blob: f619c74955e44a76fae01584cdb15456b1be3f06 [file] [log] [blame]
Edward O'Callaghan37a6a452009-08-07 20:30:09 +00001/* ===-- udivmodti4.c - Implement __udivmodti4 -----------------------------===
2 *
3 * The LLVM Compiler Infrastructure
4 *
Howard Hinnant9ad441f2010-11-16 22:13:33 +00005 * This file is dual licensed under the MIT and the University of Illinois Open
6 * Source Licenses. See LICENSE.TXT for details.
Edward O'Callaghan37a6a452009-08-07 20:30:09 +00007 *
8 * ===----------------------------------------------------------------------===
9 *
10 * This file implements __udivmodti4 for the compiler_rt library.
11 *
12 * ===----------------------------------------------------------------------===
13 */
Daniel Dunbarb3a69012009-06-26 16:47:03 +000014
Daniel Dunbarb3a69012009-06-26 16:47:03 +000015#include "int_lib.h"
16
Chandler Carruth7f2d7c72012-06-22 21:09:22 +000017#if __x86_64
18
Edward O'Callaghan37a6a452009-08-07 20:30:09 +000019/* Effects: if rem != 0, *rem = a % b
20 * Returns: a / b
21 */
Daniel Dunbarb3a69012009-06-26 16:47:03 +000022
Edward O'Callaghan37a6a452009-08-07 20:30:09 +000023/* Translated from Figure 3-40 of The PowerPC Compiler Writer's Guide */
Daniel Dunbarb3a69012009-06-26 16:47:03 +000024
25tu_int
26__udivmodti4(tu_int a, tu_int b, tu_int* rem)
27{
28 const unsigned n_udword_bits = sizeof(du_int) * CHAR_BIT;
29 const unsigned n_utword_bits = sizeof(tu_int) * CHAR_BIT;
30 utwords n;
31 n.all = a;
32 utwords d;
33 d.all = b;
34 utwords q;
35 utwords r;
36 unsigned sr;
Edward O'Callaghan37a6a452009-08-07 20:30:09 +000037 /* special cases, X is unknown, K != 0 */
Edward O'Callaghan8bf1e092009-08-09 18:41:02 +000038 if (n.s.high == 0)
Daniel Dunbarb3a69012009-06-26 16:47:03 +000039 {
Edward O'Callaghan8bf1e092009-08-09 18:41:02 +000040 if (d.s.high == 0)
Daniel Dunbarb3a69012009-06-26 16:47:03 +000041 {
Edward O'Callaghan37a6a452009-08-07 20:30:09 +000042 /* 0 X
43 * ---
44 * 0 X
45 */
Daniel Dunbarb3a69012009-06-26 16:47:03 +000046 if (rem)
Edward O'Callaghan8bf1e092009-08-09 18:41:02 +000047 *rem = n.s.low % d.s.low;
48 return n.s.low / d.s.low;
Daniel Dunbarb3a69012009-06-26 16:47:03 +000049 }
Edward O'Callaghan37a6a452009-08-07 20:30:09 +000050 /* 0 X
51 * ---
52 * K X
53 */
Daniel Dunbarb3a69012009-06-26 16:47:03 +000054 if (rem)
Edward O'Callaghan8bf1e092009-08-09 18:41:02 +000055 *rem = n.s.low;
Daniel Dunbarb3a69012009-06-26 16:47:03 +000056 return 0;
57 }
Edward O'Callaghan8bf1e092009-08-09 18:41:02 +000058 /* n.s.high != 0 */
59 if (d.s.low == 0)
Daniel Dunbarb3a69012009-06-26 16:47:03 +000060 {
Edward O'Callaghan8bf1e092009-08-09 18:41:02 +000061 if (d.s.high == 0)
Daniel Dunbarb3a69012009-06-26 16:47:03 +000062 {
Edward O'Callaghan37a6a452009-08-07 20:30:09 +000063 /* K X
64 * ---
65 * 0 0
66 */
Daniel Dunbarb3a69012009-06-26 16:47:03 +000067 if (rem)
Edward O'Callaghan8bf1e092009-08-09 18:41:02 +000068 *rem = n.s.high % d.s.low;
