xref: /netbsd-src/external/lgpl3/gmp/dist/mpn/x86/k6/mul_basecase.asm (revision d25ffa98a4bfca1fe272f3c182496ec9934faac7)
1dnl  AMD K6 mpn_mul_basecase -- multiply two mpn numbers.
2
3dnl  Copyright 1999, 2000, 2001, 2002, 2003 Free Software Foundation, Inc.
4dnl
5dnl  This file is part of the GNU MP Library.
6dnl
7dnl  The GNU MP Library is free software; you can redistribute it and/or
8dnl  modify it under the terms of the GNU Lesser General Public License as
9dnl  published by the Free Software Foundation; either version 3 of the
10dnl  License, or (at your option) any later version.
11dnl
12dnl  The GNU MP Library is distributed in the hope that it will be useful,
13dnl  but WITHOUT ANY WARRANTY; without even the implied warranty of
14dnl  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15dnl  Lesser General Public License for more details.
16dnl
17dnl  You should have received a copy of the GNU Lesser General Public License
18dnl  along with the GNU MP Library.  If not, see http://www.gnu.org/licenses/.
19
20include(`../config.m4')
21
22
23C K6: approx 9.0 cycles per cross product on 30x30 limbs (with 16 limbs/loop
24C     unrolling).
25
26
27
28dnl  K6: UNROLL_COUNT cycles/product (approx)
29dnl           8           9.75
30dnl          16           9.3
31dnl          32           9.3
32dnl  Maximum possible with the current code is 32.
33dnl
34dnl  With 16 the inner unrolled loop fits exactly in a 256 byte block, which
35dnl  might explain it's good performance.
36
37deflit(UNROLL_COUNT, 16)
38
39
40C void mpn_mul_basecase (mp_ptr wp,
41C                        mp_srcptr xp, mp_size_t xsize,
42C                        mp_srcptr yp, mp_size_t ysize);
43C
44C Calculate xp,xsize multiplied by yp,ysize, storing the result in
45C wp,xsize+ysize.
46C
47C This routine is essentially the same as mpn/generic/mul_basecase.c, but
48C it's faster because it does most of the mpn_addmul_1() entry code only
49C once.  The saving is about 10-20% on typical sizes coming from the
50C Karatsuba multiply code.
51C
52C Enhancements:
53C
54C The mul_1 loop is about 8.5 c/l, which is slower than mpn_mul_1 at 6.25
55C c/l.  Could call mpn_mul_1 when ysize is big enough to make it worthwhile.
56C
57C The main unrolled addmul loop could be shared by mpn_addmul_1, using some
58C extra stack setups and maybe 2 or 3 wasted cycles at the end.  Code saving
59C would be 256 bytes.
60
61ifdef(`PIC',`
62deflit(UNROLL_THRESHOLD, 8)
63',`
64deflit(UNROLL_THRESHOLD, 8)
65')
66
67defframe(PARAM_YSIZE,20)
68defframe(PARAM_YP,   16)
69defframe(PARAM_XSIZE,12)
70defframe(PARAM_XP,   8)
71defframe(PARAM_WP,   4)
72
73	TEXT
74	ALIGN(32)
