xref: /linux/arch/m68k/math-emu/multi_arith.h (revision 0d3b051adbb72ed81956447d0d1e54d5943ee6f5)
1 /* SPDX-License-Identifier: GPL-2.0-or-later */
2 /* multi_arith.h: multi-precision integer arithmetic functions, needed
3    to do extended-precision floating point.
4 
5    (c) 1998 David Huggins-Daines.
6 
7    Somewhat based on arch/alpha/math-emu/ieee-math.c, which is (c)
8    David Mosberger-Tang.
9 
10  */
11 
12 /* Note:
13 
14    These are not general multi-precision math routines.  Rather, they
15    implement the subset of integer arithmetic that we need in order to
16    multiply, divide, and normalize 128-bit unsigned mantissae.  */
17 
18 #ifndef MULTI_ARITH_H
19 #define MULTI_ARITH_H
20 
21 static inline void fp_denormalize(struct fp_ext *reg, unsigned int cnt)
22 {
23 	reg->exp += cnt;
24 
25 	switch (cnt) {
26 	case 0 ... 8:
27 		reg->lowmant = reg->mant.m32[1] << (8 - cnt);
28 		reg->mant.m32[1] = (reg->mant.m32[1] >> cnt) |
29 				   (reg->mant.m32[0] << (32 - cnt));
30 		reg->mant.m32[0] = reg->mant.m32[0] >> cnt;
31 		break;
32 	case 9 ... 32:
33 		reg->lowmant = reg->mant.m32[1] >> (cnt - 8);
34 		if (reg->mant.m32[1] << (40 - cnt))
35 			reg->lowmant |= 1;
36 		reg->mant.m32[1] = (reg->mant.m32[1] >> cnt) |
37 				   (reg->mant.m32[0] << (32 - cnt));
38 		reg->mant.m32[0] = reg->mant.m32[0] >> cnt;
39 		break;
40 	case 33 ... 39:
41 		asm volatile ("bfextu %1{%2,#8},%0" : "=d" (reg->lowmant)
42 			: "m" (reg->mant.m32[0]), "d" (64 - cnt));
43 		if (reg->mant.m32[1] << (40 - cnt))
44 			reg->lowmant |= 1;
45 		reg->mant.m32[1] = reg->mant.m32[0] >> (cnt - 32);
46 		reg->mant.m32[0] = 0;
47 		break;
48 	case 40 ... 71:
49 		reg->lowmant = reg->mant.m32[0] >> (cnt - 40);
50 		if ((reg->mant.m32[0] << (72 - cnt)) || reg->mant.m32[1])
51 			reg->lowmant |= 1;
52 		reg->mant.m32[1] = reg->mant.m32[0] >> (cnt - 32);
53 		reg->mant.m32[0] = 0;
54 		break;
55 	default:
56 		reg->lowmant = reg->mant.m32[0] || reg->mant.m32[1];
57 		reg->mant.m32[0] = 0;
58 		reg->mant.m32[1] = 0;
59 		break;
60 	}
61 }
62 
63 static inline int fp_overnormalize(struct fp_ext *reg)
64 {
65 	int shift;
66 
67 	if (reg->mant.m32[0]) {
68 		asm ("bfffo %1{#0,#32},%0" : "=d" (shift) : "dm" (reg->mant.m32[0]));
69 		reg->mant.m32[0] = (reg->mant.m32[0] << shift) | (reg->mant.m32[1] >> (32 - shift));
70 		reg->mant.m32[1] = (reg->mant.m32[1] << shift);
71 	} else {
72 		asm ("bfffo %1{#0,#32},%0" : "=d" (shift) : "dm" (reg->mant.m32[1]));
73 		reg->mant.m32[0] = (reg->mant.m32[1] << shift);
74 		reg->mant.m32[1] = 0;
75 		shift += 32;
76 	}
77 
78 	return shift;
79 }
80 
81 static inline int fp_addmant(struct fp_ext *dest, struct fp_ext *src)
82 {
83 	int carry;
84 
85 	/* we assume here, gcc only insert move and a clr instr */
86 	asm volatile ("add.b %1,%0" : "=d,g" (dest->lowmant)
87 		: "g,d" (src->lowmant), "0,0" (dest->lowmant));
88 	asm volatile ("addx.l %1,%0" : "=d" (dest->mant.m32[1])
89 		: "d" (src->mant.m32[1]), "0" (dest->mant.m32[1]));
90 	asm volatile ("addx.l %1,%0" : "=d" (dest->mant.m32[0])
91 		: "d" (src->mant.m32[0]), "0" (dest->mant.m32[0]));
92 	asm volatile ("addx.l %0,%0" : "=d" (carry) : "0" (0));
93 
94 	return carry;
95 }
96 
97 static inline int fp_addcarry(struct fp_ext *reg)
98 {
99 	if (++reg->exp == 0x7fff) {
100 		if (reg->mant.m64)
101 			fp_set_sr(FPSR_EXC_INEX2);