69 return n.s.high / d.s.low;
Daniel Dunbarb3a69012009-06-26 16:47:03 +000070 }
Edward O'Callaghan8bf1e092009-08-09 18:41:02 +000071 /* d.s.high != 0 */
72 if (n.s.low == 0)
Daniel Dunbarb3a69012009-06-26 16:47:03 +000073 {
Edward O'Callaghan37a6a452009-08-07 20:30:09 +000074 /* K 0
75 * ---
76 * K 0
77 */
Daniel Dunbarb3a69012009-06-26 16:47:03 +000078 if (rem)
79 {
Edward O'Callaghan8bf1e092009-08-09 18:41:02 +000080 r.s.high = n.s.high % d.s.high;
81 r.s.low = 0;
Daniel Dunbarb3a69012009-06-26 16:47:03 +000082 *rem = r.all;
83 }
Edward O'Callaghan8bf1e092009-08-09 18:41:02 +000084 return n.s.high / d.s.high;
Daniel Dunbarb3a69012009-06-26 16:47:03 +000085 }
Edward O'Callaghan37a6a452009-08-07 20:30:09 +000086 /* K K
87 * ---
88 * K 0
89 */
Edward O'Callaghan8bf1e092009-08-09 18:41:02 +000090 if ((d.s.high & (d.s.high - 1)) == 0) /* if d is a power of 2 */
Daniel Dunbarb3a69012009-06-26 16:47:03 +000091 {
92 if (rem)
93 {
Edward O'Callaghan8bf1e092009-08-09 18:41:02 +000094 r.s.low = n.s.low;
95 r.s.high = n.s.high & (d.s.high - 1);
Daniel Dunbarb3a69012009-06-26 16:47:03 +000096 *rem = r.all;
97 }
Edward O'Callaghan8bf1e092009-08-09 18:41:02 +000098 return n.s.high >> __builtin_ctzll(d.s.high);
Daniel Dunbarb3a69012009-06-26 16:47:03 +000099 }
Edward O'Callaghan37a6a452009-08-07 20:30:09 +0000100 /* K K
101 * ---
102 * K 0
103 */
Edward O'Callaghan8bf1e092009-08-09 18:41:02 +0000104 sr = __builtin_clzll(d.s.high) - __builtin_clzll(n.s.high);
Edward O'Callaghan37a6a452009-08-07 20:30:09 +0000105 /* 0 <= sr <= n_udword_bits - 2 or sr large */
Daniel Dunbarb3a69012009-06-26 16:47:03 +0000106 if (sr > n_udword_bits - 2)
107 {
108 if (rem)
109 *rem = n.all;
110 return 0;
111 }
112 ++sr;
Edward O'Callaghan37a6a452009-08-07 20:30:09 +0000113 /* 1 <= sr <= n_udword_bits - 1 */
114 /* q.all = n.all << (n_utword_bits - sr); */
Edward O'Callaghan8bf1e092009-08-09 18:41:02 +0000115 q.s.low = 0;
116 q.s.high = n.s.low << (n_udword_bits - sr);
Edward O'Callaghan37a6a452009-08-07 20:30:09 +0000117 /* r.all = n.all >> sr; */
Edward O'Callaghan8bf1e092009-08-09 18:41:02 +0000118 r.s.high = n.s.high >> sr;
119 r.s.low = (n.s.high << (n_udword_bits - sr)) | (n.s.low >> sr);
Daniel Dunbarb3a69012009-06-26 16:47:03 +0000120 }
Edward O'Callaghan8bf1e092009-08-09 18:41:02 +0000121 else /* d.s.low != 0 */
Daniel Dunbarb3a69012009-06-26 16:47:03 +0000122 {
Edward O'Callaghan8bf1e092009-08-09 18:41:02 +0000123 if (d.s.high == 0)
Daniel Dunbarb3a69012009-06-26 16:47:03 +0000124 {
Edward O'Callaghan37a6a452009-08-07 20:30:09 +0000125 /* K X
126 * ---
127 * 0 K
128 */
Edward O'Callaghan8bf1e092009-08-09 18:41:02 +0000129 if ((d.s.low & (d.s.low - 1)) == 0) /* if d is a power of 2 */
Daniel Dunbarb3a69012009-06-26 16:47:03 +0000130 {
131 if (rem)
Edward O'Callaghan8bf1e092009-08-09 18:41:02 +0000132 *rem = n.s.low & (d.s.low - 1);
133 if (d.s.low == 1)
Daniel Dunbarb3a69012009-06-26 16:47:03 +0000134 return n.all;