75PROLOGUE(mpn_mul_basecase)
76deflit(`FRAME',0)
77
78	movl	PARAM_XSIZE, %ecx
79	movl	PARAM_YP, %eax
80
81	movl	PARAM_XP, %edx
82	movl	(%eax), %eax	C yp low limb
83
84	cmpl	$2, %ecx
85	ja	L(xsize_more_than_two_limbs)
86	je	L(two_by_something)
87
88
89	C one limb by one limb
90
91	movl	(%edx), %edx	C xp low limb
92	movl	PARAM_WP, %ecx
93
94	mull	%edx
95
96	movl	%eax, (%ecx)
97	movl	%edx, 4(%ecx)
98	ret
99
100
101C -----------------------------------------------------------------------------
102L(two_by_something):
103	decl	PARAM_YSIZE
104	pushl	%ebx
105deflit(`FRAME',4)
106
107	movl	PARAM_WP, %ebx
108	pushl	%esi
109deflit(`FRAME',8)
110
111	movl	%eax, %ecx	C yp low limb
112	movl	(%edx), %eax	C xp low limb
113
114	movl	%edx, %esi	C xp
115	jnz	L(two_by_two)
116
117
118	C two limbs by one limb
119
120	mull	%ecx
121
122	movl	%eax, (%ebx)
123	movl	4(%esi), %eax
124
125	movl	%edx, %esi	C carry
126
127	mull	%ecx
128
129	addl	%eax, %esi
130	movl	%esi, 4(%ebx)
131
132	adcl	$0, %edx
133
134	movl	%edx, 8(%ebx)
135	popl	%esi
136
137	popl	%ebx
138	ret
139
140
141
142C -----------------------------------------------------------------------------
143	ALIGN(16)
144L(two_by_two):
145	C eax	xp low limb
146	C ebx	wp
147	C ecx	yp low limb
148	C edx
149	C esi	xp
150	C edi
151	C ebp
152deflit(`FRAME',8)
153
154	mull	%ecx		C xp[0] * yp[0]
155
156	push	%edi
157deflit(`FRAME',12)
158	movl	%eax, (%ebx)
159
160	movl	4(%esi), %eax
161	movl	%edx, %edi	C carry, for wp[1]
162
163	mull	%ecx		C xp[1] * yp[0]
164
165	addl	%eax, %edi
166	movl	PARAM_YP, %ecx
167
168	adcl	$0, %edx
169
170	movl	%edi, 4(%ebx)
171	movl	4(%ecx), %ecx	C yp[1]
172
173	movl	4(%esi), %eax	C xp[1]
174	movl	%edx, %edi	C carry, for wp[2]
175
176	mull	%ecx		C xp[1] * yp[1]
177
178	addl	%eax, %edi
179
180	adcl	$0, %edx
181
182	movl	(%esi), %eax	C xp[0]
183	movl	%edx, %esi	C carry, for wp[3]
184
185	mull	%ecx		C xp[0] * yp[1]
186
187	addl	%eax, 4(%ebx)
188	adcl	%edx, %edi
189	adcl	$0, %esi
190
191	movl	%edi, 8(%ebx)
192	popl	%edi
193
194	movl	%esi, 12(%ebx)
195	popl	%esi
196
197	popl	%ebx
198	ret
199
200
201C -----------------------------------------------------------------------------
202	ALIGN(16)
203L(xsize_more_than_two_limbs):
204
205C The first limb of yp is processed with a simple mpn_mul_1 style loop
206C inline.  Unrolling this doesn't seem worthwhile since it's only run once
207C (whereas the addmul below is run ysize-1 many times).  A call to the
208C actual mpn_mul_1 will be slowed down by the call and parameter pushing and
209C popping, and doesn't seem likely to be worthwhile on the typical 10-20
210C limb operations the Karatsuba code calls here with.
