
22: include(`../config.m4')
23:
24:
25: C P6: 21.5 cycles/limb
26:
27:
28: C mp_limb_t mpn_mod_1 (mp_srcptr src, mp_size_t size, mp_limb_t divisor);
29: C mp_limb_t mpn_mod_1c (mp_srcptr src, mp_size_t size, mp_limb_t divisor,
30: C mp_limb_t carry);
31: C mp_limb_t mpn_preinv_mod_1 (mp_srcptr src, mp_size_t size, mp_limb_t divisor,
32: C mp_limb_t inverse);
33: C
34: C The code here is in two parts, a simple divl loop and a mul-by-inverse.
35: C The divl is used by mod_1 and mod_1c for small sizes, until the savings in
36: C the mul-by-inverse can overcome the time to calculate an inverse.
37: C preinv_mod_1 goes straight to the mul-by-inverse.
38: C
39: C The mul-by-inverse normalizes the divisor (or for preinv_mod_1 it's
40: C already normalized). The calculation done is r=a%(d*2^n) followed by a
41: C final (r*2^n)%(d*2^n), where a is the dividend, d the divisor, and n is
42: C the number of leading zero bits on d. This means there's no bit shifts in
43: C the main loop, at the cost of an extra divide step at the end.
44: C
45: C The simple divl for mod_1 is able to skip one divide step if high<divisor.
46: C For mod_1c the carry parameter is the high of the first divide step, and
47: C no attempt is make to skip that step since carry==0 will be very rare.
48: C
49: C The mul-by-inverse always skips one divide step, but then needs an extra
50: C step at the end, unless the divisor was already normalized (n==0). This
51: C leads to different mul-by-inverse thresholds for normalized and
52: C unnormalized divisors, in mod_1 and mod_1c.
53: C
54: C Alternatives:
55: C
56: C If n is small then the extra divide step could be done by a few shift and
57: C trial subtract steps instead of a full divide. That would probably be 3
58: C or 4 cycles/bit, so say up to n=8 might benefit from that over a 21 cycle
59: C divide. However it's considered that small divisors, meaning biggish n,
60: C are more likely than small n, and that it's not worth the branch
61: C mispredicts of a loop.
62: C
63: C Past:
64: C
65: C There used to be some MMX based code for P-II and P-III, roughly following
66: C the K7 form, but it was slower (about 24.0 c/l) than the code here. That
67: C code did have an advantage that mod_1 was able to do one less divide step
68: C when high<divisor and the divisor unnormalized, but the speed advantage of
69: C the current code soon overcomes that.
70: C
71: C Future:
72: C
73: C It's not clear whether what's here is optimal. A rough count of micro-ops
74: C on the dependent chain would suggest a couple of cycles could be shaved,
75: C perhaps.
76:
77:
78: dnl The following thresholds are the sizes where the multiply by inverse
79: dnl method is used instead of plain divl's. Minimum value 2 each.
80: dnl
81: dnl MUL_NORM_THRESHOLD is for normalized divisors (high bit set),
82: dnl MUL_UNNORM_THRESHOLD for unnormalized divisors.
83: dnl
84: dnl With the divl loop at 39 c/l, and the inverse loop at 21.5 c/l but
85: dnl setups for the inverse of about 50, the threshold should be around
86: dnl 50/(39-21.5)==2.85. An unnormalized divisor gets an extra divide step
87: dnl at the end, so if that's about 25 cycles then that threshold might be
88: dnl around (50+25)/(39-21.5) == 4.3.