Joerg Sonnenberger4fcc9aa2011-07-28 23:15:41 +0000135 sr = __builtin_ctzll(d.s.low);
Edward O'Callaghan8bf1e092009-08-09 18:41:02 +0000136 q.s.high = n.s.high >> sr;
137 q.s.low = (n.s.high << (n_udword_bits - sr)) | (n.s.low >> sr);
Daniel Dunbarb3a69012009-06-26 16:47:03 +0000138 return q.all;
139 }
Edward O'Callaghan37a6a452009-08-07 20:30:09 +0000140 /* K X
141 * ---
142 * 0 K
143 */
Edward O'Callaghan8bf1e092009-08-09 18:41:02 +0000144 sr = 1 + n_udword_bits + __builtin_clzll(d.s.low)
145 - __builtin_clzll(n.s.high);
Edward O'Callaghan37a6a452009-08-07 20:30:09 +0000146 /* 2 <= sr <= n_utword_bits - 1
147 * q.all = n.all << (n_utword_bits - sr);
148 * r.all = n.all >> sr;
149 * if (sr == n_udword_bits)
150 * {
Edward O'Callaghan8bf1e092009-08-09 18:41:02 +0000151 * q.s.low = 0;
152 * q.s.high = n.s.low;
153 * r.s.high = 0;
154 * r.s.low = n.s.high;
Edward O'Callaghan37a6a452009-08-07 20:30:09 +0000155 * }
156 * else if (sr < n_udword_bits) // 2 <= sr <= n_udword_bits - 1
157 * {
Edward O'Callaghan8bf1e092009-08-09 18:41:02 +0000158 * q.s.low = 0;
159 * q.s.high = n.s.low << (n_udword_bits - sr);
160 * r.s.high = n.s.high >> sr;
161 * r.s.low = (n.s.high << (n_udword_bits - sr)) | (n.s.low >> sr);
Edward O'Callaghan37a6a452009-08-07 20:30:09 +0000162 * }
163 * else // n_udword_bits + 1 <= sr <= n_utword_bits - 1
164 * {
Edward O'Callaghan8bf1e092009-08-09 18:41:02 +0000165 * q.s.low = n.s.low << (n_utword_bits - sr);
166 * q.s.high = (n.s.high << (n_utword_bits - sr)) |
167 * (n.s.low >> (sr - n_udword_bits));
168 * r.s.high = 0;
169 * r.s.low = n.s.high >> (sr - n_udword_bits);
Edward O'Callaghan37a6a452009-08-07 20:30:09 +0000170 * }
171 */
Edward O'Callaghan8bf1e092009-08-09 18:41:02 +0000172 q.s.low = (n.s.low << (n_utword_bits - sr)) &
Daniel Dunbarb3a69012009-06-26 16:47:03 +0000173 ((di_int)(int)(n_udword_bits - sr) >> (n_udword_bits-1));
Edward O'Callaghan8bf1e092009-08-09 18:41:02 +0000174 q.s.high = ((n.s.low << ( n_udword_bits - sr)) &
Daniel Dunbarb3a69012009-06-26 16:47:03 +0000175 ((di_int)(int)(sr - n_udword_bits - 1) >> (n_udword_bits-1))) |
Edward O'Callaghan8bf1e092009-08-09 18:41:02 +0000176 (((n.s.high << (n_utword_bits - sr)) |
177 (n.s.low >> (sr - n_udword_bits))) &
Daniel Dunbarb3a69012009-06-26 16:47:03 +0000178 ((di_int)(int)(n_udword_bits - sr) >> (n_udword_bits-1)));
Edward O'Callaghan8bf1e092009-08-09 18:41:02 +0000179 r.s.high = (n.s.high >> sr) &
Daniel Dunbarb3a69012009-06-26 16:47:03 +0000180 ((di_int)(int)(sr - n_udword_bits) >> (n_udword_bits-1));
Edward O'Callaghan8bf1e092009-08-09 18:41:02 +0000181 r.s.low = ((n.s.high >> (sr - n_udword_bits)) &
Daniel Dunbarb3a69012009-06-26 16:47:03 +0000182 ((di_int)(int)(n_udword_bits - sr - 1) >> (n_udword_bits-1))) |
Edward O'Callaghan8bf1e092009-08-09 18:41:02 +0000183 (((n.s.high << (n_udword_bits - sr)) |
184 (n.s.low >> sr)) &