211
212	C eax	yp[0]
213	C ebx
214	C ecx	xsize
215	C edx	xp
216	C esi
217	C edi
218	C ebp
219deflit(`FRAME',0)
220
221	pushl	%edi		defframe_pushl(SAVE_EDI)
222	pushl	%ebp		defframe_pushl(SAVE_EBP)
223
224	movl	PARAM_WP, %edi
225	pushl	%esi		defframe_pushl(SAVE_ESI)
226
227	movl	%eax, %ebp
228	pushl	%ebx		defframe_pushl(SAVE_EBX)
229
230	leal	(%edx,%ecx,4), %ebx	C xp end
231	xorl	%esi, %esi
232
233	leal	(%edi,%ecx,4), %edi	C wp end of mul1
234	negl	%ecx
235
236
237L(mul1):
238	C eax	scratch
239	C ebx	xp end
240	C ecx	counter, negative
241	C edx	scratch
242	C esi	carry
243	C edi	wp end of mul1
244	C ebp	multiplier
245
246	movl	(%ebx,%ecx,4), %eax
247
248	mull	%ebp
249
250	addl	%esi, %eax
251	movl	$0, %esi
252
253	adcl	%edx, %esi
254
255	movl	%eax, (%edi,%ecx,4)
256	incl	%ecx
257
258	jnz	L(mul1)
259
260
261	movl	PARAM_YSIZE, %edx
262	movl	%esi, (%edi)		C final carry
263
264	movl	PARAM_XSIZE, %ecx
265	decl	%edx
266
267	jnz	L(ysize_more_than_one_limb)
268
269	popl	%ebx
270	popl	%esi
271	popl	%ebp
272	popl	%edi
273	ret
274
275
276L(ysize_more_than_one_limb):
277	cmpl	$UNROLL_THRESHOLD, %ecx
278	movl	PARAM_YP, %eax
279
280	jae	L(unroll)
281
282
283C -----------------------------------------------------------------------------
284C Simple addmul loop.
285C
286C Using ebx and edi pointing at the ends of their respective locations saves
287C a couple of instructions in the outer loop.  The inner loop is still 11
288C cycles, the same as the simple loop in aorsmul_1.asm.
289
290	C eax	yp
291	C ebx	xp end
292	C ecx	xsize
293	C edx	ysize-1
294	C esi
295	C edi	wp end of mul1
296	C ebp
297
298	movl	4(%eax), %ebp		C multiplier
299	negl	%ecx
300
301	movl	%ecx, PARAM_XSIZE	C -xsize
302	xorl	%esi, %esi		C initial carry
303
304	leal	4(%eax,%edx,4), %eax	C yp end
305	negl	%edx
306
307	movl	%eax, PARAM_YP
308	movl	%edx, PARAM_YSIZE
309
310	jmp	L(simple_outer_entry)
311
312
313	C aligning here saves a couple of cycles
314	ALIGN(16)
315L(simple_outer_top):
316	C edx	ysize counter, negative
317
318	movl	PARAM_YP, %eax		C yp end
319	xorl	%esi, %esi		C carry
320
321	movl	PARAM_XSIZE, %ecx	C -xsize
322	movl	%edx, PARAM_YSIZE
323
324	movl	(%eax,%edx,4), %ebp	C yp limb multiplier
325L(simple_outer_entry):
326	addl	$4, %edi
327
328
329L(simple_inner):
330	C eax	scratch
331	C ebx	xp end
332	C ecx	counter, negative
333	C edx	scratch
334	C esi	carry
335	C edi	wp end of this addmul
336	C ebp	multiplier
337
338	movl	(%ebx,%ecx,4), %eax
339
340	mull	%ebp
341
342	addl	%esi, %eax
343	movl	$0, %esi
344
345	adcl	$0, %edx
346	addl	%eax, (%edi,%ecx,4)
347	adcl	%edx, %esi
348
349	incl	%ecx
350	jnz	L(simple_inner)
351
352
353	movl	PARAM_YSIZE, %edx
354	movl	%esi, (%edi)
355
356	incl	%edx
357	jnz	L(simple_outer_top)
358
359
360	popl	%ebx
361	popl	%esi
362	popl	%ebp
363	popl	%edi
364	ret
365
366
367C -----------------------------------------------------------------------------
368C Unrolled loop.
369C
370C The unrolled inner loop is the same as in aorsmul_1.asm, see that code for
371C some comments.
372C
373C VAR_COUNTER is for the inner loop, running from VAR_COUNTER_INIT down to
374C 0, inclusive.
375C
376C VAR_JMP is the computed jump into the unrolled loop.