89:
90: deflit(MUL_NORM_THRESHOLD, 4)
91: deflit(MUL_UNNORM_THRESHOLD, 5)
92:
93: deflit(MUL_NORM_DELTA, eval(MUL_NORM_THRESHOLD - MUL_UNNORM_THRESHOLD))
94:
95:
96: defframe(PARAM_INVERSE, 16) dnl mpn_preinv_mod_1
97: defframe(PARAM_CARRY, 16) dnl mpn_mod_1c
98: defframe(PARAM_DIVISOR, 12)
99: defframe(PARAM_SIZE, 8)
100: defframe(PARAM_SRC, 4)
101:
102: defframe(SAVE_EBX, -4)
103: defframe(SAVE_ESI, -8)
104: defframe(SAVE_EDI, -12)
105: defframe(SAVE_EBP, -16)
106:
107: defframe(VAR_NORM, -20)
108: defframe(VAR_INVERSE, -24)
109:
110: deflit(STACK_SPACE, 24)
111:
112: TEXT
113:
114: ALIGN(16)
115: PROLOGUE(mpn_preinv_mod_1)
116: deflit(`FRAME',0)
117:
118: movl PARAM_SRC, %edx
119: subl $STACK_SPACE, %esp FRAME_subl_esp(STACK_SPACE)
120:
121: movl %ebx, SAVE_EBX
122: movl PARAM_SIZE, %ebx
123:
124: movl %ebp, SAVE_EBP
125: movl PARAM_DIVISOR, %ebp
126:
127: movl %esi, SAVE_ESI
128: movl PARAM_INVERSE, %eax
129:
130: movl %edi, SAVE_EDI
131: movl -4(%edx,%ebx,4), %edi C src high limb
132:
133: movl $0, VAR_NORM
134: leal -8(%edx,%ebx,4), %ecx C &src[size-2]
135:
136: C
137:
138: movl %edi, %esi
139: subl %ebp, %edi C high-divisor
140:
141: cmovc( %esi, %edi) C restore if underflow
142: decl %ebx
143: jnz L(preinv_entry)
144:
145: jmp L(done_edi)
146:
147: EPILOGUE()
148:
149:
150: ALIGN(16)
151: PROLOGUE(mpn_mod_1c)
152: deflit(`FRAME',0)
153:
154: movl PARAM_SIZE, %ecx
155: subl $STACK_SPACE, %esp FRAME_subl_esp(STACK_SPACE)
156:
157: movl %ebp, SAVE_EBP
158: movl PARAM_DIVISOR, %eax
159:
160: movl %esi, SAVE_ESI
161: movl PARAM_CARRY, %edx
162:
163: movl PARAM_SRC, %esi
164: orl %ecx, %ecx
165: jz L(done_edx) C result==carry if size==0
166:
167: sarl $31, %eax
168: movl PARAM_DIVISOR, %ebp
169:
170: andl $MUL_NORM_DELTA, %eax
171:
172: addl $MUL_UNNORM_THRESHOLD, %eax
173:
174: cmpl %eax, %ecx
175: jb L(divide_top)
176:
177:
178: C The carry parameter pretends to be the src high limb.
179:
180: movl %ebx, SAVE_EBX
181: leal 1(%ecx), %ebx C size+1
182:
183: movl %edx, %eax C carry
184: jmp L(mul_by_inverse_1c)
185:
186: EPILOGUE()
187:
188:
189: ALIGN(16)
190: PROLOGUE(mpn_mod_1)
191: deflit(`FRAME',0)
192:
193: movl PARAM_SIZE, %ecx
194: subl $STACK_SPACE, %esp FRAME_subl_esp(STACK_SPACE)
195: movl $0, %edx C initial carry (if can't skip a div)
196:
197: movl %esi, SAVE_ESI
198: movl PARAM_SRC, %eax
199:
200: movl %ebp, SAVE_EBP
201: movl PARAM_DIVISOR, %ebp
202:
203: movl PARAM_DIVISOR, %esi
204: orl %ecx, %ecx
205: jz L(done_edx)
206:
207: movl -4(%eax,%ecx,4), %eax C src high limb
208:
209: sarl $31, %ebp
210:
211: andl $MUL_NORM_DELTA, %ebp
212:
213: addl $MUL_UNNORM_THRESHOLD, %ebp
214: cmpl %esi, %eax C carry flag if high<divisor
215:
216: cmovc( %eax, %edx) C src high limb as initial carry
217: movl PARAM_SRC, %esi
218:
219: sbbl $0, %ecx C size-1 to skip one div
220: jz L(done_eax) C done if had size==1
221:
222: cmpl %ebp, %ecx
223: movl PARAM_DIVISOR, %ebp
224: jae L(mul_by_inverse)
225:
226:
227: L(divide_top):
228: C eax scratch (quotient)
229: C ebx
230: C ecx counter, limbs, decrementing
231: C edx scratch (remainder)