102 		reg->mant.m64 = 0;
103 		fp_set_sr(FPSR_EXC_OVFL);
104 		return 0;
105 	}
106 	reg->lowmant = (reg->mant.m32[1] << 7) | (reg->lowmant ? 1 : 0);
107 	reg->mant.m32[1] = (reg->mant.m32[1] >> 1) |
108 			   (reg->mant.m32[0] << 31);
109 	reg->mant.m32[0] = (reg->mant.m32[0] >> 1) | 0x80000000;
110 
111 	return 1;
112 }
113 
114 static inline void fp_submant(struct fp_ext *dest, struct fp_ext *src1,
115 			      struct fp_ext *src2)
116 {
117 	/* we assume here, gcc only insert move and a clr instr */
118 	asm volatile ("sub.b %1,%0" : "=d,g" (dest->lowmant)
119 		: "g,d" (src2->lowmant), "0,0" (src1->lowmant));
120 	asm volatile ("subx.l %1,%0" : "=d" (dest->mant.m32[1])
121 		: "d" (src2->mant.m32[1]), "0" (src1->mant.m32[1]));
122 	asm volatile ("subx.l %1,%0" : "=d" (dest->mant.m32[0])
123 		: "d" (src2->mant.m32[0]), "0" (src1->mant.m32[0]));
124 }
125 
126 #define fp_mul64(desth, destl, src1, src2) ({				\
127 	asm ("mulu.l %2,%1:%0" : "=d" (destl), "=d" (desth)		\
128 		: "dm" (src1), "0" (src2));				\
129 })
130 #define fp_div64(quot, rem, srch, srcl, div)				\
131 	asm ("divu.l %2,%1:%0" : "=d" (quot), "=d" (rem)		\
132 		: "dm" (div), "1" (srch), "0" (srcl))
133 #define fp_add64(dest1, dest2, src1, src2) ({				\
134 	asm ("add.l %1,%0" : "=d,dm" (dest2)				\
135 		: "dm,d" (src2), "0,0" (dest2));			\
136 	asm ("addx.l %1,%0" : "=d" (dest1)				\
137 		: "d" (src1), "0" (dest1));				\
138 })
139 #define fp_addx96(dest, src) ({						\
140 	/* we assume here, gcc only insert move and a clr instr */	\
141 	asm volatile ("add.l %1,%0" : "=d,g" (dest->m32[2])		\
142 		: "g,d" (temp.m32[1]), "0,0" (dest->m32[2]));		\
143 	asm volatile ("addx.l %1,%0" : "=d" (dest->m32[1])		\
144 		: "d" (temp.m32[0]), "0" (dest->m32[1]));		\
145 	asm volatile ("addx.l %1,%0" : "=d" (dest->m32[0])		\
146 		: "d" (0), "0" (dest->m32[0]));				\
147 })
148 #define fp_sub64(dest, src) ({						\
149 	asm ("sub.l %1,%0" : "=d,dm" (dest.m32[1])			\
150 		: "dm,d" (src.m32[1]), "0,0" (dest.m32[1]));		\
151 	asm ("subx.l %1,%0" : "=d" (dest.m32[0])			\
152 		: "d" (src.m32[0]), "0" (dest.m32[0]));			\
153 })
154 #define fp_sub96c(dest, srch, srcm, srcl) ({				\
155 	char carry;							\
156 	asm ("sub.l %1,%0" : "=d,dm" (dest.m32[2])			\
157 		: "dm,d" (srcl), "0,0" (dest.m32[2]));			\
158 	asm ("subx.l %1,%0" : "=d" (dest.m32[1])			\
159 		: "d" (srcm), "0" (dest.m32[1]));			\
160 	asm ("subx.l %2,%1; scs %0" : "=d" (carry), "=d" (dest.m32[0])	\
161 		: "d" (srch), "1" (dest.m32[0]));			\
162 	carry;								\
163 })
164 
165 static inline void fp_multiplymant(union fp_mant128 *dest, struct fp_ext *src1,
166 				   struct fp_ext *src2)
167 {
168 	union fp_mant64 temp;
169 
170 	fp_mul64(dest->m32[0], dest->m32[1], src1->mant.m32[0], src2->mant.m32[0]);
171 	fp_mul64(dest->m32[2], dest->m32[3], src1->mant.m32[1], src2->mant.m32[1]);
172 
173 	fp_mul64(temp.m32[0], temp.m32[1], src1->mant.m32[0], src2->mant.m32[1]);
174 	fp_addx96(dest, temp);
175 
176 	fp_mul64(temp.m32[0], temp.m32[1], src1->mant.m32[1], src2->mant.m32[0]);
177 	fp_addx96(dest, temp);
178 }
179 
180 static inline void fp_dividemant(union fp_mant128 *dest, struct fp_ext *src,
181 				 struct fp_ext *div)
182 {
183 	union fp_mant128 tmp;
184 	union fp_mant64 tmp64;
185 	unsigned long *mantp = dest->m32;
186 	unsigned long fix, rem, first, dummy;