377C
378C PARAM_XP and PARAM_WP get offset appropriately for where the unrolled loop
379C is entered.
380C
381C VAR_XP_LOW is the least significant limb of xp, which is needed at the
382C start of the unrolled loop.  This can't just be fetched through the xp
383C pointer because of the offset applied to it.
384C
385C PARAM_YSIZE is the outer loop counter, going from -(ysize-1) up to -1,
386C inclusive.
387C
388C PARAM_YP is offset appropriately so that the PARAM_YSIZE counter can be
389C added to give the location of the next limb of yp, which is the multiplier
390C in the unrolled loop.
391C
392C PARAM_WP is similarly offset so that the PARAM_YSIZE counter can be added
393C to give the starting point in the destination for each unrolled loop (this
394C point is one limb upwards for each limb of yp processed).
395C
396C Having PARAM_YSIZE count negative to zero means it's not necessary to
397C store new values of PARAM_YP and PARAM_WP on each loop.  Those values on
398C the stack remain constant and on each loop an leal adjusts them with the
399C PARAM_YSIZE counter value.
400
401
402defframe(VAR_COUNTER,      -20)
403defframe(VAR_COUNTER_INIT, -24)
404defframe(VAR_JMP,          -28)
405defframe(VAR_XP_LOW,       -32)
406deflit(VAR_STACK_SPACE, 16)
407
408dnl  For some strange reason using (%esp) instead of 0(%esp) is a touch
409dnl  slower in this code, hence the defframe empty-if-zero feature is
410dnl  disabled.
411dnl
412dnl  If VAR_COUNTER is at (%esp), the effect is worse.  In this case the
413dnl  unrolled loop is 255 instead of 256 bytes, but quite how this affects
414dnl  anything isn't clear.
415dnl
416define(`defframe_empty_if_zero_disabled',1)
417
418L(unroll):
419	C eax	yp (not used)
420	C ebx	xp end (not used)
421	C ecx	xsize
422	C edx	ysize-1
423	C esi
424	C edi	wp end of mul1 (not used)
425	C ebp
426deflit(`FRAME', 16)
427
428	leal	-2(%ecx), %ebp	C one limb processed at start,
429	decl	%ecx		C and ebp is one less
430
431	shrl	$UNROLL_LOG2, %ebp
432	negl	%ecx
433
434	subl	$VAR_STACK_SPACE, %esp
435deflit(`FRAME', 16+VAR_STACK_SPACE)
436	andl	$UNROLL_MASK, %ecx
437
438	movl	%ecx, %esi
439	shll	$4, %ecx
440
441	movl	%ebp, VAR_COUNTER_INIT
442	negl	%esi
443
444	C 15 code bytes per limb
445ifdef(`PIC',`
446	call	L(pic_calc)
447L(unroll_here):
448',`
449	leal	L(unroll_entry) (%ecx,%esi,1), %ecx
450')
451
452	movl	PARAM_XP, %ebx
453	movl	%ebp, VAR_COUNTER
454
455	movl	PARAM_WP, %edi
456	movl	%ecx, VAR_JMP
457
458	movl	(%ebx), %eax
459	leal	4(%edi,%esi,4), %edi	C wp adjust for unrolling and mul1
460
461	leal	(%ebx,%esi,4), %ebx	C xp adjust for unrolling
462
463	movl	%eax, VAR_XP_LOW
464
465	movl	%ebx, PARAM_XP
466	movl	PARAM_YP, %ebx
467
468	leal	(%edi,%edx,4), %ecx	C wp adjust for ysize indexing
469	movl	4(%ebx), %ebp		C multiplier (yp second limb)
470
471	leal	4(%ebx,%edx,4), %ebx	C yp adjust for ysize indexing
472
473	movl	%ecx, PARAM_WP
474
475	leal	1(%esi), %ecx	C adjust parity for decl %ecx above
476
477	movl	%ebx, PARAM_YP
478	negl	%edx
479
480	movl	%edx, PARAM_YSIZE
481	jmp	L(unroll_outer_entry)
482
483
484ifdef(`PIC',`
485L(pic_calc):
486	C See mpn/x86/README about old gas bugs
487	leal	(%ecx,%esi,1), %ecx
488	addl	$L(unroll_entry)-L(unroll_here), %ecx
489	addl	(%esp), %ecx
490	ret_internal
491')