Daniel Dunbarb3a69012009-06-26 16:47:03 +0000185 ((di_int)(int)(sr - n_udword_bits) >> (n_udword_bits-1)));
186 }
187 else
188 {
Edward O'Callaghan37a6a452009-08-07 20:30:09 +0000189 /* K X
190 * ---
191 * K K
192 */
Edward O'Callaghan8bf1e092009-08-09 18:41:02 +0000193 sr = __builtin_clzll(d.s.high) - __builtin_clzll(n.s.high);
Edward O'Callaghan37a6a452009-08-07 20:30:09 +0000194 /*0 <= sr <= n_udword_bits - 1 or sr large */
Daniel Dunbarb3a69012009-06-26 16:47:03 +0000195 if (sr > n_udword_bits - 1)
196 {
197 if (rem)
198 *rem = n.all;
199 return 0;
200 }
201 ++sr;
Edward O'Callaghan37a6a452009-08-07 20:30:09 +0000202 /* 1 <= sr <= n_udword_bits */
203 /* q.all = n.all << (n_utword_bits - sr); */
Edward O'Callaghan8bf1e092009-08-09 18:41:02 +0000204 q.s.low = 0;
205 q.s.high = n.s.low << (n_udword_bits - sr);
Edward O'Callaghan37a6a452009-08-07 20:30:09 +0000206 /* r.all = n.all >> sr;
207 * if (sr < n_udword_bits)
208 * {
Edward O'Callaghan8bf1e092009-08-09 18:41:02 +0000209 * r.s.high = n.s.high >> sr;
210 * r.s.low = (n.s.high << (n_udword_bits - sr)) | (n.s.low >> sr);
Edward O'Callaghan37a6a452009-08-07 20:30:09 +0000211 * }
212 * else
213 * {
Edward O'Callaghan8bf1e092009-08-09 18:41:02 +0000214 * r.s.high = 0;
215 * r.s.low = n.s.high;
Edward O'Callaghan37a6a452009-08-07 20:30:09 +0000216 * }
217 */
Edward O'Callaghan8bf1e092009-08-09 18:41:02 +0000218 r.s.high = (n.s.high >> sr) &
Daniel Dunbarb3a69012009-06-26 16:47:03 +0000219 ((di_int)(int)(sr - n_udword_bits) >> (n_udword_bits-1));
Edward O'Callaghan8bf1e092009-08-09 18:41:02 +0000220 r.s.low = (n.s.high << (n_udword_bits - sr)) |
221 ((n.s.low >> sr) &
Daniel Dunbarb3a69012009-06-26 16:47:03 +0000222 ((di_int)(int)(sr - n_udword_bits) >> (n_udword_bits-1)));
223 }
224 }
Edward O'Callaghan37a6a452009-08-07 20:30:09 +0000225 /* Not a special case
226 * q and r are initialized with:
227 * q.all = n.all << (n_utword_bits - sr);
228 * r.all = n.all >> sr;
229 * 1 <= sr <= n_utword_bits - 1
230 */
Daniel Dunbarb3a69012009-06-26 16:47:03 +0000231 su_int carry = 0;
232 for (; sr > 0; --sr)
233 {
Edward O'Callaghan37a6a452009-08-07 20:30:09 +0000234 /* r:q = ((r:q) << 1) | carry */
Edward O'Callaghan8bf1e092009-08-09 18:41:02 +0000235 r.s.high = (r.s.high << 1) | (r.s.low >> (n_udword_bits - 1));
236 r.s.low = (r.s.low << 1) | (q.s.high >> (n_udword_bits - 1));
237 q.s.high = (q.s.high << 1) | (q.s.low >> (n_udword_bits - 1));
238 q.s.low = (q.s.low << 1) | carry;
Edward O'Callaghan37a6a452009-08-07 20:30:09 +0000239 /* carry = 0;
240 * if (r.all >= d.all)
241 * {
242 * r.all -= d.all;
243 * carry = 1;
244 * }
245 */
Daniel Dunbarb3a69012009-06-26 16:47:03 +0000246 const ti_int s = (ti_int)(d.all - r.all - 1) >> (n_utword_bits - 1);
247 carry = s & 1;
248 r.all -= d.all & s;
249 }
250 q.all = (q.all << 1) | carry;
251 if (rem)
252 *rem = r.all;
253 return q.all;
254}
255
Edward O'Callaghan8bf1e092009-08-09 18:41:02 +0000256#endif /* __x86_64 */