232: C esi src
233: C edi
234: C ebp divisor
235:
236: movl -4(%esi,%ecx,4), %eax
237:
238: divl %ebp
239:
240: decl %ecx
241: jnz L(divide_top)
242:
243:
244: L(done_edx):
245: movl %edx, %eax
246: L(done_eax):
247: movl SAVE_ESI, %esi
248:
249: movl SAVE_EBP, %ebp
250: addl $STACK_SPACE, %esp
251:
252: ret
253:
254:
255: C -----------------------------------------------------------------------------
256:
257: L(mul_by_inverse):
258: C eax src high limb
259: C ebx
260: C ecx
261: C edx
262: C esi src
263: C edi
264: C ebp divisor
265:
266: movl %ebx, SAVE_EBX
267: movl PARAM_SIZE, %ebx
268:
269: L(mul_by_inverse_1c):
270: bsrl %ebp, %ecx C 31-l
271:
272: movl %edi, SAVE_EDI
273: xorl $31, %ecx C l
274:
275: movl %ecx, VAR_NORM
276: shll %cl, %ebp C d normalized
277:
278: movl %eax, %edi C src high -> n2
279: subl %ebp, %eax
280:
281: cmovnc( %eax, %edi) C n2-divisor if no underflow
282:
283: movl $-1, %eax
284: movl $-1, %edx
285:
286: subl %ebp, %edx C (b-d)-1 so edx:eax = b*(b-d)-1
287: leal -8(%esi,%ebx,4), %ecx C &src[size-2]
288:
289: divl %ebp C floor (b*(b-d)-1) / d
290:
291: L(preinv_entry):
292: movl %eax, VAR_INVERSE
293:
294:
295:
296: C No special scheduling of loads is necessary in this loop, out of order
297: C execution hides the latencies already.
298: C
299: C The way q1+1 is generated in %ebx and d is moved to %eax for the multiply
300: C seems fastest. The obvious change to generate q1+1 in %eax and then just
301: C multiply by %ebp (as per mpn/x86/pentium/mod_1.asm in fact) runs 1 cycle
302: C slower, for no obvious reason.
303:
304:
305: ALIGN(16)
306: L(inverse_top):
307: C eax n10 (then scratch)
308: C ebx scratch (nadj, q1)
309: C ecx src pointer, decrementing
310: C edx scratch
311: C esi n10
312: C edi n2
313: C ebp divisor
314:
315: movl (%ecx), %eax C next src limb
316: movl %eax, %esi
317:
318: sarl $31, %eax C -n1
319: movl %ebp, %ebx
320:
321: andl %eax, %ebx C -n1 & d
322: negl %eax C n1
323:
324: addl %edi, %eax C n2+n1
325:
326: mull VAR_INVERSE C m*(n2+n1)
327:
328: addl %esi, %ebx C nadj = n10 + (-n1 & d), ignoring overflow
329: subl $4, %ecx
330:
331: C
332:
333: addl %ebx, %eax C m*(n2+n1) + nadj, low giving carry flag
334: leal 1(%edi), %ebx C n2+1
335: movl %ebp, %eax C d
336:
337: adcl %edx, %ebx C 1 + high(n2<<32 + m*(n2+n1) + nadj) = q1+1
338: jz L(q1_ff)
339:
340: mull %ebx C (q1+1)*d
341:
342: C
343:
344: subl %eax, %esi C low n - (q1+1)*d
345:
346: sbbl %edx, %edi C high n - (q1+1)*d, 0 or -1
347:
348: andl %ebp, %edi C d if underflow
349:
350: addl %esi, %edi C remainder with addback if necessary
351:
352: cmpl PARAM_SRC, %ecx
353: jae L(inverse_top)
354:
355:
356: C -----------------------------------------------------------------------------
357: L(inverse_loop_done):
358:
359: C %edi is the remainder modulo d*2^n and now must be reduced to
360: C 0<=r<d by calculating r*2^n mod d*2^n and then right shifting by
361: C n. If d was already normalized on entry so that n==0 then nothing
362: C is needed here. The chance of n==0 is low, but it's true of say
363: C PP from gmp-impl.h.