492
493
494C -----------------------------------------------------------------------------
495	C Aligning here saves a couple of cycles per loop.  Using 32 doesn't
496	C cost any extra space, since the inner unrolled loop below is
497	C aligned to 32.
498	ALIGN(32)
499L(unroll_outer_top):
500	C edx	ysize
501
502	movl	PARAM_YP, %eax
503	movl	%edx, PARAM_YSIZE	C incremented ysize counter
504
505	movl	PARAM_WP, %edi
506
507	movl	VAR_COUNTER_INIT, %ebx
508	movl	(%eax,%edx,4), %ebp	C next multiplier
509
510	movl	PARAM_XSIZE, %ecx
511	leal	(%edi,%edx,4), %edi	C adjust wp for where we are in yp
512
513	movl	VAR_XP_LOW, %eax
514	movl	%ebx, VAR_COUNTER
515
516L(unroll_outer_entry):
517	mull	%ebp
518
519	C using testb is a tiny bit faster than testl
520	testb	$1, %cl
521
522	movl	%eax, %ecx	C low carry
523	movl	VAR_JMP, %eax
524
525	movl	%edx, %esi	C high carry
526	movl	PARAM_XP, %ebx
527
528	jnz	L(unroll_noswap)
529	movl	%ecx, %esi	C high,low carry other way around
530
531	movl	%edx, %ecx
532L(unroll_noswap):
533
534	jmp	*%eax
535
536
537
538C -----------------------------------------------------------------------------
539	ALIGN(32)
540L(unroll_top):
541	C eax	scratch
542	C ebx	xp
543	C ecx	carry low
544	C edx	scratch
545	C esi	carry high
546	C edi	wp
547	C ebp	multiplier
548	C VAR_COUNTER  loop counter
549	C
550	C 15 code bytes each limb
551
552	leal	UNROLL_BYTES(%edi), %edi
553
554L(unroll_entry):
555deflit(CHUNK_COUNT,2)
556forloop(`i', 0, UNROLL_COUNT/CHUNK_COUNT-1, `
557	deflit(`disp0', eval(i*CHUNK_COUNT*4))
558	deflit(`disp1', eval(disp0 + 4))
559	deflit(`disp2', eval(disp1 + 4))
560
561	movl	disp1(%ebx), %eax
562	mull	%ebp
563Zdisp(	addl,	%ecx, disp0,(%edi))
564	adcl	%eax, %esi
565	movl	%edx, %ecx
566	jadcl0( %ecx)
567
568	movl	disp2(%ebx), %eax
569	mull	%ebp
570	addl	%esi, disp1(%edi)
571	adcl	%eax, %ecx
572	movl	%edx, %esi
573	jadcl0( %esi)
574')
575
576	decl	VAR_COUNTER
577	leal	UNROLL_BYTES(%ebx), %ebx
578
579	jns	L(unroll_top)
580
581
582	movl	PARAM_YSIZE, %edx
583	addl	%ecx, UNROLL_BYTES(%edi)
584
585	adcl	$0, %esi
586
587	incl	%edx
588	movl	%esi, UNROLL_BYTES+4(%edi)
589
590	jnz	L(unroll_outer_top)
591
592
593	movl	SAVE_ESI, %esi
594	movl	SAVE_EBP, %ebp
595	movl	SAVE_EDI, %edi
596	movl	SAVE_EBX, %ebx
597
598	addl	$FRAME, %esp
599	ret
600
601EPILOGUE()
602