1; mc88100 __gmpn_mul_1 -- Multiply a limb vector with a single limb and
2; store the product in a second limb vector.
3
4; Copyright 1992, 1994, 1995, 2000 Free Software Foundation, Inc.
5
6; This file is part of the GNU MP Library.
7
8; The GNU MP Library is free software; you can redistribute it and/or modify
9; it under the terms of the GNU Lesser General Public License as published by
10; the Free Software Foundation; either version 3 of the License, or (at your
11; option) any later version.
12
13; The GNU MP Library is distributed in the hope that it will be useful, but
14; WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
15; or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU Lesser General Public
16; License for more details.
17
18; You should have received a copy of the GNU Lesser General Public License
19; along with the GNU MP Library.  If not, see http://www.gnu.org/licenses/.
20
21
22; INPUT PARAMETERS
23; res_ptr	r2
24; s1_ptr	r3
25; size		r4
26; s2_limb	r5
27
28; Common overhead is about 11 cycles/invocation.
29
30; The speed for S2_LIMB >= 0x10000 is approximately 21 cycles/limb.  (The
31; pipeline stalls 2 cycles due to WB contention.)
32
33; The speed for S2_LIMB < 0x10000 is approximately 16 cycles/limb.  (The
34; pipeline stalls 2 cycles due to WB contention and 1 cycle due to latency.)
35
36; To enhance speed:
37; 1. Unroll main loop 4-8 times.
38; 2. Schedule code to avoid WB contention.  It might be tempting to move the
39;    ld instruction in the loops down to save 2 cycles (less WB contention),
40;    but that looses because the ultimate value will be read from outside
41;    the allocated space.  But if we handle the ultimate multiplication in
42;    the tail, we can do this.
43; 3. Make the multiplication with less instructions.  I think the code for
44;    (S2_LIMB >= 0x10000) is not minimal.
45; With these techniques the (S2_LIMB >= 0x10000) case would run in 17 or
46; less cycles/limb; the (S2_LIMB < 0x10000) case would run in 11
47; cycles/limb.  (Assuming infinite unrolling.)
48
49	text
50	align	 16
51	global	 ___gmpn_mul_1
52___gmpn_mul_1:
53
54	; Make S1_PTR and RES_PTR point at the end of their blocks
55	; and negate SIZE.
56	lda	 r3,r3[r4]
57	lda	 r6,r2[r4]	; RES_PTR in r6 since r2 is retval
58	subu	 r4,r0,r4
59
60	addu.co	 r2,r0,r0	; r2 = cy = 0
61	ld	 r9,r3[r4]
62	mask	 r7,r5,0xffff	; r7 = lo(S2_LIMB)
63	extu	 r8,r5,16	; r8 = hi(S2_LIMB)
64	bcnd.n	 eq0,r8,Lsmall	; jump if (hi(S2_LIMB) == 0)
65	 subu	 r6,r6,4
66
67; General code for any value of S2_LIMB.
68
69	; Make a stack frame and save r25 and r26
70	subu	 r31,r31,16
71	st.d	 r25,r31,8
72
73	; Enter the loop in the middle
74	br.n	L1
75	addu	 r4,r4,1
76
77Loop:	ld	 r9,r3[r4]
78	st	 r26,r6[r4]
79; bcnd	ne0,r0,0		; bubble
80	addu	 r4,r4,1
81L1:	mul	 r26,r9,r5	; low word of product	mul_1	WB ld
82	mask	 r12,r9,0xffff	; r12 = lo(s1_limb)	mask_1
83	mul	 r11,r12,r7	; r11 =  prod_0		mul_2	WB mask_1
84	mul	 r10,r12,r8	; r10 = prod_1a		mul_3
85	extu	 r13,r9,16	; r13 = hi(s1_limb)	extu_1	WB mul_1
86	mul	 r12,r13,r7	; r12 = prod_1b		mul_4	WB extu_1
87	mul	 r25,r13,r8	; r25  = prod_2		mul_5	WB mul_2
88	extu	 r11,r11,16	; r11 = hi(prod_0)	extu_2	WB mul_3
89	addu	 r10,r10,r11	;			addu_1	WB extu_2
90; bcnd	ne0,r0,0		; bubble			WB addu_1
91	addu.co	 r10,r10,r12	;				WB mul_4
92	mask.u	 r10,r10,0xffff	; move the 16 most significant bits...
93	addu.ci	 r10,r10,r0	; ...to the low half of the word...
94	rot	 r10,r10,16	; ...and put carry in pos 16.
95	addu.co	 r26,r26,r2	; add old carry limb
96	bcnd.n	 ne0,r4,Loop
97	 addu.ci r2,r25,r10	; compute new carry limb
98
99	st	 r26,r6[r4]
100	ld.d	 r25,r31,8
101	jmp.n	 r1
102	 addu	 r31,r31,16
103
104; Fast code for S2_LIMB < 0x10000
105Lsmall:
106	; Enter the loop in the middle
107	br.n	SL1
108	addu	 r4,r4,1
109
110SLoop:	ld	 r9,r3[r4]	;
111	st	 r8,r6[r4]	;
112	addu	 r4,r4,1	;
113SL1:	mul	 r8,r9,r5	; low word of product
114	mask	 r12,r9,0xffff	; r12 = lo(s1_limb)
115	extu	 r13,r9,16	; r13 = hi(s1_limb)
116	mul	 r11,r12,r7	; r11 =  prod_0
117	mul	 r12,r13,r7	; r12 = prod_1b
118	addu.cio r8,r8,r2	; add old carry limb
119	extu	 r10,r11,16	; r11 = hi(prod_0)
120	addu	 r10,r10,r12	;
121	bcnd.n	 ne0,r4,SLoop
122	extu	 r2,r10,16	; r2 = new carry limb
123
124	jmp.n	 r1
125	st	 r8,r6[r4]
126