187 	int i;
188 
189 	/* the algorithm below requires dest to be smaller than div,
190 	   but both have the high bit set */
191 	if (src->mant.m64 >= div->mant.m64) {
192 		fp_sub64(src->mant, div->mant);
193 		*mantp = 1;
194 	} else
195 		*mantp = 0;
196 	mantp++;
197 
198 	/* basic idea behind this algorithm: we can't divide two 64bit numbers
199 	   (AB/CD) directly, but we can calculate AB/C0, but this means this
200 	   quotient is off by C0/CD, so we have to multiply the first result
201 	   to fix the result, after that we have nearly the correct result
202 	   and only a few corrections are needed. */
203 
204 	/* C0/CD can be precalculated, but it's an 64bit division again, but
205 	   we can make it a bit easier, by dividing first through C so we get
206 	   10/1D and now only a single shift and the value fits into 32bit. */
207 	fix = 0x80000000;
208 	dummy = div->mant.m32[1] / div->mant.m32[0] + 1;
209 	dummy = (dummy >> 1) | fix;
210 	fp_div64(fix, dummy, fix, 0, dummy);
211 	fix--;
212 
213 	for (i = 0; i < 3; i++, mantp++) {
214 		if (src->mant.m32[0] == div->mant.m32[0]) {
215 			fp_div64(first, rem, 0, src->mant.m32[1], div->mant.m32[0]);
216 
217 			fp_mul64(*mantp, dummy, first, fix);
218 			*mantp += fix;
219 		} else {
220 			fp_div64(first, rem, src->mant.m32[0], src->mant.m32[1], div->mant.m32[0]);
221 
222 			fp_mul64(*mantp, dummy, first, fix);
223 		}
224 
225 		fp_mul64(tmp.m32[0], tmp.m32[1], div->mant.m32[0], first - *mantp);
226 		fp_add64(tmp.m32[0], tmp.m32[1], 0, rem);
227 		tmp.m32[2] = 0;
228 
229 		fp_mul64(tmp64.m32[0], tmp64.m32[1], *mantp, div->mant.m32[1]);
230 		fp_sub96c(tmp, 0, tmp64.m32[0], tmp64.m32[1]);
231 
232 		src->mant.m32[0] = tmp.m32[1];
233 		src->mant.m32[1] = tmp.m32[2];
234 
235 		while (!fp_sub96c(tmp, 0, div->mant.m32[0], div->mant.m32[1])) {
236 			src->mant.m32[0] = tmp.m32[1];
237 			src->mant.m32[1] = tmp.m32[2];
238 			*mantp += 1;
239 		}
240 	}
241 }
242 
243 static inline void fp_putmant128(struct fp_ext *dest, union fp_mant128 *src,
244 				 int shift)
245 {
246 	unsigned long tmp;
247 
248 	switch (shift) {
249 	case 0:
250 		dest->mant.m64 = src->m64[0];
251 		dest->lowmant = src->m32[2] >> 24;
252 		if (src->m32[3] || (src->m32[2] << 8))
253 			dest->lowmant |= 1;
254 		break;
255 	case 1:
256 		asm volatile ("lsl.l #1,%0"
257 			: "=d" (tmp) : "0" (src->m32[2]));
258 		asm volatile ("roxl.l #1,%0"
259 			: "=d" (dest->mant.m32[1]) : "0" (src->m32[1]));
260 		asm volatile ("roxl.l #1,%0"
261 			: "=d" (dest->mant.m32[0]) : "0" (src->m32[0]));
262 		dest->lowmant = tmp >> 24;
263 		if (src->m32[3] || (tmp << 8))
264 			dest->lowmant |= 1;
265 		break;
266 	case 31:
267 		asm volatile ("lsr.l #1,%1; roxr.l #1,%0"
268 			: "=d" (dest->mant.m32[0])
269 			: "d" (src->m32[0]), "0" (src->m32[1]));
270 		asm volatile ("roxr.l #1,%0"
271 			: "=d" (dest->mant.m32[1]) : "0" (src->m32[2]));
272 		asm volatile ("roxr.l #1,%0"
273 			: "=d" (tmp) : "0" (src->m32[3]));
274 		dest->lowmant = tmp >> 24;
275 		if (src->m32[3] << 7)
276 			dest->lowmant |= 1;
277 		break;
278 	case 32:
279 		dest->mant.m32[0] = src->m32[1];
280 		dest->mant.m32[1] = src->m32[2];
281 		dest->lowmant = src->m32[3] >> 24;
282 		if (src->m32[3] << 8)
283 			dest->lowmant |= 1;
284 		break;
285 	}
286 }
287 
288 #endif	/* MULTI_ARITH_H */
289