364: C
365: C eax
366: C ebx
367: C ecx
368: C edx
369: C esi
370: C edi remainder
371: C ebp divisor (normalized)
372:
373: movl VAR_NORM, %ecx
374: movl $0, %esi
375:
376: orl %ecx, %ecx
377: jz L(done_edi)
378:
379:
380: C Here use %edi=n10 and %esi=n2, opposite to the loop above.
381: C
382: C The q1=0xFFFFFFFF case is handled with an sbbl to adjust q1+1
383: C back, rather than q1_ff special case code. This is simpler and
384: C costs only 2 uops.
385:
386: shldl( %cl, %edi, %esi)
387:
388: shll %cl, %edi
389:
390: movl %edi, %eax C n10
391: movl %ebp, %ebx C d
392:
393: sarl $31, %eax C -n1
394:
395: andl %eax, %ebx C -n1 & d
396: negl %eax C n1
397:
398: addl %edi, %ebx C nadj = n10 + (-n1 & d), ignoring overflow
399: addl %esi, %eax C n2+n1
400:
401: mull VAR_INVERSE C m*(n2+n1)
402:
403: C
404:
405: addl %ebx, %eax C m*(n2+n1) + nadj, low giving carry flag
406: leal 1(%esi), %ebx C n2+1
407:
408: adcl %edx, %ebx C 1 + high(n2<<32 + m*(n2+n1) + nadj) = q1+1
409:
410: sbbl $0, %ebx
411: movl %ebp, %eax C d
412:
413: mull %ebx C (q1+1)*d
414:
415: movl SAVE_EBX, %ebx
416:
417: C
418:
419: subl %eax, %edi C low n - (q1+1)*d is remainder
420:
421: sbbl %edx, %esi C high n - (q1+1)*d, 0 or -1
422:
423: andl %ebp, %esi
424: movl SAVE_EBP, %ebp
425:
426: leal (%esi,%edi), %eax C remainder
427: movl SAVE_ESI, %esi
428:
429: shrl %cl, %eax C denorm remainder
430: movl SAVE_EDI, %edi
431: addl $STACK_SPACE, %esp
432:
433: ret
434:
435:
436: L(done_edi):
437: movl SAVE_EBX, %ebx
438: movl %edi, %eax
439:
440: movl SAVE_ESI, %esi
441:
442: movl SAVE_EDI, %edi
443:
444: movl SAVE_EBP, %ebp
445: addl $STACK_SPACE, %esp
446:
447: ret
448:
449:
450: C -----------------------------------------------------------------------------
451: C
452: C Special case for q1=0xFFFFFFFF, giving q=0xFFFFFFFF meaning the low dword
453: C of q*d is simply -d and the remainder n-q*d = n10+d.
454: C
455: C This is reached only very rarely.
456:
457: L(q1_ff):
458: C eax (divisor)
459: C ebx (q1+1 == 0)
460: C ecx src pointer
461: C edx
462: C esi n10
463: C edi (n2)
464: C ebp divisor
465:
466: leal (%ebp,%esi), %edi C n-q*d remainder -> next n2
467:
468: cmpl PARAM_SRC, %ecx
469: jae L(inverse_top)
470:
471: jmp L(inverse_loop_done)
472:
473:
474: EPILOGUE()
FreeBSD-CVSweb <freebsd-cvsweb@FreeBSD.org>